Correct the http benchmarks

[andrewthad]

There two two unrelated issues with them. The first was that the samples had extra newline breaks in them which would cause the request and the response to be parsed incorrectly. The second was that the warp request parser was used incorrectly. It needs to be fed a list of bytestring that has already been split on newlines. The results of the changes are that the request parsers for both warp and attoparsec are significantly slower. The warp one is about 15x slower than it previously was because it previously did almost no work. It is not totally clear why the attoparsec one is slower, but it is 3x slower than it previously was.

[andrewthad]

I just realized why the parsing code is three times slower. It’s because the parser was previously failing when it tried to parse a header that had been line wrapped.

[andrewthad]

Also, it’s a little strange that the parser uses inClass, which is specifically discouraged in the documentation. I suspect that it also contributes to the slowness. I am going to change that as well.

For what it’s worth, the warp http header parser does not appear to do any validation of the header name. This difference is probably worth noting in the benchmarking code. It makes it nearly impossible for attoparsec to match the performance of warp in this case.

[andrewthad]

I've added a note about warp. Also, I've removed the use of inClass.

[bgamari]

Thanks! I've opened bgamari#1 to test under CI.