On 7/28/2025 12:38 PM, David Laight wrote:

quoted

quoted

...

Or just write a byte copy loop in C with (eg) barrier() inside it
to stop gcc converting it to memcpy().

	David  

Great. It's rep movsb without any of the performance.

And without the massive setup overhead that dominates short copies.
Given the rest of the code I'm sure a byte copy loop won't make
any difference to the overall performance.

Wouldn't it be better to introduce a generic mechanism than something
customized for this scenario?

PeterZ had suggested that inline memcpy could have more usages:
https://lore.kernel.org/lkml/20241029113611.GS14555@noisy.programming.kicks-ass.net/ (local)

Is there a concern that the inline versions might get optimized into
standard memcpy/memset calls by GCC? Wouldn't the volatile keyword
prevent that?

static __always_inline void *__inline_memcpy(void *to, const void *from,
size_t len)
{
	void *ret = to;

	asm volatile("rep movsb"
		     : "+D" (to), "+S" (from), "+c" (len)
		     : : "memory");
	return ret;
}

static __always_inline void *__inline_memset(void *s, int v, size_t n)
{
	void *ret = s;

	asm volatile("rep stosb"
		     : "+D" (s), "+c" (n)
		     : "a" ((uint8_t)v)
		     : "memory");
	return ret;
}