llvm-project

87069dfc - [MISched] Implement `UnanalyzableFrontier`-based DAG construction algorithm (#227375)

[MISched] Implement `UnanalyzableFrontier`-based DAG construction algorithm (#227375) Introduces a new algorithm for constructing the control dependencies in the schedule DAG. The new algorithm is gated behind a (temporary) `cl::opt` (`-enable-unanalyzable-store-sequencing`), currently off by default. With the option disabled, the change is an effective NFC: edge insertion order changes slightly, but the DAGs do not change materially. The difference between this algorithm and the existing algorithm is that, rather than maintaining all unanalyzable memory operations and repeatedly querying all of them, we instead maintain a frontier of loads and stores since a 'sequencing store'. A sequencing store is either (a) a store that writes to different set of base objects to the last seen sequencing store or (b) any store if there has been a load from another base object since the last sequencing store. We treat the sequencing stores carefully to ensure that all previously seen unanalyzable memory operations that do not belong to the frontier necessarily transitively succeed the current sequencing store. This makes it sound to sequence preceding memory operations against the sequencing store and only those other memory operations in the frontier at the cost of some spurious control dependencies. In large basic blocks this can dramatically reduce the number of AA queries we issue. The plan is to land this change first, and issue a follow-up to enable, giving us a chance to deal with any performance bugs on other backends. Once enabled, I will revisit the interaction with `-dag-maps-huge-region` as, at present, the `-dag-maps-huge-region` logic frequently fires needlessly in situations where the frontier already successfully dealt with the complexity.