diff --git a/tinygrad/codegen/__init__.py b/tinygrad/codegen/__init__.py
index fb7876fe56..056680bf70 100644
--- a/tinygrad/codegen/__init__.py
+++ b/tinygrad/codegen/__init__.py
@@ -56,19 +56,20 @@ def _get_rewrites_for_renderer(opts:Renderer, linearizer:bool, _QUANTIZE, _DEVEC
   # view pushing
   ret.extend(rewrites_for_views)
 
-  # this is kernel.py
-  if not _RANGEIFY: ret.append(RewriteStep(pm_get_optimization, ctx=lambda _: opts, name="get optimization"))
-  if not _POSTOPT and not _RANGEIFY: ret.append(RewriteStep(pm_do_optimize, ctx=lambda _: opts, name="optimize ast"))
+  if not _POSTOPT and not _RANGEIFY:
+    # this is kernel.py
+    ret.append(RewriteStep(pm_get_optimization, ctx=lambda _: opts, name="get optimization"))
+    ret.append(RewriteStep(pm_do_optimize, ctx=lambda _: opts, name="optimize ast"))
 
   if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
   ret.append(RewriteStep(pm_lowerer, get_index, name="lowerer", bottom_up=True))
 
+  # symbolic before post opt
+  ret.append(RewriteStep(sym+migrate_indexing, name="initial symbolic"))
+
   if _POSTOPT or _RANGEIFY: ret.append(RewriteStep(pm_postrange_opt, ctx=lambda _: opts, name="post optimize ast"))
 
   # ** expander (expand_rewrite) **
-  ret.append(RewriteStep(sym+migrate_indexing, name="initial symbolic"))
-
-  # expand
   ret.append(RewriteStep(sym+pm_pre_expander+expander, name="expander"))
 
   # add locals
diff --git a/tinygrad/codegen/gpudims.py b/tinygrad/codegen/gpudims.py
index 9e06847560..2bd962632f 100644
--- a/tinygrad/codegen/gpudims.py
+++ b/tinygrad/codegen/gpudims.py
@@ -59,8 +59,8 @@ def add_gpudims(ctx:Renderer, s:UOp):
   all_ranges = {x.arg[0]%1000:x for x in s_topo if x.op is Ops.RANGE}
 
   # extract global/local dims
-  global_dims = sorted(dedup([x.arg[0]%1000 for x in all_ranges.values() if x.arg[1] is AxisType.GLOBAL]))
-  local_dims = sorted(dedup([x.arg[0]%1000 for x in all_ranges.values() if x.arg[1] in (AxisType.LOCAL, AxisType.GROUP_REDUCE)]))
+  global_dims = sorted(dedup([x.arg[0]%1000 for x in all_ranges.values() if x.arg[-1] is AxisType.GLOBAL]))
+  local_dims = sorted(dedup([x.arg[0]%1000 for x in all_ranges.values() if x.arg[-1] in (AxisType.LOCAL, AxisType.GROUP_REDUCE)]))
   if not global_dims and not local_dims: return None
 
   # get global and local shape
diff --git a/tinygrad/codegen/late/expander.py b/tinygrad/codegen/late/expander.py
index e5944c8c29..0fa6182a5e 100644
--- a/tinygrad/codegen/late/expander.py
+++ b/tinygrad/codegen/late/expander.py
@@ -134,11 +134,11 @@ def fix_store_unroll(x:UOp):
   return UOp(Ops.CONTRACT, dtypes.void, (x.replace(src=x.src[:2]+tuple(store_range)),), tuple(flatten(x.arg for x in store_expand)), tag=1)
 
 def fix_group_for_reduce(x:UOp):
-  reduce_gfr, reduce_r = partition(x.src[1:], lambda u: u.op is Ops.RANGE and u.arg[1] == AxisType.GROUP_REDUCE)
+  reduce_gfr, reduce_r = partition(x.src[1:], lambda u: u.op is Ops.RANGE and u.arg[-1] == AxisType.GROUP_REDUCE)
   if len(reduce_gfr) == 0: return None
 
   # NOTE: if there's other locals here, we need them in the buffer too
-  upstream_locals = [u for u in x.toposort() if u.op is Ops.RANGE and u.arg[1] == AxisType.LOCAL]
+  upstream_locals = [u for u in x.toposort() if u.op is Ops.RANGE and u.arg[-1] == AxisType.LOCAL]
 
   # do only the non grouped reduces early
   ret = x.replace(src=(x.src[0],)+tuple(reduce_r))
@@ -153,7 +153,7 @@ pm_pre_expander = PatternMatcher([
   # rewrite UPCAST/UNROLL range to something to be expanded
   (UPat(Ops.RANGE, name="r"),
    lambda r: UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(s:=r.vmax+1), tuple(range(s))),), ((r.arg[0],s),)) \
-    if r.arg[1] in {AxisType.UNROLL, AxisType.UPCAST} else None),
+    if r.arg[-1] in {AxisType.UNROLL, AxisType.UPCAST} else None),
   # fix REDUCEs with UNROLLs
   (UPat(Ops.REDUCE, name="x"), fix_reduce_unroll),
   (UPat(Ops.STORE, name="x"), fix_store_unroll),
diff --git a/tinygrad/codegen/opt/postrange.py b/tinygrad/codegen/opt/postrange.py
index 83fc60f213..74f9e676a3 100644
--- a/tinygrad/codegen/opt/postrange.py
+++ b/tinygrad/codegen/opt/postrange.py
@@ -1,7 +1,9 @@
+from typing import cast
 from dataclasses import replace
+from tinygrad.dtype import dtypes, AddrSpace, PtrDType
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo
 from tinygrad.helpers import colored
-from tinygrad.codegen.opt.kernel import axis_colors
+from tinygrad.codegen.opt.kernel import axis_colors, AxisType
 
 def rename_sink(s:UOp):
   if s.arg is not None and s.arg.name != "test": return None
@@ -10,9 +12,35 @@ def rename_sink(s:UOp):
   rngs = sorted([u for u in s.parents if u.op is Ops.RANGE], key=lambda x: x.arg[0:-1])
 
   # add name to kernel
-  name = "k" + colored('_', 'BLACK').join(['']+[colored(x.src[0].render(), axis_colors[x.arg[1]]) for x in rngs])
+  name = "k" + colored('_', 'BLACK').join(['']+[colored(x.src[0].render(), axis_colors[x.arg[-1]]) for x in rngs])
   return s.replace(arg=KernelInfo(name=name) if s.arg is None else replace(s.arg, name=name))
 
+def global_stores_are_global(s:UOp):
+  if cast(PtrDType, s.src[0].dtype).addrspace != AddrSpace.GLOBAL: return None
+  return s.substitute({u:u.replace(arg=u.arg[0:-1]+(AxisType.GLOBAL,)) for u in s.src[2:] if u.op is Ops.RANGE and u.arg[-1] is AxisType.LOOP})
+
+def split_range(r:UOp):
+  if len(r.arg) > 2: return None
+  N = 4
+  rd = r.src[0].divides(N)
+  if rd is None: return None
+  sr = r.replace(src=(rd,), arg=r.arg[0:-1]+(0,r.arg[-1]))
+  er = UOp(Ops.RANGE, dtypes.int, src=(UOp.const(dtypes.int, N),),
+           arg=r.arg[0:-1]+(1,AxisType.UNROLL if r.arg[-1] is AxisType.REDUCE else AxisType.UPCAST))
+  return sr*N+er
+
+def flatten_range_in_terminators(r:UOp):
+  off = 2 if r.op is Ops.STORE else 1
+  rngs = r.src[off:]
+  if not len(rngs): return None
+  new_rngs = [x for x in UOp.sink(*rngs).toposort() if x.op is Ops.RANGE]
+  return r.replace(src=r.src[:off]+tuple(new_rngs))
+
 pm_postrange_opt = PatternMatcher([
+  # flatten ranges
+  (UPat((Ops.REDUCE, Ops.STORE), name="r"), flatten_range_in_terminators),
+
+  (UPat(Ops.RANGE, name="r"), split_range),
+  (UPat(Ops.STORE, name="s"), global_stores_are_global),
   (UPat(Ops.SINK, name="s"), rename_sink),
 ])
diff --git a/tinygrad/viz/serve.py b/tinygrad/viz/serve.py
index 61ee3da416..4659436c86 100755
--- a/tinygrad/viz/serve.py
+++ b/tinygrad/viz/serve.py
@@ -80,7 +80,7 @@ def uop_to_json(x:UOp) -> dict[int, dict]:
       if u.op not in {Ops.VIEW, Ops.BUFFER, Ops.KERNEL, Ops.ASSIGN, Ops.COPY, Ops.SINK, *GroupOp.Buffer} and u.st is not None:
         label += f"\n{shape_to_str(u.shape)}"
       elif len(rngs:=u.ranges):
-        label += f"\n({','.join([colored(str(x.arg[0]), axis_colors[x.arg[1]]) for x in sorted(rngs, key=lambda x: x.arg[0])])})"
+        label += f"\n({','.join([colored(str(x.arg[0]), axis_colors[x.arg[-1]]) for x in sorted(rngs, key=lambda x: x.arg[0:-1])])})"
     except Exception:
       label += "\n<ISSUE GETTING LABEL>"
     if (ref:=ref_map.get(u.arg.ast) if u.op is Ops.KERNEL else None) is not None: label += f"\ncodegen@{ctxs[ref]['name']}"