use buf_target in expand_index

* Revert "Revert "merge to fold_divmod_general [p] (#13359)""

This reverts commit 05ccc69248.

* Revert "Revert "actually merge to fold_divmod_general [pr] (#13363)""

This reverts commit 90e5752199.

* Revert "Revert "add cache to fold_divmod_general (#13365)""

This reverts commit 8e17bd6791.

* bring back fold_divmod_general with bugfix and test

extra/thunder/tiny/tk/tiles.py

@@ -11,9 +11,9 @@ def unwrap(x):
   if isinstance(x, dict): return {k: unwrap(v) for k,v in x.items()}
   return x
 
-def wrap(x, ker, cls):
-  if isinstance(x, UOp): return cls(x, ker)
-  if isinstance(x, (list, tuple)): return type(x)(wrap(y, ker, cls) for y in x)
+def wrap(x, s):
+  if isinstance(x, UOp): return s.ruop(x)
+  if isinstance(x, (list, tuple)): return type(x)(wrap(y, s) for y in x)
   return x
 
 def autowrap(source_cls, blacklist=None):
@@ -31,10 +31,10 @@ def autowrap(source_cls, blacklist=None):
       if callable(val):
         @functools.wraps(val)
         def proxy(*args, **kwargs):
-          return wrap(val(*unwrap(args), **unwrap(kwargs)), self.ker, cls)
+          return wrap(val(*unwrap(args), **unwrap(kwargs)), self)
         return proxy
       if name in UOp.__slots__: return val
-      return wrap(val, self.ker, cls)
+      return wrap(val, self)
     cls.__getattr__ = __getattr__
 
     for name in dir(source_cls):
@@ -46,9 +46,9 @@ def autowrap(source_cls, blacklist=None):
       else:
         original = getattr(source_cls, name)
         if callable(original):
-          def make_proxy(op_name, func):
+          def make_proxy(_, func):
             def proxy(self, *args, **kwargs):
-              return wrap(func(self._uop, *unwrap(args), **unwrap(kwargs)), self.ker, cls)
+              return wrap(func(self._uop, *unwrap(args), **unwrap(kwargs)), self)
             return proxy
           setattr(cls, name, make_proxy(name, original))
 
@@ -69,7 +69,7 @@ class TileMathMixin(MathMixin):
         if isinstance(self, RT) and isinstance(src[0], RV): uop = self.ker.warp.map(self._uop, lambda x, idx: UOp.alu(x, op, inner_op(src[0]._uop[idx[0], 0, (idx[2]%4)//2])))
         else: uop = self.ker.warp.map(self._uop, lambda x, idx: UOp.alu(x, op, inner_op(src[0]._uop[*idx])))
     else: raise NotImplementedError
-    return type(self)(uop, self.ker)
+    return self.ruop(uop)
   def const_like(self, b): return b
 
   # override ops that do compute on the src uop
@@ -83,6 +83,9 @@ class GL:
   def __init__(self, uop, ker):
     self._uop, self.ker = uop, ker
 
+  def ruop(self, uop):
+    return GL(uop, self.ker)
+
   @classmethod
   def create(cls, shape, dtype, ker):
     uop = ker.alloc(shape, dtype, AddrSpace.GLOBAL)
@@ -93,6 +96,9 @@ class ST:
   def __init__(self, uop, ker):
     self._uop, self.ker = uop, ker
 
+  def ruop(self, uop):
+    return ST(uop, self.ker)
+
   @classmethod
   def create(cls, shape, dtype, ker):
     uop = ker.alloc(shape, dtype, AddrSpace.LOCAL)
@@ -107,6 +113,9 @@ class RT(TileMathMixin):
   def __init__(self, uop, ker):
     self._uop, self.ker = uop, ker
 
+  def ruop(self, uop):
+    return RT(uop, self.ker)
+
   @classmethod
   def create(cls, shape, dtype, ker):
     assert len(shape) == 2
@@ -121,8 +130,11 @@ class RT(TileMathMixin):
 
 @autowrap(UOp)
 class RV(TileMathMixin):
-  def __init__(self, uop, ker):
-    self._uop, self.ker = uop, ker
+  def __init__(self, uop, layout, ker):
+    self._uop, self.layout, self.ker = uop, layout, ker
+
+  def ruop(self, uop):
+    return RV(uop, self.layout, self.ker)
 
   @classmethod
   def create(cls, length, dtype, layout, ker):
@@ -138,6 +150,6 @@ class RV(TileMathMixin):
       case _: raise NotImplementedError(f"rv layout {layout} not implemented")
 
     uop = ker.alloc((outer_dim, inner_dim, 2), dtype, AddrSpace.REG)
-    return RV(uop, ker)
+    return RV(uop, layout, ker)
 
 ALL_TILES = UOp | GL | ST | RT | RV

test/external/external_uop_gc.py

@@ -2,6 +2,7 @@ import gc
 from tinygrad import Tensor, UOp, Device, nn
 from tinygrad.engine.realize import method_cache, get_program
 from tinygrad.schedule.indexing import apply_movement_op
+from tinygrad.uop.divandmod import fold_divmod_general
 from test.test_tiny import TestTiny
 
 def uops_allocated(): return sum([isinstance(x, UOp) for x in gc.get_objects()])
@@ -69,6 +70,7 @@ if __name__ == "__main__":
     # these caches will keep uops alive
     method_cache.clear()
     apply_movement_op.cache_clear()
+    fold_divmod_general.cache_clear()
     Tensor._device_seeds.clear()
     Tensor._device_rng_counters.clear()
 

test/external/process_replay/process_replay.py

@@ -38,7 +38,7 @@ def trunc_log(x):
 # user config
 # NOTE: process replay is slow so it's now disabled by default. add [pr] to enable it
 #SKIP_PROCESS_REPLAY = (k:="[skip_process_replay]") in os.getenv("COMMIT_MESSAGE", "") or k in os.getenv("PR_TITLE", "")
-SKIP_PROCESS_REPLAY = not ASSERT_DIFF
+SKIP_PROCESS_REPLAY = not ASSERT_DIFF and not ((k:="[p]") in os.getenv("COMMIT_MESSAGE", "") or k in os.getenv("PR_TITLE", ""))
 if REF == "master": SKIP_PROCESS_REPLAY = True
 class ProcessReplayWarning(Warning): pass
 

test/test_uops.py

@@ -517,7 +517,7 @@ class TestUOpStr(unittest.TestCase):
 
 class TestUPatHelpers(unittest.TestCase):
   def test_location(self):
-    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "math.py")
+    self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "symbolic.py")
     self.assertEqual(shared_spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
     test_upat = UPat(Ops.CONST, dtypes.bool)
     self.assertEqual(test_upat.location[0].split("/")[-1], __file__.replace("\\", "/").split("/")[-1])

test/unit/test_microbenchmarks.py

@@ -1,4 +1,5 @@
 import unittest, time
+from tinygrad.helpers import Profiling
 from tinygrad.uop.ops import UOp
 from tinygrad.dtype import dtypes
 
@@ -38,6 +39,14 @@ class TestMicrobenchmarks(unittest.TestCase):
     a = UOp.const(dtypes.int, 2)
     for _ in range(N): (a+a).simplify()
 
+class TestMicroprofile(unittest.TestCase):
+  def test_uop_simplify_complex(self):
+    x = UOp.variable("x", 0, 10)
+    y = UOp.variable("y", 0, 10)
+    expr = (x*2)+5+(x*4)+(y*2)+y
+    with Profiling():
+      for _ in range(1000): expr.simplify()
+
 if __name__ == '__main__':
   unittest.main()
 

test/unit/test_symbolic_failures.py

@@ -159,3 +159,38 @@ class TestFuzzFailure(unittest.TestCase):
     num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
     rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
     self.assertEqual(num, rn)
+
+  def test_fuzz_failure11(self):
+    v1=Variable("v1", 0, 16)
+    v2=Variable("v2", 0, 128)
+    v3=Variable("v3", 0, 5)
+    expr = UOp(Ops.MOD, dtypes.index, arg=None, src=(
+      UOp(Ops.ADD, dtypes.index, arg=None, src=(
+        UOp(Ops.MOD, dtypes.index, arg=None, src=(
+          UOp(Ops.ADD, dtypes.index, arg=None, src=(
+            UOp(Ops.MAX, dtypes.index, arg=None, src=(
+              UOp(Ops.MUL, dtypes.index, arg=None, src=(
+                x5:=UOp(Ops.DEFINE_VAR, dtypes.index, arg=('v2', 0, 128), src=()),
+                UOp(Ops.CONST, dtypes.index, arg=0, src=()),)),
+              UOp(Ops.CONST, dtypes.index, arg=8, src=()),)),
+            UOp(Ops.MUL, dtypes.index, arg=None, src=(
+              x5,
+              UOp(Ops.CONST, dtypes.index, arg=-2, src=()),)),)),
+          x10:=UOp(Ops.CONST, dtypes.index, arg=5, src=()),)),
+        UOp(Ops.ADD, dtypes.index, arg=None, src=(
+          UOp(Ops.ADD, dtypes.index, arg=None, src=(
+            UOp(Ops.IDIV, dtypes.index, arg=None, src=(
+              x14:=UOp(Ops.DEFINE_VAR, dtypes.index, arg=('v1', 0, 16), src=()),
+              UOp(Ops.CONST, dtypes.index, arg=6, src=()),)),
+            UOp(Ops.CONST, dtypes.index, arg=4, src=()),)),
+          UOp(Ops.ADD, dtypes.index, arg=None, src=(
+            x14,
+            UOp(Ops.CONST, dtypes.index, arg=1, src=()),)),)),)),
+      x10,))
+    v1_val, v2_val, v3_val = UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 7),UOp.const(dtypes.int, 0)
+    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
+    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
+    self.assertEqual(num, rn)
+
+if __name__ == '__main__':
+  unittest.main()

test/unit/test_tqdm.py

@@ -128,6 +128,7 @@ class TestProgressBar(unittest.TestCase):
         self._compare_bars(tinytqdm_output, tqdm_output)
         if n > 5: break
 
+  @unittest.skip("this is flaky")
   @patch('sys.stderr', new_callable=StringIO)
   @patch('shutil.get_terminal_size')
   def test_set_description(self, mock_terminal_size, mock_stderr):

tinygrad/codegen/late/devectorizer.py

@@ -60,7 +60,9 @@ load_store_indexing = PatternMatcher([
 def expand_index(buf:UOp, vec:UOp):
   if getenv("UNSAFE_DISABLE_MASK", 0): vec = vec.get_idx()
   # generate the individual indexes
-  midx = graph_rewrite(UOp.sink(*[buf.index(vec.gep(i), ptr=True) for i in range(vec.dtype.count)]),
+  # we use `.buf_target()` here to avoid traversing into the AFTER
+  buf_target = buf.buf_target().rtag() if buf.op is Ops.AFTER else buf
+  midx = graph_rewrite(UOp.sink(*[buf_target.index(vec.gep(i), ptr=True) for i in range(vec.dtype.count)]),
                        symbolic+load_store_indexing, name=f"index_buf_{buf.arg}")
   # extract all the relevant offsets
   offsets_rootsrc: defaultdict[Any, dict[int, list[int]]] = defaultdict(dict)
@@ -93,7 +95,7 @@ def expand_index(buf:UOp, vec:UOp):
   assert None not in idxs, f"some idxs are missing {idxs}"
   # this base thing is for image, we want the CAT to be a normal pointer
   post_cat = UOp(Ops.PTRCAT, buf.ptrdtype.base.ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace).vec(global_offset), tuple(ret))
-  return post_cat.gep(tuple(cast(list[int], idxs)))
+  return post_cat.gep(tuple(cast(list[int], idxs))).substitute({buf_target:buf})
 
 def cat_after_store(cat:UOp, data:UOp, sto:UOp):
   # TODO: this is written in many places

tinygrad/uop/divandmod.py

@@ -0,0 +1,112 @@
+import functools
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp
+from tinygrad.dtype import dtypes
+from tinygrad.helpers import cdiv, cmod, CORRECT_DIVMOD_FOLDING, unwrap
+
+# NOTE: this cache is only on index UOps and matches the cache in the old ShapeTracker in spirit
+@functools.cache
+def fold_divmod_general(d: UOp, correct_divmod_folding: bool) -> UOp|None:
+  x, y = d.src
+
+  # cancel_divmod: simple cancel div/mod case when the range of the numerator lies within a single denominator interval
+  x_min, x_max, y_min, y_max = x.vmin, x.vmax, y.vmin, y.vmax
+  assert isinstance(x_min, int) and isinstance(x_max, int) and isinstance(y_min, int) and isinstance(y_max, int)
+  if y_min==y_max==0: raise ZeroDivisionError(f"{'Division' if d.op is Ops.IDIV else 'Mod'} by zero trying to rewrite {x.alu(d.op, y)}")
+  if y_min*y_max > 0 and (q:=cdiv(x_min,y_min)) == cdiv(x_min,y_max) == cdiv(x_max,y_min) == cdiv(x_max,y_max):
+    return x - q*y if d.op is Ops.MOD else d.const_like(q)
+
+  # split uops for the rest of the processing
+  x_peeled, const = x.pop_const()
+  uops_no_const = list(x_peeled.split_uop(Ops.ADD))
+
+  # ** Constant Denominator Rules **
+  # these rules strictly require y to be a scalar constant > 0
+  if y.op is Ops.CONST and (c := y.arg) > 0:
+    # remove_nested_mod: remove nested mod in case the inner mod is a multiple of the outer mod, example: (a%4 + b)%2 -> (a+b)%2
+    if d.op is Ops.MOD and x.vmin >= 0:
+      new_xs, changed = [], False
+      for u in uops_no_const:
+        if u.op is Ops.MOD and u.src[1].divides(c) is not None:
+          u = u.src[0]
+          changed = True
+        new_xs.append(u)
+      if changed and (new_x:=(UOp.sum(*new_xs) + const)).vmin >= 0: return new_x % y
+
+    # Shared decomposition for folding rules
+    decomp = [(u.divides(f:=u.const_factor()),f) for u in uops_no_const]
+    terms, factors = zip(*decomp)
+
+    # fold_binary_numerator: fold if expression has one non-constant term that takes on two values
+    if len(terms)==1 and (v:=terms[0]).vmax-v.vmin == 1:
+      y1 = cmod(factors[0]*v.vmin+const, c) if d.op is Ops.MOD else cdiv(factors[0]*v.vmin+const, c)
+      y2 = cmod(factors[0]*v.vmax+const, c) if d.op is Ops.MOD else cdiv(factors[0]*v.vmax+const, c)
+      return (y2-y1)*(v-v.vmin) + y1
+
+    # fold_divmod_congruence: fold if a is congruent to an expression whose range is between 0 and c
+    if not (x.vmin<0 and correct_divmod_folding):
+      rems = [min((r:=f%c), r-c, key=abs) for f in factors]
+      if (rem:=sum(r*v for r,v in zip(rems,terms))+const%c).vmin//c==rem.vmax//c:
+        if d.op is Ops.MOD: return rem - rem.vmin//c*c
+        return sum((f-r)//c * v for f,r,v in zip(factors,rems,terms)) + (const-const%c+rem.vmin//c*c)//c
+
+    # gcd_with_remainder: factor out common gcd from numerator
+    # Note: this rule uses uops_no_const to exclude the additive constant from the GCD calculation
+    if x.vmin >= 0:
+      gcd = UOp.gcd(*uops_no_const, y).simplify()
+      if gcd.op is Ops.CONST and gcd.arg > 1:
+        new_x = unwrap(x_peeled.divide_exact(gcd)).simplify() + (const%c)//gcd.arg
+        if new_x.vmin >= 0:
+          ret = new_x.alu(d.op, x.ufix(c//gcd.arg))
+          return ret*gcd + const%gcd.arg if d.op is Ops.MOD else ret+const//c
+
+    # nest_div_by_smallest_factor: try and nest the div and see if it allows the numerator to be simplified
+    if d.op is Ops.IDIV and x.vmin >= 0:
+      div = min([c] + [abs(f) for u, f in zip(uops_no_const, factors) if u.op not in (Ops.CONST, Ops.VCONST) and abs(f) > 1 and (c%f)==0])
+      # NOTE: this is recursive!
+      if div < c and (newxs := fold_divmod_general(x//div, correct_divmod_folding)) is not None and newxs.vmin >= 0:
+        return newxs // (c // div)
+
+  # ** Variable Denominator / Fallback Rules **
+  # These rules apply to variables OR constants that failed the checks above.
+  # Reconstruct all uops including const for these checks.
+  all_uops = uops_no_const + ([x.const_like(const)] if const != 0 else [])
+
+  # divide_by_gcd: x//y -> (x//gcd)//(y//gcd)
+  gcd = UOp.gcd(*all_uops, y).simplify()
+  if not (gcd.op is Ops.CONST and gcd.arg==1):
+    ret = unwrap(x.divide_exact(gcd)).alu(d.op, unwrap(y.divide_exact(gcd)))
+    return ret*gcd if d.op is Ops.MOD else ret
+
+  # factor_remainder: (d*x+y)//d -> x+y//d
+  if y.vmin<0 or x.vmin<0: return None
+  quo, rem = [], []
+  for u in all_uops:
+    if (q:=u.divide_exact(y)) is not None: quo.append(q)
+    elif d.op is Ops.MOD and y.op is Ops.CONST and (c:=u.const_factor())%y.arg!=c:
+      rem.append(u.divides(c)*(c%y.arg))
+      quo.append(u.const_like(0))
+    else: rem.append(u)
+
+  if not quo: return None
+  new_x = sum(rem)+x.const_like(0)
+  if new_x.vmin<0: return None
+  return new_x%y if d.op is Ops.MOD else new_x//y+sum(quo)
+
+div_and_mod_symbolic = PatternMatcher([
+  # ** 1. Fast Inline Rules **
+  ((UPat.var("x")//UPat.cvar("c") + UPat.cvar("a"))//UPat.cvar("d"), lambda x,c,a,d: (x+a*c)//(c*d)
+    if c.vmin>0 and d.vmin>0 and ((x.vmin>=0 and a.vmin>=0) or (x.vmax<=0 and a.vmax<=0)) else None),  # (x//c+a)//d -> (x+a*c)//(c*d)
+  (UPat.var("x", dtypes.index) // UPat.var("d"), lambda x,d: -(x//(-d)) if d.vmax < 0 else None),
+  (UPat.var("x", dtypes.index) // UPat.var("d"), lambda x,d: -((-x)//d) if x.vmax <= 0 else None),
+  ((UPat.var("x", dtypes.index)+UPat.cvar("c", vec=False)).named("n")//UPat.cvar("d", vec=False),
+    lambda x,c,n,d: ((x+c.arg%d.arg)//d + c.arg//d.arg) if c.arg%d.arg!=c.arg and x.vmin>=0 and n.vmin>=0 and d.arg>0 else None),
+  ((UPat.var("x", dtypes.index)+UPat.cvar("c", vec=False)).named("n")//UPat.cvar("d", vec=False),
+    lambda x,c,n,d: (-(-(c.arg%d.arg + x - (d.arg-1))//d) + c.arg//d.arg) if x.vmax<=0 and n.vmin>=0 and d.arg>0 else None),
+
+  # ** 2. Slow Rules **
+  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d"), lambda d: fold_divmod_general(d, bool(CORRECT_DIVMOD_FOLDING))),
+
+  # NOTE: these have to go at the bottom or TestSymbolicOps.test_var loops
+  (UPat.var("x", dtypes.index) % UPat.var("d"), lambda x,d: -((-x)%d) if x.vmax <= 0 else None),
+  (UPat.var("x", dtypes.index) % UPat.var("d"), lambda x,d: (x%(-d)) if d.vmax < 0 else None),
+])

tinygrad/uop/ops.py

@@ -615,6 +615,7 @@ class UOp(OpMixin, metaclass=UOpMetaClass):
 
   def buf_target(self) -> UOp:
     # the buffer that's being loaded from or store to
+    # NOTE: this is the good one to keep
     match self.op:
       case Ops.DEFINE_GLOBAL | Ops.DEFINE_LOCAL | Ops.DEFINE_REG: return self
       case Ops.AFTER | Ops.INDEX | Ops.STORE | Ops.LOAD: return self.src[0].buf_target()
@@ -866,8 +867,8 @@ def print_uops(uops:list[UOp]):
 
 def get_location() -> tuple[str, int]:
   frm = sys._getframe(1)
-  # skip over ops.py/mathtraits.py (unless there's nothing but ops.py/mathtraits.py)
-  while pathlib.Path(frm.f_code.co_filename).name in ("ops.py", "mathtraits.py") and frm.f_back is not None and \
+  # skip over ops.py and anything in mixin
+  while ((codepath:=pathlib.Path(frm.f_code.co_filename)).name == "ops.py" or codepath.parent.name == "mixin") and frm.f_back is not None and \
       not frm.f_back.f_code.co_filename.startswith("<frozen"):
     frm = frm.f_back
   return frm.f_code.co_filename, frm.f_lineno
@@ -1077,20 +1078,22 @@ def track_rewrites(name:Callable[..., str|TracingKey]|bool=True, replay:bool=Fal
 
 active_rewrites:list[TrackedGraphRewrite] = []
 def profile_matches(fxn:Callable):
-  def wrap(*args, **kwargs):
-    name = str(kwargs.get("name", None) or fxn.__name__)
-    assert args and isinstance(args[0], UOp), f"invalid match tracing inputs for {name} with {args}"
-    if tracking:=(TRACK_MATCH_STATS >= 2):
+  def wrap_profile_matches(*args, **kwargs):
+    if TRACK_MATCH_STATS >= 2:
+      name = str(kwargs.get("name", None) or fxn.__name__)
+      assert args and isinstance(args[0], UOp), f"invalid match tracing inputs for {name} with {args}"
       loc = ((frm:=sys._getframe(1)).f_code.co_filename, frm.f_lineno)
       depth = len(active_rewrites)
       if not tracked_ctxs: add_trace_group(TracingKey(f"default {fxn.__name__}"))
       tracked_ctxs[-1].append(ctx:=TrackedGraphRewrite(loc, args[0].trace_num, [], name, depth, kwargs.get("bottom_up", False)))
       active_rewrites.append(ctx)
-    with cpu_profile(name, "TINY", display=tracking):
-      ret = fxn(*args, **kwargs)
-    if tracking: active_rewrites.pop()
-    return ret
-  return wrap
+      with cpu_profile(name, "TINY"):
+        ret = fxn(*args, **kwargs)
+      active_rewrites.pop()
+      return ret
+    # without tracking, we just call the function
+    return fxn(*args, **kwargs)
+  return wrap_profile_matches
 
 class TrackedPatternMatcher(PatternMatcher):
   def rewrite(self, uop:UOp, ctx=None) -> UOp|None:

tinygrad/uop/symbolic.py

@@ -3,8 +3,9 @@ import math, operator, struct, functools
 from collections import defaultdict
 from tinygrad.uop.ops import Ops, PatternMatcher, UPat, UOp, GroupOp, exec_alu
 from tinygrad.dtype import ConstType, dtypes, PtrDType, can_safe_cast, Invalid
-from tinygrad.helpers import partition, all_same, prod, flatten, get_single_element, cdiv, cmod, CORRECT_DIVMOD_FOLDING, unwrap
+from tinygrad.helpers import partition, all_same, prod, flatten, get_single_element, unwrap
 from tinygrad.uop.decompositions import xpow
+from tinygrad.uop.divandmod import div_and_mod_symbolic
 
 # ******** phase 1 of symbolic used to live in ops, it's the most generic folding rules ********
 
@@ -102,13 +103,11 @@ symbolic_simple = propagate_invalid + PatternMatcher([
   # positive const ** x
   (UPat.cvar("c", vec=False).alu(Ops.POW, UPat.var("x")), lambda c,x: c if c.arg == 1 else (x*math.log2(c.arg)).exp2() if c.arg > 0 else None),
   # rules for threefry
-  ((UPat.var('x', dtypes.uint64)&0xFFFFFFFF).cast(dtypes.uint32), lambda x: x.cast(dtypes.uint32)&0xFFFFFFFF), # TODO: why is the and needed?
+  ((UPat.var('x', dtypes.uint64)&0xFFFFFFFF).cast(dtypes.uint32), lambda x: x.cast(dtypes.uint32)),
   (((UPat.var(None, dtypes.uint64)*(1<<32)) | UPat.var('y',  dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
   (((UPat.var('x',  dtypes.uint64)*(1<<32)) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))//(1<<32), lambda x: x),
-  # new decomp rules for threefry
   (((UPat.var(None, dtypes.uint64)<<32) | UPat.var('y',  dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
   (((UPat.var('x',  dtypes.uint64)<<32) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))>>32, lambda x: x),
-  (UPat.var('b').where(UPat.var('x', dtypes.uint32).cast(dtypes.uint64), UPat.const(dtypes.uint64, 0)).cast(dtypes.uint32), lambda b,x: b.where(x,0)),
   # ** simple where folding **
   # a conditional with the same results either way is a noop, also fold const conditionals
   (UPat.var().where(UPat.var("val"), UPat.var("val")), lambda val: val),
@@ -138,101 +137,6 @@ def canonicalize_simplex(X:UOp) -> UOp|None:
     ret.append(u)
   return UOp.sum(*ret) if changed else None
 
-def cancel_divmod(d: UOp, x: UOp, y: UOp) -> UOp|None:
-  # simple cancel div/mod case when the range of the numerator lies within a single denominator interval
-  x_min, x_max, y_min, y_max = x.vmin, x.vmax, y.vmin, y.vmax
-  assert isinstance(x_min, int) and isinstance(x_max, int) and isinstance(y_min, int) and isinstance(y_max, int)
-  if y_min==y_max==0: raise ZeroDivisionError(f"{'Division' if d.op is Ops.IDIV else 'Mod'} by zero trying to rewrite {x.alu(d.op, y)}")
-  if y_min*y_max > 0 and (q:=cdiv(x_min,y_min)) == cdiv(x_min,y_max) == cdiv(x_max,y_min) == cdiv(x_max,y_max):
-    return x - q*y if d.op is Ops.MOD else d.const_like(q)
-  return None
-
-def remove_nested_mod(m: UOp, x: UOp, y: UOp) -> UOp|None:
-  # remove nested mod in case the inner mod is a multiple of the outer mod
-  # example: (a%4 + b)%2 -> (a+b)%2
-  if ((c := y.arg) < 0) or x.vmin<0: return None
-  new_xs = []
-  something_changed = False
-  for u in x.split_uop(Ops.ADD):
-    if u.op is Ops.MOD:
-      if u.src[1].divides(c) is not None:
-        something_changed = True
-        u = u.src[0]
-    new_xs.append(u)
-  new_x: UOp = UOp.sum(*new_xs)
-  if something_changed and new_x.vmin>=0: return new_x % y
-  return None
-
-def fold_binary_numerator(d: UOp, x: UOp, y: UOp) -> UOp|None:
-  # we can fold if the expression has only one non-constant term and this term can only take on two values
-  if ((c := y.arg) < 0): return None
-  x,const = x.pop_const()
-  terms, factors = zip(*[(u.divides(f:=u.const_factor()),f) for u in x.split_uop(Ops.ADD)])
-  if len(terms)==1 and (v:=terms[0]).vmax-v.vmin == 1:
-    y1 = cmod(factors[0]*v.vmin+const, c) if d.op is Ops.MOD else cdiv(factors[0]*v.vmin+const, c)
-    y2 = cmod(factors[0]*v.vmax+const, c) if d.op is Ops.MOD else cdiv(factors[0]*v.vmax+const, c)
-    return (y2-y1)*(v-v.vmin) + y1
-  return None
-
-def fold_divmod_congruence(d: UOp, x: UOp, y: UOp) -> UOp|None:
-  # within a mod we can freely subtract multiples of c, we use this to see if a is congruent to an expression whose vmin/vmax are between 0 and c
-  if (x.vmin<0 and CORRECT_DIVMOD_FOLDING) or ((c := y.arg) < 0): return None
-  x,const = x.pop_const()
-  terms, factors = zip(*[(u.divides(f:=u.const_factor()),f) for u in x.split_uop(Ops.ADD)])
-  # a//c = (a-a%c)/c, if we can fold a%c, we can fold a//c
-  rems = [min((r:=f%c), r-c, key=abs) for f in factors]
-  if (rem:=sum(r*v for r,v in zip(rems,terms))+const%c).vmin//c!=rem.vmax//c: return None
-  if d.op is Ops.MOD: return rem - rem.vmin//c*c
-  return sum((f-r)//c * v for f,r,v in zip(factors,rems,terms)) + (const-const%c+rem.vmin//c*c)//c
-
-def divide_by_gcd(d: UOp, x: UOp, y: UOp) -> UOp|None:
-  # x//y -> (x//gcd)//(y//gcd) or x%y -> gcd*(x//gcd)%(y//gcd)
-  gcd = UOp.gcd(*x.split_uop(Ops.ADD), y).simplify()
-  if gcd.op is Ops.CONST and gcd.arg==1: return None
-  ret = unwrap(x.divide_exact(gcd)).alu(d.op, unwrap(y.divide_exact(gcd)))
-  return ret*gcd if d.op is Ops.MOD else ret
-
-def gcd_with_remainder(d: UOp, x: UOp, y: UOp):
-  # (gcd*x+r)//(gcd*d) -> (x+(r%d)//gcd)//d + r//(gcd*d)
-  # (gcd*x+r)%(gcd*d) -> gcd*(x+(r%d)//gcd)%d + r%gcd
-  # These only work for floordiv (and the corresponding remainder)! Thats why we check the sign of x,y and new_x
-  if ((c := y.arg) < 0) or x.vmin<0: return None
-  x_no_const, const = x.pop_const()
-  gcd = UOp.gcd(*x_no_const.split_uop(Ops.ADD), y).simplify()
-  assert gcd.op is Ops.CONST
-  if gcd.arg==1: return None
-  new_x = unwrap(x_no_const.divide_exact(gcd)).simplify() + (const%c)//gcd
-  if new_x.vmin<0: return None
-  ret = new_x.alu(d.op, x.ufix(c//gcd.arg))
-  return ret*gcd + const%gcd.arg if d.op is Ops.MOD else ret+const//c
-
-def factor_remainder(d: UOp, x: UOp, y: UOp) -> UOp|None:
-  # (d*x+y)//d -> x+y//d  or  (d*x+y)%d
-  # for mod we go further and take the remainder of all factors to reduce their size
-  # These only work for floordiv (and the corresponding remainder)! Thats why we check the sign of x,y and new_x
-  if y.vmin<0 or x.vmin<0: return None
-  quo, rem = [], []
-  for u in x.split_uop(Ops.ADD):
-    if (q:=u.divide_exact(y)) is not None: quo.append(q)
-    # if this is mod and y is a const, we can make the remainder factor sm
-    elif d.op is Ops.MOD and y.op is Ops.CONST and (c:=u.const_factor())%y.arg!=c:
-      rem.append(u.divides(c)*(c%y.arg))
-      quo.append(u.const_like(0))  # we append this so we can check if something changed
-    else: rem.append(u)
-  new_x = sum(rem)+x.const_like(0)
-  if len(quo)==0 or new_x.vmin<0: return None
-  return new_x%y if d.op is Ops.MOD else new_x//y+sum(quo)
-
-def nest_div_by_smallest_factor(d: UOp, x: UOp, y: UOp) -> UOp|None:
-  # we try and nest the div and see if it allows the numerator to be simplified
-  if ((c := y.arg) < 0): return None
-  factors = [u.const_factor() for u in x.split_uop(Ops.ADD) if u.op not in (Ops.CONST, Ops.VCONST)]
-  div = min([y.arg]+[abs(f) for f in factors if abs(f) > 1 and (c%f)==0])
-  newxs = fold_divmod_congruence(newx:=(x//div), x, y.const_like(div))
-  if newxs is None: newxs = factor_remainder(newx, x, y.const_like(div))
-  if div==y.arg or newxs is None or x.vmin<0 or newx.vmin<0: return None
-  return newxs//(c//div)
-
 def gep_through_wmma(gep:UOp, wmma:UOp):
   out_sz = prod(x[1] for x in wmma.arg[6][-1])
   wmma_idxs = gep.arg[::out_sz]
@@ -335,31 +239,9 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   # canonicalize a simplex with positive coefficients > 0
   # not x < 1 -> X > 0
   ((UPat.var("x", dtypes.index)<1).ne(True), lambda x: (newx<1).ne(True) if (newx:=canonicalize_simplex(x)) is not None else None),
-  # ** div **
-  # div folding
-  ((UPat.var("x")//UPat.cvar("c") + UPat.cvar("a"))//UPat.cvar("d"), lambda x,c,a,d: (x+a*c)//(c*d)
-    if c.vmin>0 and d.vmin>0 and ((x.vmin>=0 and a.vmin>=0) or (x.vmax<=0 and a.vmax<=0)) else None),  # (x//c+a)//d -> (x+a*c)//(c*d)
   # a range mod its own upper bound is just the range
   (UPat(Ops.RANGE, src=UPat.var("end"), name="r")%UPat.var("end"), lambda r,end: r),
   (UPat(Ops.RANGE, src=UPat.var("end"), name="r")//UPat.var("end"), lambda r,end: r.const_like(0)),
-  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d", src=(UPat.var("x"), UPat.var("y"))), cancel_divmod),
-  (UPat.var("x", dtypes.index) // UPat.var("d"), lambda x,d: -(x//(-d)) if d.vmax < 0 else None),
-  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d", src=(UPat.var("x"), UPat.cvar("y", vec=False))), fold_binary_numerator),
-  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d", src=(UPat.var("x"), UPat.cvar("y", vec=False))), fold_divmod_congruence),
-  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d", src=(UPat.var("x"), UPat.var("y"))), divide_by_gcd),
-  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d", src=(UPat.var("x"), UPat.cvar("y", vec=False))), gcd_with_remainder),
-  (UPat(Ops.MOD, dtypes.index, name="m", src=(UPat.var("x"), UPat.cvar("y", vec=False))), remove_nested_mod),
-  (UPat((Ops.IDIV), dtypes.index, name="d", src=(UPat.var("x"), UPat.cvar("y", vec=False))), nest_div_by_smallest_factor),
-  (UPat((Ops.IDIV, Ops.MOD), dtypes.index, name="d", src=(UPat.var("x"), UPat.var("y"))), factor_remainder),
-  (UPat.var("x", dtypes.index) // UPat.var("d"), lambda x,d: -((-x)//d) if x.vmax<=0 else None),
-  ((UPat.var("x", dtypes.index)+UPat.cvar("c", vec=False)).named("n")//UPat.cvar("d", vec=False),
-    lambda x,c,n,d: ((x+c.arg%d.arg)//d + c.arg//d.arg) if c.arg%d.arg!=c.arg and x.vmin>=0 and n.vmin>=0 and d.arg>0 else None),
-  ((UPat.var("x", dtypes.index)+UPat.cvar("c", vec=False)).named("n")//UPat.cvar("d", vec=False),
-    lambda x,c,n,d: (-(-(c.arg%d.arg + x - (d.arg-1))//d) + c.arg//d.arg) if x.vmax<=0 and n.vmin>=0 and d.arg>0 else None),
-  # ** mod **
-  # mod folding
-  (UPat.var("x", dtypes.index) % UPat.var("d"), lambda x,d: -((-x)%d) if x.vmax <= 0 else None),
-  (UPat.var("x", dtypes.index) % UPat.var("d"), lambda x,d: (x%(-d)) if d.vmax <  0 else None),
   # cast/long folding
   # if the intermediate cast doesnt narrow we can do it in one cast
   (UPat.var('x').cast(name="a").cast(name="b"), lambda x,a,b: x.cast(b.dtype) if can_safe_cast(x.dtype, a.dtype) else None),
@@ -376,7 +258,7 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
   (UPat(Ops.AFTER, src=(UPat.var("s"),)), lambda s: s),
   # VECTORIZE/CONST
   (UPat(Ops.VECTORIZE, src=UPat(Ops.CONST), name="vec"), lambda vec: UOp.const(vec.dtype, tuple(x.arg for x in vec.src))),
-])+gep_pushing
+])+div_and_mod_symbolic+gep_pushing
 
 # ******** we take a small aside to "simplify_valid" to rewrite valids ********
 
@@ -490,18 +372,9 @@ pm_simplify_valid = PatternMatcher([
 # this is symbolic 2.0
 REMOVE_FROM_SINK_LIKE = {Ops.UNROLL, Ops.NOOP, Ops.VECTORIZE, Ops.SINK}
 sym = symbolic+pm_simplify_valid+PatternMatcher([
-  # VECTORIZE/GEP
-  (UPat(Ops.VECTORIZE, src=UPat(Ops.GEP, src=(UPat.var("x"),)), name="vec"), lambda vec,x: x.gep(tuple(y.arg[0] for y in vec.src))),
   # reorder ALU/VECTORIZE
   (UPat(GroupOp.ALU, src=(UPat(Ops.VECTORIZE, src=UPat(name='x')), UPat(Ops.VECTORIZE, src=UPat(name='y'))), name='alu'),
    lambda x,y,alu: UOp(Ops.VECTORIZE, alu.dtype, (UOp(alu.op, alu.dtype.scalar(), (x,y)),)*alu.dtype.count)),
-  # VECTORIZE of a single element is just that element
-  (UPat(Ops.VECTORIZE, src=(UPat(name='x'),)), lambda x: x),
-  # VECTORIZE void is GROUP
-  (UPat(Ops.VECTORIZE, dtype=dtypes.void, name='x'), lambda x: UOp.group(*x.src)),
-  # tensor core with a 0 input is acc
-  (UPat(Ops.WMMA, src=(UPat.const(None, 0.0), UPat.var(), UPat.var("acc"))), lambda acc: acc),
-  (UPat(Ops.WMMA, src=(UPat.var(), UPat.const(None, 0.0), UPat.var("acc"))), lambda acc: acc),
   # ** self folding **
   # x!=0 -> (bool)x
   (UPat.var("x")!=0, lambda x: x.cast(dtypes.bool.vec(x.dtype.count))),