John Keiser

I actually think we should have a (templated?) version of on demand that validates everything as you go. I think validate() is a good idea too. My change suggestions here...

Perhaps more succinctly: "simdjson will detect any JSON errors that would change the output of your program."

I think the default user will want some assurance it is safe. I can imagine a flag that goes the other way, too: "assume the document is valid."

> > Do we actually want to release a JSON library to users, or a performant JSON reader that is used internally by other JSON libraries? (I assume the former,...

Funny that it should come up right now, but [this](https://github.com/eddelbuettel/rcppsimdjson/issues/7) is one example where a 64-bit fallback would help. Enough other software is working fine on the test machine in...

> @jkeiser I think that this is one example where we had a Penryn processor. Penryn procesors lacked POPCNT (starts with Nehalem), and CLMUL (starts with Westmere). Just to be...

It would almost certainly be worse than other parsers in terms of performance, too, because the naive implementation I have in mind would still be split up into stage 1...

> Almost all x64 binaries rely on SIMD instructions. One simple reason for this is that 128-bit SSE registers are part of what defines x64. For fun, try to compile...

Yeah, it looks like sse2 is the minimum: what I'm reading indicates that all x64 instruction sets include this. Or at least that when you compile using g++ by default,...

So I'm super interested in a lot of the work here, but as I have a lot of experience in API design/usability, I'm probably going to focus on that. I've...