[llvm] [InstCombine] Drop Range attribute when simplifying 'fshl' based on demanded bits (PR #124429)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Fangrui Song (MaskRay)

<details>
<summary>Changes</summary>

When simplifying operands based on demanded bits, the return value range
of llvm.fshl might change. Keeping the Range attribute might cause
llvm.fshl to generate a poison and lead to miscompile (#<!-- -->124387). Drop the Range
attribute similar to `dropPosonGeneratingFlags` elsewhere.

---
Full diff: https://github.com/llvm/llvm-project/pull/124429.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp (+6-3) 
- (modified) llvm/test/Transforms/InstCombine/fsh.ll (+15) 


``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
index 943598a30f040f..1b89839a72b94b 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -1039,11 +1039,14 @@ Value *InstCombinerImpl::SimplifyDemandedUseBits(Instruction *I,
         APInt DemandedMaskLHS(DemandedMask.lshr(ShiftAmt));
         APInt DemandedMaskRHS(DemandedMask.shl(BitWidth - ShiftAmt));
         if (I->getOperand(0) != I->getOperand(1)) {
-          if (SimplifyDemandedBits(I, 0, DemandedMaskLHS, LHSKnown,
-                                   Depth + 1, Q) ||
+          if (SimplifyDemandedBits(I, 0, DemandedMaskLHS, LHSKnown, Depth + 1,
+                                   Q) ||
               SimplifyDemandedBits(I, 1, DemandedMaskRHS, RHSKnown, Depth + 1,
-                                   Q))
+                                   Q)) {
+            // Range attribute may not longer hold.
+            I->dropPoisonGeneratingReturnAttributes();
             return I;
+          }
         } else { // fshl is a rotate
           // Avoid converting rotate into funnel shift.
           // Only simplify if one operand is constant.
diff --git a/llvm/test/Transforms/InstCombine/fsh.ll b/llvm/test/Transforms/InstCombine/fsh.ll
index 236c69e7a5bcb7..e52e8af5f65ca6 100644
--- a/llvm/test/Transforms/InstCombine/fsh.ll
+++ b/llvm/test/Transforms/InstCombine/fsh.ll
@@ -1068,3 +1068,18 @@ entry:
   %res = call <2 x i31> @llvm.fshl.v2i31(<2 x i31> %x, <2 x i31> zeroinitializer, <2 x i31>  %y)
   ret <2 x i31>  %res
 }
+
+define i8 @fshl_range_trunc(i1 %x) {
+; CHECK-LABEL: @fshl_range_trunc(
+; CHECK-NEXT:    [[ZEXT:%.*]] = zext i1 [[X:%.*]] to i32
+; CHECK-NEXT:    [[OR:%.*]] = or disjoint i32 [[ZEXT]], 126
+; CHECK-NEXT:    [[FSHL:%.*]] = call i32 @llvm.fshl.i32(i32 [[OR]], i32 -2, i32 1)
+; CHECK-NEXT:    [[TR:%.*]] = trunc nuw i32 [[FSHL]] to i8
+; CHECK-NEXT:    ret i8 [[TR]]
+;
+  %zext = zext i1 %x to i32
+  %or = or disjoint i32 %zext, -2
+  %fshl = call range(i32 -4, 2) i32 @llvm.fshl.i32(i32 %or, i32 %or, i32 1)
+  %tr = trunc nsw i32 %fshl to i8
+  ret i8 %tr
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/124429