Canonical type nodes, or, comptypes considered harmful

[Gabriel Dos Reis]

Douglas Gregor <doug.gregor@gmail.com> writes:

[...]

| What would it take to get canonical type nodes in GCC? Nathan has
| experimented with some pieces of it, and I have experimented with
| others. It will likely be a large effort, but the payoff will be
| worth it in the end. Is the GCC community interested in working on
| this, and how would we go about doing it?

As a data point: in fall 2004, Bjarne Stroustrup and myself did
compare GCC's type representation with the type representation in the
Pivot (which is based on what you call "canonical type nodes" -- we
called it "unified type nodes") and measured about 90% of saving of
duplicate nodes for types, when processing the canonical "hello world"
program.  A draft of the paper circulated at the last Redmond meeting. 
That was before GCC did a move to a more systematic partial unified
type nodes. 

So, I'm all for it.

-- Gaby