LLVMCompiledProgram LLVM_compile(ASTNode *node)
{
char *error = NULL; // Used to retrieve messages from functions
LLVMLinkInJIT();
LLVMInitializeNativeTarget();
LLVMModuleRef mod = LLVMModuleCreateWithName("calc_module");
LLVMTypeRef program_args[] = { };
LLVMValueRef program = LLVMAddFunction(mod, "program", LLVMFunctionType(LLVMInt32Type(), program_args, 0, 0));
LLVMSetFunctionCallConv(program, LLVMCCallConv);
LLVMBuilderRef builder = LLVMCreateBuilder();
LLVMBasicBlockRef entry = LLVMAppendBasicBlock(program, "entry");
LLVMPositionBuilderAtEnd(builder, entry);
LLVMValueRef res = LLVM_visit(node, builder);
LLVMBuildRet(builder, res);
LLVMVerifyModule(mod, LLVMAbortProcessAction, &error);
LLVMDisposeMessage(error); // Handler == LLVMAbortProcessAction -> No need to check errors
LLVMDisposeBuilder(builder);
return (LLVMCompiledProgram) { .module = mod, .function = program };
}
int main(int argc, char const *argv[]) {
LLVMModuleRef mod = LLVMModuleCreateWithName("sum");
LLVMTypeRef param_types[] = { LLVMInt32Type(), LLVMInt32Type() };
LLVMTypeRef ret_type = LLVMFunctionType(LLVMInt32Type(), /* ret type */
param_types, /* arg types */
2, /* arg count */
0 /* is variadic */);
LLVMValueRef sum = LLVMAddFunction(mod, "sum", ret_type);
LLVMBasicBlockRef entry = LLVMAppendBasicBlock(sum, "entry");
LLVMBuilderRef builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, entry);
LLVMValueRef tmp = LLVMBuildAdd(builder,
LLVMGetParam(sum, 0),
LLVMGetParam(sum, 1), "tmp");
LLVMBuildRet(builder, tmp);
char *error = NULL;
LLVMVerifyModule(mod, LLVMAbortProcessAction, &error);
LLVMDisposeMessage(error);
LLVMExecutionEngineRef engine;
error = NULL;
LLVMLinkInJIT();
LLVMInitializeNativeTarget();
if (LLVMCreateExecutionEngineForModule(&engine, mod, &error) != 0) {
fprintf(stderr, "failed to create execution engine\n");
abort();
}
if (error) {
fprintf(stderr, "error: %s\n", error);
LLVMDisposeMessage(error);
exit(EXIT_FAILURE);
}
if (argc < 3) {
fprintf(stderr, "usage: %s x y\n", argv[0]);
exit(EXIT_FAILURE);
}
long long x = strtoll(argv[1], NULL, 10);
long long y = strtoll(argv[2], NULL, 10);
LLVMGenericValueRef args[] = {
LLVMCreateGenericValueOfInt(LLVMInt32Type(), x, 0),
LLVMCreateGenericValueOfInt(LLVMInt32Type(), y, 0),
};
LLVMGenericValueRef res = LLVMRunFunction(engine, sum, 2, args);
printf("%d\n", (int)LLVMGenericValueToInt(res, 0));
// write bitcode to file
if (LLVMWriteBitcodeToFile(mod, "sum.bc") != 0) {
fprintf(stderr, "error writing bitcode to file\n");
}
LLVMDisposeBuilder(builder);
LLVMDisposeExecutionEngine(engine);
}
void llvm_new_compiler(lua_State *L) {
global_State *g = G(L);
if(g_need_init) {
LLVMLinkInJIT();
LLVMInitializeNativeTarget();
g_need_init = 0;
}
g->llvm_compiler = new LLVMCompiler(g_useJIT);
}
void
lp_build_init(void)
{
if (gallivm_initialized)
return;
#ifdef DEBUG
gallivm_debug = debug_get_option_gallivm_debug();
#endif
lp_set_target_options();
#if USE_MCJIT
LLVMLinkInMCJIT();
#else
LLVMLinkInJIT();
#endif
util_cpu_detect();
/* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
* 8-wide vector needs more floating ops than 4-wide (due to padding), it is
* actually more efficient to use 4-wide vectors on this processor.
*
* See also:
* - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
*/
if (HAVE_AVX &&
util_cpu_caps.has_avx &&
util_cpu_caps.has_intel) {
lp_native_vector_width = 256;
} else {
/* Leave it at 128, even when no SIMD extensions are available.
* Really needs to be a multiple of 128 so can fit 4 floats.
*/
lp_native_vector_width = 128;
}
lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
lp_native_vector_width);
gallivm_initialized = TRUE;
#if 0
/* For simulating less capable machines */
util_cpu_caps.has_sse3 = 0;
util_cpu_caps.has_ssse3 = 0;
util_cpu_caps.has_sse4_1 = 0;
#endif
}
static void jit_init_llvm(const char *path)
{
char *error;
LLVMMemoryBufferRef buf;
if (LLVMCreateMemoryBufferWithContentsOfFile(path, &buf, &error))
fatal("error reading bitcode from %s: %s", path, error);
if (LLVMParseBitcode(buf, &module, &error))
fatal("error parsing bitcode: %s", error);
LLVMDisposeMemoryBuffer(buf);
LLVMInitializeNativeTarget();
LLVMLinkInJIT();
if (LLVMCreateExecutionEngineForModule(&exec_engine, module, &error))
fatal("error creating execution engine: %s", error);
}
void
lp_build_init(void)
{
if (gallivm_initialized)
return;
#ifdef DEBUG
gallivm_debug = debug_get_option_gallivm_debug();
#endif
lp_set_target_options();
#if USE_MCJIT
LLVMLinkInMCJIT();
#else
LLVMLinkInJIT();
#endif
util_cpu_detect();
if (HAVE_AVX &&
util_cpu_caps.has_avx) {
lp_native_vector_width = 256;
} else {
/* Leave it at 128, even when no SIMD extensions are available.
* Really needs to be a multiple of 128 so can fit 4 floats.
*/
lp_native_vector_width = 128;
}
lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
lp_native_vector_width);
gallivm_initialized = TRUE;
#if 0
/* For simulating less capable machines */
util_cpu_caps.has_sse3 = 0;
util_cpu_caps.has_ssse3 = 0;
util_cpu_caps.has_sse4_1 = 0;
#endif
}
void JITImpl::init()
{
if (initialized)
return;
LLVMLinkInJIT();
LLVMInitializeNativeTarget();
LLVMMemoryBufferRef memBuffer =
LLVMExtraCreateMemoryBufferWithPtr(instructionBitcode,
instructionBitcodeSize);
char *outMessage;
if (LLVMParseBitcode(memBuffer, &module, &outMessage)) {
std::cerr << "Error loading bitcode: " << outMessage << '\n';
std::abort();
}
// TODO experiment with opt level.
if (LLVMCreateJITCompilerForModule(&executionEngine, module, 1,
&outMessage)) {
std::cerr << "Error creating JIT compiler: " << outMessage << '\n';
std::abort();
}
builder = LLVMCreateBuilder();
LLVMValueRef callee = LLVMGetNamedFunction(module, "jitInstructionTemplate");
assert(callee && "jitInstructionTemplate() not found in module");
jitFunctionType = LLVMGetElementType(LLVMTypeOf(callee));
functions.init(module);
FPM = LLVMCreateFunctionPassManagerForModule(module);
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(executionEngine), FPM);
LLVMAddBasicAliasAnalysisPass(FPM);
LLVMAddJumpThreadingPass(FPM);
LLVMAddGVNPass(FPM);
LLVMAddJumpThreadingPass(FPM);
LLVMAddCFGSimplificationPass(FPM);
LLVMAddDeadStoreEliminationPass(FPM);
LLVMAddInstructionCombiningPass(FPM);
LLVMInitializeFunctionPassManager(FPM);
if (DEBUG_JIT) {
LLVMExtraRegisterJitDisassembler(executionEngine, LLVMGetTarget(module));
}
initialized = true;
}
int main(int argc, char **argv)
{
unsigned verbose = 0;
FILE *fp = NULL;
unsigned long n = 1000;
unsigned i;
boolean success;
for(i = 1; i < argc; ++i) {
if(strcmp(argv[i], "-v") == 0)
++verbose;
else if(strcmp(argv[i], "-o") == 0)
fp = fopen(argv[++i], "wt");
else
n = atoi(argv[i]);
}
LLVMLinkInJIT();
LLVMInitializeNativeTarget();
util_cpu_detect();
if(fp) {
/* Warm up the caches */
test_some(0, NULL, 100);
write_tsv_header(fp);
}
if(n)
success = test_some(verbose, fp, n);
else
success = test_all(verbose, fp);
if(fp)
fclose(fp);
return success ? 0 : 1;
}
/* Force the LLVM interpreter and JIT to be linked in. */
void llvm_initialize(void) {
LLVMLinkInInterpreter();
LLVMLinkInJIT();
}
SWIGEXPORT void JNICALL Java_org_jllvm_bindings_ExecutionEngineJNI_LLVMLinkInJIT(JNIEnv *jenv, jclass jcls) {
(void)jenv;
(void)jcls;
LLVMLinkInJIT();
}
void
lp_build_init(void)
{
#ifdef DEBUG
gallivm_debug = debug_get_option_gallivm_debug();
#endif
lp_set_target_options();
LLVMInitializeNativeTarget();
LLVMLinkInJIT();
if (!lp_build_module)
lp_build_module = LLVMModuleCreateWithName("gallivm");
if (!lp_build_provider)
lp_build_provider = LLVMCreateModuleProviderForExistingModule(lp_build_module);
if (!lp_build_engine) {
enum LLVM_CodeGenOpt_Level optlevel;
char *error = NULL;
if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
optlevel = None;
}
else {
optlevel = Default;
}
if (LLVMCreateJITCompiler(&lp_build_engine, lp_build_provider,
(unsigned)optlevel, &error)) {
_debug_printf("%s\n", error);
LLVMDisposeMessage(error);
assert(0);
}
#if defined(DEBUG) || defined(PROFILE)
lp_register_oprofile_jit_event_listener(lp_build_engine);
#endif
}
if (!lp_build_target)
lp_build_target = LLVMGetExecutionEngineTargetData(lp_build_engine);
if (!lp_build_pass) {
lp_build_pass = LLVMCreateFunctionPassManager(lp_build_provider);
LLVMAddTargetData(lp_build_target, lp_build_pass);
if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
/* TODO: Add more passes */
LLVMAddCFGSimplificationPass(lp_build_pass);
LLVMAddPromoteMemoryToRegisterPass(lp_build_pass);
LLVMAddConstantPropagationPass(lp_build_pass);
if(util_cpu_caps.has_sse4_1) {
/* FIXME: There is a bug in this pass, whereby the combination of fptosi
* and sitofp (necessary for trunc/floor/ceil/round implementation)
* somehow becomes invalid code.
*/
LLVMAddInstructionCombiningPass(lp_build_pass);
}
LLVMAddGVNPass(lp_build_pass);
} else {
/* We need at least this pass to prevent the backends to fail in
* unexpected ways.
*/
LLVMAddPromoteMemoryToRegisterPass(lp_build_pass);
}
}
util_cpu_detect();
#if 0
/* For simulating less capable machines */
util_cpu_caps.has_sse3 = 0;
util_cpu_caps.has_ssse3 = 0;
util_cpu_caps.has_sse4_1 = 0;
#endif
}
