feature[next]: Non-tree-size-increasing collapse tuple on ifs

src/gt4py/next/iterator/transforms/collapse_tuple.py

@@ -28,10 +28,11 @@
 from gt4py.next.type_system import type_info, type_specifications as ts
 
 
-def _with_altered_arg(node: ir.FunCall, arg_idx: int, new_arg: ir.Expr):
+def _with_altered_arg(node: ir.FunCall, arg_idx: int, new_arg: ir.Expr | str):
     """Given a itir.FunCall return a new call with one of its argument replaced."""
     return ir.FunCall(
-        fun=node.fun, args=[arg if i != arg_idx else new_arg for i, arg in enumerate(node.args)]
+        fun=node.fun,
+        args=[arg if i != arg_idx else im.ensure_expr(new_arg) for i, arg in enumerate(node.args)],
     )
 
 
@@ -47,6 +48,32 @@ def _is_trivial_make_tuple_call(node: ir.Expr):
     return True
 
 
+def _is_trivial_or_tuple_thereof_expr(node: ir.Node) -> bool:
+    """
+    Return `true` if the expr is a trivial expression or tuple thereof.
+
+    >>> _is_trivial_or_tuple_thereof_expr(im.make_tuple("a", "b"))
+    True
+    >>> _is_trivial_or_tuple_thereof_expr(im.tuple_get(1, "a"))
+    True
+    >>> _is_trivial_or_tuple_thereof_expr(
+    ...     im.let("t", im.make_tuple("a", "b"))(im.tuple_get(1, "t"))
+    ... )
+    True
+    """
+    if cpm.is_call_to(node, "make_tuple"):
+        return all(_is_trivial_or_tuple_thereof_expr(arg) for arg in node.args)
+    if cpm.is_call_to(node, "tuple_get"):
+        return _is_trivial_or_tuple_thereof_expr(node.args[1])
+    if isinstance(node, (ir.SymRef, ir.Literal)):
+        return True
+    if cpm.is_let(node):
+        return _is_trivial_or_tuple_thereof_expr(node.fun.expr) and all(  # type: ignore[attr-defined]  # ensured by is_let
+            _is_trivial_or_tuple_thereof_expr(arg) for arg in node.args
+        )
+    return False
+
+
 # TODO(tehrengruber): Conceptually the structure of this pass makes sense: Visit depth first,
 #  transform each node until no transformations apply anymore, whenever a node is to be transformed
 #  go through all available transformation and apply them. However the final result here still
@@ -76,28 +103,42 @@ class Flag(enum.Flag):
         #: `let(tup, {trivial_expr1, trivial_expr2})(foo(tup))`
         #:  -> `foo({trivial_expr1, trivial_expr2})`
         INLINE_TRIVIAL_MAKE_TUPLE = enum.auto()
+        #:  Similar as `PROPAGATE_TO_IF_ON_TUPLES`, but propagates in the opposite direction, i.e.
+        #:  into the tree, allowing removal of tuple expressions across `if_` calls without
+        #:  increasing the size of the tree. This is particularly important for `if` statements
+        #:  in the frontend, where outwards propagation can have devastating effects on the tree
+        #:  size, without any gained optimization potential. For example
+        #:  ```
+        #:   complex_lambda(if cond1
+        #:     if cond2
+        #:       {...}
+        #:     else:
+        #:       {...}
+        #:   else
+        #:     {...})
+        #:  ```
+        #:  is problematic, since `PROPAGATE_TO_IF_ON_TUPLES` would propagate and hence duplicate
+        #:  `complex_lambda` three times, while we only want to get rid of the tuple expressions
+        #:  inside of the `if_`s.
+        #:  Note that this transformation is not mutually exclusive to `PROPAGATE_TO_IF_ON_TUPLES`.
+        PROPAGATE_TO_IF_ON_TUPLES_CPS = enum.auto()
         #: `(if cond then {1, 2} else {3, 4})[0]` -> `if cond then {1, 2}[0] else {3, 4}[0]`
         PROPAGATE_TO_IF_ON_TUPLES = enum.auto()
         #: `let((a, let(b, 1)(a_val)))(a)`-> `let(b, 1)(let(a, a_val)(a))`
         PROPAGATE_NESTED_LET = enum.auto()
-        #: `let(a, 1)(a)` -> `1`
+        #: `let(a, 1)(a)` -> `1` or `let(a, b)(f(a))` -> `f(a)`
         INLINE_TRIVIAL_LET = enum.auto()
 
         @classmethod
         def all(self) -> CollapseTuple.Flag:
             return functools.reduce(operator.or_, self.__members__.values())
 
+    uids: eve_utils.UIDGenerator
     ignore_tuple_size: bool
     flags: Flag = Flag.all()  # noqa: RUF009 [function-call-in-dataclass-default-argument]
 
     PRESERVED_ANNEX_ATTRS = ("type",)
 
-    # we use one UID generator per instance such that the generated ids are
-    # stable across multiple runs (required for caching to properly work)
-    _letify_make_tuple_uids: eve_utils.UIDGenerator = dataclasses.field(
-        init=False, repr=False, default_factory=lambda: eve_utils.UIDGenerator(prefix="_tuple_el")
-    )
-
     @classmethod
     def apply(
         cls,
@@ -111,6 +152,7 @@ def apply(
         flags: Optional[Flag] = None,
         # allow sym references without a symbol declaration, mostly for testing
         allow_undeclared_symbols: bool = False,
+        uids: Optional[eve_utils.UIDGenerator] = None,
     ) -> ir.Node:
         """
         Simplifies `make_tuple`, `tuple_get` calls.
@@ -127,6 +169,7 @@ def apply(
         """
         flags = flags or cls.flags
         offset_provider_type = offset_provider_type or {}
+        uids = uids or eve_utils.UIDGenerator()
 
         if isinstance(node, (ir.Program, ir.FencilDefinition)):
             within_stencil = False
@@ -145,6 +188,7 @@ def apply(
         new_node = cls(
             ignore_tuple_size=ignore_tuple_size,
             flags=flags,
+            uids=uids,
         ).visit(node, within_stencil=within_stencil)
 
         # inline to remove left-overs from LETIFY_MAKE_TUPLE_ELEMENTS. this is important
@@ -185,6 +229,8 @@ def transform(self, node: ir.Node, **kwargs) -> Optional[ir.Node]:
                 method = getattr(self, f"transform_{transformation.name.lower()}")
                 result = method(node, **kwargs)
                 if result is not None:
+                    assert result is not node
+                    itir_type_inference.reinfer(result)
                     return result
         return None
 
@@ -263,13 +309,13 @@ def transform_letify_make_tuple_elements(self, node: ir.FunCall, **kwargs) -> Op
         if node.fun == ir.SymRef(id="make_tuple"):
             # `make_tuple(expr1, expr1)`
             # -> `let((_tuple_el_1, expr1), (_tuple_el_2, expr2))(make_tuple(_tuple_el_1, _tuple_el_2))`
-            bound_vars: dict[str, ir.Expr] = {}
+            bound_vars: dict[ir.Sym, ir.Expr] = {}
             new_args: list[ir.Expr] = []
             for arg in node.args:
                 if cpm.is_call_to(node, "make_tuple") and not _is_trivial_make_tuple_call(node):
-                    el_name = self._letify_make_tuple_uids.sequential_id()
-                    new_args.append(im.ref(el_name))
-                    bound_vars[el_name] = arg
+                    el_name = self.uids.sequential_id(prefix="__ct_el")
+                    new_args.append(im.ref(el_name, arg.type))
+                    bound_vars[im.sym(el_name, arg.type)] = arg
                 else:
                     new_args.append(arg)
 
@@ -312,6 +358,73 @@ def transform_propagate_to_if_on_tuples(self, node: ir.FunCall, **kwargs) -> Opt
                     return im.if_(cond, new_true_branch, new_false_branch)
         return None
 
+    def transform_propagate_to_if_on_tuples_cps(
+        self, node: ir.FunCall, **kwargs
+    ) -> Optional[ir.Node]:
+        if not cpm.is_call_to(node, "if_"):
+            for i, arg in enumerate(node.args):
+                if cpm.is_call_to(arg, "if_"):
+                    itir_type_inference.reinfer(arg)
+                    if not any(isinstance(branch.type, ts.TupleType) for branch in arg.args[1:]):
+                        continue
+
+                    cond, true_branch, false_branch = arg.args
+                    tuple_type: ts.TupleType = true_branch.type  # type: ignore[assignment]  # type ensured above
+                    tuple_len = len(tuple_type.types)
+                    itir_type_inference.reinfer(node)
+                    assert node.type
+
+                    # transform function into continuation-passing-style
+                    f_type = ts.FunctionType(
+                        pos_only_args=tuple_type.types,
+                        pos_or_kw_args={},
+                        kw_only_args={},
+                        returns=node.type,
+                    )
+                    f_params = [
+                        im.sym(self.uids.sequential_id(prefix="__ct_el_cps"), type_)
+                        for type_ in tuple_type.types
+                    ]
+                    f_args = [im.ref(param.id, param.type) for param in f_params]
+                    f_body = _with_altered_arg(node, i, im.make_tuple(*f_args))
+                    # simplify, e.g., inline trivial make_tuple args
+                    new_f_body = self.fp_transform(f_body, **kwargs)
+                    # if the function did not simplify there is nothing to gain. Skip
+                    # transformation.
+                    if new_f_body is f_body:
+                        continue
+                    # if the function is not trivial the transformation would still work, but
+                    # inlining would result in a larger tree again and we didn't didn't gain
+                    # anything compared to regular `propagate_to_if_on_tuples`. Not inling also
+                    # works, but we don't want bound lambda functions in our tree (at least right
+                    # now).
+                    if not _is_trivial_or_tuple_thereof_expr(new_f_body):
+                        continue
+                    f = im.lambda_(*f_params)(new_f_body)
+
+                    tuple_var = self.uids.sequential_id(prefix="__ct_tuple_cps")
+                    f_var = self.uids.sequential_id(prefix="__ct_cont")
+                    new_branches = []
+                    for branch in arg.args[1:]:
+                        new_branch = im.let(tuple_var, branch)(
+                            im.call(im.ref(f_var, f_type))(
+                                *(
+                                    im.tuple_get(i, im.ref(tuple_var, branch.type))
+                                    for i in range(tuple_len)
+                                )
+                            )
+                        )
+                        new_branches.append(self.fp_transform(new_branch, **kwargs))
+
+                    new_node = im.let(f_var, f)(im.if_(cond, *new_branches))
+                    new_node = inline_lambda(new_node, eligible_params=[True])
+                    assert cpm.is_call_to(new_node, "if_")
+                    new_node = im.if_(
+                        cond, *(self.fp_transform(branch, **kwargs) for branch in new_node.args[1:])
+                    )
+                    return new_node
+        return None
+
     def transform_propagate_nested_let(self, node: ir.FunCall, **kwargs) -> Optional[ir.Node]:
         if cpm.is_let(node):
             # `let((a, let(b, 1)(a_val)))(a)`-> `let(b, 1)(let(a, a_val)(a))`
@@ -339,9 +452,13 @@ def transform_propagate_nested_let(self, node: ir.FunCall, **kwargs) -> Optional
         return None
 
     def transform_inline_trivial_let(self, node: ir.FunCall, **kwargs) -> Optional[ir.Node]:
-        if cpm.is_let(node) and isinstance(node.fun.expr, ir.SymRef):  # type: ignore[attr-defined]  # ensured by is_let
-            # `let(a, 1)(a)` -> `1`
-            for arg_sym, arg in zip(node.fun.params, node.args):  # type: ignore[attr-defined]  # ensured by is_let
-                if isinstance(node.fun.expr, ir.SymRef) and node.fun.expr.id == arg_sym.id:  # type: ignore[attr-defined]  # ensured by is_let
-                    return arg
+        if cpm.is_let(node):
+            if isinstance(node.fun.expr, ir.SymRef):  # type: ignore[attr-defined]  # ensured by is_let
+                # `let(a, 1)(a)` -> `1`
+                for arg_sym, arg in zip(node.fun.params, node.args):  # type: ignore[attr-defined]  # ensured by is_let
+                    if isinstance(node.fun.expr, ir.SymRef) and node.fun.expr.id == arg_sym.id:  # type: ignore[attr-defined]  # ensured by is_let
+                        return arg
+            if any(trivial_args := [isinstance(arg, (ir.SymRef, ir.Literal)) for arg in node.args]):
+                return inline_lambda(node, eligible_params=trivial_args)
+
         return None

src/gt4py/next/iterator/transforms/pass_manager.py

@@ -69,6 +69,7 @@ def apply_common_transforms(
 
     tmp_uids = eve_utils.UIDGenerator(prefix="__tmp")
     mergeasfop_uids = eve_utils.UIDGenerator()
+    collapse_tuple_uids = eve_utils.UIDGenerator()
 
     ir = MergeLet().visit(ir)
     ir = inline_fundefs.InlineFundefs().visit(ir)
@@ -80,7 +81,12 @@ def apply_common_transforms(
     # Inline. The domain inference can not handle "user" functions, e.g. `let f = λ(...) → ... in f(...)`
     ir = InlineLambdas.apply(ir, opcount_preserving=True, force_inline_lambda_args=True)
     # required in order to get rid of expressions without a domain (e.g. when a tuple element is never accessed)
-    ir = CollapseTuple.apply(ir, offset_provider_type=offset_provider_type)  # type: ignore[assignment]  # always an itir.Program
+    ir = CollapseTuple.apply(
+        ir,
+        flags=~CollapseTuple.Flag.PROPAGATE_TO_IF_ON_TUPLES,
+        uids=collapse_tuple_uids,
+        offset_provider_type=offset_provider_type,
+    )  # type: ignore[assignment]  # always an itir.Program
     ir = infer_domain.infer_program(
         ir,  # type: ignore[arg-type]  # always an itir.Program
         offset_provider=offset_provider,
@@ -94,7 +100,12 @@ def apply_common_transforms(
         inlined = ConstantFolding.apply(inlined)  # type: ignore[assignment]  # always an itir.Program
         # This pass is required to be in the loop such that when an `if_` call with tuple arguments
         # is constant-folded the surrounding tuple_get calls can be removed.
-        inlined = CollapseTuple.apply(inlined, offset_provider_type=offset_provider_type)  # type: ignore[assignment]  # always an itir.Program
+        inlined = CollapseTuple.apply(
+            inlined,
+            flags=~CollapseTuple.Flag.PROPAGATE_TO_IF_ON_TUPLES,
+            uids=collapse_tuple_uids,
+            offset_provider_type=offset_provider_type,
+        )  # type: ignore[assignment]  # always an itir.Program
         inlined = InlineScalar.apply(inlined, offset_provider_type=offset_provider_type)
 
         # This pass is required to run after CollapseTuple as otherwise we can not inline
@@ -126,7 +137,10 @@ def apply_common_transforms(
     # only run the unconditional version here instead of in the loop above.
     if unconditionally_collapse_tuples:
         ir = CollapseTuple.apply(
-            ir, ignore_tuple_size=True, offset_provider_type=offset_provider_type
+            ir,
+            ignore_tuple_size=True,
+            uids=collapse_tuple_uids,
+            offset_provider_type=offset_provider_type,
         )  # type: ignore[assignment]  # always an itir.Program
 
     ir = NormalizeShifts().visit(ir)
@@ -164,7 +178,9 @@ def apply_fieldview_transforms(
     ir = inline_fundefs.prune_unreferenced_fundefs(ir)
     ir = InlineLambdas.apply(ir, opcount_preserving=True, force_inline_lambda_args=True)
     ir = CollapseTuple.apply(
-        ir, offset_provider_type=common.offset_provider_to_type(offset_provider)
+        ir,
+        flags=~CollapseTuple.Flag.PROPAGATE_TO_IF_ON_TUPLES,
+        offset_provider_type=common.offset_provider_to_type(offset_provider),
     )  # type: ignore[assignment] # type is still `itir.Program`
     ir = infer_domain.infer_program(ir, offset_provider=offset_provider)
     return ir