/* * Copyright (C) 2005-2007 Brian Paul All Rights Reserved. * Copyright (C) 2008 VMware, Inc. All Rights Reserved. * Copyright © 2010 Intel Corporation * Copyright © 2010 Luca Barbieri * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER * DEALINGS IN THE SOFTWARE. */ /** * \file ir_to_llvm.cpp * * Translates the IR to LLVM */ /* this tends to get set as part of LLVM_CFLAGS, but we definitely want asserts */ #ifdef NDEBUG #undef NDEBUG #endif #include "llvm/ADT/ArrayRef.h" #include "llvm/DerivedTypes.h" #include "llvm/IRBuilder.h" #include "llvm/LLVMContext.h" #include "llvm/Module.h" #include "llvm/Analysis/Verifier.h" //#include "llvm/Intrinsics.h" #include <vector> #include <stdio.h> #include <map> /* #ifdef _MSC_VER #include <unordered_map> #else #include <tr1/unordered_map> #endif // use C++0x/Microsoft convention namespace std { using namespace tr1; } //*/ #include "ir.h" #include "ir_visitor.h" #include "glsl_types.h" #include "src/mesa/main/mtypes.h" // Helper function to convert array to llvm::ArrayRef template <typename T, size_t N> static inline llvm::ArrayRef<T> pack(T const (&array)[N]) { return llvm::ArrayRef<T>(array); } // Helper function to convert pointer + size to llvm::ArrayRef template <typename T> static inline llvm::ArrayRef<T> pack(T const *ptr, size_t n) { return llvm::ArrayRef<T>(ptr, n); } struct GGLState; llvm::Value * tex2D(llvm::IRBuilder<> & builder, llvm::Value * in1, const unsigned sampler, const GGLState * gglCtx); llvm::Value * texCube(llvm::IRBuilder<> & builder, llvm::Value * in1, const unsigned sampler, const GGLState * gglCtx); class ir_to_llvm_visitor : public ir_visitor { ir_to_llvm_visitor(); public: llvm::LLVMContext& ctx; llvm::Module* mod; llvm::Function* fun; // could easily support more loops, but GLSL doesn't support multiloop break/continue std::pair<llvm::BasicBlock*, llvm::BasicBlock*> loop; llvm::BasicBlock* bb; llvm::Value* result; llvm::IRBuilder<> bld; const GGLState * gglCtx; const char * shaderSuffix; llvm::Value * inputsPtr, * outputsPtr, * constantsPtr; // internal globals to store inputs/outputs/constants pointers llvm::Value * inputs, * outputs, * constants; ir_to_llvm_visitor(llvm::Module* p_mod, const GGLState * GGLCtx, const char * suffix) : ctx(p_mod->getContext()), mod(p_mod), fun(0), loop(std::make_pair((llvm::BasicBlock*)0, (llvm::BasicBlock*)0)), bb(0), bld(ctx), gglCtx(GGLCtx), shaderSuffix(suffix), inputsPtr(NULL), outputsPtr(NULL), constantsPtr(NULL), inputs(NULL), outputs(NULL), constants(NULL) { llvm::PointerType * const floatVecPtrType = llvm::PointerType::get(llvm::VectorType::get(bld.getFloatTy(),4), 0); llvm::Constant * const nullFloatVecPtr = llvm::Constant::getNullValue(floatVecPtrType); // make input, output and consts global pointers so they can be used in // different LLVM functions since the shader shares these "registers" across "functions" inputsPtr = new llvm::GlobalVariable(*mod, floatVecPtrType, false, llvm::GlobalValue::InternalLinkage, nullFloatVecPtr, "gl_inputPtr"); outputsPtr = new llvm::GlobalVariable(*mod, floatVecPtrType, false, llvm::GlobalValue::InternalLinkage, nullFloatVecPtr, "gl_outputsPtr"); constantsPtr = new llvm::GlobalVariable(*mod, floatVecPtrType, false, llvm::GlobalValue::InternalLinkage, nullFloatVecPtr, "gl_constantsPtr"); } llvm::Type* llvm_base_type(unsigned base_type) { switch(base_type) { case GLSL_TYPE_VOID: return llvm::Type::getVoidTy(ctx); case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return llvm::Type::getInt32Ty(ctx); case GLSL_TYPE_FLOAT: return llvm::Type::getFloatTy(ctx); case GLSL_TYPE_BOOL: return llvm::Type::getInt1Ty(ctx); case GLSL_TYPE_SAMPLER: return llvm::PointerType::getUnqual(llvm::Type::getVoidTy(ctx)); default: assert(0); return 0; } } llvm::Type* llvm_vec_type(const glsl_type* type) { if (type->is_array()) return llvm::ArrayType::get(llvm_type(type->fields.array), type->array_size()); if (type->is_record()) { std::vector<llvm::Type*> fields; for (unsigned i = 0; i < type->length; i++) fields.push_back(llvm_type(type->fields.structure[i].type)); return llvm::StructType::get(ctx, llvm::ArrayRef<llvm::Type*>( fields)); } llvm::Type* base_type = llvm_base_type(type->base_type); if (type->vector_elements <= 1) { return base_type; } else { return llvm::VectorType::get(base_type, type->vector_elements); } } llvm::Type* llvm_type(const glsl_type* type) { llvm::Type* vec_type = llvm_vec_type(type); if (type->matrix_columns <= 1) { return vec_type; } else { return llvm::ArrayType::get(vec_type, type->matrix_columns); } } typedef std::map<ir_variable*, llvm::Value*> llvm_variables_t; //typedef std::unordered_map<ir_variable*, llvm::Value*> llvm_variables_t; llvm_variables_t llvm_variables; llvm::Value* llvm_variable(class ir_variable* var) { llvm_variables_t::iterator vari = llvm_variables.find(var); if (vari != llvm_variables.end()) { return vari->second; } else { llvm::Type* type = llvm_type(var->type); llvm::Value* v = NULL; if(fun) { if (ir_var_in == var->mode) { assert(var->location >= 0); v = bld.CreateConstGEP1_32(inputs, var->location); v = bld.CreateBitCast(v, llvm::PointerType::get(llvm_type(var->type), 0), var->name); } else if (ir_var_out == var->mode) { assert(var->location >= 0); v = bld.CreateConstGEP1_32(outputs, var->location); v = bld.CreateBitCast(v, llvm::PointerType::get(llvm_type(var->type), 0), var->name); } else if (ir_var_uniform == var->mode) { assert(var->location >= 0); v = bld.CreateConstGEP1_32(constants, var->location); v = bld.CreateBitCast(v, llvm::PointerType::get(llvm_type(var->type), 0), var->name); } else { if(bb == &fun->getEntryBlock()) v = bld.CreateAlloca(type, 0, var->name); else v = new llvm::AllocaInst(type, 0, var->name, fun->getEntryBlock().getTerminator()); } } else { // TODO: can anything global be non-constant in GLSL?; fix linkage //printf("var '%s' mode=%d location=%d \n", var->name, var->mode, var->location); switch(var->mode) { case ir_var_auto: // fall through case ir_var_temporary: { llvm::Constant * init = llvm::UndefValue::get(llvm_type(var->type)); if(var->constant_value) init = llvm_constant(var->constant_value); v = new llvm::GlobalVariable(*mod, type, var->read_only, llvm::GlobalValue::InternalLinkage, init, var->name); break; } case ir_var_in: // fall through case ir_var_out: // fall through case ir_var_uniform: // fall through assert(var->location >= 0); return NULL; // variable outside of function means declaration default: assert(0); } // llvm::Function::LinkageTypes linkage; // if(var->mode == ir_var_auto || var->mode == ir_var_temporary) // linkage = llvm::GlobalValue::InternalLinkage; // else // linkage = llvm::GlobalValue::ExternalLinkage; // llvm::Constant* init = 0; // if(var->constant_value) // { // init = llvm_constant(var->constant_value); // // this constants need to be external (ie. written to output) // if (llvm::GlobalValue::ExternalLinkage == linkage) // linkage = llvm::GlobalValue::AvailableExternallyLinkage; // } // else if(linkage == llvm::GlobalValue::InternalLinkage) // init = llvm::UndefValue::get(llvm_type(var->type)); // v = new llvm::GlobalVariable(*mod, type, var->read_only, linkage, init, var->name); } assert(v); llvm_variables[var] = v; return v; } } //typedef std::map<ir_function_signature*, llvm::Function*> llvm_functions_t; //typedef std::unordered_map<ir_function_signature*, llvm::Function*> llvm_functions_t; //llvm_functions_t llvm_functions; llvm::Function* llvm_function(class ir_function_signature* sig) { const char* name = sig->function_name(); char * functionName = (char *)malloc(strlen(name) + strlen(shaderSuffix) + 1); strcpy(functionName, name); strcat(functionName, shaderSuffix); llvm::Function * function = mod->getFunction(functionName); if (function) { free(functionName); return function; } else { llvm::Function::LinkageTypes linkage; std::vector<llvm::Type*> params; foreach_iter(exec_list_iterator, iter, sig->parameters) { ir_variable* arg = (ir_variable*)iter.get(); params.push_back(llvm_type(arg->type)); } if(!strcmp(name, "main") || !sig->is_defined) { linkage = llvm::Function::ExternalLinkage; llvm::PointerType * vecPtrTy = llvm::PointerType::get(llvm::VectorType::get(bld.getFloatTy(), 4), 0); assert(0 == params.size()); params.push_back(vecPtrTy); // inputs params.push_back(vecPtrTy); // outputs params.push_back(vecPtrTy); // constants } else { linkage = llvm::Function::InternalLinkage; } llvm::FunctionType* ft = llvm::FunctionType::get(llvm_type(sig->return_type), llvm::ArrayRef<llvm::Type*>(params), false); function = llvm::Function::Create(ft, linkage, functionName, mod); free(functionName); return function; } } llvm::Value* llvm_value(class ir_instruction* ir) { result = 0; ir->accept(this); return result; } llvm::Constant* llvm_constant(class ir_instruction* ir) { return (llvm::Constant *)llvm_value(ir); //return &dynamic_cast<llvm::Constant&>(*llvm_value(ir)); } llvm::Constant* llvm_int(unsigned v) { return llvm::ConstantInt::get(llvm::Type::getInt32Ty(ctx), v); } llvm::Value* llvm_pointer(class ir_rvalue* ir) { if(ir_dereference_variable* deref = ir->as_dereference_variable()) return llvm_variable(deref->variable_referenced()); else if(ir_dereference_array* deref = ir->as_dereference_array()) { llvm::Value* gep[2] = {llvm_int(0), llvm_value(deref->array_index)}; return bld.CreateInBoundsGEP(llvm_pointer(deref->array), gep); } else if(ir->as_dereference()) { ir_dereference_record* deref = (ir_dereference_record*)ir; int idx = deref->record->type->field_index(deref->field); assert(idx >= 0); return bld.CreateConstInBoundsGEP2_32(llvm_pointer(deref->record), 0, idx); } else { assert(0); return 0; } } // llvm::Value* llvm_intrinsic(llvm::Intrinsic::ID id, llvm::Value* a) // { // llvm::Type* types[1] = {a->getType()}; // return bld.CreateCall(llvm::Intrinsic::getDeclaration(mod, id, types, 1), a); // } // // llvm::Value* llvm_intrinsic(llvm::Intrinsic::ID id, llvm::Value* a, llvm::Value* b) // { // llvm::Type* types[2] = {a->getType(), b->getType()}; // /* only one type suffix is usually needed, so pass 1 here */ // return bld.CreateCall2(llvm::Intrinsic::getDeclaration(mod, id, types, 1), a, b); // } llvm::Value* llvm_intrinsic_unop(ir_expression_operation op, llvm::Value * op0) { llvm::Type * floatType = llvm::Type::getFloatTy(ctx); const char * name = NULL; switch (op) { case ir_unop_sin: name = "sinf"; break; case ir_unop_cos: name = "cosf"; break; default: assert(0); } llvm::Function * function = mod->getFunction(name); if (!function) { // predeclare the intrinsic std::vector<llvm::Type*> args; args.push_back(floatType); llvm::FunctionType* type = llvm::FunctionType::get(floatType, llvm::ArrayRef<llvm::Type*>(args), false); function = llvm::Function::Create(type, llvm::Function::ExternalLinkage, name, mod); function->setCallingConv(llvm::CallingConv::C); } return bld.CreateCall(function, op0); } llvm::Value* llvm_intrinsic_binop(ir_expression_operation op, llvm::Value * op0, llvm::Value * op1) { llvm::Type * floatType = llvm::Type::getFloatTy(ctx); const char * name = NULL; switch (op) { case ir_binop_pow: name = "powf"; break; default: assert(0); } llvm::Function * function = mod->getFunction(name); if (!function) { // predeclare the intrinsic std::vector<llvm::Type*> args; args.push_back(floatType); args.push_back(floatType); llvm::FunctionType* type = llvm::FunctionType::get(floatType, llvm::ArrayRef<llvm::Type*>(args), false); function = llvm::Function::Create(type, llvm::Function::ExternalLinkage, name, mod); function->setCallingConv(llvm::CallingConv::C); } return bld.CreateCall2(function, op0, op1); } llvm::Constant* llvm_imm(llvm::Type* type, double v) { if(type->isVectorTy()) { std::vector<llvm::Constant*> values; values.push_back(llvm_imm(((llvm::VectorType*)type)->getElementType(), v)); for(unsigned i = 1; i < ((llvm::VectorType*)type)->getNumElements(); ++i) values.push_back(values[0]); return llvm::ConstantVector::get(values); } else if(type->isIntegerTy()) return llvm::ConstantInt::get(type, v); else if(type->isFloatingPointTy()) return llvm::ConstantFP::get(type, v); else { assert(0); return 0; } } static llvm::Value* create_shuffle3(llvm::IRBuilder<>& bld, llvm::Value* v, unsigned a, unsigned b, unsigned c, const llvm::Twine& name = "") { llvm::Type* int_ty = llvm::Type::getInt32Ty(v->getContext()); llvm::Constant* vals[3] = {llvm::ConstantInt::get(int_ty, a), llvm::ConstantInt::get(int_ty, b), llvm::ConstantInt::get(int_ty, c)}; return bld.CreateShuffleVector(v, llvm::UndefValue::get(v->getType()), llvm::ConstantVector::get(pack(vals)), name); } llvm::Value* create_select(unsigned width, llvm::Value * cond, llvm::Value * tru, llvm::Value * fal, const char * name = "") { if (1 == width) return bld.CreateSelect(cond, tru, fal, name); llvm::Type * vectorType = tru->getType(); llvm::Value * vector = llvm::Constant::getNullValue(vectorType); for (unsigned int i = 0; i < width; i++) { llvm::Value * c = bld.CreateExtractElement(cond, llvm_int(i)); llvm::Value * t = bld.CreateExtractElement(tru, llvm_int(i)); llvm::Value * f = bld.CreateExtractElement(fal, llvm_int(i)); llvm::Value * v = bld.CreateSelect(c, t, f, name); vector = bld.CreateInsertElement(vector, v, llvm_int(i), "vslct"); } return vector; } llvm::Value* create_dot_product(llvm::Value* ops0, llvm::Value* ops1, glsl_base_type type, unsigned width) { llvm::Value* prod; switch (type) { case GLSL_TYPE_UINT: case GLSL_TYPE_INT: prod = bld.CreateMul(ops0, ops1, "dot.mul"); break; case GLSL_TYPE_FLOAT: prod = bld.CreateFMul(ops0, ops1, "dot.mul"); break; default: assert(0); } if (width<= 1) return prod; llvm::Value* sum = 0; for (unsigned i = 0; i < width; ++i) { llvm::Value* elem = bld.CreateExtractElement(prod, llvm_int(i), "dot.elem"); if (sum) { if (type == GLSL_TYPE_FLOAT) sum = bld.CreateFAdd(sum, elem, "dot.add"); else sum = bld.CreateAdd(sum, elem, "dot.add"); } else sum = elem; } return sum; } llvm::Value* llvm_expression(ir_expression* ir) { llvm::Value* ops[2]; for(unsigned i = 0; i < ir->get_num_operands(); ++i) ops[i] = llvm_value(ir->operands[i]); if(ir->get_num_operands() == 2) { int vecidx = -1; int scaidx = -1; if(ir->operands[0]->type->vector_elements <= 1 && ir->operands[1]->type->vector_elements > 1) { scaidx = 0; vecidx = 1; } else if(ir->operands[0]->type->vector_elements > 1 && ir->operands[1]->type->vector_elements <= 1) { scaidx = 1; vecidx = 0; } else assert(ir->operands[0]->type->vector_elements == ir->operands[1]->type->vector_elements); if(scaidx >= 0) { llvm::Value* vec; vec = llvm::UndefValue::get(ops[vecidx]->getType()); for(unsigned i = 0; i < ir->operands[vecidx]->type->vector_elements; ++i) vec = bld.CreateInsertElement(vec, ops[scaidx], llvm_int(i), "sca2vec"); ops[scaidx] = vec; } } switch (ir->operation) { case ir_unop_logic_not: return bld.CreateNot(ops[0]); case ir_unop_neg: switch (ir->operands[0]->type->base_type) { case GLSL_TYPE_UINT: case GLSL_TYPE_BOOL: case GLSL_TYPE_INT: return bld.CreateNeg(ops[0]); case GLSL_TYPE_FLOAT: return bld.CreateFNeg(ops[0]); default: assert(0); } case ir_unop_abs: switch (ir->operands[0]->type->base_type) { case GLSL_TYPE_UINT: case GLSL_TYPE_BOOL: return ops[0]; case GLSL_TYPE_INT: return create_select(ir->operands[0]->type->vector_elements, bld.CreateICmpSGE(ops[0], llvm_imm(ops[0]->getType(), 0), "sabs.ge"), ops[0], bld.CreateNeg(ops[0], "sabs.neg"), "sabs.select"); case GLSL_TYPE_FLOAT: return create_select(ir->operands[0]->type->vector_elements, bld.CreateFCmpUGE(ops[0], llvm_imm(ops[0]->getType(), 0), "fabs.ge"), ops[0], bld.CreateFNeg(ops[0], "fabs.neg"), "fabs.select"); default: assert(0); } case ir_unop_sign: switch (ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: return ops[0]; case GLSL_TYPE_UINT: return bld.CreateZExt(bld.CreateICmpNE(ops[0], llvm_imm(ops[0]->getType(), 0), "usign.ne"), ops[0]->getType(), "usign.zext"); case GLSL_TYPE_INT: return bld.CreateSelect(bld.CreateICmpNE(ops[0], llvm_imm(ops[0]->getType(), 0), "ssign.ne"), bld.CreateSelect(bld.CreateICmpSGE(ops[0], llvm_imm(ops[0]->getType(), 0), "ssign.ge"), llvm_imm(ops[0]->getType(), 1), llvm_imm(ops[0]->getType(), -1), "sabs.selects"), llvm_imm(ops[0]->getType(), 0), "sabs.select0"); case GLSL_TYPE_FLOAT: return bld.CreateSelect(bld.CreateFCmpONE(ops[0], llvm_imm(ops[0]->getType(), 0), "fsign.ne"), bld.CreateSelect(bld.CreateFCmpUGE(ops[0], llvm_imm(ops[0]->getType(), 0), "fsign.ge"), llvm_imm(ops[0]->getType(), 1), llvm_imm(ops[0]->getType(), -1), "fabs.selects"), llvm_imm(ops[0]->getType(), 0), "fabs.select0"); default: assert(0); } case ir_unop_rcp: assert(ir->operands[0]->type->base_type == GLSL_TYPE_FLOAT); return bld.CreateFDiv(llvm_imm(ops[0]->getType(), 1), ops[0]); case ir_unop_exp: // fall through case ir_unop_exp2: // fall through case ir_unop_log: // fall through case ir_unop_log2: // fall through case ir_unop_sin: // fall through case ir_unop_cos: assert(ir->operands[0]->type->base_type == GLSL_TYPE_FLOAT); return llvm_intrinsic_unop(ir->operation, ops[0]); // TODO: implement these somehow case ir_unop_dFdx: assert(0); //return llvm_intrinsic(llvm::Intrinsic::ddx, ops[0]); case ir_unop_dFdy: assert(0); //return llvm_intrinsic(llvm::Intrinsic::ddy, ops[0]); case ir_binop_add: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return bld.CreateAdd(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFAdd(ops[0], ops[1]); default: assert(0); } case ir_binop_sub: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return bld.CreateSub(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFSub(ops[0], ops[1]); default: assert(0); } case ir_binop_mul: if (ir->operands[0]->type->is_matrix() && ir->operands[1]->type->is_vector()) assert(0); else if (ir->operands[0]->type->is_vector() && ir->operands[1]->type->is_matrix()) { assert(0); // matrix multiplication should have been lowered to vector ops llvm::VectorType * vectorType = llvm::VectorType::get(llvm_base_type(ir->operands[1]->type->base_type), ir->operands[1]->type->matrix_columns); llvm::Value * vector = llvm::Constant::getNullValue(vectorType); for (unsigned int i = 0; i < ir->operands[1]->type->matrix_columns; i++) { llvm::Value * value = bld.CreateExtractValue(ops[1], i, "vec*mat_col"); value = create_dot_product(value, ops[0], ir->operands[1]->type->base_type, ir->operands[1]->type->vector_elements); vector = bld.CreateInsertElement(vector, value, llvm_int(i), "vec*mat_res"); } return vector; } else if (ir->operands[0]->type->is_matrix() && ir->operands[1]->type->is_matrix()) assert(0); switch (ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: return bld.CreateAnd(ops[0], ops[1]); case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return bld.CreateMul(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFMul(ops[0], ops[1]); default: assert(0); } case ir_binop_div: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateUDiv(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateSDiv(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFDiv(ops[0], ops[1]); default: assert(0); } case ir_binop_mod: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateURem(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateSRem(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFRem(ops[0], ops[1]); default: assert(0); } case ir_binop_less: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateICmpULT(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateICmpSLT(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFCmpOLT(ops[0], ops[1]); default: assert(0); } case ir_binop_greater: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateICmpUGT(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateICmpSGT(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFCmpOGT(ops[0], ops[1]); default: assert(0); } case ir_binop_lequal: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateICmpULE(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateICmpSLE(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFCmpOLE(ops[0], ops[1]); default: assert(0); } case ir_binop_gequal: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateICmpUGE(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateICmpSGE(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFCmpOGE(ops[0], ops[1]); default: assert(0); } case ir_binop_equal: // fall through case ir_binop_all_equal: // TODO: check op same as ir_binop_equal switch (ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return bld.CreateICmpEQ(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFCmpOEQ(ops[0], ops[1]); default: assert(0); } case ir_binop_nequal: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return bld.CreateICmpNE(ops[0], ops[1]); case GLSL_TYPE_FLOAT: return bld.CreateFCmpONE(ops[0], ops[1]); default: assert(0); } case ir_binop_logic_xor: assert(ir->operands[0]->type->base_type == GLSL_TYPE_BOOL); return bld.CreateICmpNE(ops[0], ops[1]); case ir_binop_logic_or: assert(ir->operands[0]->type->base_type == GLSL_TYPE_BOOL); return bld.CreateOr(ops[0], ops[1]); case ir_binop_logic_and: assert(ir->operands[0]->type->base_type == GLSL_TYPE_BOOL); return bld.CreateAnd(ops[0], ops[1]); case ir_binop_dot: return create_dot_product(ops[0], ops[1], ir->operands[0]->type->base_type, ir->operands[0]->type->vector_elements); // case ir_binop_cross: this op does not exist in ir.h // assert(ir->operands[0]->type->vector_elements == 3); // switch(ir->operands[0]->type->base_type) // { // case GLSL_TYPE_UINT: // case GLSL_TYPE_INT: // return bld.CreateSub( // bld.CreateMul(create_shuffle3(bld, ops[0], 1, 2, 0, "cross.a120"), create_shuffle3(bld, ops[1], 2, 0, 1, "cross.a201"), "cross.ab"), // bld.CreateMul(create_shuffle3(bld, ops[1], 1, 2, 0, "cross.b120"), create_shuffle3(bld, ops[0], 2, 0, 1, "cross.b201"), "cross.ba"), // "cross.sub"); // case GLSL_TYPE_FLOAT: // return bld.CreateFSub( // bld.CreateFMul(create_shuffle3(bld, ops[0], 1, 2, 0, "cross.a120"), create_shuffle3(bld, ops[1], 2, 0, 1, "cross.a201"), "cross.ab"), // bld.CreateFMul(create_shuffle3(bld, ops[1], 1, 2, 0, "cross.b120"), create_shuffle3(bld, ops[0], 2, 0, 1, "cross.b201"), "cross.ba"), // "cross.sub"); // default: // assert(0); // } case ir_unop_sqrt: assert(ir->operands[0]->type->base_type == GLSL_TYPE_FLOAT); return llvm_intrinsic_unop(ir->operation, ops[0]); case ir_unop_rsq: assert(ir->operands[0]->type->base_type == GLSL_TYPE_FLOAT); return bld.CreateFDiv(llvm_imm(ops[0]->getType(), 1), llvm_intrinsic_unop(ir_unop_sqrt, ops[0]), "rsqrt.rcp"); case ir_unop_i2f: return bld.CreateSIToFP(ops[0], llvm_type(ir->type)); case ir_unop_u2f: case ir_unop_b2f: return bld.CreateUIToFP(ops[0], llvm_type(ir->type)); case ir_unop_b2i: return bld.CreateZExt(ops[0], llvm_type(ir->type)); case ir_unop_f2i: return bld.CreateFPToSI(ops[0], llvm_type(ir->type)); case ir_unop_f2b: return bld.CreateFCmpONE(ops[0], llvm_imm(ops[0]->getType(), 0)); case ir_unop_i2b: return bld.CreateICmpNE(ops[0], llvm_imm(ops[0]->getType(), 0)); case ir_unop_trunc: { if(ir->operands[0]->type->base_type != GLSL_TYPE_FLOAT) return ops[0]; glsl_type int_type = *ir->operands[0]->type; int_type.base_type = GLSL_TYPE_INT; return bld.CreateSIToFP(bld.CreateFPToSI(ops[0], llvm_type(&int_type), "trunc.fptosi"),ops[0]->getType(), "trunc.sitofp"); } case ir_unop_floor: { if(ir->operands[0]->type->base_type != GLSL_TYPE_FLOAT) return ops[0]; llvm::Value* one = llvm_imm(ops[0]->getType(), 1); return bld.CreateFSub(ops[0], bld.CreateFRem(ops[0], one)); } case ir_unop_ceil: { if(ir->operands[0]->type->base_type != GLSL_TYPE_FLOAT) return ops[0]; llvm::Value* one = llvm_imm(ops[0]->getType(), 1); return bld.CreateFAdd(bld.CreateFSub(ops[0], bld.CreateFRem(ops[0], one)), one); } case ir_unop_fract: { if(ir->operands[0]->type->base_type != GLSL_TYPE_FLOAT) return llvm_imm(ops[0]->getType(), 0); llvm::Value* one = llvm_imm(ops[0]->getType(), 1); return bld.CreateFRem(ops[0], one); } // TODO: NaNs might be wrong in min/max, not sure how to fix it case ir_binop_min: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: return bld.CreateAnd(ops[0], ops[1], "bmin"); case GLSL_TYPE_UINT: return bld.CreateSelect(bld.CreateICmpULE(ops[0], ops[1], "umin.le"), ops[0], ops[1], "umin.select"); case GLSL_TYPE_INT: return bld.CreateSelect(bld.CreateICmpSLE(ops[0], ops[1], "smin.le"), ops[0], ops[1], "smin.select"); case GLSL_TYPE_FLOAT: return bld.CreateSelect(bld.CreateFCmpULE(ops[0], ops[1], "fmin.le"), ops[0], ops[1], "fmin.select"); default: assert(0); } case ir_binop_max: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: return bld.CreateOr(ops[0], ops[1], "bmax"); case GLSL_TYPE_UINT: return bld.CreateSelect(bld.CreateICmpUGE(ops[0], ops[1], "umax.ge"), ops[0], ops[1], "umax.select"); case GLSL_TYPE_INT: return bld.CreateSelect(bld.CreateICmpSGE(ops[0], ops[1], "smax.ge"), ops[0], ops[1], "smax.select"); case GLSL_TYPE_FLOAT: return bld.CreateSelect(bld.CreateFCmpUGE(ops[0], ops[1], "fmax.ge"), ops[0], ops[1], "fmax.select"); default: assert(0); } case ir_binop_pow: assert(GLSL_TYPE_FLOAT == ir->operands[0]->type->base_type); assert(GLSL_TYPE_FLOAT == ir->operands[1]->type->base_type); return llvm_intrinsic_binop(ir_binop_pow, ops[0], ops[1]); case ir_unop_bit_not: return bld.CreateNot(ops[0]); case ir_binop_bit_and: return bld.CreateAnd(ops[0], ops[1]); case ir_binop_bit_xor: return bld.CreateXor(ops[0], ops[1]); case ir_binop_bit_or: return bld.CreateOr(ops[0], ops[1]); case ir_binop_lshift: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: case GLSL_TYPE_INT: return bld.CreateLShr(ops[0], ops[1]); default: assert(0); } case ir_binop_rshift: switch(ir->operands[0]->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: return bld.CreateLShr(ops[0], ops[1]); case GLSL_TYPE_INT: return bld.CreateAShr(ops[0], ops[1]); default: assert(0); return 0; } default: printf("ir->operation=%d \n", ir->operation); assert(0); return 0; } } virtual void visit(class ir_expression * ir) { result = llvm_expression(ir); } virtual void visit(class ir_dereference_array *ir) { result = bld.CreateLoad(llvm_pointer(ir)); } virtual void visit(class ir_dereference_record *ir) { result = bld.CreateLoad(llvm_pointer(ir)); } virtual void visit(class ir_dereference_variable *ir) { result = bld.CreateLoad(llvm_pointer(ir), ir->variable_referenced()->name); } virtual void visit(class ir_texture * ir) { llvm::Value * coordinate = llvm_value(ir->coordinate); if (ir->projector) { llvm::Value * proj = llvm_value(ir->projector); unsigned width = ((llvm::VectorType*)coordinate->getType())->getNumElements(); llvm::Value * div = llvm::Constant::getNullValue(coordinate->getType()); for (unsigned i = 0; i < width; i++) div = bld.CreateInsertElement(div, proj, bld.getInt32(i), "texProjDup"); coordinate = bld.CreateFDiv(coordinate, div, "texProj"); } ir_variable * sampler = NULL; if(ir_dereference_variable* deref = ir->sampler->as_dereference_variable()) sampler = deref->variable_referenced(); else if(ir_dereference_array* deref = ir->sampler->as_dereference_array()) { assert(0); // not implemented return; deref->array_index; deref->array; } else if(ir->sampler->as_dereference()) { assert(0); // not implemented ir_dereference_record* deref = (ir_dereference_record*)ir->sampler; int idx = deref->record->type->field_index(deref->field); assert(idx >= 0); } else assert(0); assert(sampler->location >= 0 && sampler->location < 64); // TODO: proper limit // ESSL texture LOD is only for 2D texture in vert shader, and it's explicit // bias used only in frag shader, and added to computed LOD assert(ir_tex == ir->op); assert(GLSL_TYPE_FLOAT == sampler->type->sampler_type); printf("sampler '%s' location=%d dim=%d type=%d proj=%d lod=%d \n", sampler->name, sampler->location, sampler->type->sampler_dimensionality, sampler->type->sampler_type, ir->projector ? 1 : 0, ir->lod_info.lod ? 1 : 0); if (GLSL_SAMPLER_DIM_CUBE == sampler->type->sampler_dimensionality) result = texCube(bld, coordinate, sampler->location, gglCtx); else if (GLSL_SAMPLER_DIM_2D == sampler->type->sampler_dimensionality) result = tex2D(bld, coordinate, sampler->location, gglCtx); else assert(0); } virtual void visit(class ir_discard * ir) { llvm::BasicBlock* discard = llvm::BasicBlock::Create(ctx, "discard", fun); llvm::BasicBlock* after; if(ir->condition) { after = llvm::BasicBlock::Create(ctx, "discard.survived", fun); bld.CreateCondBr(llvm_value(ir->condition), discard, after); } else { after = llvm::BasicBlock::Create(ctx, "dead_code.discard", fun); bld.CreateBr(discard); } bld.SetInsertPoint(discard); // FIXME: According to the LLVM mailing list, UnwindInst should not // be used by the frontend since LLVM 3.0, and 'CreateUnwind' // method has been removed from the IRBuilder. Here's the // temporary workaround. But it would be better to remove // this in the future. // // A solution after LLVM 3.0: To add a global boolean in the shader to // store whether it was discarded or not and just continue on normally, // and handle the discard outside the shader, in the scanline function. // The discard instruction is not used frequently, so it should be okay // performance wise. //new llvm::UnwindInst(ctx, discard); /// Deprecated bb = after; bld.SetInsertPoint(bb); } virtual void visit(class ir_loop_jump *ir) { llvm::BasicBlock* target; if(ir->mode == ir_loop_jump::jump_continue) target = loop.first; else if(ir->mode == ir_loop_jump::jump_break) target = loop.second; assert(target); bld.CreateBr(target); bb = llvm::BasicBlock::Create(ctx, "dead_code.jump", fun); bld.SetInsertPoint(bb); } virtual void visit(class ir_loop * ir) { llvm::BasicBlock* body = llvm::BasicBlock::Create(ctx, "loop", fun); llvm::BasicBlock* header = body; llvm::BasicBlock* after = llvm::BasicBlock::Create(ctx, "loop.after", fun); llvm::Value* ctr; if(ir->counter) { ctr = llvm_variable(ir->counter); if(ir->from) bld.CreateStore(llvm_value(ir->from), ctr); if(ir->to) header = llvm::BasicBlock::Create(ctx, "loop.header", fun); } bld.CreateBr(header); if(ir->counter && ir->to) { bld.SetInsertPoint(header); llvm::Value* cond; llvm::Value* load = bld.CreateLoad(ctr); llvm::Value* to = llvm_value(ir->to); switch(ir->counter->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: cond = bld.CreateICmpULT(load, to); break; case GLSL_TYPE_INT: cond = bld.CreateICmpSLT(load, to); break; case GLSL_TYPE_FLOAT: cond = bld.CreateFCmpOLT(load, to); break; } bld.CreateCondBr(cond, body, after); } bld.SetInsertPoint(body); std::pair<llvm::BasicBlock*, llvm::BasicBlock*> saved_loop = loop; loop = std::make_pair(header, after); visit_exec_list(&ir->body_instructions, this); loop = saved_loop; if(ir->counter && ir->increment) { switch(ir->counter->type->base_type) { case GLSL_TYPE_BOOL: case GLSL_TYPE_UINT: case GLSL_TYPE_INT: bld.CreateStore(bld.CreateAdd(bld.CreateLoad(ctr), llvm_value(ir->increment)), ctr); break; case GLSL_TYPE_FLOAT: bld.CreateStore(bld.CreateFAdd(bld.CreateLoad(ctr), llvm_value(ir->increment)), ctr); break; } } bld.CreateBr(header); bb = after; bld.SetInsertPoint(bb); } virtual void visit(class ir_if *ir) { llvm::BasicBlock* bbt = llvm::BasicBlock::Create(ctx, "if", fun); llvm::BasicBlock* bbf = llvm::BasicBlock::Create(ctx, "else", fun); llvm::BasicBlock* bbe = llvm::BasicBlock::Create(ctx, "endif", fun); bld.CreateCondBr(llvm_value(ir->condition), bbt, bbf); bld.SetInsertPoint(bbt); visit_exec_list(&ir->then_instructions, this); bld.CreateBr(bbe); bld.SetInsertPoint(bbf); visit_exec_list(&ir->else_instructions, this); bld.CreateBr(bbe); bb = bbe; bld.SetInsertPoint(bb); } virtual void visit(class ir_return * ir) { if(!ir->value) bld.CreateRetVoid(); else bld.CreateRet(llvm_value(ir->value)); bb = llvm::BasicBlock::Create(ctx, "dead_code.return", fun); bld.SetInsertPoint(bb); } virtual void visit(class ir_call * ir) { std::vector<llvm::Value*> args; foreach_iter(exec_list_iterator, iter, *ir) { ir_rvalue *arg = (ir_constant *)iter.get(); args.push_back(llvm_value(arg)); } result = bld.CreateCall(llvm_function(ir->get_callee()), llvm::ArrayRef<llvm::Value*>(args)); llvm::AttrListPtr attr; ((llvm::CallInst*)result)->setAttributes(attr); } virtual void visit(class ir_constant * ir) { if (ir->type->base_type == GLSL_TYPE_STRUCT) { std::vector<llvm::Constant*> fields; foreach_iter(exec_list_iterator, iter, ir->components) { ir_constant *field = (ir_constant *)iter.get(); fields.push_back(llvm_constant(field)); } result = llvm::ConstantStruct::get((llvm::StructType*)llvm_type(ir->type), fields); } else if (ir->type->base_type == GLSL_TYPE_ARRAY) { std::vector<llvm::Constant*> elems; for (unsigned i = 0; i < ir->type->length; i++) elems.push_back(llvm_constant(ir->array_elements[i])); result = llvm::ConstantArray::get((llvm::ArrayType*)llvm_type(ir->type), elems); } else { llvm::Type* base_type = llvm_base_type(ir->type->base_type); llvm::Type* vec_type = llvm_vec_type(ir->type); llvm::Type* type = llvm_type(ir->type); std::vector<llvm::Constant*> vecs; unsigned idx = 0; for (unsigned i = 0; i < ir->type->matrix_columns; ++i) { std::vector<llvm::Constant*> elems; for (unsigned j = 0; j < ir->type->vector_elements; ++j) { llvm::Constant* elem; switch(ir->type->base_type) { case GLSL_TYPE_FLOAT: elem = llvm::ConstantFP::get(base_type, ir->value.f[idx]); break; case GLSL_TYPE_UINT: elem = llvm::ConstantInt::get(base_type, ir->value.u[idx]); break; case GLSL_TYPE_INT: elem = llvm::ConstantInt::get(base_type, ir->value.i[idx]); break; case GLSL_TYPE_BOOL: elem = llvm::ConstantInt::get(base_type, ir->value.b[idx]); break; } elems.push_back(elem); ++idx; } llvm::Constant* vec; if(ir->type->vector_elements > 1) { llvm::ArrayRef<llvm::Constant*> ConstantArray(elems); vec = llvm::ConstantVector::get(ConstantArray); } else { vec = elems[0]; } vecs.push_back(vec); } if(ir->type->matrix_columns > 1) result = llvm::ConstantArray::get((llvm::ArrayType*)type, vecs); else result = vecs[0]; } } llvm::Value* llvm_shuffle(llvm::Value* val, int* shuffle_mask, unsigned res_width, const llvm::Twine &name = "") { llvm::Type* elem_type = val->getType(); llvm::Type* res_type = elem_type;; unsigned val_width = 1; if(val->getType()->isVectorTy()) { val_width = ((llvm::VectorType*)val->getType())->getNumElements(); elem_type = ((llvm::VectorType*)val->getType())->getElementType(); } if(res_width > 1) res_type = llvm::VectorType::get(elem_type, res_width); llvm::Constant* shuffle_mask_values[4]; assert(res_width <= 4); bool any_def = false; for(unsigned i = 0; i < res_width; ++i) { if(shuffle_mask[i] < 0) shuffle_mask_values[i] = llvm::UndefValue::get(llvm::Type::getInt32Ty(ctx)); else { any_def = true; shuffle_mask_values[i] = llvm_int(shuffle_mask[i]); } } llvm::Value* undef = llvm::UndefValue::get(res_type); if(!any_def) return undef; if(val_width > 1) { if(res_width > 1) { if(val_width == res_width) { bool nontrivial = false; for(unsigned i = 0; i < val_width; ++i) { if(shuffle_mask[i] != (int)i) nontrivial = true; } if(!nontrivial) return val; } return bld.CreateShuffleVector(val, llvm::UndefValue::get(val->getType()), llvm::ConstantVector::get(pack(shuffle_mask_values, res_width)), name); } else return bld.CreateExtractElement(val, llvm_int(shuffle_mask[0]), name); } else { if(res_width > 1) { llvm::Value* tmp = undef; for(unsigned i = 0; i < res_width; ++i) { if(shuffle_mask[i] >= 0) tmp = bld.CreateInsertElement(tmp, val, llvm_int(i), name); } return tmp; } else if(shuffle_mask[0] >= 0) return val; else return undef; } } virtual void visit(class ir_swizzle * swz) { llvm::Value* val = llvm_value(swz->val); int mask[4] = {swz->mask.x, swz->mask.y, swz->mask.z, swz->mask.w}; result = llvm_shuffle(val, mask, swz->mask.num_components, "swizzle"); } virtual void visit(class ir_assignment * ir) { llvm::Value* lhs = llvm_pointer(ir->lhs); llvm::Value* rhs = llvm_value(ir->rhs); unsigned width = ir->lhs->type->vector_elements; unsigned mask = (1 << width) - 1; assert(rhs); // TODO: masking for matrix assignment if (ir->rhs->type->is_matrix()) { bld.CreateStore(rhs, lhs, "mat_str"); return; } if (!(ir->write_mask & mask)) return; if (ir->rhs->type->vector_elements < width) { int expand_mask[4] = {-1, -1, -1, -1}; for (unsigned i = 0; i < ir->lhs->type->vector_elements; ++i) expand_mask[i] = i; // printf("ve: %u w %u issw: %i\n", ir->rhs->type->vector_elements, width, !!ir->rhs->as_swizzle()); rhs = llvm_shuffle(rhs, expand_mask, width, "assign.expand"); } if (width > 1 && (ir->write_mask & mask) != mask) { llvm::Constant* blend_mask[4]; // refer to ir.h: ir_assignment::write_mask // A partially-set write mask means that each enabled channel gets // the value from a consecutive channel of the rhs. unsigned rhsChannel = 0; for (unsigned i = 0; i < width; ++i) { if (ir->write_mask & (1 << i)) blend_mask[i] = llvm_int(width + rhsChannel++); else blend_mask[i] = llvm_int(i); } rhs = bld.CreateShuffleVector(bld.CreateLoad(lhs), rhs, llvm::ConstantVector::get(pack(blend_mask, width)), "assign.writemask"); } if(ir->condition) rhs = bld.CreateSelect(llvm_value(ir->condition), rhs, bld.CreateLoad(lhs), "assign.conditional"); bld.CreateStore(rhs, lhs); } virtual void visit(class ir_variable * var) { llvm_variable(var); } virtual void visit(ir_function_signature *sig) { if(!sig->is_defined) return; assert(!fun); fun = llvm_function(sig); bb = llvm::BasicBlock::Create(ctx, "entry", fun); bld.SetInsertPoint(bb); llvm::Function::arg_iterator ai = fun->arg_begin(); if (!strcmp("main",sig->function_name())) { assert(3 == fun->arg_size()); bld.CreateStore(ai, inputsPtr); inputs = ai; ai++; bld.CreateStore(ai, outputsPtr); outputs = ai; ai++; bld.CreateStore(ai, constantsPtr); constants = ai; ai++; } else { foreach_iter(exec_list_iterator, iter, sig->parameters) { ir_variable* arg = (ir_variable*)iter.get(); ai->setName(arg->name); bld.CreateStore(ai, llvm_variable(arg)); ++ai; } inputs = bld.CreateLoad(inputsPtr); outputs = bld.CreateLoad(outputsPtr); constants = bld.CreateLoad(constantsPtr); } inputs->setName("gl_inputs"); outputs->setName("gl_outputs"); constants->setName("gl_constants"); foreach_iter(exec_list_iterator, iter, sig->body) { ir_instruction *ir = (ir_instruction *)iter.get(); ir->accept(this); } if(fun->getReturnType()->isVoidTy()) bld.CreateRetVoid(); else bld.CreateRet(llvm::UndefValue::get(fun->getReturnType())); bb = NULL; fun = NULL; } virtual void visit(class ir_function * funs) { foreach_iter(exec_list_iterator, iter, *funs) { ir_function_signature* sig = (ir_function_signature*)iter.get(); sig->accept(this); } } }; struct llvm::Module * glsl_ir_to_llvm_module(struct exec_list *ir, llvm::Module * mod, const struct GGLState * gglCtx, const char * shaderSuffix) { ir_to_llvm_visitor v(mod, gglCtx, shaderSuffix); visit_exec_list(ir, &v); // mod->dump(); if(llvm::verifyModule(*mod, llvm::PrintMessageAction, 0)) { puts("**\n module verification failed **\n"); mod->dump(); assert(0); return NULL; } return mod; //v.ir_to_llvm_emit_op1(NULL, OPCODE_END, ir_to_llvm_undef_dst, ir_to_llvm_undef); }