incorrect code generation for i686 release build for 1.47 beta and nightly

[tspiteri]

/// ```rust
/// assert_eq!(fixed::i124f4_wrapping_mul(3 << 4, 2 << 4), 6 << 4);
/// ```
#[inline]
pub fn i124f4_wrapping_mul(lhs: i128, rhs: i128) -> i128 {
    let (ans, _) = mul_overflow(lhs, rhs, 4);
    ans
}

#[inline]
fn mul_overflow(lhs: i128, rhs: i128, frac_nbits: u32) -> (i128, bool) {
    let (lh, ll) = hi_lo(lhs);
    let (rh, rl) = hi_lo(rhs);
    let ll_rl = ll.wrapping_mul(rl);
    let lh_rl = lh.wrapping_mul(rl);
    let ll_rh = ll.wrapping_mul(rh);
    let lh_rh = lh.wrapping_mul(rh);

    let col01 = ll_rl as u128;
    let (col01_hi, col01_lo) = (col01 >> 64, col01 & !(!0 << 64));
    let partial_col12 = lh_rl + col01_hi as i128;
    let (col12, carry_col3) = carrying_add(partial_col12, ll_rh);
    let (col12_hi, col12_lo) = hi_lo(col12);
    let ans01 = shift_lo_up_unsigned(col12_lo) + col01_lo;
    let ans23 = lh_rh + col12_hi + shift_lo_up(carry_col3);
    let ret = combine_lo_then_shl(ans23, ans01, frac_nbits);
    println!(
        "combine_lo_then_shl({}, {}, {}) -> ({}, {})",
        ans23, ans01, frac_nbits, ret.0, ret.1
    );
    ret
}

#[inline]
fn hi_lo(this: i128) -> (i128, i128) {
    (this >> 64, this & !(!0 << 64))
}

#[inline]
fn combine_lo_then_shl(this: i128, lo: u128, shift: u32) -> (i128, bool) {
    if shift == 128 {
        (this, false)
    } else if shift == 0 {
        let ans = lo as i128;
        (ans, this != if ans < 0 { -1 } else { 0 })
    } else {
        let lo = (lo >> shift) as i128;
        let hi = this << (128 - shift);
        let ans = lo | hi;
        (ans, this >> shift != if ans < 0 { -1 } else { 0 })
    }
}

#[inline]
fn carrying_add(this: i128, rhs: i128) -> (i128, i128) {
    let (sum, overflow) = this.overflowing_add(rhs);
    let carry = if overflow {
        if sum < 0 {
            1
        } else {
            -1
        }
    } else {
        0
    };
    (sum, carry)
}

#[inline]
fn shift_lo_up(this: i128) -> i128 {
    debug_assert!(this >> 64 == 0);
    this << 64
}

#[inline]
fn shift_lo_up_unsigned(this: i128) -> u128 {
    debug_assert!(this >> 64 == 0);
    (this << 64) as u128
}

The test passes for 1.46.0 i686 stable release builds, for i686 beta/nightly debug builds, and for x86_64 beta/nightly release builds, but fails for i686 beta/nightly release builds.

CI logs from this code on GitLab:

https://gitlab.com/tspiteri/fixed/-/pipelines/183281142

[tspiteri]

The println! prints combine_lo_then_shl(0, 1536, 4) -> (1536, true), but combine_lo_then_shl(0, 1536, 4) should return (96, false).

[LeSeulArtichaut]

Let's try to find the culprit for this regression. Trying to reduce the example would be nice too.
@rustbot ping cleanup

[rustbot]

Hey Cleanup Crew ICE-breakers! This bug has been identified as a good
"Cleanup ICE-breaking candidate". In case it's useful, here are some
instructions for tackling these sorts of bugs. Maybe take a look?
Thanks! <3

cc @AminArria @camelid @chrissimpkins @contrun @DutchGhost @elshize @ethanboxx @h-michael @HallerPatrick @hdhoang @hellow554 @imtsuki @kanru @KarlK90 @LeSeulArtichaut @MAdrianMattocks @matheus-consoli @mental32 @nmccarty @Noah-Kennedy @pard68 @PeytonT @pierreN @Redblueflame @RobbieClarken @RobertoSnap @robjtede @SarthakSingh31 @senden9 @shekohex @sinato @spastorino @turboladen @woshilapin @yerke

[Dylan-DPC-zz]

Marking this as P-critical based on the wg-prioritization discussion

[tspiteri]

I have a reduced test case.

/// ```rust
/// assert_eq!(rust76042::foo(), 8);
/// ```
#[inline]
pub fn foo() -> i128 {
    bar().0
}

#[inline]
fn bar() -> (i128, bool) {
    let a = 0;
    let b = 128;
    let shift = 4;
    let ret = baz(a, b, shift);
    // should print "baz(0, 128, 4) -> (8, false)"
    println!("baz({}, {}, {}) -> ({}, {})", a, b, shift, ret.0, ret.1);
    ret
}

#[inline]
fn baz(a: i128, b: i128, shift: u32) -> (i128, bool) {
    if shift == 128 {
        (a, false)
    } else {
        (b >> shift, a >> shift != 0)
    }
}

Note that if I move the doc test to a lib test, I cannot trigger this bug. On i686 beta/nightly release test, the doc test fails with:

---- src/lib.rs - foo (line 1) stdout ----
Test executable failed (exit code 101).
stdout:
baz(0, 128, 4) -> (128, true)
stderr:
thread 'main' panicked at 'assertion failed: `(left == right)`
  left: `128`,
 right: `8`', src/lib.rs:4:1
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

baz is somehow returning (128, true) instead of (8, false).

https://gitlab.com/tspiteri/rust76042/-/pipelines/183379486

[tmiasko]

Minimized:

fn foo(a: i128, b: i128, s: u32) -> (i128, i128) {
    if s == 128 {
        (0, 0)
    } else {
        (b >> s, a >> s)
    }
}
fn main() {
    let r = foo(0, 8, 1);
    if r.0 != 4 {
        panic!();
    }
}

rustc --crate-type=bin  a.rs --target i686-unknown-linux-gnu -Coverflow-checks=off -Ccodegen-units=1 -Copt-level=1
./a
rustc --crate-type=bin  a.rs --target i686-unknown-linux-gnu -Coverflow-checks=off -Ccodegen-units=1 -Copt-level=0
./a
thread 'main' panicked at 'explicit panic', a.rs:11:9
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

[tmiasko]

LLVM bug report https://bugs.llvm.org/show_bug.cgi?id=47278.

EDIT:

• Introduced by upgrade to LLVM 11.
• Triggers assert in fast register allocator used at opt-level=0:

Unexpected reg unit state
UNREACHABLE executed at /rustsrc/src/llvm-project/llvm/lib/CodeGen/RegAllocFast.cpp:485!

[tspiteri]

LLVM bug 47278 has apparently been fixed in the LLVM tree and its 11.x branch. (I don't really know LLVM, I'm only mentioning this as a comment above mentioned that bug as a possible cause of this issue.)

[tspiteri]

I cannot reproduce the failure with rustc 1.48.0-nightly (8b40853 2020-09-23) which includes #77063.

[LeSeulArtichaut]

Letting this issue open to track a beta backport of #77063.
Also, it might be good to have add a test in the suite for this issue. @tspiteri Would you mind filing a PR with your reproduction? Otherwise I can do it myself.

[tspiteri]

@LeSeulArtichaut I don't know how to write this kind of test PR, please go ahead.

[tspiteri]

Now the beta is working fine as well for me.