sink64[4] = math.RoundToEven(x)
}

func sqrt(x float64) float64 {
	// amd64:"SQRTSD"
	// 386/sse2:"SQRTSD" 386/softfloat:-"SQRTD"
	// arm64:"FSQRTD"
	// arm/7:"SQRTD"
	// mips/hardfloat:"SQRTD" mips/softfloat:-"SQRTD"
	// mips64/hardfloat:"SQRTD" mips64/softfloat:-"SQRTD"
	// wasm:"F64Sqrt"
	// ppc64x:"FSQRT"
	// riscv64: "FSQRTD"
	return math.Sqrt(x)
}

func sqrt32(x float32) float32 {
	// amd64:"SQRTSS"

	Phi = 1.61803398874989484820458683436563811772030917980576286213544862 // https://oeis.org/A001622

	Sqrt2   = 1.41421356237309504880168872420969807856967187537694807317667974 // https://oeis.org/A002193
	SqrtE   = 1.64872127070012814684865078781416357165377610071014801157507931 // https://oeis.org/A019774
	SqrtPi  = 1.77245385090551602729816748334114518279754945612238712821380779 // https://oeis.org/A002161

	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

// 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.

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 }

regexp to be matched. For example, the following test:

  func Sqrt(x float64) float64 {
  	   // amd64:"SQRTSD"
  	   // arm64:"FSQRTD"
  	   return math.Sqrt(x)
  }

verifies that math.Sqrt calls are intrinsified to a SQRTSD instruction
on amd64, and to a FSQRTD instruction on arm64.

It is possible to put multiple architectures checks into the same
line, as:

  // amd64:"SQRTSD" arm64:"FSQRTD"

	// amd64/v3:"BLSRL",-"TESTL",-"CALL"
	if isNotPowerOfTwo32(x) {
		a(true)
	} else {
		b("false")
	}
}

func sarx64(x, y int64) int64 {
	// amd64/v3:"SARXQ"
	return x >> y
}

func sarx32(x, y int32) int32 {
	// amd64/v3:"SARXL"
	return x >> y
}

func sarx64_load(x []int64, i int) int64 {
	// amd64/v3: `SARXQ\t[A-Z]+[0-9]*, \([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*8\), [A-Z]+[0-9]*`

TAC_REGISTER_GPU_OP(SinOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(SliceOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(SoftmaxOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(SpaceToDepthOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(SqrtOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(SquareOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(SquaredDifferenceOp, CreateBasicOpNoCost);
TAC_REGISTER_GPU_OP(StridedSliceOp, CreateBasicOpNoCost);

			return true
		}
	}
	return false
}

// alpha returns 1 if i is 0 and returns √2 otherwise.
func alpha(i int) float64 {
	if i == 0 {
		return 1
	}
	return math.Sqrt2
}

var cosines [32]float64 // cosines[k] = cos(π/2 * k/8)

func init() {
	for k := range cosines {
		cosines[k] = math.Cos(math.Pi * float64(k) / 16)
	}
}