if err != nil {
			return nil, err
		}
		if allocateRemainingFrom(preferred.Intersection(aligned)) {
			return allocated, nil
		}
		// Then fallback to allocate from the aligned set if no preferred list
		// is returned (or not enough devices are returned in that list).
		if allocateRemainingFrom(aligned) {
			return allocated, nil
		}

	off  atomic.Uintptr
}

const traceRegionAllocBlockData = 64<<10 - unsafe.Sizeof(traceRegionAllocBlockHeader{})

// alloc allocates n-byte block. The block is always aligned to 8 bytes, regardless of platform.
func (a *traceRegionAlloc) alloc(n uintptr) *notInHeap {
	n = alignUp(n, 8)
	if n > traceRegionAllocBlockData {
		throw("traceRegion: alloc too large")
	}
	if a.dropping.Load() {
		throw("traceRegion: alloc with concurrent drop")
	}

		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)

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

//
// *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

		// caller reserved spill space.
		p.dstSpill = alignUp(p.dstSpill, uintptr(t.Align_))
		p.dstSpill += t.Size_
	} else {
		// Register assignment failed.
		// Undo the work and stack assign.
		p.parts = oldParts

		// The Go ABI aligns arguments.
		p.dstStackSize = alignUp(p.dstStackSize, uintptr(t.Align_))

		// Copy just the size of the argument. Note that this

	} else if minimum > maxPagesPerPhysPage {
		print("runtime: min = ", minimum, "\n")
		throw("min too large")
	}
	// max may not be min-aligned, so we might accidentally truncate to
	// a max value which causes us to return a non-min-aligned value.
	// To prevent this, align max up to a multiple of min (which is always
	// a power of 2). This also prevents max from ever being less than

// atomic functions (types [Int64] and [Uint64] are automatically aligned).
// The first word in an allocated struct, array, or slice; in a global
// variable; or in a local variable (because the subject of all atomic operations
// will escape to the heap) can be relied upon to be 64-bit aligned.

// SwapInt32 atomically stores new into *addr and returns the previous *addr value.

//   ((unknown): f32[1,2], (unknown): s64[3,4], (unknown): f32[1], (unknown): f32[1], (unknown): s32[5]) -> (u32[5,6], f32[1], s32[5])
//
// Memory stats:
//   arg bytes total:    392
//   arg bytes aligned:  576
//   temp bytes total:   171
//   temp bytes aligned: 512
class MyClass final : public tensorflow::XlaCompiledCpuFunction {
 public:
  // Number of input arguments for the compiled computation.
  static constexpr size_t kNumArgs = 5;

	// On some platforms we need to provide physical page aligned stack
	// allocations. Where the page size is less than the physical page
	// size, we already manage to do this by default.
	needPhysPageAlign := physPageAlignedStacks && typ == spanAllocStack && pageSize < physPageSize

	// If the allocation is small enough, try the page cache!
	// The page cache does not support aligned allocations, so we cannot use