[flang-commits] [flang] [Flang] Generate inline reduction loops for elemental count intrinsics (PR #75774)

[via flang-commits]

================
@@ -659,6 +659,124 @@ mlir::LogicalResult VariableAssignBufferization::matchAndRewrite(
   return mlir::success();
 }
 
+using GenBodyFn =
+    std::function<mlir::Value(fir::FirOpBuilder &, mlir::Location, mlir::Value,
+                              const llvm::SmallVectorImpl<mlir::Value> &)>;
+static mlir::Value generateReductionLoop(fir::FirOpBuilder &builder,
+                                         mlir::Location loc, mlir::Value init,
+                                         mlir::Value shape, GenBodyFn genBody) {
+  auto extents = hlfir::getIndexExtents(loc, builder, shape);
+  mlir::Value reduction = init;
+  mlir::IndexType idxTy = builder.getIndexType();
+  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;
+  for (unsigned i = 0; i < extents.size(); ++i) {
+    auto loop =
+        builder.create<fir::DoLoopOp>(loc, oneIdx, extents[i], oneIdx, false,
+                                      /*finalCountValue=*/false, reduction);
----------------
jeanPerier wrote:

Fortran is column major, so this should be ` for (auto extent : llvm::reverse(extents)) {` and you will need to push the indices from the back.
This is not required for correctness, but in order to get optimal memory accesses. See hlfir::genLoops.



https://github.com/llvm/llvm-project/pull/75774