[RISCV] Remove unnecessary dependencies on vtype register.

[mgudim]

Note that this is not 100% functionally correct, but all spec benchmarks seem ok, at least in my testing.

I've only measured cycles on small benchmark extracted from x264. I compiled this benchmark for a couple of upstreamed cpus, but I didn't notice significant difference. It does make a significant difference for our cpu though

[lukel97]

Do you have a test diff which shows the scheduling change?

[preames]

Note that this is not 100% functionally correct,

This seems like a rather major problem. Why should we spend review time on something which is known not to be correct? Is there an purpose to having the off-by-default option?

[mgudim]

Note that this is not 100% functionally correct,

This seems like a rather major problem. Why should we spend review time on something which is known not to be correct? Is there an purpose to having the off-by-default option?

@preames Sorry, I didn't mean to ask for a review on this or merge this. This is the "quick and dirty" solution. In case you wanted to see the performance impact.

I posted the draft of what this code should look like here: #149574 (comment)