cmd/compile: intrinsify Sub64 on mips64

This CL intrinsify Sub64 on mips64.

pkg: math/bits
                  _   sec/op    _   sec/op     vs base               _
Sub-4               2.849n _ 0%   1.948n _ 0%  -31.64% (p=0.000 n=8)
Sub32-4             3.447n _ 0%   3.446n _ 0%        ~ (p=0.982 n=8)
Sub64-4             2.815n _ 0%   1.948n _ 0%  -30.78% (p=0.000 n=8)
Sub64multiple-4     6.124n _ 0%   3.340n _ 0%  -45.46% (p=0.000 n=8)

Change-Id: Ibba91a4350e4a549ae0b60d8cafc4bca05034b84
Reviewed-on: https://go-review.googlesource.com/c/go/+/498497
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Joel Sing <joel@sing.id.au>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Heschi Kreinick <heschi@google.com>

src/cmd/compile/internal/ssa/_gen/MIPS64.rules

@@ -42,6 +42,10 @@
 (Select1 <t> (Add64carry x y c)) =>
 	(OR (SGTU <t> x s:(ADDV <t> x y)) (SGTU <t> s (ADDV <t> s c)))
 
+(Select0 <t> (Sub64borrow x y c)) => (SUBV (SUBV <t> x y) c)
+(Select1 <t> (Sub64borrow x y c)) =>
+	(OR (SGTU <t> s:(SUBV <t> x y) x) (SGTU <t> (SUBV <t> s c) s))
+
 // math package intrinsics
 (Abs ...) => (ABSD ...)
 

src/cmd/compile/internal/ssa/rewriteMIPS64.go

@@ -7353,6 +7353,22 @@ func rewriteValueMIPS64_OpSelect0(v *Value) bool {
 		v.AddArg2(v0, c)
 		return true
 	}
+	// match: (Select0 <t> (Sub64borrow x y c))
+	// result: (SUBV (SUBV <t> x y) c)
+	for {
+		t := v.Type
+		if v_0.Op != OpSub64borrow {
+			break
+		}
+		c := v_0.Args[2]
+		x := v_0.Args[0]
+		y := v_0.Args[1]
+		v.reset(OpMIPS64SUBV)
+		v0 := b.NewValue0(v.Pos, OpMIPS64SUBV, t)
+		v0.AddArg2(x, y)
+		v.AddArg2(v0, c)
+		return true
+	}
 	// match: (Select0 (DIVVU _ (MOVVconst [1])))
 	// result: (MOVVconst [0])
 	for {
@@ -7487,6 +7503,28 @@ func rewriteValueMIPS64_OpSelect1(v *Value) bool {
 		v.AddArg2(v0, v2)
 		return true
 	}
+	// match: (Select1 <t> (Sub64borrow x y c))
+	// result: (OR (SGTU <t> s:(SUBV <t> x y) x) (SGTU <t> (SUBV <t> s c) s))
+	for {
+		t := v.Type
+		if v_0.Op != OpSub64borrow {
+			break
+		}
+		c := v_0.Args[2]
+		x := v_0.Args[0]
+		y := v_0.Args[1]
+		v.reset(OpMIPS64OR)
+		v0 := b.NewValue0(v.Pos, OpMIPS64SGTU, t)
+		s := b.NewValue0(v.Pos, OpMIPS64SUBV, t)
+		s.AddArg2(x, y)
+		v0.AddArg2(s, x)
+		v2 := b.NewValue0(v.Pos, OpMIPS64SGTU, t)
+		v3 := b.NewValue0(v.Pos, OpMIPS64SUBV, t)
+		v3.AddArg2(s, c)
+		v2.AddArg2(v3, s)
+		v.AddArg2(v0, v2)
+		return true
+	}
 	// match: (Select1 (MULVU x (MOVVconst [-1])))
 	// result: (NEGV x)
 	for {

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

@@ -4849,7 +4849,7 @@ func InitTables() {
 		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {
 			return s.newValue3(ssa.OpSub64borrow, types.NewTuple(types.Types[types.TUINT64], types.Types[types.TUINT64]), args[0], args[1], args[2])
 		},
-		sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64)
+		sys.AMD64, sys.ARM64, sys.PPC64, sys.S390X, sys.RISCV64, sys.Loong64, sys.MIPS64)
 	alias("math/bits", "Sub", "math/bits", "Sub64", p8...)
 	addF("math/bits", "Div64",
 		func(s *state, n *ir.CallExpr, args []*ssa.Value) *ssa.Value {

test/codegen/mathbits.go

@@ -636,6 +636,7 @@ func Sub(x, y, ci uint) (r, co uint) {
 	// loong64:"SUBV","SGTU"
 	// ppc64x:"SUBC", "SUBE", "SUBZE", "NEG"
 	// s390x:"SUBE"
+	// mips64:"SUBV","SGTU"
 	// riscv64: "SUB","SLTU"
 	return bits.Sub(x, y, ci)
 }
@@ -646,6 +647,7 @@ func SubC(x, ci uint) (r, co uint) {
 	// loong64:"SUBV","SGTU"
 	// ppc64x:"SUBC", "SUBE", "SUBZE", "NEG"
 	// s390x:"SUBE"
+	// mips64:"SUBV","SGTU"
 	// riscv64: "SUB","SLTU"
 	return bits.Sub(x, 7, ci)
 }
@@ -656,6 +658,7 @@ func SubZ(x, y uint) (r, co uint) {
 	// loong64:"SUBV","SGTU"
 	// ppc64x:"SUBC", -"SUBE", "SUBZE", "NEG"
 	// s390x:"SUBC"
+	// mips64:"SUBV","SGTU"
 	// riscv64: "SUB","SLTU"
 	return bits.Sub(x, y, 0)
 }
@@ -687,6 +690,7 @@ func Sub64(x, y, ci uint64) (r, co uint64) {
 	// loong64:"SUBV","SGTU"
 	// ppc64x:"SUBC", "SUBE", "SUBZE", "NEG"
 	// s390x:"SUBE"
+	// mips64:"SUBV","SGTU"
 	// riscv64: "SUB","SLTU"
 	return bits.Sub64(x, y, ci)
 }
@@ -697,6 +701,7 @@ func Sub64C(x, ci uint64) (r, co uint64) {
 	// loong64:"SUBV","SGTU"
 	// ppc64x:"SUBC", "SUBE", "SUBZE", "NEG"
 	// s390x:"SUBE"
+	// mips64:"SUBV","SGTU"
 	// riscv64: "SUB","SLTU"
 	return bits.Sub64(x, 7, ci)
 }
@@ -707,6 +712,7 @@ func Sub64Z(x, y uint64) (r, co uint64) {
 	// loong64:"SUBV","SGTU"
 	// ppc64x:"SUBC", -"SUBE", "SUBZE", "NEG"
 	// s390x:"SUBC"
+	// mips64:"SUBV","SGTU"
 	// riscv64: "SUB","SLTU"
 	return bits.Sub64(x, y, 0)
 }