`parse(Float32,x)` doesn't detect overflow/underflow on windows

[c42f]

On linux, we have

julia> parse(Float32, "10e100")
ERROR: ArgumentError: cannot parse "10e100" as Float32

However, on windows with Julia 1.8, we have

julia> parse(Float32, "10e100")
Inf32

Similarly for underflow, as in parse(Float32, "1e-100") (though arguably it's less clear that we should throw in this case? We generally accept loss of precision as an unavoidable part of float parsing in general.)

---

I think this can be traced to a failure in whichever C runtime we're using on windows, as a ccall to jl_strtof_c results in:

julia> Libc.errno(0)
       @ccall jl_strtof_c("10e100"::Cstring, Ref{Ptr{UInt8}}(C_NULL)::Ptr{Ptr{UInt8}})::Cfloat
Inf32

julia> Libc.errno()
0

(on linux, we have Libc.errno() == Libc.ERANGE after the same code is run.)

[oscardssmith]

what behavior do we want here? I can see arguments for error, inf, or warning.

[gbaraldi]

I would follow linux by principle, Also couldn't we move that parsing to inside julia? Or do we lose too much performance/isn't feaseable.

[KristofferC]

Also couldn't we move that parsing to inside julia? Or do we lose too much performance/isn't feaseable.

#45301

[c42f]

what behavior do we want here

Good question. The Julia parser Meta.parse() is inconsistent with Base.parse() for floats - Meta.parse() will silently truncate underflow to 0.0. This also appears to be the behavior in Parsers.jl.

I think the most useful behavior for general text->float parsing is to silently underflow to 0.0 by default, as an extension of the fact that loss of precision is acceptable. We shouldn't emit warnings from Base.parse but we could have a keyword to change the behavior on underflow I guess? Technically changing this would be breaking but it seems relatively harmless?

For overflow I feel like throwing is a good default? Though being able to get Inf by using a keyword could be nice.

For Julia source code I think it would be best to emit a warning for underflow but this is a bit of a different case from general text parsing.

[metab0t]

See #45612 (comment) and https://github.com/JuliaLang/julia/blob/master/src/support/strtod.c

On Windows, jl_strtof_c is implemented by converting result of the fallback implementation of strtod_l to float instead of calling strtof_l in libc directly.