[Bug rtl-optimization/46200] New: optimization regression in simple pointer loop

[Bug rtl-optimization/46200] New: optimization regression in simple pointer loop

[miles at gnu dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

           Summary: optimization regression in simple pointer loop
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: miles@gnu.org


I compiled this simple test function with both g++ 4.5.1 and g++-4.6
(2010-10-16 snapshot):

  void test (int *b, int *e, int stride)
  {
    for (int *p = b; p != e; p += stride)
      *p = 1;
  }

Using compiler options "-O2 -march=core2 -S".

I got the following asm output;

g++ 4.5.1:

  .globl test(int*, int*, int)
  test(int*, int*, int):
          cmpq    %rsi, %rdi
          je      .L1
          movslq  %edx, %rax
          salq    $2, %rax
  .L3:
          movl    $1, (%rdi)
          addq    %rax, %rdi
          cmpq    %rdi, %rsi
          jne     .L3
  .L1:
          rep
          ret


g++-4.6 (2010-10-16 snapshot):

  test(int*, int*, int):
          cmpq    %rsi, %rdi
          je      .L1
          movslq  %edx, %rdx
          salq    $2, %rdx
          movq    %rdx, %rcx
          addq    %rdx, %rdi
          negq    %rcx
          jmp     .L3
  .L4:
          movq    %rax, %rdi
  .L3:
          movl    $1, (%rdi,%rcx)
          leaq    (%rdi,%rdx), %rax
          cmpq    %rdi, %rsi
          jne     .L4
  .L1:
          rep
          ret


The output of 4.6 seems clearly worse than the output of 4.5.

Thanks,

-miles

[Bug target/46200] [4.6 Regression] optimization regression in simple pointer loop

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-*-*
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.10.28 14:53:08
          Component|rtl-optimization            |target
   Target Milestone|---                         |4.6.0
            Summary|optimization regression in  |[4.6 Regression]
                   |simple pointer loop         |optimization regression in
                   |                            |simple pointer loop
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-10-28 14:53:08 UTC ---
Confirmed.

[Bug target/46200] [4.6 Regression] optimization regression in simple pointer loop

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING
                 CC|                            |hjl.tools at gmail dot com,
                   |                            |mkuvyrkov at gmail dot com

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-28 17:14:28 UTC ---
Does -mtune=generic have regressions? The current -mtune=core2 is slower
than -mtune=generic.

Maxim, could you please try this tescase with your Core 2/Core i7 changes?

[Bug target/46200] [4.6 Regression] optimization regression in simple pointer loop

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-10-28 17:42:55 UTC ---
I saw this with generic tuning also.

[Bug target/46200] [4.6 Regression] optimization regression in simple pointer loop

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |xinliangli at gmail dot com

--- Comment #4 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-28 17:53:50 UTC ---
It is caused by revision 162653:

http://gcc.gnu.org/ml/gcc-cvs/2010-07/msg01007.html

With -o2, revision 162651 gave:

test:
.LFB0:
    .cfi_startproc
    cmpq    %rsi, %rdi
    je    .L1
    movslq    %edx, %rdx
    salq    $2, %rdx
    .p2align 4,,10
    .p2align 3
.L3:
    movl    $1, (%rdi)
    addq    %rdx, %rdi
    cmpq    %rdi, %rsi
    jne    .L3
.L1:
    rep
    ret
    .cfi_endproc

Revision 162653 gave

test:
.LFB0:
    .cfi_startproc
    cmpq    %rsi, %rdi
    je    .L1
    movslq    %edx, %rdx
    salq    $2, %rdx
    movq    %rdx, %rcx
    addq    %rdx, %rdi
    negq    %rcx
    .p2align 4,,10
    .p2align 3
.L3:
    movl    $1, (%rdi,%rcx)
    addq    %rdx, %rdi
    movq    %rdi, %rax
    subq    %rdx, %rax
    cmpq    %rax, %rsi
    jne    .L3
.L1:
    rep
    ret

[Bug target/46200] [4.6 Regression] optimization regression in simple pointer loop

[davidxl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

davidxl <davidxl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
                 CC|                            |davidxl at gcc dot gnu.org

--- Comment #5 from davidxl <davidxl at gcc dot gnu.org> 2010-10-28 19:01:16 UTC ---
Confirmed.

The problem seems to be in the cost computation for loop exit tests -- the cost
associated with iv update seem to be double counted (already considered as iv
cost, but included again in testing cost). If this is the root cause, it is
there since day-1, but exposed by the ivopt enhancement patch.

David

[Bug target/46200] [4.6 Regression] optimization regression in simple pointer loop

[davidxl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46200

davidxl <davidxl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #6 from davidxl <davidxl at gcc dot gnu.org> 2010-11-03 22:40:56 UTC ---
Fix in r166280.