{name: "PopCount32", argLength: 1}, // Count bits in arg[0]
	{name: "PopCount64", argLength: 1}, // Count bits in arg[0]

	// RotateLeftX instructions rotate the X bits of arg[0] to the left
	// by the low lg_2(X) bits of arg[1], interpreted as an unsigned value.
	// Note that this works out regardless of the bit width or signedness of
	// arg[1]. In particular, RotateLeft by x is the same as RotateRight by -x.

	r1 := mul64(l0_2, l1)
	r1 = addMul64(r1, l2_38, l4)
	r1 = addMul64(r1, l3_19, l3)

	// r2 = l0×l2 + l1×l1 + l2×l0 + 19×(l3×l4 + l4×l3) = 2×l0×l2 + l1×l1 + 19×2×l3×l4
	r2 := mul64(l0_2, l2)
	r2 = addMul64(r2, l1, l1)
	r2 = addMul64(r2, l3_38, l4)

	// r3 = l0×l3 + l1×l2 + l2×l1 + l3×l0 + 19×l4×l4 = 2×l0×l3 + 2×l1×l2 + 19×l4×l4
	r3 := mul64(l0_2, l3)
	r3 = addMul64(r3, l1_2, l2)

	SUB	$4, R2, R6	// offset of the last 4 bytes
	MOVWU	(R0)(R6), R4
	MOVWU	(R1)(R6), R5
	EOR	R4, R5
	CBNZ	R5, not_equal
	B	equal
lt_4:
	TBZ	$1, R2, lt_2
	MOVHU.P	2(R0), R4
	MOVHU.P	2(R1), R5
	CMP	R4, R5
	BNE	not_equal
lt_2:
	TBZ	$0, R2, equal
one:
	MOVBU	(R0), R4
	MOVBU	(R1), R5
	CMP	R4, R5
	BNE	not_equal
equal:
	MOVD	$1, R0
	RET
not_equal:
	MOVB	ZR, R0

	TBZ	$2, R6, lt_4
	MOVWU	(R0), R4
	MOVWU	(R2), R5
	CMPW	R4, R5
	BNE	cmp
	SUBS	$4, R6
	BEQ	samebytes
	ADD	$4, R0
	ADD	$4, R2
lt_4:
	TBZ	$1, R6, lt_2
	MOVHU	(R0), R4
	MOVHU	(R2), R5
	CMPW	R4, R5
	BNE	cmp
	ADD	$2, R0
	ADD	$2, R2
lt_2:
	TBZ	$0, R6, samebytes
one:
	MOVBU	(R0), R4
	MOVBU	(R2), R5
	CMPW	R4, R5
	BNE	ret
samebytes:
	CMP	R3, R1
	CSET	NE, R0
	CNEG	LO, R0, R0

	resize   string // code to shrink to sub-power-of-two size (takes size as fmt arg)
	clear    string // code for clearing object before putting it on the free list
	minLog   int    // log_2 of minimum allocation size
	maxLog   int    // log_2 of maximum allocation size
}

type derived struct {
	name string // name for alloc/free functions
	typ  string // the type they return/accept
	base string // underlying allocator
}

  // Running example:
  // indices = {2, [1, 0]}
  // data = {[d_1, d_2], [[d_3, d_4], [d_5, d_6]]}
  // out = [[d_5, d_6], [d_3, d_4], [d_1, d_2]]
  // grad = [[g_1, g_2], [g_3, g_4], [g_5, g_6]]

  // indices and data are two equal-sized lists passed
  // into DynamicStitch.
  // num_values = 2
  int32_t num_values = op.num_inputs() / 2;

// https://github.com/llvm/llvm-project/blob/704d92607d26e696daba596b72cb70effe79a872/compiler-rt/lib/fuzzer/FuzzerValueBitMap.h#L34
// ValueProfileMap.AddValue() truncates its argument to 16 bits and shifts the
// PC to the left by log_2(128)=7, which means that only the lowest 16 - 7 bits
// of the return address matter. String compare hooks use the lowest 12 bits,
// but take the return address as an argument and thus don't require the

	// ~70ms on a 1.6GHz Zeon.
	// The iterate/delete idiom currently takes expected
	// O(n lg n) time.  Fortunately, the checkLinear test
	// leaves enough wiggle room to include n lg n time
	// (it actually tests for O(n^log_2(3)).
	// To prevent false positives, average away variation
	// by doing multiple rounds within a single run.
	checkLinear("iterdelete", 2500, func(n int) {
		for round := 0; round < 4; round++ {
			m := map[int]int{}

	// an initial estimate for sqrt(2), and then iterate:
	//     x_{n+1} = 1/2 * ( x_n + (2.0 / x_n) )

	// Since Newton's Method doubles the number of correct digits at each
	// iteration, we need at least log_2(prec) steps.
	steps := int(math.Log2(prec))

	// Initialize values we need for the computation.
	two := new(big.Float).SetPrec(prec).SetInt64(2)
	half := new(big.Float).SetPrec(prec).SetFloat64(0.5)

  public static int log10(BigInteger x, RoundingMode mode) {
    checkPositive("x", x);
    if (fitsInLong(x)) {
      return LongMath.log10(x.longValue(), mode);
    }

    int approxLog10 = (int) (log2(x, FLOOR) * LN_2 / LN_10);
    BigInteger approxPow = BigInteger.TEN.pow(approxLog10);
    int approxCmp = approxPow.compareTo(x);

    /*
     * We adjust approxLog10 and approxPow until they're equal to floor(log10(x)) and