Add miss stats for specialized instructions. (GH-31108)

This commit is contained in:
Mark Shannon 2022-02-04 09:56:46 +00:00 committed by GitHub
parent ba650af7d6
commit 832876b992
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 43 additions and 31 deletions

View file

@ -38,6 +38,33 @@
<instr N-1>
*/
/* Map from opcode to adaptive opcode.
Values of zero are ignored. */
static uint8_t adaptive_opcodes[256] = {
[LOAD_ATTR] = LOAD_ATTR_ADAPTIVE,
[LOAD_GLOBAL] = LOAD_GLOBAL_ADAPTIVE,
[LOAD_METHOD] = LOAD_METHOD_ADAPTIVE,
[BINARY_SUBSCR] = BINARY_SUBSCR_ADAPTIVE,
[STORE_SUBSCR] = STORE_SUBSCR_ADAPTIVE,
[CALL] = CALL_ADAPTIVE,
[STORE_ATTR] = STORE_ATTR_ADAPTIVE,
[BINARY_OP] = BINARY_OP_ADAPTIVE,
[COMPARE_OP] = COMPARE_OP_ADAPTIVE,
};
/* The number of cache entries required for a "family" of instructions. */
static uint8_t cache_requirements[256] = {
[LOAD_ATTR] = 2, /* _PyAdaptiveEntry and _PyAttrCache */
[LOAD_GLOBAL] = 2, /* _PyAdaptiveEntry and _PyLoadGlobalCache */
[LOAD_METHOD] = 3, /* _PyAdaptiveEntry, _PyAttrCache and _PyObjectCache */
[BINARY_SUBSCR] = 2, /* _PyAdaptiveEntry, _PyObjectCache */
[STORE_SUBSCR] = 0,
[CALL] = 2, /* _PyAdaptiveEntry and _PyObjectCache/_PyCallCache */
[STORE_ATTR] = 2, /* _PyAdaptiveEntry and _PyAttrCache */
[BINARY_OP] = 1, // _PyAdaptiveEntry
[COMPARE_OP] = 1, /* _PyAdaptiveEntry */
};
Py_ssize_t _Py_QuickenedCount = 0;
#ifdef Py_STATS
PyStats _py_stats = { 0 };
@ -144,7 +171,14 @@ _Py_GetSpecializationStats(void) {
static void
print_spec_stats(FILE *out, OpcodeStats *stats)
{
/* Mark some opcodes as specializable for stats,
* even though we don't specialize them yet. */
fprintf(out, " opcode[%d].specializable : 1\n", FOR_ITER);
fprintf(out, " opcode[%d].specializable : 1\n", UNPACK_SEQUENCE);
for (int i = 0; i < 256; i++) {
if (adaptive_opcodes[i]) {
fprintf(out, " opcode[%d].specializable : 1\n", i);
}
PRINT_STAT(i, specialization.success);
PRINT_STAT(i, specialization.failure);
PRINT_STAT(i, specialization.hit);
@ -266,33 +300,6 @@ get_cache_count(SpecializedCacheOrInstruction *quickened) {
return quickened[0].entry.zero.cache_count;
}
/* Map from opcode to adaptive opcode.
Values of zero are ignored. */
static uint8_t adaptive_opcodes[256] = {
[LOAD_ATTR] = LOAD_ATTR_ADAPTIVE,
[LOAD_GLOBAL] = LOAD_GLOBAL_ADAPTIVE,
[LOAD_METHOD] = LOAD_METHOD_ADAPTIVE,
[BINARY_SUBSCR] = BINARY_SUBSCR_ADAPTIVE,
[STORE_SUBSCR] = STORE_SUBSCR_ADAPTIVE,
[CALL] = CALL_ADAPTIVE,
[STORE_ATTR] = STORE_ATTR_ADAPTIVE,
[BINARY_OP] = BINARY_OP_ADAPTIVE,
[COMPARE_OP] = COMPARE_OP_ADAPTIVE,
};
/* The number of cache entries required for a "family" of instructions. */
static uint8_t cache_requirements[256] = {
[LOAD_ATTR] = 2, /* _PyAdaptiveEntry and _PyAttrCache */
[LOAD_GLOBAL] = 2, /* _PyAdaptiveEntry and _PyLoadGlobalCache */
[LOAD_METHOD] = 3, /* _PyAdaptiveEntry, _PyAttrCache and _PyObjectCache */
[BINARY_SUBSCR] = 2, /* _PyAdaptiveEntry, _PyObjectCache */
[STORE_SUBSCR] = 0,
[CALL] = 2, /* _PyAdaptiveEntry and _PyObjectCache/_PyCallCache */
[STORE_ATTR] = 2, /* _PyAdaptiveEntry and _PyAttrCache */
[BINARY_OP] = 1, // _PyAdaptiveEntry
[COMPARE_OP] = 1, /* _PyAdaptiveEntry */
};
/* Return the oparg for the cache_offset and instruction index.
*
* If no cache is needed then return the original oparg.