AI/rocm-systems
2
0
Форкнуть 0
Код Задачи Запросы на слияние Действия Пакеты Проекты Релизы Вики Активность
Files
db6baeefb622afafade801ec6d4d55573f2642e2
rocm-systems/rocclr/compiler/lib/backends/common/linker.cpp
T
foreman 00cd00cbee P4 to Git Change 1054428 by yaxunl@yaxunl_stg_win50 on 2014/07/11 11:42:03 
ECR #377625 - Allow function call for function with internal linkage.

	Internal linkage correponds to static function in C. The function could be large and should be allowed to be not inlined.

Affected files ...

... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/linker.cpp#106 edit
2014-07-11 11:51:37 -04:00

1300 строки
42 KiB
C++
Исходник Ответственный История

//
// Copyright (c) 2008 Advanced Micro Devices, Inc. All rights reserved.
//
// TODO: The entire linker implementation should be a pass in LLVM and
// the code in the compiler library should only call this pass.
#include "top.hpp"
#include "library.hpp"
#include "linker.hpp"
#include "os/os.hpp"
#include "thread/monitor.hpp"
#include "utils/libUtils.h"
#include "utils/options.hpp"
#include "utils/target_mappings.h"
#include "acl.h"
#include "llvm/Instructions.h"
#include "llvm/Linker.h"
#include "llvm/GlobalValue.h"
#include "llvm/GlobalVariable.h"
#include "llvm/AMDResolveLinker.h"
#include "llvm/AMDPrelinkOpt.h"
#include "llvm/ADT/Triple.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/Analysis/AMDLocalArrayUsage.h"
#include "llvm/Analysis/CodeMetrics.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/Passes.h"
#include "llvm/Analysis/Verifier.h"
#include "llvm/Bitcode/ReaderWriter.h"
#include "llvm/CodeGen/LinkAllAsmWriterComponents.h"
#include "llvm/CodeGen/LinkAllCodegenComponents.h"
#if 1 || LLVM_TRUNK_INTEGRATION_CL >= 2270
#else
#include "llvm/CodeGen/ObjectCodeEmitter.h"
#endif
#include "llvm/Config/config.h"
#include "llvm/MC/SubtargetFeature.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/FileUtilities.h"
#include "llvm/Support/FormattedStream.h"
#include "llvm/Support/Host.h"
#include "llvm/Support/ManagedStatic.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/Support/Path.h"
#include "llvm/Support/PluginLoader.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Signals.h"
#include "llvm/Support/system_error.h"
#include "llvm/Support/TargetRegistry.h"
#include "llvm/Support/TargetSelect.h"
#include "llvm/DataLayout.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Transforms/IPO.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/ValueSymbolTable.h"
#ifdef _DEBUG
#include "llvm/Assembly/Writer.h"
#endif
// need to undef DEBUG before using DEBUG macro in llvm/Support/Debug.h
#ifdef DEBUG
#undef DEBUG
#endif
#include "llvm/Support/Debug.h"
#include <cassert>
#include <cstdlib>
#include <cstdio>
#include <cstring>
#include <string>
#include <iostream>
#include <fstream>
#include <sstream>
#include <list>
#include <map>
#include <set>
#ifdef _WIN32
#include <windows.h>
#endif // _WIN32
#ifdef DEBUG_TYPE
#undef DEBUG_TYPE
#endif
#define DEBUG_TYPE "ocl_linker"
static const char* OptionMaskFName = "__option_mask";
extern llvm::Module*
clpVectorExpansion(llvm::Module *srcModules[], std::string &errorMsg);
namespace amd {
namespace {
using namespace llvm;
// LoadFile - Read the specified bitcode file in and return it. This routine
// searches the link path for the specified file to try to find it...
//
inline llvm::Module*
LoadFile(const std::string &Filename, LLVMContext& Context)
{
bool Exists;
if (sys::fs::exists(Filename, Exists) || !Exists) {
// dbgs() << "Bitcode file: '" << Filename.c_str() << "' does not exist.\n";
return 0;
}
llvm::Module* M;
std::string ErrorMessage;
OwningPtr<MemoryBuffer> Buffer;
if (error_code ec = MemoryBuffer::getFileOrSTDIN(Filename, Buffer)) {
// Error
M = NULL;
}
else {
M = ParseBitcodeFile(Buffer.get(), Context, &ErrorMessage);
}
return M;
}
inline llvm::Module*
LoadLibrary(const std::string& libFile, LLVMContext& Context, MemoryBuffer** Buffer) {
bool Exists;
if (sys::fs::exists(libFile, Exists) || !Exists) {
// dbgs() << "Bitcode file: '" << Filename.c_str() << "' does not exist.\n";
return 0;
}
llvm::Module* M = NULL;
std::string ErrorMessage;
static Monitor mapLock;
static std::map<std::string, void*> FileMap;
MemoryBuffer* statBuffer;
{
ScopedLock sl(mapLock);
statBuffer = (MemoryBuffer*) FileMap[libFile];
if (statBuffer == NULL) {
OwningPtr<MemoryBuffer> PtrBuffer;
if (error_code ec = MemoryBuffer::getFileOrSTDIN(libFile, PtrBuffer)) {
// Error
return NULL;
}
else
statBuffer = PtrBuffer.take();
M = ParseBitcodeFile(statBuffer, Context, &ErrorMessage);
FileMap[libFile] = statBuffer;
}
}
*Buffer = MemoryBuffer::getMemBufferCopy(StringRef(statBuffer->getBufferStart(), statBuffer->getBufferSize()), "");
if ( *Buffer ) {
M = getLazyBitcodeModule(*Buffer, Context, &ErrorMessage);
if (!M) {
delete *Buffer;
*Buffer = 0;
}
}
return M;
}
// Load bitcode libary from an array of const char. This assumes that
// the array has a valid ending zero !
llvm::Module*
LoadLibrary(const char* libBC, size_t libBCSize,
LLVMContext& Context, MemoryBuffer** Buffer)
{
llvm::Module* M = 0;
std::string ErrorMessage;
*Buffer = MemoryBuffer::getMemBuffer(StringRef(libBC, libBCSize), "");
if ( *Buffer ) {
M = getLazyBitcodeModule(*Buffer, Context, &ErrorMessage);
if (!M) {
delete *Buffer;
*Buffer = 0;
}
}
return M;
}
static std::set<std::string> *getAmdRtFunctions()
{
std::set<std::string> *result = new std::set<std::string>();
for (size_t i = 0; i < sizeof(amdRTFuns)/sizeof(amdRTFuns[0]); ++i)
result->insert(amdRTFuns[i]);
return result;
}
// Remove NoInline attribute to functions in a module
void
RemoveNoInlineAttr(llvm::Module* M)
{
LLVMContext &Context = M->getContext();
for (llvm::Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
I->removeFnAttr(Attributes::get(Context, Attributes::NoInline));
}
}
bool
IsKernel(llvm::Function* F)
{
return F->getName().startswith("__OpenCL_") &&
F->getName().endswith("_kernel");
}
// Add NoInline attribute to functions in a module
void
AddNoInlineAttr(llvm::Module* M)
{
LLVMContext &Context = M->getContext();
for (llvm::Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
if (I->hasName() &&
!I->isDeclaration() &&
!I->isIntrinsic() &&
!I->getName().startswith("__amdil") &&
!I->getFnAttributes().hasAttribute(Attributes::AlwaysInline) &&
!IsKernel(I)) {
DEBUG_WITH_TYPE("noinline",
dbgs() << "[Candidate] " << I->getName() << '\n');
I->addFnAttr(Attributes::NoInline);
}
}
}
unsigned
CountCallSites(llvm::Function* F, llvm::Module* M,
std::map<llvm::Function*, unsigned>& counts) {
std::map<llvm::Function*, unsigned>::iterator iter = counts.find(F);
if (iter != counts.end())
return iter->second;
unsigned numCalled = 0;
for (Function::use_iterator I = F->use_begin(), E = F->use_end(); I != E;
++I) {
User *UI = *I;
if (isa<CallInst>(UI) || isa<InvokeInst>(UI)) {
ImmutableCallSite CS(cast<Instruction>(UI));
Function* caller = const_cast<llvm::Function*>(CS.getCaller());
unsigned callerCount = CountCallSites(caller, M, counts);
if (caller->getFnAttributes().hasAttribute(Attributes::NoInline) &&
callerCount > 0)
numCalled++;
else
numCalled += callerCount;
}
}
if (numCalled == 0 && IsKernel(F))
numCalled = 1;
counts[F] = numCalled;
return numCalled;
}
unsigned
CalculateSize(llvm::Function* F, llvm::Module* M,
std::map<llvm::Function*, unsigned>& sizes) {
std::map<llvm::Function*, unsigned>::iterator iter = sizes.find(F);
if (iter != sizes.end())
return iter->second;
CodeMetrics metrics;
metrics.analyzeFunction(F);
unsigned size = metrics.NumInsts;
for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I) {
for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE; ++BI) {
if (CallInst* callInst = dyn_cast<CallInst>(BI)) {
Function* called = callInst->getCalledFunction();
if (called &&
!called->getFnAttributes().hasAttribute(Attributes::NoInline))
size += CalculateSize(called, M, sizes);
}
}
}
sizes[F] = size;
return size;
}
// Identify functions with image arguments.
// Callers may pass images with different resource ids to the callee.
// Currently pointer manager cannot handle this.
// ToDo: Should remove this after we find a way to handle image in function.
bool
IsImageFunc(Function* F) {
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
++I) {
if (PointerType *PT = dyn_cast<PointerType>(I->getType())) {
if (PT->getAddressSpace() != 1) {
continue;
}
if (StructType *ST = dyn_cast<StructType>(PT->getElementType())) {
if (ST->getName().startswith("struct._image")) {
DEBUG_WITH_TYPE("noinline", dbgs() << "[image function] " <<
F->getName() << " inline\n");
return true;
}
}
}
}
return false;
}
bool
MustInline(Function* F) {
if (F->getFnAttributes().hasAttribute(Attributes::AlwaysInline))
return true;
return IsImageFunc(F);
}
bool
CallerMustInline(Function* F) {
return IsImageFunc(F);
}
bool
CallsNoInlineFunc(Function* F, std::map<Function*, bool>& work) {
DEBUG_WITH_TYPE("noinline", dbgs() << "[CallsNoInlineFunc:" << F->getName() << " ");
std::map<Function*, bool>::iterator loc = work.find(F);
if (loc != work.end()) {
DEBUG_WITH_TYPE("noinline", dbgs() << loc->second << "(cached)]\n");
return loc->second;
}
for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I) {
for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE; ++BI) {
if (CallInst* callInst = dyn_cast<CallInst>(BI)) {
Function* called = callInst->getCalledFunction();
if (called) {
if (called->getFnAttributes().hasAttribute(Attributes::NoInline) ||
CallerMustInline(called) ||
CallsNoInlineFunc(called, work)) {
work[F] = true;
DEBUG_WITH_TYPE("noinline", dbgs() << "1(" << called->getName() <<")]\n");
return true;
}
}
}
}
}
work[F] = false;
DEBUG_WITH_TYPE("noinline", dbgs() << "0]\n");
return false;
}
bool
CalledByNoInlineFunc(Function* F, std::map<Function*, bool>& work) {
DEBUG_WITH_TYPE("noinline", dbgs() << "[CalledByNoInlineFunc: " << F->getName() << " ");
std::map<Function*, bool>::iterator loc = work.find(F);
if (loc != work.end()) {
DEBUG_WITH_TYPE("noinline", dbgs() << loc->second << "]\n");
return loc->second;
}
for (Function::use_iterator I = F->use_begin(), E = F->use_end(); I != E;
++I) {
User *UI = *I;
if (isa<CallInst>(UI) || isa<InvokeInst>(UI)) {
ImmutableCallSite CS(cast<Instruction>(UI));
Function* caller = const_cast<llvm::Function*>(CS.getCaller());
if (caller->getFnAttributes().hasAttribute(Attributes::NoInline) ||
CalledByNoInlineFunc(caller, work)) {
work[F] = true;
DEBUG_WITH_TYPE("noinline", dbgs() << "1(" << caller->getName() <<")]\n");
return true;
}
}
}
work[F] = false;
DEBUG_WITH_TYPE("noinline", dbgs() << "0]\n");
return false;
}
bool
CanBeNoInline(Function* F, std::map<Function*, bool>& callsNoInline,
std::map<Function*, bool>& calledByNoInline, bool allowMultiLevelCall) {
return !MustInline(F) && (allowMultiLevelCall ||
(!CallsNoInlineFunc(F, callsNoInline) &&
!CalledByNoInlineFunc(F, calledByNoInline)));
}
struct CostInfo {
unsigned count;
unsigned size;
unsigned cost;
};
unsigned
CalculateMaxKernelSize(llvm::Module* M) {
std::map<llvm::Function*, unsigned> sizes;
unsigned maxSize = 0;
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
if (IsKernel(I)) {
unsigned kernelSize = CalculateSize(I, M, sizes);
DEBUG_WITH_TYPE("noinlines", dbgs() << "[Kernel size] " <<
I->getName() << " : " << kernelSize << '\n');
if (maxSize < kernelSize)
maxSize = kernelSize;
}
}
return maxSize;
}
void
RefineNoInlineAttr(llvm::Module* M, int thresh, int sizeThresh,
int kernelSizeThresh, bool allowMultiLevelCall)
{
if (thresh == 0 && sizeThresh == 0)
return;
std::set<Function*> candidates;
LLVMContext &Context = M->getContext();
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
if (I->getFnAttributes().hasAttribute(Attributes::NoInline)) {
candidates.insert(I);
I->removeFnAttr(Attributes::get(Context, Attributes::NoInline));
}
}
unsigned maxKernelSize = CalculateMaxKernelSize(M);
if (maxKernelSize < unsigned(kernelSizeThresh))
return;
while (true) {
std::map<Function*, unsigned> counts;
std::map<Function*, unsigned> sizes;
std::map<Function*, CostInfo> costInfos;
std::map<Function*, bool > callsNoInline;
std::map<Function*, bool > calledByNoInline;
for (std::set<Function*>::iterator I = candidates.begin(),
E = candidates.end(); I != E; ++I) {
Function* F = *I;
unsigned count = CountCallSites(F, M, counts);
if (count > 0 && CanBeNoInline(F, callsNoInline, calledByNoInline,
allowMultiLevelCall)) {
unsigned size = CalculateSize(F, M, sizes);
if (size > unsigned(sizeThresh)) {
CostInfo& info = costInfos[F];
info.count = count;
info.size = size;
info.cost = (count - 1) * size;
DEBUG_WITH_TYPE("noinline", dbgs() << F->getName() <<
" : " << count - 1 << " * " << size << " = " << (count-1) * size <<
"\n");
}
}
}
int maxCost = -1;
Function* select = NULL;
for (std::map<Function*, CostInfo>::iterator I = costInfos.begin(),
E = costInfos.end(); I != E; ++I) {
CostInfo& info = I->second;
if (int(info.cost) > maxCost) {
maxCost = int(info.cost);
select = I->first;
}
}
if (select == NULL || maxCost < thresh)
break;
CostInfo& info = costInfos[select];
DEBUG_WITH_TYPE("noinlines", llvm::dbgs() << "select " << select->getName().str()
<< " cost = " << info.count << " x " << info.size << " = " <<
info.cost << "\n");
select->addFnAttr(Attributes::NoInline);
candidates.erase(select);
if (candidates.empty())
break;
}
if (getenv("AMD_OCL_INLINE")) {
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
if(I->hasName() && strstr(getenv("AMD_OCL_INLINE"),
I->getName().str().c_str())) {
I->removeFnAttr(Attributes::get(Context, Attributes::NoInline));
printf("force inline %s\n", I->getName().data());
}
}
}
if (getenv("AMD_OCL_NOINLINE")) {
for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
if(I->hasName() && strstr(getenv("AMD_OCL_NOINLINE"),
I->getName().str().c_str())) {
I->addFnAttr(Attributes::NoInline);
printf("force noinline %s\n", I->getName().data());
}
}
}
}
} // unnamed namespace
} // namespace amd
// create a llvm function which simply returns the given mask
static void createConstIntFunc(const char* fname,
int mask,
llvm::Module* module)
{
llvm::LLVMContext& context = module->getContext();
llvm::Type* int32Ty = llvm::Type::getInt32Ty(context);
llvm::FunctionType* fType = llvm::FunctionType::get(int32Ty, false);
llvm::Function* function
= llvm::cast<llvm::Function>(module->getOrInsertFunction(fname, fType));
function->setDoesNotThrow();
function->setDoesNotAccessMemory();
function->addFnAttr(llvm::Attributes::AlwaysInline);
llvm::BasicBlock* bb = llvm::BasicBlock::Create(context, "entry", function);
llvm::Value* retVal = llvm::ConstantInt::get(int32Ty, mask);
llvm::ReturnInst* retInst = llvm::ReturnInst::Create(context, retVal);
bb->getInstList().push_back(retInst);
assert(!verifyFunction(*function) && "verifyFunction failed");
}
// create a llvm function that returns a mask of several compile options
// which are used by the built-in library
void amdcl::OCLLinker::createOptionMaskFunction(llvm::Module* module)
{
unsigned mask = 0;
if (Options()->oVariables->NoSignedZeros) {
mask |= MASK_NO_SIGNED_ZEROES;
}
if (Options()->oVariables->UnsafeMathOpt) {
mask |= MASK_UNSAFE_MATH_OPTIMIZATIONS;
mask |= MASK_NO_SIGNED_ZEROES;
}
if (Options()->oVariables->FiniteMathOnly) {
mask |= MASK_FINITE_MATH_ONLY;
}
if (Options()->oVariables->FastRelaxedMath) {
mask |= MASK_FAST_RELAXED_MATH;
mask |= MASK_FINITE_MATH_ONLY;
mask |= MASK_UNSAFE_MATH_OPTIMIZATIONS;
mask |= MASK_NO_SIGNED_ZEROES;
}
if (Options()->oVariables->UniformWorkGroupSize) {
mask |= MASK_UNIFORM_WORK_GROUP_SIZE;
}
createConstIntFunc(OptionMaskFName, mask, module);
}
// Create functions that returns true or false for some features which
// are used by the built-in library
void amdcl::OCLLinker::createASICIDFunctions(llvm::Module* module)
{
if (!isAMDILTarget(Elf()->target))
return;
uint64_t features = aclGetChipOptions(Elf()->target);
llvm::StringRef chip(aclGetChip(Elf()->target));
llvm::StringRef family(aclGetFamily(Elf()->target));
createConstIntFunc("__amdil_have_hw_fma32",
chip == "Cypress"
|| chip == "Cayman"
|| family == "SI"
|| family == "CI"
|| family == "KV"
|| family == "TN"
|| family == "VI"
|| family == "CZ",
module);
createConstIntFunc("__amdil_have_fast_fma32",
chip == "Cypress"
|| chip == "Cayman"
|| chip == "Tahiti"
|| chip == "Hawaii",
module);
createConstIntFunc("__amdil_have_bitalign", !!(features & F_EG_BASE), module);
createConstIntFunc("__amdil_is_cypress", chip == "Cypress", module);
createConstIntFunc("__amdil_is_ni",
chip == "Cayman"
|| family == "TN",
module);
createConstIntFunc("__amdil_is_gcn",
family == "SI"
|| family == "CI"
|| family == "VI"
|| family == "KV"
|| family == "CZ",
module);
}
bool
amdcl::OCLLinker::linkWithModule(
llvm::Module* Dst, llvm::Module* Src,
std::map<const llvm::Value*, bool> *ModuleRefMap)
{
#ifndef NDEBUG
if (Options()->oVariables->EnableDebugLinker) {
llvm::DebugFlag = true;
llvm::setCurrentDebugType(DEBUG_TYPE);
}
#endif
std::string ErrorMessage;
if (llvm::linkWithModule(Dst, Src, ModuleRefMap, &ErrorMessage)) {
DEBUG(llvm::dbgs() << "Error: " << ErrorMessage << "\n");
BuildLog() += "\nInternal Error: linking libraries failed!\n";
LogError("linkWithModule(): linking bc libraries failed!");
return true;
}
return false;
}
static void delete_llvm_module(llvm::Module *a)
{
delete a;
}
bool
amdcl::OCLLinker::linkLLVMModules(std::vector<llvm::Module*> &libs)
{
// Load input modules first
bool Failed = false;
for (size_t i = 0; i < libs.size(); ++i) {
std::string ErrorMsg;
if (!libs[i]) {
char ErrStr[128];
sprintf(ErrStr,
"Error: cannot load input %d bc for linking: %s\n",
(int)i, ErrorMsg.c_str());
BuildLog() += ErrStr;
Failed = true;
break;
}
if (Options()->isDumpFlagSet(amd::option::DUMP_BC_ORIGINAL)) {
std::string MyErrorInfo;
char buf[128];
sprintf(buf, "_original%d.bc", (int)i);
std::string fileName = Options()->getDumpFileName(buf);
llvm::raw_fd_ostream outs(fileName.c_str(), MyErrorInfo,
llvm::raw_fd_ostream::F_Binary);
if (MyErrorInfo.empty())
llvm::WriteBitcodeToFile(libs[i], outs);
else
printf(MyErrorInfo.c_str());
}
}
if (!Failed) {
// Link input modules together
for (size_t i = 0; i < libs.size(); ++i) {
DEBUG(llvm::dbgs() << "LinkWithModule " << i << ":\n");
if (amdcl::OCLLinker::linkWithModule(LLVMBinary(), libs[i], NULL)) {
Failed = true;
}
}
}
if (Failed) {
delete LLVMBinary();
}
std::for_each(libs.begin(), libs.end(), std::ptr_fun(delete_llvm_module));
libs.clear();
return Failed;
}
void amdcl::OCLLinker::fixupOldTriple(llvm::Module *module)
{
llvm::Triple triple(module->getTargetTriple());
// Bug 9357: "amdopencl" used to be a hacky "OS" that was Linux or Windows
// depending on the host. It only really matters for x86. If we are trying to
// use an old binary module still using the old triple, replace it with a new
// one.
if (triple.getOSName() == "amdopencl") {
if (triple.getArch() == llvm::Triple::amdil ||
triple.getArch() == llvm::Triple::amdil64) {
triple.setOS(llvm::Triple::UnknownOS);
} else {
llvm::Triple hostTriple(llvm::sys::getDefaultTargetTriple());
triple.setOS(hostTriple.getOS());
}
triple.setEnvironment(llvm::Triple::AMDOpenCL);
module->setTargetTriple(triple.str());
}
}
//Modify module for targets before linking.
//Report error by buildLog.
//Return false on error.
static bool fixUpModule(llvm::Module *M,
llvm::StringRef TargetTriple,
llvm::StringRef TargetLayout,
bool RunSPIRLoader,
bool DemangleBuiltins,
bool RunEDGAdapter,
bool SetSPIRCallingConv,
bool RunX86Adpater,
bool RunPrintfRuntimeBinding,
bool RunPrintfCpuLowering,
bool RunLowerEnqueueKernel,
const amd::option::OptionVariables *oVariables,
std::string& buildLog) {
llvm::PassManager Passes;
DEBUG_WITH_TYPE("linkTriple", llvm::dbgs() <<
"[fixUpModule] module triple: " << M->getTargetTriple() <<
" target triple: " << TargetTriple);
llvm::Triple triple(M->getTargetTriple());
#if OPENCL_MAJOR < 2
if (triple.getArch() == llvm::Triple::spir ||
triple.getArch() == llvm::Triple::spir64 ||
triple.getArch() == llvm::Triple::x86 ||
triple.getArch() == llvm::Triple::x86_64 ||
M->getTargetTriple().empty())
#endif
{
M->setTargetTriple(TargetTriple);
M->setDataLayout(TargetLayout);
}
#if OPENCL_MAJOR < 2
if (M->getTargetTriple() != TargetTriple) {
//ToDo: There is bug 9996 in compiler library about converting BIF30 to BIF21
//which causes regressions in ocltst if the following check is enabled.
//Fix the bugs then enable the following check
#if 0
assert(0 && "Inconsistent module and library target");
buildLog += "Internal Error: failed to link modules correctly.\n";
return false;
#else
LogWarning("Inconsistent module and library target");
return true;
#endif
}
#endif
Passes.add(new llvm::DataLayout(M));
if (RunPrintfRuntimeBinding == true || RunPrintfCpuLowering == true)
Passes.add(llvm::createAMDPrintfRuntimeBinding(RunPrintfCpuLowering));
if (oVariables->LowerAtomics)
Passes.add(llvm::createAMDLowerAtomicsPass());
if (oVariables->LowerPipeBuiltins)
Passes.add(llvm::createAMDLowerPipeBuiltinsPass());
if (RunEDGAdapter) {
assert(!RunSPIRLoader);
Passes.add(llvm::createAMDEDGToIA64TranslatorPass(SetSPIRCallingConv));
}
if (RunSPIRLoader) {
assert(!RunEDGAdapter);
Passes.add(llvm::createSPIRLoader(DemangleBuiltins));
}
if (RunX86Adpater) {
// One of them should run before the AMDX86Adapter Pass.
assert(RunSPIRLoader || RunEDGAdapter);
Passes.add(llvm::createAMDX86AdapterPass());
}
if (RunLowerEnqueueKernel) {
Passes.add(llvm::createAMDLowerEnqueueKernelPass());
Passes.add(llvm::createAMDGenerateDevEnqMetadataPass());
}
Passes.run(*M);
return true;
}
static void CheckSPIRVersion(const llvm::Module *M,
const aclTargetInfo& Target) {
const llvm::NamedMDNode *SPIRVersion
= M->getNamedMetadata("opencl.spir.version");
assert(SPIRVersion);
// When multiple llvm modules are linked together to create a single module
// Metadata's of llvm modules are added into destination module and
// it results in a more than one SPIR MDNode value.
// Marking this fix as temporary and it will be tracked in bugzilla id 9775
if (SPIRVersion->getNumOperands() > 1)
LogWarning("\nIncorrect SPIR MDNode value");
assert(SPIRVersion->getNumOperands() >= 1);
const llvm::MDNode *VersionMD = SPIRVersion->getOperand(0);
assert(VersionMD->getNumOperands() == 2);
const llvm::ConstantInt *CMajor
= llvm::cast<llvm::ConstantInt>(VersionMD->getOperand(0));
assert(CMajor->getType()->getIntegerBitWidth() == 32);
unsigned VersionMajor = CMajor->getZExtValue();
switch (VersionMajor) {
case 1:
break;
case 2:
assert(!isAMDILTarget(Target));
break;
default:
llvm_unreachable("Unknown SPIR version");
break;
}
}
// On 64 bit device, aclBinary target is set to 64 bit by default. When 32 bit
// LLVM or SPIR binary is loaded, aclBinary target needs to be modified to
// match LLVM or SPIR bitness.
// Returns false on error.
static bool
checkAndFixAclBinaryTarget(llvm::Module* module, aclBinary* elf,
std::string& buildLog) {
if (module->getTargetTriple().empty()) {
LogWarning("Module has no target triple");
return true;
}
llvm::Triple triple(module->getTargetTriple());
const char* newArch = NULL;
if (elf->target.arch_id == aclAMDIL64 &&
(triple.getArch() == llvm::Triple::amdil ||
triple.getArch() == llvm::Triple::spir))
newArch = "amdil";
else if (elf->target.arch_id == aclX64 &&
(triple.getArch() == llvm::Triple::x86 ||
triple.getArch() == llvm::Triple::spir))
newArch = "x86";
else if (elf->target.arch_id == aclHSAIL64 &&
(triple.getArch() == llvm::Triple::hsail ||
triple.getArch() == llvm::Triple::spir))
newArch = "hsail";
if (newArch != NULL) {
acl_error errorCode;
elf->target = aclGetTargetInfo(newArch, aclGetChip(elf->target),
&errorCode);
if (errorCode != ACL_SUCCESS) {
assert(0 && "Invalid arch id or chip id in elf target");
buildLog += "Internal Error: failed to link modules correctlty.\n";
return false;
}
}
reinterpret_cast<amd::option::Options*>(elf->options)->libraryType_ =
getLibraryType(&elf->target);
// Check consistency between module triple and aclBinary target
if (elf->target.arch_id == aclAMDIL64 &&
(triple.getArch() == llvm::Triple::amdil64 ||
triple.getArch() == llvm::Triple::spir64))
return true;
if (elf->target.arch_id == aclAMDIL &&
(triple.getArch() == llvm::Triple::amdil ||
triple.getArch() == llvm::Triple::spir))
return true;
if (elf->target.arch_id == aclHSAIL64 &&
(triple.getArch() == llvm::Triple::hsail_64 ||
triple.getArch() == llvm::Triple::spir64))
return true;
if (elf->target.arch_id == aclHSAIL &&
(triple.getArch() == llvm::Triple::hsail ||
triple.getArch() == llvm::Triple::spir))
return true;
if (elf->target.arch_id == aclX64 &&
(triple.getArch() == llvm::Triple::x86_64 ||
triple.getArch() == llvm::Triple::spir64))
return true;
if (elf->target.arch_id == aclX86 &&
(triple.getArch() == llvm::Triple::x86 ||
triple.getArch() == llvm::Triple::spir))
return true;
DEBUG_WITH_TYPE("linkTriple", llvm::dbgs() <<
"[checkAndFixAclBinaryTarget] " <<
" aclBinary target: " << elf->target.arch_id <<
" chipId: " << elf->target.chip_id <<
" module triple: " << module->getTargetTriple() <<
'\n');
//ToDo: There is bug 9996 in compiler library about converting BIF30 to BIF21
//which causes regressions in ocltst if the following check is enabled.
//Fix the bugs then enable the following check
#if 0
assert(0 && "Inconsistent LLVM target and elf target");
buildLog += "Internal Error: failed to link modules correctlty.\n";
return false;
#else
LogWarning("Inconsistent LLVM target and elf target");
return true;
#endif
}
int
amdcl::OCLLinker::link(llvm::Module* input, std::vector<llvm::Module*> &libs)
{
bool IsGPUTarget = isGpuTarget(Elf()->target);
uint64_t start_time = 0ULL, time_link = 0ULL, time_prelinkopt = 0ULL;
if (Options()->oVariables->EnableBuildTiming) {
start_time = amd::Os::timeNanos();
}
fixupOldTriple(input);
if (!checkAndFixAclBinaryTarget(input, Elf(), BuildLog()))
return 1;
int ret = 0;
if (Options()->oVariables->UseJIT) {
hookup_.amdrtFunctions = amd::getAmdRtFunctions();
} else {
hookup_.amdrtFunctions = NULL;
}
if (Options()->isOptionSeen(amd::option::OID_LUThreshold) || !IsGPUTarget) {
setUnrollScratchThreshold(Options()->oVariables->LUThreshold);
} else {
setUnrollScratchThreshold(500);
}
setGPU(IsGPUTarget);
setPreLinkOpt(false);
// We are doing whole program optimization
setWholeProgram(true);
llvmbinary_ = input;
if ( !LLVMBinary() ) {
BuildLog() += "Internal Error: cannot load bc application for linking\n";
return 1;
}
if (linkLLVMModules(libs)) {
BuildLog() += "Internal Error: failed to link modules correctlty.\n";
return 1;
}
// Don't link in built-in libraries if we are only creating the library.
if (Options()->oVariables->clCreateLibrary) {
return 0;
}
if (Options()->isDumpFlagSet(amd::option::DUMP_BC_ORIGINAL)) {
std::string MyErrorInfo;
std::string fileName = Options()->getDumpFileName("_original.bc");
llvm::raw_fd_ostream outs(fileName.c_str(), MyErrorInfo, llvm::raw_fd_ostream::F_Binary);
if (MyErrorInfo.empty())
WriteBitcodeToFile(LLVMBinary(), outs);
else
printf(MyErrorInfo.c_str());
}
std::vector<llvm::Module*> LibMs;
// The AMDIL GPU libraries include 32 bit specific, 64 bit specific and common
// libraries. The common libraries do not have target triple. A search is
// performed to find the first library containing non-empty target triple
// and use it for translating SPIR.
amd::LibraryDescriptor LibDescs[
amd::LibraryDescriptor::MAX_NUM_LIBRARY_DESCS];
int sz;
std::string LibTargetTriple;
std::string LibDataLayout;
if (amd::getLibDescs(Options()->libraryType_, LibDescs, sz) != 0) {
// FIXME: If we error here, we don't clean up, so we crash in debug build
// on compilerfini().
BuildLog() += "Internal Error: finding libraries failed!\n";
return 1;
}
for (int i=0; i < sz; i++) {
llvm::MemoryBuffer* Buffer = 0;
llvm::Module* Library = amd::LoadLibrary(LibDescs[i].start, LibDescs[i].size, Context(), &Buffer);
DEBUG(llvm::dbgs() << "Loaded library " << i << "\n");
if ( !Library ) {
BuildLog() += "Internal Error: cannot load library!\n";
delete LLVMBinary();
for (int j = 0; j < i; ++j) {
delete LibMs[j];
}
LibMs.clear();
return 1;
#ifndef NDEBUG
} else {
if ( llvm::verifyModule( *Library ) ) {
BuildLog() += "Internal Error: library verification failed!\n";
exit(1);
}
#endif
}
DEBUG_WITH_TYPE("linkTriple", llvm::dbgs() << "Library[" << i << "] " <<
Library->getTargetTriple() << ' ' << Library->getDataLayout() << '\n');
// Find the first library whose target triple is not empty.
if (LibTargetTriple.empty() && !Library->getTargetTriple().empty()) {
LibTargetTriple = Library->getTargetTriple();
LibDataLayout = Library->getDataLayout();
}
LibMs.push_back(Library);
}
// Check consistency of target and data layout
assert (!LibTargetTriple.empty() && "At least one library should have triple");
#ifndef NDEBUG
for (size_t i = 0, e = LibMs.size(); i < e; ++i) {
if (LibMs[i]->getTargetTriple().empty())
continue;
assert (LibMs[i]->getTargetTriple() == LibTargetTriple &&
"Library target triple should match");
assert (LibMs[i]->getDataLayout() == LibDataLayout &&
"Library data layout should match");
}
#endif
// Under various situations, the LLVM dialect used in the kernel
// module does not match the dialect used in the builtin library. We
// need to fix-up the kernel module to eliminate this mismatch.
//
// SPIRLoader is required to consume a SPIR kernel:
// SPIR 1.2 on all targets.
// SPIR 2.0 on x86 and HSAIL only.
//
// The AMDIL libary is compiled by EDG, and hence it does not use
// the SPIR mangling scheme. To allow a SPIR 1.2 kernel to link with
// this library, the SPIRLoader must fix the mangling in the kernel.
//
// EDGAdapter is required to consume a non-SPIR (EDG) kernel on x86
// and HSAIL targets. The builtins library for these targets are
// built by Clang, but OpenCL 1.2 kernels are compiled by EDG.
//
// A non-SPIR kernel module is not expected on the HSAIL target in a
// normal OpenCL 2.0 build. We should actually flag an error if this
// occurs, but we let it through to facilitate custom builds created
// to test this combination. In this situation, the EDGAdapter must
// additionally set the calling conventions correctly, because the
// HSAIL library is in SPIR format.
//
// RunX86Adpater is required to run only on the CPU path. It is
// expected to the solve the link issues between the user kernel
// (SPIR/EDG) vs. Clang compiled x86 builtins library.
// Enabled for:
bool RunSPIRLoader = false; // SPIR -> x86/HSAIL/AMDIL
bool DemangleBuiltins = false; // SPIR -> AMDIL
bool RunEDGAdapter = false; // EDG -> x86/HSAIL
bool SetSPIRCallingConv = false; // EDG -> HSAIL
bool RunX86Adapter = false; // SPIR/EDG -> x86
bool RunLowerEnqueueKernel = false;
bool RunPrintfRuntimeBinding = false;
bool RunPrintfCpuLowering = false;
bool LowerToPreciseFunctions = false;
llvm::Triple ModuleTriple(LLVMBinary()->getTargetTriple());
bool isSPIRModuleTriple = ((ModuleTriple.getArch() == llvm::Triple::spir) ||
(ModuleTriple.getArch() == llvm::Triple::spir64));
if(isSPIRModuleTriple) {
CheckSPIRVersion(LLVMBinary(), Elf()->target);
RunSPIRLoader = true;
#if OPENCL_MAJOR >= 2 // this will become default
DemangleBuiltins |= isAMDILTarget(Elf()->target);
#ifdef BUILD_HSA_TARGET // special case for HSA build
DemangleBuiltins |= isHSAILTarget(Elf()->target);
#endif
// Never demangle for x86 target on 200 build.
#else // OpenCL 1.2 build (this will go away)
DemangleBuiltins = true;
#endif
} else {
#if OPENCL_MAJOR >= 2
// Decide if we need to adapt the non-SPIR (EDG) kernel module.
//
// FIXME: Remove the #ifdef when x86 and HSAIL libraries are
// always built by Clang.
#ifndef BUILD_HSA_TARGET
// Run the adapter for HSAIL, only if this is an ORCA build!
//
// On an HSA build, the HSAIL library is always built with EDG.
// This assumption must match the settings in
// "opencl/library/hsa/hsail/build/Makefile.hsail"
RunEDGAdapter |= isHSAILTarget(Elf()->target);
#endif
// HSAIL requires SPIR calling conventions since the library is in
// SPIR format. This doesn't matter if the EDGAdapter is not run.
SetSPIRCallingConv = isHSAILTarget(Elf()->target);
// Run the EDG Adapter if OPENCL_MAJOR >= 2 and for x86 target.
RunEDGAdapter |= isCpuTarget(Elf()->target);
#endif // OPENCL_MAJOR >= 2
}
// It should run for both EDG generated LLVM IR and SPIR for x86 path.
// FIXME: Remove the #ifdef when x86 is always built by Clang on
// OpenCL 1.2 builds.
#if OPENCL_MAJOR >=2
RunX86Adapter = isCpuTarget(Elf()->target);
RunLowerEnqueueKernel = isSPIRModuleTriple;
// For HSAIL targets, when the option -cl-fp32-correctly-rounded-divide-sqrt
// lower divide and sqrt functions to precise HSAIL builtin library functions.
LowerToPreciseFunctions = (isHSAILTarget(Elf()->target)
&& Options()->oVariables->FP32RoundDivideSqrt);
#endif
if (strcmp(Options()->oVariables->CLStd, "CL2.0") == 0) {
if (isHSAILTarget(Elf()->target)) {
RunPrintfRuntimeBinding = true;
} else if (isCpuTarget(Elf()->target)) {
RunPrintfCpuLowering = true;
}
}
// The first member in the list of libraries is assumed to be
// representative of the target device.
if(!fixUpModule(LLVMBinary(), LibTargetTriple, LibDataLayout,
RunSPIRLoader, DemangleBuiltins,
RunEDGAdapter, SetSPIRCallingConv,
RunX86Adapter, RunPrintfRuntimeBinding, RunPrintfCpuLowering,
RunLowerEnqueueKernel, Options()->oVariables,
BuildLog()))
return 1;
// Before doing anything else, quickly optimize Module
if (Options()->oVariables->OptLevel) {
if (Options()->oVariables->EnableBuildTiming) {
time_prelinkopt = amd::Os::timeNanos();
}
AMDPrelinkOpt(LLVMBinary(), true /*Whole*/,
!Options()->oVariables->OptSimplifyLibCall,
Options()->oVariables->UnsafeMathOpt,
Options()->oVariables->OptUseNative,
LowerToPreciseFunctions);
if (Options()->oVariables->EnableBuildTiming) {
time_prelinkopt = amd::Os::timeNanos() - time_prelinkopt;
}
}
// Now, do linking by extracting from the builtins library only those
// functions that are used in the kernel(s).
if (Options()->oVariables->EnableBuildTiming) {
time_link = amd::Os::timeNanos();
}
std::string ErrorMessage;
// CL pre-link processing
llvm::Module *clp_inputs[2];
clp_inputs[0] = LLVMBinary();
clp_inputs[1] = NULL;
std::string clp_errmsg;
llvm::Module *OnFlyLib = clpVectorExpansion (clp_inputs, clp_errmsg);
if (clp_errmsg.empty() == false) {
delete LLVMBinary();
for (unsigned int i = 0; i < LibMs.size(); ++ i) {
delete LibMs[i];
}
LibMs.clear();
BuildLog() += clp_errmsg;
BuildLog() += "Internal Error: on-fly library generation failed\n";
return 1;
}
unsigned int offset = (unsigned int)LibMs.size();
if (OnFlyLib) {
// OnFlyLib must be the last!
LibMs.push_back(OnFlyLib);
}
// build the reference map
llvm::ReferenceMapBuilder RefMapBuilder(LLVMBinary(), LibMs);
RefMapBuilder.InitReferenceMap();
if (IsGPUTarget && RefMapBuilder.isInExternFuncs("printf")) {
DEBUG(llvm::dbgs() << "Adding printf funs:\n");
// The following functions need forcing as printf-conversion happens
// after this link stage
static const char* forcedRefs[] = {
"___initDumpBuf",
"___dumpBytes_v1b8",
"___dumpBytes_v1b16",
"___dumpBytes_v1b32",
"___dumpBytes_v1b64",
"___dumpBytes_v1b128",
"___dumpBytes_v1b256",
"___dumpBytes_v1b512",
"___dumpBytes_v1b1024",
"___dumpBytes_v1bs",
"___dumpStringID"
};
RefMapBuilder.AddForcedReferences(forcedRefs,
sizeof(forcedRefs)/sizeof(forcedRefs[0]));
}
if (!IsGPUTarget && Options()->oVariables->UseJIT) {
RefMapBuilder.AddForcedReferences(amd::amdRTFuns,
sizeof(amd::amdRTFuns)/sizeof(amd::amdRTFuns[0]));
}
RefMapBuilder.AddReferences();
// inject an llvm function that returns the mask of several compile
// options, which are used by the built-in library
const std::list<std::string>& ExternFuncs
= RefMapBuilder.getExternFunctions();
const std::list<std::string>::const_iterator it
= std::find(ExternFuncs.begin(), ExternFuncs.end(), OptionMaskFName);
if (it != ExternFuncs.end()) {
createOptionMaskFunction(LLVMBinary());
}
createASICIDFunctions(LLVMBinary());
if (!isHSAILTarget(Elf()->target)) {
// Add NoInline attribute to user functions
llvm::StringRef family(aclGetFamily(Elf()->target));
llvm::StringRef chip(aclGetChip(Elf()->target));
// Add NoInline attribute to library functions so that they
// can be considered for not inlining in codegen.
if (IsGPUTarget &&
(Options()->oVariables->OptMem2reg || Options()->oVariables->DebugCall) &&
!Options()->oVariables->clInternalKernel &&
!(family == "NI" || family == "Evergreen" || family == "Sumo" ||
family == "TN")) {
if (Options()->oVariables->AddUserNoInline)
amd::AddNoInlineAttr(LLVMBinary());
if (Options()->oVariables->AddLibNoInline)
for (unsigned int i=0; i < LibMs.size(); i++)
amd::AddNoInlineAttr(LibMs[i]);
}
// Disable outline macro for mem2reg=0 unless -fdebug-call
// is on.
if (!Options()->oVariables->OptMem2reg && !Options()->oVariables->DebugCall) {
Options()->oVariables->UseMacroForCall = false;
}
}
// Link libraries to get every functions that are referenced.
std::string ErrorMsg;
if (resolveLink(LLVMBinary(), LibMs, RefMapBuilder.getModuleRefMaps(),
&ErrorMsg)) {
BuildLog() += ErrorMsg;
BuildLog() += "\nInternal Error: linking libraries failed!\n";
return 1;
}
LibMs.clear();
if (Options()->oVariables->EnableBuildTiming) {
time_link = amd::Os::timeNanos() - time_link;
std::stringstream tmp_ss;
tmp_ss << " LLVM time (link+opt): "
<< (amd::Os::timeNanos() - start_time)/1000ULL
<< " us\n"
<< " prelinkopt: " << time_prelinkopt/1000ULL << " us\n"
<< " link: " << time_link/1000ULL << " us\n"
;
appendLogToCL(CL(), tmp_ss.str());
}
if (!isHSAILTarget(Elf()->target)) {
// Refine NoInline attribute of functions
if (IsGPUTarget && !Options()->oVariables->clInternalKernel) {
amd::RefineNoInlineAttr(LLVMBinary(),
Options()->oVariables->InlineCostThreshold,
Options()->oVariables->InlineSizeThreshold,
Options()->oVariables->InlineKernelSizeThreshold,
Options()->oVariables->AllowMultiLevelCall &&
Options()->oVariables->UseMacroForCall );
}
}
if (Options()->isDumpFlagSet(amd::option::DUMP_BC_LINKED)) {
std::string MyErrorInfo;
std::string fileName = Options()->getDumpFileName("_linked.bc");
llvm::raw_fd_ostream outs(fileName.c_str(), MyErrorInfo, llvm::raw_fd_ostream::F_Binary);
// FIXME: Need to add this to the elf binary!
if (MyErrorInfo.empty())
WriteBitcodeToFile(LLVMBinary(), outs);
else
printf(MyErrorInfo.c_str());
}
// Check if kernels containing local arrays are called by other kernels.
std::string localArrayUsageError;
if (!llvm::AMDCheckLocalArrayUsage(*LLVMBinary(), &localArrayUsageError)) {
BuildLog() += "Error: " + localArrayUsageError + '\n';
return 1;
}
return 0;
}
Ссылка в новой задаче Показать git blame Копировать постоянную ссылку