09 July 2007

Tagged unboxed floating point numbers

Several modern programming language implementations employ a representation of object reference slots that is self-describing, in order to facilitate run-time type checks and automatic garbage collection. By reserving one or more bits to indicate the type of data stored within the slot, it is possible to differentiate a pointer (also known as a reference to a "boxed" object) from, say, an integer (also known as an "unboxed" integer).

Suppose that reference slots are 64-bits wide, and that 61 bits can be used to store an unboxed integer. Assuming signed integers, we can store an integer in [-2^60..2^60), but outside that range, we are forced to create a boxed integer object and store a reference to that object. This is in fact how Lyken implements integers (though it preserves 62 bits of accuracy for integers).

Now, suppose that we want to support double-precision floating point numbers. The fundamental approaches taken by every implementation I have found in the literature are to either 1) box all floating point numbers, or 2) to use a combination of boxed floating point numbers and untagged floating point numbers. As one might imagine, (1) can cause serious performance degradation for numerically intensive programs, due to the need to create new boxed objects to store the result of each floating point computation. As for (2), there are numerous papers that discuss various compilation strategies for finding opportunities to use untagged unboxed floating point numbers, but these techniques appear to to be limited to particular problem domains, since they mainly try to convert vectors of floating point numbers to be untagged and unboxed. Nowhere have I found any mention whatsoever of using tagged unboxed floating point numbers.

Let us consider the IEEE 754 floating point number format to see what challenges there are to tagged unboxed double-precision floating point numbers. (If you are unfamiliar with the format, I suggest taking a look at the Wikipedia page for an overview.) There are three fields: 1) sign, 2) exponent, and 3) fraction.

seeeeeee eeeeffff ffffffff ffffffff ffffffff ffffffff ffffffff ffffffff

Suppose we were to steal 3 bits from the fraction field. In order to avoid losing precision, we would have to box all numbers that did not have a particular bit pattern for those 3 stolen bits, thus allowing us to unbox perhaps 12.5% of the time. This is not compelling.

What if we were to instead steal bits from the exponent? This is much more useful, because it allows us to accurately store all values except those with the most extreme exponent values. Of course, there are programs that actually need the full range of exponent values, but they are by no means the common case.

There are some details that make such unboxing more work than for integers:
• The exponent must be re-biased.
• It is harder to remove the exponent bits, since they are internal.
• There are special values that require special handling (+-0.0, +-Inf, NaN).
Explaining the nuances in words is rather tedious, so actual code follows instead.
`typedef union {  uint64_t u;  int64_t i;  double r;} LktRealUnion;voidLkRealNew(LktSlot *aReal, double aVal) {    LktRealUnion val;    val.r = aVal;    // Check whether +-0.0.    if (val.u & 0x7fffffffffffffffLLU) {        LktRealUnion unboxed;        // Re-bias the exponent by subtracting 896.  This makes the useful        // exponent range for unboxed reals [-127..128].        unboxed.u = val.u - 0x3800000000000000LLU;        // Check that the most significant 3 exponent bits are 0.        if (unboxed.u & 0x7000000000000000LLU) {            if ((val.u & 0x7ff0000000000000LLU) == 0x7ff0000000000000LLU) {                // Special value (Inf or NaN).                uint64_t sign = (val.u & 0x8000000000000000LLU);                unboxed.u <<= 3;                unboxed.u &= 0x7fffffffffffffffLLU; // Clear sign bit.                unboxed.u |= sign;                unboxed.u |= 0x3; // Tag.                aReal->u.b = unboxed.u;            } else {                // Overflow; box.                // [...]            }        } else {            uint64_t sign = (val.u & 0x8000000000000000LLU);            unboxed.u <<= 3;            // Sign bit is already cleared as a result of exponent re-biasing.            unboxed.u |= sign;            unboxed.u |= 0x3; // Tag.            aReal->u.b = unboxed.u;        }    } else {        // +-0.0.        aReal->u.b = val.u | 0x3; // Tag.    }}doubleLkRealGet(LktSlot *aReal) {    LktRealUnion val;    LkmAssert(LkSlotTypeGet(aReal) == LkRealType());    // Checked whether boxed.    val.u = aReal->u.b;    if ((val.u & 0x7) == 0x3) {        // Check whether +-0.0.        if (val.u & 0x7ffffffffffffff8LLU) {            val.i >>= 3; // Sign-extended shift preserves the sign bit.            val.u &= 0x8fffffffffffffffLLU; // Clear upper exponent bits.            // Check whether a special value (Inf or NaN).            if ((val.u & 0x0ff0000000000000LLU) != 0x0ff0000000000000LLU) {                // Re-bias the exponent by adding 896.                val.u += 0x3800000000000000LLU;            } else {                // Special value.  Set all exponent bits.                val.u |= 0x7ff0000000000000LLU;            }        } else {            // +-0.0.            val.u &= 0x8000000000000000LLU;        }        return val.r;    } else {        // Boxed.        LktReal *r = (LktReal *) aReal->u.p;        return r->val;    }}`
As you can see, unboxed floating point numbers do incur some overhead, but for typical applications, they appear to me to be a big improvement over uniformly boxed floating point numbers.

At February 3, 2008 12:47 AM ,  Anonymous said...

You should take a look at http://article.gmane.org/gmane.comp.lang.lua.general/44823 for an even more clever approach.

At February 3, 2008 9:37 AM ,  Jason said...

I recently encountered the described technique in Tamarin. It is indeed clever, but it has the problem that it requires several more bits than my method, thus reducing the range of other unboxed values. Pointers on 64-bit systems are troublesome, though in practice modern systems only use 40-48 bits for pointers, thus allowing for some bit stealing tricks.

Anyway, thank you for the reference. I am certainly going to need to benchmark both methods to decide which one to use.

At February 20, 2008 3:44 PM ,  Emanuele Ruffaldi said...

I've played with the NaN technique some time ago and written some example code.

In my case the 64bit value can be a double precision floating point value or a tagged value, but unluckly I was not handling 64 bit pointers and 64 bit integers.

http://www.teslacore.it/wiki/index.php?title=FloatVariant

At May 14, 2008 6:13 PM ,  edwsmith said...

We're starting to face the x64 demon in Tamarin tracing, which uses the NaN trick on x86/ARM.
https://bugzilla.mozilla.org/show_bug.cgi?id=433793

Any outcome from experimenting with both formats?

At May 14, 2008 9:25 PM ,  Jason said...

No, I haven't conducted these experiments yet for Lyken, though I've given the issue a bit of thought in the context of Tamarin. I suspect that it will be faster to use the method described in my blog post than it would be to collapse pointers into the fraction bits of type double.

This hunch is based on the (unconfirmed) assumption that reference pointers are substantially more common than unboxed floating point numbers. Collapsing pointers is faster than re-biasing floating point numbers and dealing with boxing on overflow, but I doubt that even numerically intensive code has a large proportion of floating point objects, as compared to reference pointers.