Bug 56125 - [4.7 Regression] -O2 -ffast-math generates bad code when dividing a double by the square of another double.
Summary: [4.7 Regression] -O2 -ffast-math generates bad code when dividing a double by...
Status: RESOLVED FIXED
Alias: None
Product: gcc
Classification: Unclassified
Component: tree-optimization (show other bugs)
Version: 4.7.2
: P3 normal
Target Milestone: 4.7.3
Assignee: Jakub Jelinek
URL:
Keywords: wrong-code
Depends on:
Blocks:
 
Reported: 2013-01-28 02:58 UTC by mednafen
Modified: 2013-02-01 14:38 UTC (History)
2 users (show)

See Also:
Host:
Target:
Build:
Known to work:
Known to fail:
Last reconfirmed: 2013-01-28 00:00:00


Attachments
Test program. (158 bytes, text/x-csrc)
2013-01-28 02:58 UTC, mednafen
Details
gcc48-pr56125.patch (1.23 KB, patch)
2013-01-28 10:27 UTC, Jakub Jelinek
Details | Diff

Note You need to log in before you can comment on or make changes to this bug.
Description mednafen 2013-01-28 02:58:07 UTC
Created attachment 29288 [details]
Test program.

XXX@willow:~$ gcc -O0 -o weird weird.c 
XXX@willow:~$ ./weird 
7.250000

XXX@willow:~$ gcc -O2 -o weird weird.c 
XXX@willow:~$ ./weird 
7.250000

XXX@willow:~$ gcc -ffast-math -O2 -o weird weird.c 
XXX@willow:~$ ./weird 
5.126524



Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/4.7/lto-wrapper
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Debian 4.7.2-5' --with-bugurl=file:///usr/share/doc/gcc-4.7/README.Bugs --enable-languages=c,c++,go,fortran,objc,obj-c++ --prefix=/usr --program-suffix=-4.7 --enable-shared --enable-linker-build-id --with-system-zlib --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --with-gxx-include-dir=/usr/include/c++/4.7 --libdir=/usr/lib --enable-nls --with-sysroot=/ --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --enable-gnu-unique-object --enable-plugin --enable-objc-gc --with-arch-32=i586 --with-tune=generic --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 4.7.2 (Debian 4.7.2-5)
Comment 1 Marc Glisse 2013-01-28 07:23:24 UTC
It is turned to x*pow(n,-2.) and the expansion of pow notices that 2*2 is an integer and uses the code meant for 2.5...
Comment 2 Marc Glisse 2013-01-28 07:30:05 UTC
4.6 generated a very suboptimal call to the libm pow function, but it didn't have the completely wrong result.
Comment 3 Jakub Jelinek 2013-01-28 09:15:09 UTC
Regressed with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=174446
Looking into it.
Comment 4 Jakub Jelinek 2013-01-28 10:27:19 UTC
Created attachment 29292 [details]
gcc48-pr56125.patch

The bug is that the last two optimizations of pow (where 2c resp. 3c is a non-zero integer) silently assume that the earlier optimizations already handled the cases where already c (or 2c for the last optimization) is an integer.
But that doesn't have to be the case, as shown by the testcase, the integer optimization is guarded by c in [-1,2] or optimization for speed.

The && optimize_function_for_speed_p () (or should that be optimize_insn_for_speed_p?, the pow folding is inconsistent in that, and I don't see e.g. rtl_profile_for_bb being called during this pass to make it accurate)
is up for discussions, say on x86_64
__attribute__((cold)) double
foo (double x, double n)
{
  double u = __builtin_pow (x, -1.5);
  return u;
}
with it we get smaller code:
        movsd   .LC0(%rip), %xmm1
        jmp     pow
compared to:
        sqrtsd  %xmm0, %xmm1
        mulsd   %xmm0, %xmm1
        movsd   .LC0(%rip), %xmm0
        divsd   %xmm1, %xmm0
without it, 7 bytes shorter.
Comment 5 Jakub Jelinek 2013-01-28 14:43:07 UTC
Author: jakub
Date: Mon Jan 28 14:43:03 2013
New Revision: 195507

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195507
Log:
	PR tree-optimization/56125
	* tree-ssa-math-opts.c (gimple_expand_builtin_pow): Don't optimize
	pow(x,c) into sqrt(x) * powi(x, n/2) or
	1.0 / (sqrt(x) * powi(x, abs(n/2))) if c is an integer or when
	optimizing for size.
	Don't optimize pow(x,c) into powi(x, n/3) * powi(cbrt(x), n%3) or
	1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)) if 2c is an
	integer.

	* gcc.dg/pr56125.c: New test.

Added:
    trunk/gcc/testsuite/gcc.dg/pr56125.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-math-opts.c
Comment 6 Jakub Jelinek 2013-01-28 15:07:24 UTC
Fixed on the trunk so far.
Comment 7 Jakub Jelinek 2013-02-01 14:17:58 UTC
Author: jakub
Date: Fri Feb  1 14:17:50 2013
New Revision: 195664

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195664
Log:
	Backported from mainline
	2013-01-28  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/56125
	* tree-ssa-math-opts.c (gimple_expand_builtin_pow): Don't optimize
	pow(x,c) into sqrt(x) * powi(x, n/2) or
	1.0 / (sqrt(x) * powi(x, abs(n/2))) if c is an integer or when
	optimizing for size.
	Don't optimize pow(x,c) into powi(x, n/3) * powi(cbrt(x), n%3) or
	1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)) if 2c is an
	integer.

	* gcc.dg/pr56125.c: New test.

Added:
    branches/gcc-4_7-branch/gcc/testsuite/gcc.dg/pr56125.c
Modified:
    branches/gcc-4_7-branch/gcc/ChangeLog
    branches/gcc-4_7-branch/gcc/testsuite/ChangeLog
    branches/gcc-4_7-branch/gcc/tree-ssa-math-opts.c
Comment 8 Jakub Jelinek 2013-02-01 14:38:51 UTC
Fixed.