From: Arvind Sankar > Sent: 29 October 2020 21:35 > > On Thu, Oct 29, 2020 at 09:41:13PM +0100, Thomas Gleixner wrote: > > On Thu, Oct 29 2020 at 17:59, Paolo Bonzini wrote: > > > On 29/10/20 17:56, Arvind Sankar wrote: > > >>> For those two just add: > > >>> struct apic *apic = x86_system_apic; > > >>> before all the assignments. > > >>> Less churn and much better code. > > >>> > > >> Why would it be better code? > > >> > > > > > > I think he means the compiler produces better code, because it won't > > > read the global variable repeatedly. Not sure if that's true,(*) but I > > > think I do prefer that version if Arnd wants to do that tweak. > > > > It's not true. > > > > foo *p = bar; > > > > p->a = 1; > > p->b = 2; > > > > The compiler is free to reload bar after accessing p->a and with > > > > bar->a = 1; > > bar->b = 1; > > > > it can either cache bar in a register or reread it after bar->a > > > > The generated code is the same as long as there is no reason to reload, > > e.g. register pressure. > > > > Thanks, > > > > tglx > > It's not quite the same. > > https://godbolt.org/z/4dzPbM > > With -fno-strict-aliasing, the compiler reloads the pointer if you write > to the start of what it points to, but not if you write to later > elements. I guess it assumes that global data doesn't overlap. But in general they are sort of opposites: With the local variable it can reload if it knows the write cannot have affected the global - but is unlikely to do so. Using the global it must reload if it is possible the write might have affected the global. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
