[tree-ssa] Death to WFL nodes

Jan Hubicka jh@suse.cz
Wed Jan 15 21:00:00 GMT 2003


> Frank Ch. Eigler wrote:-
> 
> > I disagree.  Consider the desire to provide a precise location string
> > to a warning that relates to the "+" operator (as distinct from the
> > assignment).  If the encoding of an original source location was
> > efficient enough, it would be great to track and propagate such
> > precise information throughout.
> 
> Agreed.  I have caret diagnostics working in my embryonic C parser,
> in such a way that it's easy to point at any token and give a diagnostic
> referring to it, and they're cute.  The quality of diagnostics is an order
> of magnitude higher as a result.  However, if things are done "properly"
> (IMO), then the front end issues 99% of all diagnostics ever issued by
> the compiler and the middle end doesn't need much location info (just
> decls, statements and some parts of expressions, mostly for debug info).
> 
> The middle end should only need to issue diagnostics about things only
> it can discover, such as control flow implying dead code or
Hopefully with AST these will be recognized at higher level than we are
donig right now.  There is still need for RTL to have prety exact
information for debug output.

Dwarf2 debug output has notion of statements and substatements as long
as I can remember.  So backend needs to know it at least in this
granuality.  I still plan to redesign the way debug output is done and
move the information from notes into the insns themselves (so each insn
will have attached statement that has produced her).

Honza
> uninitialized variables or whatever.  There are very few such things
> to warn about.  I suspect the middle end would not need very little
> location info about expressions, maybe just the start of each full
> expression.
> 
> At present, the front end and middle end are too intertwined to do this
> properly.  Further, a bison parser doesn't allow you to get the correct
> location for most things, I think, since adding the necessary actions
> would introduce a gazillion shift / reduce conflicts.  You might be able
> to do it with a lot of ugly hacks outside the actions.
> 
> Neil.



More information about the Gcc-patches mailing list