| blob | 7a6b5d71b35d240ef528d6e727a9cec2d91841c2 |
1 ===========================
2 Defining Undefined Behavior
3 ===========================
5 .. contents:: Table of Contents
6 :depth: 2
7 :local:
9 The C standard leaves behavior undefined or implementation defined in many
10 places. Undefined behavior is behavior that the standards leave up to the
11 implementation. As an implementation, LLVM's libc must provide a result for any
12 input, including inputs for which the result is undefined. This page provides
13 examples of how these situations are handled in existing code, followed by
14 guidelines to help determine the right approach for new situations.
16 Guidelines
17 ==========
18 Most undefined behavior is outside the scope of normal use. Follow these
19 guidelines and the resulting code should behave predictably even in unexpected
20 situations.
22 #. Follow the standards.
23 #. Avoid giving an incorrect answer.
24 #. In general, correct answer > correct answer (wrong format) > no answer > crash the program >>>>>>> incorrect answer.
25 #. The C library is called frequently in performance critical situations, and so can't afford to do thorough error checking and correction.
26 #. It also cannot give the incorrect answer for any reasonable input, since it is so foundational.
27 #. This leaves crashing or address space corruption as a probable option for a libc function in an ambiguous state.
28 #. Don't overcomplicate undefined situations.
29 #. It's better to have a slightly confusing result for an undefined input than 100 extra lines of code that are never used for a well defined input.
30 #. LLVM's libc is also used for embedded systems that care a lot about code size.
31 #. Unreasonable inputs can have unreasonable outputs.
32 #. Match other implementations when it makes sense.
33 #. Every libc has to make these choices, and sometimes others have already found the right choice.
34 #. Be careful, just because there is a consensus doesn't make that consensus right.
35 #. LLVM's libc should be consistent with itself.
36 #. Similar inputs to the same function should yield similar results, even when the inputs are undefined.
37 #. The same input to similar functions should also yield similar results.
38 #. The same input to the same function on different platforms should yield the same result, unless there's a specific reason not to (e.g. 64 bit long vs 32 bit long).
39 #. Write down the decision.
40 #. Every libc has to make a decision on how to handle undefined inputs. Users should be able to find what LLVM's libc does.
41 #. While users shouldn't rely on undefined behavior, it shouldn't surprise them.
43 Approaches
44 ==========
46 Matching Behavior Against Existing Implementations
47 --------------------------------------------------
48 Existing implementations have already chosen how to handle undefined situations, and sometimes there are benefits to matching those decisions, such as in the case of atoi. The C Standard defines atoi as being equivalent to a call to strtol, with the result cast from long to int. The standard also clarifies that any input that cannot be represented as an int causes undefined behavior. For the strtol function, the standard instead defines inputs that cannot be represented in a long int as returning LONG_MAX or LONG_MIN, according to their sign. The decision of whether to cast the result from strtol or to handle integer overflow like strtol does is left to the implementation. LLVM's libc performs the raw cast, since the atoi function is fuzz tested against the implementation from glibc. By matching a known good implementation, LLVM's libc can more effectively be checked for correctness in this case.
50 Simplifying Handling Invalid Inputs
51 -----------------------------------
52 When handling invalid inputs, the output should be simple to code, and simple for the user to understand. An example of this is how the printf function handles invalid conversion specifiers. A conversion specifier is a segment of the format string that starts with a %. At the end of a conversion specifier is the character that determines the behavior for the conversion, called the conversion name. As an example, the conversion specifier %d has the conversion name of d which represents an integer conversion. If the conversion name is instead an invalid character such as ? then the behavior is undefined. When passed an invalid conversion specifier like %? LLVM's libc defines the output as the raw text of the conversion specifier. This simplifies the algorithm and makes the result obvious and predictable for the user.
54 Conforming to Existing Practice
55 -------------------------------
56 There are some behaviors that are technically undefined, but are otherwise consistent across implementations, such as how printf handles length modifiers on inappropriate conversions. For each conversion name there is a list of length modifiers that can apply to it. If a length modifier is applied to a conversion specifier that it doesn't apply to, then the behavior is undefined. For most conversions, LLVM's libc ignores any length modifier that doesn't apply. As an example, a conversion of %hf would be read as an f float conversion with the h length modifier. The h length modifier doesn't apply to floating point conversions and so %hf is the same as %f. There is one exception, which is the L length modifier on integer conversions. Many libcs handle the L length modifier like the ll length modifier when applied to integer conversions, despite L only applying to float conversions in the standard. LLVM's libc follows this convention because it is a useful feature that is simple to implement and has a predictable outcome for the user.
58 Interpreting the Standard's Reasoning
59 -------------------------------------
60 Often the standard will imply an intended behavior through what it states is undefined, such as in the case of printf's handling of the %% conversion. The %% conversion is used to write a % character, since it's used as the start of a conversion specifier. The standard specifies that %% must be the complete conversion specifier, and any options would make the conversion undefined. The conversion specifier %10% can therefore be interpreted as a % conversion with a width of 10, but the standard implies that this is not necessary. By making the options undefined, the standard implies a desired behavior for %% with options. The implied behavior is to ignore all options and always print %. This still leaves the behavior of %*% ambiguous, since the star normally consumes an argument to be used as the width. Since % conversions ignore the width, it would be reasonable to not read the argument in this case, but it would add additional complexity to the parsing logic. For that reason, the implementation in LLVM's libc will consume an argument for %*%, although the value is ignored. Adding additional logic for unreasonable edge cases, such as this one, is unnecessary.
62 Ignoring Bug-For-Bug Compatibility
63 ----------------------------------
64 Any long running implementations will have bugs and deviations from the standard. Hyrum's Law states that “all observable behaviors of your system will be depended on by somebody” which includes these bugs. An example of a long-standing bug is glibc's scanf float parsing behavior. The behavior is specifically defined in the standard, but it isn't adhered to by all libc implementations. There is a longstanding bug in glibc where it incorrectly parses the string 100er and this caused the C standard to add that specific example to the definition for scanf. The intended behavior is for scanf, when parsing a float, to parse the longest possibly valid prefix and then accept it if and only if that complete parsed value is a float. In the case of 100er the longest possibly valid prefix is 100e but the float parsed from that string is only 100. Since there is no number after the e it shouldn't be included in the float, so scanf should return a parsing error. For LLVM's libc it was decided to follow the standard, even though glibc's version is slightly simpler to implement and this edge case is rare. Following the standard must be the first priority, since that's the goal of the library.