[llvm] 3c0096a - [MergeICmps] Don't call comesBefore() if in different blocks (PR53959)

[Nikita Popov via llvm-commits]

Author: Nikita Popov
Date: 2022-02-22T12:27:20+01:00
New Revision: 3c0096a1d45689389e11c385fd4ab98fdec80323

URL: https://github.com/llvm/llvm-project/commit/3c0096a1d45689389e11c385fd4ab98fdec80323
DIFF: https://github.com/llvm/llvm-project/commit/3c0096a1d45689389e11c385fd4ab98fdec80323.diff

LOG: [MergeICmps] Don't call comesBefore() if in different blocks (PR53959)

Only call comesBefore() if the instructions are in the same block.
Otherwise make a conservative assumption.

Fixes https://github.com/llvm/llvm-project/issues/53959.

Added: 
    llvm/test/Transforms/MergeICmps/X86/pr53959.ll

Modified: 
    llvm/lib/Transforms/Scalar/MergeICmps.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Scalar/MergeICmps.cpp b/llvm/lib/Transforms/Scalar/MergeICmps.cpp
index aac0deea5be3d..d38e362015ed7 100644
--- a/llvm/lib/Transforms/Scalar/MergeICmps.cpp
+++ b/llvm/lib/Transforms/Scalar/MergeICmps.cpp
@@ -244,7 +244,7 @@ bool BCECmpBlock::canSinkBCECmpInst(const Instruction *Inst,
     auto MayClobber = [&](LoadInst *LI) {
       // If a potentially clobbering instruction comes before the load,
       // we can still safely sink the load.
-      return !Inst->comesBefore(LI) &&
+      return (Inst->getParent() != LI->getParent() || !Inst->comesBefore(LI)) &&
              isModSet(AA.getModRefInfo(Inst, MemoryLocation::get(LI)));
     };
     if (MayClobber(Cmp.Lhs.LoadI) || MayClobber(Cmp.Rhs.LoadI))

diff  --git a/llvm/test/Transforms/MergeICmps/X86/pr53959.ll b/llvm/test/Transforms/MergeICmps/X86/pr53959.ll
new file mode 100644
index 0000000000000..b7ba8bd39ba53
--- /dev/null
+++ b/llvm/test/Transforms/MergeICmps/X86/pr53959.ll
@@ -0,0 +1,48 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -mergeicmps < %s | FileCheck %s
+
+target triple = "x86_64-unknown-linux-gnu"
+
+ at c = external global i32, align 4
+
+define i1 @d() {
+; CHECK-LABEL: @d(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[G:%.*]] = alloca [8 x i64], align 16
+; CHECK-NEXT:    [[IDX1:%.*]] = getelementptr inbounds [8 x i64], [8 x i64]* [[G]], i64 0, i64 0
+; CHECK-NEXT:    [[V1:%.*]] = load i64, i64* [[IDX1]], align 8
+; CHECK-NEXT:    [[IDX2:%.*]] = getelementptr inbounds [8 x i64], [8 x i64]* [[G]], i64 0, i64 0
+; CHECK-NEXT:    [[V2:%.*]] = load i64, i64* [[IDX2]], align 8
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[V1]], [[V2]]
+; CHECK-NEXT:    br label [[SPLIT:%.*]]
+; CHECK:       split:
+; CHECK-NEXT:    [[X:%.*]] = load volatile i32, i32* @c, align 4
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF:%.*]], label [[EXIT:%.*]]
+; CHECK:       if:
+; CHECK-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[X]], 0
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[P:%.*]] = phi i1 [ false, [[SPLIT]] ], [ [[TOBOOL]], [[IF]] ]
+; CHECK-NEXT:    ret i1 [[P]]
+;
+entry:
+  %g = alloca [8 x i64], align 16
+  %idx1 = getelementptr inbounds [8 x i64], [8 x i64]* %g, i64 0, i64 0
+  %v1 = load i64, i64* %idx1, align 8
+  %idx2 = getelementptr inbounds [8 x i64], [8 x i64]* %g, i64 0, i64 0
+  %v2 = load i64, i64* %idx2, align 8
+  %cmp = icmp eq i64 %v1, %v2
+  br label %split
+
+split:
+  %x = load volatile i32, i32* @c, align 4
+  br i1 %cmp, label %if, label %exit
+
+if:
+  %tobool = icmp ne i32 %x, 0
+  br label %exit
+
+exit:
+  %p = phi i1 [ false, %split ], [ %tobool, %if ]
+  ret i1 %p
+}