!invariant.loads with potentially clobbering calls

[reames]

Summary

A load from an invariant location is assumed to not alias any otherwise potentially aliasing stores. Our implementation only applied this rule to store instructions themselves whereas they it should apply for any memory accessing instruction. This results in both FRE and PRE becoming more effective at eliminating invariant loads.

Note that as a follow on change I will likely move this into AliasAnalysis itself. That's where the TBAA constant flag is handled and the semantics are essentially the same. I'd like to separate the semantic change from the refactoring and thus have extended the hack that's already in MemoryDependenceAnalysis for this change.

[• dberlin]

Looks right to me

llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp

getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
   // being 42. A key property of this program however is that if either
   // 1 or 4 were missing, there would be a race between the store of 42
   // either the store of 0 or the load (making the whole progam racy).
   // The paper mentionned above shows that the same property is respected
   // by every program that can detect any optimisation of that kind: either
   // it is racy (undefined) or there is a release followed by an acquire
   // between the pair of accesses under consideration.
 
+  // If the load is invariant, we "know" that it doesn't alias *any* write. We
+  // do want to respect mustalias results since defs are useful for value
+  // forwarding, but any mayalias write can be assumed to be noalias.
+  // Arguably, this logic should be pushed inside AliasAnalysis itself.
   if (isLoad && QueryInst) {
     LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
     if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
       isInvariantLoad = true;
   }
 
   const DataLayout &DL = BB->getModule()->getDataLayout();
 
#endif
     // the malloced memory between the malloc call and its bitcast uses, and we
     // need to continue scanning until the malloc call.
     const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
     if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
       const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
 
       if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
         return MemDepResult::getDef(Inst);
+      if (isInvariantLoad)
+        continue;
       // Be conservative if the accessed pointer may alias the allocation.
       if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
         return MemDepResult::getClobber(Inst);
       // If the allocation is not aliased and does not read memory (like
       // strdup), it is safe to ignore.
       if (isa<AllocaInst>(Inst) ||
           isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
         continue;
     }
 
+    if (isInvariantLoad)
+       continue;
+
     // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
     AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
     // If necessary, perform additional analysis.
     if (MR == AliasAnalysis::ModRef)
       MR = AA->callCapturesBefore(Inst, MemLoc, DT);
     switch (MR) {
     case AliasAnalysis::NoModRef:
       // If the call has no effect on the queried pointer, just ignore it.

llvm/trunk/test/Transforms/GVN/invariant-load.ll

bb1:
   br label %bb2
 
 bb2:
   %v2 = load i32, i32* %p
   %res = sub i32 %v1, %v2
   ret i32 %res
 }
 
+; Checks that we return the mustalias store as a def
+; so that it contributes to value forwarding.  Note
+; that we could and should remove the store too.
+define i32 @test5(i1 %cnd, i32* %p) {
+; CHECK-LABEL: test5
+; CHECK-LABEL: entry:
+; CHECK-NEXT: store i32 5, i32* %p
+; CHECK-NEXT: ret i32 5
+entry:
+  %v1 = load i32, i32* %p, !invariant.load !0
+  store i32 5, i32* %p ;; must alias store, want to exploit
+  %v2 = load i32, i32* %p, !invariant.load !0
+  ret i32 %v2
+}
+
+
+declare void @foo()
+
+; Clobbering (mayalias) stores, even in function calls, can be ignored
+define i32 @test6(i1 %cnd, i32* %p) {
+; CHECK-LABEL: test6
+; CHECK-LABEL: entry:
+; CHECK-NEXT: @foo
+; CHECK-NEXT: ret i32 0
+entry:
+  %v1 = load i32, i32* %p, !invariant.load !0
+  call void @foo()
+  %v2 = load i32, i32* %p, !invariant.load !0
+  %res = sub i32 %v1, %v2
+  ret i32 %res
+}
+
+declare noalias i32* @bar(...) 
+
+; Same as previous, but a function with a noalias result (since they're handled
+; differently in MDA)
+define i32 @test7(i1 %cnd, i32* %p) {
+; CHECK-LABEL: test7
+; CHECK-LABEL: entry:
+; CHECK-NEXT: @bar
+; CHECK-NEXT: ret i32 0
+entry:
+  %v1 = load i32, i32* %p, !invariant.load !0
+  call i32* (...)* @bar(i32* %p)
+  %v2 = load i32, i32* %p, !invariant.load !0
+  %res = sub i32 %v1, %v2
+  ret i32 %res
+}
+
+
 !0 = !{ }