[pypy-dev] s390x jit backend the first results and status update

Richard Plangger planrichi at gmail.com
Thu Feb 11 12:17:46 EST 2016


Hi,

I have been looking quite elaborately to find the last remaining issues
in the JIT backend. I identified some problems (related to the ABI for
floats and ffi issues). Those are fixed.

There are two major left:

1) Bug that occurs very late in the translation of PyPy. As you are
aware I'm not able to reproduce it with the ssh access to my s390x
machine. It looks to me like a double free of a memory chunk, or a
miscalculation that only occurs when a lot of memory is allocated >4GB.
When I can my hands on a 8GB RAM machine I'm sure I can fix it easily.

2) Speed is currently faster than CPython on all benchmarks but probably
not on translation (which I could not run) and bm_mdp.
See here [1]. If you compare it to the results reported by Armin for PPC
& X86 ([2]) then I get around 50%-80% of the speedup.
I have had a look at a lot of profiles (generated with valgrind) but did
not find anything that is too slow. Would there be a difference when I
run the benchmark on real hardware?

Sadly not all tests on build bot are green. The ones that fail, fail
because some thread did not start in time. Or some lock was hold too
long. Eventually these tests will pass, but the suite just stops them. I
cannot reproduce that on my s390x VM either, because it seems this only
occurs under heavy load while running the test suite.

At the end I want to mention that it seems that there is a problem
related to the some special macros expanded by GCC (e.g. WCOREDUMP).
I have reported the problem to the glibc bug tracker [3]. I have fixed
the problem in pypy by altering the type provided to the macro, but
there might be more problems while compiling pypy.

Cheers,
Richard

[1] https://bpaste.net/show/d1af45831627
[2] http://morepypy.blogspot.co.at/2015/10/powerpc-backend-for-jit.html
[3] https://sourceware.org/bugzilla/show_bug.cgi?id=19613

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: OpenPGP digital signature
URL: <http://mail.python.org/pipermail/pypy-dev/attachments/20160211/aafc7575/attachment.sig>


More information about the pypy-dev mailing list