cmd/compile: simplify div/mod on ARM

On ARM, DIV, DIVU, MOD, MODU are pseudo instructions that makes
runtime calls _div/_udiv/_mod/_umod, which themselves are wrappers
of udiv. The udiv function does the real thing.

Instead of generating these pseudo instructions, call to udiv
directly. This removes one layer of wrappers (which has an awkward
way of passing argument), and also allows combining DIV and MOD
if both results are needed.

Change-Id: I118afc3986db3a1daabb5c1e6e57430888c91817
Reviewed-on: https://go-review.googlesource.com/29390Reviewed-by: David Chase <drchase@google.com>

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

@@ -196,21 +196,11 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		} else {
 			p.To.Name = obj.NAME_AUTO
 		}
-	case ssa.OpARMDIV,
-		ssa.OpARMDIVU,
-		ssa.OpARMMOD,
-		ssa.OpARMMODU:
-		// Note: for software division the assembler rewrite these
-		// instructions to sequence of instructions:
-		// - it puts numerator in R11 and denominator in g.m.divmod
-		//	and call (say) _udiv
-		// - _udiv saves R0-R3 on stack and call udiv, restores R0-R3
-		//	before return
-		// - udiv does the actual work
-		//TODO: set approperiate regmasks and call udiv directly?
-		// need to be careful for negative case
-		// Or, as soft div is already expensive, we don't care?
-		fallthrough
+	case ssa.OpARMUDIVrtcall:
+		p := gc.Prog(obj.ACALL)
+		p.To.Type = obj.TYPE_MEM
+		p.To.Name = obj.NAME_EXTERN
+		p.To.Sym = obj.Linklookup(gc.Ctxt, "udiv", 0)
 	case ssa.OpARMADD,
 		ssa.OpARMADC,
 		ssa.OpARMSUB,

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

@@ -230,12 +230,8 @@ func NewConfig(arch string, fe Frontend, ctxt *obj.Link, optimize bool) *Config
 	if c.nacl {
 		c.noDuffDevice = true // Don't use Duff's device on NaCl
 
-		// ARM assembler rewrites DIV/MOD to runtime calls, which
-		// clobber R12 on nacl
-		opcodeTable[OpARMDIV].reg.clobbers |= 1 << 12  // R12
-		opcodeTable[OpARMDIVU].reg.clobbers |= 1 << 12 // R12
-		opcodeTable[OpARMMOD].reg.clobbers |= 1 << 12  // R12
-		opcodeTable[OpARMMODU].reg.clobbers |= 1 << 12 // R12
+		// runtime call clobber R12 on nacl
+		opcodeTable[OpARMUDIVrtcall].reg.clobbers |= 1 << 12 // R12
 	}
 
 	// Assign IDs to preallocated values/blocks.

src/cmd/compile/internal/ssa/gen/ARM.rules

@@ -37,21 +37,31 @@
 
 (Mul32uhilo x y) -> (MULLU x y)
 
-(Div32 x y) -> (DIV x y)
-(Div32u x y) -> (DIVU x y)
-(Div16 x y) -> (DIV (SignExt16to32 x) (SignExt16to32 y))
-(Div16u x y) -> (DIVU (ZeroExt16to32 x) (ZeroExt16to32 y))
-(Div8 x y) -> (DIV (SignExt8to32 x) (SignExt8to32 y))
-(Div8u x y) -> (DIVU (ZeroExt8to32 x) (ZeroExt8to32 y))
+(Div32 x y) ->
+	(SUB (XOR <config.fe.TypeUInt32()>                                                                  // negate the result if one operand is negative
+		(Select0 <config.fe.TypeUInt32()> (UDIVrtcall
+			(SUB <config.fe.TypeUInt32()> (XOR x <config.fe.TypeUInt32()> (Signmask x)) (Signmask x))   // negate x if negative
+			(SUB <config.fe.TypeUInt32()> (XOR y <config.fe.TypeUInt32()> (Signmask y)) (Signmask y)))) // negate y if negative
+		(Signmask (XOR <config.fe.TypeUInt32()> x y))) (Signmask (XOR <config.fe.TypeUInt32()> x y)))
+(Div32u x y) -> (Select0 <config.fe.TypeUInt32()> (UDIVrtcall x y))
+(Div16 x y) -> (Div32 (SignExt16to32 x) (SignExt16to32 y))
+(Div16u x y) -> (Div32u (ZeroExt16to32 x) (ZeroExt16to32 y))
+(Div8 x y) -> (Div32 (SignExt8to32 x) (SignExt8to32 y))
+(Div8u x y) -> (Div32u (ZeroExt8to32 x) (ZeroExt8to32 y))
 (Div32F x y) -> (DIVF x y)
 (Div64F x y) -> (DIVD x y)
 
-(Mod32 x y) -> (MOD x y)
-(Mod32u x y) -> (MODU x y)
-(Mod16 x y) -> (MOD (SignExt16to32 x) (SignExt16to32 y))
-(Mod16u x y) -> (MODU (ZeroExt16to32 x) (ZeroExt16to32 y))
-(Mod8 x y) -> (MOD (SignExt8to32 x) (SignExt8to32 y))
-(Mod8u x y) -> (MODU (ZeroExt8to32 x) (ZeroExt8to32 y))
+(Mod32 x y) ->
+	(SUB (XOR <config.fe.TypeUInt32()>                                                                  // negate the result if x is negative
+		(Select1 <config.fe.TypeUInt32()> (UDIVrtcall
+			(SUB <config.fe.TypeUInt32()> (XOR <config.fe.TypeUInt32()> x (Signmask x)) (Signmask x))   // negate x if negative
+			(SUB <config.fe.TypeUInt32()> (XOR <config.fe.TypeUInt32()> y (Signmask y)) (Signmask y)))) // negate y if negative
+		(Signmask x)) (Signmask x))
+(Mod32u x y) -> (Select1 <config.fe.TypeUInt32()> (UDIVrtcall x y))
+(Mod16 x y) -> (Mod32 (SignExt16to32 x) (SignExt16to32 y))
+(Mod16u x y) -> (Mod32u (ZeroExt16to32 x) (ZeroExt16to32 y))
+(Mod8 x y) -> (Mod32 (SignExt8to32 x) (SignExt8to32 y))
+(Mod8u x y) -> (Mod32u (ZeroExt8to32 x) (ZeroExt8to32 y))
 
 (And32 x y) -> (AND x y)
 (And16 x y) -> (AND x y)
@@ -586,8 +596,10 @@
 (MULA (MOVWconst [c]) x a) && c%9 == 0 && isPowerOfTwo(c/9) && is32Bit(c) -> (ADD (SLLconst <x.Type> [log2(c/9)] (ADDshiftLL <x.Type> x x [3])) a)
 
 // div by constant
-(DIVU x (MOVWconst [1])) -> x
-(DIVU x (MOVWconst [c])) && isPowerOfTwo(c) -> (SRLconst [log2(c)] x)
+(Select0 (UDIVrtcall x (MOVWconst [1]))) -> x
+(Select1 (UDIVrtcall _ (MOVWconst [1]))) -> (MOVWconst [0])
+(Select0 (UDIVrtcall x (MOVWconst [c]))) && isPowerOfTwo(c) -> (SRLconst [log2(c)] x)
+(Select1 (UDIVrtcall x (MOVWconst [c]))) && isPowerOfTwo(c) -> (ANDconst [c-1] x)
 
 // constant comparisons
 (CMPconst (MOVWconst [x]) [y]) && int32(x)==int32(y) -> (FlagEQ)
@@ -805,8 +817,8 @@
 (SRAconst [c] (MOVWconst [d])) -> (MOVWconst [int64(int32(d)>>uint64(c))])
 (MUL (MOVWconst [c]) (MOVWconst [d])) -> (MOVWconst [int64(int32(c*d))])
 (MULA (MOVWconst [c]) (MOVWconst [d]) a) -> (ADDconst [int64(int32(c*d))] a)
-(DIV (MOVWconst [c]) (MOVWconst [d])) -> (MOVWconst [int64(int32(c)/int32(d))])
-(DIVU (MOVWconst [c]) (MOVWconst [d])) -> (MOVWconst [int64(uint32(c)/uint32(d))])
+(Select0 (UDIVrtcall (MOVWconst [c]) (MOVWconst [d]))) -> (MOVWconst [int64(uint32(c)/uint32(d))])
+(Select1 (UDIVrtcall (MOVWconst [c]) (MOVWconst [d]))) -> (MOVWconst [int64(uint32(c)%uint32(d))])
 (ANDconst [c] (MOVWconst [d])) -> (MOVWconst [c&d])
 (ANDconst [c] (ANDconst [d] x)) -> (ANDconst [c&d] x)
 (ORconst [c] (MOVWconst [d])) -> (MOVWconst [c|d])

src/cmd/compile/internal/ssa/gen/ARMOps.go

@@ -138,10 +138,21 @@ func init() {
 		{name: "MUL", argLength: 2, reg: gp21, asm: "MUL", commutative: true},     // arg0 * arg1
 		{name: "HMUL", argLength: 2, reg: gp21, asm: "MULL", commutative: true},   // (arg0 * arg1) >> 32, signed
 		{name: "HMULU", argLength: 2, reg: gp21, asm: "MULLU", commutative: true}, // (arg0 * arg1) >> 32, unsigned
-		{name: "DIV", argLength: 2, reg: gp21, asm: "DIV", clobberFlags: true},    // arg0 / arg1, signed, soft div clobbers flags
-		{name: "DIVU", argLength: 2, reg: gp21, asm: "DIVU", clobberFlags: true},  // arg0 / arg1, unsighed
-		{name: "MOD", argLength: 2, reg: gp21, asm: "MOD", clobberFlags: true},    // arg0 % arg1, signed
-		{name: "MODU", argLength: 2, reg: gp21, asm: "MODU", clobberFlags: true},  // arg0 % arg1, unsigned
+
+		// udiv runtime call for soft division
+		// output0 = arg0/arg1, output1 = arg0%arg1
+		// see ../../../../../runtime/vlop_arm.s
+		{
+			name:      "UDIVrtcall",
+			argLength: 2,
+			reg: regInfo{
+				inputs:   []regMask{buildReg("R1"), buildReg("R0")},
+				outputs:  []regMask{buildReg("R0"), buildReg("R1")},
+				clobbers: buildReg("R2 R3"), // also clobbers R12 on NaCl (modified in ../config.go)
+			},
+			clobberFlags: true,
+			typ:          "(UInt32,UInt32)",
+		},
 
 		{name: "ADDS", argLength: 2, reg: gp21carry, asm: "ADD", commutative: true}, // arg0 + arg1, set carry flag
 		{name: "ADDSconst", argLength: 1, reg: gp11carry, asm: "ADD", aux: "Int32"}, // arg0 + auxInt, set carry flag

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

@@ -623,10 +623,7 @@ const (
 	OpARMMUL
 	OpARMHMUL
 	OpARMHMULU
-	OpARMDIV
-	OpARMDIVU
-	OpARMMOD
-	OpARMMODU
+	OpARMUDIVrtcall
 	OpARMADDS
 	OpARMADDSconst
 	OpARMADC
@@ -7367,62 +7364,18 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
-		name:         "DIV",
+		name:         "UDIVrtcall",
 		argLen:       2,
 		clobberFlags: true,
-		asm:          arm.ADIV,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-				{1, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-			},
-			outputs: []outputInfo{
-				{0, 5119}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12
-			},
-		},
-	},
-	{
-		name:         "DIVU",
-		argLen:       2,
-		clobberFlags: true,
-		asm:          arm.ADIVU,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{0, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-				{1, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-			},
-			outputs: []outputInfo{
-				{0, 5119}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12
-			},
-		},
-	},
-	{
-		name:         "MOD",
-		argLen:       2,
-		clobberFlags: true,
-		asm:          arm.AMOD,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{0, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-				{1, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-			},
-			outputs: []outputInfo{
-				{0, 5119}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12
-			},
-		},
-	},
-	{
-		name:         "MODU",
-		argLen:       2,
-		clobberFlags: true,
-		asm:          arm.AMODU,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{0, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
-				{1, 6143}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12
+				{0, 2}, // R1
+				{1, 1}, // R0
 			},
+			clobbers: 12, // R2 R3
 			outputs: []outputInfo{
-				{0, 5119}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12
+				{0, 1}, // R0
+				{1, 2}, // R1
 			},
 		},
 	},

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

@@ -110,10 +110,6 @@ func rewriteValueARM(v *Value, config *Config) bool {
 		return rewriteValueARM_OpARMCMPshiftRL(v, config)
 	case OpARMCMPshiftRLreg:
 		return rewriteValueARM_OpARMCMPshiftRLreg(v, config)
-	case OpARMDIV:
-		return rewriteValueARM_OpARMDIV(v, config)
-	case OpARMDIVU:
-		return rewriteValueARM_OpARMDIVU(v, config)
 	case OpARMEqual:
 		return rewriteValueARM_OpARMEqual(v, config)
 	case OpARMGreaterEqual:
@@ -676,6 +672,10 @@ func rewriteValueARM(v *Value, config *Config) bool {
 		return rewriteValueARM_OpRsh8x64(v, config)
 	case OpRsh8x8:
 		return rewriteValueARM_OpRsh8x8(v, config)
+	case OpSelect0:
+		return rewriteValueARM_OpSelect0(v, config)
+	case OpSelect1:
+		return rewriteValueARM_OpSelect1(v, config)
 	case OpSignExt16to32:
 		return rewriteValueARM_OpSignExt16to32(v, config)
 	case OpSignExt8to16:
@@ -4436,87 +4436,6 @@ func rewriteValueARM_OpARMCMPshiftRLreg(v *Value, config *Config) bool {
 	}
 	return false
 }
-func rewriteValueARM_OpARMDIV(v *Value, config *Config) bool {
-	b := v.Block
-	_ = b
-	// match: (DIV (MOVWconst [c]) (MOVWconst [d]))
-	// cond:
-	// result: (MOVWconst [int64(int32(c)/int32(d))])
-	for {
-		v_0 := v.Args[0]
-		if v_0.Op != OpARMMOVWconst {
-			break
-		}
-		c := v_0.AuxInt
-		v_1 := v.Args[1]
-		if v_1.Op != OpARMMOVWconst {
-			break
-		}
-		d := v_1.AuxInt
-		v.reset(OpARMMOVWconst)
-		v.AuxInt = int64(int32(c) / int32(d))
-		return true
-	}
-	return false
-}
-func rewriteValueARM_OpARMDIVU(v *Value, config *Config) bool {
-	b := v.Block
-	_ = b
-	// match: (DIVU x (MOVWconst [1]))
-	// cond:
-	// result: x
-	for {
-		x := v.Args[0]
-		v_1 := v.Args[1]
-		if v_1.Op != OpARMMOVWconst {
-			break
-		}
-		if v_1.AuxInt != 1 {
-			break
-		}
-		v.reset(OpCopy)
-		v.Type = x.Type
-		v.AddArg(x)
-		return true
-	}
-	// match: (DIVU x (MOVWconst [c]))
-	// cond: isPowerOfTwo(c)
-	// result: (SRLconst [log2(c)] x)
-	for {
-		x := v.Args[0]
-		v_1 := v.Args[1]
-		if v_1.Op != OpARMMOVWconst {
-			break
-		}
-		c := v_1.AuxInt
-		if !(isPowerOfTwo(c)) {
-			break
-		}
-		v.reset(OpARMSRLconst)
-		v.AuxInt = log2(c)
-		v.AddArg(x)
-		return true
-	}
-	// match: (DIVU (MOVWconst [c]) (MOVWconst [d]))
-	// cond:
-	// result: (MOVWconst [int64(uint32(c)/uint32(d))])
-	for {
-		v_0 := v.Args[0]
-		if v_0.Op != OpARMMOVWconst {
-			break
-		}
-		c := v_0.AuxInt
-		v_1 := v.Args[1]
-		if v_1.Op != OpARMMOVWconst {
-			break
-		}
-		d := v_1.AuxInt
-		v.reset(OpARMMOVWconst)
-		v.AuxInt = int64(uint32(c) / uint32(d))
-		return true
-	}
-	return false
-}
 func rewriteValueARM_OpARMEqual(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b
@@ -13347,11 +13266,11 @@ func rewriteValueARM_OpDiv16(v *Value, config *Config) bool {
 	_ = b
 	// match: (Div16 x y)
 	// cond:
-	// result: (DIV (SignExt16to32 x) (SignExt16to32 y))
+	// result: (Div32 (SignExt16to32 x) (SignExt16to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMDIV)
+		v.reset(OpDiv32)
 		v0 := b.NewValue0(v.Line, OpSignExt16to32, config.fe.TypeInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -13366,11 +13285,11 @@ func rewriteValueARM_OpDiv16u(v *Value, config *Config) bool {
 	_ = b
 	// match: (Div16u x y)
 	// cond:
-	// result: (DIVU (ZeroExt16to32 x) (ZeroExt16to32 y))
+	// result: (Div32u (ZeroExt16to32 x) (ZeroExt16to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMDIVU)
+		v.reset(OpDiv32u)
 		v0 := b.NewValue0(v.Line, OpZeroExt16to32, config.fe.TypeUInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -13385,13 +13304,51 @@ func rewriteValueARM_OpDiv32(v *Value, config *Config) bool {
 	_ = b
 	// match: (Div32 x y)
 	// cond:
-	// result: (DIV x y)
+	// result: (SUB (XOR <config.fe.TypeUInt32()> 		(Select0 <config.fe.TypeUInt32()> (UDIVrtcall 			(SUB <config.fe.TypeUInt32()> (XOR x <config.fe.TypeUInt32()> (Signmask x)) (Signmask x)) 			(SUB <config.fe.TypeUInt32()> (XOR y <config.fe.TypeUInt32()> (Signmask y)) (Signmask y)))) 		(Signmask (XOR <config.fe.TypeUInt32()> x y))) (Signmask (XOR <config.fe.TypeUInt32()> x y)))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMDIV)
-		v.AddArg(x)
-		v.AddArg(y)
+		v.reset(OpARMSUB)
+		v0 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v1 := b.NewValue0(v.Line, OpSelect0, config.fe.TypeUInt32())
+		v2 := b.NewValue0(v.Line, OpARMUDIVrtcall, MakeTuple(config.fe.TypeUInt32(), config.fe.TypeUInt32()))
+		v3 := b.NewValue0(v.Line, OpARMSUB, config.fe.TypeUInt32())
+		v4 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v4.AddArg(x)
+		v5 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v5.AddArg(x)
+		v4.AddArg(v5)
+		v3.AddArg(v4)
+		v6 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v6.AddArg(x)
+		v3.AddArg(v6)
+		v2.AddArg(v3)
+		v7 := b.NewValue0(v.Line, OpARMSUB, config.fe.TypeUInt32())
+		v8 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v8.AddArg(y)
+		v9 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v9.AddArg(y)
+		v8.AddArg(v9)
+		v7.AddArg(v8)
+		v10 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v10.AddArg(y)
+		v7.AddArg(v10)
+		v2.AddArg(v7)
+		v1.AddArg(v2)
+		v0.AddArg(v1)
+		v11 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v12 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v12.AddArg(x)
+		v12.AddArg(y)
+		v11.AddArg(v12)
+		v0.AddArg(v11)
+		v.AddArg(v0)
+		v13 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v14 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v14.AddArg(x)
+		v14.AddArg(y)
+		v13.AddArg(v14)
+		v.AddArg(v13)
 		return true
 	}
 }
@@ -13415,13 +13372,16 @@ func rewriteValueARM_OpDiv32u(v *Value, config *Config) bool {
 	_ = b
 	// match: (Div32u x y)
 	// cond:
-	// result: (DIVU x y)
+	// result: (Select0 <config.fe.TypeUInt32()> (UDIVrtcall x y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMDIVU)
-		v.AddArg(x)
-		v.AddArg(y)
+		v.reset(OpSelect0)
+		v.Type = config.fe.TypeUInt32()
+		v0 := b.NewValue0(v.Line, OpARMUDIVrtcall, MakeTuple(config.fe.TypeUInt32(), config.fe.TypeUInt32()))
+		v0.AddArg(x)
+		v0.AddArg(y)
+		v.AddArg(v0)
 		return true
 	}
 }
@@ -13445,11 +13405,11 @@ func rewriteValueARM_OpDiv8(v *Value, config *Config) bool {
 	_ = b
 	// match: (Div8 x y)
 	// cond:
-	// result: (DIV (SignExt8to32 x) (SignExt8to32 y))
+	// result: (Div32 (SignExt8to32 x) (SignExt8to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMDIV)
+		v.reset(OpDiv32)
 		v0 := b.NewValue0(v.Line, OpSignExt8to32, config.fe.TypeInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -13464,11 +13424,11 @@ func rewriteValueARM_OpDiv8u(v *Value, config *Config) bool {
 	_ = b
 	// match: (Div8u x y)
 	// cond:
-	// result: (DIVU (ZeroExt8to32 x) (ZeroExt8to32 y))
+	// result: (Div32u (ZeroExt8to32 x) (ZeroExt8to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMDIVU)
+		v.reset(OpDiv32u)
 		v0 := b.NewValue0(v.Line, OpZeroExt8to32, config.fe.TypeUInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -14926,11 +14886,11 @@ func rewriteValueARM_OpMod16(v *Value, config *Config) bool {
 	_ = b
 	// match: (Mod16 x y)
 	// cond:
-	// result: (MOD (SignExt16to32 x) (SignExt16to32 y))
+	// result: (Mod32 (SignExt16to32 x) (SignExt16to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMMOD)
+		v.reset(OpMod32)
 		v0 := b.NewValue0(v.Line, OpSignExt16to32, config.fe.TypeInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -14945,11 +14905,11 @@ func rewriteValueARM_OpMod16u(v *Value, config *Config) bool {
 	_ = b
 	// match: (Mod16u x y)
 	// cond:
-	// result: (MODU (ZeroExt16to32 x) (ZeroExt16to32 y))
+	// result: (Mod32u (ZeroExt16to32 x) (ZeroExt16to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMMODU)
+		v.reset(OpMod32u)
 		v0 := b.NewValue0(v.Line, OpZeroExt16to32, config.fe.TypeUInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -14964,13 +14924,45 @@ func rewriteValueARM_OpMod32(v *Value, config *Config) bool {
 	_ = b
 	// match: (Mod32 x y)
 	// cond:
-	// result: (MOD x y)
+	// result: (SUB (XOR <config.fe.TypeUInt32()> 		(Select1 <config.fe.TypeUInt32()> (UDIVrtcall 			(SUB <config.fe.TypeUInt32()> (XOR <config.fe.TypeUInt32()> x (Signmask x)) (Signmask x)) 			(SUB <config.fe.TypeUInt32()> (XOR <config.fe.TypeUInt32()> y (Signmask y)) (Signmask y)))) 		(Signmask x)) (Signmask x))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMMOD)
-		v.AddArg(x)
-		v.AddArg(y)
+		v.reset(OpARMSUB)
+		v0 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v1 := b.NewValue0(v.Line, OpSelect1, config.fe.TypeUInt32())
+		v2 := b.NewValue0(v.Line, OpARMUDIVrtcall, MakeTuple(config.fe.TypeUInt32(), config.fe.TypeUInt32()))
+		v3 := b.NewValue0(v.Line, OpARMSUB, config.fe.TypeUInt32())
+		v4 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v4.AddArg(x)
+		v5 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v5.AddArg(x)
+		v4.AddArg(v5)
+		v3.AddArg(v4)
+		v6 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v6.AddArg(x)
+		v3.AddArg(v6)
+		v2.AddArg(v3)
+		v7 := b.NewValue0(v.Line, OpARMSUB, config.fe.TypeUInt32())
+		v8 := b.NewValue0(v.Line, OpARMXOR, config.fe.TypeUInt32())
+		v8.AddArg(y)
+		v9 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v9.AddArg(y)
+		v8.AddArg(v9)
+		v7.AddArg(v8)
+		v10 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v10.AddArg(y)
+		v7.AddArg(v10)
+		v2.AddArg(v7)
+		v1.AddArg(v2)
+		v0.AddArg(v1)
+		v11 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v11.AddArg(x)
+		v0.AddArg(v11)
+		v.AddArg(v0)
+		v12 := b.NewValue0(v.Line, OpSignmask, config.fe.TypeInt32())
+		v12.AddArg(x)
+		v.AddArg(v12)
 		return true
 	}
 }
@@ -14979,13 +14971,16 @@ func rewriteValueARM_OpMod32u(v *Value, config *Config) bool {
 	_ = b
 	// match: (Mod32u x y)
 	// cond:
-	// result: (MODU x y)
+	// result: (Select1 <config.fe.TypeUInt32()> (UDIVrtcall x y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMMODU)
-		v.AddArg(x)
-		v.AddArg(y)
+		v.reset(OpSelect1)
+		v.Type = config.fe.TypeUInt32()
+		v0 := b.NewValue0(v.Line, OpARMUDIVrtcall, MakeTuple(config.fe.TypeUInt32(), config.fe.TypeUInt32()))
+		v0.AddArg(x)
+		v0.AddArg(y)
+		v.AddArg(v0)
 		return true
 	}
 }
@@ -14994,11 +14989,11 @@ func rewriteValueARM_OpMod8(v *Value, config *Config) bool {
 	_ = b
 	// match: (Mod8 x y)
 	// cond:
-	// result: (MOD (SignExt8to32 x) (SignExt8to32 y))
+	// result: (Mod32 (SignExt8to32 x) (SignExt8to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMMOD)
+		v.reset(OpMod32)
 		v0 := b.NewValue0(v.Line, OpSignExt8to32, config.fe.TypeInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -15013,11 +15008,11 @@ func rewriteValueARM_OpMod8u(v *Value, config *Config) bool {
 	_ = b
 	// match: (Mod8u x y)
 	// cond:
-	// result: (MODU (ZeroExt8to32 x) (ZeroExt8to32 y))
+	// result: (Mod32u (ZeroExt8to32 x) (ZeroExt8to32 y))
 	for {
 		x := v.Args[0]
 		y := v.Args[1]
-		v.reset(OpARMMODU)
+		v.reset(OpMod32u)
 		v0 := b.NewValue0(v.Line, OpZeroExt8to32, config.fe.TypeUInt32())
 		v0.AddArg(x)
 		v.AddArg(v0)
@@ -16344,6 +16339,144 @@ func rewriteValueARM_OpRsh8x8(v *Value, config *Config) bool {
 		return true
 	}
 }
+func rewriteValueARM_OpSelect0(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (Select0 (UDIVrtcall x (MOVWconst [1])))
+	// cond:
+	// result: x
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMUDIVrtcall {
+			break
+		}
+		x := v_0.Args[0]
+		v_0_1 := v_0.Args[1]
+		if v_0_1.Op != OpARMMOVWconst {
+			break
+		}
+		if v_0_1.AuxInt != 1 {
+			break
+		}
+		v.reset(OpCopy)
+		v.Type = x.Type
+		v.AddArg(x)
+		return true
+	}
+	// match: (Select0 (UDIVrtcall x (MOVWconst [c])))
+	// cond: isPowerOfTwo(c)
+	// result: (SRLconst [log2(c)] x)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMUDIVrtcall {
+			break
+		}
+		x := v_0.Args[0]
+		v_0_1 := v_0.Args[1]
+		if v_0_1.Op != OpARMMOVWconst {
+			break
+		}
+		c := v_0_1.AuxInt
+		if !(isPowerOfTwo(c)) {
+			break
+		}
+		v.reset(OpARMSRLconst)
+		v.AuxInt = log2(c)
+		v.AddArg(x)
+		return true
+	}
+	// match: (Select0 (UDIVrtcall (MOVWconst [c]) (MOVWconst [d])))
+	// cond:
+	// result: (MOVWconst [int64(uint32(c)/uint32(d))])
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMUDIVrtcall {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpARMMOVWconst {
+			break
+		}
+		c := v_0_0.AuxInt
+		v_0_1 := v_0.Args[1]
+		if v_0_1.Op != OpARMMOVWconst {
+			break
+		}
+		d := v_0_1.AuxInt
+		v.reset(OpARMMOVWconst)
+		v.AuxInt = int64(uint32(c) / uint32(d))
+		return true
+	}
+	return false
+}
+func rewriteValueARM_OpSelect1(v *Value, config *Config) bool {
+	b := v.Block
+	_ = b
+	// match: (Select1 (UDIVrtcall _ (MOVWconst [1])))
+	// cond:
+	// result: (MOVWconst [0])
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMUDIVrtcall {
+			break
+		}
+		v_0_1 := v_0.Args[1]
+		if v_0_1.Op != OpARMMOVWconst {
+			break
+		}
+		if v_0_1.AuxInt != 1 {
+			break
+		}
+		v.reset(OpARMMOVWconst)
+		v.AuxInt = 0
+		return true
+	}
+	// match: (Select1 (UDIVrtcall x (MOVWconst [c])))
+	// cond: isPowerOfTwo(c)
+	// result: (ANDconst [c-1] x)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMUDIVrtcall {
+			break
+		}
+		x := v_0.Args[0]
+		v_0_1 := v_0.Args[1]
+		if v_0_1.Op != OpARMMOVWconst {
+			break
+		}
+		c := v_0_1.AuxInt
+		if !(isPowerOfTwo(c)) {
+			break
+		}
+		v.reset(OpARMANDconst)
+		v.AuxInt = c - 1
+		v.AddArg(x)
+		return true
+	}
+	// match: (Select1 (UDIVrtcall (MOVWconst [c]) (MOVWconst [d])))
+	// cond:
+	// result: (MOVWconst [int64(uint32(c)%uint32(d))])
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMUDIVrtcall {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpARMMOVWconst {
+			break
+		}
+		c := v_0_0.AuxInt
+		v_0_1 := v_0.Args[1]
+		if v_0_1.Op != OpARMMOVWconst {
+			break
+		}
+		d := v_0_1.AuxInt
+		v.reset(OpARMMOVWconst)
+		v.AuxInt = int64(uint32(c) % uint32(d))
+		return true
+	}
+	return false
+}
 func rewriteValueARM_OpSignExt16to32(v *Value, config *Config) bool {
 	b := v.Block
 	_ = b

src/runtime/vlop_arm.s

@@ -107,6 +107,7 @@ TEXT runtime·_sfloatpanic(SB),NOSPLIT,$-4
 	B	runtime·sigpanic(SB)
 
 // func udiv(n, d uint32) (q, r uint32)
+// compiler knowns the register usage of this function
 // Reference: 
 // Sloss, Andrew et. al; ARM System Developer's Guide: Designing and Optimizing System Software
 // Morgan Kaufmann; 1 edition (April 8, 2004), ISBN 978-1558608740
@@ -117,7 +118,7 @@ TEXT runtime·_sfloatpanic(SB),NOSPLIT,$-4
 #define Ra	R11
 
 // Be careful: Ra == R11 will be used by the linker for synthesized instructions.
-TEXT udiv<>(SB),NOSPLIT,$-4
+TEXT udiv(SB),NOSPLIT,$-4
 	CLZ 	Rq, Rs // find normalizing shift
 	MOVW.S	Rq<<Rs, Ra
 	MOVW	$fast_udiv_tab<>-64(SB), RM
@@ -227,7 +228,7 @@ TEXT _divu(SB), NOSPLIT, $16-0
 	MOVW	RTMP, Rr		/* numerator */
 	MOVW	g_m(g), Rq
 	MOVW	m_divmod(Rq), Rq	/* denominator */
-	BL  	udiv<>(SB)
+	BL  	udiv(SB)
 	MOVW	Rq, RTMP
 	MOVW	4(R13), Rq
 	MOVW	8(R13), Rr
@@ -245,7 +246,7 @@ TEXT _modu(SB), NOSPLIT, $16-0
 	MOVW	RTMP, Rr		/* numerator */
 	MOVW	g_m(g), Rq
 	MOVW	m_divmod(Rq), Rq	/* denominator */
-	BL  	udiv<>(SB)
+	BL  	udiv(SB)
 	MOVW	Rr, RTMP
 	MOVW	4(R13), Rq
 	MOVW	8(R13), Rr
@@ -269,7 +270,7 @@ TEXT _div(SB),NOSPLIT,$16-0
 	BGE 	d2
 	RSB 	$0, Rq, Rq
 d0:
-	BL  	udiv<>(SB)  		/* none/both neg */
+	BL  	udiv(SB)  		/* none/both neg */
 	MOVW	Rq, RTMP
 	B		out1
 d1:
@@ -277,8 +278,8 @@ d1:
 	BGE 	d0
 	RSB 	$0, Rq, Rq
 d2:
-	BL  	udiv<>(SB)  		/* one neg */
-	RSB		$0, Rq, RTMP
+	BL  	udiv(SB)  		/* one neg */
+	RSB	$0, Rq, RTMP
 out1:
 	MOVW	4(R13), Rq
 	MOVW	8(R13), Rr
@@ -300,11 +301,11 @@ TEXT _mod(SB),NOSPLIT,$16-0
 	CMP 	$0, Rr
 	BGE 	m1
 	RSB 	$0, Rr, Rr
-	BL  	udiv<>(SB)  		/* neg numerator */
+	BL  	udiv(SB)  		/* neg numerator */
 	RSB 	$0, Rr, RTMP
 	B   	out
 m1:
-	BL  	udiv<>(SB)  		/* pos numerator */
+	BL  	udiv(SB)  		/* pos numerator */
 	MOVW	Rr, RTMP
 out:
 	MOVW	4(R13), Rq