[PATCH] D148185: Add more efficient bitwise vector reductions on AArch64

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: llvm/lib/Target/AArch64/AArch64ISelLowering.cpp:13252
+    SDValue Extended =
+        DAG.getAnyExtOrTrunc(Vec, DL, VecVT.changeVectorElementType(MVT::i8));
+    switch (ScalarOpcode) {
----------------
Sp00ph wrote:
> Using either zext or sext here adds a few extra instructions in the generated code. Is it guaranteed that any-extending an i1 vector results in a vector whose elements are all either 0 or -1? It seems reasonable because afaik mask vector elements on AArch64 are always either 0 or -1, but it could also introduce some subtle incorrectness if there is some case where any-extending an i1 vector does not result in such a mask vector.
No, no guarantee here.  I mean, there are restrictions related to boolean operands certain specific operations (like the condition of a VSELECT), but there isn't any restriction that applies to arithmetic operations.  An easy way to get a vector with arbitrary data in the high bits is truncating from nxi8 to nxi1.

You could generate a different sequence if the operand is known to be sign-extended (ComputeNumSignBits).


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D148185/new/

https://reviews.llvm.org/D148185