[MemCpyOpt] Don't sink LoadInst below possible clobber.

[bryant]

Summary

Currently, MemCpyOpt::processStore will convert this

val = load A
store _, B      ; may-alias A
store _, C      ; may-alias A + D
store val, D

into

store _, C      ; may-alias A + D
val = load A
store val, D
store _, B      ; may-alias A

and then into

store _, C      ; may-alias A + D
memcpy(D, A)
store _, B      ; may-alias A

This is incorrect, since C and A may alias and therefore cannot be re-ordered.

Adding Amaury since this was introduced in http://reviews.llvm.org/D16523 .

[bryant]

For the sake of clarity, I've broken down the precise sequence of
transformations.

1. Original:

val = load A
store _, B      ; may-alias A
store _, C      ; may-alias A + D
store val, D

2. Lift store val, D and its dependents and clobbers above the store to B:

val = load A
store _, C      ; may-alias A + D
store val, D
store _, B      ; may-alias A

3. Implicitly lower the load to its counterpart store (this is the bug):

store _, C      ; may-alias A + D
val = load A
store val, D
store _, B      ; may-alias A

4. Replace the load and store with a memcpy:

store _, C      ; may-alias A + D
memcpy(D, A)
store _, B      ; may-alias A

EDIT: Formatting.
EDIT 2: Ordering.

[efriedma]

lib/Transforms/Scalar/MemCpyOptimizer.cpp

619 ↗

(On Diff #78398)

I'm not really following what this patch is doing. Whether it's legal to sink the load to location P isn't related to whether it's legal to hoist the store to location P.

On a side-note, this whole block of code will never run in a normal pass pipeline; "T->isAggregateType()" will never be true because instcombine breaks up aggregate loads and stores.

test/Transforms/MemCpyOpt/load-store-to-memcpy.ll

2 ↗

(On Diff #78398)

-aa-eval doesn't do anything useful here.

[bryant]

better test:

• inject some padding into the agg to thwart instcombine from exploding load/store.
• remove -aa-eval.

[bryant]

lib/Transforms/Scalar/MemCpyOptimizer.cpp

619 ↗

(On Diff #78398)

I'm not really following what this patch is doing. Whether it's legal to sink
the load to location P isn't related to whether it's legal to hoist the store
to location P.

The store isn't necessarily the only thing to be hoisted to P. Instructions that
the store 1) depends on or 2) aliases also need to be hoisted. That's moveUp's
job: To figure out which instructions, in addition to the store, need to be
hoisted.

Once all the hoisting is done, a memcpy/memmove is created to replace the load
and store. However, this implies sinking the load past the hoisted stuff to just
before the store, which may not be legal.

So this patch ensures that it's legal for the load to sink past the hoisted
stuff, even though no explicit sinking is needed.

Does that make sense?

On a side-note, this whole block of code will never run in a normal pass
pipeline; "T->isAggregateType()" will never be true because instcombine
breaks up aggregate loads and stores.

InstCombine refrains from exploding aggregate loads and stores when there's
padding. I've updated the test case accordingly.

[bryant]

Update FileChecks to the new aggregate.

[bryant]

lib/Transforms/Scalar/MemCpyOptimizer.cpp

619 ↗

(On Diff #78398)

Here's an example, in case the explanation wasn't clear:

%x = load
P   ; clobbers the load
X   ; clobbers the load and store
store %x

MCO will turn this into

X   ; clobbers the load and store
%x = load
store %x       ; merged with the load into a memcpy
P   ; clobbers the load

which is bad, since the load is moved past X.

[efriedma]

LGTM with the testcase fixed... but please don't commit changes to this code unless you're intending to write a followup which actually makes it useful. (If you're going to leave it in its current state, we might as well just delete it.)

lib/Transforms/Scalar/MemCpyOptimizer.cpp

619 ↗

(On Diff #78398)

Oh, wait, I see, I was misunderstanding how the algorithm works (I somehow thought one of the loops was iterating the wrong way). Yes, your approach makes sense.

test/Transforms/MemCpyOpt/load-store-to-memcpy.ll

2 ↗

(On Diff #78863)

evaluate-aa-metadata doesn't do anything here.

[bryant]

In D26811#603131, @efriedma wrote:

LGTM with the testcase fixed... but please don't commit changes to this code unless you're intending to write a followup which actually makes it useful. (If you're going to leave it in its current state, we might as well just delete it.)

Not sure that I understand. Are you suggesting that we leave the bug unfixed?

[deadalnix]

In D26811#603131, @efriedma wrote:

LGTM with the testcase fixed... but please don't commit changes to this code unless you're intending to write a followup which actually makes it useful. (If you're going to leave it in its current state, we might as well just delete it.)

What makes you think this isn't useful and should be deleted ? This is the only thing that can do load/store forwarding in the codebase when aggregates contains padding, so this ends up being important. I think it is very much worth committing.

However, if you have a better alternative to suggest, please do, I can even probably spend some time on it. Improving aggregates support is important to me a several other non clang front end devs.

[efriedma]

Oh, we special-case stores of aggregate types which specifically contain padding? Ugh, I missed that... that's *really* subtle. Also, there's probably some better way to set padding bits to undef, but that's an argument for another time.

[deadalnix]

@efriedma load/store of aggregate cannot be optimized the same way other load/store are done for various reasons.

If the aggregate do not contains padding, then it is possible to deaggregate in a serie of load/store and recompose the integrate. But in the case the aggregate has some padding, this transformation is lossy. However, doing it is memcpy and alike is not lossy so here you go. This is also useful for large aggregates. This transformation is already gated on aggregates types.

lib/Transforms/Scalar/MemCpyOptimizer.cpp

598 ↗

(On Diff #78863)

This is where the transformation is gated on aggregates.

[bryant]

Removed -evaluate-aa-metadata.

[bryant]

test/Transforms/MemCpyOpt/load-store-to-memcpy.ll

6 ↗

(On Diff #79340)

fix typo.

[bryant]

• "hould" to "should."
• remove unneeded stderr redirect.

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

static unsigned findCommonAlignment(const DataLayout &DL, const StoreInst *SI,
 
   return std::min(StoreAlign, LoadAlign);
 }
 
 // This method try to lift a store instruction before position P.
 // It will lift the store and its argument + that anything that
 // may alias with these.
 // The method returns true if it was successful.
-static bool moveUp(AliasAnalysis &AA, StoreInst *SI, Instruction *P) {
+static bool moveUp(AliasAnalysis &AA, StoreInst *SI, Instruction *P,
+                   const LoadInst *LI) {
   // If the store alias this position, early bail out.
   MemoryLocation StoreLoc = MemoryLocation::get(SI);
   if (AA.getModRefInfo(P, StoreLoc) != MRI_NoModRef)
     return false;
 
   // Keep track of the arguments of all instruction we plan to lift
   // so we can make sure to lift them as well if apropriate.
   DenseSet<Instruction*> Args;
   if (auto *Ptr = dyn_cast<Instruction>(SI->getPointerOperand()))
     if (Ptr->getParent() == SI->getParent())
       Args.insert(Ptr);
 
   // Instruction to lift before P.
   SmallVector<Instruction*, 8> ToLift;
 
   // Memory locations of lifted instructions.
-  SmallVector<MemoryLocation, 8> MemLocs;
-  MemLocs.push_back(StoreLoc);
+  SmallVector<MemoryLocation, 8> MemLocs{StoreLoc};
 
   // Lifted callsites.
   SmallVector<ImmutableCallSite, 8> CallSites;
 
+  const MemoryLocation LoadLoc = MemoryLocation::get(LI);
+
   for (auto I = --SI->getIterator(), E = P->getIterator(); I != E; --I) {
     auto *C = &*I;
 
     bool MayAlias = AA.getModRefInfo(C) != MRI_NoModRef;
 
     bool NeedLift = false;
     if (Args.erase(C))
       NeedLift = true;
     else if (MayAlias) {
       NeedLift = any_of(MemLocs, [C, &AA](const MemoryLocation &ML) {
         return AA.getModRefInfo(C, ML);
       });
 
       if (!NeedLift)
         NeedLift = any_of(CallSites, [C, &AA](const ImmutableCallSite &CS) {
           return AA.getModRefInfo(C, CS);
         });
     }
 
     if (!NeedLift)
       continue;
 
     if (MayAlias) {
-      if (auto CS = ImmutableCallSite(C)) {
+      // Since LI is implicitly moved downwards past the lifted instructions,
+      // none of them may modify its source.
+      if (AA.getModRefInfo(C, LoadLoc) & MRI_Mod)
+        return false;
+      else if (auto CS = ImmutableCallSite(C)) {
         // If we can't lift this before P, it's game over.
         if (AA.getModRefInfo(P, CS) != MRI_NoModRef)
           return false;
 
         CallSites.push_back(CS);
       } else if (isa<LoadInst>(C) || isa<StoreInst>(C) || isa<VAArgInst>(C)) {
         // If we can't lift this before P, it's game over.
         auto ML = MemoryLocation::get(C);
if (LI->isSimple() && LI->hasOneUse() &&
           }
         }
 
         // We found an instruction that may write to the loaded memory.
         // We can try to promote at this position instead of the store
         // position if nothing alias the store memory after this and the store
         // destination is not in the range.
         if (P && P != SI) {
-          if (!moveUp(AA, SI, P))
+          if (!moveUp(AA, SI, P, LI))
             P = nullptr;
         }
 
         // If a valid insertion position is found, then we can promote
         // the load/store pair to a memcpy.
         if (P) {
           // If we load from memory that may alias the memory we store to,
           // memmove must be used to preserve semantic. If not, memcpy can

llvm/trunk/test/Transforms/MemCpyOpt/load-store-to-memcpy.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -basicaa -scoped-noalias -memcpyopt -S %s | FileCheck %s
+
+%T = type { i8, i32 }
+
+; memcpy(%d, %a) should not be generated since store2 may-aliases load %a.
+define void @f(%T* %a, %T* %b, %T* %c, %T* %d) {
+; CHECK-LABEL: @f(
+; CHECK-NEXT:    [[VAL:%.*]] = load %T, %T* %a, !alias.scope !0
+; CHECK-NEXT:    store %T { i8 23, i32 23 }, %T* %b, !alias.scope !3
+; CHECK-NEXT:    store %T { i8 44, i32 44 }, %T* %c, !alias.scope !6, !noalias !3
+; CHECK-NEXT:    store %T [[VAL]], %T* %d, !alias.scope !9, !noalias !12
+; CHECK-NEXT:    ret void
+;
+  %val = load %T, %T* %a, !alias.scope !{!10}
+
+  ; store1 may-aliases the load
+  store %T { i8 23, i32 23 }, %T* %b, !alias.scope !{!11}
+
+  ; store2 may-aliases the load and store3
+  store %T { i8 44, i32 44 }, %T* %c, !alias.scope !{!12}, !noalias !{!11}
+
+  ; store3
+  store %T %val, %T* %d, !alias.scope !{!13}, !noalias !{!10, !11}
+  ret void
+}
+
+!0 = !{!0}
+!1 = !{!1}
+!2 = !{!2}
+!3 = !{!3}
+
+!10 = !{ !10, !0 }
+!11 = !{ !11, !1 }
+!12 = !{ !12, !2 }
+!13 = !{ !13, !3 }