Fix a bug in SCEV's isSafeToExpand around speculation safety

isSafeToExpand was making a common, but dangerously wrong, mistake in assuming that if any instruction within a basic block executes, that all instructions within that block must execute.  This can be trivially shown to be false by considering the following small example:
bb:
  add x, y  <-- InsertionPoint
  call @throws()
  udiv x, y <-- SCEV* S
  br ...

It's clearly not legal to expand S above the throwing call, but the previous logic would do so since S dominates (but not properlyDominates) the block containing the InsertionPoint.

Since iterating instructions w/in a block is expensive, this change special cases two cases: 1) S is an operand of InsertionPoint, and 2) InsertionPoint is the terminator of it's block.  These two together are enough to keep all current optimizations triggering while fixing the latent correctness issue.

As best I can tell, this is a silent bug in current ToT.  Given that, there's no tests with this change.  It was noticed in an upcoming optimization change (D60093), and was reviewed as part of that.  That change will include the test which caused me to notice the issue.  I'm submitting this seperately so that anyone bisecting a problem gets a clear explanation.

llvm-svn: 358680

1 files changed, 19 insertions, 1 deletions

diff --git a/llvm/lib/Analysis/ScalarEvolutionExpander.cpp b/llvm/lib/Analysis/ScalarEvolutionExpander.cpp
index 1469eae3b40..ccf031f7eae 100644
--- a/llvm/lib/Analysis/ScalarEvolutionExpander.cpp
+++ b/llvm/lib/Analysis/ScalarEvolutionExpander.cpp
@@ -2347,6 +2347,24 @@ bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
 
 bool isSafeToExpandAt(const SCEV *S, const Instruction *InsertionPoint,
                       ScalarEvolution &SE) {
-  return isSafeToExpand(S, SE) && SE.dominates(S, InsertionPoint->getParent());
+  if (!isSafeToExpand(S, SE))
+    return false;
+  // We have to prove that the expanded site of S dominates InsertionPoint.
+  // This is easy when not in the same block, but hard when S is an instruction
+  // to be expanded somewhere inside the same block as our insertion point.
+  // What we really need here is something analogous to an OrderedBasicBlock,
+  // but for the moment, we paper over the problem by handling two common and
+  // cheap to check cases.
+  if (SE.properlyDominates(S, InsertionPoint->getParent()))
+    return true;
+  if (SE.dominates(S, InsertionPoint->getParent())) {
+    if (InsertionPoint->getParent()->getTerminator() == InsertionPoint)
+      return true;
+    if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S))
+      for (const Value *V : InsertionPoint->operand_values())
+        if (V == U->getValue())
+          return true;
+  }
+  return false;
 }
 }