[flang-commits] [flang] [flang][OpenMP] fix lastprivate for allocatables (PR #99686)
via flang-commits
flang-commits at lists.llvm.org
Fri Jul 19 11:49:37 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-flang-fir-hlfir
@llvm/pr-subscribers-flang-openmp
Author: Tom Eccles (tblah)
<details>
<summary>Changes</summary>
1) Do not deallocate allocatable private variables if they are
lastprivate at the end of the openmp region because their address
will have been transferred to the host variable when the variable
is copied.
2) Don't use `copyHostAssociateVar` for allocatable variables. It isn't
clear to me whether or not this should be addressed in
`copyHostAssociateVar` instead of inside OpenMP. I opted for OpenMP
to minimise how many things I effected. `copyHostAssociateVar` will
not update the destination variable if the destination variable
was unallocated. This is incorrect because assignment inside of the
openmp block can cause the allocation status of the variable to
change. Furthermore, `copyHostAssociateVar` seems to only copy the
variable address not other metadata like the size of the allocation.
Reallocation by assignment could cause this to change.
Non-OpenMP reviewers: should this actually be done inside of `copyHostAssociateVar`?
OpenMP reviewers: I am aware that this will change with delayed privatization. My intention here is to squash the bug before the LLVM branch
---
Full diff: https://github.com/llvm/llvm-project/pull/99686.diff
2 Files Affected:
- (modified) flang/lib/Lower/OpenMP/DataSharingProcessor.cpp (+47-3)
- (added) flang/test/Lower/OpenMP/lastprivate-allocatable.f90 (+44)
``````````diff
diff --git a/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp b/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
index c1d4c089df3b2..597d62f469d75 100644
--- a/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
+++ b/flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
@@ -79,7 +79,8 @@ void DataSharingProcessor::processStep2(mlir::Operation *op, bool isLoop) {
void DataSharingProcessor::insertDeallocs() {
for (const semantics::Symbol *sym : allPrivatizedSymbols)
- if (semantics::IsAllocatable(sym->GetUltimate())) {
+ if (semantics::IsAllocatable(sym->GetUltimate()) &&
+ !sym->test(semantics::Symbol::Flag::OmpLastPrivate)) {
if (!useDelayedPrivatization) {
converter.createHostAssociateVarCloneDealloc(*sym);
continue;
@@ -127,8 +128,51 @@ void DataSharingProcessor::copyFirstPrivateSymbol(
void DataSharingProcessor::copyLastPrivateSymbol(
const semantics::Symbol *sym,
[[maybe_unused]] mlir::OpBuilder::InsertPoint *lastPrivIP) {
- if (sym->test(semantics::Symbol::Flag::OmpLastPrivate))
- converter.copyHostAssociateVar(*sym, lastPrivIP);
+ if (sym->test(semantics::Symbol::Flag::OmpLastPrivate)) {
+ bool allocatable = semantics::IsAllocatable(sym->GetUltimate());
+ if (!allocatable) {
+ converter.copyHostAssociateVar(*sym, lastPrivIP);
+ return;
+ }
+
+ // copyHostAssociateVar doesn't work properly if the privatised copy was
+ // reallocated (e.g. by assignment): it will only copy if the ultimate
+ // symbol was already allocated, and it only copies data so any reallocated
+ // lengths etc are lost
+
+ // 1) Fetch the original copy of the variable.
+ assert(sym->has<Fortran::semantics::HostAssocDetails>() &&
+ "No host-association found");
+ const Fortran::semantics::Symbol &hsym = sym->GetUltimate();
+ Fortran::lower::SymbolBox hsb = symTable->lookupOneLevelUpSymbol(hsym);
+ assert(hsb && "Host symbol box not found");
+
+ // 2) Fetch the copied one that will mask the original.
+ Fortran::lower::SymbolBox sb = symTable->shallowLookupSymbol(sym);
+ assert(sb && "Host-associated symbol box not found");
+ assert(hsb.getAddr() != sb.getAddr() &&
+ "Host and associated symbol boxes are the same");
+
+ // 3) Perform the assignment.
+ fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+ mlir::Location loc = converter.genLocation(sym->name());
+ mlir::OpBuilder::InsertPoint insPt = builder.saveInsertionPoint();
+ if (lastPrivIP && lastPrivIP->isSet())
+ builder.restoreInsertionPoint(*lastPrivIP);
+ else
+ builder.setInsertionPointAfter(sb.getAddr().getDefiningOp());
+
+ hlfir::Entity dst{hsb.getAddr()};
+ hlfir::Entity src{sb.getAddr()};
+ builder.create<hlfir::AssignOp>(
+ loc, src, dst, /*isWholeAllocatableAssignment=*/allocatable,
+ /*keepLhsLengthInAllocatableAssignment=*/false, /*temporary_lhs=*/true);
+
+ if (lastPrivIP && lastPrivIP->isSet() &&
+ sym->test(Fortran::semantics::Symbol::Flag::OmpLastPrivate)) {
+ builder.restoreInsertionPoint(insPt);
+ }
+ }
}
void DataSharingProcessor::collectOmpObjectListSymbol(
diff --git a/flang/test/Lower/OpenMP/lastprivate-allocatable.f90 b/flang/test/Lower/OpenMP/lastprivate-allocatable.f90
new file mode 100644
index 0000000000000..201c93576bda2
--- /dev/null
+++ b/flang/test/Lower/OpenMP/lastprivate-allocatable.f90
@@ -0,0 +1,44 @@
+! RUN: %flang_fc1 -emit-hlfir -o - -fopenmp %s | FileCheck %s
+! RUN: bbc -emit-hlfir -o - -fopenmp %s | FileCheck %s
+
+program lastprivate_allocatable
+ integer, allocatable :: a
+ integer :: i
+ ! a is unallocated here
+ !$omp parallel do lastprivate(a)
+ do i=1,1
+ a = 42
+ enddo
+ !$omp end parallel do
+ ! a should be allocated here
+end program
+
+! CHECK-LABEL: func.func @_QQmain()
+! CHECK: %[[VAL_0:.*]] = fir.alloca !fir.box<!fir.heap<i32>> {bindc_name = "a", uniq_name = "_QFEa"}
+! CHECK: %[[VAL_1:.*]] = fir.zero_bits !fir.heap<i32>
+! CHECK: %[[VAL_2:.*]] = fir.embox %[[VAL_1]] : (!fir.heap<i32>) -> !fir.box<!fir.heap<i32>>
+! CHECK: fir.store %[[VAL_2]] to %[[VAL_0]] : !fir.ref<!fir.box<!fir.heap<i32>>>
+! CHECK: %[[VAL_3:.*]]:2 = hlfir.declare %[[VAL_0]] {fortran_attrs = {{.*}}<allocatable>, uniq_name = "_QFEa"} : (!fir.ref<!fir.box<!fir.heap<i32>>>) -> (!fir.ref<!fir.box<!fir.heap<i32>>>, !fir.ref<!fir.box<!fir.heap<i32>>>)
+! CHECK: omp.parallel {
+! create original copy of private variable
+! CHECK: %[[VAL_16:.*]]:2 = hlfir.declare %{{.*}} {fortran_attrs = {{.*}}<allocatable>, uniq_name = "_QFEa"} : (!fir.ref<!fir.box<!fir.heap<i32>>>) -> (!fir.ref<!fir.box<!fir.heap<i32>>>, !fir.ref<!fir.box<!fir.heap<i32>>>)
+! CHECK: %[[VAL_17:.*]] = fir.alloca i32 {bindc_name = "i", pinned, uniq_name = "_QFEi"}
+! CHECK: %[[VAL_18:.*]]:2 = hlfir.declare %[[VAL_17]] {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+! CHECK: omp.wsloop {
+! CHECK: omp.loop_nest
+! [...]
+! if this is the last iteration
+! CHECK: fir.if %{{.*}} {
+! store loop IV
+! CHECK: fir.store %{{.*}} to %[[VAL_18]]#1 : !fir.ref<i32>
+! assign private variable to original copy: realloc and temporary_lhs
+! CHECK: hlfir.assign %[[VAL_16]]#0 to %[[VAL_3]]#0 realloc temporary_lhs : !fir.ref<!fir.box<!fir.heap<i32>>>, !fir.ref<!fir.box<!fir.heap<i32>>>
+! CHECK-NEXT: }
+! CHECK-NEXT: omp.yield
+! CHECK-NEXT: }
+! CHECK-NEXT: omp.terminator
+! CHECK-NEXT: }
+! [no deallocation here]
+! CHECK-NEXT: omp.terminator
+! CHECK-NEXT: }
+
``````````
</details>
https://github.com/llvm/llvm-project/pull/99686
More information about the flang-commits
mailing list