BPSectionOrderer: stabilize iteration order and node order

Exposed by the test added in the reverted llvm#120514

* Fix libstdc++/libc++ differences due to nth_element. llvm#125450 (comment)
* Fix LLVM_ENABLE_REVERSE_ITERATION=1 differences
* Fix potential issue in `currentSize += D::getSize(*sections[*sectionIdxs.begin()])` where DenseSet was used, though not covered by a test

Author: MaskRay

lld/MachO/BPSectionOrderer.cpp

@@ -110,7 +110,7 @@ DenseMap<const InputSection *, int> lld::macho::runBalancedPartitioning(
     bool compressionSortStartupFunctions, bool verbose) {
   // Collect candidate sections and associated symbols.
   SmallVector<InputSection *> sections;
-  DenseMap<CachedHashStringRef, DenseSet<unsigned>> rootSymbolToSectionIdxs;
+  DenseMap<CachedHashStringRef, std::set<unsigned>> rootSymbolToSectionIdxs;
   for (const auto *file : inputFiles) {
     for (auto *sec : file->sections) {
       for (auto &subsec : sec->subsections) {

lld/include/lld/Common/BPSectionOrdererBase.inc

@@ -22,7 +22,7 @@
 #include "lld/Common/ErrorHandler.h"
 #include "llvm/ADT/CachedHashString.h"
 #include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/MapVector.h"
 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
@@ -35,6 +35,7 @@
 #include "llvm/Support/VirtualFileSystem.h"
 #include <memory>
 #include <optional>
+#include <set>
 
 #define DEBUG_TYPE "bp-section-orderer"
 
@@ -60,7 +61,7 @@ template <class D> struct BPOrderer {
                     bool forDataCompression,
                     bool compressionSortStartupFunctions, bool verbose,
                     llvm::ArrayRef<Section *> sections,
-                    const DenseMap<CachedHashStringRef, DenseSet<unsigned>>
+                    const DenseMap<CachedHashStringRef, std::set<unsigned>>
                         &rootSymbolToSectionIdxs)
       -> llvm::DenseMap<const Section *, int>;
 };
@@ -73,7 +74,7 @@ static SmallVector<std::pair<unsigned, UtilityNodes>> getUnsForCompression(
     ArrayRef<const typename D::Section *> sections,
     const DenseMap<const void *, uint64_t> &sectionToIdx,
     ArrayRef<unsigned> sectionIdxs,
-    DenseMap<unsigned, SmallVector<unsigned>> *duplicateSectionIdxs,
+    DenseMap<unsigned, SmallVector<unsigned, 0>> *duplicateSectionIdxs,
     BPFunctionNode::UtilityNodeT &maxUN) {
   TimeTraceScope timeScope("Build nodes for compression");
 
@@ -88,7 +89,7 @@ static SmallVector<std::pair<unsigned, UtilityNodes>> getUnsForCompression(
     hashes.clear();
   }
 
-  DenseMap<uint64_t, unsigned> hashFrequency;
+  MapVector<uint64_t, unsigned> hashFrequency;
   for (auto &[sectionIdx, hashes] : sectionHashes)
     for (auto hash : hashes)
       ++hashFrequency[hash];
@@ -156,7 +157,7 @@ auto BPOrderer<D>::computeOrder(
     StringRef profilePath, bool forFunctionCompression, bool forDataCompression,
     bool compressionSortStartupFunctions, bool verbose,
     ArrayRef<Section *> sections,
-    const DenseMap<CachedHashStringRef, DenseSet<unsigned>>
+    const DenseMap<CachedHashStringRef, std::set<unsigned>>
         &rootSymbolToSectionIdxs) -> DenseMap<const Section *, int> {
   TimeTraceScope timeScope("Setup Balanced Partitioning");
   DenseMap<const void *, uint64_t> sectionToIdx;
@@ -257,7 +258,7 @@ auto BPOrderer<D>::computeOrder(
 
   // Map a section index (order directly) to a list of duplicate section indices
   // (not ordered directly).
-  DenseMap<unsigned, SmallVector<unsigned>> duplicateSectionIdxs;
+  DenseMap<unsigned, SmallVector<unsigned, 0>> duplicateSectionIdxs;
   auto unsForFunctionCompression = getUnsForCompression<D>(
       sections, sectionToIdx, sectionIdxsForFunctionCompression,
       &duplicateSectionIdxs, maxUN);

llvm/lib/Support/BalancedPartitioning.cpp

@@ -305,7 +305,7 @@ void BalancedPartitioning::split(const FunctionNodeRange Nodes,
   unsigned NumNodes = std::distance(Nodes.begin(), Nodes.end());
   auto NodesMid = Nodes.begin() + (NumNodes + 1) / 2;
 
-  std::nth_element(Nodes.begin(), NodesMid, Nodes.end(), [](auto &L, auto &R) {
+  llvm::sort(Nodes.begin(), Nodes.end(), [](auto &L, auto &R) {
     return L.InputOrderIndex < R.InputOrderIndex;
   });