There are a number of test crashes on the Windows platform that seem to be caused by the FastMalloc implementation. As one would expect, these are tied to mismatched new/free pairs, and seem to be related to the visibility of the FastMalloc overloads in template expansions, and perhaps when data is allocated and deallocated across shared library boundaries. If I force use of the system allocator, these crashes go away. I cannot measure the performance penalty this might introduce, because we do not currently track Windows performance. However, there does not seem to be a noticeable change in using the MiniBrowser with the system allocator activated. Perhaps I am not using test cases that create/destroy enough memory for this to be noticeable. I would like to do the following: 1. Turn off FastMalloc on the Windows platform. 2. Reactivate the various tests that currently fail due to issues with the allocator. 3. Get a performance bot running 4. Port ggaren’s new allocator to run on Windows, but only use it if we see no change in stability.
Created attachment 213420 [details] Patch
Committed r156933: <http://trac.webkit.org/changeset/156933>
<rdar://problem/15321791>
(In reply to comment #0) > There are a number of test crashes on the Windows platform that seem to be > caused by the FastMalloc implementation. As one would expect, these are tied > to mismatched new/free pairs, and seem to be related to the visibility of > the FastMalloc overloads in template expansions, and perhaps when data is > allocated and deallocated across shared library boundaries. > > If I force use of the system allocator, these crashes go away. > > I cannot measure the performance penalty this might introduce, because we do > not currently track Windows performance. However, there does not seem to be > a noticeable change in using the MiniBrowser with the system allocator > activated. Perhaps I am not using test cases that create/destroy enough > memory for this to be noticeable. > > I would like to do the following: > > 1. Turn off FastMalloc on the Windows platform. > 2. Reactivate the various tests that currently fail due to issues with the > allocator. > 3. Get a performance bot running > 4. Port ggaren’s new allocator to run on Windows, but only use it if we see > no change in stability. Is there any plan to switch the Windows port to bmalloc too in the near future?