[BUG] Wrong limb width when cross compiling

[alexhuszagh]

[BUG] Wrong limb width when cross compiling

[dtolnay]

Thanks -- Personally I don't use lexical-core so I don't have any stake in a release. I fixed this a long time ago on my end in serde-rs/json@6e00729. Serde uses a fork because I wasn't confident in the maintenance of this crate.

[myrrlyn]

Is this the correct resolution?

[dtolnay]

No. The target arch of the build script should have no bearing. Only the target arch which the library is being built for is relevant.

[dtolnay]

In other words absence of CARGO_CFG_TARGET_ARCH needs to be either a build error (preferred), or should guess a consistent limb width regardless of what the host architecture happens to be.

[myrrlyn]

I was about to ask what the resolution should be in the Err case. I'll fail the build.

[myrrlyn]

Once you're satisfied that this is correct, I'll publish an update, probably in May.

I don't know the versioning history of lexical-core or what its current dependents are. Do you know of a need for me to backport to 0.6, or will an 0.7 patch release suffice?

[Alexhuszagh]

@dtolnay I'm glad it was of use to you, sorry, for mental health reasons I haven't exactly been the best maintainer so I'm very glad an adaptation has suited your needs. Thank you for reporting this.

Are there any cases where CARGO_CFG_TARGET_ARCH will not be defined? If so, it's pretty easy to default to 32-bit limbs if the environment variable is not set. We use an optimization when 64-bit integer operations are faster on the CPU, but if that's unknown, it would be better to default to 32-bit limbs than error during build.

[Alexhuszagh]

I think I've satisfactorily fixed this issue. I've removed the target_arch of the build script from build.rs, and defaulted to 32-bit limbs if CARGO_CFG_TARGET_ARCH is not set, since that should be reasonably performant on most architectures.

[Alexhuszagh]

Closed as of v0.4.8, v0.5.1, v0.6.8, and v0.7.6.