Re: [RFC] GLOBAL_REMOVE_IF_UNREFERENCED, multiply defined symbols, and MSFT/GCC tool chains.

[Felix Poludov <Felixp@ami.com>]

> -----Original Message-----
> From: Ard Biesheuvel [mailto:ard.biesheuvel@linaro.org]
> Sent: Friday, March 24, 2017 1:32 PM
> To: Felix Poludov
> Cc: edk2-devel@lists.01.org
> Subject: Re: [edk2] [RFC] GLOBAL_REMOVE_IF_UNREFERENCED, multiply
> defined symbols, and MSFT/GCC tool chains.
> 
> On 24 March 2017 at 12:53, Felix Poludov <Felixp@ami.com> wrote:
> > Trying to add GCC support to projects based on MSFT tool chain, I'm keep
> stumbling into multiply defined symbol errors reported by GCC linker.
> > An attempt to understand why the errors are not reported by the
> Microsoft linker lead me to GLOBAL_REMOVE_IF_UNREFERENCED macro.
> > The purpose of the macro is to enable link time optimization of global
> variables.
> > However, the way it's defined for MSFT tool chain (__declspec(selectany) )
> has a side effect of explicitly allowing multiple instances of a symbol defined
> with GLOBAL_REMOVE_IF_UNREFERENCED.
> 
> This smells like a variant of the GCC COMMON issue that you quote
> below. 'Select any' presumes that globally visible symbols with the
> same name are guaranteed to refer to the same data item, and the
> nature and architecture of EDK2 make it impossible to ever be sure
> about that (i.e., cross-package library class resolution)
> 
> > For a while usage of the macro was the only option to enable global
> variable optimization.
> > Starting from VS2013 compiler supports /Gw flag that enables global
> variable optimization without a special declarator.
> >
> > I propose to make the following modifications:
> >
> > 1.      Change GLOBAL_REMOVE_IF_UNREFERENCED definition to an empty
> macro.
> >
> > Or more specifically, update macro definition in Base.h as follows:
> >
> > #ifndef GLOBAL_REMOVE_IF_UNREFERENCED
> >
> > #define GLOBAL_REMOVE_IF_UNREFERENCED
> >
> > #endif
> >
> > 2.      Update VS2013 and VS2015 compiler flags to add /Gw option
> >
> > 3.      Update compiler flags for older MSFT tool chains to define
> GLOBAL_REMOVE_IF_UNREFERENCED in a backward compatible manner for
> targets that enable optimization.
> >
> > /D GLOBAL_REMOVE_IF_UNREFERENCED =_declspec(selectany)
> >
> >
> > The advantages of these modifications are:
> >
> > -        Better detection of on potential errors by breaking the build when
> symbol is defined more than once.
> >
> > -        Improved consistency between MSFT and GCC tool chains
> >
> > -        Improved link time optimization with VS2013 and newer MSFT tool
> chains.
> >
> > For example, mGaugeData in
> MdeModulePkg/Library/DxeCorePerformanceLib/DxeCorePerformanceLib.c is
> not declared as GLOBAL_REMOVE_IF_UNREFERENCED, so
> >
> > today performance library is linked with DXE Core even when performance
> measurements are disabled.
> >
> > The alternative option is to enable support of multiply defined symbols on
> GCC tool chain.
> > One way to do it is by defining the macro as
> > #define GLOBAL_REMOVE_IF_UNREFERENCED __attribute__((weak))
> >
> 
> No! That may fix your build, but it only papers over the problem.
> 
> > However, I'm not sure that embracing multiple symbol definitions is a good
> idea.
> > For example, see Ard's arguments in this commit comment
> >
> https://github.com/tianocore/edk2/commit/214a3b79417f64bf2faae74af42c
> 1b9d23f50dc8
> >
> 
> GCC used -ffunction-sections and -fdata-sections, and performs link
> time garbage collection (--gc-sections), and so globals that are never
> referenced will ultimately be dropped anyway.
> 
> So in the mGaugeData case you mention above, GCC (or rather, GNU ld),
> should notice that no references to it exists, and the section
> containing just mGaugeData will be dropped from the build.
> 

I agree. Essentially, GLOBAL_REMOVE_IF_UNREFERENCED was a crutch created to work-around MSFT compiler limitation.
Now that MSFT compiler natively supports global data optimization, there no reason to have it.

> As I mentioned in the commit log of the above patch, STATIC is the key
> here. In my opinion, STATIC should be mandatory for all function and
> variables that are only referenced from the same compilation unit. Not
> only does it help the compiler produce better code (in the absence of
> LTO), it prevents namespace pollution and generally results in better
> structured code (given that you can't easily link to some symbol in
> another object)

Agree again.

> Given that each of the linker errors you get points to a potential
> problem in your code, perhaps the best approach would be to
> temporarily define GLOBAL_REMOVE_IF_UNREFERENCED to STATIC instead,
> so
> you can track down the occurrences that really require external
> linkage, but have multiple definitions, and fix those up manually.

Fixing is not a problem. My pain is that the bugs are in a third party code and I have to keep fixing them whenever new version is released,
which is not a rare event.
That's why I'm trying to come up with the modifications to break the MSFT compiler build when multiply defined symbol is encountered.

Please consider the environment before printing this email.

The information contained in this message may be confidential and proprietary to American Megatrends, Inc.  This communication is intended to be read only by the individual or entity to whom it is addressed or by their designee. If the reader of this message is not the intended recipient, you are on notice that any distribution of this message, in any form, is strictly prohibited.  Please promptly notify the sender by reply e-mail or by telephone at 770-246-8600, and then delete or destroy all copies of the transmission.