Optimize HTTPHdr conversion of HTTP/1.1 to HTTP/2

[masaori335]

http2_generate_h2_header_from_1_1() is one of bottlenecks of HTTP/2 subsystem because of extra HdrHeap allocation and all header copies.

• master
  http2_generate_h2_header_from_1_1() is 32% of Http2ConnectionState::send_headers_frame()

• 6635
  http2_convert_header_from_1_1_to_2() is 3% of Http2ConnectionState::send_headers_frame()

[maskit]

I'm not sure if this is still true, but I think I had to generate a new HTTPHdr when I add/modify the function. The reason was probably that the original HTTPHdr was read from somewhere else later and the code expected HTTP/1.1 style.

[shinrich]

Looks good. Updating in place makes a lot of sense.

[masaori335]

@maskit Was it FetchSM::check_chunk()? We got rid of it. Also, AuTests covers chunked contents cases well.

TS-4321 - Masakazu Kitajo added a comment - 06/Apr/16 02:19
I confirmed that it is certainly caused by the last HPACK change, and is triggered by TE: chunked.
The last HPACK change removes TE header from a header list which is contained in a FetchSM, so check_chunk() in FetchSM doesn't return true.

[maskit]

@masaori335 Yes, you found it! Then it should be fine, but let me review this a bit more.

[maskit]

I'm not a big fun of the additional initialization call for pseudo headers, but I don't have a better idea. Ideally, HTTPHdr should probably have a concept of pseudo header.

[zwoop]

Cherry-picked to v9.0.x branch.