add an inductive range check elimination pass

[pchintalapudi]

As stated in #42521, using @inbounds to denote areas where bounds checks are necessary is dangerous. LLVM provides an InductiveRangeCheckElimination pass which can extract some bounds checks which are known to be safe.

Godbolt Links:

@noinline function f!(ys, xs)
                       for i in eachindex(ys, xs)
                         x = xs[i]
                         if -0.5 < x < 0.5
                           ys[i] = 2*x
                         end
                      end
              end

Original: https://godbolt.org/z/YY3WrYThh
With @inbounds: https://godbolt.org/z/cEW9GTsT3
Original+IRCE: https://godbolt.org/z/j1YndbzrE

The pass makes the control flow graph a little messier than @inbounds does, but does allow vectorization that is not allowable when bounds checks are present.

[vchuravy]

Nice! Could you time the passes during sysimage compilation? So that we can see how much additional time this takes.

[vchuravy]

@nanosoldier runtests(ALL, vs = ":master")

[vchuravy]

Can you post some benchmark numbers from your test-cases?

[pchintalapudi]

Nice! Could you time the passes during sysimage compilation? So that we can see how much additional time this takes.

Core.Compiler: 45.6049 seconds
Base ───────────── 23.928286 seconds
ArgTools ───────── 4.344550 seconds
Artifacts ──────── 0.106307 seconds
Base64 ─────────── 0.106040 seconds
CRC32c ─────────── 0.007023 seconds
FileWatching ───── 0.098718 seconds
Libdl ──────────── 0.001533 seconds
Logging ────────── 0.036474 seconds
Mmap ───────────── 0.083878 seconds
NetworkOptions ─── 0.094496 seconds
SHA ────────────── 0.187798 seconds
Serialization ──── 0.288951 seconds
Sockets ────────── 0.335865 seconds
Unicode ────────── 0.007169 seconds
DelimitedFiles ─── 0.103579 seconds
LinearAlgebra ──── 7.549069 seconds
Markdown ───────── 0.824861 seconds
Printf ─────────── 0.267628 seconds
Random ─────────── 1.139277 seconds
Tar ────────────── 0.275640 seconds
Dates ──────────── 1.595620 seconds
Distributed ────── 0.787926 seconds
Future ─────────── 0.004703 seconds
InteractiveUtils ─ 0.526222 seconds
LibGit2 ────────── 1.342156 seconds
Profile ────────── 0.271113 seconds
SparseArrays ───── 2.652343 seconds
UUIDs ──────────── 0.015049 seconds
REPL ───────────── 3.619972 seconds
SharedArrays ───── 0.520581 seconds
Statistics ─────── 0.193729 seconds
SuiteSparse ────── 1.590027 seconds
TOML ───────────── 0.068915 seconds
Test ───────────── 0.296838 seconds
LibCURL ────────── 0.415338 seconds
Downloads ──────── 0.385399 seconds
Pkg ────────────── 4.512018 seconds
LazyArtifacts ──── 0.001587 seconds
Stdlibs total ──── 34.670334 seconds
Sysimage built. Summary:
Total ─────── 58.600201 seconds
Base: ─────── 23.928286 seconds 40.8331%
Stdlibs: ──── 34.670334 seconds 59.1642%

Generating REPL precompile statements... 31/31
Executing precompile statements... 1221/1253
Precompilation complete. Summary:
Total ─────── 100.895169 seconds
Generation ── 76.441509 seconds 75.7633%
Execution ─── 24.453660 seconds 24.2367%

[pchintalapudi]

Can you post some benchmark numbers from your test-cases?

Test Case 1:

@noinline function f!(ys, xs)
                       for i in eachindex(ys, xs)
                         x = xs[i]
                         if -0.5 < x < 0.5
                           ys[i] = 2*x
                         end
                      end
              end

Original:

julia> @benchmark f!(zeros(Float64, 1000), rand(Float64, 1000))
BenchmarkTools.Trial: 10000 samples with 3 evaluations.
 Range (min … max):  7.703 μs … 274.071 μs  ┊ GC (min … max): 0.00% … 95.81%
 Time  (median):     8.200 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   8.682 μs ±   6.729 μs  ┊ GC (mean ± σ):  2.09% ±  2.66%

   ▁▆▇█▇▆▄▁    ▁▃▂▂▁                   ▁▁▁▁▁                  ▂
  ▅████████▇▆▇▇███████▇▆▅▄▃▃▃▃▁▁▅▄▆▆▇████████▇▆▆▄▅▄▃▄▃▄▄▄▁▄▅▃ █
  7.7 μs       Histogram: log(frequency) by time      13.9 μs <

 Memory estimate: 15.88 KiB, allocs estimate: 2.

Original + IRCE:

julia> @benchmark f!(zeros(Float64, 1000), rand(Float64, 1000))
BenchmarkTools.Trial: 10000 samples with 9 evaluations.
 Range (min … max):  2.650 μs … 272.520 μs  ┊ GC (min … max): 0.00% … 97.36%
 Time  (median):     2.946 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   3.972 μs ±   7.472 μs  ┊ GC (mean ± σ):  9.37% ±  5.22%

  ▂▆██▇▄                                  ▁▂▃▃▃▂▁▁  ▁▂▃▃▃▂▁▁  ▂
  ███████▇▆▄▄▁▃▄▁▁▁▃▁▃▅▅▄▅▄▃▁▃▁▁▃▃▃▁▁▁▁▄▆███████████████████▇ █
  2.65 μs      Histogram: log(frequency) by time      7.64 μs <

 Memory estimate: 15.88 KiB, allocs estimate: 2.

Original + inbounds:

julia> @benchmark f!(zeros(Float64, 1000), rand(Float64, 1000))
BenchmarkTools.Trial: 10000 samples with 9 evaluations.
 Range (min … max):  2.593 μs … 300.730 μs  ┊ GC (min … max):  0.00% … 97.46%
 Time  (median):     2.934 μs               ┊ GC (median):     0.00%
 Time  (mean ± σ):   4.029 μs ±   8.744 μs  ┊ GC (mean ± σ):  11.03% ±  5.24%

  ▆▇▆██▆▂                                 ▁▂▂▃▃▃▂▁  ▁▂▂▃▃▃▂▂  ▂
  ███████▇▆▆▅▁▃▅▅▃▆▆▆▆▅▆▅▅▃▁▁▅▃▁▁▁▃▁▁▃▅▅▅▇███████████████████ █
  2.59 μs      Histogram: log(frequency) by time      7.73 μs <

 Memory estimate: 15.88 KiB, allocs estimate: 2.

Test Case 2:

function f2!(a, xs)
                  xs = Base.Experimental.Const(xs)
                  Base.Experimental.@aliasscope begin
                      for i in eachindex(xs)
                          a[1] += xs[i]
                      end
                  end
              end

Original:

julia> @benchmark f2!(zeros(Float64, 1), rand(Float64, 1000))
BenchmarkTools.Trial: 10000 samples with 7 evaluations.
 Range (min … max):  4.666 μs … 299.508 μs  ┊ GC (min … max): 0.00% … 97.83%
 Time  (median):     6.269 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   5.914 μs ±   7.616 μs  ┊ GC (mean ± σ):  3.72% ±  3.04%

   █▂
  ▇██▅▃▂▂▂▂▂▂▂▂▂▂▁▂▂▂▂▂▂▂▂▂▂▁▁▂▁▁▁▂▂▂▂▃▄▆▇▇▆▄▃▂▂▂▂▂▂▂▃▃▃▃▃▃▂▂ ▃
  4.67 μs         Histogram: frequency by time        7.23 μs <

 Memory estimate: 8.00 KiB, allocs estimate: 2.

Original + IRCE:

julia> @benchmark f2!(zeros(Float64, 1), rand(Float64, 1000))
BenchmarkTools.Trial: 10000 samples with 9 evaluations.
 Range (min … max):  2.306 μs … 237.869 μs  ┊ GC (min … max): 0.00% … 98.01%
 Time  (median):     2.399 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   2.922 μs ±   4.680 μs  ┊ GC (mean ± σ):  4.97% ±  3.55%

  ▅█▇▄▁                              ▂▃▃▃▂▂▁           ▁▁▁    ▂
  ██████▇▆▄▅▃▁▁▅▅▅▄▄▃▁▃▃▃▁▃▁▁▁▁▁▁▁▃▆██████████▇▆▆▃▅▃▅▆██████▇ █
  2.31 μs      Histogram: log(frequency) by time      5.34 μs <

 Memory estimate: 8.00 KiB, allocs estimate: 2.

Original + inbounds:

BenchmarkTools.Trial: 10000 samples with 9 evaluations.
 Range (min … max):  2.330 μs … 253.731 μs  ┊ GC (min … max): 0.00% … 98.23%
 Time  (median):     2.448 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   2.927 μs ±   4.951 μs  ┊ GC (mean ± σ):  5.70% ±  3.60%

  ▅▇██▄                                  ▂▃▃▃▃▂▁▁▁  ▁▁▁▂▁ ▁   ▂
  ██████▇▇▆▆▆▅▅▃▃▁▁▄▁▁▄▄▅▅▄▁▄▃▄▁▁▄▁▁▃▃▁▆████████████████████▇ █
  2.33 μs      Histogram: log(frequency) by time      4.75 μs <

 Memory estimate: 8.00 KiB, allocs estimate: 2.

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[vtjnash]

@nanosoldier runbenchmarks(ALL, vs = ":master")

[nanosoldier]

Something went wrong when running your job:

NanosoldierError: error when preparing/pushing to report repo: failed process: Process(setenv(`git push`; dir="/run/media/system/data/nanosoldier/workdir/NanosoldierReports"), ProcessExited(1)) [1]

Unfortunately, the logs could not be uploaded.
cc @christopher-dG

[vchuravy]

@vtjnash can you upload the logs? and maybe we should remove @christopher-dG from the ping list.

[vtjnash]

https://github.com/JuliaCI/NanosoldierReports/tree/master/benchmark/by_hash/125170e_vs_1389c2f/report.md

[vtjnash]

Is this related: ["misc", "allocation elision view", "no conditional"] 2.14 (5%) ❌ 1.00 (1%)

Otherwise, all others seem to be noise