diff --git a/test/test_fusion_op.py b/test/test_fusion_op.py
index 372fc46056..8c720f079c 100644
--- a/test/test_fusion_op.py
+++ b/test/test_fusion_op.py
@@ -27,7 +27,7 @@ class TestFusionOp(unittest.TestCase):
   def test_recursive_add(self):
     st = time.perf_counter()
     a = Tensor([1,2,3,4])
-    for _ in range(12): a = a + a
+    for _ in range(20): a = a + a
     sched = create_schedule([a.lazydata], None)
     ji = lower_schedule_item(sched[-1])
     et = time.perf_counter()
diff --git a/tinygrad/codegen/linearizer.py b/tinygrad/codegen/linearizer.py
index f30a7d45f2..03d1bcaa1e 100644
--- a/tinygrad/codegen/linearizer.py
+++ b/tinygrad/codegen/linearizer.py
@@ -516,7 +516,9 @@ class Linearizer(Kernel):
     if cachable: self.saved_exprs[key] = ret
     return ret
 
-  def ast_parse(self, x:LazyOp, acc: List[UOp], offs:Optional[List[int]], loaded_buffers:Dict[Union[MemBuffer, ConstBuffer, LocalBuffer], List[UOp]], do_reduce=False, loop_ctx=tuple()) -> List[UOp]:  # noqa: E501
+  def ast_parse(self, x:LazyOp, acc: List[UOp], offs:Optional[List[int]], loaded_buffers:Dict[Union[MemBuffer, ConstBuffer, LocalBuffer], List[UOp]], do_reduce=False, loop_ctx=tuple(), cache=None) -> List[UOp]:  # noqa: E501
+    if cache is None: cache = {}
+    if x in cache: return cache[x]
     if x.op in BufferOps: return loaded_buffers[x.arg]
     if x.op == UnaryOps.CAST: return [self.uop(UOps.CAST, x.arg[0], (u,), x.arg) if not isinstance(x.arg[0], ImageDType) else u for u in self.ast_parse(x.src[0], acc, offs, loaded_buffers)]  # noqa: E501
     if x.op in ReduceOps and not do_reduce:
@@ -527,7 +529,7 @@ class Linearizer(Kernel):
       x = LazyOp(TernaryOps.MULACC, x.src[0].src, x.arg)
     if x.op == ReduceOps.SUM and x.src[0].__class__ is LazyOp and x.src[0].op == UnaryOps.CAST and x.src[0].src[0].__class__ is LazyOp and x.src[0].src[0].op == BinaryOps.MUL:  # noqa: E501
       x = LazyOp(TernaryOps.MULACC, x.src[0].src[0].src, x.arg)
-    values = [self.ast_parse(v, acc, offs, loaded_buffers, loop_ctx=loop_ctx) for v in x.src]
+    values = [self.ast_parse(v, acc, offs, loaded_buffers, loop_ctx=loop_ctx, cache=cache) for v in x.src]
     ops = {ReduceOps.SUM:BinaryOps.ADD, ReduceOps.MAX:BinaryOps.MAX, TernaryOps.MULACC:TernaryOps.MULACC}
     if x.op in ops:
       ret: List[UOp] = []
@@ -540,4 +542,5 @@ class Linearizer(Kernel):
           acc[off] = self.uop(UOps.PHI, input_acc[off].dtype, (input_acc[off], acc[off]) + tuple(loop_ctx))
     else:
       ret = [self.uop(UOps.ALU, dtype=dtypes.bool if x.op == BinaryOps.CMPLT else None, vin=val, arg=x.op) for val in zip(*values)]
+    cache[x] = ret
     return ret
diff --git a/tinygrad/lazy.py b/tinygrad/lazy.py
index 345ef2ce08..c23da40664 100644
--- a/tinygrad/lazy.py
+++ b/tinygrad/lazy.py
@@ -36,20 +36,20 @@ class LazyBuffer:
     self.device, self.st, self.dtype, self.shape = device, st, dtype, st.shape
     if base is None:
       # properties on base
-      self.children: WeakSet[LazyBuffer] = WeakSet()
-      for x in srcs: x.base.children.add(self.base)
       self.op, self.arg, self.srcs = op, arg, srcs  # this is a LazyOp, except the src is LazyBuffers and not LazyOps
       self.realized: Optional[Buffer] = None
       self.output_buffer: Optional[Buffer] = None
       self.forced_realize = False
       self.contiguous_child: Optional[Tuple[ReferenceType[LazyBuffer], ShapeTracker]] = None
+      self.children: WeakSet[LazyBuffer] = WeakSet()
+      for x in srcs: x.base.children.add(self.base)
     else:
       # properties on view
       assert base.base == base, "base must be a base itself"
       self._base = base
 
   def __repr__(self) -> str:
-    return f"<LB {self.device} {self.shape} contig:{self.st.contiguous} {self.st if self.base != self else (self.op, self.realized)}>"
+    return f"<LB {self.device} {self.shape} contig:{self.st.contiguous} {self.st if hasattr(self, '_base') else (self.op, self.realized)}>"
 
   @property
   def base(self) -> LazyBuffer: return self._base if hasattr(self, '_base') else self
@@ -141,7 +141,10 @@ class LazyBuffer:
 # *** schedule creation ***
 
 # recursively create a lazyop
-def _recursive_lazyop(buf:LazyBuffer, inputs:List[LazyBuffer], var_vals:Dict[Variable, int], st:ShapeTracker, realizes:Set[LazyBuffer], first=True):
+def _recursive_lazyop(buf:LazyBuffer, inputs:List[LazyBuffer], var_vals:Dict[Variable, int], st:ShapeTracker,
+                      realizes:Set[LazyBuffer], first=True, cache=None) -> LazyOp:
+  if cache is None: cache = {}
+  if (buf, st) in cache: return cache[(buf, st)]
   if buf != buf.base:
     var_vals.update(merge_dicts([var_vals, buf.st.var_vals]))
     st = buf.st.unbind()+st
@@ -161,7 +164,7 @@ def _recursive_lazyop(buf:LazyBuffer, inputs:List[LazyBuffer], var_vals:Dict[Var
   # if a CONTIGUOUS made it all the way here, just skip it
   if buf.op == LoadOps.CONTIGUOUS:
     assert first
-    return _recursive_lazyop(buf.srcs[0], inputs, var_vals, st, realizes, False)
+    return _recursive_lazyop(buf.srcs[0], inputs, var_vals, st, realizes, False, cache)
 
   # if it's a reduce, we have to change the shapetracker
   if buf.op in ReduceOps:
@@ -169,7 +172,8 @@ def _recursive_lazyop(buf:LazyBuffer, inputs:List[LazyBuffer], var_vals:Dict[Var
     st = ShapeTracker.from_shape(buf.srcs[0].shape).unbind()
 
   # otherwise we fuse it like normal
-  return LazyOp(buf.op, tuple(_recursive_lazyop(x, inputs, var_vals, st, realizes, False) for x in buf.srcs), buf.arg)
+  cache[(buf, st)] = ret = LazyOp(buf.op, tuple(_recursive_lazyop(x, inputs, var_vals, st, realizes, False, cache) for x in buf.srcs), buf.arg)
+  return ret
 
 # recursively walk back in the graph to create the schedule
 def _recursive_schedule(out:LazyBuffer, seen:Set[LazyBuffer], realizes:Set[LazyBuffer],