Summary:
The issue type `ZERO_EXECUTION_TIME` actually corresponds to bottom state but has been mistakenly used to mean
- unreachable nodes (program never reaching exit state)
- having zero cost (e.g. for allocations).
Note that, for execution costs, the latter doesn't make sense since we always incur a unit cost for the start node. Hence, a function with empty body will have unit cost. For allocations or IO however, we only incur costs for specific primitives, so a function with no allocations/IO could have a zero cost. However, there is no point reporting functions with zero cost as a specific issue type. Instead, what we want to track is the former, i.e. functions whose cost becomes 0 due to program never reaching exit state.
This diff aims to split these cases into two by only reporting on the latter and adds traces to bottom/unreachable cost by creating a special category in polynomials.
Next diff will rename `ZERO_XXX` to `XXX_UNREACHABLE_AT_EXIT`.
Reviewed By: skcho
Differential Revision: D20005774
fbshipit-source-id: 46b9abd5a
Summary:
> We don't report when the cost is Top as it corresponds to subsequent 'don't know's. Instead, we
> report Top cost only at the top level per function
The previous code just ignored top costed nodes, so it was able to report a non-top cost that was
from another node. For example,
```
void foo() {
linear-cost();
top-cost();
}
```
It reported inconsistent reports: `EXPENSIVE_EXECUTION_TIME` with a linear cost and
`INFINITE_EXECUTION_TIME` at the same time.
This diff fixes it not to report `EXPENSIVE_EXECUTION_TIME` when there is a node with the top cost.
Reviewed By: ezgicicek
Differential Revision: D20139408
fbshipit-source-id: 9fedd4aec
Summary:
In the previous report, it reported the first cost of node that exceeds a threshold. However, this
may hide a bigger cost of node that appears later. This diff changes this to report the biggest
cost of node among the costs exceeding the threshold.
Reviewed By: ezgicicek
Differential Revision: D20116162
fbshipit-source-id: 06199fb46
Summary: The semantics of the `values` function of Java enum class was missing, when it is called outside the class initializer. This diff gets the size of the enum elements from the summary of class initializer function, `<clinit>`.
Reviewed By: ezgicicek
Differential Revision: D20094880
fbshipit-source-id: 7362bba1c
Summary: We don't use allocation costs in prod at the moment. There is no plan to do so in the near future. Let's not report them anymore and also save some space in `costs-report.json`.
Reviewed By: skcho
Differential Revision: D19766828
fbshipit-source-id: 06dffa61d
Summary: This diff returns non-symbolic value (top) for unknown external function calls because the symbolic values sometimes make it hard to understand costs.
Reviewed By: ezgicicek
Differential Revision: D18685715
fbshipit-source-id: 1b39c718b
Summary:
`String` and `StringBuilder` both implement `CharSequence`. Let's generalize the model for `String` to `CharSequence` wherever possible and add missing models for
- `StringBuilder.append`
- `StringBuilder.toString`
Reviewed By: skcho
Differential Revision: D19558009
fbshipit-source-id: 0dfdb21af
Summary:
Java's String models were broken for
- initializing a String object with a locally defined constant string (which is an `Object*` in SIL).
- initializing a String object with a `char`/`byte` array
This diff fixes them and also adds models for `new String ()`.
Reviewed By: skcho
Differential Revision: D19662180
fbshipit-source-id: 23968d0aa
Summary: `String.split(regexp)` returns an array that is split by the given regexp. If the regexp doesn't match, the original string is returned. Hence, the resulting array's length must be in `[1, max(1, n_u -1)]` where`n_u` is the upper bound of the string's length.
Reviewed By: dulmarod
Differential Revision: D19578318
fbshipit-source-id: 675af7376
Summary: This diff use actuall call path in the cost results instead of `class name + method name`.
Reviewed By: ngorogiannis
Differential Revision: D19194969
fbshipit-source-id: b72018586
Summary:
This diff updates the relation between iterator (offset) and integer value not only at
assignments (`x += 1`), but also at function calls (`foo()`) that increase integer values by one in
their side effects.
Reviewed By: ezgicicek
Differential Revision: D19163214
fbshipit-source-id: 47e52f939
Summary: This diff extends the domain to express the relation between iterator's offset and integer value.
Reviewed By: ezgicicek
Differential Revision: D19143670
fbshipit-source-id: 6223bc934
Summary:
Old versions of sawja/javalib got the line numbers slightly wrong. The workaround was to do a regexp search in the source file for the right line.
My understanding is that this is no longer necessary. This diff removes it.
Reviewed By: jvillard
Differential Revision: D19033415
fbshipit-source-id: 2da19d66d
Summary: This diff extends the bound domain to express multiplication of bounds in some simple cases.
Reviewed By: ezgicicek
Differential Revision: D18745246
fbshipit-source-id: 4f2dcb42c
Summary:
This diff fixes the model of substring.
Problem: The cost model of the substring function was to return `size of string - start index` as a
cost. However, sometimes this was a negative number, because of state abstractions on paths, array
elements, call contexts, etc, which caused an exception inadvertently.
This diff changes the model to return just `size of string`, when it cannot say `size of string` is
bigger than `start index`.
Reviewed By: ezgicicek
Differential Revision: D18707954
fbshipit-source-id: 63f27e461
Summary:
This diff avoids unqualified variables by `ItvUpdatedBy` are qualified later. For example,
```
z = x & y;
z = z + 1;
```
While `z` should not be selected as a control variable, it wasn't, because it was qualified by the addition. This pattern introduces FPs in many cases.
Reviewed By: ngorogiannis
Differential Revision: D18505894
fbshipit-source-id: 13aec3008
Summary:
This diff excludes integer variables from control variables when their values are calculated by
binary operators.
Reviewed By: ezgicicek
Differential Revision: D18505826
fbshipit-source-id: 710533d4c
Summary:
There was a precision loss during the substitution of array block. For example:
Callee's abstract memory includes an array block as follows, where `a` is a parameter.
```
a.elements -> { a.elements[*] with a.elements.size }
```
Callers' abstract memory includes a pointer that may point to multiple array blocks.
```
c -> { x, y }
x.elements -> { x.elements[*] with x.elements.size }
y.elements -> { y.elements[*] with y.elements.size }
```
When the callee is called with the parameter `c`, the callees memory is substituted to:
```
x.elements -> { x.elements[*] with top , y.elements[*] with top }
y.elements -> { x.elements[*] with top , y.elements[*] with top }
```
because `a.elements[*]` was substituted to `{ x.elements[*] , y.elements[*] }`
and `a.elements.size` was substituted to `top ( = x.elements.size join y.elements.size )`.
This diff tries to keep the precision in the specific case, not to join the sizes of array blocks.
So now the same callee's abstract memory is substituted to:
```
x.elements -> { x.elements[*] with x.elements.size }
y.elements -> { y.elements[*] with y.elements.size }
```
Reviewed By: ngorogiannis
Differential Revision: D18480585
fbshipit-source-id: b70e63c22
Summary:
It returns non-top value when one of the parameters of band is positive, i.e., `x & 255` returns
`[0, 255]` instead of top.
Reviewed By: ezgicicek
Differential Revision: D18448614
fbshipit-source-id: aaa298a66
Summary:
Let's introduce a set of new cost analysis issue types that are raised when the function is statically determined to run on the UI thread. For this, we rely on the existing `runs_on_ui_thread` check that is developed for RacerD. We also update the cost summary and `jsonbug.cost_item` to include whether a method is on the ui thread so that we don't repeatedly compute this at diff time for complexity increase issues.
Note that `*_UI_THREAD` cost issues are assumed to be more strict than `*_COLD_START` reports at the moment. Next, we can also consider adding a new issue type that combines both such as `*_UI_THREAD_AND_COLD_START` (i.e. for methods that are both on cold start and run on ui thread).
Reviewed By: ngorogiannis
Differential Revision: D18428408
fbshipit-source-id: f18805716
Summary:
This diff tries more narrowing during analysis in order to get preciser results on nested loops.
In the widening phase, it does narrowing a loop right after its widening, for each loops. In general, this may make the widening phase non-terminating because it keeps the abstract state from monotonely increasing to the fixed point in a finite number of iterations. To avoid that situation, this diff applies the narrowing only when the first visit of the loop in the widening phase.
Reviewed By: ngorogiannis
Differential Revision: D18400631
fbshipit-source-id: cc76f7e85
Summary:
This diff adds semantics of Java function calls of enum `values` inside class initializers.
* Java class initializer function initializes a specific field `$VALUES`, which points to the list
of enum values.
* The `values` function of enum class returns the value of `$VALUES`.
The problem is when the `values` function is called inside the class initializer, for example:
```
enum Color {
RED,
GREEN,
BLUE;
static {
for (Color c : Color.values()) {}
}
}
```
This introduces a recursive dependency: the class initializer calls `Color.values` and the function
returns `Color.$VALUES` the value of which should be initialized in the class initializer.
To address the problem, this diff finds the value of `$VALUES` in its abstract memory when
`values` is called inside the class initializer.
Reviewed By: ezgicicek
Differential Revision: D18349281
fbshipit-source-id: 21766c20f
Summary:
This diff extends bound domain to express Min/Max of another bounds, so it can keep some more
precision in `Math.min/max`.
limitation: `MinMaxB`, the constructor of the bound, can contain only linear expressions or
previous min/max expressions.
Reviewed By: ezgicicek
Differential Revision: D18395365
fbshipit-source-id: fc90d27fd
Summary:
This diff extends the alias domain, so each variable can have multiple aliases.
It changed `KeyLhs` can be mapped to multiple alias targets in the `AliasMap` domain:
```
before : KeyLhs.t -> KeyRhs.t * AliasTarget.t
after : KeyLhs.t -> KeyRhs.t -> AliasTarget.t
```
Reviewed By: ezgicicek
Differential Revision: D18062178
fbshipit-source-id: b325a6055
Summary:
The zero cost of node does not make sense especially when the abstract memory is non-bottom. This
resulted in unreasonable zero cost results sometimes, e.g. when the checker could not find
appropriate control varaibles having interval values of iteration. This diff fixes this, so sets
the minimum basic cost as 1, if the abstract memory at the node is non-bottom.
Reviewed By: ezgicicek
Differential Revision: D18199291
fbshipit-source-id: b215d10e5
Summary:
This diff avoids making top values on unknown non-static function,
such as abstract function, calls. This is necessary because the
generated top values ruin the precision of the cost checker.
Reviewed By: ezgicicek
Differential Revision: D17418611
fbshipit-source-id: aeb759bdd
Summary:
This diff adopts an array length evaluation function that is conservative. It is useful when our
domain cannot express length result precisely.
For example, suppose there is an array pointer `arr_locs` that may point to two arrays `a` and `b`,
and their lengths are `a.length` and `b.length` (symbols), respectively. Using the usual
evaluation, our current domain cannot express `a.length join b.length` (join of two symbolic
values), so it returns top.
In this case, we can use the conservative function intead. It evaluates the length as `[0,
a.length.ub + b.length.ub]`, since we know every array length is positive. The result is not
precise, but better than top.
Reviewed By: ezgicicek
Differential Revision: D17908859
fbshipit-source-id: 7c0b1591b
Summary:
This diff introduces inequality for the iterator alias target, as we
did for the size target before.
Reviewed By: ezgicicek
Differential Revision: D17879208
fbshipit-source-id: cc2f6a723
Summary:
This diff revises the semantics of hasNext model to add the lengths of
arrays, rather than join them to top.
Reviewed By: ezgicicek
Differential Revision: D17882388
fbshipit-source-id: f5edaedb3
Summary:
This diff tries to narrowing the fixpoint of outermost loops, so that over-approximated widened values do not flow to the following code.
Problem: There are two phases for finding a fixpoint, widening and narrowing. First, it finds a fixpoint with widening, in function level. After that, it finds a fixpoint with narrowing. A problem is that sometimes an overly-approximated, imprecise, values by widening are flowed to the following loops. They are hard to narrow in the narrowing phase because there is a cycle preventing it.
To mitigate the problem, it tries to do narrowing, in loop level, right after it found a fixpoint of a loop. Thus, it narrows before the widened values are flowed to the following loops. In order to guarantee the termination of the analysis, this eager narrowing is applied only to the outermost loops.
Reviewed By: ezgicicek
Differential Revision: D17740265
fbshipit-source-id: e2d454036
Summary:
This diff generates a symbolic value when a function returns only
exceptions. Previously, the exception expression is evaluated to top,
thus it was propagated to other functions, which made those costs as
top. For preventing that situation, this diff changed:
* exception expressions are evaluated to bottom, and
* if callee's return value is bottom, it generates a symbolic value
for it.
Reviewed By: ezgicicek
Differential Revision: D17500386
fbshipit-source-id: 0fdcc710d
Summary: This diff introduces an inequality for the size alias targets, in order to get preciser array lengths after loops. The alias domain in inferbo was able to express strict equality between alias source and its targets, e.g. x=size(array). Now, for the size alias target, it can express less than or equal relations, e.g. x>=size(array).
Reviewed By: ezgicicek
Differential Revision: D17606222
fbshipit-source-id: 2557d3bd0
Summary:
This diff avoids giving the top value to unknown globals in Java,
because they harm precision of the cost checker. Instead, it doesn't
subst the global symbols at function calls.
Reviewed By: ezgicicek
Differential Revision: D17498714
fbshipit-source-id: d1215b3aa
Summary:
This diff adds an eval mode for the substitutions of the cost results, in order to avoid precision
loss by joining two symbols.
The usual join of two different symbolic values, `s1` and `s2`, becomes top due to the limitation of
our domain. On the other hand, in the new eval mode, it returns an upperbound `s1+s2`, because the
cost values only care about the upperbounds.
Reviewed By: ezgicicek
Differential Revision: D17573400
fbshipit-source-id: 2c84743d5
Summary:
In the cost checker, the range of selected control variables are used to estimate the number of loop iteration. However, sometimes the ranges of control variables are not related to how many times the loop iteration. This diff strengthens the condition for them as:
1. integers from `size` models
2. integers constructed from `+` or `-`
3. integers constructed from `*`
For the last one, the loop iteration is likely to be log scale of the range of the control variable:
```
while (i < c) {
i *= 2;
}
```
We will address this in the future.
Reviewed By: ezgicicek
Differential Revision: D17365796
fbshipit-source-id: c1e709ae8
Summary:
`ModeledRange` represents how many times the interval value can be updated by modeled functions. This
domain is to support the case where there are mismatches between value of a control variable and
actual number of loop iterations. For example,
```
while((c = file_channel.read(buf)) != -1) { ... }
```
the loop will iterates as the file size, but the control variable `c` does not have that value. In
these cases, it assigns a symbolic value of the file size to the modeled range of `c`, then which
is used when calculating the overall cost.
Reviewed By: jvillard
Differential Revision: D17476621
fbshipit-source-id: 9a81376e8
Summary:
This diff revises widening functions of bounds that have a linear form and a min/max form.
For example, for lower bounds,
* 3 ▽ (1+min(2, x)) = (1+min(2, x))
* 3+x ▽ (3+min(2, x)) = (3+min(2, x))
Reviewed By: jvillard
Differential Revision: D17420786
fbshipit-source-id: ff9eebed3
Summary:
This diff ignores character symbols in the cost results, in order to
avoid FPs from parser code.
Reviewed By: ezgicicek
Differential Revision: D17132053
fbshipit-source-id: d9cf8bd26
Summary:
Since it does not make sense to get ranges of non-integer values and
use them as approximate iteration numbers, this diff ignores control
values that only contain non-integer symbols.
Reviewed By: ezgicicek
Differential Revision: D17130967
fbshipit-source-id: f5ba58d52
Summary: This diff extends size alias domain for keeping one more alias of a Java temporary variable.
Reviewed By: ezgicicek
Differential Revision: D16984082
fbshipit-source-id: 244bbd0ee
Summary: This diff prunes array sizes in Java by adding the size alias on the `get_array_length` function calls.
Reviewed By: ezgicicek
Differential Revision: D16983501
fbshipit-source-id: a924af09d
Summary:
It revises Java's cast model to keep type in the location when it has a field.
The type information is useful especially when generating ondemand values of Collection elements.
Depends on D16807299
Reviewed By: ezgicicek
Differential Revision: D16807378
fbshipit-source-id: 636e54429
Summary: Ideally, we should be able to handle them like pruning if statements but for now, let's add the test.
Reviewed By: skcho
Differential Revision: D16938842
fbshipit-source-id: 04fae9559
Summary:
Since it is non-sense to get ranges of boolean values, this diff
ignores control values that only contain boolean symbols.
Depends on D16804802
Reviewed By: ezgicicek
Differential Revision: D16804808
fbshipit-source-id: ccb25db4d
Summary: Functions with empty body have unit cost, not zero. The unit cost comes from the start node.
Reviewed By: skcho
Differential Revision: D16855642
fbshipit-source-id: 6b5181faf
Summary:
Before this diff it returned `[0,size-1]`, but which became bottom
when size was given by 0. As a result, it made the both branches of
`if(iterator.hasNext())` unreachable. Similarly, if the size was 1,
it only visited the false branch of the if condition because the
condition value was `[0,0]` at that time.
This diff changes it to return `[0,size]`, so that
* the false branch is reachable when the size is 0
* the both branches are reachable when the size is 1
Reviewed By: ezgicicek
Differential Revision: D16803000
fbshipit-source-id: f8772be27
Summary: We want to keep big O notation as simple as possible in cost analysis reports (especially in diff time). Therefore, let's not show constants/min/max in big O notations even though the resulting asymptotic bound might be inaccurate. Developers can click on the trace and see the actual cost.
Reviewed By: skcho
Differential Revision: D16731351
fbshipit-source-id: 2e16f7eca
Summary: In order to test changes to bigO notation, let's record them in test results.
Reviewed By: skcho
Differential Revision: D16763972
fbshipit-source-id: c1376909b
Summary:
Correct the models of ArrayList initialization. Basically, there are two ways to initialize:
- by setting an initial capacity, which creates an empty list
- by passing another collection as an argument
Before, we had only modeled the second case which was resulting in wrong memory model for the first case. This diff fixes that.
Reviewed By: skcho
Differential Revision: D16711055
fbshipit-source-id: e82faf191
Summary: Models of Java's Collection mistakenly assumed that there was an argument for empty set whereas `Collections.emptySet()` doesn't have any actuals. This diff fixes that an also removes the type argument from the corresponding model definition.
Reviewed By: skcho
Differential Revision: D16582314
fbshipit-source-id: d4304dc60
Summary:
The `represents_multiple_values` flag was adopted to decide whether updating abstract value strongly or weakly. However, the flag was included in the `Val` domain, which is strange, because it is a property of abstract locations, rather than abstract values. This makes the behavior of memory update function depend on the abstract value to update, making its code complicated.
This diff detach the `represents_multiple_values` flag from the `Value` domain, thus the memory update does not depend on the abstract value. Since this is a refactoring, I believe the diff should not make many semantic changes.
Reviewed By: ezgicicek
Differential Revision: D16441734
fbshipit-source-id: 4c10779d7
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 diff adds a constant to the set of widening thresholds if the
constant is compared to an abstract value in condition expressions.
Each abstract value has its own set of thresholds.
Reviewed By: mbouaziz
Differential Revision: D14147150
fbshipit-source-id: ca0db34d4
Summary: Record where each symbol in a polynomial is coming from: either a loop, function call or a modeled call.
Reviewed By: mbouaziz
Differential Revision: D14047420
fbshipit-source-id: 56d0bd926
Summary: In SIL, Java's array member is a pointer to an array, while C++'s is the array itself. This diff differentiate them in evaluating abstract locations.
Reviewed By: ezgicicek, mbouaziz
Differential Revision: D14021451
fbshipit-source-id: 00f14fe3b
Summary:
- There is no need to use AI to compute a dot product: let's just fold over all nodes, but still do it in order (using the WTO) to report at the right place
- The previous version was computing a dot product on nodes for each node, which was quadratic, the new version is linear
- Report only once, the first time the threshold is reached (if in a loop, report at the loop head)
Reviewed By: ddino
Differential Revision: D14028171
fbshipit-source-id: b4a840c6e
Ezgi Çiçek
Sungkeun Cho
Nikos Gorogiannis
Josh Berdine
Dino Distefano
Mehdi Bouaziz