cmd/compile/internal/ssa: correct MOVDnop handling for arm64

The extension-removing rules for ARM64 were moved to late lower in
CL 568616. This means that the late lower pass can now generate
MOVDreg, however the rules that potentially eliminate MOVDreg only
exist in the earlier pass. Fix this by duplicating the MOVDreg/NOVDnop
rules in late lower, such that we can potentially eliminate conversions.

Removes 400+ instructions from the Go binary on openbsd/arm64.

Change-Id: I14aad06b994c9179f3ecdda566629793ba167511
Reviewed-on: https://go-review.googlesource.com/c/go/+/651819
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Keith Randall <[email protected]>
Reviewed-by: Dmitri Shuralyov <[email protected]>
Auto-Submit: Dmitri Shuralyov <[email protected]>
Reviewed-by: Keith Randall <[email protected]>

Author: 4a6f656c

src/cmd/compile/internal/ssa/_gen/ARM64latelower.rules

@@ -85,3 +85,11 @@
 (MOVWUreg x:(MOVBUreg _)) => (MOVDreg x)
 (MOVWUreg x:(MOVHUreg _)) => (MOVDreg x)
 (MOVWUreg x:(MOVWUreg _)) => (MOVDreg x)
+
+// if a register move has only 1 use, just use the same register without emitting instruction
+// MOVDnop doesn't emit instruction, only for ensuring the type.
+(MOVDreg x) && x.Uses == 1 => (MOVDnop x)
+
+// TODO: we should be able to get rid of MOVDnop all together.
+// But for now, this is enough to get rid of lots of them.
+(MOVDnop (MOVDconst [c])) => (MOVDconst [c])