Stowage/cpython
2
0
Fork 0
mirror of https://github.com/python/cpython.git synced 2026-04-14 15:50:50 +00:00
Code Issues Projects Releases Packages Wiki Activity

gh-144888: JIT executor bloom filter wide-type optimization and function Inlining (GH-146114)

Browse source
This commit is contained in:
Hai Zhu 2026-03-19 00:58:14 +08:00 • committed by GitHub
parent 2c6afb935a
commit 6fe91a9e80
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 92 additions and 89 deletions
Download patch file Download diff file Expand all files Collapse all files

85
Python/optimizer.c
View file

@ -1601,91 +1601,6 @@ uop_optimize(
* Executor management
****************************************/
/* We use a bloomfilter with k = 6, m = 256
* The choice of k and the following constants
* could do with a more rigorous analysis,
* but here is a simple analysis:
*
* We want to keep the false positive rate low.
* For n = 5 (a trace depends on 5 objects),
* we expect 30 bits set, giving a false positive
* rate of (30/256)**6 == 2.5e-6 which is plenty
* good enough.
*
* However with n = 10 we expect 60 bits set (worst case),
* giving a false positive of (60/256)**6 == 0.0001
*
* We choose k = 6, rather than a higher number as
* it means the false positive rate grows slower for high n.
*
* n = 5, k = 6 => fp = 2.6e-6
* n = 5, k = 8 => fp = 3.5e-7
* n = 10, k = 6 => fp = 1.6e-4
* n = 10, k = 8 => fp = 0.9e-4
* n = 15, k = 6 => fp = 0.18%
* n = 15, k = 8 => fp = 0.23%
* n = 20, k = 6 => fp = 1.1%
* n = 20, k = 8 => fp = 2.3%
*
* The above analysis assumes perfect hash functions,
* but those don't exist, so the real false positive
* rates may be worse.
*/
#define K 6
#define SEED 20221211
/* TO DO -- Use more modern hash functions with better distribution of bits */
static uint64_t
address_to_hash(void *ptr) {
assert(ptr != NULL);
uint64_t uhash = SEED;
uintptr_t addr = (uintptr_t)ptr;
for (int i = 0; i < SIZEOF_VOID_P; i++) {
uhash ^= addr & 255;
uhash *= (uint64_t)PyHASH_MULTIPLIER;
addr >>= 8;
}
return uhash;
}
void
_Py_BloomFilter_Init(_PyBloomFilter *bloom)
{
for (int i = 0; i < _Py_BLOOM_FILTER_WORDS; i++) {
bloom->bits[i] = 0;
}
}
/* We want K hash functions that each set 1 bit.
* A hash function that sets 1 bit in M bits can be trivially
* derived from a log2(M) bit hash function.
* So we extract 8 (log2(256)) bits at a time from
* the 64bit hash. */
void
_Py_BloomFilter_Add(_PyBloomFilter *bloom, void *ptr)
{
uint64_t hash = address_to_hash(ptr);
assert(K <= 8);
for (int i = 0; i < K; i++) {
uint8_t bits = hash & 255;
bloom->bits[bits >> 5] |= (1 << (bits&31));
hash >>= 8;
}
}
static bool
bloom_filter_may_contain(_PyBloomFilter *bloom, _PyBloomFilter *hashes)
{
for (int i = 0; i < _Py_BLOOM_FILTER_WORDS; i++) {
if ((bloom->bits[i] & hashes->bits[i]) != hashes->bits[i]) {
return false;
}
}
return true;
}
static int
link_executor(_PyExecutorObject *executor, const _PyBloomFilter *bloom)
{