regex-automata v0.3.2: test failures on 32-bit architectures

[decathorpe]

I'm trying to update the Fedora Linux packages for the regex crate and its dependencies to the latest version, but I'm getting stuck on test failures on 32-bit x86 (i.e. i686-unknown-linux-gnu). Unit tests seem to be fine, but some doctests fail:

failures:
    src/dfa/dense.rs - dfa::dense::Config::determinize_size_limit (line 880)
    src/dfa/dense.rs - dfa::dense::Config::determinize_size_limit (line 913)
    src/util/captures.rs - util::captures::Captures::get_group (line 434)
test result: FAILED. 447 passed; 3 failed; 2 ignored; 0 measured; 0 filtered out; finished in 85.38s

The first two failures produce absolutely gigantic log output on failure, I've pasted the complete output of "cargo test" in a Gist here (though GitHub seems to cut off the content after only 12,000 of 19,000 lines): https://gist.github.com/decathorpe/a5e715d566c5b75838ccd1fbfd4cf4fa

But the gist (pun intended) seems to be:

---- src/dfa/dense.rs - dfa::dense::Config::determinize_size_limit (line 913) stdout ----
Test executable failed (exit status: 101).
stderr:
thread 'main' panicked at 'called `Result::unwrap_err()` on an `Ok` value: dense::DFA(
(... snip 6000 lines of text output ...)

and

---- src/dfa/dense.rs - dfa::dense::Config::determinize_size_limit (line 880) stdout ----
Test executable failed (exit status: 101).
stderr:
thread 'main' panicked at 'called `Result::unwrap_err()` on an `Ok` value: dense::DFA(
(... snip 12000 lines of text output ...)

The third failure is a simple one (numeric literal too large for usize, which is typically only 32-bit integer on 32-bit architectures):

---- src/util/captures.rs - util::captures::Captures::get_group (line 434) stdout ----
error: literal out of range for `usize`
  --> src/util/captures.rs:447:33
   |
16 | assert_eq!(None, caps.get_group(9944060567225171988));
   |                                 ^^^^^^^^^^^^^^^^^^^
   |
   = note: the literal `9944060567225171988` does not fit into the type `usize` whose range is `0..=4294967295`
   = note: `#[deny(overflowing_literals)]` on by default
error: aborting due to previous error
Couldn't compile the test.

[decathorpe]

build.log (it appears that uploading the log as an attachment works better)

[BurntSushi]

Thankfully all of these failures are benign. I've put up #1040 to just skip these tests on non-64-bit.

The thornier issue here is that cross doesn't seem to run doc tests, which is why these tests don't cause this repo's CI to fail. We do indeed test on i686-unknown-linux-gnu, but it doesn't capture the (significant number) of doc tests.

[decathorpe]

Thanks for looking into it! I'll just skip those three doctests on 32-bit x86 for now if they're harmless. 👍🏼

[BurntSushi]

Yup it's basically just a matter of the doc tests exercising the effects of certain limits on heap usage. 32 and 64 bit will vary there and thus the limits just have different effects. And the capture group one is harmless for the reason you point out: it only exercises a case that fails in 64-bit.