Support running cargo miri in situations where target != host (e.g. "cross-interpretation")

[thomcc]

The playground has the ability to run code under miri, but it would nice to be able to choose other targets.

A somewhat unknown (but highly useful) feature of miri is that it can run as a cross-interpreter. That is, even on x86_64-apple-darwin (or whatever), if I have, say the powerpc-unknown-linux-gnu target installed, I can do:

cargo miri test --target=powerpc-unknown-linux-gnu

and it will simulate running the test on that architecture, including emulating 32-bit big-endian linux (well, the set of linux syscalls that miri supports, anyway).

Right now I can't do this on the playground, and it would be useful to do quick checks and examples of big-endian code.

---

The downside is needing UI for this, and needing to rustup target add on the playground servers. So I understand some hesitancy.

Personally, I think it's most valuable for targets that are very different than x86_64 — e.g. I feel like just saying having an option for big-endian 32 and 64 bit targets would be good enough.

(It would be nice to also have one for target_pointer_width="16", like msp430-* or avr-*, but these being no_std only, it's hard enough to actually run the code that users probably should just do it locally)

---

Anyway, I thought I'd mention it. I kind of suspect it's the kind of "that'd be nice but doing it in a non-hacky way is likely not worth the effort" thing, but figured I'd suggest anyway. No worries if you'd rather wontfix this.

Note that this is related to #446, but not the same, as --target isn't a RUSTFLAG (or a MIRIFLAG), its a flag for cargo.

[thomcc]

Hm, I guess I did something wrong here, since it says I added wontfix. I didn't mean to.

If it's just github being weird, can I get some justification there?

EDIT: I see what happened. I missed the small link to not use the templat on the "new issue" page, so just started with the "plz add my crate" template and deleted it and started fresh. Unfortunately, That template also has you add a wontfix to your own issue.

[shepmaster]

I missed the small link

Yeah, you'd be surprised at how many people don't read the template before opening an issue.

[shepmaster]

The downside is needing UI for this, and needing to rustup target add on the playground servers

Does cargo miri setup need to be run for each target?

[RalfJung]

if I have, say the powerpc-unknown-linux-gnu target installed

You don't even need to have the target installed. The docs don't tell you to install a target and I never said so either when advertising this feature, so I wonder how this implicit assumption crept in.^^

Does cargo miri setup need to be run for each target?

Yes (but this also happens automatically when needed).

[RalfJung]

Note that this is related to #446, but not the same, as --target isn't a RUSTFLAG (or a MIRIFLAG), its a flag for cargo.

Ah, good point... but in principle we could use similar infrastructure for both of these kinds of flags.

[shepmaster]

happens automatically when needed

If I'm understanding you correctly, that'd be a bad thing on the playground, as the setup we do now takes ~150 seconds which would blow far past the request timeout of 10 seconds.

To make sure we are on the same page, what do you think the timing of each of these commands would be if run sequentially?

cargo miri setup
cargo miri test --target=x86_64-unknown-linux-gnu
cargo miri test --target=powerpc-unknown-linux-gnu

[RalfJung]

If I'm understanding you correctly, that'd be a bad thing on the playground, as the setup we do now takes ~150 seconds which would blow far past the request timeout of 10 seconds.

Yeah, it'd probably be way too long.

To make sure we are on the same page, what do you think the timing of each of these commands would be if run sequentially?

On a "x86_64-unknown-linux-gnu" host I assume? The last command doesn't benefit from the previous ones, it builds a new standard library, which takes around 50-60s on my system.

[shepmaster]

So to make this performant, we'd need to do this when pre-building the container:

cargo miri setup --target=x86_64-unknown-linux-gnu
cargo miri setup --target=powerpc-unknown-linux-gnu
...

Once for each playground-supported target. Does that sound correct?

[RalfJung]

Yes, that sounds right. However, don't you also need to pre-build the dependencies? So you can just do cargo miri run --target=X and that will do the setup part automatically.

[shepmaster]

Maybe, I dunno! I assume that's changed over time and the build has never been updated since it worked:

https://github.com/integer32llc/rust-playground/blob/d03009c20002b21d76f2f01b1efdd2749a0559f1/compiler/miri/Dockerfile#L6-L8

[RalfJung]

Yeah the cargo miri setup part of this is redundant (but doesn't hurt).