On Wed, 14 May 2014, Andrey Belevantsev wrote:
> Hello, > > As described in the PR in more details, this ICE is about the scheduler not > being able to initialize its data structures for the new unconditional jump > created when redirecting an edge and simplifying control flow. It so happens > that the new jump is the only unscheduled instruction being left so we need to > pass some of its data from the old jump. The difficulty is to support only > this particular situation to retain the existing checking code and to pass > down the required information to the point where it is needed. > > I assume that we are able to use default parameters as I haven't heard > otherwise. Default parameters are used now in several functions such as 'build_assign', 'c_build_vec_perm_expr', so that should be fine. > Bootstrapped and tested on x86-64, OK for trunk and branches? Branches should > also be safe as we're fixing only very specific corner case (hey, we regressed > four releases and nobody noticed until now). OK for trunk and 4.8, but not applicable for 4.7 since that branch does not use C++ as implementation language. Alexander