Address folding is disabled in CL42172, the commit message of which
said that "In shared library, load/store of global is rewritten to
using GOT and temp register, which conflicts with the use of temp
register for assembling large offset.". Actually this doesn't happen
because the sequence of instructions when rewritten to use Got looks
like this:
  MOVD $sym, Rx becomes
  MOVD sym@GOT, Rx
If there is an offset off, there will be one more instruction:
  ADD $off, Rx, Rx

And MOVD sym, Rx becomes
  MOVD sym@GOT, REGTMP
  MOVx (REGTMP), Ry

If there is a small offset off, it becomes:
  MOVD sym@GOT, REGTMP
  MOVx (REGTMP)off, Ry

If off is very large, it becomes:
  MOVD sym@GOT, REGTMP
  MOVD $off, Rt
  ADD  Rt, REGTMP
  MOVx (REGTMP), Ry

We can see that the address can be calculated correctly, and testing
on darwin/arm64 confirms this.

Removing this restriction is beneficial to further optimize the sequence
of "ADRP+ADD+LD/ST" to "ADRP+LD/ST(offset), so this CL removes it.

Change-Id: I0e9f7bc1723e0a027f32cf0ae2c41cd6df49defe
Reviewed-on: https://go-review.googlesource.com/c/go/+/445535
Reviewed-by: Cherry Mui <[email protected]>
Run-TryBot: Eric Fang <[email protected]>
Reviewed-by: Heschi Kreinick <[email protected]>
TryBot-Result: Gopher Robot <[email protected]>