| Message ID | 20190821074200.2203-1-efremov@ispras.ru (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | lib/memweight.c: optimize by inlining bitmap_weight() | expand |
On Wed, 21 Aug 2019 10:42:00 +0300 Denis Efremov <efremov@ispras.ru> wrote: > This patch inlines bitmap_weight() call. It is better to say the patch "open codes" the bitmap_weight() call. > Thus, removing the BUG_ON, Why is that OK to do? I expect all the code size improvements are from doing this? > and 'longs to bits -> bits to longs' conversion by directly calling > hweight_long(). > > ./scripts/bloat-o-meter lib/memweight.o.old lib/memweight.o.new > add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-10 (-10) > Function old new delta > memweight 162 152 -10 >
On 22.08.2019 04:25, Andrew Morton wrote: > On Wed, 21 Aug 2019 10:42:00 +0300 Denis Efremov <efremov@ispras.ru> wrote: > >> This patch inlines bitmap_weight() call. > > It is better to say the patch "open codes" the bitmap_weight() call. > >> Thus, removing the BUG_ON, > > Why is that OK to do? BUG_ON was necessary here to check that bitmap_weight will return a correct value, i.e. the computed weight will fit the int type: static __always_inline int bitmap_weight(const unsigned long *src, unsigned int nbits); BUG_ON was added in the memweight v2 https://lore.kernel.org/lkml/20120523092113.GG10452@quack.suse.cz/ Jan Kara wrote: >> + >> + for (longs = bytes / sizeof(long); longs > 0; ) { >> + size_t bits = min_t(size_t, INT_MAX & ~(BITS_PER_LONG - 1), > + longs * BITS_PER_LONG); > I find it highly unlikely that someone would have such a large bitmap > (256 MB or more on 32-bit). Also the condition as you wrote it can just > overflow so it won't have the desired effect. Just do > BUG_ON(longs >= ULONG_MAX / BITS_PER_LONG); > and remove the loop completely. If someone comes with such a huge bitmap, > the code can be modified easily (after really closely inspecting whether > such a huge bitmap is really well justified). >> + >> + w += bitmap_weight(bitmap.ptr, bits); >> + bytes -= bits / BITS_PER_BYTE; >> + bitmap.address += bits / BITS_PER_BYTE; >> + longs -= bits / BITS_PER_LONG; Akinobu Mita wrote: > The bits argument of bitmap_weight() is int type. So this should be > > BUG_ON(longs >= INT_MAX / BITS_PER_LONG); We don't need this check, since we removed the bitmap_weight call and control the computation directly with size_t everywhere. We could add BUG_ON(bytes >= SIZE_MAX / BITS_PER_BYTE); at the very beginning of the function to check that the array is not very big (>2000PiB), but it seems excessive. > > I expect all the code size improvements are from doing this? Yes, but I thought it's good to show that the total size is not increasing because of the manual "inlining". > >> and 'longs to bits -> bits to longs' conversion by directly calling >> hweight_long(). >> >> ./scripts/bloat-o-meter lib/memweight.o.old lib/memweight.o.new >> add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-10 (-10) >> Function old new delta >> memweight 162 152 -10 >> > Regards, Denis
diff --git a/lib/memweight.c b/lib/memweight.c index 94dd72ccaa7f..f050b2b4c5e2 100644 --- a/lib/memweight.c +++ b/lib/memweight.c @@ -20,11 +20,13 @@ size_t memweight(const void *ptr, size_t bytes) longs = bytes / sizeof(long); if (longs) { - BUG_ON(longs >= INT_MAX / BITS_PER_LONG); - ret += bitmap_weight((unsigned long *)bitmap, - longs * BITS_PER_LONG); + const unsigned long *bitmap_long = + (const unsigned long *)bitmap; + bytes -= longs * sizeof(long); - bitmap += longs * sizeof(long); + for (; longs > 0; longs--, bitmap_long++) + ret += hweight_long(*bitmap_long); + bitmap = (const unsigned char *)bitmap_long; } /* * The reason that this last loop is distinct from the preceding