int main()
{
LLVMContextRef context = LLVMGetGlobalContext();
LLVMModuleRef module = LLVMModuleCreateWithName("test-101");
LLVMBuilderRef builder = LLVMCreateBuilder();
// LLVMInt32Type()
// LLVMFunctionType(rtnType, paramType, parmCnt, isVarArg)
// LLVMAddFunction(module, name, functionType)
LLVMTypeRef main = LLVMFunctionType(LLVMInt32Type(), NULL, 0, 0);
LLVMValueRef mainFn = LLVMAddFunction(module, "main", main);
LLVMBasicBlockRef mainBlk = LLVMAppendBasicBlock(mainFn, "entry");
LLVMPositionBuilderAtEnd(builder, mainBlk);
LLVMValueRef str =
LLVMBuildGlobalStringPtr(builder, "Hello World!", "str");
LLVMTypeRef args[1];
args[0] = LLVMPointerType(LLVMInt8Type(), 0);
LLVMTypeRef puts = LLVMFunctionType(LLVMInt32Type(), args, 1, 0);
LLVMValueRef putsFn = LLVMAddFunction(module, "puts", puts);
LLVMBuildCall(builder, putsFn, &str, 1, "");
LLVMBuildRet(builder, LLVMConstInt(LLVMInt32Type(), 0, 0));
LLVMDumpModule(module);
return 0;
}
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);
}
int main (void) {
LLVMModuleRef module = LLVMModuleCreateWithName("kal");
LLVMBuilderRef builder = LLVMCreateBuilder();
// LLVMInitializeNativeTarget();
LLVMTypeRef funcType = LLVMFunctionType(LLVMVoidType(), NULL, 0, 0);
LLVMValueRef func = LLVMAddFunction(module, "main", funcType);
LLVMSetLinkage(func, LLVMExternalLinkage);
LLVMBasicBlockRef block = LLVMAppendBasicBlock(func, "entry");
LLVMPositionBuilderAtEnd(builder, block);
LLVMValueRef cond = LLVMBuildICmp(builder, LLVMIntNE, LLVMConstInt(LLVMInt32Type(), 2, 0), LLVMConstInt(LLVMInt32Type(), 1, 0), "ifcond");
LLVMValueRef owning_block = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)); //TODO: WRONG??
//LLVMValueRef owning_block = LLVMBasicBlockAsValue(LLVMGetPreviousBasicBlock(LLVMGetInsertBlock(builder)));
// 2. Generate new blocks for cases.
LLVMBasicBlockRef then_ref = LLVMAppendBasicBlock(owning_block, "then");
LLVMBasicBlockRef else_ref = LLVMAppendBasicBlock(owning_block, "else");
LLVMBasicBlockRef merge_ref = LLVMAppendBasicBlock(owning_block, "ifmerge");
// 3. Branch conditionally on then or else.
LLVMBuildCondBr(builder, cond, then_ref, else_ref);
// 4. Build then branch prologue.
LLVMPositionBuilderAtEnd(builder, then_ref);
LLVMValueRef hi1 = LLVMBuildXor(builder, LLVMGetUndef(LLVMInt32Type()), LLVMGetUndef(LLVMInt32Type()), "subtmp");
// 5. Connect then branch to merge block.
LLVMBuildBr(builder, merge_ref);
then_ref = LLVMGetInsertBlock(builder);
// 6. Build else branch prologue.
LLVMPositionBuilderAtEnd(builder, else_ref);
LLVMValueRef hi2 = LLVMBuildXor(builder, LLVMGetUndef(LLVMInt32Type()), LLVMGetUndef(LLVMInt32Type()), "subtmp2");
// 7. Connect else branch to merge block.
LLVMBuildBr(builder, merge_ref);
else_ref = LLVMGetInsertBlock(builder);
// 8. Position ourselves after the merge block.
LLVMPositionBuilderAtEnd(builder, merge_ref);
// 9. Build the phi node.
// LLVMValueRef phi = LLVMBuildPhi(builder, LLVMDoubleType(), "phi");
// 10. Add incoming edges.
// LLVMAddIncoming(phi, &hi1, &then_ref, 1);
// LLVMAddIncoming(phi, &hi2, &else_ref, 1);
LLVMDumpModule(module);
LLVMDisposeBuilder(builder);
LLVMDisposeModule(module);
return 0;
}
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;
}
unsigned r600_llvm_compile(
LLVMModuleRef mod,
enum radeon_family family,
struct r600_bytecode *bc,
boolean *use_kill,
unsigned dump,
struct pipe_debug_callback *debug)
{
unsigned r;
struct radeon_shader_binary binary;
const char * gpu_family = r600_get_llvm_processor_name(family);
radeon_shader_binary_init(&binary);
if (dump)
LLVMDumpModule(mod);
r = radeon_llvm_compile(mod, &binary, gpu_family, NULL, debug);
r = r600_create_shader(bc, &binary, use_kill);
radeon_shader_binary_clean(&binary);
return r;
}
void
lp_build_tgsi_aos(struct gallivm_state *gallivm,
const struct tgsi_token *tokens,
struct lp_type type,
const unsigned char swizzles[4],
LLVMValueRef consts_ptr,
const LLVMValueRef *inputs,
LLVMValueRef *outputs,
struct lp_build_sampler_aos *sampler,
const struct tgsi_shader_info *info)
{
struct lp_build_tgsi_aos_context bld;
struct tgsi_parse_context parse;
uint num_immediates = 0;
unsigned chan;
int pc = 0;
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.bld_base.base, gallivm, type);
lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
for (chan = 0; chan < 4; ++chan) {
bld.swizzles[chan] = swizzles[chan];
bld.inv_swizzles[swizzles[chan]] = chan;
}
bld.inputs = inputs;
bld.outputs = outputs;
bld.consts_ptr = consts_ptr;
bld.sampler = sampler;
bld.indirect_files = info->indirect_files;
bld.bld_base.emit_swizzle = swizzle_aos;
bld.bld_base.info = info;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
/* Set opcode actions */
lp_set_default_actions_cpu(&bld.bld_base);
if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
return;
}
tgsi_parse_init(&parse, tokens);
while (!tgsi_parse_end_of_tokens(&parse)) {
tgsi_parse_token(&parse);
switch(parse.FullToken.Token.Type) {
case TGSI_TOKEN_TYPE_DECLARATION:
/* Inputs already interpolated */
lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
break;
case TGSI_TOKEN_TYPE_INSTRUCTION:
/* save expanded instruction */
lp_bld_tgsi_add_instruction(&bld.bld_base,
&parse.FullToken.FullInstruction);
break;
case TGSI_TOKEN_TYPE_IMMEDIATE:
/* simply copy the immediate values into the next immediates[] slot */
{
const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
float imm[4];
assert(size <= 4);
assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
for (chan = 0; chan < 4; ++chan) {
imm[chan] = 0.0f;
}
for (chan = 0; chan < size; ++chan) {
unsigned swizzle = bld.swizzles[chan];
imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
}
bld.immediates[num_immediates] =
lp_build_const_aos(gallivm, type,
imm[0], imm[1], imm[2], imm[3],
NULL);
num_immediates++;
}
break;
case TGSI_TOKEN_TYPE_PROPERTY:
break;
default:
assert(0);
}
}
while (pc != -1) {
struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
const struct tgsi_opcode_info *opcode_info =
tgsi_get_opcode_info(instr->Instruction.Opcode);
if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
_debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
opcode_info->mnemonic);
}
if (0) {
LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
LLVMValueRef function = LLVMGetBasicBlockParent(block);
debug_printf("11111111111111111111111111111 \n");
tgsi_dump(tokens, 0);
lp_debug_dump_value(function);
debug_printf("2222222222222222222222222222 \n");
}
tgsi_parse_free(&parse);
FREE(bld.bld_base.instructions);
if (0) {
LLVMModuleRef module = LLVMGetGlobalParent(
LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
LLVMDumpModule(module);
}
}
/**
* 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;
}
SCM llvm_dump_module(SCM scm_self)
{
struct llvm_module_t *self = get_llvm(scm_self);
LLVMDumpModule(self->module);
return SCM_UNSPECIFIED;
}
static void
lp_jit_create_types(struct lp_fragment_shader_variant *lp)
{
struct gallivm_state *gallivm = lp->gallivm;
LLVMContextRef lc = gallivm->context;
LLVMTypeRef viewport_type, texture_type, sampler_type;
/* struct lp_jit_viewport */
{
LLVMTypeRef elem_types[LP_JIT_VIEWPORT_NUM_FIELDS];
elem_types[LP_JIT_VIEWPORT_MIN_DEPTH] =
elem_types[LP_JIT_VIEWPORT_MAX_DEPTH] = LLVMFloatTypeInContext(lc);
viewport_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_viewport, min_depth,
gallivm->target, viewport_type,
LP_JIT_VIEWPORT_MIN_DEPTH);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_viewport, max_depth,
gallivm->target, viewport_type,
LP_JIT_VIEWPORT_MAX_DEPTH);
LP_CHECK_STRUCT_SIZE(struct lp_jit_viewport,
gallivm->target, viewport_type);
}
/* struct lp_jit_texture */
{
LLVMTypeRef elem_types[LP_JIT_TEXTURE_NUM_FIELDS];
elem_types[LP_JIT_TEXTURE_WIDTH] =
elem_types[LP_JIT_TEXTURE_HEIGHT] =
elem_types[LP_JIT_TEXTURE_DEPTH] =
elem_types[LP_JIT_TEXTURE_FIRST_LEVEL] =
elem_types[LP_JIT_TEXTURE_LAST_LEVEL] = LLVMInt32TypeInContext(lc);
elem_types[LP_JIT_TEXTURE_BASE] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
elem_types[LP_JIT_TEXTURE_ROW_STRIDE] =
elem_types[LP_JIT_TEXTURE_IMG_STRIDE] =
elem_types[LP_JIT_TEXTURE_MIP_OFFSETS] =
LLVMArrayType(LLVMInt32TypeInContext(lc), LP_MAX_TEXTURE_LEVELS);
texture_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, width,
gallivm->target, texture_type,
LP_JIT_TEXTURE_WIDTH);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, height,
gallivm->target, texture_type,
LP_JIT_TEXTURE_HEIGHT);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, depth,
gallivm->target, texture_type,
LP_JIT_TEXTURE_DEPTH);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, first_level,
gallivm->target, texture_type,
LP_JIT_TEXTURE_FIRST_LEVEL);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, last_level,
gallivm->target, texture_type,
LP_JIT_TEXTURE_LAST_LEVEL);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, base,
gallivm->target, texture_type,
LP_JIT_TEXTURE_BASE);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, row_stride,
gallivm->target, texture_type,
LP_JIT_TEXTURE_ROW_STRIDE);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, img_stride,
gallivm->target, texture_type,
LP_JIT_TEXTURE_IMG_STRIDE);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, mip_offsets,
gallivm->target, texture_type,
LP_JIT_TEXTURE_MIP_OFFSETS);
LP_CHECK_STRUCT_SIZE(struct lp_jit_texture,
gallivm->target, texture_type);
}
/* struct lp_jit_sampler */
{
LLVMTypeRef elem_types[LP_JIT_SAMPLER_NUM_FIELDS];
elem_types[LP_JIT_SAMPLER_MIN_LOD] =
elem_types[LP_JIT_SAMPLER_MAX_LOD] =
elem_types[LP_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(lc);
elem_types[LP_JIT_SAMPLER_BORDER_COLOR] =
LLVMArrayType(LLVMFloatTypeInContext(lc), 4);
sampler_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, min_lod,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_MIN_LOD);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, max_lod,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_MAX_LOD);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, lod_bias,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_LOD_BIAS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, border_color,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_BORDER_COLOR);
LP_CHECK_STRUCT_SIZE(struct lp_jit_sampler,
gallivm->target, sampler_type);
}
/* struct lp_jit_context */
{
LLVMTypeRef elem_types[LP_JIT_CTX_COUNT];
LLVMTypeRef context_type;
elem_types[LP_JIT_CTX_CONSTANTS] =
LLVMArrayType(LLVMPointerType(LLVMFloatTypeInContext(lc), 0), LP_MAX_TGSI_CONST_BUFFERS);
elem_types[LP_JIT_CTX_NUM_CONSTANTS] =
LLVMArrayType(LLVMInt32TypeInContext(lc), LP_MAX_TGSI_CONST_BUFFERS);
elem_types[LP_JIT_CTX_ALPHA_REF] = LLVMFloatTypeInContext(lc);
elem_types[LP_JIT_CTX_STENCIL_REF_FRONT] =
elem_types[LP_JIT_CTX_STENCIL_REF_BACK] = LLVMInt32TypeInContext(lc);
elem_types[LP_JIT_CTX_U8_BLEND_COLOR] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
elem_types[LP_JIT_CTX_F_BLEND_COLOR] = LLVMPointerType(LLVMFloatTypeInContext(lc), 0);
elem_types[LP_JIT_CTX_VIEWPORTS] = LLVMPointerType(viewport_type, 0);
elem_types[LP_JIT_CTX_TEXTURES] = LLVMArrayType(texture_type,
PIPE_MAX_SHADER_SAMPLER_VIEWS);
elem_types[LP_JIT_CTX_SAMPLERS] = LLVMArrayType(sampler_type,
PIPE_MAX_SAMPLERS);
context_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, constants,
gallivm->target, context_type,
LP_JIT_CTX_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, num_constants,
gallivm->target, context_type,
LP_JIT_CTX_NUM_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, alpha_ref_value,
gallivm->target, context_type,
LP_JIT_CTX_ALPHA_REF);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, stencil_ref_front,
gallivm->target, context_type,
LP_JIT_CTX_STENCIL_REF_FRONT);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, stencil_ref_back,
gallivm->target, context_type,
LP_JIT_CTX_STENCIL_REF_BACK);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, u8_blend_color,
gallivm->target, context_type,
LP_JIT_CTX_U8_BLEND_COLOR);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, f_blend_color,
gallivm->target, context_type,
LP_JIT_CTX_F_BLEND_COLOR);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, viewports,
gallivm->target, context_type,
LP_JIT_CTX_VIEWPORTS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, textures,
gallivm->target, context_type,
LP_JIT_CTX_TEXTURES);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, samplers,
gallivm->target, context_type,
LP_JIT_CTX_SAMPLERS);
LP_CHECK_STRUCT_SIZE(struct lp_jit_context,
gallivm->target, context_type);
lp->jit_context_ptr_type = LLVMPointerType(context_type, 0);
}
/* struct lp_jit_thread_data */
{
LLVMTypeRef elem_types[LP_JIT_THREAD_DATA_COUNT];
LLVMTypeRef thread_data_type;
elem_types[LP_JIT_THREAD_DATA_COUNTER] = LLVMInt64TypeInContext(lc);
elem_types[LP_JIT_THREAD_DATA_RASTER_STATE_VIEWPORT_INDEX] =
LLVMInt32TypeInContext(lc);
thread_data_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
lp->jit_thread_data_ptr_type = LLVMPointerType(thread_data_type, 0);
}
if (gallivm_debug & GALLIVM_DEBUG_IR) {
LLVMDumpModule(gallivm->module);
}
}
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;
}
/**
* 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;
}
int main(int argc, char** argv) {
llvm::llvm_shutdown_obj Y;
llvm::error_code ec;
std::string se;
std::string file;
if (argc < 2) {
llvm::errs() << "Usage llvm-disasm <file>\n";
return 1;
}
llvm::InitializeAllTargetInfos();
llvm::InitializeAllTargetMCs();
llvm::InitializeAllAsmParsers();
llvm::InitializeAllDisassemblers();
file = argv[1];
llvm::OwningPtr<llvm::object::Binary> bin;
ec = llvm::object::createBinary(file, bin);
if (ec) {
llvm::errs() << file << ": " << ec.message() << "\n";
return 1;
}
if (!bin->isELF()) {
llvm::errs() << file << " isn't an object file\n";
return 1;
}
llvm::object::ObjectFile* obj = llvm::dyn_cast<llvm::object::ObjectFile>(bin.get());
if (!obj) {
llvm::errs() << file << ": failed to cast to llvm::ObjectFile\n";
return 1;
}
llvm::Triple tri;
tri.setArch(llvm::Triple::ArchType(obj->getArch()));
std::string targetName = tri.str();
const llvm::Target* target = llvm::TargetRegistry::lookupTarget(targetName, se);
if (!target) {
llvm::errs() << file << ": failed to get the target descriptor for " << targetName << "\n";
return 1;
}
STI = target->createMCSubtargetInfo(targetName, "", "");
if (!STI) {
llvm::errs() << file << ": " << ": to get the subtarget info!\n";
return 1;
}
disasm = target->createMCDisassembler(*STI);
if (!disasm) {
llvm::errs() << file << ": " << ": to get the disassembler!\n";
return 1;
}
MII = target->createMCInstrInfo();
if (!MII) {
llvm::errs() << file << ": no instruction info for target\n";
return 1;
}
MRI = target->createMCRegInfo(targetName);
if (!MRI) {
llvm::errs() << file << ": no register info for target\n";
return 1;
}
llvmBuilder = LLVMCreateBuilder();
LLVMModuleRef llvmModule = LLVMModuleCreateWithName("test");
LLVMTypeRef mainType = LLVMFunctionType(LLVMInt32Type(), NULL, 0, 0);
LLVMValueRef mainFn = LLVMAddFunction(llvmModule, "main", mainType);
LLVMBasicBlockRef blk = LLVMAppendBasicBlock(mainFn, "");
LLVMPositionBuilderAtEnd(llvmBuilder, blk);
for (llvm::object::section_iterator i = obj->begin_sections(), e = obj->end_sections();
i != e;
i.increment(ec)) {
if (ec) {
llvm::errs() << "Failed to increment the section iterator!\n";
return 1;
}
bool isText;
llvm::StringRef secName;
if (i->getName(secName)) {
llvm::errs() << file << ": failed to get the section name\n";
break;
}
if (i->isText(isText)) {
llvm::errs() << file << ": " << secName << ": failed to determine the section type\n";
break;
}
if (!isText) {
continue;
}
std::set<llvm::object::SymbolRef> symbols;
for (llvm::object::symbol_iterator isym = obj->begin_symbols();
isym != obj->end_symbols();
isym.increment(ec)) {
bool res;
llvm::StringRef symName;
llvm::object::SymbolRef::Type symType;
if (ec) {
llvm::errs() << "Failed to increment the symbol iterator!\n";
return 1;
}
if (isym->getName(symName)) {
llvm::errs() << file << ": " << secName << ": failed to get the symbol name!\n";
return 1;
}
/*
uint64_t secSize, secBase, symAddr;
i->getAddress(secBase);
i->getSize(secSize);
isym->getAddress(symAddr);
if (i->containsSymbol(*isym, res)) {
llvm::errs() << file << ": " << secName << ": " << symName << ": failed to check whether the symbol is in the section!\n";
return 1;
}
if (!res) {
continue;
}
if (symAddr < secBase || symAddr >= secBase + secSize) {
continue;
}
*/
llvm::object::section_iterator i2 = llvm::object::section_iterator(llvm::object::SectionRef());
isym->getSection(i2);
if (i2 != i) {
continue;
}
if (isym->getType(symType)) {
llvm::errs() << file << ": " << secName << ": " << symName << ": failed to get the symbol type!\n";
return 1;
}
if (symType != llvm::object::SymbolRef::ST_Function) {
continue;
}
symbols.insert(*isym);
}
for (std::set<llvm::object::SymbolRef>::const_iterator isym = symbols.begin();
isym != symbols.end();
++isym) {
if (analyzeSymbol(*isym)) {
return 1;
}
}
}
LLVMDumpModule(llvmModule);
LLVMDisposeModule(llvmModule);
LLVMDisposeBuilder(llvmBuilder);
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;
}