id	summary	reporter	owner	description	type	status	priority	component	version	resolution	keywords	cc
139	Unification with less cloning	a3moss		"I have selectively re-introduced deeper clones in the new AST port to handle mutation of forall-parameterized types in a way that keeps the `TypeInstType` instances pointing to the correct `TypeDecl`. This was done as an attempt to get the new AST working without making deeper changes to the compiler architecture, but Thierry's profiling results suggest these deeper clones are actually a performance hot spot.

The core reason for these clones is to support distinct type variables //per function call instance//. Consider the following:
{{{
forall(otype T) T * foo( T );

foo( foo( 42 ) );
}}}
The inner `foo` has `T => int`, the outer foo has `T => int *` -- since this is legal Cforall, both `T` variables need to exist within the same type environment and be able to be disambiguated.

The current solution is to clone the type of `foo` and rename the type variables, which is effective, but has undesirable memory-management implications. An alternate approach which avoids this cloning would be to give each `ApplicationExpr` a unique identifier (the pointer may work, but would introduce memory-management complications), and store type variables in an environment or unification calculation as the pair `(<variable identifier>, <application identifier>)`. This change should be investigated if cloning types continues to be a performance bottleneck."	task	new	minor	cfa-cc	1.0