[llvm-commits] [llvm] r139765 - in /llvm/trunk: include/llvm/CodeGen/LiveInterval.h lib/CodeGen/RegisterCoalescer.cpp test/CodeGen/X86/2011-09-14-valcoalesce.ll

[Andrew Trick]

On Sep 14, 2011, at 7:46 PM, Jakob Stoklund Olesen wrote:

> 
> On Sep 14, 2011, at 6:09 PM, Andrew Trick wrote:
> 
>> Author: atrick
>> Date: Wed Sep 14 20:09:33 2011
>> New Revision: 139765
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=139765&view=rev
>> Log:
>> [regcoalescing] bug fix for RegistersDefinedFromSameValue.
>> 
>> +    /// getVNInfoBefore - Return the VNInfo that is live up to but not
>> +    /// necessarilly including Idx, or NULL. Use this to find the reaching def
>> +    /// used by an instruction at this SlotIndex position.
>> +    VNInfo *getVNInfoBefore(SlotIndex Idx) const {
>> +      const_iterator I = FindLiveRangeContaining(Idx.getPrevIndex());
>> +      return I == end() ? 0 : I->valno;
>> +    }
> 
> Nice!
> 
>> --- llvm/trunk/test/CodeGen/X86/2011-09-14-valcoalesce.ll (added)
>> +++ llvm/trunk/test/CodeGen/X86/2011-09-14-valcoalesce.ll Wed Sep 14 20:09:33 2011
>> @@ -0,0 +1,115 @@
>> +; RUN: llc < %s -march=x86
>> +;
>> +; Test RegistersDefinedFromSameValue. We have multiple copies of the same vreg:
>> +; while.body85.i:
>> +;   vreg1 = copy vreg2
>> +;   vreg2 = add
>> +; critical edge from land.lhs.true.i -> if.end117.i:
>> +;   vreg27 = vreg2
>> +; critical edge from land.lhs.true103.i -> if.end117.i:
>> +;   vreg27 = vreg2
>> +; if.then108.i:
>> +;   vreg27 = vreg1
>> +;
>> +; Prior to fixing PR10920 401.bzip miscompile, the coalescer would
>> +; consider vreg1 and vreg27 to be copies of the same value. It would
>> +; then remove one of the critical edge copes, which cannot safely be removed.
>> +
> 
> How would this test actually fail? The miscompilation didn't cause llc to crash, right?
> 
> /jakob
> 

Nope. And verification does catch it. We really need post-coalescing interference verification based on SSA values.

I struggled with how to add CHECK lines that weren't ridiculously fragile. Block's are "randomly" shuffled, copies "randomly" coalesced on different paths, and registers "randomly" renamed. I can try to add some though if you think it's worthwhile.

I figured at least the test provides an example of an interesting case to go along with the PR.

-Andy