sna/trapezoids: Only use a single thread to emit rectilinear spans

If the trapezoids are rectilinear, they should hit a fast path through
the span compositors and so threading them seems pointless. Expect
possibily for inplace pixman operations.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

5 files changed, 113 insertions, 69 deletions

diff --git a/src/sna/gen4_vertex.c b/src/sna/gen4_vertex.c
index 85e74137..e2a48083 100644
--- a/src/sna/gen4_vertex.c
+++ b/src/sna/gen4_vertex.c
@@ -45,7 +45,9 @@ void gen4_vertex_flush(struct sna *sna)
 	     sna->render.vertex_index - sna->render.vertex_start));
 
 	assert(sna->render.vertex_offset);
+	assert(sna->render.vertex_offset <= sna->kgem.nbatch);
 	assert(sna->render.vertex_index > sna->render.vertex_start);
+	assert(sna->render.vertex_used <= sna->render.vertex_size);
 
 	sna->kgem.batch[sna->render.vertex_offset] =
 		sna->render.vertex_index - sna->render.vertex_start;
@@ -62,6 +64,7 @@ int gen4_vertex_finish(struct sna *sna)
 	     sna->render.vertex_used, sna->render.vertex_size));
 	assert(sna->render.vertex_offset == 0);
 	assert(sna->render.vertex_used);
+	assert(sna->render.vertex_used <= sna->render.vertex_size);
 
 	sna_vertex_wait__locked(&sna->render);
 
diff --git a/src/sna/sna_trapezoids.c b/src/sna/sna_trapezoids.c
index 7c02840d..c9c031a4 100644
--- a/src/sna/sna_trapezoids.c
+++ b/src/sna/sna_trapezoids.c
@@ -660,7 +660,7 @@ sna_composite_trapezoids(CARD8 op,
 
 	if (trapezoid_spans_maybe_inplace(sna, op, src, dst, maskFormat)) {
 		flags |= COMPOSITE_SPANS_INPLACE_HINT;
-		if (trapezoid_span_inplace(sna, op, src, dst, maskFormat,
+		if (trapezoid_span_inplace(sna, op, src, dst, maskFormat, flags,
 					   xSrc, ySrc, ntrap, traps,
 					   false))
 			return;
@@ -670,22 +670,22 @@ sna_composite_trapezoids(CARD8 op,
 				     xSrc, ySrc, ntrap, traps))
 		return;
 
-	if (trapezoid_span_inplace(sna, op, src, dst, maskFormat,
+	if (trapezoid_span_inplace(sna, op, src, dst, maskFormat, flags,
 				   xSrc, ySrc, ntrap, traps,
 				   false))
 		return;
 
-	if (trapezoid_mask_converter(op, src, dst, maskFormat,
+	if (trapezoid_mask_converter(op, src, dst, maskFormat, flags,
 				     xSrc, ySrc, ntrap, traps))
 		return;
 
 fallback:
-	if (trapezoid_span_inplace(sna, op, src, dst, maskFormat,
+	if (trapezoid_span_inplace(sna, op, src, dst, maskFormat, flags,
 				   xSrc, ySrc, ntrap, traps,
 				   true))
 		return;
 
-	if (trapezoid_span_fallback(op, src, dst, maskFormat,
+	if (trapezoid_span_fallback(op, src, dst, maskFormat, flags,
 				    xSrc, ySrc, ntrap, traps))
 		return;
 
diff --git a/src/sna/sna_trapezoids.h b/src/sna/sna_trapezoids.h
index 729334cb..d1017268 100644
--- a/src/sna/sna_trapezoids.h
+++ b/src/sna/sna_trapezoids.h
@@ -62,7 +62,8 @@ mono_triangles_span_converter(struct sna *sna,
 bool
 imprecise_trapezoid_span_inplace(struct sna *sna,
 				 CARD8 op, PicturePtr src, PicturePtr dst,
-				 PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				 PictFormatPtr maskFormat, unsigned flags,
+				 INT16 src_x, INT16 src_y,
 				 int ntrap, xTrapezoid *traps,
 				 bool fallback);
 
@@ -75,18 +76,21 @@ imprecise_trapezoid_span_converter(struct sna *sna,
 
 bool
 imprecise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
-				   PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				   PictFormatPtr maskFormat, unsigned flags,
+				   INT16 src_x, INT16 src_y,
 				   int ntrap, xTrapezoid *traps);
 
 bool
 imprecise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
-				  PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				  PictFormatPtr maskFormat, unsigned flags,
+				  INT16 src_x, INT16 src_y,
 				  int ntrap, xTrapezoid *traps);
 
 bool
 precise_trapezoid_span_inplace(struct sna *sna,
 				 CARD8 op, PicturePtr src, PicturePtr dst,
-				 PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				 PictFormatPtr maskFormat, unsigned flags,
+				 INT16 src_x, INT16 src_y,
 				 int ntrap, xTrapezoid *traps,
 				 bool fallback);
 
@@ -99,13 +103,15 @@ precise_trapezoid_span_converter(struct sna *sna,
 
 bool
 precise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
-				   PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				   PictFormatPtr maskFormat, unsigned flags,
+				   INT16 src_x, INT16 src_y,
 				   int ntrap, xTrapezoid *traps);
 
 bool
 precise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
-				  PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
-				  int ntrap, xTrapezoid *traps);
+				PictFormatPtr maskFormat, unsigned flags,
+				INT16 src_x, INT16 src_y,
+				int ntrap, xTrapezoid *traps);
 
 static inline bool is_mono(PicturePtr dst, PictFormatPtr mask)
 {
@@ -120,7 +126,8 @@ static inline bool is_precise(PicturePtr dst, PictFormatPtr mask)
 static inline bool
 trapezoid_span_inplace(struct sna *sna,
 		       CARD8 op, PicturePtr src, PicturePtr dst,
-		       PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+		       PictFormatPtr maskFormat, unsigned flags,
+		       INT16 src_x, INT16 src_y,
 		       int ntrap, xTrapezoid *traps,
 		       bool fallback)
 {
@@ -143,9 +150,9 @@ trapezoid_span_inplace(struct sna *sna,
 	if (is_mono(dst, maskFormat))
 		return mono_trapezoid_span_inplace(sna, op, src, dst, src_x, src_y, ntrap, traps);
 	else if (is_precise(dst, maskFormat))
-		return precise_trapezoid_span_inplace(sna, op, src, dst, maskFormat, src_x, src_y, ntrap, traps, fallback);
+		return precise_trapezoid_span_inplace(sna, op, src, dst, maskFormat, flags, src_x, src_y, ntrap, traps, fallback);
 	else
-		return imprecise_trapezoid_span_inplace(sna, op, src, dst, maskFormat, src_x, src_y, ntrap, traps, fallback);
+		return imprecise_trapezoid_span_inplace(sna, op, src, dst, maskFormat, flags, src_x, src_y, ntrap, traps, fallback);
 }
 
 static inline bool
@@ -168,30 +175,32 @@ trapezoid_span_converter(struct sna *sna,
 
 static inline bool
 trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
-			 PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+			 PictFormatPtr maskFormat, unsigned flags,
+			 INT16 src_x, INT16 src_y,
 			 int ntrap, xTrapezoid *traps)
 {
 	if (NO_SCAN_CONVERTER)
 		return false;
 
 	if (is_precise(dst, maskFormat))
-		return precise_trapezoid_mask_converter(op, src, dst, maskFormat, src_x, src_y, ntrap, traps);
+		return precise_trapezoid_mask_converter(op, src, dst, maskFormat, flags, src_x, src_y, ntrap, traps);
 	else
-		return imprecise_trapezoid_mask_converter(op, src, dst, maskFormat, src_x, src_y, ntrap, traps);
+		return imprecise_trapezoid_mask_converter(op, src, dst, maskFormat, flags, src_x, src_y, ntrap, traps);
 }
 
 static inline bool
 trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
-			PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+			PictFormatPtr maskFormat, unsigned flags,
+			INT16 src_x, INT16 src_y,
 			int ntrap, xTrapezoid *traps)
 {
 	if (NO_SCAN_CONVERTER)
 		return false;
 
 	if (is_precise(dst, maskFormat))
-		return precise_trapezoid_span_fallback(op, src, dst, maskFormat, src_x, src_y, ntrap, traps);
+		return precise_trapezoid_span_fallback(op, src, dst, maskFormat, flags, src_x, src_y, ntrap, traps);
 	else
-		return imprecise_trapezoid_span_fallback(op, src, dst, maskFormat, src_x, src_y, ntrap, traps);
+		return imprecise_trapezoid_span_fallback(op, src, dst, maskFormat, flags, src_x, src_y, ntrap, traps);
 }
 
 bool
diff --git a/src/sna/sna_trapezoids_imprecise.c b/src/sna/sna_trapezoids_imprecise.c
index 6ca3fedb..88cb660e 100644
--- a/src/sna/sna_trapezoids_imprecise.c
+++ b/src/sna/sna_trapezoids_imprecise.c
@@ -389,7 +389,7 @@ cell_list_find(struct cell_list *cells, int x)
 	} while (1);
 
 	if (tail->x != x)
-		tail = cell_list_alloc (cells, tail, x);
+		tail = cell_list_alloc(cells, tail, x);
 
 	return cells->cursor = tail;
 }
@@ -777,7 +777,7 @@ merge_unsorted_edges(struct edge *head, struct edge *unsorted)
 
 /* Test if the edges on the active list can be safely advanced by a
  * full row without intersections or any edges ending. */
-inline static bool
+inline static int
 can_full_step(struct active_list *active)
 {
 	const struct edge *e;
@@ -785,7 +785,7 @@ can_full_step(struct active_list *active)
 
 	assert(active->head.next != &active->tail);
 	for (e = active->head.next; &active->tail != e; e = e->next) {
-		assert(e->height_left >= 0);
+		assert(e->height_left > 0);
 
 		if (e->dy != 0)
 			return 0;
@@ -842,6 +842,7 @@ nonzero_subrow(struct active_list *active, struct cell_list *coverages)
 			xstart = edge->next->x.quo;
 		}
 
+		assert(edge->height_left > 0);
 		if (--edge->height_left) {
 			if (edge->dy) {
 				edge->x.quo += edge->dxdy.quo;
@@ -885,7 +886,7 @@ nonzero_row(struct active_list *active, struct cell_list *coverages)
 
 		left->height_left -= FAST_SAMPLES_Y;
 		assert(left->height_left >= 0);
-		if (! left->height_left) {
+		if (!left->height_left) {
 			left->prev->next = left->next;
 			left->next->prev = left->prev;
 		}
@@ -967,6 +968,7 @@ step_edges(struct active_list *active, int count)
 	count *= FAST_SAMPLES_Y;
 	for (edge = active->head.next; edge != &active->tail; edge = edge->next) {
 		edge->height_left -= count;
+		assert(edge->height_left >= 0);
 		if (!edge->height_left) {
 			edge->prev->next = edge->next;
 			edge->next->prev = edge->prev;
@@ -1266,6 +1268,7 @@ inplace_row(struct active_list *active, uint8_t *row, int width)
 		int lix, rix;
 
 		left->height_left -= FAST_SAMPLES_Y;
+		assert(left->height_left >= 0);
 		if (!left->height_left) {
 			left->prev->next = left->next;
 			left->next->prev = left->prev;
@@ -1274,6 +1277,7 @@ inplace_row(struct active_list *active, uint8_t *row, int width)
 		right = left->next;
 		do {
 			right->height_left -= FAST_SAMPLES_Y;
+			assert(right->height_left >= 0);
 			if (!right->height_left) {
 				right->prev->next = right->next;
 				right->next->prev = right->prev;
@@ -1414,6 +1418,7 @@ inplace_subrow(struct active_list *active, int8_t *row,
 			xstart = MAX(edge->x.quo, 0);
 		}
 
+		assert(edge->height_left > 0);
 		if (--edge->height_left) {
 			if (edge->dy) {
 				edge->x.quo += edge->dxdy.quo;
@@ -1704,8 +1709,8 @@ struct span_thread {
 #define SPAN_THREAD_MAX_BOXES (8192/sizeof(struct sna_opacity_box))
 struct span_thread_boxes {
 	const struct sna_composite_spans_op *op;
-	struct sna_opacity_box boxes[SPAN_THREAD_MAX_BOXES];
 	int num_boxes;
+	struct sna_opacity_box boxes[SPAN_THREAD_MAX_BOXES];
 };
 
 static void span_thread_add_boxes(struct sna *sna, void *data,
@@ -1943,7 +1948,9 @@ imprecise_trapezoid_span_converter(struct sna *sna,
 	dy *= FAST_SAMPLES_Y;
 
 	num_threads = 1;
-	if (!NO_GPU_THREADS && tmp.thread_boxes &&
+	if (!NO_GPU_THREADS &&
+	    (flags & COMPOSITE_SPANS_RECTILINEAR) == 0 &&
+	    tmp.thread_boxes &&
 	    thread_choose_span(&tmp, dst, maskFormat, &clip))
 		num_threads = sna_use_threads(clip.extents.x2-clip.extents.x1,
 					      clip.extents.y2-clip.extents.y1,
@@ -2063,7 +2070,8 @@ tor_blt_mask_mono(struct sna *sna,
 
 bool
 imprecise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
-				   PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				   PictFormatPtr maskFormat, unsigned flags,
+				   INT16 src_x, INT16 src_y,
 				   int ntrap, xTrapezoid *traps)
 {
 	struct tor tor;
@@ -2083,7 +2091,7 @@ imprecise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!imprecise_trapezoid_mask_converter(op, src, dst, NULL,
+			if (!imprecise_trapezoid_mask_converter(op, src, dst, NULL, flags,
 								src_x, src_y, 1, traps++))
 				return false;
 		} while (--ntrap);
@@ -2183,8 +2191,8 @@ imprecise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 }
 
 struct inplace {
-	uint32_t stride;
 	uint8_t *ptr;
+	uint32_t stride;
 	union {
 		uint8_t opacity;
 		uint32_t color;
@@ -2917,7 +2925,8 @@ static void inplace_thread(void *arg)
 bool
 imprecise_trapezoid_span_inplace(struct sna *sna,
 				 CARD8 op, PicturePtr src, PicturePtr dst,
-				 PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				 PictFormatPtr maskFormat, unsigned flags,
+				 INT16 src_x, INT16 src_y,
 				 int ntrap, xTrapezoid *traps,
 				 bool fallback)
 {
@@ -3007,7 +3016,7 @@ imprecise_trapezoid_span_inplace(struct sna *sna,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!imprecise_trapezoid_span_inplace(sna, op, src, dst, NULL,
+			if (!imprecise_trapezoid_span_inplace(sna, op, src, dst, NULL, flags,
 							      src_x, src_y, 1, traps++,
 							      fallback))
 				return false;
@@ -3064,16 +3073,17 @@ imprecise_trapezoid_span_inplace(struct sna *sna,
 	dx = dst->pDrawable->x * FAST_SAMPLES_X;
 	dy = dst->pDrawable->y * FAST_SAMPLES_Y;
 
-
 	inplace.ptr = pixmap->devPrivate.ptr;
 	if (get_drawable_deltas(dst->pDrawable, pixmap, &dst_x, &dst_y))
 		inplace.ptr += dst_y * pixmap->devKind + dst_x;
 	inplace.stride = pixmap->devKind;
 	inplace.opacity = color >> 24;
 
-	num_threads = sna_use_threads(region.extents.x2 - region.extents.x1,
-				      region.extents.y2 - region.extents.y1,
-				      16);
+	num_threads = 1;
+	if ((flags & COMPOSITE_SPANS_RECTILINEAR) == 0)
+		num_threads = sna_use_threads(region.extents.x2 - region.extents.x1,
+					      region.extents.y2 - region.extents.y1,
+					      16);
 	if (num_threads == 1) {
 		struct tor tor;
 
@@ -3143,7 +3153,8 @@ imprecise_trapezoid_span_inplace(struct sna *sna,
 
 bool
 imprecise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
-				  PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				  PictFormatPtr maskFormat, unsigned flags,
+				  INT16 src_x, INT16 src_y,
 				  int ntrap, xTrapezoid *traps)
 {
 	struct tor tor;
@@ -3163,7 +3174,7 @@ imprecise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!imprecise_trapezoid_span_fallback(op, src, dst, NULL,
+			if (!imprecise_trapezoid_span_fallback(op, src, dst, NULL, flags,
 							       src_x, src_y, 1, traps++))
 				return false;
 		} while (--ntrap);
diff --git a/src/sna/sna_trapezoids_precise.c b/src/sna/sna_trapezoids_precise.c
index bfafca03..b4d7592d 100644
--- a/src/sna/sna_trapezoids_precise.c
+++ b/src/sna/sna_trapezoids_precise.c
@@ -782,7 +782,7 @@ merge_unsorted_edges(struct edge *head, struct edge *unsorted)
 
 /* Test if the edges on the active list can be safely advanced by a
  * full row without intersections or any edges ending. */
-inline static bool
+inline static int
 can_full_step(struct active_list *active)
 {
 	const struct edge *e;
@@ -790,7 +790,7 @@ can_full_step(struct active_list *active)
 
 	assert(active->head.next != &active->tail);
 	for (e = active->head.next; &active->tail != e; e = e->next) {
-		assert(e->height_left >= 0);
+		assert(e->height_left > 0);
 
 		if (e->dy != 0)
 			return 0;
@@ -848,6 +848,7 @@ nonzero_subrow(struct active_list *active, struct cell_list *coverages)
 			xstart = edge->next->x.quo;
 		}
 
+		assert(edge->height_left > 0);
 		if (--edge->height_left) {
 			if (edge->dy) {
 				edge->x.quo += edge->dxdy.quo;
@@ -972,6 +973,7 @@ step_edges(struct active_list *active, int count)
 	count *= SAMPLES_Y;
 	for (edge = active->head.next; edge != &active->tail; edge = edge->next) {
 		edge->height_left -= count;
+		assert(edge->height_left >= 0);
 		if (!edge->height_left) {
 			edge->prev->next = edge->next;
 			edge->next->prev = edge->prev;
@@ -1223,6 +1225,7 @@ inplace_row(struct active_list *active, uint8_t *row, int width)
 		int lix, rix;
 
 		left->height_left -= SAMPLES_Y;
+		assert(left->height_left >= 0);
 		if (!left->height_left) {
 			left->prev->next = left->next;
 			left->next->prev = left->prev;
@@ -1231,6 +1234,7 @@ inplace_row(struct active_list *active, uint8_t *row, int width)
 		right = left->next;
 		do {
 			right->height_left -= SAMPLES_Y;
+			assert(right->height_left >= 0);
 			if (!right->height_left) {
 				right->prev->next = right->next;
 				right->next->prev = right->prev;
@@ -1342,6 +1346,7 @@ inplace_subrow(struct active_list *active, int8_t *row, int width)
 		} else
 			SAMPLES_X_TO_INT_FRAC(edge->x.quo, lix, lfx);
 
+		assert(edge->height_left > 0);
 		if (--edge->height_left) {
 			if (edge->dy) {
 				edge->x.quo += edge->dxdy.quo;
@@ -1377,6 +1382,7 @@ inplace_subrow(struct active_list *active, int8_t *row, int width)
 			if (0 == winding && edge->x.quo != next->x.quo)
 				break;
 
+			assert(edge->height_left > 0);
 			if (--edge->height_left) {
 				if (edge->dy) {
 					edge->x.quo += edge->dxdy.quo;
@@ -1416,6 +1422,7 @@ inplace_subrow(struct active_list *active, int8_t *row, int width)
 		} else
 			SAMPLES_X_TO_INT_FRAC(edge->x.quo, rix, rfx);
 
+		assert(edge->height_left > 0);
 		if (--edge->height_left) {
 			if (edge->dy) {
 				edge->x.quo += edge->dxdy.quo;
@@ -1509,9 +1516,8 @@ tor_inplace(struct tor *converter, PixmapPtr scratch)
 			do_full_step = can_full_step(active);
 		}
 
-		__DBG(("%s: y=%d, do_full_step=%d, new edges=%d, min_height=%d, vertical=%d\n",
-		       __FUNCTION__,
-		       i, do_full_step,
+		__DBG(("%s: y=%d, do_full_step=%d, new edges=%d\n",
+		       __FUNCTION__, i, do_full_step,
 		       polygon->y_buckets[i] != NULL));
 		if (do_full_step) {
 			memset(ptr, 0, width);
@@ -1633,8 +1639,8 @@ struct span_thread {
 #define SPAN_THREAD_MAX_BOXES (8192/sizeof(struct sna_opacity_box))
 struct span_thread_boxes {
 	const struct sna_composite_spans_op *op;
-	struct sna_opacity_box boxes[SPAN_THREAD_MAX_BOXES];
 	int num_boxes;
+	struct sna_opacity_box boxes[SPAN_THREAD_MAX_BOXES];
 };
 
 static void span_thread_add_boxes(struct sna *sna, void *data,
@@ -1792,8 +1798,9 @@ precise_trapezoid_span_converter(struct sna *sna,
 
 #if 1
 	if (((clip.extents.y2 - clip.extents.y1) | (clip.extents.x2 - clip.extents.x1)) < 32) {
-		DBG(("%s: fallback -- traps extents too small %dx%d\n",
-		     __FUNCTION__, extents.y2 - extents.y1, extents.x2 - extents.x1));
+		DBG(("%s: fallback -- traps extents too small %dx%d\n", __FUNCTION__,
+		     clip.extents.y2 - clip.extents.y1,
+		     clip.extents.x2 - clip.extents.x1));
 		return false;
 	}
 #endif
@@ -1867,7 +1874,9 @@ precise_trapezoid_span_converter(struct sna *sna,
 	dy *= SAMPLES_Y;
 
 	num_threads = 1;
-	if (!NO_GPU_THREADS && tmp.thread_boxes &&
+	if (!NO_GPU_THREADS &&
+	    (flags & COMPOSITE_SPANS_RECTILINEAR) == 0 &&
+	    tmp.thread_boxes &&
 	    thread_choose_span(&tmp, dst, maskFormat, &clip))
 		num_threads = sna_use_threads(clip.extents.x2-clip.extents.x1,
 					      clip.extents.y2-clip.extents.y1,
@@ -2022,7 +2031,8 @@ mask_thread(void *arg)
 
 bool
 precise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
-				 PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				 PictFormatPtr maskFormat, unsigned flags,
+				 INT16 src_x, INT16 src_y,
 				 int ntrap, xTrapezoid *traps)
 {
 	ScreenPtr screen = dst->pDrawable->pScreen;
@@ -2042,7 +2052,7 @@ precise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!precise_trapezoid_mask_converter(op, src, dst, NULL,
+			if (!precise_trapezoid_mask_converter(op, src, dst, NULL, flags,
 							      src_x, src_y, 1, traps++))
 				return false;
 		} while (--ntrap);
@@ -2088,9 +2098,11 @@ precise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 	DBG(("%s: created buffer %p, stride %d\n",
 	     __FUNCTION__, scratch->devPrivate.ptr, scratch->devKind));
 
-	num_threads = sna_use_threads(extents.x2 - extents.x1,
-				      extents.y2 - extents.y1,
-				      4);
+	num_threads = 1;
+	if ((flags & COMPOSITE_SPANS_RECTILINEAR) == 0)
+		num_threads = sna_use_threads(extents.x2 - extents.x1,
+					      extents.y2 - extents.y1,
+					      4);
 	if (num_threads == 1) {
 		struct tor tor;
 
@@ -2180,8 +2192,8 @@ precise_trapezoid_mask_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 }
 
 struct inplace {
-	uint32_t stride;
 	uint8_t *ptr;
+	uint32_t stride;
 	union {
 		uint8_t opacity;
 		uint32_t color;
@@ -2624,7 +2636,7 @@ static bool
 trapezoid_span_inplace__x8r8g8b8(CARD8 op,
 				 PicturePtr dst,
 				 PicturePtr src, int16_t src_x, int16_t src_y,
-				 PictFormatPtr maskFormat,
+				 PictFormatPtr maskFormat, unsigned flags,
 				 int ntrap, xTrapezoid *traps)
 {
 	uint32_t color;
@@ -2664,7 +2676,8 @@ trapezoid_span_inplace__x8r8g8b8(CARD8 op,
 			/* XXX unwind errors? */
 			if (!trapezoid_span_inplace__x8r8g8b8(op, dst,
 							      src, src_x, src_y,
-							      NULL, 1, traps++))
+							      NULL, flags,
+							      1, traps++))
 				return false;
 		} while (--ntrap);
 		return true;
@@ -2711,9 +2724,11 @@ trapezoid_span_inplace__x8r8g8b8(CARD8 op,
 	dx = dst->pDrawable->x * SAMPLES_X;
 	dy = dst->pDrawable->y * SAMPLES_Y;
 
-	num_threads = sna_use_threads(4*(region.extents.x2 - region.extents.x1),
-				      region.extents.y2 - region.extents.y1,
-				      4);
+	num_threads = 1;
+	if ((flags & COMPOSITE_SPANS_RECTILINEAR) == 0 && (lerp || is_solid))
+		num_threads = sna_use_threads(4*(region.extents.x2 - region.extents.x1),
+					      region.extents.y2 - region.extents.y1,
+					      4);
 
 	DBG(("%s: %dx%d, format=%x, op=%d, lerp?=%d, num_threads=%d\n",
 	     __FUNCTION__,
@@ -2907,7 +2922,8 @@ static void inplace_thread(void *arg)
 bool
 precise_trapezoid_span_inplace(struct sna *sna,
 			       CARD8 op, PicturePtr src, PicturePtr dst,
-			       PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+			       PictFormatPtr maskFormat, unsigned flags,
+			       INT16 src_x, INT16 src_y,
 			       int ntrap, xTrapezoid *traps,
 			       bool fallback)
 {
@@ -2928,7 +2944,7 @@ precise_trapezoid_span_inplace(struct sna *sna,
 	if (dst->format == PICT_a8r8g8b8 || dst->format == PICT_x8r8g8b8)
 		return trapezoid_span_inplace__x8r8g8b8(op, dst,
 							src, src_x, src_y,
-							maskFormat,
+							maskFormat, flags,
 							ntrap, traps);
 
 	if (!sna_picture_is_solid(src, &color)) {
@@ -2997,7 +3013,7 @@ precise_trapezoid_span_inplace(struct sna *sna,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!precise_trapezoid_span_inplace(sna, op, src, dst, NULL,
+			if (!precise_trapezoid_span_inplace(sna, op, src, dst, NULL, flags,
 							    src_x, src_y, 1, traps++,
 							    fallback))
 				return false;
@@ -3061,9 +3077,11 @@ precise_trapezoid_span_inplace(struct sna *sna,
 	inplace.stride = pixmap->devKind;
 	inplace.opacity = color >> 24;
 
-	num_threads = sna_use_threads(region.extents.x2 - region.extents.x1,
-				      region.extents.y2 - region.extents.y1,
-				      4);
+	num_threads = 1;
+	if ((flags & COMPOSITE_SPANS_RECTILINEAR) == 0)
+		num_threads = sna_use_threads(region.extents.x2 - region.extents.x1,
+					      region.extents.y2 - region.extents.y1,
+					      4);
 	if (num_threads == 1) {
 		struct tor tor;
 
@@ -3133,7 +3151,8 @@ precise_trapezoid_span_inplace(struct sna *sna,
 
 bool
 precise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
-				PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
+				PictFormatPtr maskFormat, unsigned flags,
+				INT16 src_x, INT16 src_y,
 				int ntrap, xTrapezoid *traps)
 {
 	ScreenPtr screen = dst->pDrawable->pScreen;
@@ -3152,7 +3171,7 @@ precise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!precise_trapezoid_span_fallback(op, src, dst, NULL,
+			if (!precise_trapezoid_span_fallback(op, src, dst, NULL, flags,
 							     src_x, src_y, 1, traps++))
 				return false;
 		} while (--ntrap);
@@ -3197,9 +3216,11 @@ precise_trapezoid_span_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
 	DBG(("%s: created buffer %p, stride %d\n",
 	     __FUNCTION__, scratch->devPrivate.ptr, scratch->devKind));
 
-	num_threads = sna_use_threads(extents.x2 - extents.x1,
-				      extents.y2 - extents.y1,
-				      4);
+	num_threads = 1;
+	if ((flags & COMPOSITE_SPANS_RECTILINEAR) == 0)
+		num_threads = sna_use_threads(extents.x2 - extents.x1,
+					      extents.y2 - extents.y1,
+					      4);
 	if (num_threads == 1) {
 		struct tor tor;