[ARM] Add NEON support for ISD::ABDS/ABDU nodes.

As noted on #94466, NEON has ABDS/ABDU instructions but only handles then via intrinsics, plus some VABDL custom patterns.

I'm not clear how ARM handles intrinsic -> ISD mapping, so I've left all the intrinsics handling in place so far at the moment.

Ideally all the VABD/VABA/VABDL/VABAL handling should be moved to using the abds/abdu nodes - but am I on the right track?

Fixes #94466

Author: RKSimon

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -205,9 +205,9 @@ void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT) {
   setOperationAction(ISD::SDIVREM, VT, Expand);
   setOperationAction(ISD::UDIVREM, VT, Expand);
 
-  if (!VT.isFloatingPoint() &&
-      VT != MVT::v2i64 && VT != MVT::v1i64)
-    for (auto Opcode : {ISD::ABS, ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
+  if (!VT.isFloatingPoint() && VT != MVT::v2i64 && VT != MVT::v1i64)
+    for (auto Opcode : {ISD::ABS, ISD::ABDS, ISD::ABDU, ISD::SMIN, ISD::SMAX,
+                        ISD::UMIN, ISD::UMAX})
       setOperationAction(Opcode, VT, Legal);
   if (!VT.isFloatingPoint())
     for (auto Opcode : {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT})

llvm/lib/Target/ARM/ARMInstrNEON.td

@@ -5655,16 +5655,44 @@ def  VABDhq   : N3VQInt<1, 0, 0b11, 0b1101, 0, N3RegFrm, IIC_VBINQ,
                         "vabd", "f16", v8f16, v8f16, int_arm_neon_vabds, 1>,
                 Requires<[HasNEON, HasFullFP16]>;
 
+let Predicates = [HasNEON] in {
+def : Pat<(v2i32 (abds (v2i32 DPR:$opA), (v2i32 DPR:$opB))),
+          (VABDsv2i32 DPR:$opA, DPR:$opB)>;
+def : Pat<(v4i16 (abds (v4i16 DPR:$opA), (v4i16 DPR:$opB))),
+          (VABDsv4i16 DPR:$opA, DPR:$opB)>;
+def : Pat<(v8i8 (abds (v8i8 DPR:$opA), (v8i8 DPR:$opB))),
+          (VABDsv8i8 DPR:$opA, DPR:$opB)>;
+def : Pat<(v4i32 (abds (v4i32 QPR:$opA), (v4i32 QPR:$opB))),
+          (VABDsv4i32 QPR:$opA, QPR:$opB)>;
+def : Pat<(v8i16 (abds (v8i16 QPR:$opA), (v8i16 QPR:$opB))),
+          (VABDsv8i16 QPR:$opA, QPR:$opB)>;
+def : Pat<(v16i8 (abds (v16i8 QPR:$opA), (v16i8 QPR:$opB))),
+          (VABDsv16i8 QPR:$opA, QPR:$opB)>;
+
+def : Pat<(v2i32 (abdu (v2i32 DPR:$opA), (v2i32 DPR:$opB))),
+          (VABDuv2i32 DPR:$opA, DPR:$opB)>;
+def : Pat<(v4i16 (abdu (v4i16 DPR:$opA), (v4i16 DPR:$opB))),
+          (VABDuv4i16 DPR:$opA, DPR:$opB)>;
+def : Pat<(v8i8 (abdu (v8i8 DPR:$opA), (v8i8 DPR:$opB))),
+          (VABDuv8i8 DPR:$opA, DPR:$opB)>;
+def : Pat<(v4i32 (abdu (v4i32 QPR:$opA), (v4i32 QPR:$opB))),
+          (VABDuv4i32 QPR:$opA, QPR:$opB)>;
+def : Pat<(v8i16 (abdu (v8i16 QPR:$opA), (v8i16 QPR:$opB))),
+          (VABDuv8i16 QPR:$opA, QPR:$opB)>;
+def : Pat<(v16i8 (abdu (v16i8 QPR:$opA), (v16i8 QPR:$opB))),
+          (VABDuv16i8 QPR:$opA, QPR:$opB)>;
+}
+
 //   VABDL    : Vector Absolute Difference Long (Q = | D - D |)
 defm VABDLs   : N3VLIntExt_QHS<0,1,0b0111,0, IIC_VSUBi4Q,
                                "vabdl", "s", int_arm_neon_vabds, zext, 1>;
 defm VABDLu   : N3VLIntExt_QHS<1,1,0b0111,0, IIC_VSUBi4Q,
                                "vabdl", "u", int_arm_neon_vabdu, zext, 1>;
 
 let Predicates = [HasNEON] in {
-def : Pat<(v8i16 (abs (sub (zext (v8i8 DPR:$opA)), (zext (v8i8 DPR:$opB))))),
+def : Pat<(v8i16 (zext (abdu (v8i8 DPR:$opA), (v8i8 DPR:$opB)))),
           (VABDLuv8i16 DPR:$opA, DPR:$opB)>;
-def : Pat<(v4i32 (abs (sub (zext (v4i16 DPR:$opA)), (zext (v4i16 DPR:$opB))))),
+def : Pat<(v4i32 (zext (abdu (v4i16 DPR:$opA), (v4i16 DPR:$opB)))),
           (VABDLuv4i32 DPR:$opA, DPR:$opB)>;
 }