Cast from u128 to f32 is broken in release mode

[antoyo]

The test tests/ui/numbers-arithmetic/u128-as-f32.rs was disabled in bb3ee76.

This used to work with GCC version b8f4bbf1b995d9511bad6652e8d774620b2bd833 , but with the rebase to GCC 3e3878e7183f1d1e1ccf5aba99da7fbb36c012c8, this test now fails when compiled in release mode.

Something weird is that if I print the f32 value, the test passes.

[FractalFir]

Something weird is that if I print the f32 value, the test passes.

This is also something that happens for most(if not all) floating point issues I see on ARM.

[antoyo]

Also, the code that checks if the two values are equal seems to be removed by GCC's optimizer.
That made me wonder if we implement black_box properly.

[tgross35]

Could you post the output of the failing test? Maybe some pattern will stand out

[antoyo]

@tgross35: Here's the output:

thread 'main' panicked at src/main.rs:38:5:
assertion `left == right` failed: run-time u128 -> f32
  left: inf
 right: inf
stack backtrace:
   0: __rustc::rust_begin_unwind
   1: core::panicking::panic_fmt.localalias
   2: core::panicking::assert_failed_inner
   3: core::panicking::assert_failed.constprop.0
   4: test_rust::main
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.

[tgross35]

Oh, well unfortunately that's not very helpful 🙃