Happy path

[Roland Westrelin]

>  *) On the ye-old SpscQueue benchmark, and this is how it looks with the patch:
>  http://cr.openjdk.java.net/~shade/shenandoah/happy-path/happypath-1.perfasm
>
>  It does coalescing well on non-membar paths. There are things to improve, e.g. back-to-back tests:
>
>   0.00%    0.02%   │╰│││  0x00007f8ab0b717f2: jmpq   0x00007f8ab0b7176f
>   1.88%    2.40%   ↘ │││  0x00007f8ab0b717f7: cmpb   $0x0,0x400(%r15)
>   0.21%    0.31%     │││  0x00007f8ab0b717ff: jne    0x00007f8ab0b71871
>   0.00%    0.00%     │││  0x00007f8ab0b71801: cmpb   $0x0,0x400(%r15)
>   0.67%    0.82%     │││  0x00007f8ab0b71809: jne    0x00007f8ab0b71889
>   1.88%    2.15%     │││  0x00007f8ab0b7180b: mov    %r12d,0x10(%rcx,%r14,4)
>   2.19%    2.27%     │││  0x00007f8ab0b71810: cmpb   $0x0,0x400(%r15)
>   0.01%    0.00%     ╰││  0x00007f8ab0b71818: je     0x00007f8ab0b71730
>                       ││  0x00007f8ab0b7181e: testb  $0x34,0x400(%r15)
>                       ╰│  0x00007f8ab0b71826: je     0x00007f8ab0b71720

I think they are caused by membars that don't show up in the generated
code because they are nop on x86. Anyway, we can also clone membars:

http://cr.openjdk.java.net/~roland/shenandoah/happypath-membars/webrev.00/

with the hack below.

Roland.

diff --git a/src/hotspot/share/opto/loopopts.cpp b/src/hotspot/share/opto/loopopts.cpp
--- a/src/hotspot/share/opto/loopopts.cpp
+++ b/src/hotspot/share/opto/loopopts.cpp
@@ -1078,8 +1078,8 @@
         if (m->is_FastLock())
           return false;
 #ifdef _LP64
-        if (m->Opcode() == Op_ConvI2L)
-          return false;
+        // if (m->Opcode() == Op_ConvI2L)
+        //   return false;
         if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
           return false;
         }