adjust SVP64RM class to output more PowerDecoder-friendly csv augmentation
[soc.git] / src / soc / decoder / power_decoder.py
index 01af735293c22599b89899938e1d863bbffa754d..6712cd1d63fc5f518776423ecff2d3fbf1470270 100644 (file)
@@ -1,5 +1,10 @@
 """Cascading Power ISA Decoder
 
+License: LGPLv3+
+
+# Copyright (C) 2020 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
+# Copyright (C) 2020 Michael Nolan <mtnolan2640@gmail.com>
+
 This module uses CSV tables in a hierarchical/peer cascading fashion,
 to create a multi-level instruction decoder by recognising appropriate
 patterns.  The output is a wide, flattened (1-level) series of bitfields,
@@ -33,6 +38,7 @@ this includes specifying the information sufficient to perform subdecoding.
 create_pdecode()
 
     the full hierarchical tree for decoding POWER9 is specified here
+    subsetting is possible by specifying col_subset (row_subset TODO)
 
 PowerDecoder
 
@@ -78,66 +84,116 @@ Top Level:
     ]
 
 
-Copyrights:
-# Copyright (C) 2020 Michael Nolan <mtnolan2640@gmail.com>
 """
 
+import gc
 from collections import namedtuple
 from nmigen import Module, Elaboratable, Signal, Cat, Mux
 from nmigen.cli import rtlil
 from soc.decoder.power_enums import (Function, Form, MicrOp,
                                      In1Sel, In2Sel, In3Sel, OutSel,
-                                     RC, LdstLen, LDSTMode, CryIn, get_csv,
+                                     SVEXTRA, SVEtype, SVPtype, # Simple-V
+                                     RC, LdstLen, LDSTMode, CryIn,
                                      single_bit_flags, CRInSel,
                                      CROutSel, get_signal_name,
                                      default_values, insns, asmidx)
 from soc.decoder.power_fields import DecodeFields
 from soc.decoder.power_fieldsn import SigDecode, SignalBitRange
-
+from soc.decoder.power_svp64 import SVP64RM
 
 # key data structure in which the POWER decoder is specified,
 # in a hierarchical fashion
-Subdecoder = namedtuple("Subdecoder",
-                        ["pattern",    # the major pattern to search for (e.g. major opcode)
-                         "opcodes",    # a dictionary of minor patterns to find
-                         "opint",      # true => the pattern must not be in "10----11" format
-                         # the bits (as a range) against which "pattern" matches
-                         "bitsel",
-                         "suffix",     # shift the opcode down before decoding
-                         "subdecoders"  # list of further subdecoders for *additional* matches,
-                         # *ONLY* after "pattern" has *ALSO* been matched against.
-                         ])
+Subdecoder = namedtuple(  # fix autoformatter
+    "Subdecoder",
+    ["pattern",    # the major pattern to search for (e.g. major opcode)
+     "opcodes",    # a dictionary of minor patterns to find
+     "opint",      # true => the pattern must not be in "10----11" format
+     # the bits (as a range) against which "pattern" matches
+     "bitsel",
+     "suffix",     # shift the opcode down before decoding
+     "subdecoders"  # list of further subdecoders for *additional* matches,
+     # *ONLY* after "pattern" has *ALSO* been matched against.
+     ])
+
+power_op_types = {'function_unit': Function,
+                  'internal_op': MicrOp,
+                  'form': Form,
+                  'asmcode': 8,
+                  'SV_Etype': SVEtype,
+                  'SV_Ptype': SVPtype,
+                  'in1_sel': In1Sel,
+                  'in2_sel': In2Sel,
+                  'in3_sel': In3Sel,
+                  'out_sel': OutSel,
+                  'cr_in': CRInSel,
+                  'cr_out': CROutSel,
+                  'sv_in1': SVEXTRA,
+                  'sv_in2': SVEXTRA,
+                  'sv_in3': SVEXTRA,
+                  'sv_out': SVEXTRA,
+                  'sv_cr_in': SVEXTRA,
+                  'sv_cr_out': SVEXTRA,
+                  'ldst_len': LdstLen,
+                  'upd': LDSTMode,
+                  'rc_sel': RC,
+                  'cry_in': CryIn
+                  }
+
+power_op_csvmap = {'function_unit': 'unit',
+                   'form': 'form',
+                   'internal_op': 'internal op',
+                   'in1_sel': 'in1',
+                   'in2_sel': 'in2',
+                   'in3_sel': 'in3',
+                   'out_sel': 'out',
+                   'sv_in1': 'sv_in1',
+                   'sv_in2': 'sv_in2',
+                   'sv_in3': 'sv_in3',
+                   'sv_out': 'sv_out',
+                   'sv_cr_in': 'sv_cr_in',
+                   'sv_cr_out': 'sv_cr_out',
+                   'cr_in': 'CR in',
+                   'cr_out': 'CR out',
+                   'ldst_len': 'ldst len',
+                   'upd': 'upd',
+                   'rc_sel': 'rc',
+                   'cry_in': 'cry in',
+                   }
+
+
+def get_pname(field, pname):
+    if pname is None:
+        return field
+    return "%s_%s" % (pname, field)
 
 
 class PowerOp:
-    """PowerOp: spec for execution.  op type (ADD etc.) reg specs etc.
-
-    this is an internal data structure, set up by reading CSV files
-    (which uses _eq to initialise each instance, not eq)
-
-    the "public" API (as far as actual usage as a useful decoder is concerned)
-    is Decode2ToExecute1Type
+    """PowerOp - a dynamic class that stores (subsets of) CSV rows of data
+    about a PowerISA instruction.  this is a "micro-code" expanded format
+    which generates an awful lot of wires, hence the subsetting
     """
 
-    def __init__(self, incl_asm=True):
-        self.function_unit = Signal(Function, reset_less=True)
-        self.internal_op = Signal(MicrOp, reset_less=True)
-        self.form = Signal(Form, reset_less=True)
-        if incl_asm:  # for simulator only
-            self.asmcode = Signal(8, reset_less=True)
-        self.in1_sel = Signal(In1Sel, reset_less=True)
-        self.in2_sel = Signal(In2Sel, reset_less=True)
-        self.in3_sel = Signal(In3Sel, reset_less=True)
-        self.out_sel = Signal(OutSel, reset_less=True)
-        self.cr_in = Signal(CRInSel, reset_less=True)
-        self.cr_out = Signal(CROutSel, reset_less=True)
-        self.ldst_len = Signal(LdstLen, reset_less=True)
-        self.upd = Signal(LDSTMode, reset_less=True)
-        self.rc_sel = Signal(RC, reset_less=True)
-        self.cry_in = Signal(CryIn, reset_less=True)
+    def __init__(self, incl_asm=True, name=None, subset=None):
+        self.subset = subset
+        debug_report = set()
+        fields = set()
+        for field, ptype in power_op_types.items():
+            fields.add(field)
+            if subset and field not in subset:
+                continue
+            fname = get_pname(field, name)
+            setattr(self, field, Signal(ptype, reset_less=True, name=fname))
+            debug_report.add(field)
         for bit in single_bit_flags:
-            name = get_signal_name(bit)
-            setattr(self, name, Signal(reset_less=True, name=name))
+            field = get_signal_name(bit)
+            fields.add(field)
+            if subset and field not in subset:
+                continue
+            debug_report.add(field)
+            fname = get_pname(field, name)
+            setattr(self, field, Signal(reset_less=True, name=fname))
+        print("PowerOp debug", name, debug_report)
+        print("        fields", fields)
 
     def _eq(self, row=None):
         if row is None:
@@ -157,84 +213,94 @@ class PowerOp:
             print(row)
         ldst_mode = row['upd']
         if ldst_mode.isdigit():
-            ldst_mode = LDSTMode(int(ldst_mode))
-        else:
-            ldst_mode = LDSTMode[ldst_mode]
-        res = [self.function_unit.eq(Function[row['unit']]),
-               self.form.eq(Form[row['form']]),
-               self.internal_op.eq(MicrOp[row['internal op']]),
-               self.in1_sel.eq(In1Sel[row['in1']]),
-               self.in2_sel.eq(In2Sel[row['in2']]),
-               self.in3_sel.eq(In3Sel[row['in3']]),
-               self.out_sel.eq(OutSel[row['out']]),
-               self.cr_in.eq(CRInSel[row['CR in']]),
-               self.cr_out.eq(CROutSel[row['CR out']]),
-               self.ldst_len.eq(LdstLen[row['ldst len']]),
-               self.upd.eq(ldst_mode),
-               self.rc_sel.eq(RC[row['rc']]),
-               self.cry_in.eq(CryIn[row['cry in']]),
-               ]
+            row['upd'] = int(ldst_mode)
+        res = []
+        for field, ptype in power_op_types.items():
+            if not hasattr(self, field):
+                continue
+            if field not in power_op_csvmap:
+                continue
+            csvname = power_op_csvmap[field]
+            print (field, ptype, csvname, row)
+            val = row[csvname]
+            if csvname == 'upd' and isinstance(val, int):  # LDSTMode different
+                val = ptype(val)
+            else:
+                val = ptype[val]
+            res.append(getattr(self, field).eq(val))
         if False:
             print(row.keys())
         asmcode = row['comment']
         if hasattr(self, "asmcode") and asmcode in asmidx:
             res.append(self.asmcode.eq(asmidx[asmcode]))
         for bit in single_bit_flags:
-            sig = getattr(self, get_signal_name(bit))
+            field = get_signal_name(bit)
+            if not hasattr(self, field):
+                continue
+            sig = getattr(self, field)
             res.append(sig.eq(int(row.get(bit, 0))))
         return res
 
+    def _get_eq(self, res, field, otherop):
+        copyfrom = getattr(otherop, field, None)
+        copyto = getattr(self, field, None)
+        if copyfrom is not None and copyto is not None:
+            res.append(copyto.eq(copyfrom))
+
     def eq(self, otherop):
-        res = [self.function_unit.eq(otherop.function_unit),
-               self.form.eq(otherop.form),
-               self.internal_op.eq(otherop.internal_op),
-               self.in1_sel.eq(otherop.in1_sel),
-               self.in2_sel.eq(otherop.in2_sel),
-               self.in3_sel.eq(otherop.in3_sel),
-               self.out_sel.eq(otherop.out_sel),
-               self.cr_in.eq(otherop.cr_in),
-               self.cr_out.eq(otherop.cr_out),
-               self.rc_sel.eq(otherop.rc_sel),
-               self.ldst_len.eq(otherop.ldst_len),
-               self.upd.eq(otherop.upd),
-               self.cry_in.eq(otherop.cry_in)]
+        res = []
+        for field in power_op_types.keys():
+            self._get_eq(res, field, otherop)
         for bit in single_bit_flags:
-            sig = getattr(self, get_signal_name(bit))
-            res.append(sig.eq(getattr(otherop, get_signal_name(bit))))
-        if hasattr(self, "asmcode"):
-            res.append(self.asmcode.eq(otherop.asmcode))
+            self._get_eq(res, get_signal_name(bit), otherop)
         return res
 
     def ports(self):
-        regular = [self.function_unit,
-                   self.in1_sel,
-                   self.in2_sel,
-                   self.in3_sel,
-                   self.out_sel,
-                   self.cr_in,
-                   self.cr_out,
-                   self.ldst_len,
-                   self.rc_sel,
-                   self.internal_op,
-                   self.form]
+        res = []
+        for field in power_op_types.keys():
+            if hasattr(self, field):
+                res.append(getattr(self, field))
         if hasattr(self, "asmcode"):
-            regular.append(self.asmcode)
-        single_bit_ports = [getattr(self, get_signal_name(x))
-                            for x in single_bit_flags]
-        return regular + single_bit_ports
+            res.append(self.asmcode)
+        for field in single_bit_flags:
+            field = get_signal_name(field)
+            if hasattr(self, field):
+                res.append(getattr(self, field))
+        return res
 
 
 class PowerDecoder(Elaboratable):
     """PowerDecoder - decodes an incoming opcode into the type of operation
+
+    this is a recursive algorithm, creating Switch statements that can
+    have further match-and-decode on other parts of the opcode field before
+    finally landing at a "this CSV entry details gets returned" thing.
+
+    the complicating factor is the row and col subsetting.  column subsetting
+    dynamically chooses only the CSV columns requested, whilst row subsetting
+    allows a function to be called on the row to determine if the Case
+    statement is to be generated for that row.  this not only generates
+    completely different Decoders, it also means that some sub-decoders
+    will turn up blank (empty switch statements).  if that happens we do
+    not want the parent to include a Mux for an entirely blank switch statement
+    so we have to store the switch/case statements in a tree, and
+    post-analyse it.
+
+    the reason for the tree is because elaborate can only be called *after*
+    the constructor is called.  all quite messy.
     """
 
-    def __init__(self, width, dec):
+    def __init__(self, width, dec, name=None, col_subset=None, row_subset=None):
+        self.actually_does_something = False
+        self.pname = name
+        self.col_subset = col_subset
+        self.row_subsetfn = row_subset
         if not isinstance(dec, list):
             dec = [dec]
         self.dec = dec
         self.opcode_in = Signal(width, reset_less=True)
 
-        self.op = PowerOp()
+        self.op = PowerOp(name=name, subset=col_subset)
         for d in dec:
             if d.suffix is not None and d.suffix >= width:
                 d.suffix = None
@@ -260,60 +326,121 @@ class PowerDecoder(Elaboratable):
             divided[key].append(r)
         return divided
 
-    def elaborate(self, platform):
-        m = Module()
-        comb = m.d.comb
-
-        # note: default opcode is "illegal" as this is a combinatorial block
-        # this only works because OP_ILLEGAL=0 and the default (unset) is 0
+    def tree_analyse(self):
+        self.decs = decs = []
+        self.submodules = submodules = {}
+        self.eqs = eqs = []
 
         # go through the list of CSV decoders first
         for d in self.dec:
+            cases = []
             opcode_switch = Signal(d.bitsel[1] - d.bitsel[0],
                                    reset_less=True)
-            comb += opcode_switch.eq(self.opcode_in[d.bitsel[0]:d.bitsel[1]])
+            eq = []
+            case_does_something = False
+            eq.append(opcode_switch.eq(
+                self.opcode_in[d.bitsel[0]:d.bitsel[1]]))
             if d.suffix:
                 opcodes = self.divide_opcodes(d)
                 opc_in = Signal(d.suffix, reset_less=True)
-                comb += opc_in.eq(opcode_switch[:d.suffix])
+                eq.append(opc_in.eq(opcode_switch[:d.suffix]))
                 # begin the dynamic Switch statement here
-                with m.Switch(opc_in):
-                    for key, row in opcodes.items():
-                        bitsel = (d.suffix+d.bitsel[0], d.bitsel[1])
-                        sd = Subdecoder(pattern=None, opcodes=row,
-                                        bitsel=bitsel, suffix=None,
-                                        opint=False, subdecoders=[])
-                        subdecoder = PowerDecoder(width=32, dec=sd)
-                        setattr(m.submodules, "dec_sub%d" % key, subdecoder)
-                        comb += subdecoder.opcode_in.eq(self.opcode_in)
-                        # add in the dynamic Case statement here
-                        with m.Case(key):
-                            comb += self.op.eq(subdecoder.op)
+                switch_case = {}
+                cases.append([opc_in, switch_case])
+                sub_eqs = []
+                for key, row in opcodes.items():
+                    bitsel = (d.suffix+d.bitsel[0], d.bitsel[1])
+                    sd = Subdecoder(pattern=None, opcodes=row,
+                                    bitsel=bitsel, suffix=None,
+                                    opint=False, subdecoders=[])
+                    mname = get_pname("dec_sub%d" % key, self.pname)
+                    subdecoder = PowerDecoder(width=32, dec=sd,
+                                              name=mname,
+                                              col_subset=self.col_subset,
+                                              row_subset=self.row_subsetfn)
+                    if not subdecoder.tree_analyse():
+                        del subdecoder
+                        continue
+                    submodules[mname] = subdecoder
+                    sub_eqs.append(subdecoder.opcode_in.eq(self.opcode_in))
+                    # add in the dynamic Case statement here
+                    switch_case[key] = self.op.eq(subdecoder.op)
+                    self.actually_does_something = True
+                    case_does_something = True
+                if case_does_something:
+                    eq += sub_eqs
             else:
                 # TODO: arguments, here (all of them) need to be a list.
                 # a for-loop around the *list* of decoder args.
-                with m.Switch(opcode_switch):
-                    self.handle_subdecoders(m, d)
-                    for row in d.opcodes:
-                        opcode = row['opcode']
-                        if d.opint and '-' not in opcode:
-                            opcode = int(opcode, 0)
-                        if not row['unit']:
+                switch_case = {}
+                cases.append([opcode_switch, switch_case])
+                seqs = self.handle_subdecoders(switch_case, submodules, d)
+                if seqs:
+                    case_does_something = True
+                eq += seqs
+                for row in d.opcodes:
+                    opcode = row['opcode']
+                    if d.opint and '-' not in opcode:
+                        opcode = int(opcode, 0)
+                    if not row['unit']:
+                        continue
+                    if self.row_subsetfn:
+                        if not self.row_subsetfn(opcode, row):
                             continue
-                        # add in the dynamic Case statement here
-                        with m.Case(opcode):
-                            comb += self.op._eq(row)
-        return m
+                    # add in the dynamic Case statement here
+                    switch_case[opcode] = self.op._eq(row)
+                    self.actually_does_something = True
+                    case_does_something = True
+
+            if cases:
+                decs.append(cases)
+            if case_does_something:
+                eqs += eq
+                print("submodule eqs", self.pname, eq)
+
+        print("submodules", self.pname, submodules)
+
+        gc.collect()
+        return self.actually_does_something
 
-    def handle_subdecoders(self, m, d):
+    def handle_subdecoders(self, switch_case, submodules, d):
+        eqs = []
         for dec in d.subdecoders:
-            subdecoder = PowerDecoder(self.width, dec)
             if isinstance(dec, list):  # XXX HACK: take first pattern
                 dec = dec[0]
-            setattr(m.submodules, "dec%d" % dec.pattern, subdecoder)
-            m.d.comb += subdecoder.opcode_in.eq(self.opcode_in)
-            with m.Case(dec.pattern):
-                m.d.comb += self.op.eq(subdecoder.op)
+            print("subdec", dec.pattern, self.pname)
+            mname = get_pname("dec%d" % dec.pattern, self.pname)
+            subdecoder = PowerDecoder(self.width, dec,
+                                      name=mname,
+                                      col_subset=self.col_subset,
+                                      row_subset=self.row_subsetfn)
+            if not subdecoder.tree_analyse():  # doesn't do anything
+                del subdecoder
+                continue                      # skip
+            submodules[mname] = subdecoder
+            eqs.append(subdecoder.opcode_in.eq(self.opcode_in))
+            switch_case[dec.pattern] = self.op.eq(subdecoder.op)
+            self.actually_does_something = True
+
+        return eqs
+
+    def elaborate(self, platform):
+        print("decoder elaborate", self.pname, self.submodules)
+        m = Module()
+        comb = m.d.comb
+
+        comb += self.eqs
+
+        for mname, subdecoder in self.submodules.items():
+            setattr(m.submodules, mname, subdecoder)
+
+        for switch_case in self.decs:
+            for (switch, cases) in switch_case:
+                with m.Switch(switch):
+                    for key, eqs in cases.items():
+                        with m.Case(key):
+                            comb += eqs
+        return m
 
     def ports(self):
         return [self.opcode_in] + self.op.ports()
@@ -327,19 +454,20 @@ class TopPowerDecoder(PowerDecoder):
     (reverses byte order).  See V3.0B p44 1.11.2
     """
 
-    def __init__(self, width, dec):
-        PowerDecoder.__init__(self, width, dec)
+    def __init__(self, width, dec, name=None, col_subset=None, row_subset=None):
+        PowerDecoder.__init__(self, width, dec, name, col_subset, row_subset)
         self.fields = df = DecodeFields(SignalBitRange, [self.opcode_in])
         self.fields.create_specs()
         self.raw_opcode_in = Signal.like(self.opcode_in, reset_less=True)
         self.bigendian = Signal(reset_less=True)
 
-        for name, value in self.fields.common_fields.items():
-            sig = Signal(value[0:-1].shape(), reset_less=True, name=name)
-            setattr(self, name, sig)
+        for fname, value in self.fields.common_fields.items():
+            signame = get_pname(fname, name)
+            sig = Signal(value[0:-1].shape(), reset_less=True, name=signame)
+            setattr(self, fname, sig)
 
         # create signals for all field forms
-        self.form_names = forms = self.fields.instrs.keys()
+        forms = self.form_names
         self.sigforms = {}
         for form in forms:
             fields = self.fields.instrs[form]
@@ -347,13 +475,19 @@ class TopPowerDecoder(PowerDecoder):
             Fields = namedtuple("Fields", fk)
             sf = {}
             for k, value in fields.items():
-                name = "%s_%s" % (form, k)
-                sig = Signal(value[0:-1].shape(), reset_less=True, name=name)
+                fname = "%s_%s" % (form, k)
+                sig = Signal(value[0:-1].shape(), reset_less=True, name=fname)
                 sf[k] = sig
             instr = Fields(**sf)
             setattr(self, "Form%s" % form, instr)
             self.sigforms[form] = instr
 
+        self.tree_analyse()
+
+    @property
+    def form_names(self):
+        return self.fields.instrs.keys()
+
     def elaborate(self, platform):
         m = PowerDecoder.elaborate(self, platform)
         comb = m.d.comb
@@ -367,8 +501,8 @@ class TopPowerDecoder(PowerDecoder):
         comb += self.opcode_in.eq(Mux(self.bigendian, raw_be, raw_le))
 
         # add all signal from commonly-used fields
-        for name, value in self.fields.common_fields.items():
-            sig = getattr(self, name)
+        for fname, value in self.fields.common_fields.items():
+            sig = getattr(self, fname)
             comb += sig.eq(value[0:-1])
 
         # link signals for all field forms
@@ -389,16 +523,25 @@ class TopPowerDecoder(PowerDecoder):
 ####################################################
 # PRIMARY FUNCTION SPECIFYING THE FULL POWER DECODER
 
-def create_pdecode():
+def create_pdecode(name=None, col_subset=None, row_subset=None):
     """create_pdecode - creates a cascading hierarchical POWER ISA decoder
+
+    subsetting of the PowerOp decoding is possible by setting col_subset
     """
 
+    # some alteration to the CSV files is required for SV so we use
+    # a class to do it
+    isa = SVP64RM()
+    get_csv = isa.get_svp64_csv
+
     # minor 19 has extra patterns
     m19 = []
     m19.append(Subdecoder(pattern=19, opcodes=get_csv("minor_19.csv"),
-                          opint=True, bitsel=(1, 11), suffix=None, subdecoders=[]))
+                          opint=True, bitsel=(1, 11), suffix=None,
+                          subdecoders=[]))
     m19.append(Subdecoder(pattern=19, opcodes=get_csv("minor_19_00000.csv"),
-                          opint=True, bitsel=(1, 6), suffix=None, subdecoders=[]))
+                          opint=True, bitsel=(1, 6), suffix=None,
+                          subdecoders=[]))
 
     # minor opcodes.
     pminor = [
@@ -422,10 +565,35 @@ def create_pdecode():
     dec.append(Subdecoder(pattern=None, opint=False, opcodes=opcodes,
                           bitsel=(0, 32), suffix=None, subdecoders=[]))
 
-    return TopPowerDecoder(32, dec)
+    return TopPowerDecoder(32, dec, name=name, col_subset=col_subset,
+                           row_subset=row_subset)
 
 
 if __name__ == '__main__':
+
+    if True:
+        # row subset
+
+        def rowsubsetfn(opcode, row):
+            print("row_subset", opcode, row)
+            return row['unit'] == 'ALU'
+
+        pdecode = create_pdecode(name="rowsub",
+                                 col_subset={'function_unit', 'in1_sel'},
+                                 row_subset=rowsubsetfn)
+        vl = rtlil.convert(pdecode, ports=pdecode.ports())
+        with open("row_subset_decoder.il", "w") as f:
+            f.write(vl)
+
+        # col subset
+
+        pdecode = create_pdecode(name="fusubset", col_subset={'function_unit'})
+        vl = rtlil.convert(pdecode, ports=pdecode.ports())
+        with open("col_subset_decoder.il", "w") as f:
+            f.write(vl)
+
+    # full decoder
+
     pdecode = create_pdecode()
     vl = rtlil.convert(pdecode, ports=pdecode.ports())
     with open("decoder.il", "w") as f: