Missing if optimization

Missing if optimization

[Uros Bizjak]

Hello!

The if optimization, which produces quite optimized code for integer 
case is missing for floatin-point compares.

This code:

int test(int a) {
        return a == 0 ? 0 : (a > 0 ? 1 : -1);
}

produces (with -O2  -fomit-frame-pointer):

test:
        xorl    %eax, %eax
        cmpl    $0, 4(%esp)
        je      .L4
        setg    %al
        movzbl  %al, %eax
        leal    -1(%eax,%eax), %eax
.L4:
        ret

However, fp compares:

double testf(double a) {
        return a == 0.0 ? 0.0 : (a > 0.0 ? 1.0 : -1.0);
}

Produce somehow unoptimized code (-O2 -ffast-math -fomit-frame-pointer):

testf:
        fldl    4(%esp)
        ftst
        fnstsw  %ax
        sahf
        jne     .L9
        fstp    %st(0)
        fldz
        ret
        .p2align 4,,7
.L9:
        ftst
        fnstsw  %ax
        fstp    %st(0)
        sahf
        ja      .L12
        flds    .LC1
        ret
        .p2align 4,,7
.L12:
        fld1
        .p2align 4,,2
        ret

There is no need for second fp compare, two conditional jumps after 
first fp compare would be enough.

Uros.

Re: Missing if optimization

[Andrew Pinski]

On Sep 8, 2004, at 10:19 PM, Uros Bizjak wrote:

> Hello!
>
> The if optimization, which produces quite optimized code for integer 
> case is missing for floatin-point compares.
>
> This code:
>
> int test(int a) {
>        return a == 0 ? 0 : (a > 0 ? 1 : -1);
> }
>
> produces (with -O2  -fomit-frame-pointer):
>
> test:
>        xorl    %eax, %eax
>        cmpl    $0, 4(%esp)
>        je      .L4
>        setg    %al
>        movzbl  %al, %eax
>        leal    -1(%eax,%eax), %eax
> .L4:
>        ret
>
> However, fp compares:
>
> double testf(double a) {
>        return a == 0.0 ? 0.0 : (a > 0.0 ? 1.0 : -1.0);
> }

On PPC we get only one compare:
(This is with a month old compiler):
_testf:
         lis r2,ha16(LC0)
         la r2,lo16(LC0)(r2)
         lfd f0,0(r2)
         fcmpu cr7,f1,f0 <-- the one and only compare
         beq- cr7,L4
         lis r2,ha16(LC1)
         la r2,lo16(LC1)(r2)
         lfd f0,0(r2)
         bgt- cr7,L8
L4:
         fmr f1,f0
         blr
L8:
         lis r2,ha16(LC2)
         la r2,lo16(LC2)(r2)
         lfd f0,0(r2)
         fmr f1,f0
         blr

Thanks,
Andrew Pinski

Re: Missing if optimization

[James E Wilson]

Uros Bizjak wrote:
> The if optimization, which produces quite optimized code for integer 
> case is missing for floatin-point compares.

This is very target dependent, and you didn't mention the target.  From 
the assembly code, it appears to be x86.

Some targets have comparisons that can be used to optimize this 
testcase.  Some don't.

For those that do, whether it gets optimized depends on how the backend 
represents compares, which is done differently for different targets.

As for the optimization itself, there should be no problem there, as has 
already been pointed out we get this right for powerpc.

You might consider filing a bug report for the missed optimization if 
you haven't already.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com