[Bug rtl-optimization/22152] New: Poor loop optimization when using sse2 builtins - regression from 3.3

[Bug rtl-optimization/22152] New: Poor loop optimization when using sse2 builtins - regression from 3.3

[fjahanian at apple dot com]

In the following trivial test case, gcc-4.1 produces very ineffecient code for the loop. gcc-3.3 produces 
much better code.

typedef int __m64 __attribute__ ((__vector_size__ (8)));

__m64 unsigned_add3( const __m64 *a, const __m64 *b, unsigned long count )
{
                __m64 sum;
                unsigned int i;

                for( i = 1; i < count; i++ )
                {
                        sum = (__m64) __builtin_ia32_paddq ((long long)a[i], (long long)b[i]);
                }
                return sum;
}

1) Loop when compiled with gcc-4.1 -O2 -msse2 (note in particular the extra movq to memory):

L4:
        movl    12(%ebp), %esi
        movq    (%eax,%edx,8), %mm0
        paddq   (%esi,%edx,8), %mm0
        incl    %edx
        cmpl    %edx, %ecx
        movq    %mm0, -16(%ebp)
        movl    -16(%ebp), %esi
        movl    -12(%ebp), %edi
        jne     L4

2) Loop using gcc-3.3 compiled with -O2 -msse2:

L6:
        movq    (%esi,%edx,8), %mm0
        paddq   (%eax,%edx,8), %mm0
        addl    $1, %edx
        cmpl    %ecx, %edx
        jb      L6

AFAICT, culprit is reload which generates extra load and store of %mm0:

(insn 62 30 63 2 (set (mem:V2SI (plus:SI (reg/f:SI 6 bp)
                (const_int -16 [0xfffffffffffffff0])) [0 S8 A8])
        (reg:V2SI 29 mm0)) 736 {*movv2si_internal} (nil)
    (nil))

(insn 63 62 32 2 (set (reg/v:V2SI 4 si [orig:61 sum ] [61])
        (mem:V2SI (plus:SI (reg/f:SI 6 bp)
                (const_int -16 [0xfffffffffffffff0])) [0 S8 A8])) 736 {*movv2si_internal} (nil)
    (nil))

Here is the larger test case from which above test was extracted:

#include <xmmintrin.h>

__m64 unsigned_add3( const __m64 *a, const __m64 *b, __m64 *result, unsigned long count )
{
        __m64 carry, temp, sum, one, onesCarry, _a, _b;
        unsigned int i;

        if( count > 0 )
        {
                _a = a[0];
                _b = b[0];

                one = _mm_cmpeq_pi8( _a, _a );  //-1
                one = _mm_sub_si64( _mm_xor_si64( one, one ), one );    //1
                sum = _mm_add_si64( _a, _b );

                onesCarry = _mm_and_si64( _a, _b );             //the 1's bit is set only if the 1's bit add 
generates a carry
                onesCarry = _mm_and_si64( onesCarry, one );             //onesCarry &= 1

                //Trim off the one's bit on both vA and vB to make room for a carry bit at the top after the 
add
                _a = _mm_srli_si64( _a, 1 );                                            //vA >>= 1
                _b = _mm_srli_si64( _b, 1 );                                            //vB >>= 1

                //Add vA to vB and add the carry bit
                carry = _mm_add_si64( _a, _b );
                carry = _mm_add_si64( carry, onesCarry );

                //right shift by 63 bits to get the carry bit for the high 64 bit quantity
                carry = _mm_srli_si64( carry, 63 );

                for( i = 1; i < count; i++ )
                {
                        result[i-1] = sum;
                        _a = a[i];
                        _b = b[i];
                        onesCarry = _mm_and_si64( _a, _b );
                        onesCarry = _mm_and_si64( onesCarry, one );
                        sum = _mm_add_si64( _a, _b );
                        _a = _mm_add_si64( _a, onesCarry );
                        onesCarry = _mm_and_si64( carry, _a );  //find low bit carry
                        sum = _mm_add_si64( sum, carry );               //add in carry bit to low word sum 
                        carry = _mm_add_si64( _a, onesCarry );  //add in low bit carry to high result
                }

                result[i-1] = sum;
        }

        return carry;
}

Again, gcc-3.3 produces much better code for this loop.

-- 
           Summary: Poor loop optimization when using sse2 builtins -
                    regression from 3.3
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: fjahanian at apple dot com
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: apple-x86-darwin
  GCC host triplet: apple-x86-darwin
GCC target triplet: apple-x86-darwin


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22152

[Bug rtl-optimization/22152] Poor loop optimization when using mmx builtins

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-06-22 20:16 -------
This is a MMX builtin and not a SSE builtin.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
  GCC build triplet|apple-x86-darwin            |
   GCC host triplet|apple-x86-darwin            |
 GCC target triplet|apple-x86-darwin            |i?86-*-*, x86_64-*-*
           Keywords|                            |missed-optimization
            Summary|Poor loop optimization when |Poor loop optimization when
                   |using sse2 builtins -       |using mmx builtins
                   |regression from 3.3         |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22152

[Bug target/22152] Poor loop optimization when using mmx builtins

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-06-22 20:26 -------
This comes down to PR 14552 and PR 19161.  Since this is MMX code, it is very hard to get correct as 
GCC does not currently output emms and vector code without you doing it which is what PR 19161 is 
about.  This is a target problem.  Also see PR 22076.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
  BugsThisDependsOn|                            |14552, 19161, 22076
             Status|UNCONFIRMED                 |NEW
          Component|rtl-optimization            |target
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2005-06-22 20:26:23
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22152

[Bug target/22152] Poor loop optimization when using mmx builtins

[fjahanian at apple dot com]

------- Additional Comments From fjahanian at apple dot com  2005-09-13 00:52 -------
Has there been any progress toward fixing the problems addressed by these PRs?
- thanks.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22152