Re: optimization/8156: -mcpu=pentium4 generates wrong fp results in glibc-2.3

[Andreas Jaeger <aj@suse.de>]

The following reply was made to PR optimization/8156; it has been noted by GNATS.

From: Andreas Jaeger <aj@suse.de>
To: bregor@sf.anu.edu.au
Cc: gcc-gnats@gcc.gnu.org
Subject: Re: optimization/8156: -mcpu=pentium4 generates wrong fp results in
 glibc-2.3
Date: Mon, 07 Oct 2002 06:45:15 +0200

 bregor@sf.anu.edu.au writes:
 
 >>Number:         8156
 >>Category:       optimization
 >>Synopsis:       -mcpu=pentium4 generates wrong fp results in glibc-2.3
 >>Confidential:   no
 >>Severity:       serious
 >>Priority:       medium
 >>Responsible:    unassigned
 >>State:          open
 >>Class:          wrong-code
 >>Submitter-Id:   net
 >>Arrival-Date:   Sun Oct 06 20:06:01 PDT 2002
 >>Closed-Date:
 >>Last-Modified:
 >>Originator:     Roger Brown
 >>Release:        gcc (GCC) 3.2.1 20021004 (prerelease)
 >>Organization:
 >>Environment:
 > Linux-2.4.19-pre10
 > HJL's binutils-2.13.90.0.4 
 > glibc-2.3 release source code
 >>Description:
 > *** This may be a duplicate report but the original seems
 >     to have got lost.
 >
 >   After building the released source for glibc-2.3,
 >   with "gcc (GCC) 3.2.1 20021004 (prerelease)",  and the
 >   -mcpu=pentium4 -march=pentium4 switches, make check
 >  discovered many serious errors for the floating-point
 >  special functions (see below).  The same build, with
 >   -mcpu=i686 -march=i686 failed with an assembler error.
 
 Can you send information about the assembler error?  Do you have a
 small testcase for this?
 
 >   With -mcpu and -march left at the default values, both
 >   the build and check are clean.
 >
 >   Other software details: O/S linux-2.4.19-pre10 and
 >   HJL's Linux-binutils-2.13.90.0.4.
 >
 >   The actual "check" failure was:
 >
 > GCONV PATH=/Main/Glibc/glibc-2.3/Build/iconvdata LC ALL=C
 > /Main/Glibc/glibc-2.3/Build/elf/ld-linux.so.2 --library-path
 > /Main/Glibc/glibc-2.3/Build:/Main/Glibc/glibc-2.3/Build/math:
 > /Main/Glibc/glibc-2.3/Build/elf:/Main/Glibc/glibc-2.3/Build/dlfcn:
 > /Main/Glibc/glibc-2.3/Build/nss:/Main/Glibc/glibc-2.3/Build/nis:
 > /Main/Glibc/glibc-2.3/Build/rt:/Main/Glibc/glibc-2.3/Build/resolv:
 > /Main/Glibc/glibc-2.3/Build/crypt:/Main/Glibc/glibc-2.3/Build/linuxthreads
 > /Main/Glibc/glibc-2.3/Build/math/test-idouble  >
 > /Main/Glibc/glibc-2.3/Build/math/test-idouble.out
 > make[2]: *** [/Main/Glibc/glibc-2.3/Build/math/test-idouble.out] Error 139
 
 A segmentation fault as result of test-idouble.out?  That's strange.
 
 >  The errors reported in test-idouble.out are alarming !
 >  The first part reads:
 >
 > testing double (without inline functions)
 > Failure: Test: cosh (0.75) == 1.29468328467684468784170818539018176
 > Result:
 >  is:          2.11700001661267478426e+00   0x1.0ef9db467dcf80000000p+1
 >  should be:   1.29468328467684479222e+00   0x1.4b705d1e5d6a80000000p+0
 >  difference:  8.22316731935829992040e-01   0x1.a506b2dd3c6900000000p-1
 >  ulp       :  3703385327526728.0000
 >  max.ulp   :  0.0000
 
 Can you nail down what goes wrong?  Is cosh miscompiled (the source is
 in sysdeps/ieee754/dbl-64/e_cosh.c) or one of the functions it's
 calling?
 
 > Maximal error of `cosh'
 >  is      : 3703385327526728 ulp
 >  accepted: 0 ulp
 > Failure: sinh (-inf) == -inf: Exception "Invalid operation" set
 > Failure: Test: sinh (-inf) == -inf
 > Result:
 >  is:          nan   nan
 >  should be:  -inf  -inf
 > Failure: Test: sinh (0.75) == 0.822316731935829980703661634446913849
 >
 >   and it gets much worse, much worse !!
 >
 >  Specific *.i files etc can be provided.
 
 Andreas
 -- 
  Andreas Jaeger
   SuSE Labs aj@suse.de
    private aj@arthur.inka.de
     http://www.suse.de/~aj