[llvm-dev] Supporting RTLIB calls (memset, memcpy, etc)

Alex Susu via llvm-dev llvm-dev at lists.llvm.org
Wed Sep 28 14:15:02 PDT 2016

     Could you please tell me how can I make my new back end (inspired from BPF LLVM back 
end) select runtime libc calls like memset, etc from LLVM IR intrinsics.
     More exactly, I'm getting the following error:
       LLVM ERROR: Cannot select: t18: ch,glue = ConnexISD::CALL t16, 
TargetExternalSymbol:i64'memset', Register:i64 %R1, Register:i64 %R2, Register:i64 %R3, t16:1
     This happens after I:
         - edited method ConnexTargetLowering::ConnexTargetLowering() from file 
lib/Target/Connex/ConnexISelLowering.cpp and registered the RTLIB calls like this:
       // Inspired from lib/Target/ARM/ARMISelLowering.cpp, constructor ARMTargetLowering()
       static const struct {
         const RTLIB::Libcall Op;
         const char *const Name;
         const CallingConv::ID CC;
         const ISD::CondCode Cond;
       } MemOpsLibraryCalls[] = {
         // Memory operations
         { RTLIB::MEMMOVE, "memmove", CallingConv::C, ISD::SETCC_INVALID },
         { RTLIB::MEMSET, "memset", CallingConv::C, ISD::SETCC_INVALID },

       for (const auto &LC : MemOpsLibraryCalls) {
         setLibcallName(LC.Op, LC.Name);
         setLibcallCallingConv(LC.Op, LC.CC);
         if (LC.Cond != ISD::SETCC_INVALID)
           setCmpLibcallCC(LC.Op, LC.Cond);

         - added files
             lib/Target/Connex/ConnexSelectionDAGInfo.cpp, .h inspiring from 
lib/Target/ARM/ARMSelectionDAGInfo.* - I can provide, if required, excerpts from these 
source files.

     Note that I had some success to make opt avoid generating these calls in LLVM IR, 
which obviously avoids instruction selection also, by following advice from:
         - http://lists.llvm.org/pipermail/llvm-dev/2012-May/050122.html : "How do I 
disable that feature? I've tried -fno-builtin and/or -ffreestanding with no success. clang 
(as well as gcc) requires that freestanding environment provides memcpy, memmove, memset 
and memcmp."
         - (http://stackoverflow.com/questions/27511899/llvm-intrinsic-functions: "So to 
prevent (at least the memset) intrinsics in your code don't run instcombine on your IR. 
However, instcombine is a mighty opt pass that realy shortens the code.")

   Thank you,

More information about the llvm-dev mailing list