times.txt

s_1 - stdlib hashmap 
real	5m38.348s
user	5m28.129s
sys	0m6.525s

s_2 - int math
real	3m47.678s
user	3m40.952s
sys	0m5.362s

s_2.1 better itoc and ctoi + inline
real	3m14.754s
user	3m7.526s
sys	0m5.771s

s_3 branch avoidance
real	1m32.719s
user	1m25.706s
sys	0m5.805s

ANOMALY: 
avoiding branching in int_to_char and char_to_int helped a lot. 
avoiding branching in padding function didn't help at all. Must investigate.
Distribution:
count: 1000000000

negative: 87780411 ->    8.8%   // ex. -99

point1: 178191570  ->   17.8%   // ex.   9.9
point2: 801638107  ->   80.1%   // ex.  99.9
no point1 0                     // ex.   9
no point2 20170323 ->    2.0%   // ex.  99

Thesis: having a shorter calculation for 80% of the time migth be worth the
occasional pipeline flush. -> Goal should be to optimize for point2 case.
Branch prediction should do the rest. __builtin_expect?