LoopIdiom does not produce memcpy/memset with runtime checks if pointers may alias

[davidbolvansky]

Bugzilla Link	51550
Version	trunk
OS	Linux
CC	@fhahn,@RKSimon,@nikic,@zhuhan0

Extended Description

typedef SIZE_TYPE size_t;

void idiom_memcpy_aliasing(unsigned char * dst, size_t d_len, unsigned char *src)
{
// no restrict, possible aliasing
unsigned char *d_end = dst + d_len;
// icc seems to produce memcpy when alias checks are OK and d_len j < 96, jumps to trivial copy loop otherwise.
while (dst < d_end)
*dst++ = *src++;
}

ICC -O2+ is able to do this kind of transformation.

https://godbolt.org/z/eK1P4YjMf

[llvmbot]

That's interesting. If we would like to extend LIR with that kind of transformation then it's about making processLoopStoreOfLoopLoad/processLoopStridedStore smarter.
In case of memcpy for now we bail out when loop access load (and alias checks fails). Instead giving up we could reuse precise offsets validation added in scope on memmove transformation recently.
The tricky part is building pointers distance checking code as ICC does. IIRC for now we don't do anything like that when we build memset/memcpy/memmove calls.

[davidbolvansky]

Maybe there is some easy to use API to generate such checks? Vectorizer (loop; slp soon?) generates them, not sure if reusable..

CC Florian as he works on something similar for SLP and may provide some useful hints.

[kasuga-fj]

Hi. I recently ran into the same situation and found this issue. I think now we can use LoopVersioning to generate a pointer overlap check.

[kasuga-fj]

Seems like no one is working on it, so I will.

[kasuga-fj]

As mentioned in #125043, we should add new another pass to address this issue. I don't have enough time to work on it, so I'm sorry, but I give up for now.