On 7/30/19 2:44 AM, Richard Biener wrote:
On Tue, Jul 30, 2019 at 10:35 AM Jakub Jelinek <ja...@redhat.com> wrote:
On Tue, Jul 30, 2019 at 10:21:15AM +0200, Richard Biener wrote:
Would you need to LTO stream & merge the bitmaps / maps somehow?
Yes. And if we do not throw unneeded warnings from the sets normally, LTO
streaming might be a good time to do that, so that we merge in only warnings
that will be tested during/post IPA.
(maybe "unmap" to sth else when streaming individual stmts) Not sure if
a bitmap is really necessary, we could simply have a tree/gimple * -> vec<>
mapping with records about emitted (or queued) diagnostics, like
OPT_, message pairs or so.
Right now we use it both for the case we've already emitted some diagnostics
and don't want to emit it again later, or for the case where we insert
something in the IL that a warning could be diagnosed on but we know we
don't want that. The latter is the case e.g. for when we add
TREE_NO_WARNING on the last value in a statement expression so that we don't
diagnose it as unused, or the case we are discussing here.
If we need queued diagnostics, sure, we could add it in, but I don't see a
need for that right now. vecs compared to bitmap might be larger and harder
to avoid duplicates. I guess we'd need to do some analysis though, and e.g.
if 99% of cases we need to disable just one warning and not multiple,
perhaps optimize for that case.
I was thinking we may want to use the same facility to record warnings we
want to not emit if we later figure a stmt was in an unreachable region. For
this we need to record the actual warning, not only the warning kind.
I was thinking of introducing a __builtin_warning() for this and
issuing it if it's not eliminated just before RTL expansion. That
would let programs like libstdc++ use it too.
The downside of these solutions is that warnings may be issued out
of order as code moves around. To have the issued in the order of
increasing source lines they would need to be queued.
Martin
