mm/util: reduce stack usage of folio_mapcount

folio_test_hugetlb() will call PageHeadHuge which is a function call,
and blocks the compiler from recognizing this redundant load.

After rearranging the code, stack usage is dropped from 32 to 24, and
the function size is smaller (tested on GCC 12):

Before:
Stack usage:
mm/util.c:845:5:folio_mapcount  32      static
Size:
0000000000000ea0 00000000000000c7 T folio_mapcount

After:
Stack usage:
mm/util.c:845:5:folio_mapcount  24      static
Size:
0000000000000ea0 00000000000000b0 T folio_mapcount

Link: https://lkml.kernel.org/r/20220801173155.92008-1-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

mm/util.c

@@ -850,10 +850,10 @@ int folio_mapcount(struct folio *folio)
 		return atomic_read(&folio->_mapcount) + 1;
 
 	compound = folio_entire_mapcount(folio);
-	nr = folio_nr_pages(folio);
 	if (folio_test_hugetlb(folio))
 		return compound;
 	ret = compound;
+	nr = folio_nr_pages(folio);
 	for (i = 0; i < nr; i++)
 		ret += atomic_read(&folio_page(folio, i)->_mapcount) + 1;
 	/* File pages has compound_mapcount included in _mapcount */