Inefficient code generation of `@mod` and `@divFloor` about zig HOT 4 OPEN

i dont think these transformations are valid when the number is negative. depending on what youre doing and what you can guarantee, using @divExact or adding if (num < 0) unreachable; will give the codegen youre asking for. in any case unless you see an inefficiency in the LLVM IR zig emits, optimizations like this are generally on LLVM not us

for comparison heres the equivalent in C https://godbolt.org/z/EnP6MEaP1 in all cases its identical when you use the builtins that map to what C does

from zig.

that's not quite the same, these transformations work for negative numerators as well (note that it's using sar instead of shr), I can't speak for the LLVM IR as I'm not as well versed in it. In any case, I'm not sure we should expect LLVM to be able to find a transformation like that.

from zig.

hm yeah my bad that looks right. i dont know where the logic in zig is for this stuff or if it generally attempts any kind of optimization on its own.

from zig.

In case it helps, I uploaded the code I was writing when I found this to a repo, which has 2 branches, one uses the builtins and the other uses the alternatives mentioned here.

from zig.