Improving Parser Performance Using SSE Instructions

Using SSE4.2 or more is trivial. Same goes for fast hashes with the crc32 intrinsic. Mostly the compiler does it for you (-march=native)

But you cannot assume everybody has such a CPU. Hence you'd either need to compile your own (a la macports, gentoo, perl, ...) or do run-time checks for the CPU feature and switch to the fast version then.

I wonder if there is any HTML parsers that use SSE4.2 instructions

IIRC Servo is using html5ever for HTML parsing, which (at least in one branch) uses SSE4.2 for parsing.

It's an interesting idea, but I found this blog post unenlightening. Does anyone have links to other writeups?

clang does a very similar thing in the preprocessor (looking for new lines if my memory serves correctly).

can http parsing be done on a GPU?

Maybe, but the transfer time to/from the gpu is probably going to kill any gains you might get.