diff --git a/src/cmd/6g/ggen.c b/src/cmd/6g/ggen.c
index 987473ccab60908acd7d4cbea6fe74c8bbd79c90..363620769d911d414c57d85ab6055e4b16c68a97 100644
--- a/src/cmd/6g/ggen.c
+++ b/src/cmd/6g/ggen.c
@@ -1102,26 +1102,54 @@ clearfat(Node *nl)
c = w % 8; // bytes
q = w / 8; // quads
+ if(q < 4) {
+ // Write sequence of MOV 0, off(base) instead of using STOSQ.
+ // The hope is that although the code will be slightly longer,
+ // the MOVs will have no dependencies and pipeline better
+ // than the unrolled STOSQ loop.
+ // NOTE: Must use agen, not igen, so that optimizer sees address
+ // being taken. We are not writing on field boundaries.
+ agenr(nl, &n1, N);
+ n1.op = OINDREG;
+ nodconst(&z, types[TUINT64], 0);
+ while(q-- > 0) {
+ n1.type = z.type;
+ gins(AMOVQ, &z, &n1);
+ n1.xoffset += 8;
+ }
+ if(c >= 4) {
+ nodconst(&z, types[TUINT32], 0);
+ n1.type = z.type;
+ gins(AMOVL, &z, &n1);
+ n1.xoffset += 4;
+ c -= 4;
+ }
+ nodconst(&z, types[TUINT8], 0);
+ while(c-- > 0) {
+ n1.type = z.type;
+ gins(AMOVB, &z, &n1);
+ n1.xoffset++;
+ }
+ regfree(&n1);
+ return;
+ }
+
savex(D_DI, &n1, &oldn1, N, types[tptr]);
agen(nl, &n1);
savex(D_AX, &ax, &oldax, N, types[tptr]);
gconreg(AMOVL, 0, D_AX);
- if(q > 128 || (q >= 4 && nacl)) {
+ if(q > 128 || nacl) {
gconreg(movptr, q, D_CX);
gins(AREP, N, N); // repeat
gins(ASTOSQ, N, N); // STOQ AL,*(DI)+
- } else if(q >= 4) {
+ } else {
p = gins(ADUFFZERO, N, N);
p->to.type = D_ADDR;
p->to.sym = linksym(pkglookup("duffzero", runtimepkg));
// 2 and 128 = magic constants: see ../../runtime/asm_amd64.s
p->to.offset = 2*(128-q);
- } else
- while(q > 0) {
- gins(ASTOSQ, N, N); // STOQ AL,*(DI)+
- q--;
}
z = ax;
diff --git a/src/cmd/8g/ggen.c b/src/cmd/8g/ggen.c
index 7c986cc6457b4c7d207800fefcd1f6ba9e1e5790..6333a60bb8a7d433dee97f4bfdc606933924fce6 100644
--- a/src/cmd/8g/ggen.c
+++ b/src/cmd/8g/ggen.c
@@ -157,7 +157,7 @@ void
clearfat(Node *nl)
{
uint32 w, c, q;
- Node n1;
+ Node n1, z;
Prog *p;
/* clear a fat object */
@@ -172,6 +172,32 @@ clearfat(Node *nl)
c = w % 4; // bytes
q = w / 4; // quads
+ if(q < 4) {
+ // Write sequence of MOV 0, off(base) instead of using STOSL.
+ // The hope is that although the code will be slightly longer,
+ // the MOVs will have no dependencies and pipeline better
+ // than the unrolled STOSL loop.
+ // NOTE: Must use agen, not igen, so that optimizer sees address
+ // being taken. We are not writing on field boundaries.
+ regalloc(&n1, types[tptr], N);
+ agen(nl, &n1);
+ n1.op = OINDREG;
+ nodconst(&z, types[TUINT64], 0);
+ while(q-- > 0) {
+ n1.type = z.type;
+ gins(AMOVL, &z, &n1);
+ n1.xoffset += 4;
+ }
+ nodconst(&z, types[TUINT8], 0);
+ while(c-- > 0) {
+ n1.type = z.type;
+ gins(AMOVB, &z, &n1);
+ n1.xoffset++;
+ }
+ regfree(&n1);
+ return;
+ }
+
nodreg(&n1, types[tptr], D_DI);
agen(nl, &n1);
gconreg(AMOVL, 0, D_AX);
diff --git a/src/cmd/gc/gen.c b/src/cmd/gc/gen.c
index eb9eacca8f133d39c73ff2553d32b3dc27fc6b9f..a7db833a1a32a8a2bac91031ea41636e88190f8f 100644
--- a/src/cmd/gc/gen.c
+++ b/src/cmd/gc/gen.c
@@ -731,7 +731,7 @@ cgen_as(Node *nl, Node *nr)
return;
}
- if(nr == N || isnil(nr)) {
+ if(nr == N || iszero(nr)) {
// externals and heaps should already be clear
if(nr == N) {
if(nl->class == PEXTERN)
diff --git a/src/cmd/gc/go.h b/src/cmd/gc/go.h
index 8178f7272f5427f7409e912f1f0c613fb0baf06d..475754145b18216ff990f7ebde0e1cd80d163ae7 100644
--- a/src/cmd/gc/go.h
+++ b/src/cmd/gc/go.h
@@ -1374,6 +1374,7 @@ int isnilinter(Type *t);
int isptrto(Type *t, int et);
int isslice(Type *t);
int istype(Type *t, int et);
+int iszero(Node *n);
void linehist(char *file, int32 off, int relative);
NodeList* list(NodeList *l, Node *n);
NodeList* list1(Node *n);
diff --git a/src/cmd/gc/mparith2.c b/src/cmd/gc/mparith2.c
index 5cf98c62c634cdfd91ae3328a69f57c662b3452c..fd9f591ceae2ac164b48b6be4ffdad8a04af04b4 100644
--- a/src/cmd/gc/mparith2.c
+++ b/src/cmd/gc/mparith2.c
@@ -656,7 +656,7 @@ mpdivmodfixfix(Mpint *q, Mpint *r, Mpint *n, Mpint *d)
}
static int
-iszero(Mpint *a)
+mpiszero(Mpint *a)
{
long *a1;
int i;
@@ -687,7 +687,7 @@ mpdivfract(Mpint *a, Mpint *b)
for(j=0; j<Mpscale; j++) {
x <<= 1;
if(mpcmp(&d, &n) <= 0) {
- if(!iszero(&d))
+ if(!mpiszero(&d))
x |= 1;
mpsubfixfix(&n, &d);
}
diff --git a/src/cmd/gc/sinit.c b/src/cmd/gc/sinit.c
index f050026d9d07d570e71b909657f698b0b1a9d64e..2a811513c9b3e257f43e81b13f2d65fe71214e5f 100644
--- a/src/cmd/gc/sinit.c
+++ b/src/cmd/gc/sinit.c
@@ -17,7 +17,6 @@ enum
InitPending = 2,
};
-static int iszero(Node*);
static void initplan(Node*);
static NodeList *initlist;
static void init2(Node*, NodeList**);
@@ -1356,7 +1355,6 @@ no:
return 0;
}
-static int iszero(Node*);
static int isvaluelit(Node*);
static InitEntry* entry(InitPlan*);
static void addvalue(InitPlan*, vlong, Node*, Node*);
@@ -1440,7 +1438,7 @@ addvalue(InitPlan *p, vlong xoffset, Node *key, Node *n)
e->expr = n;
}
-static int
+int
iszero(Node *n)
{
NodeList *l;
diff --git a/src/cmd/gc/walk.c b/src/cmd/gc/walk.c
index 241d7d74adbc9029c12a02c12de63e0c0fbbd628..7f2748c668d05f0e428a1f2d4bafc0fdf25e436e 100644
--- a/src/cmd/gc/walk.c
+++ b/src/cmd/gc/walk.c
@@ -1390,7 +1390,12 @@ walkexpr(Node **np, NodeList **init)
case OMAPLIT:
case OSTRUCTLIT:
case OPTRLIT:
- // XXX TODO do we need to clear var?
+ // NOTE(rsc): Race detector cannot handle seeing
+ // a STRUCTLIT or ARRAYLIT representing a zero value,
+ // so make a temporary for those always in race mode.
+ // Otherwise, leave zero values in place.
+ if(iszero(n) && !flag_race)
+ goto ret;
var = temp(n->type);
anylit(0, n, var, init);
n = var;
@@ -2009,8 +2014,8 @@ needwritebarrier(Node *l, Node *r)
if(isstack(l))
return 0;
- // No write barrier for zeroing.
- if(r == N)
+ // No write barrier for implicit or explicit zeroing.
+ if(r == N || iszero(r))
return 0;
// No write barrier for initialization to constant.