[SM6.10] LinAlg Validation: MatrixLoadFromMemory

[Ashley Coleman (V-FEXrt)]

Fixes #8499

Implements LinAlg MatrixLoadFromMemory validation rules

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[Damyan Pepper (damyanp)]

LGTM - although this also looks extremely similar to the MatrixStoreToMemory version.

[Copilot]

🟡 Changes recommended

Pointer handling can crash the validator, while bounds and alignment calculations use incorrect memory units.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds SM 6.10 validation for LinAlg matrix loads from groupshared memory.

Changes:

• Validates matrix scope, storage type, capacity, offset, and stride.
• Adds DXIL validation coverage.
• Updates code-generation fixtures.

File summaries

File	Description
lib/DxilValidation/DxilValidation.cpp	Implements load validation rules.
tools/clang/test/LitDXILValidation/LinAlgMatrix/linalgmatrix-matrixloadfrommemory.ll	Adds validator tests.
tools/clang/test/CodeGenDXIL/hlsl/linalg/builtins/matrixloadfrommemory/nominal.hlsl	Updates scalar-array expectations.
tools/clang/test/CodeGenDXIL/hlsl/linalg/builtins/matrixloadfrommemory/vector-array.hlsl	Updates vector-array expectations.

Review details

Suppressed comments (3)

lib/DxilValidation/DxilValidation.cpp:1643

• This is not the memory footprint of the read: it ignores the GEP/start offset, runtime Offset, Stride, and row/column layout, and it also undercounts 64-bit matrices stored through i32 (for example, a 4x4 I64 matrix needs 32 i32 slots, not 16). As a result, the new test's “okay” load starts at element 128 of a 64-element global and is accepted. Compute the last byte/slot touched from the resolved base, storage element width, offset, stride, layout, and matrix dimensions, then compare that footprint with the global allocation.

  unsigned ElementsPerScalar = ComponentTypeElementsPerScalar(RetMat->Type);
  unsigned ExpectedScalarCount =
      (RetMat->N + ElementsPerScalar - 1) / ElementsPerScalar * RetMat->M;
  if (ExpectedScalarCount > GSScalarCount)

lib/DxilValidation/DxilValidation.cpp:1653

• Offset is an element index, not a byte count: the public API names it StartIdx, and LinAlgTests.cpp:7896 divides OffsetBytes by the element size before passing it. Checking the raw index against 128 rejects valid offsets such as 32 for float (128 bytes). Convert the index to a byte offset using the memory pointee's allocation size before enforcing 128-byte alignment.

  if (ConstantInt *OffsetV = dyn_cast<ConstantInt>(Op.get_offset())) {
    unsigned Offset = OffsetV->getLimitedValue();
    if (Offset % 128 != 0)
      ValCtx.EmitInstrFormatError(CI, ValidationRule::InstrParamMultiple,
                                  {"Offset", "128", std::to_string(Offset)});

lib/DxilValidation/DxilValidation.cpp:1661

• Like Offset, the groupshared Stride is expressed in memory elements (LinAlgTests.cpp:7897 passes StrideBytes / ElementBytes). Applying % 16 directly rejects a valid float stride of 4 elements (16 bytes), and is even more restrictive for vector arrays. Scale by the pointee allocation size before checking the 16-byte requirement.

  if (ConstantInt *StrideV = dyn_cast<ConstantInt>(Op.get_stride())) {
    unsigned Stride = StrideV->getLimitedValue();
    if (Stride % 16 != 0)
      ValCtx.EmitInstrFormatError(CI, ValidationRule::InstrParamMultiple,
                                  {"Stride", "16", std::to_string(Stride)});

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Joshua Batista (bob80905)]

lgtm