ORPD    X0, X2 // X2= f1
	SHRQ    $52, BX
	ANDL    $0x7FF, BX
	SUBL    $0x3FE, BX
	XORPS   X1, X1 // break dependency for CVTSL2SD
	CVTSL2SD BX, X1 // x1= k, x2= f1
	// if f1 < math.Sqrt2/2 { k -= 1; f1 *= 2 }
	MOVSD   $HSqrt2, X0 // x0= 0.7071, x1= k, x2= f1
	CMPSD   X2, X0, 5 // cmpnlt; x0= 0 or ^0, x1= k, x2 = f1
	MOVSD   $1.0, X3 // x0= 0 or ^0, x1= k, x2 = f1, x3= 1
	ANDPD   X0, X3 // x0= 0 or ^0, x1= k, x2 = f1, x3= 0 or 1

// math package intrinsics
(Sqrt      ...) => (FSQRT ...)
(Floor       x) => (FIDBR [7] x)
(Ceil        x) => (FIDBR [6] x)
(Trunc       x) => (FIDBR [5] x)
(RoundToEven x) => (FIDBR [4] x)
(Round       x) => (FIDBR [1] x)
(FMA     x y z) => (FMADD z x y)

(Sqrt32    ...) => (FSQRTS ...)

// Atomic loads and stores.

(Bswap16 x) => (ROLWconst [8] x)

(PopCount(64|32) ...) => (POPCNT(Q|L) ...)
(PopCount16 x) => (POPCNTL (MOVWQZX <typ.UInt32> x))
(PopCount8 x) => (POPCNTL (MOVBQZX <typ.UInt32> x))

(Sqrt ...) => (SQRTSD ...)
(Sqrt32 ...) => (SQRTSS ...)

(RoundToEven x) => (ROUNDSD [0] x)
(Floor x)       => (ROUNDSD [1] x)
(Ceil x)        => (ROUNDSD [2] x)
(Trunc x)       => (ROUNDSD [3] x)

(FMA x y z) => (VFMADD231SD z x y)

(Or(32|16|8) ...) => (OR ...)
(Xor(32|16|8) ...) => (XOR ...)

// unary ops
(Neg(32|16|8) x) => (RSBconst [0] x)
(Neg(32|64)F ...) => (NEG(F|D) ...)

(Com(32|16|8) ...) => (MVN ...)

(Sqrt ...) => (SQRTD ...)
(Sqrt32 ...) => (SQRTF ...)
(Abs ...) => (ABSD ...)

// TODO: optimize this for ARMv5 and ARMv6
(Ctz32NonZero ...) => (Ctz32 ...)
(Ctz16NonZero ...) => (Ctz32 ...)
(Ctz8NonZero ...) => (Ctz32 ...)

(Com(64|32|16|8) ...) => (MVN ...)

// math package intrinsics
(Abs         ...) => (FABSD   ...)
(Sqrt        ...) => (FSQRTD  ...)
(Ceil        ...) => (FRINTPD ...)
(Floor       ...) => (FRINTMD ...)
(Round       ...) => (FRINTAD ...)
(RoundToEven ...) => (FRINTND ...)
(Trunc       ...) => (FRINTZD ...)
(FMA       x y z) => (FMADDD z x y)

(Sqrt32 ...) => (FSQRTS ...)

(Min(64|32)F ...) => (FMIN(D|S) ...)

func zero[T Complex]() T {
	return T(0)
}
func pi[T Complex]() T {
	return T(3.14)
}
func sqrtN1[T Complex]() T {
	return T(-1i)
}

func main() {
	fmt.Println(zero[complex128]())
	fmt.Println(pi[complex128]())
	fmt.Println(sqrtN1[complex128]())
	fmt.Println(zero[complex64]())
	fmt.Println(pi[complex64]())
	fmt.Println(sqrtN1[complex64]())

(Div64F x (Const64F <t> [c])) && reciprocalExact64(c) => (Mul64F x (Const64F <t> [1/c]))

// rewrite single-precision sqrt expression "float32(math.Sqrt(float64(x)))"
(Cvt64Fto32F sqrt0:(Sqrt (Cvt32Fto64F x))) && sqrt0.Uses==1 => (Sqrt32 x)

(Sqrt (Const64F [c])) && !math.IsNaN(math.Sqrt(c)) => (Const64F [math.Sqrt(c)])

// for rewriting results of some late-expanded rewrites (below)
(SelectN [0] (MakeResult x ___)) => x

// and "bias" are found in src/math/bits.go

// Sqrt returns the square root of x.
//
// Special cases are:
//
//	Sqrt(+Inf) = +Inf
//	Sqrt(±0) = ±0
//	Sqrt(x < 0) = NaN
//	Sqrt(NaN) = NaN
func Sqrt(x float64) float64 {
	return sqrt(x)
}

// Note: On systems where Sqrt is a single instruction, the compiler
// may turn a direct call into a direct use of that instruction instead.

import (
	"math"
	"sync"
)

var threeOnce struct {
	sync.Once
	v *Float
}

func three() *Float {
	threeOnce.Do(func() {
		threeOnce.v = NewFloat(3.0)
	})
	return threeOnce.v
}

// Sqrt sets z to the rounded square root of x, and returns it.
//
// If z's precision is 0, it is changed to x's precision before the
// operation. Rounding is performed according to z's precision and

type Pos struct {
	base      *PosBase
	line, col uint32
}

// MakePos returns a new Pos for the given PosBase, line and column.
func MakePos(base *PosBase, line, col uint) Pos { return Pos{base, sat32(line), sat32(col)} }

// TODO(gri) IsKnown makes an assumption about linebase < 1.
// Maybe we should check for Base() != nil instead.

func (pos Pos) Pos() Pos       { return pos }