+"""core of the python-based POWER9 simulator
+
+this is part of a cycle-accurate POWER9 simulator. its primary purpose is
+not speed, it is for both learning and educational purposes, as well as
+a method of verifying the HDL.
+"""
+
from functools import wraps
from soc.decoder.orderedset import OrderedSet
from soc.decoder.selectable_int import (FieldSelectableInt, SelectableInt,
selectconcat)
-from soc.decoder.power_enums import spr_dict, XER_bits
-from soc.decoder.helpers import exts
+from soc.decoder.power_enums import spr_dict, XER_bits, insns, InternalOp
+from soc.decoder.helpers import exts, trunc_div, trunc_rem
from collections import namedtuple
import math
+import sys
instruction_info = namedtuple('instruction_info',
'func read_regs uninit_regs write_regs ' + \
'VRSAVE': 256}
+def swap_order(x, nbytes):
+ x = x.to_bytes(nbytes, byteorder='little')
+ x = int.from_bytes(x, byteorder='big', signed=False)
+ return x
+
+
def create_args(reglist, extra=None):
args = OrderedSet()
for reg in reglist:
class Mem:
- def __init__(self, bytes_per_word=8, initial_mem=None):
+ def __init__(self, row_bytes=8, initial_mem=None):
self.mem = {}
- self.bytes_per_word = bytes_per_word
- self.word_log2 = math.ceil(math.log2(bytes_per_word))
+ self.bytes_per_word = row_bytes
+ self.word_log2 = math.ceil(math.log2(row_bytes))
+ print ("Sim-Mem", initial_mem, self.bytes_per_word, self.word_log2)
if not initial_mem:
return
- print ("Sim-Mem", initial_mem, self.bytes_per_word)
+
+ # different types of memory data structures recognised (for convenience)
+ if isinstance(initial_mem, list):
+ initial_mem = (0, initial_mem)
+ if isinstance(initial_mem, tuple):
+ startaddr, mem = initial_mem
+ initial_mem = {}
+ for i, val in enumerate(mem):
+ initial_mem[startaddr + row_bytes*i] = (val, row_bytes)
+
for addr, (val, width) in initial_mem.items():
- self.st(addr, val, width)
+ #val = swap_order(val, width)
+ self.st(addr, val, width, swap=False)
def _get_shifter_mask(self, wid, remainder):
shifter = ((self.bytes_per_word - wid) - remainder) * \
8 # bits per byte
+ # XXX https://bugs.libre-soc.org/show_bug.cgi?id=377
+ # BE/LE mode?
+ shifter = remainder * 8
mask = (1 << (wid * 8)) - 1
print ("width,rem,shift,mask", wid, remainder, hex(shifter), hex(mask))
return shifter, mask
# TODO: Implement ld/st of lesser width
- def ld(self, address, width=8):
+ def ld(self, address, width=8, swap=True, check_in_mem=False):
print("ld from addr 0x{:x} width {:d}".format(address, width))
remainder = address & (self.bytes_per_word - 1)
address = address >> self.word_log2
assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
if address in self.mem:
val = self.mem[address]
+ elif check_in_mem:
+ return None
else:
val = 0
print("mem @ 0x{:x} rem {:d} : 0x{:x}".format(address, remainder, val))
print ("masking", hex(val), hex(mask<<shifter), shifter)
val = val & (mask << shifter)
val >>= shifter
+ if swap:
+ val = swap_order(val, width)
print("Read 0x{:x} from addr 0x{:x}".format(val, address))
return val
- def st(self, addr, v, width=8):
+ def st(self, addr, v, width=8, swap=True):
+ staddr = addr
remainder = addr & (self.bytes_per_word - 1)
addr = addr >> self.word_log2
- print("Writing 0x{:x} to addr 0x{:x}/{:x}".format(v, addr, remainder))
+ print("Writing 0x{:x} to ST 0x{:x} memaddr 0x{:x}/{:x}".format(v,
+ staddr, addr, remainder, swap))
assert remainder & (width - 1) == 0, "Unaligned access unsupported!"
+ if swap:
+ v = swap_order(v, width)
if width != self.bytes_per_word:
if addr in self.mem:
val = self.mem[addr]
def __call__(self, ridx):
return self[ridx]
-
-
+
class ISACaller:
# decoder2 - an instance of power_decoder2
# regfile - a list of initial values for the registers
+ # initial_{etc} - initial values for SPRs, Condition Register, Mem, MSR
+ # respect_pc - tracks the program counter. requires initial_insns
def __init__(self, decoder2, regfile, initial_sprs=None, initial_cr=0,
- initial_mem=None, initial_msr=0):
+ initial_mem=None, initial_msr=0,
+ initial_insns=None, respect_pc=False,
+ disassembly=None):
+
+ self.respect_pc = respect_pc
if initial_sprs is None:
initial_sprs = {}
if initial_mem is None:
initial_mem = {}
+ if initial_insns is None:
+ initial_insns = {}
+ assert self.respect_pc == False, "instructions required to honor pc"
+
+ print ("ISACaller insns", respect_pc, initial_insns, disassembly)
+
+ # "fake program counter" mode (for unit testing)
+ self.fake_pc = 0
+ if not respect_pc:
+ if isinstance(initial_mem, tuple):
+ self.fake_pc = initial_mem[0]
+
+ # disassembly: we need this for now (not given from the decoder)
+ self.disassembly = {}
+ if disassembly:
+ for i, code in enumerate(disassembly):
+ self.disassembly[i*4 + self.fake_pc] = code
+
+ # set up registers, instruction memory, data memory, PC, SPRs, MSR
self.gpr = GPR(decoder2, regfile)
- self.mem = Mem(initial_mem=initial_mem)
+ self.mem = Mem(row_bytes=8, initial_mem=initial_mem)
+ self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
self.pc = PC()
self.spr = SPR(decoder2, initial_sprs)
self.msr = SelectableInt(initial_msr, 64) # underlying reg
+
# TODO, needed here:
# FPR (same as GPR except for FP nums)
# 4.2.2 p124 FPSCR (definitely "separate" - not in SPR)
def TRAP(self, trap_addr=0x700):
print ("TRAP: TODO")
+ #self.namespace['NIA'] = trap_addr
+ #self.SRR0 = self.namespace['CIA'] + 4
+ #self.SRR1 = self.namespace['MSR']
+ #self.namespace['MSR'][45] = 1
# store CIA(+4?) in SRR0, set NIA to 0x700
# store MSR in SRR1, set MSR to um errr something, have to check spec
self.namespace['CA'] = self.spr['XER'][XER_bits['CA']].value
self.namespace['CA32'] = self.spr['XER'][XER_bits['CA32']].value
- def handle_carry_(self, inputs, outputs):
+ def handle_carry_(self, inputs, outputs, already_done):
inv_a = yield self.dec2.e.invert_a
if inv_a:
inputs[0] = ~inputs[0]
imm = yield self.dec2.e.imm_data.data
inputs.append(SelectableInt(imm, 64))
assert len(outputs) >= 1
- output = outputs[0]
- gts = [(x > output) for x in inputs]
+ print ("outputs", repr(outputs))
+ if isinstance(outputs, list) or isinstance(outputs, tuple):
+ output = outputs[0]
+ else:
+ output = outputs
+ gts = []
+ for x in inputs:
+ print ("gt input", x, output)
+ gt = (x > output)
+ gts.append(gt)
print(gts)
cy = 1 if any(gts) else 0
- self.spr['XER'][XER_bits['CA']] = cy
+ if not (1 & already_done):
+ self.spr['XER'][XER_bits['CA']] = cy
print ("inputs", inputs)
# 32 bit carry
- gts = [(x[32:64] > output[32:64]) == SelectableInt(1, 1)
- for x in inputs]
+ gts = []
+ for x in inputs:
+ print ("input", x, output)
+ gt = (x[32:64] > output[32:64]) == SelectableInt(1, 1)
+ gts.append(gt)
cy32 = 1 if any(gts) else 0
- self.spr['XER'][XER_bits['CA32']] = cy32
+ if not (2 & already_done):
+ self.spr['XER'][XER_bits['CA32']] = cy32
- def handle_overflow(self, inputs, outputs):
+ def handle_overflow(self, inputs, outputs, div_overflow):
inv_a = yield self.dec2.e.invert_a
if inv_a:
inputs[0] = ~inputs[0]
imm = yield self.dec2.e.imm_data.data
inputs.append(SelectableInt(imm, 64))
assert len(outputs) >= 1
- if len(inputs) >= 2:
+ print ("handle_overflow", inputs, outputs, div_overflow)
+ if len(inputs) < 2 and div_overflow != 1:
+ return
+
+ # div overflow is different: it's returned by the pseudo-code
+ # because it's more complex than can be done by analysing the output
+ if div_overflow == 1:
+ ov, ov32 = 1, 1
+ # arithmetic overflow can be done by analysing the input and output
+ elif len(inputs) >= 2:
output = outputs[0]
# OV (64-bit)
ov32 = 1 if input32_sgn[0] == input32_sgn[1] and \
output32_sgn != input32_sgn[0] else 0
- self.spr['XER'][XER_bits['OV']] = ov
- self.spr['XER'][XER_bits['OV32']] = ov32
- so = self.spr['XER'][XER_bits['SO']]
- so = so | ov
- self.spr['XER'][XER_bits['SO']] = so
-
-
+ self.spr['XER'][XER_bits['OV']] = ov
+ self.spr['XER'][XER_bits['OV32']] = ov32
+ so = self.spr['XER'][XER_bits['SO']]
+ so = so | ov
+ self.spr['XER'][XER_bits['SO']] = so
def handle_comparison(self, outputs):
out = outputs[0]
def set_pc(self, pc_val):
self.namespace['NIA'] = SelectableInt(pc_val, 64)
self.pc.update(self.namespace)
-
+
+ def setup_one(self):
+ """set up one instruction
+ """
+ if self.respect_pc:
+ pc = self.pc.CIA.value
+ else:
+ pc = self.fake_pc
+ self._pc = pc
+ ins = self.imem.ld(pc, 4, False, True)
+ if ins is None:
+ raise KeyError("no instruction at 0x%x" % pc)
+ print("setup: 0x%x 0x%x %s" % (pc, ins & 0xffffffff, bin(ins)))
+ print ("NIA, CIA", self.pc.CIA.value, self.pc.NIA.value)
+
+ yield self.dec2.dec.raw_opcode_in.eq(ins & 0xffffffff)
+ yield self.dec2.dec.bigendian.eq(0) # little / big?
+
+ def execute_one(self):
+ """execute one instruction
+ """
+ # get the disassembly code for this instruction
+ code = self.disassembly[self._pc]
+ print("sim-execute", hex(self._pc), code)
+ opname = code.split(' ')[0]
+ yield from self.call(opname)
+
+ if not self.respect_pc:
+ self.fake_pc += 4
+ print ("NIA, CIA", self.pc.CIA.value, self.pc.NIA.value)
+
+ def get_assembly_name(self):
+ # TODO, asmregs is from the spec, e.g. add RT,RA,RB
+ # see http://bugs.libre-riscv.org/show_bug.cgi?id=282
+ asmcode = yield self.dec2.dec.op.asmcode
+ asmop = insns.get(asmcode, None)
+
+ # sigh reconstruct the assembly instruction name
+ ov_en = yield self.dec2.e.oe.oe
+ ov_ok = yield self.dec2.e.oe.ok
+ if ov_en & ov_ok:
+ asmop += "."
+ lk = yield self.dec2.e.lk
+ if lk:
+ asmop += "l"
+ int_op = yield self.dec2.dec.op.internal_op
+ print ("int_op", int_op)
+ if int_op in [InternalOp.OP_B.value, InternalOp.OP_BC.value]:
+ AA = yield self.dec2.dec.fields.FormI.AA[0:-1]
+ print ("AA", AA)
+ if AA:
+ asmop += "a"
+ if int_op == InternalOp.OP_MFCR.value:
+ dec_insn = yield self.dec2.e.insn
+ if dec_insn & (1<<20) != 0: # sigh
+ asmop = 'mfocrf'
+ else:
+ asmop = 'mfcr'
+ # XXX TODO: for whatever weird reason this doesn't work
+ # https://bugs.libre-soc.org/show_bug.cgi?id=390
+ if int_op == InternalOp.OP_MTCRF.value:
+ dec_insn = yield self.dec2.e.insn
+ if dec_insn & (1<<20) != 0: # sigh
+ asmop = 'mtocrf'
+ else:
+ asmop = 'mtcrf'
+ return asmop
def call(self, name):
# TODO, asmregs is from the spec, e.g. add RT,RA,RB
# see http://bugs.libre-riscv.org/show_bug.cgi?id=282
+ asmop = yield from self.get_assembly_name()
+ print ("call", name, asmop)
+ if name not in ['mtcrf', 'mtocrf']:
+ assert name == asmop, "name %s != %s" % (name, asmop)
+
info = self.instrs[name]
yield from self.prep_namespace(info.form, info.op_fields)
results = info.func(self, *inputs)
print(results)
+ # detect if CA/CA32 already in outputs (sra*, basically)
+ already_done = 0
+ if info.write_regs:
+ output_names = create_args(info.write_regs)
+ for name in output_names:
+ if name == 'CA':
+ already_done |= 1
+ if name == 'CA32':
+ already_done |= 2
+
+ print ("carry already done?", bin(already_done))
carry_en = yield self.dec2.e.output_carry
if carry_en:
- yield from self.handle_carry_(inputs, results)
- ov_en = yield self.dec2.e.oe
- if ov_en:
- yield from self.handle_overflow(inputs, results)
+ yield from self.handle_carry_(inputs, results, already_done)
+
+ # detect if overflow was in return result
+ overflow = None
+ if info.write_regs:
+ for name, output in zip(output_names, results):
+ if name == 'overflow':
+ overflow = output
+
+ ov_en = yield self.dec2.e.oe.oe
+ ov_ok = yield self.dec2.e.oe.ok
+ print ("internal overflow", overflow)
+ if ov_en & ov_ok:
+ yield from self.handle_overflow(inputs, results, overflow)
+
rc_en = yield self.dec2.e.rc.data
if rc_en:
self.handle_comparison(results)
# any modified return results?
if info.write_regs:
- output_names = create_args(info.write_regs)
for name, output in zip(output_names, results):
+ if name == 'overflow': # ignore, done already (above)
+ continue
if isinstance(output, int):
output = SelectableInt(output, 256)
if name in ['CA', 'CA32']: