Traverse PDF page tree only once in CPDF_Document Try 3

Now, we do not start traversal from where we were at, but from the top.
This makes the code less prone to bugs, as now there is no need to call
methods to recursively fix things. This will save a lot of time when
the trees are rather flat, as in the PDF file in the bug. It can still
be slow, for instance if we have a chain of page nodes, and the last
in the chain contains all of the pages (this is artificial).

Try 2 at https://codereview.chromium.org/2442403002/

Also added test where Try 2 would have failed.

Tested the pdf from the bug on my Mac:
With this CL: load in 21 seconds
Without this CL: did not load in 4 minutes, got tired of waiting

BUG=chromium:638513

Review-Url: https://codereview.chromium.org/2470803003

1 files changed, 62 insertions, 26 deletions

diff --git a/core/fpdfapi/parser/cpdf_document.cpp b/core/fpdfapi/parser/cpdf_document.cpp
index 64574047e5..8e181de97c 100644
--- a/core/fpdfapi/parser/cpdf_document.cpp
+++ b/core/fpdfapi/parser/cpdf_document.cpp
@@ -336,6 +336,7 @@ CPDF_Document::CPDF_Document(std::unique_ptr<CPDF_Parser> pParser)
       m_pParser(std::move(pParser)),
       m_pRootDict(nullptr),
       m_pInfoDict(nullptr),
+      m_iNextPageToTraverse(0),
       m_bLinearized(false),
       m_iFirstPageNo(0),
       m_dwFirstPageObjNum(0),
@@ -400,40 +401,72 @@ void CPDF_Document::LoadPages() {
   m_PageList.SetSize(RetrievePageCount());
 }
 
-CPDF_Dictionary* CPDF_Document::FindPDFPage(CPDF_Dictionary* pPages,
-                                            int iPage,
-                                            int nPagesToGo,
-                                            int level) {
+CPDF_Dictionary* CPDF_Document::TraversePDFPages(int iPage,
+                                                 int* nPagesToGo,
+                                                 size_t level) {
+  if (*nPagesToGo < 0)
+    return nullptr;
+  CPDF_Dictionary* pPages = m_pTreeTraversal[level].first;
   CPDF_Array* pKidList = pPages->GetArrayFor("Kids");
-  if (!pKidList)
-    return nPagesToGo == 0 ? pPages : nullptr;
+  if (!pKidList) {
+    if (*nPagesToGo != 1)
+      return nullptr;
+    m_PageList.SetAt(iPage, pPages->GetObjNum());
+    return pPages;
+  }
 
-  if (level >= FX_MAX_PAGE_LEVEL)
+  if (level >= FX_MAX_PAGE_LEVEL) {
+    m_pTreeTraversal.pop_back();
     return nullptr;
+  }
 
-  for (size_t i = 0; i < pKidList->GetCount(); i++) {
+  CPDF_Dictionary* page = nullptr;
+  for (size_t i = m_pTreeTraversal[level].second; i < pKidList->GetCount();
+       i++) {
+    if (*nPagesToGo == 0)
+      break;
     CPDF_Dictionary* pKid = pKidList->GetDictAt(i);
     if (!pKid) {
-      nPagesToGo--;
+      (*nPagesToGo)--;
+      m_pTreeTraversal[level].second++;
       continue;
     }
-    if (pKid == pPages)
+    if (pKid == pPages) {
+      m_pTreeTraversal[level].second++;
       continue;
+    }
     if (!pKid->KeyExist("Kids")) {
-      if (nPagesToGo == 0)
-        return pKid;
-
-      m_PageList.SetAt(iPage - nPagesToGo, pKid->GetObjNum());
-      nPagesToGo--;
+      m_PageList.SetAt(iPage - (*nPagesToGo) + 1, pKid->GetObjNum());
+      (*nPagesToGo)--;
+      m_pTreeTraversal[level].second++;
+      if (*nPagesToGo == 0) {
+        page = pKid;
+        break;
+      }
     } else {
-      int nPages = pKid->GetIntegerFor("Count");
-      if (nPagesToGo < nPages)
-        return FindPDFPage(pKid, iPage, nPagesToGo, level + 1);
-
-      nPagesToGo -= nPages;
+      // If the vector has size level+1, the child is not in yet
+      if (m_pTreeTraversal.size() == level + 1)
+        m_pTreeTraversal.push_back(std::make_pair(pKid, 0));
+      // Now m_pTreeTraversal[level+1] should exist and be equal to pKid.
+      CPDF_Dictionary* pageKid = TraversePDFPages(iPage, nPagesToGo, level + 1);
+      // Check if child was completely processed, i.e. it popped itself out
+      if (m_pTreeTraversal.size() == level + 1)
+        m_pTreeTraversal[level].second++;
+      // If child did not finish or if no pages to go, we are done
+      if (m_pTreeTraversal.size() != level + 1 || *nPagesToGo == 0) {
+        page = pageKid;
+        break;
+      }
     }
   }
-  return nullptr;
+  if (m_pTreeTraversal[level].second == pKidList->GetCount())
+    m_pTreeTraversal.pop_back();
+  return page;
+}
+
+void CPDF_Document::ResetTraversal() {
+  m_iNextPageToTraverse = 0;
+  m_pTreeTraversal.clear();
 }
 
 CPDF_Dictionary* CPDF_Document::GetPagesDict() const {
@@ -460,17 +493,18 @@ CPDF_Dictionary* CPDF_Document::GetPage(int iPage) {
   if (objnum) {
     if (CPDF_Dictionary* pDict = ToDictionary(GetOrParseIndirectObject(objnum)))
       return pDict;
+    return nullptr;
   }
 
   CPDF_Dictionary* pPages = GetPagesDict();
   if (!pPages)
     return nullptr;
 
-  CPDF_Dictionary* pPage = FindPDFPage(pPages, iPage, iPage, 0);
-  if (!pPage)
-    return nullptr;
-
-  m_PageList.SetAt(iPage, pPage->GetObjNum());
+  if (m_pTreeTraversal.empty())
+    m_pTreeTraversal.push_back(std::make_pair(pPages, 0));
+  int nPagesToGo = iPage - m_iNextPageToTraverse + 1;
+  CPDF_Dictionary* pPage = TraversePDFPages(iPage, &nPagesToGo, 0);
+  m_iNextPageToTraverse = iPage + 1;
   return pPage;
 }
 
@@ -664,6 +698,7 @@ bool CPDF_Document::InsertDeletePDFPage(CPDF_Dictionary* pPages,
       }
       pPages->SetIntegerFor(
           "Count", pPages->GetIntegerFor("Count") + (bInsert ? 1 : -1));
+      ResetTraversal();
       break;
     }
     int nPages = pKid->GetIntegerFor("Count");
@@ -704,6 +739,7 @@ bool CPDF_Document::InsertNewPage(int iPage, CPDF_Dictionary* pPageDict) {
     pPagesList->Add(new CPDF_Reference(this, pPageDict->GetObjNum()));
     pPages->SetIntegerFor("Count", nPages + 1);
     pPageDict->SetReferenceFor("Parent", this, pPages->GetObjNum());
+    ResetTraversal();
   } else {
     std::set<CPDF_Dictionary*> stack = {pPages};
     if (!InsertDeletePDFPage(pPages, iPage, pPageDict, true, &stack))