[Bug tree-optimization/118132] New: Vectorize optimization generates obsolete code, which causes significant performance drop

[German.Andryeyev at amd dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118132

            Bug ID: 118132
           Summary: Vectorize optimization generates obsolete code, which
                    causes significant performance drop
           Product: gcc
           Version: 11.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: German.Andryeyev at amd dot com
  Target Milestone: ---

Created attachment 59926
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59926&action=edit
A C++ sample that exposes a problem with tree-vectorize optimization

- The attached zip file contains a trivial program(main.cpp, dispatch.cpp,
dispatch.hpp) and a script to compile the program (build.sh)
- The script accepts 2 options 0 and 1. 0 - disables vectorize for one function
and 1 - enables
- The portion of the code that will disable tree-vectorize
#if !VECTORIZE
__attribute__((optimize("no-tree-vectorize")))
#endif
const DispatchTable* GetDispatchTable() {
  static auto* _v = &GetDispatchTableImpl<DispatchTable>();
  return _v;
}
- The issue: compiler adds extra instructions which will fill the struct on the
stack unconditionally for every invocation of GetDispatchTable() and then drops
that.