[Bug c/63307] New: [4.9/5 Regression] Cilk+ breaks -fcompare-debug bootstrap

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63307

            Bug ID: 63307
           Summary: [4.9/5 Regression] Cilk+ breaks -fcompare-debug
                    bootstrap
           Product: gcc
           Version: 4.9.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
                CC: izamyatin at gmail dot com, kyukhin at gcc dot gnu.org

I've just tried a -fcompare-debug bootstrap, and one of the two reasons (other
was Ada) that it failed is Cilk+:
../configure --enable-languages=all,ada,obj-c++,lto,go
--enable-checking=release; GCC_COMPARE_DEBUG=1 make -j48 bootstrap > LOG 2>&1
&& GCC_COMPARE_DEBUG=1 make -j48 -k check > LOGC 2>&1; ../contrib/test_summary
> LOGT 2>&1
...
xg++: error: ../../../libcilkrts/runtime/cilk-abi-cilk-for.cpp: -fcompare-debug
failure (length)

>From what I can see, the bug is (at least) in c-family/cilk.c, where it seems
in all 4 traverse handlers on decl_map it creates decls, adds fields etc.

That is a big NO NO in GCC, hash table traversal order should never affect code
generation.  So, seeing changes like:
-          _cilk_spn_1 (low, grain, mid, body, loop_root_pedigree, data, w);
+          _cilk_spn_1 (data, body, loop_root_pedigree, w, mid, low, grain);
or:
-<built-in> (unsigned int D.3190, int D.3189, unsigned int D.3188, void
(*<T39b>) (void *, unsigned int, unsigned int) D.3187, struct
__cilkrts_pedigree * D.3186, void * D.3185, struct __cilkrts_worker * D.3184)
+<built-in> (void * D.3190, void (*<T39b>) (void *, unsigned int, unsigned int)
D.3189, struct __cilkrts_pedigree * D.3188, struct __cilkrts_worker * D.3187,
unsigned int D.3186, unsigned int D.3185, int D.3184)
in *.gimple dump between -g and -g0 is just wrong (and in this case it isn't
really debug related, just different order in the hash map).

If you don't have the decls (but, apparently you put also BLOCKs in there, what
else?) you want to traverse in some other data structure, you can e.g. traverse
the hash table but only collect the decls (do you need anything else, e.g. the
BLOCKs?, types, etc.) you saw into some vector, then qsort the vector (for
decls e.g. by DECL_UID, if you have also other trees, it might be harder) and
then traverse the sorted vector insert of traversing random order hash map.