XCHG is slow ?

Fri Jul 6 05:42:48 PDT 2012

Hi Joseph,

As far as I know, lock is automatically applied only if a memory operand is
present; reg to reg should not.  As mentioned earlier, xchg eax eax
translates to same thing as nop - reg can be used to change instruction
size for alignment purposes.

Cheers,
Vitaly

Sent from my phone
On Jul 6, 2012 3:20 AM, "changren" <changren at taobao.com> wrote:

> I noticed this
>
> 0x00007fdf91061644: xchg %ax,%ax ;...666690
>
> on modern Intel and AMD cpus, it is recommended to use 0F 1F 00H for 3
> nops padding. 66 66 90H is for older AMD cpus.
> I saw a related comment regarding the change in assembler_x86.cpp
>
> Don't use "0x0F 0x1F 0x00" - need patching safe padding
>
> what does this mean? if it is related to ax register, we can choose
> another register with oldest value, for instance, 0F 1F 01 or choose
> other instruction, for instance, lea.
> BTW, @Vitaly lock is automatically applied to xchg.
>
> Thanks,
> Joseph
>
>
> ÓÚ 2012-7-4 9:30, Vladimir Kozlov Ð´µÀ:
>
>> I agree with you that it is strange. I looked on code and, as you
>> said, one with overflow detection executes a lot more instructions and
>> stack operations. What is time difference? I ran latest 7u6 (linux64)
>> promoted build and got this numbers:
>>
>> $ bin/jruby -X+C bench/bench_fib_recursive.rb
>> 0.170000 0.070000 0.240000 ( 0.100000)
>> 0.030000 0.020000 0.050000 ( 0.036000)
>> 0.040000 0.020000 0.060000 ( 0.036000)
>> 0.040000 0.030000 0.070000 ( 0.036000)
>> 0.040000 0.020000 0.060000 ( 0.036000)
>>
>> Vladimir
>>
>> R¨¦mi Forax wrote:
>>
>>> On 07/04/2012 01:44 AM, Vladimir Kozlov wrote:
>>>
>>>> R¨¦mi,
>>>>
>>>> NOP and xchg %ax,%ax are and always were the same instruction and
>>>> have the same encoding 0x90. What you see is multibytes nop (2 bytes
>>>> in your case) to align following call instruction: 0x66 0x90.
>>>>
>>>> Vladimir
>>>>
>>>
>>> I see and I suppose that a multibytes nop is not slower than a single
>>> byte nop.
>>>
>>> So to explain my problem clearly, I've two codes the first one is
>>> fibonacci (ClassicFibo)
>>> with an int, the second one is fibonacci with an int plus the
>>> overflow detection (FiboSample).
>>>
>>> My problem is that the code with the overflow detection is faster the
>>> code
>>> without the overflow detection :)
>>>
>>> R¨¦mi
>>>
>>> # {method} 'fibo' '(I)I' in 'ClassicFibo'
>>> # parm0: rsi = int
>>> # [sp+0x30] (sp of caller)
>>> 0x00007fdf91061620: mov %eax,-0x14000(%rsp) ;...89842400 c0feff
>>> 0x00007fdf91061627: push %rbp ;...55
>>> 0x00007fdf91061628: sub $0x20,%rsp ;...4883ec20
>>> ;*synchronization entry
>>> ; - ClassicFibo::fibo at -1 (line 4)
>>> 0x00007fdf9106162c: mov %esi,%ebp ;...8bee
>>> 0x00007fdf9106162e: cmp $0x2,%esi ;...83fe02
>>> 0x00007fdf91061631: jl 0x00007fdf91061651 ;...7c1e
>>> ;*if_icmpge
>>> ; - ClassicFibo::fibo at 2 (line 4)
>>> 0x00007fdf91061633: dec %esi ;...ffce
>>> ;*isub
>>> ; - ClassicFibo::fibo at 9 (line 7)
>>> 0x00007fdf91061635: xchg %ax,%ax ;...6690
>>> 0x00007fdf91061637: callq 0x00007fdf91038060 ;...e8246afd ff
>>> ; OopMap{off=28}
>>> ;*invokestatic fibo
>>> ; - ClassicFibo::fibo at 10 (line 7)
>>> ; {static_call}
>>> 0x00007fdf9106163c: mov %eax,(%rsp) ;...890424
>>> 0x00007fdf9106163f: mov %ebp,%esi ;...8bf5
>>> 0x00007fdf91061641: add $0xfffffffffffffffe,%esi ;...83c6fe
>>> ;*isub
>>> ; - ClassicFibo::fibo at 15 (line 7)
>>> 0x00007fdf91061644: xchg %ax,%ax ;...666690
>>> 0x00007fdf91061647: callq 0x00007fdf91038060 ;...e8146afd ff
>>> ; OopMap{off=44}
>>> ;*invokestatic fibo
>>> ; - ClassicFibo::fibo at 16 (line 7)
>>> ; {static_call}
>>> 0x00007fdf9106164c: add (%rsp),%eax ;...030424
>>> ;*iadd
>>> ; - ClassicFibo::fibo at 19 (line 7)
>>> 0x00007fdf9106164f: jmp 0x00007fdf91061656 ;...eb05
>>> 0x00007fdf91061651: mov $0x1,%eax ;...b8010000 00
>>> 0x00007fdf91061656: add $0x20,%rsp ;...4883c420
>>> 0x00007fdf9106165a: pop %rbp ;...5d
>>> 0x00007fdf9106165b: test %eax,0x91c899f(%rip) # 0x00007fdf9a22a000
>>> ;...85059f89 1c09
>>> ; {poll_return}
>>> 0x00007fdf91061661: retq ;...c3
>>> ;*invokestatic fibo
>>> ; - ClassicFibo::fibo at 10 (line 7)
>>> 0x00007fdf91061662: mov %rax,%rsi ;...488bf0
>>> 0x00007fdf91061665: jmp 0x00007fdf9106166a ;...eb03
>>> 0x00007fdf91061667: mov %rax,%rsi ;...488bf0
>>> ;*invokestatic fibo
>>> ; - ClassicFibo::fibo at 16 (line 7)
>>> 0x00007fdf9106166a: add $0x20,%rsp ;...4883c420
>>> 0x00007fdf9106166e: pop %rbp ;...5d
>>> 0x00007fdf9106166f: jmpq 0x00007fdf91061220 ;...e9acfbff ff
>>> ; {runtime_call}
>>> 0x00007fdf91061674: hlt ;...f4
>>> 0x00007fdf91061675: hlt ;...f4
>>> 0x00007fdf91061676: hlt ;...f4
>>> 0x00007fdf91061677: hlt ;...f4
>>> 0x00007fdf91061678: hlt ;...f4
>>> 0x00007fdf91061679: hlt ;...f4
>>> 0x00007fdf9106167a: hlt ;...f4
>>> 0x00007fdf9106167b: hlt ;...f4
>>> 0x00007fdf9106167c: hlt ;...f4
>>> 0x00007fdf9106167d: hlt ;...f4
>>> 0x00007fdf9106167e: hlt ;...f4
>>> 0x00007fdf9106167f: hlt ;...f4
>>> [Stub Code]
>>> 0x00007fdf91061680: mov $0x0,%rbx ;...48bb0000 000000
>>> ;...000000
>>> ; {no_reloc}
>>> 0x00007fdf9106168a: jmpq 0x00007fdf9106168a ;...e9fbffff ff
>>> ; {runtime_call}
>>> 0x00007fdf9106168f: mov $0x0,%rbx ;...48bb0000 000000
>>> ;...000000
>>> ; {static_stub}
>>> 0x00007fdf91061699: jmpq 0x00007fdf91061699 ;...e9fbffff ff
>>> ; {runtime_call}
>>> [Exception Handler]
>>> 0x00007fdf9106169e: jmpq 0x00007fdf9105e4a0 ;...e9fdcdff ff
>>> ; {runtime_call}
>>> [Deopt Handler Code]
>>> 0x00007fdf910616a3: callq 0x00007fdf910616a8 ;...e8000000 00
>>> 0x00007fdf910616a8: subq $0x5,(%rsp) ;...48832c24 05
>>> 0x00007fdf910616ad: jmpq 0x00007fdf91038c00 ;...e94e75fd ff
>>> ; {runtime_call}
>>> 0x00007fdf910616b2: hlt ;...f4
>>> 0x00007fdf910616b3: hlt ;...f4
>>> 0x00007fdf910616b4: hlt ;...f4
>>> 0x00007fdf910616b5: hlt ;...f4
>>> 0x00007fdf910616b6: hlt ;...f4
>>> 0x00007fdf910616b7: hlt ;...f4
>>>
>>>
>>> # {method} 'fibo' '(I)I' in 'FiboSample'
>>> # parm0: rsi = int
>>> # [sp+0x30] (sp of caller)
>>> 0x00007faadd05fd40: mov %eax,-0x14000(%rsp) ;...89842400 c0feff
>>> 0x00007faadd05fd47: push %rbp ;...55
>>> 0x00007faadd05fd48: sub $0x20,%rsp ;...4883ec20
>>> ;*synchronization entry
>>> ; - FiboSample::fibo at -1
>>> 0x00007faadd05fd4c: mov %esi,(%rsp) ;...893424
>>> 0x00007faadd05fd4f: cmp $0x2,%esi ;...83fe02
>>> 0x00007faadd05fd52: jl 0x00007faadd05fda1 ;...7c4d
>>> ;*if_icmpge
>>> ; - FiboSample::fibo at 2
>>> 0x00007faadd05fd54: dec %esi ;...ffce
>>> ;*isub
>>> ; - FiboSample::fibo at 9
>>> 0x00007faadd05fd56: nop ;...90
>>> 0x00007faadd05fd57: callq 0x00007faadd038060 ;...e80483fd ff
>>> ; OopMap{off=28}
>>> ;*invokestatic fibo
>>> ; - FiboSample::fibo at 10
>>> ; {static_call}
>>> 0x00007faadd05fd5c: mov %eax,0x4(%rsp) ;...89442404
>>> ;*goto
>>> ; - FiboSample::fibo at 61
>>> 0x00007faadd05fd60: mov (%rsp),%esi ;...8b3424
>>> 0x00007faadd05fd63: add $0xfffffffffffffffe,%esi ;...83c6fe
>>> ;*isub
>>> ; - FiboSample::fibo at 18
>>> 0x00007faadd05fd66: nop ;...90
>>> 0x00007faadd05fd67: callq 0x00007faadd038060 ;...e8f482fd ff
>>> ; OopMap{off=44}
>>> ;*invokestatic fibo
>>> ; - FiboSample::fibo at 19
>>> ; {static_call}
>>> 0x00007faadd05fd6c: mov %eax,%r9d ;...448bc8
>>> 0x00007faadd05fd6f: mov 0x4(%rsp),%eax ;...8b442404
>>> 0x00007faadd05fd73: add %r9d,%eax ;...4103c1
>>> ;*goto
>>> ; - FiboSample::fibo at 75
>>> 0x00007faadd05fd76: mov 0x4(%rsp),%r11d ;...448b5c24 04
>>> 0x00007faadd05fd7b: xor %eax,%r11d ;...4433d8
>>> 0x00007faadd05fd7e: mov %r9d,%r8d ;...458bc1
>>> 0x00007faadd05fd81: xor %eax,%r8d ;...4433c0
>>> 0x00007faadd05fd84: and %r8d,%r11d ;...4523d8
>>> 0x00007faadd05fd87: test %r11d,%r11d ;...4585db
>>> 0x00007faadd05fd8a: jge 0x00007faadd05fda6 ;...7d1a
>>> ;*ifge
>>> ; - jdart.runtime.RT::addExact at 11 (line 139)
>>> ; - FiboSample::fibo at 37
>>> 0x00007faadd05fd8c: mov $0xa5,%esi ;...bea50000 00
>>> 0x00007faadd05fd91: mov %r9d,(%rsp) ;...44890c24
>>> 0x00007faadd05fd95: xchg %ax,%ax ;...6690
>>> 0x00007faadd05fd97: callq 0x00007faadd039020 ;...e88492fd ff
>>> ; OopMap{off=92}
>>> ;*new ; - jdart.runtime.RT::addExact at 14 (line 140)
>>> ; - FiboSample::fibo at 37
>>> ; {runtime_call}
>>> 0x00007faadd05fd9c: callq 0x00007faae7394800 ;...e85f4a33 0a
>>> ;*new ; - jdart.runtime.RT::addExact at 14 (line 140)
>>> ; - FiboSample::fibo at 37
>>> ; {runtime_call}
>>> 0x00007faadd05fda1: mov $0x1,%eax ;...b8010000 00
>>> 0x00007faadd05fda6: add $0x20,%rsp ;...4883c420
>>> 0x00007faadd05fdaa: pop %rbp ;...5d
>>> 0x00007faadd05fdab: test %eax,0xab5524f(%rip) # 0x00007faae7bb5000
>>> ;...85054f52 b50a
>>> ; {poll_return}
>>> 0x00007faadd05fdb1: retq ;...c3
>>> 0x00007faadd05fdb2: mov 0x8(%rax),%r11d ;...448b5808
>>> 0x00007faadd05fdb6: cmp $0xefe4b124,%r11d ;...4181fb24 b1e4ef
>>> ; {oop('jdart/runtime/**ControlFlowException')}
>>> 0x00007faadd05fdbd: jne 0x00007faadd05fdf5 ;...7536
>>> ;*invokestatic fibo
>>> ; - FiboSample::fibo at 19
>>> 0x00007faadd05fdbf: mov 0x20(%rax),%ebp ;...8b6820
>>> 0x00007faadd05fdc2: test %ebp,%ebp ;...85ed
>>> 0x00007faadd05fdc4: jne 0x00007faadd05fe11 ;...754b
>>> ;*getfield value
>>> ; - FiboSample::fibo at 70
>>> 0x00007faadd05fdc6: mov 0x4(%rsp),%eax ;...8b442404
>>> 0x00007faadd05fdca: xor %r9d,%r9d ;...4533c9
>>> 0x00007faadd05fdcd: jmp 0x00007faadd05fd76 ;...eba7
>>> 0x00007faadd05fdcf: mov 0x8(%rax),%r11d ;...448b5808
>>> 0x00007faadd05fdd3: cmp $0xefe4b124,%r11d ;...4181fb24 b1e4ef
>>> ; {oop('jdart/runtime/**ControlFlowException')}
>>> 0x00007faadd05fdda: jne 0x00007faadd05fdf0 ;...7514
>>> ;*invokestatic fibo
>>> ; - FiboSample::fibo at 10
>>> 0x00007faadd05fddc: mov 0x20(%rax),%ebp ;...8b6820
>>> 0x00007faadd05fddf: test %ebp,%ebp ;...85ed
>>> 0x00007faadd05fde1: jne 0x00007faadd05fe02 ;...751f
>>> ;*getfield value
>>> ; - FiboSample::fibo at 57
>>> 0x00007faadd05fde3: xor %r11d,%r11d ;...4533db
>>> 0x00007faadd05fde6: mov %r11d,0x4(%rsp) ;...44895c24 04
>>> 0x00007faadd05fdeb: jmpq 0x00007faadd05fd60 ;...e970ffff ff
>>> 0x00007faadd05fdf0: mov %rax,%rsi ;...488bf0
>>> 0x00007faadd05fdf3: jmp 0x00007faadd05fdf8 ;...eb03
>>> 0x00007faadd05fdf5: mov %rax,%rsi ;...488bf0
>>> ;*invokestatic fibo
>>> ; - FiboSample::fibo at 19
>>> 0x00007faadd05fdf8: add $0x20,%rsp ;...4883c420
>>> 0x00007faadd05fdfc: pop %rbp ;...5d
>>> 0x00007faadd05fdfd: jmpq 0x00007faadd061660 ;...e95e1800 00
>>> ; {runtime_call}
>>> 0x00007faadd05fe02: mov $0xffffffec,%esi ;...beecffff ff
>>> 0x00007faadd05fe07: callq 0x00007faadd039020 ;...e81492fd ff
>>> ; OopMap{rbp=NarrowOop off=204}
>>> ;*astore_2
>>> ; - FiboSample::fibo at 60
>>> ; {runtime_call}
>>> 0x00007faadd05fe0c: callq 0x00007faae7394800 ;...e8ef4933 0a
>>> ;*getfield value
>>> ; - FiboSample::fibo at 57
>>> ; {runtime_call}
>>> 0x00007faadd05fe11: mov $0xffffffec,%esi ;...beecffff ff
>>> 0x00007faadd05fe16: nop ;...90
>>> 0x00007faadd05fe17: callq 0x00007faadd039020 ;...e80492fd ff
>>> ; OopMap{rbp=NarrowOop off=220}
>>> ;*astore
>>> ; - FiboSample::fibo at 73
>>> ; {runtime_call}
>>> 0x00007faadd05fe1c: callq 0x00007faae7394800 ;...e8df4933 0a
>>> ;*getfield value
>>> ; - FiboSample::fibo at 70
>>> ; {runtime_call}
>>> 0x00007faadd05fe21: hlt ;...f4
>>> 0x00007faadd05fe22: hlt ;...f4
>>> 0x00007faadd05fe23: hlt ;...f4
>>> 0x00007faadd05fe24: hlt ;...f4
>>> 0x00007faadd05fe25: hlt ;...f4
>>> 0x00007faadd05fe26: hlt ;...f4
>>> 0x00007faadd05fe27: hlt ;...f4
>>> 0x00007faadd05fe28: hlt ;...f4
>>> 0x00007faadd05fe29: hlt ;...f4
>>> 0x00007faadd05fe2a: hlt ;...f4
>>> 0x00007faadd05fe2b: hlt ;...f4
>>> 0x00007faadd05fe2c: hlt ;...f4
>>> 0x00007faadd05fe2d: hlt ;...f4
>>> 0x00007faadd05fe2e: hlt ;...f4
>>> 0x00007faadd05fe2f: hlt ;...f4
>>> 0x00007faadd05fe30: hlt ;...f4
>>> 0x00007faadd05fe31: hlt ;...f4
>>> 0x00007faadd05fe32: hlt ;...f4
>>> 0x00007faadd05fe33: hlt ;...f4
>>> 0x00007faadd05fe34: hlt ;...f4
>>> 0x00007faadd05fe35: hlt ;...f4
>>> 0x00007faadd05fe36: hlt ;...f4
>>> 0x00007faadd05fe37: hlt ;...f4
>>> 0x00007faadd05fe38: hlt ;...f4
>>> 0x00007faadd05fe39: hlt ;...f4
>>> 0x00007faadd05fe3a: hlt ;...f4
>>> 0x00007faadd05fe3b: hlt ;...f4
>>> 0x00007faadd05fe3c: hlt ;...f4
>>> 0x00007faadd05fe3d: hlt ;...f4
>>> 0x00007faadd05fe3e: hlt ;...f4
>>> 0x00007faadd05fe3f: hlt ;...f4
>>> [Stub Code]
>>> 0x00007faadd05fe40: mov $0x0,%rbx ;...48bb0000 000000
>>> ;...000000
>>> ; {no_reloc}
>>> 0x00007faadd05fe4a: jmpq 0x00007faadd05fe4a ;...e9fbffff ff
>>> ; {runtime_call}
>>> 0x00007faadd05fe4f: mov $0x0,%rbx ;...48bb0000 000000
>>> ;...000000
>>> ; {static_stub}
>>> 0x00007faadd05fe59: jmpq 0x00007faadd05fe59 ;...e9fbffff ff
>>> ; {runtime_call}
>>> [Exception Handler]
>>> 0x00007faadd05fe5e: jmpq 0x00007faadd05e8e0 ;...e97deaff ff
>>> ; {runtime_call}
>>> [Deopt Handler Code]
>>> 0x00007faadd05fe63: callq 0x00007faadd05fe68 ;...e8000000 00
>>> 0x00007faadd05fe68: subq $0x5,(%rsp) ;...48832c24 05
>>> 0x00007faadd05fe6d: jmpq 0x00007faadd038c00 ;...e98e8dfd ff
>>> ; {runtime_call}
>>> 0x00007faadd05fe72: hlt ;...f4
>>> 0x00007faadd05fe73: hlt ;...f4
>>> 0x00007faadd05fe74: hlt ;...f4
>>> 0x00007faadd05fe75: hlt ;...f4
>>> 0x00007faadd05fe76: hlt ;...f4
>>> 0x00007faadd05fe77: hlt ;...f4
>>>
>>>
>>>
>>>> R¨¦mi Forax wrote:
>>>>
>>>>> Hi guys,
>>>>> I've found something really weird, c2 sometimes generates
>>>>> the assembler instruction xchg %ax, %ax (see the assembly code
>>>>> of fibo just before the first recursive call)
>>>>> which I believe is equivalent to a nop but slower.
>>>>>
>>>>> In fact, xchg is really slow in my laptop (Nehalem), slower than
>>>>> at least 5/6 instructions like move/xor/and.
>>>>> I think it's because xchg is atomic see [1]
>>>>>
>>>>> I think c2 should never generate xchg or at least replace all xchg
>>>>> %r, %r by nop.
>>>>>
>>>>> R¨¦mi
>>>>>
>>>>> [1]
>>>>> http://www.intel.ru/content/**dam/doc/white-paper/intel-**
>>>>> microarchitecture-white-paper.**pdf<http://www.intel.ru/content/dam/doc/white-paper/intel-microarchitecture-white-paper.pdf>
>>>>>
>>>>>
>>>>>
>>>>> public class ClassicFibo {
>>>>> private static int fibo(int n) {
>>>>> if (n < 2) {
>>>>> return 1;
>>>>> }
>>>>> return fibo(n - 1) + fibo(n - 2);
>>>>> }
>>>>>
>>>>> public static void main(String[] args) {
>>>>> System.out.println(fibo(40));
>>>>> }
>>>>> }
>>>>>
>>>>> # {method} 'fibo' '(I)I' in 'ClassicFibo'
>>>>> # parm0: rsi = int
>>>>> # [sp+0x30] (sp of caller)
>>>>> 0x00007fb409061620: mov %eax,-0x14000(%rsp)
>>>>> 0x00007fb409061627: push %rbp
>>>>> 0x00007fb409061628: sub $0x20,%rsp ;*synchronization entry
>>>>> ; - ClassicFibo::fibo at -1 (line 4)
>>>>> 0x00007fb40906162c: mov %esi,%ebp
>>>>> 0x00007fb40906162e: cmp $0x2,%esi
>>>>> 0x00007fb409061631: jl 0x00007fb409061651 ;*if_icmpge
>>>>> ; - ClassicFibo::fibo at 2 (line 4)
>>>>> 0x00007fb409061633: dec %esi ;*isub
>>>>> ; - ClassicFibo::fibo at 9 (line 7)
>>>>> 0x00007fb409061635: xchg %ax,%ax
>>>>> 0x00007fb409061637: callq 0x00007fb409038060 ; OopMap{off=28}
>>>>> ;*invokestatic fibo
>>>>> ; - ClassicFibo::fibo at 10 (line 7)
>>>>> ; {static_call}
>>>>> ...
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>
>>>  .
>>
>>
>
> ______________________________**__
>
> This email (including any attachments) is confidential and may be legally
> privileged. If you received this email in error, please delete it
> immediately and do not copy it or use it for any purpose or disclose its
> contents to any other person. Thank you.
>
> ±¾µçÓÊ(°üÀ¨ÈÎºÎ¸½¼þ)¿ÉÄÜº¬ÓÐ»úÃÜ×ÊÁÏ²¢ÊÜ·¨ÂÉ±£»¤¡£**ÈçÄú²»ÊÇÕýÈ·µÄÊÕ¼þÈË£¬ÇëÄúÁ¢¼´É¾³ý±¾ÓÊ¼þ¡£**Çë²»Òª½«±¾µçÓÊ½øÐÐ¸´ÖÆ²¢ÓÃ×÷ÈÎºÎÆäËûÓÃÍ¾¡¢**
> »òÍ¸Â¶±¾ÓÊ¼þÖ®ÄÚÈÝ¡£Ð»Ð»¡£
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20120706/b91a4568/attachment-0001.html 