Sandy Bridge-specific double load ports optimization

[Martin Millnert]

Hi gcc-help,

I have a question regarding an optimization example for the Sandy Bridge
march.

In the Intel 64 and IA-32 Architectures Optimization Reference Manual
[1] on page 183, sec 3.6.1.2, there is a piece about optimizing L1D
cache latency w.r.t utilizing the double load ports.

In example 3-32, assembly from Intel's compiler is shown.

I'm trying to reproduce this with Debian Wheezy's packaged gcc (gcc
version 4.6.2 (Debian 4.6.2-12)).
I get assembly which is very similar, but it's not taking advantage of
the double load ports.  Is there any way to turn this on in my gcc
version, or is it not implemented yet?

Code, b.c:
#define BUFF_SIZE 1024

int main() {

  int buff[BUFF_SIZE];
  int sum = 0;
  int i;

  for(i = 0; i < BUFF_SIZE; i++) {
    sum+=buff[i];
  }
  return sum;
}

Most closely resembling compile option I've found:
gcc -march=corei7 -msse -O3 -funroll-loops -g -o b b.c

(or, to really find sandy bridge, but deactivate the AVX encoding
because that's not what I want:
gcc -march=corei7-avx -mno-avx -msse -O3 -funroll-loops -g -o b b.c )

Assembly, in both cases:
<snip>
00000000004003b0 <main>:
#define BUFF_SIZE 1024

int main() {
  4003b0:       48 81 ec 90 0f 00 00    sub    $0xf90,%rsp
  4003b7:       66 0f ef c0             pxor   %xmm0,%xmm0
  4003bb:       48 8d 44 24 88          lea    -0x78(%rsp),%rax
  4003c0:       48 8d 94 24 88 0f 00    lea    0xf88(%rsp),%rdx
  4003c7:       00 
  4003c8:       0f 1f 84 00 00 00 00    nopl   0x0(%rax,%rax,1)
  4003cf:       00 
  int buff[BUFF_SIZE];
  int sum = 0;
  int i;

  for(i = 0; i < BUFF_SIZE; i++) {
    sum+=buff[i];
  4003d0:       66 0f fe 00             paddd  (%rax),%xmm0
  4003d4:       66 0f fe 40 10          paddd  0x10(%rax),%xmm0
  4003d9:       66 0f fe 40 20          paddd  0x20(%rax),%xmm0
  4003de:       66 0f fe 40 30          paddd  0x30(%rax),%xmm0
  4003e3:       66 0f fe 40 40          paddd  0x40(%rax),%xmm0
  4003e8:       66 0f fe 40 50          paddd  0x50(%rax),%xmm0
  4003ed:       66 0f fe 40 60          paddd  0x60(%rax),%xmm0
  4003f2:       66 0f fe 40 70          paddd  0x70(%rax),%xmm0
  4003f7:       48 83 e8 80             sub    $0xffffffffffffff80,%rax
  4003fb:       48 39 d0                cmp    %rdx,%rax
  4003fe:       75 d0                   jne    4003d0 <main+0x20>
  400400:       66 0f 6f d0             movdqa %xmm0,%xmm2
  }
  return sum;
}
  400404:       48 81 c4 90 0f 00 00    add    $0xf90,%rsp
  int buff[BUFF_SIZE];
  int sum = 0;
  int i;

  for(i = 0; i < BUFF_SIZE; i++) {
    sum+=buff[i];
  40040b:       66 0f 73 da 08          psrldq $0x8,%xmm2
  400410:       66 0f fe c2             paddd  %xmm2,%xmm0
  400414:       66 0f 6f c8             movdqa %xmm0,%xmm1
  400418:       66 0f 73 d9 04          psrldq $0x4,%xmm1
  }
  return sum;
  40041d:       66 0f fe c1             paddd  %xmm1,%xmm0
  400421:       66 0f 7e c0             movd   %xmm0,%eax
}
  400425:       c3                      retq   
  400426:       90                      nop
  400427:       90                      nop
<snip>

Best regards,
Martin

[1]
http://www.intel.com/content/dam/doc/manual/64-ia-32-architectures-optimization-manual.pdf