Fix LVDocView::getBookmark() which could be slow or wrong (#313)

Since 5967f72 and some change to be able to traverse RTL tables
for text selection, LVDocView::getBookmark(), which is used to
get the current page xpointer, could be slow on some pages, or
wrong on others.
So, make it clear that, with createXPointer(lvPoint, direction)
and elementFromPoint(lvPoint, direction), using direction=0
will check for x, while direction=1/-1 will not.

Wherever we are only interested in y, and provide a dummy x=0,
we should use direction=1 or -1:
- getPageDocumentRange() was already using 1/-1 exclusively.
- getBookmark() now does too.
- getNodeByPoint() uses 0 to find the exact node at x/y.
- (Others LVDocView methods have not been updated nor checked,
  as they are not used by KOReader).

crengine/src/lvdocview.cpp

@@ -4703,7 +4703,9 @@ ldomXPointer LVDocView::getBookmark() {
 				bool found = false;
 				ldomXPointer fallback_ptr;
 				for (int y = page->start; y < page->start + page->height; y++) {
-					ptr = m_doc->createXPointer(lvPoint(0, y));
+					// Use direction=1 to avoid any x check (in case there
+					// is some left margin)
+					ptr = m_doc->createXPointer(lvPoint(0, y), 1);
 					lvPoint pt = ptr.toPoint();
 					if (pt.y >= page->start) {
 						if (!fallback_ptr)
@@ -4747,7 +4749,7 @@ ldomXPointer LVDocView::getBookmark() {
 			// Let's do the same in that case: get the previous text node
 			// position
 			for (int y = _pos; y >= 0; y--) {
-				ptr = m_doc->createXPointer(lvPoint(0, y));
+				ptr = m_doc->createXPointer(lvPoint(0, y), -1);
 				if (!ptr.isNull())
 					break;
 			}

crengine/src/lvtinydom.cpp

@@ -13430,36 +13430,44 @@ ldomNode * ldomNode::elementFromPoint( lvPoint pt, int direction )
         }
     }
 
-    if (!direction && pt.x != 0) {
+    if ( !direction ) {
         // We shouldn't do the following if we are given a direction
         // (full text search) as we may get locked on some page.
-        // We also don't need to do it if pt.x=0, which is often used
-        // to get current page top or range xpointers.
-        // We are given a x>0 when tap/hold to highlight text or find
-        // a link, and checking x vs fmt.x and width allows for doing
-        // that correctly in 2nd+ table cells.
+        // When interested only in finding the slice of a page
+        // at y (to get the current page top or range xpointers),
+        // use direction=1 or -1.
+        // Use direction=0 to exactly find the final node at y AND x,
+        // which is necessary when selecting text or finding links
+        // in table cells or floats.
         if ( pt.x >= fmt.getX() + fmt.getWidth() ) {
             return NULL;
         }
-        // No more true:
-        //   (No need to check if ( pt.x < fmt.getX() ): we probably
-        //   meet the multiple elements that can be formatted on a same
-        //   line in the order they appear as children of their parent,
-        //   we can simply just ignore those who end before our pt.x)
-        // But check x if we happen to be on a floating node (which,
-        // with float:right, can appear first in the DOM but be
-        // displayed at a higher x)
-        //  if ( pt.x < fmt.getX() && enode->isFloatingBox() ) {
-        //      return NULL;
-        //  }
-        //
-        // Now that we support RTL tables, we can meet cells
-        // no more in logical order.
-        // We could add more conditions (like parentNode->getRendMethod()>=erm_table),
-        // but let's just check this in all cases.
         if ( pt.x < fmt.getX() ) {
             return NULL;
         }
+        // We now do this above check in all cases.
+        // Previously:
+        //
+        //   We also don't need to do it if pt.x=0, which is often used
+        //   to get current page top or range xpointers.
+        //   We are given a x>0 when tap/hold to highlight text or find
+        //   a link, and checking x vs fmt.x and width allows for doing
+        //   that correctly in 2nd+ table cells.
+        //
+        //   No need to check if ( pt.x < fmt.getX() ): we probably
+        //   meet the multiple elements that can be formatted on a same
+        //   line in the order they appear as children of their parent,
+        //   we can simply just ignore those who end before our pt.x.
+        //   But check x if we happen to be on a floating node (which,
+        //   with float:right, can appear first in the DOM but be
+        //   displayed at a higher x)
+        //    if ( pt.x < fmt.getX() && enode->isFloatingBox() ) {
+        //        return NULL;
+        //    }
+        // This is no more true, now that we support RTL tables and
+        // we can meet cells in the reverse of their logical order.
+        // We could add more conditions (like parentNode->getRendMethod()>=erm_table),
+        // but let's just check this in all cases when direction=0.
     }
     if ( rm == erm_final || rm == erm_list_item || rm == erm_table_caption ) {
         // Final node, that's what we looked for