cmd/compile: use STP/LDP around morestack on arm64

The spill/restore code around morestack is almost never exectued, so
we should make it as small as possible. Using 2-register loads/stores
makes sense here. Also, the offsets from SP are pretty small so the
offset almost always fits in the (smaller than a normal load/store)
offset field of the instruction.

Makes cmd/go 0.6% smaller.

Change-Id: I8845283c1b269a259498153924428f6173bda293
Reviewed-on: https://go-review.googlesource.com/c/go/+/621556
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

src/cmd/compile/internal/arm64/ssa.go

@@ -78,6 +78,48 @@ func storeByType(t *types.Type) obj.As {
 	panic("bad store type")
 }
 
+// loadByType2 returns an opcode that can load consecutive memory locations into 2 registers with type t.
+// returns obj.AXXX if no such opcode exists.
+func loadByType2(t *types.Type) obj.As {
+	if t.IsFloat() {
+		switch t.Size() {
+		case 4:
+			return arm64.AFLDPS
+		case 8:
+			return arm64.AFLDPD
+		}
+	} else {
+		switch t.Size() {
+		case 4:
+			return arm64.ALDPW
+		case 8:
+			return arm64.ALDP
+		}
+	}
+	return obj.AXXX
+}
+
+// storeByType2 returns an opcode that can store registers with type t into 2 consecutive memory locations.
+// returns obj.AXXX if no such opcode exists.
+func storeByType2(t *types.Type) obj.As {
+	if t.IsFloat() {
+		switch t.Size() {
+		case 4:
+			return arm64.AFSTPS
+		case 8:
+			return arm64.AFSTPD
+		}
+	} else {
+		switch t.Size() {
+		case 4:
+			return arm64.ASTPW
+		case 8:
+			return arm64.ASTP
+		}
+	}
+	return obj.AXXX
+}
+
 // makeshift encodes a register shifted by a constant, used as an Offset in Prog.
 func makeshift(v *ssa.Value, reg int16, typ int64, s int64) int64 {
 	if s < 0 || s >= 64 {
@@ -167,17 +209,38 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p.From.Reg = v.Args[0].Reg()
 		ssagen.AddrAuto(&p.To, v)
 	case ssa.OpArgIntReg, ssa.OpArgFloatReg:
+		ssagen.CheckArgReg(v)
 		// The assembler needs to wrap the entry safepoint/stack growth code with spill/unspill
 		// The loop only runs once.
-		for _, a := range v.Block.Func.RegArgs {
-			// Pass the spill/unspill information along to the assembler, offset by size of
-			// the saved LR slot.
-			addr := ssagen.SpillSlotAddr(a, arm64.REGSP, base.Ctxt.Arch.FixedFrameSize)
-			s.FuncInfo().AddSpill(
-				obj.RegSpill{Reg: a.Reg, Addr: addr, Unspill: loadByType(a.Type), Spill: storeByType(a.Type)})
+		args := v.Block.Func.RegArgs
+		if len(args) == 0 {
+			break
 		}
-		v.Block.Func.RegArgs = nil
-		ssagen.CheckArgReg(v)
+		v.Block.Func.RegArgs = nil // prevent from running again
+
+		for i := 0; i < len(args); i++ {
+			a := args[i]
+			// Offset by size of the saved LR slot.
+			addr := ssagen.SpillSlotAddr(a, arm64.REGSP, base.Ctxt.Arch.FixedFrameSize)
+			// Look for double-register operations if we can.
+			if i < len(args)-1 {
+				b := args[i+1]
+				if a.Type.Size() == b.Type.Size() &&
+					a.Type.IsFloat() == b.Type.IsFloat() &&
+					b.Offset == a.Offset+a.Type.Size() {
+					ld := loadByType2(a.Type)
+					st := storeByType2(a.Type)
+					if ld != obj.AXXX && st != obj.AXXX {
+						s.FuncInfo().AddSpill(obj.RegSpill{Reg: a.Reg, Reg2: b.Reg, Addr: addr, Unspill: ld, Spill: st})
+						i++ // b is done also, skip it.
+						continue
+					}
+				}
+			}
+			// Pass the spill/unspill information along to the assembler.
+			s.FuncInfo().AddSpill(obj.RegSpill{Reg: a.Reg, Addr: addr, Unspill: loadByType(a.Type), Spill: storeByType(a.Type)})
+		}
+
 	case ssa.OpARM64ADD,
 		ssa.OpARM64SUB,
 		ssa.OpARM64AND,

src/cmd/internal/obj/link.go

@@ -1100,6 +1100,7 @@ type Auto struct {
 type RegSpill struct {
 	Addr           Addr
 	Reg            int16
+	Reg2           int16 // If not 0, a second register to spill at Addr+regSize. Only for some archs.
 	Spill, Unspill As
 }
 
@@ -1192,6 +1193,10 @@ func (fi *FuncInfo) SpillRegisterArgs(last *Prog, pa ProgAlloc) *Prog {
 		spill.As = ra.Spill
 		spill.From.Type = TYPE_REG
 		spill.From.Reg = ra.Reg
+		if ra.Reg2 != 0 {
+			spill.From.Type = TYPE_REGREG
+			spill.From.Offset = int64(ra.Reg2)
+		}
 		spill.To = ra.Addr
 		last = spill
 	}
@@ -1208,6 +1213,10 @@ func (fi *FuncInfo) UnspillRegisterArgs(last *Prog, pa ProgAlloc) *Prog {
 		unspill.From = ra.Addr
 		unspill.To.Type = TYPE_REG
 		unspill.To.Reg = ra.Reg
+		if ra.Reg2 != 0 {
+			unspill.To.Type = TYPE_REGREG
+			unspill.To.Offset = int64(ra.Reg2)
+		}
 		last = unspill
 	}
 	return last

test/codegen/spills.go

@@ -0,0 +1,31 @@
+// asmcheck
+
+// Copyright 2024 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package codegen
+
+func i64(a, b int64) int64 { // arm64:`STP\s`,`LDP\s`
+	g()
+	return a + b
+}
+
+func i32(a, b int32) int32 { // arm64:`STPW`,`LDPW`
+	g()
+	return a + b
+}
+
+func f64(a, b float64) float64 { // arm64:`FSTPD`,`FLDPD`
+	g()
+	return a + b
+}
+
+func f32(a, b float32) float32 { // arm64:`FSTPS`,`FLDPS`
+	g()
+	return a + b
+}
+
+//go:noinline
+func g() {
+}