PHP RFC: Reclassifying engine warnings
- Date: 2019-08-27
- Author: Nikita Popov email@example.com
- Status: Voting
- Target Version: PHP 8.0
While newly introduced error conditions in the engine typically use
Error exceptions, we have many old error conditions that use inappropriate severity levels for historical reasons. For example, accessing an undefined variable, while being a very severe programming error, only generates a notice. This RFC proposes to reevaluate existing error conditions and reclassify their severity level as appropriate.
As we don't have any existing rules on the matter, here are some general guidelines I try to follow in the following reclassification:
- Error exceptions should be the baseline for error conditions that indicate a programming error.
- If there is an expectation that a certain error condition is commonly intentionally suppressed, especially in legacy code, an exception should not be used.
- If the error condition is data-dependent, it may be preferable not to use an exception.
- For error conditions that have known false positives, a notice should be used.
- Avoid promoting from notice directly to Error exception. I'm only proposing this for the case of undefined variables, because it is so severely misclassified right now.
The following table contains a list of errors with notice or warning severity generated in the engine, excluding warnings that are generated by functions which have an optimized opcode implementation.
The table shows both the current error level, as well as the proposed level. A rationale for the proposed change (or non-change) is provided below each group of errors.
The “undefined variable”, “undefined array index” and “division by zero” error conditions are discussed separately below, because they are more controversial.
|Message||Current Level||Proposed Level|
|Attempt to increment/decrement property '%s' of non-object||Warning||Error exception|
|Attempt to modify property '%s' of non-object||Warning||Error exception|
|Attempt to assign property '%s' of non-object||Warning||Error exception|
|Creating default object from empty value||Warning||Error exception|
|Rationale: These errors are generated when a property is accessed on a non-object inside a write context. If the non-object is “truthy” a warning is generated and the operation is ignored, if it is “falsy” an empty stdClass object is created instead. While auto-vivification is a core part of the language for arrays, the same is not the case for objects, and creating a property on a non-object is almost certainly a programming error rather than an intentional action.|
|Trying to get property '%s' of non-object||Notice||Warning|
|Undefined property: %s::$%s||Notice||Warning|
|Rationale: The first warning is for the same case as above, but for read contexts. This is classified as a warning, because it usually indicates a programming error (in modern code, all non-magic properties tend to be known and fixed). However, object properties can also be dynamic (e.g. JSON in object form), in which case accessing an undefined property may be a less severe issue. Generally, PHP is somewhat lenient with read accesses to “missing” data.|
|Cannot add element to the array as the next element is already occupied||Warning||Error exception|
| Rationale: This error condition occurs when trying to push to an array for which the
|Cannot unset offset in a non-array variable||Warning||Error exception|
|Cannot use a scalar value as an array||Warning||Error exception|
|Trying to access array offset on value of type %s||Notice||Warning|
|Rationale: These diagnostics are generated when trying to use scalars as arrays. The first two occur in write contexts, the latter in read contexts. The latter was introduced in PHP 7.4 as a notice with express intention to elevate the severity in PHP 8.0. In line with the symmetrical case on objects, the write case is treated more severely here, as it usually implies data loss.|
|Only arrays and Traversables can be unpacked||Warning||TypeError exception|
|Invalid argument supplied for foreach()||Warning||TypeError exception|
|Rationale: These are simple type errors and should be treated as such.|
|Illegal offset type||Warning||TypeError exception|
|Illegal offset type in isset or empty||Warning||TypeError exception|
|Illegal offset type in unset||Warning||TypeError exception|
|Rationale: These are generated if an array or object is used as an array key. Once again this is a simple type error.|
|Indirect modification of overloaded element of %s has no effect||Notice||(Notice)|
|Indirect modification of overloaded property %s::$%s has no effect||Notice||(Notice)|
| Rationale: These notices occur if
|Object of class %s could not be converted to int/float/number||Notice||(Notice)|
| Rationale: Comparison between objects and scalars currently works by casting the object to the appropriate type, which is why comparisons like
|A non-numeric value encountered||Warning||(Warning)|
|A non well formed numeric value encountered||Notice||(Notice)|
|Rationale: The difference between these two warnings is whether a string is completely non-numeric, or whether it has a numeric prefix. This is a runtime issue based on the specific string value involved in an operation, which may be user-controlled. For this reason we don't promote to exceptions.|
|Accessing static property %s::$%s as non static||Notice||(Notice)|
| Rationale: This notice is somewhat confusing in what it does: It is thrown when accessing
|Array to string conversion||Notice||Warning|
|Rationale: This is generally a bug (and the “Array” string you get is meaningless), but in many cases also not a particularly severe one. Since string conversion exceptions are supported now, we could also promote this to an Error exception, and I'm generally open to that.|
|Resource ID#%d used as offset, casting to integer (%d)||Notice||Warning|
|Rationale: This is in principle a meaningful operation, but exotic enough that intent should be indicated with an explicit integer cast.|
|String offset cast occurred||Notice||Warning|
|Illegal string offset '%s'||Warning||(Warning)|
|Rationale: The former is thrown when using null/bool/float as a string offset, the latter if the string is not integral. Both of these should use the same severity.|
|Uninitialized string offset: %d||Notice||Warning|
|Illegal string offset: %d||Warning||(Warning)|
|Rationale: The former is used when reading an out-of-bounds string offset, the latter when writing to an out-of-bounds negative string offset (for positive offsets, the string is extended instead). In line with undefined index/property, we consistently generate a warning here.|
|Cannot assign an empty string to a string offset||Warning||Error exception|
|Rationale: This operation is not meaningful and indicates some kind of logic error.|
|Only variables should be passed by reference||Notice||(Notice)|
|Only variable references should be returned by reference||Notice||(Notice)|
|Only variable references should be yielded by reference||Notice||(Notice)|
|Only variables should be assigned by reference||Notice||(Notice)|
|Attempting to set reference to non referenceable value||Notice||(Notice)|
|Cannot pass by-reference argument %d of %s%s%s() by unpacking a Traversable, passing by-value instead||Warning||(Warning)|
|Rationale: The use of values where a reference is expected is currently somewhat inconsistent, with everything from compiler errors, Error exceptions, warnings and notices being possible depending on the specific case. Passing a non-variable to a reference argument is often a programming error, because it will not be possible to modify the passed value and the reference cannot serve its purpose. However, this is complicated by optional reference arguments or return values that are optionally references. In both cases the warning may be a false positive. It's not really clear what to do here, so I'm retaining the current classification for now.|
In most cases, accessing an undefined variable is a severe programming error. The current low classification is a legacy from the Dark Ages of PHP, where features like register_globals made conditionally defined variables more typical, and code quality standards were lower.
Ideally, undefined variables should be compile errors, but as the dynamic nature of PHP precludes a reliable compile-time analysis, this RFC proposes to generate an Error exception instead.
However, throwing an exception may complicate the upgrading of legacy code that currently suppresses the generation of notices wholesale, as the issue can no longer be ignored. Some people have even suggested that the use of undefined variables is a legitimate coding style choice.
For this reason, a separate vote will decide whether we should throw an Error exception, generate a warning or keep the current notice.
Undefined array index
Similarly to undefined variables or an undefined object properties, reading an undefined array index/key would generally be considered a programming error in modern PHP code. However, while variables and object properties are predominantly statically known (i.e., when variable variables and dynamic object properties are not used), the same is not true to array keys, which tend to be dynamic.
A separate vote will decide whether to elevate undefined array offset/index conditions to a warning, or leave them as notices.
Division by zero
Division by zero currently has somewhat inconsistent behavior. The
% operator throws a
DivisionByZeroError. However, the
/ throws a “Division by zero” warning and returns one of +Inf, -Inf or NaN, following IEEE 754 semantics.
The rationale for this behavioral discrepancy is as follows: The
% operator works on integers. As such the conjugated operation is actually not
/, but rather
intdiv(), which does throw
DivisionByZeroError. An integer operation shouldn't return a floating point number, so throwing an Error exception is the only choice here.
Not throwing an Error exception for division by zero using
/ is motivated by the fact that such division does have a well-defined result under IEEE 754. In some areas of application (such as numerics) it may be useful to not treat division by zero as an error condition at all, though such applications are unusual for PHP. Similarly, Bob Weinand argued that for reporting code that makes heavy use of divisions, it may be preferable to have the ability to suppress this error condition.
On the other hand, the current behavior, and especially the discrepancy with
% is quite unexpected, and many people expect that a division by zero error will in fact generate a DivisionByZeroError (duh). This discussion has already come up multiple times on the internals list and in pull requests.
As I think that both sides have a reasonable argument here, there will be a separate vote on whether to change the division by zero behavior.
Backward Incompatible Changes
Conversion of notices to warnings is fairly harmless, because both continue execution after the diagnostic has been generated. Conversion to exceptions implies that the current control flow will be aborted.
This may impact code that makes very liberal use of the error suppression operator
@ or disables error reporting wholesale. The proposal does try to avoid changing notices that are more likely to be suppressed into exceptions.
If desired, an error handler can be provided that filters out the error conditions that will be turned into an exception, so that projects can focus on addressing them prior to an upgrade.
All the following votes are independent. Each requires a 2/3 majority and may pass/fail independent of other votes. Voting closes 2019-09-26.
As the “undefined variable” vote is a 3-way vote, acceptance is determined as follows: “Error exception” is accepted if it has 2/3 majority. Otherwise, “Warning” is accepted if the first two options together have 2/3 majority. Otherwise, “Keep Notice” applies.
- 2019-09-12: Split out undefined index/offset into a separate section.
- 2019-09-10: Keep current classification for reference errors.
- 2019-08-28: Split off the “undefined variable” case into a separate vote, as it was a major point of contention on-list.