[Bug target/57057] New: Bad optimization on x86 for minps and maxps

[Bug target/57057] New: Bad optimization on x86 for minps and maxps

[mathias at gaunard dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57057

             Bug #: 57057
           Summary: Bad optimization on x86 for minps and maxps
    Classification: Unclassified
           Product: gcc
           Version: 4.7.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: mathias@gaunard.com


When optimization -O1 or greater is enabled, minps and maxps will incorrectly
deal with nan values.
icc gives the correct result.

See the attached C testcase, with -O0 it displays the correct result:
res1 = (1.000000 1.000000 1.000000 1.000000)
res2 = (-nan -nan -nan -nan)
res3 = (1.000000 1.000000 1.000000 1.000000)
res4 = (-nan -nan -nan -nan)

With -O1 it displays
res1 = (1.000000 1.000000 1.000000 1.000000)
res2 = (1.000000 1.000000 1.000000 1.000000)
res3 = (-nan -nan -nan -nan)
res4 = (-nan -nan -nan -nan)

The same also happens with the double-precision and AVX variants.

[Bug target/57057] Bad optimization on x86 for minps and maxps

[mathias at gaunard dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57057

--- Comment #1 from Mathias Gaunard <mathias at gaunard dot com> 2013-04-24 16:42:23 UTC ---
Created attachment 29931
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29931
bad optimisation testcase

[Bug target/57057] Bad optimization on x86 for minps and maxps

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57057

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Target|                            |x86_64-*-*, i?86-*-*
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-04-25
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-25 08:14:23 UTC ---
That it's "fixed" with -O1 is because the different MINPS invocations are
CSEd and only the first one is carried out.

The issue here is that MINPS is not commutative, but the SMIN RTX code
is specified as being commutative.  Which means with !flag_finite_math_only
the x86 backend has to use an UNSPEC to expand __builtin_ia32_minps and
friends.

[Bug target/57057] Bad optimization on x86 for minps and maxps

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57057

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |DUPLICATE
             Status|NEW                         |RESOLVED

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
This was fixed for GCC 6.3+ and is a dup of bug 72867.

*** This bug has been marked as a duplicate of bug 72867 ***