[Bug c/11327] New: Non-optimal code when using MMX/SSE builtins

[kevina at gnu dot org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11327

           Summary: Non-optimal code when using MMX/SSE builtins
           Product: gcc
           Version: 3.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: kevina at gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu

Gcc generates non optimal code when the builtin MMX/SSE functions are used.  In
particular it has a tendency to insert unnecessary movq and sometimes
unnecessary memory reads.

For example in the following code:

#include <stddef.h>

typedef int v8qi __attribute__ ((mode(V8QI)));
typedef long long unsigned int ullint;

#define peq  __builtin_ia32_pcmpeqb
#define pmin __builtin_ia32_pminub
#define por(a,b) (v8qi)__builtin_ia32_por((ullint)a, (ullint)b)
#define psubs __builtin_ia32_psubusb

void foo(v8qi * a, v8qi * b, v8qi * c, size_t s)
{
  size_t i;
  v8qi thres;
  memset(&thres, 10, 8);
  for (i = 0; i != s; ++i)
  {
    c[i] = peq( pmin( por(psubs(a[i],b[i]), psubs(b[i],a[i])),
                      thres),
                thres);
  }
}

When compiled with "-O2 -march=pentium3" Gcc generates:

        # %mm2 is the constant thres
        ...
	movq	(%ecx,%eax,8), %mm0
	psubusb	(%edx,%eax,8), %mm0
	movq	%mm0, %mm1
	movq	(%edx,%eax,8), %mm0
	psubusb	(%ecx,%eax,8), %mm0
	por	%mm0, %mm1
	movq	%mm1, %mm0
        pminub	%mm2, %mm0
	pcmpeqb	%mm2, %mm0
	movq	%mm0, (%esi,%eax,8)
        ...

which involves 2 unnecessary memory reads and 1 unnecessary movq.  An optimal
version of the above code:

	movq	(%ecx,%eax,8), %mm0
        movq    (%edx,%eax,8), %mm1
        movq    %mm0, %mm3
        psubusb %mm1, %mm0
        psubusb %mm3, %mm1
        por     %mm1, %mm0
        pminub	%mm2, %mm0
	pcmpeqb	%mm2, %mm0
	movq	%mm0, (%esi,%eax,8)

This is just a simple example.  In more complex code there are more unnecessary
movq.  Spelling out exactly what to do for the inner loop:

    m1 = a[i];
    m2 = b[i];
    m3 = m1;
    m1 = psubs(m1, m2);
    m2 = psubs(m2, m3);
    m1 = por(m1,m2);
    m1 = pmin(m1, thres);
    m1 = peq(m1,thres);
    c[i] = m1;

Does not help.  It avoids the unnecessary memory reads but adds several
unnecessary movq.

The attached files include the example code and the generated code gcc produces
with a "diff" to my optimal version.  The ineffect code is marked with a '-'
while my code is marked with a '+'.