More hotspot tests and some backported fixes

[Andrew John Hughes]

On 13:20 Fri 26 Sep     , Mark Wielaard wrote:
> Hi,
> 
> I noticed openjdk6 and openjdk7 seem to have a mutually exclusive set of
> tests under hotspot/test/compiler. Since I believe having more tests is
> a good thing I merged them in icedtea6. That uncovered two crasher bugs
> in the hotspot form openjdk6 that we use. So I tracked down the fixes
> for those and backported them (patches attached). All 20 tests now PASS.
> 
> It is probably a good idea to also merge the tests from 6 to 7. But I
> haven't done that yet.
> 
> Note that again I did only backport for the hotspot from openjdk6, not
> the one that zero/shark uses (7b24).
> 
> Cheers,
> 
> Mark

> 6700047: C2 failed in idom_no_update
> Summary: partial peeling shouldn't place clones into loop
> 
> diff -r 4aebfff4f8a2 hotspot/src/share/vm/opto/loopopts.cpp
> --- openjdk.orig/hotspot/src/share/vm/opto/loopopts.cpp	Mon Sep 15 11:38:34 2008 +0200
> +++ openjdk/hotspot/src/share/vm/opto/loopopts.cpp	Fri Sep 26 11:29:06 2008 +0200
> @@ -1865,18 +1865,19 @@
>      _igvn.hash_delete(use);
>      use->set_req(j, n_clone);
>      _igvn._worklist.push(use);
> +    Node* use_c;
>      if (!use->is_Phi()) {
> -      Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
> -      set_ctrl(n_clone, use_c);
> -      assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
> -      get_loop(use_c)->_body.push(n_clone);
> +      use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
>      } else {
>        // Use in a phi is considered a use in the associated predecessor block
> -      Node *prevbb = use->in(0)->in(j);
> -      set_ctrl(n_clone, prevbb);
> -      assert(!loop->is_member(get_loop(prevbb)), "should be outside loop");
> -      get_loop(prevbb)->_body.push(n_clone);
> +      use_c = use->in(0)->in(j);
>      }
> +    if (use_c->is_CountedLoop()) {
> +      use_c = use_c->in(LoopNode::EntryControl);
> +    }
> +    set_ctrl(n_clone, use_c);
> +    assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
> +    get_loop(use_c)->_body.push(n_clone);
>      _igvn.register_new_node_with_optimizer(n_clone);
>  #if !defined(PRODUCT)
>      if (TracePartialPeeling) {

> 6712835: Server compiler fails with assertion (loop_count < K,"infinite loop in PhaseIterGVN::transform")
> 
> diff -r 4aebfff4f8a2 hotspot/src/share/vm/opto/ifnode.cpp
> --- openjdk.orig/hotspot/src/share/vm/opto/ifnode.cpp	Mon Sep 15 11:38:34 2008 +0200
> +++ openjdk/hotspot/src/share/vm/opto/ifnode.cpp	Fri Sep 26 11:56:32 2008 +0200
> @@ -569,6 +569,11 @@
>    int true_path = phi->is_diamond_phi();
>    if( true_path == 0 ) return NULL;
>  
> +  // Make sure that iff and the control of the phi are different. This
> +  // should really only happen for dead control flow since it requires
> +  // an illegal cycle.
> +  if (phi->in(0)->in(1)->in(0) == iff) return NULL;
> +
>    // phi->region->if_proj->ifnode->bool->cmp
>    BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool();
>    
> @@ -595,6 +600,7 @@
>    }
>  
>    Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2);
> +  assert(new_bol != iff->in(1), "must make progress");
>    iff->set_req(1, new_bol);
>    // Intervening diamond probably goes dead
>    phase->C->set_major_progress();

Given that the b24 patch for zero only does a partial replacement of HotSpot
(the old make and test directories) remain, the work probably still applies.
On this note, though, I think it best if, instead of maintaining this patch,
we download the last OpenJDK7 HotSpot for a zero build and use this.  This
also ensures zero continues working with 7 at no extra cost.

If this seems a worthwhile proposal, I'll look into it over the next few
days.  This would of course mean that the zero HotSpot then has all the
tests and fixes found in the OpenJDK7 one, instead of using a rapidly aging
snapshot (b24 is 24 weeks old, at a rough guess - there's a build drop every
two weeks and we are now on 36 - so about six months...)
-- 
Andrew :)

Support Free Java!
Contribute to GNU Classpath and the OpenJDK
http://www.gnu.org/software/classpath
http://openjdk.java.net
PGP Key: 94EFD9D8 (http://subkeys.pgp.net)
Fingerprint = F8EF F1EA 401E 2E60 15FA  7927 142C 2591 94EF D9D8