SPIR-V `DebugTypeMatrix` metadata differs between Slang, DXC, and glslang

[pdeayton-nv]

Issue Description

For a column-major HLSL float2x3 buffer member, Slang emits a different
NonSemantic.Shader.DebugInfo.100 DebugTypeMatrix shape from DXC and glslang. The repro also
includes a row-major buffer member and local matrices as comparison cases.

Reproducer Code

repro.hlsl:

cbuffer ColumnParameters
{
    column_major float2x3 columnMajorValue;
};

cbuffer RowParameters
{
    row_major float2x3 rowMajorValue;
};

RWStructuredBuffer<float3> result;

[numthreads(1, 1, 1)]
void main()
{
    // Control case: Slang, DXC, and glslang emit the same DebugTypeMatrix
    // shape for these local float2x3 matrices.
    float2x3 localFromColumnMajor = columnMajorValue;
    float2x3 localFromRowMajor = rowMajorValue;

    float2 inputValue = float2(1.0, 2.0);
    result[0] =
        mul(inputValue, localFromColumnMajor) +
        mul(inputValue, localFromRowMajor);
}

slangc repro.hlsl -target spirv-asm -entry main -stage compute -g2

Expected Behavior

Ideally, Slang's DebugTypeMatrix for columnMajorValue would be consistent with DXC and
glslang: a three-component vector, vector count two, and a true Column Major operand.

Actual Behavior

The observed debug matrix types are:

Declaration	Slang	DXC	glslang
column_major float2x3 columnMajorValue	DebugTypeMatrix(vec2, 3, false)	DebugTypeMatrix(vec3, 2, 1)	DebugTypeMatrix(vec3, 2, true)
row_major float2x3 rowMajorValue	DebugTypeMatrix(vec3, 2, true)	DebugTypeMatrix(vec3, 2, 1)	DebugTypeMatrix(vec3, 2, true)
local float2x3 matrices	DebugTypeMatrix(vec3, 2, true)	DebugTypeMatrix(vec3, 2, 1)	DebugTypeMatrix(vec3, 2, true)

DXC expresses the true Column Major operand as integer 1.

All three compilers emit the same core matrix type. Their member decorations also agree, changing
with the source layout independently of the debug type:

%mat2v3float = OpTypeMatrix %v3float 2
OpMemberDecorate %ColumnParameters 0 RowMajor
OpMemberDecorate %ColumnParameters 0 MatrixStride 16
OpMemberDecorate %RowParameters 0 ColMajor
OpMemberDecorate %RowParameters 0 MatrixStride 16

Environment

• Slang v2026.16-21-g30d3e0f3e (30d3e0f3e8c1fae72826988aa8576db344683834)
• DXC 1.8.0.5049 (-fspv-debug=vulkan-with-source)
• glslangValidator 15.4.0 (-gVS)

[nv-slang-bot]

Triage — SPIR-V DebugTypeMatrix wrong for column_major cbuffer matrix members

• Status: Reproduced on v2026.16-21-g30d3e0f3e (compile-time, -target spirv-asm -g2, no GPU needed). Classified bug / low / P3 / target-emit (SPIR-V). Type set to Bug; reproduced applied. Not a regression — see below.
• Root cause: Single emit site source/slang/slang-emit-spirv.cpp:11108-11138. The DebugTypeMatrix operands are computed from the source matrix-layout modifier: the Slang-column-major branch (:11115-11120) uses vector<elem, getRowCount()>, count = getColumnCount(), columnMajor = false → (vec2, 3, false); the row-major/else branch (:11122-11128) uses vector<elem, getColumnCount()>, count = getRowCount(), columnMajor = true → (vec3, 2, true). normalizeMatrixDebugType (:10899, early-returns unchanged for buffer types) forces row-major only for non-buffer types, so only column_major buffer members reach the buggy branch — locals and row_major members are already correct.
• Why the physical shape is the target: For all three layouts the core type is identical — %mat2v3float = OpTypeMatrix %v3float 2 — and a SPIR-V OpTypeMatrix is inherently an array of column vectors; the buffer packing layout is carried separately by the RowMajor/ColMajor member decoration + MatrixStride. So the debug type of the value should be layout-independent: (vec3, 2, columnMajor=true), matching DXC and glslang. The current code conflates value-type with storage-packing.
• Not a regression: this exact behavior was introduced deliberately by Fix emission of SPIR-V DebugTypeMatrix matrix layout operand #9708 (merged 2026-01-26, Fixes #9325), whose test tests/spirv/debug-matrix-layout.slang (lines 54/63) locks in the column_major buffer member → %false variant. A fix revisits that intentional decision and must update that test (and re-verify debug-matrix-layout-non-buffer.slang). Lineage DebugInfo for locally-stored matrices is incorrect when using -matrix-layout-column-major #9325 / Fix debug info for local matrices with -matrix-layout-column-major #9326 / Fix emission of SPIR-V DebugTypeMatrix matrix layout operand #9708 are related, not duplicates.
• Next-action: Handed to slang-fixer with a briefing memo. Recommended fix: make DebugTypeMatrix unconditionally describe the physical OpTypeMatrix (vec(getColumnCount()), getRowCount(), columnMajor=true), matching DXC/glslang — smallest correct change; also evaluate removing the now-redundant normalizeMatrixDebugType. Because this reverses part of Fix emission of SPIR-V DebugTypeMatrix matrix layout operand #9708, a maintainer/author semantics sign-off is advisable (cc the DebugInfo for locally-stored matrices is incorrect when using -matrix-layout-column-major #9325 reviewer). Blocker: none for investigation; fix pending review.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

@nv-slang-bot can this issue be resolved by the following PR?

• Canonicalize SPIR-V matrix debug types #12860

[nv-slang-bot]

@jkwak-work Yes — PR #12860 ("Canonicalize SPIR-V matrix debug types", Fixes #12757) is the appropriate resolution. I read its diff against the root cause and it addresses this issue directly:

• Emit site (the IRMatrixType arm of emitDebugTypeImpl in slang-emit-spirv.cpp): the previous layout-keyed branch is removed, so DebugTypeMatrix is now emitted unconditionally in SPIR-V's column-vector form — vector<elem, columnCount>, count rowCount, Column Major = true. For the reported column_major float2x3 cbuffer member this yields DebugTypeMatrix(vec3, 2, true) instead of the old (vec2, 3, false), matching DXC and glslang.
• Layout is still carried independently by the RowMajor/ColMajor member decorations and MatrixStride (unchanged), so buffer storage semantics are unaffected — only the debug metadata is canonicalized.
• normalizeMatrixDebugType drops its buffer/non-buffer split and now canonicalizes every matrix type to one row-major cache key, so all layout variants of the same logical matrix share a single debug type (no more per-layout divergence).
• Tests (tests/spirv/debug-matrix-layout.slang, debug-matrix-layout-non-buffer.slang) are updated so buffer members and locals all reference the one canonical DebugTypeMatrix, with an added check rejecting the row-vector (Column Major = false) form anywhere in the module.

This is exactly the canonical/physical approach our triage recommended (the debug type should describe the physical OpTypeMatrix, which is layout-independent; packing lives in the member decorations). We also flagged that a fix reverses the intentional behavior from #9708, so author/maintainer sign-off was warranted — #12860 being the original author's own PR provides that. A human reviewer should still confirm the diff and CI before merge.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

@nv-slang-bot I am sorry, I meant this PR:

• Fix bug in matrix specialization #12840

[nv-slang-bot]

@jkwak-work Thanks for the correction. I read #12840's diff, and it does not resolve this issue — it fixes a different matrix-layout bug:

• Fix bug in matrix specialization #12840 ("Fix bug in matrix specialization", pr: breaking change) targets a functional bug: a matrix written without a row_major/column_major keyword keeps its layout unresolved until a later pass; if it reaches a generic first, specialization bakes in the unresolved layout and builds a matrix type the pass already ran past, so two versions of one type reach codegen and the SPIR-V fails validation. Its fix resolves layout before specialization. Its only change to the debug-info path is a mechanical one — in normalizeMatrixDebugType it retypes the row-major literal to the new MatrixLayoutMode enum (getLayout()->getFullType()). It does not change the DebugTypeMatrix shape logic: the IRMatrixType arm of emitDebugTypeImpl (the layout-keyed branch that emits (vec2, 3, Column Major=false) for a column_major cbuffer member) is untouched. Since this issue's repro uses an explicit column_major (an already-resolved layout), Fix bug in matrix specialization #12840 leaves the reported debug metadata unchanged.

• This issue (the DebugTypeMatrix shape) is resolved by Canonicalize SPIR-V matrix debug types #12860 ("Canonicalize SPIR-V matrix debug types"), which declares Fixes #12757 and rewrites the emit to always use SPIR-V's column-vector form (vec3, count 2, Column Major = true for the reported float2x3), matching DXC and glslang. That is the PR that closes this issue.

• Relationship / heads-up: the two PRs are orthogonal (functional specialization vs debug-type metadata) but they both edit normalizeMatrixDebugType, so whichever merges second will need a small rebase. There's no semantic conflict; Canonicalize SPIR-V matrix debug types #12860's canonicalization should just preserve Fix bug in matrix specialization #12840's enum-typed layout literal rather than reintroducing the plain-int one.

So: #12840 for the specialization/validation bug, #12860 for this debug-type issue. A human should still confirm both diffs and CI before merge.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}