[Codegen] X86 codegen does not generate unpck{h|l}ps instructions for x86-64 CPU target.

[karthik-senthil]

For functionally similar LLVM-IR between two cases, it was noticed that X86 codegen does not generate unpck{h|l}ps instructions (for x86-64 cpu target) in one of them -

Case 1 : Scalarized stores with no extracts for GEP index (https://godbolt.org/z/qcqr8f461)
Case 2 : Scalarized stores with extracts for GEP index (https://godbolt.org/z/58n95Tv5q) - no unpck{h|l}ps instructions

I would like to confirm if there are any issues with the LLVM-IR for Case 2 that prohibits X86 codegen from generating optimal ASM like Case 1.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Karthik Senthil (karthik-senthil)

For functionally similar LLVM-IR between two cases, it was noticed that X86 codegen does not generate unpck{h|l}ps instructions (for x86-64 cpu target) in one of them -

Case 1 : Scalarized stores with no extracts for GEP index (https://godbolt.org/z/qcqr8f461)
Case 2 : Scalarized stores with extracts for GEP index (https://godbolt.org/z/58n95Tv5q) - no unpck{h|l}ps instructions

I would like to confirm if there are any issues with the LLVM-IR for Case 2 that prohibits X86 codegen from generating optimal ASM like Case 1.

[karthik-senthil]

@RKSimon we're actually observing that this issue happens on AVX2 target as well, specifically after LLVM 21.1.0 release. Please check here - https://godbolt.org/z/ojGoPa86K . If we change llc version to 20.1.0 then vunpck{h|l}ps instructions are generated.