reimplement string with shared_ptr instead of concrete macro

[jingkaimori]

Motivation

Concrete struct (defined as CONCRETE and *_rep) is used as reference counter. For example, type T is implemented as two class: T_rep handles properties of given type, and T handles reference and counter of references to T_rep. When T is assigned, only reference is assigned with counter in/decreased.

std::shared_ptr plays a similar role like concrete struct, and maintain a spin lock on reference count, thus avoid data-race in parallel execution environments. However, the spin lock drags performance down when code is executed in a single thread serially.

Works

• rewrite string with shared_ptr
• add standard allocator for shared_ptr

Performance

Table below shows performance measured by string_bench.

Before

ns/op	op/s	err%	total	benchmark
83.29	12,006,624.83	0.7%	0.01	construct string
7.61	131,490,532.01	2.7%	0.01	equality of string
22.99	43,492,682.93	0.5%	0.01	compare string
70.59	14,165,431.64	4.5%	0.01	slice string
120.95	8,268,132.84	0.2%	0.06	concat string
39.24	25,481,767.96	4.2%	0.03	append string
12.55	79,671,371.62	3.5%	0.01	is quoted

After

ns/op	op/s	err%	total	benchmark
192.94	5,183,072.68	3.7%	0.01	construct string
21.47	46,582,660.07	0.9%	0.01	equality of string
59.28	16,868,608.16	2.1%	0.01	compare string
130.40	7,668,699.19	1.8%	0.01	slice string
214.34	4,665,550.51	1.8%	0.10	concat string
68.27	14,648,365.31	4.9%	0.04	append string
30.65	32,629,029.70	1.3%	0.02	is quoted

[da-liii]

@PikachuHy Could you spare some time to help us review this pr?

[jingkaimori]

Because of unnecessary spin lock, std::shared_ptr will not be used in lolly, see #326