`-Zmir-opt-level` miscompiles code with panic

[shao-hua-li]

I tried this code:

pub fn myfunc() -> i32 {
    let mut c :i32 = 1;
    c = 1 ;
    if !c != 0 {
        return 1;
    }
    panic!("Reached end of non-void function");
}

pub fn main() {
    let e = myfunc();
    println!("e={}", e);
}

I expected to see this happen: always panic

Instead, this happened: only -Zmir-opt-level=2 panic

For mir-opt-level 0, 1 and 3, they all normally return

% rustc -Awarnings -Zmir-opt-level=2 test.rs && ./test
thread 'main' panicked at test.rs:7:5:
Reached end of non-void function
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
%
% rustc -Awarnings -Zmir-opt-level=0 test.rs && ./test
e=1

Meta

rustc --version --verbose:

rustc 1.83.0-nightly (18b1161ec 2024-10-02)
binary: rustc
commit-hash: 18b1161ec9eeab8927f91405bca0ddf59a4a26c9
commit-date: 2024-10-02
host: x86_64-unknown-linux-gnu
release: 1.83.0-nightly
LLVM version: 19.1.0

[clubby789]

!1 equals -2, so I think the panic case is incorrect?

[shao-hua-li]

!1 equals -2, so I think the panic case is incorrect?

Yeah, I think so.

[clubby789]

@@ -1,4 +1,4 @@
-// MIR for `myfunc` before JumpThreading
+// MIR for `myfunc` after JumpThreading
 
 fn myfunc() -> i32 {
     let mut _0: i32;
@@ -17,7 +17,7 @@ fn myfunc() -> i32 {
         _4 = copy _1;
         _3 = Not(move _4);
         _2 = Ne(move _3, const 0_i32);
-        switchInt(move _2) -> [0: bb2, otherwise: bb1];
+        goto -> bb2;
     }
 
     bb1: {

Looks like it's JumpThreading that incorrectly removes the condition

[shao-hua-li]

I managed to create an example without panic, just wrong value

pub fn myfunc() -> i32 {
    let mut a: i32 = 0;
    a = 1;
    if !a != 0 {
        return 1
    }
    return 0;
}

pub fn main() {
    let mut e = myfunc();
    println!("e={}", e);
}

% rustc -Awarnings test.rs -Zmir-opt-level=0 && ./test
e=1
% rustc -Awarnings test.rs -Zmir-opt-level=2 && ./test
e=0

[clubby789]

found 8 bors merge commits in the specified range
  commit[0] 2024-02-10: Auto merge of #119614 - RalfJung:const-refs-to-static, r=oli-obk
  commit[1] 2024-02-10: Auto merge of #117206 - cjgillot:jump-threading-default, r=tmiasko
  commit[2] 2024-02-11: Auto merge of #120232 - c272:json-buildstd, r=Mark-Simulacrum
  commit[3] 2024-02-11: Auto merge of #120405 - cjgillot:gvn-pointer, r=oli-obk
  commit[4] 2024-02-11: Auto merge of #120914 - ehuss:downgrade-xcode, r=Mark-Simulacrum
  commit[5] 2024-02-11: Auto merge of #120920 - matthiaskrgr:rollup-jsryr8e, r=matthiaskrgr
  commit[6] 2024-02-11: Auto merge of #120903 - matthiaskrgr:rollup-tmsuzth, r=matthiaskrgr
  commit[7] 2024-02-11: Auto merge of #120803 - onur-ozkan:fix-compilation-cache, r=Mark-Simulacrum

Most likely #117206

[compiler-errors]

ha, I found the bug. We're optimizing UnOp::Not incorrectly in the jump threading code. Right now it's only written to handle bool correctly. I'll put up a fix 🤔

@rustbot claim