This is Info file f/g77.info, produced by Makeinfo version 1.68 from the input file ../../../src/gcc-2.95.3/gcc/f/g77.texi. INFO-DIR-SECTION Programming START-INFO-DIR-ENTRY * g77: (g77). The GNU Fortran compiler. END-INFO-DIR-ENTRY This file documents the use and the internals of the GNU Fortran (`g77') compiler. It corresponds to the GCC-2.95 version of `g77'. Published by the Free Software Foundation 59 Temple Place - Suite 330 Boston, MA 02111-1307 USA Copyright (C) 1995-1999 Free Software Foundation, Inc. Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies. Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided also that the sections entitled "GNU General Public License," "Funding for Free Software," and "Protect Your Freedom--Fight `Look And Feel'" are included exactly as in the original, and provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one. Permission is granted to copy and distribute translations of this manual into another language, under the above conditions for modified versions, except that the sections entitled "GNU General Public License," "Funding for Free Software," and "Protect Your Freedom--Fight `Look And Feel'", and this permission notice, may be included in translations approved by the Free Software Foundation instead of in the original English. Contributed by James Craig Burley (). Inspired by a first pass at translating `g77-0.5.16/f/DOC' that was contributed to Craig by David Ronis ().  File: g77.info, Node: Gotchas (Transforming), Next: TBD (Transforming), Prev: ste.c, Up: Overview of Translation Process Gotchas (Transforming) ---------------------- This section is not about transforming "gotchas" into something else. It is about the weirder aspects of transforming Fortran, however that's defined, into a more modern, canonical form. Multi-character Lexemes ....................... Each lexeme carries with it a pointer to where it appears in the source. To provide the ability for diagnostics to point to column numbers, in addition to line numbers and names, lexemes that represent more than one (significant) character in the source code need, generally, to provide pointers to where each *character* appears in the source. This provides the ability to properly identify the precise location of the problem in code like SUBROUTINE X END BLOCK DATA X END which, in fixed-form source, would result in single lexemes consisting of the strings `SUBROUTINEX' and `BLOCKDATAX'. (The problem is that `X' is defined twice, so a pointer to the `X' in the second definition, as well as a follow-up pointer to the corresponding pointer in the first, would be preferable to pointing to the beginnings of the statements.) This need also arises when parsing (and diagnosing) `FORMAT' statements. Further, it arises when diagnosing `FMT=' specifiers that contain constants (or partial constants, or even propagated constants!) in I/O statements, as in: PRINT '(I2, 3HAB)', J (A pointer to the beginning of the prematurely-terminated Hollerith constant, and/or to the close parenthese, is preferable to a pointer to the open-parenthese or the apostrophe that precedes it.) Multi-character lexemes, which would seem to naturally include at least digit strings, alphanumeric strings, `CHARACTER' constants, and Hollerith constants, therefore need to provide location information on each character. (Maybe Hollerith constants don't, but it's unnecessary to except them.) The question then arises, what about *other* multi-character lexemes, such as `**' and `//', and Fortran 90's `(/', `/)', `::', and so on? Turns out there's a need to identify the location of the second character of these two-character lexemes. For example, in `I(/J) = K', the slash needs to be diagnosed as the problem, not the open parenthese. Similarly, it is preferable to diagnose the second slash in `I = J // K' rather than the first, given the implicit typing rules, which would result in the compiler disallowing the attempted concatenation of two integers. (Though, since that's more of a semantic issue, it's not *that* much preferable.) Even sequences that could be parsed as digit strings could use location info, for example, to diagnose the `9' in the octal constant `O'129''. (This probably will be parsed as a character string, to be consistent with the parsing of `Z'129A''.) To avoid the hassle of recording the location of the second character, while also preserving the general rule that each significant character is distinctly pointed to by the lexeme that contains it, it's best to simply not have any fixed-size lexemes larger than one character. This new design is expected to make checking for two `*' lexemes in a row much easier than the old design, so this is not much of a sacrifice. It probably makes the lexer much easier to implement than it makes the parser harder. Space-padding Lexemes ..................... Certain lexemes need to be padded with virtual spaces when the end of the line (or file) is encountered. This is necessary in fixed form, to handle lines that don't extend to column 72, assuming that's the line length in effect. Bizarre Free-form Hollerith Constants ..................................... Last I checked, the Fortran 90 standard actually required the compiler to silently accept something like FORMAT ( 1 2 Htwelve chars ) as a valid `FORMAT' statement specifying a twelve-character Hollerith constant. The implication here is that, since the new lexer is a zero-feedback one, it won't know that the special case of a `FORMAT' statement being parsed requires apparently distinct lexemes `1' and `2' to be treated as a single lexeme. (This is a horrible misfeature of the Fortran 90 language. It's one of many such misfeatures that almost make me want to not support them, and forge ahead with designing a new "GNU Fortran" language that has the features, but not the misfeatures, of Fortran 90, and provide utility programs to do the conversion automatically.) So, the lexer must gather distinct chunks of decimal strings into a single lexeme in contexts where a single decimal lexeme might start a Hollerith constant. (Which probably means it might as well do that all the time for all multi-character lexemes, even in free-form mode, leaving it to subsequent phases to pull them apart as they see fit.) Compare the treatment of this to how CHARACTER * 4 5 HEY and CHARACTER * 12 HEY must be treated--the former must be diagnosed, due to the separation between lexemes, the latter must be accepted as a proper declaration. Hollerith Constants ................... Recognizing a Hollerith constant--specifically, that an `H' or `h' after a digit string begins such a constant--requires some knowledge of context. Hollerith constants (such as `2HAB') can appear after: * `(' * `,' * `=' * `+', `-', `/' * `*', except as noted below Hollerith constants don't appear after: * `CHARACTER*', which can be treated generally as any `*' that is the second lexeme of a statement Confusing Function Keyword .......................... While REAL FUNCTION FOO () must be a `FUNCTION' statement and REAL FUNCTION FOO (5) must be a type-definition statement, REAL FUNCTION FOO (NAMES) where NAMES is a comma-separated list of names, can be one or the other. The only way to disambiguate that statement (short of mandating free-form source or a short maximum length for name for external procedures) is based on the context of the statement. In particular, the statement is known to be within an already-started program unit (but not at the outer level of the `CONTAINS' block), it is a type-declaration statement. Otherwise, the statement is a `FUNCTION' statement, in that it begins a function program unit (external, or, within `CONTAINS', nested). Weird READ .......... The statement READ (N) is equivalent to either READ (UNIT=(N)) or READ (FMT=(N)) depending on which would be valid in context. Specifically, if `N' is type `INTEGER', `READ (FMT=(N))' would not be valid, because parentheses may not be used around `N', whereas they may around it in `READ (UNIT=(N))'. Further, if `N' is type `CHARACTER', the opposite is true--`READ (UNIT=(N))' is not valid, but `READ (FMT=(N))' is. Strictly speaking, if anything follows READ (N) in the statement, whether the first lexeme after the close parenthese is a comma could be used to disambiguate the two cases, without looking at the type of `N', because the comma is required for the `READ (FMT=(N))' interpretation and disallowed for the `READ (UNIT=(N))' interpretation. However, in practice, many Fortran compilers allow the comma for the `READ (UNIT=(N))' interpretation anyway (in that they generally allow a leading comma before an I/O list in an I/O statement), and much code takes advantage of this allowance. (This is quite a reasonable allowance, since the juxtaposition of a comma-separated list immediately after an I/O control-specification list, which is also comma-separated, without an intervening comma, looks sufficiently "wrong" to programmers that they can't resist the itch to insert the comma. `READ (I, J), K, L' simply looks cleaner than `READ (I, J) K, L'.) So, type-based disambiguation is needed unless strict adherence to the standard is always assumed, and we're not going to assume that.  File: g77.info, Node: TBD (Transforming), Prev: Gotchas (Transforming), Up: Overview of Translation Process TBD (Transforming) ------------------ Continue researching gotchas, designing the transformational process, and implementing it. Specific issues to resolve: * Just where should `INCLUDE' processing take place? Clearly before (or part of) statement identification (`sta.c'), since determining whether `I(J)=K' is a statement-function definition or an assignment statement requires knowing the context, which in turn requires having processed `INCLUDE' files. * Just where should (if it was implemented) `USE' processing take place? This gets into the whole issue of how `g77' should handle the concept of modules. I think GNAT already takes on this issue, but don't know more than that. Jim Giles has written extensively on `comp.lang.fortran' about his opinions on module handling, as have others. Jim's views should be taken into account. Actually, Richard M. Stallman (RMS) also has written up some guidelines for implementing such things, but I'm not sure where I read them. Perhaps the old list. If someone could dig references to these up and get them to me, that would be much appreciated! Even though modules are not on the short-term list for implementation, it'd be helpful to know *now* how to avoid making them harder to implement them *later*. * Should the `g77' command become just a script that invokes all the various preprocessing that might be needed, thus making it seem slower than necessary for legacy code that people are unwilling to convert, or should we provide a separate script for that, thus encouraging people to convert their code once and for all? At least, a separate script to behave as old `g77' did, perhaps named `g77old', might ease the transition, as might a corresponding one that converts source codes named `g77oldnew'. These scripts would take all the pertinent options `g77' used to take and run the appropriate filters, passing the results to `g77' or just making new sources out of them (in a subdirectory, leaving the user to do the dirty deed of moving or copying them over the old sources). * Do other Fortran compilers provide a prefix syntax to govern the treatment of backslashes in `CHARACTER' (or Hollerith) constants? Knowing what other compilers provide would help. * Is it okay to drop support for the `-fintrin-case-initcap', `-fmatch-case-initcap', `-fsymbol-case-initcap', and `-fcase-initcap' options? I've asked for input on this. Not having to support these makes it easier to write the new front end, and might also avoid complicated its design.  File: g77.info, Node: Philosophy of Code Generation, Next: Two-pass Design, Prev: Overview of Translation Process, Up: Front End Philosophy of Code Generation ============================= Don't poke the bear. The `g77' front end generates code via the `gcc' back end. The `gcc' back end (GBE) is a large, complex labyrinth of intricate code written in a combination of the C language and specialized languages internal to `gcc'. While the *code* that implements the GBE is written in a combination of languages, the GBE itself is, to the front end for a language like Fortran, best viewed as a *compiler* that compiles its own, unique, language. The GBE's "source", then, is written in this language, which consists primarily of a combination of calls to GBE functions and "tree" nodes (which are, themselves, created by calling GBE functions). So, the `g77' generates code by, in effect, translating the Fortran code it reads into a form "written" in the "language" of the `gcc' back end. This language will heretofore be referred to as "GBEL", for GNU Back End Language. GBEL is an evolving language, not fully specified in any published form as of this writing. It offers many facilities, but its "core" facilities are those that corresponding most directly to those needed to support `gcc' (compiling code written in GNU C). The `g77' Fortran Front End (FFE) is designed and implemented to navigate the currents and eddies of ongoing GBEL and `gcc' development while also delivering on the potential of an integrated FFE (as compared to using a converter like `f2c' and feeding the output into `gcc'). Goals of the FFE's code-generation strategy include: * High likelihood of generation of correct code, or, failing that, producing a fatal diagnostic or crashing. * Generation of highly optimized code, as directed by the user via GBE-specific (versus `g77'-specific) constructs, such as command-line options. * Fast overall (FFE plus GBE) compilation. * Preservation of source-level debugging information. The strategies historically, and currently, used by the FFE to achieve these goals include: * Use of GBEL constructs that most faithfully encapsulate the semantics of Fortran. * Avoidance of GBEL constructs that are so rarely used, or limited to use in specialized situations not related to Fortran, that their reliability and performance has not yet been established as sufficient for use by the FFE. * Flexible design, to readily accommodate changes to specific code-generation strategies, perhaps governed by command-line options. "Don't poke the bear" somewhat summarizes the above strategies. The GBE is the bear. The FFE is designed and implemented to avoid poking it in ways that are likely to just annoy it. The FFE usually either tackles it head-on, or avoids treating it in ways dissimilar to how the `gcc' front end treats it. For example, the FFE uses the native array facility in the back end instead of the lower-level pointer-arithmetic facility used by `gcc' when compiling `f2c' output). Theoretically, this presents more opportunities for optimization, faster compile times, and the production of more faithful debugging information. These benefits were not, however, immediately realized, mainly because `gcc' itself makes little or no use of the native array facility. Complex arithmetic is a case study of the evolution of this strategy. When originally implemented, the GBEL had just evolved its own native complex-arithmetic facility, so the FFE took advantage of that. When porting `g77' to 64-bit systems, it was discovered that the GBE didn't really implement its native complex-arithmetic facility properly. The short-term solution was to rewrite the FFE to instead use the lower-level facilities that'd be used by `gcc'-compiled code (assuming that code, itself, didn't use the native complex type provided, as an extension, by `gcc'), since these were known to work, and, in any case, if shown to not work, would likely be rapidly fixed (since they'd likely not work for vanilla C code in similar circumstances). However, the rewrite accommodated the original, native approach as well by offering a command-line option to select it over the emulated approach. This allowed users, and especially GBE maintainers, to try out fixes to complex-arithmetic support in the GBE while `g77' continued to default to compiling more code correctly, albeit producing (typically) slower executables. As of April 1999, it appeared that the last few bugs in the GBE's support of its native complex-arithmetic facility were worked out. The FFE was changed back to default to using that native facility, leaving emulation as an option. Other Fortran constructs--arrays, character strings, complex division, `COMMON' and `EQUIVALENCE' aggregates, and so on--involve issues similar to those pertaining to complex arithmetic. So, it is possible that the history of how the FFE handled complex arithmetic will be repeated, probably in modified form (and hopefully over shorter timeframes), for some of these other facilities.  File: g77.info, Node: Two-pass Design, Next: Challenges Posed, Prev: Philosophy of Code Generation, Up: Front End Two-pass Design =============== The FFE does not tell the GBE anything about a program unit until after the last statement in that unit has been parsed. (A program unit is a Fortran concept that corresponds, in the C world, mostly closely to functions definitions in ISO C. That is, a program unit in Fortran is like a top-level function in C. Nested functions, found among the extensions offered by GNU C, correspond roughly to Fortran's statement functions.) So, while parsing the code in a program unit, the FFE saves up all the information on statements, expressions, names, and so on, until it has seen the last statement. At that point, the FFE revisits the saved information (in what amounts to a second "pass" over the program unit) to perform the actual translation of the program unit into GBEL, ultimating in the generation of assembly code for it. Some lookahead is performed during this second pass, so the FFE could be viewed as a "two-plus-pass" design. * Menu: * Two-pass Code:: * Why Two Passes::  File: g77.info, Node: Two-pass Code, Next: Why Two Passes, Up: Two-pass Design Two-pass Code ------------- Most of the code that turns the first pass (parsing) into a second pass for code generation is in `egcs/gcc/f/std.c'. It has external functions, called mainly by siblings in `egcs/gcc/f/stc.c', that record the information on statements and expressions in the order they are seen in the source code. These functions save that information. It also has an external function that revisits that information, calling the siblings in `egcs/gcc/f/ste.c', which handles the actual code generation (by generating GBEL code, that is, by calling GBE routines to represent and specify expressions, statements, and so on).  File: g77.info, Node: Why Two Passes, Prev: Two-pass Code, Up: Two-pass Design Why Two Passes -------------- The need for two passes was not immediately evident during the design and implementation of the code in the FFE that was to produce GBEL. Only after a few kludges, to handle things like incorrectly-guessed `ASSIGN' label nature, had been implemented, did enough evidence pile up to make it clear that `std.c' had to be introduced to intercept, save, then revisit as part of a second pass, the digested contents of a program unit. Other such missteps have occurred during the evolution of the FFE, because of the different goals of the FFE and the GBE. Because the GBE's original, and still primary, goal was to directly support the GNU C language, the GBEL, and the GBE itself, requires more complexity on the part of most front ends than it requires of `gcc''s. For example, the GBEL offers an interface that permits the `gcc' front end to implement most, or all, of the language features it supports, without the front end having to make use of non-user-defined variables. (It's almost certainly the case that all of K&R C, and probably ANSI C as well, is handled by the `gcc' front end without declaring such variables.) The FFE, on the other hand, must resort to a variety of "tricks" to achieve its goals. Consider the following C code: int foo (int a, int b) { int c = 0; if ((c = bar (c)) == 0) goto done; quux (c << 1); done: return c; } Note what kinds of objects are declared, or defined, before their use, and before any actual code generation involving them would normally take place: * Return type of function * Entry point(s) of function * Dummy arguments * Variables * Initial values for variables Whereas, the following items can, and do, suddenly appear "out of the blue" in C: * Label references * Function references Not surprisingly, the GBE faithfully permits the latter set of items to be "discovered" partway through GBEL "programs", just as they are permitted to in C. Yet, the GBE has tended, at least in the past, to be reticent to fully support similar "late" discovery of items in the former set. This makes Fortran a poor fit for the "safe" subset of GBEL. Consider: FUNCTION X (A, ARRAY, ID1) CHARACTER*(*) A DOUBLE PRECISION X, Y, Z, TMP, EE, PI REAL ARRAY(ID1*ID2) COMMON ID2 EXTERNAL FRED ASSIGN 100 TO J CALL FOO (I) IF (I .EQ. 0) PRINT *, A(0) GOTO 200 ENTRY Y (Z) ASSIGN 101 TO J 200 PRINT *, A(1) READ *, TMP GOTO J 100 X = TMP * EE RETURN 101 Y = TMP * PI CALL FRED DATA EE, PI /2.71D0, 3.14D0/ END Here are some observations about the above code, which, while somewhat contrived, conforms to the FORTRAN 77 and Fortran 90 standards: * The return type of function `X' is not known until the `DOUBLE PRECISION' line has been parsed. * Whether `A' is a function or a variable is not known until the `PRINT *, A(0)' statement has been parsed. * The bounds of the array of argument `ARRAY' depend on a computation involving the subsequent argument `ID1' and the blank-common member `ID2'. * Whether `Y' and `Z' are local variables, additional function entry points, or dummy arguments to additional entry points is not known until the `ENTRY' statement is parsed. * Similarly, whether `TMP' is a local variable is not known until the `READ *, TMP' statement is parsed. * The initial values for `EE' and `PI' are not known until after the `DATA' statement is parsed. * Whether `FRED' is a function returning type `REAL' or a subroutine (which can be thought of as returning type `void' *or*, to support alternate returns in a simple way, type `int') is not known until the `CALL FRED' statement is parsed. * Whether `100' is a `FORMAT' label or the label of an executable statement is not known until the `X =' statement is parsed. (These two types of labels get *very* different treatment, especially when `ASSIGN''ed.) * That `J' is a local variable is not known until the first `ASSIGN' statement is parsed. (This happens *after* executable code has been seen.) Very few of these "discoveries" can be accommodated by the GBE as it has evolved over the years. The GBEL doesn't support several of them, and those it might appear to support don't always work properly, especially in combination with other GBEL and GBE features, as implemented in the GBE. (Had the GBE and its GBEL originally evolved to support `g77', the shoe would be on the other foot, so to speak--most, if not all, of the above would be directly supported by the GBEL, and a few C constructs would probably not, as they are in reality, be supported. Both this mythical, and today's real, GBE caters to its GBEL by, sometimes, scrambling around, cleaning up after itself--after discovering that assumptions it made earlier during code generation are incorrect.) So, the FFE handles these discrepancies--between the order in which it discovers facts about the code it is compiling, and the order in which the GBEL and GBE support such discoveries--by performing what amounts to two passes over each program unit. (A few ambiguities can remain at that point, such as whether, given `EXTERNAL BAZ' and no other reference to `BAZ' in the program unit, it is a subroutine, a function, or a block-data--which, in C-speak, governs its declared return type. Fortunately, these distinctions are easily finessed for the procedure, library, and object-file interfaces supported by `g77'.)  File: g77.info, Node: Challenges Posed, Next: Transforming Statements, Prev: Two-pass Design, Up: Front End Challenges Posed ================ Consider the following Fortran code, which uses various extensions (including some to Fortran 90): SUBROUTINE X(A) CHARACTER*(*) A COMPLEX CFUNC INTEGER*2 CLOCKS(200) INTEGER IFUNC CALL SYSTEM_CLOCK (CLOCKS (IFUNC (CFUNC ('('//A//')')))) The above poses the following challenges to any Fortran compiler that uses run-time interfaces, and a run-time library, roughly similar to those used by `g77': * Assuming the library routine that supports `SYSTEM_CLOCK' expects to set an `INTEGER*4' variable via its `COUNT' argument, the compiler must make available to it a temporary variable of that type. * Further, after the `SYSTEM_CLOCK' library routine returns, the compiler must ensure that the temporary variable it wrote is copied into the appropriate element of the `CLOCKS' array. (This assumes the compiler doesn't just reject the code, which it should if it is compiling under some kind of a "strict" option.) * To determine the correct index into the `CLOCKS' array, (putting aside the fact that the index, in this particular case, need not be computed until after the `SYSTEM_CLOCK' library routine returns), the compiler must ensure that the `IFUNC' function is called. That requires evaluating its argument, which requires, for `g77' (assuming `-ff2c' is in force), reserving a temporary variable of type `COMPLEX' for use as a repository for the return value being computed by `CFUNC'. * Before invoking `CFUNC', is argument must be evaluated, which requires allocating, at run time, a temporary large enough to hold the result of the concatenation, as well as actually performing the concatenation. * The large temporary needed during invocation of `CFUNC' should, ideally, be deallocated (or, at least, left to the GBE to dispose of, as it sees fit) as soon as `CFUNC' returns, which means before `IFUNC' is called (as it might need a lot of dynamically allocated memory). `g77' currently doesn't support all of the above, but, so that it might someday, it has evolved to handle at least some of the above requirements. Meeting the above requirements is made more challenging by conforming to the requirements of the GBEL/GBE combination.  File: g77.info, Node: Transforming Statements, Next: Transforming Expressions, Prev: Challenges Posed, Up: Front End Transforming Statements ======================= Most Fortran statements are given their own block, and, for temporary variables they might need, their own scope. (A block is what distinguishes `{ foo (); }' from just `foo ();' in C. A scope is included with every such block, providing a distinct name space for local variables.) Label definitions for the statement precede this block, so `10 PRINT *, I' is handled more like `fl10: { ... }' than `{ fl10: ... }' (where `fl10' is just a notation meaning "Fortran Label 10" for the purposes of this document). * Menu: * Statements Needing Temporaries:: * Transforming DO WHILE:: * Transforming Iterative DO:: * Transforming Block IF:: * Transforming SELECT CASE::  File: g77.info, Node: Statements Needing Temporaries, Next: Transforming DO WHILE, Up: Transforming Statements Statements Needing Temporaries ------------------------------ Any temporaries needed during, but not beyond, execution of a Fortran statement, are made local to the scope of that statement's block. This allows the GBE to share storage for these temporaries among the various statements without the FFE having to manage that itself. (The GBE could, of course, decide to optimize management of these temporaries. For example, it could, theoretically, schedule some of the computations involving these temporaries to occur in parallel. More practically, it might leave the storage for some temporaries "live" beyond their scopes, to reduce the number of manipulations of the stack pointer at run time.) Temporaries needed across distinct statement boundaries usually are associated with Fortran blocks (such as `DO'/`END DO'). (Also, there might be temporaries not associated with blocks at all--these would be in the scope of the entire program unit.) Each Fortran block *should* get its own block/scope in the GBE. This is best, because it allows temporaries to be more naturally handled. However, it might pose problems when handling labels (in particular, when they're the targets of `GOTO's outside the Fortran block), and generally just hassling with replicating parts of the `gcc' front end (because the FFE needs to support an arbitrary number of nested back-end blocks if each Fortran block gets one). So, there might still be a need for top-level temporaries, whose "owning" scope is that of the containing procedure. Also, there seems to be problems declaring new variables after generating code (within a block) in the back end, leading to, e.g., `label not defined before binding contour' or similar messages, when compiling with `-fstack-check' or when compiling for certain targets. Because of that, and because sometimes these temporaries are not discovered until in the middle of of generating code for an expression statement (as in the case of the optimization for `X**I'), it seems best to always pre-scan all the expressions that'll be expanded for a block before generating any of the code for that block. This pre-scan then handles discovering and declaring, to the back end, the temporaries needed for that block. It's also important to treat distinct items in an I/O list as distinct statements deserving their own blocks. That's because there's a requirement that each I/O item be fully processed before the next one, which matters in cases like `READ (*,*), I, A(I)'--the element of `A' read in the second item *must* be determined from the value of `I' read in the first item.  File: g77.info, Node: Transforming DO WHILE, Next: Transforming Iterative DO, Prev: Statements Needing Temporaries, Up: Transforming Statements Transforming DO WHILE --------------------- `DO WHILE(expr)' *must* be implemented so that temporaries needed to evaluate `expr' are generated just for the test, each time. Consider how `DO WHILE (A//B .NE. 'END'); ...; END DO' is transformed: for (;;) { int temp0; { char temp1[large]; libg77_catenate (temp1, a, b); temp0 = libg77_ne (temp1, 'END'); } if (! temp0) break; ... } In this case, it seems like a time/space tradeoff between allocating and deallocating `temp1' for each iteration and allocating it just once for the entire loop. However, if `temp1' is allocated just once for the entire loop, it could be the wrong size for subsequent iterations of that loop in cases like `DO WHILE (A(I:J)//B .NE. 'END')', because the body of the loop might modify `I' or `J'. So, the above implementation is used, though a more optimal one can be used in specific circumstances.  File: g77.info, Node: Transforming Iterative DO, Next: Transforming Block IF, Prev: Transforming DO WHILE, Up: Transforming Statements Transforming Iterative DO ------------------------- An iterative `DO' loop (one that specifies an iteration variable) is required by the Fortran standards to be implemented as though an iteration count is computed before entering the loop body, and that iteration count used to determine the number of times the loop body is to be performed (assuming the loop isn't cut short via `GOTO' or `EXIT'). The FFE handles this by allocating a temporary variable to contain the computed number of iterations. Since this variable must be in a scope that includes the entire loop, a GBEL block is created for that loop, and the variable declared as belonging to the scope of that block.  File: g77.info, Node: Transforming Block IF, Next: Transforming SELECT CASE, Prev: Transforming Iterative DO, Up: Transforming Statements Transforming Block IF --------------------- Consider: SUBROUTINE X(A,B,C) CHARACTER*(*) A, B, C LOGICAL LFUNC IF (LFUNC (A//B)) THEN CALL SUBR1 ELSE IF (LFUNC (A//C)) THEN CALL SUBR2 ELSE CALL SUBR3 END The arguments to the two calls to `LFUNC' require dynamic allocation (at run time), but are not required during execution of the `CALL' statements. So, the scopes of those temporaries must be within blocks inside the block corresponding to the Fortran `IF' block. This cannot be represented "naturally" in vanilla C, nor in GBEL. The `if', `elseif', `else', and `endif' constructs provided by both languages must, for a given `if' block, share the same C/GBE block. Therefore, any temporaries needed during evaluation of `expr' while executing `ELSE IF(expr)' must either have been predeclared at the top of the corresponding `IF' block, or declared within a new block for that `ELSE IF'--a block that, since it cannot contain the `else' or `else if' itself (due to the above requirement), actually implements the rest of the `IF' block's `ELSE IF' and `ELSE' statements within an inner block. The FFE takes the latter approach.  File: g77.info, Node: Transforming SELECT CASE, Prev: Transforming Block IF, Up: Transforming Statements Transforming SELECT CASE ------------------------ `SELECT CASE' poses a few interesting problems for code generation, if efficiency and frugal stack management are important. Consider `SELECT CASE (I('PREFIX'//A))', where `A' is `CHARACTER*(*)'. In a case like this--basically, in any case where largish temporaries are needed to evaluate the expression--those temporaries should not be "live" during execution of any of the `CASE' blocks. So, evaluation of the expression is best done within its own block, which in turn is within the `SELECT CASE' block itself (which contains the code for the CASE blocks as well, though each within their own block). Otherwise, we'd have the rough equivalent of this pseudo-code: { char temp[large]; libg77_catenate (temp, 'prefix', a); switch (i (temp)) { case 0: ... } } And that would leave temp[large] in scope during the CASE blocks (although a clever back end *could* see that it isn't referenced in them, and thus free that temp before executing the blocks). So this approach is used instead: { int temp0; { char temp1[large]; libg77_catenate (temp1, 'prefix', a); temp0 = i (temp1); } switch (temp0) { case 0: ... } } Note how `temp1' goes out of scope before starting the switch, thus making it easy for a back end to free it. The problem *that* solution has, however, is with `SELECT CASE('prefix'//A)' (which is currently not supported). Unless the GBEL is extended to support arbitrarily long character strings in its `case' facility, the FFE has to implement `SELECT CASE' on `CHARACTER' (probably excepting `CHARACTER*1') using a cascade of `if', `elseif', `else', and `endif' constructs in GBEL. To prevent the (potentially large) temporary, needed to hold the selected expression itself (`'prefix'//A'), from being in scope during execution of the `CASE' blocks, two approaches are available: * Pre-evaluate all the `CASE' tests, producing an integer ordinal that is used, a la `temp0' in the earlier example, as if `SELECT CASE(temp0)' had been written. Each corresponding `CASE' is replaced with `CASE(I)', where I is the ordinal for that case, determined while, or before, generating the cascade of `if'-related constructs to cope with `CHARACTER' selection. * Make `temp0' above just large enough to hold the longest `CASE' string that'll actually be compared against the expression (in this case, `'prefix'//A'). Since that length must be constant (because `CASE' expressions are all constant), it won't be so large, and, further, `temp1' need not be dynamically allocated, since normal `CHARACTER' assignment can be used into the fixed-length `temp0'. Both of these solutions require `SELECT CASE' implementation to be changed so all the corresponding `CASE' statements are seen during the actual code generation for `SELECT CASE'.  File: g77.info, Node: Transforming Expressions, Next: Internal Naming Conventions, Prev: Transforming Statements, Up: Front End Transforming Expressions ======================== The interactions between statements, expressions, and subexpressions at program run time can be viewed as: ACTION(EXPR) Here, ACTION is the series of steps performed to effect the statement, and EXPR is the expression whose value is used by ACTION. Expanding the above shows a typical order of events at run time: Evaluate EXPR Perform ACTION, using result of evaluation of EXPR Clean up after evaluating EXPR So, if evaluating EXPR requires allocating memory, that memory can be freed before performing ACTION only if it is not needed to hold the result of evaluating EXPR. Otherwise, it must be freed no sooner than after ACTION has been performed. The above are recursive definitions, in the sense that they apply to subexpressions of EXPR. That is, evaluating EXPR involves evaluating all of its subexpressions, performing the ACTION that computes the result value of EXPR, then cleaning up after evaluating those subexpressions. The recursive nature of this evaluation is implemented via recursive-descent transformation of the top-level statements, their expressions, *their* subexpressions, and so on. However, that recursive-descent transformation is, due to the nature of the GBEL, focused primarily on generating a *single* stream of code to be executed at run time. Yet, from the above, it's clear that multiple streams of code must effectively be simultaneously generated during the recursive-descent analysis of statements. The primary stream implements the primary ACTION items, while at least two other streams implement the evaluation and clean-up items. Requirements imposed by expressions include: * Whether the caller needs to have a temporary ready to hold the value of the expression. * Other stuff???  File: g77.info, Node: Internal Naming Conventions, Prev: Transforming Expressions, Up: Front End Internal Naming Conventions =========================== Names exported by FFE modules have the following (regular-expression) forms. Note that all names beginning `ffeMOD' or `FFEMOD', where MOD is lowercase or uppercase alphanumerics, respectively, are exported by the module `ffeMOD', with the source code doing the exporting in `MOD.h'. (Usually, the source code for the implementation is in `MOD.c'.) Identifiers that don't fit the following forms are not considered exported, even if they are according to the C language. (For example, they might be made available to other modules solely for use within expansions of exported macros, not for use within any source code in those other modules.) `ffeMOD' The single typedef exported by the module. `FFEUMOD_[A-Z][A-Z0-9_]*' (Where UMOD is the uppercase for of MOD.) A `#define' or `enum' constant of the type `ffeMOD'. `ffeMOD[A-Z][A-Z][a-z0-9]*' A typedef exported by the module. The portion of the identifier after `ffeMOD' is referred to as `ctype', a capitalized (mixed-case) form of `type'. `FFEUMOD_TYPE[A-Z][A-Z0-9_]*[A-Z0-9]?' (Where UMOD is the uppercase for of MOD.) A `#define' or `enum' constant of the type `ffeMODTYPE', where TYPE is the lowercase form of CTYPE in an exported typedef. `ffeMOD_VALUE' A function that does or returns something, as described by VALUE (see below). `ffeMOD_VALUE_INPUT' A function that does or returns something based primarily on the thing described by INPUT (see below). Below are names used for VALUE and INPUT, along with their definitions. `col' A column number within a line (first column is number 1). `file' An encapsulation of a file's name. `find' Looks up an instance of some type that matches specified criteria, and returns that, even if it has to create a new instance or crash trying to find it (as appropriate). `initialize' Initializes, usually a module. No type. `int' A generic integer of type `int'. `is' A generic integer that contains a true (non-zero) or false (zero) value. `len' A generic integer that contains the length of something. `line' A line number within a source file, or a global line number. `lookup' Looks up an instance of some type that matches specified criteria, and returns that, or returns nil. `name' A `text' that points to a name of something. `new' Makes a new instance of the indicated type. Might return an existing one if appropriate--if so, similar to `find' without crashing. `pt' Pointer to a particular character (line, column pairs) in the input file (source code being compiled). `run' Performs some herculean task. No type. `terminate' Terminates, usually a module. No type. `text' A `char *' that points to generic text.  File: g77.info, Node: Diagnostics, Next: Index, Prev: Front End, Up: Top Diagnostics *********** Some diagnostics produced by `g77' require sufficient explanation that the explanations are given below, and the diagnostics themselves identify the appropriate explanation. Identification uses the GNU Info format--specifically, the `info' command that displays the explanation is given within square brackets in the diagnostic. For example: foo.f:5: Invalid statement [info -f g77 M FOOEY] More details about the above diagnostic is found in the `g77' Info documentation, menu item `M', submenu item `FOOEY', which is displayed by typing the UNIX command `info -f g77 M FOOEY'. Other Info readers, such as EMACS, may be just as easily used to display the pertinent node. In the above example, `g77' is the Info document name, `M' is the top-level menu item to select, and, in that node (named `Diagnostics', the name of this chapter, which is the very text you're reading now), `FOOEY' is the menu item to select. * Menu: * CMPAMBIG:: Ambiguous use of intrinsic. * EXPIMP:: Intrinsic used explicitly and implicitly. * INTGLOB:: Intrinsic also used as name of global. * LEX:: Various lexer messages * GLOBALS:: Disagreements about globals. * LINKFAIL:: When linking `f771' fails. * Y2KBAD:: Use of non-Y2K-compliant intrinsic.  File: g77.info, Node: CMPAMBIG, Next: EXPIMP, Up: Diagnostics `CMPAMBIG' ========== Ambiguous use of intrinsic INTRINSIC ... The type of the argument to the invocation of the INTRINSIC intrinsic is a `COMPLEX' type other than `COMPLEX(KIND=1)'. Typically, it is `COMPLEX(KIND=2)', also known as `DOUBLE COMPLEX'. The interpretation of this invocation depends on the particular dialect of Fortran for which the code was written. Some dialects convert the real part of the argument to `REAL(KIND=1)', thus losing precision; other dialects, and Fortran 90, do no such conversion. So, GNU Fortran rejects such invocations except under certain circumstances, to avoid making an incorrect assumption that results in generating the wrong code. To determine the dialect of the program unit, perhaps even whether that particular invocation is properly coded, determine how the result of the intrinsic is used. The result of INTRINSIC is expected (by the original programmer) to be `REAL(KIND=1)' (the non-Fortran-90 interpretation) if: * It is passed as an argument to a procedure that explicitly or implicitly declares that argument `REAL(KIND=1)'. For example, a procedure with no `DOUBLE PRECISION' or `IMPLICIT DOUBLE PRECISION' statement specifying the dummy argument corresponding to an actual argument of `REAL(Z)', where `Z' is declared `DOUBLE COMPLEX', strongly suggests that the programmer expected `REAL(Z)' to return `REAL(KIND=1)' instead of `REAL(KIND=2)'. * It is used in a context that would otherwise not include any `REAL(KIND=2)' but where treating the INTRINSIC invocation as `REAL(KIND=2)' would result in unnecessary promotions and (typically) more expensive operations on the wider type. For example: DOUBLE COMPLEX Z ... R(1) = T * REAL(Z) The above example suggests the programmer expected the real part of `Z' to be converted to `REAL(KIND=1)' before being multiplied by `T' (presumed, along with `R' above, to be type `REAL(KIND=1)'). Otherwise, the conversion would have to be delayed until after the multiplication, requiring not only an extra conversion (of `T' to `REAL(KIND=2)'), but a (typically) more expensive multiplication (a double-precision multiplication instead of a single-precision one). The result of INTRINSIC is expected (by the original programmer) to be `REAL(KIND=2)' (the Fortran 90 interpretation) if: * It is passed as an argument to a procedure that explicitly or implicitly declares that argument `REAL(KIND=2)'. For example, a procedure specifying a `DOUBLE PRECISION' dummy argument corresponding to an actual argument of `REAL(Z)', where `Z' is declared `DOUBLE COMPLEX', strongly suggests that the programmer expected `REAL(Z)' to return `REAL(KIND=2)' instead of `REAL(KIND=1)'. * It is used in an expression context that includes other `REAL(KIND=2)' operands, or is assigned to a `REAL(KIND=2)' variable or array element. For example: DOUBLE COMPLEX Z DOUBLE PRECISION R, T ... R(1) = T * REAL(Z) The above example suggests the programmer expected the real part of `Z' to *not* be converted to `REAL(KIND=1)' by the `REAL()' intrinsic. Otherwise, the conversion would have to be immediately followed by a conversion back to `REAL(KIND=2)', losing the original, full precision of the real part of `Z', before being multiplied by `T'. Once you have determined whether a particular invocation of INTRINSIC expects the Fortran 90 interpretation, you can: * Change it to `DBLE(EXPR)' (if INTRINSIC is `REAL') or `DIMAG(EXPR)' (if INTRINSIC is `AIMAG') if it expected the Fortran 90 interpretation. This assumes EXPR is `COMPLEX(KIND=2)'--if it is some other type, such as `COMPLEX*32', you should use the appropriate intrinsic, such as the one to convert to `REAL*16' (perhaps `DBLEQ()' in place of `DBLE()', and `QIMAG()' in place of `DIMAG()'). * Change it to `REAL(INTRINSIC(EXPR))', otherwise. This converts to `REAL(KIND=1)' in all working Fortran compilers. If you don't want to change the code, and you are certain that all ambiguous invocations of INTRINSIC in the source file have the same expectation regarding interpretation, you can: * Compile with the `g77' option `-ff90', to enable the Fortran 90 interpretation. * Compile with the `g77' options `-fno-f90 -fugly-complex', to enable the non-Fortran-90 interpretations. *Note REAL() and AIMAG() of Complex::, for more information on this issue. Note: If the above suggestions don't produce enough evidence as to whether a particular program expects the Fortran 90 interpretation of this ambiguous invocation of INTRINSIC, there is one more thing you can try. If you have access to most or all the compilers used on the program to create successfully tested and deployed executables, read the documentation for, and *also* test out, each compiler to determine how it treats the INTRINSIC intrinsic in this case. (If all the compilers don't agree on an interpretation, there might be lurking bugs in the deployed versions of the program.) The following sample program might help: PROGRAM JCB003 C C Written by James Craig Burley 1997-02-23. C C Determine how compilers handle non-standard REAL C and AIMAG on DOUBLE COMPLEX operands. C DOUBLE COMPLEX Z REAL R Z = (3.3D0, 4.4D0) R = Z CALL DUMDUM(Z, R) R = REAL(Z) - R IF (R .NE. 0.) PRINT *, 'REAL() is Fortran 90' IF (R .EQ. 0.) PRINT *, 'REAL() is not Fortran 90' R = 4.4D0 CALL DUMDUM(Z, R) R = AIMAG(Z) - R IF (R .NE. 0.) PRINT *, 'AIMAG() is Fortran 90' IF (R .EQ. 0.) PRINT *, 'AIMAG() is not Fortran 90' END C C Just to make sure compiler doesn't use naive flow C analysis to optimize away careful work above, C which might invalidate results.... C SUBROUTINE DUMDUM(Z, R) DOUBLE COMPLEX Z REAL R END If the above program prints contradictory results on a particular compiler, run away!