arm64: Add SVE ptrue reuse pass after lowering

src/coreclr/jit/CMakeLists.txt

@@ -168,6 +168,7 @@ set( JIT_SOURCES
   optimizer.cpp
   patchpoint.cpp
   phase.cpp
+  constantmaskreuse.cpp
   promotion.cpp
   promotiondecomposition.cpp
   promotionliveness.cpp
@@ -338,6 +339,7 @@ set( JIT_HEADERS
   compilerbitsettraits.hpp
   compmemkind.h
   compphases.h
+  constantmaskreuse.h
   handlekinds.h
   dataflow.h
   debuginfo.h

src/coreclr/jit/compiler.cpp

@@ -5034,6 +5034,12 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
     m_pLowering = new (this, CMK_LSRA) Lowering(this, m_regAlloc); // PHASE_LOWERING
     m_pLowering->Run();
 
+#if defined(TARGET_ARM64) && defined(FEATURE_MASKED_HW_INTRINSICS)
+    DoPhase(this, PHASE_ARM64_CONSTANT_MASK_REUSE, &Compiler::fgOptimizeConstantMaskReuse);
+#endif
+
+    DoPhase(this, PHASE_POST_LOWERING, &Compiler::fgPostLowering);
+
     // Set stack levels and analyze throw helper usage.
     StackLevelSetter stackLevelSetter(this);
     stackLevelSetter.Run();
@@ -5177,6 +5183,56 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
 #endif // FUNC_INFO_LOGGING
 }
 
+//------------------------------------------------------------------------
+// fgPostLowering: Run cleanup that depends on lowered IR.
+//
+// Returns:
+//    PhaseStatus indicating whether the IR may have been changed.
+//
+PhaseStatus Compiler::fgPostLowering()
+{
+    // Recompute local var ref counts before potentially sorting for liveness.
+    // Note this does minimal work in cases where we are not going to sort.
+    const bool isRecompute    = true;
+    const bool setSlotNumbers = false;
+    lvaComputeRefCounts(isRecompute, setSlotNumbers);
+
+    if (m_dfsTree == nullptr)
+    {
+        // Compute DFS tree. We want to remove dead blocks even in MinOpts, so we
+        // do this everywhere. The dead blocks are removed below.
+        m_dfsTree = fgComputeDfs();
+    }
+
+    // Remove dead blocks. We want to remove unreachable blocks even in MinOpts.
+    fgRemoveBlocksOutsideDfsTree();
+
+    if (backendRequiresLocalVarLifetimes())
+    {
+        assert(opts.OptimizationEnabled());
+
+        fgPostLowerLiveness();
+        // local var liveness can delete code, which may create empty blocks
+        bool modified = fgUpdateFlowGraph(/* doTailDuplication */ false, /* isPhase */ false);
+
+        if (modified)
+        {
+            fgDfsBlocksAndRemove();
+            JITDUMP("had to run another liveness pass:\n");
+            fgPostLowerLiveness();
+        }
+
+        // Recompute local var ref counts again after liveness to reflect
+        // impact of any dead code removal. Note this may leave us with
+        // tracked vars that have zero refs.
+        lvaComputeRefCounts(isRecompute, setSlotNumbers);
+    }
+
+    fgInvalidateDfsTree();
+
+    return PhaseStatus::MODIFIED_EVERYTHING;
+}
+
 //----------------------------------------------------------------------------------------------
 // FinalizeEH: Finalize EH information
 //

src/coreclr/jit/compiler.h

@@ -7388,6 +7388,8 @@ class Compiler
     void fgExposeLocalsInBitVec(BitVec_ValArg_T bitVec);
 
     PhaseStatus fgOptimizeMaskConversions();
+    PhaseStatus fgOptimizeConstantMaskReuse();
+    PhaseStatus fgPostLowering();
 
     PhaseStatus PhysicalPromotion();
 

src/coreclr/jit/compphases.h

@@ -138,6 +138,8 @@ CompPhaseNameMacro(PHASE_LCLVARLIVENESS_INTERBLOCK,  "Global local var liveness"
 
 CompPhaseNameMacro(PHASE_LOWERING_DECOMP,            "Lowering decomposition",         false, -1, false)
 CompPhaseNameMacro(PHASE_LOWERING,                   "Lowering nodeinfo",              false, -1, true)
+CompPhaseNameMacro(PHASE_ARM64_CONSTANT_MASK_REUSE,  "Arm64 SVE constant mask reuse",  false, -1, true)
+CompPhaseNameMacro(PHASE_POST_LOWERING,              "Post-lowering cleanup",          false, -1, true)
 CompPhaseNameMacro(PHASE_STACK_LEVEL_SETTER,         "Calculate stack level slots",    false, -1, false)
 CompPhaseNameMacro(PHASE_LINEAR_SCAN,                "Linear scan register alloc",     true,  -1, true)
 CompPhaseNameMacro(PHASE_LINEAR_SCAN_BUILD,          "LSRA build intervals",           false, PHASE_LINEAR_SCAN, false)