Re: [RFC, PATCH] Avoid manual shift-and-test logic in AllocSetFreeIndex

Hi,

> Also, are you still seeing the same improvement with the __builtin_clz
> as your inline asm implementation?

In my benchmark program, it is a little different performance
in fls implementation and inline asm implementation.
However, the result of a pgbench is almost the same improvement.

Here is the result of my benchmark.

Xeon(Core architecture)
bytes : 4 8 16 32 64 128 256 512 1024 mix
original : 0.780 0.790 0.820 0.870 0.930 0.980 1.040 1.080 1.140 0.910
inline asm: 0.320 0.180 0.190 0.180 0.190 0.180 0.190 0.180 0.190 0.170
fls : 0.270 0.260 0.290 0.290 0.290 0.290 0.290 0.300 0.290 0.380

Xeon(P4 architecrure)
bytes : 4 8 16 32 64 128 256 512 1024 mix
original : 0.520 0.520 0.670 0.780 0.950 1.000 1.060 1.190 1.250 0.940
inline asm: 0.610 0.530 0.530 0.520 0.520 0.540 0.540 0.580 0.540 0.600
fls : 0.390 0.370 0.780 0.780 0.780 0.790 0.780 0.780 0.780 0.520

pgbench result (measured by oprofile)

CPU: Xeon(P4 architecrure)
test program: pgbench -c 1 -t 50000 (fsync=off)

original
samples % symbol name
66854 6.6725 AllocSetAlloc
11817 1.1794 AllocSetFree

inline asm
samples % symbol name
47610 4.9333 AllocSetAlloc
6248 0.6474 AllocSetFree

fls
samples % symbol name
48779 4.9954 AllocSetAlloc
7648 0.7832 AllocSetFree

Best regards,

---
Atsushi Ogawa