[3.13] gh-119258: Backport optimizer frame fixes in GH-119365 (GH-120699)

(cherry picked from commit 55402d3)
This commit is contained in:
Ken Jin 2024-06-20 23:55:20 +08:00 committed by GitHub
parent b8fd80f91b
commit 7c7aa5a99c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 21 additions and 35 deletions

View file

@ -107,9 +107,9 @@ extern void _Py_uop_abstractcontext_fini(_Py_UOpsContext *ctx);
extern _Py_UOpsAbstractFrame *_Py_uop_frame_new( extern _Py_UOpsAbstractFrame *_Py_uop_frame_new(
_Py_UOpsContext *ctx, _Py_UOpsContext *ctx,
PyCodeObject *co, PyCodeObject *co,
_Py_UopsSymbol **localsplus_start, int curr_stackentries,
int n_locals_already_filled, _Py_UopsSymbol **args,
int curr_stackentries); int arg_len);
extern int _Py_uop_frame_pop(_Py_UOpsContext *ctx); extern int _Py_uop_frame_pop(_Py_UOpsContext *ctx);
PyAPI_FUNC(PyObject *) _Py_uop_symbols_test(PyObject *self, PyObject *ignored); PyAPI_FUNC(PyObject *) _Py_uop_symbols_test(PyObject *self, PyObject *ignored);

View file

@ -411,7 +411,7 @@ optimize_uops(
if (_Py_uop_abstractcontext_init(ctx) < 0) { if (_Py_uop_abstractcontext_init(ctx) < 0) {
goto out_of_space; goto out_of_space;
} }
_Py_UOpsAbstractFrame *frame = _Py_uop_frame_new(ctx, co, ctx->n_consumed, 0, curr_stacklen); _Py_UOpsAbstractFrame *frame = _Py_uop_frame_new(ctx, co, curr_stacklen, NULL, 0);
if (frame == NULL) { if (frame == NULL) {
return -1; return -1;
} }

View file

@ -616,17 +616,12 @@ dummy_func(void) {
argcount++; argcount++;
} }
_Py_UopsSymbol **localsplus_start = ctx->n_consumed;
int n_locals_already_filled = 0;
// Can determine statically, so we interleave the new locals
// and make the current stack the new locals.
// This also sets up for true call inlining.
if (sym_is_null(self_or_null) || sym_is_not_null(self_or_null)) { if (sym_is_null(self_or_null) || sym_is_not_null(self_or_null)) {
localsplus_start = args; OUT_OF_SPACE_IF_NULL(new_frame = frame_new(ctx, co, 0, args, argcount));
n_locals_already_filled = argcount; } else {
OUT_OF_SPACE_IF_NULL(new_frame = frame_new(ctx, co, 0, NULL, 0));
} }
OUT_OF_SPACE_IF_NULL(new_frame =
frame_new(ctx, co, localsplus_start, n_locals_already_filled, 0));
} }
op(_PY_FRAME_GENERAL, (callable, self_or_null, args[oparg] -- new_frame: _Py_UOpsAbstractFrame *)) { op(_PY_FRAME_GENERAL, (callable, self_or_null, args[oparg] -- new_frame: _Py_UOpsAbstractFrame *)) {

View file

@ -1696,17 +1696,11 @@
args--; args--;
argcount++; argcount++;
} }
_Py_UopsSymbol **localsplus_start = ctx->n_consumed;
int n_locals_already_filled = 0;
// Can determine statically, so we interleave the new locals
// and make the current stack the new locals.
// This also sets up for true call inlining.
if (sym_is_null(self_or_null) || sym_is_not_null(self_or_null)) { if (sym_is_null(self_or_null) || sym_is_not_null(self_or_null)) {
localsplus_start = args; OUT_OF_SPACE_IF_NULL(new_frame = frame_new(ctx, co, 0, args, argcount));
n_locals_already_filled = argcount; } else {
OUT_OF_SPACE_IF_NULL(new_frame = frame_new(ctx, co, 0, NULL, 0));
} }
OUT_OF_SPACE_IF_NULL(new_frame =
frame_new(ctx, co, localsplus_start, n_locals_already_filled, 0));
stack_pointer[-2 - oparg] = (_Py_UopsSymbol *)new_frame; stack_pointer[-2 - oparg] = (_Py_UopsSymbol *)new_frame;
stack_pointer += -1 - oparg; stack_pointer += -1 - oparg;
break; break;

View file

@ -303,9 +303,9 @@ _Py_UOpsAbstractFrame *
_Py_uop_frame_new( _Py_uop_frame_new(
_Py_UOpsContext *ctx, _Py_UOpsContext *ctx,
PyCodeObject *co, PyCodeObject *co,
_Py_UopsSymbol **localsplus_start, int curr_stackentries,
int n_locals_already_filled, _Py_UopsSymbol **args,
int curr_stackentries) int arg_len)
{ {
assert(ctx->curr_frame_depth < MAX_ABSTRACT_FRAME_DEPTH); assert(ctx->curr_frame_depth < MAX_ABSTRACT_FRAME_DEPTH);
_Py_UOpsAbstractFrame *frame = &ctx->frames[ctx->curr_frame_depth]; _Py_UOpsAbstractFrame *frame = &ctx->frames[ctx->curr_frame_depth];
@ -313,21 +313,21 @@ _Py_uop_frame_new(
frame->stack_len = co->co_stacksize; frame->stack_len = co->co_stacksize;
frame->locals_len = co->co_nlocalsplus; frame->locals_len = co->co_nlocalsplus;
frame->locals = localsplus_start; frame->locals = ctx->n_consumed;
frame->stack = frame->locals + co->co_nlocalsplus; frame->stack = frame->locals + co->co_nlocalsplus;
frame->stack_pointer = frame->stack + curr_stackentries; frame->stack_pointer = frame->stack + curr_stackentries;
ctx->n_consumed = localsplus_start + (co->co_nlocalsplus + co->co_stacksize); ctx->n_consumed = ctx->n_consumed + (co->co_nlocalsplus + co->co_stacksize);
if (ctx->n_consumed >= ctx->limit) { if (ctx->n_consumed >= ctx->limit) {
return NULL; return NULL;
} }
// Initialize with the initial state of all local variables // Initialize with the initial state of all local variables
for (int i = n_locals_already_filled; i < co->co_nlocalsplus; i++) { for (int i = 0; i < arg_len; i++) {
frame->locals[i] = args[i];
}
for (int i = arg_len; i < co->co_nlocalsplus; i++) {
_Py_UopsSymbol *local = _Py_uop_sym_new_unknown(ctx); _Py_UopsSymbol *local = _Py_uop_sym_new_unknown(ctx);
if (local == NULL) {
return NULL;
}
frame->locals[i] = local; frame->locals[i] = local;
} }
@ -335,9 +335,6 @@ _Py_uop_frame_new(
// Initialize the stack as well // Initialize the stack as well
for (int i = 0; i < curr_stackentries; i++) { for (int i = 0; i < curr_stackentries; i++) {
_Py_UopsSymbol *stackvar = _Py_uop_sym_new_unknown(ctx); _Py_UopsSymbol *stackvar = _Py_uop_sym_new_unknown(ctx);
if (stackvar == NULL) {
return NULL;
}
frame->stack[i] = stackvar; frame->stack[i] = stackvar;
} }