https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81196
Bug ID: 81196
Summary: Number of iterations found for p!=q but not for p<q
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: glisse at gcc dot gnu.org
Target Milestone: ---
void f(short*p){
p=(short*)__builtin_assume_aligned(p,64);
short*q=p+256;
for(;p!=q;++p,--q){
short t=*p;*p=*q;*q=t;
}
}
compiled with -O3 -march=skylake, this gets vectorized. However, if I replace
p!=q with p<q (much safer in case 256 was actually 255), it isn't vectorized
anymore, apparently because the number of iterations cannot be determined.
Replacing p!=q with p!=q&&(p+1)!=q (or using std::reverse) also breaks
vectorization, but that's less surprising.
