[PATCH] Refactor normalization for invokes in placestatepoints pass.

[Philip Reames]

REPOSITORY
  rL LLVM

================
Comment at: lib/Transforms/Scalar/PlaceSafepoints.cpp:698
@@ -697,3 @@
-                                nullptr);
-        assert(
-            !isa<PHINode>(cast<Instruction>(GCResult)->getParent()->begin()));
----------------
Leaving the assertion here would be good.  

================
Comment at: lib/Transforms/Scalar/PlaceSafepoints.cpp:939
@@ +938,3 @@
+    if (isa<PHINode>(toReplace->getNormalDest()->begin())) {
+      normalizeForInvokeSafepoint(toReplace->getNormalDest(),
+                                  toReplace->getParent());
----------------
Please just do this as a preprocessing step.  If the statepoint we're about to insert is for an invoke, normalize it before starting the insertion process.  Having this functionality inside a function called ReplaceWithStatepoint seems misleading.

In RewriteStatepointsForGC, we have:
  // When inserting gc.relocates for invokes, we need to be able to insert at
  // the top of the successor blocks.  See the comment on
  // normalForInvokeSafepoint on exactly what is needed.  Note that this step
  // may restructure the CFG.
  for (CallSite CS : toUpdate) {
    if (!CS.isInvoke())
      continue;
    InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
    normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
                                P);
    normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
                                P);
  }

Doing that same thing before calling ReplaceWithStatepoint would seem reasonable.

http://reviews.llvm.org/D9497

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/