p = alignUp(p, align)
		p2 := sysReserve(unsafe.Pointer(p), size)
		if p != uintptr(p2) {
			// Must have raced. Try again.
			sysFreeOS(p2, size)
			if retries++; retries == 100 {
				throw("failed to allocate aligned heap memory; too many retries")
			}
			goto retry
		}
		// Success.
		return p2, size
	default:
		// Trim off the unaligned parts.
		pAligned := alignUp(p, align)

}

// The pointer tag, OR-ed into the XlaTensor's address to distinguish it from
// device-side tensors, which are either CPU or GPU memory pointers. This works
// because we're guaranteed that CPU and GPU pointers are aligned to > 1 bits.
namespace {
constexpr uintptr_t kTag = 0x1ULL;
}

/*static*/ XlaTensor* XlaTensor::FromOpaquePointer(void* ptr) {
  uintptr_t value = reinterpret_cast<uintptr_t>(ptr);
  if (value & kTag) {

//
// This method is a helper for dealing with the endianness of different CPU
// architectures, since sub-word-sized arguments in big endian architectures
// need to be "aligned" to the upper edge of the register to be interpreted
// by the CPU correctly.
func (r *RegArgs) IntRegArgAddr(reg int, argSize uintptr) unsafe.Pointer {
	if argSize > goarch.PtrSize || argSize == 0 || argSize&(argSize-1) != 0 {

TEXT ·Or8(SB), NOSPLIT, $0-9
	MOVV	ptr+0(FP), R1
	MOVBU	val+8(FP), R2
	// Align ptr down to 4 bytes so we can use 32-bit load/store.
	MOVV	$~3, R3
	AND	R1, R3
	// Compute val shift.
#ifdef GOARCH_mips64
	// Big endian.  ptr = ptr ^ 3
	XOR	$3, R1
#endif
	// R4 = ((ptr & 3) * 8)
	AND	$3, R1, R4
	SLLV	$3, R4
	// Shift val for aligned ptr. R2 = val << R4
	SLLV	R4, R2

	SYNC
	LL	(R3), R4
	OR	R2, R4

	// get to really high addresses and panic if it does.
	addrBits = 48

	// In addition to the 16 bits taken from the top, we can take 3 from the
	// bottom, because node must be pointer-aligned, giving a total of 19 bits
	// of count.
	tagBits = 64 - addrBits + 3

	// On AIX, 64-bit addresses are split into 36-bit segment number and 28-bit

	// Flag to evenly distribute CPUs across NUMA nodes in cases where more
	// than one NUMA node is required to satisfy the allocation.
	DistributeCPUsAcrossNUMA bool
	// Flag to ensure CPUs are considered aligned at socket boundary rather than
	// NUMA boundary
	AlignBySocket bool
}

// NewStaticPolicyOptions creates a StaticPolicyOptions struct from the user configuration.

	// Typically a non-fixed seed should be used, such as time.Now().UnixNano().
	// Using a fixed seed will produce the same output on every run.
	r := rand.New(rand.NewSource(99))

	// The tabwriter here helps us generate aligned output.
	w := tabwriter.NewWriter(os.Stdout, 1, 1, 1, ' ', 0)
	defer w.Flush()
	show := func(name string, v1, v2, v3 any) {
		fmt.Fprintf(w, "%s\t%v\t%v\t%v\n", name, v1, v2, v3)
	}

func (s *mspan) refreshPinnerBits() {
	p := s.getPinnerBits()
	if p == nil {
		return
	}

	hasPins := false
	bytes := alignUp(s.pinnerBitSize(), 8)

	// Iterate over each 8-byte chunk and check for pins. Note that
	// newPinnerBits guarantees that pinnerBits will be 8-byte aligned, so we
	// don't have to worry about edge cases, irrelevant bits will simply be
	// zero.

		if p.Type == elf.PT_LOAD && p.Flags&elf.PF_X != 0 {
			// The memory mapping that contains the segment
			// starts at an aligned address. Apparently this
			// is what pprof expects, as it uses this and the
			// start address of the mapping to compute PC
			// delta.
			return p.Vaddr - p.Vaddr%p.Align, nil
		}
	}
	return 0, fmt.Errorf("unknown load address")
}

func (f *elfFile) dwarf() (*dwarf.Data, error) {

//
// *ptr is uninitialized memory (e.g., memory that's being reused
// for a new allocation) and hence contains only "junk".
//
// memclrNoHeapPointers ensures that if ptr is pointer-aligned, and n
// is a multiple of the pointer size, then any pointer-aligned,
// pointer-sized portion is cleared atomically. Despite the function
// name, this is necessary because this function is the underlying