[OpenMP] Fix assertion when array slice is passed as an argument in parallel

Author: kiranchandramohan

flang/lib/Optimizer/CodeGen/CodeGen.cpp

@@ -983,8 +983,22 @@ struct EmboxCommonConversion : public FIROpConversion<OP> {
                     mlir::ConversionPatternRewriter &rewriter) const {
     auto thisPt = rewriter.saveInsertionPoint();
     auto *thisBlock = rewriter.getInsertionBlock();
-    auto func = mlir::cast<mlir::LLVM::LLVMFuncOp>(thisBlock->getParentOp());
-    rewriter.setInsertionPointToStart(&func.front());
+    auto op = thisBlock->getParentOp();
+    // Order to find the Op in whose entry block the alloca should be inserted.
+    // The parent Op if it is an LLVM Function Op.
+    // The ancestor OpenMP Op which is outlineable.
+    // The ancestor LLVM Function Op.
+    if (auto iface =
+             thisBlock->getParent()
+                 ->getParentOfType<
+                     mlir::omp::OutlineableOpenMPOpInterface>()) {
+       rewriter.setInsertionPointToStart(iface.getAllocaBlock());
+     } else {
+       auto func = mlir::isa<mlir::LLVM::LLVMFuncOp>(op) ? 
+           mlir::cast<mlir::LLVM::LLVMFuncOp>(op) : 
+           op->getParentOfType<mlir::LLVM::LLVMFuncOp>();
+       rewriter.setInsertionPointToStart(&func.front());
+     }
     auto sz = this->genConstantOffset(loc, rewriter, 1);
     auto al = rewriter.create<mlir::LLVM::AllocaOp>(loc, toTy, sz, alignment);
     rewriter.restoreInsertionPoint(thisPt);

flang/test/Lower/OpenMP/omp-parallel-alloca-fix.f90

@@ -0,0 +1,64 @@
+! This test checks passing an array slice in a parallel region
+
+! RUN: bbc -fopenmp %s -o - | tco 2>&1 | FileCheck %s
+
+! CHECK-LABEL: @_QPsb1..omp_par
+! CHECK-LABEL: omp.par.region1
+! CHECK:   %[[SB1_MAT:.*]] = alloca { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, align 8
+! CHECK:   %[[SB1_MAT1:.*]] = getelementptr { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }* %[[SB1_MAT]], i32 0, i32 0
+! CHECK:   %[[SB1_MAT2:.*]] = load i32*, i32** %[[SB1_MAT1]], align 8
+! CHECK:   %[[SB1_MAT3:.*]] = bitcast i32* %[[SB1_MAT2]] to [3 x i32]*
+! CHECK:   call void @_QPouter_src_calc([3 x i32]* %[[SB1_MAT3]])
+
+subroutine sb1
+  IMPLICIT NONE
+  INTEGER, DIMENSION(3, 3) :: mat
+  INTEGER :: k
+
+  !$OMP PARALLEL
+  DO k = 1, 2
+     CALL outer_src_calc (  mat(:,k) )
+  END DO
+  !$OMP END PARALLEL
+end subroutine
+
+! CHECK-LABEL: @_QPsb2..omp_par
+! CHECK-LABEL: omp.par.region1
+! CHECK:   %[[SB2_MAT:.*]] = alloca { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, align 8
+! CHECK:   call i32 @__kmpc_master
+! CHECK:   %[[SB2_MAT1:.*]] = getelementptr { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }* %[[SB2_MAT]], i32 0, i32 0
+! CHECK:   %[[SB2_MAT2:.*]] = load i32*, i32** %[[SB2_MAT1]], align 8
+! CHECK:   %[[SB2_MAT3:.*]] = bitcast i32* %[[SB2_MAT2]] to [3 x i32]*
+! CHECK:   call void @_QPouter_src_calc([3 x i32]* %[[SB2_MAT3]])
+subroutine sb2
+  IMPLICIT NONE
+  INTEGER, DIMENSION(3, 3) :: mat
+  INTEGER :: k
+
+  !$OMP PARALLEL
+  !$OMP MASTER
+  DO k = 1, 2
+     CALL outer_src_calc (  mat(:,k) )
+  END DO
+  !$OMP END MASTER
+  !$OMP END PARALLEL
+end subroutine
+
+! CHECK-LABEL: @_QPsb3
+! CHECK:   %[[SB3_MAT:.*]] = alloca { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, align 8
+! CHECK:   call i32 @__kmpc_master
+! CHECK:   %[[SB3_MAT1:.*]] = getelementptr { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }, { i32*, i64, i32, i8, i8, i8, i8, [1 x [3 x i64]] }* %[[SB3_MAT]], i32 0, i32 0
+! CHECK:   %[[SB3_MAT2:.*]] = load i32*, i32** %[[SB3_MAT1]], align 8
+! CHECK:   %[[SB3_MAT3:.*]] = bitcast i32* %[[SB3_MAT2]] to [3 x i32]*
+! CHECK:   call void @_QPouter_src_calc([3 x i32]* %[[SB3_MAT3]])
+subroutine sb3
+  IMPLICIT NONE
+  INTEGER, DIMENSION(3, 3) :: mat
+  INTEGER :: k
+
+  !$OMP MASTER
+  DO k = 1, 2
+     CALL outer_src_calc (  mat(:,k) )
+  END DO
+  !$OMP END MASTER
+end subroutine