[PATCH 4/6] arm: Use the WEAK assembly entry point consistently

[Tom Rini]

On Wed, Nov 23, 2022 at 11:50:59PM +0100, Pali Rohár wrote:
> On Tuesday 22 November 2022 12:31:56 Tom Rini wrote:
> > It is a bad idea, and more modern toolchains will fail, if you declare
> > an assembly function to be global and then weak, instead of declaring it
> > weak to start with. Update assorted assembly files to use the WEAK macro
> > directly.
> > 
> > Signed-off-by: Tom Rini <trini at konsulko.com>
> 
> During debugging of Nokia N900 code I was looking at this and I
> originally thought that this redefinition is the issue why N900 u-boot
> did not work... (but as we now know, the n900 issue was somewhere else).
> 
> So I agree with this change, feel free to add my:
> 
> Reviewed-by: Pali Rohár <pali at kernel.org>
> 
> ... but even after this change, linked u-boot.bin binary is
> not-so-correct. It works but has an issue: In final u-boot.bin binary
> there is both weak and non-weak version of every weak function. You can
> verify it for example by looking at "save_boot_params" code (really
> code, not just symbol) in u-boot ELF binary.
> 
> The reason for this is that linker (even LTO enabled) cannot eliminate
> code for weak version of function because it does not know how to
> "drop" it from binary/assembly code. So linker just set that non-weak
> version of function is active and let non-weak version present in binary
> as probably dead code.
> 
> This is affected only by assembly files, not by C files, because gcc is
> called with -ffunction-sections -fdata-sections flags which cause that
> every (weak) function is in its separate section (so C function
> "int abc() { ... }" is put into the section ".text.abc" instead of
> ".text") and linker's flag --gc-sections (or LTO optimization) then drop
> all unreferenced sections.
> 
> I do not know how fix this issue in assembly files. But cannot be WEAK
> macro modified to change section to ".text.<entry_name>" to mimic C
> compiler behavior? Would this cause any issues?

Yes, you're right about the cause, and potential solution. If you can
come up with a way to get each assembly function put in to a separate
.text.funcname section, that would be great and much appreciated. I
think I tried this at one point a long long time ago and it did work,
but I didn't have something clean, either. I think I was hoping that the
linux kernel folks would solve it in time, but they decided the
time/effort for --gc-sections wasn't worth it, in the end.

-- 
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20221123/acfa3045/attachment.sig>