RFR: 8264054: Bad XMM performance on java.lang.MathBench.sqrtDouble

[Sandhya Viswanathan]

On Tue, 30 Mar 2021 17:53:18 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> For the j.l.Math JMH at https://github.com/openjdk/jmh-jdk-microbenchmarks/blob/master/micros-jdk11/src/main/java/org/openjdk/bench/java/lang/MathBench.java, the performance for sqrt benchmark could be improved.  Thanks a lot to Eric Caspole for finding this issue.
>> 
>> Benchmark:
>>     @Benchmark
>>     public double  sqrtDouble() {
>>         return  Math.sqrt(double4Dot1);
>>     }
>> 
>> Current code generated (linux format) by c2 JIT is:
>>      vsqrtsd 0x50(%r10),%xmm0,%xmm0
>> 
>> The vsqrtsd instruction operation is specified as below: 
>>      VSQRTSD (VEX.128 encoded version)
>>      DEST[63:0] := SQRT(SRC2[63:0])
>>      DEST[127:64] := SRC1[127:64]
>>      DEST[MAXVL-1:128] := 0
>> 
>> The upper 127:64 bits are set from previous contents of xmm0. As the destination xmm0 register was not initialized prior to use by c2 JIT, this causes stall and lower performance.
>> 
>> By adding xmm0 initialization prior to use, the performance of the above benchmark improves significantly.
>> 
>> Code generated after patch:
>>      vxorpd %xmm0,%xmm0,%xmm0
>>      vsqrtsd 0x50(%r10),%xmm0,%xmm0
>> 
>> Performance before patch:
>> Benchmark                    Mode  Cnt       Score      Error   Units
>> MathBench.sqrtDouble        thrpt    8  193612.396 ±  95.807  ops/ms
>> 
>> Performance after patch:
>> MathBench.sqrtDouble        thrpt    8  276388.024 ± 846.372  ops/ms
>> 
>> Best Regards,
>> Sandhya
>
> What is faster: `xor(xmm) + sqrt(xmm, mem)` or `mov(xmm, mem) + sqrt(xmm,xmm)` ?
> 
> Do we have other instructions which may suffer stall too if `dst` register is not zeroed?

@vnkozlov  Both xor(xmm) + sqrt(xmm, mem) or mov(xmm, mem) + sqrt(xmm,xmm) have same performance. I could look into simplifying the patch by just keeping the register version of these rules.

I looked through x86.ad for other usages of unary instructions but didn't find any other instances. We implement negate and abs using binary instructions.

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

PR: https://git.openjdk.java.net/jdk/pull/3256