From 1ea79e8626740dd43b9ab1cbf389b86a1e5f9e72 Mon Sep 17 00:00:00 2001
From: Bill Currie <bill@taniwha.org>
Date: Sun, 17 Mar 2013 16:37:27 +0900
Subject: [PATCH] Allocate stack blocks and mightsee in one block.

This bypasses set_new, but completely removes the use of the global lock
from within RecursiveClusterFlow. This seems to give a small speedup: 203
seconds threaded.
---
 tools/qfvis/include/vis.h |  4 ++--
 tools/qfvis/source/flow.c | 42 ++++++++++++++++++++-------------------
 2 files changed, 24 insertions(+), 22 deletions(-)

diff --git a/tools/qfvis/include/vis.h b/tools/qfvis/include/vis.h
index b4a29ab4f..8c5c1b22d 100644
--- a/tools/qfvis/include/vis.h
+++ b/tools/qfvis/include/vis.h
@@ -102,8 +102,8 @@ typedef struct pstack_s {
 	portal_t   *portal;		// portal exiting
 	winding_t  *source, *pass;
 	plane_t     portalplane;
-	set_t      *mightsee;
 	sep_t      *separators[2];
+	set_t       mightsee;
 } pstack_t;
 
 typedef struct {
@@ -126,7 +126,7 @@ typedef struct threaddata_s {
 	visstat_t   stats;			///< per-thread statistics merged on completion
 	set_t      *clustervis;		///< clusters base portal can see
 	portal_t   *base;			///< portal for which this thread is being run
-	pstack_t    pstack_head;
+	pstack_t   *pstack_head;
 	sep_t      *sep_freelist;	///< per-thread list of free separators
 } threaddata_t;
 
diff --git a/tools/qfvis/source/flow.c b/tools/qfvis/source/flow.c
index 732351aa6..ae94c9b1d 100644
--- a/tools/qfvis/source/flow.c
+++ b/tools/qfvis/source/flow.c
@@ -63,7 +63,7 @@ CheckStack (cluster_t *cluster, threaddata_t *thread)
 {
 	pstack_t   *portal;
 
-	for (portal = thread->pstack_head.next; portal; portal = portal->next) {
+	for (portal = thread->pstack_head->next; portal; portal = portal->next) {
 		if (portal->cluster == cluster) {
 			printf ("CheckStack: cluster recursion\n");
 			return 1;
@@ -78,12 +78,12 @@ static pstack_t *
 new_stack (void)
 {
 	pstack_t   *stack;
+	unsigned    size = SET_SIZE (portalclusters - 1);
 
-	stack = malloc (sizeof (pstack_t));
-	stack->next = 0;
-	LOCK;
-	stack->mightsee = set_new_size (portalclusters);
-	UNLOCK;
+	stack = calloc (1, field_offset (pstack_t,
+									 mightsee.defmap[size / SET_BITS]));
+	stack->mightsee.size = size;
+	stack->mightsee.map = stack->mightsee.defmap;
 	return stack;
 }
 
@@ -301,19 +301,19 @@ RecursiveClusterFlow (int clusternum, threaddata_t *thread, pstack_t *prevstack)
 	stack->portal = NULL;
 	stack->separators[0] = 0;
 	stack->separators[1] = 0;
-	might = stack->mightsee;
+	might = &stack->mightsee;
 	vis = thread->clustervis;
 
 	// check all portals for flowing into other clusters
 	for (i = 0; i < cluster->numportals; i++) {
 		portal = cluster->portals[i];
 
-		if (!set_is_member (prevstack->mightsee, portal->cluster))
+		if (!set_is_member (&prevstack->mightsee, portal->cluster))
 			continue;		// can't possibly see it
 
 		// if the portal can't see anything we haven't already seen, skip it
 		test = select_test_set (portal, thread);
-		if (!mightsee_more (might, prevstack->mightsee, test, vis)) {
+		if (!mightsee_more (might, &prevstack->mightsee, test, vis)) {
 			// can't see anything new
 			continue;
 		}
@@ -331,7 +331,7 @@ RecursiveClusterFlow (int clusternum, threaddata_t *thread, pstack_t *prevstack)
 		stack->portal = portal;
 
 		target = ClipWinding (portal->winding,
-							  &thread->pstack_head.portalplane, false);
+							  &thread->pstack_head->portalplane, false);
 		if (!target)
 			continue;
 
@@ -367,7 +367,7 @@ RecursiveClusterFlow (int clusternum, threaddata_t *thread, pstack_t *prevstack)
 			winding_t  *old = target;
 			if (!stack->separators[0])
 				stack->separators[0] = FindSeparators (thread, source,
-											  thread->pstack_head.portalplane,
+											  thread->pstack_head->portalplane,
 													  prevstack->pass, 0);
 			target = ClipToSeparators (stack->separators[0], target);
 			if (!target) {
@@ -464,14 +464,16 @@ PortalFlow (threaddata_t *data, portal_t *portal)
 	data->clustervis = portal->visbits;
 	data->base = portal;
 
-	data->pstack_head.cluster = 0;
-	data->pstack_head.portal = portal;
-	data->pstack_head.source = portal->winding;
-	data->pstack_head.pass = 0;
-	data->pstack_head.portalplane = portal->plane;
-	data->pstack_head.mightsee = portal->mightsee;
-	data->pstack_head.separators[0] = 0;
-	data->pstack_head.separators[1] = 0;
+	if (!data->pstack_head)
+		data->pstack_head = new_stack ();
+	data->pstack_head->cluster = 0;
+	data->pstack_head->portal = portal;
+	data->pstack_head->source = portal->winding;
+	data->pstack_head->pass = 0;
+	data->pstack_head->portalplane = portal->plane;
+	set_assign (&data->pstack_head->mightsee, portal->mightsee);
+	data->pstack_head->separators[0] = 0;
+	data->pstack_head->separators[1] = 0;
 
-	RecursiveClusterFlow (portal->cluster, data, &data->pstack_head);
+	RecursiveClusterFlow (portal->cluster, data, data->pstack_head);
 }