Before this commit, weak symbols were not overridden by non-weak symbols
found in archive libraries when linking with recent versions of
binutils. As stated in the System V ABI, "the link editor does not
extract archive members to resolve undefined weak symbols".
This commit changes all Makefiles to use partial linking (ld -r) instead
of creating library archives, which forces all symbols to participate in
linking, allowing non-weak symbols to override weak symbols as intended.
This approach is also used by Linux, from which the gmake function
cmd_link_o_target (defined in config.mk and used in all Makefiles) is
inspired.
The name of each former library archive is preserved except for
extensions which change from ".a" to ".o". This commit updates
references accordingly where needed, in particular in some linker
scripts.
This commit reveals board configurations that exclude some features but
include source files that depend these disabled features in the build,
resulting in undefined symbols. Known such cases include:
- disabling CMD_NET but not CMD_NFS;
- enabling CONFIG_OF_LIBFDT but not CONFIG_QE.
Signed-off-by: Sebastien Carlier <sebastien.carlier@gmail.com>
Commit 14d0a02a "Rename TEXT_BASE into CONFIG_SYS_TEXT_BASE" missed a
few places, especially for boards that were added inbetween. Fix the
remaining issues.
Signed-off-by: Wolfgang Denk <wd@denx.de>
The change is currently needed to be able to remove the board
configuration scripting from the top level Makefile and replace it by
a simple, table driven script.
Moving this configuration setting into the "CONFIG_*" name space is
also desirable because it is needed if we ever should move forward to
a Kconfig driven configuration system.
Signed-off-by: Wolfgang Denk <wd@denx.de>
Provides a small speed increase and prepares for fully relocatable image.
Downside is the TEXT_BASE, bss, load address etc must ALL be aligned on a
a 4-byte boundary which is not such a terrible restriction as everything
is already 4-byte aligned anyway
Use TEXT_BASE rather than a hard-coded base address on x86 linker scripts.
This will allow any board to define its base link address without having
to modify the linker script
The eNET uses the sc520 software timers rather than the PC/AT clones
Set all interrupts and timers up to be PC/AT compatible
Signed-off-by: Graeme Russ <graeme.russ@gmail.com>
Add a parameter to the 32-bit entry to indicate if entry is from Real
Mode or not. If entry is from Real Mode, execute the destructive 'sizer'
routine to determine memory size as we are booting cold and running in
Flash. If not entering from Real Mode, we are executing a U-Boot image
from RAM and therefore the memory size is already known (and running
'sizer' will destroy the running image)
There are now two 32-bit entry points. The first is the 'in RAM' entry
point which exists at the start of the U-Boot binary image. As such,
you can load u-boot.bin in RAM and jump directly to the load address
without needing to calculate any offsets. The second entry point is
used by the real-to-protected mode switch
This patch also changes TEXT_BASE to 0x6000000 (in RAM). You can load
the resulting image at 0x6000000 and simple go 0x6000000 from the u-boot
prompt
Hopefully a later patch will completely elliminate any dependency on
TEXT_BASE like a relocatable linux kernel (perfect world)
Signed-off-by: Graeme Russ <graeme.russ@gmail.com>
Previously, a specific file or directory could be compiled with custom
CFLAGS by adding a Makefile variable such as:
CFLAGS_dlmalloc.o = <custom flags for common/dlmalloc.c>
or
CFLAGS_lib = <custom flags for lib directory>
This method breaks down once multiple files or directories share the
same path. Eg FLAGS_fileA = <custom flags> would incorrectly result in
both dir1/fileA.c and dir2/fileA.c being compiled with <custom flags>.
This change allows finer grained control which we need once we move
lib_$ARCH to arch/$ARCH/lib/ and lib_generic/ to lib/. Without this
change all lib/ directories would share the same custom CFLAGS.
Signed-off-by: Peter Tyser <ptyser@xes-inc.com>
gcc 4.3.2 optimiser creates multiple copies of inline asm (who knows why)
Remove use of global names for labels to prevent 'symbol already defined'
errors
Signed-off-by: Graeme Russ <graeme.russ@gmail.com>
A recent gcc added a new unaligned rodata section called '.rodata.str1.1',
which needs to be added the the linker script. Instead of just adding this
one section, we use a wildcard ".rodata*" to get all rodata linker section
gcc has now and might add in the future.
However, '*(.rodata*)' by itself will result in sub-optimal section
ordering. The sections will be sorted by object file, which causes extra
padding between the unaligned rodata.str.1.1 of one object file and the
aligned rodata of the next object file. This is easy to fix by using the
SORT_BY_ALIGNMENT command.
This patch has not be tested one most of the boards modified. Some boards
have a linker script that looks something like this:
*(.text)
. = ALIGN(16);
*(.rodata)
*(.rodata.str1.4)
*(.eh_frame)
I change this to:
*(.text)
. = ALIGN(16);
*(.eh_frame)
*(SORT_BY_ALIGNMENT(SORT_BY_NAME(.rodata*)))
This means the start of rodata will no longer be 16 bytes aligned.
However, the boundary between text and rodata/eh_frame is still aligned to
16 bytes, which is what I think the real purpose of the ALIGN call is.
Signed-off-by: Trent Piepho <xyzzy@speakeasy.org>
Brings i386 in line with other CPUs with a reset vector and frees up reset.c
for CPU reset functions
Signed-off-by: Graeme Russ <graeme.russ@gmail.com>
Sebastien Carlier