Back end with special loop instructions

Project Infrastructure LLVM Dev List Archives
1

Hello.
     I'm writing a back end for my research SIMD processor that has an assembly language that is blocked structured, with one-level loops. An example program with my assembly language:
       REPEAT_X_TIMES(Param2)
         R0 = LS[offset_A];
       END_REPEAT;

     The LLVM code somewhat equivalent to the above ASM program is:
       vector.body:
         %index = phi i64 [ %index.unr, %vector.body.preheader.split.split ], [ %index.next.3, %vector.body ]
         %20 = getelementptr inbounds i32, i32* %A, i64 %index
         %21 = bitcast i32* %20 to <16 x i32>*
         %wide.load = load <16 x i32>, <16 x i32>* %21, align 4
         br i1 %48, label %middle.block.unr-lcssa, label %vector.body, !llvm.loop !3

     How do you suggest to attack this problem?
     I guess I need to provide custom matching code in the Select() function of the back end prior to the SelectCode() invocation in order to translate the label vector.body in LLVM/Machine Instr(DAG) before selection to a REPEAT_X_TIMES(...) instruction - are you aware if anybody else has done such a thing? Also, the br LLVM instruction will be translated to an END_REPEAT.

   Thank you,
     Alex

2

Hi Alex,

You might find it useful to look at how lib/Target/PowerPC/PPCCTRLoops.cpp works.

-Hal

3

Hello.
     Hal, the source file you mention (lib/Target/PowerPC/PPCCTRLoops.cpp) makes use of LLVM IR intrinsics, in this case defined at [LLVM_repo]/llvm/include/llvm/IR/IntrinsicsPowerPC.td, such as:
           // Intrinsics used to generate ctr-based loops. These should only be
    // generated by the PowerPC backend!
    def int_ppc_mtctr : Intrinsic<, [llvm_anyint_ty], >;
    def int_ppc_is_decremented_ctr_nonzero : Intrinsic<[llvm_i1_ty], , >;

     And these intrinsics are instantiated in the LLVM program, for example with the following code in PPCCTRLoops.cpp:
       IRBuilder<> CountBuilder(Preheader->getTerminator());
       Module *M = Preheader->getParent()->getParent();
       Value *MTCTRFunc = Intrinsic::getDeclaration(M, Intrinsic::ppc_mtctr,CountType);
       CountBuilder.CreateCall(MTCTRFunc, ECValue);

     I have defined also some intrinsics for my loop instructions in my file Intrinsics_Connex.td: 1 intrinsic for REPEAT_X_TIMES and 1 for END_REPEAT.
       /* following Intrinsics.td:
       class Intrinsic<list<LLVMType> ret_types,
                 list<LLVMType> param_types = ,
                 list<IntrinsicProperty> properties = ,
                 string name = "">
       */
       def int_connex_repeat_x_times : Intrinsic<, , >;
       def int_connex_end_repeat : Intrinsic<[llvm_i1_ty], , >;
    and added C++ code doing CreateCall() like the one above.

     I'm looking now at http://llvm.org/docs/ExtendingLLVM.html on how to specify the instruction selection of this intrinsic. They write there:
       "Once the intrinsic has been added to the system, you must add code generator support for it. Generally you must do the following steps:
       Add support to the .td file for the target(s) of your choice in lib/Target/*/*.td.
       This is usually a matter of adding a pattern to the .td file that matches the intrinsic, though it may obviously require adding the instructions you want to generate as well. There are lots of examples in the PowerPC and X86 backend to follow."

     Then in my pass I create a call to the respective intrinsic:
      /* See http://llvm.org/docs/doxygen/html/classllvm_1_1TargetIntrinsicInfo.html : "Create or insert an LLVM Function declaration for an intrinsic, and return it."
       * This creates a line in the LLVM program like: declare void @llvm.connex.repeat.x.times() #2 .
       * This line is required, otherwise llc will complain:
       * <<error: use of undefined value '@llvm.connex.repeat.x.times'
             call void @llvm.connex.repeat.x.times()>>
       */
       Value *repeatFunc = Intrinsic::getDeclaration(M,
                                                 Intrinsic::connex_repeat_x_times);

       // See http://llvm.org/docs/doxygen/html/classllvm_1_1IRBuilder.html
       aB.CreateCall(repeatFunc); //, ECValue);

     Then, in the back end, in InstrInfo.td I write:
       let hasSideEffects = 1, isCodeGenOnly = 1 in {
       //let Pattern = [(int_connex_repeat_x_times)] in
       def REPEAT_X_TIMES : ImmediateInstruction< 0b111111,
                                             (outs),
                                             (ins),
                                             "REPEAT_X_TIMES(1001)",
                                             [(int_connex_repeat_x_times)] >;
       }

     Then, when I run opt and llc I obtain the expected behavior.

   Thank you,
     Alex

4

Hello.
     I come back to this thread. But I want to ask a slightly different question.
     Is there a way to have LLVM IR language intrinsics that are given at construction time a string that is written at assembly generation time as it is? (so, basically having placeholders of strings in LLVM that remain untouched until the end, including code generation time.)

     More exactly, I would like to give something LIKE this in a middle-tier pass (I haven't tried this code in LLVM):
      Value *instrinsicFunc = Intrinsic::getDeclaration(M,
                                                 Intrinsic::my_intrinsic);
      myIRBuilder.CreateCall(repeatFunc,
          "string_that_llc_codegen_should_output_as_assembly");
    So my intrinsic would be accepting a general string which will be used at codegen to be output. I guess I would have to define it like this in a .td file:
     def int_my_intrinsic : Intrinsic<, [string], >;

     The other option that can work but is more complex without any benefit is to define several intrinsics in the .td files, one for EACH possible string:
       // To put in the Intrinsics_....td file:
       def int_my_intrinsic1 : Intrinsic<, , >;
       ...
       def int_my_intrinsicN : Intrinsic<, , >;
       // To put in the InstrInfo.td file something like:
       def STR_PLACEHOLDER1 : ImmediateInstruction< 0b111111,
                                               (outs),
                                               (ins),
                                "string_that_llc_codegen_should_output_as_assembly1",
                                             [(int_my_intrinsic1)] >;
       def STR_PLACEHOLDER2 : ImmediateInstruction< 0b111110,
                                               (outs),
                                               (ins),
                                "string_that_llc_codegen_should_output_as_assembly2",
                                             [(int_my_intrinsic2)] >;

   Best regards,
     Alex

5

Hi,

Is there a reason that you can’t use inline assembly for this?

David

6

Hello.
     David, thank you for pointing out inline assembly expressions.
     I found: http://clang.llvm.org/compatibility.html#inline-asm and http://llvm.org/docs/LangRef.html#inline-assembler-expressions .

     So I added in my LLVM program a line like
      call void asm sideeffect "MY OWN STRING", "~{dirflag},~{fpsr},~{flags}"() #1
     and compiled it with llc and it worked as expected.

     More importantly, to programmatically insert assembly expressions in an LLVM pass we need to give something like:
       // Inspired from http://reviews.llvm.org/D15399?id=42353 (see also Insert inline assembly expressions using Llvm pass - Stack Overflow)
       std::vector<llvm::Value*> Args;
       Type *ResultType;
       ResultType = Type::getVoidTy(getGlobalContext());

       // See http://llvm.org/docs/doxygen/html/classllvm_1_1FunctionType.html
       FunctionType *fTy = FunctionType::get(
             /* Result */ ResultType,
             /* NO Params */
             false);
       StringRef asmString = "test_ASM_inline_string!";
       bool hasSideEffect = true;
       InlineAsm::AsmDialect asmDialect = InlineAsm::AD_ATT;

       /*
        * Note: Extreme care with the constraints.
        * Info about them at http://llvm.org/docs/LangRef.html#inline-asm-constraint-string
        * (and https://doc.rust-lang.org/book/inline-assembly.html)

7

Hello.
     I come back to this older thread.
     I would like rephrase the quesion: I want to add strings (containing arithmetic expressions, for example) for immediate operands for MachineInstr (possibly also MachineSDNode and LLVM IR Instruction) in LLVM since my assembly is let's say interpreted at runtime.
     For this I already used inline assembly expressions (or directly SDNode with the opcode ISD::INLINEASM), but it's difficult to make sure they stick together with their associated Connex instructions (VLOAD, setlc, etc) during all the scheduling and Register-allocation passes for llc.

     Did anybody else deal with a similar issue?

   Thank you,
     Alex