JIT: Add a head merging transformation alongside the tail merging tra…

…nsformation (#90468)

Add a pass that does head merging to compliment the existing tail
merging pass. Unlike tail merging this requires reordering the first
statement with the terminator node of the predecessor, which requires
some interference checking.

The head merging is enabled only for BBJ_COND since the CQ benefit
for enabling it more generally does not seem to pay off. Additionally
we mainly try to target the IR produced by the importer due to spill
cliques, which usually happen with the IL that results from using
ternaries.

Fix #90017

src/coreclr/jit/block.h

@@ -722,6 +722,11 @@ struct BasicBlock : private LIR::Range
         return KindIs(kind) || KindIs(rest...);
     }
 
+    bool HasTerminator()
+    {
+        return KindIs(BBJ_EHFINALLYRET, BBJ_EHFAULTRET, BBJ_EHFILTERRET, BBJ_COND, BBJ_SWITCH, BBJ_RETURN);
+    }
+
     // NumSucc() gives the number of successors, and GetSucc() returns a given numbered successor.
     //
     // There are two versions of these functions: ones that take a Compiler* and ones that don't. You must

src/coreclr/jit/compiler.cpp

@@ -4718,7 +4718,7 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
     {
         // Tail merge
         //
-        DoPhase(this, PHASE_TAIL_MERGE, &Compiler::fgTailMerge);
+        DoPhase(this, PHASE_HEAD_TAIL_MERGE, [this]() { return fgHeadTailMerge(true); });
 
         // Merge common throw blocks
         //
@@ -4835,7 +4835,7 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
 
         // Second pass of tail merge
         //
-        DoPhase(this, PHASE_TAIL_MERGE2, &Compiler::fgTailMerge);
+        DoPhase(this, PHASE_HEAD_TAIL_MERGE2, [this]() { return fgHeadTailMerge(false); });
 
         // Compute reachability sets and dominators.
         //

src/coreclr/jit/compiler.h

@@ -5519,7 +5519,11 @@ class Compiler
 
     void fgMoveBlocksAfter(BasicBlock* bStart, BasicBlock* bEnd, BasicBlock* insertAfterBlk);
 
-    PhaseStatus fgTailMerge();
+    PhaseStatus fgHeadTailMerge(bool early);
+    bool fgHeadMerge(BasicBlock* block, bool early);
+    bool fgTryOneHeadMerge(BasicBlock* block, bool early);
+    bool gtTreeContainsTailCall(GenTree* tree);
+    bool fgCanMoveFirstStatementIntoPred(bool early, Statement* firstStmt, BasicBlock* pred);
 
     enum FG_RELOCATE_TYPE
     {

src/coreclr/jit/compiler.hpp

@@ -676,6 +676,8 @@ inline bool BasicBlock::HasPotentialEHSuccs(Compiler* comp)
         return false;
     }
 
+    // Throwing in a filter is the same as returning "continue search", which
+    // causes enclosed finally/fault handlers to be executed.
     return hndDesc->InFilterRegionBBRange(this);
 }
 

src/coreclr/jit/compphases.h

@@ -55,10 +55,10 @@ CompPhaseNameMacro(PHASE_COMPUTE_EDGE_WEIGHTS,       "Compute edge weights (1, f
 #if defined(FEATURE_EH_FUNCLETS)
 CompPhaseNameMacro(PHASE_CREATE_FUNCLETS,            "Create EH funclets",             false, -1, false)
 #endif // FEATURE_EH_FUNCLETS
-CompPhaseNameMacro(PHASE_TAIL_MERGE,                 "Tail merge",                     false, -1, false)
+CompPhaseNameMacro(PHASE_HEAD_TAIL_MERGE,            "Head and tail merge",            false, -1, false)
 CompPhaseNameMacro(PHASE_MERGE_THROWS,               "Merge throw blocks",             false, -1, false)
 CompPhaseNameMacro(PHASE_INVERT_LOOPS,               "Invert loops",                   false, -1, false)
-CompPhaseNameMacro(PHASE_TAIL_MERGE2,                "Post-morph tail merge",          false, -1, false)
+CompPhaseNameMacro(PHASE_HEAD_TAIL_MERGE2,           "Post-morph head and tail merge", false, -1, false)
 CompPhaseNameMacro(PHASE_OPTIMIZE_FLOW,              "Optimize control flow",          false, -1, false)
 CompPhaseNameMacro(PHASE_OPTIMIZE_LAYOUT,            "Optimize layout",                false, -1, false)
 CompPhaseNameMacro(PHASE_COMPUTE_REACHABILITY,       "Compute blocks reachability",    false, -1, false)