[PATCH 1/2] compiler.h: add support for function attribute assume_aligned

[PATCH 1/2] compiler.h: add support for function attribute assume_aligned

[Rasmus Villemoes]

gcc 4.9 added the function attribute assume_aligned, indicating to the
caller that the returned pointer may be assumed to have a certain
minimal alignment. This is useful if, for example, the return value is
passed to memset(). Add a shorthand macro for that.

Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
---
 include/linux/compiler-gcc.h | 17 +++++++++++++++++
 include/linux/compiler.h     |  8 ++++++++
 2 files changed, 25 insertions(+)

diff --git a/include/linux/compiler-gcc.h b/include/linux/compiler-gcc.h
index dfaa7b3e9ae9..18606a460baf 100644
--- a/include/linux/compiler-gcc.h
+++ b/include/linux/compiler-gcc.h
@@ -210,6 +210,23 @@
 #define __visible	__attribute__((externally_visible))
 #endif
 
+
+#if GCC_VERSION >= 40900
+/*
+ * __assume_aligned(n, k): Tell the optimizer that the returned
+ * pointer can be assumed to be k modulo n. The second argument is
+ * optional (default 0), so we use a variadic macro to make the
+ * shorthand.
+ *
+ * Beware: Do not apply this to functions which may return
+ * ERR_PTRs. Also, it is probably unwise to apply it to functions
+ * returning extra information in the low bits (but in that case the
+ * compiler should see some alignment anyway, when the return value is
+ * massaged by 'flags = ptr & 3; ptr &= ~3;').
+ */
+#define __assume_aligned(a, ...) __attribute__((__assume_aligned__(a, ## __VA_ARGS__)))
+#endif
+
 /*
  * GCC 'asm goto' miscompiles certain code sequences:
  *
diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index c836eb2dc44d..6167ca663ad9 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -393,6 +393,14 @@ static __always_inline void __write_once_size(volatile void *p, void *res, int s
 #define __visible
 #endif
 
+/*
+ * Assume alignment of return value.
+ */
+#ifndef __assume_aligned
+#define __assume_aligned(a, ...)
+#endif
+
+
 /* Are two types/vars the same type (ignoring qualifiers)? */
 #ifndef __same_type
 # define __same_type(a, b) __builtin_types_compatible_p(typeof(a), typeof(b))
-- 
2.1.3

[PATCH 2/2] slab.h: sprinkle __assume_aligned attributes

[Rasmus Villemoes]

The various allocators return aligned memory. Telling the compiler
that allows it to generate better code in many cases, for example when
the return value is immediately passed to memset().

Some code does become larger, but at least we win twice as much as we lose:

$ scripts/bloat-o-meter /tmp/vmlinux vmlinux
add/remove: 0/0 grow/shrink: 13/52 up/down: 995/-2140 (-1145)

An example of the different (and smaller) code can be seen in mm_alloc(). Before:

:       48 8d 78 08             lea    0x8(%rax),%rdi
:       48 89 c1                mov    %rax,%rcx
:       48 89 c2                mov    %rax,%rdx
:       48 c7 00 00 00 00 00    movq   $0x0,(%rax)
:       48 c7 80 48 03 00 00    movq   $0x0,0x348(%rax)
:       00 00 00 00
:       31 c0                   xor    %eax,%eax
:       48 83 e7 f8             and    $0xfffffffffffffff8,%rdi
:       48 29 f9                sub    %rdi,%rcx
:       81 c1 50 03 00 00       add    $0x350,%ecx
:       c1 e9 03                shr    $0x3,%ecx
:       f3 48 ab                rep stos %rax,%es:(%rdi)

After:

:       48 89 c2                mov    %rax,%rdx
:       b9 6a 00 00 00          mov    $0x6a,%ecx
:       31 c0                   xor    %eax,%eax
:       48 89 d7                mov    %rdx,%rdi
:       f3 48 ab                rep stos %rax,%es:(%rdi)

So gcc's strategy is to do two possibly (but not really, of course)
unaligned stores to the first and last word, then do an aligned rep
stos covering the middle part with a little overlap. Maybe arches
which do not allow unaligned stores gain even more.

I don't know if gcc can actually make use of alignments greater than 8
for anything, so one could probably drop the __assume_xyz_alignment
macros and just use __assume_aligned(8).

Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
---

The increases in code size are mostly caused by gcc deciding to
opencode strlen() using the check-four-bytes-at-a-time trick when it
knows the buffer is sufficiently aligned (one function grew by 200
bytes). Now it turns out that many of these strlen() calls showing up
were in fact redundant, and they're gone from -next. Applying the two
patches to next-20151001 bloat-o-meter instead says

add/remove: 0/0 grow/shrink: 6/52 up/down: 244/-2140 (-1896)


 include/linux/slab.h | 43 ++++++++++++++++++++++++++-----------------
 1 file changed, 26 insertions(+), 17 deletions(-)

diff --git a/include/linux/slab.h b/include/linux/slab.h
index 7e37d448ed91..b89a67d297f1 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h
@@ -158,6 +158,24 @@ size_t ksize(const void *);
 #endif
 
 /*
+ * Setting ARCH_SLAB_MINALIGN in arch headers allows a different alignment.
+ * Intended for arches that get misalignment faults even for 64 bit integer
+ * aligned buffers.
+ */
+#ifndef ARCH_SLAB_MINALIGN
+#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
+#endif
+
+/*
+ * kmalloc and friends return ARCH_KMALLOC_MINALIGN aligned
+ * pointers. kmem_cache_alloc and friends return ARCH_SLAB_MINALIGN
+ * aligned pointers.
+ */
+#define __assume_kmalloc_alignment __assume_aligned(ARCH_KMALLOC_MINALIGN)
+#define __assume_slab_alignment __assume_aligned(ARCH_SLAB_MINALIGN)
+#define __assume_page_alignment __assume_aligned(PAGE_SIZE)
+
+/*
  * Kmalloc array related definitions
  */
 
@@ -286,8 +304,8 @@ static __always_inline int kmalloc_index(size_t size)
 }
 #endif /* !CONFIG_SLOB */
 
-void *__kmalloc(size_t size, gfp_t flags);
-void *kmem_cache_alloc(struct kmem_cache *, gfp_t flags);
+void *__kmalloc(size_t size, gfp_t flags) __assume_kmalloc_alignment;
+void *kmem_cache_alloc(struct kmem_cache *, gfp_t flags) __assume_slab_alignment;
 void kmem_cache_free(struct kmem_cache *, void *);
 
 /*
@@ -301,8 +319,8 @@ void kmem_cache_free_bulk(struct kmem_cache *, size_t, void **);
 bool kmem_cache_alloc_bulk(struct kmem_cache *, gfp_t, size_t, void **);
 
 #ifdef CONFIG_NUMA
-void *__kmalloc_node(size_t size, gfp_t flags, int node);
-void *kmem_cache_alloc_node(struct kmem_cache *, gfp_t flags, int node);
+void *__kmalloc_node(size_t size, gfp_t flags, int node) __assume_kmalloc_alignment;
+void *kmem_cache_alloc_node(struct kmem_cache *, gfp_t flags, int node) __assume_slab_alignment;
 #else
 static __always_inline void *__kmalloc_node(size_t size, gfp_t flags, int node)
 {
@@ -316,12 +334,12 @@ static __always_inline void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t f
 #endif
 
 #ifdef CONFIG_TRACING
-extern void *kmem_cache_alloc_trace(struct kmem_cache *, gfp_t, size_t);
+extern void *kmem_cache_alloc_trace(struct kmem_cache *, gfp_t, size_t) __assume_slab_alignment;
 
 #ifdef CONFIG_NUMA
 extern void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
 					   gfp_t gfpflags,
-					   int node, size_t size);
+					   int node, size_t size) __assume_slab_alignment;
 #else
 static __always_inline void *
 kmem_cache_alloc_node_trace(struct kmem_cache *s,
@@ -354,10 +372,10 @@ kmem_cache_alloc_node_trace(struct kmem_cache *s,
 }
 #endif /* CONFIG_TRACING */
 
-extern void *kmalloc_order(size_t size, gfp_t flags, unsigned int order);
+extern void *kmalloc_order(size_t size, gfp_t flags, unsigned int order) __assume_page_alignment;
 
 #ifdef CONFIG_TRACING
-extern void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order);
+extern void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) __assume_page_alignment;
 #else
 static __always_inline void *
 kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
@@ -482,15 +500,6 @@ static __always_inline void *kmalloc_node(size_t size, gfp_t flags, int node)
 	return __kmalloc_node(size, flags, node);
 }
 
-/*
- * Setting ARCH_SLAB_MINALIGN in arch headers allows a different alignment.
- * Intended for arches that get misalignment faults even for 64 bit integer
- * aligned buffers.
- */
-#ifndef ARCH_SLAB_MINALIGN
-#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
-#endif
-
 struct memcg_cache_array {
 	struct rcu_head rcu;
 	struct kmem_cache *entries[0];
-- 
2.1.3

Re: [PATCH 2/2] slab.h: sprinkle __assume_aligned attributes

[Christoph Lameter]

On Fri, 2 Oct 2015, Rasmus Villemoes wrote:

> The various allocators return aligned memory. Telling the compiler
> that allows it to generate better code in many cases, for example when
> the return value is immediately passed to memset().


Looks good.

Acked-by: Christoph Lameter <cl@linux.com>