Re: [PATCH 2/4] allow poisoning input_location in ranges it should not be used

[Trevor Saunders <tbsaunde@tbsaunde.org>]

On Thu, Jul 01, 2021 at 11:40:55AM -0400, David Malcolm via Gcc-patches wrote:
> On Thu, 2021-07-01 at 14:53 +0200, Richard Biener wrote:
> > On Thu, Jul 1, 2021 at 12:16 PM Trevor Saunders <
> > tbsaunde@tbsaunde.org> wrote:
> > > 
> > > On Wed, Jun 30, 2021 at 11:13:23AM -0400, David Malcolm wrote:
> > > > On Wed, 2021-06-30 at 01:35 -0400, Trevor Saunders wrote:
> > > > > This makes it possible to assert if input_location is used
> > > > > during the
> > > > > lifetime
> > > > > of a scope.  This will allow us to find places that currently
> > > > > use it
> > > > > within a
> > > > > function and its callees, or prevent adding uses within the
> > > > > lifetime
> > > > > of a
> > > > > function after all existing uses are removed.
> > > > > 
> > > > > bootstrapped and regtested on x86_64-linux-gnu, ok?
> > > > > 
> > > > > Trev
> > > > 
> > > > [...snip...]
> > > > 
> > > > > diff --git a/gcc/diagnostic.c b/gcc/diagnostic.c
> > > > > index d58586f2526..3f68d1d79eb 100644
> > > > > --- a/gcc/diagnostic.c
> > > > > +++ b/gcc/diagnostic.c
> > > > > @@ -1835,7 +1835,7 @@ internal_error (const char *gmsgid, ...)
> > > > >    auto_diagnostic_group d;
> > > > >    va_list ap;
> > > > >    va_start (ap, gmsgid);
> > > > > -  rich_location richloc (line_table, input_location);
> > > > > +  rich_location richloc (line_table, UNKNOWN_LOCATION);
> > > > >    diagnostic_impl (&richloc, NULL, -1, gmsgid, &ap, DK_ICE);
> > > > >    va_end (ap);
> > > > > 
> > > > 
> > > > I actually make use of this in the analyzer: the analyzer sets
> > > > input_location to stmt->location when analyzing a given stmt -
> > > > that
> > > > way, if the analyzer ICEs, the ICE is shown at the code construct
> > > > that
> > > > crashed the analyzer.
> > > > 
> > > > This behavior is useful to me, and would be lost with the
> > > > proposed
> > > > patch.
> > > 
> > > I made this change because otherwise if the compiler ICE's while
> > > access
> > > to input_location is blocked we end up infinitely recursing
> > > complaining
> > > we can't access it while trying to say where the last error was.  I
> > > was
> > > nervous about the change before, and now I agree we need something
> > > else.
> > > 
> > > > Is there a better way of doing what I'm doing?
> > > > 
> > > > Is the long-term goal of the patch kit to reduce our reliance on
> > > > global
> > > > variables?  Are we ultimately still going to need a variable for
> > > > "where
> > > > to show the ICE if gcc crashes"?  (perhaps stashing it in the
> > > > diagnostic_context???)
> > > 
> > > Yes, the goal is ultimately removal of global state, however I'm
> > > not
> > > really ure what the better approach to your problem is, after all
> > > even
> > > moving it to the diagnostic context is sort of a global state, and
> > > sort
> > > of dupplicates input_location.  That said it is somewhat more
> > > constrained, so if it removes usage of input_location perhaps its
> > > worthwhile?
> > 
> > Reduction of global state is of course good - but in particular
> > input_location
> > should be something only used during parsing because it's a quite
> > broken concept otherwise.  And fiddling with it tends to be quite
> > fragile...
> > for example see g:7d6f7e92c3b737736a2d8ff97a71af9f230c2f88
> > for the "fun" you can have with "stale" values in input_location ...
> 
> Yeah.  Another example, from the analyzer, is
> g:2fbea4190e76a59c4880727cf84706fe083c00ae (PR 93349)

So, one rather useful this patch allows doing, even if you apply it
locally, is that you can block off access to input_location in the same
scope you currently save and restore it, and then once you've fixed up
all the asserts you find with a bootstrap and test cycle, you can be
somewhat confident nothing uses it and the save and restore dance can
go.  I may have a look at doing that with some of these.

> 
> > IMHO users should have their own "copy", for example the gimplifier
> > instead of mucking with and using input_location could use a
> > similar state in its gimplify_ctx.
> 
> Some ideas (not necessarily good ones):
> 
> (a) the diagnostic_context could have an ice_location field, and use
> that in internal_error (and maybe an RAII class for setting/clearing
> it).

If this is useful to people as it seems to be, this seems sensible, my
one concern is when exactly should ice_location be set? other than "it
should match input_location".  There's still the global state of the
diagnostic context, but at least its more understandable and isolated.
On the other hand it offers the opertunity to have a stack of
ice_locations, if you walk up the stack looking at all the RAII objects,
perhaps that would be useful, especially with things like template
instantiation?

> (b) move input_location to diagnostic_context, and add:
>        #define input_location (global_dc->x_input_location)
> or:
>        #define input_location (global_dc->x_default_location)
> which add an indirection everywhere.  I don't love these ideas, in that
> we already overuse the preprocessor IMHO.

There's that, also it doesn't seem to really reduce the complexity? it
just merges too globals into one block of memory?

Thanks

Trev

> 
> Trevor: BTW, if you're looking for global state to eliminate, it might
> be nice to move the globals in input.c for caching source lines
> (fcache_tab etc) into a new source_cache class, and have the
> diagnostic_context own it via a new "source_cache *" field.
> 
> Dave
>