RFR(S): Implementation of aarch64 interpreter matrix barrier

[Andrew Haley]

On 21/06/17 09:14, Aleksey Shipilev wrote:
> On 06/21/2017 01:10 AM, Zhengyu Gu wrote:
>> Interpreter's matrix barrier is missing in aarch64 port.
>>
>> Webrev:
>> http://cr.openjdk.java.net/~zgu/shenandoah/aarch64_int_matrix_barrier/webrev.00/index.html
> 
> Looks okay to me.
> 
> You can move this block inside Shenandoah branch:
> 
>  180           // For Shenandoah, make sure we only store refs into to-space.
>  181           oopDesc::bs()->interpreter_read_barrier(_masm, val);
> 
> I would probably write the checks explicitly too, like this:
> 
> if (UseG1GC) {
>   ...
> } else if (UseShenandoahGC) {
>   ...
> } else {
>   ShouldNotReachHere();
> }

+void MacroAssembler::shenandoah_write_barrier_post(Register store_addr,
+                                                   Register new_val,
+                                                   Register thread,
+                                                   Register tmp,
+                                                   Register tmp2) {
+  assert(thread == rthread, "must be");
+  assert(UseShenandoahGC, "expect Shenandoah GC");
+
+  if (! UseShenandoahMatrix) {
+    // No need for that barrier if not using matrix.
+    return;
+  }
+
+  assert_different_registers(store_addr, new_val, thread, tmp, tmp2, rscratch2);
+
+  Label done;
+  cbz(new_val, done);
+
+  ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix();
+  address matrix_addr = matrix->matrix_addr();
+  movptr(rscratch2, (intptr_t)ShenandoahHeap::heap()->base());

If an address isn't going to be relocated it makes more sense to use
adrp, not movptr, here:

  unsigned long offset;
  adrp(rscratch2, ExternalAddress(ShenandoahHeap::heap()->base(), offset);
  lea(rscratch2, Address(rscratch2, offset));

The heap base should be page aligned, so the lea won't generate any code.

+  mov(tmp, new_val);
+  sub(tmp, tmp, rscratch2);

sub(tmp, new_val, rscratch2);

+  lsr(tmp, tmp, ShenandoahHeapRegion::region_size_shift_jint());
+
+  mov(tmp2, store_addr);
+  sub(tmp2, tmp2, rscratch2);

sub(tmp2, store_addr, rscratch2)

+  lsr(tmp2, tmp2, ShenandoahHeapRegion::region_size_shift_jint());
+
+  movptr(rscratch2, (intptr_t)matrix->stride_jint());

Don't use movptr here: it generates a patchable address, which you
don't need.  Use mov(reg, uint64_t) .

+  madd(tmp, tmp, rscratch2, tmp2);
+
+  movptr(rscratch2, (intptr_t)matrix_addr);

Use mov(dst, address) .

+  Address loc(rscratch2, tmp);
+
+  ldrb(tmp2, loc);
+  cbnz(tmp2, done);
+  mov(tmp2, 1);
+  strb(tmp2, loc);
+  bind(done);
+}
+

-- 
Andrew Haley
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671