They can, yes.
For example, here's a simple Compiler Explorer link with a simple loop conditionally copying from one of two values:
https://godbolt.org/z/TbeK99sx1.
If you look at the assembly code produced, you'll see in lines 9-11, it's doing a similar thing as the swap_if() function: ANDing one value with the mask, ANDing the other with the negation of the mask, and then ORing them together. And it produced that automatically from the ternary in the loop.
They can also do things like automatically generate epilogues with a serial loop to handle the leftover items if the loop count isn't an even multiple of the SIMD width. That's something else that I remember writing by hand before autovectorizers.
That said, coaxing an autovectorizer to actually vectorize a loop can be an art in itself. They're notoriously finicky. My favorite used to be the one in the old ("classic") ICC compiler where you use compiler flags to get a nice optimization report which would explain exactly why it chose what it did for each loop. You could also set #pragma's which would noisily warn or error out if the associated loop failed to vectorize for some reason.
(And of course, a lot of this stuff you might just use GPU compute for now.)
For example, here's a simple Compiler Explorer link with a simple loop conditionally copying from one of two values:
https://godbolt.org/z/TbeK99sx1.
If you look at the assembly code produced, you'll see in lines 9-11, it's doing a similar thing as the swap_if() function: ANDing one value with the mask, ANDing the other with the negation of the mask, and then ORing them together. And it produced that automatically from the ternary in the loop.
They can also do things like automatically generate epilogues with a serial loop to handle the leftover items if the loop count isn't an even multiple of the SIMD width. That's something else that I remember writing by hand before autovectorizers.
That said, coaxing an autovectorizer to actually vectorize a loop can be an art in itself. They're notoriously finicky. My favorite used to be the one in the old ("classic") ICC compiler where you use compiler flags to get a nice optimization report which would explain exactly why it chose what it did for each loop. You could also set #pragma's which would noisily warn or error out if the associated loop failed to vectorize for some reason.
(And of course, a lot of this stuff you might just use GPU compute for now.)