cmd/gc: do not copy via temporary for writebarrierfat{2,3,4}

The general writebarrierfat needs a temporary for src,
because we need to pass the address of the temporary
to the writebarrierfat routine. But the new fixed-size
ones pass the value directly and don't need to introduce
the temporary.

Magnifies some of the effect of the custom write barrier change.

Comparing best of 5 with TurboBoost turned off,
on a 2012 Retina MacBook Pro Core i5.
Still not completely confident in these numbers,
but the fmt, regexp, and revcomp improvements seem real.

benchmark                      old ns/op  new ns/op  delta
BenchmarkBinaryTree17          3942965521 3929654940 -0.34%
BenchmarkFannkuch11            3707543350 3699566011 -0.22%
BenchmarkFmtFprintfEmpty       119        119        +0.00%
BenchmarkFmtFprintfString      295        296        +0.34%
BenchmarkFmtFprintfInt         313        314        +0.32%
BenchmarkFmtFprintfIntInt      517        484        -6.38%
BenchmarkFmtFprintfPrefixedInt 439        429        -2.28%
BenchmarkFmtFprintfFloat       571        569        -0.35%
BenchmarkFmtManyArgs           1899       1820       -4.16%
BenchmarkGobDecode             15507208   15325649   -1.17%
BenchmarkGobEncode             14811710   14715434   -0.65%
BenchmarkGzip                  561144467  549624323  -2.05%
BenchmarkGunzip                137377667  137691087  +0.23%
BenchmarkHTTPClientServer      126632     124717     -1.51%
BenchmarkJSONEncode            29944112   29526629   -1.39%
BenchmarkJSONDecode            108954913  107339551  -1.48%
BenchmarkMandelbrot200         5828755    5821659    -0.12%
BenchmarkGoParse               5577437    5521895    -1.00%
BenchmarkRegexpMatchEasy0_32   198        193        -2.53%
BenchmarkRegexpMatchEasy0_1K   486        469        -3.50%
BenchmarkRegexpMatchEasy1_32   175        167        -4.57%
BenchmarkRegexpMatchEasy1_1K   1450       1419       -2.14%
BenchmarkRegexpMatchMedium_32  344        338        -1.74%
BenchmarkRegexpMatchMedium_1K  112088     109855     -1.99%
BenchmarkRegexpMatchHard_32    6078       6003       -1.23%
BenchmarkRegexpMatchHard_1K    191166     187499     -1.92%
BenchmarkRevcomp               854870445  799012851  -6.53%
BenchmarkTemplate              141572691  141508105  -0.05%
BenchmarkTimeParse             604        603        -0.17%
BenchmarkTimeFormat            579        560        -3.28%

LGTM=r
R=r
CC=golang-codereviews
https://golang.org/cl/155450043

src/cmd/gc/order.c

@@ -438,6 +438,9 @@ ordercall(Node *n, Order *order)
 // cases they are also typically registerizable, so not much harm done.
 // And this only applies to the multiple-assignment form.
 // We could do a more precise analysis if needed, like in walk.c.
+//
+// Ordermapassign also inserts these temporaries if needed for
+// calling writebarrierfat with a pointer to n->right.
 static void
 ordermapassign(Node *n, Order *order)
 {
@@ -451,7 +454,8 @@ ordermapassign(Node *n, Order *order)
 
 	case OAS:
 		order->out = list(order->out, n);
-		if((n->left->op == OINDEXMAP || (needwritebarrier(n->left, n->right) && n->left->type->width > widthptr)) && !isaddrokay(n->right)) {
+		// We call writebarrierfat only for values > 4 pointers long. See walk.c.
+		if((n->left->op == OINDEXMAP || (needwritebarrier(n->left, n->right) && n->left->type->width > 4*widthptr)) && !isaddrokay(n->right)) {
 			m = n->left;
 			n->left = ordertemp(m->type, order, 0);
 			a = nod(OAS, m, n->left);