[flang-commits] [flang] [llvm] [mlir] [Flang][OpenMP][MLIR] Fix common block mapping for regular and declare target link (PR #91829)
via flang-commits
flang-commits at lists.llvm.org
Fri May 10 18:15:03 PDT 2024
https://github.com/agozillon created https://github.com/llvm/llvm-project/pull/91829
This PR attempts to fix common block mapping for regular mapping of these types as well as when they have been marked as "declare target link". This PR should allow correct mapping of both the members of a common block and the full common block via its block symbol.
The main changes were some adjustments to the Fortran OpenMP lowering to HLFIR/FIR, the lowering of the LLVM+OpenMP dialect to LLVM-IR and adjustments to the way the we handle target kernel map argument rebinding inside of the OMPIRBuilder.
For the Fortran OpenMP lowering were two changes, one to prevent the implicit capture of common block members when the common block symbol itself has been marked and the other creates intermediate member access inside of the target region to be used in-place of those external to the target region, this prevents external usages breaking the IsolatedFromAbove pact.
In the latter case, there was an adjustment to the size calculation for types to better handle cases where we pass an array as the type of a map (as opposed to the bounds and the type of the element), which occurs in the case of common blocks. There is also some adjustment to how handleDeclareTargetMapVar handles renaming of declare target symbols in the module to the reference pointer, now it will only apply to those within the kernel that is currently being generated and we also perform a modification to replace constants with instructions as necessary as we cannot replace these with our reference pointer (non-constant and constants do not mix nicely).
In the case of the OpenMPIRBuilder some changes were made to defer global symbol rebinding to kernel arguments until all other arguments have been rebound. This makes sure we do not replace uses that may refer to the global (e.g. a GEP) but are themselves actually a separate argument that needs bound.
Currently "declare target to" still needs some work, but this may be the case for all types in conjunction with "declare target to" at the moment.
>From 0c09c3a505a00d00a1f88322ab24ce5201a165dd Mon Sep 17 00:00:00 2001
From: agozillon <Andrew.Gozillon at amd.com>
Date: Fri, 10 May 2024 20:10:52 -0500
Subject: [PATCH] [Flang][OpenMP][MLIR] Fix common block mapping for regular
and declare target link
This PR attempts to fix common block mapping for regular mapping of these types as
well as when they have been marked as "declare target link". This PR should allow correct
mapping of both the members of a common block and the full common block via its
block symbol.
The main changes were some adjustments to the Fortran OpenMP lowering to HLFIR/FIR,
the lowering of the LLVM+OpenMP dialect to LLVM-IR and adjustments to the way the
we handle target kernel map argument rebinding inside of the OMPIRBuilder.
For the Fortran OpenMP lowering were two changes, one to prevent the implicit capture
of common block members when the common block symbol itself has been marked and
the other creates intermediate member access inside of the target region to be used
in-place of those external to the target region, this prevents external usages breaking the
IsolatedFromAbove pact.
In the latter case, there was an adjustment to the size calculation for types to better
handle cases where we pass an array as the type of a map (as opposed to the
bounds and the type of the element), which occurs in the case of common blocks. There
is also some adjustment to how handleDeclareTargetMapVar handles renaming of declare
target symbols in the module to the reference pointer, now it will only apply to those
within the kernel that is currently being generated and we also perform a modification
to replace constants with instructions as necessary as we cannot replace these with our
reference pointer (non-constant and constants do not mix nicely).
In the case of the OpenMPIRBuilder some changes were mde to defer global symbol
rebinding to kernel arguments until all other arguments have been rebound. This
makes sure we do not replace uses that may refer to the global (e.g. a GEP) but are
themselves actually a separate argument that needs bound.
Currently "declare target to" still needs some work, but this may be the case for all
types in conjunction with "declare target to" at the moment.
---
flang/lib/Lower/OpenMP/OpenMP.cpp | 46 +++++++++
.../OpenMP/map-types-and-sizes.f90 | 41 ++++++++
flang/test/Lower/OpenMP/common-block-map.f90 | 83 ++++++++++++++++
.../llvm/Frontend/OpenMP/OMPIRBuilder.h | 5 +
llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp | 74 +++++++++++----
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 42 ++++----
.../target-map-all-common-block-members.f90 | 55 +++++++++++
.../fortran/target-map-common-block.f90 | 50 ++++++++++
...t-map-declare-target-link-common-block.f90 | 95 +++++++++++++++++++
.../target-map-first-common-block-member.f90 | 47 +++++++++
...t-map-mix-imp-exp-common-block-members.f90 | 58 +++++++++++
.../target-map-second-common-block-member.f90 | 47 +++++++++
12 files changed, 607 insertions(+), 36 deletions(-)
create mode 100644 flang/test/Lower/OpenMP/common-block-map.f90
create mode 100644 offload/test/offloading/fortran/target-map-all-common-block-members.f90
create mode 100644 offload/test/offloading/fortran/target-map-common-block.f90
create mode 100644 offload/test/offloading/fortran/target-map-declare-target-link-common-block.f90
create mode 100644 offload/test/offloading/fortran/target-map-first-common-block-member.f90
create mode 100644 offload/test/offloading/fortran/target-map-mix-imp-exp-common-block-members.f90
create mode 100644 offload/test/offloading/fortran/target-map-second-common-block-member.f90
diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index f23902d6a8239..07a8535d7bc6a 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -830,6 +830,33 @@ static void genBodyOfTargetDataOp(
genNestedEvaluations(converter, eval);
}
+// This generates intermediate common block member accesses within a region
+// and then rebinds the members symbol to the intermediate accessors we have
+// generated so that subsequent code generation will utilise these instead.
+//
+// When the scope changes, the bindings to the intermediate accessors should
+// be dropped in place of the original symbol bindings.
+//
+// This is for utilisation with TargetOp.
+static void genIntermediateCommonBlockAccessors(
+ Fortran::lower::AbstractConverter &converter,
+ const mlir::Location ¤tLocation, mlir::Region ®ion,
+ llvm::ArrayRef<const Fortran::semantics::Symbol *> mapSyms) {
+ for (auto [argIndex, argSymbol] : llvm::enumerate(mapSyms)) {
+ if (auto *details =
+ argSymbol->detailsIf<Fortran::semantics::CommonBlockDetails>()) {
+ for (auto obj : details->objects()) {
+ auto targetCBMemberBind = Fortran::lower::genCommonBlockMember(
+ converter, currentLocation, *obj, region.getArgument(argIndex));
+ fir::ExtendedValue sexv = converter.getSymbolExtendedValue(*obj);
+ fir::ExtendedValue targetCBExv =
+ getExtendedValue(sexv, targetCBMemberBind);
+ converter.bindSymbol(*obj, targetCBExv);
+ }
+ }
+ }
+}
+
// This functions creates a block for the body of the targetOp's region. It adds
// all the symbols present in mapSymbols as block arguments to this block.
static void
@@ -983,6 +1010,18 @@ genBodyOfTargetOp(Fortran::lower::AbstractConverter &converter,
// Create the insertion point after the marker.
firOpBuilder.setInsertionPointAfter(undefMarker.getDefiningOp());
+
+ // If we map a common block using it's symbol e.g. map(tofrom: /common_block/)
+ // and accessing it's members within the target region, there is a large
+ // chance we will end up with uses external to the region accessing the common
+ // block. As target regions are IsolatedFromAbove, we must make sure to
+ // resolve these, we do so by generating new common block member accesses
+ // within the region, binding them to the member symbol for the scope of the
+ // region so that subsequent code generation within the region will utilise
+ // our new member accesses we have created.
+ genIntermediateCommonBlockAccessors(converter, currentLocation, region,
+ mapSyms);
+
if (genNested)
genNestedEvaluations(converter, eval);
}
@@ -1574,6 +1613,13 @@ genTargetOp(Fortran::lower::AbstractConverter &converter,
// symbols used inside the region that have not been explicitly mapped using
// the map clause.
auto captureImplicitMap = [&](const Fortran::semantics::Symbol &sym) {
+ // if the symbol is part of an already mapped common block, do not make a
+ // map for it.
+ if (const Fortran::semantics::Symbol *common =
+ Fortran::semantics::FindCommonBlockContaining(sym.GetUltimate()))
+ if (llvm::find(mapSyms, common) != mapSyms.end())
+ return;
+
if (llvm::find(mapSyms, &sym) == mapSyms.end()) {
mlir::Value baseOp = converter.getSymbolAddress(sym);
if (!baseOp)
diff --git a/flang/test/Integration/OpenMP/map-types-and-sizes.f90 b/flang/test/Integration/OpenMP/map-types-and-sizes.f90
index f3a20690f05a9..591be0b680a51 100644
--- a/flang/test/Integration/OpenMP/map-types-and-sizes.f90
+++ b/flang/test/Integration/OpenMP/map-types-and-sizes.f90
@@ -231,6 +231,31 @@ subroutine mapType_char
!$omp end target
end subroutine mapType_char
+!CHECK: @.offload_sizes{{.*}} = private unnamed_addr constant [1 x i64] [i64 8]
+!CHECK: @.offload_maptypes{{.*}} = private unnamed_addr constant [1 x i64] [i64 35]
+subroutine mapType_common_block
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: /var_common/)
+ var1 = var1 + 20
+ var2 = var2 + 30
+!$omp end target
+end subroutine mapType_common_block
+
+!CHECK: @.offload_sizes{{.*}} = private unnamed_addr constant [2 x i64] [i64 4, i64 4]
+!CHECK: @.offload_maptypes{{.*}} = private unnamed_addr constant [2 x i64] [i64 35, i64 35]
+subroutine mapType_common_block_members
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+
+!$omp target map(tofrom: var1, var2)
+ var2 = var1
+!$omp end target
+end subroutine mapType_common_block_members
+
+
!CHECK-LABEL: define {{.*}} @{{.*}}maptype_ptr_explicit_{{.*}}
!CHECK: %[[ALLOCA:.*]] = alloca { ptr, i64, i32, i8, i8, i8, i8 }, i64 1, align 8
!CHECK: %[[ALLOCA_GEP:.*]] = getelementptr { ptr, i64, i32, i8, i8, i8, i8 }, ptr %[[ALLOCA]], i32 1
@@ -346,3 +371,19 @@ end subroutine mapType_char
!CHECK: store ptr %[[ALLOCA]], ptr %[[BASE_PTR_ARR]], align 8
!CHECK: %[[OFFLOAD_PTR_ARR:.*]] = getelementptr inbounds [1 x ptr], ptr %.offload_ptrs, i32 0, i32 0
!CHECK: store ptr %[[ARR_OFF]], ptr %[[OFFLOAD_PTR_ARR]], align 8
+
+!CHECK-LABEL: define {{.*}} @{{.*}}maptype_common_block_{{.*}}
+!CHECK: %[[BASE_PTR_ARR:.*]] = getelementptr inbounds [1 x ptr], ptr %.offload_baseptrs, i32 0, i32 0
+!CHECK: store ptr @var_common_, ptr %[[BASE_PTR_ARR]], align 8
+!CHECK: %[[OFFLOAD_PTR_ARR:.*]] = getelementptr inbounds [1 x ptr], ptr %.offload_ptrs, i32 0, i32 0
+!CHECK: store ptr @var_common_, ptr %[[OFFLOAD_PTR_ARR]], align 8
+
+!CHECK-LABEL: define {{.*}} @{{.*}}maptype_common_block_members_{{.*}}
+!CHECK: %[[BASE_PTR_ARR:.*]] = getelementptr inbounds [2 x ptr], ptr %.offload_baseptrs, i32 0, i32 0
+!CHECK: store ptr @var_common_, ptr %[[BASE_PTR_ARR]], align 8
+!CHECK: %[[OFFLOAD_PTR_ARR:.*]] = getelementptr inbounds [2 x ptr], ptr %.offload_ptrs, i32 0, i32 0
+!CHECK: store ptr @var_common_, ptr %[[OFFLOAD_PTR_ARR]], align 8
+!CHECK: %[[BASE_PTR_ARR_1:.*]] = getelementptr inbounds [2 x ptr], ptr %.offload_baseptrs, i32 0, i32 1
+!CHECK: store ptr getelementptr (i8, ptr @var_common_, i64 4), ptr %[[BASE_PTR_ARR_1]], align 8
+!CHECK: %[[OFFLOAD_PTR_ARR_1:.*]] = getelementptr inbounds [2 x ptr], ptr %.offload_ptrs, i32 0, i32 1
+!CHECK: store ptr getelementptr (i8, ptr @var_common_, i64 4), ptr %[[OFFLOAD_PTR_ARR_1]], align 8
diff --git a/flang/test/Lower/OpenMP/common-block-map.f90 b/flang/test/Lower/OpenMP/common-block-map.f90
new file mode 100644
index 0000000000000..56a5e775b1b65
--- /dev/null
+++ b/flang/test/Lower/OpenMP/common-block-map.f90
@@ -0,0 +1,83 @@
+!RUN: %flang_fc1 -emit-hlfir -fopenmp %s -o - | FileCheck %s
+
+!CHECK: fir.global common @var_common_(dense<0> : vector<8xi8>) : !fir.array<8xi8>
+!CHECK: fir.global common @var_common_link_(dense<0> : vector<8xi8>) {omp.declare_target = #omp.declaretarget<device_type = (any), capture_clause = (link)>} : !fir.array<8xi8>
+
+!CHECK-LABEL: func.func @_QPmap_full_block
+!CHECK: %[[CB_ADDR:.*]] = fir.address_of(@var_common_) : !fir.ref<!fir.array<8xi8>>
+!CHECK: %[[MAP:.*]] = omp.map.info var_ptr(%[[CB_ADDR]] : !fir.ref<!fir.array<8xi8>>, !fir.array<8xi8>) map_clauses(tofrom) capture(ByRef) -> !fir.ref<!fir.array<8xi8>> {name = "var_common"}
+!CHECK: omp.target map_entries(%[[MAP]] -> %[[MAP_ARG:.*]] : !fir.ref<!fir.array<8xi8>>) {
+!CHECK: ^bb0(%[[MAP_ARG]]: !fir.ref<!fir.array<8xi8>>):
+!CHECK: %[[CONV:.*]] = fir.convert %[[MAP_ARG]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[INDEX:.*]] = arith.constant 0 : index
+!CHECK: %[[COORD:.*]] = fir.coordinate_of %[[CONV]], %[[INDEX]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[CONV2:.*]] = fir.convert %[[COORD]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[CB_MEMBER_1:.*]]:2 = hlfir.declare %[[CONV2]] {uniq_name = "_QFmap_full_blockEvar1"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+!CHECK: %[[CONV3:.*]] = fir.convert %[[MAP_ARG]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[INDEX2:.*]] = arith.constant 4 : index
+!CHECK: %[[COORD2:.*]] = fir.coordinate_of %[[CONV3]], %[[INDEX2]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[CONV4:.*]] = fir.convert %[[COORD2]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[CB_MEMBER_2:.*]]:2 = hlfir.declare %[[CONV4]] {uniq_name = "_QFmap_full_blockEvar2"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+subroutine map_full_block
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: /var_common/)
+ var1 = var1 + 20
+ var2 = var2 + 30
+!$omp end target
+end
+
+!CHECK-LABEL: @_QPmap_mix_of_members
+!CHECK: %[[COMMON_BLOCK:.*]] = fir.address_of(@var_common_) : !fir.ref<!fir.array<8xi8>>
+!CHECK: %[[CB_CONV:.*]] = fir.convert %[[COMMON_BLOCK]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[INDEX:.*]] = arith.constant 0 : index
+!CHECK: %[[COORD:.*]] = fir.coordinate_of %[[CB_CONV]], %[[INDEX]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[CONV:.*]] = fir.convert %[[COORD]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[CB_MEMBER_1:.*]]:2 = hlfir.declare %[[CONV]] {uniq_name = "_QFmap_mix_of_membersEvar1"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+!CHECK: %[[CB_CONV:.*]] = fir.convert %[[COMMON_BLOCK]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[INDEX:.*]] = arith.constant 4 : index
+!CHECK: %[[COORD:.*]] = fir.coordinate_of %[[CB_CONV]], %[[INDEX]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[CONV:.*]] = fir.convert %[[COORD]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[CB_MEMBER_2:.*]]:2 = hlfir.declare %[[CONV]] {uniq_name = "_QFmap_mix_of_membersEvar2"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+!CHECK: %[[MAP_EXP:.*]] = omp.map.info var_ptr(%[[CB_MEMBER_2]]#0 : !fir.ref<i32>, i32) map_clauses(tofrom) capture(ByRef) -> !fir.ref<i32> {name = "var2"}
+!CHECK: %[[MAP_IMP:.*]] = omp.map.info var_ptr(%[[CB_MEMBER_1]]#1 : !fir.ref<i32>, i32) map_clauses(implicit, exit_release_or_enter_alloc) capture(ByCopy) -> !fir.ref<i32> {name = "var1"}
+!CHECK: omp.target map_entries(%[[MAP_EXP]] -> %[[ARG_EXP:.*]], %[[MAP_IMP]] -> %[[ARG_IMP:.*]] : !fir.ref<i32>, !fir.ref<i32>) {
+!CHECK: ^bb0(%[[ARG_EXP]]: !fir.ref<i32>, %[[ARG_IMP]]: !fir.ref<i32>):
+!CHECK: %[[EXP_MEMBER:.*]]:2 = hlfir.declare %[[ARG_EXP]] {uniq_name = "_QFmap_mix_of_membersEvar2"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+!CHECK: %[[IMP_MEMBER:.*]]:2 = hlfir.declare %[[ARG_IMP]] {uniq_name = "_QFmap_mix_of_membersEvar1"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+subroutine map_mix_of_members
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+
+!$omp target map(tofrom: var2)
+ var2 = var1
+!$omp end target
+end
+
+!CHECK-LABEL: @_QQmain
+!CHECK: %[[DECL_TAR_CB:.*]] = fir.address_of(@var_common_link_) : !fir.ref<!fir.array<8xi8>>
+!CHECK: %[[MAP_DECL_TAR_CB:.*]] = omp.map.info var_ptr(%[[DECL_TAR_CB]] : !fir.ref<!fir.array<8xi8>>, !fir.array<8xi8>) map_clauses(tofrom) capture(ByRef) -> !fir.ref<!fir.array<8xi8>> {name = "var_common_link"}
+!CHECK: omp.target map_entries(%[[MAP_DECL_TAR_CB]] -> %[[MAP_DECL_TAR_ARG:.*]] : !fir.ref<!fir.array<8xi8>>) {
+!CHECK: ^bb0(%[[MAP_DECL_TAR_ARG]]: !fir.ref<!fir.array<8xi8>>):
+!CHECK: %[[CONV:.*]] = fir.convert %[[MAP_DECL_TAR_ARG]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[INDEX:.*]] = arith.constant 0 : index
+!CHECK: %[[COORD:.*]] = fir.coordinate_of %[[CONV]], %[[INDEX]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[CONV:.*]] = fir.convert %[[COORD]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[MEMBER_ONE:.*]]:2 = hlfir.declare %[[CONV]] {uniq_name = "_QFElink1"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+!CHECK: %[[CONV:.*]] = fir.convert %[[MAP_DECL_TAR_ARG]] : (!fir.ref<!fir.array<8xi8>>) -> !fir.ref<!fir.array<?xi8>>
+!CHECK: %[[INDEX:.*]] = arith.constant 4 : index
+!CHECK: %[[COORD:.*]] = fir.coordinate_of %[[CONV]], %[[INDEX]] : (!fir.ref<!fir.array<?xi8>>, index) -> !fir.ref<i8>
+!CHECK: %[[CONV:.*]] = fir.convert %[[COORD]] : (!fir.ref<i8>) -> !fir.ref<i32>
+!CHECK: %[[MEMBER_TWO:.*]]:2 = hlfir.declare %[[CONV]] {uniq_name = "_QFElink2"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+program main
+ implicit none
+ common /var_common_link/ link1, link2
+ integer :: link1, link2
+ !$omp declare target link(/var_common_link/)
+
+!$omp target map(tofrom: /var_common_link/)
+ link1 = link2 + 20
+!$omp end target
+end program
diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
index c9ee0c25194c2..0f078b03457b0 100644
--- a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
@@ -2110,6 +2110,11 @@ class OpenMPIRBuilder {
int32_t UB);
///}
+ /// Replaces constant values with instruction equivelants where possible
+ /// inside of a function.
+ static void replaceConstantValueUsesInFuncWithInstr(llvm::Value *Input,
+ Function *Func);
+
private:
// Sets the function attributes expected for the outlined function
void setOutlinedTargetRegionFunctionAttributes(Function *OutlinedFn);
diff --git a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
index 42ea20919a5e1..4d803e65ae542 100644
--- a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
+++ b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
@@ -5085,8 +5085,8 @@ static void replaceConstatExprUsesInFuncWithInstr(ConstantExpr *ConstExpr,
}
}
-static void replaceConstantValueUsesInFuncWithInstr(llvm::Value *Input,
- Function *Func) {
+void OpenMPIRBuilder::replaceConstantValueUsesInFuncWithInstr(
+ llvm::Value *Input, Function *Func) {
for (User *User : make_early_inc_range(Input->users()))
if (auto *Const = dyn_cast<Constant>(User))
if (auto *ConstExpr = dyn_cast<ConstantExpr>(Const))
@@ -5160,6 +5160,31 @@ static Function *createOutlinedFunction(
? make_range(Func->arg_begin() + 1, Func->arg_end())
: Func->args();
+ auto ReplaceValue = [&OMPBuilder](Value *Input, Value *InputCopy,
+ Function *Func) {
+ // Things like GEP's can come in the form of Constants. Constants and
+ // ConstantExpr's do not have access to the knowledge of what they're
+ // contained in, so we must dig a little to find an instruction so we
+ // can tell if they're used inside of the function we're outlining. We
+ // also replace the original constant expression with a new instruction
+ // equivalent; an instruction as it allows easy modification in the
+ // following loop, as we can now know the constant (instruction) is
+ // owned by our target function and replaceUsesOfWith can now be invoked
+ // on it (cannot do this with constants it seems). A brand new one also
+ // allows us to be cautious as it is perhaps possible the old expression
+ // was used inside of the function but exists and is used externally
+ // (unlikely by the nature of a Constant, but still).
+ OMPBuilder.replaceConstantValueUsesInFuncWithInstr(Input, Func);
+
+ // Collect all the instructions
+ for (User *User : make_early_inc_range(Input->users()))
+ if (auto *Instr = dyn_cast<Instruction>(User))
+ if (Instr->getFunction() == Func)
+ Instr->replaceUsesOfWith(Input, InputCopy);
+ };
+
+ SmallVector<std::pair<Value *, Value *>> DeferredReplacement;
+
// Rewrite uses of input valus to parameters.
for (auto InArg : zip(Inputs, ArgRange)) {
Value *Input = std::get<0>(InArg);
@@ -5169,27 +5194,36 @@ static Function *createOutlinedFunction(
Builder.restoreIP(
ArgAccessorFuncCB(Arg, Input, InputCopy, AllocaIP, Builder.saveIP()));
- // Things like GEP's can come in the form of Constants. Constants and
- // ConstantExpr's do not have access to the knowledge of what they're
- // contained in, so we must dig a little to find an instruction so we can
- // tell if they're used inside of the function we're outlining. We also
- // replace the original constant expression with a new instruction
- // equivalent; an instruction as it allows easy modification in the
- // following loop, as we can now know the constant (instruction) is owned by
- // our target function and replaceUsesOfWith can now be invoked on it
- // (cannot do this with constants it seems). A brand new one also allows us
- // to be cautious as it is perhaps possible the old expression was used
- // inside of the function but exists and is used externally (unlikely by the
- // nature of a Constant, but still).
- replaceConstantValueUsesInFuncWithInstr(Input, Func);
+ // In certain cases a Global may be set up for replacement, however, this
+ // Global may be used in multiple arguments to the kernel, just segmented
+ // apart, for example, if we have a global array, that is sectioned into
+ // multiple mappings (technically not legal in OpenMP, but there is a case
+ // in Fortran for Common Blocks where this is neccesary), we will end up
+ // with GEP's into this array inside the kernel, that refer to the Global
+ // but are technically seperate arguments to the kernel for all intents and
+ // purposes. If we have mapped a segment that requires a GEP into the 0-th
+ // index, it will fold into an referal to the Global, if we then encounter
+ // this folded GEP during replacement all of the references to the
+ // Global in the kernel will be replaced with the argument we have generated
+ // that corresponds to it, including any other GEP's that refer to the
+ // Global that may be other arguments. This will invalidate all of the other
+ // preceding mapped arguments that refer to the same global that may be
+ // seperate segments. To prevent this, we defer global processing until all
+ // other processing has been performed.
+ if (llvm::isa<llvm::GlobalValue>(std::get<0>(InArg)) ||
+ llvm::isa<llvm::GlobalObject>(std::get<0>(InArg)) ||
+ llvm::isa<llvm::GlobalVariable>(std::get<0>(InArg))) {
+ DeferredReplacement.push_back(std::make_pair(Input, InputCopy));
+ continue;
+ }
- // Collect all the instructions
- for (User *User : make_early_inc_range(Input->users()))
- if (auto *Instr = dyn_cast<Instruction>(User))
- if (Instr->getFunction() == Func)
- Instr->replaceUsesOfWith(Input, InputCopy);
+ ReplaceValue(Input, InputCopy, Func);
}
+ // Replace all of our deferred Input values, currently just Globals.
+ for (auto Deferred : DeferredReplacement)
+ ReplaceValue(std::get<0>(Deferred), std::get<1>(Deferred), Func);
+
// Restore insert point.
Builder.restoreIP(OldInsertPoint);
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 0c34126667324..d2b4d77cfdcb8 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2021,12 +2021,6 @@ llvm::Value *getSizeInBytes(DataLayout &dl, const mlir::Type &type,
Operation *clauseOp, llvm::Value *basePointer,
llvm::Type *baseType, llvm::IRBuilderBase &builder,
LLVM::ModuleTranslation &moduleTranslation) {
- // utilising getTypeSizeInBits instead of getTypeSize as getTypeSize gives
- // the size in inconsistent byte or bit format.
- uint64_t underlyingTypeSzInBits = dl.getTypeSizeInBits(type);
- if (auto arrTy = llvm::dyn_cast_if_present<LLVM::LLVMArrayType>(type))
- underlyingTypeSzInBits = getArrayElementSizeInBits(arrTy, dl);
-
if (auto memberClause =
mlir::dyn_cast_if_present<mlir::omp::MapInfoOp>(clauseOp)) {
// This calculates the size to transfer based on bounds and the underlying
@@ -2052,6 +2046,12 @@ llvm::Value *getSizeInBytes(DataLayout &dl, const mlir::Type &type,
}
}
+ // utilising getTypeSizeInBits instead of getTypeSize as getTypeSize gives
+ // the size in inconsistent byte or bit format.
+ uint64_t underlyingTypeSzInBits = dl.getTypeSizeInBits(type);
+ if (auto arrTy = llvm::dyn_cast_if_present<LLVM::LLVMArrayType>(type))
+ underlyingTypeSzInBits = getArrayElementSizeInBits(arrTy, dl);
+
// The size in bytes x number of elements, the sizeInBytes stored is
// the underyling types size, e.g. if ptr<i32>, it'll be the i32's
// size, so we do some on the fly runtime math to get the size in
@@ -2062,7 +2062,7 @@ llvm::Value *getSizeInBytes(DataLayout &dl, const mlir::Type &type,
}
}
- return builder.getInt64(underlyingTypeSzInBits / 8);
+ return builder.getInt64(dl.getTypeSizeInBits(type) / 8);
}
void collectMapDataFromMapOperands(MapInfoData &mapData,
@@ -2948,7 +2948,7 @@ static bool targetOpSupported(Operation &opInst) {
static void
handleDeclareTargetMapVar(MapInfoData &mapData,
LLVM::ModuleTranslation &moduleTranslation,
- llvm::IRBuilderBase &builder) {
+ llvm::IRBuilderBase &builder, llvm::Function *func) {
for (size_t i = 0; i < mapData.MapClause.size(); ++i) {
// In the case of declare target mapped variables, the basePointer is
// the reference pointer generated by the convertDeclareTargetAttr
@@ -2963,19 +2963,27 @@ handleDeclareTargetMapVar(MapInfoData &mapData,
// reference pointer and the pointer are assigned in the kernel argument
// structure for the host.
if (mapData.IsDeclareTarget[i]) {
+
+ moduleTranslation.getOpenMPBuilder()
+ ->replaceConstantValueUsesInFuncWithInstr(mapData.OriginalValue[i],
+ func);
+
// The users iterator will get invalidated if we modify an element,
- // so we populate this vector of uses to alter each user on an individual
- // basis to emit its own load (rather than one load for all).
+ // so we populate this vector of uses to alter each user on an
+ // individual basis to emit its own load (rather than one load for
+ // all).
llvm::SmallVector<llvm::User *> userVec;
for (llvm::User *user : mapData.OriginalValue[i]->users())
userVec.push_back(user);
for (llvm::User *user : userVec) {
if (auto *insn = dyn_cast<llvm::Instruction>(user)) {
- auto *load = builder.CreateLoad(mapData.BasePointers[i]->getType(),
- mapData.BasePointers[i]);
- load->moveBefore(insn);
- user->replaceUsesOfWith(mapData.OriginalValue[i], load);
+ if (insn->getFunction() == func) {
+ auto *load = builder.CreateLoad(mapData.BasePointers[i]->getType(),
+ mapData.BasePointers[i]);
+ load->moveBefore(insn);
+ user->replaceUsesOfWith(mapData.OriginalValue[i], load);
+ }
}
}
}
@@ -3093,6 +3101,7 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
auto &targetRegion = targetOp.getRegion();
DataLayout dl = DataLayout(opInst.getParentOfType<ModuleOp>());
SmallVector<Value> mapOperands = targetOp.getMapOperands();
+ llvm::Function *llvmOutlinedFn = nullptr;
LogicalResult bodyGenStatus = success();
using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
@@ -3102,7 +3111,7 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
// original function to the new outlined function.
llvm::Function *llvmParentFn =
moduleTranslation.lookupFunction(parentFn.getName());
- llvm::Function *llvmOutlinedFn = codeGenIP.getBlock()->getParent();
+ llvmOutlinedFn = codeGenIP.getBlock()->getParent();
assert(llvmParentFn && llvmOutlinedFn &&
"Both parent and outlined functions must exist at this point");
@@ -3197,7 +3206,8 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
// Remap access operations to declare target reference pointers for the
// device, essentially generating extra loadop's as necessary
if (moduleTranslation.getOpenMPBuilder()->Config.isTargetDevice())
- handleDeclareTargetMapVar(mapData, moduleTranslation, builder);
+ handleDeclareTargetMapVar(mapData, moduleTranslation, builder,
+ llvmOutlinedFn);
return bodyGenStatus;
}
diff --git a/offload/test/offloading/fortran/target-map-all-common-block-members.f90 b/offload/test/offloading/fortran/target-map-all-common-block-members.f90
new file mode 100644
index 0000000000000..def1e7c663073
--- /dev/null
+++ b/offload/test/offloading/fortran/target-map-all-common-block-members.f90
@@ -0,0 +1,55 @@
+! Offloading test checking interaction of
+! mapping all the members of a common block
+! to a target region
+! REQUIRES: flang, amdgcn-amd-amdhsa
+! UNSUPPORTED: nvptx64-nvidia-cuda
+! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
+! UNSUPPORTED: aarch64-unknown-linux-gnu
+! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
+! UNSUPPORTED: x86_64-pc-linux-gnu
+! UNSUPPORTED: x86_64-pc-linux-gnu-LTO
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ implicit none
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+
+ call modify_1
+
+ !$omp target map(tofrom: var1, var2, var3)
+ var3 = var3 * 10
+ var2 = var2 * 10
+ var1 = var1 * 10
+ !$omp end target
+
+ call modify_2
+
+ print *, var1
+ print *, var2
+ print *, var3
+end program
+
+subroutine modify_1
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+!$omp target map(tofrom: var2, var1, var3)
+ var3 = var3 + 40
+ var2 = var2 + 20
+ var1 = var1 + 30
+!$omp end target
+end
+
+subroutine modify_2
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+!$omp target map(tofrom: var2, var3, var1)
+ var3 = var3 + 20
+ var1 = var1 + 10
+ var2 = var2 + 15
+!$omp end target
+end
+
+!CHECK: 310
+!CHECK: 215
+!CHECK: 420
diff --git a/offload/test/offloading/fortran/target-map-common-block.f90 b/offload/test/offloading/fortran/target-map-common-block.f90
new file mode 100644
index 0000000000000..8fcd504950515
--- /dev/null
+++ b/offload/test/offloading/fortran/target-map-common-block.f90
@@ -0,0 +1,50 @@
+! Offloading test checking interaction of
+! mapping a full common block in a target
+! region
+! REQUIRES: flang, amdgcn-amd-amdhsa
+! UNSUPPORTED: nvptx64-nvidia-cuda
+! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
+! UNSUPPORTED: aarch64-unknown-linux-gnu
+! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
+! UNSUPPORTED: x86_64-pc-linux-gnu
+! UNSUPPORTED: x86_64-pc-linux-gnu-LTO
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+
+ call modify_1
+
+ !$omp target map(tofrom: /var_common/)
+ var1 = var1 + 20
+ var2 = var2 + 50
+ !$omp end target
+
+ call modify_2
+
+ print *, var1
+ print *, var2
+end program
+
+subroutine modify_1
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: /var_common/)
+ var1 = var1 + 20
+ var2 = var2 + 30
+!$omp end target
+end
+
+subroutine modify_2
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: /var_common/)
+ var1 = var1 * 10
+ var2 = var2 * 10
+!$omp end target
+end
+
+!CHECK: 400
+!CHECK: 800
diff --git a/offload/test/offloading/fortran/target-map-declare-target-link-common-block.f90 b/offload/test/offloading/fortran/target-map-declare-target-link-common-block.f90
new file mode 100644
index 0000000000000..758f7e8970246
--- /dev/null
+++ b/offload/test/offloading/fortran/target-map-declare-target-link-common-block.f90
@@ -0,0 +1,95 @@
+! Offloading test checking interaction of
+! mapping a declare target link common
+! block with device_type any to a target
+! region
+! REQUIRES: flang, amdgcn-amd-amdhsa
+! UNSUPPORTED: nvptx64-nvidia-cuda
+! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
+! UNSUPPORTED: aarch64-unknown-linux-gnu
+! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
+! UNSUPPORTED: x86_64-pc-linux-gnu
+! UNSUPPORTED: x86_64-pc-linux-gnu-LTO
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ implicit none
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+ !$omp declare target link(/var_common/)
+
+ call modify_1
+
+ !$omp target map(tofrom: var2)
+ var2 = var2 + var3
+ !$omp end target
+
+ call modify_2
+
+ print *, var1
+ print *, var2
+ print *, var3
+
+ call modify_3
+
+ if (var1 /= 20) then
+ print*, "======= FORTRAN Test Failed! ======="
+ stop 1
+ end if
+
+ if (var2 /= 100) then
+ print*, "======= FORTRAN Test Failed! ======="
+ stop 1
+ end if
+
+ if (var3 /= 60) then
+ print*, "======= FORTRAN Test Failed! ======="
+ stop 1
+ end if
+
+ print*, "======= FORTRAN Test Passed! ======="
+end program
+
+subroutine modify_1
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+
+!$omp target map(tofrom: /var_common/)
+ var1 = 10
+ var2 = 20
+ var3 = 30
+!$omp end target
+
+end
+
+subroutine modify_2
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+ integer :: copy
+
+!$omp target map(tofrom: copy)
+ copy = var2 + var3
+!$omp end target
+
+ print *, copy
+
+ if (copy /= 80) then
+ print*, "======= FORTRAN Test Failed! ======="
+ stop 1
+ end if
+end
+
+subroutine modify_3
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+
+!$omp target map(tofrom: /var_common/)
+ var1 = var1 + var1
+ var2 = var2 + var2
+ var3 = var3 + var3
+!$omp end target
+end
+
+!CHECK: 80
+!CHECK: 20
+!CHECK: 100
+!CHECK: 60
\ No newline at end of file
diff --git a/offload/test/offloading/fortran/target-map-first-common-block-member.f90 b/offload/test/offloading/fortran/target-map-first-common-block-member.f90
new file mode 100644
index 0000000000000..69c1fa13501d2
--- /dev/null
+++ b/offload/test/offloading/fortran/target-map-first-common-block-member.f90
@@ -0,0 +1,47 @@
+! Offloading test checking interaction of
+! mapping a member of a common block to a
+! target region
+! REQUIRES: flang, amdgcn-amd-amdhsa
+! UNSUPPORTED: nvptx64-nvidia-cuda
+! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
+! UNSUPPORTED: aarch64-unknown-linux-gnu
+! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
+! UNSUPPORTED: x86_64-pc-linux-gnu
+! UNSUPPORTED: x86_64-pc-linux-gnu-LTO
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+
+ call modify_1
+
+ !$omp target map(tofrom: var1)
+ var1 = var1 + 20
+ !$omp end target
+
+ call modify_2
+
+ print *, var1
+ print *, var2
+end program
+
+subroutine modify_1
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: var1)
+ var1 = var1 + 20
+!$omp end target
+end
+
+subroutine modify_2
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: var1)
+ var1 = var1 * 10
+!$omp end target
+end
+
+!CHECK: 400
+!CHECK: 0
\ No newline at end of file
diff --git a/offload/test/offloading/fortran/target-map-mix-imp-exp-common-block-members.f90 b/offload/test/offloading/fortran/target-map-mix-imp-exp-common-block-members.f90
new file mode 100644
index 0000000000000..672630aec7d7c
--- /dev/null
+++ b/offload/test/offloading/fortran/target-map-mix-imp-exp-common-block-members.f90
@@ -0,0 +1,58 @@
+! Offloading test checking interaction of
+! mapping all the members of a common block
+! with a mix of explicit and implicit
+! mapping to a target region
+! REQUIRES: flang, amdgcn-amd-amdhsa
+! UNSUPPORTED: nvptx64-nvidia-cuda
+! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
+! UNSUPPORTED: aarch64-unknown-linux-gnu
+! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
+! UNSUPPORTED: x86_64-pc-linux-gnu
+! UNSUPPORTED: x86_64-pc-linux-gnu-LTO
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ implicit none
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+
+ call modify_1
+
+ !$omp target map(tofrom: var2)
+ var2 = var2 + var3
+ !$omp end target
+
+ call modify_2
+
+ print *, var1
+ print *, var2
+ print *, var3
+end program
+
+subroutine modify_1
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+
+ !$omp target map(tofrom: /var_common/)
+ var1 = 10
+ var2 = 20
+ var3 = 30
+ !$omp end target
+end
+
+subroutine modify_2
+ common /var_common/ var1, var2, var3
+ integer :: var1, var2, var3
+ integer :: copy
+
+ !$omp target map(tofrom: copy)
+ copy = var2 + var3
+ !$omp end target
+
+ print *, copy
+end
+
+!CHECK: 80
+!CHECK: 10
+!CHECK: 50
+!CHECK: 30
diff --git a/offload/test/offloading/fortran/target-map-second-common-block-member.f90 b/offload/test/offloading/fortran/target-map-second-common-block-member.f90
new file mode 100644
index 0000000000000..3aa937eb9adc6
--- /dev/null
+++ b/offload/test/offloading/fortran/target-map-second-common-block-member.f90
@@ -0,0 +1,47 @@
+! Offloading test checking interaction of
+! mapping a member of a common block to a
+! target region
+! REQUIRES: flang, amdgcn-amd-amdhsa
+! UNSUPPORTED: nvptx64-nvidia-cuda
+! UNSUPPORTED: nvptx64-nvidia-cuda-LTO
+! UNSUPPORTED: aarch64-unknown-linux-gnu
+! UNSUPPORTED: aarch64-unknown-linux-gnu-LTO
+! UNSUPPORTED: x86_64-pc-linux-gnu
+! UNSUPPORTED: x86_64-pc-linux-gnu-LTO
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ implicit none
+ common /var_common/ var1, var2
+ integer :: var1, var2
+
+ call modify_1
+
+ !$omp target map(tofrom: var2)
+ var2 = var2 + 20
+ !$omp end target
+
+ call modify_2
+
+ print *, var1
+ print *, var2
+end program
+
+subroutine modify_1
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: var2)
+ var2 = var2 + 20
+!$omp end target
+end
+
+subroutine modify_2
+ common /var_common/ var1, var2
+ integer :: var1, var2
+!$omp target map(tofrom: var2)
+ var2 = var2 * 10
+!$omp end target
+end
+
+!CHECK: 0
+!CHECK: 400
More information about the flang-commits
mailing list