Merge branch 'dev' into dev-exp

This commit is contained in:
daan 2019-08-11 14:46:53 -07:00
commit 0c912445c4
5 changed files with 84 additions and 30 deletions

View file

@ -91,7 +91,7 @@ if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang|GNU")
if(CMAKE_C_COMPILER_ID MATCHES "GNU")
list(APPEND mi_cflags -Wno-invalid-memory-model)
list(APPEND mi_cflags -fvisibility=hidden)
list(APPEND mi_cflags -fbranch-target-load-optimize )
list(APPEND mi_cflags -fbranch-target-load-optimize)
endif()
endif()

View file

@ -167,9 +167,9 @@ typedef struct mi_page_s {
#if MI_SECURE
uintptr_t cookie; // random cookie to encode the free lists
#endif
size_t used; // number of blocks in use (including blocks in `local_free` and `thread_free`)
mi_page_flags_t flags; // threadid:62 | has_aligned:1 | in_full:1
size_t used; // number of blocks in use (including blocks in `local_free` and `thread_free`)
mi_block_t* local_free; // list of deferred free blocks by this thread (migrates to `free`)
volatile uintptr_t thread_freed; // at least this number of blocks are in `thread_free`
volatile mi_thread_free_t thread_free; // list of deferred free blocks freed by other threads
@ -383,7 +383,8 @@ typedef struct mi_segments_tld_s {
} mi_segments_tld_t;
// OS thread local data
typedef struct mi_os_tld_s {
typedef struct mi_os_tld_s {
size_t region_idx; // start point for next allocation
mi_stats_t* stats; // points to tld stats
} mi_os_tld_t;

View file

@ -72,7 +72,7 @@ extern inline void* mi_heap_malloc(mi_heap_t* heap, size_t size) mi_attr_noexcep
void* p;
if (mi_likely(size <= MI_SMALL_SIZE_MAX)) {
p = mi_heap_malloc_small(heap, size);
}
}
else {
p = _mi_malloc_generic(heap, size);
}
@ -200,16 +200,16 @@ static void mi_decl_noinline mi_free_generic(const mi_segment_t* segment, mi_pag
// Free a block
void mi_free(void* p) mi_attr_noexcept
{
{
#if (MI_DEBUG>0)
if (mi_unlikely(((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0)) {
_mi_error_message("trying to free an invalid (unaligned) pointer: %p\n", p);
return;
}
#endif
const mi_segment_t* const segment = _mi_ptr_segment(p);
if (segment == NULL) return; // checks for (p==NULL)
if (segment == NULL) return; // checks for (p==NULL)
#if (MI_DEBUG>0)
if (mi_unlikely(!mi_is_in_heap_region(p))) {
@ -224,8 +224,8 @@ void mi_free(void* p) mi_attr_noexcept
return;
}
#endif
mi_page_t* page = _mi_segment_page_of(segment, p);
mi_page_t* const page = _mi_segment_page_of(segment, p);
#if (MI_STAT>1)
mi_heap_t* heap = mi_heap_get_default();
@ -236,11 +236,11 @@ void mi_free(void* p) mi_attr_noexcept
// huge page stat is accounted for in `_mi_page_retire`
#endif
uintptr_t tid = _mi_thread_id();
const uintptr_t tid = _mi_thread_id();
if (mi_likely(tid == page->flags)) { // if equal, the thread id matches and it is not a full page, nor has aligned blocks
// local, and not full or aligned
mi_block_t* block = (mi_block_t*)p;
mi_block_set_next(page, block, page->local_free);
mi_block_set_next(page, block, page->local_free);
page->local_free = block;
page->used--;
if (mi_unlikely(mi_page_all_free(page))) { _mi_page_retire(page); }

View file

@ -12,16 +12,16 @@ terms of the MIT license. A copy of the license can be found in the file
// Empty page used to initialize the small free pages array
const mi_page_t _mi_page_empty = {
0, false, false, false, 0, 0,
0, false, false, false, 0, 0,
NULL, // free
#if MI_SECURE
0,
#endif
0, {0}, // used, flags
0, 0, // flags, used
NULL, 0, 0,
0, NULL, NULL, NULL
#if (MI_INTPTR_SIZE==8 && MI_SECURE==0)
, { NULL }
, { NULL }
#endif
};
@ -97,7 +97,7 @@ static mi_tld_t tld_main = {
0,
&_mi_heap_main,
{ { NULL, NULL }, {NULL ,NULL}, 0, 0, 0, 0, 0, 0, NULL, tld_main_stats }, // segments
{ tld_main_stats }, // os
{ 0, tld_main_stats }, // os
{ MI_STATS_NULL } // stats
};
@ -421,7 +421,7 @@ static void mi_process_load(void) {
// show message from the redirector (if present)
const char* msg = NULL;
mi_allocator_init(&msg);
if (msg != NULL) _mi_verbose_message(msg);
if (msg != NULL) _mi_verbose_message(msg);
}
// Initialize the process; called by thread_init or the process loader
@ -433,7 +433,7 @@ void mi_process_init(void) mi_attr_noexcept {
// when using dynamic linking with interpose.
mi_heap_t* h = _mi_heap_default;
_mi_process_is_initialized = true;
_mi_heap_main.thread_id = _mi_thread_id();
_mi_verbose_message("process init: 0x%zx\n", _mi_heap_main.thread_id);
uintptr_t random = _mi_random_init(_mi_heap_main.thread_id) ^ (uintptr_t)h;
@ -442,7 +442,7 @@ void mi_process_init(void) mi_attr_noexcept {
#endif
_mi_heap_main.random = _mi_random_shuffle(random);
mi_process_setup_auto_thread_done();
_mi_os_init();
_mi_os_init();
#if (MI_DEBUG)
_mi_verbose_message("debug level : %d\n", MI_DEBUG);
#endif
@ -478,7 +478,7 @@ static void mi_process_done(void) {
__declspec(dllexport) BOOL WINAPI DllMain(HINSTANCE inst, DWORD reason, LPVOID reserved) {
UNUSED(reserved);
UNUSED(inst);
if (reason==DLL_PROCESS_ATTACH) {
if (reason==DLL_PROCESS_ATTACH) {
mi_process_load();
}
else if (reason==DLL_THREAD_DETACH) {

View file

@ -79,7 +79,6 @@ typedef struct mem_region_s {
static mem_region_t regions[MI_REGION_MAX];
static volatile size_t regions_count = 0; // allocated regions
static volatile uintptr_t region_next_idx = 0; // good place to start searching
/* ----------------------------------------------------------------------------
@ -180,12 +179,45 @@ static bool mi_region_commit_blocks(mem_region_t* region, size_t idx, size_t bit
}
// and return the allocation
mi_atomic_write(&region_next_idx,idx); // next search from here
*p = blocks_start;
*id = (idx*MI_REGION_MAP_BITS) + bitidx;
return true;
}
// Use bit scan forward to quickly find the first zero bit if it is available
#if defined(_MSC_VER)
#define MI_HAVE_BITSCAN
#include <intrin.h>
static inline size_t mi_bsf(uintptr_t x) {
if (x==0) return 8*MI_INTPTR_SIZE;
DWORD idx;
#if (MI_INTPTR_SIZE==8)
_BitScanForward64(&idx, x);
#else
_BitScanForward(&idx, x);
#endif
return idx;
}
static inline size_t mi_bsr(uintptr_t x) {
if (x==0) return 8*MI_INTPTR_SIZE;
DWORD idx;
#if (MI_INTPTR_SIZE==8)
_BitScanReverse64(&idx, x);
#else
_BitScanReverse(&idx, x);
#endif
return idx;
}
#elif defined(__GNUC__) || defined(__clang__)
#define MI_HAVE_BITSCAN
static inline size_t mi_bsf(uintptr_t x) {
return (x==0 ? 8*MI_INTPTR_SIZE : __builtin_ctzl(x));
}
static inline size_t mi_bsr(uintptr_t x) {
return (x==0 ? 8*MI_INTPTR_SIZE : (8*MI_INTPTR_SIZE - 1) - __builtin_clzl(x));
}
#endif
// Allocate `blocks` in a `region` at `idx` of a given `size`.
// Returns `false` on an error (OOM); `true` otherwise. `p` and `id` are only written
// if the blocks were successfully claimed so ensure they are initialized to NULL/SIZE_MAX before the call.
@ -194,21 +226,28 @@ static bool mi_region_alloc_blocks(mem_region_t* region, size_t idx, size_t bloc
{
mi_assert_internal(p != NULL && id != NULL);
mi_assert_internal(blocks < MI_REGION_MAP_BITS);
const uintptr_t mask = mi_region_block_mask(blocks,0);
const uintptr_t mask = mi_region_block_mask(blocks, 0);
const size_t bitidx_max = MI_REGION_MAP_BITS - blocks;
// scan linearly for a free range of zero bits
uintptr_t map = mi_atomic_read(&region->map);
uintptr_t m = mask; // the mask shifted by bitidx
for(size_t bitidx = 0; bitidx <= bitidx_max; bitidx++, m <<= 1) {
#ifdef MI_HAVE_BITSCAN
size_t bitidx = mi_bsf(~map); // quickly find the first zero bit if possible
#else
size_t bitidx = 0; // otherwise start at 0
#endif
uintptr_t m = (mask << bitidx); // invariant: m == mask shifted by bitidx
// scan linearly for a free range of zero bits
while(bitidx <= bitidx_max) {
if ((map & m) == 0) { // are the mask bits free at bitidx?
mi_assert_internal((m >> bitidx) == mask); // no overflow?
uintptr_t newmap = map | m;
mi_assert_internal((newmap^map) >> bitidx == mask);
if (!mi_atomic_compare_exchange(&region->map, newmap, map)) {
// no success, another thread claimed concurrently.. keep going
map = mi_atomic_read(&region->map);
map = mi_atomic_read(&region->map);
continue;
}
else {
// success, we claimed the bits
@ -216,6 +255,17 @@ static bool mi_region_alloc_blocks(mem_region_t* region, size_t idx, size_t bloc
return mi_region_commit_blocks(region, idx, bitidx, blocks, size, commit, p, id, tld);
}
}
else {
// on to the next bit range
#ifdef MI_HAVE_BITSCAN
size_t shift = (blocks == 1 ? 1 : mi_bsr(map & m) - bitidx + 1);
mi_assert_internal(shift > 0 && shift <= blocks);
#else
size_t shift = 1;
#endif
bitidx += shift;
m <<= shift;
}
}
// no error, but also no bits found
return true;
@ -267,7 +317,7 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool commit, size_t*
// find a range of free blocks
void* p = NULL;
size_t count = mi_atomic_read(&regions_count);
size_t idx = mi_atomic_read(&region_next_idx);
size_t idx = tld->region_idx; // start index is per-thread to reduce contention
for (size_t visited = 0; visited < count; visited++, idx++) {
if (idx >= count) idx = 0; // wrap around
if (!mi_region_try_alloc_blocks(idx, blocks, size, commit, &p, id, tld)) return NULL; // error
@ -286,6 +336,9 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool commit, size_t*
// we could not find a place to allocate, fall back to the os directly
p = _mi_os_alloc_aligned(size, alignment, commit, tld);
}
else {
tld->region_idx = idx; // next start of search
}
mi_assert_internal( p == NULL || (uintptr_t)p % alignment == 0);
return p;