aarch64: C2 fast lock/unlock issues

[Andrew Haley]

I'm rejecting this patch.

With biased locking there is path through the code with no barriers of
any kind.

This code:

        for (int i = 0; i < 10000; i++) {
            synchronized(this) { n++; }
        }

generates this:

 ;; B5: #	B17 B6 <- B4 B8 	Loop: B5-B8 inner main of N92 Freq: 10082.5

  0x000003ffa8496e90: ldr	x10, [x19]
  0x000003ffa8496e94: and	x11, x10, #0x7
  0x000003ffa8496e98: cmp	x11, #0x5
  0x000003ffa8496e9c: b.ne	0x000003ffa8496f74
 ;; B6: #	B21 B7 <- B5  Freq: 10082.5

  0x000003ffa8496ea0: ldr	x11, [xesp,#176]
  0x000003ffa8496ea4: mov	xmethod, xthread
  0x000003ffa8496ea8: orr	xmethod, xmethod, x11
  0x000003ffa8496eac: eor	x13, xmethod, x10
  0x000003ffa8496eb0: and	x14, x13, #0xffffffffffffff87
  0x000003ffa8496eb4: cbnz	x14, 0x000003ffa8497098
                                                ;*monitorenter
                                                ; - Twiddle::run at 12 (line 7)

 ;; B7: #	B19 B8 <- B18 B17 B6 B24  Freq: 10082.5

  0x000003ffa8496eb8: ldr	w11, [x19,#12]  ;*getfield n
                                                ; - Twiddle::run at 15 (line 7)

  0x000003ffa8496ebc: add	w10, w11, #0x1
  0x000003ffa8496ec0: add	w12, w11, #0x4
  0x000003ffa8496ec4: str	w10, [x19,#12]
  0x000003ffa8496ec8: str	w12, [x19,#12]
  0x000003ffa8496ecc: ldr	x10, [x19]
  0x000003ffa8496ed0: and	x10, x10, #0x7
  0x000003ffa8496ed4: cmp	x10, #0x5
  0x000003ffa8496ed8: b.ne	0x000003ffa8496ffc  ;*aload_0
                                                ; - Twiddle::run at 9 (line 7)

 ;; B8: #	B5 B9 <- B20 B19 B7  Freq: 10082.5

  0x000003ffa8496edc: add	w29, w29, #0x4  ;*iinc
                                                ; - Twiddle::run at 33 (line 6)

 ;; 0x270D
  0x000003ffa8496ee0: mov	wscratch1, #0x270d                	// #9997
  0x000003ffa8496ee4: cmp	w29, wscratch1
  0x000003ffa8496ee8: b.lt	0x000003ffa8496e90  ;*if_icmpge
                                                ; - Twiddle::run at 6 (line 6)

So any other thread waiting for the counter to hit a value would not
see it.  And we do not see the writes of other threads.  This is
wrong.

(It's also, IIRC, why I rejected this approach before.)

Andrew.