8272014: Better array indexing

Reviewed-by: andrew

Author: TianminShi

hotspot/src/cpu/x86/vm/c1_LIRGenerator_x86.cpp

@@ -166,8 +166,32 @@ LIR_Address* LIRGenerator::emit_array_address(LIR_Opr array_opr, LIR_Opr index_o
   LIR_Address* addr;
   if (index_opr->is_constant()) {
     int elem_size = type2aelembytes(type);
-    addr = new LIR_Address(array_opr,
-                           offset_in_bytes + index_opr->as_jint() * elem_size, type);
+#ifdef _LP64
+    jint index = index_opr->as_jint();
+    jlong disp = offset_in_bytes + (jlong)(index) * elem_size;
+    if (disp > max_jint) {
+      // Displacement overflow. Cannot directly use instruction with 32-bit displacement for 64-bit addresses.
+      // Convert array index to long to do array offset computation with 64-bit values.
+      index_opr = new_register(T_LONG);
+      __ move(LIR_OprFact::longConst(index), index_opr);
+      addr = new LIR_Address(array_opr, index_opr, LIR_Address::scale(type), offset_in_bytes, type);
+    } else {
+      addr = new LIR_Address(array_opr, disp, type);
+    }
+#else
+    // A displacement overflow can also occur for x86 but that is not a problem due to the 32-bit address range!
+    // Let's assume an array 'a' and an access with displacement 'disp'. When disp overflows, then "a + disp" will
+    // always be negative (i.e. underflows the 32-bit address range):
+    // Let N = 2^32: a + signed_overflow(disp) = a + disp - N.
+    // "a + disp" is always smaller than N. If an index was chosen which would point to an address beyond N, then
+    // range checks would catch that and throw an exception. Thus, a + disp < 0 holds which means that it always
+    // underflows the 32-bit address range:
+    // unsigned_underflow(a + signed_overflow(disp)) = unsigned_underflow(a + disp - N)
+    //                                              = (a + disp - N) + N = a + disp
+    // This shows that we still end up at the correct address with a displacement overflow due to the 32-bit address
+    // range limitation. This overflow only needs to be handled if addresses can be larger as on 64-bit platforms.
+    addr = new LIR_Address(array_opr, offset_in_bytes + index_opr->as_jint() * elem_size, type);
+#endif // _LP64
   } else {
 #ifdef _LP64
     if (index_opr->type() == T_INT) {