--- old/src/hotspot/share/opto/mulnode.cpp	2018-11-07 12:23:48.115336259 +0100
+++ new/src/hotspot/share/opto/mulnode.cpp	2018-11-07 12:23:47.973337054 +0100
@@ -478,15 +478,6 @@
   Node *load = in(1);
   uint lop = load->Opcode();
 
-#if INCLUDE_SHENANDOAHGC
-  if (UseShenandoahGC && ShenandoahWriteBarrierNode::is_gc_state_load(load)) {
-    // Do not touch the load+mask, we would match the whole sequence exactly.
-    // Converting the load to LoadUB/LoadUS would mismatch and waste a register
-    // on the barrier fastpath.
-    return NULL;
-  }
-#endif
-
   // Masking bits off of a Character?  Hi bits are already zero.
   if( lop == Op_LoadUS &&
       (mask & 0xFFFF0000) )     // Can we make a smaller mask?
@@ -503,7 +494,10 @@
 
     // Masking sign bits off of a Byte?  Do an unsigned byte load plus
     // an and.
-    if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
+    // Prevent transform if it feeds to CmpI. We have special matcher
+    // for LoadB+AndI+CmpI that generates a single instruction.
+    bool feeds_to_cmpi = outcnt() == 1 && unique_out()->Opcode() == Op_CmpI;
+    if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0 && !feeds_to_cmpi) {
       Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
       ldub = phase->transform(ldub);
       return new AndINode(ldub, phase->intcon(mask));