fixes for Darwin/x86

[Stuart Hastings <stuart at apple dot com>]

We've been neglecting the Darwin/x86 port.  Mea culpa.

My Darwin/x86 box is currently running version 8F1111 of OS X,  
upgraded with cctools v590.20 (cctools means assembler, static and  
dynamic linkers, etc).

The current Darwin/x86 mainline GCC currently fails about 750 DejaGnu  
testcases for C.  That's using the 590.20 linker; older linkers  
generate a warning that breaks an additional 70-odd tests.

The attached patch reduces this to 80 failures for C.  ObjC gains  
about 160 passes, no regressions, and ObjC++ show no regressions.

C++ shows no regressions, but about 30 currently PASSing tests become  
UNSUPPORTED; I think this is a DejaGnu issue. The patch adds a Darwin/ 
x86-only warning that says "-fpic is not supported; -fPIC assumed",  
just like Darwin/PPC, and the warning frightens DejaGnu into avoiding  
g++.dg/debug/debug3.C (see the -fpic checking in testsuite/lib/target- 
supports.exp).  DejaGnu has a regular expression that should match  
and ignore the warning (testsuite/lib/prune.exp); obviously it's not  
working.  Somebody (probably me) should figure this out.  For now,  
the attached patch duplicates the Darwin/PPC UNSUPPORTED behavior;  
Darwin/PPC and Darwin/x86 become "bug-for-bug compatible."  ;-)

AFAICT, Darwin/x86 must be the only GCC port where the default stack  
alignment can be /lowered/ from the commandline ("-mpreferred-stack- 
boundary=2").  Darwin requires a 16-byte (SSE) stack alignment, yet  
the Darwin/x86 GCC is obliged to support a few projects  
(cough*kernel*cough) that require 4-byte alignment.  It's trivial to  
make the STACK_BOUNDARY macro expand to a variable, but that won't  
bootstrap, as the middle-end uses STACK_BOUNDARY in signed and  
unsigned comparisons.  This hasn't been a problem before, as every  
other port seems to define STACK_BOUNDARY to an integral constant (or  
a ?: yielding multiple integer constants) of indeterminate  
signedness.  Since we must pick one signedness, I arbitrarily chose  
to make stack alignments "unsigned" in the attached patch; if the  
reviewers prefer otherwise, I'll change it.  IMHO, the consistency is  
important; the choice is not.

I've revised the x86 Mach-O "picbase" stuff to support multiple  
instances of "builtin_setjmp_receiver," including a new "darwin.md"  
file, all slavishly copied from GeoffKs PPC implementation.  This  
fixes many DejaGnu EH failures.

The patch adds Mach-O support for TARGET_DEEP_BRANCH_PREDICTION,  
since it was tied in with all the picbase wrangling.  Not a big deal.

On Darwin/x86: Bootstrapped, DejaGnu, about 30 "UNSUPPORTED"  
regressions in DejaGnu/C++ (described above).
On Darwin/PPC:  Bootstrapped.
On Linux/x86:  Bootstrapped.

2005-11-30  Stuart Hastings  <stuart@apple.com>

        * gcc/function.c (pad_to_arg_alignment): Stack alignments are
        unsigned.
        * gcc/calls.c (compute_argument_block_size): Ditto.
        * gcc/cfgexpand.c (expand_one_stack_var_at): Ditto.
        * gcc/config/i386/darwin.md: New file.
        * gcc/config/i386/i386.md (darwin.md): Include it.
        (UNSPEC_LD_MPIC): New.  (builtin_setjmp_receiver):
        Re-materialize Mach-O picbase.
        * gcc/config/i386/darwin.h (SUBTARGET_OVERRIDE_OPTIONS):
        Equate -fpic to -fPIC. (STACK_BOUNDARY): Darwin 16-byte stack
        alignment.  (TARGET_ASM_FILE_END):  New.
	(TARGET_DEEP_BRANCH_PREDICTION): Remove.
        (GOT_SYMBOL_NAME): Mach-O support.
        * gcc/config/i386/i386.c (override_options): Simplify stack
        boundary computation. (get_pc_thunk_name): Mach-O picbase
	re-materialization support.  (darwin_x86_file_end): New.

Patch:

Attachment: gcc.fsf.bugs.diffs.txt
Description: Text document

New file (config/i386/darwin.md):

Attachment: darwin.md
Description: Binary data

O.K. to commit to mainline?

stuart hastings
Apple Computer