Inline ResolveMethod and ResolveField in nterp

They are on very hot path (field and method opcodes).

This makes up startup ~1% faster with interpreter.

The methods are fairly big and spill most registers,
but the DexCache-hit fastpath requires no spills.
Rather than refactoring the methods, it is easier
to just inline them, since the caller already spills.

Test: test.py -b --host
Change-Id: I60221e3aa819378ed2f353044cfdfe6cbc51237d
diff --git a/runtime/interpreter/mterp/nterp.cc b/runtime/interpreter/mterp/nterp.cc
index ddef31d..fd57ece 100644
--- a/runtime/interpreter/mterp/nterp.cc
+++ b/runtime/interpreter/mterp/nterp.cc
@@ -20,6 +20,7 @@
 #include "nterp.h"
 
 #include "base/quasi_atomic.h"
+#include "class_linker-inl.h"
 #include "dex/dex_instruction_utils.h"
 #include "debugger.h"
 #include "entrypoints/entrypoint_utils-inl.h"
@@ -264,6 +265,7 @@
   return dex_file->GetShorty(proto_idx);
 }
 
+FLATTEN
 extern "C" size_t NterpGetMethod(Thread* self, ArtMethod* caller, uint16_t* dex_pc_ptr)
     REQUIRES_SHARED(Locks::mutator_lock_) {
   UpdateHotness(caller);
@@ -422,6 +424,7 @@
   }
 }
 
+FLATTEN
 static ArtField* ResolveFieldWithAccessChecks(Thread* self,
                                               ClassLinker* class_linker,
                                               uint16_t field_index,