Bug 3637 – Array append patch to prevent stomping and to enhance thread-local append performance

Created attachment 528 Patch to prevent array stomping and increase append performance The attached patch fixes 2 problems: 1. Array stomping. A slice of an array can stomp on another array if the slice starts at the beginning of the other array. This patch prevents such problems by storing the "allocated length" or the length of valid data at the end of the array. 2. Array append previously required acquiring the GC lock, even when the array can be extended in-place. The patch makes use of an 8-element MRU (most recently used) cache in thread-local storage that allows the runtime to avoid taking the global lock to look up an array's allocated length. This allows one to append to up to 8 arrays per thread without taking the global lock unless an actual re-allocation is required. Shared array appending still requires a global lock for appending, and does not have a cache associated with it. However, shared appending is still safe as far as stomping goes. Note that with this patch, the following trick *no longer works*: int[] bigarray = new int[10000]; bigarray.length = 0; foreach(i; 0..10000) bigarray ~= i; Because truncating the array this way can cause stomping, the first append will reallocate bigarray and you will lose the benefit of preallocating the original array. A (yet unwritten) library function is required to preallocate an array. To apply the patch, download the 2.037 distribution (this code was only tested against the distribution), and apply the patch from the src/druntime directory: cd dmd2/src/druntime patch -p0 < append.patch then build and re-install druntime and phobos libraries. The patch was tested on Linux dmd version 2.037. An earlier version of the patch which was mostly the same was tested on Windows dmd 2.036 and Mac OSX 10.6 dmd 2.037. The difference between the earlier and this version of the patch is the handling of appending dchar to char[] and wchar[].

Comment on attachment 528 Patch to prevent array stomping and increase append performance Fix the MIME type and mark the attachment as a patch.

It is really necessary change the GC API to fix this? I think it's a really bad idea to modify the GC API (add a function) just because dynamic arrays have problems (I don't want to start over the discussion about dynamic arrays being broken :). If you absolutely need this new function: BlkInfo gc_malloc_bi(size_t sz, uint ba = 0); then maybe it's better to change the existing regular gc_malloc() to: void* gc_malloc(size_t sz, uint ba = 0, BlkInfo bi = null); The advantage is that it's API-backward compatible and doesn't introduce a new function to the GC API and the downside is it's not binary-backward compatible, but I don't think people care much about this in D. The inability to pre-allocate can be a little bad too, but it was awkward anyway, so providing a better way to do so can be a good thing after all.

It's not *necessary* to change the API to fix this, but it is hugely advantageous. If you want to know the block size of the chunk you just allocated, the current API requires *another* lock of the GC, and a search through the pools. All the info is there in malloc, it's just not returned. BTW, just an additional size parameter would suffice (this is how gcx handles the allocation). Having it return a BlkInfo struct is convenient because that is the data type I'm working with when setting allocated length. And introducing a new function is just as backwards compatible as adding a new optional parameter to a current function. Preallocation needs to be a runtime function that is yet to be written which allocates a large block but sets the "allocated" size to 0. It can also avoid pre-initializing the block if the block has no pointers (something that the old trick didn't do).

(In reply to comment #3) > It's not *necessary* to change the API to fix this, but it is hugely > advantageous. If you want to know the block size of the chunk you just > allocated, the current API requires *another* lock of the GC, and a search > through the pools. All the info is there in malloc, it's just not returned. > > BTW, just an additional size parameter would suffice (this is how gcx handles > the allocation). Having it return a BlkInfo struct is convenient because that > is the data type I'm working with when setting allocated length. Yes, that seems reasonable since BlkInfo is already part of the API. > And introducing a new function is just as backwards compatible as adding a new > optional parameter to a current function. I know that, but it adds complexity to the API, and seems a little weird since gc_malloc() and gc_malloc_bi() are basically the same.

Created attachment 569 Patch to druntime revision 245 to implement array appending generated with svn diff.

Note that shared data isn't properly synchronized when being copied, but this also happens in all the other array functions (not just appending). I will file a separate bug on this.

Created attachment 570 Patch to druntime revision 245 to implement array appending The only addition to this patch is I added comments identifying where the compiler fails to deliver the proper typeinfo for shared appending.

changeset 252 http://www.dsource.org/projects/druntime/changeset/252

Fixed dmd 2.041