[llvm] [AArch64][SelectionDAG][SVE] Fold following extend(icmp) patterns (PR #192052)

[Sushant Gokhale via llvm-commits]

================
@@ -24422,6 +24437,105 @@ static SDValue performExtendCombine(SDNode *N,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     SelectionDAG &DAG,
                                     const AArch64Subtarget *Subtarget) {
+  // sext(icmp slt a, b) --> asr (shsub a, b), bitwidth-1
+  // sext(icmp ult a, b) --> asr (uhsub a, b), bitwidth-1
+  // zext(icmp slt a, b) --> lsr (shsub a, b), bitwidth-1
+  // zext(icmp ult a, b) --> lsr (uhsub a, b), bitwidth-1
+  // sext(icmp sgt a, b) --> asr (shsub b, a), bitwidth-1
+  // sext(icmp ugt a, b) --> asr (uhsub b, a), bitwidth-1
+  // zext(icmp sgt a, b) --> lsr (shsub b, a), bitwidth-1
+  // zext(icmp ugt a, b) --> lsr (uhsub b, a), bitwidth-1
+  SDValue Compare = N->getOperand(0);
+  SDLoc DL(N);
+  if (Compare.getOpcode() == AArch64ISD::SETCC_MERGE_ZERO &&
----------------
sushgokh wrote:

Having all true pred creates one more `ptrue`  instruction in addition to what is already present for `cmphi` instruction and ptrue is not very cheap(lat=2). Rather, I am using the same pred as the `cmp` instruction.

Regarding one use, the compare instruction isnt eliminated here. It remains if there are other uses of it. I am just intercepting the extend instruction.

https://github.com/llvm/llvm-project/pull/192052