"MULU",
	"MULH",
	"MULHU",
	"MULW",
	"NEGD",
	"NEGF",
	"NEGW",
	"NEGV",
	"NOOP",
	"NOR",
	"OR",
	"REM",
	"REMU",
	"RFE",
	"SC",
	"SCV",
	"SGT",
	"SGTU",
	"SLL",
	"SQRTD",
	"SQRTF",
	"SRA",
	"SRL",
	"ROTR",
	"SUB",
	"SUBD",
	"SUBF",
	"SUBU",
	"SUBW",
	"DBAR",
	"SYSCALL",
	"TEQ",
	"TNE",
	"WORD",
	"XOR",
	"MASKEQZ",
	"MASKNEZ",
	"MOVV",

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

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

 public:
  explicit SqrtGradientFunction(AbstractTensorHandle* sqrt) : sqrt_(sqrt) {
    sqrt->Ref();
  }
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    std::string name = "Sqrt_Grad";
    TF_RETURN_IF_ERROR(SqrtGrad(ctx, sqrt_.get(), grad_outputs[0],

subsequent non-comment line. Inline comments are also supported; the
regexp will be matched against the code found on the same line:

  func Sqrt(x float64) float64 {
  	   return math.Sqrt(x) // arm:"SQRTD"
  }

It's possible to specify a comma-separated list of regexps to be
matched. For example, the following test:

  func TZ8(n uint8) int {
  	   // amd64:"BSFQ","ORQ\t\\$256"

//
// __ieee754_log(x)
// Return the logarithm of x
//
// Method :
//   1. Argument Reduction: find k and f such that
//			x = 2**k * (1+f),
//	   where  sqrt(2)/2 < 1+f < sqrt(2) .
//
//   2. Approximation of log(1+f).
//	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
//		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
//	     	 = 2s + s*R

	DIVF.NE	F0, F2        // 002a821e
	DIVD	F3, F5        // 035b85ee
	NEGF	F0, F1        // 401ab1ee
	NEGD	F4, F5        // 445bb1ee
	ABSF	F0, F1        // c01ab0ee
	ABSD	F4, F5        // c45bb0ee
	SQRTF	F0, F1        // c01ab1ee
	SQRTD	F4, F5        // c45bb1ee
	MOVFD	F0, F1        // c01ab7ee
	MOVDF	F4, F5        // c45bb7ee

	LDREX	(R8), R9      // 9f9f98e1
	LDREXD	(R11), R12    // 9fcfbbe1

		{name: "NEGD", argLength: 1, reg: fp11, asm: "NEGD"},   // -arg0, float64
		{name: "ABSD", argLength: 1, reg: fp11, asm: "ABSD"},   // abs(arg0), float64
		{name: "SQRTD", argLength: 1, reg: fp11, asm: "SQRTD"}, // sqrt(arg0), float64
		{name: "SQRTF", argLength: 1, reg: fp11, asm: "SQRTF"}, // sqrt(arg0), float32

		// shifts
		{name: "SLL", argLength: 2, reg: gp21, asm: "SLL"},                    // arg0 << arg1, shift amount is mod 32

		{name: "NEGF", argLength: 1, reg: fp11, asm: "NEGF"},   // -arg0, float32
		{name: "NEGD", argLength: 1, reg: fp11, asm: "NEGD"},   // -arg0, float64
		{name: "SQRTD", argLength: 1, reg: fp11, asm: "SQRTD"}, // sqrt(arg0), float64
		{name: "SQRTF", argLength: 1, reg: fp11, asm: "SQRTF"}, // sqrt(arg0), float32

		{name: "MASKEQZ", argLength: 2, reg: gp21, asm: "MASKEQZ"}, // returns 0 if arg1 == 0, otherwise returns arg0

		{name: "NEGD", argLength: 1, reg: fp11, asm: "NEGD"},   // -arg0, float64
		{name: "ABSD", argLength: 1, reg: fp11, asm: "ABSD"},   // abs(arg0), float64
		{name: "SQRTD", argLength: 1, reg: fp11, asm: "SQRTD"}, // sqrt(arg0), float64
		{name: "SQRTF", argLength: 1, reg: fp11, asm: "SQRTF"}, // sqrt(arg0), float32

		// shifts
		{name: "SLLV", argLength: 2, reg: gp21, asm: "SLLV"},                    // arg0 << arg1, shift amount is mod 64

	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