Try to speed up LateLowerGCFrame::ComputeLiveness

[chriselrod]

Co-authored-by: Jameson Nash[email protected]
Co-authored-by: Oscar Smith <oscardssmith.gmail.com>

Reduces allocations (malloc) in late lower gc frame.

[oscardssmith]

I can confirm this is a significant improvement. I'll see if I can come up with a simple example that exhibits this.

[oscardssmith]

I haven't been able to get a great mwe for this, but the following shows a 2x improvement (on the pass runtime, not the whole runtime).

n = 1500
f(var1, var2, var3) = (var1>1.1) ? (var2+var3) : (var2*var1)
vars = [Symbol("var$i") for i in 1:n]
s = :(struct x
    $([:($(vars[i])::Any) for i in 1:n]...)
    function x(;d...)
        var1 = get(d, :var1, rand(1:4))
        var2 = get(d, :var2, -rand())
        $(Expr[:(
            $(vars[i]) = get(d, $(QuoteNode(vars[i]))) do;
                f($(vars[i-1]), $(vars[i-2]), $(vars[rand(1:div(i,2))]))
        end) for i in 3:n]...)
        new($(vars...))
    end
end)
eval(s)
x(var3=4)

On master we get

oscardssmith:julia$ JULIA_LLVM_ARGS=--time-passes julia test.jl tee 2>&1 |head -n20
===-------------------------------------------------------------------------===
                      ... Pass execution timing report ...
===-------------------------------------------------------------------------===
  Total Execution Time: 27.8093 seconds (27.3734 wall clock)
   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  14.5054 ( 57.6%)   0.4789 ( 18.3%)  14.9843 ( 53.9%)  14.9414 ( 54.6%)  X86 DAG->DAG Instruction Selection
   1.2221 (  4.9%)   0.4556 ( 17.4%)   1.6777 (  6.0%)   1.6346 (  6.0%)  X86 Assembly Printer
   1.4528 (  5.8%)   0.0712 (  2.7%)   1.5240 (  5.5%)   1.5226 (  5.6%)  Late Lower GCFrame Pass

and on the PR, we get

oscardssmith:julia$ JULIA_LLVM_ARGS=--time-passes julia test.jl tee 2>&1 |head -n20
===-------------------------------------------------------------------------===
                      ... Pass execution timing report ...
===-------------------------------------------------------------------------===
  Total Execution Time: 14.5055 seconds (14.0426 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
  10.1672 ( 73.3%)   0.1061 ( 16.8%)  10.2733 ( 70.8%)  10.2115 ( 72.7%)  X86 DAG->DAG Instruction Selection
   2.2113 ( 15.9%)   0.0085 (  1.3%)   2.2198 ( 15.3%)   2.2188 ( 15.8%)  SLP Vectorizer
   0.3562 (  2.6%)   0.0467 (  7.4%)   0.4029 (  2.8%)   0.3893 (  2.8%)  Greedy Register Allocator
   0.2686 (  1.9%)   0.0202 (  3.2%)   0.2889 (  2.0%)   0.2839 (  2.0%)  Late Lower GCFrame Pass

There's definitely more to improve here, but this does appear to be a noticeable and welcome speedup.

[oscardssmith]

Wait, I just looked at this again and it makes absolutely no sense. The late-lower GCFrame pass did get a bunch faster, but the total execution time also almost halved which makes absolutely no sense. @vtjnash any ideas?