This file lists recent changes to the GNU Fortran compiler. Copyright (C) 1995, 1996 Free Software Foundation, Inc. You may copy, distribute, and modify it freely as long as you preserve this copyright notice and permission notice. News About GNU Fortran ********************** In 0.5.18: ========== * Add some rudimentary support for `INTEGER*1', `INTEGER*2', `INTEGER*8', and their `LOGICAL' equivalents. (This support works on most, maybe all, `gcc' targets.) Thanks to Scott Snyder (`snyder@d0sgif.fnal.gov') for providing the patch for this! Among the missing elements from the support for these features are full intrinsic support and constants. * Add some rudimentary support for the `BYTE' and `WORD' type-declaration statements. `BYTE' corresponds to `INTEGER*1', while `WORD' corresponds to `INTEGER*2'. Thanks to Scott Snyder (`snyder@d0sgif.fnal.gov') for providing the patch for this! * The compiler code handling intrinsics has been largely rewritten to accommodate the new types. No new intrinsics or arguments for existing intrinsics have been added, so there is, at this point, no intrinsic to convert to `INTEGER*8', for example. * Support automatic arrays in procedures. * Reduce space/time requirements for handling large *sparsely* initialized aggregate arrays. This improvement applies to only a subset of the general problem to be addressed in 0.6. * Treat initial values of zero as if they weren't specified (in DATA and type-declaration statements). The initial values will be set to zero anyway, but the amount of compile time processing them will be reduced, in some cases significantly (though, again, this is only a subset of the general problem to be addressed in 0.6). A new option, `-fzeros', is introduced to enable the traditional treatment of zeros as any other value. * With `-ff90' in force, `g77' incorrectly interpreted `REAL(Z)' as returning a `REAL' result, instead of as a `DOUBLE PRECISION' result. (Here, `Z' is `DOUBLE COMPLEX'.) With `-fno-f90' in force, the interpretation remains unchanged, since this appears to be how at least some F77 code using the `DOUBLE COMPLEX' extension expected it to work. Essentially, `REAL(Z)' in F90 is the same as `DBLE(Z)', while in extended F77, it appears to be the same as `REAL(REAL(Z))'. * An expression involving exponentiation, where both operands were type `INTEGER' and the right-hand operand was negative, was erroneously evaluated. * Fix bugs involving `DATA' implied-`DO' constructs (these involved an errant diagnostic and a crash, both on good code, one involving subsequent statement-function definition). * Close `INCLUDE' files after processing them, so compiling source files with lots of `INCLUDE' statements does not result in being unable to open `INCLUDE' files after all the available file descriptors are used up. * Speed up compiling, especially of larger programs, and perhaps slightly reduce memory utilization while compiling (this is *not* the improvement planned for 0.6 involving large aggregate areas)--these improvements result from simply turning off some low-level code to do self-checking that hasn't been triggered in a long time. * Introduce three new options that implement optimizations in the `gcc' back end (GBE). These options are `-fmove-all-movables', `-freduce-all-givs', and `-frerun-loop-opt', which are enabled, by default, for Fortran compilations. These optimizations are intended to help toon Fortran programs. * Patch the GBE to do a better job optimizing certain kinds of references to array elements. * Due to patches to the GBE, the version number of `gcc' also is patched to make it easier to manage installations, especially useful if it turns out a `g77' change to the GBE has a bug. The `g77'-modified version number is the `gcc' version number with the string `.f.N' appended, where `f' identifies the version as enhanced for Fortran, and N is `1' for the first Fortran patch for that version of `gcc', `2' for the second, and so on. So, this introduces version `2.7.2.f.1' of `gcc'. * Make several improvements and fixes to diagnostics, including the removal of two that were inappropriate or inadequate. * Warning about two successive arithmetic operators, produced by `-Wsurprising', now produced *only* when both operators are, indeed, arithmetic (not relational/boolean). * `-Wsurprising' now warns about the remaining cases of using non-integral variables for implied-`DO' loops, instead of these being rejected unless `-fpedantic' or `-fugly' specified. * Allow `SAVE' of a local variable or array, even after it has been given an initial value via `DATA', for example. * Introduce an Info version of `g77' documentation, which supercedes `gcc/f/CREDITS', `gcc/f/DOC', and `gcc/f/PROJECTS'. These files will be removed in a future release. The files `gcc/f/BUGS', `gcc/f/INSTALL', and `gcc/f/NEWS' now are automatically built from the texinfo source when distributions are made. This effort was inspired by a first pass at translating `g77-0.5.16/f/DOC' that was contributed to Craig by David Ronis (`ronis@onsager.chem.mcgill.ca'). * New `-fno-second-underscore' option to specify that, when `-funderscoring' is in effect, a second underscore is not to be appended to Fortran names already containing an underscore. * Change the way iterative `DO' loops work to follow the F90 standard. In particular, calculation of the iteration count is still done by converting the start, end, and increment parameters to the type of the `DO' variable, but the result of the calculation is always converted to the default `INTEGER' type. (This should have no effect on existing code compiled by `g77', but code written to assume that use of a *wider* type for the `DO' variable will result in an iteration count being fully calculated using that wider type (wider than default `INTEGER') must be rewritten.) * Upgrade to `libf2c' as of 1996-03-23, and fix up some of the build procedures. Note that the email addresses related to `f2c' have changed--the distribution site now is named `netlib.bell-labs.com', and the maintainer's new address is `dmg@bell-labs.com'. In 0.5.17: ========== * *Fix serious bug* in `g77 -v' command that can cause removal of a system's `/dev/null' special file if run by user `root'. *All users* of version 0.5.16 should ensure that they have not removed `/dev/null' or replaced it with an ordinary file (e.g. by comparing the output of `ls -l /dev/null' with `ls -l /dev/zero'. If the output isn't basically the same, contact your system administrator about restoring `/dev/null' to its proper status). This bug is particularly insidious because removing `/dev/null' as a special file can go undetected for quite a while, aside from various applications and programs exhibiting sudden, strange behaviors. I sincerely apologize for not realizing the implications of the fact that when `g77 -v' runs the `ld' command with `-o /dev/null' that `ld' tries to *remove* the executable it is supposed to build (especially if it reports unresolved references, which it should in this case)! * Fix crash on `CHARACTER*(*) FOO' in a main or block data program unit. * Fix crash that can occur when diagnostics given outside of any program unit (such as when input file contains `@foo'). * Fix crashes, infinite loops (hangs), and such involving diagnosed code. * Fix `ASSIGN''ed variables so they can be `SAVE''d or dummy arguments, and issue clearer error message in cases where target of `ASSIGN' or `ASSIGN'ed `GOTO'/`FORMAT' is too small (which should never happen). * Make `libf2c' build procedures work on more systems again by eliminating unnecessary invocations of `ld -r -x' and `mv'. * Fix omission of `-funix-intrinsics-...' options in list of permitted options to compiler. * Fix failure to always diagnose missing type declaration for `IMPLICIT NONE'. * Fix compile-time performance problem (which could sometimes crash the compiler, cause a hang, or whatever, due to a bug in the back end) involving exponentiation with a large `INTEGER' constant for the right-hand operator (e.g. `I**32767'). * Fix build procedures so cross-compiling `g77' (the `fini' utility in particular) is properly built using the host compiler. * Add new `-Wsurprising' option to warn about constructs that are interpreted by the Fortran standard (and `g77') in ways that are surprising to many programmers. * Add `ERF()' and `ERFC()' as generic intrinsics mapping to existing `ERF'/`DERF' and `ERFC'/`DERFC' specific intrinsics. *Note:* You should specify `INTRINSIC ERF,ERFC' in any code where you might use these as generic intrinsics, to improve likelihood of diagnostics (instead of subtle run-time bugs) when using a compiler that doesn't support these as intrinsics (e.g. `f2c'). * Remove from `-fno-pedantic' the diagnostic about `DO' with non-`INTEGER' index variable; issue that under `-Wsurprising' instead. * Clarify some diagnostics that say things like "ignored" when that's misleading. * Clarify diagnostic on use of `.EQ.'/`.NE.' on `LOGICAL' operands. * Minor improvements to code generation for various operations on `LOGICAL' operands. * Minor improvement to code generation for some `DO' loops on some machines. * Support `gcc' version 2.7.1. * Upgrade to `libf2c' as of 1995-11-15. In 0.5.16: ========== * Fix a code-generation bug involving complicated `EQUIVALENCE' statements not involving `COMMON' * Fix code-generation bugs involving invoking "gratis" library procedures in `libf2c' from code compiled with `-fno-f2c' by making these procedures known to `g77' as intrinsics (not affected by -fno-f2c). This is known to fix code invoking `ERF()', `ERFC()', `DERF()', and `DERFC()'. * Update `libf2c' to include netlib patches through 1995-08-16, and `#define' `WANT_LEAD_0' to 1 to make `g77'-compiled code more consistent with other Fortran implementations by outputting leading zeros in formatted and list-directed output. * Fix a code-generation bug involving adjustable dummy arrays with high bounds whose primaries are changed during procedure execution, and which might well improve code-generation performance for such arrays compared to `f2c' plus `gcc' (but apparently only when using `gcc-2.7.0' or later). * Fix a code-generation bug involving invocation of `COMPLEX' and `DOUBLE COMPLEX' `FUNCTION's and doing `COMPLEX' and `DOUBLE COMPLEX' divides, when the result of the invocation or divide is assigned directly to a variable that overlaps one or more of the arguments to the invocation or divide. * Fix crash by not generating new optimal code for `X**I' if `I' is nonconstant and the expression is used to dimension a dummy array, since the `gcc' back end does not support the necessary mechanics (and the `gcc' front end rejects the equivalent construct, as it turns out). * Fix crash on expressions like `COMPLEX**INTEGER'. * Fix crash on expressions like `(1D0,2D0)**2', i.e. raising a `DOUBLE COMPLEX' constant to an `INTEGER' constant power. * Fix crashes and such involving diagnosed code. * Diagnose, instead of crashing on, statement function definitions having duplicate dummy argument names. * Fix bug causing rejection of good code involving statement function definitions. * Fix bug resulting in debugger not knowing size of local equivalence area when any member of area has initial value (via `DATA', for example). * Fix installation bug that prevented installation of `g77' driver. Provide for easy selection of whether to install copy of `g77' as `f77' to replace the broken code. * Fix `gcc' driver (affects `g77' thereby) to not gratuitously invoke the `f771' program (e.g. when `-E' is specified). * Fix diagnostic to point to correct source line when it immediately follows an `INCLUDE' statement. * Support more compiler options in `gcc'/`g77' when compiling Fortran files. These options include `-p', `-pg', `-aux-info', `-P', correct setting of version-number macros for preprocessing, full recognition of `-O0', and automatic insertion of configuration-specific linker specs. * Add new intrinsics that interface to existing routines in `libf2c': `ABORT', `DERF', `DERFC', `ERF', `ERFC', `EXIT', `FLUSH', `GETARG', `GETENV', `IARGC', `SIGNAL', and `SYSTEM'. Note that `ABORT', `EXIT', `FLUSH', `SIGNAL', and `SYSTEM' are intrinsic subroutines, not functions (since they have side effects), so to get the return values from `SIGNAL' and `SYSTEM', append a final argument specifying an `INTEGER' variable or array element (e.g. `CALL SYSTEM('rm foo',ISTAT)'). * Add new intrinsic group named `unix' to contain the new intrinsics, and by default enable this new group. * Move `LOC()' intrinsic out of the `vxt' group to the new `unix' group. * Improve `g77' so that `g77 -v' by itself (or with certain other options, including `-B', `-b', `-i', `-nostdlib', and `-V') reports lots more useful version info, and so that long-form options `gcc' accepts are understood by `g77' as well (even in truncated, unambiguous forms). * Add new `g77' option `--driver=name' to specify driver when default, `gcc', isn't appropriate. * Add support for `#' directives (as output by the preprocessor) in the compiler, and enable generation of those directives by the preprocessor (when compiling `.F' files) so diagnostics and debugging info are more useful to users of the preprocessor. * Produce better diagnostics, more like `gcc', with info such as `In function `foo':' and `In file included from...:'. * Support `gcc''s `-fident' and `-fno-ident' options. * When `-Wunused' in effect, don't warn about local variables used as statement-function dummy arguments or `DATA' implied-`DO' iteration variables, even though, strictly speaking, these are not uses of the variables themselves. * When `-W -Wunused' in effect, don't warn about unused dummy arguments at all, since there's no way to turn this off for individual cases (`g77' might someday start warning about these)--applies to `gcc' versions 2.7.0 and later, since earlier versions didn't warn about unused dummy arguments. * New option `-fno-underscoring' that inhibits transformation of names (by appending one or two underscores) so users may experiment with implications of such an environment. * Minor improvement to `gcc/f/info' module to make it easier to build `g77' using the native (non-`gcc') compiler on certain machines (but definitely not all machines nor all non-`gcc' compilers). Please do not report bugs showing problems compilers have with macros defined in `gcc/f/target.h' and used in places like `gcc/f/expr.c'. * Add warning to be printed for each invocation of the compiler if the target machine `INTEGER', REAL, or `LOGICAL' size is not 32 bits, since `g77' is known to not work well for such cases (to be fixed in Version 0.6--*note Actual Bugs We Haven't Fixed Yet: Actual Bugs.). * Lots of new documentation (though work is still needed to put it into canonical GNU format). * Build `libf2c' with `-g0', not `-g2', in effect (by default), to produce smaller library without lots of debugging clutter. In 0.5.15: ========== * Fix bad code generation involving `X**I' and temporary, internal variables generated by `g77' and the back end (such as for `DO' loops). * Fix crash given `CHARACTER A;DATA A/.TRUE./'. * Replace crash with diagnostic given `CHARACTER A;DATA A/1.0/'. * Fix crash or other erratic behavior when null character constant (`''') is encountered. * Fix crash or other erratic behavior involving diagnosed code. * Fix code generation for external functions returning type `REAL' when the `-ff2c' option is in force (which it is by default) so that `f2c' compatibility is indeed provided. * Disallow `COMMON I(10)' if `I' has previously been specified with an array declarator. * New `-ffixed-line-length-N' option, where N is the maximum length of a typical fixed-form line, defaulting to 72 columns, such that characters beyond column N are ignored, or N is `none', meaning no characters are ignored. does not affect lines with `&' in column 1, which are always processed as if `-ffixed-line-length-none' was in effect. * No longer generate better code for some kinds of array references, as `gcc' back end is to be fixed to do this even better, and it turned out to slow down some code in some cases after all. * In `COMMON' and `EQUIVALENCE' areas with any members given initial values (e.g. via `DATA'), uninitialized members now always initialized to binary zeros (though this is not required by the standard, and might not be done in future versions of `g77'). Previously, in some `COMMON'/`EQUIVALENCE' areas (essentially those with members of more than one type), the uninitialized members were initialized to spaces, to cater to `CHARACTER' types, but it seems no existing code expects that, while much existing code expects binary zeros. In 0.5.14: ========== * Don't emit bad code when low bound of adjustable array is nonconstant and thus might vary as an expression at run time. * Emit correct code for calculation of number of trips in `DO' loops for cases where the loop should not execute at all. (This bug affected cases where the difference between the begin and end values was less than the step count, though probably not for floating-point cases.) * Fix crash when extra parentheses surround item in `DATA' implied-`DO' list. * Fix crash over minor internal inconsistencies in handling diagnostics, just substitute dummy strings where necessary. * Fix crash on some systems when compiling call to `MVBITS()' intrinsic. * Fix crash on array assignment `TYPEDDD(...)=...', where DDD is a string of one or more digits. * Fix crash on `DCMPLX()' with a single `INTEGER' argument. * Fix various crashes involving code with diagnosed errors. * Support `-I' option for `INCLUDE' statement, plus `gcc''s `header.gcc' facility for handling systems like MS-DOS. * Allow `INCLUDE' statement to be continued across multiple lines, even allow it to coexist with other statements on the same line. * Incorporate Bellcore fixes to `libf2c' through 1995-03-15--this fixes a bug involving infinite loops reading EOF with empty list-directed I/O list. * Remove all the `g77'-specific auto-configuration scripts, code, and so on, except for temporary substitutes for bsearch() and strtoul(), as too many configure/build problems were reported in these areas. People will have to fix their systems' problems themselves, or at least somewhere other than `g77', which expects a working ANSI C environment (and, for now, a GNU C compiler to compile `g77' itself). * Complain if initialized common redeclared as larger in subsequent program unit. * Warn if blank common initialized, since its size can vary and hence related warnings that might be helpful won't be seen. * New `-fbackslash' option, on by default, that causes `\' within `CHARACTER' and Hollerith constants to be interpreted a la GNU C. Note that this behavior is somewhat different from `f2c''s, which supports only a limited subset of backslash (escape) sequences. * Make `-fugly-args' the default. * New `-fugly-init' option, on by default, that allows typeless/Hollerith to be specified as initial values for variables or named constants (`PARAMETER'), and also allows character<->numeric conversion in those contexts--turn off via `-fno-ugly-init'. * New `-finit-local-zero' option to initialize local variables to binary zeros. This does not affect whether they are `SAVE'd, i.e. made automatic or static. * New `-Wimplicit' option to warn about implicitly typed variables, arrays, and functions. (Basically causes all program units to default to `IMPLICIT NONE'.) * `-Wall' now implies `-Wuninitialized' as with `gcc' (i.e. unless `-O' not specified, since `-Wuninitialized' requires `-O'), and implies `-Wunused' as well. * `-Wunused' no longer gives spurious messages for unused `EXTERNAL' names (since they are assumed to refer to block data program units, to make use of libraries more reliable). * Support `%LOC()' and `LOC()' of character arguments. * Support null (zero-length) character constants and expressions. * Support `f2c''s `IMAG()' generic intrinsic. * Support `ICHAR()', `IACHAR()', and `LEN()' of character expressions that are valid in assignments but not normally as actual arguments. * Support `f2c'-style `&' in column 1 to mean continuation line. * Allow `NAMELIST', `EXTERNAL', `INTRINSIC', and `VOLATILE' in `BLOCK DATA', even though these are not allowed by the standard. * Allow `RETURN' in main program unit. * Changes to Hollerith-constant support to obey Appendix C of the standard: - Now padded on the right with zeros, not spaces. - Hollerith "format specifications" in the form of arrays of non-character allowed. - Warnings issued when non-blank truncation occurs when converting to another type. - When specified as actual argument, now passed by reference to `INTEGER' (padded on right with spaces if constant too small, otherwise fully intact if constant wider the `INTEGER' type) instead of by value. *Warning:* `f2c' differs on the interpretation of `CALL FOO(1HX)', which it treats exactly the same as `CALL FOO('X')', but which the standard and `g77' treat as `CALL FOO(%REF('X '))' (padded with as many blanks as necessary to widen to `INTEGER'), essentially. * Changes and fixes to typeless-constant support: - Now treated as a typeless double-length `INTEGER' value. - Warnings issued when overflow occurs. - Padded on the left with zeros when converting to a larger type. - Should be properly aligned and ordered on the target machine for whatever type it is turned into. - When specified as actual argument, now passed as reference to a default `INTEGER' constant. * `%DESCR()' of a non-`CHARACTER' expression now passes a pointer to the expression plus a length for the expression just as if it were a `CHARACTER' expression. For example, `CALL FOO(%DESCR(D))', where `D' is `REAL*8', is the same as `CALL FOO(D,%VAL(8)))'. * Name of multi-entrypoint master function changed to incorporate the name of the primary entry point instead of a decimal value, so the name of the master function for `SUBROUTINE X' with alternate entry points is now `__g77_masterfun_x'. * Remove redundant message about zero-step-count `DO' loops. * Clean up diagnostic messages, shortening many of them. * Fix typo in `g77' man page. * Clarify implications of constant-handling bugs in `f/BUGS'. * Generate better code for `**' operator with a right-hand operand of type `INTEGER'. * Generate better code for `SQRT()' and `DSQRT()', also when `-ffast-math' specified, enable better code generation for `SIN()' and `COS()'. * Generate better code for some kinds of array references. * Speed up lexing somewhat (this makes the compilation phase noticably faster).