[Bug target/30315] optimize unsigned-add overflow test on x86 to use cpu flags from addl

[ubizjak at gmail dot com]

------- Comment #19 from ubizjak at gmail dot com  2009-11-16 19:05 -------
(In reply to comment #18)
> This bug also still appears in 4.4.2 with --with-arch=pentium3.

pentium3 fails because it is not TARGET_HIMODE_MATH and TARGET_QIMODE_MATH.

So, it fails following testcase:

unsigned short
plusccsa (unsigned short a, unsigned short b)
{
  unsigned short sum = a + b;
  if (sum < a)
    abort ();
  return sum;
}

It used to generate:

plusccsa:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $8, %esp
        movzwl  8(%ebp), %edx
        movzwl  12(%ebp), %eax
        addl    %edx, %eax
        movzwl  %ax, %eax
>>      cmpw    %ax, %dx
        ja      .L97
        leave
        ret
.L97:
        call    abort

And after the fix for wrong mode of compare insn for these targets [1]:

plusccsa:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $8, %esp
        movzwl  8(%ebp), %edx
        movzwl  12(%ebp), %eax
        addl    %edx, %eax
        movzwl  %ax, %eax
>>      cmpl    %eax, %edx
        ja      .L52
        leave
        ret
.L52:
        call    abort

You don't want instructions that operate on bytes or words on pentium3...

At the end of the day, no bug here.

[1] http://gcc.gnu.org/ml/gcc-patches/2009-11/msg00787.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30315