optimizer: refactor SROA pass

[aviatesk]

• use BitSet instead of IdSet{Int}
• reduce # of dynamic allocations
• separate some computations into individual functions

---

TTFP improvements:

before (df0080c)

❯ ./usr/bin/julia -e "@time using Plots; @time plot(rand(10,3));"
  4.588537 seconds (9.34 M allocations: 624.761 MiB, 6.73% gc time, 15.28% compilation time)
  2.875475 seconds (3.84 M allocations: 214.448 MiB, 5.90% gc time, 99.77% compilation time)

after (cef9efa)

❯ ./usr/bin/julia -e "@time using Plots; @time plot(rand(10,3));"
  4.543190 seconds (9.24 M allocations: 615.898 MiB, 6.70% gc time, 15.24% compilation time)
  2.823435 seconds (3.78 M allocations: 209.089 MiB, 5.43% gc time, 99.76% compilation time)

[aviatesk]

I just wonder: we may want to have some nice support for this sort of optimization?
Like: @delay_alloc defuses = IdDict{Int, Tuple{BitSet, SSADefUse}}()

[Keno]

BitSet can be dangerous on large functions with high sparsity

[codecov]

Codecov Report

Merging #43232 (df0080c) into master (7107ced) will decrease coverage by 0.10%.
The diff coverage is 93.33%.

❗ Current head df0080c differs from pull request most recent head 6c4e203. Consider uploading reports for the commit 6c4e203 to get more accurate results

@@            Coverage Diff             @@
##           master   #43232      +/-   ##
==========================================
- Coverage   89.62%   89.52%   -0.11%     
==========================================
  Files         343      343              
  Lines       80294    80301       +7     
==========================================
- Hits        71961    71886      -75     
- Misses       8333     8415      +82     

Impacted Files	Coverage Δ
base/compiler/ssair/ir.jl	79.86% <0.00%> (-0.55%)	⬇️
base/compiler/ssair/passes.jl	91.97% <94.02%> (-1.06%)	⬇️
base/compiler/ssair/show.jl	86.88% <100.00%> (ø)
base/compiler/ssair/slot2ssa.jl	87.72% <100.00%> (+0.02%)	⬆️
base/compiler/stmtinfo.jl	78.57% <0.00%> (-14.29%)	⬇️
stdlib/REPL/src/REPLCompletions.jl	88.76% <0.00%> (-2.72%)	⬇️
base/compiler/abstractinterpretation.jl	87.99% <0.00%> (-2.15%)	⬇️
base/compiler/inferencestate.jl	84.01% <0.00%> (-1.37%)	⬇️
base/compiler/ssair/queries.jl	92.20% <0.00%> (-1.30%)	⬇️
base/compiler/typeutils.jl	82.01% <0.00%> (-1.06%)	⬇️
... and 15 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2b1ece9...6c4e203. Read the comment docs.

[aviatesk]

BitSet can be dangerous on large functions with high sparsity

Aha, I see the point. SROA works on IR after inlining, which can be very large sometimes, and program counters analyzed by it are often very sparse.
I made them IdSet{Int} back, also added some note for this choice.