TF_DataType, const int64_t* dims, int num_dims, void* data, size_t len,
    void (*deallocator)(void* data, size_t len, void* arg),
    void* deallocator_arg);

// Returns the alignment, in bytes, required for allocating aligned tensors.
//
// This can be used in combination with TF_NewTensor to manually manage
// memory while ensuring the resulting tensors satisfy TensorFlow's
// memory alignment preferences.

entry:
	ADDV	R4, R14, R12	// R6 start of a, R14 end of a
	BEQ	R4, R12, samebytes // length is 0

	SRLV	$4, R14		// R14 is number of chunks
	BEQ	R0, R14, byte_loop

	// make sure both a and b are aligned.
	OR	R4, R6, R15
	AND	$7, R15
	BNE	R0, R15, byte_loop

	PCALIGN	$16
chunk16_loop:
	BEQ	R0, R14, byte_loop
	MOVV	(R4), R8
	MOVV	(R6), R9
	BNE	R8, R9, byte_loop
	MOVV	8(R4), R16
	MOVV	8(R6), R17

	// A hint is considered socket aligned if sockets across which numa nodes span is equal to minSockets
	minSockets := (minAffinitySize + numaNodesPerSocket - 1) / numaNodesPerSocket
	return p.topology.CPUDetails.SocketsInNUMANodes(numaNodesBitMask...).Size() == minSockets
}

// getAlignedCPUs return set of aligned CPUs based on numa affinity mask and configured policy options.

	case *Struct:
		if len(t.fields) == 0 && IsSyncAtomicAlign64(T) {
			// Special case: sync/atomic.align64 is an
			// empty struct we recognize as a signal that
			// the struct it contains must be
			// 64-bit-aligned.
			//
			// This logic is equivalent to the logic in
			// cmd/compile/internal/types/size.go:calcStructOffset
			return 8
		}

		// spec: "For a variable x of struct type: unsafe.Alignof(x)

	// Alignment is maximum alignment required by the architecture
	// for any (compiler-generated) load or store instruction.
	// Loads or stores smaller than Alignment must be naturally aligned.
	// Loads or stores larger than Alignment need only be Alignment-aligned.
	Alignment int8

	// CanMergeLoads reports whether the backend optimization passes
	// can combine adjacent loads into a single larger, possibly unaligned, load.

	AND	R1, R3
#ifdef GOARCH_mips
	// Big endian.  ptr = ptr ^ 3
	XOR	$3, R1
#endif
	AND	$3, R1, R4	// R4 = ((ptr & 3) * 8)
	SLL	$3, R4
	SLL	R4, R2, R2	// Shift val for aligned ptr. R2 = val << R4
	SYNC
try_or8:
	LL	(R3), R4	// R4 = *R3
	OR	R2, R4
	SC	R4, (R3)	// *R3 = R4
	BEQ	R4, try_or8
	SYNC
	RET

// void	And8(byte volatile*, byte);
TEXT ·And8(SB),NOSPLIT,$0-5

	// bitfield types. The order in which bitfields are laid out
	// in memory is implementation defined, so we can't easily
	// know how a bitfield should correspond to a Go type, even if
	// it appears to be aligned correctly.
	bitfieldType := reflect.TypeOf(bitfields{})
	check := func(name string) {
		_, ok := bitfieldType.FieldByName(name)
		if ok {
			fmt.Fprintf(os.Stderr, "found unexpected bitfields field %s\n", name)

// the compiler word, the link editor word, and the TOC save word.
const MinFrameSize = goarch.MinFrameSize

// StackAlign is the required alignment of the SP register.
// The stack must be at least word aligned, but some architectures require more.

// There are 2 arg bitmaps for this function, each with 2 bits.
// In the first, p and q are both live, so that bitmap is 11.
// In the second, only p is live, so that bitmap is 10.
// Bitmaps are byte aligned, so if the first bitmap is interpreted as
// extending across the entire argument area, we incorrectly concatenate
// the bitmaps and end up using 110000001. That bad bitmap causes a6
// to be considered a pointer.

 * units away from 0 and 0.499 towards 0. The smallest unit this returns is "µs"; the largest unit
 * it returns is "s". For values in [-499..499] this returns "  0 µs".
 *
 * The returned string attempts to be column-aligned to 6 characters. For negative and large values
 * the returned string may be longer.
 */
fun formatDuration(ns: Long): String {
  val s =
    when {
      ns <= -999_500_000 -> "${(ns - 500_000_000) / 1_000_000_000} s "