You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
47 lines
3.1 KiB
47 lines
3.1 KiB
5 years ago
|
This is an experimental inter-procedural analysis that detects pure (side-effect free) functions. For each function, purity analysis keeps track of not only the purity of the function but also some additional information such as whether the function modifies a global variable or which of the parameters are modified. It models functions with no summary/model as modifying the global state (hence impure).
|
||
|
|
|
||
|
|
If the function is pure (i.e. doesn't modify any global state or its parameters and doesn't call any unknown functions), then it reports an [`PURE_FUNCTION`](/docs/next/all-issue-types#pure_function) issue.
|
||
|
|
|
||
|
|
|
||
|
|
## Weaknesses
|
||
|
|
|
||
|
|
There are two issues with the existing purity analysis:
|
||
|
|
- In order to detect which parameters are modified, we need an alias analysis which is difficult to get right.
|
||
|
|
- Just keeping track of modified arguments doesn't suffice.
|
||
|
|
|
||
|
|
Too see the issue with the first point, consider the following simple program:
|
||
|
|
|
||
|
|
```java
|
||
|
|
void foo(Foo a){
|
||
|
|
Foo b = a;
|
||
|
|
b.x = 10;
|
||
|
|
}
|
||
|
|
```
|
||
|
|
|
||
|
|
in order to determine that `foo` is impure, we need to know that the write to `b`'s field is actually changing the function parameter `a`, i.e. we need to check if `b` is aliasing `a`. This is known as alias analysis and is hard to get right in a scalable manner. When this analysis was being developed, Infer didn't have a unified alias analysis and using biabduction seemed like a too daunting task at the time. Hence, we relied on [InferBo](/docs/next/checker-bufferoverrun)'s aliasing mechanism which was easy to invoke and integrate with. However, InferBo's aliasing analysis is far from perfect and causes issues for purity.
|
||
|
|
To see the issue with the second point, consider the following program:
|
||
|
|
|
||
|
|
```java
|
||
|
|
boolean contains(Integer i, ArrayList<Integer> list){
|
||
|
|
Iterator<Integer> listIterator = list.iterator();
|
||
|
|
while(listIterator.hasNext()) {
|
||
|
|
Integer el = listIterator.next();
|
||
|
|
if (i.equals(el)){
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
```
|
||
|
|
|
||
|
|
The existing purity analysis concludes that the above function `contains` is impure because it calls an impure function `next()` which modifies the iterator (hence it thinks it also modifies the `list`). However, notice that `contains` doesn't have an observable side-effect: `list.iterator()` returns a new object, `hasNext()` and `equals()` are pure, and `next()` only modifies the fields of the fresh object `listIterator`. Therefore, `contains` should be considered as pure.
|
||
|
|
|
||
|
|
|
||
|
|
To alleviate this problem, we have developed an [Impurity](/docs/next/checker-impurity) analysis which uses [pulse](/docs/next/checker-pulse) which can successfully analyze this program as pure \o/
|
||
|
|
|
||
|
|
|
||
|
|
The analysis is used by:
|
||
|
|
|
||
|
|
- [Loop-hoisting](/docs/next/checker-loop-hoisting) analysis which identifies loop-invariant function calls, i.e. functions that are pure and have loop-invariant arguments.
|
||
|
|
- [Cost](/docs/next/checker-cost) analysis which identifies control variables in the loop that affect how many times a loop is executed. In this computation, we need to prune control variables that do not affect how many times a loop is executed. In this pruning step, we need to compute loop-invariant variables (which requires the above analysis).
|