Skip to content

[X86] lowerV64I8Shuffle - prefer VPERMV3 byte shuffles to OR(PSHUFB,PSHUFB) on VBMI targets #182852

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
RKSimon merged 1 commit into from
Feb 23, 2026
Merged

[X86] lowerV64I8Shuffle - prefer VPERMV3 byte shuffles to OR(PSHUFB,PSHUFB) on VBMI targets #182852

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions llvm/lib/Target/X86/X86ISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18267,6 +18267,12 @@ static SDValue lowerV64I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
Mask, Subtarget, DAG))
return V;

// VBMI can use VPERMV/VPERMV3 byte shuffles more efficiently than
// OR(PSHUFB,PSHUFB).
if (Subtarget.hasVBMI())
return lowerShuffleWithPERMV(DL, MVT::v64i8, Mask, V1, V2, Subtarget,
DAG);

// If we can't directly blend but can use PSHUFB, that will be better as it
// can both shuffle and set up the inefficient blend.
bool V1InUse, V2InUse;
Expand Down
43 changes: 19 additions & 24 deletions llvm/test/CodeGen/X86/vector-shuffle-512-v64.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1951,33 +1951,33 @@ define <512 x i8> @PR153457(<512 x i8> %a0, <512 x i8> %a1) nounwind {
; AVX512VBMI-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,1,2,3,4,5,6,66,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,67,0,1,2,3,4,5,6,66,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,67]
; AVX512VBMI-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
; AVX512VBMI-NEXT: vpermi2b %zmm0, %zmm2, %zmm3
; AVX512VBMI-NEXT: vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[1,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512VBMI-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14],zero,zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63]
; AVX512VBMI-NEXT: vporq %zmm2, %zmm1, %zmm1
; AVX512VBMI-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,71]
; AVX512VBMI-NEXT: vpermi2b %zmm0, %zmm6, %zmm2
; AVX512VBMI-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,65,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,65,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63]
; AVX512VBMI-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]
; AVX512VBMI-NEXT: vpermi2b %zmm0, %zmm1, %zmm2
; AVX512VBMI-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,71]
; AVX512VBMI-NEXT: vpermi2b %zmm0, %zmm6, %zmm1
; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,0,64,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,58,59,60,61,62,63]
; AVX512VBMI-NEXT: vpermi2b %zmm7, %zmm0, %zmm6
; AVX512VBMI-NEXT: vpmovsxbw {{.*#+}} zmm0 = [0,1,2,3,33,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,20,21,22,23,24,25,26,27,28,29,30,31]
; AVX512VBMI-NEXT: vpermi2w %zmm7, %zmm3, %zmm0
; AVX512VBMI-NEXT: vinserti32x4 $1, %xmm7, %zmm2, %zmm0
; AVX512VBMI-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,17,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63]
; AVX512VBMI-NEXT: vpmovsxbw {{.*#+}} zmm2 = [0,1,2,3,33,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,20,21,22,23,24,25,26,27,28,29,30,31]
; AVX512VBMI-NEXT: vpermi2w %zmm7, %zmm3, %zmm2
; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm3 = [67,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,68,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vpermi2b %zmm7, %zmm4, %zmm3
; AVX512VBMI-NEXT: vinserti32x4 $3, %xmm7, %zmm5, %zmm4
; AVX512VBMI-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,53,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [16,17,18,19,35,0,0,0,8,9,10,11,12,13,14,15,16,17,18,19,35,0,0,0,8,9,10,11,12,13,14,15]
; AVX512VBMI-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
; AVX512VBMI-NEXT: vpermi2w %zmm7, %zmm8, %zmm5
; AVX512VBMI-NEXT: vinserti64x4 $1, %ymm7, %zmm2, %zmm2
; AVX512VBMI-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,39,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm8 = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,65,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63]
; AVX512VBMI-NEXT: vpermi2b %zmm7, %zmm1, %zmm8
; AVX512VBMI-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
; AVX512VBMI-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,39,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vmovdqa64 %zmm5, 320(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm4, 256(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm3, 192(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm0, 128(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm8, 64(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm2, 128(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm0, 64(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm6, (%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm2, 384(%rdi)
; AVX512VBMI-NEXT: vmovdqa64 %zmm1, 384(%rdi)
; AVX512VBMI-NEXT: movq %rbp, %rsp
; AVX512VBMI-NEXT: popq %rbp
; AVX512VBMI-NEXT: vzeroupper
Expand Down Expand Up @@ -2043,16 +2043,11 @@ define <64 x i8> @narrow_u32x16x4_to_u8x64(<64 x i8> %x0, <64 x i8> %x1, <64 x i
;
; AVX512VBMI-LABEL: narrow_u32x16x4_to_u8x64:
; AVX512VBMI: # %bb.0:
; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm4 = [128,128,128,128,0,4,8,12,u,u,u,u,u,u,u,u,128,128,128,128,16,20,24,28,u,u,u,u,u,u,u,u,128,128,128,128,32,36,40,44,u,u,u,u,u,u,u,u,128,128,128,128,48,52,56,60,u,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vpshufb %zmm4, %zmm1, %zmm1
; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,4,8,12,128,128,128,128,u,u,u,u,u,u,u,u,16,20,24,28,128,128,128,128,u,u,u,u,u,u,u,u,32,36,40,44,128,128,128,128,u,u,u,u,u,u,u,u,48,52,56,60,128,128,128,128,u,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vpshufb %zmm5, %zmm0, %zmm0
; AVX512VBMI-NEXT: vporq %zmm1, %zmm0, %zmm1
; AVX512VBMI-NEXT: vpshufb %zmm4, %zmm3, %zmm0
; AVX512VBMI-NEXT: vpshufb %zmm5, %zmm2, %zmm2
; AVX512VBMI-NEXT: vporq %zmm0, %zmm2, %zmm2
; AVX512VBMI-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,4,8,12,1,5,9,13,16,20,24,28,17,21,25,29]
; AVX512VBMI-NEXT: vpermi2d %zmm2, %zmm1, %zmm0
; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,4,8,12,64,68,72,76,u,u,u,u,u,u,u,u,16,20,24,28,80,84,88,92,u,u,u,u,u,u,u,u,32,36,40,44,96,100,104,108,u,u,u,u,u,u,u,u,48,52,56,60,112,116,120,124,u,u,u,u,u,u,u,u]
; AVX512VBMI-NEXT: vpermt2b %zmm1, %zmm4, %zmm0
; AVX512VBMI-NEXT: vpermt2b %zmm3, %zmm4, %zmm2
; AVX512VBMI-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,4,8,12,1,5,9,13,16,20,24,28,17,21,25,29]
; AVX512VBMI-NEXT: vpermt2d %zmm2, %zmm1, %zmm0
; AVX512VBMI-NEXT: retq
%lo = shufflevector <64 x i8> %x0, <64 x i8> %x1, <64 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 32, i32 36, i32 40, i32 44, i32 48, i32 52, i32 56, i32 60, i32 64, i32 68, i32 72, i32 76, i32 80, i32 84, i32 88, i32 92, i32 96, i32 100, i32 104, i32 108, i32 112, i32 116, i32 120, i32 124, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
%hi = shufflevector <64 x i8> %x2, <64 x i8> %x3, <64 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 32, i32 36, i32 40, i32 44, i32 48, i32 52, i32 56, i32 60, i32 64, i32 68, i32 72, i32 76, i32 80, i32 84, i32 88, i32 92, i32 96, i32 100, i32 104, i32 108, i32 112, i32 116, i32 120, i32 124>
Expand Down
Loading