rustc_apfloat: Panic in Double::mul_add_r

[mjdr]

I tried this code and get a panic:

#[test]
fn issue_93224() {
    let r1 = Double::from_bits(0x000000000000002D as u128);
    let r2 = Double::from_bits(0xC2D6C16C166666DE as u128);
    let r3 = Double::from_bits(0x0000000000000055 as u128);

    let _ = r1.mul_add_r(r2, r3, Round::NearestTiesToEven);
}

Meta

Commit: master 17d29dc

Backtrace

thread 'issue_93224' panicked at 'assertion failed: `(left == right)`
  left: `1`,
 right: `0`', compiler/rustc_apfloat/src/ieee.rs:2545:17
stack backtrace:
   0: rust_begin_unwind
             at /rustc/092e1c9d23158d81be27bb6f71bdd0c6282478fb/library/std/src/panicking.rs:498:5
   1: core::panicking::panic_fmt
             at /rustc/092e1c9d23158d81be27bb6f71bdd0c6282478fb/library/core/src/panicking.rs:110:14
   2: core::panicking::assert_failed_inner
   3: core::panicking::assert_failed
   4: rustc_apfloat::ieee::sig::add_or_sub
   5: <rustc_apfloat::ieee::IeeeFloat<S> as rustc_apfloat::Float>::mul_add_r
   6: ieee::issue_93224
   7: ieee::issue_93224::{{closure}}
   8: core::ops::function::FnOnce::call_once
   9: core::ops::function::FnOnce::call_once
             at /rustc/092e1c9d23158d81be27bb6f71bdd0c6282478fb/library/core/src/ops/function.rs:227:5

[eddyb]

Can you reproduce with LLVM's APFloat? Might even be able to trigger it by getting LLVM to constant-fold FMA (assuming they have as many checks as we do).

[mjdr]

This test passes

Add to unittests/ADT/APFloatTest.cpp

TEST(APFloatTest, FMA_issue_93224) {
    uint64_t u1 = 0x000000000000002D;
    uint64_t u2 = 0xC2D6C16C166666DE;
    uint64_t u3 = 0x0000000000000055;

    APFloat f1(*((double*)&u1));
    APFloat f2(*((double*)&u2));
    APFloat f3(*((double*)&u3));
    f1.fusedMultiplyAdd(f2, f3, APFloat::rmNearestTiesToEven);
}

[eddyb]

This test passes

@mjdr Did you enable LLVM assertions? I think I independently (re)investigated this bug just now (because fuzzing found it again), and a C++ assert would trigger once I got rid of NDEBUG - see #100233 (comment)

EDIT: forgot it's not fixed yet, but will be once I advance the port enough to reach the upstream fix.

[eddyb]

I posted a longer update to #55993 (comment) regarding the progress with the whole rustc_apfloat situation, but of note for this issue is that the WIP repo already has this bug fixed.