javac/x86/JIT incorrectly treats some double constant-expressions

Name: dkC59003 Date: 09/23/98



Javac/x86/JIT (JDK 1.2fcs-J and -K) fails to pass lang/FP/fpl03502m2 and
lang/FP/fpl03502m4 tests from JCK 1.2(beta4), which check if compile-time
constant expression is treated FP-strict.

These tests calculate a number of double constant expressions, and each of
those expressions result in a double value slightly greater than 1. That tiny
appendix to 1.0 vanishes, if constant-expression is incorrectly calculated
using extended floating-point format suffering second rounding error when
rounded to standard format. (The tests are designed to catch mantissa
extensions upto 105 significand bits.)

Below is the shortest check extracted from fpl03502m2, which produces error
message:

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

public class fpl352m24 {

    public static void main (String args[]) {

double test_value = 1 + (double)((1L<<11) + 1)/(1L<<11)/(1L<<53);

double correct_value = Double.longBitsToDouble(0x3ff0000000000001L);

if (test_value != correct_value) {
println("test_value = " + test_value);
println("instead of: " + correct_value +
" == Double.longBitsToDouble(0x3ffe000000000001)");

};
    };

    static void println (Object x) {
System.out.println( x );
    };
}

---- Prints:

test_value = 1.0
instead of: 1.0000000000000002 == Double.longBitsToDouble(0x3ffe000000000001)

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

The following disassembled fragment compiled by javac/x86 with JIT shows,
that test_value is incorrectly computed at compile time:

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

public static Method main:"([Ljava/lang/String;)V"
stack 4 locals 5
{
// test_value = 1 + (double)((1L<<11) + 1)/(1L<<11)/(1L<<53):
dconst_1; // incorrectly rounded to 1.0 !!!
dstore_1;

// correct_result = Double.longBitsToDouble(0x3ff0000000000001L):
ldc2_w long 4607182418800017409l;
invokestatic Method java/lang/Double.longBitsToDouble:"(J)D";
dstore_3;

// if (test_value == correct_result) goto O'Key;
dload_1;
dload_3;
dcmpl;
ifeq L58;

// ...print error message...

L58: return;
}

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

Nevertheless, the same fragment translated with no JIT using the command:
    javac -J-Djava.compiler=none ...
is compiled correctly, and the test does not print error message.

So, javac/x86/JIT (JDK 1.2 fcs-J & -K) incorrectly calculates some double
constant expressions. Maybe, javac does not says to JIT to calculate those
expressions FP-strict.

But, see also BugId 4175714: "JIT/x86 suffers more than one rounding error
in FP-strict addition".

======================================================================