[llvm-branch-commits] [llvm] [X86] Don't let BUILD_VECTOR pick a value for multi-use freeze(undef) (PR #240471)

Mitch Briles via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Tue Nov 10 01:54:03 PST 2026


https://github.com/MitchBriles updated https://github.com/llvm/llvm-project/pull/240471

>From ac1b89c72662eb2bde51cbc68bdd97f3701385a3 Mon Sep 17 00:00:00 2001
From: Mitch <mitchbriles at gmail.com>
Date: Fri, 6 Nov 2026 21:14:46 -0700
Subject: [PATCH 1/2] Precommit X86 BUILD_VECTOR freeze tests

Assisted-By: Claude Opus 5.5 <noreply at anthropic.com>
---
 llvm/test/CodeGen/X86/freeze-vector.ll | 74 ++++++++++++++++++++++++++
 1 file changed, 74 insertions(+)

diff --git a/llvm/test/CodeGen/X86/freeze-vector.ll b/llvm/test/CodeGen/X86/freeze-vector.ll
index a97161a657e86c..f491a10f394758 100644
--- a/llvm/test/CodeGen/X86/freeze-vector.ll
+++ b/llvm/test/CodeGen/X86/freeze-vector.ll
@@ -1120,4 +1120,78 @@ define void @freeze_dynamic_insertelement_chain_wide_vector(ptr %vp, i32 %x, i32
   ret void
 }
 
+; The stored value must be returned in element 1.
+define <4 x i32> @freeze_poison_multi_use_with_zeros(ptr %p) nounwind {
+; X86-LABEL: freeze_poison_multi_use_with_zeros:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl %eax, (%eax)
+; X86-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; X86-NEXT:    retl
+;
+; X64-LABEL: freeze_poison_multi_use_with_zeros:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %eax, (%rdi)
+; X64-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; X64-NEXT:    retq
+  %f = freeze i32 poison
+  store i32 %f, ptr %p
+  %v = insertelement <4 x i32> zeroinitializer, i32 %f, i32 1
+  ret <4 x i32> %v
+}
+
+; The stored value must be returned in elements 0, 2 and 3.
+define <4 x i32> @freeze_poison_multi_use_splat(ptr %p, i32 %x) nounwind {
+; X86-LABEL: freeze_poison_multi_use_splat:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl %eax, (%eax)
+; X86-NEXT:    vbroadcastss {{[0-9]+}}(%esp), %xmm0
+; X86-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; X86-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
+; X86-NEXT:    retl
+;
+; X64-LABEL: freeze_poison_multi_use_splat:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %eax, (%rdi)
+; X64-NEXT:    vmovd %esi, %xmm0
+; X64-NEXT:    vpbroadcastd %xmm0, %xmm0
+; X64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; X64-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
+; X64-NEXT:    retq
+  %f = freeze i32 poison
+  store i32 %f, ptr %p
+  %v0 = insertelement <4 x i32> poison, i32 %f, i32 0
+  %v1 = insertelement <4 x i32> %v0, i32 %x, i32 1
+  %v2 = insertelement <4 x i32> %v1, i32 %f, i32 2
+  %v3 = insertelement <4 x i32> %v2, i32 %f, i32 3
+  ret <4 x i32> %v3
+}
+
+; The stored value must be returned in the upper elements.
+define <8 x i32> @freeze_poison_multi_use_upper(ptr %p, i32 %x) nounwind {
+; X86-LABEL: freeze_poison_multi_use_upper:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl %eax, (%eax)
+; X86-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; X86-NEXT:    retl
+;
+; X64-LABEL: freeze_poison_multi_use_upper:
+; X64:       # %bb.0:
+; X64-NEXT:    movl %eax, (%rdi)
+; X64-NEXT:    vmovd %esi, %xmm0
+; X64-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; X64-NEXT:    retq
+  %f = freeze i32 poison
+  store i32 %f, ptr %p
+  %v0 = insertelement <8 x i32> poison, i32 %x, i32 0
+  %v1 = insertelement <8 x i32> %v0, i32 %f, i32 4
+  %v2 = insertelement <8 x i32> %v1, i32 %f, i32 5
+  %v3 = insertelement <8 x i32> %v2, i32 %f, i32 6
+  %v4 = insertelement <8 x i32> %v3, i32 %f, i32 7
+  ret <8 x i32> %v4
+}
+
 declare i32 @llvm.ctlz.i32(i32, i1)

>From 523726084aa8d449c2afd87e4c8204e81fe6c3f1 Mon Sep 17 00:00:00 2001
From: Mitch <mitchbriles at gmail.com>
Date: Fri, 6 Nov 2026 21:14:47 -0700
Subject: [PATCH 2/2] [X86] Don't let BUILD_VECTOR pick a value for multi-use
 freeze(undef)

LowerBUILD_VECTOR treats freeze(undef) operands as free to choose. It
folds them to zero, blends in a new freeze, or zeroes them in the upper
half. If the freeze has other users, they may see a different value
than the BUILD_VECTOR does. Only treat a freeze(undef) as free when the
BUILD_VECTOR is its only user, and only zero the upper half when each
one has a single use, so that a freeze used by several lanes keeps one
value.

Assisted-By: Claude Opus 5.5 <noreply at anthropic.com>
---
 llvm/lib/Target/X86/X86ISelLowering.cpp |  7 ++-
 llvm/test/CodeGen/X86/freeze-vector.ll  | 70 +++++++++++++------------
 2 files changed, 42 insertions(+), 35 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 4d081b7ecbf43d..fb6f13afd12eba 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -9512,7 +9512,8 @@ X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
       UndefMask.setBit(i);
       continue;
     }
-    if (ISD::isFreezeUndef(Elt.getNode())) {
+    // A freeze(undef) with other users must keep its value.
+    if (ISD::isFreezeUndef(Elt.getNode()) && Op->isOnlyUserOf(Elt.getNode())) {
       OneUseFrozenUndefs = OneUseFrozenUndefs && Elt->hasOneUse();
       FrozenUndefMask.setBit(i);
       continue;
@@ -9575,7 +9576,9 @@ X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
   if ((VT.is256BitVector() || VT.is512BitVector()) &&
       !isFoldableUseOfShuffle(BV)) {
     unsigned UpperElems = NumElems / 2;
-    APInt UndefOrZeroMask = FrozenUndefMask | UndefMask | ZeroMask;
+    APInt UndefOrZeroMask = UndefMask | ZeroMask;
+    if (OneUseFrozenUndefs)
+      UndefOrZeroMask |= FrozenUndefMask;
     unsigned NumUpperUndefsOrZeros = UndefOrZeroMask.countl_one();
     if (NumUpperUndefsOrZeros >= UpperElems) {
       if (VT.is512BitVector() &&
diff --git a/llvm/test/CodeGen/X86/freeze-vector.ll b/llvm/test/CodeGen/X86/freeze-vector.ll
index f491a10f394758..b97d75802d050e 100644
--- a/llvm/test/CodeGen/X86/freeze-vector.ll
+++ b/llvm/test/CodeGen/X86/freeze-vector.ll
@@ -389,20 +389,17 @@ define void @freeze_two_frozen_buildvectors(ptr %origin0, ptr %origin1, ptr %dst
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
-; X86-NEXT:    movl (%esi), %esi
-; X86-NEXT:    andl $15, %esi
 ; X86-NEXT:    movl (%edx), %edx
-; X86-NEXT:    andl $15, %edx
-; X86-NEXT:    vmovd %esi, %xmm0
-; X86-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
-; X86-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X86-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
+; X86-NEXT:    vmovd %eax, %xmm0
+; X86-NEXT:    vpinsrd $1, (%esi), %xmm0, %xmm1
+; X86-NEXT:    vpinsrd $2, %eax, %xmm1, %xmm1
+; X86-NEXT:    vpinsrd $3, %eax, %xmm1, %xmm1
 ; X86-NEXT:    vbroadcastss {{.*#+}} xmm2 = [7,7,7,7]
-; X86-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; X86-NEXT:    vmovdqa %xmm0, (%ecx)
-; X86-NEXT:    vmovd %edx, %xmm0
-; X86-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
-; X86-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5],xmm1[6,7]
+; X86-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; X86-NEXT:    vmovdqa %xmm1, (%ecx)
+; X86-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
+; X86-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
+; X86-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
 ; X86-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; X86-NEXT:    vmovdqa %xmm0, (%eax)
 ; X86-NEXT:    popl %esi
@@ -411,16 +408,16 @@ define void @freeze_two_frozen_buildvectors(ptr %origin0, ptr %origin1, ptr %dst
 ; X64-LABEL: freeze_two_frozen_buildvectors:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl (%rsi), %eax
-; X64-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X64-NEXT:    vpbroadcastd %xmm0, %xmm0
-; X64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X64-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
-; X64-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [7,7,7,7]
-; X64-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; X64-NEXT:    vmovdqa %xmm0, (%rdx)
 ; X64-NEXT:    vmovd %eax, %xmm0
-; X64-NEXT:    vpbroadcastd %xmm0, %xmm0
-; X64-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3]
+; X64-NEXT:    vpinsrd $1, (%rdi), %xmm0, %xmm1
+; X64-NEXT:    vpinsrd $2, %eax, %xmm1, %xmm1
+; X64-NEXT:    vpinsrd $3, %eax, %xmm1, %xmm1
+; X64-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [7,7,7,7]
+; X64-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; X64-NEXT:    vmovdqa %xmm1, (%rdx)
+; X64-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
+; X64-NEXT:    vpinsrd $2, %eax, %xmm0, %xmm0
+; X64-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
 ; X64-NEXT:    vpand %xmm2, %xmm0, %xmm0
 ; X64-NEXT:    vmovdqa %xmm0, (%rcx)
 ; X64-NEXT:    retq
@@ -1126,13 +1123,15 @@ define <4 x i32> @freeze_poison_multi_use_with_zeros(ptr %p) nounwind {
 ; X86:       # %bb.0:
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movl %eax, (%eax)
-; X86-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; X86-NEXT:    vmovd %eax, %xmm0
+; X86-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,0,1,1]
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: freeze_poison_multi_use_with_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %eax, (%rdi)
-; X64-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; X64-NEXT:    vmovd %eax, %xmm0
+; X64-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,0,1,1]
 ; X64-NEXT:    retq
   %f = freeze i32 poison
   store i32 %f, ptr %p
@@ -1146,18 +1145,19 @@ define <4 x i32> @freeze_poison_multi_use_splat(ptr %p, i32 %x) nounwind {
 ; X86:       # %bb.0:
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movl %eax, (%eax)
-; X86-NEXT:    vbroadcastss {{[0-9]+}}(%esp), %xmm0
-; X86-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; X86-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
+; X86-NEXT:    vmovd %eax, %xmm0
+; X86-NEXT:    vpinsrd $1, {{[0-9]+}}(%esp), %xmm0, %xmm0
+; X86-NEXT:    vpinsrd $2, %eax, %xmm0, %xmm0
+; X86-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: freeze_poison_multi_use_splat:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %eax, (%rdi)
-; X64-NEXT:    vmovd %esi, %xmm0
-; X64-NEXT:    vpbroadcastd %xmm0, %xmm0
-; X64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; X64-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
+; X64-NEXT:    vmovd %eax, %xmm0
+; X64-NEXT:    vpinsrd $1, %esi, %xmm0, %xmm0
+; X64-NEXT:    vpinsrd $2, %eax, %xmm0, %xmm0
+; X64-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
 ; X64-NEXT:    retq
   %f = freeze i32 poison
   store i32 %f, ptr %p
@@ -1174,15 +1174,19 @@ define <8 x i32> @freeze_poison_multi_use_upper(ptr %p, i32 %x) nounwind {
 ; X86:       # %bb.0:
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movl %eax, (%eax)
-; X86-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X86-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; X86-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-NEXT:    vmovd %eax, %xmm1
+; X86-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
+; X86-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: freeze_poison_multi_use_upper:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %eax, (%rdi)
 ; X64-NEXT:    vmovd %esi, %xmm0
-; X64-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; X64-NEXT:    vmovd %eax, %xmm1
+; X64-NEXT:    vpbroadcastd %xmm1, %xmm1
+; X64-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
 ; X64-NEXT:    retq
   %f = freeze i32 poison
   store i32 %f, ptr %p



More information about the llvm-branch-commits mailing list