The case of division is not nearly so symmetric. I also have some bad news for you:
All modern 16-bit CPU's support integer divide. The manufacturer's data sheet will describe this operation as a 32 x 16-bit divide, meaning that you can divide a 32-bit dividend by a 16-bit divisor. Here's the bad news:
THEY'RE LYING TO YOU!!!
If you don't believe it, try dividing any large 32-bit number (meaning that it has non-zero bits in the upper 16 bits) by the integer 1. You are guaranteed to get an overflow exception.
The problem is that the instruction really requires that the resulting quotient fit into a 16-bit result. This won't happen unless the divisor is sufficiently large. When any number is divided by unity, the quotient will of course be the same as the dividend, which had better fit into a 16-bit word.
Since the beginning of time (well, computers, anyway), CPU architects have provided this little gotcha in the division circuitry. It provides a certain amount of symmetry in things, since it is sort of the inverse of the way a multiply works. But since unity is a perfectly valid (and rather common) number to use as a divisor, the division as implemented in hardware needs some help from us programmers.
The implications are as follows:
This looks like a job for another table, to summarize the required actions:
T1 --> | | | | | | | | | | B | W | L | T2 V | | | | ----------+-----------------+-----------------+-----------------+ | | | | B | Convert D0 to W | Convert D0 to W | Convert D0 to L | | Convert D7 to L | Convert D7 to L | | | DIVS | DIVS | JSR DIV32 | | Result = B | Result = W | Result = L | | | | | ----------+-----------------+-----------------+-----------------+ | | | | W | Convert D7 to L | Convert D7 to L | Convert D0 to L | | DIVS | DIVS | JSR DIV32 | | Result = B | Result = W | Result = L | | | | | ----------+-----------------+-----------------+-----------------+ | | | | L | Convert D7 to L | Convert D7 to L | | | JSR DIV32 | JSR DIV32 | JSR DIV32 | | Result = B | Result = W | Result = L | | | | | ----------+-----------------+-----------------+-----------------+ |
You may wonder why it's necessary to do a 32-bit division, when the dividend is, say, only a byte in the first place. Since the number of bits in the result can only be as many as that in the dividend, why bother? The reason is that, if the divisor is a longword, and there are any high bits set in it, the result of the division must be zero. We might not get that if we only use the lower word of the divisor.
The following code provides the correct function for PopDiv:
{ Generate Code to Divide Stack by the Primary }
function PopDiv(T1, T2: char): char;
begin
Pop(T1);
Convert(T1, 'L', 'D7');
if (T1 = 'L') or (T2 = 'L') then begin
Convert(T2, 'L', 'D0');
GenLongDiv;
PopDiv := 'L';
end
else begin
Convert(T2, 'W', 'D0');
GenDiv;
PopDiv := T1;
end;
end; |
The two code generation procedures are:
{ Divide Top of Stack by Primary (Word) } procedure GenDiv; begin EmitLn('DIVS D0,D7'); Move('W', 'D7', 'D0'); end; { Divide Top of Stack by Primary (Long) } procedure GenLongDiv; begin EmitLn('JSR DIV32'); end; |
Note that we assume that DIV32 leaves the (longword) result in D0.
OK, install the new procedures for division. At this point you should be able to generate code for any kind of arithmetic expression. Give it a whirl!