src/soc/fu/mul/post_stage.py

   1 # This stage is intended to do most of the work of analysing the multiply result
   2
   3 from nmigen import (Module, Signal, Cat, Repl, Mux, signed)
   4 from nmutil.pipemodbase import PipeModBase
   5 from soc.fu.alu.pipe_data import ALUOutputData
   6 from soc.fu.mul.pipe_data import MulOutputData
   7 from ieee754.part.partsig import PartitionedSignal
   8 from soc.decoder.power_enums import InternalOp
   9
  10
  11 class MulMainStage3(PipeModBase):
  12     def __init__(self, pspec):
  13         super().__init__(pspec, "mul3")
  14
  15     def ispec(self):
  16         return MulOutputData(self.pspec) # pipeline stage output format
  17
  18     def ospec(self):
  19         return ALUOutputData(self.pspec) # defines pipeline stage output format
  20
  21     def elaborate(self, platform):
  22         m = Module()
  23         comb = m.d.comb
  24
  25         # convenience variables
  26         cry_o, o, cr0 = self.o.xer_ca, self.o.o, self.o.cr0
  27         ov_o = self.o.xer_ov
  28         o_i, cry_i, op = self.i.o, self.i.xer_ca, self.i.ctx.op
  29
  30         # check if op is 32-bit, and get sign bit from operand a
  31         is_32bit = Signal(reset_less=True)
  32         comb += is_32bit.eq(op.is_32bit)
  33
  34         # check negate: select signed/unsigned
  35         mul_o = Signal(o_i.width, reset_less=True)
  36         comb += mul_o.eq(Mux(self.i.neg_res, -o_i, o_i))
  37         comb += o.ok.eq(1)
  38
  39         # OP_MUL_nnn - select hi32/hi64/lo64 from result
  40         with m.Switch(op.insn_type):
  41             # hi-32 replicated twice
  42             with m.Case(InternalOp.OP_MUL_H32):
  43                 comb += o.data.eq(Repl(mul_o[32:64], 2))
  44             # hi-64
  45             with m.Case(InternalOp.OP_MUL_H64):
  46                 comb += o.data.eq(mul_o[64:128])
  47             # lo-64 - overflow
  48             with m.Default():
  49                 comb += o.data.eq(mul_o[0:64])
  50
  51                 # compute overflow
  52                 mul_ov = Signal(reset_less=True)
  53                 with m.If(is_32bit):
  54                     m32 = mul_o[32:64]
  55                     comb += mul_ov.eq(m32.bool() & ~m32.all())
  56                 with m.Else():
  57                     m64 = mul_o[64:128]
  58                     comb += mul_ov.eq(m64.bool() & ~m64.all())
  59
  60                 # 32-bit (ov[1]) and 64-bit (ov[0]) overflow
  61                 ov = Signal(2, reset_less=True)
  62                 comb += ov[0].eq(mul_ov)
  63                 comb += ov[1].eq(mul_ov)
  64                 comb += ov_o.data.eq(ov)
  65                 comb += ov_o.ok.eq(1)
  66
  67         # https://bugs.libre-soc.org/show_bug.cgi?id=319#c5
  68         ca = Signal(2, reset_less=True)
  69         comb += ca[0].eq(mul_o[-1])                      # XER.CA - XXX more?
  70         comb += ca[1].eq(mul_o[32] ^ (self.i.neg_res32)) # XER.CA32
  71         comb += cry_o.data.eq(ca)
  72         comb += cry_o.ok.eq(1)
  73
  74         ###### sticky overflow and context, both pass-through #####
  75
  76         comb += self.o.xer_so.data.eq(self.i.xer_so)
  77         comb += self.o.ctx.eq(self.i.ctx)
  78
  79         return m
  80