[Bug tree-optimization/70317] New: [6 Regression] folding x<=x to true for large vectors, ignoring nans

["glisse at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70317

            Bug ID: 70317
           Summary: [6 Regression] folding x<=x to true for large vectors,
                    ignoring nans
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org
  Target Milestone: ---

typedef double vecf __attribute__((vector_size(256)));
auto f(vecf x){
  return x <= x;
}

g++ x.c -O3 -S -fdump-tree-all

the .optimized dump has:
  <retval> = { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 };

This is because the corresponding match.pd pattern uses:
HONOR_NANS (TYPE_MODE (TREE_TYPE (@0)))
instead of
HONOR_NANS (@0)

(don't know if I missed it when I redefined HONOR_NANS or if it was added
later)

By the way, the pattern canonicalizes x==x, x<=x and x>=x to x==x, but
ORDERED_EXPR would be a nicer alternative IMHO (if we really prefer x==x, then
we should also canonicalize x ord x to x == x).