from soc.decoder.power_enums import (MicrOp, CryIn, Function,
CRInSel, CROutSel,
LdstLen, In1Sel, In2Sel, In3Sel,
- OutSel, SPR, RC, LDSTMode)
+ 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
from soc.regfile.regfiles import FastRegs
return m
+class SVP64RegExtra(Elaboratable):
+ """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):
+ self.extra = Signal(10, 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.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 = Module()
+ comb = m.d.comb
+
+ # first get the spec. if not changed it's "scalar identity behaviour"
+ # which is zero which is ok.
+ spec = Signal(3)
+
+ # 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 how
+ 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
+ comb += spec[1:3].eq(self.extra[0:2])
+ with m.Case(SVEXTRA.Idx1): # 2nd 2 bits
+ comb += spec[1:3].eq(self.extra[2:4])
+ with m.Case(SVEXTRA.Idx2): # 3rd 2 bits
+ comb += spec[1:3].eq(self.extra[4:6])
+ with m.Case(SVEXTRA.Idx3): # 4th 2 bits
+ comb += spec[1:3].eq(self.extra[6:8])
+ # 3-bit index selection mode
+ with m.Case(SVEtype.EXTRA3):
+ with m.Switch(self.idx):
+ with m.Case(SVEXTRA.Idx0): # 1st 3 bits
+ comb += spec.eq(self.extra[0:3])
+ with m.Case(SVEXTRA.Idx1): # 2nd 3 bits
+ comb += spec.eq(self.extra[3:6])
+ with m.Case(SVEXTRA.Idx2): # 3rd 3 bits
+ comb += spec.eq(self.extra[6:9])
+ # cannot fit more than 9 bits so there is no 4th thing
+
+ # now decode it. bit 2 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[2])
+
+ # decode vector differently from scalar
+ with m.If(self.isvec):
+ # Vector: shifted up, extra in LSBs (RA << 2) | spec[0:1]
+ comb += self.reg_out.eq(Cat(spec[:2], self.reg_in))
+ with m.Else():
+ # Scalar: not shifted up, extra in MSBs RA | (spec[0:1] << 5)
+ comb += self.reg_out.eq(Cat(self.reg_in, spec[:2]))
+
+ return m
+
+
class DecodeA(Elaboratable):
"""DecodeA from instruction
def __init__(self, dec):
self.dec = dec
+ self.sv_rm = SVP64Rec() # SVP64 RM field
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.reg_out = Data(7, name="reg_a")
+ self.reg_isvec = Signal(1, name="reg_a_isvec") # TODO: in reg_out
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
m.submodules.sprmap = sprmap = SPRMap()
+ 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)
# select Register A field
ra = Signal(5, reset_less=True)
with m.If((self.sel_in == In1Sel.RA) |
((self.sel_in == In1Sel.RA_OR_ZERO) &
(ra != Const(0, 5)))):
- comb += self.reg_out.data.eq(ra)
+ comb += reg.eq(ra)
comb += self.reg_out.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 += self.reg_out.data.eq(self.dec.RS)
+ comb += reg.eq(rs)
comb += self.reg_out.ok.eq(1)
+ # 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
+ comb += svdec.extra.eq(extra) # EXTRA field of SVP64 RM
+ comb += svdec.etype.eq(op.SV_Etype) # EXTRA2/3 for this insn
+ comb += svdec.idx.eq(op.sv_in1) # SVP64 reg #1 (matches in1_sel)
+ comb += svdec.reg_in.eq(reg) # 5-bit (RA, RS)
+
+ # outputs: 7-bit reg number and whether it's vectorised
+ comb += self.reg_out.data.eq(svdec.reg_out)
+ comb += self.reg_isvec.eq(svdec.isvec)
+
# decode Fast-SPR based on instruction type
- op = self.dec.op
with m.Switch(op.internal_op):
# BC or BCREG: implicit register (CTR) NOTE: same in DecodeOut
def __init__(self, dec):
self.dec = dec
+ self.sv_rm = SVP64Rec() # SVP64 RM field
self.sel_in = Signal(In2Sel, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
- self.reg_out = Data(5, "reg_b")
+ 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
+ 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)
# select Register B field
with m.Switch(self.sel_in):
with m.Case(In2Sel.RB):
- comb += self.reg_out.data.eq(self.dec.RB)
+ comb += reg.eq(self.dec.RB)
comb += self.reg_out.ok.eq(1)
with m.Case(In2Sel.RS):
# for M-Form shiftrot
- comb += self.reg_out.data.eq(self.dec.RS)
+ comb += reg.eq(self.dec.RS)
comb += self.reg_out.ok.eq(1)
+ # now do the SVP64 munging. different from DecodeA only by sv_in2
+
+ extra = self.sv_rm.extra # SVP64 extra bits 10:18
+ comb += svdec.extra.eq(extra) # EXTRA field of SVP64 RM
+ comb += svdec.etype.eq(op.SV_Etype) # EXTRA2/3 for this insn
+ comb += svdec.idx.eq(op.sv_in2) # SVP64 reg #2 (matches in2_sel)
+ comb += svdec.reg_in.eq(reg) # 5-bit (RA, RS)
+
+ # outputs: 7-bit reg number and whether it's vectorised
+ comb += self.reg_out.data.eq(svdec.reg_out)
+ comb += self.reg_isvec.eq(svdec.isvec)
+
# decode SPR2 based on instruction type
- op = self.dec.op
# 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):
def __init__(self, dec):
self.dec = dec
+ self.sv_rm = SVP64Rec() # SVP64 RM field
self.sel_in = Signal(In3Sel, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
- self.reg_out = Data(5, "reg_c")
+ self.reg_out = Data(7, "reg_c")
+ self.reg_isvec = Signal(1, name="reg_c_isvec") # TODO: in reg_out
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)
# select Register C field
with m.Switch(self.sel_in):
with m.Case(In3Sel.RB):
# for M-Form shiftrot
- comb += self.reg_out.data.eq(self.dec.RB)
+ comb += reg.eq(self.dec.RB)
comb += self.reg_out.ok.eq(1)
with m.Case(In3Sel.RS):
- comb += self.reg_out.data.eq(self.dec.RS)
+ comb += reg.eq(self.dec.RS)
comb += self.reg_out.ok.eq(1)
+ # now do the SVP64 munging. different from DecodeA only by sv_in3
+
+ extra = self.sv_rm.extra # SVP64 extra bits 10:18
+ comb += svdec.extra.eq(extra) # EXTRA field of SVP64 RM
+ comb += svdec.etype.eq(op.SV_Etype) # EXTRA2/3 for this insn
+ comb += svdec.idx.eq(op.sv_in3) # SVP64 reg #3 (matches in3_sel)
+ comb += svdec.reg_in.eq(reg) # 5-bit (RA, RS)
+
+ # outputs: 7-bit reg number and whether it's vectorised
+ comb += self.reg_out.data.eq(svdec.reg_out)
+ comb += self.reg_isvec.eq(svdec.isvec)
+
return m
def __init__(self, dec):
self.dec = dec
+ self.sv_rm = SVP64Rec() # SVP64 RM field
self.sel_in = Signal(OutSel, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
- self.reg_out = Data(5, "reg_o")
+ self.reg_out = Data(7, "reg_o")
+ self.reg_isvec = Signal(1, name="reg_o_isvec") # TODO: in reg_out
self.spr_out = Data(SPR, "spr_o")
self.fast_out = Data(3, "fast_o")
comb = m.d.comb
m.submodules.sprmap = sprmap = SPRMap()
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)
# select Register out field
with m.Switch(self.sel_in):
with m.Case(OutSel.RT):
- comb += self.reg_out.data.eq(self.dec.RT)
+ comb += reg.eq(self.dec.RT)
comb += self.reg_out.ok.eq(1)
with m.Case(OutSel.RA):
- comb += self.reg_out.data.eq(self.dec.RA)
+ comb += reg.eq(self.dec.RA)
comb += self.reg_out.ok.eq(1)
with m.Case(OutSel.SPR):
spr = Signal(10, reset_less=True)
comb += self.spr_out.eq(sprmap.spr_o)
comb += self.fast_out.eq(sprmap.fast_o)
+ # now do the SVP64 munging. different from DecodeA only by sv_out
+
+ extra = self.sv_rm.extra # SVP64 extra bits 10:18
+ comb += svdec.extra.eq(extra) # EXTRA field of SVP64 RM
+ comb += svdec.etype.eq(op.SV_Etype) # EXTRA2/3 for this insn
+ comb += svdec.idx.eq(op.sv_out) # SVP64 reg out1 (matches out_sel)
+ comb += svdec.reg_in.eq(reg) # 5-bit (RA, RS)
+
+ # outputs: 7-bit reg number and whether it's vectorised
+ comb += self.reg_out.data.eq(svdec.reg_out)
+ comb += self.reg_isvec.eq(svdec.isvec)
+
+ # determine Fast Reg
with m.Switch(op.internal_op):
# BC or BCREG: implicit register (CTR) NOTE: same in DecodeA
class DecodeOut2(Elaboratable):
"""DecodeOut2 from instruction
- decodes output registers
+ 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.sv_rm = SVP64Rec() # SVP64 RM field
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_o")
- self.fast_out = Data(3, "fast_o")
+ self.reg_out = Data(7, "reg_o2")
+ #self.reg_isvec = Signal(1, name="reg_o2_isvec") # TODO: in reg_out
+ 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.eq(self.dec.RA)
+ 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.
- op = self.dec.op
with m.Switch(op.internal_op):
# BC* implicit register (LR)
def __init__(self, dec):
self.dec = dec
+ self.sv_rm = SVP64Rec() # SVP64 RM field
self.sel_in = Signal(CRInSel, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
self.cr_bitfield = Data(3, "cr_bitfield")
def __init__(self, dec):
self.dec = dec
+ self.sv_rm = SVP64Rec() # SVP64 RM field
self.rc_in = Signal(reset_less=True)
self.sel_in = Signal(CROutSel, reset_less=True)
self.insn_in = Signal(32, reset_less=True)
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)
- col_subset = self.get_col_subset(self.do)
# create decoder if one not already given
if dec is None:
return row['unit'] == self.fn_name
def ports(self):
- return self.dec.ports() + self.e.ports()
+ return self.dec.ports() + self.e.ports() + self.sv_rm.ports()
def needs_field(self, field, op_field):
if self.final:
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_in.sv_rm.eq(self.sv_rm)
+ comb += self.dec_cr_out.sv_rm.eq(self.sv_rm)
comb += self.dec_cr_out.sel_in.eq(op.cr_out)
comb += self.dec_cr_out.rc_in.eq(dec_rc.rc_out.data)
comb += self.do_copy("msr", msr)
comb += self.do_copy("cia", cia)
- # set up instruction, pick fn unit
+ # set up instruction type
# no op: defaults to OP_ILLEGAL
comb += self.do_copy("insn_type", self.op_get("internal_op"))
- comb += self.do_copy("fn_unit", self.op_get("function_unit"))
+
+ # 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
+
+ # for first test only forward SPRs 18 and 19 to MMU, when
+ # operation is MTSPR or MFSPR. TODO: add other MMU SPRs
+ 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))):
+ 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"):
def get_col_subset(self, opkls):
subset = super().get_col_subset(opkls)
- subset.add("in1_sel")
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_Etype")
+ subset.add("SV_Ptype")
subset.add("lk")
subset.add("internal_op")
subset.add("form")
dec_c.insn_in, dec_o.insn_in, dec_o2.insn_in]:
comb += i.eq(self.dec.opcode_in)
+ # ... and svp64 rm
+ for i in [dec_a.insn_in, dec_b.insn_in,
+ dec_c.insn_in, dec_o.insn_in, dec_o2.insn_in]:
+ comb += i.eq(self.sv_rm)
+
# ...and subdecoders' input fields
comb += dec_a.sel_in.eq(op.in1_sel)
comb += dec_b.sel_in.eq(op.in2_sel)
comb += dec_o2.lk.eq(do.lk)
# registers a, b, c and out and out2 (LD/ST EA)
- comb += e.read_reg1.eq(dec_a.reg_out)
- comb += e.read_reg2.eq(dec_b.reg_out)
- comb += e.read_reg3.eq(dec_c.reg_out)
- comb += e.write_reg.eq(dec_o.reg_out)
- comb += e.write_ea.eq(dec_o2.reg_out)
+ for to_reg, fromreg in (
+ (e.read_reg1, dec_a.reg_out),
+ (e.read_reg2, dec_b.reg_out),
+ (e.read_reg3, dec_c.reg_out),
+ (e.write_reg, dec_o.reg_out),
+ (e.write_ea, dec_o2.reg_out)):
+ comb += to_reg.data.eq(fromreg.data)
+ comb += to_reg.ok.eq(fromreg.ok)
# SPRs out
comb += e.read_spr1.eq(dec_a.spr_out)
projects / soc.git / blobdiff