Branches returning errors are not weighted against

[dweiller]

Zig Version

0.15.0-dev.936+fc2c1883b

Steps to Reproduce and Observed Behavior

copy this to repro.zig:

export fn foo(a: usize, b: usize) usize {
    return do_thing(a, b) catch 0;
}

fn do_thing(a: usize, b: usize) !usize {
    if (b == 0) {
        //@branchHint(.unlikely);
        return error.bad;
    }
    return a + b;
}

and run zig build-obj -OReleaseFast -fno-emit-bin --verbose-llvm-ir=repro.ll and then grep repro.ll for branch_weights to observe that there is no branch weight metadata. If you uncomment the @branchHint(.unlikely) there will be branch weight metadata.

Expected Behavior

I expect the above, with or without the @branchHint, to weight against the branch returning an error. Currently it seems that only branches introduced by try/catch are implicitly weighted, but this means that only points that errors propagate to are weighted against by default, while the origin of an error isn't.

I'd also expect that if every execution path in a branch (containing sub-branches) ends up returning an error, then that branch should be weighted against (I'm not sure what llvm will do if you just label every error-returning sub-branch with an .unlikely).

[mlugg]

Note that implementing this in the compiler will require care not to propagate too far. For instance:

if (foo) {
    if (bar) {
        return 123;
    }
    return error.Bad;
}

The if (foo) branch should not be weighted against, even though the body ends with an error return. Ideally, this would actually weight in favour of the if (bar) "then" branch. It's possible it'll turn out easier to do this error-weighting analysis during AstGen rather than Sema where it currently lives (in which case the existing error-weighting analysis in Sema would probably also be moved there).