[Bug c/31661] New: mmintrin calls are slower than plain C

[Bug c/31661] New: mmintrin calls are slower than plain C

[effbiae at gmail dot com]

hi,

this thread:
 http://gcc.gnu.org/ml/gcc-help/2007-04/msg00201.html
details my problems.  (duplicated here)

i want to sum an array of longs using mmx.  i use the functions:
   _mm_set_pi32 and _m_paddd
but the resultant binary contains significantly less efficient code
than inline asm or even plain C ( for(i=0;i<n;i++)total+=a[i]; ).
here's the relevant function:

simd_mmintrin(n, is)
I *is;
{   __m64 q,r;
  I i;
  _m_empty();
  q=_m_from_int(0);
  for (i=0; i < n; i+=W) {
      r=_mm_set_pi32(is[i],is[i+1]);
      q=_m_paddd(q,r);
  }
  union {long a[2];__m64 m;}u;
  u.m=q;
  return u.a[0]+u.a[1];
}

i have a script RUNME.sh:

$ sh RUNME.sh
---
expect: 199990000
impl: C (SISD)
199990000


real    0m0.604s
user    0m0.580s
sys     0m0.004s



---
expect: 199990000
impl: ASM (SIMD)
199990000



real    0m0.377s
user    0m0.360s
sys     0m0.008s



---
expect: 199990000
impl: MMINTRIN (SIMD)
199990000


real    0m1.235s
user    0m1.228s
sys     0m0.004s




$ cat RUNME.sh
#!/bin/sh
repeats=4000        # number of times to repeat the test
vectorsize=10000   # size of the vector in 32 bit ints
gcc -O -mmmx v.c -o v
for which in 0 1 2; do time ./v $repeats $vectorsize $which; done


$ cat v.c
#include <string.h>
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <mmintrin.h>


typedef long I;typedef unsigned long J;
typedef char C;
#define IZ sizeof(I)
#define W 2



simd_mmintrin(n, is)
I *is;
{   __v2si q,r;
  I i;
  _m_empty();
  q=_m_from_int(0);
  for (i=0; i < n; i+=W) {
      memcpy(&r,is+i,IZ*W);
      q=_m_paddd(q,r);
  }
  I*qq=(I*)&q;
  return qq[0]+qq[1];
}


simd_asm(n, is)
I *is;
{   I i,*r=malloc(IZ*W*8);
  asm("emms");
  asm("pxor %mm0,%mm0");
  for (i=0; i < n; i+=W) {
      asm("movq %0,%%mm1\n\t"
          "paddd %%mm1,%%mm0"
          :
          :"m"(is[i])           );
  }
  asm("movq %%mm0,%0":"=m"(*(__m64*)r));
  return r[0]+r[1];
}


sisd(n, is)
I *is;
{
  I i = 0, j = 0;
  for (i = 0; i < n; i++)
      j += is[i];
  return j;
}


main(c, v)
C **v;
{
  I n=atol(v[1]), z=atol(v[2]), m=atol(v[3]);
  I result, *is=malloc(IZ*(z*=2)), i;
  int(*fs[])()={sisd,simd_asm,simd_mmintrin,0};
  C*ss[]={"C (SISD)","ASM (SIMD)","MMINTRIN (SIMD)"};
  for(i=0;i<z;i++)is[i]=i;
  printf("\n\n---\nexpect: %d\n",(z)*(z-1)/2);
  printf("impl: %s\n",ss[m]);
  while (n--)
      result=fs[m](z, is);
  printf("%d\n",result);
}



[jack@fedora i]$ gcc -v
Using built-in specs.
Target: i386-redhat-linux
Configured with: ../configure --prefix=/usr --mandir=/usr/share/man
--infodir=/u
sr/share/info --enable-shared --enable-threads=posix --enable-checking=release
-
-with-system-zlib --enable-__cxa_atexit --disable-libunwind-exceptions
--enable-
libgcj-multifile --enable-languages=c,c++,objc,obj-c++,java,fortran,ada
--enable
-java-awt=gtk --disable-dssi
--with-java-home=/usr/lib/jvm/java-1.4.2-gcj-1.4.2.
0/jre --with-cpu=generic --host=i386-redhat-linux
Thread model: posix
gcc version 4.1.0 20060304 (Red Hat 4.1.0-3)


-- 
           Summary: mmintrin calls are slower than plain C
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: effbiae at gmail dot com
  GCC host triplet: fedora core 5; pentium III


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31661

[Bug target/31661] mmintrin calls are slower than plain C

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2007-04-23 03:04 -------
the main reason why it is slower is because taking the address messes up the
lots of stuff.

And also MMX is slower because we try not to reload it so this is a target
issue.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
          Component|c                           |target
   GCC host triplet|fedora core 5; pentium III  |
 GCC target triplet|                            |i686-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31661