[Bug target/50131] New: Optimize x = -1 with "or" for -O

[Bug target/50131] New: Optimize x = -1 with "or" for -O

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

             Bug #: 50131
           Summary: Optimize x = -1 with "or" for -O
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: hjl.tools@gmail.com
                CC: ubizjak@gmail.com


[hjl@gnu-6 rtm-1]$ cat x.c
int
foo1 ()
{
  return -1;
}

short
foo2 ()
{
  return -1;
}

long long
foo3 ()
{
  return -1;
}
[hjl@gnu-6 rtm-1]$ gcc -S -Os x.c 
[hjl@gnu-6 rtm-1]$ cat x.s
    .file    "x.c"
    .text
    .globl    foo1
    .type    foo1, @function
foo1:
.LFB0:
    .cfi_startproc
    orl    $-1, %eax
    ret
    .cfi_endproc
.LFE0:
    .size    foo1, .-foo1
    .globl    foo2
    .type    foo2, @function
foo2:
.LFB1:
    .cfi_startproc
    orl    $-1, %eax
    ret
    .cfi_endproc
.LFE1:
    .size    foo2, .-foo2
    .globl    foo3
    .type    foo3, @function
foo3:
.LFB2:
    .cfi_startproc
    orq    $-1, %rax
    ret
    .cfi_endproc
.LFE2:
    .size    foo3, .-foo3
    .ident    "GCC: (GNU) 4.6.0 20110603 (Red Hat 4.6.0-10)"
    .section    .note.GNU-stack,"",@progbits
[hjl@gnu-6 rtm-1]$ gcc -S -O2 x.c 
[hjl@gnu-6 rtm-1]$ cat x.s
    .file    "x.c"
    .text
    .p2align 4,,15
    .globl    foo1
    .type    foo1, @function
foo1:
.LFB0:
    .cfi_startproc
    movl    $-1, %eax
    ret
    .cfi_endproc
.LFE0:
    .size    foo1, .-foo1
    .p2align 4,,15
    .globl    foo2
    .type    foo2, @function
foo2:
.LFB1:
    .cfi_startproc
    movl    $-1, %eax
    ret
    .cfi_endproc
.LFE1:
    .size    foo2, .-foo2
    .p2align 4,,15
    .globl    foo3
    .type    foo3, @function
foo3:
.LFB2:
    .cfi_startproc
    movq    $-1, %rax
    ret
    .cfi_endproc
.LFE2:
    .size    foo3, .-foo3
    .ident    "GCC: (GNU) 4.6.0 20110603 (Red Hat 4.6.0-10)"
    .section    .note.GNU-stack,"",@progbits
[hjl@gnu-6 rtm-1]$ 

I was expecting "or" for -O.

[Bug target/50131] Optimize x = -1 with "or" for -O

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

--- Comment #1 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-19 20:58:00 UTC ---
There are 1273 "movl $0xffffffff," and 924 "movq 0xffffffffffffffff,"
in cc1plus.  This change can reduce cc1plus size by 6242 bytes.

[Bug target/50131] Optimize x = -1 with "or" for -O

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-20 08:32:12 UTC ---
Is or $-1, reg on all CPUs equally expensive to mov $-1, reg though (as or
generally needs the previous reg content while mov does not; I know some CPUs
special case xor reg, reg, but I doubt or $-1, reg is special cased)?

[Bug target/50131] Optimize x = -1 with "or" for -O

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

--- Comment #3 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-20 19:44:49 UTC ---
(In reply to comment #2)
> Is or $-1, reg on all CPUs equally expensive to mov $-1, reg though (as or
> generally needs the previous reg content while mov does not; I know some CPUs
> special case xor reg, reg, but I doubt or $-1, reg is special cased)?

We should add a new x86 optimization flag and turn it on
only if it helps for the processor to be optimized.

[Bug target/50131] Optimize x = -1 with "or" for -O

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-08-21 09:19:28 UTC ---
I'm sure or $-1,reg doesn't avoid the data dependence on reg and may even
result in hitting partial reg stall issues.  Surely xor reg,reg; not reg
might be another alternative?

[Bug target/50131] Optimize x = -1 with "or" for -O

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |sergos.gnu at gmail dot com

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-23 14:38:58 UTC ---
We do have

  /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
     than a MOV.  */
  m_PENT,

We should check how it performs on the current processors.

[Bug target/50131] Optimize x = -1 with "or" for -O

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=50131

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-12-15
           Keywords|                            |missed-optimization
     Ever confirmed|0                           |1

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
.