RFR: 8282221: x86 intrinsics for divideUnsigned and remainderUnsigned methods in java.lang.Integer and java.lang.Long

[Jatin Bhateja]

On Tue, 22 Feb 2022 09:24:47 GMT, Vamsi Parasa <duke at openjdk.java.net> wrote:

> Optimizes the divideUnsigned() and remainderUnsigned() methods in java.lang.Integer and java.lang.Long classes using x86 intrinsics. This change shows 3x improvement for Integer methods and upto 25% improvement for Long. This change also implements the DivMod optimization which fuses division and modulus operations if needed. The DivMod optimization shows 3x improvement for Integer and ~65% improvement for Long.

src/hotspot/cpu/x86/x86_64.ad line 8602:

> 8600:     __ jmp(done);
> 8601:     __ bind(neg_divisor_fastpath); 
> 8602:     // Fastpath for divisor < 0: 

Move in macro assembly routine.

src/hotspot/cpu/x86/x86_64.ad line 8633:

> 8631:     __ jmp(done);
> 8632:     __ bind(neg_divisor_fastpath);
> 8633:     // Fastpath for divisor < 0: 

Move in macro assembly rountine.

src/hotspot/cpu/x86/x86_64.ad line 8722:

> 8720:     __ shrl(rax, 31); // quotient
> 8721:     __ sarl(tmp, 31);
> 8722:     __ andl(tmp, divisor);

Move in macro assembly routine.

src/hotspot/cpu/x86/x86_64.ad line 8763:

> 8761:     __ andnq(rax, rax, rdx);
> 8762:     __ movq(tmp, rax);
> 8763:     __ shrq(rax, 63); // quotient

Move in macro assembly routine.

src/hotspot/cpu/x86/x86_64.ad line 8902:

> 8900:     __ subl(tmp_rax, divisor);
> 8901:     __ andnl(tmp_rax, tmp_rax, rdx);
> 8902:     __ sarl(tmp_rax, 31);

Please move this into a macro assembly routine.

src/hotspot/cpu/x86/x86_64.ad line 8932:

> 8930:     // Fastpath when divisor < 0: 
> 8931:     // remainder = dividend - (((dividend & ~(dividend - divisor)) >> (Long.SIZE - 1)) & divisor)
> 8932:     // See Hacker's Delight (2nd ed), section 9.3 which is implemented in java.lang.Long.remainderUnsigned()

Please move it into a macro assembly routine.

src/hotspot/share/opto/compile.cpp line 3499:

> 3497:       Node* d = n->find_similar(Op_UDivI);
> 3498:       if (d) {
> 3499:         // Replace them with a fused unsigned divmod if supported

Can you explain a bit here, why can't this transformation be handled earlier ?

src/hotspot/share/opto/divnode.cpp line 1350:

> 1348:     return NULL;
> 1349:   }
> 1350: 

Please remove Value and Ideal routines if no explicit transforms are being done.

src/hotspot/share/opto/divnode.cpp line 1362:

> 1360:   }
> 1361: 
> 1362: //=============================================================================

You can remove Ideal routine is not transformation is being done.

test/micro/org/openjdk/bench/java/lang/IntegerDivMod.java line 76:

> 74:         return quotients;
> 75:     }
> 76: 

Return seems redundant here.

test/micro/org/openjdk/bench/java/lang/IntegerDivMod.java line 83:

> 81:         }
> 82:         return remainders;
> 83:     }

Return seems redundant here.

test/micro/org/openjdk/bench/java/lang/LongDivMod.java line 75:

> 73:         }
> 74:         return quotients;
> 75:     }

Do we need to return quotients, since it's a field  being explicitly modified.

test/micro/org/openjdk/bench/java/lang/LongDivMod.java line 82:

> 80:             remainders[i] = Long.remainderUnsigned(dividends[i], divisors[i]);
> 81:         }
> 82:         return remainders;

Same as above

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

PR: https://git.openjdk.java.net/jdk/pull/7572