Seeing error in one test on ARM: org.openjdk.jcstress.tests.volatiles.ReadTwiceOverVolatileReadTest

[Aleksey Shipilev]

On 02/23/2017 02:42 PM, Andrew Haley wrote:
> On 23/02/17 10:19, Aleksey Shipilev wrote:
>> What I find alarming though, is that generated code has reads in actor2() in
>> order.
> 
> What exactly does this comment mean?

It means that in:

    int x;
    volatile int y;

    @Actor
    public void actor1() {
        x = 1;
        y = 1;
    }
    @Actor
    public void actor2(IntResult3 r) {
        r.r1 = x;
        r.r2 = y;
        r.r3 = x;
    }

actor() is generated like:

--- C2 ---:
   ldr     w0, [x1,#12]    ;*getfield x {reexecute=0 rethrow=0 return_oop=0}
   str     w0, [x2,#12]    ;*putfield r1 {reexecute=0 rethrow=0 return_oop=0}
   ldr     w0, [x1,#16]
   dmb     ishld           ;*getfield y {reexecute=0 rethrow=0 return_oop=0}
   str     w0, [x2,#16]    ;*putfield r2 {reexecute=0 rethrow=0 return_oop=0}
   ldr     w0, [x1,#12]    ;*getfield x {reexecute=0 rethrow=0 return_oop=0}
   str     w0, [x2,#20]    ;*putfield r3 {reexecute=0 rethrow=0 return_oop=0}

In this code, the reads of $x into r1 and r3 are in "program order". But this is
the read of the same variable, and they should be _coherent_ on ARM. The result
(1, 0, 0) is impossible then. So either this is a hardware bug, or there is some
other version of generated code where reads of $x have swapped.

Thanks,
-Aleksey