[PATCH] D109149: [DAG] Fold select_cc setgt X, -1, C, ~C -> xor (ashr X, BW-1), C

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:22873
+        DAG.getConstant(CmpOpVT.getScalarSizeInBits() - 1, DL, CmpOpVT));
+    return DAG.getNode(ISD::XOR, DL, VT, DAG.getSExtOrTrunc(ASR, DL, VT),
+                       DAG.getSExtOrTrunc(CC == ISD::SETLT ? N3 : N2, DL, VT));
----------------
dmgreen wrote:
> arsenm wrote:
> > Is it really beneficial if this cast is necessary? Should this check the compared and selected types match (or maybe allow truncate, if it's free)
> Any of the selecti32i64 and selecti8i32 tests will involve sign extending, and all look like they produce less or equal numbers of instructions. I've not seen any cases where it made it unprofitable (but may not have tried all possibilities).
For AMDGPU if the compared type was 64-bits, and the selected type was 32-bits, this would be worse. The other way around would be beneficial


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109149/new/

https://reviews.llvm.org/D109149