[llvm] [MemCpyOpt] Avoid infinite loops in `MemCpyOptPass::processMemCpyMemCpyDependence` (PR #103218)

[Yingwei Zheng via llvm-commits]

https://github.com/dtcxzyw updated https://github.com/llvm/llvm-project/pull/103218

>From 7bc99d38c051c67517696aef2644725e140a57f0 Mon Sep 17 00:00:00 2001
From: Yingwei Zheng <dtcxzyw2333 at gmail.com>
Date: Tue, 13 Aug 2024 22:36:00 +0800
Subject: [PATCH 1/2] [MemCpyOpt] Avoid infinite loops in
 `MemCpyOptPass::processMemCpyMemCpyDependence`

---
 llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp |  4 ++++
 llvm/test/Transforms/MemCpyOpt/pr102994.ll     | 18 ++++++++++++++++++
 2 files changed, 22 insertions(+)
 create mode 100644 llvm/test/Transforms/MemCpyOpt/pr102994.ll

diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
index 1d779128e454c1..3f15fa2163d270 100644
--- a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -1193,6 +1193,10 @@ bool MemCpyOptPass::processMemCpyMemCpyDependence(MemCpyInst *M,
       CopySourceAlign = commonAlignment(*CopySourceAlign, MForwardOffset);
   }
 
+  // Avoid infinite loops
+  if (BAA.isMustAlias(M->getSource(), CopySource))
+    return false;
+
   // Verify that the copied-from memory doesn't change in between the two
   // transfers.  For example, in:
   //    memcpy(a <- b)
diff --git a/llvm/test/Transforms/MemCpyOpt/pr102994.ll b/llvm/test/Transforms/MemCpyOpt/pr102994.ll
new file mode 100644
index 00000000000000..53d6335f6e9d0b
--- /dev/null
+++ b/llvm/test/Transforms/MemCpyOpt/pr102994.ll
@@ -0,0 +1,18 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes=memcpyopt < %s | FileCheck %s
+
+ at g1 = external global i8
+ at g2 = external global [64 x i8]
+
+define void @func() {
+; CHECK-LABEL: define void @func() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr getelementptr inbounds (i8, ptr @g2, i64 16), ptr getelementptr inbounds nuw (i8, ptr @g2, i64 16), i64 20, i1 false)
+; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr @g1, ptr getelementptr inbounds (i8, ptr @g2, i64 24), i64 1, i1 false)
+; CHECK-NEXT:    ret void
+;
+entry:
+  call void @llvm.memcpy.p0.p0.i64(ptr getelementptr inbounds (i8, ptr @g2, i64 16), ptr getelementptr inbounds nuw (i8, ptr @g2, i64 16), i64 20, i1 false)
+  call void @llvm.memcpy.p0.p0.i64(ptr @g1, ptr getelementptr inbounds (i8, ptr @g2, i64 24), i64 1, i1 false)
+  ret void
+}

>From bd8d6dd19ca3f424751875d3d026749d6d935bd6 Mon Sep 17 00:00:00 2001
From: Yingwei Zheng <dtcxzyw2333 at gmail.com>
Date: Thu, 22 Aug 2024 15:47:09 +0800
Subject: [PATCH 2/2] [MemCpyOpt] Add new tests without constantexpr

---
 llvm/test/Transforms/MemCpyOpt/pr102994.ll | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/llvm/test/Transforms/MemCpyOpt/pr102994.ll b/llvm/test/Transforms/MemCpyOpt/pr102994.ll
index 53d6335f6e9d0b..9a782b9b3752a2 100644
--- a/llvm/test/Transforms/MemCpyOpt/pr102994.ll
+++ b/llvm/test/Transforms/MemCpyOpt/pr102994.ll
@@ -3,6 +3,7 @@
 
 @g1 = external global i8
 @g2 = external global [64 x i8]
+ at g3 = global i8 0, align 1
 
 define void @func() {
 ; CHECK-LABEL: define void @func() {
@@ -16,3 +17,23 @@ entry:
   call void @llvm.memcpy.p0.p0.i64(ptr @g1, ptr getelementptr inbounds (i8, ptr @g2, i64 24), i64 1, i1 false)
   ret void
 }
+
+define void @func2(ptr %p) {
+; CHECK-LABEL: define void @func2(
+; CHECK-SAME: ptr [[P:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr i8, ptr [[P]], i64 32
+; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr i8, ptr [[P]], i64 34
+; CHECK-NEXT:    [[GEP3:%.*]] = getelementptr i8, ptr [[P]], i64 32
+; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr [[GEP1]], ptr [[GEP3]], i64 32, i1 false)
+; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr @g3, ptr [[GEP2]], i64 1, i1 false)
+; CHECK-NEXT:    ret void
+;
+entry:
+  %gep1 = getelementptr i8, ptr %p, i64 32
+  %gep2 = getelementptr i8, ptr %p, i64 34
+  %gep3 = getelementptr i8, ptr %p, i64 32
+  call void @llvm.memcpy.p0.p0.i64(ptr %gep1, ptr %gep3, i64 32, i1 false)
+  call void @llvm.memcpy.p0.p0.i64(ptr @g3, ptr %gep2, i64 1, i1 false)
+  ret void
+}