fix cast in cstyle (#18070)

test/backend/test_call.py

@@ -1,7 +1,8 @@
 import unittest, ctypes
 from tinygrad import Tensor, UOp
 from tinygrad.device import Device
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, AddrSpace
+from tinygrad.codegen import to_program
 from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.uop.ops import KernelInfo
 
@@ -37,4 +38,10 @@ class TestCall(unittest.TestCase):
     c.realize()
     self.assertEqual(c.item(), 44)
 
+  def test_call_stack_pointer(self):
+    slot = UOp.placeholder((1,), dtypes.uint32, addrspace=AddrSpace.REG)
+    call = UOp.custom_function("callback", UOp.const(0, dtypes.uint64)).call(slot[0], ret_dtype=dtypes.void)
+    prg = to_program(call.sink(arg=KernelInfo("call_stack")), Device["CPU"].renderer)
+    self.assertIn("(unsigned int*)((buf", prg.src[2].arg)
+
 if __name__ == "__main__": unittest.main()

tinygrad/renderer/cstyle.py

@@ -187,7 +187,8 @@ class CStyleLanguage(Renderer):
       return prefix + self.type_map.get(dtype, dtype.name).replace(" ", "_") + str(sz) + suffix
     return prefix + self.type_map.get(dtype, dtype.name) + suffix
 
-  def render_type(self, u:UOp): return self._render_dtype(u.dtype, u.max_numel(), u.addrspace, shape=u._shape)
+  def render_type(self, u:UOp):
+    return self._render_dtype(u.dtype, u.max_numel(), u.addrspace, shape=u._shape, override_ptr=u.op is Ops.INDEX and u.addrspace is AddrSpace.REG)
   def render_ptr(self, u:UOp):
     # the address of an access, vector-cast if the access reads/writes more lanes than the pointer's scalar type
     if u.max_numel() > 1 or u.dtype != u.src[0].dtype: