Misuse of Math library without strictfp

[andelf]

In java8, the default behavior of Math is using any available FPU like 8087. The results of float operations might be platform-dependent.
In later versions, the default behavior changes and will use IEEE-float(SSE2 instructions). So you can't update JDK/JRE without a proposal-fork.

In detail, java-tron's double arithmetic will use x87 instructions. And is difficult to port to other languages, since all modern languages use ieee-compatible and stable arithmetic for float numbers.

[andelf]

In ExchangeProcessor.java:

  private long exchangeToSupply(long balance, long quant) {
    long newBalance = balance + quant;
    double issuedSupply = -supply * (1.0 - Math.pow(1.0 + (double) quant / newBalance, 0.0005));
    // ...
}

The Math.pow(1.0 + (double) quant / newBalance, 0.0005) is the first step to calculate exchange amount.

let balance = 4732214, quant = 4732214, then:

Math.pow(1.0 + (double) quant / newBalance, 0.0005)
= Math.pow(1.0061363892207218, 0.0005)

= 1.000003058823805400000000 in java8, javatron.
But.
= 1.000003058823805100000000 in Python, Rust, java13, etc.

The final calculation result(exchange amount) in java8(mainnet) is 13516579896.

While in other languages, it's 13516579895. THIS IS A 1 SUN DIFF!!. 🐛

This is what happens in block #4137160, 0xa33ca08a8cffff532fa43771527e145a63a37fca50dd1468632acc146e8e2d7a.

[andelf]

To simulate the calculation in Rust:

Rust Playground

#![feature(asm)]

fn main() {
    let a = 1.0061363892207218_f64;
    let b = 0.0005_f64;

    let val0: f64 = a.powf(0.0005);
    let mut v = 0_f64;

    let mut cntr = 0_u16;
    unsafe {
        asm!(
            "fnstcw word ptr [{cntr}]",
            // calculate
            "fld qword ptr [{y}]",
            "fld qword ptr [{x}]",
            "fyl2x",
            "f2xm1",
            "fld1",
            "faddp st(1), st(0)",
            "fstp qword ptr [{v}]",
            x = in(reg) &a,
            y = in(reg) &b,
            v = in(reg) &mut v,
            cntr = in(reg) &mut cntr,
            options(nostack)
        );
    }

    println!("val0 => {}", val0);
    println!("val1 => {}", v);

    println!("high precision => {}", (cntr & 0x300) != 0); // high precision flag
}

Output

val0 => 1.0000030588238051
val1 => 1.0000030588238054
high precision => true

[andelf]

See-also: jdk8/hotspot/src/cpu/x86/vm/macroAssembler_x86.cpp

[andelf]

This is fixed by tronprotocol/tips#697

But one can not sync from 0-block without java 8. It's heavily relies on the "downloadable" snapshot.