llvm: Use no-builtins attribute instead of nobuiltin.

[alexrp]

The former prevents recognizing code patterns and turning them into libcalls, which is what we want for compiler-rt. The latter is meant to be used on call sites to prevent them from being turned into intrinsics.

Context: #21833

[alexrp]

cc @dweiller

[dweiller]

This sounds great - I'll test it out on some code that llvm would turn into stack overflows.

[alexrp]

Sounds good. Do you feel like doing some testing before I merge?

[dweiller]

Sounds good. Do you feel like doing some testing before I merge?

I'm doing some testing, but I'm struggling to get llvm to make memcpy recursively call itself even without this patch. It could be that the code patterns that I had doing this no longer trigger this behaviour (I originally saw it happen when zig was on llvm 17 I believe).

[alexrp]

Maybe try translating one of the Clang examples for this attribute to Zig? https://clang.llvm.org/docs/AttributeReference.html#no-builtin

[alexrp]

pub fn naive_memset_zero(data: [*]u8, count: usize) callconv(.c) void {
    for (0..count) |i| {
        data[i] = 0;
    }
}

Putting this function in compiler-rt before this PR, its body is replaced with memset(). After this PR, it turns into an auto-vectorized loop instead.