Loss of inlining after strictness analysis

While investigating something else I saw this in the final Tidy Core after optimisation

join {
  wild_X5jU [Dmd=<S,1*U>]
    :: (# State# RealWorld, Array Int HValue #)
  [LclId[JoinId(0)],
   Unf=Unf{Src=InlineStable, TopLvl=False, Value=False,
           ConLike=False, WorkFree=False,
           Expandable=False,
           Guidance=ALWAYS_IF(arity=0,unsat_ok=True,boring_ok=False)
           Tmpl= jump $j1_s5RP w_s5Qs}]
  wild_X5jU = jump $j1_s5RP ww_s5Qw } in
jump wild_X5jU;

That's ridiculous! Why wasn't it inlined at its (only!) use site?

It turns out that

• The original InlineStable came from when wild was lambda-bound.

  See Note [Case binders and join points] in Simplify.hs

• But then worker-wrapper let-binds that previously-lambda-boudn arg

  in WwLib.mkWWstr_one. But it does not remove the unfolding.

  That's bad; in contrast, when we beta-reduce we are careful to

  remove the unfolding; see Note [Zap unfolding when beta-reducing]

  in Simplify.hs.

• We don't preInlineUnconditionally things with stable unfoldings; see Note [Stable unfoldings and preInlineUnconditionally] in SimplUtils

Net result: it doesn't get inlined at all!

Easy fix: zap the unfolding in mkWWstr_one as well.

I'm unhappy with the whole business of passing both boxed and unboxed versions, described in Note [Case binders and join points]. It smells wrong, and this ticket is another canary.

Trac metadata

Trac field	Value
Version	8.0.1
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture