Add loop-aware RPO and use in LSRA

src/coreclr/jit/compiler.cpp

@@ -4524,7 +4524,7 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
     compFunctionTraceStart();
 
     // Enable flow graph checks
-    activePhaseChecks |= PhaseChecks::CHECK_FG;
+    activePhaseChecks |= PhaseChecks::CHECK_FG | PhaseChecks::CHECK_FG_ANNOTATIONS;
 
     // Prepare for importation
     //
@@ -5239,6 +5239,10 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
     lvaTrackedFixed                    = true;
     const unsigned numBlocksBeforeLSRA = fgBBcount;
 
+    // Backend phases will use a loop-aware RPO traversal of the flowgraph,
+    // so skip checking pre/postorder numbers for correctness.
+    activePhaseChecks &= ~PhaseChecks::CHECK_FG_ANNOTATIONS;
+
     // Now that lowering is completed we can proceed to perform register allocation
     //
     auto linearScanPhase = [this] {

src/coreclr/jit/compiler.h

@@ -1527,15 +1527,16 @@ extern const char* PhaseEnums[];
 // clang-format off
 enum class PhaseChecks : unsigned int
 {
-    CHECK_NONE          = 0,
-    CHECK_IR            = 1 << 0, // ir flags, etc
-    CHECK_UNIQUE        = 1 << 1, // tree node uniqueness
-    CHECK_FG            = 1 << 2, // flow graph integrity
-    CHECK_EH            = 1 << 3, // eh table integrity
-    CHECK_LOOPS         = 1 << 4, // loop integrity/canonicalization
-    CHECK_LIKELIHOODS   = 1 << 5, // profile data likelihood integrity
-    CHECK_PROFILE       = 1 << 6, // profile data full integrity
-    CHECK_LINKED_LOCALS = 1 << 7, // check linked list of locals
+    CHECK_NONE           = 0,
+    CHECK_IR             = 1 << 0, // ir flags, etc
+    CHECK_UNIQUE         = 1 << 1, // tree node uniqueness
+    CHECK_FG             = 1 << 2, // flow graph integrity
+    CHECK_EH             = 1 << 3, // eh table integrity
+    CHECK_LOOPS          = 1 << 4, // loop integrity/canonicalization
+    CHECK_LIKELIHOODS    = 1 << 5, // profile data likelihood integrity
+    CHECK_PROFILE        = 1 << 6, // profile data full integrity
+    CHECK_LINKED_LOCALS  = 1 << 7, // check linked list of locals
+    CHECK_FG_ANNOTATIONS = 1 << 8, // check flowgraph annotation data structures
 };
 
 inline constexpr PhaseChecks operator ~(PhaseChecks a)
@@ -6286,6 +6287,7 @@ class Compiler
 
     template <const bool useProfile = false>
     FlowGraphDfsTree* fgComputeDfs();
+    FlowGraphDfsTree* fgComputeLoopAwareDfs();
     void fgInvalidateDfsTree();
 
     void fgRemoveReturnBlock(BasicBlock* block);

src/coreclr/jit/flowgraph.cpp

@@ -4120,6 +4120,62 @@ FlowGraphDfsTree* Compiler::fgComputeDfs()
 template FlowGraphDfsTree* Compiler::fgComputeDfs<false>();
 template FlowGraphDfsTree* Compiler::fgComputeDfs<true>();
 
+FlowGraphDfsTree* Compiler::fgComputeLoopAwareDfs()
+{
+    if (m_dfsTree == nullptr)
+    {
+        m_dfsTree = fgComputeDfs</* useProfile */ true>();
+    }
+
+    if (!m_dfsTree->HasCycle())
+    {
+        return m_dfsTree;
+    }
+
+    if (m_loops == nullptr)
+    {
+        m_loops = FlowGraphNaturalLoops::Find(m_dfsTree);
+    }
+
+    m_blockToLoop = BlockToNaturalLoopMap::Build(m_loops);
+
+    EnsureBasicBlockEpoch();
+    BlockSet visitedBlocks(BlockSetOps::MakeEmpty(this));
+
+    BasicBlock**   loopAwarePostOrder = new (this, CMK_DepthFirstSearch) BasicBlock*[fgBBcount];
+    const unsigned numBlocks          = m_dfsTree->GetPostOrderCount();
+    unsigned       newIndex           = numBlocks - 1;
+
+    auto visitBlock = [this, loopAwarePostOrder, &visitedBlocks, &newIndex,
+                       numBlocks](BasicBlock* block) -> BasicBlockVisit {
+        if (!BlockSetOps::IsMember(this, visitedBlocks, block->bbNum))
+        {
+            assert(newIndex < numBlocks);
+            loopAwarePostOrder[newIndex--] = block;
+            BlockSetOps::AddElemD(this, visitedBlocks, block->bbNum);
+        }
+
+        return BasicBlockVisit::Continue;
+    };
+
+    for (unsigned i = numBlocks; i != 0; i--)
+    {
+        BasicBlock* const           block = m_dfsTree->GetPostOrder(i - 1);
+        FlowGraphNaturalLoop* const loop  = m_blockToLoop->GetLoop(block);
+
+        if ((loop != nullptr) && (block == loop->GetHeader()))
+        {
+            loop->VisitLoopBlocksReversePostOrder(visitBlock);
+        }
+        else
+        {
+            visitBlock(block);
+        }
+    }
+
+    return new (this, CMK_DepthFirstSearch) FlowGraphDfsTree(this, loopAwarePostOrder, numBlocks, /* hasCycle */ true);
+}
+
 //------------------------------------------------------------------------
 // fgInvalidateDfsTree: Invalidate computed DFS tree and dependent annotations
 // (like loops, dominators and SSA).

src/coreclr/jit/lsra.cpp

@@ -949,10 +949,11 @@ void LinearScan::setBlockSequence()
 #endif // DEBUG
 
     // Initialize the "visited" blocks set.
-    bbVisitedSet = BlockSetOps::MakeEmpty(compiler);
+    bbVisitedSet        = BlockSetOps::MakeEmpty(compiler);
+    compiler->m_dfsTree = compiler->fgComputeLoopAwareDfs();
 
     assert((blockSequence == nullptr) && (bbSeqCount == 0));
-    FlowGraphDfsTree* const dfsTree = compiler->fgComputeDfs</* useProfile */ true>();
+    FlowGraphDfsTree* const dfsTree = compiler->m_dfsTree;
     blockSequence                   = dfsTree->GetPostOrder();
     bbNumMaxBeforeResolution        = compiler->fgBBNumMax;
     blockInfo                       = new (compiler, CMK_LSRA) LsraBlockInfo[bbNumMaxBeforeResolution + 1];

src/coreclr/jit/phase.cpp

@@ -189,7 +189,10 @@ void Phase::PostPhase(PhaseStatus status)
             comp->fgDebugCheckLinkedLocals();
         }
 
-        comp->fgDebugCheckFlowGraphAnnotations();
+        if (hasFlag(checks, PhaseChecks::CHECK_FG_ANNOTATIONS))
+        {
+            comp->fgDebugCheckFlowGraphAnnotations();
+        }
     }
 #endif // DEBUG
 }