azanella

= test.c =

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>

uint32_t test_rldimi ();
uint32_t test_rlwimi ();

typedef unsigned long long int hp_timing_t;

#define HP_TIMING_ZERO(Var)     (Var) = (0)

#define HP_TIMING_NOW(Var)                                              \
  do {                                                                  \
    unsigned int hi, lo, tmp;                                           \
    __asm__ __volatile__ ("1:   mfspr   %0,269;"                        \
                          "     mfspr   %1,268;"                        \
                          "     mfspr   %2,269;"                        \
                          "     cmpw    %0,%2;"                         \
                          "     bne     1b;"                            \
                          : "=&r" (hi), "=&r" (lo), "=&r" (tmp)         \
                          : : "cr0");                                   \
    Var = ((hp_timing_t) hi << 32) | lo;                                \
  } while (0)

static hp_timing_t hp_timing_overhead = 0ull;

#define HP_TIMING_DIFF_INIT() \
  do {                                                                  \
    if (hp_timing_overhead == 0)                                        \
      {                                                                 \
        int __cnt = 5;                                                  \
        hp_timing_overhead = ~0ull;                                     \
        do                                                              \
          {                                                             \
            hp_timing_t __t1, __t2;                                     \
            HP_TIMING_NOW (__t1);                                       \
            HP_TIMING_NOW (__t2);                                       \
            if (__t2 - __t1 < hp_timing_overhead)                       \
              hp_timing_overhead = __t2 - __t1;                         \
          }                                                             \
        while (--__cnt > 0);                                            \
      }                                                                 \
  } while (0)

#define HP_TIMING_DIFF(Diff, Start, End)        (Diff) = ((End) - (Start))

#define MIN(a,b) (((a)<(b))?(a):(b))
#define MAX(a,b) (((a)<(b))?(b):(a))

#define NCALLS 1024

int main ()
{
  hp_timing_t start, end, diff;
  hp_timing_t min, max;
  int i=0;

  HP_TIMING_DIFF_INIT ();

  min = ~0ull;
  max = 0ull;

  for (i=0; i<NCALLS; ++i)
    {
      HP_TIMING_NOW (start);
      test_rldimi ();
      HP_TIMING_NOW (end);
      HP_TIMING_DIFF (diff, start, end);
      min = MIN (min, diff);
      max = MAX (max, diff);
    }
  printf ("rldimi: min: %llu | max: %llu\n", min, max);

  for (i=0; i<NCALLS; ++i)
    {
      HP_TIMING_NOW (start);
      test_rlwimi ();
      HP_TIMING_NOW (end);
      HP_TIMING_DIFF (diff, start, end);
      min = MIN (min, diff);
      max = MAX (max, diff);
    }
  printf ("rlwimi: min: %llu | max: %llu\n", min, max);

  return 0;
}

= test_impl.S =

#define CALL_MCOUNT
#define ALIGNARG(X) X
#define C_SYMBOL_NAME(name) name
#define C_LABEL(name)  name##:
#define ASM_SIZE_DIRECTIVE(name) .size name,.-name

#define	ENTRY(name)							      \
  .globl C_SYMBOL_NAME(name);						      \
  .type C_SYMBOL_NAME(name),@function;					      \
  .align ALIGNARG(2);							      \
  C_LABEL(name)								      \
  .cfi_startproc;							      \
  CALL_MCOUNT

#define END(name)							      \
  .cfi_endproc;								      \
  ASM_SIZE_DIRECTIVE(name)


ENTRY (test_rldimi)
	rldimi	3, 3, 8, 16
	blr
END (test_rldimi)

ENTRY (test_rlwimi)
	rlwimi	3, 3, 16, 32
	blr
END (test_rlwimi)