from openpower.decoder.power_enums import SVPtype
-from openpower.decoder.helpers import exts, gtu, ltu, undefined
+from openpower.decoder.helpers import (exts, gtu, ltu, undefined, bitrev)
from openpower.consts import PIb, MSRb # big-endian (PowerISA versions)
from openpower.consts import SVP64CROffs
from openpower.decoder.power_svp64 import SVP64RM, decode_extra
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.util import log
"TAR": 0,
"MSR": 0,
"SVSTATE": 0,
+ "SVSHAPE0": 0,
+ "SVSHAPE1": 0,
+ "SVSHAPE2": 0,
+ "SVSHAPE3": 0,
"CA": 0,
"CA32": 0,
self[i] = SelectableInt(regfile[i], 64)
def __call__(self, ridx):
+ if isinstance(ridx, SelectableInt):
+ ridx = ridx.value
return self[ridx]
def set_form(self, form):
self.form = form
+ def __setitem__(self, rnum, value):
+ # rnum = rnum.value # only SelectableInt allowed
+ log("GPR setitem", rnum, value)
+ if isinstance(rnum, SelectableInt):
+ rnum = rnum.value
+ dict.__setitem__(self, rnum, value)
+
def getz(self, rnum):
# rnum = rnum.value # only SelectableInt allowed
log("GPR getzero?", rnum)
namespace['NIA'] = self.NIA
-# Simple-V: see https://libre-soc.org/openpower/sv
-class SVP64State:
- def __init__(self, init=0):
- self.spr = SelectableInt(init, 32)
- # fields of SVSTATE, see https://libre-soc.org/openpower/sv/sprs/
- self.maxvl = FieldSelectableInt(self.spr, tuple(range(0,7)))
- self.vl = FieldSelectableInt(self.spr, tuple(range(7,14)))
- self.srcstep = FieldSelectableInt(self.spr, tuple(range(14,21)))
- self.dststep = FieldSelectableInt(self.spr, tuple(range(21,28)))
- self.subvl = FieldSelectableInt(self.spr, tuple(range(28,30)))
- self.svstep = FieldSelectableInt(self.spr, tuple(range(30,32)))
-
-
# SVP64 ReMap field
class SVP64RMFields:
def __init__(self, init=0):
return mask
+# TODO, really should just be using PowerDecoder2
def get_pdecode_idx_in(dec2, name):
op = dec2.dec.op
in1_sel = yield op.in1_sel
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
return None, False
+# TODO, really should just be using PowerDecoder2
def get_pdecode_cr_out(dec2, name):
op = dec2.dec.op
out_sel = yield op.cr_out
return None, False
+# TODO, really should just be using PowerDecoder2
def get_pdecode_idx_out(dec2, name):
op = dec2.dec.op
out_sel = yield op.out_sel
return out, o_isvec
elif name == 'RT':
log ("get_pdecode_idx_out", out_sel, OutSel.RT.value,
- OutSel.RT_OR_ZERO.value, out, o_isvec)
+ OutSel.RT_OR_ZERO.value, out, o_isvec,
+ dec2.dec.RT)
if out_sel == OutSel.RT.value:
return out, o_isvec
+ 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
elif name == 'FRA':
log ("get_pdecode_idx_out", out_sel, OutSel.FRA.value, out, o_isvec)
if out_sel == OutSel.FRA.value:
return None, False
+# TODO, really should just be using PowerDecoder2
def get_pdecode_idx_out2(dec2, name):
# check first if register is activated for write
- out_ok = yield dec2.e.write_ea.ok
- if not out_ok:
- return None, False
-
op = dec2.dec.op
out_sel = yield op.out_sel
out = yield dec2.e.write_ea.data
o_isvec = yield dec2.o2_isvec
- log ("get_pdecode_idx_out2", name, out_sel, out, o_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
+
if name == 'RA':
if hasattr(op, "upd"):
# update mode LD/ST uses read-reg A also as an output
out, o_isvec)
if upd == LDSTMode.update.value:
return out, o_isvec
+ 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:
+ if fft_en:
+ log ("get_pdecode_idx_out2", out_sel, OutSel.FRS.value,
+ out, o_isvec)
+ return out, o_isvec
return None, False
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
+
+ # 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)
+ 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)
'memassign': self.memassign,
'NIA': self.pc.NIA,
'CIA': self.pc.CIA,
- 'SVSTATE': self.svstate.spr,
+ 'SVSTATE': self.svstate,
+ 'SVSHAPE0': self.spr['SVSHAPE0'],
+ 'SVSHAPE1': self.spr['SVSHAPE1'],
+ 'SVSHAPE2': self.spr['SVSHAPE2'],
+ 'SVSHAPE3': self.spr['SVSHAPE3'],
'CR': self.cr,
'MSR': self.msr,
'undefined': undefined,
# then "yield" fields only from op_fields rather than hard-coded
# list, here.
fields = self.decoder.sigforms[formname]
+ log("prep_namespace", formname, op_fields)
for name in op_fields:
if name == 'spr':
sig = getattr(fields, name.upper())
self.namespace['CA32'] = self.spr['XER'][XER_bits['CA32']].value
# add some SVSTATE convenience variables
- vl = self.svstate.vl.asint(msb0=True)
- srcstep = self.svstate.srcstep.asint(msb0=True)
+ vl = self.svstate.vl
+ srcstep = self.svstate.srcstep
self.namespace['VL'] = vl
self.namespace['srcstep'] = srcstep
SO = self.spr['XER'][XER_bits['SO']]
log("handle_comparison SO", SO)
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)
def set_pc(self, pc_val):
yield self.dec2.state.msr.eq(self.msr.value)
yield self.dec2.state.pc.eq(pc)
if self.svstate is not None:
- yield self.dec2.state.svstate.eq(self.svstate.spr.value)
+ yield self.dec2.state.svstate.eq(self.svstate.value)
# SVP64. first, check if the opcode is EXT001, and SVP64 id bits set
yield Settle()
pfx.insn[7].value == 0b1 and
pfx.insn[9].value == 0b1)
self.pc.update_nia(self.is_svp64_mode)
+ yield self.dec2.is_svp64_mode.eq(self.is_svp64_mode) # set SVP64 decode
self.namespace['NIA'] = self.pc.NIA
- self.namespace['SVSTATE'] = self.svstate.spr
+ self.namespace['SVSTATE'] = self.svstate
if not 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 (" svstate.vl", self.svstate.vl.asint(msb0=True))
- log (" svstate.mvl", self.svstate.maxvl.asint(msb0=True))
+ 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 Settle()
+ # store this for use in get_src_dststeps()
+ self.ldstmode = yield self.dec2.rm_dec.ldstmode
def execute_one(self):
"""execute one instruction
asmop = 'mtcrf'
return asmop
+ 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()
+ # get four SVSHAPEs. here we are hard-coding
+ SVSHAPE0 = self.spr['SVSHAPE0']
+ SVSHAPE1 = self.spr['SVSHAPE1']
+ SVSHAPE2 = self.spr['SVSHAPE2']
+ SVSHAPE3 = self.spr['SVSHAPE3']
+ # set up the iterators
+ remaps = [(SVSHAPE0, SVSHAPE0.get_iterator()),
+ (SVSHAPE1, SVSHAPE1.get_iterator()),
+ (SVSHAPE2, SVSHAPE2.get_iterator()),
+ (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"
+ if shape.value == 0x0:
+ self.remap_idxs[i] = 0
+ # pick src or dststep depending on reg num (0-2=in, 3-4=out)
+ step = dststep if (i in [3, 4]) else srcstep
+ # this is terrible. O(N^2) looking for the match. but hey.
+ for idx, (remap_idx, loopends) in enumerate(remap):
+ if idx == step:
+ break
+ self.remap_idxs[i] = remap_idx
+ self.remap_loopends[i] = loopends
+ dbg.append((i, step, remap_idx, loopends))
+ for (i, step, remap_idx, loopends) in dbg:
+ log ("SVSHAPE %d idx, end" % i, step, remap_idx, bin(loopends))
+ return remaps
+
def get_spr_msb(self):
dec_insn = yield self.dec2.e.do.insn
return dec_insn & (1 << 20) != 0 # sigh - XFF.spr[-1]?
self.last_st_addr = None # reset the last known store address
self.last_ld_addr = None # etc.
- name = name.strip() # remove spaces if not already done so
+ ins_name = name.strip() # remove spaces if not already done so
if self.halted:
- log("halted - not executing", name)
+ log("halted - not executing", ins_name)
return
# 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", name, asmop)
+ log("call", ins_name, asmop)
# check privileged
int_op = yield self.dec2.dec.op.internal_op
return
# check halted condition
- if name == 'attn':
+ if ins_name == 'attn':
self.halted = True
return
# check illegal instruction
illegal = False
- if name not in ['mtcrf', 'mtocrf']:
- illegal = name != asmop
+ 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
- name = 'setvl'
+ 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'
if illegal:
- print("illegal", name, asmop)
+ print("illegal", ins_name, asmop)
self.call_trap(0x700, PIb.ILLEG)
print("name %s != %s - calling ILLEGAL trap, PC: %x" %
- (name, asmop, self.pc.CIA.value))
+ (ins_name, asmop, self.pc.CIA.value))
return
+ # this is for setvl "Vertical" mode: if set true,
+ # srcstep/dststep is explicitly advanced. mode says which SVSTATE to
+ # test for Rc=1 end condition. 3 bits of all 3 loops are put into CR0
+ self.allow_next_step_inc = False
+ self.svstate_next_mode = 0
+
# nop has to be supported, we could let the actual op calculate
# but PowerDecoder has a pattern for nop
- if name is 'nop':
+ if ins_name is 'nop':
self.update_pc_next()
return
- info = self.instrs[name]
+ info = self.instrs[ins_name]
yield from self.prep_namespace(info.form, info.op_fields)
# preserve order of register names
input_names = create_args(list(info.read_regs) +
list(info.uninit_regs))
- log(input_names)
+ log("input names", input_names)
# get SVP64 entry for the current instruction
- sv_rm = self.svp64rm.instrs.get(name)
+ sv_rm = self.svp64rm.instrs.get(ins_name)
if sv_rm is not None:
dest_cr, src_cr, src_byname, dest_byname = decode_extra(sv_rm)
else:
dest_cr, src_cr, src_byname, dest_byname = False, False, {}, {}
log ("sv rm", sv_rm, dest_cr, src_cr, src_byname, dest_byname)
- # get SVSTATE VL (oh and print out some debug stuff)
- if self.is_svp64_mode:
- vl = self.svstate.vl.asint(msb0=True)
- srcstep = self.svstate.srcstep.asint(msb0=True)
- dststep = self.svstate.dststep.asint(msb0=True)
- sv_a_nz = yield self.dec2.sv_a_nz
- in1 = yield self.dec2.e.read_reg1.data
- log ("SVP64: VL, srcstep, dststep, sv_a_nz, in1",
- vl, srcstep, dststep, sv_a_nz, in1)
-
- # get predicate mask
- srcmask = dstmask = 0xffff_ffff_ffff_ffff
+ # see if srcstep/dststep need skipping over masked-out predicate bits
+ if (self.is_svp64_mode or ins_name == 'setvl' or
+ ins_name.startswith("sv")):
+ yield from self.svstate_pre_inc()
if self.is_svp64_mode:
- 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
-
- # update SVSTATE with new srcstep
- self.svstate.srcstep[0:7] = srcstep
- self.svstate.dststep[0:7] = dststep
- self.namespace['SVSTATE'] = self.svstate.spr
- yield self.dec2.state.svstate.eq(self.svstate.spr.value)
- yield Settle() # let decoder update
- srcstep = self.svstate.srcstep.asint(msb0=True)
- dststep = self.svstate.dststep.asint(msb0=True)
- log (" srcstep", srcstep)
- log (" dststep", dststep)
-
- # check if end reached (we let srcstep overrun, above)
- # nothing needs doing (TODO zeroing): just do next instruction
- if srcstep == vl or dststep == vl:
+ pre = yield from self.update_new_svstate_steps()
+ if pre:
self.svp64_reset_loop()
+ self.update_nia()
self.update_pc_next()
return
+ srcstep, dststep = self.get_src_dststeps()
+ pred_dst_zero = self.pred_dst_zero
+ pred_src_zero = self.pred_src_zero
+ vl = self.svstate.vl
# VL=0 in SVP64 mode means "do nothing: skip instruction"
if self.is_svp64_mode and vl == 0:
self.namespace['NIA'])
return
+ # for when SVREMAP is active, using pre-arranged schedule.
+ # note: modifying PowerDecoder2 needs to "settle"
+ remap_en = self.svstate.SVme
+ persist = self.svstate.RMpst
+ active = (persist or self.last_op_svshape) and remap_en != 0
+ yield self.dec2.remap_active.eq(remap_en if active else 0)
+ yield Settle()
+ 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)
+ # after that, settle down (combinatorial) to let Vector reg numbers
+ # work themselves out
+ yield Settle()
+ remap_active = yield self.dec2.remap_active
+ log ("remap active", bin(remap_active))
+
# main input registers (RT, RA ...)
inputs = []
for name in input_names:
log('reading reg %s %s' % (name, str(regnum)), is_vec)
if name in fregs:
reg_val = self.fpr(regnum)
- else:
+ elif name is not None:
reg_val = self.gpr(regnum)
else:
log('zero input reg %s %s' % (name, str(regnum)), is_vec)
reg_val = 0
inputs.append(reg_val)
+ # arrrrgh, awful hack, to get _RT into namespace
+ if ins_name in ['setvl', 'svstep']:
+ regname = "_RT"
+ RT = yield self.dec2.dec.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")
+ 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:
- D = yield self.dec2.dec.fields.FormD.D[0:16]
- D = exts(D, 16) # sign-extend to integer
ldstmode = yield self.dec2.rm_dec.ldstmode
+ # bitreverse 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.BITREVERSE.value:
+ imm = yield self.dec2.dec.fields.FormSVD.SVD[0:11]
+ imm = exts(imm, 11) # sign-extend to integer
+ log ("bitrev 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:
- offsmul = srcstep
- log("D-field src", D, offsmul)
+ 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", D, offsmul)
+ log("D-field dst", imm, offsmul)
+ # bit-reverse mode, rev already done through get_src_dst_steps()
+ if ldstmode == SVP64LDSTmode.BITREVERSE.value:
+ # manually look up RC, sigh
+ RC = yield self.dec2.dec.RC[0:5]
+ RC = self.gpr(RC)
+ log ("LD-BITREVERSE:", "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
- if ldstmode == SVP64LDSTmode.UNITSTRIDE.value:
+ elif ldstmode == SVP64LDSTmode.UNITSTRIDE.value:
ldst_len = yield self.dec2.e.do.data_len
- D = SelectableInt(D + offsmul * ldst_len, 32)
+ imm = SelectableInt(imm + offsmul * ldst_len, 32)
replace_d = True
# Element-strided multiplies the immediate by element step
elif ldstmode == SVP64LDSTmode.ELSTRIDE.value:
- D = SelectableInt(D * offsmul, 32)
+ imm = SelectableInt(imm * offsmul, 32)
replace_d = True
- ldst_ra_vec = yield self.dec2.rm_dec.ldst_ra_vec
- ldst_imz_in = yield self.dec2.rm_dec.ldst_imz_in
- log("LDSTmode", ldstmode, offsmul, D, ldst_ra_vec, ldst_imz_in)
- # new replacement D
+ 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:
- self.namespace['D'] = 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
# "special" registers
for special in info.special_regs:
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
- if rc_en:
+ 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)
ftype = 'fpr'
else:
ftype = 'gpr'
- log('writing %s %s %s' % (regnum, ftype, str(output)),
+ log('writing %s %s %s' % (ftype, regnum, str(output)),
is_vec)
if output.bits > 64:
output = SelectableInt(output.value, 64)
# 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 self.is_svp64_mode:
- # XXX twin predication TODO
- vl = self.svstate.vl.asint(msb0=True)
- mvl = self.svstate.maxvl.asint(msb0=True)
- srcstep = self.svstate.srcstep.asint(msb0=True)
- dststep = self.svstate.dststep.asint(msb0=True)
- 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
- out_vec = not (yield self.dec2.no_out_vec)
- in_vec = not (yield self.dec2.no_in_vec)
- log (" svstate.vl", vl)
- log (" svstate.mvl", mvl)
- log (" svstate.srcstep", srcstep)
- log (" svstate.dststep", dststep)
- 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)
+ pre = False
+ post = False
+ 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
else:
- svp64_is_vector = out_vec
- 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.pc.NIA.value = self.pc.CIA.value
- self.namespace['NIA'] = self.pc.NIA
+ if self.allow_next_step_inc == 2:
+ log ("SVSTATE_NEXT: read")
+ yield from self.svstate_post_inc()
+ 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
- log("end of sub-pc call", self.namespace['CIA'],
- self.namespace['NIA'])
- return # DO NOT allow PC to update whilst Sub-PC loop running
- # reset loop to zero
- self.svp64_reset_loop()
+ # 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:
+ yield from self.svstate_post_inc()
+ else:
+ # XXX only in non-SVP64 mode!
+ # record state of whether the current operation was an svshape,
+ # 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.update_pc_next()
+ def SVSTATE_NEXT(self, mode, submode):
+ """explicitly moves srcstep/dststep on to next element, for
+ "Vertical-First" mode. this function is called from
+ setvl pseudo-code, as a pseudo-op "svstep"
+
+ WARNING: this function uses information that was created EARLIER
+ due to it being in the middle of a yield, but this function is
+ *NOT* called from yield (it's called from compiled pseudocode).
+ """
+ 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:
+ shape_idx = self.svstate_next_mode.value-1
+ return SelectableInt(self.remap_idxs[shape_idx], 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
+ 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 but performs bit-reversal on srcstep if
+ required. use this ONLY to perform calculations, do NOT update
+ SVSTATE with the bit-reversed value of srcstep
+
+ ARGH, had to store self.ldstmode and VL due to yield issues
+ """
+ srcstep, dststep = self.new_srcstep, self.new_dststep
+ if self.is_svp64_mode:
+ if self.ldstmode == SVP64LDSTmode.BITREVERSE.value:
+ vl = self.svstate.vl
+ log ("SRCSTEP-BITREVERSE:", "VL", vl, "srcstep", srcstep,
+ "rev", bin(bitrev(srcstep, vl)))
+ srcstep = bitrev(srcstep, vl)
+
+ return (srcstep, 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
+ self.namespace['SVSTATE'] = self.svstate
+ yield self.dec2.state.svstate.eq(self.svstate.value)
+ yield Settle() # let decoder update
+ srcstep = self.svstate.srcstep
+ dststep = self.svstate.dststep
+ vl = self.svstate.vl
+ log (" srcstep", srcstep)
+ log (" dststep", dststep)
+ log (" vl", vl)
+
+ # check if end reached (we let srcstep overrun, above)
+ # nothing needs doing (TODO zeroing): just do next instruction
+ return srcstep == vl or dststep == vl
+
+ def svstate_post_inc(self, 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:
+ self.update_nia()
+ return True
+
+ # check if it is the SVSTATE.src/dest step that needs incrementing
+ # this is our Sub-Program-Counter loop from 0 to VL-1
+ # XXX twin predication TODO
+ vl = self.svstate.vl
+ mvl = self.svstate.maxvl
+ srcstep = self.svstate.srcstep
+ dststep = self.svstate.dststep
+ 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
+ out_vec = not (yield self.dec2.no_out_vec)
+ in_vec = not (yield self.dec2.no_in_vec)
+ log (" svstate.vl", vl)
+ log (" svstate.mvl", mvl)
+ log (" svstate.srcstep", srcstep)
+ log (" svstate.dststep", dststep)
+ 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
+ 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.pc.NIA.value = self.pc.CIA.value
+ self.namespace['NIA'] = self.pc.NIA
+ self.namespace['SVSTATE'] = self.svstate
+ 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
+
def update_pc_next(self):
# UPDATE program counter
self.pc.update(self.namespace, self.is_svp64_mode)
self.namespace['SVSTATE'])
def svp64_reset_loop(self):
- self.svstate.srcstep[0:7] = 0
- self.svstate.dststep[0:7] = 0
+ self.svstate.srcstep = 0
+ self.svstate.dststep = 0
log (" svstate.srcstep loop end (PC to update)")
+ self.namespace['SVSTATE'] = self.svstate
+
+ def update_nia(self):
self.pc.update_nia(self.is_svp64_mode)
self.namespace['NIA'] = self.pc.NIA
- self.namespace['SVSTATE'] = self.svstate.spr
-
def inject():
"""Decorator factory.
context = args[0].namespace # variables to be injected
saved_values = func_globals.copy() # Shallow copy of dict.
+ log("globals before", context.keys())
func_globals.update(context)
result = func(*args, **kwargs)
log("globals after", func_globals['CIA'], func_globals['NIA'])