GH-111485: Separate out parsing, analysis and code-gen phases of tier 1 code generator (GH-112299)

Include/internal/pycore_opcode_metadata.h

@@ -1285,11 +1285,11 @@ int _PyOpcode_num_pushed(int opcode, int oparg, bool jump)  {
         case _INIT_CALL_PY_EXACT_ARGS:
             return 1;
         case _PUSH_FRAME:
-            return 1;
+            return 0;
         case CALL_BOUND_METHOD_EXACT_ARGS:
-            return 1;
+            return 0;
         case CALL_PY_EXACT_ARGS:
-            return 1;
+            return 0;
         case CALL_PY_WITH_DEFAULTS:
             return 1;
         case CALL_TYPE_1:

Makefile.pre.in

@@ -1587,7 +1587,6 @@ regen-cases:
 	$(PYTHON_FOR_REGEN) \
 	    $(srcdir)/Tools/cases_generator/generate_cases.py \
 		$(CASESFLAG) \
-		-o $(srcdir)/Python/generated_cases.c.h.new \
 		-n $(srcdir)/Include/opcode_ids.h.new \
 		-t $(srcdir)/Python/opcode_targets.h.new \
 		-m $(srcdir)/Include/internal/pycore_opcode_metadata.h.new \
@@ -1595,6 +1594,8 @@ regen-cases:
 		-p $(srcdir)/Lib/_opcode_metadata.py.new \
 		-a $(srcdir)/Python/abstract_interp_cases.c.h.new \
 		$(srcdir)/Python/bytecodes.c
+	$(PYTHON_FOR_REGEN) \
+	    $(srcdir)/Tools/cases_generator/tier1_generator.py -o $(srcdir)/Python/generated_cases.c.h.new $(srcdir)/Python/bytecodes.c
 	$(UPDATE_FILE) $(srcdir)/Python/generated_cases.c.h $(srcdir)/Python/generated_cases.c.h.new
 	$(UPDATE_FILE) $(srcdir)/Include/opcode_ids.h $(srcdir)/Include/opcode_ids.h.new
 	$(UPDATE_FILE) $(srcdir)/Python/opcode_targets.h $(srcdir)/Python/opcode_targets.h.new

Python/abstract_interp_cases.c.h

@@ -774,7 +774,8 @@
         }
 
         case _PUSH_FRAME: {
-            PARTITIONNODE_OVERWRITE((_Py_PARTITIONNODE_t *)PARTITIONNODE_NULLROOT, PEEK(-(-1)), true);
+            STACK_SHRINK(1);
+            PARTITIONNODE_OVERWRITE((_Py_PARTITIONNODE_t *)PARTITIONNODE_NULLROOT, PEEK(-(0)), true);
             break;
         }
 

Python/bytecodes.c

@@ -800,11 +800,11 @@ dummy_func(
         // We also push it onto the stack on exit, but that's a
         // different frame, and it's accounted for by _PUSH_FRAME.
         op(_POP_FRAME, (retval --)) {
-            assert(EMPTY());
             #if TIER_ONE
             assert(frame != &entry_frame);
             #endif
             STORE_SP();
+            assert(EMPTY());
             _Py_LeaveRecursiveCallPy(tstate);
             // GH-99729: We need to unlink the frame *before* clearing it:
             _PyInterpreterFrame *dying = frame;
@@ -1165,7 +1165,6 @@ dummy_func(
             }
         }
 
-
         inst(STORE_NAME, (v -- )) {
             PyObject *name = GETITEM(FRAME_CO_NAMES, oparg);
             PyObject *ns = LOCALS();
@@ -3130,7 +3129,7 @@ dummy_func(
         // The 'unused' output effect represents the return value
         // (which will be pushed when the frame returns).
         // It is needed so CALL_PY_EXACT_ARGS matches its family.
-        op(_PUSH_FRAME, (new_frame: _PyInterpreterFrame* -- unused)) {
+        op(_PUSH_FRAME, (new_frame: _PyInterpreterFrame* -- unused if (0))) {
             // Write it out explicitly because it's subtly different.
             // Eventually this should be the only occurrence of this code.
             assert(tstate->interp->eval_frame == NULL);

Python/executor_cases.c.h

@@ -661,11 +661,11 @@
             PyObject *retval;
             retval = stack_pointer[-1];
             STACK_SHRINK(1);
-            assert(EMPTY());
             #if TIER_ONE
             assert(frame != &entry_frame);
             #endif
             STORE_SP();
+            assert(EMPTY());
             _Py_LeaveRecursiveCallPy(tstate);
             // GH-99729: We need to unlink the frame *before* clearing it:
             _PyInterpreterFrame *dying = frame;

Tools/cases_generator/analyzer.py

@@ -0,0 +1,456 @@
+from dataclasses import dataclass
+import lexer
+import parser
+from typing import Optional
+
+
+@dataclass
+class Properties:
+    escapes: bool
+    infallible: bool
+    deopts: bool
+    oparg: bool
+    jumps: bool
+    ends_with_eval_breaker: bool
+    needs_this: bool
+    always_exits: bool
+    stores_sp: bool
+
+    def dump(self, indent: str) -> None:
+        print(indent, end="")
+        text = ", ".join([f"{key}: {value}" for (key, value) in self.__dict__.items()])
+        print(indent, text, sep="")
+
+    @staticmethod
+    def from_list(properties: list["Properties"]) -> "Properties":
+        return Properties(
+            escapes=any(p.escapes for p in properties),
+            infallible=all(p.infallible for p in properties),
+            deopts=any(p.deopts for p in properties),
+            oparg=any(p.oparg for p in properties),
+            jumps=any(p.jumps for p in properties),
+            ends_with_eval_breaker=any(p.ends_with_eval_breaker for p in properties),
+            needs_this=any(p.needs_this for p in properties),
+            always_exits=any(p.always_exits for p in properties),
+            stores_sp=any(p.stores_sp for p in properties),
+        )
+
+
+SKIP_PROPERTIES = Properties(
+    escapes=False,
+    infallible=True,
+    deopts=False,
+    oparg=False,
+    jumps=False,
+    ends_with_eval_breaker=False,
+    needs_this=False,
+    always_exits=False,
+    stores_sp=False,
+)
+
+
+@dataclass
+class Skip:
+    "Unused cache entry"
+    size: int
+
+    @property
+    def name(self) -> str:
+        return f"unused/{self.size}"
+
+    @property
+    def properties(self) -> Properties:
+        return SKIP_PROPERTIES
+
+
+@dataclass
+class StackItem:
+    name: str
+    type: str | None
+    condition: str | None
+    size: str
+    peek: bool = False
+
+    def __str__(self) -> str:
+        cond = f" if ({self.condition})" if self.condition else ""
+        size = f"[{self.size}]" if self.size != "1" else ""
+        type = "" if self.type is None else f"{self.type} "
+        return f"{type}{self.name}{size}{cond} {self.peek}"
+
+    def is_array(self) -> bool:
+        return self.type == "PyObject **"
+
+
+@dataclass
+class StackEffect:
+    inputs: list[StackItem]
+    outputs: list[StackItem]
+
+    def __str__(self) -> str:
+        return f"({', '.join([str(i) for i in self.inputs])} -- {', '.join([str(i) for i in self.outputs])})"
+
+
+@dataclass
+class CacheEntry:
+    name: str
+    size: int
+
+    def __str__(self) -> str:
+        return f"{self.name}/{self.size}"
+
+
+@dataclass
+class Uop:
+    name: str
+    context: parser.Context | None
+    annotations: list[str]
+    stack: StackEffect
+    caches: list[CacheEntry]
+    body: list[lexer.Token]
+    properties: Properties
+    _size: int = -1
+
+    def dump(self, indent: str) -> None:
+        print(
+            indent, self.name, ", ".join(self.annotations) if self.annotations else ""
+        )
+        print(indent, self.stack, ", ".join([str(c) for c in self.caches]))
+        self.properties.dump("    " + indent)
+
+    @property
+    def size(self) -> int:
+        if self._size < 0:
+            self._size = sum(c.size for c in self.caches)
+        return self._size
+
+
+Part = Uop | Skip
+
+
+@dataclass
+class Instruction:
+    name: str
+    parts: list[Part]
+    _properties: Properties | None
+    is_target: bool = False
+    family: Optional["Family"] = None
+
+    @property
+    def properties(self) -> Properties:
+        if self._properties is None:
+            self._properties = self._compute_properties()
+        return self._properties
+
+    def _compute_properties(self) -> Properties:
+        return Properties.from_list([part.properties for part in self.parts])
+
+    def dump(self, indent: str) -> None:
+        print(indent, self.name, "=", ", ".join([part.name for part in self.parts]))
+        self.properties.dump("    " + indent)
+
+    @property
+    def size(self) -> int:
+        return 1 + sum(part.size for part in self.parts)
+
+
+@dataclass
+class PseudoInstruction:
+    name: str
+    targets: list[Instruction]
+    flags: list[str]
+
+    def dump(self, indent: str) -> None:
+        print(indent, self.name, "->", " or ".join([t.name for t in self.targets]))
+
+
+@dataclass
+class Family:
+    name: str
+    size: str
+    members: list[Instruction]
+
+    def dump(self, indent: str) -> None:
+        print(indent, self.name, "= ", ", ".join([m.name for m in self.members]))
+
+
+@dataclass
+class Analysis:
+    instructions: dict[str, Instruction]
+    uops: dict[str, Uop]
+    families: dict[str, Family]
+    pseudos: dict[str, PseudoInstruction]
+
+
+def analysis_error(message: str, tkn: lexer.Token) -> SyntaxError:
+    # To do -- support file and line output
+    # Construct a SyntaxError instance from message and token
+    return lexer.make_syntax_error(message, "", tkn.line, tkn.column, "")
+
+
+def override_error(
+    name: str,
+    context: parser.Context | None,
+    prev_context: parser.Context | None,
+    token: lexer.Token,
+) -> SyntaxError:
+    return analysis_error(
+        f"Duplicate definition of '{name}' @ {context} "
+        f"previous definition @ {prev_context}",
+        token,
+    )
+
+
+def convert_stack_item(item: parser.StackEffect) -> StackItem:
+    return StackItem(item.name, item.type, item.cond, (item.size or "1"))
+
+
+def analyze_stack(op: parser.InstDef) -> StackEffect:
+    inputs: list[StackItem] = [
+        convert_stack_item(i) for i in op.inputs if isinstance(i, parser.StackEffect)
+    ]
+    outputs: list[StackItem] = [convert_stack_item(i) for i in op.outputs]
+    for input, output in zip(inputs, outputs):
+        if input.name == output.name:
+            input.peek = output.peek = True
+    return StackEffect(inputs, outputs)
+
+
+def analyze_caches(op: parser.InstDef) -> list[CacheEntry]:
+    caches: list[parser.CacheEffect] = [
+        i for i in op.inputs if isinstance(i, parser.CacheEffect)
+    ]
+    return [CacheEntry(i.name, int(i.size)) for i in caches]
+
+
+def variable_used(node: parser.InstDef, name: str) -> bool:
+    """Determine whether a variable with a given name is used in a node."""
+    return any(
+        token.kind == "IDENTIFIER" and token.text == name for token in node.tokens
+    )
+
+
+def is_infallible(op: parser.InstDef) -> bool:
+    return not (
+        variable_used(op, "ERROR_IF")
+        or variable_used(op, "error")
+        or variable_used(op, "pop_1_error")
+        or variable_used(op, "exception_unwind")
+        or variable_used(op, "resume_with_error")
+    )
+
+
+from flags import makes_escaping_api_call
+
+EXITS = {
+    "DISPATCH",
+    "GO_TO_INSTRUCTION",
+    "Py_UNREACHABLE",
+    "DISPATCH_INLINED",
+    "DISPATCH_GOTO",
+}
+
+
+def eval_breaker_at_end(op: parser.InstDef) -> bool:
+    return op.tokens[-5].text == "CHECK_EVAL_BREAKER"
+
+
+def always_exits(op: parser.InstDef) -> bool:
+    depth = 0
+    tkn_iter = iter(op.tokens)
+    for tkn in tkn_iter:
+        if tkn.kind == "LBRACE":
+            depth += 1
+        elif tkn.kind == "RBRACE":
+            depth -= 1
+        elif depth > 1:
+            continue
+        elif tkn.kind == "GOTO" or tkn.kind == "RETURN":
+            return True
+        elif tkn.kind == "KEYWORD":
+            if tkn.text in EXITS:
+                return True
+        elif tkn.kind == "IDENTIFIER":
+            if tkn.text in EXITS:
+                return True
+            if tkn.text == "DEOPT_IF" or tkn.text == "ERROR_IF":
+                next(tkn_iter)  # '('
+                t = next(tkn_iter)
+                if t.text == "true":
+                    return True
+    return False
+
+
+def compute_properties(op: parser.InstDef) -> Properties:
+    return Properties(
+        escapes=makes_escaping_api_call(op),
+        infallible=is_infallible(op),
+        deopts=variable_used(op, "DEOPT_IF"),
+        oparg=variable_used(op, "oparg"),
+        jumps=variable_used(op, "JUMPBY"),
+        ends_with_eval_breaker=eval_breaker_at_end(op),
+        needs_this=variable_used(op, "this_instr"),
+        always_exits=always_exits(op),
+        stores_sp=variable_used(op, "STORE_SP"),
+    )
+
+
+def make_uop(name: str, op: parser.InstDef) -> Uop:
+    return Uop(
+        name=name,
+        context=op.context,
+        annotations=op.annotations,
+        stack=analyze_stack(op),
+        caches=analyze_caches(op),
+        body=op.block.tokens,
+        properties=compute_properties(op),
+    )
+
+
+def add_op(op: parser.InstDef, uops: dict[str, Uop]) -> None:
+    assert op.kind == "op"
+    if op.name in uops:
+        if "override" not in op.annotations:
+            raise override_error(
+                op.name, op.context, uops[op.name].context, op.tokens[0]
+            )
+    uops[op.name] = make_uop(op.name, op)
+
+
+def add_instruction(
+    name: str, parts: list[Part], instructions: dict[str, Instruction]
+) -> None:
+    instructions[name] = Instruction(name, parts, None)
+
+
+def desugar_inst(
+    inst: parser.InstDef, instructions: dict[str, Instruction], uops: dict[str, Uop]
+) -> None:
+    assert inst.kind == "inst"
+    name = inst.name
+    uop = make_uop("_" + inst.name, inst)
+    uops[inst.name] = uop
+    add_instruction(name, [uop], instructions)
+
+
+def add_macro(
+    macro: parser.Macro, instructions: dict[str, Instruction], uops: dict[str, Uop]
+) -> None:
+    parts: list[Uop | Skip] = []
+    for part in macro.uops:
+        match part:
+            case parser.OpName():
+                if part.name not in uops:
+                    analysis_error(f"No Uop named {part.name}", macro.tokens[0])
+                parts.append(uops[part.name])
+            case parser.CacheEffect():
+                parts.append(Skip(part.size))
+            case _:
+                assert False
+    assert parts
+    add_instruction(macro.name, parts, instructions)
+
+
+def add_family(
+    pfamily: parser.Family,
+    instructions: dict[str, Instruction],
+    families: dict[str, Family],
+) -> None:
+    family = Family(
+        pfamily.name,
+        pfamily.size,
+        [instructions[member_name] for member_name in pfamily.members],
+    )
+    for member in family.members:
+        member.family = family
+    # The head of the family is an implicit jump target for DEOPTs
+    instructions[family.name].is_target = True
+    families[family.name] = family
+
+
+def add_pseudo(
+    pseudo: parser.Pseudo,
+    instructions: dict[str, Instruction],
+    pseudos: dict[str, PseudoInstruction],
+) -> None:
+    pseudos[pseudo.name] = PseudoInstruction(
+        pseudo.name,
+        [instructions[target] for target in pseudo.targets],
+        pseudo.flags,
+    )
+
+
+def analyze_forest(forest: list[parser.AstNode]) -> Analysis:
+    instructions: dict[str, Instruction] = {}
+    uops: dict[str, Uop] = {}
+    families: dict[str, Family] = {}
+    pseudos: dict[str, PseudoInstruction] = {}
+    for node in forest:
+        match node:
+            case parser.InstDef(name):
+                if node.kind == "inst":
+                    desugar_inst(node, instructions, uops)
+                else:
+                    assert node.kind == "op"
+                    add_op(node, uops)
+            case parser.Macro():
+                pass
+            case parser.Family():
+                pass
+            case parser.Pseudo():
+                pass
+            case _:
+                assert False
+    for node in forest:
+        if isinstance(node, parser.Macro):
+            add_macro(node, instructions, uops)
+    for node in forest:
+        match node:
+            case parser.Family():
+                add_family(node, instructions, families)
+            case parser.Pseudo():
+                add_pseudo(node, instructions, pseudos)
+            case _:
+                pass
+    for uop in uops.values():
+        tkn_iter = iter(uop.body)
+        for tkn in tkn_iter:
+            if tkn.kind == "IDENTIFIER" and tkn.text == "GO_TO_INSTRUCTION":
+                if next(tkn_iter).kind != "LPAREN":
+                    continue
+                target = next(tkn_iter)
+                if target.kind != "IDENTIFIER":
+                    continue
+                if target.text in instructions:
+                    instructions[target.text].is_target = True
+    # Hack
+    instructions["BINARY_OP_INPLACE_ADD_UNICODE"].family = families["BINARY_OP"]
+    return Analysis(instructions, uops, families, pseudos)
+
+
+def analyze_files(filenames: list[str]) -> Analysis:
+    return analyze_forest(parser.parse_files(filenames))
+
+
+def dump_analysis(analysis: Analysis) -> None:
+    print("Uops:")
+    for u in analysis.uops.values():
+        u.dump("    ")
+    print("Instructions:")
+    for i in analysis.instructions.values():
+        i.dump("    ")
+    print("Families:")
+    for f in analysis.families.values():
+        f.dump("    ")
+    print("Pseudos:")
+    for p in analysis.pseudos.values():
+        p.dump("    ")
+
+
+if __name__ == "__main__":
+    import sys
+
+    if len(sys.argv) < 2:
+        print("No input")
+    else:
+        filenames = sys.argv[1:]
+        dump_analysis(analyze_files(filenames))

Tools/cases_generator/cwriter.py

@@ -0,0 +1,111 @@
+from lexer import Token
+from typing import TextIO
+
+
+class CWriter:
+    "A writer that understands tokens and how to format C code"
+
+    last_token: Token | None
+
+    def __init__(self, out: TextIO, indent: int, line_directives: bool):
+        self.out = out
+        self.base_column = indent * 4
+        self.indents = [i * 4 for i in range(indent + 1)]
+        self.line_directives = line_directives
+        self.last_token = None
+        self.newline = True
+
+    def set_position(self, tkn: Token) -> None:
+        if self.last_token is not None:
+            if self.last_token.line < tkn.line:
+                self.out.write("\n")
+                if self.line_directives:
+                    self.out.write(f'#line {tkn.line} "{tkn.filename}"\n')
+                self.out.write(" " * self.indents[-1])
+            else:
+                gap = tkn.column - self.last_token.end_column
+                self.out.write(" " * gap)
+        elif self.newline:
+            self.out.write(" " * self.indents[-1])
+        self.last_token = tkn
+        self.newline = False
+
+    def emit_at(self, txt: str, where: Token) -> None:
+        self.set_position(where)
+        self.out.write(txt)
+
+    def maybe_dedent(self, txt: str) -> None:
+        parens = txt.count("(") - txt.count(")")
+        if parens < 0:
+            self.indents.pop()
+        elif "}" in txt or is_label(txt):
+            self.indents.pop()
+
+    def maybe_indent(self, txt: str) -> None:
+        parens = txt.count("(") - txt.count(")")
+        if parens > 0 and self.last_token:
+            offset = self.last_token.end_column - 1
+            if offset <= self.indents[-1] or offset > 40:
+                offset = self.indents[-1] + 4
+            self.indents.append(offset)
+        elif "{" in txt or is_label(txt):
+            self.indents.append(self.indents[-1] + 4)
+
+    def emit_text(self, txt: str) -> None:
+        self.out.write(txt)
+
+    def emit_multiline_comment(self, tkn: Token) -> None:
+        self.set_position(tkn)
+        lines = tkn.text.splitlines(True)
+        first = True
+        for line in lines:
+            text = line.lstrip()
+            if first:
+                spaces = 0
+            else:
+                spaces = self.indents[-1]
+                if text.startswith("*"):
+                    spaces += 1
+                else:
+                    spaces += 3
+            first = False
+            self.out.write(" " * spaces)
+            self.out.write(text)
+
+    def emit_token(self, tkn: Token) -> None:
+        if tkn.kind == "COMMENT" and "\n" in tkn.text:
+            return self.emit_multiline_comment(tkn)
+        self.maybe_dedent(tkn.text)
+        self.set_position(tkn)
+        self.emit_text(tkn.text)
+        self.maybe_indent(tkn.text)
+
+    def emit_str(self, txt: str) -> None:
+        self.maybe_dedent(txt)
+        if self.newline and txt:
+            if txt[0] != "\n":
+                self.out.write(" " * self.indents[-1])
+            self.newline = False
+        self.emit_text(txt)
+        if txt.endswith("\n"):
+            self.newline = True
+        self.maybe_indent(txt)
+        self.last_token = None
+
+    def emit(self, txt: str | Token) -> None:
+        if isinstance(txt, Token):
+            self.emit_token(txt)
+        elif isinstance(txt, str):
+            self.emit_str(txt)
+        else:
+            assert False
+
+    def start_line(self) -> None:
+        if not self.newline:
+            self.out.write("\n")
+        self.newline = True
+        self.last_token = None
+
+
+def is_label(txt: str) -> bool:
+    return not txt.startswith("//") and txt.endswith(":")

Tools/cases_generator/generate_cases.py

@@ -883,7 +883,6 @@ def main() -> None:
         return
 
     # 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)

Tools/cases_generator/lexer.py

@@ -112,7 +112,7 @@ def choice(*opts: str) -> str:
 char = r"\'.\'"  # TODO: escape sequence
 CHARACTER = "CHARACTER"
 
-comment_re = r"//.*|/\*([^*]|\*[^/])*\*/"
+comment_re = r"(//.*)|/\*([^*]|\*[^/])*\*/"
 COMMENT = "COMMENT"
 
 newline = r"\n"
@@ -234,6 +234,7 @@ def make_syntax_error(
 
 @dataclass(slots=True)
 class Token:
+    filename: str
     kind: str
     text: str
     begin: tuple[int, int]
@@ -261,7 +262,7 @@ def width(self) -> int:
 
     def replaceText(self, txt: str) -> "Token":
         assert isinstance(txt, str)
-        return Token(self.kind, txt, self.begin, self.end)
+        return Token(self.filename, self.kind, txt, self.begin, self.end)
 
     def __repr__(self) -> str:
         b0, b1 = self.begin
@@ -272,7 +273,7 @@ def __repr__(self) -> str:
             return f"{self.kind}({self.text!r}, {b0}:{b1}, {e0}:{e1})"
 
 
-def tokenize(src: str, line: int = 1, filename: str | None = None) -> Iterator[Token]:
+def tokenize(src: str, line: int = 1, filename: str = "") -> Iterator[Token]:
     linestart = -1
     for m in matcher.finditer(src):
         start, end = m.span()
@@ -323,7 +324,7 @@ def tokenize(src: str, line: int = 1, filename: str | None = None) -> Iterator[T
         else:
             begin = line, start - linestart
         if kind != "\n":
-            yield Token(kind, text, begin, (line, start - linestart + len(text)))
+            yield Token(filename, kind, text, begin, (line, start - linestart + len(text)))
 
 
 def to_text(tkns: list[Token], dedent: int = 0) -> str:

Tools/cases_generator/mypy.ini

@@ -11,3 +11,5 @@ strict = True
 strict_concatenate = True
 enable_error_code = ignore-without-code,redundant-expr,truthy-bool,possibly-undefined
 warn_unreachable = True
+allow_redefinition = True
+implicit_reexport = True

Tools/cases_generator/parser.py

@@ -0,0 +1,55 @@
+from parsing import (
+    InstDef,
+    Macro,
+    Pseudo,
+    Family,
+    Parser,
+    Context,
+    CacheEffect,
+    StackEffect,
+    OpName,
+    AstNode,
+)
+from formatting import prettify_filename
+
+
+BEGIN_MARKER = "// BEGIN BYTECODES //"
+END_MARKER = "// END BYTECODES //"
+
+
+def parse_files(filenames: list[str]) -> list[AstNode]:
+    result: list[AstNode] = []
+    for filename in filenames:
+        with open(filename) as file:
+            src = file.read()
+
+        psr = Parser(src, filename=prettify_filename(filename))
+
+        # Skip until begin marker
+        while tkn := psr.next(raw=True):
+            if tkn.text == BEGIN_MARKER:
+                break
+        else:
+            raise psr.make_syntax_error(
+                f"Couldn't find {BEGIN_MARKER!r} in {psr.filename}"
+            )
+        start = psr.getpos()
+
+        # Find end marker, then delete everything after it
+        while tkn := psr.next(raw=True):
+            if tkn.text == END_MARKER:
+                break
+        del psr.tokens[psr.getpos() - 1 :]
+
+        # Parse from start
+        psr.setpos(start)
+        thing_first_token = psr.peek()
+        while node := psr.definition():
+            assert node is not None
+            result.append(node)  # type: ignore[arg-type]
+        if not psr.eof():
+            psr.backup()
+            raise psr.make_syntax_error(
+                f"Extra stuff at the end of {filename}", psr.next(True)
+            )
+    return result

Tools/cases_generator/parsing.py

@@ -141,10 +141,11 @@ class Pseudo(Node):
     flags: list[str]   # instr flags to set on the pseudo instruction
     targets: list[str] # opcodes this can be replaced by
 
+AstNode = InstDef | Macro | Pseudo | Family
 
 class Parser(PLexer):
     @contextual
-    def definition(self) -> InstDef | Macro | Pseudo | Family | None:
+    def definition(self) -> AstNode | None:
         if macro := self.macro_def():
             return macro
         if family := self.family_def():

Tools/cases_generator/stack.py

@@ -0,0 +1,81 @@
+import sys
+from analyzer import StackItem
+from dataclasses import dataclass
+from formatting import maybe_parenthesize
+
+
+def var_size(var: StackItem) -> str:
+    if var.condition:
+        # Special case simplification
+        if var.condition == "oparg & 1" and var.size == "1":
+            return f"({var.condition})"
+        else:
+            return f"(({var.condition}) ? {var.size} : 0)"
+    else:
+        return var.size
+
+
+class StackOffset:
+    "The stack offset of the virtual base of the stack from the physical stack pointer"
+
+    def __init__(self) -> None:
+        self.popped: list[str] = []
+        self.pushed: list[str] = []
+
+    def pop(self, item: StackItem) -> None:
+        self.popped.append(var_size(item))
+
+    def push(self, item: StackItem) -> None:
+        self.pushed.append(var_size(item))
+
+    def simplify(self) -> None:
+        "Remove matching values from both the popped and pushed list"
+        if not self.popped or not self.pushed:
+            return
+        # Sort the list so the lexically largest element is last.
+        popped = sorted(self.popped)
+        pushed = sorted(self.pushed)
+        self.popped = []
+        self.pushed = []
+        while popped and pushed:
+            pop = popped.pop()
+            push = pushed.pop()
+            if pop == push:
+                pass
+            elif pop > push:
+                # if pop > push, there can be no element in pushed matching pop.
+                self.popped.append(pop)
+                pushed.append(push)
+            else:
+                self.pushed.append(push)
+                popped.append(pop)
+        self.popped.extend(popped)
+        self.pushed.extend(pushed)
+
+    def to_c(self) -> str:
+        self.simplify()
+        int_offset = 0
+        symbol_offset = ""
+        for item in self.popped:
+            try:
+                int_offset -= int(item)
+            except ValueError:
+                symbol_offset += f" - {maybe_parenthesize(item)}"
+        for item in self.pushed:
+            try:
+                int_offset += int(item)
+            except ValueError:
+                symbol_offset += f" + {maybe_parenthesize(item)}"
+        if symbol_offset and not int_offset:
+            res = symbol_offset
+        else:
+            res = f"{int_offset}{symbol_offset}"
+        if res.startswith(" + "):
+            res = res[3:]
+        if res.startswith(" - "):
+            res = "-" + res[3:]
+        return res
+
+    def clear(self) -> None:
+        self.popped = []
+        self.pushed = []

Tools/cases_generator/tier1_generator.py

@@ -0,0 +1,417 @@
+"""Generate the main interpreter switch.
+Reads the instruction definitions from bytecodes.c.
+Writes the cases to generated_cases.c.h, which is #included in ceval.c.
+"""
+
+import argparse
+import os.path
+import sys
+
+from analyzer import (
+    Analysis,
+    Instruction,
+    Uop,
+    Part,
+    analyze_files,
+    Skip,
+    StackItem,
+    analysis_error,
+)
+from cwriter import CWriter
+from typing import TextIO, Iterator
+from lexer import Token
+from stack import StackOffset
+
+
+HERE = os.path.dirname(__file__)
+ROOT = os.path.join(HERE, "../..")
+THIS = os.path.relpath(__file__, ROOT).replace(os.path.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"))
+
+
+def write_header(filename: str, outfile: TextIO) -> None:
+    outfile.write(
+        f"""// This file is generated by {THIS}
+// from:
+//   {filename}
+// Do not edit!
+
+#ifdef TIER_TWO
+    #error "This file is for Tier 1 only"
+#endif
+#define TIER_ONE 1
+"""
+    )
+
+
+FOOTER = "#undef TIER_ONE\n"
+
+
+class SizeMismatch(Exception):
+    pass
+
+
+class Stack:
+    def __init__(self) -> None:
+        self.top_offset = StackOffset()
+        self.base_offset = StackOffset()
+        self.peek_offset = StackOffset()
+        self.variables: list[StackItem] = []
+        self.defined: set[str] = set()
+
+    def pop(self, var: StackItem) -> str:
+        self.top_offset.pop(var)
+        if not var.peek:
+            self.peek_offset.pop(var)
+        indirect = "&" if var.is_array() else ""
+        if self.variables:
+            popped = self.variables.pop()
+            if popped.size != var.size:
+                raise SizeMismatch(
+                    f"Size mismatch when popping '{popped.name}' from stack to assign to {var.name}. "
+                    f"Expected {var.size} got {popped.size}"
+                )
+            if popped.name == var.name:
+                return ""
+            elif popped.name == "unused":
+                self.defined.add(var.name)
+                return (
+                    f"{var.name} = {indirect}stack_pointer[{self.top_offset.to_c()}];\n"
+                )
+            elif var.name == "unused":
+                return ""
+            else:
+                self.defined.add(var.name)
+                return f"{var.name} = {popped.name};\n"
+        self.base_offset.pop(var)
+        if var.name == "unused":
+            return ""
+        else:
+            self.defined.add(var.name)
+        assign = f"{var.name} = {indirect}stack_pointer[{self.base_offset.to_c()}];"
+        if var.condition:
+            return f"if ({var.condition}) {{ {assign} }}\n"
+        return f"{assign}\n"
+
+    def push(self, var: StackItem) -> str:
+        self.variables.append(var)
+        if var.is_array() and var.name not in self.defined and var.name != "unused":
+            c_offset = self.top_offset.to_c()
+            self.top_offset.push(var)
+            self.defined.add(var.name)
+            return f"{var.name} = &stack_pointer[{c_offset}];\n"
+        else:
+            self.top_offset.push(var)
+            return ""
+
+    def flush(self, out: CWriter) -> None:
+        for var in self.variables:
+            if not var.peek:
+                if var.name != "unused" and not var.is_array():
+                    if var.condition:
+                        out.emit(f" if ({var.condition}) ")
+                    out.emit(
+                        f"stack_pointer[{self.base_offset.to_c()}] = {var.name};\n"
+                    )
+            self.base_offset.push(var)
+        if self.base_offset.to_c() != self.top_offset.to_c():
+            print("base", self.base_offset.to_c(), "top", self.top_offset.to_c())
+            assert False
+        number = self.base_offset.to_c()
+        if number != "0":
+            out.emit(f"stack_pointer += {number};\n")
+        self.variables = []
+        self.base_offset.clear()
+        self.top_offset.clear()
+        self.peek_offset.clear()
+
+    def as_comment(self) -> str:
+        return f"/* Variables: {[v.name for v in self.variables]}. Base offset: {self.base_offset.to_c()}. Top offset: {self.top_offset.to_c()} */"
+
+
+def declare_variables(inst: Instruction, out: CWriter) -> None:
+    variables = {"unused"}
+    for uop in inst.parts:
+        if isinstance(uop, Uop):
+            for var in reversed(uop.stack.inputs):
+                if var.name not in variables:
+                    type = var.type if var.type else "PyObject *"
+                    variables.add(var.name)
+                    if var.condition:
+                        out.emit(f"{type}{var.name} = NULL;\n")
+                    else:
+                        out.emit(f"{type}{var.name};\n")
+            for var in uop.stack.outputs:
+                if var.name not in variables:
+                    variables.add(var.name)
+                    type = var.type if var.type else "PyObject *"
+                    if var.condition:
+                        out.emit(f"{type}{var.name} = NULL;\n")
+                    else:
+                        out.emit(f"{type}{var.name};\n")
+
+
+def emit_to(out: CWriter, tkn_iter: Iterator[Token], end: str) -> None:
+    parens = 0
+    for tkn in tkn_iter:
+        if tkn.kind == end and parens == 0:
+            return
+        if tkn.kind == "LPAREN":
+            parens += 1
+        if tkn.kind == "RPAREN":
+            parens -= 1
+        out.emit(tkn)
+
+
+def replace_deopt(
+    out: CWriter,
+    tkn: Token,
+    tkn_iter: Iterator[Token],
+    uop: Uop,
+    unused: Stack,
+    inst: Instruction,
+) -> None:
+    out.emit_at("DEOPT_IF", tkn)
+    out.emit(next(tkn_iter))
+    emit_to(out, tkn_iter, "RPAREN")
+    next(tkn_iter)  # Semi colon
+    out.emit(", ")
+    assert inst.family is not None
+    out.emit(inst.family.name)
+    out.emit(");\n")
+
+
+def replace_error(
+    out: CWriter,
+    tkn: Token,
+    tkn_iter: Iterator[Token],
+    uop: Uop,
+    stack: Stack,
+    inst: Instruction,
+) -> None:
+    out.emit_at("if ", tkn)
+    out.emit(next(tkn_iter))
+    emit_to(out, tkn_iter, "COMMA")
+    label = next(tkn_iter).text
+    next(tkn_iter)  # RPAREN
+    next(tkn_iter)  # Semi colon
+    out.emit(") ")
+    c_offset = stack.peek_offset.to_c()
+    try:
+        offset = -int(c_offset)
+        close = ";\n"
+    except ValueError:
+        offset = None
+        out.emit(f"{{ stack_pointer += {c_offset}; ")
+        close = "; }\n"
+    out.emit("goto ")
+    if offset:
+        out.emit(f"pop_{offset}_")
+    out.emit(label)
+    out.emit(close)
+
+
+def replace_decrefs(
+    out: CWriter,
+    tkn: Token,
+    tkn_iter: Iterator[Token],
+    uop: Uop,
+    stack: Stack,
+    inst: Instruction,
+) -> None:
+    next(tkn_iter)
+    next(tkn_iter)
+    next(tkn_iter)
+    out.emit_at("", tkn)
+    for var in uop.stack.inputs:
+        if var.name == "unused" or var.name == "null" or var.peek:
+            continue
+        if var.size != "1":
+            out.emit(f"for (int _i = {var.size}; --_i >= 0;) {{\n")
+            out.emit(f"Py_DECREF({var.name}[_i]);\n")
+            out.emit("}\n")
+        elif var.condition:
+            out.emit(f"Py_XDECREF({var.name});\n")
+        else:
+            out.emit(f"Py_DECREF({var.name});\n")
+
+
+def replace_store_sp(
+    out: CWriter,
+    tkn: Token,
+    tkn_iter: Iterator[Token],
+    uop: Uop,
+    stack: Stack,
+    inst: Instruction,
+) -> None:
+    next(tkn_iter)
+    next(tkn_iter)
+    next(tkn_iter)
+    out.emit_at("", tkn)
+    stack.flush(out)
+    out.emit("_PyFrame_SetStackPointer(frame, stack_pointer);\n")
+
+
+def replace_check_eval_breaker(
+    out: CWriter,
+    tkn: Token,
+    tkn_iter: Iterator[Token],
+    uop: Uop,
+    stack: Stack,
+    inst: Instruction,
+) -> None:
+    next(tkn_iter)
+    next(tkn_iter)
+    next(tkn_iter)
+    if not uop.properties.ends_with_eval_breaker:
+        out.emit_at("CHECK_EVAL_BREAKER();", tkn)
+
+
+REPLACEMENT_FUNCTIONS = {
+    "DEOPT_IF": replace_deopt,
+    "ERROR_IF": replace_error,
+    "DECREF_INPUTS": replace_decrefs,
+    "CHECK_EVAL_BREAKER": replace_check_eval_breaker,
+    "STORE_SP": replace_store_sp,
+}
+
+
+# Move this to formatter
+def emit_tokens(out: CWriter, uop: Uop, stack: Stack, inst: Instruction) -> None:
+    tkns = uop.body[1:-1]
+    if not tkns:
+        return
+    tkn_iter = iter(tkns)
+    out.start_line()
+    for tkn in tkn_iter:
+        if tkn.kind == "IDENTIFIER" and tkn.text in REPLACEMENT_FUNCTIONS:
+            REPLACEMENT_FUNCTIONS[tkn.text](out, tkn, tkn_iter, uop, stack, inst)
+        else:
+            out.emit(tkn)
+
+
+def write_uop(
+    uop: Part, out: CWriter, offset: int, stack: Stack, inst: Instruction, braces: bool
+) -> int:
+    # out.emit(stack.as_comment() + "\n")
+    if isinstance(uop, Skip):
+        entries = "entries" if uop.size > 1 else "entry"
+        out.emit(f"/* Skip {uop.size} cache {entries} */\n")
+        return offset + uop.size
+    try:
+        out.start_line()
+        if braces:
+            out.emit(f"// {uop.name}\n")
+        for var in reversed(uop.stack.inputs):
+            out.emit(stack.pop(var))
+        if braces:
+            out.emit("{\n")
+        if not uop.properties.stores_sp:
+            for i, var in enumerate(uop.stack.outputs):
+                out.emit(stack.push(var))
+        for cache in uop.caches:
+            if cache.name != "unused":
+                if cache.size == 4:
+                    type = "PyObject *"
+                    reader = "read_obj"
+                else:
+                    type = f"uint{cache.size*16}_t "
+                    reader = f"read_u{cache.size*16}"
+                out.emit(
+                    f"{type}{cache.name} = {reader}(&this_instr[{offset}].cache);\n"
+                )
+            offset += cache.size
+        emit_tokens(out, uop, stack, inst)
+        if uop.properties.stores_sp:
+            for i, var in enumerate(uop.stack.outputs):
+                out.emit(stack.push(var))
+        if braces:
+            out.start_line()
+            out.emit("}\n")
+        # out.emit(stack.as_comment() + "\n")
+        return offset
+    except SizeMismatch as ex:
+        raise analysis_error(ex.args[0], uop.body[0])
+
+
+def uses_this(inst: Instruction) -> bool:
+    if inst.properties.needs_this:
+        return True
+    for uop in inst.parts:
+        if isinstance(uop, Skip):
+            continue
+        for cache in uop.caches:
+            if cache.name != "unused":
+                return True
+    return False
+
+
+def generate_tier1(
+    filenames: str, analysis: Analysis, outfile: TextIO, lines: bool
+) -> None:
+    write_header(filenames, outfile)
+    out = CWriter(outfile, 2, lines)
+    out.emit("\n")
+    for name, inst in sorted(analysis.instructions.items()):
+        needs_this = uses_this(inst)
+        out.emit("\n")
+        out.emit(f"TARGET({name}) {{\n")
+        if needs_this and not inst.is_target:
+            out.emit(f"_Py_CODEUNIT *this_instr = frame->instr_ptr = next_instr;\n")
+        else:
+            out.emit(f"frame->instr_ptr = next_instr;\n")
+        out.emit(f"next_instr += {inst.size};\n")
+        out.emit(f"INSTRUCTION_STATS({name});\n")
+        if inst.is_target:
+            out.emit(f"PREDICTED({name});\n")
+            if needs_this:
+                out.emit(f"_Py_CODEUNIT *this_instr = next_instr - {inst.size};\n")
+        if inst.family is not None:
+            out.emit(
+                f"static_assert({inst.family.size} == {inst.size-1}"
+                ', "incorrect cache size");\n'
+            )
+        declare_variables(inst, out)
+        offset = 1  # The instruction itself
+        stack = Stack()
+        for part in inst.parts:
+            # Only emit braces if more than one uop
+            offset = write_uop(part, out, offset, stack, inst, len(inst.parts) > 1)
+        out.start_line()
+        if not inst.parts[-1].properties.always_exits:
+            stack.flush(out)
+            if inst.parts[-1].properties.ends_with_eval_breaker:
+                out.emit("CHECK_EVAL_BREAKER();\n")
+            out.emit("DISPATCH();\n")
+        out.start_line()
+        out.emit("}")
+        out.emit("\n")
+    outfile.write(FOOTER)
+
+
+arg_parser = argparse.ArgumentParser(
+    description="Generate the code for the interpreter switch.",
+    formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+)
+
+arg_parser.add_argument(
+    "-o", "--output", type=str, help="Generated code", default=DEFAULT_OUTPUT
+)
+
+arg_parser.add_argument(
+    "-l", "--emit-line-directives", help="Emit #line directives", action="store_true"
+)
+
+arg_parser.add_argument(
+    "input", nargs=argparse.REMAINDER, help="Instruction definition file(s)"
+)
+
+if __name__ == "__main__":
+    args = arg_parser.parse_args()
+    if len(args.input) == 0:
+        args.input.append(DEFAULT_INPUT)
+    data = analyze_files(args.input)
+    with open(args.output, "w") as outfile:
+        generate_tier1(args.input, data, outfile, args.emit_line_directives)