Hacky way to make FOR_ITER a viable uop

gvanrossum · Nov 15, 2023 · b28effa · b28effa
1 parent b838435
commit b28effa
Show file tree

Hide file tree

Showing 8 changed files with 137 additions and 44 deletions.
diff --git a/Include/internal/pycore_opcode_metadata.h b/Include/internal/pycore_opcode_metadata.h
diff --git a/Python/abstract_interp_cases.c.h b/Python/abstract_interp_cases.c.h
diff --git a/Python/bytecodes.c b/Python/bytecodes.c
@@ -2368,7 +2368,7 @@ dummy_func(
             goto enter_tier_one;
         }
 
-       replaced op(_POP_JUMP_IF_FALSE, (unused/1, cond -- )) {
+        replaced op(_POP_JUMP_IF_FALSE, (unused/1, cond -- )) {
             assert(PyBool_Check(cond));
             int flag = Py_IsFalse(cond);
             #if ENABLE_SPECIALIZATION
@@ -2512,7 +2512,7 @@ dummy_func(
             #endif  /* ENABLE_SPECIALIZATION */
         }
 
-        op(_FOR_ITER, (iter -- iter, next)) {
+        replaced op(_FOR_ITER, (iter -- iter, next)) {
             /* before: [iter]; after: [iter, iter()] *or* [] (and jump over END_FOR.) */
             next = (*Py_TYPE(iter)->tp_iternext)(iter);
             if (next == NULL) {
@@ -2535,6 +2535,31 @@ dummy_func(
             // Common case: no jump, leave it to the code generator
         }
 
+        op(_FOR_ITER_TIER_TWO, (iter -- iter, next)) {
+            /* before: [iter]; after: [iter, iter()] *or* [] (and jump over END_FOR.) */
+            next = (*Py_TYPE(iter)->tp_iternext)(iter);
+            if (next == NULL) {
+                if (_PyErr_Occurred(tstate)) {
+                    if (!_PyErr_ExceptionMatches(tstate, PyExc_StopIteration)) {
+                        GOTO_ERROR(error);
+                    }
+                    _PyErr_Clear(tstate);
+                }
+                /* iterator ended normally */
+                Py_DECREF(iter);
+                STACK_SHRINK(1);
+                /* HACK: Emulate DEOPT_IF to jump over END_FOR */
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                frame->instr_ptr += 1 + INLINE_CACHE_ENTRIES_FOR_ITER + oparg + 1;
+                assert(frame->instr_ptr[-1].op.code == END_FOR ||
+                       frame->instr_ptr[-1].op.code == INSTRUMENTED_END_FOR);
+                Py_DECREF(current_executor);
+                OPT_HIST(trace_uop_execution_counter, trace_run_length_hist);
+                goto enter_tier_one;
+            }
+            // Common case: no jump, leave it to the code generator
+        }
+
         macro(FOR_ITER) = _SPECIALIZE_FOR_ITER + _FOR_ITER;
 
         inst(INSTRUMENTED_FOR_ITER, (unused/1 -- )) {

diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
diff --git a/Python/optimizer.c b/Python/optimizer.c
@@ -392,6 +392,7 @@ _PyUop_Replacements[OPCODE_METADATA_SIZE] = {
     [_ITER_JUMP_RANGE] = _GUARD_NOT_EXHAUSTED_RANGE,
     [_ITER_JUMP_LIST] = _GUARD_NOT_EXHAUSTED_LIST,
     [_ITER_JUMP_TUPLE] = _GUARD_NOT_EXHAUSTED_TUPLE,
+    [_FOR_ITER] = _FOR_ITER_TIER_TWO,
 };
 
 static const uint16_t

diff --git a/Tools/cases_generator/flags.py b/Tools/cases_generator/flags.py
@@ -175,7 +175,7 @@ def variable_used_unspecialized(node: parsing.Node, name: str) -> bool:
     tokens: list[lx.Token] = []
     skipping = False
     for i, token in enumerate(node.tokens):
-        if token.kind == "MACRO":
+        if token.kind == "CMACRO":
             text = "".join(token.text.split())
             # TODO: Handle nested #if
             if text == "#if":

diff --git a/Tools/cases_generator/generate_cases.py b/Tools/cases_generator/generate_cases.py
@@ -658,7 +658,7 @@ def write_macro_expansions(
                 if not part.instr.is_viable_uop() and "replaced" not in part.instr.annotations:
                     # This note just reminds us about macros that cannot
                     # be expanded to Tier 2 uops. It is not an error.
-                    # It is sometimes emitted for macros that have a
+                    # Suppress it using 'replaced op(...)' for macros having
                     # manual translation in translate_bytecode_to_trace()
                     # in Python/optimizer.c.
                     if len(parts) > 1 or part.instr.name != name:

diff --git a/Tools/cases_generator/instructions.py b/Tools/cases_generator/instructions.py
@@ -115,7 +115,7 @@ def __init__(self, inst: parsing.InstDef):
     def is_viable_uop(self) -> bool:
         """Whether this instruction is viable as a uop."""
         dprint: typing.Callable[..., None] = lambda *args, **kwargs: None
-        if "FRAME" in self.name:
+        if self.name == "_FOR_ITER_TIER_TWO":
             dprint = print
 
         if self.name == "_EXIT_TRACE":