mirrors/cpython
1
0
Fork 0
mirror of https://github.com/python/cpython.git synced 2025-09-10 18:58:35 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 13

[3.13] gh-136541: Fix several problems of perf trampolines in x86_64 and aarch64 (GH-136500) (#136545)

Browse source 
This commit fixes the following problems:

* The x86_64 trampolines are not preserving frame pointers
* The hardcoded offsets to the code segment from the FDE only worked properly for x64_64
* The CIE data was not following conventions of aarch64
* The eh_frame for aarch64 was not fully correct
(cherry picked from commit 236f733d8f)
This commit is contained in:
Pablo Galindo Salgado 2025-07-11 15:02:19 +01:00 • committed by GitHub
parent 175ada2806
commit cba2974a54
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 147 additions and 41 deletions
Download patch file Download diff file Expand all files Collapse all files

16
Python/perf_trampoline.c
View file

@ -186,6 +186,8 @@ struct code_arena_st {
*prev; // Pointer to the arena or NULL if this is the first arena.
};
#define CODE_ALIGNMENT 32
typedef struct code_arena_st code_arena_t;
typedef struct trampoline_api_st trampoline_api_t;
@ -291,7 +293,9 @@ new_code_arena(void)
void *start = &_Py_trampoline_func_start;
void *end = &_Py_trampoline_func_end;
size_t code_size = end - start;
size_t chunk_size = round_up(code_size + trampoline_api.code_padding, 16);
size_t unaligned_size = code_size + trampoline_api.code_padding;
size_t chunk_size = round_up(unaligned_size, CODE_ALIGNMENT);
assert(chunk_size % CODE_ALIGNMENT == 0);
// TODO: Check the effect of alignment of the code chunks. Initial investigation
// showed that this has no effect on performance in x86-64 or aarch64 and the current
// version has the advantage that the unwinder in GDB can unwind across JIT-ed code.
@ -356,7 +360,9 @@ static inline py_trampoline
code_arena_new_code(code_arena_t *code_arena)
{
py_trampoline trampoline = (py_trampoline)code_arena->current_addr;
size_t total_code_size = round_up(code_arena->code_size + trampoline_api.code_padding, 16);
size_t total_code_size = round_up(code_arena->code_size + trampoline_api.code_padding,
CODE_ALIGNMENT);
assert(total_code_size % CODE_ALIGNMENT == 0);
code_arena->size_left -= total_code_size;
code_arena->current_addr += total_code_size;
return trampoline;
@ -489,9 +495,6 @@ _PyPerfTrampoline_Init(int activate)
}
else {
tstate->interp->eval_frame = py_trampoline_evaluator;
if (new_code_arena() < 0) {
return -1;
}
extra_code_index = _PyEval_RequestCodeExtraIndex(NULL);
if (extra_code_index == -1) {
return -1;
@ -499,6 +502,9 @@ _PyPerfTrampoline_Init(int activate)
if (trampoline_api.state == NULL && trampoline_api.init_state != NULL) {
trampoline_api.state = trampoline_api.init_state();
}
if (new_code_arena() < 0) {
return -1;
}
perf_status = PERF_STATUS_OK;
}
#endif