Stream: git-wasmtime

Topic: wasmtime / issue #4463 Add a `frem` instruction to cranelift

view this post on Zulip Wasmtime GitHub notifications bot (Jul 17 2022 at 14:06):

afonso360 opened issue #4463:

Feature

We are missing a frem x, y instruction providing the floating-point remainder of x/y.

Benefit

This is a strange omission to our instruction set. We have srem and urem for integers and the rest of the commonly available floating point operations but no remainder.

Some of our users are already lowering it to libcall's themselves:

Implementation

Currently the idea would be to implement this as a libcall on all 3 backends, this seems to be what llvm does (on x86 and aarch64 at least) and should be fairly easy to implement.

I'd be available to work on this.

Alternatives

We can ignore this and let implementers lower it to a libcall themselves, which is what is already happening in practice.

view this post on Zulip Wasmtime GitHub notifications bot (Jul 17 2022 at 15:37):

sunfishcode commented on issue #4463:

I can give some background on why frem was not originally included.

"Remainder" isn't a natural operation for floating-point values. With integers, division rounds to the nearest integer toward zero, so there is an inherent quantity which is left over from division, which we call the remainder. In floating point, division rounds to the nearest representable value in the floating point type, so there is no representable quantity which is left over from division.

Consequently, it's somewhat ambiguous what "remainder" for floating point even means. There isn't a single common definition in common use. The thing LLVM calls "frem" is ends up being different from the IEEE 754 remainder operation and the C remainder function. It does have uses, but they're uncommon in practice, and far less common than operations like sin or cos, for example. It isn't present in any modern hardware. It will probably always be implemented as a library call, and it will always have subtle behavior.

I witnessed the existence of frem in LLVM contribute to a high-level source language deciding to support % on floating-point values in the language, and considered that to have been a mistake.

view this post on Zulip Wasmtime GitHub notifications bot (Jul 17 2022 at 16:15):

afonso360 commented on issue #4463:

Thanks for clarifying! I didn't know there was such a disagreement about the implementation of the remainder.

view this post on Zulip Wasmtime GitHub notifications bot (Jul 17 2022 at 16:15):

afonso360 closed issue #4463:

Feature

We are missing a frem x, y instruction providing the floating-point remainder of x/y.

Benefit

This is a strange omission to our instruction set. We have srem and urem for integers and the rest of the commonly available floating point operations but no remainder.

Some of our users are already lowering it to libcall's themselves:

Implementation

Currently the idea would be to implement this as a libcall on all 3 backends, this seems to be what llvm does (on x86 and aarch64 at least) and should be fairly easy to implement.

I'd be available to work on this.

Alternatives

We can ignore this and let implementers lower it to a libcall themselves, which is what is already happening in practice.

Last updated: Oct 23 2024 at 20:03 UTC