[llvm-commits] [llvm] r56326 - in /llvm/trunk: include/llvm/CodeGen/LiveInterval.h include/llvm/CodeGen/LiveIntervalAnalysis.h lib/CodeGen/LiveInterval.cpp lib/CodeGen/LiveIntervalAnalysis.cpp lib/CodeGen/RegAllocLinearScan.cpp

[Dale Johannesen]

On Sep 23, 2008, at 4:46 PMPDT, Dan Gohman wrote:
> On Sep 23, 2008, at 4:37 PM, Dale Johannesen wrote:
>> On Sep 23, 2008, at 4:11 PMPDT, Dan Gohman wrote:
>>
>>> Hi Dale,
>>>
>>> This commit caused a regression in
>>> test/CodeGen/PowerPC/2007-09-11-RegCoalescerAssert.ll
>>> in the "martini x86_64" nightly tester.
>>>
>>> It's an llc abort, and it's reproducible by running
>>> llc with the following explicit triple:
>>> -march=ppc64 -mtriple=powerpc-unknown-linux-gnu
>>>
>>> Can you investigate?
>>
>> It works correctly in my local llc, with the latest rewrite of this
>> stuff.
>> So I think the thing to do is finish that and check it in.
>> (Odd that it didn't crash when I ran this before checkin, though,  
>> only
>> obvious difference between linux & darwin is the output register
>> syntax,
>> and I don't see that causing a crash.  What does the crash look  
>> like?)
>
> I just double-checked the target triple. It passes with
> powerpc-apple-darwin and aborts with powerpc-unknown-linux-gnu.
> The abort looks like this:
>
> Assertion failed: (DefMI->getOpcode() ==
> TargetInstrInfo::IMPLICIT_DEF), function TurnCopyIntoImpDef, file /
> Users/gohman/LLVM/llvm-orange/lib/CodeGen/
> SimpleRegisterCoalescing.cpp, line 2239.

OK.  Weird.  Anyway, my next round of changes should fix it.
Thanks for pointing this out.