poor optimisation case

[Tim Prince]

tristan@wibberley.org wrote:
> Hi
> 
> I've found a case which looks like it should be possible to optimise but
> gcc (very recent trunk) isn't doing which could give improvements in
> many cases - certainly in a case I've come across:
> 
>     #ifdef NEW
>     unsigned int fn(unsigned int n, unsigned int dmax) throw()
>     {
>       for (unsigned int d = 0; d < dmax; ++d) {
>         n += d?d:1;
>       }
>       return n;
>     }
>     #else
>     unsigned int fn(unsigned int n, unsigned int dmax) throw()
>     {
>       unsigned int add = 1;
>       for (unsigned int d = 0; d < dmax; add = ++d) {
>         n += add;
>       }
>       return n;
>     }
>     #endif
> 
> When compiled with -O3 -DOLD I get:
> 
>         .p2align 4,,15
>     .globl _Z2fnjj
>         .type   _Z2fnjj, @function
>     _Z2fnjj:
>     .LFB2:
>         testl   %esi, %esi
>         je  .L2
>         movl    $1, %edx
>         xorl    %eax, %eax
>         .p2align 4,,10
>         .p2align 3
>     .L3:
>         addl    $1, %eax
>         addl    %edx, %edi
>         cmpl    %esi, %eax
>         movl    %eax, %edx
>         jne .L3
>     .L2:
>         movl    %edi, %eax
>         ret
>     .LFE2:
>         .size   _Z2fnjj, .-_Z2fnjj
>     
> but with -DNEW I get:
> 
>         .p2align 4,,15
>     .globl _Z2fnjj
>         .type   _Z2fnjj, @function
>     _Z2fnjj:
>     .LFB2:
>         testl   %esi, %esi
>         je  .L2
>         movl    $1, %edx
>         xorl    %eax, %eax
>         movl    $1, %ecx
>         jmp .L7
>         .p2align 4,,10
>         .p2align 3
>     .L5:
>         testl   %eax, %eax
>         movl    %ecx, %edx
>         cmovne  %eax, %edx
>     .L7:
>         addl    $1, %eax
>         addl    %edx, %edi
>         cmpl    %esi, %eax
>         jne .L5
>     .L2:
>         movl    %edi, %eax
>         ret
>     .LFE2:
>         .size   _Z2fnjj, .-_Z2fnjj
> 
> The performance difference is about 50% with -DNEW taking 1.5 times as
> long as -DOLD (that was with dmax == 1000000000).
> 
> The loop unfortunately can't always be written as in -DOLD as the
> implementation of an iterator adapter might use ?: to special case the
> first element of a sequence and when used in a generic algorithm which
> just has the simple loop of -DNEW it ought to be optimised like -DOLD if
> inlining occurs.
> 
I don't see why you special case the first iteration of a loop with ? 
inside the loop.  Simply write the first iteration separately, and begin 
the loop with the next iteration.  It should be a lot clearer both to us 
and to the compiler what is your intention.
Doesn't this belong on gcc-help?  Better peeling optimization needs more 
justification than this.