inference: add missing limit-cycle poisoning

[vtjnash]

Seems to have been missing since the original implementation of IPO constant-prop.
It's impressive how long we got away with just poisoning the cache with Any
if there was a self-recursive call.

Fix #29923

[ararslan]

Is it possible to add a test for this?

[martinholters]

Can the type assertions in int.jl (#30032) be dropped then? (That would also act as a test, albeit a rather indirect one.)

[martinholters]

Can the type assertions in int.jl (#30032) be dropped then?

And the answer is ... almost. Those added in #30032 seem safe to drop, but

     function div(x::UInt128, y::UInt128)
-        return UInt128(div(BigInt(x), BigInt(y)))::UInt128
+        return UInt128(div(BigInt(x), BigInt(y)))
     end

still results in

julia> @inferred div(one(UInt128), one(UInt128))
ERROR: return type UInt128 does not match inferred return type Any
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] top-level scope at none:0

julia> @code_warntype div(one(UInt128), one(UInt128))
Body::UInt128
771 1 ─ %1 = invoke Base.BigInt(_2::UInt128)::BigInt                                                                            │ 
    │   %2 = invoke Base.BigInt(_3::UInt128)::BigInt                                                                            │ 
    │   %3 = Base.GMP.MPZ.tdiv_q::typeof(Base.GMP.MPZ.tdiv_q)                                                                   │╻ div
    │   %4 = invoke %3(%1::BigInt, %2::BigInt)::BigInt                                                                          ││
    │   %5 = invoke Base.UInt128(%4::BigInt)::UInt128                                                                           │ 
    └──      return %5  

So that's an instance which, without beneficially primed caches, is inferred as Any (and that result is cached and leads to the @inferred failing), but after some of the callees have been inferred to tighter results (and cached), inference invoked with @code_warntype also gives a tighter result?

[vtjnash]

Is it possible to add a test for this?

Yes, it was just tricky to figure out. To answer my original question, we've gotten away with this for a long time because there's many fast-path efforts to ensure we don't get here in the majority of cases. In particular, the const-prop code needs to run into call-stack recursion that did not get observed without const-prop (e.g. would not be in the cache yet). The result of that recursion then gets set to Any, but it needs to be the case that without const-prop, we would have been able to figure out the result type anyways (even though without const-prop, we didn't even know this method would get called—but inside const-prop this method could affect the final answer!)

Here's the min repro of my above prose rendered as code:

_false = false
f() = _false ? g() : 3
g() = (h(:f); 4)
h(f) = getfield(Main, f)()

julia> Base.return_types(g, ())
1-element Array{Any,1}:
 Int64

julia> Base.return_types(f, ())
1-element Array{Any,1}:
 Any

julia> Base.return_types(f, ())
1-element Array{Any,1}:
 Int64

[StefanKarpinski]

@vtjnash: Thank you for finding and fixing this and finding such a tricky repro/test case. ❤️

[KristofferC]

Does not backport cleanly. Please push backport to "backport-1.0.3" or make a rebased branch on top of that and I will backport the commits. @vtjnash

[vtjnash]

I think you're likely missing a few commits (tested myself these should work):

* ff87181ebc N* - inference: set limited flag on entire cycle (10 seconds ago) <Jameson Nash>
* 6aa0a64d35 N* - interpreter: fix pref regression from #29795 (#29873) (5 minutes ago) <Jameson Nash>
* 266d6cde79 N* - partially address #29293 (5 minutes ago) <Jameson Nash>
* f375eb16e7 N* - inference: improve method caching (5 minutes ago) <Jameson Nash>

[KristofferC]

I don't have those commits:

➜  julia git:(backport-1.0.3) ✗ git cherry-pick -x ff87181ebc
fatal: bad revision 'ff87181ebc'

[vtjnash]

oh, sorry, those hashes are my local branch. The commits are aaa9bd9, e1fbd0a, 92ab600, 439f798

[KristofferC]

Hmm, doesn't seem to work cleanly for me:

➜  julia git:(backport-1.0.3) ✗ git cherry-pick -x aaa9bd977815dc910be
[backport-1.0.3 70d5acdb74] inference: set limited flag on entire cycle
 Author: Jameson Nash <[email protected]>
 Date: Wed Oct 24 17:47:22 2018 -0400
 1 file changed, 10 insertions(+), 3 deletions(-)
➜  julia git:(backport-1.0.3) ✗ git cherry-pick -x e1fbd0a
error: could not apply e1fbd0a217... interpreter: fix pref regression from #29795 (#29873)
hint: after resolving the conflicts, mark the corrected paths

[KristofferC]

Removing backport because it doesnt seem like it will happen.