# HG changeset patch
# User Jin
# Date 1462760274 14400
# Node ID e26ad49b719414c091797674dfc1bf20485e9bfa
# Parent  fbd9470e188d645f7bfcd8225bc69c1725ae75e2
[C2] disjoint_byte_copy: fix unaligned LW/SW

/mnt/j2sdk-image/bin/java \
  -XX:+PrintCompilation \
  -XX:CompileCommand="compileonly,java/util/jar/Manifest\$FastInputStream.*" \
  -jar SPECjvm2008.jar -ikv -coe -ict -bt 1 \
  compress

It triggers 35000 unaligned instructions.

280 CALL_LEAF_NOFP,runtime jbyte_arraycopy

a0: (-122610748, .) fc127a6b pc: ffe9809004   <-- src是个奇地址
a1: (-122610748, .) fc128100 pc: ffe980910c   <-- dest是个偶地址
a2: (-122610748, .) 2b pc: ffe9809210
t3: (-122610748, .) fc127a6c pc: ffe98093a4   <-- 试图把src对齐4字节，
t0: (-122610748, .) fc128101 pc: ffe98094a8   <-- 则dest就不对齐了，使用sw就出了unalign
t3: (-122610747, *) fc127a70 pc: ffe98093a4
t0: (-122610747, .) fc128105 pc: ffe98094a8   <-- 后面所有对dest的访问都会unalign
t3: (-122610746, *) fc127a74 pc: ffe98093a4
t0: (-122610746, .) fc128109 pc: ffe98094a8

Effect:
* 0 unaligned instructions.

diff -r fbd9470e188d -r e26ad49b7194 src/cpu/mips/vm/stubGenerator_mips_64.cpp
--- a/src/cpu/mips/vm/stubGenerator_mips_64.cpp	Sun May 08 17:53:51 2016 -0400
+++ b/src/cpu/mips/vm/stubGenerator_mips_64.cpp	Sun May 08 22:17:54 2016 -0400
@@ -638,6 +638,12 @@
 	  __ blez(AT, l_4);  
 	  __ delayed()->nop();	
 	  if (!aligned) {
+	    // 2016/5/8 Jin: only when src and dest has the same alignment can we do lw/sw */
+	    __ andi(AT, T3, 3); 
+	    __ andi(T9, T0, 3); 
+	    __ bne(AT, T9, l_5); 
+	    __ delayed()->nop();	
+	  
 	    // align source address at dword address boundary
 	    __ move(T1, 4); 
 	    __ sub(T1, T1, T3); 
@@ -711,6 +717,7 @@
   //
   address generate_conjoint_byte_copy(bool aligned, const char *name) {
 		Label l_1, l_2, l_3, l_4, l_5;
+		Label l_unaligned, l_aligned;
 		StubCodeMark mark(this, "StubRoutines", name);
 		__ align(CodeEntryAlignment);
 		address start = __ pc();
@@ -725,19 +732,42 @@
 		__ push(T1);	
 		__ push(T8);	
 
-
 		// copy from high to low
 		__ move(T3, A0); 
 		__ move(T0, A1);
 		__ move(T1, A2);  
-		__ dadd(AT, T3, T1);  
-		__ lea(T3, Address(AT, -4));
-		__ dadd(AT, T0, T1);  
-		__ lea(T0, Address(AT, -4));
+		__ dadd(T3, T3, T1);  
+		__ dadd(T0, T0, T1);  
+
+		// 2016/5/8 Jin: copy starting unalinged bytes
+		__ bind(l_unaligned);
+		__ beq(T1, R0, l_5); 
+		__ delayed()->nop(); 
+
+		__ andi(AT, T3, 3);    
+		__ beq(AT, R0, l_aligned); 
+		__ delayed()->nop(); 
+		__ lb(AT, T3, -1);   
+		__ sb(AT, T0, -1); 
+		__ daddi(AT, T1, -1); 
+		__ daddi(AT, T3, -1); 
+		__ daddi(AT, T0, -1); 
+		__ b(l_unaligned); 
+		__ delayed()->nop(); 
+		
+		// now T0, T3 point to 4-byte aligned high-ends
+		//     T1 contains byte count that is not copied.
+		__ bind(l_aligned);
+
 		__ move(T8, T1); 
 		__ daddi(AT, T1, -3); 
 		__ blez(AT, l_3); 
 		__ delayed()->nop();	
+
+		__ andi(T8, T8, 3); 
+		__ lea(T3, Address(T3, -4));
+		__ lea(T0, Address(T0, -4));
+
 		__ dsrl(T1, T1, 2); 
 		__ align(16);
 		__ bind(l_1);