minor speedup in pcg32_random_r() possible?

[jason-s]

return (xorshifted >> rot) | (xorshifted << ((-rot) & 31));

Wouldn't this be equivalent (and hence be slightly faster)?

return (xorshifted >> rot) | (xorshifted << (32-rot));

[imneme]

No. Not when rot == 0. Your code would shift by 32, and that’s undefined behavior.

[imneme]

FWIW, when the compiler's idiom-detection is working properly, it should look at this code and say “Oh, that's an idiomatic C/C++ rotate, I'll compile that to a single rotate instruction”. If the optimizer doesn't do that, it's not really doing its job and should be reported as a bug.

[jason-s]

ah -- those darn standards writers. Thanks!

(editorial comment: Seems like the C standard library should have contained a rotate intrinsic, rather than forcing programmers to jump through tricky hoops to do what they really want.)

[imneme]

Agreed!

[travisdowns]

@jason-s - in this case I think the standard writers chose a reasonable tack: ISA do actually differ in what they do when do a shift by 32. Some make the result zero, x86 leaves the result unchanged (effectively, the shift amount is taken mod 32), ARM is a mix of both.