[FRONTEND][BACKEND] Fixes for cat / reshape / addptr (#959)

Most notably, this PR:
- changes the traits (and assembly format) of addptr so it can handle offsets that have arbitrary integer width.
- adds support for `cat`

lib/Dialect/Triton/IR/Traits.cpp

@@ -19,7 +19,7 @@ mlir::OpTrait::impl::verifySameOperandsAndResultEncoding(Operation *op) {
   for (auto resultType : op->getResultTypes())
     if (failed(verifySameEncoding(resultType, type)))
       return op->emitOpError()
-             << "requires the same shape for all operands and results";
+             << "requires the same encoding for all operands and results";
   return verifySameOperandsEncoding(op);
 }
 

lib/Dialect/Triton/Transforms/Combine.cpp

@@ -196,7 +196,7 @@ public:
     patterns.add<CombineDotAddFRevPattern>(context);
     // %}
     patterns.add<CombineSelectMaskedLoadPattern>(context);
-    patterns.add<CombineAddPtrPattern>(context);
+    // patterns.add<CombineAddPtrPattern>(context);
     patterns.add<CombineBroadcastConstantPattern>(context);
 
     if (applyPatternsAndFoldGreedily(m, std::move(patterns)).failed())

lib/Dialect/Triton/Transforms/Combine.td

@@ -29,13 +29,14 @@ def CombineDotAddFRevPattern : Pat<
         (TT_DotOp $a, $b, $d, $allowTF32),
         [(Constraint<CPred<"isZero($0)">> $c)]>;
 
-
+// TODO: this fails for addptr(addptr(ptr, i32), i64)
+// Commented out until fixed
 // addptr(addptr(%ptr, %idx0), %idx1) => addptr(%ptr, AddI(%idx0, %idx1))
 //   Note: leave (sub %c0, %c0) canceling to ArithmeticDialect
 //         (ref: ArithmeticCanonicalization.td)
-def CombineAddPtrPattern : Pat<
-        (TT_AddPtrOp (TT_AddPtrOp $ptr, $idx0), $idx1),
-        (TT_AddPtrOp $ptr, (Arith_AddIOp $idx0, $idx1))>;
+// def CombineAddPtrPattern : Pat<
+//         (TT_AddPtrOp (TT_AddPtrOp $ptr, $idx0), $idx1),
+//         (TT_AddPtrOp $ptr, (Arith_AddIOp $idx0, $idx1))>;
 
 // broadcast(cst) => cst
 def getConstantValue : NativeCodeCall<"getConstantValue($_builder, $0, $1)">;