optimizer: SROA mutable(immutable(...)) case correctly

[aviatesk]

Our SROA should be able to handle mutable(immutable(...)) case even
without "proper" alias analysis, since we eliminate immutable first
and then process mutable within our iterative approach.

It turns out it hasn't worked because after the immutable handling,
we keep the reference count before DCE, but don't update dead reference
count. This commit fixes it and now we should be able to eliminate more
allocations, e.g.:

julia> simple_sroa(s) = broadcast(identity, Ref(s))
simple_sroa (generic function with 1 method)

julia> s = "julia";

julia> simple_sroa(s); # compile

julia> @allocated(simple_sroa(s))
0

---

@nanosoldier runbenchmarks("broadcast" || "sparse" || "array" || "union", vs=":master")

(... and now I lose yet another motivative example for EscapeAnalysis.jl-based SROA :P)

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[chriselrod]

It turns out it hasn't worked because after the immutable handling,
we keep the reference count before DCE, but don't update dead reference
count. This commit fixes it and now we should be able to eliminate more
allocations, e.g.:

Was that a regression?

julia> simple_sroa(s) = broadcast(identity, Ref(s))
simple_sroa (generic function with 1 method)

julia> simple_sroa("compile"); @allocated(simple_sroa("julia"))
0

julia> versioninfo()
Julia Version 1.6.4-pre.2
Commit 1a92e56a4d* (2021-10-09 19:47 UTC)

[aviatesk]

Ah, it seems like the entire chunk of @allocated(simple_sroa("julia)) is fully compiled (and so it's not a very good example):

julia> code_typed() do
           simple_sroa("julia")
       end
1-element Vector{Any}:
 CodeInfo(
1 ─     goto #3 if not true
2 ─     nothing::Nothing
3 ┄     goto #4
4 ─     goto #5
5 ─     goto #6
6 ─     goto #7
7 ─     goto #8
8 ─     return "julia"
) => String

Allocation happens if we use the compiled code of simple_sroa:

julia> s = "julia"
"julia"

julia> @allocated(simple_sroa(s))
16

but with this PR, no allocation should happen.

[Keno]

Looks sensible to me though, of course this entire bit of code is a bit messy and we may want better abstractions in the future.