SWIGEXPORT void JNICALL Java_org_jllvm_bindings_ScalarJNI_LLVMAddGVNPass(JNIEnv *jenv, jclass jcls, jlong jarg1) {
LLVMPassManagerRef arg1 = (LLVMPassManagerRef) 0 ;
(void)jenv;
(void)jcls;
arg1 = *(LLVMPassManagerRef *)&jarg1;
LLVMAddGVNPass(arg1);
}
/**
* Create the LLVM (optimization) pass manager and install
* relevant optimization passes.
* \return TRUE for success, FALSE for failure
*/
static boolean
create_pass_manager(struct gallivm_state *gallivm)
{
assert(!gallivm->passmgr);
assert(gallivm->target);
gallivm->passmgr = LLVMCreateFunctionPassManager(gallivm->provider);
if (!gallivm->passmgr)
return FALSE;
LLVMAddTargetData(gallivm->target, gallivm->passmgr);
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(gallivm->passmgr);
if (HAVE_LLVM >= 0x207 && sizeof(void*) == 4) {
/* For LLVM >= 2.7 and 32-bit build, use this order of passes to
* avoid generating bad code.
* Test with piglit glsl-vs-sqrt-zero test.
*/
LLVMAddConstantPropagationPass(gallivm->passmgr);
LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
}
else {
LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
LLVMAddConstantPropagationPass(gallivm->passmgr);
}
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(gallivm->passmgr);
}
LLVMAddGVNPass(gallivm->passmgr);
}
else {
/* We need at least this pass to prevent the backends to fail in
* unexpected ways.
*/
LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
}
return TRUE;
}
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 LLVM_execute(LLVMCompiledProgram program)
{
LLVMModuleRef module = program.module;
LLVMValueRef function = program.function;
char *error = NULL; // Used to retrieve messages from functions
LLVMExecutionEngineRef engine;
LLVMModuleProviderRef provider = LLVMCreateModuleProviderForExistingModule(module);
error = NULL;
if(LLVMCreateJITCompiler(&engine, provider, 2, &error) != 0) {
fprintf(stderr, "%s\n", error);
LLVMDisposeMessage(error);
abort();
}
LLVMPassManagerRef pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#ifdef NDEBUG
#else
LLVMDumpModule(module);
#endif
LLVMGenericValueRef exec_args[] = {};
LLVMGenericValueRef exec_res = LLVMRunFunction(engine, function, 0, exec_args);
int result = LLVMGenericValueToInt(exec_res, 0);
LLVMDisposePassManager(pass);
LLVMDisposeExecutionEngine(engine);
return result;
}
SCM llvm_compile_module(SCM scm_llvm, SCM scm_name)
{
struct llvm_module_t *self = get_llvm(scm_llvm);
if (self->engine != NULL)
scm_misc_error("llvm-compile", "LLVM module already compiled", SCM_EOL);
char *error = NULL;
if (LLVMCreateJITCompilerForModule(&self->engine, self->module, 2, &error)) {
SCM scm_error = scm_from_locale_string(error);
LLVMDisposeMessage(error);
scm_misc_error("llvm-compile", "Error initialising JIT engine: ~a", scm_list_1(scm_error));
};
LLVMPassManagerRef pass_manager = LLVMCreatePassManager();
LLVMAddConstantPropagationPass(pass_manager);
LLVMAddInstructionCombiningPass(pass_manager);
LLVMAddPromoteMemoryToRegisterPass(pass_manager);
LLVMAddGVNPass(pass_manager);
LLVMAddCFGSimplificationPass(pass_manager);
LLVMRunPassManager(pass_manager, self->module);
LLVMDisposePassManager(pass_manager);
return SCM_UNSPECIFIED;
}
ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
const struct pipe_blend_state *blend,
enum vector_mode mode,
struct lp_type type)
{
LLVMModuleRef module = NULL;
LLVMValueRef func = NULL;
LLVMExecutionEngineRef engine = NULL;
LLVMModuleProviderRef provider = NULL;
LLVMPassManagerRef pass = NULL;
char *error = NULL;
blend_test_ptr_t blend_test_ptr;
boolean success;
const unsigned n = LP_TEST_NUM_SAMPLES;
int64_t cycles[LP_TEST_NUM_SAMPLES];
double cycles_avg = 0.0;
unsigned i, j;
if(verbose >= 1)
dump_blend_type(stdout, blend, mode, type);
module = LLVMModuleCreateWithName("test");
func = add_blend_test(module, blend, mode, type);
if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
LLVMDumpModule(module);
abort();
}
LLVMDisposeMessage(error);
provider = LLVMCreateModuleProviderForExistingModule(module);
if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) {
if(verbose < 1)
dump_blend_type(stderr, blend, mode, type);
fprintf(stderr, "%s\n", error);
LLVMDisposeMessage(error);
abort();
}
#if 0
pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#else
(void)pass;
#endif
if(verbose >= 2)
LLVMDumpModule(module);
blend_test_ptr = (blend_test_ptr_t)LLVMGetPointerToGlobal(engine, func);
if(verbose >= 2)
lp_disassemble(blend_test_ptr);
success = TRUE;
for(i = 0; i < n && success; ++i) {
if(mode == AoS) {
ALIGN16_ATTRIB uint8_t src[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t dst[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t con[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t res[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t ref[LP_NATIVE_VECTOR_WIDTH/8];
int64_t start_counter = 0;
int64_t end_counter = 0;
random_vec(type, src);
random_vec(type, dst);
random_vec(type, con);
{
double fsrc[LP_MAX_VECTOR_LENGTH];
double fdst[LP_MAX_VECTOR_LENGTH];
double fcon[LP_MAX_VECTOR_LENGTH];
double fref[LP_MAX_VECTOR_LENGTH];
read_vec(type, src, fsrc);
read_vec(type, dst, fdst);
read_vec(type, con, fcon);
for(j = 0; j < type.length; j += 4)
compute_blend_ref(blend, fsrc + j, fdst + j, fcon + j, fref + j);
write_vec(type, ref, fref);
}
start_counter = rdtsc();
blend_test_ptr(src, dst, con, res);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
if(!compare_vec(type, res, ref)) {
success = FALSE;
if(verbose < 1)
dump_blend_type(stderr, blend, mode, type);
fprintf(stderr, "MISMATCH\n");
fprintf(stderr, " Src: ");
dump_vec(stderr, type, src);
fprintf(stderr, "\n");
fprintf(stderr, " Dst: ");
dump_vec(stderr, type, dst);
fprintf(stderr, "\n");
fprintf(stderr, " Con: ");
dump_vec(stderr, type, con);
fprintf(stderr, "\n");
fprintf(stderr, " Res: ");
dump_vec(stderr, type, res);
fprintf(stderr, "\n");
fprintf(stderr, " Ref: ");
dump_vec(stderr, type, ref);
fprintf(stderr, "\n");
}
}
if(mode == SoA) {
const unsigned stride = type.length*type.width/8;
ALIGN16_ATTRIB uint8_t src[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t dst[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t con[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t res[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t ref[4*LP_NATIVE_VECTOR_WIDTH/8];
int64_t start_counter = 0;
int64_t end_counter = 0;
boolean mismatch;
for(j = 0; j < 4; ++j) {
random_vec(type, src + j*stride);
random_vec(type, dst + j*stride);
random_vec(type, con + j*stride);
}
{
double fsrc[4];
double fdst[4];
double fcon[4];
double fref[4];
unsigned k;
for(k = 0; k < type.length; ++k) {
for(j = 0; j < 4; ++j) {
fsrc[j] = read_elem(type, src + j*stride, k);
fdst[j] = read_elem(type, dst + j*stride, k);
fcon[j] = read_elem(type, con + j*stride, k);
}
compute_blend_ref(blend, fsrc, fdst, fcon, fref);
for(j = 0; j < 4; ++j)
write_elem(type, ref + j*stride, k, fref[j]);
}
}
start_counter = rdtsc();
blend_test_ptr(src, dst, con, res);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
mismatch = FALSE;
for (j = 0; j < 4; ++j)
if(!compare_vec(type, res + j*stride, ref + j*stride))
mismatch = TRUE;
if (mismatch) {
success = FALSE;
if(verbose < 1)
dump_blend_type(stderr, blend, mode, type);
fprintf(stderr, "MISMATCH\n");
for(j = 0; j < 4; ++j) {
char channel = "RGBA"[j];
fprintf(stderr, " Src%c: ", channel);
dump_vec(stderr, type, src + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Dst%c: ", channel);
dump_vec(stderr, type, dst + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Con%c: ", channel);
dump_vec(stderr, type, con + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Res%c: ", channel);
dump_vec(stderr, type, res + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Ref%c: ", channel);
dump_vec(stderr, type, ref + j*stride);
fprintf(stderr, "\n");
}
}
}
}
/*
* Unfortunately the output of cycle counter is not very reliable as it comes
* -- sometimes we get outliers (due IRQs perhaps?) which are
* better removed to avoid random or biased data.
*/
{
double sum = 0.0, sum2 = 0.0;
double avg, std;
unsigned m;
for(i = 0; i < n; ++i) {
sum += cycles[i];
sum2 += cycles[i]*cycles[i];
}
avg = sum/n;
std = sqrtf((sum2 - n*avg*avg)/n);
m = 0;
sum = 0.0;
for(i = 0; i < n; ++i) {
if(fabs(cycles[i] - avg) <= 4.0*std) {
sum += cycles[i];
++m;
}
}
cycles_avg = sum/m;
}
if(fp)
write_tsv_row(fp, blend, mode, type, cycles_avg, success);
if (!success) {
if(verbose < 2)
LLVMDumpModule(module);
LLVMWriteBitcodeToFile(module, "blend.bc");
fprintf(stderr, "blend.bc written\n");
fprintf(stderr, "Invoke as \"llc -o - blend.bc\"\n");
abort();
}
LLVMFreeMachineCodeForFunction(engine, func);
LLVMDisposeExecutionEngine(engine);
if(pass)
LLVMDisposePassManager(pass);
return success;
}
PIPE_ALIGN_STACK
static boolean
test_round(unsigned verbose, FILE *fp)
{
LLVMModuleRef module = NULL;
LLVMValueRef test_round = NULL, test_trunc, test_floor, test_ceil;
LLVMExecutionEngineRef engine = lp_build_engine;
LLVMPassManagerRef pass = NULL;
char *error = NULL;
test_round_t round_func, trunc_func, floor_func, ceil_func;
float unpacked[4];
unsigned packed;
boolean success = TRUE;
int i;
module = LLVMModuleCreateWithName("test");
test_round = add_test(module, "round", lp_build_round);
test_trunc = add_test(module, "trunc", lp_build_trunc);
test_floor = add_test(module, "floor", lp_build_floor);
test_ceil = add_test(module, "ceil", lp_build_ceil);
if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
printf("LLVMVerifyModule: %s\n", error);
LLVMDumpModule(module);
abort();
}
LLVMDisposeMessage(error);
#if 0
pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#else
(void)pass;
#endif
round_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_round));
trunc_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_trunc));
floor_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_floor));
ceil_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_ceil));
memset(unpacked, 0, sizeof unpacked);
packed = 0;
if (0)
LLVMDumpModule(module);
for (i = 0; i < 3; i++) {
v4sf xvals[3] = {
{-10.0, -1, 0, 12.0},
{-1.5, -0.25, 1.25, 2.5},
{-0.99, -0.01, 0.01, 0.99}
};
v4sf x = xvals[i];
v4sf y, ref;
float *xp = (float *) &x;
float *refp = (float *) &ref;
printf("\n");
printv("x ", x);
refp[0] = round(xp[0]);
refp[1] = round(xp[1]);
refp[2] = round(xp[2]);
refp[3] = round(xp[3]);
y = round_func(x);
printv("C round(x) ", ref);
printv("LLVM round(x)", y);
compare(ref, y);
refp[0] = trunc(xp[0]);
refp[1] = trunc(xp[1]);
refp[2] = trunc(xp[2]);
refp[3] = trunc(xp[3]);
y = trunc_func(x);
printv("C trunc(x) ", ref);
printv("LLVM trunc(x)", y);
compare(ref, y);
refp[0] = floor(xp[0]);
refp[1] = floor(xp[1]);
refp[2] = floor(xp[2]);
refp[3] = floor(xp[3]);
y = floor_func(x);
printv("C floor(x) ", ref);
printv("LLVM floor(x)", y);
compare(ref, y);
refp[0] = ceil(xp[0]);
refp[1] = ceil(xp[1]);
refp[2] = ceil(xp[2]);
refp[3] = ceil(xp[3]);
y = ceil_func(x);
printv("C ceil(x) ", ref);
printv("LLVM ceil(x) ", y);
compare(ref, y);
}
LLVMFreeMachineCodeForFunction(engine, test_round);
LLVMFreeMachineCodeForFunction(engine, test_trunc);
LLVMFreeMachineCodeForFunction(engine, test_floor);
LLVMFreeMachineCodeForFunction(engine, test_ceil);
LLVMDisposeExecutionEngine(engine);
if(pass)
LLVMDisposePassManager(pass);
return success;
}
/* [<Llvm.PassManager.any] Llvm.PassManager.t -> unit */
CAMLprim value llvm_add_gvn(LLVMPassManagerRef PM) {
LLVMAddGVNPass(PM);
return Val_unit;
}
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
}
int main(int c, char **v)
{
LLVMContextRef *contexts;
LLVMModuleRef *modules;
char *error;
const char *mode = "opt";
const char **filenames;
unsigned numFiles;
unsigned i;
bool moreOptions;
static int verboseFlag = 0;
static int timingFlag = 0;
static int disassembleFlag = 0;
bool manyContexts = true;
double beforeAll;
if (c == 1)
usage();
moreOptions = true;
while (moreOptions) {
static struct option longOptions[] = {
{"verbose", no_argument, &verboseFlag, 1},
{"timing", no_argument, &timingFlag, 1},
{"disassemble", no_argument, &disassembleFlag, 1},
{"mode", required_argument, 0, 0},
{"contexts", required_argument, 0, 0},
{"help", no_argument, 0, 0}
};
int optionIndex;
int optionValue;
optionValue = getopt_long(c, v, "", longOptions, &optionIndex);
switch (optionValue) {
case -1:
moreOptions = false;
break;
case 0: {
const char* thisOption = longOptions[optionIndex].name;
if (!strcmp(thisOption, "help"))
usage();
if (!strcmp(thisOption, "contexts")) {
if (!strcasecmp(optarg, "one"))
manyContexts = false;
else if (!strcasecmp(optarg, "many"))
manyContexts = true;
else {
fprintf(stderr, "Invalid argument for --contexts.\n");
exit(1);
}
break;
}
if (!strcmp(thisOption, "mode")) {
mode = strdup(optarg);
break;
}
break;
}
case '?':
exit(0);
break;
default:
printf("optionValue = %d\n", optionValue);
abort();
break;
}
}
LLVMLinkInMCJIT();
LLVMInitializeNativeTarget();
LLVMInitializeX86AsmPrinter();
LLVMInitializeX86Disassembler();
filenames = (const char **)(v + optind);
numFiles = c - optind;
contexts = malloc(sizeof(LLVMContextRef) * numFiles);
modules = malloc(sizeof(LLVMModuleRef) * numFiles);
if (manyContexts) {
for (i = 0; i < numFiles; ++i)
contexts[i] = LLVMContextCreate();
} else {
LLVMContextRef context = LLVMContextCreate();
for (i = 0; i < numFiles; ++i)
contexts[i] = context;
}
for (i = 0; i < numFiles; ++i) {
LLVMMemoryBufferRef buffer;
const char* filename = filenames[i];
if (LLVMCreateMemoryBufferWithContentsOfFile(filename, &buffer, &error)) {
fprintf(stderr, "Error reading file %s: %s\n", filename, error);
exit(1);
}
if (LLVMParseBitcodeInContext(contexts[i], buffer, modules + i, &error)) {
fprintf(stderr, "Error parsing file %s: %s\n", filename, error);
exit(1);
}
LLVMDisposeMemoryBuffer(buffer);
if (verboseFlag) {
printf("Module #%u (%s) after parsing:\n", i, filename);
LLVMDumpModule(modules[i]);
}
}
if (verboseFlag)
printf("Generating code for modules...\n");
if (timingFlag)
beforeAll = currentTime();
for (i = 0; i < numFiles; ++i) {
LLVMModuleRef module;
LLVMExecutionEngineRef engine;
struct LLVMMCJITCompilerOptions options;
LLVMValueRef value;
LLVMPassManagerRef functionPasses = 0;
LLVMPassManagerRef modulePasses = 0;
double before;
if (timingFlag)
before = currentTime();
module = modules[i];
LLVMInitializeMCJITCompilerOptions(&options, sizeof(options));
options.OptLevel = 2;
options.EnableFastISel = 0;
options.MCJMM = LLVMCreateSimpleMCJITMemoryManager(
0, mmAllocateCodeSection, mmAllocateDataSection, mmApplyPermissions, mmDestroy);
if (LLVMCreateMCJITCompilerForModule(&engine, module, &options, sizeof(options), &error)) {
fprintf(stderr, "Error building MCJIT: %s\n", error);
exit(1);
}
if (!strcasecmp(mode, "simple")) {
modulePasses = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), modulePasses);
LLVMAddConstantPropagationPass(modulePasses);
LLVMAddInstructionCombiningPass(modulePasses);
LLVMAddPromoteMemoryToRegisterPass(modulePasses);
LLVMAddBasicAliasAnalysisPass(modulePasses);
LLVMAddTypeBasedAliasAnalysisPass(modulePasses);
LLVMAddGVNPass(modulePasses);
LLVMAddCFGSimplificationPass(modulePasses);
LLVMRunPassManager(modulePasses, module);
} else if (!strcasecmp(mode, "opt")) {
LLVMPassManagerBuilderRef passBuilder;
passBuilder = LLVMPassManagerBuilderCreate();
LLVMPassManagerBuilderSetOptLevel(passBuilder, 2);
LLVMPassManagerBuilderSetSizeLevel(passBuilder, 0);
functionPasses = LLVMCreateFunctionPassManagerForModule(module);
modulePasses = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), modulePasses);
LLVMPassManagerBuilderPopulateFunctionPassManager(passBuilder, functionPasses);
LLVMPassManagerBuilderPopulateModulePassManager(passBuilder, modulePasses);
LLVMPassManagerBuilderDispose(passBuilder);
LLVMInitializeFunctionPassManager(functionPasses);
for (value = LLVMGetFirstFunction(module); value; value = LLVMGetNextFunction(value))
LLVMRunFunctionPassManager(functionPasses, value);
LLVMFinalizeFunctionPassManager(functionPasses);
LLVMRunPassManager(modulePasses, module);
} else {
fprintf(stderr, "Bad optimization mode: %s.\n", mode);
fprintf(stderr, "Valid modes are: \"simple\" or \"opt\".\n");
exit(1);
}
if (verboseFlag) {
printf("Module #%d (%s) after optimization:\n", i, filenames[i]);
LLVMDumpModule(module);
}
for (value = LLVMGetFirstFunction(module); value; value = LLVMGetNextFunction(value)) {
if (LLVMIsDeclaration(value))
continue;
LLVMGetPointerToGlobal(engine, value);
}
if (functionPasses)
LLVMDisposePassManager(functionPasses);
if (modulePasses)
LLVMDisposePassManager(modulePasses);
LLVMDisposeExecutionEngine(engine);
if (timingFlag) {
double after = currentTime();
printf("Module #%d (%s) took %lf ms.\n", i, filenames[i], (after - before) * 1000);
}
}
if (timingFlag) {
double after = currentTime();
printf("Compilation took a total of %lf ms.\n", (after - beforeAll) * 1000);
}
if (disassembleFlag) {
LLVMDisasmContextRef disassembler;
struct MemorySection *section;
disassembler = LLVMCreateDisasm("x86_64-apple-darwin", 0, 0, 0, symbolLookupCallback);
if (!disassembler) {
fprintf(stderr, "Error building disassembler.\n");
exit(1);
}
for (section = sectionHead; section; section = section->next) {
printf("Disassembly for section %p:\n", section);
char pcString[20];
char instructionString[1000];
uint8_t *pc;
uint8_t *end;
pc = section->start;
end = pc + section->size;
while (pc < end) {
snprintf(
pcString, sizeof(pcString), "0x%lx",
(unsigned long)(uintptr_t)pc);
size_t instructionSize = LLVMDisasmInstruction(
disassembler, pc, end - pc, (uintptr_t)pc,
instructionString, sizeof(instructionString));
if (!instructionSize)
snprintf(instructionString, sizeof(instructionString), ".byte 0x%02x", *pc++);
else
pc += instructionSize;
printf(" %16s: %s\n", pcString, instructionString);
}
}
}
return 0;
}
PIPE_ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
struct lp_type src_type,
struct lp_type dst_type)
{
LLVMModuleRef module = NULL;
LLVMValueRef func = NULL;
LLVMExecutionEngineRef engine = NULL;
LLVMModuleProviderRef provider = NULL;
LLVMPassManagerRef pass = NULL;
char *error = NULL;
conv_test_ptr_t conv_test_ptr;
boolean success;
const unsigned n = LP_TEST_NUM_SAMPLES;
int64_t cycles[LP_TEST_NUM_SAMPLES];
double cycles_avg = 0.0;
unsigned num_srcs;
unsigned num_dsts;
double eps;
unsigned i, j;
if(verbose >= 1)
dump_conv_types(stdout, src_type, dst_type);
if(src_type.length > dst_type.length) {
num_srcs = 1;
num_dsts = src_type.length/dst_type.length;
}
else {
num_dsts = 1;
num_srcs = dst_type.length/src_type.length;
}
assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
eps = MAX2(lp_const_eps(src_type), lp_const_eps(dst_type));
module = LLVMModuleCreateWithName("test");
func = add_conv_test(module, src_type, num_srcs, dst_type, num_dsts);
if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
LLVMDumpModule(module);
abort();
}
LLVMDisposeMessage(error);
provider = LLVMCreateModuleProviderForExistingModule(module);
if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) {
if(verbose < 1)
dump_conv_types(stderr, src_type, dst_type);
fprintf(stderr, "%s\n", error);
LLVMDisposeMessage(error);
abort();
}
#if 0
pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#else
(void)pass;
#endif
if(verbose >= 2)
LLVMDumpModule(module);
conv_test_ptr = (conv_test_ptr_t)LLVMGetPointerToGlobal(engine, func);
if(verbose >= 2)
lp_disassemble(conv_test_ptr);
success = TRUE;
for(i = 0; i < n && success; ++i) {
unsigned src_stride = src_type.length*src_type.width/8;
unsigned dst_stride = dst_type.length*dst_type.width/8;
PIPE_ALIGN_VAR(16) uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
PIPE_ALIGN_VAR(16) uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
double fref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
int64_t start_counter = 0;
int64_t end_counter = 0;
for(j = 0; j < num_srcs; ++j) {
random_vec(src_type, src + j*src_stride);
read_vec(src_type, src + j*src_stride, fref + j*src_type.length);
}
for(j = 0; j < num_dsts; ++j) {
write_vec(dst_type, ref + j*dst_stride, fref + j*dst_type.length);
}
start_counter = rdtsc();
conv_test_ptr(src, dst);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
for(j = 0; j < num_dsts; ++j) {
if(!compare_vec_with_eps(dst_type, dst + j*dst_stride, ref + j*dst_stride, eps))
success = FALSE;
}
if (!success) {
if(verbose < 1)
dump_conv_types(stderr, src_type, dst_type);
fprintf(stderr, "MISMATCH\n");
for(j = 0; j < num_srcs; ++j) {
fprintf(stderr, " Src%u: ", j);
dump_vec(stderr, src_type, src + j*src_stride);
fprintf(stderr, "\n");
}
#if 1
fprintf(stderr, " Ref: ");
for(j = 0; j < src_type.length*num_srcs; ++j)
fprintf(stderr, " %f", fref[j]);
fprintf(stderr, "\n");
#endif
for(j = 0; j < num_dsts; ++j) {
fprintf(stderr, " Dst%u: ", j);
dump_vec(stderr, dst_type, dst + j*dst_stride);
fprintf(stderr, "\n");
fprintf(stderr, " Ref%u: ", j);
dump_vec(stderr, dst_type, ref + j*dst_stride);
fprintf(stderr, "\n");
}
}
}
/*
* Unfortunately the output of cycle counter is not very reliable as it comes
* -- sometimes we get outliers (due IRQs perhaps?) which are
* better removed to avoid random or biased data.
*/
{
double sum = 0.0, sum2 = 0.0;
double avg, std;
unsigned m;
for(i = 0; i < n; ++i) {
sum += cycles[i];
sum2 += cycles[i]*cycles[i];
}
avg = sum/n;
std = sqrtf((sum2 - n*avg*avg)/n);
m = 0;
sum = 0.0;
for(i = 0; i < n; ++i) {
if(fabs(cycles[i] - avg) <= 4.0*std) {
sum += cycles[i];
++m;
}
}
cycles_avg = sum/m;
}
if(fp)
write_tsv_row(fp, src_type, dst_type, cycles_avg, success);
if (!success) {
static boolean firsttime = TRUE;
if(firsttime) {
if(verbose < 2)
LLVMDumpModule(module);
LLVMWriteBitcodeToFile(module, "conv.bc");
fprintf(stderr, "conv.bc written\n");
fprintf(stderr, "Invoke as \"llc -o - conv.bc\"\n");
firsttime = FALSE;
/* abort(); */
}
}
LLVMFreeMachineCodeForFunction(engine, func);
LLVMDisposeExecutionEngine(engine);
if(pass)
LLVMDisposePassManager(pass);
return success;
}