Reapply and fix - [flang][OpenMP] Support delayed privatisation for composite distribute simd (llvm#151169) (llvm#3450)

Author: ronlieb

flang/lib/Lower/OpenMP/OpenMP.cpp

@@ -3088,12 +3088,20 @@ static mlir::omp::DistributeOp genCompositeDistributeSimd(
   genSimdClauses(converter, semaCtx, simdItem->clauses, loc, simdClauseOps,
                  simdReductionSyms);
 
-  // TODO: Support delayed privatization.
-  DataSharingProcessor dsp(converter, semaCtx, simdItem->clauses, eval,
-                           /*shouldCollectPreDeterminedSymbols=*/true,
-                           /*useDelayedPrivatization=*/false, symTable);
-  dsp.processStep1();
-  dsp.processStep2();
+
+  DataSharingProcessor distributeItemDSP(
+      converter, semaCtx, distributeItem->clauses, eval,
+      /*shouldCollectPreDeterminedSymbols=*/false,
+      /*useDelayedPrivatization=*/true, symTable);
+  distributeItemDSP.processStep1();
+  distributeItemDSP.processStep2(&distributeClauseOps);
+
+  DataSharingProcessor simdItemDSP(converter, semaCtx, simdItem->clauses, eval,
+                                   /*shouldCollectPreDeterminedSymbols=*/true,
+                                   /*useDelayedPrivatization=*/true, symTable);
+  simdItemDSP.processStep1();
+  simdItemDSP.processStep2(&simdClauseOps);
+
   // Pass the innermost leaf construct's clauses because that's where COLLAPSE
   // is placed by construct decomposition.
   mlir::omp::LoopNestOperands loopNestClauseOps;
@@ -3103,13 +3111,15 @@ static mlir::omp::DistributeOp genCompositeDistributeSimd(
 
   // Operation creation.
   EntryBlockArgs distributeArgs;
-  // TODO: Add private syms and vars.
+  distributeArgs.priv.syms = distributeItemDSP.getDelayedPrivSymbols();
+  distributeArgs.priv.vars = distributeClauseOps.privateVars;
   auto distributeOp = genWrapperOp<mlir::omp::DistributeOp>(
       converter, loc, distributeClauseOps, distributeArgs);
   distributeOp.setComposite(/*val=*/true);
 
   EntryBlockArgs simdArgs;
-  // TODO: Add private syms and vars.
+  simdArgs.priv.syms = simdItemDSP.getDelayedPrivSymbols();
+  simdArgs.priv.vars = simdClauseOps.privateVars;
   simdArgs.reduction.syms = simdReductionSyms;
   simdArgs.reduction.vars = simdClauseOps.reductionVars;
   auto simdOp =
@@ -3119,7 +3129,7 @@ static mlir::omp::DistributeOp genCompositeDistributeSimd(
   genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, simdItem,
                 loopNestClauseOps, iv,
                 {{distributeOp, distributeArgs}, {simdOp, simdArgs}},
-                llvm::omp::Directive::OMPD_distribute_simd, dsp);
+                llvm::omp::Directive::OMPD_distribute_simd, simdItemDSP);
   return distributeOp;
 }
 

flang/test/Lower/OpenMP/distribute-simd.f90

@@ -7,7 +7,7 @@
 subroutine distribute_simd_aligned(A)
   use iso_c_binding
   type(c_ptr) :: A
-  
+
   !$omp teams
 
   ! CHECK:      omp.distribute
@@ -57,3 +57,22 @@ subroutine distribute_simd_simdlen()
 
   !$omp end teams
 end subroutine distribute_simd_simdlen
+
+! CHECK-LABEL: func.func @_QPdistribute_simd_private(
+subroutine distribute_simd_private()
+  integer, allocatable :: tmp
+  ! CHECK:      omp.teams
+  !$omp teams
+  ! CHECK:      omp.distribute
+  ! CHECK:      omp.simd
+  ! CHECK-SAME: private(@[[PRIV_BOX_SYM:.*]] %{{.*}} -> %[[PRIV_BOX:.*]], @[[PRIV_IVAR_SYM:.*]] %{{.*}} -> %[[PRIV_IVAR:.*]] : !fir.ref<!fir.box<!fir.heap<i32>>>, !fir.ref<i32>)
+  ! CHECK-NEXT: omp.loop_nest (%[[IVAR:.*]]) : i32
+  !$omp distribute simd private(tmp)
+  do index_ = 1, 10
+  ! CHECK:      %[[PRIV_BOX_DECL:.*]]:2 = hlfir.declare %[[PRIV_BOX]]
+  ! CHECK:      %[[PRIV_IVAR_DECL:.*]]:2 = hlfir.declare %[[PRIV_IVAR]]
+  ! CHECK:      hlfir.assign %[[IVAR]] to %[[PRIV_IVAR_DECL]]#0
+  end do
+  !$omp end distribute simd
+  !$omp end teams
+end subroutine distribute_simd_private