[Bug tree-optimization/54717] [4.8 Regression] Runtime regression: polyhedron test "rnflow" degraded

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54717

--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-27 10:43:00 UTC ---
I can reproduce the slowdown.  Code differences appear first in early FRE,
good ones like:

-  _84 = &*a_56(D)[_83];
+  _84 = _75;

which was the intention of the patch (and that is also likely the
reason for the inliner code size/time estimate changes).

It would be nice to get a smaller testcase for the PRE change you quote.

Unfortunately the big slowdown does not reproduce with -fno-inline which makes
it harder to track down.

The real differences do appear in PRE, some of the kind you quote and
some where we perform more PRE like:

@@ -19695,11 +19720,13 @@
   <bb 289>:
   pretmp_ = stride.258_ * _;
   pretmp_ = offset.259_ + pretmp_;
+  pretmp_ = stride.258_ * _;
+  pretmp_ = offset.259_ + pretmp_;

   <bb 123>:
   # i_ = PHI <1(289), i_(292)>
-  _ = stride.258_ * _;
-  _ = _ + offset.259_;
+  _ = pretmp_;
+  _ = pretmp_;

Aside from that the differences you quote result in less if-conversion
applied:

   # ival2_ = PHI <ival2_(39), ival2_(41)>
   # ival2_ = PHI <ival2_(39), ival2_(41)>
-  # prephitmp_ = PHI <pretmp_(39), prephitmp_(41)>
   _ = (integer(kind=8)) ival2_;
   _ = _ + -1;
   _ = *xxtrt_(D)[_];
-  ival2_ = _ < prephitmp_ ? ival2_ : ival2_;
-  prephitmp_ = MIN_EXPR <_, prephitmp_>;
+  _ = (integer(kind=8)) ival2_;
+  _ = _ + -1;
+  _ = *xxtrt_(D)[_];
+  ival2_ = _ < _ ? ival2_ : ival2_;

but that does not result in any extra or missed vectorization.

Btw, dropping to -O2 also fixes the regression.

So, it's not at all clear what we are chasing here (the PRE seems to be
a partial antic expression).