Re: [RFC] Improve memset

[Linus Torvalds]

On Fri, Sep 13, 2019 at 8:22 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> since the merge window is closing in and y'all are on a conference, I
> thought I should take another stab at it. It being something which Ingo,
> Linus and Peter have suggested in the past at least once.
>
> Instead of calling memset:
>
> ffffffff8100cd8d:       e8 0e 15 7a 00          callq  ffffffff817ae2a0 <__memset>
>
> and having a JMP inside it depending on the feature supported, let's simply
> have the REP; STOSB directly in the code:

That's probably fine for when the memset *is* a call, but:

> The result is this:
>
> static __always_inline void *memset(void *dest, int c, size_t n)
> {
>         void *ret, *dummy;
>
>         asm volatile(ALTERNATIVE_2_REVERSE("rep; stosb",

Forcing this code means that if you do

     struct { long hi, low; } a;
     memset(&a, 0, sizeof(a));

you force that "rep stosb". Which is HORRID.

The compiler should turn it into just one single 8-byte store. But
because you took over all of memset(), now that doesn't happen.

In fact, the compiler should be able to keep a structure like that in
registers if the use of it is fairly simple. Which again wouldn't
happen due to forcing that inline asm.

And "rep movsb" is ok for variable-sized memsets (well, honestly,
generally only when size is sufficient, but it's been getting
progressively better). But "rep movsb"  is absolutely disastrous for
small constant-sized memset() calls. It serializes the pipeline, it
takes tens of cycles etc - for something that can take one single
cycle and be easily hidden in the instruction stream among other
changes.

And we do have a number of small structs etc in the kernel.

So we do need to have gcc do the __builtin_memset() for the simple cases..

                 Linus