If decoding 32 bytes of code (256 bits, somewhere between 8 and 16 instructions) You can figure out where all the actual instructions start (yes, even the 16th instruction) with 2 layers of LUT6.
You can then use those outputs to mux two possible starting positions for 8 decoders that do 16 or 32 bit instructions, plus 8 decoders what will only ever do 16 bit instructions from fixed start positions (and might output a NOP or in some other way indicate they don't have an input).
OR you can use those outputs to mux the outputs of a 8 decoders that only do 32 bit instructions and 8 decoders that do 16 or 32 (all with fixed starting positions), plus again 8 decoders that only do 16 bit instructions from fixed start positions (possibly not used / NOP).
The first option uses less hardware but has higher latency.
That, again, is for decoding between 8 and 16 instructions per cycle, with an average on real code of close to 12.
That is more than is actually useful on normally branchy code.
In short: not a problem. Unlike x86 decoding.
I’m assuming the decode window is aligned, but that may be a bad assumption.
Unlike x86 where instructions can range from 1 up to 15 byte, current RISC-V ISA only has 2 instruction sizes.
Today x86 decoding is limiting because we want to decode more than ~4 instructions each cycle, for RISC-V to cause same decoding difficulty it would probably be required to decode more than ~20 instructions each cycle
What I've been wondering is how difficult it is to fuse instructions when the compressed instructions you're trying to fuse isn't aligned to a 32-bit word.
Uh? In RISC-V with the C extension, 32bit instructions are 16-bits aligned, so no you don't.
However, trade offs matter. Compressing instructions may improve density, but it makes them variable length. This is a big barrier to decoding in parallel, which is very important to high performance cores.