So I managed to re-introduce PR41043 by implementing the
fold_unary (T1)(X * Y) -> (T1)X * (T1)Y pattern on GIMPLE.
That is because it re-applies recursively on a large
multiplication tree and because if there are multiple uses
of SSA names doing that will expand (aka un-CSE) the
multiplication tree. Oops. The following restricts it
to at most one SSA name or two single-use SSA names.
In the end this kind of demotion should be done by a
pass, not by expression simplification (it's not really
a simplification after all).
Applied.
Richard.
2014-09-11 Richard Biener <rguent...@suse.de>
PR middle-end/41043
* match-conversions.pd ((T1)(X * Y) -> (T1)X * (T1)Y): Restrict
to a single or single-use SSA names.
Index: gcc/match-conversions.pd
===================================================================
--- gcc/match-conversions.pd (revision 215110)
+++ gcc/match-conversions.pd (working copy)
@@ -56,7 +56,15 @@
(convert (mult @0 @1))
(if (INTEGRAL_TYPE_P (type)
&& INTEGRAL_TYPE_P (TREE_TYPE (@0))
- && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (@0)))
+ && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (@0))
+ /* ??? These kind of patterns are a bad idea - see PR41043. We
+ create a lot of redundant statements if operands are used multiple
+ times. Maybe we want a flag for this. But eventually these
+ kind of transforms should be done in a pass. */
+ && (GENERIC
+ || TREE_CODE (@0) != SSA_NAME || TREE_CODE (@1) != SSA_NAME
+ || ((TREE_CODE (@0) != SSA_NAME || has_single_use (@0))
+ && (TREE_CODE (@1) != SSA_NAME || has_single_use (@1)))))
(if (TYPE_OVERFLOW_WRAPS (type))
(mult (convert @0) (convert @1)))
(with { tree utype = unsigned_type_for (type); }
