X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fexperiment%2Fscore6600.py;h=05de08fa1de491d280b0c635a54b705024396a6a;hb=a7b8337867b15252dc1c63b7e3dc757d3449a6e1;hp=1c0caca0a74acf1f3139fd823bbcc18bcce78440;hpb=a878168fb22776aa7a76f617255f5d88952782e6;p=soc.git

diff --git a/src/experiment/score6600.py b/src/experiment/score6600.py
index 1c0caca0..05de08fa 100644
--- a/src/experiment/score6600.py
+++ b/src/experiment/score6600.py
@@ -1,62 +1,146 @@
 from nmigen.compat.sim import run_simulation
 from nmigen.cli import verilog, rtlil
-from nmigen import Module, Const, Signal, Array, Cat, Elaboratable
+from nmigen import Module, Const, Signal, Array, Cat, Elaboratable, Memory
 
 from regfile.regfile import RegFileArray, treereduce
-from scoreboard.fn_unit import IntFnUnit, FPFnUnit, LDFnUnit, STFnUnit
 from scoreboard.fu_fu_matrix import FUFUDepMatrix
 from scoreboard.fu_reg_matrix import FURegDepMatrix
 from scoreboard.global_pending import GlobalPending
 from scoreboard.group_picker import GroupPicker
-from scoreboard.issue_unit import IntFPIssueUnit, RegDecode
+from scoreboard.issue_unit import IssueUnitGroup, IssueUnitArray, RegDecode
+from scoreboard.shadow import ShadowMatrix, BranchSpeculationRecord
+from scoreboard.instruction_q import Instruction, InstructionQ
+from scoreboard.memfu import MemFunctionUnits
 
 from compalu import ComputationUnitNoDelay
+from compldst import LDSTCompUnit
 
-from alu_hier import ALU
+from alu_hier import ALU, BranchALU
 from nmutil.latch import SRLatch
+from nmutil.nmoperator import eq
 
-from random import randint
+from random import randint, seed
+from copy import deepcopy
+from math import log
 
-class CompUnits(Elaboratable):
 
-    def __init__(self, rwid, n_units):
+class TestMemory(Elaboratable):
+    def __init__(self, regwid, addrw):
+        self.ddepth = 1 # regwid //8
+        depth = (1<<addrw) // self.ddepth
+        self.adr   = Signal(addrw)
+        self.dat_r = Signal(regwid)
+        self.dat_w = Signal(regwid)
+        self.we    = Signal()
+        self.mem   = Memory(width=regwid, depth=depth, init=range(0, depth))
+
+    def elaborate(self, platform):
+        m = Module()
+        m.submodules.rdport = rdport = self.mem.read_port()
+        m.submodules.wrport = wrport = self.mem.write_port()
+        m.d.comb += [
+            rdport.addr.eq(self.adr[self.ddepth:]), # ignore low bits
+            self.dat_r.eq(rdport.data),
+            wrport.addr.eq(self.adr),
+            wrport.data.eq(self.dat_w),
+            wrport.en.eq(self.we),
+        ]
+        return m
+
+
+class MemSim:
+    def __init__(self, regwid, addrw):
+        self.regwid = regwid
+        self.ddepth = 1 # regwid//8
+        depth = (1<<addrw) // self.ddepth
+        self.mem = list(range(0, depth))
+
+    def ld(self, addr):
+        return self.mem[addr>>self.ddepth]
+
+    def st(self, addr, data):
+        self.mem[addr>>self.ddepth] = data & ((1<<self.regwid)-1)
+
+
+class CompUnitsBase(Elaboratable):
+    """ Computation Unit Base class.
+
+        Amazingly, this class works recursively.  It's supposed to just
+        look after some ALUs (that can handle the same operations),
+        grouping them together, however it turns out that the same code
+        can also group *groups* of Computation Units together as well.
+
+        Basically it was intended just to concatenate the ALU's issue,
+        go_rd etc. signals together, which start out as bits and become
+        sequences.  Turns out that the same trick works just as well
+        on Computation Units!
+
+        So this class may be used recursively to present a top-level
+        sequential concatenation of all the signals in and out of
+        ALUs, whilst at the same time making it convenient to group
+        ALUs together.
+
+        At the lower level, the intent is that groups of (identical)
+        ALUs may be passed the same operation.  Even beyond that,
+        the intent is that that group of (identical) ALUs actually
+        share the *same pipeline* and as such become a "Concurrent
+        Computation Unit" as defined by Mitch Alsup (see section
+        11.4.9.3)
+    """
+    def __init__(self, rwid, units, ldstmode=False):
         """ Inputs:
 
             * :rwid:   bit width of register file(s) - both FP and INT
-            * :n_units: number of ALUs
+            * :units: sequence of ALUs (or CompUnitsBase derivatives)
         """
-        self.n_units = n_units
+        self.units = units
+        self.ldstmode = ldstmode
+        self.rwid = rwid
         self.rwid = rwid
+        if units and isinstance(units[0], CompUnitsBase):
+            self.n_units = 0
+            for u in self.units:
+                self.n_units += u.n_units
+        else:
+            self.n_units = len(units)
 
+        n_units = self.n_units
+
+        # inputs
         self.issue_i = Signal(n_units, reset_less=True)
         self.go_rd_i = Signal(n_units, reset_less=True)
         self.go_wr_i = Signal(n_units, reset_less=True)
+        self.shadown_i = Signal(n_units, reset_less=True)
+        self.go_die_i = Signal(n_units, reset_less=True)
+        if ldstmode:
+            self.go_ad_i = Signal(n_units, reset_less=True)
+            self.go_st_i = Signal(n_units, reset_less=True)
+
+        # outputs
         self.busy_o = Signal(n_units, reset_less=True)
         self.rd_rel_o = Signal(n_units, reset_less=True)
         self.req_rel_o = Signal(n_units, reset_less=True)
-
-        self.dest_o = Signal(rwid, reset_less=True)
-        self.src1_data_i = Signal(rwid, reset_less=True)
-        self.src2_data_i = Signal(rwid, reset_less=True)
+        if ldstmode:
+            self.ld_o = Signal(n_units, reset_less=True) # op is LD
+            self.st_o = Signal(n_units, reset_less=True) # op is ST
+            self.adr_rel_o = Signal(n_units, reset_less=True)
+            self.sto_rel_o = Signal(n_units, reset_less=True)
+            self.req_rel_o = Signal(n_units, reset_less=True)
+            self.load_mem_o = Signal(n_units, reset_less=True)
+            self.stwd_mem_o = Signal(n_units, reset_less=True)
+
+        # in/out register data (note: not register#, actual data)
+        self.data_o = Signal(rwid, reset_less=True)
+        self.src1_i = Signal(rwid, reset_less=True)
+        self.src2_i = Signal(rwid, reset_less=True)
+        # input operand
 
     def elaborate(self, platform):
         m = Module()
+        comb = m.d.comb
 
-        # Int ALUs
-        add = ALU(self.rwid)
-        sub = ALU(self.rwid)
-        mul = ALU(self.rwid)
-        shf = ALU(self.rwid)
-        m.submodules.comp1 = comp1 = ComputationUnitNoDelay(self.rwid, 2, add)
-        m.submodules.comp2 = comp2 = ComputationUnitNoDelay(self.rwid, 2, sub)
-        m.submodules.comp3 = comp3 = ComputationUnitNoDelay(self.rwid, 2, mul)
-        m.submodules.comp4 = comp4 = ComputationUnitNoDelay(self.rwid, 2, shf)
-        int_alus = [comp1, comp2, comp3, comp4]
-
-        m.d.comb += comp1.oper_i.eq(Const(0, 2)) # op=add
-        m.d.comb += comp2.oper_i.eq(Const(1, 2)) # op=sub
-        m.d.comb += comp3.oper_i.eq(Const(2, 2)) # op=mul
-        m.d.comb += comp4.oper_i.eq(Const(3, 2)) # op=shf
+        for i, alu in enumerate(self.units):
+            setattr(m.submodules, "comp%d" % i, alu)
 
         go_rd_l = []
         go_wr_l = []
@@ -64,30 +148,176 @@ class CompUnits(Elaboratable):
         busy_l = []
         req_rel_l = []
         rd_rel_l = []
-        for alu in int_alus:
+        shadow_l = []
+        godie_l = []
+        for alu in self.units:
             req_rel_l.append(alu.req_rel_o)
             rd_rel_l.append(alu.rd_rel_o)
+            shadow_l.append(alu.shadown_i)
+            godie_l.append(alu.go_die_i)
             go_wr_l.append(alu.go_wr_i)
             go_rd_l.append(alu.go_rd_i)
             issue_l.append(alu.issue_i)
             busy_l.append(alu.busy_o)
-        m.d.comb += self.rd_rel_o.eq(Cat(*rd_rel_l))
-        m.d.comb += self.req_rel_o.eq(Cat(*req_rel_l))
-        m.d.comb += self.busy_o.eq(Cat(*busy_l))
-        m.d.comb += Cat(*go_wr_l).eq(self.go_wr_i)
-        m.d.comb += Cat(*go_rd_l).eq(self.go_rd_i)
-        m.d.comb += Cat(*issue_l).eq(self.issue_i)
+        comb += self.rd_rel_o.eq(Cat(*rd_rel_l))
+        comb += self.req_rel_o.eq(Cat(*req_rel_l))
+        comb += self.busy_o.eq(Cat(*busy_l))
+        comb += Cat(*godie_l).eq(self.go_die_i)
+        comb += Cat(*shadow_l).eq(self.shadown_i)
+        comb += Cat(*go_wr_l).eq(self.go_wr_i)
+        comb += Cat(*go_rd_l).eq(self.go_rd_i)
+        comb += Cat(*issue_l).eq(self.issue_i)
 
         # connect data register input/output
 
         # merge (OR) all integer FU / ALU outputs to a single value
-        # bit of a hack: treereduce needs a list with an item named "dest_o"
-        dest_o = treereduce(int_alus)
-        m.d.comb += self.dest_o.eq(dest_o)
+        # bit of a hack: treereduce needs a list with an item named "data_o"
+        if self.units:
+            data_o = treereduce(self.units)
+            comb += self.data_o.eq(data_o)
+
+        for i, alu in enumerate(self.units):
+            comb += alu.src1_i.eq(self.src1_i)
+            comb += alu.src2_i.eq(self.src2_i)
+
+        if not self.ldstmode:
+            return m
+
+        ldmem_l = []
+        stmem_l = []
+        go_ad_l = []
+        go_st_l = []
+        ld_l = []
+        st_l = []
+        adr_rel_l = []
+        sto_rel_l = []
+        for alu in self.units:
+            ld_l.append(alu.ld_o)
+            st_l.append(alu.st_o)
+            adr_rel_l.append(alu.adr_rel_o)
+            sto_rel_l.append(alu.sto_rel_o)
+            ldmem_l.append(alu.load_mem_o)
+            stmem_l.append(alu.stwd_mem_o)
+            go_ad_l.append(alu.go_ad_i)
+            go_st_l.append(alu.go_st_i)
+        comb += self.adr_rel_o.eq(Cat(*adr_rel_l))
+        comb += self.sto_rel_o.eq(Cat(*sto_rel_l))
+        comb += self.load_mem_o.eq(Cat(*ldmem_l))
+        comb += self.stwd_mem_o.eq(Cat(*stmem_l))
+        comb += Cat(*go_ad_l).eq(self.go_ad_i)
+        comb += Cat(*go_st_l).eq(self.go_st_i)
+
+        return m
+
+
+class CompUnitLDSTs(CompUnitsBase):
+
+    def __init__(self, rwid, opwid, n_ldsts, mem):
+        """ Inputs:
+
+            * :rwid:   bit width of register file(s) - both FP and INT
+            * :opwid:  operand bit width
+        """
+        self.opwid = opwid
+
+        # inputs
+        self.oper_i = Signal(opwid, reset_less=True)
+        self.imm_i = Signal(rwid, reset_less=True)
+
+        # Int ALUs
+        self.alus = []
+        for i in range(n_ldsts):
+            self.alus.append(ALU(rwid))
+
+        units = []
+        for alu in self.alus:
+            aluopwid = 4 # see compldst.py for "internal" opcode
+            units.append(LDSTCompUnit(rwid, aluopwid, alu, mem))
+
+        CompUnitsBase.__init__(self, rwid, units, ldstmode=True)
+
+    def elaborate(self, platform):
+        m = CompUnitsBase.elaborate(self, platform)
+        comb = m.d.comb
+
+        # hand the same operation to all units, 4 lower bits though
+        for alu in self.units:
+            comb += alu.oper_i[0:4].eq(self.oper_i)
+            comb += alu.imm_i.eq(self.imm_i)
+            comb += alu.isalu_i.eq(0)
+
+        return m
+
+
+class CompUnitALUs(CompUnitsBase):
+
+    def __init__(self, rwid, opwid, n_alus):
+        """ Inputs:
+
+            * :rwid:   bit width of register file(s) - both FP and INT
+            * :opwid:  operand bit width
+        """
+        self.opwid = opwid
+
+        # inputs
+        self.oper_i = Signal(opwid, reset_less=True)
+        self.imm_i = Signal(rwid, reset_less=True)
+
+        # Int ALUs
+        alus = []
+        for i in range(n_alus):
+            alus.append(ALU(rwid))
+
+        units = []
+        for alu in alus:
+            aluopwid = 3 # extra bit for immediate mode
+            units.append(ComputationUnitNoDelay(rwid, aluopwid, alu))
+
+        CompUnitsBase.__init__(self, rwid, units)
+
+    def elaborate(self, platform):
+        m = CompUnitsBase.elaborate(self, platform)
+        comb = m.d.comb
+
+        # hand the same operation to all units, only lower 3 bits though
+        for alu in self.units:
+            comb += alu.oper_i[0:3].eq(self.oper_i)
+            comb += alu.imm_i.eq(self.imm_i)
+
+        return m
+
+
+class CompUnitBR(CompUnitsBase):
+
+    def __init__(self, rwid, opwid):
+        """ Inputs:
+
+            * :rwid:   bit width of register file(s) - both FP and INT
+            * :opwid:  operand bit width
+
+            Note: bgt unit is returned so that a shadow unit can be created
+            for it
+        """
+        self.opwid = opwid
+
+        # inputs
+        self.oper_i = Signal(opwid, reset_less=True)
+        self.imm_i = Signal(rwid, reset_less=True)
+
+        # Branch ALU and CU
+        self.bgt = BranchALU(rwid)
+        aluopwid = 3 # extra bit for immediate mode
+        self.br1 = ComputationUnitNoDelay(rwid, aluopwid, self.bgt)
+        CompUnitsBase.__init__(self, rwid, [self.br1])
 
-        for i, alu in enumerate(int_alus):
-            m.d.comb += alu.src1_i.eq(self.src1_data_i)
-            m.d.comb += alu.src2_i.eq(self.src2_data_i)
+    def elaborate(self, platform):
+        m = CompUnitsBase.elaborate(self, platform)
+        comb = m.d.comb
+
+        # hand the same operation to all units
+        for alu in self.units:
+            comb += alu.oper_i.eq(self.oper_i)
+            comb += alu.imm_i.eq(self.imm_i)
 
         return m
 
@@ -109,54 +339,60 @@ class FunctionUnits(Elaboratable):
         self.src1_rsel_o = Signal(n_regs, reset_less=True) # src1 reg (bot)
         self.src2_rsel_o = Signal(n_regs, reset_less=True) # src2 reg (bot)
 
-        self.req_rel_i = Signal(n_int_alus, reset_less = True)
         self.readable_o = Signal(n_int_alus, reset_less=True)
         self.writable_o = Signal(n_int_alus, reset_less=True)
 
         self.go_rd_i = Signal(n_int_alus, reset_less=True)
         self.go_wr_i = Signal(n_int_alus, reset_less=True)
-        self.req_rel_o = Signal(n_int_alus, reset_less=True)
+        self.go_die_i = Signal(n_int_alus, reset_less=True)
         self.fn_issue_i = Signal(n_int_alus, reset_less=True)
 
+        # Note: FURegs wr_pend_o is also outputted from here, for use in WaWGrid
+
     def elaborate(self, platform):
         m = Module()
+        comb = m.d.comb
+        sync = m.d.sync
 
-        n_int_fus = self.n_int_alus
+        n_intfus = self.n_int_alus
 
         # Integer FU-FU Dep Matrix
-        intfudeps = FUFUDepMatrix(n_int_fus, n_int_fus)
+        intfudeps = FUFUDepMatrix(n_intfus, n_intfus)
         m.submodules.intfudeps = intfudeps
         # Integer FU-Reg Dep Matrix
-        intregdeps = FURegDepMatrix(n_int_fus, self.n_regs)
+        intregdeps = FURegDepMatrix(n_intfus, self.n_regs, 2)
         m.submodules.intregdeps = intregdeps
 
-        m.d.comb += self.g_int_rd_pend_o.eq(intregdeps.rd_rsel_o)
-        m.d.comb += self.g_int_wr_pend_o.eq(intregdeps.wr_rsel_o)
+        comb += self.g_int_rd_pend_o.eq(intregdeps.v_rd_rsel_o)
+        comb += self.g_int_wr_pend_o.eq(intregdeps.v_wr_rsel_o)
 
-        m.d.comb += intregdeps.rd_pend_i.eq(intregdeps.rd_rsel_o)
-        m.d.comb += intregdeps.wr_pend_i.eq(intregdeps.wr_rsel_o)
+        comb += intregdeps.rd_pend_i.eq(intregdeps.v_rd_rsel_o)
+        comb += intregdeps.wr_pend_i.eq(intregdeps.v_wr_rsel_o)
 
-        m.d.comb += intfudeps.rd_pend_i.eq(intregdeps.rd_pend_o)
-        m.d.comb += intfudeps.wr_pend_i.eq(intregdeps.wr_pend_o)
+        comb += intfudeps.rd_pend_i.eq(intregdeps.rd_pend_o)
+        comb += intfudeps.wr_pend_i.eq(intregdeps.wr_pend_o)
+        self.wr_pend_o = intregdeps.wr_pend_o # also output for use in WaWGrid
 
-        m.d.comb += intfudeps.issue_i.eq(self.fn_issue_i)
-        m.d.comb += intfudeps.go_rd_i.eq(self.go_rd_i)
-        m.d.comb += intfudeps.go_wr_i.eq(self.go_wr_i)
-        m.d.comb += self.readable_o.eq(intfudeps.readable_o)
-        m.d.comb += self.writable_o.eq(intfudeps.writable_o)
+        comb += intfudeps.issue_i.eq(self.fn_issue_i)
+        comb += intfudeps.go_rd_i.eq(self.go_rd_i)
+        comb += intfudeps.go_wr_i.eq(self.go_wr_i)
+        comb += intfudeps.go_die_i.eq(self.go_die_i)
+        comb += self.readable_o.eq(intfudeps.readable_o)
+        comb += self.writable_o.eq(intfudeps.writable_o)
 
         # Connect function issue / arrays, and dest/src1/src2
-        m.d.comb += intregdeps.dest_i.eq(self.dest_i)
-        m.d.comb += intregdeps.src1_i.eq(self.src1_i)
-        m.d.comb += intregdeps.src2_i.eq(self.src2_i)
+        comb += intregdeps.dest_i.eq(self.dest_i)
+        comb += intregdeps.src_i[0].eq(self.src1_i)
+        comb += intregdeps.src_i[1].eq(self.src2_i)
 
-        m.d.comb += intregdeps.go_rd_i.eq(self.go_rd_i)
-        m.d.comb += intregdeps.go_wr_i.eq(self.go_wr_i)
-        m.d.comb += intregdeps.issue_i.eq(self.fn_issue_i)
+        comb += intregdeps.go_rd_i.eq(self.go_rd_i)
+        comb += intregdeps.go_wr_i.eq(self.go_wr_i)
+        comb += intregdeps.go_die_i.eq(self.go_die_i)
+        comb += intregdeps.issue_i.eq(self.fn_issue_i)
 
-        m.d.comb += self.dest_rsel_o.eq(intregdeps.dest_rsel_o)
-        m.d.comb += self.src1_rsel_o.eq(intregdeps.src1_rsel_o)
-        m.d.comb += self.src2_rsel_o.eq(intregdeps.src2_rsel_o)
+        comb += self.dest_rsel_o.eq(intregdeps.dest_rsel_o)
+        comb += self.src1_rsel_o.eq(intregdeps.src_rsel_o[0])
+        comb += self.src2_rsel_o.eq(intregdeps.src_rsel_o[1])
 
         return m
 
@@ -175,18 +411,40 @@ class Scoreboard(Elaboratable):
         self.intregs = RegFileArray(rwid, n_regs)
         self.fpregs = RegFileArray(rwid, n_regs)
 
+        # issue q needs to get at these
+        self.aluissue = IssueUnitGroup(2)
+        self.lsissue = IssueUnitGroup(2)
+        self.brissue = IssueUnitGroup(1)
+        # and these
+        self.alu_oper_i = Signal(4, reset_less=True)
+        self.alu_imm_i = Signal(rwid, reset_less=True)
+        self.br_oper_i = Signal(4, reset_less=True)
+        self.br_imm_i = Signal(rwid, reset_less=True)
+        self.ls_oper_i = Signal(4, reset_less=True)
+        self.ls_imm_i = Signal(rwid, reset_less=True)
+
         # inputs
-        self.int_store_i = Signal(reset_less=True) # instruction is a store
         self.int_dest_i = Signal(max=n_regs, reset_less=True) # Dest R# in
         self.int_src1_i = Signal(max=n_regs, reset_less=True) # oper1 R# in
         self.int_src2_i = Signal(max=n_regs, reset_less=True) # oper2 R# in
         self.reg_enable_i = Signal(reset_less=True) # enable reg decode
 
+        # outputs
         self.issue_o = Signal(reset_less=True) # instruction was accepted
         self.busy_o = Signal(reset_less=True) # at least one CU is busy
 
+        # for branch speculation experiment.  branch_direction = 0 if
+        # the branch hasn't been met yet.  1 indicates "success", 2 is "fail"
+        # branch_succ and branch_fail are requests to have the current
+        # instruction be dependent on the branch unit "shadow" capability.
+        self.branch_succ_i = Signal(reset_less=True)
+        self.branch_fail_i = Signal(reset_less=True)
+        self.branch_direction_o = Signal(2, reset_less=True)
+
     def elaborate(self, platform):
         m = Module()
+        comb = m.d.comb
+        sync = m.d.sync
 
         m.submodules.intregs = self.intregs
         m.submodules.fpregs = self.fpregs
@@ -200,27 +458,54 @@ class Scoreboard(Elaboratable):
         fp_src1 = self.fpregs.read_port("src1")
         fp_src2 = self.fpregs.read_port("src2")
 
-        # Int ALUs and Comp Units
-        n_int_alus = 4
-        m.submodules.cu = cu = CompUnits(self.rwid, n_int_alus)
+        # Int ALUs and BR ALUs
+        n_int_alus = 5
+        cua = CompUnitALUs(self.rwid, 3, n_alus=self.aluissue.n_insns)
+        cub = CompUnitBR(self.rwid, 3) # 1 BR ALUs
+
+        # LDST Comp Units
+        n_ldsts = 2
+        cul = CompUnitLDSTs(self.rwid, 4, self.lsissue.n_insns, None)
+
+        # Comp Units
+        m.submodules.cu = cu = CompUnitsBase(self.rwid, [cua, cul, cub])
+        bgt = cub.bgt # get at the branch computation unit
+        br1 = cub.br1
 
         # Int FUs
         m.submodules.intfus = intfus = FunctionUnits(self.n_regs, n_int_alus)
 
+        # Memory FUs
+        m.submodules.memfus = memfus = MemFunctionUnits(n_ldsts, 5)
+
         # Count of number of FUs
-        n_int_fus = n_int_alus
+        n_intfus = n_int_alus
         n_fp_fus = 0 # for now
 
-        # Integer Priority Picker 1: Adder + Subtractor
-        intpick1 = GroupPicker(n_int_fus) # picks between add, sub, mul and shf
+        # Integer Priority Picker 1: Adder + Subtractor (and LD/ST)
+        intpick1 = GroupPicker(n_intfus) # picks 1 reader and 1 writer to intreg
         m.submodules.intpick1 = intpick1
 
         # INT/FP Issue Unit
         regdecode = RegDecode(self.n_regs)
         m.submodules.regdecode = regdecode
-        issueunit = IntFPIssueUnit(self.n_regs, n_int_fus, n_fp_fus)
+        issueunit = IssueUnitArray([self.aluissue, self.lsissue, self.brissue])
         m.submodules.issueunit = issueunit
 
+        # Shadow Matrix.  currently n_intfus shadows, to be used for
+        # write-after-write hazards.  NOTE: there is one extra for branches,
+        # so the shadow width is increased by 1
+        m.submodules.shadows = shadows = ShadowMatrix(n_intfus, n_intfus, True)
+        m.submodules.bshadow = bshadow = ShadowMatrix(n_intfus, 1, False)
+
+        # record previous instruction to cast shadow on current instruction
+        prev_shadow = Signal(n_intfus)
+
+        # Branch Speculation recorder.  tracks the success/fail state as
+        # each instruction is issued, so that when the branch occurs the
+        # allow/cancel can be issued as appropriate.
+        m.submodules.specrec = bspec = BranchSpeculationRecord(n_intfus)
+
         #---------
         # ok start wiring things together...
         # "now hear de word of de looord... dem bones dem bones dem dryy bones"
@@ -230,117 +515,392 @@ class Scoreboard(Elaboratable):
         #---------
         # Issue Unit is where it starts.  set up some in/outs for this module
         #---------
-        m.d.comb += [issueunit.i.store_i.eq(self.int_store_i),
-                     regdecode.dest_i.eq(self.int_dest_i),
+        comb += [    regdecode.dest_i.eq(self.int_dest_i),
                      regdecode.src1_i.eq(self.int_src1_i),
                      regdecode.src2_i.eq(self.int_src2_i),
                      regdecode.enable_i.eq(self.reg_enable_i),
-                     issueunit.i.dest_i.eq(regdecode.dest_o),
                      self.issue_o.eq(issueunit.issue_o)
                     ]
-        self.int_insn_i = issueunit.i.insn_i # enabled by instruction decode
 
-        # connect global rd/wr pending vector (for WaW detection)
-        m.d.sync += issueunit.i.g_wr_pend_i.eq(intfus.g_int_wr_pend_o)
+        # take these to outside (issue needs them)
+        comb += cua.oper_i.eq(self.alu_oper_i)
+        comb += cua.imm_i.eq(self.alu_imm_i)
+        comb += cub.oper_i.eq(self.br_oper_i)
+        comb += cub.imm_i.eq(self.br_imm_i)
+        comb += cul.oper_i.eq(self.ls_oper_i)
+        comb += cul.imm_i.eq(self.ls_imm_i)
+
         # TODO: issueunit.f (FP)
 
         # and int function issue / busy arrays, and dest/src1/src2
-        m.d.comb += intfus.dest_i.eq(regdecode.dest_o)
-        m.d.comb += intfus.src1_i.eq(regdecode.src1_o)
-        m.d.comb += intfus.src2_i.eq(regdecode.src2_o)
+        comb += intfus.dest_i.eq(regdecode.dest_o)
+        comb += intfus.src1_i.eq(regdecode.src1_o)
+        comb += intfus.src2_i.eq(regdecode.src2_o)
 
-        fn_issue_o = issueunit.i.fn_issue_o
+        fn_issue_o = issueunit.fn_issue_o
 
-        m.d.comb += intfus.fn_issue_i.eq(fn_issue_o)
-        m.d.comb += issueunit.i.busy_i.eq(cu.busy_o)
-        m.d.comb += self.busy_o.eq(cu.busy_o.bool())
+        comb += intfus.fn_issue_i.eq(fn_issue_o)
+        comb += issueunit.busy_i.eq(cu.busy_o)
+        comb += self.busy_o.eq(cu.busy_o.bool())
+
+        #---------
+        # Memory Function Unit
+        #---------
+        reset_b = Signal(cul.n_units, reset_less=True)
+        sync += reset_b.eq(cul.go_st_i | cul.go_wr_i | cul.go_die_i)
+
+
+        comb += memfus.fn_issue_i.eq(cul.issue_i) # Comp Unit Issue -> Mem FUs
+        comb += memfus.addr_en_i.eq(cul.adr_rel_o) # Match enable on adr rel
+        comb += memfus.addr_rs_i.eq(reset_b) # reset same as LDSTCompUnit
+
+        # LD/STs have to accumulate prior LD/STs (TODO: multi-issue as well,
+        # in a transitive fashion).  This cycle activates based on LDSTCompUnit
+        # issue_i.  multi-issue gets a bit more complex but not a lot.
+        prior_ldsts = Signal(cul.n_units, reset_less=True)
+        sync += prior_ldsts.eq(memfus.g_int_ld_pend_o | memfus.g_int_st_pend_o)
+        with m.If(self.ls_oper_i[2]): # LD bit of operand
+            comb += memfus.ld_i.eq(cul.issue_i | prior_ldsts)
+        with m.If(self.ls_oper_i[3]): # ST bit of operand
+            comb += memfus.st_i.eq(cul.issue_i | prior_ldsts)
+
+        # TODO: adr_rel_o needs to go into L1 Cache.  for now,
+        # just immediately activate go_adr
+        comb += cul.go_ad_i.eq(cul.adr_rel_o)
+
+        # connect up address data
+        comb += memfus.addrs_i[0].eq(cul.units[0].data_o)
+        comb += memfus.addrs_i[1].eq(cul.units[1].data_o)
+
+        # connect loadable / storable to go_ld/go_st.
+        # XXX should only be done when the memory ld/st has actually happened!
+        go_st_i = Signal(cul.n_units, reset_less=True)
+        go_ld_i = Signal(cul.n_units, reset_less=True)
+        comb += go_ld_i.eq(memfus.storable_o & memfus.addr_nomatch_o &\
+                                  cul.req_rel_o & cul.ld_o)
+        comb += go_st_i.eq(memfus.storable_o & memfus.addr_nomatch_o &\
+                                  cul.sto_rel_o & cul.st_o)
+        comb += memfus.go_ld_i.eq(go_ld_i)
+        comb += memfus.go_st_i.eq(go_st_i)
+        #comb += cul.go_wr_i.eq(memfus.loadable_o & memfus.addr_nomatch_o)
+        comb += cul.go_st_i.eq(go_st_i)
+
+        #comb += cu.go_rd_i[0:n_intfus].eq(go_rd_o[0:n_intfus])
+        #comb += cu.go_wr_i[0:n_intfus].eq(go_wr_o[0:n_intfus])
+        #comb += cu.issue_i[0:n_intfus].eq(fn_issue_o[0:n_intfus])
+
+        #---------
+        # merge shadow matrices outputs
+        #---------
+
+        # these are explained in ShadowMatrix docstring, and are to be
+        # connected to the FUReg and FUFU Matrices, to get them to reset
+        anydie = Signal(n_intfus, reset_less=True)
+        allshadown = Signal(n_intfus, reset_less=True)
+        shreset = Signal(n_intfus, reset_less=True)
+        comb += allshadown.eq(shadows.shadown_o & bshadow.shadown_o)
+        comb += anydie.eq(shadows.go_die_o | bshadow.go_die_o)
+        comb += shreset.eq(bspec.match_g_o | bspec.match_f_o)
 
         #---------
         # connect fu-fu matrix
         #---------
 
-        # Group Picker... done manually for now.  TODO: cat array of pick sigs
+        # Group Picker... done manually for now.
         go_rd_o = intpick1.go_rd_o
         go_wr_o = intpick1.go_wr_o
         go_rd_i = intfus.go_rd_i
         go_wr_i = intfus.go_wr_i
-        m.d.comb += go_rd_i[0:n_int_fus].eq(go_rd_o[0:n_int_fus]) # rd
-        m.d.comb += go_wr_i[0:n_int_fus].eq(go_wr_o[0:n_int_fus]) # wr
+        go_die_i = intfus.go_die_i
+        # NOTE: connect to the shadowed versions so that they can "die" (reset)
+        comb += go_rd_i[0:n_intfus].eq(go_rd_o[0:n_intfus]) # rd
+        comb += go_wr_i[0:n_intfus].eq(go_wr_o[0:n_intfus]) # wr
+        comb += go_die_i[0:n_intfus].eq(anydie[0:n_intfus]) # die
 
         # Connect Picker
         #---------
-        m.d.comb += intpick1.rd_rel_i[0:n_int_fus].eq(cu.rd_rel_o[0:n_int_fus])
-        m.d.comb += intpick1.req_rel_i[0:n_int_fus].eq(cu.req_rel_o[0:n_int_fus])
+        comb += intpick1.rd_rel_i[0:n_intfus].eq(cu.rd_rel_o[0:n_intfus])
+        comb += intpick1.req_rel_i[0:n_intfus].eq(cu.req_rel_o[0:n_intfus])
         int_rd_o = intfus.readable_o
         int_wr_o = intfus.writable_o
-        m.d.comb += intpick1.readable_i[0:n_int_fus].eq(int_rd_o[0:n_int_fus])
-        m.d.comb += intpick1.writable_i[0:n_int_fus].eq(int_wr_o[0:n_int_fus])
+        comb += intpick1.readable_i[0:n_intfus].eq(int_rd_o[0:n_intfus])
+        comb += intpick1.writable_i[0:n_intfus].eq(int_wr_o[0:n_intfus])
+
+        #---------
+        # Shadow Matrix
+        #---------
+
+        comb += shadows.issue_i.eq(fn_issue_o)
+        #comb += shadows.reset_i[0:n_intfus].eq(bshadow.go_die_o[0:n_intfus])
+        comb += shadows.reset_i[0:n_intfus].eq(bshadow.go_die_o[0:n_intfus])
+        #---------
+        # NOTE; this setup is for the instruction order preservation...
+
+        # connect shadows / go_dies to Computation Units
+        comb += cu.shadown_i[0:n_intfus].eq(allshadown)
+        comb += cu.go_die_i[0:n_intfus].eq(anydie)
+
+        # ok connect first n_int_fu shadows to busy lines, to create an
+        # instruction-order linked-list-like arrangement, using a bit-matrix
+        # (instead of e.g. a ring buffer).
+
+        # when written, the shadow can be cancelled (and was good)
+        for i in range(n_intfus):
+            comb += shadows.s_good_i[i][0:n_intfus].eq(go_wr_o[0:n_intfus])
+
+        # *previous* instruction shadows *current* instruction, and, obviously,
+        # if the previous is completed (!busy) don't cast the shadow!
+        comb += prev_shadow.eq(~fn_issue_o & cu.busy_o)
+        for i in range(n_intfus):
+            comb += shadows.shadow_i[i][0:n_intfus].eq(prev_shadow)
+
+        #---------
+        # ... and this is for branch speculation.  it uses the extra bit
+        # tacked onto the ShadowMatrix (hence shadow_wid=n_intfus+1)
+        # only needs to set shadow_i, s_fail_i and s_good_i
+
+        # issue captures shadow_i (if enabled)
+        comb += bshadow.reset_i[0:n_intfus].eq(shreset[0:n_intfus])
+
+        bactive = Signal(reset_less=True)
+        comb += bactive.eq((bspec.active_i | br1.issue_i) & ~br1.go_wr_i)
+
+        # instruction being issued (fn_issue_o) has a shadow cast by the branch
+        with m.If(bactive & (self.branch_succ_i | self.branch_fail_i)):
+            comb += bshadow.issue_i.eq(fn_issue_o)
+            for i in range(n_intfus):
+                with m.If(fn_issue_o & (Const(1<<i))):
+                    comb += bshadow.shadow_i[i][0].eq(1)
+
+        # finally, we need an indicator to the test infrastructure as to
+        # whether the branch succeeded or failed, plus, link up to the
+        # "recorder" of whether the instruction was under shadow or not
+
+        with m.If(br1.issue_i):
+            sync += bspec.active_i.eq(1)
+        with m.If(self.branch_succ_i):
+            comb += bspec.good_i.eq(fn_issue_o & 0x1f) # XXX MAGIC CONSTANT
+        with m.If(self.branch_fail_i):
+            comb += bspec.fail_i.eq(fn_issue_o & 0x1f) # XXX MAGIC CONSTANT
+
+        # branch is active (TODO: a better signal: this is over-using the
+        # go_write signal - actually the branch should not be "writing")
+        with m.If(br1.go_wr_i):
+            sync += self.branch_direction_o.eq(br1.data_o+Const(1, 2))
+            sync += bspec.active_i.eq(0)
+            comb += bspec.br_i.eq(1)
+            # branch occurs if data == 1, failed if data == 0
+            comb += bspec.br_ok_i.eq(br1.data_o == 1)
+            for i in range(n_intfus):
+                # *expected* direction of the branch matched against *actual*
+                comb += bshadow.s_good_i[i][0].eq(bspec.match_g_o[i])
+                # ... or it didn't
+                comb += bshadow.s_fail_i[i][0].eq(bspec.match_f_o[i])
 
         #---------
         # Connect Register File(s)
         #---------
-        print ("intregdeps wen len", len(intfus.dest_rsel_o))
-        m.d.comb += int_dest.wen.eq(intfus.dest_rsel_o)
-        m.d.comb += int_src1.ren.eq(intfus.src1_rsel_o)
-        m.d.comb += int_src2.ren.eq(intfus.src2_rsel_o)
+        comb += int_dest.wen.eq(intfus.dest_rsel_o)
+        comb += int_src1.ren.eq(intfus.src1_rsel_o)
+        comb += int_src2.ren.eq(intfus.src2_rsel_o)
 
         # connect ALUs to regfule
-        m.d.comb += int_dest.data_i.eq(cu.dest_o)
-        m.d.comb += cu.src1_data_i.eq(int_src1.data_o)
-        m.d.comb += cu.src2_data_i.eq(int_src2.data_o)
+        comb += int_dest.data_i.eq(cu.data_o)
+        comb += cu.src1_i.eq(int_src1.data_o)
+        comb += cu.src2_i.eq(int_src2.data_o)
 
         # connect ALU Computation Units
-        m.d.comb += cu.go_rd_i[0:n_int_fus].eq(go_rd_o[0:n_int_fus])
-        m.d.comb += cu.go_wr_i[0:n_int_fus].eq(go_wr_o[0:n_int_fus])
-        m.d.comb += cu.issue_i[0:n_int_fus].eq(fn_issue_o[0:n_int_fus])
+        comb += cu.go_rd_i[0:n_intfus].eq(go_rd_o[0:n_intfus])
+        comb += cu.go_wr_i[0:n_intfus].eq(go_wr_o[0:n_intfus])
+        comb += cu.issue_i[0:n_intfus].eq(fn_issue_o[0:n_intfus])
 
         return m
 
-
     def __iter__(self):
         yield from self.intregs
         yield from self.fpregs
-        yield self.int_store_i
         yield self.int_dest_i
         yield self.int_src1_i
         yield self.int_src2_i
         yield self.issue_o
-        #yield from self.int_src1
-        #yield from self.int_dest
-        #yield from self.int_src1
-        #yield from self.int_src2
-        #yield from self.fp_dest
-        #yield from self.fp_src1
-        #yield from self.fp_src2
+        yield self.branch_succ_i
+        yield self.branch_fail_i
+        yield self.branch_direction_o
+
+    def ports(self):
+        return list(self)
+
+
+class IssueToScoreboard(Elaboratable):
+
+    def __init__(self, qlen, n_in, n_out, rwid, opwid, n_regs):
+        self.qlen = qlen
+        self.n_in = n_in
+        self.n_out = n_out
+        self.rwid = rwid
+        self.opw = opwid
+        self.n_regs = n_regs
+
+        mqbits = (int(log(qlen) / log(2))+2, False)
+        self.p_add_i = Signal(mqbits) # instructions to add (from data_i)
+        self.p_ready_o = Signal() # instructions were added
+        self.data_i = Instruction.nq(n_in, "data_i", rwid, opwid)
+
+        self.busy_o = Signal(reset_less=True) # at least one CU is busy
+        self.qlen_o = Signal(mqbits, reset_less=True)
+
+    def elaborate(self, platform):
+        m = Module()
+        comb = m.d.comb
+        sync = m.d.sync
+
+        iq = InstructionQ(self.rwid, self.opw, self.qlen, self.n_in, self.n_out)
+        sc = Scoreboard(self.rwid, self.n_regs)
+        mem = TestMemory(self.rwid, 8) # not too big, takes too long
+        m.submodules.iq = iq
+        m.submodules.sc = sc
+        m.submodules.mem = mem
+
+        # get at the regfile for testing
+        self.intregs = sc.intregs
+
+        # and the "busy" signal and instruction queue length
+        comb += self.busy_o.eq(sc.busy_o)
+        comb += self.qlen_o.eq(iq.qlen_o)
+
+        # link up instruction queue
+        comb += iq.p_add_i.eq(self.p_add_i)
+        comb += self.p_ready_o.eq(iq.p_ready_o)
+        for i in range(self.n_in):
+            comb += eq(iq.data_i[i], self.data_i[i])
+
+        # take instruction and process it.  note that it's possible to
+        # "inspect" the queue contents *without* actually removing the
+        # items.  items are only removed when the
+
+        # in "waiting" state
+        wait_issue_br = Signal()
+        wait_issue_alu = Signal()
+        wait_issue_ls = Signal()
+
+        with m.If(wait_issue_br | wait_issue_alu | wait_issue_ls):
+            # set instruction pop length to 1 if the unit accepted
+            with m.If(wait_issue_ls & (sc.lsissue.fn_issue_o != 0)):
+                with m.If(iq.qlen_o != 0):
+                    comb += iq.n_sub_i.eq(1)
+            with m.If(wait_issue_br & (sc.brissue.fn_issue_o != 0)):
+                with m.If(iq.qlen_o != 0):
+                    comb += iq.n_sub_i.eq(1)
+            with m.If(wait_issue_alu & (sc.aluissue.fn_issue_o != 0)):
+                with m.If(iq.qlen_o != 0):
+                    comb += iq.n_sub_i.eq(1)
+
+        # see if some instruction(s) are here.  note that this is
+        # "inspecting" the in-place queue.  note also that on the
+        # cycle following "waiting" for fn_issue_o to be set, the
+        # "resetting" done above (insn_i=0) could be re-ASSERTed.
+        with m.If(iq.qlen_o != 0):
+            # get the operands and operation
+            imm = iq.data_o[0].imm_i
+            dest = iq.data_o[0].dest_i
+            src1 = iq.data_o[0].src1_i
+            src2 = iq.data_o[0].src2_i
+            op = iq.data_o[0].oper_i
+            opi = iq.data_o[0].opim_i # immediate set
+
+            # set the src/dest regs
+            comb += sc.int_dest_i.eq(dest)
+            comb += sc.int_src1_i.eq(src1)
+            comb += sc.int_src2_i.eq(src2)
+            comb += sc.reg_enable_i.eq(1) # enable the regfile
+
+            # choose a Function-Unit-Group
+            with m.If((op & (0x3<<2)) != 0): # branch
+                comb += sc.br_oper_i.eq(Cat(op[0:2], opi))
+                comb += sc.br_imm_i.eq(imm)
+                comb += sc.brissue.insn_i.eq(1)
+                comb += wait_issue_br.eq(1)
+            with m.Elif((op & (0x3<<4)) != 0): # ld/st
+                # see compldst.py
+                # bit 0: ADD/SUB
+                # bit 1: immed
+                # bit 4: LD
+                # bit 5: ST
+                comb += sc.ls_oper_i.eq(Cat(op[0], opi[0], op[4:6]))
+                comb += sc.ls_imm_i.eq(imm)
+                comb += sc.lsissue.insn_i.eq(1)
+                comb += wait_issue_ls.eq(1)
+            with m.Else(): # alu
+                comb += sc.alu_oper_i.eq(Cat(op[0:2], opi))
+                comb += sc.alu_imm_i.eq(imm)
+                comb += sc.aluissue.insn_i.eq(1)
+                comb += wait_issue_alu.eq(1)
+
+            # XXX TODO
+            # these indicate that the instruction is to be made
+            # shadow-dependent on
+            # (either) branch success or branch fail
+            #yield sc.branch_fail_i.eq(branch_fail)
+            #yield sc.branch_succ_i.eq(branch_success)
+
+        return m
+
+    def __iter__(self):
+        yield self.p_ready_o
+        for o in self.data_i:
+            yield from list(o)
+        yield self.p_add_i
 
     def ports(self):
         return list(self)
 
+
 IADD = 0
 ISUB = 1
 IMUL = 2
 ISHF = 3
+IBGT = 4
+IBLT = 5
+IBEQ = 6
+IBNE = 7
+
 
 class RegSim:
     def __init__(self, rwidth, nregs):
         self.rwidth = rwidth
         self.regs = [0] * nregs
 
-    def op(self, op, src1, src2, dest):
-        src1 = self.regs[src1]
-        src2 = self.regs[src2]
+    def op(self, op, op_imm, imm, src1, src2, dest):
+        maxbits = (1 << self.rwidth) - 1
+        src1 = self.regs[src1] & maxbits
+        if op_imm:
+            src2 = imm
+        else:
+            src2 = self.regs[src2] & maxbits
         if op == IADD:
-            val = (src1 + src2)
+            val = src1 + src2
         elif op == ISUB:
-            val = (src1 - src2)
+            val = src1 - src2
         elif op == IMUL:
-            val = (src1 * src2)
+            val = src1 * src2
         elif op == ISHF:
-            val = (src1 << (src2 & self.rwidth))
-        val &= ((1<<(self.rwidth))-1)
-        self.regs[dest] = val
+            val = src1 >> (src2 & maxbits)
+        elif op == IBGT:
+            val = int(src1 > src2)
+        elif op == IBLT:
+            val = int(src1 < src2)
+        elif op == IBEQ:
+            val = int(src1 == src2)
+        elif op == IBNE:
+            val = int(src1 != src2)
+        else:
+            return 0 # LD/ST TODO
+        val &= maxbits
+        self.setval(dest, val)
+        return val
 
     def setval(self, dest, val):
+        print ("sim setval", dest, hex(val))
         self.regs[dest] = val
 
     def dump(self, dut):
@@ -357,15 +917,50 @@ class RegSim:
                 yield from self.dump(dut)
                 assert False
 
-def int_instr(dut, alusim, op, src1, src2, dest):
-    for i in range(len(dut.int_insn_i)):
-        yield dut.int_insn_i[i].eq(0)
+def instr_q(dut, op, op_imm, imm, src1, src2, dest,
+            branch_success, branch_fail):
+    instrs = [{'oper_i': op, 'dest_i': dest, 'imm_i': imm, 'opim_i': op_imm,
+               'src1_i': src1, 'src2_i': src2}]
+
+    sendlen = 1
+    for idx in range(sendlen):
+        yield from eq(dut.data_i[idx], instrs[idx])
+        di = yield dut.data_i[idx]
+        print ("senddata %d %x" % (idx, di))
+    yield dut.p_add_i.eq(sendlen)
+    yield
+    o_p_ready = yield dut.p_ready_o
+    while not o_p_ready:
+        yield
+        o_p_ready = yield dut.p_ready_o
+
+    yield dut.p_add_i.eq(0)
+
+
+def int_instr(dut, op, imm, src1, src2, dest, branch_success, branch_fail):
+    yield from disable_issue(dut)
     yield dut.int_dest_i.eq(dest)
     yield dut.int_src1_i.eq(src1)
     yield dut.int_src2_i.eq(src2)
-    yield dut.int_insn_i[op].eq(1)
+    if (op & (0x3<<2)) != 0: # branch
+        yield dut.brissue.insn_i.eq(1)
+        yield dut.br_oper_i.eq(Const(op & 0x3, 2))
+        yield dut.br_imm_i.eq(imm)
+        dut_issue = dut.brissue
+    else:
+        yield dut.aluissue.insn_i.eq(1)
+        yield dut.alu_oper_i.eq(Const(op & 0x3, 2))
+        yield dut.alu_imm_i.eq(imm)
+        dut_issue = dut.aluissue
     yield dut.reg_enable_i.eq(1)
-    alusim.op(op, src1, src2, dest)
+
+    # these indicate that the instruction is to be made shadow-dependent on
+    # (either) branch success or branch fail
+    yield dut.branch_fail_i.eq(branch_fail)
+    yield dut.branch_succ_i.eq(branch_success)
+
+    yield
+    yield from wait_for_issue(dut, dut_issue)
 
 
 def print_reg(dut, rnums):
@@ -377,41 +972,205 @@ def print_reg(dut, rnums):
     print ("reg %s: %s" % (','.join(rnums), ','.join(rs)))
 
 
+def create_random_ops(dut, n_ops, shadowing=False, max_opnums=3):
+    insts = []
+    for i in range(n_ops):
+        src1 = randint(1, dut.n_regs-1)
+        src2 = randint(1, dut.n_regs-1)
+        imm = randint(1, (1<<dut.rwid)-1)
+        dest = randint(1, dut.n_regs-1)
+        op = randint(0, max_opnums)
+        opi = 0 if randint(0, 2) else 1 # set true if random is nonzero
+
+        if shadowing:
+            insts.append((src1, src2, dest, op, opi, imm, (0, 0)))
+        else:
+            insts.append((src1, src2, dest, op, opi, imm))
+    return insts
+
+
+def wait_for_busy_clear(dut):
+    while True:
+        busy_o = yield dut.busy_o
+        if not busy_o:
+            break
+        print ("busy",)
+        yield
+
+def disable_issue(dut):
+    yield dut.aluissue.insn_i.eq(0)
+    yield dut.brissue.insn_i.eq(0)
+    yield dut.lsissue.insn_i.eq(0)
+
+
+def wait_for_issue(dut, dut_issue):
+    while True:
+        issue_o = yield dut_issue.fn_issue_o
+        if issue_o:
+            yield from disable_issue(dut)
+            yield dut.reg_enable_i.eq(0)
+            break
+        print ("busy",)
+        #yield from print_reg(dut, [1,2,3])
+        yield
+    #yield from print_reg(dut, [1,2,3])
+
+def scoreboard_branch_sim(dut, alusim):
+
+    iseed = 3
+
+    for i in range(1):
+
+        print ("rseed", iseed)
+        seed(iseed)
+        iseed += 1
+
+        yield dut.branch_direction_o.eq(0)
+
+        # set random values in the registers
+        for i in range(1, dut.n_regs):
+            val = 31+i*3
+            val = randint(0, (1<<alusim.rwidth)-1)
+            yield dut.intregs.regs[i].reg.eq(val)
+            alusim.setval(i, val)
+
+        if False:
+            # create some instructions: branches create a tree
+            insts = create_random_ops(dut, 1, True, 1)
+            #insts.append((6, 6, 1, 2, (0, 0)))
+            #insts.append((4, 3, 3, 0, (0, 0)))
+
+            src1 = randint(1, dut.n_regs-1)
+            src2 = randint(1, dut.n_regs-1)
+            #op = randint(4, 7)
+            op = 4 # only BGT at the moment
+
+            branch_ok = create_random_ops(dut, 1, True, 1)
+            branch_fail = create_random_ops(dut, 1, True, 1)
+
+            insts.append((src1, src2, (branch_ok, branch_fail), op, (0, 0)))
+
+        if True:
+            insts = []
+            insts.append( (3, 5, 2, 0, (0, 0)) )
+            branch_ok = []
+            branch_fail = []
+            #branch_ok.append  ( (5, 7, 5, 1, (1, 0)) )
+            branch_ok.append( None )
+            branch_fail.append( (1, 1, 2, 0, (0, 1)) )
+            #branch_fail.append( None )
+            insts.append( (6, 4, (branch_ok, branch_fail), 4, (0, 0)) )
+
+        siminsts = deepcopy(insts)
+
+        # issue instruction(s)
+        i = -1
+        instrs = insts
+        branch_direction = 0
+        while instrs:
+            yield
+            yield
+            i += 1
+            branch_direction = yield dut.branch_direction_o # way branch went
+            (src1, src2, dest, op, (shadow_on, shadow_off)) = insts.pop(0)
+            if branch_direction == 1 and shadow_on:
+                print ("skip", i, src1, src2, dest, op, shadow_on, shadow_off)
+                continue # branch was "success" and this is a "failed"... skip
+            if branch_direction == 2 and shadow_off:
+                print ("skip", i, src1, src2, dest, op, shadow_on, shadow_off)
+                continue # branch was "fail" and this is a "success"... skip
+            if branch_direction != 0:
+                shadow_on = 0
+                shadow_off = 0
+            is_branch = op >= 4
+            if is_branch:
+                branch_ok, branch_fail = dest
+                dest = src2
+                # ok zip up the branch success / fail instructions and
+                # drop them into the queue, one marked "to have branch success"
+                # the other to be marked shadow branch "fail".
+                # one out of each of these will be cancelled
+                for ok, fl in zip(branch_ok, branch_fail):
+                    if ok:
+                        instrs.append((ok[0], ok[1], ok[2], ok[3], (1, 0)))
+                    if fl:
+                        instrs.append((fl[0], fl[1], fl[2], fl[3], (0, 1)))
+            print ("instr %d: (%d, %d, %d, %d, (%d, %d))" % \
+                            (i, src1, src2, dest, op, shadow_on, shadow_off))
+            yield from int_instr(dut, op, src1, src2, dest,
+                                 shadow_on, shadow_off)
+
+        # wait for all instructions to stop before checking
+        yield
+        yield from wait_for_busy_clear(dut)
+
+        i = -1
+        while siminsts:
+            instr = siminsts.pop(0)
+            if instr is None:
+                continue
+            (src1, src2, dest, op, (shadow_on, shadow_off)) = instr
+            i += 1
+            is_branch = op >= 4
+            if is_branch:
+                branch_ok, branch_fail = dest
+                dest = src2
+            print ("sim %d: (%d, %d, %d, %d, (%d, %d))" % \
+                            (i, src1, src2, dest, op, shadow_on, shadow_off))
+            branch_res = alusim.op(op, src1, src2, dest)
+            if is_branch:
+                if branch_res:
+                    siminsts += branch_ok
+                else:
+                    siminsts += branch_fail
+
+        # check status
+        yield from alusim.check(dut)
+        yield from alusim.dump(dut)
+
+
 def scoreboard_sim(dut, alusim):
 
-    yield dut.int_store_i.eq(0)
+    seed(0)
 
     for i in range(1):
 
         # set random values in the registers
         for i in range(1, dut.n_regs):
-            yield dut.intregs.regs[i].reg.eq(31+i*3)
-            alusim.setval(i, 31+i*3)
+            val = randint(0, (1<<alusim.rwidth)-1)
+            #val = 31+i*3
+            #val = i
+            yield dut.intregs.regs[i].reg.eq(val)
+            alusim.setval(i, val)
 
         # create some instructions (some random, some regression tests)
         instrs = []
         if False:
-            for i in range(10):
-                src1 = randint(1, dut.n_regs-1)
-                src2 = randint(1, dut.n_regs-1)
-                while True:
-                    dest = randint(1, dut.n_regs-1)
-                    break
-                    if dest not in [src1, src2]:
-                        break
-                #src1 = 2
-                #src2 = 3
-                #dest = 2
-
-                op = randint(0, 2)
-                #op = i % 2
-                #op = 0
-
-                instrs.append((src1, src2, dest, op))
+            instrs = create_random_ops(dut, 15, True, 4)
+
+        if True: # LD/ST test (with immediate)
+            instrs.append( (1, 2, 2, 0x10, 1, 1, (0, 0)) )
+            instrs.append( (1, 2, 7, 0x12, 1, 1, (0, 0)) )
+
+        if False:
+            instrs.append( (1, 2, 2, 1, 1, 20, (0, 0)) )
+
+        if False:
+            instrs.append( (7, 3, 2, 4, (0, 0)) )
+            instrs.append( (7, 6, 6, 2, (0, 0)) )
+            instrs.append( (1, 7, 2, 2, (0, 0)) )
+
+        if False:
+            instrs.append((2, 3, 3, 0, 0, 0, (0, 0)))
+            instrs.append((5, 3, 3, 1, 0, 0, (0, 0)))
+            instrs.append((3, 5, 5, 2, 0, 0, (0, 0)))
+            instrs.append((5, 3, 3, 3, 0, 0, (0, 0)))
+            instrs.append((3, 5, 5, 0, 0, 0, (0, 0)))
 
         if False:
-            instrs.append((2, 3, 3, 0))
-            instrs.append((5, 3, 3, 1))
+            instrs.append( (3, 3, 4, 0, 0, 13979, (0, 0)))
+            instrs.append( (6, 4, 1, 2, 0, 40976, (0, 0)))
+            instrs.append( (1, 4, 7, 4, 1, 23652, (0, 0)))
 
         if False:
             instrs.append((5, 6, 2, 1))
@@ -456,65 +1215,72 @@ def scoreboard_sim(dut, alusim):
             instrs.append((1, 1, 1, 1))
             instrs.append((1, 5, 3, 0))
 
-        if True:
-            instrs.append( (7, 1, 2, 0) )
-            instrs.append( (1, 1, 4, 2) )
-            instrs.append( (2, 3, 2, 2) )
-            instrs.append( (5, 3, 1, 0) )
-            instrs.append( (7, 3, 5, 2) )
-            instrs.append( (1, 2, 6, 2) )
+        if False:
+            # very weird failure
             instrs.append( (5, 2, 5, 2) )
-            instrs.append( (2, 2, 3, 0) )
+            instrs.append( (2, 6, 3, 0) )
             instrs.append( (4, 2, 2, 1) )
-            instrs.append( (2, 4, 6, 1) )
+
+        if False:
+            v1 = 4
+            yield dut.intregs.regs[5].reg.eq(v1)
+            alusim.setval(5, v1)
+            yield dut.intregs.regs[3].reg.eq(5)
+            alusim.setval(3, 5)
+            instrs.append((5, 3, 3, 4, (0, 0)))
+            instrs.append((4, 2, 1, 2, (0, 1)))
+
+        if False:
+            v1 = 6
+            yield dut.intregs.regs[5].reg.eq(v1)
+            alusim.setval(5, v1)
+            yield dut.intregs.regs[3].reg.eq(5)
+            alusim.setval(3, 5)
+            instrs.append((5, 3, 3, 4, (0, 0)))
+            instrs.append((4, 2, 1, 2, (1, 0)))
+
+        if False:
+            instrs.append( (4, 3, 5, 1, 0, (0, 0)) )
+            instrs.append( (5, 2, 3, 1, 0, (0, 0)) )
+            instrs.append( (7, 1, 5, 2, 0, (0, 0)) )
+            instrs.append( (5, 6, 6, 4, 0, (0, 0)) )
+            instrs.append( (7, 5, 2, 2, 0, (1, 0)) )
+            instrs.append( (1, 7, 5, 0, 0, (0, 1)) )
+            instrs.append( (1, 6, 1, 2, 0, (1, 0)) )
+            instrs.append( (1, 6, 7, 3, 0, (0, 0)) )
+            instrs.append( (6, 7, 7, 0, 0, (0, 0)) )
 
         # issue instruction(s), wait for issue to be free before proceeding
-        for i, (src1, src2, dest, op) in enumerate(instrs):
+        for i, instr in enumerate(instrs):
+            src1, src2, dest, op, opi, imm, (br_ok, br_fail) = instr
 
-            print ("instr %d: (%d, %d, %d, %d)" % (i, src1, src2, dest, op))
-            yield from int_instr(dut, alusim, op, src1, src2, dest)
-            yield
-            while True:
-                issue_o = yield dut.issue_o
-                if issue_o:
-                    for i in range(len(dut.int_insn_i)):
-                        yield dut.int_insn_i[i].eq(0)
-                        yield dut.reg_enable_i.eq(0)
-                    break
-                #print ("busy",)
-                #yield from print_reg(dut, [1,2,3])
-                yield
-            #yield from print_reg(dut, [1,2,3])
+            print ("instr %d: (%d, %d, %d, %d, %d, %d)" % \
+                    (i, src1, src2, dest, op, opi, imm))
+            alusim.op(op, opi, imm, src1, src2, dest)
+            yield from instr_q(dut, op, opi, imm, src1, src2, dest,
+                               br_ok, br_fail)
 
         # wait for all instructions to stop before checking
-        yield
         while True:
-            busy_o = yield dut.busy_o
-            if not busy_o:
+            iqlen = yield dut.qlen_o
+            if iqlen == 0:
                 break
-            print ("busy",)
             yield
+        yield
+        yield
+        yield
+        yield
+        yield from wait_for_busy_clear(dut)
 
         # check status
         yield from alusim.check(dut)
         yield from alusim.dump(dut)
 
 
-def explore_groups(dut):
-    from nmigen.hdl.ir import Fragment
-    from nmigen.hdl.xfrm import LHSGroupAnalyzer
-
-    fragment = dut.elaborate(platform=None)
-    fr = Fragment.get(fragment, platform=None)
-
-    groups = LHSGroupAnalyzer()(fragment._statements)
-
-    print (groups)
-
-
 def test_scoreboard():
-    dut = Scoreboard(16, 8)
+    dut = IssueToScoreboard(2, 1, 1, 16, 8, 8)
     alusim = RegSim(16, 8)
+    memsim = MemSim(16, 16)
     vl = rtlil.convert(dut, ports=dut.ports())
     with open("test_scoreboard6600.il", "w") as f:
         f.write(vl)
@@ -522,6 +1288,9 @@ def test_scoreboard():
     run_simulation(dut, scoreboard_sim(dut, alusim),
                         vcd_name='test_scoreboard6600.vcd')
 
+    #run_simulation(dut, scoreboard_branch_sim(dut, alusim),
+    #                    vcd_name='test_scoreboard6600.vcd')
+
 
 if __name__ == '__main__':
     test_scoreboard()