Now they want to introduce yet another (sic!) extension Oilsm... It maaaaaay become part of RVA30, so in the best case scenario it will be decades before we will be able to rely on it widely (especially considering that RVA23 is likely to become heavily entrenched as "the default").
IMO the spec authors should've mandated that the base load/store instructions work only with aligned pointers and introduced misaligned instructions in a separate early extension. (After all, passing a misaligned pointer where your code does not expect it is a correctness issue.) But I would've been fine as well if they mandated that misaligned pointers should be always accepted. Instead we have to deal the terrible middle ground.
>atomic memory operations are made mandatory in Ziccamoa
In other words, forget about potential performance advantages of load-link/store-conditional instructions. `compare_exchange` and `compare_exchange_weak` will always compile into the same instructions.
And I guess you are fine with the page size part. I know there are huge-page-like proposals, but they do not resolve the fundamental issue.
I have other minor performance-related nits such `seed` CSR being allowed to produce poor quality entropy which means that we have bring a whole CSPRNG if we want to generate a cryptographic key or nonce on a low-powered micro-controller.
By no means I consider myself a RISC-V expert, if anything my familiarity with the ISA as a systems language programmer is quite shallow, but the number of accumulated disappointments even from such shallow familiarity has cooled my enthusiasm for RISC-V quite significantly.
RISC-V truly is the RyanAir of processors: Oh, you want FP maths? That's an optional extra, did you check that when you booked? And was that single or double-precision, all optional extras at an extra charge. Atomic instructions, that's an extra too, have your credit card details handy. Multiply and divide? Yeah, extras. Now, let me tell you about our high-end customer options, packed SIMD and user-level interrupts, only for business class users. And then there's our first-class benefits, hypervisor extensions for big spenders, and even more, all optional extras.
So it's modular. This is normally considered a good thing. It means you don't have to pay for features you don't need.
The ISA is open so there's no greedy corporation trying to upsell you. I mean there's an implementation and die area cost for each extension but it's not being set at an artificial level by a monopolist.
There's a good chance you're actually paying more for the features you don't need. Preparing an EUV mask set costs something like 30 million dollars (that figure may be out of date, i.e. it could be more now). So instead of a single mask set with everything on the device, whether you need it or not, you're paying $30 million for each special-snowflake variant. This is why vendors do a one-size-fits-all version of many of their products and then disable the extra functionality for the cheaper market segments, because it's much, much cheaper than making separate reduced-functionality devices.
It's a good thing in many cases but not if you're going to be running applications distributed as binaries. Maybe if we go the Gentoo route of everybody always recompiling everything for their own system?
RVA23 is, finally, the belated admission that maybe we shouldn't have everything as optional extras. Hopefully it'll take off, I can't imagine what sort of a headache it is for maintainers of repos who have to track a dozen different variants of binaries depending on which flavour of RISC-V the apt-get is coming from.
The "G" extension for everything you want to run shrink-wrapped binaries on a standard OS has been there since the May 7 2014 "User Level ISA, Version 2.0", which is before RISC-V started to be promoted outside of Berkeley e.g. at Hot Chips 26 in August 2014, and the first RISC-V workshop in January 2015 in Monterey.
The name "G" has morphed into now (along with the C extension) being called "RVA20", which led to "RVA22" and "RVA23", but the principle is unchanged.
"An integer base plus these four standard extensions (“IMAFD”) is given the abbreviation “G” and provides a general-purpose scalar instruction set. RV32G and RV64G are currently the default target of our compiler toolchains."
"Making everything optional" is for the embedded space.
As for general purpose processors, RISC-V has always had the idea of profiles (mandatory set of extensions). Just look at the G extension, which mandated floating point, multiply/division, atomics, ... things that you expect to see on user-facing general-purpose processors.
> the belated admission that maybe we shouldn't have everything as optional extras
That's why I disagree with the above claim.
(1) The optionality is a feature of RISC-V and it allows RISC-V to shine on different ecosystems. The desktop isn't everything.
(2) RISC-V has always addressed the fear of fragmentation on the desktop by using profiles.
RVA23 (and RVA20 before it) aren't an admission that Risc-V got it wrong. It's a necessary step to make Risc-V competetive in the desktop space as opposed to micro-controllers where the flexibility is hugely valuable.
The "G" extension for everything you want to run shrink-wrapped binaries on a standard OS has been there since the May 7 2014 "User Level ISA, Version 2.0", which is before RISC-V started to be promoted outside of Berkeley e.g. at Hot Chips 26 in August 2014, and the first RISC-V workshop in January 2015 in Monterey.
The name "G" has morphed into now (along with the C extension) being called "RVA20", which led to "RVA22" and "RVA23", but the principle is unchanged.
"An integer base plus these four standard extensions (“IMAFD”) is given the abbreviation “G” and provides a general-purpose scalar instruction set. RV32G and RV64G are currently the default target of our compiler toolchains."
But that means a port of Linux can’t be to RISC-V, it has to be to a specific implementation of RISC-V, or if sufficient (which seems still debatable) to a specific common RISC-V profile.
In what way are RISC-V profiles debatable? Canonical is spearheading the RVA23-as-a-default movement and so far, it seems that there are no heavy objections towards that effort (beyond the usual "Canonical sucks" shtick that you see in every discussion involving Canonical)
You can target the minimum instruction set and it'll run everywhere. Albeit very slowly. Perhaps you use a fat binary to get reasonable performance in most cases.
This isn't easy but it can be done (and it is being done on x86, despite constantly evolving variations of AVX).
Interestingly, RISC-V vector extensions are variable length.
So, you can compile your RISC-V software to require the equivalent of AVX and it will run on whatever size vectors the hardwre supports.
So, on x86-64, if I write AVX2 software and run it on AVX512 capable hardware, I am leaving performance on the table. But if I write software that uses AVX512, it will not run on hardware that does not support those extensions (flags).
On RISC-V, the same binary that uses 256 bit vectors on hardware that only supports that will use 512 bit vectors on hardware that supports it, or even 1024 bit vectors on hardware like the A100 cores of the SpacemiT K3.
So, I guess X86-64 is is the RyanAir of processors.
(Personal opinion)
I get the impression that RISC-V-related discussions often lack of awareness of prior work/alternatives. A large amount of (x86) software actually uses our Highway library to run on whatever size vectors and instructions the CPU offers.
This works quite well in practice. As to leaving performance on the table, it seems RVV has some egregious performance differences/cliffs. For example, should we use vrgather (with what LMUL), or interesting workarounds such as widening+slide1, to implement a basic operation such as interleaving two vectors?
> For example, should we use vrgather (with what LMUL), or interesting workarounds such as widening+slide1, to implement a basic operation such as interleaving two vectors?
Use Zvzip, in the mean time:
zip: vwmaccu.vx(vwaddu.vv(a, b), -1, b), or segmented load/store when you are touching memory anyways
unzip: vsnrl
trn1/trn2: masked vslide1up/vslide1down with even/odd mask
The only thing base RVV does bad in those is register to register zip, which takes twice as many instructions as other ISAs. Zvzip gives you dedicated instructions of the above.
Looks like the ratification plan for Zvzip is November. So maybe 3y until HW is actually usable?
That's a neat trick with wmacc, congrats. But still, half the speed for quite a fundamental operation that has been heavily used in other ISAs for 20+ years :(
Great that you did a gap analysis [1]. I'm curious if one of the inputs for that was the list of Highway ops [2]?
RyanAir is about exploiting consumers, with bait-and-switch and shitty terms and conditions.
RISC-V's modularity is about giving choice to hardware designers, so they can pick and choose just those features that their solution needs, and even allow for custom extensions.
RISC-V's modularity is for academia. 1) for education, where students learn/use/work on simple processors, 2) for research in new types of hardware and extensions, where ease of implementation or ease of creating a custom extension is important.
Extensiosn are not just for academia. If I am building a microcontroller to control the storage media I am selling (eg. hard drives), why do I need to implement a bunch of features I am not going to use? What about my flow rate monitor? Or my pacemaker?
In some of these, less silicon means less power means more better. Like that last example.
Then x86_64 is the cable television service of processors. "Oh, you want channel 5? Then you have to buy this bundle with 40 other channels you will never watch, including 7 channels in languages you do not speak."
And where it actually mattered they did not introduce a separate extension. Integer division is significantly more complex than multiplication, so it may make sense for low-end microcontrollers to implement in hardware only the latter.
RyanAir is the least expensive right? And it still gets you there?
I would be ok with that if it was a valid analogy.
It is valid in microcontroller land. There, the chip and the software are provided by the same party. So you can select for exactly the RISC-V features you need and save yourself some silicon. That sounds like a win to me.
At the application level, like a server or a desktop, that would be a disaster because I get my hardware and software from different people. How do the software guys know what hardware to target? Well, that is exacly why RVA23 exists.
What does RVA23 mean? It is the RISC-V "Application" profile. It allows you to build software to a single hardware target and trust that hardware makers will target the same proifle. RVA23 is like saying x86-64v4. Both are simple names for a long list of extensions (flags) and assumptions that you expect the hardware to honour. So, when Ubuntu 26.04 says it requires RVA23, it means that all the software built on it can assume those features. No a la carte.
The reason RVA23 is geting so much attention is that it has essentially the same feature set as modern ARM64 or x86-64. Software will be able to target this profile for a long time. There may be a new profile in a few years time, like RVA30, but hardware that implements that will still run RVA23 software (just as x86-64v4 hardware will run x86-64v1 software). Hardware built for profiles before RVA23 may be missing features modern applications expect.
I guess you could say that RVA23 is British Airways Business Class.
If you really want to support hardware designed before RVA23, almost everything you would want to run pre-built software on supports RVA20. And again, your RVA20 stuff will run fine on RVA23 hardware (but with fewer features--like no vectors). So maybe no in-flight meal, but it will get you there.
I think having separate unaligned load/store instructions would be a much worse design, not least because they use a lot of the opcode space. I don't understand why you don't just have an option to not generate misaligned loads for people that happen to be running on CPUs where it's really slow. You don't need to wait for a profile for that.
As for `seed`, if you're running on a microcontroller you can just look up the data sheet to see if it's seed entropy is sufficient. By the time you get to CPUs where portable code is important a CSPRNG is probably fine.
I agree about page size though. Svnapot seems overly complicated and gives only a fraction of the advantages of actually bigger pages.
>As for `seed`, if you're running on a microcontroller you can just look up the data sheet to see if it's seed entropy is sufficient.
It's a terrible attitude to have towards programmers, but looking at misaligned ops, I guess we can see a pattern from RISC-V authors here.
Most programmers do not target a concrete microcontroller and develop every line of code from scratch. They either develop portable libraries (e.g. https://docs.rs/getrandom) or build their projects using those libraries.
The whole raison d'être of an ISA is to provide a portable contract between hardware vendors and programmers . RISC-V authors shirk this responsibility with "just look at your micro specs, lol" attitude.
The option to generate or not generate misaligned loads/stores does exist (-mno-strict-align / -mstrict-align). But of course that's a compile-time option, and of course the preferred state would be to have use of them on by default, but RVA23 doesn't sufficiently guarantee/encourage them not being unreasonably-slow, leaving native misaligned loads/stores still effectively-unusable (and off by default on clang/gcc on -march=rva23u64).
aka, Zicclsm / RVA23 are entirely-useless as far as actually getting to make use of native misaligned loads/stores goes.
Yeah, that is quite funky; and indeed gcc does that. Relatedly, super-annoying is that `vle64.v` & co could then also make use of that same hardware, but that's not guaranteed. (I suppose there could be awful hardware that does vle8.v via single-byte loads, which wouldn't translate to vle64.v?)
> RVA23 doesn't guatantee them not being unreasonably-slow
Right but it doesn't guarantee that anything is unreasonably slow does it? I am free to make an RVA23 compliant CPU with a div instruction that takes 10k cycles. Does that mean LLVM won't output div? At some point you're left with either -mcpu=<specific cpu> and falling back to reasonable assumptions about the actual hardware landscape.
Do ARM or x86 make any guarantees about the performance of misaligned loads/stores? I couldn't find anything.
Exactly, I 100% agree, and IMO toolchains should default to assuming fast misaligned load/store for RISC-V.
However, the spec has the explicit note:
> Even though mandated, misaligned loads and stores might execute extremely slowly. Standard software distributions should assume their existence only
for correctness, not for performance.
Which was a mistake. As you said any instruction could be arbitrarily slow, and in other aspects where performance recommendations could actually be useful RVI usually says "we can't mandate implementation".
I don't think x86/ARM particularly guarantee fastness, but at least they effectively encourage making use of them via their contributions to compilers that do. They also don't really need to given that they mostly control who can make hardware anyway. (at the very least, if general-purpose HW with horribly-slow misaligned loads/stores came out from them, people would laugh at it, and assume/hope that that's because of some silicon defect requiring chicken-bit-ing it off, instead of just not bothering to implement it)
Indeed one can make any instruction take basically-forever, but I think it's a fairly reasonable expectation that all supported hardware instructions/behaviors (at least non-deprecated ones) are not slower than a software implementation (on at least some inputs), else having said instruction is strictly-redundant.
And if any significant general-purpose hardware actually did a 10k-cycle div around the time the respective compiler defaults were decided, I think there's a good chance that software would have defaulted to calling division through a function such that an implementation can be picked depending on the running hardware. (let's ignore whether 10k-cycle-division and general-purpose-hardware would ever go together... but misaligned-mem-ops+general-purpose-hardware definitely do)
> if general-purpose HW with horribly-slow misaligned loads/stores came out from them
How is that different for RISC-V?
> I think it's a fairly reasonable expectation that all supported hardware instructions/behaviors (at least non-deprecated ones) are not slower than a software implementation
I agree! So just use misaligned loads if Zicclsm is supported. As you observed there's a feedback loop between what compilers output and what gets optimised in hardware. Since RVA23 hardware is basically non-existent at the moment you kind of have the opportunity to dictate to hardware "LLVM will use misaligned accesses on RVA23; if you make an RVA23 chip where this is horribly slow then people will laugh at you and assume it's some sort of silicon defect".
RISC-V hardware with slow misaligned mem ops does exist to non-insignificant extent, and it seems not enough people have laughed at them, and instead compilers did just surrender and default to not using them.
> As you observed there's a feedback loop between what compilers output and what gets optimised in hardware.
Well, that loop needs to start somewhere, and it has already started, and started wrong. I suppose we'll see what happens with real RVA23 hardware; at the very least, even if it takes a decade for most hardware to support misaligned well, software could retroactively change its defaults while still remaining technically-RVA23-compatible, so I suppose that's good.
> RISC-V hardware with slow misaligned mem ops does exist to non-insignificant extent
Only U74 and P550, old RV64GC CPUs.
SiFive's RVA23 cores have fast misaligned accesses, as do all THead and SpacemiT cores.
I can't imagine that all the Tenstorrent and Ventana and so forth people doing massively OoO 8-wide cores won't also have fast misaligned accesses.
As a previous poster said: if you're targeting RVA23 then just assume misaligned is fast and if someone one day makes one that isn't then sucks to be them.
P550 is, like, what, only a year old? I suppose there has been some laughing at it at least.
Also Kendryte K230 / C908, but only on vector mem ops, which adds a whole another mess onto this.
I'd hope all the massive OoO will have fast misaligned mem ops, anything else would immediately cause infinite pain for decades.
But of course there'll be plenty of RVA23 hardware that's much smaller eventually too, once it becomes a general expectation instead of "cool thing for the very-top-end to have".
I do agree that it'd be reasonable to just assume fast misaligned ops, but for whatever reason gcc and clang just don't, and that's what we have for defaults.
It has take a while for this core to appear in an SoC suitable for SBCs, as Intel was originally announced as doing that and got as far as showing a working SoC/Board at the Intel Innovation 2022 event in September 2022.
Someone who attended that event was able to download the source code for my primes benchmark and compile and run it, at the show, and was kind enough to send me the results. They were fine.
For reasons known only to Intel, they subsequently cancelled mass production of the chip.
ESWIN stepped up and made the EIC7700X, as used in the Milk-V Megrez and SiFive HiFive Premier P550, which did indeed ship just over a year ago.
But technically we could have had boards with the Intel chip three years ago.
Heck we should have had the far better/faster Milk-V Oasis with the P670 core (and 16 of them!) two years ago. Again, that was business/politics that prevented it, not technology.
> No, it was released to customers in June 2021, almost five years ago.
Ah, okay. (still, like, at least a couple decades newer than the last x86-64 chip with slow unaligned mem ops, if such ever existed at all? Haven't heard of / can't find anything saying any aarch64 ever had problems with them either, so still much worse for the RISC-V side).
Well, I suppose we can hope that business/politics messes will all never happen again and won't affect anything RVA23.
> I do agree that it'd be reasonable to just assume fast misaligned ops, but for whatever reason gcc and clang just don't, and that's what we have for defaults.
This very much has a "for now" on it. Once there is actually widespread hardware with the feature, I would be very surprised if the compilers don't update their heuristics (at least for RVA23 chips)
Indeed we shall hope heuristics update; but of course if no compilers emit it hardware has no reason to actually bother making fast misaligned ops, so it's primed for going wrong.
hardware devs traditionally have been pretty good at helping the compiler teams with things like this (because its a lot cheaper to improve the compiler than your chip).
>So just use misaligned loads if Zicclsm is supported.
LLVM and GCC developers clearly disagree with you. In other words, re-iterating the previously raised point: Zicclsm is effectively useless and we have to wait decades for hypothetical Oilsm.
Most programmers will not know that the misaligned issue even exists, even less about options like -mno-strict-align. They just will compile their project with default settings and blame RISC-V for being slow.
RISC-V could've easily avoided all this mess by properly mandating misaligned pointer handling as part of the I extension.
Well, we don't necessarily have to wait for Oilsm; software that wants to could just choose to be opinionated and run massively-worse on suboptimal hardware. And, of course, once Oilsm hardware becomes the standard, it'd be fine to recompile RVA23-targeting software to it too.
> RISC-V could've easily avoided all this mess by properly mandating misaligned pointer handling as part of the I extension.
Rather hard to mandate performance by an open ISA. Especially considering that there could actually be scenarios where it may be necessary to chicken-bit it off; and of course the fact that there's already some questionability on ops crossing pages, where even ARM/x86 are very slow.
I am not saying that RISC-V should mandate performance. If anything, we wouldn't had the problem with Zicclsm if they did not bother with the stupid performance note.
I would be fine with any of the following 3 approaches:
1) Mandate that store/loads do not support misaligned pointers and introduce separate misaligned instructions (good for correctness, so its my personal preference).
2) Mandate that store/loads always support misaligned pointers.
3) Mandate that store/loads do not support misaligned pointers unless Zicclsm/Oilsm/whatever is available.
If hardware wants to implement a slow handling of misaligned pointers for some reason, it's squarely responsibility of the hardware's vendor. And everyone would know whom to blame for poor performance on some workloads.
We are effectively going to end up with 3, but many years later and with a lot of additional unnecessary mess associated with it. Arguably, this issue should've been long sorted out in the age of ratification of the I extension.
2 is basically infeasible with RISC-V being intended for a wide range of use-cases. 1 might be ok but introduces a bunch of opcode space waste.
Indeed extremely sad that Zicclsm wasn't a thing in the spec, from the very start (never mind that even now it only lives in the profiles spec); going through the git history, seems that the text around misaligned handling optionality goes all the way back to the very start of the riscv/riscv-isa-manual repo, before `Z*` extensions existed at all.
More broadly, it's rather sad that there aren't similar extensions for other forms of optional behavior (thing that was recently brought up is RVV vsetvli with e.g. `e64,mf2`, useful for massive-VLEN>DLEN hardware).
>1 might be ok but introduces a bunch of opcode space waste.
I wouldn't call it "waste". Moreover, it's fine for misaligned instructions to use a wider encoding or be less rich than their aligned counterparts. For example, they may not have the immediate offset or have a shorter one. One fun potential possibility is to encode the misaligned variant into aligned instructions using the immediate offset with all bits set to one, as a side effect it also would make the offset fully symmetric.
Of course that'd result in entirely-avoidable slowdown for the potentially-misaligned ops. Perhaps fine for a program that doesn't use them frequently, but quite bad for ones that need misaligned ops everywhere.
In terms of correctness, there's also the possibility of partially-misaligned ops (e.g. an 8B load with 4B alignment, loading two adjacent int32_t fields) so you're not handling everything with correct faults anyways.
I don't think it's too bad. The compressed extension was arguably a mistake (and shouldn't be in RVA23 IMO), but apart from that there aren't any major blunders. You're probably thinking about how JAL(R) basically always uses x1/x5 (or whatever it is), but I don't think that's a huge deal.
About 1/3 of the opcode space is used currently so there's a decent amount of space left.
Nope. See https://github.com/llvm/llvm-project/issues/110454 which was linked in the first issue. The spec authors have managed to made a mess even here.
Now they want to introduce yet another (sic!) extension Oilsm... It maaaaaay become part of RVA30, so in the best case scenario it will be decades before we will be able to rely on it widely (especially considering that RVA23 is likely to become heavily entrenched as "the default").
IMO the spec authors should've mandated that the base load/store instructions work only with aligned pointers and introduced misaligned instructions in a separate early extension. (After all, passing a misaligned pointer where your code does not expect it is a correctness issue.) But I would've been fine as well if they mandated that misaligned pointers should be always accepted. Instead we have to deal the terrible middle ground.
>atomic memory operations are made mandatory in Ziccamoa
In other words, forget about potential performance advantages of load-link/store-conditional instructions. `compare_exchange` and `compare_exchange_weak` will always compile into the same instructions.
And I guess you are fine with the page size part. I know there are huge-page-like proposals, but they do not resolve the fundamental issue.
I have other minor performance-related nits such `seed` CSR being allowed to produce poor quality entropy which means that we have bring a whole CSPRNG if we want to generate a cryptographic key or nonce on a low-powered micro-controller.
By no means I consider myself a RISC-V expert, if anything my familiarity with the ISA as a systems language programmer is quite shallow, but the number of accumulated disappointments even from such shallow familiarity has cooled my enthusiasm for RISC-V quite significantly.