DateTime parsing regression

[gbaraldi]

#50367 Caused a regression in the run(BaseBenchmarks.SUITE[["dates", "parse", "DateTime"]]) test of BaseBenchmarks.jl @BioTurboNick

[BioTurboNick]

Is it possible to figure out exactly what method(s) slowed down?

[gbaraldi]

Probably, the benchmark is here https://github.com/JuliaCI/BaseBenchmarks.jl/blob/0ec1edace2d731de8d926f4b48a22cf2e5cc99c2/src/misc/MiscellaneousBenchmarks.jl#L73

[BioTurboNick]

1.9.2:

using BenchmarkTools, Dates
@btime map(string, range(DateTime("2016-02-19T12:34:56"), step = Dates.Millisecond(123), length = 200));
  126.700 μs (4437 allocations: 218.20 KiB)

#50367:

@btime map(string, range(DateTime("2016-02-19T12:34:56"), step = Dates.Millisecond(123), length = 200));
  103.700 μs (4437 allocations: 218.20 KiB)

I'm not seeing a regression? Or I didn't find the right one.

[gbaraldi]

I sent you the wrong link sorry. It's actually https://github.com/JuliaCI/BaseBenchmarks.jl/blob/0ec1edace2d731de8d926f4b48a22cf2e5cc99c2/src/dates/DatesBenchmarks.jl#L88

[BioTurboNick]

Hmm... It's not super obvious what's going on here. Profiling some of the functions doesn't show anything obviously passing through reinterpret. Inspected a couple levels of @code_native but they're identical.

_datetime_str = "2016-02-19T12:34:56.78"
_date_str = "2016-02-19"

DateTime($_datetime_str) # 51.976 ns (0 allocations: 0 bytes) vs. 97.574 ns (0 allocations: 0 bytes)
Date($_date_str) # 38.004 ns (0 allocations: 0 bytes) vs. 57.841 ns (0 allocations: 0 bytes)

_custom_date_str = "20160219"
_custom_date_fmt = "yyyymmdd"
_custom_datetime_str = "20160219 123456.78"
_custom_datetime_fmt = "yyyymmdd HHMMSS.sss"

Date($_custom_date_str, $_custom_date_fmt) # 2.144 μs (21 allocations: 1.16 KiB) vs. 2.411 μs (21 allocations: 1.16 KiB)
DateTime($_custom_datetime_str, $_custom_datetime_fmt) # 4.157 μs (46 allocations: 2.72 KiB) vs. 4.629 μs (46 allocations: 2.72 KiB)

DateTime($_datetime_str, $(Dates.ISODateTimeFormat)) # 51.521 ns (0 allocations: 0 bytes)
Date($_date_str, $(Dates.ISODateFormat)) # 38.951 ns (0 allocations: 0 bytes) vs. 97.992 ns (0 allocations: 0 bytes)
DateTime("Sat, 12 Nov 2016 07:45:36", $(Dates.RFC1123Format)) # 168.477 ns (0 allocations: 0 bytes) vs. 206.903 ns (0 allocations: 0 bytes)
DateTime("sat, 12 Nov 2016 07:45:36", $(Dates.RFC1123Format)) # 184.501 ns (0 allocations: 0 bytes) vs. 206.964 ns (0 allocations: 0 bytes)
DateTime("sAt, 12 Nov 2016 07:45:36", $(Dates.RFC1123Format)) # 272.136 ns (3 allocations: 120 bytes) vs. 301.229 ns (3 allocations: 120 bytes)

[gbaraldi]

The bisect is at least confirmed? It might an inlining change or something. I didn't take too deep a look for now.