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 codegen_shutdown(pass_opt_t* opt)
{
LLVMDisposeMessage(opt->triple);
LLVMDisposeMessage(opt->cpu);
LLVMDisposeMessage(opt->features);
LLVMShutdown();
}
void codegen_pass_cleanup(pass_opt_t* opt)
{
LLVMDisposeMessage(opt->triple);
LLVMDisposeMessage(opt->cpu);
LLVMDisposeMessage(opt->features);
opt->triple = NULL;
opt->cpu = NULL;
opt->features = NULL;
}
void codegen_shutdown(pass_opt_t* opt)
{
LLVMDisposeMessage(opt->triple);
LLVMDisposeMessage(opt->cpu);
LLVMDisposeMessage(opt->features);
LLVMResetFatalErrorHandler();
LLVMShutdown();
}
static LLVMTargetMachineRef make_machine(pass_opt_t* opt)
{
LLVMTargetRef target;
char* err;
if(LLVMGetTargetFromTriple(opt->triple, &target, &err) != 0)
{
errorf(opt->check.errors, NULL, "couldn't create target: %s", err);
LLVMDisposeMessage(err);
return NULL;
}
LLVMCodeGenOptLevel opt_level =
opt->release ? LLVMCodeGenLevelAggressive : LLVMCodeGenLevelNone;
LLVMRelocMode reloc =
(opt->pic || opt->library) ? LLVMRelocPIC : LLVMRelocDefault;
LLVMTargetMachineRef machine = LLVMCreateTargetMachine(target, opt->triple,
opt->cpu, opt->features, opt_level, reloc, LLVMCodeModelDefault);
if(machine == NULL)
{
errorf(opt->check.errors, NULL, "couldn't create target machine");
return NULL;
}
return machine;
}
static void handle_line(char **tokens, int ntokens) {
char *name = tokens[0];
LLVMValueRef param;
LLVMValueRef res;
LLVMModuleRef M = LLVMModuleCreateWithName(name);
LLVMTypeRef I64ty = LLVMInt64Type();
LLVMTypeRef I64Ptrty = LLVMPointerType(I64ty, 0);
LLVMTypeRef Fty = LLVMFunctionType(I64ty, &I64Ptrty, 1, 0);
LLVMValueRef F = LLVMAddFunction(M, name, Fty);
LLVMBuilderRef builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, LLVMAppendBasicBlock(F, "entry"));
LLVMGetParams(F, ¶m);
LLVMSetValueName(param, "in");
res = build_from_tokens(tokens + 1, ntokens - 1, builder, param);
if (res) {
char *irstr = LLVMPrintModuleToString(M);
puts(irstr);
LLVMDisposeMessage(irstr);
}
LLVMDisposeBuilder(builder);
LLVMDisposeModule(M);
}
/* unit -> string */
CAMLprim value llvm_target_default_triple(value Unit) {
char *TripleCStr = LLVMGetDefaultTargetTriple();
value TripleStr = caml_copy_string(TripleCStr);
LLVMDisposeMessage(TripleCStr);
return TripleStr;
}
void
vm_state_destroy(struct vm_state *vm)
{
LLVMBasicBlockRef current_block, return_block;
char *error;
current_block = LLVMGetInsertBlock(vm->builder);
return_block = LLVMInsertBasicBlock(current_block, "ret");
LLVMPositionBuilderAtEnd(vm->builder, current_block);
LLVMBuildBr(vm->builder, return_block);
LLVMPositionBuilderAtEnd(vm->builder, return_block);
LLVMBuildRetVoid(vm->builder);
LLVMMoveBasicBlockAfter(return_block, current_block);
LLVMDumpModule(vm->module);
error = NULL;
LLVMVerifyModule(vm->module, LLVMAbortProcessAction, &error);
LLVMDisposeMessage(error);
LLVMDisposeBuilder(vm->builder);
LLVMDisposeModule(vm->module);
symbol_table_destroy(vm->symtab);
}
void
ac_dump_module(LLVMModuleRef module)
{
char *str = LLVMPrintModuleToString(module);
fprintf(stderr, "%s", str);
LLVMDisposeMessage(str);
}
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 };
}
static void radeonDiagnosticHandler(LLVMDiagnosticInfoRef di, void *context)
{
struct radeon_llvm_diagnostics *diag = (struct radeon_llvm_diagnostics *)context;
LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
char *description = LLVMGetDiagInfoDescription(di);
const char *severity_str = NULL;
switch (severity) {
case LLVMDSError:
severity_str = "error";
break;
case LLVMDSWarning:
severity_str = "warning";
break;
case LLVMDSRemark:
severity_str = "remark";
break;
case LLVMDSNote:
severity_str = "note";
break;
default:
severity_str = "unknown";
}
pipe_debug_message(diag->debug, SHADER_INFO,
"LLVM diagnostic (%s): %s", severity_str, description);
if (severity == LLVMDSError) {
diag->retval = 1;
fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", description);
}
LLVMDisposeMessage(description);
}
static void init_module(compile_t* c, ast_t* program, pass_opt_t* opt)
{
c->opt = opt;
// Get the first package and the builtin package.
ast_t* package = ast_child(program);
ast_t* builtin = ast_sibling(package);
// If we have only one package, we are compiling builtin itself.
if(builtin == NULL)
builtin = package;
c->reach = reach_new();
// The name of the first package is the name of the program.
c->filename = package_filename(package);
// LLVM context and machine settings.
if(c->opt->library || target_is_ilp32(opt->triple))
c->callconv = LLVMCCallConv;
else
c->callconv = LLVMFastCallConv;
if(!c->opt->release || c->opt->library || c->opt->extfun)
c->linkage = LLVMExternalLinkage;
else
c->linkage = LLVMPrivateLinkage;
c->context = LLVMContextCreate();
c->machine = make_machine(opt);
c->target_data = LLVMGetTargetMachineData(c->machine);
// Create a module.
c->module = LLVMModuleCreateWithNameInContext(c->filename, c->context);
// Set the target triple.
LLVMSetTarget(c->module, opt->triple);
// Set the data layout.
char* layout = LLVMCopyStringRepOfTargetData(c->target_data);
LLVMSetDataLayout(c->module, layout);
LLVMDisposeMessage(layout);
// IR builder.
c->builder = LLVMCreateBuilderInContext(c->context);
c->di = LLVMNewDIBuilder(c->module);
// TODO: what LANG id should be used?
c->di_unit = LLVMDIBuilderCreateCompileUnit(c->di, 0x0004,
package_filename(package), package_path(package), "ponyc-" PONY_VERSION,
c->opt->release);
// Empty frame stack.
c->frame = NULL;
}
SCM llvm_verify_module(SCM scm_llvm)
{
struct llvm_module_t *llvm = get_llvm(scm_llvm);
char *error = NULL;
if (LLVMVerifyModule(llvm->module, LLVMPrintMessageAction, &error)) {
SCM scm_error = scm_from_locale_string(error);
LLVMDisposeMessage(error);
scm_misc_error("verify-module", "Module is not valid: ~a", scm_list_1(scm_error));
};
return SCM_UNSPECIFIED;
}
int module_dump(bool Lazy, bool New) {
LLVMModuleRef M = load_module(Lazy, New);
char *irstr = LLVMPrintModuleToString(M);
puts(irstr);
LLVMDisposeMessage(irstr);
LLVMDisposeModule(M);
return 0;
}
static void radeonDiagnosticHandler(LLVMDiagnosticInfoRef di, void *context)
{
if (LLVMGetDiagInfoSeverity(di) == LLVMDSError) {
unsigned int *diagnosticflag = (unsigned int *)context;
char *diaginfo_message = LLVMGetDiagInfoDescription(di);
*diagnosticflag = 1;
fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", diaginfo_message);
LLVMDisposeMessage(diaginfo_message);
}
}
/* ExecutionEngine.t -> Llvm_target.DataLayout.t */
CAMLprim value llvm_ee_get_data_layout(LLVMExecutionEngineRef EE) {
value DataLayout;
LLVMTargetDataRef OrigDataLayout;
char* TargetDataCStr;
OrigDataLayout = LLVMGetExecutionEngineTargetData(EE);
TargetDataCStr = LLVMCopyStringRepOfTargetData(OrigDataLayout);
DataLayout = llvm_alloc_data_layout(LLVMCreateTargetData(TargetDataCStr));
LLVMDisposeMessage(TargetDataCStr);
return DataLayout;
}
static void llvm_raise(value Prototype, char *Message) {
CAMLparam1(Prototype);
CAMLlocal1(CamlMessage);
CamlMessage = copy_string(Message);
LLVMDisposeMessage(Message);
raise_with_arg(Prototype, CamlMessage);
abort(); /* NOTREACHED */
#ifdef CAMLnoreturn
CAMLnoreturn; /* Silences warnings, but is missing in some versions. */
#endif
}
/* TargetMachine.t -> DataLayout.t */
CAMLprim value llvm_targetmachine_data_layout(value Machine) {
CAMLparam1(Machine);
CAMLlocal1(DataLayout);
/* LLVMGetTargetMachineData returns a pointer owned by the TargetMachine,
so it is impossible to wrap it with llvm_alloc_target_data, which assumes
that OCaml owns the pointer. */
LLVMTargetDataRef OrigDataLayout;
OrigDataLayout = LLVMGetTargetMachineData(TargetMachine_val(Machine));
char* TargetDataCStr;
TargetDataCStr = LLVMCopyStringRepOfTargetData(OrigDataLayout);
DataLayout = llvm_alloc_data_layout(LLVMCreateTargetData(TargetDataCStr));
LLVMDisposeMessage(TargetDataCStr);
CAMLreturn(DataLayout);
}
LLVMTargetRef ac_get_llvm_target(const char *triple)
{
LLVMTargetRef target = NULL;
char *err_message = NULL;
call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
fprintf(stderr, "Cannot find target for triple %s ", triple);
if (err_message) {
fprintf(stderr, "%s\n", err_message);
}
LLVMDisposeMessage(err_message);
return NULL;
}
return target;
}
LLVMTargetRef radeon_llvm_get_r600_target(const char *triple)
{
LLVMTargetRef target = NULL;
char *err_message = NULL;
init_r600_target();
if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
fprintf(stderr, "Cannot find target for triple %s ", triple);
if (err_message) {
fprintf(stderr, "%s\n", err_message);
}
LLVMDisposeMessage(err_message);
return NULL;
}
return target;
}
SCM llvm_module_destroy(SCM scm_self)
{
struct llvm_module_t *self = get_llvm_no_check(scm_self);
if (self->engine) {
if (self->module) {
char *error = NULL;
LLVMRemoveModule(self->engine, self->module, &self->module, &error);
if (error) LLVMDisposeMessage(error);
};
LLVMDisposeExecutionEngine(self->engine);
self->engine = NULL;
};
if (self->module) {
LLVMDisposeModule(self->module);
self->module = NULL;
};
return SCM_UNSPECIFIED;
}
LLVMTypeRef ty_get_counted(LLVMTypeRef in)
{
char *translated = LLVMPrintTypeToString(in);
LLVMTypeRef ref = hst_get(&counted_table, translated, NULL, NULL);
if(!ref)
{
ref = LLVMStructCreateNamed(utl_get_current_context(), "");
LLVMTypeRef tys[6];
LLVMGetStructElementTypes(in, tys);
LLVMStructSetBody(ref, tys, 6, 0);
hst_put(&counted_table, translated, ref, NULL, NULL);
}
LLVMDisposeMessage(translated);
return ref;
}
static void init_module(compile_t* c, ast_t* program, pass_opt_t* opt)
{
c->opt = opt;
// Get the first package and the builtin package.
ast_t* package = ast_child(program);
ast_t* builtin = ast_sibling(package);
// If we have only one package, we are compiling builtin itself.
if(builtin == NULL)
builtin = package;
c->reachable = reach_new();
reach_primitives(c->reachable, opt, builtin);
// The name of the first package is the name of the program.
c->filename = package_filename(package);
// LLVM context and machine settings.
c->context = LLVMContextCreate();
c->machine = make_machine(opt);
c->target_data = LLVMGetTargetMachineData(c->machine);
// Create a module.
c->module = LLVMModuleCreateWithNameInContext(c->filename, c->context);
// Set the target triple.
LLVMSetTarget(c->module, opt->triple);
// Set the data layout.
char* layout = LLVMCopyStringRepOfTargetData(c->target_data);
LLVMSetDataLayout(c->module, layout);
LLVMDisposeMessage(layout);
// IR builder.
c->builder = LLVMCreateBuilderInContext(c->context);
// Empty frame stack.
c->frame = NULL;
}
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;
}
int module_list_globals(void) {
LLVMModuleRef M = load_module(false, false);
LLVMValueRef g;
g = LLVMGetFirstGlobal(M);
while (g) {
LLVMTypeRef T = LLVMTypeOf(g);
char *s = LLVMPrintTypeToString(T);
printf("Global%s: %s %s\n",
LLVMIsDeclaration(g) ? "Declaration" : "Definition",
LLVMGetValueName(g), s);
LLVMDisposeMessage(s);
g = LLVMGetNextGlobal(g);
}
LLVMDisposeModule(M);
return 0;
}
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;
}
/**
* Allocate gallivm LLVM objects.
* \return TRUE for success, FALSE for failure
*/
static boolean
init_gallivm_state(struct gallivm_state *gallivm)
{
assert(!gallivm->context);
assert(!gallivm->module);
assert(!gallivm->provider);
lp_build_init();
gallivm->context = LLVMContextCreate();
if (!gallivm->context)
goto fail;
gallivm->module = LLVMModuleCreateWithNameInContext("gallivm",
gallivm->context);
if (!gallivm->module)
goto fail;
gallivm->provider =
LLVMCreateModuleProviderForExistingModule(gallivm->module);
if (!gallivm->provider)
goto fail;
if (!GlobalEngine) {
/* We can only create one LLVMExecutionEngine (w/ LLVM 2.6 anyway) */
enum LLVM_CodeGenOpt_Level optlevel;
char *error = NULL;
if (gallivm_debug & GALLIVM_DEBUG_NO_OPT) {
optlevel = None;
}
else {
optlevel = Default;
}
if (LLVMCreateJITCompiler(&GlobalEngine, gallivm->provider,
(unsigned) optlevel, &error)) {
_debug_printf("%s\n", error);
LLVMDisposeMessage(error);
goto fail;
}
#if defined(DEBUG) || defined(PROFILE)
lp_register_oprofile_jit_event_listener(GlobalEngine);
#endif
}
gallivm->engine = GlobalEngine;
LLVMAddModuleProvider(gallivm->engine, gallivm->provider);//new
gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
if (!gallivm->target)
goto fail;
if (!create_pass_manager(gallivm))
goto fail;
gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
if (!gallivm->builder)
goto fail;
return TRUE;
fail:
free_gallivm_state(gallivm);
return FALSE;
}
/* TargetData.t -> string */
CAMLprim value llvm_targetdata_as_string(LLVMTargetDataRef TD) {
char *StringRep = LLVMCopyStringRepOfTargetData(TD);
value Copy = copy_string(StringRep);
LLVMDisposeMessage(StringRep);
return Copy;
}
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;
}
static void diagnosticHandler(LLVMDiagnosticInfoRef DI, void *C) {
char *CErr = LLVMGetDiagInfoDescription(DI);
fprintf(stderr, "Error with new bitcode parser: %s\n", CErr);
LLVMDisposeMessage(CErr);
exit(1);
}