/************************************************************************** * * Copyright 2010 VMware, Inc. * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the * "Software"), to deal in the Software without restriction, including * without limitation the rights to use, copy, modify, merge, publish, * distribute, sub license, and/or sell copies of the Software, and to * permit persons to whom the Software is furnished to do so, subject to * the following conditions: * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE * USE OR OTHER DEALINGS IN THE SOFTWARE. * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. * **************************************************************************/ /** * The purpose of this module is to expose LLVM functionality not available * through the C++ bindings. */ #ifndef __STDC_LIMIT_MACROS #define __STDC_LIMIT_MACROS #endif #ifndef __STDC_CONSTANT_MACROS #define __STDC_CONSTANT_MACROS #endif // Undef these vars just to silence warnings #undef PACKAGE_BUGREPORT #undef PACKAGE_NAME #undef PACKAGE_STRING #undef PACKAGE_TARNAME #undef PACKAGE_VERSION #include // Workaround http://llvm.org/PR23628 #if HAVE_LLVM >= 0x0307 # pragma push_macro("DEBUG") # undef DEBUG #endif #include #include #include #include #include #if HAVE_LLVM >= 0x0307 #include #else #include #endif #if HAVE_LLVM < 0x0306 #include #else #include #endif #include #include #include #include #include #include #include #include #if LLVM_USE_INTEL_JITEVENTS #include #endif // Workaround http://llvm.org/PR23628 #if HAVE_LLVM >= 0x0307 # pragma pop_macro("DEBUG") #endif #include "c11/threads.h" #include "os/os_thread.h" #include "pipe/p_config.h" #include "util/u_debug.h" #include "util/u_cpu_detect.h" #include "lp_bld_misc.h" namespace { class LLVMEnsureMultithreaded { public: LLVMEnsureMultithreaded() { if (!LLVMIsMultithreaded()) { LLVMStartMultithreaded(); } } }; static LLVMEnsureMultithreaded lLVMEnsureMultithreaded; } static once_flag init_native_targets_once_flag = ONCE_FLAG_INIT; static void init_native_targets() { // If we have a native target, initialize it to ensure it is linked in and // usable by the JIT. llvm::InitializeNativeTarget(); llvm::InitializeNativeTargetAsmPrinter(); llvm::InitializeNativeTargetDisassembler(); } /** * The llvm target registry is not thread-safe, so drivers and state-trackers * that want to initialize targets should use the gallivm_init_llvm_targets() * function to safely initialize targets. * * LLVM targets should be initialized before the driver or state-tracker tries * to access the registry. */ extern "C" void gallivm_init_llvm_targets(void) { call_once(&init_native_targets_once_flag, init_native_targets); } extern "C" void lp_set_target_options(void) { #if HAVE_LLVM < 0x0304 /* * By default LLVM adds a signal handler to output a pretty stack trace. * This signal handler is never removed, causing problems when unloading the * shared object where the gallium driver resides. */ llvm::DisablePrettyStackTrace = true; #endif gallivm_init_llvm_targets(); } extern "C" LLVMTargetLibraryInfoRef gallivm_create_target_library_info(const char *triple) { return reinterpret_cast( #if HAVE_LLVM < 0x0307 new llvm::TargetLibraryInfo( #else new llvm::TargetLibraryInfoImpl( #endif llvm::Triple(triple))); } extern "C" void gallivm_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info) { delete reinterpret_cast< #if HAVE_LLVM < 0x0307 llvm::TargetLibraryInfo #else llvm::TargetLibraryInfoImpl #endif *>(library_info); } #if HAVE_LLVM < 0x0304 extern "C" void LLVMSetAlignmentBackport(LLVMValueRef V, unsigned Bytes) { switch (LLVMGetInstructionOpcode(V)) { case LLVMLoad: llvm::unwrap(V)->setAlignment(Bytes); break; case LLVMStore: llvm::unwrap(V)->setAlignment(Bytes); break; default: assert(0); break; } } #endif #if HAVE_LLVM < 0x0306 typedef llvm::JITMemoryManager BaseMemoryManager; #else typedef llvm::RTDyldMemoryManager BaseMemoryManager; #endif /* * Delegating is tedious but the default manager class is hidden in an * anonymous namespace in LLVM, so we cannot just derive from it to change * its behavior. */ class DelegatingJITMemoryManager : public BaseMemoryManager { protected: virtual BaseMemoryManager *mgr() const = 0; public: #if HAVE_LLVM < 0x0306 /* * From JITMemoryManager */ virtual void setMemoryWritable() { mgr()->setMemoryWritable(); } virtual void setMemoryExecutable() { mgr()->setMemoryExecutable(); } virtual void setPoisonMemory(bool poison) { mgr()->setPoisonMemory(poison); } virtual void AllocateGOT() { mgr()->AllocateGOT(); /* * isManagingGOT() is not virtual in base class so we can't delegate. * Instead we mirror the value of HasGOT in our instance. */ HasGOT = mgr()->isManagingGOT(); } virtual uint8_t *getGOTBase() const { return mgr()->getGOTBase(); } virtual uint8_t *startFunctionBody(const llvm::Function *F, uintptr_t &ActualSize) { return mgr()->startFunctionBody(F, ActualSize); } virtual uint8_t *allocateStub(const llvm::GlobalValue *F, unsigned StubSize, unsigned Alignment) { return mgr()->allocateStub(F, StubSize, Alignment); } virtual void endFunctionBody(const llvm::Function *F, uint8_t *FunctionStart, uint8_t *FunctionEnd) { mgr()->endFunctionBody(F, FunctionStart, FunctionEnd); } virtual uint8_t *allocateSpace(intptr_t Size, unsigned Alignment) { return mgr()->allocateSpace(Size, Alignment); } virtual uint8_t *allocateGlobal(uintptr_t Size, unsigned Alignment) { return mgr()->allocateGlobal(Size, Alignment); } virtual void deallocateFunctionBody(void *Body) { mgr()->deallocateFunctionBody(Body); } #if HAVE_LLVM < 0x0304 virtual uint8_t *startExceptionTable(const llvm::Function *F, uintptr_t &ActualSize) { return mgr()->startExceptionTable(F, ActualSize); } virtual void endExceptionTable(const llvm::Function *F, uint8_t *TableStart, uint8_t *TableEnd, uint8_t *FrameRegister) { mgr()->endExceptionTable(F, TableStart, TableEnd, FrameRegister); } virtual void deallocateExceptionTable(void *ET) { mgr()->deallocateExceptionTable(ET); } #endif virtual bool CheckInvariants(std::string &s) { return mgr()->CheckInvariants(s); } virtual size_t GetDefaultCodeSlabSize() { return mgr()->GetDefaultCodeSlabSize(); } virtual size_t GetDefaultDataSlabSize() { return mgr()->GetDefaultDataSlabSize(); } virtual size_t GetDefaultStubSlabSize() { return mgr()->GetDefaultStubSlabSize(); } virtual unsigned GetNumCodeSlabs() { return mgr()->GetNumCodeSlabs(); } virtual unsigned GetNumDataSlabs() { return mgr()->GetNumDataSlabs(); } virtual unsigned GetNumStubSlabs() { return mgr()->GetNumStubSlabs(); } #endif /* * From RTDyldMemoryManager */ #if HAVE_LLVM >= 0x0304 virtual uint8_t *allocateCodeSection(uintptr_t Size, unsigned Alignment, unsigned SectionID, llvm::StringRef SectionName) { return mgr()->allocateCodeSection(Size, Alignment, SectionID, SectionName); } #else virtual uint8_t *allocateCodeSection(uintptr_t Size, unsigned Alignment, unsigned SectionID) { return mgr()->allocateCodeSection(Size, Alignment, SectionID); } #endif virtual uint8_t *allocateDataSection(uintptr_t Size, unsigned Alignment, unsigned SectionID, #if HAVE_LLVM >= 0x0304 llvm::StringRef SectionName, #endif bool IsReadOnly) { return mgr()->allocateDataSection(Size, Alignment, SectionID, #if HAVE_LLVM >= 0x0304 SectionName, #endif IsReadOnly); } #if HAVE_LLVM >= 0x0304 virtual void registerEHFrames(uint8_t *Addr, uint64_t LoadAddr, size_t Size) { mgr()->registerEHFrames(Addr, LoadAddr, Size); } virtual void deregisterEHFrames(uint8_t *Addr, uint64_t LoadAddr, size_t Size) { mgr()->deregisterEHFrames(Addr, LoadAddr, Size); } #else virtual void registerEHFrames(llvm::StringRef SectionData) { mgr()->registerEHFrames(SectionData); } #endif virtual void *getPointerToNamedFunction(const std::string &Name, bool AbortOnFailure=true) { return mgr()->getPointerToNamedFunction(Name, AbortOnFailure); } #if HAVE_LLVM <= 0x0303 virtual bool applyPermissions(std::string *ErrMsg = 0) { return mgr()->applyPermissions(ErrMsg); } #else virtual bool finalizeMemory(std::string *ErrMsg = 0) { return mgr()->finalizeMemory(ErrMsg); } #endif }; /* * Delegate memory management to one shared manager for more efficient use * of memory than creating a separate pool for each LLVM engine. * Keep generated code until freeGeneratedCode() is called, instead of when * memory manager is destroyed, which happens during engine destruction. * This allows additional memory savings as we don't have to keep the engine * around in order to use the code. * All methods are delegated to the shared manager except destruction and * deallocating code. For the latter we just remember what needs to be * deallocated later. The shared manager is deleted once it is empty. */ class ShaderMemoryManager : public DelegatingJITMemoryManager { BaseMemoryManager *TheMM; struct GeneratedCode { typedef std::vector Vec; Vec FunctionBody, ExceptionTable; BaseMemoryManager *TheMM; GeneratedCode(BaseMemoryManager *MM) { TheMM = MM; } ~GeneratedCode() { /* * Deallocate things as previously requested and * free shared manager when no longer used. */ #if HAVE_LLVM < 0x0306 Vec::iterator i; assert(TheMM); for ( i = FunctionBody.begin(); i != FunctionBody.end(); ++i ) TheMM->deallocateFunctionBody(*i); #if HAVE_LLVM < 0x0304 for ( i = ExceptionTable.begin(); i != ExceptionTable.end(); ++i ) TheMM->deallocateExceptionTable(*i); #endif /* HAVE_LLVM < 0x0304 */ #endif /* HAVE_LLVM < 0x0306 */ } }; GeneratedCode *code; BaseMemoryManager *mgr() const { return TheMM; } public: ShaderMemoryManager(BaseMemoryManager* MM) { TheMM = MM; code = new GeneratedCode(MM); } virtual ~ShaderMemoryManager() { /* * 'code' is purposely not deleted. It is the user's responsibility * to call getGeneratedCode() and freeGeneratedCode(). */ } struct lp_generated_code *getGeneratedCode() { return (struct lp_generated_code *) code; } static void freeGeneratedCode(struct lp_generated_code *code) { delete (GeneratedCode *) code; } #if HAVE_LLVM < 0x0304 virtual void deallocateExceptionTable(void *ET) { // remember for later deallocation code->ExceptionTable.push_back(ET); } #endif virtual void deallocateFunctionBody(void *Body) { // remember for later deallocation code->FunctionBody.push_back(Body); } }; /** * Same as LLVMCreateJITCompilerForModule, but: * - allows using MCJIT and enabling AVX feature where available. * - set target options * * See also: * - llvm/lib/ExecutionEngine/ExecutionEngineBindings.cpp * - llvm/tools/lli/lli.cpp * - http://markmail.org/message/ttkuhvgj4cxxy2on#query:+page:1+mid:aju2dggerju3ivd3+state:results */ extern "C" LLVMBool lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT, lp_generated_code **OutCode, LLVMModuleRef M, LLVMMCJITMemoryManagerRef CMM, unsigned OptLevel, int useMCJIT, char **OutError) { using namespace llvm; std::string Error; #if HAVE_LLVM >= 0x0306 EngineBuilder builder(std::unique_ptr(unwrap(M))); #else EngineBuilder builder(unwrap(M)); #endif /** * LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and * friends for configuring code generation options, like stack alignment. */ TargetOptions options; #if defined(PIPE_ARCH_X86) options.StackAlignmentOverride = 4; #if HAVE_LLVM < 0x0304 options.RealignStack = true; #endif #endif #if defined(DEBUG) && HAVE_LLVM < 0x0307 options.JITEmitDebugInfo = true; #endif /* XXX: Workaround http://llvm.org/PR21435 */ #if defined(DEBUG) || defined(PROFILE) || \ (HAVE_LLVM >= 0x0303 && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))) #if HAVE_LLVM < 0x0304 options.NoFramePointerElimNonLeaf = true; #endif #if HAVE_LLVM < 0x0307 options.NoFramePointerElim = true; #endif #endif builder.setEngineKind(EngineKind::JIT) .setErrorStr(&Error) .setTargetOptions(options) .setOptLevel((CodeGenOpt::Level)OptLevel); if (useMCJIT) { #if HAVE_LLVM < 0x0306 builder.setUseMCJIT(true); #endif #ifdef _WIN32 /* * MCJIT works on Windows, but currently only through ELF object format. * * XXX: We could use `LLVM_HOST_TRIPLE "-elf"` but LLVM_HOST_TRIPLE has * different strings for MinGW/MSVC, so better play it safe and be * explicit. */ # ifdef _WIN64 LLVMSetTarget(M, "x86_64-pc-win32-elf"); # else LLVMSetTarget(M, "i686-pc-win32-elf"); # endif #endif } llvm::SmallVector MAttrs; #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) /* * We need to unset attributes because sometimes LLVM mistakenly assumes * certain features are present given the processor name. * * https://bugs.freedesktop.org/show_bug.cgi?id=92214 * http://llvm.org/PR25021 * http://llvm.org/PR19429 * http://llvm.org/PR16721 */ MAttrs.push_back(util_cpu_caps.has_sse ? "+sse" : "-sse" ); MAttrs.push_back(util_cpu_caps.has_sse2 ? "+sse2" : "-sse2" ); MAttrs.push_back(util_cpu_caps.has_sse3 ? "+sse3" : "-sse3" ); MAttrs.push_back(util_cpu_caps.has_ssse3 ? "+ssse3" : "-ssse3" ); #if HAVE_LLVM >= 0x0304 MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse4.1" : "-sse4.1"); #else MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse41" : "-sse41" ); #endif #if HAVE_LLVM >= 0x0304 MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse4.2" : "-sse4.2"); #else MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse42" : "-sse42" ); #endif /* * AVX feature is not automatically detected from CPUID by the X86 target * yet, because the old (yet default) JIT engine is not capable of * emitting the opcodes. On newer llvm versions it is and at least some * versions (tested with 3.3) will emit avx opcodes without this anyway. */ MAttrs.push_back(util_cpu_caps.has_avx ? "+avx" : "-avx"); MAttrs.push_back(util_cpu_caps.has_f16c ? "+f16c" : "-f16c"); if (HAVE_LLVM >= 0x0304) { MAttrs.push_back(util_cpu_caps.has_fma ? "+fma" : "-fma"); } else { /* * The old JIT in LLVM 3.3 has a bug encoding llvm.fmuladd.f32 and * llvm.fmuladd.v2f32 intrinsics when FMA is available. */ MAttrs.push_back("-fma"); } MAttrs.push_back(util_cpu_caps.has_avx2 ? "+avx2" : "-avx2"); /* disable avx512 and all subvariants */ #if HAVE_LLVM >= 0x0304 MAttrs.push_back("-avx512cd"); MAttrs.push_back("-avx512er"); MAttrs.push_back("-avx512f"); MAttrs.push_back("-avx512pf"); #endif #if HAVE_LLVM >= 0x0305 MAttrs.push_back("-avx512bw"); MAttrs.push_back("-avx512dq"); MAttrs.push_back("-avx512vl"); #endif #endif #if defined(PIPE_ARCH_PPC) MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec"); #if HAVE_LLVM >= 0x0304 /* * Make sure VSX instructions are disabled * See LLVM bug https://llvm.org/bugs/show_bug.cgi?id=25503#c7 */ if (util_cpu_caps.has_altivec) { MAttrs.push_back("-vsx"); } #endif #endif builder.setMAttrs(MAttrs); #if HAVE_LLVM >= 0x0305 StringRef MCPU = llvm::sys::getHostCPUName(); /* * The cpu bits are no longer set automatically, so need to set mcpu manually. * Note that the MAttrs set above will be sort of ignored (since we should * not set any which would not be set by specifying the cpu anyway). * It ought to be safe though since getHostCPUName() should include bits * not only from the cpu but environment as well (for instance if it's safe * to use avx instructions which need OS support). According to * http://llvm.org/bugs/show_bug.cgi?id=19429 however if I understand this * right it may be necessary to specify older cpu (or disable mattrs) though * when not using MCJIT so no instructions are generated which the old JIT * can't handle. Not entirely sure if we really need to do anything yet. */ builder.setMCPU(MCPU); #endif ShaderMemoryManager *MM = NULL; if (useMCJIT) { BaseMemoryManager* JMM = reinterpret_cast(CMM); MM = new ShaderMemoryManager(JMM); *OutCode = MM->getGeneratedCode(); #if HAVE_LLVM >= 0x0306 builder.setMCJITMemoryManager(std::unique_ptr(MM)); MM = NULL; // ownership taken by std::unique_ptr #elif HAVE_LLVM > 0x0303 builder.setMCJITMemoryManager(MM); #else builder.setJITMemoryManager(MM); #endif } else { #if HAVE_LLVM < 0x0306 BaseMemoryManager* JMM = reinterpret_cast(CMM); MM = new ShaderMemoryManager(JMM); *OutCode = MM->getGeneratedCode(); builder.setJITMemoryManager(MM); #else assert(0); #endif } ExecutionEngine *JIT; JIT = builder.create(); #if LLVM_USE_INTEL_JITEVENTS JITEventListener *JEL = JITEventListener::createIntelJITEventListener(); JIT->RegisterJITEventListener(JEL); #endif if (JIT) { *OutJIT = wrap(JIT); return 0; } lp_free_generated_code(*OutCode); *OutCode = 0; delete MM; *OutError = strdup(Error.c_str()); return 1; } extern "C" void lp_free_generated_code(struct lp_generated_code *code) { ShaderMemoryManager::freeGeneratedCode(code); } extern "C" LLVMMCJITMemoryManagerRef lp_get_default_memory_manager() { BaseMemoryManager *mm; #if HAVE_LLVM < 0x0306 mm = llvm::JITMemoryManager::CreateDefaultMemManager(); #else mm = new llvm::SectionMemoryManager(); #endif return reinterpret_cast(mm); } extern "C" void lp_free_memory_manager(LLVMMCJITMemoryManagerRef memorymgr) { delete reinterpret_cast(memorymgr); } extern "C" void lp_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes) { #if HAVE_LLVM >= 0x0306 llvm::Argument *A = llvm::unwrap(val); llvm::AttrBuilder B; B.addDereferenceableAttr(bytes); A->addAttr(llvm::AttributeSet::get(A->getContext(), A->getArgNo() + 1, B)); #endif }