Added a new "splitting" spiller.

When a call is placed to spill an interval this spiller will first try to
break the interval up into its component values. Single value intervals and
intervals which have already been split (or are the result of previous splits)
are spilled by the default spiller.

Splitting intervals as described above may improve the performance of generated
code in some circumstances. This work is experimental however, and it still
miscompiles many benchmarks. It's not recommended for general use yet.




git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@90951 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 2 insertions, 2 deletions

diff --git a/lib/CodeGen/LiveIntervalAnalysis.cpp b/lib/CodeGen/LiveIntervalAnalysis.cpp
index 35337ef..c16ddc5 100644
--- a/lib/CodeGen/LiveIntervalAnalysis.cpp
+++ b/lib/CodeGen/LiveIntervalAnalysis.cpp
@@ -401,7 +401,7 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb,
         interval.removeRange(Start, End);        
         assert(interval.ranges.size() == 1 &&
                "Newly discovered PHI interval has >1 ranges.");
-        MachineBasicBlock *killMBB = getMBBFromIndex(interval.endIndex());
+        MachineBasicBlock *killMBB = getMBBFromIndex(VNI->def);
         VNI->addKill(indexes_->getTerminatorGap(killMBB));
         VNI->setHasPHIKill(true);
         DEBUG({
@@ -412,7 +412,7 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb,
         // Replace the interval with one of a NEW value number.  Note that this
         // value number isn't actually defined by an instruction, weird huh? :)
         LiveRange LR(Start, End,
-                     interval.getNextValue(SlotIndex(getMBBStartIdx(mbb), true),
+                     interval.getNextValue(SlotIndex(getMBBStartIdx(Killer->getParent()), true),
                        0, false, VNInfoAllocator));
         LR.valno->setIsPHIDef(true);
         DEBUG(errs() << " replace range with " << LR);