SpinPause() should be inlined?

[Yasumasa Suenaga]

Hi all,

I saw lock contention in SecureRandom. When I was analyzing it, I had a 
question.

ObjectMonitor::TrySpin() calls SpinPause(). SpinPause() would issue 
PAUSE, but it is not inlined as following:

```
  1587e90:       f0 4d 0f b1 2e          lock cmpxchg %r13,(%r14)
  1587e95:       48 85 c0                test   %rax,%rax
  1587e98:       74 16                   je     1587eb0 
<ObjectMonitor::TrySpin(JavaThread*)+0x60>
  1587e9a:       e8 31 7b d4 ff          callq  12cf9d0 <SpinPause>
  1587e9f:       83 eb 01                sub    $0x1,%ebx
  1587ea2:       72 da                   jb     1587e7e 
<ObjectMonitor::TrySpin(JavaThread*)+0x2e>
```

I found following comment about it in os.hpp. It says SpinPause() should 
be inlined.

```
// Note that "PAUSE" is almost always used with synchronization
// so arguably we should provide Atomic::SpinPause() instead
// of the global SpinPause() with C linkage.
// It'd also be eligible for inlining on many platforms.
```

According to Intel Software Developer's Manual, PAUSE seems to need to 
be inlined, but I'm not sure it can be allow through function call.
I've fixed it to do so for Linux x64 [1] and I benchmarked with [2], 
then I saw some advantage in PAUSE on my Core i3-8145U, but it may be 
within the margin of error.

* original
Benchmark                       (algo)  (bytes)   Mode  Cnt       Score  
       Error  Units
RandomBenchmark.fillRandomInMT    DRBG       16  thrpt    3  510141.578 
± 138543.261  ops/s

* with inlined PAUSE
Benchmark                       (algo)  (bytes)   Mode  Cnt       Score  
      Error  Units
RandomBenchmark.fillRandomInMT    DRBG       16  thrpt    3  531589.958 
± 66942.549  ops/s


Should be inlined PAUSE as `Atomic::SpinPause()`?


Thanks,

Yasumasa


[1] 
https://github.com/YaSuenag/jdk/commit/66b31d35fac6fb0537bb9d85957157342fec564d
[2] 
https://github.com/YaSuenag/hwrand/blob/master/benchmark/src/main/java/com/yasuenag/hwrand/benchmark/RandomBenchmark.java#L59-L64