[Regression] LLVM18 and 19 performance regression when using an AVX-512 path

[Yahfz]

When testing for a performance regression we figured out that the performance regression is only there for AVX-512 paths.

The performance got worse in LLVM18 and then worse again in LLVM19 when using znver4, znver5 and tigerlake targets (AVX-512 paths)

But the opposite happened when using AVX2 paths, LLVM18 and 19 are both faster than LLVM16 when using the skylake target (AVX2 path)

AVX512 PATH (ZNVER4 / TIGERLAKE TARGET - 9800X3D CPU):

LLVM16 - 72FPS AVG
LLVM19 - 67FPS AVG

AVX2 PATH (SKYLAKE TARGET - 9800X3D CPU):

LLVM16 - 56FPS AVG
LLVM19 - 59FPS AVG

[Whatcookie]

I've opened one issue at upstream LLVM: llvm/llvm-project#128237, but there should be more regressions to find.