cmd/compile: regalloc fails to reserve key register (CX) for future code

[josharian]

After CLs 297049 and 297050, generate code for shrVU_g in math/big.

The tight inner loop compiles to:

v94     00020 (+182) MOVQ -8(DX)(CX*8), R8
v106    00021 (182) MOVQ (DX)(CX*8), R9
v133    00022 (181) MOVQ CX, R10
v117    00023 (182) MOVQ SI, CX
v109    00024 (182) SHRQ CX, R9, R8
v119    00025 (182) MOVQ R8, -8(DI)(R10*8)
v121    00026 (+181) LEAQ 1(R10), R8
v107    00027 (181) MOVQ R8, CX

This contains unnecessary copies. One is #44670.

The others are CX contention. Regalloc decided to put i in CX. Unfortunately, CX is the required register for the shift amount input for the SHRDQ instruction. So we shuffle i in and out of CX on every iteration.

I thought regalloc had some amount of lookahead to avoid using highly restricted registers in upcoming code, to avoid this situation. Either I was wrong, or it isn't working as desired here.

I believe that fixing this issue would make the generated code as fast (or nearly so) as the hand-written assembly, allowing us to delete the assembly.

cc @randall77 @cherrymui