[Mlir-commits] [mlir] [mlir] Don't hoist transfers from potentially zero trip loops (PR #112752)

[Andrzej Warzyński]

================
@@ -308,6 +308,61 @@ module attributes {transform.with_named_sequence} {
 
 // -----
 
+// CHECK-LABEL:  func.func @no_hoisting_zero_trip_loop
+func.func @no_hoisting_zero_trip_loop(%arg0: memref<20xi32>, %lb: index, %ub: index) {
+  %c0_i32 = arith.constant 0 : i32
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  // %lb and %ub are unbounded, so do not hoist.
+
+  // CHECK:       scf.for {{.*}} {
+  // CHECK-NEXT:    vector.transfer_read
+  // CHECK-NEXT:    "prevent.dce"
+  scf.for %arg2 = %lb to %ub step %c1 {
+    %read = vector.transfer_read %arg0[%c0], %c0_i32 {in_bounds = [true]} : memref<20xi32>, vector<4xi32>
+    "prevent.dce"(%read) : (vector<4xi32>) ->()
----------------
banach-space wrote:

[nit] "test.some_use" works without requiring `--allow-unregistered-dialect` and as such, it feels like better practice.

https://github.com/llvm/llvm-project/pull/112752