1 """ Load / Store partial address matcher
3 Loads and Stores do not need a full match (CAM), they need "good enough"
4 avoidance. Around 11 bits on a 64-bit address is "good enough".
6 The simplest way to use this module is to ignore not only the top bits,
7 but also the bottom bits as well: in this case (this RV64 processor),
8 enough to cover a DWORD (64-bit). that means ignore the bottom 4 bits,
9 due to the possibility of 64-bit LD/ST being misaligned.
11 To reiterate: the use of this module is an *optimisation*. All it has
12 to do is cover the cases that are *definitely* matches (by checking 11
13 bits or so), and if a few opportunities for parallel LD/STs are missed
14 because the top (or bottom) bits weren't checked, so what: all that
15 happens is: the mis-matched addresses are LD/STd on single-cycles. Big Deal.
17 However, if we wanted to enhance this algorithm (without using a CAM and
18 without using expensive comparators) probably the best way to do so would
19 be to turn the last 16 bits into a byte-level bitmap. LD/ST on a byte
20 would have 1 of the 16 bits set. LD/ST on a DWORD would have 8 of the 16
21 bits set (offset if the LD/ST was misaligned). TODO.
25 > I have used bits <11:6> as they are not translated (4KB pages)
26 > and larger than a cache line (64 bytes).
27 > I have used bits <11:4> when the L1 cache was QuadW sized and
28 > the L2 cache was Line sized.
31 from nmigen
.compat
.sim
import run_simulation
32 from nmigen
.cli
import verilog
, rtlil
33 from nmigen
import Module
, Signal
, Const
, Array
, Cat
, Elaboratable
34 from nmigen
.lib
.coding
import Decoder
36 from nmutil
.latch
import latchregister
, SRLatch
39 class PartialAddrMatch(Elaboratable
):
40 """A partial address matcher
42 def __init__(self
, n_adr
, bitwid
):
46 self
.addrs_i
= Array(Signal(bitwid
, name
="addr") for i
in range(n_adr
))
47 self
.addr_we_i
= Signal(n_adr
) # write-enable for incoming address
48 self
.addr_en_i
= Signal(n_adr
) # address latched in
49 self
.addr_rs_i
= Signal(n_adr
) # address deactivated
52 self
.addr_nomatch_o
= Signal(n_adr
, name
="nomatch_o")
53 self
.addr_nomatch_a_o
= Array(Signal(n_adr
, name
="nomatch_array_o") \
54 for i
in range(n_adr
))
56 def elaborate(self
, platform
):
58 return self
._elaborate
(m
, platform
)
60 def _elaborate(self
, m
, platform
):
64 # array of address-latches
65 m
.submodules
.l
= self
.l
= l
= SRLatch(llen
=self
.n_adr
, sync
=False)
66 self
.addrs_r
= addrs_r
= Array(Signal(self
.bitwid
, name
="a_r") \
67 for i
in range(self
.n_adr
))
70 comb
+= l
.s
.eq(self
.addr_en_i
)
71 comb
+= l
.r
.eq(self
.addr_rs_i
)
73 # copy in addresses (and "enable" signals)
74 for i
in range(self
.n_adr
):
75 latchregister(m
, self
.addrs_i
[i
], addrs_r
[i
], l
.q
[i
])
77 # is there a clash, yes/no
79 for i
in range(self
.n_adr
):
81 for j
in range(self
.n_adr
):
82 match
.append(self
.is_match(i
, j
))
83 comb
+= self
.addr_nomatch_a_o
[i
].eq(~
Cat(*match
) & l
.q
)
84 matchgrp
.append(self
.addr_nomatch_a_o
[i
] == l
.q
)
85 comb
+= self
.addr_nomatch_o
.eq(Cat(*matchgrp
) & l
.q
)
89 def is_match(self
, i
, j
):
91 return Const(0) # don't match against self!
92 return self
.addrs_r
[i
] == self
.addrs_r
[j
]
95 yield from self
.addrs_i
98 yield from self
.addr_nomatch_a_o
99 yield self
.addr_nomatch_o
105 class PartialAddrBitmap(PartialAddrMatch
):
106 def __init__(self
, n_adr
, bitwid
, bit_len
):
107 PartialAddrMatch
.__init
__(self
, n_adr
, bitwid
)
108 self
.bitlen
= bitlen
# number of bits to turn into unary
110 # inputs: length of the LOAD/STORE
111 self
.len_i
= Array(Signal(bitwid
, name
="len") for i
in range(n_adr
))
113 def elaborate(self
, platform
):
114 m
= PartialAddrMatch
.elaborate(self
, platform
)
117 addrs_r
, l
= self
.addrs_r
, self
.l
118 expwid
= 8 + (1<<self
.bitlen
) # XXX assume LD/ST no greater than 8
119 explen_i
= Array(Signal(expwid
, name
="a_l") \
120 for i
in range(self
.n_adr
))
121 lenexp_r
= Array(Signal(expwid
, name
="a_l") \
122 for i
in range(self
.n_adr
))
124 # the mapping between length, address and lenexp_r is that the
125 # length and address creates a bytemap which a LD/ST covers.
128 # copy in lengths and latch them
129 for i
in range(self
.n_adr
):
130 latchregister(m
, explen_i
[i
], lenexp_r
[i
], l
.q
[i
])
135 def part_addr_sim(dut
):
136 yield dut
.dest_i
.eq(1)
137 yield dut
.issue_i
.eq(1)
139 yield dut
.issue_i
.eq(0)
141 yield dut
.src1_i
.eq(1)
142 yield dut
.issue_i
.eq(1)
144 yield dut
.issue_i
.eq(0)
146 yield dut
.go_rd_i
.eq(1)
148 yield dut
.go_rd_i
.eq(0)
150 yield dut
.go_wr_i
.eq(1)
152 yield dut
.go_wr_i
.eq(0)
155 def test_part_addr():
156 dut
= PartialAddrMatch(3, 10)
157 vl
= rtlil
.convert(dut
, ports
=dut
.ports())
158 with
open("test_part_addr.il", "w") as f
:
161 run_simulation(dut
, part_addr_sim(dut
), vcd_name
='test_part_addr.vcd')
163 if __name__
== '__main__':