<html><head><meta http-equiv="Content-Type" content="text/html; charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class=""><div><blockquote type="cite" class=""><div class="">On Dec 3, 2020, at 10:26 AM, David Blaikie <<a href="mailto:dblaikie@gmail.com" class="">dblaikie@gmail.com</a>> wrote:</div><div class=""><div dir="ltr" class=""><div class="gmail_quote"><div class="">I was meaning to highlight that I think a better conceptual abstraction than "small" (1 or more inlined objects) and "not small" (0 inlined objects) it'd be better to have an abstraction more like std::string - where the smallness isn't part of the API, as such. <br class=""></div></div></div></div></blockquote><div><br class=""></div>But we have SmallString as a dual to std::string precisely because we need that distinction. Speaking of, the whole default argument thing should probably be applied to SmallString, SmallDenseMap and other types as well.</div><div><br class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="gmail_quote"><div class="">I'm suggesting that we should have two types:<br class=""><br class="">Reduced sizeof(t) and guaranteed iterator validity over swap/move (ie: must not use an inline buffer)<br class="">The rest: Like std::string. It might have an inline buffer, it might have zero, depending on size, etc. But importantly it's not guaranteed to maintain iterator/pointer validity on move/swap, and sizeof is optimized for stack usage.<br class=""><br class="">Neither of these really make sense being called "SmallVector" I suppose. And I'd dislike calling (2) "Vector" even though it's likely the more popular one - because it'd have subtly different semantics from std::vector regarding iterator invalidation.</div></div></div></div></blockquote><div><br class=""></div><div>I think I see what you’re going for here, and I agree that std::string is a slightly different situation than std::vector given that some implementations have a small string optimizations already.</div><div><br class=""></div><div>I feel like you’re prioritizing the iterator invalidation behavior as the core difference between the two types, whereas I’m prioritizing the performance/allocation-behavior aspect. I feel like this is a major difference in practice that API users should think about, and they should be prepared to deal with the iterator invalidation issues as necessary.</div><div><br class=""></div><div>Is the "Reduced sizeof(t) and guaranteed iterator validity over swap/move (ie: must not use an inline buffer)” use case important enough to have a “string-y” type? As you say, we don’t have a solution for this in the current code other than std::vector. I haven’t heard of this being a significant enough problem to be worth “fixing”, and I don’t think we can drop the inline vs out-of-line distinction.</div><div><br class=""></div><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div class=""><div class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="">So I think (2) has some legs - but the problem for me is that if we name the type llvm::Vector, which already gets a bit close to std::vector (unqualified "Vector" seems like fairly subtle code to me - but I'm leaning towards being OK with that aspect if other peopel are) and it has different iterator invalidation/move semantics than std::vector, I think that could be problematic.</div></div></div></blockquote><div class=""><br class=""></div>Not sure if I want to defend this, but we have lots of precedent for this, including llvm::sort etc.</div></div></blockquote><div class=""><br class="">For sure - I'm not trying to advocate for avoiding having a std::vector replacement in LLVM (excuse the several negatives there). And some differences between the C++ standard APIs and the LLVM ones is expected - llvm::sort's and many of the other alternatives are fairly "obvious" I'd say, it sorts a range instead of a pair of iterators - I don't think anyone would find that surprising. Something with subtly different iterator/pointer invalidation semantics under a near exactly matching name wouldn't be such a good fit, though.<br class=""></div></div></div></div></blockquote><div><br class=""></div><div>The difference I meant was that it forwards transparently to array_pod_sort / qsort which is a pretty big behavioral difference.</div></div><br class=""><div class="">-Chris</div></body></html>