fixing -fsplit-stack

[Thomas Neumann via Gcc]

Hi,

I would like to use -fsplit-stack for some recursion heavy C++ code, but 
I noticed that it currently has severe issues, which prevents practical 
usage, but which are fixable: 1) currently a stack switch costs about 
167ns due to signal blocking, which is prohibitive in hot loops 2) the 
stack computation for signal handlers is wrong, which leads to crashes 
on machines with AVX-512 support, and 3) a signaler handler with split 
stack causes a crash if the signal happens during exception unwinding.

I have prototype fixes for all three (x86-64 only for now). By making 
__morestack re-entrant safe we can avoid signal blocking, bringing stack 
switch cost to about 5ns, which fixes the hot loop problem. The other 
two are largely classical bug fixes.

Before working on this more, are you actually interested in reviving 
this functionality? I got the impression that the code is hardly used at 
all (except for go, which uses a separate signal stack and thus avoids 
some of these problems). But I think it would be useful to make it work, 
as it makes recursive functions much more practical.

Thomas

PS: I noticed that alloca/VLA with split stack is broken, too, because 
the memory allocated by __morestack_allocate_stack_space is not freed 
properly. I don't care about this too much, and fixing it requires 
compiler changes, thus I did not try to fix it. But perhaps it should be 
addressed, too.