https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90269
Bug ID: 90269
Summary: loop distribution defeated by clobbers
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: glisse at gcc dot gnu.org
Target Milestone: ---
#include <memory>
#ifndef EASY
typedef std::unique_ptr<int> T;
#else
typedef int* T;
#endif
void f(T*__restrict a,T*__restrict b){
for(int i=0;i<1024;++i){
new(a+i)T(std::move(b[i]));
b[i].~T();
}
}
Compiling with -O3, with -DEASY I get a nice call to memcpy, while without it I
keep a loop. The difference, before the ldist pass, is that the version with
unique_ptr has clobbers:
_8 = MEM[(int * const &)_3];
MEM[(struct &)_10] ={v} {CLOBBER};
MEM[(struct _Head_base *)_10]._M_head_impl = _8;
MEM[(struct &)_3] ={v} {CLOBBER};
ldist checks gimple_has_side_effects (stmt) on the first clobber and gives up.
Vectorization does not seem to have any problem with those clobbers.
(by the way, I believe DSE could remove the first clobber, it seems pretty
useless when it is immediately followed by a write to the full object)
