[Bug rtl-optimization/3507] appalling optimisation with sub/cmp on multiple targets

[rask at gcc dot gnu dot org]

------- Comment #8 from rask at gcc dot gnu dot org  2007-11-27 18:04 -------
Created an attachment (id=14647)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14647&action=view)
Patch to enhance cse.c

The attached patch can optimize this testcase:

void foo (int);

int bar2 (int a, int b)
{
  int c = a - b;
  if (a < b)
    {
      foo (c);
    }
  return c;
}

Before:
bar2:
        pushl   %ebx            # 33    *pushsi2        [length = 1]
        subl    $8, %esp        # 34    pro_epilogue_adjust_stack_1/1   [length
= 3]
        movl    16(%esp), %edx  # 2     *movsi_1/1      [length = 4]
        movl    20(%esp), %eax  # 3     *movsi_1/1      [length = 4]
        movl    %edx, %ebx      # 32    *movsi_1/1      [length = 2]
        subl    %eax, %ebx      # 7     *subsi_1/1      [length = 2]
        cmpl    %eax, %edx      # 8     *cmpsi_1_insn/1 [length = 2]
        jge     .L2             # 9     *jcc_1  [length = 2]
        movl    %ebx, (%esp)    # 11    *movsi_1/2      [length = 3]
        call    foo             # 12    *call_0 [length = 5]
.L2:
        movl    %ebx, %eax      # 19    *movsi_1/1      [length = 2]
        addl    $8, %esp        # 37    pro_epilogue_adjust_stack_1/1   [length
= 3]
        popl    %ebx            # 38    popsi1  [length = 1]
        ret                     # 39    return_internal [length = 1]

After:
bar2:
        pushl   %ebx            # 33    *pushsi2        [length = 1]
        subl    $8, %esp        # 34    pro_epilogue_adjust_stack_1/1   [length
= 3]
        movl    16(%esp), %eax  # 2     *movsi_1/1      [length = 4]
        movl    %eax, %ebx      # 32    *movsi_1/1      [length = 2]
        subl    20(%esp), %ebx  # 8     *subsi_2/2      [length = 4]
        jns     .L2             # 9     *jcc_1  [length = 2]
        movl    %ebx, (%esp)    # 11    *movsi_1/2      [length = 3]
        call    foo             # 12    *call_0 [length = 5]
.L2:
        movl    %ebx, %eax      # 19    *movsi_1/1      [length = 2]
        addl    $8, %esp        # 37    pro_epilogue_adjust_stack_1/1   [length
= 3]
        popl    %ebx            # 38    popsi1  [length = 1]
        ret             # 39    return_internal [length = 1]

One of the difficulties is that by the time the code gets to the cse1 pass, the
statement "c = a - b" might have been moved below the "a < b" test, making it
harder to optimize. The "bar2" testcase was crafted so this doesn't happen. But
given the need to potentially replace both the sub/cmp insn and the jump insn,
it would be better to move this optimization to combine.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=3507