from copy import deepcopy
from functools import wraps
import os
+import errno
+import struct
+from openpower.syscalls import ppc_flags
import sys
+from elftools.elf.elffile import ELFFile # for isinstance
from nmigen.sim import Settle
import openpower.syscalls
SVP64CROffs, SVP64MODEb)
from openpower.decoder.helpers import (ISACallerHelper, ISAFPHelpers, exts,
gtu, undefined, copy_assign_rhs)
-from openpower.decoder.isa.mem import Mem, MemMMap, MemException
+from openpower.decoder.isa.mem import Mem, MemMMap, MemException, LoadedELF
from openpower.decoder.isa.radixmmu import RADIX
from openpower.decoder.isa.svshape import SVSHAPE
from openpower.decoder.isa.svstate import SVP64State
*read_regs, *uninit_regs, *write_regs, *special_regs], extra=extra)
+def is_ffirst_mode(dec2):
+ rm_mode = yield dec2.rm_dec.mode
+ return rm_mode == SVP64RMMode.FFIRST.value
+
+
class GPR(dict):
def __init__(self, decoder, isacaller, svstate, regfile):
dict.__init__(self)
rnum = rnum.value
dict.__setitem__(self, rnum, value)
- def getz(self, rnum):
+ def getz(self, rnum, rvalue=None):
# rnum = rnum.value # only SelectableInt allowed
- log("GPR getzero?", rnum)
+ log("GPR getzero?", rnum, rvalue)
+ if rvalue is not None:
+ if rnum == 0:
+ return SelectableInt(0, rvalue.bits)
+ return rvalue
if rnum == 0:
return SelectableInt(0, 64)
return self[rnum]
for j in range(8):
s.append("%08x" % res[i+j])
s = ' '.join(s)
- print("reg", "%2d" % i, s)
+ log("reg", "%2d" % i, s, kind=LogType.InstrInOuts)
return res
class SPR(dict):
- def __init__(self, dec2, initial_sprs={}):
+ def __init__(self, dec2, initial_sprs={}, gpr=None):
self.sd = dec2
+ self.gpr = gpr # for SVSHAPE[0-3]
dict.__init__(self)
for key, v in initial_sprs.items():
if isinstance(key, SelectableInt):
self.__setitem__('SRR1', value)
if key == 1:
value = XERState(value)
+ if key in ('SVSHAPE0', 'SVSHAPE1', 'SVSHAPE2', 'SVSHAPE3'):
+ value = SVSHAPE(value, self.gpr)
log("setting spr", key, value)
dict.__setitem__(self, key, value)
log(" new dststep", dststep)
+class ExitSyscallCalled(Exception):
+ pass
+
+
class SyscallEmulator(openpower.syscalls.Dispatcher):
def __init__(self, isacaller):
self.__isacaller = isacaller
(identifier, *arguments) = map(int, (identifier, *arguments))
return super().__call__(identifier, *arguments)
+ def sys_exit_group(self, status, *rest):
+ self.__isacaller.halted = True
+ raise ExitSyscallCalled(status)
+
+ def sys_write(self, fd, buf, count, *rest):
+ buf = self.__isacaller.mem.get_ctypes(buf, count, is_write=False)
+ try:
+ return os.write(fd, buf)
+ except OSError as e:
+ return -e.errno
+
+ def sys_read(self, fd, buf, count, *rest):
+ buf = self.__isacaller.mem.get_ctypes(buf, count, is_write=True)
+ try:
+ return os.readv(fd, [buf])
+ except OSError as e:
+ return -e.errno
+
+ def sys_mmap(self, addr, length, prot, flags, fd, offset, *rest):
+ return self.__isacaller.mem.mmap_syscall(
+ addr, length, prot, flags, fd, offset, is_mmap2=False)
+
+ def sys_mmap2(self, addr, length, prot, flags, fd, offset, *rest):
+ return self.__isacaller.mem.mmap_syscall(
+ addr, length, prot, flags, fd, offset, is_mmap2=True)
+
+ def sys_brk(self, addr, *rest):
+ return self.__isacaller.mem.brk_syscall(addr)
+
+ def sys_munmap(self, addr, length, *rest):
+ return -errno.ENOSYS # TODO: implement
+
+ def sys_mprotect(self, addr, length, prot, *rest):
+ return -errno.ENOSYS # TODO: implement
+
+ def sys_pkey_mprotect(self, addr, length, prot, pkey, *rest):
+ return -errno.ENOSYS # TODO: implement
+
+ def sys_openat(self, dirfd, pathname, flags, mode, *rest):
+ try:
+ path = self.__isacaller.mem.read_cstr(pathname)
+ except (ValueError, MemException):
+ return -errno.EFAULT
+ try:
+ if dirfd == ppc_flags.AT_FDCWD:
+ return os.open(path, flags, mode)
+ else:
+ return os.open(path, flags, mode, dir_fd=dirfd)
+ except OSError as e:
+ return -e.errno
+
+ def _uname(self):
+ uname = os.uname()
+ sysname = b'Linux'
+ nodename = uname.nodename.encode()
+ release = b'5.6.0-1-powerpc64le'
+ version = b'#1 SMP Debian 5.6.7-1 (2020-04-29)'
+ machine = b'ppc64le'
+ domainname = b''
+ return sysname, nodename, release, version, machine, domainname
+
+ def sys_uname(self, buf, *rest):
+ s = struct.Struct("<65s65s65s65s65s")
+ try:
+ buf = self.__isacaller.mem.get_ctypes(buf, s.size, is_write=True)
+ except (ValueError, MemException):
+ return -errno.EFAULT
+ sysname, nodename, release, version, machine, domainname = \
+ self._uname()
+ s.pack_into(buf, 0, sysname, nodename, release, version, machine)
+ return 0
+
+ def sys_newuname(self, buf, *rest):
+ name_len = ppc_flags.__NEW_UTS_LEN + 1
+ s = struct.Struct("<%ds%ds%ds%ds%ds%ds" % ((name_len,) * 6))
+ try:
+ buf = self.__isacaller.mem.get_ctypes(buf, s.size, is_write=True)
+ except (ValueError, MemException):
+ return -errno.EFAULT
+ sysname, nodename, release, version, machine, domainname = \
+ self._uname()
+ s.pack_into(buf, 0,
+ sysname, nodename, release, version, machine, domainname)
+ return 0
+
+ def sys_readlink(self, pathname, buf, bufsiz, *rest):
+ dirfd = ppc_flags.AT_FDCWD
+ return self.sys_readlinkat(dirfd, pathname, buf, bufsiz)
+
+ def sys_readlinkat(self, dirfd, pathname, buf, bufsiz, *rest):
+ try:
+ path = self.__isacaller.mem.read_cstr(pathname)
+ buf = self.__isacaller.mem.get_ctypes(buf, bufsiz, is_write=True)
+ except (ValueError, MemException):
+ return -errno.EFAULT
+ try:
+ if dirfd == ppc_flags.AT_FDCWD:
+ result = os.readlink(path)
+ else:
+ result = os.readlink(path, dir_fd=dirfd)
+ retval = min(len(result), len(buf))
+ buf[:retval] = result[:retval]
+ return retval
+ except OSError as e:
+ return -e.errno
+
class ISACaller(ISACallerHelper, ISAFPHelpers, StepLoop):
# decoder2 - an instance of power_decoder2
initial_fpscr=0,
insnlog=None,
use_mmap_mem=False,
- use_syscall_emu=False):
+ use_syscall_emu=False,
+ emulating_mmap=False):
if use_syscall_emu:
self.syscall = SyscallEmulator(isacaller=self)
if not use_mmap_mem:
else:
self.syscall = None
+ # we will eventually be able to load ELF files without use_syscall_emu
+ # (e.g. the linux kernel), so do it in a separate if block
+ if isinstance(initial_insns, ELFFile):
+ if not use_mmap_mem:
+ log("forcing use_mmap_mem due to loading an ELF file")
+ use_mmap_mem = True
+ if not emulating_mmap:
+ log("forcing emulating_mmap due to loading an ELF file")
+ emulating_mmap = True
+
# trace log file for model output. if None do nothing
self.insnlog = insnlog
self.insnlog_is_file = hasattr(insnlog, "write")
initial_sprs = deepcopy(initial_sprs) # so as not to get modified
self.gpr = GPR(decoder2, self, self.svstate, regfile)
self.fpr = GPR(decoder2, self, self.svstate, fpregfile)
- self.spr = SPR(decoder2, initial_sprs) # initialise SPRs before MMU
+ # initialise SPRs before MMU
+ self.spr = SPR(decoder2, initial_sprs, gpr=self.gpr)
# set up 4 dummy SVSHAPEs if they aren't already set up
for i in range(4):
sname = 'SVSHAPE%d' % i
val = self.spr.get(sname, 0)
- # make sure it's an SVSHAPE
- self.spr[sname] = SVSHAPE(val, self.gpr)
+ # make sure it's an SVSHAPE -- conversion done by SPR.__setitem__
+ self.spr[sname] = val
self.last_op_svshape = False
# "raw" memory
if use_mmap_mem:
self.mem = MemMMap(row_bytes=8,
initial_mem=initial_mem,
- misaligned_ok=True)
+ misaligned_ok=True,
+ emulating_mmap=emulating_mmap)
self.imem = self.mem
- self.mem.initialize(row_bytes=4, initial_mem=initial_insns)
+ lelf = self.mem.initialize(row_bytes=4, initial_mem=initial_insns)
+ if isinstance(lelf, LoadedELF): # stuff parsed from ELF
+ initial_pc = lelf.pc
+ for k, v in lelf.gprs.items():
+ self.gpr[k] = SelectableInt(v, 64)
+ initial_fpscr = lelf.fpscr
self.mem.log_fancy(kind=LogType.InstrInOuts)
else:
self.mem = Mem(row_bytes=8, initial_mem=initial_mem,
TRAP function is callable from inside the pseudocode itself,
hence the default arguments. when calling from inside ISACaller
it is best to use call_trap()
+
+ trap_addr: int | SelectableInt
+ the address to go to (before any modifications from `KAIVB`)
+ trap_bit: int | None
+ the bit in `SRR1` to set, `None` means don't set any bits.
"""
if isinstance(trap_addr, SelectableInt):
trap_addr = trap_addr.value
if self.is_svp64_mode:
self.spr['SVSRR0'] = self.namespace['SVSTATE'].value
self.trap_nia = SelectableInt(trap_addr | (kaivb & ~0x1fff), 64)
- self.spr['SRR1'][trap_bit] = 1 # change *copy* of MSR in SRR1
+ if trap_bit is not None:
+ self.spr['SRR1'][trap_bit] = 1 # change *copy* of MSR in SRR1
# set exception bits. TODO: this should, based on the address
# in figure 66 p1065 V3.0B and the table figure 65 p1063 set these
self.cr_backup = self.cr.value
# sv.bc* need some extra fields
- if self.is_svp64_mode and insn_name.startswith("sv.bc"):
- # blegh grab bits manually
- mode = yield self.dec2.rm_dec.rm_in.mode
- # convert to SelectableInt before test
- mode = SelectableInt(mode, 5)
- bc_vlset = mode[SVP64MODEb.BC_VLSET] != 0
- bc_vli = mode[SVP64MODEb.BC_VLI] != 0
- bc_snz = mode[SVP64MODEb.BC_SNZ] != 0
- bc_vsb = yield self.dec2.rm_dec.bc_vsb
- bc_lru = yield self.dec2.rm_dec.bc_lru
- bc_gate = yield self.dec2.rm_dec.bc_gate
- sz = yield self.dec2.rm_dec.pred_sz
- self.namespace['mode'] = SelectableInt(mode, 5)
- self.namespace['ALL'] = SelectableInt(bc_gate, 1)
- self.namespace['VSb'] = SelectableInt(bc_vsb, 1)
- self.namespace['LRu'] = SelectableInt(bc_lru, 1)
- self.namespace['VLSET'] = SelectableInt(bc_vlset, 1)
- self.namespace['VLI'] = SelectableInt(bc_vli, 1)
- self.namespace['sz'] = SelectableInt(sz, 1)
- self.namespace['SNZ'] = SelectableInt(bc_snz, 1)
+ if not self.is_svp64_mode or not insn_name.startswith("sv.bc"):
+ return
+
+ # blegh grab bits manually
+ mode = yield self.dec2.rm_dec.rm_in.mode
+ # convert to SelectableInt before test
+ mode = SelectableInt(mode, 5)
+ bc_vlset = mode[SVP64MODEb.BC_VLSET] != 0
+ bc_vli = mode[SVP64MODEb.BC_VLI] != 0
+ bc_snz = mode[SVP64MODEb.BC_SNZ] != 0
+ bc_vsb = yield self.dec2.rm_dec.bc_vsb
+ bc_ctrtest = yield self.dec2.rm_dec.bc_ctrtest
+ bc_lru = yield self.dec2.rm_dec.bc_lru
+ bc_gate = yield self.dec2.rm_dec.bc_gate
+ sz = yield self.dec2.rm_dec.pred_sz
+ self.namespace['mode'] = SelectableInt(mode, 5)
+ self.namespace['ALL'] = SelectableInt(bc_gate, 1)
+ self.namespace['VSb'] = SelectableInt(bc_vsb, 1)
+ self.namespace['LRu'] = SelectableInt(bc_lru, 1)
+ self.namespace['CTRtest'] = SelectableInt(bc_ctrtest, 1)
+ self.namespace['VLSET'] = SelectableInt(bc_vlset, 1)
+ self.namespace['VLI'] = SelectableInt(bc_vli, 1)
+ self.namespace['sz'] = SelectableInt(sz, 1)
+ self.namespace['SNZ'] = SelectableInt(bc_snz, 1)
def get_kludged_op_add_ca_ov(self, inputs, inp_ca_ov):
""" this was not at all necessary to do. this function massively
return ca64, ca32, ov64, ov32
def handle_carry_(self, inputs, output, ca, ca32, inp_ca_ov):
+ if ca is not None and ca32 is not None:
+ return
op = yield self.dec2.e.do.insn_type
if op == MicrOp.OP_ADD.value and ca is None and ca32 is None:
retval = yield from self.get_kludged_op_add_ca_ov(
cr_field = selectconcat(negative, positive, zero, SO)
log("handle_comparison cr_field", self.cr, cr_idx, cr_field)
self.crl[cr_idx].eq(cr_field)
+ return cr_field
def set_pc(self, pc_val):
self.namespace['NIA'] = SelectableInt(pc_val, 64)
# 2. Call the HDL implementation which invokes trap.
# 3. Reroute the guest system call to host system call.
# 4. Force return from the interrupt as if we had guest OS.
- # "executing" rfid requires putting 0x4c000024 temporarily
- # into the program at the PC. TODO investigate and remove
if ((asmop in ("sc", "scv")) and
(self.syscall is not None) and
not syscall_emu_active):
self.gpr.write(3, result, False, self.namespace["XLEN"])
# Return from interrupt
- backup = self.imem.ld(pc, 4, False, True, instr_fetch=True)
- self.imem.st(pc, 0x4c000024, width=4, swap=True)
yield from self.call("rfid", syscall_emu_active=True)
- self.imem.st(pc, backup, width=4, swap=True)
+ return
elif ((name in ("rfid", "hrfid")) and syscall_emu_active):
asmop = "rfid"
"brh", "brw", "brd",
'setvl', 'svindex', 'svremap', 'svstep',
'svshape', 'svshape2',
- 'ternlogi', 'bmask', 'cprop',
+ 'ternlogi', 'bmask', 'cprop', 'gbbd',
'absdu', 'absds', 'absdacs', 'absdacu', 'avgadd',
'fmvis', 'fishmv', 'pcdec', "maddedu", "divmod2du",
"dsld", "dsrd", "maddedus",
illegal = False
ins_name = dotstrp
+ # match against instructions treated as nop, see nop below
+ if asmop.startswith("dcbt"):
+ illegal = False
+ ins_name = "nop"
+
# branch-conditional redirects to sv.bc
if asmop.startswith('bc') and self.is_svp64_mode:
ins_name = 'sv.%s' % ins_name
ew_src = 8 << (3-int(ew_src)) # convert to bitlength
ew_dst = 8 << (3-int(ew_dst)) # convert to bitlength
xlen = max(ew_src, ew_dst)
- log("elwdith", ew_src, ew_dst)
+ log("elwidth", ew_src, ew_dst)
log("XLEN:", self.is_svp64_mode, xlen)
# look up instruction in ISA.instrs, prepare namespace
remap_active = yield self.dec2.remap_active
else:
remap_active = False
- log("remap active", bin(remap_active))
+ log("remap active", bin(remap_active), self.is_svp64_mode)
+
+ # LDST does *not* allow elwidth overrides on RA (Effective Address).
+ # this has to be detected. XXX TODO: RB for ldst-idx *may* need
+ # conversion (to 64-bit) also.
+ # see write reg this *HAS* to also override XLEN to 64 on LDST/Update
+ sv_mode = yield self.dec2.rm_dec.sv_mode
+ is_ldst = (sv_mode in [SVMode.LDST_IDX.value, SVMode.LDST_IMM.value] \
+ and self.is_svp64_mode)
+ log("is_ldst", sv_mode, is_ldst)
# main input registers (RT, RA ...)
for name in input_names:
inputs[name] = self.crl[0]
elif name in spr_byname:
inputs[name] = self.spr[name]
+ elif is_ldst and name == 'RA':
+ regval = (yield from self.get_input(name, ew_src, 64))
+ log("EA (RA) regval name", name, regval)
+ inputs[name] = regval
else:
- regval = (yield from self.get_input(name, ew_src))
+ regval = (yield from self.get_input(name, ew_src, xlen))
log("regval name", name, regval)
inputs[name] = regval
# check if this was an sv.bc* and create an indicator that
# this is the last check to be made as a loop. combined with
- # the ALL/ANY mode we can early-exit
+ # the ALL/ANY mode we can early-exit. note that BI (to test)
+ # is an input so there is no termination if BI is scalar
+ # (because early-termination is for *output* scalars)
if self.is_svp64_mode and ins_name.startswith("sv.bc"):
- no_in_vec = yield self.dec2.no_in_vec # BI is scalar
- end_loop = no_in_vec or srcstep == vl-1 or dststep == vl-1
+ end_loop = srcstep == vl-1 or dststep == vl-1
self.namespace['end_loop'] = SelectableInt(end_loop, 1)
inp_ca_ov = (self.spr['XER'][XER_bits['CA']].value,
results = info.func(self, *inputs)
output_names = create_args(info.write_regs)
outs = {}
+ # record .ok before anything after the pseudo-code can modify it
+ outs_ok = {}
for out, n in zip(results or [], output_names):
outs[n] = out
+ outs_ok[n] = True
+ if isinstance(out, SelectableInt):
+ outs_ok[n] = out.ok
log("results", outs)
+ log("results ok", outs_ok)
# "inject" decorator takes namespace from function locals: we need to
# overwrite NIA being overwritten (sigh)
# XXX TODO: now that CR0 is supported, sort out svstep's pseudocode
# to write directly to CR0 instead of in ISACaller. hooyahh.
if rc_en and ins_name not in ['svstep']:
+ if outs_ok.get('FPSCR', False):
+ FPSCR = outs['FPSCR']
+ else:
+ FPSCR = self.FPSCR
yield from self.do_rc_ov(
- ins_name, results[0], overflow, cr0, cr1, output_names)
+ ins_name, results[0], overflow, cr0, cr1, FPSCR)
# check failfirst
ffirst_hit = False, False
sv_mode = yield self.dec2.rm_dec.sv_mode
is_cr = sv_mode == SVMode.CROP.value
chk = rc_en or is_cr
+ if outs_ok.get('CR', False):
+ # early write so check_ffirst can see value
+ self.namespace['CR'].eq(outs['CR'])
ffirst_hit = (yield from self.check_ffirst(info, chk, srcstep))
+ # any modified return results?
+ yield from self.do_outregs(
+ info, outs, carry_en, ffirst_hit, ew_dst, outs_ok)
+
# check if a FP Exception occurred. TODO for DD-FFirst, check VLi
# and raise the exception *after* if VLi=1 but if VLi=0 then
# truncate and make the exception "disappear".
self.call_trap(0x700, PIb.FP)
return
- # any modified return results?
- yield from self.do_outregs_nia(asmop, ins_name, info, outs,
- carry_en, rc_en, ffirst_hit, ew_dst)
+ yield from self.do_nia(asmop, ins_name, rc_en, ffirst_hit)
def check_ffirst(self, info, rc_en, srcstep):
"""fail-first mode: checks a bit of Rc Vector, truncates VL
log(" vli", vli_)
log(" cr_bit", cr_bit)
log(" rc_en", rc_en)
- if not rc_en or rm_mode != SVP64RMMode.FFIRST.value:
+ ffirst = yield from is_ffirst_mode(self.dec2)
+ if not rc_en or not ffirst:
return False, False
# get the CR vevtor, do BO-test
crf = "CR0"
yield Settle() # let decoder update
return True, vli_
- def do_rc_ov(self, ins_name, result, overflow, cr0, cr1, output_names):
+ def do_rc_ov(self, ins_name, result, overflow, cr0, cr1, FPSCR):
cr_out = yield self.dec2.op.cr_out
if cr_out == CROutSel.CR1.value:
rc_reg = "CR1"
if rc_reg == "CR1":
if cr1 is None:
- cr1 = int(self.FPSCR.FX) << 3
- cr1 |= int(self.FPSCR.FEX) << 2
- cr1 |= int(self.FPSCR.VX) << 1
- cr1 |= int(self.FPSCR.OX)
+ cr1 = int(FPSCR.FX) << 3
+ cr1 |= int(FPSCR.FEX) << 2
+ cr1 |= int(FPSCR.VX) << 1
+ cr1 |= int(FPSCR.OX)
log("default fp cr1", cr1)
else:
log("explicit cr1", cr1)
elif cr0 is None:
# if there was not an explicit CR0 in the pseudocode,
# do implicit Rc=1
- self.handle_comparison(result, regnum, overflow, no_so=is_setvl)
+ c = self.handle_comparison(result, regnum, overflow, no_so=is_setvl)
+ log("implicit cr0", c)
else:
# otherwise we just blat CR0 into the required regnum
- log("explicit rc0", cr0)
+ log("explicit cr0", cr0)
self.crl[regnum].eq(cr0)
- def do_outregs_nia(self, asmop, ins_name, info, outs,
- ca_en, rc_en, ffirst_hit, ew_dst):
+ def do_outregs(self, info, outs, ca_en, ffirst_hit, ew_dst, outs_ok):
ffirst_hit, vli = ffirst_hit
# write out any regs for this instruction, but only if fail-first is ok
# XXX TODO: allow CR-vector to be written out even if ffirst fails
if not ffirst_hit or vli:
for name, output in outs.items():
+ if not outs_ok[name]:
+ log("skipping writing output with .ok=False", name, output)
+ continue
yield from self.check_write(info, name, output, ca_en, ew_dst)
# restore the CR value on non-VLI failfirst (from sv.cmp and others
# which write directly to CR in the pseudocode (gah, what a mess)
# if ffirst_hit and not vli:
# self.cr.value = self.cr_backup
+ def do_nia(self, asmop, ins_name, rc_en, ffirst_hit):
+ ffirst_hit, vli = ffirst_hit
if ffirst_hit:
self.svp64_reset_loop()
nia_update = True
else:
# check advancement of src/dst/sub-steps and if PC needs updating
- nia_update = (yield from self.check_step_increment(rc_en,
- asmop, ins_name))
+ nia_update = (yield from self.check_step_increment(
+ rc_en, asmop, ins_name))
if nia_update:
self.update_pc_next()
else:
self.namespace['D'] = imm
- def get_input(self, name, ew_src):
+ def get_input(self, name, ew_src, xlen):
# using PowerDecoder2, first, find the decoder index.
# (mapping name RA RB RC RS to in1, in2, in3)
regnum, is_vec = yield from get_idx_in(self.dec2, name, True)
regname = "_" + name
if not self.is_svp64_mode or ew_src == 64:
self.namespace[regname] = regnum
- elif regname in self.namespace:
- del self.namespace[regname]
+ else:
+ # FIXME: we're trying to access a sub-register, plain register
+ # numbers don't work for that. for now, just pass something that
+ # can be compared to 0 and probably will cause an error if misused.
+ # see https://bugs.libre-soc.org/show_bug.cgi?id=1221
+ self.namespace[regname] = regnum * 10000
if not self.is_svp64_mode or not self.pred_src_zero:
log('reading reg %s %s' % (name, str(regnum)), is_vec)
if name in fregs:
- reg_val = SelectableInt(self.fpr(base, is_vec, offs, ew_src))
- log("read reg %d/%d: 0x%x" % (base, offs, reg_val.value),
- kind=LogType.InstrInOuts)
+ fval = self.fpr(base, is_vec, offs, ew_src)
+ reg_val = SelectableInt(fval)
+ assert ew_src == self.XLEN, "TODO fix elwidth conversion"
self.trace("r:FPR:%d:%d:%d " % (base, offs, ew_src))
+ log("read fp reg %d/%d: 0x%x" % (base, offs, reg_val.value),
+ kind=LogType.InstrInOuts)
elif name is not None:
- reg_val = SelectableInt(self.gpr(base, is_vec, offs, ew_src))
+ gval = self.gpr(base, is_vec, offs, ew_src)
+ reg_val = SelectableInt(gval.value, bits=xlen)
self.trace("r:GPR:%d:%d:%d " % (base, offs, ew_src))
- log("read reg %d/%d: 0x%x" % (base, offs, reg_val.value),
+ log("read int reg %d/%d: 0x%x" % (base, offs, reg_val.value),
kind=LogType.InstrInOuts)
else:
log('zero input reg %s %s' % (name, str(regnum)), is_vec)
if name in fregs:
self.fpr.write(regnum, output, is_vec, ew_dst)
self.trace("w:FPR:%d:%d:%d " % (rnum, offset, ew_dst))
- else:
- self.gpr.write(regnum, output, is_vec, ew_dst)
- self.trace("w:GPR:%d:%d:%d " % (rnum, offset, ew_dst))
+ return
+
+ # LDST/Update does *not* allow elwidths on RA (Effective Address).
+ # this has to be detected, and overridden. see get_input (related)
+ sv_mode = yield self.dec2.rm_dec.sv_mode
+ is_ldst = (sv_mode in [SVMode.LDST_IDX.value, SVMode.LDST_IMM.value] \
+ and self.is_svp64_mode)
+ if is_ldst and name in ['EA', 'RA']:
+ op = self.dec2.dec.op
+ if hasattr(op, "upd"):
+ # update mode LD/ST uses read-reg A also as an output
+ upd = yield op.upd
+ log("write is_ldst is_update", sv_mode, is_ldst, upd)
+ if upd == LDSTMode.update.value:
+ ew_dst = 64 # override for RA (EA) to 64-bit
+
+ self.gpr.write(regnum, output, is_vec, ew_dst)
+ self.trace("w:GPR:%d:%d:%d " % (rnum, offset, ew_dst))
def check_step_increment(self, rc_en, asmop, ins_name):
# check if it is the SVSTATE.src/dest step that needs incrementing
vfirst = self.svstate.vfirst
log(" SV Vertical First", vf, vfirst)
if not vf and vfirst == 1:
+ # SV Branch-Conditional required to be as-if-vector
+ # because there *is* no destination register
+ # (SV normally only terminates on 1st scalar reg written
+ # except in [slightly-misnamed] mapreduce mode)
+ ffirst = yield from is_ffirst_mode(self.dec2)
+ if insn_name.startswith("sv.bc") or ffirst:
+ self.update_pc_next()
+ return False
self.update_nia()
return True
sv_ptype = yield self.dec2.dec.op.SV_Ptype
out_vec = not (yield self.dec2.no_out_vec)
in_vec = not (yield self.dec2.no_in_vec)
+ rm_mode = yield self.dec2.rm_dec.mode
log(" svstate.vl", vl)
log(" svstate.mvl", mvl)
log(" rm.subvl", subvl)
log(" out_vec", out_vec)
log(" in_vec", in_vec)
log(" sv_ptype", sv_ptype, sv_ptype == SVPType.P2.value)
+ log(" rm_mode", rm_mode)
# check if this was an sv.bc* and if so did it succeed
if self.is_svp64_mode and insn_name.startswith("sv.bc"):
end_loop = self.namespace['end_loop']
svp64_is_vector = (out_vec or in_vec)
else:
svp64_is_vector = out_vec
+ # also if data-dependent fail-first is used, only in_vec is tested,
+ # allowing *scalar destinations* to be used as an accumulator.
+ # effectively this implies /mr (mapreduce mode) is 100% on with ddffirst
+ # see https://bugs.libre-soc.org/show_bug.cgi?id=1183#c16
+ ffirst = yield from is_ffirst_mode(self.dec2)
+ if ffirst:
+ svp64_is_vector = in_vec
+
# loops end at the first "hit" (source or dest)
yield from self.advance_svstate_steps()
loopend = self.loopend
# not an SVP64 branch, so fix PC (NIA==CIA) for next loop
# (by default, NIA is CIA+4 if v3.0B or CIA+8 if SVP64)
# this way we keep repeating the same instruction (with new steps)
- self.pc.NIA.value = self.pc.CIA.value
+ self.pc.NIA.eq(self.pc.CIA)
self.namespace['NIA'] = self.pc.NIA
log("end of sub-pc call", self.namespace['CIA'], self.namespace['NIA'])
return False # DO NOT allow PC update whilst Sub-PC loop running