From: Arnd Bergmann > Sent: 30 August 2019 16:59 ... > > Or even better, it would be great if we could get Clang to change their > > memset() insertion heuristics, so that KASAN acts more like non-KASAN > > code in that regard. > > I suspect that's going to be harder. The clang-9 release is going to be > soon, and that change probably wouldn't be considered a regression fix. > > Maybe Nick can find what happens, but I don't actually see any reference > to KASAN in the llvm source code related to the memset generation. > > https://github.com/llvm-mirror/clang/blob/master/lib/CodeGen/CGExprAgg.cpp#L1803 > has a check for >16 bytes, but that again does not match my observation.
That looks like the code for initialising a local structure variable, not for merging 'random' writes into a memset(). David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)