[Bug c++/17642] [4.0 Regression] internal compiler error: in invert_truthvalue, at fold-const.c:2997

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

------- Additional Comments From mmitchel at gcc dot gnu dot org  2004-09-24 23:24 -------
Andrew's analysis is overly simplistic.

During the initial parsing of templates, we build up trees which do not obey the
GCC tying rules to represent the users input.  At template-substitution time,
they are converted so that they do follow those rules.

Unfortunately, we sometimes call "fold" during the initial parsing phase from
build_binary_op.  It's not clear if this is necessary.  It is clearly that
constant-expressions (like "6 - 4") must be resolved at template parse time. 
The addition of the assert in fold is causing the compiler to crash.  It is
quite probable that if the compiler merely did not assert that by the time this
code was ever presented to the gimplifier all the types would be correct.

In the long run, we want to avoid calling "fold" at all, except on integral
constant-expressions.  In all other cases, we should just let the optimizers do
their thing.  That would make the front end both simpler and faster.  However,
that's not pratical for 4.0.

I will experiment with turning off the call to "fold" in build_binary_op, but I
suspect that we many need to remove the assert.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17642