[Bug rtl-optimization/46279] New: cmov not hoisted out of the loop

[Bug rtl-optimization/46279] New: cmov not hoisted out of the loop

[davidxl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

           Summary: cmov not hoisted out of the loop
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: davidxl@gcc.gnu.org


Simple test case:

extern int gen_int(int);
extern void ref_int_p(int*);

void kernel3 ()
{
  int i;
  int j;
  int k;
  int l;
  int m;
  int a[200];

  j = gen_int (0);
  k = gen_int (0);

  for (i = 0; i < 200; i++)
    {
      if (j < k)
        a[i] = 1;
      else
        a[i] = j;
    }

  ref_int_p (&a[0]);

  return;
}


Code generated by trunk gcc at O2:

kernel3:
.LFB0:
    .cfi_startproc
    pushq    %rbx
    .cfi_def_cfa_offset 16
    .cfi_offset 3, -16
    xorl    %edi, %edi
    subq    $800, %rsp
    .cfi_def_cfa_offset 816
    call    gen_int
    xorl    %edi, %edi
    movl    %eax, %ebx
    call    gen_int
    movq    %rsp, %rdx
    leaq    800(%rsp), %rdi
    movl    $1, %esi
    .p2align 4,,10
    .p2align 3
.L4:
    cmpl    %eax, %ebx
    movl    %esi, %ecx
    cmovge    %ebx, %ecx
    movl    %ecx, (%rdx)
    addq    $4, %rdx
    cmpq    %rdi, %rdx
    jne    .L4
    movq    %rsp, %rdi
    call    ref_int_p
    addq    $800, %rsp
    .cfi_def_cfa_offset 16
    popq    %rbx
    .cfi_def_cfa_offset 8
    ret

The loop header is L4.


LLVM generates:

.Leh_func_begin0:
    pushq    %rbx
.Ltmp0:
    subq    $800, %rsp
.Ltmp1:
    xorl    %edi, %edi
    callq    gen_int
    movl    %eax, %ebx
    xorl    %edi, %edi
    callq    gen_int
    cmpl    %eax, %ebx
    movl    $1, %eax
    cmovgel    %ebx, %eax
    xorl    %ecx, %ecx
    .align    16, 0x90
.LBB0_1:
    movl    %eax, (%rsp,%rcx,4)
    incq    %rcx
    cmpq    $200, %rcx
    jne    .LBB0_1
    leaq    (%rsp), %rdi
    callq    ref_int_p
    addq    $800, %rsp
    popq    %rbx

The loop (LBB0_1) is much tighter.

David

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-11-02 21:31:03 UTC ---
Try -O3.  -O3 enables -funswitch-loops which does this.

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[xinliangli at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

--- Comment #3 from davidxl <xinliangli at gmail dot com> 2010-11-02 21:33:37 UTC ---
Hoisting cmov out of the loop is more efficient ( I tried O3 before filing
the bug).

David

On Tue, Nov 2, 2010 at 2:31 PM, pinskia at gcc dot gnu.org <
gcc-bugzilla@gcc.gnu.org> wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279
>
> --- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-11-02
> 21:31:03 UTC ---
> Try -O3.  -O3 enables -funswitch-loops which does this.
>
> --
> Configure bugmail: http://gcc.gnu.org/bugzilla/userprefs.cgi?tab=email
> ------- You are receiving this mail because: -------
> You are on the CC list for the bug.
>

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-11-02 21:33:18 UTC ---
Though -O3 generates much larger code.  What needs to be done is that one thing
really:
      if (j < k)
        a[i] = 1;
      else
        a[i] = j;

Needs to be transformed into:
      if (j < k)
        tmp = 1;
      else
        tmp = j;
      a[i] = tmp;

And then it should work correctly.  I think there is already a bug about that.

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

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

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

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-11-02 21:44:57 UTC ---
PR46009 ?

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-11-02 21:46:18 UTC ---
(In reply to comment #4)
> PR46009 ?

That one and I thought there was one more but of course I cannot find it.

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.11.03 10:58:33
     Ever Confirmed|0                           |1

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-03 10:58:33 UTC ---
Store sinking.

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
         Depends on|                            |89430
   Target Milestone|---                         |8.0
             Status|NEW                         |RESOLVED

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed for GCC 8, by one of the patches to tree-ssa-phiopt.c
(cond_store_replacement).


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89430
[Bug 89430] A missing ifcvt optimization to generate csel

[Bug rtl-optimization/46279] cmov not hoisted out of the loop

[tnfchris at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=46279
Bug 46279 depends on bug 89430, which changed state.

Bug 89430 Summary: A missing ifcvt optimization to generate csel
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89430

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