`test_convert_f32` test failed on `i686-pc-windows-*` target

test failed on i686-pc-windows-msvc, but succeeded on x86_64-pc-windows-msvc.

toolchain:

$ rustup --version
rustup 1.25.1 (bb60b1e89 2022-07-12)
info: This is the version for the rustup toolchain manager, not the rustc compiler.
info: The currently active `rustc` version is `rustc 1.65.0 (897e37553 2022-11-02)`

command:

$ cargo test --target=i686-pc-windows-msvc
    Finished test [unoptimized + debuginfo] target(s) in 0.01s
     Running unittests src\lib.rs (target\i686-pc-windows-msvc\debug\deps\num_bigint-4a5286ac62f80819.exe)

running 13 tests
test bigint::test_assign_from_slice ... ok
test bigint::test_from_biguint ... ok
test bigint::test_from_slice ... ok
test biguint::iter::test_iter_u32_digits ... ok
test biguint::iter::test_iter_u32_digits_be ... ok
test biguint::iter::test_iter_u64_digits ... ok
test biguint::iter::test_iter_u64_digits_be ... ok
test biguint::power::test_pow_biguint ... ok
test biguint::multiplication::test_sub_sign ... ok
test biguint::test_from_slice ... ok
test biguint::test_u128_u32_roundtrip ... ok
test biguint::test_u32_u128 ... ok
test biguint::power::test_plain_modpow ... ok

test result: ok. 13 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

     Running tests\bigint.rs (target\i686-pc-windows-msvc\debug\deps\bigint-f180ad9c24c76948.exe)

running 52 tests
test test_abs_sub ... ok
test test_binary ... ok
test test_bit ... ok
test test_checked_add ... ok
test test_checked_sub ... ok
test test_checked_mul ... ok
test test_cmp ... ok
test test_convert_from_biguint ... ok
test test_checked_div ... ok
test test_convert_from_int ... ok
test test_add ... ok
test test_convert_from_uint ... ok
test test_convert_i128 ... ok
test test_convert_f32 ... FAILED
test test_convert_i64 ... ok
test test_convert_to_biguint ... ok
test test_convert_u128 ... ok
test test_convert_f64 ... ok
test test_convert_u64 ... ok
test test_display ... ok
test test_from_bytes_be ... ok
test test_from_bytes_le ... ok
test test_div_ceil ... ok
test test_from_signed_bytes_be ... ok
test test_from_signed_bytes_le ... ok
test test_from_str_radix ... ok
test test_hash ... ok
test test_gcd ... ok
test test_iter_product ... ok
test test_iter_product_generic ... ok
test test_iter_sum ... ok
test test_iter_sum_generic ... ok
test test_lcm ... ok
test test_lower_hex ... ok
test test_neg ... ok
test test_negative_shr ... ok
test test_next_multiple_of ... ok
test test_div_mod_floor ... ok
test test_octal ... ok
test test_prev_multiple_of ... ok
test test_mul ... ok
test test_set_bit ... ok
test test_to_bytes_be ... ok
test test_to_bytes_le ... ok
test test_to_signed_bytes_be ... ok
test test_to_signed_bytes_le ... ok
test test_upper_hex ... ok
test test_sub ... ok
test test_div_rem ... ok
test test_pow ... ok
test test_signed_bytes_le_round_trip ... ok
test test_signed_bytes_be_round_trip ... ok

failures:

---- test_convert_f32 stdout ----
thread 'test_convert_f32' panicked at 'assertion failed: `(left == right)`
  left: `Some(1.8014399e16)`,
 right: `Some(1.8014397e16)`', tests\bigint.rs:410:5
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace


failures:
    test_convert_f32

test result: FAILED. 51 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.50s

error: test failed, to rerun pass `--test bigint`

It seems that n as i64 as f32 and n as u64 as f32 give different results on a 32-bit target.

num-bigint/tests/bigint.rs

Lines 408 to 413 in a730396

	let mut n: i64 = 0b0000000000111111111111111111111111011111111111111111111111111111;
	assert!((n as f64) as f32 != n as f32);
	assert_eq!(BigInt::from(n).to_f32(), Some(n as f32));
	n = -n;
	assert!((n as f64) as f32 != n as f32);
	assert_eq!(BigInt::from(n).to_f32(), Some(n as f32));

[cuviper]

FWIW, all tests do pass on i686-unknown-linux-gnu, and this is even tested in CI.

I haven't tested other 32-bit targets yet, but it seems to be an issue only with 32-bit Windows targets.

[cuviper]

Hmm, I was able to reproduce this on i686-pc-windows-gnu as well, so at least I can debug that via Wine in my Linux development environment.

[CryZe]

We concluded that it's the FPU Control Word that is configured differently by default on Windows environments to use 53 bits instead of 64 bits like on Linux. This would need to be special cased in LLVM or so.

[cuviper]

By "we", do you mean that you are working with @icedrocket?

I was thinking this should be closed as a compiler issue, and it sounds like you would agree, right?

[CryZe]

No, I simply tried reproducing it once I saw the issue in the rust repo and then figured out what caused it. I'm not affiliated with them. But yeah this for sure is a compiler / LLVM issue.

[cuviper]

Ah, now I see the linked issue rust-lang/rust#105626 -- then let's leave it there.