[llvm] [X86] isSplatValueForTargetNode - test source value for vector uniform shift ops (PR #186619)
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Sat Mar 14 13:12:36 PDT 2026
https://github.com/RKSimon created https://github.com/llvm/llvm-project/pull/186619
For old SSE style vector shifts, we just need to check the shifted value is a splat as the shift amount is uniform
Avoids an unnecessary variable shuffle in i512 ashr expansion
>From ec0f8251c981a66e3841b599e6a42f0ba806e216 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Sat, 14 Mar 2026 20:11:22 +0000
Subject: [PATCH] [X86] isSplatValueForTargetNode - test source value for
vector uniform shift ops
For old SSE style vector shifts, we just need to check the shifted value is splat as the shift amount is uniform
Avoids an unnecessary variable shuffle in i512 ashr expansion
---
llvm/lib/Target/X86/X86ISelLowering.cpp | 8 +++
llvm/test/CodeGen/X86/shift-i512.ll | 70 ++++++++++++-------------
2 files changed, 42 insertions(+), 36 deletions(-)
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ddba1c7dbdf00..4fbbf63c39065 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -45827,6 +45827,14 @@ bool X86TargetLowering::isSplatValueForTargetNode(SDValue Op,
case X86ISD::VBROADCAST_LOAD:
UndefElts = APInt::getZero(NumElts);
return true;
+ case X86ISD::VSHL:
+ case X86ISD::VSRA:
+ case X86ISD::VSRL:
+ case X86ISD::VSHLI:
+ case X86ISD::VSRAI:
+ case X86ISD::VSRLI:
+ return DAG.isSplatValue(Op.getOperand(0), DemandedElts, UndefElts,
+ Depth + 1);
}
return TargetLowering::isSplatValueForTargetNode(Op, DemandedElts, UndefElts,
diff --git a/llvm/test/CodeGen/X86/shift-i512.ll b/llvm/test/CodeGen/X86/shift-i512.ll
index 01cd11b9e712c..fa854663d38f9 100644
--- a/llvm/test/CodeGen/X86/shift-i512.ll
+++ b/llvm/test/CodeGen/X86/shift-i512.ll
@@ -604,8 +604,8 @@ define i512 @ashr_i512(i512 %a0, i512 %a1) nounwind {
; AVX512F-NEXT: andl $56, %eax
; AVX512F-NEXT: vmovdqu64 -128(%rsp,%rax), %zmm1
; AVX512F-NEXT: vpsrlq %xmm0, %zmm1, %zmm2
-; AVX512F-NEXT: vpsraq $63, -72(%rsp,%rax){1to8}, %zmm3
-; AVX512F-NEXT: valignq {{.*#+}} zmm3 = zmm3[7,0,1,2,3,4,5,6]
+; AVX512F-NEXT: vpbroadcastq -72(%rsp,%rax), %xmm3
+; AVX512F-NEXT: vpsraq $63, %zmm3, %zmm3
; AVX512F-NEXT: valignq {{.*#+}} zmm1 = zmm1[1,2,3,4,5,6,7],zmm3[0]
; AVX512F-NEXT: vpaddq %zmm1, %zmm1, %zmm1
; AVX512F-NEXT: vpandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
@@ -639,20 +639,19 @@ define i512 @ashr_i512(i512 %a0, i512 %a1) nounwind {
; AVX512VL-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
; AVX512VL-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
; AVX512VL-NEXT: movl %eax, %ecx
-; AVX512VL-NEXT: shrl $3, %ecx
-; AVX512VL-NEXT: andl $56, %ecx
-; AVX512VL-NEXT: vpsraq $63, -72(%rsp,%rcx){1to8}, %zmm0
-; AVX512VL-NEXT: vmovdqu64 -128(%rsp,%rcx), %zmm1
-; AVX512VL-NEXT: vmovdqa64 {{.*#+}} zmm2 = [1,2,3,4,5,6,7,15]
-; AVX512VL-NEXT: vpermi2q %zmm0, %zmm1, %zmm2
-; AVX512VL-NEXT: vpaddq %zmm2, %zmm2, %zmm0
-; AVX512VL-NEXT: andl $63, %eax
-; AVX512VL-NEXT: vpbroadcastq %rax, %xmm2
-; AVX512VL-NEXT: vpandnq {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to2}, %xmm2, %xmm3
-; AVX512VL-NEXT: vpsllq %xmm3, %zmm0, %zmm0
-; AVX512VL-NEXT: vpsrlq %xmm2, %zmm1, %zmm1
+; AVX512VL-NEXT: andl $63, %ecx
+; AVX512VL-NEXT: vpbroadcastq %rcx, %xmm0
+; AVX512VL-NEXT: shrl $3, %eax
+; AVX512VL-NEXT: andl $56, %eax
+; AVX512VL-NEXT: vmovdqu64 -128(%rsp,%rax), %zmm1
+; AVX512VL-NEXT: vpsrlq %xmm0, %zmm1, %zmm2
+; AVX512VL-NEXT: vpsraq $63, -72(%rsp,%rax){1to2}, %xmm3
+; AVX512VL-NEXT: valignq {{.*#+}} zmm1 = zmm1[1,2,3,4,5,6,7],zmm3[0]
+; AVX512VL-NEXT: vpaddq %zmm1, %zmm1, %zmm1
+; AVX512VL-NEXT: vpandnq {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to2}, %xmm0, %xmm0
+; AVX512VL-NEXT: vpsllq %xmm0, %zmm1, %zmm0
; AVX512VL-NEXT: movq %rdi, %rax
-; AVX512VL-NEXT: vporq %zmm1, %zmm0, %zmm0
+; AVX512VL-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512VL-NEXT: vextracti64x4 $1, %zmm0, 32(%rdi)
; AVX512VL-NEXT: vmovdqu %ymm0, (%rdi)
; AVX512VL-NEXT: popq %rcx
@@ -663,34 +662,33 @@ define i512 @ashr_i512(i512 %a0, i512 %a1) nounwind {
; AVX512VBMI: # %bb.0:
; AVX512VBMI-NEXT: pushq %rax
; AVX512VBMI-NEXT: movq %rdi, %rax
-; AVX512VBMI-NEXT: vmovaps {{[0-9]+}}(%rsp), %xmm0
; AVX512VBMI-NEXT: movq {{[0-9]+}}(%rsp), %rdi
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: vmovups %xmm0, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: vmovaps {{[0-9]+}}(%rsp), %xmm0
; AVX512VBMI-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: vmovups %xmm0, -{{[0-9]+}}(%rsp)
; AVX512VBMI-NEXT: movq %r9, -{{[0-9]+}}(%rsp)
; AVX512VBMI-NEXT: movq %r8, -{{[0-9]+}}(%rsp)
; AVX512VBMI-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)
; AVX512VBMI-NEXT: movq %rdx, -{{[0-9]+}}(%rsp)
; AVX512VBMI-NEXT: movq %rsi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: sarq $63, %rdi
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)
-; AVX512VBMI-NEXT: vpbroadcastq %r10, %zmm0
-; AVX512VBMI-NEXT: # kill: def $r10d killed $r10d killed $r10 def $r10
-; AVX512VBMI-NEXT: shrl $3, %r10d
-; AVX512VBMI-NEXT: andl $56, %r10d
-; AVX512VBMI-NEXT: vpsraq $63, -72(%rsp,%r10){1to8}, %zmm1
-; AVX512VBMI-NEXT: vmovdqu64 -128(%rsp,%r10), %zmm2
-; AVX512VBMI-NEXT: vmovdqa64 {{.*#+}} zmm3 = [1,2,3,4,5,6,7,15]
-; AVX512VBMI-NEXT: vpermi2q %zmm1, %zmm2, %zmm3
-; AVX512VBMI-NEXT: vpshrdvq %zmm0, %zmm3, %zmm2
+; AVX512VBMI-NEXT: sarq $63, %r10
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: movq %r10, -{{[0-9]+}}(%rsp)
+; AVX512VBMI-NEXT: vpbroadcastq %rdi, %zmm0
+; AVX512VBMI-NEXT: # kill: def $edi killed $edi killed $rdi def $rdi
+; AVX512VBMI-NEXT: shrl $3, %edi
+; AVX512VBMI-NEXT: andl $56, %edi
+; AVX512VBMI-NEXT: vpsraq $63, -72(%rsp,%rdi){1to2}, %xmm1
+; AVX512VBMI-NEXT: vmovdqu64 -128(%rsp,%rdi), %zmm2
+; AVX512VBMI-NEXT: valignq {{.*#+}} zmm1 = zmm2[1,2,3,4,5,6,7],zmm1[0]
+; AVX512VBMI-NEXT: vpshrdvq %zmm0, %zmm1, %zmm2
; AVX512VBMI-NEXT: vextracti64x4 $1, %zmm2, 32(%rax)
; AVX512VBMI-NEXT: vmovdqu %ymm2, (%rax)
; AVX512VBMI-NEXT: popq %rcx
More information about the llvm-commits
mailing list