gh-105481: generate op IDs from bytecode.c instead of hard coding them in opcode.py (#107971)

1 files changed, 200 insertions, 6 deletions

diff --git a/Tools/cases_generator/generate_cases.py b/Tools/cases_generator/generate_cases.py
index e170e11..d991cb4 100644
--- a/Tools/cases_generator/generate_cases.py
+++ b/Tools/cases_generator/generate_cases.py
@@ -5,6 +5,7 @@ Writes the cases to generated_cases.c.h, which is #included in ceval.c.
 
 import argparse
 import contextlib
+import itertools
 import os
 import posixpath
 import sys
@@ -36,6 +37,12 @@ THIS = os.path.relpath(__file__, ROOT).replace(os.path.sep, posixpath.sep)
 
 DEFAULT_INPUT = os.path.relpath(os.path.join(ROOT, "Python/bytecodes.c"))
 DEFAULT_OUTPUT = os.path.relpath(os.path.join(ROOT, "Python/generated_cases.c.h"))
+DEFAULT_OPCODE_IDS_H_OUTPUT = os.path.relpath(
+    os.path.join(ROOT, "Include/opcode_ids.h")
+)
+DEFAULT_OPCODE_TARGETS_H_OUTPUT = os.path.relpath(
+    os.path.join(ROOT, "Python/opcode_targets.h")
+)
 DEFAULT_METADATA_OUTPUT = os.path.relpath(
     os.path.join(ROOT, "Include/internal/pycore_opcode_metadata.h")
 )
@@ -87,6 +94,20 @@ arg_parser.add_argument(
     "-o", "--output", type=str, help="Generated code", default=DEFAULT_OUTPUT
 )
 arg_parser.add_argument(
+    "-n",
+    "--opcode_ids_h",
+    type=str,
+    help="Header file with opcode number definitions",
+    default=DEFAULT_OPCODE_IDS_H_OUTPUT,
+)
+arg_parser.add_argument(
+    "-t",
+    "--opcode_targets_h",
+    type=str,
+    help="File with opcode targets for computed gotos",
+    default=DEFAULT_OPCODE_TARGETS_H_OUTPUT,
+)
+arg_parser.add_argument(
     "-m",
     "--metadata",
     type=str,
@@ -225,6 +246,129 @@ class Generator(Analyzer):
         self.out.write_raw(self.from_source_files())
         self.out.write_raw(f"{self.out.comment} Do not edit!\n")
 
+    def assign_opcode_ids(self):
+        """Assign IDs to opcodes"""
+
+        ops: list[(bool, str)] = []  # (has_arg, name) for each opcode
+        instrumented_ops: list[str] = []
+
+        for instr in itertools.chain(
+            [instr for instr in self.instrs.values() if instr.kind != "op"],
+            self.macro_instrs.values()):
+
+            name = instr.name
+            if name.startswith('INSTRUMENTED_'):
+                instrumented_ops.append(name)
+            else:
+                ops.append((instr.instr_flags.HAS_ARG_FLAG, name))
+
+        # Special case: this instruction is implemented in ceval.c
+        # rather than bytecodes.c, so we need to add it explicitly
+        # here (at least until we add something to bytecodes.c to
+        # declare external instructions).
+        instrumented_ops.append('INSTRUMENTED_LINE')
+
+        # assert lists are unique
+        assert len(set(ops)) == len(ops)
+        assert len(set(instrumented_ops)) == len(instrumented_ops)
+
+        opname: list[str or None] = [None] * 512
+        opmap: dict = {}
+        markers: dict = {}
+
+        def map_op(op, name):
+            assert op < len(opname)
+            assert opname[op] is None
+            assert name not in opmap
+            opname[op] = name
+            opmap[name] = op
+
+
+        # 0 is reserved for cache entries. This helps debugging.
+        map_op(0, 'CACHE')
+
+        # 17 is reserved as it is the initial value for the specializing counter.
+        # This helps catch cases where we attempt to execute a cache.
+        map_op(17, 'RESERVED')
+
+        # 166 is RESUME - it is hard coded as such in Tools/build/deepfreeze.py
+        map_op(166, 'RESUME')
+
+        next_opcode = 1
+
+        for has_arg, name in sorted(ops):
+            if name in opmap:
+                continue  # an anchored name, like CACHE
+            while opname[next_opcode] is not None:
+                next_opcode += 1
+            assert next_opcode < 255
+            map_op(next_opcode, name)
+
+            if has_arg and 'HAVE_ARGUMENT' not in markers:
+                markers['HAVE_ARGUMENT'] = next_opcode
+
+        # Instrumented opcodes are at the end of the valid range
+        min_instrumented = 254 - (len(instrumented_ops) - 1)
+        assert next_opcode <= min_instrumented
+        markers['MIN_INSTRUMENTED_OPCODE'] = min_instrumented
+        for i, op in enumerate(instrumented_ops):
+            map_op(min_instrumented + i, op)
+
+        # Pseudo opcodes are after the valid range
+        for i, op in enumerate(sorted(self.pseudos)):
+            map_op(256 + i, op)
+
+        assert 255 not in opmap  # 255 is reserved
+        self.opmap = opmap
+        self.markers = markers
+
+    def write_opcode_ids(self, opcode_ids_h_filename, opcode_targets_filename):
+        """Write header file that defined the opcode IDs"""
+
+        with open(opcode_ids_h_filename, "w") as f:
+            # Create formatter
+            self.out = Formatter(f, 0)
+
+            self.write_provenance_header()
+
+            self.out.emit("")
+            self.out.emit("#ifndef Py_OPCODE_IDS_H")
+            self.out.emit("#define Py_OPCODE_IDS_H")
+            self.out.emit("#ifdef __cplusplus")
+            self.out.emit("extern \"C\" {")
+            self.out.emit("#endif")
+            self.out.emit("")
+            self.out.emit("/* Instruction opcodes for compiled code */")
+
+            def define(name, opcode):
+                self.out.emit(f"#define {name:<38} {opcode:>3}")
+
+            all_pairs = []
+            # the second item in the tuple sorts the markers before the ops
+            all_pairs.extend((i, 1, name) for (name, i) in self.markers.items())
+            all_pairs.extend((i, 2, name) for (name, i) in self.opmap.items())
+            for i, _, name in sorted(all_pairs):
+                assert name is not None
+                define(name, i)
+
+            self.out.emit("")
+            self.out.emit("#ifdef __cplusplus")
+            self.out.emit("}")
+            self.out.emit("#endif")
+            self.out.emit("#endif /* !Py_OPCODE_IDS_H */")
+
+        with open(opcode_targets_filename, "w") as f:
+            # Create formatter
+            self.out = Formatter(f, 0)
+
+            with self.out.block("static void *opcode_targets[256] =", ";"):
+                targets = ["_unknown_opcode"] * 256
+                for name, op in self.opmap.items():
+                    if op < 256:
+                        targets[op] = f"TARGET_{name}"
+                f.write(",\n".join([f"    &&{s}" for s in targets]))
+
+
     def write_metadata(self, metadata_filename: str, pymetadata_filename: str) -> None:
         """Write instruction metadata to output file."""
 
@@ -378,12 +522,46 @@ class Generator(Analyzer):
             ):
                 self.write_uop_items(lambda name, counter: f'[{name}] = "{name}",')
 
+            with self.metadata_item(
+                f"const char *const _PyOpcode_OpName[{1 + max(self.opmap.values())}]", "=", ";"
+            ):
+                for name in self.opmap:
+                    self.out.emit(f'[{name}] = "{name}",')
+
+            deoptcodes = {}
+            for name, op in self.opmap.items():
+                if op < 256:
+                    deoptcodes[name] = name
+            for name, family in self.families.items():
+                for m in family.members:
+                    deoptcodes[m] = name
+            # special case:
+            deoptcodes['BINARY_OP_INPLACE_ADD_UNICODE'] = 'BINARY_OP'
+
+            with self.metadata_item(
+                f"const uint8_t _PyOpcode_Deopt[256]", "=", ";"
+            ):
+                for opt, deopt in sorted(deoptcodes.items()):
+                    self.out.emit(f"[{opt}] = {deopt},")
+
+            self.out.emit("")
+            self.out.emit("#define EXTRA_CASES \\")
+            valid_opcodes = set(self.opmap.values())
+            with self.out.indent():
+                for op in range(256):
+                    if op not in valid_opcodes:
+                        self.out.emit(f"case {op}: \\")
+                self.out.emit("    ;\n")
+
         with open(pymetadata_filename, "w") as f:
             # Create formatter
             self.out = Formatter(f, 0, comment="#")
 
             self.write_provenance_header()
 
+            # emit specializations
+            specialized_ops = set()
+
             self.out.emit("")
             self.out.emit("_specializations = {")
             for name, family in self.families.items():
@@ -392,6 +570,7 @@ class Generator(Analyzer):
                     with self.out.indent():
                         for m in family.members:
                             self.out.emit(f'"{m}",')
+                        specialized_ops.update(family.members)
                     self.out.emit(f"],")
             self.out.emit("}")
 
@@ -402,14 +581,26 @@ class Generator(Analyzer):
                 '_specializations["BINARY_OP"].append('
                 '"BINARY_OP_INPLACE_ADD_UNICODE")'
             )
+            specialized_ops.add("BINARY_OP_INPLACE_ADD_UNICODE")
 
-            # Make list of specialized instructions
+            ops = sorted((id, name) for (name, id) in self.opmap.items())
+            # emit specialized opmap
             self.out.emit("")
-            self.out.emit(
-                "_specialized_instructions = ["
-                "opcode for family in _specializations.values() for opcode in family"
-                "]"
-            )
+            with self.out.block("_specialized_opmap ="):
+                for op, name in ops:
+                    if name in specialized_ops:
+                        self.out.emit(f"'{name}': {op},")
+
+            # emit opmap
+            self.out.emit("")
+            with self.out.block("opmap ="):
+                for op, name in ops:
+                    if name not in specialized_ops:
+                        self.out.emit(f"'{name}': {op},")
+
+            for name in ['MIN_INSTRUMENTED_OPCODE', 'HAVE_ARGUMENT']:
+                self.out.emit(f"{name} = {self.markers[name]}")
+
 
     def write_pseudo_instrs(self) -> None:
         """Write the IS_PSEUDO_INSTR macro"""
@@ -683,6 +874,9 @@ def main():
 
     # These raise OSError if output can't be written
     a.write_instructions(args.output, args.emit_line_directives)
+
+    a.assign_opcode_ids()
+    a.write_opcode_ids(args.opcode_ids_h, args.opcode_targets_h)
     a.write_metadata(args.metadata, args.pymetadata)
     a.write_executor_instructions(args.executor_cases, args.emit_line_directives)
     a.write_abstract_interpreter_instructions(args.abstract_interpreter_cases,