src/soc/fu/alu/main_stage.py

   1 # This stage is intended to do most of the work of executing the Arithmetic
   2 # instructions. This would be like the additions, compares, and sign-extension
   3 # as well as carry and overflow generation. This module
   4 # however should not gate the carry or overflow, that's up to the
   5 # output stage
   6 from nmigen import (Module, Signal, Cat, Repl, Mux, Const)
   7 from nmutil.pipemodbase import PipeModBase
   8 from soc.fu.alu.pipe_data import ALUInputData, ALUOutputData
   9 from ieee754.part.partsig import PartitionedSignal
  10 from soc.decoder.power_enums import InternalOp
  11
  12
  13 class ALUMainStage(PipeModBase):
  14     def __init__(self, pspec):
  15         super().__init__(pspec, "main")
  16
  17     def ispec(self):
  18         return ALUInputData(self.pspec)
  19
  20     def ospec(self):
  21         return ALUOutputData(self.pspec) # TODO: ALUIntermediateData
  22
  23     def elaborate(self, platform):
  24         m = Module()
  25         comb = m.d.comb
  26         carry_out, o = self.o.xer_co, self.o.o
  27
  28         # check if op is 32-bit, and get sign bit from operand a
  29         is_32bit = Signal(reset_less=True)
  30         sign_bit = Signal(reset_less=True)
  31         comb += is_32bit.eq(self.i.ctx.op.is_32bit)
  32         comb += sign_bit.eq(Mux(is_32bit, self.i.a[31], self.i.a[63]))
  33
  34         # little trick: do the add using only one add (not 2)
  35         add_a = Signal(self.i.a.width + 2, reset_less=True)
  36         add_b = Signal(self.i.a.width + 2, reset_less=True)
  37         add_output = Signal(self.i.a.width + 2, reset_less=True)
  38         with m.If((self.i.ctx.op.insn_type == InternalOp.OP_ADD) |
  39                   (self.i.ctx.op.insn_type == InternalOp.OP_CMP)):
  40             # in bit 0, 1+carry_in creates carry into bit 1 and above
  41             comb += add_a.eq(Cat(self.i.carry_in, self.i.a, Const(0, 1)))
  42             comb += add_b.eq(Cat(Const(1, 1), self.i.b, Const(0, 1)))
  43             comb += add_output.eq(add_a + add_b)
  44
  45         ##########################
  46         # main switch-statement for handling arithmetic operations
  47
  48         with m.Switch(self.i.ctx.op.insn_type):
  49             #### CMP, CMPL ####
  50             with m.Case(InternalOp.OP_CMP):
  51                 # this is supposed to be inverted (b-a, not a-b)
  52                 # however we have a trick: instead of adding either 2x 64-bit
  53                 # MUXes to invert a and b, or messing with a 64-bit output,
  54                 # swap +ve and -ve test in the *output* stage using an XOR gate
  55                 comb += o.eq(add_output[1:-1])
  56
  57             #### add ####
  58             with m.Case(InternalOp.OP_ADD):
  59                 # bit 0 is not part of the result, top bit is the carry-out
  60                 comb += o.eq(add_output[1:-1])
  61                 comb += carry_out.data[0].eq(add_output[-1]) # XER.CO
  62
  63                 # XXX no!  wrongggg, see microwatt OP_ADD code
  64                 # https://bugs.libre-soc.org/show_bug.cgi?id=319#c5
  65                 comb += carry_out.data[1].eq(add_output[-1]) # XER.CO32
  66
  67             #### exts (sign-extend) ####
  68             with m.Case(InternalOp.OP_EXTS):
  69                 with m.If(self.i.ctx.op.data_len == 1):
  70                     comb += o.eq(Cat(self.i.a[0:8], Repl(self.i.a[7], 64-8)))
  71                 with m.If(self.i.ctx.op.data_len == 2):
  72                     comb += o.eq(Cat(self.i.a[0:16], Repl(self.i.a[15], 64-16)))
  73                 with m.If(self.i.ctx.op.data_len == 4):
  74                     comb += o.eq(Cat(self.i.a[0:32], Repl(self.i.a[31], 64-32)))
  75             with m.Case(InternalOp.OP_CMPEQB):
  76                 eqs = Signal(8, reset_less=True)
  77                 src1 = Signal(8, reset_less=True)
  78                 comb += src1.eq(self.i.a[0:8])
  79                 for i in range(8):
  80                     comb += eqs[i].eq(src1 == self.i.b[8*i:8*(i+1)])
  81                 comb += self.o.cr0.eq(Cat(Const(0, 2), eqs.any(), Const(0, 1)))
  82
  83         ###### sticky overflow and context, both pass-through #####
  84
  85         comb += self.o.xer_so.data.eq(self.i.so)
  86         comb += self.o.ctx.eq(self.i.ctx)
  87
  88         return m