| # Copyright lowRISC contributors. |
| # Licensed under the Apache License, Version 2.0, see LICENSE for details. |
| # SPDX-License-Identifier: Apache-2.0 |
| |
| from enum import IntEnum |
| from typing import Dict, List, Optional |
| |
| from shared.mem_layout import get_memory_layout |
| |
| from .csr import CSRFile |
| from .dmem import Dmem |
| from .constants import ErrBits, Status |
| from .ext_regs import OTBNExtRegs |
| from .flags import FlagReg |
| from .gpr import GPRs |
| from .loop import LoopStack |
| from .reg import RegFile |
| from .trace import Trace, TracePC |
| from .wsr import WSRFile |
| |
| |
| class FsmState(IntEnum): |
| r'''State of the internal start/stop FSM |
| |
| The FSM diagram looks like: |
| |
| /----------------------------------------------------------\ |
| | | |
| v /-> POST_EXEC --/ |
| IDLE -> PRE_EXEC -> FETCH_WAIT -> EXEC < |
| \-> LOCKING --\ |
| | |
| /----------------------------------------------------------/ |
| | |
| v |
| LOCKED |
| |
| IDLE represents the state when nothing is going on but there have been no |
| fatal errors. It matches Status.IDLE. LOCKED represents the state when |
| there has been a fatal error. It matches Status.LOCKED. |
| |
| PRE_EXEC, FETCH_WAIT, EXEC, POST_EXEC and LOCKING correspond to |
| Status.BUSY_EXECUTE. PRE_EXEC is the period after starting OTBN where we're |
| still waiting for an EDN value to seed URND. FETCH_WAIT is the single cycle |
| delay after seeding URND to fill the prefetch stage. EXEC is the period |
| where we start fetching and executing instructions. |
| |
| POST_EXEC and LOCKING are both used for the single cycle after we finish |
| executing where the STATUS register gets updated. The difference between |
| them is that POST_EXEC goes back to IDLE and LOCKING goes to LOCKED. |
| |
| This is a refinement of the Status enum and the integer values are picked |
| so that you can divide by 10 to get the corresponding Status entry. (This |
| isn't used in the code, but makes debugging slightly more convenient when |
| you just have the numeric values available). |
| |
| ''' |
| IDLE = 0 |
| PRE_EXEC = 10 |
| FETCH_WAIT = 11 |
| EXEC = 12 |
| POST_EXEC = 13 |
| LOCKING = 13 |
| LOCKED = 2550 |
| |
| |
| class OTBNState: |
| def __init__(self) -> None: |
| self.gprs = GPRs() |
| self.wdrs = RegFile('w', 256, 32) |
| |
| self.wsrs = WSRFile() |
| self.csrs = CSRFile() |
| |
| self.pc = 0 |
| self._pc_next_override = None # type: Optional[int] |
| |
| _, imem_size = get_memory_layout()['IMEM'] |
| self.imem_size = imem_size |
| |
| self.dmem = Dmem() |
| |
| self.fsm_state = FsmState.IDLE |
| |
| self.loop_stack = LoopStack() |
| self.ext_regs = OTBNExtRegs() |
| |
| self._err_bits = 0 |
| self.pending_halt = False |
| |
| self.rnd_256b_counter = 0 |
| self.urnd_256b_counter = 0 |
| |
| self.rnd_set_flag = False |
| |
| self.rnd_cdc_pending = False |
| self.urnd_cdc_pending = False |
| |
| self.rnd_cdc_counter = 0 |
| self.urnd_cdc_counter = 0 |
| |
| self.rnd_256b = 0 |
| self.urnd_256b = 4 * [0] |
| self.urnd_64b = 0 |
| |
| # This flag is set to true if we've injected integrity errors, trashing |
| # the whole of IMEM. The next fetch should fail. |
| self.invalidated_imem = False |
| |
| def get_next_pc(self) -> int: |
| if self._pc_next_override is not None: |
| return self._pc_next_override |
| return self.pc + 4 |
| |
| def set_next_pc(self, next_pc: int) -> None: |
| '''Overwrite the next program counter, e.g. as result of a jump.''' |
| assert(self.is_pc_valid(next_pc)) |
| self._pc_next_override = next_pc |
| |
| def edn_urnd_step(self, urnd_data: int) -> None: |
| # Take the new data |
| assert 0 <= urnd_data < (1 << 32) |
| |
| # There should not be a pending URND result before an EDN step. |
| assert not self.urnd_cdc_pending |
| |
| # Collect 32b packages in a 64b array of 4 elements |
| shift_num = 32 * (self.urnd_256b_counter % 2) |
| self.urnd_64b = self.urnd_64b | (urnd_data << shift_num) |
| |
| if self.urnd_256b_counter % 2: |
| idx = self.urnd_256b_counter // 2 |
| self.urnd_256b[idx] = self.urnd_64b |
| self.urnd_64b = 0 |
| |
| if self.urnd_256b_counter == 7: |
| # Reset the 32b package counter and wait until receiving done |
| # signal from RTL |
| self.urnd_256b_counter = 0 |
| self.urnd_cdc_pending = True |
| else: |
| # Count until 8 valid packages are received |
| self.urnd_256b_counter += 1 |
| return |
| |
| # Reset the 32b package counter and wait until receiving done |
| # signal from RTL |
| self.urnd_256b_counter = 0 |
| |
| def edn_rnd_step(self, rnd_data: int) -> None: |
| # Take the new data |
| assert 0 <= rnd_data < (1 << 32) |
| |
| # There should not be a pending RND result before an EDN step. |
| assert not self.rnd_cdc_pending |
| |
| # Collect 32b packages in a 256b variable |
| shift_num = 32 * self.rnd_256b_counter |
| self.rnd_256b = (self.rnd_256b | (rnd_data << shift_num)) & ((1 << 256) - 1) |
| |
| if self.rnd_256b_counter == 7: |
| # Reset the 32b package counter and wait until receiving done |
| # signal from RTL |
| self.rnd_256b_counter = 0 |
| self.rnd_cdc_pending = True |
| else: |
| # Count until 8 valid packages are received |
| self.rnd_256b_counter += 1 |
| return |
| |
| # Reset the 32b package counter and wait until receiving done |
| # signal from RTL |
| self.rnd_256b_counter = 0 |
| |
| def edn_flush(self) -> None: |
| # EDN Flush gets called after a reset signal from EDN clock domain |
| # arrives. It clears out internals of the model regarding EDN data |
| # processing on both RND and URND side. |
| self.rnd_256b = 0 |
| self.rnd_cdc_pending = False |
| self.rnd_cdc_counter = 0 |
| self.rnd_256b_counter = 0 |
| |
| self.urnd_64b = 0 |
| self.urnd_256b = 4 * [0] |
| self.urnd_256b_counter = 0 |
| self.urnd_cdc_pending = False |
| self.urnd_cdc_counter = 0 |
| |
| def rnd_reg_set(self) -> None: |
| # This sets RND register inside WSR immediately. Calling this with |
| # using DPI causes timing problems so it is best to it when we want |
| # to actually set RND register (at commit method and at the start if |
| # RND processing is completed after OTBN stopped running) |
| if self.rnd_set_flag: |
| self.wsrs.RND.set_unsigned(self.rnd_256b) |
| self.rnd_256b = 0 |
| self.rnd_cdc_pending = False |
| self.rnd_set_flag = False |
| self.rnd_cdc_counter = 0 |
| |
| def rnd_completed(self) -> None: |
| # This will be called when all the packages are received and processed |
| # by RTL. Model will set RND register, pending flag and internal |
| # variables will be cleared. |
| |
| # These must be true since model calculates RND data faster than RTL. |
| # But the synchronisation of the data should not take more than |
| # 5 cycles ideally. |
| assert self.rnd_cdc_counter < 6 |
| |
| # TODO: Assert rnd_cdc_pending when request is correctly modelled. |
| self.rnd_set_flag = True |
| |
| def urnd_completed(self) -> None: |
| # URND completed gets called after RTL signals that the processing |
| # of incoming EDN data is done. This also sets up EXEC state of the |
| # FSM of the model. This includes a dirty hack which disables |
| # fsm_state assertion because we are always calling this method |
| # while we are doing system level tests. This will be removed after |
| # request modelling of EDN is done. |
| assert self.urnd_cdc_counter < 6 |
| |
| # TODO: Assert urnd_cdc_pending when request is correctly modelled. |
| self.wsrs.URND.set_seed(self.urnd_256b) |
| self.urnd_256b = 4 * [0] |
| self.urnd_cdc_pending = False |
| self.urnd_cdc_counter = 0 |
| |
| def loop_start(self, iterations: int, bodysize: int) -> None: |
| self.loop_stack.start_loop(self.pc + 4, iterations, bodysize) |
| |
| def loop_step(self, loop_warps: Dict[int, int]) -> None: |
| back_pc = self.loop_stack.step(self.pc, loop_warps) |
| if back_pc is not None: |
| self.set_next_pc(back_pc) |
| |
| def in_loop(self) -> bool: |
| '''The processor is currently executing a loop.''' |
| |
| # A loop is executed if the loop stack is not empty. |
| return bool(self.loop_stack.stack) |
| |
| def changes(self) -> List[Trace]: |
| c = [] # type: List[Trace] |
| c += self.gprs.changes() |
| if self._pc_next_override is not None: |
| # Only append the next program counter to the trace if it is special |
| c.append(TracePC(self.get_next_pc())) |
| c += self.dmem.changes() |
| c += self.loop_stack.changes() |
| c += self.ext_regs.changes() |
| c += self.wsrs.changes() |
| c += self.csrs.flags.changes() |
| c += self.wdrs.changes() |
| return c |
| |
| def running(self) -> bool: |
| return self.fsm_state not in [FsmState.IDLE, FsmState.LOCKED] |
| |
| def commit(self, sim_stalled: bool) -> None: |
| # We shouldn't be running commit() in IDLE or LOCKED mode |
| assert self.running() |
| |
| # Check if we processed the RND data, if so set the register. This is |
| # done seperately from the rnd_completed method because in other case |
| # we are exiting stall caused by RND waiting by one cycle too early. |
| self.rnd_reg_set() |
| |
| # If model is waiting for the RND register to cross CDC, increment a |
| # counter to say how long we've waited. This lets us spot if the CDC |
| # gets stuck for some reason. |
| if self.rnd_cdc_pending: |
| self.rnd_cdc_counter += 1 |
| |
| # If model is waiting for the RND register to cross CDC, increment a |
| # counter to say how long we've waited. This lets us spot if the CDC |
| # gets stuck for some reason. |
| if self.urnd_cdc_pending: |
| self.urnd_cdc_counter += 1 |
| |
| # If we are in PRE_EXEC mode, we should commit external registers |
| # (which lets us reflect things like the update to the STATUS |
| # register). Then we wait until URND processing is done, at which |
| # point, we'll switch to EXEC mode. |
| if self.fsm_state == FsmState.PRE_EXEC: |
| self.ext_regs.commit() |
| if self.wsrs.URND.running: |
| # This part is strictly for standalone simulation. Otherwise |
| # we would set fsm_state before commit (at urnd_completed) |
| self.fsm_state = FsmState.FETCH_WAIT |
| |
| return |
| |
| # FETCH_WAIT works like PRE_EXEC, but it's only ever a single cycle |
| # wait. |
| if self.fsm_state == FsmState.FETCH_WAIT: |
| self.ext_regs.commit() |
| self.fsm_state = FsmState.EXEC |
| |
| return |
| |
| # If we are in POST_EXEC mode, this is the single cycle after the end |
| # of execution after either completion or a recoverable error. Commit |
| # external registers (to update STATUS) and then switch to IDLE. |
| if self.fsm_state == FsmState.POST_EXEC: |
| self.ext_regs.commit() |
| self.fsm_state = FsmState.IDLE |
| return |
| |
| # If we are in LOCKING mode, this is the single cycle after the end of |
| # execution after a fatal error. Commit external registers (to update |
| # STATUS) and then switch to LOCKED. |
| if self.fsm_state == FsmState.LOCKING: |
| self.ext_regs.commit() |
| self.fsm_state = FsmState.LOCKED |
| return |
| |
| # Otherwise, we're in EXEC mode. |
| assert self.fsm_state == FsmState.EXEC |
| |
| # In case of a pending halt, commit the external registers, which |
| # contain e.g. the ERR_BITS field, but nothing else. Switch to |
| # POST_EXEC to allow one more cycle. |
| if self.pending_halt: |
| self.ext_regs.commit() |
| self.fsm_state = (FsmState.LOCKING |
| if self._err_bits >> 16 else FsmState.POST_EXEC) |
| return |
| |
| # As pending_halt wasn't set, there shouldn't be any pending error bits |
| assert self._err_bits == 0 |
| |
| self.ext_regs.commit() |
| |
| # If we're stalled, there's nothing more to do: we only commit the rest |
| # of the architectural state when we finish our stall cycles. |
| if sim_stalled: |
| return |
| |
| self.dmem.commit() |
| self.gprs.commit() |
| self.pc = self.get_next_pc() |
| self._pc_next_override = None |
| self.loop_stack.commit() |
| self.ext_regs.commit() |
| self.wsrs.commit() |
| self.csrs.flags.commit() |
| self.wdrs.commit() |
| |
| def _abort(self) -> None: |
| '''Abort any pending state changes''' |
| self.gprs.abort() |
| self._pc_next_override = None |
| self.dmem.abort() |
| self.loop_stack.abort() |
| self.ext_regs.abort() |
| self.wsrs.abort() |
| self.csrs.flags.abort() |
| self.wdrs.abort() |
| |
| def start(self) -> None: |
| '''Start running; perform state init''' |
| self.ext_regs.write('STATUS', Status.BUSY_EXECUTE, True) |
| self.pending_halt = False |
| self._err_bits = 0 |
| |
| self.fsm_state = FsmState.PRE_EXEC |
| |
| self.pc = 0 |
| |
| # Reset CSRs, WSRs, loop stack and call stack |
| self.csrs = CSRFile() |
| |
| # Save the RND value when starting another run because when a SW |
| # run finishes we still keep RND data. |
| self.rnd_reg_set() |
| old_rnd = self.wsrs.RND._random_value |
| |
| self.wsrs = WSRFile() |
| if old_rnd is not None: |
| self.wsrs.RND.set_unsigned(old_rnd) |
| |
| self.loop_stack = LoopStack() |
| self.gprs.start() |
| |
| def stop(self) -> None: |
| '''Set flags to stop the processor and maybe abort the instruction. |
| |
| If the current instruction has caused an error (so self._err_bits is |
| nonzero), abort all its pending changes, including changes to external |
| registers. |
| |
| If not, we've just executed an ECALL. The only pending change will be |
| the increment of INSN_CNT that we want to keep. |
| |
| Either way, set the appropriate bits in the external ERR_CODE register, |
| clear the busy flag and write STOP_PC. |
| |
| ''' |
| |
| if self._err_bits: |
| # Abort all pending changes, including changes to external registers. |
| self._abort() |
| |
| # INTR_STATE is the interrupt state register. Bit 0 (which is being |
| # set) is the 'done' flag. |
| self.ext_regs.set_bits('INTR_STATE', 1 << 0) |
| |
| # STATUS is a status register. If there are any pending error bits |
| # greater than 16, this was a fatal error so we should lock ourselves. |
| # Otherwise, go back to IDLE. |
| new_status = Status.LOCKED if self._err_bits >> 16 else Status.IDLE |
| self.ext_regs.write('STATUS', new_status, True) |
| |
| # Make any error bits visible |
| self.ext_regs.write('ERR_BITS', self._err_bits, True) |
| |
| # Make the final PC visible. This isn't currently in the RTL, but is |
| # useful in simulations that want to track whether we stopped where we |
| # expected to stop. |
| self.ext_regs.write('STOP_PC', self.pc, True) |
| |
| def set_flags(self, fg: int, flags: FlagReg) -> None: |
| '''Update flags for a flag group''' |
| self.csrs.flags[fg] = flags |
| |
| def set_mlz_flags(self, fg: int, result: int) -> None: |
| '''Update M, L, Z flags for a flag group using the given result''' |
| self.csrs.flags[fg] = \ |
| FlagReg.mlz_for_result(self.csrs.flags[fg].C, result) |
| |
| def pre_insn(self, insn_affects_control: bool) -> None: |
| '''Run before running an instruction''' |
| self.loop_stack.check_insn(self.pc, insn_affects_control) |
| |
| def is_pc_valid(self, pc: int) -> bool: |
| '''Return whether pc is a valid program counter.''' |
| # The PC should always be non-negative since it's represented as an |
| # unsigned value. (It's an error in the simulator if that's come |
| # unstuck) |
| assert 0 <= pc |
| |
| # Check the new PC is word-aligned |
| if pc & 3: |
| return False |
| |
| # Check the new PC lies in instruction memory |
| if pc >= self.imem_size: |
| return False |
| |
| return True |
| |
| def post_insn(self, loop_warps: Dict[int, int]) -> None: |
| '''Update state after running an instruction but before commit''' |
| self.ext_regs.increment_insn_cnt() |
| self.loop_step(loop_warps) |
| self.gprs.post_insn() |
| |
| self._err_bits |= self.gprs.err_bits() | self.loop_stack.err_bits() |
| if self._err_bits: |
| self.pending_halt = True |
| |
| # Check that the next PC is valid, but only if we're not stopping |
| # anyway. This handles the case where we have a straight-line |
| # instruction at the top of memory. Jumps and branches to invalid |
| # addresses are handled in the instruction definition. |
| # |
| # This check is squashed if we're already halting, which avoids a |
| # problem when you have an ECALL instruction at the top of memory (the |
| # next address is bogus, but we don't care because we're stopping |
| # anyway). |
| if not self.is_pc_valid(self.get_next_pc()) and not self.pending_halt: |
| self._err_bits |= ErrBits.BAD_INSN_ADDR |
| self.pending_halt = True |
| |
| def read_csr(self, idx: int) -> int: |
| '''Read the CSR with index idx as an unsigned 32-bit number''' |
| return self.csrs.read_unsigned(self.wsrs, idx) |
| |
| def write_csr(self, idx: int, value: int) -> None: |
| '''Write value (an unsigned 32-bit number) to the CSR with index idx''' |
| self.csrs.write_unsigned(self.wsrs, idx, value) |
| |
| def peek_call_stack(self) -> List[int]: |
| '''Return the current call stack, bottom-first''' |
| return self.gprs.peek_call_stack() |
| |
| def stop_at_end_of_cycle(self, err_bits: int) -> None: |
| '''Tell the simulation to stop at the end of the cycle |
| |
| Any bits set in err_bits will be set in the ERR_BITS register when |
| we're done. |
| |
| ''' |
| self._err_bits |= err_bits |
| self.pending_halt = True |
