[CIR][CIRGen][Builtin][Neon] Lower neon_vqmovns_s32 and add CIR PoisonAttr #1199
Conversation
ghehg
force-pushed
the
bigMac
branch
3 times, most recently
from
3dc47ce to
7157833
Compare
|
Please introduce PoisonAttr, just like we have a UndefAttr (introduced by @smeenai). The idea is that such attribute could be used by both a full vector type or a regular type used in a vector lane, just like UndefAttr can. |
Sure, sounds best approach to me. let's do it. I'll create a vector type constantOP with thsi PoisonAttr, and it will become LLVM::PoisonOp after lowering. That CIR constant OP will be used as input to VecCreateOp, so we also change VecCreateOp lowering, so it won't generate initialization code during lowering when the input is poison constant. |
ghehg
force-pushed
the
bigMac
branch
2 times, most recently
from
7ad1b26 to
3c21cec
Compare
|
Updated the PR using |
ghehg
changed the title
…nAttr (#1199) CIR PoisonOp is needed in this context as alternative would be to use VecCreateOp to prepare an arg for VecInsertElement, but VecCreate is for different purpose and [it would insert all elements](https://github.com/llvm/clangir/blob/eacaabba76ebdbf87217fefaa77f92c45cf4509c/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp#L1679) which is not totally unnecessary in this context. Here is the [intrinsic def ](https://developer.arm.com/architectures/instruction-sets/intrinsics/#f:@navigationhierarchiessimdisa=[Neon]&q=vqmovns_)
CIR PoisonOp is needed in this context as alternative would be to use VecCreateOp to prepare an arg for VecInsertElement, but VecCreate is for different purpose and it would insert all elements which is not totally unnecessary in this context.
Here is the intrinsic def