On Fri, 06 Jun 2025 11:44:02 +0200, Christophe Leroy wrote: > > At the time being recalculate_boundary() is implemented with a > loop which shows up as costly in a perf profile, as depicted by > the annotate below: > > 0.00 : c057e934: 3d 40 7f ff lis r10,32767 > 0.03 : c057e938: 61 4a ff ff ori r10,r10,65535 > 0.21 : c057e93c: 7d 49 50 50 subf r10,r9,r10 > 5.39 : c057e940: 7d 3c 4b 78 mr r28,r9 > 2.11 : c057e944: 55 29 08 3c slwi r9,r9,1 > 3.04 : c057e948: 7c 09 50 40 cmplw r9,r10 > 2.47 : c057e94c: 40 81 ff f4 ble c057e940 > <snd_pcm_ioctl+0xee0> > > Total: 13.2% on that simple loop. > > But what the loop does is to multiply the boundary by 2 until it is > over the wanted border. This can be avoided by using fls() to get the > boundary value order and shift it by the appropriate number of bits at > once. > > This change provides the following profile: > > 0.04 : c057f6e8: 3d 20 7f ff lis r9,32767 > 0.02 : c057f6ec: 61 29 ff ff ori r9,r9,65535 > 0.34 : c057f6f0: 7d 5a 48 50 subf r10,r26,r9 > 0.23 : c057f6f4: 7c 1a 50 40 cmplw r26,r10 > 0.02 : c057f6f8: 41 81 00 20 bgt c057f718 > <snd_pcm_ioctl+0xf08> > 0.26 : c057f6fc: 7f 47 00 34 cntlzw r7,r26 > 0.09 : c057f700: 7d 48 00 34 cntlzw r8,r10 > 0.22 : c057f704: 7d 08 38 50 subf r8,r8,r7 > 0.04 : c057f708: 7f 5a 40 30 slw r26,r26,r8 > 0.35 : c057f70c: 7c 0a d0 40 cmplw r10,r26 > 0.13 : c057f710: 40 80 05 f8 bge c057fd08 > <snd_pcm_ioctl+0x14f8> > 0.00 : c057f714: 57 5a f8 7e srwi r26,r26,1 > > Total: 1.7% with that loopless alternative. > > Signed-off-by: Christophe Leroy <christophe.le...@csgroup.eu>
Applied now to for-next branch. Thanks. Takashi
