[DSE] Teach the pass that atomic memory intrinsics are stores.

[dneilson]

Summary

This change teaches DSE that the atomic memory intrinsics are stores
that can be eliminated, and can allow other stores to be eliminated.
This change specifically does not teach DSE that these intrinsics
can be partially eliminated (i.e. length reduced, and dest/src changed);
that will be handled in another change.

[mkazantsev]

I don't see any problems other than the comment inlined, but I don't know this part of code as well. Please wait for someone more familiar with it to take a look.

lib/Analysis/MemoryLocation.cpp

88 ↗

(On Diff #142071)

What was the reason of renaming the parameter? Can it go as a separate NFC?

[dneilson]

ping - any reviewers able to take a crack at this?

[dmgreen]

Am I right in saying these are treated in the same way as unordered atomic stores are? In that they can be deleted by following non-atomic stores, even if they are "stronger" atomically? If so, this looks fine to me.

lib/Transforms/Scalar/DeadStoreElimination.cpp

868 ↗

(On Diff #142071)

This one? This function does it's own special version of DSE. It's about removing stores to dead stack objects iirc.

[dneilson]

In D45535#1071336, @dmgreen wrote:

Am I right in saying these are treated in the same way as unordered atomic stores are? In that they can be deleted by following non-atomic stores, even if they are "stronger" atomically? If so, this looks fine to me.

Yes, I believe that it correct -- at least that's the reasoning I've been going off of. The tests in test/Transforms/DeadStoreElimination/atomic.ll show unordered atomic stores essentially being treated the same as non-atomic stores, so if it's good there then it should be good here.

lib/Transforms/Scalar/DeadStoreElimination.cpp

868 ↗

(On Diff #142071)

Good catch. Thanks!

Looks like there might not be a test for this case at all. I'll try to add one of those as well.

[dneilson]

• Update for reviewer comment.

[dneilson]

lib/Transforms/Scalar/DeadStoreElimination.cpp

868 ↗

(On Diff #142071)

Made the change, but I'm having difficulty creating a test that gets to here. Calls to the memtransfer intrinsics end up getting into the conditional at line 838, above:

// If the call might load from any of our allocas, then any store above
// the call is live.
DeadStackObjects.remove_if([&](Value *I) {
  // See if the call site touches the value.
  return isRefSet(AA->getModRefInfo(CS, I, getPointerSize(I, DL, *TLI)));
});

The source argument for the intrinsic aliases the stack object(s), and this bit of code at 868 ends up having no effect.

[efriedma]

lib/Transforms/Scalar/DeadStoreElimination.cpp

868 ↗

(On Diff #142071)

Yes, I agree it's impossible to get here because of the condition on line 825; an AnyMemTransferInst is also a CallSite. Please just delete this line, then.

[dneilson]

• Deleted memtransferinst case in handleEndBlock.

[efriedma]

LGTM

[dneilson]

In D45535#1075797, @efriedma wrote:

LGTM

Thanks, Eli!

llvm/trunk/include/llvm/Analysis/MemoryLocation.h

 #include "llvm/IR/Metadata.h"
 
 namespace llvm {
 
 class LoadInst;
 class StoreInst;
 class MemTransferInst;
 class MemIntrinsic;
+class AtomicMemTransferInst;
+class AtomicMemIntrinsic;
+class AnyMemTransferInst;
+class AnyMemIntrinsic;
 class TargetLibraryInfo;
 
 /// Representation for a specific memory location.
 ///
 /// This abstraction can be used to represent a specific location in memory.
 /// The goal of the location is to represent enough information to describe
 /// abstract aliasing, modification, and reference behaviors of whatever
 /// value(s) are stored in memory at the particular location.
case Instruction::AtomicRMW:
       return get(cast<AtomicRMWInst>(Inst));
     default:
       return None;
     }
   }
 
   /// Return a location representing the source of a memory transfer.
   static MemoryLocation getForSource(const MemTransferInst *MTI);
+  static MemoryLocation getForSource(const AtomicMemTransferInst *MTI);
+  static MemoryLocation getForSource(const AnyMemTransferInst *MTI);
 
   /// Return a location representing the destination of a memory set or
   /// transfer.
   static MemoryLocation getForDest(const MemIntrinsic *MI);
+  static MemoryLocation getForDest(const AtomicMemIntrinsic *MI);
+  static MemoryLocation getForDest(const AnyMemIntrinsic *MI);
 
   /// Return a location representing a particular argument of a call.
   static MemoryLocation getForArgument(ImmutableCallSite CS, unsigned ArgIdx,
                                        const TargetLibraryInfo &TLI);
 
   explicit MemoryLocation(const Value *Ptr = nullptr,
                           uint64_t Size = UnknownSize,
                           const AAMDNodes &AATags = AAMDNodes())

llvm/trunk/lib/Analysis/MemoryLocation.cpp

MemoryLocation MemoryLocation::get(const AtomicRMWInst *RMWI) {
   const auto &DL = RMWI->getModule()->getDataLayout();
 
   return MemoryLocation(RMWI->getPointerOperand(),
                         DL.getTypeStoreSize(RMWI->getValOperand()->getType()),
                         AATags);
 }
 
 MemoryLocation MemoryLocation::getForSource(const MemTransferInst *MTI) {
+  return getForSource(cast<AnyMemTransferInst>(MTI));
+}
+
+MemoryLocation MemoryLocation::getForSource(const AtomicMemTransferInst *MTI) {
+  return getForSource(cast<AnyMemTransferInst>(MTI));
+}
+
+MemoryLocation MemoryLocation::getForSource(const AnyMemTransferInst *MTI) {
   uint64_t Size = UnknownSize;
   if (ConstantInt *C = dyn_cast<ConstantInt>(MTI->getLength()))
     Size = C->getValue().getZExtValue();
 
   // memcpy/memmove can have AA tags. For memcpy, they apply
   // to both the source and the destination.
   AAMDNodes AATags;
   MTI->getAAMetadata(AATags);
 
   return MemoryLocation(MTI->getRawSource(), Size, AATags);
 }
 
-MemoryLocation MemoryLocation::getForDest(const MemIntrinsic *MTI) {
+MemoryLocation MemoryLocation::getForDest(const MemIntrinsic *MI) {
+  return getForDest(cast<AnyMemIntrinsic>(MI));
+}
+
+MemoryLocation MemoryLocation::getForDest(const AtomicMemIntrinsic *MI) {
+  return getForDest(cast<AnyMemIntrinsic>(MI));
+}
+
+MemoryLocation MemoryLocation::getForDest(const AnyMemIntrinsic *MI) {
   uint64_t Size = UnknownSize;
-  if (ConstantInt *C = dyn_cast<ConstantInt>(MTI->getLength()))
+  if (ConstantInt *C = dyn_cast<ConstantInt>(MI->getLength()))
     Size = C->getValue().getZExtValue();
 
   // memcpy/memmove can have AA tags. For memcpy, they apply
   // to both the source and the destination.
   AAMDNodes AATags;
-  MTI->getAAMetadata(AATags);
+  MI->getAAMetadata(AATags);
 
-  return MemoryLocation(MTI->getRawDest(), Size, AATags);
+  return MemoryLocation(MI->getRawDest(), Size, AATags);
 }
 
 MemoryLocation MemoryLocation::getForArgument(ImmutableCallSite CS,
                                               unsigned ArgIdx,
                                               const TargetLibraryInfo &TLI) {
   AAMDNodes AATags;
   CS->getAAMetadata(AATags);
   const Value *Arg = CS.getArgument(ArgIdx);

llvm/trunk/lib/Transforms/Scalar/DeadStoreElimination.cpp

if (isa<StoreInst>(I))
     return true;
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
     switch (II->getIntrinsicID()) {
     default:
       return false;
     case Intrinsic::memset:
     case Intrinsic::memmove:
     case Intrinsic::memcpy:
+    case Intrinsic::memcpy_element_unordered_atomic:
+    case Intrinsic::memmove_element_unordered_atomic:
+    case Intrinsic::memset_element_unordered_atomic:
     case Intrinsic::init_trampoline:
     case Intrinsic::lifetime_end:
       return true;
     }
   }
   if (auto CS = CallSite(I)) {
     if (Function *F = CS.getCalledFunction()) {
       StringRef FnName = F->getName();

 /// Return a Location stored to by the specified instruction. If isRemovable
 /// returns true, this function and getLocForRead completely describe the memory
 /// operations for this instruction.
 static MemoryLocation getLocForWrite(Instruction *Inst) {
   
   if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
     return MemoryLocation::get(SI);
 
-  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
+  if (auto *MI = dyn_cast<AnyMemIntrinsic>(Inst)) {
     // memcpy/memmove/memset.
     MemoryLocation Loc = MemoryLocation::getForDest(MI);
     return Loc;
   }
 
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
     switch (II->getIntrinsicID()) {
     default:

 /// Return the location read by the specified "hasAnalyzableMemoryWrite"
 /// instruction if any.
 static MemoryLocation getLocForRead(Instruction *Inst,
                                     const TargetLibraryInfo &TLI) {
   assert(hasAnalyzableMemoryWrite(Inst, TLI) && "Unknown instruction case");
 
   // The only instructions that both read and write are the mem transfer
   // instructions (memcpy/memmove).
-  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Inst))
+  if (auto *MTI = dyn_cast<AnyMemTransferInst>(Inst))
     return MemoryLocation::getForSource(MTI);
   return MemoryLocation();
 }
 
 /// If the value of this instruction and the memory it writes to is unused, may
 /// we delete this instruction?
 static bool isRemovable(Instruction *I) {
   // Don't remove volatile/atomic stores.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
     case Intrinsic::init_trampoline:
       // Always safe to remove init_trampoline.
       return true;
     case Intrinsic::memset:
     case Intrinsic::memmove:
     case Intrinsic::memcpy:
       // Don't remove volatile memory intrinsics.
       return !cast<MemIntrinsic>(II)->isVolatile();
+    case Intrinsic::memcpy_element_unordered_atomic:
+    case Intrinsic::memmove_element_unordered_atomic:
+    case Intrinsic::memset_element_unordered_atomic:
+      return true;
     }
   }
 
   // note: only get here for calls with analyzable writes - i.e. libcalls
   if (auto CS = CallSite(I))
     return CS.getInstruction()->use_empty();
 
   return false;
 }
 
 /// Returns true if the end of this instruction can be safely shortened in
 /// length.
 static bool isShortenableAtTheEnd(Instruction *I) {
   // Don't shorten stores for now
   if (isa<StoreInst>(I))
     return false;
 
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
     switch (II->getIntrinsicID()) {
       default: return false;
       case Intrinsic::memset:
       case Intrinsic::memcpy:
         // Do shorten memory intrinsics.
         // FIXME: Add memmove if it's also safe to transform.
+        // TODO: Add atomic memcpy/memset
         return true;
     }
   }
 
   // Don't shorten libcalls calls for now.
 
   return false;
 }
 
 /// Returns true if the beginning of this instruction can be safely shortened
 /// in length.
 static bool isShortenableAtTheBeginning(Instruction *I) {
   // FIXME: Handle only memset for now. Supporting memcpy/memmove should be
   // easily done by offsetting the source address.
+  // TODO: Handle atomic memory intrinsics
   IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
   return II && II->getIntrinsicID() == Intrinsic::memset;
 }
 
 /// Return the pointer that is being written to.
 static Value *getStoredPointerOperand(Instruction *I) {
   //TODO: factor this to reuse getLocForWrite
   MemoryLocation Loc = getLocForWrite(I);
static bool isPossibleSelfRead(Instruction *Inst,
   MemoryLocation InstReadLoc = getLocForRead(Inst, TLI);
   if (!InstReadLoc.Ptr)
     return false; // Not a reading instruction.
 
   // If the read and written loc obviously don't alias, it isn't a read.
   if (AA.isNoAlias(InstReadLoc, InstStoreLoc))
     return false;
 
-  if (isa<MemCpyInst>(Inst)) {
+  if (isa<AnyMemCpyInst>(Inst)) {
     // LLVM's memcpy overlap semantics are not fully fleshed out (see PR11763)
     // but in practice memcpy(A <- B) either means that A and B are disjoint or
     // are equal (i.e. there are not partial overlaps).  Given that, if we have:
     //
     //   memcpy/memmove(A <- B)  // DepWrite
     //   memcpy(A <- B)  // Inst
     //
     // with Inst reading/writing a >= size than DepWrite, we can reason as
for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
 
     // If we encounter a use of the pointer, it is no longer considered dead
     if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
       if (!L->isUnordered()) // Be conservative with atomic/volatile load
         break;
       LoadedLoc = MemoryLocation::get(L);
     } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
       LoadedLoc = MemoryLocation::get(V);
-    } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
-      LoadedLoc = MemoryLocation::getForSource(MTI);
     } else if (!BBI->mayReadFromMemory()) {
       // Instruction doesn't read memory.  Note that stores that weren't removed
       // above will hit this case.
       continue;
     } else {
       // Unknown inst; assume it clobbers everything.
       break;
     }

llvm/trunk/test/Transforms/DeadStoreElimination/memintrinsics.ll

 
 declare void @llvm.memcpy.element.unordered.atomic.p0i16.p0i16.i16(i16* nocapture, i16* nocapture, i16, i32) nounwind
 declare void @llvm.memmove.element.unordered.atomic.p0i16.p0i16.i16(i16* nocapture, i16* nocapture, i16, i32) nounwind
 declare void @llvm.memset.element.unordered.atomic.p0i16.i16(i16* nocapture, i8, i16, i32) nounwind
 
 
 define void @test4() {
 ; CHECK-LABEL: @test4(
-; CHECK-NEXT:    [[A:%.*]] = alloca i16, i16 1024, align 2
-; CHECK-NEXT:    [[B:%.*]] = alloca i16, i16 1024, align 2
-; CHECK-NEXT:    store atomic i16 0, i16* [[B]] unordered, align 2
-; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i16.p0i16.i16(i16* align 2 [[A]], i16* align 2 [[B]], i16 1024, i32 2)
 ; CHECK-NEXT:    ret void
 ;
   %A = alloca i16, i16 1024, align 2
   %B = alloca i16, i16 1024, align 2
 
   store atomic i16 0, i16* %A unordered, align 2 ;; Written to by memcpy
   store atomic i16 0, i16* %B unordered, align 2 ;; Read by memcpy
 
   call void @llvm.memcpy.element.unordered.atomic.p0i16.p0i16.i16(i16* align 2 %A, i16* align 2 %B, i16 1024, i32 2)
 
   ret void
 }
 
 define void @test5() {
 ; CHECK-LABEL: @test5(
-; CHECK-NEXT:    [[A:%.*]] = alloca i16, i16 1024, align 2
-; CHECK-NEXT:    [[B:%.*]] = alloca i16, i16 1024, align 2
-; CHECK-NEXT:    store atomic i16 0, i16* [[B]] unordered, align 2
-; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i16.p0i16.i16(i16* align 2 [[A]], i16* align 2 [[B]], i16 1024, i32 2)
 ; CHECK-NEXT:    ret void
 ;
   %A = alloca i16, i16 1024, align 2
   %B = alloca i16, i16 1024, align 2
 
   store atomic i16 0, i16* %A unordered, align 2 ;; Written to by memmove
   store atomic i16 0, i16* %B unordered, align 2 ;; Read by memmove
 
   call void @llvm.memmove.element.unordered.atomic.p0i16.p0i16.i16(i16* align 2 %A, i16* align 2 %B, i16 1024, i32 2)
 
   ret void
 }
 
 define void @test6() {
 ; CHECK-LABEL: @test6(
-; CHECK-NEXT:    [[A:%.*]] = alloca i16, i16 1024, align 2
-; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i16.i16(i16* align 2 [[A]], i8 0, i16 1024, i32 2)
 ; CHECK-NEXT:    ret void
 ;
   %A = alloca i16, i16 1024, align 2
   %B = alloca i16, i16 1024, align 2
 
   store atomic i16 0, i16* %A unordered, align 2 ;; Written to by memset
 
   call void @llvm.memset.element.unordered.atomic.p0i16.i16(i16* align 2 %A, i8 0, i16 1024, i32 2)
 
   ret void
 }

llvm/trunk/test/Transforms/DeadStoreElimination/simple.ll

;
   store i32 30, i32* %p, align 4
   ret void
 }
 
 ; Should delete store of 10 even though memset is a may-store to P (P and Q may
 ; alias).
 define void @test6_atomic(i32* align 4 %p, i8* align 4 %q) {
 ; CHECK-LABEL: @test6_atomic(
-; CHECK-NEXT:    store atomic i32 10, i32* [[P:%.*]] unordered, align 4
 ; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 4 [[Q:%.*]], i8 42, i64 900, i32 4)
-; CHECK-NEXT:    store atomic i32 30, i32* [[P]] unordered, align 4
+; CHECK-NEXT:    store atomic i32 30, i32* [[P:%.*]] unordered, align 4
 ; CHECK-NEXT:    ret void
 ;
   store atomic i32 10, i32* %p unordered, align 4       ;; dead.
   call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 4 %q, i8 42, i64 900, i32 4)
   store atomic i32 30, i32* %p unordered, align 4
   ret void
 }
 
;
   store i32 30, i32* %p, align 4
   ret void
 }
 
 ; Should delete store of 10 even though memcpy is a may-store to P (P and Q may
 ; alias).
 define void @test7_atomic(i32* align 4 %p, i8* align 4 %q, i8* noalias align 4 %r) {
 ; CHECK-LABEL: @test7_atomic(
-; CHECK-NEXT:    store atomic i32 10, i32* [[P:%.*]] unordered, align 4
 ; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 4 [[Q:%.*]], i8* align 4 [[R:%.*]], i64 900, i32 4)
-; CHECK-NEXT:    store atomic i32 30, i32* [[P]] unordered, align 4
+; CHECK-NEXT:    store atomic i32 30, i32* [[P:%.*]] unordered, align 4
 ; CHECK-NEXT:    ret void
 ;
   store atomic i32 10, i32* %p unordered, align 4       ;; dead.
   call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 4 %q, i8* align 4 %r, i64 900, i32 4)
   store atomic i32 30, i32* %p unordered, align 4
   ret void
 }
 
;
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false)
   ret void
 }
 
 ;; Fully dead overwrite of memcpy.
 define void @test15_atomic(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test15_atomic(
 ; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i32 1)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   ret void
 }
 
-; It would only be valid to remove the non-atomic memcpy
+;; Fully dead overwrite of memcpy.
 define void @test15_atomic_weaker(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test15_atomic_weaker(
-; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false)
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i32 1)
+; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false)
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   ret void
 }
 
-; It would only be valid to remove the non-atomic memcpy
+;; Fully dead overwrite of memcpy.
 define void @test15_atomic_weaker_2(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test15_atomic_weaker_2(
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
-; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i1 false)
+; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false)
   ret void
 }
 
 ;; Full overwrite of smaller memcpy.
 define void @test16(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test16(
 ; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 8, i1 false)
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false)
   ret void
 }
 
 ;; Full overwrite of smaller memcpy.
 define void @test16_atomic(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test16_atomic(
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 8, i32 1)
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i32 1)
+; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 8, i32 1)
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   ret void
 }
 
 ;; Full overwrite of smaller memory where overwrite has stronger atomicity
 define void @test16_atomic_weaker(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test16_atomic_weaker(
-; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 8, i1 false)
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i32 1)
+; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 8, i1 false)
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   ret void
 }
 
 ;; Full overwrite of smaller memory where overwrite has weaker atomicity.
 define void @test16_atomic_weaker_2(i8* %P, i8* %Q) nounwind ssp {
 ; CHECK-LABEL: @test16_atomic_weaker_2(
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 8, i32 1)
-; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i1 false)
+; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 8, i32 1)
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false)
   ret void
 }
 
 ;; Overwrite of memset by memcpy.
 define void @test17(i8* %P, i8* noalias %Q) nounwind ssp {
 ; CHECK-LABEL: @test17(
 ; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memset.p0i8.i64(i8* %P, i8 42, i64 8, i1 false)
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false)
   ret void
 }
 
 ;; Overwrite of memset by memcpy.
 define void @test17_atomic(i8* %P, i8* noalias %Q) nounwind ssp {
 ; CHECK-LABEL: @test17_atomic(
-; CHECK-NEXT:    tail call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 1 [[P:%.*]], i8 42, i64 8, i32 1)
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
+; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 1 %P, i8 42, i64 8, i32 1)
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   ret void
 }
 
 ;; Overwrite of memset by memcpy. Overwrite is stronger atomicity. We can
 ;; remove the memset.
 define void @test17_atomic_weaker(i8* %P, i8* noalias %Q) nounwind ssp {
 ; CHECK-LABEL: @test17_atomic_weaker(
-; CHECK-NEXT:    tail call void @llvm.memset.p0i8.i64(i8* align 1 [[P:%.*]], i8 42, i64 8, i1 false)
-; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
+; CHECK-NEXT:    tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memset.p0i8.i64(i8* align 1 %P, i8 42, i64 8, i1 false)
   tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1)
   ret void
 }
 
 ;; Overwrite of memset by memcpy. Overwrite is weaker atomicity. We can remove
 ;; the memset.
 define void @test17_atomic_weaker_2(i8* %P, i8* noalias %Q) nounwind ssp {
 ; CHECK-LABEL: @test17_atomic_weaker_2(
-; CHECK-NEXT:    tail call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 1 [[P:%.*]], i8 42, i64 8, i32 1)
-; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q:%.*]], i64 12, i1 false)
+; CHECK-NEXT:    tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false)
 ; CHECK-NEXT:    ret void
 ;
   tail call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 1 %P, i8 42, i64 8, i32 1)
   tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false)
   ret void
 }
 
 ; Should not delete the volatile memset.