diff options
| author | Mazdak Farrokhzad <twingoow@gmail.com> | 2019-05-14 22:00:13 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2019-05-14 22:00:13 +0200 |
| commit | b4c340e4bb246572029d235e46850faa9823d312 (patch) | |
| tree | 6d4a00e80c0285020fe513ec4c0f9ec1bb39a34f /src/test/ui/thinlto | |
| parent | 088c99410b8516a4fc639507ab6c27184875d005 (diff) | |
| parent | decd6d366018823a8b1116b346bc778eb010accd (diff) | |
| download | rust-b4c340e4bb246572029d235e46850faa9823d312.tar.gz rust-b4c340e4bb246572029d235e46850faa9823d312.zip | |
Rollup merge of #60444 - nikomatsakis:issue-60010-cycle-error-investigation, r=pnkfelix
forego caching for all participants in cycles, apart from root node This is a targeted fix for #60010, which uncovered a pretty bad failure of our caching strategy in the face of coinductive cycles. The problem is explained in the comment in the PR on the new field, `in_cycle`, but I'll reproduce it here: > Starts out as false -- if, during evaluation, we encounter a > cycle, then we will set this flag to true for all participants > in the cycle (apart from the "head" node). These participants > will then forego caching their results. This is not the most > efficient solution, but it addresses #60010. The problem we > are trying to prevent: > > - If you have `A: AutoTrait` requires `B: AutoTrait` and `C: NonAutoTrait` > - `B: AutoTrait` requires `A: AutoTrait` (coinductive cycle, ok) > - `C: NonAutoTrait` requires `A: AutoTrait` (non-coinductive cycle, not ok) > > you don't want to cache that `B: AutoTrait` or `A: AutoTrait` > is `EvaluatedToOk`; this is because they were only considered > ok on the premise that if `A: AutoTrait` held, but we indeed > encountered a problem (later on) with `A: AutoTrait. So we > currently set a flag on the stack node for `B: AutoTrait` (as > well as the second instance of `A: AutoTrait`) to supress > caching. > > This is a simple, targeted fix. The correct fix requires > deeper changes, but would permit more caching: we could > basically defer caching until we have fully evaluated the > tree, and then cache the entire tree at once. I'm not sure what the impact of this fix will be in terms of existing crates or performance: we were accepting incorrect code before, so there will perhaps be some regressions, and we are now caching less. As the comment above notes, we could do a lot better than this fix, but that would involve more invasive rewrites. I thought it best to start with something simple. r? @pnkfelix -- but let's do crater/perf run cc @arielb1
Diffstat (limited to 'src/test/ui/thinlto')
0 files changed, 0 insertions, 0 deletions
