[llvm] [X86, SimplifyCFG] Support hoisting load/store with conditional faulting (Part II) (PR #108812)

[Shengchen Kan via llvm-commits]

KanRobert wrote:

> > For cases like this why aren't we folding to:
> > ```
> > void src(int a, int *c, int *d) {
> >   if (a)
> >    *c = a;
> >   else
> >    *d = a;
> > }
> > 
> > void tgt(int a, int *c, int *d) {
> >   int *p = a ? c : d;
> >   *p = a;
> > }
> > ```
> > 
> > 
> >     
> >   
> > or the even more general:
> > ```
> > void src(int s, int a, int b, int *c, int *d) {
> >   if (s)
> >    *c = a;
> >   else
> >    *d = b;
> > }
> > 
> > void tgt(int s, int a, int b, int *c, int *d) {
> >   int *p = s ? c : d;
> >   int v = s ? a : b;
> >   *p = v;
> > }
> > ```
> 
> That's a good question. It doesn't rely on hardware feature, so can be a general branch to select transformation. I'm also wondering why didn't we do it before, especially the first one. Is there any concern on possible sideeffect of memory operand?

>From this perspective, should we implement this optimization w/o any HW feature and do further optimization w/ CFCMOV at backend?



https://github.com/llvm/llvm-project/pull/108812