[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