Bring back div_rem

[alexcrichton]

Tracking issue for #850

[alexcrichton]

cc @quantheory

[Boscop]

Any update on this? :)

[vi]

Shall it avoid giving negative remainders unlike %?

[Lokathor]

When you divide -1234 by 10, your remainder is -4.

I think the advantage of div_rem is that the computer is already doing both at once, it's just throwing away part of the work instead of instead of letting us keep both parts.

[dignifiedquire]

What exactly would it need to push this forward?

[Centril]

@dignifiedquire I'd just write a PR against https://github.com/rust-lang/rust/ and then T-libs can decide what to do.

[Lokathor]

Make sure to spec it as const from the start (for integers at least)

Edit: oh crap no const-if yet, nevermind i guess

[AaronKutch]

I want to add a set of uX::wrapping_div_rem(lhs, rhs) -> (uX, uX) functions as a way to guarantee that __udivmodti4 and the like are called, and not have to rely on the compiler backend to combine operations. However, someone encountered an LLVM error where the target sparc-unknown-linux-gnu cannot lower any function that returns a tuple of two u128s. If the function were added, wouldn't that completely break 32-bit SPARC targets unless the LLVM issue is resolved?

[programmerjake]

I'd imagine rustc could add a workaround to do the return lowering in rustc if the LLVM bug is not easy to fix or we want to support old LLVM versions.

[AaronKutch]

Is there someone who wants to tackle the problem? I might try myself, although I have no idea where to begin. I've never messed with code lowering before.

[programmerjake]

It seems to work fine from clang: https://gcc.godbolt.org/z/8xbrWE

[parraman]

I don't know if it's useful, but I have compiled the following code and generated the corresponding LLVM IR:

#[no_mangle]
fn test_128(one: u128, two: u128) -> (u128, u128) {
    let division = one / two;
    (division + 40000000, division - 100000)
}

The LLVM IR is:

; Function Attrs: minsize nounwind optsize
define { i128, i128 } @test_128(i128 %one, i128 %two) unnamed_addr #0 {
start:
  %_6 = icmp eq i128 %two, 0
  br i1 %_6, label %panic, label %bb1, !prof !1, !misexpect !2

bb1:                                              ; preds = %start
  %division = udiv i128 %one, %two
  %_7 = add i128 %division, 40000000
  %_9 = add i128 %division, -100000
  %0 = insertvalue { i128, i128 } undef, i128 %_7, 0
  %1 = insertvalue { i128, i128 } %0, i128 %_9, 1
  ret { i128, i128 } %1

NOTE: I have removed the panic block since I believe it shouldn't be what is causing the error.

The error output generated by rustc is:

LLVM ERROR: unable to allocate function return #6

I have generated the LLVM IR with the flag --emit=llvm-ir. The equivalent function in LLVM IR generated from clang is here.

Thanks a lot!

[NCGThompson]

I would like to point out that div_rem methods may have value even without intrinsic functions to implement them.

• Most importantly, the methods might just help with convenience or readability.
• The standard library is always compiled with opt-level=3, so the optimization will exist even in debug mode.
• There is a little bit of room for optimization outside of the intrinsics. Ex: wrapping_div_rem_euclid()

Since the compiler doesn't recognize the hard coded div function, ethnum-rs has a set of div_rem methods analogous to the integer methods in the standard library. I suggest taking a look at the API and documentation as an example of what the div_rem methods could look like. The implementing code itself isn't as applicable because there were different intrinsics to work with.