Incremental compilation fails for valid source when correctly widening the return error set

[Zirunis]

Zig Version

0.16.0-dev.1246+4b593a6c2

Steps to Reproduce and Observed Behavior

Wrong code:

const std = @import("std");

pub fn main() void {  //no error set declared here!
    try std.testing.expect(2 + 2 == 5);
}

zig build --watch -fincremental correctly points out the error:

src\main.zig:4:5: error: expected type 'void', found '@typeInfo(@typeInfo(@TypeOf(testing.expect)).@"fn".return_type.?).error_union.error_set'
    try std.testing.expect(2 + 2 == 5);
    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
src\main.zig:3:15: note: function cannot return an error
pub fn main() void {
              ^~~~

Fix the code:

const std = @import("std");

pub fn main() !void {  //now correct
    try std.testing.expect(2 + 2 == 5);
}

zig build --watch -fincremental incorrectly throws an error despite the source being valid:

error: Found return instr that returns non-void in Function of void return type!
         ret i16 %1, !dbg !50393
        voidFound return instr that returns non-void in Function of void return type!
         ret i16 0, !dbg !50393
        voidLLVM ERROR: Broken module found, compilation aborted!

error: process exited with error code 9

Expected Behavior

Incremental compilation should lead to an equivalent result as running zig build again. In this case that means building successfully.

Related: #25874

[mlugg]

Given the error message, it sounds like you are using the LLVM backend, which does not currently support incremental compilation. It looks like you're on Windows, where that backend is still the default due to deficiencies in the self-hosted COFF linker. You can try setting .use_llvm = false in your build script to override this, though you might then run into those linker bugs.

I'm closing #25872 for the same reason; the emitted binary is not expected to be correct when performing an incremental update using the LLVM backend.

[Zirunis]

Oh no! I thought opting into incremental compilation would automatically use the self-hosted backend (although I should've clocked that that's not the case given the llvm error). Maybe that would be a good idea? Or to at least emit a warning recommending to opt out of LLVM when using incremental compilation? Would've saved us both time @mlugg
Thanks for the hint and apologies for the misleading report!

[mlugg]

No problem! Thanks for reporting nonetheless. I think it would probably make sense for the compiler to spit out a warning when combining -fincremental with the LLVM backend; I'll PR that shortly.

[Zirunis]

Nice, thank you!