[Bug tree-optimization/79389] [7 Regression] 30% performance regression in SciMark2 MonteCarlo

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79389

--- Comment #5 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 22 Feb 2017, ubizjak at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79389
> 
> Uroš Bizjak <ubizjak at gmail dot com> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>                  CC|                            |jakub at gcc dot gnu.org
> 
> --- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
> The testcase for missing if-conversion:
> 
> --cut here--
> extern double Random_nextDouble (void);
> 
> int foo (int Num_samples)
> {
>   int count;
>   int under_curve = 0;
> 
>   for (count=0; count<Num_samples; count++)
>     {
>       double x= Random_nextDouble ();
>       double y= Random_nextDouble ();
>       if ( x*x + y*y <= 1.0)
>         under_curve ++;
>     }
> 
>   return under_curve;
> }
> --cut here--
> 
> gcc -O2:
> 
>         xorl    %ebx, %ebx
>         ...
> .L5:
>         call    Random_nextDouble
>         movsd   %xmm0, 8(%rsp)
>         call    Random_nextDouble
>         movsd   8(%rsp), %xmm1
> ++>     leal    1(%rbx), %eax
>         mulsd   %xmm0, %xmm0
>         mulsd   %xmm1, %xmm1
>         movsd   .LC0(%rip), %xmm2
>         addsd   %xmm0, %xmm1
>         ucomisd %xmm1, %xmm2
> -->     cmovnb  %eax, %ebx
>         addl    $1, %ebp
>         cmpl    %ebp, %r12d
>         jne     .L5
> 
> gcc -O3 is even more "innovative":
> 
>         xorl    %r12d, %r12d
>         ...
> .L2:
>         call    Random_nextDouble
>         movsd   %xmm0, 8(%rsp)
>         call    Random_nextDouble
>         movsd   8(%rsp), %xmm1
>         mulsd   %xmm0, %xmm0
>         mulsd   %xmm1, %xmm1
>         movsd   .LC0(%rip), %xmm2
>         addsd   %xmm0, %xmm1
>         ucomisd %xmm1, %xmm2
> -->     jnb     .L13
>         addl    $1, %ebx
>         cmpl    %ebx, %ebp
>         jne     .L2
> 
> .L13:
>         addl    $1, %ebx
> ++>     addl    $1, %r12d
>         cmpl    %ebp, %ebx
>         jne     .L2
> 
> IF-conversion to SETcc + ZEXT + ADD would avoid cmove/jump in this short loop.
> 
> (I'm not sure if tree or RTL ifconvert pass should catch this situation.)

There is only RTL ifconvert left.