Commit e35735b9 authored by Anton Blanchard's avatar Anton Blanchard Committed by Michael Ellerman

powerpc: Speed up clear_page by unrolling it

Unroll clear_page 8 times. A simple microbenchmark which
allocates and frees a zeroed page:

for (i = 0; i < iterations; i++) {
	unsigned long p = __get_free_page(GFP_KERNEL | __GFP_ZERO);
	free_page(p);
}

improves 20% on POWER8.

This assumes cacheline sizes won't grow beyond 512 bytes or
page sizes wont drop below 1kB, which is unlikely, but we could
add a runtime check during early init if it makes people nervous.

Michael found that some versions of gcc produce quite bad code
(all multiplies), so we give gcc a hand by using shifts and adds.
Signed-off-by: default avatarAnton Blanchard <anton@samba.org>
Signed-off-by: default avatarMichael Ellerman <mpe@ellerman.id.au>
parent 2013add4
...@@ -42,20 +42,40 @@ ...@@ -42,20 +42,40 @@
typedef unsigned long pte_basic_t; typedef unsigned long pte_basic_t;
static __inline__ void clear_page(void *addr) static inline void clear_page(void *addr)
{ {
unsigned long lines, line_size; unsigned long iterations;
unsigned long onex, twox, fourx, eightx;
line_size = ppc64_caches.dline_size; iterations = ppc64_caches.dlines_per_page / 8;
lines = ppc64_caches.dlines_per_page;
__asm__ __volatile__( /*
* Some verisions of gcc use multiply instructions to
* calculate the offsets so lets give it a hand to
* do better.
*/
onex = ppc64_caches.dline_size;
twox = onex << 1;
fourx = onex << 2;
eightx = onex << 3;
asm volatile(
"mtctr %1 # clear_page\n\ "mtctr %1 # clear_page\n\
.balign 16\n\
1: dcbz 0,%0\n\ 1: dcbz 0,%0\n\
add %0,%0,%3\n\ dcbz %3,%0\n\
dcbz %4,%0\n\
dcbz %5,%0\n\
dcbz %6,%0\n\
dcbz %7,%0\n\
dcbz %8,%0\n\
dcbz %9,%0\n\
add %0,%0,%10\n\
bdnz+ 1b" bdnz+ 1b"
: "=r" (addr) : "=&r" (addr)
: "r" (lines), "0" (addr), "r" (line_size) : "r" (iterations), "0" (addr), "b" (onex), "b" (twox),
"b" (twox+onex), "b" (fourx), "b" (fourx+onex),
"b" (twox+fourx), "b" (eightx-onex), "r" (eightx)
: "ctr", "memory"); : "ctr", "memory");
} }
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment