gmp-4.2 check failure with experimental gcc

[Torbjorn Granlund]

Martin Reinecke <martin at MPA-Garching.MPG.DE> writes:

  when gmp-4.2 is compiled with the current experimental version of gcc 4.2
  on a Pentium 4 (pentium4-pc-linux-gnu/i686-pc-linux-gnu),
  the test t-mul in the mpz/ directory fails:
  
  /scratch/gmp-4.2/tests/mpz>./t-mul
  ERROR: incorrect plain product in test 53
  op1          = -FFFFFFFFFFFFFFFFFFFFFFFFF...0000000000000000000000000 [71992]
  op2          = 1FFFFFFFFFFFFFFFFFFFFFFFF...FFFFFFFFFFFFFFFFFFFFFFFFF [202386]
       product  = -61EDD19F23160438880BD99C5...-6701F030DAFAEAD82176BAB51 [274384]
  ref_product  = -1FFFFFFFFFFFFFFFFFFFFFFFF...0000000000000000000000000 [274378]
  Abort
  
  GMP was configured and compiled with a simple "./configure; make; make check"
  This problem does not occur with gcc 4.1 and the head of the gcc 4.1-branch.
  
  It appears to be related to processor-specific optimisation; the minimum set
  of optimisation flags which trigger the bug appears to be "-O1 -march=pentium4".
  
  Since it is very likely that this is a bug in gcc and not in GMP, I filed a bug report
  there (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27322).
  On the other hand, the gcc developers would like to see a standalone testcase
  that demonstrates the bug, which is very hard for me to provide, since I'm not
  familiar with GMP at all. Could you please give me a hint from which GMP files
  this particular test case depends, so that I can try to build a single-file
  testcase? Or do you have an idea how this could be done more easily?

This would quite likely be mpn/generic/mul_fft.c.  This complex
file is often miscompiled.

-- 
Torbjörn