[Bug debug/67355] [5 Regression] ICE compiling LTP testcase, endless cselib recursion from var-tracking

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67355

--- Comment #9 from Alexandre Oliva <aoliva at gcc dot gnu.org> ---
Author: aoliva
Date: Sun Dec 13 20:37:26 2015
New Revision: 231599

URL: https://gcc.gnu.org/viewcvs?rev=231599&root=gcc&view=rev
Log:
[PR67355] drop dummy zero from reverse VTA ops, fix infinite recursion

VTA's cselib expression hashing compares expressions with the same
hash before adding them to the hash table.  When there is a collision
involving a self-referencing expression, we could get infinite
recursion, in spite of the cycle breakers already in place.  The
problem is currently latent in the trunk, because by chance we don't
get a collision.

Such value cycles are often introduced by reverse_op; most often,
they're indirect, and then value canonicalization takes care of the
cycle, but if the reverse operation simplifies to the original value,
we used to issue a (plus V (const_int 0)), because at some point
adding a plain value V to a location list as a reverse_op equivalence
caused other problems.

This dummy zero, in turn, caused the value canonicalizer to not fully
realize the equivalence, leading to more complex graphs and,
occasionally, to infinite recursion when comparing such
value-plus-zero expressions recursively.

Simply using V solves the infinite recursion from the PR testcase,
since the extra equivalence and the preexisting value canonicalization
together prevent recursion while the unrecognized equivalence
wouldn't, but it exposed another infinite recursion in
memrefs_conflict_p: get_addr had a cycle breaker in place, to skip RTL
referencing values introduced after the one we're examining, but it
wouldn't break the cycle if the value itself appeared in the
expression being examined.

After removing the dummy zero above, this kind of cycle in the
equivalence graph is no longer introduced by VTA itself, but dummy
zeros are also present in generated code, such as in the 32-bit x86's
pro_epilogue_adjust_stack_si_add epilogue insn generated as part of
the builtin longjmp in _Unwind_RaiseException building libgcc's
unwind-dw2.o.  So, break the recursion cycle for them too.


for  gcc/ChangeLog

        PR debug/67355
        * var-tracking.c (reverse_op): Don't add dummy zero to reverse
        ops that simplify back to the original value.
        * alias.c (refs_newer_value_p): Cut off recursion for
        expressions containing the original value.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/alias.c
    trunk/gcc/var-tracking.c