Re-encode the multi-destination register instructions.

[ben-marshall]

The current encoding of the multi-precision and xc.mror instructions is very wasteful.

Current usage patterns are:

• xc.madd.3 (a,b), c, d, a
• xc.msub.3 (a,b), c, d, a
• xc.mmul.3 (a,b), c, d, a
• xc.macc.1 (a,b), a, b, c

Most of the time, a source and a destination register (or two) are the same.

Suggestions:

• For xc.madd.3, xc.msub.3, xc.mmul.3, make what is now rs3 an implicit argument, which is always equal to the "low" destination register.
  • This frees up 32 code points per instruction.
• For xc.macc.1, either:
  • Remove rs3, make rs1 a multi-register address and add rs2 to the catenation of rs1 registers, placing the result in rdm.
  • Frees up 32-code points.
  • Make rdm an implicit source register pair.
    • Frees up 256 code points.

For xc.mror:

• make rs1 a multi-source register.
• make rs2 the rotate ammount.