Restrict -G option support to relevant targets

In <http://gcc.gnu.org/ml/gcc-patches/2010-07/msg02225.html> I noted
there was an issue with common.opt having a -G option taking an
argument that was incompatible with a linker option -G not taking an
argument on some targets.

This patch fixes this issue by making the common.opt -G option into a
target option - which is logically is - that is only present on
targets that actually support the relevant feature.  The targets that
use g_switch_* (that is, that meaningfully support -G) are alpha frv
ia64 lm32 m32r mips rs6000 score.  All of these have appropriate specs
to pass such options to cc1 and the assembler and linker as needed.

In addition to those targets, bfin has specs to pass -G options to the
assembler and linker but not cc1; as far as I can see, those specs are
useless as the relevant options don't exist.

The -G option to cc1 takes an operand that may be joined or separate
(-Gn or -G n).  For this to work with the separate operand passed to
the driver, SWITCH_TAKES_ARG must be defined to include -G.  This is
the case for six of the eight architectures with -G support; alpha and
ia64 do not have the required SWITCH_TAKES_ARG support so -G will only
work with a joined argument on those targets.  iq2000 has a stray
SWITCH_TAKES_ARG definition with -G support despite no other small
data support.

This patch limits -G support in cc1 to the listed eight architectures,
by moving the option to a new g.opt file.  As a target option the code
supporting it can no longer be in opts.c, so it is moved into the
target handle_option hooks.  The duplication is perhaps unfortunate,
but having a function to replace two statements in each place seems
excessive.  A future patch in this series will cause (effectively)
*_set variables to be created and set automatically, so at that point
-G can be entirely described in the .opt file without handle_option
hook support for it at all.

This patch does not address the SWITCH_TAKES_ARG issues.  Once the
driver shares option processing machinery with the core compiler, all
the places in the driver using SWITCH_TAKES_ARG can be adapted so that
it is eventually only used for interpreting unknown options
(i.e. those handled solely in specs) and existing definitions can be
largely garbage-collected, but a fix may come earlier in a general
cleanup patch for options issues found.  This patch does not do
anything with the bfin specs; again, that may be fixed in general
cleanup.

Bootstrapped with no regressions on x86_64-unknown-linux-gnu.  I have
also built cc1 for the following targets as a sanity check on the
target-specific changes: alpha-linux-gnu frv-linux-gnu ia64-linux-gnu
lm32-elf m32r-linux-gnu mips-linux-gnu powerpc-linux-gnu
score-elf.  OK to commit?

2010-07-30  Joseph Myers  <joseph@codesourcery.com>

	* common.opt (-G): Don't define option here.
	* config/g.opt: New.
	* config.gcc: Use g.opt for alpha, frv, ia64, lm32, m32r, mips,
	rs6000/powerpc and score targets.
	* opts.c (common_handle_option): Don't handle -G here.
	* config/alpha/alpha.c (alpha_handle_option): Handle -G.
	* config/frv/frv.c (frv_handle_option): Handle -G.
	* config/ia64/ia64.c (ia64_handle_option): Handle -G.
	* config/lm32/lm32.c (lm32_handle_option, TARGET_HANDLE_OPTION):
	New.
	* config/m32r/m32r.c (m32r_handle_option): Handle -G.
	* config/mips/mips.c (mips_handle_option): Handle -G.
	* config/rs6000/rs6000.c (rs6000_handle_option) Handle -G.
	* config/score/score.c (score_handle_option): Handle -G.

Restrict -G option support to relevant targets

Commit Message

Comments

Patch