[Bug tree-optimization/107639] New: GCC unable to reason about range of idx/len

[Bug tree-optimization/107639] New: GCC unable to reason about range of idx/len

[jmuizelaar at mozilla dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107639

            Bug ID: 107639
           Summary: GCC unable to reason about range of idx/len
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jmuizelaar at mozilla dot com
  Target Milestone: ---

Clang 14 is able to optimize this function to just 'ret'. GCC 12.2 is not.

#include <cstddef>

void do_checks(const int* begin, const size_t len){
    size_t idx = 0;
    const auto end = begin + len;
    for (const int* it = begin; it!=end; ++it, ++idx){
        if (idx <= len){
            // Do something useful
        }
        else
        {
            throw 5;
        }
    }
}

https://godbolt.org/z/f7PjjqG9T

[Bug tree-optimization/107639] GCC unable to reason about range of idx/len

[jmuizelaar at mozilla dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107639

--- Comment #1 from Jeff Muizelaar <jmuizelaar at mozilla dot com> ---
This test case comes from https://github.com/llvm/llvm-project/issues/56612

[Bug tree-optimization/107639] GCC unable to reason about range of idx/len

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107639

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |missed-optimization
             Blocks|                            |85316
     Ever confirmed|0                           |1
            Version|unknown                     |13.0
   Last reconfirmed|                            |2022-11-14

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  Something for ranger/VRP "symbolic" handling.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85316
[Bug 85316] [meta-bug] VRP range propagation missed cases

[Bug tree-optimization/107639] GCC unable to reason about range of idx/len

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107639

--- Comment #3 from Andrew Macleod <amacleod at redhat dot com> ---
(In reply to Richard Biener from comment #2)
> Confirmed.  Something for ranger/VRP "symbolic" handling.

I'm not so sure.  what is related?  there are 2 loop values being bumped by
different amounts.  the loop is controlled by the it pointer, which is bumped
by 4.  The only connection between len5 and idx_14 is thru the calculation of
end_7, which boils down to is begin+len_5 * 4.  theres no way ranger is making
that connection now.

The loop boils down to:

    <bb 2> [local count: 118111600]:
    _1 = (long unsigned int) len_5(D);
    _2 = _1 * 4;
    end_7 = begin_6(D) + _2;
    if (begin_6(D) != end_7)
      goto <bb 5>; [89.00%]

    <bb 3> [local count: 850510901]:
    if (len_5(D) >= idx_14)              // how to know this is always true
      goto <bb 5>; [100.00%]
    else
      goto <bb 4>; [0.00%]

    <bb 5> [local count: 955630225]:
    # idx_17 = PHI <idx_14(3), 0(5)>
    # it_18 = PHI <it_13(3), begin_6(D)(5)>
    it_13 = it_18 + 4;
    idx_14 = idx_17 + 1;
    if (end_7 != it_13)
      goto <bb 3>; [89.00%]
    else
      goto <bb 6>; [11.00%]

This seems more like a loop thing.  both values are being bumped by the loop,
and the end condition is based on a calculation using len_5.. it seems like it
should now the bounds of both ir and idx?  or maybe it does and cant
communicate it any more?