Please review fix for 7120481 storeStore barrier in constructor with final field

[Jiangli Zhou]

Hi Vitaly,

Thanks for the comments! The LIR_Assembler APIs only supports membar(), 
membar_acquire() and membar_release() currently. Looks like we need to 
add more APIs and LIR_Code in order to support more fine grained memory 
barriers in compiled code for ARM. I will make the change. The ARM 
interpreter is using the correct barrier as we can choose the specific 
one to use.

Thanks,

Jiangli


On 02/08/2012 04:06 PM, Vitaly Davidovich wrote:
>
> Hi Jiangli,
>
> Isn't the "sy" option a full membar? I think the "st" option is 
> equivalent to StoreStore.
>
> Vitaly
>
> Sent from my phone
>
> On Feb 8, 2012 6:37 PM, "Jiangli Zhou" <jiangli.zhou at oracle.com 
> <mailto:jiangli.zhou at oracle.com>> wrote:
>
>     Hi,
>
>     Please review the following webrevs for 7120481 storeStore barrier
>     in constructor with final field:
>
>     http://javaweb.sfbay.sun.com/~jianzhou/webrev.storestore/
>     <http://javaweb.sfbay.sun.com/%7Ejianzhou/webrev.storestore/>
>     http://javaweb.sfbay.sun.com/~jianzhou/webrev.storestore.closed/
>     <http://javaweb.sfbay.sun.com/%7Ejianzhou/webrev.storestore.closed/>
>
>     Both c1 and the template interpreter are fixed to add the needed
>     memory barrier. For a constructor that writes final field, a
>     release barrier is added in the compiled method before the
>     constructor returns if the constructor is not inlined. When the
>     constructor is inlined, a release barrier is added at the end of
>     the inlined constructor. The changes are done in the shared c1
>     code only, the platform specific code already knows whether memory
>     barrier instructions should be generated for the platform.
>     Following are ARM examples of a compiled constructor that writes
>     final field after the fix. For the template interpreter, a new
>     opcode '_return_with_membar' is added. During method rewriting,
>     the VM detects if a constructor writes final. If so, the _return
>     opcode is changed into _return_with_membar for the constructor.
>     The ARM and PPC template interpreter code are changed to issue a
>     storesotre barrier for the new opcode.
>
>     Not inlined:
>
>       0x41386450: mov    r1, r0
>       0x41386454: mov    r0, r1          ;*invokespecial <init>
>                                             ; - HasFinal::<init>@1
>     (line 19)
>       0x41386458: str    r1, [sp, #16]
>       0x4138645c: bl    0x41325d00      ; OopMap{[16]=Oop off=48}
>                                             ;*invokespecial <init>
>                                             ; - HasFinal::<init>@1
>     (line 19)
>                                             ;   {optimized virtual_call}
>       0x41386460: mov    r0, #100    ; 0x64
>       0x41386464: ldr    r1, [sp, #16]
>       0x41386468: str    r0, [r1, #8]    ;*putfield x
>                                             ; - HasFinal::<init>@7
>     (line 20)
>     0x4138646c: dmb    sy
>       0x41386470: add    sp, sp, #24
>       0x41386474: pop    {fp, lr}
>       0x41386478: movw    ip, #53248    ; 0xd000
>       0x4138647c: movt    ip, #16397    ; 0x400d
>       0x41386480: ldr    ip, [ip]        ;   {poll_return}
>       0x41386484: bx    lr
>       0x41386488: nop            ; (mov r0, r0)
>
>
>     Inlined:
>
>       0x413797e0: mov    r1, r0
>       0x413797e4: mov    r0, r1          ;*invokespecial <init>
>                                             ; - HasFinal::<init>@1
>     (line 19)
>                                             ; - FinalTest::main at 25
>     (line 11)
>       0x413797e8: str    r1, [sp, #32]
>       0x413797ec: bl    0x41318d00      ; OopMap{[32]=Oop off=272}
>                                             ;*invokespecial <init>
>                                             ; - HasFinal::<init>@1
>     (line 19)
>                                             ; - FinalTest::main at 25
>     (line 11)
>                                             ;   {optimized virtual_call}
>       0x413797f0: mov    r0, #100    ; 0x64
>       0x413797f4: ldr    r1, [sp, #32]
>       0x413797f8: str    r0, [r1, #8]    ;*putfield x
>                                             ; - HasFinal::<init>@7
>     (line 20)
>                                             ; - FinalTest::main at 25
>     (line 11)
>     0x413797fc: dmb    sy
>       0x41379800: movw    r0, #23416    ; 0x5b78
>                                             ;   {oop(a
>     'java/lang/Class' = 'FinalTest')}
>       0x41379804: movt    r0, #17206    ; 0x4336
>       0x41379808: str    r1, [r0, #104]    ; 0x68
>
>     Thanks,
>
>     Jiangli
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20120208/9903e9ae/attachment-0001.html