[Bug target/47186] New: -O2 moves invariant address load INTO loop

[Bug target/47186] New: -O2 moves invariant address load INTO loop

[gcc.hall at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47186

           Summary: -O2 moves invariant address load INTO loop
           Product: gcc
           Version: 4.5.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: gcc.hall@gmail.com
              Host: Fedora 14
            Target: i?86-*-*
             Build: 4.5.1 20100924 (Red Hat 4.5.1-4)


gcc -O1 try.c -S -masm=intel -o try.s

In the example below, the address load to esi for the movsd is moved into the
loop when -O2 is used.  With -O1 its outside the loop.

Perhaps this is something to do with scheduling?  I tried with various x86
targets and it always happened, although the position of the address load
within the loop changed between atom and core2 for example.

--------------------------------------
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

int mike[100], joe[100];

int main( int argc, char *argv[] )
{
  int i;
  for( i = 0; i < 100; ++i )
    mike[i] = rand();
  memcpy( joe, mike, sizeof(joe) );
}
--------------------------------------
Compile with -O1 produces ...

.L2:
    call    rand
    mov DWORD PTR mike[0+ebx*4], eax
    add ebx, 1
    cmp ebx, 100
    jne .L2
    mov edi, OFFSET FLAT:joe
    mov esi, OFFSET FLAT:mike
    mov ecx, 100
    rep movsd

Now repeat with -O2, and esi is loaded within the loop

.L2:
    call    rand
    mov esi, OFFSET FLAT:mike
    mov DWORD PTR mike[0+ebx*4], eax
    add ebx, 1
    cmp ebx, 100
    jne .L2
    mov ecx, ebx
    mov edi, OFFSET FLAT:joe
    rep movsd

[Bug target/47186] -O2 moves invariant address load INTO loop

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47186

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Looks to be fixed in 5.1+.

5.1-7.3 produces:
.L2:
        call    rand
        addl    $4, %ebx
        movl    %eax, -4(%ebx)
        cmpl    $mike+400, %ebx
        jne     .L2

8.1+ produces
.L2:
        call    rand
        movl    %eax, mike(,%ebx,4)
        addl    $1, %ebx
        cmpl    $100, %ebx
        jne     .L2