cmd/compile: improve register overwrite decision for resultInArg0 ops

When we're compiling a resultInArg0 op, we need to clobber the
register containing the input value. So we first make a register copy
of the input value. We can then clobber either of the two registers
the value is in and still have the original input value in a register
for future uses.

Before this CL, we always clobbered the original, not the copy.
But that's not always the right decision - if the original is already
in a specific register that it needs to be in later (typically, a
return value register), clobber the copy instead.

This optimization can remove a mov instruction. It saves 1376 bytes
of instructions in cmd/go.

Redo of CL 460656, reverted at CL 463475, with a fix for s390x.

The new code just ensures that the copied value is in a register
which is a valid input register for the instruction.

Change-Id: Id570b8a60a6d2da9090de80a90b6bb0266e9e38a
Reviewed-on: https://go-review.googlesource.com/c/go/+/463221
Auto-Submit: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

src/cmd/compile/internal/ssa/regalloc.go

@@ -1544,6 +1544,7 @@ func (s *regAllocState) regalloc(f *Func) {
 						}
 					}
 				}
+
 				// Avoid future fixed uses if we can.
 				if m&^desired.avoid != 0 {
 					m &^= desired.avoid
@@ -1551,6 +1552,21 @@ func (s *regAllocState) regalloc(f *Func) {
 				// Save input 0 to a new register so we can clobber it.
 				c := s.allocValToReg(v.Args[0], m, true, v.Pos)
 				s.copies[c] = false
+
+				// Normally we use the register of the old copy of input 0 as the target.
+				// However, if input 0 is already in its desired register then we use
+				// the register of the new copy instead.
+				if regspec.outputs[0].regs>>s.f.getHome(c.ID).(*Register).num&1 != 0 {
+					if rp, ok := s.f.getHome(args[0].ID).(*Register); ok {
+						r := register(rp.num)
+						for _, r2 := range dinfo[idx].in[0] {
+							if r == r2 {
+								args[0] = c
+								break
+							}
+						}
+					}
+				}
 			}
 
 		ok: