Encode ADIW/SDIW instructions correctly #13
Description
In the machine code generator, the ADIW
and SDIW
instructions encode their immediate operand as the source operand.
The instructions are defined like so:
let Constraints = "$src = $dst",
Defs = [SREG] in
def ADIWRdK : FWRdK<0b0,
(outs IWREGS:$dst),
(ins IWREGS:$src, i16imm:$src2),
"adiw\t$dst, $src2",
[(set IWREGS:$dst, (add IWREGS:$src, uimm6:$src2)),
(implicit SREG)]>;
There are three operands: $dst
, $src
, and $src2
, with $dst
being constrained to be the same register as $src
.
In the FWRdK
format:
class FWRdK<bit f, dag outs, dag ins, string asmstr, list<dag> pattern>
: AVRInst16<outs, ins, asmstr, pattern>
{
bits<5> d; // mapped to dst register
bits<6> K; // mapped to immediate value
let Inst{15-9} = 0b1001011;
let Inst{8} = f;
let Inst{7-6} = K{5-4};
let Inst{5-4} = d{2-1};
let Inst{3-0} = K{3-0};
}
$src
and $dst
are different operands, even though they are equal. d
is mapped to the first operand, and K
is mapped to the second. The first operand is actually $dst
and the second is $src
, and the third is the immediate value.
Thus, the register encoding is written when the immediate value should be, and the immediate value is ignored.
This is the code generated by tablegen
:
case AVR::ADIWRdK:
case AVR::SBIWRdK: {
// op: d
op = getMachineOpValue(MI, MI.getOperand(0), Fixups, STI);
Value |= (op & UINT64_C(6)) << 3;
// op: K
op = getMachineOpValue(MI, MI.getOperand(1), Fixups, STI);
Value |= (op & UINT64_C(48)) << 2;
Value |= op & UINT64_C(15);
break;
}
One way to fix the problem is to manually edit the
getMachineOpValue(MI, MI.getOperand(1), Fixups, STI)
into
getMachineOpValue(MI, MI.getOperand(2), Fixups, STI)
in the generated file, but this is not an acceptable solution.
This problem could have a simple solution, but I am still in the process of learning the tablegen format and there isn't much documentation.