Remove c-11 memcpy checks from perf-critical code

Change-Id: Id4f37f5d4a03160572954a416efa1ef9b3d79ad1
Signed-off-by: Dave Barach <dave@barachs.net>

1 files changed, 212 insertions, 210 deletions

diff --git a/src/vnet/srv6/sr_policy_rewrite.c b/src/vnet/srv6/sr_policy_rewrite.c
index c80b9c4505a..e9c42211209 100755
--- a/src/vnet/srv6/sr_policy_rewrite.c
+++ b/src/vnet/srv6/sr_policy_rewrite.c
@@ -115,7 +115,7 @@ static ip6_address_t sr_pr_encaps_src;
 void
 sr_set_source (ip6_address_t * address)
 {
-  clib_memcpy (&sr_pr_encaps_src, address, sizeof (sr_pr_encaps_src));
+  clib_memcpy_fast (&sr_pr_encaps_src, address, sizeof (sr_pr_encaps_src));
 }
 
 static clib_error_t *
@@ -192,8 +192,8 @@ compute_rewrite_encaps (ip6_address_t * sl)
       addrp = srh->segments + vec_len (sl) - 1;
       vec_foreach (this_address, sl)
       {
-	clib_memcpy (addrp->as_u8, this_address->as_u8,
-		     sizeof (ip6_address_t));
+	clib_memcpy_fast (addrp->as_u8, this_address->as_u8,
+			  sizeof (ip6_address_t));
 	addrp--;
       }
     }
@@ -234,7 +234,8 @@ compute_rewrite_insert (ip6_address_t * sl)
   addrp = srh->segments + vec_len (sl);
   vec_foreach (this_address, sl)
   {
-    clib_memcpy (addrp->as_u8, this_address->as_u8, sizeof (ip6_address_t));
+    clib_memcpy_fast (addrp->as_u8, this_address->as_u8,
+		      sizeof (ip6_address_t));
     addrp--;
   }
   return rs;
@@ -272,7 +273,8 @@ compute_rewrite_bsid (ip6_address_t * sl)
   addrp = srh->segments + vec_len (sl) - 1;
   vec_foreach (this_address, sl)
   {
-    clib_memcpy (addrp->as_u8, this_address->as_u8, sizeof (ip6_address_t));
+    clib_memcpy_fast (addrp->as_u8, this_address->as_u8,
+		      sizeof (ip6_address_t));
     addrp--;
   }
   return rs;
@@ -589,7 +591,7 @@ sr_policy_add (ip6_address_t * bsid, ip6_address_t * segments,
   /* Add an SR policy object */
   pool_get (sm->sr_policies, sr_policy);
   clib_memset (sr_policy, 0, sizeof (*sr_policy));
-  clib_memcpy (&sr_policy->bsid, bsid, sizeof (ip6_address_t));
+  clib_memcpy_fast (&sr_policy->bsid, bsid, sizeof (ip6_address_t));
   sr_policy->type = behavior;
   sr_policy->fib_table = (fib_table != (u32) ~ 0 ? fib_table : 0);	//Is default FIB 0 ?
   sr_policy->is_encap = is_encap;
@@ -852,8 +854,8 @@ sr_policy_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	if (unformat (input, "next %U", unformat_ip6_address, &next_address))
 	{
 	  vec_add2 (segments, this_seg, 1);
-	  clib_memcpy (this_seg->as_u8, next_address.as_u8,
-		       sizeof (*this_seg));
+	  clib_memcpy_fast (this_seg->as_u8, next_address.as_u8,
+			    sizeof (*this_seg));
 	}
       else if (unformat (input, "add sl"))
 	operation = 1;
@@ -1159,14 +1161,14 @@ sr_policy_rewrite_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  ip2_encap = vlib_buffer_get_current (b2);
 	  ip3_encap = vlib_buffer_get_current (b3);
 
-	  clib_memcpy (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
-		       sl0->rewrite, vec_len (sl0->rewrite));
-	  clib_memcpy (((u8 *) ip1_encap) - vec_len (sl1->rewrite),
-		       sl1->rewrite, vec_len (sl1->rewrite));
-	  clib_memcpy (((u8 *) ip2_encap) - vec_len (sl2->rewrite),
-		       sl2->rewrite, vec_len (sl2->rewrite));
-	  clib_memcpy (((u8 *) ip3_encap) - vec_len (sl3->rewrite),
-		       sl3->rewrite, vec_len (sl3->rewrite));
+	  clib_memcpy_fast (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) ip1_encap) - vec_len (sl1->rewrite),
+			    sl1->rewrite, vec_len (sl1->rewrite));
+	  clib_memcpy_fast (((u8 *) ip2_encap) - vec_len (sl2->rewrite),
+			    sl2->rewrite, vec_len (sl2->rewrite));
+	  clib_memcpy_fast (((u8 *) ip3_encap) - vec_len (sl3->rewrite),
+			    sl3->rewrite, vec_len (sl3->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b1, -(word) vec_len (sl1->rewrite));
@@ -1189,40 +1191,40 @@ sr_policy_rewrite_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b0, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b1->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b1, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip1->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip1->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip1->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip1->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b2->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b2, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip2->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip2->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip2->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip2->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b3->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b3, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip3->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip3->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip3->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip3->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 	    }
 
@@ -1257,8 +1259,8 @@ sr_policy_rewrite_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  ip0_encap = vlib_buffer_get_current (b0);
 
-	  clib_memcpy (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
-		       sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 
 	  ip0 = vlib_buffer_get_current (b0);
@@ -1270,10 +1272,10 @@ sr_policy_rewrite_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      sr_policy_rewrite_trace_t *tr =
 		vlib_add_trace (vm, node, b0, sizeof (*tr));
-	      clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			   sizeof (tr->src.as_u8));
-	      clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			   sizeof (tr->dst.as_u8));
+	      clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				sizeof (tr->src.as_u8));
+	      clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				sizeof (tr->dst.as_u8));
 	    }
 
 	  encap_pkts++;
@@ -1442,14 +1444,14 @@ sr_policy_rewrite_encaps_v4 (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  ip2_encap = vlib_buffer_get_current (b2);
 	  ip3_encap = vlib_buffer_get_current (b3);
 
-	  clib_memcpy (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
-		       sl0->rewrite, vec_len (sl0->rewrite));
-	  clib_memcpy (((u8 *) ip1_encap) - vec_len (sl1->rewrite),
-		       sl1->rewrite, vec_len (sl1->rewrite));
-	  clib_memcpy (((u8 *) ip2_encap) - vec_len (sl2->rewrite),
-		       sl2->rewrite, vec_len (sl2->rewrite));
-	  clib_memcpy (((u8 *) ip3_encap) - vec_len (sl3->rewrite),
-		       sl3->rewrite, vec_len (sl3->rewrite));
+	  clib_memcpy_fast (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) ip1_encap) - vec_len (sl1->rewrite),
+			    sl1->rewrite, vec_len (sl1->rewrite));
+	  clib_memcpy_fast (((u8 *) ip2_encap) - vec_len (sl2->rewrite),
+			    sl2->rewrite, vec_len (sl2->rewrite));
+	  clib_memcpy_fast (((u8 *) ip3_encap) - vec_len (sl3->rewrite),
+			    sl3->rewrite, vec_len (sl3->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b1, -(word) vec_len (sl1->rewrite));
@@ -1472,40 +1474,40 @@ sr_policy_rewrite_encaps_v4 (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b0, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b1->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b1, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip1->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip1->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip1->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip1->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b2->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b2, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip2->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip2->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip2->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip2->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b3->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b3, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip3->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip3->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip3->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip3->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 	    }
 
@@ -1541,8 +1543,8 @@ sr_policy_rewrite_encaps_v4 (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  ip0_encap = vlib_buffer_get_current (b0);
 
-	  clib_memcpy (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
-		       sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 
 	  ip0 = vlib_buffer_get_current (b0);
@@ -1554,10 +1556,10 @@ sr_policy_rewrite_encaps_v4 (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      sr_policy_rewrite_trace_t *tr =
 		vlib_add_trace (vm, node, b0, sizeof (*tr));
-	      clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			   sizeof (tr->src.as_u8));
-	      clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			   sizeof (tr->dst.as_u8));
+	      clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				sizeof (tr->src.as_u8));
+	      clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				sizeof (tr->dst.as_u8));
 	    }
 
 	  encap_pkts++;
@@ -1799,14 +1801,14 @@ sr_policy_rewrite_encaps_l2 (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  en2 = vlib_buffer_get_current (b2);
 	  en3 = vlib_buffer_get_current (b3);
 
-	  clib_memcpy (((u8 *) en0) - vec_len (sl0->rewrite), sl0->rewrite,
-		       vec_len (sl0->rewrite));
-	  clib_memcpy (((u8 *) en1) - vec_len (sl1->rewrite), sl1->rewrite,
-		       vec_len (sl1->rewrite));
-	  clib_memcpy (((u8 *) en2) - vec_len (sl2->rewrite), sl2->rewrite,
-		       vec_len (sl2->rewrite));
-	  clib_memcpy (((u8 *) en3) - vec_len (sl3->rewrite), sl3->rewrite,
-		       vec_len (sl3->rewrite));
+	  clib_memcpy_fast (((u8 *) en0) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) en1) - vec_len (sl1->rewrite),
+			    sl1->rewrite, vec_len (sl1->rewrite));
+	  clib_memcpy_fast (((u8 *) en2) - vec_len (sl2->rewrite),
+			    sl2->rewrite, vec_len (sl2->rewrite));
+	  clib_memcpy_fast (((u8 *) en3) - vec_len (sl3->rewrite),
+			    sl3->rewrite, vec_len (sl3->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b1, -(word) vec_len (sl1->rewrite));
@@ -1868,40 +1870,40 @@ sr_policy_rewrite_encaps_l2 (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b0, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b1->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b1, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip1->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip1->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip1->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip1->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b2->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b2, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip2->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip2->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip2->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip2->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b3->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b3, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip3->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip3->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip3->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip3->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 	    }
 
@@ -1955,8 +1957,8 @@ sr_policy_rewrite_encaps_l2 (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  en0 = vlib_buffer_get_current (b0);
 
-	  clib_memcpy (((u8 *) en0) - vec_len (sl0->rewrite), sl0->rewrite,
-		       vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) en0) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 
@@ -1978,10 +1980,10 @@ sr_policy_rewrite_encaps_l2 (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      sr_policy_rewrite_trace_t *tr =
 		vlib_add_trace (vm, node, b0, sizeof (*tr));
-	      clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			   sizeof (tr->src.as_u8));
-	      clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			   sizeof (tr->dst.as_u8));
+	      clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				sizeof (tr->src.as_u8));
+	      clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				sizeof (tr->dst.as_u8));
 	    }
 
 	  encap_pkts++;
@@ -2145,23 +2147,23 @@ sr_policy_rewrite_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  else
 	    sr3 = (ip6_sr_header_t *) (ip3 + 1);
 
-	  clib_memcpy ((u8 *) ip0 - vec_len (sl0->rewrite), (u8 *) ip0,
-		       (void *) sr0 - (void *) ip0);
-	  clib_memcpy ((u8 *) ip1 - vec_len (sl1->rewrite), (u8 *) ip1,
-		       (void *) sr1 - (void *) ip1);
-	  clib_memcpy ((u8 *) ip2 - vec_len (sl2->rewrite), (u8 *) ip2,
-		       (void *) sr2 - (void *) ip2);
-	  clib_memcpy ((u8 *) ip3 - vec_len (sl3->rewrite), (u8 *) ip3,
-		       (void *) sr3 - (void *) ip3);
-
-	  clib_memcpy (((u8 *) sr0 - vec_len (sl0->rewrite)), sl0->rewrite,
-		       vec_len (sl0->rewrite));
-	  clib_memcpy (((u8 *) sr1 - vec_len (sl1->rewrite)), sl1->rewrite,
-		       vec_len (sl1->rewrite));
-	  clib_memcpy (((u8 *) sr2 - vec_len (sl2->rewrite)), sl2->rewrite,
-		       vec_len (sl2->rewrite));
-	  clib_memcpy (((u8 *) sr3 - vec_len (sl3->rewrite)), sl3->rewrite,
-		       vec_len (sl3->rewrite));
+	  clib_memcpy_fast ((u8 *) ip0 - vec_len (sl0->rewrite), (u8 *) ip0,
+			    (void *) sr0 - (void *) ip0);
+	  clib_memcpy_fast ((u8 *) ip1 - vec_len (sl1->rewrite), (u8 *) ip1,
+			    (void *) sr1 - (void *) ip1);
+	  clib_memcpy_fast ((u8 *) ip2 - vec_len (sl2->rewrite), (u8 *) ip2,
+			    (void *) sr2 - (void *) ip2);
+	  clib_memcpy_fast ((u8 *) ip3 - vec_len (sl3->rewrite), (u8 *) ip3,
+			    (void *) sr3 - (void *) ip3);
+
+	  clib_memcpy_fast (((u8 *) sr0 - vec_len (sl0->rewrite)),
+			    sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) sr1 - vec_len (sl1->rewrite)),
+			    sl1->rewrite, vec_len (sl1->rewrite));
+	  clib_memcpy_fast (((u8 *) sr2 - vec_len (sl2->rewrite)),
+			    sl2->rewrite, vec_len (sl2->rewrite));
+	  clib_memcpy_fast (((u8 *) sr3 - vec_len (sl3->rewrite)),
+			    sl3->rewrite, vec_len (sl3->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b1, -(word) vec_len (sl1->rewrite));
@@ -2284,40 +2286,40 @@ sr_policy_rewrite_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b0, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b1->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b1, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip1->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip1->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip1->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip1->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b2->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b2, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip2->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip2->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip2->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip2->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b3->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b3, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip3->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip3->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip3->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip3->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 	    }
 
@@ -2360,10 +2362,10 @@ sr_policy_rewrite_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  else
 	    sr0 = (ip6_sr_header_t *) (ip0 + 1);
 
-	  clib_memcpy ((u8 *) ip0 - vec_len (sl0->rewrite), (u8 *) ip0,
-		       (void *) sr0 - (void *) ip0);
-	  clib_memcpy (((u8 *) sr0 - vec_len (sl0->rewrite)), sl0->rewrite,
-		       vec_len (sl0->rewrite));
+	  clib_memcpy_fast ((u8 *) ip0 - vec_len (sl0->rewrite), (u8 *) ip0,
+			    (void *) sr0 - (void *) ip0);
+	  clib_memcpy_fast (((u8 *) sr0 - vec_len (sl0->rewrite)),
+			    sl0->rewrite, vec_len (sl0->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 
@@ -2400,10 +2402,10 @@ sr_policy_rewrite_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      sr_policy_rewrite_trace_t *tr =
 		vlib_add_trace (vm, node, b0, sizeof (*tr));
-	      clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			   sizeof (tr->src.as_u8));
-	      clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			   sizeof (tr->dst.as_u8));
+	      clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				sizeof (tr->src.as_u8));
+	      clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				sizeof (tr->dst.as_u8));
 	    }
 
 	  insert_pkts++;
@@ -2567,23 +2569,23 @@ sr_policy_rewrite_b_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  else
 	    sr3 = (ip6_sr_header_t *) (ip3 + 1);
 
-	  clib_memcpy ((u8 *) ip0 - vec_len (sl0->rewrite_bsid), (u8 *) ip0,
-		       (void *) sr0 - (void *) ip0);
-	  clib_memcpy ((u8 *) ip1 - vec_len (sl1->rewrite_bsid), (u8 *) ip1,
-		       (void *) sr1 - (void *) ip1);
-	  clib_memcpy ((u8 *) ip2 - vec_len (sl2->rewrite_bsid), (u8 *) ip2,
-		       (void *) sr2 - (void *) ip2);
-	  clib_memcpy ((u8 *) ip3 - vec_len (sl3->rewrite_bsid), (u8 *) ip3,
-		       (void *) sr3 - (void *) ip3);
-
-	  clib_memcpy (((u8 *) sr0 - vec_len (sl0->rewrite_bsid)),
-		       sl0->rewrite_bsid, vec_len (sl0->rewrite_bsid));
-	  clib_memcpy (((u8 *) sr1 - vec_len (sl1->rewrite_bsid)),
-		       sl1->rewrite_bsid, vec_len (sl1->rewrite_bsid));
-	  clib_memcpy (((u8 *) sr2 - vec_len (sl2->rewrite_bsid)),
-		       sl2->rewrite_bsid, vec_len (sl2->rewrite_bsid));
-	  clib_memcpy (((u8 *) sr3 - vec_len (sl3->rewrite_bsid)),
-		       sl3->rewrite_bsid, vec_len (sl3->rewrite_bsid));
+	  clib_memcpy_fast ((u8 *) ip0 - vec_len (sl0->rewrite_bsid),
+			    (u8 *) ip0, (void *) sr0 - (void *) ip0);
+	  clib_memcpy_fast ((u8 *) ip1 - vec_len (sl1->rewrite_bsid),
+			    (u8 *) ip1, (void *) sr1 - (void *) ip1);
+	  clib_memcpy_fast ((u8 *) ip2 - vec_len (sl2->rewrite_bsid),
+			    (u8 *) ip2, (void *) sr2 - (void *) ip2);
+	  clib_memcpy_fast ((u8 *) ip3 - vec_len (sl3->rewrite_bsid),
+			    (u8 *) ip3, (void *) sr3 - (void *) ip3);
+
+	  clib_memcpy_fast (((u8 *) sr0 - vec_len (sl0->rewrite_bsid)),
+			    sl0->rewrite_bsid, vec_len (sl0->rewrite_bsid));
+	  clib_memcpy_fast (((u8 *) sr1 - vec_len (sl1->rewrite_bsid)),
+			    sl1->rewrite_bsid, vec_len (sl1->rewrite_bsid));
+	  clib_memcpy_fast (((u8 *) sr2 - vec_len (sl2->rewrite_bsid)),
+			    sl2->rewrite_bsid, vec_len (sl2->rewrite_bsid));
+	  clib_memcpy_fast (((u8 *) sr3 - vec_len (sl3->rewrite_bsid)),
+			    sl3->rewrite_bsid, vec_len (sl3->rewrite_bsid));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite_bsid));
 	  vlib_buffer_advance (b1, -(word) vec_len (sl1->rewrite_bsid));
@@ -2697,40 +2699,40 @@ sr_policy_rewrite_b_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b0, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b1->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b1, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip1->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip1->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip1->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip1->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b2->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b2, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip2->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip2->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip2->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip2->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b3->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b3, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip3->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip3->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip3->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip3->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 	    }
 
@@ -2773,10 +2775,10 @@ sr_policy_rewrite_b_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  else
 	    sr0 = (ip6_sr_header_t *) (ip0 + 1);
 
-	  clib_memcpy ((u8 *) ip0 - vec_len (sl0->rewrite_bsid), (u8 *) ip0,
-		       (void *) sr0 - (void *) ip0);
-	  clib_memcpy (((u8 *) sr0 - vec_len (sl0->rewrite_bsid)),
-		       sl0->rewrite_bsid, vec_len (sl0->rewrite_bsid));
+	  clib_memcpy_fast ((u8 *) ip0 - vec_len (sl0->rewrite_bsid),
+			    (u8 *) ip0, (void *) sr0 - (void *) ip0);
+	  clib_memcpy_fast (((u8 *) sr0 - vec_len (sl0->rewrite_bsid)),
+			    sl0->rewrite_bsid, vec_len (sl0->rewrite_bsid));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite_bsid));
 
@@ -2811,10 +2813,10 @@ sr_policy_rewrite_b_insert (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      sr_policy_rewrite_trace_t *tr =
 		vlib_add_trace (vm, node, b0, sizeof (*tr));
-	      clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			   sizeof (tr->src.as_u8));
-	      clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			   sizeof (tr->dst.as_u8));
+	      clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				sizeof (tr->src.as_u8));
+	      clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				sizeof (tr->dst.as_u8));
 	    }
 
 	  insert_pkts++;
@@ -2997,14 +2999,14 @@ sr_policy_rewrite_b_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  end_bsid_encaps_srh_processing (node, b2, ip2_encap, sr2, &next2);
 	  end_bsid_encaps_srh_processing (node, b3, ip3_encap, sr3, &next3);
 
-	  clib_memcpy (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
-		       sl0->rewrite, vec_len (sl0->rewrite));
-	  clib_memcpy (((u8 *) ip1_encap) - vec_len (sl1->rewrite),
-		       sl1->rewrite, vec_len (sl1->rewrite));
-	  clib_memcpy (((u8 *) ip2_encap) - vec_len (sl2->rewrite),
-		       sl2->rewrite, vec_len (sl2->rewrite));
-	  clib_memcpy (((u8 *) ip3_encap) - vec_len (sl3->rewrite),
-		       sl3->rewrite, vec_len (sl3->rewrite));
+	  clib_memcpy_fast (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) ip1_encap) - vec_len (sl1->rewrite),
+			    sl1->rewrite, vec_len (sl1->rewrite));
+	  clib_memcpy_fast (((u8 *) ip2_encap) - vec_len (sl2->rewrite),
+			    sl2->rewrite, vec_len (sl2->rewrite));
+	  clib_memcpy_fast (((u8 *) ip3_encap) - vec_len (sl3->rewrite),
+			    sl3->rewrite, vec_len (sl3->rewrite));
 
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b1, -(word) vec_len (sl1->rewrite));
@@ -3027,40 +3029,40 @@ sr_policy_rewrite_b_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b0, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b1->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b1, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip1->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip1->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip1->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip1->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b2->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b2, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip2->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip2->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip2->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip2->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 
 	      if (PREDICT_FALSE (b3->flags & VLIB_BUFFER_IS_TRACED))
 		{
 		  sr_policy_rewrite_trace_t *tr =
 		    vlib_add_trace (vm, node, b3, sizeof (*tr));
-		  clib_memcpy (tr->src.as_u8, ip3->src_address.as_u8,
-			       sizeof (tr->src.as_u8));
-		  clib_memcpy (tr->dst.as_u8, ip3->dst_address.as_u8,
-			       sizeof (tr->dst.as_u8));
+		  clib_memcpy_fast (tr->src.as_u8, ip3->src_address.as_u8,
+				    sizeof (tr->src.as_u8));
+		  clib_memcpy_fast (tr->dst.as_u8, ip3->dst_address.as_u8,
+				    sizeof (tr->dst.as_u8));
 		}
 	    }
 
@@ -3100,8 +3102,8 @@ sr_policy_rewrite_b_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 				 IP_PROTOCOL_IPV6_ROUTE);
 	  end_bsid_encaps_srh_processing (node, b0, ip0_encap, sr0, &next0);
 
-	  clib_memcpy (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
-		       sl0->rewrite, vec_len (sl0->rewrite));
+	  clib_memcpy_fast (((u8 *) ip0_encap) - vec_len (sl0->rewrite),
+			    sl0->rewrite, vec_len (sl0->rewrite));
 	  vlib_buffer_advance (b0, -(word) vec_len (sl0->rewrite));
 
 	  ip0 = vlib_buffer_get_current (b0);
@@ -3113,10 +3115,10 @@ sr_policy_rewrite_b_encaps (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      sr_policy_rewrite_trace_t *tr =
 		vlib_add_trace (vm, node, b0, sizeof (*tr));
-	      clib_memcpy (tr->src.as_u8, ip0->src_address.as_u8,
-			   sizeof (tr->src.as_u8));
-	      clib_memcpy (tr->dst.as_u8, ip0->dst_address.as_u8,
-			   sizeof (tr->dst.as_u8));
+	      clib_memcpy_fast (tr->src.as_u8, ip0->src_address.as_u8,
+				sizeof (tr->src.as_u8));
+	      clib_memcpy_fast (tr->dst.as_u8, ip0->dst_address.as_u8,
+				sizeof (tr->dst.as_u8));
 	    }
 
 	  encap_pkts++;