[AMDGPU][GISel] Add Identity BUILD_VECTOR Combines

[Pierre-vh]

Summary

Folds-away BUILD_VECTOR-related noops in the post-legalizer combiner.

Depends on D134433

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

48–52 ↗

(On Diff #464226)

Can move this to the generic combiner

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

333 ↗

(On Diff #464226)

Shouldn't need to bother with getDefIgnoringCopies, copy of copy is folded out during combining. We should have an m_Undef or m_ImplicitDef too if there isn't one already

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.1d.d16.ll

947–949

Should remove this and remove the unused prefixes

llvm/test/CodeGen/AMDGPU/GlobalISel/postlegalizer-combiner-buildvector-identities.mir

3

Don't need -global-isel with -run-pass

llvm/test/CodeGen/AMDGPU/GlobalISel/postlegalizer-combiner-trunc-bitcast-buildvector.mir

3

Don't need -global-isel with -run-pass

[Pierre-vh]

Comments

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

48 ↗

(On Diff #464237)

Left behind AMDGPU copy

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

308–309 ↗

(On Diff #464237)

Left behind the AMDGPU copy

[arsenm]

LGTM except for the leftover copy

[Pierre-vh]

Comment

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

48 ↗

(On Diff #464237)

That's another combine for (G_TRUNC (G_BITCAST (G_BUILD_VECTOR x, y), unless that one can also go in the generic combiner?

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

48 ↗

(On Diff #464237)

Yes, all of this can. It's not doing anything that target specific and uses no custom opcodes

[Pierre-vh]

Move all to Helper