// to the buffer space in R25.
// It clobbers R23 (the linker temp register).
// The act of CALLing gcWriteBarrier will clobber R31 (LR).
// It does not clobber any other general-purpose registers,
// but may clobber others (e.g., floating point registers).
TEXT gcWriteBarrier<>(SB),NOSPLIT,$104
	// Save the registers clobbered by the fast path.
	MOVW	R1, 100(R29)
	MOVW	R2, 104(R29)
retry:

// to the buffer space in R25.
// It clobbers condition codes.
// It does not clobber any general-purpose registers except R27,
// but may clobber others (e.g., floating point registers)
// The act of CALLing gcWriteBarrier will clobber R30 (LR).
TEXT gcWriteBarrier<>(SB),NOSPLIT,$200
	// Save the registers clobbered by the fast path.
	STP	(R0, R1), 184(RSP)
retry:
	MOVD	g_m(g), R0

// madd/msub
(ADD a l:(MUL  x y)) && l.Uses==1 && clobber(l) => (MADD a x y)
(SUB a l:(MUL  x y)) && l.Uses==1 && clobber(l) => (MSUB a x y)
(ADD a l:(MNEG x y)) && l.Uses==1 && clobber(l) => (MSUB a x y)
(SUB a l:(MNEG x y)) && l.Uses==1 && clobber(l) => (MADD a x y)

(ADD a l:(MULW  x y)) && v.Type.Size() <= 4 && l.Uses==1 && clobber(l) => (MADDW a x y)

// gcWriteBarrier does NOT follow the Go ABI. It accepts the
// number of bytes of buffer needed in R11, and returns a pointer
// to the buffer space in R11.
// It clobbers FLAGS. It does not clobber any general-purpose registers,
// but may clobber others (e.g., SSE registers).
// Typical use would be, when doing *(CX+88) = AX
//     CMPL    $0, runtime.writeBarrier(SB)
//     JEQ     dowrite

		return
	}

	// We only care about global data: NAME_EXTERN means a global
	// symbol in the Go sense, and p.Sym.Local is true for a few
	// internally defined symbols.
	// Rewrites must not clobber flags and therefore cannot use the
	// ADD instruction.
	if p.From.Type == obj.TYPE_ADDR && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local() {
		// MOVD $sym, Rx becomes MOVD sym@GOT, Rx

(MOVBUreg <t> x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVBUload <t> [off] {sym} ptr mem)
(MOVHreg <t> x:(MOVHUload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHload  <t> [off] {sym} ptr mem)
(MOVHUreg <t> x:(MOVHload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVHUload <t> [off] {sym} ptr mem)

(MFVSRD (FMOVDconst [c])) => (MOVDconst [int64(math.Float64bits(c))])

(MTVSRD x:(MOVDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (FMOVDload [off] {sym} ptr mem)
(MFVSRD x:(FMOVDload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) => @x.Block (MOVDload [off] {sym} ptr mem)

// Rules for MOV* or FMOV* ops determine when indexed (MOV*loadidx or MOV*storeidx)

// number of bytes of buffer needed in R29, and returns a pointer
// to the buffer space in R29.
// It clobbers condition codes.
// It does not clobber R0 through R17 (except special registers),
// but may clobber any other register, *including* R31.
TEXT gcWriteBarrier<>(SB),NOSPLIT,$120
	// The standard prologue clobbers R31.
	// We use R18, R19, and R31 as scratch registers.
retry:
	MOVD	g_m(g), R18
	MOVD	m_p(R18), R18

  SavedModelV2Bundle bundle;
  TF_RETURN_IF_ERROR(SavedModelV2Bundle::Load(directory, &bundle));

  // TODO(bmzhao): Mangle loaded function names so that different
  // models loaded in the same runtime Context don't clobber eachother.
  // This occurs in python here:
  // https://github.com/tensorflow/tensorflow/blob/285b5fa15405c5e2c084080f52a1818be8648079/tensorflow/python/saved_model/function_deserialization.py#L438-L454

// file deals only with the buffer.
//
// The write barrier has a fast path and a slow path. The fast path
// simply enqueues to a per-P write barrier buffer. It's written in
// assembly and doesn't clobber any general purpose registers, so it
// doesn't have the usual overheads of a Go call.
//
// When the buffer fills up, the write barrier invokes the slow path
// (wbBufFlush) to flush the buffer to the GC work queues. In this