Subject: Re: rep questions
From: Keunwoo Lee (klee@cs.washington.edu)
Date: Mon Oct 30 2000 - 17:08:55 PST
OK. 1 question:
We discussed the fact that, as optimizations operate on the IR, the AST
may eventually become invalid relative to the current state of the code.
Does this mean that we will have to write two cases for most lowerings,
one that works with the AST pseudo-DFG and one that works with a real DFG?
The amount of info provided by the two is clearly not identical. For
example, I am working on lowering UnaryConvertOps. If I want to know
where the converted value goes under the AST, I have to look at the parent
AST node of the current node, etc. However, with a DFG I can just ask for
the successor.
~k
On Mon, 30 Oct 2000, Craig Chambers wrote:
> Yes. The idea is that, when we're not doing optimizations, we don't
> build the DFG (to make compilation as fast as possible). When we care
> about dataflow links, we consult the AST as a coarse approximation of
> the DFG. You can ask a WindIR for its dfg_kind, which is either
> WindDataflowKind or WindASTKind, depending on which slice is
> representing the dataflow links. You then access your dataflow
> arguments using node.arg1(kind), where kind was computed from
> ir.dfg_kind. See wil-codegen-c.cecil for lots of examples of
> accessing dataflow predecessors indexed by analysis.kind (which is a
> cached copy of ir.dfg_kind).
>
> ast-helpers, by the way, defines a bunch of functions to give
> convenient names to particular dataflow predecessors, including the
> functions indexed by the dataflow slice to access. (These functions
> should probably be moved to ir-node, since they're generally useful,
> not just when manipulating ASTs. The constructors, on the other hand,
> should stay, since they're explicitly building up AST links.)
>
> -- Craig
>
>
> Keunwoo Lee wrote:
> >
> > Craig: (cc'd to list for benefit of all)
> >
> > Is there any particular reason method lower(ir:WindIR) in lower.cecil
> > takes either a CFG or a CDFG? It seems the lowering pass needs data flow
> > information. Certainly, DFG info would be useful for the lower_ops I'm
> > working on now.
> >
> > ~k
>
This archive was generated by hypermail 2b25 : Mon Oct 30 2000 - 17:09:01 PST