Ragnar {Groot Koerkamp} @curiouscoding.nl · Dec 13

I know mov is cheap is sometimes a no-op, but... wtaf is going on here?? Yes, let's rotate these 4 registers up rbp->r13->r14->rbx->rcx then do a load from memory with address in rbp then rotate everything back down rbp<-r13<-r14<-rbx<-rcx ??? Must `vpbroadcastq`'s address be relative to rbp?

1 likes 1 replies

?

Replies

Ragnar {Groot Koerkamp} · Dec 13

A good old std::hint::black_box around the computation of the address of the inner load seems to fix it :shrug: