// real(x) < 0 is fine for the complex case
    auto log_x = Where3(scope, NotEqual(scope, x, zero), Log(scope, x),
                        ZerosLike(scope, x));
    auto gy_1 = Mul(scope, Mul(scope, grad, z), log_x);
    return BinaryGradCommon(scope, op, grad_outputs, gx_1, gy_1);
  } else {
    // There's no sensible real value to return if x < 0, so return 0
    auto log_x = Where3(scope, Greater(scope, x, zero), Log(scope, x),

	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
}

// The README file says, "The software is in public domain.
// You can use the software without any obligation."
//
// This code is a simplified version of the original.

#define LN2 0.6931471805599453094172321214581766 // log_e(2)
#define LOG2E 1.4426950408889634073599246810018920 // 1/LN2
#define LN2U 0.69314718055966295651160180568695068359375 // upper half LN2
#define LN2L 0.28235290563031577122588448175013436025525412068e-12 // lower half LN2

                const char* name = nullptr,
                const char* raw_device_name = nullptr);

// Computes natural logarithm of (1 + x) element-wise.
Status Log1p(AbstractContext* ctx, AbstractTensorHandle* const x,
             AbstractTensorHandle** y, const char* name = nullptr,
             const char* raw_device_name = nullptr);

}  // namespace ops
}  // namespace tensorflow

	// The probability distribution function is mean*exp(-mean*x), so the CDF is
	// p = 1 - exp(-mean*x), so
	// q = 1 - p == exp(-mean*x)
	// log_e(q) = -mean*x
	// -log_e(q)/mean = x
	// x = -log_e(q) * mean
	// x = log_2(q) * (-log_e(2)) * mean    ; Using log_2 for efficiency
	const randomBitCount = 26
	q := cheaprandn(1<<randomBitCount) + 1
	qlog := fastlog2(float64(q)) - randomBitCount
	if qlog > 0 {

		p(b, "	if l%ssel++; l%ssel >= 2 { o.log(`loop L%ssel`); return -1 }\n", s, s, s)
		p(b, "	l%s := 0\n", s)
		p(b, "goto L%s; L%s:	for %s i%s := range %s {\n", s, s, prefix, s, rangeExpr)
		p(b, "	o.log1(`L%s top`, i%s)\n", s, s)
		p(b, "	if l%s++; l%s >= 4 { o.log(`loop L%s`); return -1 }\n", s, s, s)
		printTests := func() {
			if code++; allowed(code) {
				p(b, "	if code == %v { break }\n", code)
			}

	// 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)

	// ~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{}

// 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

//CHECK-NEXT:}

func.func @log(%arg0: tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32> {
  %0 = "vhlo.log_v1" (%arg0) : (tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32>
  func.return %0 : tensor<1x1x1x96xf32>
}

//CHECK:func.func private @log(%arg0: tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32> {
//CHECK-NEXT: %0 = "vhlo.log_v1"(%arg0) : (tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32>
//CHECK-NEXT: return %0 : tensor<1x1x1x96xf32>