llvm dev - Jan 2015 - [LLVMdev] Backend Tablegen Instruction Definition

All, in working through the RISCV LLVM backend, I’m running into some trouble in
defining the instruction formats for the system instruction.  The system
instructions follow a pre-defined instruction template (type-I), but differ in
that they have no input registers (only the target).  The system instructions
are defined as:

rdcycle Rt

I’ve defined a stand-alone instruction definition (as opposed to using the
type-I template) that is as follows:
//rdcycle
def RDCYCLE: InstRISCV<4, (outs GR32:$dst), (ins), ([set GC32:$dst])>{

        field bits<32> Inst;

        bits<5> dst;

        let Inst{31-20} = 0b000000000000;
        let Inst{19-15} = 0b00000;
        let Inst{14-12} = 0b010;
        let Inst{11- 7} = dst;
        let Inst{6 - 0} = 0b1110011;
      }

However, the compilation fails with the follow: 
llvm[3]: Building RISCV.td instruction information with tblgen
llvm-tblgen:
/home/jleidel/dev/working/jleidel-riscv-llvm/riscv-trunk/utils/TableGen/CodeGenDAGPatterns.cpp:161
ode::ApplyTypeConstraints(llvm::TreePattern &, bool): Assertion
`getNumChildren() >= 2 && "Missing RHS of a set

Does the codegen permit me to write instruction definitions manually in this
manner, or do I need to virtualize the instruction given the lack of RHS args?

cheers
john 

John D. Leidel

Hi John,

On 10 January 2015 at 17:33, John Leidel (jleidel) <jleidel at micron.com>
wrote:
> def RDCYCLE: InstRISCV<4, (outs GR32:$dst), (ins), ([set
GC32:$dst])>{
I think the problem is that pattern. For your purposes just using an
empty pattern, "[]", ought to be fine to begin with.

If and when you do decide it needs to be selected, you'll probably go
via an intrinsic at the IR level. Something like "(set GR32:$dst,
int_riscv_read_cycle)", though I've not actually ever tried to match
an intrinsic with *no* arguments before, so you may need to modify
that slightly.

Cheers.

Tim.

Tim, thanks for the response. It seems to build with the pattern omitted. It ll
flys well with the the MC tests, this should finish out the system calls.

Cheers
John

--John D. Leidel
> On Jan 10, 2015, at 8:21 PM, Tim Northover <t.p.northover at
gmail.com> wrote:
> 
> Hi John,
> 
>> On 10 January 2015 at 17:33, John Leidel (jleidel) <jleidel at
micron.com> wrote:
>> def RDCYCLE: InstRISCV<4, (outs GR32:$dst), (ins), ([set
GC32:$dst])>{
> 
> I think the problem is that pattern. For your purposes just using an
> empty pattern, "[]", ought to be fine to begin with.
> 
> If and when you do decide it needs to be selected, you'll probably go
> via an intrinsic at the IR level. Something like "(set GR32:$dst,
> int_riscv_read_cycle)", though I've not actually ever tried to
match
> an intrinsic with *no* arguments before, so you may need to modify
> that slightly.
> 
> Cheers.
> 
> Tim.