fix parse(ComplexF64, "inf")

[stevengj]

I noticed that the parsing code in #24713 was getting confused by the i in "inf" when parsing a purely real infinity. This PR fixes that and adds some more tests for floating-point values.

Probably could be backported?

[StefanKarpinski]

I've added backport labels and triage to discuss but maybe the latter is unnecessary since this seems like a straightforward bugfix so long as it applies.

[JeffBezanson]

Seems to have relevant failures on windows.

[mbauman]

Test failures from the windows buildbots look highly relevant. For example:

Error During Test at C:\cygwin\home\Administrator\buildbot\worker-tabularasa\tester_win64\build\share\julia\test\parse.jl:299
  Test threw exception
  Expression: n === parse(ComplexF64, s)
  ArgumentError: cannot parse " nan" as Float64

[stevengj]

@mbauman, is that a Windows bug? On my Mac I get:

julia> parse(Float64, " nan")
NaN

We should definitely be consistent across platforms here.

[mbauman]

I'm just copy-pasting the error message from the buildbots. I don't know anything more.

[GregPlowman]

On Windows 10 with Julia 1.1.0 I get this:

Julia-1.1.0> parse(Float64, "nan")
NaN

Julia-1.1.0> parse(Float64, "nan ")
NaN

Julia-1.1.0> parse(Float64, " nan")
ERROR: ArgumentError: cannot parse " nan" as Float64
Stacktrace:
 [1] _parse_failure(::Type, ::String, ::Int64, ::Int64) at .\parse.jl:370 (repeats 2 times)
 [2] #tryparse_internal#349 at .\parse.jl:366 [inlined]
 [3] tryparse_internal at .\parse.jl:364 [inlined]
 [4] #parse#350 at .\parse.jl:376 [inlined]
 [5] parse(::Type{Float64}, ::String) at .\parse.jl:376
 [6] top-level scope at none:0

[stevengj]

Sounds like a bug in on jl_strtod_c in strtod.c, which is called on Windows because it doesn't have strtod_l (though it does have _strtod_l… not sure why we don't call that? edit: @tknopp explained in #5928 (comment) that mingw doesn't/didn't have _strtod_l).

Seems like the bug dates to #5988 in Julia 0.3.

[stevengj]

added a bugfix for windows strtod

[stevengj]

AppVeyor is now passing. Travis Linux x86 failures (in zillions of tests) look unrelated. win32 buildbot timeout looks unrelated.

Note that this PR does not need to be squashed — it's two commits that fix two bugs.