Opt compiler: Speedup div/rem by constants on arm32 and arm64.

This patch also includes:
1. Add java test for div/rem negative constants.
2. Fix a thumb2 encoding issue where the last operand is
   "reg, shift #amount" in some instructions.
3. Support a simple filter in arm32 assembler test to filter out
   unsupported cases, such as "smull r0, r0, r1, r2".
4. Add smull arm32 assembler test.
5. Add smull/umull thumb2 test.
6. Add test for the thumb2 encoding issue which is fixed in this
   patch.

Change-Id: I1601bc9c38f70f11909f2816fe3ec105a158951e
diff --git a/compiler/optimizing/code_generator_arm.h b/compiler/optimizing/code_generator_arm.h
index 1a498e1..2edbcf8 100644
--- a/compiler/optimizing/code_generator_arm.h
+++ b/compiler/optimizing/code_generator_arm.h
@@ -189,6 +189,10 @@
                              Label* true_target,
                              Label* false_target,
                              Label* always_true_target);
+  void DivRemOneOrMinusOne(HBinaryOperation* instruction);
+  void DivRemByPowerOfTwo(HBinaryOperation* instruction);
+  void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
+  void GenerateDivRemConstantIntegral(HBinaryOperation* instruction);
 
   ArmAssembler* const assembler_;
   CodeGeneratorARM* const codegen_;