* Optional macro MEMSET_TURBO for faster low-level memory initializations.

git-svn-id: https://rathena.svn.sourceforge.net/svnroot/rathena/trunk@8499 54d463be-8e91-2dee-dedb-b68131a5f0ec

10 files changed, 120 insertions, 23 deletions

diff --git a/src/common/graph.c b/src/common/graph.c
index 0a49fa709..4caa830e6 100644
--- a/src/common/graph.c
+++ b/src/common/graph.c
@@ -71,7 +71,7 @@ void graph_pallet(struct graph* g, int index,unsigned long c) {
 	if(g == NULL || c >= 256) return;

 

 	if(g->pallet_count <= index) {

-		memset(g->png_data + 0x29 + 3 * g->pallet_count,0,(index - g->pallet_count) * 3);

+		malloc_set(g->png_data + 0x29 + 3 * g->pallet_count,0,(index - g->pallet_count) * 3);

 		g->pallet_count = index + 1;

 	}

 	g->png_data[0x29 + index * 3    ] = (unsigned char)((c >> 16) & 0xFF); // R

diff --git a/src/common/grfio.c b/src/common/grfio.c
index aa677a0e5..6cfc7eeaa 100644
--- a/src/common/grfio.c
+++ b/src/common/grfio.c
@@ -169,7 +169,7 @@ static void BitConvert(BYTE *Src,char *BitSwapTable)
 	int lop,prm;

 	BYTE tmp[8];

 //	*(DWORD*)tmp=*(DWORD*)(tmp+4)=0;

-	memset(tmp,0,8);

+	malloc_tsetdword(tmp,0,8);

 	for(lop=0;lop!=64;lop++) {

 		prm = BitSwapTable[lop]-1;

 		if (Src[(prm >> 3) & 7] & BitMaskTable[prm & 7]) {

@@ -299,7 +299,7 @@ int decode_zip(unsigned char *dest, unsigned long* destLen, const unsigned char*
 int encode_zip(unsigned char *dest, unsigned long* destLen, const unsigned char* source, unsigned long sourceLen) {

 	z_stream stream;

 	int err;

-	memset(&stream, 0, sizeof(stream));

+	malloc_tsetdword(&stream, 0, sizeof(stream));

 	stream.next_in = (Bytef*)source;

 	stream.avail_in = (uInt)sourceLen;

 	/* Check for source > 64K on 16-bit machine: */

@@ -479,7 +479,7 @@ static FILELIST* filelist_add(FILELIST *entry)
 

 	if (filelist_entrys >= filelist_maxentry) {

 		filelist = (FILELIST *)aRealloc(filelist, (filelist_maxentry + FILELIST_ADDS) * sizeof(FILELIST));

-		memset(filelist + filelist_maxentry, '\0', FILELIST_ADDS * sizeof(FILELIST));

+		malloc_tsetdword(filelist + filelist_maxentry, '\0', FILELIST_ADDS * sizeof(FILELIST));

 		filelist_maxentry += FILELIST_ADDS;

 	}

 

@@ -949,7 +949,7 @@ char *grfio_alloc_ptr(char *fname)
 	if (gentry_entrys >= gentry_maxentry) {

 		gentry_maxentry += GENTRY_ADDS;

 		gentry_table = (char**)aRealloc(gentry_table, gentry_maxentry * sizeof(char*));

-		memset(gentry_table + (gentry_maxentry - GENTRY_ADDS), 0, sizeof(char*) * GENTRY_ADDS);

+		malloc_tsetdword(gentry_table + (gentry_maxentry - GENTRY_ADDS), 0, sizeof(char*) * GENTRY_ADDS);

 	}

 	len = strlen( fname );

 	buf = (char*)aMallocA(len + 1);

diff --git a/src/common/malloc.c b/src/common/malloc.c
index 6eaa994d8..700da0f88 100644
--- a/src/common/malloc.c
+++ b/src/common/malloc.c
@@ -4,7 +4,7 @@
 #include <stdio.h>

 #include <stdlib.h>

 #include <string.h>

-#include "malloc.h"

+#include "../common/malloc.h"

 #include "../common/core.h"

 #include "../common/showmsg.h"

 

@@ -116,13 +116,15 @@ void aFree_ (void *p, const char *file, int line, const char *func)
 void* _bcallocA(size_t size, size_t cnt)

 {

 	void *ret = MALLOCA(size * cnt);

-	if (ret) memset(ret, 0, size * cnt);

+	if (ret) //memset(ret, 0, size * cnt);

+		malloc_set(ret, 0, size*cnt);

 	return ret;

 }

 void* _bcalloc(size_t size, size_t cnt)

 {

 	void *ret = MALLOC(size * cnt);

-	if (ret) memset(ret, 0, size * cnt);

+	if (ret) //memset(ret, 0, size * cnt);

+		malloc_set(ret, 0, size*cnt);

 	return ret;

 }

 char* _bstrdup(const char *chr)

@@ -325,7 +327,8 @@ void* _mmalloc(size_t size, const char *file, int line, const char *func ) {
 

 void* _mcalloc(size_t num, size_t size, const char *file, int line, const char *func ) {

 	void *p = _mmalloc(num * size,file,line,func);

-	memset(p,0,num * size);

+	//memset(p,0,num * size);

+	malloc_set(p,0,num*size);

 	return p;

 }

 

@@ -683,6 +686,81 @@ static void memmgr_init (void)
 }

 #endif

 

+#ifdef MEMSET_TURBO

+	// This function is practically useless if the the size of the data is

+	// static. It decides whether to use setword or setdword.

+	void malloc_set(void *dest, int value, int size)

+	{

+		if(size%4 == 0)

+			malloc_tsetdword(dest, value, size);

+		else if(size%2 == 0)

+			malloc_tsetword(dest, (short)value, size);

+		else

+			memset(dest, value, (size_t) size);

+	}

+

+	// Sets 32-bit aligned memory.

+	void malloc_tsetdword(void *dest, int value, int count){

+#ifdef _WIN32

+		_asm

+			{

+				mov edx, 0

+				mov eax, count

+				mov ebx, 4

+				idiv ebx

+				mov edi, dest

+				mov ecx, eax

+				mov eax, value

+				rep stosd

+			}

+#else

+		__asm__("movl $0, %%edx;

+				 movl %1, %%eax;

+				 movl $4, %%ebx;

+				 idivl %%ebx;

+				 movl %0, %%edi; 

+				 movl %%eax, %%ecx; 

+				 movl %2, %%eax; 

+				 rep;

+				 stosd;"

+			: 

+			: "g" (dest), "g" (count), "g" (value)

+			: "edx", "eax", "ebx", "edi", "ecx"

+			);

+#endif

+	}

+

+	// Sets 16-bit aligned memory.

+	void malloc_tsetword(void *dest, short value, int count){

+#ifdef _WIN32

+		_asm

+			{

+				mov edx, 0

+				mov eax, count

+				mov ebx, 2

+				idiv ebx

+				mov edi, dest

+				mov ecx, eax

+				mov ax, value

+				rep stosw

+			}

+#else

+		__asm__("movl $0, %%edx;

+				 movl %1, %%eax;

+				 movl $2, %%ebx;

+				 idivl %%ebx;

+				 movl %0, %%edi; 

+				 movl %%eax, %%ecx; 

+				 movw %2, %%ax; 

+				 rep;

+				 stosw;"

+			: 

+			: "g" (dest), "g" (count), "g" (value)

+			: "edx", "eax", "ebx", "edi", "ecx", "ax"

+			);

+#endif

+	}

+#endif

 

 /*======================================

  * Initialise

diff --git a/src/common/malloc.h b/src/common/malloc.h
index e9dbb9d44..d3df667a6 100644
--- a/src/common/malloc.h
+++ b/src/common/malloc.h
@@ -4,6 +4,8 @@
 #ifndef _MALLOC_H_

 #define _MALLOC_H_

 

+//#define MEMSET_TURBO

+

 #ifndef __NETBSD__

 #if __STDC_VERSION__ < 199901L

 #	if __GNUC__ >= 2

@@ -147,6 +149,22 @@
 ////////////////////////////////////////////////

 

 unsigned int malloc_usage (void);

+#ifndef INLINE

+	#ifdef _WIN32

+		#define INLINE 

+	#else

+		#define INLINE inline

+	#endif

+#endif

+#ifdef MEMSET_TURBO

+	INLINE void malloc_tsetdword(void *dest, int value, int count);

+	INLINE void malloc_tsetword(void *dest, short value, int count);

+	INLINE void malloc_set(void *dest, int value, int size);

+#else

+	#define malloc_tsetdword(x,y,z) memset(x,y,z)

+	#define malloc_tsetword(x,y,z) memset(x,y,z)

+	#define malloc_set(x,y,z) memset(x,y,z)

+#endif

 void malloc_init (void);

 void malloc_final (void);

 

diff --git a/src/common/mapindex.c b/src/common/mapindex.c
index 8fbf2da12..0c3faec75 100644
--- a/src/common/mapindex.c
+++ b/src/common/mapindex.c
@@ -3,6 +3,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include "showmsg.h"
+#include "../common/malloc.h"
 
 #define MAX_MAPINDEX 2000
 
@@ -70,7 +71,7 @@ void mapindex_init(void) {
 	int index, length;
 	char map_name[1024];
 	
-	memset (&indexes, 0, sizeof (indexes));
+	malloc_tsetdword (&indexes, 0, sizeof (indexes));
 	fp=fopen(mapindex_cfgfile,"r");
 	if(fp==NULL){
 		ShowFatalError("Unable to read mapindex config file %s!\n", mapindex_cfgfile);
diff --git a/src/common/plugins.c b/src/common/plugins.c
index fbadca065..a057190b9 100644
--- a/src/common/plugins.c
+++ b/src/common/plugins.c
@@ -133,7 +133,7 @@ int export_symbol (void *var, int offset)
 		plugin_call_table = (void**)aRealloc(plugin_call_table, max_call_table*sizeof(void*));

 		

 		// clear the new alloced block

-		memset(plugin_call_table + call_table_size, 0, (max_call_table-call_table_size)*sizeof(void*));

+		malloc_tsetdword(plugin_call_table + call_table_size, 0, (max_call_table-call_table_size)*sizeof(void*));

 	}

 

 	// the new table size is delimited by the new element at the end

diff --git a/src/common/socket.c b/src/common/socket.c
index dc8162ee1..3d7684972 100644
--- a/src/common/socket.c
+++ b/src/common/socket.c
@@ -424,7 +424,7 @@ int make_listen_port(int port)
 

 	CREATE(session[fd], struct socket_data, 1);

 

-	memset(session[fd],0,sizeof(*session[fd]));

+	malloc_set(session[fd],0,sizeof(*session[fd]));

 	session[fd]->func_recv = connect_client;

 

 	return fd;

@@ -502,7 +502,7 @@ int make_listen_bind(long ip,int port)
 

 	CREATE(session[fd], struct socket_data, 1);

 

-	memset(session[fd],0,sizeof(*session[fd]));

+	malloc_set(session[fd],0,sizeof(*session[fd]));

 	session[fd]->func_recv = connect_client;

 

 	ShowStatus("Open listen port on %d.%d.%d.%d:%i\n",

@@ -517,7 +517,7 @@ int console_recieve(int i) {
 	char *buf;

 

 	CREATE_A(buf, char, 64);

-	memset(buf,0,sizeof(64));

+	malloc_tsetdword(buf,0,sizeof(64));

 

 	n = read(0, buf , 64);

 	if ( n < 0 )

@@ -580,7 +580,7 @@ int start_console(void) {
 	if (!session[0]) {	// dummy socket already uses fd 0

 		CREATE(session[0], struct socket_data, 1);

 	}

-	memset(session[0],0,sizeof(*session[0]));

+	malloc_set(session[0],0,sizeof(*session[0]));

 

 	session[0]->func_recv = console_recieve;

 	session[0]->func_console = default_console_parse;

@@ -1369,7 +1369,7 @@ void socket_init (void)
 	session[0]->max_rdata   = (int)2*rfifo_size;

 	session[0]->max_wdata   = (int)2*wfifo_size;

 

-	memset (func_parse_table, 0, sizeof(func_parse_table));

+	malloc_set (func_parse_table, 0, sizeof(func_parse_table));

 	func_parse_table[SESSION_RAW].check = default_func_check;

 	func_parse_table[SESSION_RAW].func = default_func_parse;

 

diff --git a/src/common/socket.h b/src/common/socket.h
index 158759f9e..ca4b20c91 100644
--- a/src/common/socket.h
+++ b/src/common/socket.h
@@ -16,7 +16,7 @@ typedef long in_addr_t;
 #include <netinet/in.h>

 #endif

 #include <time.h>

-#include "malloc.h"

+#include "../common/malloc.h"

 #include "cbasetypes.h"

 

 extern time_t last_tick;

diff --git a/src/common/strlib.c b/src/common/strlib.c
index d62bf4f46..8cff2a878 100644
--- a/src/common/strlib.c
+++ b/src/common/strlib.c
@@ -8,7 +8,7 @@
 

 #include "strlib.h"

 #include "utils.h"

-#include "malloc.h"

+#include "../common/malloc.h"

 

 //-----------------------------------------------

 // string lib.

@@ -123,7 +123,7 @@ char *trim(char *str, const char *delim)
 	char *strp = strtok(str,delim);

 	char buf[1024];

 	char *bufp = buf;

-	memset(buf,0,sizeof buf);

+	malloc_tsetdword(buf,0,sizeof buf);

 

 	while(strp) {

 		strcpy(bufp, strp);

diff --git a/src/common/timer.c b/src/common/timer.c
index 9baa33e08..1e9ff25aa 100644
--- a/src/common/timer.c
+++ b/src/common/timer.c
@@ -20,7 +20,7 @@
 #include <time.h>

 

 #include "timer.h"

-#include "malloc.h"

+#include "../common/malloc.h"

 #include "showmsg.h"

 

 // タイマー間隔の最小値。モンスターの大量召還時、多数のクライアント接続時に

@@ -137,7 +137,7 @@ static void push_timer_heap(int index)
 		} else {

 			timer_heap_max += 256;

 			timer_heap = (int *) aRealloc( timer_heap, sizeof(int) * timer_heap_max);

-			memset(timer_heap + (timer_heap_max - 256), 0, sizeof(int) * 256);

+			malloc_tsetdword(timer_heap + (timer_heap_max - 256), 0, sizeof(int) * 256);

 		}

 	}

 

@@ -210,7 +210,7 @@ int acquire_timer (void)
 		} else {

 			timer_data_max += 256;

 			timer_data = (struct TimerData *) aRealloc( timer_data, sizeof(struct TimerData) * timer_data_max);

-			memset(timer_data + (timer_data_max - 256), 0, sizeof(struct TimerData) * 256);

+			malloc_tsetdword(timer_data + (timer_data_max - 256), 0, sizeof(struct TimerData) * 256);

 		}

 	}

 

@@ -383,7 +383,7 @@ int do_timer(unsigned int tick)
 				if (free_timer_list_pos >= free_timer_list_max) {

 					free_timer_list_max += 256;

 					free_timer_list = (int *) aRealloc(free_timer_list, sizeof(int) * free_timer_list_max);

-					memset(free_timer_list + (free_timer_list_max - 256), 0, 256 * sizeof(int));

+					malloc_tsetdword(free_timer_list + (free_timer_list_max - 256), 0, 256 * sizeof(int));

 				}

 				free_timer_list[free_timer_list_pos++] = i;

 				break;