VM compilation question

[Norman Ramsey <nr@...>]

I've been poking around at the VM and its compiler, trying to learn
how constants are handled.  It appears that

 1. A typical infix operator can take B and C operands that might be
    constant, so for each operand, distinguishing a constant from a
    register requires a conditional test (e.g., whether a particular
    bit is set).

 2. Every function has its own pool of constants.

 3. As long as the number of constants in a function is small, source
    code like

      n = n + 1   -- assume `n` is local here

    will compile to a single VM instruction.

 4. If the number of constants in a function gets large, to the point
    where there aren't enough bits to pack them all into a B or C
    operand, *from that point onward* no more constants are packed
    into a B or C operand.  In other words, if the assignment `n = n + 1`
    is compiled *after* the compiler has seen a large number of
    constants, then it compiles to two VM instructions: one to load
    the literal 1 into a register, and one to do the addition.

Does this summary seem accurate?  It's the last point that I'm really
curious about?


Norman Ramsey


P.S. Why I want to know: last year I taught a course in which students
built a compiler and VM inspired by Lua's VM.  Performance was quite
good except around the matter of literals; my students' compilers
always compiled code like `n = n + 1` into two VM instructions.
Their codes were otherwise competitive with Lua.  So I'm trying to
figure out if they can eliminate that extra instruction without making
things too complicated.