Every time new extensions get added to x86 new patents and copyrights are issued to cover those extensions. If you want to make a CPU compatible with what a current compiler produces, you need most of those extensions.
Legally, could a CPU manufacturer implement the unencumbered ISA in hardware and have a separate corporate entity provide a low-level software compatibility trap for the missing instructions? The CPU could even have functional equivalent (but ISA-incompatible) instructions to make it almost as fast. Kind of like third-party microcode?
In theory yes, the problem is that even x86 emulation in hardware in order to run x86 code natively without recompiling can drag you into a legal mess which any western company will avoid.
NVIDIA got pinched for this over a decade ago.
I’m not entirely sure how Qualcomm and Apple didn’t.
But overall the more you try to make an x86 enabled alternative viable the more likely you’ll get served with papers and even if you’ll win it would take a decade and cost 100’s of millions to fight.
There are further versions of SSE (SSE4 is pretty much a hard requirement on Windows) and a follow-on series, AVX. AVX-512 is from 2016 and AVX10 is from 2023.
That makes me wonder if all those vector extensions pilling on top of each other were really that necessary, or if they are mostly a means of keep churning out patents to delay expiration.
Is it possible to just improve the original SSE extensions in a logical backward compatible way? Similar to what AMD did to x86, widening it to x86-64, dooming Intel efforts to push the incompatible Itanium architecture?
SSE3+ & AVX{,2,-512} & co improve on SSE in pretty much the same way that x86-64 improves on x86 - the old thing still works just fine, but the new one is wider, adds new (very useful!) instructions, doesn't copy over others, and (at least partly) uses different encodings.
And an important thing to remember is that there is and never as a single "x86" before x86-64; both Intel and AMD added new instructions as was seen useful in new generations. AVX & co just continue the pattern that's been going on for four decades.
No, the newer extensions are different opcodes. It's like extending an API, you can't change old function signatures, you have to add new ones. The new ones are legitimately useful, most video games and media production software use them a lot.
Predicting what you will want in a few years is tricky at best. Some things that seem like a great idea are not worth it in the real world and so you pay the price for flexibility nobody uses. Some use case you didn't think of comes along that could really be helped with some tweak you didn't anticipate. thus your flexible architecture is both too flexible and not enough at the same time.
the above is a constant problem in engineering projects more than about 6 months old.
RISC-V uses a length-agnostic approach, so that would've at least bypassed the need for width-expansion upgrades. But it's something you have to take into account from the very start...
And even that only helps with the length problem, and doesn't help with doing new operations.
For SIMD, baseline x86-64 (i.e. SSE + SSE2) didn't have dynamic shuffles & shifts & blend, float floor/ceil, integer conversions & min/max & 64-bit comparisons & 32-bit mul, just to name things useful for even very boring SIMD; then in AVX2 we also get gather/masked load/store, FMA, and in AVX-512 we get a bunch of neat mask stuff, integer narrowing & rotates, compress.
(much of those things RVV has in its base extension, but RISC-V already has a good number of extensions on top of base RVV for things like like float16/bfloat16, expanded bitwise stuff (Zvbb - rotates/popcount/lzcnt/widening shift), clmul, and a bunch of crypto things; and presumably in a decade there'll be a bunch more things that people will want in their CPUs that'll have no choice but to be new extensions)
If the scalable VLEN is the same as the ALU width, which should generally be the target, small vectors would still perform optimally.
Of course if you need less than a VLEN-sized vector you're wasting throughput, but that applies just as much when using 128-bit vectors on AVX-capable hardware, and even worse so on AVX-512-capable (which, while double-pumped or equivalent to some extent on most impls, still has 512-bit-exclusive throughput on most).
Some of SSE is required as part of the x86_64 ABI, and also new versions of Windows (infamously, now) add required CPU extensions so software will often base its requirements on that. And SSE4x is ubiquitous enough (99% of PCs) that some software/games will just require it and simply crash if it can't use those instructions.
Considering there are no meaningful patent-free x86 CPUs in the wild, why should they?
It's just the default optimization level for those distros. If patent-free x86 CPUs become relevant, compiling another set of binaries would be trivial. Until then it doesn't make any sense to kneecap the >99% of x86 deployments by deliberately refusing to use faster and more efficient instructions.
Wait, you can patent an operation? Is it not considered an API? I assumed the Java case would meant you couldn't. I would think it would be limited to the hardware implementation, or maybe some specifics of the alg.
They said that, but my understanding was that they were really trying to scare apple back on to x86-64. It didn't work, and it was pretty specious anyway.
