Bug 750811

I accidentally added my notes on this to bug 757021, rather than here, so I'll repeat the information here.

Attachment 541076 [details] is a debugging patch to capture the extent of the problem.  It hacks up test_cmath so that test_specific_values shows all failing cases (and ultimately "succeeds"), rather than aborting on the first failed case, so that we can see the extent of the problem.

Attachment 541077 [details] is the log from running "python -m test.test_cmath" with the patch.  Summarizing that log, and comparing against cmath_testcases.txt:

  * 3 of the 20 tests for "acos() near infinity" fail due to the imaginary
    component of the result being unexpectedly either "+/-inf"

  * 4 of the 20 tests for "acosh() near infinity" fail due to the real
    component of the result being unexpectedly "inf"

  * 4 of the 20 tests for "asin() near infinity" fail due to the imaginary
    component of the result being unexpectedly either "+/-inf"

  * 4 of the 20 tests for "asinh() near infinity" fail due to the real
    component of the result being unexpectedly either "+/-inf"

  * 4 of the 20 tests for "atan() near infinity" fail due to the imaginary
    component of the result being unexpectedly either "+/-0.0"

  * 4 of the 20 tests for "atanh() near infinity" fail due to the real
    component of the result being unexpectedly either "+/-0.0"

  * 4 of the 20 tests for "log() near infinity" fail due to the real
    component of the result being unexpectedly either "inf"

  * 4 of the 20 tests for "log10() near infinity" fail due to the real
    component of the result being unexpectedly either "inf"

  * both of the 2 tests for "sqrt() with input whose abs value overflows" fail,
    one expectedly collapsing to (inf, -0.0), the other to (0.0, -inf)

This is out of 1885 tests within test_specific_values (__main__.CMathTests)

Attachment 541092 [details] is a crude patch to log all calls to "hypot (3)" from Modules/cmathmodule.c

Attachment 541093 [details] is a log from running "python -m test.test_cmath" with that patch, and shows the calls to hypot() interleaved with the cmath test cases.

Python's "complex" type uses a pair of C "double" values internally (real and imag), which on ppc64 should be standard IEEE754.

ppc64's glibc has an optimized implementation of "hypot" (in libm):
  sysdeps/powerpc/fpu/e_hypot.c

All of the failing Python test cases involve a call to hypot().

When examining the results in comment #2, a comparison of calls to hypot with the failing test cases, reveals that all of the failing cases are preceded by a call to hypot that returns "inf".

Python's math.hypot() is a thin wrapper around hypot(3); a Python "float" uses a C "double" internally.

Attachment 541097 [details] is a Python script to exercise hypot(3).  It calls hypot() for all of the x,y pairs that fail within the Python test case on ppc64.

On x86_64, with glibc-2.14-5.x86_64 (and python-2.7.1-7.fc15.x86_64), all results are non-infinite, with exponent e+307 or e+308 (see attachment 541102 [details] for the results on this box).

On ppc64, with glibc-2.14.90-19.ppc64 (and python-2.7.2-4.2.fc16.ppc64), all results are "inf".

Filed in upstream bug tracker for CPython as: http://bugs.python.org/issue13534

Filed in glibc's bug tracker as: http://sourceware.org/bugzilla/show_bug.cgi?id=13472

as there's already a upstream fix available, can you add that to the f16 glibc git branch, please ? It affects only powerpc, so an update on the primary archs won't be necessary, we just need the git commit.

I was hoping to start to throttle back the updates now that I'm starting to get a handle on things in glibc-land.  There's a small handful that I've got queued for the next update, I'll make sure this one gets added.

glibc-2.14.90-24.fc16.4 has been submitted as an update for Fedora 16.
https://admin.fedoraproject.org/updates/glibc-2.14.90-24.fc16.4

glibc-2.14.90-24.fc16.4 has been pushed to the Fedora 16 stable repository.  If problems still persist, please make note of it in this bug report.