[all-commits] [llvm/llvm-project] 9d5881: [AMDGPU] Fall back for FP8-to-half conversions wit...

[Chinmay Deshpande via All-commits]

  Branch: refs/heads/users/chinmaydd/fix-fp8-2
  Home:   https://github.com/llvm/llvm-project
  Commit: 9d5881f8ccf22c7ff2fc3c0777ca56f33dbefa79
      https://github.com/llvm/llvm-project/commit/9d5881f8ccf22c7ff2fc3c0777ca56f33dbefa79
  Author: Chinmay Deshpande <chdeshpa at amd.com>
  Date:   2026-09-22 (Tue, 22 Sep 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/SIISelLowering.cpp
    A llvm/test/CodeGen/AMDGPU/arbitrary-fp-to-half-fallback.ll

  Log Message:
  -----------
  [AMDGPU] Fall back for FP8-to-half conversions without native support

The custom v2i8 source action also handles half results on gfx950, which
has native OCP FP8-to-f32 conversions but lacks FP8-to-f16 instructions.
Converting FP8 vectors to half therefore reached lowerFromFP8 and asserted
that FP8F16ConversionInsts was available.

Use generic legalization for half results when the subtarget lacks the
required instructions. Preserve native f32 conversions and native f16
conversions on targets that support them.

Add gfx950 regression coverage for both OCP encodings, including vectors
that widen or split into pairs, and check that native f32 instructions
remain selected.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications