o3: use new_k (size-class free lists) like plain code

Constant-size typed new inside o3 now calls rt.new_k instead of always
taking new_t/new_raw. new_k and recycle_leaf skip the o3 argument-fatten
path the way free/recycle already do. trees o3 new/free on the x86-64
jit dropped from about 4.1 ms to about 2.7 ms per iter (matches plain).

1 files changed, 4 insertions, 4 deletions

diff --git a/src/compiler/expr.zc b/src/compiler/expr.zc
index 1cd765b..9d7cc2b 100644
--- a/src/compiler/expr.zc
+++ b/src/compiler/expr.zc
@@ -144,7 +144,7 @@ int stamp_site() {
 // new(T) and new(T, count): a typed object, its type stamped into the header
 // new(T), new(T, n), new(T[?]), new(fixed T[?], n). a growable array goes to
 // dnew. otherwise the call is new_t(size, type), unless the size is known at
-// compile time and we are not in o3 or a dry run: then the size class is
+// compile time and we are not in a dry run: then the size class is
 // computed here and the call is new_k(size, type, class), which pops a block
 // straight off the class list. newsz and newleaf record what was allocated so
 // the variable it lands in can later be freed with the matching fast path.
@@ -176,7 +176,7 @@ int newtyped() {
 		lit(tsz(t)); lit(t);
 		newsz = tsz(t); newleaf = 0;
 		if (isleaf(t)) newleaf = t;
-		if (newsz <= 0x7FFFFFFF - rt.HDR && !o3on && !dry && lookup(s_new_k) >= 0) { lit(szclass(newsz)); fast = 1; }
+		if (newsz <= 0x7FFFFFFF - rt.HDR && !dry && lookup(s_new_k) >= 0) { lit(szclass(newsz)); fast = 1; }
 	}
 	expect(s_rp);
 	stamp_site();
@@ -439,7 +439,7 @@ int call(int s) {
 	rt.fill_w(bs, 0, 8);
 	if (ngbind) { rt.copy_w(bs, gbind, ngbind); ngbind = 0; }
 	np = 0;
-	if (o3on && !(sflag[s] & F_O3) && sname[s] != s_free && sname[s] != s_recycle) np = 1;	// o3 calling normal code: fatten the arguments
+	if (o3on && !(sflag[s] & F_O3) && sname[s] != s_free && sname[s] != s_recycle && sname[s] != s_new_k && sname[s] != s_recycle_leaf) np = 1;	// o3 calling normal code: fatten the arguments
 	ifcall = s; effect_argument_base = 0;
 	np = args(tags, ft, bs, np);
 	storage = argfact; origin = argorigin; origins = argorigin_mask; temporary = argtemporary; owner = argowner; owner_fact = argowner_fact;
@@ -456,7 +456,7 @@ int call(int s) {
 		xaf = af_new(newsz, 2, cp);
 		if (lsp > 0) newloop(cp);
 		// a constant size: the class is known now, so use the allocator's direct path
-		if (newsz >= 0 && newsz <= 0x7FFFFFFF - rt.HDR && !o3on && !dry && lookup(s_new_k) >= 0) {	// (a bigger size has no class: new_t stops it)
+		if (newsz >= 0 && newsz <= 0x7FFFFFFF - rt.HDR && !dry && lookup(s_new_k) >= 0) {	// (a bigger size has no class: new_t stops it)
 			lit(0);
 			lit(szclass(newsz));
 			s = lookup(s_new_k);