Reverse loop generation order

llvm · Jan 9, 2024 · a377173 · a377173
1 parent ed74ee7
commit a377173
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 9 deletions.
diff --git a/flang/lib/Optimizer/HLFIR/Transforms/OptimizedBufferization.cpp b/flang/lib/Optimizer/HLFIR/Transforms/OptimizedBufferization.cpp
@@ -671,14 +671,15 @@ static mlir::Value generateReductionLoop(fir::FirOpBuilder &builder,
   mlir::Value oneIdx = builder.createIntegerConstant(loc, idxTy, 1);
 
   // Create a reduction loop nest. We use one-based indices so that they can be
-  // passed to the elemental.
-  llvm::SmallVector<mlir::Value> indices;
+  // passed to the elemental, and reverse the order so that they can be
+  // generated in column-major order for better performance.
+  llvm::SmallVector<mlir::Value> indices(extents.size(), mlir::Value{});
   for (unsigned i = 0; i < extents.size(); ++i) {
-    auto loop =
-        builder.create<fir::DoLoopOp>(loc, oneIdx, extents[i], oneIdx, false,
-                                      /*finalCountValue=*/false, reduction);
+    auto loop = builder.create<fir::DoLoopOp>(
+        loc, oneIdx, extents[extents.size() - i - 1], oneIdx, false,
+        /*finalCountValue=*/false, reduction);
     reduction = loop.getRegionIterArgs()[0];
-    indices.push_back(loop.getInductionVar());
+    indices[extents.size() - i - 1] = loop.getInductionVar();
     // Set insertion point to the loop body so that the next loop
     // is inserted inside the current one.
     builder.setInsertionPointToStart(loop.getBody());

diff --git a/flang/test/HLFIR/count-elemental.fir b/flang/test/HLFIR/count-elemental.fir
@@ -191,10 +191,10 @@ func.func @_QFPtest_multi(%arg0: !fir.ref<!fir.array<4x7x2xi32>> {fir.bindc_name
 // CHECK-NEXT:     %[[V5:.*]]:2 = hlfir.declare %arg2 {uniq_name = "_QFFtestEval"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
 // CHECK-NEXT:     %[[V6:.*]] = hlfir.designate %[[V1]]#0 (%c1:%c4:%c1, %c1:%c7:%c1, %c1:%c2:%c1)  shape %[[V0]] : (!fir.ref<!fir.array<4x7x2xi32>>, index, index, index, index, index, index, index, index, index, !fir.shape<3>) -> !fir.ref<!fir.array<4x7x2xi32>>
 // CHECK-NEXT:     %[[V7:.*]] = fir.load %[[V5]]#0 : !fir.ref<i32>
-// CHECK-NEXT:     %[[V8:.*]] = fir.do_loop %arg3 = %c1 to %c4 step %c1 iter_args(%arg4 = %c0_i32) -> (i32) {
+// CHECK-NEXT:     %[[V8:.*]] = fir.do_loop %arg3 = %c1 to %c2 step %c1 iter_args(%arg4 = %c0_i32) -> (i32) {
 // CHECK-NEXT:       %[[V10:.*]] = fir.do_loop %arg5 = %c1 to %c7 step %c1 iter_args(%arg6 = %arg4) -> (i32) {
-// CHECK-NEXT:         %[[V11:.*]] = fir.do_loop %arg7 = %c1 to %c2 step %c1 iter_args(%arg8 = %arg6) -> (i32) {
-// CHECK-NEXT:           %[[V12:.*]] = hlfir.designate %[[V6]] (%arg3, %arg5, %arg7)  : (!fir.ref<!fir.array<4x7x2xi32>>, index, index, index) -> !fir.ref<i32>
+// CHECK-NEXT:         %[[V11:.*]] = fir.do_loop %arg7 = %c1 to %c4 step %c1 iter_args(%arg8 = %arg6) -> (i32) {
+// CHECK-NEXT:           %[[V12:.*]] = hlfir.designate %[[V6]] (%arg7, %arg5, %arg3)  : (!fir.ref<!fir.array<4x7x2xi32>>, index, index, index) -> !fir.ref<i32>
 // CHECK-NEXT:           %[[V13:.*]] = fir.load %[[V12]] : !fir.ref<i32>
 // CHECK-NEXT:           %[[V14:.*]] = arith.cmpi sge, %[[V13]], %[[V7]] : i32
 // CHECK-NEXT:           %[[V15:.*]] = arith.addi %arg8, %c1_i32 : i32