RFR: 8271883: Math CopySign optimization for x86 [v5]

[Sandhya Viswanathan]

On Mon, 9 Aug 2021 15:40:59 GMT, Marcus G K Williams <mgkwill at openjdk.org> wrote:

>> Intrinsic for Math.copySign is disabled on x86_64.
>> 
>> We can improve on generated c2 instructions for float and double, and this change adds optimized intrinsics for float and double Math.copySign.
>> 
>> ### **Math.copySign(double)**
>> _From:_
>>   0x00007f7d606e5dac:   vmovq  %xmm1,%r10
>>   0x00007f7d606e5db1:   vmovq  %xmm0,%r11
>>   0x00007f7d606e5db6:   movabs $0x7fffffffffffffff,%r8
>>   0x00007f7d606e5dc0:   and    %r8,%r11
>>   0x00007f7d606e5dc3:   movabs $0x8000000000000000,%r8
>>   0x00007f7d606e5dcd:   and    %r8,%r10
>>   0x00007f7d606e5dd0:   or     %r11,%r10
>>   0x00007f7d606e5dd3:   vmovq  %r10,%xmm0                   
>> 
>> _To:_
>>   0x00007fc3c14c63ac:   movabs $0x7fffffffffffffff,%r10
>>   0x00007fc3c14c63b6:   vmovq  %r10,%xmm2
>>   0x00007fc3c14c63bb:   vpternlogq $0xe4,%xmm2,%xmm1,%xmm0
>> 
>> ### **Math.copySign(float)**
>> _From:_
>>   0x00007ff8886e60ac:   vmovd  %xmm1,%r11d
>>   0x00007ff8886e60b1:   vmovd  %xmm0,%r10d
>>   0x00007ff8886e60b6:   and    $0x80000000,%r11d
>>   0x00007ff8886e60bd:   and    $0x7fffffff,%r10d
>>   0x00007ff8886e60c4:   or     %r10d,%r11d
>>   0x00007ff8886e60c7:   vmovd  %r11d,%xmm0
>> _To:_
>>   0x00007fc7d94c63ac:   mov    $0x7fffffff,%r10d
>>   0x00007fc7d94c63b2:   vmovd  %r10d,%xmm3
>>   0x00007fc7d94c63b7:   vpternlogd $0xe4,%xmm3,%xmm1,%xmm0
>> 
>> #### _**Performance of patch using updated test/micro/org/openjdk/bench/vm/compiler/Signum.java:**_
>> #### **BEFORE**
>> Signum._5_copySignFloatTest       avgt    5  2.442 ? 0.024  ns/op
>> Signum._7_copySignDoubleTest      avgt    5  2.400 ? 0.033  ns/op
>> 
>> #### **PATCH**
>> Signum._5_copySignFloatTest       avgt    5  2.029 ? 0.011  ns/op
>> Signum._7_copySignDoubleTest      avgt    5  2.029 ? 0.024  ns/op
>> 
>> #### **JTREG that covers this case:**
>> test/hotspot/jtreg/compiler/intrinsics/math/TestSignumIntrinsic.java
>> 
>> Signed-off-by: Marcus G K Williams <marcus.williams at intel.com>
>
> Marcus G K Williams has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Remove predicates
>   
>   Signed-off-by: Marcus G K Williams <marcus.williams at intel.com>

src/hotspot/cpu/x86/x86.ad line 5813:

> 5811: // Desired Truth Table: A -> xmm0 bit, B -> xmm1 bit, C -> xmm2 bit
> 5812: // Wherever xmm2 is 0, we want to pick from B (sign)
> 5813: // Wherever xmm2 is 1, we want to pick from A (src)

Good to add then C (xmm2) is set to 0x7FFFFFFF.

src/hotspot/cpu/x86/x86.ad line 5835:

> 5833:   ins_encode %{
> 5834:     __ movl($tmp2$$Register, 0x7FFFFFFF);
> 5835:     __ vmovd($tmp1$$XMMRegister, $tmp2$$Register);

You could use movdl() instead of adding a new vmod instruction.

src/hotspot/cpu/x86/x86.ad line 5848:

> 5846:   ins_encode %{
> 5847:     __ mov64($tmp2$$Register, 0x7FFFFFFFFFFFFFFF);
> 5848:     __ vmovq($tmp1$$XMMRegister, $tmp2$$Register);

You could use movq() instead of adding a new vmovq instruction.

-------------

PR: https://git.openjdk.java.net/jdk/pull/5005