[llvm-dev] replacing sequence of IR instruction to intrinsic/builtin call

[Oleg Tsibulnyak via llvm-dev]

Hi,please guide me in solving the next task.I want to replace sequence of IR instruction to call of my own procedure,specifically this sequence (it is adding 2 vectors, each consists from 8 elements):
  %31 = getelementptr inbounds i32, i32* %0, i64 %29
  %32 = bitcast i32* %31 to <8 x i32>*
  %33 = load <8 x i32>, <8 x i32>* %32, align 4, !tbaa !1, !alias.scope !5
  %34 = getelementptr inbounds i32, i32* %1, i64 %29
  %35 = bitcast i32* %34 to <8 x i32>*
  %36 = load <8 x i32>, <8 x i32>* %35, align 4, !tbaa !1, !alias.scope !8
  %37 = add nsw <8 x i32> %36, %33
  %38 = getelementptr inbounds i32, i32* %2, i64 %29
  %39 = bitcast i32* %38 to <8 x i32>*
  store <8 x i32> %37, <8 x i32>* %39, align 4, !tbaa !1, !alias.scope !10, !noalias !12

to C procdure likeexec_vect8add(add1stPtr, add2ndPtr, resPtr)

I tried to do this via tablegen (I tried to replace to intrinsic), but cannot compile LLVM.I added to the instruction definition file Cpu0InstrInfo.td next lines:def : Pat<(store (v8i32 (add (v8i32 (load addr:$adr1)), (v8i32 (load addr:$adr2))
                        )
                 ),
             addr:$adrdst
          ),
          (int_cpu0_GEadd addr:$adr1, addr:$adr2, addr:$adrdst)>;

Error during compilation:Included from /home/tsib/LLVM/test/src/lib/Target/Cpu0/Cpu0.td:15:
/home/tsib/LLVM/test/src/lib/Target/Cpu0/Cpu0InstrInfo.td:1125:1: error: In anonymous_537: Cannot use 'int_cpu0_GEadd' in an output pattern!
def : Pat<(store (v8i32 (add (v8i32 (load addr:$adr1)), (v8i32 (load addr:$adr2))


Thanks,Oleg
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170320/fb8d4fcf/attachment.html>