[LV] #149042 caused performance regressions on avx512

[Andarwinux]

#149042 seems to cause a lot of regressions on avx512, and since my loads are based on LTO and PGO, I'm share the samples as lld reproduce.

149042.zip

This example is mujs, running v8-v7 benchmark.
before #149042

./mujs.exe .\run.js
Richards: 559
DeltaBlue: 738
Crypto: 514
RayTrace: 809
EarleyBoyer: 1008
RegExp: 569
Splay: 1405
NavierStokes: 1315
----
Score (version 7): 808

after #149042

./mujs.exe .\run.js
Richards: 530
DeltaBlue: 698
Crypto: 365
RayTrace: 688
EarleyBoyer: 938
RegExp: 565
Splay: 1407
NavierStokes: 1295
----
Score (version 7): 740

The main difference seems to be jsR_run, which seems to be partially vectorized after #149042, but gets slower on tigerlake.

[fhahn]

@Andarwinux I'm not sure how to get the IR before vectorization for the interesting function. It would be great if you could provide any pointers,

[Andarwinux]

@Andarwinux I'm not sure how to get the IR before vectorization for the interesting function. It would be great if you could provide any pointers,

You can get the lld reproduce from my 149042.zip, where one.c.obj in “build\packages\mujs-prefix\src\mujs-build\build\libmujs.a” includes the LTO bitcode.

Below are all the IRs I saved via -save-temps:

after_149042.zip
before_149042.zip