summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorScott Shawcroft <scott@tannewt.org>2020-02-11 17:06:24 -0800
committerScott Shawcroft <scott@tannewt.org>2020-02-11 17:06:24 -0800
commit36e6cc8febf7f744f61772d29eed68d9ebb57200 (patch)
treed7ca610709815421707a1b8cee4b2ec06b25af8e
parenta4d197a6efeb8af9fc59be95689b2c703c5f9cdb (diff)
Track first free atb for multiple block sizes.
This speeds up cases where no single block allocations happen while Python code is allocating other block sizes.
-rwxr-xr-xpy/gc.c81
-rwxr-xr-xpy/mpconfig.h8
-rw-r--r--py/mpstate.h2
3 files changed, 59 insertions, 32 deletions
diff --git a/py/gc.c b/py/gc.c
index fef67f61b..271bc9462 100755
--- a/py/gc.c
+++ b/py/gc.c
@@ -150,9 +150,13 @@ void gc_init(void *start, void *end) {
#endif
// Set first free ATB index to the start of the heap.
- MP_STATE_MEM(gc_first_free_atb_index) = 0;
+ for (size_t i = 0; i < MICROPY_ATB_INDICES; i++) {
+ MP_STATE_MEM(gc_first_free_atb_index)[i] = 0;
+ }
+
// Set last free ATB index to the end of the heap.
MP_STATE_MEM(gc_last_free_atb_index) = MP_STATE_MEM(gc_alloc_table_byte_len) - 1;
+
// Set the lowest long lived ptr to the end of the heap to start. This will be lowered as long
// lived objects are allocated.
MP_STATE_MEM(gc_lowest_long_lived_ptr) = (void*) PTR_FROM_BLOCK(MP_STATE_MEM(gc_alloc_table_byte_len * BLOCKS_PER_ATB));
@@ -387,7 +391,9 @@ void gc_collect_root(void **ptrs, size_t len) {
void gc_collect_end(void) {
gc_deal_with_stack_overflow();
gc_sweep();
- MP_STATE_MEM(gc_first_free_atb_index) = 0;
+ for (size_t i = 0; i < MICROPY_ATB_INDICES; i++) {
+ MP_STATE_MEM(gc_first_free_atb_index)[i] = 0;
+ }
MP_STATE_MEM(gc_last_free_atb_index) = MP_STATE_MEM(gc_alloc_table_byte_len) - 1;
MP_STATE_MEM(gc_lock_depth)--;
GC_EXIT();
@@ -513,14 +519,16 @@ void *gc_alloc(size_t n_bytes, bool has_finaliser, bool long_lived) {
size_t crossover_block = BLOCK_FROM_PTR(MP_STATE_MEM(gc_lowest_long_lived_ptr));
while (keep_looking) {
int8_t direction = 1;
- size_t start = MP_STATE_MEM(gc_first_free_atb_index);
+ size_t bucket = MIN(n_blocks, MICROPY_ATB_INDICES) - 1;
+ size_t first_free = MP_STATE_MEM(gc_first_free_atb_index)[bucket];
+ size_t start = first_free;
if (long_lived) {
direction = -1;
start = MP_STATE_MEM(gc_last_free_atb_index);
}
n_free = 0;
// look for a run of n_blocks available blocks
- for (size_t i = start; keep_looking && MP_STATE_MEM(gc_first_free_atb_index) <= i && i <= MP_STATE_MEM(gc_last_free_atb_index); i += direction) {
+ for (size_t i = start; keep_looking && first_free <= i && i <= MP_STATE_MEM(gc_last_free_atb_index); i += direction) {
byte a = MP_STATE_MEM(gc_alloc_table_start)[i];
// Four ATB states are packed into a single byte.
int j = 0;
@@ -565,22 +573,24 @@ void *gc_alloc(size_t n_bytes, bool has_finaliser, bool long_lived) {
// Found free space ending at found_block inclusive.
// Also, set last free ATB index to block after last block we found, for start of
- // next scan. To reduce fragmentation, we only do this if we were looking
- // for a single free block, which guarantees that there are no free blocks
- // before this one. Also, whenever we free or shrink a block we must check
- // if this index needs adjusting (see gc_realloc and gc_free).
+ // next scan. Also, whenever we free or shrink a block we must check if this index needs
+ // adjusting (see gc_realloc and gc_free).
if (!long_lived) {
end_block = found_block;
start_block = found_block - n_free + 1;
- if (n_blocks == 1) {
- MP_STATE_MEM(gc_first_free_atb_index) = (found_block + 1) / BLOCKS_PER_ATB;
+ if (n_blocks < MICROPY_ATB_INDICES) {
+ size_t next_free_atb = (found_block + n_blocks) / BLOCKS_PER_ATB;
+ // Update all atb indices for larger blocks too.
+ for (size_t i = n_blocks - 1; i < MICROPY_ATB_INDICES; i++) {
+ MP_STATE_MEM(gc_first_free_atb_index)[i] = next_free_atb;
+ }
}
} else {
start_block = found_block;
end_block = found_block + n_free - 1;
- if (n_blocks == 1) {
- MP_STATE_MEM(gc_last_free_atb_index) = (found_block - 1) / BLOCKS_PER_ATB;
- }
+ // Always update the bounds of the long lived area because we assume it is contiguous. (It
+ // can still be reset by a sweep.)
+ MP_STATE_MEM(gc_last_free_atb_index) = (found_block - 1) / BLOCKS_PER_ATB;
}
#ifdef LOG_HEAP_ACTIVITY
@@ -676,30 +686,37 @@ void gc_free(void *ptr) {
}
// get the GC block number corresponding to this pointer
assert(VERIFY_PTR(ptr));
- size_t block = BLOCK_FROM_PTR(ptr);
- assert(ATB_GET_KIND(block) == AT_HEAD);
+ size_t start_block = BLOCK_FROM_PTR(ptr);
+ assert(ATB_GET_KIND(start_block) == AT_HEAD);
#if MICROPY_ENABLE_FINALISER
- FTB_CLEAR(block);
+ FTB_CLEAR(start_block);
#endif
- // set the last_free pointer to this block if it's earlier in the heap
- if (block / BLOCKS_PER_ATB < MP_STATE_MEM(gc_first_free_atb_index)) {
- MP_STATE_MEM(gc_first_free_atb_index) = block / BLOCKS_PER_ATB;
- }
- if (block / BLOCKS_PER_ATB > MP_STATE_MEM(gc_last_free_atb_index)) {
- MP_STATE_MEM(gc_last_free_atb_index) = block / BLOCKS_PER_ATB;
- }
-
// free head and all of its tail blocks
- #ifdef LOG_HEAP_ACTIVITY
- gc_log_change(block, 0);
- #endif
+ #ifdef LOG_HEAP_ACTIVITY
+ gc_log_change(start_block, 0);
+ #endif
+ size_t block = start_block;
do {
ATB_ANY_TO_FREE(block);
block += 1;
} while (ATB_GET_KIND(block) == AT_TAIL);
+ // Update the first free pointer for our size only. Not much calls gc_free directly so there
+ // is decent chance we'll want to allocate this size again. By only updating the specific
+ // size we don't risk something smaller fitting in.
+ size_t n_blocks = block - start_block;
+ size_t bucket = MIN(n_blocks, MICROPY_ATB_INDICES) - 1;
+ size_t new_free_atb = start_block / BLOCKS_PER_ATB;
+ if (new_free_atb < MP_STATE_MEM(gc_first_free_atb_index)[bucket]) {
+ MP_STATE_MEM(gc_first_free_atb_index)[bucket] = new_free_atb;
+ }
+ // set the last_free pointer to this block if it's earlier in the heap
+ if (new_free_atb > MP_STATE_MEM(gc_last_free_atb_index)) {
+ MP_STATE_MEM(gc_last_free_atb_index) = new_free_atb;
+ }
+
GC_EXIT();
#if EXTENSIVE_HEAP_PROFILING
@@ -870,11 +887,13 @@ void *gc_realloc(void *ptr_in, size_t n_bytes, bool allow_move) {
}
// set the last_free pointer to end of this block if it's earlier in the heap
- if ((block + new_blocks) / BLOCKS_PER_ATB < MP_STATE_MEM(gc_first_free_atb_index)) {
- MP_STATE_MEM(gc_first_free_atb_index) = (block + new_blocks) / BLOCKS_PER_ATB;
+ size_t new_free_atb = (block + new_blocks) / BLOCKS_PER_ATB;
+ size_t bucket = MIN(n_blocks - new_blocks, MICROPY_ATB_INDICES) - 1;
+ if (new_free_atb < MP_STATE_MEM(gc_first_free_atb_index)[bucket]) {
+ MP_STATE_MEM(gc_first_free_atb_index)[bucket] = new_free_atb;
}
- if ((block + new_blocks) / BLOCKS_PER_ATB > MP_STATE_MEM(gc_last_free_atb_index)) {
- MP_STATE_MEM(gc_last_free_atb_index) = (block + new_blocks) / BLOCKS_PER_ATB;
+ if (new_free_atb > MP_STATE_MEM(gc_last_free_atb_index)) {
+ MP_STATE_MEM(gc_last_free_atb_index) = new_free_atb;
}
GC_EXIT();
diff --git a/py/mpconfig.h b/py/mpconfig.h
index a0d211bfa..1512c7d3a 100755
--- a/py/mpconfig.h
+++ b/py/mpconfig.h
@@ -244,6 +244,14 @@
#define alloca(x) m_malloc(x)
#endif
+// Number of atb indices to cache. Allocations of fewer blocks will be faster
+// because the search will be accelerated by the index cache. This only applies
+// to short lived allocations because we assume the long lived allocations are
+// contiguous.
+#ifndef MICROPY_ATB_INDICES
+#define MICROPY_ATB_INDICES (8)
+#endif
+
/*****************************************************************************/
/* MicroPython emitters */
diff --git a/py/mpstate.h b/py/mpstate.h
index a3d7e5dcc..a5815776a 100644
--- a/py/mpstate.h
+++ b/py/mpstate.h
@@ -92,7 +92,7 @@ typedef struct _mp_state_mem_t {
size_t gc_alloc_threshold;
#endif
- size_t gc_first_free_atb_index;
+ size_t gc_first_free_atb_index[MICROPY_ATB_INDICES];
size_t gc_last_free_atb_index;
#if MICROPY_PY_GC_COLLECT_RETVAL