CmdGet.Run = runGet // break init loop
	CmdGet.Flag.Var(&getD, "d", "")
	CmdGet.Flag.Var(&getU, "u", "")
}

func runGet(ctx context.Context, cmd *base.Command, args []string) {
	switch getU.version {
	case "", "upgrade", "patch":
		// ok
	default:
		base.Fatalf("go: unknown upgrade flag -u=%s", getU.rawVersion)
	}
	if getD.set {
		if !getD.value {
			base.Fatalf("go: -d flag may not be set to false")
		}

		//	38	eattr[4]
		//	42	off[4]
		//	46	name[namelen]
		//	46+namelen+xlen+fclen - next header
		//
		if get4(buf) != zcheader {
			break
		}
		meth := get2(buf[10:])
		size := get4(buf[24:])
		namelen := get2(buf[28:])
		xlen := get2(buf[30:])
		fclen := get2(buf[32:])
		off := get4(buf[42:])
		zname := buf[46 : 46+namelen]
		buf = buf[46+namelen+xlen+fclen:]
		if string(zname) != name {
			continue

  );
}

def TF_GlobalIterIdOp : TF_Op<"GlobalIterId", [DeclareOpInterfaceMethods<TF_GetResourceInstanceInterface>, TF_GlobalIterIdEffect]> {
  let summary = [{
Op that gets the global step id.
  }];

  let description = [{
This op gets the step id for each loop iteration.
  }];

  let arguments = (ins);

  let results = (outs
    TF_Int64Tensor:$iter_id
  );
}

// See go.dev/issue/67401.
//
//go:linkname atomicwb
//go:nosplit
func atomicwb(ptr *unsafe.Pointer, new unsafe.Pointer) {
	slot := (*uintptr)(unsafe.Pointer(ptr))
	buf := getg().m.p.ptr().wbBuf.get2()
	buf[0] = *slot
	buf[1] = uintptr(new)
}

// atomicstorep performs *ptr = new atomically and invokes a write barrier.
//
//go:nosplit
func atomicstorep(ptr unsafe.Pointer, new unsafe.Pointer) {

func (b *wbBuf) empty() bool {
	return b.next == uintptr(unsafe.Pointer(&b.buf[0]))
}

// getX returns space in the write barrier buffer to store X pointers.
// getX will flush the buffer if necessary. Callers should use this as:
//
//	buf := &getg().m.p.ptr().wbBuf
//	p := buf.get2()
//	p[0], p[1] = old, new
//	... actual memory write ...
//

}

func notetsleep(n *note, ns int64) bool {
	gp := getg()
	if gp != gp.m.g0 && gp.m.preemptoff != "" {
		throw("notetsleep not on g0")
	}

	return notetsleep_internal(n, ns)
}

// same as runtime·notetsleep, but called on user g (not g0)
// calls only nosplit functions between entersyscallblock/exitsyscall.
func notetsleepg(n *note, ns int64) bool {
	gp := getg()
	if gp == gp.m.g0 {
		throw("notetsleepg on g0")

				break
			}
			dstx := (*uintptr)(unsafe.Pointer(addr))
			p := buf.get1()
			p[0] = *dstx
		}
	} else {
		for {
			var addr uintptr
			if tp, addr = tp.next(dst + size); addr == 0 {
				break
			}
			dstx := (*uintptr)(unsafe.Pointer(addr))
			srcx := (*uintptr)(unsafe.Pointer(src + (addr - dst)))
			p := buf.get2()
			p[0] = *dstx
			p[1] = *srcx
		}
	}
}

// RUN: tac-opt-all-backends -tfl-get-alternative-subgraph='device-specs=GPU' %s -split-input-file -verify-diagnostics | FileCheck %s

module {
  func.func @simpleTest(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = func.call @func_0_GPU_FLOAT(%arg0, %arg1, %arg2) {tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

jaystone776 <******@****.***> 1711838806 +0800

			"(*mspan).base",
			"(*mspan).markBitsForBase",
			"(*mspan).markBitsForIndex",
			"(*mspan).writeUserArenaHeapBits",
			"(*muintptr).set",
			"(*puintptr).set",
			"(*wbBuf).get1",
			"(*wbBuf).get2",

			// Trace-related ones.
			"traceLocker.ok",
			"traceEnabled",
		},
		"runtime/internal/sys": {},
		"runtime/internal/math": {
			"MulUintptr",
		},
		"bytes": {