cmd/compile: improve LoweredMove performance on ppc64x

This change improves the performance for LoweredMove on ppc64le
and ppc64.

benchmark                   old ns/op     new ns/op     delta
BenchmarkCopyFat8-16        0.93          0.69          -25.81%
BenchmarkCopyFat12-16       2.61          1.85          -29.12%
BenchmarkCopyFat16-16       9.68          1.89          -80.48%
BenchmarkCopyFat24-16       4.48          1.85          -58.71%
BenchmarkCopyFat32-16       6.12          1.82          -70.26%
BenchmarkCopyFat64-16       21.2          2.70          -87.26%
BenchmarkCopyFat128-16      29.6          3.97          -86.59%
BenchmarkCopyFat256-16      52.6          13.4          -74.52%
BenchmarkCopyFat512-16      97.1          18.7          -80.74%
BenchmarkCopyFat1024-16     186           35.3          -81.02%

BenchmarkAssertE2TLarge-16      14.2          5.06          -64.37%

Fixes #19785

Change-Id: I7d5e0052712b75811c02c7d86c5112e5649ad782
Reviewed-on: https://go-review.googlesource.com/38950
Reviewed-by: Keith Randall <khr@golang.org>

src/cmd/compile/internal/ppc64/ssa.go

@@ -917,75 +917,171 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		}
 
 	case ssa.OpPPC64LoweredMove:
-		// Similar to how this is done on ARM,
-		// except that PPC MOVDU x,off(y) is *(y+off) = x; y=y+off,
-		// not store-and-increment.
-		// Inputs must be valid pointers to memory,
-		// so adjust arg0 and arg1 as part of the expansion.
-		// arg2 should be src+size-align,
-		//
-		// ADD    -8,R3,R3
-		// ADD    -8,R4,R4
-		// MOVDU	8(R4), Rtmp
-		// MOVDU 	Rtmp, 8(R3)
-		// CMP	R4, Rarg2
-		// BL	-3(PC)
-		// arg2 is the address of the last element of src
-		// auxint is alignment
-		var sz int64
-		var movu obj.As
-		switch {
-		case v.AuxInt%8 == 0:
-			sz = 8
-			movu = ppc64.AMOVDU
-		case v.AuxInt%4 == 0:
-			sz = 4
-			movu = ppc64.AMOVWZU // MOVWU instruction not implemented
-		case v.AuxInt%2 == 0:
-			sz = 2
-			movu = ppc64.AMOVHU
-		default:
-			sz = 1
-			movu = ppc64.AMOVBU
-		}
 
-		p := s.Prog(ppc64.AADD)
+		// This will be used when moving more
-		p.Reg = v.Args[0].Reg()
+		// than 8 bytes.  Moves start with as
+		// as many 8 byte moves as possible, then
+		// 4, 2, or 1 byte(s) as remaining.  This will
+		// work and be efficient for power8 or later.
+		// If there are 64 or more bytes, then a
+		// loop is generated to move 32 bytes and
+		// update the src and dst addresses on each
+		// iteration. When < 64 bytes, the appropriate
+		// number of moves are generated based on the
+		// size.
+		// When moving >= 64 bytes a loop is used
+		//	MOVD len/32,REG_TMP
+		//	MOVD REG_TMP,CTR
+		// top:
+		//	MOVD (R4),R7
+		//	MOVD 8(R4),R8
+		//	MOVD 16(R4),R9
+		//	MOVD 24(R4),R10
+		//	ADD  R4,$32
+		//	MOVD R7,(R3)
+		//	MOVD R8,8(R3)
+		//	MOVD R9,16(R3)
+		//	MOVD R10,24(R3)
+		//	ADD  R3,$32
+		//	BC 16,0,top
+		// Bytes not moved by this loop are moved
+		// with a combination of the following instructions,
+		// starting with the largest sizes and generating as
+		// many as needed, using the appropriate offset value.
+		//	MOVD  n(R4),R7
+		//	MOVD  R7,n(R3)
+		//	MOVW  n1(R4),R7
+		//	MOVW  R7,n1(R3)
+		//	MOVH  n2(R4),R7
+		//	MOVH  R7,n2(R3)
+		//	MOVB  n3(R4),R7
+		//	MOVB  R7,n3(R3)
+
+		// Each loop iteration moves 32 bytes
+		ctr := v.AuxInt / 32
+
+		// Remainder after the loop
+		rem := v.AuxInt % 32
+
+		dst_reg := v.Args[0].Reg()
+		src_reg := v.Args[1].Reg()
+
+		// The set of registers used here, must match the clobbered reg list
+		// in PPC64Ops.go.
+		useregs := []int16{ppc64.REG_R7, ppc64.REG_R8, ppc64.REG_R9, ppc64.REG_R10}
+		offset := int64(0)
+
+		// top of the loop
+		var top *obj.Prog
+		// Only generate looping code when loop counter is > 1 for >= 64 bytes
+		if ctr > 1 {
+			// Set up the CTR
+			p := s.Prog(ppc64.AMOVD)
 			p.From.Type = obj.TYPE_CONST
-		p.From.Offset = -sz
+			p.From.Offset = ctr
-		p.To.Type = obj.TYPE_REG
-		p.To.Reg = v.Args[0].Reg()
-
-		p = s.Prog(ppc64.AADD)
-		p.Reg = v.Args[1].Reg()
-		p.From.Type = obj.TYPE_CONST
-		p.From.Offset = -sz
-		p.To.Type = obj.TYPE_REG
-		p.To.Reg = v.Args[1].Reg()
-
-		p = s.Prog(movu)
-		p.From.Type = obj.TYPE_MEM
-		p.From.Reg = v.Args[1].Reg()
-		p.From.Offset = sz
 			p.To.Type = obj.TYPE_REG
 			p.To.Reg = ppc64.REGTMP
 
-		p2 := s.Prog(movu)
+			p = s.Prog(ppc64.AMOVD)
-		p2.From.Type = obj.TYPE_REG
+			p.From.Type = obj.TYPE_REG
-		p2.From.Reg = ppc64.REGTMP
+			p.From.Reg = ppc64.REGTMP
-		p2.To.Type = obj.TYPE_MEM
+			p.To.Type = obj.TYPE_REG
-		p2.To.Reg = v.Args[0].Reg()
+			p.To.Reg = ppc64.REG_CTR
-		p2.To.Offset = sz
 
-		p3 := s.Prog(ppc64.ACMPU)
+			// Generate all the MOVDs for loads
-		p3.From.Reg = v.Args[1].Reg()
+			// based off the same register, increasing
-		p3.From.Type = obj.TYPE_REG
+			// the offset by 8 for each instruction
-		p3.To.Reg = v.Args[2].Reg()
+			for _, rg := range useregs {
-		p3.To.Type = obj.TYPE_REG
+				p := s.Prog(ppc64.AMOVD)
+				p.From.Type = obj.TYPE_MEM
+				p.From.Reg = src_reg
+				p.From.Offset = offset
+				p.To.Type = obj.TYPE_REG
+				p.To.Reg = rg
+				if top == nil {
+					top = p
+				}
+				offset += 8
+			}
+			// increment the src_reg for next iteration
+			p = s.Prog(ppc64.AADD)
+			p.Reg = src_reg
+			p.From.Type = obj.TYPE_CONST
+			p.From.Offset = 32
+			p.To.Type = obj.TYPE_REG
+			p.To.Reg = src_reg
 
-		p4 := s.Prog(ppc64.ABLT)
+			// generate the MOVDs for stores, based
-		p4.To.Type = obj.TYPE_BRANCH
+			// off the same register, using the same
-		gc.Patch(p4, p)
+			// offsets as in the loads.
+			offset = int64(0)
+			for _, rg := range useregs {
+				p := s.Prog(ppc64.AMOVD)
+				p.From.Type = obj.TYPE_REG
+				p.From.Reg = rg
+				p.To.Type = obj.TYPE_MEM
+				p.To.Reg = dst_reg
+				p.To.Offset = offset
+				offset += 8
+			}
+			// increment the dst_reg for next iteration
+			p = s.Prog(ppc64.AADD)
+			p.Reg = dst_reg
+			p.From.Type = obj.TYPE_CONST
+			p.From.Offset = 32
+			p.To.Type = obj.TYPE_REG
+			p.To.Reg = dst_reg
+
+			// BC with BO_BCTR generates bdnz to branch on nonzero CTR
+			// to loop top.
+			p = s.Prog(ppc64.ABC)
+			p.From.Type = obj.TYPE_CONST
+			p.From.Offset = ppc64.BO_BCTR
+			p.Reg = ppc64.REG_R0
+			p.To.Type = obj.TYPE_BRANCH
+			gc.Patch(p, top)
+
+			// src_reg and dst_reg were incremented in the loop, so
+			// later instructions start with offset 0.
+			offset = int64(0)
+		}
+
+		// No loop was generated for one iteration, so
+		// add 32 bytes to the remainder to move those bytes.
+		if ctr == 1 {
+			rem += 32
+		}
+
+		// Generate all the remaining load and store pairs, starting with
+		// as many 8 byte moves as possible, then 4, 2, 1.
+		for rem > 0 {
+			op, size := ppc64.AMOVB, int64(1)
+			switch {
+			case rem >= 8:
+				op, size = ppc64.AMOVD, 8
+			case rem >= 4:
+				op, size = ppc64.AMOVW, 4
+			case rem >= 2:
+				op, size = ppc64.AMOVH, 2
+			}
+			// Load
+			p := s.Prog(op)
+			p.To.Type = obj.TYPE_REG
+			p.To.Reg = ppc64.REG_R7
+			p.From.Type = obj.TYPE_MEM
+			p.From.Reg = src_reg
+			p.From.Offset = offset
+
+			// Store
+			p = s.Prog(op)
+			p.From.Type = obj.TYPE_REG
+			p.From.Reg = ppc64.REG_R7
+			p.To.Type = obj.TYPE_MEM
+			p.To.Reg = dst_reg
+			p.To.Offset = offset
+			rem -= size
+			offset += size
+		}
 
 	case ssa.OpPPC64CALLstatic:
 		s.Call(v)

src/cmd/compile/internal/ssa/gen/PPC64.rules

@@ -556,46 +556,29 @@
 // moves
 (Move [0] _ _ mem) -> mem
 (Move [1] dst src mem) -> (MOVBstore dst (MOVBZload src mem) mem)
-(Move [2] {t} dst src mem) && t.(Type).Alignment()%2 == 0 ->
-	(MOVHstore dst (MOVHZload src mem) mem)
 (Move [2] dst src mem) ->
-	(MOVBstore [1] dst (MOVBZload [1] src mem)
+        (MOVHstore dst (MOVHZload src mem) mem)
-		(MOVBstore dst (MOVBZload src mem) mem))
-(Move [4] {t} dst src mem) && t.(Type).Alignment()%4 == 0 ->
-	(MOVWstore dst (MOVWload src mem) mem)
-(Move [4] {t} dst src mem) && t.(Type).Alignment()%2 == 0 ->
-	(MOVHstore [2] dst (MOVHZload [2] src mem)
-		(MOVHstore dst (MOVHZload src mem) mem))
-(Move [4] dst src mem) ->
-	(MOVBstore [3] dst (MOVBZload [3] src mem)
-		(MOVBstore [2] dst (MOVBZload [2] src mem)
-			(MOVBstore [1] dst (MOVBZload [1] src mem)
-				(MOVBstore dst (MOVBZload src mem) mem))))
-
-(Move [8] {t} dst src mem) && t.(Type).Alignment()%8 == 0 ->
-	(MOVDstore dst (MOVDload src mem) mem)
-(Move [8] {t} dst src mem) && t.(Type).Alignment()%4 == 0 ->
-	(MOVWstore [4] dst (MOVWZload [4] src mem)
-		(MOVWstore dst (MOVWZload src mem) mem))
-(Move [8] {t} dst src mem) && t.(Type).Alignment()%2 == 0->
-	(MOVHstore [6] dst (MOVHZload [6] src mem)
-		(MOVHstore [4] dst (MOVHZload [4] src mem)
-			(MOVHstore [2] dst (MOVHZload [2] src mem)
-				(MOVHstore dst (MOVHZload src mem) mem))))
-
 (Move [3] dst src mem) ->
         (MOVBstore [2] dst (MOVBZload [2] src mem)
-		(MOVBstore [1] dst (MOVBZload [1] src mem)
+                (MOVHstore dst (MOVHload src mem) mem))
-			(MOVBstore dst (MOVBZload src mem) mem)))
+(Move [4] dst src mem) ->
+        (MOVWstore dst (MOVWload src mem) mem)
+(Move [5] dst src mem) ->
+        (MOVBstore [4] dst (MOVBZload [4] src mem)
+                (MOVWstore dst (MOVWload src mem) mem))
+(Move [6] dst src mem) ->
+        (MOVHstore [4] dst (MOVHZload [4] src mem)
+                (MOVWstore dst (MOVWload src mem) mem))
+(Move [7] dst src mem) ->
+        (MOVBstore [6] dst (MOVBZload [6] src mem)
+                (MOVHstore [4] dst (MOVHZload [4] src mem)
+                        (MOVWstore dst (MOVWload src mem) mem)))
+(Move [8] dst src mem) ->
+        (MOVDstore dst (MOVDload src mem) mem)
 
 // Large move uses a loop
-(Move [s] {t} dst src mem)
+(Move [s] dst src mem) && s > 8 ->
-	&& (s > 512 || config.noDuffDevice) || t.(Type).Alignment()%8 != 0 ->
+        (LoweredMove [s] dst src mem)
-	(LoweredMove [t.(Type).Alignment()]
-		dst
-		src
-		(ADDconst <src.Type> src [s-moveSize(t.(Type).Alignment(), config)])
-		mem)
 
 // Calls
 // Lowering calls

src/cmd/compile/internal/ssa/gen/PPC64Ops.go

@@ -349,26 +349,41 @@ func init() {
 			typ:            "Mem",
 			faultOnNilArg0: true,
 		},
+		// Loop code:
+		//	MOVD len/32,REG_TMP  only for loop
+		//	MOVD REG_TMP,CTR     only for loop
+		// loop:
+		//	MOVD (R4),R7
+		//	MOVD 8(R4),R8
+		//	MOVD 16(R4),R9
+		//	MOVD 24(R4),R10
+		//	ADD  R4,$32          only with loop
+		//	MOVD R7,(R3)
+		//	MOVD R8,8(R3)
+		//	MOVD R9,16(R3)
+		//	MOVD R10,24(R3)
+		//	ADD  R3,$32          only with loop
+		//	BC 16,0,loop         only with loop
+		// Bytes not moved by this loop are moved
+		// with a combination of the following instructions,
+		// starting with the largest sizes and generating as
+		// many as needed, using the appropriate offset value.
+		//	MOVD  n(R4),R7
+		//	MOVD  R7,n(R3)
+		//	MOVW  n1(R4),R7
+		//	MOVW  R7,n1(R3)
+		//	MOVH  n2(R4),R7
+		//	MOVH  R7,n2(R3)
+		//	MOVB  n3(R4),R7
+		//	MOVB  R7,n3(R3)
 
-		// large or unaligned move
-		// arg0 = address of dst memory (in R3, changed as side effect)
-		// arg1 = address of src memory (in R4, changed as side effect)
-		// arg2 = address of the last element of src
-		// arg3 = mem
-		// returns mem
-		//  ADD -8,R3,R3 // intermediate value not valid GC ptr, cannot expose to opt+GC
-		//  ADD -8,R4,R4 // intermediate value not valid GC ptr, cannot expose to opt+GC
-		//	MOVDU	8(R4), Rtmp
-		//	MOVDU	Rtmp, 8(R3)
-		//	CMP	R4, Rarg2
-		//	BLT	-3(PC)
 		{
 			name:      "LoweredMove",
 			aux:       "Int64",
-			argLength: 4,
+			argLength: 3,
 			reg: regInfo{
-				inputs:   []regMask{buildReg("R3"), buildReg("R4"), gp},
+				inputs:   []regMask{buildReg("R3"), buildReg("R4")},
-				clobbers: buildReg("R3 R4"),
+				clobbers: buildReg("R3 R4 R7 R8 R9 R10"),
 			},
 			clobberFlags:   true,
 			typ:            "Mem",

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

@@ -17409,7 +17409,7 @@ var opcodeTable = [...]opInfo{
 	{
 		name:           "LoweredMove",
 		auxType:        auxInt64,
-		argLen:         4,
+		argLen:         3,
 		clobberFlags:   true,
 		faultOnNilArg0: true,
 		faultOnNilArg1: true,
@@ -17417,9 +17417,8 @@ var opcodeTable = [...]opInfo{
 			inputs: []inputInfo{
 				{0, 8},  // R3
 				{1, 16}, // R4
-				{2, 1073733624}, // R3 R4 R5 R6 R7 R8 R9 R10 R11 R12 R14 R15 R16 R17 R18 R19 R20 R21 R22 R23 R24 R25 R26 R27 R28 R29
 			},
-			clobbers: 24, // R3 R4
+			clobbers: 1944, // R3 R4 R7 R8 R9 R10
 		},
 	},
 	{

src/cmd/compile/internal/ssa/rewritePPC64.go

@@ -3686,8 +3686,6 @@ func rewriteValuePPC64_OpMod8u(v *Value) bool {
 func rewriteValuePPC64_OpMove(v *Value) bool {
 	b := v.Block
 	_ = b
-	config := b.Func.Config
-	_ = config
 	types := &b.Func.Config.Types
 	_ = types
 	// match: (Move [0] _ _ mem)
@@ -3722,20 +3720,16 @@ func rewriteValuePPC64_OpMove(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (Move [2] {t} dst src mem)
+	// match: (Move [2] dst src mem)
-	// cond: t.(Type).Alignment()%2 == 0
+	// cond:
 	// result: (MOVHstore dst (MOVHZload src mem) mem)
 	for {
 		if v.AuxInt != 2 {
 			break
 		}
-		t := v.Aux
 		dst := v.Args[0]
 		src := v.Args[1]
 		mem := v.Args[2]
-		if !(t.(Type).Alignment()%2 == 0) {
-			break
-		}
 		v.reset(OpPPC64MOVHstore)
 		v.AddArg(dst)
 		v0 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
@@ -3745,243 +3739,9 @@ func rewriteValuePPC64_OpMove(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (Move [2] dst src mem)
-	// cond:
-	// result: (MOVBstore [1] dst (MOVBZload [1] src mem) 		(MOVBstore dst (MOVBZload src mem) mem))
-	for {
-		if v.AuxInt != 2 {
-			break
-		}
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		v.reset(OpPPC64MOVBstore)
-		v.AuxInt = 1
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
-		v0.AuxInt = 1
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVBstore, TypeMem)
-		v1.AddArg(dst)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
-		v2.AddArg(src)
-		v2.AddArg(mem)
-		v1.AddArg(v2)
-		v1.AddArg(mem)
-		v.AddArg(v1)
-		return true
-	}
-	// match: (Move [4] {t} dst src mem)
-	// cond: t.(Type).Alignment()%4 == 0
-	// result: (MOVWstore dst (MOVWload src mem) mem)
-	for {
-		if v.AuxInt != 4 {
-			break
-		}
-		t := v.Aux
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		if !(t.(Type).Alignment()%4 == 0) {
-			break
-		}
-		v.reset(OpPPC64MOVWstore)
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVWload, types.Int32)
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v.AddArg(mem)
-		return true
-	}
-	// match: (Move [4] {t} dst src mem)
-	// cond: t.(Type).Alignment()%2 == 0
-	// result: (MOVHstore [2] dst (MOVHZload [2] src mem) 		(MOVHstore dst (MOVHZload src mem) mem))
-	for {
-		if v.AuxInt != 4 {
-			break
-		}
-		t := v.Aux
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		if !(t.(Type).Alignment()%2 == 0) {
-			break
-		}
-		v.reset(OpPPC64MOVHstore)
-		v.AuxInt = 2
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
-		v0.AuxInt = 2
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVHstore, TypeMem)
-		v1.AddArg(dst)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
-		v2.AddArg(src)
-		v2.AddArg(mem)
-		v1.AddArg(v2)
-		v1.AddArg(mem)
-		v.AddArg(v1)
-		return true
-	}
-	// match: (Move [4] dst src mem)
-	// cond:
-	// result: (MOVBstore [3] dst (MOVBZload [3] src mem) 		(MOVBstore [2] dst (MOVBZload [2] src mem) 			(MOVBstore [1] dst (MOVBZload [1] src mem) 				(MOVBstore dst (MOVBZload src mem) mem))))
-	for {
-		if v.AuxInt != 4 {
-			break
-		}
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		v.reset(OpPPC64MOVBstore)
-		v.AuxInt = 3
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
-		v0.AuxInt = 3
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVBstore, TypeMem)
-		v1.AuxInt = 2
-		v1.AddArg(dst)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
-		v2.AuxInt = 2
-		v2.AddArg(src)
-		v2.AddArg(mem)
-		v1.AddArg(v2)
-		v3 := b.NewValue0(v.Pos, OpPPC64MOVBstore, TypeMem)
-		v3.AuxInt = 1
-		v3.AddArg(dst)
-		v4 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
-		v4.AuxInt = 1
-		v4.AddArg(src)
-		v4.AddArg(mem)
-		v3.AddArg(v4)
-		v5 := b.NewValue0(v.Pos, OpPPC64MOVBstore, TypeMem)
-		v5.AddArg(dst)
-		v6 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
-		v6.AddArg(src)
-		v6.AddArg(mem)
-		v5.AddArg(v6)
-		v5.AddArg(mem)
-		v3.AddArg(v5)
-		v1.AddArg(v3)
-		v.AddArg(v1)
-		return true
-	}
-	// match: (Move [8] {t} dst src mem)
-	// cond: t.(Type).Alignment()%8 == 0
-	// result: (MOVDstore dst (MOVDload src mem) mem)
-	for {
-		if v.AuxInt != 8 {
-			break
-		}
-		t := v.Aux
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		if !(t.(Type).Alignment()%8 == 0) {
-			break
-		}
-		v.reset(OpPPC64MOVDstore)
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVDload, types.Int64)
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v.AddArg(mem)
-		return true
-	}
-	// match: (Move [8] {t} dst src mem)
-	// cond: t.(Type).Alignment()%4 == 0
-	// result: (MOVWstore [4] dst (MOVWZload [4] src mem) 		(MOVWstore dst (MOVWZload src mem) mem))
-	for {
-		if v.AuxInt != 8 {
-			break
-		}
-		t := v.Aux
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		if !(t.(Type).Alignment()%4 == 0) {
-			break
-		}
-		v.reset(OpPPC64MOVWstore)
-		v.AuxInt = 4
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVWZload, types.UInt32)
-		v0.AuxInt = 4
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVWstore, TypeMem)
-		v1.AddArg(dst)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVWZload, types.UInt32)
-		v2.AddArg(src)
-		v2.AddArg(mem)
-		v1.AddArg(v2)
-		v1.AddArg(mem)
-		v.AddArg(v1)
-		return true
-	}
-	// match: (Move [8] {t} dst src mem)
-	// cond: t.(Type).Alignment()%2 == 0
-	// result: (MOVHstore [6] dst (MOVHZload [6] src mem) 		(MOVHstore [4] dst (MOVHZload [4] src mem) 			(MOVHstore [2] dst (MOVHZload [2] src mem) 				(MOVHstore dst (MOVHZload src mem) mem))))
-	for {
-		if v.AuxInt != 8 {
-			break
-		}
-		t := v.Aux
-		dst := v.Args[0]
-		src := v.Args[1]
-		mem := v.Args[2]
-		if !(t.(Type).Alignment()%2 == 0) {
-			break
-		}
-		v.reset(OpPPC64MOVHstore)
-		v.AuxInt = 6
-		v.AddArg(dst)
-		v0 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
-		v0.AuxInt = 6
-		v0.AddArg(src)
-		v0.AddArg(mem)
-		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVHstore, TypeMem)
-		v1.AuxInt = 4
-		v1.AddArg(dst)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
-		v2.AuxInt = 4
-		v2.AddArg(src)
-		v2.AddArg(mem)
-		v1.AddArg(v2)
-		v3 := b.NewValue0(v.Pos, OpPPC64MOVHstore, TypeMem)
-		v3.AuxInt = 2
-		v3.AddArg(dst)
-		v4 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
-		v4.AuxInt = 2
-		v4.AddArg(src)
-		v4.AddArg(mem)
-		v3.AddArg(v4)
-		v5 := b.NewValue0(v.Pos, OpPPC64MOVHstore, TypeMem)
-		v5.AddArg(dst)
-		v6 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
-		v6.AddArg(src)
-		v6.AddArg(mem)
-		v5.AddArg(v6)
-		v5.AddArg(mem)
-		v3.AddArg(v5)
-		v1.AddArg(v3)
-		v.AddArg(v1)
-		return true
-	}
 	// match: (Move [3] dst src mem)
 	// cond:
-	// result: (MOVBstore [2] dst (MOVBZload [2] src mem) 		(MOVBstore [1] dst (MOVBZload [1] src mem) 			(MOVBstore dst (MOVBZload src mem) mem)))
+	// result: (MOVBstore [2] dst (MOVBZload [2] src mem)                 (MOVHstore dst (MOVHload src mem) mem))
 	for {
 		if v.AuxInt != 3 {
 			break
@@ -3997,17 +3757,120 @@ func rewriteValuePPC64_OpMove(v *Value) bool {
 		v0.AddArg(src)
 		v0.AddArg(mem)
 		v.AddArg(v0)
-		v1 := b.NewValue0(v.Pos, OpPPC64MOVBstore, TypeMem)
+		v1 := b.NewValue0(v.Pos, OpPPC64MOVHstore, TypeMem)
-		v1.AuxInt = 1
 		v1.AddArg(dst)
-		v2 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
+		v2 := b.NewValue0(v.Pos, OpPPC64MOVHload, types.Int16)
-		v2.AuxInt = 1
 		v2.AddArg(src)
 		v2.AddArg(mem)
 		v1.AddArg(v2)
-		v3 := b.NewValue0(v.Pos, OpPPC64MOVBstore, TypeMem)
+		v1.AddArg(mem)
+		v.AddArg(v1)
+		return true
+	}
+	// match: (Move [4] dst src mem)
+	// cond:
+	// result: (MOVWstore dst (MOVWload src mem) mem)
+	for {
+		if v.AuxInt != 4 {
+			break
+		}
+		dst := v.Args[0]
+		src := v.Args[1]
+		mem := v.Args[2]
+		v.reset(OpPPC64MOVWstore)
+		v.AddArg(dst)
+		v0 := b.NewValue0(v.Pos, OpPPC64MOVWload, types.Int32)
+		v0.AddArg(src)
+		v0.AddArg(mem)
+		v.AddArg(v0)
+		v.AddArg(mem)
+		return true
+	}
+	// match: (Move [5] dst src mem)
+	// cond:
+	// result: (MOVBstore [4] dst (MOVBZload [4] src mem)                 (MOVWstore dst (MOVWload src mem) mem))
+	for {
+		if v.AuxInt != 5 {
+			break
+		}
+		dst := v.Args[0]
+		src := v.Args[1]
+		mem := v.Args[2]
+		v.reset(OpPPC64MOVBstore)
+		v.AuxInt = 4
+		v.AddArg(dst)
+		v0 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
+		v0.AuxInt = 4
+		v0.AddArg(src)
+		v0.AddArg(mem)
+		v.AddArg(v0)
+		v1 := b.NewValue0(v.Pos, OpPPC64MOVWstore, TypeMem)
+		v1.AddArg(dst)
+		v2 := b.NewValue0(v.Pos, OpPPC64MOVWload, types.Int32)
+		v2.AddArg(src)
+		v2.AddArg(mem)
+		v1.AddArg(v2)
+		v1.AddArg(mem)
+		v.AddArg(v1)
+		return true
+	}
+	// match: (Move [6] dst src mem)
+	// cond:
+	// result: (MOVHstore [4] dst (MOVHZload [4] src mem)                 (MOVWstore dst (MOVWload src mem) mem))
+	for {
+		if v.AuxInt != 6 {
+			break
+		}
+		dst := v.Args[0]
+		src := v.Args[1]
+		mem := v.Args[2]
+		v.reset(OpPPC64MOVHstore)
+		v.AuxInt = 4
+		v.AddArg(dst)
+		v0 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
+		v0.AuxInt = 4
+		v0.AddArg(src)
+		v0.AddArg(mem)
+		v.AddArg(v0)
+		v1 := b.NewValue0(v.Pos, OpPPC64MOVWstore, TypeMem)
+		v1.AddArg(dst)
+		v2 := b.NewValue0(v.Pos, OpPPC64MOVWload, types.Int32)
+		v2.AddArg(src)
+		v2.AddArg(mem)
+		v1.AddArg(v2)
+		v1.AddArg(mem)
+		v.AddArg(v1)
+		return true
+	}
+	// match: (Move [7] dst src mem)
+	// cond:
+	// result: (MOVBstore [6] dst (MOVBZload [6] src mem)                 (MOVHstore [4] dst (MOVHZload [4] src mem)                         (MOVWstore dst (MOVWload src mem) mem)))
+	for {
+		if v.AuxInt != 7 {
+			break
+		}
+		dst := v.Args[0]
+		src := v.Args[1]
+		mem := v.Args[2]
+		v.reset(OpPPC64MOVBstore)
+		v.AuxInt = 6
+		v.AddArg(dst)
+		v0 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
+		v0.AuxInt = 6
+		v0.AddArg(src)
+		v0.AddArg(mem)
+		v.AddArg(v0)
+		v1 := b.NewValue0(v.Pos, OpPPC64MOVHstore, TypeMem)
+		v1.AuxInt = 4
+		v1.AddArg(dst)
+		v2 := b.NewValue0(v.Pos, OpPPC64MOVHZload, types.UInt16)
+		v2.AuxInt = 4
+		v2.AddArg(src)
+		v2.AddArg(mem)
+		v1.AddArg(v2)
+		v3 := b.NewValue0(v.Pos, OpPPC64MOVWstore, TypeMem)
 		v3.AddArg(dst)
-		v4 := b.NewValue0(v.Pos, OpPPC64MOVBZload, types.UInt8)
+		v4 := b.NewValue0(v.Pos, OpPPC64MOVWload, types.Int32)
 		v4.AddArg(src)
 		v4.AddArg(mem)
 		v3.AddArg(v4)
@@ -4016,26 +3879,40 @@ func rewriteValuePPC64_OpMove(v *Value) bool {
 		v.AddArg(v1)
 		return true
 	}
-	// match: (Move [s] {t} dst src mem)
+	// match: (Move [8] dst src mem)
-	// cond: (s > 512 || config.noDuffDevice) || t.(Type).Alignment()%8 != 0
+	// cond:
-	// result: (LoweredMove [t.(Type).Alignment()] 		dst 		src 		(ADDconst <src.Type> src [s-moveSize(t.(Type).Alignment(), config)]) 		mem)
+	// result: (MOVDstore dst (MOVDload src mem) mem)
 	for {
-		s := v.AuxInt
+		if v.AuxInt != 8 {
-		t := v.Aux
+			break
+		}
 		dst := v.Args[0]
 		src := v.Args[1]
 		mem := v.Args[2]
-		if !((s > 512 || config.noDuffDevice) || t.(Type).Alignment()%8 != 0) {
+		v.reset(OpPPC64MOVDstore)
+		v.AddArg(dst)
+		v0 := b.NewValue0(v.Pos, OpPPC64MOVDload, types.Int64)
+		v0.AddArg(src)
+		v0.AddArg(mem)
+		v.AddArg(v0)
+		v.AddArg(mem)
+		return true
+	}
+	// match: (Move [s] dst src mem)
+	// cond: s > 8
+	// result: (LoweredMove [s] dst src mem)
+	for {
+		s := v.AuxInt
+		dst := v.Args[0]
+		src := v.Args[1]
+		mem := v.Args[2]
+		if !(s > 8) {
 			break
 		}
 		v.reset(OpPPC64LoweredMove)
-		v.AuxInt = t.(Type).Alignment()
+		v.AuxInt = s
 		v.AddArg(dst)
 		v.AddArg(src)
-		v0 := b.NewValue0(v.Pos, OpPPC64ADDconst, src.Type)
-		v0.AuxInt = s - moveSize(t.(Type).Alignment(), config)
-		v0.AddArg(src)
-		v.AddArg(v0)
 		v.AddArg(mem)
 		return true
 	}