/**
* Generate color blending and color output.
* \param rt the render target index (to index blend, colormask state)
* \param type the pixel color type
* \param context_ptr pointer to the runtime JIT context
* \param mask execution mask (active fragment/pixel mask)
* \param src colors from the fragment shader
* \param dst_ptr the destination color buffer pointer
*/
static void
generate_blend(const struct pipe_blend_state *blend,
unsigned rt,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
LLVMValueRef mask,
LLVMValueRef *src,
LLVMValueRef dst_ptr)
{
struct lp_build_context bld;
struct lp_build_flow_context *flow;
struct lp_build_mask_context mask_ctx;
LLVMTypeRef vec_type;
LLVMValueRef const_ptr;
LLVMValueRef con[4];
LLVMValueRef dst[4];
LLVMValueRef res[4];
unsigned chan;
lp_build_context_init(&bld, builder, type);
flow = lp_build_flow_create(builder);
/* we'll use this mask context to skip blending if all pixels are dead */
lp_build_mask_begin(&mask_ctx, flow, type, mask);
vec_type = lp_build_vec_type(type);
const_ptr = lp_jit_context_blend_color(builder, context_ptr);
const_ptr = LLVMBuildBitCast(builder, const_ptr,
LLVMPointerType(vec_type, 0), "");
/* load constant blend color and colors from the dest color buffer */
for(chan = 0; chan < 4; ++chan) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
lp_build_name(con[chan], "con.%c", "rgba"[chan]);
lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
}
/* do blend */
lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
/* store results to color buffer */
for(chan = 0; chan < 4; ++chan) {
if(blend->rt[rt].colormask & (1 << chan)) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
lp_build_name(res[chan], "res.%c", "rgba"[chan]);
res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
}
}
lp_build_mask_end(&mask_ctx);
lp_build_flow_destroy(flow);
}
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;
}
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);
}
static void translateBlock(std::deque<llvm::MCInst>& block) {
typedef std::map<std::string, LLVMValueRef>::const_iterator ValRef;
std::map<std::string, LLVMValueRef> locals;
std::map<std::string, LLVMValueRef> regs;
for (InstIter it = block.begin(); it != block.end(); ++it) {
llvm::MCInst& inst = *it;
const llvm::MCInstrDesc& id = MII->get(inst.getOpcode());
llvm::StringRef iname = MII->getName(inst.getOpcode());
if (iname.startswith("MOV")) {
LLVMValueRef lhs;
unsigned iop = 0;
if (id.OpInfo[0].OperandType == llvm::MCOI::OPERAND_MEMORY) {
std::string localName = getLocalName(inst, 0);
ValRef pval = locals.find(localName);
if (pval == locals.end()) {
lhs = LLVMBuildAlloca(llvmBuilder, LLVMInt32Type(), localName.c_str());
locals[localName] = lhs;
} else {
lhs = pval->second;
}
if (id.OpInfo[5].OperandType == llvm::MCOI::OPERAND_IMMEDIATE) {
const llvm::MCOperand& op = inst.getOperand(5);
LLVMBuildStore(llvmBuilder, lhs, LLVMConstInt(LLVMInt32Type(), op.getImm(), 0));
}
if (id.OpInfo[5].OperandType == llvm::MCOI::OPERAND_REGISTER) {
LLVMBuildStore(llvmBuilder, lhs, regs[getRegName(inst, 5)]);
}
} else if (id.OpInfo[0].OperandType == llvm::MCOI::OPERAND_REGISTER) {
LLVMValueRef rhs;
printInst(inst);
if (id.OpInfo[1].OperandType == llvm::MCOI::OPERAND_IMMEDIATE) {
rhs = LLVMConstInt(LLVMInt32Type(), inst.getOperand(1).getImm(), 0);
} else if (id.OpInfo[1].OperandType == llvm::MCOI::OPERAND_MEMORY) {
ValRef pval = locals.find(getLocalName(inst, 1));
if (pval == locals.end()) {
llvm::outs() << "No such local " << getLocalName(inst, 1) << "\n";
break;
}
rhs = LLVMBuildLoad(llvmBuilder, pval->second, getRegName(inst, 0));
} else {
continue;
}
regs[getRegName(inst, 0)] = rhs;
}
}
}
}
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;
}
/**
* Generate color blending and color output.
*/
static void
generate_blend(const struct pipe_blend_state *blend,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
LLVMValueRef mask,
LLVMValueRef *src,
LLVMValueRef dst_ptr)
{
struct lp_build_context bld;
struct lp_build_flow_context *flow;
struct lp_build_mask_context mask_ctx;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
LLVMValueRef const_ptr;
LLVMValueRef con[4];
LLVMValueRef dst[4];
LLVMValueRef res[4];
unsigned chan;
lp_build_context_init(&bld, builder, type);
flow = lp_build_flow_create(builder);
lp_build_mask_begin(&mask_ctx, flow, type, mask);
vec_type = lp_build_vec_type(type);
int_vec_type = lp_build_int_vec_type(type);
const_ptr = lp_jit_context_blend_color(builder, context_ptr);
const_ptr = LLVMBuildBitCast(builder, const_ptr,
LLVMPointerType(vec_type, 0), "");
for(chan = 0; chan < 4; ++chan) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
lp_build_name(con[chan], "con.%c", "rgba"[chan]);
lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
}
lp_build_blend_soa(builder, blend, type, src, dst, con, res);
for(chan = 0; chan < 4; ++chan) {
if(blend->colormask & (1 << chan)) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
lp_build_name(res[chan], "res.%c", "rgba"[chan]);
res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
}
}
lp_build_mask_end(&mask_ctx);
lp_build_flow_destroy(flow);
}
static LLVMValueRef
translateArrayExpr(SymbolTable *TyTable, SymbolTable *ValTable, ASTNode *Node) {
PtrVector *V = &(Node->Child);
ASTNode *SizeNode = (ASTNode*) ptrVectorGet(V, 0),
*InitNode = (ASTNode*) ptrVectorGet(V, 1);
Type *ThisType = createType(IdTy, Node->Value);
LLVMTypeRef ArrayType = getLLVMTypeFromType(TyTable, ThisType);
LLVMValueRef SizeVal = translateExpr(TyTable, ValTable, SizeNode),
InitVal = translateExpr(TyTable, ValTable, InitNode);
LLVMValueRef ArrayVal = LLVMBuildArrayMalloc(Builder, ArrayType, SizeVal, "");
// This BasicBlock and ThisFunction
LLVMBasicBlockRef ThisBB = LLVMGetInsertBlock(Builder);
LLVMValueRef ThisFn = LLVMGetBasicBlockParent(ThisBB);
LLVMValueRef Counter = LLVMBuildAlloca(Builder, LLVMInt32Type(), "");
LLVMBuildStore(Builder, LLVMConstInt(LLVMInt32Type(), 0, 1), Counter);
LLVMTargetDataRef DataRef = LLVMCreateTargetData(LLVMGetDataLayout(Module));
unsigned long long Size = LLVMStoreSizeOfType(DataRef, ArrayType);
LLVMBasicBlockRef InitBB, MidBB, EndBB;
InitBB = LLVMAppendBasicBlock(ThisFn, "for.init");
EndBB = LLVMAppendBasicBlock(ThisFn, "for.end");
MidBB = LLVMAppendBasicBlock(ThisFn, "for.mid");
LLVMBuildBr(Builder, InitBB);
LLVMPositionBuilderAtEnd(Builder, InitBB);
LLVMValueRef CurrentCounter = LLVMBuildLoad(Builder, Counter, "");
LLVMValueRef Comparation = LLVMBuildICmp(Builder, LLVMIntSLT, CurrentCounter, SizeVal, "");
LLVMBuildCondBr(Builder, Comparation, MidBB, EndBB);
LLVMPositionBuilderAtEnd(Builder, MidBB);
CurrentCounter = LLVMBuildLoad(Builder, Counter, "");
LLVMValueRef TheValue = LLVMBuildLoad(Builder, InitVal, "");
LLVMValueRef ElemIdx[] = { LLVMConstInt(LLVMInt32Type(), 0, 1), CurrentCounter };
LLVMValueRef Elem = LLVMBuildInBoundsGEP(Builder, ArrayVal, ElemIdx, 2, "");
copyMemory(Elem, TheValue, getSConstInt(Size));
LLVMBuildBr(Builder, InitBB);
LLVMPositionBuilderAtEnd(Builder, EndBB);
return ArrayVal;
}
/*
* Derive from the quad's upper left scalar coordinates the coordinates for
* all other quad pixels
*/
static void
generate_pos0(LLVMBuilderRef builder,
LLVMValueRef x,
LLVMValueRef y,
LLVMValueRef *x0,
LLVMValueRef *y0)
{
LLVMTypeRef int_elem_type = LLVMInt32Type();
LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
LLVMTypeRef elem_type = LLVMFloatType();
LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
LLVMValueRef x_offsets[QUAD_SIZE];
LLVMValueRef y_offsets[QUAD_SIZE];
unsigned i;
x = lp_build_broadcast(builder, int_vec_type, x);
y = lp_build_broadcast(builder, int_vec_type, y);
for(i = 0; i < QUAD_SIZE; ++i) {
x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
}
x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
*x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
*y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
}
void expand_vectors(struct list_t *elem_list)
{
int index;
int vec_index;
struct cl2llvm_val_t *cl2llvm_index;
struct cl2llvm_val_t *current_vec_elem;
struct cl2llvm_val_t *current_elem;
LIST_FOR_EACH(elem_list, index)
{
current_elem = list_get(elem_list, index);
if (LLVMGetTypeKind(cl2llvmTypeWrapGetLlvmType(current_elem->type)) == LLVMVectorTypeKind)
{
for(vec_index = 0; vec_index < LLVMGetVectorSize(cl2llvmTypeWrapGetLlvmType(current_elem->type)); vec_index++)
{
cl2llvm_index = cl2llvm_val_create_w_init( LLVMConstInt(
LLVMInt32Type(), vec_index, 0), 1);
snprintf(temp_var_name, sizeof(temp_var_name),
"tmp_%d", temp_var_count++);
current_vec_elem = cl2llvm_val_create_w_init( LLVMBuildExtractElement(cl2llvm_builder, current_elem->val, cl2llvm_index->val, temp_var_name), cl2llvmTypeWrapGetSign(current_elem->type));
list_insert(elem_list, index + vec_index, current_vec_elem);
cl2llvm_val_free(cl2llvm_index);
}
cl2llvm_val_free(current_elem);
list_remove(elem_list, current_elem);
}
}
static LLVMTypeRef llvm_type(int type)
{
switch (type) {
case SCM_FOREIGN_TYPE_FLOAT:
return LLVMFloatType();
case SCM_FOREIGN_TYPE_DOUBLE:
return LLVMDoubleType();
case SCM_FOREIGN_TYPE_BOOL:
return LLVMInt1Type();
case SCM_FOREIGN_TYPE_UINT8:
case SCM_FOREIGN_TYPE_INT8:
return LLVMInt8Type();
case SCM_FOREIGN_TYPE_UINT16:
case SCM_FOREIGN_TYPE_INT16:
return LLVMInt16Type();
case SCM_FOREIGN_TYPE_UINT32:
case SCM_FOREIGN_TYPE_INT32:
return LLVMInt32Type();
case SCM_FOREIGN_TYPE_UINT64:
case SCM_FOREIGN_TYPE_INT64:
return LLVMInt64Type();
default:
return LLVMVoidType();
};
}
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 INLINE LLVMValueRef
lp_build_round_sse41(struct lp_build_context *bld,
LLVMValueRef a,
enum lp_build_round_sse41_mode mode)
{
const struct lp_type type = bld->type;
LLVMTypeRef vec_type = lp_build_vec_type(type);
const char *intrinsic;
assert(type.floating);
assert(type.width*type.length == 128);
assert(lp_check_value(type, a));
assert(util_cpu_caps.has_sse4_1);
switch(type.width) {
case 32:
intrinsic = "llvm.x86.sse41.round.ps";
break;
case 64:
intrinsic = "llvm.x86.sse41.round.pd";
break;
default:
assert(0);
return bld->undef;
}
return lp_build_intrinsic_binary(bld->builder, intrinsic, vec_type, a,
LLVMConstInt(LLVMInt32Type(), mode, 0));
}
static LLVMValueRef
translateBinOp(SymbolTable *TyTable, SymbolTable *ValTable, ASTNode *Node) {
ASTNode *NodeE1 = (ASTNode*) ptrVectorGet(&(Node->Child), 0),
*NodeE2 = (ASTNode*) ptrVectorGet(&(Node->Child), 1);
LLVMValueRef ValueE1Ptr = translateExpr(TyTable, ValTable, NodeE1),
ValueE2Ptr = translateExpr(TyTable, ValTable, NodeE2);
LLVMValueRef ValueE1 = LLVMBuildLoad(Builder, ValueE1Ptr, "binop.ld.e1."),
ValueE2 = LLVMBuildLoad(Builder, ValueE2Ptr, "binop.ld.e2.");
LLVMValueRef ResultVal = NULL;
switch (LLVMGetTypeKind(LLVMTypeOf(ValueE1))) {
case LLVMIntegerTypeKind: ResultVal = translateIntBinOp (Node->Kind, ValueE1, ValueE2); break;
case LLVMFloatTypeKind: ResultVal = translateFloatBinOp (Node->Kind, ValueE1, ValueE2); break;
case LLVMStructTypeKind: ResultVal = translateStructBinOp(Node->Kind, ValueE1Ptr, ValueE2Ptr); break;
case LLVMPointerTypeKind: ResultVal = translateStringBinOp(Node->Kind, ValueE1, ValueE2); break;
default: return NULL;
}
switch (LLVMGetTypeKind(LLVMTypeOf(ResultVal))) {
case LLVMIntegerTypeKind: ResultVal = LLVMBuildZExt(Builder, ResultVal, LLVMInt32Type(), ""); break;
default: break;
}
return wrapValue(ResultVal);
}
static LLVMValueRef
add_conv_test(LLVMModuleRef module,
struct lp_type src_type, unsigned num_srcs,
struct lp_type dst_type, unsigned num_dsts)
{
LLVMTypeRef args[2];
LLVMValueRef func;
LLVMValueRef src_ptr;
LLVMValueRef dst_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef src[LP_MAX_VECTOR_LENGTH];
LLVMValueRef dst[LP_MAX_VECTOR_LENGTH];
unsigned i;
args[0] = LLVMPointerType(lp_build_vec_type(src_type), 0);
args[1] = LLVMPointerType(lp_build_vec_type(dst_type), 0);
func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidType(), args, 2, 0));
LLVMSetFunctionCallConv(func, LLVMCCallConv);
src_ptr = LLVMGetParam(func, 0);
dst_ptr = LLVMGetParam(func, 1);
block = LLVMAppendBasicBlock(func, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
for(i = 0; i < num_srcs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef ptr = LLVMBuildGEP(builder, src_ptr, &index, 1, "");
src[i] = LLVMBuildLoad(builder, ptr, "");
}
lp_build_conv(builder, src_type, dst_type, src, num_srcs, dst, num_dsts);
for(i = 0; i < num_dsts; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef ptr = LLVMBuildGEP(builder, dst_ptr, &index, 1, "");
LLVMBuildStore(builder, dst[i], ptr);
}
LLVMBuildRetVoid(builder);;
LLVMDisposeBuilder(builder);
return func;
}
int llvm_add_named_metadata_operand(void) {
LLVMModuleRef m = LLVMModuleCreateWithName("Mod");
LLVMValueRef values[] = { LLVMConstInt(LLVMInt32Type(), 0, 0) };
// This used to trigger an assertion
LLVMAddNamedMetadataOperand(m, "name", LLVMMDNode(values, 1));
LLVMDisposeModule(m);
return 0;
}
int llvm_set_metadata(void) {
LLVMBuilderRef b = LLVMCreateBuilder();
LLVMValueRef values[] = { LLVMConstInt(LLVMInt32Type(), 0, 0) };
// This used to trigger an assertion
LLVMSetMetadata(
LLVMBuildRetVoid(b),
LLVMGetMDKindID("kind", 4),
LLVMMDNode(values, 1));
LLVMDisposeBuilder(b);
return 0;
}
static void
emit_declaration(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_declaration *decl)
{
LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
unsigned first = decl->Range.First;
unsigned last = decl->Range.Last;
unsigned idx;
for (idx = first; idx <= last; ++idx) {
switch (decl->Declaration.File) {
case TGSI_FILE_TEMPORARY:
assert(idx < LP_MAX_TGSI_TEMPS);
if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
last + 1, 0);
bld->temps_array = lp_build_array_alloca(bld->base.builder,
vec_type, array_size, "");
} else {
bld->temps[idx] = lp_build_alloca(bld->base.builder,
vec_type, "");
}
break;
case TGSI_FILE_OUTPUT:
bld->outputs[idx] = lp_build_alloca(bld->base.builder,
vec_type, "");
break;
case TGSI_FILE_ADDRESS:
assert(idx < LP_MAX_TGSI_ADDRS);
bld->addr[idx] = lp_build_alloca(bld->base.builder,
vec_type, "");
break;
case TGSI_FILE_PREDICATE:
assert(idx < LP_MAX_TGSI_PREDS);
bld->preds[idx] = lp_build_alloca(bld->base.builder,
vec_type, "");
break;
default:
/* don't need to declare other vars */
break;
}
}
}
static LLVMTypeRef
get_llvm_type(int type_specifier)
{
switch (type_specifier) {
case TYPE_INT:
return LLVMInt32Type();
case TYPE_FLOAT:
return LLVMFloatType();
default:
fprintf(stderr, "Unknown type specifier: %d\n", type_specifier);
exit(EXIT_FAILURE);
}
return NULL;
}
static LLVMValueRef
translateStringBinOp(NodeKind Op, LLVMValueRef ValueE1, LLVMValueRef ValueE2) {
LLVMValueRef StrCmpFn = LLVMGetNamedFunction(Module, "strcmp");
LLVMValueRef CmpArgs[] = { ValueE1, ValueE2 },
CallStrCmp = LLVMBuildCall(Builder, StrCmpFn, CmpArgs, 2, ""),
ZeroConst = LLVMConstInt(LLVMInt32Type(), 0, 1);
switch (Op) {
case LtOp: return LLVMBuildICmp(Builder, LLVMIntSLT, CallStrCmp, ZeroConst, "");
case LeOp: return LLVMBuildICmp(Builder, LLVMIntSLE, CallStrCmp, ZeroConst, "");
case GtOp: return LLVMBuildICmp(Builder, LLVMIntSGT, CallStrCmp, ZeroConst, "");
case GeOp: return LLVMBuildICmp(Builder, LLVMIntSGE, CallStrCmp, ZeroConst, "");
case EqOp: return LLVMBuildICmp(Builder, LLVMIntEQ, CallStrCmp, ZeroConst, "");
case DiffOp: return LLVMBuildICmp(Builder, LLVMIntNE, CallStrCmp, ZeroConst, "");
default: return NULL;
}
}
struct cl2llvm_val_t *cl2llvm_val_create(void)
{
int i;
struct cl2llvm_val_t *cl2llvm_val;
cl2llvm_val = xcalloc(1, sizeof(struct cl2llvm_val_t));
struct cl2llvmTypeWrap *cl2llvm_type;
cl2llvm_type = cl2llvmTypeWrapCreate(LLVMInt32Type(), 1);
cl2llvm_val->type = cl2llvm_type;
cl2llvm_val->vector_indices = xmalloc(sizeof(struct cl2llvm_val_t *) * 17);
/* Initialize vector indices to NULL */
for (i = 0; i < 17; i++)
cl2llvm_val->vector_indices[i] = NULL;
return cl2llvm_val;
}
static inline LLVMValueRef LLVM_visit(ASTNode *node, LLVMBuilderRef builder) {
switch(node->type) {
case AST_BINARY_OP: {
ASTBinaryOp *binary_op = (ASTBinaryOp*)node;
LLVMValueRef a = LLVM_visit(binary_op->lhs, builder);
LLVMValueRef b = LLVM_visit(binary_op->rhs, builder);
switch(binary_op->op) {
case '+': return LLVMBuildAdd(builder, a, b, "a + b");
case '-': return LLVMBuildSub(builder, a, b, "a - b");
case '*': return LLVMBuildMul(builder, a, b, "a * b");
case '/': return LLVMBuildSDiv(builder, a, b, "a / b");
}
}
case AST_INT: {
return LLVMConstInt(LLVMInt32Type(), ((ASTInt*)node)->value, 0);
}
}
}
/**
* lp_build_assert.
*
* Build an assertion in LLVM IR by building a function call to the
* lp_assert() function above.
*
* \param condition should be an 'i1' or 'i32' value
* \param msg a string to print if the assertion fails.
*/
LLVMValueRef
lp_build_assert(LLVMBuilderRef builder, LLVMValueRef condition,
const char *msg)
{
LLVMModuleRef module;
LLVMTypeRef arg_types[2];
LLVMValueRef msg_string, assert_func, params[2], r;
module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(
LLVMGetInsertBlock(builder)));
msg_string = lp_build_const_string_variable(module, msg, strlen(msg) + 1);
arg_types[0] = LLVMInt32Type();
arg_types[1] = LLVMPointerType(LLVMInt8Type(), 0);
/* lookup the lp_assert function */
assert_func = LLVMGetNamedFunction(module, "lp_assert");
/* Create the assertion function if not found */
if (!assert_func) {
LLVMTypeRef func_type =
LLVMFunctionType(LLVMVoidType(), arg_types, 2, 0);
assert_func = LLVMAddFunction(module, "lp_assert", func_type);
LLVMSetFunctionCallConv(assert_func, LLVMCCallConv);
LLVMSetLinkage(assert_func, LLVMExternalLinkage);
LLVMAddGlobalMapping(lp_build_engine, assert_func,
func_to_pointer((func_pointer)lp_assert));
}
assert(assert_func);
/* build function call param list */
params[0] = LLVMBuildZExt(builder, condition, arg_types[0], "");
params[1] = LLVMBuildBitCast(builder, msg_string, arg_types[1], "");
/* check arg types */
assert(LLVMTypeOf(params[0]) == arg_types[0]);
assert(LLVMTypeOf(params[1]) == arg_types[1]);
r = LLVMBuildCall(builder, assert_func, params, 2, "");
return r;
}
static void analyzeStackReferences(std::deque<llvm::MCInst>& block) {
for (InstIter it = block.begin(); it != block.end(); ++it) {
llvm::MCInst& inst = *it;
const llvm::MCInstrDesc& id = MII->get(inst.getOpcode());
llvm::StringRef iname = MII->getName(inst.getOpcode());
for (unsigned iop = 0; iop < inst.getNumOperands(); ++iop) {
if (id.OpInfo[iop].OperandType == llvm::MCOI::OPERAND_MEMORY) {
const llvm::MCOperand& op = inst.getOperand(iop);
if (op.isReg() && strcmp(MRI->getName(op.getReg()), "RBP") == 0) {
LLVMBuildAlloca(llvmBuilder, LLVMInt32Type(), "");
//printInst(inst);
}
iop += 5;
break;
}
}
}
}
/**
* Gather elements from scatter positions in memory into a single vector.
*
* @param src_width src element width
* @param dst_width result element width (source will be expanded to fit)
* @param length length of the offsets,
* @param base_ptr base pointer, should be a i8 pointer type.
* @param offsets vector with offsets
*/
LLVMValueRef
lp_build_gather(LLVMBuilderRef builder,
unsigned length,
unsigned src_width,
unsigned dst_width,
LLVMValueRef base_ptr,
LLVMValueRef offsets)
{
LLVMTypeRef src_type = LLVMIntType(src_width);
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
LLVMTypeRef dst_elem_type = LLVMIntType(dst_width);
LLVMTypeRef dst_vec_type = LLVMVectorType(dst_elem_type, length);
LLVMValueRef res;
unsigned i;
res = LLVMGetUndef(dst_vec_type);
for(i = 0; i < length; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef elem_offset;
LLVMValueRef elem_ptr;
LLVMValueRef elem;
elem_offset = LLVMBuildExtractElement(builder, offsets, index, "");
elem_ptr = LLVMBuildGEP(builder, base_ptr, &elem_offset, 1, "");
elem_ptr = LLVMBuildBitCast(builder, elem_ptr, src_ptr_type, "");
elem = LLVMBuildLoad(builder, elem_ptr, "");
assert(src_width <= dst_width);
if(src_width > dst_width)
elem = LLVMBuildTrunc(builder, elem, dst_elem_type, "");
if(src_width < dst_width)
elem = LLVMBuildZExt(builder, elem, dst_elem_type, "");
res = LLVMBuildInsertElement(builder, res, elem, index, "");
}
return res;
}
/**
* Perform the occlusion test and increase the counter.
* Test the depth mask. Add the number of channel which has none zero mask
* into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
* The counter will add 4.
*
* \param type holds element type of the mask vector.
* \param maskvalue is the depth test mask.
* \param counter is a pointer of the uint32 counter.
*/
static void
lp_build_occlusion_count(LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef maskvalue,
LLVMValueRef counter)
{
LLVMValueRef countmask = lp_build_const_int_vec(type, 1);
LLVMValueRef countv = LLVMBuildAnd(builder, maskvalue, countmask, "countv");
LLVMTypeRef i8v16 = LLVMVectorType(LLVMInt8Type(), 16);
LLVMValueRef counti = LLVMBuildBitCast(builder, countv, i8v16, "counti");
LLVMValueRef maskarray[4] = {
LLVMConstInt(LLVMInt32Type(), 0, 0),
LLVMConstInt(LLVMInt32Type(), 4, 0),
LLVMConstInt(LLVMInt32Type(), 8, 0),
LLVMConstInt(LLVMInt32Type(), 12, 0),
};
LLVMValueRef shufflemask = LLVMConstVector(maskarray, 4);
LLVMValueRef shufflev = LLVMBuildShuffleVector(builder, counti, LLVMGetUndef(i8v16), shufflemask, "shufflev");
LLVMValueRef shuffle = LLVMBuildBitCast(builder, shufflev, LLVMInt32Type(), "shuffle");
LLVMValueRef count = lp_build_intrinsic_unary(builder, "llvm.ctpop.i32", LLVMInt32Type(), shuffle);
LLVMValueRef orig = LLVMBuildLoad(builder, counter, "orig");
LLVMValueRef incr = LLVMBuildAdd(builder, orig, count, "incr");
LLVMBuildStore(builder, incr, counter);
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned do_tri_test)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[14];
LLVMTypeRef func_type;
LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
LLVMValueRef function;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[8] = LLVMInt32Type(); /* c0 */
arg_types[9] = LLVMInt32Type(); /* c1 */
arg_types[10] = LLVMInt32Type(); /* c2 */
/* Note: the step arrays are built as int32[16] but we interpret
* them here as int32_vec4[4].
*/
arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[do_tri_test] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
a0_ptr = LLVMGetParam(function, 3);
dadx_ptr = LLVMGetParam(function, 4);
dady_ptr = LLVMGetParam(function, 5);
color_ptr_ptr = LLVMGetParam(function, 6);
depth_ptr = LLVMGetParam(function, 7);
c0 = LLVMGetParam(function, 8);
c1 = LLVMGetParam(function, 9);
c2 = LLVMGetParam(function, 10);
step0_ptr = LLVMGetParam(function, 11);
step1_ptr = LLVMGetParam(function, 12);
step2_ptr = LLVMGetParam(function, 13);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(c0, "c0");
lp_build_name(c1, "c1");
lp_build_name(c2, "c2");
lp_build_name(step0_ptr, "step0");
lp_build_name(step1_ptr, "step1");
lp_build_name(step2_ptr, "step2");
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp,
shader->base.tokens,
key->flatshade,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
int cbuf;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
do_tri_test,
c0, c1, c2,
step0_ptr, step1_ptr, step2_ptr);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
LLVMDumpValue(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
if (1)
LLVMRunFunctionPassManager(screen->pass, function);
if (LP_DEBUG & DEBUG_JIT) {
/* Print the LLVM IR to stderr */
LLVMDumpValue(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
variant->jit_function[do_tri_test] = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, function);
if (LP_DEBUG & DEBUG_ASM)
lp_disassemble(variant->jit_function[do_tri_test]);
}
static LLVMValueRef
add_blend_test(LLVMModuleRef module,
const struct pipe_blend_state *blend,
enum vector_mode mode,
struct lp_type type)
{
LLVMTypeRef ret_type;
LLVMTypeRef vec_type;
LLVMTypeRef args[4];
LLVMValueRef func;
LLVMValueRef src_ptr;
LLVMValueRef dst_ptr;
LLVMValueRef const_ptr;
LLVMValueRef res_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
ret_type = LLVMInt64Type();
vec_type = lp_build_vec_type(type);
args[3] = args[2] = args[1] = args[0] = LLVMPointerType(vec_type, 0);
func = LLVMAddFunction(module, "test", LLVMFunctionType(LLVMVoidType(), args, 4, 0));
LLVMSetFunctionCallConv(func, LLVMCCallConv);
src_ptr = LLVMGetParam(func, 0);
dst_ptr = LLVMGetParam(func, 1);
const_ptr = LLVMGetParam(func, 2);
res_ptr = LLVMGetParam(func, 3);
block = LLVMAppendBasicBlock(func, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
if (mode == AoS) {
LLVMValueRef src;
LLVMValueRef dst;
LLVMValueRef con;
LLVMValueRef res;
src = LLVMBuildLoad(builder, src_ptr, "src");
dst = LLVMBuildLoad(builder, dst_ptr, "dst");
con = LLVMBuildLoad(builder, const_ptr, "const");
res = lp_build_blend_aos(builder, blend, type, src, dst, con, 3);
lp_build_name(res, "res");
LLVMBuildStore(builder, res, res_ptr);
}
if (mode == SoA) {
LLVMValueRef src[4];
LLVMValueRef dst[4];
LLVMValueRef con[4];
LLVMValueRef res[4];
unsigned i;
for(i = 0; i < 4; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
src[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, src_ptr, &index, 1, ""), "");
dst[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
con[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
lp_build_name(src[i], "src.%c", "rgba"[i]);
lp_build_name(con[i], "con.%c", "rgba"[i]);
lp_build_name(dst[i], "dst.%c", "rgba"[i]);
}
lp_build_blend_soa(builder, blend, type, src, dst, con, res);
for(i = 0; i < 4; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
lp_build_name(res[i], "res.%c", "rgba"[i]);
LLVMBuildStore(builder, res[i], LLVMBuildGEP(builder, res_ptr, &index, 1, ""));
}
}
LLVMBuildRetVoid(builder);;
LLVMDisposeBuilder(builder);
return func;
}
/**
* Register fetch.
*/
static LLVMValueRef
emit_fetch(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
unsigned src_op)
{
struct lp_type type = bld->base.type;
const struct tgsi_full_src_register *reg = &inst->Src[src_op];
LLVMValueRef res;
unsigned chan;
assert(!reg->Register.Indirect);
/*
* Fetch the from the register file.
*/
switch (reg->Register.File) {
case TGSI_FILE_CONSTANT:
/*
* Get the constants components
*/
res = bld->base.undef;
for (chan = 0; chan < 4; ++chan) {
LLVMValueRef index;
LLVMValueRef scalar_ptr;
LLVMValueRef scalar;
LLVMValueRef swizzle;
index = LLVMConstInt(LLVMInt32Type(),
reg->Register.Index*4 + chan,
0);
scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
&index, 1, "");
scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
/*
* NOTE: constants array is always assumed to be RGBA
*/
swizzle = LLVMConstInt(LLVMInt32Type(), chan, 0);
res = LLVMBuildInsertElement(bld->base.builder, res, scalar, swizzle, "");
}
/*
* Broadcast the first quaternion to all others.
*
* XXX: could be factored into a reusable function.
*/
if (type.length > 4) {
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
unsigned i;
for (chan = 0; chan < 4; ++chan) {
shuffles[chan] = LLVMConstInt(LLVMInt32Type(), chan, 0);
}
for (i = 4; i < type.length; ++i) {
shuffles[i] = shuffles[i % 4];
}
res = LLVMBuildShuffleVector(bld->base.builder,
res, bld->base.undef,
LLVMConstVector(shuffles, type.length),
"");
}
break;
case TGSI_FILE_IMMEDIATE:
res = bld->immediates[reg->Register.Index];
assert(res);
break;
case TGSI_FILE_INPUT:
res = bld->inputs[reg->Register.Index];
assert(res);
break;
case TGSI_FILE_TEMPORARY:
{
LLVMValueRef temp_ptr;
temp_ptr = bld->temps[reg->Register.Index];
res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
if (!res)
return bld->base.undef;
}
break;
default:
assert(0 && "invalid src register in emit_fetch()");
return bld->base.undef;
}
/*
* Apply sign modifier.
*/
if (reg->Register.Absolute) {
res = lp_build_abs(&bld->base, res);
}
if(reg->Register.Negate) {
res = lp_build_negate(&bld->base, res);
}
/*
* Swizzle the argument
*/
res = swizzle_aos(bld, res,
reg->Register.SwizzleX,
reg->Register.SwizzleY,
reg->Register.SwizzleZ,
reg->Register.SwizzleW);
return res;
}
static LLVMValueRef
translateIntLit(ASTNode *Node) {
LLVMValueRef Container = LLVMBuildAlloca(Builder, LLVMInt32Type(), "");
LLVMBuildStore(Builder, getSConstInt(*((int*)Node->Value)), Container);
return Container;
}
/**
* Generate the code to do inside/outside triangle testing for the
* four pixels in a 2x2 quad. This will set the four elements of the
* quad mask vector to 0 or ~0.
* \param i which quad of the quad group to test, in [0,3]
*/
static void
generate_tri_edge_mask(LLVMBuilderRef builder,
unsigned i,
LLVMValueRef *mask, /* ivec4, out */
LLVMValueRef c0, /* int32 */
LLVMValueRef c1, /* int32 */
LLVMValueRef c2, /* int32 */
LLVMValueRef step0_ptr, /* ivec4 */
LLVMValueRef step1_ptr, /* ivec4 */
LLVMValueRef step2_ptr) /* ivec4 */
{
#define OPTIMIZE_IN_OUT_TEST 0
#if OPTIMIZE_IN_OUT_TEST
struct lp_build_if_state ifctx;
LLVMValueRef not_draw_all;
#endif
struct lp_build_flow_context *flow;
struct lp_type i32_type;
LLVMTypeRef i32vec4_type, mask_type;
LLVMValueRef c0_vec, c1_vec, c2_vec;
LLVMValueRef in_out_mask;
assert(i < 4);
/* int32 vector type */
memset(&i32_type, 0, sizeof i32_type);
i32_type.floating = FALSE; /* values are integers */
i32_type.sign = TRUE; /* values are signed */
i32_type.norm = FALSE; /* values are not normalized */
i32_type.width = 32; /* 32-bit int values */
i32_type.length = 4; /* 4 elements per vector */
i32vec4_type = lp_build_int32_vec4_type();
mask_type = LLVMIntType(32 * 4);
/*
* Use a conditional here to do detailed pixel in/out testing.
* We only have to do this if c0 != INT_MIN.
*/
flow = lp_build_flow_create(builder);
lp_build_flow_scope_begin(flow);
{
#if OPTIMIZE_IN_OUT_TEST
/* not_draw_all = (c0 != INT_MIN) */
not_draw_all = LLVMBuildICmp(builder,
LLVMIntNE,
c0,
LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
"");
in_out_mask = lp_build_int_const_scalar(i32_type, ~0);
lp_build_flow_scope_declare(flow, &in_out_mask);
/* if (not_draw_all) {... */
lp_build_if(&ifctx, flow, builder, not_draw_all);
#endif
{
LLVMValueRef step0_vec, step1_vec, step2_vec;
LLVMValueRef m0_vec, m1_vec, m2_vec;
LLVMValueRef index, m;
/* c0_vec = {c0, c0, c0, c0}
* Note that we emit this code four times but LLVM optimizes away
* three instances of it.
*/
c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
lp_build_name(c0_vec, "edgeconst0vec");
lp_build_name(c1_vec, "edgeconst1vec");
lp_build_name(c2_vec, "edgeconst2vec");
/* load step0vec, step1, step2 vec from memory */
index = LLVMConstInt(LLVMInt32Type(), i, 0);
step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
lp_build_name(step0_vec, "step0vec");
lp_build_name(step1_vec, "step1vec");
lp_build_name(step2_vec, "step2vec");
/* m0_vec = step0_ptr[i] > c0_vec */
m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
/* in_out_mask = m0_vec & m1_vec & m2_vec */
m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
lp_build_name(in_out_mask, "inoutmaskvec");
}
#if OPTIMIZE_IN_OUT_TEST
lp_build_endif(&ifctx);
#endif
}
lp_build_flow_scope_end(flow);
lp_build_flow_destroy(flow);
/* This is the initial alive/dead pixel mask for a quad of four pixels.
* It's an int[4] vector with each word set to 0 or ~0.
* Words will get cleared when pixels faile the Z test, etc.
*/
*mask = in_out_mask;
}
