wiki:SemiTagging

Version 14 (modified by alexey, 7 years ago) (diff)

--

The semi-tagging optimisation

Here I describe the design of the semi-tagging optimisation. Originally the text comes from http://hackage.haskell.org/trac/summer-of-code/ticket/48

This page reflects my current understanding on the compiler and the RTS, so if there is something wrong, just yell!

The starting point

Currently when evaluating an expression that is the scrutinee of a case:

case x of { ... }

GHC jumps to the code for (i.e. "enters") the x closure, which returns when x is evaluated. Commonly, x is already evaluated, and the code for an evaluated constructor just (vector) returns immediately.

The code for the not function:

not x = case x of
        False -> True
        True  -> False

jumps to the boolean argument, passed in R2, after pushing a case frame (the continuation of the function):

        <stack check omitted>
        R1 = R2;
        I64[Sp + (-8)] = sej_info;
        Sp = Sp + (-8);
        jump I64[R1];

Before looking at the rest of the not function, let's look at the code for the True and False closures

True_info:
        jump <address to True alternative>;

False_info:
        jump <address to False alternative>;

they just jump to the appropriate case alternative that is evaluating the closure. These addresses are calculated from the case frame that is on the top of the stack. In this case they select the alternatives from the jump table that is referred to by the not case frame. Below you see the True alternative

sej_0_alt() {
        R1 = False_closure;
        Sp = Sp + 8;
        jump <address to False alternative>;
}

and the False alternative of the not function.

sej_1_alt() {
        R1 = True_closure;
        Sp = Sp + 8;
        jump <address to True alternative>;
}

Just like the constructor closures, they jump to the appropriate branch of the case expression that is evaluating the not function.

Testing before jumping

The simplest optimisation is this. Instead of entering the closure, grab its info pointer, and follow the info pointer to get the tag. Now test the tag; if it's evaluated, don't enter the closure.

The benefit is that processors are typically faster at "test-and-jump to known location" than they are at "jump to this pointer".

Under this scheme, the entry code for the not function would look as follows:

        if(R2 & 1 == 1) goto tagged
        <stack check omitted>
        R1 = R2;
        I64[Sp + (-8)] = sej_info;
        Sp = Sp + (-8);
        jump I64[R1];
tagged:
        R1 = R2 & ~1;  // mask pointer tag out
        <extract constructor tag from pointer>
        if(tag==0) goto sej_0_alt
        goto sej_1_alt

Tagging the LSB of an evaluated closure

The idea is to encode the fact that a pointer points to an evaluated object by setting the LSB of the pointer. If the case expression detects that the closure is evaluated, it can avoid the jump and return, which are expensive on modern processors (indirect jumps).

bits 31..2 bits 1 0
unevaluated closure ptr 00
evaluated constructor closure ptr 01

This would require modifying

  • the code generation so that when allocating a constructor, the pointer to it has the appropriate bits set (just a matter of adjusting the offset from Hp)
  • perhaps, make the entry code for a constructor return a pointer with the appropriate bits set. That way, a function like f xs = head (tail xs) would enter the second element of the list, and return to the caller with appropriate tag bits set.
  • the GC to set the LSB bit of constructor closure pointers,
  • the GC and the RTS code to mask out the LSB pointer when dereferencing it,
  • the code generation to test the LSB bit and case expressions and avoid the indirect jump.

Using more than one bit

We can go a bit further than this, too. Since there are 2 spare bits (4 on a 64-bit machine), we can encode 4 (16) states. Taking 0 to mean "unevaluted", that leaves 3 (15) states to encode the values for the "tag" of the constructor. eg. an evaluated Bool would use 1 to indicate False and 2 to indicate True. An evaluated list cell would use 1 to indicate [] and 2 to indicate (:).

bits 31..2 bits 1 0
unevaluated closure ptr 00
cons. no. 1 ptr 01
cons. no. 2 ptr 10
cons. no. 3 ptr 11

The nice thing about the current approach is that code size is small; implementing the test and jump will certainly add extra code to compiled case expressions. But the gains might be worth it. Complexity-wise this means masking out these bits when following any pointer to a heap object, which means carefully checking most of the runtime.

This would require modifying all of the above plus modifying

  • the code generator so that it checks whether the number of constructors is smaller or equal than 3/15.

Using a tag directly in the pointer

Constructors without children (such as False and True) only need their tag to be represented. Hence we can drop the pointer altogether as follows:

bits 31..2 bits 1 0
unevaluated closure ptr 00
cons. w/no children tag 01
cons. w/children no. 1 ptr 10
cons. w/children no. 2 ptr 11

This still limits us to data types that have no more than two constructors with children. We can improve on this by noting that pointers will not point to low addresses. So we can make a simple test to distinguish between tags and pointers:

bits 31..2 bits 1 0
unevaluated closure ptr 00
cons. w/no children tag 01 tag < no. of constructors
cons. w/ children ptr 01 ptr >= no. of constructors
cons. w/children no. 1 ptr 10
cons. w/children no. 2 ptr 11

Of course this assumes that we don't have data types with too many thousands of constructors.

It might be possible that the case code for the alternatives above is becoming too complex. We might settle for the following "simple" option:

bits 31..2 bits 1 0
unevaluated closure ptr 00
cons. w/no children tag 01 tag < no. of constructors
cons. w/ children ptr 01 ptr >= no. of constructors