[PATCH] D20739: Better fix for PR27903.

[Than McIntosh via llvm-commits]

thanm added inline comments.

================
Comment at: lib/CodeGen/StackColoring.cpp:563
@@ +562,3 @@
+  for (unsigned slot = 0; slot < NumSlot; ++slot)
+    if (NumStartLifetimes[slot] > 1)
+      ConservativeSlots.set(slot);
----------------
wmi wrote:
> It may be a stupid question. I am wondering is it possible for one slot to have one Lifetime_start and two Livetime_end? Could it be more safe to use NumEndLifetimes[] > 1 as the condition to fall back?
I think you're right, this could come up. I will post a revised patch that checks both start and end counts.

I experimented a bit with trying to assert in cases where there is a mismatch (end with no start, for example) but eventually gave up (there are number of tests written with imbalanced lifetime markers already).



http://reviews.llvm.org/D20739