This article is about Go, but I wonder how many C/C++ developers realize that you've always had the ability to allocate on the stack using alloca() rather than malloc().
Of course use cases are limited (variable length buffers/strings, etc) since the lifetime of anything on the stack has to match the lifetime of the stack frame (i.e the calling function), but it's super fast since it's just bumping up the stack pointer.
Nice to see common and natural patterns to have their performance improved. Theoretically appending to a slice would be possible to handle with just stack growth, but that would require having large gaps between goroutine stacks and mapping them lazily upon access instead of moving goroutines to the new contiguous blocks as it's implemented right now. But given how many questionable changes it requires from runtime it's certainly not going to happen :)
Awesome stuff! Does Go have profile-guided optimization? I'm wondering whether a profile could hint to the compiler how large to make the pre-reserved stack space.
> ...
> On the third loop iteration, the backing store of size 2 is full. append again has to allocate a new backing store, this time of size 4. The old backing store of size 2 is now garbage.
Correct me if I'm wrong, but isn't this a worst-case scenario? realloc can, iirc, extend in place. Your original pointer is still invalid then, but no copy is needed then.
Unless I'm missing something?
Equally, what happens to the ordering of variables on the stack? Is this new one pushed as the last one? Or is there space kept open?
[]task is a pointer to a range of elements. TFA says if you initialize it to point to a new array of 10 elements, that array of 10 elements may be stack–allocated. If you allocate another array dynamically, that one won't be.
It's kind of like the small string optimization you see in C++[1] where all the string metadata to account for heap pointer, size and capacity is union'ed with char*. Getting the stack allocation doesn't costs extra memory, but does cost a bit check. Not sure if slices in go use the same method. 32 bytes is a lot so maybe they fattened slice representations a bit to get a bit more bang for your buck?
It is actually a bad design when compiler go this far into a micro optimization but assume it understands the context so it can make decisions for you.
I want to like this, and it's directionally good work...
But it's hard to see this as very useful unless we also start to see some increases in legibility, and ways to make sure these optimizations are being used (and that textually minor changes don't cause non-obvious performance regressions).
I've written a lot of golang code that was benchmarked to shreds, and in which we absolutely cared about stack-vs-heap allocations because they were crucial to overall performance. I've spent a lot of time pouring over assembler dumps, because grepping those for indications of new object creation was sometimes clearer (and certainly more definitive) than trying to infer it from the source code level. The one thing I've learned from this?
It's very, very easy for all those efforts to come to naught if the rules change slightly.
And it's very, very, VERY easy for a co-maintainer on a project to stroll in and make seemingly textually trivial changes that have outsized impacts. (I'm looking at inliner thresholds, specifically. Hoo boy.)
The best balm we have for this right now is writing benchmarks and making sure they report zero allocs. (Or unit tests using the runtime memstats harness; potato potato.) But that is a very fragile balm, and relatively complex to maintain, and (if DX is considered) is not textually local to the code in question -- which means someone changing the code can easily miss the criticality of a section (until the tests yell at them, at least).
I really yearn for some markup that can say "I expect this code to contain zero heap allocations; please flunk the compile if that is not the case".
15 comments
[ 4.0 ms ] story [ 36.0 ms ] threadOf course use cases are limited (variable length buffers/strings, etc) since the lifetime of anything on the stack has to match the lifetime of the stack frame (i.e the calling function), but it's super fast since it's just bumping up the stack pointer.
Correct me if I'm wrong, but isn't this a worst-case scenario? realloc can, iirc, extend in place. Your original pointer is still invalid then, but no copy is needed then.
Unless I'm missing something?
Equally, what happens to the ordering of variables on the stack? Is this new one pushed as the last one? Or is there space kept open?
E.g.:
[1] https://github.com/elliotgoodrich/SSO-23
But it's hard to see this as very useful unless we also start to see some increases in legibility, and ways to make sure these optimizations are being used (and that textually minor changes don't cause non-obvious performance regressions).
I've written a lot of golang code that was benchmarked to shreds, and in which we absolutely cared about stack-vs-heap allocations because they were crucial to overall performance. I've spent a lot of time pouring over assembler dumps, because grepping those for indications of new object creation was sometimes clearer (and certainly more definitive) than trying to infer it from the source code level. The one thing I've learned from this?
It's very, very easy for all those efforts to come to naught if the rules change slightly.
And it's very, very, VERY easy for a co-maintainer on a project to stroll in and make seemingly textually trivial changes that have outsized impacts. (I'm looking at inliner thresholds, specifically. Hoo boy.)
The best balm we have for this right now is writing benchmarks and making sure they report zero allocs. (Or unit tests using the runtime memstats harness; potato potato.) But that is a very fragile balm, and relatively complex to maintain, and (if DX is considered) is not textually local to the code in question -- which means someone changing the code can easily miss the criticality of a section (until the tests yell at them, at least).
I really yearn for some markup that can say "I expect this code to contain zero heap allocations; please flunk the compile if that is not the case".