Fix incorrect constant folding of CompareScalar True/False comparisons

[tannergooding]

Why

Avx.CompareScalar with a constant "always true"/"always false" float comparison mode (OrderedFalse*, UnorderedTrue*) was being constant folded to a full Zero/AllBitsSet vector. That is wrong for the scalar form: CompareScalar only writes the lowest element and must copy the remaining upper elements unchanged from op1. The fold discarded those upper elements, producing incorrect results (e.g. CompareScalar(<1,2,3,4>, ..., OrderedFalse) yielded <0,0,0,0> instead of <0,2,3,4>).

Separately, gtNewZeroConNode and gtNewAllBitsSetConNode did not support TYP_MASK (they hit unreached()), so folding the AVX512 mask compare (NI_AVX512_CompareMask, whose result type is TYP_MASK) for these modes crashed.

Fixes #125160

Approach

• gtNewZeroConNode now handles TYP_MASK by returning a zero GenTreeMskCon (guarded by FEATURE_MASKED_HW_INTRINSICS).
• gtFoldExprHWIntrinsic True/False handling for NI_AVX_Compare / NI_AVX_CompareScalar / NI_AVX512_CompareMask:
  • For CompareScalar, bail out of folding entirely unless op1 is itself a constant. When it is, set only the lowest element (0 for false modes, all-ones for true modes) and leave the upper elements of op1 intact.
  • For the non-scalar false case, use the now mask-safe gtNewZeroConNode(retType).
  • For the non-scalar true case producing a mask, build a canonical true mask via simdmask_t::AllBitsSet(elementCount) (matching VNAllBitsForType) rather than relying on gtNewAllBitsSetConNode, which cannot know the element count. Vector (non-mask) results still use gtNewAllBitsSetConNode.

Notes

• The scalar fold reuses the constant op1 node as the result and wraps it with gtWrapWithSideEffects(tree, ...); because op1 is a side-effect-free constant, this preserves any side effects from the other operands without duplicating op1.
• Direct bit writes (gtSimdVal.u32[0] / u64[0]) are used for the true-mode all-ones pattern since SetElementIntegral asserts on floating base types.

Testing

Added Runtime_125160 regression test covering all four True/False modes for float and double Avx.CompareScalar (asserting both the low-element result and upper-element preservation) plus Avx512F.Compare for Vector512<float>/<double>. Verified the test fails without the fix (Expected 2, Actual 0) and passes with it. Baseline clr+libs (checked) build is clean.

Note

This pull request was authored by GitHub Copilot.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR corrects JIT constant-folding for AVX compare intrinsics in the “always true/always false” comparison modes, ensuring Avx.CompareScalar preserves the upper elements from op1 and avoiding crashes when folding AVX512 mask-typed compares.

Changes:

• Extend gtNewZeroConNode to support TYP_MASK (via GenTreeMskCon) under FEATURE_MASKED_HW_INTRINSICS.
• Update gtFoldExprHWIntrinsic folding for NI_AVX_Compare / NI_AVX_CompareScalar / NI_AVX512_CompareMask, with special handling for CompareScalar to only fold when op1 is a constant vector and to preserve upper elements.
• Add and wire up a new JIT regression test Runtime_125160.

Show a summary per file

File	Description
src/coreclr/jit/gentree.cpp	Fixes compare folding behavior for scalar compares and makes zero-constant creation mask-safe.
src/tests/JIT/Regression/JitBlue/Runtime_125160/Runtime_125160.cs	Adds regression coverage for CompareScalar true/false modes and AVX512 compare folding scenarios.
src/tests/JIT/Regression/Regression_ro_2.csproj	Includes the new Runtime_125160 test in the regression suite.

Copilot's findings

• Files reviewed: 3/3 changed files
• Comments generated: 3

[Copilot]

Copilot's findings

• Files reviewed: 3/3 changed files
• Comments generated: 3