LLVMModuleRef radeon_llvm_parse_bitcode(LLVMContextRef ctx,
const char * bitcode, unsigned bitcode_len)
{
LLVMMemoryBufferRef buf;
LLVMModuleRef module;
buf = LLVMCreateMemoryBufferWithMemoryRangeCopy((const char*)bitcode,
bitcode_len, "radeon");
LLVMParseBitcodeInContext(ctx, buf, &module, NULL);
LLVMDisposeMemoryBuffer(buf);
return module;
}
static LLVMModuleRef load_module(bool Lazy, bool New) {
LLVMMemoryBufferRef MB;
LLVMModuleRef M;
char *msg = NULL;
if (LLVMCreateMemoryBufferWithSTDIN(&MB, &msg)) {
fprintf(stderr, "Error reading file: %s\n", msg);
exit(1);
}
LLVMBool Ret;
if (New) {
LLVMContextRef C = LLVMGetGlobalContext();
LLVMContextSetDiagnosticHandler(C, diagnosticHandler, NULL);
if (Lazy)
Ret = LLVMGetBitcodeModule2(MB, &M);
else
Ret = LLVMParseBitcode2(MB, &M);
} else {
if (Lazy)
Ret = LLVMGetBitcodeModule(MB, &M, &msg);
else
Ret = LLVMParseBitcode(MB, &M, &msg);
}
if (Ret) {
fprintf(stderr, "Error parsing bitcode: %s\n", msg);
LLVMDisposeMemoryBuffer(MB);
exit(1);
}
if (!Lazy)
LLVMDisposeMemoryBuffer(MB);
return M;
}
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);
}
/**
* Compile an LLVM module to machine code.
*
* @returns 0 for success, 1 for failure
*/
unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
const char *gpu_family, unsigned dump, LLVMTargetMachineRef tm)
{
char cpu[CPU_STRING_LEN];
char fs[FS_STRING_LEN];
char *err;
bool dispose_tm = false;
LLVMContextRef llvm_ctx;
unsigned rval = 0;
LLVMMemoryBufferRef out_buffer;
unsigned buffer_size;
const char *buffer_data;
char triple[TRIPLE_STRING_LEN];
LLVMBool mem_err;
if (!tm) {
strncpy(triple, "r600--", TRIPLE_STRING_LEN);
LLVMTargetRef target = radeon_llvm_get_r600_target(triple);
if (!target) {
return 1;
}
strncpy(cpu, gpu_family, CPU_STRING_LEN);
memset(fs, 0, sizeof(fs));
if (dump) {
strncpy(fs, "+DumpCode", FS_STRING_LEN);
}
tm = LLVMCreateTargetMachine(target, triple, cpu, fs,
LLVMCodeGenLevelDefault, LLVMRelocDefault,
LLVMCodeModelDefault);
dispose_tm = true;
}
if (dump) {
LLVMDumpModule(M);
}
/* Setup Diagnostic Handler*/
llvm_ctx = LLVMGetModuleContext(M);
#if HAVE_LLVM >= 0x0305
LLVMContextSetDiagnosticHandler(llvm_ctx, radeonDiagnosticHandler, &rval);
#endif
rval = 0;
/* Compile IR*/
mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
&out_buffer);
/* Process Errors/Warnings */
if (mem_err) {
fprintf(stderr, "%s: %s", __FUNCTION__, err);
FREE(err);
LLVMDisposeTargetMachine(tm);
return 1;
}
if (0 != rval) {
fprintf(stderr, "%s: Processing Diag Flag\n", __FUNCTION__);
}
/* Extract Shader Code*/
buffer_size = LLVMGetBufferSize(out_buffer);
buffer_data = LLVMGetBufferStart(out_buffer);
radeon_elf_read(buffer_data, buffer_size, binary, dump);
/* Clean up */
LLVMDisposeMemoryBuffer(out_buffer);
if (dispose_tm) {
LLVMDisposeTargetMachine(tm);
}
return rval;
}
/**
* Compile an LLVM module to machine code.
*
* @returns 0 for success, 1 for failure
*/
unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_llvm_binary *binary,
const char * gpu_family, unsigned dump) {
LLVMTargetRef target;
LLVMTargetMachineRef tm;
char cpu[CPU_STRING_LEN];
char fs[FS_STRING_LEN];
char *err;
LLVMMemoryBufferRef out_buffer;
unsigned buffer_size;
const char *buffer_data;
char triple[TRIPLE_STRING_LEN];
char *elf_buffer;
Elf *elf;
Elf_Scn *section = NULL;
size_t section_str_index;
LLVMBool r;
init_r600_target();
target = get_r600_target();
if (!target) {
return 1;
}
strncpy(cpu, gpu_family, CPU_STRING_LEN);
memset(fs, 0, sizeof(fs));
if (dump) {
LLVMDumpModule(M);
strncpy(fs, "+DumpCode", FS_STRING_LEN);
}
strncpy(triple, "r600--", TRIPLE_STRING_LEN);
tm = LLVMCreateTargetMachine(target, triple, cpu, fs,
LLVMCodeGenLevelDefault, LLVMRelocDefault,
LLVMCodeModelDefault);
r = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
&out_buffer);
if (r) {
fprintf(stderr, "%s", err);
FREE(err);
return 1;
}
buffer_size = LLVMGetBufferSize(out_buffer);
buffer_data = LLVMGetBufferStart(out_buffer);
/* One of the libelf implementations
* (http://www.mr511.de/software/english.htm) requires calling
* elf_version() before elf_memory().
*/
elf_version(EV_CURRENT);
elf_buffer = MALLOC(buffer_size);
memcpy(elf_buffer, buffer_data, buffer_size);
elf = elf_memory(elf_buffer, buffer_size);
elf_getshdrstrndx(elf, §ion_str_index);
binary->disassembled = 0;
while ((section = elf_nextscn(elf, section))) {
const char *name;
Elf_Data *section_data = NULL;
GElf_Shdr section_header;
if (gelf_getshdr(section, §ion_header) != §ion_header) {
fprintf(stderr, "Failed to read ELF section header\n");
return 1;
}
name = elf_strptr(elf, section_str_index, section_header.sh_name);
if (!strcmp(name, ".text")) {
section_data = elf_getdata(section, section_data);
binary->code_size = section_data->d_size;
binary->code = MALLOC(binary->code_size * sizeof(unsigned char));
memcpy(binary->code, section_data->d_buf, binary->code_size);
} else if (!strcmp(name, ".AMDGPU.config")) {
section_data = elf_getdata(section, section_data);
binary->config_size = section_data->d_size;
binary->config = MALLOC(binary->config_size * sizeof(unsigned char));
memcpy(binary->config, section_data->d_buf, binary->config_size);
} else if (dump && !strcmp(name, ".AMDGPU.disasm")) {
binary->disassembled = 1;
section_data = elf_getdata(section, section_data);
fprintf(stderr, "\nShader Disassembly:\n\n");
fprintf(stderr, "%.*s\n", (int)section_data->d_size,
(char *)section_data->d_buf);
}
}
LLVMDisposeMemoryBuffer(out_buffer);
LLVMDisposeTargetMachine(tm);
return 0;
}
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;
}
/**
* Compile an LLVM module to machine code.
*
* @returns 0 for success, 1 for failure
*/
unsigned radeon_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
const char *gpu_family,
LLVMTargetMachineRef tm,
struct pipe_debug_callback *debug)
{
struct radeon_llvm_diagnostics diag;
char cpu[CPU_STRING_LEN];
char fs[FS_STRING_LEN];
char *err;
bool dispose_tm = false;
LLVMContextRef llvm_ctx;
LLVMMemoryBufferRef out_buffer;
unsigned buffer_size;
const char *buffer_data;
char triple[TRIPLE_STRING_LEN];
LLVMBool mem_err;
diag.debug = debug;
diag.retval = 0;
if (!tm) {
strncpy(triple, "r600--", TRIPLE_STRING_LEN);
LLVMTargetRef target = radeon_llvm_get_r600_target(triple);
if (!target) {
return 1;
}
strncpy(cpu, gpu_family, CPU_STRING_LEN);
memset(fs, 0, sizeof(fs));
strncpy(fs, "+DumpCode", FS_STRING_LEN);
tm = LLVMCreateTargetMachine(target, triple, cpu, fs,
LLVMCodeGenLevelDefault, LLVMRelocDefault,
LLVMCodeModelDefault);
dispose_tm = true;
}
/* Setup Diagnostic Handler*/
llvm_ctx = LLVMGetModuleContext(M);
LLVMContextSetDiagnosticHandler(llvm_ctx, radeonDiagnosticHandler, &diag);
/* Compile IR*/
mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
&out_buffer);
/* Process Errors/Warnings */
if (mem_err) {
fprintf(stderr, "%s: %s", __FUNCTION__, err);
pipe_debug_message(debug, SHADER_INFO,
"LLVM emit error: %s", err);
FREE(err);
diag.retval = 1;
goto out;
}
/* Extract Shader Code*/
buffer_size = LLVMGetBufferSize(out_buffer);
buffer_data = LLVMGetBufferStart(out_buffer);
radeon_elf_read(buffer_data, buffer_size, binary);
/* Clean up */
LLVMDisposeMemoryBuffer(out_buffer);
out:
if (dispose_tm) {
LLVMDisposeTargetMachine(tm);
}
if (diag.retval != 0)
pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
return diag.retval;
}
