* https://bugs.libre-soc.org/show_bug.cgi?id=424
"""
-import re
-from nmigen.back.pysim import Settle
+from collections import namedtuple
+from copy import deepcopy
from functools import wraps
-from copy import copy, deepcopy
-from openpower.decoder.orderedset import OrderedSet
-from openpower.decoder.selectable_int import (
- SelectableIntMapping,
- FieldSelectableInt,
- SelectableInt,
- selectconcat,
-)
-from openpower.decoder.power_enums import (spr_dict, spr_byname, XER_bits,
- insns, MicrOp,
- In1Sel, In2Sel, In3Sel,
- OutSel, CRInSel, CROutSel, LDSTMode,
- SVP64RMMode, SVP64PredMode,
- SVP64PredInt, SVP64PredCR,
- SVP64LDSTmode)
-
-from openpower.decoder.power_enums import SVPtype
-
-from openpower.decoder.helpers import (exts, gtu, ltu, undefined,
- ISACallerHelper, ISAFPHelpers)
-from openpower.consts import PIb, MSRb # big-endian (PowerISA versions)
-from openpower.consts import (SVP64MODE,
- SVP64CROffs,
- )
-from openpower.decoder.power_svp64 import SVP64RM, decode_extra
-
+import os
+import sys
+from elftools.elf.elffile import ELFFile # for isinstance
+
+from nmigen.sim import Settle
+import openpower.syscalls
+from openpower.consts import (MSRb, PIb, # big-endian (PowerISA versions)
+ SVP64CROffs, SVP64MODEb)
+from openpower.decoder.helpers import (ISACallerHelper, ISAFPHelpers, exts,
+ gtu, undefined, copy_assign_rhs)
+from openpower.decoder.isa.mem import Mem, MemMMap, MemException, LoadedELF
from openpower.decoder.isa.radixmmu import RADIX
-from openpower.decoder.isa.mem import Mem, swap_order, MemException
from openpower.decoder.isa.svshape import SVSHAPE
from openpower.decoder.isa.svstate import SVP64State
+from openpower.decoder.orderedset import OrderedSet
+from openpower.decoder.power_enums import (FPTRANS_INSNS, CRInSel, CROutSel,
+ In1Sel, In2Sel, In3Sel, LDSTMode,
+ MicrOp, OutSel, SVMode,
+ SVP64LDSTmode, SVP64PredCR,
+ SVP64PredInt, SVP64PredMode,
+ SVP64RMMode, SVPType, XER_bits,
+ insns, spr_byname, spr_dict,
+ BFP_FLAG_NAMES)
+from openpower.insndb.core import SVP64Instruction
+from openpower.decoder.power_svp64 import SVP64RM, decode_extra
+from openpower.decoder.selectable_int import (FieldSelectableInt,
+ SelectableInt, selectconcat,
+ EFFECTIVELY_UNLIMITED)
+from openpower.consts import DEFAULT_MSR
+from openpower.fpscr import FPSCRState
+from openpower.xer import XERState
+from openpower.util import LogType, log
+
+LDST_UPDATE_INSNS = ['ldu', 'lwzu', 'lbzu', 'lhzu', 'lhau', 'lfsu', 'lfdu',
+ 'stwu', 'stbu', 'sthu', 'stfsu', 'stfdu', 'stdu',
+ ]
-from openpower.util import log
-
-from collections import namedtuple
-import math
-import sys
-
instruction_info = namedtuple('instruction_info',
'func read_regs uninit_regs write_regs ' +
'special_regs op_fields form asmregs')
'VRSAVE': 256}
+# rrright. this is here basically because the compiler pywriter returns
+# results in a specific priority order. to make sure regs match up they
+# need partial sorting. sigh.
REG_SORT_ORDER = {
# TODO (lkcl): adjust other registers that should be in a particular order
# probably CA, CA32, and CR
"CA": 0,
"CA32": 0,
+ "FPSCR": 1,
+
"overflow": 7, # should definitely be last
+ "CR0": 8, # likewise
}
fregs = ['FRA', 'FRB', 'FRC', 'FRS', 'FRT']
+def get_masked_reg(regs, base, offs, ew_bits):
+ # rrrright. start by breaking down into row/col, based on elwidth
+ gpr_offs = offs // (64 // ew_bits)
+ gpr_col = offs % (64 // ew_bits)
+ # compute the mask based on ew_bits
+ mask = (1 << ew_bits) - 1
+ # now select the 64-bit register, but get its value (easier)
+ val = regs[base + gpr_offs]
+ # shift down so element we want is at LSB
+ val >>= gpr_col * ew_bits
+ # mask so we only return the LSB element
+ return val & mask
+
+
+def set_masked_reg(regs, base, offs, ew_bits, value):
+ # rrrright. start by breaking down into row/col, based on elwidth
+ gpr_offs = offs // (64//ew_bits)
+ gpr_col = offs % (64//ew_bits)
+ # compute the mask based on ew_bits
+ mask = (1 << ew_bits)-1
+ # now select the 64-bit register, but get its value (easier)
+ val = regs[base+gpr_offs]
+ # now mask out the bit we don't want
+ val = val & ~(mask << (gpr_col*ew_bits))
+ # then wipe the bit we don't want from the value
+ value = value & mask
+ # OR the new value in, shifted up
+ val |= value << (gpr_col*ew_bits)
+ regs[base+gpr_offs] = val
+
+
def create_args(reglist, extra=None):
retval = list(OrderedSet(reglist))
retval.sort(key=lambda reg: REG_SORT_ORDER.get(reg, 0))
return retval
+def create_full_args(*, read_regs, special_regs, uninit_regs, write_regs,
+ extra=None):
+ return create_args([
+ *read_regs, *uninit_regs, *write_regs, *special_regs], extra=extra)
+
+
class GPR(dict):
def __init__(self, decoder, isacaller, svstate, regfile):
dict.__init__(self)
for i in range(len(regfile)):
self[i] = SelectableInt(regfile[i], 64)
- def __call__(self, ridx):
+ def __call__(self, ridx, is_vec=False, offs=0, elwidth=64):
if isinstance(ridx, SelectableInt):
ridx = ridx.value
- return self[ridx]
+ if elwidth == 64:
+ return self[ridx+offs]
+ # rrrright. start by breaking down into row/col, based on elwidth
+ gpr_offs = offs // (64//elwidth)
+ gpr_col = offs % (64//elwidth)
+ # now select the 64-bit register, but get its value (easier)
+ val = self[ridx+gpr_offs].value
+ # now shift down and mask out
+ val = val >> (gpr_col*elwidth) & ((1 << elwidth)-1)
+ # finally, return a SelectableInt at the required elwidth
+ log("GPR call", ridx, "isvec", is_vec, "offs", offs,
+ "elwid", elwidth, "offs/col", gpr_offs, gpr_col, "val", hex(val))
+ return SelectableInt(val, elwidth)
def set_form(self, form):
self.form = form
+ def write(self, rnum, value, is_vec=False, elwidth=64):
+ # get internal value
+ if isinstance(rnum, SelectableInt):
+ rnum = rnum.value
+ if isinstance(value, SelectableInt):
+ value = value.value
+ # compatibility...
+ if isinstance(rnum, tuple):
+ rnum, base, offs = rnum
+ else:
+ base, offs = rnum, 0
+ # rrrright. start by breaking down into row/col, based on elwidth
+ gpr_offs = offs // (64//elwidth)
+ gpr_col = offs % (64//elwidth)
+ # compute the mask based on elwidth
+ mask = (1 << elwidth)-1
+ # now select the 64-bit register, but get its value (easier)
+ val = self[base+gpr_offs].value
+ # now mask out the bit we don't want
+ val = val & ~(mask << (gpr_col*elwidth))
+ # then wipe the bit we don't want from the value
+ value = value & mask
+ # OR the new value in, shifted up
+ val |= value << (gpr_col*elwidth)
+ # finally put the damn value into the regfile
+ log("GPR write", base, "isvec", is_vec, "offs", offs,
+ "elwid", elwidth, "offs/col", gpr_offs, gpr_col, "val", hex(val),
+ "@", base+gpr_offs)
+ dict.__setitem__(self, base+gpr_offs, SelectableInt(val, 64))
+
def __setitem__(self, rnum, value):
# rnum = rnum.value # only SelectableInt allowed
log("GPR setitem", rnum, value)
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[key] = v
def __getitem__(self, key):
- log("get spr", key)
- log("dict", self.items())
+ #log("get spr", key)
+ #log("dict", self.items())
# if key in special_sprs get the special spr, otherwise return key
if isinstance(key, SelectableInt):
key = key.value
info = spr_dict[key]
else:
info = spr_byname[key]
- dict.__setitem__(self, key, SelectableInt(0, info.length))
+ self[key] = SelectableInt(0, info.length)
res = dict.__getitem__(self, key)
- log("spr returning", key, res)
+ #log("spr returning", key, res)
return res
def __setitem__(self, key, value):
self.__setitem__('SRR0', value)
if key == 'HSRR1': # HACK!
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)
namespace['NIA'] = self.NIA
-# SVP64 ReMap field
-class SVP64RMFields(SelectableIntMapping, bits=24, fields={
- "spr": range(24),
- # SVP64 RM fields: see https://libre-soc.org/openpower/sv/svp64/
- "mmode": (0,),
- "mask": range(1, 4),
- "elwidth": range(4, 6),
- "ewsrc": range(6, 8),
- "subvl": range(8, 10),
- "extra": range(10, 19),
- "mode": range(19, 24),
- # these cover the same extra field, split into parts as EXTRA2
- "extra2": dict(enumerate([
- range(10, 12),
- range(12, 14),
- range(14, 16),
- range(16, 18),
- ])),
- "smask": range(16, 19),
- # and here as well, but EXTRA3
- "extra3": dict(enumerate([
- range(10, 13),
- range(13, 16),
- range(16, 19),
- ])),
-}):
-
- def __init__(self, value=0):
- super().__init__(value=value)
- self.spr = self
-
-
-SVP64RM_MMODE_SIZE = len(SVP64RMFields.mmode)
-SVP64RM_MASK_SIZE = len(SVP64RMFields.mask)
-SVP64RM_ELWIDTH_SIZE = len(SVP64RMFields.elwidth)
-SVP64RM_EWSRC_SIZE = len(SVP64RMFields.ewsrc)
-SVP64RM_SUBVL_SIZE = len(SVP64RMFields.subvl)
-SVP64RM_EXTRA2_SPEC_SIZE = len(SVP64RMFields.extra2[0])
-SVP64RM_EXTRA3_SPEC_SIZE = len(SVP64RMFields.extra3[0])
-SVP64RM_SMASK_SIZE = len(SVP64RMFields.smask)
-SVP64RM_MODE_SIZE = len(SVP64RMFields.mode)
-
-
-# SVP64 Prefix fields: see https://libre-soc.org/openpower/sv/svp64/
-class SVP64PrefixFields(SelectableIntMapping, bits=32, fields={
- "insn": range(32),
- # 6 bit major opcode EXT001, 2 bits "identifying" (7, 9), 24 SV ReMap
- "major": range(0, 6),
- "pid": (7, 9),
- # SVP64 24-bit RM (ReMap)
- "rm": ((6, 8) + tuple(range(10, 32))),
-}):
-
- def __init__(self, value=0):
- super().__init__(value=value)
- self.insn = self
-
-
-SV64P_MAJOR_SIZE = len(SVP64PrefixFields.major)
-SV64P_PID_SIZE = len(SVP64PrefixFields.pid)
-SV64P_RM_SIZE = len(SVP64PrefixFields.rm)
-
-
# CR register fields
# See PowerISA Version 3.0 B Book 1
# Section 2.3.1 Condition Register pages 30 - 31
_cr = FieldSelectableInt(self.cr, bits)
self.crl.append(_cr)
-# decode SVP64 predicate integer to reg number and invert
-
+# decode SVP64 predicate integer to reg number and invert
def get_predint(gpr, mask):
+ r3 = gpr(3)
r10 = gpr(10)
r30 = gpr(30)
log("get_predint", mask, SVP64PredInt.ALWAYS.value)
if mask == SVP64PredInt.ALWAYS.value:
return 0xffff_ffff_ffff_ffff # 64 bits of 1
if mask == SVP64PredInt.R3_UNARY.value:
- return 1 << (gpr(3).value & 0b111111)
+ return 1 << (r3.value & 0b111111)
if mask == SVP64PredInt.R3.value:
- return gpr(3).value
+ return r3.value
if mask == SVP64PredInt.R3_N.value:
- return ~gpr(3).value
+ return ~r3.value
if mask == SVP64PredInt.R10.value:
- return gpr(10).value
+ return r10.value
if mask == SVP64PredInt.R10_N.value:
- return ~gpr(10).value
+ return ~r10.value
if mask == SVP64PredInt.R30.value:
- return gpr(30).value
+ return r30.value
if mask == SVP64PredInt.R30_N.value:
- return ~gpr(30).value
-
-# decode SVP64 predicate CR to reg number and invert status
+ return ~r30.value
+# decode SVP64 predicate CR to reg number and invert status
def _get_predcr(mask):
if mask == SVP64PredCR.LT.value:
return 0, 1
if mask == SVP64PredCR.NS.value:
return 3, 0
+
# read individual CR fields (0..VL-1), extract the required bit
# and construct the mask
-
-
def get_predcr(crl, mask, vl):
idx, noninv = _get_predcr(mask)
mask = 0
# TODO, really should just be using PowerDecoder2
-def get_pdecode_idx_in(dec2, name):
+def get_idx_map(dec2, name):
op = dec2.dec.op
in1_sel = yield op.in1_sel
in2_sel = yield op.in2_sel
in3_sel = yield op.in3_sel
- # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
in1 = yield dec2.e.read_reg1.data
- in2 = yield dec2.e.read_reg2.data
- in3 = yield dec2.e.read_reg3.data
- in1_isvec = yield dec2.in1_isvec
- in2_isvec = yield dec2.in2_isvec
- in3_isvec = yield dec2.in3_isvec
- log("get_pdecode_idx_in in1", name, in1_sel, In1Sel.RA.value,
- in1, in1_isvec)
- log("get_pdecode_idx_in in2", name, in2_sel, In2Sel.RB.value,
- in2, in2_isvec)
- log("get_pdecode_idx_in in3", name, in3_sel, In3Sel.RS.value,
- in3, in3_isvec)
- log("get_pdecode_idx_in FRS in3", name, in3_sel, In3Sel.FRS.value,
- in3, in3_isvec)
- log("get_pdecode_idx_in FRB in2", name, in2_sel, In2Sel.FRB.value,
- in2, in2_isvec)
- log("get_pdecode_idx_in FRC in3", name, in3_sel, In3Sel.FRC.value,
- in3, in3_isvec)
# identify which regnames map to in1/2/3
- if name == 'RA':
+ if name == 'RA' or name == 'RA_OR_ZERO':
if (in1_sel == In1Sel.RA.value or
(in1_sel == In1Sel.RA_OR_ZERO.value and in1 != 0)):
- return in1, in1_isvec
+ return 1
if in1_sel == In1Sel.RA_OR_ZERO.value:
- return in1, in1_isvec
+ return 1
elif name == 'RB':
if in2_sel == In2Sel.RB.value:
- return in2, in2_isvec
+ return 2
if in3_sel == In3Sel.RB.value:
- return in3, in3_isvec
+ return 3
# XXX TODO, RC doesn't exist yet!
elif name == 'RC':
- assert False, "RC does not exist yet"
- elif name == 'RS':
+ if in3_sel == In3Sel.RC.value:
+ return 3
+ elif name in ['EA', 'RS']:
if in1_sel == In1Sel.RS.value:
- return in1, in1_isvec
+ return 1
if in2_sel == In2Sel.RS.value:
- return in2, in2_isvec
+ return 2
if in3_sel == In3Sel.RS.value:
- return in3, in3_isvec
+ return 3
elif name == 'FRA':
if in1_sel == In1Sel.FRA.value:
- return in1, in1_isvec
+ return 1
+ if in3_sel == In3Sel.FRA.value:
+ return 3
elif name == 'FRB':
if in2_sel == In2Sel.FRB.value:
- return in2, in2_isvec
+ return 2
elif name == 'FRC':
if in3_sel == In3Sel.FRC.value:
- return in3, in3_isvec
+ return 3
elif name == 'FRS':
if in1_sel == In1Sel.FRS.value:
- return in1, in1_isvec
+ return 1
if in3_sel == In3Sel.FRS.value:
- return in3, in3_isvec
+ return 3
+ elif name == 'FRT':
+ if in1_sel == In1Sel.FRT.value:
+ return 1
+ elif name == 'RT':
+ if in1_sel == In1Sel.RT.value:
+ return 1
+ return None
+
+
+# TODO, really should just be using PowerDecoder2
+def get_idx_in(dec2, name, ewmode=False):
+ idx = yield from get_idx_map(dec2, name)
+ if idx is None:
+ return None, False
+ op = dec2.dec.op
+ in1_sel = yield op.in1_sel
+ in2_sel = yield op.in2_sel
+ in3_sel = yield op.in3_sel
+ # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
+ in1 = yield dec2.e.read_reg1.data
+ in2 = yield dec2.e.read_reg2.data
+ in3 = yield dec2.e.read_reg3.data
+ if ewmode:
+ in1_base = yield dec2.e.read_reg1.base
+ in2_base = yield dec2.e.read_reg2.base
+ in3_base = yield dec2.e.read_reg3.base
+ in1_offs = yield dec2.e.read_reg1.offs
+ in2_offs = yield dec2.e.read_reg2.offs
+ in3_offs = yield dec2.e.read_reg3.offs
+ in1 = (in1, in1_base, in1_offs)
+ in2 = (in2, in2_base, in2_offs)
+ in3 = (in3, in3_base, in3_offs)
+
+ in1_isvec = yield dec2.in1_isvec
+ in2_isvec = yield dec2.in2_isvec
+ in3_isvec = yield dec2.in3_isvec
+ log("get_idx_in in1", name, in1_sel, In1Sel.RA.value,
+ in1, in1_isvec)
+ log("get_idx_in in2", name, in2_sel, In2Sel.RB.value,
+ in2, in2_isvec)
+ log("get_idx_in in3", name, in3_sel, In3Sel.RS.value,
+ in3, in3_isvec)
+ log("get_idx_in FRS in3", name, in3_sel, In3Sel.FRS.value,
+ in3, in3_isvec)
+ log("get_idx_in FRB in2", name, in2_sel, In2Sel.FRB.value,
+ in2, in2_isvec)
+ log("get_idx_in FRC in3", name, in3_sel, In3Sel.FRC.value,
+ in3, in3_isvec)
+ if idx == 1:
+ return in1, in1_isvec
+ if idx == 2:
+ return in2, in2_isvec
+ if idx == 3:
+ return in3, in3_isvec
return None, False
# TODO, really should just be using PowerDecoder2
-def get_pdecode_cr_in(dec2, name):
+def get_cr_in(dec2, name):
op = dec2.dec.op
in_sel = yield op.cr_in
in_bitfield = yield dec2.dec_cr_in.cr_bitfield.data
# get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
in1 = yield dec2.e.read_cr1.data
cr_isvec = yield dec2.cr_in_isvec
- log("get_pdecode_cr_in", in_sel, CROutSel.CR0.value, in1, cr_isvec)
+ log("get_cr_in", in_sel, CROutSel.CR0.value, in1, cr_isvec)
log(" sv_cr_in", sv_cr_in)
log(" cr_bf", in_bitfield)
log(" spec", spec)
if name == 'BI':
if in_sel == CRInSel.BI.value:
return in1, cr_isvec
- log("get_pdecode_cr_in not found", name)
+ log("get_cr_in not found", name)
return None, False
# TODO, really should just be using PowerDecoder2
-def get_pdecode_cr_out(dec2, name):
+def get_cr_out(dec2, name):
op = dec2.dec.op
out_sel = yield op.cr_out
out_bitfield = yield dec2.dec_cr_out.cr_bitfield.data
sv_override = yield dec2.dec_cr_out.sv_override
# get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
out = yield dec2.e.write_cr.data
- o_isvec = yield dec2.o_isvec
- log("get_pdecode_cr_out", out_sel, CROutSel.CR0.value, out, o_isvec)
+ o_isvec = yield dec2.cr_out_isvec
+ log("get_cr_out", out_sel, CROutSel.CR0.value, out, o_isvec)
log(" sv_cr_out", sv_cr_out)
log(" cr_bf", out_bitfield)
log(" spec", spec)
log(" override", sv_override)
# identify which regnames map to out / o2
+ if name == 'BF':
+ if out_sel == CROutSel.BF.value:
+ return out, o_isvec
if name == 'CR0':
if out_sel == CROutSel.CR0.value:
return out, o_isvec
- log("get_pdecode_cr_out not found", name)
+ if name == 'CR1': # these are not actually calculated correctly
+ if out_sel == CROutSel.CR1.value:
+ return out, o_isvec
+ # check RC1 set? if so return implicit vector, this is a REAL bad hack
+ RC1 = yield dec2.rm_dec.RC1
+ if RC1:
+ log("get_cr_out RC1 mode")
+ if name == 'CR0':
+ return 0, True # XXX TODO: offset CR0 from SVSTATE SPR
+ if name == 'CR1':
+ return 1, True # XXX TODO: offset CR1 from SVSTATE SPR
+ # nope - not found.
+ log("get_cr_out not found", name)
return None, False
# TODO, really should just be using PowerDecoder2
-def get_pdecode_idx_out(dec2, name):
+def get_out_map(dec2, name):
op = dec2.dec.op
out_sel = yield op.out_sel
# get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
out = yield dec2.e.write_reg.data
- o_isvec = yield dec2.o_isvec
# identify which regnames map to out / o2
if name == 'RA':
- log("get_pdecode_idx_out", out_sel, OutSel.RA.value, out, o_isvec)
if out_sel == OutSel.RA.value:
- return out, o_isvec
+ return True
elif name == 'RT':
- log("get_pdecode_idx_out", out_sel, OutSel.RT.value,
- OutSel.RT_OR_ZERO.value, out, o_isvec,
- dec2.dec.RT)
if out_sel == OutSel.RT.value:
- return out, o_isvec
+ return True
+ if out_sel == OutSel.RT_OR_ZERO.value and out != 0:
+ return True
elif name == 'RT_OR_ZERO':
- log("get_pdecode_idx_out", out_sel, OutSel.RT.value,
- OutSel.RT_OR_ZERO.value, out, o_isvec,
- dec2.dec.RT)
if out_sel == OutSel.RT_OR_ZERO.value:
- return out, o_isvec
+ return True
elif name == 'FRA':
- log("get_pdecode_idx_out", out_sel, OutSel.FRA.value, out, o_isvec)
if out_sel == OutSel.FRA.value:
- return out, o_isvec
+ return True
+ elif name == 'FRS':
+ if out_sel == OutSel.FRS.value:
+ return True
elif name == 'FRT':
- log("get_pdecode_idx_out", out_sel, OutSel.FRT.value,
- OutSel.FRT.value, out, o_isvec)
if out_sel == OutSel.FRT.value:
- return out, o_isvec
- log("get_pdecode_idx_out not found", name, out_sel, out, o_isvec)
+ return True
+ return False
+
+
+# TODO, really should just be using PowerDecoder2
+def get_idx_out(dec2, name, ewmode=False):
+ op = dec2.dec.op
+ out_sel = yield op.out_sel
+ # get the IN1/2/3 from the decoder (includes SVP64 remap and isvec)
+ out = yield dec2.e.write_reg.data
+ o_isvec = yield dec2.o_isvec
+ if ewmode:
+ offs = yield dec2.e.write_reg.offs
+ base = yield dec2.e.write_reg.base
+ out = (out, base, offs)
+ # identify which regnames map to out / o2
+ ismap = yield from get_out_map(dec2, name)
+ if ismap:
+ log("get_idx_out", name, out_sel, out, o_isvec)
+ return out, o_isvec
+ log("get_idx_out not found", name, out_sel, out, o_isvec)
return None, False
# TODO, really should just be using PowerDecoder2
-def get_pdecode_idx_out2(dec2, name):
+def get_out2_map(dec2, name):
# check first if register is activated for write
op = dec2.dec.op
out_sel = yield op.out_sel
out = yield dec2.e.write_ea.data
- o_isvec = yield dec2.o2_isvec
out_ok = yield dec2.e.write_ea.ok
- log("get_pdecode_idx_out2", name, out_sel, out, out_ok, o_isvec)
if not out_ok:
- return None, False
+ return False
- if name == 'RA':
+ if name in ['EA', 'RA']:
if hasattr(op, "upd"):
# update mode LD/ST uses read-reg A also as an output
upd = yield op.upd
- log("get_pdecode_idx_out2", upd, LDSTMode.update.value,
+ log("get_idx_out2", upd, LDSTMode.update.value,
out_sel, OutSel.RA.value,
- out, o_isvec)
+ out)
if upd == LDSTMode.update.value:
- return out, o_isvec
+ return True
+ if name == 'RS':
+ fft_en = yield dec2.implicit_rs
+ if fft_en:
+ log("get_idx_out2", out_sel, OutSel.RS.value,
+ out)
+ return True
if name == 'FRS':
- int_op = yield dec2.dec.op.internal_op
- fft_en = yield dec2.use_svp64_fft
- # if int_op == MicrOp.OP_FP_MADD.value and fft_en:
+ fft_en = yield dec2.implicit_rs
if fft_en:
- log("get_pdecode_idx_out2", out_sel, OutSel.FRS.value,
- out, o_isvec)
- return out, o_isvec
+ log("get_idx_out2", out_sel, OutSel.FRS.value,
+ out)
+ return True
+ return False
+
+
+# TODO, really should just be using PowerDecoder2
+def get_idx_out2(dec2, name, ewmode=False):
+ # check first if register is activated for write
+ op = dec2.dec.op
+ out_sel = yield op.out_sel
+ out = yield dec2.e.write_ea.data
+ if ewmode:
+ offs = yield dec2.e.write_ea.offs
+ base = yield dec2.e.write_ea.base
+ out = (out, base, offs)
+ o_isvec = yield dec2.o2_isvec
+ ismap = yield from get_out2_map(dec2, name)
+ if ismap:
+ log("get_idx_out2", name, out_sel, out, o_isvec)
+ return out, o_isvec
return None, False
-class ISACaller(ISACallerHelper, ISAFPHelpers):
+class StepLoop:
+ """deals with svstate looping.
+ """
+
+ def __init__(self, svstate):
+ self.svstate = svstate
+ self.new_iterators()
+
+ def new_iterators(self):
+ self.src_it = self.src_iterator()
+ self.dst_it = self.dst_iterator()
+ self.loopend = False
+ self.new_srcstep = 0
+ self.new_dststep = 0
+ self.new_ssubstep = 0
+ self.new_dsubstep = 0
+ self.pred_dst_zero = 0
+ self.pred_src_zero = 0
+
+ def src_iterator(self):
+ """source-stepping iterator
+ """
+ pack = self.svstate.pack
+
+ # source step
+ if pack:
+ # pack advances subvl in *outer* loop
+ while True: # outer subvl loop
+ while True: # inner vl loop
+ vl = self.svstate.vl
+ subvl = self.subvl
+ srcmask = self.srcmask
+ srcstep = self.svstate.srcstep
+ pred_src_zero = ((1 << srcstep) & srcmask) != 0
+ if self.pred_sz or pred_src_zero:
+ self.pred_src_zero = not pred_src_zero
+ log(" advance src", srcstep, vl,
+ self.svstate.ssubstep, subvl)
+ # yield actual substep/srcstep
+ yield (self.svstate.ssubstep, srcstep)
+ # the way yield works these could have been modified.
+ vl = self.svstate.vl
+ subvl = self.subvl
+ srcstep = self.svstate.srcstep
+ log(" advance src check", srcstep, vl,
+ self.svstate.ssubstep, subvl, srcstep == vl-1,
+ self.svstate.ssubstep == subvl)
+ if srcstep == vl-1: # end-point
+ self.svstate.srcstep = SelectableInt(0, 7) # reset
+ if self.svstate.ssubstep == subvl: # end-point
+ log(" advance pack stop")
+ return
+ break # exit inner loop
+ self.svstate.srcstep += SelectableInt(1, 7) # advance ss
+ subvl = self.subvl
+ if self.svstate.ssubstep == subvl: # end-point
+ self.svstate.ssubstep = SelectableInt(0, 2) # reset
+ log(" advance pack stop")
+ return
+ self.svstate.ssubstep += SelectableInt(1, 2)
+
+ else:
+ # these cannot be done as for-loops because SVSTATE may change
+ # (srcstep/substep may be modified, interrupted, subvl/vl change)
+ # but they *can* be done as while-loops as long as every SVSTATE
+ # "thing" is re-read every single time a yield gives indices
+ while True: # outer vl loop
+ while True: # inner subvl loop
+ vl = self.svstate.vl
+ subvl = self.subvl
+ srcmask = self.srcmask
+ srcstep = self.svstate.srcstep
+ pred_src_zero = ((1 << srcstep) & srcmask) != 0
+ if self.pred_sz or pred_src_zero:
+ self.pred_src_zero = not pred_src_zero
+ log(" advance src", srcstep, vl,
+ self.svstate.ssubstep, subvl)
+ # yield actual substep/srcstep
+ yield (self.svstate.ssubstep, srcstep)
+ if self.svstate.ssubstep == subvl: # end-point
+ self.svstate.ssubstep = SelectableInt(0, 2) # reset
+ break # exit inner loop
+ self.svstate.ssubstep += SelectableInt(1, 2)
+ vl = self.svstate.vl
+ if srcstep == vl-1: # end-point
+ self.svstate.srcstep = SelectableInt(0, 7) # reset
+ self.loopend = True
+ return
+ self.svstate.srcstep += SelectableInt(1, 7) # advance srcstep
+
+ def dst_iterator(self):
+ """dest-stepping iterator
+ """
+ unpack = self.svstate.unpack
+
+ # dest step
+ if unpack:
+ # pack advances subvl in *outer* loop
+ while True: # outer subvl loop
+ while True: # inner vl loop
+ vl = self.svstate.vl
+ subvl = self.subvl
+ dstmask = self.dstmask
+ dststep = self.svstate.dststep
+ pred_dst_zero = ((1 << dststep) & dstmask) != 0
+ if self.pred_dz or pred_dst_zero:
+ self.pred_dst_zero = not pred_dst_zero
+ log(" advance dst", dststep, vl,
+ self.svstate.dsubstep, subvl)
+ # yield actual substep/dststep
+ yield (self.svstate.dsubstep, dststep)
+ # the way yield works these could have been modified.
+ vl = self.svstate.vl
+ dststep = self.svstate.dststep
+ log(" advance dst check", dststep, vl,
+ self.svstate.ssubstep, subvl)
+ if dststep == vl-1: # end-point
+ self.svstate.dststep = SelectableInt(0, 7) # reset
+ if self.svstate.dsubstep == subvl: # end-point
+ log(" advance unpack stop")
+ return
+ break
+ self.svstate.dststep += SelectableInt(1, 7) # advance ds
+ subvl = self.subvl
+ if self.svstate.dsubstep == subvl: # end-point
+ self.svstate.dsubstep = SelectableInt(0, 2) # reset
+ log(" advance unpack stop")
+ return
+ self.svstate.dsubstep += SelectableInt(1, 2)
+ else:
+ # these cannot be done as for-loops because SVSTATE may change
+ # (dststep/substep may be modified, interrupted, subvl/vl change)
+ # but they *can* be done as while-loops as long as every SVSTATE
+ # "thing" is re-read every single time a yield gives indices
+ while True: # outer vl loop
+ while True: # inner subvl loop
+ subvl = self.subvl
+ dstmask = self.dstmask
+ dststep = self.svstate.dststep
+ pred_dst_zero = ((1 << dststep) & dstmask) != 0
+ if self.pred_dz or pred_dst_zero:
+ self.pred_dst_zero = not pred_dst_zero
+ log(" advance dst", dststep, self.svstate.vl,
+ self.svstate.dsubstep, subvl)
+ # yield actual substep/dststep
+ yield (self.svstate.dsubstep, dststep)
+ if self.svstate.dsubstep == subvl: # end-point
+ self.svstate.dsubstep = SelectableInt(0, 2) # reset
+ break
+ self.svstate.dsubstep += SelectableInt(1, 2)
+ subvl = self.subvl
+ vl = self.svstate.vl
+ if dststep == vl-1: # end-point
+ self.svstate.dststep = SelectableInt(0, 7) # reset
+ return
+ self.svstate.dststep += SelectableInt(1, 7) # advance dststep
+
+ def src_iterate(self):
+ """source-stepping iterator
+ """
+ subvl = self.subvl
+ vl = self.svstate.vl
+ pack = self.svstate.pack
+ unpack = self.svstate.unpack
+ ssubstep = self.svstate.ssubstep
+ end_ssub = ssubstep == subvl
+ end_src = self.svstate.srcstep == vl-1
+ log(" pack/unpack/subvl", pack, unpack, subvl,
+ "end", end_src,
+ "sub", end_ssub)
+ # first source step
+ srcstep = self.svstate.srcstep
+ srcmask = self.srcmask
+ if pack:
+ # pack advances subvl in *outer* loop
+ while True:
+ assert srcstep <= vl-1
+ end_src = srcstep == vl-1
+ if end_src:
+ if end_ssub:
+ self.loopend = True
+ else:
+ self.svstate.ssubstep += SelectableInt(1, 2)
+ srcstep = 0 # reset
+ break
+ else:
+ srcstep += 1 # advance srcstep
+ if not self.srcstep_skip:
+ break
+ if ((1 << srcstep) & srcmask) != 0:
+ break
+ else:
+ log(" sskip", bin(srcmask), bin(1 << srcstep))
+ else:
+ # advance subvl in *inner* loop
+ if end_ssub:
+ while True:
+ assert srcstep <= vl-1
+ end_src = srcstep == vl-1
+ if end_src: # end-point
+ self.loopend = True
+ srcstep = 0
+ break
+ else:
+ srcstep += 1
+ if not self.srcstep_skip:
+ break
+ if ((1 << srcstep) & srcmask) != 0:
+ break
+ else:
+ log(" sskip", bin(srcmask), bin(1 << srcstep))
+ self.svstate.ssubstep = SelectableInt(0, 2) # reset
+ else:
+ # advance ssubstep
+ self.svstate.ssubstep += SelectableInt(1, 2)
+
+ self.svstate.srcstep = SelectableInt(srcstep, 7)
+ log(" advance src", self.svstate.srcstep, self.svstate.ssubstep,
+ self.loopend)
+
+ def dst_iterate(self):
+ """dest step iterator
+ """
+ vl = self.svstate.vl
+ subvl = self.subvl
+ pack = self.svstate.pack
+ unpack = self.svstate.unpack
+ dsubstep = self.svstate.dsubstep
+ end_dsub = dsubstep == subvl
+ dststep = self.svstate.dststep
+ end_dst = dststep == vl-1
+ dstmask = self.dstmask
+ log(" pack/unpack/subvl", pack, unpack, subvl,
+ "end", end_dst,
+ "sub", end_dsub)
+ # now dest step
+ if unpack:
+ # unpack advances subvl in *outer* loop
+ while True:
+ assert dststep <= vl-1
+ end_dst = dststep == vl-1
+ if end_dst:
+ if end_dsub:
+ self.loopend = True
+ else:
+ self.svstate.dsubstep += SelectableInt(1, 2)
+ dststep = 0 # reset
+ break
+ else:
+ dststep += 1 # advance dststep
+ if not self.dststep_skip:
+ break
+ if ((1 << dststep) & dstmask) != 0:
+ break
+ else:
+ log(" dskip", bin(dstmask), bin(1 << dststep))
+ else:
+ # advance subvl in *inner* loop
+ if end_dsub:
+ while True:
+ assert dststep <= vl-1
+ end_dst = dststep == vl-1
+ if end_dst: # end-point
+ self.loopend = True
+ dststep = 0
+ break
+ else:
+ dststep += 1
+ if not self.dststep_skip:
+ break
+ if ((1 << dststep) & dstmask) != 0:
+ break
+ else:
+ log(" dskip", bin(dstmask), bin(1 << dststep))
+ self.svstate.dsubstep = SelectableInt(0, 2) # reset
+ else:
+ # advance ssubstep
+ self.svstate.dsubstep += SelectableInt(1, 2)
+
+ self.svstate.dststep = SelectableInt(dststep, 7)
+ log(" advance dst", self.svstate.dststep, self.svstate.dsubstep,
+ self.loopend)
+
+ def at_loopend(self):
+ """tells if this is the last possible element. uses the cached values
+ for src/dst-step and sub-steps
+ """
+ subvl = self.subvl
+ vl = self.svstate.vl
+ srcstep, dststep = self.new_srcstep, self.new_dststep
+ ssubstep, dsubstep = self.new_ssubstep, self.new_dsubstep
+ end_ssub = ssubstep == subvl
+ end_dsub = dsubstep == subvl
+ if srcstep == vl-1 and end_ssub:
+ return True
+ if dststep == vl-1 and end_dsub:
+ return True
+ return False
+
+ def advance_svstate_steps(self):
+ """ advance sub/steps. note that Pack/Unpack *INVERTS* the order.
+ TODO when Pack/Unpack is set, substep becomes the *outer* loop
+ """
+ self.subvl = yield self.dec2.rm_dec.rm_in.subvl
+ if self.loopend: # huhn??
+ return
+ self.src_iterate()
+ self.dst_iterate()
+
+ def read_src_mask(self):
+ """read/update pred_sz and src mask
+ """
+ # get SVSTATE VL (oh and print out some debug stuff)
+ vl = self.svstate.vl
+ srcstep = self.svstate.srcstep
+ ssubstep = self.svstate.ssubstep
+
+ # get predicate mask (all 64 bits)
+ srcmask = 0xffff_ffff_ffff_ffff
+
+ pmode = yield self.dec2.rm_dec.predmode
+ sv_ptype = yield self.dec2.dec.op.SV_Ptype
+ srcpred = yield self.dec2.rm_dec.srcpred
+ dstpred = yield self.dec2.rm_dec.dstpred
+ pred_sz = yield self.dec2.rm_dec.pred_sz
+ if pmode == SVP64PredMode.INT.value:
+ srcmask = dstmask = get_predint(self.gpr, dstpred)
+ if sv_ptype == SVPType.P2.value:
+ srcmask = get_predint(self.gpr, srcpred)
+ elif pmode == SVP64PredMode.CR.value:
+ srcmask = dstmask = get_predcr(self.crl, dstpred, vl)
+ if sv_ptype == SVPType.P2.value:
+ srcmask = get_predcr(self.crl, srcpred, vl)
+ # work out if the ssubsteps are completed
+ ssubstart = ssubstep == 0
+ log(" pmode", pmode)
+ log(" ptype", sv_ptype)
+ log(" srcpred", bin(srcpred))
+ log(" srcmask", bin(srcmask))
+ log(" pred_sz", bin(pred_sz))
+ log(" ssubstart", ssubstart)
+
+ # store all that above
+ self.srcstep_skip = False
+ self.srcmask = srcmask
+ self.pred_sz = pred_sz
+ self.new_ssubstep = ssubstep
+ log(" new ssubstep", ssubstep)
+ # until the predicate mask has a "1" bit... or we run out of VL
+ # let srcstep==VL be the indicator to move to next instruction
+ if not pred_sz:
+ self.srcstep_skip = True
+
+ def read_dst_mask(self):
+ """same as read_src_mask - check and record everything needed
+ """
+ # get SVSTATE VL (oh and print out some debug stuff)
+ # yield Delay(1e-10) # make changes visible
+ vl = self.svstate.vl
+ dststep = self.svstate.dststep
+ dsubstep = self.svstate.dsubstep
+
+ # get predicate mask (all 64 bits)
+ dstmask = 0xffff_ffff_ffff_ffff
+
+ pmode = yield self.dec2.rm_dec.predmode
+ reverse_gear = yield self.dec2.rm_dec.reverse_gear
+ sv_ptype = yield self.dec2.dec.op.SV_Ptype
+ dstpred = yield self.dec2.rm_dec.dstpred
+ pred_dz = yield self.dec2.rm_dec.pred_dz
+ if pmode == SVP64PredMode.INT.value:
+ dstmask = get_predint(self.gpr, dstpred)
+ elif pmode == SVP64PredMode.CR.value:
+ dstmask = get_predcr(self.crl, dstpred, vl)
+ # work out if the ssubsteps are completed
+ dsubstart = dsubstep == 0
+ log(" pmode", pmode)
+ log(" ptype", sv_ptype)
+ log(" dstpred", bin(dstpred))
+ log(" dstmask", bin(dstmask))
+ log(" pred_dz", bin(pred_dz))
+ log(" dsubstart", dsubstart)
+
+ self.dststep_skip = False
+ self.dstmask = dstmask
+ self.pred_dz = pred_dz
+ self.new_dsubstep = dsubstep
+ log(" new dsubstep", dsubstep)
+ if not pred_dz:
+ self.dststep_skip = True
+
+ def svstate_pre_inc(self):
+ """check if srcstep/dststep need to skip over masked-out predicate bits
+ note that this is not supposed to do anything to substep,
+ it is purely for skipping masked-out bits
+ """
+
+ self.subvl = yield self.dec2.rm_dec.rm_in.subvl
+ yield from self.read_src_mask()
+ yield from self.read_dst_mask()
+
+ self.skip_src()
+ self.skip_dst()
+
+ def skip_src(self):
+
+ srcstep = self.svstate.srcstep
+ srcmask = self.srcmask
+ pred_src_zero = self.pred_sz
+ vl = self.svstate.vl
+ # srcstep-skipping opportunity identified
+ if self.srcstep_skip:
+ # cannot do this with sv.bc - XXX TODO
+ if srcmask == 0:
+ self.loopend = True
+ while (((1 << srcstep) & srcmask) == 0) and (srcstep != vl):
+ log(" sskip", bin(1 << srcstep))
+ srcstep += 1
+
+ # now work out if the relevant mask bits require zeroing
+ if pred_src_zero:
+ pred_src_zero = ((1 << srcstep) & srcmask) == 0
+
+ # store new srcstep / dststep
+ self.new_srcstep = srcstep
+ self.pred_src_zero = pred_src_zero
+ log(" new srcstep", srcstep)
+
+ def skip_dst(self):
+ # dststep-skipping opportunity identified
+ dststep = self.svstate.dststep
+ dstmask = self.dstmask
+ pred_dst_zero = self.pred_dz
+ vl = self.svstate.vl
+ if self.dststep_skip:
+ # cannot do this with sv.bc - XXX TODO
+ if dstmask == 0:
+ self.loopend = True
+ while (((1 << dststep) & dstmask) == 0) and (dststep != vl):
+ log(" dskip", bin(1 << dststep))
+ dststep += 1
+
+ # now work out if the relevant mask bits require zeroing
+ if pred_dst_zero:
+ pred_dst_zero = ((1 << dststep) & dstmask) == 0
+
+ # store new srcstep / dststep
+ self.new_dststep = dststep
+ self.pred_dst_zero = pred_dst_zero
+ log(" new dststep", dststep)
+
+
+class ExitSyscallCalled(Exception):
+ pass
+
+
+class SyscallEmulator(openpower.syscalls.Dispatcher):
+ def __init__(self, isacaller):
+ self.__isacaller = isacaller
+
+ host = os.uname().machine
+ bits = (64 if (sys.maxsize > (2**32)) else 32)
+ host = openpower.syscalls.architecture(arch=host, bits=bits)
+
+ return super().__init__(guest="ppc64", host=host)
+
+ def __call__(self, identifier, *arguments):
+ (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
+
+
+class ISACaller(ISACallerHelper, ISAFPHelpers, StepLoop):
# 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
initial_pc=0,
bigendian=False,
mmu=False,
- icachemmu=False):
+ icachemmu=False,
+ initial_fpscr=0,
+ insnlog=None,
+ use_mmap_mem=False,
+ use_syscall_emu=False,
+ emulating_mmap=False):
+ if use_syscall_emu:
+ self.syscall = SyscallEmulator(isacaller=self)
+ if not use_mmap_mem:
+ log("forcing use_mmap_mem due to use_syscall_emu active")
+ use_mmap_mem = True
+ 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")
+ if not self.insnlog_is_file and self.insnlog:
+ self.insnlog = open(self.insnlog, "w")
self.bigendian = bigendian
self.halted = False
if initial_insns is None:
initial_insns = {}
assert self.respect_pc == False, "instructions required to honor pc"
+ if initial_msr is None:
+ initial_msr = DEFAULT_MSR
log("ISACaller insns", respect_pc, initial_insns, disassembly)
log("ISACaller initial_msr", initial_msr)
if isinstance(initial_svstate, int):
initial_svstate = SVP64State(initial_svstate)
# SVSTATE, MSR and PC
- self.svstate = initial_svstate
+ StepLoop.__init__(self, initial_svstate)
self.msr = SelectableInt(initial_msr, 64) # underlying reg
self.pc = PC()
# GPR FPR SPR registers
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
- if sname not in self.spr:
- self.spr[sname] = SVSHAPE(0)
- else:
- # make sure it's an SVSHAPE
- val = self.spr[sname].value
- self.spr[sname] = SVSHAPE(val)
+ val = self.spr.get(sname, 0)
+ # make sure it's an SVSHAPE -- conversion done by SPR.__setitem__
+ self.spr[sname] = val
self.last_op_svshape = False
# "raw" memory
- self.mem = Mem(row_bytes=8, initial_mem=initial_mem)
- self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
+ if use_mmap_mem:
+ self.mem = MemMMap(row_bytes=8,
+ initial_mem=initial_mem,
+ misaligned_ok=True,
+ emulating_mmap=emulating_mmap)
+ self.imem = self.mem
+ 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,
+ misaligned_ok=True)
+ self.mem.log_fancy(kind=LogType.InstrInOuts)
+ self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
# MMU mode, redirect underlying Mem through RADIX
if mmu:
self.mem = RADIX(self.mem, self)
# FPR (same as GPR except for FP nums)
# 4.2.2 p124 FPSCR (definitely "separate" - not in SPR)
# note that mffs, mcrfs, mtfsf "manage" this FPSCR
+ self.fpscr = FPSCRState(initial_fpscr)
+
# 2.3.1 CR (and sub-fields CR0..CR6 - CR0 SO comes from XER.SO)
# note that mfocrf, mfcr, mtcr, mtocrf, mcrxrx "manage" CRs
# -- Done
# create CR then allow portions of it to be "selectable" (below)
self.cr_fields = CRFields(initial_cr)
self.cr = self.cr_fields.cr
+ self.cr_backup = 0 # sigh, dreadful hack: for fail-first (VLi)
# "undefined", just set to variable-bit-width int (use exts "max")
- # self.undefined = SelectableInt(0, 256) # TODO, not hard-code 256!
+ # self.undefined = SelectableInt(0, EFFECTIVELY_UNLIMITED)
self.namespace = {}
self.namespace.update(self.spr)
'SVSHAPE3': self.spr['SVSHAPE3'],
'CR': self.cr,
'MSR': self.msr,
+ 'FPSCR': self.fpscr,
'undefined': undefined,
'mode_is_64bit': True,
'SO': XER_bits['SO'],
- 'XLEN': 64 # elwidth overrides, later
+ 'XLEN': 64 # elwidth overrides
})
+ for name in BFP_FLAG_NAMES:
+ setattr(self, name, 0)
+
# update pc to requested start point
self.set_pc(initial_pc)
self.decoder = decoder2.dec
self.dec2 = decoder2
- super().__init__(XLEN=self.namespace["XLEN"])
+ super().__init__(XLEN=self.namespace["XLEN"], FPSCR=self.fpscr)
+
+ def trace(self, out):
+ if self.insnlog is None:
+ return
+ self.insnlog.write(out)
@property
def XLEN(self):
return self.namespace["XLEN"]
+ @property
+ def FPSCR(self):
+ return self.fpscr
+
def call_trap(self, trap_addr, trap_bit):
"""calls TRAP and sets up NIA to the new execution location.
next instruction will begin at trap_addr.
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.
"""
- log("TRAP:", hex(trap_addr), hex(self.namespace['MSR'].value))
+ if isinstance(trap_addr, SelectableInt):
+ trap_addr = trap_addr.value
+ # https://bugs.libre-soc.org/show_bug.cgi?id=859
+ kaivb = self.spr['KAIVB'].value
+ msr = self.namespace['MSR'].value
+ log("TRAP:", hex(trap_addr), hex(msr), "kaivb", hex(kaivb))
# store CIA(+4?) in SRR0, set NIA to 0x700
# store MSR in SRR1, set MSR to um errr something, have to check spec
# store SVSTATE (if enabled) in SVSRR0
self.spr['SRR0'].value = self.pc.CIA.value
- self.spr['SRR1'].value = self.namespace['MSR'].value
+ self.spr['SRR1'].value = msr
if self.is_svp64_mode:
self.spr['SVSRR0'] = self.namespace['SVSTATE'].value
- self.trap_nia = SelectableInt(trap_addr, 64)
- self.spr['SRR1'][trap_bit] = 1 # change *copy* of MSR in SRR1
+ self.trap_nia = SelectableInt(trap_addr | (kaivb & ~0x1fff), 64)
+ 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
def memassign(self, ea, sz, val):
self.mem.memassign(ea, sz, val)
- def prep_namespace(self, insn_name, formname, op_fields):
+ def prep_namespace(self, insn_name, formname, op_fields, xlen):
# TODO: get field names from form in decoder*1* (not decoder2)
# decoder2 is hand-created, and decoder1.sigform is auto-generated
# from spec
# then "yield" fields only from op_fields rather than hard-coded
# list, here.
fields = self.decoder.sigforms[formname]
- log("prep_namespace", formname, op_fields)
+ log("prep_namespace", formname, op_fields, insn_name)
for name in op_fields:
# CR immediates. deal with separately. needs modifying
# pseudocode
if self.is_svp64_mode and name in ['BI']: # TODO, more CRs
# BI is a 5-bit, must reconstruct the value
- regnum, is_vec = yield from get_pdecode_cr_in(self.dec2, name)
+ regnum, is_vec = yield from get_cr_in(self.dec2, name)
sig = getattr(fields, name)
val = yield sig
# low 2 LSBs (CR field selector) remain same, CR num extended
assert regnum <= 7, "sigh, TODO, 128 CR fields"
val = (val & 0b11) | (regnum << 2)
+ elif self.is_svp64_mode and name in ['BF']: # TODO, more CRs
+ regnum, is_vec = yield from get_cr_out(self.dec2, "BF")
+ log('hack %s' % name, regnum, is_vec)
+ val = regnum
else:
- if name == 'spr':
- sig = getattr(fields, name.upper())
- else:
- sig = getattr(fields, name)
+ sig = getattr(fields, name)
val = yield sig
# these are all opcode fields involved in index-selection of CR,
# and need to do "standard" arithmetic. CR[BA+32] for example
self.namespace['XER'] = self.spr['XER']
self.namespace['CA'] = self.spr['XER'][XER_bits['CA']].value
self.namespace['CA32'] = self.spr['XER'][XER_bits['CA32']].value
+ self.namespace['OV'] = self.spr['XER'][XER_bits['OV']].value
+ self.namespace['OV32'] = self.spr['XER'][XER_bits['OV32']].value
+ self.namespace['XLEN'] = xlen
# add some SVSTATE convenience variables
vl = self.svstate.vl
self.namespace['VL'] = vl
self.namespace['srcstep'] = srcstep
- # 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
- bc_vlset = (mode & SVP64MODE.BC_VLSET) != 0
- bc_vli = (mode & SVP64MODE.BC_VLI) != 0
- bc_snz = (mode & SVP64MODE.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['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)
-
- def handle_carry_(self, inputs, outputs, already_done):
+ # take a copy of the CR field value: if non-VLi fail-first fails
+ # this is because the pseudocode writes *directly* to CR. sigh
+ self.cr_backup = self.cr.value
+
+ # sv.bc* need some extra fields
+ 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
+ duplicates - in a laborious and complex fashion - the contents of
+ the CSV files that were extracted two years ago from microwatt's
+ source code. A-inversion is the "inv A" column, output inversion
+ is the "inv out" column, carry-in equal to 0 or 1 or CA is the
+ "cry in" column
+
+ all of that information is available in
+ self.instrs[ins_name].op_fields
+ where info is usually assigned to self.instrs[ins_name]
+
+ https://git.libre-soc.org/?p=openpower-isa.git;a=blob;f=openpower/isatables/minor_31.csv;hb=HEAD
+
+ the immediate constants are *also* decoded correctly and placed
+ usually by DecodeIn2Imm into operand2, as part of power_decoder2.py
+ """
+ def ca(a, b, ca_in, width):
+ mask = (1 << width) - 1
+ y = (a & mask) + (b & mask) + ca_in
+ return y >> width
+
+ asmcode = yield self.dec2.dec.op.asmcode
+ insn = insns.get(asmcode)
+ SI = yield self.dec2.dec.SI
+ SI &= 0xFFFF
+ CA, OV = inp_ca_ov
+ inputs = [i.value for i in inputs]
+ if SI & 0x8000:
+ SI -= 0x10000
+ if insn in ("add", "addo", "addc", "addco"):
+ a = inputs[0]
+ b = inputs[1]
+ ca_in = 0
+ elif insn == "addic" or insn == "addic.":
+ a = inputs[0]
+ b = SI
+ ca_in = 0
+ elif insn in ("subf", "subfo", "subfc", "subfco"):
+ a = ~inputs[0]
+ b = inputs[1]
+ ca_in = 1
+ elif insn == "subfic":
+ a = ~inputs[0]
+ b = SI
+ ca_in = 1
+ elif insn == "adde" or insn == "addeo":
+ a = inputs[0]
+ b = inputs[1]
+ ca_in = CA
+ elif insn == "subfe" or insn == "subfeo":
+ a = ~inputs[0]
+ b = inputs[1]
+ ca_in = CA
+ elif insn == "addme" or insn == "addmeo":
+ a = inputs[0]
+ b = ~0
+ ca_in = CA
+ elif insn == "addze" or insn == "addzeo":
+ a = inputs[0]
+ b = 0
+ ca_in = CA
+ elif insn == "subfme" or insn == "subfmeo":
+ a = ~inputs[0]
+ b = ~0
+ ca_in = CA
+ elif insn == "subfze" or insn == "subfzeo":
+ a = ~inputs[0]
+ b = 0
+ ca_in = CA
+ elif insn == "addex":
+ # CA[32] aren't actually written, just generate so we have
+ # something to return
+ ca64 = ov64 = ca(inputs[0], inputs[1], OV, 64)
+ ca32 = ov32 = ca(inputs[0], inputs[1], OV, 32)
+ return ca64, ca32, ov64, ov32
+ elif insn == "neg" or insn == "nego":
+ a = ~inputs[0]
+ b = 0
+ ca_in = 1
+ else:
+ raise NotImplementedError(
+ "op_add kludge unimplemented instruction: ", asmcode, insn)
+
+ ca64 = ca(a, b, ca_in, 64)
+ ca32 = ca(a, b, ca_in, 32)
+ ov64 = ca64 != ca(a, b, ca_in, 63)
+ ov32 = ca32 != ca(a, b, ca_in, 31)
+ 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(
+ inputs, inp_ca_ov)
+ ca, ca32, ov, ov32 = retval
+ asmcode = yield self.dec2.dec.op.asmcode
+ if insns.get(asmcode) == 'addex':
+ # TODO: if 32-bit mode, set ov to ov32
+ self.spr['XER'][XER_bits['OV']] = ov
+ self.spr['XER'][XER_bits['OV32']] = ov32
+ log(f"write OV/OV32 OV={ov} OV32={ov32}",
+ kind=LogType.InstrInOuts)
+ else:
+ # TODO: if 32-bit mode, set ca to ca32
+ self.spr['XER'][XER_bits['CA']] = ca
+ self.spr['XER'][XER_bits['CA32']] = ca32
+ log(f"write CA/CA32 CA={ca} CA32={ca32}",
+ kind=LogType.InstrInOuts)
+ return
inv_a = yield self.dec2.e.do.invert_in
if inv_a:
inputs[0] = ~inputs[0]
if imm_ok:
imm = yield self.dec2.e.do.imm_data.data
inputs.append(SelectableInt(imm, 64))
- assert len(outputs) >= 1
- log("outputs", repr(outputs))
- if isinstance(outputs, list) or isinstance(outputs, tuple):
- output = outputs[0]
- else:
- output = outputs
gts = []
for x in inputs:
log("gt input", x, output)
log(gts)
cy = 1 if any(gts) else 0
log("CA", cy, gts)
- if not (1 & already_done):
+ if ca is None: # already written
self.spr['XER'][XER_bits['CA']] = cy
- log("inputs", already_done, inputs)
# 32 bit carry
# ARGH... different for OP_ADD... *sigh*...
op = yield self.dec2.e.do.insn_type
gts.append(gt)
cy32 = 1 if any(gts) else 0
log("CA32", cy32, gts)
- if not (2 & already_done):
+ if ca32 is None: # already written
self.spr['XER'][XER_bits['CA32']] = cy32
- def handle_overflow(self, inputs, outputs, div_overflow):
+ def handle_overflow(self, inputs, output, div_overflow, inp_ca_ov):
+ op = yield self.dec2.e.do.insn_type
+ if op == MicrOp.OP_ADD.value:
+ retval = yield from self.get_kludged_op_add_ca_ov(
+ inputs, inp_ca_ov)
+ ca, ca32, ov, ov32 = retval
+ # TODO: if 32-bit mode, set ov to ov32
+ self.spr['XER'][XER_bits['OV']] = ov
+ self.spr['XER'][XER_bits['OV32']] = ov32
+ self.spr['XER'][XER_bits['SO']] |= ov
+ return
if hasattr(self.dec2.e.do, "invert_in"):
inv_a = yield self.dec2.e.do.invert_in
if inv_a:
if imm_ok:
imm = yield self.dec2.e.do.imm_data.data
inputs.append(SelectableInt(imm, 64))
- assert len(outputs) >= 1
- log("handle_overflow", inputs, outputs, div_overflow)
+ log("handle_overflow", inputs, output, div_overflow)
if len(inputs) < 2 and div_overflow is None:
return
ov, ov32 = div_overflow, div_overflow
# arithmetic overflow can be done by analysing the input and output
elif len(inputs) >= 2:
- output = outputs[0]
-
# OV (64-bit)
input_sgn = [exts(x.value, x.bits) < 0 for x in inputs]
output_sgn = exts(output.value, output.bits) < 0
ov32 = 1 if input32_sgn[0] == input32_sgn[1] and \
output32_sgn != input32_sgn[0] else 0
+ # now update XER OV/OV32/SO
+ so = self.spr['XER'][XER_bits['SO']]
+ new_so = so | ov # sticky overflow ORs in old with new
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['SO']] = new_so
+ log(" set overflow", ov, ov32, so, new_so)
- def handle_comparison(self, outputs, cr_idx=0):
- out = outputs[0]
+ def handle_comparison(self, out, cr_idx=0, overflow=None, no_so=False):
assert isinstance(out, SelectableInt), \
"out zero not a SelectableInt %s" % repr(outputs)
log("handle_comparison", out.bits, hex(out.value))
# print ("handle_comparison exts 32 bit", hex(o32))
out = exts(out.value, out.bits)
log("handle_comparison exts", hex(out))
+ # create the three main CR flags, EQ GT LT
zero = SelectableInt(out == 0, 1)
positive = SelectableInt(out > 0, 1)
negative = SelectableInt(out < 0, 1)
- SO = self.spr['XER'][XER_bits['SO']]
- log("handle_comparison SO", SO)
+ # get (or not) XER.SO. for setvl this is important *not* to read SO
+ if no_so:
+ SO = SelectableInt(1, 0)
+ else:
+ SO = self.spr['XER'][XER_bits['SO']]
+ log("handle_comparison SO", SO.value,
+ "overflow", overflow,
+ "zero", zero.value,
+ "+ve", positive.value,
+ "-ve", negative.value)
+ # alternative overflow checking (setvl mainly at the moment)
+ if overflow is not None and overflow == 1:
+ SO = SelectableInt(1, 1)
+ # create the four CR field values and set the required CR field
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)
pc, insn = self.get_next_insn()
yield from self.setup_next_insn(pc, insn)
+ # cache since it's really slow to construct
+ __PREFIX_CACHE = SVP64Instruction.Prefix(SelectableInt(value=0, bits=32))
+
+ def __decode_prefix(self, opcode):
+ pfx = self.__PREFIX_CACHE
+ pfx.storage.eq(opcode)
+ return pfx
+
def setup_next_insn(self, pc, ins):
"""set up next instruction
"""
# SVP64. first, check if the opcode is EXT001, and SVP64 id bits set
yield Settle()
opcode = yield self.dec2.dec.opcode_in
- pfx = SVP64PrefixFields() # TODO should probably use SVP64PrefixDecoder
- pfx.insn.value = opcode
- major = pfx.major.asint(msb0=True) # MSB0 inversion
- log("prefix test: opcode:", major, bin(major),
- pfx.insn[7] == 0b1, pfx.insn[9] == 0b1)
- self.is_svp64_mode = ((major == 0b000001) and
- pfx.insn[7].value == 0b1 and
- pfx.insn[9].value == 0b1)
+ opcode = SelectableInt(value=opcode, bits=32)
+ pfx = self.__decode_prefix(opcode)
+ log("prefix test: opcode:", pfx.PO, bin(pfx.PO), pfx.id)
+ self.is_svp64_mode = bool((pfx.PO == 0b000001) and (pfx.id == 0b11))
self.pc.update_nia(self.is_svp64_mode)
# set SVP64 decode
yield self.dec2.is_svp64_mode.eq(self.is_svp64_mode)
return
# in SVP64 mode. decode/print out svp64 prefix, get v3.0B instruction
- log("svp64.rm", bin(pfx.rm.asint(msb0=True)))
+ log("svp64.rm", bin(pfx.rm))
log(" svstate.vl", self.svstate.vl)
log(" svstate.mvl", self.svstate.maxvl)
- sv_rm = pfx.rm.asint(msb0=True)
ins = self.imem.ld(pc+4, 4, False, True, instr_fetch=True)
log(" svsetup: 0x%x 0x%x %s" % (pc+4, ins & 0xffffffff, bin(ins)))
yield self.dec2.dec.raw_opcode_in.eq(ins & 0xffffffff) # v3.0B suffix
- yield self.dec2.sv_rm.eq(sv_rm) # svp64 prefix
+ yield self.dec2.sv_rm.eq(int(pfx.rm)) # svp64 prefix
yield Settle()
def execute_one(self):
"""execute one instruction
"""
# get the disassembly code for this instruction
- if self.is_svp64_mode:
- if not self.disassembly:
- code = yield from self.get_assembly_name()
- else:
- code = self.disassembly[self._pc+4]
- log(" svp64 sim-execute", hex(self._pc), code)
+ if not self.disassembly:
+ code = yield from self.get_assembly_name()
else:
- if not self.disassembly:
- code = yield from self.get_assembly_name()
- else:
- code = self.disassembly[self._pc]
- log("sim-execute", hex(self._pc), code)
+ offs, dbg = 0, ""
+ if self.is_svp64_mode:
+ offs, dbg = 4, "svp64 "
+ code = self.disassembly[self._pc+offs]
+ log(" %s sim-execute" % dbg, hex(self._pc), code)
opname = code.split(' ')[0]
try:
yield from self.call(opname) # execute the instruction
except MemException as e: # check for memory errors
if e.args[0] == 'unaligned': # alignment error
- # run a Trap but set DAR first
- print("memory unaligned exception, DAR", e.dar)
+ # run a Trap but set DAR first
+ print("memory unaligned exception, DAR", e.dar, repr(e))
self.spr['DAR'] = SelectableInt(e.dar, 64)
self.call_trap(0x600, PIb.PRIV) # 0x600, privileged
return
elif e.args[0] == 'invalid': # invalid
- # run a Trap but set DAR first
+ # run a Trap but set DAR first
log("RADIX MMU memory invalid error, mode %s" % e.mode)
if e.mode == 'EXECUTE':
# XXX TODO: must set a few bits in SRR1,
# not supported yet:
raise e # ... re-raise
+ # append to the trace log file
+ self.trace(" # %s\n" % code)
+
+ log("gprs after code", code)
+ self.gpr.dump()
+ crs = []
+ for i in range(len(self.crl)):
+ crs.append(bin(self.crl[i].asint()))
+ log("crs", " ".join(crs))
+ log("vl,maxvl", self.svstate.vl, self.svstate.maxvl)
+
# don't use this except in special circumstances
if not self.respect_pc:
self.fake_pc += 4
else:
rc_en = False
rc_ok = False
+ # annoying: ignore rc_ok if RC1 is set (for creating *assembly name*)
+ RC1 = yield self.dec2.rm_dec.RC1
+ if RC1:
+ rc_en = False
+ rc_ok = False
# grrrr have to special-case MUL op (see DecodeOE)
log("ov %d en %d rc %d en %d op %d" %
(ov_ok, ov_en, rc_ok, rc_en, int_op))
asmop = 'mtcrf'
return asmop
+ def reset_remaps(self):
+ self.remap_loopends = [0] * 4
+ self.remap_idxs = [0, 1, 2, 3]
+
def get_remap_indices(self):
"""WARNING, this function stores remap_idxs and remap_loopends
in the class for later use. this to avoid problems with yield
"""
# go through all iterators in lock-step, advance to next remap_idx
- srcstep, dststep = self.get_src_dststeps()
+ srcstep, dststep, ssubstep, dsubstep = self.get_src_dststeps()
# get four SVSHAPEs. here we are hard-coding
+ self.reset_remaps()
SVSHAPE0 = self.spr['SVSHAPE0']
SVSHAPE1 = self.spr['SVSHAPE1']
SVSHAPE2 = self.spr['SVSHAPE2']
(SVSHAPE3, SVSHAPE3.get_iterator()),
]
- self.remap_loopends = [0] * 4
- self.remap_idxs = [0, 1, 2, 3]
dbg = []
for i, (shape, remap) in enumerate(remaps):
# zero is "disabled"
dec_insn = yield self.dec2.e.do.insn
return dec_insn & (1 << 20) != 0 # sigh - XFF.spr[-1]?
- def call(self, name):
+ def call(self, name, syscall_emu_active=False):
"""call(opcode) - the primary execution point for instructions
"""
self.last_st_addr = None # reset the last known store address
# 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()
- log("call", ins_name, asmop)
+ log("call", ins_name, asmop,
+ kind=LogType.InstrInOuts)
+
+ # sv.setvl is *not* a loop-function. sigh
+ log("is_svp64_mode", self.is_svp64_mode, asmop)
# check privileged
int_op = yield self.dec2.dec.op.internal_op
self.halted = True
return
+ # User mode system call emulation consists of several steps:
+ # 1. Detect whether instruction is sc or scv.
+ # 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.
+ if ((asmop in ("sc", "scv")) and
+ (self.syscall is not None) and
+ not syscall_emu_active):
+ # Memoize PC and trigger an interrupt
+ if self.respect_pc:
+ pc = self.pc.CIA.value
+ else:
+ pc = self.fake_pc
+ yield from self.call(asmop, syscall_emu_active=True)
+
+ # Reroute the syscall to host OS
+ identifier = self.gpr(0)
+ arguments = map(self.gpr, range(3, 9))
+ result = self.syscall(identifier, *arguments)
+ self.gpr.write(3, result, False, self.namespace["XLEN"])
+
+ # Return from interrupt
+ yield from self.call("rfid", syscall_emu_active=True)
+ return
+ elif ((name in ("rfid", "hrfid")) and syscall_emu_active):
+ asmop = "rfid"
+
# check illegal instruction
illegal = False
if ins_name not in ['mtcrf', 'mtocrf']:
illegal = ins_name != asmop
- # sigh deal with setvl not being supported by binutils (.long)
- if asmop.startswith('setvl'):
- illegal = False
- ins_name = 'setvl'
-
- # and svstep not being supported by binutils (.long)
- if asmop.startswith('svstep'):
- illegal = False
- ins_name = 'svstep'
-
- # and svremap not being supported by binutils (.long)
- if asmop.startswith('svremap'):
- illegal = False
- ins_name = 'svremap'
-
- # and svshape not being supported by binutils (.long)
- if asmop.startswith('svshape'):
- illegal = False
- ins_name = 'svshape'
-
- # and fsin and fcos
- if asmop == 'fsins':
- illegal = False
- ins_name = 'fsins'
- if asmop == 'fcoss':
- illegal = False
- ins_name = 'fcoss'
-
- # sigh also deal with ffmadds not being supported by binutils (.long)
- if asmop == 'ffmadds':
- illegal = False
- ins_name = 'ffmadds'
-
- # and fdmadds not being supported by binutils (.long)
- if asmop == 'fdmadds':
- illegal = False
- ins_name = 'fdmadds'
-
- # and ffadds not being supported by binutils (.long)
- if asmop == 'ffadds':
- illegal = False
- ins_name = 'ffadds'
-
- # and min/max/su
- if asmop in ['mins', 'maxs', 'minu', 'maxu',
- 'mins.', 'maxs.', 'minu.', 'maxu.']:
- illegal = False
- ins_name = asmop
-
- # and anything avgadd
- if asmop.startswith('avgadd'):
- illegal = False
- ins_name = asmop
-
- # and anything absdu
- if asmop.startswith('absdu'):
- illegal = False
- ins_name = asmop
-
- # and anything absadd
- if asmop.startswith('absadd'):
- illegal = False
- ins_name = asmop
-
- # and anything ternlog
- if asmop.startswith('ternlog'):
+ # list of instructions not being supported by binutils (.long)
+ dotstrp = asmop[:-1] if asmop[-1] == '.' else asmop
+ if dotstrp in [*FPTRANS_INSNS,
+ *LDST_UPDATE_INSNS,
+ 'ffmadds', 'fdmadds', 'ffadds',
+ 'minmax',
+ "brh", "brw", "brd",
+ 'setvl', 'svindex', 'svremap', 'svstep',
+ 'svshape', 'svshape2',
+ 'ternlogi', 'bmask', 'cprop', 'gbbd',
+ 'absdu', 'absds', 'absdacs', 'absdacu', 'avgadd',
+ 'fmvis', 'fishmv', 'pcdec', "maddedu", "divmod2du",
+ "dsld", "dsrd", "maddedus",
+ "sadd", "saddw", "sadduw",
+ "cffpr", "cffpro",
+ "mffpr", "mffprs",
+ "ctfpr", "ctfprs",
+ "mtfpr", "mtfprs",
+ "maddsubrs", "maddrs", "msubrs",
+ "cfuged", "cntlzdm", "cnttzdm", "pdepd", "pextd",
+ "setbc", "setbcr", "setnbc", "setnbcr",
+ ]:
illegal = False
- ins_name = asmop
+ ins_name = dotstrp
- # and anything grev
- if asmop.startswith('grev'):
+ # match against instructions treated as nop, see nop below
+ if asmop.startswith("dcbt"):
illegal = False
- ins_name = asmop
+ ins_name = "nop"
# branch-conditional redirects to sv.bc
if asmop.startswith('bc') and self.is_svp64_mode:
ins_name = 'sv.%s' % ins_name
- log(" post-processed name", ins_name, asmop)
+ # ld-immediate-with-pi mode redirects to ld-with-postinc
+ ldst_imm_postinc = False
+ if 'u' in ins_name and self.is_svp64_mode:
+ ldst_pi = yield self.dec2.rm_dec.ldst_postinc
+ if ldst_pi:
+ ins_name = ins_name.replace("u", "up")
+ ldst_imm_postinc = True
+ log(" enable ld/st postinc", ins_name)
+
+ log(" post-processed name", dotstrp, ins_name, asmop)
# illegal instructions call TRAP at 0x700
if illegal:
# nop has to be supported, we could let the actual op calculate
# but PowerDecoder has a pattern for nop
- if ins_name is 'nop':
+ if ins_name == 'nop':
self.update_pc_next()
return
+ # get elwidths, defaults to 64
+ xlen = 64
+ ew_src = 64
+ ew_dst = 64
+ if self.is_svp64_mode:
+ ew_src = yield self.dec2.rm_dec.ew_src
+ ew_dst = yield self.dec2.rm_dec.ew_dst
+ 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("elwidth", ew_src, ew_dst)
+ log("XLEN:", self.is_svp64_mode, xlen)
+
# look up instruction in ISA.instrs, prepare namespace
- info = self.instrs[ins_name]
- yield from self.prep_namespace(ins_name, info.form, info.op_fields)
+ if ins_name == 'pcdec': # grrrr yes there are others ("stbcx." etc.)
+ info = self.instrs[ins_name+"."]
+ elif asmop[-1] == '.' and asmop in self.instrs:
+ info = self.instrs[asmop]
+ else:
+ info = self.instrs[ins_name]
+ yield from self.prep_namespace(ins_name, info.form, info.op_fields,
+ xlen)
+
+ # dict retains order
+ inputs = dict.fromkeys(create_full_args(
+ read_regs=info.read_regs, special_regs=info.special_regs,
+ uninit_regs=info.uninit_regs, write_regs=info.write_regs))
# preserve order of register names
- input_names = create_args(list(info.read_regs) +
- list(info.uninit_regs))
+ write_without_special_regs = OrderedSet(info.write_regs)
+ write_without_special_regs -= OrderedSet(info.special_regs)
+ input_names = create_args([
+ *info.read_regs, *info.uninit_regs, *write_without_special_regs])
log("input names", input_names)
# get SVP64 entry for the current instruction
log("sv rm", sv_rm, dest_cr, src_cr, src_byname, dest_byname)
# see if srcstep/dststep need skipping over masked-out predicate bits
- if (self.is_svp64_mode or ins_name == 'setvl' or
- ins_name in ['svremap', 'svstate']):
+ # svstep also needs advancement because it calls SVSTATE_NEXT.
+ # bit the remaps get computed just after pre_inc moves them on
+ # with remap_set_steps substituting for PowerDecider2 not doing it,
+ # and SVSTATE_NEXT not being able to.use yield, the preinc on
+ # svstep is necessary for now.
+ self.reset_remaps()
+ if (self.is_svp64_mode or ins_name in ['svstep']):
yield from self.svstate_pre_inc()
if self.is_svp64_mode:
pre = yield from self.update_new_svstate_steps()
self.update_nia()
self.update_pc_next()
return
- srcstep, dststep = self.get_src_dststeps()
+ srcstep, dststep, ssubstep, dsubstep = self.get_src_dststeps()
pred_dst_zero = self.pred_dst_zero
pred_src_zero = self.pred_src_zero
vl = self.svstate.vl
+ subvl = yield self.dec2.rm_dec.rm_in.subvl
# VL=0 in SVP64 mode means "do nothing: skip instruction"
if self.is_svp64_mode and vl == 0:
self.pc.update(self.namespace, self.is_svp64_mode)
log("SVP64: VL=0, end of call", self.namespace['CIA'],
- self.namespace['NIA'])
+ self.namespace['NIA'], kind=LogType.InstrInOuts)
return
# for when SVREMAP is active, using pre-arranged schedule.
if persist or self.last_op_svshape:
remaps = self.get_remap_indices()
if self.is_svp64_mode and (persist or self.last_op_svshape):
- # just some convenient debug info
- for i in range(4):
- sname = 'SVSHAPE%d' % i
- shape = self.spr[sname]
- log(sname, bin(shape.value))
- log(" lims", shape.lims)
- log(" mode", shape.mode)
- log(" skip", shape.skip)
-
- # set up the list of steps to remap
- mi0 = self.svstate.mi0
- mi1 = self.svstate.mi1
- mi2 = self.svstate.mi2
- mo0 = self.svstate.mo0
- mo1 = self.svstate.mo1
- steps = [(self.dec2.in1_step, mi0), # RA
- (self.dec2.in2_step, mi1), # RB
- (self.dec2.in3_step, mi2), # RC
- (self.dec2.o_step, mo0), # RT
- (self.dec2.o2_step, mo1), # EA
- ]
- remap_idxs = self.remap_idxs
- rremaps = []
- # now cross-index the required SHAPE for each of 3-in 2-out regs
- rnames = ['RA', 'RB', 'RC', 'RT', 'EA']
- for i, (dstep, shape_idx) in enumerate(steps):
- (shape, remap) = remaps[shape_idx]
- remap_idx = remap_idxs[shape_idx]
- # zero is "disabled"
- if shape.value == 0x0:
- continue
- # now set the actual requested step to the current index
- yield dstep.eq(remap_idx)
-
- # debug printout info
- rremaps.append((shape.mode, i, rnames[i], shape_idx,
- remap_idx))
- for x in rremaps:
- log("shape remap", x)
+ yield from self.remap_set_steps(remaps)
# after that, settle down (combinatorial) to let Vector reg numbers
# work themselves out
yield Settle()
log("remap active", bin(remap_active))
# main input registers (RT, RA ...)
- inputs = []
for name in input_names:
- # using PowerDecoder2, first, find the decoder index.
- # (mapping name RA RB RC RS to in1, in2, in3)
- regnum, is_vec = yield from get_pdecode_idx_in(self.dec2, name)
- if regnum is None:
- # doing this is not part of svp64, it's because output
- # registers, to be modified, need to be in the namespace.
- regnum, is_vec = yield from get_pdecode_idx_out(self.dec2, name)
- if regnum is None:
- regnum, is_vec = yield from get_pdecode_idx_out2(self.dec2,
- name)
-
- # in case getting the register number is needed, _RA, _RB
- regname = "_" + name
- self.namespace[regname] = regnum
- if not self.is_svp64_mode or not pred_src_zero:
- log('reading reg %s %s' % (name, str(regnum)), is_vec)
- if name in fregs:
- reg_val = SelectableInt(self.fpr(regnum))
- elif name is not None:
- reg_val = SelectableInt(self.gpr(regnum))
+ if name == "overflow":
+ inputs[name] = SelectableInt(0, 1)
+ elif name == "FPSCR":
+ inputs[name] = self.FPSCR
+ elif name in ("CA", "CA32", "OV", "OV32"):
+ inputs[name] = self.spr['XER'][XER_bits[name]]
+ elif name in "CR0":
+ inputs[name] = self.crl[0]
+ elif name in spr_byname:
+ inputs[name] = self.spr[name]
else:
- log('zero input reg %s %s' % (name, str(regnum)), is_vec)
- reg_val = 0
- inputs.append(reg_val)
+ regval = (yield from self.get_input(name, ew_src, xlen))
+ log("regval name", name, regval)
+ inputs[name] = regval
+
# arrrrgh, awful hack, to get _RT into namespace
if ins_name in ['setvl', 'svstep']:
regname = "_RT"
self.namespace[regname] = SelectableInt(RT, 5)
if RT == 0:
self.namespace["RT"] = SelectableInt(0, 5)
- regnum, is_vec = yield from get_pdecode_idx_out(self.dec2, "RT")
+ regnum, is_vec = yield from get_idx_out(self.dec2, "RT")
log('hack input reg %s %s' % (name, str(regnum)), is_vec)
# in SVP64 mode for LD/ST work out immediate
# XXX TODO: replace_ds for DS-Form rather than D-Form.
# use info.form to detect
- replace_d = False # update / replace constant in pseudocode
- if self.is_svp64_mode:
- ldstmode = yield self.dec2.rm_dec.ldstmode
- # shift mode reads SVD (or SVDS - TODO)
- # *BUT*... because this is "overloading" of LD operations,
- # it gets *STORED* into D (or DS, TODO)
- if ldstmode == SVP64LDSTmode.SHIFT.value:
- imm = yield self.dec2.dec.fields.FormSVD.SVD[0:11]
- imm = exts(imm, 11) # sign-extend to integer
- log("shift SVD", imm)
- replace_d = True
- else:
- if info.form == 'DS':
- # DS-Form, multiply by 4 then knock 2 bits off after
- imm = yield self.dec2.dec.fields.FormDS.DS[0:14] * 4
- else:
- imm = yield self.dec2.dec.fields.FormD.D[0:16]
- imm = exts(imm, 16) # sign-extend to integer
- # get the right step. LD is from srcstep, ST is dststep
- op = yield self.dec2.e.do.insn_type
- offsmul = 0
- if op == MicrOp.OP_LOAD.value:
- if remap_active:
- offsmul = yield self.dec2.in1_step
- log("D-field REMAP src", imm, offsmul)
- else:
- offsmul = srcstep
- log("D-field src", imm, offsmul)
- elif op == MicrOp.OP_STORE.value:
- # XXX NOTE! no bit-reversed STORE! this should not ever be used
- offsmul = dststep
- log("D-field dst", imm, offsmul)
- # bit-reverse mode, rev already done through get_src_dst_steps()
- if ldstmode == SVP64LDSTmode.SHIFT.value:
- # manually look up RC, sigh
- RC = yield self.dec2.dec.RC[0:5]
- RC = self.gpr(RC)
- log("LD-SHIFT:", "VL", vl,
- "RC", RC.value, "imm", imm,
- "offs", bin(offsmul),
- )
- imm = SelectableInt((imm * offsmul) << RC.value, 32)
- # Unit-Strided LD/ST adds offset*width to immediate
- elif ldstmode == SVP64LDSTmode.UNITSTRIDE.value:
- ldst_len = yield self.dec2.e.do.data_len
- imm = SelectableInt(imm + offsmul * ldst_len, 32)
- replace_d = True
- # Element-strided multiplies the immediate by element step
- elif ldstmode == SVP64LDSTmode.ELSTRIDE.value:
- imm = SelectableInt(imm * offsmul, 32)
- replace_d = True
- if replace_d:
- ldst_ra_vec = yield self.dec2.rm_dec.ldst_ra_vec
- ldst_imz_in = yield self.dec2.rm_dec.ldst_imz_in
- log("LDSTmode", SVP64LDSTmode(ldstmode),
- offsmul, imm, ldst_ra_vec, ldst_imz_in)
- # new replacement D... errr.. DS
- if replace_d:
- if info.form == 'DS':
- # TODO: assert 2 LSBs are zero?
- log("DS-Form, TODO, assert 2 LSBs zero?", bin(imm.value))
- imm.value = imm.value >> 2
- self.namespace['DS'] = imm
- else:
- self.namespace['D'] = imm
+ if self.is_svp64_mode and not ldst_imm_postinc:
+ yield from self.check_replace_d(info, remap_active)
# "special" registers
for special in info.special_regs:
if special in special_sprs:
- inputs.append(self.spr[special])
+ inputs[special] = self.spr[special]
else:
- inputs.append(self.namespace[special])
+ inputs[special] = self.namespace[special]
# clear trap (trap) NIA
self.trap_nia = None
# 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,
+ self.spr['XER'][XER_bits['OV']].value)
+
+ for k, v in inputs.items():
+ if v is None:
+ v = SelectableInt(0, self.XLEN)
+ # prevent pseudo-code from modifying input registers
+ v = copy_assign_rhs(v)
+ if isinstance(v, SelectableInt):
+ v.ok = False
+ inputs[k] = v
+
# execute actual instruction here (finally)
log("inputs", inputs)
+ inputs = list(inputs.values())
results = info.func(self, *inputs)
- log("results", results)
+ 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)
self.last_st_addr, self.last_ld_addr)
# 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
-
- log("carry already done?", bin(already_done))
- if hasattr(self.dec2.e.do, "output_carry"):
- carry_en = yield self.dec2.e.do.output_carry
- else:
- carry_en = False
- if carry_en:
- yield from self.handle_carry_(inputs, results, already_done)
+ ca = outs.get("CA")
+ ca32 = outs.get("CA32")
- if not self.is_svp64_mode: # yeah just no. not in parallel processing
+ log("carry already done?", ca, ca32, output_names)
+ # soc test_pipe_caller tests don't have output_carry
+ has_output_carry = hasattr(self.dec2.e.do, "output_carry")
+ carry_en = has_output_carry and (yield self.dec2.e.do.output_carry)
+ if carry_en:
+ yield from self.handle_carry_(
+ inputs, results[0], ca, ca32, inp_ca_ov=inp_ca_ov)
+
+ # get output named "overflow" and "CR0"
+ overflow = outs.get('overflow')
+ cr0 = outs.get('CR0')
+ cr1 = outs.get('CR1')
+
+ # soc test_pipe_caller tests don't have oe
+ has_oe = hasattr(self.dec2.e.do, "oe")
+ # yeah just no. not in parallel processing
+ if has_oe and not self.is_svp64_mode:
# 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
-
- if hasattr(self.dec2.e.do, "oe"):
- ov_en = yield self.dec2.e.do.oe.oe
- ov_ok = yield self.dec2.e.do.oe.ok
- else:
- ov_en = False
- ov_ok = False
- log("internal overflow", overflow, ov_en, ov_ok)
+ ov_en = yield self.dec2.e.do.oe.oe
+ ov_ok = yield self.dec2.e.do.oe.ok
+ log("internal overflow", ins_name, overflow, "en?", ov_en, ov_ok)
if ov_en & ov_ok:
- yield from self.handle_overflow(inputs, results, overflow)
+ yield from self.handle_overflow(
+ inputs, results[0], overflow, inp_ca_ov=inp_ca_ov)
# only do SVP64 dest predicated Rc=1 if dest-pred is not enabled
rc_en = False
if not self.is_svp64_mode or not pred_dst_zero:
if hasattr(self.dec2.e.do, "rc"):
rc_en = yield self.dec2.e.do.rc.rc
+ # don't do Rc=1 for svstep it is handled explicitly.
+ # 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']:
- regnum, is_vec = yield from get_pdecode_cr_out(self.dec2, "CR0")
- self.handle_comparison(results, regnum)
+ 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, FPSCR)
+
+ # check failfirst
+ ffirst_hit = False, False
+ if self.is_svp64_mode:
+ 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?
- if info.write_regs:
- for name, output in zip(output_names, results):
- if name == 'overflow': # ignore, done already (above)
+ 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".
+ if self.FPSCR.FEX and (self.msr[MSRb.FE0] or self.msr[MSRb.FE1]):
+ self.call_trap(0x700, PIb.FP)
+ return
+
+ 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
+ """
+ rm_mode = yield self.dec2.rm_dec.mode
+ ff_inv = yield self.dec2.rm_dec.inv
+ cr_bit = yield self.dec2.rm_dec.cr_sel
+ RC1 = yield self.dec2.rm_dec.RC1
+ vli_ = yield self.dec2.rm_dec.vli # VL inclusive if truncated
+ log(" ff rm_mode", rc_en, rm_mode, SVP64RMMode.FFIRST.value)
+ log(" inv", ff_inv)
+ log(" RC1", RC1)
+ log(" vli", vli_)
+ log(" cr_bit", cr_bit)
+ log(" rc_en", rc_en)
+ if not rc_en or rm_mode != SVP64RMMode.FFIRST.value:
+ return False, False
+ # get the CR vevtor, do BO-test
+ crf = "CR0"
+ log("asmregs", info.asmregs[0], info.write_regs)
+ if 'CR' in info.write_regs and 'BF' in info.asmregs[0]:
+ crf = 'BF'
+ regnum, is_vec = yield from get_cr_out(self.dec2, crf)
+ crtest = self.crl[regnum]
+ ffirst_hit = crtest[cr_bit] != ff_inv
+ log("cr test", crf, regnum, int(crtest), crtest, cr_bit, ff_inv)
+ log("cr test?", ffirst_hit)
+ if not ffirst_hit:
+ return False, False
+ # Fail-first activated, truncate VL
+ vli = SelectableInt(int(vli_), 7)
+ self.svstate.vl = srcstep + vli
+ yield self.dec2.state.svstate.eq(self.svstate.value)
+ yield Settle() # let decoder update
+ return True, vli_
+
+ 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"
+ else:
+ rc_reg = "CR0"
+ regnum, is_vec = yield from get_cr_out(self.dec2, rc_reg)
+ # hang on... for `setvl` actually you want to test SVSTATE.VL
+ is_setvl = ins_name in ('svstep', 'setvl')
+ if is_setvl:
+ result = SelectableInt(result.vl, 64)
+ # else:
+ # overflow = None # do not override overflow except in setvl
+
+ if rc_reg == "CR1":
+ if cr1 is None:
+ 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)
+ self.crl[regnum].eq(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)
+ else:
+ # otherwise we just blat CR0 into the required regnum
+ log("explicit rc0", cr0)
+ self.crl[regnum].eq(cr0)
+
+ 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
- if isinstance(output, int):
- output = SelectableInt(output, 256)
- if name in ['CA', 'CA32']:
- if carry_en:
- log("writing %s to XER" % name, output)
- self.spr['XER'][XER_bits[name]] = output.value
- else:
- log("NOT writing %s to XER" % name, output)
- elif name in info.special_regs:
- log('writing special %s' % name, output, special_sprs)
- if name in special_sprs:
- self.spr[name] = output
- else:
- self.namespace[name].eq(output)
- if name == 'MSR':
- log('msr written', hex(self.msr.value))
- else:
- regnum, is_vec = yield from get_pdecode_idx_out(self.dec2,
- name)
- if regnum is None:
- regnum, is_vec = yield from get_pdecode_idx_out2(
- self.dec2, name)
- if regnum is None:
- # temporary hack for not having 2nd output
- regnum = yield getattr(self.decoder, name)
- is_vec = False
- if self.is_svp64_mode and pred_dst_zero:
- log('zeroing reg %d %s' % (regnum, str(output)),
- is_vec)
- output = SelectableInt(0, 256)
- else:
- if name in fregs:
- ftype = 'fpr'
- else:
- ftype = 'gpr'
- log('writing %s %s %s' % (ftype, regnum, str(output)),
- is_vec)
- if output.bits > 64:
- output = SelectableInt(output.value, 64)
- if name in fregs:
- self.fpr[regnum] = output
- else:
- self.gpr[regnum] = output
+ 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))
+ if nia_update:
+ self.update_pc_next()
- # check if it is the SVSTATE.src/dest step that needs incrementing
- # this is our Sub-Program-Counter loop from 0 to VL-1
- pre = False
- post = False
- nia_update = True
- if self.allow_next_step_inc:
- log("SVSTATE_NEXT: inc requested, mode",
- self.svstate_next_mode, self.allow_next_step_inc)
- yield from self.svstate_pre_inc()
- pre = yield from self.update_new_svstate_steps()
- if pre:
- # reset at end of loop including exit Vertical Mode
- log("SVSTATE_NEXT: end of loop, reset")
- self.svp64_reset_loop()
- self.svstate.vfirst = 0
- self.update_nia()
- if rc_en:
- results = [SelectableInt(0, 64)]
- self.handle_comparison(results) # CR0
+ def check_replace_d(self, info, remap_active):
+ replace_d = False # update / replace constant in pseudocode
+ ldstmode = yield self.dec2.rm_dec.ldstmode
+ vl = self.svstate.vl
+ subvl = yield self.dec2.rm_dec.rm_in.subvl
+ srcstep, dststep = self.new_srcstep, self.new_dststep
+ ssubstep, dsubstep = self.new_ssubstep, self.new_dsubstep
+ if info.form == 'DS':
+ # DS-Form, multiply by 4 then knock 2 bits off after
+ imm = yield self.dec2.dec.fields.FormDS.DS[0:14] * 4
+ else:
+ imm = yield self.dec2.dec.fields.FormD.D[0:16]
+ imm = exts(imm, 16) # sign-extend to integer
+ # get the right step. LD is from srcstep, ST is dststep
+ op = yield self.dec2.e.do.insn_type
+ offsmul = 0
+ if op == MicrOp.OP_LOAD.value:
+ if remap_active:
+ offsmul = yield self.dec2.in1_step
+ log("D-field REMAP src", imm, offsmul, ldstmode)
else:
- if self.allow_next_step_inc == 2:
- log("SVSTATE_NEXT: read")
- nia_update = (yield from self.svstate_post_inc(ins_name))
- else:
- log("SVSTATE_NEXT: post-inc")
- # use actual src/dst-step here to check end, do NOT
- # use bit-reversed version
- srcstep, dststep = self.new_srcstep, self.new_dststep
- remaps = self.get_remap_indices()
- remap_idxs = self.remap_idxs
- vl = self.svstate.vl
- end_src = srcstep == vl-1
- end_dst = dststep == vl-1
- if self.allow_next_step_inc != 2:
- if not end_src:
- self.svstate.srcstep += SelectableInt(1, 7)
- if not end_dst:
- self.svstate.dststep += SelectableInt(1, 7)
- self.namespace['SVSTATE'] = self.svstate.spr
- # set CR0 (if Rc=1) based on end
- if rc_en:
- srcstep = self.svstate.srcstep
- dststep = self.svstate.srcstep
- endtest = 1 if (end_src or end_dst) else 0
- #results = [SelectableInt(endtest, 64)]
- # self.handle_comparison(results) # CR0
-
- # see if svstep was requested, if so, which SVSTATE
- endings = 0b111
- if self.svstate_next_mode > 0:
- shape_idx = self.svstate_next_mode.value-1
- endings = self.remap_loopends[shape_idx]
- cr_field = SelectableInt((~endings) << 1 | endtest, 4)
- print("svstep Rc=1, CR0", cr_field)
- self.crl[0].eq(cr_field) # CR0
- if end_src or end_dst:
- # reset at end of loop including exit Vertical Mode
- log("SVSTATE_NEXT: after increments, reset")
- self.svp64_reset_loop()
- self.svstate.vfirst = 0
-
- elif self.is_svp64_mode:
- nia_update = (yield from self.svstate_post_inc(ins_name))
+ offsmul = (srcstep * (subvl+1)) + ssubstep
+ log("D-field src", imm, offsmul, ldstmode)
+ elif op == MicrOp.OP_STORE.value:
+ # XXX NOTE! no bit-reversed STORE! this should not ever be used
+ offsmul = (dststep * (subvl+1)) + dsubstep
+ log("D-field dst", imm, offsmul, ldstmode)
+ # Unit-Strided LD/ST adds offset*width to immediate
+ if ldstmode == SVP64LDSTmode.UNITSTRIDE.value:
+ ldst_len = yield self.dec2.e.do.data_len
+ imm = SelectableInt(imm + offsmul * ldst_len, 32)
+ replace_d = True
+ # Element-strided multiplies the immediate by element step
+ elif ldstmode == SVP64LDSTmode.ELSTRIDE.value:
+ imm = SelectableInt(imm * offsmul, 32)
+ replace_d = True
+ if replace_d:
+ ldst_ra_vec = yield self.dec2.rm_dec.ldst_ra_vec
+ ldst_imz_in = yield self.dec2.rm_dec.ldst_imz_in
+ log("LDSTmode", SVP64LDSTmode(ldstmode),
+ offsmul, imm, ldst_ra_vec, ldst_imz_in)
+ # new replacement D... errr.. DS
+ if replace_d:
+ if info.form == 'DS':
+ # TODO: assert 2 LSBs are zero?
+ log("DS-Form, TODO, assert 2 LSBs zero?", bin(imm.value))
+ imm.value = imm.value >> 2
+ self.namespace['DS'] = imm
+ else:
+ self.namespace['D'] = imm
+
+ 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)
+ if regnum is None:
+ # doing this is not part of svp64, it's because output
+ # registers, to be modified, need to be in the namespace.
+ regnum, is_vec = yield from get_idx_out(self.dec2, name, True)
+ if regnum is None:
+ regnum, is_vec = yield from get_idx_out2(self.dec2, name, True)
+
+ if isinstance(regnum, tuple):
+ (regnum, base, offs) = regnum
+ else:
+ base, offs = regnum, 0 # temporary HACK
+
+ # in case getting the register number is needed, _RA, _RB
+ # (HACK: only in straight non-svp64-mode for now, or elwidth == 64)
+ 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]
+
+ 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:
+ fval = self.fpr(base, is_vec, offs, ew_src)
+ reg_val = SelectableInt(fval)
+ assert ew_src == 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:
+ 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 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)
+ reg_val = SelectableInt(0, ew_src)
+ return reg_val
+
+ def remap_set_steps(self, remaps):
+ """remap_set_steps sets up the in1/2/3 and out1/2 steps.
+ they work in concert with PowerDecoder2 at the moment,
+ there is no HDL implementation of REMAP. therefore this
+ function, because ISACaller still uses PowerDecoder2,
+ will *explicitly* write the dec2.XX_step values. this has
+ to get sorted out.
+ """
+ # just some convenient debug info
+ for i in range(4):
+ sname = 'SVSHAPE%d' % i
+ shape = self.spr[sname]
+ log(sname, bin(shape.value))
+ log(" lims", shape.lims)
+ log(" mode", shape.mode)
+ log(" skip", shape.skip)
+
+ # set up the list of steps to remap
+ mi0 = self.svstate.mi0
+ mi1 = self.svstate.mi1
+ mi2 = self.svstate.mi2
+ mo0 = self.svstate.mo0
+ mo1 = self.svstate.mo1
+ steps = [[self.dec2.in1_step, mi0], # RA
+ [self.dec2.in2_step, mi1], # RB
+ [self.dec2.in3_step, mi2], # RC
+ [self.dec2.o_step, mo0], # RT
+ [self.dec2.o2_step, mo1], # EA
+ ]
+ if False: # TODO
+ rnames = ['RA', 'RB', 'RC', 'RT', 'RS']
+ for i, reg in enumerate(rnames):
+ idx = yield from get_idx_map(self.dec2, reg)
+ if idx is None:
+ idx = yield from get_idx_map(self.dec2, "F"+reg)
+ if idx == 1: # RA
+ steps[i][0] = self.dec2.in1_step
+ elif idx == 2: # RB
+ steps[i][0] = self.dec2.in2_step
+ elif idx == 3: # RC
+ steps[i][0] = self.dec2.in3_step
+ log("remap step", i, reg, idx, steps[i][1])
+ remap_idxs = self.remap_idxs
+ rremaps = []
+ # now cross-index the required SHAPE for each of 3-in 2-out regs
+ rnames = ['RA', 'RB', 'RC', 'RT', 'EA']
+ for i, (dstep, shape_idx) in enumerate(steps):
+ (shape, remap) = remaps[shape_idx]
+ remap_idx = remap_idxs[shape_idx]
+ # zero is "disabled"
+ if shape.value == 0x0:
+ continue
+ # now set the actual requested step to the current index
+ if dstep is not None:
+ yield dstep.eq(remap_idx)
+
+ # debug printout info
+ rremaps.append((shape.mode, hex(shape.value), dstep,
+ i, rnames[i], shape_idx, remap_idx))
+ for x in rremaps:
+ log("shape remap", x)
+
+ def check_write(self, info, name, output, carry_en, ew_dst):
+ if name == 'overflow': # ignore, done already (above)
+ return
+ if name == 'CR0': # ignore, done already (above)
+ return
+ if isinstance(output, int):
+ output = SelectableInt(output, EFFECTIVELY_UNLIMITED)
+ # write FPSCR
+ if name in ['FPSCR', ]:
+ log("write FPSCR 0x%x" % (output.value))
+ self.FPSCR.eq(output)
+ return
+ # write carry flags
+ if name in ['CA', 'CA32']:
+ if carry_en:
+ log("writing %s to XER" % name, output)
+ log("write XER %s 0x%x" % (name, output.value))
+ self.spr['XER'][XER_bits[name]] = output.value
+ else:
+ log("NOT writing %s to XER" % name, output)
+ return
+ # write special SPRs
+ if name in info.special_regs:
+ log('writing special %s' % name, output, special_sprs)
+ log("write reg %s 0x%x" % (name, output.value),
+ kind=LogType.InstrInOuts)
+ if name in special_sprs:
+ self.spr[name] = output
+ else:
+ self.namespace[name].eq(output)
+ if name == 'MSR':
+ log('msr written', hex(self.msr.value))
+ return
+ # find out1/out2 PR/FPR
+ regnum, is_vec = yield from get_idx_out(self.dec2, name, True)
+ if regnum is None:
+ regnum, is_vec = yield from get_idx_out2(self.dec2, name, True)
+ if regnum is None:
+ # temporary hack for not having 2nd output
+ regnum = yield getattr(self.decoder, name)
+ is_vec = False
+ # convenient debug prefix
+ if name in fregs:
+ reg_prefix = 'f'
+ else:
+ reg_prefix = 'r'
+ # check zeroing due to predicate bit being zero
+ if self.is_svp64_mode and self.pred_dst_zero:
+ log('zeroing reg %s %s' % (str(regnum), str(output)), is_vec)
+ output = SelectableInt(0, EFFECTIVELY_UNLIMITED)
+ log("write reg %s%s 0x%x ew %d" % (reg_prefix, str(regnum),
+ output.value, ew_dst),
+ kind=LogType.InstrInOuts)
+ # zero-extend tov64 bit begore storing (should use EXT oh well)
+ if output.bits > 64:
+ output = SelectableInt(output.value, 64)
+ rnum, base, offset = regnum
+ 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))
+
+ def check_step_increment(self, rc_en, asmop, ins_name):
+ # check if it is the SVSTATE.src/dest step that needs incrementing
+ # this is our Sub-Program-Counter loop from 0 to VL-1
+ if not self.allow_next_step_inc:
+ if self.is_svp64_mode:
+ return (yield from self.svstate_post_inc(ins_name))
+
# XXX only in non-SVP64 mode!
# record state of whether the current operation was an svshape,
+ # OR svindex!
# to be able to know if it should apply in the next instruction.
# also (if going to use this instruction) should disable ability
# to interrupt in between. sigh.
- self.last_op_svshape = asmop == 'svremap'
+ self.last_op_svshape = asmop in ['svremap', 'svindex',
+ 'svshape2']
+ return True
- if nia_update:
- self.update_pc_next()
+ pre = False
+ post = False
+ nia_update = True
+ log("SVSTATE_NEXT: inc requested, mode",
+ self.svstate_next_mode, self.allow_next_step_inc)
+ yield from self.svstate_pre_inc()
+ pre = yield from self.update_new_svstate_steps()
+ if pre:
+ # reset at end of loop including exit Vertical Mode
+ log("SVSTATE_NEXT: end of loop, reset")
+ self.svp64_reset_loop()
+ self.svstate.vfirst = 0
+ self.update_nia()
+ if not rc_en:
+ return True
+ self.handle_comparison(SelectableInt(0, 64)) # CR0
+ return True
+ if self.allow_next_step_inc == 2:
+ log("SVSTATE_NEXT: read")
+ nia_update = (yield from self.svstate_post_inc(ins_name))
+ else:
+ log("SVSTATE_NEXT: post-inc")
+ # use actual (cached) src/dst-step here to check end
+ remaps = self.get_remap_indices()
+ remap_idxs = self.remap_idxs
+ vl = self.svstate.vl
+ subvl = yield self.dec2.rm_dec.rm_in.subvl
+ if self.allow_next_step_inc != 2:
+ yield from self.advance_svstate_steps()
+ #self.namespace['SVSTATE'] = self.svstate.spr
+ # set CR0 (if Rc=1) based on end
+ endtest = 1 if self.at_loopend() else 0
+ if rc_en:
+ #results = [SelectableInt(endtest, 64)]
+ # self.handle_comparison(results) # CR0
+
+ # see if svstep was requested, if so, which SVSTATE
+ endings = 0b111
+ if self.svstate_next_mode > 0:
+ shape_idx = self.svstate_next_mode.value-1
+ endings = self.remap_loopends[shape_idx]
+ cr_field = SelectableInt((~endings) << 1 | endtest, 4)
+ log("svstep Rc=1, CR0", cr_field, endtest)
+ self.crl[0].eq(cr_field) # CR0
+ if endtest:
+ # reset at end of loop including exit Vertical Mode
+ log("SVSTATE_NEXT: after increments, reset")
+ self.svp64_reset_loop()
+ self.svstate.vfirst = 0
+ return nia_update
def SVSTATE_NEXT(self, mode, submode):
"""explicitly moves srcstep/dststep on to next element, for
self.allow_next_step_inc = submode.value + 1
log("SVSTATE_NEXT mode", mode, submode, self.allow_next_step_inc)
self.svstate_next_mode = mode
- if self.svstate_next_mode > 0:
+ if self.svstate_next_mode > 0 and self.svstate_next_mode < 5:
shape_idx = self.svstate_next_mode.value-1
return SelectableInt(self.remap_idxs[shape_idx], 7)
+ if self.svstate_next_mode == 5:
+ self.svstate_next_mode = 0
+ return SelectableInt(self.svstate.srcstep, 7)
+ if self.svstate_next_mode == 6:
+ self.svstate_next_mode = 0
+ return SelectableInt(self.svstate.dststep, 7)
+ if self.svstate_next_mode == 7:
+ self.svstate_next_mode = 0
+ return SelectableInt(self.svstate.ssubstep, 7)
+ if self.svstate_next_mode == 8:
+ self.svstate_next_mode = 0
+ return SelectableInt(self.svstate.dsubstep, 7)
return SelectableInt(0, 7)
- def svstate_pre_inc(self):
- """check if srcstep/dststep need to skip over masked-out predicate bits
- """
- # get SVSTATE VL (oh and print out some debug stuff)
- vl = self.svstate.vl
- srcstep = self.svstate.srcstep
- dststep = self.svstate.dststep
- sv_a_nz = yield self.dec2.sv_a_nz
- fft_mode = yield self.dec2.use_svp64_fft
- in1 = yield self.dec2.e.read_reg1.data
- log("SVP64: VL, srcstep, dststep, sv_a_nz, in1 fft, svp64",
- vl, srcstep, dststep, sv_a_nz, in1, fft_mode,
- self.is_svp64_mode)
-
- # get predicate mask (all 64 bits)
- srcmask = dstmask = 0xffff_ffff_ffff_ffff
-
- pmode = yield self.dec2.rm_dec.predmode
- reverse_gear = yield self.dec2.rm_dec.reverse_gear
- sv_ptype = yield self.dec2.dec.op.SV_Ptype
- srcpred = yield self.dec2.rm_dec.srcpred
- dstpred = yield self.dec2.rm_dec.dstpred
- pred_src_zero = yield self.dec2.rm_dec.pred_sz
- pred_dst_zero = yield self.dec2.rm_dec.pred_dz
- if pmode == SVP64PredMode.INT.value:
- srcmask = dstmask = get_predint(self.gpr, dstpred)
- if sv_ptype == SVPtype.P2.value:
- srcmask = get_predint(self.gpr, srcpred)
- elif pmode == SVP64PredMode.CR.value:
- srcmask = dstmask = get_predcr(self.crl, dstpred, vl)
- if sv_ptype == SVPtype.P2.value:
- srcmask = get_predcr(self.crl, srcpred, vl)
- log(" pmode", pmode)
- log(" reverse", reverse_gear)
- log(" ptype", sv_ptype)
- log(" srcpred", bin(srcpred))
- log(" dstpred", bin(dstpred))
- log(" srcmask", bin(srcmask))
- log(" dstmask", bin(dstmask))
- log(" pred_sz", bin(pred_src_zero))
- log(" pred_dz", bin(pred_dst_zero))
-
- # okaaay, so here we simply advance srcstep (TODO dststep)
- # until the predicate mask has a "1" bit... or we run out of VL
- # let srcstep==VL be the indicator to move to next instruction
- if not pred_src_zero:
- while (((1 << srcstep) & srcmask) == 0) and (srcstep != vl):
- log(" skip", bin(1 << srcstep))
- srcstep += 1
- # same for dststep
- if not pred_dst_zero:
- while (((1 << dststep) & dstmask) == 0) and (dststep != vl):
- log(" skip", bin(1 << dststep))
- dststep += 1
-
- # now work out if the relevant mask bits require zeroing
- if pred_dst_zero:
- pred_dst_zero = ((1 << dststep) & dstmask) == 0
- if pred_src_zero:
- pred_src_zero = ((1 << srcstep) & srcmask) == 0
-
- # store new srcstep / dststep
- self.new_srcstep, self.new_dststep = srcstep, dststep
- self.pred_dst_zero, self.pred_src_zero = pred_dst_zero, pred_src_zero
- log(" new srcstep", srcstep)
- log(" new dststep", dststep)
-
def get_src_dststeps(self):
- """gets srcstep and dststep
+ """gets srcstep, dststep, and ssubstep, dsubstep
"""
- return self.new_srcstep, self.new_dststep
-
- def update_new_svstate_steps(self):
- # note, do not get the bit-reversed srcstep here!
- srcstep, dststep = self.new_srcstep, self.new_dststep
-
- # update SVSTATE with new srcstep
- self.svstate.srcstep = srcstep
- self.svstate.dststep = dststep
+ return (self.new_srcstep, self.new_dststep,
+ self.new_ssubstep, self.new_dsubstep)
+
+ def update_svstate_namespace(self, overwrite_svstate=True):
+ if overwrite_svstate:
+ # note, do not get the bit-reversed srcstep here!
+ srcstep, dststep = self.new_srcstep, self.new_dststep
+ ssubstep, dsubstep = self.new_ssubstep, self.new_dsubstep
+
+ # update SVSTATE with new srcstep
+ self.svstate.srcstep = srcstep
+ self.svstate.dststep = dststep
+ self.svstate.ssubstep = ssubstep
+ self.svstate.dsubstep = dsubstep
self.namespace['SVSTATE'] = self.svstate
yield self.dec2.state.svstate.eq(self.svstate.value)
yield Settle() # let decoder update
+
+ def update_new_svstate_steps(self, overwrite_svstate=True):
+ yield from self.update_svstate_namespace(overwrite_svstate)
srcstep = self.svstate.srcstep
dststep = self.svstate.dststep
+ ssubstep = self.svstate.ssubstep
+ dsubstep = self.svstate.dsubstep
+ pack = self.svstate.pack
+ unpack = self.svstate.unpack
vl = self.svstate.vl
+ sv_mode = yield self.dec2.rm_dec.sv_mode
+ subvl = yield self.dec2.rm_dec.rm_in.subvl
+ rm_mode = yield self.dec2.rm_dec.mode
+ ff_inv = yield self.dec2.rm_dec.inv
+ cr_bit = yield self.dec2.rm_dec.cr_sel
log(" srcstep", srcstep)
log(" dststep", dststep)
+ log(" pack", pack)
+ log(" unpack", unpack)
+ log(" ssubstep", ssubstep)
+ log(" dsubstep", dsubstep)
log(" vl", vl)
+ log(" subvl", subvl)
+ log(" rm_mode", rm_mode)
+ log(" sv_mode", sv_mode)
+ log(" inv", ff_inv)
+ log(" cr_bit", cr_bit)
# check if end reached (we let srcstep overrun, above)
# nothing needs doing (TODO zeroing): just do next instruction
- return srcstep == vl or dststep == vl
+ if self.loopend:
+ return True
+ return ((ssubstep == subvl and srcstep == vl) or
+ (dsubstep == subvl and dststep == vl))
def svstate_post_inc(self, insn_name, vf=0):
# check if SV "Vertical First" mode is enabled
vfirst = self.svstate.vfirst
log(" SV Vertical First", vf, vfirst)
if not vf and vfirst == 1:
+ if insn_name.startswith("sv.bc"):
+ self.update_pc_next()
+ return False
self.update_nia()
return True
# this is our Sub-Program-Counter loop from 0 to VL-1
# XXX twin predication TODO
vl = self.svstate.vl
+ subvl = yield self.dec2.rm_dec.rm_in.subvl
mvl = self.svstate.maxvl
srcstep = self.svstate.srcstep
dststep = self.svstate.dststep
+ ssubstep = self.svstate.ssubstep
+ dsubstep = self.svstate.dsubstep
+ pack = self.svstate.pack
+ unpack = self.svstate.unpack
rm_mode = yield self.dec2.rm_dec.mode
reverse_gear = yield self.dec2.rm_dec.reverse_gear
sv_ptype = yield self.dec2.dec.op.SV_Ptype
in_vec = not (yield self.dec2.no_in_vec)
log(" svstate.vl", vl)
log(" svstate.mvl", mvl)
+ log(" rm.subvl", subvl)
log(" svstate.srcstep", srcstep)
log(" svstate.dststep", dststep)
+ log(" svstate.ssubstep", ssubstep)
+ log(" svstate.dsubstep", dsubstep)
+ log(" svstate.pack", pack)
+ log(" svstate.unpack", unpack)
log(" mode", rm_mode)
log(" reverse", reverse_gear)
log(" out_vec", out_vec)
log(" in_vec", in_vec)
- log(" sv_ptype", sv_ptype, sv_ptype == SVPtype.P2.value)
- # check if srcstep needs incrementing by one, stop PC advancing
- # svp64 loop can end early if the dest is scalar for single-pred
- # but for 2-pred both src/dest have to be checked.
- # XXX this might not be true! it may just be LD/ST
- if sv_ptype == SVPtype.P2.value:
- svp64_is_vector = (out_vec or in_vec)
- else:
- svp64_is_vector = out_vec
+ log(" sv_ptype", sv_ptype, sv_ptype == SVPType.P2.value)
# 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']
self.svp64_reset_loop()
self.update_pc_next()
return False
- if svp64_is_vector and srcstep != vl-1 and dststep != vl-1:
- self.svstate.srcstep += SelectableInt(1, 7)
- self.svstate.dststep += SelectableInt(1, 7)
- self.namespace['SVSTATE'] = self.svstate
- # 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.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
-
- # reset loop to zero and update NIA
- self.svp64_reset_loop()
- self.update_nia()
-
- return True
+ # check if srcstep needs incrementing by one, stop PC advancing
+ # but for 2-pred both src/dest have to be checked.
+ # XXX this might not be true! it may just be LD/ST
+ if sv_ptype == SVPType.P2.value:
+ svp64_is_vector = (out_vec or in_vec)
+ else:
+ svp64_is_vector = out_vec
+ # loops end at the first "hit" (source or dest)
+ yield from self.advance_svstate_steps()
+ loopend = self.loopend
+ log("loopend", svp64_is_vector, loopend)
+ if not svp64_is_vector or loopend:
+ # reset loop to zero and update NIA
+ self.svp64_reset_loop()
+ self.update_nia()
+
+ return True
+
+ # still looping, advance and update NIA
+ self.namespace['SVSTATE'] = self.svstate
+
+ # 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.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
def update_pc_next(self):
# UPDATE program counter
self.pc.update(self.namespace, self.is_svp64_mode)
- self.svstate.spr = self.namespace['SVSTATE']
+ #self.svstate.spr = self.namespace['SVSTATE']
log("end of call", self.namespace['CIA'],
self.namespace['NIA'],
self.namespace['SVSTATE'])
def svp64_reset_loop(self):
self.svstate.srcstep = 0
self.svstate.dststep = 0
+ self.svstate.ssubstep = 0
+ self.svstate.dsubstep = 0
+ self.loopend = False
log(" svstate.srcstep loop end (PC to update)")
self.namespace['SVSTATE'] = self.svstate