adding backend agnostic IR

.clang-format

@@ -0,0 +1,295 @@
+---
+BasedOnStyle: Mozilla
+AccessModifierOffset: 0
+AlignAfterOpenBracket: Align
+AlignArrayOfStructures: None
+AlignConsecutiveAssignments:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: false
+  AlignFunctionPointers: false
+  PadOperators: true
+AlignConsecutiveBitFields:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: false
+  AlignFunctionPointers: false
+  PadOperators: false
+AlignConsecutiveDeclarations:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: true
+  AlignFunctionPointers: false
+  PadOperators: false
+AlignConsecutiveMacros:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: false
+  AlignFunctionPointers: false
+  PadOperators: false
+AlignConsecutiveShortCaseStatements:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCaseArrows: false
+  AlignCaseColons: false
+AlignConsecutiveTableGenBreakingDAGArgColons:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: false
+  AlignFunctionPointers: false
+  PadOperators: false
+AlignConsecutiveTableGenCondOperatorColons:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: false
+  AlignFunctionPointers: false
+  PadOperators: false
+AlignConsecutiveTableGenDefinitionColons:
+  Enabled: false
+  AcrossEmptyLines: false
+  AcrossComments: false
+  AlignCompound: false
+  AlignFunctionDeclarations: false
+  AlignFunctionPointers: false
+  PadOperators: false
+AlignEscapedNewlines: Right
+AlignOperands: Align
+AlignTrailingComments:
+  Kind: Always
+  OverEmptyLines: 0
+AllowAllArgumentsOnNextLine: true
+AllowAllParametersOfDeclarationOnNextLine: false
+AllowBreakBeforeNoexceptSpecifier: Never
+AllowShortBlocksOnASingleLine: Never
+AllowShortCaseExpressionOnASingleLine: true
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortCompoundRequirementOnASingleLine: true
+AllowShortEnumsOnASingleLine: true
+AllowShortFunctionsOnASingleLine: None
+AllowShortIfStatementsOnASingleLine: Never
+AllowShortLambdasOnASingleLine: All
+AllowShortLoopsOnASingleLine: false
+AllowShortNamespacesOnASingleLine: false
+AlwaysBreakAfterDefinitionReturnType: None
+AlwaysBreakBeforeMultilineStrings: false
+AttributeMacros:
+  - __capability
+BinPackArguments: false
+BinPackLongBracedList: true
+BinPackParameters: OnePerLine
+BitFieldColonSpacing: Both
+BraceWrapping:
+  AfterCaseLabel: true
+  AfterClass: true
+  AfterControlStatement: Never
+  AfterEnum: true
+  AfterFunction: true
+  AfterNamespace: true
+  AfterObjCDeclaration: false
+  AfterStruct: true
+  AfterUnion: true
+  AfterExternBlock: true
+  BeforeCatch: true
+  BeforeElse: true
+  BeforeLambdaBody: false
+  BeforeWhile: true
+  IndentBraces: true
+  SplitEmptyFunction: true
+  SplitEmptyRecord: false
+  SplitEmptyNamespace: true
+BracedInitializerIndentWidth: -1
+BreakAdjacentStringLiterals: true
+BreakAfterAttributes: Leave
+BreakAfterJavaFieldAnnotations: false
+BreakAfterReturnType: None
+BreakArrays: true
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Allman
+BreakBeforeConceptDeclarations: Always
+BreakBeforeInlineASMColon: OnlyMultiline
+BreakBeforeTemplateCloser: false
+BreakBeforeTernaryOperators: true
+BreakBinaryOperations: Never
+BreakConstructorInitializers: BeforeComma
+BreakFunctionDefinitionParameters: false
+BreakInheritanceList: BeforeComma
+BreakStringLiterals: true
+BreakTemplateDeclarations: Yes
+ColumnLimit: 128
+CommentPragmas: "^ IWYU pragma:"
+CompactNamespaces: false
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: false
+DerivePointerAlignment: false
+DisableFormat: false
+EmptyLineAfterAccessModifier: Never
+EmptyLineBeforeAccessModifier: LogicalBlock
+EnumTrailingComma: Leave
+ExperimentalAutoDetectBinPacking: false
+FixNamespaceComments: false
+ForEachMacros:
+  - foreach
+  - Q_FOREACH
+  - BOOST_FOREACH
+IfMacros:
+  - KJ_IF_MAYBE
+IncludeBlocks: Preserve
+IncludeCategories:
+  - Regex: ^"(llvm|llvm-c|clang|clang-c)/
+    Priority: 2
+    SortPriority: 0
+    CaseSensitive: false
+  - Regex: ^(<|"(gtest|gmock|isl|json)/)
+    Priority: 3
+    SortPriority: 0
+    CaseSensitive: false
+  - Regex: .*
+    Priority: 1
+    SortPriority: 0
+    CaseSensitive: false
+IncludeIsMainRegex: (Test)?$
+IncludeIsMainSourceRegex: ""
+IndentAccessModifiers: false
+IndentCaseBlocks: true
+IndentCaseLabels: true
+IndentExportBlock: true
+IndentExternBlock: AfterExternBlock
+IndentGotoLabels: true
+IndentPPDirectives: None
+IndentRequiresClause: true
+IndentWidth: 4
+IndentWrappedFunctionNames: false
+InsertBraces: false
+InsertNewlineAtEOF: false
+InsertTrailingCommas: None
+IntegerLiteralSeparator:
+  Binary: 0
+  BinaryMinDigits: 0
+  Decimal: 0
+  DecimalMinDigits: 0
+  Hex: 0
+  HexMinDigits: 0
+JavaScriptQuotes: Leave
+JavaScriptWrapImports: true
+KeepEmptyLines:
+  AtEndOfFile: false
+  AtStartOfBlock: true
+  AtStartOfFile: true
+KeepFormFeed: false
+LambdaBodyIndentation: Signature
+LineEnding: DeriveLF
+MacroBlockBegin: ""
+MacroBlockEnd: ""
+MainIncludeChar: Quote
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: All
+ObjCBinPackProtocolList: Auto
+ObjCBlockIndentWidth: 4
+ObjCBreakBeforeNestedBlockParam: true
+ObjCSpaceAfterProperty: true
+ObjCSpaceBeforeProtocolList: false
+OneLineFormatOffRegex: ""
+PPIndentWidth: -1
+PackConstructorInitializers: BinPack
+PenaltyBreakAssignment: 2
+PenaltyBreakBeforeFirstCallParameter: 19
+PenaltyBreakBeforeMemberAccess: 150
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakOpenParenthesis: 0
+PenaltyBreakScopeResolution: 500
+PenaltyBreakString: 1000
+PenaltyBreakTemplateDeclaration: 10
+PenaltyExcessCharacter: 1000000
+PenaltyIndentedWhitespace: 0
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Left
+QualifierAlignment: Leave
+ReferenceAlignment: Pointer
+ReflowComments: Always
+RemoveBracesLLVM: false
+RemoveEmptyLinesInUnwrappedLines: false
+RemoveParentheses: Leave
+RemoveSemicolon: false
+RequiresClausePosition: OwnLine
+RequiresExpressionIndentation: OuterScope
+SeparateDefinitionBlocks: Leave
+ShortNamespaceLines: 1
+SkipMacroDefinitionBody: false
+SortIncludes:
+  Enabled: true
+  IgnoreCase: false
+SortJavaStaticImport: Before
+SortUsingDeclarations: LexicographicNumeric
+SpaceAfterCStyleCast: false
+SpaceAfterLogicalNot: false
+SpaceAfterOperatorKeyword: false
+SpaceAfterTemplateKeyword: false
+SpaceAroundPointerQualifiers: Default
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeCaseColon: false
+SpaceBeforeCpp11BracedList: false
+SpaceBeforeCtorInitializerColon: true
+SpaceBeforeInheritanceColon: true
+SpaceBeforeJsonColon: false
+SpaceBeforeParens: Never
+SpaceBeforeParensOptions:
+  AfterControlStatements: false
+  AfterForeachMacros: false
+  AfterFunctionDeclarationName: false
+  AfterFunctionDefinitionName: false
+  AfterIfMacros: true
+  AfterNot: false
+  AfterOverloadedOperator: false
+  AfterPlacementOperator: true
+  AfterRequiresInClause: false
+  AfterRequiresInExpression: false
+  BeforeNonEmptyParentheses: false
+SpaceBeforeRangeBasedForLoopColon: false
+SpaceBeforeSquareBrackets: false
+SpaceInEmptyBlock: false
+SpacesBeforeTrailingComments: 1
+SpacesInAngles: Never
+SpacesInContainerLiterals: true
+SpacesInLineCommentPrefix:
+  Minimum: 1
+  Maximum: -1
+SpacesInParens: Never
+SpacesInParensOptions:
+  ExceptDoubleParentheses: false
+  InConditionalStatements: false
+  InCStyleCasts: false
+  InEmptyParentheses: false
+  Other: false
+SpacesInSquareBrackets: false
+Standard: Latest
+StatementAttributeLikeMacros:
+  - Q_EMIT
+StatementMacros:
+  - Q_UNUSED
+  - QT_REQUIRE_VERSION
+TabWidth: 4
+TableGenBreakInsideDAGArg: DontBreak
+UseTab: ForIndentation
+VerilogBreakBetweenInstancePorts: true
+WhitespaceSensitiveMacros:
+  - BOOST_PP_STRINGIZE
+  - CF_SWIFT_NAME
+  - NS_SWIFT_NAME
+  - PP_STRINGIZE
+  - STRINGIZE
+WrapNamespaceBodyWithEmptyLines: Leave

.gitea/workflows/Build.yml

@@ -18,14 +18,54 @@ jobs:
 
       - uses: https://codeberg.org/mlugg/setup-zig@v2
 
+      - name: Install MPSS
+        run: |
+          set -euxo pipefail
+
+          apt-get update
+          apt-get install -y unzip curl
+
+          MPSS_URL="https://drive.kbz8.me/dav/public-files/KSKeiOKxnFPyfws/mpss.zip"
+          MPSS_ZIP="/tmp/mpss.zip"
+          MPSS_BIN="/opt/mpss/3.8.6/sysroots/k1om-mpss-linux/bin"
+
+          curl --fail --location --show-error --progress-bar "$MPSS_URL" -o "$MPSS_ZIP"
+
+          size="$(stat -c%s "$MPSS_ZIP")"
+          test "$size" -gt 300000000
+
+          unzip -t "$MPSS_ZIP" >/dev/null
+
+          rm -rf /opt/mpss
+          unzip -q "$MPSS_ZIP" -d /opt
+          chmod -R a+rX /opt/mpss
+
+          test -d "$MPSS_BIN"
+
+          # Append as fallback for following CI steps.
+          echo "PATH=$PATH:$MPSS_BIN" >> "$GITHUB_ENV"
+
+          echo "MPSS installed."
+
+      - name: Check MPSS PATH
+        run: |
+          echo "$PATH"
+          ls -la /opt/mpss/3.8.6/sysroots/k1om-mpss-linux/bin | head -50
+
+      - name: ZLint pass
+        uses: DonIsaac/zlint-action@v0.0.1
+
       - name: Building Ape
-        run: zig build ape
+        run: zig build ape --release=safe
 
       - name: Building Soft
-        run: zig build soft
+        run: zig build soft --release=safe
 
       - name: Building Flint
-        run: zig build flint
+        run: zig build flint --release=safe
+
+      - name: Building Phi
+        run: zig build phi --release=safe
 
       - name: Generating docs
         run: zig build docs
@@ -42,3 +82,19 @@ jobs:
           local-path: "./zig-out/docs"
           remote-path: "/www"
           sync: full
+
+      - name: Generating IR docs
+        run: zig build docs-ir
+
+      - name: Deploying IR docs
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: milanmk/actions-file-deployer@master
+        with:
+          remote-protocol: sftp
+          remote-host: ${{ secrets.SFTP_HOST_DOCS_IR }}
+          remote-user: ${{ secrets.SFTP_USER_DOCS_IR }}
+          remote-password: ${{ secrets.SFTP_PASSWORD_DOCS_IR  }}
+          remote-port: 6969
+          local-path: "./zig-out/docs-ir"
+          remote-path: "/www"
+          sync: full

.gitea/workflows/Test.yml

@@ -15,11 +15,54 @@ jobs:
       - uses: actions/checkout@v4
       - uses: https://codeberg.org/mlugg/setup-zig@v2
 
+      - name: Install MPSS
+        run: |
+          set -euxo pipefail
+
+          apt-get update
+          apt-get install -y unzip curl
+
+          MPSS_URL="https://drive.kbz8.me/dav/public-files/KSKeiOKxnFPyfws/mpss.zip"
+          MPSS_ZIP="/tmp/mpss.zip"
+          MPSS_BIN="/opt/mpss/3.8.6/sysroots/k1om-mpss-linux/bin"
+
+          curl --fail --location --show-error --progress-bar "$MPSS_URL" -o "$MPSS_ZIP"
+
+          size="$(stat -c%s "$MPSS_ZIP")"
+          test "$size" -gt 300000000
+
+          unzip -t "$MPSS_ZIP" >/dev/null
+
+          rm -rf /opt/mpss
+          unzip -q "$MPSS_ZIP" -d /opt
+          chmod -R a+rX /opt/mpss
+
+          test -d "$MPSS_BIN"
+
+          # Append as fallback for following CI steps.
+          echo "PATH=$PATH:$MPSS_BIN" >> "$GITHUB_ENV"
+
+          echo "MPSS installed."
+
+      - name: Check MPSS PATH
+        run: |
+          echo "$PATH"
+          ls -la /opt/mpss/3.8.6/sysroots/k1om-mpss-linux/bin | head -50
+
+      - name: Installing dependencies
+        run: apt update && apt install spirv-tools
+
       - name: Ape Tests
-      run: zig build test-ape
+        run: zig build test-ape --release=safe
 
       - name: Soft Tests
-      run: zig build test-soft
+        run: zig build test-soft --release=safe
 
       - name: Flint Tests
-      run: zig build test-flint
+        run: zig build test-flint --release=safe
+
+      - name: Phi Tests
+        run: zig build test-phi --release=safe
+
+      - name: IR Tests
+        run: zig build test-ir --release=safe

.gitignore

@@ -8,8 +8,6 @@ scripts/__pycache__/
 *.o
 .gdb_history
 *.json
-*.png
-!logo.png
 *.bin
 *.qpa
 *.xml

README.md

@@ -14,7 +14,7 @@ To understand Vulkan - not as a humble API mere mortals call upon, but as a laby
 It does not seek to produce a performant or production-worthy driver. \
 *The gods are merciful, but not that merciful.*
 
-## Soft [software implementation]
+## Soft [software]
 
 Soft be a software implementation of the Vulkan specification, abiding within this driver's own codebase.\
 It maketh use of a bespoke [SPIR-V interpreter](https://git.kbz8.me/kbz_8/SPIRV-Interpreter) and renderer, by whose workings its labours are carried forth.
@@ -89,11 +89,11 @@ vkCmdResetEvent                                  | ✅ Implemented
 vkCmdResetQueryPool                              | ✅ Implemented
 vkCmdResolveImage                                | ✅ Implemented
 vkCmdSetBlendConstants                           | ✅ Implemented
-vkCmdSetDepthBias                                | ⚙️ WIP
-vkCmdSetDepthBounds                              | ⚙️ WIP
+vkCmdSetDepthBias                                | ✅ Implemented
+vkCmdSetDepthBounds                              | ✅ Implemented
 vkCmdSetDeviceMaskKHR                            | ✅ Implemented
 vkCmdSetEvent                                    | ✅ Implemented
-vkCmdSetLineWidth                                | ⚙️ WIP
+vkCmdSetLineWidth                                | ✅ Implemented
 vkCmdSetScissor                                  | ✅ Implemented
 vkCmdSetStencilCompareMask                       | ✅ Implemented
 vkCmdSetStencilReference                         | ✅ Implemented
@@ -101,7 +101,7 @@ vkCmdSetStencilWriteMask                         | ✅ Implemented
 vkCmdSetViewport                                 | ✅ Implemented
 vkCmdUpdateBuffer                                | ✅ Implemented
 vkCmdWaitEvents                                  | ✅ Implemented
-vkCmdWriteTimestamp                              | ⚙️ WIP
+vkCmdWriteTimestamp                              | ✅ Implemented
 vkCreateBuffer                                   | ✅ Implemented
 vkCreateBufferView                               | ✅ Implemented
 vkCreateCommandPool                              | ✅ Implemented
@@ -162,11 +162,14 @@ vkFlushMappedMemoryRanges                        | ✅ Implemented
 vkFreeCommandBuffers                             | ✅ Implemented
 vkFreeDescriptorSets                             | ✅ Implemented
 vkFreeMemory                                     | ✅ Implemented
+vkGetBufferDeviceAddress                         | ✅ Implemented
+vkGetBufferDeviceAddressEXT                      | ✅ Implemented
+vkGetBufferDeviceAddressKHR                      | ✅ Implemented
 vkGetBufferMemoryRequirements                    | ✅ Implemented
 vkGetDeviceGroupPeerMemoryFeaturesKHR            | ✅ Implemented
 vkGetDeviceGroupPresentCapabilitiesKHR           | ✅ Implemented
 vkGetDeviceGroupSurfacePresentModesKHR           | ✅ Implemented
-vkGetDeviceMemoryCommitment                      | ⚙️ WIP
+vkGetDeviceMemoryCommitment                      | ✅ Implemented
 vkGetDeviceProcAddr                              | ✅ Implemented
 vkGetDeviceQueue                                 | ✅ Implemented
 vkGetEventStatus                                 | ✅ Implemented
@@ -215,6 +218,406 @@ vkWaitForFences                                  | ✅ Implemented
 
 [Here](https://vulkan-driver.kbz8.me/cts/soft/) shalt thou find a most meticulous account of the Vulkan 1.0 conformance trials, set forth for thy scrutiny.
 
+## Phi [Xeon Phi KNC]
+
+Phi be an implementation of the Vulkan specification, wrought for the Xeon Phi Knights Corner cards.
+Whether the Knights Landing cards shall one day receive the same providence remaineth to be seen.
+
+### Build
+
+To bring forth the driver:
+```
+zig build phi --release=[fast|safe|small]
+```
+
+And for those who seek another manner of building:
+```
+zig build --help
+```
+
+The same rites apply as for the Soft build.
+
+#### Vulkan 1.0 specification
+<details>
+    <summary>
+        Here standeth the present reckoning of thy Vulkan 1.0 implementation
+    </summary>
+
+\
+⚠️ Implemented, yet perchance not fully tested nor proven conformant, but rather working in a manner most general to thee and thine.\
+Assume thou that functions lacking in this array are, for now, not intended to be wrought.
+
+Name                                             | Status
+-------------------------------------------------|--------
+vkAcquireNextImage2KHR                           | ⚙️ WIP
+vkAcquireNextImageKHR                            | ⚙️ WIP
+vkAllocateCommandBuffers                         | ✅ Implemented
+vkAllocateDescriptorSets                         | ⚙️ WIP
+vkAllocateMemory                                 | ✅ Implemented
+vkBeginCommandBuffer                             | ✅ Implemented
+vkBindBufferMemory                               | ⚙️ WIP
+vkBindImageMemory                                | ✅ Implemented
+vkCmdBeginQuery                                  | ⚙️ WIP
+vkCmdBeginRenderPass                             | ⚙️ WIP
+vkCmdBindDescriptorSets                          | ⚙️ WIP
+vkCmdBindIndexBuffer                             | ⚙️ WIP
+vkCmdBindPipeline                                | ⚙️ WIP
+vkCmdBindVertexBuffers                           | ⚙️ WIP
+vkCmdBlitImage                                   | ✅ Implemented
+vkCmdClearAttachments                            | ⚙️ WIP
+vkCmdClearColorImage                             | ⚙️ WIP
+vkCmdClearDepthStencilImage                      | ⚙️ WIP
+vkCmdCopyBuffer                                  | ✅ Implemented
+vkCmdCopyBufferToImage                           | ✅ Implemented
+vkCmdCopyImage                                   | ⚙️ WIP
+vkCmdCopyImageToBuffer                           | ✅ Implemented
+vkCmdCopyQueryPoolResults                        | ⚙️ WIP
+vkCmdDispatch                                    | ⚙️ WIP
+vkCmdDispatchBaseKHR                             | ⚙️ WIP
+vkCmdDispatchIndirect                            | ⚙️ WIP
+vkCmdDraw                                        | ⚙️ WIP
+vkCmdDrawIndexed                                 | ⚙️ WIP
+vkCmdDrawIndexedIndirect                         | ⚙️ WIP
+vkCmdDrawIndirect                                | ⚙️ WIP
+vkCmdEndQuery                                    | ⚙️ WIP
+vkCmdEndRenderPass                               | ⚙️ WIP
+vkCmdExecuteCommands                             | ⚙️ WIP
+vkCmdFillBuffer                                  | ✅ Implemented
+vkCmdNextSubpass                                 | ⚙️ WIP
+vkCmdPipelineBarrier                             | ⚙️ WIP
+vkCmdPushConstants                               | ⚙️ WIP
+vkCmdResetEvent                                  | ⚙️ WIP
+vkCmdResetQueryPool                              | ⚙️ WIP
+vkCmdResolveImage                                | ⚙️ WIP
+vkCmdSetBlendConstants                           | ⚙️ WIP
+vkCmdSetDepthBias                                | ⚙️ WIP
+vkCmdSetDepthBounds                              | ⚙️ WIP
+vkCmdSetDeviceMaskKHR                            | ⚙️ WIP
+vkCmdSetEvent                                    | ⚙️ WIP
+vkCmdSetLineWidth                                | ⚙️ WIP
+vkCmdSetScissor                                  | ⚙️ WIP
+vkCmdSetStencilCompareMask                       | ⚙️ WIP
+vkCmdSetStencilReference                         | ⚙️ WIP
+vkCmdSetStencilWriteMask                         | ⚙️ WIP
+vkCmdSetViewport                                 | ⚙️ WIP
+vkCmdUpdateBuffer                                | ⚙️ WIP
+vkCmdWaitEvents                                  | ⚙️ WIP
+vkCmdWriteTimestamp                              | ⚙️ WIP
+vkCreateBuffer                                   | ⚙️ WIP
+vkCreateBufferView                               | ⚙️ WIP
+vkCreateCommandPool                              | ✅ Implemented
+vkCreateComputePipelines                         | ⚙️ WIP
+vkCreateDescriptorPool                           | ⚙️ WIP
+vkCreateDescriptorSetLayout                      | ⚙️ WIP
+vkCreateDevice                                   | ✅ Implemented
+vkCreateEvent                                    | ⚙️ WIP
+vkCreateFence                                    | ⚙️ WIP
+vkCreateFramebuffer                              | ⚙️ WIP
+vkCreateGraphicsPipelines                        | ⚙️ WIP
+vkCreateImage                                    | ✅ Implemented
+vkCreateImageView                                | ⚙️ WIP
+vkCreateInstance                                 | ✅ Implemented
+vkCreatePipelineCache                            | ⚙️ WIP
+vkCreatePipelineLayout                           | ⚙️ WIP
+vkCreateQueryPool                                | ⚙️ WIP
+vkCreateRenderPass                               | ⚙️ WIP
+vkCreateSampler                                  | ⚙️ WIP
+vkCreateSemaphore                                | ⚙️ WIP
+vkCreateShaderModule                             | ⚙️ WIP
+vkCreateSwapchainKHR                             | ⚙️ WIP
+vkCreateWaylandSurfaceKHR                        | ⚙️ WIP
+vkCreateWin32SurfaceKHR                          | ⚙️ WIP
+vkCreateXcbSurfaceKHR                            | ⚙️ WIP
+vkCreateXlibSurfaceKHR                           | ⚙️ WIP
+vkDestroyBuffer                                  | ⚙️ WIP
+vkDestroyBufferView                              | ⚙️ WIP
+vkDestroyCommandPool                             | ✅ Implemented
+vkDestroyDescriptorPool                          | ⚙️ WIP
+vkDestroyDescriptorSetLayout                     | ⚙️ WIP
+vkDestroyDevice                                  | ✅ Implemented
+vkDestroyEvent                                   | ⚙️ WIP
+vkDestroyFence                                   | ⚙️ WIP
+vkDestroyFramebuffer                             | ⚙️ WIP
+vkDestroyImage                                   | ✅ Implemented
+vkDestroyImageView                               | ⚙️ WIP
+vkDestroyInstance                                | ✅ Implemented
+vkDestroyPipeline                                | ⚙️ WIP
+vkDestroyPipelineCache                           | ⚙️ WIP
+vkDestroyPipelineLayout                          | ⚙️ WIP
+vkDestroyQueryPool                               | ⚙️ WIP
+vkDestroyRenderPass                              | ⚙️ WIP
+vkDestroySampler                                 | ⚙️ WIP
+vkDestroySemaphore                               | ⚙️ WIP
+vkDestroyShaderModule                            | ⚙️ WIP
+vkDestroySurfaceKHR                              | ⚙️ WIP
+vkDestroySwapchainKHR                            | ⚙️ WIP
+vkDeviceWaitIdle                                 | ⚙️ WIP
+vkEndCommandBuffer                               | ✅ Implemented
+vkEnumerateDeviceExtensionProperties             | ⚙️ WIP
+vkEnumerateDeviceLayerProperties                 | ⚙️ WIP
+vkEnumerateInstanceExtensionProperties           | ⚙️ WIP
+vkEnumerateInstanceLayerProperties               | ⚙️ WIP
+vkEnumeratePhysicalDeviceGroupsKHR               | ⚙️ WIP
+vkEnumeratePhysicalDevices                       | ✅ Implemented
+vkFlushMappedMemoryRanges                        | ⚙️ WIP
+vkFreeCommandBuffers                             | ⚙️ WIP
+vkFreeDescriptorSets                             | ⚙️ WIP
+vkFreeMemory                                     | ✅ Implemented
+vkGetBufferDeviceAddress                         | ⚙️ WIP
+vkGetBufferDeviceAddressEXT                      | ⚙️ WIP
+vkGetBufferDeviceAddressKHR                      | ⚙️ WIP
+vkGetBufferMemoryRequirements                    | ⚙️ WIP
+vkGetDeviceGroupPeerMemoryFeaturesKHR            | ⚙️ WIP
+vkGetDeviceGroupPresentCapabilitiesKHR           | ⚙️ WIP
+vkGetDeviceGroupSurfacePresentModesKHR           | ⚙️ WIP
+vkGetDeviceMemoryCommitment                      | ⚙️ WIP
+vkGetDeviceProcAddr                              | ⚙️ WIP
+vkGetDeviceQueue                                 | ⚙️ WIP
+vkGetEventStatus                                 | ⚙️ WIP
+vkGetFenceStatus                                 | ⚙️ WIP
+vkGetImageMemoryRequirements                     | ✅ Implemented
+vkGetImageSparseMemoryRequirements               | ⚙️ WIP
+vkGetImageSubresourceLayout                      | ✅ Implemented
+vkGetInstanceProcAddr                            | ⚙️ WIP
+vkGetPhysicalDeviceFeatures                      | ✅ Implemented
+vkGetPhysicalDeviceFormatProperties              | ⚙️ WIP
+vkGetPhysicalDeviceImageFormatProperties         | ✅ Implemented
+vkGetPhysicalDeviceMemoryProperties              | ✅ Implemented
+vkGetPhysicalDeviceProperties                    | ✅ Implemented
+vkGetPhysicalDeviceQueueFamilyProperties         | ⚙️ WIP
+vkGetPhysicalDeviceSparseImageFormatProperties   | ⚙️ WIP
+vkGetPhysicalDeviceSurfaceCapabilitiesKHR        | ⚙️ WIP
+vkGetPhysicalDeviceSurfaceFormatsKHR             | ⚙️ WIP
+vkGetPhysicalDeviceSurfacePresentModesKHR        | ⚙️ WIP
+vkGetPhysicalDeviceSurfaceSupportKHR             | ⚙️ WIP
+vkGetPhysicalDeviceWaylandPresentationSupportKHR | ⚙️ WIP
+vkGetPhysicalDeviceWin32PresentationSupportKHR   | ⚙️ WIP
+vkGetPhysicalDeviceXcbPresentationSupportKHR     | ⚙️ WIP
+vkGetPhysicalDeviceXlibPresentationSupportKHR    | ⚙️ WIP
+vkGetPipelineCacheData                           | ⚙️ WIP
+vkGetQueryPoolResults                            | ⚙️ WIP
+vkGetRenderAreaGranularity                       | ⚙️ WIP
+vkGetSwapchainImagesKHR                          | ⚙️ WIP
+vkInvalidateMappedMemoryRanges                   | ⚙️ WIP
+vkMapMemory                                      | ⚙️ WIP
+vkMergePipelineCaches                            | ⚙️ WIP
+vkQueueBindSparse                                | ⚙️ WIP
+vkQueuePresentKHR                                | ⚙️ WIP
+vkQueueSubmit                                    | ✅ Implemented
+vkQueueWaitIdle                                  | ⚙️ WIP
+vkResetCommandBuffer                             | ⚙️ WIP
+vkResetCommandPool                               | ✅ Implemented
+vkResetDescriptorPool                            | ⚙️ WIP
+vkResetEvent                                     | ⚙️ WIP
+vkResetFences                                    | ⚙️ WIP
+vkResetQueryPool                                 | ⚙️ WIP
+vkSetEvent                                       | ⚙️ WIP
+vkUnmapMemory                                    | ⚙️ WIP
+vkUpdateDescriptorSets                           | ⚙️ WIP
+vkWaitForFences                                  | ⚙️ WIP
+</details>
+
+## Flint [Intel]
+
+Flint be an implementation of the Vulkan specification, fashioned for the elder Intel iGPUs of generations 9 through 11.
+Whether Haswell and those that followed-even the Xe kindred-shall one day partake of its workings remaineth unwritten.
+
+### Build
+
+The rite of forging the driver:
+```
+zig build flint --release=[fast|safe|small]
+```
+
+Should the foregoing not suffice, further build options await thee:
+```
+zig build --help
+```
+
+The same rites apply as for the Soft and Phi build.
+
+#### Vulkan 1.0 specification
+<details>
+    <summary>
+        Behold the present state of thy Vulkan 1.0 implementation
+    </summary>
+
+\
+⚠️ Implemented, yet perchance not fully tested nor proven conformant, but rather working in a manner most general to thee and thine.\
+Assume thou that functions lacking in this array are, for now, not intended to be wrought.
+
+Name                                             | Status
+-------------------------------------------------|--------
+vkAcquireNextImage2KHR                           | ⚙️ WIP
+vkAcquireNextImageKHR                            | ⚙️ WIP
+vkAllocateCommandBuffers                         | ✅ Implemented
+vkAllocateDescriptorSets                         | ⚙️ WIP
+vkAllocateMemory                                 | ✅ Implemented
+vkBeginCommandBuffer                             | ✅ Implemented
+vkBindBufferMemory                               | ✅ Implemented
+vkBindImageMemory                                | ✅ Implemented
+vkCmdBeginQuery                                  | ⚙️ WIP
+vkCmdBeginRenderPass                             | ⚙️ WIP
+vkCmdBindDescriptorSets                          | ⚙️ WIP
+vkCmdBindIndexBuffer                             | ⚙️ WIP
+vkCmdBindPipeline                                | ⚙️ WIP
+vkCmdBindVertexBuffers                           | ⚙️ WIP
+vkCmdBlitImage                                   | ⚙️ WIP
+vkCmdClearAttachments                            | ⚙️ WIP
+vkCmdClearColorImage                             | ⚙️ WIP
+vkCmdClearDepthStencilImage                      | ⚙️ WIP
+vkCmdCopyBuffer                                  | ✅ Implemented
+vkCmdCopyBufferToImage                           | ✅ Implemented
+vkCmdCopyImage                                   | ✅ Implemented
+vkCmdCopyImageToBuffer                           | ✅ Implemented
+vkCmdCopyQueryPoolResults                        | ⚙️ WIP
+vkCmdDispatch                                    | ⚙️ WIP
+vkCmdDispatchBaseKHR                             | ⚙️ WIP
+vkCmdDispatchIndirect                            | ⚙️ WIP
+vkCmdDraw                                        | ⚙️ WIP
+vkCmdDrawIndexed                                 | ⚙️ WIP
+vkCmdDrawIndexedIndirect                         | ⚙️ WIP
+vkCmdDrawIndirect                                | ⚙️ WIP
+vkCmdEndQuery                                    | ⚙️ WIP
+vkCmdEndRenderPass                               | ⚙️ WIP
+vkCmdExecuteCommands                             | ✅ Implemented
+vkCmdFillBuffer                                  | ✅ Implemented
+vkCmdNextSubpass                                 | ⚙️ WIP
+vkCmdPipelineBarrier                             | ⚙️ WIP
+vkCmdPushConstants                               | ⚙️ WIP
+vkCmdResetEvent                                  | ⚙️ WIP
+vkCmdResetQueryPool                              | ⚙️ WIP
+vkCmdResolveImage                                | ⚙️ WIP
+vkCmdSetBlendConstants                           | ⚙️ WIP
+vkCmdSetDepthBias                                | ⚙️ WIP
+vkCmdSetDepthBounds                              | ⚙️ WIP
+vkCmdSetDeviceMaskKHR                            | ⚙️ WIP
+vkCmdSetEvent                                    | ⚙️ WIP
+vkCmdSetLineWidth                                | ⚙️ WIP
+vkCmdSetScissor                                  | ⚙️ WIP
+vkCmdSetStencilCompareMask                       | ⚙️ WIP
+vkCmdSetStencilReference                         | ⚙️ WIP
+vkCmdSetStencilWriteMask                         | ⚙️ WIP
+vkCmdSetViewport                                 | ⚙️ WIP
+vkCmdUpdateBuffer                                | ⚙️ WIP
+vkCmdWaitEvents                                  | ⚙️ WIP
+vkCmdWriteTimestamp                              | ⚙️ WIP
+vkCreateBuffer                                   | ✅ Implemented
+vkCreateBufferView                               | ⚙️ WIP
+vkCreateCommandPool                              | ✅ Implemented
+vkCreateComputePipelines                         | ⚙️ WIP
+vkCreateDescriptorPool                           | ⚙️ WIP
+vkCreateDescriptorSetLayout                      | ⚙️ WIP
+vkCreateDevice                                   | ✅ Implemented
+vkCreateEvent                                    | ⚙️ WIP
+vkCreateFence                                    | ✅ Implemented
+vkCreateFramebuffer                              | ⚙️ WIP
+vkCreateGraphicsPipelines                        | ⚙️ WIP
+vkCreateImage                                    | ✅ Implemented
+vkCreateImageView                                | ⚙️ WIP
+vkCreateInstance                                 | ✅ Implemented
+vkCreatePipelineCache                            | ⚙️ WIP
+vkCreatePipelineLayout                           | ⚙️ WIP
+vkCreateQueryPool                                | ⚙️ WIP
+vkCreateRenderPass                               | ⚙️ WIP
+vkCreateSampler                                  | ⚙️ WIP
+vkCreateSemaphore                                | ✅ Implemented
+vkCreateShaderModule                             | ⚙️ WIP
+vkCreateSwapchainKHR                             | ⚙️ WIP
+vkCreateWaylandSurfaceKHR                        | ⚙️ WIP
+vkCreateWin32SurfaceKHR                          | ⚙️ WIP
+vkCreateXcbSurfaceKHR                            | ⚙️ WIP
+vkCreateXlibSurfaceKHR                           | ⚙️ WIP
+vkDestroyBuffer                                  | ✅ Implemented
+vkDestroyBufferView                              | ⚙️ WIP
+vkDestroyCommandPool                             | ✅ Implemented
+vkDestroyDescriptorPool                          | ⚙️ WIP
+vkDestroyDescriptorSetLayout                     | ⚙️ WIP
+vkDestroyDevice                                  | ✅ Implemented
+vkDestroyEvent                                   | ⚙️ WIP
+vkDestroyFence                                   | ✅ Implemented
+vkDestroyFramebuffer                             | ⚙️ WIP
+vkDestroyImage                                   | ✅ Implemented
+vkDestroyImageView                               | ⚙️ WIP
+vkDestroyInstance                                | ✅ Implemented
+vkDestroyPipeline                                | ⚙️ WIP
+vkDestroyPipelineCache                           | ⚙️ WIP
+vkDestroyPipelineLayout                          | ⚙️ WIP
+vkDestroyQueryPool                               | ⚙️ WIP
+vkDestroyRenderPass                              | ⚙️ WIP
+vkDestroySampler                                 | ⚙️ WIP
+vkDestroySemaphore                               | ✅ Implemented
+vkDestroyShaderModule                            | ⚙️ WIP
+vkDestroySurfaceKHR                              | ⚙️ WIP
+vkDestroySwapchainKHR                            | ⚙️ WIP
+vkDeviceWaitIdle                                 | ✅ Implemented
+vkEndCommandBuffer                               | ✅ Implemented
+vkEnumerateDeviceExtensionProperties             | ✅ Implemented
+vkEnumerateDeviceLayerProperties                 | ⚙️ WIP
+vkEnumerateInstanceExtensionProperties           | ⚙️ WIP
+vkEnumerateInstanceLayerProperties               | ⚙️ WIP
+vkEnumeratePhysicalDeviceGroupsKHR               | ⚙️ WIP
+vkEnumeratePhysicalDevices                       | ✅ Implemented
+vkFlushMappedMemoryRanges                        | ✅ Implemented
+vkFreeCommandBuffers                             | ✅ Implemented
+vkFreeDescriptorSets                             | ⚙️ WIP
+vkFreeMemory                                     | ✅ Implemented
+vkGetBufferDeviceAddress                         | ⚙️ WIP
+vkGetBufferDeviceAddressEXT                      | ⚙️ WIP
+vkGetBufferDeviceAddressKHR                      | ⚙️ WIP
+vkGetBufferMemoryRequirements                    | ⚙️ WIP
+vkGetDeviceGroupPeerMemoryFeaturesKHR            | ⚙️ WIP
+vkGetDeviceGroupPresentCapabilitiesKHR           | ⚙️ WIP
+vkGetDeviceGroupSurfacePresentModesKHR           | ⚙️ WIP
+vkGetDeviceMemoryCommitment                      | ⚙️ WIP
+vkGetDeviceProcAddr                              | ✅ Implemented
+vkGetDeviceQueue                                 | ✅ Implemented
+vkGetEventStatus                                 | ⚙️ WIP
+vkGetFenceStatus                                 | ✅ Implemented
+vkGetImageMemoryRequirements                     | ⚙️ WIP
+vkGetImageSparseMemoryRequirements               | ⚙️ WIP
+vkGetImageSubresourceLayout                      | ⚙️ WIP
+vkGetInstanceProcAddr                            | ⚙️ WIP
+vkGetPhysicalDeviceFeatures                      | ✅ Implemented
+vkGetPhysicalDeviceFormatProperties              | ⚙️ WIP
+vkGetPhysicalDeviceImageFormatProperties         | ⚙️ WIP
+vkGetPhysicalDeviceMemoryProperties              | ✅ Implemented
+vkGetPhysicalDeviceProperties                    | ✅ Implemented
+vkGetPhysicalDeviceQueueFamilyProperties         | ⚙️ WIP
+vkGetPhysicalDeviceSparseImageFormatProperties   | ⚙️ WIP
+vkGetPhysicalDeviceSurfaceCapabilitiesKHR        | ⚙️ WIP
+vkGetPhysicalDeviceSurfaceFormatsKHR             | ⚙️ WIP
+vkGetPhysicalDeviceSurfacePresentModesKHR        | ⚙️ WIP
+vkGetPhysicalDeviceSurfaceSupportKHR             | ⚙️ WIP
+vkGetPhysicalDeviceWaylandPresentationSupportKHR | ⚙️ WIP
+vkGetPhysicalDeviceWin32PresentationSupportKHR   | ⚙️ WIP
+vkGetPhysicalDeviceXcbPresentationSupportKHR     | ⚙️ WIP
+vkGetPhysicalDeviceXlibPresentationSupportKHR    | ⚙️ WIP
+vkGetPipelineCacheData                           | ⚙️ WIP
+vkGetQueryPoolResults                            | ⚙️ WIP
+vkGetRenderAreaGranularity                       | ⚙️ WIP
+vkGetSwapchainImagesKHR                          | ⚙️ WIP
+vkInvalidateMappedMemoryRanges                   | ✅ Implemented
+vkMapMemory                                      | ✅ Implemented
+vkMergePipelineCaches                            | ⚙️ WIP
+vkQueueBindSparse                                | ⚙️ WIP
+vkQueuePresentKHR                                | ⚙️ WIP
+vkQueueSubmit                                    | ✅ Implemented
+vkQueueWaitIdle                                  | ✅ Implemented
+vkResetCommandBuffer                             | ✅ Implemented
+vkResetCommandPool                               | ✅ Implemented
+vkResetDescriptorPool                            | ⚙️ WIP
+vkResetEvent                                     | ⚙️ WIP
+vkResetFences                                    | ✅ Implemented
+vkResetQueryPool                                 | ⚙️ WIP
+vkSetEvent                                       | ⚙️ WIP
+vkUnmapMemory                                    | ✅ Implemented
+vkUpdateDescriptorSets                           | ⚙️ WIP
+vkWaitForFences                                  | ✅ Implemented
+</details>
+
+[Herein](https://vulkan-driver.kbz8.me/cts/flint/) is the reckoning of the Vulkan 1.0 conformance trials laid bare. Judge it as thou wilt.
+
 ## License
 
 Released unto the world as MIT for study, experimentation, and the occasional horrified whisper.

build.zig

@@ -2,7 +2,7 @@ const std = @import("std");
 const Step = std.Build.Step;
 const builtin = @import("builtin");
 
-const driver_version: std.SemanticVersion = .{ .major = 1, .minor = 0, .patch = 0 };
+const driver_version: std.SemanticVersion = .{ .major = 26, .minor = 0, .patch = 0 };
 
 const ImplementationDesc = struct {
     name: []const u8,
@@ -45,6 +45,13 @@ const implementations = [_]ImplementationDesc{
         .custom = customFlint,
         .options = optionsFlint,
     },
+    .{
+        .name = "phi",
+        .root_source_file = "src/phi/lib.zig",
+        .vulkan_version = .{ .major = 1, .minor = 0, .patch = 0 },
+        .custom = customPhi,
+        .options = optionsPhi,
+    },
 };
 
 const RunningMode = enum {
@@ -64,6 +71,37 @@ pub fn build(b: *std.Build) !void {
     const target = b.standardTargetOptions(.{});
     const optimize = b.standardOptimizeOption(.{});
 
+    const ir_mod = b.createModule(.{
+        .root_source_file = b.path("src/compiler/root.zig"),
+        .target = target,
+        .optimize = optimize,
+    });
+
+    const ir_tests = b.addTest(.{
+        .root_module = ir_mod,
+        .test_runner = .{
+            .path = b.path("test/test_runner.zig"),
+            .mode = .simple,
+        },
+    });
+    const run_ir_tests = b.addRunArtifact(ir_tests);
+    const ir_test_step = b.step("test-ir", "Run shared shader ir tests");
+    ir_test_step.dependOn(&run_ir_tests.step);
+
+    const ir_autodoc_test = b.addObject(.{
+        .name = "lib",
+        .root_module = ir_mod,
+    });
+
+    const ir_install_docs = b.addInstallDirectory(.{
+        .source_dir = ir_autodoc_test.getEmittedDocs(),
+        .install_dir = .prefix,
+        .install_subdir = "docs-ir",
+    });
+
+    const ir_docs_step = b.step("docs-ir", "Build and install the documentation or shader IR");
+    ir_docs_step.dependOn(&ir_install_docs.step);
+
     const base_mod = b.createModule(.{
         .root_source_file = b.path("src/vulkan/lib.zig"),
         .target = target,
@@ -81,10 +119,12 @@ pub fn build(b: *std.Build) !void {
     const drm = b.dependency("drm", .{}).module("drm");
 
     const logs_option: LogType = b.option(LogType, "logs", "Driver logs") orelse .none;
+    const debug_allocator_option = b.option(bool, "device-debug-allocator", "Debug device allocator") orelse false;
 
     const options = b.addOptions();
     options.addOption(std.SemanticVersion, "driver_version", driver_version);
     options.addOption(LogType, "logs", logs_option);
+    options.addOption(bool, "device_debug_allocator", debug_allocator_option);
 
     base_mod.addImport("vulkan", vulkan);
     base_mod.addImport("zmath", zmath);
@@ -166,7 +206,13 @@ pub fn build(b: *std.Build) !void {
         const install_step = b.step(impl.name, b.fmt("Build libvulkan_{s}", .{impl.name}));
         install_step.dependOn(&lib_install.step);
 
-        const lib_tests = b.addTest(.{ .root_module = lib_mod });
+        const lib_tests = b.addTest(.{
+            .root_module = lib_mod,
+            .test_runner = .{
+                .path = b.path("test/test_runner.zig"),
+                .mode = .simple,
+            },
+        });
 
         const run_tests = b.addRunArtifact(lib_tests);
         const test_step = b.step(b.fmt("test-{s}", .{impl.name}), b.fmt("Run libvulkan_{s} tests", .{impl.name}));
@@ -209,97 +255,6 @@ pub fn build(b: *std.Build) !void {
     docs_step.dependOn(&install_docs.step);
 }
 
-fn customApe(
-    b: *std.Build,
-    lib: *Step.Compile,
-    lib_mod: *std.Build.Module,
-    base_mod: *std.Build.Module,
-    vulkan: *std.Build.Module,
-    base_c_mod: *std.Build.Module,
-    target: std.Build.ResolvedTarget,
-    optimize: std.builtin.OptimizeMode,
-    use_llvm: bool,
-) !void {
-    for (implementations) |impl| {
-        if (std.mem.eql(u8, impl.name, "ape"))
-            continue;
-
-        const mod = b.createModule(.{
-            .root_source_file = b.path(impl.root_source_file),
-            .target = target,
-            .optimize = optimize,
-            .imports = &.{
-                .{ .name = "base", .module = base_mod },
-                .{ .name = "vulkan", .module = vulkan },
-            },
-        });
-
-        if (impl.custom) |func| {
-            func(b, lib, mod, base_mod, vulkan, base_c_mod, target, optimize, use_llvm) catch continue;
-        }
-
-        lib_mod.addImport(impl.name, mod);
-    }
-}
-
-fn customSoft(
-    b: *std.Build,
-    _: *Step.Compile,
-    lib_mod: *std.Build.Module,
-    _: *std.Build.Module,
-    _: *std.Build.Module,
-    base_c_mod: *std.Build.Module,
-    _: std.Build.ResolvedTarget,
-    _: std.builtin.OptimizeMode,
-    use_llvm: bool,
-) !void {
-    const spv = b.lazyDependency("SPIRV_Interpreter", .{
-        .@"no-example" = true,
-        .@"no-test" = true,
-        .@"use-llvm" = use_llvm,
-    }) orelse return error.UnresolvedDependency;
-
-    lib_mod.addImport("soft_c", base_c_mod);
-    lib_mod.addImport("spv", spv.module("spv"));
-}
-
-fn optionsSoft(b: *std.Build, options: *Step.Options) !void {
-    const single_threaded_option = b.option(bool, "soft-single-threaded", "Single threaded runtime mode") orelse false;
-    const debug_allocator_option = b.option(bool, "soft-debug-allocator", "Debug device allocator") orelse false;
-    const shaders_simd_option = b.option(bool, "soft-shader-simd", "Shaders SIMD acceleration") orelse true;
-    const compute_dump_early_results_table_option = b.option(u32, "soft-compute-dump-early-results-table", "Dump compute shaders results table before invocation");
-    const compute_dump_final_results_table_option = b.option(u32, "soft-compute-dump-final-results-table", "Dump compute shaders results table after invocation");
-    const approxiamte_rgb_option = b.option(bool, "soft-approximates-rgb", "Approximate sRGB <-> RGB conversions") orelse true;
-
-    options.addOption(bool, "soft_single_threaded", single_threaded_option);
-    options.addOption(bool, "soft_debug_allocator", debug_allocator_option);
-    options.addOption(bool, "soft_shaders_simd", shaders_simd_option);
-    options.addOption(?u32, "soft_compute_dump_early_results_table", compute_dump_early_results_table_option);
-    options.addOption(?u32, "soft_compute_dump_final_results_table", compute_dump_final_results_table_option);
-    options.addOption(bool, "soft_approximates_rgb", approxiamte_rgb_option);
-}
-
-fn customFlint(
-    _: *std.Build,
-    _: *Step.Compile,
-    lib_mod: *std.Build.Module,
-    _: *std.Build.Module,
-    _: *std.Build.Module,
-    base_c_mod: *std.Build.Module,
-    _: std.Build.ResolvedTarget,
-    _: std.builtin.OptimizeMode,
-    _: bool,
-) !void {
-    lib_mod.addImport("intel_c", base_c_mod);
-
-    lib_mod.addImport("soft_c", base_c_mod);
-}
-
-fn optionsFlint(b: *std.Build, options: *Step.Options) !void {
-    _ = b;
-    _ = options;
-}
-
 fn addCTS(b: *std.Build, target: std.Build.ResolvedTarget, impl: *const ImplementationDesc, impl_lib: *Step.Compile, comptime mode: RunningMode) !*Step {
     const cts = b.dependency("cts_bin", .{});
 
@@ -432,7 +387,8 @@ fn addMultithreadedCTS(b: *std.Build, target: std.Build.ResolvedTarget, impl: *c
     }
 
     run.addArg("run");
-    run.addArg("--verbose");
+    run.addArg("--timeout");
+    run.addArg("60");
     run.addArg("--deqp");
     run.addArg(cts_exe_path);
     run.addArg("--caselist");
@@ -469,3 +425,254 @@ fn addMultithreadedCTS(b: *std.Build, target: std.Build.ResolvedTarget, impl: *c
 
     return &run.step;
 }
+
+// Ape specialized functions
+
+fn customApe(
+    b: *std.Build,
+    lib: *Step.Compile,
+    lib_mod: *std.Build.Module,
+    base_mod: *std.Build.Module,
+    vulkan: *std.Build.Module,
+    base_c_mod: *std.Build.Module,
+    target: std.Build.ResolvedTarget,
+    optimize: std.builtin.OptimizeMode,
+    use_llvm: bool,
+) !void {
+    for (implementations) |impl| {
+        if (std.mem.eql(u8, impl.name, "ape"))
+            continue;
+
+        const mod = b.createModule(.{
+            .root_source_file = b.path(impl.root_source_file),
+            .target = target,
+            .optimize = optimize,
+            .imports = &.{
+                .{ .name = "base", .module = base_mod },
+                .{ .name = "vulkan", .module = vulkan },
+            },
+        });
+
+        if (impl.custom) |func| {
+            func(b, lib, mod, base_mod, vulkan, base_c_mod, target, optimize, use_llvm) catch continue;
+        }
+
+        lib_mod.addImport(impl.name, mod);
+    }
+}
+
+// Soft specialized functions
+
+fn customSoft(
+    b: *std.Build,
+    _: *Step.Compile,
+    lib_mod: *std.Build.Module,
+    _: *std.Build.Module,
+    _: *std.Build.Module,
+    base_c_mod: *std.Build.Module,
+    target: std.Build.ResolvedTarget,
+    optimize: std.builtin.OptimizeMode,
+    use_llvm: bool,
+) !void {
+    const spv = b.lazyDependency("SPIRV_Interpreter", .{
+        .target = target,
+        .optimize = optimize,
+        .@"use-llvm" = use_llvm,
+    }) orelse return error.UnresolvedDependency;
+
+    lib_mod.addImport("soft_c", base_c_mod);
+    lib_mod.addImport("spv", spv.module("spv"));
+}
+
+fn optionsSoft(b: *std.Build, options: *Step.Options) !void {
+    const single_threaded_option = b.option(bool, "soft-single-threaded", "Single threaded runtime mode") orelse false;
+    const shaders_simd_option = b.option(bool, "soft-shader-simd", "Shaders SIMD acceleration") orelse true;
+    const compute_dump_early_results_table_option = b.option(u32, "soft-compute-dump-early-results-table", "Dump compute shaders results table before invocation");
+    const compute_dump_final_results_table_option = b.option(u32, "soft-compute-dump-final-results-table", "Dump compute shaders results table after invocation");
+    const approxiamte_rgb_option = b.option(bool, "soft-approximates-rgb", "Approximate sRGB <-> RGB conversions") orelse true;
+
+    options.addOption(bool, "soft_single_threaded", single_threaded_option);
+    options.addOption(bool, "soft_shaders_simd", shaders_simd_option);
+    options.addOption(?u32, "soft_compute_dump_early_results_table", compute_dump_early_results_table_option);
+    options.addOption(?u32, "soft_compute_dump_final_results_table", compute_dump_final_results_table_option);
+    options.addOption(bool, "soft_approximates_rgb", approxiamte_rgb_option);
+}
+
+// Flint specialized functions
+
+fn customFlint(
+    b: *std.Build,
+    _: *Step.Compile,
+    lib_mod: *std.Build.Module,
+    _: *std.Build.Module,
+    _: *std.Build.Module,
+    base_c_mod: *std.Build.Module,
+    target: std.Build.ResolvedTarget,
+    optimize: std.builtin.OptimizeMode,
+    _: bool,
+) !void {
+    lib_mod.addImport("intel_c", base_c_mod);
+    lib_mod.addImport("shader_compiler", b.createModule(.{
+        .root_source_file = b.path("src/compiler/root.zig"),
+        .target = target,
+        .optimize = optimize,
+    }));
+}
+
+fn optionsFlint(b: *std.Build, options: *Step.Options) !void {
+    _ = b;
+    _ = options;
+}
+
+// Phi specialized functions
+
+fn customPhi(
+    b: *std.Build,
+    lib: *Step.Compile,
+    lib_mod: *std.Build.Module,
+    _: *std.Build.Module,
+    _: *std.Build.Module,
+    base_c_mod: *std.Build.Module,
+    target: std.Build.ResolvedTarget,
+    optimize: std.builtin.OptimizeMode,
+    use_llvm: bool,
+) !void {
+    lib_mod.addImport("phi_c", base_c_mod);
+
+    const miclib = b.lazyDependency("miclib", .{
+        .target = target,
+        .optimize = optimize,
+        .@"use-llvm" = use_llvm,
+    }) orelse return error.UnresolvedDependency;
+
+    lib_mod.addImport("miclib", miclib.module("miclib"));
+
+    const phi_protocol_c = b.addTranslateC(.{
+        .root_source_file = b.path("src/phi/shared/Protocol.h"),
+        .target = target,
+        .optimize = optimize,
+        .link_libc = false,
+    });
+
+    lib_mod.addImport("phi_protocol_c", phi_protocol_c.createModule());
+
+    // To avoid duplicated options due to Ape's custom function
+    if (!std.mem.eql(u8, lib.name, "vulkan_phi")) {
+        const daemon = try addPhiCardDaemon(b, optimize, "k1om-mpss-linux-gcc", null);
+        const embedded_daemon = addEmbeddedPhiDaemon(b, daemon);
+        lib_mod.addAnonymousImport("phi_daemon", .{
+            .root_source_file = embedded_daemon,
+        });
+        return;
+    }
+
+    const build_card = b.option(
+        bool,
+        "phi-build-daemon",
+        "Build Xeon Phi card daemon",
+    ) orelse true;
+
+    if (!build_card)
+        return;
+
+    const cc = b.option(
+        []const u8,
+        "phi-card-cc",
+        "Path to k1om-mpss-linux-gcc",
+    ) orelse "k1om-mpss-linux-gcc";
+
+    const sysroot = b.option(
+        []const u8,
+        "phi-card-sysroot",
+        "MPSS sysroot path",
+    );
+
+    const daemon = try addPhiCardDaemon(b, optimize, cc, sysroot);
+    const install_daemon = b.addInstallFile(daemon, "lib/phi_device.mic");
+    lib.step.dependOn(&install_daemon.step);
+
+    const embedded_daemon = addEmbeddedPhiDaemon(b, daemon);
+    lib_mod.addAnonymousImport("phi_daemon", .{
+        .root_source_file = embedded_daemon,
+    });
+}
+
+fn optionsPhi(b: *std.Build, options: *Step.Options) !void {
+    const daemon_remote_path = b.option(
+        []const u8,
+        "phi-daemon-remote-path",
+        "Path where the Xeon Phi daemon is copied on the card",
+    ) orelse "/tmp/phi_device.mic";
+
+    const daemon_host_prefix = b.option(
+        []const u8,
+        "phi-daemon-host-prefix",
+        "Host prefix used to reach cards over ssh/scp; card N uses <prefix>N",
+    ) orelse "mic";
+
+    options.addOption([]const u8, "phi_daemon_remote_path", daemon_remote_path);
+    options.addOption([]const u8, "phi_daemon_host_prefix", daemon_host_prefix);
+}
+
+fn addPhiCardDaemon(
+    b: *std.Build,
+    optimize: std.builtin.OptimizeMode,
+    cc: []const u8,
+    sysroot: ?[]const u8,
+) !std.Build.LazyPath {
+    const cmd = b.addSystemCommand(&.{cc});
+
+    cmd.addArgs(&.{
+        "-std=c11",
+        "-Wall",
+        "-Wextra",
+        "-Wno-unused-parameter",
+        "-pthread",
+    });
+
+    cmd.addArg("-I");
+    cmd.addDirectoryArg(b.path("src/phi/mic"));
+    cmd.addArg("-I");
+    cmd.addDirectoryArg(b.path("src/phi/shared"));
+
+    if (sysroot) |path| {
+        cmd.addArg("--sysroot");
+        cmd.addArg(path);
+    }
+
+    switch (optimize) {
+        .Debug => cmd.addArgs(&.{ "-O0", "-g3" }),
+        .ReleaseSafe => cmd.addArgs(&.{ "-O2", "-g", "-DNDEBUG" }),
+        .ReleaseFast => cmd.addArgs(&.{ "-O3", "-DNDEBUG" }),
+        .ReleaseSmall => cmd.addArgs(&.{ "-Os", "-DNDEBUG" }),
+    }
+
+    const sources = [_][]const u8{
+        "src/phi/mic/main.c",
+        "src/phi/mic/Buffer.c",
+        "src/phi/mic/CommandBuffer.c",
+        "src/phi/mic/Daemon.c",
+        "src/phi/mic/Logger.c",
+        "src/phi/mic/Memory.c",
+        // Add new files here
+    };
+
+    for (sources) |source| {
+        cmd.addFileArg(b.path(source));
+    }
+
+    cmd.addArgs(&.{
+        "-lscif",
+        "-o",
+    });
+
+    return cmd.addOutputFileArg("phi_device.mic");
+}
+
+fn addEmbeddedPhiDaemon(b: *std.Build, daemon: std.Build.LazyPath) std.Build.LazyPath {
+    const wf = b.addWriteFiles();
+    _ = wf.addCopyFile(daemon, "phi_device.mic");
+    return wf.add("phi_daemon.zig",
+        \\pub const data = @embedFile("phi_device.mic");
+    );
+}

build.zig.zon

@@ -1,6 +1,7 @@
 .{
     .name = .VulkanDriver,
-    .version = "0.0.1",
+    .version = "2026.0.0",
+
     .fingerprint = 0x52cb73649f1107de,
     .minimum_zig_version = "0.16.0",
 
@@ -26,14 +27,14 @@
             .hash = "N-V-__8AAF9uOh0I4P_99za7N822J3JwsDaqONrFVrcEQo59",
         },
         .drm = .{
-            .url = "git+https://git.kbz8.me/kbz_8/zig-drm.git#409f58daa8f5174b2fcb8897f1c30f0b0729b611",
+            .url = "git+https://github.com/Kbz-8/zig-drm#409f58daa8f5174b2fcb8897f1c30f0b0729b611",
             .hash = "drm-0.0.1-uWWar5YwAQCiTNOcqtYqt_yL19B0_Q-YLjPNkRysFSnR",
         },
 
         // Soft dependencies
         .SPIRV_Interpreter = .{
-            .url = "git+https://github.com/Kbz-8/SPIRV-Interpreter#f82e3e1629af4dc751a80a3a4b04ff1b43725c24",
-            .hash = "SPIRV_Interpreter-0.0.1-ajmpn6mrBwDl75NH4NndKIOjnn-1ooUY-0NRyWR2eW-P",
+            .url = "git+https://github.com/Kbz-8/SPIRV-Interpreter#92e86ba34fccf32a30e6f717a9cafe0b9e95bb2a",
+            .hash = "SPIRV_Interpreter-1.0.0-ajmpnwSKCQDjTAGgiw6N0rWdZiD3VdEMAE352RSpN4vl",
             .lazy = true,
         },
         //.SPIRV_Interpreter = .{
@@ -41,6 +42,13 @@
         //    .path = "../SPIRV-Interpreter",
         //    .lazy = true,
         //},
+
+        // Phi dependencies
+        .miclib = .{
+            .url = "git+https://git.kbz8.me/kbz_8/miclib-zig.git#f46bd2a058db7acbc9012a657f3fbe79d49c017f",
+            .hash = "miclib-1.0.0-AbDwSzB6AgDVMQ3muWEwaBdfYJBDgm3pJKNf2EEUnmNp",
+            .lazy = true,
+        },
     },
 
     .paths = .{

compile_flags.txt

@@ -0,0 +1,5 @@
+-xc
+-std=c11
+-Isrc/phi/shared
+-Isrc/phi/mic
+-isystem/opt/mpss/3.8.6/sysroots/k1om-mpss-linux/usr/include/

src/ape/ApeInstance.zig

@@ -4,8 +4,8 @@ const base = @import("base");
 
 const soft = @import("soft");
 const flint = @import("flint");
+const phi = @import("phi");
 
-const Dispatchable = base.Dispatchable;
 const VkError = base.VkError;
 
 const Self = @This();
@@ -14,7 +14,7 @@ pub const Interface = base.Instance;
 interface: Interface,
 backend_instances: std.ArrayList(*Interface),
 
-pub const EXTENSIONS = soft.Instance.EXTENSIONS;
+pub const extensions = soft.Instance.extensions;
 
 pub fn create(allocator: std.mem.Allocator, infos: *const vk.InstanceCreateInfo) VkError!*Interface {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
@@ -36,13 +36,17 @@ pub fn create(allocator: std.mem.Allocator, infos: *const vk.InstanceCreateInfo)
     };
 
     const soft_instance = try soft.Instance.create(allocator, infos);
-    errdefer soft_instance.deinit(allocator) catch {};
+    errdefer soft_instance.deinit(allocator) catch @panic("Caught an error while handling an error");
     self.backend_instances.append(allocator, soft_instance) catch return VkError.OutOfHostMemory;
 
     const flint_instance = try flint.Instance.create(allocator, infos);
-    errdefer flint_instance.deinit(allocator) catch {};
+    errdefer flint_instance.deinit(allocator) catch @panic("Caught an error while handling an error");
     self.backend_instances.append(allocator, flint_instance) catch return VkError.OutOfHostMemory;
 
+    const phi_instance = try phi.Instance.create(allocator, infos);
+    errdefer phi_instance.deinit(allocator) catch @panic("Caught an error while handling an error");
+    self.backend_instances.append(allocator, phi_instance) catch return VkError.OutOfHostMemory;
+
     return &self.interface;
 }
 
@@ -57,13 +61,13 @@ fn requestPhysicalDevices(interface: *Interface, allocator: std.mem.Allocator, _
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
 
     for (self.backend_instances.items) |backend| {
-        try appendBackendPhysicalDevices(self, allocator, backend);
+        appendBackendPhysicalDevices(self, allocator, backend) catch continue;
     }
 }
 
 fn appendBackendPhysicalDevices(self: *Self, allocator: std.mem.Allocator, backend: *Interface) VkError!void {
     try backend.requestPhysicalDevices(allocator);
-    errdefer backend.releasePhysicalDevices(allocator) catch {};
+    errdefer backend.releasePhysicalDevices(allocator) catch @panic("Caught an error while handling an error");
 
     self.interface.physical_devices.appendSlice(allocator, backend.physical_devices.items) catch return VkError.OutOfHostMemory;
     backend.physical_devices.deinit(allocator);

src/ape/lib.zig

@@ -10,9 +10,9 @@ pub const ApeInstance = @import("ApeInstance.zig");
 
 pub const Instance = ApeInstance;
 
-pub const DRIVER_NAME = "Ape";
+pub const driver_name = "Ape";
 
-pub const VULKAN_VERSION = vk.makeApiVersion(
+pub const vulkan_version = vk.makeApiVersion(
     0,
     config.ape_vulkan_version.major,
     config.ape_vulkan_version.minor,

src/compiler/README.md

@@ -0,0 +1,612 @@
+# Backend-Agnostic Shader IR
+
+> Note: this IR is still foundational and incomplete. Its representation may
+> change as the compiler gains features. Backends should not treat it as a
+> stable ABI yet.
+
+This directory contains the backend-agnostic shader intermediate representation.
+It sits between SPIR-V and the backends that consume it.
+Format-specific details are removed while types, values, control flow,
+interfaces, and semantic operations remain.
+
+## Overview
+
+- One `Module` describes one selected shader entry point.
+- The current stages are `vertex`, `fragment`, and `compute`.
+- Values are typed and use static single assignment (SSA).
+- Blocks act as control-flow graph labels, but also own instructions,
+  parameters, a terminator, and structured-control metadata.
+- Block parameters serve the same purpose as `OpPhi` in SPIR-V.
+- Instructions speak normalized meanings, not source-format opcodes.
+- Types and constants are interned. Equal things share one identity.
+- IDs are stable. Erasure leaves a tombstone; no dead ID is reused for another object.
+- The printer is for debugging and tests. Its output can be parsed back into a
+  validated module, but it is not yet a stable interchange format.
+
+## Printer syntax
+
+The printer uses these prefixes:
+
+| Prefix  | Meaning                                                          | Example               |
+| ------- | ---------------------------------------------------------------- | --------------------- |
+| `%id`   | An SSA value, whether constant, parameter, or instruction result | `%3`, `%merged_value` |
+| `@name` | A function or interface declaration                              | `@main`, `@out_color` |
+| `.name` | A basic block                                                    | `.entry`, `.merge`    |
+| `#N`    | A constant-store identity used within composite constants        | `#2`                  |
+
+Names are annotations rather than identity. When a name is absent, invalid for
+the textual grammar, or duplicated, the printer uses a numeric `%N` value
+reference. Other unnamed objects use forms such as `@fn0`, `@interface1`, and
+`.b2`. The parser accepts both numeric and identifier-shaped value references.
+
+An instruction that produces a value prints its result type explicitly:
+
+```text
+%result: <type> = <opcode> <operands>
+```
+
+The type annotation makes operations such as `bitcast` and heterogeneous
+`composite_construct` unambiguous when the text is parsed. Constants, function
+parameters, and block parameters already carry their types in their own forms.
+
+The outer structure has this shape:
+
+```text
+shader <stage> @<entry-point>
+{
+    <interface declarations>
+    <constant declarations>
+
+    fn @<name>(<parameters>) -> <type>
+    {
+        .<block>(<block parameters>):
+            <instructions>
+            <terminator>
+
+    }
+}
+```
+
+Execution modes, resources, source locations, and structured-control metadata
+exist in memory, but the printer does not display them yet.
+
+## Parsing
+
+`ir.parser` accepts the complete syntax emitted by the printer:
+
+```zig
+var from_string = try ir.parser.parseString(allocator, source);
+defer from_string.deinit();
+
+var from_file = try ir.parser.parseFile(allocator, io, "shader.ir");
+defer from_file.deinit();
+```
+
+Use `parseFileInDir` when the path is relative to an existing `std.Io.Dir`.
+Each parser entry point owns the returned module with the supplied allocator and
+runs the IR validator before returning it. Parse, reference-resolution, or
+validation failures are returned as errors. Because the printer omits the
+metadata listed above, a print/parse round trip preserves the displayed IR but
+cannot recover those hidden fields.
+
+## Types
+
+Types are interned in the module and printed inline; their `TypeId` is hidden.
+
+| Kind             | Printed form                    | Meaning                                                                   |
+| ---------------- | ------------------------------- | ------------------------------------------------------------------------- |
+| Void             | `void`                          | No value. Used mainly for functions that return nothing.                  |
+| Boolean          | `bool`                          | A truth value.                                                            |
+| Signed integer   | `i32`                           | A signed integer of the written bit width.                                |
+| Unsigned integer | `u32`                           | An unsigned integer of the written bit width.                             |
+| Floating point   | `f32`                           | A floating value of the written bit width.                                |
+| Vector           | `vec4[f32]`                     | A fixed number of equal scalar elements. Its length must be at least two. |
+| Array            | `array[u32, 8]`                 | A fixed number of equal elements. Its length must not be zero.            |
+| Structure        | `struct[f32, vec4[f32]]`        | An ordered sequence of potentially different member types.                |
+| Pointer          | `ptr[workgroup, u32]`           | A pointer to a type within an address space.                              |
+| Resource handle  | `resourceHandle[sampled_image]` | An opaque handle for later resource operations.                           |
+
+The current address spaces are `function`, `private`, `workgroup`,
+`input`, `output`, `uniform`, `storage`, `push_constant`, and `physical`.
+
+The current resource kinds are `uniform_buffer`, `storage_buffer`,
+`sampled_image`, `storage_image`, and `sampler`. A resource handle may also
+carry an optional data type in memory, although the printer omits that type.
+
+## Constants
+
+Constants live at module scope and also have ordinary numeric or named `%id` value identities.
+The parser accepts direct decimal integer and floating-point values, including
+signed values and floating-point exponents, as source-level convenience syntax:
+
+```text
+%0: constant u32 = 42
+%1: constant i32 = -7
+%2: constant f32 = 1.5e2
+```
+
+Direct integers must fit their declared width and signedness. Direct floats are
+rounded to their declared `f16`, `f32`, or `f64` representation. The canonical
+printer always emits integer and float bit patterns so reparsing cannot silently
+change the stored value.
+
+```text
+%id: constant <type> = <value>
+```
+
+| Form         | Meaning                             | Printed example                         |
+| ------------ | ----------------------------------- | --------------------------------------- |
+| Boolean      | `true` or `false`                   | `%0: constant bool = true`              |
+| Integer bits | The fixed-width integer bit pattern | `%1: constant u32 = bits(0x2a)`         |
+| Float bits   | The IEEE-like bit pattern as stored | `%2: constant f32 = bits(0x3f800000)`   |
+| Null         | The null value of its type          | `%3: constant ptr[private, u32] = null` |
+| Undef        | An unconstrained value              | `%4: constant u32 = undef`              |
+| Composite    | A sequence of other constants       | `%5: constant vec2[u32] = [#1, #1]`     |
+
+## Functions, blocks, and SSA
+
+A function owns typed parameters, an ordered list of blocks, one entry block,
+and one return type. The first block created by the builder becomes the entry
+block. The validator requires the entry block to have no predecessor.
+
+Every block must end in exactly one terminator. A value defined by an instruction
+must dominate every use, and within one block it must be written before it is
+used. Constants and standalone undef values are module-wide; function and block
+parameters cannot be used by another function.
+
+### Block parameters and Phi lowering
+
+A merge block does not contain a `phi` instruction. Instead, it declares a
+parameter, and every incoming edge passes one argument of the same type:
+
+```text
+.left():
+    branch .merge(%3)
+
+.right():
+    branch .merge(%4)
+
+.merge(%5: u32):
+    %6: u32 = integer_multiply %5, %2
+    return
+```
+
+`%5` therefore receives the value supplied by the selected edge. The number and
+types of edge arguments must exactly match the target block's parameters.
+
+### Structured control
+
+A block may carry one of these unprinted metadata values:
+
+- `none`: no structured-control promise.
+- `selection`: names one merge block.
+- `loop`: names both merge and continue blocks.
+
+The SPIR-V translator preserves `OpSelectionMerge` and `OpLoopMerge` in this
+metadata. They are not terminators and do not create graph edges themselves.
+
+## Common instruction rules
+
+An instruction belongs to one block, has zero or one result, and may carry a
+source location. Except for `store_interface` and `call`, current operations are
+treated as side-effect free by the rewriter. A block's terminator is stored
+separately from its ordinary instructions.
+
+Most arithmetic operations are intended for scalars or vectors of their named
+category and act component by component where vectors are allowed. The current
+foundational validator often checks only that operand and result types match.
+The stricter integer, float, boolean, bit-width, and vector-shape requirements
+below describe semantic intent and still need more complete validation.
+
+## Unary opcodes
+
+Form:
+
+```text
+%result: <type> = <opcode> %operand
+```
+
+| Opcode        | Arity | Description                  | Usage                                                             | Small printed example       |
+| ------------- | ----: | ---------------------------- | ----------------------------------------------------------------- | --------------------------- |
+| `negate`      |     1 | Changes the arithmetic sign. | Signed integer or floating operand; the result has the same type. | `%2: i32 = negate %1`       |
+| `logical_not` |     1 | Inverts a boolean value.     | Boolean operand and boolean result.                               | `%2: bool = logical_not %1` |
+| `bitwise_not` |     1 | Inverts every bit.           | Integer operand; the result has the same type.                    | `%2: u32 = bitwise_not %1`  |
+
+`negate` is one normalized opcode: the operand type distinguishes integer
+negation from floating negation.
+
+## Binary opcodes
+
+Form:
+
+```text
+%result: <type> = <opcode> %lhs, %rhs
+```
+
+The two operands and result currently must have the same IR type.
+
+### Integer arithmetic
+
+| Opcode             | Description                                             | Usage                                                                       | Small printed example               |
+| ------------------ | ------------------------------------------------------- | --------------------------------------------------------------------------- | ----------------------------------- |
+| `integer_add`      | Adds fixed-width integers.                              | Integer operands of one type.                                               | `%3: u32 = integer_add %1, %2`      |
+| `integer_subtract` | Subtracts the right operand from the left.              | Integer operands of one type.                                               | `%3: u32 = integer_subtract %1, %2` |
+| `integer_multiply` | Multiplies fixed-width integers.                        | Integer operands of one type.                                               | `%3: u32 = integer_multiply %1, %2` |
+| `unsigned_divide`  | Divides unsigned integers.                              | Unsigned integer operands.                                                  | `%3: u32 = unsigned_divide %1, %2`  |
+| `signed_divide`    | Divides signed integers.                                | Signed integer operands.                                                    | `%3: i32 = signed_divide %1, %2`    |
+| `unsigned_modulo`  | Produces the unsigned remainder.                        | Unsigned integer operands.                                                  | `%3: u32 = unsigned_modulo %1, %2`  |
+| `signed_modulo`    | Produces signed modulo, whose sign follows the divisor. | Signed integer operands; this corresponds to SPIR-V `OpSMod`, not `OpSRem`. | `%3: i32 = signed_modulo %1, %2`    |
+
+Integer addition, subtraction, and multiplication are signedness-neutral at the
+opcode level; the type retains signedness. Exceptional division, overflow,
+and poison rules are not yet separately recorded by the IR.
+
+### Floating arithmetic
+
+| Opcode           | Description                                                  | Usage                                                               | Small printed example             |
+| ---------------- | ------------------------------------------------------------ | ------------------------------------------------------------------- | --------------------------------- |
+| `float_add`      | Adds floating-point values.                                  | Floating operands of one type.                                      | `%3: f32 = float_add %1, %2`      |
+| `float_subtract` | Subtracts the right operand from the left.                   | Floating operands of one type.                                      | `%3: f32 = float_subtract %1, %2` |
+| `float_multiply` | Multiplies floating-point values.                            | Floating operands of one type.                                      | `%3: f32 = float_multiply %1, %2` |
+| `float_divide`   | Divides the left operand by the right.                       | Floating operands of one type.                                      | `%3: f32 = float_divide %1, %2`   |
+| `float_modulo`   | Produces floating modulo, similar to `x - y * floor(x / y)`. | Floating operands of one type; this corresponds to SPIR-V `OpFMod`. | `%3: f32 = float_modulo %1, %2`   |
+
+Instructions do not yet carry fast-math flags, rounding modes, contraction
+permission, or NaN guarantees.
+
+### Shifts and bitwise arithmetic
+
+| Opcode                   | Description                                        | Usage                                                         | Small printed example                     |
+| ------------------------ | -------------------------------------------------- | ------------------------------------------------------------- | ----------------------------------------- |
+| `shift_left`             | Shifts bits left and fills the low bits with zero. | Integer operands; the right operand supplies the shift count. | `%3: u32 = shift_left %1, %2`             |
+| `logical_shift_right`    | Shifts right and fills high bits with zero.        | Integer operands interpreted without sign extension.          | `%3: u32 = logical_shift_right %1, %2`    |
+| `arithmetic_shift_right` | Shifts right while repeating the sign bit.         | Signed integer value and integer shift count.                 | `%3: i32 = arithmetic_shift_right %1, %2` |
+| `bitwise_and`            | Keeps bits set in both operands.                   | Integer operands of one type.                                 | `%3: u32 = bitwise_and %1, %2`            |
+| `bitwise_or`             | Keeps bits set in either operand.                  | Integer operands of one type.                                 | `%3: u32 = bitwise_or %1, %2`             |
+| `bitwise_xor`            | Keeps bits set in exactly one operand.             | Integer operands of one type.                                 | `%3: u32 = bitwise_xor %1, %2`            |
+
+The validator currently requires the shift count to have the same complete IR
+type as the shifted value. More flexible shift typing is not implemented yet.
+
+### Boolean conjunction
+
+| Opcode        | Description                               | Usage                        | Small printed example           |
+| ------------- | ----------------------------------------- | ---------------------------- | ------------------------------- |
+| `logical_and` | Is true only when both operands are true. | Boolean operands and result. | `%3: bool = logical_and %1, %2` |
+| `logical_or`  | Is true when either operand is true.      | Boolean operands and result. | `%3: bool = logical_or %1, %2`  |
+
+## Comparison opcodes
+
+All comparisons are printed with the `cmp_` prefix as one opcode token:
+
+```text
+%result: bool = cmp_<opcode> %lhs, %rhs
+```
+
+The operands must share one type, and the current validator requires the result
+to be the scalar `bool` type.
+
+| Opcode                          | Description                                             | Usage                                    | Small printed example                             |
+| ------------------------------- | ------------------------------------------------------- | ---------------------------------------- | ------------------------------------------------- |
+| `cmp_equal`                     | Tests whether two booleans or integers are equal.       | Equal-typed boolean or integer operands. | `%3: bool = cmp_equal %1, %2`                     |
+| `cmp_not_equal`                 | Tests whether two booleans or integers differ.          | Equal-typed boolean or integer operands. | `%3: bool = cmp_not_equal %1, %2`                 |
+| `cmp_unsigned_less`             | Compares integer bit patterns as unsigned.              | Unsigned integer operands.               | `%3: bool = cmp_unsigned_less %1, %2`             |
+| `cmp_signed_less`               | Compares integers as signed.                            | Signed integer operands.                 | `%3: bool = cmp_signed_less %1, %2`               |
+| `cmp_ordered_float_equal`       | Is true when neither operand is NaN and they are equal. | Floating operands.                       | `%3: bool = cmp_ordered_float_equal %1, %2`       |
+| `cmp_unordered_float_equal`     | Is true when either operand is NaN, or they are equal.  | Floating operands.                       | `%3: bool = cmp_unordered_float_equal %1, %2`     |
+| `cmp_ordered_float_not_equal`   | Is true when neither operand is NaN and they differ.    | Floating operands.                       | `%3: bool = cmp_ordered_float_not_equal %1, %2`   |
+| `cmp_unordered_float_not_equal` | Is true when either operand is NaN, or they differ.     | Floating operands.                       | `%3: bool = cmp_unordered_float_not_equal %1, %2` |
+| `cmp_ordered_float_less`        | Is true when neither operand is NaN and left is less.   | Floating operands.                       | `%3: bool = cmp_ordered_float_less %1, %2`        |
+| `cmp_unordered_float_less`      | Is true when either operand is NaN, or left is less.    | Floating operands.                       | `%3: bool = cmp_unordered_float_less %1, %2`      |
+
+There are no greater-than opcodes in the current instruction set. Swap the
+operands and use the appropriate less-than form. Less-or-equal forms are also
+not defined yet.
+
+## Other opcodes
+
+### `select`
+
+Selects one of two equal-typed values using a boolean condition. It does not
+change control flow.
+
+```text
+%4: u32 = select %1, %2, %3
+```
+
+Here `%1` is `bool`; `%2`, `%3`, and `%4` share one type.
+
+### `bitcast`
+
+Reinterprets an operand's bits as the result type without performing a numeric
+conversion.
+
+```text
+%2: f32 = bitcast %1
+```
+
+The intended source and destination have equal total bit width. The current
+validator only requires that both values exist; it does not yet prove equal
+width.
+
+### `composite_construct`
+
+Constructs a vector or structure from its immediate elements.
+
+```text
+%5: vec4[f32] = composite_construct %1, %2, %3, %4
+```
+
+For a vector, every element must have the vector's element type and their count
+must equal its length. For a structure, each element must match the member at
+the same position. The validator does not support array construction yet.
+
+### `composite_extract`
+
+Traverses one or more literal indices through a vector, array, or structure and
+returns the selected nested member.
+
+```text
+%4: f32 = composite_extract %3[1][0]
+```
+
+At least one index is required. Every index must lie within its composite, and
+the result type must equal the selected member type.
+
+### `load_interface`
+
+Reads one declared shader input. It cannot read an output declaration.
+
+```text
+%1: vec4[f32] = load_interface @in_color
+```
+
+The result type must equal the interface variable's type. An optional dynamic
+`element_index` exists in memory for future arrayed interfaces, but the current
+printer does not show it and the validator does not use it to change the
+result type.
+
+### `store_interface`
+
+Writes one declared shader output. It produces no SSA result and cannot write an
+input declaration.
+
+```text
+store_interface @out_color, %1
+```
+
+The stored value must equal the interface variable's type. As with
+`load_interface`, an optional unprinted `element_index` is reserved for later
+arrayed-interface work. This operation has side effects.
+
+### `call`
+
+Invokes another IR function. Arguments must match the callee's parameters in
+number, order, and type.
+
+```text
+%4: vec4[f32] = call @shade(%1, %2)
+call @observe(%4)
+```
+
+A non-void callee requires a result of its return type; a void callee forbids
+one. Calls are conservatively treated as side-effecting. The operation exists in
+the common IR, although the current SPIR-V translator rejects
+`OpFunctionCall`.
+
+## Terminators
+
+Terminators yield no ordinary instruction result. They alone determine outgoing
+control-flow edges.
+
+| Terminator           | Description                                     | Usage                                                                           | Small printed example                      |
+| -------------------- | ----------------------------------------------- | ------------------------------------------------------------------------------- | ------------------------------------------ |
+| `branch`             | Unconditionally transfers control to one block. | Pass exactly one argument for every target parameter.                           | `branch .merge(%3)`                        |
+| `conditional_branch` | Selects one of two edges using a boolean.       | The condition is `bool`; each edge independently matches its target parameters. | `conditional_branch %1, .yes(%2), .no(%3)` |
+| `return` (void)      | Ends a void function.                           | The enclosing return type is `void`.                                            | `return`                                   |
+| `return` (value)     | Ends a function and returns a value.            | The value type equals the function return type.                                 | `return %3`                                |
+| `discard`            | Discards the fragment invocation.               | Fragment stage only.                                                            | `discard`                                  |
+| `unreachable`        | States that execution cannot reach this point.  | Any function; no successors.                                                    | `unreachable`                              |
+
+The Zig union names the return forms `return_void` and `return_value`; the
+printer renders both as the overloaded `return` spelling shown above.
+
+## Interfaces and builtins
+
+An interface variable has a type, a direction (`input` or `output`), and one
+semantic. Its semantic attributes are enclosed in the direction's brackets:
+
+- A location: `location(N), component(C), index(I)`.
+- A builtin: `builtin(name)`.
+
+The currently supported builtins are `position`, `vertex_index`, `instance_index`,
+`frag_coord`, `frag_depth`, and `global_invocation_id`.
+
+Printed declarations resemble these:
+
+```text
+@in_color: vec4[f32] = input[location(0), component(0), index(0)]
+@position: vec4[f32] = output[builtin(position)]
+```
+
+## Complete examples
+
+These examples use the printer's exact grammar and indentation. Their numeric
+value IDs are illustrative but follow the same module-wide numbering used by
+the printer.
+
+### A compute shader that adds two constants
+
+```text
+shader compute @main
+{
+    %0: constant u32 = bits(0x1)
+    %1: constant u32 = bits(0x2)
+
+    fn @main() -> void
+    {
+        .entry():
+            %2: u32 = integer_add %0, %1
+            return
+
+    }
+}
+```
+
+### A vertex interface passed through
+
+```text
+shader vertex @main
+{
+    @in_color: vec4[f32] = input[location(0), component(0), index(0)]
+    @out_color: vec4[f32] = output[location(0), component(0), index(0)]
+
+    fn @main() -> void
+    {
+        .entry():
+            %0: vec4[f32] = load_interface @in_color
+            store_interface @out_color, %0
+            return
+
+    }
+}
+```
+
+### A selection whose Phi becomes a block parameter
+
+```text
+shader compute @main
+{
+    %0: constant bool = true
+    %1: constant u32 = bits(0x1)
+    %2: constant u32 = bits(0x2)
+
+    fn @main() -> void
+    {
+        .entry():
+            conditional_branch %0, .left(), .right()
+
+        .left():
+            %3: u32 = integer_add %1, %2
+            branch .merge(%3)
+
+        .right():
+            %4: u32 = integer_subtract %2, %1
+            branch .merge(%4)
+
+        .merge(%5: u32):
+            %6: u32 = integer_multiply %5, %2
+            return
+
+    }
+}
+```
+
+The selection's merge metadata is not visible in this output, although it
+remains attached to `.entry` in memory.
+
+## Validator guarantees
+
+When `validator.validate` succeeds, it has proved the following:
+
+- The module has a live entry point.
+- All referenced types, constants, values, functions, blocks, and instructions
+  are live.
+- Parent links and SSA definition links agree in both directions.
+- Every function has an entry block, and no edge targets that entry block.
+- Every block has a terminator.
+- Edges remain within their function and exactly match target block parameters.
+- Returns agree with function return types; `discard` appears only in a fragment shader.
+- Interface loads read inputs; interface stores write outputs.
+- Operation-specific result presence and the foundational type equalities hold.
+- Structured merge and continue targets belong to the same function.
+
+The validator does not yet prove every semantic category listed in the opcode
+reference. In particular, several arithmetic opcodes can currently be built
+with an inappropriate but equal operand type; bitcast widths are not compared;
+shift-count rules are rudimentary; and floating-point execution modes are not
+attached to operations. Backends should explicitly require the properties and
+validation needed by their lowering.
+
+## Properties and passes
+
+The module carries independent property bits:
+
+- `valid_cfg`
+- `valid_ssa`
+- `structured_control_flow`
+- `no_function_calls`
+- `no_local_memory`
+- `no_matrix_types`
+- `no_large_composites`
+- `explicit_resource_offsets`
+
+A pass declares properties that it requires, produces, and invalidates. The pass
+manager rejects a pass whose requirements are missing, applies its property
+changes, and runs the validator after every pass by default.
+
+## Builder, rewriter, and visitor
+
+`Builder.zig` is the normal entry point for construction. It interns types and
+constants, stores copied slices in the module arena, adds functions and blocks,
+appends instructions, and assigns block terminators.
+
+`parser/root.zig` owns the public parsing entry points and recursive-descent grammar.
+Its implementation details are split by responsibility: `parser/Lexer.zig`
+tokenizes input, `parser/ast.zig` holds the temporary syntax model, and
+`parser/lower.zig` resolves that model into the common IR.
+
+`Rewriter.zig` provides the first safe mutation operations:
+
+- Count and replace SSA uses without changing definitions.
+- Erase a dead, side-effect-free instruction.
+- Redirect edges with a complete new argument list.
+- Add a block parameter while adding every incoming edge argument.
+- Remove a block parameter while removing its incoming arguments.
+
+`visitor.zig` walks module declarations, functions, blocks, instructions,
+terminators, and SSA uses in hierarchical order. `cfg.zig` computes
+predecessors, reachability, and dominance with a deliberately simple quadratic
+matrix. `validator/dominance.zig` contains the SSA dominance checks built on
+that analysis. This is suitable for the foundation but is not intended as the
+final large-shader implementation.
+
+## SPIR-V frontend
+
+The compiler currently provides a word parser and an initial translator in
+`spirv/`. The parser validates the header, word counts, truncation, and literal
+strings. The translator selects one entry point and lowers a defined subset:
+
+- Vertex, fragment, and compute stages.
+- Basic scalar, vector, array, structure, pointer, and function types.
+- Ordinary and composite constants; unapplied specialization constants are
+  refused.
+- Functions, blocks, branches, structured merge marks, and returns.
+- `OpPhi` into block parameters and edge arguments.
+- The arithmetic, comparison, select, bitcast, and composite operations named
+  in the reference above where mappings currently exist.
+- Decorated stage inputs and outputs, with interface load and store.
+- `OpName` debug names for functions, blocks, parameters, constants, and
+  instruction results when they are valid textual IR identifiers.
+
+Symbolic identifiers in SPIR-V assembly are assembler syntax and are not stored
+in the binary by `spirv-as`; add `OpName` instructions when those names must
+survive translation. Unsupported source instructions return an error; they are
+not preserved as opaque SPIR-V. This prevents silent mistranslation.
+
+## Running tests
+
+From the repository root, run:
+
+```sh
+zig build test-ir
+```
+
+The SPIR-V translation tests keep their assembly as multiline strings beside
+their assertions and pipe it through `spirv-as`. Therefore SPIRV-Tools must be
+on `PATH` when those tests run. Only malformed-binary parser tests use raw words,
+because an assembler cannot produce intentionally malformed instructions.
+
+## Documentation
+
+A complete codebase documentation can be found [here](https://vulkan-driver.kbz8.me/docs/ir/).

src/compiler/ir/Builder.zig

@@ -0,0 +1,227 @@
+const std = @import("std");
+const ids = @import("id.zig");
+const type_ir = @import("type.zig");
+const constant_ir = @import("constant.zig");
+const inst_ir = @import("instruction.zig");
+const module_ir = @import("module.zig");
+
+const Self = @This();
+
+module: *module_ir.Module,
+
+pub fn init(module: *module_ir.Module) Self {
+    return .{ .module = module };
+}
+
+fn copyName(self: *Self, name: ?[]const u8) !?[]const u8 {
+    return if (name) |text| try self.module.allocator().dupe(u8, text) else null;
+}
+
+pub fn internType(self: *Self, ty: type_ir.Type) !ids.TypeId {
+    return self.module.internType(ty);
+}
+
+pub fn internConstant(self: *Self, ty: ids.TypeId, candidate: constant_ir.ConstantValue) !ids.ValueId {
+    for (self.module.constants.entries.items, 0..) |entry, index| {
+        const existing = entry orelse continue;
+
+        if (existing.type == ty and constantEql(existing.value, candidate)) {
+            const constant_id = ids.ConstantId.fromIndex(index);
+
+            for (self.module.values.entries.items, 0..) |value_entry, value_index| {
+                const value = value_entry orelse continue;
+
+                if (value.definition == .constant and value.definition.constant == constant_id)
+                    return ids.ValueId.fromIndex(value_index);
+            }
+        }
+    }
+
+    var owned = candidate;
+    if (candidate == .composite)
+        owned.composite = try self.module.allocator().dupe(ids.ConstantId, candidate.composite);
+
+    const constant_id = try self.module.constants.add(self.module.allocator(), .{ .type = ty, .value = owned });
+    return self.module.values.add(self.module.allocator(), .{
+        .type = ty,
+        .definition = .{ .constant = constant_id },
+    });
+}
+
+pub fn setValueName(self: *Self, value_id: ids.ValueId, name: ?[]const u8) !void {
+    const text = name orelse return;
+    const value = self.module.values.getMut(value_id) orelse return error.InvalidValue;
+
+    if (value.name == null)
+        value.name = try self.copyName(text);
+}
+
+pub fn addFunction(self: *Self, return_type: ids.TypeId, name: ?[]const u8) !ids.FunctionId {
+    return self.module.functions.add(self.module.allocator(), .{
+        .return_type = return_type,
+        .name = try self.copyName(name),
+    });
+}
+
+pub fn setEntryPoint(self: *Self, function: ids.FunctionId) void {
+    self.module.entry_point = function;
+}
+
+pub fn addFunctionParameter(self: *Self, function_id: ids.FunctionId, ty: ids.TypeId, name: ?[]const u8) !ids.ValueId {
+    const function = self.module.functions.getMut(function_id) orelse return error.InvalidFunction;
+    const index: u32 = @intCast(function.parameters.items.len);
+
+    const value_id = try self.module.values.add(self.module.allocator(), .{
+        .type = ty,
+        .definition = .{ .function_parameter = .{ .function = function_id, .index = index } },
+        .name = try self.copyName(name),
+    });
+
+    try function.parameter_types.append(self.module.allocator(), ty);
+    try function.parameters.append(self.module.allocator(), value_id);
+
+    return value_id;
+}
+
+pub fn addBlock(self: *Self, function_id: ids.FunctionId, name: ?[]const u8) !ids.BlockId {
+    const function = self.module.functions.getMut(function_id) orelse return error.InvalidFunction;
+
+    const block_id = try self.module.blocks.add(self.module.allocator(), .{
+        .parent_function = function_id,
+        .name = try self.copyName(name),
+    });
+
+    try function.blocks.append(self.module.allocator(), block_id);
+    if (function.entry_block == null)
+        function.entry_block = block_id;
+
+    return block_id;
+}
+
+pub fn addBlockParameter(self: *Self, block_id: ids.BlockId, ty: ids.TypeId, name: ?[]const u8) !ids.ValueId {
+    const block = self.module.blocks.getMut(block_id) orelse return error.InvalidBlock;
+    const index: u32 = @intCast(block.parameters.items.len);
+
+    const value_id = try self.module.values.add(self.module.allocator(), .{
+        .type = ty,
+        .definition = .{ .block_parameter = .{ .block = block_id, .index = index } },
+        .name = try self.copyName(name),
+    });
+
+    try block.parameters.append(self.module.allocator(), value_id);
+
+    return value_id;
+}
+
+pub fn appendInstruction(
+    self: *Self,
+    block_id: ids.BlockId,
+    result_type: ?ids.TypeId,
+    operation: inst_ir.Operation,
+    name: ?[]const u8,
+) !?ids.ValueId {
+    const block = self.module.blocks.getMut(block_id) orelse return error.InvalidBlock;
+    const owned_operation = try self.copyOperation(operation);
+
+    const instruction_id = try self.module.instructions.add(self.module.allocator(), .{
+        .parent_block = block_id,
+        .result = null,
+        .operation = owned_operation,
+    });
+    errdefer _ = self.module.instructions.remove(instruction_id);
+
+    const result = if (result_type) |ty|
+        try self.module.values.add(self.module.allocator(), .{
+            .type = ty,
+            .definition = .{ .instruction = instruction_id },
+            .name = try self.copyName(name),
+        })
+    else
+        null;
+
+    self.module.instructions.getMut(instruction_id).?.result = result;
+    try block.instructions.append(self.module.allocator(), instruction_id);
+
+    return result;
+}
+
+pub fn setTerminator(self: *Self, block_id: ids.BlockId, terminator: module_ir.Terminator) !void {
+    const block = self.module.blocks.getMut(block_id) orelse return error.InvalidBlock;
+
+    if (block.terminator != null)
+        return error.TerminatorAlreadySet;
+
+    block.terminator = try self.copyTerminator(terminator);
+}
+
+pub fn addInterfaceVariable(
+    self: *Self,
+    ty: ids.TypeId,
+    direction: module_ir.InterfaceDirection,
+    semantic: module_ir.InterfaceSemantic,
+    name: ?[]const u8,
+) !ids.InterfaceVariableId {
+    return self.module.interface_variables.add(self.module.allocator(), .{
+        .type = ty,
+        .direction = direction,
+        .semantic = semantic,
+        .name = try self.copyName(name),
+    });
+}
+
+pub fn edge(self: *Self, target: ids.BlockId, arguments: []const ids.ValueId) !module_ir.Edge {
+    return .{
+        .target = target,
+        .arguments = try self.module.allocator().dupe(ids.ValueId, arguments),
+    };
+}
+
+fn copyOperation(self: *Self, operation: inst_ir.Operation) !inst_ir.Operation {
+    return switch (operation) {
+        .composite_construct => |op| .{
+            .composite_construct = .{
+                .elements = try self.module.allocator().dupe(ids.ValueId, op.elements),
+            },
+        },
+        .composite_extract => |op| .{
+            .composite_extract = .{
+                .composite = op.composite,
+                .indices = try self.module.allocator().dupe(u32, op.indices),
+            },
+        },
+        .call => |op| .{
+            .call = .{
+                .function = op.function,
+                .arguments = try self.module.allocator().dupe(ids.ValueId, op.arguments),
+            },
+        },
+        else => operation,
+    };
+}
+
+fn copyTerminator(self: *Self, terminator: module_ir.Terminator) !module_ir.Terminator {
+    return switch (terminator) {
+        .branch => |edge_value| .{
+            .branch = try self.edge(edge_value.target, edge_value.arguments),
+        },
+        .conditional_branch => |branch| .{
+            .conditional_branch = .{
+                .condition = branch.condition,
+                .true_edge = try self.edge(branch.true_edge.target, branch.true_edge.arguments),
+                .false_edge = try self.edge(branch.false_edge.target, branch.false_edge.arguments),
+            },
+        },
+        else => terminator,
+    };
+}
+
+fn constantEql(a: constant_ir.ConstantValue, b: constant_ir.ConstantValue) bool {
+    return switch (a) {
+        .boolean => |value| b == .boolean and value == b.boolean,
+        .integer_bits => |value| b == .integer_bits and value == b.integer_bits,
+        .float_bits => |value| b == .float_bits and value == b.float_bits,
+        .null => b == .null,
+        .undef => b == .undef,
+        .composite => |value| b == .composite and std.mem.eql(ids.ConstantId, value, b.composite),
+    };
+}

src/compiler/ir/Rewriter.zig

@@ -0,0 +1,323 @@
+const std = @import("std");
+const ids = @import("id.zig");
+const module_ir = @import("module.zig");
+const Builder = @import("Builder.zig");
+
+const Self = @This();
+
+pub const Error = std.mem.Allocator.Error || error{
+    InvalidValue,
+    InvalidInstruction,
+    InvalidBlock,
+    InvalidFunction,
+    TypeMismatch,
+    ResultStillUsed,
+    SideEffectingInstruction,
+    InstructionNotOwnedByBlock,
+    InvalidParameterIndex,
+    MissingIncomingValue,
+    UnexpectedIncomingValue,
+};
+
+pub const IncomingValue = struct {
+    predecessor: ids.BlockId,
+    value: ids.ValueId,
+};
+
+const UseCountContext = struct {
+    needle: ids.ValueId,
+    count: usize = 0,
+};
+
+module: *module_ir.Module,
+
+pub fn init(module: *module_ir.Module) Self {
+    return .{ .module = module };
+}
+
+pub fn countUses(self: *const Self, value: ids.ValueId) usize {
+    var context: UseCountContext = .{ .needle = value };
+    for (self.module.instructions.entries.items) |entry| {
+        const instruction = entry orelse continue;
+        instruction.operation.visitValueUses(&context, countUse);
+    }
+
+    for (self.module.blocks.entries.items) |entry| {
+        const block = entry orelse continue;
+        if (block.terminator) |terminator|
+            module_ir.visitTerminatorValueUses(terminator, &context, countUse);
+    }
+
+    return context.count;
+}
+
+pub fn replaceAllUses(self: *Self, old: ids.ValueId, replacement: ids.ValueId) Error!usize {
+    const old_value = self.module.values.get(old) orelse return error.InvalidValue;
+    const replacement_value = self.module.values.get(replacement) orelse return error.InvalidValue;
+
+    if (old_value.type != replacement_value.type)
+        return error.TypeMismatch;
+
+    if (old == replacement)
+        return 0;
+
+    var count: usize = 0;
+    for (self.module.instructions.entries.items) |*entry| {
+        const instruction = if (entry.*) |*value| value else continue;
+        count += try instruction.operation.replaceValueUses(self.module.allocator(), old, replacement);
+    }
+
+    for (self.module.blocks.entries.items) |*entry| {
+        const block = if (entry.*) |*value| value else continue;
+        if (block.terminator) |*terminator|
+            count += try module_ir.replaceTerminatorValueUses(self.module.allocator(), terminator, old, replacement);
+    }
+
+    return count;
+}
+
+pub fn eraseInstruction(self: *Self, instruction_id: ids.InstructionId) Error!void {
+    const instruction = self.module.instructions.get(instruction_id) orelse return error.InvalidInstruction;
+
+    if (instruction.operation.hasSideEffects())
+        return error.SideEffectingInstruction;
+
+    if (instruction.result) |result| {
+        if (self.countUses(result) != 0)
+            return error.ResultStillUsed;
+    }
+
+    const block = self.module.blocks.getMut(instruction.parent_block) orelse return error.InvalidBlock;
+    var owned_index: ?usize = null;
+
+    for (block.instructions.items, 0..) |candidate, index| {
+        if (candidate == instruction_id) {
+            owned_index = index;
+            break;
+        }
+    }
+
+    _ = block.instructions.orderedRemove(owned_index orelse return error.InstructionNotOwnedByBlock);
+    if (instruction.result) |result|
+        _ = self.module.values.remove(result);
+    _ = self.module.instructions.remove(instruction_id);
+}
+
+pub fn redirectEdges(
+    self: *Self,
+    source: ids.BlockId,
+    old_target: ids.BlockId,
+    new_target: ids.BlockId,
+    new_arguments: []const ids.ValueId,
+) Error!usize {
+    const source_block = self.module.blocks.get(source) orelse return error.InvalidBlock;
+    const target_block = self.module.blocks.get(new_target) orelse return error.InvalidBlock;
+
+    if (source_block.parent_function != target_block.parent_function)
+        return error.InvalidFunction;
+
+    try self.validateArguments(target_block, new_arguments);
+
+    const mutable_source = self.module.blocks.getMut(source).?;
+    const terminator = if (mutable_source.terminator) |*value| value else return error.InvalidBlock;
+
+    var count: usize = 0;
+
+    switch (terminator.*) {
+        .branch => |*edge| {
+            if (try self.redirectOne(edge, old_target, new_target, new_arguments))
+                count += 1;
+        },
+        .conditional_branch => |*branch| {
+            if (try self.redirectOne(&branch.true_edge, old_target, new_target, new_arguments))
+                count += 1;
+
+            if (try self.redirectOne(&branch.false_edge, old_target, new_target, new_arguments))
+                count += 1;
+        },
+        else => {},
+    }
+
+    return count;
+}
+
+pub fn addBlockParameter(
+    self: *Self,
+    block_id: ids.BlockId,
+    ty: ids.TypeId,
+    name: ?[]const u8,
+    incoming: []const IncomingValue,
+) Error!ids.ValueId {
+    const block = self.module.blocks.get(block_id) orelse return error.InvalidBlock;
+    const function = self.module.functions.get(block.parent_function) orelse return error.InvalidFunction;
+
+    for (incoming) |item| {
+        const value = self.module.values.get(item.value) orelse return error.InvalidValue;
+
+        if (value.type != ty)
+            return error.TypeMismatch;
+
+        if (!functionHasEdgeTo(self.module, function, item.predecessor, block_id))
+            return error.UnexpectedIncomingValue;
+    }
+    for (function.blocks.items) |predecessor| {
+        const edge_count = countEdgesTo(self.module.blocks.get(predecessor).?, block_id);
+        if (edge_count != 0 and findIncoming(incoming, predecessor) == null)
+            return error.MissingIncomingValue;
+    }
+
+    var builder = Builder.init(self.module);
+    const parameter = try builder.addBlockParameter(block_id, ty, name);
+
+    for (function.blocks.items) |predecessor| {
+        const incoming_value = findIncoming(incoming, predecessor) orelse continue;
+        try self.appendArgumentToEdges(predecessor, block_id, incoming_value);
+    }
+
+    return parameter;
+}
+
+pub fn removeBlockParameter(
+    self: *Self,
+    block_id: ids.BlockId,
+    parameter_index: usize,
+    replacement: ids.ValueId,
+) Error!void {
+    const block = self.module.blocks.get(block_id) orelse return error.InvalidBlock;
+    if (parameter_index >= block.parameters.items.len) return error.InvalidParameterIndex;
+    const parameter = block.parameters.items[parameter_index];
+    if (parameter == replacement) return error.InvalidValue;
+    _ = try self.replaceAllUses(parameter, replacement);
+
+    const function = self.module.functions.get(block.parent_function) orelse return error.InvalidFunction;
+    for (function.blocks.items) |predecessor| {
+        try self.removeArgumentFromEdges(predecessor, block_id, parameter_index);
+    }
+
+    const mutable_block = self.module.blocks.getMut(block_id).?;
+    _ = mutable_block.parameters.orderedRemove(parameter_index);
+
+    for (mutable_block.parameters.items[parameter_index..], parameter_index..) |value_id, index| {
+        const value = self.module.values.getMut(value_id) orelse return error.InvalidValue;
+        value.definition.block_parameter.index = @intCast(index);
+    }
+
+    _ = self.module.values.remove(parameter);
+}
+
+fn validateArguments(self: *const Self, target: *const module_ir.Block, arguments: []const ids.ValueId) Error!void {
+    if (arguments.len != target.parameters.items.len) return error.TypeMismatch;
+    for (arguments, target.parameters.items) |argument, parameter| {
+        const argument_value = self.module.values.get(argument) orelse return error.InvalidValue;
+        const parameter_value = self.module.values.get(parameter) orelse return error.InvalidValue;
+        if (argument_value.type != parameter_value.type) return error.TypeMismatch;
+    }
+}
+
+fn redirectOne(
+    self: *Self,
+    edge: *module_ir.Edge,
+    old_target: ids.BlockId,
+    new_target: ids.BlockId,
+    arguments: []const ids.ValueId,
+) !bool {
+    if (edge.target != old_target)
+        return false;
+
+    edge.target = new_target;
+    edge.arguments = try self.module.allocator().dupe(ids.ValueId, arguments);
+    return true;
+}
+
+fn appendArgumentToEdges(self: *Self, predecessor: ids.BlockId, target: ids.BlockId, value: ids.ValueId) !void {
+    const block = self.module.blocks.getMut(predecessor) orelse return error.InvalidBlock;
+    const terminator = if (block.terminator) |*item| item else return error.InvalidBlock;
+
+    switch (terminator.*) {
+        .branch => |*edge| {
+            if (edge.target == target)
+                try self.appendEdgeArgument(edge, value);
+        },
+        .conditional_branch => |*branch| {
+            if (branch.true_edge.target == target)
+                try self.appendEdgeArgument(&branch.true_edge, value);
+
+            if (branch.false_edge.target == target)
+                try self.appendEdgeArgument(&branch.false_edge, value);
+        },
+        else => {},
+    }
+}
+
+fn appendEdgeArgument(self: *Self, edge: *module_ir.Edge, value: ids.ValueId) !void {
+    const arguments = try self.module.allocator().alloc(ids.ValueId, edge.arguments.len + 1);
+    @memcpy(arguments[0..edge.arguments.len], edge.arguments);
+    arguments[edge.arguments.len] = value;
+    edge.arguments = arguments;
+}
+
+fn removeArgumentFromEdges(self: *Self, predecessor: ids.BlockId, target: ids.BlockId, index: usize) !void {
+    const block = self.module.blocks.getMut(predecessor) orelse return error.InvalidBlock;
+    const terminator = if (block.terminator) |*item| item else return error.InvalidBlock;
+
+    switch (terminator.*) {
+        .branch => |*edge| {
+            if (edge.target == target)
+                try self.removeEdgeArgument(edge, index);
+        },
+        .conditional_branch => |*branch| {
+            if (branch.true_edge.target == target)
+                try self.removeEdgeArgument(&branch.true_edge, index);
+
+            if (branch.false_edge.target == target)
+                try self.removeEdgeArgument(&branch.false_edge, index);
+        },
+        else => {},
+    }
+}
+
+fn removeEdgeArgument(self: *Self, edge: *module_ir.Edge, index: usize) !void {
+    if (index >= edge.arguments.len)
+        return error.InvalidParameterIndex;
+
+    const arguments = try self.module.allocator().alloc(ids.ValueId, edge.arguments.len - 1);
+    @memcpy(arguments[0..index], edge.arguments[0..index]);
+    @memcpy(arguments[index..], edge.arguments[index + 1 ..]);
+    edge.arguments = arguments;
+}
+
+fn countUse(context: *UseCountContext, value: ids.ValueId) void {
+    if (value == context.needle)
+        context.count += 1;
+}
+
+fn findIncoming(incoming: []const IncomingValue, predecessor: ids.BlockId) ?ids.ValueId {
+    for (incoming) |item| {
+        if (item.predecessor == predecessor)
+            return item.value;
+    }
+    return null;
+}
+
+fn functionHasEdgeTo(
+    module: *const module_ir.Module,
+    function: *const module_ir.Function,
+    predecessor: ids.BlockId,
+    target: ids.BlockId,
+) bool {
+    for (function.blocks.items) |block_id| {
+        if (block_id != predecessor)
+            continue;
+        return countEdgesTo(module.blocks.get(block_id) orelse return false, target) != 0;
+    }
+    return false;
+}
+
+fn countEdgesTo(block: *const module_ir.Block, target: ids.BlockId) usize {
+    const terminator = block.terminator orelse return 0;
+    return switch (terminator) {
+        .branch => |edge| @intFromBool(edge.target == target),
+        .conditional_branch => |branch| @intFromBool(branch.true_edge.target == target) + @intFromBool(branch.false_edge.target == target),
+        else => 0,
+    };
+}

src/compiler/ir/cfg.zig

@@ -0,0 +1,198 @@
+const std = @import("std");
+const ids = @import("id.zig");
+const module_ir = @import("module.zig");
+
+pub const Self = @This();
+
+pub const Error = std.mem.Allocator.Error || error{
+    InvalidFunction,
+    MissingEntryBlock,
+    InvalidBlock,
+    MissingTerminator,
+    CrossFunctionEdge,
+};
+
+allocator: std.mem.Allocator,
+blocks: []ids.BlockId,
+predecessors_by_block: []std.ArrayList(ids.BlockId),
+reachable: []bool,
+dominators: []bool,
+
+pub fn init(allocator: std.mem.Allocator, module: *const module_ir.Module, function_id: ids.FunctionId) Error!Self {
+    const function = module.functions.get(function_id) orelse return error.InvalidFunction;
+    const entry = function.entry_block orelse return error.MissingEntryBlock;
+    const blocks = try allocator.dupe(ids.BlockId, function.blocks.items);
+    errdefer allocator.free(blocks);
+
+    const predecessor_lists = try allocator.alloc(std.ArrayList(ids.BlockId), blocks.len);
+    errdefer allocator.free(predecessor_lists);
+
+    for (predecessor_lists) |*list|
+        list.* = .empty;
+
+    errdefer for (predecessor_lists) |*list| list.deinit(allocator);
+
+    const reachable = try allocator.alloc(bool, blocks.len);
+    errdefer allocator.free(reachable);
+    @memset(reachable, false);
+
+    const dominators = try allocator.alloc(bool, blocks.len * blocks.len);
+    errdefer allocator.free(dominators);
+    @memset(dominators, false);
+
+    var self: Self = .{
+        .allocator = allocator,
+        .blocks = blocks,
+        .predecessors_by_block = predecessor_lists,
+        .reachable = reachable,
+        .dominators = dominators,
+    };
+
+    try self.buildPredecessors(module);
+    try self.buildReachability(module, entry);
+    self.buildDominators(entry);
+
+    return self;
+}
+
+pub fn deinit(self: *Self) void {
+    for (self.predecessors_by_block) |*list|
+        list.deinit(self.allocator);
+
+    self.allocator.free(self.predecessors_by_block);
+    self.allocator.free(self.blocks);
+    self.allocator.free(self.reachable);
+    self.allocator.free(self.dominators);
+    self.* = undefined;
+}
+
+pub fn predecessors(self: *const Self, block: ids.BlockId) ?[]const ids.BlockId {
+    const index = self.indexOf(block) orelse return null;
+    return self.predecessors_by_block[index].items;
+}
+
+pub fn isReachable(self: *const Self, block: ids.BlockId) bool {
+    const index = self.indexOf(block) orelse return false;
+    return self.reachable[index];
+}
+
+pub fn dominates(self: *const Self, dominator: ids.BlockId, block: ids.BlockId) bool {
+    const dominator_index = self.indexOf(dominator) orelse return false;
+    const block_index = self.indexOf(block) orelse return false;
+    return self.dominators[block_index * self.blocks.len + dominator_index];
+}
+
+fn buildPredecessors(self: *Self, module: *const module_ir.Module) Error!void {
+    for (self.blocks) |source| {
+        const block = module.blocks.get(source) orelse return error.InvalidBlock;
+        const terminator = block.terminator orelse return error.MissingTerminator;
+        switch (terminator) {
+            .branch => |edge| try self.addPredecessor(edge.target, source),
+            .conditional_branch => |branch| {
+                try self.addPredecessor(branch.true_edge.target, source);
+                try self.addPredecessor(branch.false_edge.target, source);
+            },
+            else => {},
+        }
+    }
+}
+
+fn buildReachability(self: *Self, module: *const module_ir.Module, entry: ids.BlockId) Error!void {
+    var queue: std.ArrayList(ids.BlockId) = .empty;
+    defer queue.deinit(self.allocator);
+    try queue.append(self.allocator, entry);
+    self.reachable[self.indexOf(entry) orelse return error.InvalidBlock] = true;
+
+    var cursor: usize = 0;
+    while (cursor < queue.items.len) : (cursor += 1) {
+        const block = module.blocks.get(queue.items[cursor]) orelse return error.InvalidBlock;
+        const terminator = block.terminator orelse return error.MissingTerminator;
+        switch (terminator) {
+            .branch => |edge| try self.markReachable(&queue, edge.target),
+            .conditional_branch => |branch| {
+                try self.markReachable(&queue, branch.true_edge.target);
+                try self.markReachable(&queue, branch.false_edge.target);
+            },
+            else => {},
+        }
+    }
+}
+
+fn buildDominators(self: *Self, entry: ids.BlockId) void {
+    const entry_index = self.indexOf(entry).?;
+    const count = self.blocks.len;
+
+    for (0..count) |block_index| {
+        if (!self.reachable[block_index]) {
+            self.setDominates(block_index, block_index, true);
+        } else if (block_index == entry_index) {
+            self.setDominates(block_index, entry_index, true);
+        } else {
+            for (0..count) |candidate| {
+                if (self.reachable[candidate])
+                    self.setDominates(block_index, candidate, true);
+            }
+        }
+    }
+
+    var changed = true;
+    while (changed) {
+        changed = false;
+        for (0..count) |block_index| {
+            if (!self.reachable[block_index] or block_index == entry_index)
+                continue;
+
+            for (0..count) |candidate| {
+                var new_value = candidate == block_index;
+                if (!new_value) {
+                    var saw_reachable_predecessor = false;
+                    new_value = true;
+                    for (self.predecessors_by_block[block_index].items) |predecessor| {
+                        const predecessor_index = self.indexOf(predecessor).?;
+
+                        if (!self.reachable[predecessor_index])
+                            continue;
+
+                        saw_reachable_predecessor = true;
+                        new_value = new_value and self.getDominates(predecessor_index, candidate);
+                    }
+                    new_value = new_value and saw_reachable_predecessor;
+                }
+                if (self.getDominates(block_index, candidate) != new_value) {
+                    self.setDominates(block_index, candidate, new_value);
+                    changed = true;
+                }
+            }
+        }
+    }
+}
+
+fn addPredecessor(self: *Self, target: ids.BlockId, source: ids.BlockId) Error!void {
+    const target_index = self.indexOf(target) orelse return error.CrossFunctionEdge;
+    try self.predecessors_by_block[target_index].append(self.allocator, source);
+}
+
+fn markReachable(self: *Self, queue: *std.ArrayList(ids.BlockId), target: ids.BlockId) Error!void {
+    const target_index = self.indexOf(target) orelse return error.CrossFunctionEdge;
+
+    if (self.reachable[target_index])
+        return;
+
+    self.reachable[target_index] = true;
+    try queue.append(self.allocator, target);
+}
+
+fn indexOf(self: *const Self, block: ids.BlockId) ?usize {
+    for (self.blocks, 0..) |candidate, index| {
+        if (candidate == block) return index;
+    }
+    return null;
+}
+
+fn getDominates(self: *const Self, block_index: usize, candidate_index: usize) bool {
+    return self.dominators[block_index * self.blocks.len + candidate_index];
+}
+
+fn setDominates(self: *Self, block_index: usize, candidate_index: usize, value: bool) void {
+    self.dominators[block_index * self.blocks.len + candidate_index] = value;
+}

src/compiler/ir/constant.zig

@@ -0,0 +1,18 @@
+const ids = @import("id.zig");
+
+pub const ConstantId = ids.ConstantId;
+pub const TypeId = ids.TypeId;
+
+pub const ConstantValue = union(enum) {
+    boolean: bool,
+    integer_bits: u64,
+    float_bits: u64,
+    null,
+    undef,
+    composite: []const ConstantId,
+};
+
+pub const Constant = struct {
+    type: TypeId,
+    value: ConstantValue,
+};

src/compiler/ir/id.zig

@@ -0,0 +1,84 @@
+const std = @import("std");
+
+pub const TypeTag = opaque {};
+pub const ConstantTag = opaque {};
+pub const ValueTag = opaque {};
+pub const InstructionTag = opaque {};
+pub const BlockTag = opaque {};
+pub const FunctionTag = opaque {};
+pub const InterfaceVariableTag = opaque {};
+pub const ResourceTag = opaque {};
+
+pub const TypeId = Id(TypeTag);
+pub const ConstantId = Id(ConstantTag);
+pub const ValueId = Id(ValueTag);
+pub const InstructionId = Id(InstructionTag);
+pub const BlockId = Id(BlockTag);
+pub const FunctionId = Id(FunctionTag);
+pub const InterfaceVariableId = Id(InterfaceVariableTag);
+pub const ResourceId = Id(ResourceTag);
+
+pub fn Id(comptime Tag: type) type {
+    return enum(u32) {
+        _,
+
+        pub const tag_type = Tag;
+
+        pub fn fromIndex(item_index: usize) @This() {
+            std.debug.assert(item_index <= std.math.maxInt(u32));
+            return @enumFromInt(item_index);
+        }
+
+        pub fn index(self: @This()) usize {
+            return @intFromEnum(self);
+        }
+    };
+}
+
+pub fn Store(comptime IdType: type, comptime T: type) type {
+    return struct {
+        const Self = @This();
+
+        entries: std.ArrayList(?T) = .empty,
+
+        pub fn add(self: *Self, allocator: std.mem.Allocator, value: T) !IdType {
+            const id = IdType.fromIndex(self.entries.items.len);
+            try self.entries.append(allocator, value);
+            return id;
+        }
+
+        pub fn get(self: *const Self, id: IdType) ?*const T {
+            if (id.index() >= self.entries.items.len)
+                return null;
+
+            const entry = &self.entries.items[id.index()];
+            return if (entry.*) |*value| value else null;
+        }
+
+        pub fn getMut(self: *Self, id: IdType) ?*T {
+            if (id.index() >= self.entries.items.len)
+                return null;
+
+            const entry = &self.entries.items[id.index()];
+            return if (entry.*) |*value| value else null;
+        }
+
+        /// Removing an object leaves a tombstone as IDs are deliberately not recycled
+        /// so they are never silently redirected to a different object.
+        pub fn remove(self: *Self, id: IdType) bool {
+            if (id.index() >= self.entries.items.len)
+                return false;
+
+            const entry = &self.entries.items[id.index()];
+            if (entry.* == null)
+                return false;
+
+            entry.* = null;
+            return true;
+        }
+
+        pub fn isLive(self: *const Self, id: IdType) bool {
+            return self.get(id) != null;
+        }
+    };
+}

src/compiler/ir/instruction.zig

@@ -0,0 +1,228 @@
+const std = @import("std");
+const ids = @import("id.zig");
+
+pub const TypeId = ids.TypeId;
+pub const ValueId = ids.ValueId;
+pub const BlockId = ids.BlockId;
+pub const FunctionId = ids.FunctionId;
+pub const InterfaceVariableId = ids.InterfaceVariableId;
+
+pub const SourceLocation = struct {
+    file: ?[]const u8 = null,
+    line: u32,
+    column: u32,
+};
+
+pub const UnaryOpcode = enum {
+    negate,
+    logical_not,
+    bitwise_not,
+};
+
+pub const BinaryOpcode = enum {
+    integer_add,
+    integer_subtract,
+    integer_multiply,
+    unsigned_divide,
+    signed_divide,
+    unsigned_modulo,
+    signed_modulo,
+    float_add,
+    float_subtract,
+    float_multiply,
+    float_divide,
+    float_modulo,
+    shift_left,
+    logical_shift_right,
+    arithmetic_shift_right,
+    bitwise_and,
+    bitwise_or,
+    bitwise_xor,
+    logical_and,
+    logical_or,
+};
+
+pub const CompareOpcode = enum {
+    equal,
+    not_equal,
+    unsigned_less,
+    signed_less,
+    ordered_float_equal,
+    unordered_float_equal,
+    ordered_float_not_equal,
+    unordered_float_not_equal,
+    ordered_float_less,
+    unordered_float_less,
+};
+
+pub const Unary = struct {
+    opcode: UnaryOpcode,
+    operand: ValueId,
+};
+
+pub const Binary = struct {
+    opcode: BinaryOpcode,
+    lhs: ValueId,
+    rhs: ValueId,
+};
+
+pub const Compare = struct {
+    opcode: CompareOpcode,
+    lhs: ValueId,
+    rhs: ValueId,
+};
+
+pub const Select = struct {
+    condition: ValueId,
+    true_value: ValueId,
+    false_value: ValueId,
+};
+
+pub const CompositeConstruct = struct {
+    elements: []const ValueId,
+};
+
+pub const CompositeExtract = struct {
+    composite: ValueId,
+    indices: []const u32,
+};
+
+pub const LoadInterface = struct {
+    variable: InterfaceVariableId,
+    element_index: ?ValueId = null,
+};
+
+pub const StoreInterface = struct {
+    variable: InterfaceVariableId,
+    value: ValueId,
+    element_index: ?ValueId = null,
+};
+
+pub const Call = struct {
+    function: FunctionId,
+    arguments: []const ValueId,
+};
+
+pub const Operation = union(enum) {
+    unary: Unary,
+    binary: Binary,
+    compare: Compare,
+    select: Select,
+    bitcast: ValueId,
+    composite_construct: CompositeConstruct,
+    composite_extract: CompositeExtract,
+    load_interface: LoadInterface,
+    store_interface: StoreInterface,
+    call: Call,
+
+    pub fn visitValueUses(self: Operation, context: anytype, comptime visitor: anytype) void {
+        switch (self) {
+            .unary => |op| visitor(context, op.operand),
+            .binary => |op| {
+                visitor(context, op.lhs);
+                visitor(context, op.rhs);
+            },
+            .compare => |op| {
+                visitor(context, op.lhs);
+                visitor(context, op.rhs);
+            },
+            .select => |op| {
+                visitor(context, op.condition);
+                visitor(context, op.true_value);
+                visitor(context, op.false_value);
+            },
+            .bitcast => |operand| visitor(context, operand),
+            .composite_construct => |op| for (op.elements) |element| visitor(context, element),
+            .composite_extract => |op| visitor(context, op.composite),
+            .load_interface => |op| if (op.element_index) |index| visitor(context, index),
+            .store_interface => |op| {
+                visitor(context, op.value);
+                if (op.element_index) |index|
+                    visitor(context, index);
+            },
+            .call => |op| {
+                for (op.arguments) |argument|
+                    visitor(context, argument);
+            },
+        }
+    }
+
+    pub fn replaceValueUses(self: *Operation, allocator: std.mem.Allocator, old: ValueId, replacement: ValueId) !usize {
+        var count: usize = 0;
+        switch (self.*) {
+            .unary => |*op| replaceOne(&op.operand, old, replacement, &count),
+            .binary => |*op| {
+                replaceOne(&op.lhs, old, replacement, &count);
+                replaceOne(&op.rhs, old, replacement, &count);
+            },
+            .compare => |*op| {
+                replaceOne(&op.lhs, old, replacement, &count);
+                replaceOne(&op.rhs, old, replacement, &count);
+            },
+            .select => |*op| {
+                replaceOne(&op.condition, old, replacement, &count);
+                replaceOne(&op.true_value, old, replacement, &count);
+                replaceOne(&op.false_value, old, replacement, &count);
+            },
+            .bitcast => |*operand| replaceOne(operand, old, replacement, &count),
+            .composite_construct => |*op| op.elements = try replaceSlice(allocator, op.elements, old, replacement, &count),
+            .composite_extract => |*op| replaceOne(&op.composite, old, replacement, &count),
+            .load_interface => |*op| {
+                if (op.element_index) |*index|
+                    replaceOne(index, old, replacement, &count);
+            },
+            .store_interface => |*op| {
+                replaceOne(&op.value, old, replacement, &count);
+                if (op.element_index) |*index|
+                    replaceOne(index, old, replacement, &count);
+            },
+            .call => |*op| op.arguments = try replaceSlice(allocator, op.arguments, old, replacement, &count),
+        }
+        return count;
+    }
+
+    pub fn hasSideEffects(self: Operation) bool {
+        return switch (self) {
+            .store_interface, .call => true,
+            else => false,
+        };
+    }
+};
+
+pub const Instruction = struct {
+    parent_block: BlockId,
+    result: ?ValueId,
+    operation: Operation,
+    source: ?SourceLocation = null,
+};
+
+fn replaceOne(operand: *ValueId, old: ValueId, replacement: ValueId, count: *usize) void {
+    if (operand.* != old) return;
+    operand.* = replacement;
+    count.* += 1;
+}
+
+fn replaceSlice(
+    allocator: std.mem.Allocator,
+    operands: []const ValueId,
+    old: ValueId,
+    replacement: ValueId,
+    count: *usize,
+) ![]const ValueId {
+    var occurrences: usize = 0;
+    for (operands) |operand| if (operand == old) {
+        occurrences += 1;
+    };
+
+    if (occurrences == 0)
+        return operands;
+
+    const copy = try allocator.dupe(ValueId, operands);
+    for (copy) |*operand| {
+        if (operand.* == old)
+            operand.* = replacement;
+    }
+
+    count.* += occurrences;
+    return copy;
+}

src/compiler/ir/ir.zig

@@ -0,0 +1,58 @@
+//! ## Intermediate Representation
+//!
+//! The IR is the compiler's target-independent shader representation. It keeps
+//! shader structure explicit while hiding SPIR-V's binary encoding and large
+//! instruction surface.
+//!
+//! `module.Module` owns types, constants, values, instructions, blocks, functions,
+//! interfaces, and resources in typed ID stores. Values are in SSA form: each value
+//! is defined by a constant, parameter, instruction, or `undef`.
+//!
+//! Control flow is represented with basic blocks and terminators. Phi-like values
+//! are modeled as block parameters, with branch edge arguments supplying incoming
+//! values. Blocks can also record structured selection or loop metadata.
+//!
+//! Use `Builder` to construct modules, `validator.validate` to check invariants,
+//! `cfg` for control-flow queries, `Rewriter` for common edits, and
+//! `parser`/`printer` for the textual IR format used by tests and debugging.
+//!
+//! Here's a simple text representation of a shader module that `parser.Parser` and `printer` can handle/produce:
+//! ```
+//! shader vertex @main
+//! {
+//!     @color: vec4[f32] = input[location(0), component(0), index(0)]
+//!     @out_color: vec4[f32] = output[location(0), component(0), index(0)]
+//!     %0: constant bool = true
+//!     %1: constant f32 = bits(0x3f800000)
+//!
+//!     fn @main() -> void
+//!     {
+//!         .entry():
+//!             %3: vec4[f32] = load_interface @color
+//!             conditional_branch %0, .pass(), .merge(%3)
+//!
+//!         .pass():
+//!             %4: vec4[f32] = composite_construct %1, %1, %1, %1
+//!             branch .merge(%4)
+//!
+//!         .merge(%2: vec4[f32]):
+//!             store_interface @out_color, %2
+//!             return
+//!     }
+//! }
+//! ```
+
+pub const Builder = @import("Builder.zig");
+pub const Rewriter = @import("Rewriter.zig");
+pub const cfg = @import("cfg.zig");
+pub const constant = @import("constant.zig");
+pub const id = @import("id.zig");
+pub const instruction = @import("instruction.zig");
+pub const module = @import("module.zig");
+pub const parser = @import("parser/parser.zig");
+pub const pass_manager = @import("pass_manager.zig");
+pub const printer = @import("printer.zig");
+pub const types = @import("type.zig");
+pub const validator = @import("validator/validator.zig");
+pub const value = @import("value.zig");
+pub const visitor = @import("visitor.zig");

src/compiler/ir/module.zig

@@ -0,0 +1,235 @@
+const std = @import("std");
+const ids = @import("id.zig");
+const types = @import("type.zig");
+const constants = @import("constant.zig");
+const values = @import("value.zig");
+const instructions = @import("instruction.zig");
+
+pub const Stage = enum {
+    vertex,
+    fragment,
+    compute,
+};
+
+pub const ExecutionModes = struct {
+    workgroup_size: ?[3]u32 = null,
+    early_fragment_tests: bool = false,
+};
+
+pub const Properties = packed struct {
+    valid_cfg: bool = false,
+    valid_ssa: bool = false,
+    structured_control_flow: bool = false,
+    no_function_calls: bool = false,
+    no_local_memory: bool = false,
+    no_matrix_types: bool = false,
+    no_large_composites: bool = false,
+    explicit_resource_offsets: bool = false,
+    _padding: u24 = 0,
+};
+
+pub const ConstantStore = ids.Store(ids.ConstantId, constants.Constant);
+pub const ValueStore = ids.Store(ids.ValueId, values.Value);
+pub const InstructionStore = ids.Store(ids.InstructionId, instructions.Instruction);
+pub const BlockStore = ids.Store(ids.BlockId, Block);
+pub const FunctionStore = ids.Store(ids.FunctionId, Function);
+pub const InterfaceVariableStore = ids.Store(ids.InterfaceVariableId, InterfaceVariable);
+pub const ResourceStore = ids.Store(ids.ResourceId, Resource);
+pub const TypeStore = ids.Store(ids.TypeId, types.Type);
+
+pub const Edge = struct {
+    target: ids.BlockId,
+    arguments: []const ids.ValueId,
+};
+
+pub const Terminator = union(enum) {
+    branch: Edge,
+    conditional_branch: struct {
+        condition: ids.ValueId,
+        true_edge: Edge,
+        false_edge: Edge,
+    },
+    return_void,
+    return_value: ids.ValueId,
+    discard,
+    @"unreachable",
+};
+
+pub const StructuredControl = union(enum) {
+    none,
+    selection: struct { merge_block: ids.BlockId },
+    loop: struct {
+        merge_block: ids.BlockId,
+        continue_block: ids.BlockId,
+    },
+};
+
+pub const Block = struct {
+    parent_function: ids.FunctionId,
+    parameters: std.ArrayList(ids.ValueId) = .empty,
+    instructions: std.ArrayList(ids.InstructionId) = .empty,
+    terminator: ?Terminator = null,
+    structured_control: StructuredControl = .none,
+    name: ?[]const u8 = null,
+};
+
+pub const Function = struct {
+    return_type: ids.TypeId,
+    parameter_types: std.ArrayList(ids.TypeId) = .empty,
+    parameters: std.ArrayList(ids.ValueId) = .empty,
+    blocks: std.ArrayList(ids.BlockId) = .empty,
+    entry_block: ?ids.BlockId = null,
+    name: ?[]const u8 = null,
+};
+
+pub const InterfaceDirection = enum {
+    input,
+    output,
+};
+
+pub const Builtin = enum {
+    position,
+    vertex_index,
+    instance_index,
+    frag_coord,
+    frag_depth,
+    global_invocation_id,
+};
+
+pub const InterfaceSemantic = union(enum) {
+    location: struct { location: u32, component: u8 = 0, index: u8 = 0 },
+    builtin: Builtin,
+};
+
+pub const InterfaceVariable = struct {
+    type: ids.TypeId,
+    direction: InterfaceDirection,
+    semantic: InterfaceSemantic,
+    name: ?[]const u8 = null,
+};
+
+pub const Resource = struct {
+    kind: types.ResourceKind,
+    set: u32,
+    binding: u32,
+    type: ids.TypeId,
+    name: ?[]const u8 = null,
+};
+
+pub const Module = struct {
+    arena: std.heap.ArenaAllocator,
+    stage: Stage,
+    entry_point: ?ids.FunctionId = null,
+    execution_modes: ExecutionModes = .{},
+    types: TypeStore = .{},
+    constants: ConstantStore = .{},
+    values: ValueStore = .{},
+    instructions: InstructionStore = .{},
+    blocks: BlockStore = .{},
+    functions: FunctionStore = .{},
+    interface_variables: InterfaceVariableStore = .{},
+    resources: ResourceStore = .{},
+    properties: Properties = .{},
+
+    pub fn init(backing_allocator: std.mem.Allocator, stage: Stage) Module {
+        return .{
+            .arena = std.heap.ArenaAllocator.init(backing_allocator),
+            .stage = stage,
+        };
+    }
+
+    pub fn deinit(self: *Module) void {
+        self.arena.deinit();
+        self.* = undefined;
+    }
+
+    pub fn allocator(self: *Module) std.mem.Allocator {
+        return self.arena.allocator();
+    }
+
+    pub fn backingAllocator(self: *const Module) std.mem.Allocator {
+        return self.arena.child_allocator;
+    }
+
+    pub fn internType(self: *Module, candidate: types.Type) !ids.TypeId {
+        for (self.types.entries.items, 0..) |entry, index| {
+            if (entry) |existing| {
+                if (existing.eql(candidate))
+                    return ids.TypeId.fromIndex(index);
+            }
+        }
+
+        var owned = candidate;
+        if (candidate == .structure) {
+            owned.structure.members = try self.allocator().dupe(ids.TypeId, candidate.structure.members);
+        }
+
+        return self.types.add(self.allocator(), owned);
+    }
+
+    pub fn typeOf(self: *const Module, value_id: ids.ValueId) ?ids.TypeId {
+        const value = self.values.get(value_id) orelse return null;
+        return value.type;
+    }
+};
+
+pub fn visitTerminatorValueUses(terminator: Terminator, context: anytype, comptime visitor: anytype) void {
+    switch (terminator) {
+        .branch => |edge| {
+            for (edge.arguments) |argument|
+                visitor(context, argument);
+        },
+        .conditional_branch => |branch| {
+            visitor(context, branch.condition);
+
+            for (branch.true_edge.arguments) |argument|
+                visitor(context, argument);
+
+            for (branch.false_edge.arguments) |argument|
+                visitor(context, argument);
+        },
+        .return_value => |value| visitor(context, value),
+        else => {},
+    }
+}
+
+pub fn replaceTerminatorValueUses(allocator: std.mem.Allocator, terminator: *Terminator, old: ids.ValueId, replacement: ids.ValueId) !usize {
+    var count: usize = 0;
+    switch (terminator.*) {
+        .branch => |*edge| try replaceEdgeUses(allocator, edge, old, replacement, &count),
+        .conditional_branch => |*branch| {
+            replaceOne(&branch.condition, old, replacement, &count);
+            try replaceEdgeUses(allocator, &branch.true_edge, old, replacement, &count);
+            try replaceEdgeUses(allocator, &branch.false_edge, old, replacement, &count);
+        },
+        .return_value => |*value| replaceOne(value, old, replacement, &count),
+        else => {},
+    }
+    return count;
+}
+
+fn replaceEdgeUses(allocator: std.mem.Allocator, edge: *Edge, old: ids.ValueId, replacement: ids.ValueId, count: *usize) !void {
+    var occurrences: usize = 0;
+    for (edge.arguments) |argument| if (argument == old) {
+        occurrences += 1;
+    };
+
+    if (occurrences == 0)
+        return;
+
+    const copy = try allocator.dupe(ids.ValueId, edge.arguments);
+    for (copy) |*argument| {
+        if (argument.* == old)
+            argument.* = replacement;
+    }
+
+    edge.arguments = copy;
+    count.* += occurrences;
+}
+
+fn replaceOne(operand: *ids.ValueId, old: ids.ValueId, replacement: ids.ValueId, count: *usize) void {
+    if (operand.* != old)
+        return;
+    operand.* = replacement;
+    count.* += 1;
+}

src/compiler/ir/parser/Lexer.zig

@@ -0,0 +1,196 @@
+const std = @import("std");
+
+const Self = @This();
+
+source: []const u8,
+cursor: usize = 0,
+lookahead: ?Token = null,
+
+pub const TokenTag = enum {
+    eof,
+    invalid,
+    identifier,
+    number,
+    value_ref,
+    constant_ref,
+    at_name,
+    dot_name,
+    left_brace,
+    right_brace,
+    left_paren,
+    right_paren,
+    left_square,
+    right_square,
+    colon,
+    comma,
+    equal,
+    arrow,
+};
+
+pub const Token = struct {
+    tag: TokenTag,
+    text: []const u8,
+};
+
+pub fn init(source: []const u8) Self {
+    return .{
+        .source = source,
+    };
+}
+
+pub fn peek(self: *Self) Token {
+    if (self.lookahead == null)
+        self.lookahead = self.lex();
+    return self.lookahead.?;
+}
+
+pub fn take(self: *Self) Token {
+    const token = self.peek();
+    self.lookahead = null;
+    return token;
+}
+
+fn lex(self: *Self) Token {
+    while (self.cursor < self.source.len and std.ascii.isWhitespace(self.source[self.cursor]))
+        self.cursor += 1;
+
+    if (self.cursor == self.source.len) {
+        return .{
+            .tag = .eof,
+            .text = self.source[self.cursor..self.cursor],
+        };
+    }
+
+    const start = self.cursor;
+    const byte = self.source[self.cursor];
+    self.cursor += 1;
+
+    switch (byte) {
+        '{' => return self.simpleToken(.left_brace, start),
+        '}' => return self.simpleToken(.right_brace, start),
+        '(' => return self.simpleToken(.left_paren, start),
+        ')' => return self.simpleToken(.right_paren, start),
+        '[' => return self.simpleToken(.left_square, start),
+        ']' => return self.simpleToken(.right_square, start),
+        ':' => return self.simpleToken(.colon, start),
+        ',' => return self.simpleToken(.comma, start),
+        '=' => return self.simpleToken(.equal, start),
+        '-', '+' => {
+            if (byte == '-' and self.cursor < self.source.len and self.source[self.cursor] == '>') {
+                self.cursor += 1;
+                return .{
+                    .tag = .arrow,
+                    .text = self.source[start..self.cursor],
+                };
+            }
+
+            if (self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor]))
+                return self.numberToken(start);
+
+            return self.simpleToken(.invalid, start);
+        },
+        '%', '@', '.' => {
+            const tag: TokenTag = switch (byte) {
+                '%' => .value_ref,
+                '@' => .at_name,
+                '.' => .dot_name,
+                else => unreachable,
+            };
+            const content_start = self.cursor;
+
+            if (byte == '%' and self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor])) {
+                while (self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor])) self.cursor += 1;
+            } else {
+                while (self.cursor < self.source.len and isNameByte(self.source[self.cursor])) self.cursor += 1;
+            }
+
+            if (self.cursor == content_start)
+                return self.simpleToken(.invalid, start);
+
+            return .{
+                .tag = tag,
+                .text = self.source[content_start..self.cursor],
+            };
+        },
+        '#' => {
+            const number_start = self.cursor;
+            while (self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor]))
+                self.cursor += 1;
+
+            if (self.cursor == number_start)
+                return self.simpleToken(.invalid, start);
+
+            return .{
+                .tag = .constant_ref,
+                .text = self.source[number_start..self.cursor],
+            };
+        },
+        else => {},
+    }
+
+    if (std.ascii.isDigit(byte))
+        return self.numberToken(start);
+
+    if (isNameStart(byte)) {
+        while (self.cursor < self.source.len and isNameByte(self.source[self.cursor]))
+            self.cursor += 1;
+
+        return .{
+            .tag = .identifier,
+            .text = self.source[start..self.cursor],
+        };
+    }
+
+    return self.simpleToken(.invalid, start);
+}
+
+fn numberToken(self: *Self, start: usize) Token {
+    var number_start = start;
+    if (self.source[number_start] == '-' or self.source[number_start] == '+')
+        number_start += 1;
+
+    self.cursor = number_start;
+
+    if (self.source[number_start] == '0' and number_start + 1 < self.source.len and self.source[number_start + 1] == 'x') {
+        self.cursor = number_start + 2;
+        while (self.cursor < self.source.len and std.ascii.isHex(self.source[self.cursor]))
+            self.cursor += 1;
+    } else {
+        while (self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor]))
+            self.cursor += 1;
+
+        if (self.cursor < self.source.len and self.source[self.cursor] == '.') {
+            self.cursor += 1;
+            while (self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor]))
+                self.cursor += 1;
+        }
+
+        if (self.cursor < self.source.len and (self.source[self.cursor] == 'e' or self.source[self.cursor] == 'E')) {
+            self.cursor += 1;
+            if (self.cursor < self.source.len and (self.source[self.cursor] == '-' or self.source[self.cursor] == '+'))
+                self.cursor += 1;
+            while (self.cursor < self.source.len and std.ascii.isDigit(self.source[self.cursor]))
+                self.cursor += 1;
+        }
+    }
+
+    return .{
+        .tag = .number,
+        .text = self.source[start..self.cursor],
+    };
+}
+
+fn simpleToken(self: *Self, tag: TokenTag, start: usize) Token {
+    return .{
+        .tag = tag,
+        .text = self.source[start..self.cursor],
+    };
+}
+
+fn isNameStart(byte: u8) bool {
+    return std.ascii.isAlphabetic(byte) or byte == '_';
+}
+
+fn isNameByte(byte: u8) bool {
+    return std.ascii.isAlphanumeric(byte) or byte == '_';
+}

src/compiler/ir/parser/ast.zig

@@ -0,0 +1,93 @@
+const std = @import("std");
+const ids = @import("../id.zig");
+const inst_ir = @import("../instruction.zig");
+const module_ir = @import("../module.zig");
+
+pub const ValueRef = []const u8;
+
+pub const ParsedModule = struct {
+    entry_point_name: ?[]const u8,
+    interfaces: std.ArrayList(ParsedInterface) = .empty,
+    constants: std.ArrayList(ParsedConstant) = .empty,
+    functions: std.ArrayList(ParsedFunction) = .empty,
+};
+
+pub const ParsedInterface = struct {
+    direction: module_ir.InterfaceDirection,
+    name: []const u8,
+    ty: ids.TypeId,
+    semantic: module_ir.InterfaceSemantic,
+};
+
+pub const ParsedConstantValue = union(enum) {
+    boolean: bool,
+    integer_bits: u64,
+    float_bits: u64,
+    null_value,
+    undef,
+    composite: []const u32,
+};
+
+pub const ParsedConstant = struct {
+    printed_value: ValueRef,
+    ty: ids.TypeId,
+    value: ParsedConstantValue,
+};
+
+pub const ParsedParameter = struct {
+    printed_value: ValueRef,
+    ty: ids.TypeId,
+};
+
+pub const ParsedInstruction = struct {
+    printed_result: ?ValueRef,
+    result_type: ?ids.TypeId,
+    operation: ParsedOperation,
+};
+
+pub const ParsedBlock = struct {
+    name: []const u8,
+    parameters: std.ArrayList(ParsedParameter) = .empty,
+    instructions: std.ArrayList(ParsedInstruction) = .empty,
+    terminator: ?ParsedTerminator = null,
+    actual: ?ids.BlockId = null,
+};
+
+pub const ParsedFunction = struct {
+    name: []const u8,
+    return_type: ids.TypeId,
+    parameters: std.ArrayList(ParsedParameter) = .empty,
+    blocks: std.ArrayList(ParsedBlock) = .empty,
+    actual: ?ids.FunctionId = null,
+};
+
+pub const ParsedEdge = struct {
+    block_name: []const u8,
+    arguments: []const ValueRef,
+};
+
+pub const ParsedTerminator = union(enum) {
+    branch: ParsedEdge,
+    conditional_branch: struct {
+        condition: ValueRef,
+        true_edge: ParsedEdge,
+        false_edge: ParsedEdge,
+    },
+    return_void,
+    return_value: ValueRef,
+    discard,
+    unreachable_value,
+};
+
+pub const ParsedOperation = union(enum) {
+    unary: struct { opcode: inst_ir.UnaryOpcode, operand: ValueRef },
+    binary: struct { opcode: inst_ir.BinaryOpcode, lhs: ValueRef, rhs: ValueRef },
+    compare: struct { opcode: inst_ir.CompareOpcode, lhs: ValueRef, rhs: ValueRef },
+    select: struct { condition: ValueRef, true_value: ValueRef, false_value: ValueRef },
+    bitcast: ValueRef,
+    composite_construct: []const ValueRef,
+    composite_extract: struct { composite: ValueRef, indices: []const u32 },
+    load_interface: []const u8,
+    store_interface: struct { interface_name: []const u8, value: ValueRef },
+    call: struct { function_name: []const u8, arguments: []const ValueRef },
+};

src/compiler/ir/parser/lower.zig

@@ -0,0 +1,380 @@
+const std = @import("std");
+const Builder = @import("../Builder.zig");
+const constant_ir = @import("../constant.zig");
+const ids = @import("../id.zig");
+const inst_ir = @import("../instruction.zig");
+const module_ir = @import("../module.zig");
+const ast = @import("ast.zig");
+
+const ValueRef = ast.ValueRef;
+const ParsedModule = ast.ParsedModule;
+const ParsedOperation = ast.ParsedOperation;
+const ParsedTerminator = ast.ParsedTerminator;
+const ParsedEdge = ast.ParsedEdge;
+
+const LoweredOperation = struct {
+    operation: inst_ir.Operation,
+    inferred_type: ?ids.TypeId,
+};
+
+pub fn lower(allocator: std.mem.Allocator, module: *module_ir.Module, parsed: *ParsedModule) !void {
+    var builder = Builder.init(module);
+    var values: std.StringHashMapUnmanaged(ids.ValueId) = .empty;
+    var constants: std.AutoHashMapUnmanaged(u32, ids.ConstantId) = .empty;
+    var interfaces: std.StringHashMapUnmanaged(ids.InterfaceVariableId) = .empty;
+    var functions: std.StringHashMapUnmanaged(ids.FunctionId) = .empty;
+
+    for (parsed.interfaces.items) |interface| {
+        if (interfaces.contains(interface.name))
+            return error.DuplicateName;
+
+        const id = try builder.addInterfaceVariable(interface.ty, interface.direction, interface.semantic, interface.name);
+        try interfaces.put(allocator, interface.name, id);
+    }
+
+    for (parsed.constants.items, 0..) |constant, constant_index| {
+        const value: constant_ir.ConstantValue = switch (constant.value) {
+            .boolean => |item| .{ .boolean = item },
+            .integer_bits => |item| .{ .integer_bits = item },
+            .float_bits => |item| .{ .float_bits = item },
+            .null_value => .null,
+            .undef => .undef,
+            .composite => |printed_elements| blk: {
+                var elements: std.ArrayList(ids.ConstantId) = .empty;
+                for (printed_elements) |printed_element| {
+                    try elements.append(allocator, constants.get(printed_element) orelse return error.UnknownConstant);
+                }
+                break :blk .{ .composite = elements.items };
+            },
+        };
+
+        const value_id = try builder.internConstant(constant.ty, value);
+        try builder.setValueName(value_id, valueName(constant.printed_value));
+        try putValue(allocator, &values, constant.printed_value, value_id);
+        const definition = module.values.get(value_id).?.definition;
+
+        if (definition != .constant)
+            return error.InvalidResult;
+
+        try constants.put(allocator, @intCast(constant_index), definition.constant);
+    }
+
+    for (parsed.functions.items) |*function| {
+        if (functions.contains(function.name))
+            return error.DuplicateName;
+
+        const function_id = try builder.addFunction(function.return_type, function.name);
+        function.actual = function_id;
+        try functions.put(allocator, function.name, function_id);
+
+        for (function.parameters.items) |parameter| {
+            const value_id = try builder.addFunctionParameter(function_id, parameter.ty, valueName(parameter.printed_value));
+            try putValue(allocator, &values, parameter.printed_value, value_id);
+        }
+    }
+
+    if (parsed.entry_point_name) |entry_name|
+        builder.setEntryPoint(functions.get(entry_name) orelse return error.UnknownFunction);
+
+    for (parsed.functions.items) |*function| {
+        var block_names: std.StringHashMapUnmanaged(ids.BlockId) = .empty;
+        for (function.blocks.items) |*block| {
+            if (block_names.contains(block.name))
+                return error.DuplicateName;
+
+            const block_id = try builder.addBlock(function.actual.?, block.name);
+            block.actual = block_id;
+            try block_names.put(allocator, block.name, block_id);
+
+            for (block.parameters.items) |parameter| {
+                const value_id = try builder.addBlockParameter(block_id, parameter.ty, valueName(parameter.printed_value));
+                try putValue(allocator, &values, parameter.printed_value, value_id);
+            }
+        }
+
+        for (function.blocks.items) |block| {
+            for (block.instructions.items) |instruction| {
+                const lowered = try lowerOperation(allocator, module, &values, &interfaces, &functions, instruction.operation);
+                const result_type = instruction.result_type orelse lowered.inferred_type;
+
+                if (instruction.printed_result != null and result_type == null)
+                    return error.MissingResultType;
+
+                if (instruction.printed_result == null and result_type != null)
+                    return error.InvalidResult;
+
+                const result = try builder.appendInstruction(
+                    block.actual.?,
+                    result_type,
+                    lowered.operation,
+                    if (instruction.printed_result) |printed_result| valueName(printed_result) else null,
+                );
+                if (instruction.printed_result) |printed_result|
+                    try putValue(allocator, &values, printed_result, result orelse return error.InvalidResult)
+                else if (result != null)
+                    return error.InvalidResult;
+            }
+            const terminator = try lowerTerminator(allocator, &builder, &values, &block_names, block.terminator orelse return error.MissingTerminator);
+            try builder.setTerminator(block.actual.?, terminator);
+        }
+    }
+}
+
+fn lowerOperation(
+    allocator: std.mem.Allocator,
+    module: *module_ir.Module,
+    values: *const std.StringHashMapUnmanaged(ids.ValueId),
+    interfaces: *const std.StringHashMapUnmanaged(ids.InterfaceVariableId),
+    functions: *const std.StringHashMapUnmanaged(ids.FunctionId),
+    parsed: ParsedOperation,
+) !LoweredOperation {
+    return switch (parsed) {
+        .unary => |op| blk: {
+            const operand = resolveValue(values, op.operand) orelse return error.UnknownValue;
+
+            break :blk .{
+                .operation = .{
+                    .unary = .{
+                        .opcode = op.opcode,
+                        .operand = operand,
+                    },
+                },
+                .inferred_type = module.typeOf(operand),
+            };
+        },
+        .binary => |op| blk: {
+            const lhs = resolveValue(values, op.lhs) orelse return error.UnknownValue;
+            const rhs = resolveValue(values, op.rhs) orelse return error.UnknownValue;
+
+            break :blk .{
+                .operation = .{
+                    .binary = .{
+                        .opcode = op.opcode,
+                        .lhs = lhs,
+                        .rhs = rhs,
+                    },
+                },
+                .inferred_type = module.typeOf(lhs),
+            };
+        },
+        .compare => |op| blk: {
+            const lhs = resolveValue(values, op.lhs) orelse return error.UnknownValue;
+            const rhs = resolveValue(values, op.rhs) orelse return error.UnknownValue;
+
+            break :blk .{
+                .operation = .{
+                    .compare = .{
+                        .opcode = op.opcode,
+                        .lhs = lhs,
+                        .rhs = rhs,
+                    },
+                },
+                .inferred_type = try module.internType(.boolean),
+            };
+        },
+        .select => |op| blk: {
+            const condition = resolveValue(values, op.condition) orelse return error.UnknownValue;
+            const true_value = resolveValue(values, op.true_value) orelse return error.UnknownValue;
+            const false_value = resolveValue(values, op.false_value) orelse return error.UnknownValue;
+
+            break :blk .{
+                .operation = .{
+                    .select = .{
+                        .condition = condition,
+                        .true_value = true_value,
+                        .false_value = false_value,
+                    },
+                },
+                .inferred_type = module.typeOf(true_value),
+            };
+        },
+        .bitcast => |printed_operand| blk: {
+            const operand = resolveValue(values, printed_operand) orelse return error.UnknownValue;
+            break :blk .{ .operation = .{ .bitcast = operand }, .inferred_type = module.typeOf(operand) };
+        },
+        .composite_construct => |printed_elements| blk: {
+            var elements: std.ArrayList(ids.ValueId) = .empty;
+            var element_types: std.ArrayList(ids.TypeId) = .empty;
+            for (printed_elements) |printed_element| {
+                const element = resolveValue(values, printed_element) orelse return error.UnknownValue;
+                try elements.append(allocator, element);
+                try element_types.append(allocator, module.typeOf(element) orelse return error.UnknownValue);
+            }
+
+            break :blk .{
+                .operation = .{
+                    .composite_construct = .{
+                        .elements = elements.items,
+                    },
+                },
+                .inferred_type = try inferCompositeType(module, element_types.items),
+            };
+        },
+        .composite_extract => |op| blk: {
+            const composite = resolveValue(values, op.composite) orelse return error.UnknownValue;
+
+            break :blk .{
+                .operation = .{
+                    .composite_extract = .{
+                        .composite = composite,
+                        .indices = op.indices,
+                    },
+                },
+                .inferred_type = try extractedType(module, module.typeOf(composite) orelse return error.UnknownValue, op.indices),
+            };
+        },
+        .load_interface => |name| blk: {
+            const interface_id = interfaces.get(name) orelse return error.UnknownInterface;
+
+            break :blk .{
+                .operation = .{
+                    .load_interface = .{
+                        .variable = interface_id,
+                    },
+                },
+                .inferred_type = module.interface_variables.get(interface_id).?.type,
+            };
+        },
+        .store_interface => |op| blk: {
+            const interface_id = interfaces.get(op.interface_name) orelse return error.UnknownInterface;
+            const value = resolveValue(values, op.value) orelse return error.UnknownValue;
+
+            break :blk .{
+                .operation = .{
+                    .store_interface = .{
+                        .variable = interface_id,
+                        .value = value,
+                    },
+                },
+                .inferred_type = null,
+            };
+        },
+        .call => |op| blk: {
+            const function_id = functions.get(op.function_name) orelse return error.UnknownFunction;
+            var arguments: std.ArrayList(ids.ValueId) = .empty;
+
+            for (op.arguments) |printed_argument|
+                try arguments.append(allocator, resolveValue(values, printed_argument) orelse return error.UnknownValue);
+
+            const return_type = module.functions.get(function_id).?.return_type;
+            const return_ir_type = module.types.get(return_type) orelse return error.InvalidType;
+
+            break :blk .{
+                .operation = .{
+                    .call = .{
+                        .function = function_id,
+                        .arguments = arguments.items,
+                    },
+                },
+                .inferred_type = if (return_ir_type.* == .void) null else return_type,
+            };
+        },
+    };
+}
+
+fn lowerTerminator(
+    allocator: std.mem.Allocator,
+    builder: *Builder,
+    values: *const std.StringHashMapUnmanaged(ids.ValueId),
+    blocks: *const std.StringHashMapUnmanaged(ids.BlockId),
+    parsed: ParsedTerminator,
+) !module_ir.Terminator {
+    return switch (parsed) {
+        .branch => |edge| .{
+            .branch = try lowerEdge(allocator, builder, values, blocks, edge),
+        },
+        .conditional_branch => |branch| .{
+            .conditional_branch = .{
+                .condition = resolveValue(values, branch.condition) orelse return error.UnknownValue,
+                .true_edge = try lowerEdge(allocator, builder, values, blocks, branch.true_edge),
+                .false_edge = try lowerEdge(allocator, builder, values, blocks, branch.false_edge),
+            },
+        },
+        .return_void => .return_void,
+        .return_value => |printed_value| .{
+            .return_value = resolveValue(values, printed_value) orelse return error.UnknownValue,
+        },
+        .discard => .discard,
+        .unreachable_value => .@"unreachable",
+    };
+}
+
+fn lowerEdge(
+    allocator: std.mem.Allocator,
+    builder: *Builder,
+    values: *const std.StringHashMapUnmanaged(ids.ValueId),
+    blocks: *const std.StringHashMapUnmanaged(ids.BlockId),
+    parsed: ParsedEdge,
+) !module_ir.Edge {
+    var arguments: std.ArrayList(ids.ValueId) = .empty;
+    for (parsed.arguments) |printed_argument|
+        try arguments.append(allocator, resolveValue(values, printed_argument) orelse return error.UnknownValue);
+
+    return builder.edge(blocks.get(parsed.block_name) orelse return error.UnknownBlock, arguments.items);
+}
+
+fn inferCompositeType(module: *module_ir.Module, element_types: []const ids.TypeId) !ids.TypeId {
+    if (element_types.len >= 2 and element_types.len <= std.math.maxInt(u8)) {
+        const first = element_types[0];
+        for (element_types[1..]) |element_type| {
+            if (element_type != first)
+                return module.internType(.{
+                    .structure = .{
+                        .members = element_types,
+                    },
+                });
+        }
+        return module.internType(.{
+            .vector = .{
+                .element_type = first,
+                .length = @intCast(element_types.len),
+            },
+        });
+    }
+    return module.internType(.{
+        .structure = .{
+            .members = element_types,
+        },
+    });
+}
+
+fn extractedType(module: *const module_ir.Module, root_type: ids.TypeId, indices: []const u32) !ids.TypeId {
+    var current = root_type;
+    for (indices) |index| {
+        const ty = module.types.get(current) orelse return error.InvalidType;
+        current = switch (ty.*) {
+            .vector => |vector| if (index < vector.length)
+                vector.element_type
+            else
+                return error.InvalidCompositeIndex,
+            .array => |array| if (index < array.length)
+                array.element_type
+            else
+                return error.InvalidCompositeIndex,
+            .structure => |structure| if (index < structure.members.len)
+                structure.members[index]
+            else
+                return error.InvalidCompositeIndex,
+            else => return error.InvalidCompositeIndex,
+        };
+    }
+    return current;
+}
+
+fn putValue(allocator: std.mem.Allocator, values: *std.StringHashMapUnmanaged(ids.ValueId), printed: ValueRef, actual: ids.ValueId) !void {
+    if (values.contains(printed))
+        return error.DuplicateValue;
+    try values.put(allocator, printed, actual);
+}
+
+fn resolveValue(values: *const std.StringHashMapUnmanaged(ids.ValueId), printed: ValueRef) ?ids.ValueId {
+    return values.get(printed);
+}
+
+fn valueName(reference: ValueRef) ?[]const u8 {
+    for (reference) |byte| {
+        if (!std.ascii.isDigit(byte))
+            return reference;
+    }
+    return null;
+}

src/compiler/ir/parser/parser.zig

@@ -0,0 +1,767 @@
+const std = @import("std");
+const ids = @import("../id.zig");
+const inst_ir = @import("../instruction.zig");
+const module_ir = @import("../module.zig");
+const type_ir = @import("../type.zig");
+const validator = @import("../validator/validator.zig");
+const Lexer = @import("Lexer.zig");
+
+const ast = @import("ast.zig");
+const lowerer = @import("lower.zig");
+
+pub const Error = error{
+    UnexpectedToken,
+    InvalidNumber,
+    InvalidStage,
+    InvalidType,
+    InvalidOpcode,
+    InvalidSemantic,
+    DuplicateValue,
+    DuplicateName,
+    UnknownValue,
+    UnknownConstant,
+    UnknownInterface,
+    UnknownFunction,
+    UnknownBlock,
+    MissingTerminator,
+    MissingResultType,
+    InvalidResult,
+    InvalidCompositeIndex,
+};
+
+pub const max_file_size = 64 * 1024 * 1024;
+
+const ValueRef = ast.ValueRef;
+const ParsedModule = ast.ParsedModule;
+const ParsedInterface = ast.ParsedInterface;
+const ParsedConstantValue = ast.ParsedConstantValue;
+const ParsedConstant = ast.ParsedConstant;
+const ParsedParameter = ast.ParsedParameter;
+const ParsedInstruction = ast.ParsedInstruction;
+const ParsedBlock = ast.ParsedBlock;
+const ParsedFunction = ast.ParsedFunction;
+const ParsedEdge = ast.ParsedEdge;
+const ParsedTerminator = ast.ParsedTerminator;
+const ParsedOperation = ast.ParsedOperation;
+const Token = Lexer.Token;
+const TokenTag = Lexer.TokenTag;
+
+const Parser = struct {
+    lexer: Lexer,
+    allocator: std.mem.Allocator,
+    module: ?*module_ir.Module = null,
+
+    fn parseInterface(self: *Parser) !ParsedInterface {
+        const name = (try self.expect(.at_name)).text;
+
+        try self.expectDiscard(.colon);
+        const ty = try self.parseType();
+        try self.expectDiscard(.equal);
+
+        const direction_token = try self.expect(.identifier);
+        const direction = std.meta.stringToEnum(module_ir.InterfaceDirection, direction_token.text) orelse return error.InvalidSemantic;
+        try self.expectDiscard(.left_square);
+
+        const semantic_name = (try self.expect(.identifier)).text;
+        const semantic: module_ir.InterfaceSemantic = if (std.mem.eql(u8, semantic_name, "location")) blk: {
+            try self.expectDiscard(.left_paren);
+            const location = try self.parseUnsigned(u32, .number);
+            try self.expectDiscard(.right_paren);
+
+            try self.expectDiscard(.comma);
+            try self.expectIdentifier("component");
+            try self.expectDiscard(.left_paren);
+            const component = try self.parseUnsigned(u8, .number);
+            try self.expectDiscard(.right_paren);
+
+            try self.expectDiscard(.comma);
+            try self.expectIdentifier("index");
+            try self.expectDiscard(.left_paren);
+            const index = try self.parseUnsigned(u8, .number);
+            try self.expectDiscard(.right_paren);
+
+            break :blk .{
+                .location = .{
+                    .location = location,
+                    .component = component,
+                    .index = index,
+                },
+            };
+        } else if (std.mem.eql(u8, semantic_name, "builtin")) blk: {
+            try self.expectDiscard(.left_paren);
+            const builtin_name = (try self.expect(.identifier)).text;
+            try self.expectDiscard(.right_paren);
+
+            const builtin = std.meta.stringToEnum(module_ir.Builtin, builtin_name) orelse return error.InvalidSemantic;
+            break :blk .{ .builtin = builtin };
+        } else return error.InvalidSemantic;
+
+        try self.expectDiscard(.right_square);
+
+        return .{
+            .direction = direction,
+            .name = name,
+            .ty = ty,
+            .semantic = semantic,
+        };
+    }
+
+    fn parseConstant(self: *Parser) !ParsedConstant {
+        const printed_value = try self.parseValueRef();
+
+        try self.expectDiscard(.colon);
+        try self.expectIdentifier("constant");
+
+        const ty = try self.parseType();
+        try self.expectDiscard(.equal);
+
+        const token = try self.peek();
+
+        const value: ParsedConstantValue = switch (token.tag) {
+            .identifier => blk: {
+                const word = (try self.take()).text;
+                if (std.mem.eql(u8, word, "true"))
+                    break :blk .{ .boolean = true };
+
+                if (std.mem.eql(u8, word, "false"))
+                    break :blk .{ .boolean = false };
+
+                if (std.mem.eql(u8, word, "null"))
+                    break :blk .null_value;
+
+                if (std.mem.eql(u8, word, "undef"))
+                    break :blk .undef;
+
+                if (std.mem.eql(u8, word, "bits")) {
+                    try self.expectDiscard(.left_paren);
+                    const bits = try self.parseUnsigned(u64, .number);
+                    try self.expectDiscard(.right_paren);
+
+                    const ir_type = self.module.?.types.get(ty) orelse return error.InvalidType;
+
+                    break :blk switch (ir_type.*) {
+                        .integer => .{
+                            .integer_bits = bits,
+                        },
+                        .floating => .{
+                            .float_bits = bits,
+                        },
+                        else => return error.InvalidType,
+                    };
+                }
+                return error.UnexpectedToken;
+            },
+            .left_square => .{
+                .composite = try self.parseConstantList(),
+            },
+            .number => try self.parseDirectConstant(ty),
+            else => return error.UnexpectedToken,
+        };
+
+        return .{
+            .printed_value = printed_value,
+            .ty = ty,
+            .value = value,
+        };
+    }
+
+    fn parseDirectConstant(self: *Parser, ty: ids.TypeId) !ParsedConstantValue {
+        const text = (try self.expect(.number)).text;
+        const ir_type = self.module.?.types.get(ty) orelse return error.InvalidType;
+
+        return switch (ir_type.*) {
+            .integer => |integer| .{ .integer_bits = try parseIntegerLiteral(integer, text) },
+            .floating => |float| .{ .float_bits = try parseFloatLiteral(float.bits, text) },
+            else => error.InvalidType,
+        };
+    }
+
+    fn parseFunction(self: *Parser) !ParsedFunction {
+        try self.expectIdentifier("fn");
+        const name = (try self.expect(.at_name)).text;
+        try self.expectDiscard(.left_paren);
+
+        var parameters: std.ArrayList(ParsedParameter) = .empty;
+        if ((try self.peek()).tag != .right_paren) {
+            while (true) {
+                const printed_value = try self.parseValueRef();
+                try self.expectDiscard(.colon);
+
+                const ty = try self.parseType();
+                try parameters.append(self.allocator, .{ .printed_value = printed_value, .ty = ty });
+
+                if (!try self.consume(.comma))
+                    break;
+            }
+        }
+
+        try self.expectDiscard(.right_paren);
+        try self.expectDiscard(.arrow);
+
+        const return_type = try self.parseType();
+
+        var function: ParsedFunction = .{
+            .name = name,
+            .return_type = return_type,
+            .parameters = parameters,
+        };
+
+        try self.expectDiscard(.left_brace);
+
+        while ((try self.peek()).tag != .right_brace) {
+            if ((try self.peek()).tag != .dot_name)
+                return error.UnexpectedToken;
+            try function.blocks.append(self.allocator, try self.parseBlock());
+        }
+
+        try self.expectDiscard(.right_brace);
+
+        return function;
+    }
+
+    fn parseBlock(self: *Parser) !ParsedBlock {
+        var block: ParsedBlock = .{
+            .name = (try self.expect(.dot_name)).text,
+        };
+
+        try self.expectDiscard(.left_paren);
+
+        if ((try self.peek()).tag != .right_paren) {
+            while (true) {
+                const printed_value = try self.parseValueRef();
+
+                try self.expectDiscard(.colon);
+                const ty = try self.parseType();
+                try block.parameters.append(self.allocator, .{ .printed_value = printed_value, .ty = ty });
+
+                if (!try self.consume(.comma))
+                    break;
+            }
+        }
+
+        try self.expectDiscard(.right_paren);
+        try self.expectDiscard(.colon);
+
+        while (block.terminator == null) {
+            const token = try self.peek();
+
+            if (token.tag == .right_brace or token.tag == .dot_name)
+                return error.MissingTerminator;
+
+            if (token.tag == .value_ref) {
+                try block.instructions.append(self.allocator, try self.parseInstruction(true));
+                continue;
+            }
+
+            if (token.tag != .identifier)
+                return error.UnexpectedToken;
+
+            if (isTerminatorName(token.text)) {
+                block.terminator = try self.parseTerminator();
+            } else {
+                try block.instructions.append(self.allocator, try self.parseInstruction(false));
+            }
+        }
+
+        return block;
+    }
+
+    fn parseInstruction(self: *Parser, has_result: bool) !ParsedInstruction {
+        const printed_result = if (has_result) try self.parseValueRef() else null;
+        const result_type = if (has_result and try self.consume(.colon)) try self.parseType() else null;
+
+        if (has_result)
+            try self.expectDiscard(.equal);
+
+        return .{
+            .printed_result = printed_result,
+            .result_type = result_type,
+            .operation = try self.parseOperation(),
+        };
+    }
+
+    fn parseOperation(self: *Parser) !ParsedOperation {
+        const name = (try self.expect(.identifier)).text;
+
+        if (std.mem.startsWith(u8, name, "cmp_")) {
+            const opcode_name = name["cmp_".len..];
+            const opcode = std.meta.stringToEnum(inst_ir.CompareOpcode, opcode_name) orelse return error.InvalidOpcode;
+            const lhs = try self.parseValueRef();
+
+            try self.expectDiscard(.comma);
+
+            return .{
+                .compare = .{
+                    .opcode = opcode,
+                    .lhs = lhs,
+                    .rhs = try self.parseValueRef(),
+                },
+            };
+        }
+
+        if (std.meta.stringToEnum(inst_ir.UnaryOpcode, name)) |opcode| {
+            return .{ .unary = .{ .opcode = opcode, .operand = try self.parseValueRef() } };
+        }
+
+        if (std.meta.stringToEnum(inst_ir.BinaryOpcode, name)) |opcode| {
+            const lhs = try self.parseValueRef();
+            try self.expectDiscard(.comma);
+            return .{ .binary = .{ .opcode = opcode, .lhs = lhs, .rhs = try self.parseValueRef() } };
+        }
+
+        if (std.mem.eql(u8, name, "select")) {
+            const condition = try self.parseValueRef();
+            try self.expectDiscard(.comma);
+            const true_value = try self.parseValueRef();
+            try self.expectDiscard(.comma);
+            return .{ .select = .{
+                .condition = condition,
+                .true_value = true_value,
+                .false_value = try self.parseValueRef(),
+            } };
+        }
+
+        if (std.mem.eql(u8, name, "bitcast"))
+            return .{ .bitcast = try self.parseValueRef() };
+
+        if (std.mem.eql(u8, name, "composite_construct"))
+            return .{ .composite_construct = try self.parseTrailingValueList() };
+
+        if (std.mem.eql(u8, name, "composite_extract")) {
+            const composite = try self.parseValueRef();
+            var indices: std.ArrayList(u32) = .empty;
+
+            while (try self.consume(.left_square)) {
+                try indices.append(self.allocator, try self.parseUnsigned(u32, .number));
+                try self.expectDiscard(.right_square);
+            }
+
+            if (indices.items.len == 0)
+                return error.InvalidCompositeIndex;
+            return .{
+                .composite_extract = .{
+                    .composite = composite,
+                    .indices = indices.items,
+                },
+            };
+        }
+
+        if (std.mem.eql(u8, name, "load_interface"))
+            return .{
+                .load_interface = (try self.expect(.at_name)).text,
+            };
+
+        if (std.mem.eql(u8, name, "store_interface")) {
+            const interface_name = (try self.expect(.at_name)).text;
+            try self.expectDiscard(.comma);
+            return .{
+                .store_interface = .{
+                    .interface_name = interface_name,
+                    .value = try self.parseValueRef(),
+                },
+            };
+        }
+
+        if (std.mem.eql(u8, name, "call")) {
+            const function_name = (try self.expect(.at_name)).text;
+            try self.expectDiscard(.left_paren);
+
+            const arguments = try self.parseDelimitedValueList(.right_paren);
+            try self.expectDiscard(.right_paren);
+
+            return .{
+                .call = .{
+                    .function_name = function_name,
+                    .arguments = arguments,
+                },
+            };
+        }
+
+        return error.InvalidOpcode;
+    }
+
+    fn parseTerminator(self: *Parser) !ParsedTerminator {
+        const name = (try self.expect(.identifier)).text;
+
+        if (std.mem.eql(u8, name, "branch"))
+            return .{ .branch = try self.parseEdge() };
+
+        if (std.mem.eql(u8, name, "conditional_branch")) {
+            const condition = try self.parseValueRef();
+            try self.expectDiscard(.comma);
+
+            const true_edge = try self.parseEdge();
+            try self.expectDiscard(.comma);
+
+            return .{
+                .conditional_branch = .{
+                    .condition = condition,
+                    .true_edge = true_edge,
+                    .false_edge = try self.parseEdge(),
+                },
+            };
+        }
+
+        if (std.mem.eql(u8, name, "return")) {
+            if ((try self.peek()).tag == .value_ref)
+                return .{ .return_value = try self.parseValueRef() };
+            return .return_void;
+        }
+
+        if (std.mem.eql(u8, name, "discard"))
+            return .discard;
+
+        if (std.mem.eql(u8, name, "unreachable"))
+            return .unreachable_value;
+
+        return error.UnexpectedToken;
+    }
+
+    fn parseEdge(self: *Parser) !ParsedEdge {
+        const block_name = (try self.expect(.dot_name)).text;
+        try self.expectDiscard(.left_paren);
+
+        const arguments = try self.parseDelimitedValueList(.right_paren);
+        try self.expectDiscard(.right_paren);
+
+        return .{
+            .block_name = block_name,
+            .arguments = arguments,
+        };
+    }
+
+    fn parseType(self: *Parser) !ids.TypeId {
+        const token = try self.expect(.identifier);
+        const module = self.module.?;
+
+        if (std.mem.eql(u8, token.text, "void"))
+            return module.internType(.void);
+
+        if (std.mem.eql(u8, token.text, "bool"))
+            return module.internType(.boolean);
+
+        if (std.mem.startsWith(u8, token.text, "vec")) {
+            const length = parseTextUnsigned(u8, token.text[3..]) catch return error.InvalidType;
+            try self.expectDiscard(.left_square);
+
+            const element_type = try self.parseType();
+            try self.expectDiscard(.right_square);
+
+            return module.internType(.{
+                .vector = .{
+                    .element_type = element_type,
+                    .length = length,
+                },
+            });
+        }
+
+        if (std.mem.eql(u8, token.text, "array")) {
+            try self.expectDiscard(.left_square);
+
+            const element_type = try self.parseType();
+            try self.expectDiscard(.comma);
+
+            const length = try self.parseUnsigned(u32, .number);
+            try self.expectDiscard(.right_square);
+
+            return module.internType(.{
+                .array = .{
+                    .element_type = element_type,
+                    .length = length,
+                },
+            });
+        }
+
+        if (std.mem.eql(u8, token.text, "struct")) {
+            try self.expectDiscard(.left_square);
+            var members: std.ArrayList(ids.TypeId) = .empty;
+
+            if ((try self.peek()).tag != .right_square) {
+                while (true) {
+                    try members.append(self.allocator, try self.parseType());
+                    if (!try self.consume(.comma))
+                        break;
+                }
+            }
+
+            try self.expectDiscard(.right_square);
+
+            return module.internType(.{
+                .structure = .{
+                    .members = members.items,
+                },
+            });
+        }
+
+        if (std.mem.eql(u8, token.text, "ptr")) {
+            try self.expectDiscard(.left_square);
+
+            const address_name = (try self.expect(.identifier)).text;
+            const address_space = std.meta.stringToEnum(type_ir.AddressSpace, address_name) orelse return error.InvalidType;
+            try self.expectDiscard(.comma);
+
+            const pointee_type = try self.parseType();
+            try self.expectDiscard(.right_square);
+
+            return module.internType(.{
+                .pointer = .{
+                    .address_space = address_space,
+                    .pointee_type = pointee_type,
+                },
+            });
+        }
+
+        if (std.mem.eql(u8, token.text, "resourceHandle")) {
+            try self.expectDiscard(.left_square);
+
+            const kind_name = (try self.expect(.identifier)).text;
+            const kind = std.meta.stringToEnum(type_ir.ResourceKind, kind_name) orelse return error.InvalidType;
+
+            try self.expectDiscard(.right_square);
+
+            return module.internType(.{
+                .resource_handle = .{
+                    .kind = kind,
+                },
+            });
+        }
+
+        if (token.text.len > 1 and (token.text[0] == 'i' or token.text[0] == 'u')) {
+            const bits = parseTextUnsigned(u16, token.text[1..]) catch return error.InvalidType;
+            return module.internType(.{ .integer = .{
+                .bits = bits,
+                .signedness = if (token.text[0] == 'i') .signed else .unsigned,
+            } });
+        }
+
+        if (token.text.len > 1 and token.text[0] == 'f') {
+            const bits = parseTextUnsigned(u16, token.text[1..]) catch return error.InvalidType;
+            return module.internType(.{
+                .floating = .{
+                    .bits = bits,
+                },
+            });
+        }
+        return error.InvalidType;
+    }
+
+    fn parseConstantList(self: *Parser) ![]const u32 {
+        try self.expectDiscard(.left_square);
+        var values: std.ArrayList(u32) = .empty;
+
+        if ((try self.peek()).tag != .right_square) {
+            while (true) {
+                try values.append(self.allocator, try self.parseUnsigned(u32, .constant_ref));
+                if (!try self.consume(.comma))
+                    break;
+            }
+        }
+
+        try self.expectDiscard(.right_square);
+        return values.items;
+    }
+
+    fn parseTrailingValueList(self: *Parser) ![]const ValueRef {
+        var values: std.ArrayList(ValueRef) = .empty;
+        if ((try self.peek()).tag != .value_ref)
+            return values.items;
+
+        while (true) {
+            try values.append(self.allocator, try self.parseValueRef());
+            if (!try self.consume(.comma))
+                break;
+        }
+
+        return values.items;
+    }
+
+    fn parseDelimitedValueList(self: *Parser, closing: TokenTag) ![]const ValueRef {
+        var values: std.ArrayList(ValueRef) = .empty;
+        if ((try self.peek()).tag == closing)
+            return values.items;
+
+        while (true) {
+            try values.append(self.allocator, try self.parseValueRef());
+            if (!try self.consume(.comma))
+                break;
+        }
+
+        return values.items;
+    }
+
+    fn parseValueRef(self: *Parser) !ValueRef {
+        return (try self.expect(.value_ref)).text;
+    }
+
+    fn parseUnsigned(self: *Parser, comptime T: type, tag: TokenTag) !T {
+        const token = try self.expect(tag);
+        return parseTextUnsigned(T, token.text) catch error.InvalidNumber;
+    }
+
+    fn expectIdentifier(self: *Parser, expected: []const u8) !void {
+        const token = try self.expect(.identifier);
+        if (!std.mem.eql(u8, token.text, expected))
+            return error.UnexpectedToken;
+    }
+
+    fn expectDiscard(self: *Parser, tag: TokenTag) !void {
+        _ = try self.expect(tag);
+    }
+
+    fn expect(self: *Parser, tag: TokenTag) !Token {
+        const token = try self.take();
+        if (token.tag != tag)
+            return error.UnexpectedToken;
+        return token;
+    }
+
+    fn consume(self: *Parser, tag: TokenTag) !bool {
+        if ((try self.peek()).tag != tag)
+            return false;
+        _ = try self.take();
+        return true;
+    }
+
+    fn peek(self: *Parser) !Token {
+        return self.lexer.peek();
+    }
+
+    fn take(self: *Parser) !Token {
+        const token = self.lexer.take();
+        if (token.tag == .invalid)
+            return error.UnexpectedToken;
+        return token;
+    }
+};
+
+fn isTerminatorName(name: []const u8) bool {
+    return std.mem.eql(u8, name, "branch") or
+        std.mem.eql(u8, name, "conditional_branch") or
+        std.mem.eql(u8, name, "return") or
+        std.mem.eql(u8, name, "discard") or
+        std.mem.eql(u8, name, "unreachable");
+}
+
+fn parseIntegerLiteral(integer: type_ir.IntegerType, text: []const u8) !u64 {
+    if (integer.bits == 0 or integer.bits > 64)
+        return error.InvalidType;
+
+    if (integer.signedness == .unsigned) {
+        const value = std.fmt.parseInt(u64, text, 10) catch return error.InvalidNumber;
+        if (integer.bits < 64) {
+            const shift: u6 = @intCast(integer.bits);
+            const maximum = (@as(u64, 1) << shift) - 1;
+            if (value > maximum)
+                return error.InvalidNumber;
+        }
+        return value;
+    }
+
+    const value = std.fmt.parseInt(i64, text, 10) catch return error.InvalidNumber;
+    if (integer.bits < 64) {
+        const sign_shift: u6 = @intCast(integer.bits - 1);
+        const magnitude = @as(i64, 1) << sign_shift;
+        if (value < -magnitude or value > magnitude - 1)
+            return error.InvalidNumber;
+
+        const width: u6 = @intCast(integer.bits);
+        const mask = (@as(u64, 1) << width) - 1;
+        return @as(u64, @bitCast(value)) & mask;
+    }
+
+    return @bitCast(value);
+}
+
+fn parseFloatLiteral(bits: u16, text: []const u8) !u64 {
+    return switch (bits) {
+        16 => blk: {
+            const value = std.fmt.parseFloat(f16, text) catch return error.InvalidNumber;
+            break :blk @as(u16, @bitCast(value));
+        },
+        32 => blk: {
+            const value = std.fmt.parseFloat(f32, text) catch return error.InvalidNumber;
+            break :blk @as(u32, @bitCast(value));
+        },
+        64 => blk: {
+            const value = std.fmt.parseFloat(f64, text) catch return error.InvalidNumber;
+            break :blk @as(u64, @bitCast(value));
+        },
+        else => error.InvalidType,
+    };
+}
+
+fn parseTextUnsigned(comptime T: type, text: []const u8) !T {
+    const base: u8 = if (std.mem.startsWith(u8, text, "0x")) 16 else 10;
+    const digits = if (base == 16) text[2..] else text;
+
+    if (digits.len == 0)
+        return error.InvalidNumber;
+
+    return std.fmt.parseInt(T, digits, base);
+}
+
+pub fn parseString(backing_allocator: std.mem.Allocator, source: []const u8) !module_ir.Module {
+    var temporary = std.heap.ArenaAllocator.init(backing_allocator);
+    defer temporary.deinit();
+    const temporary_allocator = temporary.allocator();
+
+    var parser: Parser = .{
+        .lexer = .init(source),
+        .allocator = temporary_allocator,
+    };
+
+    try parser.expectIdentifier("shader");
+    const stage_token = try parser.expect(.identifier);
+    const stage = std.meta.stringToEnum(module_ir.Stage, stage_token.text) orelse return error.InvalidStage;
+
+    var module = module_ir.Module.init(backing_allocator, stage);
+    errdefer module.deinit();
+    parser.module = &module;
+
+    const entry_point_name = if ((try parser.peek()).tag == .at_name)
+        (try parser.take()).text
+    else
+        null;
+
+    try parser.expectDiscard(.left_brace);
+
+    var parsed: ParsedModule = .{ .entry_point_name = entry_point_name };
+    while ((try parser.peek()).tag != .right_brace) {
+        const token = try parser.peek();
+        switch (token.tag) {
+            .value_ref => try parsed.constants.append(temporary_allocator, try parser.parseConstant()),
+            .at_name => try parsed.interfaces.append(temporary_allocator, try parser.parseInterface()),
+            .identifier => {
+                if (std.mem.eql(u8, token.text, "fn")) {
+                    try parsed.functions.append(temporary_allocator, try parser.parseFunction());
+                } else {
+                    return error.UnexpectedToken;
+                }
+            },
+            else => return error.UnexpectedToken,
+        }
+    }
+    try parser.expectDiscard(.right_brace);
+    try parser.expectDiscard(.eof);
+
+    try lowerer.lower(temporary_allocator, &module, &parsed);
+    try validator.validate(&module);
+    return module;
+}
+
+pub fn parseFile(backing_allocator: std.mem.Allocator, io: std.Io, path: []const u8) !module_ir.Module {
+    return parseFileInDir(backing_allocator, io, std.Io.Dir.cwd(), path);
+}
+
+pub fn parseFileInDir(backing_allocator: std.mem.Allocator, io: std.Io, directory: std.Io.Dir, path: []const u8) !module_ir.Module {
+    const file = try directory.openFile(io, path, .{});
+    defer file.close(io);
+
+    var buffer: [4096]u8 = @splat(0);
+    var reader = file.reader(io, &buffer);
+
+    const source = try reader.interface.allocRemaining(backing_allocator, .limited(max_file_size));
+    defer backing_allocator.free(source);
+
+    return parseString(backing_allocator, source);
+}

src/compiler/ir/pass_manager.zig

@@ -0,0 +1,83 @@
+const std = @import("std");
+const module_ir = @import("module.zig");
+const validator = @import("validator/validator.zig");
+
+pub const Context = struct {
+    allocator: std.mem.Allocator,
+    validate_after_each_pass: bool = true,
+};
+
+pub const Pass = struct {
+    name: []const u8,
+    required: module_ir.Properties = .{},
+    produced: module_ir.Properties = .{},
+    invalidated: module_ir.Properties = .{},
+    run: *const fn (module: *module_ir.Module, context: *Context) anyerror!bool,
+};
+
+pub const Manager = struct {
+    allocator: std.mem.Allocator,
+    passes: std.ArrayList(Pass) = .empty,
+
+    pub fn init(allocator: std.mem.Allocator) Manager {
+        return .{ .allocator = allocator };
+    }
+
+    pub fn deinit(self: *Manager) void {
+        self.passes.deinit(self.allocator);
+        self.* = undefined;
+    }
+
+    pub fn add(self: *Manager, pass: Pass) !void {
+        try self.passes.append(self.allocator, pass);
+    }
+
+    pub fn run(self: *Manager, module: *module_ir.Module, context: *Context) !bool {
+        var changed = false;
+        for (self.passes.items) |pass| {
+            if (!satisfies(module.properties, pass.required))
+                return error.RequiredPropertyMissing;
+
+            changed = (try pass.run(module, context)) or changed;
+            applyInvalidated(&module.properties, pass.invalidated);
+            applyProduced(&module.properties, pass.produced);
+
+            if (context.validate_after_each_pass)
+                try validator.validate(module);
+        }
+        return changed;
+    }
+};
+
+fn satisfies(actual: module_ir.Properties, required: module_ir.Properties) bool {
+    inline for (property_names) |name| {
+        if (@field(required, name) and !@field(actual, name))
+            return false;
+    }
+    return true;
+}
+
+fn applyProduced(properties: *module_ir.Properties, produced: module_ir.Properties) void {
+    inline for (property_names) |name| {
+        if (@field(produced, name))
+            @field(properties, name) = true;
+    }
+}
+
+fn applyInvalidated(properties: *module_ir.Properties, invalidated: module_ir.Properties) void {
+    inline for (property_names) |name| {
+        if (@field(invalidated, name))
+            @field(properties, name) = false;
+    }
+}
+
+const property_names = .{
+    "valid_cfg",
+    "valid_ssa",
+    "structured_control_flow",
+    "no_function_calls",
+    "no_local_memory",
+    "no_matrix_types",
+    "no_large_composites",
+    "explicit_resource_offsets",
+};

src/compiler/ir/printer.zig

@@ -0,0 +1,336 @@
+const std = @import("std");
+const ids = @import("id.zig");
+const inst_ir = @import("instruction.zig");
+const module_ir = @import("module.zig");
+
+const indent = "    ";
+
+pub fn write(module: *const module_ir.Module, writer: *std.Io.Writer) std.Io.Writer.Error!void {
+    try writer.print("shader {t}", .{module.stage});
+
+    if (module.entry_point) |entry| {
+        try writer.writeByte(' ');
+        try writeFunctionRef(module, writer, entry);
+    }
+
+    try writer.writeAll("\n{\n");
+
+    for (module.interface_variables.entries.items, 0..) |entry, index| {
+        const variable = entry orelse continue;
+
+        try writer.writeAll(indent);
+        try writeNamedRef(writer, variable.name, "interface", index);
+        try writer.writeAll(": ");
+        try writeType(module, writer, variable.type);
+        try writer.print(" = {t}[", .{variable.direction});
+
+        switch (variable.semantic) {
+            .location => |location| try writer.print("location({d}), component({d}), index({d})", .{ location.location, location.component, location.index }),
+            .builtin => |builtin| try writer.print("builtin({t})", .{builtin}),
+        }
+
+        try writer.writeAll("]\n");
+    }
+
+    for (module.constants.entries.items, 0..) |entry, constant_index| {
+        const constant = entry orelse continue;
+        const value_id = constantValueId(module, ids.ConstantId.fromIndex(constant_index)) orelse continue;
+
+        try writer.writeAll(indent);
+        try writeValueRef(module, writer, value_id);
+        try writer.writeAll(": constant ");
+        try writeType(module, writer, constant.type);
+        try writer.writeAll(" = ");
+
+        switch (constant.value) {
+            .boolean => |value| try writer.print("{}", .{value}),
+            .integer_bits => |bits| try writer.print("bits(0x{x})", .{bits}),
+            .float_bits => |bits| try writer.print("bits(0x{x})", .{bits}),
+            .null => try writer.writeAll("null"),
+            .undef => try writer.writeAll("undef"),
+            .composite => |elements| {
+                try writer.writeByte('[');
+                for (elements, 0..) |element, index| {
+                    if (index != 0) try writer.writeAll(", ");
+                    try writer.print("#{d}", .{element.index()});
+                }
+                try writer.writeByte(']');
+            },
+        }
+
+        try writer.writeByte('\n');
+    }
+
+    for (module.functions.entries.items, 0..) |entry, function_index| {
+        const function = entry orelse continue;
+
+        try writer.writeAll("\n" ++ indent ++ "fn ");
+        try writeNamedRef(writer, function.name, "fn", function_index);
+        try writer.writeByte('(');
+        for (function.parameters.items, 0..) |parameter, index| {
+            if (index != 0)
+                try writer.writeAll(", ");
+            try writeValueRef(module, writer, parameter);
+            try writer.writeAll(": ");
+            try writeType(module, writer, function.parameter_types.items[index]);
+        }
+        try writer.writeAll(") -> ");
+        try writeType(module, writer, function.return_type);
+        try writer.writeAll("\n" ++ indent ++ "{\n");
+
+        for (function.blocks.items) |block_id| {
+            const block = module.blocks.get(block_id) orelse continue;
+
+            try writer.writeAll(indent ** 2);
+            try writeBlockRef(module, writer, block_id);
+            try writer.writeByte('(');
+            for (block.parameters.items, 0..) |parameter, index| {
+                if (index != 0) try writer.writeAll(", ");
+                try writeValueRef(module, writer, parameter);
+                try writer.writeAll(": ");
+                try writeType(module, writer, module.typeOf(parameter).?);
+            }
+            try writer.writeAll("):\n");
+
+            for (block.instructions.items) |instruction_id| {
+                const instruction = module.instructions.get(instruction_id) orelse continue;
+                try writer.writeAll(indent ** 3);
+                if (instruction.result) |result| {
+                    try writeValueRef(module, writer, result);
+                    try writer.writeAll(": ");
+                    try writeType(module, writer, module.typeOf(result).?);
+                    try writer.writeAll(" = ");
+                }
+                try writeOperation(module, writer, instruction.operation);
+                try writer.writeByte('\n');
+            }
+
+            if (block.terminator) |terminator| {
+                try writer.writeAll(indent ** 3);
+                try writeTerminator(module, writer, terminator);
+                try writer.writeAll("\n\n");
+            } else {
+                try writer.writeAll(indent ** 3 ++ "<missing terminator>\n\n");
+            }
+        }
+        try writer.writeAll(indent ++ "}\n");
+    }
+    try writer.writeAll("}\n");
+}
+
+pub fn allocPrint(allocator: std.mem.Allocator, module: *const module_ir.Module) ![]u8 {
+    var output: std.Io.Writer.Allocating = .init(allocator);
+    defer output.deinit();
+    try write(module, &output.writer);
+    return output.toOwnedSlice();
+}
+
+fn writeType(module: *const module_ir.Module, writer: *std.Io.Writer, type_id: ids.TypeId) !void {
+    const ty = module.types.get(type_id) orelse {
+        try writer.print("<invalid-type-{d}>", .{type_id.index()});
+        return;
+    };
+
+    switch (ty.*) {
+        .void => try writer.writeAll("void"),
+        .boolean => try writer.writeAll("bool"),
+        .integer => |integer| try writer.print("{s}{d}", .{ if (integer.signedness == .signed) "i" else "u", integer.bits }),
+        .floating => |float| try writer.print("f{d}", .{float.bits}),
+        .vector => |vector| {
+            try writer.print("vec{d}[", .{vector.length});
+            try writeType(module, writer, vector.element_type);
+            try writer.writeByte(']');
+        },
+        .array => |array| {
+            try writer.writeAll("array[");
+            try writeType(module, writer, array.element_type);
+            try writer.print(", {d}]", .{array.length});
+        },
+        .structure => |structure| {
+            try writer.writeAll("struct[");
+            for (structure.members, 0..) |member, index| {
+                if (index != 0) try writer.writeAll(", ");
+                try writeType(module, writer, member);
+            }
+            try writer.writeByte(']');
+        },
+        .pointer => |pointer| {
+            try writer.print("ptr[{t}, ", .{pointer.address_space});
+            try writeType(module, writer, pointer.pointee_type);
+            try writer.writeByte(']');
+        },
+        .resource_handle => |handle| try writer.print("resourceHandle[{t}]", .{handle.kind}),
+    }
+}
+
+fn writeOperation(module: *const module_ir.Module, writer: *std.Io.Writer, operation: inst_ir.Operation) !void {
+    switch (operation) {
+        .unary => |op| {
+            try writer.print("{t} ", .{op.opcode});
+            try writeValueRef(module, writer, op.operand);
+        },
+        .binary => |op| {
+            try writer.print("{t} ", .{op.opcode});
+            try writeValueRef(module, writer, op.lhs);
+            try writer.writeAll(", ");
+            try writeValueRef(module, writer, op.rhs);
+        },
+        .compare => |op| {
+            try writer.print("cmp_{t} ", .{op.opcode});
+            try writeValueRef(module, writer, op.lhs);
+            try writer.writeAll(", ");
+            try writeValueRef(module, writer, op.rhs);
+        },
+        .select => |op| {
+            try writer.writeAll("select ");
+            try writeValueRef(module, writer, op.condition);
+            try writer.writeAll(", ");
+            try writeValueRef(module, writer, op.true_value);
+            try writer.writeAll(", ");
+            try writeValueRef(module, writer, op.false_value);
+        },
+        .bitcast => |value| {
+            try writer.writeAll("bitcast ");
+            try writeValueRef(module, writer, value);
+        },
+        .composite_construct => |op| {
+            try writer.writeAll("composite_construct ");
+            try writeValueList(module, writer, op.elements);
+        },
+        .composite_extract => |op| {
+            try writer.writeAll("composite_extract ");
+            try writeValueRef(module, writer, op.composite);
+            for (op.indices) |index| try writer.print("[{d}]", .{index});
+        },
+        .load_interface => |op| {
+            try writer.writeAll("load_interface ");
+            const variable = module.interface_variables.get(op.variable);
+            try writeNamedRef(writer, if (variable) |v| v.name else null, "interface", op.variable.index());
+        },
+        .store_interface => |op| {
+            try writer.writeAll("store_interface ");
+            const variable = module.interface_variables.get(op.variable);
+            try writeNamedRef(writer, if (variable) |v| v.name else null, "interface", op.variable.index());
+            try writer.writeAll(", ");
+            try writeValueRef(module, writer, op.value);
+        },
+        .call => |op| {
+            try writer.writeAll("call ");
+            try writeFunctionRef(module, writer, op.function);
+            try writer.writeByte('(');
+            try writeValueList(module, writer, op.arguments);
+            try writer.writeByte(')');
+        },
+    }
+}
+
+fn writeTerminator(module: *const module_ir.Module, writer: *std.Io.Writer, terminator: module_ir.Terminator) !void {
+    switch (terminator) {
+        .branch => |edge| {
+            try writer.writeAll("branch ");
+            try writeEdge(module, writer, edge);
+        },
+        .conditional_branch => |branch| {
+            try writer.writeAll("conditional_branch ");
+            try writeValueRef(module, writer, branch.condition);
+            try writer.writeAll(", ");
+            try writeEdge(module, writer, branch.true_edge);
+            try writer.writeAll(", ");
+            try writeEdge(module, writer, branch.false_edge);
+        },
+        .return_void => try writer.writeAll("return"),
+        .return_value => |value| {
+            try writer.writeAll("return ");
+            try writeValueRef(module, writer, value);
+        },
+        .discard => try writer.writeAll("discard"),
+        .@"unreachable" => try writer.writeAll("unreachable"),
+    }
+}
+
+fn writeEdge(module: *const module_ir.Module, writer: *std.Io.Writer, edge: module_ir.Edge) !void {
+    try writeBlockRef(module, writer, edge.target);
+    try writer.writeByte('(');
+    try writeValueList(module, writer, edge.arguments);
+    try writer.writeByte(')');
+}
+
+fn writeValueList(module: *const module_ir.Module, writer: *std.Io.Writer, value_ids: []const ids.ValueId) !void {
+    for (value_ids, 0..) |value, index| {
+        if (index != 0)
+            try writer.writeAll(", ");
+        try writeValueRef(module, writer, value);
+    }
+}
+
+fn writeValueRef(module: *const module_ir.Module, writer: *std.Io.Writer, value_id: ids.ValueId) !void {
+    try writer.writeByte('%');
+
+    const value = module.values.get(value_id);
+    if (value) |data| {
+        if (data.name) |name| {
+            if (isValidName(name) and isUniqueValueName(module, value_id, name)) {
+                try writer.writeAll(name);
+                return;
+            }
+        }
+    }
+
+    try writer.print("{d}", .{value_id.index()});
+}
+
+fn isValidName(name: []const u8) bool {
+    if (name.len == 0 or (!std.ascii.isAlphabetic(name[0]) and name[0] != '_'))
+        return false;
+
+    for (name[1..]) |byte| {
+        if (!std.ascii.isAlphanumeric(byte) and byte != '_')
+            return false;
+    }
+    return true;
+}
+
+fn isUniqueValueName(module: *const module_ir.Module, value_id: ids.ValueId, name: []const u8) bool {
+    for (module.values.entries.items, 0..) |entry, index| {
+        if (index == value_id.index())
+            continue;
+
+        const other = entry orelse continue;
+        if (other.name) |other_name| {
+            if (std.mem.eql(u8, name, other_name))
+                return false;
+        }
+    }
+    return true;
+}
+
+fn writeBlockRef(module: *const module_ir.Module, writer: *std.Io.Writer, block: ids.BlockId) !void {
+    const value = module.blocks.get(block);
+    try writeNamedRef(writer, if (value) |b| b.name else null, "b", block.index());
+}
+
+fn writeFunctionRef(module: *const module_ir.Module, writer: *std.Io.Writer, function: ids.FunctionId) !void {
+    const value = module.functions.get(function);
+    try writeNamedRef(writer, if (value) |f| f.name else null, "fn", function.index());
+}
+
+fn writeNamedRef(writer: *std.Io.Writer, name: ?[]const u8, fallback: []const u8, index: usize) !void {
+    try writer.writeByte(if (std.mem.eql(u8, fallback, "b")) '.' else '@');
+    if (name) |text| {
+        if (isValidName(text)) {
+            try writer.writeAll(text);
+            return;
+        }
+    }
+    try writer.print("{s}{d}", .{ fallback, index });
+}
+
+fn constantValueId(module: *const module_ir.Module, constant_id: ids.ConstantId) ?ids.ValueId {
+    for (module.values.entries.items, 0..) |entry, index| {
+        const value = entry orelse continue;
+        if (value.definition == .constant and value.definition.constant == constant_id)
+            return ids.ValueId.fromIndex(index);
+    }
+
+    return null;
+}

src/compiler/ir/type.zig

@@ -0,0 +1,106 @@
+const std = @import("std");
+const ids = @import("id.zig");
+
+pub const TypeId = ids.TypeId;
+
+pub const Signedness = enum { signed, unsigned };
+
+pub const IntegerType = struct {
+    bits: u16,
+    signedness: Signedness,
+};
+
+pub const FloatType = struct {
+    bits: u16,
+};
+
+pub const VectorType = struct {
+    element_type: TypeId,
+    length: u8,
+};
+
+pub const ArrayType = struct {
+    element_type: TypeId,
+    length: u32,
+};
+
+pub const StructureType = struct {
+    members: []const TypeId,
+};
+
+pub const AddressSpace = enum {
+    function,
+    private,
+    workgroup,
+    input,
+    output,
+    uniform,
+    storage,
+    push_constant,
+    physical,
+};
+
+pub const PointerType = struct {
+    address_space: AddressSpace,
+    pointee_type: TypeId,
+};
+
+pub const ResourceKind = enum {
+    uniform_buffer,
+    storage_buffer,
+    sampled_image,
+    storage_image,
+    sampler,
+};
+
+pub const ResourceHandleType = struct {
+    kind: ResourceKind,
+    data_type: ?TypeId = null,
+};
+
+pub const Type = union(enum) {
+    void,
+    boolean,
+    integer: IntegerType,
+    floating: FloatType,
+    vector: VectorType,
+    array: ArrayType,
+    structure: StructureType,
+    pointer: PointerType,
+    resource_handle: ResourceHandleType,
+
+    pub fn eql(a: Type, b: Type) bool {
+        return switch (a) {
+            .void => b == .void,
+            .boolean => b == .boolean,
+            .integer => |value| switch (b) {
+                .integer => |other| std.meta.eql(value, other),
+                else => false,
+            },
+            .floating => |value| switch (b) {
+                .floating => |other| std.meta.eql(value, other),
+                else => false,
+            },
+            .vector => |value| switch (b) {
+                .vector => |other| std.meta.eql(value, other),
+                else => false,
+            },
+            .array => |value| switch (b) {
+                .array => |other| std.meta.eql(value, other),
+                else => false,
+            },
+            .structure => |value| switch (b) {
+                .structure => |other| std.mem.eql(TypeId, value.members, other.members),
+                else => false,
+            },
+            .pointer => |value| switch (b) {
+                .pointer => |other| std.meta.eql(value, other),
+                else => false,
+            },
+            .resource_handle => |value| switch (b) {
+                .resource_handle => |other| std.meta.eql(value, other),
+                else => false,
+            },
+        };
+    }
+};

src/compiler/ir/validator/dominance.zig

@@ -0,0 +1,93 @@
+const std = @import("std");
+const cfg = @import("../cfg.zig");
+const ids = @import("../id.zig");
+const module_ir = @import("../module.zig");
+
+pub const Error = std.mem.Allocator.Error || error{
+    InvalidBlock,
+    DefinitionDoesNotDominateUse,
+};
+
+const DominanceUseContext = struct {
+    module: *const module_ir.Module,
+    analysis: *const cfg,
+    function_id: ids.FunctionId,
+    use_block: ids.BlockId,
+    use_index: usize,
+    valid: bool = true,
+};
+
+pub fn validate(module: *const module_ir.Module, function_id: ids.FunctionId) Error!void {
+    var analysis = cfg.init(module.backingAllocator(), module, function_id) catch |err| switch (err) {
+        error.OutOfMemory => return error.OutOfMemory,
+        else => return error.InvalidBlock,
+    };
+    defer analysis.deinit();
+
+    const function = module.functions.get(function_id).?;
+    for (function.blocks.items) |block_id| {
+        const block = module.blocks.get(block_id).?;
+
+        for (block.instructions.items, 0..) |instruction_id, instruction_index| {
+            const instruction = module.instructions.get(instruction_id).?;
+            var context: DominanceUseContext = .{
+                .module = module,
+                .analysis = &analysis,
+                .function_id = function_id,
+                .use_block = block_id,
+                .use_index = instruction_index,
+            };
+
+            instruction.operation.visitValueUses(&context, checkDominanceUse);
+
+            if (!context.valid)
+                return error.DefinitionDoesNotDominateUse;
+        }
+
+        var context: DominanceUseContext = .{
+            .module = module,
+            .analysis = &analysis,
+            .function_id = function_id,
+            .use_block = block_id,
+            .use_index = block.instructions.items.len,
+        };
+
+        module_ir.visitTerminatorValueUses(block.terminator.?, &context, checkDominanceUse);
+
+        if (!context.valid)
+            return error.DefinitionDoesNotDominateUse;
+    }
+}
+
+fn checkDominanceUse(context: *DominanceUseContext, value_id: ids.ValueId) void {
+    if (!context.valid)
+        return;
+
+    const value = context.module.values.get(value_id) orelse {
+        context.valid = false;
+        return;
+    };
+
+    context.valid = switch (value.definition) {
+        .constant, .undef => true,
+        .function_parameter => |definition| definition.function == context.function_id,
+        .block_parameter => |definition| context.analysis.dominates(definition.block, context.use_block),
+        .instruction => |instruction_id| blk: {
+            const definition = context.module.instructions.get(instruction_id) orelse break :blk false;
+
+            if (!context.analysis.dominates(definition.parent_block, context.use_block))
+                break :blk false;
+
+            if (definition.parent_block != context.use_block)
+                break :blk true;
+
+            const block = context.module.blocks.get(context.use_block) orelse break :blk false;
+            for (block.instructions.items, 0..) |candidate, definition_index| {
+                if (candidate == instruction_id)
+                    break :blk definition_index < context.use_index;
+            }
+
+            break :blk false;
+        },
+    };
+}

src/compiler/ir/validator/validator.zig

@@ -0,0 +1,470 @@
+const std = @import("std");
+const ids = @import("../id.zig");
+const type_ir = @import("../type.zig");
+const inst_ir = @import("../instruction.zig");
+const module_ir = @import("../module.zig");
+const dominance = @import("dominance.zig");
+
+pub const ValidationError = error{
+    MissingEntryPoint,
+    InvalidEntryPoint,
+    InvalidType,
+    InvalidConstant,
+    InvalidValue,
+    InvalidFunction,
+    InvalidBlock,
+    InvalidInstruction,
+    MissingFunctionEntryBlock,
+    MissingTerminator,
+    EntryBlockHasPredecessor,
+    WrongParent,
+    WrongDefinition,
+    WrongParameterIndex,
+    WrongResultPresence,
+    WrongOperandType,
+    WrongResultType,
+    WrongBranchArgumentCount,
+    WrongBranchArgumentType,
+    CrossFunctionReference,
+    WrongReturnType,
+    WrongInterfaceDirection,
+    InvalidStructuredControl,
+    DefinitionDoesNotDominateUse,
+};
+
+pub const Error = ValidationError || std.mem.Allocator.Error;
+
+/// Early validator for the foundational IR. It covers object ownership, CFG
+/// edges, single definitions, function boundaries, and the currently modeled
+/// operation types, and SSA dominance.
+pub fn validate(module: *const module_ir.Module) Error!void {
+    const entry_point = module.entry_point orelse return error.MissingEntryPoint;
+    if (!module.functions.isLive(entry_point))
+        return error.InvalidEntryPoint;
+
+    for (module.types.entries.items) |entry| {
+        const ty = entry orelse continue;
+        try validateType(module, ty);
+    }
+
+    for (module.constants.entries.items) |entry| {
+        const constant = entry orelse continue;
+
+        if (!module.types.isLive(constant.type))
+            return error.InvalidType;
+
+        if (constant.value == .composite) {
+            for (constant.value.composite) |element| {
+                if (!module.constants.isLive(element))
+                    return error.InvalidConstant;
+            }
+        }
+    }
+
+    for (module.values.entries.items, 0..) |entry, value_index| {
+        const value = entry orelse continue;
+
+        if (!module.types.isLive(value.type))
+            return error.InvalidType;
+
+        const value_id = ids.ValueId.fromIndex(value_index);
+
+        switch (value.definition) {
+            .constant => |id| {
+                const constant = module.constants.get(id) orelse return error.InvalidConstant;
+                if (constant.type != value.type)
+                    return error.WrongResultType;
+            },
+            .function_parameter => |definition| {
+                const function = module.functions.get(definition.function) orelse return error.InvalidFunction;
+                if (definition.index >= function.parameters.items.len or function.parameters.items[definition.index] != value_id)
+                    return error.WrongParameterIndex;
+            },
+            .block_parameter => |definition| {
+                const block = module.blocks.get(definition.block) orelse return error.InvalidBlock;
+                if (definition.index >= block.parameters.items.len or block.parameters.items[definition.index] != value_id)
+                    return error.WrongParameterIndex;
+            },
+            .instruction => |instruction_id| {
+                const instruction = module.instructions.get(instruction_id) orelse return error.InvalidInstruction;
+                if (instruction.result != value_id)
+                    return error.WrongDefinition;
+            },
+            .undef => {},
+        }
+    }
+
+    for (module.interface_variables.entries.items) |entry| {
+        const variable = entry orelse continue;
+        if (!module.types.isLive(variable.type))
+            return error.InvalidType;
+    }
+
+    for (module.resources.entries.items) |entry| {
+        const resource = entry orelse continue;
+        if (!module.types.isLive(resource.type))
+            return error.InvalidType;
+    }
+
+    for (module.functions.entries.items, 0..) |entry, function_index| {
+        const function = entry orelse continue;
+        const function_id = ids.FunctionId.fromIndex(function_index);
+
+        if (!module.types.isLive(function.return_type))
+            return error.InvalidType;
+
+        if (function.parameter_types.items.len != function.parameters.items.len)
+            return error.WrongParameterIndex;
+
+        for (function.parameter_types.items, function.parameters.items, 0..) |parameter_type, parameter_id, index| {
+            const parameter = module.values.get(parameter_id) orelse return error.InvalidValue;
+
+            if (parameter.type != parameter_type)
+                return error.WrongResultType;
+
+            if (parameter.definition != .function_parameter or
+                parameter.definition.function_parameter.function != function_id or
+                parameter.definition.function_parameter.index != index)
+                return error.WrongDefinition;
+        }
+
+        const entry_block = function.entry_block orelse return error.MissingFunctionEntryBlock;
+        const entry_block_value = module.blocks.get(entry_block) orelse return error.InvalidBlock;
+        if (entry_block_value.parent_function != function_id) return error.WrongParent;
+
+        for (function.blocks.items) |block_id| {
+            const block = module.blocks.get(block_id) orelse return error.InvalidBlock;
+
+            if (block.parent_function != function_id)
+                return error.WrongParent;
+
+            try validateBlock(module, function_id, block_id, block);
+        }
+
+        for (function.blocks.items) |block_id| {
+            const block = module.blocks.get(block_id).?;
+
+            if (block.terminator) |terminator| {
+                if (targetsBlock(terminator, entry_block))
+                    return error.EntryBlockHasPredecessor;
+            }
+        }
+
+        try dominance.validate(module, function_id);
+    }
+}
+
+fn validateType(module: *const module_ir.Module, ty: type_ir.Type) ValidationError!void {
+    switch (ty) {
+        .vector => |vector| {
+            if (!module.types.isLive(vector.element_type) or vector.length < 2)
+                return error.InvalidType;
+        },
+        .array => |array| {
+            if (!module.types.isLive(array.element_type) or array.length == 0)
+                return error.InvalidType;
+        },
+        .structure => |structure| for (structure.members) |member| {
+            if (!module.types.isLive(member))
+                return error.InvalidType;
+        },
+        .pointer => |pointer| {
+            if (!module.types.isLive(pointer.pointee_type))
+                return error.InvalidType;
+        },
+        .resource_handle => |handle| if (handle.data_type) |data_type| {
+            if (!module.types.isLive(data_type))
+                return error.InvalidType;
+        },
+        else => {},
+    }
+}
+
+fn validateBlock(
+    module: *const module_ir.Module,
+    function_id: ids.FunctionId,
+    block_id: ids.BlockId,
+    block: *const module_ir.Block,
+) ValidationError!void {
+    for (block.parameters.items, 0..) |parameter_id, index| {
+        const parameter = module.values.get(parameter_id) orelse return error.InvalidValue;
+        if (parameter.definition != .block_parameter or
+            parameter.definition.block_parameter.block != block_id or
+            parameter.definition.block_parameter.index != index)
+            return error.WrongDefinition;
+    }
+
+    switch (block.structured_control) {
+        .none => {},
+        .selection => |selection| try validateTarget(module, function_id, selection.merge_block),
+        .loop => |loop| {
+            try validateTarget(module, function_id, loop.merge_block);
+            try validateTarget(module, function_id, loop.continue_block);
+        },
+    }
+
+    for (block.instructions.items) |instruction_id| {
+        const instruction = module.instructions.get(instruction_id) orelse return error.InvalidInstruction;
+
+        if (instruction.parent_block != block_id)
+            return error.WrongParent;
+
+        if (instruction.result) |result_id| {
+            const result = module.values.get(result_id) orelse return error.InvalidValue;
+
+            if (result.definition != .instruction or result.definition.instruction != instruction_id)
+                return error.WrongDefinition;
+        }
+
+        try validateOperation(module, function_id, instruction);
+    }
+
+    const terminator = block.terminator orelse return error.MissingTerminator;
+    try validateTerminator(module, function_id, terminator);
+}
+
+fn validateOperation(module: *const module_ir.Module, function_id: ids.FunctionId, instruction: *const inst_ir.Instruction) ValidationError!void {
+    const result_type = if (instruction.result) |result| module.typeOf(result) orelse return error.InvalidValue else null;
+
+    switch (instruction.operation) {
+        .unary => |op| {
+            const operand_type = try operandType(module, function_id, op.operand);
+
+            if (result_type == null)
+                return error.WrongResultPresence;
+
+            if (result_type.? != operand_type)
+                return error.WrongResultType;
+        },
+        .binary => |op| {
+            const lhs_type = try operandType(module, function_id, op.lhs);
+            const rhs_type = try operandType(module, function_id, op.rhs);
+
+            if (lhs_type != rhs_type)
+                return error.WrongOperandType;
+
+            if (result_type == null or result_type.? != lhs_type)
+                return error.WrongResultType;
+        },
+        .compare => |op| {
+            const lhs_type = try operandType(module, function_id, op.lhs);
+            if (try operandType(module, function_id, op.rhs) != lhs_type)
+                return error.WrongOperandType;
+
+            const result = result_type orelse return error.WrongResultPresence;
+            if (!isBoolean(module, result))
+                return error.WrongResultType;
+        },
+        .select => |op| {
+            if (!isBoolean(module, try operandType(module, function_id, op.condition)))
+                return error.WrongOperandType;
+
+            const true_type = try operandType(module, function_id, op.true_value);
+            if (try operandType(module, function_id, op.false_value) != true_type)
+                return error.WrongOperandType;
+
+            if (result_type == null or result_type.? != true_type)
+                return error.WrongResultType;
+        },
+        .bitcast => |operand| {
+            _ = try operandType(module, function_id, operand);
+            if (result_type == null)
+                return error.WrongResultPresence;
+        },
+        .composite_construct => |op| {
+            const result = result_type orelse return error.WrongResultPresence;
+            const ty = module.types.get(result) orelse return error.InvalidType;
+
+            switch (ty.*) {
+                .vector => |vector| {
+                    if (op.elements.len != vector.length)
+                        return error.WrongOperandType;
+
+                    for (op.elements) |element| {
+                        if (try operandType(module, function_id, element) != vector.element_type)
+                            return error.WrongOperandType;
+                    }
+                },
+                .structure => |structure| {
+                    if (op.elements.len != structure.members.len)
+                        return error.WrongOperandType;
+
+                    for (op.elements, structure.members) |element, member_type| {
+                        if (try operandType(module, function_id, element) != member_type)
+                            return error.WrongOperandType;
+                    }
+                },
+                else => return error.WrongResultType,
+            }
+        },
+        .composite_extract => |op| {
+            const composite_type = try operandType(module, function_id, op.composite);
+            const extracted_type = try indexedType(module, composite_type, op.indices);
+
+            if (result_type == null or result_type.? != extracted_type)
+                return error.WrongResultType;
+        },
+        .load_interface => |op| {
+            const variable = module.interface_variables.get(op.variable) orelse return error.InvalidValue;
+
+            if (variable.direction != .input)
+                return error.WrongInterfaceDirection;
+
+            if (op.element_index) |index|
+                _ = try operandType(module, function_id, index);
+
+            if (result_type == null or result_type.? != variable.type)
+                return error.WrongResultType;
+        },
+        .store_interface => |op| {
+            if (result_type != null)
+                return error.WrongResultPresence;
+
+            const variable = module.interface_variables.get(op.variable) orelse return error.InvalidValue;
+
+            if (variable.direction != .output)
+                return error.WrongInterfaceDirection;
+
+            if (try operandType(module, function_id, op.value) != variable.type)
+                return error.WrongOperandType;
+
+            if (op.element_index) |index|
+                _ = try operandType(module, function_id, index);
+        },
+        .call => |op| {
+            const callee = module.functions.get(op.function) orelse return error.InvalidFunction;
+
+            if (op.arguments.len != callee.parameter_types.items.len)
+                return error.WrongOperandType;
+
+            for (op.arguments, callee.parameter_types.items) |argument, parameter_type| {
+                if (try operandType(module, function_id, argument) != parameter_type)
+                    return error.WrongOperandType;
+            }
+
+            const return_type = module.types.get(callee.return_type) orelse return error.InvalidType;
+
+            if (return_type.* == .void) {
+                if (result_type != null)
+                    return error.WrongResultPresence;
+            } else if (result_type == null or result_type.? != callee.return_type)
+                return error.WrongResultType;
+        },
+    }
+}
+
+fn validateTerminator(module: *const module_ir.Module, function_id: ids.FunctionId, terminator: module_ir.Terminator) ValidationError!void {
+    const function = module.functions.get(function_id) orelse return error.InvalidFunction;
+
+    switch (terminator) {
+        .branch => |edge| try validateEdge(module, function_id, edge),
+        .conditional_branch => |branch| {
+            if (!isBoolean(module, try operandType(module, function_id, branch.condition)))
+                return error.WrongOperandType;
+
+            try validateEdge(module, function_id, branch.true_edge);
+            try validateEdge(module, function_id, branch.false_edge);
+        },
+        .return_void => {
+            if (module.types.get(function.return_type).?.* != .void)
+                return error.WrongReturnType;
+        },
+        .return_value => |value| {
+            if (try operandType(module, function_id, value) != function.return_type)
+                return error.WrongReturnType;
+        },
+        .discard => {
+            if (module.stage != .fragment)
+                return error.WrongReturnType;
+        },
+        .@"unreachable" => {},
+    }
+}
+
+fn validateEdge(module: *const module_ir.Module, function_id: ids.FunctionId, edge: module_ir.Edge) ValidationError!void {
+    const target = module.blocks.get(edge.target) orelse return error.InvalidBlock;
+
+    if (target.parent_function != function_id)
+        return error.CrossFunctionReference;
+
+    if (edge.arguments.len != target.parameters.items.len)
+        return error.WrongBranchArgumentCount;
+
+    for (edge.arguments, target.parameters.items) |argument, parameter| {
+        if (try operandType(module, function_id, argument) != module.typeOf(parameter).?)
+            return error.WrongBranchArgumentType;
+    }
+}
+
+fn validateTarget(module: *const module_ir.Module, function_id: ids.FunctionId, target_id: ids.BlockId) ValidationError!void {
+    const target = module.blocks.get(target_id) orelse return error.InvalidStructuredControl;
+    if (target.parent_function != function_id)
+        return error.InvalidStructuredControl;
+}
+
+fn operandType(module: *const module_ir.Module, function_id: ids.FunctionId, value_id: ids.ValueId) ValidationError!ids.TypeId {
+    const value = module.values.get(value_id) orelse return error.InvalidValue;
+    const owner = valueFunction(module, value_id) catch return error.InvalidValue;
+
+    if (owner) |actual| {
+        if (actual != function_id)
+            return error.CrossFunctionReference;
+    }
+    return value.type;
+}
+
+fn valueFunction(module: *const module_ir.Module, value_id: ids.ValueId) ValidationError!?ids.FunctionId {
+    const value = module.values.get(value_id) orelse return error.InvalidValue;
+
+    return switch (value.definition) {
+        .constant, .undef => null,
+        .function_parameter => |definition| definition.function,
+        .block_parameter => |definition| (module.blocks.get(definition.block) orelse return error.InvalidBlock).parent_function,
+        .instruction => |instruction_id| blk: {
+            const instruction = module.instructions.get(instruction_id) orelse return error.InvalidInstruction;
+            const block = module.blocks.get(instruction.parent_block) orelse return error.InvalidBlock;
+            break :blk block.parent_function;
+        },
+    };
+}
+
+fn indexedType(module: *const module_ir.Module, root: ids.TypeId, indices: []const u32) ValidationError!ids.TypeId {
+    if (indices.len == 0)
+        return error.WrongOperandType;
+
+    var current = root;
+
+    for (indices) |index| {
+        const ty = module.types.get(current) orelse return error.InvalidType;
+        current = switch (ty.*) {
+            .vector => |vector| if (index < vector.length)
+                vector.element_type
+            else
+                return error.WrongOperandType,
+            .array => |array| if (index < array.length)
+                array.element_type
+            else
+                return error.WrongOperandType,
+            .structure => |structure| if (index < structure.members.len)
+                structure.members[index]
+            else
+                return error.WrongOperandType,
+
+            else => return error.WrongOperandType,
+        };
+    }
+    return current;
+}
+
+fn isBoolean(module: *const module_ir.Module, type_id: ids.TypeId) bool {
+    const ty = module.types.get(type_id) orelse return false;
+    return ty.* == .boolean;
+}
+
+fn targetsBlock(terminator: module_ir.Terminator, target: ids.BlockId) bool {
+    return switch (terminator) {
+        .branch => |edge| edge.target == target,
+        .conditional_branch => |branch| branch.true_edge.target == target or branch.false_edge.target == target,
+        else => false,
+    };
+}

src/compiler/ir/value.zig

@@ -0,0 +1,31 @@
+const ids = @import("id.zig");
+
+pub const TypeId = ids.TypeId;
+pub const ConstantId = ids.ConstantId;
+pub const InstructionId = ids.InstructionId;
+pub const FunctionId = ids.FunctionId;
+pub const BlockId = ids.BlockId;
+
+pub const FunctionParameterDefinition = struct {
+    function: FunctionId,
+    index: u32,
+};
+
+pub const BlockParameterDefinition = struct {
+    block: BlockId,
+    index: u32,
+};
+
+pub const Definition = union(enum) {
+    constant: ConstantId,
+    function_parameter: FunctionParameterDefinition,
+    block_parameter: BlockParameterDefinition,
+    instruction: InstructionId,
+    undef,
+};
+
+pub const Value = struct {
+    type: TypeId,
+    definition: Definition,
+    name: ?[]const u8 = null,
+};

src/compiler/ir/visitor.zig

@@ -0,0 +1,98 @@
+const ids = @import("id.zig");
+const instruction_ir = @import("instruction.zig");
+const module_ir = @import("module.zig");
+
+pub const Visitor = struct {
+    context: ?*anyopaque = null,
+    visitInterfaceVariable: ?*const fn (?*anyopaque, ids.InterfaceVariableId, *const module_ir.InterfaceVariable) anyerror!void = null,
+    visitResource: ?*const fn (?*anyopaque, ids.ResourceId, *const module_ir.Resource) anyerror!void = null,
+    visitFunction: ?*const fn (?*anyopaque, ids.FunctionId, *const module_ir.Function) anyerror!void = null,
+    visitBlock: ?*const fn (?*anyopaque, ids.BlockId, *const module_ir.Block) anyerror!void = null,
+    visitInstruction: ?*const fn (?*anyopaque, ids.InstructionId, *const instruction_ir.Instruction) anyerror!void = null,
+    visitTerminator: ?*const fn (?*anyopaque, ids.BlockId, module_ir.Terminator) anyerror!void = null,
+    visitValueUse: ?*const fn (?*anyopaque, ids.BlockId, ?ids.InstructionId, ids.ValueId) anyerror!void = null,
+};
+
+const UseContext = struct {
+    visitor_context: ?*anyopaque,
+    callback: *const fn (?*anyopaque, ids.BlockId, ?ids.InstructionId, ids.ValueId) anyerror!void,
+    block: ids.BlockId,
+    instruction: ?ids.InstructionId,
+    failure: ?anyerror = null,
+};
+
+pub fn walk(module: *const module_ir.Module, visitor: Visitor) !void {
+    for (module.interface_variables.entries.items, 0..) |entry, index| {
+        const variable = entry orelse continue;
+        if (visitor.visitInterfaceVariable) |callback|
+            try callback(visitor.context, ids.InterfaceVariableId.fromIndex(index), &variable);
+    }
+
+    for (module.resources.entries.items, 0..) |entry, index| {
+        const resource = entry orelse continue;
+        if (visitor.visitResource) |callback|
+            try callback(visitor.context, ids.ResourceId.fromIndex(index), &resource);
+    }
+
+    for (module.functions.entries.items, 0..) |entry, function_index| {
+        const function = entry orelse continue;
+        const function_id = ids.FunctionId.fromIndex(function_index);
+
+        if (visitor.visitFunction) |callback|
+            try callback(visitor.context, function_id, &function);
+
+        for (function.blocks.items) |block_id| {
+            const block = module.blocks.get(block_id) orelse continue;
+
+            if (visitor.visitBlock) |callback|
+                try callback(visitor.context, block_id, block);
+
+            for (block.instructions.items) |instruction_id| {
+                const instruction = module.instructions.get(instruction_id) orelse continue;
+
+                if (visitor.visitInstruction) |callback|
+                    try callback(visitor.context, instruction_id, instruction);
+
+                if (visitor.visitValueUse) |callback| {
+                    var context: UseContext = .{
+                        .visitor_context = visitor.context,
+                        .callback = callback,
+                        .block = block_id,
+                        .instruction = instruction_id,
+                    };
+                    instruction.operation.visitValueUses(&context, visitUse);
+
+                    if (context.failure) |failure|
+                        return failure;
+                }
+            }
+
+            if (block.terminator) |terminator| {
+                if (visitor.visitTerminator) |callback|
+                    try callback(visitor.context, block_id, terminator);
+
+                if (visitor.visitValueUse) |callback| {
+                    var context: UseContext = .{
+                        .visitor_context = visitor.context,
+                        .callback = callback,
+                        .block = block_id,
+                        .instruction = null,
+                    };
+                    module_ir.visitTerminatorValueUses(terminator, &context, visitUse);
+
+                    if (context.failure) |failure|
+                        return failure;
+                }
+            }
+        }
+    }
+}
+
+fn visitUse(context: *UseContext, value: ids.ValueId) void {
+    if (context.failure != null)
+        return;
+
+    context.callback(context.visitor_context, context.block, context.instruction, value) catch |err| {
+        context.failure = err;
+    };
+}

src/compiler/root.zig

@@ -0,0 +1,826 @@
+//! ## Shader compiler infrastructure.
+//!
+//! This module exposes the project-specific intermediate representation in
+//! `ir` and the SPIR-V frontend in `spirv`.
+//!
+//! Together they form
+//! the first stage of the compiler pipeline: SPIR-V binary modules are decoded,
+//! translated into a smaller and easier-to-transform IR, validated, and then made
+//! available to later optimization or code-generation passes.
+
+const std = @import("std");
+
+pub const ir = @import("ir/ir.zig");
+pub const spirv = @import("spirv/root.zig");
+
+const VisitorStatistics = struct {
+    functions: usize = 0,
+    blocks: usize = 0,
+};
+
+test "IR builder generation" {
+    // shader vertex @main
+    // {
+    //     @color: vec4[f32] = input[location(0), component(0), index(0)]
+    //     @out_color: vec4[f32] = output[location(0), component(0), index(0)]
+    //     %0: constant bool = true
+    //     %1: constant f32 = bits(0x3f800000)
+    //
+    //     fn @main() -> void
+    //     {
+    //         .entry():
+    //             %3: vec4[f32] = load_interface @color
+    //             conditional_branch %0, .pass(), .merge(%3)
+    //
+    //         .pass():
+    //             %4: vec4[f32] = composite_construct %1, %1, %1, %1
+    //             branch .merge(%4)
+    //
+    //         .merge(%2: vec4[f32]):
+    //             store_interface @out_color, %2
+    //             return
+    //     }
+    // }
+
+    var module = ir.module.Module.init(std.testing.allocator, .vertex);
+    defer module.deinit();
+    var builder = ir.Builder.init(&module);
+
+    const void_type = try builder.internType(.void);
+    const bool_type = try builder.internType(.boolean);
+    const f32_type = try builder.internType(.{ .floating = .{ .bits = 32 } });
+    const duplicate_f32 = try builder.internType(.{ .floating = .{ .bits = 32 } });
+    try std.testing.expectEqual(f32_type, duplicate_f32);
+    const vec4_type = try builder.internType(.{ .vector = .{ .element_type = f32_type, .length = 4 } });
+
+    const true_value = try builder.internConstant(bool_type, .{ .boolean = true });
+    const one = try builder.internConstant(f32_type, .{ .float_bits = @as(u32, @bitCast(@as(f32, 1.0))) });
+
+    const input = try builder.addInterfaceVariable(vec4_type, .input, .{ .location = .{ .location = 0 } }, "color");
+    const output = try builder.addInterfaceVariable(vec4_type, .output, .{ .location = .{ .location = 0 } }, "out_color");
+    const main = try builder.addFunction(void_type, "main");
+    builder.setEntryPoint(main);
+    const entry = try builder.addBlock(main, "entry");
+    const pass = try builder.addBlock(main, "pass");
+    const merge = try builder.addBlock(main, "merge");
+    const merged = try builder.addBlockParameter(merge, vec4_type, "merged");
+
+    const loaded = (try builder.appendInstruction(entry, vec4_type, .{
+        .load_interface = .{ .variable = input },
+    }, "loaded")).?;
+    try builder.setTerminator(entry, .{ .conditional_branch = .{
+        .condition = true_value,
+        .true_edge = try builder.edge(pass, &.{}),
+        .false_edge = try builder.edge(merge, &.{loaded}),
+    } });
+
+    const splat = (try builder.appendInstruction(pass, vec4_type, .{
+        .composite_construct = .{ .elements = &.{ one, one, one, one } },
+    }, "white")).?;
+    try builder.setTerminator(pass, .{ .branch = try builder.edge(merge, &.{splat}) });
+    _ = try builder.appendInstruction(merge, null, .{
+        .store_interface = .{ .variable = output, .value = merged },
+    }, null);
+    try builder.setTerminator(merge, .return_void);
+
+    try ir.validator.validate(&module);
+
+    var control_flow = try ir.cfg.init(std.testing.allocator, &module, main);
+    defer control_flow.deinit();
+    try std.testing.expectEqual(@as(usize, 2), control_flow.predecessors(merge).?.len);
+    try std.testing.expect(control_flow.dominates(entry, merge));
+    try std.testing.expect(!control_flow.dominates(pass, merge));
+
+    const text = try ir.printer.allocPrint(std.testing.allocator, &module);
+
+    defer std.testing.allocator.free(text);
+    try std.testing.expect(std.mem.indexOf(u8, text, "shader vertex @main") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "@color: vec4[f32] = input[location(0), component(0), index(0)]") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "@out_color: vec4[f32] = output[location(0), component(0), index(0)]") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "conditional_branch %0, .pass(), .merge(%loaded)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, ".merge(%merged: vec4[f32])") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "store_interface @out_color, %merged") != null);
+
+    var parsed = try ir.parser.parseString(std.testing.allocator, text);
+    defer parsed.deinit();
+    const round_trip = try ir.printer.allocPrint(std.testing.allocator, &parsed);
+    defer std.testing.allocator.free(round_trip);
+    try std.testing.expectEqualStrings(text, round_trip);
+
+    const io = std.Options.debug_io;
+    const path = ".zig-cache/ir-parser-round-trip.ir";
+    const file = try std.Io.Dir.cwd().createFile(io, path, .{ .truncate = true });
+    {
+        defer file.close(io);
+        var file_buffer: [4096]u8 = @splat(0);
+        var file_writer = file.writer(io, &file_buffer);
+        try file_writer.interface.writeAll(text);
+        try file_writer.interface.flush();
+    }
+    defer std.Io.Dir.cwd().deleteFile(io, path) catch @panic("Caught an error while handling an error");
+
+    var parsed_file = try ir.parser.parseFile(std.testing.allocator, io, path);
+    defer parsed_file.deinit();
+    const file_round_trip = try ir.printer.allocPrint(std.testing.allocator, &parsed_file);
+    defer std.testing.allocator.free(file_round_trip);
+    try std.testing.expectEqualStrings(text, file_round_trip);
+}
+
+test "IR parse interface" {
+    const source =
+        \\ shader vertex @main
+        \\ {
+        \\     @in_color: vec4[f32] = input[location(0), component(0), index(0)]
+        \\     @out_color: vec4[f32] = output[location(0), component(0), index(0)]
+        \\     @position: vec4[f32] = output[builtin(position)]
+        \\
+        \\     fn @main() -> void
+        \\     {
+        \\         .entry():
+        \\             return
+        \\     }
+        \\ }
+    ;
+
+    var module = try ir.parser.parseString(std.testing.allocator, source);
+    defer module.deinit();
+    const printed = try ir.printer.allocPrint(std.testing.allocator, &module);
+    defer std.testing.allocator.free(printed);
+
+    try std.testing.expect(std.mem.indexOf(u8, printed, "@in_color: vec4[f32] = input[location(0), component(0), index(0)]") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "@out_color: vec4[f32] = output[location(0), component(0), index(0)]") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "@position: vec4[f32] = output[builtin(position)]") != null);
+}
+
+test "IR parse types, operations, calls, terminators" {
+    const source =
+        \\ shader fragment @main
+        \\ {
+        \\     %0: constant bool = true
+        \\     %1: constant u32 = bits(0x1)
+        \\     %2: constant u32 = bits(0x2)
+        \\     %3: constant f32 = bits(0x3f800000)
+        \\     %4: constant array[u32, 2] = [#1, #2]
+        \\     %5: constant struct[u32, u32] = [#1, #2]
+        \\     %6: constant ptr[private, u32] = null
+        \\     %7: constant resourceHandle[sampler] = null
+        \\
+        \\     fn @main() -> void
+        \\     {
+        \\         .entry():
+        \\             %9: u32 = bitwise_not %1
+        \\             %10: u32 = integer_add %9, %2
+        \\             %11: bool = cmp_equal %1, %2
+        \\             %12: u32 = select %11, %1, %2
+        \\             %13: u32 = bitcast %12
+        \\             %14: vec2[u32] = composite_construct %1, %2
+        \\             %15: u32 = composite_extract %14[0]
+        \\             %16: f32 = negate %3
+        \\             %17: f32 = float_add %3, %16
+        \\             %18: u32 = call @helper(%15)
+        \\             return
+        \\     }
+        \\
+        \\     fn @helper(%8: u32) -> u32
+        \\     {
+        \\         .entry():
+        \\             return %8
+        \\     }
+        \\
+        \\     fn @discarder() -> void
+        \\     {
+        \\         .entry():
+        \\             discard
+        \\     }
+        \\
+        \\     fn @dead() -> void
+        \\     {
+        \\         .entry():
+        \\             unreachable
+        \\     }
+        \\ }
+    ;
+
+    var module = try ir.parser.parseString(std.testing.allocator, source);
+    defer module.deinit();
+    const printed = try ir.printer.allocPrint(std.testing.allocator, &module);
+    defer std.testing.allocator.free(printed);
+
+    var reparsed = try ir.parser.parseString(std.testing.allocator, printed);
+    defer reparsed.deinit();
+    const printed_again = try ir.printer.allocPrint(std.testing.allocator, &reparsed);
+    defer std.testing.allocator.free(printed_again);
+    try std.testing.expectEqualStrings(printed, printed_again);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "cmp_equal %1, %2") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "cmp.") == null);
+}
+
+test "IR parse named value IDs" {
+    const source =
+        \\ shader compute @main
+        \\ {
+        \\     %one_value: constant u32 = bits(0x1)
+        \\
+        \\     fn @main() -> void
+        \\     {
+        \\         .entry():
+        \\             %sum_value: u32 = integer_add %one_value, %one_value
+        \\             branch .merge(%sum_value)
+        \\
+        \\         .merge(%merged_value: u32):
+        \\             %product_value: u32 = integer_multiply %merged_value, %one_value
+        \\             return
+        \\     }
+        \\ }
+    ;
+
+    var module = try ir.parser.parseString(std.testing.allocator, source);
+    defer module.deinit();
+    const printed = try ir.printer.allocPrint(std.testing.allocator, &module);
+    defer std.testing.allocator.free(printed);
+
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%one_value: constant u32 = bits(0x1)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%sum_value: u32 = integer_add %one_value, %one_value") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "branch .merge(%sum_value)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, ".merge(%merged_value: u32)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%product_value: u32 = integer_multiply %merged_value, %one_value") != null);
+
+    var reparsed = try ir.parser.parseString(std.testing.allocator, printed);
+    defer reparsed.deinit();
+    const printed_again = try ir.printer.allocPrint(std.testing.allocator, &reparsed);
+    defer std.testing.allocator.free(printed_again);
+    try std.testing.expectEqualStrings(printed, printed_again);
+}
+
+test "IR parse numeric constants" {
+    const source =
+        \\ shader compute @main
+        \\ {
+        \\     %0: constant u8 = 255
+        \\     %1: constant i8 = -1
+        \\     %2: constant f16 = 1.5
+        \\     %3: constant f32 = -0.0
+        \\     %4: constant f64 = 2.5e0
+        \\
+        \\     fn @main() -> void
+        \\     {
+        \\         .entry():
+        \\             return
+        \\     }
+        \\ }
+    ;
+
+    var module = try ir.parser.parseString(std.testing.allocator, source);
+    defer module.deinit();
+    const printed = try ir.printer.allocPrint(std.testing.allocator, &module);
+    defer std.testing.allocator.free(printed);
+
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%0: constant u8 = bits(0xff)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%1: constant i8 = bits(0xff)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%2: constant f16 = bits(0x3e00)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%3: constant f32 = bits(0x80000000)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, printed, "%4: constant f64 = bits(0x4004000000000000)") != null);
+
+    const out_of_range =
+        \\ shader compute @main
+        \\ {
+        \\     %0: constant u8 = 256
+        \\     fn @main() -> void
+        \\     {
+        \\         .entry():
+        \\             return
+        \\     }
+        \\ }
+    ;
+    try std.testing.expectError(error.InvalidNumber, ir.parser.parseString(std.testing.allocator, out_of_range));
+}
+
+test "IR parser error: unknown value" {
+    const source =
+        \\ shader compute @main
+        \\ {
+        \\     fn @main() -> void
+        \\     {
+        \\         .entry():
+        \\             return %99
+        \\     }
+        \\ }
+    ;
+    try std.testing.expectError(error.UnknownValue, ir.parser.parseString(std.testing.allocator, source));
+}
+
+test "Validator error: wrong block argument count" {
+    // shader compute @main
+    // {
+    //     fn @main() -> void
+    //     {
+    //         .entry():
+    //             branch .merge()
+    //
+    //         .merge(%0: u32):
+    //             return
+    //     }
+    // }
+
+    var module = ir.module.Module.init(std.testing.allocator, .compute);
+    defer module.deinit();
+    var builder = ir.Builder.init(&module);
+
+    const void_type = try builder.internType(.void);
+    const u32_type = try builder.internType(.{ .integer = .{ .bits = 32, .signedness = .unsigned } });
+    const main = try builder.addFunction(void_type, "main");
+    builder.setEntryPoint(main);
+    const entry = try builder.addBlock(main, "entry");
+    const merge = try builder.addBlock(main, "merge");
+    _ = try builder.addBlockParameter(merge, u32_type, null);
+    try builder.setTerminator(entry, .{ .branch = try builder.edge(merge, &.{}) });
+    try builder.setTerminator(merge, .return_void);
+
+    try std.testing.expectError(error.WrongBranchArgumentCount, ir.validator.validate(&module));
+}
+
+test "Central store IDs disposal" {
+    var module = ir.module.Module.init(std.testing.allocator, .fragment);
+    defer module.deinit();
+    const first = try module.internType(.boolean);
+    try std.testing.expect(module.types.remove(first));
+    const second = try module.internType(.boolean);
+    try std.testing.expect(first.index() != second.index());
+    try std.testing.expect(module.types.get(first) == null);
+}
+
+test "Validator error: SSA definition does not dominate its use" {
+    // shader compute @main
+    // {
+    //     %0: constant bool = true
+    //     %1: constant u32 = bits(0x1)
+    //
+    //     fn @main() -> void
+    //     {
+    //         .entry():
+    //             conditional_branch %0, .left(), .right()
+    //
+    //         .left():
+    //             %2: u32 = integer_add %1, %1
+    //             branch .merge()
+    //
+    //         .right():
+    //             branch .merge()
+    //
+    //         .merge():
+    //             %3: u32 = integer_multiply %2, %1
+    //             return
+    //     }
+    // }
+
+    var module = ir.module.Module.init(std.testing.allocator, .compute);
+    defer module.deinit();
+    var builder = ir.Builder.init(&module);
+
+    const void_type = try builder.internType(.void);
+    const bool_type = try builder.internType(.boolean);
+    const u32_type = try builder.internType(.{ .integer = .{ .bits = 32, .signedness = .unsigned } });
+    const condition = try builder.internConstant(bool_type, .{ .boolean = true });
+    const one = try builder.internConstant(u32_type, .{ .integer_bits = 1 });
+    const main = try builder.addFunction(void_type, "main");
+    builder.setEntryPoint(main);
+    const entry = try builder.addBlock(main, "entry");
+    const left = try builder.addBlock(main, "left");
+    const right = try builder.addBlock(main, "right");
+    const merge = try builder.addBlock(main, "merge");
+
+    try builder.setTerminator(
+        entry,
+        .{
+            .conditional_branch = .{
+                .condition = condition,
+                .true_edge = try builder.edge(left, &.{}),
+                .false_edge = try builder.edge(right, &.{}),
+            },
+        },
+    );
+
+    const left_value = (try builder.appendInstruction(left, u32_type, .{
+        .binary = .{
+            .opcode = .integer_add,
+            .lhs = one,
+            .rhs = one,
+        },
+    }, null)).?;
+
+    try builder.setTerminator(left, .{ .branch = try builder.edge(merge, &.{}) });
+    try builder.setTerminator(right, .{ .branch = try builder.edge(merge, &.{}) });
+
+    _ = try builder.appendInstruction(merge, u32_type, .{
+        .binary = .{
+            .opcode = .integer_multiply,
+            .lhs = left_value,
+            .rhs = one,
+        },
+    }, null);
+
+    try builder.setTerminator(merge, .return_void);
+
+    try std.testing.expectError(error.DefinitionDoesNotDominateUse, ir.validator.validate(&module));
+}
+
+test "Rewriter replace all ID uses, safely erase dead instruction" {
+    // shader compute @main
+    // {
+    //     %0: constant u32 = bits(0x1)
+    //     %1: constant u32 = bits(0x2)
+    //
+    //     fn @main() -> void
+    //     {
+    //         .entry():
+    //             %2: u32 = integer_add %0, %1
+    //             %3: u32 = integer_multiply %2, %1
+    //             return
+    //     }
+    // }
+
+    var module = ir.module.Module.init(std.testing.allocator, .compute);
+    defer module.deinit();
+
+    var builder = ir.Builder.init(&module);
+
+    const void_type = try builder.internType(.void);
+    const u32_type = try builder.internType(.{ .integer = .{ .bits = 32, .signedness = .unsigned } });
+
+    const one = try builder.internConstant(u32_type, .{ .integer_bits = 1 });
+    const two = try builder.internConstant(u32_type, .{ .integer_bits = 2 });
+
+    const main = try builder.addFunction(void_type, "main");
+    builder.setEntryPoint(main);
+
+    const entry = try builder.addBlock(main, "entry");
+    const sum = (try builder.appendInstruction(entry, u32_type, .{
+        .binary = .{
+            .opcode = .integer_add,
+            .lhs = one,
+            .rhs = two,
+        },
+    }, null)).?;
+    _ = try builder.appendInstruction(entry, u32_type, .{
+        .binary = .{
+            .opcode = .integer_multiply,
+            .lhs = sum,
+            .rhs = two,
+        },
+    }, null);
+    try builder.setTerminator(entry, .return_void);
+
+    try ir.validator.validate(&module);
+
+    const sum_instruction = module.values.get(sum).?.definition.instruction;
+    var rewriter = ir.Rewriter.init(&module);
+
+    try std.testing.expectEqual(@as(usize, 1), try rewriter.replaceAllUses(sum, one));
+    try rewriter.eraseInstruction(sum_instruction);
+
+    try std.testing.expect(module.values.get(sum) == null);
+    try std.testing.expect(module.instructions.get(sum_instruction) == null);
+
+    try ir.validator.validate(&module);
+}
+
+test "Rewriter add block parameter and sync branch calls" {
+    // shader compute @main
+    // {
+    //     %0: constant u32 = bits(0x1)
+    //
+    //     fn @main() -> void
+    //     {
+    //         .entry():
+    //             branch .merge()
+    //
+    //         .merge():
+    //             return
+    //
+    //         .alternate():
+    //             return
+    //     }
+    // }
+
+    var module = ir.module.Module.init(std.testing.allocator, .compute);
+    defer module.deinit();
+
+    var builder = ir.Builder.init(&module);
+
+    const void_type = try builder.internType(.void);
+    const u32_type = try builder.internType(.{ .integer = .{ .bits = 32, .signedness = .unsigned } });
+
+    const one = try builder.internConstant(u32_type, .{ .integer_bits = 1 });
+
+    const main = try builder.addFunction(void_type, "main");
+    builder.setEntryPoint(main);
+
+    const entry = try builder.addBlock(main, "entry");
+    const merge = try builder.addBlock(main, "merge");
+    const alternate = try builder.addBlock(main, "alternate");
+
+    try builder.setTerminator(entry, .{ .branch = try builder.edge(merge, &.{}) });
+    try builder.setTerminator(merge, .return_void);
+    try builder.setTerminator(alternate, .return_void);
+
+    var rewriter = ir.Rewriter.init(&module);
+
+    const parameter = try rewriter.addBlockParameter(merge, u32_type, "incoming", &.{
+        .{
+            .predecessor = entry,
+            .value = one,
+        },
+    });
+    const merge_edge = module.blocks.get(entry).?.terminator.?.branch;
+    try std.testing.expectEqualSlices(ir.id.ValueId, &.{one}, merge_edge.arguments);
+
+    _ = try builder.appendInstruction(merge, u32_type, .{
+        .binary = .{
+            .opcode = .integer_add,
+            .lhs = parameter,
+            .rhs = one,
+        },
+    }, null);
+    try ir.validator.validate(&module);
+
+    try rewriter.removeBlockParameter(merge, 0, one);
+
+    try std.testing.expectEqual(@as(usize, 0), module.blocks.get(merge).?.parameters.items.len);
+    try std.testing.expectEqual(@as(usize, 0), module.blocks.get(entry).?.terminator.?.branch.arguments.len);
+
+    try ir.validator.validate(&module);
+
+    try std.testing.expectEqual(@as(usize, 1), try rewriter.redirectEdges(entry, merge, alternate, &.{}));
+    try std.testing.expectEqual(alternate, module.blocks.get(entry).?.terminator.?.branch.target);
+
+    try ir.validator.validate(&module);
+}
+
+fn establishNoCalls(_: *ir.module.Module, _: *ir.pass_manager.Context) !bool {
+    return false;
+}
+
+fn countVisitedFunction(context: ?*anyopaque, _: ir.id.FunctionId, _: *const ir.module.Function) !void {
+    const statistics: *VisitorStatistics = @ptrCast(@alignCast(context.?));
+    statistics.functions += 1;
+}
+
+fn countVisitedBlock(context: ?*anyopaque, _: ir.id.BlockId, _: *const ir.module.Block) !void {
+    const statistics: *VisitorStatistics = @ptrCast(@alignCast(context.?));
+    statistics.blocks += 1;
+}
+
+test "Pass manager track independent IR properties" {
+    // shader compute @main
+    // {
+    //     fn @main() -> void
+    //     {
+    //         .entry():
+    //             return
+    //     }
+    // }
+
+    var module = ir.module.Module.init(std.testing.allocator, .compute);
+    defer module.deinit();
+
+    var builder = ir.Builder.init(&module);
+
+    const void_type = try builder.internType(.void);
+
+    const main = try builder.addFunction(void_type, "main");
+    builder.setEntryPoint(main);
+
+    const entry = try builder.addBlock(main, "entry");
+    try builder.setTerminator(entry, .return_void);
+
+    module.properties.valid_cfg = true;
+
+    var manager = ir.pass_manager.Manager.init(std.testing.allocator);
+    defer manager.deinit();
+
+    try manager.add(.{
+        .name = "establish-no-calls",
+        .required = .{ .valid_cfg = true },
+        .produced = .{ .no_function_calls = true },
+        .run = establishNoCalls,
+    });
+
+    var context: ir.pass_manager.Context = .{ .allocator = std.testing.allocator };
+
+    try std.testing.expect(!try manager.run(&module, &context));
+    try std.testing.expect(module.properties.no_function_calls);
+
+    var statistics: VisitorStatistics = .{};
+
+    try ir.visitor.walk(&module, .{
+        .context = &statistics,
+        .visitFunction = countVisitedFunction,
+        .visitBlock = countVisitedBlock,
+    });
+
+    try std.testing.expectEqual(@as(usize, 1), statistics.functions);
+    try std.testing.expectEqual(@as(usize, 1), statistics.blocks);
+}
+
+test "SPIR-V parser error: zero-word instruction" {
+    const words = [_]u32{
+        spirv.spec.magic_number,
+        0x0001_0000,
+        0,
+        2,
+        0,
+        instructionWord(.nop, 0),
+    };
+    try std.testing.expectError(error.ZeroWordInstruction, spirv.Parser.init(&words));
+
+    const truncated = [_]u32{
+        spirv.spec.magic_number,
+        0x0001_0000,
+        0,
+        2,
+        0,
+        instructionWord(.i_add, 5),
+        1,
+    };
+    try std.testing.expectError(error.TruncatedInstruction, spirv.Parser.init(&truncated));
+}
+
+test "SPIR-V structured branches and OpPhi to block parameters" {
+    const assembly =
+        \\ OpCapability Shader
+        \\ OpMemoryModel Logical GLSL450
+        \\ OpEntryPoint GLCompute %main "main"
+        \\ OpExecutionMode %main LocalSize 1 1 1
+        \\ OpName %main "main"
+        \\ OpName %entry "entry"
+        \\ OpName %true "true"
+        \\ OpName %one "one"
+        \\ OpName %then "then"
+        \\ OpName %then_value "then_value"
+        \\ OpName %else "else"
+        \\ OpName %else_value "else_value"
+        \\ OpName %merge "merge"
+        \\ OpName %merged "merged"
+        \\ OpName %product "product"
+        \\
+        \\ %void = OpTypeVoid
+        \\ %bool = OpTypeBool
+        \\ %uint = OpTypeInt 32 0
+        \\ %fn_void = OpTypeFunction %void
+        \\ %true = OpConstantTrue %bool
+        \\ %one = OpConstant %uint 1
+        \\
+        \\ %main = OpFunction %void None %fn_void
+        \\     %entry = OpLabel
+        \\     OpSelectionMerge %merge None
+        \\     OpBranchConditional %true %then %else
+        \\     %then = OpLabel
+        \\     %then_value = OpIAdd %uint %one %one
+        \\     OpBranch %merge
+        \\     %else = OpLabel
+        \\     %else_value = OpISub %uint %one %one
+        \\     OpBranch %merge
+        \\     %merge = OpLabel
+        \\     %merged = OpPhi %uint %then_value %then %else_value %else
+        \\     %product = OpIMul %uint %merged %one
+        \\     OpReturn
+        \\ OpFunctionEnd
+    ;
+    const words = try assembleSpirv(std.testing.allocator, assembly);
+    defer std.testing.allocator.free(words);
+
+    var module = try spirv.translator.translate(std.testing.allocator, words, .{ .entry_point = "main" });
+    defer module.deinit();
+
+    try std.testing.expectEqual(ir.module.Stage.compute, module.stage);
+    try std.testing.expectEqual([3]u32{ 1, 1, 1 }, module.execution_modes.workgroup_size.?);
+    try std.testing.expect(module.properties.valid_cfg);
+    try std.testing.expect(module.properties.valid_ssa);
+
+    const function = module.functions.get(module.entry_point.?).?;
+    try std.testing.expectEqual(@as(usize, 4), function.blocks.items.len);
+    const entry = module.blocks.get(function.blocks.items[0]).?;
+    try std.testing.expect(entry.structured_control == .selection);
+    const merge = module.blocks.get(function.blocks.items[3]).?;
+    try std.testing.expectEqual(@as(usize, 1), merge.parameters.items.len);
+    try std.testing.expectEqual(@as(usize, 1), merge.instructions.items.len);
+    const multiply = module.instructions.get(merge.instructions.items[0]).?;
+    try std.testing.expectEqual(ir.instruction.BinaryOpcode.integer_multiply, multiply.operation.binary.opcode);
+
+    const text = try ir.printer.allocPrint(std.testing.allocator, &module);
+    defer std.testing.allocator.free(text);
+
+    try std.testing.expect(std.mem.indexOf(u8, text, "%one: constant u32 = bits(0x1)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "%true: constant bool = true") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "conditional_branch %true, .then(), .else()") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "%then_value: u32 = integer_add %one, %one") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "branch .merge(%then_value)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "%else_value: u32 = integer_subtract %one, %one") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, ".merge(%merged: u32)") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "%product: u32 = integer_multiply %merged, %one") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "integerMultiply") == null);
+
+    var parsed = try ir.parser.parseString(std.testing.allocator, text);
+    defer parsed.deinit();
+    const round_trip = try ir.printer.allocPrint(std.testing.allocator, &parsed);
+    defer std.testing.allocator.free(round_trip);
+    try std.testing.expectEqualStrings(text, round_trip);
+}
+
+test "SPIR-V decorated vertex interfaces and load-store operations" {
+    const assembly =
+        \\ OpCapability Shader
+        \\ OpMemoryModel Logical GLSL450
+        \\ OpEntryPoint Vertex %main "main" %in_color %out_color
+        \\ OpName %in_color "in_color"
+        \\ OpName %out_color "out_color"
+        \\ OpDecorate %in_color Location 0
+        \\ OpDecorate %out_color Location 0
+        \\
+        \\ %void = OpTypeVoid
+        \\ %float = OpTypeFloat 32
+        \\ %vec4 = OpTypeVector %float 4
+        \\ %input_vec4 = OpTypePointer Input %vec4
+        \\ %output_vec4 = OpTypePointer Output %vec4
+        \\ %fn_void = OpTypeFunction %void
+        \\ %in_color = OpVariable %input_vec4 Input
+        \\ %out_color = OpVariable %output_vec4 Output
+        \\
+        \\ %main = OpFunction %void None %fn_void
+        \\     %entry = OpLabel
+        \\     %color = OpLoad %vec4 %in_color
+        \\     OpStore %out_color %color
+        \\     OpReturn
+        \\ OpFunctionEnd
+    ;
+    const words = try assembleSpirv(std.testing.allocator, assembly);
+    defer std.testing.allocator.free(words);
+
+    var module = try spirv.translator.translate(std.testing.allocator, words, .{ .entry_point = "main" });
+    defer module.deinit();
+    try std.testing.expectEqual(ir.module.Stage.vertex, module.stage);
+    try std.testing.expectEqual(@as(usize, 2), module.interface_variables.entries.items.len);
+
+    const text = try ir.printer.allocPrint(std.testing.allocator, &module);
+    defer std.testing.allocator.free(text);
+    try std.testing.expect(std.mem.indexOf(u8, text, "load_interface @in_color") != null);
+    try std.testing.expect(std.mem.indexOf(u8, text, "store_interface @out_color") != null);
+}
+
+fn instructionWord(opcode: spirv.spec.Opcode, word_count: u16) u32 {
+    return (@as(u32, word_count) << 16) | @intFromEnum(opcode);
+}
+
+fn assembleSpirv(allocator: std.mem.Allocator, assembly: []const u8) ![]u32 {
+    var io_backend: std.Io.Threaded = .init(allocator, .{});
+    defer io_backend.deinit();
+    const io = io_backend.io();
+
+    var child = try std.process.spawn(io, .{
+        .argv = &.{ "spirv-as", "--target-env", "spv1.0", "-o", "-", "-" },
+        .stdin = .pipe,
+        .stdout = .pipe,
+        .stderr = .pipe,
+    });
+    defer child.kill(io);
+
+    {
+        const stdin = child.stdin.?;
+        var stdin_writer = stdin.writer(io, &.{});
+        try stdin_writer.interface.writeAll(assembly);
+        try stdin_writer.interface.flush();
+        stdin.close(io);
+        child.stdin = null;
+    }
+
+    var stdout_buffer: [4096]u8 = undefined;
+    var stdout_reader = child.stdout.?.reader(io, &stdout_buffer);
+    const binary = try stdout_reader.interface.allocRemaining(allocator, .limited(1024 * 1024));
+    defer allocator.free(binary);
+
+    var stderr_buffer: [4096]u8 = undefined;
+    var stderr_reader = child.stderr.?.reader(io, &stderr_buffer);
+    const stderr = try stderr_reader.interface.allocRemaining(allocator, .limited(64 * 1024));
+    defer allocator.free(stderr);
+
+    const term = try child.wait(io);
+    switch (term) {
+        .exited => |code| if (code != 0) {
+            std.log.err("spirv-as failed:\n{s}", .{stderr});
+            return error.SpirvAssemblyFailed;
+        },
+        else => {
+            std.log.err("spirv-as terminated unexpectedly:\n{s}", .{stderr});
+            return error.SpirvAssemblyFailed;
+        },
+    }
+
+    if (binary.len % @sizeOf(u32) != 0) return error.InvalidSpirvBinaryLength;
+    const words = try allocator.alloc(u32, binary.len / @sizeOf(u32));
+    errdefer allocator.free(words);
+    for (words, 0..) |*word, index| {
+        const offset = index * @sizeOf(u32);
+        word.* = std.mem.readInt(u32, binary[offset..][0..4], .little);
+    }
+    return words;
+}

src/compiler/spirv/Parser.zig

@@ -0,0 +1,155 @@
+const spirv = @import("spirv.zig");
+
+const Self = @This();
+
+pub const Error = error{
+    HeaderTooShort,
+    InvalidMagic,
+    ByteSwappedModule,
+    InvalidVersion,
+    InvalidIdBound,
+    InvalidSchema,
+    ZeroWordInstruction,
+    TruncatedInstruction,
+    UnterminatedString,
+};
+
+pub const Header = struct {
+    version: u32,
+    generator: u32,
+    bound: u32,
+    schema: u32,
+
+    pub inline fn major(self: Header) u8 {
+        return @truncate(self.version >> 16);
+    }
+
+    pub inline fn minor(self: Header) u8 {
+        return @truncate(self.version >> 8);
+    }
+};
+
+pub const Instruction = struct {
+    opcode: spirv.Opcode,
+    operands: []const u32,
+    word_offset: usize,
+
+    pub fn operand(self: Instruction, index: usize) ?u32 {
+        return if (index < self.operands.len) self.operands[index] else null;
+    }
+};
+
+pub const Iterator = struct {
+    words: []const u32,
+    cursor: usize = spirv.header_word_count,
+
+    pub fn next(self: *Iterator) Error!?Instruction {
+        if (self.cursor == self.words.len)
+            return null;
+
+        const first_word = self.words[self.cursor];
+        const word_count: usize = first_word >> 16;
+
+        if (word_count == 0)
+            return error.ZeroWordInstruction;
+        if (word_count > self.words.len - self.cursor)
+            return error.TruncatedInstruction;
+
+        const instruction: Instruction = .{
+            .opcode = @enumFromInt(@as(u16, @truncate(first_word))),
+            .operands = self.words[self.cursor + 1 .. self.cursor + word_count],
+            .word_offset = self.cursor,
+        };
+        self.cursor += word_count;
+        return instruction;
+    }
+};
+
+words: []const u32,
+header: Header,
+
+pub fn init(words: []const u32) Error!Self {
+    if (words.len < spirv.header_word_count)
+        return error.HeaderTooShort;
+    if (words[0] == spirv.byte_swapped_magic_number)
+        return error.ByteSwappedModule;
+    if (words[0] != spirv.magic_number)
+        return error.InvalidMagic;
+
+    const header: Header = .{
+        .version = words[1],
+        .generator = words[2],
+        .bound = words[3],
+        .schema = words[4],
+    };
+
+    if (header.major() != 1 or header.minor() > 6 or (header.version & 0xff00_00ff) != 0)
+        return error.InvalidVersion;
+    if (header.bound == 0)
+        return error.InvalidIdBound;
+    if (header.schema != 0)
+        return error.InvalidSchema;
+
+    var self: Self = .{ .words = words, .header = header };
+    var instruction_iterator = self.iterator();
+    while (try instruction_iterator.next()) |_| {}
+    return self;
+}
+
+pub fn iterator(self: Self) Iterator {
+    return .{ .words = self.words };
+}
+
+pub fn literalStringWordCount(words: []const u32) Error!usize {
+    for (words, 0..) |word, word_index| {
+        inline for (0..4) |byte_index| {
+            if (@as(u8, @truncate(word >> (byte_index * 8))) == 0)
+                return word_index + 1;
+        }
+    }
+    return error.UnterminatedString;
+}
+
+pub fn literalStringEquals(words: []const u32, expected: []const u8) Error!bool {
+    var byte_cursor: usize = 0;
+    for (words) |word| {
+        inline for (0..4) |byte_index| {
+            const byte: u8 = @truncate(word >> (byte_index * 8));
+
+            if (byte == 0)
+                return byte_cursor == expected.len;
+            if (byte_cursor >= expected.len or byte != expected[byte_cursor])
+                return false;
+
+            byte_cursor += 1;
+        }
+    }
+    return error.UnterminatedString;
+}
+
+pub fn copyLiteralString(allocator: anytype, words: []const u32) ![]u8 {
+    const word_count = try literalStringWordCount(words);
+    var byte_count: usize = 0;
+    outer: for (words[0..word_count]) |word| {
+        inline for (0..4) |byte_index| {
+            if (@as(u8, @truncate(word >> (byte_index * 8))) == 0)
+                break :outer;
+
+            byte_count += 1;
+        }
+    }
+
+    const result = try allocator.alloc(u8, byte_count);
+    var cursor: usize = 0;
+    outer: for (words[0..word_count]) |word| {
+        inline for (0..4) |byte_index| {
+            const byte: u8 = @truncate(word >> (byte_index * 8));
+            if (byte == 0)
+                break :outer;
+
+            result[cursor] = byte;
+            cursor += 1;
+        }
+    }
+    return result;
+}

src/compiler/spirv/root.zig

@@ -0,0 +1,16 @@
+//! ## SPIR-V frontend
+//!
+//! This namespace contains the SPIR-V parser and translator used to import shader
+//! modules into the compiler IR.
+//!
+//! `Parser` validates the SPIR-V header and iterates over binary instructions.
+//! `spec` exposes a minimalistic SPIR-V header translation.
+//!
+//! The main entry point is `translator.translate`, which finds the requested entry
+//! point, maps its execution model to an IR shader stage, lowers supported types,
+//! constants, interfaces, instructions, and structured control flow, then validates
+//! the generated IR module.
+
+pub const Parser = @import("Parser.zig");
+pub const translator = @import("translator.zig");
+pub const spec = @import("spirv.zig");

src/compiler/spirv/spirv.zig

@@ -0,0 +1,169 @@
+pub const magic_number: u32 = 0x07230203;
+pub const byte_swapped_magic_number: u32 = 0x03022307;
+pub const header_word_count: usize = 5;
+
+pub const Opcode = enum(u32) {
+    nop = 0,
+    undef = 1,
+    name = 5,
+    member_name = 6,
+    string = 7,
+    line = 8,
+    extension = 10,
+    ext_inst_import = 11,
+    ext_inst = 12,
+    memory_model = 14,
+    entry_point = 15,
+    execution_mode = 16,
+    capability = 17,
+
+    type_void = 19,
+    type_bool = 20,
+    type_int = 21,
+    type_float = 22,
+    type_vector = 23,
+    type_matrix = 24,
+    type_image = 25,
+    type_sampler = 26,
+    type_sampled_image = 27,
+    type_array = 28,
+    type_runtime_array = 29,
+    type_struct = 30,
+    type_opaque = 31,
+    type_pointer = 32,
+    type_function = 33,
+
+    constant_true = 41,
+    constant_false = 42,
+    constant = 43,
+    constant_composite = 44,
+    constant_null = 46,
+    spec_constant_true = 48,
+    spec_constant_false = 49,
+    spec_constant = 50,
+    spec_constant_composite = 51,
+    spec_constant_op = 52,
+
+    function = 54,
+    function_parameter = 55,
+    function_end = 56,
+    function_call = 57,
+    variable = 59,
+    load = 61,
+    store = 62,
+    access_chain = 65,
+
+    decorate = 71,
+    member_decorate = 72,
+    vector_shuffle = 79,
+    composite_construct = 80,
+    composite_extract = 81,
+    composite_insert = 82,
+    copy_object = 83,
+
+    convert_f_to_u = 109,
+    convert_f_to_s = 110,
+    convert_s_to_f = 111,
+    convert_u_to_f = 112,
+    u_convert = 113,
+    s_convert = 114,
+    f_convert = 115,
+    bitcast = 124,
+
+    s_negate = 126,
+    f_negate = 127,
+    i_add = 128,
+    f_add = 129,
+    i_sub = 130,
+    f_sub = 131,
+    i_mul = 132,
+    f_mul = 133,
+    u_div = 134,
+    s_div = 135,
+    f_div = 136,
+    u_mod = 137,
+    s_rem = 138,
+    s_mod = 139,
+    f_rem = 140,
+    f_mod = 141,
+    shift_right_logical = 194,
+    shift_right_arithmetic = 195,
+    shift_left_logical = 196,
+    bitwise_or = 197,
+    bitwise_xor = 198,
+    bitwise_and = 199,
+    logical_equal = 164,
+    logical_not_equal = 165,
+    logical_or = 166,
+    logical_and = 167,
+    logical_not = 168,
+    select = 169,
+    i_equal = 170,
+    i_not_equal = 171,
+    u_less_than = 176,
+    s_less_than = 177,
+    f_ord_equal = 180,
+    f_unord_equal = 181,
+    f_ord_not_equal = 182,
+    f_unord_not_equal = 183,
+    f_ord_less_than = 184,
+    f_unord_less_than = 185,
+
+    phi = 245,
+    loop_merge = 246,
+    selection_merge = 247,
+    label = 248,
+    branch = 249,
+    branch_conditional = 250,
+    @"switch" = 251,
+    kill = 252,
+    return_ = 253,
+    return_value = 254,
+    @"unreachable" = 255,
+    no_line = 317,
+
+    _,
+};
+
+pub const ExecutionModel = enum(u32) {
+    vertex = 0,
+    tessellation_control = 1,
+    tessellation_evaluation = 2,
+    geometry = 3,
+    fragment = 4,
+    gl_compute = 5,
+    kernel = 6,
+    _,
+};
+
+pub const StorageClass = enum(u32) {
+    uniform_constant = 0,
+    input = 1,
+    uniform = 2,
+    output = 3,
+    workgroup = 4,
+    cross_workgroup = 5,
+    private = 6,
+    function = 7,
+    generic = 8,
+    push_constant = 9,
+    atomic_counter = 10,
+    image = 11,
+    storage_buffer = 12,
+    physical_storage_buffer = 5349,
+    _,
+};
+
+pub const ExecutionMode = enum(u32) {
+    early_fragment_tests = 9,
+    local_size = 17,
+    _,
+};
+
+pub const Decoration = enum(u32) {
+    built_in = 11,
+    location = 30,
+    component = 31,
+    index = 32,
+    _,
+};

src/intel/FlintBinarySemaphore.zig

@@ -1,20 +1,70 @@
 const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
+const kmd = @import("kmd.zig");
 
 const VkError = base.VkError;
-const Device = base.Device;
+const FlintDevice = @import("FlintDevice.zig");
+
+const drm_syncobj_create = 0xbf;
+const drm_syncobj_destroy = 0xc0;
+const drm_syncobj_wait = 0xc3;
+const drm_syncobj_reset = 0xc4;
+const drm_syncobj_signal = 0xc5;
+
+const syncobj_wait_all: u32 = 1 << 0;
+const syncobj_wait_for_submit: u32 = 1 << 1;
+
+const SyncObjCreate = extern struct {
+    handle: u32,
+    flags: u32,
+};
+
+const SyncObjDestroy = extern struct {
+    handle: u32,
+    pad: u32,
+};
+
+const SyncObjWait = extern struct {
+    handles: u64,
+    timeout_nsec: i64,
+    count_handles: u32,
+    flags: u32,
+    first_signaled: u32,
+    pad: u32,
+    deadline_nsec: u64,
+};
+
+const SyncObjArray = extern struct {
+    handles: u64,
+    count_handles: u32,
+    pad: u32,
+};
 
 const Self = @This();
 pub const Interface = base.BinarySemaphore;
 
 interface: Interface,
+handle: u32,
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.SemaphoreCreateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(device, allocator, info);
+    const flint_device: *FlintDevice = @alignCast(@fieldParentPtr("interface", device));
+    var create_info = SyncObjCreate{
+        .handle = 0,
+        .flags = 0,
+    };
+
+    base.utils.ioctl(
+        try flint_device.kmd.file(),
+        device.io(),
+        kmd.drmIoctlIowr(drm_syncobj_create, SyncObjCreate),
+        &create_info,
+    ) catch return VkError.DeviceLost;
+    errdefer destroyHandle(flint_device, device.io(), create_info.handle);
 
     interface.vtable = &.{
         .destroy = destroy,
@@ -24,21 +74,89 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
 
     self.* = .{
         .interface = interface,
+        .handle = create_info.handle,
     };
     return self;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    destroyHandle(device, interface.owner.io(), self.handle);
     allocator.destroy(self);
 }
 
 pub fn signal(interface: *Interface) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var signal_info = SyncObjArray{
+        .handles = @intFromPtr(&handles),
+        .count_handles = handles.len,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_signal, SyncObjArray),
+        &signal_info,
+    ) catch return VkError.DeviceLost;
 }
 
 pub fn wait(interface: *Interface) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var wait_info = SyncObjWait{
+        .handles = @intFromPtr(&handles),
+        .timeout_nsec = std.math.maxInt(i64),
+        .count_handles = handles.len,
+        .flags = syncobj_wait_all | syncobj_wait_for_submit,
+        .first_signaled = 0,
+        .pad = 0,
+        .deadline_nsec = 0,
+    };
+
+    const errno = base.utils.ioctlErrno(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_wait, SyncObjWait),
+        &wait_info,
+    ) catch return VkError.DeviceLost;
+
+    switch (errno) {
+        .SUCCESS => {},
+        else => return VkError.DeviceLost,
+    }
+    try reset(interface);
+}
+
+pub fn reset(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var reset_info = SyncObjArray{
+        .handles = @intFromPtr(&handles),
+        .count_handles = handles.len,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_reset, SyncObjArray),
+        &reset_info,
+    ) catch return VkError.DeviceLost;
+}
+
+fn destroyHandle(device: *FlintDevice, io: std.Io, handle: u32) void {
+    var destroy_info = SyncObjDestroy{
+        .handle = handle,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        device.kmd.file() catch return,
+        io,
+        kmd.drmIoctlIowr(drm_syncobj_destroy, SyncObjDestroy),
+        &destroy_info,
+    ) catch @panic("Caught an error while handling an error");
 }

src/intel/FlintBuffer.zig

@@ -2,10 +2,8 @@ const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
 
-const lib = @import("lib.zig");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.Buffer;
@@ -17,6 +15,8 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(device, allocator, info);
+    interface.allowed_memory_types = std.bit_set.IntegerBitSet(32).initEmpty();
+    interface.allowed_memory_types.set(0);
 
     interface.vtable = &.{
         .destroy = destroy,
@@ -35,6 +35,14 @@ pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
 }
 
 pub fn getMemoryRequirements(interface: *Interface, requirements: *vk.MemoryRequirements) void {
-    _ = interface;
-    _ = requirements;
+    requirements.alignment = 16;
+    if (interface.usage.uniform_texel_buffer_bit or interface.usage.storage_texel_buffer_bit) {
+        requirements.alignment = @max(requirements.alignment, interface.owner.physical_device.props.limits.min_texel_buffer_offset_alignment);
+    }
+    if (interface.usage.storage_buffer_bit) {
+        requirements.alignment = @max(requirements.alignment, interface.owner.physical_device.props.limits.min_storage_buffer_offset_alignment);
+    }
+    if (interface.usage.uniform_buffer_bit) {
+        requirements.alignment = @max(requirements.alignment, interface.owner.physical_device.props.limits.min_uniform_buffer_offset_alignment);
+    }
 }

src/intel/FlintBufferView.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.BufferView;

src/intel/FlintCommandBuffer.zig

@@ -1,13 +1,26 @@
 const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
+const kmd = @import("kmd.zig");
 
 const VkError = base.VkError;
+const FlintDevice = @import("FlintDevice.zig");
+const FlintDescriptorSet = @import("FlintDescriptorSet.zig");
+const FlintPipeline = @import("FlintPipeline.zig");
+
+const MemoryRange = @import("MemoryRange.zig");
+
+const copy = @import("copy_commands.zig");
+const blitter = @import("blitter.zig");
 
 const Self = @This();
 pub const Interface = base.CommandBuffer;
 
 interface: Interface,
+batch: std.ArrayList(u32),
+relocations: std.ArrayList(kmd.Relocation),
+bound_compute_pipeline: ?*FlintPipeline,
+bound_compute_descriptor_sets: [base.vulkan_max_descriptor_sets]?*FlintDescriptorSet,
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.CommandBufferAllocateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
@@ -53,7 +66,10 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
         .resolveImage = resolveImage,
         .setEvent = setEvent,
         .setBlendConstants = setBlendConstants,
+        .setDepthBias = setDepthBias,
+        .setDepthBounds = setDepthBounds,
         .setDeviceMask = setDeviceMask,
+        .setLineWidth = setLineWidth,
         .setScissor = setScissor,
         .setStencilCompareMask = setStencilCompareMask,
         .setStencilReference = setStencilReference,
@@ -61,20 +77,35 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
         .setViewport = setViewport,
         .updateBuffer = updateBuffer,
         .waitEvent = waitEvent,
+        .writeTimestamp = writeTimestamp,
     };
 
-    self.* = .{ .interface = interface };
+    self.* = .{
+        .interface = interface,
+        .batch = .empty,
+        .relocations = .empty,
+        .bound_compute_pipeline = null,
+        .bound_compute_descriptor_sets = @splat(null),
+    };
     return self;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const command_allocator = self.interface.host_allocator.allocator();
+    self.batch.deinit(command_allocator);
+    self.relocations.deinit(command_allocator);
     allocator.destroy(self);
 }
 
-pub fn execute(self: *Self) VkError!void {
+pub fn submitGpuBatch(self: *Self, syncs: []const kmd.SyncDependency) VkError!void {
     try self.interface.submit();
-    self.interface.finish() catch {};
+    defer self.interface.finish() catch @panic("Caught an error while handling an error");
+
+    // Empty command buffers still need a no-op submission to carry queue synchronization.
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", self.interface.owner));
+    const allocator = self.interface.host_allocator.allocator();
+    try device.kmd.submitBatch(self.interface.owner.io(), allocator, self.batch.items, self.relocations.items, syncs);
 }
 
 pub fn begin(interface: *Interface, info: *const vk.CommandBufferBeginInfo) VkError!void {
@@ -87,8 +118,34 @@ pub fn end(interface: *Interface) VkError!void {
 }
 
 pub fn reset(interface: *Interface, flags: vk.CommandBufferResetFlags) VkError!void {
-    _ = interface;
-    _ = flags;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    if (flags.release_resources_bit) {
+        const command_allocator = self.interface.host_allocator.allocator();
+        self.batch.clearAndFree(command_allocator);
+        self.relocations.clearAndFree(command_allocator);
+    } else {
+        self.batch.clearRetainingCapacity();
+        self.relocations.clearRetainingCapacity();
+    }
+    self.bound_compute_pipeline = null;
+    self.bound_compute_descriptor_sets = @splat(null);
+}
+
+pub fn emit(self: *Self, dword: u32) VkError!void {
+    self.batch.append(self.interface.host_allocator.allocator(), dword) catch return VkError.OutOfHostMemory;
+}
+
+pub fn emitRelocatedAddress(self: *Self, range: MemoryRange, read: bool, write: bool) VkError!void {
+    const address_offset = self.batch.items.len * @sizeOf(u32);
+    try self.emit(@intCast(range.offset));
+    try self.emit(0);
+    self.relocations.append(self.interface.host_allocator.allocator(), .{
+        .target_handle = try range.memory.allocation.handle(),
+        .offset = @intCast(address_offset),
+        .delta = @intCast(range.offset),
+        .read = read,
+        .write = write,
+    }) catch return VkError.OutOfHostMemory;
 }
 
 pub fn beginQuery(interface: *Interface, pool: *base.QueryPool, query: u32, flags: vk.QueryControlFlags) VkError!void {
@@ -115,18 +172,27 @@ pub fn beginRenderPass(interface: *Interface, render_pass: *base.RenderPass, fra
     _ = clear_values;
 }
 
-pub fn bindDescriptorSets(interface: *Interface, bind_point: vk.PipelineBindPoint, first_set: u32, sets: [base.VULKAN_MAX_DESCRIPTOR_SETS]?*base.DescriptorSet, dynamic_offsets: []const u32) VkError!void {
-    _ = interface;
-    _ = bind_point;
-    _ = first_set;
-    _ = sets;
+pub fn bindDescriptorSets(interface: *Interface, bind_point: vk.PipelineBindPoint, first_set: u32, sets: [base.vulkan_max_descriptor_sets]?*base.DescriptorSet, dynamic_offsets: []const u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    if (bind_point != .compute) return;
+    if (first_set >= base.vulkan_max_descriptor_sets) return VkError.ValidationFailed;
+
+    for (sets, 0..) |set, index| {
+        const base_set = set orelse break;
+        const destination = first_set + index;
+        if (destination >= base.vulkan_max_descriptor_sets) return VkError.ValidationFailed;
+        self.bound_compute_descriptor_sets[destination] = @alignCast(@fieldParentPtr("interface", base_set));
+    }
     _ = dynamic_offsets;
 }
 
 pub fn bindPipeline(interface: *Interface, bind_point: vk.PipelineBindPoint, pipeline: *base.Pipeline) VkError!void {
-    _ = interface;
-    _ = bind_point;
-    _ = pipeline;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    if (bind_point != .compute)
+        return;
+
+    const flint_pipeline: *FlintPipeline = @alignCast(@fieldParentPtr("interface", pipeline));
+    self.bound_compute_pipeline = flint_pipeline;
 }
 
 pub fn bindIndexBuffer(interface: *Interface, buffer: *base.Buffer, offset: usize, index_type: vk.IndexType) VkError!void {
@@ -144,13 +210,13 @@ pub fn bindVertexBuffer(interface: *Interface, index: usize, buffer: *base.Buffe
 }
 
 pub fn blitImage(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Image, dst_layout: vk.ImageLayout, regions: []const vk.ImageBlit, filter: vk.Filter) VkError!void {
-    _ = interface;
-    _ = src;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     _ = src_layout;
-    _ = dst;
     _ = dst_layout;
-    _ = regions;
-    _ = filter;
+    if (filter != .nearest) return VkError.FeatureNotPresent;
+
+    for (regions) |region|
+        try blitter.blitImageRegion(self, src, dst, region);
 }
 
 pub fn clearAttachment(interface: *Interface, attachment: vk.ClearAttachment, rect: vk.ClearRect) VkError!void {
@@ -176,35 +242,35 @@ pub fn clearDepthStencilImage(interface: *Interface, image: *base.Image, layout:
 }
 
 pub fn copyBuffer(interface: *Interface, src: *base.Buffer, dst: *base.Buffer, regions: []const vk.BufferCopy) VkError!void {
-    _ = interface;
-    _ = src;
-    _ = dst;
-    _ = regions;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+
+    for (regions) |region| {
+        const src_range = try copy.copyRangeFromRegion(src, region.src_offset, region.size);
+        const dst_range = try copy.copyRangeFromRegion(dst, region.dst_offset, region.size);
+        try copy.emitLinearCopy(self, src_range, dst_range);
+    }
 }
 
 pub fn copyBufferToImage(interface: *Interface, src: *base.Buffer, dst: *base.Image, dst_layout: vk.ImageLayout, regions: []const vk.BufferImageCopy) VkError!void {
-    _ = interface;
-    _ = src;
-    _ = dst;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     _ = dst_layout;
-    _ = regions;
+    for (regions) |region|
+        try copy.copyBufferImage(self, src, dst, region, true);
 }
 
 pub fn copyImage(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Image, dst_layout: vk.ImageLayout, regions: []const vk.ImageCopy) VkError!void {
-    _ = interface;
-    _ = src;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     _ = src_layout;
-    _ = dst;
     _ = dst_layout;
-    _ = regions;
+    for (regions) |region|
+        try copy.copyImage(self, src, dst, region);
 }
 
 pub fn copyImageToBuffer(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Buffer, regions: []const vk.BufferImageCopy) VkError!void {
-    _ = interface;
-    _ = src;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     _ = src_layout;
-    _ = dst;
-    _ = regions;
+    for (regions) |region|
+        try copy.copyBufferImage(self, dst, src, region, false);
 }
 
 pub fn copyQueryPoolResults(interface: *Interface, pool: *base.QueryPool, first: u32, count: u32, dst: *base.Buffer, offset: vk.DeviceSize, stride: vk.DeviceSize, flags: vk.QueryResultFlags) VkError!void {
@@ -219,10 +285,7 @@ pub fn copyQueryPoolResults(interface: *Interface, pool: *base.QueryPool, first:
 }
 
 pub fn dispatch(interface: *Interface, group_count_x: u32, group_count_y: u32, group_count_z: u32) VkError!void {
-    _ = interface;
-    _ = group_count_x;
-    _ = group_count_y;
-    _ = group_count_z;
+    try dispatchBase(interface, 0, 0, 0, group_count_x, group_count_y, group_count_z);
 }
 
 pub fn dispatchBase(interface: *Interface, base_group_x: u32, base_group_y: u32, base_group_z: u32, group_count_x: u32, group_count_y: u32, group_count_z: u32) VkError!void {
@@ -284,16 +347,37 @@ pub fn endRenderPass(interface: *Interface) VkError!void {
 }
 
 pub fn executeCommands(interface: *Interface, commands: *Interface) VkError!void {
-    _ = interface;
-    _ = commands;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const secondary: *Self = @alignCast(@fieldParentPtr("interface", commands));
+    const allocator = self.interface.host_allocator.allocator();
+    const relocation_offset = self.batch.items.len * @sizeOf(u32);
+
+    self.batch.appendSlice(allocator, secondary.batch.items) catch return VkError.OutOfHostMemory;
+    for (secondary.relocations.items) |relocation| {
+        self.relocations.append(allocator, .{
+            .target_handle = relocation.target_handle,
+            .offset = relocation.offset + relocation_offset,
+            .delta = relocation.delta,
+            .read = relocation.read,
+            .write = relocation.write,
+        }) catch return VkError.OutOfHostMemory;
+    }
 }
 
 pub fn fillBuffer(interface: *Interface, buffer: *base.Buffer, offset: vk.DeviceSize, size: vk.DeviceSize, data: u32) VkError!void {
-    _ = interface;
-    _ = buffer;
-    _ = offset;
-    _ = size;
-    _ = data;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const dst_range = try copy.fillRange(buffer, offset, size);
+
+    var filled: vk.DeviceSize = 0;
+    while (filled < dst_range.size) {
+        const dst_chunk: MemoryRange = .{ .memory = dst_range.memory, .offset = dst_range.offset + filled, .size = @sizeOf(u32) };
+
+        try self.emit(kmd.mi_store_data_imm_dword);
+        try self.emitRelocatedAddress(dst_chunk, false, true);
+        try self.emit(data);
+
+        filled += @sizeOf(u32);
+    }
 }
 
 pub fn updateBuffer(interface: *Interface, buffer: *base.Buffer, offset: vk.DeviceSize, data: []const u8) VkError!void {
@@ -363,6 +447,24 @@ pub fn setBlendConstants(interface: *Interface, constants: [4]f32) VkError!void
     _ = constants;
 }
 
+pub fn setDepthBias(interface: *Interface, constant_factor: f32, clamp: f32, slope_factor: f32) VkError!void {
+    _ = interface;
+    _ = constant_factor;
+    _ = clamp;
+    _ = slope_factor;
+}
+
+pub fn setDepthBounds(interface: *Interface, min: f32, max: f32) VkError!void {
+    _ = interface;
+    _ = min;
+    _ = max;
+}
+
+pub fn setLineWidth(interface: *Interface, width: f32) VkError!void {
+    _ = interface;
+    _ = width;
+}
+
 pub fn setStencilCompareMask(interface: *Interface, face_mask: vk.StencilFaceFlags, compare_mask: u32) VkError!void {
     _ = interface;
     _ = face_mask;
@@ -390,3 +492,9 @@ pub fn waitEvent(interface: *Interface, event: *base.Event, src_stage: vk.Pipeli
     _ = buffer_barriers;
     _ = image_barriers;
 }
+
+pub fn writeTimestamp(interface: *Interface, stage: vk.PipelineStageFlags, pool: *base.QueryPool, query: u32) VkError!void {
+    _ = interface;
+    _ = stage;
+    try pool.writeTimestamp(query, 0);
+}

src/intel/FlintCommandPool.zig

@@ -2,9 +2,7 @@ const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
 
-const NonDispatchable = base.NonDispatchable;
 const VkError = base.VkError;
-const Device = base.Device;
 
 const FlintCommandBuffer = @import("FlintCommandBuffer.zig");
 

src/intel/FlintDescriptorPool.zig

@@ -5,12 +5,13 @@ const base = @import("base");
 const VkError = base.VkError;
 const VulkanAllocator = base.VulkanAllocator;
 
-const Device = base.Device;
+const FlintDescriptorSet = @import("FlintDescriptorSet.zig");
 
 const Self = @This();
 pub const Interface = base.DescriptorPool;
 
 interface: Interface,
+sets: std.ArrayList(*FlintDescriptorSet),
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.DescriptorPoolCreateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
@@ -27,29 +28,42 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
 
     self.* = .{
         .interface = interface,
+        .sets = std.ArrayList(*FlintDescriptorSet).initCapacity(allocator, info.max_sets) catch return VkError.OutOfHostMemory,
     };
     return self;
 }
 
 pub fn allocateDescriptorSet(interface: *Interface, layout: *base.DescriptorSetLayout) VkError!*base.DescriptorSet {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
-    _ = layout;
-    return VkError.Unknown;
+    if (self.sets.items.len == self.sets.capacity) return VkError.OutOfPoolMemory;
+
+    const allocator = VulkanAllocator.init(null, .object).allocator();
+    const set = try FlintDescriptorSet.create(interface.owner, allocator, layout);
+    self.sets.appendAssumeCapacity(set);
+    return &set.interface;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const set_allocator = VulkanAllocator.init(null, .object).allocator();
+    for (self.sets.items) |set| set.interface.destroy(set_allocator);
+    self.sets.deinit(allocator);
     allocator.destroy(self);
 }
 
 pub fn freeDescriptorSet(interface: *Interface, set: *base.DescriptorSet) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
-    _ = set;
+    const flint_set: *FlintDescriptorSet = @alignCast(@fieldParentPtr("interface", set));
+    const index = std.mem.indexOfScalar(*FlintDescriptorSet, self.sets.items, flint_set) orelse return VkError.ValidationFailed;
+    _ = self.sets.orderedRemove(index);
+
+    const allocator = VulkanAllocator.init(null, .object).allocator();
+    set.destroy(allocator);
 }
 
 pub fn reset(interface: *Interface, _: vk.DescriptorPoolResetFlags) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const allocator = VulkanAllocator.init(null, .object).allocator();
+    for (self.sets.items) |set| set.interface.destroy(allocator);
+    self.sets.clearRetainingCapacity();
 }

src/intel/FlintDescriptorSet.zig

@@ -3,11 +3,25 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
+const NonDispatchable = base.NonDispatchable;
 
 const Self = @This();
 pub const Interface = base.DescriptorSet;
 
+pub const DescriptorBuffer = struct {
+    buffer: ?*base.Buffer,
+    offset: vk.DeviceSize,
+    size: vk.DeviceSize,
+};
+
+const Descriptor = union(enum) {
+    buffer: []DescriptorBuffer,
+    unsupported,
+};
+
 interface: Interface,
+heap: []u8,
+descriptors: []Descriptor,
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, layout: *base.DescriptorSetLayout) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
@@ -21,26 +35,139 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, layout: *base.
         .write = write,
     };
 
+    var heap_size = layout.bindings.len * @sizeOf(Descriptor);
+    for (layout.bindings) |binding| {
+        heap_size += switch (binding.descriptor_type) {
+            .uniform_buffer,
+            .uniform_buffer_dynamic,
+            .storage_buffer,
+            .storage_buffer_dynamic,
+            => binding.array_size * @sizeOf(DescriptorBuffer),
+
+            else => 0,
+        };
+    }
+
+    const heap = allocator.alloc(u8, heap_size) catch return VkError.OutOfHostMemory;
+    errdefer allocator.free(heap);
+
+    var fixed = std.heap.FixedBufferAllocator.init(heap);
+    const descriptors = fixed.allocator().alloc(Descriptor, layout.bindings.len) catch return VkError.OutOfHostMemory;
+    for (descriptors, layout.bindings) |*descriptor, binding| {
+        descriptor.* = switch (binding.descriptor_type) {
+            .uniform_buffer, .uniform_buffer_dynamic, .storage_buffer, .storage_buffer_dynamic => blk: {
+                const buffers = fixed.allocator().alloc(DescriptorBuffer, binding.array_size) catch return VkError.OutOfHostMemory;
+
+                for (buffers) |*buffer| {
+                    buffer.* = .{
+                        .buffer = null,
+                        .offset = 0,
+                        .size = 0,
+                    };
+                }
+
+                break :blk .{ .buffer = buffers };
+            },
+            else => .unsupported,
+        };
+    }
+
     self.* = .{
         .interface = interface,
+        .heap = heap,
+        .descriptors = descriptors,
     };
     return self;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.free(self.heap);
     allocator.destroy(self);
 }
 
 pub fn copy(interface: *Interface, src_interface: *const Interface, data: vk.CopyDescriptorSet) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
-    _ = src_interface;
-    _ = data;
+    const src: *const Self = @alignCast(@fieldParentPtr("interface", src_interface));
+
+    if (data.dst_binding >= self.descriptors.len or data.src_binding >= src.descriptors.len)
+        return VkError.ValidationFailed;
+
+    const dst = switch (self.descriptors[data.dst_binding]) {
+        .buffer => |buffers| buffers,
+        .unsupported => return VkError.FeatureNotPresent,
+    };
+
+    const source = switch (src.descriptors[data.src_binding]) {
+        .buffer => |buffers| buffers,
+        .unsupported => return VkError.FeatureNotPresent,
+    };
+
+    const dst_start: usize = @intCast(data.dst_array_element);
+    const src_start: usize = @intCast(data.src_array_element);
+    const count: usize = @intCast(data.descriptor_count);
+
+    if (dst_start > dst.len or count > dst.len - dst_start)
+        return VkError.ValidationFailed;
+    if (src_start > source.len or count > source.len - src_start)
+        return VkError.ValidationFailed;
+
+    @memcpy(dst[dst_start .. dst_start + count], source[src_start .. src_start + count]);
 }
 
 pub fn write(interface: *Interface, write_data: vk.WriteDescriptorSet) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
-    _ = write_data;
+
+    switch (write_data.descriptor_type) {
+        .uniform_buffer,
+        .uniform_buffer_dynamic,
+        .storage_buffer,
+        .storage_buffer_dynamic,
+        => {
+            if (write_data.dst_binding >= self.descriptors.len)
+                return VkError.ValidationFailed;
+
+            const descriptors = switch (self.descriptors[write_data.dst_binding]) {
+                .buffer => |buffers| buffers,
+                .unsupported => return VkError.FeatureNotPresent,
+            };
+
+            const start: usize = @intCast(write_data.dst_array_element);
+            const count: usize = @intCast(write_data.descriptor_count);
+
+            if (start > descriptors.len or count > descriptors.len - start)
+                return VkError.ValidationFailed;
+
+            for (write_data.p_buffer_info, 0..write_data.descriptor_count) |buffer_info, index| {
+                const descriptor = &descriptors[start + index];
+                descriptor.* = .{ .buffer = null, .offset = buffer_info.offset, .size = buffer_info.range };
+
+                if (buffer_info.buffer == .null_handle)
+                    continue;
+
+                const buffer = try NonDispatchable(base.Buffer).fromHandleObject(buffer_info.buffer);
+
+                if (descriptor.offset > buffer.size)
+                    return VkError.ValidationFailed;
+                if (descriptor.size == vk.WHOLE_SIZE)
+                    descriptor.size = buffer.size - descriptor.offset;
+                if (descriptor.size > buffer.size - descriptor.offset)
+                    return VkError.ValidationFailed;
+
+                descriptor.buffer = buffer;
+            }
+        },
+
+        else => return VkError.FeatureNotPresent,
+    }
+}
+
+pub fn getBuffer(self: *const Self, binding: u32, array_element: u32) VkError!DescriptorBuffer {
+    if (binding >= self.descriptors.len) return VkError.ValidationFailed;
+    const buffers = switch (self.descriptors[binding]) {
+        .buffer => |items| items,
+        .unsupported => return VkError.FeatureNotPresent,
+    };
+    if (array_element >= buffers.len) return VkError.ValidationFailed;
+    return buffers[array_element];
 }

src/intel/FlintDescriptorSetLayout.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.DescriptorSetLayout;

src/intel/FlintDevice.zig

@@ -3,6 +3,8 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const FlintQueue = @import("FlintQueue.zig");
+const FlintPhysicalDevice = @import("FlintPhysicalDevice.zig");
+const Kmd = @import("kmd.zig");
 
 pub const FlintBinarySemaphore = @import("FlintBinarySemaphore.zig");
 pub const FlintBuffer = @import("FlintBuffer.zig");
@@ -32,12 +34,16 @@ const Self = @This();
 pub const Interface = base.Device;
 
 interface: Interface,
+kmd: Kmd.Device,
 
 pub fn create(instance: *base.Instance, physical_device: *base.PhysicalDevice, allocator: std.mem.Allocator, info: *const vk.DeviceCreateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(allocator, instance, physical_device, info);
+    const flint_physical_device: *const FlintPhysicalDevice = @alignCast(@fieldParentPtr("interface", physical_device));
+    var kmd_device = try Kmd.Device.open(instance.io(), flint_physical_device);
+    errdefer kmd_device.close(instance.io());
 
     interface.vtable = &.{
         .createQueue = FlintQueue.create,
@@ -73,6 +79,7 @@ pub fn create(instance: *base.Instance, physical_device: *base.PhysicalDevice, a
 
     self.* = .{
         .interface = interface,
+        .kmd = kmd_device,
     };
 
     try self.interface.createQueues(allocator, info);
@@ -81,6 +88,7 @@ pub fn create(instance: *base.Instance, physical_device: *base.PhysicalDevice, a
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.kmd.close(interface.io());
     allocator.destroy(self);
 }
 

src/intel/FlintDeviceMemory.zig

@@ -1,9 +1,9 @@
 const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
-const lib = @import("lib.zig");
 
 const FlintDevice = @import("FlintDevice.zig");
+const kmd = @import("kmd.zig");
 
 const VkError = base.VkError;
 
@@ -11,12 +11,19 @@ const Self = @This();
 pub const Interface = base.DeviceMemory;
 
 interface: Interface,
+allocation: kmd.Memory,
 
 pub fn create(device: *FlintDevice, allocator: std.mem.Allocator, size: vk.DeviceSize, memory_type_index: u32) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
+    if (memory_type_index >= device.interface.physical_device.mem_props.memory_type_count) {
+        return VkError.ValidationFailed;
+    }
+
     var interface = try Interface.init(&device.interface, size, memory_type_index);
+    var allocation = try device.kmd.allocateMemory(device.interface.io(), size);
+    errdefer allocation.deinit(&device.kmd, device.interface.io());
 
     interface.vtable = &.{
         .destroy = destroy,
@@ -28,33 +35,42 @@ pub fn create(device: *FlintDevice, allocator: std.mem.Allocator, size: vk.Devic
 
     self.* = .{
         .interface = interface,
+        .allocation = allocation,
     };
     return self;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    self.allocation.deinit(&device.kmd, interface.owner.io());
     allocator.destroy(self);
 }
 
 pub fn flushRange(interface: *Interface, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
-    _ = interface;
-    _ = offset;
-    _ = size;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    try self.allocation.flushRange(&device.kmd, interface.owner.io(), offset, size);
 }
 
 pub fn invalidateRange(interface: *Interface, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
-    _ = interface;
-    _ = offset;
-    _ = size;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    try self.allocation.invalidateRange(&device.kmd, interface.owner.io(), offset, size);
 }
 
 pub fn map(interface: *Interface, offset: vk.DeviceSize, size: vk.DeviceSize) VkError![]u8 {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
-    _ = offset;
-    _ = size;
-    return VkError.Unknown;
+    if (interface.is_mapped) return VkError.MemoryMapFailed;
+
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    const data = try self.allocation.map(&device.kmd, interface.owner.io(), offset, size);
+    interface.is_mapped = true;
+    return data;
 }
 
-pub fn unmap(_: *Interface) void {}
+pub fn unmap(interface: *Interface) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.allocation.unmap();
+    interface.is_mapped = false;
+}

src/intel/FlintEvent.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.Event;

src/intel/FlintFence.zig

@@ -1,20 +1,73 @@
 const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
+const kmd = @import("kmd.zig");
 
 const VkError = base.VkError;
 const Device = base.Device;
+const FlintDevice = @import("FlintDevice.zig");
+
+const drm_syncobj_create = 0xbf;
+const drm_syncobj_destroy = 0xc0;
+const drm_syncobj_wait = 0xc3;
+const drm_syncobj_reset = 0xc4;
+const drm_syncobj_signal = 0xc5;
+
+const syncobj_create_signaled: u32 = 1 << 0;
+const syncobj_wait_all: u32 = 1 << 0;
+const syncobj_wait_for_submit: u32 = 1 << 1;
+
+const SyncObjCreate = extern struct {
+    handle: u32,
+    flags: u32,
+};
+
+const SyncObjDestroy = extern struct {
+    handle: u32,
+    pad: u32,
+};
+
+const SyncObjWait = extern struct {
+    handles: u64,
+    timeout_nsec: i64,
+    count_handles: u32,
+    flags: u32,
+    first_signaled: u32,
+    pad: u32,
+    deadline_nsec: u64,
+};
+
+const SyncObjArray = extern struct {
+    handles: u64,
+    count_handles: u32,
+    pad: u32,
+};
 
 const Self = @This();
 pub const Interface = base.Fence;
 
 interface: Interface,
+handle: u32,
 
 pub fn create(device: *Device, allocator: std.mem.Allocator, info: *const vk.FenceCreateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(device, allocator, info);
+    const flint_device: *FlintDevice = @alignCast(@fieldParentPtr("interface", device));
+    var create_info = SyncObjCreate{
+        .handle = 0,
+        .flags = if (info.flags.signaled_bit) syncobj_create_signaled else 0,
+    };
+
+    base.utils.ioctl(
+        try flint_device.kmd.file(),
+        device.io(),
+        kmd.drmIoctlIowr(drm_syncobj_create, SyncObjCreate),
+        &create_info,
+    ) catch return VkError.DeviceLost;
+
+    errdefer destroyHandle(flint_device, device.io(), create_info.handle);
 
     interface.vtable = &.{
         .destroy = destroy,
@@ -26,32 +79,104 @@ pub fn create(device: *Device, allocator: std.mem.Allocator, info: *const vk.Fen
 
     self.* = .{
         .interface = interface,
+        .handle = create_info.handle,
     };
     return self;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    destroyHandle(device, interface.owner.io(), self.handle);
     allocator.destroy(self);
 }
 
 pub fn getStatus(interface: *Interface) VkError!void {
-    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    wait(interface, 0) catch |err| switch (err) {
+        VkError.Timeout => return VkError.NotReady,
+        else => return err,
+    };
 }
 
 pub fn reset(interface: *Interface) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var reset_info = SyncObjArray{
+        .handles = @intFromPtr(&handles),
+        .count_handles = handles.len,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_reset, SyncObjArray),
+        &reset_info,
+    ) catch return VkError.DeviceLost;
 }
 
 pub fn signal(interface: *Interface) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var signal_info = SyncObjArray{
+        .handles = @intFromPtr(&handles),
+        .count_handles = handles.len,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_signal, SyncObjArray),
+        &signal_info,
+    ) catch return VkError.DeviceLost;
 }
 
 pub fn wait(interface: *Interface, timeout: u64) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
-    _ = self;
-    _ = timeout;
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    var handles = [_]u32{self.handle};
+    var wait_info = SyncObjWait{
+        .handles = @intFromPtr(&handles),
+        .timeout_nsec = absoluteTimeout(interface.owner.io(), timeout),
+        .count_handles = handles.len,
+        .flags = syncobj_wait_all | syncobj_wait_for_submit,
+        .first_signaled = 0,
+        .pad = 0,
+        .deadline_nsec = 0,
+    };
+
+    const errno = base.utils.ioctlErrno(
+        try device.kmd.file(),
+        interface.owner.io(),
+        kmd.drmIoctlIowr(drm_syncobj_wait, SyncObjWait),
+        &wait_info,
+    ) catch return VkError.DeviceLost;
+
+    return switch (errno) {
+        .SUCCESS => {},
+        .TIME => VkError.Timeout,
+        else => VkError.DeviceLost,
+    };
+}
+
+fn destroyHandle(device: *FlintDevice, io: std.Io, handle: u32) void {
+    var destroy_info = SyncObjDestroy{
+        .handle = handle,
+        .pad = 0,
+    };
+    base.utils.ioctl(
+        device.kmd.file() catch return,
+        io,
+        kmd.drmIoctlIowr(drm_syncobj_destroy, SyncObjDestroy),
+        &destroy_info,
+    ) catch @panic("ioctl failed");
+}
+
+fn absoluteTimeout(io: std.Io, timeout: u64) i64 {
+    if (timeout == std.math.maxInt(u64)) return std.math.maxInt(i64);
+
+    const now = std.Io.Clock.awake.now(io).nanoseconds;
+    const deadline: i96 = now + @as(i96, timeout);
+    return @intCast(@min(deadline, std.math.maxInt(i64)));
 }

src/intel/FlintImage.zig

@@ -1,16 +1,21 @@
+//! Flint images currently use a tightly packed, linear representation.
+//! Aspects are stored consecutively. Within each aspect, every array layer
+//! contains all mip levels, and every mip level contains its depth slices and
+//! samples. Keeping this layout description here gives command encoding a
+//! single source of truth for image addresses and pitches.
+
 const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
 const lib = @import("lib.zig");
 
+const FlintDeviceMemory = @import("FlintDeviceMemory.zig");
+
 const VkError = base.VkError;
 
 const Self = @This();
 pub const Interface = base.Image;
 
-pub const F32x4 = @Vector(4, f32);
-pub const U32x4 = @Vector(4, u32);
-
 interface: Interface,
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.ImageCreateInfo) VkError!*Self {
@@ -18,6 +23,9 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(device, allocator, info);
+    interface.allowed_memory_types = std.bit_set.IntegerBitSet(32).initEmpty();
+    interface.allowed_memory_types.set(0);
+
     interface.vtable = &.{
         .destroy = destroy,
         .getMemoryRequirements = getMemoryRequirements,
@@ -28,9 +36,7 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
         .copyToMemory = copyToMemory,
     };
 
-    self.* = .{
-        .interface = interface,
-    };
+    self.* = .{ .interface = interface };
     return self;
 }
 
@@ -40,39 +46,141 @@ pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
 }
 
 pub fn getMemoryRequirements(_: *Interface, requirements: *vk.MemoryRequirements) VkError!void {
-    _ = requirements;
+    requirements.alignment = lib.image_memory_alignment;
+    requirements.size = std.mem.alignForward(vk.DeviceSize, requirements.size, lib.image_memory_alignment);
 }
 
-pub fn copyToMemory(interface: *const Interface, memory: []u8, subresource: vk.ImageSubresourceLayers) VkError!void {
-    _ = interface;
-    _ = subresource;
-    @memset(memory, 0);
+pub fn copyToMemory(interface: *const Interface, dst: []u8, subresource: vk.ImageSubresourceLayers) VkError!void {
+    const self: *const Self = @alignCast(@fieldParentPtr("interface", interface));
+    const memory_interface = interface.memory orelse return VkError.InvalidDeviceMemoryDrv;
+    const memory: *FlintDeviceMemory = @alignCast(@fieldParentPtr("interface", memory_interface));
+
+    try validateSingleAspect(interface.format, subresource.aspect_mask);
+    if (subresource.mip_level >= interface.mip_levels or
+        subresource.base_array_layer >= interface.array_layers or
+        subresource.layer_count == 0)
+        return VkError.ValidationFailed;
+    const layer_count = if (subresource.layer_count == vk.REMAINING_ARRAY_LAYERS)
+        interface.array_layers - subresource.base_array_layer
+    else
+        subresource.layer_count;
+    if (layer_count > interface.array_layers - subresource.base_array_layer)
+        return VkError.ValidationFailed;
+    const level_size = self.getMultiSampledLevelSize(subresource.aspect_mask, subresource.mip_level);
+    const required_size, const size_overflow = @mulWithOverflow(level_size, @as(usize, layer_count));
+    if (size_overflow != 0 or dst.len < required_size) return VkError.ValidationFailed;
+
+    const first_offset = try self.getSubresourceOffset(
+        subresource.aspect_mask,
+        subresource.mip_level,
+        subresource.base_array_layer,
+    );
+    const absolute_offset, const offset_overflow = @addWithOverflow(interface.memory_offset, first_offset);
+    if (offset_overflow != 0) return VkError.ValidationFailed;
+
+    const device: *@import("FlintDevice.zig") = @alignCast(@fieldParentPtr("interface", interface.owner));
+    const mapped = try memory.allocation.map(&device.kmd, interface.owner.io(), absolute_offset, vk.WHOLE_SIZE);
+
+    const layer_pitch = self.getLayerSize(subresource.aspect_mask);
+    var dst_offset: usize = 0;
+    var src_offset: usize = 0;
+    for (0..layer_count) |_| {
+        if (src_offset > mapped.len or level_size > mapped.len - src_offset)
+            return VkError.InvalidDeviceMemoryDrv;
+        @memcpy(dst[dst_offset..][0..level_size], mapped[src_offset..][0..level_size]);
+        dst_offset += level_size;
+        src_offset += layer_pitch;
+    }
+}
+
+pub fn getSubresourceOffset(self: *const Self, aspect_mask: vk.ImageAspectFlags, mip_level: u32, layer: u32) VkError!usize {
+    if (mip_level >= self.interface.mip_levels or layer >= self.interface.array_layers)
+        return VkError.ValidationFailed;
+
+    var offset = try self.getAspectOffset(aspect_mask);
+    offset += layer * self.getLayerSize(aspect_mask);
+    for (0..mip_level) |mip|
+        offset += self.getMultiSampledLevelSize(aspect_mask, @intCast(mip));
+    return offset;
+}
+
+fn getAspectOffset(self: *const Self, aspect_mask: vk.ImageAspectFlags) VkError!usize {
+    try validateSingleAspect(self.interface.format, aspect_mask);
+    return switch (self.interface.format) {
+        .d16_unorm_s8_uint,
+        .d24_unorm_s8_uint,
+        .d32_sfloat_s8_uint,
+        => if (aspect_mask.stencil_bit)
+            self.interface.getTotalSizeForAspect(.{ .depth_bit = true })
+        else
+            0,
+        else => 0,
+    };
 }
 
 pub fn getTotalSizeForAspect(interface: *const Interface, aspect_mask: vk.ImageAspectFlags) VkError!usize {
-    _ = aspect_mask;
-    return interface.extent.width * interface.extent.height * interface.extent.depth * base.format.texelSize(interface.format);
+    const self: *const Self = @alignCast(@fieldParentPtr("interface", interface));
+    const valid_aspects = base.format.toAspect(interface.format);
+    if (aspect_mask.toInt() == 0 or aspect_mask.subtract(valid_aspects).toInt() != 0)
+        return VkError.ValidationFailed;
+
+    var size: usize = 0;
+    if (aspect_mask.color_bit) size += self.getLayerSize(.{ .color_bit = true });
+    if (aspect_mask.depth_bit) size += self.getLayerSize(.{ .depth_bit = true });
+    if (aspect_mask.stencil_bit) size += self.getLayerSize(.{ .stencil_bit = true });
+    return size * interface.array_layers;
 }
 
 pub fn getSubresourceLayout(interface: *const Interface, subresource: vk.ImageSubresource) VkError!vk.SubresourceLayout {
-    _ = subresource;
+    const self: *const Self = @alignCast(@fieldParentPtr("interface", interface));
+    try validateSingleAspect(interface.format, subresource.aspect_mask);
+
     return .{
-        .offset = 0,
-        .size = try getTotalSizeForAspect(interface, base.format.toAspect(interface.format)),
-        .row_pitch = getRowPitchMemSizeForMipLevel(interface, base.format.toAspect(interface.format), 0),
-        .array_pitch = getSliceMemSizeForMipLevel(interface, base.format.toAspect(interface.format), 0),
-        .depth_pitch = getSliceMemSizeForMipLevel(interface, base.format.toAspect(interface.format), 0),
+        .offset = try self.getSubresourceOffset(subresource.aspect_mask, subresource.mip_level, subresource.array_layer),
+        .size = self.getMultiSampledLevelSize(subresource.aspect_mask, subresource.mip_level),
+        .row_pitch = getRowPitchMemSizeForMipLevel(interface, subresource.aspect_mask, subresource.mip_level),
+        .array_pitch = self.getLayerSize(subresource.aspect_mask),
+        .depth_pitch = getSliceMemSizeForMipLevel(interface, subresource.aspect_mask, subresource.mip_level),
+    };
+}
+
+pub fn getLayerSize(self: *const Self, aspect_mask: vk.ImageAspectFlags) usize {
+    var size: usize = 0;
+    for (0..self.interface.mip_levels) |mip_level|
+        size += self.getMultiSampledLevelSize(aspect_mask, @intCast(mip_level));
+    return size;
+}
+
+pub inline fn getMultiSampledLevelSize(self: *const Self, aspect_mask: vk.ImageAspectFlags, mip_level: u32) usize {
+    return self.getMipLevelSize(aspect_mask, mip_level) * self.interface.samples.toInt();
+}
+
+pub inline fn getMipLevelSize(self: *const Self, aspect_mask: vk.ImageAspectFlags, mip_level: u32) usize {
+    return getSliceMemSizeForMipLevel(&self.interface, aspect_mask, mip_level) * self.getMipLevelExtent(mip_level).depth;
+}
+
+pub fn getMipLevelExtent(self: *const Self, mip_level: u32) vk.Extent3D {
+    return .{
+        .width = @max(1, self.interface.extent.width >> @intCast(mip_level)),
+        .height = @max(1, self.interface.extent.height >> @intCast(mip_level)),
+        .depth = @max(1, self.interface.extent.depth >> @intCast(mip_level)),
     };
 }
 
 pub fn getSliceMemSizeForMipLevel(interface: *const Interface, aspect_mask: vk.ImageAspectFlags, mip_level: u32) usize {
-    _ = aspect_mask;
-    _ = mip_level;
-    return interface.extent.width * interface.extent.height * base.format.texelSize(interface.format);
+    const self: *const Self = @alignCast(@fieldParentPtr("interface", interface));
+    const extent = self.getMipLevelExtent(mip_level);
+    return base.format.sliceMemSize(base.format.fromAspect(interface.format, aspect_mask), extent.width, extent.height);
 }
 
 pub fn getRowPitchMemSizeForMipLevel(interface: *const Interface, aspect_mask: vk.ImageAspectFlags, mip_level: u32) usize {
-    _ = aspect_mask;
-    _ = mip_level;
-    return interface.extent.width * base.format.texelSize(interface.format);
+    const self: *const Self = @alignCast(@fieldParentPtr("interface", interface));
+    const extent = self.getMipLevelExtent(mip_level);
+    return base.format.pitchMemSize(base.format.fromAspect(interface.format, aspect_mask), extent.width);
+}
+
+fn validateSingleAspect(format: vk.Format, aspect_mask: vk.ImageAspectFlags) VkError!void {
+    const valid_aspects = base.format.toAspect(format);
+    if (aspect_mask.toInt() == 0 or @popCount(aspect_mask.toInt()) != 1 or aspect_mask.subtract(valid_aspects).toInt() != 0)
+        return VkError.ValidationFailed;
 }

src/intel/FlintImageView.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.ImageView;

src/intel/FlintInstance.zig

@@ -26,7 +26,7 @@ fn castExtension(comptime ext: vk.ApiInfo) vk.ExtensionProperties {
     return props;
 }
 
-pub const EXTENSIONS = [_]vk.ExtensionProperties{
+pub const extensions = [_]vk.ExtensionProperties{
     castExtension(vk.extensions.khr_device_group_creation),
     castExtension(vk.extensions.khr_get_physical_device_properties_2),
     castExtension(vk.extensions.khr_surface),
@@ -73,11 +73,24 @@ fn requestPhysicalDevices(interface: *Interface, allocator: std.mem.Allocator, d
 
         if (drm_device.node_type != .render or
             std.meta.activeTag(drm_device.device_info) != .pci or
-            drm_device.device_info.pci.vendor_id != lib.INTEL_PCI_VENDOR_ID)
+            drm_device.device_info.pci.vendor_id != lib.intel_pci_vendor_id)
             continue;
 
-        const physical_device = try FlintPhysicalDevice.create(allocator, interface, &drm_device);
-        errdefer physical_device.interface.release(allocator) catch {};
+        const version = device.getVersion(io_var, allocator) catch continue;
+        defer version.deinit(allocator);
+
+        const kmd_type: lib.KmdType = if (std.mem.eql(u8, version.name, "i915"))
+            .i915
+        else if (std.mem.eql(u8, version.name, "xe"))
+            .xe
+        else
+            .invalid;
+
+        if (kmd_type == .invalid)
+            continue;
+
+        const physical_device = try FlintPhysicalDevice.create(allocator, interface, &drm_device, kmd_type);
+        errdefer physical_device.interface.release(allocator) catch @panic("Caught an error while handling an error");
 
         const dispatchable = try Dispatchable(base.PhysicalDevice).wrap(allocator, &physical_device.interface);
         errdefer dispatchable.destroy(allocator);

src/intel/FlintPhysicalDevice.zig

@@ -9,7 +9,6 @@ const pci_ids = @import("pci_ids.zig").map;
 const FlintDevice = @import("FlintDevice.zig");
 
 const VkError = base.VkError;
-const VulkanAllocator = base.VulkanAllocator;
 const SurfaceKHR = base.SurfaceKHR;
 
 const Self = @This();
@@ -24,14 +23,15 @@ fn castExtension(comptime ext: vk.ApiInfo) vk.ExtensionProperties {
     return props;
 }
 
-pub const EXTENSIONS = [_]vk.ExtensionProperties{
-    castExtension(vk.extensions.khr_device_group),
+pub const extensions = [_]vk.ExtensionProperties{
     castExtension(vk.extensions.khr_swapchain),
 };
 
 interface: Interface,
+kmd_type: lib.KmdType,
+node_path: [base.drm.max_node_name:0]u8,
 
-pub fn create(allocator: std.mem.Allocator, instance: *base.Instance, drm_device: *const base.drm.Device) VkError!*Self {
+pub fn create(allocator: std.mem.Allocator, instance: *base.Instance, drm_device: *const base.drm.Device, kmd_type: lib.KmdType) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
@@ -53,9 +53,9 @@ pub fn create(allocator: std.mem.Allocator, instance: *base.Instance, drm_device
         .getSurfaceSupportKHR = getSurfaceSupportKHR,
     };
 
-    interface.props.api_version = @bitCast(lib.VULKAN_VERSION);
-    interface.props.vendor_id = lib.INTEL_PCI_VENDOR_ID;
-    interface.props.driver_version = @bitCast(base.DRIVER_VERSION);
+    interface.props.api_version = @bitCast(lib.vulkan_version);
+    interface.props.vendor_id = lib.intel_pci_vendor_id;
+    interface.props.driver_version = @bitCast(base.driver_version);
     interface.props.device_id = drm_device.device_info.pci.device_id;
     interface.props.device_type = .integrated_gpu;
 
@@ -78,7 +78,7 @@ pub fn create(allocator: std.mem.Allocator, instance: *base.Instance, drm_device
         break;
     }
 
-    interface.props.pipeline_cache_uuid = undefined;
+    interface.props.pipeline_cache_uuid = @splat(0);
     interface.props.limits = .{
         .max_image_dimension_1d = 4096,
         .max_image_dimension_2d = 4096,
@@ -223,11 +223,19 @@ pub fn create(allocator: std.mem.Allocator, instance: *base.Instance, drm_device
 
     self.* = .{
         .interface = interface,
+        .kmd_type = kmd_type,
+        .node_path = @splat(0),
     };
+    const node_path = drm_device.nodePath();
+    @memcpy(self.node_path[0..node_path.len], node_path);
 
     return self;
 }
 
+pub fn getNodePath(self: *const Self) [:0]const u8 {
+    return std.mem.sliceTo(&self.node_path, 0);
+}
+
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     allocator.destroy(self);
@@ -248,10 +256,10 @@ pub fn enumerateExtensionProperties(_: *const Interface, layer_name: ?[]const u8
         return VkError.LayerNotPresent;
     }
 
-    const available = EXTENSIONS.len;
+    const available = extensions.len;
     if (p_properties) |properties| {
         const write_count = @min(count.*, available);
-        for (EXTENSIONS[0..write_count], properties[0..write_count]) |ext, *prop| {
+        for (extensions[0..write_count], properties[0..write_count]) |ext, *prop| {
             prop.* = ext;
         }
         count.* = @intCast(write_count);
@@ -368,7 +376,6 @@ pub fn getFormatProperties(interface: *Interface, format: vk.Format) VkError!vk.
             properties.optimal_tiling_features.sampled_image_bit = true;
             properties.optimal_tiling_features.transfer_dst_bit = true;
             properties.optimal_tiling_features.transfer_src_bit = true;
-            properties.optimal_tiling_features.sampled_image_filter_linear_bit = true;
         },
 
         // Formats which can be sampled, but don't support filtering
@@ -408,7 +415,6 @@ pub fn getFormatProperties(interface: *Interface, format: vk.Format) VkError!vk.
         .g10x6_b10x6r10x6_2plane_420_unorm_3pack16,
         => {
             properties.optimal_tiling_features.sampled_image_bit = true;
-            properties.optimal_tiling_features.sampled_image_filter_linear_bit = true;
             properties.optimal_tiling_features.sampled_image_ycbcr_conversion_linear_filter_bit = true;
             properties.optimal_tiling_features.transfer_src_bit = true;
             properties.optimal_tiling_features.transfer_dst_bit = true;
@@ -678,6 +684,11 @@ pub fn getFormatProperties(interface: *Interface, format: vk.Format) VkError!vk.
         properties.linear_tiling_features.transfer_dst_bit = true;
     }
 
+    if (properties.optimal_tiling_features.blit_src_bit or properties.optimal_tiling_features.blit_dst_bit) {
+        properties.optimal_tiling_features.blit_src_bit = true;
+        properties.optimal_tiling_features.blit_dst_bit = true;
+    }
+
     return properties;
 }
 
@@ -687,20 +698,48 @@ pub fn getImageFormatProperties(
     image_type: vk.ImageType,
     tiling: vk.ImageTiling,
     usage: vk.ImageUsageFlags,
-    _: vk.ImageCreateFlags,
+    flags: vk.ImageCreateFlags,
 ) VkError!vk.ImageFormatProperties {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
     if (!try self.isFormatSupported(format, image_type, tiling, usage))
         return VkError.FormatNotSupported;
 
-    const properties: vk.ImageFormatProperties = .{
-        .max_extent = .{ .width = 0, .height = 0, .depth = 1 },
+    var properties: vk.ImageFormatProperties = .{
+        .max_extent = .{ .width = 1, .height = 1, .depth = 1 },
         .max_mip_levels = 1,
-        .max_array_layers = 1,
+        .max_array_layers = interface.props.limits.max_image_array_layers,
         .sample_counts = .{ .@"1_bit" = true },
         .max_resource_size = std.math.maxInt(u32),
     };
 
+    switch (image_type) {
+        .@"1d" => {
+            properties.max_extent.width = interface.props.limits.max_image_dimension_1d;
+            properties.max_mip_levels = std.math.log2_int(u32, properties.max_extent.width) + 1;
+        },
+        .@"2d" => {
+            const dimension = if (flags.cube_compatible_bit)
+                interface.props.limits.max_image_dimension_cube
+            else
+                interface.props.limits.max_image_dimension_2d;
+            properties.max_extent.width = dimension;
+            properties.max_extent.height = dimension;
+            properties.max_mip_levels = std.math.log2_int(u32, dimension) + 1;
+        },
+        .@"3d" => {
+            const dimension = interface.props.limits.max_image_dimension_3d;
+            properties.max_extent = .{ .width = dimension, .height = dimension, .depth = dimension };
+            properties.max_mip_levels = std.math.log2_int(u32, dimension) + 1;
+            properties.max_array_layers = 1;
+        },
+        else => return VkError.FormatNotSupported,
+    }
+
+    if (tiling == .linear) {
+        properties.max_mip_levels = 1;
+        properties.max_array_layers = 1;
+    }
+
     return properties;
 }
 
@@ -827,10 +866,10 @@ fn cpuid(leaf_id: u32, subleaf_id: u32) CpuidRegs {
         }
     }
 
-    var eax: u32 = undefined;
-    var ebx: u32 = undefined;
-    var ecx: u32 = undefined;
-    var edx: u32 = undefined;
+    var eax: u32 = 0;
+    var ebx: u32 = 0;
+    var ecx: u32 = 0;
+    var edx: u32 = 0;
 
     asm volatile ("cpuid"
         : [_] "={eax}" (eax),

src/intel/FlintPipeline.zig

@@ -8,6 +8,7 @@ const Self = @This();
 pub const Interface = base.Pipeline;
 
 interface: Interface,
+host_allocator: base.VulkanAllocator,
 
 pub fn createCompute(device: *base.Device, allocator: std.mem.Allocator, cache: ?*base.PipelineCache, info: *const vk.ComputePipelineCreateInfo) VkError!*Self {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
@@ -15,8 +16,12 @@ pub fn createCompute(device: *base.Device, allocator: std.mem.Allocator, cache:
 
     var interface = try Interface.initCompute(device, allocator, cache, info);
     interface.vtable = &.{ .destroy = destroy };
+    self.* = .{
+        .interface = interface,
+        .host_allocator = base.VulkanAllocator.from(allocator).clone(),
+    };
+    errdefer self.interface.layout.unref(allocator);
 
-    self.* = .{ .interface = interface };
     return self;
 }
 
@@ -27,7 +32,12 @@ pub fn createGraphics(device: *base.Device, allocator: std.mem.Allocator, cache:
     var interface = try Interface.initGraphics(device, allocator, cache, info);
     interface.vtable = &.{ .destroy = destroy };
 
-    self.* = .{ .interface = interface };
+    self.* = .{
+        .interface = interface,
+        .host_allocator = base.VulkanAllocator.from(allocator).clone(),
+    };
+    errdefer self.interface.layout.unref(allocator);
+
     return self;
 }
 

src/intel/FlintPipelineLayout.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.PipelineLayout;

src/intel/FlintQueryPool.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.QueryPool;

src/intel/FlintQueue.zig

@@ -2,7 +2,11 @@ const std = @import("std");
 const vk = @import("vulkan");
 const base = @import("base");
 
+const FlintBinarySemaphore = @import("FlintBinarySemaphore.zig");
 const FlintCommandBuffer = @import("FlintCommandBuffer.zig");
+const FlintDevice = @import("FlintDevice.zig");
+const FlintFence = @import("FlintFence.zig");
+const kmd = @import("kmd.zig");
 
 const VkError = base.VkError;
 
@@ -10,24 +14,35 @@ const Self = @This();
 pub const Interface = base.Queue;
 
 interface: Interface,
+completion: *FlintFence,
 
 pub fn create(allocator: std.mem.Allocator, device: *base.Device, index: u32, family_index: u32, flags: vk.DeviceQueueCreateFlags) VkError!*Interface {
     const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
     errdefer allocator.destroy(self);
 
     var interface = try Interface.init(allocator, device, index, family_index, flags);
+    const completion = try FlintFence.create(device, allocator, &.{
+        .s_type = .fence_create_info,
+        .p_next = null,
+        .flags = .{ .signaled_bit = true },
+    });
+    errdefer completion.interface.destroy(allocator);
     interface.dispatch_table = &.{
         .bindSparse = bindSparse,
         .submit = submit,
         .waitIdle = waitIdle,
     };
 
-    self.* = .{ .interface = interface };
+    self.* = .{
+        .interface = interface,
+        .completion = completion,
+    };
     return &self.interface;
 }
 
 pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
     const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.completion.interface.destroy(allocator);
     allocator.destroy(self);
 }
 
@@ -39,26 +54,89 @@ pub fn bindSparse(interface: *Interface, info: []const vk.BindSparseInfo, fence:
 }
 
 pub fn submit(interface: *Interface, infos: []Interface.SubmitInfo, fence: ?*base.Fence) VkError!void {
-    _ = interface;
-    for (infos) |info| {
-        for (info.wait_semaphores.items) |semaphore| {
-            try semaphore.wait();
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *FlintDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    const allocator = interface.host_allocator.allocator();
+
+    try self.completion.interface.reset();
+
+    for (infos, 0..) |info, info_index| {
+        const last_info = info_index + 1 == infos.len;
+        const request_count = @max(info.command_buffers.items.len, 1);
+
+        for (0..request_count) |request_index| {
+            const first_request = request_index == 0;
+            const last_request = request_index + 1 == request_count;
+
+            var syncs = std.ArrayList(kmd.SyncDependency).empty;
+            defer syncs.deinit(allocator);
+
+            if (first_request) {
+                for (info.wait_semaphores.items) |base_semaphore| {
+                    const semaphore: *FlintBinarySemaphore = @alignCast(@fieldParentPtr("interface", base_semaphore));
+                    syncs.append(allocator, .{ .handle = semaphore.handle, .wait = true }) catch return VkError.OutOfHostMemory;
+                }
             }
 
-        for (info.command_buffers.items) |command_buffer| {
+            if (last_request) {
+                for (info.signal_semaphores.items) |base_semaphore| {
+                    const semaphore: *FlintBinarySemaphore = @alignCast(@fieldParentPtr("interface", base_semaphore));
+                    syncs.append(allocator, .{ .handle = semaphore.handle, .signal = true }) catch return VkError.OutOfHostMemory;
+                }
+
+                if (last_info) {
+                    syncs.append(allocator, .{ .handle = self.completion.handle, .signal = true }) catch return VkError.OutOfHostMemory;
+                    if (fence) |base_fence| {
+                        const flint_fence: *FlintFence = @alignCast(@fieldParentPtr("interface", base_fence));
+                        syncs.append(allocator, .{ .handle = flint_fence.handle, .signal = true }) catch return VkError.OutOfHostMemory;
+                    }
+                }
+            }
+
+            if (info.command_buffers.items.len == 0) {
+                try device.kmd.submitBatch(
+                    interface.owner.io(),
+                    allocator,
+                    &.{},
+                    &.{},
+                    syncs.items,
+                );
+            } else {
+                const command_buffer = info.command_buffers.items[request_index];
                 const intel_command_buffer: *FlintCommandBuffer = @alignCast(@fieldParentPtr("interface", command_buffer));
-            _ = intel_command_buffer;
+                try intel_command_buffer.submitGpuBatch(syncs.items);
             }
 
-        for (info.signal_semaphores.items) |semaphore| {
-            try semaphore.signal();
+            if (first_request) {
+                for (info.wait_semaphores.items) |base_semaphore| {
+                    const semaphore: *FlintBinarySemaphore = @alignCast(@fieldParentPtr("interface", base_semaphore));
+                    try FlintBinarySemaphore.reset(&semaphore.interface);
                 }
             }
-    if (fence) |value| {
-        try value.signal();
+        }
+    }
+
+    if (infos.len == 0) {
+        var syncs: [2]kmd.SyncDependency = undefined;
+        var sync_count: usize = 1;
+        syncs[0] = .{ .handle = self.completion.handle, .signal = true };
+        if (fence) |base_fence| {
+            const flint_fence: *FlintFence = @alignCast(@fieldParentPtr("interface", base_fence));
+            syncs[sync_count] = .{ .handle = flint_fence.handle, .signal = true };
+            sync_count += 1;
+        }
+
+        try device.kmd.submitBatch(
+            interface.owner.io(),
+            allocator,
+            &.{},
+            &.{},
+            syncs[0..sync_count],
+        );
     }
 }
 
 pub fn waitIdle(interface: *Interface) VkError!void {
-    _ = interface;
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    try self.completion.interface.wait(std.math.maxInt(u64));
 }

src/intel/FlintRenderPass.zig

@@ -3,7 +3,6 @@ const vk = @import("vulkan");
 const base = @import("base");
 
 const VkError = base.VkError;
-const Device = base.Device;
 
 const Self = @This();
 pub const Interface = base.RenderPass;

src/intel/FlintShaderModule.zig

@@ -8,6 +8,7 @@ const Self = @This();
 pub const Interface = base.ShaderModule;
 
 interface: Interface,
+code: []u32,
 ref_count: std.atomic.Value(usize),
 
 pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.ShaderModuleCreateInfo) VkError!*Self {
@@ -16,9 +17,13 @@ pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const v
 
     var interface = try Interface.init(device, allocator, info);
     interface.vtable = &.{ .destroy = destroy };
+    if (info.code_size % @sizeOf(u32) != 0) return VkError.ValidationFailed;
+    const code = allocator.dupe(u32, info.p_code[0 .. info.code_size / @sizeOf(u32)]) catch return VkError.OutOfHostMemory;
+    errdefer allocator.free(code);
 
     self.* = .{
         .interface = interface,
+        .code = code,
         .ref_count = std.atomic.Value(usize).init(1),
     };
     return self;
@@ -30,6 +35,7 @@ pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
 }
 
 pub fn drop(self: *Self, allocator: std.mem.Allocator) void {
+    allocator.free(self.code);
     allocator.destroy(self);
 }
 

src/intel/MemoryRange.zig

@@ -0,0 +1,42 @@
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+const FlintDeviceMemory = @import("FlintDeviceMemory.zig");
+
+const Self = @This();
+
+memory: *FlintDeviceMemory,
+offset: vk.DeviceSize,
+size: vk.DeviceSize,
+
+pub fn fromBuffer(buffer: *base.Buffer, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!Self {
+    const base_memory = buffer.memory orelse return VkError.InvalidDeviceMemoryDrv;
+    const bound, const bound_overflow = @addWithOverflow(offset, size);
+    if (bound_overflow != 0 or bound > buffer.size) return VkError.ValidationFailed;
+
+    const memory_offset, const memory_offset_overflow = @addWithOverflow(buffer.offset, offset);
+    if (memory_offset_overflow != 0) return VkError.ValidationFailed;
+
+    return fromMemory(base_memory, memory_offset, size);
+}
+
+pub fn fromImage(image: *base.Image, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!Self {
+    const base_memory = image.memory orelse return VkError.InvalidDeviceMemoryDrv;
+    const memory_offset, const memory_offset_overflow = @addWithOverflow(image.memory_offset, offset);
+    if (memory_offset_overflow != 0) return VkError.ValidationFailed;
+
+    return fromMemory(base_memory, memory_offset, size);
+}
+
+pub fn fromMemory(base_memory: *base.DeviceMemory, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!Self {
+    const memory: *FlintDeviceMemory = @alignCast(@fieldParentPtr("interface", base_memory));
+    if (offset > base_memory.size or size > base_memory.size - offset)
+        return VkError.ValidationFailed;
+
+    return .{
+        .memory = memory,
+        .offset = offset,
+        .size = size,
+    };
+}

src/intel/blitter.zig

@@ -0,0 +1,171 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+const FlintImage = @import("FlintImage.zig");
+const FlintCommandBuffer = @import("FlintCommandBuffer.zig");
+const MemoryRange = @import("MemoryRange.zig");
+
+const copy = @import("copy_commands.zig");
+
+pub fn blitImageRegion(cmd: *FlintCommandBuffer, src: *base.Image, dst: *base.Image, region: vk.ImageBlit) VkError!void {
+    if (region.src_subresource.mip_level >= src.mip_levels or
+        region.dst_subresource.mip_level >= dst.mip_levels)
+        return VkError.ValidationFailed;
+    if (src.samples.toInt() != 1 or dst.samples.toInt() != 1)
+        return VkError.ValidationFailed;
+    if (src.image_type != dst.image_type)
+        return VkError.FeatureNotPresent;
+    try validateAspect(src, region.src_subresource.aspect_mask);
+    try validateAspect(dst, region.dst_subresource.aspect_mask);
+    if (region.src_subresource.aspect_mask != region.dst_subresource.aspect_mask)
+        return VkError.ValidationFailed;
+
+    const src_image: *FlintImage = @alignCast(@fieldParentPtr("interface", src));
+    const dst_image: *FlintImage = @alignCast(@fieldParentPtr("interface", dst));
+    const src_format = src.formatFromAspect(region.src_subresource.aspect_mask);
+    const dst_format = dst.formatFromAspect(region.dst_subresource.aspect_mask);
+    const src_texel_size = base.format.texelSize(src_format);
+    const dst_texel_size = base.format.texelSize(dst_format);
+
+    if (base.format.isCompressed(src_format) or base.format.isCompressed(dst_format))
+        return VkError.FormatNotSupported;
+
+    // xy_src_copy_blt does not perform component conversion. Different Vulkan
+    // names are safe only when they describe the same bytes and values.
+    if (!bitwiseCompatibleFormats(src_format, dst_format) or src_texel_size != dst_texel_size)
+        return VkError.FormatNotSupported;
+
+    if ((base.format.isDepth(src.format) or base.format.isStencil(src.format) or
+        base.format.isDepth(dst.format) or base.format.isStencil(dst.format)) and
+        src.format != dst.format)
+        return VkError.FormatNotSupported;
+
+    const src_extent = src_image.getMipLevelExtent(region.src_subresource.mip_level);
+    const dst_extent = dst_image.getMipLevelExtent(region.dst_subresource.mip_level);
+    try validateOffsets(region.src_offsets, src_extent);
+    try validateOffsets(region.dst_offsets, dst_extent);
+
+    const src_layer_count = try resolveLayerCount(src, region.src_subresource);
+    const dst_layer_count = try resolveLayerCount(dst, region.dst_subresource);
+    if (src_layer_count != dst_layer_count)
+        return VkError.ValidationFailed;
+
+    var src_0 = region.src_offsets[0];
+    var src_1 = region.src_offsets[1];
+    var dst_0 = region.dst_offsets[0];
+    var dst_1 = region.dst_offsets[1];
+    inline for (.{ "x", "y", "z" }) |field| {
+        if (@field(dst_0, field) > @field(dst_1, field)) {
+            std.mem.swap(i32, &@field(src_0, field), &@field(src_1, field));
+            std.mem.swap(i32, &@field(dst_0, field), &@field(dst_1, field));
+        }
+    }
+
+    if (dst_0.x == dst_1.x or dst_0.y == dst_1.y or dst_0.z == dst_1.z or
+        src_0.x == src_1.x or src_0.y == src_1.y or src_0.z == src_1.z)
+        return VkError.ValidationFailed;
+
+    const copy_whole_rows = src_1.x - src_0.x == dst_1.x - dst_0.x and src_1.x > src_0.x;
+    for (0..src_layer_count) |layer| {
+        const src_layer = region.src_subresource.base_array_layer + @as(u32, @intCast(layer));
+        const dst_layer = region.dst_subresource.base_array_layer + @as(u32, @intCast(layer));
+
+        var dst_z = dst_0.z;
+        while (dst_z < dst_1.z) : (dst_z += 1) {
+            const src_z = nearestBlitCoordinate(src_0.z, src_1.z, dst_0.z, dst_1.z, dst_z, src_extent.depth);
+            var dst_y = dst_0.y;
+            while (dst_y < dst_1.y) : (dst_y += 1) {
+                const src_y = nearestBlitCoordinate(src_0.y, src_1.y, dst_0.y, dst_1.y, dst_y, src_extent.height);
+
+                if (copy_whole_rows) {
+                    const src_offset = try imageTexelOffset(src_image, region.src_subresource.aspect_mask, region.src_subresource.mip_level, src_layer, @intCast(src_0.x), src_y, src_z);
+                    const dst_offset = try imageTexelOffset(dst_image, region.dst_subresource.aspect_mask, region.dst_subresource.mip_level, dst_layer, @intCast(dst_0.x), @intCast(dst_y), @intCast(dst_z));
+                    const size: usize = @as(usize, @intCast(dst_1.x - dst_0.x)) * dst_texel_size;
+                    try copy.emitLinearCopy(
+                        cmd,
+                        try MemoryRange.fromImage(src, src_offset, size),
+                        try MemoryRange.fromImage(dst, dst_offset, size),
+                    );
+                    continue;
+                }
+
+                var dst_x = dst_0.x;
+                while (dst_x < dst_1.x) : (dst_x += 1) {
+                    const src_x = nearestBlitCoordinate(src_0.x, src_1.x, dst_0.x, dst_1.x, dst_x, src_extent.width);
+                    const src_offset = try imageTexelOffset(src_image, region.src_subresource.aspect_mask, region.src_subresource.mip_level, src_layer, src_x, src_y, src_z);
+                    const dst_offset = try imageTexelOffset(dst_image, region.dst_subresource.aspect_mask, region.dst_subresource.mip_level, dst_layer, @intCast(dst_x), @intCast(dst_y), @intCast(dst_z));
+                    try copy.emitLinearCopy(
+                        cmd,
+                        try MemoryRange.fromImage(src, src_offset, src_texel_size),
+                        try MemoryRange.fromImage(dst, dst_offset, dst_texel_size),
+                    );
+                }
+            }
+        }
+    }
+}
+
+fn validateAspect(image: *const base.Image, aspect: vk.ImageAspectFlags) VkError!void {
+    const valid_aspects = base.format.toAspect(image.format);
+    if (aspect.toInt() == 0 or @popCount(aspect.toInt()) != 1 or
+        aspect.subtract(valid_aspects).toInt() != 0)
+        return VkError.ValidationFailed;
+}
+
+fn validateOffsets(offsets: [2]vk.Offset3D, extent: vk.Extent3D) VkError!void {
+    for (offsets) |offset| {
+        if (offset.x < 0 or offset.y < 0 or offset.z < 0 or
+            offset.x > extent.width or offset.y > extent.height or offset.z > extent.depth)
+            return VkError.ValidationFailed;
+    }
+}
+
+fn resolveLayerCount(image: *const base.Image, subresource: vk.ImageSubresourceLayers) VkError!u32 {
+    if (subresource.base_array_layer >= image.array_layers)
+        return VkError.ValidationFailed;
+    const available = image.array_layers - subresource.base_array_layer;
+    const count = if (subresource.layer_count == vk.REMAINING_ARRAY_LAYERS)
+        available
+    else
+        subresource.layer_count;
+    if (count == 0 or count > available)
+        return VkError.ValidationFailed;
+    return count;
+}
+
+fn bitwiseCompatibleFormats(src: vk.Format, dst: vk.Format) bool {
+    if (src == dst) return true;
+
+    // On little-endian the packed A8B8G8R8 formats have the same byte
+    // representation and component interpretation as their R8G8B8A8 peers.
+    return switch (src) {
+        .r8g8b8a8_unorm => dst == .a8b8g8r8_unorm_pack32,
+        .a8b8g8r8_unorm_pack32 => dst == .r8g8b8a8_unorm,
+        .r8g8b8a8_snorm => dst == .a8b8g8r8_snorm_pack32,
+        .a8b8g8r8_snorm_pack32 => dst == .r8g8b8a8_snorm,
+        .r8g8b8a8_uint => dst == .a8b8g8r8_uint_pack32,
+        .a8b8g8r8_uint_pack32 => dst == .r8g8b8a8_uint,
+        .r8g8b8a8_sint => dst == .a8b8g8r8_sint_pack32,
+        .a8b8g8r8_sint_pack32 => dst == .r8g8b8a8_sint,
+        .r8g8b8a8_srgb => dst == .a8b8g8r8_srgb_pack32,
+        .a8b8g8r8_srgb_pack32 => dst == .r8g8b8a8_srgb,
+        else => false,
+    };
+}
+
+fn nearestBlitCoordinate(src_0: i32, src_1: i32, dst_0: i32, dst_1: i32, dst: i32, extent: u32) usize {
+    const numerator = @as(i64, 2 * (dst - dst_0) + 1) * @as(i64, src_1 - src_0);
+    const denominator: i64 = 2 * (dst_1 - dst_0);
+    const coordinate = @as(i64, src_0) + @divFloor(numerator, denominator);
+    return @intCast(std.math.clamp(coordinate, 0, @as(i64, extent) - 1));
+}
+
+fn imageTexelOffset(image: *const FlintImage, aspect: vk.ImageAspectFlags, mip_level: u32, layer: u32, x: usize, y: usize, z: usize) VkError!usize {
+    const format = image.interface.formatFromAspect(aspect);
+    return try image.getSubresourceOffset(aspect, mip_level, layer) +
+        z * image.interface.getSliceMemSizeForMipLevel(aspect, mip_level) +
+        y * image.interface.getRowPitchMemSizeForMipLevel(aspect, mip_level) +
+        x * base.format.texelSize(format);
+}

src/intel/copy_commands.zig

@@ -0,0 +1,246 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+const kmd = @import("kmd.zig");
+
+const VkError = base.VkError;
+const FlintImage = @import("FlintImage.zig");
+const FlintCommandBuffer = @import("FlintCommandBuffer.zig");
+const MemoryRange = @import("MemoryRange.zig");
+
+pub fn emitLinearCopy(cmd: *FlintCommandBuffer, src: MemoryRange, dst: MemoryRange) VkError!void {
+    if (src.size != dst.size) return VkError.ValidationFailed;
+
+    var copied: vk.DeviceSize = 0;
+    while (copied < src.size) {
+        const chunk = @min(src.size - copied, kmd.max_blt_span);
+        const src_chunk: MemoryRange = .{ .memory = src.memory, .offset = src.offset + copied, .size = chunk };
+        const dst_chunk: MemoryRange = .{ .memory = dst.memory, .offset = dst.offset + copied, .size = chunk };
+        const width: u32 = @intCast(chunk);
+
+        try cmd.emit(kmd.xy_src_copy_blt | kmd.xy_blt_write_alpha | kmd.xy_blt_write_rgb);
+        try cmd.emit(kmd.blt_depth_8 | kmd.rop_source_copy | width);
+        try cmd.emit(0);
+        try cmd.emit((1 << 16) | width);
+        try cmd.emitRelocatedAddress(dst_chunk, false, true);
+        try cmd.emit(0);
+        try cmd.emit(width);
+        try cmd.emitRelocatedAddress(src_chunk, true, false);
+
+        copied += chunk;
+    }
+}
+
+pub fn copyBufferImage(cmd: *FlintCommandBuffer, buffer: *base.Buffer, image_interface: *base.Image, region: vk.BufferImageCopy, image_is_dst: bool) VkError!void {
+    if (region.image_extent.width == 0 or region.image_extent.height == 0 or region.image_extent.depth == 0)
+        return;
+    if (region.image_offset.x < 0 or region.image_offset.y < 0 or region.image_offset.z < 0)
+        return VkError.ValidationFailed;
+
+    const image: *FlintImage = @alignCast(@fieldParentPtr("interface", image_interface));
+    const format = image_interface.formatFromAspect(region.image_subresource.aspect_mask);
+    const bytes_per_block = base.format.texelSize(format);
+    const block_width = base.format.blockWidth(format);
+    const block_height = base.format.blockHeight(format);
+
+    const buffer_width = if (region.buffer_row_length == 0) region.image_extent.width else region.buffer_row_length;
+    const buffer_height = if (region.buffer_image_height == 0) region.image_extent.height else region.buffer_image_height;
+    const buffer_row_pitch = base.format.pitchMemSize(format, buffer_width);
+    const buffer_slice_pitch = base.format.sliceMemSize(format, buffer_width, buffer_height);
+    const copy_row_size = base.format.pitchMemSize(format, region.image_extent.width);
+    const copy_rows = base.format.blockCountY(format, region.image_extent.height);
+
+    const layer_count = if (region.image_subresource.layer_count == vk.REMAINING_ARRAY_LAYERS)
+        image_interface.array_layers - region.image_subresource.base_array_layer
+    else
+        region.image_subresource.layer_count;
+    if (layer_count == 0 or layer_count > image_interface.array_layers - region.image_subresource.base_array_layer)
+        return VkError.ValidationFailed;
+
+    const image_row_pitch = image_interface.getRowPitchMemSizeForMipLevel(region.image_subresource.aspect_mask, region.image_subresource.mip_level);
+    const image_slice_pitch = image_interface.getSliceMemSizeForMipLevel(region.image_subresource.aspect_mask, region.image_subresource.mip_level);
+    const image_x_offset = @divFloor(@as(usize, @intCast(region.image_offset.x)), block_width) * bytes_per_block;
+    const image_y_offset = @divFloor(@as(usize, @intCast(region.image_offset.y)), block_height);
+
+    for (0..layer_count) |layer| {
+        const image_subresource_offset = try image.getSubresourceOffset(
+            region.image_subresource.aspect_mask,
+            region.image_subresource.mip_level,
+            region.image_subresource.base_array_layer + @as(u32, @intCast(layer)),
+        );
+
+        for (0..region.image_extent.depth) |z| {
+            for (0..copy_rows) |row| {
+                const buffer_offset = region.buffer_offset +
+                    (layer * region.image_extent.depth + z) * buffer_slice_pitch +
+                    row * buffer_row_pitch;
+                const image_offset = image_subresource_offset +
+                    (@as(usize, @intCast(region.image_offset.z)) + z) * image_slice_pitch +
+                    (image_y_offset + row) * image_row_pitch +
+                    image_x_offset;
+
+                const buffer_range = try MemoryRange.fromBuffer(buffer, buffer_offset, copy_row_size);
+                const image_range = try MemoryRange.fromImage(image_interface, image_offset, copy_row_size);
+                if (image_is_dst)
+                    try emitLinearCopy(cmd, buffer_range, image_range)
+                else
+                    try emitLinearCopy(cmd, image_range, buffer_range);
+            }
+        }
+    }
+}
+
+pub fn copyImage(cmd: *FlintCommandBuffer, src_interface: *base.Image, dst_interface: *base.Image, region: vk.ImageCopy) VkError!void {
+    const depth_stencil: vk.ImageAspectFlags = .{ .depth_bit = true, .stencil_bit = true };
+    if (region.src_subresource.aspect_mask == depth_stencil and region.dst_subresource.aspect_mask == depth_stencil) {
+        var single_aspect_region = region;
+        single_aspect_region.src_subresource.aspect_mask = .{ .depth_bit = true };
+        single_aspect_region.dst_subresource.aspect_mask = .{ .depth_bit = true };
+        try copyImageSingleAspect(cmd, src_interface, dst_interface, single_aspect_region);
+
+        single_aspect_region.src_subresource.aspect_mask = .{ .stencil_bit = true };
+        single_aspect_region.dst_subresource.aspect_mask = .{ .stencil_bit = true };
+        try copyImageSingleAspect(cmd, src_interface, dst_interface, single_aspect_region);
+        return;
+    }
+
+    try copyImageSingleAspect(cmd, src_interface, dst_interface, region);
+}
+
+fn copyImageSingleAspect(cmd: *FlintCommandBuffer, src_interface: *base.Image, dst_interface: *base.Image, region: vk.ImageCopy) VkError!void {
+    if (region.extent.width == 0 or region.extent.height == 0 or region.extent.depth == 0)
+        return;
+    if (region.src_offset.x < 0 or region.src_offset.y < 0 or region.src_offset.z < 0 or
+        region.dst_offset.x < 0 or region.dst_offset.y < 0 or region.dst_offset.z < 0)
+        return VkError.ValidationFailed;
+    if (@popCount(region.src_subresource.aspect_mask.toInt()) != 1 or
+        @popCount(region.dst_subresource.aspect_mask.toInt()) != 1)
+        return VkError.ValidationFailed;
+    if (region.src_subresource.aspect_mask.subtract(base.format.toAspect(src_interface.format)).toInt() != 0 or
+        region.dst_subresource.aspect_mask.subtract(base.format.toAspect(dst_interface.format)).toInt() != 0)
+        return VkError.ValidationFailed;
+    if (src_interface.samples.toInt() != dst_interface.samples.toInt())
+        return VkError.ValidationFailed;
+
+    const src: *FlintImage = @alignCast(@fieldParentPtr("interface", src_interface));
+    const dst: *FlintImage = @alignCast(@fieldParentPtr("interface", dst_interface));
+    const src_format = src_interface.formatFromAspect(region.src_subresource.aspect_mask);
+    const dst_format = dst_interface.formatFromAspect(region.dst_subresource.aspect_mask);
+    const bytes_per_block = base.format.texelSize(src_format);
+
+    if (bytes_per_block != base.format.texelSize(dst_format))
+        return VkError.FormatNotSupported;
+
+    const src_block_width = base.format.blockWidth(src_format);
+    const src_block_height = base.format.blockHeight(src_format);
+    const dst_block_width = base.format.blockWidth(dst_format);
+    const dst_block_height = base.format.blockHeight(dst_format);
+    if (base.format.isCompressed(src_format) and base.format.isCompressed(dst_format) and
+        (src_block_width != dst_block_width or src_block_height != dst_block_height))
+        return VkError.FormatNotSupported;
+
+    const src_x: usize = @intCast(region.src_offset.x);
+    const src_y: usize = @intCast(region.src_offset.y);
+    const src_z: usize = @intCast(region.src_offset.z);
+    const dst_x: usize = @intCast(region.dst_offset.x);
+    const dst_y: usize = @intCast(region.dst_offset.y);
+    const dst_z: usize = @intCast(region.dst_offset.z);
+    if (@mod(src_x, src_block_width) != 0 or @mod(src_y, src_block_height) != 0 or
+        @mod(dst_x, dst_block_width) != 0 or @mod(dst_y, dst_block_height) != 0)
+        return VkError.ValidationFailed;
+
+    if (region.src_subresource.mip_level >= src_interface.mip_levels or
+        region.dst_subresource.mip_level >= dst_interface.mip_levels)
+        return VkError.ValidationFailed;
+    const src_extent = src.getMipLevelExtent(region.src_subresource.mip_level);
+    const dst_extent = dst.getMipLevelExtent(region.dst_subresource.mip_level);
+    const copy_blocks_x = base.format.blockCountX(src_format, region.extent.width);
+    const copy_blocks_y = base.format.blockCountY(src_format, region.extent.height);
+    const src_block_x = src_x / src_block_width;
+    const src_block_y = src_y / src_block_height;
+    const dst_block_x = dst_x / dst_block_width;
+    const dst_block_y = dst_y / dst_block_height;
+    if (src_block_x + copy_blocks_x > base.format.blockCountX(src_format, src_extent.width) or
+        src_block_y + copy_blocks_y > base.format.blockCountY(src_format, src_extent.height) or
+        dst_block_x + copy_blocks_x > base.format.blockCountX(dst_format, dst_extent.width) or
+        dst_block_y + copy_blocks_y > base.format.blockCountY(dst_format, dst_extent.height) or
+        src_z + region.extent.depth > src_extent.depth or
+        dst_z + region.extent.depth > dst_extent.depth)
+        return VkError.ValidationFailed;
+
+    const src_layer_count = try resolveLayerCount(src_interface, region.src_subresource);
+    const dst_layer_count = try resolveLayerCount(dst_interface, region.dst_subresource);
+    if (src_layer_count != dst_layer_count)
+        return VkError.ValidationFailed;
+
+    const src_row_pitch = src_interface.getRowPitchMemSizeForMipLevel(region.src_subresource.aspect_mask, region.src_subresource.mip_level);
+    const src_slice_pitch = src_interface.getSliceMemSizeForMipLevel(region.src_subresource.aspect_mask, region.src_subresource.mip_level);
+    const src_sample_pitch = src.getMipLevelSize(region.src_subresource.aspect_mask, region.src_subresource.mip_level);
+    const dst_row_pitch = dst_interface.getRowPitchMemSizeForMipLevel(region.dst_subresource.aspect_mask, region.dst_subresource.mip_level);
+    const dst_slice_pitch = dst_interface.getSliceMemSizeForMipLevel(region.dst_subresource.aspect_mask, region.dst_subresource.mip_level);
+    const dst_sample_pitch = dst.getMipLevelSize(region.dst_subresource.aspect_mask, region.dst_subresource.mip_level);
+    const copy_row_size = copy_blocks_x * bytes_per_block;
+
+    for (0..src_layer_count) |layer| {
+        const src_subresource_offset = try src.getSubresourceOffset(
+            region.src_subresource.aspect_mask,
+            region.src_subresource.mip_level,
+            region.src_subresource.base_array_layer + @as(u32, @intCast(layer)),
+        );
+        const dst_subresource_offset = try dst.getSubresourceOffset(
+            region.dst_subresource.aspect_mask,
+            region.dst_subresource.mip_level,
+            region.dst_subresource.base_array_layer + @as(u32, @intCast(layer)),
+        );
+
+        for (0..src_interface.samples.toInt()) |sample| {
+            for (0..region.extent.depth) |z| {
+                for (0..copy_blocks_y) |row| {
+                    const src_offset = src_subresource_offset +
+                        sample * src_sample_pitch +
+                        (src_z + z) * src_slice_pitch +
+                        (src_block_y + row) * src_row_pitch +
+                        src_block_x * bytes_per_block;
+                    const dst_offset = dst_subresource_offset +
+                        sample * dst_sample_pitch +
+                        (dst_z + z) * dst_slice_pitch +
+                        (dst_block_y + row) * dst_row_pitch +
+                        dst_block_x * bytes_per_block;
+                    try emitLinearCopy(
+                        cmd,
+                        try MemoryRange.fromImage(src_interface, src_offset, copy_row_size),
+                        try MemoryRange.fromImage(dst_interface, dst_offset, copy_row_size),
+                    );
+                }
+            }
+        }
+    }
+}
+
+fn resolveLayerCount(image: *const base.Image, subresource: vk.ImageSubresourceLayers) VkError!u32 {
+    if (subresource.base_array_layer >= image.array_layers)
+        return VkError.ValidationFailed;
+    const layer_count = if (subresource.layer_count == vk.REMAINING_ARRAY_LAYERS)
+        image.array_layers - subresource.base_array_layer
+    else
+        subresource.layer_count;
+    if (layer_count == 0 or layer_count > image.array_layers - subresource.base_array_layer)
+        return VkError.ValidationFailed;
+    return layer_count;
+}
+
+pub fn copyRangeFromRegion(buffer: *base.Buffer, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!MemoryRange {
+    return MemoryRange.fromBuffer(buffer, offset, size);
+}
+
+pub fn fillRange(buffer: *base.Buffer, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!MemoryRange {
+    const resolved_size = if (size == vk.WHOLE_SIZE) blk: {
+        if (offset > buffer.size) return VkError.ValidationFailed;
+        break :blk std.mem.alignBackward(vk.DeviceSize, buffer.size - offset, @sizeOf(u32));
+    } else blk: {
+        if (size % @sizeOf(u32) != 0) return VkError.ValidationFailed;
+        break :blk size;
+    };
+
+    return MemoryRange.fromBuffer(buffer, offset, resolved_size);
+}

src/intel/i915/i915.zig

@@ -0,0 +1,80 @@
+pub const command_base = 0x40;
+pub const gem_create = 0x1b;
+pub const gem_mmap_gtt = 0x24;
+pub const gem_set_domain = 0x1f;
+pub const gem_execbuffer2 = 0x29;
+pub const gem_close = 0x09;
+
+pub const mmap_offset_wb = 2;
+pub const gem_domain_cpu = 0x00000001;
+pub const gem_domain_gtt = 0x00000040;
+pub const exec_blt = 3 << 0;
+pub const exec_fence_array: u64 = 1 << 19;
+pub const exec_fence_wait: u32 = 1 << 0;
+pub const exec_fence_signal: u32 = 1 << 1;
+pub const exec_object_write = 1 << 2;
+pub const mi_flush_dw: u32 = (0x26 << 23) | 3;
+
+pub const GemCreate = extern struct {
+    size: u64,
+    handle: u32,
+    pad: u32,
+};
+
+pub const GemMmapOffset = extern struct {
+    handle: u32,
+    pad: u32,
+    offset: u64,
+    flags: u64,
+    extensions: u64,
+};
+
+pub const GemClose = extern struct {
+    handle: u32,
+    pad: u32,
+};
+
+pub const GemSetDomain = extern struct {
+    handle: u32,
+    read_domains: u32,
+    write_domain: u32,
+};
+
+pub const RelocationEntry = extern struct {
+    target_handle: u32,
+    delta: u32,
+    offset: u64,
+    presumed_offset: u64,
+    read_domains: u32,
+    write_domain: u32,
+};
+
+pub const ExecObject2 = extern struct {
+    handle: u32,
+    relocation_count: u32,
+    relocs_ptr: u64,
+    alignment: u64,
+    offset: u64,
+    flags: u64,
+    rsvd1: u64,
+    rsvd2: u64,
+};
+
+pub const ExecBuffer2 = extern struct {
+    buffers_ptr: u64,
+    buffer_count: u32,
+    batch_start_offset: u32,
+    batch_len: u32,
+    DR1: u32,
+    DR4: u32,
+    num_cliprects: u32,
+    cliprects_ptr: u64,
+    flags: u64,
+    rsvd1: u64,
+    rsvd2: u64,
+};
+
+pub const ExecFence = extern struct {
+    handle: u32,
+    flags: u32,
+};

src/intel/i915/kmd.zig

@@ -0,0 +1,247 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const _i915 = @import("i915.zig");
+const common_kmd = @import("../kmd.zig");
+
+const VkError = base.VkError;
+
+const Mapping = struct {
+    bytes: []align(std.heap.page_size_min) u8,
+
+    inline fn slice(self: Mapping, offset: vk.DeviceSize, size: vk.DeviceSize) VkError![]u8 {
+        const start: usize = @intCast(offset);
+        const len: usize = @intCast(size);
+        return self.bytes[start .. start + len];
+    }
+};
+
+pub const Device = struct {
+    card: base.drm.Card,
+
+    pub fn open(io: std.Io, node_path: []const u8) VkError!Device {
+        return .{
+            .card = base.drm.Card.open(io, node_path) catch return VkError.InitializationFailed,
+        };
+    }
+
+    pub fn close(self: *Device, io: std.Io) void {
+        self.card.close(io);
+    }
+
+    pub fn allocateMemory(self: *Device, io: std.Io, size: vk.DeviceSize) VkError!Memory {
+        var create = _i915.GemCreate{
+            .size = size,
+            .handle = 0,
+            .pad = 0,
+        };
+        base.utils.ioctl(
+            self.card.handle,
+            io,
+            common_kmd.drmIoctlIowr(_i915.command_base + _i915.gem_create, _i915.GemCreate),
+            &create,
+        ) catch return VkError.OutOfDeviceMemory;
+
+        var memory = Memory{
+            .handle = create.handle,
+            .size = create.size,
+            .mapping = null,
+        };
+        errdefer memory.deinit(self, io);
+
+        try memory.setDomain(self, io, _i915.gem_domain_cpu, 0);
+        return memory;
+    }
+
+    pub fn submitBatch(self: *Device, io: std.Io, allocator: std.mem.Allocator, commands: []const u32, relocations: []const common_kmd.Relocation, syncs: []const common_kmd.SyncDependency) VkError!void {
+        const trailer_words = 6;
+        const batch_size = (commands.len + trailer_words) * @sizeOf(u32);
+        var batch = try self.allocateMemory(io, batch_size);
+        defer batch.deinit(self, io);
+
+        {
+            const batch_map = try batch.map(self, io, 0, batch_size);
+            const batch_words = std.mem.bytesAsSlice(u32, batch_map);
+            @memcpy(batch_words[0..commands.len], commands);
+            batch_words[commands.len + 0] = _i915.mi_flush_dw;
+            batch_words[commands.len + 1] = 0;
+            batch_words[commands.len + 2] = 0;
+            batch_words[commands.len + 3] = 0;
+            batch_words[commands.len + 4] = 0;
+            batch_words[commands.len + 5] = 0x05000000;
+            batch.unmap();
+        }
+        try batch.flushRange(self, io, 0, batch_size);
+
+        var objects = std.ArrayList(_i915.ExecObject2).empty;
+        defer objects.deinit(allocator);
+
+        var object_handles = std.ArrayList(u32).empty;
+        defer object_handles.deinit(allocator);
+
+        for (relocations) |relocation| {
+            if (std.mem.indexOfScalar(u32, object_handles.items, relocation.target_handle) == null) {
+                object_handles.append(allocator, relocation.target_handle) catch return VkError.OutOfHostMemory;
+                objects.append(allocator, .{
+                    .handle = relocation.target_handle,
+                    .relocation_count = 0,
+                    .relocs_ptr = 0,
+                    .alignment = 0,
+                    .offset = 0,
+                    .flags = if (relocation.write) _i915.exec_object_write else 0,
+                    .rsvd1 = 0,
+                    .rsvd2 = 0,
+                }) catch return VkError.OutOfHostMemory;
+            } else if (relocation.write) {
+                const index = std.mem.indexOfScalar(u32, object_handles.items, relocation.target_handle).?;
+                objects.items[index].flags |= _i915.exec_object_write;
+            }
+        }
+
+        var i915_relocations = std.ArrayList(_i915.RelocationEntry).empty;
+        defer i915_relocations.deinit(allocator);
+
+        for (relocations) |relocation| {
+            i915_relocations.append(allocator, .{
+                .target_handle = relocation.target_handle,
+                .delta = relocation.delta,
+                .offset = relocation.offset,
+                .presumed_offset = 0,
+                .read_domains = 0,
+                .write_domain = 0,
+            }) catch return VkError.OutOfHostMemory;
+        }
+
+        objects.append(allocator, .{
+            .handle = batch.handle,
+            .relocation_count = @intCast(i915_relocations.items.len),
+            .relocs_ptr = @intFromPtr(i915_relocations.items.ptr),
+            .alignment = 0,
+            .offset = 0,
+            .flags = 0,
+            .rsvd1 = 0,
+            .rsvd2 = 0,
+        }) catch return VkError.OutOfHostMemory;
+
+        var exec_fences = std.ArrayList(_i915.ExecFence).empty;
+        defer exec_fences.deinit(allocator);
+        for (syncs) |sync| {
+            exec_fences.append(allocator, .{
+                .handle = sync.handle,
+                .flags = (if (sync.wait) _i915.exec_fence_wait else 0) | (if (sync.signal) _i915.exec_fence_signal else 0),
+            }) catch return VkError.OutOfHostMemory;
+        }
+
+        var execbuffer = _i915.ExecBuffer2{
+            .buffers_ptr = @intFromPtr(objects.items.ptr),
+            .buffer_count = @intCast(objects.items.len),
+            .batch_start_offset = 0,
+            .batch_len = @intCast(batch_size),
+            .DR1 = 0,
+            .DR4 = 0,
+            .num_cliprects = @intCast(exec_fences.items.len),
+            .cliprects_ptr = if (exec_fences.items.len == 0) 0 else @intFromPtr(exec_fences.items.ptr),
+            .flags = _i915.exec_blt | (if (exec_fences.items.len == 0) 0 else _i915.exec_fence_array),
+            .rsvd1 = 0,
+            .rsvd2 = 0,
+        };
+        base.utils.ioctl(
+            self.card.handle,
+            io,
+            common_kmd.drmIoctlIowr(_i915.command_base + _i915.gem_execbuffer2, _i915.ExecBuffer2),
+            &execbuffer,
+        ) catch return VkError.DeviceLost;
+    }
+};
+
+pub const Memory = struct {
+    handle: u32,
+    size: vk.DeviceSize,
+    mapping: ?Mapping,
+
+    pub fn deinit(self: *Memory, device: *Device, io: std.Io) void {
+        self.unmap();
+
+        var close = _i915.GemClose{
+            .handle = self.handle,
+            .pad = 0,
+        };
+        base.utils.ioctl(device.card.handle, io, common_kmd.drmIoctlIow(_i915.gem_close, _i915.GemClose), &close) catch @panic("Caught an error while handling an error");
+
+        self.* = undefined;
+    }
+
+    pub fn map(self: *Memory, device: *Device, io: std.Io, offset: vk.DeviceSize, size: vk.DeviceSize) VkError![]u8 {
+        if (offset > self.size) return VkError.MemoryMapFailed;
+        const available = self.size - offset;
+        const map_size = if (size == vk.WHOLE_SIZE) available else size;
+        if (map_size > available) return VkError.MemoryMapFailed;
+        if (map_size > std.math.maxInt(usize)) return VkError.MemoryMapFailed;
+
+        if (self.mapping) |mapping| {
+            return mapping.slice(offset, map_size);
+        }
+
+        var mmap_offset = _i915.GemMmapOffset{
+            .handle = self.handle,
+            .pad = 0,
+            .offset = 0,
+            .flags = _i915.mmap_offset_wb,
+            .extensions = 0,
+        };
+        base.utils.ioctl(
+            device.card.handle,
+            io,
+            common_kmd.drmIoctlIowr(_i915.command_base + _i915.gem_mmap_gtt, _i915.GemMmapOffset),
+            &mmap_offset,
+        ) catch return VkError.MemoryMapFailed;
+
+        if (self.size > std.math.maxInt(usize)) return VkError.MemoryMapFailed;
+        const full_size: usize = @intCast(self.size);
+        const bytes = std.posix.mmap(
+            null,
+            full_size,
+            .{ .READ = true, .WRITE = true },
+            .{ .TYPE = .SHARED },
+            device.card.handle.handle,
+            @intCast(mmap_offset.offset),
+        ) catch return VkError.MemoryMapFailed;
+
+        self.mapping = .{ .bytes = bytes };
+        return self.mapping.?.slice(offset, map_size);
+    }
+
+    pub fn unmap(self: *Memory) void {
+        if (self.mapping) |mapping| {
+            std.posix.munmap(mapping.bytes);
+            self.mapping = null;
+        }
+    }
+
+    pub fn flushRange(self: *Memory, device: *Device, io: std.Io, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
+        _ = offset;
+        _ = size;
+        try self.setDomain(device, io, _i915.gem_domain_cpu, 0);
+    }
+
+    pub fn invalidateRange(self: *Memory, device: *Device, io: std.Io, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
+        _ = offset;
+        _ = size;
+        try self.setDomain(device, io, _i915.gem_domain_cpu, 0);
+    }
+
+    fn setDomain(self: *Memory, device: *Device, io: std.Io, read_domains: u32, write_domain: u32) VkError!void {
+        var domain = _i915.GemSetDomain{
+            .handle = self.handle,
+            .read_domains = read_domains,
+            .write_domain = write_domain,
+        };
+        base.utils.ioctl(
+            device.card.handle,
+            io,
+            common_kmd.drmIoctlIow(_i915.command_base + _i915.gem_set_domain, _i915.GemSetDomain),
+            &domain,
+        ) catch return VkError.DeviceLost;
+    }
+};

src/intel/kmd.zig

@@ -0,0 +1,168 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const KmdType = @import("lib.zig").KmdType;
+
+const FlintPhysicalDevice = @import("FlintPhysicalDevice.zig");
+const i915_kmd = @import("i915/kmd.zig");
+const xe = @import("xe/kmd.zig");
+
+const VkError = base.VkError;
+const ioctl = std.os.linux.IOCTL;
+
+pub const xy_src_copy_blt: u32 = (2 << 29) | (0x53 << 22) | 8;
+pub const xy_blt_write_alpha: u32 = 1 << 21;
+pub const xy_blt_write_rgb: u32 = 1 << 20;
+pub const mi_store_data_imm_dword: u32 = (0x20 << 23) | 2;
+pub const blt_depth_8: u32 = 0 << 24;
+pub const rop_source_copy: u32 = 0xcc << 16;
+pub const max_blt_span: vk.DeviceSize = 32 * 1024 - 1;
+
+pub const Relocation = struct {
+    target_handle: u32,
+    offset: u64,
+    delta: u32,
+    read: bool = false,
+    write: bool = false,
+};
+
+pub const SyncDependency = struct {
+    handle: u32,
+    wait: bool = false,
+    signal: bool = false,
+};
+
+pub const Device = union(KmdType) {
+    invalid: void,
+    i915: i915_kmd.Device,
+    xe: xe.Device,
+
+    pub fn open(io: std.Io, physical_device: *const FlintPhysicalDevice) VkError!Device {
+        return switch (physical_device.kmd_type) {
+            .i915 => .{ .i915 = try i915_kmd.Device.open(io, physical_device.getNodePath()) },
+            .xe => .{ .xe = try xe.Device.open(io, physical_device.getNodePath()) },
+            .invalid => VkError.InitializationFailed,
+        };
+    }
+
+    pub fn close(self: *Device, io: std.Io) void {
+        switch (self.*) {
+            .i915 => |*device| device.close(io),
+            .xe => |*device| device.close(io),
+            .invalid => {},
+        }
+        self.* = .{ .invalid = {} };
+    }
+
+    pub fn allocateMemory(self: *Device, io: std.Io, size: vk.DeviceSize) VkError!Memory {
+        return switch (self.*) {
+            .i915 => |*device| .{ .i915 = try device.allocateMemory(io, size) },
+            .xe => |*device| .{ .xe = try device.allocateMemory(io, size) },
+            .invalid => VkError.OutOfDeviceMemory,
+        };
+    }
+
+    pub fn submitBatch(self: *Device, io: std.Io, allocator: std.mem.Allocator, commands: []const u32, relocations: []const Relocation, syncs: []const SyncDependency) VkError!void {
+        return switch (self.*) {
+            .i915 => |*device| device.submitBatch(io, allocator, commands, relocations, syncs),
+            .xe => |*device| device.submitBatch(io, allocator, commands, relocations, syncs),
+            .invalid => VkError.DeviceLost,
+        };
+    }
+
+    pub fn file(self: *Device) VkError!std.Io.File {
+        return switch (self.*) {
+            .i915 => |*device| device.card.handle,
+            .xe => |*device| device.card.handle,
+            .invalid => VkError.DeviceLost,
+        };
+    }
+};
+
+pub const Memory = union(KmdType) {
+    invalid: void,
+    i915: i915_kmd.Memory,
+    xe: xe.Memory,
+
+    pub fn deinit(self: *Memory, device: *Device, io: std.Io) void {
+        switch (self.*) {
+            .i915 => |*memory| switch (device.*) {
+                .i915 => |*adapter| memory.deinit(adapter, io),
+                else => {},
+            },
+            .xe => |*memory| switch (device.*) {
+                .xe => |*adapter| memory.deinit(adapter, io),
+                else => {},
+            },
+            .invalid => {},
+        }
+        self.* = .{ .invalid = {} };
+    }
+
+    pub fn map(self: *Memory, device: *Device, io: std.Io, offset: vk.DeviceSize, size: vk.DeviceSize) VkError![]u8 {
+        return switch (self.*) {
+            .i915 => |*memory| switch (device.*) {
+                .i915 => |*adapter| memory.map(adapter, io, offset, size),
+                else => VkError.MemoryMapFailed,
+            },
+            .xe => |*memory| switch (device.*) {
+                .xe => |*adapter| memory.map(adapter, io, offset, size),
+                else => VkError.MemoryMapFailed,
+            },
+            .invalid => VkError.MemoryMapFailed,
+        };
+    }
+
+    pub fn unmap(self: *Memory) void {
+        switch (self.*) {
+            .i915 => |*memory| memory.unmap(),
+            .xe => |*memory| memory.unmap(),
+            .invalid => {},
+        }
+    }
+
+    pub fn flushRange(self: *Memory, device: *Device, io: std.Io, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
+        return switch (self.*) {
+            .i915 => |*memory| switch (device.*) {
+                .i915 => |*adapter| memory.flushRange(adapter, io, offset, size),
+                else => VkError.InvalidDeviceMemoryDrv,
+            },
+            .xe => |*memory| switch (device.*) {
+                .xe => |*adapter| memory.flushRange(adapter, io, offset, size),
+                else => VkError.InvalidDeviceMemoryDrv,
+            },
+            .invalid => VkError.InvalidDeviceMemoryDrv,
+        };
+    }
+
+    pub fn invalidateRange(self: *Memory, device: *Device, io: std.Io, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
+        return switch (self.*) {
+            .i915 => |*memory| switch (device.*) {
+                .i915 => |*adapter| memory.invalidateRange(adapter, io, offset, size),
+                else => VkError.InvalidDeviceMemoryDrv,
+            },
+            .xe => |*memory| switch (device.*) {
+                .xe => |*adapter| memory.invalidateRange(adapter, io, offset, size),
+                else => VkError.InvalidDeviceMemoryDrv,
+            },
+            .invalid => VkError.InvalidDeviceMemoryDrv,
+        };
+    }
+
+    pub fn handle(self: *const Memory) VkError!u32 {
+        return switch (self.*) {
+            .i915 => |*memory| memory.handle,
+            .xe => VkError.FeatureNotPresent,
+            .invalid => VkError.InvalidDeviceMemoryDrv,
+        };
+    }
+};
+
+pub inline fn drmIoctlIow(nr: u8, comptime T: type) u32 {
+    return ioctl.IOW('d', nr, T);
+}
+
+pub inline fn drmIoctlIowr(nr: u8, comptime T: type) u32 {
+    return ioctl.IOWR('d', nr, T);
+}

src/intel/lib.zig

@@ -9,6 +9,7 @@ pub const FlintInstance = @import("FlintInstance.zig");
 pub const FlintDevice = @import("FlintDevice.zig");
 pub const FlintPhysicalDevice = @import("FlintPhysicalDevice.zig");
 pub const FlintQueue = @import("FlintQueue.zig");
+pub const kmd = @import("kmd.zig");
 
 pub const FlintBinarySemaphore = @import("FlintBinarySemaphore.zig");
 pub const FlintBuffer = @import("FlintBuffer.zig");
@@ -34,23 +35,26 @@ pub const FlintShaderModule = @import("FlintShaderModule.zig");
 
 pub const Instance = FlintInstance;
 
-pub const DRIVER_NAME = "Flint";
+pub const driver_name = "Flint";
 
-pub const PHYSICAL_DEVICE_DEFAULT_NAME = "Unkown Intel device";
+pub const physical_device_default_name = "Unkown Intel device";
 
-pub const INTEL_PCI_VENDOR_ID = 0x8086;
+pub const intel_pci_vendor_id = 0x8086;
 
-pub const VULKAN_VERSION = vk.makeApiVersion(
+pub const vulkan_version = vk.makeApiVersion(
     0,
     config.flint_vulkan_version.major,
     config.flint_vulkan_version.minor,
     config.flint_vulkan_version.patch,
 );
 
+/// GEM buffer objects are page based
+pub const image_memory_alignment = std.heap.page_size_max;
+
 pub const KmdType = enum {
-    Invalid,
-    I915,
-    Xe,
+    invalid,
+    i915,
+    xe,
 };
 
 pub const std_options = base.std_options;
@@ -85,5 +89,6 @@ test {
     std.testing.refAllDecls(FlintRenderPass);
     std.testing.refAllDecls(FlintSampler);
     std.testing.refAllDecls(FlintShaderModule);
+    std.testing.refAllDecls(kmd);
     std.testing.refAllDecls(base);
 }

src/intel/pci_ids.zig

@@ -1,5 +1,3 @@
-const std = @import("std");
-
 const PciInfo = struct {
     id: u16,
     name: []const u8,

src/intel/xe/kmd.zig

@@ -0,0 +1,47 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const common_kmd = @import("../kmd.zig");
+
+const VkError = base.VkError;
+
+pub const Device = struct {
+    card: base.drm.Card,
+
+    pub fn open(io: std.Io, node_path: []const u8) VkError!Device {
+        return .{
+            .card = base.drm.Card.open(io, node_path) catch return VkError.InitializationFailed,
+        };
+    }
+
+    pub fn close(self: *Device, io: std.Io) void {
+        self.card.close(io);
+    }
+
+    pub fn allocateMemory(_: *Device, _: std.Io, _: vk.DeviceSize) VkError!Memory {
+        return VkError.OutOfDeviceMemory;
+    }
+
+    pub fn submitBatch(_: *Device, _: std.Io, _: std.mem.Allocator, _: []const u32, _: []const common_kmd.Relocation, _: []const common_kmd.SyncDependency) VkError!void {
+        return VkError.FeatureNotPresent;
+    }
+};
+
+pub const Memory = struct {
+    pub fn deinit(_: *Memory, _: *Device, _: std.Io) void {}
+
+    pub fn map(_: *Memory, _: *Device, _: std.Io, _: vk.DeviceSize, _: vk.DeviceSize) VkError![]u8 {
+        return VkError.MemoryMapFailed;
+    }
+
+    pub fn unmap(_: *Memory) void {}
+
+    pub fn flushRange(_: *Memory, _: *Device, _: std.Io, _: vk.DeviceSize, _: vk.DeviceSize) VkError!void {
+        return VkError.FeatureNotPresent;
+    }
+
+    pub fn invalidateRange(_: *Memory, _: *Device, _: std.Io, _: vk.DeviceSize, _: vk.DeviceSize) VkError!void {
+        return VkError.FeatureNotPresent;
+    }
+};

src/phi/PhiBinarySemaphore.zig

@@ -0,0 +1,43 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.BinarySemaphore;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.SemaphoreCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+        .signal = signal,
+        .wait = wait,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn signal(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn wait(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}

src/phi/PhiBuffer.zig

@@ -0,0 +1,38 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Buffer;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.BufferCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+        .getMemoryRequirements = getMemoryRequirements,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn getMemoryRequirements(interface: *Interface, requirements: *vk.MemoryRequirements) void {
+    _ = interface;
+    _ = requirements;
+}

src/phi/PhiBufferView.zig

@@ -0,0 +1,31 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.BufferView;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.BufferViewCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiCommandBuffer.zig

@@ -0,0 +1,516 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+const lib = @import("lib.zig");
+const proto = lib.proto;
+
+const VkError = base.VkError;
+const PhiDeviceMemory = @import("PhiDeviceMemory.zig");
+
+const Self = @This();
+pub const Interface = base.CommandBuffer;
+
+interface: Interface,
+cmd_count: usize,
+serialized_cmd_count: usize,
+commands: std.ArrayList(u8),
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.CommandBufferAllocateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+    interface.vtable = &.{ .destroy = destroy };
+    interface.dispatch_table = &.{
+        .begin = begin,
+        .beginQuery = beginQuery,
+        .beginRenderPass = beginRenderPass,
+        .bindDescriptorSets = bindDescriptorSets,
+        .bindPipeline = bindPipeline,
+        .bindIndexBuffer = bindIndexBuffer,
+        .bindVertexBuffer = bindVertexBuffer,
+        .blitImage = blitImage,
+        .clearAttachment = clearAttachment,
+        .clearColorImage = clearColorImage,
+        .clearDepthStencilImage = clearDepthStencilImage,
+        .copyBuffer = copyBuffer,
+        .copyBufferToImage = copyBufferToImage,
+        .copyImage = copyImage,
+        .copyImageToBuffer = copyImageToBuffer,
+        .copyQueryPoolResults = copyQueryPoolResults,
+        .dispatch = dispatch,
+        .dispatchBase = dispatchBase,
+        .dispatchIndirect = dispatchIndirect,
+        .draw = draw,
+        .drawIndexed = drawIndexed,
+        .drawIndexedIndirect = drawIndexedIndirect,
+        .drawIndirect = drawIndirect,
+        .end = end,
+        .endQuery = endQuery,
+        .endRenderPass = endRenderPass,
+        .executeCommands = executeCommands,
+        .fillBuffer = fillBuffer,
+        .nextSubpass = nextSubpass,
+        .pipelineBarrier = pipelineBarrier,
+        .pushConstants = pushConstants,
+        .reset = reset,
+        .resetQueryPool = resetQueryPool,
+        .resetEvent = resetEvent,
+        .resolveImage = resolveImage,
+        .setEvent = setEvent,
+        .setBlendConstants = setBlendConstants,
+        .setDepthBias = setDepthBias,
+        .setDepthBounds = setDepthBounds,
+        .setDeviceMask = setDeviceMask,
+        .setLineWidth = setLineWidth,
+        .setScissor = setScissor,
+        .setStencilCompareMask = setStencilCompareMask,
+        .setStencilReference = setStencilReference,
+        .setStencilWriteMask = setStencilWriteMask,
+        .setViewport = setViewport,
+        .updateBuffer = updateBuffer,
+        .waitEvent = waitEvent,
+        .writeTimestamp = writeTimestamp,
+    };
+
+    self.* = .{
+        .interface = interface,
+        .cmd_count = 0,
+        .serialized_cmd_count = 0,
+        .commands = .empty,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.commands.deinit(allocator);
+    allocator.destroy(self);
+}
+
+pub fn begin(interface: *Interface, info: *const vk.CommandBufferBeginInfo) VkError!void {
+    _ = interface;
+    _ = info;
+}
+
+pub fn end(interface: *Interface) VkError!void {
+    _ = interface;
+}
+
+pub fn reset(interface: *Interface, flags: vk.CommandBufferResetFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count = 0;
+    self.serialized_cmd_count = 0;
+    self.commands.clearRetainingCapacity();
+    _ = flags;
+}
+
+fn appendCommand(self: *Self, comptime T: type, command_type: c_int, payload: T) VkError!void {
+    const allocator = self.interface.host_allocator.allocator();
+    const header: proto.PhiCmdHeader = .{
+        .magic = proto.PHI_COMMAND_MAGIC,
+        .type = @intCast(command_type),
+    };
+    self.commands.appendSlice(allocator, std.mem.asBytes(&header)) catch return VkError.OutOfHostMemory;
+    self.commands.appendSlice(allocator, std.mem.asBytes(&payload)) catch return VkError.OutOfHostMemory;
+    self.cmd_count += 1;
+    self.serialized_cmd_count += 1;
+}
+
+fn remoteMemory(buffer: *base.Buffer) VkError!*PhiDeviceMemory {
+    const memory = buffer.memory orelse return VkError.ValidationFailed;
+    const phi_memory: *PhiDeviceMemory = @alignCast(@fieldParentPtr("interface", memory));
+    if (phi_memory.remote_handle == 0) {
+        return VkError.ValidationFailed;
+    }
+    return phi_memory;
+}
+
+pub fn beginQuery(interface: *Interface, pool: *base.QueryPool, query: u32, flags: vk.QueryControlFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = flags;
+    try pool.begin(query);
+}
+
+pub fn endQuery(interface: *Interface, pool: *base.QueryPool, query: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    try pool.end(query);
+}
+
+pub fn resetQueryPool(interface: *Interface, pool: *base.QueryPool, first: u32, count: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    try pool.reset(first, count);
+}
+
+pub fn beginRenderPass(interface: *Interface, render_pass: *base.RenderPass, framebuffer: *base.Framebuffer, render_area: vk.Rect2D, clear_values: ?[]const vk.ClearValue) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = render_pass;
+    _ = framebuffer;
+    _ = render_area;
+    _ = clear_values;
+}
+
+pub fn bindDescriptorSets(interface: *Interface, bind_point: vk.PipelineBindPoint, first_set: u32, sets: [base.vulkan_max_descriptor_sets]?*base.DescriptorSet, dynamic_offsets: []const u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = bind_point;
+    _ = first_set;
+    _ = sets;
+    _ = dynamic_offsets;
+}
+
+pub fn bindPipeline(interface: *Interface, bind_point: vk.PipelineBindPoint, pipeline: *base.Pipeline) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = bind_point;
+    _ = pipeline;
+}
+
+pub fn bindIndexBuffer(interface: *Interface, buffer: *base.Buffer, offset: usize, index_type: vk.IndexType) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = buffer;
+    _ = offset;
+    _ = index_type;
+}
+
+pub fn bindVertexBuffer(interface: *Interface, index: usize, buffer: *base.Buffer, offset: usize) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = index;
+    _ = buffer;
+    _ = offset;
+}
+
+pub fn blitImage(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Image, dst_layout: vk.ImageLayout, regions: []const vk.ImageBlit, filter: vk.Filter) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = src;
+    _ = src_layout;
+    _ = dst;
+    _ = dst_layout;
+    _ = regions;
+    _ = filter;
+}
+
+pub fn clearAttachment(interface: *Interface, attachment: vk.ClearAttachment, rect: vk.ClearRect) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = attachment;
+    _ = rect;
+}
+
+pub fn clearColorImage(interface: *Interface, image: *base.Image, layout: vk.ImageLayout, color: *const vk.ClearColorValue, range: vk.ImageSubresourceRange) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = image;
+    _ = layout;
+    _ = color;
+    _ = range;
+}
+
+pub fn clearDepthStencilImage(interface: *Interface, image: *base.Image, layout: vk.ImageLayout, value: *const vk.ClearDepthStencilValue, range: vk.ImageSubresourceRange) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = image;
+    _ = layout;
+    _ = value;
+    _ = range;
+}
+
+pub fn copyBuffer(interface: *Interface, src: *base.Buffer, dst: *base.Buffer, regions: []const vk.BufferCopy) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const src_memory = try remoteMemory(src);
+    const dst_memory = try remoteMemory(dst);
+
+    for (regions) |region| {
+        const src_offset, const src_overflow = @addWithOverflow(src.offset, region.src_offset);
+        const dst_offset, const dst_overflow = @addWithOverflow(dst.offset, region.dst_offset);
+        if (src_overflow != 0 or dst_overflow != 0) {
+            return VkError.ValidationFailed;
+        }
+
+        try self.appendCommand(proto.PhiCmdCopyBuffer, proto.PHI_CMD_COPY_BUFFER, .{
+            .size = region.size,
+            .src_memory = @intCast(src_memory.remote_handle),
+            .dst_memory = @intCast(dst_memory.remote_handle),
+            .src_offset = src_offset,
+            .dst_offset = dst_offset,
+        });
+    }
+}
+
+pub fn copyBufferToImage(interface: *Interface, src: *base.Buffer, dst: *base.Image, dst_layout: vk.ImageLayout, regions: []const vk.BufferImageCopy) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = src;
+    _ = dst;
+    _ = dst_layout;
+    _ = regions;
+}
+
+pub fn copyImage(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Image, dst_layout: vk.ImageLayout, regions: []const vk.ImageCopy) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = src;
+    _ = src_layout;
+    _ = dst;
+    _ = dst_layout;
+    _ = regions;
+}
+
+pub fn copyImageToBuffer(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Buffer, regions: []const vk.BufferImageCopy) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = src;
+    _ = src_layout;
+    _ = dst;
+    _ = regions;
+}
+
+pub fn copyQueryPoolResults(interface: *Interface, pool: *base.QueryPool, first: u32, count: u32, dst: *base.Buffer, offset: vk.DeviceSize, stride: vk.DeviceSize, flags: vk.QueryResultFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = pool;
+    _ = first;
+    _ = count;
+    _ = dst;
+    _ = offset;
+    _ = stride;
+    _ = flags;
+}
+
+pub fn dispatch(interface: *Interface, group_count_x: u32, group_count_y: u32, group_count_z: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = group_count_x;
+    _ = group_count_y;
+    _ = group_count_z;
+}
+
+pub fn dispatchBase(interface: *Interface, base_group_x: u32, base_group_y: u32, base_group_z: u32, group_count_x: u32, group_count_y: u32, group_count_z: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = base_group_x;
+    _ = base_group_y;
+    _ = base_group_z;
+    _ = group_count_x;
+    _ = group_count_y;
+    _ = group_count_z;
+}
+
+pub fn setDeviceMask(interface: *Interface, device_mask: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = device_mask;
+}
+
+pub fn dispatchIndirect(interface: *Interface, buffer: *base.Buffer, offset: vk.DeviceSize) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = buffer;
+    _ = offset;
+}
+
+pub fn draw(interface: *Interface, vertex_count: usize, instance_count: usize, first_vertex: usize, first_instance: usize) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = vertex_count;
+    _ = instance_count;
+    _ = first_vertex;
+    _ = first_instance;
+}
+
+pub fn drawIndexed(interface: *Interface, index_count: usize, instance_count: usize, first_index: usize, vertex_offset: i32, first_instance: usize) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = index_count;
+    _ = instance_count;
+    _ = first_index;
+    _ = vertex_offset;
+    _ = first_instance;
+}
+
+pub fn drawIndexedIndirect(interface: *Interface, buffer: *base.Buffer, offset: usize, count: usize, stride: usize) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = buffer;
+    _ = offset;
+    _ = count;
+    _ = stride;
+}
+
+pub fn drawIndirect(interface: *Interface, buffer: *base.Buffer, offset: usize, count: usize, stride: usize) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = buffer;
+    _ = offset;
+    _ = count;
+    _ = stride;
+}
+
+pub fn endRenderPass(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+}
+
+pub fn executeCommands(interface: *Interface, commands: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = commands;
+}
+
+pub fn fillBuffer(interface: *Interface, buffer: *base.Buffer, offset: vk.DeviceSize, size: vk.DeviceSize, data: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+
+    const memory = try remoteMemory(buffer);
+
+    try self.appendCommand(proto.PhiCmdFillBuffer, proto.PHI_CMD_FILL_BUFFER, .{
+        .size = if (size == vk.WHOLE_SIZE) buffer.size - offset else size,
+        .memory = @intCast(memory.remote_handle),
+        .offset = offset,
+        .data = data,
+    });
+}
+
+pub fn updateBuffer(interface: *Interface, buffer: *base.Buffer, offset: vk.DeviceSize, data: []const u8) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = buffer;
+    _ = offset;
+    _ = data;
+}
+
+pub fn nextSubpass(interface: *Interface, contents: vk.SubpassContents) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = contents;
+}
+
+pub fn pipelineBarrier(interface: *Interface, src_stage: vk.PipelineStageFlags, dst_stage: vk.PipelineStageFlags, dependency: vk.DependencyFlags, memory: []const vk.MemoryBarrier, buffers: []const vk.BufferMemoryBarrier, images: []const vk.ImageMemoryBarrier) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = src_stage;
+    _ = dst_stage;
+    _ = dependency;
+    _ = memory;
+    _ = buffers;
+    _ = images;
+}
+
+pub fn pushConstants(interface: *Interface, stages: vk.ShaderStageFlags, offset: u32, blob: []const u8) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = stages;
+    _ = offset;
+    _ = blob;
+}
+
+pub fn resetEvent(interface: *Interface, event: *base.Event, stage: vk.PipelineStageFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = stage;
+    try event.reset();
+}
+
+pub fn resolveImage(interface: *Interface, src: *base.Image, src_layout: vk.ImageLayout, dst: *base.Image, dst_layout: vk.ImageLayout, region: vk.ImageResolve) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = src;
+    _ = src_layout;
+    _ = dst;
+    _ = dst_layout;
+    _ = region;
+}
+
+pub fn setEvent(interface: *Interface, event: *base.Event, stage: vk.PipelineStageFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = stage;
+    try event.signal();
+}
+
+pub fn setScissor(interface: *Interface, first: u32, scissor: []const vk.Rect2D) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = first;
+    _ = scissor;
+}
+
+pub fn setViewport(interface: *Interface, first: u32, viewports: []const vk.Viewport) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = first;
+    _ = viewports;
+}
+
+pub fn setBlendConstants(interface: *Interface, constants: [4]f32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = constants;
+}
+
+pub fn setDepthBias(interface: *Interface, constant_factor: f32, clamp: f32, slope_factor: f32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = constant_factor;
+    _ = clamp;
+    _ = slope_factor;
+}
+
+pub fn setDepthBounds(interface: *Interface, min: f32, max: f32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = min;
+    _ = max;
+}
+
+pub fn setLineWidth(interface: *Interface, width: f32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = width;
+}
+
+pub fn setStencilCompareMask(interface: *Interface, face_mask: vk.StencilFaceFlags, compare_mask: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = face_mask;
+    _ = compare_mask;
+}
+
+pub fn setStencilReference(interface: *Interface, face_mask: vk.StencilFaceFlags, reference: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = face_mask;
+    _ = reference;
+}
+
+pub fn setStencilWriteMask(interface: *Interface, face_mask: vk.StencilFaceFlags, write_mask: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = face_mask;
+    _ = write_mask;
+}
+
+pub fn waitEvent(interface: *Interface, event: *base.Event, src_stage: vk.PipelineStageFlags, dst_stage: vk.PipelineStageFlags, memory_barriers: []const vk.MemoryBarrier, buffer_barriers: []const vk.BufferMemoryBarrier, image_barriers: []const vk.ImageMemoryBarrier) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = event;
+    _ = src_stage;
+    _ = dst_stage;
+    _ = memory_barriers;
+    _ = buffer_barriers;
+    _ = image_barriers;
+}
+
+pub fn writeTimestamp(interface: *Interface, stage: vk.PipelineStageFlags, pool: *base.QueryPool, query: u32) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.cmd_count += 1;
+    _ = stage;
+    try pool.writeTimestamp(query, 0);
+}

src/phi/PhiCommandPool.zig

@@ -0,0 +1,46 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const PhiCommandBuffer = @import("PhiCommandBuffer.zig");
+
+const Self = @This();
+pub const Interface = base.CommandPool;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.CommandPoolCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .createCommandBuffer = createCommandBuffer,
+        .destroy = destroy,
+        .reset = reset,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn createCommandBuffer(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.CommandBufferAllocateInfo) VkError!*base.CommandBuffer {
+    const cmd = try PhiCommandBuffer.create(interface.owner, allocator, info);
+    return &cmd.interface;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn reset(interface: *Interface, flags: vk.CommandPoolResetFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+    _ = flags;
+}

src/phi/PhiDescriptorPool.zig

@@ -0,0 +1,53 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+
+const Self = @This();
+pub const Interface = base.DescriptorPool;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.DescriptorPoolCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .allocateDescriptorSet = allocateDescriptorSet,
+        .destroy = destroy,
+        .freeDescriptorSet = freeDescriptorSet,
+        .reset = reset,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn allocateDescriptorSet(interface: *Interface, layout: *base.DescriptorSetLayout) VkError!*base.DescriptorSet {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+    _ = layout;
+    return VkError.Unknown;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn freeDescriptorSet(interface: *Interface, set: *base.DescriptorSet) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+    _ = set;
+}
+
+pub fn reset(interface: *Interface, _: vk.DescriptorPoolResetFlags) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}

src/phi/PhiDescriptorSet.zig

@@ -0,0 +1,46 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.DescriptorSet;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, layout: *base.DescriptorSetLayout) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, layout);
+
+    interface.vtable = &.{
+        .copy = copy,
+        .destroy = destroy,
+        .write = write,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn copy(interface: *Interface, src_interface: *const Interface, data: vk.CopyDescriptorSet) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+    _ = src_interface;
+    _ = data;
+}
+
+pub fn write(interface: *Interface, write_data: vk.WriteDescriptorSet) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+    _ = write_data;
+}

src/phi/PhiDescriptorSetLayout.zig

@@ -0,0 +1,31 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.DescriptorSetLayout;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.DescriptorSetLayoutCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiDevice.zig

@@ -0,0 +1,287 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+const lib = @import("lib.zig");
+
+const PhiQueue = @import("PhiQueue.zig");
+const PhiPhysicalDevice = @import("PhiPhysicalDevice.zig");
+const PhiTransport = @import("PhiTransport.zig");
+const phi_daemon = @import("phi_daemon");
+
+pub const PhiBinarySemaphore = @import("PhiBinarySemaphore.zig");
+pub const PhiBuffer = @import("PhiBuffer.zig");
+pub const PhiBufferView = @import("PhiBufferView.zig");
+pub const PhiCommandBuffer = @import("PhiCommandBuffer.zig");
+pub const PhiCommandPool = @import("PhiCommandPool.zig");
+pub const PhiDescriptorPool = @import("PhiDescriptorPool.zig");
+pub const PhiDescriptorSetLayout = @import("PhiDescriptorSetLayout.zig");
+pub const PhiDeviceMemory = @import("PhiDeviceMemory.zig");
+pub const PhiEvent = @import("PhiEvent.zig");
+pub const PhiFence = @import("PhiFence.zig");
+pub const PhiFramebuffer = @import("PhiFramebuffer.zig");
+pub const PhiImage = @import("PhiImage.zig");
+pub const PhiInstance = @import("PhiInstance.zig");
+pub const PhiImageView = @import("PhiImageView.zig");
+pub const PhiPipeline = @import("PhiPipeline.zig");
+pub const PhiPipelineCache = @import("PhiPipelineCache.zig");
+pub const PhiPipelineLayout = @import("PhiPipelineLayout.zig");
+pub const PhiQueryPool = @import("PhiQueryPool.zig");
+pub const PhiRenderPass = @import("PhiRenderPass.zig");
+pub const PhiSampler = @import("PhiSampler.zig");
+pub const PhiShaderModule = @import("PhiShaderModule.zig");
+
+const config = lib.config;
+const daemon_binary = phi_daemon.data;
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Device;
+
+interface: Interface,
+transport: PhiTransport,
+
+pub fn create(instance: *base.Instance, physical_device: *base.PhysicalDevice, allocator: std.mem.Allocator, info: *const vk.DeviceCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(allocator, instance, physical_device, info);
+
+    interface.vtable = &.{
+        .createQueue = PhiQueue.create,
+        .destroyQueue = PhiQueue.destroy,
+    };
+
+    interface.dispatch_table = &.{
+        .allocateMemory = allocateMemory,
+        .createBuffer = createBuffer,
+        .createBufferView = createBufferView,
+        .createCommandPool = createCommandPool,
+        .createComputePipeline = createComputePipeline,
+        .createDescriptorPool = createDescriptorPool,
+        .createDescriptorSetLayout = createDescriptorSetLayout,
+        .createEvent = createEvent,
+        .createFence = createFence,
+        .createFramebuffer = createFramebuffer,
+        .createGraphicsPipeline = createGraphicsPipeline,
+        .createImage = createImage,
+        .createImageView = createImageView,
+        .createPipelineCache = createPipelineCache,
+        .createPipelineLayout = createPipelineLayout,
+        .createQueryPool = createQueryPool,
+        .createRenderPass = createRenderPass,
+        .createSampler = createSampler,
+        .createSemaphore = createSemaphore,
+        .createShaderModule = createShaderModule,
+        .destroy = destroy,
+        .getDeviceGroupPeerMemoryFeatures = getDeviceGroupPeerMemoryFeatures,
+        .getDeviceGroupPresentCapabilitiesKHR = getDeviceGroupPresentCapabilitiesKHR,
+        .getDeviceGroupSurfacePresentModesKHR = getDeviceGroupSurfacePresentModesKHR,
+    };
+
+    const phi_physical_device: *PhiPhysicalDevice = @alignCast(@fieldParentPtr("interface", physical_device));
+
+    const transport = PhiTransport.init(instance, phi_physical_device.scif_node_id) catch blk: {
+        // If first connect failed try to upload the daemon to the card
+        try uploadAndLaunchDaemon(instance, allocator, phi_physical_device.mic_device_num);
+        break :blk try PhiTransport.init(instance, phi_physical_device.scif_node_id);
+    };
+
+    self.* = .{
+        .interface = interface,
+        .transport = transport,
+    };
+
+    try self.interface.createQueues(allocator, info);
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.transport.deinit();
+    allocator.destroy(self);
+}
+
+pub fn allocateMemory(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.MemoryAllocateInfo) VkError!*base.DeviceMemory {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device_memory = try PhiDeviceMemory.create(self, allocator, info.allocation_size, info.memory_type_index);
+    return &device_memory.interface;
+}
+
+pub fn createBuffer(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.BufferCreateInfo) VkError!*base.Buffer {
+    const buffer = try PhiBuffer.create(interface, allocator, info);
+    return &buffer.interface;
+}
+
+pub fn createDescriptorPool(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.DescriptorPoolCreateInfo) VkError!*base.DescriptorPool {
+    const pool = try PhiDescriptorPool.create(interface, allocator, info);
+    return &pool.interface;
+}
+
+pub fn createDescriptorSetLayout(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.DescriptorSetLayoutCreateInfo) VkError!*base.DescriptorSetLayout {
+    const layout = try PhiDescriptorSetLayout.create(interface, allocator, info);
+    return &layout.interface;
+}
+
+pub fn createFence(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.FenceCreateInfo) VkError!*base.Fence {
+    const fence = try PhiFence.create(interface, allocator, info);
+    return &fence.interface;
+}
+
+pub fn createCommandPool(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.CommandPoolCreateInfo) VkError!*base.CommandPool {
+    const pool = try PhiCommandPool.create(interface, allocator, info);
+    return &pool.interface;
+}
+
+pub fn createImage(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.ImageCreateInfo) VkError!*base.Image {
+    const image = try PhiImage.create(interface, allocator, info);
+    return &image.interface;
+}
+
+pub fn createImageView(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.ImageViewCreateInfo) VkError!*base.ImageView {
+    const view = try PhiImageView.create(interface, allocator, info);
+    return &view.interface;
+}
+
+pub fn createBufferView(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.BufferViewCreateInfo) VkError!*base.BufferView {
+    const view = try PhiBufferView.create(interface, allocator, info);
+    return &view.interface;
+}
+
+pub fn createComputePipeline(interface: *Interface, allocator: std.mem.Allocator, cache: ?*base.PipelineCache, info: *const vk.ComputePipelineCreateInfo) VkError!*base.Pipeline {
+    const pipeline = try PhiPipeline.createCompute(interface, allocator, cache, info);
+    return &pipeline.interface;
+}
+
+pub fn createEvent(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.EventCreateInfo) VkError!*base.Event {
+    const event = try PhiEvent.create(interface, allocator, info);
+    return &event.interface;
+}
+
+pub fn createFramebuffer(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.FramebufferCreateInfo) VkError!*base.Framebuffer {
+    const framebuffer = try PhiFramebuffer.create(interface, allocator, info);
+    return &framebuffer.interface;
+}
+
+pub fn createGraphicsPipeline(interface: *Interface, allocator: std.mem.Allocator, cache: ?*base.PipelineCache, info: *const vk.GraphicsPipelineCreateInfo) VkError!*base.Pipeline {
+    const pipeline = try PhiPipeline.createGraphics(interface, allocator, cache, info);
+    return &pipeline.interface;
+}
+
+pub fn createPipelineCache(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.PipelineCacheCreateInfo) VkError!*base.PipelineCache {
+    const cache = try PhiPipelineCache.create(interface, allocator, info);
+    return &cache.interface;
+}
+
+pub fn createPipelineLayout(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.PipelineLayoutCreateInfo) VkError!*base.PipelineLayout {
+    const layout = try PhiPipelineLayout.create(interface, allocator, info);
+    return &layout.interface;
+}
+
+pub fn createQueryPool(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.QueryPoolCreateInfo) VkError!*base.QueryPool {
+    const pool = try PhiQueryPool.create(interface, allocator, info);
+    return &pool.interface;
+}
+
+pub fn createRenderPass(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.RenderPassCreateInfo) VkError!*base.RenderPass {
+    const pass = try PhiRenderPass.create(interface, allocator, info);
+    return &pass.interface;
+}
+
+pub fn createSampler(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.SamplerCreateInfo) VkError!*base.Sampler {
+    const sampler = try PhiSampler.create(interface, allocator, info);
+    return &sampler.interface;
+}
+
+pub fn createSemaphore(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.SemaphoreCreateInfo) VkError!*base.BinarySemaphore {
+    const semaphore = try PhiBinarySemaphore.create(interface, allocator, info);
+    return &semaphore.interface;
+}
+
+pub fn createShaderModule(interface: *Interface, allocator: std.mem.Allocator, info: *const vk.ShaderModuleCreateInfo) VkError!*base.ShaderModule {
+    const module = try PhiShaderModule.create(interface, allocator, info);
+    return &module.interface;
+}
+
+pub fn getDeviceGroupPeerMemoryFeatures(interface: *Interface, heap_index: u32, local_device_index: u32, remote_device_index: u32) VkError!vk.PeerMemoryFeatureFlags {
+    if (heap_index >= interface.physical_device.mem_props.memory_heap_count) return VkError.ValidationFailed;
+    if (local_device_index != 0 or remote_device_index != 0) return VkError.ValidationFailed;
+
+    return .{
+        .copy_src_bit = true,
+        .copy_dst_bit = true,
+        .generic_src_bit = true,
+        .generic_dst_bit = true,
+    };
+}
+
+pub fn getDeviceGroupPresentCapabilitiesKHR(_: *Interface, capabilities: *vk.DeviceGroupPresentCapabilitiesKHR) VkError!void {
+    capabilities.present_mask = @splat(0);
+    capabilities.present_mask[0] = 1;
+    capabilities.modes = .{ .local_bit_khr = true };
+}
+
+pub fn getDeviceGroupSurfacePresentModesKHR(_: *Interface, _: *base.SurfaceKHR) VkError!vk.DeviceGroupPresentModeFlagsKHR {
+    return .{ .local_bit_khr = true };
+}
+
+fn uploadAndLaunchDaemon(instance: *base.Instance, allocator: std.mem.Allocator, mic_device_num: u32) VkError!void {
+    const io = instance.io();
+
+    const local_path = std.fmt.allocPrint(allocator, "/tmp/ape_phi_device_{d}_{d}.mic", .{ std.os.linux.getpid(), mic_device_num }) catch return VkError.OutOfHostMemory;
+    defer allocator.free(local_path);
+    defer std.Io.Dir.deleteFileAbsolute(io, local_path) catch @panic("Caught an error while handling an error");
+
+    std.Io.Dir.writeFile(.cwd(), io, .{
+        .sub_path = local_path,
+        .data = daemon_binary,
+    }) catch |err| {
+        std.log.scoped(.PhiDevice).err("Failed to write embedded Phi daemon: {s}", .{@errorName(err)});
+        return VkError.InitializationFailed;
+    };
+
+    const host = std.fmt.allocPrint(allocator, "{s}{d}", .{ config.phi_daemon_host_prefix, mic_device_num }) catch return VkError.OutOfHostMemory;
+    defer allocator.free(host);
+
+    const remote_target = std.fmt.allocPrint(allocator, "{s}:{s}", .{ host, config.phi_daemon_remote_path }) catch return VkError.OutOfHostMemory;
+    defer allocator.free(remote_target);
+
+    try runHostCommand(instance, allocator, &.{
+        "scp",
+        local_path,
+        remote_target,
+    });
+
+    const launch_command = std.fmt.allocPrint(
+        allocator,
+        "chmod +x {s} && nohup {s} >/tmp/phi_device.log 2>&1 </dev/null &",
+        .{ config.phi_daemon_remote_path, config.phi_daemon_remote_path },
+    ) catch return VkError.OutOfHostMemory;
+    defer allocator.free(launch_command);
+
+    try runHostCommand(instance, allocator, &.{
+        "ssh",
+        host,
+        launch_command,
+    });
+}
+
+fn runHostCommand(instance: *base.Instance, allocator: std.mem.Allocator, argv: []const []const u8) VkError!void {
+    const result = std.process.run(allocator, instance.io(), .{
+        .argv = argv,
+        .stdout_limit = .limited(4096),
+        .stderr_limit = .limited(4096),
+    }) catch |err| {
+        std.log.scoped(.PhiDevice).err("Failed to run {s}: {s}", .{ argv[0], @errorName(err) });
+        return VkError.InitializationFailed;
+    };
+    defer allocator.free(result.stdout);
+    defer allocator.free(result.stderr);
+
+    switch (result.term) {
+        .exited => |code| if (code == 0) return,
+        else => {},
+    }
+
+    std.log.scoped(.PhiDevice).err("{s} failed: stdout=\"{s}\" stderr=\"{s}\"", .{ argv[0], result.stdout, result.stderr });
+    return VkError.InitializationFailed;
+}

src/phi/PhiDeviceMemory.zig

@@ -0,0 +1,127 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+const lib = @import("lib.zig");
+const proto = lib.proto;
+
+const PhiDevice = @import("PhiDevice.zig");
+const PhiTransport = @import("PhiTransport.zig");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.DeviceMemory;
+
+interface: Interface,
+remote_handle: u64,
+data: ?[]u8,
+
+pub fn create(device: *PhiDevice, allocator: std.mem.Allocator, size: vk.DeviceSize, memory_type_index: u32) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(&device.interface, size, memory_type_index);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+        .map = map,
+        .unmap = unmap,
+        .flushRange = flushRange,
+        .invalidateRange = invalidateRange,
+    };
+
+    if (memory_type_index >= device.interface.physical_device.mem_props.memory_type_count) {
+        return VkError.ValidationFailed;
+    }
+
+    const memory_type = device.interface.physical_device.mem_props.memory_types[memory_type_index];
+    const host_visible = memory_type.property_flags.host_visible_bit;
+    const device_local = memory_type.property_flags.device_local_bit;
+    const allocation_size = std.math.cast(usize, size) orelse return VkError.OutOfDeviceMemory;
+
+    const remote_handle = if (device_local) blk: {
+        const alloc_request: proto.PhiAllocMemoryRequest = .{
+            .size = size,
+            .memory_type_index = memory_type_index,
+            .flags = 0,
+        };
+
+        var reply = std.mem.zeroes(proto.PhiAllocMemoryReply);
+        try device.transport.request(proto.PHI_PACKET_ALLOC_MEMORY, std.mem.asBytes(&alloc_request), std.mem.asBytes(&reply));
+
+        if (reply.result.status != proto.PHI_STATUS_OK) {
+            return PhiTransport.statusToErr(reply.result.status);
+        }
+
+        std.log.scoped(.PhiDeviceMemory).info("Recieved remote handle 0x{X}", .{reply.remote_handle});
+
+        break :blk reply.remote_handle;
+    } else 0;
+    errdefer if (remote_handle != 0) self.interface.destroy(allocator);
+
+    const data = if (host_visible)
+        device.interface.device_allocator.allocator().alloc(u8, allocation_size) catch return VkError.OutOfDeviceMemory
+    else
+        null;
+
+    self.* = .{
+        .interface = interface,
+        .remote_handle = remote_handle,
+        .data = data,
+    };
+
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const device: *PhiDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+    if (self.data) |data| {
+        interface.owner.device_allocator.allocator().free(data);
+    }
+    if (self.remote_handle != 0) {
+        const request_payload: proto.PhiFreeMemoryRequest = .{
+            .remote_handle = self.remote_handle,
+        };
+        var reply: proto.PhiFreeMemoryReply = undefined;
+        device.transport.request(proto.PHI_PACKET_FREE_MEMORY, std.mem.asBytes(&request_payload), std.mem.asBytes(&reply)) catch |err| {
+            std.log.scoped(.PhiTransport).err("Remote free failed: {s}", .{@errorName(err)});
+            return;
+        };
+        if (reply.result.status != proto.PHI_STATUS_OK) {
+            std.log.scoped(.PhiTransport).err("Remote free returned status {d}", .{reply.result.status});
+        }
+    }
+    allocator.destroy(self);
+}
+
+pub fn flushRange(interface: *Interface, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
+    _ = interface;
+    _ = offset;
+    _ = size;
+}
+
+pub fn invalidateRange(interface: *Interface, offset: vk.DeviceSize, size: vk.DeviceSize) VkError!void {
+    _ = interface;
+    _ = offset;
+    _ = size;
+}
+
+pub fn map(interface: *Interface, offset: vk.DeviceSize, size: vk.DeviceSize) VkError![]u8 {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    const data = self.data orelse return VkError.MemoryMapFailed;
+    const map_offset = std.math.cast(usize, offset) orelse return VkError.MemoryMapFailed;
+    if (map_offset >= data.len) {
+        return VkError.MemoryMapFailed;
+    }
+    const map_size = if (size == vk.WHOLE_SIZE)
+        data.len - map_offset
+    else
+        std.math.cast(usize, size) orelse return VkError.MemoryMapFailed;
+    if (map_size > data.len - map_offset) {
+        return VkError.MemoryMapFailed;
+    }
+    return data[map_offset..(map_offset + map_size)];
+}
+
+pub fn unmap(_: *Interface) void {}

src/phi/PhiEvent.zig

@@ -0,0 +1,55 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Event;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.EventCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+        .getStatus = getStatus,
+        .reset = reset,
+        .signal = signal,
+        .wait = wait,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn getStatus(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn reset(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn signal(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn wait(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}

src/phi/PhiFence.zig

@@ -0,0 +1,57 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+const Device = base.Device;
+
+const Self = @This();
+pub const Interface = base.Fence;
+
+interface: Interface,
+
+pub fn create(device: *Device, allocator: std.mem.Allocator, info: *const vk.FenceCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+        .getStatus = getStatus,
+        .reset = reset,
+        .signal = signal,
+        .wait = wait,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn getStatus(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn reset(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn signal(interface: *Interface) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+}
+
+pub fn wait(interface: *Interface, timeout: u64) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    _ = self;
+    _ = timeout;
+}

src/phi/PhiFramebuffer.zig

@@ -0,0 +1,26 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Framebuffer;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.FramebufferCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+    interface.vtable = &.{ .destroy = destroy };
+
+    self.* = .{ .interface = interface };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiImage.zig

@@ -0,0 +1,77 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Image;
+
+pub const F32x4 = @Vector(4, f32);
+pub const U32x4 = @Vector(4, u32);
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.ImageCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+    interface.vtable = &.{
+        .destroy = destroy,
+        .getMemoryRequirements = getMemoryRequirements,
+        .getSubresourceLayout = getSubresourceLayout,
+        .getTotalSizeForAspect = getTotalSizeForAspect,
+        .getSliceMemSizeForMipLevel = getSliceMemSizeForMipLevel,
+        .getRowPitchMemSizeForMipLevel = getRowPitchMemSizeForMipLevel,
+        .copyToMemory = copyToMemory,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn getMemoryRequirements(_: *Interface, requirements: *vk.MemoryRequirements) VkError!void {
+    _ = requirements;
+}
+
+pub fn copyToMemory(interface: *const Interface, memory: []u8, subresource: vk.ImageSubresourceLayers) VkError!void {
+    _ = interface;
+    _ = subresource;
+    @memset(memory, 0);
+}
+
+pub fn getTotalSizeForAspect(interface: *const Interface, aspect_mask: vk.ImageAspectFlags) VkError!usize {
+    _ = aspect_mask;
+    return interface.extent.width * interface.extent.height * interface.extent.depth * base.format.texelSize(interface.format);
+}
+
+pub fn getSubresourceLayout(interface: *const Interface, subresource: vk.ImageSubresource) VkError!vk.SubresourceLayout {
+    _ = subresource;
+    return .{
+        .offset = 0,
+        .size = try getTotalSizeForAspect(interface, base.format.toAspect(interface.format)),
+        .row_pitch = getRowPitchMemSizeForMipLevel(interface, base.format.toAspect(interface.format), 0),
+        .array_pitch = getSliceMemSizeForMipLevel(interface, base.format.toAspect(interface.format), 0),
+        .depth_pitch = getSliceMemSizeForMipLevel(interface, base.format.toAspect(interface.format), 0),
+    };
+}
+
+pub fn getSliceMemSizeForMipLevel(interface: *const Interface, aspect_mask: vk.ImageAspectFlags, mip_level: u32) usize {
+    _ = aspect_mask;
+    _ = mip_level;
+    return interface.extent.width * interface.extent.height * base.format.texelSize(interface.format);
+}
+
+pub fn getRowPitchMemSizeForMipLevel(interface: *const Interface, aspect_mask: vk.ImageAspectFlags, mip_level: u32) usize {
+    _ = aspect_mask;
+    _ = mip_level;
+    return interface.extent.width * base.format.texelSize(interface.format);
+}

src/phi/PhiImageView.zig

@@ -0,0 +1,31 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.ImageView;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.ImageViewCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiInstance.zig

@@ -0,0 +1,140 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+const lib = @import("lib.zig");
+const mic = lib.mic;
+
+const PhiPhysicalDevice = @import("PhiPhysicalDevice.zig");
+
+const Dispatchable = base.Dispatchable;
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Instance;
+
+interface: Interface,
+threaded: std.Io.Threaded,
+io_impl: std.Io,
+allocator: std.mem.Allocator,
+
+fn castExtension(comptime ext: vk.ApiInfo) vk.ExtensionProperties {
+    var props: vk.ExtensionProperties = .{
+        .extension_name = @splat(0),
+        .spec_version = @bitCast(ext.version),
+    };
+    @memcpy(props.extension_name[0..ext.name.len], ext.name);
+    return props;
+}
+
+pub const extensions = [_]vk.ExtensionProperties{
+    castExtension(vk.extensions.khr_device_group_creation),
+    castExtension(vk.extensions.khr_get_physical_device_properties_2),
+    castExtension(vk.extensions.khr_surface),
+    castExtension(vk.extensions.khr_wayland_surface),
+};
+
+pub fn create(allocator: std.mem.Allocator, infos: *const vk.InstanceCreateInfo) VkError!*Interface {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    self.allocator = std.heap.smp_allocator;
+    self.threaded = std.Io.Threaded.init(self.allocator, .{});
+    self.io_impl = self.threaded.io();
+
+    self.interface = try base.Instance.init(allocator, infos);
+    self.interface.dispatch_table = &.{
+        .destroy = destroy,
+    };
+
+    self.interface.vtable = &.{
+        .requestPhysicalDevices = requestPhysicalDevices,
+        .releasePhysicalDevices = releasePhysicalDevices,
+        .io = io,
+    };
+
+    return &self.interface;
+}
+
+fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.threaded.deinit();
+    allocator.destroy(self);
+
+    mic.unload();
+}
+
+fn requestPhysicalDevices(interface: *Interface, allocator: std.mem.Allocator, _: []base.drm.Card) VkError!void {
+    if (interface.physical_devices.items.len != 0) {
+        return;
+    }
+
+    mic.load() catch |err| {
+        std.log.scoped(.MIC).err("Failed to load libmicmgmt: {s}", .{@errorName(err)});
+        return VkError.InitializationFailed;
+    };
+
+    var devices = mic.DeviceList.init() catch |err| {
+        std.log.scoped(.MIC).err("Failed to create device list: {s}", .{@errorName(err)});
+        return VkError.InitializationFailed;
+    };
+    defer devices.deinit();
+
+    const count = devices.count() catch |err| {
+        std.log.scoped(.MIC).err("Failed to fetch device list count: {s}", .{@errorName(err)});
+        return VkError.InitializationFailed;
+    };
+
+    for (0..count) |index| {
+        const device_num = devices.deviceAtIndex(index) catch |err| {
+            std.log.scoped(.MIC).err("Failed to fetch device: {s}", .{@errorName(err)});
+            continue;
+        };
+
+        var device = mic.Device.open(device_num) catch |err| {
+            std.log.scoped(.MIC).err("Failed to open device {d}: {s}", .{ device_num, @errorName(err) });
+            continue;
+        };
+        defer device.deinit();
+
+        const physical_device = try PhiPhysicalDevice.create(allocator, interface, device, device_num);
+        errdefer physical_device.interface.release(allocator) catch @panic("Caught an error while handling an error");
+
+        const dispatchable = try Dispatchable(base.PhysicalDevice).wrap(allocator, &physical_device.interface);
+        errdefer dispatchable.destroy(allocator);
+
+        interface.physical_devices.append(allocator, dispatchable) catch return VkError.OutOfHostMemory;
+    }
+}
+
+fn releasePhysicalDevices(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
+    var result: ?VkError = null;
+
+    for (interface.physical_devices.items) |physical_device| {
+        physical_device.object.release(allocator) catch |err| {
+            if (result == null) {
+                result = err;
+            }
+        };
+        physical_device.destroy(allocator);
+    }
+
+    interface.physical_devices.deinit(allocator);
+    interface.physical_devices = .empty;
+
+    if (result) |err| {
+        return err;
+    }
+}
+
+fn io(interface: *Interface) std.Io {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    return self.io_impl;
+}
+
+fn mapDeviceEnumerationError(err: anyerror) VkError {
+    return switch (err) {
+        error.OutOfMemory => VkError.OutOfHostMemory,
+        else => VkError.InitializationFailed,
+    };
+}

src/phi/PhiPhysicalDevice.zig

@@ -0,0 +1,901 @@
+const std = @import("std");
+const builtin = @import("builtin");
+const vk = @import("vulkan");
+const base = @import("base");
+const lib = @import("lib.zig");
+const mic = lib.mic;
+
+const pci_ids = @import("pci_ids.zig").map;
+
+const PhiDevice = @import("PhiDevice.zig");
+
+const VkError = base.VkError;
+const SurfaceKHR = base.SurfaceKHR;
+
+const Self = @This();
+pub const Interface = base.PhysicalDevice;
+
+fn castExtension(comptime ext: vk.ApiInfo) vk.ExtensionProperties {
+    var props: vk.ExtensionProperties = .{
+        .extension_name = @splat(0),
+        .spec_version = @bitCast(ext.version),
+    };
+    @memcpy(props.extension_name[0..ext.name.len], ext.name);
+    return props;
+}
+
+pub const extensions = [_]vk.ExtensionProperties{
+    castExtension(vk.extensions.khr_device_group),
+    castExtension(vk.extensions.khr_swapchain),
+};
+
+interface: Interface,
+scif_node_id: u16,
+mic_device_num: u32,
+
+pub fn create(allocator: std.mem.Allocator, instance: *base.Instance, mic_device: mic.Device, mic_device_num: u32) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(allocator, instance);
+
+    interface.dispatch_table = &.{
+        .createDevice = createDevice,
+        .getFormatProperties = getFormatProperties,
+        .getImageFormatProperties = getImageFormatProperties,
+        .getSparseImageFormatProperties = getSparseImageFormatProperties,
+        .enumerateLayerProperties = enumerateLayerProperties,
+        .enumerateExtensionProperties = enumerateExtensionProperties,
+        .release = destroy,
+
+        // VK_KHR_get_physical_device_properties_2
+        .getSparseImageFormatProperties2 = getSparseImageFormatProperties2,
+
+        // VK_KHR_surface
+        .getSurfaceSupportKHR = getSurfaceSupportKHR,
+    };
+
+    interface.props.api_version = @bitCast(lib.vulkan_version);
+    interface.props.driver_version = @bitCast(base.driver_version);
+    interface.props.device_type = .other;
+
+    @memset(interface.props.device_name[0..], 0);
+
+    if (mic_device.pciConfig()) |pci_value| {
+        var pci = pci_value;
+        defer pci.deinit();
+
+        interface.props.vendor_id = pci.vendorId() catch 0;
+        interface.props.device_id = pci.deviceId() catch 0;
+
+        for (pci_ids[0..]) |pci_info| {
+            if (pci_info.id != pci.deviceId() catch 0)
+                continue;
+
+            const len = @min(vk.MAX_PHYSICAL_DEVICE_NAME_SIZE, pci_info.name.len);
+            @memcpy(interface.props.device_name[0..len], pci_info.name[0..len]);
+
+            const driver_mark = " [Phi ApeDriver]";
+
+            @memcpy(interface.props.device_name[len .. len + driver_mark.len], driver_mark);
+
+            break;
+        }
+    } else |err| {
+        std.log.scoped(.MIC).err("Failed to fetch device PCI config: {s}", .{@errorName(err)});
+        return VkError.InitializationFailed;
+    }
+
+    interface.props.pipeline_cache_uuid = @splat(0);
+    interface.props.limits = .{
+        .max_image_dimension_1d = 4096,
+        .max_image_dimension_2d = 4096,
+        .max_image_dimension_3d = 256,
+        .max_image_dimension_cube = 4096,
+        .max_image_array_layers = 256,
+        .max_texel_buffer_elements = 65536,
+        .max_uniform_buffer_range = 16384,
+        .max_storage_buffer_range = 134217728,
+        .max_push_constants_size = 256,
+        .max_memory_allocation_count = 1024,
+        .max_sampler_allocation_count = 4096,
+        .buffer_image_granularity = 131072,
+        .sparse_address_space_size = 0,
+        .max_bound_descriptor_sets = 8,
+        .max_per_stage_descriptor_samplers = 16,
+        .max_per_stage_descriptor_uniform_buffers = 12,
+        .max_per_stage_descriptor_storage_buffers = 4,
+        .max_per_stage_descriptor_sampled_images = 16,
+        .max_per_stage_descriptor_storage_images = 4,
+        .max_per_stage_descriptor_input_attachments = 4,
+        .max_per_stage_resources = 128,
+        .max_descriptor_set_samplers = 96,
+        .max_descriptor_set_uniform_buffers = 72,
+        .max_descriptor_set_uniform_buffers_dynamic = 8,
+        .max_descriptor_set_storage_buffers = 24,
+        .max_descriptor_set_storage_buffers_dynamic = 4,
+        .max_descriptor_set_sampled_images = 96,
+        .max_descriptor_set_storage_images = 24,
+        .max_descriptor_set_input_attachments = 4,
+        .max_vertex_input_attributes = 32,
+        .max_vertex_input_bindings = 32,
+        .max_vertex_input_attribute_offset = 2047,
+        .max_vertex_input_binding_stride = 2048,
+        .max_vertex_output_components = 64,
+        .max_tessellation_generation_level = 0,
+        .max_tessellation_patch_size = 0,
+        .max_tessellation_control_per_vertex_input_components = 0,
+        .max_tessellation_control_per_vertex_output_components = 0,
+        .max_tessellation_control_per_patch_output_components = 0,
+        .max_tessellation_control_total_output_components = 0,
+        .max_tessellation_evaluation_input_components = 0,
+        .max_tessellation_evaluation_output_components = 0,
+        .max_geometry_shader_invocations = 0,
+        .max_geometry_input_components = 0,
+        .max_geometry_output_components = 0,
+        .max_geometry_output_vertices = 0,
+        .max_geometry_total_output_components = 0,
+        .max_fragment_input_components = 64,
+        .max_fragment_output_attachments = 4,
+        .max_fragment_dual_src_attachments = 0,
+        .max_fragment_combined_output_resources = 4,
+        .max_compute_shared_memory_size = 16384,
+        .max_compute_work_group_count = .{ 65535, 65535, 65535 },
+        .max_compute_work_group_invocations = 128,
+        .max_compute_work_group_size = .{ 128, 128, 64 },
+        .sub_pixel_precision_bits = 4,
+        .sub_texel_precision_bits = 4,
+        .mipmap_precision_bits = 4,
+        .max_draw_indexed_index_value = 4294967295,
+        .max_draw_indirect_count = 65535,
+        .max_sampler_lod_bias = 2.0,
+        .max_sampler_anisotropy = 1.0,
+        .max_viewports = 1,
+        .max_viewport_dimensions = .{ 4096, 4096 },
+        .viewport_bounds_range = .{ -8192.0, 8191.0 },
+        .viewport_sub_pixel_bits = 0,
+        .min_memory_map_alignment = 64,
+        .min_texel_buffer_offset_alignment = 256,
+        .min_uniform_buffer_offset_alignment = 256,
+        .min_storage_buffer_offset_alignment = 256,
+        .min_texel_offset = -8,
+        .max_texel_offset = 7,
+        .min_texel_gather_offset = 0,
+        .max_texel_gather_offset = 0,
+        .min_interpolation_offset = 0.0,
+        .max_interpolation_offset = 0.0,
+        .sub_pixel_interpolation_offset_bits = 0,
+        .max_framebuffer_width = 4096,
+        .max_framebuffer_height = 4096,
+        .max_framebuffer_layers = 256,
+        .framebuffer_color_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .framebuffer_depth_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .framebuffer_stencil_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .framebuffer_no_attachments_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .max_color_attachments = 4,
+        .sampled_image_color_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .sampled_image_integer_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .sampled_image_depth_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .sampled_image_stencil_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .storage_image_sample_counts = .{ .@"1_bit" = true, .@"4_bit" = true },
+        .max_sample_mask_words = 1,
+        .timestamp_compute_and_graphics = .false,
+        .timestamp_period = 1.0,
+        .max_clip_distances = 0,
+        .max_cull_distances = 0,
+        .max_combined_clip_and_cull_distances = 0,
+        .discrete_queue_priorities = 2,
+        .point_size_range = .{ 1.0, 1.0 },
+        .line_width_range = .{ 1.0, 1.0 },
+        .point_size_granularity = 0.0,
+        .line_width_granularity = 0.0,
+        .strict_lines = .false,
+        .standard_sample_locations = .true,
+        .optimal_buffer_copy_offset_alignment = 1,
+        .optimal_buffer_copy_row_pitch_alignment = 1,
+        .non_coherent_atom_size = 256,
+    };
+
+    {
+        interface.mem_props.memory_type_count = 3;
+
+        interface.mem_props.memory_types[0] = .{
+            .heap_index = 0,
+            .property_flags = .{
+                .device_local_bit = true,
+            },
+        };
+
+        interface.mem_props.memory_types[1] = .{
+            .heap_index = 1,
+            .property_flags = .{
+                .host_visible_bit = true,
+                .host_coherent_bit = true,
+            },
+        };
+
+        interface.mem_props.memory_types[2] = .{
+            .heap_index = 1,
+            .property_flags = .{
+                .host_visible_bit = true,
+                .host_coherent_bit = true,
+                .host_cached_bit = true,
+            },
+        };
+    }
+
+    if (mic_device.memoryInfo()) |memory_value| {
+        var memory = memory_value;
+        defer memory.deinit();
+
+        interface.mem_props.memory_heap_count = 2;
+
+        interface.mem_props.memory_heaps[0] = .{
+            .size = memory.size() catch 0,
+            .flags = .{ .device_local_bit = true },
+        };
+        interface.mem_props.memory_heaps[1] = .{
+            .size = std.process.totalSystemMemory() catch 0,
+            .flags = .{},
+        };
+    } else |err| {
+        std.log.scoped(.MIC).err("Failed to fetch device memory infos: {s}", .{@errorName(err)});
+        return VkError.InitializationFailed;
+    }
+
+    interface.features = .{
+        .shader_float_64 = .true,
+        .shader_int_64 = .true,
+        .shader_int_16 = .true,
+    };
+
+    var queue_family_props = [_]vk.QueueFamilyProperties{
+        .{
+            .queue_flags = .{ .graphics_bit = true, .compute_bit = true, .transfer_bit = true },
+            .queue_count = 1,
+            .timestamp_valid_bits = 0,
+            .min_image_transfer_granularity = .{ .width = 1, .height = 1, .depth = 1 },
+        },
+    };
+    interface.queue_family_props.appendSlice(allocator, queue_family_props[0..]) catch return VkError.OutOfHostMemory;
+
+    self.* = .{
+        .interface = interface,
+        .scif_node_id = deviceNumToScifNode(mic_device_num),
+        .mic_device_num = mic_device_num,
+    };
+
+    return self;
+}
+
+fn deviceNumToScifNode(device_num: u32) u16 {
+    return @intCast(device_num + 1);
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn createDevice(interface: *Interface, allocator: std.mem.Allocator, infos: *const vk.DeviceCreateInfo) VkError!*base.Device {
+    const device = try PhiDevice.create(interface.instance, interface, allocator, infos);
+    return &device.interface;
+}
+
+pub fn enumerateLayerProperties(_: *const Interface, count: *u32, p_properties: ?[*]vk.LayerProperties) VkError!void {
+    count.* = 0;
+    _ = p_properties;
+}
+
+pub fn enumerateExtensionProperties(_: *const Interface, layer_name: ?[]const u8, count: *u32, p_properties: ?[*]vk.ExtensionProperties) VkError!void {
+    if (layer_name) |_| {
+        return VkError.LayerNotPresent;
+    }
+
+    const available = extensions.len;
+    if (p_properties) |properties| {
+        const write_count = @min(count.*, available);
+        for (extensions[0..write_count], properties[0..write_count]) |ext, *prop| {
+            prop.* = ext;
+        }
+        count.* = @intCast(write_count);
+        if (write_count < available) return VkError.Incomplete;
+    } else {
+        count.* = @intCast(available);
+    }
+}
+
+pub fn getFormatProperties(interface: *Interface, format: vk.Format) VkError!vk.FormatProperties {
+    _ = interface;
+    var properties: vk.FormatProperties = .{};
+
+    switch (format) {
+        // Formats which can be sampled *and* filtered
+        .r4g4b4a4_unorm_pack16,
+        .b4g4r4a4_unorm_pack16,
+        .a4r4g4b4_unorm_pack16,
+        .a4b4g4r4_unorm_pack16,
+        .r5g6b5_unorm_pack16,
+        .b5g6r5_unorm_pack16,
+        .r5g5b5a1_unorm_pack16,
+        .b5g5r5a1_unorm_pack16,
+        .a1r5g5b5_unorm_pack16,
+        .r8_unorm,
+        .r8_srgb,
+        .r8_snorm,
+        .r8g8_unorm,
+        .r8g8_srgb,
+        .r8g8_snorm,
+        .r8g8b8a8_unorm,
+        .r8g8b8a8_snorm,
+        .r8g8b8a8_srgb,
+        .b8g8r8a8_unorm,
+        .b8g8r8a8_srgb,
+        .a8b8g8r8_unorm_pack32,
+        .a8b8g8r8_snorm_pack32,
+        .a8b8g8r8_srgb_pack32,
+        .a2b10g10r10_unorm_pack32,
+        .a2r10g10b10_unorm_pack32,
+        .r16_unorm,
+        .r16_snorm,
+        .r16_sfloat,
+        .r16g16_unorm,
+        .r16g16_snorm,
+        .r16g16_sfloat,
+        .r16g16b16a16_unorm,
+        .r16g16b16a16_snorm,
+        .r16g16b16a16_sfloat,
+        .r32_sfloat,
+        .r32g32_sfloat,
+        .r32g32b32a32_sfloat,
+        .b10g11r11_ufloat_pack32,
+        .e5b9g9r9_ufloat_pack32,
+        //.bc1_rgb_unorm_block,
+        //.bc1_rgb_srgb_block,
+        //.bc1_rgba_unorm_block,
+        //.bc1_rgba_srgb_block,
+        //.bc2_unorm_block,
+        //.bc2_srgb_block,
+        //.bc3_unorm_block,
+        //.bc3_srgb_block,
+        //.bc4_unorm_block,
+        //.bc4_snorm_block,
+        //.bc5_unorm_block,
+        //.bc5_snorm_block,
+        //.bc6h_ufloat_block,
+        //.bc6h_sfloat_block,
+        //.bc7_unorm_block,
+        //.bc7_srgb_block,
+        //.etc2_r8g8b8_unorm_block,
+        //.etc2_r8g8b8_srgb_block,
+        //.etc2_r8g8b8a1_unorm_block,
+        //.etc2_r8g8b8a1_srgb_block,
+        //.etc2_r8g8b8a8_unorm_block,
+        //.etc2_r8g8b8a8_srgb_block,
+        //.eac_r11_unorm_block,
+        //.eac_r11_snorm_block,
+        //.eac_r11g11_unorm_block,
+        //.eac_r11g11_snorm_block,
+        //.astc_4x_4_unorm_block,
+        //.astc_5x_4_unorm_block,
+        //.astc_5x_5_unorm_block,
+        //.astc_6x_5_unorm_block,
+        //.astc_6x_6_unorm_block,
+        //.astc_8x_5_unorm_block,
+        //.astc_8x_6_unorm_block,
+        //.astc_8x_8_unorm_block,
+        //.astc_1_0x_5_unorm_block,
+        //.astc_1_0x_6_unorm_block,
+        //.astc_1_0x_8_unorm_block,
+        //.astc_1_0x_10_unorm_block,
+        //.astc_1_2x_10_unorm_block,
+        //.astc_1_2x_12_unorm_block,
+        //.astc_4x_4_srgb_block,
+        //.astc_5x_4_srgb_block,
+        //.astc_5x_5_srgb_block,
+        //.astc_6x_5_srgb_block,
+        //.astc_6x_6_srgb_block,
+        //.astc_8x_5_srgb_block,
+        //.astc_8x_6_srgb_block,
+        //.astc_8x_8_srgb_block,
+        //.astc_1_0x_5_srgb_block,
+        //.astc_1_0x_6_srgb_block,
+        //.astc_1_0x_8_srgb_block,
+        //.astc_1_0x_10_srgb_block,
+        //.astc_1_2x_10_srgb_block,
+        //.astc_1_2x_12_srgb_block,
+        .d16_unorm,
+        .d32_sfloat,
+        .d32_sfloat_s8_uint,
+        => {
+            properties.optimal_tiling_features.blit_src_bit = true;
+            properties.optimal_tiling_features.sampled_image_bit = true;
+            properties.optimal_tiling_features.transfer_dst_bit = true;
+            properties.optimal_tiling_features.transfer_src_bit = true;
+            properties.optimal_tiling_features.sampled_image_filter_linear_bit = true;
+        },
+
+        // Formats which can be sampled, but don't support filtering
+        .r8_uint,
+        .r8_sint,
+        .r8g8_uint,
+        .r8g8_sint,
+        .r8g8b8a8_uint,
+        .r8g8b8a8_sint,
+        .a8b8g8r8_uint_pack32,
+        .a8b8g8r8_sint_pack32,
+        .a2b10g10r10_uint_pack32,
+        .a2r10g10b10_uint_pack32,
+        .r16_uint,
+        .r16_sint,
+        .r16g16_uint,
+        .r16g16_sint,
+        .r16g16b16a16_uint,
+        .r16g16b16a16_sint,
+        .r32_uint,
+        .r32_sint,
+        .r32g32_uint,
+        .r32g32_sint,
+        .r32g32b32a32_uint,
+        .r32g32b32a32_sint,
+        .s8_uint,
+        => {
+            properties.optimal_tiling_features.blit_src_bit = true;
+            properties.optimal_tiling_features.sampled_image_bit = true;
+            properties.optimal_tiling_features.transfer_dst_bit = true;
+            properties.optimal_tiling_features.transfer_src_bit = true;
+        },
+
+        // YCbCr formats
+        .g8_b8_r8_3plane_420_unorm,
+        .g8_b8r8_2plane_420_unorm,
+        .g10x6_b10x6r10x6_2plane_420_unorm_3pack16,
+        => {
+            properties.optimal_tiling_features.sampled_image_bit = true;
+            properties.optimal_tiling_features.sampled_image_filter_linear_bit = true;
+            properties.optimal_tiling_features.sampled_image_ycbcr_conversion_linear_filter_bit = true;
+            properties.optimal_tiling_features.transfer_src_bit = true;
+            properties.optimal_tiling_features.transfer_dst_bit = true;
+            properties.optimal_tiling_features.cosited_chroma_samples_bit = true;
+        },
+        else => {},
+    }
+
+    switch (format) {
+        // Vulkan 1.0 mandatory storage image formats supporting atomic operations
+        .r32_uint,
+        .r32_sint,
+        => {
+            properties.buffer_features.storage_texel_buffer_bit = true;
+            properties.buffer_features.storage_texel_buffer_atomic_bit = true;
+            properties.optimal_tiling_features.storage_image_bit = true;
+            properties.optimal_tiling_features.storage_image_atomic_bit = true;
+        },
+        // vulkan 1.0 mandatory storage image formats
+        .r8g8b8a8_unorm,
+        .r8g8b8a8_snorm,
+        .r8g8b8a8_uint,
+        .r8g8b8a8_sint,
+        .r16g16b16a16_uint,
+        .r16g16b16a16_sint,
+        .r16g16b16a16_sfloat,
+        .r32_sfloat,
+        .r32g32_uint,
+        .r32g32_sint,
+        .r32g32_sfloat,
+        .r32g32b32a32_uint,
+        .r32g32b32a32_sint,
+        .r32g32b32a32_sfloat,
+        .a2b10g10r10_unorm_pack32,
+        .a2b10g10r10_uint_pack32,
+        // vulkan 1.0 shaderstorageimageextendedformats
+        .r16g16_sfloat,
+        .b10g11r11_ufloat_pack32,
+        .r16_sfloat,
+        .r16g16b16a16_unorm,
+        .r16g16_unorm,
+        .r8g8_unorm,
+        .r16_unorm,
+        .r8_unorm,
+        .r16g16b16a16_snorm,
+        .r16g16_snorm,
+        .r8g8_snorm,
+        .r16_snorm,
+        .r8_snorm,
+        .r16g16_sint,
+        .r8g8_sint,
+        .r16_sint,
+        .r8_sint,
+        .r16g16_uint,
+        .r8g8_uint,
+        .r16_uint,
+        .r8_uint,
+        // additional formats not listed under "formats without shader storage format"
+        .a8b8g8r8_unorm_pack32,
+        .a8b8g8r8_snorm_pack32,
+        .a8b8g8r8_uint_pack32,
+        .a8b8g8r8_sint_pack32,
+        .b8g8r8a8_unorm,
+        .b8g8r8a8_srgb,
+        => {
+            properties.optimal_tiling_features.storage_image_bit = true;
+            properties.buffer_features.storage_texel_buffer_bit = true;
+        },
+
+        else => {},
+    }
+
+    switch (format) {
+        .r5g6b5_unorm_pack16,
+        .a1r5g5b5_unorm_pack16,
+        .r4g4b4a4_unorm_pack16,
+        .b4g4r4a4_unorm_pack16,
+        .a4r4g4b4_unorm_pack16,
+        .a4b4g4r4_unorm_pack16,
+        .b5g6r5_unorm_pack16,
+        .r5g5b5a1_unorm_pack16,
+        .b5g5r5a1_unorm_pack16,
+        .r8_unorm,
+        .r8g8_unorm,
+        .r8g8b8a8_unorm,
+        .r8g8b8a8_srgb,
+        .b8g8r8a8_unorm,
+        .b8g8r8a8_srgb,
+        .a8b8g8r8_unorm_pack32,
+        .a8b8g8r8_srgb_pack32,
+        .a2b10g10r10_unorm_pack32,
+        .a2r10g10b10_unorm_pack32,
+        .r16_sfloat,
+        .r16g16_sfloat,
+        .r16g16b16a16_sfloat,
+        .r32_sfloat,
+        .r32g32_sfloat,
+        .r32g32b32a32_sfloat,
+        .b10g11r11_ufloat_pack32,
+        .r8_uint,
+        .r8_sint,
+        .r8g8_uint,
+        .r8g8_sint,
+        .r8g8b8a8_uint,
+        .r8g8b8a8_sint,
+        .a8b8g8r8_uint_pack32,
+        .a8b8g8r8_sint_pack32,
+        .a2b10g10r10_uint_pack32,
+        .a2r10g10b10_uint_pack32,
+        .r16_unorm,
+        .r16_uint,
+        .r16_sint,
+        .r16g16_unorm,
+        .r16g16_uint,
+        .r16g16_sint,
+        .r16g16b16a16_unorm,
+        .r16g16b16a16_uint,
+        .r16g16b16a16_sint,
+        .r32_uint,
+        .r32_sint,
+        .r32g32_uint,
+        .r32g32_sint,
+        .r32g32b32a32_uint,
+        .r32g32b32a32_sint,
+        => {
+            properties.optimal_tiling_features.color_attachment_bit = true;
+            properties.optimal_tiling_features.blit_dst_bit = true;
+        },
+        .s8_uint,
+        .d16_unorm,
+        .d32_sfloat, // note: either vk_format_d32_sfloat or vk_format_x8_d24_unorm_pack32 must be supported
+        .d32_sfloat_s8_uint,
+        => { // note: either vk_format_d24_unorm_s8_uint or vk_format_d32_sfloat_s8_uint must be supported
+            properties.optimal_tiling_features.depth_stencil_attachment_bit = true;
+        },
+
+        else => {},
+    }
+
+    if (base.format.supportsColorAttachemendBlend(format)) {
+        properties.optimal_tiling_features.color_attachment_blend_bit = true;
+    }
+
+    switch (format) {
+        .r8_unorm,
+        .r8_snorm,
+        .r8_uscaled,
+        .r8_sscaled,
+        .r8_uint,
+        .r8_sint,
+        .r8g8_unorm,
+        .r8g8_snorm,
+        .r8g8_uscaled,
+        .r8g8_sscaled,
+        .r8g8_uint,
+        .r8g8_sint,
+        .r8g8b8a8_unorm,
+        .r8g8b8a8_snorm,
+        .r8g8b8a8_uscaled,
+        .r8g8b8a8_sscaled,
+        .r8g8b8a8_uint,
+        .r8g8b8a8_sint,
+        .b8g8r8a8_unorm,
+        .a8b8g8r8_unorm_pack32,
+        .a8b8g8r8_snorm_pack32,
+        .a8b8g8r8_uscaled_pack32,
+        .a8b8g8r8_sscaled_pack32,
+        .a8b8g8r8_uint_pack32,
+        .a8b8g8r8_sint_pack32,
+        .a2r10g10b10_unorm_pack32,
+        .a2r10g10b10_snorm_pack32,
+        .a2r10g10b10_uint_pack32,
+        .a2r10g10b10_sint_pack32,
+        .a2b10g10r10_unorm_pack32,
+        .a2b10g10r10_snorm_pack32,
+        .a2b10g10r10_uint_pack32,
+        .a2b10g10r10_sint_pack32,
+        .r16_unorm,
+        .r16_snorm,
+        .r16_uscaled,
+        .r16_sscaled,
+        .r16_uint,
+        .r16_sint,
+        .r16_sfloat,
+        .r16g16_unorm,
+        .r16g16_snorm,
+        .r16g16_uscaled,
+        .r16g16_sscaled,
+        .r16g16_uint,
+        .r16g16_sint,
+        .r16g16_sfloat,
+        .r16g16b16a16_unorm,
+        .r16g16b16a16_snorm,
+        .r16g16b16a16_uscaled,
+        .r16g16b16a16_sscaled,
+        .r16g16b16a16_uint,
+        .r16g16b16a16_sint,
+        .r16g16b16a16_sfloat,
+        .r32_uint,
+        .r32_sint,
+        .r32_sfloat,
+        .r32g32_uint,
+        .r32g32_sint,
+        .r32g32_sfloat,
+        .r32g32b32_uint,
+        .r32g32b32_sint,
+        .r32g32b32_sfloat,
+        .r32g32b32a32_uint,
+        .r32g32b32a32_sint,
+        .r32g32b32a32_sfloat,
+        => properties.buffer_features.vertex_buffer_bit = true,
+        else => {},
+    }
+
+    switch (format) {
+        // Vulkan 1.1 mandatory
+        .r8_unorm,
+        .r8_snorm,
+        .r8_uint,
+        .r8_sint,
+        .r8g8_unorm,
+        .r8g8_snorm,
+        .r8g8_uint,
+        .r8g8_sint,
+        .r8g8b8a8_unorm,
+        .r8g8b8a8_snorm,
+        .r8g8b8a8_uint,
+        .r8g8b8a8_sint,
+        .b8g8r8a8_unorm,
+        .a8b8g8r8_unorm_pack32,
+        .a8b8g8r8_snorm_pack32,
+        .a8b8g8r8_uint_pack32,
+        .a8b8g8r8_sint_pack32,
+        .a2b10g10r10_unorm_pack32,
+        .a2b10g10r10_uint_pack32,
+        .r16_uint,
+        .r16_sint,
+        .r16_sfloat,
+        .r16g16_uint,
+        .r16g16_sint,
+        .r16g16_sfloat,
+        .r16g16b16a16_uint,
+        .r16g16b16a16_sint,
+        .r16g16b16a16_sfloat,
+        .r32_uint,
+        .r32_sint,
+        .r32_sfloat,
+        .r32g32_uint,
+        .r32g32_sint,
+        .r32g32_sfloat,
+        .r32g32b32a32_uint,
+        .r32g32b32a32_sint,
+        .r32g32b32a32_sfloat,
+        .b10g11r11_ufloat_pack32,
+        // optional
+        .a2r10g10b10_unorm_pack32,
+        .a2r10g10b10_uint_pack32,
+        => properties.buffer_features.uniform_texel_buffer_bit = true,
+        else => {},
+    }
+
+    if (properties.optimal_tiling_features.toInt() != 0) {
+        // "Formats that are required to support VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT must also support
+        //  VK_FORMAT_FEATURE_TRANSFER_SRC_BIT and VK_FORMAT_FEATURE_TRANSFER_DST_BIT."
+
+        properties.linear_tiling_features.transfer_src_bit = true;
+        properties.linear_tiling_features.transfer_dst_bit = true;
+    }
+
+    return properties;
+}
+
+pub fn getImageFormatProperties(
+    interface: *Interface,
+    format: vk.Format,
+    image_type: vk.ImageType,
+    tiling: vk.ImageTiling,
+    usage: vk.ImageUsageFlags,
+    _: vk.ImageCreateFlags,
+) VkError!vk.ImageFormatProperties {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    if (!try self.isFormatSupported(format, image_type, tiling, usage))
+        return VkError.FormatNotSupported;
+
+    const properties: vk.ImageFormatProperties = .{
+        .max_extent = .{ .width = 0, .height = 0, .depth = 1 },
+        .max_mip_levels = 1,
+        .max_array_layers = 1,
+        .sample_counts = .{ .@"1_bit" = true },
+        .max_resource_size = std.math.maxInt(u32),
+    };
+
+    return properties;
+}
+
+/// Phi does not support sparse images.
+pub fn getSparseImageFormatProperties(
+    interface: *Interface,
+    format: vk.Format,
+    image_type: vk.ImageType,
+    samples: vk.SampleCountFlags,
+    tiling: vk.ImageTiling,
+    usage: vk.ImageUsageFlags,
+    properties: ?[*]vk.SparseImageFormatProperties,
+) VkError!u32 {
+    _ = interface;
+    _ = format;
+    _ = image_type;
+    _ = samples;
+    _ = tiling;
+    _ = usage;
+    _ = properties;
+    return 0;
+}
+
+/// Phi does not support sparse images.
+pub fn getSparseImageFormatProperties2(
+    interface: *Interface,
+    format: vk.Format,
+    image_type: vk.ImageType,
+    samples: vk.SampleCountFlags,
+    tiling: vk.ImageTiling,
+    usage: vk.ImageUsageFlags,
+    properties: ?[*]vk.SparseImageFormatProperties2,
+) VkError!u32 {
+    _ = interface;
+    _ = format;
+    _ = image_type;
+    _ = samples;
+    _ = tiling;
+    _ = usage;
+    _ = properties;
+    return 0;
+}
+
+fn isFormatSupported(
+    self: *Self,
+    format: vk.Format,
+    image_type: vk.ImageType,
+    tiling: vk.ImageTiling,
+    usage: vk.ImageUsageFlags,
+) VkError!bool {
+    const format_properties = try self.interface.getFormatProperties(format);
+    const format_features = switch (tiling) {
+        .linear => format_properties.linear_tiling_features,
+        .optimal => format_properties.optimal_tiling_features,
+        else => return false,
+    };
+
+    if (!checkFormatUsage(usage, format_features))
+        return false;
+
+    const all_recognized_usages: vk.ImageUsageFlags = .{
+        .sampled_bit = true,
+        .storage_bit = true,
+        .color_attachment_bit = true,
+        .depth_stencil_attachment_bit = true,
+        .input_attachment_bit = true,
+        .transfer_src_bit = true,
+        .transfer_dst_bit = true,
+        .transient_attachment_bit = true,
+    };
+
+    if (usage.subtract(all_recognized_usages).toInt() != 0)
+        return false;
+
+    if (usage.sampled_bit) {
+        if (tiling != .linear and !format_features.sampled_image_bit)
+            return false;
+    }
+
+    if (tiling == .linear) {
+        if (image_type != .@"2d")
+            return false;
+        if (base.format.isDepth(format) or base.format.isStencil(format))
+            return false;
+    }
+
+    return true;
+}
+
+fn checkFormatUsage(usage: vk.ImageUsageFlags, features: vk.FormatFeatureFlags) bool {
+    if (usage.sampled_bit and !features.sampled_image_bit)
+        return false;
+    if (usage.storage_bit and !features.storage_image_bit)
+        return false;
+    if (usage.color_attachment_bit and !features.color_attachment_bit)
+        return false;
+    if (usage.depth_stencil_attachment_bit and !features.depth_stencil_attachment_bit)
+        return false;
+    if (usage.input_attachment_bit and !(features.color_attachment_bit or features.depth_stencil_attachment_bit))
+        return false;
+    if (usage.transfer_src_bit and !features.transfer_src_bit)
+        return false;
+    if (usage.transfer_dst_bit and !features.transfer_dst_bit)
+        return false;
+    return true;
+}
+
+pub fn getSurfaceSupportKHR(_: *Interface, _: u32, _: *SurfaceKHR) VkError!bool {
+    return true;
+}
+
+const CpuidRegs = packed struct {
+    eax: u32,
+    ebx: u32,
+    ecx: u32,
+    edx: u32,
+};
+
+fn cpuid(leaf_id: u32, subleaf_id: u32) CpuidRegs {
+    comptime {
+        switch (builtin.cpu.arch) {
+            .x86, .x86_64 => {},
+            else => @compileError("cpuid is only available on x86/x86_64"),
+        }
+    }
+
+    var eax: u32 = 0;
+    var ebx: u32 = 0;
+    var ecx: u32 = 0;
+    var edx: u32 = 0;
+
+    asm volatile ("cpuid"
+        : [_] "={eax}" (eax),
+          [_] "={ebx}" (ebx),
+          [_] "={ecx}" (ecx),
+          [_] "={edx}" (edx),
+        : [_] "{eax}" (leaf_id),
+          [_] "{ecx}" (subleaf_id),
+    );
+
+    return .{
+        .eax = eax,
+        .ebx = ebx,
+        .ecx = ecx,
+        .edx = edx,
+    };
+}
+
+fn writeU32Le(dst: []u8, offset: usize, value: u32) void {
+    dst[offset + 0] = @truncate(value);
+    dst[offset + 1] = @truncate(value >> 8);
+    dst[offset + 2] = @truncate(value >> 16);
+    dst[offset + 3] = @truncate(value >> 24);
+}

src/phi/PhiPipeline.zig

@@ -0,0 +1,37 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Pipeline;
+
+interface: Interface,
+
+pub fn createCompute(device: *base.Device, allocator: std.mem.Allocator, cache: ?*base.PipelineCache, info: *const vk.ComputePipelineCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.initCompute(device, allocator, cache, info);
+    interface.vtable = &.{ .destroy = destroy };
+
+    self.* = .{ .interface = interface };
+    return self;
+}
+
+pub fn createGraphics(device: *base.Device, allocator: std.mem.Allocator, cache: ?*base.PipelineCache, info: *const vk.GraphicsPipelineCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.initGraphics(device, allocator, cache, info);
+    interface.vtable = &.{ .destroy = destroy };
+
+    self.* = .{ .interface = interface };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiPipelineCache.zig

@@ -0,0 +1,26 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.PipelineCache;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.PipelineCacheCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+    interface.vtable = &.{ .destroy = destroy };
+
+    self.* = .{ .interface = interface };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiPipelineLayout.zig

@@ -0,0 +1,31 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.PipelineLayout;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.PipelineLayoutCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiQueryPool.zig

@@ -0,0 +1,32 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.QueryPool;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.QueryPoolCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.free(interface.queries);
+    allocator.destroy(self);
+}

src/phi/PhiQueue.zig

@@ -0,0 +1,88 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+const lib = @import("lib.zig");
+const proto = lib.proto;
+
+const PhiCommandBuffer = @import("PhiCommandBuffer.zig");
+const PhiDevice = @import("PhiDevice.zig");
+const PhiTransport = @import("PhiTransport.zig");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Queue;
+
+interface: Interface,
+
+pub fn create(allocator: std.mem.Allocator, device: *base.Device, index: u32, family_index: u32, flags: vk.DeviceQueueCreateFlags) VkError!*Interface {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(allocator, device, index, family_index, flags);
+    interface.dispatch_table = &.{
+        .bindSparse = bindSparse,
+        .submit = submit,
+        .waitIdle = waitIdle,
+    };
+
+    self.* = .{ .interface = interface };
+    return &self.interface;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) VkError!void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn bindSparse(interface: *Interface, info: []const vk.BindSparseInfo, fence: ?*base.Fence) VkError!void {
+    _ = interface;
+    _ = info;
+    _ = fence;
+    return VkError.FeatureNotPresent;
+}
+
+pub fn submit(interface: *Interface, infos: []Interface.SubmitInfo, fence: ?*base.Fence) VkError!void {
+    const device: *PhiDevice = @alignCast(@fieldParentPtr("interface", interface.owner));
+
+    for (infos) |info| {
+        for (info.wait_semaphores.items) |semaphore| {
+            try semaphore.wait();
+        }
+
+        for (info.command_buffers.items) |command_buffer| {
+            const phi_command_buffer: *PhiCommandBuffer = @alignCast(@fieldParentPtr("interface", command_buffer));
+
+            const work_execution_request: proto.PhiWorkExecutionRequest = .{
+                .cmd_count = phi_command_buffer.serialized_cmd_count,
+                .command_buffer_size = phi_command_buffer.commands.items.len,
+            };
+            const payload_size = @sizeOf(proto.PhiWorkExecutionRequest) + phi_command_buffer.commands.items.len;
+            const allocator = interface.host_allocator.allocator();
+            const payload = allocator.alloc(u8, payload_size) catch return VkError.OutOfHostMemory;
+            defer allocator.free(payload);
+
+            @memcpy(payload[0..@sizeOf(proto.PhiWorkExecutionRequest)], std.mem.asBytes(&work_execution_request));
+            @memcpy(payload[@sizeOf(proto.PhiWorkExecutionRequest)..], phi_command_buffer.commands.items);
+
+            // Synchronous queues for now
+            var reply = std.mem.zeroes(proto.PhiWorkExecutionReply);
+            try device.transport.request(proto.PHI_PACKET_WORK_EXECUTION, payload, std.mem.asBytes(&reply));
+
+            if (reply.result.status != proto.PHI_STATUS_OK) {
+                return PhiTransport.statusToErr(reply.result.status);
+            }
+        }
+
+        for (info.signal_semaphores.items) |semaphore| {
+            try semaphore.signal();
+        }
+    }
+    if (fence) |value| {
+        try value.signal();
+    }
+}
+
+pub fn waitIdle(interface: *Interface) VkError!void {
+    _ = interface;
+}

src/phi/PhiRenderPass.zig

@@ -0,0 +1,39 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.RenderPass;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.RenderPassCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+
+    interface.vtable = &.{
+        .destroy = destroy,
+        .getRenderAreaGranularity = getRenderAreaGranularity,
+    };
+
+    self.* = .{
+        .interface = interface,
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}
+
+pub fn getRenderAreaGranularity(_: *Interface) vk.Extent2D {
+    return .{
+        .width = 1,
+        .height = 1,
+    };
+}

src/phi/PhiSampler.zig

@@ -0,0 +1,26 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.Sampler;
+
+interface: Interface,
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.SamplerCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+    interface.vtable = &.{ .destroy = destroy };
+
+    self.* = .{ .interface = interface };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    allocator.destroy(self);
+}

src/phi/PhiShaderModule.zig

@@ -0,0 +1,44 @@
+const std = @import("std");
+const vk = @import("vulkan");
+const base = @import("base");
+
+const VkError = base.VkError;
+
+const Self = @This();
+pub const Interface = base.ShaderModule;
+
+interface: Interface,
+ref_count: std.atomic.Value(usize),
+
+pub fn create(device: *base.Device, allocator: std.mem.Allocator, info: *const vk.ShaderModuleCreateInfo) VkError!*Self {
+    const self = allocator.create(Self) catch return VkError.OutOfHostMemory;
+    errdefer allocator.destroy(self);
+
+    var interface = try Interface.init(device, allocator, info);
+    interface.vtable = &.{ .destroy = destroy };
+
+    self.* = .{
+        .interface = interface,
+        .ref_count = std.atomic.Value(usize).init(1),
+    };
+    return self;
+}
+
+pub fn destroy(interface: *Interface, allocator: std.mem.Allocator) void {
+    const self: *Self = @alignCast(@fieldParentPtr("interface", interface));
+    self.unref(allocator);
+}
+
+pub fn drop(self: *Self, allocator: std.mem.Allocator) void {
+    allocator.destroy(self);
+}
+
+pub fn ref(self: *Self) void {
+    _ = self.ref_count.fetchAdd(1, .monotonic);
+}
+
+pub fn unref(self: *Self, allocator: std.mem.Allocator) void {
+    if (self.ref_count.fetchSub(1, .release) == 1) {
+        self.drop(allocator);
+    }
+}

src/phi/PhiTransport.zig

@@ -0,0 +1,134 @@
+const std = @import("std");
+const base = @import("base");
+const lib = @import("lib.zig");
+const scif = @import("scif.zig");
+
+const VkError = base.VkError;
+const proto = lib.proto;
+
+const Self = @This();
+
+epd: scif.epd_t,
+sequence: u64 = 1,
+mutex: std.Io.Mutex = .init,
+instance: *base.Instance,
+
+pub fn init(instance: *base.Instance, node_id: u16) VkError!Self {
+    try scif.load();
+    errdefer scif.unload();
+
+    const epd = scif.open();
+    if (epd < 0) {
+        std.log.scoped(.PhiTransport).err("SCIF open failed", .{});
+        return VkError.InitializationFailed;
+    }
+    errdefer _ = scif.close(epd);
+
+    var dst: scif.PortId = .{
+        .node = node_id,
+        .port = @intCast(proto.PHI_SCIF_PORT),
+    };
+
+    if (scif.connect(epd, &dst) < 0) {
+        std.log.scoped(.PhiTransport).err("SCIF connection to node {d} port {d} failed", .{ dst.node, dst.port });
+        return VkError.InitializationFailed;
+    }
+
+    var self: Self = .{
+        .epd = epd,
+        .instance = instance,
+    };
+    try self.handshake();
+
+    std.log.scoped(.PhiTransport).info("Successfully connected", .{});
+    return self;
+}
+
+pub fn deinit(self: *Self) void {
+    _ = scif.close(self.epd);
+    scif.unload();
+    std.log.scoped(.PhiTransport).info("Closed connection", .{});
+}
+
+pub fn request(self: *Self, command: c_uint, payload: []const u8, reply_payload: []u8) VkError!void {
+    self.mutex.lock(self.instance.io()) catch return VkError.DeviceLost;
+    defer self.mutex.unlock(self.instance.io());
+
+    const sequence = self.sequence;
+    self.sequence += 1;
+
+    const header: proto.PhiMessageHeader = .{
+        .magic = proto.PHI_PROTOCOL_MAGIC,
+        .version = proto.PHI_PROTOCOL_VERSION,
+        .type = @intCast(command),
+        .sequence = sequence,
+        .payload_size = payload.len,
+    };
+
+    try self.writeAll(std.mem.asBytes(&header));
+    try self.writeAll(payload);
+
+    // SAFETY: will be entirely written with the readAll
+    var reply_header: proto.PhiMessageHeader = undefined;
+    try self.readAll(std.mem.asBytes(&reply_header));
+
+    if (reply_header.magic != proto.PHI_PROTOCOL_MAGIC or
+        reply_header.version != proto.PHI_PROTOCOL_VERSION or
+        reply_header.type != header.type or
+        reply_header.sequence != sequence or
+        reply_header.payload_size != reply_payload.len)
+    {
+        std.log.scoped(.PhiTransport).err("Invalid Phi reply header", .{});
+        return VkError.InitializationFailed;
+    }
+
+    try self.readAll(reply_payload);
+}
+
+pub fn statusToErr(status: c_int) VkError {
+    return switch (status) {
+        proto.PHI_STATUS_OUT_OF_MEMORY => VkError.OutOfDeviceMemory,
+        proto.PHI_STATUS_UNSUPPORTED_VERSION => VkError.InitializationFailed,
+        else => VkError.Unknown,
+    };
+}
+
+fn writeAll(self: *Self, bytes: []const u8) VkError!void {
+    var offset: usize = 0;
+    while (offset < bytes.len) {
+        const written = scif.send(self.epd, bytes[offset..].ptr, bytes.len - offset, scif.send_block);
+        if (written <= 0) {
+            return VkError.InitializationFailed;
+        }
+        offset += @intCast(written);
+    }
+}
+
+fn readAll(self: *Self, bytes: []u8) VkError!void {
+    var offset: usize = 0;
+    while (offset < bytes.len) {
+        const read = scif.recv(self.epd, bytes[offset..].ptr, bytes.len - offset, scif.recv_block);
+        if (read <= 0) {
+            return VkError.InitializationFailed;
+        }
+        offset += @intCast(read);
+    }
+}
+
+fn handshake(self: *Self) VkError!void {
+    const request_payload: proto.PhiHelloRequest = .{
+        .host_protocol_version = proto.PHI_PROTOCOL_VERSION,
+        .reserved = 0,
+    };
+
+    // SAFETY: will be entirely written by the request
+    var reply: proto.PhiHelloReply = undefined;
+    try self.request(proto.PHI_PACKET_HELLO, std.mem.asBytes(&request_payload), std.mem.asBytes(&reply));
+    if (reply.result.status != proto.PHI_STATUS_OK) {
+        return statusToErr(reply.result.status);
+    }
+    if (reply.device_protocol_version != proto.PHI_PROTOCOL_VERSION) {
+        std.log.scoped(.PhiTransport).err("Unsupported Phi protocol version {d}", .{reply.device_protocol_version});
+        return VkError.InitializationFailed;
+    }
+}

src/phi/lib.zig

@@ -0,0 +1,87 @@
+const std = @import("std");
+const vk = @import("vulkan");
+pub const base = @import("base");
+
+pub const c = @import("phi_c");
+pub const proto = @import("phi_protocol_c");
+pub const config = base.config;
+pub const mic = @import("miclib");
+pub const scif = @import("scif.zig");
+
+pub const PhiInstance = @import("PhiInstance.zig");
+pub const PhiDevice = @import("PhiDevice.zig");
+pub const PhiPhysicalDevice = @import("PhiPhysicalDevice.zig");
+pub const PhiQueue = @import("PhiQueue.zig");
+pub const PhiTransport = @import("PhiTransport.zig");
+
+pub const PhiBinarySemaphore = @import("PhiBinarySemaphore.zig");
+pub const PhiBuffer = @import("PhiBuffer.zig");
+pub const PhiBufferView = @import("PhiBufferView.zig");
+pub const PhiCommandBuffer = @import("PhiCommandBuffer.zig");
+pub const PhiCommandPool = @import("PhiCommandPool.zig");
+pub const PhiDescriptorPool = @import("PhiDescriptorPool.zig");
+pub const PhiDescriptorSet = @import("PhiDescriptorSet.zig");
+pub const PhiDescriptorSetLayout = @import("PhiDescriptorSetLayout.zig");
+pub const PhiDeviceMemory = @import("PhiDeviceMemory.zig");
+pub const PhiEvent = @import("PhiEvent.zig");
+pub const PhiFence = @import("PhiFence.zig");
+pub const PhiFramebuffer = @import("PhiFramebuffer.zig");
+pub const PhiImage = @import("PhiImage.zig");
+pub const PhiImageView = @import("PhiImageView.zig");
+pub const PhiPipeline = @import("PhiPipeline.zig");
+pub const PhiPipelineCache = @import("PhiPipelineCache.zig");
+pub const PhiPipelineLayout = @import("PhiPipelineLayout.zig");
+pub const PhiQueryPool = @import("PhiQueryPool.zig");
+pub const PhiRenderPass = @import("PhiRenderPass.zig");
+pub const PhiSampler = @import("PhiSampler.zig");
+pub const PhiShaderModule = @import("PhiShaderModule.zig");
+
+pub const Instance = PhiInstance;
+
+pub const driver_name = "Phi";
+
+pub const physical_device_default_name = "Intel(R) Xeon Phi(TM) Coprocessor";
+
+pub const vulkan_version = vk.makeApiVersion(
+    0,
+    config.phi_vulkan_version.major,
+    config.phi_vulkan_version.minor,
+    config.phi_vulkan_version.patch,
+);
+
+pub const std_options = base.std_options;
+
+comptime {
+    _ = base;
+}
+
+test {
+    std.testing.refAllDecls(PhiBinarySemaphore);
+    std.testing.refAllDecls(PhiBuffer);
+    std.testing.refAllDecls(PhiBufferView);
+    std.testing.refAllDecls(PhiCommandBuffer);
+    std.testing.refAllDecls(PhiCommandPool);
+    std.testing.refAllDecls(PhiDescriptorPool);
+    std.testing.refAllDecls(PhiDescriptorSet);
+    std.testing.refAllDecls(PhiDescriptorSetLayout);
+    std.testing.refAllDecls(PhiDevice);
+    std.testing.refAllDecls(PhiDeviceMemory);
+    std.testing.refAllDecls(PhiEvent);
+    std.testing.refAllDecls(PhiFence);
+    std.testing.refAllDecls(PhiFramebuffer);
+    std.testing.refAllDecls(PhiImage);
+    std.testing.refAllDecls(PhiImageView);
+    std.testing.refAllDecls(PhiInstance);
+    std.testing.refAllDecls(PhiPhysicalDevice);
+    std.testing.refAllDecls(PhiTransport);
+    std.testing.refAllDecls(scif);
+    std.testing.refAllDecls(PhiPipeline);
+    std.testing.refAllDecls(PhiPipelineCache);
+    std.testing.refAllDecls(PhiPipelineLayout);
+    std.testing.refAllDecls(PhiQueryPool);
+    std.testing.refAllDecls(PhiQueue);
+    std.testing.refAllDecls(PhiRenderPass);
+    std.testing.refAllDecls(PhiSampler);
+    std.testing.refAllDecls(PhiShaderModule);
+    std.testing.refAllDecls(base);
+}

src/phi/mic/Buffer.c

@@ -0,0 +1,66 @@
+#include <Buffer.h>
+
+#include <string.h>
+
+int PhiIsBufferCommand(const PhiCmdHeader* header)
+{
+	switch((PhiCmdType)header->type)
+	{
+		case PHI_CMD_COPY_BUFFER:
+		case PHI_CMD_FILL_BUFFER:
+			return 1;
+
+		default:
+			return 0;
+	}
+}
+
+static PhiStatus CopyBuffer(PhiCommandReader* reader)
+{
+	PhiCmdCopyBuffer command;
+	PhiStatus status = PhiReadCommandData(reader, &command, sizeof(command));
+	if(status != PHI_STATUS_OK)
+		return status;
+
+	if(command.src_memory == 0 || command.dst_memory == 0)
+		return PHI_STATUS_INVALID_HANDLE;
+
+	void* dst = (void*)((uintptr_t)command.dst_memory + (uintptr_t)command.dst_offset);
+	const void* src = (const void*)((uintptr_t)command.src_memory + (uintptr_t)command.src_offset);
+
+	memcpy(dst, src, (size_t)command.size);
+
+	return PHI_STATUS_OK;
+}
+
+static PhiStatus FillBuffer(PhiCommandReader* reader)
+{
+	PhiCmdFillBuffer command;
+	PhiStatus status = PhiReadCommandData(reader, &command, sizeof(command));
+	if(status != PHI_STATUS_OK)
+		return status;
+
+	if(command.memory == 0)
+		return PHI_STATUS_INVALID_HANDLE;
+
+	uint32_t* dst = (uint32_t*)((uintptr_t)command.memory + (uintptr_t)command.offset);
+
+	for(; command.size >= 4; command.size -= 4, dst++)
+		*dst = command.data;
+
+	return PHI_STATUS_OK;
+}
+
+PhiStatus PhiExecuteBufferCommand(PhiCommandReader* reader, const PhiCmdHeader* header)
+{
+	switch((PhiCmdType)header->type)
+	{
+		case PHI_CMD_COPY_BUFFER:
+			return CopyBuffer(reader);
+		case PHI_CMD_FILL_BUFFER:
+			return FillBuffer(reader);
+
+		default:
+			return PHI_STATUS_BAD_MESSAGE;
+	}
+}

src/phi/mic/Buffer.h

@@ -0,0 +1,9 @@
+#ifndef APE_PHI_BUFFER_H
+#define APE_PHI_BUFFER_H
+
+#include <CommandBuffer.h>
+
+int PhiIsBufferCommand(const PhiCmdHeader* header);
+PhiStatus PhiExecuteBufferCommand(PhiCommandReader* reader, const PhiCmdHeader* header);
+
+#endif

src/phi/mic/CommandBuffer.c

@@ -0,0 +1,96 @@
+#include <CommandBuffer.h>
+
+#include <Buffer.h>
+
+PhiStatus PhiReadCommandData(PhiCommandReader* reader, void* data, uint64_t size)
+{
+	if(reader->remaining < size)
+		return PHI_STATUS_BAD_MESSAGE;
+
+	if(ReadAll(reader->endpoint, data, (size_t)size) < 0)
+		return PHI_STATUS_BAD_MESSAGE;
+
+	reader->remaining -= size;
+	return PHI_STATUS_OK;
+}
+
+int PhiDrainCommandReader(PhiCommandReader* reader)
+{
+	if(reader->remaining == 0)
+		return 0;
+
+	int result = DrainPayload(reader->endpoint, reader->remaining);
+	reader->remaining = 0;
+	return result;
+}
+
+static PhiStatus ReadCommandHeader(PhiCommandReader* reader, PhiCmdHeader* command_header)
+{
+	PhiStatus status = PhiReadCommandData(reader, command_header, sizeof(*command_header));
+	if(status != PHI_STATUS_OK)
+		return status;
+
+	if(command_header->magic != PHI_COMMAND_MAGIC)
+		return PHI_STATUS_BAD_MESSAGE;
+
+	return PHI_STATUS_OK;
+}
+
+static PhiStatus ExecuteCommand(PhiCommandReader* reader, const PhiCmdHeader* command_header)
+{
+	if(PhiIsBufferCommand(command_header))
+		return PhiExecuteBufferCommand(reader, command_header);
+
+	return PHI_STATUS_BAD_MESSAGE;
+}
+
+int HandleWorkExecution(scif_epd_t endpoint, const PhiMessageHeader* header)
+{
+	PhiWorkExecutionRequest request;
+	PhiWorkExecutionReply reply = {
+		.result = {
+			.status = PHI_STATUS_OK,
+			.reserved = 0,
+		},
+	};
+
+	if(header->payload_size < sizeof(request))
+	{
+		if(DrainPayload(endpoint, header->payload_size) < 0)
+			return -1;
+		reply.result.status = PHI_STATUS_BAD_MESSAGE;
+		return SendReply(endpoint, header, &reply, sizeof(reply));
+	}
+	if(ReadAll(endpoint, &request, sizeof(request)) < 0)
+		return -1;
+
+	PhiCommandReader reader = {
+		.endpoint = endpoint,
+		.remaining = header->payload_size - sizeof(request),
+	};
+
+	if(reader.remaining != request.command_buffer_size)
+	{
+		if(PhiDrainCommandReader(&reader) < 0)
+			return -1;
+		reply.result.status = PHI_STATUS_BAD_MESSAGE;
+		return SendReply(endpoint, header, &reply, sizeof(reply));
+	}
+
+	for(uint64_t cmd_index = 0; cmd_index < request.cmd_count; ++cmd_index)
+	{
+		PhiCmdHeader cmd_header;
+		reply.result.status = ReadCommandHeader(&reader, &cmd_header);
+		if(reply.result.status != PHI_STATUS_OK)
+			break;
+
+		reply.result.status = ExecuteCommand(&reader, &cmd_header);
+		if(reply.result.status != PHI_STATUS_OK)
+			break;
+	}
+
+	if(reader.remaining > 0 && PhiDrainCommandReader(&reader) < 0)
+		return -1;
+
+	return SendReply(endpoint, header, &reply, sizeof(reply));
+}

src/phi/mic/CommandBuffer.h

@@ -0,0 +1,16 @@
+#ifndef APE_PHI_COMMAND_BUFFER_H
+#define APE_PHI_COMMAND_BUFFER_H
+
+#include <Daemon.h>
+
+typedef struct PhiCommandReader
+{
+	scif_epd_t endpoint;
+	uint64_t remaining;
+} PhiCommandReader;
+
+int HandleWorkExecution(scif_epd_t endpoint, const PhiMessageHeader* header);
+int PhiDrainCommandReader(PhiCommandReader* reader);
+PhiStatus PhiReadCommandData(PhiCommandReader* reader, void* data, uint64_t size);
+
+#endif

src/phi/mic/Daemon.c

@@ -0,0 +1,199 @@
+#include <Daemon.h>
+#include <CommandBuffer.h>
+#include <Logger.h>
+#include <Memory.h>
+
+static int HandleHello(scif_epd_t endpoint, const PhiMessageHeader* header)
+{
+	PhiHelloRequest request;
+	PhiHelloReply reply = {
+		.result = {
+			.status = PHI_STATUS_OK,
+			.reserved = 0,
+		},
+		.device_protocol_version = PHI_PROTOCOL_VERSION,
+		.pointer_bits = (uint32_t)(sizeof(void *) * 8u),
+	};
+
+	if(header->payload_size != sizeof(request))
+	{
+		if(DrainPayload(endpoint, header->payload_size) < 0)
+			return -1;
+		reply.result.status = PHI_STATUS_BAD_MESSAGE;
+		return SendReply(endpoint, header, &reply, sizeof(reply));
+	}
+
+	if(ReadAll(endpoint, &request, sizeof(request)) < 0)
+		return -1;
+
+	if(request.host_protocol_version != PHI_PROTOCOL_VERSION)
+		reply.result.status = PHI_STATUS_UNSUPPORTED_VERSION;
+
+	return SendReply(endpoint, header, &reply, sizeof(reply));
+}
+
+scif_epd_t StartDaemon()
+{
+	PhiLogInfo("Starting the daemon...");
+
+	scif_epd_t endpoint = scif_open();
+	if(endpoint < 0)
+	{
+		PhiLogError("Failed to create SCIF endpoint");
+		return 0;
+	}
+
+	if(scif_bind(endpoint, PHI_SCIF_PORT) < 0)
+	{
+		PhiLogError("Failed to bind SCIF port");
+		scif_close(endpoint);
+		return 0;
+	}
+
+	if(scif_listen(endpoint, 1) < 0)
+	{
+		PhiLogError("Could not listen to SCIF port");
+		scif_close(endpoint);
+		return 0;
+	}
+
+	return endpoint;
+}
+
+void ShutdownDaemon(scif_epd_t endpoint)
+{
+	PhiLogInfo("Shuting down the daemon...");
+	scif_close(endpoint);
+}
+
+int HandlePacket(scif_epd_t endpoint)
+{
+	for(;;)
+	{
+		PhiMessageHeader header;
+
+		if(ReadAll(endpoint, &header, sizeof(header)) < 0)
+			return -1;
+
+		if(header.magic != PHI_PROTOCOL_MAGIC || header.version != PHI_PROTOCOL_VERSION)
+		{
+			if(DrainPayload(endpoint, header.payload_size) < 0)
+				return -1;
+			if(SendStatus(endpoint, &header, PHI_STATUS_BAD_MESSAGE) < 0)
+				return -1;
+			continue;
+		}
+
+		switch((PhiPacketType)header.type)
+		{
+			case PHI_PACKET_HELLO:
+				if(HandleHello(endpoint, &header) < 0)
+					return -1;
+				break;
+
+			case PHI_PACKET_ALLOC_MEMORY:
+				if(HandleAllocMemory(endpoint, &header) < 0)
+					return -1;
+				break;
+
+			case PHI_PACKET_FREE_MEMORY:
+				if(HandleFreeMemory(endpoint, &header) < 0)
+					return -1;
+				break;
+
+			case PHI_PACKET_WORK_EXECUTION:
+				if(HandleWorkExecution(endpoint, &header) < 0)
+					return -1;
+				break;
+
+			case PHI_PACKET_SHUTDOWN:
+				if(DrainPayload(endpoint, header.payload_size) < 0)
+					return -1;
+				if(SendStatus(endpoint, &header, PHI_STATUS_OK) < 0)
+					return -1;
+				return 0;
+
+			default:
+				if(DrainPayload(endpoint, header.payload_size) < 0)
+					return -1;
+				if(SendStatus(endpoint, &header, PHI_STATUS_UNSUPPORTED_PACKET) < 0)
+					return -1;
+				break;
+		}
+	}
+}
+
+int ReadAll(scif_epd_t endpoint, void* data, size_t size)
+{
+	uint8_t* bytes = data;
+	size_t offset = 0;
+
+	while(offset < size)
+	{
+		int got = scif_recv(endpoint, bytes + offset, size - offset, SCIF_RECV_BLOCK);
+		if(got <= 0)
+			return -1;
+		offset += (size_t)got;
+	}
+
+	return 0;
+}
+
+int WriteAll(scif_epd_t endpoint, const void* data, size_t size)
+{
+	const uint8_t* bytes = data;
+	size_t offset = 0;
+
+	while(offset < size)
+	{
+		int sent = scif_send(endpoint, (void*)(bytes + offset), size - offset, SCIF_SEND_BLOCK);
+		if(sent <= 0)
+			return -1;
+		offset += (size_t)sent;
+	}
+
+	return 0;
+}
+
+int SendReply(scif_epd_t endpoint, const PhiMessageHeader* request, const void* payload, uint64_t payload_size)
+{
+	PhiMessageHeader reply = {
+		.magic = PHI_PROTOCOL_MAGIC,
+		.version = PHI_PROTOCOL_VERSION,
+		.type = request->type,
+		.sequence = request->sequence,
+		.payload_size = payload_size,
+	};
+
+	if(WriteAll(endpoint, &reply, sizeof(reply)) < 0)
+		return -1;
+
+	return WriteAll(endpoint, payload, (size_t)payload_size);
+}
+
+int SendStatus(scif_epd_t endpoint, const PhiMessageHeader* request, PhiStatus status)
+{
+	PhiFreeMemoryReply reply = {
+		.result = {
+			.status = status,
+			.reserved = 0,
+		},
+	};
+
+	return SendReply(endpoint, request, &reply, sizeof(reply));
+}
+
+int DrainPayload(scif_epd_t endpoint, uint64_t size)
+{
+	uint8_t buffer[256];
+
+	while(size > 0)
+	{
+		size_t chunk = size < sizeof(buffer) ? (size_t)size : sizeof(buffer);
+		if(ReadAll(endpoint, buffer, chunk) < 0)
+			return -1;
+		size -= chunk;
+	}
+
+	return 0;
+}

src/phi/mic/Daemon.h

@@ -0,0 +1,20 @@
+#ifndef APE_PHI_DAEMON_H
+#define APE_PHI_DAEMON_H
+
+#include <sys/types.h>
+
+#include <scif.h>
+
+#include <Protocol.h>
+
+scif_epd_t StartDaemon();
+void ShutdownDaemon(scif_epd_t endpoint);
+
+int DrainPayload(scif_epd_t endpoint, uint64_t size);
+int HandlePacket(scif_epd_t endpoint);
+int ReadAll(scif_epd_t endpoint, void* data, size_t size);
+int SendReply(scif_epd_t endpoint, const PhiMessageHeader* request, const void* payload, uint64_t payload_size);
+int SendStatus(scif_epd_t endpoint, const PhiMessageHeader* request, PhiStatus status);
+int WriteAll(scif_epd_t endpoint, const void* data, size_t size);
+
+#endif

src/phi/mic/Logger.c

@@ -0,0 +1,95 @@
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <time.h>
+#include <unistd.h>
+
+#include <Logger.h>
+
+#define RED 31
+#define GREEN 32
+#define BLUE 34
+#define DEF 0
+#define BLACK 30
+#define YELLOW 33
+#define MAGENTA 35
+#define CYAN 36
+#define WHITE 37
+#define BG_RED 41
+#define BG_GREEN 42
+#define BG_BLUE 44
+#define BG_DEF 0
+#define BG_BLACK 40
+#define BG_YELLOW 43
+#define BG_MAGENTA 45
+#define BG_CYAN 46
+#define BG_WHITE 47
+#define RESET 0
+#define BOLD 1
+#define UNDERLINE 4
+#define INVERSE 7
+#define BOLD_OFF 21
+#define UNDERLINE_OFF 24
+#define INVERSE_OFF 27
+
+inline static void SetConsoleColor(FILE* file, int code)
+{
+	fprintf(file, "\033[1;%dm", code);
+}
+
+void PhiLog(PhiLogLevel level, const char* fmt, const char* file, const char* function, int line, ...)
+{
+	time_t now = time(0);
+	struct tm tstruct = *localtime(&now);
+	char buffer[128];
+	strftime(buffer, sizeof(buffer), "%X", &tstruct);
+
+	FILE* out = stdout;
+
+	if(level != PHI_LOG_LEVEL_INFO)
+		out = stderr;
+
+	// Way too much printf calls
+	SetConsoleColor(out, MAGENTA);
+	fprintf(out, "[ApeDriver ");
+	SetConsoleColor(out, GREEN);
+	fprintf(out, "Phi ");
+	SetConsoleColor(out, YELLOW);
+	fprintf(out, "%s", buffer);
+	SetConsoleColor(out, MAGENTA);
+	fputc(']', out);
+
+	switch(level)
+	{
+		case PHI_LOG_LEVEL_INFO:
+			SetConsoleColor(out, BLUE);
+			fprintf(out, "[info] ");
+			break;
+		case PHI_LOG_LEVEL_WARN:
+			SetConsoleColor(out, MAGENTA);
+			fprintf(out, "[warn] ");
+			break;
+		case PHI_LOG_LEVEL_ERR:
+		case PHI_LOG_LEVEL_FATAL:
+			SetConsoleColor(out, RED);
+			fprintf(out, "[err]  ");
+			break;
+	}
+
+	SetConsoleColor(out, RESET);
+
+	va_list argptr;
+	va_start(argptr, line);
+	vfprintf(out, fmt, argptr);
+	va_end(argptr);
+
+	fputc('\n', out);
+
+	if(level == PHI_LOG_LEVEL_FATAL)
+	{
+		SetConsoleColor(out, BG_RED);
+		fprintf(out, "Fatal Error: emergency exit\n");
+		SetConsoleColor(out, BG_DEF);
+		abort();
+	}
+}