[flang][OpenMP] Support for common block in copyin clause

[NimishMishra]

Summary

This patch adds lowering support for threadprivate variables encapsulated in a common block and marked inside a copyin clause.

[kiranchandramohan]

flang/lib/Lower/OpenMP.cpp

1517–1518

Is a break or something missing here?

[NimishMishra]

flang/lib/Lower/OpenMP.cpp

1517–1518

Thank you for the spot. Indeed I missed. sym here represents the common block, and we need to copy host association details for all members of the block, and not just the last (which was what happened in previous iteration of the patch).

omp.parallel {
%[[val_6:.*]] = fir.alloca i32 {adapt.valuebyref, pinned}
%[[val_7:.*]] = omp.threadprivate %[[val_1]] : !fir.ref<!fir.array<8xi8>> -> !fir.ref<!fir.array<8xi8>>
%[[val_8:.*]] = fir.convert %[[val_7]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
%[[val_c0:.*]] = arith.constant 0 : index
%[[val_9:.*]] = fir.coordinate_of %[[val_8]], %[[val_c0]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
%[[val_10:.*]] = fir.convert %[[val_9]] : (!fir.ref<i8>) -> !fir.ref<i32>
%[[val_11:.*]] = fir.convert %[[val_7]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
%[[val_c4_0:.*]] = arith.constant 4 : index
%[[val_12:.*]] = fir.coordinate_of %[[val_11]], %[[val_c4_0]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
%[[val_13:.*]] = fir.convert %[[val_12]] : (!fir.ref<i8>) -> !fir.ref<i32>
%[[val_14:.*]] = fir.load %[[val_5]] : !fir.ref<i32>
fir.store %[[val_14]] to %[[val_13]] : !fir.ref<i32>
omp.barrier

This was the IR generated when both x and y were members of the common block. As evident, the conversion happens for both indices 0 and 4 (since both x and y are integers), but the fir.store happens only for y, and not for x.

The present version of the patch does it for both. Please have a look.

[kiranchandramohan]

LG.

[NimishMishra]

Hi @kiranchandramohan

Thanks for the LG. Small hiccup. The status here shows build failed, but when I go to https://buildkite.com/llvm-project/diff-checks/builds/190652, both linux and windows builds have passed. Can I go ahead with the merge? Seems like a buildbot issue.

flang/lib/Lower/OpenMP.cpp

return findRepeatableClause<ClauseTy::Allocate>(
                           allocateOperands);
       });
 }
 
 bool ClauseProcessor::processCopyin() const {
   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
   mlir::OpBuilder::InsertPoint insPt = firOpBuilder.saveInsertionPoint();
   firOpBuilder.setInsertionPointToStart(firOpBuilder.getAllocaBlock());
-
+  auto checkAndCopyHostAssociateVar =
+      [&](Fortran::semantics::Symbol *sym,
+          mlir::OpBuilder::InsertPoint *copyAssignIP = nullptr) {
+        assert(sym->has<Fortran::semantics::HostAssocDetails>() &&
+               "No host-association found");
+        converter.copyHostAssociateVar(*sym, copyAssignIP);
+      };
   bool hasCopyin = findRepeatableClause<ClauseTy::Copyin>(
       [&](const ClauseTy::Copyin *copyinClause,
           const Fortran::parser::CharBlock &) {
         const Fortran::parser::OmpObjectList &ompObjectList = copyinClause->v;
         for (const Fortran::parser::OmpObject &ompObject : ompObjectList.v) {
           Fortran::semantics::Symbol *sym = getOmpObjectSymbol(ompObject);
-          if (sym->has<Fortran::semantics::CommonBlockDetails>())
-            TODO(converter.getCurrentLocation(),
-                 "common block in Copyin clause");
+          if (const auto *commonDetails =
+                  sym->detailsIf<Fortran::semantics::CommonBlockDetails>()) {
+            for (const auto &mem : commonDetails->objects())
+              checkAndCopyHostAssociateVar(&*mem, &insPt);
+            break;
+          }
           if (Fortran::semantics::IsAllocatableOrPointer(sym->GetUltimate()))
             TODO(converter.getCurrentLocation(),
                  "pointer or allocatable variables in Copyin clause");
           assert(sym->has<Fortran::semantics::HostAssocDetails>() &&
                  "No host-association found");
-          converter.copyHostAssociateVar(*sym);
+          checkAndCopyHostAssociateVar(sym);
         }
       });
 
   // [OMP 5.0, 2.19.6.1] The copy is done after the team is formed and prior to
   // the execution of the associated structured block. Emit implicit barrier to
   // synchronize threads and avoid data races on propagation master's thread
   // values of threadprivate variables to local instances of that variables of
   // all other implicit threads.

flang/test/Lower/OpenMP/copyin.f90

subroutine combined_parallel_sections()
   !$omp parallel sections copyin(x7)
     !$omp section
       call sub5(x7)
     !$omp section
       call sub6(x7)
   !$omp end parallel sections
 
 end
+
+
+!CHECK: func.func @_QPcommon_1() {
+!CHECK: %[[val_0:.*]] = fir.address_of(@_QCc) : !fir.ref<!fir.array<4xi8>>
+!CHECK: %[[val_1:.*]] = omp.threadprivate %[[val_0]] : !fir.ref<!fir.array<4xi8>> -> !fir.ref<!fir.array<4xi8>>
+!CHECK: %[[val_2:.*]] = fir.convert %[[val_1]] : (!fir.ref<!fir.array<4xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[val_c0:.*]] = arith.constant 0 : index
+!CHECK: %[[val_3:.*]] = fir.coordinate_of %[[val_2]], %[[val_c0]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[val_4:.*]] = fir.convert %[[val_3]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[val_5:.*]] = fir.alloca i32 {bindc_name = "y", uniq_name = "_QFcommon_1Ey"}
+!CHECK: omp.parallel {
+!CHECK: %[[val_6:.*]] = omp.threadprivate %[[val_0]] : !fir.ref<!fir.array<4xi8>> -> !fir.ref<!fir.array<4xi8>>
+!CHECK: %[[val_7:.*]] = fir.convert %[[val_6]] : (!fir.ref<!fir.array<4xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[val_c0_0:.*]] = arith.constant 0 : index
+!CHECK: %[[val_8:.*]] = fir.coordinate_of %[[val_7]], %[[val_c0_0]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[val_9:.*]] = fir.convert %[[val_8]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[val_10:.*]] = fir.load %[[val_4]] : !fir.ref<i32>
+!CHECK: fir.store %[[val_10]] to %[[val_9]] : !fir.ref<i32>
+!CHECK: omp.barrier
+!CHECK: omp.sections   {
+!CHECK: omp.section {
+!CHECK: %[[val_11:.*]] = fir.load %[[val_9]] : !fir.ref<i32>
+!CHECK: %[[val_c1_i32:.*]] = arith.constant 1 : i32
+!CHECK: %[[val_12:.*]] = arith.addi %[[val_11]], %[[val_c1_i32]] : i32
+!CHECK: fir.store %[[val_12]] to %[[val_5]] : !fir.ref<i32>
+!CHECK: omp.terminator
+!CHECK: }
+!CHECK: omp.section {
+!CHECK: %[[val_11:.*]] = fir.load %[[val_5]] : !fir.ref<i32>
+!CHECK: %[[val_12:.*]] = fir.load %[[val_5]] : !fir.ref<i32>
+!CHECK: %[[val_13:.*]] = arith.muli %[[val_11]], %[[val_12]] : i32
+!CHECK: fir.store %[[val_13]] to %[[val_9]] : !fir.ref<i32>
+!CHECK: omp.terminator
+!CHECK: }
+!CHECK: omp.terminator
+!CHECK: }
+!CHECK: return
+!CHECK: }
+subroutine common_1()
+  integer :: x
+  integer :: y
+  common /c/ x
+  !$omp threadprivate(/c/)
+
+  !$omp parallel sections copyin(/c/)
+      !$omp section
+        y = x + 1
+      !$omp section
+        x = y * y
+  !$omp end parallel sections
+end subroutine
+
+!CHECK: func.func @_QPcommon_2() {
+!CHECK: %[[val_0:.*]] = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFcommon_2Ei"}
+!CHECK: %[[val_1:.*]] = fir.address_of(@_QCd) : !fir.ref<!fir.array<8xi8>>
+!CHECK: %[[val_2:.*]] = omp.threadprivate %[[val_1]] : !fir.ref<!fir.array<8xi8>> -> !fir.ref<!fir.array<8xi8>>
+!CHECK: %[[val_3:.*]] = fir.convert %[[val_2]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[val_c0:.*]] = arith.constant 0 : index
+!CHECK: %[[val_4:.*]] = fir.coordinate_of %[[val_3]], %[[val_c0]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[val_5:.*]] = fir.convert %[[val_4]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[val_6:.*]] = fir.convert %[[val_2]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[val_c4:.*]] = arith.constant 4 : index
+!CHECK: %[[val_7:.*]] = fir.coordinate_of %[[val_6]], %[[val_c4]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[val_8:.*]] = fir.convert %[[val_7]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: omp.parallel {
+!CHECK: %[[val_9:.*]] = fir.alloca i32 {adapt.valuebyref, pinned}
+!CHECK: %[[val_10:.*]] = omp.threadprivate %[[val_1]] : !fir.ref<!fir.array<8xi8>> -> !fir.ref<!fir.array<8xi8>>
+!CHECK: %[[val_11:.*]] = fir.convert %[[val_10]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[val_c0_0:.*]] = arith.constant 0 : index
+!CHECK: %[[val_12:.*]] = fir.coordinate_of %[[val_11]], %[[val_c0_0]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[val_13:.*]] = fir.convert %[[val_12]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[val_14:.*]] = fir.convert %[[val_10]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[val_c4_1:.*]] = arith.constant 4 : index
+!CHECK: %[[val_15:.*]] = fir.coordinate_of %[[val_14]], %[[val_c4_1]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[val_16:.*]] = fir.convert %[[val_15]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[val_17:.*]] = fir.load %[[val_5]] : !fir.ref<i32>
+!CHECK: fir.store %[[val_17]] to %[[val_13]] : !fir.ref<i32>
+!CHECK: %[[val_18:.*]] = fir.load %[[val_8]] : !fir.ref<i32>
+!CHECK: fir.store %[[val_18]] to %[[val_16]] : !fir.ref<i32>
+!CHECK: omp.barrier
+!CHECK: %[[val_c1_i32:.*]] = arith.constant 1 : i32
+!CHECK: %[[val_19:.*]] = fir.load %[[val_13]] : !fir.ref<i32>
+!CHECK: %[[val_c1_i32_2:.*]] = arith.constant 1 : i32
+!CHECK: omp.wsloop   for (%[[arg:.*]]) : i32 = (%[[val_c1_i32]]) to (%[[val_19]]) inclusive step (%[[val_c1_i32_2]]) {
+!CHECK: fir.store %[[arg]] to %[[val_9]] : !fir.ref<i32>
+!CHECK: %[[val_20:.*]] = fir.load %[[val_16]] : !fir.ref<i32>
+!CHECK: %[[val_21:.*]] = fir.load %[[val_9]] : !fir.ref<i32>
+!CHECK: %[[val_22:.*]] = arith.addi %[[val_20]], %[[val_21]] : i32
+!CHECK: fir.store %[[val_22]] to %[[val_16]] : !fir.ref<i32>
+!CHECK: omp.yield
+!CHECK: }
+!CHECK: omp.terminator
+!CHECK: }
+!CHECK: return
+!CHECK: }
+subroutine common_2()
+  integer :: x
+  integer :: y
+  common /d/ x, y
+  !$omp threadprivate(/d/)
+  
+  !$omp parallel do copyin(/d/)
+     do i = 1, x
+        y = y + i
+     end do
+  !$omp end parallel do
+end subroutine