Add Overview

[rossberg]

With some delay, this is a proposal repo for language profiles, as discussed at the 2021-03-02 meeting.

[rossberg]

With some delay, this is a proposal repo for language profiles, as discussed at the 2021-03-02 meeting.

[sunfishcode]

Overall, this looks like a useful direction, thanks for putting this together!

One question I have is about whether modules should have a way of depending on a particular profile. For example, some floating-point algorithms depend on a true fma, and the nondeterministic qfma proposed in the relaxed-simd proposal would not be reliable. The deterministic profile might restrict qfma to deterministically behave like true fma, however would applications that need a true fma have a way of depending on that? Or, should we propose a separate true-fma instruction, so that applications that need to depend on it can do so?

[rossberg]

@sunfishcode:

One question I have is about whether modules should have a way of depending on a particular profile. For example, some floating-point algorithms depend on a true fma, and the nondeterministic qfma proposed in the relaxed-simd proposal would not be reliable. The deterministic profile might restrict qfma to deterministically behave like true fma, however would applications that need a true fma have a way of depending on that? Or, should we propose a separate true-fma instruction, so that applications that need to depend on it can do so?

Hm, interesting question. I think the latter.

Generally, a producer should not (have to) depend on certain behaviours being absent in a profile. They should view profiles as possible restrictions to cope with, not as guarantees to depend on. Although we cannot enforce this view, we should still enable it.

But I believe that this question can actually be answered independently of profiles. Profiles are chosen globally by an ecosystem, not by an individual producer. If producers have, say, a programmatic need for true fma as you describe, then why would that need only arise in deterministic ecosystems? That's unlikely, so any solution should be applicable in an unrestricted environment with "full" profile as well. That suggests that the full language should have two different instructions.

[lukewagner]

Overall, looks great; this seems valuable.