[PATCH 1/6] powerpc: Speed up clear_page by unrolling it
From: Anton Blanchard <hidden>
Date: 2014-09-17 04:39:41
Subsystem:
linux for powerpc (32-bit and 64-bit), the rest · Maintainers:
Madhavan Srinivasan, Linus Torvalds
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 and
page sizes wont drop below 1kB, which is unlikely, but we could
add a runtime check during early init if it makes people nervous.
Signed-off-by: Anton Blanchard <redacted>
---
arch/powerpc/include/asm/page_64.h | 18 ++++++++++++++----
1 file changed, 14 insertions(+), 4 deletions(-)
diff --git a/arch/powerpc/include/asm/page_64.h b/arch/powerpc/include/asm/page_64.h
index d0d6afb..b6e78cb 100644
--- a/arch/powerpc/include/asm/page_64.h
+++ b/arch/powerpc/include/asm/page_64.h@@ -51,11 +51,21 @@ static __inline__ void clear_page(void *addr) __asm__ __volatile__( "mtctr %1 # clear_page\n\ -1: dcbz 0,%0\n\ - add %0,%0,%3\n\ + .balign 16\n\ +1: dcbz 0,%0\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" - : "=r" (addr) - : "r" (lines), "0" (addr), "r" (line_size) + : "=&r" (addr) + : "r" (lines/8), "0" (addr), "b" (line_size), "b" (line_size*2), + "b" (line_size*3), "b" (line_size*4), "b" (line_size*5), + "b" (line_size*6), "b" (line_size*7), "r" (line_size*8) : "ctr", "memory"); }
--
1.9.1