mtop := int((logM - 2) / _W) // -2 because the top word of N bits is the (N-1)/W'th word.
	mmask := ^Word(0)
	if mbits := (logM - 1) & (_W - 1); mbits != 0 {
		mmask = (1 << mbits) - 1
	}
	if i > mtop {
		i = mtop
	}
	advance := false
	z = z.setWord(1)
	for ; i >= 0; i-- {
		yi := y[i]
		if i == mtop {
			yi &= mmask
		}
		for j := 0; j < _W; j += n {
			if advance {

//go:linkname abigen_sync_atomic_OrInt64 sync/atomic.OrInt64
func abigen_sync_atomic_OrInt64(addr *int64, mask int64) (old int64)

//go:linkname abigen_sync_atomic_OrUint64 sync/atomic.OrUint64
func abigen_sync_atomic_OrUint64(addr *uint64, mask uint64) (old uint64)

//go:linkname abigen_sync_atomic_OrUintptr sync/atomic.OrUintptr
func abigen_sync_atomic_OrUintptr(addr *uintptr, mask uintptr) (old uintptr)

the SIGPROF signal in particular.

The non-Go code should not change the signal mask on any threads
created by the Go runtime. If the non-Go code starts new threads
itself, those threads may set the signal mask as they please.

If the non-Go code starts a new thread, changes the signal mask, and
then invokes a Go function in that thread, the Go runtime will

	separator := byte(0x80) // gets reset to 0x00 once used
	for i := byte(0); i < chunk; i++ {
		mask := byte(int8(i-nx) >> 7) // 0x00 after the end of data

		// if we reached the end of the data, replace with 0x80 or 0x00
		d.x[i] = (^mask & separator) | (mask & d.x[i])

		// zero the separator once used
		separator &= mask

		if i >= 56 {
			// we might have to write the length here if all fit in one block

// The filename is generated by taking pattern and adding a random string to the end.
// If pattern includes a "*", the random string replaces the last "*".
// The file is created with mode 0o600 (before umask).
// If dir is the empty string, CreateTemp uses the default directory for temporary files, as returned by [TempDir].
// Multiple programs or goroutines calling CreateTemp simultaneously will not choose the same file.

	ADD	$16, OUTENC, OUTENC
	ADD	$-16, OUTDEC, OUTDEC
	VSPLTISB	$8, KEY                  // vspltisb 3,8
	MOVD	CNT, CTR                         // mtctr 7
	VSUBUBM	MASK, KEY, MASK                  // vsububm 5,5,3

loop192:
	VPERM	IN1, IN1, MASK, KEY // vperm 3,2,2,5
	VSLDOI	$12, ZERO, IN0, TMP // vsldoi 6,0,1,12
	VCIPHERLAST	KEY, RCON, KEY      // vcipherlast 3,3,4

	VXOR	IN0, TMP, IN0       // vxor 1,1,6

    ReducedPrecisionSupport mask = ReducedPrecisionSupport::None;
    if (toco_flags.quantize_to_float16()) {
      mask |= ReducedPrecisionSupport::Float16Inference;
    }
    if (toco_flags.allow_bfloat16()) {
      mask |= ReducedPrecisionSupport::Bfloat16Inference;
    }
    if (toco_flags.accumulation_type() == toco::IODataType::FLOAT16) {
      mask |= ReducedPrecisionSupport::Float16Accumulation;

		}
	}

	// The perm argument is meant to be adjusted according to umask,
	// but we don't know what the umask is.
	// Create a dummy file to find out.
	// This avoids build tags and works even on systems like Plan 9
	// where the file mask computation incorporates other information.
	mode := perm
	f, err := os.OpenFile(filepath.Clean(dst)+"-go-tmp-umask", os.O_WRONLY|os.O_CREATE|os.O_EXCL, perm)
	if err == nil {

	// for VPERMXOR
	MOVD $consts<>+0xa0(SB), MASK
	MOVD $16, R20
	// V16
	LXVW4X (CONSTBASE)(R0), VS48
	ADD $80,CONSTBASE

	// Load key into V17,V18
	LXVW4X (KEY)(R0), VS49
	LXVW4X (KEY)(R8), VS50

	// Load CNT, NONCE into V19
	LXVW4X (CNT)(R0), VS51

	// Clear V27
	VXOR V27, V27, V27

	// V28
	LXVW4X (CONSTBASE)(R11), VS60

	// Load mask constants for VPERMXOR
	LXVW4X (MASK)(R0), V20

}

func bitSetTest(x int) bool {
	// amd64:"ANDL\t[$]9, AX"
	// amd64:"CMPQ\tAX, [$]9"
	return x&9 == 9
}

// mask contiguous one bits
func cont1Mask64U(x uint64) uint64 {
	// s390x:"RISBGZ\t[$]16, [$]47, [$]0,"
	return x & 0x0000ffffffff0000
}

// mask contiguous zero bits
func cont0Mask64U(x uint64) uint64 {
	// s390x:"RISBGZ\t[$]48, [$]15, [$]0,"
	return x & 0xffff00000000ffff
}