This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Replace divide and multiply with mask in sYSTRIm
- From: Anton Blanchard <anton at samba dot org>
- To: libc-alpha at sourceware dot org
- Date: Mon, 26 Jul 2010 17:21:09 +1000
- Subject: Re: [PATCH] Replace divide and multiply with mask in sYSTRIm
- References: <20100726043651.GE29316@kryten>
> When profiling a simple malloc/free testcase, sYSTRIm was found to
> have a divide in it. Looking closer we are using a divide and multiply
> to round something to mp_.pagesize bytes. Since this is always a
> power of two and we have assertions in a number of places to enforce
> this, replace with a mask.
>
> A simple testcase of malloc/free of 513 bytes on a POWER7 box
> shows a decent improvement:
>
> baseline: 6869000 ops/sec
> patched: 8174000 ops/sec
>
> Almost 19% faster
>
> Anton
I missed the arena version, found when running the same test multithreaded.
Updated patch below.
Anton
--
2010-07-26 Anton Blanchard <anton@samba.org>
* malloc/malloc.c (sYSTRIm): Replace divide and multiply with mask.
* malloc/arena.c (heap_trim): Replace divide and multiply with mask.
Index: glibc/malloc/malloc.c
===================================================================
--- glibc.orig/malloc/malloc.c 2010-07-26 13:42:35.470741337 +1000
+++ glibc/malloc/malloc.c 2010-07-26 13:58:29.870741493 +1000
@@ -3466,7 +3466,7 @@ static int sYSTRIm(pad, av) size_t pad;
top_size = chunksize(av->top);
/* Release in pagesize units, keeping at least one page */
- extra = ((top_size - pad - MINSIZE + (pagesz-1)) / pagesz - 1) * pagesz;
+ extra = (top_size - pad - MINSIZE + (pagesz-1)) & ~(pagesz-1);
if (extra > 0) {
Index: glibc/malloc/arena.c
===================================================================
--- glibc.orig/malloc/arena.c 2010-07-26 17:05:22.311992028 +1000
+++ glibc/malloc/arena.c 2010-07-26 17:06:50.400741083 +1000
@@ -879,7 +879,7 @@ heap_trim(heap, pad) heap_info *heap; si
/*check_chunk(ar_ptr, top_chunk);*/
}
top_size = chunksize(top_chunk);
- extra = ((top_size - pad - MINSIZE + (pagesz-1))/pagesz - 1) * pagesz;
+ extra = (top_size - pad - MINSIZE + (pagesz-1)) & ~(pagesz-1);
if(extra < (long)pagesz)
return 0;
/* Try to shrink. */