[flang][openacc] Keep region when applying data operand conversion

[clementval]

Summary

Similar to D148039 but for the FIR to LLVM IR
conversion pass.

The inner part of the acc.loop has been removed since the rest of the
pipeline is not ready and would raise an error here. This was passing
until now because the acc.loop was discarded completely.

[PeteSteinfeld]

All builds and tests correctly and looks good.

flang/lib/Optimizer/Transforms/OpenACC/OpenACCDataOperandConversion.cpp

for (unsigned idx = 0; idx < numDataOperands; ++idx) {
                 .getResult(0);
         convertedOperands.push_back(castedOperand);
       } else {
         // Type not supported.
         return builder.notifyMatchFailure(op, "expecting a reference type");
       }
     }
 
+    if constexpr (std::is_same_v<Op, acc::ParallelOp> ||
+                  std::is_same_v<Op, acc::DataOp>) {
+      auto newOp =
+          builder.create<Op>(op.getLoc(), TypeRange(), convertedOperands,
+                             op.getOperation()->getAttrs());
+      builder.inlineRegionBefore(op.getRegion(), newOp.getRegion(),
+                                 newOp.getRegion().end());
+      if (failed(builder.convertRegionTypes(&newOp.getOperation()->getRegion(0),
+                                            *this->getTypeConverter())))
+        return failure();
+      builder.eraseOp(op);
+    } else {
       builder.replaceOpWithNewOp<Op>(op, TypeRange(), convertedOperands,
                                      op.getOperation()->getAttrs());
+    }
 
     return success();
   }
 };
 } // namespace
 
 namespace {
 struct OpenACCDataOperandConversion

flang/test/Transforms/OpenACC/convert-data-operands-to-llvmir.fir

 // RUN: fir-opt -fir-openacc-data-operand-conversion='use-opaque-pointers=1' -split-input-file %s | FileCheck %s
-// RUN: fir-opt -fir-openacc-data-operand-conversion='use-opaque-pointers=1' -split-input-file %s | fir-opt --fir-to-llvm-ir | FileCheck %s --check-prefix=LLVMIR
+// RUN: fir-opt -fir-openacc-data-operand-conversion='use-opaque-pointers=1' -split-input-file %s | fir-opt -split-input-file --fir-to-llvm-ir | FileCheck %s --check-prefix=LLVMIR
 
 fir.global internal @_QFEa : !fir.array<10xf32> {
   %0 = fir.undefined !fir.array<10xf32>
   fir.has_value %0 : !fir.array<10xf32>
 }
 
 func.func @_QQsub1() attributes {fir.bindc_name = "arr"} {
   %0 = fir.address_of(@_QFEa) : !fir.ref<!fir.array<10xf32>>
fir.global internal @_QFEa : !fir.array<10xf32> {
   fir.has_value %0 : !fir.array<10xf32>
 }
 
 func.func @_QQsub_parallel() attributes {fir.bindc_name = "test"} {
   %0 = fir.address_of(@_QFEa) : !fir.ref<!fir.array<10xf32>>
   %1 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
   acc.parallel copyin(%0: !fir.ref<!fir.array<10xf32>>) {
     acc.loop {
-      %c1_i32 = arith.constant 1 : i32
-      %2 = fir.convert %c1_i32 : (i32) -> index
-      %c10_i32 = arith.constant 10 : i32
-      %3 = fir.convert %c10_i32 : (i32) -> index
-      %c1 = arith.constant 1 : index
-      %4 = fir.convert %2 : (index) -> i32
-      %5:2 = fir.do_loop %arg0 = %2 to %3 step %c1 iter_args(%arg1 = %4) -> (index, i32) {
-        fir.store %arg1 to %1 : !fir.ref<i32>
-        %6 = fir.load %1 : !fir.ref<i32>
-        %7 = fir.convert %6 : (i32) -> f32
-        %c10_i64 = arith.constant 10 : i64
-        %c1_i64 = arith.constant 1 : i64
-        %8 = arith.subi %c10_i64, %c1_i64 : i64
-        %9 = fir.coordinate_of %0, %8 : (!fir.ref<!fir.array<10xf32>>, i64) -> !fir.ref<f32>
-        fir.store %7 to %9 : !fir.ref<f32>
-        %10 = arith.addi %arg0, %c1 : index
-        %11 = fir.convert %c1 : (index) -> i32
-        %12 = fir.load %1 : !fir.ref<i32>
-        %13 = arith.addi %12, %11 : i32
-        fir.result %10, %13 : index, i32
-      }
-      fir.store %5#1 to %1 : !fir.ref<i32>
       acc.yield
     }
     acc.yield
   }
   return
 }
 
 // CHECK-LABEL: func.func @_QQsub_parallel() attributes {fir.bindc_name = "test"} {
 // CHECK: %[[ADDR:.*]] = fir.address_of(@_QFEa) : !fir.ref<!fir.array<10xf32>>
 // CHECK: %[[CAST:.*]] = builtin.unrealized_conversion_cast %[[ADDR]] : !fir.ref<!fir.array<10xf32>> to !llvm.ptr<array<10 x f32>>
 // CHECK: acc.parallel copyin(%[[CAST]]: !llvm.ptr<array<10 x f32>>) {
+// CHECK:   acc.loop
 
 // LLVMIR-LABEL: llvm.func @_QQsub_parallel() attributes {fir.bindc_name = "test"} {
 // LLVMIR: %[[ADDR:.*]] = llvm.mlir.addressof @_QFEa : !llvm.ptr<array<10 x f32>>
 // LLVMIR: acc.parallel copyin(%[[ADDR]]: !llvm.ptr<array<10 x f32>>) {