// If there are less than size records, copyFn is invoked for each record, and
// ok returns true.
func threadCreateProfileInternal(size int, copyFn func(profilerecord.StackRecord)) (n int, ok bool) {
	first := (*m)(atomic.Loadp(unsafe.Pointer(&allm)))
	for mp := first; mp != nil; mp = mp.alllink {
		n++
	}
	if n <= size {
		ok = true
		for mp := first; mp != nil; mp = mp.alllink {

				changed := false
				state.changedSlots.clear()

				// Update locs/registers with the effects of each Value.
				for _, v := range b.Values {
					slots := state.valueNames[v.ID]

					// Loads and stores inherit the names of their sources.
					var source *Value
					switch v.Op {
					case OpStoreReg:
						source = v.Args[0]
					case OpLoadReg:
						switch a := v.Args[0]; a.Op {

// changes would be needed to make this work for big
// endian; however additional changes beyond what I
// have noted are most likely needed to make it work.
// - The string used with VPERM to swap the byte order
//   for loads and stores.
// - The constants that are loaded from CPOOL.
//

// The following constants are defined in an order
// that is correct for use with LXVD2X/STXVD2X
// on little endian.

}

// atomicScavChunkData is an atomic wrapper around a scavChunkData
// that stores it in its packed form.
type atomicScavChunkData struct {
	value atomic.Uint64
}

// load loads and unpacks a scavChunkData.
func (sc *atomicScavChunkData) load() scavChunkData {
	return unpackScavChunkData(sc.value.Load())
}

// store packs and writes a new scavChunkData. store must be serialized

    TF_Tensor:$output
  );
}

def TF_TPUCompileMlirAndExecuteOp : TF_Op<"TPUCompileMlirAndExecute", [AttrSizedOperandSegments]> {
  let summary = "Op that compiles a computation in MLIR into a TPU program, and loads and executes it on a TPU device.";

  let description = [{
For the internal use of the TPU compiler.

	unlock(&h.lock)

	// N.B. The arenas L1 map is quite small on all platforms, so it's fine to
	// just iterate over the whole thing.
	for i := range h.arenas {
		l2 := (*[1 << arenaL2Bits]*heapArena)(atomic.Loadp(unsafe.Pointer(&h.arenas[i])))
		if l2 == nil {
			continue
		}
		sysHugePage(unsafe.Pointer(l2), unsafe.Sizeof(*l2))
	}
}

// base address for all 0-byte allocations
var zerobase uintptr

// to the new table.

// Picking loadFactor: too large and we have lots of overflow
// buckets, too small and we waste a lot of space. I wrote
// a simple program to check some stats for different loads:
// (64-bit, 8 byte keys and elems)
//  loadFactor    %overflow  bytes/entry     hitprobe    missprobe
//        4.00         2.13        20.77         3.00         4.00

	var sl notInHeapSlice
	sl = notInHeapSlice{(*notInHeap)(unsafe.Pointer(spanBase + spanSize - bitmapSize)), elems, elems}
	return *(*[]uintptr)(unsafe.Pointer(&sl))
}

// heapBitsSmallForAddr loads the heap bits for the object stored at addr from span.heapBits.
//
// addr must be the base pointer of an object in the span. heapBitsInSpan(span.elemsize)
// must be true.
//
//go:nosplit

	// into a deterministic-but-arbitrary order.
	sort.Slice(upgrades, func(i, j int) bool {
		return upgrades[i].path < upgrades[j].path
	})
	return upgrades
}

// loadPackages loads the packages matching the given patterns, invoking the
// findPackage function for each package that may require a change to the
// build list.
//

Following the configuration phase, Gradle writes a snapshot of the task graph to a new configuration cache entry, for later Gradle invocations.
Gradle then loads the task graph from the configuration cache, so that it can apply optimizations to the tasks, and then runs the execution phase as normal.
Configuration time will still be spent the first time you run a particular set of tasks.