fix: ignore load_ptr_tko padding when no mask is given

src/cuda/tile/_ir/ops.py

@@ -1326,7 +1326,8 @@ def generate_bytecode(self, ctx: BytecodeContext) -> tuple[bc.Value, bc.Value]:
             result_token_type=ctx.type_table.Token,
             source=ctx.get_value(self.pointer),
             mask=None if self.mask is None else ctx.get_value(self.mask),
-            paddingValue=ctx.get_value(self.padding_value),
+            paddingValue=(None if self.mask is None or self.padding_value is None
+                          else ctx.get_value(self.padding_value)),
             token=None if self.token is None else ctx.get_value(self.token),
             memory_ordering_semantics=bc.MemoryOrderingSemantics.WEAK,
             memory_scope=None,

test/test_gather_scatter.py

@@ -16,7 +16,7 @@
 from cuda.tile._ir.cast_ops import _is_implicit_cast_ok
 from cuda.tile._ir.typing_support import to_dtype
 from cuda.tile._compile import compile_tile
-from util import assert_equal, raises_if
+from util import assert_equal, filecheck, raises_if
 from conftest import float_dtypes, bool_dtypes, int_dtypes, dtype_id
 from torch.testing import make_tensor
 
@@ -216,6 +216,27 @@ def test_ir_checked_vs_unchecked(kernel, expected_mask):
     assert (store_ops[0].mask is not None) == expected_mask
 
 
+@ct.kernel
+def load_offset_unmasked(x, y):
+    ind = ct.arange(8, dtype=ct.int32)
+    t = x.get_raw_memory().load_offset(ind)
+    ct.scatter(y, ind, t)
+
+
+@pytest.mark.parametrize("kernel", [copy_8_unchecked, load_offset_unmasked])
+def test_unmasked_load_has_no_padding_operand(kernel):
+    x = torch.arange(10, 18, dtype=torch.float32, device="cuda:0")
+    y = torch.zeros_like(x)
+    sig = ct.compilation.KernelSignature.from_kernel_args(
+            kernel, (x, y),
+            ct.compilation.CallingConvention.cutile_python_v1())
+    bytecode = compile_tile(kernel._pyfunc, [sig], return_bytecode=True,
+                            return_cubin=False).bytecode
+
+    # The only operand of the load is the pointer: no mask, and so no padding either
+    filecheck(bytecode, "// CHECK: load_ptr_tko weak %{{[^,]*}} token=")
+
+
 # ============================================================================
 # Tests for custom mask parameter
 # ============================================================================