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).
