[llvm-dev] Special vselect code generation case using TableGen
Alex Susu via llvm-dev
llvm-dev at lists.llvm.org
Wed Oct 12 16:50:21 PDT 2016
Hello.
I would like to ask a few more things related to the subject of manually generating
SelectionDAG nodes (in C++, not with TableGen):
- can I add new use edges to an already created node (that is the existing node to
use some more values from some other node)?
- is it OK to add a chain or glue edge between a node with a chain/glue output and
a node that has no chain/glue input?
- is it OK to add a chain or glue output to a machine instruction defined in
TableGen that does not have any output?
Thank you,
Alex
On 10/10/2016 9:54 PM, Alex Susu wrote:
> Hello.
> I addressed the problem in the end without using TableGen - in
> ISelLowering::LowerOperation(). It turned out that writing code to generate about 10
> SDNode and especially linking them correctly is quite complicated (a few details at
> https://sites.google.com/site/alexsusu/home/backend-llvm, look for "P27. Making VSELECT
> instruction selection work").
> I guess that one could do in TableGen, but I'm not very good at it and there are some
> difficulties: I need to create an SDNode of a certain target instruction depending on the
> predicate of the SETCC node being fed to the VSELECT.
>
> I would like to mention that an interesting example is given in the book Mayur Pandey,
> Suyog_Sarda, "LLVM Cookbook", 2015, page 226, section "Lowering to multiple instructions":
> it describes how we can do in InstrInfo correct instruction selection by first lowering a
> MOVi32 to MOVLOi16 and MOVHIi16 .
>
> Best regards,
> Alex
>
>
>> Could you do this in ISelLowering/DAGToDAG? This is probably where I would do this. I
>> think it would be cleaner there.
>>
>> On Oct 3, 2016 19:06, "Alex Susu via llvm-dev" <llvm-dev at lists.llvm.org
>> <mailto:llvm-dev at lists.llvm.org>> wrote:
>>
>> Hello.
>> I want to implement in TableGen code generation for vselect for the Connex
>> research SIMD processor. The problem is that my SIMD processor does not have a vselect
>> instruction and therefore we need to use the other existing instructions. More
>> exactly, if we have in LLVM IR a program like this:
>> %pred = cmp eq %op1, %op2
>> %vres = vselect %pred, %if_set, %if_clear
>> we can implement it in Connex like this:
>> EQ op1, op2
>> Rdst = Rif_clear
>> WHEREEQ
>> Rdst = Rif_set
>> ENDWHERE
>>
>> Normally TableGen allows specifying single instructions that match part of a given
>> DAG.
>> But in include/llvm/Target/TargetSelectionDAG.td it is given the possibility to
>> match a given DAG node with several target nodes by using Pattern<> records:
>> // Selection DAG Pattern Support.
>> // Patterns are what are actually matched against by the target-flavored
>> // instruction selection DAG. Instructions defined by the target implicitly
>> // define patterns in most cases, but patterns can also be explicitly added when
>> // an operation is defined by a sequence of instructions (e.g. loading a large
>> // immediate value on RISC targets that do not support immediates as large as
>> // their GPRs).
>> //
>> class Pattern<dag patternToMatch, list<dag> resultInstrs> {
>> dag PatternToMatch = patternToMatch;
>> list<dag> ResultInstrs = resultInstrs;
>> list<Predicate> Predicates = []; // See class Instruction in Target.td.
>> int AddedComplexity = 0; // See class Instruction in Target.td.
>> }
>>
>> // From MispMSAInstrInfo.td: (vselect cond, if_set, if_clear)
>> def : Pattern<(set MSA128BOpnd:$wd, (vselect (MSA128BOpnd:$pred),
>> (MSA128BOpnd:$ws_true), MSA128BOpnd:$ws_false)),
>> [ // assign value $ws_false to vregx, which keeps the result of
>> vselect - use CopyToReg
>> (WHEREEQ), // no inputs, no outputs
>> // assign value $ws_true to vregx, which keeps the result of
>> vselect - use CopyToReg
>> (ENDWHERE) // no inputs, no outputs
>> ]
>> >;
>>
>> I'm having difficulties in specifying this TableGen Pattern.
>> For example, does the "set" TableGen pattern operation specify CopyToReg?
>>
>> Thank you,
>> Alex
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>> <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
>>
More information about the llvm-dev
mailing list