long time. Go ahead and commit the patch since a co-worker has been
working up a code patch in this area and wondered why it hasn't been
committed yet.
PR: 35800
having any effect. This restores a calling convention traditionally used
by FreeBSD/i386 and which is expected by a number of assembly sources
in libc.
Reviewed by: obrien
Approved by: re (rwatson)
*sigh* GCC developers can't seem to make up their mind which pre-defined arch
symbols are the official ones. So make them, and the Sparc64 using Martians
happy with every permutation.
This fixes f77 and possibly changes some 'long long' handling.
Approved by: re(bmah)
now only produce ELF objects. It also makes us closer to stock GCC, and
simplifies the set of changes we still need from stock GCC on every import.
Applauded by: peter
Approved by: re
- Make DDB use %y instead of %z.
- Teach GCC about %y.
- Implement support for the C99 %z format modifier.
Approved by: re@
Reviewed by: peter
Tested on: i386, sparc64
1.\\{2,15\\} FREEBSD_NATIVE
1.\\{5,13\\} ELF, and objformat support
1.\\{16,23,25\\} Better cross building control
1.21 'GCC_OPTIONS'
1.27 cross-arch MD_EXEC_PREFIX fixes
cc -print-search-dir fixes
1.28 Read specs from /usr/libdata/gcc/specs, if available
Approved by: obrien
From current testsuite results, the optimizer bugs don't appear to exist
anymore. RTH@cygnus.com did a lot of work on the Alpha ELF code generator
for GCC 3.2[.0]. A recent FreeBSD/AXP GCC bootstrap is at
http://gcc.gnu.org/ml/gcc-testresults/2002-09/msg00604.html
In this bootstraps, all gcc libraries are built with -O2 and c-torture
gives -O2 a real workout. None of the remaining failures have anything
to do with -O2 optimizer bugs.
Submitted by: Loren James Rittle <rittle@latour.rsch.comm.mot.com>
code will know not to try to use `long long'.
Unfortunately the GCC spec parser will not allow us to properly detect the
"iso9899:1990" and "iso9899:199409" forms of the acceptable -std= arguments,
because of the ':' in the -std argument. :-( I have left them in the spec
as a place holder in hopes someone knows a way to make the detection of
them work.
Desired by: wollman
1.\{2,15\} FREEBSD_NATIVE
1.\{5,13\} ELF, and objformat support
1.\{16,23,25\} Better cross building control
1.21 'GCC_OPTIONS'
1.27 cross-arch MD_EXEC_PREFIX fixes
cc -print-search-dir fixes
1.28 Read specs from /usr/libdata/gcc/specs,
if available
Approved by: obrien
(put the function stabs in traditional order on a.out, or gdb doesn't see
function local variables), I failed to remove the related knobs here.
Effectively were overrode the ELF-wide definition in elfos.h w/o providing
new infrastructure. This is what caused GDB to fail to debug applications
compiled and linked with -stabs. This is because GCC was unconditionally
inserts .stabs instruction for functions after the function body. GDB was
getting confused because what it thinks is function beginning address is
actually function ending address.
Submitted by: Alexander Kabaev <ak03@gte.com>
for space. -Os could do this, but it was easy to hack an MD version.
This saves a whole 32 bytes in boot2, so I think it is worth using it.
(keep how much worse gcc 3.2 will compile boot2...)
Submitted by: bde (minus gcc 3.2 commentary)
This broke newfs (newfs left some garbage in a bitmap).
The ASM for:
#include <string.h>
int x, foo[100];
main()
{
memset(&foo[0], 0, x);
}
is (at least if you have fixed function alignment):
.file "z.c"
.text
.p2align 2,,3
.globl main
.type main,@function
main:
pushl %ebp
movl %esp, %ebp
pushl %edi
pushl %eax
movl x, %ecx
xorl %eax, %eax
shrl $2, %ecx
movl $foo, %edi
cld
rep
stosl
andl $-16, %esp
<-- the lower bits of `len' should be loaded
near here
testl $2, %edi <-- this seems to be meant to test the 2^1
bit in `len' (not alignment of the pointer
like it actually does). %edi is the wrong
register for holding the bits, since it is
still needed for the pointer.
je .L2
stosw
.L2:
testl $1, %edi <-- similarly for the 2^0 bit.
je .L3
stosb
.L3:
movl -4(%ebp), %edi
leave
ret
.Lfe1:
.size main,.Lfe1-main
.comm foo,400,32
.comm x,4,4
.ident "GCC: (GNU) 3.1 [FreeBSD] 20020509 (prerelease)"
This seems to only result in (len % 3) bytes not being cleared, since gcc
doesn't seem to use the builtin memset unless it knows that the pointer is
aligned. If %edi could be misaligned, then too many bytes would be set.
Submitted by: BDE
compiler.
* Undo the diking out of cross compiler logic from gcc.c rev 1.16.
* Add the `CROSS_STARTFILE_PREFIX' knob.
* Add our own definition of `STANDARD_INCLUDE_DIR'. This should have been
included in freebsd-native.h rev 1.5.
The changes from GCC 2.95.3 to 3.1 are large and it is hard to figure
out how to apply our changes. Once BDE can properly test GCC 3.1 he'll
add back in our changes.
Until we can figure out how to reimpliment our custom format warnings
since the FSF developers totally revamped them, and took away the hooks
we were using to accomplish our desires.
Note that the "NO_PROFILE_DATA"/"NO_PROFILE_COUNTERS" bits were added
to the stock FSF GCC 2.97 (and thus 3.1) source as i386 Linux needed them
also. (amazing what can get committed to the FSF GCC when needed by Linux...)
Remove our definition for FUNCTION_PROFILER as it is wrong.
Note that "jsr $28,_mcount" is a macro for
ldq $27, _mcount($29) !literal!1
jsr $28, ($27), _mcount !lituse_jsr!1
1. The call to _mcount is added by alpha_expand_prologue after we load the gp.
Our _mcount uses $27 for the incoming address, unlike OSF/1 and Linux,
which use $28. This probably doesn't matter since we probably don't use $27
within _mcount itself.
2. You can't use this insn with _mcount because it uses the PLT, which clobbers
the return address in $28. Note that the prologue_mcount pattern carefully
avoids adding the lituse_jsr relocation so that we call through the GOT
directly.
Submitted by: Richard Henderson <rth@redhat.com>
When rtld runs the .fini section in a shared lib (C++), the code in
question from .../contrib/gdb/config/alpha/crtbegin.asm first calls
__do_globals_dtors_aux and then __do_frame_takedown. Unfortunately, the
value of gp after a jsr is undefined and in this case had changed from before
the call, probably as a result of calling code in some other shared library.
The normal calling convention for alpha is to re-initialize gp using
'ldgp gp,0(ra)' after a jsr instruction but in this case no such
re-initialization is done. This leads to a bogus value being read for the
address of __do_frame_takedown and a quick segfault.
Submitted by: dfr
Obtained from: GCC 3.0
value, it forces GCC to not optimize above this level. For intance, GCC
made with "WANT_FORCE_OPTIMIZATION_DOWNGRADE=1" is a good setting for the
Alpha platform when building ports.
default options for GCC. These options are interpreted first and can be
overwritten by explicit command line parameters. This provides one way of
adding [temporary] options to your world build w/o editing /etc/make.conf.
If "WANT_DWARF2_UNWIND" you get it. Note that this is a different C++ ABI
than the FreeBSD default. So you will need to recompile all your C++ apps
that uses exceptions if you turn this on. I am adding it here for Yahoo!'s
use and for those that want to be early adopters to what I will make as the
default with the GCC 3.0 import.
that were removed from GCC 2.95.3.test4 and the subsequent release due
to problems on HP-UX. However, they work just fine on all the BSD's.
W/o these patches the following program segmentation faults if compiled
with -O2 (but not -Os or -O or -O0):
#include <stdio.h>
class A {
public:
A() { printf("c'tor A\n"); }
~A(){ printf("d'tor A\n"); }
};
class foo : public A {
public:
foo() { printf("C'tor foo\n"); throw 8; }
~foo() { printf("D'tor foo\n"); }
};
int main(){
try { foo fii; }
catch (int){ printf("catch ...\n"); }
return 0;
}
header to reduce the difference of our sources to the stock GNU/FSF ones.
While the mix binary format support was nice to have in the FreeBSD MI
header as a frame work, it just clutters up too much and makes the FreeBSD
MI header more different from the FSF/GNU stock one than it needs to be.
into the i386 MD FreeBSD header. Also comment tweaking, continuation line
style changes, and other minor changes to make this closer to the latest
FSF/GNU 3.0 sources (to reduce my headache when 3.0 is imported).
linked in addition to libc rather than instead of libc.
Ideally, "-pthread" would now be equivalent to adding "-lc_r" to the
end of the link command. But it is slightly different in this
implementation. Adding "-lc_r" to the link command would produce a
"ld" command line containing this:
... -lc_r /usr/lib/libgcc.a -lc /usr/lib/libgcc.a ...
but this implementation of the "-pthread" option produces this:
... /usr/lib/libgcc.a -lc_r -lc /usr/lib/libgcc.a ...
It would be possible to make them identical, but that doesn't fit
as nicely into GCC's way of doing things. I don't think the ordering
change will make any difference in practice.
This option depended on bits not part of the base system and required
people to install the LinuxThreads port in a manner non-consistent with
the workings of our Ports Collection.
The directions for properly linking with LinuxThreads are given by that
port at install time.
Requested by: jasone
before importing new versions of GCC. This differs from FREEBSD-Xlist
in that this is for use only with anoncvs checkouts, not tarball'ed
releases [snapshots].
This delete list applies to the 3-June-2000 import.
to generate the special .type and .size directives which are used to set
the corresponding fields of the linker symbol table entries in the ELF
object file. As such they are not i386-specific and thus belong in our
MI header. Otherwise on the Alpha we don't properly give the type and
size of dynamic symbols. Bintuil versions past 2.9.1 warn of this and
w/o this change, `ld' generates a lot of warnings during a `make world'.
I did not catch this on the EGCS 1.1.x --> GCC 2.95 upgrade.
So propogate this change to our custom hacks.
PR: 15549
Submitted by: Naohiko Tsuji <yakisoba@osaka.interq.or.jp>
Patrick Bihan-Faou <patrick@mindstep.com>
"The problem is that egcs/gcc-2.95's reorganisation of the prologue and
epilogue code to use rtl instead of output_asm_insn() completely broke our
hooks. rtl is emitted in a different order, only after optimisation, while
output_asm_insn() is emitted immediately. rtl is presumably used so that
the prologue and epilogue can be optimised.
I couldn't find any good examples to copy. gcc's own
FUNCTION_BLOCK_PROFILER still uses output_asm_insn() and seems to be
completely broken. One of the XXX comments points to this.
IIRC, the hacks here basically arrange to emit magic label names; then when
the magic names are output, they are transformed into prologue and epilogue
code."
Submitted by: bde
If one wishes to anchor the compiler toolchain tree somewhere other than /,
all one needs to do is set "TOOLS_PREFIX" to a different rooting.
Submitted by: marcel (in a different format and reworked by me)
of changing the search dirs. This also removes an used search dir,
removes unneeded redundancy, and a bugus dir we enherited on the i386
by baseing off of svr4.h.
We went from:
install: /usr/libexec/(null)
programs: /usr/libexec/<OBJFORMAT>/:/usr/libexec/:/usr/bin/:/usr/libexec/
libraries: /usr/libdata/gcc/:/usr/libexec/:/usr/ccs/lib/:/usr/lib/
to:
install: /usr/libexec/(null)
programs: /usr/libexec/<OBJFORMAT>/:/usr/libexec/
libraries: /usr/libexec/:/usr/lib/
