LICM: Don't sink stores out of loops that may throw.

[eli.friedman]

Summary

This hasn't been caught before because it requires noalias or similarly
strong alias analysis to actually reproduce.

Fixes http://llvm.org/PR27952 .

[hfinkel]

lib/Transforms/Scalar/LICM.cpp

936 ↗

(On Diff #59485)

The assumption here is that stores to allocas much be unobservable after the function unwinds, and I agree with that. However, what if the exception is caught in this function. In that case, such a store might still be observable.

Also, it would be friendlier to call GetUnderlyingObjects and check whether any are not allocas (if we can determine that there are no relevant landingpads?).

[eli.friedman]

lib/Transforms/Scalar/LICM.cpp

936 ↗

(On Diff #59485)

The tricky case here is specifically call instructions, which unwind without an explicit unwind edge . If there's an invoke instruction in a loop, the unwind edge is visible to the optimizer; we can treat it the same way as any other loop exit.

Changing this to use GetUnderlyingObjects seems fine.

[eli.friedman]

lib/Transforms/Scalar/LICM.cpp

936 ↗

(On Diff #59485)

Hmm... actually, just tried GetUnderlyingObjects; can't get it to do anything useful because alias analysis can't figure out that x ? &a2 : &a2 doesn't alias an arbitrary function call.

[hfinkel]

lib/Transforms/Scalar/LICM.cpp

936 ↗

(On Diff #59485)

Okay, we should fix that too :( - This LGTM.

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

bool llvm::promoteLoopAccessesToScalars(
   // make sure we catch that. An additional load may be generated in the
   // preheader for SSA updater, so also avoid sinking when no preheader
   // is available.
   if (!HasDedicatedExits || !Preheader)
     return false;
 
   const DataLayout &MDL = Preheader->getModule()->getDataLayout();
 
+  if (SafetyInfo->MayThrow) {
+    // If a loop can throw, we have to insert a store along each unwind edge.
+    // That said, we can't actually make the unwind edge explicit. Therefore,
+    // we have to prove that the store is dead along the unwind edge.
+    //
+    // Currently, this code just special-cases alloca instructions.
+    if (!isa<AllocaInst>(GetUnderlyingObject(SomePtr, MDL)))
+      return false;
+  }
+
   // Check that all of the pointers in the alias set have the same type.  We
   // cannot (yet) promote a memory location that is loaded and stored in
   // different sizes.  While we are at it, collect alignment and AA info.
   bool Changed = false;
   for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
     Value *ASIV = ASI->getValue();
     PointerMustAliases.insert(ASIV);
 

llvm/trunk/test/Transforms/LICM/scalar_promote-unwind.ll

+; RUN: opt < %s -basicaa -licm -S | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Make sure we don't hoist the store out of the loop; %a would
+; have the wrong value if f() unwinds
+
+define void @test1(i32* nocapture noalias %a, i1 zeroext %y) uwtable {
+entry:
+  br label %for.body
+
+for.body:
+  %i.03 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]
+  %0 = load i32, i32* %a, align 4
+  %add = add nsw i32 %0, 1
+  store i32 %add, i32* %a, align 4
+  br i1 %y, label %if.then, label %for.inc
+
+; CHECK: define void @test1
+; CHECK: load i32, i32*
+; CHECK-NEXT: add
+; CHECK-NEXT: store i32
+
+if.then:
+  tail call void @f()
+  br label %for.inc
+
+for.inc:
+  %inc = add nuw nsw i32 %i.03, 1
+  %exitcond = icmp eq i32 %inc, 10000
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+
+for.cond.cleanup:
+  ret void
+}
+
+; We can hoist the store out of the loop here; if f() unwinds,
+; the lifetime of %a ends.
+
+define void @test2(i1 zeroext %y) uwtable {
+entry:
+  %a = alloca i32
+  br label %for.body
+
+for.body:
+  %i.03 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]
+  %0 = load i32, i32* %a, align 4
+  %add = add nsw i32 %0, 1
+  store i32 %add, i32* %a, align 4
+  br i1 %y, label %if.then, label %for.inc
+
+if.then:
+  tail call void @f()
+  br label %for.inc
+
+for.inc:
+  %inc = add nuw nsw i32 %i.03, 1
+  %exitcond = icmp eq i32 %inc, 10000
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+
+for.cond.cleanup:
+  ret void
+
+; CHECK: define void @test2
+; CHECK: store i32
+; CHECK-NEXT: ret void
+  ret void
+}
+
+declare void @f() uwtable