[phire]

> Instructions are more easily added than taken away.

That's not saying much, it's basically impossible to remove an instruction. Just because something is easier than impossible doesn't mean that it's easy.

And sure, from a technical perspective, it's quite easy to add new instructions to RISC-V. Anyone can draft up a spec and implement it in their core.

But if you actually want wide-spread adoption of a new instruction, to the point where compilers can actually emit it by default and expect it to run everywhere, that's really, really hard. First you have to prove that this instruction is worthwhile standardizing, then debate the details and actually agree on a spec. Then you have to repeat the process and argue the extension is worth including in the next RVA profile, which is highly contentious.

Then you have to wait. Not just for the first CPUs to support that profile. You have to wait for every single processor that doesn't support that profile to become irrelevant. It might be over a decade before a compiler can safely switch on that instruction by default.

[brucehoult]

It's not THAT hard. Heck, I've done it myself. But, as I said, the burden of proof that something new is truly useful quite rightly lies with the proposer.

The ORC.B instruction in Zbb was my idea, never done anywhere before as far as anyone has been able to find. I proposed it in late 2019, it was in the ratified spec in later 2021, and implemented in the very popular JH7110 quad core 1.5 GHz SoC in the VisionFive 2 (and many others later on) that was delivered to pre-order customers in Dec 2022 / Jan 2023.

You might say that's a long time, but that's pretty fast in the microprocessor industry -- just over three years from proposal (by an individual member of RISC-V International) to mass-produced hardware.

Compare that to Arm who published the spec for SVE in 2016 and SVE 2 in 2019. The first time you've been able to buy an SBC with SVE was early 2025 with the Radxa Orion O6.

In contrast RISC-V Vector extension (RVV) 1.0 was published in late 2021 and was available on the CanMV-K230 development board in November 2023, just two years later, and in a flood of much more powerful octa-core SpacemiT K1/M1 boards (BPI-F3, Milk-V Jupiter, Sipeed LicheePi 3A, Muse Pi, DC-Roma II laptop) starting around six months later.

[phire]

The question is not so much when the first CPU ships with the instruction, but when the last CPU without it stops being relevant.

It varies from instruction to instruction, but alternative code paths are expensive, and not well supported by compilers, so new instructions tend to go unused (unless you are compiling code with -march=native).

In one way, RISC-V is lucky. It's not that currently widely deployed anywhere, so RVA23 should be picked up as the default target, and anything included in it will have widespread support.

But RVA23 is kind of pulling the door closed after itself. It will probably become the default target that all binary distributions will target for the next decade, and anything that didn't make it into RVA23 will have a hard time gaining adoption.

[brucehoult]

I'm confused. You appear to be against adding new instructions, but also against picking a baseline such as RVA23 and sticking with it for a long time.

Every ISA adds new instructions over time. Exactly the same considerations apply to all of them.

Some Linux distros are still built for original AMD64 spec published in August 2000, while some now require the x86-64-v2 spec defined in 2020 but actually met by CPUs from Nehalem and Jaguar on.

The ARMv8-A ecosystem (other than Apple) seems to have been very reluctant to move past the 8.2 spec published in January 2016, even on the hardware side, and no Linux distro I'm aware of requires anything past original October 2011 ARMv8.0-A spec.

[phire]

I'm not against adding new instructions. I love new instructions, even considered trying to push for a few myself.

What I'm against is the idea that it's easy to add instructions. Or more the idea that it's a good idea to start with the minimum subset of instructions and add them later as needed.

It seems like a good idea; Save yourself some upfront work. Be able to respond to actual real-world needs rather than trying to predict them all in advance. But IMO it just doesn't work in the real world.

The fact that distros get stuck on the older spec is the exact problem that drives me mad, and it's not even their fault. For example, compilers are forced generate some absolute horrid ARMv8.0-A exclusive load/store loops when it comes to atomics, yet there are some excellent atomic instructions right there in ARMv8.1-A, which most ARM SoCs support.

But they can't emit them because that code would then fail on the (substantial) minority of SoCs that are stuck on ARMv8.0-A. So those wonderful instructions end up largely unused on ARMv8 android/linux, simply because they arrived 11 years ago instead of 14 years ago.

At least I can use them on my Mac, or any linux code I compile myself.

-------

There isn't really a solution. Ecosystems getting stuck on increasingly outdated baseline is a necessary evil. It has happened to every single ecosystem to some extent or another, and it will happen to the various RISC-V ecosystems too.

I just disagree with the implication that the RISC-V approach was the right approach [1]. I think ARMv8.0-A did a much better job, including almost all the instructions you need in the very first version, if only they had included proper atomics.

[1] That is, not the right approach for creating a modern, commercially relevant ISA. RISC-V was originally intended as more of an academic ISA, so focusing on minimalism and "RISCness" was probably the best approach for that field.

[brucehoult]

It takes a heck of a lot longer if you wait until all the advanced features are ready before you publish anything at all.

I think RISC-V did pretty well to get everything in RVA23 -- which is more equivalent to ARMv9.0-A than to ARMv8.0-A -- out after RV64GC aka RVA20 in the 2nd half of 2019.

We don't know how long Arm was cooking up ARMv8 in secret before they announced it in 2011. Was it five years? Was it 10? More? It would not surprise me at all if it was kicked off when AMD demonstrated that Itanium was not going to be the only 64 bit future by starting to talk about AMD64 in 1999, publishing the spec in 2001, and shipping Opteron in April 2003 and Athlon64 five months later.

It's pretty hard to do that with an open and community-developed specification. By which I mean impossible.

I can't even imagine the mess if everyone knew RISC-V was being developed from 2015 but no official spec was published until late 2024.

I am sure it would not have the momentum that it has now.

[klelatti]

> We don't know how long Arm was cooking up ARMv8 in secret before they announced it in 2011. Was it five years? Was it 10? More?

Arm says "We started work on ARMv8-A in 2007"

Source: Arm plc Strategic Report 2014

[NetMageSCW]

When were these extensions available and used by popular compilers for higher level languages?

[chongli]

it's basically impossible to remove an instruction.

Of course not. You can replace an instruction with a polyfill. This will generally be a lot slower, but it won't break any code if you implement it correctly.

[snvzz]

It'll continue to take instruction encoding space.

[LeFantome]

While I agree with you, the original comment was still valuable for understanding why RISC-V has evolved the way it has and the philosophy behind the extension idea.

Also, it seems at least some of the RISC-V ecosystem is willing to be a little bit more aggressive. With Ubuntu making RVA23 the minimum profile for Ubuntu, perhaps we will not be waiting a decade for it to become the default. RVA23 was only ratafied a year ago.

[yellowapple]

> it's basically impossible to remove an instruction

laughs and/or cries in one of the myriad OISC ISAs