summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorChristophe Leroy <christophe.leroy@csgroup.eu>2022-03-24 02:05:44 +0300
committerLinus Torvalds <torvalds@linux-foundation.org>2022-03-24 05:00:33 +0300
commitf334f5668bedf7307f6df1d98b14f55902931926 (patch)
treeb4139b928db3e7084e0fdc97b3940db67853153a
parent14e83077d55ff4b88fe39f5e98fb8230c2ccb4fb (diff)
downloadlinux-f334f5668bedf7307f6df1d98b14f55902931926.tar.xz
ilog2: force inlining of __ilog2_u32() and __ilog2_u64()
Building a kernel with CONFIG_CC_OPTIMISE_FOR_SIZE leads to __ilog2_u32() being duplicated 50 times and __ilog2_u64() 3 times in vmlinux on a tiny powerpc32 config. __ilog2_u32() being 2 instructions it is not worth being kept out of line, so force inlining. Allthough the u64 version is a bit bigger, there is still a small benefit in keeping it inlined. On a 64 bits config there's a real benefit. With this change the size of vmlinux text is reduced by 1 kbytes, which is approx 50% more than the size of the removed functions. Before the patch there is for instance: c00d2a94 <__ilog2_u32>: c00d2a94: 7c 63 00 34 cntlzw r3,r3 c00d2a98: 20 63 00 1f subfic r3,r3,31 c00d2a9c: 4e 80 00 20 blr c00d36d8 <__order_base_2>: c00d36d8: 28 03 00 01 cmplwi r3,1 c00d36dc: 40 81 00 2c ble c00d3708 <__order_base_2+0x30> c00d36e0: 94 21 ff f0 stwu r1,-16(r1) c00d36e4: 7c 08 02 a6 mflr r0 c00d36e8: 38 63 ff ff addi r3,r3,-1 c00d36ec: 90 01 00 14 stw r0,20(r1) c00d36f0: 4b ff f3 a5 bl c00d2a94 <__ilog2_u32> c00d36f4: 80 01 00 14 lwz r0,20(r1) c00d36f8: 38 63 00 01 addi r3,r3,1 c00d36fc: 7c 08 03 a6 mtlr r0 c00d3700: 38 21 00 10 addi r1,r1,16 c00d3704: 4e 80 00 20 blr c00d3708: 38 60 00 00 li r3,0 c00d370c: 4e 80 00 20 blr With the patch it has become: c00d356c <__order_base_2>: c00d356c: 28 03 00 01 cmplwi r3,1 c00d3570: 40 81 00 14 ble c00d3584 <__order_base_2+0x18> c00d3574: 38 63 ff ff addi r3,r3,-1 c00d3578: 7c 63 00 34 cntlzw r3,r3 c00d357c: 20 63 00 20 subfic r3,r3,32 c00d3580: 4e 80 00 20 blr c00d3584: 38 60 00 00 li r3,0 c00d3588: 4e 80 00 20 blr No more need for __order_base_2() to setup a stack frame and save/restore caller address. And the following 'add 1' is merged in the subtract. Another typical use of it: c080ff28 <hugepagesz_setup>: ... c080fff8: 7f c3 f3 78 mr r3,r30 c080fffc: 4b 8f 81 f1 bl c01081ec <__ilog2_u32> c0810000: 38 63 ff f2 addi r3,r3,-14 ... Becomes c080ff1c <hugepagesz_setup>: ... c080ffec: 7f c3 00 34 cntlzw r3,r30 c080fff0: 20 63 00 11 subfic r3,r3,17 ... Here no need to move r30 argument to r3 then substract 14 to result. Just work on r30 and merge the 'sub 14' with the 'sub from 31'. Link: https://lkml.kernel.org/r/803a2ac3d923ebcfd0dd40f5886b05cae7bb0aba.1644243860.git.christophe.leroy@csgroup.eu Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
-rw-r--r--include/linux/log2.h4
1 files changed, 2 insertions, 2 deletions
diff --git a/include/linux/log2.h b/include/linux/log2.h
index df0b155c2141..9f30d087a128 100644
--- a/include/linux/log2.h
+++ b/include/linux/log2.h
@@ -18,7 +18,7 @@
* - the arch is not required to handle n==0 if implementing the fallback
*/
#ifndef CONFIG_ARCH_HAS_ILOG2_U32
-static inline __attribute__((const))
+static __always_inline __attribute__((const))
int __ilog2_u32(u32 n)
{
return fls(n) - 1;
@@ -26,7 +26,7 @@ int __ilog2_u32(u32 n)
#endif
#ifndef CONFIG_ARCH_HAS_ILOG2_U64
-static inline __attribute__((const))
+static __always_inline __attribute__((const))
int __ilog2_u64(u64 n)
{
return fls64(n) - 1;