JIT: Clean up liveness (#103809)

We have a DFS tree available in both early liveness and SSA's liveness,
and we can use it to make the data flow cheaper by running in an RPO
over the DFS tree. This allows us to propagate the maximal amount of
knowledge in each iteration and also to stop the data flow early when
there is no cycle in the DFS tree.

We do not have the DFS tree available in lowering where we also call
liveness. However, lowering was already iterating all blocks to remove
dead blocks; switch this to `fgDfsBlocksAndRemove` to remove dead blocks
and compute the DFS tree in one go, and remove the old code doing this.

Additionally there was a bunch of logic in liveness to consider debug
scopes for debug codegen, left-over from a time where we tracked GC
pointers even in MinOpts. This code can be deleted since we do not do
liveness in MinOpts anymore.

src/coreclr/jit/block.cpp

@@ -909,7 +909,6 @@ void BasicBlock::CloneBlockState(Compiler* compiler, BasicBlock* to, const Basic
     to->bbStkDepth    = from->bbStkDepth;
     to->bbCodeOffs    = from->bbCodeOffs;
     to->bbCodeOffsEnd = from->bbCodeOffsEnd;
-    VarSetOps::AssignAllowUninitRhs(compiler, to->bbScope, from->bbScope);
 #ifdef DEBUG
     to->bbTgtStkDepth = from->bbTgtStkDepth;
 #endif // DEBUG
@@ -1470,7 +1469,6 @@ void BasicBlock::InitVarSets(Compiler* comp)
     VarSetOps::AssignNoCopy(comp, bbVarDef, VarSetOps::MakeEmpty(comp));
     VarSetOps::AssignNoCopy(comp, bbLiveIn, VarSetOps::MakeEmpty(comp));
     VarSetOps::AssignNoCopy(comp, bbLiveOut, VarSetOps::MakeEmpty(comp));
-    VarSetOps::AssignNoCopy(comp, bbScope, VarSetOps::MakeEmpty(comp));
 
     bbMemoryUse     = emptyMemoryKindSet;
     bbMemoryDef     = emptyMemoryKindSet;
@@ -1672,15 +1670,13 @@ BasicBlock* BasicBlock::New(Compiler* compiler)
         VarSetOps::AssignNoCopy(compiler, block->bbVarDef, VarSetOps::MakeEmpty(compiler));
         VarSetOps::AssignNoCopy(compiler, block->bbLiveIn, VarSetOps::MakeEmpty(compiler));
         VarSetOps::AssignNoCopy(compiler, block->bbLiveOut, VarSetOps::MakeEmpty(compiler));
-        VarSetOps::AssignNoCopy(compiler, block->bbScope, VarSetOps::MakeEmpty(compiler));
     }
     else
     {
         VarSetOps::AssignNoCopy(compiler, block->bbVarUse, VarSetOps::UninitVal());
         VarSetOps::AssignNoCopy(compiler, block->bbVarDef, VarSetOps::UninitVal());
         VarSetOps::AssignNoCopy(compiler, block->bbLiveIn, VarSetOps::UninitVal());
         VarSetOps::AssignNoCopy(compiler, block->bbLiveOut, VarSetOps::UninitVal());
-        VarSetOps::AssignNoCopy(compiler, block->bbScope, VarSetOps::UninitVal());
     }
 
     block->bbMemoryUse     = emptyMemoryKindSet;

src/coreclr/jit/block.h

@@ -1611,8 +1611,6 @@ struct BasicBlock : private LIR::Range
     unsigned bbMemorySsaNumIn[MemoryKindCount];  // The SSA # of memory on entry to the block.
     unsigned bbMemorySsaNumOut[MemoryKindCount]; // The SSA # of memory on exit from the block.
 
-    VARSET_TP bbScope; // variables in scope over the block
-
     void InitVarSets(class Compiler* comp);
 
     /* The following are the standard bit sets for dataflow analysis.

src/coreclr/jit/compiler.cpp

@@ -10923,9 +10923,6 @@ void Compiler::EnregisterStats::RecordLocal(const LclVarDsc* varDsc)
             case DoNotEnregisterReason::NoRegVars:
                 m_noRegVars++;
                 break;
-            case DoNotEnregisterReason::MinOptsGC:
-                m_minOptsGC++;
-                break;
 #if !defined(TARGET_64BIT)
             case DoNotEnregisterReason::LongParamField:
                 m_longParamField++;
@@ -11080,7 +11077,6 @@ void Compiler::EnregisterStats::Dump(FILE* fout) const
     PRINT_STATS(m_structArg, notEnreg);
     PRINT_STATS(m_depField, notEnreg);
     PRINT_STATS(m_noRegVars, notEnreg);
-    PRINT_STATS(m_minOptsGC, notEnreg);
 #if !defined(TARGET_64BIT)
     PRINT_STATS(m_longParamField, notEnreg);
 #endif // !TARGET_64BIT

src/coreclr/jit/compiler.h

@@ -452,7 +452,6 @@ enum class DoNotEnregisterReason
     IsStructArg,        // Is a struct passed as an argument in a way that requires a stack location.
     DepField,           // It is a field of a dependently promoted struct
     NoRegVars,          // opts.compFlags & CLFLG_REGVAR is not set
-    MinOptsGC,          // It is a GC Ref and we are compiling MinOpts
 #if !defined(TARGET_64BIT)
     LongParamField, // It is a decomposed field of a long parameter.
 #endif
@@ -5525,7 +5524,7 @@ class Compiler
 
     void fgAddHandlerLiveVars(BasicBlock* block, VARSET_TP& ehHandlerLiveVars, MemoryKindSet& memoryLiveness);
 
-    void fgLiveVarAnalysis(bool updateInternalOnly = false);
+    void fgLiveVarAnalysis();
 
     void fgComputeLifeCall(VARSET_TP& life, GenTreeCall* call);
 
@@ -5545,10 +5544,10 @@ class Compiler
     void fgComputeLife(VARSET_TP&       life,
                        GenTree*         startNode,
                        GenTree*         endNode,
-                       VARSET_VALARG_TP volatileVars,
+                       VARSET_VALARG_TP keepAliveVars,
                        bool* pStmtInfoDirty DEBUGARG(bool* treeModf));
 
-    void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP volatileVars);
+    void fgComputeLifeLIR(VARSET_TP& life, BasicBlock* block, VARSET_VALARG_TP keepAliveVars);
 
     bool fgTryRemoveNonLocal(GenTree* node, LIR::Range* blockRange);
 
@@ -5916,8 +5915,6 @@ class Compiler
 
     PhaseStatus fgComputeDominators(); // Compute dominators
 
-    bool fgRemoveDeadBlocks(); // Identify and remove dead blocks.
-
 public:
     enum GCPollType
     {
@@ -6678,19 +6675,6 @@ class Compiler
 
     void fgMarkUseDef(GenTreeLclVarCommon* tree);
 
-    void fgBeginScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
-    void fgEndScopeLife(VARSET_TP* inScope, VarScopeDsc* var);
-
-    void fgMarkInScope(BasicBlock* block, VARSET_VALARG_TP inScope);
-    void fgUnmarkInScope(BasicBlock* block, VARSET_VALARG_TP unmarkScope);
-
-    void fgExtendDbgScopes();
-    void fgExtendDbgLifetimes();
-
-#ifdef DEBUG
-    void fgDispDebugScopes();
-#endif // DEBUG
-
     //-------------------------------------------------------------------------
     //
     //  The following keeps track of any code we've added for things like array
@@ -10891,7 +10875,6 @@ class Compiler
         unsigned m_liveInOutHndlr;
         unsigned m_depField;
         unsigned m_noRegVars;
-        unsigned m_minOptsGC;
 #ifdef JIT32_GCENCODER
         unsigned m_PinningRef;
 #endif // JIT32_GCENCODER

src/coreclr/jit/fgopt.cpp

@@ -162,111 +162,6 @@ bool Compiler::fgRemoveUnreachableBlocks(CanRemoveBlockBody canRemoveBlock)
     return changed;
 }
 
-//------------------------------------------------------------------------
-// fgRemoveDeadBlocks: Identify all the unreachable blocks and remove them.
-//
-bool Compiler::fgRemoveDeadBlocks()
-{
-    JITDUMP("\n*************** In fgRemoveDeadBlocks()");
-
-    unsigned prevFgCurBBEpoch = fgCurBBEpoch;
-    EnsureBasicBlockEpoch();
-
-    BlockSet visitedBlocks(BlockSetOps::MakeEmpty(this));
-
-    jitstd::list<BasicBlock*> worklist(jitstd::allocator<void>(getAllocator(CMK_Reachability)));
-    worklist.push_back(fgFirstBB);
-
-    // Visit all the reachable blocks, everything else can be removed
-    while (!worklist.empty())
-    {
-        BasicBlock* block = *(worklist.begin());
-        worklist.pop_front();
-
-        if (BlockSetOps::IsMember(this, visitedBlocks, block->bbNum))
-        {
-            continue;
-        }
-
-        BlockSetOps::AddElemD(this, visitedBlocks, block->bbNum);
-
-        for (BasicBlock* succ : block->Succs(this))
-        {
-            worklist.push_back(succ);
-        }
-
-        // Add all the "EH" successors. For every `try`, add its handler (including filter) to the worklist.
-        if (bbIsTryBeg(block))
-        {
-            // Due to EH normalization, a block can only be the start of a single `try` region, with the exception
-            // of mutually-protect regions.
-            assert(block->hasTryIndex());
-            unsigned  tryIndex = block->getTryIndex();
-            EHblkDsc* ehDsc    = ehGetDsc(tryIndex);
-            for (;;)
-            {
-                worklist.push_back(ehDsc->ebdHndBeg);
-                if (ehDsc->HasFilter())
-                {
-                    worklist.push_back(ehDsc->ebdFilter);
-                }
-                tryIndex = ehDsc->ebdEnclosingTryIndex;
-                if (tryIndex == EHblkDsc::NO_ENCLOSING_INDEX)
-                {
-                    break;
-                }
-                ehDsc = ehGetDsc(tryIndex);
-                if (ehDsc->ebdTryBeg != block)
-                {
-                    break;
-                }
-            }
-        }
-    }
-
-    // Track if there is any unreachable block. Even if it is marked with
-    // BBF_DONT_REMOVE, fgRemoveUnreachableBlocks() still removes the code
-    // inside the block. So this variable tracks if we ever found such blocks
-    // or not.
-    bool hasUnreachableBlock = false;
-
-    auto isBlockRemovable = [&](BasicBlock* block) -> bool {
-        const bool isVisited   = BlockSetOps::IsMember(this, visitedBlocks, block->bbNum);
-        const bool isRemovable = !isVisited || (block->bbRefs == 0);
-
-        hasUnreachableBlock |= isRemovable;
-        return isRemovable;
-    };
-
-    bool     changed        = false;
-    unsigned iterationCount = 1;
-    do
-    {
-        JITDUMP("\nRemoving unreachable blocks for fgRemoveDeadBlocks iteration #%u\n", iterationCount);
-
-        // Just to be paranoid, avoid infinite loops; fall back to minopts.
-        if (iterationCount++ > 10)
-        {
-            noway_assert(!"Too many unreachable block removal loops");
-        }
-        changed = fgRemoveUnreachableBlocks(isBlockRemovable);
-    } while (changed);
-
-#ifdef DEBUG
-    if (verbose && hasUnreachableBlock)
-    {
-        printf("\nAfter dead block removal:\n");
-        fgDispBasicBlocks(verboseTrees);
-        printf("\n");
-    }
-
-    fgVerifyHandlerTab();
-    fgDebugCheckBBlist(false);
-#endif // DEBUG
-
-    return hasUnreachableBlock;
-}
-
 //-------------------------------------------------------------
 // fgComputeDominators: Compute dominators
 //

src/coreclr/jit/gcencode.cpp

@@ -4065,9 +4065,13 @@ void GCInfo::gcMakeRegPtrTable(
 {
     GCENCODER_WITH_LOGGING(gcInfoEncoderWithLog, gcInfoEncoder);
 
-    const bool noTrackedGCSlots =
-        (compiler->opts.MinOpts() && !compiler->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT) &&
-         !JitConfig.JitMinOptsTrackGCrefs());
+    // TODO: Decide on whether we should enable this optimization for all
+    // targets: https://github.com/dotnet/runtime/issues/103917
+#ifdef TARGET_XARCH
+    const bool noTrackedGCSlots = compiler->opts.MinOpts() && !compiler->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT);
+#else
+    const bool noTrackedGCSlots = false;
+#endif
 
     if (mode == MAKE_REG_PTR_MODE_ASSIGN_SLOTS)
     {

src/coreclr/jit/jitconfigvalues.h

@@ -521,14 +521,6 @@ RELEASE_CONFIG_INTEGER(JitEnableNoWayAssert, W("JitEnableNoWayAssert"), 0)
 RELEASE_CONFIG_INTEGER(JitEnableNoWayAssert, W("JitEnableNoWayAssert"), 1)
 #endif // !defined(DEBUG) && !defined(_DEBUG)
 
-// Track GC roots
-#if defined(TARGET_AMD64) || defined(TARGET_X86)
-#define JitMinOptsTrackGCrefs_Default 0 // Not tracking GC refs in MinOpts is new behavior
-#else
-#define JitMinOptsTrackGCrefs_Default 1
-#endif
-RELEASE_CONFIG_INTEGER(JitMinOptsTrackGCrefs, W("JitMinOptsTrackGCrefs"), JitMinOptsTrackGCrefs_Default)
-
 // The following should be wrapped inside "#if MEASURE_MEM_ALLOC / #endif", but
 // some files include this one without bringing in the definitions from "jit.h"
 // so we don't always know what the "true" value of that flag should be. For now

src/coreclr/jit/lclvars.cpp

@@ -3212,10 +3212,6 @@ void Compiler::lvaSetVarDoNotEnregister(unsigned varNum DEBUGARG(DoNotEnregister
             JITDUMP("opts.compFlags & CLFLG_REGVAR is not set\n");
             assert(!compEnregLocals());
             break;
-        case DoNotEnregisterReason::MinOptsGC:
-            JITDUMP("it is a GC Ref and we are compiling MinOpts\n");
-            assert(!JitConfig.JitMinOptsTrackGCrefs() && varTypeIsGC(varDsc->TypeGet()));
-            break;
 #if !defined(TARGET_64BIT)
         case DoNotEnregisterReason::LongParamField:
             JITDUMP("it is a decomposed field of a long parameter\n");
@@ -4147,11 +4143,6 @@ void Compiler::lvaSortByRefCount()
             lvaSetVarDoNotEnregister(lclNum DEBUGARG(DoNotEnregisterReason::PinningRef));
 #endif
         }
-        if (opts.MinOpts() && !JitConfig.JitMinOptsTrackGCrefs() && varTypeIsGC(varDsc->TypeGet()))
-        {
-            varDsc->lvTracked = 0;
-            lvaSetVarDoNotEnregister(lclNum DEBUGARG(DoNotEnregisterReason::MinOptsGC));
-        }
         if (!compEnregLocals())
         {
             lvaSetVarDoNotEnregister(lclNum DEBUGARG(DoNotEnregisterReason::NoRegVars));