[202x][unroll] Only run the default LLVM loop unroll for 202x onward

[Farzon Lotfi (farzonl)]

resolves #8789

This change maintains the old loop unroll behavior for language modes below 202x.

The 202x unroll behavior is for this to become a hint that is mostly used for O3 optimizations.

We will do this by passing a language-version-derived flag through PassManagerBuilder to DxilLoopUnroll. For HLSL 202x and later, DxilLoopUnroll will leave loops carrying llvm.loop.unroll.count unchanged. The standard unroller may then consume the metadata at -O3. For older language versions, DxilLoopUnroll will continue using the count as an execution limit.

The custom pass will continue handling llvm.loop.unroll.full in every language version. This keeps bare [unroll] behavior and its diagnostics unchanged.

Assisted by GPT 5.6-Sol via Copilot

[Copilot]

Pull request overview

Updates HLSL 202x [unroll(N)] semantics while preserving legacy behavior.

Changes:

• Treats unroll counts as optimization hints in HLSL 202x.
• Adds O2/O3 coverage for fixed and runtime loop counts.
• Documents the user-visible behavior change.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tools/clang/test/HLSLFileCheck/hlsl/control_flow/attributes/unroll/count_runtime_hint_202x.hlsl	Tests runtime-count loops.
tools/clang/test/HLSLFileCheck/hlsl/control_flow/attributes/unroll/count_less_than_i.hlsl	Pins legacy semantics to HLSL 2021.
tools/clang/test/HLSLFileCheck/hlsl/control_flow/attributes/unroll/count_hint_202x.hlsl	Tests fixed-count loops.
tools/clang/test/HLSLFileCheck/hlsl/control_flow/attributes/unroll/count_greater_than_i.hlsl	Pins legacy semantics to HLSL 2021.
tools/clang/test/HLSLFileCheck/hlsl/control_flow/attributes/unroll/count_cbuff.hlsl	Pins runtime legacy behavior to HLSL 2021.
tools/clang/lib/CodeGen/BackendUtil.cpp	Derives hint behavior from language version.
lib/Transforms/Scalar/DxilLoopUnroll.cpp	Preserves count metadata in HLSL 202x.
lib/Transforms/IPO/PassManagerBuilder.cpp	Passes the language-derived setting to unrolling.
include/llvm/Transforms/Scalar.h	Extends the pass factory interface.
include/llvm/Transforms/IPO/PassManagerBuilder.h	Adds the pass-manager setting.
docs/ReleaseNotes.md	Records the language behavior change.

---

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

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 1 comment.

[Finn Plummer (inbelic)]

Makes sense as a hint

[Farzon Lotfi (farzonl)]

Need to investigate the following failures before we take out of draft

    Clang :: DXC/dxopt_test.test
    Clang-Unit :: HLSL/ClangHLSLTests/CompilerTest.BatchHLSL
    Clang-Unit :: HLSL/ClangHLSLTests/CompilerTest.CompileWhenODumpThenOptimizerMatch
    Clang-Unit :: HLSL/ClangHLSLTests/OptimizerTest.OptimizerWhenSlice0ThenOK
    Clang-Unit :: HLSL/ClangHLSLTests/OptimizerTest.OptimizerWhenSlice1ThenOK
    Clang-Unit :: HLSL/ClangHLSLTests/OptimizerTest.OptimizerWhenSlice2ThenOK
    Clang-Unit :: HLSL/ClangHLSLTests/OptimizerTest.OptimizerWhenSlice3ThenOK
    Clang-Unit :: HLSL/ClangHLSLTests/OptimizerTest.OptimizerWhenSliceWithIntermediateOptionsThenOK

[Copilot]

Pull request overview

Copilot reviewed 13 out of 13 changed files in this pull request and generated no new comments.

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

tools/clang/test/HLSLFileCheck/hlsl/control_flow/attributes/unroll/count_runtime_hint_202x.hlsl:8

• Use sentence-case capitalization here and remove the trailing whitespace.

// O2 Doesn't unroll and O3 Runtime unrolling only supports 
// power-of-two factors, so the count hint is
// consumed without cloning the loop body.

[Chris B (llvm-beanz)]

Can you also put up a PR to tc57 to add the proposed spec language into the proposal?

We had language in the hlsl-specs draft, but it hasn't been run through the committee yet.