JIT: Tune CSE/Constant propagation to fix this SIMD example

[EgorBo]

Found during review of #70654

Vector128<byte> Test(Vector128<byte> a, Vector128<byte> b) 
 => Vector128.GreaterThan(a, b);

Current codegen:

G_M23213_IG01:
 vzeroupper 
G_M23213_IG02:
 vmovupd xmm0, xmmword ptr [r8]
 vpsubb xmm0, xmm0, xmmword ptr [reloc @RWD00]
 vmovupd xmm1, xmmword ptr [r9]
 vpsubb xmm1, xmm1, xmmword ptr [reloc @RWD00]
 vpcmpgtb xmm0, xmm0, xmm1
 vmovupd xmmword ptr [rdx], xmm0
 mov rax, rdx
G_M23213_IG03:
 ret 
RWD00 	dq	8080808080808080h, 8080808080808080h
; Total bytes of code: 41

Here, before vpcmpgtb we subtract a constant vector from both xmm0 and xmm1 - CSE should be able to handle them and give constant propagation a hint that it should not propagate them back.

category:cq
theme:assertion-prop
skill-level:intermediate
cost:small
impact:medium

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Found during review of #70654

Vector128<byte> Test(Vector128<byte> a, Vector128<byte> b) 
 => Vector128.GreaterThan(a, b);

Current codegen:

G_M23213_IG01:
 vzeroupper 
G_M23213_IG02:
 vmovupd xmm0, xmmword ptr [r8]
 vpsubb xmm0, xmm0, xmmword ptr [reloc @RWD00]
 vmovupd xmm1, xmmword ptr [r9]
 vpsubb xmm1, xmm1, xmmword ptr [reloc @RWD00]
 vpcmpgtb xmm0, xmm0, xmm1
 vmovupd xmmword ptr [rdx], xmm0
 mov rax, rdx
G_M23213_IG03:
 ret 
RWD00 	dq	8080808080808080h, 8080808080808080h
; Total bytes of code: 41

Here, before vpcmpgtb we subtract a constant vector from both xmm0 and xmm1 - CSE should be able to handle them and give constant propagation a hint that it should not propagate them back.

Author:	EgorBo
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[jakobbotsch]

Current codegen is:

; Method Program:Test(System.Runtime.Intrinsics.Vector128`1[ubyte],System.Runtime.Intrinsics.Vector128`1[ubyte]):System.Runtime.Intrinsics.Vector128`1[ubyte]:this
G_M40440_IG01:
 vzeroupper 
						;; size=3 bbWeight=1 PerfScore 1.00

G_M40440_IG02:
 vmovupd xmm0, xmmword ptr [r8]
 vmovupd xmm1, xmmword ptr [reloc @RWD00]
 vpsubb xmm0, xmm0, xmm1
 vmovupd xmm2, xmmword ptr [r9]
 vpsubb xmm1, xmm2, xmm1
 vpcmpgtb xmm0, xmm0, xmm1
 vmovupd xmmword ptr [rdx], xmm0
 mov rax, rdx
						;; size=37 bbWeight=1 PerfScore 14.42

G_M40440_IG03:
 ret 

I assume this was fixed by one of your changes @EgorBo so this can be closed?