Emulate i64 add with 3 instructions

Project Infrastructure LLVM Dev List Archives
#1

Hello!

Is there a way to tell LLVM how to emulate an instruction with multiple others? Specifically, in our processor, there is no instruction for adding two i64s; it has to be done like this

dst_high32:dst_low32 = src1_low32 + src2_low32 (unsigned add; dst_high might contain the overflow bit)
dst_high32 = dst_high32 + src1_high32
dst_high32 = dst_high32 + src2_high32

I tried it with patterns like the following in InstrInfo.td (this is obviously wrong, but you have to start somewhere):

def : Pattern<(set LLRegs:$dst, (add LLRegs:$src1, LLRegs:$src2)), [(set LLRegs:$dst, (anyext (add (trunc LLRegs:$src1), (trunc LLRegs:$src2))))]>;

... but tablegen complains: "Cannot use 'add' in an output pattern!". Is this the right way to do it - that is, defining a pattern for an operation? ... Or better: What is the right way to do it?

Any help appreciated!

Thanks,
Johannes

#2

you need to use machine instructions and not SDNodes in the pattern.

Here is an example from our backend:
// integer subtraction
// a - b ==> a + (-b)
def SUB_i8 : Pat<(sub GPRI8:$src0, GPRI8:$src1),
    (ADD_i8 GPRI8:$src0, (NEGATE_i8 GPRI8:$src1))>;

Where ADD_i8 and NEGATE_i8 are machine instructions.

Mica

#3

Hi Johannes,

The usual way to achieve this is to use carry-setting / carry-using instructions, like addc/adde or subc/sube nodes in a dag.

Your target ISelLowering can use them to emulate an i64 unsupported add with 2 x i32 supported add.

You can look how this is done in the MSP430 or ARM backends for example.

Regards,