Re: RFC: midlevel RTL

[law at redhat dot com]

In message <20020121175550.GJ25957@atrey.karlin.mff.cuni.cz>, Jan Hubicka write
s:
 > > On Sun, Jan 20, 2002 at 06:24:13PM +0100, Jan Hubicka wrote:
 > > > 	* gensupport.c (init_md_reader): Implicitly inlucde midrtl.md file.
 > > 
 > > This I think is a mistake.  I think mid-level rtl and target rtl
 > > should have completely separate recognizer functions.
 > 
 > As I've described, I want to do that later.  I can't simply change whole
 > code generation process to not include any lowlevel RTL instruction in one
 > day, as there is way too many expanders called from everwhere - machine
 > specific builtins, call convention, string operations, etc.
IMHO the best way to deal with this today, right now, is to have your
midrtl.md include a few selected backend patterns.  Lowering of those
patterns is trivial since they should be 100% identical to the patterns
in the target md file.

Doing that allows you to concentrate on the bits necessary to build two
independent recognizers and the easy lowering bits like logicals.

Over time you remove those selected backend patterns from midrtl.md until
none remain.

I think this accomplishes the same goal you were trying to achieve, but
allows some initial concentration on the infrastructure for lowering to
happen now.  IMHO it's critical that we get the right lowering infrastructure
in place early.

jeff