Lookup tables and addressing modes

[Ernst Blofeld]

What's the biggest lookup table I can do and still use it efficiently with (d0, a0) style addressing?

Assuming a table of words, I think the answer is 16K table entries.

This is based on the d0 part being signed, so if the table only goes up, then the max is 0x7fff, but words mean 0x3fff, which is the 16K above.

But, if I'm allowed signed values, then I can go between -16K and 16K-1, right?

So, if I really believe I need all that accuracy, then should I change all my code to use values of [-pi..pi) rather than [0..2pi)?

But, I can't do this in C, can I?

This question has come up today, twice, because the atan2 code I was thinking of using assumes pi is 16384, and returns values between 0 and 32,767 inclusive, and because converting yaw, pitch, roll to a quaternion involves half angles, so 1/2 = 0 becomes a problem.

[ross]

It's not a problem also if you got d0.w containing unsigned values and use 32K word entries.

Simply init a0 to [start+32768] and construct the table considering that $8000 is at a0-$8000 and on.

No idea how to do it in C.

[Ernst Blofeld]

Originally Posted by ross

No idea how to do it in C.

I can vasm in a bit of assembly, and put the label somewhere in the middle?

[ross]

Originally Posted by Ernst Blofeld

I can vasm in a bit of assembly, and put the label somewhere in the middle?

If you add vasm code there is no problem at all (you can pass directly the right address ad use the a0,d0.w addressing mode).

In C probably full 32 bit from d0 are used (and generated code to handle this is sure slower)..

Anyway, a0,d0.w and a0,d0.l execute in the same number of cycles.

[Ernst Blofeld]

Originally Posted by ross

If you add vasm code there is no problem at all (you can pass directly the right address ad use the a0,d0.w addressing mode).

In C probably full 32 bit from d0 are used (and generated code to handle this is sure slower)..

Anyway, a0,d0.w and a0,d0.l execute in the same number of cycles.

There's an a0,d0.l?

[Don_Adan]

Originally Posted by Ernst Blofeld

There's an a0,d0.l?

Yes, theoretical maximum table can be 4GB. Exist also a0,a1.w or a0,a1.l.

[Ernst Blofeld]

Originally Posted by Don_Adan

Yes, theoretical maximum table can be 4GB. Exist also a0,a1.w or a0,a1.l.

You guys are talking 68K, not 020+, right?

[Don_Adan]

Originally Posted by Ernst Blofeld

You guys are talking 68K, not 020+, right?

Yes, 68000. 4GB is theory only, because 68000 can not address more than 16MB. f.e move.b (A0,D0.L),D1 for 68000 is limited to 16MB, but for 68020+ for 4GB.

[Ernst Blofeld]

Originally Posted by Don_Adan

Yes, 68000. 4GB is theory only, because 68000 can not address more than 16MB. f.e move.b (A0,D0.L),D1 for 68000 is limited to 16MB, but for 68020+ for 4GB.

Right then. I wonder where my assumption that the offset was limited to word size came from.

[a/b]

That's why I'm so annoyed when I see (xx,ax,d0). Why do people insist on writing ambiguous and potentially buggy code and relying on assembler defaults (that may or may not work with different assemblers), and then eventually unexperienced people see that and start doing the same?
And if I see moveq.l or lea.l in the same code, it's time for painkillers.

[Jobbo]

Originally Posted by Ernst Blofeld

Right then. I wonder where my assumption that the offset was limited to word size came from.

When writing optmized code, it's often slightly more work to make sure the top half of the index register isn't trash. So there is some cost involved if you use a long index.