Skip to content

Commit 1a69056

Browse files
authored
Backport [DAGCombine] Fix multi-use miscompile in load combine (#81586) (#81633)
(cherry picked from commit 25b9ed6)
1 parent d01a4ab commit 1a69056

File tree

2 files changed

+33
-1
lines changed

2 files changed

+33
-1
lines changed

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9253,7 +9253,7 @@ SDValue DAGCombiner::MatchLoadCombine(SDNode *N) {
92539253

92549254
// Transfer chain users from old loads to the new load.
92559255
for (LoadSDNode *L : Loads)
9256-
DAG.ReplaceAllUsesOfValueWith(SDValue(L, 1), SDValue(NewLoad.getNode(), 1));
9256+
DAG.makeEquivalentMemoryOrdering(L, NewLoad);
92579257

92589258
if (!NeedsBswap)
92599259
return NewLoad;

llvm/test/CodeGen/X86/load-combine.ll

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1282,3 +1282,35 @@ define i32 @zext_load_i32_by_i8_bswap_shl_16(ptr %arg) {
12821282
%tmp8 = or i32 %tmp7, %tmp30
12831283
ret i32 %tmp8
12841284
}
1285+
1286+
define i32 @pr80911_vector_load_multiuse(ptr %ptr, ptr %clobber) nounwind {
1287+
; CHECK-LABEL: pr80911_vector_load_multiuse:
1288+
; CHECK: # %bb.0:
1289+
; CHECK-NEXT: pushl %esi
1290+
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx
1291+
; CHECK-NEXT: movl {{[0-9]+}}(%esp), %edx
1292+
; CHECK-NEXT: movl (%edx), %esi
1293+
; CHECK-NEXT: movzwl (%edx), %eax
1294+
; CHECK-NEXT: movl $0, (%ecx)
1295+
; CHECK-NEXT: movl %esi, (%edx)
1296+
; CHECK-NEXT: popl %esi
1297+
; CHECK-NEXT: retl
1298+
;
1299+
; CHECK64-LABEL: pr80911_vector_load_multiuse:
1300+
; CHECK64: # %bb.0:
1301+
; CHECK64-NEXT: movl (%rdi), %ecx
1302+
; CHECK64-NEXT: movzwl (%rdi), %eax
1303+
; CHECK64-NEXT: movl $0, (%rsi)
1304+
; CHECK64-NEXT: movl %ecx, (%rdi)
1305+
; CHECK64-NEXT: retq
1306+
%load = load <4 x i8>, ptr %ptr, align 16
1307+
store i32 0, ptr %clobber
1308+
store <4 x i8> %load, ptr %ptr, align 16
1309+
%e1 = extractelement <4 x i8> %load, i64 1
1310+
%e1.ext = zext i8 %e1 to i32
1311+
%e1.ext.shift = shl nuw nsw i32 %e1.ext, 8
1312+
%e0 = extractelement <4 x i8> %load, i64 0
1313+
%e0.ext = zext i8 %e0 to i32
1314+
%res = or i32 %e1.ext.shift, %e0.ext
1315+
ret i32 %res
1316+
}

0 commit comments

Comments
 (0)