Pattern match incorrectly optimized out (bad range info handling in LLVM?)

[rprichard]

#[derive(Copy, Clone, Eq, PartialEq, Debug)]
#[allow(dead_code)]
enum BarMode {
    Bar1,
    Bar2,
}

#[derive(Copy, Clone, Eq, PartialEq, Debug)]
#[allow(dead_code)]
enum Mode {
    Foo(u8),
    Bar(BarMode),
}

#[inline(never)]
fn broken(mode: &Mode) -> u32 {
    for _ in 0..1 {
        match *mode {
            Mode::Bar(BarMode::Bar1) => { return 17 }
            Mode::Foo(5) => { return 19 }
            _ => {}
        }
    }
    return 42;
}

fn main() {
    let mode = Mode::Foo(5);
    assert_eq!(broken(&mode), 19);
}

This test program should exit quietly, but with rustc -O, it instead fails the assertion:

thread '<main>' panicked at 'assertion failed: `(left == right) && (right == left)` (left: `42`, right: `19`)', redux.rs:29

rprichard@ryan:~/mess/redux$ rustc --version --verbose
rustc 1.2.0-nightly (20d23d8e5 2015-06-18)
binary: rustc
commit-hash: 20d23d8e57c0313c811135535d6872463cc6525d
commit-date: 2015-06-18
host: x86_64-unknown-linux-gnu
release: 1.2.0-nightly

I also tested a really old 32-bit compiler, and it also failed:

rprichard@ryan:~/mess/redux$ rustn32 --version --verbose
rustc 1.0.0-nightly (c89de2c56 2015-03-28) (built 2015-03-29)
binary: rustc
commit-hash: c89de2c56baeb61e7cc434924dcc8bedd32b26b8
commit-date: 2015-03-28
build-date: 2015-03-29
host: i686-unknown-linux-gnu
release: 1.0.0-nightly

This test case is reduced from #25919.

[Stebalien]

Probably a dup of #26251 (and #18060)

[rprichard]

Perhaps it's a dup. The other two issues are broken with and without optimizations, though. This issue only fails with optimizations.

[rprichard]

Also: the match expression in this issue doesn't actually require any overlap of the match patterns. (I'll update the test case.)

[rprichard]

The bug also happens without a match expression:

#[inline(never)]
fn broken(mode: &Mode) -> u32 {
    for _ in 0..1 {
        if let Mode::Bar(BarMode::Bar1) = *mode { return 17 }
        if let Mode::Foo(5) = *mode { return 19 }
    }
    return 42;
}

[Stebalien]

Good point. Also, this only happens with opt-level=2 (not 0 or 3).

[rprichard]

cc @dotdash

Further refinement of the test case:

#[allow(dead_code)]
#[repr(u16)]
enum FooMode {
    First = 0x1000,
    Check = 0x1001,
    Last  = 0x100f,
}

#[allow(dead_code)]
#[repr(u16)]
enum BarMode {
    First = 0x2000,
    Check = 0x2001,
    Last  = 0x200f,
}

#[allow(dead_code)]
enum Mode {
    Foo(FooMode),
    Bar(BarMode),
}

#[inline(never)]
fn broken(mode: &Mode) -> u32 {
    for _ in 0..1 {
        if let Mode::Foo(FooMode::Check) = *mode { return 17 }
        if let Mode::Bar(BarMode::Check) = *mode { return 19 }
    }
    return 42;
}

fn main() {
    let mode = Mode::Bar(BarMode::Check);
    assert_eq!(broken(&mode), 19);
}

Changing FooMode::Last from 0x100f to a large enough value avoids the bug. Apparently, the value needs to be large enough that, when rounded to the next highest power-of-two-minus-one, it is at least as large as BarMode::Check.

I think this is an LLVM bug. It involves the !range metadata. It's as if the compiler is reusing the range information from the Mode::Foo(FooMode::Check) check to elide the Mode::Bar(BarMode::Check) check (which is missing from the x64 assembly). Relevant unoptimized LLVM IR:

match_case7:                                      ; preds = %case_body2
  %26 = bitcast %Mode* %21 to { i16, i16 }*
  %27 = getelementptr inbounds { i16, i16 }, { i16, i16 }* %26, i32 0, i32 1
  %28 = load i16, i16* %27, !range !3
  switch i16 %28, label %match_else8 [
    i16 4097, label %match_case9
  ]

...

match_case14:                                     ; preds = %join
  %32 = bitcast %Mode* %29 to { i16, i16 }*
  %33 = getelementptr inbounds { i16, i16 }, { i16, i16 }* %32, i32 0, i32 1
  %34 = load i16, i16* %33, !range !4
  switch i16 %34, label %match_else15 [
    i16 8193, label %match_case16
  ]

...

!3 = !{i16 4096, i16 4112}
!4 = !{i16 8192, i16 8208}

FooMode::Last isn't used in the test program. When I change its value, the only difference I notice in the unoptimized LLVM IR is in the range metadata (i.e. !3).

[dotdash]

Thanks Ryan! Indeed LLVM replaces the load with another disregarding the range information.

[dotdash]

Patch is at http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20150622/282940.html

[dotdash]

Patch landed upstream.

[MagaTailor]

Using yesterday's nightly, make test fails on armv7 at this single test. The problem seems to be local llvm 3.6 is still allowed by the build system.