On Sun, Mar 15, 2020 at 09:36:11AM +0100, Thomas Monjalon wrote: > 14/03/2020 00:38, Dmitry Kozlyuk: > > > I suggest this change (I can send a patch fixing the issue in other .h > > > files): > > > > > > +/* > > > + * RTE_TOOLCHAIN_GCC is true if the target is built with GCC, > > > + * while a host application (like pmdinfogen) may have another compiler. > > > + * RTE_CC_IS_GNU is true if the file is compiled with GCC, > > > + * no matter it is a target or host application. > > > + */ > > > +#if defined __GNUC__ && !defined __clang__ && !defined __INTEL_COMPILER > > > +#define RTE_CC_IS_GNU > > > +#endif > > > + > > > +#ifdef RTE_CC_IS_GNU > > > -/** Define GCC_VERSION **/ > > > -#ifdef RTE_TOOLCHAIN_GCC > > > #define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + \ > > > __GNUC_PATCHLEVEL__) > > > #endif > > > @@ -96,7 +105,7 @@ typedef uint16_t unaligned_uint16_t; > > > * even if the underlying stdio implementation is ANSI-compliant, > > > * so this must be overridden. > > > */ > > > -#if defined(RTE_TOOLCHAIN_GCC) > > > +#ifdef RTE_CC_IS_GNU > > > #define __rte_format_printf(format_index, first_arg) \ > > > __attribute__((format(gnu_printf, format_index, first_arg))) > > > #else > > > > The code you propose LGTM itself. If you think it's a better solution than > > the one proposed below, I see no problem going with it. > > > > What I wonder is whether pmdinfogen should include the problematic code in > > the > > first place. The errors come from declarations in rte_debug.h, but > > pmdinfogen > > really can't use them, because definitions are compiled for different > > machine. Pmdinfogen pulls rte_debug.h via rte_pci.h, which is only needed > > for > > struct rte_pci_id. Shouldn't we instead break this bogus dependency chain by > > moving struct rte_pci_id to a separate header? > > Splitting headers to avoid EAL dependency looks to be a bad precedent to me. >
Rather than splitting, we can still fix this by breaking the dependency by just not having rte_debug.h included in rte_pci.h. From what I see, there is no need for that include to be there, and DPDK pretty much compiles fine with it removed - the only other change I had to make was add an extra include into mlx5_common.h to compensate for it not getting pulled in via rte_pci.h. Generally, while we should ensure that each header includes any other headers it depends upon, we must be careful that headers don't include other headers that they don't directly use. /Bruce
