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>

diff --git a/src/cmd/compile/internal/ssa/regalloc.go b/src/cmd/compile/internal/ssa/regalloc.go
index 294c522a90a87073100c63c1612367598b133621..80f6434e76876af7c3543f62c811902a0b276402 100644 (file)
--- a/src/cmd/compile/internal/ssa/regalloc.go
+++ b/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: