Summary:
Adds `-mergefunc` and `-dce` passes to `Frontend.translate` to match
the `buck link` flow with `opt`
Reviewed By: ngorogiannis
Differential Revision: D15938641
fbshipit-source-id: 128cb89cd
Summary:
Using `Marshal.to_string` to create SQLite values used in comparisons is brittle as there is no guarantee that it will return the same value for structurally equal values.
When adding sharing, this will definitely break.
From the SQLite queries I found, only `SourceFile` and `Procname` are used in comparisons.
I haven't tested performance.
It shouldn't change anything for `SourceFile` as there is no possible sharing.
It shouldn't change much for `Procname` as they are pretty small anyway.
Reviewed By: ngorogiannis
Differential Revision: D15923122
fbshipit-source-id: ce4af1fe3
Summary:
The current handling of the formal return variable scope is not
correct. Since it is passed as an actual argument to the return
continuation, it is manipulated as if it was a local variable of the
caller. However, its scope is not ended with the caller's locals,
leading to clashes.
This diff reworks the passing of return values to avoid this problem,
mainly by introducing a notion of temporary variables during parameter
passing. This essentially has the effect of taking a function spec
{ P } f(x) { λv. Q }
and generating a "temporary" variable v, applying the post λv. Q to it
to obtain the pre-state for the call to the return continuation
k(v). Being a temporary variable just means that it goes out of scope
just after parameter passing. This amounts to a long-winded way of
applying the post-state to the formal parameter of the return
continuation without violating scopes or SSA.
This diff also separates the manipulation of the symbolic states as they
proceed from:
1. the pre-state before the return instruction;
2. the exit-state after the return instruction (including the binding
of the returned value to the return formal variable);
3. the post-state, where the locals are existentially quantified; and
4. the return-state, which is expressed in terms of actual args
instead of formal parameters.
Also in support of summarization, formal return and throw parameters
are no longer tracked on the analyzer's stack.
Note that these changes involve changing the locals of blocks and
functions to no longer include the formal parameters.
Reviewed By: kren1
Differential Revision: D15912148
fbshipit-source-id: e41dd6e42
Summary:
The solver couldn't deal with `∃ a,b . a = b` , so this diff adds
a special case to deal with it.
Reviewed By: ngorogiannis
Differential Revision: D15897953
fbshipit-source-id: d841d3557
Summary: Inject destructor calls to destroy a temporary when its lifetime ends.
Reviewed By: mbouaziz
Differential Revision: D15674209
fbshipit-source-id: 0f783a906
Summary:
Now that HIL doesn't help us anymore we need to reconstruct its mapping
"SIL logical var -> program access path". We already have everything we
need in pulse: it suffices to walk the current memory graph starting
from program variables until we find the value of the temporary we are
interested in.
This diff also builds some type machinery to make sure all accesses are
explained.
Reviewed By: mbouaziz
Differential Revision: D15824959
fbshipit-source-id: 722c81b39
Summary:
It turns out HIL gets in the way of a precise heap analysis. For
instance, instead of:
```
n$0 = *&x.f
_ = delete(&x)
*&y = n$0
```
HIL tries hard to forget about intermediate variables and shows instead
```
_ = delete(&x)
*&y = *&x.f
```
Oops, that's a use-after-delete, whereas the original code was safe.
While it's easy to write SIL programs that are completely unsound for
HIL, they are not generated very often from the frontends. In fact, the
problem became apparent only when making the clang frontend translate
C++ temporaries destructors, which produces the situation above
routinely.
This diff makes the minimal amount of change to make Pulse build and
produce equivalent results (minus HIL bugs) starting from SIL instead of
HIL. The reporting sucks for now because we need to translate SIL
temporaries back into program access paths. This is done in the next
diff.
Reviewed By: mbouaziz
Differential Revision: D15824961
fbshipit-source-id: 8e4e2a3ed
Summary:
Just moving code around.
This is needed later to make some types in `PulseTrace` depend on
a new that I'll have to define in `PulseDomain`.
Also, this gives better names all around I think
Reviewed By: mbouaziz
Differential Revision: D15881281
fbshipit-source-id: e86c1472e
Summary:
:
This patch adds several passes that reduce the amount of bitcode
making sledge's job easier, more info:
https://llvm.org/docs/Passes.html
`-mergefunc`
This pass merges functions that do the same thing, this can be because of
templating or casts (ie. same functionality but on 32bit and 64bit ints,
which is the same in machine code). More details at
http://llvm.org/docs/MergeFunctions.html
Note that this pass is currently not available through C/OCaml API.
`-constmerge`
This merges constants that have the same value, this is possible to do
when the constants are internalized.
`-argpromotion`
```
This pass promotes “by reference” arguments to be “by value” arguments.
In practice, this means looking for internal functions that have pointer
arguments. If it can prove, through the use of alias analysis, that an
argument is only loaded, then it can pass the value into the function
instead of the address of the value. This can cause recursive
simplification of code and lead to the elimination of allocas
(especially in C++ template code like the STL).
```
`-ipsccp`
```
Sparse conditional constant propagation and merging, which can be
summarized as:
Assumes values are constant unless proven otherwise
Assumes BasicBlocks are dead unless proven otherwise
Proves values to be constant, and replaces them with constants
Proves conditional branches to be unconditional
```
`-deadargelim`
Removes dead arguments of internal functions, good to run after other inter-procedural
passes. Seems to crash llvm if run directly after `ipsccp`.
Note that while this might look like doing full link-time optimisation,
we are actually picking relatively cheap optimisations that mostly look
at globals and walk their use chains. The main reason link-time
optimisations are expensive is due to inlining and then running the full
optimisation again from there.
Reviewed By: jberdine
Differential Revision: D15851408
fbshipit-source-id: be7191683
Summary:
Just moving code around.
This is needed later to make some types in `PulseInvalidation` depend on
a new type that I'll have to define in `PulseDomain`.
Reviewed By: mbouaziz
Differential Revision: D15824962
fbshipit-source-id: 86cba2bfb
Summary:
Make it possible to re-use the graph visitor to compute all sorts of
things with a flexible API where you can pass a function that folds over
all addresses reachable from certain stack variables (specified with a
filter) and gets passed the access path that leads to each address.
This is used in later commits.
Reviewed By: mbouaziz
Differential Revision: D15824960
fbshipit-source-id: c424a71cb
Summary: Preanalysis is performed at the frontend now. Hence, we don't need to repeatedly check/set when/if it is performed.
Reviewed By: mbouaziz
Differential Revision: D15863175
fbshipit-source-id: f9c6b7ae1
Summary:
One "interesting" feature of the approach of merging the captured targets in Java, is that we union their type environments, as opposed to store partial tenvs together with each source file, which is the case for Clang.
This means
- the final global type environment is potentially huge because it contains all the types in all targets.
- all analysis workers start by loading that tenv in memory, meaning we consume `|size of tenv| x #cpus` memory, which can tip the balance towards OOMs
This diff attempts to economise on global tenv size. This is done by increasing sharing which is then preserved by marshalling. It's done in a brute force way, with hashtables for each struct component, and is not fully effective due to the recursion amongst types and types names, as well types appearing inside other constructs such as procnames.
This is done when calling `Tenv.store` so that
- the computation can be parallelised somewhat (capture is parallel, merging is not)
- buck caching will benefit from smaller tenvs.
This saves about 24% of total memory devoted to the type environment.
Reviewed By: mbouaziz
Differential Revision: D15840054
fbshipit-source-id: 6f03be1a4
Summary:
- Add allocation costs to `costs-report.json` and enable diffing over allocation costs.
- Also, let's be more consistent and modular in naming our cost issues.
- introduce a generic issue type `X_TIME_COMPLEXITY_INCREASE` where `X` can be one of the cost kinds. If the function is on the cold start, issue can have the `COLD_START` suffix. Similarly for infinite/zero/expensive calls.
- Change `PERFORMANCE_VARIATION` -> `EXECUTION_TIME_COMPLEXITY_INCREASE`
- Add new issue type for `ALLOCATION_COMPLEXITY_INCREASE_COLD_START` which will be enabled by default
- Refactor cost issues to be more modular and succinct. This also makes addition of a new cost kind very easy by adding the kind into the `enabled_cost_kinds` list in `CostKind.ml`
Reviewed By: mbouaziz
Differential Revision: D15822681
fbshipit-source-id: cf89ece59
Summary:
This one isn't caught because we don't destruct temporaries that are
bound to a const reference. According to the C++ standard these should
get destroyed when the const reference gets destroyed but instead we
just don't destroy them for now.
Reviewed By: mbouaziz
Differential Revision: D15760209
fbshipit-source-id: 32c935ec0
Summary:
In a next diff temporaries will get destructed at the end of their
lifetimes and that naive model would be causing false positives.
The flipside is that we lose all reports on closures for now, will need
to model them separately later.
Reviewed By: mbouaziz
Differential Revision: D15695943
fbshipit-source-id: c2c482c02
Summary:
Needed for next diff: we'll need to do 2 passes on the AST to collect
the temporaries to destroy at the end of an `ExprWithCleanups`, but the
SIL names of these temporaries are generated freshly on the fly so they
would get different names if we do it naively.
This adds a hashmap to the translation context so the temporary
corresponding to a given `MaterializeTemporyExpr` is only generated once
and then reused.
Reviewed By: mbouaziz
Differential Revision: D15674212
fbshipit-source-id: 0e16062d9
Summary:
This started as an attempt to understand how to modify the frontend to
inject destructors for C++ temporaries (see next diffs).
This diff rewrites the existing logic for computing the list of
variables that should be destroyed at the end of each statement, either
because it's the end of their syntactic scope or because control flow
branches outside of their syntactic scope.
The frontend translates a function from the last instructions to the
first, but scope computation needs to be done in the other direction, so
it's done in a separate pass *before* the main translation happens. That
first pass creates a map from statements in the AST to the list of
variables that should be destroyed at the end of these statements. This
is still the case now.
Before, that map would be computed in a bit of a weird way: scopes are
naturally a stack but instead of that the structure maintained was a
flat list + a counter to know where the current scope ended in that
list.
In this diff, redo the computation maintaining a stack of scopes
instead, which is a bit cleaner. Also treat more instructions as
introducing a new scope, eg if, for, ...
Reviewed By: mbouaziz
Differential Revision: D15674208
fbshipit-source-id: c92429e82
Summary:
Somewhat trivial: add a string to "Destruction" nodes to indicate why
they were created. Rename the main `instruction_aux` function into
`instruction_translate` (see next diff for why).
Reviewed By: mbouaziz
Differential Revision: D15674211
fbshipit-source-id: 8a7eda72c
Summary:
I rewrote the test so it doesn't need any C++ headers so that:
- it's easier to see what's going on
- it's easier to debug: the whole AST is now somewhat readable vs before
the headers made it impossibly long
Reviewed By: ezgicicek
Differential Revision: D15674213
fbshipit-source-id: d98941983
Summary:
This diff introduces a `-lib-fuzz` flag to `buck link`, which links in a
simple main that calls the LLVMFuzzerTestOneInput function, which is the
entry point of libFuzzer fuzzer.
Reviewed By: jberdine, jvillard
Differential Revision: D15821512
fbshipit-source-id: cff731ed3
Summary: This allows to match `foo<int_&>` and many other horrible names.
Reviewed By: mbouaziz
Differential Revision: D15825403
fbshipit-source-id: c892033aa
Summary:
I realized that there was a discrepancy in the # of instructions between whether we run a single analysis or multiple analyses at the same time. It turns out that in biabduction, bufferoverrun and other HIL analyses we did Preanalysis step (which adds scope instructions and invokes liveness etc.) but not in others. This discrepancy results in inconsistent analysis results (e.g. in the new inefficient-keyset-iterator) that rely on instructions. We should be consistent. Hence, we now invoke Preanalysis in the frontend and remove all other uses in the rest of the checkers.
Consequently, I had to update the inefficient-keyset-checker to take the CFG resulting from Preanalysis with extra scoping instructions.
Reviewed By: mbouaziz, ngorogiannis, jvillard
Differential Revision: D15803492
fbshipit-source-id: 4e21eb610
Summary:
Previous change to allow bitcasts in call instructions was too strict
and did not allow for indirect calls.
Reviewed By: jberdine
Differential Revision: D15803262
fbshipit-source-id: 40d828b59
Summary:
Currently printing symbolic heaps is unreadable, because there are too
many quantified variables, that are mostly just equal to other
variables.
This diff tries to replace all variables in an equivalence class with a
single variable and remove the unneccesary variables.
It also introduces two modes for printing state domains:
`-t +State_domain.pp_full` prints the state domain as is
`-t +State_domain.pp` uses the simplification before printing.
Reviewed By: jberdine
Differential Revision: D15738748
fbshipit-source-id: 7c85b580e
Summary:
Print pre- and post- conditions (aka, summaries) when analyzer hits a
function return
- plumbing the precondition through the analyzer
so that it is available when return is hit
Reviewed By: jberdine
Differential Revision: D15713725
fbshipit-source-id: b10b6206f
Summary:
This diff adds a `__llair_alloc` intrinsic which is modeled
as a non-failing malloc. Using it instead of `malloc` increases
the readbility of symbolic heaps, because it removes all the cases
where malloc failed.
Note that `assert(malloc())` does not have the desired effect.
Reviewed By: ngorogiannis
Differential Revision: D15778817
fbshipit-source-id: d02784077
Summary:
Some call instructions in LLVM bitcast the function,
for example
`%call = call i32 (i64, ...) bitcast (i32 (...)* @__llair_alloc to i32 (i64, ...)*)(i64 %conv)`
This would cause sledge to crash in LLVM when build with assertions.
Reviewed By: jberdine
Differential Revision: D15779003
fbshipit-source-id: c273f92db
Summary:
This is a simple checker that identifies inefficient uses of `keySet` iterator where (not only the key but also) the value is accessed via `get(key)`. It is more efficient to use `entrySet` iterator which already returns both key-value pairs. This optimization would get rid of many extra lookups which can be expensive.
We simply traverse the CFG starting from the loop head upwards and pick up the map that is iterated over. Then, we check in the loop nodes if there is a call to `get(...)` over this map. If, so we report.
Reviewed By: ngorogiannis
Differential Revision: D15737779
fbshipit-source-id: 702465b4e
Summary:
This diff adds a formal parameter to each non-void-returning function
to name the return value, and similarly a formal parameter for the
thrown exception value. These are interpreted as call-by-reference
parameters, so that they can be constrained in formulas to e.g. be
equal to the return value, and are still in scope when the function
returns, and so can be passed to the return block. Prior to
summarizing functions, this means that these formals need to be
tracked on the analyzer's control stack.
This will be needed to express function specs/summaries in terms of
formals, and fixes a bug where in some cases return values were not
tracked correctly.
Reviewed By: kren1
Differential Revision: D15738026
fbshipit-source-id: fff2d107c
Summary:
Previously the locals of a function were computed after backpatching
the blocks in its cfg. This resulted in loss of sharing, and incorrect
locals if queried through the parent of a block.
Reviewed By: kren1
Differential Revision: D15738027
fbshipit-source-id: d7e70530a
Summary:
Disable exceptional control flow
- treat throw as unreachable
- confidence in the correctness of the frontend's treatment of
exception handling is very low, and making summaries that are
expressive enough to talk about exceptions is a complication
that isn't needed for the first iteration
To facilitate, start on a struct that holds all the CL options.
Reviewed By: jberdine, jvillard
Differential Revision: D15713601
fbshipit-source-id: ee92dfbd8
Summary:
Move genrule capture integration logic from shell to OCaml.
Also, stop relying on side-effects of buck compilation for constructing the infer-deps.txt file used for merging. Now this is obtained by passing `--show-output` to buck, which spits out the `buck-out` output paths to the targets we asked to build.
Reviewed By: ezgicicek
Differential Revision: D15715608
fbshipit-source-id: 8fa896ba6
Summary:
The synthetic methods from `topl.Property` are now nonempty: they
simulate a nondeterministic automaton.
Reviewed By: jvillard
Differential Revision: D15668471
fbshipit-source-id: 050408283
Summary:
Instrument SIL according to TOPL properties. Roughly, the
instrumentation is a set of calls into procedures that simulate a
nondeterministic automaton. For now, those procedures are NOP dummies.
Reviewed By: jvillard
Differential Revision: D15063942
fbshipit-source-id: d22c2f6fa
Summary:
This diff adapts the test scripts to the new sledge CLI, and reworks
them to enable checking changes with respect to a baseline. In
particular, now
```
make -C test
```
has exit code 0 if the current test results match the expected ones,
and otherwise prints the diff. Also,
```
make -C test promote
```
promotes the current test results to the new baseline.
Reviewed By: kren1
Differential Revision: D15706573
fbshipit-source-id: 0cbf3231e
Timotej Kapus
Mehdi Bouaziz
Josh Berdine
Ezgi Çiçek
Jules Villard
Nikos Gorogiannis
Dino Distefano
Radu Grigore