Lint, a C Program Checker S. C. Johnson is a command which exam- ines C source programs, detecting a number of bugs and obscuri- ties. It enforces the type rules of C more strictly than the C compilers. It may also be used to enforce a number of portabili- ty restrictions involved in moving programs between different machines and/or operating systems. Another option detects a number of wasteful, or error prone, constructions which neverthe- less are, strictly speaking, legal. accepts multiple input files and library specifications, and checks them for consistency. The separation of function between and the C compilers has both his- torical and practical rationale. The compilers turn C programs into executable files rapidly and efficiently. This is possible in part because the compilers do not do sophisticated type check- ing, especially between separately compiled programs. takes a more global, leisurely view of the program, looking much more carefully at the compatibilities. This document discusses the use of gives an overview of the implementation, and gives some hints on the writing of machine independent C code. Introduction and Usage Suppose there are two C Kernighan Ritchie Programming Prentice 1978 source files, and which are ordinarily compiled and loaded together. Then the command lint file1.c file2.c pro- duces messages describing inconsistencies and inefficiencies in the programs. The program enforces the typing rules of C more strictly than the C compilers (for both historical and practical reasons) enforce them. The command lint -p file1.c file2.c will produce, in addition to the above messages, additional mes- sages which relate to the portability of the programs to other operating systems and machines. Replacing the by will produce messages about various error-prone or wasteful constructions which, strictly speaking, are not bugs. Saying gets the whole works. The next several sections describe the major messages; the document closes with sections discussing the implementation and giving suggestions for writing portable C. An appendix gives a summary of the options. A Word About Philosophy Many of the facts which needs may be impossible to discover. For example, whether a given function in a program ever gets called may depend on the input data. Deciding whether is ever called is equivalent to solving the famous ``halting problem,'' known to be recursive- ly undecidable. Thus, most of the algorithms are a compromise. If a function is never mentioned, it can never be called. If a function is mentioned, assumes it can be called; this is not necessarily so, but in practice is quite reasonable. tries to give information with a high degree of relevance. Messages of the form ``_x_x_x might be a bug'' are easy to generate, but are ac- ceptable only in proportion to the fraction of real bugs they un- cover. If this fraction of real bugs is too small, the messages lose their credibility and serve merely to clutter up the output, obscuring the more important messages. Keeping these issues in mind, we now consider in more detail the classes of messages which produces. Unused Variables and Functions As sets of pro- grams evolve and develop, previously used variables and arguments to functions may become unused; it is not uncommon for external variables, or even entire functions, to become unnecessary, and yet not be removed from the source. These ``errors of commis- sion'' rarely cause working programs to fail, but they are a source of inefficiency, and make programs harder to understand and change. Moreover, information about such unused variables and functions can occasionally serve to discover bugs; if a func- tion does a necessary job, and is never called, something is wrong! complains about variables and functions which are defined but not otherwise mentioned. An exception is variables which are declared through explicit statements but are never referenced; thus the statement extern float sin(); will evoke no comment if is never used. Note that this agrees with the semantics of the C compiler. In some cases, these unused external declarations might be of some interest; they can be discovered by adding the flag to the invocation. Certain styles of programming require many functions to be written with similar interfaces; frequently, some of the arguments may be unused in many of the calls. The option is available to suppress the printing of complaints about unused arguments. When is in effect, no messages are produced about unused arguments except for those arguments which are unused and also declared as register arguments; this can be con- sidered an active (and preventable) waste of the register resources of the machine. There is one case where information about unused, or undefined, variables is more distracting than helpful. This is when is applied to some, but not all, files out of a collection which are to be loaded together. In this case, many of the functions and variables defined may not be used, and, conversely, many functions and variables defined elsewhere may be used. The flag may be used to suppress the spurious messages which might otherwise appear. Set/Used Information attempts to detect cases where a variable is used before it is set. This is very difficult to do well; many algorithms take a good deal of time and space, and still produce messages about perfectly valid programs. detects local variables (automatic and register storage classes) whose first use appears physically earlier in the input file than the first assignment to the variable. It as- sumes that taking the address of a variable constitutes a ``use,'' since the actual use may occur at any later time, in a data dependent fashion. The restriction to the physical appear- ance of variables in the file makes the algorithm very simple and quick to implement, since the true flow of control need not be discovered. It does mean that can complain about some programs which are legal, but these programs would probably be considered bad on stylistic grounds (e.g. might contain at least two goto's). Because static and external variables are initialized to 0, no meaningful information can be discovered about their uses. The algorithm deals correctly, however, with initialized automatic variables, and variables which are used in the expres- sion which first sets them. The set/used information also per- mits recognition of those local variables which are set and never used; these form a frequent source of inefficiencies, and may also be symptomatic of bugs. Flow of Control attempts to detect unreachable portions of the programs which it processes. It will complain about unlabeled statements immediately following goto, break, continue, or return statements. An attempt is made to detect loops which can never be left at the bottom, detecting the special cases while( 1 ) and for(;;) as infinite loops. also complains about loops which cannot be entered at the top; some valid programs may have such loops, but at best they are bad style, at worst bugs. has an important area of blindness in the flow of control algorithm: it has no way of detecting functions which are called and never return. Thus, a call to may cause un- reachable code which does not detect; the most serious effects of this are in the determination of returned function values (see the next section). One form of unreachable statement is not usu- ally complained about by a statement that cannot be reached causes no message. Programs generated by Johnson Yacc 1975 and especially Lesk Lex may have literally hundreds of unreachable statements. The flag in the C compiler will often eliminate the resulting object code inefficiency. Thus, these unreached state- ments are of little importance, there is typically nothing the user can do about them, and the resulting messages would clutter up the output. If these messages are desired, can be invoked with the option. Function Values Sometimes functions return values which are never used; sometimes programs incorrectly use function ``values'' which have never been returned. addresses this problem in a number of ways. Locally, within a function de- finition, the appearance of both return( _e_x_p_r ); and return ; statements is cause for alarm; will give the message function _n_a_m_e contains return(e) and return The most serious difficulty with this is detecting when a function return is implied by flow of control reaching the end of the function. This can be seen with a simple example: f ( a ) { if ( a ) return ( 3 ); g (); } Notice that, if _a tests false, _f will call _g and then return with no defined return value; this will trigger a complaint from If _g, like _e_x_i_t, never returns, the message will still be produced when in fact nothing is wrong. In practice, some potentially serious bugs have been discovered by this feature; it also accounts for a substantial fraction of the ``noise'' messages produced by On a global scale, detects cases where a function returns a value, but this value is sometimes, or always, unused. When the value is always unused, it may consti- tute an inefficiency in the function definition. When the value is sometimes unused, it may represent bad style (e.g., not test- ing for error conditions). The dual problem, using a function value when the function does not return one, is also detected. This is a serious problem. Amazingly, this bug has been observed on a couple of occasions in ``working'' programs; the desired function value just happened to have been computed in the func- tion return register! Type Checking enforces the type checking rules of C more strictly than the compilers do. The additional checking is in four major areas: across certain binary operators and implied assignments, at the structure selection operators, between the definition and uses of functions, and in the use of enumerations. There are a number of operators which have an im- plied balancing between types of the operands. The assignment, conditional ( ?: ), and relational operators have this property; the argument of a return statement, and expressions used in ini- tialization also suffer similar conversions. In these opera- tions, char, short, int, long, unsigned, float, and double types may be freely intermixed. The types of pointers must agree ex- actly, except that arrays of _x's can, of course, be intermixed with pointers to _x's. The type checking rules also require that, in structure references, the left operand of the -> be a pointer to structure, the left operand of the . be a structure, and the right operand of these operators be a member of the structure im- plied by the left operand. Similar checking is done for refer- ences to unions. Strict rules apply to function argument and re- turn value matching. The types float and double may be freely matched, as may the types char, short, int, and unsigned. Also, pointers can be matched with the associated arrays. Aside from this, all actual arguments must agree in type with their declared counterparts. With enumerations, checks are made that enumera- tion variables or members are not mixed with other types, or oth- er enumerations, and that the only operations applied are =, ini- tialization, ==, !=, and function arguments and return values. Type Casts The type cast feature in C was introduced largely as an aid to producing more portable programs. Consider the assign- ment p = 1 ; where is a character pointer. will quite rightly complain. Now, consider the assignment p = (char *)1 ; in which a cast has been used to convert the integer to a character pointer. The programmer obviously had a strong motivation for doing this, and has clearly signaled his intentions. It seems harsh for to continue to complain about this. On the other hand, if this code is moved to another machine, such code should be looked at carefully. The flag controls the printing of comments about casts. When is in effect, casts are treated as though they were assignments subject to complaint; otherwise, all legal casts are passed without comment, no matter how strange the type mixing seems to be. Nonportable Character Use On the PDP-11, characters are signed quantities, with a range from -128 to 127. On most of the other C implementations, characters take on only positive values. Thus, will flag certain comparisons and assignments as being illegal or nonportable. For example, the fragment char c; ... if( (c = getchar()) < 0 ) .... works on the PDP-11, but will fail on machines where characters always take on posi- tive values. The real solution is to declare an integer, since is actually returning integer values. In any case, will say ``nonportable character comparison''. A similar issue arises with bitfields; when assignments of constant values are made to bitfields, the field may be too small to hold the value. This is especially true because on some machines bitfields are considered as signed quantities. While it may seem unintuitive to consider that a two bit field declared of type cannot hold the value 3, the problem disappears if the bitfield is declared to have type Assignments of longs to ints Bugs may arise from the assignment of to an which loses accuracy. This may happen in programs which have been incompletely converted to use When a variable is changed from int to long, the program can stop working because some intermediate results may be assigned to ints, losing accura- cy. Since there are a number of legitimate reasons for assigning longs to ints, the detection of these assignments is enabled by the flag. Strange Constructions Several perfectly legal, but somewhat strange, constructions are flagged by the messages hope- fully encourage better code quality, clearer style, and may even point out bugs. The flag is used to enable these checks. For example, in the statement *p++ ; the * does nothing; this pro- vokes the message ``null effect'' from The program fragment un- signed x ; if( x < 0 ) ... is clearly somewhat strange; the test will never succeed. Similarly, the test if( x > 0 ) ... is equivalent to if( x != 0 ) which may not be the intended action. will say ``degenerate unsigned comparison'' in these cases. If one says if( 1 != 0 ) .... will report ``constant in conditional context'', since the comparison of 1 with 0 gives a constant result. Another construction detected by involves operator pre- cedence. Bugs which arise from misunderstandings about the pre- cedence of operators can be accentuated by spacing and format- ting, making such bugs extremely hard to find. For example, the statements if( x&077 == 0 ) ... or x<<2 + 40 probably do not do what was intended. The best solution is to parenthesize such ex- pressions, and encourages this by an appropriate message. Final- ly, when the flag is in force complains about variables which are redeclared in inner blocks in a way that conflicts with their use in outer blocks. This is legal, but is considered by many (in- cluding the author) to be bad style, usually unnecessary, and frequently a bug. Ancient History There are several forms of older syntax which are being officially discouraged. These fall into two classes, assignment operators and initialization. The older forms of assignment operators (e.g., =+, =-, . . . ) could cause ambiguous expressions, such as a =-1 ; which could be tak- en as either a =- 1 ; or a = -1 ; The situation is especially perplexing if this kind of ambiguity arises as the result of a macro substitution. The newer, and preferred operators (+=, -=, etc. ) have no such ambiguities. To spur the abandonment of the older forms, complains about these old fashioned operators. A similar issue arises with initialization. The older language al- lowed int x 1 ; to initialize to 1. This also caused syntactic difficulties: for example, int x ( -1 ) ; looks somewhat like the beginning of a function declaration: int x ( y ) { . . . and the compiler must read a fair ways past in order to sure what the declaration really is.. Again, the problem is even more per- plexing when the initializer involves a macro. The current syn- tax places an equals sign between the variable and the initializ- er: int x = -1 ; This is free of any possible syntactic ambi- guity. Pointer Alignment Certain pointer assignments may be rea- sonable on some machines, and illegal on others, due entirely to alignment restrictions. For example, on the PDP-11, it is rea- sonable to assign integer pointers to double pointers, since dou- ble precision values may begin on any integer boundary. On the Honeywell 6000, double precision values must begin on even word boundaries; thus, not all such assignments make sense. tries to detect cases where pointers are assigned to other pointers, and such alignment problems might arise. The message ``possible pointer alignment problem'' results from this situation whenever either the or flags are in effect. Multiple Uses and Side Ef- fects In complicated expressions, the best order in which to evaluate subexpressions may be highly machine dependent. For ex- ample, on machines (like the PDP-11) in which the stack runs backwards, function arguments will probably be best evaluated from right-to-left; on machines with a stack running forward, left-to-right seems most attractive. Function calls embedded as arguments of other functions may or may not be treated similarly to ordinary arguments. Similar issues arise with other operators which have side effects, such as the assignment operators and the increment and decrement operators. In order that the efficiency of C on a particular machine not be unduly compromised, the C language leaves the order of evaluation of complicated expres- sions up to the local compiler, and, in fact, the various C com- pilers have considerable differences in the order in which they will evaluate complicated expressions. In particular, if any variable is changed by a side effect, and also used elsewhere in the same expression, the result is explicitly undefined. checks for the important special case where a simple scalar variable is affected. For example, the statement _a[_i] = _b[_i++] ; will draw the complaint: warning: _i evaluation order undefined Implementa- tion consists of two programs and a driver. The first program is a version of the Portable C Compiler Johnson Ritchie BSTJ Porta- bility Programs System Johnson portable compiler 1978 which is the basis of the IBM 370, Honeywell 6000, and Interdata 8/32 C compilers. This compiler does lexical and syntax analysis on the input text, constructs and maintains symbol tables, and builds trees for expressions. Instead of writing an intermediate file which is passed to a code generator, as the other compilers do, produces an intermediate file which consists of lines of ascii text. Each line contains an external variable name, an encoding of the context in which it was seen (use, definition, declara- tion, etc.), a type specifier, and a source file name and line number. The information about variables local to a function or file is collected by accessing the symbol table, and examining the expression trees. Comments about local problems are produced as detected. The information about external names is collected onto an intermediate file. After all the source files and li- brary descriptions have been collected, the intermediate file is sorted to bring all information collected about a given external name together. The second, rather small, program then reads the lines from the intermediate file and compares all of the defini- tions, declarations, and uses for consistency. The driver con- trols this process, and is also responsible for making the op- tions available to both passes of Portability C on the Honeywell and IBM systems is used, in part, to write system code for the host operating system. This means that the implementation of C tends to follow local conventions rather than adhere strictly to system conventions. Despite these differences, many C programs have been successfully moved to GCOS and the various IBM instal- lations with little effort. This section describes some of the differences between the implementations, and discusses the features which encourage portability. Uninitialized external variables are treated differently in different implementations of C. Suppose two files both contain a declaration without initial- ization, such as int a ; outside of any function. The loader will resolve these declarations, and cause only a single word of storage to be set aside for _a. Under the GCOS and IBM implemen- tations, this is not feasible (for various stupid reasons!) so each such declaration causes a word of storage to be set aside and called _a. When loading or library editing takes place, this causes fatal conflicts which prevent the proper operation of the program. If is invoked with the -p flag, it will detect such multiple definitions. A related difficulty comes from the amount of information retained about external names during the loading process. On the system, externally known names have seven signi- ficant characters, with the upper/lower case distinction kept. On the IBM systems, there are eight significant characters, but the case distinction is lost. On GCOS, there are only six char- acters, of a single case. This leads to situations where pro- grams run on the system, but encounter loader problems on the IBM or GCOS systems. causes all external symbols to be mapped to one case and truncated to six characters, providing a worst-case analysis. A number of differences arise in the area of character handling: characters in the system are eight bit ascii, while they are eight bit ebcdic on the IBM, and nine bit ascii on GCOS. Moreover, character strings go from high to low bit positions (``left to right'') on GCOS and IBM, and low to high (``right to left'') on the PDP-11. This means that code attempting to con- struct strings out of character constants, or attempting to use characters as indices into arrays, must be looked at with great suspicion. is of little help here, except to flag multi- character character constants. Of course, the word sizes are different! This causes less trouble than might be expected, at least when moving from the system (16 bit words) to the IBM (32 bits) or GCOS (36 bits). The main problems are likely to arise in shifting or masking. C now supports a bit-field facility, which can be used to write much of this code in a reasonably portable way. Frequently, portability of such code can be enhanced by slight rearrangements in coding style. Many of the incompatibilities seem to have the flavor of writing x &= 0177700 ; to clear the low order six bits of _x. This suffices on the PDP-11, but fails badly on GCOS and IBM. If the bit field feature cannot be used, the same effect can be obtained by writ- ing x &= 9~8 077 ; which will work on all these machines. The right shift operator is arithmetic shift on the PDP-11, and logi- cal shift on most other machines. To obtain a logical shift on all machines, the left operand can be typed unsigned. Characters are considered signed integers on the PDP-11, and unsigned on the other machines. This persistence of the sign bit may be reason- ably considered a bug in the PDP-11 hardware which has infiltrat- ed itself into the C language. If there were a good way to dis- cover the programs which would be affected, C could be changed; in any case, is no help here. The above discussion may have made the problem of portability seem bigger than it in fact is. The issues involved here are rarely subtle or mysterious, at least to the implementor of the program, although they can involve some work to straighten out. The most serious bar to the portability of system utilities has been the inability to mimic essential system functions on the other systems. The inability to seek to a random character position in a text file, or to establish a pipe between processes, has involved far more rewriting and de- bugging than any of the differences in C compilers. On the other hand, has been very helpful in moving the operating system and associated utility programs to other machines. Shutting Lint Up There are occasions when the programmer is smarter than There may be valid reasons for ``illegal'' type casts, functions with a variable number of arguments, etc. Moreover, as specified above, the flow of control information produced by often has blind spots, causing occasional spurious messages about perfectly rea- sonable programs. Thus, some way of communicating with typically to shut it up, is desirable. The form which this mechanism should take is not at all clear. New keywords would require current and old compilers to recognize these keywords, if only to ignore them. This has both philosophical and practical problems. New preprocessor syntax suffers from similar problems. What was finally done was to cause a number of words to be recognized by when they were embedded in comments. This required minimal preprocessor changes; the preprocessor just had to agree to pass comments through to its output, instead of deleting them as had been previously done. Thus, directives are invisible to the com- pilers, and the effect on systems with the older preprocessors is merely that the directives don't work. The first directive is concerned with flow of control information; if a particular place in the program cannot be reached, but this is not apparent to this can be asserted by the directive /* NOTREACHED */ at the ap- propriate spot in the program. Similarly, if it is desired to turn off strict type checking for the next expression, the direc- tive /* NOSTRICT */ can be used; the situation reverts to the previous default after the next expression. The flag can be turned on for one function by the directive /* ARGSUSED */ Com- plaints about variable number of arguments in calls to a function can be turned off by the directive /* VARARGS */ preceding the function definition. In some cases, it is desirable to check the first several arguments, and leave the later arguments unchecked. This can be done by following the VARARGS keyword immediately with a digit giving the number of arguments which should be checked; thus, /* VARARGS2 */ will cause the first two arguments to be checked, the others unchecked. Finally, the directive /* LINTLIBRARY */ at the head of a file identifies this file as a library declaration file; this topic is worth a section by it- self. Library Declaration Files accepts certain library direc- tives, such as -ly and tests the source files for compatibility with these libraries. This is done by accessing library descrip- tion files whose names are constructed from the library direc- tives. These files all begin with the directive /* LINTLIBRARY */ which is followed by a series of dummy function definitions. The critical parts of these definitions are the declaration of the function return type, whether the dummy function returns a value, and the number and types of arguments to the function. The VARARGS and ARGSUSED directives can be used to specify features of the library functions. library files are processed almost exactly like ordinary source files. The only difference is that functions which are defined on a library file, but are not used on a source file, draw no complaints. does not simulate a full library search algorithm, and complains if the source files contain a redefinition of a library routine (this is a feature!). By default, checks the programs it is given against a standard library file, which contains descriptions of the pro- grams which are normally loaded when a C program is run. When the flag is in effect, another file is checked containing descriptions of the standard I/O library routines which are ex- pected to be portable across various machines. The flag can be used to suppress all library checking. Bugs, etc. was a diffi- cult program to write, partially because it is closely connected with matters of programming style, and partially because users usually don't notice bugs which cause to miss errors which it should have caught. (By contrast, if incorrectly complains about something that is correct, the programmer reports that immediate- ly!) A number of areas remain to be further developed. The checking of structures and arrays is rather inadequate; size in- compatibilities go unchecked, and no attempt is made to match up structure and union declarations across files. Some stricter checking of the use of the is clearly desirable, but what check- ing is appropriate, and how to carry it out, is still to be determined. shares the preprocessor with the C compiler. At some point it may be appropriate for a special version of the preprocessor to be constructed which checks for things such as unused macro definitions, macro arguments which have side effects which are not expanded at all, or are expanded more than once, etc. The central problem with is the packaging of the informa- tion which it collects. There are many options which serve only to turn off, or slightly modify, certain features. There are pressures to add even more of these options. In conclusion, it appears that the general notion of having two programs is a good one. The compiler concentrates on quickly and accurately turning the program text into bits which can be run; concentrates on is- sues of portability, style, and efficiency. can afford to be wrong, since incorrectness and over-conservatism are merely an- noying, not fatal. The compiler can be fast since it knows that will cover its flanks. Finally, the programmer can concentrate at one stage of the programming process solely on the algorithms, data structures, and correctness of the program, and then later retrofit, with the aid of the desirable properties of universali- ty and portability. $LIST$ Appendix: Current Lint Options The command currently has the form lint [-options ] files... library-descriptors... The op- tions are Perform heuristic checks Perform portability checks Don't report unused arguments Don't report unused or undefined externals Report unreachable statements. Report unused external declarations Report assignments of to or shorter. Complain about questionable casts No library checking is done Same as (for his- torical reasons)