Finer-grained recompilation checking for exports

Ticket: #25881 (closed)

This MR refines the recompilation checking logic, to avoid recompiling modules with an explicit import list when the modules they import start exporting new items. More specifically, when:

1. module N imports module M,
2. M is changed, but in a way that:
   a. preserves the exports that N imports
   b. does not introduce anything that forces recompilation downstream, such as orphan instances

then we no longer require recompilation of N.

Note that there is more to (2a) as initially meets the eye:

• if N includes a whole module or import hiding import of M, then we require that the export list of M does not change,

• if N only includes explicit imports, we check that the imported items don't change, e.g.

  • if we have import M(T(K, f), g), we must check that N continues to export all these identifiers, with the same Avail structure (i.e. we should error if N stops bundling K or f with T)
  • if we have import M(T(..)), we must check that the children of T have not changed

See Note [When to recompile when export lists change?] in GHC.Iface.Recomp.

This is all tested in the new tests RecompExports{1,2,3,4,5}

mentioned in issue #25881 (closed)

Looks very good. I briefly skimmed. Could we also have a Note to describe this scheme?

Could we have a test where there is an explicit import list, but the invisible hash changes?

added 1 commit

• cc426c11 - Finer-grained recompilation checking for exports

Compare with previous version

added 1 commit

• f29ea1f2 - Finer-grained recompilation checking for exports

Compare with previous version

changed the description

I have:

• Updated the nomenclature to focus on "orphan/deps" vs "exported avails". This is because the previous "visible"/"invisible" naming scheme was misleading, as we don't deal with non-orphan class instances here.
• Extended the logic to avoid recompilation in cases such as import M(T(..)) and the set of children exported by T does not change; this is tested in ExportRecomp1 (happy path) and ExportRecomp3 (we do need to recompile, because the children change)
• Added the ExportRecomp5 test which checks that we do recompile in the case of orphans.
• Added Note [When to recompile when export lists change?] in GHC.Iface.Recomp which explains the new mechanism for recompilation checking.

added 1 commit

• d2c822cf - Finer-grained recompilation checking for exports

Compare with previous version

added 5 commits

• d2c822cf...d0fd9370 - 4 commits from branch ghc:master
• e213f378 - Finer-grained recompilation checking for exports

Compare with previous version

added 2 commits

• 39eba292 - Finer-grained recompilation checking for exports
• 4b915798 - Add second test for COMPLETE pragma recompilation

Compare with previous version

Latest changes:

• spring cleaning of the notes in GHC.Iface.Recomp
• added new RecompCompletePragma2 test which tests the tricky case explained in Note [Fingerprinting complete matches]

added 2 commits

• e5ab6946 - Finer-grained recompilation checking for exports
• 1efeb490 - Add second test for COMPLETE pragma recompilation

Compare with previous version

resolved all threads

added 1 commit

• e51900a0 - Finer-grained recompilation checking for exports

Compare with previous version

Looks great. Thanks Sam. I left a few small comments.

approved this merge request

resolved all threads