JIT: Add loop-aware RPO, and use as LSRA's block sequence (#108086)

Part of #107749, and follow-up to #107927. When computing a RPO of the flow graph, ensuring that the entirety of a loop body is visited before any of the loop's successors has the benefit of keeping the loop body compact in the traversal. This is certainly ideal when computing an initial block layout, and may be preferable for register allocation, too. Thus, this change formalizes loop-aware RPO creation as part of the flowgraph API surface, and uses it for LSRA's block sequence.

src/coreclr/jit/compiler.h

@@ -6286,6 +6286,9 @@ class Compiler
     FlowGraphDfsTree* fgComputeDfs();
     void fgInvalidateDfsTree();
 
+    template <typename TFunc>
+    void fgVisitBlocksInLoopAwareRPO(FlowGraphDfsTree* dfsTree, FlowGraphNaturalLoops* loops, TFunc func);
+
     void fgRemoveReturnBlock(BasicBlock* block);
 
     void fgConvertBBToThrowBB(BasicBlock* block);

src/coreclr/jit/compiler.hpp

@@ -4974,6 +4974,75 @@ unsigned Compiler::fgRunDfs(VisitPreorder visitPreorder, VisitPostorder visitPos
     return preOrderIndex;
 }
 
+//------------------------------------------------------------------------
+// fgVisitBlocksInLoopAwareRPO: Visit the blocks in 'dfsTree' in reverse post-order,
+// but ensure loop bodies are visited before loop successors.
+//
+// Type parameters:
+//   TFunc - Callback functor type
+//
+// Parameters:
+//   dfsTree - The DFS tree of the flow graph
+//   loops   - A collection of the loops in the flow graph
+//   func    - Callback functor that operates on a BasicBlock*
+//
+// Returns:
+//   A postorder traversal with compact loop bodies.
+//
+template <typename TFunc>
+void Compiler::fgVisitBlocksInLoopAwareRPO(FlowGraphDfsTree* dfsTree, FlowGraphNaturalLoops* loops, TFunc func)
+{
+    assert(dfsTree != nullptr);
+    assert(loops != nullptr);
+
+    // We will start by visiting blocks in reverse post-order.
+    // If we encounter the header of a loop, we will visit the loop's remaining blocks next
+    // to keep the loop body compact in the visitation order.
+    // We have to do this recursively to handle nested loops.
+    // Since the presence of loops implies we will try to visit some blocks more than once,
+    // we need to track visited blocks.
+    struct LoopAwareVisitor
+    {
+        BitVecTraits           traits;
+        BitVec                 visitedBlocks;
+        FlowGraphNaturalLoops* loops;
+        TFunc                  func;
+
+        LoopAwareVisitor(FlowGraphDfsTree* dfsTree, FlowGraphNaturalLoops* loops, TFunc func)
+            : traits(dfsTree->PostOrderTraits())
+            , visitedBlocks(BitVecOps::MakeEmpty(&traits))
+            , loops(loops)
+            , func(func)
+        {
+        }
+
+        void VisitBlock(BasicBlock* block)
+        {
+            if (BitVecOps::TryAddElemD(&traits, visitedBlocks, block->bbPostorderNum))
+            {
+                func(block);
+
+                FlowGraphNaturalLoop* const loop = loops->GetLoopByHeader(block);
+                if (loop != nullptr)
+                {
+                    loop->VisitLoopBlocksReversePostOrder([&](BasicBlock* block) {
+                        VisitBlock(block);
+                        return BasicBlockVisit::Continue;
+                    });
+                }
+            }
+        }
+    };
+
+    LoopAwareVisitor visitor(dfsTree, loops, func);
+
+    for (unsigned i = dfsTree->GetPostOrderCount(); i != 0; i--)
+    {
+        BasicBlock* const block = dfsTree->GetPostOrder(i - 1);
+        visitor.VisitBlock(block);
+    }
+}
+
 //------------------------------------------------------------------------------
 // FlowGraphNaturalLoop::VisitLoopBlocksReversePostOrder: Visit all of the
 // loop's blocks in reverse post order.

src/coreclr/jit/lsra.cpp

@@ -953,9 +953,29 @@ void LinearScan::setBlockSequence()
 
     assert((blockSequence == nullptr) && (bbSeqCount == 0));
     FlowGraphDfsTree* const dfsTree = compiler->fgComputeDfs</* useProfile */ true>();
-    blockSequence                   = dfsTree->GetPostOrder();
-    bbNumMaxBeforeResolution        = compiler->fgBBNumMax;
-    blockInfo                       = new (compiler, CMK_LSRA) LsraBlockInfo[bbNumMaxBeforeResolution + 1];
+
+    if (compiler->opts.OptimizationEnabled() && dfsTree->HasCycle())
+    {
+        // Ensure loop bodies are compact in the visitation order
+        FlowGraphNaturalLoops* const loops = FlowGraphNaturalLoops::Find(dfsTree);
+        blockSequence                      = new (compiler, CMK_LSRA) BasicBlock*[compiler->fgBBcount];
+        unsigned index                     = dfsTree->GetPostOrderCount();
+
+        auto addToSequence = [this, &index](BasicBlock* block) {
+            assert(index != 0);
+            blockSequence[--index] = block;
+        };
+
+        compiler->fgVisitBlocksInLoopAwareRPO(dfsTree, loops, addToSequence);
+    }
+    else
+    {
+        // TODO: Just use lexical block order in MinOpts
+        blockSequence = dfsTree->GetPostOrder();
+    }
+
+    bbNumMaxBeforeResolution = compiler->fgBBNumMax;
+    blockInfo                = new (compiler, CMK_LSRA) LsraBlockInfo[bbNumMaxBeforeResolution + 1];
 
     // Flip the DFS traversal to get the reverse post-order traversal
     // (this is the order in which blocks will be allocated)