[Bug target/67032] Geode optimizations incorrectly return -NaN

["ubizjak at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67032

UroÅ Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ra
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-07-27
                 CC|                            |ubizjak at gmail dot com,
                   |                            |vmakarov at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #6 from UroÅ Bizjak <ubizjak at gmail dot com> ---
Found the problem.

geode is one of the targets that enable Ym and Yn constraints:

(define_register_constraint "Ym"
 "TARGET_MMX && TARGET_INTER_UNIT_MOVES_TO_VEC ? MMX_REGS : NO_REGS"
 "@internal Any MMX register, when inter-unit moves to vector registers are
enabled.")

(define_register_constraint "Yn"
 "TARGET_MMX && TARGET_INTER_UNIT_MOVES_FROM_VEC ? MMX_REGS : NO_REGS"
 "@internal Any MMX register, when inter-unit moves from vector registers are
enabled.")

And ... the compiler generates insns that involve MMX registers.

$ ./cc1plus -O2 -march=geode -m32 -dp main.ii

$ grep %mm main.s
        movd    %eax, %mm0      # 131   *zero_extendsidi2/6     [length = 3]
        movq    %mm0, (%esp)    # 330   *movdi_internal/10      [length = 4]
        movq    %mm0, (%esp)    # 333   *movdi_internal/10      [length = 4]
        movd    %eax, %mm0      # 288   *zero_extendsidi2/6     [length = 3]

GCC uses many tricks to penalize moves to/from MMX registers. Actually, MMX
registers should not be allocated unless absolutely needed, that is the reason
that every single constraint involving MMX regs is decorated by * and most of
them are penalized by ? and/or !.

One of the most important tricks is the one in inline_secondary_memory_needed,
where:

  /* ??? This is a lie.  We do have moves between mmx/general, and for
     mmx/sse2.  But by saying we need secondary memory we discourage the
     register allocator from using the mmx registers unless needed.  */
  if (MMX_CLASS_P (class1) != MMX_CLASS_P (class2))
    return true;

It looks that with new and improved RA, this trick is not effective any more
and delicate balance that prevented MMX regs to be allocated is broken.

CC register allocator expert for opinion.