#include <llvm-c/ExecutionEngine.h>
#include <llvm/Target/TargetOptions.h>
#include <llvm/ExecutionEngine/ExecutionEngine.h>
-#if HAVE_LLVM >= 0x0301
#include <llvm/ADT/Triple.h>
+#if HAVE_LLVM < 0x0306
#include <llvm/ExecutionEngine/JITMemoryManager.h>
+#else
+#include <llvm/ExecutionEngine/SectionMemoryManager.h>
#endif
#include <llvm/Support/CommandLine.h>
+#include <llvm/Support/Host.h>
#include <llvm/Support/PrettyStackTrace.h>
-#if HAVE_LLVM >= 0x0300
#include <llvm/Support/TargetSelect.h>
-#else /* HAVE_LLVM < 0x0300 */
-#include <llvm/Target/TargetSelect.h>
-#endif /* HAVE_LLVM < 0x0300 */
+
+#include <llvm/IR/IRBuilder.h>
+#include <llvm/IR/Module.h>
+#include <llvm/Support/CBindingWrapping.h>
#include "pipe/p_config.h"
#include "util/u_debug.h"
#include "lp_bld_misc.h"
+namespace {
-extern "C" void
-lp_set_target_options(void)
-{
-#if HAVE_LLVM <= 0x0300
-#if defined(DEBUG)
-#if HAVE_LLVM >= 0x0207
- llvm::JITEmitDebugInfo = true;
-#endif
-#endif
-
- /*
- * LLVM revision 123367 switched the default stack alignment to 16 bytes on
- * Linux (and several other Unices in later revisions), to match recent gcc
- * versions.
- *
- * However our drivers can be loaded by old binary applications, still
- * maintaining a 4 bytes stack alignment. Therefore we must tell LLVM here
- * to only assume a 4 bytes alignment for backwards compatibility.
- */
-#if defined(PIPE_ARCH_X86)
-#if HAVE_LLVM == 0x0300
- llvm::StackAlignmentOverride = 4;
-#else
- llvm::StackAlignment = 4;
-#endif
-#endif
-
-#if defined(DEBUG) || defined(PROFILE)
- llvm::NoFramePointerElim = true;
-#if HAVE_LLVM >= 0x0208
- llvm::NoFramePointerElimNonLeaf = true;
-#endif
-#endif
-
- llvm::NoExcessFPPrecision = false;
+class LLVMEnsureMultithreaded {
+public:
+ LLVMEnsureMultithreaded()
+ {
+ if (!LLVMIsMultithreaded()) {
+ LLVMStartMultithreaded();
+ }
+ }
+};
- /* XXX: Investigate this */
-#if 0
- llvm::UnsafeFPMath = true;
-#endif
-#endif /* HAVE_LLVM <= 0x0300 */
+static LLVMEnsureMultithreaded lLVMEnsureMultithreaded;
-#if HAVE_LLVM < 0x0209
- /*
- * LLVM will generate MMX instructions for vectors <= 64 bits, leading to
- * innefficient code, and in 32bit systems, to the corruption of the FPU
- * stack given that it expects the user to generate the EMMS instructions.
- *
- * See also:
- * - http://llvm.org/bugs/show_bug.cgi?id=3287
- * - http://l4.me.uk/post/2009/06/07/llvm-wrinkle-3-configuration-what-configuration/
- *
- * The -disable-mmx global option can be specified only once since we
- * dynamically link against LLVM it will reside in a separate shared object,
- * which may or not be delete when this shared object is, so we use the
- * llvm::DisablePrettyStackTrace variable (which we set below and should
- * reside in the same shared library) to determine whether the -disable-mmx
- * option has been set or not.
- *
- * Thankfully this ugly hack is not necessary on LLVM 2.9 onwards.
- */
- if (!llvm::DisablePrettyStackTrace) {
- static boolean first = TRUE;
- static const char* options[] = {
- "prog",
- "-disable-mmx"
- };
- assert(first);
- llvm::cl::ParseCommandLineOptions(2, const_cast<char**>(options));
- first = FALSE;
- }
-#endif
+}
+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
// If we have a native target, initialize it to ensure it is linked in and
// usable by the JIT.
llvm::InitializeNativeTarget();
-#if HAVE_LLVM >= 0x0208
llvm::InitializeNativeTargetAsmPrinter();
-#elif defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
- LLVMInitializeX86AsmPrinter();
-#elif defined(PIPE_ARCH_ARM)
- LLVMInitializeARMAsmPrinter();
-#elif defined(PIPE_ARCH_PPC)
- LLVMInitializePowerPCAsmPrinter();
-#endif
-#if HAVE_LLVM >= 0x0207
-# if HAVE_LLVM >= 0x0301
llvm::InitializeNativeTargetDisassembler();
-# elif defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
- LLVMInitializeX86Disassembler();
-# elif defined(PIPE_ARCH_ARM)
- LLVMInitializeARMDisassembler();
-# endif
-#endif
-}
-
-
-extern "C" void
-lp_func_delete_body(LLVMValueRef FF)
-{
- llvm::Function *func = llvm::unwrap<llvm::Function>(FF);
- func->deleteBody();
}
extern "C"
void
lp_set_store_alignment(LLVMValueRef Inst,
- unsigned Align)
+ unsigned Align)
{
llvm::unwrap<llvm::StoreInst>(Inst)->setAlignment(Align);
}
-#if HAVE_LLVM >= 0x301
+#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<void *> 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:
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<Module>(unwrap(M)));
+#else
EngineBuilder builder(unwrap(M));
+#endif
/**
* LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and
TargetOptions options;
#if defined(PIPE_ARCH_X86)
options.StackAlignmentOverride = 4;
+#if HAVE_LLVM < 0x0304
options.RealignStack = true;
#endif
+#endif
#if defined(DEBUG)
options.JITEmitDebugInfo = true;
#endif
-#if defined(DEBUG) || defined(PROFILE)
+ /* 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
options.NoFramePointerElim = true;
#endif
.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.
+ */
+ std::string targetTriple = llvm::sys::getProcessTriple();
+ targetTriple.append("-elf");
+ unwrap(M)->setTargetTriple(targetTriple);
+#endif
}
llvm::SmallVector<std::string, 1> MAttrs;
/*
* 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. But as we're using MCJIT here, it is safe to
- * add set this attribute.
+ * 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("+avx");
if (util_cpu_caps.has_f16c) {
}
builder.setMAttrs(MAttrs);
}
- builder.setJITMemoryManager(JITMemoryManager::CreateDefaultMemManager());
- ExecutionEngine *JIT;
-#if 0
- JIT = builder.create();
-#else
+#if HAVE_LLVM >= 0x0305
+ StringRef MCPU = llvm::sys::getHostCPUName();
/*
- * Workaround http://llvm.org/bugs/show_bug.cgi?id=12833
+ * 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.
*/
- StringRef MArch = "";
- StringRef MCPU = "";
- Triple TT(unwrap(M)->getTargetTriple());
- JIT = builder.create(builder.selectTarget(TT, MArch, MCPU, MAttrs));
+ builder.setMCPU(MCPU);
#endif
+
+ ShaderMemoryManager *MM = NULL;
+ if (useMCJIT) {
+#if HAVE_LLVM > 0x0303
+ BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
+ MM = new ShaderMemoryManager(JMM);
+ *OutCode = MM->getGeneratedCode();
+
+#if HAVE_LLVM >= 0x0306
+ builder.setMCJITMemoryManager(std::unique_ptr<RTDyldMemoryManager>(MM));
+#else
+ builder.setMCJITMemoryManager(MM);
+#endif
+#endif
+ } else {
+#if HAVE_LLVM < 0x0306
+ BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
+ MM = new ShaderMemoryManager(JMM);
+ *OutCode = MM->getGeneratedCode();
+
+ builder.setJITMemoryManager(MM);
+#else
+ assert(0);
+#endif
+ }
+
+ ExecutionEngine *JIT;
+
+ JIT = builder.create();
if (JIT) {
*OutJIT = wrap(JIT);
return 0;
}
+ lp_free_generated_code(*OutCode);
+ *OutCode = 0;
+ delete MM;
*OutError = strdup(Error.c_str());
return 1;
}
-#endif /* HAVE_LLVM >= 0x301 */
+
+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<LLVMMCJITMemoryManagerRef>(mm);
+}
+
+extern "C"
+void
+lp_free_memory_manager(LLVMMCJITMemoryManagerRef memorymgr)
+{
+ delete reinterpret_cast<BaseMemoryManager*>(memorymgr);
+}
projects / mesa.git / blobdiff