[PATCH] D124657: [AMDGPU][ISEL] Directly custom lower INSERT_SUBVECTOR instead of via INSERT_VECTOR_ELT

[Mahesha S via Phabricator via llvm-commits]

hsmhsm added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIISelLowering.cpp:5801
+      // instead of again taking it through INSERT_VECTOR_ELT.
+      Vec = InsertVecElt(Vec, Elt, CurIdx, SL, DAG);
+    } else {
----------------
foad wrote:
> I don't understand this. CurIdx is a ConstantSDNode here, so InsertVecElt will not do anything useful - it just returns SDValue().
Hmm, I did not realize it. 

But, I was told that returning SDValue() means treat it as legal.

And, hence, no default expansion takes place which otherwise would introduce stack access?

I myself do not understand it - I need to further explore it in detail.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D124657/new/

https://reviews.llvm.org/D124657