Jitted array-length checks before entering a loop
Vitaly Davidovich
vitalyd at gmail.com
Fri Oct 9 20:14:58 UTC 2015
So basically, the deopt blobs are unreachable -- not sure if that's
intentional or not. Perhaps it's intentional in that deopt'ing in this
case is a waste of time (the simple exit is just fine), but the code for it
is left behind.
On Fri, Oct 9, 2015 at 2:51 PM, Vitaly Davidovich <vitalyd at gmail.com> wrote:
> Chris,
>
> Below is a sample dump (array is in %rsi). You can see the 2 tests in
> question that jump to deopt blobs (I presume). If I change the code to
> always pass a zero length array (the below assembly is for non-empty array
> 100% of the time), the code generation changes to favor that case, and
> these tests are not present.
>
> 0x00007f63476c7da0: mov %eax,-0x14000(%rsp)
> 0x00007f63476c7da7: push %rbp
> 0x00007f63476c7da8: sub $0x10,%rsp ;*synchronization entry
>
> 0x00007f63476c7dac: mov 0xc(%rsi),%r10d ;*arraylength
>
> ; implicit exception:
> dispatches to 0x00007f63476c7e11
> 0x00007f63476c7db0: test %r10d,%r10d
> 0x00007f63476c7db3: jle 0x00007f63476c7de2 ;*if_icmplt
>
> 0x00007f63476c7db5: test %r10d,%r10d
> 0x00007f63476c7db8: jbe 0x00007f63476c7dfd
> 0x00007f63476c7dba: mov %r10d,%r8d
> 0x00007f63476c7dbd: dec %r8d
> 0x00007f63476c7dc0: cmp %r10d,%r8d
> 0x00007f63476c7dc3: jae 0x00007f63476c7dfd
> 0x00007f63476c7dc5: xor %r11d,%r11d
> 0x00007f63476c7dc8: nopl 0x0(%rax,%rax,1) ;*aload_0
>
> 0x00007f63476c7dd0: mov 0x10(%rsi,%r11,4),%r8d ;*aaload
>
> 0x00007f63476c7dd5: test %r8d,%r8d
> 0x00007f63476c7dd8: je 0x00007f63476c7dee ;*invokevirtual amethod
>
> 0x00007f63476c7dda: inc %r11d ;*iinc
>
> 0x00007f63476c7ddd: cmp %r10d,%r11d
> 0x00007f63476c7de0: jl 0x00007f63476c7dd0 ;*return
>
> 0x00007f63476c7de2: add $0x10,%rsp
> 0x00007f63476c7de6: pop %rbp
> 0x00007f63476c7de7: test %eax,0x5cd3213(%rip) #
> 0x00007f634d39b000
> ; {poll_return}
> 0x00007f63476c7ded: retq
> 0x00007f63476c7dee: mov $0xfffffff6,%esi
> 0x00007f63476c7df3: callq 0x00007f6347684320 ; OopMap{off=88}
> ;*invokevirtual amethod
> ; {runtime_call}
> 0x00007f63476c7df8: callq 0x00007f634c17f570 ;*invokevirtual amethod
> ; {runtime_call}
> 0x00007f63476c7dfd: mov %rsi,%rbp
> 0x00007f63476c7e00: mov $0xffffff86,%esi
> 0x00007f63476c7e05: xchg %ax,%ax
> 0x00007f63476c7e07: callq 0x00007f6347684320 ; OopMap{rbp=Oop off=108}
> ;*aload_0
> ; {runtime_call}
> 0x00007f63476c7e0c: callq 0x00007f634c17f570 ;*aload_0
> ; {runtime_call}
> 0x00007f63476c7e11: mov $0xfffffff6,%esi
> 0x00007f63476c7e16: nop
> 0x00007f63476c7e17: callq 0x00007f6347684320 ; OopMap{off=124}
> ;*arraylength
> ; {runtime_call}
> 0x00007f63476c7e1c: callq 0x00007f634c17f570 ;*arraylength
> ; {runtime_call}
>
>
> On Fri, Oct 9, 2015 at 2:31 PM, Christian Thalinger <
> christian.thalinger at oracle.com> wrote:
>
>>
>> On Sep 28, 2015, at 3:37 AM, Nassim Halli <nassim.halli at gmail.com> wrote:
>>
>> Hi guys,
>>
>> I have a question relative to some code generated by the C2 compiler.
>>
>> When I look at the assembly codes for this method:
>>
>> static void compiledMethod (Atype[] data) {
>> int n = data.length
>> for (int i = 0; i < n; ++i)
>> data[i].amethod();
>> }
>>
>> I get the this assembly on Linux X86_64 before entering the loop (not
>> OSR):
>>
>> # data is in rbx
>> mov 0xc(%rbx),%r9d
>>
>> # data.length is in r9d
>> test %r9d,%r9d
>> jle END_OF_LOOP
>>
>> test %r9d,%r9d
>> jbe BB0
>>
>> mov %r9d,%r11d
>> dec %r11d
>> cmp %r9d,%r11d
>> jae BB0
>>
>>
>> Did you leave out any assembly? Where is BB0?
>>
>>
>> It seems to me the second and third check are useless and the
>> corresponding branches are never taken.
>> Could you please tell me more about these checks, If and why are they
>> required ?
>>
>> Thanks a lot.
>>
>> Nassim H.
>>
>>
>> *Test conditions:*
>>
>> java version "1.8.0_51"
>> Java(TM) SE Runtime Environment (build 1.8.0_51-b16)
>> Java HotSpot(TM) 64-Bit Server VM (build 25.51-b03, mixed mode)
>> Default options.
>> Linux, Intel Sandy Bridge core i5-2000
>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20151009/78c1da42/attachment.html>
More information about the hotspot-compiler-dev
mailing list