Summary:
This diff changes the analysis exploration algorithm from considering
only sequential executions to considering executions of the
interleaving semantics. As part of this, symbolic states are changed
so that each thread has its own registers, while all memory is shared
between them.
Currently only a simple threads interface is supported: they can be
created with `thread_create(&thread_routine)`, they can exit by
returning from `thread_routine`, and they can be joined with
`thread_join`. Current simplifications include that newly created
threads are already runnable, thread routines accept no arguments and
return no result, and no failures are possible.
The concurrent exploration algorithm gives preference to executions
which have fewer context switches, thereby performing an incremental
form of context-bounded analysis.
A form of partial-order reduction is performed, where the symbolic
states are joined across (prefixes of) executions with the same number
of context switches which reach a point where the instruction pointers
and call stacks of all threads are the same. This has the effect of
"dagifying" the concurrent execution tree by merging points after
e.g. threads perform actions that commute with each other. This is
unlike traditional partial-order reduction in that it relies on the
symbolic join to combine the results of commuting operations in a way
that the following symbolic execution can take advantage of, rather
than performing some up-front analysis to identify commuting
operations and quotienting the space of executions. The current state
of the symbolic join and execution is significantly suboptimal in this
regard.
Differential Revision: D29441149
fbshipit-source-id: cf801a6b1
Summary:
The width of joins in the concurrency analysis is much wider, making
it worthwhile to use an n-ary version of join in order to avoid
repeated formula simplifiaction.
Differential Revision: D29441154
fbshipit-source-id: ae17de37b
Summary:
The only domain with a partial join is the lifting of a predicate
domain to a relation one, where the entry states are required to be
equal. This situation now indicates a programming error in the
analysis, rather than something that the domain should be responsible
for. Therefore this diff changes that check to an assertion and
simplifies the remaining join operations which are all total.
Reviewed By: jvillard
Differential Revision: D27828763
fbshipit-source-id: ec52ff741
Summary:
Currently all alarms are reported as "Invalid memory access", which is
not accurate for `abort` and hence assertion violations. This diff
adds an explicit type for alarms which distinguishes these two
cases. Further refinement is left for later.
Reviewed By: jvillard
Differential Revision: D27828754
fbshipit-source-id: 9c33f3c86
Summary:
This warning (68) triggers when a function argument pattern depends on
mutable state, which prevents the remaining arguments from being
uncurried, causing additional closure allocations.
Reviewed By: jvillard
Differential Revision: D27188311
fbshipit-source-id: a43354e15
Summary:
Revise the control-flow exploration scheduling algorithm to fix
several issues. The main difference is to change the priority queue to
keep the control edges on the frontier of exploration in sync with the
states that are waiting to be propagated. This fixes several sorts of
issue where the decision of which control and state joins to perform
was unexpected / wrong. Part of keeping the frontier edges and waiting
states in sync is that the waiting states are associated not only with
a destination block, but the stack of that block. This fixes several
issues.
Combined, these changes lead to the algorithm only attempting joins
for which the pointwise max join on depth maps is correct (with the
caveat of no mathematical proof yet).
Reviewed By: jvillard
Differential Revision: D25196733
fbshipit-source-id: db007fe1f
Summary:
Use the equality class information in the symbolic state to resolve
callees of indirect calls.
Reviewed By: jvillard
Differential Revision: D25146160
fbshipit-source-id: a1c39bbe1
Summary:
The callee function of a Call can often be resolved
statically. Currently this is resolution is only done dynamically
during symbolic execution by checking if the callee expression is a
function name and looking up the function in the program. This is
wasted and redundant work. Also, the static resolution code is
duplicated in all the domains.
This diff resolves this by resolving known callees statically at
translation time. This involves:
- add an ICall terminator that is the same as Call is currently
- change Call to use a func callee instead of Exp.t
- make callee field mutable since recursive calls can create cycles
- change the Llair.Term.call constructor to return a thunk to perform
the backpatching once the callee has been translated
- modify the Frontend
+ to determine whether to emit Call or ICall depending on whether
the callee in LLVM is already a Function
+ to record the LLVM function -- backpatch thunk pairs encountered during translation
+ record the mapping of LLVM to LLAIR functions during translation
+ to enumerate the calls to backpatch after all functions have been
translated, and find the LLAIR function corresponding to each LLVM
function and backpatch the call to use it as the callee
+ to handle direct calls to undefined functions, when backpatching
translate such function declarations into undefined functions
Reviewed By: jvillard
Differential Revision: D25146152
fbshipit-source-id: 47d2ca1ff
Summary:
Currently intrinsics are treated as functions, with Call instructions
to possibly-undefined functions with known names. This diff adds an
Intrinsic instruction form to express these more directly and:
- avoid the overhead of intrinsics needing to end blocks
- avoid the overhead of the function call machiery
- avoid the complexity of doing the string name lookup to find their
specs, repeatedly
This diff only adds the backend support, the added Intrinsic
instructions are not yet generated by the frontend.
Reviewed By: jvillard
Differential Revision: D25146155
fbshipit-source-id: f24024183
Summary:
Previously, when LLAIR was in SSA form, blocks took parameters just
like functions, and it was sometimes necessary to partially apply a
block to some of the parameters. For example, blocks to which function
calls return would need to accept the return value as an argument, and
sometimes immediately jump to another block passing the rest of the
arguments as well. These "trampoline" blocks were partial applications
of the eventual block to all but the final, return value,
argument.
This partial application mechanism meant that function parameters and
arguments were represented as a stack, with the first argument at the
bottom, that is, in reverse order.
Now that LLAIR is free of SSA, this confusion is no longer needed, and
this diff changes the representation of function formal parameters and
actual arguments to be in the natural order. This also brings Call
instructions in line with Intrinsic instructions, which will make
changing the handling of intrinsics from Calls to Intrinsic less
error-prone.
Reviewed By: jvillard
Differential Revision: D25146163
fbshipit-source-id: d3ed07a45
Summary:
Overwritten variables in move instructions are not impossible. Since
Domain_itv does not handle them, the check should be a `todo` rather
than an `assert false`.
Reviewed By: jvillard
Differential Revision: D25146168
fbshipit-source-id: 13d8587c7
Summary:
Localizing the entry of a procedure needs the globals (that the
procedure uses), but later creating a summary does not.
Reviewed By: jvillard
Differential Revision: D24886570
fbshipit-source-id: 8a7b18c58
Summary:
Distinguish expressions that name globals from registers. This leads
to clearer code, and globals are semantically distinct from general
registers. In particular, they are constant, so any machinery for
handling assignment does not need to consider them. This diff only
adds the distinction to LLAIR, it is not pushed through to FOL, which
will come later.
Reviewed By: jvillard
Differential Revision: D24846676
fbshipit-source-id: 3aca025bf
Summary:
Distinguish expressions that name functions from registers. This leads
to clearer code, and function names are semantically distinct from
general registers. In particular, they are constant, so any machinery
for handling assignment does not need to consider them. Unlike general
globals, they never have initializer expressions, and in particular
not recursive initializers. This diff only adds the distinction to
LLAIR, it is not pushed through to FOL, which will come later.
Reviewed By: jvillard
Differential Revision: D24846672
fbshipit-source-id: 2101f353f
Summary:
The support for recursive references to globals from within their
initializers is enough to handle all the cases of recursive structs
that have been encountered so far. Therefore this diff removes the
complication of recursive records entirely.
Reviewed By: jvillard
Differential Revision: D24772955
fbshipit-source-id: f59f06257
Summary:
Change the type of `fold` functions to enable them to compose
better. The guiding reasoning behind using types such as:
```
val fold : 'a t -> 's -> f:('a -> 's -> 's) -> 's
```
is:
1. The function argument should be labeled. This is so that it can be
reordered relative to the others, since it is often a multi-line
`fun` expression.
2. The function argument should come last. This enables its
arguments (which are often polymorphic) to benefit from type-based
disambiguation information determined by the types of the other
arguments at the call sites.
3. The function argument's type should produce an
accumulator-transformer when partially-applied. That is,
`f x : 's -> 's`. This composes well with other functions designed
to produce transformers/endofunctions when partially applied, and
in particular improves the common case of composing folds into
"state-passing style" code.
4. The fold function itself should produce an accumulator-transformer
when partially applied. So `'a t -> 's -> f:_ -> 's` rather than
`'s -> 'a t -> f:_ -> 's` or `'a t -> init:'s -> f:_ -> 's` etc.
Reviewed By: jvillard
Differential Revision: D24306063
fbshipit-source-id: 13bd8bbee
Summary:
The changes in set_intf.ml dictate the rest. The previous API
minimized changes when changing the backing implementation. But that
API is hostile toward composition, partial application, and
state-passing style code.
Reviewed By: jvillard
Differential Revision: D24306089
fbshipit-source-id: 00a09f486
Summary:
The treatment of comparison and exceptions in Core/Core_kernel/Base
makes them questionable as the default. This diff changes nonstdlib so
that Core is no longer opened in the global namespace, and makes a few
changes to handle the resulting minor API changes. This leads to a
lighter-touch nonstdlib, which makes a few definitions of its own, and
selects and extends modules from several libraries, including base,
core_kernel, containers, iter.
Reviewed By: jvillard
Differential Revision: D24306090
fbshipit-source-id: 42c91bd1b
Summary:
Code is expressed using Llair.Exp, which is potentially higher
fidelity than Term. So define the interval analysis directly in terms
of the Llair form.
Reviewed By: jvillard
Differential Revision: D24306085
fbshipit-source-id: f9d876eec
Summary:
Just to avoid confusion over directories named 'bin' being expected to
contain binaries, rather than the sources of executables.
Reviewed By: jvillard
Differential Revision: D23636372
fbshipit-source-id: 8593380e6
Josh Berdine