code review 81370046: runtime: get rid of most uses of REP for copying/zeroing.

src/cmd/6g/cgen.c

 // Copyright 2009 The Go Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
 #include <u.h>
 #include <libc.h>
 #include "gg.h"
 
 /*
  * generate:
@@ skipping 1327 matching lines @@
 /*
  * block copy:
  *      memmove(&ns, &n, w);
  */
 void
 sgen(Node *n, Node *ns, int64 w)
 {
         Node nodl, nodr, nodsi, noddi, cx, oldcx, tmp;
         vlong c, q, odst, osrc;
         NodeList *l;
+        Prog *p;
 
         if(debug['g']) {
                 print("\nsgen w=%lld\n", w);
                 dump("r", n);
                 dump("res", ns);
         }
 
         if(n->ullman >= UINF && ns->ullman >= UINF)
                 fatal("sgen UINF");
 
@@ skipping 82 matching lines @@
                                 gconreg(addptr, w-8, D_DI);
                         }
                         gconreg(movptr, q, D_CX);
                         gins(AREP, N, N);       // repeat
                         gins(AMOVSQ, N, N);     // MOVQ *(SI)-,*(DI)-
                 }
                 // we leave with the flag clear
                 gins(ACLD, N, N);
         } else {
                 // normal direction
-                if(q >= 4) {
+                if(q > 128) {
                         gconreg(movptr, q, D_CX);
                         gins(AREP, N, N);       // repeat
                         gins(AMOVSQ, N, N);     // MOVQ *(SI)+,*(DI)+
+                } else if (q >= 4) {
+                        p = gins(ADUFFCOPY, N, N);
+                        p->to.type = D_ADDR;
+                        p->to.sym = linksym(pkglookup("duffcopy", runtimepkg));
+                        // 14 = magic constant: see ../../pkg/runtime/asm_amd64.s

[rsc, 2014/04/01 16:55:05]

128 is a magic constant too. :-)

+                        p->to.offset = 14*(128-q);
                 } else
                 while(q > 0) {
                         gins(AMOVSQ, N, N);     // MOVQ *(SI)+,*(DI)+
                         q--;
                 }
                 // copy the remaining c bytes
                 if(w < 4 || c <= 1 || (odst < osrc && osrc < odst+w)) {
                         while(c > 0) {
                                 gins(AMOVSB, N, N);     // MOVB *(SI)+,*(DI)+
                                 c--;
@@ skipping 250 matching lines @@
                 regfree(&nodl);
         return 0;
 
 yes:
         if(freer)
                 regfree(&nodr);
         if(freel)
                 regfree(&nodl);
         return 1;
 }