xenia-canary/tools/ppc-table-gen.py

526 lines
17 KiB
Python

#!/usr/bin/env python
# Copyright 2015 Ben Vanik & shuffle2. All Rights Reserved.
"""PPC instruction table generator.
Generates various headers/sources for looking up and handling PPC instructions.
This is based on shuffle2's PPC generator:
https://gist.github.com/shuffle2/10015968
"""
__author__ = 'ben.vanik@gmail.com (Ben Vanik)'
import os
import sys
from xml.etree.ElementTree import ElementTree, Element, SubElement, tostring, dump
self_path = os.path.dirname(os.path.abspath(__file__))
def bit_extract(x, leftmost, rightmost):
return (x >> (32 - 1 - rightmost)) & ((1 << (rightmost - leftmost + 1)) - 1)
extended_opcode_bits = {
'X': [(21, 30)],
'XL': [(21, 30)],
'XFX': [(21, 30)],
'XFL': [(21, 30)],
'VX': [(21, 31)],
'VX128': [(22, 25), (27, 27)],
'VX128_1': [(21, 27), (30, 31)],
'VX128_2': [(22, 22), (27, 27)],
'VX128_3': [(21, 27)],
'VX128_4': [(21, 23), (26, 27)],
'VX128_5': [(27, 27)],
'VX128_R': [(22, 24), (27, 27)],
'VX128_P': [(21, 22), (26, 27)],
'VC': [(22, 31)],
'VA': [(26, 31)],
'XO': [(22, 30)],
'XW': [(25, 30)],
'A': [(26, 30)],
'DS': [(30, 31)],
'MD': [(27, 29)],
'MDS': [(27, 30)],
'XS': [(21, 29)],
'DCBZ': [(6, 10), (21, 30)], # like X
}
def opcode_primary(insn):
return bit_extract(insn, 0, 5)
def opcode_extended(insn, form):
if form in extended_opcode_bits:
parts = extended_opcode_bits[form]
value = 0
shift = 0
for part in parts:
shift = max(shift, part[1])
for part in parts:
part_value = bit_extract(insn, part[0], part[1])
value = value | (part_value << (shift - part[1]))
return value
else:
return -1
class Insn:
pass
def parse_insns(filename):
root = ElementTree(file = filename)
insns = []
# Convert to python types
for e in root.findall('.//insn'):
i = Insn()
i.opcode = int(e.attrib['opcode'], 16)
i.mnem = e.attrib['mnem']
i.form = e.attrib['form']
i.group = e.attrib['group']
i.desc = e.attrib['desc']
i.type = 'General'
if 'sync' in e.attrib and e.attrib['sync'] == 'true':
i.type = 'Sync'
i.op_primary = opcode_primary(i.opcode)
i.op_extended = opcode_extended(i.opcode, i.form)
i.reads = []
i.writes = []
for r in e.findall('.//in'):
is_conditional = 'conditional' in r.attrib and r.attrib['conditional'] == 'true'
i.reads.append((r.attrib['field'], is_conditional))
for w in e.findall('.//out'):
is_conditional = 'conditional' in w.attrib and w.attrib['conditional'] == 'true'
i.writes.append((w.attrib['field'], is_conditional))
i.disasm_str = None
for d in e.findall('.//disasm'):
i.disasm_str = d.text
insns.append(i)
return insns
def c_mnem(x):
return x.replace('.', 'x')
def c_group(x):
return 'k' + x[0].upper() + x[1:]
def c_bool(x):
return 'true' if x else 'false'
def c_field(x):
base_name = 'k' + x[0]
cond_name = 'cond' if x[1] else ''
return base_name + cond_name
def generate_opcodes(insns):
l = []
TAB = ' ' * 2
def w0(x): l.append(x)
def w1(x): w0(TAB * 1 + x)
def w2(x): w0(TAB * 2 + x)
def w3(x): w0(TAB * 3 + x)
w0('// This code was autogenerated by %s. Do not modify!' % (sys.argv[0]))
w0('// clang-format off')
w0('#ifndef XENIA_CPU_PPC_PPC_OPCODE_H_')
w0('#define XENIA_CPU_PPC_PPC_OPCODE_H_')
w0('')
w0('#include <cstdint>')
w0('')
w0('namespace xe {')
w0('namespace cpu {')
w0('namespace ppc {')
w0('')
for i in insns:
i.mnem = c_mnem(i.mnem)
insns = sorted(insns, key = lambda i: i.mnem)
w0('// All PPC opcodes in the same order they appear in ppc_opcode_table.h:')
w0('enum class PPCOpcode : uint32_t {')
for i in insns:
w1('%s,' % (i.mnem))
w1('kInvalid,')
w0('};')
w0('')
w0('} // namespace ppc')
w0('} // namespace cpu')
w0('} // namespace xe')
w0('')
w0('#endif // XENIA_CPU_PPC_PPC_OPCODE_H_')
w0('')
return '\n'.join(l)
def generate_table(insns):
l = []
TAB = ' ' * 2
def w0(x): l.append(x)
def w1(x): w0(TAB * 1 + x)
def w2(x): w0(TAB * 2 + x)
def w3(x): w0(TAB * 3 + x)
w0('// This code was autogenerated by %s. Do not modify!' % (sys.argv[0]))
w0('// clang-format off')
w0('#include <cstdint>')
w0('')
w0('#include "xenia/base/assert.h"')
w0('#include "xenia/cpu/ppc/ppc_opcode.h"')
w0('#include "xenia/cpu/ppc/ppc_opcode_info.h"')
w0('')
w0('namespace xe {')
w0('namespace cpu {')
w0('namespace ppc {')
w0('')
for i in insns:
i.mnem = '"' + c_mnem(i.mnem) + '"'
i.form = c_group(i.form)
i.group = c_group(i.group)
i.type = c_group(i.type)
mnem_len = len(max(insns, key = lambda i: len(i.mnem)).mnem)
form_len = len(max(insns, key = lambda i: len(i.form)).form)
group_len = len(max(insns, key = lambda i: len(i.group)).group)
type_len = len(max(insns, key = lambda i: len(i.type)).type)
insns = sorted(insns, key = lambda i: i.mnem)
w0('#define INSTRUCTION(opcode, mnem, form, group, type) \\')
w0(' {PPCOpcodeType::type, nullptr}')
w0('PPCOpcodeInfo ppc_opcode_table[] = {')
fmt = 'INSTRUCTION(' + ', '.join([
'0x%08x',
'%-' + str(mnem_len) + 's',
'%-' + str(form_len) + 's',
'%-' + str(group_len) + 's',
'%-' + str(type_len) + 's',
]) + '),'
for i in insns:
w1(fmt % (i.opcode, i.mnem, i.form, i.group, i.type))
w0('};')
w0('static_assert(sizeof(ppc_opcode_table) / sizeof(PPCOpcodeInfo) == static_cast<int>(PPCOpcode::kInvalid), "PPC table mismatch - rerun ppc-table-gen");')
w0('')
w0('const PPCOpcodeInfo& GetOpcodeInfo(PPCOpcode opcode) {')
w1('return ppc_opcode_table[static_cast<int>(opcode)];')
w0('}')
w0('void RegisterOpcodeEmitter(PPCOpcode opcode, InstrEmitFn fn) {')
w1('assert_null(ppc_opcode_table[static_cast<int>(opcode)].emit);')
w1('ppc_opcode_table[static_cast<int>(opcode)].emit = fn;')
w0('}')
w0('')
w0('} // namespace ppc')
w0('} // namespace cpu')
w0('} // namespace xe')
w0('')
return '\n'.join(l)
def literal_mnem(x):
x = x.replace('.', '_')
x = x.replace('"', '')
return x
def generate_token_append(i, token):
# Rc = . iff Rc=1
# OE = o iff OE=1
if token == 'Rc':
return 'if (d.%s.Rc()) str->Append(\'.\');' % (i.o_form)
elif token == 'OE':
return 'if (d.%s.OE()) str->Append(\'o\');' % (i.o_form)
elif token == 'LK':
return 'if (d.%s.LK()) str->Append(\'l\');' % (i.o_form)
elif token == 'AA':
return 'if (d.%s.AA()) str->Append(\'a\');' % (i.o_form)
elif token in ['RA', 'RA0', 'RB', 'RC', 'RT', 'RS', 'RD']:
return 'str->AppendFormat("r%%d", d.%s.%s());' % (i.o_form, token)
elif token in ['FA', 'FB', 'FC', 'FT', 'FS', 'FD']:
return 'str->AppendFormat("fr%%d", d.%s.%s());' % (i.o_form, token)
elif token in ['VA', 'VB', 'VC', 'VT', 'VS', 'VD']:
return 'str->AppendFormat("vr%%d", d.%s.%s());' % (i.o_form, token)
elif token in ['CRFD', 'CRFS']:
return 'str->AppendFormat("crf%%d", d.%s.%s());' % (i.o_form, token)
elif token in ['CRBA', 'CRBB', 'CRBD']:
return 'str->AppendFormat("crb%%d", d.%s.%s());' % (i.o_form, token)
elif token in ['BO', 'BI', 'TO', 'SPR', 'TBR', 'L', 'FM', 'MB', 'ME', 'SH', 'IMM', 'z']:
return 'str->AppendFormat("%%d", d.%s.%s());' % (i.o_form, token)
elif token == 'UIMM':
return 'str->AppendFormat("0x%%X", d.%s.%s());' % (i.o_form, token)
elif token in ['d', 'ds', 'SIMM']:
return 'str->AppendFormat(d.%s.%s() < 0 ? "-0x%%X" : "0x%%X", std::abs(d.%s.%s()));' % (i.o_form, token, i.o_form, token)
elif token == 'ADDR':
return 'str->AppendFormat("0x%%X", d.%s.%s());' % (i.o_form, token)
return 'str->AppendFormat("(UNHANDLED %s)");' % token
def generate_disasm(insns):
l = []
TAB = ' ' * 2
def w0(x): l.append(x)
def w1(x): w0(TAB * 1 + x)
def w2(x): w0(TAB * 2 + x)
def w3(x): w0(TAB * 3 + x)
w0('// This code was autogenerated by %s. Do not modify!' % (sys.argv[0]))
w0('// clang-format off')
w0('#include <cstdint>')
w0('')
w0('#include "xenia/base/assert.h"')
w0('#include "xenia/cpu/ppc/ppc_decode_data.h"')
w0('#include "xenia/cpu/ppc/ppc_opcode.h"')
w0('#include "xenia/cpu/ppc/ppc_opcode_info.h"')
w0('')
w0('namespace xe {')
w0('namespace cpu {')
w0('namespace ppc {')
w0('')
w0('constexpr size_t kNamePad = 11;')
w0('const uint8_t kSpaces[kNamePad] = {0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20};')
w0('void PadStringBuffer(StringBuffer* str, size_t base, size_t pad) {')
w1('size_t added_len = str->length() - base;')
w1('if (added_len < pad) str->AppendBytes(kSpaces, kNamePad - added_len);')
w0('}')
w0('')
for i in insns:
i.mnem = '"' + c_mnem(i.mnem) + '"'
i.o_form = i.form
i.form = c_group(i.form)
i.desc = '"' + i.desc + '"'
i.group = c_group(i.group)
i.type = c_group(i.type)
mnem_len = len(max(insns, key = lambda i: len(i.mnem)).mnem)
form_len = len(max(insns, key = lambda i: len(i.form)).form)
desc_len = len(max(insns, key = lambda i: len(i.desc)).desc)
group_len = len(max(insns, key = lambda i: len(i.group)).group)
type_len = len(max(insns, key = lambda i: len(i.type)).type)
insns = sorted(insns, key = lambda i: i.mnem)
# TODO(benvanik): support alts:
# <disasm cond="![RA]">li [RD], [SIMM]</disasm>
for i in insns:
if not i.disasm_str:
continue
w0('void PrintDisasm_%s(const PPCDecodeData& d, StringBuffer* str) {' % (literal_mnem(i.mnem)))
w1('// ' + i.disasm_str)
w1('size_t str_start = str->length();')
current_str = ''
j = 0
first_space = False
while j < len(i.disasm_str):
c = i.disasm_str[j]
if c == '[':
if current_str:
w1('str->Append("%s");' % (current_str))
current_str = ''
token = i.disasm_str[j + 1 : i.disasm_str.index(']', j)]
j += len(token) + 1
w1(generate_token_append(i, token))
else:
if c == ' ' and not first_space:
if current_str:
w1('str->Append("%s");' % (current_str))
current_str = ''
w1('PadStringBuffer(str, str_start, kNamePad);')
first_space = True
else:
current_str += c
j += 1
if current_str:
w1('str->Append("%s");' % (current_str))
if not first_space:
w1('PadStringBuffer(str, str_start, kNamePad);')
w0('}')
w0('#define INIT_LIST(...) {__VA_ARGS__}')
w0('#define INSTRUCTION(opcode, mnem, form, group, type, desc, reads, writes, fn) \\')
w0(' {PPCOpcodeGroup::group, PPCOpcodeFormat::form, opcode, mnem, desc, INIT_LIST reads, INIT_LIST writes, fn}')
w0('PPCOpcodeDisasmInfo ppc_opcode_disasm_table[] = {')
fmt = 'INSTRUCTION(' + ', '.join([
'0x%08x',
'%-' + str(mnem_len) + 's',
'%-' + str(form_len) + 's',
'%-' + str(group_len) + 's',
'%-' + str(type_len) + 's',
'%-' + str(desc_len) + 's',
'(%s)',
'(%s)',
'%s',
]) + '),'
for i in insns:
w1(fmt % (
i.opcode,
i.mnem,
i.form,
i.group,
i.type,
i.desc,
','.join(['PPCOpcodeField::' + c_field(r) for r in i.reads]),
','.join(['PPCOpcodeField::' + c_field(w) for w in i.writes]),
('PrintDisasm_' + literal_mnem(i.mnem)) if i.disasm_str else 'nullptr',
))
w0('};')
w0('static_assert(sizeof(ppc_opcode_disasm_table) / sizeof(PPCOpcodeDisasmInfo) == static_cast<int>(PPCOpcode::kInvalid), "PPC table mismatch - rerun ppc-table-gen");')
w0('')
w0('const PPCOpcodeDisasmInfo& GetOpcodeDisasmInfo(PPCOpcode opcode) {')
w1('return ppc_opcode_disasm_table[static_cast<int>(opcode)];')
w0('}')
w0('void RegisterOpcodeDisasm(PPCOpcode opcode, InstrDisasmFn fn) {')
w1('assert_null(ppc_opcode_disasm_table[static_cast<int>(opcode)].disasm);')
w1('ppc_opcode_disasm_table[static_cast<int>(opcode)].disasm = fn;')
w0('}')
w0('')
w0('} // namespace ppc')
w0('} // namespace cpu')
w0('} // namespace xe')
w0('')
return '\n'.join(l)
def generate_lookup(insns):
l = []
TAB = ' ' * 2
def w0(x): l.append(x)
def w1(x): w0(TAB * 1 + x)
def w2(x): w0(TAB * 2 + x)
def w3(x): w0(TAB * 3 + x)
for i in insns:
i.mnem = c_mnem(i.mnem)
w0('// This code was autogenerated by %s. Do not modify!' % (sys.argv[0]))
w0('// clang-format off')
w0('#include <cstdint>')
w0('')
w0('#include "xenia/base/assert.h"')
w0('#include "xenia/cpu/ppc/ppc_opcode.h"')
w0('#include "xenia/cpu/ppc/ppc_opcode_info.h"')
w0('')
w0('namespace xe {')
w0('namespace cpu {')
w0('namespace ppc {')
w0('')
w0('constexpr uint32_t ExtractBits(uint32_t v, uint32_t a, uint32_t b) {')
w0(' return (v >> (32 - 1 - b)) & ((1 << (b - a + 1)) - 1);')
w0('}')
w0('')
w0('#define PPC_DECODER_MISS assert_always(); return PPCOpcode::kInvalid')
w0('#define PPC_DECODER_HIT(name) return PPCOpcode::name;')
w0('')
w0('PPCOpcode LookupOpcode(uint32_t code) {')
w1('switch (ExtractBits(code, 0, 5)) {')
subtables = {}
for i in sorted(insns, key = lambda i: i.op_primary):
if i.op_primary not in subtables: subtables[i.op_primary] = []
subtables[i.op_primary].append(i)
for pri in sorted(subtables.iterkeys()):
# all the extended encodings (which we care about) end with bit 30. So we want to
# do the rest of the seach by bitscanning left from bit 30. This is simulated
# in the C switch-statement by creating leafs for each extended opcode,
# sorted by bitlength shortest to longest.
if len(subtables[pri]) == 1:
for i in subtables[pri]:
# the primary opcode field fully identifies the opcode
w1('case %i: PPC_DECODER_HIT(%s);' % (i.op_primary, i.mnem))
continue
extract_groups = {}
for i in subtables[pri]:
form_parts = extended_opcode_bits[i.form]
shift = 0
for form_part in form_parts:
shift = max(shift, form_part[1])
extract_parts = []
for form_part in form_parts:
extract_parts.append('(ExtractBits(code, %s, %s) << %s)' % (form_part[0], form_part[1], shift - form_part[1]))
extract_expression = '|'.join(extract_parts)
if extract_expression not in extract_groups:
extract_groups[extract_expression] = (i.form, extract_expression, [])
extract_groups[extract_expression][2].append(i)
w1('case %i:' % (pri))
for extract_expression in sorted(extract_groups.iterkeys()):
(form, extract_expression, group_insns) = extract_groups[extract_expression]
bit_span_low = 31
bit_span_high = 0
form_parts = extended_opcode_bits[form]
for form_part in form_parts:
bit_span_low = min(bit_span_low, form_part[0])
bit_span_high = max(bit_span_high, form_part[1])
bit_count = bit_span_high - bit_span_low + 1
w2('switch (%s) {' % (extract_expression))
for i in sorted(group_insns, key=lambda i: i.op_extended):
w3('case 0b%s: PPC_DECODER_HIT(%s);' % (
('{:0'+str(bit_count)+'b}').format(i.op_extended),
i.mnem))
w2('}')
w2('PPC_DECODER_MISS;')
w1('default: PPC_DECODER_MISS;')
w1('}')
w0('}')
w0('')
w0('} // namespace ppc')
w0('} // namespace cpu')
w0('} // namespace xe')
w0('')
# from this we can see some tables have bits which can be used to determine extended opcoded size:
# primary opcode 31:
# 01... = 9 bits (XO form), else 10 bits (X/XFX forms)
# primary opcode 63:
# 1.... = 7 bits (A form), else 10 bits (X/XFL forms)
# primary opcode 4:
# does not have small bit range to determine size, but you can just use the
# low 7 bits in order to "guess" the opcode. if you assume no invalid
# encodings are input, only the sequence ...0001000 actually *needs* the upper
# bits in order to differentiate the opcode (0100001000 = ps_abs, 0010001000 = ps_nabs)
# otherwise, the low 7bits can be used as the determinant, and a second comparison
# can be used against the real length of bits to fully match the extended opcode
#
# this approach can be generalized for all primary opcodes with extended opcodes of varying lengths:
# compare bits of smallest length, fall through to comparing larger sizes until found or failure
# with the optional optimization of discarding further compares for extended opcodes which
# share top bits with any other extended opcode (at the price of failing to detect invalid opcodes)
return '\n'.join(l)
if __name__ == '__main__':
ppc_src_path = os.path.join(self_path, '..', 'src', 'xenia', 'cpu', 'ppc')
insns = parse_insns(os.path.join(self_path, 'ppc-instructions.xml'))
with open(os.path.join(ppc_src_path, 'ppc_opcode.h'), 'w') as f:
f.write(generate_opcodes(insns))
insns = parse_insns(os.path.join(self_path, 'ppc-instructions.xml'))
with open(os.path.join(ppc_src_path, 'ppc_opcode_table.cc'), 'w') as f:
f.write(generate_table(insns))
insns = parse_insns(os.path.join(self_path, 'ppc-instructions.xml'))
with open(os.path.join(ppc_src_path, 'ppc_opcode_disasm.cc'), 'w') as f:
f.write(generate_disasm(insns))
insns = parse_insns(os.path.join(self_path, 'ppc-instructions.xml'))
with open(os.path.join(ppc_src_path, 'ppc_opcode_lookup.cc'), 'w') as f:
f.write(generate_lookup(insns))