Allow the formal engine to perform a same-cycle result in the ALU
[soc.git] / 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 nmutil.extend import exts
9 from soc.fu.alu.pipe_data import ALUInputData, ALUOutputData
10 from ieee754.part.partsig import PartitionedSignal
11 from soc.decoder.power_enums import MicrOp
12
13
14 # microwatt calc_ov function.
15 def calc_ov(msb_a, msb_b, ca, msb_r):
16 return (ca ^ msb_r) & ~(msb_a ^ msb_b)
17
18
19 class ALUMainStage(PipeModBase):
20 def __init__(self, pspec):
21 super().__init__(pspec, "main")
22
23 def ispec(self):
24 return ALUInputData(self.pspec) # defines pipeline stage input format
25
26 def ospec(self):
27 return ALUOutputData(self.pspec) # defines pipeline stage output format
28
29 def elaborate(self, platform):
30 m = Module()
31 comb = m.d.comb
32
33 # convenience variables
34 cry_o, o, cr0 = self.o.xer_ca, self.o.o, self.o.cr0
35 ov_o = self.o.xer_ov
36 a, b, cry_i, op = self.i.a, self.i.b, self.i.xer_ca, self.i.ctx.op
37
38 # check if op is 32-bit, and get sign bit from operand a
39 is_32bit = Signal(reset_less=True)
40 sign_bit = Signal(reset_less=True)
41 comb += is_32bit.eq(op.is_32bit)
42 comb += sign_bit.eq(Mux(is_32bit, a[31], a[63]))
43
44 # little trick: do the add using only one add (not 2)
45 # LSB: carry-in [0]. op/result: [1:-1]. MSB: carry-out [-1]
46 add_a = Signal(a.width + 2, reset_less=True)
47 add_b = Signal(a.width + 2, reset_less=True)
48 add_o = Signal(a.width + 2, reset_less=True)
49 with m.If((op.insn_type == MicrOp.OP_ADD) |
50 (op.insn_type == MicrOp.OP_CMP)):
51 # in bit 0, 1+carry_in creates carry into bit 1 and above
52 comb += add_a.eq(Cat(cry_i[0], a, Const(0, 1)))
53 comb += add_b.eq(Cat(Const(1, 1), b, Const(0, 1)))
54 comb += add_o.eq(add_a + add_b)
55
56 ##########################
57 # main switch-statement for handling arithmetic operations
58
59 with m.Switch(op.insn_type):
60 #### CMP, CMPL ####
61 with m.Case(MicrOp.OP_CMP):
62 # this is supposed to be inverted (b-a, not a-b)
63 # however we have a trick: instead of adding either 2x 64-bit
64 # MUXes to invert a and b, or messing with a 64-bit output,
65 # swap +ve and -ve test in the *output* stage using an XOR gate
66 comb += o.data.eq(add_o[1:-1])
67 comb += o.ok.eq(0) # use o.data but do *not* actually output
68
69 #### add ####
70 with m.Case(MicrOp.OP_ADD):
71 # bit 0 is not part of the result, top bit is the carry-out
72 comb += o.data.eq(add_o[1:-1])
73 comb += o.ok.eq(1) # output register
74
75 # see microwatt OP_ADD code
76 # https://bugs.libre-soc.org/show_bug.cgi?id=319#c5
77 ca = Signal(2, reset_less=True)
78 comb += ca[0].eq(add_o[-1]) # XER.CA
79 comb += ca[1].eq(add_o[33] ^ (a[32] ^ b[32])) # XER.CA32
80 comb += cry_o.data.eq(ca)
81 comb += cry_o.ok.eq(1)
82 # 32-bit (ov[1]) and 64-bit (ov[0]) overflow
83 ov = Signal(2, reset_less=True)
84 comb += ov[0].eq(calc_ov(a[-1], b[-1], ca[0], add_o[-2]))
85 comb += ov[1].eq(calc_ov(a[31], b[31], ca[1], add_o[32]))
86 comb += ov_o.data.eq(ov)
87 comb += ov_o.ok.eq(1)
88
89 #### exts (sign-extend) ####
90 with m.Case(MicrOp.OP_EXTS):
91 with m.If(op.data_len == 1):
92 comb += o.data.eq(exts(a, 8, 64))
93 with m.If(op.data_len == 2):
94 comb += o.data.eq(exts(a, 16, 64))
95 with m.If(op.data_len == 4):
96 comb += o.data.eq(exts(a, 32, 64))
97 comb += o.ok.eq(1) # output register
98
99 #### cmpeqb ####
100 with m.Case(MicrOp.OP_CMPEQB):
101 eqs = Signal(8, reset_less=True)
102 src1 = Signal(8, reset_less=True)
103 comb += src1.eq(a[0:8])
104 for i in range(8):
105 comb += eqs[i].eq(src1 == b[8*i:8*(i+1)])
106 comb += o.data[0].eq(eqs.any())
107 comb += o.ok.eq(0) # use o.data but do *not* actually output
108 comb += cr0.data.eq(Cat(Const(0, 2), eqs.any(), Const(0, 1)))
109 comb += cr0.ok.eq(1)
110
111 ###### sticky overflow and context, both pass-through #####
112
113 comb += self.o.xer_so.data.eq(self.i.xer_so)
114 comb += self.o.ctx.eq(self.i.ctx)
115
116 return m