🚀 go-pugleaf

Hello,

We use GCC 2.7.0 to compile binaries for R4400-based SGI IRIX 5.2
machines.

Performance of GCC 2.7.0-compiled binaries with compilation flags
-O2 -mips2 -mcpu=r4400 is well below that of binaries compiled with
/bin/cc -O2 -mips2.  For example, the following small matrix
multiplication code:

----------------------------------------------------------------------
#include <stdio.h>

#define MAT_SIZE 256

int main()
{
  register i,j,k;
  double A[MAT_SIZE][MAT_SIZE]; 
  double B[MAT_SIZE][MAT_SIZE]; 
  double C[MAT_SIZE][MAT_SIZE];

  for (i=0; i<MAT_SIZE; i++)
    for (j=0; j<MAT_SIZE; j++)
      for (C[i][j]=k=0; k<MAT_SIZE; k++)
	C[i][j] += A[i][k]*B[k][j];
}
----------------------------------------------------------------------

tcsh's builtin time reports:

/home/zola/.../tmp> time mmult.sgi
3.846u 0.473s 0:04.52 95.3% 0+0k 1+0io 0pf+0w
/home/zola/.../tmp> time mmult.gnu
6.033u 0.060s 0:06.34 96.0% 0+0k 1+0io 0pf+0w

We have seen this order of magnitude performance difference
(i.e. ~50%) in other floating-point intensive code as well.  Is the
GNU C port (or, more likely, GNU assembler port) to MIPS instruction sets
known to be not as effective as the SGI compiler/assembler?
GNU as reports GNU assembler version 2.5.2 (mips-sgi-irix5.2), using
BFD version 2.5.

Thanks,

		Pierre

--
Pierre P. Tremblay          ppt@cae.ca