[monocasa]

I don't think anyone bakes instruction alignment into their caches since the early 2000s, and adding an extra bit to the branch predictors isn't that big of a deal. It's got to be the first or second stage of their front end right before the decoders.

[phire]

Why not bake instruction alignment into the cache? When you can assume instructions will always be 32bit aligned, then you can simplify the icache read port and simplify the data path from the read port to the instruction decoder. Seems like it would be an oversight to not optimise for that.

Though, I suspect that's easy problem to fix. The more pressing issue is what happens after the decoders. I understand this is a very wide design, decoding say 10 instructions per cycle.

There might be a single 16bit instruction in the middle of that block 40 bytes, changing the alignment halfway though. To keep the same throughput, Qualcomm now need 20 decoders, one attempting to decode on every 16bit boundary. The extra decoders waste power and die space.

Even worse, they somehow need to collect the first 10 valid instructions from those 20 decoders. I really doubt they have enough slack to do that inside the decode stage, or the next stage, so Qualcomm might find them selves adding an entire extra pipeline stage, (probably before decode, so they can have 20 simpler length decoders feeding into 10 full decoders on the next) just to deal with possible misaligned instructions.

I don't know how flexible their design is, it's quite possible adding an entire extra pipeline stage is a big deal. Much bigger than just rewriting the instruction decoders to 32bit RISC-V.

[monocasa]

Because RISC-V was designed to be trivial to decode length for, you simply need to look at the top two bits of each 16bit word to tell if it's a 32bit or 16bit instruction. At that point, spending the extra I$ budget isn't worth it. Those 20 'simple decoders' are literally just each one 2nand gate. Adding complexity to the I$ hasn't even made sense for x86 in two decades, because of the extra area needed for the I$ versus the extra decode logic. And that's a place where this extra decode is legitimately an extra pipeline stage.

> I don't know how flexible their design is, it's quite possible adding an entire extra pipeline stage is a big deal. Much bigger than just rewriting the instruction decoders to 32bit RISC-V.

I'm sure it is legitimately simpler for them. I'm not sure we should bend over backwards and bring down the rest of the industry because they don't want to do it. Veyron, Tenstorrent were showing off high perf designs with RV-C.

[phire]

It doesn't matter how optimised the length decoding is. Not doing it is still faster.

For an 8-wide or 10-wide design, the propagation delays are getting too long to do it in all in single cycle. So you need the extra pipeline stage. The longer pipeline translates to more cycles wasted on branch mispredits.

RISC-V code is only about 6-14% denser than Aarch64 [1], I'm really not sure the extra complexity is worth it. Especially since Aarch64 still ends up with a lower instruction count, so it will be faster whenever you are decode limited instead of icache limited.

> Adding complexity to the I$ hasn't even made sense for x86 in two decades

Hang on. Limiting the Icache to only 32bit aligned access actually simplifies it.

And since the NUVIA core was originally an aarch64 core, why wouldn't they optimise for hardcoded 32bit alignment and get a slightly smaller Icache?

[1] https://www.bitsnbites.eu/cisc-vs-risc-code-density/

[monocasa]

> Hang on. Limiting the Icache to only 32bit aligned access actually simplifies it.

Even x86 only reads 16 or 32 byte aligned fields out of the I$, then shifts them. There's not extra I$ complexity. You still have to do that shift at some point, in case you don't jump 32 byte aligned address. You also ideally don't want to only hit peak decode bandwidth starting on aligned 32 byte program counters, so that whole shift register thing is pretty much a requirement. And that's where most of the propagation delays are.

> RISC-V code is only about 6-14% denser than Aarch64 [1], I'm really not sure the extra complexity is worth it. Especially since Aarch64 still ends up with a lower instruction count, so it will be faster whenever you are decode limited instead of icache limited.

There's heavy use of fusion, and fwiw, the M1 also heavily fuses into micro ops too (and I'm sure the AArch64 morph of NUVIA's cores do too).

[Symmetry]

Under a classic RISC architectures you can't jump to non-aligned addresses. That lets you specify jumps that are 4 times longer for the same number of bits in your jump instruction. Here's MIPS as an example:

https://en.wikibooks.org/wiki/MIPS_Assembly/Instruction_Form...

[monocasa]

Classic RISC was targeting about 20k gates and isn't really applicable here.

[hajile]

Back in 2019, RISC-V was 15-20% smaller than x86 (up to 85% smaller in some cases) and was 20-30% smaller than ARM64 (up to 50% smaller in some cases).

https://project-archive.inf.ed.ac.uk/ug4/20191424/ug4_proj.p...

Since then, RISC-V has added a bunch more instructions that ARM/x86 already had which has made RISC-V even smaller relative to them.