create_pdecode_svp64_ldst,
PowerOp)
from openpower.decoder.power_enums import (MicrOp, CryIn, Function,
- CRInSel, CROutSel,
- LdstLen, In1Sel, In2Sel, In3Sel,
- OutSel, SPRfull, SPRreduced,
- RC, SVP64LDSTmode, LDSTMode,
- SVEXTRA, SVEtype, SVPtype)
+ CRInSel, CROutSel,
+ LdstLen, In1Sel, In2Sel, In3Sel,
+ OutSel, SPRfull, SPRreduced,
+ RCOE, SVP64LDSTmode, LDSTMode,
+ SVEXTRA, SVEtype, SVPtype)
from openpower.decoder.decode2execute1 import (Decode2ToExecute1Type, Data,
Decode2ToOperand)
FastRegsEnum, XERRegsEnum, TT)
from openpower.state import CoreState
-from openpower.util import (spr_to_fast, log)
+from openpower.util import (spr_to_fast, spr_to_state, log)
def decode_spr_num(spr):
with m.Case(MicrOp.OP_ATTN, MicrOp.OP_MFMSR, MicrOp.OP_MTMSRD,
MicrOp.OP_MTMSR, MicrOp.OP_RFID):
comb += is_priv_insn.eq(1)
- with m.Case(MicrOp.OP_TLBIE) : comb += is_priv_insn.eq(1)
+ with m.Case(MicrOp.OP_TLBIE):
+ comb += is_priv_insn.eq(1)
with m.Case(MicrOp.OP_MFSPR, MicrOp.OP_MTSPR):
with m.If(insn[20]): # field XFX.spr[-1] i think
comb += is_priv_insn.eq(1)
self.spr_i = Signal(10, reset_less=True)
self.spr_o = Data(SPR, name="spr_o")
- self.fast_o = Data(3, name="fast_o")
+ self.fast_o = Data(4, name="fast_o")
+ self.state_o = Data(3, name="state_o")
def elaborate(self, platform):
m = Module()
with m.Case(x.value):
m.d.comb += self.fast_o.data.eq(v)
m.d.comb += self.fast_o.ok.eq(1)
+ for x, v in spr_to_state.items():
+ with m.Case(x.value):
+ m.d.comb += self.state_o.data.eq(v)
+ m.d.comb += self.state_o.ok.eq(1)
return m
self.insn_in = Signal(32, reset_less=True)
self.reg_out = Data(5, name="reg_a")
self.spr_out = Data(SPR, "spr_a")
- self.fast_out = Data(3, "fast_a")
+ self.fast_out = Data(4, "fast_a")
+ self.state_out = Data(3, "state_a")
self.sv_nz = Signal(1)
def elaborate(self, platform):
comb += sprmap.spr_i.eq(spr)
comb += self.spr_out.eq(sprmap.spr_o)
comb += self.fast_out.eq(sprmap.fast_o)
+ comb += self.state_out.eq(sprmap.state_o)
return m
self.dec = dec
self.sel_in = Signal(In1Sel, reset_less=True)
self.immz_out = Signal(reset_less=True)
- self.sv_nz = Signal(1) # EXTRA bits from SVP64
+ self.sv_nz = Signal(1) # EXTRA bits from SVP64
def elaborate(self, platform):
m = Module()
ra = Signal(5, reset_less=True)
comb += ra.eq(self.dec.RA)
with m.If((self.sel_in == In1Sel.RA_OR_ZERO) &
- (ra == Const(0, 5)) &
- (self.sv_nz == Const(0, 1))):
+ (ra == Const(0, 5)) &
+ (self.sv_nz == Const(0, 1))):
comb += self.immz_out.eq(1)
return m
self.sel_in = Signal(In2Sel, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
self.reg_out = Data(7, "reg_b")
- self.reg_isvec = Signal(1, name="reg_b_isvec") # TODO: in reg_out
- self.fast_out = Data(3, "fast_b")
+ self.reg_isvec = Signal(1, name="reg_b_isvec") # TODO: in reg_out
+ self.fast_out = Data(4, "fast_b")
def elaborate(self, platform):
m = Module()
class DecodeBImm(Elaboratable):
"""DecodeB immediate from instruction
"""
+
def __init__(self, dec):
self.dec = dec
self.sel_in = Signal(In2Sel, reset_less=True)
# select Register B Immediate
with m.Switch(self.sel_in):
- with m.Case(In2Sel.CONST_UI): # unsigned
+ with m.Case(In2Sel.CONST_UI): # unsigned
comb += self.imm_out.data.eq(self.dec.UI)
comb += self.imm_out.ok.eq(1)
with m.Case(In2Sel.CONST_SI): # sign-extended 16-bit
comb += si_hi.eq(self.dec.SI << 16)
comb += self.imm_out.data.eq(exts(si_hi, 32, 64))
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_UI_HI): # unsigned
+ with m.Case(In2Sel.CONST_UI_HI): # unsigned
ui = Signal(16, reset_less=True)
comb += ui.eq(self.dec.UI)
comb += self.imm_out.data.eq(ui << 16)
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_LI): # sign-extend 24+2=26 bit
+ with m.Case(In2Sel.CONST_LI): # sign-extend 24+2=26 bit
li = Signal(26, reset_less=True)
comb += li.eq(self.dec.LI << 2)
comb += self.imm_out.data.eq(exts(li, 26, 64))
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_BD): # sign-extend (14+2)=16 bit
+ with m.Case(In2Sel.CONST_BD): # sign-extend (14+2)=16 bit
bd = Signal(16, reset_less=True)
comb += bd.eq(self.dec.BD << 2)
comb += self.imm_out.data.eq(exts(bd, 16, 64))
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_DS): # sign-extended (14+2=16) bit
+ with m.Case(In2Sel.CONST_DS): # sign-extended (14+2=16) bit
ds = Signal(16, reset_less=True)
comb += ds.eq(self.dec.DS << 2)
comb += self.imm_out.data.eq(exts(ds, 16, 64))
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_M1): # signed (-1)
+ with m.Case(In2Sel.CONST_M1): # signed (-1)
comb += self.imm_out.data.eq(~Const(0, 64)) # all 1s
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_SH): # unsigned - for shift
+ with m.Case(In2Sel.CONST_SH): # unsigned - for shift
comb += self.imm_out.data.eq(self.dec.sh)
comb += self.imm_out.ok.eq(1)
- with m.Case(In2Sel.CONST_SH32): # unsigned - for shift
+ with m.Case(In2Sel.CONST_SH32): # unsigned - for shift
comb += self.imm_out.data.eq(self.dec.SH32)
comb += self.imm_out.ok.eq(1)
+ with m.Case(In2Sel.CONST_XBI): # unsigned - for grevi
+ comb += self.imm_out.data.eq(self.dec.FormXB.XBI)
+ comb += self.imm_out.ok.eq(1)
return m
comb += reg.data.eq(self.dec.RC)
comb += reg.ok.eq(1)
with m.Case(In3Sel.RT):
- # for TII-form ternary
+ # for TLI-form ternlogi
comb += reg.data.eq(self.dec.RT)
comb += reg.ok.eq(1)
class DecodeOut(Elaboratable):
"""DecodeOut from instruction
- decodes output register RA, RT or SPR
+ decodes output register RA, RT, FRS, FRT, or SPR
"""
def __init__(self, dec, op, regreduce_en):
self.insn_in = Signal(32, reset_less=True)
self.reg_out = Data(5, "reg_o")
self.spr_out = Data(SPR, "spr_o")
- self.fast_out = Data(3, "fast_o")
+ self.fast_out = Data(4, "fast_o")
+ self.state_out = Data(3, "state_o")
def elaborate(self, platform):
m = Module()
# select Register out field
with m.Switch(self.sel_in):
+ with m.Case(OutSel.FRS):
+ comb += reg.data.eq(self.dec.FRS)
+ comb += reg.ok.eq(1)
with m.Case(OutSel.FRT):
comb += reg.data.eq(self.dec.FRT)
comb += reg.ok.eq(1)
comb += sprmap.spr_i.eq(spr)
comb += self.spr_out.eq(sprmap.spr_o)
comb += self.fast_out.eq(sprmap.fast_o)
+ comb += self.state_out.eq(sprmap.state_o)
# determine Fast Reg
with m.Switch(op.internal_op):
self.dec = dec
self.op = op
self.sel_in = Signal(OutSel, reset_less=True)
- self.svp64_fft_mode = Signal(reset_less=True) # SVP64 FFT mode
+ self.svp64_fft_mode = Signal(reset_less=True) # SVP64 FFT mode
self.lk = Signal(reset_less=True)
self.insn_in = Signal(32, reset_less=True)
self.reg_out = Data(5, "reg_o2")
- self.fp_madd_en = Signal(reset_less=True) # FFT instruction detected
- self.fast_out = Data(3, "fast_o2")
- self.fast_out3 = Data(3, "fast_o3")
+ self.fp_madd_en = Signal(reset_less=True) # FFT instruction detected
+ self.fast_out = Data(4, "fast_o2")
+ self.fast_out3 = Data(4, "fast_o3")
def elaborate(self, platform):
m = Module()
with m.Case(MicrOp.OP_RFID):
comb += self.fast_out.data.eq(FastRegsEnum.SRR1) # SRR1
comb += self.fast_out.ok.eq(1)
- comb += self.fast_out3.data.eq(FastRegsEnum.SVSRR0) # SVSRR0
+ comb += self.fast_out3.data.eq(FastRegsEnum.SVSRR0) # SVSRR0
comb += self.fast_out3.ok.eq(1)
# SVP64 FFT mode, FP mul-add: 2nd output reg (FRS) same as FRT
# will be offset by VL in hardware
- #with m.Case(MicrOp.OP_FP_MADD):
+ # with m.Case(MicrOp.OP_FP_MADD):
with m.If(self.svp64_fft_mode):
comb += self.reg_out.data.eq(self.dec.FRT)
comb += self.reg_out.ok.eq(1)
def __init__(self, dec):
self.dec = dec
- self.sel_in = Signal(RC, reset_less=True)
+ self.sel_in = Signal(RCOE, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
self.rc_out = Data(1, "rc")
# select Record bit out field
with m.Switch(self.sel_in):
- with m.Case(RC.RC):
+ with m.Case(RCOE.RC, RCOE.RC_ONLY):
comb += self.rc_out.data.eq(self.dec.Rc)
comb += self.rc_out.ok.eq(1)
- with m.Case(RC.ONE):
+ with m.Case(RCOE.ONE):
comb += self.rc_out.data.eq(1)
comb += self.rc_out.ok.eq(1)
- with m.Case(RC.NONE):
+ with m.Case(RCOE.NONE):
comb += self.rc_out.data.eq(0)
comb += self.rc_out.ok.eq(1)
class DecodeOE(Elaboratable):
"""DecodeOE from instruction
- decodes OE field: uses RC decode detection which might not be good
-
- -- For now, use "rc" in the decode table to decide whether oe exists.
- -- This is not entirely correct architecturally: For mulhd and
- -- mulhdu, the OE field is reserved. It remains to be seen what an
- -- actual POWER9 does if we set it on those instructions, for now we
- -- test that further down when assigning to the multiplier oe input.
+ decodes OE field: uses RC decode detection which has now been
+ updated to separate out RC_ONLY. all cases RC_ONLY are *NOT*
+ listening to the OE field, here.
"""
def __init__(self, dec, op):
self.dec = dec
self.op = op
- self.sel_in = Signal(RC, reset_less=True)
+ self.sel_in = Signal(RCOE, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
self.oe_out = Data(1, "oe")
def elaborate(self, platform):
m = Module()
comb = m.d.comb
- op = self.op
- with m.Switch(op.internal_op):
-
- # mulhw, mulhwu, mulhd, mulhdu - these *ignore* OE
- # also rotate
- # XXX ARGH! ignoring OE causes incompatibility with microwatt
- # http://lists.libre-soc.org/pipermail/libre-soc-dev/2020-August/000302.html
- with m.Case(MicrOp.OP_MUL_H64, MicrOp.OP_MUL_H32,
- MicrOp.OP_EXTS, MicrOp.OP_CNTZ,
- MicrOp.OP_SHL, MicrOp.OP_SHR, MicrOp.OP_RLC,
- MicrOp.OP_LOAD, MicrOp.OP_STORE,
- MicrOp.OP_RLCL, MicrOp.OP_RLCR,
- MicrOp.OP_EXTSWSLI):
- pass
-
- # all other ops decode OE field
+ with m.Switch(self.sel_in):
+ with m.Case(RCOE.RC):
+ comb += self.oe_out.data.eq(self.dec.OE)
+ comb += self.oe_out.ok.eq(1)
with m.Default():
- # select OE bit out field
- with m.Switch(self.sel_in):
- with m.Case(RC.RC):
- comb += self.oe_out.data.eq(self.dec.OE)
- comb += self.oe_out.ok.eq(1)
+ # default: clear OE.
+ comb += self.oe_out.data.eq(0)
+ comb += self.oe_out.ok.eq(0)
return m
self.cr_bitfield_b = Data(3, "cr_bitfield_b")
self.cr_bitfield_o = Data(3, "cr_bitfield_o")
self.whole_reg = Data(8, "cr_fxm")
- self.sv_override = Signal(2, reset_less=True) # do not do EXTRA spec
+ self.sv_override = Signal(2, reset_less=True) # do not do EXTRA spec
def elaborate(self, platform):
m = Module()
comb = m.d.comb
op = self.op
m.submodules.ppick = ppick = PriorityPicker(8, reverse_i=True,
- reverse_o=True)
+ reverse_o=True)
# zero-initialisation
comb += self.cr_bitfield.ok.eq(0)
with m.Case(CRInSel.NONE):
pass # No bitfield activated
with m.Case(CRInSel.CR0):
- comb += self.cr_bitfield.data.eq(0) # CR0 (MSB0 numbering)
+ comb += self.cr_bitfield.data.eq(0) # CR0 (MSB0 numbering)
comb += self.cr_bitfield.ok.eq(1)
comb += self.sv_override.eq(1)
with m.Case(CRInSel.CR1):
- comb += self.cr_bitfield.data.eq(1) # CR1 (MSB0 numbering)
+ comb += self.cr_bitfield.data.eq(1) # CR1 (MSB0 numbering)
comb += self.cr_bitfield.ok.eq(1)
comb += self.sv_override.eq(2)
with m.Case(CRInSel.BI):
with m.Case(CRInSel.WHOLE_REG):
comb += self.whole_reg.ok.eq(1)
move_one = Signal(reset_less=True)
- comb += move_one.eq(self.insn_in[20]) # MSB0 bit 11
+ comb += move_one.eq(self.insn_in[20]) # MSB0 bit 11
with m.If((op.internal_op == MicrOp.OP_MFCR) & move_one):
# must one-hot the FXM field
comb += ppick.i.eq(self.dec.FXM)
self.insn_in = Signal(32, reset_less=True)
self.cr_bitfield = Data(3, "cr_bitfield")
self.whole_reg = Data(8, "cr_fxm")
- self.sv_override = Signal(2, reset_less=True) # do not do EXTRA spec
+ self.sv_override = Signal(2, reset_less=True) # do not do EXTRA spec
def elaborate(self, platform):
m = Module()
comb = m.d.comb
op = self.op
m.submodules.ppick = ppick = PriorityPicker(8, reverse_i=True,
- reverse_o=True)
+ reverse_o=True)
comb += self.cr_bitfield.ok.eq(0)
comb += self.whole_reg.ok.eq(0)
with m.Case(CROutSel.NONE):
pass # No bitfield activated
with m.Case(CROutSel.CR0):
- comb += self.cr_bitfield.data.eq(0) # CR0 (MSB0 numbering)
+ comb += self.cr_bitfield.data.eq(0) # CR0 (MSB0 numbering)
comb += self.cr_bitfield.ok.eq(self.rc_in) # only when RC=1
comb += self.sv_override.eq(1)
with m.Case(CROutSel.CR1):
- comb += self.cr_bitfield.data.eq(1) # CR1 (MSB0 numbering)
+ comb += self.cr_bitfield.data.eq(1) # CR1 (MSB0 numbering)
comb += self.cr_bitfield.ok.eq(self.rc_in) # only when RC=1
comb += self.sv_override.eq(2)
with m.Case(CROutSel.BF):
with m.If(ppick.en_o):
comb += self.whole_reg.data.eq(ppick.o)
with m.Else():
- comb += self.whole_reg.data.eq(0b00000001) # CR7
+ comb += self.whole_reg.data.eq(0b00000001) # CR7
with m.Else():
comb += self.whole_reg.data.eq(self.dec.FXM)
with m.Else():
return m
+
# dictionary of Input Record field names that, if they exist,
# will need a corresponding CSV Decoder file column (actually, PowerOp)
# to be decoded (this includes the single bit names)
'is_signed': 'sgn',
'lk': 'lk',
'data_len': 'ldst_len',
+ 'reserve': 'rsrv',
'byte_reverse': 'br',
'sign_extend': 'sgn_ext',
'ldst_mode': 'upd',
only fields actually requested are copied over. hence, "subset" (duh).
"""
+
def __init__(self, dec, opkls=None, fn_name=None, final=False, state=None,
- svp64_en=True, regreduce_en=False):
+ svp64_en=True, regreduce_en=False):
self.svp64_en = svp64_en
self.regreduce_en = regreduce_en
if svp64_en:
- self.is_svp64_mode = Signal() # mark decoding as SVP64 Mode
- self.use_svp64_ldst_dec = Signal() # must use LDST decoder
+ self.is_svp64_mode = Signal() # mark decoding as SVP64 Mode
self.use_svp64_fft = Signal() # FFT Mode
- self.sv_rm = SVP64Rec(name="dec_svp64") # SVP64 RM field
+ self.sv_rm = SVP64Rec(name="dec_svp64") # SVP64 RM field
self.rm_dec = SVP64RMModeDecode("svp64_rm_dec")
# set these to the predicate mask bits needed for the ALU
- self.pred_sm = Signal() # TODO expand to SIMD mask width
- self.pred_dm = Signal() # TODO expand to SIMD mask width
+ self.pred_sm = Signal() # TODO expand to SIMD mask width
+ self.pred_dm = Signal() # TODO expand to SIMD mask width
self.sv_a_nz = Signal(1)
self.final = final
self.opkls = opkls
# alternatives. useful for PCR (Program Compatibility Register)
# amongst other things
if svp64_en:
- conditions = {'SVP64BREV': self.use_svp64_ldst_dec,
- 'SVP64FFT': self.use_svp64_fft,
- }
+ conditions = {
+ # XXX NO 'SVP64FFT': self.use_svp64_fft,
+ }
else:
conditions = None
# create decoder if one not already given
if dec is None:
dec = create_pdecode(name=fn_name, col_subset=col_subset,
- row_subset=row_subset,
- conditions=conditions)
+ row_subset=row_subset,
+ conditions=conditions)
self.dec = dec
# set up a copy of the PowerOp
self.state = state
def get_col_subset(self, do):
- subset = { 'cr_in', 'cr_out', 'rc_sel'} # needed, non-optional
+ subset = {'cr_in', 'cr_out', 'rc_sel'} # needed, non-optional
for k, v in record_names.items():
if hasattr(do, k):
subset.add(v)
- log ("get_col_subset", self.fn_name, do.fields, subset)
+ log("get_col_subset", self.fn_name, do.fields, subset)
return subset
def rowsubsetfn(self, opcode, row):
# really this should be done by modifying the CSV syntax
# to support multiple tasks (unit column multiple entries)
# see https://bugs.libre-soc.org/show_bug.cgi?id=310
- (self.fn_name == 'MMU' and row['unit'] == 'SPR' and
- row['internal op'] in ['OP_MTSPR', 'OP_MFSPR'])
+ (self.fn_name == 'MMU' and row['unit'] == 'SPR' and
+ row['internal op'] in ['OP_MTSPR', 'OP_MFSPR']) or
+ # urrr... and the KAIVB SPR, which must also be redirected
+ # (to the TRAP pipeline)
+ # see https://bugs.libre-soc.org/show_bug.cgi?id=859
+ (self.fn_name == 'TRAP' and row['unit'] == 'SPR' and
+ row['internal op'] in ['OP_MTSPR', 'OP_MFSPR'])
)
def ports(self):
if self.svp64_en:
ports += self.sv_rm.ports()
ports.append(self.is_svp64_mode)
- ports.append(self.use_svp64_ldst_dec )
- ports.append(self.use_svp64_fft )
+ ports.append(self.use_svp64_fft)
return ports
def needs_field(self, field, op_field):
else:
name = self.fn_name + "tmp"
self.e_tmp = Decode2ToExecute1Type(name=name, opkls=self.opkls,
- regreduce_en=self.regreduce_en)
+ regreduce_en=self.regreduce_en)
# set up submodule decoders
m.submodules.dec = dec = self.dec
# for SPR set/get
fn = self.op_get("function_unit")
spr = Signal(10, reset_less=True)
- comb += spr.eq(decode_spr_num(self.dec.SPR)) # from XFX
+ comb += spr.eq(decode_spr_num(self.dec.SPR)) # from XFX
# Microwatt doesn't implement the partition table
# instead has PRTBL register (SPR) to point to process table
+ # Kestrel has a KAIVB SPR to "rebase" exceptions. rebasing is normally
+ # done with Hypervisor Mode which is not implemented (yet)
is_spr_mv = Signal()
is_mmu_spr = Signal()
+ is_trap_spr = Signal()
comb += is_spr_mv.eq((internal_op == MicrOp.OP_MTSPR) |
(internal_op == MicrOp.OP_MFSPR))
comb += is_mmu_spr.eq((spr == SPR.DSISR.value) |
(spr == SPR.DAR.value) |
(spr == SPR.PRTBL.value) |
(spr == SPR.PIDR.value))
+ comb += is_trap_spr.eq((spr == SPR.KAIVB.value)
+ )
# MMU must receive MMU SPRs
with m.If(is_spr_mv & (fn == Function.SPR) & is_mmu_spr):
comb += self.do_copy("fn_unit", Function.MMU)
comb += self.do_copy("insn_type", internal_op)
- # SPR pipe must *not* receive MMU SPRs
- with m.Elif(is_spr_mv & (fn == Function.MMU) & ~is_mmu_spr):
+ # TRAP must receive TRAP SPR KAIVB
+ with m.If(is_spr_mv & (fn == Function.SPR) & is_trap_spr):
+ comb += self.do_copy("fn_unit", Function.TRAP)
+ comb += self.do_copy("insn_type", internal_op)
+ # SPR pipe must *not* receive MMU or TRAP SPRs
+ with m.Elif(is_spr_mv & ((fn == Function.MMU) & ~is_mmu_spr) &
+ ((fn == Function.TRAP) & ~is_trap_spr)):
comb += self.do_copy("fn_unit", Function.NONE)
comb += self.do_copy("insn_type", MicrOp.OP_ILLEGAL)
# all others ok
if self.needs_field("imm_data", "in2_sel"):
m.submodules.dec_bi = dec_bi = DecodeBImm(self.dec)
comb += dec_bi.sel_in.eq(self.op_get("in2_sel"))
- comb += self.do_copy("imm_data", dec_bi.imm_out) # imm in RB
+ comb += self.do_copy("imm_data", dec_bi.imm_out) # imm in RB
# rc and oe out
comb += self.do_copy("rc", dec_rc.rc_out)
rc_out = self.dec_rc.rc_out.data
with m.Switch(self.op_get("cr_out")):
with m.Case(CROutSel.CR0, CROutSel.CR1):
- comb += self.do_copy("write_cr0", rc_out) # only when RC=1
+ comb += self.do_copy("write_cr0", rc_out) # only when RC=1
with m.Case(CROutSel.BF, CROutSel.BT):
comb += self.do_copy("write_cr0", 1)
# the alternative decoder, svdecldst. what a mess... *sigh*
sv_ptype = self.op_get("SV_Ptype")
fn = self.op_get("function_unit")
- # detect major opcode for LDs: include 58 here. from CSV files.
- # BLECH! TODO: these should be done using "mini decoders",
- # using row and column subsets
- is_major_ld = Signal()
- major = Signal(6) # bits... errr... MSB0 0..5 which is 26:32 python
- comb += major.eq(self.dec.opcode_in[26:32])
- comb += is_major_ld.eq((major == 34) | (major == 35) |
- (major == 50) | (major == 51) |
- (major == 48) | (major == 49) |
- (major == 42) | (major == 43) |
- (major == 40) | (major == 41) |
- (major == 32) | (major == 33) |
- (major == 58))
- with m.If(self.is_svp64_mode & is_major_ld):
- # straight-up: "it's a LD". this gives enough info
- # for SVP64 RM Mode decoding to detect LD/ST, and
- # consequently detect the SHIFT mode. sigh
- comb += rm_dec.fn_in.eq(Function.LDST)
- with m.Else():
- comb += rm_dec.fn_in.eq(fn) # decode needs to know Fn type
- comb += rm_dec.ptype_in.eq(sv_ptype) # Single/Twin predicated
- comb += rm_dec.rc_in.eq(rc_out) # Rc=1
- comb += rm_dec.rm_in.eq(self.sv_rm) # SVP64 RM mode
+ comb += rm_dec.fn_in.eq(fn) # decode needs to know Fn type
+ comb += rm_dec.ptype_in.eq(sv_ptype) # Single/Twin predicated
+ comb += rm_dec.rc_in.eq(rc_out) # Rc=1
+ comb += rm_dec.rm_in.eq(self.sv_rm) # SVP64 RM mode
if self.needs_field("imm_data", "in2_sel"):
bzero = dec_bi.imm_out.ok & ~dec_bi.imm_out.data.bool()
- comb += rm_dec.ldst_imz_in.eq(bzero) # B immediate is zero
+ comb += rm_dec.ldst_imz_in.eq(bzero) # B immediate is zero
# main PowerDecoder2 determines if different SVP64 modes enabled
- if not self.final:
- # if shift mode requested
- shiftmode = rm_dec.ldstmode == SVP64LDSTmode.SHIFT
- comb += self.use_svp64_ldst_dec.eq(shiftmode)
# detect if SVP64 FFT mode enabled (really bad hack),
# exclude fcfids and others
# XXX this is a REALLY bad hack, REALLY has to be done better.
# likely with a sub-decoder.
- xo5 = Signal(1) # 1 bit from Minor 59 XO field == 0b0XXXX
- comb += xo5.eq(self.dec.opcode_in[5])
- xo = Signal(5) # 5 bits from Minor 59 fcfids == 0b01110
- comb += xo.eq(self.dec.opcode_in[1:6])
- comb += self.use_svp64_fft.eq((major == 59) & (xo5 == 0b0) &
- (xo != 0b01110))
+ # what this ultimately does is enable the 2nd implicit register
+ # (FRS) for SVP64-decoding. all of these instructions are
+ # 3-in 2-out but there is not enough room either in the
+ # opcode *or* EXTRA2/3 to specify a 5th operand.
+ major = Signal(6)
+ comb += major.eq(self.dec.opcode_in[26:32])
+ xo = Signal(10)
+ comb += xo.eq(self.dec.opcode_in[1:11])
+ comb += self.use_svp64_fft.eq((major == 59) & xo.matches(
+ '-----00100', # ffmsubs
+ '-----00101', # ffmadds
+ '-----00110', # ffnmsubs
+ '-----00111', # ffnmadds
+ '1111100000', # ffadds
+ '-----11011', # fdmadds
+ ))
# decoded/selected instruction flags
comb += self.do_copy("data_len", self.op_get("ldst_len"))
comb += self.do_copy("byte_reverse", self.op_get("br"))
comb += self.do_copy("sign_extend", self.op_get("sgn_ext"))
comb += self.do_copy("ldst_mode", self.op_get("upd")) # LD/ST mode
+ comb += self.do_copy("reserve", self.op_get("rsrv")) # atomic
# copy over SVP64 input record fields (if they exist)
if self.svp64_en:
# TODO, really do we have to do these explicitly?? sigh
- #for (field, _) in sv_input_record_layout:
+ # for (field, _) in sv_input_record_layout:
# comb += self.do_copy(field, self.rm_dec.op_get(field))
comb += self.do_copy("sv_saturate", self.rm_dec.saturate)
comb += self.do_copy("sv_Ptype", self.rm_dec.ptype_in)
to make this work, TestIssuer must notice "exception.happened"
after the (failed) LD/ST and copies the LDSTException info from
the output, into here (PowerDecoder2). without incrementing PC.
+
+ also instr_fault works the same way: the instruction is "rewritten"
+ so that the "fake" op that gets created is OP_FETCH_FAILED
"""
def __init__(self, dec, opkls=None, fn_name=None, final=False,
- state=None, svp64_en=True, regreduce_en=False):
+ state=None, svp64_en=True, regreduce_en=False):
super().__init__(dec, opkls, fn_name, final, state, svp64_en,
regreduce_en=False)
- self.ldst_exc = LDSTException("dec2_exc")
+ self.ldst_exc = LDSTException("dec2_exc") # rewrites as OP_TRAP
+ self.instr_fault = Signal() # rewrites instruction as OP_FETCH_FAILED
if self.svp64_en:
self.cr_out_isvec = Signal(1, name="cr_out_isvec")
self.in3_step = Signal(7, name="reg_c_step")
self.o_step = Signal(7, name="reg_o_step")
self.o2_step = Signal(7, name="reg_o2_step")
- self.remap_active = Signal(5, name="remap_active") # per reg
- self.no_in_vec = Signal(1, name="no_in_vec") # no inputs vector
- self.no_out_vec = Signal(1, name="no_out_vec") # no outputs vector
+ self.remap_active = Signal(5, name="remap_active") # per reg
+ self.no_in_vec = Signal(1, name="no_in_vec") # no inputs vector
+ self.no_out_vec = Signal(1, name="no_out_vec") # no outputs vector
self.loop_continue = Signal(1, name="loop_continue")
else:
self.no_in_vec = Const(1, 1)
#######
# CR out
- comb += crout_svdec.idx.eq(self.op_get("sv_cr_out")) # SVP64 CR out
+ # SVP64 CR out
+ comb += crout_svdec.idx.eq(self.op_get("sv_cr_out"))
comb += self.cr_out_isvec.eq(crout_svdec.isvec)
#######
# indices are slightly different, BA/BB mess sorted above
comb += crin_svdec.idx.eq(cr_a_idx) # SVP64 CR in A
comb += crin_svdec_b.idx.eq(cr_b_idx) # SVP64 CR in B
- comb += crin_svdec_o.idx.eq(self.op_get("sv_cr_out")) # SVP64 CR out
+ # SVP64 CR out
+ comb += crin_svdec_o.idx.eq(self.op_get("sv_cr_out"))
# get SVSTATE srcstep (TODO: elwidth etc.) needed below
vl = Signal.like(self.state.svstate.vl)
+ maxvl = Signal.like(self.state.svstate.maxvl)
+ subvl = Signal.like(self.rm_dec.rm_in.subvl)
srcstep = Signal.like(self.state.svstate.srcstep)
dststep = Signal.like(self.state.svstate.dststep)
+ ssubstep = Signal.like(self.state.svstate.ssubstep)
+ dsubstep = Signal.like(self.state.svstate.ssubstep)
comb += vl.eq(self.state.svstate.vl)
+ comb += maxvl.eq(self.state.svstate.maxvl)
+ comb += subvl.eq(self.rm_dec.rm_in.subvl)
comb += srcstep.eq(self.state.svstate.srcstep)
comb += dststep.eq(self.state.svstate.dststep)
+ comb += ssubstep.eq(self.state.svstate.ssubstep)
+ comb += dsubstep.eq(self.state.svstate.dsubstep)
in1_step, in2_step = self.in1_step, self.in2_step
in3_step = self.in3_step
o_step, o2_step = self.o_step, self.o2_step
+ # multiply vl by subvl - note that this is only 7 bit!
+ # when elwidth overrides get involved this will have to go up
+ vmax = Signal(7)
+ comb += vmax.eq(vl*(subvl+1))
+
# registers a, b, c and out and out2 (LD/ST EA)
sv_etype = self.op_get("SV_Etype")
for i, stuff in enumerate((
rname, to_reg, fromreg, svdec, remapstep, out = stuff
comb += svdec.extra.eq(extra) # EXTRA field of SVP64 RM
comb += svdec.etype.eq(sv_etype) # EXTRA2/3 for this insn
- comb += svdec.reg_in.eq(fromreg.data) # 3-bit (CR0/BC/BFA)
+ comb += svdec.reg_in.eq(fromreg.data) # 3-bit (CR0/BC/BFA)
comb += to_reg.ok.eq(fromreg.ok)
# *screaam* FFT mode needs an extra offset for RB
# similar to FRS/FRT (below). all of this needs cleanup
with m.If(dec_o2.reg_out.ok & dec_o2.fp_madd_en):
with m.If(~self.remap_active[i]):
with m.If(svdec.isvec):
- comb += offs.eq(vl) # VL for Vectors
+ comb += offs.eq(maxvl) # MAXVL for Vectors
# detect if Vectorised: add srcstep/dststep if yes.
# to_reg is 7-bits, outs get dststep added, ins get srcstep
with m.If(svdec.isvec):
selectstep = dststep if out else srcstep
+ subselect = dsubstep if out else ssubstep
step = Signal(7, name="step_%s" % rname.lower())
with m.If(self.remap_active[i]):
- comb += step.eq(remapstep)
+ comb += step.eq((remapstep*(subvl+1))+subselect)
with m.Else():
- comb += step.eq(selectstep)
+ comb += step.eq((selectstep*(subvl+1))+subselect)
# reverse gear goes the opposite way
with m.If(self.rm_dec.reverse_gear):
- comb += to_reg.data.eq(offs+svdec.reg_out+(vl-1-step))
+ comb += to_reg.data.eq(offs+svdec.reg_out+(vmax-1-step))
with m.Else():
comb += to_reg.data.eq(offs+step+svdec.reg_out)
with m.Else():
comb += to_reg.data.eq(offs+svdec.reg_out)
# SVP64 in/out fields
- comb += in1_svdec.idx.eq(self.op_get("sv_in1")) # reg #1 (in1_sel)
+ comb += in1_svdec.idx.eq(self.op_get("sv_in1")) # reg #1 (in1_sel)
comb += in2_svdec.idx.eq(self.op_get("sv_in2")) # reg #2 (in2_sel)
comb += in3_svdec.idx.eq(self.op_get("sv_in3")) # reg #3 (in3_sel)
comb += o_svdec.idx.eq(self.op_get("sv_out")) # output (out_sel)
- comb += o2_svdec.idx.eq(self.op_get("sv_out2")) # output (implicit)
+ # output (implicit)
+ comb += o2_svdec.idx.eq(self.op_get("sv_out2"))
# XXX TODO - work out where this should come from. the problem is
# that LD-with-update is implied (computed from "is instruction in
# "update mode" rather than specified cleanly as its own CSV column
comb += offs.eq(0)
with m.If(~self.remap_active[4]):
with m.If(o2_svdec.isvec):
- comb += offs.eq(vl) # VL for Vectors
+ comb += offs.eq(vl) # VL for Vectors
with m.Else():
comb += offs.eq(1) # add 1 if scalar
- svdec = o_svdec # yes take source as o_svdec...
+ svdec = o_svdec # yes take source as o_svdec...
with m.If(svdec.isvec):
step = Signal(7, name="step_%s" % rname.lower())
with m.If(self.remap_active[4]):
# TODO add SPRs here. must be True when *all* are scalar
l = map(lambda svdec: svdec.isvec, [in1_svdec, in2_svdec, in3_svdec,
- crin_svdec, crin_svdec_b, crin_svdec_o])
- comb += self.no_in_vec.eq(~Cat(*l).bool()) # all input scalar
- l = map(lambda svdec: svdec.isvec, [o2_svdec, o_svdec, crout_svdec])
+ crin_svdec, crin_svdec_b,
+ crin_svdec_o])
+ comb += self.no_in_vec.eq(~Cat(*l).bool()) # all input scalar
+ l = map(lambda svdec: svdec.isvec, [
+ o2_svdec, o_svdec, crout_svdec])
# in mapreduce mode, scalar out is *allowed*
with m.If(self.rm_dec.mode == SVP64RMMode.MAPREDUCE.value):
comb += self.no_out_vec.eq(0)
with m.Else():
- comb += self.no_out_vec.eq(~Cat(*l).bool()) # all output scalar
+ # all output scalar
+ comb += self.no_out_vec.eq(~Cat(*l).bool())
# now create a general-purpose "test" as to whether looping
# should continue. this doesn't include predication bit-tests
loop = self.loop_continue
(e.read_cr1, self.dec_cr_in, "cr_bitfield", crin_svdec, 0),
(e.read_cr2, self.dec_cr_in, "cr_bitfield_b", crin_svdec_b, 0),
(e.read_cr3, self.dec_cr_in, "cr_bitfield_o", crin_svdec_o, 0),
- (e.write_cr, self.dec_cr_out, "cr_bitfield", crout_svdec, 1)):
+ (e.write_cr, self.dec_cr_out, "cr_bitfield", crout_svdec, 1)):
fromreg = getattr(cr, name)
comb += svdec.extra.eq(extra) # EXTRA field of SVP64 RM
comb += svdec.etype.eq(sv_etype) # EXTRA2/3 for this insn
- comb += svdec.cr_in.eq(fromreg.data) # 3-bit (CR0/BC/BFA)
+ comb += svdec.cr_in.eq(fromreg.data) # 3-bit (CR0/BC/BFA)
with m.If(svdec.isvec):
# check if this is CR0 or CR1: treated differently
# (does not "listen" to EXTRA2/3 spec for a start)
# also: the CRs start from completely different locations
step = dststep if out else srcstep
- with m.If(cr.sv_override == 1): # CR0
+ with m.If(cr.sv_override == 1): # CR0
offs = SVP64CROffs.CR0
comb += to_reg.data.eq(step+offs)
- with m.Elif(cr.sv_override == 2): # CR1
+ with m.Elif(cr.sv_override == 2): # CR1
offs = SVP64CROffs.CR1
comb += to_reg.data.eq(step+1)
with m.Else():
- comb += to_reg.data.eq(step+svdec.cr_out) # 7-bit out
+ comb += to_reg.data.eq(step+svdec.cr_out) # 7-bit out
with m.Else():
- comb += to_reg.data.eq(svdec.cr_out) # 7-bit output
+ comb += to_reg.data.eq(svdec.cr_out) # 7-bit output
comb += to_reg.ok.eq(fromreg.ok)
# sigh must determine if RA is nonzero (7 bit)
# connect up to/from read/write CRs
for to_reg, cr, name in (
- (e.read_cr1, self.dec_cr_in, "cr_bitfield", ),
- (e.read_cr2, self.dec_cr_in, "cr_bitfield_b", ),
- (e.read_cr3, self.dec_cr_in, "cr_bitfield_o", ),
- (e.write_cr, self.dec_cr_out, "cr_bitfield", )):
+ (e.read_cr1, self.dec_cr_in, "cr_bitfield", ),
+ (e.read_cr2, self.dec_cr_in, "cr_bitfield_b", ),
+ (e.read_cr3, self.dec_cr_in, "cr_bitfield_o", ),
+ (e.write_cr, self.dec_cr_out, "cr_bitfield", )):
fromreg = getattr(cr, name)
comb += to_reg.data.eq(fromreg.data)
comb += to_reg.ok.eq(fromreg.ok)
if self.svp64_en:
- comb += self.rm_dec.ldst_ra_vec.eq(self.in1_isvec) # RA is vector
+ comb += self.rm_dec.ldst_ra_vec.eq(self.in1_isvec) # RA is vector
# SPRs out
comb += e.read_spr1.eq(dec_a.spr_out)
comb += e.read_fast2.eq(dec_b.fast_out)
comb += e.write_fast1.eq(dec_o.fast_out) # SRR0 (OP_RFID)
comb += e.write_fast2.eq(dec_o2.fast_out) # SRR1 (ditto)
- comb += e.write_fast3.eq(dec_o2.fast_out3) # SVSRR0 (ditto)
+ comb += e.write_fast3.eq(dec_o2.fast_out3) # SVSRR0 (ditto)
+ # and State regs (DEC, TB)
+ comb += e.read_state1.eq(dec_a.state_out) # DEC/TB
+ comb += e.write_state1.eq(dec_o.state_out) # DEC/TB
# sigh this is exactly the sort of thing for which the
# decoder is designed to not need. MTSPR, MFSPR and others need
# access to the XER bits. however setting e.oe is not appropriate
internal_op = self.op_get("internal_op")
with m.If(internal_op == MicrOp.OP_MFSPR):
- comb += e.xer_in.eq(0b111) # SO, CA, OV
+ comb += e.xer_in.eq(0b111) # SO, CA, OV
with m.If(internal_op == MicrOp.OP_CMP):
- comb += e.xer_in.eq(1<<XERRegsEnum.SO) # SO
+ comb += e.xer_in.eq(1 << XERRegsEnum.SO) # SO
with m.If(internal_op == MicrOp.OP_MTSPR):
comb += e.xer_out.eq(1)
with m.If(op.internal_op == MicrOp.OP_TRAP):
# *DO NOT* call self.trap here. that would reset absolutely
# everything including destroying read of RA and RB.
- comb += self.do_copy("trapaddr", 0x70) # strip first nibble
+ comb += self.do_copy("trapaddr", 0x70) # strip first nibble
####################
# ok so the instruction's been decoded, blah blah, however
illeg_ok = Signal()
ldst_exc = self.ldst_exc
- comb += ext_irq_ok.eq(ext_irq & msr[MSR.EE]) # v3.0B p944 (MSR.EE)
- comb += dec_irq_ok.eq(dec_spr[63] & msr[MSR.EE]) # 6.5.11 p1076
+ comb += ext_irq_ok.eq(ext_irq & msr[MSR.EE]) # v3.0B p944 (MSR.EE)
+ comb += dec_irq_ok.eq(dec_spr[63] & msr[MSR.EE]) # 6.5.11 p1076
comb += priv_ok.eq(is_priv_insn & msr[MSR.PR])
comb += illeg_ok.eq(op.internal_op == MicrOp.OP_ILLEGAL)
+ # absolute top priority: check for an instruction failed
+ with m.If(self.instr_fault):
+ comb += self.e.eq(0) # reset eeeeeverything
+ comb += self.do_copy("insn", self.dec.opcode_in, True)
+ comb += self.do_copy("insn_type", MicrOp.OP_FETCH_FAILED, True)
+ comb += self.do_copy("fn_unit", Function.MMU, True)
+ comb += self.do_copy("cia", self.state.pc, True) # PC
+ comb += self.do_copy("msr", self.state.msr, True) # MSR
+ # special override on internal_op, due to being a "fake" op
+ comb += self.dec.op.internal_op.eq(MicrOp.OP_FETCH_FAILED)
+
# LD/ST exceptions. TestIssuer copies the exception info at us
# after a failed LD/ST.
- with m.If(ldst_exc.happened):
+ with m.Elif(ldst_exc.happened):
with m.If(ldst_exc.alignment):
- self.trap(m, TT.PRIV, 0x600)
+ self.trap(m, TT.MEMEXC, 0x600)
with m.Elif(ldst_exc.instr_fault):
with m.If(ldst_exc.segment_fault):
- self.trap(m, TT.PRIV, 0x480)
+ self.trap(m, TT.MEMEXC, 0x480)
with m.Else():
# pass exception info to trap to create SRR1
self.trap(m, TT.MEMEXC, 0x400, ldst_exc)
with m.Else():
with m.If(ldst_exc.segment_fault):
- self.trap(m, TT.PRIV, 0x380)
+ self.trap(m, TT.MEMEXC, 0x380)
with m.Else():
- self.trap(m, TT.PRIV, 0x300)
+ self.trap(m, TT.MEMEXC, 0x300)
# decrement counter (v3.0B p1099): TODO 32-bit version (MSR.LPCR)
with m.Elif(dec_irq_ok):
comb += self.do_copy("insn", self.dec.opcode_in, True)
comb += self.do_copy("insn_type", MicrOp.OP_TRAP, True)
comb += self.do_copy("fn_unit", Function.TRAP, True)
- comb += self.do_copy("trapaddr", trapaddr >> 4, True) # bottom 4 bits
+ comb += self.do_copy("trapaddr", trapaddr >> 4, True) # bottom 4 bits
comb += self.do_copy("traptype", traptype, True) # request type
comb += self.do_copy("ldst_exc", ldst_exc, True) # request type
- comb += self.do_copy("msr", self.state.msr, True) # copy of MSR "state"
+ comb += self.do_copy("msr", self.state.msr,
+ True) # copy of MSR "state"
comb += self.do_copy("cia", self.state.pc, True) # copy of PC "state"
comb += self.do_copy("svstate", self.state.svstate, True) # SVSTATE
-
-def get_rdflags(e, cu):
+def get_rdflags(m, e, cu):
+ """returns a sequential list of the read "ok" flags for a given FU.
+ this list is in order of the CompUnit input specs
+ """
rdl = []
for idx in range(cu.n_src):
regfile, regname, _ = cu.get_in_spec(idx)
- rdflag, read = regspec_decode_read(e, regfile, regname)
- rdl.append(rdflag)
+ decinfo = regspec_decode_read(m, e, regfile, regname)
+ rdl.append(decinfo.okflag)
log("rdflags", rdl)
return Cat(*rdl)