[PATCH] Re: Checking out PR4483 (constant overflow on PPC, problems compiling Linux kernel)

Richard Kenner kenner@vlsi1.ultra.nyu.edu
Mon Nov 19 23:36:00 GMT 2001


    This looks much more correct, but I do not have authority to approve
    it.  Hopefully someone with experience in this part of the compiler
    could review it.  Maybe even Richard Kenner as he has been interested
    in this type work.

I agree it's "more correct", but I'm still not totally convinced this still
isn't covering up some other problem.  I haven't followed this thread enough
to know for sure.

As to the larger problem, I don't that I'd want to go to a framework
where CONST_INT has a mode.  If you do that, then it's tricky to
interpret INTVAL correctly.

I do understand how we've come to need to interpret a CONST_INT as
being in some mode, but I really like the original idea of CONST_INT
as simply being a representation of a integer in the mathematical
sene.  So my own preference is to look more carefully at the
violations of that principle and see if we can resolve them some other
way.  As I recall, a large part of the issue has to do with CONST_INT vs.
CONST_DOUBLE when the target and host have different word sizes.  I'd
be more comfortable with trying to resolve that issue by merging those two
RTL nodes in some way.



More information about the Gcc-patches mailing list