[Mesa-dev] [PIGLIT,radeonsi] crash in spec/glsl-1.50/execution/geometry/max-input-components fixed (was: crash in spec/glsl-1.50/execution/geometry/max-input-components – who's bug is it?)

[Kai Wasserbäch]

Hi there,
Tom Stellard schrieb am 16.04.2014 17:07:
> On Wed, Apr 16, 2014 at 02:36:19PM +0200, Kai Wasserbäch wrote:
>> Michel Dänzer schrieb am 15.04.2014 09:27:
>>> On 23.03.2014 04:53, Kai Wasserbäch wrote:
>>>> Dear Mesa devs,
>>>> I'm not sure whether this is a bug in Mesa, LLVM or in eglibc. The crash happens
>>>> in _int_malloc, but since that is certainly one of the more often used
>>>> functions, I'm not yet convinced, the fault lies indeed with eglibc.
>>>>
>>>> Therefore I'm attaching the full backtrace of the crash in
>>>> spec/glsl-1.50/execution/geometry/max-input-components (it takes a very long
>>>> time until the crash actually happens, Piglit recorded an execution time of
>>>> 1538.0506579875946) and hope you can point me to the right bug tracker.
>>>>
>>>> I'm unable to tell, whether this is a regression or not, since today was the
>>>> first time I was able to run a full Piglit quick test, without crashing my X on
>>>> this machine with the radeonsi.
>>>
>>> It's not a regression. If you build LLVM with assertions enabled, you get:
>>>
>>> shader_runner:
>>> /home/daenzer/src/llvm-git/llvm/lib/CodeGen/RegAllocGreedy.cpp:2268:
>>> unsigned int
>>> {anonymous}::RAGreedy::selectOrSplitImpl(llvm::LiveInterval&,
>>> llvm::SmallVectorImpl<unsigned int>&,
>>> {anonymous}::RAGreedy::SmallVirtRegSet&, unsigned int): Assertion
>>> `NewVRegs.empty() && "Cannot append to existing NewVRegs"' failed.
>>>
>>> So this is an LLVM issue. It might be worth testing if Tom's register
>>> spilling patches help.
>>
>> Can you point me to those patches? Preferrably as a branch, but ML is ok as well.
>>
> 
> Here is the branch: http://cgit.freedesktop.org/~tstellar/llvm/log/?h=si-spill-fixes

I've just rerun the Piglit test
(spec/glsl-1.50/execution/geometry/max-input-components) with the patches from
the si-spill-fixes branch (I needed to massage them a bit into applying on top
of LLVM's SVN revision 206583, but that was straight forward enough) and can
report, that the crash is gone. The test (still) fails, however:

$ gdb --args <PIGLIT-DIR>/build/bin/shader_runner
<PIGLIT-DIR>/local.git/tests/spec/glsl-1.50/execution/geometry/max-input-components.shader_test
-auto
[GNU GDB boilerplate]
Reading symbols from <PIGLIT-DIR>/build/bin/shader_runner...done.
(gdb) r
Starting program: <PIGLIT-DIR>/build/bin/shader_runner
<PIGLIT-DIR>/local.git/tests/spec/glsl-1.50/execution/geometry/max-input-components.shader_test
-auto
warning: Could not load shared library symbols for linux-vdso.so.1.
Do you need "set solib-search-path" or "set sysroot"?
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff0081700 (LWP 24909)]
Probe color at (0,0)
  Expected: 0.000000 1.000000 0.000000 1.000000
  Observed: 0.000000 0.000000 0.000000 0.000000
PIGLIT: {'result': 'fail' }
[Thread 0x7ffff0081700 (LWP 24909) exited]
[Inferior 1 (process 24905) exited with code 01]


The stack used for this test is a bit farther ahead as well, in case that makes
a difference, I'm detailing it below:
GPU: "PITCAIRN" (ChipID = 0x6819)
Linux: 3.14.0
libdrm: Git:master/libdrm-2.4.53
LLVM: SVN:trunk/r206583
libclc: Git:master/1e278a7b04
Mesa: Git:master/352e06ddea
GLAMOR: Git:master/a4fbc7732a (Standalone)
DDX: Git:master/ea6d0affe5
X: 2:1.15.0.901-1

Thank you for pointing me in the right direction; if you should need me to run
another test, please let me know.

Cheers,
Kai

P.S.: Any idea, when the si-spill-fixes branch is going to land upstream?



-- 

Kai Wasserbäch (Kai Wasserbaech)

E-Mail: kai at dev.carbon-project.org

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 648 bytes
Desc: OpenPGP digital signature
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20140418/308170fc/attachment.sig>