[llvm] [AArch64][GlobalISel] Take abs scalar codegen closer to SDAG (PR #84886)

[David Green via llvm-commits]

================
@@ -3914,8 +3914,12 @@ LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT LowerHintTy) {
   case G_SSHLSAT:
   case G_USHLSAT:
     return lowerShlSat(MI);
-  case G_ABS:
+  case G_ABS: {
+    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
+    if (Ty.isScalar() &&  LI.isLegalOrCustom({G_ICMP, {Ty, Ty}}))
+        return lowerAbsToCNeg(MI);
----------------
davemgreen wrote:

I personally don't know if there is a good check that captures what we would want through the isLegalOrCustom interface. I think many targets would want the ashr/AddXor form even if they have legal sub+sel. I would remove this code and just keep the AArch64LegalizerInfo::legalizeCustom method.

https://github.com/llvm/llvm-project/pull/84886