diff --git a/include/string.c b/include/string.c
index a105b31..dc5ba4e 100644
--- a/include/string.c
+++ b/include/string.c
@@ -41,6 +41,8 @@ char * strcpy(char * dst, const char * src)
 }
 #endif
 
+#pragma native(strcpy)
+
 #if 1
 
 int strcmp(const char * ptr1, const char * ptr2)
@@ -97,6 +99,7 @@ int strcmp(const char * ptr1, const char * ptr2)
 }
 #endif
 
+#pragma native(strcmp)
 
 int strlen(const char * str)
 {
@@ -212,6 +215,8 @@ void * memcpy(void * dst, const void * src, int size)
 #endif
 }
 
+#pragma native(memcpy)
+
 void * memmove(void * dst, const void * src, int size)
 {
 	char	*	d = dst, * s = src;
diff --git a/oscar64/ByteCodeGenerator.cpp b/oscar64/ByteCodeGenerator.cpp
index 4c4726d..ee1c7aa 100644
--- a/oscar64/ByteCodeGenerator.cpp
+++ b/oscar64/ByteCodeGenerator.cpp
@@ -4952,6 +4952,20 @@ bool ByteCodeBasicBlock::PeepHoleOptimizer(int phase)
 						mBranch = TransposeBranchCondition(mBranch);
 						progress = true;
 					}
+					else if (
+						i + 3 == mIns.Size() && mFalseJump &&
+						mIns[i + 0].mCode == BC_STORE_REG_16 &&
+						mIns[i + 1].mCode == BC_LOAD_REG_16 &&
+						mIns[i + 2].mCode == BC_BINOP_CMPUR_16 && mIns[i + 0].mRegister == mIns[i + 2].mRegister && !(mExitLive & LIVE_ACCU) && mIns[i + 2].mRegisterFinal
+						)
+					{
+						mIns[i + 0].mCode = BC_NOP;
+						mIns[i + 1].mCode = BC_NOP;
+						mIns[i + 2].mRegister = mIns[i + 1].mRegister;
+						mIns[i + 2].mRegisterFinal = mIns[i + 1].mRegisterFinal;
+						mBranch = TransposeBranchCondition(mBranch);
+						progress = true;
+					}
 
 #endif
 					
@@ -5187,6 +5201,16 @@ bool ByteCodeBasicBlock::PeepHoleOptimizer(int phase)
 						mIns[i + 1].mCode = BC_NOP;
 						progress = true;
 					}
+					else if (
+						mIns[i + 0].mCode == BC_LOAD_REG_8 && !(mIns[i + 1].mLive & LIVE_ACCU) && 
+						(mIns[i + 1].mCode == BC_STORE_ABS_8 || mIns[i + 1].mCode == BC_STORE_LOCAL_8 || mIns[i + 1].mCode == BC_STORE_ADDR_8 || mIns[i + 1].mCode == BC_STORE_FRAME_8) &&
+						mIns[i + 1].mRegister == BC_REG_ACCU)
+					{
+						mIns[i + 0].mCode = BC_NOP;
+						mIns[i + 1].mRegister = mIns[i + 0].mRegister;
+						mIns[i + 1].mRegisterFinal = mIns[i + 0].mRegisterFinal;
+						progress = true;
+					}
 
 #if 1
 					else if (
@@ -5603,6 +5627,22 @@ void ByteCodeBasicBlock::CalculateOffset(int& total)
 					mTrueJump->CalculateOffset(total);
 				}
 			}
+#if 1
+			else if (!mTrueJump->mFalseJump && !mFalseJump->mFalseJump && mTrueJump->mTrueJump == mFalseJump->mTrueJump && 
+				mTrueJump->mCode.Size() < 120 && mFalseJump->mCode.Size() < 120 && mTrueJump->mTrueJump->mOffset > mOffset)
+			{
+				// Small diamond so place true then false directly behind each other
+				// with short branches
+
+				mSize = mCode.Size() + 2;
+
+				mFalseJump->mOffset = next + 2;
+				mFalseJump->mSize = mFalseJump->mCode.Size() + 2;
+
+				total = mFalseJump->mOffset + mFalseJump->mSize;
+				mTrueJump->CalculateOffset(total);
+			}
+#endif
 			else
 			{
 				// neither falseJump nor trueJump have been placed
diff --git a/oscar64/GlobalAnalyzer.cpp b/oscar64/GlobalAnalyzer.cpp
index a61ab66..d643d5f 100644
--- a/oscar64/GlobalAnalyzer.cpp
+++ b/oscar64/GlobalAnalyzer.cpp
@@ -119,7 +119,7 @@ void GlobalAnalyzer::AutoInline(void)
 	for (int i = 0; i < mFunctions.Size(); i++)
 	{
 		Declaration* f = mFunctions[i];
-		if (!(f->mFlags & DTF_INLINE) && !(f->mBase->mFlags & DTF_VARIADIC) && !(f->mFlags & DTF_FUNC_VARIABLE) && !(f->mFlags & DTF_INTRINSIC) && f->mCalled.Size() == 0)
+		if (!(f->mFlags & DTF_INLINE) && !(f->mBase->mFlags & DTF_VARIADIC) && !(f->mFlags & DTF_FUNC_VARIABLE) && f->mCalled.Size() == 0)
 		{
 			int		nparams = 0;
 			Declaration* dec = f->mBase->mParams;