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;
}
LLVMModuleRef radeon_llvm_parse_bitcode(const unsigned char * bitcode,
unsigned bitcode_len)
{
LLVMMemoryBufferRef buf;
LLVMContextRef ctx = LLVMContextCreate();
LLVMModuleRef module;
buf = LLVMCreateMemoryBufferWithMemoryRangeCopy((const char*)bitcode,
bitcode_len, "radeon");
LLVMParseBitcodeInContext(ctx, buf, &module, NULL);
LLVMDisposeMemoryBuffer(buf);
return module;
}
void *evergreen_create_compute_state(
struct pipe_context *ctx_,
const const struct pipe_compute_state *cso)
{
struct r600_context *ctx = (struct r600_context *)ctx_;
struct r600_pipe_compute *shader = CALLOC_STRUCT(r600_pipe_compute);
#ifdef HAVE_OPENCL
const struct pipe_llvm_program_header * header;
const char *code;
void *p;
boolean use_kill;
COMPUTE_DBG(ctx->screen, "*** evergreen_create_compute_state\n");
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
#if HAVE_LLVM < 0x0306
(void)use_kill;
(void)p;
shader->llvm_ctx = LLVMContextCreate();
shader->num_kernels = radeon_llvm_get_num_kernels(shader->llvm_ctx,
code, header->num_bytes);
shader->kernels = CALLOC(sizeof(struct r600_kernel),
shader->num_kernels);
{
unsigned i;
for (i = 0; i < shader->num_kernels; i++) {
struct r600_kernel *kernel = &shader->kernels[i];
kernel->llvm_module = radeon_llvm_get_kernel_module(
shader->llvm_ctx, i, code, header->num_bytes);
}
}
#else
radeon_shader_binary_init(&shader->binary);
radeon_elf_read(code, header->num_bytes, &shader->binary);
r600_create_shader(&shader->bc, &shader->binary, &use_kill);
shader->code_bo = r600_compute_buffer_alloc_vram(ctx->screen,
shader->bc.ndw * 4);
p = r600_buffer_map_sync_with_rings(&ctx->b, shader->code_bo, PIPE_TRANSFER_WRITE);
memcpy(p, shader->bc.bytecode, shader->bc.ndw * 4);
ctx->b.ws->buffer_unmap(shader->code_bo->buf);
#endif
#endif
shader->ctx = ctx;
shader->local_size = cso->req_local_mem;
shader->private_size = cso->req_private_mem;
shader->input_size = cso->req_input_mem;
return shader;
}
int
main(int argc, char **argv)
{
int n = argc > 1 ? atol(argv[1]) : 24;
LLVMInitializeNativeTarget();
LLVMLinkInInterpreter();
LLVMContextRef Context = LLVMContextCreate();
// Create some module to put our function into it.
LLVMModuleRef M = LLVMModuleCreateWithNameInContext("test", Context);
// We are about to create the "fib" function:
LLVMValueRef FibF = CreateFibFunction(M, Context);
// Now we going to create JIT
LLVMExecutionEngineRef EE;
char * outError;
if (LLVMCreateInterpreterForModule(&EE, M, &outError) != 0) {
printf("%s\n", outError);
return 1;
}
printf("verifying...\n");
if (LLVMVerifyModule(M, LLVMReturnStatusAction, &outError) != 0) {
printf("%s\n", outError);
return 1;
}
printf("OK\n");
printf("We just constructed this LLVM module:\n\n---------\n");
printf("%s\n", LLVMPrintModuleToString(M));
LLVMGenericValueRef Args = LLVMCreateGenericValueOfInt(LLVMInt32TypeInContext(Context), n, 0);
LLVMGenericValueRef Result = LLVMRunFunction(EE, FibF, 1, &Args);
printf("Result: %llu\n", LLVMGenericValueToInt(Result, 0));
return 0;
}
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;
}
void *evergreen_create_compute_state(
struct pipe_context *ctx_,
const const struct pipe_compute_state *cso)
{
struct r600_context *ctx = (struct r600_context *)ctx_;
struct r600_pipe_compute *shader = CALLOC_STRUCT(r600_pipe_compute);
#ifdef HAVE_OPENCL
const struct pipe_llvm_program_header * header;
const unsigned char * code;
unsigned i;
shader->llvm_ctx = LLVMContextCreate();
COMPUTE_DBG(ctx->screen, "*** evergreen_create_compute_state\n");
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
#endif
shader->ctx = (struct r600_context*)ctx;
shader->local_size = cso->req_local_mem;
shader->private_size = cso->req_private_mem;
shader->input_size = cso->req_input_mem;
#ifdef HAVE_OPENCL
shader->num_kernels = radeon_llvm_get_num_kernels(shader->llvm_ctx, code,
header->num_bytes);
shader->kernels = CALLOC(sizeof(struct r600_kernel), shader->num_kernels);
for (i = 0; i < shader->num_kernels; i++) {
struct r600_kernel *kernel = &shader->kernels[i];
kernel->llvm_module = radeon_llvm_get_kernel_module(shader->llvm_ctx, i,
code, header->num_bytes);
}
#endif
return shader;
}
/**
* 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;
}
PIPE_ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
struct lp_type src_type,
struct lp_type dst_type)
{
LLVMContextRef context;
struct gallivm_state *gallivm;
LLVMValueRef func = 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 ((src_type.width >= dst_type.width && src_type.length > dst_type.length) ||
(src_type.width <= dst_type.width && src_type.length < dst_type.length)) {
return TRUE;
}
/* Known failures
* - fixed point 32 -> float 32
* - float 32 -> signed normalised integer 32
*/
if ((src_type.floating && !dst_type.floating && dst_type.sign && dst_type.norm && src_type.width == dst_type.width) ||
(!src_type.floating && dst_type.floating && src_type.fixed && src_type.width == dst_type.width)) {
return TRUE;
}
/* Known failures
* - fixed point 32 -> float 32
* - float 32 -> signed normalised integer 32
*/
if ((src_type.floating && !dst_type.floating && dst_type.sign && dst_type.norm && src_type.width == dst_type.width) ||
(!src_type.floating && dst_type.floating && src_type.fixed && src_type.width == dst_type.width)) {
return TRUE;
}
if(verbose >= 1)
dump_conv_types(stderr, src_type, dst_type);
if (src_type.length > dst_type.length) {
num_srcs = 1;
num_dsts = src_type.length/dst_type.length;
}
else if (src_type.length < dst_type.length) {
num_dsts = 1;
num_srcs = dst_type.length/src_type.length;
}
else {
num_dsts = 1;
num_srcs = 1;
}
/* 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));
if (dst_type.norm && dst_type.sign && src_type.sign && !src_type.floating) {
/*
* This is quite inaccurate due to shift being used.
* I don't think it's possible to hit such conversions with
* llvmpipe though.
*/
eps *= 2;
}
context = LLVMContextCreate();
gallivm = gallivm_create("test_module", context);
func = add_conv_test(gallivm, src_type, num_srcs, dst_type, num_dsts);
gallivm_compile_module(gallivm);
conv_test_ptr = (conv_test_ptr_t)gallivm_jit_function(gallivm, func);
gallivm_free_ir(gallivm);
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(LP_MIN_VECTOR_ALIGN) uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
PIPE_ALIGN_VAR(LP_MIN_VECTOR_ALIGN) 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 || verbose >= 3) {
if(verbose < 1)
dump_conv_types(stderr, src_type, dst_type);
if (success) {
fprintf(stderr, "PASS\n");
}
else {
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);
gallivm_destroy(gallivm);
LLVMContextDispose(context);
return success;
}
/**
* 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();
if (!gallivm_context) {
gallivm_context = LLVMContextCreate();
}
gallivm->context = gallivm_context;
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;
gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
if (!gallivm->builder)
goto fail;
/* FIXME: MC-JIT only allows compiling one module at a time, and it must be
* complete when MC-JIT is created. So defer the MC-JIT engine creation for
* now.
*/
#if !USE_MCJIT
if (!init_gallivm_engine(gallivm)) {
goto fail;
}
#else
/*
* MC-JIT engine compiles the module immediately on creation, so we can't
* obtain the target data from it. Instead we create a target data layout
* from a string.
*
* The produced layout strings are not precisely the same, but should make
* no difference for the kind of optimization passes we run.
*
* For reference this is the layout string on x64:
*
* e-p:64:64:64-S128-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-f128:128:128-n8:16:32:64
*
* See also:
* - http://llvm.org/docs/LangRef.html#datalayout
*/
{
const unsigned pointer_size = 8 * sizeof(void *);
char layout[512];
util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
#ifdef PIPE_ARCH_LITTLE_ENDIAN
'e', // little endian
#else
'E', // big endian
#endif
pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
pointer_size, // aggregate preferred alignment
pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
gallivm->target = LLVMCreateTargetData(layout);
if (!gallivm->target) {
return FALSE;
}
}
#endif
if (!create_pass_manager(gallivm))
goto fail;
return TRUE;
fail:
free_gallivm_state(gallivm);
return FALSE;
}
/*
* machine_init
*
* Initializes the machine context.
*/
machinedef_t *
machine_init (const char *mspec)
{
machine_ctx_t m;
machinedef_t *mach;
char *err;
LLVMTargetRef target;
char *machspec = (char *) mspec;
unsigned long allosize;
if (machspec == 0) {
machspec = LLVM_DEFAULT_TARGET_TRIPLE;
}
allosize = (sizeof(struct machine_ctx_s) +
sizeof(struct machinedef_s) +
strlen(machspec) + 1);
m = malloc(allosize);
if (m == 0) return 0;
memset(m, 0, allosize);
m->triple = ((char *) m) + (sizeof(struct machine_ctx_s) +
sizeof(struct machinedef_s));
memcpy(m->triple, machspec, strlen(machspec));
LLVM_NATIVE_TARGETINFO();
LLVM_NATIVE_TARGET();
LLVM_NATIVE_TARGETMC();
LLVM_NATIVE_ASMPRINTER();
LLVM_NATIVE_ASMPARSER();
err = 0;
target = HelperLookupTarget(machspec, &err);
if (target == 0) {
if (err != 0) free(err);
free(m);
return 0;
}
m->target_machine = LLVMCreateTargetMachine(target, (char *)machspec, "", "",
LLVMCodeGenLevelDefault,
LLVMRelocPIC, LLVMCodeModelDefault);
if (m->target_machine == 0) {
free(m);
return 0;
}
HelperSetAsmVerbosity(m->target_machine, 1);
m->llvmctx = LLVMContextCreate();
if (m->llvmctx == 0) {
LLVMDisposeTargetMachine(m->target_machine);
free(m);
return 0;
}
m->is_macho = (strstr(machspec, "darwin") != 0); // XXX
mach = (machinedef_t *)(m + 1);
mach->machctx = m;
mach->bpaddr = sizeof(int *) * 8;
mach->bpval = sizeof(long) * 8;
mach->bpunit = 8;
mach->charsize_count = 1;
mach->charsizes[0] = 8;
mach->flags = MACH_M_SIGNEXT | MACH_M_LTC_INIT;
mach->max_align = 4;
mach->reg_count = 16;
return mach;
} /* machine_init */
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,
const struct pipe_blend_state *blend,
struct lp_type type)
{
LLVMContextRef context;
struct gallivm_state *gallivm;
LLVMValueRef func = 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;
const unsigned stride = lp_type_width(type)/8;
if(verbose >= 1)
dump_blend_type(stdout, blend, type);
context = LLVMContextCreate();
gallivm = gallivm_create("test_module", context);
func = add_blend_test(gallivm, blend, type);
gallivm_compile_module(gallivm);
blend_test_ptr = (blend_test_ptr_t)gallivm_jit_function(gallivm, func);
gallivm_free_ir(gallivm);
success = TRUE;
{
uint8_t *src, *src1, *dst, *con, *res, *ref;
src = align_malloc(stride, stride);
src1 = align_malloc(stride, stride);
dst = align_malloc(stride, stride);
con = align_malloc(stride, stride);
res = align_malloc(stride, stride);
ref = align_malloc(stride, stride);
for(i = 0; i < n && success; ++i) {
int64_t start_counter = 0;
int64_t end_counter = 0;
random_vec(type, src);
random_vec(type, src1);
random_vec(type, dst);
random_vec(type, con);
{
double fsrc[LP_MAX_VECTOR_LENGTH];
double fsrc1[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, src1, fsrc1);
read_vec(type, dst, fdst);
read_vec(type, con, fcon);
for(j = 0; j < type.length; j += 4)
compute_blend_ref(blend, fsrc + j, fsrc1 + j, fdst + j, fcon + j, fref + j);
write_vec(type, ref, fref);
}
start_counter = rdtsc();
blend_test_ptr(src, src1, 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, type);
fprintf(stderr, "MISMATCH\n");
fprintf(stderr, " Src: ");
dump_vec(stderr, type, src);
fprintf(stderr, "\n");
fprintf(stderr, " Src1: ");
dump_vec(stderr, type, src1);
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");
}
}
align_free(src);
align_free(src1);
align_free(dst);
align_free(con);
align_free(res);
align_free(ref);
}
/*
* 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, type, cycles_avg, success);
gallivm_destroy(gallivm);
LLVMContextDispose(context);
return success;
}
