Index: llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -1461,9 +1461,24 @@
       llvm::PointerMayBeCaptured(SrcAlloca, /* ReturnCaptures=*/true,
                                  /* StoreCaptures= */ true))
     return false;
+  // TODO: refine the before memcpy condition for dest alloca
+  // Bailout if any Mod except lifetime intrinsics exists before memcpy on the
+  // dest.
+  auto DestLoc = MemoryLocation(DestAlloca, LocationSize::precise(Size));
+  for (User *U : DestAlloca->users()) {
+    if (auto *I = dyn_cast<Instruction>(U)) {
+      if (!I->comesBefore(Store))
+        continue;
+      if (auto *II = dyn_cast<IntrinsicInst>(I);
+          II && II->isLifetimeStartOrEnd())
+        continue;
+      if (isModSet(AA->getModRefInfo(I, DestLoc)))
+        return false;
+    }
+  }
 
-  // 3. The src has no read and no write except lifetime intrinsics, from after
-  // the full copy to the end of the BB.
+  // 3. The src has no read and no write except lifetime intrinsics, from
+  // after the full copy to the end of the BB.
   for (auto It = ++Store->getIterator(), E = Store->getParent()->end(); It != E;
        ++It) {
     // See all users of the SrcAlloca, between full copy and end of the BB.
@@ -1485,39 +1500,35 @@
   DestAlloca->replaceAllUsesWith(SrcAlloca);
 
   // Erase redundant lifetimes
-  // FIXME: currently this doesn't work correctly for lifetime.end
   IntrinsicInst *LifetimeStart = nullptr, *LifetimeEnd = nullptr;
-  for (auto It = SrcAlloca->user_begin(), E = SrcAlloca->user_end(); It != E;
-       ++It) {
-    if (auto *II = dyn_cast<IntrinsicInst>(*It);
-        II && II->isLifetimeStartOrEnd()) {
+  SmallVector<IntrinsicInst *, 4> LifetimeToErase;
+  for (User *U : SrcAlloca->users()) {
+    if (auto *II = dyn_cast<IntrinsicInst>(U)) {
       if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
-        if (!LifetimeStart) {
+        if (!LifetimeStart)
           LifetimeStart = II;
-          continue;
-        }
-        if (II->comesBefore(LifetimeStart)) {
-          eraseInstruction(LifetimeStart);
+        else if (II->comesBefore(LifetimeStart)) {
+          LifetimeToErase.push_back(LifetimeStart);
           LifetimeStart = II;
-          continue;
-        }
+        } else
+          LifetimeToErase.push_back(II);
       } else if (II->getIntrinsicID() == Intrinsic::lifetime_end) {
-        if (!LifetimeEnd) {
+        if (!LifetimeEnd)
           LifetimeEnd = II;
-          continue;
-        }
-        if (LifetimeEnd->comesBefore(II)) {
-          eraseInstruction(LifetimeEnd);
+        else if (LifetimeEnd->comesBefore(II)) {
+          LifetimeToErase.push_back(LifetimeEnd);
           LifetimeEnd = II;
-          continue;
-        }
+        } else
+          LifetimeToErase.push_back(II);
       }
-      // FIXME: this is necessary to remove redundant lifetime.end, but breaks
-      // for some reasons
-      // eraseInstruction(II);
     }
   }
 
+  // FIXME: still crash on erase
+  for (auto *II : make_early_inc_range(LifetimeToErase)) {
+    eraseInstruction(II);
+  }
+
   // Drop metadata on the source alloca.
   SrcAlloca->dropUnknownNonDebugMetadata();