add some debug checking to get_pdecode_cr_out

[soc.git] / src / soc / decoder / power_decoder2.py

"""Power ISA Decoder second stage

based on Anton Blanchard microwatt decode2.vhdl

Note: OP_TRAP is used for exceptions and interrupts (micro-code style) by
over-riding the internal opcode when an exception is needed.
"""

from nmigen import Module, Elaboratable, Signal, Mux, Const, Cat, Repl, Record
from nmigen.cli import rtlil
from soc.regfile.regfiles import XERRegs

from nmutil.picker import PriorityPicker
from nmutil.iocontrol import RecordObject
from nmutil.extend import exts

from soc.experiment.mem_types import LDSTException

from soc.decoder.power_regspec_map import regspec_decode_read
from soc.decoder.power_regspec_map import regspec_decode_write
from soc.decoder.power_decoder import create_pdecode
from soc.decoder.power_enums import (MicrOp, CryIn, Function,
                                     CRInSel, CROutSel,
                                     LdstLen, In1Sel, In2Sel, In3Sel,
                                     OutSel, SPR, RC, LDSTMode,
                                     SVEXTRA, SVEtype)
from soc.decoder.decode2execute1 import (Decode2ToExecute1Type, Data,
                                         Decode2ToOperand)
from soc.sv.svp64 import SVP64Rec
from soc.consts import (MSR, sel, SPEC, EXTRA2, EXTRA3, SVP64P, field,
                        SPEC_SIZE, SPECb, SPEC_AUG_SIZE)

from soc.regfile.regfiles import FastRegs
from soc.consts import TT
from soc.config.state import CoreState
from soc.regfile.util import spr_to_fast


def decode_spr_num(spr):
    return Cat(spr[5:10], spr[0:5])


def instr_is_priv(m, op, insn):
    """determines if the instruction is privileged or not
    """
    comb = m.d.comb
    is_priv_insn = Signal(reset_less=True)
    with m.Switch(op):
        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_MFSPR, MicrOp.OP_MTSPR):
            with m.If(insn[20]):  # field XFX.spr[-1] i think
                comb += is_priv_insn.eq(1)
    return is_priv_insn


class SPRMap(Elaboratable):
    """SPRMap: maps POWER9 SPR numbers to internal enum values, fast and slow
    """

    def __init__(self):
        self.spr_i = Signal(10, reset_less=True)
        self.spr_o = Data(SPR, name="spr_o")
        self.fast_o = Data(3, name="fast_o")

    def elaborate(self, platform):
        m = Module()
        with m.Switch(self.spr_i):
            for i, x in enumerate(SPR):
                with m.Case(x.value):
                    m.d.comb += self.spr_o.data.eq(i)
                    m.d.comb += self.spr_o.ok.eq(1)
            for x, v in spr_to_fast.items():
                with m.Case(x.value):
                    m.d.comb += self.fast_o.data.eq(v)
                    m.d.comb += self.fast_o.ok.eq(1)
        return m


class SVP64ExtraSpec(Elaboratable):
    """SVP64ExtraSpec - decodes SVP64 Extra specification.

    selects the required EXTRA2/3 field.

    see https://libre-soc.org/openpower/sv/svp64/
    """
    def __init__(self):
        self.extra   = Signal(9, reset_less=True)
        self.etype   = Signal(SVEtype, reset_less=True) # 2 or 3 bits
        self.idx     = Signal(SVEXTRA, reset_less=True) # which part of extra
        self.spec  = Signal(3) # EXTRA spec for the register

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        spec = self.spec
        extra = self.extra

        # back in the LDSTRM-* and RM-* files generated by sv_analysis.py
        # we marked every op with an Etype: EXTRA2 or EXTRA3, and also said
        # which of the 4 (or 3 for EXTRA3) sub-fields of bits 10:18 contain
        # the register-extension information.  extract those now
        with m.Switch(self.etype):
            # 2-bit index selection mode
            with m.Case(SVEtype.EXTRA2):
                with m.Switch(self.idx):
                    with m.Case(SVEXTRA.Idx0):  # 1st 2 bits [0:1]
                        comb += spec[SPEC.VEC].eq(extra[EXTRA2.IDX0_VEC])
                        comb += spec[SPEC.MSB].eq(extra[EXTRA2.IDX0_MSB])
                    with m.Case(SVEXTRA.Idx1):  # 2nd 2 bits [2:3]
                        comb += spec[SPEC.VEC].eq(extra[EXTRA2.IDX1_VEC])
                        comb += spec[SPEC.MSB].eq(extra[EXTRA2.IDX1_MSB])
                    with m.Case(SVEXTRA.Idx2):  # 3rd 2 bits [4:5]
                        comb += spec[SPEC.VEC].eq(extra[EXTRA2.IDX2_VEC])
                        comb += spec[SPEC.MSB].eq(extra[EXTRA2.IDX2_MSB])
                    with m.Case(SVEXTRA.Idx3):  # 4th 2 bits [6:7]
                        comb += spec[SPEC.VEC].eq(extra[EXTRA2.IDX3_VEC])
                        comb += spec[SPEC.MSB].eq(extra[EXTRA2.IDX3_MSB])
            # 3-bit index selection mode
            with m.Case(SVEtype.EXTRA3):
                with m.Switch(self.idx):
                    with m.Case(SVEXTRA.Idx0):  # 1st 3 bits [0:2]
                        comb += spec.eq(sel(extra, EXTRA3.IDX0))
                    with m.Case(SVEXTRA.Idx1):  # 2nd 3 bits [3:5]
                        comb += spec.eq(sel(extra, EXTRA3.IDX1))
                    with m.Case(SVEXTRA.Idx2):  # 3rd 3 bits [6:8]
                        comb += spec.eq(sel(extra, EXTRA3.IDX2))
                    # cannot fit more than 9 bits so there is no 4th thing

        return m


class SVP64RegExtra(SVP64ExtraSpec):
    """SVP64RegExtra - decodes SVP64 Extra fields to determine reg extension

    incoming 5-bit GPR/FP is turned into a 7-bit and marked as scalar/vector
    depending on info in one of the positions in the EXTRA field.

    designed so that "no change" to the 5-bit register number occurs if
    SV either does not apply or the relevant EXTRA2/3 field bits are zero.

    see https://libre-soc.org/openpower/sv/svp64/
    """
    def __init__(self):
        SVP64ExtraSpec.__init__(self)
        self.reg_in  = Signal(5) # incoming reg number (5 bits, RA, RB)
        self.reg_out = Signal(7) # extra-augmented output (7 bits)
        self.isvec   = Signal(1) # reg is marked as vector if true

    def elaborate(self, platform):
        m = super().elaborate(platform) # select required EXTRA2/3
        comb = m.d.comb

        # first get the spec.  if not changed it's "scalar identity behaviour"
        # which is zero which is ok.
        spec = self.spec

        # now decode it. bit 0 is "scalar/vector".  note that spec could be zero
        #  from above, which (by design) has the effect of "no change", below.

        # simple: isvec is top bit of spec
        comb += self.isvec.eq(spec[SPEC.VEC])
        # extra bits for register number augmentation
        spec_aug = Signal(SPEC_AUG_SIZE)
        comb += spec_aug.eq(field(spec, SPECb.MSB, SPECb.LSB, SPEC_SIZE))

        # decode vector differently from scalar
        with m.If(self.isvec):
            # Vector: shifted up, extra in LSBs (RA << 2) | spec[1:2]
            comb += self.reg_out.eq(Cat(spec_aug, self.reg_in))
        with m.Else():
            # Scalar: not shifted up, extra in MSBs RA | (spec[1:2] << 5)
            comb += self.reg_out.eq(Cat(self.reg_in, spec_aug))

        return m


class SVP64CRExtra(SVP64ExtraSpec):
    """SVP64CRExtra - decodes SVP64 Extra fields to determine CR extension

    incoming 3-bit CR is turned into a 7-bit and marked as scalar/vector
    depending on info in one of the positions in the EXTRA field.

    yes, really, 128 CRs.  INT is 128, FP is 128, therefore CRs are 128.

    designed so that "no change" to the 3-bit CR register number occurs if
    SV either does not apply or the relevant EXTRA2/3 field bits are zero.

    see https://libre-soc.org/openpower/sv/svp64/appendix
    """
    def __init__(self):
        SVP64ExtraSpec.__init__(self)
        self.cr_in  = Signal(3) # incoming CR number (3 bits, BA[0:2], BFA)
        self.cr_out = Signal(7) # extra-augmented CR output (7 bits)
        self.isvec  = Signal(1) # reg is marked as vector if true

    def elaborate(self, platform):
        m = super().elaborate(platform) # select required EXTRA2/3
        comb = m.d.comb

        # first get the spec.  if not changed it's "scalar identity behaviour"
        # which is zero which is ok.
        spec = self.spec

        # now decode it. bit 0 is "scalar/vector".  note that spec could be zero
        #  from above, which (by design) has the effect of "no change", below.

        # simple: isvec is top bit of spec
        comb += self.isvec.eq(spec[SPEC.VEC])
        # extra bits for register number augmentation
        spec_aug = Signal(SPEC_AUG_SIZE)
        comb += spec_aug.eq(field(spec, SPECb.MSB, SPECb.LSB, SPEC_SIZE))

        # decode vector differently from scalar, insert bits 1 and 2 accordingly
        with m.If(self.isvec):
            # Vector: shifted up, extra in LSBs (CR << 4) | (spec[1:2] << 2)
            comb += self.cr_out.eq(Cat(Const(0, 2), spec_aug, self.cr_in))
        with m.Else():
            # Scalar: not shifted up, extra in MSBs CR | (spec[1:2] << 3)
            comb += self.cr_out.eq(Cat(self.cr_in, spec_aug))

        return m


class DecodeA(Elaboratable):
    """DecodeA from instruction

    decodes register RA, implicit and explicit CSRs
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(In1Sel, reset_less=True)
        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")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        op = self.dec.op
        reg = self.reg_out
        m.submodules.sprmap = sprmap = SPRMap()

        # select Register A field
        ra = Signal(5, reset_less=True)
        comb += ra.eq(self.dec.RA)
        with m.If((self.sel_in == In1Sel.RA) |
                  ((self.sel_in == In1Sel.RA_OR_ZERO) &
                   (ra != Const(0, 5)))):
            comb += reg.data.eq(ra)
            comb += reg.ok.eq(1)

        # some Logic/ALU ops have RS as the 3rd arg, but no "RA".
        # moved it to 1st position (in1_sel)... because
        rs = Signal(5, reset_less=True)
        comb += rs.eq(self.dec.RS)
        with m.If(self.sel_in == In1Sel.RS):
            comb += reg.data.eq(rs)
            comb += reg.ok.eq(1)

        # decode Fast-SPR based on instruction type
        with m.Switch(op.internal_op):

            # BC or BCREG: implicit register (CTR) NOTE: same in DecodeOut
            with m.Case(MicrOp.OP_BC):
                with m.If(~self.dec.BO[2]):  # 3.0B p38 BO2=0, use CTR reg
                    # constant: CTR
                    comb += self.fast_out.data.eq(FastRegs.CTR)
                    comb += self.fast_out.ok.eq(1)
            with m.Case(MicrOp.OP_BCREG):
                xo9 = self.dec.FormXL.XO[9]  # 3.0B p38 top bit of XO
                xo5 = self.dec.FormXL.XO[5]  # 3.0B p38
                with m.If(xo9 & ~xo5):
                    # constant: CTR
                    comb += self.fast_out.data.eq(FastRegs.CTR)
                    comb += self.fast_out.ok.eq(1)

            # MFSPR move from SPRs
            with m.Case(MicrOp.OP_MFSPR):
                spr = Signal(10, reset_less=True)
                comb += spr.eq(decode_spr_num(self.dec.SPR))  # from XFX
                comb += sprmap.spr_i.eq(spr)
                comb += self.spr_out.eq(sprmap.spr_o)
                comb += self.fast_out.eq(sprmap.fast_o)

        return m


class DecodeAImm(Elaboratable):
    """DecodeA immediate from instruction

    decodes register RA, whether immediate-zero, implicit and
    explicit CSRs
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(In1Sel, reset_less=True)
        self.immz_out = Signal(reset_less=True)

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb

        # zero immediate requested
        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))):
            comb += self.immz_out.eq(1)

        return m


class DecodeB(Elaboratable):
    """DecodeB from instruction

    decodes register RB, different forms of immediate (signed, unsigned),
    and implicit SPRs.  register B is basically "lane 2" into the CompUnits.
    by industry-standard convention, "lane 2" is where fully-decoded
    immediates are muxed in.
    """

    def __init__(self, dec):
        self.dec = dec
        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")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        op = self.dec.op
        reg = self.reg_out

        # select Register B field
        with m.Switch(self.sel_in):
            with m.Case(In2Sel.RB):
                comb += reg.data.eq(self.dec.RB)
                comb += reg.ok.eq(1)
            with m.Case(In2Sel.RS):
                # for M-Form shiftrot
                comb += reg.data.eq(self.dec.RS)
                comb += reg.ok.eq(1)

        # decode SPR2 based on instruction type
        # BCREG implicitly uses LR or TAR for 2nd reg
        # CTR however is already in fast_spr1 *not* 2.
        with m.If(op.internal_op == MicrOp.OP_BCREG):
            xo9 = self.dec.FormXL.XO[9]  # 3.0B p38 top bit of XO
            xo5 = self.dec.FormXL.XO[5]  # 3.0B p38
            with m.If(~xo9):
                comb += self.fast_out.data.eq(FastRegs.LR)
                comb += self.fast_out.ok.eq(1)
            with m.Elif(xo5):
                comb += self.fast_out.data.eq(FastRegs.TAR)
                comb += self.fast_out.ok.eq(1)

        return m


class DecodeBImm(Elaboratable):
    """DecodeB immediate from instruction
    """
    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(In2Sel, reset_less=True)
        self.imm_out = Data(64, "imm_b")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb

        # select Register B Immediate
        with m.Switch(self.sel_in):
            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
                si = Signal(16, reset_less=True)
                comb += si.eq(self.dec.SI)
                comb += self.imm_out.data.eq(exts(si, 16, 64))
                comb += self.imm_out.ok.eq(1)
            with m.Case(In2Sel.CONST_SI_HI):  # sign-extended 16+16=32 bit
                si_hi = Signal(32, reset_less=True)
                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
                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
                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
                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
                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)
                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
                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
                comb += self.imm_out.data.eq(self.dec.SH32)
                comb += self.imm_out.ok.eq(1)

        return m


class DecodeC(Elaboratable):
    """DecodeC from instruction

    decodes register RC.  this is "lane 3" into some CompUnits (not many)
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(In3Sel, reset_less=True)
        self.insn_in = Signal(32, reset_less=True)
        self.reg_out = Data(5, "reg_c")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        op = self.dec.op
        reg = self.reg_out

        # select Register C field
        with m.Switch(self.sel_in):
            with m.Case(In3Sel.RB):
                # for M-Form shiftrot
                comb += reg.data.eq(self.dec.RB)
                comb += reg.ok.eq(1)
            with m.Case(In3Sel.RS):
                comb += reg.data.eq(self.dec.RS)
                comb += reg.ok.eq(1)

        return m


class DecodeOut(Elaboratable):
    """DecodeOut from instruction

    decodes output register RA, RT or SPR
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(OutSel, reset_less=True)
        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")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        m.submodules.sprmap = sprmap = SPRMap()
        op = self.dec.op
        reg = self.reg_out

        # select Register out field
        with m.Switch(self.sel_in):
            with m.Case(OutSel.RT):
                comb += reg.data.eq(self.dec.RT)
                comb += reg.ok.eq(1)
            with m.Case(OutSel.RA):
                comb += reg.data.eq(self.dec.RA)
                comb += reg.ok.eq(1)
            with m.Case(OutSel.SPR):
                spr = Signal(10, reset_less=True)
                comb += spr.eq(decode_spr_num(self.dec.SPR))  # from XFX
                # MFSPR move to SPRs - needs mapping
                with m.If(op.internal_op == MicrOp.OP_MTSPR):
                    comb += sprmap.spr_i.eq(spr)
                    comb += self.spr_out.eq(sprmap.spr_o)
                    comb += self.fast_out.eq(sprmap.fast_o)

        # determine Fast Reg
        with m.Switch(op.internal_op):

            # BC or BCREG: implicit register (CTR) NOTE: same in DecodeA
            with m.Case(MicrOp.OP_BC, MicrOp.OP_BCREG):
                with m.If(~self.dec.BO[2]):  # 3.0B p38 BO2=0, use CTR reg
                    # constant: CTR
                    comb += self.fast_out.data.eq(FastRegs.CTR)
                    comb += self.fast_out.ok.eq(1)

            # RFID 1st spr (fast)
            with m.Case(MicrOp.OP_RFID):
                comb += self.fast_out.data.eq(FastRegs.SRR0)  # constant: SRR0
                comb += self.fast_out.ok.eq(1)

        return m


class DecodeOut2(Elaboratable):
    """DecodeOut2 from instruction

    decodes output registers (2nd one).  note that RA is *implicit* below,
    which now causes problems with SVP64

    TODO: SVP64 is a little more complex, here.  svp64 allows extending
    by one more destination by having one more EXTRA field.  RA-as-src
    is not the same as RA-as-dest.  limited in that it's the same first
    5 bits (from the v3.0B opcode), but still kinda cool.  mostly used
    for operations that have src-as-dest: mostly this is LD/ST-with-update
    but there are others.
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(OutSel, reset_less=True)
        self.lk = Signal(reset_less=True)
        self.insn_in = Signal(32, reset_less=True)
        self.reg_out = Data(5, "reg_o2")
        self.fast_out = Data(3, "fast_o2")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        op = self.dec.op
        #m.submodules.svdec = svdec = SVP64RegExtra()

        # get the 5-bit reg data before svp64-munging it into 7-bit plus isvec
        #reg = Signal(5, reset_less=True)

        if hasattr(self.dec.op, "upd"):
            # update mode LD/ST uses read-reg A also as an output
            with m.If(self.dec.op.upd == LDSTMode.update):
                comb += self.reg_out.data.eq(self.dec.RA)
                comb += self.reg_out.ok.eq(1)

        # B, BC or BCREG: potential implicit register (LR) output
        # these give bl, bcl, bclrl, etc.
        with m.Switch(op.internal_op):

            # BC* implicit register (LR)
            with m.Case(MicrOp.OP_BC, MicrOp.OP_B, MicrOp.OP_BCREG):
                with m.If(self.lk):  # "link" mode
                    comb += self.fast_out.data.eq(FastRegs.LR)  # constant: LR
                    comb += self.fast_out.ok.eq(1)

            # RFID 2nd spr (fast)
            with m.Case(MicrOp.OP_RFID):
                comb += self.fast_out.data.eq(FastRegs.SRR1)  # constant: SRR1
                comb += self.fast_out.ok.eq(1)

        return m


class DecodeRC(Elaboratable):
    """DecodeRc from instruction

    decodes Record bit Rc
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(RC, reset_less=True)
        self.insn_in = Signal(32, reset_less=True)
        self.rc_out = Data(1, "rc")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb

        # select Record bit out field
        with m.Switch(self.sel_in):
            with m.Case(RC.RC):
                comb += self.rc_out.data.eq(self.dec.Rc)
                comb += self.rc_out.ok.eq(1)
            with m.Case(RC.ONE):
                comb += self.rc_out.data.eq(1)
                comb += self.rc_out.ok.eq(1)
            with m.Case(RC.NONE):
                comb += self.rc_out.data.eq(0)
                comb += self.rc_out.ok.eq(1)

        return m


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.
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(RC, 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.dec.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.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)

        return m


class DecodeCRIn(Elaboratable):
    """Decodes input CR from instruction

    CR indices - insn fields - (not the data *in* the CR) require only 3
    bits because they refer to CR0-CR7
    """

    def __init__(self, dec):
        self.dec = dec
        self.sel_in = Signal(CRInSel, reset_less=True)
        self.insn_in = Signal(32, reset_less=True)
        self.cr_bitfield = Data(3, "cr_bitfield")
        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")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        op = self.dec.op
        m.submodules.ppick = ppick = PriorityPicker(8, reverse_i=True,
                                                       reverse_o=True)

        # zero-initialisation
        comb += self.cr_bitfield.ok.eq(0)
        comb += self.cr_bitfield_b.ok.eq(0)
        comb += self.cr_bitfield_o.ok.eq(0)
        comb += self.whole_reg.ok.eq(0)

        # select the relevant CR bitfields
        with m.Switch(self.sel_in):
            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.ok.eq(1)
            with m.Case(CRInSel.BI):
                comb += self.cr_bitfield.data.eq(self.dec.BI[2:5])
                comb += self.cr_bitfield.ok.eq(1)
            with m.Case(CRInSel.BFA):
                comb += self.cr_bitfield.data.eq(self.dec.FormX.BFA)
                comb += self.cr_bitfield.ok.eq(1)
            with m.Case(CRInSel.BA_BB):
                comb += self.cr_bitfield.data.eq(self.dec.BA[2:5])
                comb += self.cr_bitfield.ok.eq(1)
                comb += self.cr_bitfield_b.data.eq(self.dec.BB[2:5])
                comb += self.cr_bitfield_b.ok.eq(1)
                comb += self.cr_bitfield_o.data.eq(self.dec.BT[2:5])
                comb += self.cr_bitfield_o.ok.eq(1)
            with m.Case(CRInSel.BC):
                comb += self.cr_bitfield.data.eq(self.dec.BC[2:5])
                comb += self.cr_bitfield.ok.eq(1)
            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
                with m.If((op.internal_op == MicrOp.OP_MFCR) & move_one):
                    # must one-hot the FXM field
                    comb += ppick.i.eq(self.dec.FXM)
                    comb += self.whole_reg.data.eq(ppick.o)
                with m.Else():
                    # otherwise use all of it
                    comb += self.whole_reg.data.eq(0xff)

        return m


class DecodeCROut(Elaboratable):
    """Decodes input CR from instruction

    CR indices - insn fields - (not the data *in* the CR) require only 3
    bits because they refer to CR0-CR7
    """

    def __init__(self, dec):
        self.dec = dec
        self.rc_in = Signal(reset_less=True)
        self.sel_in = Signal(CROutSel, reset_less=True)
        self.insn_in = Signal(32, reset_less=True)
        self.cr_bitfield = Data(3, "cr_bitfield")
        self.whole_reg = Data(8,  "cr_fxm")

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        op = self.dec.op
        m.submodules.ppick = ppick = PriorityPicker(8, reverse_i=True,
                                                       reverse_o=True)

        comb += self.cr_bitfield.ok.eq(0)
        comb += self.whole_reg.ok.eq(0)

        with m.Switch(self.sel_in):
            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.ok.eq(self.rc_in)  # only when RC=1
            with m.Case(CROutSel.BF):
                comb += self.cr_bitfield.data.eq(self.dec.FormX.BF)
                comb += self.cr_bitfield.ok.eq(1)
            with m.Case(CROutSel.BT):
                comb += self.cr_bitfield.data.eq(self.dec.FormXL.BT[2:5])
                comb += self.cr_bitfield.ok.eq(1)
            with m.Case(CROutSel.WHOLE_REG):
                comb += self.whole_reg.ok.eq(1)
                move_one = Signal(reset_less=True)
                comb += move_one.eq(self.insn_in[20])
                with m.If((op.internal_op == MicrOp.OP_MTCRF)):
                    with m.If(move_one):
                        # must one-hot the FXM field
                        comb += ppick.i.eq(self.dec.FXM)
                        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
                    with m.Else():
                        comb += self.whole_reg.data.eq(self.dec.FXM)
                with m.Else():
                    # otherwise use all of it
                    comb += self.whole_reg.data.eq(0xff)

        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)
record_names = {'insn_type': 'internal_op',
                'fn_unit': 'function_unit',
                'rc': 'rc_sel',
                'oe': 'rc_sel',
                'zero_a': 'in1_sel',
                'imm_data': 'in2_sel',
                'invert_in': 'inv_a',
                'invert_out': 'inv_out',
                'rc': 'cr_out',
                'oe': 'cr_in',
                'output_carry': 'cry_out',
                'input_carry': 'cry_in',
                'is_32bit': 'is_32b',
                'is_signed': 'sgn',
                'lk': 'lk',
                'data_len': 'ldst_len',
                'byte_reverse': 'br',
                'sign_extend': 'sgn_ext',
                'ldst_mode': 'upd',
                }


class PowerDecodeSubset(Elaboratable):
    """PowerDecodeSubset: dynamic subset decoder

    only fields actually requested are copied over. hence, "subset" (duh).
    """
    def __init__(self, dec, opkls=None, fn_name=None, final=False, state=None):

        self.sv_rm = SVP64Rec(name="dec_svp64") # SVP64 RM field
        self.final = final
        self.opkls = opkls
        self.fn_name = fn_name
        if opkls is None:
            opkls = Decode2ToOperand
        self.do = opkls(fn_name)
        col_subset = self.get_col_subset(self.do)

        # only needed for "main" PowerDecode2
        if not self.final:
            self.e = Decode2ToExecute1Type(name=self.fn_name, do=self.do)

        # create decoder if one not already given
        if dec is None:
            dec = create_pdecode(name=fn_name, col_subset=col_subset,
                                      row_subset=self.rowsubsetfn)
        self.dec = dec

        # state information needed by the Decoder
        if state is None:
            state = CoreState("dec2")
        self.state = state

    def get_col_subset(self, do):
        subset = { 'cr_in', 'cr_out', 'rc_sel'} # needed, non-optional
        for k, v in record_names.items():
            if hasattr(do, k):
                subset.add(v)
        print ("get_col_subset", self.fn_name, do.fields, subset)
        return subset

    def rowsubsetfn(self, opcode, row):
        return row['unit'] == self.fn_name

    def ports(self):
        return self.dec.ports() + self.e.ports() + self.sv_rm.ports()

    def needs_field(self, field, op_field):
        if self.final:
            do = self.do
        else:
            do = self.e_tmp.do
        return hasattr(do, field) and self.op_get(op_field) is not None

    def do_copy(self, field, val, final=False):
        if final or self.final:
            do = self.do
        else:
            do = self.e_tmp.do
        if hasattr(do, field) and val is not None:
            return getattr(do, field).eq(val)
        return []

    def op_get(self, op_field):
        return getattr(self.dec.op, op_field, None)

    def elaborate(self, platform):
        m = Module()
        comb = m.d.comb
        state = self.state
        op, do = self.dec.op, self.do
        msr, cia = state.msr, state.pc

        # fill in for a normal instruction (not an exception)
        # copy over if non-exception, non-privileged etc. is detected
        if not self.final:
            if self.fn_name is None:
                name = "tmp"
            else:
                name = self.fn_name + "tmp"
            self.e_tmp = Decode2ToExecute1Type(name=name, opkls=self.opkls)

        # set up submodule decoders
        m.submodules.dec = self.dec
        m.submodules.dec_rc = dec_rc = DecodeRC(self.dec)
        m.submodules.dec_oe = dec_oe = DecodeOE(self.dec)
        m.submodules.dec_cr_in = self.dec_cr_in = DecodeCRIn(self.dec)
        m.submodules.dec_cr_out = self.dec_cr_out = DecodeCROut(self.dec)

        # copy instruction through...
        for i in [do.insn,
                  dec_rc.insn_in, dec_oe.insn_in,
                  self.dec_cr_in.insn_in, self.dec_cr_out.insn_in]:
            comb += i.eq(self.dec.opcode_in)

        # ...and subdecoders' input fields
        comb += dec_rc.sel_in.eq(op.rc_sel)
        comb += dec_oe.sel_in.eq(op.rc_sel)  # XXX should be OE sel
        comb += self.dec_cr_in.sel_in.eq(op.cr_in)
        comb += self.dec_cr_out.sel_in.eq(op.cr_out)
        comb += self.dec_cr_out.rc_in.eq(dec_rc.rc_out.data)

        # copy "state" over
        comb += self.do_copy("msr", msr)
        comb += self.do_copy("cia", cia)

        # set up instruction type
        # no op: defaults to OP_ILLEGAL
        if self.fn_name=="MMU":
            # mmu is special case: needs SPR opcode as well
            mmu0 = self.mmu0_spr_dec
            with m.If(((mmu0.dec.op.internal_op == MicrOp.OP_MTSPR) |
                       (mmu0.dec.op.internal_op == MicrOp.OP_MFSPR))):
                comb += self.do_copy("insn_type", mmu0.op_get("internal_op"))
            with m.Else():
                comb += self.do_copy("insn_type", self.op_get("internal_op"))
        else:
            comb += self.do_copy("insn_type", self.op_get("internal_op"))

        # function unit for decoded instruction: requires minor redirect
        # 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

        # XXX BUG - don't use hardcoded magic constants.
        # also use ".value" otherwise the test fails.  bit of a pain
        # https://bugs.libre-soc.org/show_bug.cgi?id=603

        SPR_PID   = 48  # TODO read docs for POWER9
        # Microwatt doesn't implement the partition table
        # instead has PRTBL register (SPR) to point to process table
        SPR_PRTBL = 720 # see common.vhdl in microwatt, not in POWER9
        with m.If(((self.dec.op.internal_op == MicrOp.OP_MTSPR) |
                   (self.dec.op.internal_op == MicrOp.OP_MFSPR)) &
                  ((spr == SPR.DSISR) | (spr == SPR.DAR)
                   | (spr==SPR_PRTBL) | (spr==SPR_PID))):
            comb += self.do_copy("fn_unit", Function.MMU)
        with m.Else():
            comb += self.do_copy("fn_unit",fn)

        # immediates
        if self.needs_field("zero_a", "in1_sel"):
            m.submodules.dec_ai = dec_ai = DecodeAImm(self.dec)
            comb += dec_ai.sel_in.eq(op.in1_sel)
            comb += self.do_copy("zero_a", dec_ai.immz_out)  # RA==0 detected
        if self.needs_field("imm_data", "in2_sel"):
            m.submodules.dec_bi = dec_bi = DecodeBImm(self.dec)
            comb += dec_bi.sel_in.eq(op.in2_sel)
            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)
        comb += self.do_copy("oe", dec_oe.oe_out)

        # CR in/out
        comb += self.do_copy("read_cr_whole", self.dec_cr_in.whole_reg)
        comb += self.do_copy("write_cr_whole", self.dec_cr_out.whole_reg)
        comb += self.do_copy("write_cr0", self.dec_cr_out.cr_bitfield.ok)

        comb += self.do_copy("input_cr", self.op_get("cr_in"))   # CR in
        comb += self.do_copy("output_cr", self.op_get("cr_out"))  # CR out

        # decoded/selected instruction flags
        comb += self.do_copy("data_len", self.op_get("ldst_len"))
        comb += self.do_copy("invert_in", self.op_get("inv_a"))
        comb += self.do_copy("invert_out", self.op_get("inv_out"))
        comb += self.do_copy("input_carry", self.op_get("cry_in"))
        comb += self.do_copy("output_carry", self.op_get("cry_out"))
        comb += self.do_copy("is_32bit", self.op_get("is_32b"))
        comb += self.do_copy("is_signed", self.op_get("sgn"))
        lk = self.op_get("lk")
        if lk is not None:
            with m.If(lk):
                comb += self.do_copy("lk", self.dec.LK)  # XXX TODO: accessor

        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

        return m


class PowerDecode2(PowerDecodeSubset):
    """PowerDecode2: the main instruction decoder.

    whilst PowerDecode is responsible for decoding the actual opcode, this
    module encapsulates further specialist, sparse information and
    expansion of fields that is inconvenient to have in the CSV files.
    for example: the encoding of the immediates, which are detected
    and expanded out to their full value from an annotated (enum)
    representation.

    implicit register usage is also set up, here.  for example: OP_BC
    requires implicitly reading CTR, OP_RFID requires implicitly writing
    to SRR1 and so on.

    in addition, PowerDecoder2 is responsible for detecting whether
    instructions are illegal (or privileged) or not, and instead of
    just leaving at that, *replacing* the instruction to execute with
    a suitable alternative (trap).

    LDSTExceptions are done the cycle _after_ they're detected (after
    they come out of LDSTCompUnit).  basically despite the instruction
    being decoded, the results of the decode are completely ignored
    and "exception.happened" used to set the "actual" instruction to
    "OP_TRAP".  the LDSTException data structure gets filled in,
    in the CompTrapOpSubset and that's what it fills in SRR.

    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.
    """

    def __init__(self, dec, opkls=None, fn_name=None, final=False, state=None):
        super().__init__(dec, opkls, fn_name, final, state)
        self.exc = LDSTException("dec2_exc")

        self.cr_out_isvec = Signal(1, name="cr_out_isvec")
        self.cr_in_isvec = Signal(1, name="cr_in_isvec")
        self.cr_in_b_isvec = Signal(1, name="cr_in_b_isvec")
        self.cr_in_o_isvec = Signal(1, name="cr_in_o_isvec")
        self.in1_isvec = Signal(1, name="reg_a_isvec")
        self.in2_isvec = Signal(1, name="reg_b_isvec")
        self.in3_isvec = Signal(1, name="reg_c_isvec")
        self.o_isvec = Signal(1, name="reg_o_isvec")
        self.o2_isvec = Signal(1, name="reg_o2_isvec")
        self.no_out_vec = Signal(1, name="no_out_vec") # no outputs are vectors

    def get_col_subset(self, opkls):
        subset = super().get_col_subset(opkls)
        subset.add("asmcode")
        subset.add("in1_sel")
        subset.add("in2_sel")
        subset.add("in3_sel")
        subset.add("out_sel")
        subset.add("sv_in1")
        subset.add("sv_in2")
        subset.add("sv_in3")
        subset.add("sv_out")
        subset.add("sv_cr_in")
        subset.add("sv_cr_out")
        subset.add("SV_Etype")
        subset.add("SV_Ptype")
        subset.add("lk")
        subset.add("internal_op")
        subset.add("form")
        return subset

    def elaborate(self, platform):
        m = super().elaborate(platform)
        comb = m.d.comb
        state = self.state
        e_out, op, do_out = self.e, self.dec.op, self.e.do
        dec_spr, msr, cia, ext_irq = state.dec, state.msr, state.pc, state.eint
        e = self.e_tmp
        do = e.do

        # fill in for a normal instruction (not an exception)
        # copy over if non-exception, non-privileged etc. is detected

        # set up submodule decoders
        m.submodules.dec_a = dec_a = DecodeA(self.dec)
        m.submodules.dec_b = dec_b = DecodeB(self.dec)
        m.submodules.dec_c = dec_c = DecodeC(self.dec)
        m.submodules.dec_o = dec_o = DecodeOut(self.dec)
        m.submodules.dec_o2 = dec_o2 = DecodeOut2(self.dec)

        # and SVP64 Extra decoders
        m.submodules.crout_svdec = crout_svdec = SVP64CRExtra()
        m.submodules.crin_svdec = crin_svdec = SVP64CRExtra()
        m.submodules.crin_svdec_b = crin_svdec_b = SVP64CRExtra()
        m.submodules.crin_svdec_o = crin_svdec_o = SVP64CRExtra()
        m.submodules.in1_svdec = in1_svdec = SVP64RegExtra()
        m.submodules.in2_svdec = in2_svdec = SVP64RegExtra()
        m.submodules.in3_svdec = in3_svdec = SVP64RegExtra()
        m.submodules.o_svdec = o_svdec = SVP64RegExtra()
        m.submodules.o2_svdec = o2_svdec = SVP64RegExtra()

        # debug access to crout_svdec (used in get_pdecode_cr_out)
        self.crout_svdec = crout_svdec

        # get the 5-bit reg data before svp64-munging it into 7-bit plus isvec
        reg = Signal(5, reset_less=True)

        # copy instruction through...
        for i in [do.insn, dec_a.insn_in, dec_b.insn_in,
                  dec_c.insn_in, dec_o.insn_in, dec_o2.insn_in]:
            comb += i.eq(self.dec.opcode_in)

        # now do the SVP64 munging.  op.SV_Etype and op.sv_in1 comes from
        # PowerDecoder which in turn comes from LDST-RM*.csv and RM-*.csv
        # which in turn were auto-generated by sv_analysis.py
        extra = self.sv_rm.extra            # SVP64 extra bits 10:18

        #######
        # CR out
        comb += crout_svdec.idx.eq(op.sv_cr_out)  # SVP64 CR out
        comb += self.cr_out_isvec.eq(crout_svdec.isvec)

        #######
        # CR in - index selection slightly different due to shared CR field sigh
        cr_a_idx = Signal(SVEXTRA)
        cr_b_idx = Signal(SVEXTRA)

        # these change slightly, when decoding BA/BB.  really should have
        # their own separate CSV column: sv_cr_in1 and sv_cr_in2, but hey
        comb += cr_a_idx.eq(op.sv_cr_in)
        comb += cr_b_idx.eq(SVEXTRA.NONE)
        with m.If(op.sv_cr_in == SVEXTRA.Idx_1_2.value):
            comb += cr_a_idx.eq(SVEXTRA.Idx1)
            comb += cr_b_idx.eq(SVEXTRA.Idx2)

        comb += self.cr_in_isvec.eq(crin_svdec.isvec)
        comb += self.cr_in_b_isvec.eq(crin_svdec_b.isvec)
        comb += self.cr_in_o_isvec.eq(crin_svdec_o.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(op.sv_cr_out) # SVP64 CR out

        # ...and subdecoders' input fields
        comb += dec_a.sel_in.eq(op.in1_sel)
        comb += dec_b.sel_in.eq(op.in2_sel)
        comb += dec_c.sel_in.eq(op.in3_sel)
        comb += dec_o.sel_in.eq(op.out_sel)
        comb += dec_o2.sel_in.eq(op.out_sel)
        if hasattr(do, "lk"):
            comb += dec_o2.lk.eq(do.lk)

        # get SVSTATE srcstep (TODO: elwidth, dststep etc.) needed below
        srcstep = Signal.like(self.state.svstate.srcstep)
        comb += srcstep.eq(self.state.svstate.srcstep)

        # registers a, b, c and out and out2 (LD/ST EA)
        for to_reg, fromreg, svdec in (
            (e.read_reg1, dec_a.reg_out, in1_svdec),
            (e.read_reg2, dec_b.reg_out, in2_svdec),
            (e.read_reg3, dec_c.reg_out, in3_svdec),
            (e.write_reg, dec_o.reg_out, o_svdec),
            (e.write_ea, dec_o2.reg_out, o2_svdec)):
            comb += svdec.extra.eq(extra)        # EXTRA field of SVP64 RM
            comb += svdec.etype.eq(op.SV_Etype)  # EXTRA2/3 for this insn
            comb += svdec.reg_in.eq(fromreg.data) # 3-bit (CR0/BC/BFA)
            comb += to_reg.ok.eq(fromreg.ok)
            # detect if Vectorised: add srcstep if yes.  TODO: a LOT.
            # this trick only holds when elwidth=default and in single-pred
            with m.If(svdec.isvec):
                comb += to_reg.data.eq(srcstep+svdec.reg_out) # 7-bit output
            with m.Else():
                comb += to_reg.data.eq(svdec.reg_out) # 7-bit output

        comb += in1_svdec.idx.eq(op.sv_in1)  # SVP64 reg #1 (matches in1_sel)
        comb += in2_svdec.idx.eq(op.sv_in2)  # SVP64 reg #2 (matches in2_sel)
        comb += in3_svdec.idx.eq(op.sv_in3)  # SVP64 reg #3 (matches in3_sel)
        comb += o_svdec.idx.eq(op.sv_out)    # SVP64 output (matches out_sel)
        # 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 += o2_svdec.idx.eq(op.sv_out)    # SVP64 output (implicit)

        # output reg-is-vectorised (and when no output is vectorised)
        comb += self.in1_isvec.eq(in1_svdec.isvec)
        comb += self.in2_isvec.eq(in2_svdec.isvec)
        comb += self.in3_isvec.eq(in3_svdec.isvec)
        comb += self.o_isvec.eq(o_svdec.isvec)
        comb += self.o2_isvec.eq(o2_svdec.isvec)
        # TODO: include SPRs and CRs here!  must be True when *all* are scalar
        comb += self.no_out_vec.eq((~o2_svdec.isvec) & (~o_svdec.isvec))

        # SPRs out
        comb += e.read_spr1.eq(dec_a.spr_out)
        comb += e.write_spr.eq(dec_o.spr_out)

        # Fast regs out
        comb += e.read_fast1.eq(dec_a.fast_out)
        comb += e.read_fast2.eq(dec_b.fast_out)
        comb += e.write_fast1.eq(dec_o.fast_out)
        comb += e.write_fast2.eq(dec_o2.fast_out)

        # condition registers (CR)
        for to_reg, fromreg, svdec in (
            (e.read_cr1, self.dec_cr_in.cr_bitfield, crin_svdec),
            (e.read_cr2, self.dec_cr_in.cr_bitfield_b, crin_svdec_b),
            (e.read_cr3, self.dec_cr_in.cr_bitfield_o, crin_svdec_o),
            (e.write_cr, self.dec_cr_out.cr_bitfield, crout_svdec)):
            comb += svdec.extra.eq(extra)        # EXTRA field of SVP64 RM
            comb += svdec.etype.eq(op.SV_Etype)  # EXTRA2/3 for this insn
            comb += svdec.cr_in.eq(fromreg.data) # 3-bit (CR0/BC/BFA)
            with m.If(svdec.isvec):
                comb += to_reg.data.eq(srcstep+svdec.cr_out) # 7-bit output
            with m.Else():
                comb += to_reg.data.eq(svdec.cr_out) # 7-bit output
            comb += to_reg.ok.eq(fromreg.ok)

        # 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
        with m.If(op.internal_op == MicrOp.OP_MFSPR):
            comb += e.xer_in.eq(0b111) # SO, CA, OV
        with m.If(op.internal_op == MicrOp.OP_CMP):
            comb += e.xer_in.eq(1<<XERRegs.SO) # SO
        with m.If(op.internal_op == MicrOp.OP_MTSPR):
            comb += e.xer_out.eq(1)

        # set the trapaddr to 0x700 for a td/tw/tdi/twi operation
        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

        ####################
        # ok so the instruction's been decoded, blah blah, however
        # now we need to determine if it's actually going to go ahead...
        # *or* if in fact it's a privileged operation, whether there's
        # an external interrupt, etc. etc.  this is a simple priority
        # if-elif-elif sequence.  decrement takes highest priority,
        # EINT next highest, privileged operation third.

        # check if instruction is privileged
        is_priv_insn = instr_is_priv(m, op.internal_op, e.do.insn)

        # different IRQ conditions
        ext_irq_ok = Signal()
        dec_irq_ok = Signal()
        priv_ok = Signal()
        illeg_ok = Signal()
        exc = self.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 += priv_ok.eq(is_priv_insn & msr[MSR.PR])
        comb += illeg_ok.eq(op.internal_op == MicrOp.OP_ILLEGAL)

        # LD/ST exceptions.  TestIssuer copies the exception info at us
        # after a failed LD/ST.
        with m.If(exc.happened):
            with m.If(exc.alignment):
                self.trap(m, TT.PRIV, 0x600)
            with m.Elif(exc.instr_fault):
                with m.If(exc.segment_fault):
                    self.trap(m, TT.PRIV, 0x480)
                with m.Else():
                    # pass exception info to trap to create SRR1
                    self.trap(m, TT.MEMEXC, 0x400, exc)
            with m.Else():
                with m.If(exc.segment_fault):
                    self.trap(m, TT.PRIV, 0x380)
                with m.Else():
                    self.trap(m, TT.PRIV, 0x300)

        # decrement counter (v3.0B p1099): TODO 32-bit version (MSR.LPCR)
        with m.Elif(dec_irq_ok):
            self.trap(m, TT.DEC, 0x900)   # v3.0B 6.5 p1065

        # external interrupt? only if MSR.EE set
        with m.Elif(ext_irq_ok):
            self.trap(m, TT.EINT, 0x500)

        # privileged instruction trap
        with m.Elif(priv_ok):
            self.trap(m, TT.PRIV, 0x700)

        # illegal instruction must redirect to trap. this is done by
        # *overwriting* the decoded instruction and starting again.
        # (note: the same goes for interrupts and for privileged operations,
        # just with different trapaddr and traptype)
        with m.Elif(illeg_ok):
            # illegal instruction trap
            self.trap(m, TT.ILLEG, 0x700)

        # no exception, just copy things to the output
        with m.Else():
            comb += e_out.eq(e)

        ####################
        # follow-up after trap/irq to set up SRR0/1

        # trap: (note e.insn_type so this includes OP_ILLEGAL) set up fast regs
        # Note: OP_SC could actually be modified to just be a trap
        with m.If((do_out.insn_type == MicrOp.OP_TRAP) |
                  (do_out.insn_type == MicrOp.OP_SC)):
            # TRAP write fast1 = SRR0
            comb += e_out.write_fast1.data.eq(FastRegs.SRR0)  # constant: SRR0
            comb += e_out.write_fast1.ok.eq(1)
            # TRAP write fast2 = SRR1
            comb += e_out.write_fast2.data.eq(FastRegs.SRR1)  # constant: SRR1
            comb += e_out.write_fast2.ok.eq(1)

        # RFID: needs to read SRR0/1
        with m.If(do_out.insn_type == MicrOp.OP_RFID):
            # TRAP read fast1 = SRR0
            comb += e_out.read_fast1.data.eq(FastRegs.SRR0)  # constant: SRR0
            comb += e_out.read_fast1.ok.eq(1)
            # TRAP read fast2 = SRR1
            comb += e_out.read_fast2.data.eq(FastRegs.SRR1)  # constant: SRR1
            comb += e_out.read_fast2.ok.eq(1)

        # annoying simulator bug
        if hasattr(e_out, "asmcode") and hasattr(self.dec.op, "asmcode"):
            comb += e_out.asmcode.eq(self.dec.op.asmcode)

        return m

    def trap(self, m, traptype, trapaddr, exc=None):
        """trap: this basically "rewrites" the decoded instruction as a trap
        """
        comb = m.d.comb
        op, e = self.dec.op, self.e
        comb += e.eq(0)  # reset eeeeeverything

        # start again
        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("traptype", traptype, True)  # request type
        comb += self.do_copy("ldst_exc", exc, True)  # request type
        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"


# SVP64 Prefix fields: see https://libre-soc.org/openpower/sv/svp64/
# identifies if an instruction is a SVP64-encoded prefix, and extracts
# the 24-bit SVP64 context (RM) if it is
class SVP64PrefixDecoder(Elaboratable):

    def __init__(self):
        self.opcode_in = Signal(32, reset_less=True)
        self.raw_opcode_in = Signal.like(self.opcode_in, reset_less=True)
        self.is_svp64_mode = Signal(1, reset_less=True)
        self.svp64_rm = Signal(24, reset_less=True)
        self.bigendian = Signal(reset_less=True)

    def elaborate(self, platform):
        m = Module()
        opcode_in = self.opcode_in
        comb = m.d.comb
        # sigh copied this from TopPowerDecoder
        # raw opcode in assumed to be in LE order: byte-reverse it to get BE
        raw_le = self.raw_opcode_in
        l = []
        for i in range(0, 32, 8):
            l.append(raw_le[i:i+8])
        l.reverse()
        raw_be = Cat(*l)
        comb += opcode_in.eq(Mux(self.bigendian, raw_be, raw_le))

        # start identifying if the incoming opcode is SVP64 prefix)
        major = Signal(6, reset_less=True)
        ident = Signal(2, reset_less=True)

        comb += major.eq(sel(opcode_in, SVP64P.OPC))
        comb += ident.eq(sel(opcode_in, SVP64P.SVP64_7_9))

        comb += self.is_svp64_mode.eq(
            (major == Const(1, 6)) &   # EXT01
            (ident == Const(0b11, 2))  # identifier bits
        )

        with m.If(self.is_svp64_mode):
            # now grab the 24-bit ReMap context bits,
            comb += self.svp64_rm.eq(sel(opcode_in, SVP64P.RM))

        return m

    def ports(self):
        return [self.opcode_in, self.raw_opcode_in, self.is_svp64_mode,
                self.svp64_rm, self.bigendian]

def get_rdflags(e, cu):
    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)
    print("rdflags", rdl)
    return Cat(*rdl)


if __name__ == '__main__':
    svp64 = SVP64PowerDecoder()
    vl = rtlil.convert(svp64, ports=svp64.ports())
    with open("svp64_dec.il", "w") as f:
        f.write(vl)
    pdecode = create_pdecode()
    dec2 = PowerDecode2(pdecode)
    vl = rtlil.convert(dec2, ports=dec2.ports() + pdecode.ports())
    with open("dec2.il", "w") as f:
        f.write(vl)