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
7 # Copyright (C) 2020 Michael Nolan <mtnolan2640@gmail.com>
8 from nmigen
import (Module
, Signal
, Cat
, Repl
, Mux
, Const
)
9 from nmutil
.pipemodbase
import PipeModBase
10 from nmutil
.extend
import exts
, extz
11 from soc
.fu
.alu
.pipe_data
import ALUInputData
, ALUOutputData
12 from ieee754
.part
.partsig
import PartitionedSignal
13 from soc
.decoder
.power_enums
import MicrOp
15 from soc
.decoder
.power_fields
import DecodeFields
16 from soc
.decoder
.power_fieldsn
import SignalBitRange
19 # microwatt calc_ov function.
20 def calc_ov(msb_a
, msb_b
, ca
, msb_r
):
21 return (ca ^ msb_r
) & ~
(msb_a ^ msb_b
)
24 class ALUMainStage(PipeModBase
):
25 def __init__(self
, pspec
):
26 super().__init
__(pspec
, "main")
27 self
.fields
= DecodeFields(SignalBitRange
, [self
.i
.ctx
.op
.insn
])
28 self
.fields
.create_specs()
31 return ALUInputData(self
.pspec
) # defines pipeline stage input format
34 return ALUOutputData(self
.pspec
) # defines pipeline stage output format
36 def elaborate(self
, platform
):
40 # convenience variables
41 cry_o
, o
, cr0
= self
.o
.xer_ca
, self
.o
.o
, self
.o
.cr0
42 xer_so_i
, ov_o
= self
.i
.xer_so
, self
.o
.xer_ov
43 a
, b
, cry_i
, op
= self
.i
.a
, self
.i
.b
, self
.i
.xer_ca
, self
.i
.ctx
.op
45 # get L-field for OP_CMP
46 x_fields
= self
.fields
.FormX
49 # check if op is 32-bit, and get sign bit from operand a
50 is_32bit
= Signal(reset_less
=True)
52 with m
.If(op
.insn_type
== MicrOp
.OP_CMP
):
53 comb
+= is_32bit
.eq(~L
)
55 # little trick: do the add using only one add (not 2)
56 # LSB: carry-in [0]. op/result: [1:-1]. MSB: carry-out [-1]
57 add_a
= Signal(a
.width
+ 2, reset_less
=True)
58 add_b
= Signal(a
.width
+ 2, reset_less
=True)
59 add_o
= Signal(a
.width
+ 2, reset_less
=True)
64 with m
.If(op
.is_signed
):
65 comb
+= a_i
.eq(exts(a
, 32, 64))
66 comb
+= b_i
.eq(exts(b
, 32, 64))
68 comb
+= a_i
.eq(extz(a
, 32, 64))
69 comb
+= b_i
.eq(extz(b
, 32, 64))
74 with m
.If((op
.insn_type
== MicrOp
.OP_ADD
) |
75 (op
.insn_type
== MicrOp
.OP_CMP
)):
76 # in bit 0, 1+carry_in creates carry into bit 1 and above
77 comb
+= add_a
.eq(Cat(cry_i
[0], a_i
, Const(0, 1)))
78 comb
+= add_b
.eq(Cat(Const(1, 1), b_i
, Const(0, 1)))
79 comb
+= add_o
.eq(add_a
+ add_b
)
81 ##########################
82 # main switch-statement for handling arithmetic operations
84 with m
.Switch(op
.insn_type
):
87 #### CMP, CMPL v3.0B p85-86
89 with m
.Case(MicrOp
.OP_CMP
):
90 a_n
= Signal(64) # temporary - inverted a
101 # this is supposed to be inverted (b-a, not a-b)
102 comb
+= a_n
.eq(~a
) # sigh a gets inverted
103 comb
+= carry_32
.eq(add_o
[33] ^ a
[32] ^ b
[32])
104 comb
+= carry_64
.eq(add_o
[65])
106 comb
+= zerolo
.eq(~
((a_n
[0:32] ^ b
[0:32]).bool()))
107 comb
+= zerohi
.eq(~
((a_n
[32:64] ^ b
[32:64]).bool()))
109 with m
.If(zerolo
& (is_32bit | zerohi
)):
111 comb
+= tval
[2].eq(1)
113 comb
+= msb_a
.eq(Mux(is_32bit
, a_n
[31], a_n
[63]))
114 comb
+= msb_b
.eq(Mux(is_32bit
, b
[31], b
[63]))
116 with m
.If(msb_a
!= msb_b
):
117 # Subtraction might overflow, but
118 # comparison is clear from MSB difference.
119 # for signed, 0 is greater; for unsigned, 1 is greater
120 comb
+= tval
.eq(Cat(msb_a
, msb_b
, C0
, msb_b
, msb_a
))
122 # Subtraction cannot overflow since MSBs are equal.
123 # carry = 1 indicates RA is smaller (signed or unsigned)
124 comb
+= a_lt
.eq(Mux(is_32bit
, carry_32
, carry_64
))
125 comb
+= tval
.eq(Cat(~a_lt
, a_lt
, C0
, ~a_lt
, a_lt
))
126 comb
+= cr0
.data
[0:2].eq(Cat(xer_so_i
[0], tval
[2]))
127 with m
.If(op
.is_signed
):
128 comb
+= cr0
.data
[2:4].eq(tval
[3:5])
130 comb
+= cr0
.data
[2:4].eq(tval
[0:2])
134 #### add v3.0B p67, p69-72
136 with m
.Case(MicrOp
.OP_ADD
):
137 # bit 0 is not part of the result, top bit is the carry-out
138 comb
+= o
.data
.eq(add_o
[1:-1])
139 comb
+= o
.ok
.eq(1) # output register
141 # see microwatt OP_ADD code
142 # https://bugs.libre-soc.org/show_bug.cgi?id=319#c5
143 ca
= Signal(2, reset_less
=True)
144 comb
+= ca
[0].eq(add_o
[-1]) # XER.CA
145 comb
+= ca
[1].eq(add_o
[33] ^
(a_i
[32] ^ b_i
[32])) # XER.CA32
146 comb
+= cry_o
.data
.eq(ca
)
147 comb
+= cry_o
.ok
.eq(1)
148 # 32-bit (ov[1]) and 64-bit (ov[0]) overflow
149 ov
= Signal(2, reset_less
=True)
150 comb
+= ov
[0].eq(calc_ov(a_i
[-1], b_i
[-1], ca
[0], add_o
[-2]))
151 comb
+= ov
[1].eq(calc_ov(a_i
[31], b_i
[31], ca
[1], add_o
[32]))
152 comb
+= ov_o
.data
.eq(ov
)
153 comb
+= ov_o
.ok
.eq(1)
156 #### exts (sign-extend) v3.0B p96, p99
158 with m
.Case(MicrOp
.OP_EXTS
):
159 with m
.If(op
.data_len
== 1):
160 comb
+= o
.data
.eq(exts(a
, 8, 64))
161 with m
.If(op
.data_len
== 2):
162 comb
+= o
.data
.eq(exts(a
, 16, 64))
163 with m
.If(op
.data_len
== 4):
164 comb
+= o
.data
.eq(exts(a
, 32, 64))
165 comb
+= o
.ok
.eq(1) # output register
168 #### cmpeqb v3.0B p88
170 with m
.Case(MicrOp
.OP_CMPEQB
):
171 eqs
= Signal(8, reset_less
=True)
172 src1
= Signal(8, reset_less
=True)
173 comb
+= src1
.eq(a
[0:8])
175 comb
+= eqs
[i
].eq(src1
== b
[8*i
:8*(i
+1)])
176 comb
+= o
.data
[0].eq(eqs
.any())
177 comb
+= o
.ok
.eq(0) # use o.data but do *not* actually output
178 comb
+= cr0
.data
.eq(Cat(Const(0, 2), eqs
.any(), Const(0, 1)))
181 ###### sticky overflow and context, both pass-through #####
183 comb
+= self
.o
.xer_so
.data
.eq(xer_so_i
)
184 comb
+= self
.o
.ctx
.eq(self
.i
.ctx
)