[bug] Fix the mapping from virtual axes to physical axes again

taichi/transforms/scalar_pointer_lowerer.cpp

@@ -60,6 +60,7 @@ void ScalarPointerLowerer::run() {
   if (path_length_ == 0)
     return;
 
+  auto *leaf_snode = snodes_[path_length_ - 1];
   Stmt *last = lowered_->push_back<GetRootStmt>(snodes_[0]);
   for (int i = 0; i < path_length_; i++) {
     auto *snode = snodes_[i];
@@ -72,8 +73,8 @@ void ScalarPointerLowerer::run() {
     std::vector<int> strides;
     // extract lowered indices
     for (int k_ = 0; k_ < (int)indices_.size(); k_++) {
-      int k = snode->physical_index_position[k_];
-      if (k < 0)
+      int k = leaf_snode->physical_index_position[k_];
+      if (!snode->extractors[k].active)
         continue;
       Stmt *extracted;
       if (packed_) {  // no dependence on POT

tests/python/test_indices.py

@@ -18,6 +18,21 @@ def test_indices():
     # Note that b is column-major:
     # the virtual first index exposed to the user comes second in memory layout.
 
+    @ti.kernel
+    def fill():
+        for i, j in b:
+            b[i, j] = i * 10 + j
+
+    @ti.kernel
+    def get_field_addr(i: ti.i32, j: ti.i32) -> ti.u64:
+        return ti.get_addr(b, [i, j])
+
+    fill()
+    for i in range(16):
+        for j in range(32):
+            assert b[i, j] == i * 10 + j
+    assert get_field_addr(0, 1) + 4 == get_field_addr(1, 1)
+
 
 @ti.test(arch=ti.get_host_arch_list())
 def test_float_as_index():