[riscv-port] RFR: 8279292: riscv: Intrinsify multiplyToLen and squareToLen
Fei Yang
fyang at openjdk.java.net
Mon Dec 27 11:39:43 UTC 2021
On Mon, 27 Dec 2021 08:25:22 GMT, Feilong Jiang <fjiang at openjdk.org> wrote:
> BigInteger intrinsic: MultiplyToLen and SquareToLen intrinsic are missed in current vm. They should be implemented.
> The JMH tests show that the MultiplyToLen intrinsic improve the performance by up to 2x ~ 3x and the SquareToLen intrinsic improve the performance by up to 1.8x ~ 2x when the length of BigInteger changed from 1 to 5000, compared with that of C2.
>
> Full jtreg tests on qemu and hotspot/jdk tier1 test on Unmathced are passed without new failures.
>
> JMH tests and results on D1 and Unmatched list as follows:
> [MyBenchmark.txt](https://github.com/openjdk/riscv-port/files/7779255/MyBenchmark.txt)
>
> [squareToLen_unmatched.txt](https://github.com/openjdk/riscv-port/files/7779247/squareToLen_unmatched.txt)
> [squareToLen_d1.txt](https://github.com/openjdk/riscv-port/files/7779248/squareToLen_d1.txt)
> [multiplyToLen_unmatched.txt](https://github.com/openjdk/riscv-port/files/7779249/multiplyToLen_unmatched.txt)
> [multiplyToLen_d1.txt](https://github.com/openjdk/riscv-port/files/7779250/multiplyToLen_d1.txt)
Changes requested by fyang (Lead).
src/hotspot/cpu/riscv/macroAssembler_riscv.cpp line 3249:
> 3247:
> 3248: /**
> 3249: * Multiply 128 bit by 128. Unrolled inner loop.
Should be: "Multiply 128 bit by 128 bit. Unrolled inner loop."
src/hotspot/cpu/riscv/macroAssembler_riscv.cpp line 3381:
> 3379:
> 3380: /**
> 3381: * Code for BigInteger::multiplyToLen() instrinsic.
There is a typo here. Should be: ""Code for BigInteger::multiplyToLen() intrinsic."
src/hotspot/cpu/riscv/macroAssembler_riscv.cpp line 3417:
> 3415: mv(carry, zr); // carry = 0;
> 3416:
> 3417: Label L_multiply_64_or_128, L_done;
Suggestion: rename L_multiply_64_or_128 into L_multiply_64_x_64_loop
src/hotspot/cpu/riscv/macroAssembler_riscv.cpp line 3424:
> 3422: const Register jdx = tmp1;
> 3423:
> 3424: // if x and y are both 8 bytes aligend.
There is a typo here. Maybe "// Check if x and y are both 8-byte aligned."
-------------
PR: https://git.openjdk.java.net/riscv-port/pull/38
More information about the riscv-port-dev
mailing list