[Mlir-commits] [mlir] [mlir][affine] Clamp negative static trip counts to zero (PR #227134)

[Jeremy Kun]

================
@@ -965,6 +965,27 @@ func.func @no_hoist_from_zero_trip_count_loop(%x: i32) -> i32 {
 
 // -----
 
+// The bounds are not constant, but the upper bound is below the lower bound, so
+// this loop never executes either.
+
+// CHECK-LABEL: func @no_hoist_from_zero_trip_count_loop_nonconstant_bounds
+func.func @no_hoist_from_zero_trip_count_loop_nonconstant_bounds(%x: i32, %n: index) -> i32 {
+  %alloc = memref.alloc() : memref<1xi32>
+  %c42 = arith.constant 42 : i32
+  affine.store %c42, %alloc[0] : memref<1xi32>
+  affine.for %i = affine_map<(d0) -> (d0)>(%n) to affine_map<(d0) -> (d0 - 2)>(%n) {
+    affine.store %x, %alloc[0] : memref<1xi32>
+  }
+  // CHECK:      affine.store %c42{{.*}}, %alloc[0]
+  // CHECK:      affine.for
+  // CHECK-NEXT:   affine.store %arg0, %alloc[0]
----------------
j2kun wrote:

nit: avoid hard-coding the variable names into the check statements, as the specific choice of names is not guaranteed by MLIR and it makes these tests more brittle.

https://github.com/llvm/llvm-project/pull/227134