Unchecked div/rem no longer generic

[strega-nil]

Now they are [ui](8|16|32|64)_unchecked_(div|rem)

This is more in line with the other mathematical intrinsics, and means you can't, for example, call unchecked_sdiv((),()) anymore.

[rust-highfive]

r? @aturon

(rust_highfive has picked a reviewer for you, use r? to override)

[strega-nil]

I'd also like to change overflowing_(add|sub|mul) to be non-generic if possible.

[nagisa]

Rather than this approach, I’d prefer intrinsics checked their specialisation instead. Similarly to the SSE intrinsics code. See this which results in E511.

[alexcrichton]

I would also kinda tend to err on the side of ease of implementation of intrinsics, but @GBGamer I'm curious was there a use case other than bad monomorphizations which motivated this?

[strega-nil]

@alexcrichton because it's really ugly. unchecked_udiv<T>(T, T) -> T says "Hey, I'm valid for all T, so pass anything to me." when really, it's exclusively for the integer types. It's not meant for anything else.

[huonw]

Intrinsics are regarded as things that should essentially never be called directly: just once for each type, to build up safe abstractions. I.e. inaccuracies in their type signatures aren't thought to be the end of the world.

[strega-nil]

@huonw: that may be true, that doesn't mean it doesn't bother me :P.

However, I think that switching to the simd_add<T> version would be good. Then we don't have to worry about an LLVM error popping up in rust code, and there's no bloat. Actually, this might be good for all of the other intrinsics that are currently T_function, like u8_mul_with_overflow and friends.

[strega-nil]

@huonw The issue I have is this:

#![feature(core_intrinsics)]
fn main() {
    unsafe { std::intrinsics::unchecked_sdiv((),()); }
}

ends up spitting out

/Constants.cpp:1898: static llvm::Constant* llvm::ConstantExpr::get(unsigned int, llvm::Constant*, llvm::Constant*, unsigned int, llvm::Type*): Assertion `C1->getType()->isIntOrIntVectorTy() && "Tried to create an arithmetic operation on a non-arithmetic type!"' failed.
Aborted (core dumped)

and also that you can call sdiv on an unsigned integer, and udiv on a signed integer.

[huonw]

Sure, I'm all for "last minute" (i.e. like simd_add) checks to avoid LLVM asserts like that.

[strega-nil]

I've an idea. One sec, let me try something.