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
|
"""
This is optional support code for backends: it finds which cycles
in a graph are likely to correspond to source-level 'inner loops'.
"""
from rpython.tool.algo import graphlib
from rpython.flowspace.model import Variable
from rpython.translator.backendopt.ssa import DataFlowFamilyBuilder
class Loop:
def __init__(self, headblock, links):
self.headblock = headblock
self.links = links # list of Links making the cycle,
# starting from one of the exits of headblock
def find_inner_loops(graph, check_exitswitch_type=None):
"""Enumerate what look like the innermost loops of the graph.
Returns a list of non-overlapping Loop() instances.
"""
# Heuristic (thanks Stakkars for the idea):
# to find the "best" cycle to pick,
#
# * look for variables that don't change over the whole cycle
#
# * cycles with _more_ of them are _inside_ cycles with less of them,
# because any variable that doesn't change over an outer loop will
# not change over an inner loop either, and on the other hand the
# outer loop is likely to use and modify variables that remain
# constant over the inner loop.
#
# * if the numbers are the same, fall back to a more arbitrary
# measure: loops involving less blocks may be more important
# to optimize
#
# * in a cycle, which block is the head of the loop? Somewhat
# arbitrarily we pick the first Bool-switching block that has
# two exits. The "first" means the one closest to the
# startblock of the graph.
#
startdistance = {} # {block: distance-from-startblock}
pending = [graph.startblock]
edge_list = []
dist = 0
while pending:
newblocks = []
for block in pending:
if block not in startdistance:
startdistance[block] = dist
for link in block.exits:
newblocks.append(link.target)
edge = graphlib.Edge(block, link.target)
edge.link = link
edge_list.append(edge)
dist += 1
pending = newblocks
vertices = startdistance
edges = graphlib.make_edge_dict(edge_list)
cycles = graphlib.all_cycles(graph.startblock, vertices, edges)
loops = []
variable_families = None
for cycle in cycles:
# find the headblock
candidates = []
for i in range(len(cycle)):
block = cycle[i].source
v = block.exitswitch
if isinstance(v, Variable) and len(block.exits) == 2:
if getattr(v, 'concretetype', None) is check_exitswitch_type:
dist = startdistance[block]
candidates.append((dist, i))
if not candidates:
continue
_, i = min(candidates)
links = [edge.link for edge in cycle[i:] + cycle[:i]]
loop = Loop(cycle[i].source, links)
# count the variables that remain constant across the cycle,
# detected as having its SSA family present across all blocks.
if variable_families is None:
dffb = DataFlowFamilyBuilder(graph)
variable_families = dffb.get_variable_families()
num_loop_constants = 0
for v in loop.headblock.inputargs:
v = variable_families.find_rep(v)
for link in loop.links:
block1 = link.target
for v1 in block1.inputargs:
v1 = variable_families.find_rep(v1)
if v1 is v:
break # ok, found in this block
else:
break # not found in this block, fail
else:
# found in all blocks, this variable is a loop constant
num_loop_constants += 1
# smaller keys are "better"
key = (-num_loop_constants, # maximize num_loop_constants
len(cycle)) # minimize len(cycle)
loops.append((key, loop))
loops.sort()
# returns 'loops' without overlapping blocks
result = []
blocks_seen = {}
for key, loop in loops:
for link in loop.links:
if link.target in blocks_seen:
break # overlapping
else:
# non-overlapping
result.append(loop)
for link in loop.links:
blocks_seen[link.target] = True
return result
|
