Tweak bounds check elimination

[pchintalapudi]

This is a combination of two changes:

1. Move the IRCE pass further down the pass pipeline and add an extra GVN pass
   1. This allows some additional functions to become vectorizable without an inbounds macro.
2. Change size(Vector) to delegate to arraylen instead of arraysize
   1. This converts gep inbounds {}*, {}** instructions to instead use gep inbounds { i8*, i64, i16, i16, i32 }, { i8*, i64, i16, i16, i32 }* to get the array length.
   2. Without this, LLVM does not see that the two geps are equivalent, and therefore fails to eliminate one of the loads. This has implications on whether or not some functions vectorize as well as overall code size (a useless loop inStatistics._mean(identity, Int[], :) was completely eliminated after this change).

[vchuravy]

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

[pchintalapudi]

Test case 1:

function f4!(a, xs)
                       s = a[1]
                      for i in 1:length(xs)
                           s2 = 0
                       for j in 1:length(xs[i])
                           s3 = 0
                           for k in 1:length(xs[i][j])
                               s3 += xs[i][j][k]
                           end
                           s2 += s3
                       end
                          s += s2
                      end
                       a[1] = s
                   end

The function was benchmarked with a Vector{100 x Vector{100 x Vector{100 x random Int}}} as second input.

master

Godbolt: https://godbolt.org/z/Kc8zM6616

julia> @benchmark f4!([0], $(c))
BenchmarkTools.Trial: 8743 samples with 1 evaluation.
 Range (min … max):  499.955 μs …  4.058 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     568.345 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   570.304 μs ± 62.607 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                     ▂ ▃▃  ▁▂▃▁  ▁█▂▂▄▃▃▅   ▁  ▁
  ▄▁▅▁▁▃▅▁▁▃█▁▁▃▄▁▅▁▁▇▆▆▄▅▇▆▅▁█▇▄▄▅▅▄█▆███▇████▇▇█████████▇███ █
  500 μs        Histogram: log(frequency) by time       585 μs <

 Memory estimate: 64 bytes, allocs estimate: 1.

PR

Godbolt: https://godbolt.org/z/vK73fevsM

julia> @benchmark f4!([0], $(c))
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):  69.419 μs … 115.779 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     69.430 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   69.558 μs ±   1.019 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

  █▁  ▁                                                        ▁
  ██▃▁█▁▁▁▃▁▁▃▁▁▄▁▁▃▁▁▁▃▃▄▁▁▁▁▁▁▁▁▅▄▁▁▁▄▄▄▄▅▄▄▆▄▄▄▄▄▆▆▆▅▅▄▅▄▄▄ █
  69.4 μs       Histogram: log(frequency) by time        74 μs <

 Memory estimate: 64 bytes, allocs estimate: 1.

[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.

[vtjnash]

Seems a bit noisy, but it is unclear. You might want to re-run those that showed significant deltas: https://github.com/JuliaCI/NanosoldierReports/blob/master/benchmark/by_hash/c3f5381_vs_50fcb03/report.md

[pchintalapudi]

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

[vchuravy]

Bump! I think we can rebase this now that the Statistics.jl changes are on master

[vchuravy]

@pchintalapudi can you test if this fixes #43308?

[vchuravy]

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

[nanosoldier]

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

[vchuravy]

@nanosoldier runtests(["AlphaStableDistributions", "BenchmarkTools", "CMAEvolutionStrategy", "CopEnt", "DiffEqOperators", "FaultDetectionTools", "GPUArrays", "ITensors", "IntervalTrees", "NODAL", "NumericalAlgorithms", "PartialSvdStoch", "Pitaya", "ProgressiveHedging", "QuadEig", "QuantumPropagators", "SLEEFPirates", "SimpleSDMLayers", "SnowyOwl", "StochasticRounding", "Wflow", "Zygote"], vs = ":master")

[nanosoldier]

Something went wrong when running your job:

NanosoldierError: failed to run tests: NanosoldierError: failed to upload test log: AWSException: InvalidAccessKeyId -- The AWS Access Key Id you provided does not exist in our records.
HTTP.ExceptionRequest.StatusError(403, "PUT", "/julialang-reports/nanosoldier/pkgeval/by_hash/f65ee36_vs_b3e4341/NumericalAlgorithms.1.8.0-DEV-e225ea776a0.log?x-amz-acl=public-read", HTTP.Messages.Response:
"""
HTTP/1.1 403 Forbidden
x-amz-request-id: NXY9SHDPHBSSBZ63
x-amz-id-2: Ox0mg6Med7cK17AdIZbQK0ghk2M/LsH1/aykBdneySzQ+EDpbYcyOj4rF3kpRFkIG1H2yq3YNP0=
Content-Type: application/xml
Transfer-Encoding: chunked
Date: Wed, 15 Dec 2021 21:36:47 GMT
Server: AmazonS3
Connection: close

<?xml version="1.0" encoding="UTF-8"?>
<Error><Code>InvalidAccessKeyId</Code><Message>The AWS Access Key Id you provided does not exist in our records.</Message><AWSAccessKeyId>AKIA4WZGSTHCMSSHRWL6</AWSAccessKeyId><RequestId>NXY9SHDPHBSSBZ63</RequestId><HostId>Ox0mg6Med7cK17AdIZbQK0ghk2M/LsH1/aykBdneySzQ+EDpbYcyOj4rF3kpRFkIG1H2yq3YNP0=</HostId></Error>""")

Logs and partial data can be found here
cc @maleadt

[vchuravy]

Ah great yet another AWS failure...

[pchintalapudi]

The difference between gp and gp_inbounds in #43308 should be fixed by e07c2bf, but that requires the addition of an additional DCE pass to eliminate a select undef statement. The finer difference between ge_inbounds and gp_inbounds is due to LLVM unrolling the loop by 16X for gp_inbounds but only 8X for ge_inbounds prior to SLP vectorization, which is something that is not addressed by this PR.

[vchuravy]

@nanosoldier runtests(["AlphaStableDistributions", "BenchmarkTools", "CMAEvolutionStrategy", "CopEnt", "DiffEqOperators", "FaultDetectionTools", "GPUArrays", "ITensors", "IntervalTrees", "NODAL", "NumericalAlgorithms", "PartialSvdStoch", "Pitaya", "ProgressiveHedging", "QuadEig", "QuantumPropagators", "SLEEFPirates", "SimpleSDMLayers", "SnowyOwl", "StochasticRounding", "Wflow", "Zygote"], vs = ":master")

[nanosoldier]

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