Release build emit assertion `(KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"'

[kamyuentse]

I'm a user of the rust-zookeeper, and I found it build failed in release mode.
using the command: cargo build --release --verbose

log：https://gist.github.com/kamyuentse/1bf0c8d65fd40a17060f5cca515e720c

It seems an upstream bug related the LLVM? There is any workaround to fix this bug?

bonifaido/rust-zookeeper#37

Environment:
rustc: rustc 1.22.0-nightly (0e6f4cf51 2017-09-27)
OS: CentOS Linux release 7.4.1708
Kernel: Linux 4.13.3-1.el7.elrepo.x86_64

[jason-wolfe]

I just encountered the same bug using the same library, using cargo build --release.

rustc: rustc 1.22.0-nightly (dd08c3070 2017-09-12)
toolchain: nightly-x86_64-unknown-linux-musl and nightly-x86_64-unknown-linux-gnu both trigger it.
OS: Ubuntu 14.04.5 LTS
Kernel: Linux 3.19.0-80-generic x86_64

Builds fine in debug mode.

[jonas-schievink]

Can repro on Arch with Rust nightlies from rustup, so not a packaging issue. Only one LLVM bug mentions the message: https://bugs.llvm.org/show_bug.cgi?id=24241

[jason-wolfe]

I don't know if this is entirely relevant and don't know how to debug rustc, but I tried running with RUSTFLAGS="-Zverbose" cargo build --release --verbose, which ended up resulting in

   Compiling zookeeper_derive v0.3.0 (file:///home/jwolfe/Rust/rust-zookeeper/zookeeper-derive)
     Running `rustc --crate-name zookeeper_derive zookeeper-derive/src/lib.rs --crate-type proc-macro --emit=dep-info,link -C prefer-dynamic -C opt-level=3 -C metadata=3d948e35b92c7695 -C extra-filename=-3d948e35b92c7695 --out-dir /home/jwolfe/Rust/rust-zookeeper/target/release/deps -L dependency=/home/jwolfe/Rust/rust-zookeeper/target/release/deps --extern syn=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libsyn-b00a444c8aa4acf3.rlib --extern quote=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libquote-7cd2e6e51ec0ccc5.rlib -Zverbose`
   Compiling zookeeper v0.3.0 (file:///home/jwolfe/Rust/rust-zookeeper)
     Running `rustc --crate-name zookeeper src/lib.rs --crate-type lib --emit=dep-info,link -C opt-level=3 -C metadata=b5a475f1c342e7c6 -C extra-filename=-b5a475f1c342e7c6 --out-dir /home/jwolfe/Rust/rust-zookeeper/target/release/deps -L dependency=/home/jwolfe/Rust/rust-zookeeper/target/release/deps --extern snowflake=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libsnowflake-e9a06768725f9ce6.rlib --extern bytes=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libbytes-896568c95020390b.rlib --extern byteorder=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libbyteorder-549a4dbae79b4adf.rlib --extern zookeeper_derive=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libzookeeper_derive-3d948e35b92c7695.so --extern log=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/liblog-03d566855f19b793.rlib --extern mio=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/libmio-c1cb259295b5b695.rlib --extern lazy_static=/home/jwolfe/Rust/rust-zookeeper/target/release/deps/liblazy_static-a9194fff651d7463.rlib -Zverbose`
error: /home/jwolfe/Rust/rust-zookeeper/target/release/deps/libzookeeper_derive-3d948e35b92c7695.so: undefined symbol: _ZN116_$LT$alloc..string..String$u20$as$u20$core..convert..From$LT$$RF$ReEarlyBound$LP$0$C$$u20$$u27$a$RP$$u20$str$GT$$GT$4from17h3156f71d21fd9c11E
  --> src/lib.rs:11:1
   |
11 | extern crate zookeeper_derive;
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

error: Could not compile `zookeeper`.

I was surprised to see something entirely different than the error I see without the debug flag, the surprising part to me being

error: /home/jwolfe/Rust/rust-zookeeper/target/release/deps/libzookeeper_derive-3d948e35b92c7695.so: undefined symbol: _ZN116_$LT$alloc..string..String$u20$as$u20$core..convert..From$LT$$RF$ReEarlyBound$LP$0$C$$u20$$u27$a$RP$$u20$str$GT$$GT$4from17h3156f71d21fd9c11E

[fulmicoton]

Here is a minimum code to reproduce.

https://github.com/fulmicoton/bug_known_bits/blob/master/src/lib.rs

To reproduce you can clone the repo and run cargo build --release.

[fulmicoton]

Here is the last pass that makes LLVM unhappy.

*** IR Dump Before Early CSE ***
; Function Attrs: uwtable
define i32 @_ZN3lib16bugging_function17hab1ad2b68c3d5905E(i32) unnamed_addr #2 {
start:
  %1 = icmp ult i32 %0, 3
  call void @llvm.assume(i1 %1)
  %2 = xor i32 %0, -1
  %.off.i = add i32 %2, -1
  %switch.i = icmp ult i32 %.off.i, 2
  br i1 %switch.i, label %_ZN3lib8from_u3217hcbd3743a2503d713E.exit, label %bb3.i

bb3.i:                                            ; preds = %start
  call fastcc void @_ZN3std9panicking11begin_panic17h684cdffa117ea7c2E()
  unreachable

_ZN3lib8from_u3217hcbd3743a2503d713E.exit:        ; preds = %start
  %3 = icmp ult i32 %2, 3
  call void @llvm.assume(i1 %3)
  %4 = or i32 %2, 1
  ret i32 %4
}

LLVM's value checker is correct to infer that some of the bits in the second branch are supposed to be both 0 and 1... The conclusion is however wrong : the branch is simply unreachable.

[fulmicoton]

Same error without any panics! nor any enum.

#![feature(core_intrinsics)]

extern crate core;
use core::intrinsics::assume;

#[inline(always)]
pub fn from_u32(v: u32) -> Option<u32> {
    if v >= 1 && v <= 2u32 {
        Some(v)
    } else {
        None
    }
}

pub fn bugging_function(v: bool) -> u32 {
    let input: u32 = if v { 1 } else { 2 };
    if let Some(nval) =  from_u32(!input) {
        unsafe {
            assume(nval >= 1);
            assume(nval <= 2);
        }
        return nval;
    }
    0u32
}

[arielb1]

This crashes our LLVM with -early-cse set, but doesn't crash any of the debian llvms

; Function Attrs: nounwind
declare void @llvm.assume(i1)

; Function Attrs: nounwind uwtable
define i32 @_ZN3ice16bugging_function17hec6c566c29298c46E(i1 zeroext) unnamed_addr #0 {
start:
  %. = select i1 %0, i32 -2, i32 -3
  %.off.i = add i32 %., -1
  %1 = icmp ult i32 %.off.i, 2
  %_0.sroa.0.0.i = zext i1 %1 to i64
  %_0.sroa.3.0.insert.ext.i = zext i32 %. to i64
  %_0.sroa.3.0.insert.shift.i = shl nuw i64 %_0.sroa.3.0.insert.ext.i, 32
  %_0.sroa.0.0.insert.insert.i = or i64 %_0.sroa.0.0.i, %_0.sroa.3.0.insert.shift.i
  %abi_cast.sroa.4.0.extract.shift = lshr i64 %_0.sroa.0.0.insert.insert.i, 32
  %abi_cast.sroa.4.0.extract.trunc = trunc i64 %abi_cast.sroa.4.0.extract.shift to i32
  %2 = and i64 %_0.sroa.0.0.insert.insert.i, 4294967295
  %cond = icmp eq i64 %2, 1
  br i1 %cond, label %bb6, label %bb9

bb6:                                              ; preds = %start
  %3 = icmp ne i32 %abi_cast.sroa.4.0.extract.trunc, 0
  call void @llvm.assume(i1 %3)
  %4 = icmp ult i32 %abi_cast.sroa.4.0.extract.trunc, 3
  call void @llvm.assume(i1 %4)
  br label %bb9

bb9:                                              ; preds = %start, %bb6
  %_0.0 = phi i32 [ %abi_cast.sroa.4.0.extract.trunc, %bb6 ], [ 0, %start ]
  ret i32 %_0.0
}

[arielb1]

Minified:

; Function Attrs: nounwind
declare void @llvm.assume(i1)

; Function Attrs: nounwind uwtable
define i64 @icey(i1 zeroext) {
start:
  %. = select i1 %0, i64 -1, i64 -2
  %_4 = icmp eq i64 %., 0
  call void @llvm.assume(i1 %_4)

  ret i64 %.
}

[arielb1]

Looks like https://bugs.llvm.org/show_bug.cgi?id=31809 - there's a fix in upstream LLVM, which we can backport.

[nikic]

I believe this can be closed, as the fix landed upstream and our LLVM has since been updated. Also can't repro with #44899 (comment) anymore.

[fulmicoton]

I confirm that this seems fixed.

[kamyuentse]

Thanks @fulmicoton , I am closing this now.