[PATCH v2] msvcrt: Add an ermsb-based memset for x86/64.

[Rémi Bernon]

On 8/31/21 2:18 PM, Piotr Caban wrote:
> On 8/31/21 12:52 PM, Rémi Bernon wrote:
>> For a start we could add -ffreestanding / -fno-builtin compilation 
>> flags   to msvcrt instead, although it won't be much faster, it will 
>> at least fix the issue with the builtin replacement.
> As far as I understand -fno-builtin does not guarantee that there will 
> be no call to memset introduced by compiler. Isn't it similar to the 
> memcpy problem we have while using clang?

I don't really know, but in practice it has the desired effect. And yes, 
there's probably the same problem with memcpy.

Note that apparently there's a flag for each builtin but 
-fno-builtin-memset doesn't do the trick for GCC (it does on clang).

Also note that, as far as I could see, unrolling the loop a bit like in 
my quote is already enough to make it stop optimizing it to a builtin 
memset call, without having to add specific flags (although there's 
always the risk of it becoming even more clever).
-- 
Rémi Bernon <rbernon at codeweavers.com>