Mask `rs2` when using `bext` instruction if `rex.w` is not set

[Coekjan]

Fixes #1652 .

[ptitSeb]

Mmmm, is that a fix or a workaround?

doesn't BEXT works on 32bits input?

[Coekjan]

I think this is a fix and not a workaround.

bext on rv64 does not handle 32bits input. This pseudo code describes how it works:

let index = X(rs2) & (XLEN - 1);
X(rd) = (X(rs1) >> index) & 1;

In the first step, X(rs2) & (XLEN - 1) is actually X(rs2) & (64 - 1) in rv64. When rex.w is not set, what we need to do is X(rs2) & (32 - 1) instead.

[ptitSeb]

Then why not using the 's0' scratch reg to do the mask and still use BEXT, that would still be 1 opcode less then the fallback?

[Coekjan]

Then why not using the 's0' scratch reg to do the mask and still use BEXT, that would still be 1 opcode less then the fallback?

This can be better. I will push a new version and then please review the changes again. Thanks!