search for: d21464

...n the transitive invalidation walking happens. Evidently > while a transformation pass is running, things might temporarily be > stale; what are the "checkpoints" where the mapping is guaranteed to > be valid? At the start of each transformation pass? At least > Chandler's D21464 does not stick to this because the IRUnit's > (SCC's) are only updated at the end of running potentially many > function transformation passes. I.e. all but the first function > transformation pass might observe stale IRUnit's (SCC's). > One other thing to note is that...

...; while a transformation pass is running, things might temporarily > > > be > > > stale; what are the "checkpoints" where the mapping is guaranteed > > > to > > > be valid? At the start of each transformation pass? At least > > > Chandler's D21464 does not stick to this because the IRUnit's > > > (SCC's) are only updated at the end of running potentially many > > > function transformation passes. I.e. all but the first function > > > transformation pass might observe stale IRUnit's (SCC's). > &gt...

...; while a transformation pass is running, things might temporarily > > > be > > > stale; what are the "checkpoints" where the mapping is guaranteed > > > to > > > be valid? At the start of each transformation pass? At least > > > Chandler's D21464 does not stick to this because the IRUnit's > > > (SCC's) are only updated at the end of running potentially many > > > function transformation passes. I.e. all but the first function > > > transformation pass might observe stale IRUnit's (SCC's). > &gt...

...ransitive invalidation walking happens. Evidently while a >> transformation pass is running, things might temporarily be stale; what are >> the "checkpoints" where the mapping is guaranteed to be valid? At the start >> of each transformation pass? At least Chandler's D21464 does not stick to >> this because the IRUnit's (SCC's) are only updated at the end of running >> potentially many function transformation passes. I.e. all but the first >> function transformation pass might observe stale IRUnit's (SCC's). >> >> One othe...

...validation walking happens. Evidently while a >>> transformation pass is running, things might temporarily be stale; what are >>> the "checkpoints" where the mapping is guaranteed to be valid? At the start >>> of each transformation pass? At least Chandler's D21464 does not stick to >>> this because the IRUnit's (SCC's) are only updated at the end of running >>> potentially many function transformation passes. I.e. all but the first >>> function transformation pass might observe stale IRUnit's (SCC's). >>> &...

...(module or (theoretically)CGSCC). So for any really nontrivial module analysis, we may end up having to change the interface of module passes to have some way to incrementally recompute themselves as function passes mutate the IR? In some sense, Chandler's CGSCC patch https://reviews.llvm.org/D21464 is trying to do this for a specific module analysis (lazy call graph) and even for just that one (and lots of special handling in the adaptors etc.) it still only gets incrementally updated it after a potentially large set of function passes have run. And it's already quite tricky. Broadly spe...

...tically)CGSCC). So for any really nontrivial module analysis, we may > end up having to change the interface of module passes to have some way to > incrementally recompute themselves as function passes mutate the IR? > > In some sense, Chandler's CGSCC patch https://reviews.llvm.org/D21464 is > trying to do this for a specific module analysis (lazy call graph) and even > for just that one (and lots of special handling in the adaptors etc.) it > still only gets incrementally updated it after a potentially large set of > function passes have run. And it's already quite...

On Wed, Jul 13, 2016 at 1:48 AM, Sean Silva <chisophugis at gmail.com> wrote: > > > On Wed, Jul 13, 2016 at 12:34 AM, Chandler Carruth <chandlerc at gmail.com> > wrote: > >> On Wed, Jul 13, 2016 at 12:25 AM Sean Silva <chisophugis at gmail.com> >> wrote: >> >>> On Tue, Jul 12, 2016 at 11:39 PM, Chandler Carruth <chandlerc at

...So for any really nontrivial module analysis, we may >> end up having to change the interface of module passes to have some way to >> incrementally recompute themselves as function passes mutate the IR? >> >> In some sense, Chandler's CGSCC patch https://reviews.llvm.org/D21464 is >> trying to do this for a specific module analysis (lazy call graph) and even >> for just that one (and lots of special handling in the adaptors etc.) it >> still only gets incrementally updated it after a potentially large set of >> function passes have run. And it'...

The more closely I look at this, the more it seems like there may be a useful incremental step in the transition to the new PM: use the new PM analysis machinery in the old PM. If this is possible, it will simplify the old PM and (hopefully) allow an incremental transition to the new PM instead of a flag day transition for the switch. I.e., AFAICT, the new PM transition is essentially about 2