[mypyc] Merge yield_from_except_op (#9660)

relates mypyc/mypyc#753

This PR merges yield_from_except_op into the new IR, including several changes:

A change to C wrapper's signature since the new IR currently has no unified way to represent the address of a pointer.
A change to LoadAddress, allowing it to load local reg's address.
A change to uninit pass, suppressing checks on LoadAddress

Author: TH3CHARLie

mypyc/codegen/emitfunc.py

@@ -12,7 +12,7 @@
     LoadStatic, InitStatic, TupleGet, TupleSet, Call, IncRef, DecRef, Box, Cast, Unbox,
     BasicBlock, Value, MethodCall, PrimitiveOp, EmitterInterface, Unreachable, NAMESPACE_STATIC,
     NAMESPACE_TYPE, NAMESPACE_MODULE, RaiseStandardError, CallC, LoadGlobal, Truncate,
-    BinaryIntOp, LoadMem, GetElementPtr, LoadAddress, ComparisonOp, SetMem
+    BinaryIntOp, LoadMem, GetElementPtr, LoadAddress, ComparisonOp, SetMem, Register
 )
 from mypyc.ir.rtypes import (
     RType, RTuple, is_tagged, is_int32_rprimitive, is_int64_rprimitive, RStruct,
@@ -496,7 +496,8 @@ def visit_get_element_ptr(self, op: GetElementPtr) -> None:
     def visit_load_address(self, op: LoadAddress) -> None:
         typ = op.type
         dest = self.reg(op)
-        self.emit_line('%s = (%s)&%s;' % (dest, typ._ctype, op.src))
+        src = self.reg(op.src) if isinstance(op.src, Register) else op.src
+        self.emit_line('%s = (%s)&%s;' % (dest, typ._ctype, src))
 
     # Helpers
 

mypyc/ir/ops.py

@@ -1502,19 +1502,34 @@ def accept(self, visitor: 'OpVisitor[T]') -> T:
 
 
 class LoadAddress(RegisterOp):
+    """Get the address of a value
+
+    ret = (type)&src
+
+    Attributes:
+      type: Type of the loaded address(e.g. ptr/object_ptr)
+      src: Source value, str for named constants like 'PyList_Type',
+           Register for temporary values
+    """
     error_kind = ERR_NEVER
     is_borrowed = True
 
-    def __init__(self, type: RType, src: str, line: int = -1) -> None:
+    def __init__(self, type: RType, src: Union[str, Register], line: int = -1) -> None:
         super().__init__(line)
         self.type = type
         self.src = src
 
     def sources(self) -> List[Value]:
-        return []
+        if isinstance(self.src, Register):
+            return [self.src]
+        else:
+            return []
 
     def to_str(self, env: Environment) -> str:
-        return env.format("%r = load_address %s", self, self.src)
+        if isinstance(self.src, Register):
+            return env.format("%r = load_address %r", self, self.src)
+        else:
+            return env.format("%r = load_address %s", self, self.src)
 
     def accept(self, visitor: 'OpVisitor[T]') -> T:
         return visitor.visit_load_address(self)

mypyc/ir/rtypes.py

@@ -191,6 +191,8 @@ def __init__(self,
             self.c_undefined = 'NULL'
         elif ctype == 'char':
             self.c_undefined = '2'
+        elif ctype == 'PyObject **':
+            self.c_undefined = 'NULL'
         else:
             assert False, 'Unrecognized ctype: %r' % ctype
 
@@ -223,6 +225,10 @@ def __repr__(self) -> str:
 object_rprimitive = RPrimitive('builtins.object', is_unboxed=False,
                                is_refcounted=True)  # type: Final
 
+# represents a low level pointer of an object
+object_pointer_rprimitive = RPrimitive('object_ptr', is_unboxed=False,
+                               is_refcounted=False, ctype='PyObject **')  # type: Final
+
 # Arbitrary-precision integer (corresponds to Python 'int'). Small
 # enough values are stored unboxed, while large integers are
 # represented as a tagged pointer to a Python 'int' PyObject. The

mypyc/irbuild/function.py

@@ -20,9 +20,9 @@
 
 from mypyc.ir.ops import (
     BasicBlock, Value,  Return, SetAttr, LoadInt, Environment, GetAttr, Branch, AssignmentTarget,
-    TupleGet, InitStatic
+    InitStatic, LoadAddress
 )
-from mypyc.ir.rtypes import object_rprimitive, RInstance
+from mypyc.ir.rtypes import object_rprimitive, RInstance, object_pointer_rprimitive
 from mypyc.ir.func_ir import (
     FuncIR, FuncSignature, RuntimeArg, FuncDecl, FUNC_CLASSMETHOD, FUNC_STATICMETHOD, FUNC_NORMAL
 )
@@ -523,9 +523,10 @@ def except_body() -> None:
         # The body of the except is all implemented in a C function to
         # reduce how much code we need to generate. It returns a value
         # indicating whether to break or yield (or raise an exception).
-        res = builder.primitive_op(yield_from_except_op, [builder.read(iter_reg)], o.line)
-        to_stop = builder.add(TupleGet(res, 0, o.line))
-        val = builder.add(TupleGet(res, 1, o.line))
+        val = builder.alloc_temp(object_rprimitive)
+        val_address = builder.add(LoadAddress(object_pointer_rprimitive, val))
+        to_stop = builder.call_c(yield_from_except_op,
+                                 [builder.read(iter_reg), val_address], o.line)
 
         ok, stop = BasicBlock(), BasicBlock()
         builder.add(Branch(to_stop, stop, ok, Branch.BOOL))

mypyc/primitives/misc_ops.py

@@ -2,11 +2,11 @@
 
 from mypyc.ir.ops import ERR_NEVER, ERR_MAGIC, ERR_FALSE
 from mypyc.ir.rtypes import (
-    RTuple, bool_rprimitive, object_rprimitive, str_rprimitive,
+    bool_rprimitive, object_rprimitive, str_rprimitive, object_pointer_rprimitive,
     int_rprimitive, dict_rprimitive, c_int_rprimitive, bit_rprimitive
 )
 from mypyc.primitives.registry import (
-    simple_emit, custom_op, c_function_op, c_custom_op, load_address_op, ERR_NEG_INT
+    c_function_op, c_custom_op, load_address_op, ERR_NEG_INT
 )
 
 
@@ -55,21 +55,19 @@
     error_kind=ERR_NEVER)
 
 # This is sort of unfortunate but oh well: yield_from_except performs most of the
-# error handling logic in `yield from` operations. It returns a bool and a value.
+# error handling logic in `yield from` operations. It returns a bool and passes
+# a value by address.
 # If the bool is true, then a StopIteration was received and we should return.
 # If the bool is false, then the value should be yielded.
 # The normal case is probably that it signals an exception, which gets
 # propagated.
-yield_from_rtuple = RTuple([bool_rprimitive, object_rprimitive])
-
 # Op used for "yield from" error handling.
 # See comment in CPy_YieldFromErrorHandle for more information.
-yield_from_except_op = custom_op(
-    name='yield_from_except',
-    arg_types=[object_rprimitive],
-    result_type=yield_from_rtuple,
-    error_kind=ERR_MAGIC,
-    emit=simple_emit('{dest}.f0 = CPy_YieldFromErrorHandle({args[0]}, &{dest}.f1);'))
+yield_from_except_op = c_custom_op(
+    arg_types=[object_rprimitive, object_pointer_rprimitive],
+    return_type=bool_rprimitive,
+    c_function_name='CPy_YieldFromErrorHandle',
+    error_kind=ERR_MAGIC)
 
 # Create method object from a callable object and self.
 method_new_op = c_custom_op(

mypyc/transform/uninit.py

@@ -9,7 +9,8 @@
     AnalysisDict
 )
 from mypyc.ir.ops import (
-    BasicBlock, Branch, Value, RaiseStandardError, Unreachable, Environment, Register
+    BasicBlock, Branch, Value, RaiseStandardError, Unreachable, Environment, Register,
+    LoadAddress
 )
 from mypyc.ir.func_ir import FuncIR
 
@@ -44,8 +45,13 @@ def split_blocks_at_uninits(env: Environment,
                 # If a register operand is not guaranteed to be
                 # initialized is an operand to something other than a
                 # check that it is defined, insert a check.
+
+                # Note that for register operand in a LoadAddress op,
+                # we should be able to use it without initialization
+                # as we may need to use its address to update itself
                 if (isinstance(src, Register) and src not in defined
-                        and not (isinstance(op, Branch) and op.op == Branch.IS_ERROR)):
+                        and not (isinstance(op, Branch) and op.op == Branch.IS_ERROR)
+                        and not isinstance(op, LoadAddress)):
                     new_block, error_block = BasicBlock(), BasicBlock()
                     new_block.error_handler = error_block.error_handler = cur_block.error_handler
                     new_blocks += [error_block, new_block]