This is the mail archive of the
mailing list for the GCC project.
Re: [tree-ssa] CCP fixes
On Tue, Feb 18, 2003 at 05:55:36PM -0700, email@example.com wrote:
> In fact, you have to allow for arbitrarily complex trees since this
> stuff is usually peppered with a bunch of NOP_EXPRs and CONVERT_EXPRS.
> >In rtl we have (const (plus (symbol_ref) (const_int))) to
> >record this, but nothing similar in trees.
> Yup. I'm just not sure this concept is the best thing to re-introduce
> into the gimple code.
If we don't, then the rtl optimizers will need to continue to
be smart enough to reconstruct the bits that we missed at the
tree level, due to link-time constants not being considered
Your comment about NOP and CONVERT exprs leads me to believe
that something *exactly* like a CONST_EXPR might be the right
thing. Anything that passes initializer_constant_valid_p
could be wrapped in this.