[flang][cuda] Use fir.cuda_deallocate for automatic deallocation (#89662)

clementval · web-flow · commit 7c0da7993ecf · 2024-04-24T08:43:54.000-07:00
Automatic deallocation of allocatable that are cuda device variable must
use the fir.cuda_deallocate operation. This patch update the automatic
deallocation code generation to use this operation when the variable is
a cuda variable.

This patch has also the side effect to correctly call
`attachDeclarePostDeallocAction` for OpenACC declare variable on
automatic deallocation as well. Update the code in
`attachDeclarePostDeallocAction` so we do not attach on fir.result but
on the correct last op.
diff --git a/flang/include/flang/Lower/Allocatable.h b/flang/include/flang/Lower/Allocatable.h
@@ -55,12 +55,14 @@ void genDeallocateStmt(AbstractConverter &converter,
 
 void genDeallocateBox(AbstractConverter &converter,
                       const fir::MutableBoxValue &box, mlir::Location loc,
+                      const Fortran::semantics::Symbol *sym = nullptr,
                       mlir::Value declaredTypeDesc = {});
 
 /// Deallocate an allocatable if it is allocated at the end of its lifetime.
 void genDeallocateIfAllocated(AbstractConverter &converter,
                               const fir::MutableBoxValue &box,
-                              mlir::Location loc);
+                              mlir::Location loc,
+                              const Fortran::semantics::Symbol *sym = nullptr);
 
 /// Create a MutableBoxValue for an allocatable or pointer entity.
 /// If the variables is a local variable that is not a dummy, it will be
diff --git a/flang/lib/Lower/Allocatable.cpp b/flang/lib/Lower/Allocatable.cpp
@@ -859,18 +859,20 @@ genDeallocate(fir::FirOpBuilder &builder,
 void Fortran::lower::genDeallocateBox(
     Fortran::lower::AbstractConverter &converter,
     const fir::MutableBoxValue &box, mlir::Location loc,
-    mlir::Value declaredTypeDesc) {
+    const Fortran::semantics::Symbol *sym, mlir::Value declaredTypeDesc) {
   const Fortran::lower::SomeExpr *statExpr = nullptr;
   const Fortran::lower::SomeExpr *errMsgExpr = nullptr;
   ErrorManager errorManager;
   errorManager.init(converter, loc, statExpr, errMsgExpr);
   fir::FirOpBuilder &builder = converter.getFirOpBuilder();
-  genDeallocate(builder, converter, loc, box, errorManager, declaredTypeDesc);
+  genDeallocate(builder, converter, loc, box, errorManager, declaredTypeDesc,
+                sym);
 }
 
 void Fortran::lower::genDeallocateIfAllocated(
     Fortran::lower::AbstractConverter &converter,
-    const fir::MutableBoxValue &box, mlir::Location loc) {
+    const fir::MutableBoxValue &box, mlir::Location loc,
+    const Fortran::semantics::Symbol *sym) {
   fir::FirOpBuilder &builder = converter.getFirOpBuilder();
   mlir::Value isAllocated =
       fir::factory::genIsAllocatedOrAssociatedTest(builder, loc, box);
@@ -880,9 +882,9 @@ void Fortran::lower::genDeallocateIfAllocated(
             eleType.isa<fir::RecordType>() && box.isPolymorphic()) {
           mlir::Value declaredTypeDesc = builder.create<fir::TypeDescOp>(
               loc, mlir::TypeAttr::get(eleType));
-          genDeallocateBox(converter, box, loc, declaredTypeDesc);
+          genDeallocateBox(converter, box, loc, sym, declaredTypeDesc);
         } else {
-          genDeallocateBox(converter, box, loc);
+          genDeallocateBox(converter, box, loc, sym);
         }
       })
       .end();
diff --git a/flang/lib/Lower/ConvertVariable.cpp b/flang/lib/Lower/ConvertVariable.cpp
@@ -916,13 +916,14 @@ static void instantiateLocal(Fortran::lower::AbstractConverter &converter,
       break;
     case VariableCleanUp::Deallocate:
       auto *converterPtr = &converter;
-      converter.getFctCtx().attachCleanup([converterPtr, loc, exv]() {
+      auto *sym = &var.getSymbol();
+      converter.getFctCtx().attachCleanup([converterPtr, loc, exv, sym]() {
         const fir::MutableBoxValue *mutableBox =
             exv.getBoxOf<fir::MutableBoxValue>();
         assert(mutableBox &&
                "trying to deallocate entity not lowered as allocatable");
         Fortran::lower::genDeallocateIfAllocated(*converterPtr, *mutableBox,
-                                                 loc);
+                                                 loc, sym);
       });
     }
   }
diff --git a/flang/lib/Lower/OpenACC.cpp b/flang/lib/Lower/OpenACC.cpp
@@ -4187,21 +4187,27 @@ void Fortran::lower::attachDeclarePostDeallocAction(
 
   std::stringstream fctName;
   fctName << converter.mangleName(sym) << declarePostDeallocSuffix.str();
-  mlir::Operation &op = builder.getInsertionBlock()->back();
-  if (op.hasAttr(mlir::acc::getDeclareActionAttrName())) {
-    auto attr = op.getAttrOfType<mlir::acc::DeclareActionAttr>(
+  mlir::Operation *op = &builder.getInsertionBlock()->back();
+  if (auto resOp = mlir::dyn_cast<fir::ResultOp>(*op)) {
+    assert(resOp.getOperands().size() == 0 &&
+           "expect only fir.result op with no operand");
+    op = op->getPrevNode();
+  }
+  assert(op && "expect operation to attach the post deallocation action");
+  if (op->hasAttr(mlir::acc::getDeclareActionAttrName())) {
+    auto attr = op->getAttrOfType<mlir::acc::DeclareActionAttr>(
         mlir::acc::getDeclareActionAttrName());
-    op.setAttr(mlir::acc::getDeclareActionAttrName(),
-               mlir::acc::DeclareActionAttr::get(
-                   builder.getContext(), attr.getPreAlloc(),
-                   attr.getPostAlloc(), attr.getPreDealloc(),
-                   /*postDealloc=*/builder.getSymbolRefAttr(fctName.str())));
+    op->setAttr(mlir::acc::getDeclareActionAttrName(),
+                mlir::acc::DeclareActionAttr::get(
+                    builder.getContext(), attr.getPreAlloc(),
+                    attr.getPostAlloc(), attr.getPreDealloc(),
+                    /*postDealloc=*/builder.getSymbolRefAttr(fctName.str())));
   } else {
-    op.setAttr(mlir::acc::getDeclareActionAttrName(),
-               mlir::acc::DeclareActionAttr::get(
-                   builder.getContext(),
-                   /*preAlloc=*/{}, /*postAlloc=*/{}, /*preDealloc=*/{},
-                   /*postDealloc=*/builder.getSymbolRefAttr(fctName.str())));
+    op->setAttr(mlir::acc::getDeclareActionAttrName(),
+                mlir::acc::DeclareActionAttr::get(
+                    builder.getContext(),
+                    /*preAlloc=*/{}, /*postAlloc=*/{}, /*preDealloc=*/{},
+                    /*postDealloc=*/builder.getSymbolRefAttr(fctName.str())));
   }
 }
 
diff --git a/flang/test/Lower/CUDA/cuda-allocatable.cuf b/flang/test/Lower/CUDA/cuda-allocatable.cuf
@@ -17,6 +17,15 @@ end subroutine
 
 ! CHECK: %{{.*}} = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
 
+! CHECK: %[[BOX_LOAD:.*]] = fir.load %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>
+! CHECK: %[[ADDR:.*]] = fir.box_addr %[[BOX_LOAD]] : (!fir.box<!fir.heap<!fir.array<?xf32>>>) -> !fir.heap<!fir.array<?xf32>>
+! CHECK: %[[ADDR_I64:.*]] = fir.convert %[[ADDR]] : (!fir.heap<!fir.array<?xf32>>) -> i64
+! CHECK: %[[C0:.*]] = arith.constant 0 : i64
+! CHECK: %[[NE_C0:.*]] = arith.cmpi ne, %[[ADDR_I64]], %[[C0]] : i64
+! CHECK: fir.if %[[NE_C0]] {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: }
+
 subroutine sub2()
   real, allocatable, managed :: a(:)
   integer :: istat
@@ -37,6 +46,10 @@ end subroutine
 ! CHECK: %[[STAT:.*]] = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<managed>, hasStat} -> i32
 ! CHECK: fir.store %[[STAT]] to %[[ISTAT_DECL]]#1 : !fir.ref<i32>
 
+! CHECK: fir.if %{{.*}} {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<managed>} -> i32
+! CHECK: }
+
 subroutine sub3()
   integer, allocatable, pinned :: a(:,:)
   logical :: plog
@@ -50,6 +63,9 @@ end subroutine
 ! CHECK: %[[PLOG_DECL:.*]]:2 = hlfir.declare %5 {uniq_name = "_QFsub3Eplog"} : (!fir.ref<!fir.logical<4>>) -> (!fir.ref<!fir.logical<4>>, !fir.ref<!fir.logical<4>>)
 ! CHECK-2: fir.call @_FortranAAllocatableSetBounds
 ! CHECK: %{{.*}} = fir.cuda_allocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?x?xi32>>>> pinned(%[[PLOG_DECL]]#1 : !fir.ref<!fir.logical<4>>) {cuda_attr = #fir.cuda<pinned>} -> i32
+! CHECK: fir.if %{{.*}} {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?x?xi32>>>> {cuda_attr = #fir.cuda<pinned>} -> i32
+! CHECK: }
 
 subroutine sub4()
   real, allocatable, device :: a(:)
@@ -65,6 +81,9 @@ end subroutine
 ! CHECK: fir.call @_FortranAAllocatableSetBounds
 ! CHECK: %[[STREAM:.*]] = fir.load %[[ISTREAM_DECL]]#0 : !fir.ref<i32>
 ! CHECK: %{{.*}} = fir.cuda_allocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> stream(%[[STREAM]] : i32) {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: fir.if %{{.*}} {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: }
 
 subroutine sub5()
   real, allocatable, device :: a(:)
@@ -80,6 +99,11 @@ end subroutine
 ! CHECK: %[[LOAD_B:.*]] = fir.load %[[BOX_B_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>
 ! CHECK: fir.call @_FortranAAllocatableSetBounds
 ! CHECK: %{{.*}} = fir.cuda_allocate %[[BOX_A_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> source(%[[LOAD_B]] : !fir.box<!fir.heap<!fir.array<?xf32>>>) {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: fir.if
+! CHECK: fir.freemem
+! CHECK: fir.if %{{.*}} {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_A_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: }
 
 subroutine sub6()
   real, allocatable, device :: a(:)
@@ -95,6 +119,9 @@ end subroutine
 ! CHECK: %[[LOAD_B:.*]] = fir.load %[[BOX_B_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>
 ! CHECK: fir.call @_FortranAAllocatableApplyMold
 ! CHECK: %{{.*}} = fir.cuda_allocate %[[BOX_A_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: fir.if %{{.*}} {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_A_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: }
 
 subroutine sub7()
   real, allocatable, device :: a(:)
@@ -120,3 +147,6 @@ end subroutine
 ! CHECK: %[[ERR_BOX:.*]] = fir.embox %[[ERR_DECL]]#1 : (!fir.ref<!fir.char<1,50>>) -> !fir.box<!fir.char<1,50>>
 ! CHECK: %[[STAT:.*]] = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> errmsg(%15 : !fir.box<!fir.char<1,50>>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
 ! CHECK: fir.store %[[STAT]] to %[[ISTAT_DECL]]#1 : !fir.ref<i32>
+! CHECK: fir.if %{{.*}} {
+! CHECK:   %{{.*}} = fir.cuda_deallocate %[[BOX_DECL]]#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {cuda_attr = #fir.cuda<device>} -> i32
+! CHECK: }
diff --git a/flang/test/Lower/OpenACC/acc-declare.f90 b/flang/test/Lower/OpenACC/acc-declare.f90
@@ -245,6 +245,11 @@ subroutine acc_declare_allocate()
 ! CHECK: fir.freemem %{{.*}} : !fir.heap<!fir.array<?xi32>>
 ! CHECK: fir.store %{{.*}} to %{{.*}} {acc.declare_action = #acc.declare_action<postDealloc = @_QMacc_declareFacc_declare_allocateEa_acc_declare_update_desc_post_dealloc>} : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
 
+! CHECK: fir.if
+! CHECK: fir.freemem %{{.*}} : !fir.heap<!fir.array<?xi32>>
+! CHECK: fir.store %{{.*}} to %{{.*}}#1 {acc.declare_action = #acc.declare_action<postDealloc = @_QMacc_declareFacc_declare_allocateEa_acc_declare_update_desc_post_dealloc>} : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
+! CHECK: }
+
   end subroutine
 
 ! CHECK-LABEL: func.func private @_QMacc_declareFacc_declare_allocateEa_acc_declare_update_desc_post_alloc(
