On Thu, Aug 23, 2018 at 05:09:12PM -0700, Dave Hansen wrote:

On 08/23/2018 06:21 AM, Michal Hocko wrote:

quoted

--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h

@@ -1155,9 +1155,9 @@ static inline struct mem_section *__nr_to_section(unsigned long nr)
 #ifdef CONFIG_SPARSEMEM_EXTREME
 	if (!mem_section)
 		return NULL;
-#endif
 	if (!mem_section[SECTION_NR_TO_ROOT(nr)])
 		return NULL;
+#endif
 	return &mem_section[SECTION_NR_TO_ROOT(nr)][nr & SECTION_ROOT_MASK];
 }

This patch has no practical effect and only adds unnecessary churn.

#ifdef CONFIG_SPARSEMEM_EXTREME
...
#else
struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT];
#endif

The compiler knows that NR_SECTION_ROOTS==1 and that
!mem_section[SECTION_NR_TO_ROOT(nr) is always false.  It doesn't need
our help.

I didn't know the compile would optimize the code when this is a one dimension
array. Just wrote a code and their assembly looks the same.

Thanks for pointing out.

My goal with the sparsemem code, and code in general is t avoid #ifdefs
whenever possible and limit their scope to the smallest possible area
whenever possible.

-- 
Wei Yang
Help you, Help me