LLVMExecutionEngineRef
mono_llvm_create_ee (LLVMModuleProviderRef MP, AllocCodeMemoryCb *alloc_cb, FunctionEmittedCb *emitted_cb, ExceptionTableCb *exception_cb)
{
std::string Error;
force_pass_linking ();
LLVMInitializeX86Target ();
LLVMInitializeX86TargetInfo ();
llvm::cl::ParseEnvironmentOptions("mono", "MONO_LLVM", "", false);
mono_mm = new MonoJITMemoryManager ();
mono_mm->alloc_cb = alloc_cb;
#if LLVM_MAJOR_VERSION == 2 && LLVM_MINOR_VERSION < 8
DwarfExceptionHandling = true;
#else
JITExceptionHandling = true;
#endif
// PrettyStackTrace installs signal handlers which trip up libgc
DisablePrettyStackTrace = true;
ExecutionEngine *EE = ExecutionEngine::createJIT (unwrap (MP), &Error, mono_mm, CodeGenOpt::Default);
if (!EE) {
errs () << "Unable to create LLVM ExecutionEngine: " << Error << "\n";
g_assert_not_reached ();
}
EE->InstallExceptionTableRegister (exception_cb);
EE->RegisterJITEventListener (new MonoJITEventListener (emitted_cb));
fpm = new FunctionPassManager (unwrap (MP));
fpm->add(new TargetData(*EE->getTargetData()));
/* Add a default set of passes */
//createStandardFunctionPasses (fpm, 2);
fpm->add(createInstructionCombiningPass());
fpm->add(createReassociatePass());
fpm->add(createGVNPass());
fpm->add(createCFGSimplificationPass());
/* The one used by opt is:
* -simplifycfg -domtree -domfrontier -scalarrepl -instcombine -simplifycfg -basiccg -domtree -domfrontier -scalarrepl -simplify-libcalls -instcombine -simplifycfg -instcombine -simplifycfg -reassociate -domtree -loops -loopsimplify -domfrontier -loopsimplify -lcssa -loop-rotate -licm -lcssa -loop-unswitch -instcombine -scalar-evolution -loopsimplify -lcssa -iv-users -indvars -loop-deletion -loopsimplify -lcssa -loop-unroll -instcombine -memdep -gvn -memdep -memcpyopt -sccp -instcombine -domtree -memdep -dse -adce -gvn -simplifycfg -preverify -domtree -verify
*/
/* Add passes specified by the env variable */
/* Only the passes in force_pass_linking () can be used */
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
fpm->add (P);
}
return wrap(EE);
}
LLVMExecutionEngineRef
mono_llvm_create_ee (LLVMModuleProviderRef MP, AllocCodeMemoryCb *alloc_cb, FunctionEmittedCb *emitted_cb, ExceptionTableCb *exception_cb)
{
std::string Error;
LLVMInitializeX86Target ();
LLVMInitializeX86TargetInfo ();
llvm::cl::ParseEnvironmentOptions("mono", "MONO_LLVM", "", false);
mono_mm = new MonoJITMemoryManager ();
mono_mm->alloc_cb = alloc_cb;
mono_mm->emitted_cb = emitted_cb;
DwarfExceptionHandling = true;
// PrettyStackTrace installs signal handlers which trip up libgc
DisablePrettyStackTrace = true;
ExecutionEngine *EE = ExecutionEngine::createJIT (unwrap (MP), &Error, mono_mm, CodeGenOpt::Default);
if (!EE) {
errs () << "Unable to create LLVM ExecutionEngine: " << Error << "\n";
g_assert_not_reached ();
}
EE->InstallExceptionTableRegister (exception_cb);
EE->RegisterJITEventListener (new MonoJITEventListener ());
fpm = new FunctionPassManager (unwrap (MP));
fpm->add(new TargetData(*EE->getTargetData()));
/* Add a random set of passes */
/* Make this run-time configurable */
fpm->add(createInstructionCombiningPass());
fpm->add(createReassociatePass());
fpm->add(createGVNPass());
fpm->add(createCFGSimplificationPass());
/* Add passes specified by the env variable */
/* FIXME: This can only add passes which are linked in, thus are already used */
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
if (dynamic_cast<MachineFunctionPass*>(P) != 0) {
errs () << PassInf->getPassName () << " is a machine function pass.\n";
} else {
fpm->add (P);
}
}
return wrap(EE);
}
LLVMExecutionEngineRef
mono_llvm_create_ee (LLVMModuleProviderRef MP, AllocCodeMemoryCb *alloc_cb, FunctionEmittedCb *emitted_cb, ExceptionTableCb *exception_cb, DlSymCb *dlsym_cb)
{
std::string Error;
force_pass_linking ();
#ifdef TARGET_ARM
LLVMInitializeARMTarget ();
LLVMInitializeARMTargetInfo ();
LLVMInitializeARMTargetMC ();
#else
LLVMInitializeX86Target ();
LLVMInitializeX86TargetInfo ();
LLVMInitializeX86TargetMC ();
#endif
mono_mm = new MonoJITMemoryManager ();
mono_mm->alloc_cb = alloc_cb;
mono_mm->dlsym_cb = dlsym_cb;
//JITExceptionHandling = true;
// PrettyStackTrace installs signal handlers which trip up libgc
DisablePrettyStackTrace = true;
/*
* The Default code model doesn't seem to work on amd64,
* test_0_fields_with_big_offsets (among others) crashes, because LLVM tries to call
* memset using a normal pcrel code which is in 32bit memory, while memset isn't.
*/
TargetOptions opts;
opts.JITExceptionHandling = 1;
EngineBuilder b (unwrap (MP));
#ifdef TARGET_AMD64
ExecutionEngine *EE = b.setJITMemoryManager (mono_mm).setTargetOptions (opts).setCodeModel (CodeModel::Large).setAllocateGVsWithCode (true).create ();
#else
ExecutionEngine *EE = b.setJITMemoryManager (mono_mm).setTargetOptions (opts).setAllocateGVsWithCode (true).create ();
#endif
g_assert (EE);
#if 0
ExecutionEngine *EE = ExecutionEngine::createJIT (unwrap (MP), &Error, mono_mm, CodeGenOpt::Default, true, Reloc::Default, CodeModel::Large);
if (!EE) {
errs () << "Unable to create LLVM ExecutionEngine: " << Error << "\n";
g_assert_not_reached ();
}
#endif
EE->InstallExceptionTableRegister (exception_cb);
mono_event_listener = new MonoJITEventListener (emitted_cb);
EE->RegisterJITEventListener (mono_event_listener);
fpm = new FunctionPassManager (unwrap (MP));
fpm->add(new DataLayout(*EE->getDataLayout()));
PassRegistry &Registry = *PassRegistry::getPassRegistry();
initializeCore(Registry);
initializeScalarOpts(Registry);
//initializeIPO(Registry);
initializeAnalysis(Registry);
initializeIPA(Registry);
initializeTransformUtils(Registry);
initializeInstCombine(Registry);
//initializeInstrumentation(Registry);
initializeTarget(Registry);
llvm::cl::ParseEnvironmentOptions("mono", "MONO_LLVM", "");
if (PassList.size() > 0) {
/* Use the passes specified by the env variable */
/* Only the passes in force_pass_linking () can be used */
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
fpm->add (P);
}
} else {
/* Use the same passes used by 'opt' by default, without the ipo passes */
const char *opts = "-simplifycfg -domtree -domfrontier -scalarrepl -instcombine -simplifycfg -domtree -domfrontier -scalarrepl -simplify-libcalls -instcombine -simplifycfg -instcombine -simplifycfg -reassociate -domtree -loops -loop-simplify -domfrontier -loop-simplify -lcssa -loop-rotate -licm -lcssa -loop-unswitch -instcombine -scalar-evolution -loop-simplify -lcssa -iv-users -indvars -loop-deletion -loop-simplify -lcssa -loop-unroll -instcombine -memdep -gvn -memdep -memcpyopt -sccp -instcombine -domtree -memdep -dse -adce -gvn -simplifycfg -preverify -domtree -verify";
char **args;
int i;
args = g_strsplit (opts, " ", 1000);
for (i = 0; args [i]; i++)
;
llvm::cl::ParseCommandLineOptions (i, args, "");
g_strfreev (args);
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
g_assert (P->getPassKind () == llvm::PT_Function || P->getPassKind () == llvm::PT_Loop);
fpm->add (P);
}
/*
fpm->add(createInstructionCombiningPass());
fpm->add(createReassociatePass());
fpm->add(createGVNPass());
fpm->add(createCFGSimplificationPass());
*/
}
return wrap(EE);
}
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
* 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<std::string, 16> MAttrs;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
#if HAVE_LLVM >= 0x0400
/* llvm-3.7+ implements sys::getHostCPUFeatures for x86,
* which allows us to enable/disable code generation based
* on the results of cpuid.
*/
llvm::StringMap<bool> features;
llvm::sys::getHostCPUFeatures(features);
for (StringMapIterator<bool> f = features.begin();
f != features.end();
++f) {
MAttrs.push_back(((*f).second ? "+" : "-") + (*f).first().str());
}
#else
/*
* 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
#endif
#if defined(PIPE_ARCH_PPC)
MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec");
#if (HAVE_LLVM >= 0x0304)
#if (HAVE_LLVM < 0x0400)
/*
* Make sure VSX instructions are disabled
* See LLVM bugs:
* https://llvm.org/bugs/show_bug.cgi?id=25503#c7 (fixed in 3.8.1)
* https://llvm.org/bugs/show_bug.cgi?id=26775 (fixed in 3.8.1)
* https://llvm.org/bugs/show_bug.cgi?id=33531 (fixed in 4.0)
* https://llvm.org/bugs/show_bug.cgi?id=34647 (llc performance on certain unusual shader IR; intro'd in 4.0, pending as of 5.0)
*/
if (util_cpu_caps.has_altivec) {
MAttrs.push_back("-vsx");
}
#else
/*
* Bug 25503 is fixed, by the same fix that fixed
* bug 26775, in versions of LLVM later than 3.8 (starting with 3.8.1).
* BZ 33531 actually comprises more than one bug, all of
* which are fixed in LLVM 4.0.
*
* With LLVM 4.0 or higher:
* Make sure VSX instructions are ENABLED, unless
* a) the entire -mattr option is overridden via GALLIVM_MATTRS, or
* b) VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
*/
if (util_cpu_caps.has_altivec) {
char *env_mattrs = getenv("GALLIVM_MATTRS");
if (env_mattrs) {
MAttrs.push_back(env_mattrs);
}
else {
boolean enable_vsx = true;
char *env_vsx = getenv("GALLIVM_VSX");
if (env_vsx && env_vsx[0] == '0') {
enable_vsx = false;
}
if (enable_vsx)
MAttrs.push_back("+vsx");
else
MAttrs.push_back("-vsx");
}
}
#endif
#endif
#endif
builder.setMAttrs(MAttrs);
if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
int n = MAttrs.size();
if (n > 0) {
debug_printf("llc -mattr option(s): ");
for (int i = 0; i < n; i++)
debug_printf("%s%s", MAttrs[i].c_str(), (i < n - 1) ? "," : "");
debug_printf("\n");
}
}
#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.
*/
#if defined(PIPE_ARCH_LITTLE_ENDIAN) && defined(PIPE_ARCH_PPC_64)
/*
* Versions of LLVM prior to 4.0 lacked a table entry for "POWER8NVL",
* resulting in (big-endian) "generic" being returned on
* little-endian Power8NVL systems. The result was that code that
* attempted to load the least significant 32 bits of a 64-bit quantity
* from memory loaded the wrong half. This resulted in failures in some
* Piglit tests, e.g.
* .../arb_gpu_shader_fp64/execution/conversion/frag-conversion-explicit-double-uint
*/
if (MCPU == "generic")
MCPU = "pwr8";
#endif
builder.setMCPU(MCPU);
if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
debug_printf("llc -mcpu option: %s\n", MCPU.str().c_str());
}
#endif
ShaderMemoryManager *MM = NULL;
if (useMCJIT) {
BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
MM = new ShaderMemoryManager(JMM);
*OutCode = MM->getGeneratedCode();
#if HAVE_LLVM >= 0x0306
builder.setMCJITMemoryManager(std::unique_ptr<RTDyldMemoryManager>(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<BaseMemoryManager*>(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;
}
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
* 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<std::string, 16> 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<BaseMemoryManager*>(CMM);
MM = new ShaderMemoryManager(JMM);
*OutCode = MM->getGeneratedCode();
#if HAVE_LLVM >= 0x0306
builder.setMCJITMemoryManager(std::unique_ptr<RTDyldMemoryManager>(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<BaseMemoryManager*>(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;
}
MonoEERef
mono_llvm_create_ee (LLVMModuleProviderRef MP, AllocCodeMemoryCb *alloc_cb, FunctionEmittedCb *emitted_cb, ExceptionTableCb *exception_cb, DlSymCb *dlsym_cb, LLVMExecutionEngineRef *ee)
{
std::string Error;
MonoEE *mono_ee;
init_llvm ();
mono_ee = new MonoEE ();
MonoJITMemoryManager *mono_mm = new MonoJITMemoryManager ();
mono_mm->alloc_cb = alloc_cb;
mono_mm->dlsym_cb = dlsym_cb;
mono_mm->exception_cb = exception_cb;
mono_ee->mm = mono_mm;
/*
* The Default code model doesn't seem to work on amd64,
* test_0_fields_with_big_offsets (among others) crashes, because LLVM tries to call
* memset using a normal pcrel code which is in 32bit memory, while memset isn't.
*/
TargetOptions opts;
opts.JITExceptionHandling = 1;
StringRef cpu_name = sys::getHostCPUName ();
// EngineBuilder no longer has a copy assignment operator (?)
std::unique_ptr<Module> Owner(unwrap(MP));
EngineBuilder b (std::move(Owner));
ExecutionEngine *EE = b.setJITMemoryManager (mono_mm).setTargetOptions (opts).setAllocateGVsWithCode (true).setMCPU (cpu_name).create ();
g_assert (EE);
mono_ee->EE = EE;
MonoJITEventListener *listener = new MonoJITEventListener (emitted_cb);
EE->RegisterJITEventListener (listener);
mono_ee->listener = listener;
FunctionPassManager *fpm = new FunctionPassManager (unwrap (MP));
mono_ee->fpm = fpm;
fpm->add(new DataLayoutPass(*EE->getDataLayout()));
if (PassList.size() > 0) {
/* Use the passes specified by the env variable */
/* Only the passes in force_pass_linking () can be used */
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
fpm->add (P);
}
} else {
/* Use the same passes used by 'opt' by default, without the ipo passes */
const char *opts = "-simplifycfg -domtree -domfrontier -scalarrepl -instcombine -simplifycfg -domtree -domfrontier -scalarrepl -instcombine -simplifycfg -instcombine -simplifycfg -reassociate -domtree -loops -loop-simplify -domfrontier -loop-simplify -lcssa -loop-rotate -licm -lcssa -loop-unswitch -instcombine -scalar-evolution -loop-simplify -lcssa -iv-users -indvars -loop-deletion -loop-simplify -lcssa -loop-unroll -instcombine -memdep -gvn -memdep -memcpyopt -sccp -instcombine -domtree -memdep -dse -adce -gvn -simplifycfg";
char **args;
int i;
args = g_strsplit (opts, " ", 1000);
for (i = 0; args [i]; i++)
;
llvm::cl::ParseCommandLineOptions (i, args, "");
g_strfreev (args);
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
g_assert (P->getPassKind () == llvm::PT_Function || P->getPassKind () == llvm::PT_Loop);
fpm->add (P);
}
/*
fpm->add(createInstructionCombiningPass());
fpm->add(createReassociatePass());
fpm->add(createGVNPass());
fpm->add(createCFGSimplificationPass());
*/
}
*ee = wrap (EE);
return mono_ee;
}
LLVMExecutionEngineRef
mono_llvm_create_ee (LLVMModuleProviderRef MP, AllocCodeMemoryCb *alloc_cb, FunctionEmittedCb *emitted_cb, ExceptionTableCb *exception_cb)
{
std::string Error;
force_pass_linking ();
LLVMInitializeX86Target ();
LLVMInitializeX86TargetInfo ();
mono_mm = new MonoJITMemoryManager ();
mono_mm->alloc_cb = alloc_cb;
#if LLVM_MAJOR_VERSION == 2 && LLVM_MINOR_VERSION < 8
DwarfExceptionHandling = true;
#else
JITExceptionHandling = true;
#endif
// PrettyStackTrace installs signal handlers which trip up libgc
DisablePrettyStackTrace = true;
ExecutionEngine *EE = ExecutionEngine::createJIT (unwrap (MP), &Error, mono_mm, CodeGenOpt::Default);
if (!EE) {
errs () << "Unable to create LLVM ExecutionEngine: " << Error << "\n";
g_assert_not_reached ();
}
EE->InstallExceptionTableRegister (exception_cb);
mono_event_listener = new MonoJITEventListener (emitted_cb);
EE->RegisterJITEventListener (mono_event_listener);
fpm = new FunctionPassManager (unwrap (MP));
fpm->add(new TargetData(*EE->getTargetData()));
#if LLVM_CHECK_VERSION(2, 9)
PassRegistry &Registry = *PassRegistry::getPassRegistry();
initializeCore(Registry);
initializeScalarOpts(Registry);
//initializeIPO(Registry);
initializeAnalysis(Registry);
initializeIPA(Registry);
initializeTransformUtils(Registry);
initializeInstCombine(Registry);
//initializeInstrumentation(Registry);
initializeTarget(Registry);
#endif
llvm::cl::ParseEnvironmentOptions("mono", "MONO_LLVM", "", false);
if (PassList.size() > 0) {
/* Use the passes specified by the env variable */
/* Only the passes in force_pass_linking () can be used */
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
fpm->add (P);
}
} else {
/* Use the same passes used by 'opt' by default, without the ipo passes */
const char *opts = "-simplifycfg -domtree -domfrontier -scalarrepl -instcombine -simplifycfg -basiccg -domtree -domfrontier -scalarrepl -simplify-libcalls -instcombine -simplifycfg -instcombine -simplifycfg -reassociate -domtree -loops -loopsimplify -domfrontier -loopsimplify -lcssa -loop-rotate -licm -lcssa -loop-unswitch -instcombine -scalar-evolution -loopsimplify -lcssa -iv-users -indvars -loop-deletion -loopsimplify -lcssa -loop-unroll -instcombine -memdep -gvn -memdep -memcpyopt -sccp -instcombine -domtree -memdep -dse -adce -gvn -simplifycfg -preverify -domtree -verify";
char **args;
int i;
args = g_strsplit (opts, " ", 1000);
for (i = 0; args [i]; i++)
;
llvm::cl::ParseCommandLineOptions (i, args, "", false);
g_strfreev (args);
for (unsigned i = 0; i < PassList.size(); ++i) {
const PassInfo *PassInf = PassList[i];
Pass *P = 0;
if (PassInf->getNormalCtor())
P = PassInf->getNormalCtor()();
fpm->add (P);
}
/*
fpm->add(createInstructionCombiningPass());
fpm->add(createReassociatePass());
fpm->add(createGVNPass());
fpm->add(createCFGSimplificationPass());
*/
}
return wrap(EE);
}
int main() {
ygt_print("Hello!\n");
InitializeNativeTarget();
InitializeNativeTargetAsmPrinter();
InitializeNativeTargetAsmParser();
InitializeNativeTargetDisassembler();
ygt_print("Creating LLVM context and module...\n");
LLVMContext ctx;
auto m = make_unique<Module>("the_module", ctx);
ygt_print("Creating StackMapHelper...\n");
StackMapHelper smhelper;
ygt_print("Declaring yg_stack_swap as an LLVM-level function...\n");
declare_ss(ctx, *m);
ygt_print("Constructing bar, the ss leaf...\n");
Function *bar;
smid_t bar_smid;
tie(bar, bar_smid) = make_ss_leaf("barss", &main_stack, &coro_stack, smhelper, ctx, *m);
verifyFunction(*bar);
ygt_print("Constructing foo, the caller...\n");
Function *foo;
smid_t foo_smid;
tie(foo, foo_smid) = make_caller("foo", bar, smhelper, ctx, *m);
verifyFunction(*foo);
m->dump();
Function *ss_func = m->getFunction("yg_stack_swap");
ygt_print("Creating execution engine...\n");
ExecutionEngine *ee = EngineBuilder(move(m)).setEngineKind(EngineKind::JIT).create();
ygt_print("Execution engine created.\n");
StackMapSectionRecorder smsr;
ygt_print("Registering JIT event listener...\n");
ee->RegisterJITEventListener(&smsr);
EHFrameSectionRegisterer reg;
ygt_print("Registering EHFrame registerer...\n");
ee->RegisterJITEventListener(®);
ygt_print("Adding global mapping yg_stack_swap...\n");
ee->addGlobalMapping(ss_func, reinterpret_cast<void*>(yg_stack_swap));
ygt_print("JIT compiling...\n");
ee->finalizeObject();
ygt_print("Adding stack map sections...\n");
smhelper.add_stackmap_sections(smsr, *ee);
ygt_print("Registering EH frames...\n");
reg.registerEHFrameSections();
uintptr_t yg_stack_swap_addr = ee->getFunctionAddress("yg_stack_swap");
ygt_print("yg_stack_swap_addr=%" PRIxPTR ", yg_stack_swap=%p\n", yg_stack_swap_addr, yg_stack_swap);
assert(yg_stack_swap_addr == reinterpret_cast<uintptr_t>(yg_stack_swap));
ygt_print("Getting foo...\n");
void (*the_real_foo)(int32_t, int64_t, float, double) = (void(*)(int32_t, int64_t, float, double))
ee->getFunctionAddress("foo");
ygt_print("the_real_foo == %p\n", the_real_foo);
ygt_print("Prepare corotine for introspection...\n");
coro_stack = YGStack::alloc(8*1024*1024);
coro_stack.init(reinterpret_cast<uintptr_t>(coro));
MyCtx my_ctx = { foo_smid, bar_smid, &smhelper };
ygt_print("Swap-stack to prepare the coroutine...\n");
yg_stack_swap(&main_stack, &coro_stack, reinterpret_cast<uintptr_t>(&my_ctx));
ygt_print("Back from coro.\n");
ygt_print("Calling the_real_foo...\n");
the_real_foo(100, 200, 300.0f, 400.0);
ygt_print("Returned from foo.\n");
return 0;
}
