[flang-commits] [flang] 49a813b - [flang][openacc] Keep region when applying data operand conversion

[Valentin Clement via flang-commits]

Author: Valentin Clement
Date: 2023-04-12T08:20:34-07:00
New Revision: 49a813b4a6ea98e64373a7f7e8827a7e71e19189

URL: https://github.com/llvm/llvm-project/commit/49a813b4a6ea98e64373a7f7e8827a7e71e19189
DIFF: https://github.com/llvm/llvm-project/commit/49a813b4a6ea98e64373a7f7e8827a7e71e19189.diff

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

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.

Reviewed By: PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D148057

Added: 
    

Modified: 
    flang/lib/Optimizer/Transforms/OpenACC/OpenACCDataOperandConversion.cpp
    flang/test/Transforms/OpenACC/convert-data-operands-to-llvmir.fir

Removed: 
    


################################################################################
diff  --git a/flang/lib/Optimizer/Transforms/OpenACC/OpenACCDataOperandConversion.cpp b/flang/lib/Optimizer/Transforms/OpenACC/OpenACCDataOperandConversion.cpp
index b27795eca39aa..c68b06878949a 100644
--- a/flang/lib/Optimizer/Transforms/OpenACC/OpenACCDataOperandConversion.cpp
+++ b/flang/lib/Optimizer/Transforms/OpenACC/OpenACCDataOperandConversion.cpp
@@ -74,8 +74,21 @@ class LegalizeDataOpForLLVMTranslation : public ConvertOpToLLVMPattern<Op> {
       }
     }
 
-    builder.replaceOpWithNewOp<Op>(op, TypeRange(), convertedOperands,
-                                   op.getOperation()->getAttrs());
+    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();
   }

diff  --git a/flang/test/Transforms/OpenACC/convert-data-operands-to-llvmir.fir b/flang/test/Transforms/OpenACC/convert-data-operands-to-llvmir.fir
index df78e938da9e6..4a4225e4eca0e 100644
--- a/flang/test/Transforms/OpenACC/convert-data-operands-to-llvmir.fir
+++ b/flang/test/Transforms/OpenACC/convert-data-operands-to-llvmir.fir
@@ -1,5 +1,5 @@
 // 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>
@@ -83,28 +83,6 @@ func.func @_QQsub_parallel() attributes {fir.bindc_name = "test"} {
   %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
@@ -116,6 +94,7 @@ 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>>