// Copyright 2016 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 amd64

import (
	"fmt"
	"math"

	"cmd/compile/internal/gc"
	"cmd/compile/internal/ssa"
	"cmd/internal/obj"
	"cmd/internal/obj/x86"
)

// Smallest possible faulting page at address zero.
const minZeroPage = 4096

// ssaRegToReg maps ssa register numbers to obj register numbers.
var ssaRegToReg = []int16{
	x86.REG_AX,
	x86.REG_CX,
	x86.REG_DX,
	x86.REG_BX,
	x86.REG_SP,
	x86.REG_BP,
	x86.REG_SI,
	x86.REG_DI,
	x86.REG_R8,
	x86.REG_R9,
	x86.REG_R10,
	x86.REG_R11,
	x86.REG_R12,
	x86.REG_R13,
	x86.REG_R14,
	x86.REG_R15,
	x86.REG_X0,
	x86.REG_X1,
	x86.REG_X2,
	x86.REG_X3,
	x86.REG_X4,
	x86.REG_X5,
	x86.REG_X6,
	x86.REG_X7,
	x86.REG_X8,
	x86.REG_X9,
	x86.REG_X10,
	x86.REG_X11,
	x86.REG_X12,
	x86.REG_X13,
	x86.REG_X14,
	x86.REG_X15,
	0, // SB isn't a real register.  We fill an Addr.Reg field with 0 in this case.
}

// markMoves marks any MOVXconst ops that need to avoid clobbering flags.
func ssaMarkMoves(s *gc.SSAGenState, b *ssa.Block) {
	flive := b.FlagsLiveAtEnd
	if b.Control != nil && b.Control.Type.IsFlags() {
		flive = true
	}
	for i := len(b.Values) - 1; i >= 0; i-- {
		v := b.Values[i]
		if flive && (v.Op == ssa.OpAMD64MOVLconst || v.Op == ssa.OpAMD64MOVQconst) {
			// The "mark" is any non-nil Aux value.
			v.Aux = v
		}
		if v.Type.IsFlags() {
			flive = false
		}
		for _, a := range v.Args {
			if a.Type.IsFlags() {
				flive = true
			}
		}
	}
}

// loadByType returns the load instruction of the given type.
func loadByType(t ssa.Type) obj.As {
	// Avoid partial register write
	if !t.IsFloat() && t.Size() <= 2 {
		if t.Size() == 1 {
			return x86.AMOVBLZX
		} else {
			return x86.AMOVWLZX
		}
	}
	// Otherwise, there's no difference between load and store opcodes.
	return storeByType(t)
}

// storeByType returns the store instruction of the given type.
func storeByType(t ssa.Type) obj.As {
	width := t.Size()
	if t.IsFloat() {
		switch width {
		case 4:
			return x86.AMOVSS
		case 8:
			return x86.AMOVSD
		}
	} else {
		switch width {
		case 1:
			return x86.AMOVB
		case 2:
			return x86.AMOVW
		case 4:
			return x86.AMOVL
		case 8:
			return x86.AMOVQ
		}
	}
	panic("bad store type")
}

// moveByType returns the reg->reg move instruction of the given type.
func moveByType(t ssa.Type) obj.As {
	if t.IsFloat() {
		// Moving the whole sse2 register is faster
		// than moving just the correct low portion of it.
		// There is no xmm->xmm move with 1 byte opcode,
		// so use movups, which has 2 byte opcode.
		return x86.AMOVUPS
	} else {
		switch t.Size() {
		case 1:
			// Avoids partial register write
			return x86.AMOVL
		case 2:
			return x86.AMOVL
		case 4:
			return x86.AMOVL
		case 8:
			return x86.AMOVQ
		case 16:
			return x86.AMOVUPS // int128s are in SSE registers
		default:
			panic(fmt.Sprintf("bad int register width %d:%s", t.Size(), t))
		}
	}
}

// opregreg emits instructions for
//     dest := dest(To) op src(From)
// and also returns the created obj.Prog so it
// may be further adjusted (offset, scale, etc).
func opregreg(op obj.As, dest, src int16) *obj.Prog {
	p := gc.Prog(op)
	p.From.Type = obj.TYPE_REG
	p.To.Type = obj.TYPE_REG
	p.To.Reg = dest
	p.From.Reg = src
	return p
}

// DUFFZERO consists of repeated blocks of 4 MOVUPSs + ADD,
// See runtime/mkduff.go.
func duffStart(size int64) int64 {
	x, _ := duff(size)
	return x
}
func duffAdj(size int64) int64 {
	_, x := duff(size)
	return x
}

// duff returns the offset (from duffzero, in bytes) and pointer adjust (in bytes)
// required to use the duffzero mechanism for a block of the given size.
func duff(size int64) (int64, int64) {
	if size < 32 || size > 1024 || size%dzClearStep != 0 {
		panic("bad duffzero size")
	}
	steps := size / dzClearStep
	blocks := steps / dzBlockLen
	steps %= dzBlockLen
	off := dzBlockSize * (dzBlocks - blocks)
	var adj int64
	if steps != 0 {
		off -= dzAddSize
		off -= dzMovSize * steps
		adj -= dzClearStep * (dzBlockLen - steps)
	}
	return off, adj
}

func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
	s.SetLineno(v.Line)
	switch v.Op {
	case ssa.OpAMD64ADDQ, ssa.OpAMD64ADDL:
		r := gc.SSARegNum(v)
		r1 := gc.SSARegNum(v.Args[0])
		r2 := gc.SSARegNum(v.Args[1])
		switch {
		case r == r1:
			p := gc.Prog(v.Op.Asm())
			p.From.Type = obj.TYPE_REG
			p.From.Reg = r2
			p.To.Type = obj.TYPE_REG
			p.To.Reg = r
		case r == r2:
			p := gc.Prog(v.Op.Asm())
			p.From.Type = obj.TYPE_REG
			p.From.Reg = r1
			p.To.Type = obj.TYPE_REG
			p.To.Reg = r
		default:
			var asm obj.As
			if v.Op == ssa.OpAMD64ADDQ {
				asm = x86.ALEAQ
			} else {
				asm = x86.ALEAL
			}
			p := gc.Prog(asm)
			p.From.Type = obj.TYPE_MEM
			p.From.Reg = r1
			p.From.Scale = 1
			p.From.Index = r2
			p.To.Type = obj.TYPE_REG
			p.To.Reg = r
		}
	// 2-address opcode arithmetic
	case ssa.OpAMD64SUBQ, ssa.OpAMD64SUBL,
		ssa.OpAMD64MULQ, ssa.OpAMD64MULL,
		ssa.OpAMD64ANDQ, ssa.OpAMD64ANDL,
		ssa.OpAMD64ORQ, ssa.OpAMD64ORL,
		ssa.OpAMD64XORQ, ssa.OpAMD64XORL,
		ssa.OpAMD64SHLQ, ssa.OpAMD64SHLL,
		ssa.OpAMD64SHRQ, ssa.OpAMD64SHRL, ssa.OpAMD64SHRW, ssa.OpAMD64SHRB,
		ssa.OpAMD64SARQ, ssa.OpAMD64SARL, ssa.OpAMD64SARW, ssa.OpAMD64SARB,
		ssa.OpAMD64ADDSS, ssa.OpAMD64ADDSD, ssa.OpAMD64SUBSS, ssa.OpAMD64SUBSD,
		ssa.OpAMD64MULSS, ssa.OpAMD64MULSD, ssa.OpAMD64DIVSS, ssa.OpAMD64DIVSD,
		ssa.OpAMD64PXOR:
		r := gc.SSARegNum(v)
		if r != gc.SSARegNum(v.Args[0]) {
			v.Fatalf("input[0] and output not in same register %s", v.LongString())
		}
		opregreg(v.Op.Asm(), r, gc.SSARegNum(v.Args[1]))

	case ssa.OpAMD64DIVQ, ssa.OpAMD64DIVL, ssa.OpAMD64DIVW,
		ssa.OpAMD64DIVQU, ssa.OpAMD64DIVLU, ssa.OpAMD64DIVWU,
		ssa.OpAMD64MODQ, ssa.OpAMD64MODL, ssa.OpAMD64MODW,
		ssa.OpAMD64MODQU, ssa.OpAMD64MODLU, ssa.OpAMD64MODWU:

		// Arg[0] is already in AX as it's the only register we allow
		// and AX is the only output
		x := gc.SSARegNum(v.Args[1])

		// CPU faults upon signed overflow, which occurs when most
		// negative int is divided by -1.
		var j *obj.Prog
		if v.Op == ssa.OpAMD64DIVQ || v.Op == ssa.OpAMD64DIVL ||
			v.Op == ssa.OpAMD64DIVW || v.Op == ssa.OpAMD64MODQ ||
			v.Op == ssa.OpAMD64MODL || v.Op == ssa.OpAMD64MODW {

			var c *obj.Prog
			switch v.Op {
			case ssa.OpAMD64DIVQ, ssa.OpAMD64MODQ:
				c = gc.Prog(x86.ACMPQ)
				j = gc.Prog(x86.AJEQ)
				// go ahead and sign extend to save doing it later
				gc.Prog(x86.ACQO)

			case ssa.OpAMD64DIVL, ssa.OpAMD64MODL:
				c = gc.Prog(x86.ACMPL)
				j = gc.Prog(x86.AJEQ)
				gc.Prog(x86.ACDQ)

			case ssa.OpAMD64DIVW, ssa.OpAMD64MODW:
				c = gc.Prog(x86.ACMPW)
				j = gc.Prog(x86.AJEQ)
				gc.Prog(x86.ACWD)
			}
			c.From.Type = obj.TYPE_REG
			c.From.Reg = x
			c.To.Type = obj.TYPE_CONST
			c.To.Offset = -1

			j.To.Type = obj.TYPE_BRANCH

		}

		// for unsigned ints, we sign extend by setting DX = 0
		// signed ints were sign extended above
		if v.Op == ssa.OpAMD64DIVQU || v.Op == ssa.OpAMD64MODQU ||
			v.Op == ssa.OpAMD64DIVLU || v.Op == ssa.OpAMD64MODLU ||
			v.Op == ssa.OpAMD64DIVWU || v.Op == ssa.OpAMD64MODWU {
			c := gc.Prog(x86.AXORQ)
			c.From.Type = obj.TYPE_REG
			c.From.Reg = x86.REG_DX
			c.To.Type = obj.TYPE_REG
			c.To.Reg = x86.REG_DX
		}

		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = x

		// signed division, rest of the check for -1 case
		if j != nil {
			j2 := gc.Prog(obj.AJMP)
			j2.To.Type = obj.TYPE_BRANCH

			var n *obj.Prog
			if v.Op == ssa.OpAMD64DIVQ || v.Op == ssa.OpAMD64DIVL ||
				v.Op == ssa.OpAMD64DIVW {
				// n * -1 = -n
				n = gc.Prog(x86.ANEGQ)
				n.To.Type = obj.TYPE_REG
				n.To.Reg = x86.REG_AX
			} else {
				// n % -1 == 0
				n = gc.Prog(x86.AXORQ)
				n.From.Type = obj.TYPE_REG
				n.From.Reg = x86.REG_DX
				n.To.Type = obj.TYPE_REG
				n.To.Reg = x86.REG_DX
			}

			j.To.Val = n
			j2.To.Val = s.Pc()
		}

	case ssa.OpAMD64HMULQ, ssa.OpAMD64HMULL, ssa.OpAMD64HMULW, ssa.OpAMD64HMULB,
		ssa.OpAMD64HMULQU, ssa.OpAMD64HMULLU, ssa.OpAMD64HMULWU, ssa.OpAMD64HMULBU:
		// the frontend rewrites constant division by 8/16/32 bit integers into
		// HMUL by a constant
		// SSA rewrites generate the 64 bit versions

		// Arg[0] is already in AX as it's the only register we allow
		// and DX is the only output we care about (the high bits)
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[1])

		// IMULB puts the high portion in AH instead of DL,
		// so move it to DL for consistency
		if v.Type.Size() == 1 {
			m := gc.Prog(x86.AMOVB)
			m.From.Type = obj.TYPE_REG
			m.From.Reg = x86.REG_AH
			m.To.Type = obj.TYPE_REG
			m.To.Reg = x86.REG_DX
		}

	case ssa.OpAMD64AVGQU:
		// compute (x+y)/2 unsigned.
		// Do a 64-bit add, the overflow goes into the carry.
		// Shift right once and pull the carry back into the 63rd bit.
		r := gc.SSARegNum(v)
		if r != gc.SSARegNum(v.Args[0]) {
			v.Fatalf("input[0] and output not in same register %s", v.LongString())
		}
		p := gc.Prog(x86.AADDQ)
		p.From.Type = obj.TYPE_REG
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r
		p.From.Reg = gc.SSARegNum(v.Args[1])
		p = gc.Prog(x86.ARCRQ)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 1
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r

	case ssa.OpAMD64ADDQconst, ssa.OpAMD64ADDLconst:
		r := gc.SSARegNum(v)
		a := gc.SSARegNum(v.Args[0])
		if r == a {
			if v.AuxInt == 1 {
				var asm obj.As
				// Software optimization manual recommends add $1,reg.
				// But inc/dec is 1 byte smaller. ICC always uses inc
				// Clang/GCC choose depending on flags, but prefer add.
				// Experiments show that inc/dec is both a little faster
				// and make a binary a little smaller.
				if v.Op == ssa.OpAMD64ADDQconst {
					asm = x86.AINCQ
				} else {
					asm = x86.AINCL
				}
				p := gc.Prog(asm)
				p.To.Type = obj.TYPE_REG
				p.To.Reg = r
				return
			}
			if v.AuxInt == -1 {
				var asm obj.As
				if v.Op == ssa.OpAMD64ADDQconst {
					asm = x86.ADECQ
				} else {
					asm = x86.ADECL
				}
				p := gc.Prog(asm)
				p.To.Type = obj.TYPE_REG
				p.To.Reg = r
				return
			}
			p := gc.Prog(v.Op.Asm())
			p.From.Type = obj.TYPE_CONST
			p.From.Offset = v.AuxInt
			p.To.Type = obj.TYPE_REG
			p.To.Reg = r
			return
		}
		var asm obj.As
		if v.Op == ssa.OpAMD64ADDQconst {
			asm = x86.ALEAQ
		} else {
			asm = x86.ALEAL
		}
		p := gc.Prog(asm)
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = a
		p.From.Offset = v.AuxInt
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r

	case ssa.OpAMD64CMOVQEQconst, ssa.OpAMD64CMOVLEQconst, ssa.OpAMD64CMOVWEQconst,
		ssa.OpAMD64CMOVQNEconst, ssa.OpAMD64CMOVLNEconst, ssa.OpAMD64CMOVWNEconst:
		r := gc.SSARegNum(v)
		if r != gc.SSARegNum(v.Args[0]) {
			v.Fatalf("input[0] and output not in same register %s", v.LongString())
		}

		// Constant into AX
		p := gc.Prog(moveByType(v.Type))
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = v.AuxInt
		p.To.Type = obj.TYPE_REG
		p.To.Reg = x86.REG_AX

		p = gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = x86.REG_AX
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r

	case ssa.OpAMD64MULQconst, ssa.OpAMD64MULLconst:
		r := gc.SSARegNum(v)
		if r != gc.SSARegNum(v.Args[0]) {
			v.Fatalf("input[0] and output not in same register %s", v.LongString())
		}
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = v.AuxInt
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r
		// TODO: Teach doasm to compile the three-address multiply imul $c, r1, r2
		// then we don't need to use resultInArg0 for these ops.
		//p.From3 = new(obj.Addr)
		//p.From3.Type = obj.TYPE_REG
		//p.From3.Reg = gc.SSARegNum(v.Args[0])

	case ssa.OpAMD64SUBQconst, ssa.OpAMD64SUBLconst,
		ssa.OpAMD64ANDQconst, ssa.OpAMD64ANDLconst,
		ssa.OpAMD64ORQconst, ssa.OpAMD64ORLconst,
		ssa.OpAMD64XORQconst, ssa.OpAMD64XORLconst,
		ssa.OpAMD64SHLQconst, ssa.OpAMD64SHLLconst,
		ssa.OpAMD64SHRQconst, ssa.OpAMD64SHRLconst, ssa.OpAMD64SHRWconst, ssa.OpAMD64SHRBconst,
		ssa.OpAMD64SARQconst, ssa.OpAMD64SARLconst, ssa.OpAMD64SARWconst, ssa.OpAMD64SARBconst,
		ssa.OpAMD64ROLQconst, ssa.OpAMD64ROLLconst, ssa.OpAMD64ROLWconst, ssa.OpAMD64ROLBconst:
		r := gc.SSARegNum(v)
		if r != gc.SSARegNum(v.Args[0]) {
			v.Fatalf("input[0] and output not in same register %s", v.LongString())
		}
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = v.AuxInt
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r
	case ssa.OpAMD64SBBQcarrymask, ssa.OpAMD64SBBLcarrymask:
		r := gc.SSARegNum(v)
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = r
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r
	case ssa.OpAMD64LEAQ1, ssa.OpAMD64LEAQ2, ssa.OpAMD64LEAQ4, ssa.OpAMD64LEAQ8:
		r := gc.SSARegNum(v.Args[0])
		i := gc.SSARegNum(v.Args[1])
		p := gc.Prog(x86.ALEAQ)
		switch v.Op {
		case ssa.OpAMD64LEAQ1:
			p.From.Scale = 1
			if i == x86.REG_SP {
				r, i = i, r
			}
		case ssa.OpAMD64LEAQ2:
			p.From.Scale = 2
		case ssa.OpAMD64LEAQ4:
			p.From.Scale = 4
		case ssa.OpAMD64LEAQ8:
			p.From.Scale = 8
		}
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = r
		p.From.Index = i
		gc.AddAux(&p.From, v)
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64LEAQ:
		p := gc.Prog(x86.ALEAQ)
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.From, v)
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64CMPQ, ssa.OpAMD64CMPL, ssa.OpAMD64CMPW, ssa.OpAMD64CMPB,
		ssa.OpAMD64TESTQ, ssa.OpAMD64TESTL, ssa.OpAMD64TESTW, ssa.OpAMD64TESTB:
		opregreg(v.Op.Asm(), gc.SSARegNum(v.Args[1]), gc.SSARegNum(v.Args[0]))
	case ssa.OpAMD64UCOMISS, ssa.OpAMD64UCOMISD:
		// Go assembler has swapped operands for UCOMISx relative to CMP,
		// must account for that right here.
		opregreg(v.Op.Asm(), gc.SSARegNum(v.Args[0]), gc.SSARegNum(v.Args[1]))
	case ssa.OpAMD64CMPQconst, ssa.OpAMD64CMPLconst, ssa.OpAMD64CMPWconst, ssa.OpAMD64CMPBconst:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[0])
		p.To.Type = obj.TYPE_CONST
		p.To.Offset = v.AuxInt
	case ssa.OpAMD64TESTQconst, ssa.OpAMD64TESTLconst, ssa.OpAMD64TESTWconst, ssa.OpAMD64TESTBconst:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = v.AuxInt
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v.Args[0])
	case ssa.OpAMD64MOVLconst, ssa.OpAMD64MOVQconst:
		x := gc.SSARegNum(v)
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = v.AuxInt
		p.To.Type = obj.TYPE_REG
		p.To.Reg = x
		// If flags are live at this instruction, suppress the
		// MOV $0,AX -> XOR AX,AX optimization.
		if v.Aux != nil {
			p.Mark |= x86.PRESERVEFLAGS
		}
	case ssa.OpAMD64MOVSSconst, ssa.OpAMD64MOVSDconst:
		x := gc.SSARegNum(v)
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_FCONST
		p.From.Val = math.Float64frombits(uint64(v.AuxInt))
		p.To.Type = obj.TYPE_REG
		p.To.Reg = x
	case ssa.OpAMD64MOVQload, ssa.OpAMD64MOVSSload, ssa.OpAMD64MOVSDload, ssa.OpAMD64MOVLload, ssa.OpAMD64MOVWload, ssa.OpAMD64MOVBload, ssa.OpAMD64MOVBQSXload, ssa.OpAMD64MOVWQSXload, ssa.OpAMD64MOVLQSXload, ssa.OpAMD64MOVOload:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.From, v)
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64MOVQloadidx8, ssa.OpAMD64MOVSDloadidx8:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.From, v)
		p.From.Scale = 8
		p.From.Index = gc.SSARegNum(v.Args[1])
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64MOVLloadidx4, ssa.OpAMD64MOVSSloadidx4:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.From, v)
		p.From.Scale = 4
		p.From.Index = gc.SSARegNum(v.Args[1])
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64MOVWloadidx2:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.From, v)
		p.From.Scale = 2
		p.From.Index = gc.SSARegNum(v.Args[1])
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64MOVBloadidx1, ssa.OpAMD64MOVWloadidx1, ssa.OpAMD64MOVLloadidx1, ssa.OpAMD64MOVQloadidx1, ssa.OpAMD64MOVSSloadidx1, ssa.OpAMD64MOVSDloadidx1:
		r := gc.SSARegNum(v.Args[0])
		i := gc.SSARegNum(v.Args[1])
		if i == x86.REG_SP {
			r, i = i, r
		}
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_MEM
		p.From.Reg = r
		p.From.Scale = 1
		p.From.Index = i
		gc.AddAux(&p.From, v)
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpAMD64MOVQstore, ssa.OpAMD64MOVSSstore, ssa.OpAMD64MOVSDstore, ssa.OpAMD64MOVLstore, ssa.OpAMD64MOVWstore, ssa.OpAMD64MOVBstore, ssa.OpAMD64MOVOstore:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[1])
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.To, v)
	case ssa.OpAMD64MOVQstoreidx8, ssa.OpAMD64MOVSDstoreidx8:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[2])
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = gc.SSARegNum(v.Args[0])
		p.To.Scale = 8
		p.To.Index = gc.SSARegNum(v.Args[1])
		gc.AddAux(&p.To, v)
	case ssa.OpAMD64MOVSSstoreidx4, ssa.OpAMD64MOVLstoreidx4:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[2])
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = gc.SSARegNum(v.Args[0])
		p.To.Scale = 4
		p.To.Index = gc.SSARegNum(v.Args[1])
		gc.AddAux(&p.To, v)
	case ssa.OpAMD64MOVWstoreidx2:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[2])
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = gc.SSARegNum(v.Args[0])
		p.To.Scale = 2
		p.To.Index = gc.SSARegNum(v.Args[1])
		gc.AddAux(&p.To, v)
	case ssa.OpAMD64MOVBstoreidx1, ssa.OpAMD64MOVWstoreidx1, ssa.OpAMD64MOVLstoreidx1, ssa.OpAMD64MOVQstoreidx1, ssa.OpAMD64MOVSSstoreidx1, ssa.OpAMD64MOVSDstoreidx1:
		r := gc.SSARegNum(v.Args[0])
		i := gc.SSARegNum(v.Args[1])
		if i == x86.REG_SP {
			r, i = i, r
		}
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[2])
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = r
		p.To.Scale = 1
		p.To.Index = i
		gc.AddAux(&p.To, v)
	case ssa.OpAMD64MOVQstoreconst, ssa.OpAMD64MOVLstoreconst, ssa.OpAMD64MOVWstoreconst, ssa.OpAMD64MOVBstoreconst:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_CONST
		sc := v.AuxValAndOff()
		p.From.Offset = sc.Val()
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux2(&p.To, v, sc.Off())
	case ssa.OpAMD64MOVQstoreconstidx1, ssa.OpAMD64MOVQstoreconstidx8, ssa.OpAMD64MOVLstoreconstidx1, ssa.OpAMD64MOVLstoreconstidx4, ssa.OpAMD64MOVWstoreconstidx1, ssa.OpAMD64MOVWstoreconstidx2, ssa.OpAMD64MOVBstoreconstidx1:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_CONST
		sc := v.AuxValAndOff()
		p.From.Offset = sc.Val()
		r := gc.SSARegNum(v.Args[0])
		i := gc.SSARegNum(v.Args[1])
		switch v.Op {
		case ssa.OpAMD64MOVBstoreconstidx1, ssa.OpAMD64MOVWstoreconstidx1, ssa.OpAMD64MOVLstoreconstidx1, ssa.OpAMD64MOVQstoreconstidx1:
			p.To.Scale = 1
			if i == x86.REG_SP {
				r, i = i, r
			}
		case ssa.OpAMD64MOVWstoreconstidx2:
			p.To.Scale = 2
		case ssa.OpAMD64MOVLstoreconstidx4:
			p.To.Scale = 4
		case ssa.OpAMD64MOVQstoreconstidx8:
			p.To.Scale = 8
		}
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = r
		p.To.Index = i
		gc.AddAux2(&p.To, v, sc.Off())
	case ssa.OpAMD64MOVLQSX, ssa.OpAMD64MOVWQSX, ssa.OpAMD64MOVBQSX, ssa.OpAMD64MOVLQZX, ssa.OpAMD64MOVWQZX, ssa.OpAMD64MOVBQZX,
		ssa.OpAMD64CVTSL2SS, ssa.OpAMD64CVTSL2SD, ssa.OpAMD64CVTSQ2SS, ssa.OpAMD64CVTSQ2SD,
		ssa.OpAMD64CVTTSS2SL, ssa.OpAMD64CVTTSD2SL, ssa.OpAMD64CVTTSS2SQ, ssa.OpAMD64CVTTSD2SQ,
		ssa.OpAMD64CVTSS2SD, ssa.OpAMD64CVTSD2SS:
		opregreg(v.Op.Asm(), gc.SSARegNum(v), gc.SSARegNum(v.Args[0]))
	case ssa.OpAMD64DUFFZERO:
		off := duffStart(v.AuxInt)
		adj := duffAdj(v.AuxInt)
		var p *obj.Prog
		if adj != 0 {
			p = gc.Prog(x86.AADDQ)
			p.From.Type = obj.TYPE_CONST
			p.From.Offset = adj
			p.To.Type = obj.TYPE_REG
			p.To.Reg = x86.REG_DI
		}
		p = gc.Prog(obj.ADUFFZERO)
		p.To.Type = obj.TYPE_ADDR
		p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
		p.To.Offset = off
	case ssa.OpAMD64MOVOconst:
		if v.AuxInt != 0 {
			v.Unimplementedf("MOVOconst can only do constant=0")
		}
		r := gc.SSARegNum(v)
		opregreg(x86.AXORPS, r, r)
	case ssa.OpAMD64DUFFCOPY:
		p := gc.Prog(obj.ADUFFCOPY)
		p.To.Type = obj.TYPE_ADDR
		p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
		p.To.Offset = v.AuxInt

	case ssa.OpCopy, ssa.OpAMD64MOVQconvert: // TODO: use MOVQreg for reg->reg copies instead of OpCopy?
		if v.Type.IsMemory() {
			return
		}
		x := gc.SSARegNum(v.Args[0])
		y := gc.SSARegNum(v)
		if x != y {
			opregreg(moveByType(v.Type), y, x)
		}
	case ssa.OpLoadReg:
		if v.Type.IsFlags() {
			v.Unimplementedf("load flags not implemented: %v", v.LongString())
			return
		}
		p := gc.Prog(loadByType(v.Type))
		n, off := gc.AutoVar(v.Args[0])
		p.From.Type = obj.TYPE_MEM
		p.From.Node = n
		p.From.Sym = gc.Linksym(n.Sym)
		p.From.Offset = off
		if n.Class == gc.PPARAM || n.Class == gc.PPARAMOUT {
			p.From.Name = obj.NAME_PARAM
			p.From.Offset += n.Xoffset
		} else {
			p.From.Name = obj.NAME_AUTO
		}
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)

	case ssa.OpStoreReg:
		if v.Type.IsFlags() {
			v.Unimplementedf("store flags not implemented: %v", v.LongString())
			return
		}
		p := gc.Prog(storeByType(v.Type))
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[0])
		n, off := gc.AutoVar(v)
		p.To.Type = obj.TYPE_MEM
		p.To.Node = n
		p.To.Sym = gc.Linksym(n.Sym)
		p.To.Offset = off
		if n.Class == gc.PPARAM || n.Class == gc.PPARAMOUT {
			p.To.Name = obj.NAME_PARAM
			p.To.Offset += n.Xoffset
		} else {
			p.To.Name = obj.NAME_AUTO
		}
	case ssa.OpPhi:
		// just check to make sure regalloc and stackalloc did it right
		if v.Type.IsMemory() {
			return
		}
		f := v.Block.Func
		loc := f.RegAlloc[v.ID]
		for _, a := range v.Args {
			if aloc := f.RegAlloc[a.ID]; aloc != loc { // TODO: .Equal() instead?
				v.Fatalf("phi arg at different location than phi: %v @ %v, but arg %v @ %v\n%s\n", v, loc, a, aloc, v.Block.Func)
			}
		}
	case ssa.OpInitMem:
		// memory arg needs no code
	case ssa.OpArg:
		// input args need no code
	case ssa.OpAMD64LoweredGetClosurePtr:
		// Output is hardwired to DX only,
		// and DX contains the closure pointer on
		// closure entry, and this "instruction"
		// is scheduled to the very beginning
		// of the entry block.
	case ssa.OpAMD64LoweredGetG:
		r := gc.SSARegNum(v)
		// See the comments in cmd/internal/obj/x86/obj6.go
		// near CanUse1InsnTLS for a detailed explanation of these instructions.
		if x86.CanUse1InsnTLS(gc.Ctxt) {
			// MOVQ (TLS), r
			p := gc.Prog(x86.AMOVQ)
			p.From.Type = obj.TYPE_MEM
			p.From.Reg = x86.REG_TLS
			p.To.Type = obj.TYPE_REG
			p.To.Reg = r
		} else {
			// MOVQ TLS, r
			// MOVQ (r)(TLS*1), r
			p := gc.Prog(x86.AMOVQ)
			p.From.Type = obj.TYPE_REG
			p.From.Reg = x86.REG_TLS
			p.To.Type = obj.TYPE_REG
			p.To.Reg = r
			q := gc.Prog(x86.AMOVQ)
			q.From.Type = obj.TYPE_MEM
			q.From.Reg = r
			q.From.Index = x86.REG_TLS
			q.From.Scale = 1
			q.To.Type = obj.TYPE_REG
			q.To.Reg = r
		}
	case ssa.OpAMD64CALLstatic:
		if v.Aux.(*gc.Sym) == gc.Deferreturn.Sym {
			// Deferred calls will appear to be returning to
			// the CALL deferreturn(SB) that we are about to emit.
			// However, the stack trace code will show the line
			// of the instruction byte before the return PC.
			// To avoid that being an unrelated instruction,
			// insert an actual hardware NOP that will have the right line number.
			// This is different from obj.ANOP, which is a virtual no-op
			// that doesn't make it into the instruction stream.
			ginsnop()
		}
		p := gc.Prog(obj.ACALL)
		p.To.Type = obj.TYPE_MEM
		p.To.Name = obj.NAME_EXTERN
		p.To.Sym = gc.Linksym(v.Aux.(*gc.Sym))
		if gc.Maxarg < v.AuxInt {
			gc.Maxarg = v.AuxInt
		}
	case ssa.OpAMD64CALLclosure:
		p := gc.Prog(obj.ACALL)
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v.Args[0])
		if gc.Maxarg < v.AuxInt {
			gc.Maxarg = v.AuxInt
		}
	case ssa.OpAMD64CALLdefer:
		p := gc.Prog(obj.ACALL)
		p.To.Type = obj.TYPE_MEM
		p.To.Name = obj.NAME_EXTERN
		p.To.Sym = gc.Linksym(gc.Deferproc.Sym)
		if gc.Maxarg < v.AuxInt {
			gc.Maxarg = v.AuxInt
		}
	case ssa.OpAMD64CALLgo:
		p := gc.Prog(obj.ACALL)
		p.To.Type = obj.TYPE_MEM
		p.To.Name = obj.NAME_EXTERN
		p.To.Sym = gc.Linksym(gc.Newproc.Sym)
		if gc.Maxarg < v.AuxInt {
			gc.Maxarg = v.AuxInt
		}
	case ssa.OpAMD64CALLinter:
		p := gc.Prog(obj.ACALL)
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v.Args[0])
		if gc.Maxarg < v.AuxInt {
			gc.Maxarg = v.AuxInt
		}
	case ssa.OpAMD64NEGQ, ssa.OpAMD64NEGL,
		ssa.OpAMD64BSWAPQ, ssa.OpAMD64BSWAPL,
		ssa.OpAMD64NOTQ, ssa.OpAMD64NOTL:
		r := gc.SSARegNum(v)
		if r != gc.SSARegNum(v.Args[0]) {
			v.Fatalf("input[0] and output not in same register %s", v.LongString())
		}
		p := gc.Prog(v.Op.Asm())
		p.To.Type = obj.TYPE_REG
		p.To.Reg = r
	case ssa.OpAMD64BSFQ, ssa.OpAMD64BSFL, ssa.OpAMD64BSFW,
		ssa.OpAMD64BSRQ, ssa.OpAMD64BSRL, ssa.OpAMD64BSRW,
		ssa.OpAMD64SQRTSD:
		p := gc.Prog(v.Op.Asm())
		p.From.Type = obj.TYPE_REG
		p.From.Reg = gc.SSARegNum(v.Args[0])
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
	case ssa.OpSP, ssa.OpSB:
		// nothing to do
	case ssa.OpAMD64SETEQ, ssa.OpAMD64SETNE,
		ssa.OpAMD64SETL, ssa.OpAMD64SETLE,
		ssa.OpAMD64SETG, ssa.OpAMD64SETGE,
		ssa.OpAMD64SETGF, ssa.OpAMD64SETGEF,
		ssa.OpAMD64SETB, ssa.OpAMD64SETBE,
		ssa.OpAMD64SETORD, ssa.OpAMD64SETNAN,
		ssa.OpAMD64SETA, ssa.OpAMD64SETAE:
		p := gc.Prog(v.Op.Asm())
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)

	case ssa.OpAMD64SETNEF:
		p := gc.Prog(v.Op.Asm())
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
		q := gc.Prog(x86.ASETPS)
		q.To.Type = obj.TYPE_REG
		q.To.Reg = x86.REG_AX
		// ORL avoids partial register write and is smaller than ORQ, used by old compiler
		opregreg(x86.AORL, gc.SSARegNum(v), x86.REG_AX)

	case ssa.OpAMD64SETEQF:
		p := gc.Prog(v.Op.Asm())
		p.To.Type = obj.TYPE_REG
		p.To.Reg = gc.SSARegNum(v)
		q := gc.Prog(x86.ASETPC)
		q.To.Type = obj.TYPE_REG
		q.To.Reg = x86.REG_AX
		// ANDL avoids partial register write and is smaller than ANDQ, used by old compiler
		opregreg(x86.AANDL, gc.SSARegNum(v), x86.REG_AX)

	case ssa.OpAMD64InvertFlags:
		v.Fatalf("InvertFlags should never make it to codegen %v", v.LongString())
	case ssa.OpAMD64FlagEQ, ssa.OpAMD64FlagLT_ULT, ssa.OpAMD64FlagLT_UGT, ssa.OpAMD64FlagGT_ULT, ssa.OpAMD64FlagGT_UGT:
		v.Fatalf("Flag* ops should never make it to codegen %v", v.LongString())
	case ssa.OpAMD64REPSTOSQ:
		gc.Prog(x86.AREP)
		gc.Prog(x86.ASTOSQ)
	case ssa.OpAMD64REPMOVSQ:
		gc.Prog(x86.AREP)
		gc.Prog(x86.AMOVSQ)
	case ssa.OpVarDef:
		gc.Gvardef(v.Aux.(*gc.Node))
	case ssa.OpVarKill:
		gc.Gvarkill(v.Aux.(*gc.Node))
	case ssa.OpVarLive:
		gc.Gvarlive(v.Aux.(*gc.Node))
	case ssa.OpKeepAlive:
		if !v.Args[0].Type.IsPtrShaped() {
			v.Fatalf("keeping non-pointer alive %v", v.Args[0])
		}
		n, off := gc.AutoVar(v.Args[0])
		if n == nil {
			v.Fatalf("KeepLive with non-spilled value %s %s", v, v.Args[0])
		}
		if off != 0 {
			v.Fatalf("KeepLive with non-zero offset spill location %s:%d", n, off)
		}
		gc.Gvarlive(n)
	case ssa.OpAMD64LoweredNilCheck:
		// Optimization - if the subsequent block has a load or store
		// at the same address, we don't need to issue this instruction.
		mem := v.Args[1]
		for _, w := range v.Block.Succs[0].Block().Values {
			if w.Op == ssa.OpPhi {
				if w.Type.IsMemory() {
					mem = w
				}
				continue
			}
			if len(w.Args) == 0 || !w.Args[len(w.Args)-1].Type.IsMemory() {
				// w doesn't use a store - can't be a memory op.
				continue
			}
			if w.Args[len(w.Args)-1] != mem {
				v.Fatalf("wrong store after nilcheck v=%s w=%s", v, w)
			}
			switch w.Op {
			case ssa.OpAMD64MOVQload, ssa.OpAMD64MOVLload, ssa.OpAMD64MOVWload, ssa.OpAMD64MOVBload,
				ssa.OpAMD64MOVQstore, ssa.OpAMD64MOVLstore, ssa.OpAMD64MOVWstore, ssa.OpAMD64MOVBstore,
				ssa.OpAMD64MOVBQSXload, ssa.OpAMD64MOVWQSXload, ssa.OpAMD64MOVLQSXload,
				ssa.OpAMD64MOVSSload, ssa.OpAMD64MOVSDload, ssa.OpAMD64MOVOload,
				ssa.OpAMD64MOVSSstore, ssa.OpAMD64MOVSDstore, ssa.OpAMD64MOVOstore:
				if w.Args[0] == v.Args[0] && w.Aux == nil && w.AuxInt >= 0 && w.AuxInt < minZeroPage {
					if gc.Debug_checknil != 0 && int(v.Line) > 1 {
						gc.Warnl(v.Line, "removed nil check")
					}
					return
				}
			case ssa.OpAMD64MOVQstoreconst, ssa.OpAMD64MOVLstoreconst, ssa.OpAMD64MOVWstoreconst, ssa.OpAMD64MOVBstoreconst:
				off := ssa.ValAndOff(v.AuxInt).Off()
				if w.Args[0] == v.Args[0] && w.Aux == nil && off >= 0 && off < minZeroPage {
					if gc.Debug_checknil != 0 && int(v.Line) > 1 {
						gc.Warnl(v.Line, "removed nil check")
					}
					return
				}
			}
			if w.Type.IsMemory() {
				if w.Op == ssa.OpVarDef || w.Op == ssa.OpVarKill || w.Op == ssa.OpVarLive {
					// these ops are OK
					mem = w
					continue
				}
				// We can't delay the nil check past the next store.
				break
			}
		}
		// Issue a load which will fault if the input is nil.
		// TODO: We currently use the 2-byte instruction TESTB AX, (reg).
		// Should we use the 3-byte TESTB $0, (reg) instead?  It is larger
		// but it doesn't have false dependency on AX.
		// Or maybe allocate an output register and use MOVL (reg),reg2 ?
		// That trades clobbering flags for clobbering a register.
		p := gc.Prog(x86.ATESTB)
		p.From.Type = obj.TYPE_REG
		p.From.Reg = x86.REG_AX
		p.To.Type = obj.TYPE_MEM
		p.To.Reg = gc.SSARegNum(v.Args[0])
		gc.AddAux(&p.To, v)
		if gc.Debug_checknil != 0 && v.Line > 1 { // v.Line==1 in generated wrappers
			gc.Warnl(v.Line, "generated nil check")
		}
	default:
		v.Unimplementedf("genValue not implemented: %s", v.LongString())
	}
}

var blockJump = [...]struct {
	asm, invasm obj.As
}{
	ssa.BlockAMD64EQ:  {x86.AJEQ, x86.AJNE},
	ssa.BlockAMD64NE:  {x86.AJNE, x86.AJEQ},
	ssa.BlockAMD64LT:  {x86.AJLT, x86.AJGE},
	ssa.BlockAMD64GE:  {x86.AJGE, x86.AJLT},
	ssa.BlockAMD64LE:  {x86.AJLE, x86.AJGT},
	ssa.BlockAMD64GT:  {x86.AJGT, x86.AJLE},
	ssa.BlockAMD64ULT: {x86.AJCS, x86.AJCC},
	ssa.BlockAMD64UGE: {x86.AJCC, x86.AJCS},
	ssa.BlockAMD64UGT: {x86.AJHI, x86.AJLS},
	ssa.BlockAMD64ULE: {x86.AJLS, x86.AJHI},
	ssa.BlockAMD64ORD: {x86.AJPC, x86.AJPS},
	ssa.BlockAMD64NAN: {x86.AJPS, x86.AJPC},
}

var eqfJumps = [2][2]gc.FloatingEQNEJump{
	{{Jump: x86.AJNE, Index: 1}, {Jump: x86.AJPS, Index: 1}}, // next == b.Succs[0]
	{{Jump: x86.AJNE, Index: 1}, {Jump: x86.AJPC, Index: 0}}, // next == b.Succs[1]
}
var nefJumps = [2][2]gc.FloatingEQNEJump{
	{{Jump: x86.AJNE, Index: 0}, {Jump: x86.AJPC, Index: 1}}, // next == b.Succs[0]
	{{Jump: x86.AJNE, Index: 0}, {Jump: x86.AJPS, Index: 0}}, // next == b.Succs[1]
}

func ssaGenBlock(s *gc.SSAGenState, b, next *ssa.Block) {
	s.SetLineno(b.Line)

	switch b.Kind {
	case ssa.BlockPlain, ssa.BlockCall, ssa.BlockCheck:
		if b.Succs[0].Block() != next {
			p := gc.Prog(obj.AJMP)
			p.To.Type = obj.TYPE_BRANCH
			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[0].Block()})
		}
	case ssa.BlockDefer:
		// defer returns in rax:
		// 0 if we should continue executing
		// 1 if we should jump to deferreturn call
		p := gc.Prog(x86.ATESTL)
		p.From.Type = obj.TYPE_REG
		p.From.Reg = x86.REG_AX
		p.To.Type = obj.TYPE_REG
		p.To.Reg = x86.REG_AX
		p = gc.Prog(x86.AJNE)
		p.To.Type = obj.TYPE_BRANCH
		s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[1].Block()})
		if b.Succs[0].Block() != next {
			p := gc.Prog(obj.AJMP)
			p.To.Type = obj.TYPE_BRANCH
			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[0].Block()})
		}
	case ssa.BlockExit:
		gc.Prog(obj.AUNDEF) // tell plive.go that we never reach here
	case ssa.BlockRet:
		gc.Prog(obj.ARET)
	case ssa.BlockRetJmp:
		p := gc.Prog(obj.AJMP)
		p.To.Type = obj.TYPE_MEM
		p.To.Name = obj.NAME_EXTERN
		p.To.Sym = gc.Linksym(b.Aux.(*gc.Sym))

	case ssa.BlockAMD64EQF:
		gc.SSAGenFPJump(s, b, next, &eqfJumps)

	case ssa.BlockAMD64NEF:
		gc.SSAGenFPJump(s, b, next, &nefJumps)

	case ssa.BlockAMD64EQ, ssa.BlockAMD64NE,
		ssa.BlockAMD64LT, ssa.BlockAMD64GE,
		ssa.BlockAMD64LE, ssa.BlockAMD64GT,
		ssa.BlockAMD64ULT, ssa.BlockAMD64UGT,
		ssa.BlockAMD64ULE, ssa.BlockAMD64UGE:
		jmp := blockJump[b.Kind]
		likely := b.Likely
		var p *obj.Prog
		switch next {
		case b.Succs[0].Block():
			p = gc.Prog(jmp.invasm)
			likely *= -1
			p.To.Type = obj.TYPE_BRANCH
			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[1].Block()})
		case b.Succs[1].Block():
			p = gc.Prog(jmp.asm)
			p.To.Type = obj.TYPE_BRANCH
			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[0].Block()})
		default:
			p = gc.Prog(jmp.asm)
			p.To.Type = obj.TYPE_BRANCH
			s.Branches = append(s.Branches, gc.Branch{P: p, B: b.Succs[0].Block()})
			q := gc.Prog(obj.AJMP)
			q.To.Type = obj.TYPE_BRANCH
			s.Branches = append(s.Branches, gc.Branch{P: q, B: b.Succs[1].Block()})
		}

		// liblink reorders the instruction stream as it sees fit.
		// Pass along what we know so liblink can make use of it.
		// TODO: Once we've fully switched to SSA,
		// make liblink leave our output alone.
		switch likely {
		case ssa.BranchUnlikely:
			p.From.Type = obj.TYPE_CONST
			p.From.Offset = 0
		case ssa.BranchLikely:
			p.From.Type = obj.TYPE_CONST
			p.From.Offset = 1
		}

	default:
		b.Unimplementedf("branch not implemented: %s. Control: %s", b.LongString(), b.Control.LongString())
	}
}