blob: d9ef6c2a1078945f4472b465669f51043b13e167 [file] [log] [blame]
//===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This provides a generalized class for OpenMP runtime code generation
// specialized by GPU targets NVPTX and AMDGCN.
//
//===----------------------------------------------------------------------===//
#include "CGOpenMPRuntimeGPU.h"
#include "CGOpenMPRuntimeNVPTX.h"
#include "CodeGenFunction.h"
#include "clang/AST/Attr.h"
#include "clang/AST/DeclOpenMP.h"
#include "clang/AST/StmtOpenMP.h"
#include "clang/AST/StmtVisitor.h"
#include "clang/Basic/Cuda.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/Frontend/OpenMP/OMPGridValues.h"
#include "llvm/IR/IntrinsicsNVPTX.h"
using namespace clang;
using namespace CodeGen;
using namespace llvm::omp;
namespace {
enum OpenMPRTLFunctionNVPTX {
/// Call to void __kmpc_kernel_init(kmp_int32 thread_limit,
/// int16_t RequiresOMPRuntime);
OMPRTL_NVPTX__kmpc_kernel_init,
/// Call to void __kmpc_kernel_deinit(int16_t IsOMPRuntimeInitialized);
OMPRTL_NVPTX__kmpc_kernel_deinit,
/// Call to void __kmpc_spmd_kernel_init(kmp_int32 thread_limit,
/// int16_t RequiresOMPRuntime, int16_t RequiresDataSharing);
OMPRTL_NVPTX__kmpc_spmd_kernel_init,
/// Call to void __kmpc_spmd_kernel_deinit_v2(int16_t RequiresOMPRuntime);
OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2,
/// Call to void __kmpc_kernel_prepare_parallel(void
/// *outlined_function);
OMPRTL_NVPTX__kmpc_kernel_prepare_parallel,
/// Call to bool __kmpc_kernel_parallel(void **outlined_function);
OMPRTL_NVPTX__kmpc_kernel_parallel,
/// Call to void __kmpc_kernel_end_parallel();
OMPRTL_NVPTX__kmpc_kernel_end_parallel,
/// Call to void __kmpc_serialized_parallel(ident_t *loc, kmp_int32
/// global_tid);
OMPRTL_NVPTX__kmpc_serialized_parallel,
/// Call to void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32
/// global_tid);
OMPRTL_NVPTX__kmpc_end_serialized_parallel,
/// Call to int32_t __kmpc_shuffle_int32(int32_t element,
/// int16_t lane_offset, int16_t warp_size);
OMPRTL_NVPTX__kmpc_shuffle_int32,
/// Call to int64_t __kmpc_shuffle_int64(int64_t element,
/// int16_t lane_offset, int16_t warp_size);
OMPRTL_NVPTX__kmpc_shuffle_int64,
/// Call to __kmpc_nvptx_parallel_reduce_nowait_v2(ident_t *loc, kmp_int32
/// global_tid, kmp_int32 num_vars, size_t reduce_size, void* reduce_data,
/// void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t
/// lane_offset, int16_t shortCircuit),
/// void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num));
OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2,
/// Call to __kmpc_nvptx_teams_reduce_nowait_v2(ident_t *loc, kmp_int32
/// global_tid, void *global_buffer, int32_t num_of_records, void*
/// reduce_data,
/// void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t
/// lane_offset, int16_t shortCircuit),
/// void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num), void
/// (*kmp_ListToGlobalCpyFctPtr)(void *buffer, int idx, void *reduce_data),
/// void (*kmp_GlobalToListCpyFctPtr)(void *buffer, int idx,
/// void *reduce_data), void (*kmp_GlobalToListCpyPtrsFctPtr)(void *buffer,
/// int idx, void *reduce_data), void (*kmp_GlobalToListRedFctPtr)(void
/// *buffer, int idx, void *reduce_data));
OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2,
/// Call to __kmpc_nvptx_end_reduce_nowait(int32_t global_tid);
OMPRTL_NVPTX__kmpc_end_reduce_nowait,
/// Call to void __kmpc_data_sharing_init_stack();
OMPRTL_NVPTX__kmpc_data_sharing_init_stack,
/// Call to void __kmpc_data_sharing_init_stack_spmd();
OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd,
/// Call to void* __kmpc_data_sharing_coalesced_push_stack(size_t size,
/// int16_t UseSharedMemory);
OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack,
/// Call to void* __kmpc_data_sharing_push_stack(size_t size, int16_t
/// UseSharedMemory);
OMPRTL_NVPTX__kmpc_data_sharing_push_stack,
/// Call to void __kmpc_data_sharing_pop_stack(void *a);
OMPRTL_NVPTX__kmpc_data_sharing_pop_stack,
/// Call to void __kmpc_begin_sharing_variables(void ***args,
/// size_t n_args);
OMPRTL_NVPTX__kmpc_begin_sharing_variables,
/// Call to void __kmpc_end_sharing_variables();
OMPRTL_NVPTX__kmpc_end_sharing_variables,
/// Call to void __kmpc_get_shared_variables(void ***GlobalArgs)
OMPRTL_NVPTX__kmpc_get_shared_variables,
/// Call to uint16_t __kmpc_parallel_level(ident_t *loc, kmp_int32
/// global_tid);
OMPRTL_NVPTX__kmpc_parallel_level,
/// Call to int8_t __kmpc_is_spmd_exec_mode();
OMPRTL_NVPTX__kmpc_is_spmd_exec_mode,
/// Call to void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode,
/// const void *buf, size_t size, int16_t is_shared, const void **res);
OMPRTL_NVPTX__kmpc_get_team_static_memory,
/// Call to void __kmpc_restore_team_static_memory(int16_t
/// isSPMDExecutionMode, int16_t is_shared);
OMPRTL_NVPTX__kmpc_restore_team_static_memory,
/// Call to void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
OMPRTL__kmpc_barrier,
/// Call to void __kmpc_barrier_simple_spmd(ident_t *loc, kmp_int32
/// global_tid);
OMPRTL__kmpc_barrier_simple_spmd,
/// Call to int32_t __kmpc_warp_active_thread_mask(void);
OMPRTL_NVPTX__kmpc_warp_active_thread_mask,
/// Call to void __kmpc_syncwarp(int32_t Mask);
OMPRTL_NVPTX__kmpc_syncwarp,
};
/// Pre(post)-action for different OpenMP constructs specialized for NVPTX.
class NVPTXActionTy final : public PrePostActionTy {
llvm::FunctionCallee EnterCallee = nullptr;
ArrayRef<llvm::Value *> EnterArgs;
llvm::FunctionCallee ExitCallee = nullptr;
ArrayRef<llvm::Value *> ExitArgs;
bool Conditional = false;
llvm::BasicBlock *ContBlock = nullptr;
public:
NVPTXActionTy(llvm::FunctionCallee EnterCallee,
ArrayRef<llvm::Value *> EnterArgs,
llvm::FunctionCallee ExitCallee,
ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false)
: EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
ExitArgs(ExitArgs), Conditional(Conditional) {}
void Enter(CodeGenFunction &CGF) override {
llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
if (Conditional) {
llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
ContBlock = CGF.createBasicBlock("omp_if.end");
// Generate the branch (If-stmt)
CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
CGF.EmitBlock(ThenBlock);
}
}
void Done(CodeGenFunction &CGF) {
// Emit the rest of blocks/branches
CGF.EmitBranch(ContBlock);
CGF.EmitBlock(ContBlock, true);
}
void Exit(CodeGenFunction &CGF) override {
CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
}
};
/// A class to track the execution mode when codegening directives within
/// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry
/// to the target region and used by containing directives such as 'parallel'
/// to emit optimized code.
class ExecutionRuntimeModesRAII {
private:
CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode =
CGOpenMPRuntimeGPU::EM_Unknown;
CGOpenMPRuntimeGPU::ExecutionMode &ExecMode;
bool SavedRuntimeMode = false;
bool *RuntimeMode = nullptr;
public:
/// Constructor for Non-SPMD mode.
ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode)
: ExecMode(ExecMode) {
SavedExecMode = ExecMode;
ExecMode = CGOpenMPRuntimeGPU::EM_NonSPMD;
}
/// Constructor for SPMD mode.
ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode,
bool &RuntimeMode, bool FullRuntimeMode)
: ExecMode(ExecMode), RuntimeMode(&RuntimeMode) {
SavedExecMode = ExecMode;
SavedRuntimeMode = RuntimeMode;
ExecMode = CGOpenMPRuntimeGPU::EM_SPMD;
RuntimeMode = FullRuntimeMode;
}
~ExecutionRuntimeModesRAII() {
ExecMode = SavedExecMode;
if (RuntimeMode)
*RuntimeMode = SavedRuntimeMode;
}
};
/// GPU Configuration: This information can be derived from cuda registers,
/// however, providing compile time constants helps generate more efficient
/// code. For all practical purposes this is fine because the configuration
/// is the same for all known NVPTX architectures.
enum MachineConfiguration : unsigned {
/// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target
/// specific Grid Values like GV_Warp_Size, GV_Warp_Size_Log2,
/// and GV_Warp_Size_Log2_Mask.
/// Global memory alignment for performance.
GlobalMemoryAlignment = 128,
/// Maximal size of the shared memory buffer.
SharedMemorySize = 128,
};
static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
RefExpr = RefExpr->IgnoreParens();
if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
Base = TempASE->getBase()->IgnoreParenImpCasts();
RefExpr = Base;
} else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) {
const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base))
Base = TempOASE->getBase()->IgnoreParenImpCasts();
while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
Base = TempASE->getBase()->IgnoreParenImpCasts();
RefExpr = Base;
}
RefExpr = RefExpr->IgnoreParenImpCasts();
if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr))
return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl());
const auto *ME = cast<MemberExpr>(RefExpr);
return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl());
}
static RecordDecl *buildRecordForGlobalizedVars(
ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
&MappedDeclsFields, int BufSize) {
using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>;
if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
return nullptr;
SmallVector<VarsDataTy, 4> GlobalizedVars;
for (const ValueDecl *D : EscapedDecls)
GlobalizedVars.emplace_back(
CharUnits::fromQuantity(std::max(
C.getDeclAlign(D).getQuantity(),
static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))),
D);
for (const ValueDecl *D : EscapedDeclsForTeams)
GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) {
return L.first > R.first;
});
// Build struct _globalized_locals_ty {
// /* globalized vars */[WarSize] align (max(decl_align,
// GlobalMemoryAlignment))
// /* globalized vars */ for EscapedDeclsForTeams
// };
RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
GlobalizedRD->startDefinition();
llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
for (const auto &Pair : GlobalizedVars) {
const ValueDecl *VD = Pair.second;
QualType Type = VD->getType();
if (Type->isLValueReferenceType())
Type = C.getPointerType(Type.getNonReferenceType());
else
Type = Type.getNonReferenceType();
SourceLocation Loc = VD->getLocation();
FieldDecl *Field;
if (SingleEscaped.count(VD)) {
Field = FieldDecl::Create(
C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
C.getTrivialTypeSourceInfo(Type, SourceLocation()),
/*BW=*/nullptr, /*Mutable=*/false,
/*InitStyle=*/ICIS_NoInit);
Field->setAccess(AS_public);
if (VD->hasAttrs()) {
for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
E(VD->getAttrs().end());
I != E; ++I)
Field->addAttr(*I);
}
} else {
llvm::APInt ArraySize(32, BufSize);
Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal,
0);
Field = FieldDecl::Create(
C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
C.getTrivialTypeSourceInfo(Type, SourceLocation()),
/*BW=*/nullptr, /*Mutable=*/false,
/*InitStyle=*/ICIS_NoInit);
Field->setAccess(AS_public);
llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(),
static_cast<CharUnits::QuantityType>(
GlobalMemoryAlignment)));
Field->addAttr(AlignedAttr::CreateImplicit(
C, /*IsAlignmentExpr=*/true,
IntegerLiteral::Create(C, Align,
C.getIntTypeForBitwidth(32, /*Signed=*/0),
SourceLocation()),
{}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned));
}
GlobalizedRD->addDecl(Field);
MappedDeclsFields.try_emplace(VD, Field);
}
GlobalizedRD->completeDefinition();
return GlobalizedRD;
}
/// Get the list of variables that can escape their declaration context.
class CheckVarsEscapingDeclContext final
: public ConstStmtVisitor<CheckVarsEscapingDeclContext> {
CodeGenFunction &CGF;
llvm::SetVector<const ValueDecl *> EscapedDecls;
llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls;
llvm::SmallPtrSet<const Decl *, 4> EscapedParameters;
RecordDecl *GlobalizedRD = nullptr;
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
bool AllEscaped = false;
bool IsForCombinedParallelRegion = false;
void markAsEscaped(const ValueDecl *VD) {
// Do not globalize declare target variables.
if (!isa<VarDecl>(VD) ||
OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD))
return;
VD = cast<ValueDecl>(VD->getCanonicalDecl());
// Use user-specified allocation.
if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>())
return;
// Variables captured by value must be globalized.
if (auto *CSI = CGF.CapturedStmtInfo) {
if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) {
// Check if need to capture the variable that was already captured by
// value in the outer region.
if (!IsForCombinedParallelRegion) {
if (!FD->hasAttrs())
return;
const auto *Attr = FD->getAttr<OMPCaptureKindAttr>();
if (!Attr)
return;
if (((Attr->getCaptureKind() != OMPC_map) &&
!isOpenMPPrivate(Attr->getCaptureKind())) ||
((Attr->getCaptureKind() == OMPC_map) &&
!FD->getType()->isAnyPointerType()))
return;
}
if (!FD->getType()->isReferenceType()) {
assert(!VD->getType()->isVariablyModifiedType() &&
"Parameter captured by value with variably modified type");
EscapedParameters.insert(VD);
} else if (!IsForCombinedParallelRegion) {
return;
}
}
}
if ((!CGF.CapturedStmtInfo ||
(IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) &&
VD->getType()->isReferenceType())
// Do not globalize variables with reference type.
return;
if (VD->getType()->isVariablyModifiedType())
EscapedVariableLengthDecls.insert(VD);
else
EscapedDecls.insert(VD);
}
void VisitValueDecl(const ValueDecl *VD) {
if (VD->getType()->isLValueReferenceType())
markAsEscaped(VD);
if (const auto *VarD = dyn_cast<VarDecl>(VD)) {
if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) {
const bool SavedAllEscaped = AllEscaped;
AllEscaped = VD->getType()->isLValueReferenceType();
Visit(VarD->getInit());
AllEscaped = SavedAllEscaped;
}
}
}
void VisitOpenMPCapturedStmt(const CapturedStmt *S,
ArrayRef<OMPClause *> Clauses,
bool IsCombinedParallelRegion) {
if (!S)
return;
for (const CapturedStmt::Capture &C : S->captures()) {
if (C.capturesVariable() && !C.capturesVariableByCopy()) {
const ValueDecl *VD = C.getCapturedVar();
bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion;
if (IsCombinedParallelRegion) {
// Check if the variable is privatized in the combined construct and
// those private copies must be shared in the inner parallel
// directive.
IsForCombinedParallelRegion = false;
for (const OMPClause *C : Clauses) {
if (!isOpenMPPrivate(C->getClauseKind()) ||
C->getClauseKind() == OMPC_reduction ||
C->getClauseKind() == OMPC_linear ||
C->getClauseKind() == OMPC_private)
continue;
ArrayRef<const Expr *> Vars;
if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C))
Vars = PC->getVarRefs();
else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C))
Vars = PC->getVarRefs();
else
llvm_unreachable("Unexpected clause.");
for (const auto *E : Vars) {
const Decl *D =
cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl();
if (D == VD->getCanonicalDecl()) {
IsForCombinedParallelRegion = true;
break;
}
}
if (IsForCombinedParallelRegion)
break;
}
}
markAsEscaped(VD);
if (isa<OMPCapturedExprDecl>(VD))
VisitValueDecl(VD);
IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion;
}
}
}
void buildRecordForGlobalizedVars(bool IsInTTDRegion) {
assert(!GlobalizedRD &&
"Record for globalized variables is built already.");
ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
unsigned WarpSize = CGF.getTarget().getGridValue(llvm::omp::GV_Warp_Size);
if (IsInTTDRegion)
EscapedDeclsForTeams = EscapedDecls.getArrayRef();
else
EscapedDeclsForParallel = EscapedDecls.getArrayRef();
GlobalizedRD = ::buildRecordForGlobalizedVars(
CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
MappedDeclsFields, WarpSize);
}
public:
CheckVarsEscapingDeclContext(CodeGenFunction &CGF,
ArrayRef<const ValueDecl *> TeamsReductions)
: CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) {
}
virtual ~CheckVarsEscapingDeclContext() = default;
void VisitDeclStmt(const DeclStmt *S) {
if (!S)
return;
for (const Decl *D : S->decls())
if (const auto *VD = dyn_cast_or_null<ValueDecl>(D))
VisitValueDecl(VD);
}
void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
if (!D)
return;
if (!D->hasAssociatedStmt())
return;
if (const auto *S =
dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
// Do not analyze directives that do not actually require capturing,
// like `omp for` or `omp simd` directives.
llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) {
VisitStmt(S->getCapturedStmt());
return;
}
VisitOpenMPCapturedStmt(
S, D->clauses(),
CaptureRegions.back() == OMPD_parallel &&
isOpenMPDistributeDirective(D->getDirectiveKind()));
}
}
void VisitCapturedStmt(const CapturedStmt *S) {
if (!S)
return;
for (const CapturedStmt::Capture &C : S->captures()) {
if (C.capturesVariable() && !C.capturesVariableByCopy()) {
const ValueDecl *VD = C.getCapturedVar();
markAsEscaped(VD);
if (isa<OMPCapturedExprDecl>(VD))
VisitValueDecl(VD);
}
}
}
void VisitLambdaExpr(const LambdaExpr *E) {
if (!E)
return;
for (const LambdaCapture &C : E->captures()) {
if (C.capturesVariable()) {
if (C.getCaptureKind() == LCK_ByRef) {
const ValueDecl *VD = C.getCapturedVar();
markAsEscaped(VD);
if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD))
VisitValueDecl(VD);
}
}
}
}
void VisitBlockExpr(const BlockExpr *E) {
if (!E)
return;
for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) {
if (C.isByRef()) {
const VarDecl *VD = C.getVariable();
markAsEscaped(VD);
if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture())
VisitValueDecl(VD);
}
}
}
void VisitCallExpr(const CallExpr *E) {
if (!E)
return;
for (const Expr *Arg : E->arguments()) {
if (!Arg)
continue;
if (Arg->isLValue()) {
const bool SavedAllEscaped = AllEscaped;
AllEscaped = true;
Visit(Arg);
AllEscaped = SavedAllEscaped;
} else {
Visit(Arg);
}
}
Visit(E->getCallee());
}
void VisitDeclRefExpr(const DeclRefExpr *E) {
if (!E)
return;
const ValueDecl *VD = E->getDecl();
if (AllEscaped)
markAsEscaped(VD);
if (isa<OMPCapturedExprDecl>(VD))
VisitValueDecl(VD);
else if (const auto *VarD = dyn_cast<VarDecl>(VD))
if (VarD->isInitCapture())
VisitValueDecl(VD);
}
void VisitUnaryOperator(const UnaryOperator *E) {
if (!E)
return;
if (E->getOpcode() == UO_AddrOf) {
const bool SavedAllEscaped = AllEscaped;
AllEscaped = true;
Visit(E->getSubExpr());
AllEscaped = SavedAllEscaped;
} else {
Visit(E->getSubExpr());
}
}
void VisitImplicitCastExpr(const ImplicitCastExpr *E) {
if (!E)
return;
if (E->getCastKind() == CK_ArrayToPointerDecay) {
const bool SavedAllEscaped = AllEscaped;
AllEscaped = true;
Visit(E->getSubExpr());
AllEscaped = SavedAllEscaped;
} else {
Visit(E->getSubExpr());
}
}
void VisitExpr(const Expr *E) {
if (!E)
return;
bool SavedAllEscaped = AllEscaped;
if (!E->isLValue())
AllEscaped = false;
for (const Stmt *Child : E->children())
if (Child)
Visit(Child);
AllEscaped = SavedAllEscaped;
}
void VisitStmt(const Stmt *S) {
if (!S)
return;
for (const Stmt *Child : S->children())
if (Child)
Visit(Child);
}
/// Returns the record that handles all the escaped local variables and used
/// instead of their original storage.
const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) {
if (!GlobalizedRD)
buildRecordForGlobalizedVars(IsInTTDRegion);
return GlobalizedRD;
}
/// Returns the field in the globalized record for the escaped variable.
const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const {
assert(GlobalizedRD &&
"Record for globalized variables must be generated already.");
auto I = MappedDeclsFields.find(VD);
if (I == MappedDeclsFields.end())
return nullptr;
return I->getSecond();
}
/// Returns the list of the escaped local variables/parameters.
ArrayRef<const ValueDecl *> getEscapedDecls() const {
return EscapedDecls.getArrayRef();
}
/// Checks if the escaped local variable is actually a parameter passed by
/// value.
const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const {
return EscapedParameters;
}
/// Returns the list of the escaped variables with the variably modified
/// types.
ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const {
return EscapedVariableLengthDecls.getArrayRef();
}
};
} // anonymous namespace
/// Get the id of the warp in the block.
/// We assume that the warp size is 32, which is always the case
/// on the NVPTX device, to generate more efficient code.
static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
unsigned LaneIDBits =
CGF.getTarget().getGridValue(llvm::omp::GV_Warp_Size_Log2);
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id");
}
/// Get the id of the current lane in the Warp.
/// We assume that the warp size is 32, which is always the case
/// on the NVPTX device, to generate more efficient code.
static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
unsigned LaneIDMask = CGF.getContext().getTargetInfo().getGridValue(
llvm::omp::GV_Warp_Size_Log2_Mask);
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask),
"nvptx_lane_id");
}
/// Get the value of the thread_limit clause in the teams directive.
/// For the 'generic' execution mode, the runtime encodes thread_limit in
/// the launch parameters, always starting thread_limit+warpSize threads per
/// CTA. The threads in the last warp are reserved for master execution.
/// For the 'spmd' execution mode, all threads in a CTA are part of the team.
static llvm::Value *getThreadLimit(CodeGenFunction &CGF,
bool IsInSPMDExecutionMode = false) {
CGBuilderTy &Bld = CGF.Builder;
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return IsInSPMDExecutionMode
? RT.getGPUNumThreads(CGF)
: Bld.CreateNUWSub(RT.getGPUNumThreads(CGF),
RT.getGPUWarpSize(CGF), "thread_limit");
}
/// Get the thread id of the OMP master thread.
/// The master thread id is the first thread (lane) of the last warp in the
/// GPU block. Warp size is assumed to be some power of 2.
/// Thread id is 0 indexed.
/// E.g: If NumThreads is 33, master id is 32.
/// If NumThreads is 64, master id is 32.
/// If NumThreads is 1024, master id is 992.
static llvm::Value *getMasterThreadID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
llvm::Value *NumThreads = RT.getGPUNumThreads(CGF);
// We assume that the warp size is a power of 2.
llvm::Value *Mask = Bld.CreateNUWSub(RT.getGPUWarpSize(CGF), Bld.getInt32(1));
return Bld.CreateAnd(Bld.CreateNUWSub(NumThreads, Bld.getInt32(1)),
Bld.CreateNot(Mask), "master_tid");
}
CGOpenMPRuntimeGPU::WorkerFunctionState::WorkerFunctionState(
CodeGenModule &CGM, SourceLocation Loc)
: WorkerFn(nullptr), CGFI(CGM.getTypes().arrangeNullaryFunction()),
Loc(Loc) {
createWorkerFunction(CGM);
}
void CGOpenMPRuntimeGPU::WorkerFunctionState::createWorkerFunction(
CodeGenModule &CGM) {
// Create an worker function with no arguments.
WorkerFn = llvm::Function::Create(
CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
/*placeholder=*/"_worker", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), WorkerFn, CGFI);
WorkerFn->setDoesNotRecurse();
}
CGOpenMPRuntimeGPU::ExecutionMode
CGOpenMPRuntimeGPU::getExecutionMode() const {
return CurrentExecutionMode;
}
static CGOpenMPRuntimeGPU::DataSharingMode
getDataSharingMode(CodeGenModule &CGM) {
return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA
: CGOpenMPRuntimeGPU::Generic;
}
/// Check for inner (nested) SPMD construct, if any
static bool hasNestedSPMDDirective(ASTContext &Ctx,
const OMPExecutableDirective &D) {
const auto *CS = D.getInnermostCapturedStmt();
const auto *Body =
CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NestedDir =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
switch (D.getDirectiveKind()) {
case OMPD_target:
if (isOpenMPParallelDirective(DKind))
return true;
if (DKind == OMPD_teams) {
Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
/*IgnoreCaptured=*/true);
if (!Body)
return false;
ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NND =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
DKind = NND->getDirectiveKind();
if (isOpenMPParallelDirective(DKind))
return true;
}
}
return false;
case OMPD_target_teams:
return isOpenMPParallelDirective(DKind);
case OMPD_target_simd:
case OMPD_target_parallel:
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute:
case OMPD_target_teams_distribute_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
case OMPD_parallel:
case OMPD_for:
case OMPD_parallel_for:
case OMPD_parallel_master:
case OMPD_parallel_sections:
case OMPD_for_simd:
case OMPD_parallel_for_simd:
case OMPD_cancel:
case OMPD_cancellation_point:
case OMPD_ordered:
case OMPD_threadprivate:
case OMPD_allocate:
case OMPD_task:
case OMPD_simd:
case OMPD_sections:
case OMPD_section:
case OMPD_single:
case OMPD_master:
case OMPD_critical:
case OMPD_taskyield:
case OMPD_barrier:
case OMPD_taskwait:
case OMPD_taskgroup:
case OMPD_atomic:
case OMPD_flush:
case OMPD_depobj:
case OMPD_scan:
case OMPD_teams:
case OMPD_target_data:
case OMPD_target_exit_data:
case OMPD_target_enter_data:
case OMPD_distribute:
case OMPD_distribute_simd:
case OMPD_distribute_parallel_for:
case OMPD_distribute_parallel_for_simd:
case OMPD_teams_distribute:
case OMPD_teams_distribute_simd:
case OMPD_teams_distribute_parallel_for:
case OMPD_teams_distribute_parallel_for_simd:
case OMPD_target_update:
case OMPD_declare_simd:
case OMPD_declare_variant:
case OMPD_begin_declare_variant:
case OMPD_end_declare_variant:
case OMPD_declare_target:
case OMPD_end_declare_target:
case OMPD_declare_reduction:
case OMPD_declare_mapper:
case OMPD_taskloop:
case OMPD_taskloop_simd:
case OMPD_master_taskloop:
case OMPD_master_taskloop_simd:
case OMPD_parallel_master_taskloop:
case OMPD_parallel_master_taskloop_simd:
case OMPD_requires:
case OMPD_unknown:
default:
llvm_unreachable("Unexpected directive.");
}
}
return false;
}
static bool supportsSPMDExecutionMode(ASTContext &Ctx,
const OMPExecutableDirective &D) {
OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
switch (DirectiveKind) {
case OMPD_target:
case OMPD_target_teams:
return hasNestedSPMDDirective(Ctx, D);
case OMPD_target_parallel:
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
case OMPD_target_simd:
case OMPD_target_teams_distribute_simd:
return true;
case OMPD_target_teams_distribute:
return false;
case OMPD_parallel:
case OMPD_for:
case OMPD_parallel_for:
case OMPD_parallel_master:
case OMPD_parallel_sections:
case OMPD_for_simd:
case OMPD_parallel_for_simd:
case OMPD_cancel:
case OMPD_cancellation_point:
case OMPD_ordered:
case OMPD_threadprivate:
case OMPD_allocate:
case OMPD_task:
case OMPD_simd:
case OMPD_sections:
case OMPD_section:
case OMPD_single:
case OMPD_master:
case OMPD_critical:
case OMPD_taskyield:
case OMPD_barrier:
case OMPD_taskwait:
case OMPD_taskgroup:
case OMPD_atomic:
case OMPD_flush:
case OMPD_depobj:
case OMPD_scan:
case OMPD_teams:
case OMPD_target_data:
case OMPD_target_exit_data:
case OMPD_target_enter_data:
case OMPD_distribute:
case OMPD_distribute_simd:
case OMPD_distribute_parallel_for:
case OMPD_distribute_parallel_for_simd:
case OMPD_teams_distribute:
case OMPD_teams_distribute_simd:
case OMPD_teams_distribute_parallel_for:
case OMPD_teams_distribute_parallel_for_simd:
case OMPD_target_update:
case OMPD_declare_simd:
case OMPD_declare_variant:
case OMPD_begin_declare_variant:
case OMPD_end_declare_variant:
case OMPD_declare_target:
case OMPD_end_declare_target:
case OMPD_declare_reduction:
case OMPD_declare_mapper:
case OMPD_taskloop:
case OMPD_taskloop_simd:
case OMPD_master_taskloop:
case OMPD_master_taskloop_simd:
case OMPD_parallel_master_taskloop:
case OMPD_parallel_master_taskloop_simd:
case OMPD_requires:
case OMPD_unknown:
default:
break;
}
llvm_unreachable(
"Unknown programming model for OpenMP directive on NVPTX target.");
}
/// Check if the directive is loops based and has schedule clause at all or has
/// static scheduling.
static bool hasStaticScheduling(const OMPExecutableDirective &D) {
assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) &&
isOpenMPLoopDirective(D.getDirectiveKind()) &&
"Expected loop-based directive.");
return !D.hasClausesOfKind<OMPOrderedClause>() &&
(!D.hasClausesOfKind<OMPScheduleClause>() ||
llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(),
[](const OMPScheduleClause *C) {
return C->getScheduleKind() == OMPC_SCHEDULE_static;
}));
}
/// Check for inner (nested) lightweight runtime construct, if any
static bool hasNestedLightweightDirective(ASTContext &Ctx,
const OMPExecutableDirective &D) {
assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive.");
const auto *CS = D.getInnermostCapturedStmt();
const auto *Body =
CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NestedDir =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
switch (D.getDirectiveKind()) {
case OMPD_target:
if (isOpenMPParallelDirective(DKind) &&
isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
hasStaticScheduling(*NestedDir))
return true;
if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd)
return true;
if (DKind == OMPD_parallel) {
Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
/*IgnoreCaptured=*/true);
if (!Body)
return false;
ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NND =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
DKind = NND->getDirectiveKind();
if (isOpenMPWorksharingDirective(DKind) &&
isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
return true;
}
} else if (DKind == OMPD_teams) {
Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
/*IgnoreCaptured=*/true);
if (!Body)
return false;
ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NND =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
DKind = NND->getDirectiveKind();
if (isOpenMPParallelDirective(DKind) &&
isOpenMPWorksharingDirective(DKind) &&
isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
return true;
if (DKind == OMPD_parallel) {
Body = NND->getInnermostCapturedStmt()->IgnoreContainers(
/*IgnoreCaptured=*/true);
if (!Body)
return false;
ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NND =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
DKind = NND->getDirectiveKind();
if (isOpenMPWorksharingDirective(DKind) &&
isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
return true;
}
}
}
}
return false;
case OMPD_target_teams:
if (isOpenMPParallelDirective(DKind) &&
isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
hasStaticScheduling(*NestedDir))
return true;
if (DKind == OMPD_distribute_simd || DKind == OMPD_simd)
return true;
if (DKind == OMPD_parallel) {
Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
/*IgnoreCaptured=*/true);
if (!Body)
return false;
ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
if (const auto *NND =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
DKind = NND->getDirectiveKind();
if (isOpenMPWorksharingDirective(DKind) &&
isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
return true;
}
}
return false;
case OMPD_target_parallel:
if (DKind == OMPD_simd)
return true;
return isOpenMPWorksharingDirective(DKind) &&
isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir);
case OMPD_target_teams_distribute:
case OMPD_target_simd:
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
case OMPD_parallel:
case OMPD_for:
case OMPD_parallel_for:
case OMPD_parallel_master:
case OMPD_parallel_sections:
case OMPD_for_simd:
case OMPD_parallel_for_simd:
case OMPD_cancel:
case OMPD_cancellation_point:
case OMPD_ordered:
case OMPD_threadprivate:
case OMPD_allocate:
case OMPD_task:
case OMPD_simd:
case OMPD_sections:
case OMPD_section:
case OMPD_single:
case OMPD_master:
case OMPD_critical:
case OMPD_taskyield:
case OMPD_barrier:
case OMPD_taskwait:
case OMPD_taskgroup:
case OMPD_atomic:
case OMPD_flush:
case OMPD_depobj:
case OMPD_scan:
case OMPD_teams:
case OMPD_target_data:
case OMPD_target_exit_data:
case OMPD_target_enter_data:
case OMPD_distribute:
case OMPD_distribute_simd:
case OMPD_distribute_parallel_for:
case OMPD_distribute_parallel_for_simd:
case OMPD_teams_distribute:
case OMPD_teams_distribute_simd:
case OMPD_teams_distribute_parallel_for:
case OMPD_teams_distribute_parallel_for_simd:
case OMPD_target_update:
case OMPD_declare_simd:
case OMPD_declare_variant:
case OMPD_begin_declare_variant:
case OMPD_end_declare_variant:
case OMPD_declare_target:
case OMPD_end_declare_target:
case OMPD_declare_reduction:
case OMPD_declare_mapper:
case OMPD_taskloop:
case OMPD_taskloop_simd:
case OMPD_master_taskloop:
case OMPD_master_taskloop_simd:
case OMPD_parallel_master_taskloop:
case OMPD_parallel_master_taskloop_simd:
case OMPD_requires:
case OMPD_unknown:
default:
llvm_unreachable("Unexpected directive.");
}
}
return false;
}
/// Checks if the construct supports lightweight runtime. It must be SPMD
/// construct + inner loop-based construct with static scheduling.
static bool supportsLightweightRuntime(ASTContext &Ctx,
const OMPExecutableDirective &D) {
if (!supportsSPMDExecutionMode(Ctx, D))
return false;
OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
switch (DirectiveKind) {
case OMPD_target:
case OMPD_target_teams:
case OMPD_target_parallel:
return hasNestedLightweightDirective(Ctx, D);
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
// (Last|First)-privates must be shared in parallel region.
return hasStaticScheduling(D);
case OMPD_target_simd:
case OMPD_target_teams_distribute_simd:
return true;
case OMPD_target_teams_distribute:
return false;
case OMPD_parallel:
case OMPD_for:
case OMPD_parallel_for:
case OMPD_parallel_master:
case OMPD_parallel_sections:
case OMPD_for_simd:
case OMPD_parallel_for_simd:
case OMPD_cancel:
case OMPD_cancellation_point:
case OMPD_ordered:
case OMPD_threadprivate:
case OMPD_allocate:
case OMPD_task:
case OMPD_simd:
case OMPD_sections:
case OMPD_section:
case OMPD_single:
case OMPD_master:
case OMPD_critical:
case OMPD_taskyield:
case OMPD_barrier:
case OMPD_taskwait:
case OMPD_taskgroup:
case OMPD_atomic:
case OMPD_flush:
case OMPD_depobj:
case OMPD_scan:
case OMPD_teams:
case OMPD_target_data:
case OMPD_target_exit_data:
case OMPD_target_enter_data:
case OMPD_distribute:
case OMPD_distribute_simd:
case OMPD_distribute_parallel_for:
case OMPD_distribute_parallel_for_simd:
case OMPD_teams_distribute:
case OMPD_teams_distribute_simd:
case OMPD_teams_distribute_parallel_for:
case OMPD_teams_distribute_parallel_for_simd:
case OMPD_target_update:
case OMPD_declare_simd:
case OMPD_declare_variant:
case OMPD_begin_declare_variant:
case OMPD_end_declare_variant:
case OMPD_declare_target:
case OMPD_end_declare_target:
case OMPD_declare_reduction:
case OMPD_declare_mapper:
case OMPD_taskloop:
case OMPD_taskloop_simd:
case OMPD_master_taskloop:
case OMPD_master_taskloop_simd:
case OMPD_parallel_master_taskloop:
case OMPD_parallel_master_taskloop_simd:
case OMPD_requires:
case OMPD_unknown:
default:
break;
}
llvm_unreachable(
"Unknown programming model for OpenMP directive on NVPTX target.");
}
void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
StringRef ParentName,
llvm::Function *&OutlinedFn,
llvm::Constant *&OutlinedFnID,
bool IsOffloadEntry,
const RegionCodeGenTy &CodeGen) {
ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode);
EntryFunctionState EST;
WorkerFunctionState WST(CGM, D.getBeginLoc());
Work.clear();
WrapperFunctionsMap.clear();
// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
CGOpenMPRuntimeGPU::EntryFunctionState &EST;
CGOpenMPRuntimeGPU::WorkerFunctionState &WST;
public:
NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST,
CGOpenMPRuntimeGPU::WorkerFunctionState &WST)
: EST(EST), WST(WST) {}
void Enter(CodeGenFunction &CGF) override {
auto &RT =
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
RT.emitNonSPMDEntryHeader(CGF, EST, WST);
// Skip target region initialization.
RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
}
void Exit(CodeGenFunction &CGF) override {
auto &RT =
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
RT.clearLocThreadIdInsertPt(CGF);
RT.emitNonSPMDEntryFooter(CGF, EST);
}
} Action(EST, WST);
CodeGen.setAction(Action);
IsInTTDRegion = true;
// Reserve place for the globalized memory.
GlobalizedRecords.emplace_back();
if (!KernelStaticGlobalized) {
KernelStaticGlobalized = new llvm::GlobalVariable(
CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false,
llvm::GlobalValue::InternalLinkage,
llvm::ConstantPointerNull::get(CGM.VoidPtrTy),
"_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr,
llvm::GlobalValue::NotThreadLocal,
CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
}
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
IsOffloadEntry, CodeGen);
IsInTTDRegion = false;
// Now change the name of the worker function to correspond to this target
// region's entry function.
WST.WorkerFn->setName(Twine(OutlinedFn->getName(), "_worker"));
// Create the worker function
emitWorkerFunction(WST);
}
// Setup NVPTX threads for master-worker OpenMP scheme.
void CGOpenMPRuntimeGPU::emitNonSPMDEntryHeader(CodeGenFunction &CGF,
EntryFunctionState &EST,
WorkerFunctionState &WST) {
CGBuilderTy &Bld = CGF.Builder;
llvm::BasicBlock *WorkerBB = CGF.createBasicBlock(".worker");
llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck");
llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
EST.ExitBB = CGF.createBasicBlock(".exit");
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
llvm::Value *IsWorker =
Bld.CreateICmpULT(RT.getGPUThreadID(CGF), getThreadLimit(CGF));
Bld.CreateCondBr(IsWorker, WorkerBB, MasterCheckBB);
CGF.EmitBlock(WorkerBB);
emitCall(CGF, WST.Loc, WST.WorkerFn);
CGF.EmitBranch(EST.ExitBB);
CGF.EmitBlock(MasterCheckBB);
llvm::Value *IsMaster =
Bld.CreateICmpEQ(RT.getGPUThreadID(CGF), getMasterThreadID(CGF));
Bld.CreateCondBr(IsMaster, MasterBB, EST.ExitBB);
CGF.EmitBlock(MasterBB);
IsInTargetMasterThreadRegion = true;
// SEQUENTIAL (MASTER) REGION START
// First action in sequential region:
// Initialize the state of the OpenMP runtime library on the GPU.
// TODO: Optimize runtime initialization and pass in correct value.
llvm::Value *Args[] = {getThreadLimit(CGF),
Bld.getInt16(/*RequiresOMPRuntime=*/1)};
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_init), Args);
// For data sharing, we need to initialize the stack.
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_data_sharing_init_stack));
emitGenericVarsProlog(CGF, WST.Loc);
}
void CGOpenMPRuntimeGPU::emitNonSPMDEntryFooter(CodeGenFunction &CGF,
EntryFunctionState &EST) {
IsInTargetMasterThreadRegion = false;
if (!CGF.HaveInsertPoint())
return;
emitGenericVarsEpilog(CGF);
if (!EST.ExitBB)
EST.ExitBB = CGF.createBasicBlock(".exit");
llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".termination.notifier");
CGF.EmitBranch(TerminateBB);
CGF.EmitBlock(TerminateBB);
// Signal termination condition.
// TODO: Optimize runtime initialization and pass in correct value.
llvm::Value *Args[] = {CGF.Builder.getInt16(/*IsOMPRuntimeInitialized=*/1)};
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_deinit), Args);
// Barrier to terminate worker threads.
syncCTAThreads(CGF);
// Master thread jumps to exit point.
CGF.EmitBranch(EST.ExitBB);
CGF.EmitBlock(EST.ExitBB);
EST.ExitBB = nullptr;
}
void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
StringRef ParentName,
llvm::Function *&OutlinedFn,
llvm::Constant *&OutlinedFnID,
bool IsOffloadEntry,
const RegionCodeGenTy &CodeGen) {
ExecutionRuntimeModesRAII ModeRAII(
CurrentExecutionMode, RequiresFullRuntime,
CGM.getLangOpts().OpenMPCUDAForceFullRuntime ||
!supportsLightweightRuntime(CGM.getContext(), D));
EntryFunctionState EST;
// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
CGOpenMPRuntimeGPU &RT;
CGOpenMPRuntimeGPU::EntryFunctionState &EST;
const OMPExecutableDirective &D;
public:
NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
CGOpenMPRuntimeGPU::EntryFunctionState &EST,
const OMPExecutableDirective &D)
: RT(RT), EST(EST), D(D) {}
void Enter(CodeGenFunction &CGF) override {
RT.emitSPMDEntryHeader(CGF, EST, D);
// Skip target region initialization.
RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
}
void Exit(CodeGenFunction &CGF) override {
RT.clearLocThreadIdInsertPt(CGF);
RT.emitSPMDEntryFooter(CGF, EST);
}
} Action(*this, EST, D);
CodeGen.setAction(Action);
IsInTTDRegion = true;
// Reserve place for the globalized memory.
GlobalizedRecords.emplace_back();
if (!KernelStaticGlobalized) {
KernelStaticGlobalized = new llvm::GlobalVariable(
CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false,
llvm::GlobalValue::InternalLinkage,
llvm::ConstantPointerNull::get(CGM.VoidPtrTy),
"_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr,
llvm::GlobalValue::NotThreadLocal,
CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
}
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
IsOffloadEntry, CodeGen);
IsInTTDRegion = false;
}
void CGOpenMPRuntimeGPU::emitSPMDEntryHeader(
CodeGenFunction &CGF, EntryFunctionState &EST,
const OMPExecutableDirective &D) {
CGBuilderTy &Bld = CGF.Builder;
// Setup BBs in entry function.
llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute");
EST.ExitBB = CGF.createBasicBlock(".exit");
llvm::Value *Args[] = {getThreadLimit(CGF, /*IsInSPMDExecutionMode=*/true),
/*RequiresOMPRuntime=*/
Bld.getInt16(RequiresFullRuntime ? 1 : 0),
/*RequiresDataSharing=*/Bld.getInt16(0)};
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_spmd_kernel_init), Args);
if (RequiresFullRuntime) {
// For data sharing, we need to initialize the stack.
CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd));
}
CGF.EmitBranch(ExecuteBB);
CGF.EmitBlock(ExecuteBB);
IsInTargetMasterThreadRegion = true;
}
void CGOpenMPRuntimeGPU::emitSPMDEntryFooter(CodeGenFunction &CGF,
EntryFunctionState &EST) {
IsInTargetMasterThreadRegion = false;
if (!CGF.HaveInsertPoint())
return;
if (!EST.ExitBB)
EST.ExitBB = CGF.createBasicBlock(".exit");
llvm::BasicBlock *OMPDeInitBB = CGF.createBasicBlock(".omp.deinit");
CGF.EmitBranch(OMPDeInitBB);
CGF.EmitBlock(OMPDeInitBB);
// DeInitialize the OMP state in the runtime; called by all active threads.
llvm::Value *Args[] = {/*RequiresOMPRuntime=*/
CGF.Builder.getInt16(RequiresFullRuntime ? 1 : 0)};
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2), Args);
CGF.EmitBranch(EST.ExitBB);
CGF.EmitBlock(EST.ExitBB);
EST.ExitBB = nullptr;
}
// Create a unique global variable to indicate the execution mode of this target
// region. The execution mode is either 'generic', or 'spmd' depending on the
// target directive. This variable is picked up by the offload library to setup
// the device appropriately before kernel launch. If the execution mode is
// 'generic', the runtime reserves one warp for the master, otherwise, all
// warps participate in parallel work.
static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name,
bool Mode) {
auto *GVMode =
new llvm::GlobalVariable(CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
llvm::GlobalValue::WeakAnyLinkage,
llvm::ConstantInt::get(CGM.Int8Ty, Mode ? 0 : 1),
Twine(Name, "_exec_mode"));
CGM.addCompilerUsedGlobal(GVMode);
}
void CGOpenMPRuntimeGPU::emitWorkerFunction(WorkerFunctionState &WST) {
ASTContext &Ctx = CGM.getContext();
CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, WST.WorkerFn, WST.CGFI, {},
WST.Loc, WST.Loc);
emitWorkerLoop(CGF, WST);
CGF.FinishFunction();
}
void CGOpenMPRuntimeGPU::emitWorkerLoop(CodeGenFunction &CGF,
WorkerFunctionState &WST) {
//
// The workers enter this loop and wait for parallel work from the master.
// When the master encounters a parallel region it sets up the work + variable
// arguments, and wakes up the workers. The workers first check to see if
// they are required for the parallel region, i.e., within the # of requested
// parallel threads. The activated workers load the variable arguments and
// execute the parallel work.
//
CGBuilderTy &Bld = CGF.Builder;
llvm::BasicBlock *AwaitBB = CGF.createBasicBlock(".await.work");
llvm::BasicBlock *SelectWorkersBB = CGF.createBasicBlock(".select.workers");
llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute.parallel");
llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".terminate.parallel");
llvm::BasicBlock *BarrierBB = CGF.createBasicBlock(".barrier.parallel");
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
CGF.EmitBranch(AwaitBB);
// Workers wait for work from master.
CGF.EmitBlock(AwaitBB);
// Wait for parallel work
syncCTAThreads(CGF);
Address WorkFn =
CGF.CreateDefaultAlignTempAlloca(CGF.Int8PtrTy, /*Name=*/"work_fn");
Address ExecStatus =
CGF.CreateDefaultAlignTempAlloca(CGF.Int8Ty, /*Name=*/"exec_status");
CGF.InitTempAlloca(ExecStatus, Bld.getInt8(/*C=*/0));
CGF.InitTempAlloca(WorkFn, llvm::Constant::getNullValue(CGF.Int8PtrTy));
// TODO: Optimize runtime initialization and pass in correct value.
llvm::Value *Args[] = {WorkFn.getPointer()};
llvm::Value *Ret = CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_parallel), Args);
Bld.CreateStore(Bld.CreateZExt(Ret, CGF.Int8Ty), ExecStatus);
// On termination condition (workid == 0), exit loop.
llvm::Value *WorkID = Bld.CreateLoad(WorkFn);
llvm::Value *ShouldTerminate = Bld.CreateIsNull(WorkID, "should_terminate");
Bld.CreateCondBr(ShouldTerminate, ExitBB, SelectWorkersBB);
// Activate requested workers.
CGF.EmitBlock(SelectWorkersBB);
llvm::Value *IsActive =
Bld.CreateIsNotNull(Bld.CreateLoad(ExecStatus), "is_active");
Bld.CreateCondBr(IsActive, ExecuteBB, BarrierBB);
// Signal start of parallel region.
CGF.EmitBlock(ExecuteBB);
// Skip initialization.
setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
// Process work items: outlined parallel functions.
for (llvm::Function *W : Work) {
// Try to match this outlined function.
llvm::Value *ID = Bld.CreatePointerBitCastOrAddrSpaceCast(W, CGM.Int8PtrTy);
llvm::Value *WorkFnMatch =
Bld.CreateICmpEQ(Bld.CreateLoad(WorkFn), ID, "work_match");
llvm::BasicBlock *ExecuteFNBB = CGF.createBasicBlock(".execute.fn");
llvm::BasicBlock *CheckNextBB = CGF.createBasicBlock(".check.next");
Bld.CreateCondBr(WorkFnMatch, ExecuteFNBB, CheckNextBB);
// Execute this outlined function.
CGF.EmitBlock(ExecuteFNBB);
// Insert call to work function via shared wrapper. The shared
// wrapper takes two arguments:
// - the parallelism level;
// - the thread ID;
emitCall(CGF, WST.Loc, W,
{Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)});
// Go to end of parallel region.
CGF.EmitBranch(TerminateBB);
CGF.EmitBlock(CheckNextBB);
}
// Default case: call to outlined function through pointer if the target
// region makes a declare target call that may contain an orphaned parallel
// directive.
auto *ParallelFnTy =
llvm::FunctionType::get(CGM.VoidTy, {CGM.Int16Ty, CGM.Int32Ty},
/*isVarArg=*/false);
llvm::Value *WorkFnCast =
Bld.CreateBitCast(WorkID, ParallelFnTy->getPointerTo());
// Insert call to work function via shared wrapper. The shared
// wrapper takes two arguments:
// - the parallelism level;
// - the thread ID;
emitCall(CGF, WST.Loc, {ParallelFnTy, WorkFnCast},
{Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)});
// Go to end of parallel region.
CGF.EmitBranch(TerminateBB);
// Signal end of parallel region.
CGF.EmitBlock(TerminateBB);
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_end_parallel),
llvm::None);
CGF.EmitBranch(BarrierBB);
// All active and inactive workers wait at a barrier after parallel region.
CGF.EmitBlock(BarrierBB);
// Barrier after parallel region.
syncCTAThreads(CGF);
CGF.EmitBranch(AwaitBB);
// Exit target region.
CGF.EmitBlock(ExitBB);
// Skip initialization.
clearLocThreadIdInsertPt(CGF);
}
/// Returns specified OpenMP runtime function for the current OpenMP
/// implementation. Specialized for the NVPTX device.
/// \param Function OpenMP runtime function.
/// \return Specified function.
llvm::FunctionCallee
CGOpenMPRuntimeGPU::createNVPTXRuntimeFunction(unsigned Function) {
llvm::FunctionCallee RTLFn = nullptr;
switch (static_cast<OpenMPRTLFunctionNVPTX>(Function)) {
case OMPRTL_NVPTX__kmpc_kernel_init: {
// Build void __kmpc_kernel_init(kmp_int32 thread_limit, int16_t
// RequiresOMPRuntime);
llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_init");
break;
}
case OMPRTL_NVPTX__kmpc_kernel_deinit: {
// Build void __kmpc_kernel_deinit(int16_t IsOMPRuntimeInitialized);
llvm::Type *TypeParams[] = {CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_deinit");
break;
}
case OMPRTL_NVPTX__kmpc_spmd_kernel_init: {
// Build void __kmpc_spmd_kernel_init(kmp_int32 thread_limit,
// int16_t RequiresOMPRuntime, int16_t RequiresDataSharing);
llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_spmd_kernel_init");
break;
}
case OMPRTL_NVPTX__kmpc_spmd_kernel_deinit_v2: {
// Build void __kmpc_spmd_kernel_deinit_v2(int16_t RequiresOMPRuntime);
llvm::Type *TypeParams[] = {CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_spmd_kernel_deinit_v2");
break;
}
case OMPRTL_NVPTX__kmpc_kernel_prepare_parallel: {
/// Build void __kmpc_kernel_prepare_parallel(
/// void *outlined_function);
llvm::Type *TypeParams[] = {CGM.Int8PtrTy};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_prepare_parallel");
break;
}
case OMPRTL_NVPTX__kmpc_kernel_parallel: {
/// Build bool __kmpc_kernel_parallel(void **outlined_function);
llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy};
llvm::Type *RetTy = CGM.getTypes().ConvertType(CGM.getContext().BoolTy);
auto *FnTy =
llvm::FunctionType::get(RetTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_parallel");
break;
}
case OMPRTL_NVPTX__kmpc_kernel_end_parallel: {
/// Build void __kmpc_kernel_end_parallel();
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_kernel_end_parallel");
break;
}
case OMPRTL_NVPTX__kmpc_serialized_parallel: {
// Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32
// global_tid);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel");
break;
}
case OMPRTL_NVPTX__kmpc_end_serialized_parallel: {
// Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32
// global_tid);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel");
break;
}
case OMPRTL_NVPTX__kmpc_shuffle_int32: {
// Build int32_t __kmpc_shuffle_int32(int32_t element,
// int16_t lane_offset, int16_t warp_size);
llvm::Type *TypeParams[] = {CGM.Int32Ty, CGM.Int16Ty, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_shuffle_int32");
break;
}
case OMPRTL_NVPTX__kmpc_shuffle_int64: {
// Build int64_t __kmpc_shuffle_int64(int64_t element,
// int16_t lane_offset, int16_t warp_size);
llvm::Type *TypeParams[] = {CGM.Int64Ty, CGM.Int16Ty, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.Int64Ty, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_shuffle_int64");
break;
}
case OMPRTL_NVPTX__kmpc_nvptx_parallel_reduce_nowait_v2: {
// Build int32_t kmpc_nvptx_parallel_reduce_nowait_v2(ident_t *loc,
// kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void*
// reduce_data, void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t
// lane_id, int16_t lane_offset, int16_t Algorithm Version), void
// (*kmp_InterWarpCopyFctPtr)(void* src, int warp_num));
llvm::Type *ShuffleReduceTypeParams[] = {CGM.VoidPtrTy, CGM.Int16Ty,
CGM.Int16Ty, CGM.Int16Ty};
auto *ShuffleReduceFnTy =
llvm::FunctionType::get(CGM.VoidTy, ShuffleReduceTypeParams,
/*isVarArg=*/false);
llvm::Type *InterWarpCopyTypeParams[] = {CGM.VoidPtrTy, CGM.Int32Ty};
auto *InterWarpCopyFnTy =
llvm::FunctionType::get(CGM.VoidTy, InterWarpCopyTypeParams,
/*isVarArg=*/false);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(),
CGM.Int32Ty,
CGM.Int32Ty,
CGM.SizeTy,
CGM.VoidPtrTy,
ShuffleReduceFnTy->getPointerTo(),
InterWarpCopyFnTy->getPointerTo()};
auto *FnTy =
llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(
FnTy, /*Name=*/"__kmpc_nvptx_parallel_reduce_nowait_v2");
break;
}
case OMPRTL_NVPTX__kmpc_end_reduce_nowait: {
// Build __kmpc_end_reduce_nowait(kmp_int32 global_tid);
llvm::Type *TypeParams[] = {CGM.Int32Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(
FnTy, /*Name=*/"__kmpc_nvptx_end_reduce_nowait");
break;
}
case OMPRTL_NVPTX__kmpc_nvptx_teams_reduce_nowait_v2: {
// Build int32_t __kmpc_nvptx_teams_reduce_nowait_v2(ident_t *loc, kmp_int32
// global_tid, void *global_buffer, int32_t num_of_records, void*
// reduce_data,
// void (*kmp_ShuffleReductFctPtr)(void *rhsData, int16_t lane_id, int16_t
// lane_offset, int16_t shortCircuit),
// void (*kmp_InterWarpCopyFctPtr)(void* src, int32_t warp_num), void
// (*kmp_ListToGlobalCpyFctPtr)(void *buffer, int idx, void *reduce_data),
// void (*kmp_GlobalToListCpyFctPtr)(void *buffer, int idx,
// void *reduce_data), void (*kmp_GlobalToListCpyPtrsFctPtr)(void *buffer,
// int idx, void *reduce_data), void (*kmp_GlobalToListRedFctPtr)(void
// *buffer, int idx, void *reduce_data));
llvm::Type *ShuffleReduceTypeParams[] = {CGM.VoidPtrTy, CGM.Int16Ty,
CGM.Int16Ty, CGM.Int16Ty};
auto *ShuffleReduceFnTy =
llvm::FunctionType::get(CGM.VoidTy, ShuffleReduceTypeParams,
/*isVarArg=*/false);
llvm::Type *InterWarpCopyTypeParams[] = {CGM.VoidPtrTy, CGM.Int32Ty};
auto *InterWarpCopyFnTy =
llvm::FunctionType::get(CGM.VoidTy, InterWarpCopyTypeParams,
/*isVarArg=*/false);
llvm::Type *GlobalListTypeParams[] = {CGM.VoidPtrTy, CGM.IntTy,
CGM.VoidPtrTy};
auto *GlobalListFnTy =
llvm::FunctionType::get(CGM.VoidTy, GlobalListTypeParams,
/*isVarArg=*/false);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(),
CGM.Int32Ty,
CGM.VoidPtrTy,
CGM.Int32Ty,
CGM.VoidPtrTy,
ShuffleReduceFnTy->getPointerTo(),
InterWarpCopyFnTy->getPointerTo(),
GlobalListFnTy->getPointerTo(),
GlobalListFnTy->getPointerTo(),
GlobalListFnTy->getPointerTo(),
GlobalListFnTy->getPointerTo()};
auto *FnTy =
llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(
FnTy, /*Name=*/"__kmpc_nvptx_teams_reduce_nowait_v2");
break;
}
case OMPRTL_NVPTX__kmpc_data_sharing_init_stack: {
/// Build void __kmpc_data_sharing_init_stack();
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_data_sharing_init_stack");
break;
}
case OMPRTL_NVPTX__kmpc_data_sharing_init_stack_spmd: {
/// Build void __kmpc_data_sharing_init_stack_spmd();
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
RTLFn =
CGM.CreateRuntimeFunction(FnTy, "__kmpc_data_sharing_init_stack_spmd");
break;
}
case OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack: {
// Build void *__kmpc_data_sharing_coalesced_push_stack(size_t size,
// int16_t UseSharedMemory);
llvm::Type *TypeParams[] = {CGM.SizeTy, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(
FnTy, /*Name=*/"__kmpc_data_sharing_coalesced_push_stack");
break;
}
case OMPRTL_NVPTX__kmpc_data_sharing_push_stack: {
// Build void *__kmpc_data_sharing_push_stack(size_t size, int16_t
// UseSharedMemory);
llvm::Type *TypeParams[] = {CGM.SizeTy, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(
FnTy, /*Name=*/"__kmpc_data_sharing_push_stack");
break;
}
case OMPRTL_NVPTX__kmpc_data_sharing_pop_stack: {
// Build void __kmpc_data_sharing_pop_stack(void *a);
llvm::Type *TypeParams[] = {CGM.VoidPtrTy};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(FnTy,
/*Name=*/"__kmpc_data_sharing_pop_stack");
break;
}
case OMPRTL_NVPTX__kmpc_begin_sharing_variables: {
/// Build void __kmpc_begin_sharing_variables(void ***args,
/// size_t n_args);
llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy->getPointerTo(), CGM.SizeTy};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_begin_sharing_variables");
break;
}
case OMPRTL_NVPTX__kmpc_end_sharing_variables: {
/// Build void __kmpc_end_sharing_variables();
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, llvm::None, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_sharing_variables");
break;
}
case OMPRTL_NVPTX__kmpc_get_shared_variables: {
/// Build void __kmpc_get_shared_variables(void ***GlobalArgs);
llvm::Type *TypeParams[] = {CGM.Int8PtrPtrTy->getPointerTo()};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_get_shared_variables");
break;
}
case OMPRTL_NVPTX__kmpc_parallel_level: {
// Build uint16_t __kmpc_parallel_level(ident_t *loc, kmp_int32 global_tid);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.Int16Ty, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_parallel_level");
break;
}
case OMPRTL_NVPTX__kmpc_is_spmd_exec_mode: {
// Build int8_t __kmpc_is_spmd_exec_mode();
auto *FnTy = llvm::FunctionType::get(CGM.Int8Ty, /*isVarArg=*/false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_is_spmd_exec_mode");
break;
}
case OMPRTL_NVPTX__kmpc_get_team_static_memory: {
// Build void __kmpc_get_team_static_memory(int16_t isSPMDExecutionMode,
// const void *buf, size_t size, int16_t is_shared, const void **res);
llvm::Type *TypeParams[] = {CGM.Int16Ty, CGM.VoidPtrTy, CGM.SizeTy,
CGM.Int16Ty, CGM.VoidPtrPtrTy};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_get_team_static_memory");
break;
}
case OMPRTL_NVPTX__kmpc_restore_team_static_memory: {
// Build void __kmpc_restore_team_static_memory(int16_t isSPMDExecutionMode,
// int16_t is_shared);
llvm::Type *TypeParams[] = {CGM.Int16Ty, CGM.Int16Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
RTLFn =
CGM.CreateRuntimeFunction(FnTy, "__kmpc_restore_team_static_memory");
break;
}
case OMPRTL__kmpc_barrier: {
// Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn =
CGM.CreateConvergentRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier");
break;
}
case OMPRTL__kmpc_barrier_simple_spmd: {
// Build void __kmpc_barrier_simple_spmd(ident_t *loc, kmp_int32
// global_tid);
llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty};
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false);
RTLFn = CGM.CreateConvergentRuntimeFunction(
FnTy, /*Name*/ "__kmpc_barrier_simple_spmd");
break;
}
case OMPRTL_NVPTX__kmpc_warp_active_thread_mask: {
// Build int32_t __kmpc_warp_active_thread_mask(void);
auto *FnTy =
llvm::FunctionType::get(CGM.Int32Ty, llvm::None, /*isVarArg=*/false);
RTLFn = CGM.CreateConvergentRuntimeFunction(FnTy, "__kmpc_warp_active_thread_mask");
break;
}
case OMPRTL_NVPTX__kmpc_syncwarp: {
// Build void __kmpc_syncwarp(kmp_int32 Mask);
auto *FnTy =
llvm::FunctionType::get(CGM.VoidTy, CGM.Int32Ty, /*isVarArg=*/false);
RTLFn = CGM.CreateConvergentRuntimeFunction(FnTy, "__kmpc_syncwarp");
break;
}
}
return RTLFn;
}
void CGOpenMPRuntimeGPU::createOffloadEntry(llvm::Constant *ID,
llvm::Constant *Addr,
uint64_t Size, int32_t,
llvm::GlobalValue::LinkageTypes) {
// TODO: Add support for global variables on the device after declare target
// support.
if (!isa<llvm::Function>(Addr))
return;
llvm::Module &M = CGM.getModule();
llvm::LLVMContext &Ctx = CGM.getLLVMContext();
// Get "nvvm.annotations" metadata node
llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");
llvm::Metadata *MDVals[] = {
llvm::ConstantAsMetadata::get(Addr), llvm::MDString::get(Ctx, "kernel"),
llvm::ConstantAsMetadata::get(
llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
// Append metadata to nvvm.annotations
MD->addOperand(llvm::MDNode::get(Ctx, MDVals));
}
void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction(
const OMPExecutableDirective &D, StringRef ParentName,
llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
if (!IsOffloadEntry) // Nothing to do.
return;
assert(!ParentName.empty() && "Invalid target region parent name!");
bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
if (Mode)
emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
CodeGen);
else
emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
CodeGen);
setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode);
}
namespace {
LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE();
/// Enum for accesseing the reserved_2 field of the ident_t struct.
enum ModeFlagsTy : unsigned {
/// Bit set to 1 when in SPMD mode.
KMP_IDENT_SPMD_MODE = 0x01,
/// Bit set to 1 when a simplified runtime is used.
KMP_IDENT_SIMPLE_RT_MODE = 0x02,
LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE)
};
/// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime.
static const ModeFlagsTy UndefinedMode =
(~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE;
} // anonymous namespace
unsigned CGOpenMPRuntimeGPU::getDefaultLocationReserved2Flags() const {
switch (getExecutionMode()) {
case EM_SPMD:
if (requiresFullRuntime())
return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE);
return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE;
case EM_NonSPMD:
assert(requiresFullRuntime() && "Expected full runtime.");
return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE);
case EM_Unknown:
return UndefinedMode;
}
llvm_unreachable("Unknown flags are requested.");
}
CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
: CGOpenMPRuntime(CGM, "_", "$") {
if (!CGM.getLangOpts().OpenMPIsDevice)
llvm_unreachable("OpenMP NVPTX can only handle device code.");
}
void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF,
ProcBindKind ProcBind,
SourceLocation Loc) {
// Do nothing in case of SPMD mode and L0 parallel.
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
return;
CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc);
}
void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
llvm::Value *NumThreads,
SourceLocation Loc) {
// Do nothing in case of SPMD mode and L0 parallel.
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
return;
CGOpenMPRuntime::emitNumThreadsClause(CGF, NumThreads, Loc);
}
void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
const Expr *NumTeams,
const Expr *ThreadLimit,
SourceLocation Loc) {}
llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
bool &IsInParallelRegion;
bool PrevIsInParallelRegion;
public:
NVPTXPrePostActionTy(bool &IsInParallelRegion)
: IsInParallelRegion(IsInParallelRegion) {}
void Enter(CodeGenFunction &CGF) override {
PrevIsInParallelRegion = IsInParallelRegion;
IsInParallelRegion = true;
}
void Exit(CodeGenFunction &CGF) override {
IsInParallelRegion = PrevIsInParallelRegion;
}
} Action(IsInParallelRegion);
CodeGen.setAction(Action);
bool PrevIsInTTDRegion = IsInTTDRegion;
IsInTTDRegion = false;
bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion;
IsInTargetMasterThreadRegion = false;
auto *OutlinedFun =
cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
D, ThreadIDVar, InnermostKind, CodeGen));
if (CGM.getLangOpts().Optimize) {
OutlinedFun->removeFnAttr(llvm::Attribute::NoInline);
OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone);
OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline);
}
IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion;
IsInTTDRegion = PrevIsInTTDRegion;
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD &&
!IsInParallelRegion) {
llvm::Function *WrapperFun =
createParallelDataSharingWrapper(OutlinedFun, D);
WrapperFunctionsMap[OutlinedFun] = WrapperFun;
}
return OutlinedFun;
}
/// Get list of lastprivate variables from the teams distribute ... or
/// teams {distribute ...} directives.
static void
getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
"expected teams directive.");
const OMPExecutableDirective *Dir = &D;
if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
Ctx,
D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
/*IgnoreCaptured=*/true))) {
Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
Dir = nullptr;
}
}
if (!Dir)
return;
for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
for (const Expr *E : C->getVarRefs())
Vars.push_back(getPrivateItem(E));
}
}
/// Get list of reduction variables from the teams ... directives.
static void
getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
"expected teams directive.");
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
for (const Expr *E : C->privates())
Vars.push_back(getPrivateItem(E));
}
}
llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction(
const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
SourceLocation Loc = D.getBeginLoc();
const RecordDecl *GlobalizedRD = nullptr;
llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
unsigned WarpSize = CGM.getTarget().getGridValue(llvm::omp::GV_Warp_Size);
// Globalize team reductions variable unconditionally in all modes.
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
if (!LastPrivatesReductions.empty()) {
GlobalizedRD = ::buildRecordForGlobalizedVars(
CGM.getContext(), llvm::None, LastPrivatesReductions,
MappedDeclsFields, WarpSize);
}
} else if (!LastPrivatesReductions.empty()) {
assert(!TeamAndReductions.first &&
"Previous team declaration is not expected.");
TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
std::swap(TeamAndReductions.second, LastPrivatesReductions);
}
// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
SourceLocation &Loc;
const RecordDecl *GlobalizedRD;
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
&MappedDeclsFields;
public:
NVPTXPrePostActionTy(
SourceLocation &Loc, const RecordDecl *GlobalizedRD,
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
&MappedDeclsFields)
: Loc(Loc), GlobalizedRD(GlobalizedRD),
MappedDeclsFields(MappedDeclsFields) {}
void Enter(CodeGenFunction &CGF) override {
auto &Rt =
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
if (GlobalizedRD) {
auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
I->getSecond().GlobalRecord = GlobalizedRD;
I->getSecond().MappedParams =
std::make_unique<CodeGenFunction::OMPMapVars>();
DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
for (const auto &Pair : MappedDeclsFields) {
assert(Pair.getFirst()->isCanonicalDecl() &&
"Expected canonical declaration");
Data.insert(std::make_pair(Pair.getFirst(),
MappedVarData(Pair.getSecond(),
/*IsOnePerTeam=*/true)));
}
}
Rt.emitGenericVarsProlog(CGF, Loc);
}
void Exit(CodeGenFunction &CGF) override {
static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
.emitGenericVarsEpilog(CGF);
}
} Action(Loc, GlobalizedRD, MappedDeclsFields);
CodeGen.setAction(Action);
llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
D, ThreadIDVar, InnermostKind, CodeGen);
if (CGM.getLangOpts().Optimize) {
OutlinedFun->removeFnAttr(llvm::Attribute::NoInline);
OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone);
OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline);
}
return OutlinedFun;
}
void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF,
SourceLocation Loc,
bool WithSPMDCheck) {
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
return;
CGBuilderTy &Bld = CGF.Builder;
const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
if (I == FunctionGlobalizedDecls.end())
return;
if (const RecordDecl *GlobalizedVarsRecord = I->getSecond().GlobalRecord) {
QualType GlobalRecTy = CGM.getContext().getRecordType(GlobalizedVarsRecord);
QualType SecGlobalRecTy;
// Recover pointer to this function's global record. The runtime will
// handle the specifics of the allocation of the memory.
// Use actual memory size of the record including the padding
// for alignment purposes.
unsigned Alignment =
CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
unsigned GlobalRecordSize =
CGM.getContext().getTypeSizeInChars(GlobalRecTy).getQuantity();
GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);
llvm::PointerType *GlobalRecPtrTy =
CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo();
llvm::Value *GlobalRecCastAddr;
llvm::Value *IsTTD = nullptr;
if (!IsInTTDRegion &&
(WithSPMDCheck ||
getExecutionMode() == CGOpenMPRuntimeGPU::EM_Unknown)) {
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
llvm::BasicBlock *SPMDBB = CGF.createBasicBlock(".spmd");
llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd");
if (I->getSecond().SecondaryGlobalRecord.hasValue()) {
llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
llvm::Value *ThreadID = getThreadID(CGF, Loc);
llvm::Value *PL = CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_parallel_level),
{RTLoc, ThreadID});
IsTTD = Bld.CreateIsNull(PL);
}
llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode)));
Bld.CreateCondBr(IsSPMD, SPMDBB, NonSPMDBB);
// There is no need to emit line number for unconditional branch.
(void)ApplyDebugLocation::CreateEmpty(CGF);
CGF.EmitBlock(SPMDBB);
Address RecPtr = Address(llvm::ConstantPointerNull::get(GlobalRecPtrTy),
CharUnits::fromQuantity(Alignment));
CGF.EmitBranch(ExitBB);
// There is no need to emit line number for unconditional branch.
(void)ApplyDebugLocation::CreateEmpty(CGF);
CGF.EmitBlock(NonSPMDBB);
llvm::Value *Size = llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize);
if (const RecordDecl *SecGlobalizedVarsRecord =
I->getSecond().SecondaryGlobalRecord.getValueOr(nullptr)) {
SecGlobalRecTy =
CGM.getContext().getRecordType(SecGlobalizedVarsRecord);
// Recover pointer to this function's global record. The runtime will
// handle the specifics of the allocation of the memory.
// Use actual memory size of the record including the padding
// for alignment purposes.
unsigned Alignment =
CGM.getContext().getTypeAlignInChars(SecGlobalRecTy).getQuantity();
unsigned GlobalRecordSize =
CGM.getContext().getTypeSizeInChars(SecGlobalRecTy).getQuantity();
GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);
Size = Bld.CreateSelect(
IsTTD, llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize), Size);
}
// TODO: allow the usage of shared memory to be controlled by
// the user, for now, default to global.
llvm::Value *GlobalRecordSizeArg[] = {
Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack),
GlobalRecordSizeArg);
GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
GlobalRecValue, GlobalRecPtrTy);
CGF.EmitBlock(ExitBB);
auto *Phi = Bld.CreatePHI(GlobalRecPtrTy,
/*NumReservedValues=*/2, "_select_stack");
Phi->addIncoming(RecPtr.getPointer(), SPMDBB);
Phi->addIncoming(GlobalRecCastAddr, NonSPMDBB);
GlobalRecCastAddr = Phi;
I->getSecond().GlobalRecordAddr = Phi;
I->getSecond().IsInSPMDModeFlag = IsSPMD;
} else if (!CGM.getLangOpts().OpenMPCUDATargetParallel && IsInTTDRegion) {
assert(GlobalizedRecords.back().Records.size() < 2 &&
"Expected less than 2 globalized records: one for target and one "
"for teams.");
unsigned Offset = 0;
for (const RecordDecl *RD : GlobalizedRecords.back().Records) {
QualType RDTy = CGM.getContext().getRecordType(RD);
unsigned Alignment =
CGM.getContext().getTypeAlignInChars(RDTy).getQuantity();
unsigned Size = CGM.getContext().getTypeSizeInChars(RDTy).getQuantity();
Offset =
llvm::alignTo(llvm::alignTo(Offset, Alignment) + Size, Alignment);
}
unsigned Alignment =
CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
Offset = llvm::alignTo(Offset, Alignment);
GlobalizedRecords.back().Records.push_back(GlobalizedVarsRecord);
++GlobalizedRecords.back().RegionCounter;
if (GlobalizedRecords.back().Records.size() == 1) {
assert(KernelStaticGlobalized &&
"Kernel static pointer must be initialized already.");
auto *UseSharedMemory = new llvm::GlobalVariable(
CGM.getModule(), CGM.Int16Ty, /*isConstant=*/true,
llvm::GlobalValue::InternalLinkage, nullptr,
"_openmp_static_kernel$is_shared");
UseSharedMemory->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth(
/*DestWidth=*/16, /*Signed=*/0);
llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar(
Address(UseSharedMemory,
CGM.getContext().getTypeAlignInChars(Int16Ty)),
/*Volatile=*/false, Int16Ty, Loc);
auto *StaticGlobalized = new llvm::GlobalVariable(
CGM.getModule(), CGM.Int8Ty, /*isConstant=*/false,
llvm::GlobalValue::CommonLinkage, nullptr);
auto *RecSize = new llvm::GlobalVariable(
CGM.getModule(), CGM.SizeTy, /*isConstant=*/true,
llvm::GlobalValue::InternalLinkage, nullptr,
"_openmp_static_kernel$size");
RecSize->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
llvm::Value *Ld = CGF.EmitLoadOfScalar(
Address(RecSize, CGM.getSizeAlign()), /*Volatile=*/false,
CGM.getContext().getSizeType(), Loc);
llvm::Value *ResAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
KernelStaticGlobalized, CGM.VoidPtrPtrTy);
llvm::Value *GlobalRecordSizeArg[] = {
llvm::ConstantInt::get(
CGM.Int16Ty,
getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD ? 1 : 0),
StaticGlobalized, Ld, IsInSharedMemory, ResAddr};
CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_get_team_static_memory),
GlobalRecordSizeArg);
GlobalizedRecords.back().Buffer = StaticGlobalized;
GlobalizedRecords.back().RecSize = RecSize;
GlobalizedRecords.back().UseSharedMemory = UseSharedMemory;
GlobalizedRecords.back().Loc = Loc;
}
assert(KernelStaticGlobalized && "Global address must be set already.");
Address FrameAddr = CGF.EmitLoadOfPointer(
Address(KernelStaticGlobalized, CGM.getPointerAlign()),
CGM.getContext()
.getPointerType(CGM.getContext().VoidPtrTy)
.castAs<PointerType>());
llvm::Value *GlobalRecValue =
Bld.CreateConstInBoundsGEP(FrameAddr, Offset).getPointer();
I->getSecond().GlobalRecordAddr = GlobalRecValue;
I->getSecond().IsInSPMDModeFlag = nullptr;
GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
GlobalRecValue, CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo());
} else {
// TODO: allow the usage of shared memory to be controlled by
// the user, for now, default to global.
bool UseSharedMemory =
IsInTTDRegion && GlobalRecordSize <= SharedMemorySize;
llvm::Value *GlobalRecordSizeArg[] = {
llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize),
CGF.Builder.getInt16(UseSharedMemory ? 1 : 0)};
llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
IsInTTDRegion
? OMPRTL_NVPTX__kmpc_data_sharing_push_stack
: OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack),
GlobalRecordSizeArg);
GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
GlobalRecValue, GlobalRecPtrTy);
I->getSecond().GlobalRecordAddr = GlobalRecValue;
I->getSecond().IsInSPMDModeFlag = nullptr;
}
LValue Base =
CGF.MakeNaturalAlignPointeeAddrLValue(GlobalRecCastAddr, GlobalRecTy);
// Emit the "global alloca" which is a GEP from the global declaration
// record using the pointer returned by the runtime.
LValue SecBase;
decltype(I->getSecond().LocalVarData)::const_iterator SecIt;
if (IsTTD) {
SecIt = I->getSecond().SecondaryLocalVarData->begin();
llvm::PointerType *SecGlobalRecPtrTy =
CGF.ConvertTypeForMem(SecGlobalRecTy)->getPointerTo();
SecBase = CGF.MakeNaturalAlignPointeeAddrLValue(
Bld.CreatePointerBitCastOrAddrSpaceCast(
I->getSecond().GlobalRecordAddr, SecGlobalRecPtrTy),
SecGlobalRecTy);
}
for (auto &Rec : I->getSecond().LocalVarData) {
bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
llvm::Value *ParValue;
if (EscapedParam) {
const auto *VD = cast<VarDecl>(Rec.first);
LValue ParLVal =
CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
}
LValue VarAddr = CGF.EmitLValueForField(Base, Rec.second.FD);
// Emit VarAddr basing on lane-id if required.
QualType VarTy;
if (Rec.second.IsOnePerTeam) {
VarTy = Rec.second.FD->getType();
} else {
llvm::Value *Ptr = CGF.Builder.CreateInBoundsGEP(
VarAddr.getAddress(CGF).getPointer(),
{Bld.getInt32(0), getNVPTXLaneID(CGF)});
VarTy =
Rec.second.FD->getType()->castAsArrayTypeUnsafe()->getElementType();
VarAddr = CGF.MakeAddrLValue(
Address(Ptr, CGM.getContext().getDeclAlign(Rec.first)), VarTy,
AlignmentSource::Decl);
}
Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
if (!IsInTTDRegion &&
(WithSPMDCheck ||
getExecutionMode() == CGOpenMPRuntimeGPU::EM_Unknown)) {
assert(I->getSecond().IsInSPMDModeFlag &&
"Expected unknown execution mode or required SPMD check.");
if (IsTTD) {
assert(SecIt->second.IsOnePerTeam &&
"Secondary glob data must be one per team.");
LValue SecVarAddr = CGF.EmitLValueForField(SecBase, SecIt->second.FD);
VarAddr.setAddress(
Address(Bld.CreateSelect(IsTTD, SecVarAddr.getPointer(CGF),
VarAddr.getPointer(CGF)),
VarAddr.getAlignment()));
Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
}
Address GlobalPtr = Rec.second.PrivateAddr;
Address LocalAddr = CGF.CreateMemTemp(VarTy, Rec.second.FD->getName());
Rec.second.PrivateAddr = Address(
Bld.CreateSelect(I->getSecond().IsInSPMDModeFlag,
LocalAddr.getPointer(), GlobalPtr.getPointer()),
LocalAddr.getAlignment());
}
if (EscapedParam) {
const auto *VD = cast<VarDecl>(Rec.first);
CGF.EmitStoreOfScalar(ParValue, VarAddr);
I->getSecond().MappedParams->setVarAddr(CGF, VD,
VarAddr.getAddress(CGF));
}
if (IsTTD)
++SecIt;
}
}
for (const ValueDecl *VD : I->getSecond().EscapedVariableLengthDecls) {
// Recover pointer to this function's global record. The runtime will
// handle the specifics of the allocation of the memory.
// Use actual memory size of the record including the padding
// for alignment purposes.
CGBuilderTy &Bld = CGF.Builder;
llvm::Value *Size = CGF.getTypeSize(VD->getType());
CharUnits Align = CGM.getContext().getDeclAlign(VD);
Size = Bld.CreateNUWAdd(
Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
llvm::Value *AlignVal =
llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
Size = Bld.CreateUDiv(Size, AlignVal);
Size = Bld.CreateNUWMul(Size, AlignVal);
// TODO: allow the usage of shared memory to be controlled by
// the user, for now, default to global.
llvm::Value *GlobalRecordSizeArg[] = {
Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_data_sharing_coalesced_push_stack),
GlobalRecordSizeArg);
llvm::Value *GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
GlobalRecValue, CGF.ConvertTypeForMem(VD->getType())->getPointerTo());
LValue Base = CGF.MakeAddrLValue(GlobalRecCastAddr, VD->getType(),
CGM.getContext().getDeclAlign(VD),
AlignmentSource::Decl);
I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD),
Base.getAddress(CGF));
I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(GlobalRecValue);
}
I->getSecond().MappedParams->apply(CGF);
}
void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF,
bool WithSPMDCheck) {
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
return;
const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
if (I != FunctionGlobalizedDecls.end()) {
I->getSecond().MappedParams->restore(CGF);
if (!CGF.HaveInsertPoint())
return;
for (llvm::Value *Addr :
llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_data_sharing_pop_stack),
Addr);
}
if (I->getSecond().GlobalRecordAddr) {
if (!IsInTTDRegion &&
(WithSPMDCheck ||
getExecutionMode() == CGOpenMPRuntimeGPU::EM_Unknown)) {
CGBuilderTy &Bld = CGF.Builder;
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd");
Bld.CreateCondBr(I->getSecond().IsInSPMDModeFlag, ExitBB, NonSPMDBB);
// There is no need to emit line number for unconditional branch.
(void)ApplyDebugLocation::CreateEmpty(CGF);
CGF.EmitBlock(NonSPMDBB);
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_data_sharing_pop_stack),
CGF.EmitCastToVoidPtr(I->getSecond().GlobalRecordAddr));
CGF.EmitBlock(ExitBB);
} else if (!CGM.getLangOpts().OpenMPCUDATargetParallel && IsInTTDRegion) {
assert(GlobalizedRecords.back().RegionCounter > 0 &&
"region counter must be > 0.");
--GlobalizedRecords.back().RegionCounter;
// Emit the restore function only in the target region.
if (GlobalizedRecords.back().RegionCounter == 0) {
QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth(
/*DestWidth=*/16, /*Signed=*/0);
llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar(
Address(GlobalizedRecords.back().UseSharedMemory,
CGM.getContext().getTypeAlignInChars(Int16Ty)),
/*Volatile=*/false, Int16Ty, GlobalizedRecords.back().Loc);
llvm::Value *Args[] = {
llvm::ConstantInt::get(
CGM.Int16Ty,
getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD ? 1 : 0),
IsInSharedMemory};
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_restore_team_static_memory),
Args);
}
} else {
CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_data_sharing_pop_stack),
I->getSecond().GlobalRecordAddr);
}
}
}
}
void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF,
const OMPExecutableDirective &D,
SourceLocation Loc,
llvm::Function *OutlinedFn,
ArrayRef<llvm::Value *> CapturedVars) {
if (!CGF.HaveInsertPoint())
return;
Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
/*Name=*/".zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer());
OutlinedFnArgs.push_back(ZeroAddr.getPointer());
OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
}
void CGOpenMPRuntimeGPU::emitParallelCall(
CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn,
ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
if (!CGF.HaveInsertPoint())
return;
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
emitSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond);
else
emitNonSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond);
}
void CGOpenMPRuntimeGPU::emitNonSPMDParallelCall(
CodeGenFunction &CGF, SourceLocation Loc, llvm::Value *OutlinedFn,
ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
llvm::Function *Fn = cast<llvm::Function>(OutlinedFn);
// Force inline this outlined function at its call site.
Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
/*Name=*/".zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
// ThreadId for serialized parallels is 0.
Address ThreadIDAddr = ZeroAddr;
auto &&CodeGen = [this, Fn, CapturedVars, Loc, &ThreadIDAddr](
CodeGenFunction &CGF, PrePostActionTy &Action) {
Action.Enter(CGF);
Address ZeroAddr =
CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
/*Name=*/".bound.zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
OutlinedFnArgs.push_back(ZeroAddr.getPointer());
OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
emitOutlinedFunctionCall(CGF, Loc, Fn, OutlinedFnArgs);
};
auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF,
PrePostActionTy &) {
RegionCodeGenTy RCG(CodeGen);
llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
llvm::Value *ThreadID = getThreadID(CGF, Loc);
llvm::Value *Args[] = {RTLoc, ThreadID};
NVPTXActionTy Action(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_serialized_parallel),
Args,
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_serialized_parallel),
Args);
RCG.setAction(Action);
RCG(CGF);
};
auto &&L0ParallelGen = [this, CapturedVars, Fn](CodeGenFunction &CGF,
PrePostActionTy &Action) {
CGBuilderTy &Bld = CGF.Builder;
llvm::Function *WFn = WrapperFunctionsMap[Fn];
assert(WFn && "Wrapper function does not exist!");
llvm::Value *ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
// Prepare for parallel region. Indicate the outlined function.
llvm::Value *Args[] = {ID};
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_prepare_parallel),
Args);
// Create a private scope that will globalize the arguments
// passed from the outside of the target region.
CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);
// There's something to share.
if (!CapturedVars.empty()) {
// Prepare for parallel region. Indicate the outlined function.
Address SharedArgs =
CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "shared_arg_refs");
llvm::Value *SharedArgsPtr = SharedArgs.getPointer();
llvm::Value *DataSharingArgs[] = {
SharedArgsPtr,
llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
CGF.EmitRuntimeCall(createNVPTXRuntimeFunction(
OMPRTL_NVPTX__kmpc_begin_sharing_variables),
DataSharingArgs);
// Store variable address in a list of references to pass to workers.
unsigned Idx = 0;
ASTContext &Ctx = CGF.getContext();
Address SharedArgListAddress = CGF.EmitLoadOfPointer(
SharedArgs, Ctx.getPointerType(Ctx.getPointerType(Ctx.VoidPtrTy))
.castAs<PointerType>());
for (llvm::Value *V : CapturedVars) {
Address Dst = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
llvm::Value *PtrV;
if (V->getType()->isIntegerTy())
PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
else
PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
Ctx.getPointerType(Ctx.VoidPtrTy));
++Idx;
}
}
// Activate workers. This barrier is used by the master to signal
// work for the workers.
syncCTAThreads(CGF);
// OpenMP [2.5, Parallel Construct, p.49]
// There is an implied barrier at the end of a parallel region. After the
// end of a parallel region, only the master thread of the team resumes
// execution of the enclosing task region.
//
// The master waits at this barrier until all workers are done.
syncCTAThreads(CGF);
if (!CapturedVars.empty())
CGF.EmitRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_end_sharing_variables));
// Remember for post-processing in worker loop.
Work.emplace_back(WFn);
};
auto &&LNParallelGen = [this, Loc, &SeqGen, &L0ParallelGen](
CodeGenFunction &CGF, PrePostActionTy &Action) {
if (IsInParallelRegion) {
SeqGen(CGF, Action);
} else if (IsInTargetMasterThreadRegion) {
L0ParallelGen(CGF, Action);
} else {
// Check for master and then parallelism:
// if (__kmpc_is_spmd_exec_mode() || __kmpc_parallel_level(loc, gtid)) {
// Serialized execution.
// } else {
// Worker call.
// }
CGBuilderTy &Bld = CGF.Builder;
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
llvm::BasicBlock *SeqBB = CGF.createBasicBlock(".sequential");
llvm::BasicBlock *ParallelCheckBB = CGF.createBasicBlock(".parcheck");
llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall(
createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode)));
Bld.CreateCondBr(IsSPMD, SeqBB, ParallelCheckBB);
// There is no need to emit line number for unconditional branch.
(<