performance issue with alignment-like checks

[Maurizio Cimadamore]

This issue was already known to Vlad - not sure if a bug has been 
submitted or not.

Currently Hotspot fail to hoist alignment-like checks (with either mod 
operator or bitwise-and) outside of loops. Simple benchmark:

|static final int ALIGN_MASK = 3; static final int ALIGN = ALIGN_MASK + 
1; @Benchmark @OutputTimeUnit(TimeUnit.MICROSECONDS) public int 
align_loop() { int sum = 0; for (int i = 0 ; i < ELEM_SIZE ; i+= ALIGN) 
{ if ((i & ALIGN_MASK) != 0) { throw new AssertionError(); } sum += i; } 
return sum; } @Benchmark @OutputTimeUnit(TimeUnit.MICROSECONDS) public 
int alignno_loop() { int sum = 0; for (int i = 0 ; i < ELEM_SIZE ; i+= 
ALIGN) { sum += i; } return sum; } |

Result:

|Benchmark Mode Cnt Score Error Units PanamaPointer.align_loop avgt 30 
12.589 ? 0.137 us/op PanamaPointer.alignno_loop avgt 30 7.443 ? 0.088 us/op |

I realized that this issue is the main offender behind performance 
issues in code using MemoryAddress::addOffset in tight loops. Numbers 
with Graal/JVMCI seem more consistent:

|Benchmark Mode Cnt Score Error Units PanamaPointer.align_loop avgt 30 
7.052 ? 0.067 us/op PanamaPointer.alignno_loop avgt 30 6.988 ? 0.041 us/op |

The priority of this should be raised - it is probably a relatively 
low-hanging fruit to fix, and it could help fixing our woes with 
alignment checks in Panama.

Maurizio

​