================ @@ -88,4 +88,9 @@ def int_dx_rsqrt : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>] def int_dx_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrConvergent]>; def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>; def int_dx_step : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, LLVMMatchType<0>], [IntrNoMem]>; + +def int_dx_splitdouble : DefaultAttrsIntrinsic< + [llvm_anyint_ty, LLVMMatchType<0>], ---------------- farzonl wrote:
so why `llvm_anyint_ty` instead of a vector type? This means the the return type isn't preserving the vector length provided by the double argument. I think this is part of why you are having to do your own custom scalarization. https://github.com/llvm/llvm-project/pull/109331 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
