1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
|
(*
* Partition a cluster into multiple smaller clusters for region-based
* register allocation.
*)
functor ClusterPartitioner
(structure Flowgraph : FLOWGRAPH
structure InsnProps : INSN_PROPERTIES
sharing Flowgraph.I = InsnProps.I
) : RA_FLOWGRAPH_PARTITIONER =
struct
structure F = Flowgraph
structure I = F.I
structure C = I.C
structure PQ = PriorityQueue
structure Liveness = Liveness(Flowgraph)
structure A = Array
type flowgraph = F.cluster
val debug = true
fun error msg = MLRiscErrorMsg.error("ClusterPartitioner",msg)
val maxSize = MLRiscControl.getInt "ra-max-region-size"
val _ = maxSize := 300
fun numberOfBlocks(F.CLUSTER{blkCounter,...}) = !blkCounter
(*
* Partition the cluster into a set of clusters so that each can
* be allocated independently.
*)
fun partition(F.CLUSTER{blkCounter, blocks, entry, exit,
annotations, ...})
cellkind processRegion =
(* Number of basic blocks *)
let val N = !blkCounter
val _ = if debug then
print("[Region based register allocation: "^
Int.toString N^"]\n")
else ()
val maxSize = !maxSize
(* Perform global liveness analysis first.
* Unfortunately, I know of no way of avoiding this step because
* we have to know which values are live across regions.
*)
val _ = Liveness.liveness{blocks=blocks,
defUse=InsnProps.defUse cellkind,
getCell=C.getCellsByKind cellkind,
updateCell=C.updateCellsByKind cellkind
}
val F.ENTRY{succ=entrySucc, ...} = entry
val F.EXIT{pred=exitPred, ...} = exit
val initTrail = [(entrySucc,!entrySucc), (exitPred, !exitPred)]
(* Priority queue of basic blocks in non-increasing order
* of execution frequency
*)
fun higherFreq(F.BBLOCK{freq=a,...}, F.BBLOCK{freq=b,...}) = !a > !b
| higherFreq _ = error "higherFreq"
val blocks = List.foldr (fn (b as F.BBLOCK _,l) => b::l | (_,l) => l)
[] blocks
val seedQueue = PQ.fromList higherFreq blocks
(* Current region id *)
val regionCounter = ref 0
fun newRegionId() =
let val regionId = !regionCounter
in regionCounter := !regionCounter + 1; regionId end
(* Has the block been included in any region?
* Non-negative means yes. The number is the region id in which
* the block belongs.
*)
val processed = A.array(N, ~1)
fun hasBeenProcessed n = A.sub(processed,n) >= 0
fun markAsProcessed(n, regionId) = A.update(processed,n,regionId)
(* Get an unprocessed seed block from the queue *)
fun getSeedBlock(regionId) =
case PQ.deleteMin seedQueue of
block as F.BBLOCK{blknum, insns, ...} =>
if hasBeenProcessed blknum then getSeedBlock(regionId)
else block
| _ => error "getSeedBlock"
fun resetTrail [] = ()
| resetTrail((r,x)::trail) = (r := x; resetTrail trail)
(*
* Grow a region. Currently, region growth is limited only by size.
* Note that we only select nodes with one out edges as possible
* region cut points. We also try not to make a region too small
* as it will waste initialization time. It's a delicate balance.
*)
fun growRegion() =
let val regionId = newRegionId()
fun add([], Q) = Q
| add((b as F.BBLOCK{blknum, ...},_)::bs, Q) =
if hasBeenProcessed blknum then add(bs, Q)
else add(bs, b::Q)
| add(_::bs, Q) = add(bs, Q)
fun grow((b as F.BBLOCK{blknum, succ, pred, insns, ...})::F, B,
size, blks, m) =
if hasBeenProcessed blknum
then grow(F, B, size, blks, m)
else
let val n = length(!insns)
val newSize = size + n
in if m > 0 andalso newSize > maxSize andalso length(!succ) = 1
then grow(F, B, size, blks, m)
else (markAsProcessed(blknum, regionId);
grow(F, add(!pred,add(!succ,B)), newSize,
b::blks, m+1)
)
end
| grow([], [], size, blks, m) = (size, blks, m)
| grow([], B, size, blks, m) = grow(rev B, [], size, blks, m)
| grow _ = error "grow"
(* Find a seed block *)
val seed = getSeedBlock(regionId)
(* Grow until we reach some limit *)
val (totalSize, blocks, blockCount) = grow([seed], [], 0, [], 0)
(* Now create a cluster with only these blocks
* We have to update the edges so that region-entry edges
* are made into entry edges and region-exit edges are
* made into exit edges.
*)
fun makeSubgraph(blocks) =
let fun inSubgraph(y) = A.sub(processed,y) = regionId
fun processSucc(b,x,(e as (F.BBLOCK{blknum=y, ...},freq))::es,
es', exit, exitFreq) =
if inSubgraph(y) then
processSucc(b,x,es,e::es',exit,exitFreq)
else processSucc(b,x,es,es',true, exitFreq + !freq)
| processSucc(b,x,(e as (F.EXIT{blknum=y,...},freq))::es,es',
exit, exitFreq) =
processSucc(b,x,es,es', true, exitFreq + !freq)
| processSucc(b,x,[],es',true, exitFreq) =
let val w = ref exitFreq
in exitPred := (b,w) :: !exitPred;
((exit,w)::es', true)
end
| processSucc(b,x,[],es', false, exitFreq) = (es', false)
| processSucc _ = error "processSucc"
fun processPred(b,x,(e as (F.BBLOCK{blknum=y, ...},freq))::es,
es', entry, entryFreq) =
if inSubgraph(y) then
processPred(b,x,es,e::es',entry,entryFreq)
else processPred(b,x,es,es',true,entryFreq + !freq)
| processPred(b,x,(e as (F.ENTRY{blknum=y,...},freq))::es,es',
entry, entryFreq) =
processPred(b,x,es,es',true, entryFreq + !freq)
| processPred(b,x,[], es', true, entryFreq) =
let val w = ref entryFreq
in entrySucc := (b,w) :: !entrySucc;
((entry,w)::es', true)
end
| processPred(b,x,[], es', false, entryFreq) = (es', false)
| processPred _ = error "processPred"
fun processNodes([], trail) = trail
| processNodes(
(b as F.BBLOCK{blknum=n,liveIn,liveOut,succ,pred,...})
::nodes, trail) =
let val (succ', exit) = processSucc(b,n,!succ,[],false,0)
val trail = if exit then (succ, !succ)::trail else trail
val (pred', entry) = processPred(b,n,!pred,[],false,0)
val trail = if entry then (pred, !pred)::trail else trail
in succ := succ';
pred := pred';
(* To save space, clear liveIn and
* liveOut information (if it is not an exit)
*)
liveIn := CellsBasis.CellSet.empty;
if exit then () else liveOut := CellsBasis.CellSet.empty;
processNodes(nodes, trail)
end
| processNodes _ = error "processNodes"
val _ = entrySucc := []
val _ = exitPred := []
val trail = processNodes(blocks, initTrail)
in trail
end
(* Make a subgraph with the appropriate edges *)
val trail = makeSubgraph(blocks)
val region =
F.CLUSTER{blkCounter = blkCounter,
blocks = blocks,
entry = entry,
exit = exit,
annotations = annotations
}
in (regionId, region, trail, blockCount)
end
(*
* Extract a new region to compile. Raises PQ.EmptyPriorityQueue if
* everything is finished.
*)
fun iterate() =
let val (id, region, trail, blockCount) = growRegion() (* get a region *)
in if debug then
print("[Region "^Int.toString id^" has "^Int.toString blockCount^
" blocks]\n")
else ();
processRegion region; (* allocate this region *)
resetTrail trail; (* reset the flowgraph *)
iterate() (* process next region *)
end
in (* Repeat until the entire flowgraph has been processed *)
iterate() handle PQ.EmptyPriorityQueue => ();
if debug then print "[Region based register allocation done]\n" else ()
end
end
|
