[llvm-branch-commits] [llvm] release/18.x: [AArch64][SelectionDAG] Mask for SUBS with multiple users cannot be elided (#90911) (PR #91151)

[via llvm-branch-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-aarch64

Author: AtariDreams (AtariDreams)

<details>
<summary>Changes</summary>

In DAGCombiner, the `performCONDCombine` function attempts to remove AND instructions in front of SUBS (cmp) instructions for which the AND is transparent. The rules for that are correct, but it fails to take into account the case where the SUBS instruction has multiple users with different condition codes for comparison and simply removes the AND for all of them. This causes a miscompilation in the attached test case.

(cherry picked from commit 72eaa0ed9934bfaa2449091bbc6e45648d1396d6)

---
Full diff: https://github.com/llvm/llvm-project/pull/91151.diff


2 Files Affected:

- (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+2-1) 
- (modified) llvm/test/CodeGen/AArch64/and-mask-removal.ll (+22) 


``````````diff
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 95d8ab95b2c097..bcfd0253e73c88 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -22122,7 +22122,8 @@ SDValue performCONDCombine(SDNode *N,
   SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
   unsigned CondOpcode = SubsNode->getOpcode();
 
-  if (CondOpcode != AArch64ISD::SUBS || SubsNode->hasAnyUseOfValue(0))
+  if (CondOpcode != AArch64ISD::SUBS || SubsNode->hasAnyUseOfValue(0) ||
+      !SubsNode->hasOneUse())
     return SDValue();
 
   // There is a SUBS feeding this condition. Is it fed by a mask we can
diff --git a/llvm/test/CodeGen/AArch64/and-mask-removal.ll b/llvm/test/CodeGen/AArch64/and-mask-removal.ll
index 17ff0159701689..a31355549ba87c 100644
--- a/llvm/test/CodeGen/AArch64/and-mask-removal.ll
+++ b/llvm/test/CodeGen/AArch64/and-mask-removal.ll
@@ -526,4 +526,26 @@ define i64 @pr58109b(i8 signext %0, i64 %a, i64 %b) {
   ret i64 %4
 }
 
+define i64 @test_2_selects(i8 zeroext %a) {
+; CHECK-LABEL: test_2_selects:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    add w9, w0, #24
+; CHECK-NEXT:    mov w8, #131
+; CHECK-NEXT:    and w9, w9, #0xff
+; CHECK-NEXT:    cmp w9, #81
+; CHECK-NEXT:    mov w9, #57
+; CHECK-NEXT:    csel x8, x8, xzr, lo
+; CHECK-NEXT:    csel x9, xzr, x9, eq
+; CHECK-NEXT:    add x0, x8, x9
+; CHECK-NEXT:    ret
+  %1 = add i8 %a, 24
+  %2 = zext i8 %1 to i64
+  %3 = icmp ult i8 %1, 81
+  %4 = select i1 %3, i64 131, i64 0
+  %5 = icmp eq i8 %1, 81
+  %6 = select i1 %5, i64 0, i64 57
+  %7 = add i64 %4, %6
+  ret i64 %7
+}
+
 declare i8 @llvm.usub.sat.i8(i8, i8) #0

``````````

</details>


https://github.com/llvm/llvm-project/pull/91151