[MLIR] Fix bug in addAffineParallelOpDomain upper bound constraint

Fix upper bound constraint addition in addAffineParallelOpDomain; it was
off by one in the case of constants.

Differential Revision: https://reviews.llvm.org/D144836

mlir/lib/Dialect/Affine/Analysis/AffineStructures.cpp

@@ -664,7 +664,7 @@ LogicalResult FlatAffineValueConstraints::addAffineParallelOpDomain(
 
     auto upperBound = parallelOp.getUpperBoundMap(ivPos);
     if (upperBound.isConstant())
-      addBound(BoundType::UB, pos, upperBound.getSingleConstantResult());
+      addBound(BoundType::UB, pos, upperBound.getSingleConstantResult() - 1);
     else if (failed(addBound(BoundType::UB, pos, upperBound,
                              parallelOp.getUpperBoundsOperands())))
       return failure();

mlir/test/Transforms/memref-dependence-check.mlir

@@ -1105,8 +1105,7 @@ func.func @affine_if_no_dependence() {
 
 // -----
 
-// CHECK-LABEL: func @affine_parallel_dep_check
-func.func @affine_parallel_dep_check() {
+func.func @affine_parallel_dep_check_1() {
   %memref_23 = memref.alloc() : memref<1x130xf32>
   %memref_25 = memref.alloc() : memref<1x130x130xf32>
   %cst = arith.constant 0.0 : f32
@@ -1130,3 +1129,20 @@ func.func @affine_parallel_dep_check() {
   }
   return
 }
+
+// -----
+
+func.func @affine_parallel_dep_check_2() {
+  %m = memref.alloc() : memref<128xf32>
+  %cst = arith.constant 0.0 : f32
+  affine.parallel (%arg4) = (0) to (127) {
+    affine.store %cst, %m[%arg4] : memref<128xf32>
+    // expected-remark@above {{dependence from 0 to 0 at depth 1 = false}}
+    // expected-remark@above {{dependence from 0 to 0 at depth 2 = false}}
+    // expected-remark@above {{dependence from 0 to 1 at depth 1 = false}}
+  }
+  affine.load %m[127]: memref<128xf32>
+  // expected-remark@above {{dependence from 1 to 1 at depth 1 = false}}
+  // expected-remark@above {{dependence from 1 to 0 at depth 1 = false}}
+  return
+}