Make use of getelementptr nuw in pointer::add

[nikic]

add() on pointers should generate getelementptr inbounds nuw on LLVM 20 and above, to let LLVM know that the offset cannot be negative. We'll need a new variant of intrinsic::offset to pass down this information.

[nikic]

cc @scottmcm as this seems like something up your alley :)

[scottmcm]

Shouldn't need a new intrinsic -- we already lower add to MIR Offset (#110837) -- so it's just a matter of updating

rust/compiler/rustc_codegen_ssa/src/mir/rvalue.rs

Lines 958 to 971 in 3b022d8

	mir::BinOp::Offset => {
	let pointee_type = input_ty
	.builtin_deref(true)
	.unwrap_or_else(|| bug!("deref of non-pointer {:?}", input_ty));
	let pointee_layout = bx.cx().layout_of(pointee_type);
	if pointee_layout.is_zst() {
	// `Offset` works in terms of the size of pointee,
	// so offsetting a pointer to ZST is a noop.
	lhs
	} else {
	let llty = bx.cx().backend_type(pointee_layout);
	bx.inbounds_gep(llty, lhs, &[rhs])
	}
	}

to check whether the RHS is unsigned, and add nuw if it is.

[nikic]

Oh nice, that will make things a lot simpler. Maybe I can even implement this myself...

[scottmcm]

If you're tempted, go for it! Should be pretty similar to https://github.com/rust-lang/rust/pull/136575/files#diff-76d7dff532dc6719661f41c02777d2b1f8888cd0b8113b9908ed0085ef49fff4 , just with less confusion along the way since you actually know the LLVM API (unlike me :P)

If not I'll get to it eventually, just not sure when that'll be. I'm excited for it, since I tried doing it with assumes in #123598 but perf wasn't happy with me.