#!/usr/bin/python3 # # Copyright (C) 2018 The Android Open Source Project # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # """Generate intrinsics code.""" from collections import OrderedDict import asm_defs import gen_device_insn_info_lib import json import os import re import sys # C-level intrinsic calling convention: # 1. All arguments are passed using the natural data types: # - int8_t passed as one byte argument (on the stack in IA32 mode, in GP register in x86-64 mode) # - int32_t passed as 4 bytes argument (on the stack in IA32 mode, in GP register in x86-64 mode) # - int64_t is passed as 8 byte argument (on the stack in IA32 mode, in GP register in x86-64 mode) # - float is passed as float (on the stack in IA32 mode, in XMM register in x86-64 mode) # - double is passed as double (on the stack in IA32 mode, in XMM register in x86-64 mode) # - vector formats are passed as pointers to 128bit data structure # 2. Return values. # - Values are returned as std::tuple. This means that on IA32 it's always returned on stack. INDENT = ' ' AUTOGEN = """\ // This file automatically generated by gen_intrinsics.py // DO NOT EDIT! """ class VecFormat(object): def __init__(self, num_elements, element_size, is_unsigned, is_float, index, c_type): self.num_elements = num_elements self.element_size = element_size self.is_unsigned = is_unsigned self.is_float = is_float self.index = index self.c_type = c_type # Vector format defined as: # vector_size, element_size, is_unsigned, is_float, index, ir_format, c_type # TODO(olonho): make flat numbering after removing legacy macro compat. _VECTOR_FORMATS = { 'U8x8': VecFormat(8, 1, True, False, 1, 'uint8_t'), 'U16x4': VecFormat(4, 2, True, False, 2, 'uint16_t'), 'U32x2': VecFormat(2, 4, True, False, 3, 'uint32_t'), 'U64x1': VecFormat(1, 8, True, False, 4, 'uint64_t'), 'U8x16': VecFormat(16, 1, True, False, 5, 'uint8_t'), 'U16x8': VecFormat(8, 2, True, False, 6, 'uint16_t'), 'U32x4': VecFormat(4, 4, True, False, 7, 'uint32_t'), 'U64x2': VecFormat(2, 8, True, False, 8, 'uint64_t'), 'I8x8': VecFormat(8, 1, False, False, 9, 'int8_t'), 'I16x4': VecFormat(4, 2, False, False, 10, 'int16_t'), 'I32x2': VecFormat(2, 4, False, False, 11, 'int32_t'), 'I64x1': VecFormat(1, 8, False, False, 12, 'int64_t'), 'I8x16': VecFormat(16, 1, False, False, 13, 'int8_t'), 'I16x8': VecFormat(8, 2, False, False, 14, 'int16_t'), 'I32x4': VecFormat(4, 4, False, False, 15, 'int32_t'), 'I64x2': VecFormat(2, 8, False, False, 16, 'int64_t'), 'U8x1': VecFormat(1, 1, True, False, 17, 'uint8_t'), 'I8x1': VecFormat(1, 1, False, False, 18, 'int8_t'), 'U16x1': VecFormat(1, 2, True, False, 19, 'uint16_t'), 'I16x1': VecFormat(1, 2, False, False, 20, 'int16_t'), 'U32x1': VecFormat(1, 4, True, False, 21, 'uint32_t'), 'I32x1': VecFormat(1, 4, False, False, 22, 'int32_t'), # These vector formats can never intersect with above, so can reuse index. 'F32x1': VecFormat(1, 4, False, True, 1, 'Float32'), 'F32x2': VecFormat(2, 4, False, True, 2, 'Float32'), 'F32x4': VecFormat(4, 4, False, True, 3, 'Float32'), 'F64x1': VecFormat(1, 8, False, True, 4, 'Float64'), 'F64x2': VecFormat(2, 8, False, True, 5, 'Float64'), # Those vector formats can never intersect with above, so can reuse index. 'U8x4': VecFormat(4, 1, True, False, 1, 'uint8_t'), 'U16x2': VecFormat(2, 2, True, False, 2, 'uint16_t'), 'I8x4': VecFormat(4, 1, False, False, 3, 'int8_t'), 'I16x2': VecFormat(2, 2, False, False, 4, 'int16_t'), } class VecSize(object): def __init__(self, num_elements, index): self.num_elements = num_elements self.index = index _VECTOR_SIZES = {'X64': VecSize(64, 1), 'X128': VecSize(128, 2)} _ROUNDING_MODES = ['FE_TONEAREST', 'FE_DOWNWARD', 'FE_UPWARD', 'FE_TOWARDZERO', 'FE_TIESAWAY'] def _is_imm_type(arg_type): return 'imm' in arg_type def _is_template_type(arg_type): if not arg_type.startswith('Type'): return False assert isinstance(int(arg_type[4:]), int) return True def _get_imm_c_type(arg_type): return { 'imm8' : 'int8_t', 'uimm8' : 'uint8_t', 'uimm16' : 'uint16_t', 'uimm32' : 'uint32_t', }[arg_type] def _get_c_type(arg_type): if (arg_type in ('Float16', 'Float32', 'Float64', 'int8_t', 'uint8_t', 'int16_t', 'uint16_t', 'int32_t', 'uint32_t', 'int64_t', 'uint64_t', 'volatile uint8_t*', 'volatile uint32_t*') or _is_template_type(arg_type)): return arg_type if arg_type in ('fp_flags', 'fp_control', 'int', 'flag', 'flags', 'vec32'): return 'uint32_t' if _is_imm_type(arg_type): return _get_imm_c_type(arg_type) if arg_type == 'vec': return 'SIMD128Register' if arg_type in _ROUNDING_MODES: return 'int' raise Exception('Type %s not supported' % (arg_type)) def _get_semantic_player_type(arg_type, type_map, listener=''): if type_map is not None and type_map != False and arg_type in type_map: type = type_map[arg_type] if type in ('FpRegister', 'Register', 'SimdRegister'): return listener + type return type if arg_type in ('Float16', 'Float32', 'Float64', 'vec'): return listener + 'SimdRegister' if _is_imm_type(arg_type): return _get_imm_c_type(arg_type) return listener + 'Register' def _gen_scalar_intr_decl(f, name, intr): ins = intr.get('in') outs = intr.get('out') params = [_get_c_type(op) for op in ins] if len(outs) > 0: retval = 'std::tuple<' + ', '.join(_get_c_type(out) for out in outs) + '>' else: retval = 'void' comment = intr.get('comment') if comment: print('// %s.' % (comment), file=f) if intr.get('precise_nans', False): print('template ', file=f) print('%s %s(%s);' % (retval, name, ', '.join(params)), file=f) def _gen_template_intr_decl(f, name, intr): ins = intr.get('in') outs = intr.get('out') params = [_get_c_type(op) for op in ins] if len(outs) > 0: retval = 'std::tuple<' + ', '.join(_get_c_type(out) for out in outs) + '>' else: retval = 'void' comment = intr.get('comment') if comment: print('// %s.' % (comment), file=f) print('template <%s>' % _get_template_parameters( intr.get('variants'), intr.get('precise_nans', False)), file=f) print('%s %s(%s);' % (retval, name, ', '.join(params)), file=f) def _get_template_parameters( variants, precise_nans=False, extra=['enum PreferredIntrinsicsImplementation = kUseAssemblerImplementationIfPossible']): if len(extra) == 0: typename = 'intrinsics::TemplateTypeId' else: typename = 'typename' template = None for variant in variants: counter = -1 def get_counter(): nonlocal counter counter += 1 return counter new_template = ', '.join( (["bool kPreciseNaNOperationsHandling"] if precise_nans else []) + ['bool kBool%s' % get_counter() if param.strip() in ('true', 'false') else 'int kInt%s' % get_counter() if param.strip() in _ROUNDING_MODES else '%s Type%d' % (typename, get_counter()) if re.search('[_a-zA-Z]', param) else 'int kInt%s' % get_counter() for param in variant.split(',')] + extra) assert template is None or template == new_template template = new_template return template def _gen_vector_intr_decl(f, name, intr): ins = intr.get('in') outs = intr.get('out') params = [_get_c_type(op) for op in ins] if len(outs) > 0: retval = 'std::tuple<' + ', '.join(_get_c_type(out) for out in outs) + '>' else: retval = 'void' comment = intr.get('comment') if comment: print('// %s.' % (comment), file=f) if intr.get('precise_nans', False): template_parameters = 'bool precise_nan_operations_handling, ' else: template_parameters = '' if not 'raw' in intr['variants']: template_parameters += 'typename Type, ' template_parameters += 'int size, ' template_parameters += 'enum PreferredIntrinsicsImplementation' template_parameters += ' = kUseAssemblerImplementationIfPossible' print('template <%s>' % template_parameters, file=f) print('%s %s(%s);' % (retval, name, ', '.join(params)), file=f) def _is_vector_class(intr): return intr.get('class') in ('vector_4', 'vector_8', 'vector_16', 'vector_8/16', 'vector_8/16/single', 'vector_8/single', 'vector_16/single') def _is_simd128_conversion_required(t, type_map=None): return (_get_semantic_player_type(t, type_map) == 'SimdRegister' and _get_c_type(t) != 'SIMD128Register') def _get_semantics_player_hook_result(intr, listener=''): outs = intr['out'] if len(outs) == 0: return 'void' elif len(outs) == 1: # No tuple for single result. return _get_semantic_player_type(outs[0], intr.get('sem-player-types'), listener) return 'std::tuple<' + ', '.join( _get_semantic_player_type(out, intr.get('sem-player-types'), listener) for out in outs) + '>' def _get_semantics_player_hook_proto_components(name, intr, listener=''): ins = intr['in'] args = [] if _is_vector_class(intr): if 'raw' in intr['variants']: assert len(intr['variants']) == 1, "Unexpected length of variants" args = ["uint8_t size"] else: args = ["uint8_t elem_size", "uint8_t elem_num"] if (_is_signed(intr) and _is_unsigned(intr)): args += ['bool is_signed'] args += [ '%s arg%d' % ( _get_semantic_player_type(op, intr.get('sem-player-types'), listener), num) for num, op in enumerate(ins) ] result = _get_semantics_player_hook_result(intr, listener) return result, name, ', '.join(args) def _get_semantics_player_hook_proto(name, intr, make_template=True, listener=''): if listener != '': make_template = False result, name, params = _get_semantics_player_hook_proto_components(name, intr, listener) if intr.get('class') == 'template': template_parameters = _get_template_parameters(intr.get('variants'), extra = []) if make_template: spec_arguments = _get_template_spec_arguments(intr.get('variants')) values = ', ' + ', '.join( ['intrinsics::Value<%s>' % argument for argument in spec_arguments]) return 'template<%s>\n%s %s(%s%s)' % ( template_parameters, result, name, params, values) else: return '%s %s%s(%s, %s)' % ( result, listener, name, params, template_parameters) return '%s %s(%s)' % (result, name, params) def _get_interpreter_hook_call_expr(name, intr, desc=None): ins = intr['in'] outs = intr['out'] call_params = [] for num, op in enumerate(ins): arg = 'arg%d' % (num) semantic_player_type = _get_semantic_player_type( op, intr.get('sem-player-types')) if semantic_player_type == 'FpRegister': if op.startswith('Type'): op = 'intrinsics::TypeFromId<%s>' % op call_params.append('FPRegToFloat<%s>(%s)' % (op, arg)) elif semantic_player_type == 'SimdRegister': call_params.append(_get_cast_from_simd128(arg, op, ptr_bits=64)) elif '*' in _get_c_type(op): call_params.append('berberis::bit_cast<%s>(%s)' % (_get_c_type(op), arg)) else: c_type = _get_c_type(op) if c_type.startswith('Type'): c_type = 'intrinsics::TypeFromId<%s>' % c_type call_params.append('GPRRegToInteger<%s>(%s)' % (c_type, arg)) call_expr = 'intrinsics::%s%s(%s)' % ( name, _get_desc_specializations(intr, desc), ', '.join(call_params)) if len(outs) == 1: # Unwrap tuple for single result. call_expr = 'std::get<0>(%s)' % call_expr if 'sem-player-types' in intr: out_type = _get_semantic_player_type(outs[0], intr.get('sem-player-types')) if out_type == "FpRegister": call_expr = 'FloatToFPReg(%s)' % call_expr elif out_type != "SimdRegister": assert out_type == "Register" assert not _is_simd128_conversion_required( outs[0], intr.get('sem-player-types')) call_expr = 'IntegerToGPRReg(%s)' % call_expr else: # Currently this kind of mismatch can only happen for single result, so we # can keep simple code here for now. if _is_simd128_conversion_required(outs[0]): out_type = _get_c_type(outs[0]) if out_type in ('Float16', 'Float32', 'Float64'): call_expr = 'FloatToFPReg(%s)' % call_expr else: raise Exception('Type %s is not supported' % (out_type)) else: if any(_is_simd128_conversion_required(out) for out in outs): raise Exception( 'Unsupported SIMD128Register conversion with multiple results') return call_expr def _get_interpreter_hook_return_stmt(name, intr, desc=None): return 'return ' + _get_interpreter_hook_call_expr(name, intr, desc) + ';' def _get_unused(intr): call_expr = 'UNUSED(%s);' % ', '.join('arg%d' % (num) for num, _ in enumerate(intr['in'])) return call_expr def _get_placeholder_return_stmt(intr, f): print(INDENT + _get_unused(intr), file=f) outs = intr['out'] if outs: print(INDENT + 'return {};', file=f) def _get_semantics_player_hook_raw_vector_body(name, intr, get_return_stmt): outs = intr['out'] if (len(outs) == 0): raise Exception('No result raw vector intrinsic is not supported') reg_class = intr.get('class') yield 'switch (size) {' for fmt, desc in _VECTOR_SIZES.items(): if _check_reg_class_size(reg_class, desc.num_elements / 8): yield INDENT + 'case %s:' % desc.num_elements yield 2 * INDENT + get_return_stmt(name, intr, desc) yield INDENT + 'default:' yield 2 * INDENT + 'LOG_ALWAYS_FATAL("Unsupported size");' yield '}' def _is_signed(intr): return any(v.startswith("signed") for v in intr['variants']) def _is_unsigned(intr): return any(v.startswith("unsigned") for v in intr['variants']) def _get_vector_format_init_expr(intr): variants = intr.get('variants') if ('Float16' in variants or 'Float32' in variants or 'Float64' in variants): return 'intrinsics::GetVectorFormatFP(elem_size, elem_num)' assert _is_signed(intr) or _is_unsigned(intr), "Unexpected intrinsic class" if _is_signed(intr) and _is_unsigned(intr): signed_arg = ', is_signed' else: signed_arg = ', true' if _is_signed(intr) else ', false' return 'intrinsics::GetVectorFormatInt(elem_size, elem_num%s)' % signed_arg def _get_semantics_player_hook_vector_body(name, intr, get_return_stmt): outs = intr['out'] if (len(outs) == 0): raise Exception('No result vector intrinsic is not supported') reg_class = intr.get('class') yield 'auto format = %s;' % _get_vector_format_init_expr(intr) yield 'switch (format) {' for variant in intr.get('variants'): for fmt, desc in _VECTOR_FORMATS.items(): if (_check_reg_class_size(reg_class, desc.element_size * desc.num_elements) and _check_typed_variant(variant, desc)): yield INDENT + 'case intrinsics::kVector%s:' % fmt yield 2 * INDENT + get_return_stmt(name, intr, desc) elif (reg_class in ('vector_8/single', 'vector_8/16/single', 'vector_16/single') and desc.num_elements == 1 and _check_typed_variant(variant, desc)): assert desc.element_size <= 8, "Unexpected element size" yield INDENT + 'case intrinsics::kVector%s:' % fmt yield 2 * INDENT + get_return_stmt(name, intr, desc) yield INDENT + 'default:' yield 2 * INDENT + 'LOG_ALWAYS_FATAL("Unsupported format");' yield '}' # Syntax sugar heavily used in tests. def _get_interpreter_hook_vector_body(name, intr): return _get_semantics_player_hook_vector_body( name, intr, _get_interpreter_hook_return_stmt) def _gen_interpreter_hook(f, name, intr, option): if intr.get('class') == 'template': print('#ifndef BERBERIS_INTRINSICS_HOOKS_INLINE_DEMULTIPLEXER\n' '%s const;\n#endif' % ( _get_semantics_player_hook_proto(name, intr, make_template=False)), file=f) print('%s const {' % ( _get_semantics_player_hook_proto(name, intr)), file=f) if _is_vector_class(intr): if 'raw' in intr['variants']: assert len(intr['variants']) == 1, "Unexpected length of variants" lines = _get_semantics_player_hook_raw_vector_body( name, intr, _get_interpreter_hook_return_stmt) else: lines = _get_interpreter_hook_vector_body(name, intr) lines = [INDENT + l for l in lines] print('\n'.join(lines), file=f) else: # TODO(b/363057506): Add float support and clean up the logic here. arm64_allowlist = ['AmoAdd', 'AmoAnd', 'AmoMax', 'AmoMin', 'AmoOr', 'AmoSwap', 'AmoXor', 'Bclr', 'Bclri', 'Bext', 'Bexti', 'Binv', 'Binvi', 'Bset', 'Bseti', 'DivRiscV', 'Max', 'Min', 'RemRiscV', 'Rev8', 'Rol', 'Ror', 'Sext', 'Sh1add', 'Sh1adduw', 'Sh2add', 'Sh2adduw', 'Sh3add', 'Sh3adduw', 'Zext', 'UnboxNan'] if (option == 'arm64') and (name not in arm64_allowlist): _get_placeholder_return_stmt(intr, f) else: print(INDENT + _get_interpreter_hook_return_stmt(name, intr), file=f) print('}\n', file=f) def _get_translator_hook_call_expr(name, intr, desc=None): desc_spec = _get_desc_specializations(intr, desc) args = [('arg%d' % n) for n, _ in enumerate(intr['in'])] template_params = ['&intrinsics::' + name + desc_spec] template_params += [_get_semantics_player_hook_result(intr)] return 'CallIntrinsic<%s>(%s)' % (', '.join(template_params), ', '.join(args)) def _get_translator_hook_return_stmt(name, intr, desc=None): return 'return ' + _get_translator_hook_call_expr(name, intr, desc) + ';' def _gen_translator_hook(f, name, intr): if intr.get('class') == 'template': print('#ifndef BERBERIS_INTRINSICS_HOOKS_INLINE_DEMULTIPLEXER\n' '%s;\n#endif' % ( _get_semantics_player_hook_proto(name, intr, make_template=False)), file=f) print('%s {' % ( _get_semantics_player_hook_proto(name, intr)), file=f) if _is_vector_class(intr): if 'raw' in intr['variants']: assert len(intr['variants']) == 1, "Unexpected length of variants" lines = _get_semantics_player_hook_raw_vector_body( name, intr, _get_translator_hook_return_stmt) else: lines = _get_semantics_player_hook_vector_body( name, intr, _get_translator_hook_return_stmt) lines = [INDENT + l for l in lines] print('\n'.join(lines), file=f) else: print(INDENT + _get_translator_hook_return_stmt(name, intr), file=f) print('}\n', file=f) def _get_expectations_for_variant(variant): expectations = [] for param in variant.split(','): param = param.strip() if (re.search('[_a-zA-Z]', param) and not param in (['true', 'false'] + _ROUNDING_MODES)): expectations.append('intrinsics::kIdFromType<%s>' % param) else: expectations.append(param) return expectations def _get_semantics_player_hook_template_switch(name, intr): variables = _get_template_spec_arguments(intr.get('variants')) yield 'switch (intrinsics::TrivialDemultiplexer(%s)) {' % ', '.join(variables) for variant in intr.get('variants'): expectations = _get_expectations_for_variant(variant) yield '%scase intrinsics::TrivialDemultiplexer(%s):' % ( INDENT, ', '.join(expectations)) yield 2 * INDENT + "// Disable LOG_NDEBUG to use DCHECK for debugging!" for expectation, variable in zip(expectations, variables): yield 2 * INDENT + 'DCHECK_EQ(%s, %s);' % (expectation, variable) yield 2 * INDENT + 'return %s(%s);' % ( name, ','.join(list('arg%d' % arg[0] for arg in enumerate(intr['in'])) + list(f'intrinsics::Value<{expectation}>{{}}' for expectation in expectations))) yield INDENT + 'default:' yield 2 * INDENT + 'FATAL("Unsupported size");' yield '}' def _gen_demultiplexer_hook(f, name, intr): print('%s BERBERIS_INTRINSICS_HOOKS_CONST {' % _get_semantics_player_hook_proto( name, intr, listener=' BERBERIS_INTRINSICS_HOOKS_LISTENER '), file=f) lines = _get_semantics_player_hook_template_switch(name, intr) lines = [INDENT + l for l in lines] print('\n'.join(lines), file=f) print('}\n', file=f) def _gen_mock_semantics_listener_hook(f, name, intr): result, name, params = _get_semantics_player_hook_proto_components(name, intr) if intr.get('class') == 'template': spec_arguments = _get_template_spec_arguments(intr.get('variants')) template_parameters = _get_template_parameters(intr.get('variants'), extra = []) args = ', '.join([('arg%d' % n) for n, _ in enumerate(intr['in'])] + spec_arguments) values = ', ' + ', '.join( ['intrinsics::Value<%s>' % argument for argument in spec_arguments]) print('template<%s>\n%s %s(%s%s) {' % ( template_parameters, result, name, params, values), file=f) _gen_template_parameters_verifier(f, intr) print('%sreturn %s(%s);\n}' % (INDENT, name, args), file=f) params = ', '.join( [params] + ['%s %s' % ( { 'kBoo': 'bool', 'kInt': 'int', 'Type': 'intrinsics::TemplateTypeId' }[argument[0:4]], argument) for argument in spec_arguments]) print('MOCK_METHOD((%s), %s, (%s));' % (result, name, params), file=f) def _gen_template_parameters_verifier(f, intr): received_params = ', '.join(_get_template_spec_arguments(intr.get('variants'))) print('%sstatic_assert(%s);' % ( INDENT, ' || '.join( 'std::tuple{%s} == std::tuple{%s}' % ( received_params, ', '.join(_get_expectations_for_variant(variant))) for variant in intr.get('variants'))), file=f) def _check_signed_variant(variant, desc): if variant == 'signed': return True if variant == 'signed_32': return desc.element_size == 4 if variant == 'signed_64': return desc.element_size == 8 if variant == 'signed_16/32': return desc.element_size in (2, 4) if variant == 'signed_8/16/32': return desc.element_size in (1, 2, 4) if variant == 'signed_16/32/64': return desc.element_size in (2, 4, 8) if variant == 'signed_8/16/32/64': return desc.element_size in (1, 2, 4, 8) if variant == 'signed_32/64': return desc.element_size in (4, 8) return False def _check_unsigned_variant(variant, desc): if variant == 'unsigned': return True if variant == 'unsigned_8': return desc.element_size == 1 if variant == 'unsigned_16': return desc.element_size == 2 if variant == 'unsigned_32': return desc.element_size == 4 if variant == 'unsigned_64': return desc.element_size == 8 if variant == 'unsigned_8/16': return desc.element_size in (1, 2) if variant == 'unsigned_8/16/32': return desc.element_size in (1, 2, 4) if variant == 'unsigned_16/32/64': return desc.element_size in (2, 4, 8) if variant == 'unsigned_8/16/32/64': return desc.element_size in (1, 2, 4, 8) if variant == 'unsigned_32/64': return desc.element_size in (4, 8) return False def _check_reg_class_size(reg_class, size): # Small vectors are separate namespace. if size == 4 and reg_class == 'vector_4': return True if size == 8 and reg_class in ('vector_8', 'vector_8/16', 'vector_8/16/single', 'vector_8/single'): return True if size == 16 and reg_class in ('vector_16', 'vector_8/16', 'vector_8/16/single', 'vector_16/single'): return True return False def _check_typed_variant(variant, desc): if desc.is_unsigned and not desc.is_float: return _check_unsigned_variant(variant, desc) if not desc.is_unsigned and not desc.is_float: return _check_signed_variant(variant, desc) if desc.is_float: if desc.element_size == 2: return variant == 'Float16' if desc.element_size == 4: return variant == 'Float32' if desc.element_size == 8: return variant == 'Float64' return False def _get_formats_with_descriptions(intr): reg_class = intr.get('class') for variant in intr.get('variants'): found_fmt = False for fmt, desc in _VECTOR_FORMATS.items(): if (_check_reg_class_size(reg_class, desc.element_size * desc.num_elements) and _check_typed_variant(variant, desc) and (reg_class != 'vector_4' or desc.element_size < 4)): found_fmt = True yield fmt, desc if variant == 'raw': for fmt, desc in _VECTOR_SIZES.items(): if _check_reg_class_size(reg_class, desc.num_elements / 8): found_fmt = True yield fmt, desc assert found_fmt, 'Couldn\'t expand %s' % reg_class def _get_result_type(outs): result_type = 'void' return_stmt = '' if len(outs) >= 1: result_type = ('std::tuple<' + ', '.join(_get_c_type(out) for out in outs) + '>') return_stmt = 'return ' return result_type, return_stmt def _get_in_params(params): for param_index, param in enumerate(params): yield _get_c_type(param), 'in%d' % (param_index) def _get_out_params(params): for param_index, param in enumerate(params): yield _get_c_type(param), 'out%d' % (param_index) def _get_cast_from_simd128(var, target_type, ptr_bits): if ('*' in target_type): return 'berberis::bit_cast<%s>(%s.Get(0))' % (_get_c_type(target_type), var, ptr_bits) c_type = _get_c_type(target_type) if c_type in ('Float16', 'Float32', 'Float64'): return 'FPRegToFloat<%s>(%s)' % (c_type, var) cast_map = { 'int8_t': '.Get(0)', 'uint8_t': '.Get(0)', 'int16_t': '.Get(0)', 'uint16_t': '.Get(0)', 'int32_t': '.Get(0)', 'uint32_t': '.Get(0)', 'int64_t': '.Get(0)', 'uint64_t': '.Get(0)', 'SIMD128Register': '' } return '%s%s' % (var, cast_map[c_type]) def _get_desc_specializations(intr, desc=None): if intr.get('class') == 'template': spec = _get_template_spec_arguments(intr.get('variants'), use_type_id=True) elif hasattr(desc, 'c_type'): spec = [desc.c_type, str(desc.num_elements)] elif hasattr(desc, 'num_elements'): spec = [str(desc.num_elements)] else: spec = [] if intr.get('precise_nans', False): spec = ['config::kPreciseNaNOperationsHandling'] + spec if not len(spec): return '' return '<%s>' % ', '.join(spec) def _get_template_spec_arguments(variants, use_type_id=False): spec = None for variant in variants: counter = -1 def get_counter(): nonlocal counter counter += 1 return counter new_spec = [ 'kBool%s' % get_counter() if param.strip() in ('true', 'false') else 'kInt%s' % get_counter() if param.strip() in _ROUNDING_MODES else ('intrinsics::TypeFromId' % get_counter() if use_type_id else 'Type%d' % get_counter()) if re.search('[_a-zA-Z]', param) else 'kInt%s' % get_counter() for param in variant.split(',')] assert spec is None or spec == new_spec spec = new_spec return spec def _intr_has_side_effects(intr, fmt=None): ins = intr.get('in') outs = intr.get('out') # If we have 'has_side_effects' mark in JSON file then we use it "as is". if 'has_side_effects' in intr: return intr.get('has_side_effects') # Otherwise we mark all floating-point related intrinsics as "volatile". # TODO(b/68857496): move that information in HIR/LIR and stop doing that. if 'Float16' in ins or 'Float32' in ins or 'Float64' in ins: return True if 'Float16' in outs or 'Float32' in outs or 'Float64' in outs: return True if fmt is not None and fmt.startswith('F'): return True return False def _gen_intrinsics_inl_h(f, intrs): print(AUTOGEN, file=f) for name, intr in intrs: if intr.get('class') == 'scalar': _gen_scalar_intr_decl(f, name, intr) elif intr.get('class') == 'template': _gen_template_intr_decl(f, name, intr) else: assert intr.get('class').startswith('vector') _gen_vector_intr_decl(f, name, intr) def _gen_semantic_player_types(intrs): for name, intr in intrs: if intr.get('class') == 'template': map = None for variant in intr.get('variants'): counter = -1 def get_counter(): nonlocal counter counter += 1 return counter new_map = { 'Float16': 'FpRegister', 'Float32': 'FpRegister', 'Float64': 'FpRegister', } for type in filter( lambda param: param.strip() not in ('true', 'false') and re.search('[_a-zA-Z]', param), variant.split(',')): new_map['Type%d' % get_counter()] = ( 'FpRegister' if type.strip() in ('Float16', 'Float32', 'Float64') else _get_semantic_player_type(type, None)) if map is None: map = new_map elif map != new_map: # Note: we would use literal `False` as type, which would lead to # compile-time error… that's Ok, because mix of ints and floats may # only happen with vector intrinsics where types used are # never arguments, but just specify type of vector element. # If intrinsics actually have to receive such arguments that such # intrinsics should be split in two. map = False intr['sem-player-types'] = map def _gen_interpreter_intrinsics_hooks_impl_inl_h(f, intrs, option): print(AUTOGEN, file=f) for name, intr in intrs: _gen_interpreter_hook(f, name, intr, option) def _gen_translator_intrinsics_hooks_impl_inl_h(f, intrs): print(AUTOGEN, file=f) for name, intr in intrs: _gen_translator_hook(f, name, intr) def _gen_demultiplexer_intrinsics_hooks_impl_inl_h(f, intrs): print(AUTOGEN, file=f) print(""" #ifndef BERBERIS_INTRINSICS_HOOKS_LISTENER #define BERBERIS_INTRINSICS_HOOKS_LISTENER #endif #ifndef BERBERIS_INTRINSICS_HOOKS_CONST #define BERBERIS_INTRINSICS_HOOKS_CONST #endif #ifndef BERBERIS_INTRINSICS_HOOKS_SHARDS_COUNT #define BERBERIS_INTRINSICS_HOOKS_SHARDS_COUNT 1 #define BERBERIS_INTRINSICS_HOOKS_SHARD 0 #endif """, file=f) counter = 0 for name, intr in intrs: if intr.get('class') == 'template': print('#if %d %% BERBERIS_INTRINSICS_HOOKS_SHARDS_COUNT == ' 'BERBERIS_INTRINSICS_HOOKS_SHARD' % counter, file=f) _gen_demultiplexer_hook(f, name, intr) print('#endif', file=f) counter += 1 def _gen_mock_semantics_listener_intrinsics_hooks_impl_inl_h(f, intrs): print(AUTOGEN, file=f) for name, intr in intrs: _gen_mock_semantics_listener_hook(f, name, intr) def _get_binding_info(arg): need_tmp = arg['class'] in ('EAX', 'EDX', 'CL', 'ECX', 'RAX', 'RDX') if arg['class'] == 'Imm8': return 'ImmArg<%d>' % (arg['ir_arg']) if arg['usage'] == 'use': if need_tmp: return 'InTmpArg<%d>' % arg['ir_arg'] return 'InArg<%d>' % arg['ir_arg'] if arg['usage'] in ('def', 'def_early_clobber'): assert 'ir_arg' not in arg if 'ir_res' in arg: if need_tmp: return 'OutTmpArg<%d>' % arg['ir_res'] return 'OutArg<%d>' % arg['ir_res'] return 'TmpArg' if arg['usage'] == 'use_def': if 'ir_res' in arg: if need_tmp: return 'InOutTmpArg<%s, %s>' % (arg['ir_arg'], arg['ir_res']) return 'InOutArg<%s, %s>' % (arg['ir_arg'], arg['ir_res']) return 'InTmpArg<%s>' % (arg['ir_arg']) assert False, 'unknown operand usage %s' % (arg['usage']) def _get_bindings_info(args): return 'std::tuple<%s>' % ', '.join(_get_binding_info(arg) for arg in args) def _gen_process_all_bindings(f, intrs, archs): print("%s" % AUTOGEN, file=f) callback_lines = [] for line in _gen_c_intrinsics_generator( intrs, _is_interpreter_compatible_assembler, False): callback_lines.append(line) # Put implementation into arch-specific namespace to access bindings. print("namespace %s {\n" % archs[-1], file = f) print(""" template constexpr void ProcessAllBindings([[maybe_unused]] Callback callback, [[maybe_unused]] Args&&... args) { using berberis::intrinsics::Float16; using berberis::intrinsics::Float32; using berberis::intrinsics::Float64;""", file=f) for line in callback_lines: print(line, file=f) print(""" } } // namespace %s using %s::ProcessAllBindings; """ % (archs[-1], archs[-1]), file=f) def _gen_process_bindings(f, intrs, archs): print("%s" % AUTOGEN, file=f) callback_lines = [] for line in _gen_c_intrinsics_generator( intrs, _is_translator_compatible_assembler, True): callback_lines.append(line) # Include definitions of registers for appropriate type of bindings print('#include "berberis/device_arch_info/%s/device_arch_info.h"' % archs[-1], file = f) # Put implementation into arch-specific namespace to access bindings. print('namespace berberis{\n\nnamespace %s::intrinsics::bindings {' % archs[-1], file = f) print(""" template using FunctionCompareTag = berberis::intrinsics::bindings::FunctionCompareTag; template constexpr Result ProcessBindings(Callback callback, Result def_result, Args&&... args) { using berberis::intrinsics::Float16; using berberis::intrinsics::Float32; using berberis::intrinsics::Float64;""", file=f) for line in callback_lines: print(line, file=f) print(""" } return std::forward(def_result); } } // namespace %s::intrinsics::bindings namespace intrinsics::bindings { using %s::intrinsics::bindings::ProcessBindings; } // namespace intrinsics::bindings } // namespace berberis """ % (archs[-1], archs[-1]), file=f) def _gen_c_intrinsics_generator(intrs, check_compatible_assembler, gen_builder): processed_names = set() for name, intr in intrs: ins = intr.get('in') outs = intr.get('out') params = _get_in_params(ins) formal_args = ', '.join('%s %s' % (type, param) for type, param in params) result_type, _ = _get_result_type(outs) if 'asm' not in intr: continue if 'variants' in intr: variants = _get_formats_with_descriptions(intr) # Sort by index, to keep order close to what _gen_intrs_enum produces. variants = sorted(variants, key=lambda variant: variant[1].index) # Collect intr_asms for all versions of intrinsic. # Note: not all variants are guaranteed to have asm version! # If that happens list of intr_asms for that variant would be empty. variants = [(desc, [ intr_asm for intr_asm in _gen_sorted_asms(intr) if fmt in intr_asm['variants'] ]) for fmt, desc in variants] # Print intrinsic generator for desc, intr_asms in variants: if len(intr_asms) > 0: if 'raw' in intr['variants']: spec = '%d' % (desc.num_elements) else: spec = '%s, %d' % (desc.c_type, desc.num_elements) for intr_asm in intr_asms: for line in _gen_c_intrinsic('%s<%s>' % (name, spec), intr, intr_asm, processed_names, check_compatible_assembler, gen_builder): yield line else: for intr_asm in _gen_sorted_asms(intr): for line in _gen_c_intrinsic(name, intr, intr_asm, processed_names, check_compatible_assembler, gen_builder): yield line def _gen_sorted_asms(intr): return sorted(intr['asm'], key = lambda intr: intr.get('nan', '') + _KNOWN_FEATURES_KEYS.get( intr.get('feature', ''), intr.get('feature', '')), reverse = True) _KNOWN_FEATURES_KEYS = { 'LZCNT': '001', 'BMI': '002', 'BMI2': '003', 'SSE': '010', 'SSE2': '011', 'SSE3': '012', 'SSSE3': '013', 'SSE4a': '014', 'SSE4_1': '015', 'SSE4_2': '016', 'AVX': '017', 'AVX2': '018', 'AES': '019', 'AESAVX': '020', 'VAES': '021', 'CLMUL': '012', 'CLMULAVX': '023', 'VPCLMULQD': '024', 'F16C': '025', 'FMA': '026', 'FMA4': '027', 'CustomCapability': '999' } def _gen_c_intrinsic( name, intr, asm, processed_names, check_compatible_assembler, gen_builder): if not check_compatible_assembler(asm): return cpuid_restriction = gen_device_insn_info_lib._get_cpuid_restriction(asm) nan_restriction = 'berberis::intrinsics::bindings::NoNansOperation' if 'nan' in asm: nan_restriction = 'berberis::intrinsics::bindings::%sNanOperationsHandling' % asm['nan'] template_arg = 'true' if asm['nan'] == "Precise" else "false" if '<' in name: template_pos = name.index('<') name = name[0:template_pos+1] + template_arg + ", " + name[template_pos+1:] else: name += '<' + template_arg + '>' if name not in processed_names: if gen_builder: yield ' %s if constexpr (std::is_same_v,' % ( '' if len(processed_names) == 0 else ' } else' ) yield '%s FunctionCompareTag>) {' % (' ' * 36, name) processed_names.add(name) restriction = [cpuid_restriction, nan_restriction] if gen_builder: yield ' if (auto result = callback(' else: yield ' callback(' yield ' berberis::intrinsics::bindings::IntrinsicBindingInfo<' yield ' %s,' % ( ',\n '.join( ['"%s"' % name, nan_restriction, _get_c_type_tuple(intr['in']), _get_c_type_tuple(intr['out']), _get_bindings_info(asm['args'])])) yield ' device_arch_info::DeviceInsnInfo<%s>>(),' % ( ',\n '.join( [gen_device_insn_info_lib._get_asm_reference(asm), '"%s"' % asm['mnemo'], 'true' if _intr_has_side_effects(intr) else 'false', gen_device_insn_info_lib._get_opcode_reference(asm), cpuid_restriction, gen_device_insn_info_lib._get_reg_operands_info(asm['args'])])) if gen_builder: yield ' std::forward(args)...); result.has_value()) {' yield ' return *std::move(result);' yield ' }' else: yield ' std::forward(args)...);' def _get_c_type_tuple(arguments): return 'std::tuple<%s>' % ', '.join( _get_c_type(argument) for argument in arguments) def _load_intrs_def_files(intrs_def_files): result = {} for intrs_def in intrs_def_files: with open(intrs_def) as intrs: result.update(json.load(intrs)) result.pop('License', None) return result def _load_intrs_arch_def(intrs_defs): json_data = [] for intrs_def in intrs_defs: with open(intrs_def) as intrs: json_array = json.load(intrs) while isinstance(len(json_array) > 0 and json_array[0], str): json_array.pop(0) json_data.extend(json_array) return json_data def _load_macro_def(intrs, arch_intrs, insns_def, macroassembler): arch, insns = asm_defs.load_asm_defs(insns_def) for insn in insns: insn['macroassembler'] = macroassembler insns_map = dict((insn['name'], insn) for insn in insns) unprocessed_intrs = [] for arch_intr in arch_intrs: if arch_intr['insn'] in insns_map: insn = insns_map[arch_intr['insn']] _add_asm_insn(intrs, arch_intr, insn) else: unprocessed_intrs.append(arch_intr) return arch, unprocessed_intrs def _is_interpreter_compatible_assembler(intr_asm): if intr_asm.get('usage', '') == 'inline-only': return False return True def _is_translator_compatible_assembler(intr_asm): if intr_asm.get('usage', '') == 'no-inline': return False return True def _add_asm_insn(intrs, arch_intr, insn): name = ','.join(name_part.strip() for name_part in arch_intr['name'].split(',')) # Sanity checks: MacroInstruction could implement few different intrinsics but # number of arguments in arch intrinsic and arch-independent intrinsic # should match. # # Note: we allow combining intrinsics with variants and intrinsics without # variants (e.g. AbsF32 is combined with VectorAbsoluteFP for F32x2 and F32x4), # but don't allow macroinstructions which would handle different set of # variants for different intrinsics. assert 'variants' not in insn or insn['variants'] == arch_intr['variants'] assert 'feature' not in insn or insn['feature'] == arch_intr['feature'] assert 'nan' not in insn or insn['nan'] == arch_intr['nan'] assert 'usage' not in insn or insn['usage'] == arch_intr['usage'] # Some intrinsics have extra inputs which can be ignored. e,g fpcr could be # ignored when not needed for precise emulation of NaNs. # Therefore we check that number inputs to (macro) instruction is less than # or equal to number of inputs to number of inputs to intrinsic. assert len(intrs[name]['in']) >= len(arch_intr['in']) assert len(intrs[name]['out']) == len(arch_intr['out']) if 'variants' in arch_intr: insn['variants'] = arch_intr['variants'] if 'feature' in arch_intr: insn['feature'] = arch_intr['feature'] if 'nan' in arch_intr: insn['nan'] = arch_intr['nan'] if 'usage' in arch_intr: insn['usage'] = arch_intr['usage'] for count, in_arg in enumerate(arch_intr['in']): # Sanity check: each in argument should only be used once - but if two # different intrinsics use them same macroinstruction it could be already # defined... yet it must be defined identically. assert ('ir_arg' not in insn['args'][in_arg] or insn['args'][in_arg]['ir_arg'] == count) insn['args'][in_arg]['ir_arg'] = count for count, out_arg in enumerate(arch_intr['out']): # Sanity check: each out argument should only be used once, too. assert ('ir_res' not in insn['args'][out_arg] or insn['args'][out_arg]['ir_res'] == count) insn['args'][out_arg]['ir_res'] = count # Note: one intrinsic could have more than one implementation (e.g. # SSE2 vs SSE4.2). if 'asm' not in intrs[name]: intrs[name]['asm'] = [] intrs[name]['asm'].append(insn) def _open_asm_def_files(def_files, arch_def_files, asm_def_files): intrs = _load_intrs_def_files(def_files) expanded_intrs = _expand_template_intrinsics(intrs) arch_intrs = _load_intrs_arch_def(arch_def_files) archs = [] assemblers = 0 for macro_def in asm_def_files: arch, arch_intrs = _load_macro_def(expanded_intrs, arch_intrs, macro_def, assemblers) if arch is not None: archs.append(arch) assemblers += 1 # Make sure that all intrinsics were found during processing of arch_intrs. assert arch_intrs == [] return archs, sorted(intrs.items()), sorted(expanded_intrs.items()) def _expand_template_intrinsics(intrs): expanded_intrs = {} for name, intr in intrs.items(): if intr.get('class') != 'template': expanded_intrs[name] = intr else: variants = intr.get('variants').copy() # These intrinsics serve a dual duty: # 1. They are used to implement regular instructions. # 2. They are used to implement vector instrinsics. # And there's a dilemma: 8bit and 16bit sizes are not supported in #1 but # are needed for #2. # This hack helps us to resolve it. If we would have more such intrinsics # we may need to extend JSON capabilities. if name in ('Aadd', 'Asub', 'DivRiscV', 'RemRiscV', 'Roundoff'): if 'int8_t' not in variants: variants += ['int8_t'] if 'uint8_t' not in variants: variants += ['uint8_t'] if 'int16_t' not in variants: variants += ['int16_t'] if 'uint16_t' not in variants: variants += ['uint16_t'] for variant in variants: types = {} params = [param.strip() for param in variant.split(',')] for param in params: if param in ('true', 'false'): continue if re.search('[_a-zA-Z]', param): types['Type'+str(len(types))] = param new_intr = intr.copy() del new_intr['variants'] new_intr['in'] = [types.get(param, param) for param in new_intr.get('in')] new_intr['out'] = [types.get(param, param) for param in new_intr.get('out')] expanded_intrs[name+'<'+','.join(params)+'>'] = new_intr return expanded_intrs def main(argv): # Usage: # gen_intrinsics.py --public_headers # # # # , # # , # ... # , # ... # gen_intrinsics.py --text_asm_intrinsics_bindings # , # ... # , # ... def open_out_file(name): try: os.makedirs(os.path.dirname(name)) except: pass return open(name, 'w') # Temporary special case for riscv64 to arm64. # TODO(b/362520361): generalize and combine with the below. option = argv[1] if option == 'arm64': mode = argv[2] out_files_end = 6 def_files_end = out_files_end while argv[def_files_end].endswith('intrinsic_def.json'): def_files_end += 1 if (def_files_end == len(argv)): break intrs = sorted(_load_intrs_def_files(argv[out_files_end:def_files_end]).items()) _gen_intrinsics_inl_h(open_out_file(argv[3]), intrs) _gen_semantic_player_types(intrs) _gen_interpreter_intrinsics_hooks_impl_inl_h(open_out_file(argv[4]), intrs, option) _gen_demultiplexer_intrinsics_hooks_impl_inl_h(open_out_file(argv[5]), intrs) return 0 mode = argv[1] if mode in ('--text_asm_intrinsics_bindings', '--public_headers'): out_files_end = 3 if mode == '--text_asm_intrinsics_bindings' else 8 def_files_end = out_files_end while argv[def_files_end].endswith('intrinsic_def.json'): def_files_end += 1 arch_def_files_end = def_files_end while argv[arch_def_files_end].endswith('machine_ir_intrinsic_binding.json'): arch_def_files_end += 1 archs, intrs, expanded_intrs = _open_asm_def_files( argv[out_files_end:def_files_end], argv[def_files_end:arch_def_files_end], argv[arch_def_files_end:]) if mode == '--text_asm_intrinsics_bindings': _gen_process_all_bindings(open_out_file(argv[2]), expanded_intrs, archs) else: _gen_intrinsics_inl_h(open_out_file(argv[2]), intrs) _gen_process_bindings(open_out_file(argv[3]), expanded_intrs, archs) _gen_semantic_player_types(intrs) _gen_interpreter_intrinsics_hooks_impl_inl_h(open_out_file(argv[4]), intrs, '') _gen_translator_intrinsics_hooks_impl_inl_h( open_out_file(argv[5]), intrs) _gen_demultiplexer_intrinsics_hooks_impl_inl_h( open_out_file(argv[6]), intrs) _gen_mock_semantics_listener_intrinsics_hooks_impl_inl_h( open_out_file(argv[7]), intrs) else: assert False, 'unknown option %s' % (mode) return 0 if __name__ == '__main__': sys.exit(main(sys.argv))