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
|
(** Linearization of the control-flow graph:
translation from LTL to Linear *)
Require Import Coqlib.
Require Import Maps.
Require Import Sets.
Require Import AST.
Require Import Values.
Require Import Globalenvs.
Require Import Op.
Require Import Locations.
Require Import LTL.
Require Import Linear.
Require Import Kildall.
Require Import Lattice.
(** To translate from LTL to Linear, we must layout the basic blocks
of the LTL control-flow graph in some linear order, and insert
explicit branches and conditional branches to make sure that
each basic block jumps to its successors as prescribed by the
LTL control-flow graph. However, branches are not necessary
if the fall-through behaviour of Linear instructions already
implements the desired flow of control. For instance,
consider the two LTL basic blocks
<<
L1: Bop op args res (Bgoto L2)
L2: ...
>>
If the blocks [L1] and [L2] are laid out consecutively in the Linear
code, we can generate the following Linear code:
<<
L1: Lop op args res
L2: ...
>>
However, if this is not possible, an explicit [Lgoto] is needed:
<<
L1: Lop op args res
Lgoto L2
...
L2: ...
>>
The main challenge in code linearization is therefore to pick a
``good'' order for the basic blocks that exploits well the
fall-through behavior. Many clever trace picking heuristics
have been developed for this purpose.
In this file, we present linearization in a way that clearly
separates the heuristic part (choosing an order for the basic blocks)
from the actual code transformation parts. We proceed in three passes:
- Choosing an order for the basic blocks. This returns an enumeration
of CFG nodes stating that the basic blocks must be laid out in
the order shown in the list.
- Generate naive Linear code where each basic block branches explicitly
to its successors, even if one of these successors is the next basic
block.
- Simplify the naive Linear code, removing unconditional branches to
a label that immediately follows:
<<
... ...
Igoto L1; becomes L1: ...
L1: ...
>>
The beauty of this approach is that correct code is generated
under surprisingly weak hypotheses on the enumeration of
CFG nodes: it suffices that every reachable basic block occurs
exactly once in the enumeration. While the enumeration heuristic we use
is quite trivial, it is easy to implement more sophisticated
trace picking heuristics: as long as they satisfy the property above,
we do not need to re-do the proof of semantic preservation.
The enumeration could even be performed by external, untrusted
Caml code, then verified (for the property above) by certified Coq code.
*)
(** * Determination of the order of basic blocks *)
(** We first compute a mapping from CFG nodes to booleans,
indicating whether a CFG basic block is reachable or not.
This computation is a trivial forward dataflow analysis
where the transfer function is the identity: the successors
of a reachable block are reachable, by the very
definition of reachability. *)
Module DS := Dataflow_Solver(LBoolean).
Definition reachable_aux (f: LTL.function) : option (PMap.t bool) :=
DS.fixpoint
(successors f)
(Psucc f.(fn_entrypoint))
(fun pc r => r)
((f.(fn_entrypoint), true) :: nil).
Definition reachable (f: LTL.function) : PMap.t bool :=
match reachable_aux f with
| None => PMap.init true
| Some rs => rs
end.
(** We then enumerate the nodes of reachable basic blocks. The heuristic
we take is trivial: nodes are enumerated in decreasing numerical
order. Remember that CFG nodes are positive integers, and that
the [RTLgen] pass allocated fresh nodes 1, 2, 3, ..., but generated
instructions in roughly reverse control-flow order: often,
a basic block at label [n] has [n-1] as its only successor.
Therefore, by enumerating reachable nodes in decreasing order,
we recover a reasonable layout of basic blocks that globally respects
the structure of the original Cminor code. *)
Definition enumerate (f: LTL.function) : list node :=
let reach := reachable f in
positive_rec (list node) nil
(fun pc nodes => if reach!!pc then pc :: nodes else nodes)
(Psucc f.(fn_entrypoint)).
(** * Translation from LTL to Linear *)
(** We now flatten the structure of the CFG graph, laying out
basic blocks consecutively in the order computed by [enumerate],
and inserting a branch at the end of every basic block.
For blocks ending in a conditional branch [Bcond cond args s1 s2],
we have two possible translations:
<<
Lcond cond args s1; or Lcond (not cond) args s2;
Lgoto s2 Lgoto s1
>>
We favour the first translation if [s2] is the label of the
next instruction, and the second if [s1] is the label of the
next instruction, thus exhibiting more opportunities for
fall-through optimization later. *)
Fixpoint starts_with (lbl: label) (k: code) {struct k} : bool :=
match k with
| Llabel lbl' :: k' => if peq lbl lbl' then true else starts_with lbl k'
| _ => false
end.
Fixpoint linearize_block (b: block) (k: code) {struct b} : code :=
match b with
| Bgetstack s r b =>
Lgetstack s r :: linearize_block b k
| Bsetstack r s b =>
Lsetstack r s :: linearize_block b k
| Bop op args res b =>
Lop op args res :: linearize_block b k
| Bload chunk addr args dst b =>
Lload chunk addr args dst :: linearize_block b k
| Bstore chunk addr args src b =>
Lstore chunk addr args src :: linearize_block b k
| Bcall sig ros b =>
Lcall sig ros :: linearize_block b k
| Balloc b =>
Lalloc :: linearize_block b k
| Bgoto s =>
Lgoto s :: k
| Bcond cond args s1 s2 =>
if starts_with s1 k then
Lcond (negate_condition cond) args s2 :: Lgoto s1 :: k
else
Lcond cond args s1 :: Lgoto s2 :: k
| Breturn =>
Lreturn :: k
end.
(* Linearize a function body according to an enumeration of its
nodes. *)
Fixpoint linearize_body (f: LTL.function) (enum: list node)
{struct enum} : code :=
match enum with
| nil => nil
| pc :: rem =>
match f.(LTL.fn_code)!pc with
| None => linearize_body f rem
| Some b => Llabel pc :: linearize_block b (linearize_body f rem)
end
end.
Definition linearize_function (f: LTL.function) : Linear.function :=
mkfunction
(LTL.fn_sig f)
(LTL.fn_stacksize f)
(linearize_body f (enumerate f)).
(** * Cleanup of the linearized code *)
(** We now eliminate [Lgoto] instructions that branch to an
immediately following label: these are redundant, as the fall-through
behaviour obtained by removing the [Lgoto] instruction is
semantically equivalent. *)
Fixpoint cleanup_code (c: code) {struct c} : code :=
match c with
| nil => nil
| Lgoto lbl :: c' =>
if starts_with lbl c'
then cleanup_code c'
else Lgoto lbl :: cleanup_code c'
| i :: c' =>
i :: cleanup_code c'
end.
Definition cleanup_function (f: Linear.function) : Linear.function :=
mkfunction
(fn_sig f)
(fn_stacksize f)
(cleanup_code f.(fn_code)).
(** * Entry points for code linearization *)
Definition transf_function (f: LTL.function) : Linear.function :=
cleanup_function (linearize_function f).
Definition transf_fundef (f: LTL.fundef) : Linear.fundef :=
AST.transf_fundef transf_function f.
Definition transf_program (p: LTL.program) : Linear.program :=
transform_program transf_fundef p.
|