Mercurial > octave-nkf
diff libinterp/corefcn/jit-typeinfo.cc @ 20654:b65888ec820e draft default tip gccjit
dmalcom gcc jit import
author | Stefan Mahr <dac922@gmx.de> |
---|---|
date | Fri, 27 Feb 2015 16:59:36 +0100 |
parents | d35201e5ce5d |
children |
line wrap: on
line diff
--- a/libinterp/corefcn/jit-typeinfo.cc Tue Oct 13 11:40:05 2015 +0100 +++ b/libinterp/corefcn/jit-typeinfo.cc Fri Feb 27 16:59:36 2015 +0100 @@ -504,9 +504,22 @@ } // -------------------- jit_type -------------------- -jit_type::jit_type (const std::string& aname, jit_type *aparent, - llvm::Type *allvm_type, bool askip_paren, int aid) : - mname (aname), mparent (aparent), llvm_type (allvm_type), mid (aid), +jit_type::jit_type (const std::string& aname, jit_type *aparent +#ifdef HAVE_LLVM + , llvm::Type *allvm_type +#endif +#ifdef HAVE_GCCJIT + , gccjit::type agcc_type +#endif + , bool askip_paren, int aid) : + mname (aname), mparent (aparent) +#ifdef HAVE_LLVM + , llvm_type (allvm_type) +#endif +#ifdef HAVE_GCCJIT + , gccjit_type (agcc_type) +#endif + , mid (aid), mdepth (aparent ? aparent->mdepth + 1 : 0), mskip_paren (askip_paren) { std::memset (msret, 0, sizeof (msret)); @@ -525,14 +538,21 @@ } // -------------------- jit_function -------------------- -jit_function::jit_function () : module (0), llvm_function (0), mresult (0), +jit_function::jit_function () : module (0), llvm_function (0), +#ifdef HAVE_GCCJIT + gccjit_function (), +#endif + mresult (0), call_conv (jit_convention::length), mcan_error (false) {} jit_function::jit_function (llvm::Module *amodule, +#ifdef HAVE_GCCJIT + gccjit::context gccjit_ctxt, +#endif jit_convention::type acall_conv, - const llvm::Twine& aname, jit_type *aresult, + std::string aname, jit_type *aresult, const std::vector<jit_type *>& aargs) : module (amodule), mresult (aresult), args (aargs), call_conv (acall_conv), mcan_error (false) @@ -586,17 +606,75 @@ #else llvm_function->addFnAttr (llvm::Attribute::AlwaysInline); #endif + +#ifdef HAVE_GCCJIT + if (gccjit_ctxt.get_inner_context ()) + { + std::vector<gccjit::param> gccjit_params; + for (int i = 0; i < args.size (); i++) + { + jit_type *ty = args[i]; + assert (ty); + + gccjit::type argty = ty->to_gccjit (); + if (ty->pointer_arg (call_conv)) + argty = argty.get_pointer (); + + std::stringstream paramname; + paramname << "arg" << i; + + gccjit::param param = gccjit_ctxt.new_param (argty, + paramname.str ()); + gccjit_params.push_back (param); + } + gccjit::type gccjit_return_type; + if (aresult) + gccjit_return_type = aresult->to_gccjit (); + else + gccjit_return_type = gccjit_ctxt.get_type (GCC_JIT_TYPE_VOID); + + enum gcc_jit_function_kind kind; + if (acall_conv == jit_convention::external) + kind = GCC_JIT_FUNCTION_IMPORTED; + else + { + if (0) + /* Doing this is correct, but makes the dump a little harder + to read. + Also, not fully implemented yet in libgccjit. */ + kind = GCC_JIT_FUNCTION_ALWAYS_INLINE; + else + kind = GCC_JIT_FUNCTION_EXPORTED; + } + gccjit_function = + gccjit_ctxt.new_function (kind, + gccjit_return_type, + aname, + gccjit_params, + 0); + } +#endif } jit_function::jit_function (const jit_function& fn, jit_type *aresult, const std::vector<jit_type *>& aargs) - : module (fn.module), llvm_function (fn.llvm_function), mresult (aresult), + : module (fn.module), + llvm_function (fn.llvm_function), +#ifdef HAVE_GCCJIT + gccjit_function (fn.gccjit_function), +#endif + mresult (aresult), args (aargs), call_conv (fn.call_conv), mcan_error (fn.mcan_error) { } jit_function::jit_function (const jit_function& fn) - : module (fn.module), llvm_function (fn.llvm_function), mresult (fn.mresult), + : module (fn.module), + llvm_function (fn.llvm_function), +#ifdef HAVE_GCCJIT + gccjit_function (fn.gccjit_function), +#endif + mresult (fn.mresult), args (fn.args), call_conv (fn.call_conv), mcan_error (fn.mcan_error) {} @@ -754,8 +832,76 @@ if (verify) llvm::verifyFunction (*llvm_function); + + if (0) + { + std::cout << "-------------------- llvm ir (at do_return) --------------------"; + std::cout << *llvm_function << std::endl; + } } +#ifdef HAVE_GCCJIT +#if 1 +gccjit::rvalue +jit_function::call (gccjit::context ctxt, + gccjit::block block, + const std::vector<jit_value *>& in_args) const +{ + assert (in_args.size () == args.size ()); + std::vector<gccjit::rvalue> gccjit_args (args.size ()); + for (size_t i = 0; i < in_args.size (); ++i) + gccjit_args[i] = in_args[i]->as_rvalue (); + return call (ctxt, block, gccjit_args); +} + +gccjit::rvalue +jit_function::call (gccjit::context ctxt, + gccjit::block block, + std::vector<gccjit::rvalue>& in_args) const +{ + assert (in_args.size () == args.size ()); + +#if 1 + std::vector<gccjit::rvalue> packed_args (in_args.size ()); + + for (size_t i = 0; i < in_args.size (); ++i) + { + gccjit::rvalue arg = in_args[i]; +#if 0 + jit_type::convert_fn convert = args[i]->pack (call_conv); + if (convert) + arg = convert (builder, arg); +#endif + + if (args[i]->pointer_arg (call_conv)) + { + // The LLVM implementation takes a copy using alloca, passing + // a ptr to the copy to the fn. Emulate this behavior. + gccjit::lvalue tmp = + block.get_function ().new_local (arg.get_type (), "tmp"); + block.add_assignment (tmp, arg); + arg = tmp.get_address (); + } + + packed_args[i] = arg; + } + + return ctxt.new_call (gccjit_function, + packed_args); +#else + return ctxt.new_call (gccjit_function, + in_args); +#endif +} +#endif + +gccjit::lvalue +jit_function::argument (size_t idx) const +{ + return gccjit_function.get_param (idx); +} +#endif + void jit_function::do_add_mapping (llvm::ExecutionEngine *engine, void *fn) { @@ -928,6 +1074,7 @@ return 0; } +#ifdef HAVE_LLVM llvm::Value * jit_index_operation::create_arg_array (llvm::IRBuilderD& builder, const jit_function &fn, size_t start_idx, @@ -947,6 +1094,28 @@ builder.CreateStore (array, array_mem); return builder.CreateBitCast (array_mem, scalar_t->getPointerTo ()); } +#endif // #ifdef HAVE_LLVM + +#ifdef HAVE_GCCJIT +gccjit::rvalue +jit_index_operation::create_arg_array (const jit_function &fn, + gccjit::block block, + size_t start_idx, + size_t end_idx) const +{ + size_t n = end_idx - start_idx; + gccjit::type scalar_t = jit_typeinfo::get_scalar_gccjit (); + gccjit::type array_t = block.get_context ().new_array_type (scalar_t, n); + gccjit::lvalue array = + block.get_function ().new_local (array_t, "tmp_array"); + for (size_t i = start_idx; i < end_idx; ++i) + { + gccjit::rvalue idx = fn.argument (i); + block.add_assignment (array[i - start_idx], idx); + } + return array.get_address ().cast_to (scalar_t.get_pointer ()); +} +#endif // #ifdef HAVE_GCCJIT // -------------------- jit_paren_subsref -------------------- jit_function * @@ -956,7 +1125,11 @@ ss << "jit_paren_subsref_matrix_scalar" << (types.size () - 1); jit_type *scalar = jit_typeinfo::get_scalar (); - jit_function *fn = new jit_function (module, jit_convention::internal, + jit_function *fn = new jit_function (module, +#ifdef HAVE_GCCJIT + gccjit_ctxt, +#endif + jit_convention::internal, ss.str (), scalar, types); fn->mark_can_error (); llvm::BasicBlock *body = fn->new_block (); @@ -969,6 +1142,26 @@ llvm::Value *mat = fn->argument (builder, 0); llvm::Value *ret = paren_scalar.call (builder, mat, array, nelem); fn->do_return (builder, ret); + +#ifdef HAVE_GCCJIT + // gcc implementation + { + gccjit::function gf = fn->gccjit_function; + gccjit::block body = gf.new_block (); + gccjit::rvalue array = create_arg_array (*fn, body, 1, types.size ()); + jit_type *index = jit_typeinfo::get_index (); + gccjit::rvalue nelem = gccjit_ctxt.new_rvalue (index->to_gccjit (), + (int)types.size () - 1); + gccjit::rvalue mat = gf.get_param (0); + std::vector<gccjit::rvalue> args (3); + args[0] = mat; + args[1] = array; + args[2] = nelem; + gccjit::rvalue ret = paren_scalar.call (gccjit_ctxt, body, args); + body.end_with_return (ret); + } +#endif + return fn; } @@ -981,7 +1174,11 @@ types[2] = jit_typeinfo::get_index (); jit_type *scalar = jit_typeinfo::get_scalar (); - paren_scalar = jit_function (module, jit_convention::external, + paren_scalar = jit_function (module, +#ifdef HAVE_GCCJIT + gccjit_ctxt, +#endif + jit_convention::external, "octave_jit_paren_scalar", scalar, types); paren_scalar.add_mapping (engine, &octave_jit_paren_scalar); paren_scalar.mark_can_error (); @@ -995,7 +1192,11 @@ ss << "jit_paren_subsasgn_matrix_scalar" << (types.size () - 2); jit_type *matrix = jit_typeinfo::get_matrix (); - jit_function *fn = new jit_function (module, jit_convention::internal, + jit_function *fn = new jit_function (module, +#ifdef HAVE_GCCJIT + gccjit_ctxt, +#endif + jit_convention::internal, ss.str (), matrix, types); fn->mark_can_error (); llvm::BasicBlock *body = fn->new_block (); @@ -1010,6 +1211,28 @@ llvm::Value *value = fn->argument (builder, types.size () - 1); llvm::Value *ret = paren_scalar.call (builder, mat, array, nelem, value); fn->do_return (builder, ret); + +#ifdef HAVE_GCCJIT + { + // FIXME: TODO + gccjit::function gf = fn->gccjit_function; + gccjit::block body = gf.new_block (); + gccjit::rvalue array = create_arg_array (*fn, body, 1, types.size () - 1); + jit_type *index = jit_typeinfo::get_index (); + gccjit::rvalue nelem = gccjit_ctxt.new_rvalue (index->to_gccjit (), + (int)types.size () - 2); + gccjit::rvalue mat = gf.get_param (0); + gccjit::rvalue value = gf.get_param (types.size () - 1); + std::vector<gccjit::rvalue> args(4); + args[0] = mat; + args[1] = array; + args[2] = nelem; + args[3] = value; + gccjit::rvalue ret = paren_scalar.call (gccjit_ctxt, body, args); + body.end_with_return (ret); + } +#endif + return fn; } @@ -1026,7 +1249,11 @@ types[2] = jit_typeinfo::get_index (); types[3] = jit_typeinfo::get_scalar (); - paren_scalar = jit_function (module, jit_convention::external, + paren_scalar = jit_function (module, +#ifdef HAVE_GCCJIT + gccjit_ctxt, +#endif + jit_convention::external, "octave_jit_paren_scalar", matrix, types); paren_scalar.add_mapping (engine, &octave_jit_paren_scalar_subsasgn); paren_scalar.mark_can_error (); @@ -1048,7 +1275,12 @@ { instance = this; +#ifdef HAVE_GCCJIT + gccjit_ctxt = gccjit::context::acquire (); +#endif + // FIXME: We should be registering types like in octave_value_typeinfo +#ifdef HAVE_LLVM llvm::Type *any_t = llvm::StructType::create (context, "octave_base_value"); any_t = any_t->getPointerTo (); @@ -1058,14 +1290,54 @@ string_t = string_t->getPointerTo (); llvm::Type *index_t = llvm::Type::getIntNTy (context, sizeof(octave_idx_type) * 8); - +#endif + +#ifdef HAVE_GCCJIT + gccjit::type any_t_gcc = + gccjit_ctxt.new_opaque_struct_type ("octave_base_value"); + any_t_gcc = any_t_gcc.get_pointer (); + gccjit::type scalar_t_gcc = gccjit_ctxt.get_type (GCC_JIT_TYPE_DOUBLE); + gccjit::type bool_t_gcc = gccjit_ctxt.get_type (GCC_JIT_TYPE_BOOL); + gccjit::type string_t_gcc = + gccjit_ctxt.get_type (GCC_JIT_TYPE_CHAR).get_pointer (); + gccjit::type index_t_gcc = gccjit_ctxt.get_int_type <octave_idx_type> (); + gccjit::type int_t_gcc = gccjit_ctxt.get_type (GCC_JIT_TYPE_INT); +#endif + +#ifdef HAVE_LLVM llvm::StructType *range_t = llvm::StructType::create (context, "range"); std::vector<llvm::Type *> range_contents (4, scalar_t); range_contents[3] = index_t; range_t->setBody (range_contents); - +#endif +#ifdef HAVE_GCCJIT + field_rng_base = gccjit_ctxt.new_field (scalar_t_gcc, "rng_base"); + field_rng_limit = gccjit_ctxt.new_field (scalar_t_gcc, "rng_limit"); + field_rng_inc = gccjit_ctxt.new_field (scalar_t_gcc, "rng_inc"); + field_rng_nelem = gccjit_ctxt.new_field (index_t_gcc, "rng_nelem"); + /* FIXME: what about the "mutable Matrix cache;" */ + + std::vector<gccjit::field> range_fields (4); + range_fields [0] = field_rng_base; + range_fields [1] = field_rng_limit; + range_fields [2] = field_rng_inc; + range_fields [3] = field_rng_nelem; + + gccjit::type range_t_gcc = + gccjit_ctxt.new_struct_type ( + "range", + range_fields, + gccjit_ctxt.new_location ("liboctave/array/Range.h", 33, 0)); +#endif + +#ifdef HAVE_LLVM llvm::Type *refcount_t = llvm::Type::getIntNTy (context, sizeof(int) * 8); - +#endif +#ifdef HAVE_GCCJIT + gccjit::type refcount_t_gcc = gccjit_ctxt.get_type (GCC_JIT_TYPE_INT); +#endif + +#ifdef HAVE_LLVM llvm::StructType *matrix_t = llvm::StructType::create (context, "matrix"); llvm::Type *matrix_contents[5]; matrix_contents[0] = refcount_t->getPointerTo (); @@ -1074,8 +1346,58 @@ matrix_contents[3] = index_t->getPointerTo (); matrix_contents[4] = string_t; matrix_t->setBody (llvm::makeArrayRef (matrix_contents, 5)); - +#endif +#ifdef HAVE_GCCJIT + /* + jit-typeinfo.h has, somewhat ominously: + // jit_array is compatable with the llvm array/matrix structures + typedef jit_array<NDArray, double> jit_matrix; + */ + gccjit::type matrix_t_gcc; + { + /* typedef jit_array<NDArray, double> jit_matrix; */ + /* template <typename T, typename U> struct jit_array {...}; */ + gccjit::type T = string_t_gcc; /* NDArray */ + gccjit::type U = scalar_t_gcc; /* double */ + + /* int *ref_count; */ + field_ref_count = + gccjit_ctxt.new_field (refcount_t_gcc.get_pointer (), "ref_count"); + + /* U *slice_data; */ + field_slice_data = + gccjit_ctxt.new_field (U.get_pointer (), "slice_data"); + + /* octave_idx_type slice_len; */ + field_slice_len = gccjit_ctxt.new_field (index_t_gcc, "slice_len"); + + /* octave_idx_type *dimensions; */ + field_dimensions = + gccjit_ctxt.new_field (index_t_gcc.get_pointer (), "dimensions"); + + /* T *array; */ + field_array = gccjit_ctxt.new_field (T.get_pointer (), "array"); + + std::vector<gccjit::field> matrix_fields (5); + matrix_fields[0] = field_ref_count; + matrix_fields[1] = field_slice_data; + matrix_fields[2] = field_slice_len; + matrix_fields[3] = field_dimensions; + matrix_fields[4] = field_array; + + matrix_t_gcc = gccjit_ctxt.new_struct_type ( + "jit_matrix", + matrix_fields, + gccjit_ctxt.new_location ("jit-typeinfo.h", 106, 0)); + } +#endif + +#ifdef HAVE_LLVM llvm::Type *complex_t = llvm::ArrayType::get (scalar_t, 2); +#endif +#ifdef HAVE_GCCJIT + gccjit::type complex_t_gcc = gccjit_ctxt.new_array_type (scalar_t_gcc, 2); +#endif // complex_ret is what is passed to C functions in order to get calling // convention right @@ -1089,6 +1411,20 @@ } // create types +#ifdef HAVE_GCCJIT + any = new_type ("any", 0, any_t, any_t_gcc); + matrix = new_type ("matrix", any, matrix_t, matrix_t_gcc); + complex = new_type ("complex", any, complex_t, complex_t_gcc); + scalar = new_type ("scalar", complex, scalar_t, scalar_t_gcc); + scalar_ptr = new_type ("scalar_ptr", 0, scalar_t->getPointerTo (), + scalar_t_gcc.get_pointer ()); + any_ptr = new_type ("any_ptr", 0, any_t->getPointerTo (), + any_t_gcc.get_pointer()); + range = new_type ("range", any, range_t, range_t_gcc); + string = new_type ("string", any, string_t, string_t_gcc); + boolean = new_type ("bool", any, bool_t, bool_t_gcc); + index = new_type ("index", any, index_t, index_t_gcc); +#else any = new_type ("any", 0, any_t); matrix = new_type ("matrix", any, matrix_t); complex = new_type ("complex", any, complex_t); @@ -1099,6 +1435,7 @@ string = new_type ("string", any, string_t); boolean = new_type ("bool", any, bool_t); index = new_type ("index", any, index_t); +#endif create_int (8); create_int (16); @@ -1124,10 +1461,16 @@ if (sizeof (void *) == 4) complex->mark_sret (jit_convention::external); +#ifdef HAVE_GCCJIT + paren_subsref_fn.initialize (module, engine, gccjit_ctxt); + paren_subsasgn_fn.initialize (module, engine, gccjit_ctxt); +#else paren_subsref_fn.initialize (module, engine); paren_subsasgn_fn.initialize (module, engine); +#endif // bind global variables +#ifdef HAVE_LLVM lerror_state = new llvm::GlobalVariable (*module, bool_t, false, llvm::GlobalValue::ExternalLinkage, 0, "error_state"); @@ -1142,6 +1485,20 @@ "octave_interrupt_state"); engine->addGlobalMapping (loctave_interrupt_state, reinterpret_cast<void *> (&octave_interrupt_state)); +#endif + +#ifdef HAVE_GCCJIT + // Access "error_state" (actually declared as an "int"), by + // taking its address and dereferencing: + error_state_gccjit = + *gccjit_ctxt.new_rvalue (int_t_gcc.get_pointer (), + &error_state); + sig_atomic_type_gccjit = gccjit_ctxt.get_int_type <sig_atomic_t> (); + octave_interrupt_state_gccjit = + *gccjit_ctxt.new_rvalue ( + sig_atomic_type_gccjit.get_volatile ().get_pointer (), + &octave_interrupt_state); +#endif // generic call function { @@ -1175,8 +1532,10 @@ for (int op = 0; op < octave_value::num_binary_ops; ++op) { - llvm::Twine fn_name ("octave_jit_binary_any_any_"); - fn_name = fn_name + llvm::Twine (op); + std::string fn_name ("octave_jit_binary_any_any_"); + char buf[64]; + snprintf (buf, sizeof(buf), "%i", op); + fn_name = fn_name + std::string (buf); fn = create_internal (fn_name, any, any, any); fn.mark_can_error (); @@ -1189,6 +1548,20 @@ fn.argument (builder, 0), fn.argument (builder, 1)); fn.do_return (builder, ret); + +#ifdef HAVE_GCCJIT + // gccjit implementation of octave_jit_binary_any_any_*: + gccjit::function gf = fn.gccjit_function; + gccjit::block gccjit_block = gf.new_block (); + gccjit_block.end_with_return ( + gccjit_block.add_call ( + any_binary.gccjit_function, + gccjit_ctxt.new_rvalue (binary_op_type->to_gccjit (), + op), + gf.get_param (0), + gf.get_param (1))); +#endif + binary_ops[op].add_overload (fn); } @@ -1220,9 +1593,40 @@ destroy_fn.add_overload (create_identity(index)); destroy_fn.add_overload (create_identity(complex)); - // -------------------- scalar related operations -------------------- - // now for binary scalar operations +#ifdef HAVE_GCCJIT + add_binary_op (scalar, octave_value::op_add, + llvm::Instruction::FAdd, + GCC_JIT_BINARY_OP_PLUS); + add_binary_op (scalar, octave_value::op_sub, + llvm::Instruction::FSub, + GCC_JIT_BINARY_OP_MINUS); + add_binary_op (scalar, octave_value::op_mul, + llvm::Instruction::FMul, + GCC_JIT_BINARY_OP_MULT); + add_binary_op (scalar, octave_value::op_el_mul, + llvm::Instruction::FMul, + GCC_JIT_BINARY_OP_MULT); + + add_binary_fcmp (scalar, octave_value::op_lt, + llvm::CmpInst::FCMP_ULT, + GCC_JIT_COMPARISON_LT); + add_binary_fcmp (scalar, octave_value::op_le, + llvm::CmpInst::FCMP_ULE, + GCC_JIT_COMPARISON_LE); + add_binary_fcmp (scalar, octave_value::op_eq, + llvm::CmpInst::FCMP_UEQ, + GCC_JIT_COMPARISON_EQ); + add_binary_fcmp (scalar, octave_value::op_ge, + llvm::CmpInst::FCMP_UGE, + GCC_JIT_COMPARISON_GE); + add_binary_fcmp (scalar, octave_value::op_gt, + llvm::CmpInst::FCMP_UGT, + GCC_JIT_COMPARISON_GT); + add_binary_fcmp (scalar, octave_value::op_ne, + llvm::CmpInst::FCMP_UNE, + GCC_JIT_COMPARISON_NE); +#else add_binary_op (scalar, octave_value::op_add, llvm::Instruction::FAdd); add_binary_op (scalar, octave_value::op_sub, llvm::Instruction::FSub); add_binary_op (scalar, octave_value::op_mul, llvm::Instruction::FMul); @@ -1234,6 +1638,7 @@ add_binary_fcmp (scalar, octave_value::op_ge, llvm::CmpInst::FCMP_UGE); add_binary_fcmp (scalar, octave_value::op_gt, llvm::CmpInst::FCMP_UGT); add_binary_fcmp (scalar, octave_value::op_ne, llvm::CmpInst::FCMP_UNE); +#endif jit_function gripe_div0 = create_external (JIT_FN (gripe_divide_by_zero), 0); gripe_div0.mark_can_error (); @@ -1261,6 +1666,25 @@ fn.argument (builder, 1)); fn.do_return (builder, ret); } + + // gccjit implementation of octave_jit_div_scalar_scalar: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block initial_block = gf.new_block ("initial"); + gccjit::block warn_block = gf.new_block ("warn"); + gccjit::block normal_block = gf.new_block ("normal"); + initial_block.end_with_conditional ( + gf.get_param (1) != scalar_t_gcc.zero (), + normal_block, // on_true + warn_block); // on_false + + warn_block.add_call (gripe_div0.gccjit_function); + warn_block.end_with_jump (normal_block); + + normal_block.end_with_return (gf.get_param (0) / gf.get_param (1)); + } +#endif binary_ops[octave_value::op_div].add_overload (fn); binary_ops[octave_value::op_el_div].add_overload (fn); @@ -1279,7 +1703,7 @@ // now for unary scalar operations // FIXME: Impelment not - fn = create_internal ("octave_jit_++", scalar, scalar); + fn = create_internal ("octave_jit_plusplus", scalar, scalar); body = fn.new_block (); builder.SetInsertPoint (body); { @@ -1288,9 +1712,17 @@ val = builder.CreateFAdd (val, one); fn.do_return (builder, val); } + // gccjit implementation of octave_jit_plusplus: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return (gf.get_param (0) + scalar_t_gcc.one ()); + } +#endif unary_ops[octave_value::op_incr].add_overload (fn); - fn = create_internal ("octave_jit_--", scalar, scalar); + fn = create_internal ("octave_jit_minusminus", scalar, scalar); body = fn.new_block (); builder.SetInsertPoint (body); { @@ -1298,7 +1730,15 @@ llvm::Value *val = fn.argument (builder, 0); val = builder.CreateFSub (val, one); fn.do_return (builder, val); + + // gccjit implementation of octave_jit_minusminus: +#ifdef HAVE_GCCJIT + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return (gf.get_param (0) - scalar_t_gcc.one ()); +#endif } + unary_ops[octave_value::op_decr].add_overload (fn); fn = create_internal ("octave_jit_uminus", scalar, scalar); @@ -1310,7 +1750,14 @@ val = builder.CreateFMul (val, mone); fn.do_return (builder, val); } - unary_ops[octave_value::op_uminus].add_overload (fn); + // gccjit implementation of octave_jit_uminus: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return (-gf.get_param (0)); + } +#endif fn = create_identity (scalar); unary_ops[octave_value::op_uplus].add_overload (fn); @@ -1318,7 +1765,7 @@ unary_ops[octave_value::op_hermitian].add_overload (fn); // now for binary complex operations - fn = create_internal ("octave_jit_+_complex_complex", complex, complex, + fn = create_internal ("octave_jit_plus_complex_complex", complex, complex, complex); body = fn.new_block (); builder.SetInsertPoint (body); @@ -1331,9 +1778,21 @@ complex_imag (rhs)); fn.do_return (builder, complex_new (real, imag)); } + // gccjit implementation of octave_jit_plus_complex_complex: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::rvalue lhs = gf.get_param (0); + gccjit::rvalue rhs = gf.get_param (1); + gccjit::rvalue real = complex_real (lhs) + complex_real (rhs); + gccjit::rvalue imag = complex_imag (lhs) + complex_imag (rhs); + b.end_with_return (complex_new (b, real, imag)); + } +#endif binary_ops[octave_value::op_add].add_overload (fn); - fn = create_internal ("octave_jit_-_complex_complex", complex, complex, + fn = create_internal ("octave_jit_minus_complex_complex", complex, complex, complex); body = fn.new_block (); builder.SetInsertPoint (body); @@ -1346,6 +1805,18 @@ complex_imag (rhs)); fn.do_return (builder, complex_new (real, imag)); } + // gccjit implementation of octave_jit_minus_complex_complex: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::rvalue lhs = gf.get_param (0); + gccjit::rvalue rhs = gf.get_param (1); + gccjit::rvalue real = complex_real (lhs) - complex_real (rhs); + gccjit::rvalue imag = complex_imag (lhs) - complex_imag (rhs); + b.end_with_return (complex_new (b, real, imag)); + } +#endif binary_ops[octave_value::op_sub].add_overload (fn); fn = create_external (JIT_FN (octave_jit_complex_mul), @@ -1364,7 +1835,7 @@ binary_ops[octave_value::op_pow].add_overload (fn); binary_ops[octave_value::op_el_pow].add_overload (fn); - fn = create_internal ("octave_jit_*_scalar_complex", complex, scalar, + fn = create_internal ("octave_jit_mult_scalar_complex", complex, scalar, complex); jit_function mul_scalar_complex = fn; body = fn.new_block (); @@ -1391,6 +1862,31 @@ builder.CreateFMul (lhs, complex_imag (rhs))); fn.do_return (builder, temp); } + // gccjit implementation of octave_jit_mult_scalar_complex: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + + gccjit::rvalue fzero = scalar_t_gcc.zero (); + gccjit::rvalue lhs = gf.get_param (0); + gccjit::rvalue rhs = gf.get_param (1); + + gccjit::block initial = gf.new_block ("initial"); + gccjit::block complex_mul = gf.new_block ("complex_mul"); + gccjit::block scalar_mul = gf.new_block ("scalar_mul"); + + initial.end_with_conditional (complex_imag (rhs) == fzero, + scalar_mul, + complex_mul); + + scalar_mul.end_with_return (complex_new (scalar_mul, + lhs * complex_real (rhs), + fzero)); + complex_mul.end_with_return (complex_new (complex_mul, + lhs * complex_real (rhs), + lhs * complex_imag (rhs))); + } +#endif binary_ops[octave_value::op_mul].add_overload (fn); binary_ops[octave_value::op_el_mul].add_overload (fn); @@ -1399,7 +1895,7 @@ binary_ops[octave_value::op_mul].add_overload (fn); binary_ops[octave_value::op_el_mul].add_overload (fn); - fn = create_internal ("octave_jit_+_scalar_complex", complex, scalar, + fn = create_internal ("octave_jit_plus_scalar_complex", complex, scalar, complex); body = fn.new_block (); builder.SetInsertPoint (body); @@ -1409,12 +1905,23 @@ llvm::Value *real = builder.CreateFAdd (lhs, complex_real (rhs)); fn.do_return (builder, complex_real (rhs, real)); } + // gccjit implementation of octave_jit_plus_scalar_complex: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::rvalue lhs = gf.get_param (0); + gccjit::lvalue rhs = gf.get_param (1); + gccjit::rvalue real = lhs + complex_real (rhs); + b.end_with_return (complex_real (b, rhs, real)); + } +#endif binary_ops[octave_value::op_add].add_overload (fn); fn = mirror_binary (fn); binary_ops[octave_value::op_add].add_overload (fn); - fn = create_internal ("octave_jit_-_complex_scalar", complex, complex, + fn = create_internal ("octave_jit_minus_complex_scalar", complex, complex, scalar); body = fn.new_block (); builder.SetInsertPoint (body); @@ -1424,9 +1931,20 @@ llvm::Value *real = builder.CreateFSub (complex_real (lhs), rhs); fn.do_return (builder, complex_real (lhs, real)); } + // gccjit implementation of octave_jit_minus_complex_scalar: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::lvalue lhs = gf.get_param (0); + gccjit::rvalue rhs = gf.get_param (1); + gccjit::rvalue real = complex_real (lhs) - rhs; + b.end_with_return (complex_real (b, lhs, real)); + } +#endif binary_ops[octave_value::op_sub].add_overload (fn); - fn = create_internal ("octave_jit_-_scalar_complex", complex, scalar, + fn = create_internal ("octave_jit_minus_scalar_complex", complex, scalar, complex); body = fn.new_block (); builder.SetInsertPoint (body); @@ -1436,6 +1954,17 @@ llvm::Value *real = builder.CreateFSub (lhs, complex_real (rhs)); fn.do_return (builder, complex_real (rhs, real)); } + // gccjit implementation of octave_jit_minus_scalar_complex: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::rvalue lhs = gf.get_param (0); + gccjit::lvalue rhs = gf.get_param (1); + gccjit::rvalue real = lhs - complex_real (rhs); + b.end_with_return (complex_real (b, rhs, real)); + } +#endif binary_ops[octave_value::op_sub].add_overload (fn); fn = create_external (JIT_FN (octave_jit_pow_scalar_complex), complex, scalar, @@ -1449,11 +1978,25 @@ binary_ops[octave_value::op_el_pow].add_overload (fn); // now for binary index operators +#ifdef HAVE_GCCJIT + add_binary_op (index, octave_value::op_add, + llvm::Instruction::Add, + GCC_JIT_BINARY_OP_PLUS); + + // and binary bool operators + add_binary_op (boolean, octave_value::op_el_or, + llvm::Instruction::Or, + GCC_JIT_BINARY_OP_LOGICAL_OR); + add_binary_op (boolean, octave_value::op_el_and, + llvm::Instruction::And, + GCC_JIT_BINARY_OP_LOGICAL_AND); +#else add_binary_op (index, octave_value::op_add, llvm::Instruction::Add); // and binary bool operators add_binary_op (boolean, octave_value::op_el_or, llvm::Instruction::Or); add_binary_op (boolean, octave_value::op_el_and, llvm::Instruction::And); +#endif // now for printing functions print_fn.stash_name ("print"); @@ -1470,6 +2013,13 @@ llvm::Value *zero = llvm::ConstantInt::get (index_t, 0); fn.do_return (builder, zero); } +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return (gccjit_ctxt.zero (index_t_gcc)); + } +#endif for_init_fn.add_overload (fn); // bounds check for for loop @@ -1484,7 +2034,20 @@ llvm::Value *idx = fn.argument (builder, 1); llvm::Value *ret = builder.CreateICmpULT (idx, nelem); fn.do_return (builder, ret); + } +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + // FIXME: is "range" a (struct range) or a (struct range *) ? + gccjit::rvalue nelem + = gf.get_param (0).access_field (field_rng_nelem); + gccjit::rvalue idx = gf.get_param (1); + gccjit::rvalue ret = idx < nelem; + gccjit::block b = gf.new_block (); + b.end_with_return (ret); + } +#endif for_check_fn.add_overload (fn); // index variabe for for loop @@ -1504,6 +2067,20 @@ ret = builder.CreateFAdd (base, ret); fn.do_return (builder, ret); } +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::rvalue idx = gf.get_param (1); + gccjit::rvalue didx = idx.cast_to (scalar_t_gcc); + gccjit::rvalue rng = gf.get_param (0); + gccjit::rvalue base = rng.access_field (field_rng_base); + gccjit::rvalue inc = rng.access_field (field_rng_inc); + gccjit::rvalue ret = didx * inc; + ret = base + ret; + b.end_with_return (ret); + } +#endif for_index_fn.add_overload (fn); // logically true @@ -1535,6 +2112,22 @@ llvm::Value *ret = builder.CreateFCmpONE (fn.argument (builder, 0), zero); fn.do_return (builder, ret); } +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block initial = gf.new_block ("initial"); + gccjit::block error = gf.new_block ("error"); + gccjit::block normal = gf.new_block ("normal"); + initial.add_comment ("check for NaN"); + initial.end_with_conditional (gf.get_param (0) != gf.get_param (0), + error, normal); + + error.add_call (gripe_nantl.gccjit_function); + error.end_with_jump (normal); + + normal.end_with_return ( gf.get_param (0) != scalar_t_gcc.zero ()); + } +#endif logically_true_fn.add_overload (fn); // logically_true boolean @@ -1568,6 +2161,22 @@ rng = builder.CreateInsertValue (rng, nelem, 3); fn.do_return (builder, rng); } +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::rvalue base = gf.get_param (0); + gccjit::rvalue limit = gf.get_param (1); + gccjit::rvalue inc = gf.get_param (2); + gccjit::block b = gf.new_block (); + gccjit::rvalue nelem = compute_nelem.gccjit_function (base, limit, inc); + gccjit::lvalue rng = gf.new_local (range_t_gcc, "rng"); + b.add_assignment (rng.access_field (field_rng_base), base); + b.add_assignment (rng.access_field (field_rng_limit), limit); + b.add_assignment (rng.access_field (field_rng_inc), inc); + b.add_assignment (rng.access_field (field_rng_nelem), nelem); + b.end_with_return (rng); + } +#endif make_range_fn.add_overload (fn); // paren_subsref @@ -1579,7 +2188,7 @@ 0, jit_int, jit_int, index, index); - fn = create_internal ("()subsref", scalar, matrix, scalar); + fn = create_internal ("subsref", scalar, matrix, scalar); fn.mark_can_error (); body = fn.new_block (); @@ -1638,10 +2247,55 @@ merge->addIncoming (ret, success); fn.do_return (builder, merge); } + // gccjit implementation of subsref: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block initial = gf.new_block ("initial"); + + gccjit::rvalue one_idx = index_t_gcc.one (); + gccjit::rvalue one_int = jit_int->to_gccjit ().one (); + + gccjit::rvalue mat = gf.get_param (0); + gccjit::rvalue idx = gf.get_param (1); // scalar + + // convert index to scalar to integer, and check index >= 1 + gccjit::rvalue int_idx = idx.cast_to (index_t_gcc); + gccjit::rvalue check_idx = int_idx.cast_to (scalar_t_gcc); + gccjit::rvalue cond0 = (idx != check_idx); + gccjit::rvalue cond1 = (int_idx < one_idx); + + gccjit::block conv_error = gf.new_block ("conv_error"); + gccjit::block normal = gf.new_block ("normal"); + initial.end_with_conditional (cond0 || cond1, conv_error, normal); + + // "conv_error" block: + conv_error.add_call (ginvalid_index.gccjit_function); + conv_error.end_with_return (scalar_t_gcc.zero ()); // dummy value + + // "normal" block: + gccjit::rvalue len = mat.access_field (field_slice_len); + gccjit::rvalue cond = (int_idx > len); + gccjit::block bounds_error = gf.new_block ("bounds_error"); + gccjit::block success = gf.new_block ("success"); + normal.end_with_conditional (cond, bounds_error, success); + + // "bounds_error" block: + bounds_error.add_call (gindex_range.gccjit_function, + one_int, one_int, int_idx, len); + bounds_error.end_with_return (scalar_t_gcc.zero ()); // dummy value + + // "success" block: + gccjit::rvalue data = mat.access_field (field_slice_data); + gccjit::rvalue gep = data[int_idx]; + gccjit::rvalue ret = gep; + success.end_with_return (ret); + } +#endif paren_subsref_fn.add_overload (fn); // paren subsasgn - paren_subsasgn_fn.stash_name ("()subsasgn"); + paren_subsasgn_fn.stash_name ("subsasgn"); jit_function resize_paren_subsasgn = create_external (JIT_FN (octave_jit_paren_subsasgn_impl), matrix, matrix, @@ -1710,6 +2364,61 @@ merge->addIncoming (mat, success); fn.do_return (builder, merge); } + // gccjit implementation of octave_jit_paren_subsasgn: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block initial = gf.new_block ("initial"); + gccjit::rvalue one_idx = index_t_gcc.one (); + gccjit::rvalue one_int = jit_int->to_gccjit ().one (); + + gccjit::lvalue mat = gf.get_param (0); + gccjit::rvalue idx = gf.get_param (1); + gccjit::rvalue value = gf.get_param (2); + gccjit::rvalue int_idx = idx.cast_to (index_t_gcc); + gccjit::rvalue check_idx = int_idx.cast_to (scalar_t_gcc); + gccjit::rvalue cond0 = (idx != check_idx); + gccjit::rvalue cond1 = (int_idx < one_idx); + gccjit::rvalue cond = (cond0 || cond1); + + gccjit::block conv_error = gf.new_block ("conv_error"); + gccjit::block normal = gf.new_block ("normal"); + initial.end_with_conditional (cond, conv_error, normal); + + // block: conv_error + conv_error.add_call (ginvalid_index.gccjit_function); + conv_error.end_with_return (mat); + + // block: normal + gccjit::rvalue len = mat.access_field (field_slice_len); + cond0 = (int_idx > len); + + gccjit::rvalue rcount = mat.access_field (field_ref_count); + rcount = rcount.dereference (); + cond1 = rcount > one_int; + cond = (cond0 || cond1); + + gccjit::block bounds_error = gf.new_block ("bounds_error"); + gccjit::block success = gf.new_block ("success"); + normal.end_with_conditional (cond, bounds_error, success); + + // block: bounds_error + // resize on out of bounds access + std::vector<gccjit::rvalue> args (3); + args[0] = mat; + args[1] = int_idx; + args[2] = value; + gccjit::rvalue resize_result = + resize_paren_subsasgn.call (gccjit_ctxt, bounds_error, args); + bounds_error.end_with_return (resize_result); + + // block: success + gccjit::rvalue data = mat.access_field (field_slice_data); + gccjit::lvalue gep = data[int_idx]; + success.add_assignment (gep, value); + success.end_with_return (mat); + } +#endif paren_subsasgn_fn.add_overload (fn); fn = create_external (JIT_FN (octave_jit_paren_subsasgn_matrix_range), matrix, @@ -1726,6 +2435,16 @@ llvm::Value *ret = builder.CreateExtractValue (mat, 2); fn.do_return (builder, builder.CreateSIToFP (ret, scalar_t)); } +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::rvalue mat = gf.get_param (0); + // FIXME: is this the right behavior? + gccjit::rvalue ret = mat.access_field (field_slice_len); + gccjit::block b = gf.new_block (); + b.end_with_return (ret.cast_to (scalar_t_gcc)); + } +#endif end1_fn.add_overload (fn); end_fn.stash_name ("end"); @@ -1784,6 +2503,15 @@ llvm::Value *zero = llvm::ConstantFP::get (scalar_t, 0); fn.do_return (builder, complex_new (fn.argument (builder, 0), zero)); } + // gccjit implementation of octave_jit_cast_complex_scalar: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + gccjit::rvalue zero = scalar_t_gcc.zero (); + b.end_with_return (complex_new (b, gf.get_param (0), zero)); + } +#endif casts[complex->type_id ()].add_overload (fn); // cast scalar <- complex @@ -1791,6 +2519,13 @@ body = fn.new_block (); builder.SetInsertPoint (body); fn.do_return (builder, complex_real (fn.argument (builder, 0))); +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return (complex_real (gf.get_param (0))); + } +#endif casts[scalar->type_id ()].add_overload (fn); // cast any <- any @@ -1818,7 +2553,9 @@ register_generic ("cos", matrix, matrix); add_builtin ("exp"); - register_intrinsic ("exp", llvm::Intrinsic::exp, scalar, scalar); + // FIXME: looks like a typo: "cos" here should be "exp": + // filed as http://savannah.gnu.org/bugs/index.php?41560 + register_intrinsic ("exp", llvm::Intrinsic::cos, scalar, scalar); register_generic ("exp", matrix, matrix); add_builtin ("balance"); @@ -1833,10 +2570,13 @@ add_builtin ("norm"); register_generic ("norm", scalar, matrix); + //FIXME: gccjit can't yet cope with duplicate names +#if !defined (HAVE_GCCJIT) add_builtin ("rand"); register_generic ("rand", matrix, scalar); register_generic ("rand", matrix, std::vector<jit_type *> (2, scalar)); + add_builtin ("magic"); register_generic ("magic", matrix, scalar); register_generic ("magic", matrix, std::vector<jit_type *> (2, scalar)); @@ -1844,6 +2584,7 @@ add_builtin ("eye"); register_generic ("eye", matrix, scalar); register_generic ("eye", matrix, std::vector<jit_type *> (2, scalar)); +#endif add_builtin ("mod"); register_generic ("mod", scalar, std::vector<jit_type *> (2, scalar)); @@ -1870,6 +2611,9 @@ casts[btype->type_id ()].add_overload (jit_function (any_id, btype, args)); } +#ifdef HAVE_GCCJIT + gccjit_ctxt.dump_to_file ("/tmp/jit-typeinfo-dump.c", true); +#endif } const jit_function& @@ -1883,10 +2627,23 @@ } jit_type* -jit_typeinfo::new_type (const std::string& name, jit_type *parent, - llvm::Type *llvm_type, bool skip_paren) +jit_typeinfo::new_type (const std::string& name, jit_type *parent +#ifdef HAVE_LLVM + , llvm::Type *llvm_type +#endif +#ifdef HAVE_GCCJIT + , gccjit::type gccjit_type +#endif + , bool skip_paren) { - jit_type *ret = new jit_type (name, parent, llvm_type, skip_paren, next_id++); + jit_type *ret = new jit_type (name, parent +#ifdef HAVE_LLVM + , llvm_type +#endif +#ifdef HAVE_GCCJIT + , gccjit_type +#endif + , skip_paren, next_id++); id_to_type.push_back (ret); return ret; } @@ -1903,14 +2660,21 @@ // FIXME: cp between add_binary_op, add_binary_icmp, and add_binary_fcmp void -jit_typeinfo::add_binary_op (jit_type *ty, int op, int llvm_op) +jit_typeinfo::add_binary_op (jit_type *ty, int op + , int llvm_op +#ifdef HAVE_GCCJIT + , enum gcc_jit_binary_op gccjit_op +#endif + ) { std::stringstream fname; octave_value::binary_op ov_op = static_cast<octave_value::binary_op>(op); - fname << "octave_jit_" << octave_value::binary_op_as_string (ov_op) + fname << "octave_jit_" << octave_value::binary_op_fcn_name (ov_op) << "_" << ty->name (); jit_function fn = create_internal (fname.str (), ty, ty, ty); + + // LLVM implementation: llvm::BasicBlock *block = fn.new_block (); builder.SetInsertPoint (block); llvm::Instruction::BinaryOps temp @@ -1919,18 +2683,37 @@ llvm::Value *ret = builder.CreateBinOp (temp, fn.argument (builder, 0), fn.argument (builder, 1)); fn.do_return (builder, ret); + + // gccjit implementation: +#ifdef HAVE_GCCJIT + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return ( + gf.get_context ().new_binary_op ( + gccjit_op, + ty->to_gccjit (), + gf.get_param (0), + gf.get_param (1))); +#endif binary_ops[op].add_overload (fn); } void -jit_typeinfo::add_binary_icmp (jit_type *ty, int op, int llvm_op) +jit_typeinfo::add_binary_icmp (jit_type *ty, int op + , int llvm_op +#ifdef HAVE_GCCJIT + , enum gcc_jit_comparison gccjit_op +#endif + ) { std::stringstream fname; octave_value::binary_op ov_op = static_cast<octave_value::binary_op>(op); - fname << "octave_jit" << octave_value::binary_op_as_string (ov_op) + fname << "octave_jit_" << octave_value::binary_op_fcn_name (ov_op) << "_" << ty->name (); jit_function fn = create_internal (fname.str (), boolean, ty, ty); + + // LLVM implementation: llvm::BasicBlock *block = fn.new_block (); builder.SetInsertPoint (block); llvm::CmpInst::Predicate temp @@ -1938,18 +2721,36 @@ llvm::Value *ret = builder.CreateICmp (temp, fn.argument (builder, 0), fn.argument (builder, 1)); fn.do_return (builder, ret); + + // gccjit implementation: +#ifdef HAVE_GCCJIT + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return ( + gf.get_context ().new_comparison ( + gccjit_op, + gf.get_param (0), + gf.get_param (1))); +#endif binary_ops[op].add_overload (fn); } void -jit_typeinfo::add_binary_fcmp (jit_type *ty, int op, int llvm_op) +jit_typeinfo::add_binary_fcmp (jit_type *ty, int op + , int llvm_op +#ifdef HAVE_GCCJIT + , enum gcc_jit_comparison gccjit_op +#endif + ) { std::stringstream fname; octave_value::binary_op ov_op = static_cast<octave_value::binary_op>(op); - fname << "octave_jit" << octave_value::binary_op_as_string (ov_op) + fname << "octave_jit_" << octave_value::binary_op_fcn_name (ov_op) << "_" << ty->name (); jit_function fn = create_internal (fname.str (), boolean, ty, ty); + + // LLVM implementation: llvm::BasicBlock *block = fn.new_block (); builder.SetInsertPoint (block); llvm::CmpInst::Predicate temp @@ -1957,15 +2758,30 @@ llvm::Value *ret = builder.CreateFCmp (temp, fn.argument (builder, 0), fn.argument (builder, 1)); fn.do_return (builder, ret); + + // gccjit implementation: +#ifdef HAVE_GCCJIT + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return ( + gf.get_context ().new_comparison ( + gccjit_op, + gf.get_param (0), + gf.get_param (1))); +#endif binary_ops[op].add_overload (fn); } jit_function -jit_typeinfo::create_function (jit_convention::type cc, const llvm::Twine& name, +jit_typeinfo::create_function (jit_convention::type cc, std::string name, jit_type *ret, const std::vector<jit_type *>& args) { - jit_function result (module, cc, name, ret, args); + jit_function result (module, +#ifdef HAVE_GCCJIT + gccjit_ctxt, +#endif + cc, name, ret, args); return result; } @@ -1982,15 +2798,26 @@ name << "id_" << type->name (); jit_function fn = create_internal (name.str (), type, type); + + // LLVM implementation: llvm::BasicBlock *body = fn.new_block (); builder.SetInsertPoint (body); fn.do_return (builder, fn.argument (builder, 0)); + + // gccjit implementation: +#ifdef HAVE_GCCJIT + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block (); + b.end_with_return (gf.get_param (0)); +#endif + return identities[id] = fn; } return identities[id]; } +#ifdef HAVE_LLVM llvm::Value * jit_typeinfo::do_insert_error_check (llvm::IRBuilderD& abuilder) { @@ -2004,11 +2831,31 @@ val->setVolatile (true); return abuilder.CreateICmpSGT (val, abuilder.getInt32 (0)); } +#endif /* #ifdef HAVE_LLVM */ + +#ifdef HAVE_GCCJIT +gccjit::rvalue +jit_typeinfo::do_insert_error_check (gccjit::function func) +{ + return error_state_gccjit.cast_to (gccjit_ctxt.get_type (GCC_JIT_TYPE_BOOL)); +} + +gccjit::rvalue +jit_typeinfo::do_insert_interrupt_check (gccjit::function func) +{ + return octave_interrupt_state_gccjit > sig_atomic_type_gccjit.zero (); +} +#endif /* #ifdef HAVE_GCCJIT */ + void jit_typeinfo::add_builtin (const std::string& name) { - jit_type *btype = new_type (name, any, any->to_llvm (), true); + jit_type *btype = new_type (name, any, any->to_llvm (), +#ifdef HAVE_GCCJIT + any->to_gccjit (), +#endif + true); builtins[name] = btype; octave_builtin *ov_builtin = find_builtin (name); @@ -2050,6 +2897,21 @@ llvm::Value *ret = builder.CreateCall (ifun, fargs); fn.do_return (builder, ret); + + // gcc implementation +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::function builtin = gccjit_ctxt.get_builtin_function (name); + std::vector<gccjit::rvalue> gccjit_args (nargs); + for (size_t i = 0; i < nargs; ++i) + gccjit_args[i] = gf.get_param (i + 1); + gccjit::rvalue ret = gccjit_ctxt.new_call (builtin, gccjit_args); + gccjit::block b = gf.new_block (); + b.end_with_return (ret); + } +#endif + paren_subsref_fn.add_overload (fn); } @@ -2075,6 +2937,8 @@ std::copy (args.begin (), args.end (), fn_args.begin () + 1); jit_function fn = create_internal (name, result, fn_args); fn.mark_can_error (); + + // LLVM implementation: llvm::BasicBlock *block = fn.new_block (); builder.SetInsertPoint (block); llvm::Type *any_t = any->to_llvm (); @@ -2105,6 +2969,55 @@ jit_function cast_result = cast (result, any); fn.do_return (builder, cast_result.call (builder, ret)); + + // gccjit implementation: +#ifdef HAVE_GCCJIT + { + gccjit::function gf = fn.gccjit_function; + gccjit::block b = gf.new_block ("body of register_generic"); + b.add_comment ("TODO: register_generic"); + + gccjit::type array_t = gccjit_ctxt.new_array_type (any->to_gccjit (), + args.size ()); +#if 1 + gccjit::lvalue array = gf.new_local (array_t, "tmp_array"); + for (size_t i = 0; i < args.size (); ++i) + { + gccjit::lvalue arg = gf.get_param (i + 1); + jit_function agrab = get_grab (args[i]); + if (agrab.valid ()) + { + std::vector<gccjit::rvalue> grab_args (1, arg); + b.add_assignment (arg, agrab.call (gccjit_ctxt, b, grab_args)); + } + jit_function acast = cast (any, args[i]); + std::vector<gccjit::rvalue> cast_args (1, arg); + b.add_assignment (array[i], + acast.call (gccjit_ctxt, b, cast_args)); + } +#endif + + gccjit::type int_t = intN (sizeof (octave_builtin::fcn) * 8)->to_gccjit (); + size_t fcn_int = reinterpret_cast<size_t> (builtin->function ()); + gccjit::rvalue fcn = gccjit_ctxt.new_rvalue (int_t, (int)fcn_int); + gccjit::rvalue nargin = gccjit_ctxt.new_rvalue (int_t, (int)args.size ()); + size_t result_int = reinterpret_cast<size_t> (result); + gccjit::rvalue res_gcc = gccjit_ctxt.new_rvalue (int_t, (int)result_int); + std::vector<gccjit::rvalue> call_args (4); + call_args[0] = fcn; + call_args[1] = nargin; + call_args[2] = array; + call_args[3] = res_gcc; + gccjit::rvalue ret = any_call.call (gccjit_ctxt, b, call_args); + + jit_function cast_result = cast (result, any); + std::vector<gccjit::rvalue> final_cast_args (1); + final_cast_args[0] = ret; + gccjit::rvalue final_result = cast_result.call (gccjit_ctxt, b, final_cast_args); + b.end_with_return (final_result); + } +#endif + paren_subsref_fn.add_overload (fn); } @@ -2117,6 +3030,7 @@ if (fn.can_error ()) ret.mark_can_error (); + // LLVM implementation: llvm::BasicBlock *body = ret.new_block (); builder.SetInsertPoint (body); llvm::Value *result = fn.call (builder, ret.argument (builder, 1), @@ -2126,9 +3040,27 @@ else ret.do_return (builder); + // gccjit implementation: +#ifdef HAVE_GCCJIT + gccjit::function gf = ret.gccjit_function; + gccjit::block b = gf.new_block (); + b.add_comment ("built by mirror_binary"); + //std::vector<gccjit::rvalue> gccjit_args (2); + //gccjit_args[0] = gf.get_param (1); + //gccjit_args[1] = gf.get_param (0); + if (ret.result ()) + b.end_with_return ( + gccjit_ctxt.new_call (fn.gccjit_function, + gf.get_param (1), + gf.get_param (0))); + else + b.end_with_return (); +#endif + return ret; } +#ifdef HAVE_LLVM llvm::Value * jit_typeinfo::pack_complex (llvm::IRBuilderD& bld, llvm::Value *cplx) { @@ -2189,14 +3121,65 @@ ret = complex_real (ret, real); return complex_imag (ret, imag); } +#endif // #ifdef HAVE_LLVM + +#ifdef HAVE_GCCJIT +gccjit::rvalue +jit_typeinfo::complex_real (gccjit::rvalue cx) +{ + return cx[0]; +} + +gccjit::rvalue +jit_typeinfo::complex_real (gccjit::block block, + gccjit::lvalue cx, + gccjit::rvalue real) +{ + block.add_assignment (cx[0], real); + return cx; +} + +gccjit::rvalue +jit_typeinfo::complex_imag (gccjit::rvalue cx) +{ + return cx[1]; +} + +gccjit::rvalue +jit_typeinfo::complex_imag (gccjit::block block, + gccjit::lvalue cx, + gccjit::rvalue imag) +{ + block.add_assignment (cx[1], imag); + return cx; +} + +gccjit::rvalue +jit_typeinfo::complex_new (gccjit::block block, + gccjit::rvalue real, + gccjit::rvalue imag) +{ + gccjit::rvalue tmp = + block.get_function ().new_local (complex->to_gccjit (), + "complex_new"); + block.add_assignment (tmp[0], real); + block.add_assignment (tmp[1], imag); + return tmp; +} +#endif // #ifdef HAVE_GCCJIT void jit_typeinfo::create_int (size_t nbits) { std::stringstream tname; tname << "int" << nbits; - ints[nbits] = new_type (tname.str (), any, llvm::Type::getIntNTy (context, - nbits)); + ints[nbits] = new_type (tname.str (), any + , llvm::Type::getIntNTy (context, + nbits) +#ifdef HAVE_GCCJIT + , gccjit_ctxt.get_int_type (nbits / 8, 1) +#endif + ); } jit_type *