Fixing a new regresssion with 8.4 wrt to βι-normalization of created hyps.

Formerly, mk_refgoals in logic.ml was applying full βι on new Meta-based goals.
We simulate part of this βι-normalization in pose_all_metas_as_evars.

I suspect that we don't βι-normalize goals more than in 8.4 by doing
that, since all Metas would have eventually gone to mk_refgoals, but
difficult to know for sure as there were probably metas turned to
evars (and hence a priori not βι-normalized) even when logic.ml was
used more pervasively.

However, βι-normalizing is a priori a better heuristic than no
βι-normalizing, independently of what it was in 8.4 and before (even
if, ideally, I would personally lean towards preferring a
"chirurgical" substitution with reduction only at the place of
substitution).

1 files changed, 4 insertions, 0 deletions

diff --git a/pretyping/unification.ml b/pretyping/unification.ml
index f090921e5..61160147a 100644
--- a/pretyping/unification.ml
+++ b/pretyping/unification.ml
@@ -194,6 +194,10 @@ let pose_all_metas_as_evars env evd t =
         let {rebus=ty;freemetas=mvs} = Evd.meta_ftype evd mv in
         let ty = EConstr.of_constr ty in
         let ty = if Evd.Metaset.is_empty mvs then ty else aux ty in
+        let ty =
+          if Flags.version_strictly_greater Flags.V8_6 || Flags.version_less_or_equal Flags.VOld
+          then nf_betaiota evd ty (* How it was in Coq <= 8.4 (but done in logic.ml at this time) *)
+          else ty (* some beta-iota-normalization "regression" in 8.5 and 8.6 *) in
         let src = Evd.evar_source_of_meta mv !evdref in
         let ev = Evarutil.e_new_evar env evdref ~src ty in
         evdref := meta_assign mv (EConstr.Unsafe.to_constr ev,(Conv,TypeNotProcessed)) !evdref;