mm/migrate: Mark unmap_and_move() "noinline" to avoid ICE in gcc 4.7.3

Commit Message

Andrew Morton April 1, 2015, 7:40 p.m. UTC

On Wed, 01 Apr 2015 10:47:49 +0100 Marc Zyngier <[e-mail]> wrote:

> > -static int unmap_and_move(new_page_t get_new_page, free_page_t put_new_page,
> > -			unsigned long private, struct page *page, int force,
> > -			enum migrate_mode mode)
> > +static noinline int unmap_and_move(new_page_t get_new_page,
> > +				   free_page_t put_new_page,
> > +				   unsigned long private, struct page *page,
> > +				   int force, enum migrate_mode mode)
> >  {
> >  	int rc = 0;
> >  	int *result = NULL;
> > 
> 
> Ouch. That's really ugly. And on 32bit ARM, we end-up spilling half of
> the parameters on the stack, which is not going to help performance
> either (not that this would be useful on 32bit ARM anyway...).
> 
> Any chance you could make this dependent on some compiler detection
> mechanism?

With my arm compiler (gcc-4.4.4) the patch makes no difference -
unmap_and_move() isn't being inlined anyway.

How does this look?

Kevin, could you please retest?  I might have fat-fingered something...

Patch

--- a/mm/migrate.c~mm-migrate-mark-unmap_and_move-noinline-to-avoid-ice-in-gcc-473-fix
+++ a/mm/migrate.c
@@ -901,10 +901,20 @@  out:
 }
 
 /*
+ * gcc-4.7.3 on arm gets an ICE when inlining unmap_and_move().  Work around
+ * it.
+ */
+#if GCC_VERSION == 40703 && defined(CONFIG_ARM)
+#define ICE_noinline noinline
+#else
+#define ICE_noinline
+#endif
+
+/*
  * Obtain the lock on page, remove all ptes and migrate the page
  * to the newly allocated page in newpage.
  */
-static noinline int unmap_and_move(new_page_t get_new_page,
+static ICE_noinline int unmap_and_move(new_page_t get_new_page,
 				   free_page_t put_new_page,
 				   unsigned long private, struct page *page,
 				   int force, enum migrate_mode mode)