1
0
Fork 0
mirror of https://github.com/git/git.git synced 2024-11-16 22:14:53 +01:00
git/compat/nedmalloc/nedmalloc.c

970 lines
25 KiB
C
Raw Normal View History

Add custom memory allocator to MinGW and MacOS builds The standard allocator on Windows is pretty bad prior to Windows Vista, and nedmalloc is better than the modified dlmalloc provided with newer versions of the MinGW libc. NedMalloc stats in Git ---------------------- All results are the best result out of 3 runs. The benchmarks have been done on different hardware, so the repack times are not comparable. These benchmarks are all based on 'git repack -adf' on the Linux kernel. XP ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 3.4.2 (1T) 00:12:28.422 3.4.2 + nedmalloc (1T) 00:07:25.437 1.68x 3.4.5 (1T) 00:12:20.718 3.4.5 + nedmalloc (1T) 00:07:24.809 1.67x 4.3.3-tdm (1T) 00:12:01.843 4.3.3-tdm + nedmalloc (1T) 00:07:16.468 1.65x 4.3.3-tdm (2T) 00:07:35.062 4.3.3-tdm + nedmalloc (2T) 00:04:57.874 1.54x Vista ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 4.3.3-tdm (1T) 00:07:40.844 4.3.3-tdm + nedmalloc (1T) 00:07:17.548 1.05x 4.3.3-tdm (2T) 00:05:33.746 4.3.3-tdm + nedmalloc (2T) 00:05:27.334 1.02x Mac Mini ----------------------------------------------- GCC Threads Total Time Speed ----------------------------------------------- i686-darwin9-4.0.1 (2T) 00:09:57.346 i686-darwin9-4.0.1+ned (2T) 00:08:51.072 1.12x Signed-off-by: Marius Storm-Olsen <marius@trolltech.com> Signed-off-by: Steffen Prohaska <prohaska@zib.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-05-31 18:15:23 +02:00
/* Alternative malloc implementation for multiple threads without
lock contention based on dlmalloc. (C) 2005-2006 Niall Douglas
Boost Software License - Version 1.0 - August 17th, 2003
Permission is hereby granted, free of charge, to any person or organization
obtaining a copy of the software and accompanying documentation covered by
this license (the "Software") to use, reproduce, display, distribute,
execute, and transmit the Software, and to prepare derivative works of the
Software, and to permit third-parties to whom the Software is furnished to
do so, all subject to the following:
The copyright notices in the Software and this entire statement, including
the above license grant, this restriction and the following disclaimer,
must be included in all copies of the Software, in whole or in part, and
all derivative works of the Software, unless such copies or derivative
works are solely in the form of machine-executable object code generated by
a source language processor.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
DEALINGS IN THE SOFTWARE.
*/
#ifdef _MSC_VER
/* Enable full aliasing on MSVC */
/*#pragma optimize("a", on)*/
#endif
/*#define FULLSANITYCHECKS*/
#include "nedmalloc.h"
#if defined(WIN32)
Add custom memory allocator to MinGW and MacOS builds The standard allocator on Windows is pretty bad prior to Windows Vista, and nedmalloc is better than the modified dlmalloc provided with newer versions of the MinGW libc. NedMalloc stats in Git ---------------------- All results are the best result out of 3 runs. The benchmarks have been done on different hardware, so the repack times are not comparable. These benchmarks are all based on 'git repack -adf' on the Linux kernel. XP ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 3.4.2 (1T) 00:12:28.422 3.4.2 + nedmalloc (1T) 00:07:25.437 1.68x 3.4.5 (1T) 00:12:20.718 3.4.5 + nedmalloc (1T) 00:07:24.809 1.67x 4.3.3-tdm (1T) 00:12:01.843 4.3.3-tdm + nedmalloc (1T) 00:07:16.468 1.65x 4.3.3-tdm (2T) 00:07:35.062 4.3.3-tdm + nedmalloc (2T) 00:04:57.874 1.54x Vista ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 4.3.3-tdm (1T) 00:07:40.844 4.3.3-tdm + nedmalloc (1T) 00:07:17.548 1.05x 4.3.3-tdm (2T) 00:05:33.746 4.3.3-tdm + nedmalloc (2T) 00:05:27.334 1.02x Mac Mini ----------------------------------------------- GCC Threads Total Time Speed ----------------------------------------------- i686-darwin9-4.0.1 (2T) 00:09:57.346 i686-darwin9-4.0.1+ned (2T) 00:08:51.072 1.12x Signed-off-by: Marius Storm-Olsen <marius@trolltech.com> Signed-off-by: Steffen Prohaska <prohaska@zib.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-05-31 18:15:23 +02:00
#include <malloc.h>
#endif
#define MSPACES 1
#define ONLY_MSPACES 1
#ifndef USE_LOCKS
#define USE_LOCKS 1
#endif
#define FOOTERS 1 /* Need to enable footers so frees lock the right mspace */
#undef DEBUG /* dlmalloc wants DEBUG either 0 or 1 */
#ifdef _DEBUG
#define DEBUG 1
#else
#define DEBUG 0
#endif
#ifdef NDEBUG /* Disable assert checking on release builds */
#undef DEBUG
#endif
/* The default of 64Kb means we spend too much time kernel-side */
#ifndef DEFAULT_GRANULARITY
#define DEFAULT_GRANULARITY (1*1024*1024)
#endif
/*#define USE_SPIN_LOCKS 0*/
/*#define FORCEINLINE*/
#include "malloc.c.h"
#ifdef NDEBUG /* Disable assert checking on release builds */
#undef DEBUG
#endif
/* The maximum concurrent threads in a pool possible */
#ifndef MAXTHREADSINPOOL
#define MAXTHREADSINPOOL 16
#endif
/* The maximum number of threadcaches which can be allocated */
#ifndef THREADCACHEMAXCACHES
#define THREADCACHEMAXCACHES 256
#endif
/* The maximum size to be allocated from the thread cache */
#ifndef THREADCACHEMAX
#define THREADCACHEMAX 8192
#endif
#if 0
/* The number of cache entries for finer grained bins. This is (topbitpos(THREADCACHEMAX)-4)*2 */
#define THREADCACHEMAXBINS ((13-4)*2)
#else
/* The number of cache entries. This is (topbitpos(THREADCACHEMAX)-4) */
#define THREADCACHEMAXBINS (13-4)
#endif
/* Point at which the free space in a thread cache is garbage collected */
#ifndef THREADCACHEMAXFREESPACE
#define THREADCACHEMAXFREESPACE (512*1024)
#endif
#ifdef WIN32
#define TLSVAR DWORD
#define TLSALLOC(k) (*(k)=TlsAlloc(), TLS_OUT_OF_INDEXES==*(k))
#define TLSFREE(k) (!TlsFree(k))
#define TLSGET(k) TlsGetValue(k)
#define TLSSET(k, a) (!TlsSetValue(k, a))
#ifdef DEBUG
static LPVOID ChkedTlsGetValue(DWORD idx)
{
LPVOID ret=TlsGetValue(idx);
assert(S_OK==GetLastError());
return ret;
}
#undef TLSGET
#define TLSGET(k) ChkedTlsGetValue(k)
#endif
#else
#define TLSVAR pthread_key_t
#define TLSALLOC(k) pthread_key_create(k, 0)
#define TLSFREE(k) pthread_key_delete(k)
#define TLSGET(k) pthread_getspecific(k)
#define TLSSET(k, a) pthread_setspecific(k, a)
#endif
#if 0
/* Only enable if testing with valgrind. Causes misoperation */
#define mspace_malloc(p, s) malloc(s)
#define mspace_realloc(p, m, s) realloc(m, s)
#define mspace_calloc(p, n, s) calloc(n, s)
#define mspace_free(p, m) free(m)
#endif
#if defined(__cplusplus)
#if !defined(NO_NED_NAMESPACE)
namespace nedalloc {
#else
extern "C" {
#endif
#endif
size_t nedblksize(void *mem) THROWSPEC
{
#if 0
/* Only enable if testing with valgrind. Causes misoperation */
return THREADCACHEMAX;
#else
if(mem)
{
mchunkptr p=mem2chunk(mem);
assert(cinuse(p)); /* If this fails, someone tried to free a block twice */
if(cinuse(p))
return chunksize(p)-overhead_for(p);
}
return 0;
#endif
}
void nedsetvalue(void *v) THROWSPEC { nedpsetvalue(0, v); }
void * nedmalloc(size_t size) THROWSPEC { return nedpmalloc(0, size); }
void * nedcalloc(size_t no, size_t size) THROWSPEC { return nedpcalloc(0, no, size); }
void * nedrealloc(void *mem, size_t size) THROWSPEC { return nedprealloc(0, mem, size); }
void nedfree(void *mem) THROWSPEC { nedpfree(0, mem); }
void * nedmemalign(size_t alignment, size_t bytes) THROWSPEC { return nedpmemalign(0, alignment, bytes); }
#if !NO_MALLINFO
struct mallinfo nedmallinfo(void) THROWSPEC { return nedpmallinfo(0); }
#endif
int nedmallopt(int parno, int value) THROWSPEC { return nedpmallopt(0, parno, value); }
int nedmalloc_trim(size_t pad) THROWSPEC { return nedpmalloc_trim(0, pad); }
void nedmalloc_stats(void) THROWSPEC { nedpmalloc_stats(0); }
size_t nedmalloc_footprint(void) THROWSPEC { return nedpmalloc_footprint(0); }
Add custom memory allocator to MinGW and MacOS builds The standard allocator on Windows is pretty bad prior to Windows Vista, and nedmalloc is better than the modified dlmalloc provided with newer versions of the MinGW libc. NedMalloc stats in Git ---------------------- All results are the best result out of 3 runs. The benchmarks have been done on different hardware, so the repack times are not comparable. These benchmarks are all based on 'git repack -adf' on the Linux kernel. XP ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 3.4.2 (1T) 00:12:28.422 3.4.2 + nedmalloc (1T) 00:07:25.437 1.68x 3.4.5 (1T) 00:12:20.718 3.4.5 + nedmalloc (1T) 00:07:24.809 1.67x 4.3.3-tdm (1T) 00:12:01.843 4.3.3-tdm + nedmalloc (1T) 00:07:16.468 1.65x 4.3.3-tdm (2T) 00:07:35.062 4.3.3-tdm + nedmalloc (2T) 00:04:57.874 1.54x Vista ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 4.3.3-tdm (1T) 00:07:40.844 4.3.3-tdm + nedmalloc (1T) 00:07:17.548 1.05x 4.3.3-tdm (2T) 00:05:33.746 4.3.3-tdm + nedmalloc (2T) 00:05:27.334 1.02x Mac Mini ----------------------------------------------- GCC Threads Total Time Speed ----------------------------------------------- i686-darwin9-4.0.1 (2T) 00:09:57.346 i686-darwin9-4.0.1+ned (2T) 00:08:51.072 1.12x Signed-off-by: Marius Storm-Olsen <marius@trolltech.com> Signed-off-by: Steffen Prohaska <prohaska@zib.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-05-31 18:15:23 +02:00
void **nedindependent_calloc(size_t elemsno, size_t elemsize, void **chunks) THROWSPEC { return nedpindependent_calloc(0, elemsno, elemsize, chunks); }
void **nedindependent_comalloc(size_t elems, size_t *sizes, void **chunks) THROWSPEC { return nedpindependent_comalloc(0, elems, sizes, chunks); }
struct threadcacheblk_t;
typedef struct threadcacheblk_t threadcacheblk;
struct threadcacheblk_t
{ /* Keep less than 16 bytes on 32 bit systems and 32 bytes on 64 bit systems */
#ifdef FULLSANITYCHECKS
unsigned int magic;
#endif
unsigned int lastUsed, size;
threadcacheblk *next, *prev;
};
typedef struct threadcache_t
{
#ifdef FULLSANITYCHECKS
unsigned int magic1;
#endif
int mymspace; /* Last mspace entry this thread used */
long threadid;
unsigned int mallocs, frees, successes;
size_t freeInCache; /* How much free space is stored in this cache */
threadcacheblk *bins[(THREADCACHEMAXBINS+1)*2];
#ifdef FULLSANITYCHECKS
unsigned int magic2;
#endif
} threadcache;
struct nedpool_t
{
MLOCK_T mutex;
void *uservalue;
int threads; /* Max entries in m to use */
threadcache *caches[THREADCACHEMAXCACHES];
TLSVAR mycache; /* Thread cache for this thread. 0 for unset, negative for use mspace-1 directly, otherwise is cache-1 */
mstate m[MAXTHREADSINPOOL+1]; /* mspace entries for this pool */
};
static nedpool syspool;
static FORCEINLINE unsigned int size2binidx(size_t _size) THROWSPEC
{ /* 8=1000 16=10000 20=10100 24=11000 32=100000 48=110000 4096=1000000000000 */
unsigned int topbit, size=(unsigned int)(_size>>4);
/* 16=1 20=1 24=1 32=10 48=11 64=100 96=110 128=1000 4096=100000000 */
#if defined(__GNUC__)
topbit = sizeof(size)*__CHAR_BIT__ - 1 - __builtin_clz(size);
#elif defined(_MSC_VER) && _MSC_VER>=1300
{
unsigned long bsrTopBit;
_BitScanReverse(&bsrTopBit, size);
topbit = bsrTopBit;
}
#else
#if 0
union {
unsigned asInt[2];
double asDouble;
};
int n;
asDouble = (double)size + 0.5;
topbit = (asInt[!FOX_BIGENDIAN] >> 20) - 1023;
#else
{
unsigned int x=size;
x = x | (x >> 1);
x = x | (x >> 2);
x = x | (x >> 4);
x = x | (x >> 8);
x = x | (x >>16);
x = ~x;
x = x - ((x >> 1) & 0x55555555);
x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
x = (x + (x >> 4)) & 0x0F0F0F0F;
x = x + (x << 8);
x = x + (x << 16);
topbit=31 - (x >> 24);
}
#endif
#endif
return topbit;
}
#ifdef FULLSANITYCHECKS
static void tcsanitycheck(threadcacheblk **ptr) THROWSPEC
{
assert((ptr[0] && ptr[1]) || (!ptr[0] && !ptr[1]));
if(ptr[0] && ptr[1])
{
assert(nedblksize(ptr[0])>=sizeof(threadcacheblk));
assert(nedblksize(ptr[1])>=sizeof(threadcacheblk));
assert(*(unsigned int *) "NEDN"==ptr[0]->magic);
assert(*(unsigned int *) "NEDN"==ptr[1]->magic);
assert(!ptr[0]->prev);
assert(!ptr[1]->next);
if(ptr[0]==ptr[1])
{
assert(!ptr[0]->next);
assert(!ptr[1]->prev);
}
}
}
static void tcfullsanitycheck(threadcache *tc) THROWSPEC
{
threadcacheblk **tcbptr=tc->bins;
int n;
for(n=0; n<=THREADCACHEMAXBINS; n++, tcbptr+=2)
{
threadcacheblk *b, *ob=0;
tcsanitycheck(tcbptr);
for(b=tcbptr[0]; b; ob=b, b=b->next)
{
assert(*(unsigned int *) "NEDN"==b->magic);
assert(!ob || ob->next==b);
assert(!ob || b->prev==ob);
}
}
}
#endif
static NOINLINE void RemoveCacheEntries(nedpool *p, threadcache *tc, unsigned int age) THROWSPEC
{
#ifdef FULLSANITYCHECKS
tcfullsanitycheck(tc);
#endif
if(tc->freeInCache)
{
threadcacheblk **tcbptr=tc->bins;
int n;
for(n=0; n<=THREADCACHEMAXBINS; n++, tcbptr+=2)
{
threadcacheblk **tcb=tcbptr+1; /* come from oldest end of list */
/*tcsanitycheck(tcbptr);*/
for(; *tcb && tc->frees-(*tcb)->lastUsed>=age; )
{
threadcacheblk *f=*tcb;
size_t blksize=f->size; /*nedblksize(f);*/
assert(blksize<=nedblksize(f));
assert(blksize);
#ifdef FULLSANITYCHECKS
assert(*(unsigned int *) "NEDN"==(*tcb)->magic);
#endif
*tcb=(*tcb)->prev;
if(*tcb)
(*tcb)->next=0;
else
*tcbptr=0;
tc->freeInCache-=blksize;
assert((long) tc->freeInCache>=0);
mspace_free(0, f);
/*tcsanitycheck(tcbptr);*/
}
}
}
#ifdef FULLSANITYCHECKS
tcfullsanitycheck(tc);
#endif
}
static void DestroyCaches(nedpool *p) THROWSPEC
{
if(p->caches)
{
threadcache *tc;
int n;
for(n=0; n<THREADCACHEMAXCACHES; n++)
{
if((tc=p->caches[n]))
{
tc->frees++;
RemoveCacheEntries(p, tc, 0);
assert(!tc->freeInCache);
tc->mymspace=-1;
tc->threadid=0;
mspace_free(0, tc);
p->caches[n]=0;
}
}
}
}
static NOINLINE threadcache *AllocCache(nedpool *p) THROWSPEC
{
threadcache *tc=0;
int n, end;
ACQUIRE_LOCK(&p->mutex);
for(n=0; n<THREADCACHEMAXCACHES && p->caches[n]; n++);
if(THREADCACHEMAXCACHES==n)
{ /* List exhausted, so disable for this thread */
RELEASE_LOCK(&p->mutex);
return 0;
}
tc=p->caches[n]=(threadcache *) mspace_calloc(p->m[0], 1, sizeof(threadcache));
if(!tc)
{
RELEASE_LOCK(&p->mutex);
return 0;
}
#ifdef FULLSANITYCHECKS
tc->magic1=*(unsigned int *)"NEDMALC1";
tc->magic2=*(unsigned int *)"NEDMALC2";
#endif
tc->threadid=(long)(size_t)CURRENT_THREAD;
for(end=0; p->m[end]; end++);
tc->mymspace=tc->threadid % end;
RELEASE_LOCK(&p->mutex);
if(TLSSET(p->mycache, (void *)(size_t)(n+1))) abort();
return tc;
}
static void *threadcache_malloc(nedpool *p, threadcache *tc, size_t *size) THROWSPEC
{
void *ret=0;
unsigned int bestsize;
unsigned int idx=size2binidx(*size);
size_t blksize=0;
threadcacheblk *blk, **binsptr;
#ifdef FULLSANITYCHECKS
tcfullsanitycheck(tc);
#endif
/* Calculate best fit bin size */
bestsize=1<<(idx+4);
#if 0
/* Finer grained bin fit */
idx<<=1;
if(*size>bestsize)
{
idx++;
bestsize+=bestsize>>1;
}
if(*size>bestsize)
{
idx++;
bestsize=1<<(4+(idx>>1));
}
#else
if(*size>bestsize)
{
idx++;
bestsize<<=1;
}
#endif
assert(bestsize>=*size);
if(*size<bestsize) *size=bestsize;
assert(*size<=THREADCACHEMAX);
assert(idx<=THREADCACHEMAXBINS);
binsptr=&tc->bins[idx*2];
/* Try to match close, but move up a bin if necessary */
blk=*binsptr;
if(!blk || blk->size<*size)
{ /* Bump it up a bin */
if(idx<THREADCACHEMAXBINS)
{
idx++;
binsptr+=2;
blk=*binsptr;
}
}
if(blk)
{
blksize=blk->size; /*nedblksize(blk);*/
assert(nedblksize(blk)>=blksize);
assert(blksize>=*size);
if(blk->next)
blk->next->prev=0;
*binsptr=blk->next;
if(!*binsptr)
binsptr[1]=0;
#ifdef FULLSANITYCHECKS
blk->magic=0;
#endif
assert(binsptr[0]!=blk && binsptr[1]!=blk);
assert(nedblksize(blk)>=sizeof(threadcacheblk) && nedblksize(blk)<=THREADCACHEMAX+CHUNK_OVERHEAD);
/*printf("malloc: %p, %p, %p, %lu\n", p, tc, blk, (long) size);*/
ret=(void *) blk;
}
++tc->mallocs;
if(ret)
{
assert(blksize>=*size);
++tc->successes;
tc->freeInCache-=blksize;
assert((long) tc->freeInCache>=0);
}
#if defined(DEBUG) && 0
if(!(tc->mallocs & 0xfff))
{
printf("*** threadcache=%u, mallocs=%u (%f), free=%u (%f), freeInCache=%u\n", (unsigned int) tc->threadid, tc->mallocs,
(float) tc->successes/tc->mallocs, tc->frees, (float) tc->successes/tc->frees, (unsigned int) tc->freeInCache);
}
#endif
#ifdef FULLSANITYCHECKS
tcfullsanitycheck(tc);
#endif
return ret;
}
static NOINLINE void ReleaseFreeInCache(nedpool *p, threadcache *tc, int mymspace) THROWSPEC
{
unsigned int age=THREADCACHEMAXFREESPACE/8192;
/*ACQUIRE_LOCK(&p->m[mymspace]->mutex);*/
while(age && tc->freeInCache>=THREADCACHEMAXFREESPACE)
{
RemoveCacheEntries(p, tc, age);
/*printf("*** Removing cache entries older than %u (%u)\n", age, (unsigned int) tc->freeInCache);*/
age>>=1;
}
/*RELEASE_LOCK(&p->m[mymspace]->mutex);*/
}
static void threadcache_free(nedpool *p, threadcache *tc, int mymspace, void *mem, size_t size) THROWSPEC
{
unsigned int bestsize;
unsigned int idx=size2binidx(size);
threadcacheblk **binsptr, *tck=(threadcacheblk *) mem;
assert(size>=sizeof(threadcacheblk) && size<=THREADCACHEMAX+CHUNK_OVERHEAD);
#ifdef DEBUG
{ /* Make sure this is a valid memory block */
mchunkptr p = mem2chunk(mem);
mstate fm = get_mstate_for(p);
if (!ok_magic(fm)) {
USAGE_ERROR_ACTION(fm, p);
return;
}
}
#endif
#ifdef FULLSANITYCHECKS
tcfullsanitycheck(tc);
#endif
/* Calculate best fit bin size */
bestsize=1<<(idx+4);
#if 0
/* Finer grained bin fit */
idx<<=1;
if(size>bestsize)
{
unsigned int biggerbestsize=bestsize+bestsize<<1;
if(size>=biggerbestsize)
{
idx++;
bestsize=biggerbestsize;
}
}
#endif
if(bestsize!=size) /* dlmalloc can round up, so we round down to preserve indexing */
size=bestsize;
binsptr=&tc->bins[idx*2];
assert(idx<=THREADCACHEMAXBINS);
if(tck==*binsptr)
{
fprintf(stderr, "Attempt to free already freed memory block %p - aborting!\n", tck);
abort();
}
#ifdef FULLSANITYCHECKS
tck->magic=*(unsigned int *) "NEDN";
#endif
tck->lastUsed=++tc->frees;
tck->size=(unsigned int) size;
tck->next=*binsptr;
tck->prev=0;
if(tck->next)
tck->next->prev=tck;
else
binsptr[1]=tck;
assert(!*binsptr || (*binsptr)->size==tck->size);
*binsptr=tck;
assert(tck==tc->bins[idx*2]);
assert(tc->bins[idx*2+1]==tck || binsptr[0]->next->prev==tck);
/*printf("free: %p, %p, %p, %lu\n", p, tc, mem, (long) size);*/
tc->freeInCache+=size;
#ifdef FULLSANITYCHECKS
tcfullsanitycheck(tc);
#endif
#if 1
if(tc->freeInCache>=THREADCACHEMAXFREESPACE)
ReleaseFreeInCache(p, tc, mymspace);
#endif
}
static NOINLINE int InitPool(nedpool *p, size_t capacity, int threads) THROWSPEC
{ /* threads is -1 for system pool */
ensure_initialization();
ACQUIRE_MALLOC_GLOBAL_LOCK();
if(p->threads) goto done;
if(INITIAL_LOCK(&p->mutex)) goto err;
if(TLSALLOC(&p->mycache)) goto err;
if(!(p->m[0]=(mstate) create_mspace(capacity, 1))) goto err;
p->m[0]->extp=p;
p->threads=(threads<1 || threads>MAXTHREADSINPOOL) ? MAXTHREADSINPOOL : threads;
done:
RELEASE_MALLOC_GLOBAL_LOCK();
return 1;
err:
if(threads<0)
abort(); /* If you can't allocate for system pool, we're screwed */
DestroyCaches(p);
if(p->m[0])
{
destroy_mspace(p->m[0]);
p->m[0]=0;
}
if(p->mycache)
{
if(TLSFREE(p->mycache)) abort();
p->mycache=0;
}
RELEASE_MALLOC_GLOBAL_LOCK();
return 0;
}
static NOINLINE mstate FindMSpace(nedpool *p, threadcache *tc, int *lastUsed, size_t size) THROWSPEC
{ /* Gets called when thread's last used mspace is in use. The strategy
is to run through the list of all available mspaces looking for an
unlocked one and if we fail, we create a new one so long as we don't
exceed p->threads */
int n, end;
for(n=end=*lastUsed+1; p->m[n]; end=++n)
{
if(TRY_LOCK(&p->m[n]->mutex)) goto found;
}
for(n=0; n<*lastUsed && p->m[n]; n++)
{
if(TRY_LOCK(&p->m[n]->mutex)) goto found;
}
if(end<p->threads)
{
mstate temp;
if(!(temp=(mstate) create_mspace(size, 1)))
goto badexit;
/* Now we're ready to modify the lists, we lock */
ACQUIRE_LOCK(&p->mutex);
while(p->m[end] && end<p->threads)
end++;
if(end>=p->threads)
{ /* Drat, must destroy it now */
RELEASE_LOCK(&p->mutex);
destroy_mspace((mspace) temp);
goto badexit;
}
/* We really want to make sure this goes into memory now but we
have to be careful of breaking aliasing rules, so write it twice */
{
volatile struct malloc_state **_m=(volatile struct malloc_state **) &p->m[end];
*_m=(p->m[end]=temp);
}
Add custom memory allocator to MinGW and MacOS builds The standard allocator on Windows is pretty bad prior to Windows Vista, and nedmalloc is better than the modified dlmalloc provided with newer versions of the MinGW libc. NedMalloc stats in Git ---------------------- All results are the best result out of 3 runs. The benchmarks have been done on different hardware, so the repack times are not comparable. These benchmarks are all based on 'git repack -adf' on the Linux kernel. XP ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 3.4.2 (1T) 00:12:28.422 3.4.2 + nedmalloc (1T) 00:07:25.437 1.68x 3.4.5 (1T) 00:12:20.718 3.4.5 + nedmalloc (1T) 00:07:24.809 1.67x 4.3.3-tdm (1T) 00:12:01.843 4.3.3-tdm + nedmalloc (1T) 00:07:16.468 1.65x 4.3.3-tdm (2T) 00:07:35.062 4.3.3-tdm + nedmalloc (2T) 00:04:57.874 1.54x Vista ----------------------------------------------- MinGW Threads Total Time Speed ----------------------------------------------- 4.3.3-tdm (1T) 00:07:40.844 4.3.3-tdm + nedmalloc (1T) 00:07:17.548 1.05x 4.3.3-tdm (2T) 00:05:33.746 4.3.3-tdm + nedmalloc (2T) 00:05:27.334 1.02x Mac Mini ----------------------------------------------- GCC Threads Total Time Speed ----------------------------------------------- i686-darwin9-4.0.1 (2T) 00:09:57.346 i686-darwin9-4.0.1+ned (2T) 00:08:51.072 1.12x Signed-off-by: Marius Storm-Olsen <marius@trolltech.com> Signed-off-by: Steffen Prohaska <prohaska@zib.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-05-31 18:15:23 +02:00
ACQUIRE_LOCK(&p->m[end]->mutex);
/*printf("Created mspace idx %d\n", end);*/
RELEASE_LOCK(&p->mutex);
n=end;
goto found;
}
/* Let it lock on the last one it used */
badexit:
ACQUIRE_LOCK(&p->m[*lastUsed]->mutex);
return p->m[*lastUsed];
found:
*lastUsed=n;
if(tc)
tc->mymspace=n;
else
{
if(TLSSET(p->mycache, (void *)(size_t)(-(n+1)))) abort();
}
return p->m[n];
}
nedpool *nedcreatepool(size_t capacity, int threads) THROWSPEC
{
nedpool *ret;
if(!(ret=(nedpool *) nedpcalloc(0, 1, sizeof(nedpool)))) return 0;
if(!InitPool(ret, capacity, threads))
{
nedpfree(0, ret);
return 0;
}
return ret;
}
void neddestroypool(nedpool *p) THROWSPEC
{
int n;
ACQUIRE_LOCK(&p->mutex);
DestroyCaches(p);
for(n=0; p->m[n]; n++)
{
destroy_mspace(p->m[n]);
p->m[n]=0;
}
RELEASE_LOCK(&p->mutex);
if(TLSFREE(p->mycache)) abort();
nedpfree(0, p);
}
void nedpsetvalue(nedpool *p, void *v) THROWSPEC
{
if(!p) { p=&syspool; if(!syspool.threads) InitPool(&syspool, 0, -1); }
p->uservalue=v;
}
void *nedgetvalue(nedpool **p, void *mem) THROWSPEC
{
nedpool *np=0;
mchunkptr mcp=mem2chunk(mem);
mstate fm;
if(!(is_aligned(chunk2mem(mcp))) && mcp->head != FENCEPOST_HEAD) return 0;
if(!cinuse(mcp)) return 0;
if(!next_pinuse(mcp)) return 0;
if(!is_mmapped(mcp) && !pinuse(mcp))
{
if(next_chunk(prev_chunk(mcp))!=mcp) return 0;
}
fm=get_mstate_for(mcp);
if(!ok_magic(fm)) return 0;
if(!ok_address(fm, mcp)) return 0;
if(!fm->extp) return 0;
np=(nedpool *) fm->extp;
if(p) *p=np;
return np->uservalue;
}
void neddisablethreadcache(nedpool *p) THROWSPEC
{
int mycache;
if(!p)
{
p=&syspool;
if(!syspool.threads) InitPool(&syspool, 0, -1);
}
mycache=(int)(size_t) TLSGET(p->mycache);
if(!mycache)
{ /* Set to mspace 0 */
if(TLSSET(p->mycache, (void *)-1)) abort();
}
else if(mycache>0)
{ /* Set to last used mspace */
threadcache *tc=p->caches[mycache-1];
#if defined(DEBUG)
printf("Threadcache utilisation: %lf%% in cache with %lf%% lost to other threads\n",
100.0*tc->successes/tc->mallocs, 100.0*((double) tc->mallocs-tc->frees)/tc->mallocs);
#endif
if(TLSSET(p->mycache, (void *)(size_t)(-tc->mymspace))) abort();
tc->frees++;
RemoveCacheEntries(p, tc, 0);
assert(!tc->freeInCache);
tc->mymspace=-1;
tc->threadid=0;
mspace_free(0, p->caches[mycache-1]);
p->caches[mycache-1]=0;
}
}
#define GETMSPACE(m,p,tc,ms,s,action) \
do \
{ \
mstate m = GetMSpace((p),(tc),(ms),(s)); \
action; \
RELEASE_LOCK(&m->mutex); \
} while (0)
static FORCEINLINE mstate GetMSpace(nedpool *p, threadcache *tc, int mymspace, size_t size) THROWSPEC
{ /* Returns a locked and ready for use mspace */
mstate m=p->m[mymspace];
assert(m);
if(!TRY_LOCK(&p->m[mymspace]->mutex)) m=FindMSpace(p, tc, &mymspace, size);\
/*assert(IS_LOCKED(&p->m[mymspace]->mutex));*/
return m;
}
static FORCEINLINE void GetThreadCache(nedpool **p, threadcache **tc, int *mymspace, size_t *size) THROWSPEC
{
int mycache;
if(size && *size<sizeof(threadcacheblk)) *size=sizeof(threadcacheblk);
if(!*p)
{
*p=&syspool;
if(!syspool.threads) InitPool(&syspool, 0, -1);
}
mycache=(int)(size_t) TLSGET((*p)->mycache);
if(mycache>0)
{
*tc=(*p)->caches[mycache-1];
*mymspace=(*tc)->mymspace;
}
else if(!mycache)
{
*tc=AllocCache(*p);
if(!*tc)
{ /* Disable */
if(TLSSET((*p)->mycache, (void *)-1)) abort();
*mymspace=0;
}
else
*mymspace=(*tc)->mymspace;
}
else
{
*tc=0;
*mymspace=-mycache-1;
}
assert(*mymspace>=0);
assert((long)(size_t)CURRENT_THREAD==(*tc)->threadid);
#ifdef FULLSANITYCHECKS
if(*tc)
{
if(*(unsigned int *)"NEDMALC1"!=(*tc)->magic1 || *(unsigned int *)"NEDMALC2"!=(*tc)->magic2)
{
abort();
}
}
#endif
}
void * nedpmalloc(nedpool *p, size_t size) THROWSPEC
{
void *ret=0;
threadcache *tc;
int mymspace;
GetThreadCache(&p, &tc, &mymspace, &size);
#if THREADCACHEMAX
if(tc && size<=THREADCACHEMAX)
{ /* Use the thread cache */
ret=threadcache_malloc(p, tc, &size);
}
#endif
if(!ret)
{ /* Use this thread's mspace */
GETMSPACE(m, p, tc, mymspace, size,
ret=mspace_malloc(m, size));
}
return ret;
}
void * nedpcalloc(nedpool *p, size_t no, size_t size) THROWSPEC
{
size_t rsize=size*no;
void *ret=0;
threadcache *tc;
int mymspace;
GetThreadCache(&p, &tc, &mymspace, &rsize);
#if THREADCACHEMAX
if(tc && rsize<=THREADCACHEMAX)
{ /* Use the thread cache */
if((ret=threadcache_malloc(p, tc, &rsize)))
memset(ret, 0, rsize);
}
#endif
if(!ret)
{ /* Use this thread's mspace */
GETMSPACE(m, p, tc, mymspace, rsize,
ret=mspace_calloc(m, 1, rsize));
}
return ret;
}
void * nedprealloc(nedpool *p, void *mem, size_t size) THROWSPEC
{
void *ret=0;
threadcache *tc;
int mymspace;
if(!mem) return nedpmalloc(p, size);
GetThreadCache(&p, &tc, &mymspace, &size);
#if THREADCACHEMAX
if(tc && size && size<=THREADCACHEMAX)
{ /* Use the thread cache */
size_t memsize=nedblksize(mem);
assert(memsize);
if((ret=threadcache_malloc(p, tc, &size)))
{
memcpy(ret, mem, memsize<size ? memsize : size);
if(memsize<=THREADCACHEMAX)
threadcache_free(p, tc, mymspace, mem, memsize);
else
mspace_free(0, mem);
}
}
#endif
if(!ret)
{ /* Reallocs always happen in the mspace they happened in, so skip
locking the preferred mspace for this thread */
ret=mspace_realloc(0, mem, size);
}
return ret;
}
void nedpfree(nedpool *p, void *mem) THROWSPEC
{ /* Frees always happen in the mspace they happened in, so skip
locking the preferred mspace for this thread */
threadcache *tc;
int mymspace;
size_t memsize;
assert(mem);
GetThreadCache(&p, &tc, &mymspace, 0);
#if THREADCACHEMAX
memsize=nedblksize(mem);
assert(memsize);
if(mem && tc && memsize<=(THREADCACHEMAX+CHUNK_OVERHEAD))
threadcache_free(p, tc, mymspace, mem, memsize);
else
#endif
mspace_free(0, mem);
}
void * nedpmemalign(nedpool *p, size_t alignment, size_t bytes) THROWSPEC
{
void *ret;
threadcache *tc;
int mymspace;
GetThreadCache(&p, &tc, &mymspace, &bytes);
{ /* Use this thread's mspace */
GETMSPACE(m, p, tc, mymspace, bytes,
ret=mspace_memalign(m, alignment, bytes));
}
return ret;
}
#if !NO_MALLINFO
struct mallinfo nedpmallinfo(nedpool *p) THROWSPEC
{
int n;
struct mallinfo ret={0};
if(!p) { p=&syspool; if(!syspool.threads) InitPool(&syspool, 0, -1); }
for(n=0; p->m[n]; n++)
{
struct mallinfo t=mspace_mallinfo(p->m[n]);
ret.arena+=t.arena;
ret.ordblks+=t.ordblks;
ret.hblkhd+=t.hblkhd;
ret.usmblks+=t.usmblks;
ret.uordblks+=t.uordblks;
ret.fordblks+=t.fordblks;
ret.keepcost+=t.keepcost;
}
return ret;
}
#endif
int nedpmallopt(nedpool *p, int parno, int value) THROWSPEC
{
return mspace_mallopt(parno, value);
}
int nedpmalloc_trim(nedpool *p, size_t pad) THROWSPEC
{
int n, ret=0;
if(!p) { p=&syspool; if(!syspool.threads) InitPool(&syspool, 0, -1); }
for(n=0; p->m[n]; n++)
{
ret+=mspace_trim(p->m[n], pad);
}
return ret;
}
void nedpmalloc_stats(nedpool *p) THROWSPEC
{
int n;
if(!p) { p=&syspool; if(!syspool.threads) InitPool(&syspool, 0, -1); }
for(n=0; p->m[n]; n++)
{
mspace_malloc_stats(p->m[n]);
}
}
size_t nedpmalloc_footprint(nedpool *p) THROWSPEC
{
size_t ret=0;
int n;
if(!p) { p=&syspool; if(!syspool.threads) InitPool(&syspool, 0, -1); }
for(n=0; p->m[n]; n++)
{
ret+=mspace_footprint(p->m[n]);
}
return ret;
}
void **nedpindependent_calloc(nedpool *p, size_t elemsno, size_t elemsize, void **chunks) THROWSPEC
{
void **ret;
threadcache *tc;
int mymspace;
GetThreadCache(&p, &tc, &mymspace, &elemsize);
GETMSPACE(m, p, tc, mymspace, elemsno*elemsize,
ret=mspace_independent_calloc(m, elemsno, elemsize, chunks));
return ret;
}
void **nedpindependent_comalloc(nedpool *p, size_t elems, size_t *sizes, void **chunks) THROWSPEC
{
void **ret;
threadcache *tc;
int mymspace;
size_t i, *adjustedsizes=(size_t *) alloca(elems*sizeof(size_t));
if(!adjustedsizes) return 0;
for(i=0; i<elems; i++)
adjustedsizes[i]=sizes[i]<sizeof(threadcacheblk) ? sizeof(threadcacheblk) : sizes[i];
GetThreadCache(&p, &tc, &mymspace, 0);
GETMSPACE(m, p, tc, mymspace, 0,
ret=mspace_independent_comalloc(m, elems, adjustedsizes, chunks));
return ret;
}
#ifdef OVERRIDE_STRDUP
/*
* This implementation is purely there to override the libc version, to
* avoid a crash due to allocation and free on different 'heaps'.
*/
char *strdup(const char *s1)
{
char *s2 = 0;
if (s1) {
s2 = malloc(strlen(s1) + 1);
strcpy(s2, s1);
}
return s2;
}
#endif
#if defined(__cplusplus)
}
#endif