[LLVMdev] Relative addressing

[Chris Lattner]

On Wed, 10 Aug 2005, Tzu-Chien Chiu wrote:
> The destination register and each of the source registers can be
> relatively addressed by some special "address registers": a0, a1, a2,
> a3.
> suppose a0 = 4, a1 = 2, the instruction
>  add r[a0], r[a1], r7
> equals to:
>  r4 = r2 + r7

ok

> How should the instruction be defined in TableGen *.td file? If an
> opcode is defined for each variant (like X86InstrInfo.td) there will
> be eight opcodes for each binary operation, and 16 opcodes for each
> trinary operation!
>
>  { relative, not-relative } *  { relative, not-relative } *  {
> relative, not-relative } = 8

Yup.

>  That is, opcodes:
>
>    ADD_x_x_x
>    ADD_r_x_x
>    ADD_x_r_x
>    ADD_x_x_r
>    ADD_x_r_r
>    ADD_r_x_r
>    ADD_r_r_x
>    ADD_r_r_r
>
>  (x: not relative addressing,  r: relative addressing)

I think this (though it is painful) really is the best way to go.  The 
different operations have different semantics that should be reflected by 
their opcode #.

Having a large number of opcodes should not cause any problem for the code 
generator (X86 has many as well), it is just a pain to write by hand.  I 
would suggest using a perl script or something to help generate the 
initial ones.  In time we could add some sort of macro expansion support 
to tblgen to help with this, but we don't currently have anything that 
will help.

-Chris

> Alternatively, the code selector could be designed to generate 'mov'
> instruction with relative addressing only. The above 'add' instruction
> is translated to:
>
>  mov r1, r[a0]
>  mov r2, r[a1]
>  add r1, r2, r7
>
> but it makes the register allocation/assignment and instruction
> scheduling difficult. For the register allocation, one more register
> will be needed. For the instruction scheduling, the 'add' and  the
> associated 'mov's must be moved altogehter.
>
>

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/