Re: 3rd deficiency (was: Superfluous branches due to insufficient flow analysis)

["Stefan Kanthak" <stefan.kanthak@nexgo.de>]

Gabriel Ravier <gabravier@gmail.com> wrote:

Independent from the defunct flow analysis in the presence of NaNs, my
example demonstrates another minor deficiency: know thy instruction set!
See the comments in the assembly below.

> On 8/13/21 8:58 PM, Stefan Kanthak wrote:
>> Hi,
>>
>> compile the following naive implementation of nextafter() for AMD64:
>>
>> JFTR: ignore the aliasing casts, they don't matter here!
>>
>> $ cat repro.c
>> double nextafter(double from, double to)
>> {
>>      if (to != to)
>>          return to;        // to is NAN
>>
>>      if (from != from)
>>          return from;      // from is NAN
>>
>>      if (from == to)       // neither from nor to can be NAN here!
>>          return to;
>>
>>      if (from == 0.0)      // dito!
>>          return to < 0.0 ? -0x1.0p-1074 : 0x1.0p-1074;
>>
>>      unsigned long long ull = *(unsigned long long *) &from;
>>
>>      if ((from < to) == (from < 0.0))
>>          ull--;
>>      else
>>          ull++;
>>
>>      return *(double *) &ull;
>> }
>> $ gcc -m64 -o- -O3 -S repro.c
>> ...
>> nextafter:
[...]
>> .L4:
>>          comisd  %xmm0, %xmm1    # sets CF if first comparand < second comparand
>>          movq    %xmm0, %rdx
>>          leaq    -1(%rdx), %rax  # superfluous
>>          seta    %r8b            #      sbb %rcx, %rcx
>>          comisd  %xmm0, %xmm2
>>          seta    %cl             #      sbb %rax, %rax
>>          addq    $1, %rdx        #      xor %rcx, %rax
>>          cmpb    %cl, %r8b       #      or  $1, %rax
>>          cmovne  %rdx, %rax      #      add %rdx, %rax
>>          movq    %rax, %xmm0

Stefan