cmd/compile: aggressive IMUL rewrites on AMD64 sometimes lead to performance regressions

[quasilyte]

image/draw

name        old time/op  new time/op  delta
CopyOver-4  1.18ms ± 0%  1.52ms ± 0%  +28.49%  (p=0.008 n=5+5)

Bisecting from tip leads to b1df8d6.
This is regression from go1.10.2.
Performance delta is shown for b1df8d6 itself and it's parent (previous commit).

Single IMUL is rewritten into mov+shift+lea.
I can't reliably reproduce those results in free context, but they're stable for CopyOver.

[josharian]

Duplicate of #21735?

[quasilyte]

Maybe.

I'm reporting regression and it hopefully can be fixed.
Shouldn't be unnoticed.

At the very least it adds related benchmark that can be used to check the results.

[bradfitz]

/cc @randall77