[PATCH] D53257: [ARM][NEON] Improve vector popcnt lowering with PADDL (PR39281)

[Sam Parker via Phabricator via llvm-commits]

samparker added inline comments.


================
Comment at: lib/Target/ARM/ARMISelLowering.cpp:5454
+    SmallVector<SDValue, 8> Ops;
+    Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddlu, DL,
+                                  TLI.getPointerTy(DAG.getDataLayout())));
----------------
RKSimon wrote:
> samparker wrote:
> > For the 64-bit vector case, couldn't we use vpadd instead? We don't care about signed/unsigned, but we'd have to know that the wide result isn't necessary too - which I expect is fine for most bit counting cases.
> Sorry, I don't quite understand - please can you show in the test codegen what you're trying to achieve?
Sorry, that would have been more clear and would have prevented me from asking in the first place! I thought you could use vpadd instead of vpaddl because I didn't realise the output vector properties.



Repository:
  rL LLVM

https://reviews.llvm.org/D53257