Conversation
* incorporate simd percent decoder * fixing some comments * made avx512 version match existing behavior on malformed input * made avx2 version match existing behavior on malformed input. made percent_decode_slow the fallback on non-x64_64 or x64_64 lacking avx2, which required changing the function signature of percent_decode_slow. deleted simde version since simde would require an external dependency. * got rid of a few instructions in AVX2 version --------- Co-authored-by: Jeff Plaisance <jplaisance@indeed.com>
|
I wouldn't expect improving percent decoding performance to make a huge difference in the benchmarks right now, since it seems that currently percent decoding is only performed while parsing the host. I'd expect it to make a substantial difference when parsing query strings in long URLs. |
|
I believe in either scenario, we can expose this under |
Ah ! Ah ! This was so before TEST(url_search_params, with_accents) {
auto search_params = ada::url_search_params();
search_params.append("key1", "été");
search_params.append("key2", "Céline Dion++");
ASSERT_EQ(search_params.size(), 2);
ASSERT_EQ(search_params.to_string(), "key1=%C3%A9t%C3%A9&key2=C%C3%A9line+Dion%2B%2B");
ASSERT_EQ(search_params.get("key1"), "été");
ASSERT_EQ(search_params.get("key2"), "Céline Dion++");
SUCCEED();
} |
|
@jeffplaisance and @anonrig : I think we need a new benchmark where we could make this code shine. What sort of applications tend to generate long URLs with longer percent encoded search values? It does not need to be a real use case, but it should be realistic. |
|
here are some really long examples from indeed.com, these are authentication redirects which is probably a good example of a scenario where long URLs with lots of percent encoded data are common: to generate more we could take some regular urls from an existing benchmark, percent encode them, and put them as a query parameter in a new url |
|
@jeffplaisance That's helpful. I'll work on a benchmark "soon". Update: Done by #477 |
|
We have a precursor PR: #478 |
3 participants
Incorporate simd percent decoder (Jeff Plaisance). Currently, this does not shine in our benchmarks.
credit: @jeffplaisance