[Bug tree-optimization/49583] New: Reloading stack operands in the wrong order, so needs to insert fxch

[Bug tree-optimization/49583] New: Reloading stack operands in the wrong order, so needs to insert fxch

[sgunderson at bigfoot dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49583

           Summary: Reloading stack operands in the wrong order, so needs
                    to insert fxch
           Product: gcc
           Version: 4.6.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: sgunderson@bigfoot.com


Created attachment 24638
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24638
Minimal testcase

Hi,

It seems that when generating x87 code, GCC sometimes reloads items from the
stack in the wrong order, and then goes to great lengths to swap them around. I
have an example with six loads and six fxch instructions, but attached is a
minimal example. Compiling with "gcc version 4.6.1 (Debian 4.6.1-1)" as
follows:

pannekake:~> gcc-4.6 -m32 -Wall -O2 -march=pentium3 -c fxch.c

The odd sequence is around this:

  41:    d9 44 24 48              flds   0x48(%esp)
  45:    dd 5c 24 08              fstpl  0x8(%esp)
  49:    dd 14 24                 fstl   (%esp)
  4c:    d9 5c 24 10              fstps  0x10(%esp)
  50:    e8 fc ff ff ff           call   51 <process+0x51>
  55:    d9 5c 24 1c              fstps  0x1c(%esp)
  59:    d9 44 24 1c              flds   0x1c(%esp)
  5d:    d9 44 24 10              flds   0x10(%esp)
  61:    d9 c9                    fxch   %st(1)
  63:    d9 1c b7                 fstps  (%edi,%esi,4)
  66:    46                       inc    %esi
  67:    39 ee                    cmp    %ebp,%esi

In particular, why did it use fstps immediately followed by flds of the same
value? And if it really wants to reload (in my more complex example, it really
needs to), why not just do the loads in the right order from the start instead
of doing the fxch?

[Bug target/49583] Reloading stack operands in the wrong order, so needs to insert fxch

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49583

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-30 10:07:40 UTC ---
Reload doesn't really know how to deal with the register stack.

[Bug target/49583] Reloading stack operands in the wrong order, so needs to insert fxch

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49583

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |WONTFIX

--- Comment #2 from Uros Bizjak <ubizjak at gmail dot com> 2011-07-03 14:38:54 UTC ---
You should use powf instead of pow. Many of these moves are actually
double->float and float->double x87 conversion instructions.

You can use -ffast-math to avoid these conversions, but using powf on float
values is probably what you want.

Other moves are necessary to pass arguments to the function via stack and to
save live values from registers around the call.

Also, don't worry about fxch or two, these are extremely fast instructions.

[Bug target/49583] Reloading stack operands in the wrong order, so needs to insert fxch

[sgunderson at bigfoot dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49583

--- Comment #3 from sgunderson at bigfoot dot com 2011-07-03 17:20:11 UTC ---
Hi,

My bug report was (as you can see in the title) not about the fstps/fld
sequence; it was about the extraneous fxch instructions. (My original code was
with -ffast-math, but I didn't want to burden the example with too many flags.)

In any case, even if I am to ignore “one or two”, how many fxch are too many? I
can give you can code where there are _five_ (in what is a tight inner loop for
me), just by expanding on the example in question.