reached unreachable code compiling zig

[raquentin]

Zig Version

0.14.0

Steps to Reproduce and Observed Behavior

1. Be on aarch64 linux. sudo dnf install clang-devel llvm-devel lld-devel zlib zlib-devel.
2. clone zig, cd
3. clear cache: rm -rf build; mkdir build; rm -rf ~/.cache/zig; rm -rf zig-cache; git pull
4. build cd build; cmake .. -DCMAKE_PREFIX_PATH=$(where llvm); make install

Output:

thread 6774 panic: reached unreachable code
/home/<...>/zig/lib/std/posix.zig:4686:23: 0x150b1db in mprotect (build)
            .INVAL => unreachable,
                      ^
/home/<...>/zig/lib/std/Thread.zig:1424:23: 0x153651f in spawn__anon_56775 (build)
        posix.mprotect(
                      ^
/home/<...>/zig/lib/std/Thread.zig:419:32: 0x14eb1f7 in spawn__anon_8489 (build)
    const impl = try Impl.spawn(config, function, args);
                               ^
/home/<...>/zig/lib/std/Progress.zig:423:55: 0x14d7c53 in start (build)
                .ansi_escape_codes => std.Thread.spawn(.{}, updateThreadRun, .{}),
                                                      ^
/home/<...>/zig/lib/compiler/build_runner.zig:329:50: 0x14d27f3 in main (build)
    const main_progress_node = std.Progress.start(.{
                                                 ^
/home/<...>/zig/lib/std/start.zig:656:37: 0x14beab7 in posixCallMainAndExit (build)
            const result = root.main() catch |err| {
                                    ^
???:?:?: 0x0 in ??? (???)
error: the following build command crashed:
/home/<...>/zig/.zig-cache/o/edacf39d414fa60bf436caccbad656c9/build /home/<...>/zig/build/zig2 /home/<...>/zig/lib /home/<...>/zig /home/<...>/zig/.zig-cache /home/<user>/.cache/zig --seed 0x64a0fe9c -Z94e65736300eab91 --prefix /home/<...>/zig/build/stage3 -Dversion-string=0.14.0-dev.2643+fb43e91b2 -Dtarget=native -Dcpu=native -Denable-llvm -Dconfig_h=/home/<...>/zig/build/config.h -Dno-langref -Doptimize=Debug
make[2]: *** [CMakeFiles/stage3.dir/build.make:73: stage3/bin/zig] Error 1
make[1]: *** [CMakeFiles/Makefile2:196: CMakeFiles/stage3.dir/all] Error 2
make: *** [Makefile:136: all] Error 2

Expected Behavior

Zig builds

[raquentin]

Here's the function panicking for convenience. In `lib/std/posix.zig:4666':

pub fn mprotect(memory: []align(mem.page_size) u8, protection: u32) MProtectError!void {
    if (native_os == .windows) {
        const win_prot: windows.DWORD = switch (@as(u3, @truncate(protection))) {
            0b000 => windows.PAGE_NOACCESS,
            0b001 => windows.PAGE_READONLY,
            0b010 => unreachable, // +w -r not allowed
            0b011 => windows.PAGE_READWRITE,
            0b100 => windows.PAGE_EXECUTE,
            0b101 => windows.PAGE_EXECUTE_READ,
            0b110 => unreachable, // +w -r not allowed
            0b111 => windows.PAGE_EXECUTE_READWRITE,
        };
        var old: windows.DWORD = undefined;
        windows.VirtualProtect(memory.ptr, memory.len, win_prot, &old) catch |err| switch (err) {
            error.InvalidAddress => return error.AccessDenied,
            error.Unexpected => return error.Unexpected,
        };
    } else {
        switch (errno(system.mprotect(memory.ptr, memory.len, protection))) {
            .SUCCESS => return,
            .INVAL => unreachable, // panic here
            .ACCES => return error.AccessDenied,
            .NOMEM => return error.OutOfMemory,
            else => |err| return unexpectedErrno(err),
        }
    }
}

[raquentin]

Thanks to anyone checking this out been stumping me for a bit now

[rootbeer]

The compile is failing because mprotect is returning EINVAL and the standard library assumes EINVAL cannot happen. The mprotect man page says EINVAL might be returned if the pointer is invalid or not a multiple of the system page size (see https://linux.die.net/man/2/mprotect). Since this mprotect is being invoked from Zig's thread setup, I'm going to guess the address isn't aligned to the system page size. I believe there is ongoing work to make Zig work with non-4k page systems (e.g., I think #16331) might be related. What does getconfig PAGE_SIZE say on your system?

[raquentin]

Thanks a lot. getconf PAGE_SIZE prints 16384. Looking more into it, my kernel assumes 16k page size despite the hardware technically supporting both at times.

[alexrp]

In that case, I'll close this as a duplicate of #16331. There's a workaround in that issue that you can use until #20511 is merged.