Re: [PATCH] eal: provide rte attribute macro for GCC attribute

[Tyler Retzlaff <roretzla@linux.microsoft.com>]

On Sun, Feb 18, 2024 at 05:44:48PM +0100, Thomas Monjalon wrote:
> 18/02/2024 17:38, Morten Brørup:
> > > From: Thomas Monjalon [mailto:thomas@monjalon.net]
> > > Sent: Sunday, 18 February 2024 16.35
> > > 
> > > 18/02/2024 13:53, Morten Brørup:
> > > > > From: Thomas Monjalon [mailto:thomas@monjalon.net]
> > > > > Sent: Sunday, 18 February 2024 13.24
> > > > >
> > > > > 15/02/2024 23:20, Tyler Retzlaff:
> > > > > > Provide a new macro __rte_attribute(a) that when directly used
> > > > > > compiles to empty for MSVC and to __attribute__(a) when using
> > > > > GCC/LLVM.
> > > > > >
> > > > > > Replace direct use of __attribute__ in __rte_xxx macros where
> > > there
> > > > > is
> > > > > > existing empty expansion of the macro for MSVC allowing removal
> > > of
> > > > > > repeated #ifdef RTE_TOOLCHAIN_MSVC per macro to expand empty.
> > > > >
> > > > > I'm not sure it makes sense.
> > > > > I prefer seeing clearly what is empty with MSVC.
> > > >
> > > > This topic has previously been discussed in another context - adding
> > > external libraries [1].
> > > >
> > > > Like you do here, I generally preferred #ifdefs in the code, but the
> > > majority preferred stubs "promoting improved code readability".
> > > 
> > > Stubs may make sense in many places,
> > > but here we are talking about rte_common.h
> > > where we abstract differences between arch and compilers,
> > > so it is the right place to be explicit with compilers support.
> > 
> > Very strong point. I'm convinced.
> > 
> > Should the new rte_attribute() macro still be introduced for other uses of __attribute__(), e.g. the somewhat exotic attributes in eal/include/rte_lock_annotations.h?
> 
> They are all wrapped in a meaningful macro.
> 
> > The not-so-exotic attributes could have new macros added, e.g. __rte_const and __rte_pure.
> 
> Yes we need wrappers for all attributes.

These are my expectation of what is supposed to be done based on the
comments in checkpatches.sh. The intention would be that
__rte__attribute not be directly used outside of rte_common.h (as is
required for __attribute__).

Introducing __rte_attribute was an attempt by me to appease previous
complaints of there being a conditional MSVC check for many macros which
was triggered by finding bare use of __attribute__ elsewhere
contradicting what was commented in checkpatches.sh

> 
> 
> > > > I might argue that Tyler is following that guidance here; and perhaps
> > > the decision should be reconsidered, now that we have a real-life
> > > example of how it affects code readability. ;-)
> > > >
> > > > [1]: https://inbox.dpdk.org/dev/20240109141009.497807-1-
> > > jerinj@marvell.com/
> 
>