Happy and `haskell-src-exts` differ by shift/reduce conflicts

[tbidne]

Hello.

haskell-src-exts-1.23.1 fails to build with new happy-2.2.1:

unused rules: 3
The grammar has 12 shift/reduce conflicts.
This is different from the number given in the expect directive
Error: [Cabal-7125]
Failed to build haskell-src-exts-1.23.1 (which is required by hs-src-exts-happy-0.1.0.0). See the build log above for details.

My guess is that this difference comes from this new commit, whereas haskell-src-exts currently expects 13:

> %expect 13

I documented this on a repo here, where the old index-state chooses older happy-2.2 (ok), but the newer index-state chooses happy-2.2.1 (fails). Incidentally, it would be nice if there was a way to force the older happy even with the newer index state e.g.

constraints: happy <2.2.1

But this doesn't seem to work for Build-Tools (and, I guess, build-tool-depends).

EDIT: Looks like this works:

constraints: any.happy <2.2.1

I'm not sure what should be done here e.g. this seems like it is probably a bug on haskell-src-exts. But I figured I'd make an issue here, since this may affect happy users more generally.

Thanks!

[sgraf812]

Thanks for reporting this. Indeed, removal of the unused state due to #353 is an unpredicted breaking change on all %expect <n> statements. I deprecated 2.2.1 for now to reflect this.

Note that use of %expect is discouraged anyway. See https://gitlab.haskell.org/ghc/ghc/-/blob/master/compiler/GHC/Parser.y#L104 for how shift/reduce conflicts should be resolved using %shift pragmas on the problematic production.

But the breaking change on the ecosystem in 2.2.1 is unforced; I think we should just revert #353. Unless @andreasabel has a better idea?

[andreasabel]

First lesson from this regression (and action item): add a reproducer to the Happy testsuite!

We could look into whether the %expect semantics should be changed to upper bound rather than exact number of conflicts. The docs affirm the latter and tie it to the semantics that yacc has: https://haskell-happy.readthedocs.io/en/latest/syntax.html#expect-declarations
The new semantics would be an error only if the number of s/r conflicts exceeds the expected number, and a warning when the bound is not exact.

[sgraf812]

Good call. I opened #355 with the fix. Will merge within the next week and release 2.2.2.