delay_tsc(): inefficient delay loop (2.6.16-mm1)
Hello all,
I discovered that the delay loop used there (which simply found a new place in 2.6.16-mm1; it existed much longer) is inefficient,
since it does an unnecessary subtraction *in the main loop* which also hits asm code:
00000024 <delay_tsc>:
24: 53 push %ebx 25: 89 c3 mov %eax,%ebx 27: 0f 31 rdtsc 29: 89 c1 mov %eax,%ecx 2b: f3 90 pause 2d: 0f 31 rdtsc 2f: 29 c8 sub %ecx,%eax 31: 39 d8 cmp %ebx,%eax 33: 72 f6 jb 2b <delay_tsc+0x7> 35: 5b pop %ebx 36: c3 ret
With such a patch:
arch/i386/lib/delay.c renamed from arch/i386/lib/delay.c.orig +3 -1
@@ -44,10 +44,12 @@unsigned long bclock, now;rdtscl (bclock); +/* offset with bclock to have very simple comparison below */ + loops += bclock;do {rep_nop ();rdtscl (now); - }while ((now-bclock) < loops); + }while (now < loops); }/*
the result is:
00000024 <delay_tsc>:
24: 89 c1 mov %eax,%ecx 26: 0f 31 rdtsc 28: 01 c1 add %eax,%ecx 2a: f3 90 pause 2c: 0f 31 rdtsc 2e: 39 c8 cmp %ecx,%eax 30: 72 f8 jb 2a <delay_tsc+0x6> 32: c3 ret
Improvement: no unnecessary stuff after having hit the timer target value (read: faster), better power saving, 4 bytes less opcodes.
The patch above could be considered weird since it fiddles with "loops" directly. A possibly cleaner way would be to introduce a new variable end = bclock + loops.
Comments? Anything that I'm missing here as to why it hasn't been done that way before?
If this holds water then I'll submit a final patch soon.
Thanks!
Andreas Mohr
Quoted 1 line · Andreas Mohr
+ loops += bclock;
[...]
arch/i386/lib/delay.c line 52
Quoted 2 lines
arch/i386/lib/delay.c renamed from arch/i386/lib/delay.c.orig +1
-1- }while ((now-bclock) < loops); + }while (now < loops);
Erm, aren't you introducing an overflow problem here?
if loops is 2^32-1, bclock is 1, the old version would execute the proper number of times, the new one will blow out in one tick.
Hi,
Quoted 10 lines · Ray Lee
Quoted 1 line · Andreas Mohr
+ loops += bclock;
[...]
arch/i386/lib/delay.c line 52
Quoted 2 lines
arch/i386/lib/delay.c renamed from arch/i386/lib/delay.c.orig +1
-1- }while ((now-bclock) < loops); + }while (now < loops);Erm, aren't you introducing an overflow problem here?
if loops is 2^32-1, bclock is 1, the old version would execute the proper number of times, the new one will blow out in one tick.
Doh. That's what happens if you get too excited about some new trick... Back to the drawing board, methinks.
Andreas Mohr
Signature · 2 lines
No programming skills!? Why not help translate many Linux applications! https://launchpad.ubuntu.com/rosetta
arch/i386/lib/delay.c line 53
Quoted 10 lines · Andreas Mohr
arch/i386/lib/delay.c renamed from arch/i386/lib/delay.c.orig +3
-1rdtscl (bclock); +/* offset with bclock to have very simple comparison below */ + loops += bclock;do {rep_nop ();rdtscl (now); - }while ((now-bclock) < loops); + }while (now < loops); }
Hehe, optimizing delay loops *g* But your optimization is wrong. 'loops+bclock' and/or 'now' is likely to wrap around and then the test condition becomes bogus.
Ciao, ET.