Re: [Patch] SourceLevelDebugPkg/SecPeiDebugAgentLib: Fix duplicate symbol

[Laszlo Ersek <lersek@redhat.com>]

On 05/25/17 19:38, Kinney, Michael D wrote:
> Laszlo,
> 
> I think the equivalent flag for GCC builds is --whole-archive.
> 
> I tried adding that flag to DLINK_FLAGS in GCC5, and I get the 
> following error building OVMF from edk2/master with 
> -D SOURCE_DEBUG_ENABLE set.
> 
> DxeLoad.obj (symbol from plugin): In function `InstallIplPermanentMemoryPpis':
> (.text+0x0): multiple definition of `mMemoryDiscoveredNotifyList'
> SecPeiDebugAgentLib.obj (symbol from plugin):(.text+0x0): first defined here
> collect2: error: ld returned 1 exit status

Great find! That's the error message we should get.

Unfortunately, after reading the "ld" manual on "--whole-archive", it
seems that the complete object files will actually be copied into the
resultant binary, even if several of their symbols will remain unused. I
think that's quite sub-optimal. (I haven't verified this though.) What
we'd like to get is (a) the full verification at link time, and (b)
inclusion of *only* those symbols that are actually necessary.

In your testing, when you build OVMF with and without "--whole-archive",
do you see a difference in, say, the DXEFV footprint, when the build
completes?

(If so, then I wonder if we should add "--whole-archive" only to the
NOOPT build... Not sure.)

> Visual Studio 2015 Update 2 has also added a new linker flag called
> /WHOLEARCHIVE.  I am working on evaluating that flag to see if it 
> catches the same issue.

Thanks!
Laszlo

>> -----Original Message-----
>> From: Laszlo Ersek [mailto:lersek@redhat.com]
>> Sent: Thursday, May 25, 2017 9:09 AM
>> To: Kinney, Michael D <michael.d.kinney@intel.com>; Ard Biesheuvel
>> <ard.biesheuvel@linaro.org>; Andrew Fish (afish@apple.com) <afish@apple.com>
>> Cc: Wu, Hao A <hao.a.wu@intel.com>; edk2-devel@lists.01.org; Fan, Jeff
>> <jeff.fan@intel.com>
>> Subject: Re: [edk2] [Patch] SourceLevelDebugPkg/SecPeiDebugAgentLib: Fix duplicate
>> symbol
>>
>> On 05/25/17 03:47, Kinney, Michael D wrote:
>>> Andrew,
>>>
>>> I think I have found an alternate fix for this XCODE5 specific
>>> build failure.  Since there appears to be a difference in the
>>> linker behavior between MSFT/GCC/XCODE tool chains, I reviewed
>>> the 'ld' command line options used in XCODE5 tool chain in
>>> tools_def.txt.
>>>
>>> There is a flag set call '-all_load'.  The description of this
>>> flag is 'Loads all members of static archive libraries.'.
>>>
>>> I tried removing this flag from the XCODE5 specific SLINK_FLAGS
>>> and DLINK_FLAGS statements in tools_def.txt, and the duplicate
>>> symbol build failure is no longer present.  I am able to build
>>> and boot OVMF with XCODE5 with -D SOURCE_DEBUG_ENABLE flag set.
>>>
>>> This seems to make XCODE5 linker behavior match the MSFT and GCC
>>> linker behavior.
>>>
>>> Do you know why '-all_load' is used in XCODE5 and what impacts
>>> there may be from removing it?
>>
>> Please don't remove -all_load from there; instead we should figure out
>> if the same can be brought to MSFT and GCC.
>>
>> The error message that XCODE5 emitted caught a real bug (undefined
>> behavior according to ISO C, see my previous email), and so we should
>> keep that detection enabled (we should even extend it to other
>> toolchains, if that's possible).
>>
>> As for docs, I found this:
>>
>> http://www.manpages.info/macosx/ld.1.html
>>
>>> -all_load
>>>     Loads all members of static archive libraries. This option does
>>>     not apply to dynamic shared libraries.
>>
>>
>> Thanks
>> Laszlo