In an early version of RapidJSON, [an issue](https://code.google.com/archive/p/rapidjson/issues/104) reported that the `SkipWhitespace_SIMD()` causes crash very rarely (around 1 in 500,000). After investigation, it is suspected that `_mm_loadu_si128()` accessed bytes after `'\0'`, and across a protected page boundary.
> To support algorithms requiring unaligned 128-bit SIMD memory accesses, memory buffer allocation by a caller function should consider adding some pad space so that a callee function can safely use the address pointer safely with unaligned 128-bit SIMD memory operations.
To fix this issue, currently the routine process bytes up to the next aligned address. After tha, use aligned read to perform SIMD processing. Also see [#85](https://github.com/miloyip/rapidjson/issues/85).
During optimization, it is found that some compilers cannot localize some member data access of streams into local variables or registers. Experimental results show that for some stream types, making a copy of the stream and used it in inner-loop can improve performance. For example, the actual (non-SIMD) implementation of `SkipWhitespace()` is implemented as:
Depending on the traits of stream, `StreamLocalCopy` will make (or not make) a copy of the stream object, use it locally and copy the states of stream back to the original stream.
Parsing string into `double` is difficult. The standard library function `strtod()` can do the job but it is slow. By default, the parsers use normal precision setting. This has has maximum 3 [ULP](http://en.wikipedia.org/wiki/Unit_in_the_last_place) error and implemented in `internal::StrtodNormalPrecision()`.
When using `kParseFullPrecisionFlag`, the parsers calls `internal::StrtodFullPrecision()` instead, and this function actually implemented 3 versions of conversion methods.
The naive algorithm for integer-to-string conversion involves division per each decimal digit. We have implemented various implementations and evaluated them in [itoa-benchmark](https://github.com/miloyip/itoa-benchmark).
Although SSE2 version is the fastest but the difference is minor by comparing to the first running-up `branchlut`. And `branchlut` is pure C++ implementation so we adopt `branchlut` in RapidJSON.
Originally RapidJSON uses `snprintf(..., ..., "%g")` to achieve double-to-string conversion. This is not accurate as the default precision is 6. Later we also find that this is slow and there is an alternative.
Google 的 V8 [double-conversion](https://github.com/floitsch/double-conversion
)implemented a newer, fast algorithm called Grisu3 (Loitsch, Florian. "Printing floating-point numbers quickly and accurately with integers." ACM Sigplan Notices 45.6 (2010): 233-243.).
)实现了更新的、快速的被称为 Grisu3 的算法（Loitsch, Florian. "Printing floating-point numbers quickly and accurately with integers." ACM Sigplan Notices 45.6 (2010): 233-243.）。
However, since it is not header-only so that we implemented a header-only version of Grisu2. This algorithm guarantees that the result is always accurate. And in most of cases it produces the shortest (optimal) string representation.
