program-logic.tex 23 KB
Newer Older
Ralf Jung's avatar
Ralf Jung committed
1

2
\section{Program Logic}
3
\label{sec:program-logic}
4

5
This section describes how to build a program logic for an arbitrary language (\cf \Sref{sec:language}) on top of the logic described in \Sref{sec:dc-logic}.
Ralf Jung's avatar
Ralf Jung committed
6
So in the following, we assume that some language $\Lang$ was fixed.
Ralf Jung's avatar
Ralf Jung committed
7

Ralf Jung's avatar
Ralf Jung committed
8
\subsection{World Satisfaction, Invariants, Fancy Updates}
Ralf Jung's avatar
Ralf Jung committed
9
\label{sec:invariants}
Ralf Jung's avatar
Ralf Jung committed
10
11
12
13
14
15
16

To introduce invariants into our logic, we will define weakest precondition to explicitly thread through the proof that all the invariants are maintained throughout program execution.
However, in order to be able to access invariants, we will also have to provide a way to \emph{temporarily disable} (or ``open'') them.
To this end, we use tokens that manage which invariants are currently enabled.

We assume to have the following four CMRAs available:
\begin{align*}
Robbert Krebbers's avatar
Robbert Krebbers committed
17
  \textmon{State} \eqdef{}& \authm(\maybe{\exm(\State)}) \\
Robbert Krebbers's avatar
Robbert Krebbers committed
18
19
20
  \textmon{Inv} \eqdef{}& \authm(\nat \fpfn \agm(\latert \iPreProp)) \\
  \textmon{En} \eqdef{}& \pset{\nat} \\
  \textmon{Dis} \eqdef{}& \finpset{\nat}
Ralf Jung's avatar
Ralf Jung committed
21
22
23
24
25
26
27
\end{align*}
The last two are the tokens used for managing invariants, $\textmon{Inv}$ is the monoid used to manage the invariants themselves.
Finally, $\textmon{State}$ is used to provide the program with a view of the physical state of the machine.

Furthermore, we assume that instances named $\gname_{\textmon{State}}$, $\gname_{\textmon{Inv}}$, $\gname_{\textmon{En}}$ and $\gname_{\textmon{Dis}}$ of these CMRAs have been created.
(We will discuss later how this assumption is discharged.)

Ralf Jung's avatar
Ralf Jung committed
28
\paragraph{World Satisfaction.}
Ralf Jung's avatar
Ralf Jung committed
29
30
We can now define the assertion $W$ (\emph{world satisfaction}) which ensures that the enabled invariants are actually maintained:
\begin{align*}
Robbert Krebbers's avatar
Robbert Krebbers committed
31
  W \eqdef{}& \Exists I : \nat \fpfn \Prop.
32
33
34
35
36
37
38
  \begin{array}{@{} l}
    \ownGhost{\gname_{\textmon{Inv}}}{\authfull
      \mapsingletonComp {\iname}
        {\aginj(\latertinj(\wIso(I(\iname))))}
        {\iname \in \dom(I)}} * \\
    \Sep_{\iname \in \dom(I)} \left( \later I(\iname) * \ownGhost{\gname_{\textmon{Dis}}}{\set{\iname}} \lor \ownGhost{\gname_{\textmon{En}}}{\set{\iname}} \right)
  \end{array}
Ralf Jung's avatar
Ralf Jung committed
39
40
41
42
\end{align*}

\paragraph{Invariants.}
The following assertion states that an invariant with name $\iname$ exists and maintains assertion $\prop$:
43
44
\[ \knowInv\iname\prop \eqdef \ownGhost{\gname_{\textmon{Inv}}}
  {\authfrag \mapsingleton \iname {\aginj(\latertinj(\wIso(\prop)))}} \]
Ralf Jung's avatar
Ralf Jung committed
45

Ralf Jung's avatar
Ralf Jung committed
46
47
\paragraph{Fancy Updates and View Shifts.}
Next, we define \emph{fancy updates}, which are essentially the same as the basic updates of the base logic ($\Sref{sec:base-logic}$), except that they also have access to world satisfaction and can enable and disable invariants:
48
\[ \pvs[\mask_1][\mask_2] \prop \eqdef W * \ownGhost{\gname_{\textmon{En}}}{\mask_1} \wand \upd\diamond (W * \ownGhost{\gname_{\textmon{En}}}{\mask_2} * \prop) \]
Ralf Jung's avatar
Ralf Jung committed
49
Here, $\mask_1$ and $\mask_2$ are the \emph{masks} of the view update, defining which invariants have to be (at least!) available before and after the update.
Robbert Krebbers's avatar
Robbert Krebbers committed
50
We use $\top$ as symbol for the largest possible mask, $\nat$, and $\bot$ for the smallest possible mask $\emptyset$.
51
52
We will write $\pvs[\mask] \prop$ for $\pvs[\mask][\mask]\prop$.
%
Ralf Jung's avatar
Ralf Jung committed
53
Fancy updates satisfy the following basic proof rules:
54
\begin{mathpar}
Ralf Jung's avatar
Ralf Jung committed
55
\infer[fup-mono]
56
57
58
{\prop \proves \propB}
{\pvs[\mask_1][\mask_2] \prop \proves \pvs[\mask_1][\mask_2] \propB}

Ralf Jung's avatar
Ralf Jung committed
59
\infer[fup-intro-mask]
60
{\mask_2 \subseteq \mask_1}
61
{\prop \proves \pvs[\mask_1][\mask_2]\pvs[\mask_2][\mask_1] \prop}
62

Ralf Jung's avatar
Ralf Jung committed
63
\infer[fup-trans]
64
65
66
{}
{\pvs[\mask_1][\mask_2] \pvs[\mask_2][\mask_3] \prop \proves \pvs[\mask_1][\mask_3] \prop}

Ralf Jung's avatar
Ralf Jung committed
67
\infer[fup-upd]
68
69
{}{\upd\prop \proves \pvs[\mask] \prop}

Ralf Jung's avatar
Ralf Jung committed
70
\infer[fup-frame]
71
{}{\propB * \pvs[\mask_1][\mask_2]\prop \proves \pvs[\mask_1 \uplus \mask_\f][\mask_2 \uplus \mask_\f] \propB * \prop}
72

Ralf Jung's avatar
Ralf Jung committed
73
\inferH{fup-update}
74
75
76
{\melt \mupd \meltsB}
{\ownM\melt \proves \pvs[\mask] \Exists\meltB\in\meltsB. \ownM\meltB}

Ralf Jung's avatar
Ralf Jung committed
77
\infer[fup-timeless]
78
79
{\timeless\prop}
{\later\prop \proves \pvs[\mask] \prop}
80
%
Ralf Jung's avatar
Ralf Jung committed
81
% \inferH{fup-allocI}
82
83
84
% {\text{$\mask$ is infinite}}
% {\later\prop \proves \pvs[\mask] \Exists \iname \in \mask. \knowInv\iname\prop}
%gov
Ralf Jung's avatar
Ralf Jung committed
85
% \inferH{fup-openI}
86
87
% {}{\knowInv\iname\prop \proves \pvs[\set\iname][\emptyset] \later\prop}
%
Ralf Jung's avatar
Ralf Jung committed
88
% \inferH{fup-closeI}
89
% {}{\knowInv\iname\prop \land \later\prop \proves \pvs[\emptyset][\set\iname] \TRUE}
90
\end{mathpar}
91
(There are no rules related to invariants here. Those rules will be discussed later, in \Sref{sec:invariants}.)
Ralf Jung's avatar
Ralf Jung committed
92

Ralf Jung's avatar
Ralf Jung committed
93
We can further define the notions of \emph{view shifts} and \emph{linear view shifts}:
Ralf Jung's avatar
Ralf Jung committed
94
\begin{align*}
95
96
  \prop \vsW[\mask_1][\mask_2] \propB \eqdef{}& \prop \wand \pvs[\mask_1][\mask_2] \propB \\
  \prop \vs[\mask_1][\mask_2] \propB \eqdef{}& \always(\prop \wand \pvs[\mask_1][\mask_2] \propB)
Ralf Jung's avatar
Ralf Jung committed
97
\end{align*}
Ralf Jung's avatar
Ralf Jung committed
98
These two are useful when writing down specifications and for comparing with previous versions of Iris, but for reasoning, it is typically easier to just work directly with fancy updates.
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
Still, just to give an idea of what view shifts ``are'', here are some proof rules for them:
\begin{mathparpagebreakable}
\inferH{vs-update}
  {\melt \mupd \meltsB}
  {\ownGhost\gname{\melt} \vs \exists \meltB \in \meltsB.\; \ownGhost\gname{\meltB}}
\and
\inferH{vs-trans}
  {\prop \vs[\mask_1][\mask_2] \propB \and \propB \vs[\mask_2][\mask_3] \propC}
  {\prop \vs[\mask_1][\mask_3] \propC}
\and
\inferH{vs-imp}
  {\always{(\prop \Ra \propB)}}
  {\prop \vs[\emptyset] \propB}
\and
\inferH{vs-mask-frame}
  {\prop \vs[\mask_1][\mask_2] \propB}
  {\prop \vs[\mask_1 \uplus \mask'][\mask_2 \uplus \mask'] \propB}
\and
\inferH{vs-frame}
  {\prop \vs[\mask_1][\mask_2] \propB}
  {\prop * \propC \vs[\mask_1][\mask_2] \propB * \propC}
\and
\inferH{vs-timeless}
  {\timeless{\prop}}
  {\later \prop \vs \prop}

125
126
127
128
129
130
131
132
133
134
% \inferH{vs-allocI}
%   {\infinite(\mask)}
%   {\later{\prop} \vs[\mask] \exists \iname\in\mask.\; \knowInv{\iname}{\prop}}
% \and
% \axiomH{vs-openI}
%   {\knowInv{\iname}{\prop} \proves \TRUE \vs[\{ \iname \} ][\emptyset] \later \prop}
% \and
% \axiomH{vs-closeI}
%   {\knowInv{\iname}{\prop} \proves \later \prop \vs[\emptyset][\{ \iname \} ] \TRUE }
%
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
\inferHB{vs-disj}
  {\prop \vs[\mask_1][\mask_2] \propC \and \propB \vs[\mask_1][\mask_2] \propC}
  {\prop \lor \propB \vs[\mask_1][\mask_2] \propC}
\and
\inferHB{vs-exist}
  {\All \var. (\prop \vs[\mask_1][\mask_2] \propB)}
  {(\Exists \var. \prop) \vs[\mask_1][\mask_2] \propB}
\and
\inferHB{vs-always}
  {\always\propB \proves \prop \vs[\mask_1][\mask_2] \propC}
  {\prop \land \always{\propB} \vs[\mask_1][\mask_2] \propC}
 \and
\inferH{vs-false}
  {}
  {\FALSE \vs[\mask_1][\mask_2] \prop }
\end{mathparpagebreakable}

\subsection{Weakest Precondition}
Ralf Jung's avatar
Ralf Jung committed
153
154

Finally, we can define the core piece of the program logic, the assertion that reasons about program behavior: Weakest precondition, from which Hoare triples will be derived.
155
156

\paragraph{Defining weakest precondition.}
Ralf Jung's avatar
Ralf Jung committed
157
158
159
We assume that everything making up the definition of the language, \ie values, expressions, states, the conversion functions, reduction relation and all their properties, are suitably reflected into the logic (\ie they are part of the signature $\Sig$).

\begin{align*}
160
161
162
  \textdom{wp} \eqdef{}& \MU \textdom{wp}. \Lam \mask, \expr, \pred. \\
        & (\Exists\val. \toval(\expr) = \val \land \pvs[\mask] \prop) \lor {}\\
        & \Bigl(\toval(\expr) = \bot \land \All \state. \ownGhost{\gname_{\textmon{State}}}{\authfull \state} \vsW[\mask][\emptyset] {}\\
163
164
        &\qquad \red(\expr, \state) * \later\All \expr', \state', \vec\expr. (\expr, \state \step \expr', \state', \vec\expr) \vsW[\emptyset][\mask] {}\\
        &\qquad\qquad \ownGhost{\gname_{\textmon{State}}}{\authfull \state'} * \textdom{wp}(\mask, \expr', \pred) * \Sep_{\expr'' \in \vec\expr} \textdom{wp}(\top, \expr'', \Lam \any. \TRUE)\Bigr) \\
Ralf Jung's avatar
Ralf Jung committed
165
%  (* value case *)
166
  \wpre\expr[\mask]{\Ret\val. \prop} \eqdef{}& \textdom{wp}(\mask, \expr, \Lam\val.\prop)
Ralf Jung's avatar
Ralf Jung committed
167
\end{align*}
Ralf Jung's avatar
Ralf Jung committed
168
If we leave away the mask, we assume it to default to $\top$.
Ralf Jung's avatar
Ralf Jung committed
169

Ralf Jung's avatar
Ralf Jung committed
170
171
172
173
This ties the authoritative part of \textmon{State} to the actual physical state of the reduction witnessed by the weakest precondition.
The fragment will then be available to the user of the logic, as their way of talking about the physical state:
\[ \ownPhys\state \eqdef \ownGhost{\gname_{\textmon{State}}}{\authfrag \state} \]

174
175
\paragraph{Laws of weakest precondition.}
The following rules can all be derived inside the DC logic:
176
177
178
179
180
\begin{mathpar}
\infer[wp-value]
{}{\prop[\val/\var] \proves \wpre{\val}[\mask]{\Ret\var.\prop}}

\infer[wp-mono]
181
182
{\mask_1 \subseteq \mask_2 \and \vctx,\var:\textlog{val}\mid\prop \proves \propB}
{\vctx\mid\wpre\expr[\mask_1]{\Ret\var.\prop} \proves \wpre\expr[\mask_2]{\Ret\var.\propB}}
183

Ralf Jung's avatar
Ralf Jung committed
184
\infer[fup-wp]
185
186
{}{\pvs[\mask] \wpre\expr[\mask]{\Ret\var.\prop} \proves \wpre\expr[\mask]{\Ret\var.\prop}}

Ralf Jung's avatar
Ralf Jung committed
187
\infer[wp-fup]
188
189
190
{}{\wpre\expr[\mask]{\Ret\var.\pvs[\mask] \prop} \proves \wpre\expr[\mask]{\Ret\var.\prop}}

\infer[wp-atomic]
191
{\physatomic{\expr}}
192
193
194
195
196
197
198
199
{\pvs[\mask_1][\mask_2] \wpre\expr[\mask_2]{\Ret\var. \pvs[\mask_2][\mask_1]\prop}
 \proves \wpre\expr[\mask_1]{\Ret\var.\prop}}

\infer[wp-frame]
{}{\propB * \wpre\expr[\mask]{\Ret\var.\prop} \proves \wpre\expr[\mask]{\Ret\var.\propB*\prop}}

\infer[wp-frame-step]
{\toval(\expr) = \bot \and \mask_2 \subseteq \mask_1}
200
{\wpre\expr[\mask_2]{\Ret\var.\prop} * \pvs[\mask_1][\mask_2]\later\pvs[\mask_2][\mask_1]\propB \proves \wpre\expr[\mask_1]{\Ret\var.\propB*\prop}}
201
202
203
204
205
206

\infer[wp-bind]
{\text{$\lctx$ is a context}}
{\wpre\expr[\mask]{\Ret\var. \wpre{\lctx(\ofval(\var))}[\mask]{\Ret\varB.\prop}} \proves \wpre{\lctx(\expr)}[\mask]{\Ret\varB.\prop}}
\end{mathpar}

207
208
We will also want rules that connect weakest preconditions to the operational semantics of the language.
In order to cover the most general case, those rules end up being more complicated:
209
210
\begin{mathpar}
  \infer[wp-lift-step]
211
  {}
212
  { {\begin{inbox} % for some crazy reason, LaTeX is actually sensitive to the space between the "{ {" here and the "} }" below...
213
        ~~\pvs[\mask][\emptyset] \Exists \state_1. \red(\expr_1,\state_1) * \later\ownPhys{\state_1} * {}\\\qquad~~ \later\All \expr_2, \state_2, \vec\expr. \Bigl( (\expr_1, \state_1 \step \expr_2, \state_2, \vec\expr) * \ownPhys{\state_2} \Bigr) \wand \pvs[\emptyset][\mask] \Bigl(\wpre{\expr_2}[\mask]{\Ret\var.\prop} * \Sep_{\expr_\f \in \vec\expr} \wpre{\expr_\f}[\top]{\Ret\any.\TRUE}\Bigr)  {}\\\proves \wpre{\expr_1}[\mask]{\Ret\var.\prop}
214
215
216
      \end{inbox}} }
\\\\
  \infer[wp-lift-pure-step]
217
  {\All \state_1. \red(\expr_1, \state_1) \and
218
219
   \All \state_1, \expr_2, \state_2, \vec\expr. \expr_1,\state_1 \step \expr_2,\state_2,\vec\expr \Ra \state_1 = \state_2 }
  {\later\All \state, \expr_2, \vec\expr. (\expr_1,\state \step \expr_2, \state,\vec\expr)  \Ra \wpre{\expr_2}[\mask]{\Ret\var.\prop} * \Sep_{\expr_\f \in \vec\expr} \wpre{\expr_\f}[\top]{\Ret\any.\TRUE} \proves \wpre{\expr_1}[\mask]{\Ret\var.\prop}}
220
221
\end{mathpar}

Ralf Jung's avatar
Ralf Jung committed
222
223
224
225
226
227
We can further derive some slightly simpler rules for special cases:
We can derive some specialized forms of the lifting axioms for the operational semantics.
\begin{mathparpagebreakable}
  \infer[wp-lift-atomic-step]
  {\atomic(\expr_1) \and
   \red(\expr_1, \state_1)}
228
  { {\begin{inbox}~~\later\ownPhys{\state_1} * \later\All \val_2, \state_2, \vec\expr. (\expr_1,\state_1 \step \ofval(\val),\state_2,\vec\expr)  * \ownPhys{\state_2} \wand \prop[\val_2/\var] * \Sep_{\expr_\f \in \vec\expr} \wpre{\expr_\f}[\top]{\Ret\any.\TRUE} {}\\ \proves  \wpre{\expr_1}[\mask_1]{\Ret\var.\prop}
Ralf Jung's avatar
Ralf Jung committed
229
230
231
232
233
  \end{inbox}} }

  \infer[wp-lift-atomic-det-step]
  {\atomic(\expr_1) \and
   \red(\expr_1, \state_1) \and
234
235
   \All \expr'_2, \state'_2, \vec\expr'. \expr_1,\state_1 \step \expr'_2,\state'_2,\vec\expr' \Ra \state_2 = \state_2' \land \toval(\expr_2') = \val_2 \land \vec\expr = \vec\expr'}
  {\later\ownPhys{\state_1} * \later \Bigl(\ownPhys{\state_2} \wand \prop[\val_2/\var] * \Sep_{\expr_\f \in \vec\expr} \wpre{\expr_\f}[\top]{\Ret\any.\TRUE} \Bigr) \proves \wpre{\expr_1}[\mask_1]{\Ret\var.\prop}}
Ralf Jung's avatar
Ralf Jung committed
236
237

  \infer[wp-lift-pure-det-step]
238
  {\All \state_1. \red(\expr_1, \state_1) \\
239
240
   \All \state_1, \expr_2', \state'_2, \vec\expr'. \expr_1,\state_1 \step \expr'_2,\state'_2,\vec\expr' \Ra \state_1 = \state'_2 \land \expr_2 = \expr_2' \land \vec\expr = \vec\expr'}
  {\later \Bigl( \wpre{\expr_2}[\mask_1]{\Ret\var.\prop} * \Sep_{\expr_\f \in \vec\expr} \wpre{\expr_\f}[\top]{\Ret\any.\TRUE} \Bigr) \proves \wpre{\expr_1}[\mask_1]{\Ret\var.\prop}}
Ralf Jung's avatar
Ralf Jung committed
241
\end{mathparpagebreakable}
242
243


Ralf Jung's avatar
Ralf Jung committed
244
\paragraph{Adequacy of weakest precondition.}
245

246
247
248
249
The purpose of the adequacy statement is to show that our notion of weakest preconditions is \emph{realistic} in the sense that it actually has anything to do with the actual behavior of the program.
There are two properties we are looking for: First of all, the postcondition should reflect actual properties of the values the program can terminate with.
Second, a proof of a weakest precondition with any postcondition should imply that the program is \emph{safe}, \ie that it does not get stuck.

250
To express the adequacy statement for functional correctness, we assume we are given some set $V \subseteq \Val$ of legal return values.
251
252
Furthermore, we assume that the signature $\Sig$ adds a predicate $\pred$ to the logic which reflects $V$ into the logic:
\[\begin{array}{rMcMl}
253
  \Sem\pred &:& \Sem{\Val\,} \nfn \Sem\Prop \\
254
255
256
257
258
  \Sem\pred &\eqdef& \Lam \val. \Lam \any. \setComp{n}{v \in V}
\end{array}\]
The signature can of course state arbitrary additional properties of $\pred$, as long as they are proven sound.

The adequacy statement now reads as follows:
259
\begin{align*}
260
261
 &\All \mask, \expr, \val, \pred, \state, \state', \tpool'.
 \\&( \ownPhys\state \proves \wpre{\expr}[\mask]{x.\; \pred(x)}) \Ra
262
263
 \\&\cfg{\state}{[\expr]} \step^\ast
     \cfg{\state'}{[\val] \dplus \tpool'} \Ra
264
     \\&\val \in V
265
266
\end{align*}

267
The adequacy statement for safety says that our weakest preconditions imply that every expression in the thread pool either is a value, or can reduce further.
268
\begin{align*}
269
 &\All \mask, \expr, \state, \state', \tpool'.
270
 \\&(\All n. \melt \in \mval_n) \Ra
271
 \\&( \ownPhys\state \proves \wpre{\expr}[\mask]{x.\; \pred(x)}) \Ra
272
273
274
275
276
277
 \\&\cfg{\state}{[\expr]} \step^\ast
     \cfg{\state'}{\tpool'} \Ra
     \\&\All\expr'\in\tpool'. \toval(\expr') \neq \bot \lor \red(\expr', \state')
\end{align*}
Notice that this is stronger than saying that the thread pool can reduce; we actually assert that \emph{every} non-finished thread can take a step.

Ralf Jung's avatar
Ralf Jung committed
278
279
280
281
\paragraph{Hoare triples.}
It turns out that weakest precondition is actually quite convenient to work with, in particular when perfoming these proofs in Coq.
Still, for a more traditional presentation, we can easily derive the notion of a Hoare triple:
\[
282
\hoare{\prop}{\expr}{\Ret\val.\propB}[\mask] \eqdef \always{(\prop \wand \wpre{\expr}[\mask]{\Ret\val.\propB})}
Ralf Jung's avatar
Ralf Jung committed
283
\]
Ralf Jung's avatar
Ralf Jung committed
284

Ralf Jung's avatar
Ralf Jung committed
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
We only give some of the proof rules for Hoare triples here, since we usually do all our reasoning directly with weakest preconditions and use Hoare triples only to write specifications.
\begin{mathparpagebreakable}
\inferH{Ht-ret}
  {}
  {\hoare{\TRUE}{\valB}{\Ret\val. \val = \valB}[\mask]}
\and
\inferH{Ht-bind}
  {\text{$\lctx$ is a context} \and \hoare{\prop}{\expr}{\Ret\val. \propB}[\mask] \\
   \All \val. \hoare{\propB}{\lctx(\val)}{\Ret\valB.\propC}[\mask]}
  {\hoare{\prop}{\lctx(\expr)}{\Ret\valB.\propC}[\mask]}
\and
\inferH{Ht-csq}
  {\prop \vs \prop' \\
    \hoare{\prop'}{\expr}{\Ret\val.\propB'}[\mask] \\   
   \All \val. \propB' \vs \propB}
  {\hoare{\prop}{\expr}{\Ret\val.\propB}[\mask]}
\and
% \inferH{Ht-mask-weaken}
%   {\hoare{\prop}{\expr}{\Ret\val. \propB}[\mask]}
%   {\hoare{\prop}{\expr}{\Ret\val. \propB}[\mask \uplus \mask']}
% \\\\
\inferH{Ht-frame}
  {\hoare{\prop}{\expr}{\Ret\val. \propB}[\mask]}
  {\hoare{\prop * \propC}{\expr}{\Ret\val. \propB * \propC}[\mask]}
\and
% \inferH{Ht-frame-step}
%   {\hoare{\prop}{\expr}{\Ret\val. \propB}[\mask] \and \toval(\expr) = \bot \and \mask_2 \subseteq \mask_2 \\\\ \propC_1 \vs[\mask_1][\mask_2] \later\propC_2 \and \propC_2 \vs[\mask_2][\mask_1] \propC_3}
%   {\hoare{\prop * \propC_1}{\expr}{\Ret\val. \propB * \propC_3}[\mask \uplus \mask_1]}
% \and
\inferH{Ht-atomic}
  {\prop \vs[\mask \uplus \mask'][\mask] \prop' \\
    \hoare{\prop'}{\expr}{\Ret\val.\propB'}[\mask] \\   
   \All\val. \propB' \vs[\mask][\mask \uplus \mask'] \propB \\
   \physatomic{\expr}
  }
  {\hoare{\prop}{\expr}{\Ret\val.\propB}[\mask \uplus \mask']}
\and
\inferH{Ht-false}
  {}
  {\hoare{\FALSE}{\expr}{\Ret \val. \prop}[\mask]}
\and
\inferHB{Ht-disj}
  {\hoare{\prop}{\expr}{\Ret\val.\propC}[\mask] \and \hoare{\propB}{\expr}{\Ret\val.\propC}[\mask]}
  {\hoare{\prop \lor \propB}{\expr}{\Ret\val.\propC}[\mask]}
\and
\inferHB{Ht-exist}
  {\All \var. \hoare{\prop}{\expr}{\Ret\val.\propB}[\mask]}
  {\hoare{\Exists \var. \prop}{\expr}{\Ret\val.\propB}[\mask]}
\and
\inferHB{Ht-box}
  {\always\propB \proves \hoare{\prop}{\expr}{\Ret\val.\propC}[\mask]}
  {\hoare{\prop \land \always{\propB}}{\expr}{\Ret\val.\propC}[\mask]}
% \and
% \inferH{Ht-inv}
%   {\hoare{\later\propC*\prop}{\expr}{\Ret\val.\later\propC*\propB}[\mask] \and
%    \physatomic{\expr}
%   }
%   {\knowInv\iname\propC \proves \hoare{\prop}{\expr}{\Ret\val.\propB}[\mask \uplus \set\iname]}
% \and
% \inferH{Ht-inv-timeless}
%   {\hoare{\propC*\prop}{\expr}{\Ret\val.\propC*\propB}[\mask] \and
%    \physatomic{\expr} \and \timeless\propC
%   }
%   {\knowInv\iname\propC \proves \hoare{\prop}{\expr}{\Ret\val.\propB}[\mask \uplus \set\iname]}
\end{mathparpagebreakable}
Ralf Jung's avatar
Ralf Jung committed
350

Ralf Jung's avatar
Ralf Jung committed
351
352
353
354
355
356
357
358
359
360
\subsection{Invariant Namespaces}
\label{sec:namespaces}

In \Sref{sec:invariants}, we defined an assertion $\knowInv\iname\prop$ expressing knowledge (\ie the assertion is persistent) that $\prop$ is maintained as invariant with name $\iname$.
The concrete name $\iname$ is picked when the invariant is allocated, so it cannot possibly be statically known -- it will always be a variable that's threaded through everything.
However, we hardly care about the actual, concrete name.
All we need to know is that this name is \emph{different} from the names of other invariants that we want to open at the same time.
Keeping track of the $n^2$ mutual inequalities that arise with $n$ invariants quickly gets in the way of the actual proof.

To solve this issue, instead of remembering the exact name picked for an invariant, we will keep track of the \emph{namespace} the invariant was allocated in.
361
Namespaces are sets of invariants, following a tree-like structure:
Ralf Jung's avatar
Ralf Jung committed
362
363
364
365
366
367
368
369
Think of the name of an invariant as a sequence of identifiers, much like a fully qualified Java class name.
A \emph{namespace} $\namesp$ then is like a Java package: it is a sequence of identifiers that we think of as \emph{containing} all invariant names that begin with this sequence. For example, \texttt{org.mpi-sws.iris} is a namespace containing the invariant name \texttt{org.mpi-sws.iris.heap}.

The crux is that all namespaces contain infinitely many invariants, and hence we can \emph{freely pick} the namespace an invariant is allocated in -- no further, unpredictable choice has to be made.
Furthermore, we will often know that namespaces are \emph{disjoint} just by looking at them.
The namespaces $\namesp.\texttt{iris}$ and $\namesp.\texttt{gps}$ are disjoint no matter the choice of $\namesp$.
As a result, there is often no need to track disjointness of namespaces, we just have to pick the namespaces that we allocate our invariants in accordingly.

Robbert Krebbers's avatar
Robbert Krebbers committed
370
Formally speaking, let $\namesp \in \textlog{InvNamesp} \eqdef \List(\nat)$ be the type of \emph{invariant namespaces}.
Ralf Jung's avatar
Ralf Jung committed
371
372
373
374
We use the notation $\namesp.\iname$ for the namespace $[\iname] \dplus \namesp$.
(In other words, the list is ``backwards''. This is because cons-ing to the list, like the dot does above, is easier to deal with in Coq than appending at the end.)

The elements of a namespaces are \emph{structured invariant names} (think: Java fully qualified class name).
Robbert Krebbers's avatar
Robbert Krebbers committed
375
376
377
They, too, are lists of $\nat$, the same type as namespaces.
In order to connect this up to the definitions of \Sref{sec:invariants}, we need a way to map structued invariant names to $\nat$, the type of ``plain'' invariant names.
Any injective mapping $\textlog{namesp\_inj}$ will do; and such a mapping has to exist because $\List(\nat)$ is countable.
Ralf Jung's avatar
Ralf Jung committed
378
379
380
381
Whenever needed, we (usually implicitly) coerce $\namesp$ to its encoded suffix-closure, \ie to the set of encoded structured invariant names contained in the namespace: \[\namecl\namesp \eqdef \setComp{\iname}{\Exists \namesp'. \iname = \textlog{namesp\_inj}(\namesp' \dplus \namesp)}\]

We will overload the notation for invariant assertions for using namespaces instead of names:
\[ \knowInv\namesp\prop \eqdef \Exists \iname \in \namecl\namesp. \knowInv\iname{\prop} \]
Ralf Jung's avatar
Ralf Jung committed
382
We can now derive the following rules (this involves unfolding the definition of fancy updates):
Ralf Jung's avatar
Ralf Jung committed
383
384
\begin{mathpar}
  \axiomH{inv-persist}{\knowInv\namesp\prop \proves \always\knowInv\namesp\prop}
Ralf Jung's avatar
Ralf Jung committed
385

386
  \axiomH{inv-alloc}{\later\prop \proves \pvs[\emptyset] \knowInv\namesp\prop}
Ralf Jung's avatar
Ralf Jung committed
387

Ralf Jung's avatar
Ralf Jung committed
388
389
390
  \inferH{inv-open}
  {\namesp \subseteq \mask}
  {\knowInv\namesp\prop \vs[\mask][\mask\setminus\namesp] \later\prop * (\later\prop \vsW[\mask\setminus\namesp][\mask] \TRUE)}
Ralf Jung's avatar
Ralf Jung committed
391

Ralf Jung's avatar
Ralf Jung committed
392
393
394
395
  \inferH{inv-open-timeless}
  {\namesp \subseteq \mask \and \timeless\prop}
  {\knowInv\namesp\prop \vs[\mask][\mask\setminus\namesp] \prop * (\prop \vsW[\mask\setminus\namesp][\mask] \TRUE)}
\end{mathpar}
Ralf Jung's avatar
Ralf Jung committed
396

Ralf Jung's avatar
Ralf Jung committed
397
398
399
400
401
402
403
404
405
406
407
408
409
410
\subsection{Accessors}

The two rules \ruleref{inv-open} and \ruleref{inv-open-timeless} above may look a little surprising, in the sense that it is not clear on first sight how they would be applied.
The rules are the first \emph{accessors} that show up in this document.
Accessors are assertions of the form
\[ \prop \vs[\mask_1][\mask_2] \Exists\var. \propB * (\All\varB. \propB' \vsW[\mask_2][\mask_1] \propC) \]

One way to think about such assertions is as follows:
Given some accessor, if during our verification we have the assertion $\prop$ and the mask $\mask_1$ available, we can use the accessor to \emph{access} $\propB$ and obtain the witness $\var$.
We call this \emph{opening} the accessor, and it changes the mask to $\mask_2$.
Additionally, opening the accessor provides us with $\All\varB. \propB' \vsW[\mask_2][\mask_1] \propC$, a \emph{linear view shift} (\ie a view shift that can only be used once).
This linear view shift tells us that in order to \emph{close} the accessor again and go back to mask $\mask_1$, we have to pick some $\varB$ and establish the corresponding $\propB'$.
After closing, we will obtain $\propC$.

Ralf Jung's avatar
Ralf Jung committed
411
Using \ruleref{vs-trans} and \ruleref{Ht-atomic} (or the corresponding proof rules for fancy updates and weakest preconditions), we can show that it is possible to open an accessor around any view shift and any \emph{atomic} expression.
Ralf Jung's avatar
Ralf Jung committed
412
413
414
415
416
Furthermore, in the special case that $\mask_1 = \mask_2$, the accessor can be opened around \emph{any} expression.
For this reason, we also call such accessors \emph{non-atomic}.

The reasons accessors are useful is that they let us talk about ``opening X'' (\eg ``opening invariants'') without having to care what X is opened around.
Furthermore, as we construct more sophisticated and more interesting things that can be opened (\eg invariants that can be ``cancelled'', or STSs), accessors become a useful interface that allows us to mix and match different abstractions in arbitrary ways.
417
418
419
420
421

%%% Local Variables:
%%% mode: latex
%%% TeX-master: "iris"
%%% End: