The purpose of this patch is to make kernel buildable
with "gcc -ffunction-sections -fdata-sections".
This patch fixes blackfin architecture.
Signed-off-by: Denys Vlasenko <[email protected]>
--
vda
--- 0.org/arch/blackfin/kernel/vmlinux.lds.S Wed Jul 2 00:40:40 2008
+++ 1.fixname/arch/blackfin/kernel/vmlinux.lds.S Wed Jul 2 00:45:51 2008
@@ -91,7 +91,7 @@
__sdata = .;
/* This gets done first, so the glob doesn't suck it in */
. = ALIGN(32);
- *(.data.cacheline_aligned)
+ *(.cacheline_aligned.data)
#if !L1_DATA_A_LENGTH
. = ALIGN(32);
On Tue, Jul 1, 2008 at 8:35 PM, Denys Vlasenko wrote:
> The purpose of this patch is to make kernel buildable
> with "gcc -ffunction-sections -fdata-sections".
> This patch fixes blackfin architecture.
the comment right above what you changed says it already works for
Blackfin. so you arent fixing it at all.
> /* This gets done first, so the glob doesn't suck it in */
> . = ALIGN(32);
> - *(.data.cacheline_aligned)
> + *(.cacheline_aligned.data)
ive built Blackfin kernels with function/data sections a long time ago
... but iirc, there were toolchain problems, so i havent pursued it
since.
if you're going to muck with names, you might as well do it right the
first time: move the names into asm-generic/vmlinux.lds.h so we dont
have to worry about this kind of churn again.
-mike
On Wednesday 02 July 2008 00:58, Mike Frysinger wrote:
> On Tue, Jul 1, 2008 at 8:35 PM, Denys Vlasenko wrote:
> > The purpose of this patch is to make kernel buildable
> > with "gcc -ffunction-sections -fdata-sections".
> > This patch fixes blackfin architecture.
>
> the comment right above what you changed says it already works for
> Blackfin. so you arent fixing it at all.
> > /* This gets done first, so the glob doesn't suck it in */
> > . = ALIGN(32);
> > - *(.data.cacheline_aligned)
> > + *(.cacheline_aligned.data)
This may pull in an unrelated data object named "cacheline_aligned"
(say, a static variable in a driver). If that variable is not
itself aligned to the cacheline size, it will mess up alignment of all
objects in .data.cacheline_aligned which follow. Not good.
To be safe from such weird and hard to debug problems
it's better to not use names like .data.XXXX at all.
I just uniformly renamed al such "special sections"
in the kernel to .XXXXX.data
--
vda
On Tue, Jul 1, 2008 at 8:08 PM, Denys Vlasenko wrote:
> On Wednesday 02 July 2008 00:58, Mike Frysinger wrote:
>> On Tue, Jul 1, 2008 at 8:35 PM, Denys Vlasenko wrote:
>> > The purpose of this patch is to make kernel buildable
>> > with "gcc -ffunction-sections -fdata-sections".
>> > This patch fixes blackfin architecture.
>>
>> the comment right above what you changed says it already works for
>> Blackfin. so you arent fixing it at all.
>> > /* This gets done first, so the glob doesn't suck it in */
>> > . = ALIGN(32);
>> > - *(.data.cacheline_aligned)
>> > + *(.cacheline_aligned.data)
>
> This may pull in an unrelated data object named "cacheline_aligned"
> (say, a static variable in a driver). If that variable is not
> itself aligned to the cacheline size, it will mess up alignment of all
> objects in .data.cacheline_aligned which follow. Not good.
>
> To be safe from such weird and hard to debug problems
> it's better to not use names like .data.XXXX at all.
> I just uniformly renamed al such "special sections"
> in the kernel to .XXXXX.data
you're right of course ... i obviously hadnt thought of this. please
however still abstract this stuff into the common header.
i just tested the current Blackfin kernel and it does build/link/run
fine with -ffunction-sections/-fdata-sections. however, --gc-sections
still causes crashes (but this is because all of the .init.setup
sections are flagged as unused).
-mike