diff --git a/pspamm/codegen/architectures/arm_sve/generator.py b/pspamm/codegen/architectures/arm_sve/generator.py
index d1bca7f..d5007cf 100644
--- a/pspamm/codegen/architectures/arm_sve/generator.py
+++ b/pspamm/codegen/architectures/arm_sve/generator.py
@@ -379,22 +379,14 @@ def make_microkernel(self,
             divider = 1
             elem128 = 1
             vk = bk
-            preg = 'p7/z'
-            preg_last = 'p7/z'
         else:
             max_offs = 127
             divider = 16
             elem128 = 16 // self.get_precision().size()
             vk = -(bk // -elem128)
 
-            #if isinstance(B, DenseCursor):
-            #    preg = 'p1/z' if self.has_bk_overhead else 'p7/z'
-            #    preg_last = 'p2/z' if self.has_k_overhead else preg
-            #else:
-            #    preg = 'p7/z'
-            #    preg_last = 'p7/z'
-            preg = 'p7/z'
-            preg_last = 'p7/z'
+        preg = self.pred_n_trues(elem128, elem128, 'z')
+        preg_last = preg if bk % elem128 == 0 else self.pred_n_trues(bk % elem128, elem128, 'z')
         for Vmi in range(Vm):
             # set to all v_size predicates to true, we want to replicate a B element into a whole vector
             for bni in range(bn):  # inside this n-block
@@ -404,7 +396,7 @@ def make_microkernel(self,
                     if B.has_nonzero_cell(B_ptr, to_B_block, to_cell):
                         B_cell_addr, B_comment = B.look(B_ptr, to_B_block, to_cell)
                         if B_regs[bki_reg, bni] not in bs:
-                            p_zeroing = Register_ARM(AsmType.p64x8, preg_last) if bki_reg + 1 == vk else Register_ARM(AsmType.p64x8, preg)
+                            p_zeroing = preg_last if bki_reg + 1 == vk else preg
 
                             # max_offs is the maximum allowed immediate offset when using ld1rd/ld1rw to broadcast a scalar value
                             if B_cell_addr.disp > max_offs or B_cell_addr.disp % divider != 0: