arm64: Mark functions using explicit register variables as '__always_inline'

As of ac7c3e4ff4 ("compiler: enable CONFIG_OPTIMIZE_INLINING forcibly"),
inline functions are no longer annotated with '__always_inline', which
allows the compiler to decide whether inlining is really a good idea or
not. Although this is a great idea on paper, the reality is that AArch64
GCC prior to 9.1 has been shown to get confused when creating an
out-of-line copy of a function passing explicit 'register' variables
into an inline assembly block:

  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91111

It's not clear whether this is specific to arm64 or not but, for now,
ensure that all of our functions using 'register' variables are marked
as '__always_inline' so that the old behaviour is effectively preserved.

Hopefully other architectures are luckier with their compilers.

Cc: Masahiro Yamada <yamada.masahiro@socionext.com>
Cc: Nicolas Saenz Julienne <nsaenzjulienne@suse.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Will Deacon <will@kernel.org>

arch/arm64/include/asm/atomic_lse.h

@@ -321,7 +321,8 @@ static inline s64 __lse_atomic64_dec_if_positive(atomic64_t *v)
 }
 
 #define __CMPXCHG_CASE(w, sfx, name, sz, mb, cl...)			\
-static inline u##sz __lse__cmpxchg_case_##name##sz(volatile void *ptr,	\
+static __always_inline u##sz						\
+__lse__cmpxchg_case_##name##sz(volatile void *ptr,			\
 					      u##sz old,		\
 					      u##sz new)		\
 {									\
@@ -362,7 +363,8 @@ __CMPXCHG_CASE(x,  ,  mb_, 64, al, "memory")
 #undef __CMPXCHG_CASE
 
 #define __CMPXCHG_DBL(name, mb, cl...)					\
-static inline long __lse__cmpxchg_double##name(unsigned long old1,	\
+static __always_inline long						\
+__lse__cmpxchg_double##name(unsigned long old1,				\
 					 unsigned long old2,		\
 					 unsigned long new1,		\
 					 unsigned long new2,		\