RISC-V Conditional Moves
corsix.org
corsix.org
That said, RISC-V does have a proper conditional move instruction. And the funny part: it has multiple! `xtheadcondmov` and `xmipscmove` both implement "real" conditional moves. The catch is that those are vendor-specific extensions; compared to the official narrative that "it doesn't fit the design of RISC-V" apparently the actual hardware vendors see the value of adding real cmovs to their hardware. I wonder how many more vendor-specific extensions will it take before a common cross-vendor extension is standardized, if ever?
(And yes, I'm perfectly aware of why `Zicond` was designed the way it was. I don't really want to get into a discussion whether that's the right design or not long-term.)
Only if you need the full properties of cmove. In many cases it just generates a single Zicond.
While some companies implement a 3R1W integer pipeline and use fusion, others keep the integer side 2R1W. If you use 2R1W you can get wider issue for the same area, if you have a four issue integer pipeline you may be able to add a fifth integer execition unit for cheaper than moving it to 3R1W, which may give you a higher performance gain.
Or, better yet, have the 3R extra port come from some of the 2R being split up; e.g. for a block of 3×2R1W ALUs, be able to split one up for its read ports, reusing it as 2×3R1W when needed, thereby being able to do 3R1W at 66% the throughput of 2R1W without any extra register ports (i.e. 1.3x throughput benefit of 3R1W over two 2R1W instrs). Probably has some extra costs from scheduling & co needing to handle 3R though.
P adds instructions like integer multiply-accumulate, which have a third register read (for rd). So, they're taking the opportunity to add a few forms of 3-register select instructions:
MVM Move Masked
for each bit i: X(rd)[i] = X(rs2)[i] ? X(rs1)[i] : X(rd)[i]
MVMN Move Masked Not
for each bit i: X(rd)[i] = X(rs2)[i] ? X(rd)[i] : X(rs1)[i]
MERGE Merge
for each bit i: X(rd)[i] = X(rd)[i] ? X(rs2)[i] : X(rs1)[i]
Adding more instructions is kind of non productive for a R(educed)ISC ISA. It has to be weighted with extreme care. Compressed instructions went thru for the sake of code density (marketing vs arm thumb instructions).
In the end, programs will want probably to stay conservative and will implement only the core ISA, at best giving some love to some instruction fusion patterns and that's it, unless being built knowingly for a specific risc-v hardware implementation.
As such, there are compromises for both aims.
RISC-V has cmp-and-branch in a single instruction, which with c.mv normally makes six bytes. If the cmp-and-branch instruction tests one of x8..x15 against zero then that could also be a compressed instruction: making four bytes in total.
https://developer.arm.com/documentation/ddi0403/d/Applicatio...
T32 is a pretty good encoding but far from perfect. If they had the chance to redo it I doubt they would spend a full 1/32nd of the encoding space on asrs, for example.
This is probably not the case. The core ISA doesn't include floating point, it doesn't include integer multiply or divide, it doesn't include atomic and fence instructions.
What has happened is that most compilers and programs for "normal desktop/laptop/server/phone class systems" all have some baseline set of extensions. Today, this is more or less what we call the "G" extension collection (which is short-hand for IMAFD_Zicsr_Zifencei). Though what we consider "baseline" in "normal systems" will obviously evolve over time (just like how SSE is considered a part of "baseline amd64" these days but was once a new and exotic extension).
Then lower power use cases like MCUs will have fewer instructions. There will be lots of MCUs without stuff like hardware floating point support that won't run binaries compiled for the G extension collection. In MCU use cases, you typically know at the time of compiling exactly what MCU your code will be running on, so passing the right flags to the compiler to make sure it generates only the supported instructions is not an issue.
And then HPC use cases will probably assume more exotic extensions.
And normal "desktop/phone/laptop/server" style use cases will have runtime detection of things like vector instructions in some situations, just like in aarch64/amd64.
https://riscv.org/ecosystem-news/2025/04/risc-v-rva23-a-majo...
clang, as far back as Compiler Explorer goes (i.e. clang 3.0.0, i.e. 2011), always assumes SSE for -m32; presumably because there's nothing to be backwards-compatible to, unlike gcc.
Doesn't look particularly good for "default will just change at some point", though we can hope.
Ofc, if your program uses floating point calculations you will want to use the hardware machine instructions for that.
Here, we were talking about about all those machine instructions which do not bring much more on top of the core ISA. Those would be implemented using fusion, appropriate for R(educed)ISC silicon. The trade-off is code density, and code density on modern silicon, probably in very specific niches, but there, program machine instructions would be generated (BTW, probably written instead of generated for those niches...) with those very specific niches in mind.
And RISC-V hardware implementations, with proper publishing of most common, and pertinent, machine instruction fusion patterns, will be able to "improve" step by step, targetting what they actually run and what would make real difference. Sure, this will require a bit of coordination to agree on machine instruction fusion patterns.
Re-read my post, please.
The problem is those machine instructions not bringing much more than the core ISA which do not require an ISA extension.
Stop using AIs and/or trolling, thx.
There are two possibilities here:
* Either I'm misunderstanding what you're saying, and you did not mean that most programs will use only the core ISA.
* Or you're trying to say that integer multiply/divide and floating point is part of the core ISA.
Which one is it?
If it's the first one, could you try to clarify? Because I can't see another way to interpret the phrase "programs will want probably to stay conservative and will implement only the core ISA".
I think sylware doesn’t mean the core ISA exactly, but the core with the standard extensions rather than manufacturer-specific extensions.
Let's start over for microsoft GPT-6.
It all depends on the program: if it does not need more than a conservative use of the ISA to run at a reasonable speed on targeted hardware, it should not use anything else. Those people tend to forget that large implementations of RISC-V will probably be heavy on machine instruction fusion.
In the end, adding 'new machine instructions' is only to be though about, after proper machine instruction fusion investigation.
They are jumping the gun way too easily on 'adding new machine instructions', forgetting completely about machine instruction fusion.
In an effort to show that I'm sincere and that this topic genuinely interests me, let me show you my RISC-V CPU implemented in Logisim: https://github.com/mortie/rv32i-logisim-cpu. For this project, I did actually only implement (most of) the core ISA; so in order to run C programs compiled with clang, I actually had to tell clang to generate code for the core RV32I. That means integer multiplication and division in the C source code was turned into loops which used addition, subtraction, shifts and branches to implement multiplication and division.
> It all depends on the program: if it does not need more than a conservative use of the ISA to run at a reasonable speed on targeted hardware, it should not use anything else.
Essentially all programs will benefit significantly from at the very least integer multiply and divide. And every single CPU that's even capable of running anything like a mainstream "phone/laptop/desktop/server class" operating system has the integer multiply and divide extension.
So to say that most programs will use the core ISA and not extensions is wild. Only a tiny minority of executables compiled for the absolute tiniest of RISC-V MCUs (or, y'know, my own Logisim RV32I CPU) will be compiled for the core RISC-V ISA.
Stop using AI, thx.
Honestly you're acting like an LLM instructed to produce antagonistic, bad-faith arguments. You're certainly not acting like a human who has any idea what he's talking about.
* Either I'm misunderstanding what you're saying, and you did not mean that most programs will use only the core ISA.
* Or you're trying to say that integer multiply/divide and floating point is part of the core ISA.
Which one is it?
In any case, "force hardware to do an extremely-stupid amount of fusion to get back the performance lost from intentionally not adding/using useful instructions" isn't a sane thing to target in any universe no matter ones goals; you're just wasting silicon & hardware development time that would be better spent actually doing useful things. Fusion is neat (esp. for fixing past mistakes or working around fixed-size instructions (i.e. all x86 & ARM use fusion for, but a from-scratch designed ISA with variable-length instrs (e.g. RISC-V) should need neither)), but it's still very unquestionably strictly worse than just having more actual instructions and using them.
Again, the bulk of the programs out there don't need those extensions to be reasonably performant on modern silicon hardware. In other words, all programs out there will want to stick to a conservative usage of the ISA anyway ("core-ish").
Programs requiring floating point hardware in order to be "usable" will mandate probably a cache line vector ISA extension silicon block (they won't even use the FPU ISA extension). Who would even use a FPU silicon block nowadays for floating point calculations (unless niche and small hardware implementation)?
(x86 and arm are out: they have strong IP locks in many places in the world, there are not to be considered for any sane future. Those are just legacy burden and full of "marketing" instructions)
Indeed, most sane software doesn't need most extensions to be "reasonably performant"; in fact, most sane software is reasonably-performant even on two decades old hardware!
But, unfortunately, there's a ton of software doing things quite inefficiently, and it will continue to exist forever unless something crazy happens like a non-insignificant amount of humans starting to care (impossible) or LLMs becoming functional enough to rewrite entire codebases (more possible than humans caring, at least).
You're extremely-heavily underestimating software doing random garbage in floating point (using it to compute a square root or multiplying an integer by 0.4 or something; ad-hoc game logic/physics that isn't written in a vectorizable way; doing a bunch of things where integers would do in FP (esp. languages which expose floats as the main datatype, esp. JavaScript))
It may be neat to dream about a hypothetical world where none of that garbage exists, but that dream isn't coming true today, nor is there any sign that it will at any point in the future. Basing architecture/compiler/configuration decisions around this hypothetical is just purely entirely stupid.
And even in that dream world a lot of code would benefit from sh1add/sh2add/sh3add from Zba, Zbb's min/max is useful in a ton of places, memory managers might want clz for computing bucket from size, anything doing bitwise stuff would benefit from andn and much of Zbs. And of course ideally the vast majority of code would be running in RVV instead of scalar code.
It seems to be also why the core ISA has only 32bits instructions: because smaller instructions hardly brings anything, that based the same feedback from experience. Maybe only on super small ultra tiny embedded micro-controllers with a very old silicon process... This smells more aggressive marketing using super niche or broken programs to justify itself.
Of course, there is no perfect REDUCED ISA: trade-offs were made based on the designers experience. Expect arm people to press hard on the bad side of those trade-offs (a trade-off has good sides and bad sides, definition), because risc-v is a death sentence for them (and they are making a push right now on HN, I can tell you...). Yep, arm and x86_64 have strong IP locks all around the glob... RISC-V, none, free for all to implement.
Nowadays, programs requiring floating point hardware acceleration for reasonable performance use vector machine instructions. I think this is a mistake of RVA2x: the FPU extension should not be there. Only cache line size vector machine instructions should be there. The FPU extension would be for niche/specialized/small hardware. And a scalar is a vector with one used dimension... and the "synchronous"/"inline" handling of floating point operations... yummy.
I have a lot of doubts on compressed instructions, because I don't see code density being that much of an killer feature (it sounds more like arm marketing to me), and I recall reading numbers going in this way and not the other way for the general case.
What I am very sure of: nobody wants to design a clean and modern ISA to handle the "bad" programs, come on, and in the worst case scenario that will fit only some "bads" not all of them anyway, choices will have to be made on the "accelerated bads"... sane? nope.
All that said, I am coding RISC-V and x86_64 assembly, and did a little bit of arm64: for the code I wrote, arm and risc-v were nearly the same.
What I am keeping an eye on is the memory reservation/ZACAS stuff though. Because hart(and io) synchronization in a world of (hart read/write queues) and cache memory coherency seems to be critical for "normal" performance and very quickly.
And another thing people tend to forget: RISC-V is standard across vendors/implementors, namely it is appropriate and reasonable to write fast code path variants in assembly... and that could change A LOT of things, well at least in the "system/kernel area" (extremely hard to do planned obsolescence is a killer feature...).
While some very-important software like video codecs, and various sporadic projects where some drive-by open-source dev decided to add an optimized path will use vector, that's, like, on the order of 0.001% (number out of my ass) of all software that runs slowly enough to be noticable, and the remaining 99.999% remains slow. Much as I like working with SIMD/vector, it's a very tiny minority of people that do.
RVV does also actually make good use of scalar FP, with .vf instruction variants which take one operand from the scalar registers, allowing storing constants in the scalar registers instead of wasting vector registers (and with LMUL it's very easy to exhaust the entire RVV register file). Especially important in matmul kernels.
> What I am very sure of: nobody wants to design a clean and modern ISA to handle the "bad" programs
And yet that's what RVA23 & co basically have to be, and are. And as such they have scalar FP, vector FP, and basically everything else that's potentially useful (other than 3-source-operand instructions).
I do wonder how much compressed actually benefits perf-wise, but it's very clearly true that, at least icache-wise, reducing code size by, say, 20%, is equivalent to adding 20% more icache; and 20% of a typical L1 icache is quite a lot of area to save.
I don't even mention ffmpeg.
RVAXX looks like more a grab bag to match x86_64 and aarch64, feature wise, and it includes bad features: this is very probably to ease porting only.
In a risc-v world, there would be much more assembly of code path variants (cross-vendor neutral-ish), and high level languages with assembly written interpreters.
No more C42+ , only ultra-stable-in-time core-ish ISA assembly... and Big Tech hates that because planned obsolescence is excrutiatingly harder to do.
RVA23 doesn't in any way ease porting, no clue what you're on about there; besides vectorizable code, where you necessarily do simply just need RVV or similar to get good performance, base RV64G does just cover everything needed for software to be able to run. All RVA23 does is just provide a baseline with extensions to be able to achieve good performance, and cheap instructions that software should've been already utilizing for decades but hasn't generally been able to due to legacy hardware not supporting them (importantly clz/ctz/cpop, but also to a smaller extent min/max, zicond, bit rotates).
Granted, RVA23 does have some more questionable (though still actually useful) inclusions like the cache block ones that force 64-byte cache lines to not be dysfunctional, but that also brings up the massive wart of mandatory 4K pages in base RISC-V, not even RVA23, that's explicitly chosen for ease-of-portability.
Maybe RV64 is the end-of-line for CPUs, but, for all we know, RV64 today might be what Intel 8086 was in 1978, and RV64 is to be extended and grown and eventually just replaced in the future.
The div/rem one is odd because I saw it suggested in the ISA manual, but I have yet to ever see that pattern crop up in compiled code. Usually it's just in library functions like C stdlib `div()` which returns a quotient and remainder, but why on earth are you calling that library function on a processor that has a divide instruction?
Because they rightfully expect that div() compiles down to the fastest div/rem idiom for the target hardware. Mainstream compilers go to great lengths to optimize calls to the core C math functions.
If stdlib div() were promoted to a builtin one day (it currently is not in GCC afaict), and its implementation were inlined, then the compiler would recognise the common case of one side of the struct being dead, and you'd still end up with a single div/rem instruction.
#include <cstdlib>
#include <cstdint>
#include <utility>
std::pair<int64_t, int64_t> LibDivWithRemainder(int64_t numerator, int64_t denominator) {
const auto res = std::div(numerator, denominator);
return std::make_pair(res.quot, res.rem);
}
std::pair<int64_t, int64_t> ManDivWithRemainder(int64_t numerator, int64_t denominator) {
const int64_t quot = numerator / denominator;
const int64_t rem = numerator % denominator;
return std::make_pair(quot, rem);
}
"LibDivWithRemainder(long, long)":
sub rsp, 8
call "ldiv"
add rsp, 8
ret
"ManDivWithRemainder(long, long)":
mov rax, rdi
cqo
idiv rsi
ret
LibDivWithRemainder(long, long):
jmp ldiv@PLT
ManDivWithRemainder(long, long):
mov rax, rdi
mov rcx, rdi
or rcx, rsi
shr rcx, 32
je .LBB1_1
cqo
idiv rsi
ret
.LBB1_1:
xor edx, edx
div esi
ret
mov rax, rdx
cdq
idiv r8
mov QWORD PTR [rcx], rax
mov rax, rcx
mov QWORD PTR [rcx+8], rdx
ret 0
Unlikely, as pointed out in sibling comments the core ISA is too limited. What might prevail is profiles, specifically profiles for application processors like RVA22U64 and RVA23U64, which the latter one makes a lot more sense IMHO.
I had to clarify the obvious: if a program does not need more than a conservative usage of the ISA to run at reasonable speed, no hardcore change to the hardware should be investigated.
Additionnally, the 'adding new machine instructions' fan boys tend to forget about machine instruction fusion (they probably want they names in the extension specifications) which has to be investigated first, and often in such niche cases, it may be not the CPU to think about, but specialized ASIC blocks and/or FPGA.
The conventional wisdom is that conditional moves mainly uplift in order pipelines, but I feel like there could be a benefit to increased ROB residency on OoOE cores as well with the right architecture.
But like I said, I don't have a good way to prove that or not.
At a first glance, it might seem insane to replace a simple data dependancy with a control-flow dependency, control-flow dependencies are way more expensive as they might lead to a miss-predict and pipeline flush.
But to understand modern Massively-Out-of-Order cores, you really need to get in the mindset of "The branch predictor stupidly accurate", and actually optimise for the cases when it was right.
If the data is anything other than completely random, the branch predictor will guess correctly (at least some of the time) and the dependency is now invisible to the backend. The dependency chain is broken and the execution units can execute both segments in parallel.
--------
So while more CMOVs might help with ROB residency, I'm really not sure that would translate to overall improved performance.
But this does make me wonder if it might be worth while designing an μarch that could dynamically swap between executing a CMOV style instruction as a branch or conditional move? If the CMOV is predictable, insert a fake branch into the branch predictor and handle it that way from now on.
It's pretty hard to make modern compilers reliably emit cmovs in my experience. I had to resort to inline asm.
That doesn't sound like a very well thought out argument. The moment you are conditional with respect to two independent conditions, you can run both conditional moves in parallel.
>At a first glance, it might seem insane to replace a simple data dependancy with a control-flow dependency, control-flow dependencies are way more expensive as they might lead to a miss-predict and pipeline flush.
The moment you have N parallel branches such as from unrolling a data parallel loop, you have a combinatorial explosion of 2^N possible paths to take. You have to successfully predict through all of them and you certainly can't execute them in parallel anymore.
Also, you're saying there is a miss predict and a pipeline flush, but those are concepts that relate to prefetching instructions and are completely irrelevant to conditional moves that do not change the instruction pointer. If you have nothing to execute, because you're waiting for a dependent instruction that is currently executing, then you're stalling the pipeline, which is equivalent to executing a NOP instruction. It's a waste of a cycle (not really, because you're waiting for a good reason), but it can't be more expensive than that.
IMHO this approach seems to fit modern CPU designs reasonably well. There is no explicit flag or predicate register, but it does require fusing 2 instructions with possibly different operand. But restricting which instructions can use it might help (even better if its completely orthogonal).
It can also do else clauses, instructions that get executed only when the condition fails.
I'm not sure how well this approach would work on modern CPUs; These days, Thumb-2 is generally only used on small microprocessors, and it's notable that ARM64 didn't carry that feature forwards.
I was not able to open the given link, but it's not true, at least for the U74.
Fusion means that one or more instructions are converted to one internal instruction (µop).
SiFive's optimisation [1] of a short forward conditional branch over exactly one instruction has both instructions executing as normal, the branch in pipe A and the other instruction simultaneously in pipe B. At the final stage if the branch turns out to be taken then it is not in fact physically taken, but is instead implemented by suppressing the register write-back of the 2nd instruction.
There are only a limited set of instructions that can be the 2nd instruction in this optimisation, and loads and stores do not qualify. Only simple register-register or register-immediate ALU operations are allowed, including `lui` and `auipc` as well as C aliases such as `c.mv` and `c.li`
> The whole premise of fusion is predicated on the idea that it is valid for a core to transform code similar to the branchy code on the left into code similar to the branch-free code on the right. I wish to cast doubt on this validity: it is true that the two instruction sequences compute the same thing, but details of the RISC-V memory consistency model mean that the two sequences are very much not equivalent, and therefore a core cannot blindly turn one into the other.
The presented code ...
mv rd, x0
beq rs2, x0, skip_next
mv rd, rs1
skip_next:
czero.eqz rd, rs1, rs2
mv rd, rs1 // safe even if they are the same register
bne rs2, x0, skip
mv rd, x0
skip:
Then switching to code involving loads and stores is completely irrelevant:
lw x1, 0(x2)
bne x1, x0, next
next:
sw x3, 0(x4)
Even putting the `sw` between the `bne` and the label is ludicrous. There is no branch-free code that does the same thing -- not only in RISC-V but also in arm64 or amd64. SiFive's optimisation will not trigger with a store in that position.
[1] SiFive materials consistently describe it as an optimisation not as fusion e.g. in the description of the chicken bits CSR in the U74 core complex manual.
It is not even about RISC-V but about instruction fusion in general in any ISA with a memory model at least as strong as RVWMO -- which includes x86. I'm not as familiar with the Aarch64 memory model, but I think this probably also applies to it.
The point here is that if an aggressive implementation wants to implement instruction fusion that removes conditional branches (or indirect branches) to make a branch-free µop -- for example, to turn a conditional branch over a move into something similar to the `czero` instruction -- then in order to maintain memory ordering AS SEEN BY A DIFFERENT CORE the fused µop has to also have `fence r,w` properties.
That is all.
It is irrelevant to this whether the actual RISC-V instruction set has a conditional move instruction, or the properties it has if it exists.
It is irrelevant to the situation where a human programmer or a compiler might choose to transform branchy code into branch-free code. They have a more global view of the program and can make sure things make sense. A CPU core implementing fusion has only a local view.
Finally, I'll note that instruction fusion is at present hypothetical in RISC-V processors that you can buy today while it has been used in both x86 and Arm chips for a long time.
Intel's "Core" µarch had fusion of e.g. `cmp;bCC` sequences in 2006, while AMD added it with Bulldozer in 2011. Arm introduced a limited capability -- `CMP r0, #0; BEQ label` is given as an example -- in A53 in 2012 and A57, A72 etc expanded the generality.
Upcoming RISC-V cores from companies such as Ventana and Tenstorrent are believed to implement instruction fusion for some cases.
Just for completeness, I'll again repeat that SiFive's U74 optimises execution of a condition branch and a following simple ALU instruction that execute simultaneously in two pipelines, but this is NOT fusion into a single µop.
No... It's kind of an artefact of RISC-V's memory model being weak. x86 side-steps the issue because it insists that stores always occur in program order, allowing it to fuse away conditional branches without issue.
(Note: the actual hardware implementation of x86 cpus issues the stores anyway, and then rewinds if it later detects a memory ordering violation)
RISC-V ran into this corner case because it wanted the best of both worlds: A Weak memory model, but still have strong ordering across branches.
Looks like ARM avoided this issue because its memory model is weaker, branches don't force any ordering, which means the arm compiler might need to insert a few extra memory barrier instructions.
---------
TBH, I don't think this fusing instructions edge case is a big deal. For smaller RISC-V cores, you aren't reordering memory operations in the first place.
And for larger RISC-V cores, you already need a complex mechanism for dealing with store order violationss, so you just throw your fused come instruction at it. Your core already needs to deal with sync points that aren't proper branches, because non-taken branches also enforce ordering.
> The whole premise of fusion is predicated on the idea that it is valid for a core to transform code similar to the branchy code on the left into code similar to the branch-free code on the right.
The idea of Zicond afaict is that the compiler transforms select sequences into (usually multiple) Zicond instructions, and cores with more register ports available can fuse Zicond compounds into more complex select macro-ops. It's a 2R1W vocabulary for describing selects which require more than 2 read ports.
As an aside I evaluated Zicond on my scalar 3-stage implementation and found that at 1 CPI for ALU ops and 2-cycle taken branch cost, the branchless sequences GCC produced for Zicond were never better and sometimes worse than the equivalent branching sequence. It really does seem to be targeting bigger cores, or constant-time execution
ARM went out of their way to remove it. Multiple times, with both AArch64 and the various implementations of Thumb.
A CPU for this century is not the one for last?