The plain text parser allocates a 16 KB buffer for each parsed text

PlainTextStreamParser wraps each source it parses in a BufferedReader, which allocates a 16 KB buffer, even when the source is already in memory, like a StringReader. Allocating this buffer costs more than parsing a short text.

This matters for code that parses many short texts:

• The code macro parses the text of each token of the highlighted code: highlighting a short snippet allocates several megabytes, which slows down the rendering of pages with many code macros, especially when many of them are rendered at the same time.
• Translation messages with parameters are formatted and parsed again each time they are rendered (translations without parameters are parsed once and then cached).

The buffer is only useful for a source that is costly to read character by character, so the parser shouldn't add one to a source that is already in memory or already buffered.

Measurements

Parsing a translation message with parameters like MessageFormatTranslationMessage does (formatting a 60-character message, then parsing it into an XDOM), on an 8-core laptop, in two runs:

Without the buffer, the parsing is about 1.9 times faster on 1 thread and 4 times faster on 8 threads: the allocation of the buffer was more than three quarters of the memory allocated per parse, and limited how the parsing scales on several threads (1.6 times faster on 8 threads than on 1 before, 3.4 times faster now).

For the code macro, see the measurements on XWIKI-25413.