[llvm] [InstCombine] Drop UB-implying attrs/metadata after speculating an instruction (PR #85542)

via llvm-commits llvm-commits at lists.llvm.org
Sat Mar 16 09:52:01 PDT 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

<details>
<summary>Changes</summary>

When speculating an instruction in `InstCombinerImpl::FoldOpIntoSelect`, the call may result in undefined behavior. This patch drops all UB-implying attrs/metadata to fix this.

Fixes #<!-- -->85536.


---
Full diff: https://github.com/llvm/llvm-project/pull/85542.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/InstCombine/InstructionCombining.cpp (+7) 
- (added) llvm/test/Transforms/InstCombine/pr85536.ll (+44) 


``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index edb046defbc1ca..3376653f48a5ca 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -1650,6 +1650,13 @@ static Value *foldOperationIntoSelectOperand(Instruction &I, SelectInst *SI,
                                              Value *NewOp, InstCombiner &IC) {
   Instruction *Clone = I.clone();
   Clone->replaceUsesOfWith(SI, NewOp);
+
+  // If we speculated an instruction, we need to drop any metadata that may
+  // result in undefined behavior, as the metadata might have been valid
+  // only given the branch precondition.
+  // Similarly strip attributes on call parameters that may cause UB in
+  // location the call is moved to.
+  Clone->dropUBImplyingAttrsAndMetadata();
   IC.InsertNewInstBefore(Clone, SI->getIterator());
   return Clone;
 }
diff --git a/llvm/test/Transforms/InstCombine/pr85536.ll b/llvm/test/Transforms/InstCombine/pr85536.ll
new file mode 100644
index 00000000000000..5e2bab984f9f2b
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/pr85536.ll
@@ -0,0 +1,44 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S -passes=instcombine < %s | FileCheck %s
+
+define i8 @test_drop_noundef(i1 %cond, i8 %val) {
+; CHECK-LABEL: define i8 @test_drop_noundef(
+; CHECK-SAME: i1 [[COND:%.*]], i8 [[VAL:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i8 @llvm.smin.i8(i8 [[VAL]], i8 0)
+; CHECK-NEXT:    [[RET:%.*]] = select i1 [[COND]], i8 -1, i8 [[TMP0]]
+; CHECK-NEXT:    ret i8 [[RET]]
+;
+entry:
+  %sel = select i1 %cond, i8 -1, i8 %val
+  %ret = call noundef i8 @llvm.smin.i8(i8 %sel, i8 0)
+  ret i8 %ret
+}
+
+define i1 @pr85536(i32 %a) {
+; CHECK-LABEL: define i1 @pr85536(
+; CHECK-SAME: i32 [[A:%.*]]) {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ult i32 [[A]], 31
+; CHECK-NEXT:    [[SHL1:%.*]] = shl nsw i32 -1, [[A]]
+; CHECK-NEXT:    [[ZEXT:%.*]] = zext i32 [[SHL1]] to i64
+; CHECK-NEXT:    [[SHL2:%.*]] = shl i64 [[ZEXT]], 48
+; CHECK-NEXT:    [[SHR:%.*]] = ashr exact i64 [[SHL2]], 48
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.smin.i64(i64 [[SHR]], i64 0)
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 65535
+; CHECK-NEXT:    [[RET1:%.*]] = icmp eq i64 [[TMP1]], 0
+; CHECK-NEXT:    [[RET:%.*]] = select i1 [[CMP1]], i1 [[RET1]], i1 false
+; CHECK-NEXT:    ret i1 [[RET]]
+;
+entry:
+  %cmp1 = icmp ugt i32 %a, 30
+  %shl1 = shl nsw i32 -1, %a
+  %zext = zext i32 %shl1 to i64
+  %shl2 = shl i64 %zext, 48
+  %shr = ashr exact i64 %shl2, 48
+  %sel = select i1 %cmp1, i64 -1, i64 %shr
+  %smin = call noundef i64 @llvm.smin.i64(i64 %sel, i64 0)
+  %masked = and i64 %smin, 65535
+  %ret = icmp eq i64 %masked, 0
+  ret i1 %ret
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/85542


More information about the llvm-commits mailing list