[llvm-testresults] buildbot failure in llvm on llvm-arm-linux

[Stuart Hastings]

On Apr 5, 2011, at 10:13 AM, Jakob Stoklund Olesen wrote:

> 
> On Apr 5, 2011, at 10:08 AM, Stuart Hastings wrote:
> 
>> 
>> On Apr 5, 2011, at 10:03 AM, Eric Christopher wrote:
>> 
>>> 
>>> On Apr 5, 2011, at 10:02 AM, Jakob Stoklund Olesen wrote:
>>> 
>>>> 
>>>> On Apr 5, 2011, at 9:54 AM, Devang Patel wrote:
>>>> 
>>>>> This builder is now failing a test (segmentation fault). Is anyone looking at it ?
>>>> 
>>>> Reproduce like this:
>>>> 
>>>> $ Release+Asserts/bin/llc -march=arm < /d/g/llvm/test/CodeGen/Generic/2010-11-04-BigByval.ll
>>>> 	.syntax unified
>>>> Segmentation fault
>>>> 
>>>> It is crashing during ARM isel. Did anybody change byval lowering for ARM?
>>> 
>>> I think Stuart removed the patch that did byval lowering?
>> 
>> I removed a patch that permitted nested calls in the legalizer.  This is necessary to support memcpy() on a byval parameter to another function.  However, AFAIK, no FE was generating byval parameters for ARM.
>> 
>> Actually, I'm surprised this ever worked for ARM.  ARM byval doesn't really exist outside of a patch that I've never committed.
>> 
>> I'm guessing this passed on ARM because the ARM BE currently ignores byval.  IOW, the test should avoid ARM until byval support appears.
>> 
>> Perhaps the test could exclude ARM (and other targets that don't support byval)?
> 
> If ARM doesn't support byval, it makes sense to XFAIL that test for arm.
> 
> We should fail with an assertion instead of a segfault, though.

Agreed.  I'll XFAIL the test, but I don't have time to chase down the segfault right now.  I agree the segfault is bad form, but it should go away when I finally add ARM byval support.  Additionally, AFAIK no FE generates byval for ARM right now.

If somebody else could chase down the segfault, I'd be very grateful.

stuart