Skip to contentSkip to conversation

__cacheline_aligned always in own section

PATCH In review Rusty Russell 5 participants 5 messages Feb 26, 2004 – Feb 27, 2004

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 __cacheline_aligned always in own section

Name: Always Put Cache Aligned Code in Own Section: Even Modules
Status: Tested on 2.6.3-bk7

We put ____cacheline_aligned things in their own section, simply because we waste less space that way. Otherwise we end up padding innocent variables to the next cacheline to get the required alignment.

There's no reason not to do this in modules, too.

include/linux/cache.h +0 -4
diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/current-dontdiff --minimal .16274-linux-2.6.3-bk7/include/linux/cache.h .16274-linux-2.6.3-bk7.updated/include/linux/cache.h
@@ -26,13 +26,9 @@
 #endif
 
 #ifndef __cacheline_aligned
-#ifdef MODULE
-#define __cacheline_aligned ____cacheline_aligned
-#else
 #define __cacheline_aligned					\
   __attribute__((__aligned__(SMP_CACHE_BYTES),			\
 		 __section__(".data.cacheline_aligned")))
-#endif
 #endif /* __cacheline_aligned */
 
 #ifndef __cacheline_aligned_in_smp

--
Anyone who quotes me in their sig is an idiot. -- Rusty Russell.

#1 lore raw
Quoted 9 lines · Rusty Russell

Name: Always Put Cache Aligned Code in Own Section: Even Modules
Status: Tested on 2.6.3-bk7

We put ____cacheline_aligned things in their own section, simply because we waste less space that way. Otherwise we end up padding innocent variables to the next cacheline to get the required alignment.

There's no reason not to do this in modules, too.

[snip]

while we are at it, arm-26, ppc, sparc, sparc64 and sh have per-arch definitions of __cacheline_aligned that are identical to default. And yes, removal is safe - all users of __cacheline_aligned actually pull linux/cache.h in.

include/asm-arm26/cache.h +0 -8
diff -urN RC3-bk1/include/asm-arm26/cache.h RC3-bk1-current/include/asm-arm26/cache.h
@@ -8,12 +8,4 @@
 #define        L1_CACHE_ALIGN(x)       (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
 #define        SMP_CACHE_BYTES L1_CACHE_BYTES
 
-#ifdef MODULE
-#define __cacheline_aligned __attribute__((__aligned__(L1_CACHE_BYTES)))
-#else
-#define __cacheline_aligned					\
-  __attribute__((__aligned__(L1_CACHE_BYTES),			\
-		 __section__(".data.cacheline_aligned")))
-#endif
-
 #endif
include/asm-ppc/cache.h +0 -8
diff -urN RC3-bk1/include/asm-ppc/cache.h RC3-bk1-current/include/asm-ppc/cache.h
@@ -30,14 +30,6 @@
 #define	L1_CACHE_ALIGN(x)       (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
 #define	L1_CACHE_PAGES		8
 
-#ifdef MODULE
-#define __cacheline_aligned __attribute__((__aligned__(L1_CACHE_BYTES)))
-#else
-#define __cacheline_aligned					\
-  __attribute__((__aligned__(L1_CACHE_BYTES),			\
-		 __section__(".data.cacheline_aligned")))
-#endif
-
 #ifndef __ASSEMBLY__
 extern void clean_dcache_range(unsigned long start, unsigned long stop);
 extern void flush_dcache_range(unsigned long start, unsigned long stop);
include/asm-sh/cache.h +0 -8
diff -urN RC3-bk1/include/asm-sh/cache.h RC3-bk1-current/include/asm-sh/cache.h
@@ -21,14 +21,6 @@
 
 #define L1_CACHE_ALIGN(x)	(((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
 
-#ifdef MODULE
-#define __cacheline_aligned __attribute__((__aligned__(L1_CACHE_BYTES)))
-#else
-#define __cacheline_aligned					\
-  __attribute__((__aligned__(L1_CACHE_BYTES),			\
-		 __section__(".data.cacheline_aligned")))
-#endif
-
 #define L1_CACHE_SHIFT_MAX 	5	/* largest L1 which this arch supports */
 
 struct cache_info {
include/asm-sparc/cache.h +0 -8
diff -urN RC3-bk1/include/asm-sparc/cache.h RC3-bk1-current/include/asm-sparc/cache.h
@@ -17,14 +17,6 @@
 
 #define SMP_CACHE_BYTES 32
 
-#ifdef MODULE
-#define __cacheline_aligned __attribute__((__aligned__(SMP_CACHE_BYTES)))
-#else
-#define __cacheline_aligned					\
-  __attribute__((__aligned__(SMP_CACHE_BYTES),			\
-		 __section__(".data.cacheline_aligned")))
-#endif
-
 /* Direct access to the instruction cache is provided through and
  * alternate address space.  The IDC bit must be off in the ICCR on
  * HyperSparcs for these accesses to work.  The code below does not do
include/asm-sparc64/cache.h +0 -8
diff -urN RC3-bk1/include/asm-sparc64/cache.h RC3-bk1-current/include/asm-sparc64/cache.h
@@ -14,12 +14,4 @@
 #define        SMP_CACHE_BYTES_SHIFT	6
 #define        SMP_CACHE_BYTES		(1 << SMP_CACHE_BYTES_SHIFT) /* L2 cache line size. */
 
-#ifdef MODULE
-#define __cacheline_aligned __attribute__((__aligned__(SMP_CACHE_BYTES)))
-#else
-#define __cacheline_aligned					\
-  __attribute__((__aligned__(SMP_CACHE_BYTES),			\
-		 __section__(".data.cacheline_aligned")))
-#endif
-
 #endif
#2 lore raw
In reply to Rusty Russell
Quoted 9 lines · Rusty Russell

Name: Always Put Cache Aligned Code in Own Section: Even Modules
Status: Tested on 2.6.3-bk7

We put ____cacheline_aligned things in their own section, simply because we waste less space that way. Otherwise we end up padding innocent variables to the next cacheline to get the required alignment.

There's no reason not to do this in modules, too.

On a related matter,

I compile my kernels for 486 but buils system aligns functions and labels to 16 bytes, with results like this:

00000730 <islpci_eth_tx_timeout>:

     730:       55                      push   %ebp
     731:       89 e5                   mov    %esp,%ebp
     733:       8b 45 08                mov    0x8(%ebp),%eax
     736:       8b 40 64                mov    0x64(%eax),%eax
     739:       05 14 03 00 00          add    $0x314,%eax
     73e:       ff 40 14                incl   0x14(%eax)
     741:       5d                      pop    %ebp
     742:       c3                      ret
     743:       90                      nop
     744:       90                      nop
     745:       90                      nop
     746:       90                      nop
     747:       90                      nop
     748:       90                      nop
     749:       90                      nop
     74a:       90                      nop
     74b:       90                      nop
     74c:       90                      nop
     74d:       90                      nop
     74e:       90                      nop
     74f:       90                      nop

Losing on average 15/2 bytes to alignment, my kernel lose

# echo $((`cat System.map | grep '0 ' | wc -l`*15/2))
149632
bytes only due to function alignment, not counting jump target
alighment.

Is there any way to prevent this?
--
vda

#3 lore raw
Quoted 27 lines · Denis Vlasenko

I compile my kernels for 486 but buils system aligns functions and labels to 16 bytes, with results like this:

00000730 <islpci_eth_tx_timeout>:

     730:       55                      push   %ebp
     731:       89 e5                   mov    %esp,%ebp
     733:       8b 45 08                mov    0x8(%ebp),%eax
     736:       8b 40 64                mov    0x64(%eax),%eax
     739:       05 14 03 00 00          add    $0x314,%eax
     73e:       ff 40 14                incl   0x14(%eax)
     741:       5d                      pop    %ebp
     742:       c3                      ret
     743:       90                      nop
     744:       90                      nop
     745:       90                      nop
     746:       90                      nop
     747:       90                      nop
     748:       90                      nop
     749:       90                      nop
     74a:       90                      nop
     74b:       90                      nop
     74c:       90                      nop
     74d:       90                      nop
     74e:       90                      nop
     74f:       90                      nop
Quoted 7 lines

Losing on average 15/2 bytes to alignment, my kernel lose

# echo $((`cat System.map | grep '0 ' | wc -l`*15/2))
149632
bytes only due to function alignment, not counting jump target
alighment.

Is there any way to prevent this?

Yes, there are ways of turning off a lot of this alignment padding and it makes a significant different in code size. You need to dig around in the gcc documentation for the several -*align* options.

IIRC, not all of the padding could be turned off (the gcc options were not complete) but it's a couple of years since I investigated this.

#4 lore raw
Quoted 34 lines · Andrew Morton
Quoted 33 lines · Denis Vlasenko

I compile my kernels for 486 but buils system aligns functions and labels to 16 bytes, with results like this:

00000730 <islpci_eth_tx_timeout>:

     730:       55                      push   %ebp
     731:       89 e5                   mov    %esp,%ebp
     733:       8b 45 08                mov    0x8(%ebp),%eax
     736:       8b 40 64                mov    0x64(%eax),%eax
     739:       05 14 03 00 00          add    $0x314,%eax
     73e:       ff 40 14                incl   0x14(%eax)
     741:       5d                      pop    %ebp
     742:       c3                      ret
     743:       90                      nop
     744:       90                      nop
     745:       90                      nop
     746:       90                      nop
     747:       90                      nop
     748:       90                      nop
     749:       90                      nop
     74a:       90                      nop
     74b:       90                      nop
     74c:       90                      nop
     74d:       90                      nop
     74e:       90                      nop
     74f:       90                      nop

Losing on average 15/2 bytes to alignment, my kernel lose

# echo $((`cat System.map | grep '0 ' | wc -l`*15/2))
149632
bytes only due to function alignment, not counting jump target
alighment.
Quoted 8 lines
Quoted 1 line

Is there any way to prevent this?

Yes, there are ways of turning off a lot of this alignment padding and it makes a significant different in code size. You need to dig around in the gcc documentation for the several -*align* options.

IIRC, not all of the padding could be turned off (the gcc options were not complete) but it's a couple of years since I investigated this.

Denis, there's a patch in -tiny to let you experiment with overriding CFLAGS. See http://selenic.com/tiny-about/

You probably want something like:

-falign-functions=1 -falign-jumps=1 -falign-labels=1 -falign-loops=1

Signature · 1 line
Matt Mackall : http://www.selenic.com : Linux development and consulting