Make lang_specific_driver use cl_decoded_option structures

This patch continues the transition to using central option-processing
code everywhere by moving lang_specific_driver to working with
cl_decoded_option structures rather than the argv array.

In the process of so doing, various oddities in the code have been
noted, and some fixed while others are left for discussion of the
desired semantics and fixing by front-end maintainers.

* The drivers had various different versions of logic to determine
  what is an option and what is an option argument, most of which
  could go wrong in some cases.  This is fixed by using the central
  logic in opts-common.c (which in turn uses *SWITCH_TAKES_ARG for
  unknown options) to produce the cl_decoded_option structures.  (The
  code for translating options in the driver, and do_self_spec, still
  need converting to use the central logic to eliminate the remaining
  possibilities for such inconsistencies.)  Similarly, the drivers
  would inconsistently try to give errors for missing option
  arguments; this is now unambiguously the responsibility of the
  central option processing code (while the drivers just need to check
  for CL_ERR_MISSING_ARG in some cases to avoid dereferencing null
  pointers).

* Some of the ad hoc checks for options taking arguments still handled
  the -b and -V options that were removed a while back.  Similarly,
  the Fortran driver's option processing code handled those options.
  The -b and -V handling is removed.

* The drivers would handle non-option arguments "", "-", other
  single-character arguments and longer non-option arguments in
  various ways inconsistent from driver to driver and between the
  different types of such arguments listed above.  I have tried to
  keep the logic in question the same as before this patch, although
  probably it should be fixed for greater consistency (except where
  special handling for "-" as stdin is needed, which may just be in
  cppspec.c, by treating all these cases the same as longer
  non-options, I think).

* Because of string comparison when the drivers checked for options
  such as "-lc" or "-lm" already being present, some drivers checking
  for such options would fail to detect what should be equivalent
  forms split into two arguments such as "-l c".  Fixed by using
  cl_decoded_option structures.

* The C++ driver checked for an option "-ObjC++".  No such option
  exists; it would be passed through to cc1plus, which would then
  reject it.  The code for this option is removed.

* The Fortran code has an error "overflowed output arg list for %qs"
  that is a call to fatal_error but looks like it should be an
  internal_error instead since if it triggers there is a bug in the
  calculation of how large the new option list should be.  The patch
  does not change this.  (Actually, it would be better just to resize
  dynamically as needed rather than hardcoding a calculation of the
  maximum size that might be needed, and so eliminate this error
  completely.  The same applies in other drivers)

* The Fortran driver has another relic of code for the C++ +e option
  (recall that this was removed in 2.95 and I removed further relics
  in <http://gcc.gnu.org/ml/gcc-patches/2010-06/msg02407.html>).
  This is removed by this patch.

* The Fortran driver had a comment saying "Also, if -v is specified,
  but no other options that do anything special (allowing -V version,
  etc.), remember to add special stuff to make gcc command actually
  invoke all the different phases of the compilation process so all
  the version numbers can be seen.".  Apart from the reference to -V,
  there is nothing special in the Fortran code at all regarding the
  lack of other options with -v.

* The Fortran driver has a comment about "-lg2c -lm" but it always
  appears to use -lgfortran, not -lg2c.  Left unchanged.

* There are various target macros (defaulted with #ifndef in
  individual driver files) to control the libraries for different
  languages etc. - MATH_LIBRARY is documented in tm.texi.in, but
  MATH_LIBRARY_PROFILE, LIBSTDCXX, LIBSTDCXX_PROFILE, LIBSTDCXX_STATIC
  and FORTRAN_LIBRARY are not.  (No target actually has a non-default
  value of LIBSTDCXX_PROFILE or FORTRAN_LIBRARY, but by analogy with
  the others they should still be considered target macros.)  Left
  unchanged.

* The Java driver has its own support for @FILE files but expandargv
  has already been called by the time lang_specific_driver is reached
  so I think this is dead code.  Left unchanged.

* The Java driver handles various options that were not previously
  known to the other drivers: -classpath, -bootclasspath, -CLASSPATH,
  -encoding, -extdirs.  I entered these in the .opt file, so now the
  generic driver will handle them as taking arguments - and then
  reject them later as not being processed by any spec if it wouldn't
  have accepted them anyway because of some spec that handled them.
  In general I think such a change is a good idea because differences
  between the drivers are bad - the existence of separate drivers is
  to some extent a workaround for the generic driver not being able to
  tell which runtime libraries are needed by the object files it is
  linking.  It may however change behavior in corner cases, where
  -encoding or -extdirs were used as linker options to say that a
  symbol "ncoding" or "xtdirs" is the entry point.

* The Java -d option - -d with a separate argument being an alias for
  -foutput-class-dir= - is problematic in that it is incompatible with
  the semantics of -d with a joined argument.  For now I made
  translate_options handle it (for all languages); when
  translate_options becomes generic .opt alias machinery, that
  machinery will need to handle this special case of joined and
  separate arguments having different semantics.

* Java -extdirs is also peculiar in that it is different from
  -fextdirs/--extdirs; -extdirs is processed by the driver only, not
  directly generating jc1 options, while -fextdirs is processed by jc1
  only.  This may not result in any semantic differences, but looks
  odd.

* The Java driver handled "--syntax-only" and "--indirect-dispatch"
  but those would have been translated to the -f forms before
  lang_specific_driver is called; such handling is removed as useless.

Bootstrapped with no regressions on x86_64-unknown-linux-gnu.  OK to
commit?

2010-08-18  Joseph Myers  <joseph@codesourcery.com>

	* Makefile.in (gccspec.o, cppspec.o): Update dependencies.
	* common.opt (L, nodefaultlibs, nostdlib, pg, static): New
	options.
	* config/avr/avr.h (LIBSTDCXX): Remove initial "-l".
	* config/freebsd.h (MATH_LIBRARY_PROFILE): Remove initial "-l".
	* config/i386/djgpp.h (LIBSTDCXX): Remove initial "-l".
	* config/rs6000/aix.h (LIBSTDCXX_STATIC): Remove initial "-l".
	* config/s390/tpf.h (MATH_LIBRARY, LIBSTDCXX): Remove initial
	"-l".
	* cppspec.c: Include opts.h.
	(lang_specific_driver): Use cl_decoded_option structures.
	* doc/tm.texi.in (MATH_LIBRARY): Update documentation.
	* doc/tm.texi: Regenerate.
	* gcc.c (translate_options): Translate -d to -foutput-class-dir=.
	(driver_handle_option): Allow driver options needing no special
	processing.
	(process_command): Decode options before call to
	lang_specific_driver.  Pass decoded options to
	lang_specific_driver.
	* gcc.h (lang_specific_driver): Update prototype.
	* gccspec.c: Include opts.h.
	(lang_specific_driver): Use cl_decoded_option structures.
	* opts-common.c (option_ok_for_language, generate_option,
	generate_option_input_file): New.
	(decode_cmdline_option): Use option_ok_for_language.
	(decode_cmdline_options_to_array): Use generate_option_input_file.
	(handle_generated_option): Use generate_option.
	* opts.h (generate_option, generate_option_input_file): Declare.

cp:
2010-08-18  Joseph Myers  <joseph@codesourcery.com>

	* Make-lang.in (g++spec.o): Update dependencies.
	* g++spec.c: Include opts.h
	(MATH_LIBRARY, LIBSTDCXX): Remove initial "-l".
	(lang_specific_driver): Use cl_decoded_option structures.

fortran:
2010-08-18  Joseph Myers  <joseph@codesourcery.com>

	* Make-lang.in (gfortranspec.o): Update dependencies.
	* gfortranspec.c: Include coretypes.h before gcc.h.  Include
	opts.h.
	(MATH_LIBRARY, FORTRAN_LIBRARY): Remove initial "-l".
	(ADD_ARG_LIBGFORTRAN, Option, lookup_option): Remove.
	(g77_xargc): Make unsigned.
	(g77_xargv): Change to g77_x_decoded_options.
	(g77_newargc): Make unsigned.
	(g77_newargv): Change to g77_new_decoded_options.
	(strings_same, options_same): New.
	(append_arg): Use cl_decoded_option structures.
	(append_option): New.
	(add_arg_libgfortran): New.
	(lang_specific_driver): Use cl_decoded_option structures.

java:
2010-08-18  Joseph Myers  <joseph@codesourcery.com>

	* Make-lang.in (jvspec.o): Update dependencies.
	* jvspec.c: Include opts.h.
	(PARAM_ARG): Remove.
	(find_spec_file): Do not add leading -specs=.
	(lang_specific_driver): Use cl_decoded_option structures.
	* lang.opt (C, CLASSPATH, D, bootclasspath, classpath, encoding,
	extdirs, fmain=, s-bc-abi): New options.

Make lang_specific_driver use cl_decoded_option structures

Commit Message

Comments

Patch