[llvm-dev] What pattern string corresponds to CopyToReg?

Sun Nov 5 22:04:12 PST 2017

Most instructions produced by isel use a vreg(virtual register) as their
inputs and destination. The register allocator is responsible for assigning
these virtual registers to specific physical registers later. Each virtual
register is assigned a specific register class and the register allocator
has the freedom to allocate a physical register for it within the class.
The CopyToReg and CopyFromReg nodes exist to place constraints on the
register allocator.

Most CopyToReg nodes will appear on the return value of a function or just
before a call to another function. These are used to ensure that things are
placed in the right registers to match the calling convention. It appears
that your calling convention requires result to be returned in R5 so the
CopyToReg was created to make sure the return value is in R5.

Most CopyFromReg nodes appear on the inputs to a function where the
physical register represent the input registers defined by the calling
convention.

The register allocator will see the CopyToReg after the DAG has been
converted to machine instructions at that point it will be
TargetOpcode::COPY with a physical register output and a virtual register
input. In your case the virtual register input comes from the output of
your LI instruction. If the register allocator is able to it will allocate
this virtual register to the same physical register as the output of the
COPY as it did in your case. But in more complicated functions it might not
be able to do that and will instead create a register to register move to
get the register into the right place.

~Craig

On Sun, Nov 5, 2017 at 5:23 PM, Robert Baruch <robert.c.baruch at gmail.com>
wrote:

> Hmm, well that seems to have worked, but I'm a little suspicious. This is
> my instruction def:
>
> // Load Immediate (LI)
> def LI : IImmReg<0x010, (outs REG16:$dst), (ins i16imm:$src),
>     [(set REG16:$dst, imm:$src)], "LI\t$dst, $src">;
>
>
> The rest of the debug output:
>
> Selecting: t3: ch,glue = CopyToReg t0, Register:i16 %R5, Constant:i16<127>
> Selecting: t2: i16 = Register %R5
> Selecting: t1: i16 = Constant<127>
>
> ISEL: Starting pattern match on root node: t1: i16 = Constant<127>
>   Initial Opcode index to 16
> Creating constant: t5: i16 = TargetConstant<127>
>   Morphed node: t1: i16 = LI TargetConstant:i16<127>
> ISEL: Match complete!
>
> Selected selection DAG: BB#0 'my_func:entry'
> SelectionDAG has 6 nodes:
>       t0: ch = EntryToken
>       t1: i16 = LI TargetConstant:i16<127>
>     t3: ch,glue = CopyToReg t0, Register:i16 %R5, t1
>   t4: ch,glue = RTWP Register:i16 %R5, t3
>
> So at this point, t3 is still a CopyToReg with R5 and another node that is
> an LI with the constant. That doesn't make sense to me. Shouldn't the LI
> node end up being a parent of R5 and the constant?
>
> The phases proceed, and this makes even less sense:
>
> # Machine code for function my_func: IsSSA, TracksLiveness
>
> BB#0: derived from LLVM BB %entry
> %vreg0<def> = LI 127; REG16:%vreg0
> %R5<def> = COPY %vreg0; REG16:%vreg0
> RTWP %R5
>
> # End machine code for function my_func.
>
> I'm not complaining too hard, though, because in the end I get the
> assembly output I expect:
>
> # BB#0:                                 # %entry
> LI $r5, 127
> RTWP
>
> What I'm complaining about is that I still don't know how the DAG nodes
> work. Here's the -view-sched-dags output:
>
>
> [image: dags.jpg]
>
> I just don't understand the LI node. From a dataflow perspective, LI does
> depend on the constant, and not on the register R5. Is this just how LLVM
> represents data stores? With a CopyToReg node?
>
> There's something that I'm missing, but I don't know what it is :(
>
> --Rob
>
> On Sun, Nov 5, 2017 at 1:22 AM 陳韋任 <chenwj.cs97g at g2.nctu.edu.tw> wrote:
>
>> Try match your instruction with (set REG16:$dst, i16imm:$src). Just give
>> it a shot. :)
>>
>> 2017-11-05 12:56 GMT+08:00 Robert Baruch via llvm-dev <
>> llvm-dev at lists.llvm.org>:
>>
>>> Well, that's the thing: I thought that was CopyToReg. I don't know what
>>> the name of the node is to load one value into a register, so I don't know
>>> how to construct such a pattern.
>>>
>>> On Sat, Nov 4, 2017 at 9:23 PM Craig Topper <craig.topper at gmail.com>
>>> wrote:
>>>
>>>> Do you have a pattern for loading an i16 immediate into a 16-bit
>>>> register?
>>>>
>>>> ~Craig
>>>>
>>>> On Sat, Nov 4, 2017 at 8:00 PM, Robert Baruch <
>>>> robert.c.baruch at gmail.com> wrote:
>>>>
>>>>> Hmm, okay. Then what's the problem being reported here? I'm not sure
>>>>> what I'm supposed to do with "LLVM ERROR: Cannot select: t1: i16 =
>>>>> Constant<127>".BTW, the function is:
>>>>>
>>>>> ; ModuleID = 'return.c'
>>>>> source_filename = "return.c"
>>>>> target datalayout = "E-m:e-p:16:16:16-i1:16:16-i8:
>>>>> 16:16-i16:16:16-i32:16:16-i64:16:16-S16-n16"
>>>>> target triple = "tms9900"
>>>>>
>>>>> ; Function Attrs: noinline nounwind optnone
>>>>> define signext i16 @my_func() #0 {
>>>>> entry:
>>>>>   ret i16 127
>>>>> }
>>>>>
>>>>> ------- debug output --------
>>>>>
>>>>> Optimized legalized selection DAG: BB#0 'my_func:entry'
>>>>> SelectionDAG has 5 nodes:
>>>>>     t0: ch = EntryToken
>>>>>   t3: ch,glue = CopyToReg t0, Register:i16 %R5, Constant:i16<127>
>>>>>   t4: ch = TMS9900ISD::Ret t3, Register:i16 %R5, t3:1
>>>>>
>>>>> ===== Instruction selection begins: BB#0 'entry'
>>>>> Selecting: t4: ch = TMS9900ISD::Ret t3, Register:i16 %R5, t3:1
>>>>>
>>>>> ISEL: Starting pattern match on root node: t4: ch = TMS9900ISD::Ret
>>>>> t3, Register:i16 %R5, t3:1
>>>>>
>>>>>   Morphed node: t4: ch,glue = Retr Register:i16 %R5, t3
>>>>>
>>>>> ISEL: Match complete!
>>>>> Selecting: t3: ch,glue = CopyToReg t0, Register:i16 %R5,
>>>>> Constant:i16<127>
>>>>>
>>>>> Selecting: t2: i16 = Register %R5
>>>>>
>>>>> Selecting: t1: i16 = Constant<127>
>>>>>
>>>>> ISEL: Starting pattern match on root node: t1: i16 = Constant<127>
>>>>>
>>>>>   Initial Opcode index to 0
>>>>>   Match failed at index 0
>>>>> LLVM ERROR: Cannot select: t1: i16 = Constant<127>
>>>>> In function: my_func
>>>>>
>>>>>
>>>>> On Sat, Nov 4, 2017 at 7:22 PM Craig Topper <craig.topper at gmail.com>
>>>>> wrote:
>>>>>
>>>>>> CopyToReg is not handle by patterns. It should be passed through isel
>>>>>> unchanged. It’s part of a special list of ISD opcodes that don’t change in
>>>>>> SelectioDAGISel::SelectCodeCommon
>>>>>>
>>>>>> It will then be turned into a TargetOpcode::COPY in
>>>>>> InstrEmitter::EmitSpecialNode when the DAG is turned into MachineInstrs.
>>>>>>
>>>>>> On Sat, Nov 4, 2017 at 7:02 PM Robert Baruch via llvm-dev <
>>>>>> llvm-dev at lists.llvm.org> wrote:
>>>>>>
>>>>>>> So there's a DAG that looks like this in the debug output:
>>>>>>>
>>>>>>>   Selecting: t3: ch,glue = CopyToReg t0, Register:i16 %R5,
>>>>>>> Constant:i16<127>
>>>>>>>
>>>>>>> In the instruction selection phase, what pattern would that match?
>>>>>>> I've constructed this so far:
>>>>>>>
>>>>>>>   (??? REG16:$dst, i16imm:$src)
>>>>>>>
>>>>>>> but the problem is, I can't determine what to use as ???. There is
>>>>>>> an ISD::CopyToReg enum value, but I don't know how that translates
>>>>>>> to the string to use in the pattern.
>>>>>>>
>>>>>>> And more generally, how do I find out from a DAG diagram like the
>>>>>>> ones output by -view-isel-dags which node type corresponds to which
>>>>>>> pattern string?
>>>>>>>
>>>>>>> Thanks,
>>>>>>>
>>>>>>> --Rob
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> LLVM Developers mailing list
>>>>>>> llvm-dev at lists.llvm.org
>>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>>>>
>>>>>> --
>>>>>> ~Craig
>>>>>>
>>>>>
>>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>>>
>>
>>
>> --
>> Wei-Ren Chen (陳韋任)
>> Homepage: https://people.cs.nctu.edu.tw/~chenwj
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20171105/042d1702/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dags.jpg
Type: image/jpeg
Size: 24329 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20171105/042d1702/attachment-0001.jpg>