This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[RFC] Faster memset.

I looked how is memset implemented and as it used computed jumps which
are expensive I decided to write different implementation. 

My profile shows that memset inputs are multiples of 8 and it works on 8
byte aligned strings(This is probably caused by calloc).

My original idea was handle start and end by unaligned loads, which
while bit expensive is cheaper than alternatives.

I then discovered that unrolling loop makes difference only on nehalem.
There unrolling twice has same asymptomatic characteristic as original.
Also 256 bit avx stores do not make difference.

For behaviour on unit tests (for real programs I need to also 
handle calls from dynamic linker.) see following:

I want to compare these graphs with theoretical store throughput. Is
there easy way to determine it for given processor?

Here is my current implementation(memset_new_gcc). I created
faster(memset_new) by fixing gcc bug that it aligned code only to 8


#include <emmintrin.h>
#include <stdint.h>
#include <stdlib.h>
typedef __m128i tp_vector;
typedef uint64_t tp_mask;

#define LOAD(x) _mm_load_si128 ((tp_vector *) (x))
#define LOADU(x) _mm_loadu_si128 ((tp_vector *) (x))
#define STORE(x, y) _mm_store_si128 ((tp_vector *) (x), (y))
#define STORENT(x, y) _mm_stream_si128 ((tp_vector *) (x), (y))

#define STOREU(x, y) _mm_storeu_si128 ((tp_vector *) (x), (y))
#define CONCAT(x, y, n) _mm_alignr_epi8 (y , x ,n)
#define MIN _mm_min_epu8
#define EQ  _mm_cmpeq_epi8
#define OR  _mm_or_si128

#define BROADCAST(x) _mm_set1_epi8 (x)
#define get_mask(x) ((uint64_t) _mm_movemask_epi8 (x))

/* Align VALUE down by ALIGN bytes.  */
#define ALIGN_DOWN(value, align) \
       ALIGN_DOWN_M1(value, align - 1)
/* Align VALUE down by ALIGN_M1 + 1 bytes.
   Useful if you have precomputed ALIGN - 1.  */
#define ALIGN_DOWN_M1(value, align_m1) \
       (void *)((uintptr_t)(value) \
                & ~(uintptr_t)(align_m1))

/* Align VALUE up by ALIGN bytes.  */
#define ALIGN_UP(value, align) \
       ALIGN_UP_M1(value, align - 1)
/* Align VALUE up by ALIGN_M1 + 1 bytes.
   Useful if you have precomputed ALIGN - 1.  */
#define ALIGN_UP_M1(value, align_m1) \
       (void *)(((uintptr_t)(value) + (uintptr_t)(align_m1)) \
                & ~(uintptr_t)(align_m1))

static char *memset_small (char *dest, uint64_t c, size_t no, char *ret);

void *memset(char *dest, int _c, size_t n)
  int i; 
	char *dest2 = dest;
  unsigned char c = _c;
  tp_vector vc = BROADCAST(c);
	uint64_t mc = *(uint64_t*)(&vc);
  if (n < 16)
      return memset_small(dest, mc, n, dest);
      STOREU(dest, vc);
      STOREU(dest + n - 16, vc);
      char *to   = ALIGN_DOWN(dest + n, 16);
      dest = ALIGN_DOWN(dest + 16, 16);
			/* TODO simplify test? */
      while (dest != to)
          dest += 16;
  return dest2;
static char *memset_small (char *dest, uint64_t c, size_t no, char *ret)
  if (no & (8))
      ((uint64_t *) dest)[0] = c;
      ((uint64_t *)(dest + no - 8))[0] = c;
      return ret;
  if (no & 4)
      ((uint32_t *) dest)[0] = c;
      ((uint32_t *)(dest + no - 4))[0] = c;
      return ret;
  if (no & 1)
      dest[0] = c;
  if (no & 2)
      ((uint16_t *)(dest + no - 2))[0] = c;
  return ret;

void __bzero (char *dest, size_t n)
	memset(dest, 0, n);

/* To be used by strncpy.  */
char *memset_tail (char *dest, int c, size_t n, char *ret){
	memset (dest, c, n);
	return ret;

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]