[Bug c++/104276] New: Fail to eliminate deadstore from vector constructor

[Bug c++/104276] New: Fail to eliminate deadstore from vector constructor

[Darrell.Wright at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

            Bug ID: 104276
           Summary: Fail to eliminate deadstore from vector constructor
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: Darrell.Wright at gmail dot com
  Target Milestone: ---

clang is unable to remove the memset in code like 

std::vector<int> foo() {
  auto result = std::vector<int>(SZ);
  int *ptr = result.data();
  for (std::size_t n = 0; n < SZ; ++n) {
                ptr[n] = static_cast<int>( n );
  }
  return result;
}
https://gcc.godbolt.org/z/5cbKejfqr

This is unaffected if the value is set during resize.  That may result in
inlining of the memset but does not eliminate it.

However for code that uses blessed methods like memset subsequently

std::vector<int> foo() {
  auto result = std::vector<int>(SZ);
  std::memset(result.data(), 5, sizeof(int) * SZ);
  return result;
}

https://gcc.godbolt.org/z/Kfs9x8Pe9

It is.  This seems to be the usecase of resize_and_overwrite in future string. 
Is there a way to formulate the code to do this.  Or, and I think a better way,
is there a builtin or could there be that lets the compiler assume that the
memory will be subsequently written to?  e.g. `__bultin_assume_set( void *
dest, size_t count )` ?

[Bug tree-optimization/104276] Fail to eliminate deadstore from vector constructor

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
           Severity|normal                      |enhancement

[Bug tree-optimization/104276] Fail to eliminate deadstore from vector constructor

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Full testcase:
#include <cstdint>
#include <iterator>
#include <type_traits>
#include <vector>
#include <cstring>


#define SZ 4096

std::vector<int> foo() {
  auto result = std::vector<int>(SZ);
  int *ptr = result.data();
  for (std::size_t n = 0; n < SZ; ++n) {
                ptr[n] = static_cast<int>( n );
  }
  return result;
}

[Bug tree-optimization/104276] Fail to eliminate deadstore from vector constructor

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>clang is unable to remove the memset in code like 

I think you mean GCC there :).

[Bug tree-optimization/104276] Fail to eliminate deadstore from vector constructor

[Darrell.Wright at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

--- Comment #3 from Darrell Wright <Darrell.Wright at gmail dot com> ---
(In reply to Andrew Pinski from comment #2)
> >clang is unable to remove the memset in code like 
> 
> I think you mean GCC there :).


:) both are true.  This optimization would remove the need for things like
resize_and_overwrite which means we all win

[Bug tree-optimization/104276] memset is not elimited when followed by a store loop writing to that memory location

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|Fail to eliminate deadstore |memset is not elimited when
                   |from vector constructor     |followed by a store loop
                   |                            |writing to that memory
                   |                            |location
   Last reconfirmed|                            |2022-01-28
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Simplified C testcase (without std::vector usage):

#define SZ 4096

void foo1(int *a) {
    __builtin_memset(a, 0, sizeof(int)*SZ);
    for (int n = 0; n < SZ; ++n) {
        a[n] = n;
    }
}

[Bug tree-optimization/104276] memset is not elimited when followed by a store loop writing to that memory location

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |DUPLICATE

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Exact duplicate of PR 28134.

*** This bug has been marked as a duplicate of bug 28134 ***

[Bug tree-optimization/104276] memset is not elimited when followed by a store loop writing to that memory location

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |NEW
         Resolution|DUPLICATE                   |---

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Actually not an exact sorry still only related.

[Bug tree-optimization/104276] memset is not elimited when followed by a store loop writing to that memory location

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
DSE could in theory prune live-bytes from a store inside a loop by looking at
SCEV / niter info.  That'll cost though, and it's not clear where it would
nicely fit.

[Bug tree-optimization/104276] memset is not elimited when followed by a store loop writing to that memory location

[Darrell.Wright at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104276

--- Comment #8 from Darrell Wright <Darrell.Wright at gmail dot com> ---
What about something like a __builtin_overwrite( ptr, size_t ) that tells the
compiler that the range specified will be written.  It forms a contract to do
so with the compiler and would allow the memset to be eliminated or augmented
to part of the range