[aarch64-port-dev ] Proposed patch for C2 volatile generation (resent)

[Andrew Haley]

Looks pretty good.

There is one remaining dmb(Assembler::SY).  Please make it
dmb(Assembler::ISH).

It would be super-nice also to fix getObjectVolatile and its friends
at the same time.  But you might prefer to address those with a
different patch.

compareAndSwapObject is really fugly.  I would love, LOVE, to see
it fixed to use ldaxr/stlxr.  :-)

I see this in ConcurrentHashMap::transfer:

  0x000003ffa51f5f68: dmb	ish
  0x000003ffa51f5f6c: lsr	x10, x0, #9
  0x000003ffa51f5f70: str	w17, [x0,#20]
  0x000003ffa51f5f74: adrp	x17, word_map_base  ;   {external_word}
  0x000003ffa51f5f78: strb	wzr, [x17,x10,lsl #0]
  0x000003ffa51f5f7c: dmb	ish             ;*putfield val
                                                ; - java.util.concurrent.ConcurrentHashMap$Node::<init>@16 (line 628)
                                                ; - java.util.concurrent.ConcurrentHashMap::transfer at 505 (line 2452)

  0x000003ffa51f5f80: dmb	ish
  0x000003ffa51f5f84: mov	x10, x5
  0x000003ffa51f5f88: lsr	x17, x0, #9
  0x000003ffa51f5f8c: str	w10, [x0,#24]
  0x000003ffa51f5f90: adrp	x10, word_map_base  ;   {external_word}
  0x000003ffa51f5f94: strb	wzr, [x10,x17,lsl #0]
  0x000003ffa51f5f98: dmb	ish             ;*putfield next
                                                ; - java.util.concurrent.ConcurrentHashMap$Node::<init>@22 (line 629)
                                                ; - java.util.concurrent.ConcurrentHashMap::transfer at 505 (line 2452)

  0x000003ffa51f5f9c: dmb	ish             ;*aload
                                                ; - java.util.concurrent.ConcurrentHashMap::transfer at 510 (line 2447)

That does look like a lot of DMBs!

In fact, much of ConcurrentHashMap::transfer looks pretty grim.  Many
redundant barriers.  I think this is all to do with the mix of
Unsafe operations and volatile fields.

There's a bug in C2 where elided constructors have a memory barrier which
doesn't get elided.  That bug is supposed to have been fixed, but I
can still see its effect.  It's unrelated to what you're doing here,
but it does lead to a lot of mysterious DMBs.

Andrew.