JIT: some lclvars related cleanup (#19077)

Consolidate various compiler globals used when setting local var ref
counts by folding them into the visitor:
* lvaMarkRefsCurBlock
* lvaMarkRefsCurStmt
* lvaMarkRefsWeight

Remove the largely vestigial `lvPrefReg` and associated methods to set
or modify this field. Haven't verified but this is likely a remmant of
the legacy backend.

In the one remaning use (lcl var sorting predicates), swap in `lvIsRegArg`
instead, which gets most of the same cases.

7 files changed, 51 insertions, 285 deletions

diff --git a/src/jit/compiler.h b/src/jit/compiler.h
index 7a914d42af..e5bd4436a2 100644
--- a/src/jit/compiler.h
+++ b/src/jit/compiler.h
@@ -598,8 +598,6 @@ public:
         return regMask;
     }
 
-    regMaskSmall lvPrefReg; // set of regs it prefers to live in
-
     unsigned short lvVarIndex; // variable tracking index
 
 private:
@@ -2971,16 +2969,9 @@ public:
     //=========================================================================
 
 protected:
-//---------------- Local variable ref-counting ----------------------------
-
-#if ASSERTION_PROP
-    BasicBlock* lvaMarkRefsCurBlock;
-    GenTree*    lvaMarkRefsCurStmt;
-#endif
-    BasicBlock::weight_t lvaMarkRefsWeight;
-
-    void lvaMarkLclRefs(GenTree* tree);
+    //---------------- Local variable ref-counting ----------------------------
 
+    void lvaMarkLclRefs(GenTree* tree, BasicBlock* block, GenTreeStmt* stmt);
     bool IsDominatedByExceptionalEntry(BasicBlock* block);
     void SetVolatileHint(LclVarDsc* varDsc);
 
diff --git a/src/jit/compiler.hpp b/src/jit/compiler.hpp
index 0385e1035c..165255a639 100644
--- a/src/jit/compiler.hpp
+++ b/src/jit/compiler.hpp
@@ -1885,7 +1885,7 @@ inline void LclVarDsc::decRefCnts(BasicBlock::weight_t weight, Compiler* comp, R
         {
             for (unsigned i = lvFieldLclStart; i < lvFieldLclStart + lvFieldCnt; ++i)
             {
-                comp->lvaTable[i].decRefCnts(comp->lvaMarkRefsWeight, comp, state, false); // Don't propagate
+                comp->lvaTable[i].decRefCnts(weight, comp, state, false); // Don't propagate
             }
         }
     }
@@ -1898,7 +1898,7 @@ inline void LclVarDsc::decRefCnts(BasicBlock::weight_t weight, Compiler* comp, R
         assert(!parentvarDsc->lvRegStruct);
         if (promotionType == Compiler::PROMOTION_TYPE_DEPENDENT)
         {
-            parentvarDsc->decRefCnts(comp->lvaMarkRefsWeight, comp, state, false); // Don't propagate
+            parentvarDsc->decRefCnts(weight, comp, state, false); // Don't propagate
         }
     }
 
@@ -1942,7 +1942,7 @@ inline void LclVarDsc::incRefCnts(BasicBlock::weight_t weight, Compiler* comp, R
             setLvRefCnt((unsigned short)newRefCnt, state);
         }
 
-        // This fires when an uninitialize value for 'weight' is used (see lvaMarkRefsWeight)
+        // This fires when an uninitialize value for 'weight' is used
         assert(weight != 0xdddddddd);
         //
         // Increment lvRefCntWtd
@@ -1976,7 +1976,7 @@ inline void LclVarDsc::incRefCnts(BasicBlock::weight_t weight, Compiler* comp, R
         {
             for (unsigned i = lvFieldLclStart; i < lvFieldLclStart + lvFieldCnt; ++i)
             {
-                comp->lvaTable[i].incRefCnts(comp->lvaMarkRefsWeight, comp, state, false); // Don't propagate
+                comp->lvaTable[i].incRefCnts(weight, comp, state, false); // Don't propagate
             }
         }
     }
@@ -1989,7 +1989,7 @@ inline void LclVarDsc::incRefCnts(BasicBlock::weight_t weight, Compiler* comp, R
         assert(!parentvarDsc->lvRegStruct);
         if (promotionType == Compiler::PROMOTION_TYPE_DEPENDENT)
         {
-            parentvarDsc->incRefCnts(comp->lvaMarkRefsWeight, comp, state, false); // Don't propagate
+            parentvarDsc->incRefCnts(weight, comp, state, false); // Don't propagate
         }
     }
 
@@ -2008,108 +2008,6 @@ inline void LclVarDsc::incRefCnts(BasicBlock::weight_t weight, Compiler* comp, R
 
 /*****************************************************************************
  *
- *  Set the lvPrefReg field to reg
- */
-
-inline void LclVarDsc::setPrefReg(regNumber regNum, Compiler* comp)
-{
-    regMaskTP regMask;
-    if (isFloatRegType(TypeGet()))
-    {
-        // Check for FP struct-promoted field being passed in integer register
-        //
-        if (!genIsValidFloatReg(regNum))
-        {
-            return;
-        }
-        regMask = genRegMaskFloat(regNum, TypeGet());
-    }
-    else
-    {
-        regMask = genRegMask(regNum);
-    }
-
-#ifdef _TARGET_ARM_
-    // Don't set a preferred register for a TYP_STRUCT that takes more than one register slot
-    if ((TypeGet() == TYP_STRUCT) && (lvSize() > REGSIZE_BYTES))
-        return;
-#endif
-
-    /* Only interested if we have a new register bit set */
-    if (lvPrefReg & regMask)
-    {
-        return;
-    }
-
-#ifdef DEBUG
-    if (comp->verbose)
-    {
-        if (lvPrefReg)
-        {
-            printf("Change preferred register for V%02u from ", this - comp->lvaTable);
-            dspRegMask(lvPrefReg);
-        }
-        else
-        {
-            printf("Set preferred register for V%02u", this - comp->lvaTable);
-        }
-        printf(" to ");
-        dspRegMask(regMask);
-        printf("\n");
-    }
-#endif
-
-    /* Overwrite the lvPrefReg field */
-
-    lvPrefReg = (regMaskSmall)regMask;
-}
-
-/*****************************************************************************
- *
- *  Add regMask to the lvPrefReg field
- */
-
-inline void LclVarDsc::addPrefReg(regMaskTP regMask, Compiler* comp)
-{
-    assert(regMask != RBM_NONE);
-
-#ifdef _TARGET_ARM_
-    // Don't set a preferred register for a TYP_STRUCT that takes more than one register slot
-    if ((lvType == TYP_STRUCT) && (lvSize() > REGSIZE_BYTES))
-        return;
-#endif
-
-    /* Only interested if we have a new register bit set */
-    if (lvPrefReg & regMask)
-    {
-        return;
-    }
-
-#ifdef DEBUG
-    if (comp->verbose)
-    {
-        if (lvPrefReg)
-        {
-            printf("Additional preferred register for V%02u from ", this - comp->lvaTable);
-            dspRegMask(lvPrefReg);
-        }
-        else
-        {
-            printf("Set preferred register for V%02u", this - comp->lvaTable);
-        }
-        printf(" to ");
-        dspRegMask(lvPrefReg | regMask);
-        printf("\n");
-    }
-#endif
-
-    /* Update the lvPrefReg field */
-
-    lvPrefReg |= regMask;
-}
-
-/*****************************************************************************
- *
  *  The following returns the mask of all tracked locals
  *  referenced in a statement.
  */
diff --git a/src/jit/flowgraph.cpp b/src/jit/flowgraph.cpp
index ef2a4b3a37..72f67f0132 100644
--- a/src/jit/flowgraph.cpp
+++ b/src/jit/flowgraph.cpp
@@ -8465,17 +8465,6 @@ private:
                 comp->compFloatingPointUsed = true;
             }
 
-            if (!varTypeIsFloating(comp->info.compRetType))
-            {
-                returnLocalDsc.setPrefReg(REG_INTRET, comp);
-            }
-#ifdef REG_FLOATRET
-            else
-            {
-                returnLocalDsc.setPrefReg(REG_FLOATRET, comp);
-            }
-#endif
-
 #ifdef DEBUG
             // This temporary should not be converted to a double in stress mode,
             // because we introduce assigns to it after the stress conversion
@@ -8907,8 +8896,7 @@ void Compiler::fgAddInternal()
         lvaInlinedPInvokeFrameVar = lvaGrabTempWithImplicitUse(false DEBUGARG("Pinvoke FrameVar"));
 
         LclVarDsc* varDsc = &lvaTable[lvaInlinedPInvokeFrameVar];
-        varDsc->addPrefReg(RBM_PINVOKE_TCB, this);
-        varDsc->lvType = TYP_BLK;
+        varDsc->lvType    = TYP_BLK;
         // Make room for the inlined frame.
         varDsc->lvExactSize = eeGetEEInfo()->inlinedCallFrameInfo.size;
 #if FEATURE_FIXED_OUT_ARGS
diff --git a/src/jit/lclvars.cpp b/src/jit/lclvars.cpp
index ea1976008a..9e83d11776 100644
--- a/src/jit/lclvars.cpp
+++ b/src/jit/lclvars.cpp
@@ -441,7 +441,6 @@ void Compiler::lvaInitThisPtr(InitVarDscInfo* varDscInfo)
 #if FEATURE_MULTIREG_ARGS
         varDsc->lvOtherArgReg = REG_NA;
 #endif
-        varDsc->setPrefReg(varDsc->lvArgReg, this);
         varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame
 
 #ifdef DEBUG
@@ -488,7 +487,6 @@ void Compiler::lvaInitRetBuffArg(InitVarDscInfo* varDscInfo)
 #if FEATURE_MULTIREG_ARGS
         varDsc->lvOtherArgReg = REG_NA;
 #endif
-        varDsc->setPrefReg(varDsc->lvArgReg, this);
         varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame
 
         info.compRetBuffDefStack = 0;
@@ -842,7 +840,6 @@ void Compiler::lvaInitUserArgs(InitVarDscInfo* varDscInfo)
                 if (secondEightByteType != TYP_UNDEF)
                 {
                     varDsc->lvOtherArgReg = genMapRegArgNumToRegNum(secondAllocatedRegArgNum, secondEightByteType);
-                    varDsc->addPrefReg(genRegMask(varDsc->lvOtherArgReg), this);
                 }
 #else // ARM32 or ARM64
                 varDsc->lvArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum, TYP_I_IMPL);
@@ -850,7 +847,6 @@ void Compiler::lvaInitUserArgs(InitVarDscInfo* varDscInfo)
                 if (cSlots == 2)
                 {
                     varDsc->lvOtherArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum + 1, TYP_I_IMPL);
-                    varDsc->addPrefReg(genRegMask(varDsc->lvOtherArgReg), this);
                 }
 #endif //  _TARGET_ARM64_
 #endif // defined(UNIX_AMD64_ABI)
@@ -861,13 +857,10 @@ void Compiler::lvaInitUserArgs(InitVarDscInfo* varDscInfo)
                 varDsc->lvArgReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum, argType);
             }
 
-            varDsc->setPrefReg(varDsc->lvArgReg, this);
-
 #ifdef _TARGET_ARM_
             if (varDsc->TypeGet() == TYP_LONG)
             {
                 varDsc->lvOtherReg = genMapRegArgNumToRegNum(firstAllocatedRegArgNum + 1, TYP_INT);
-                varDsc->addPrefReg(genRegMask(varDsc->lvOtherReg), this);
             }
 #endif // _TARGET_ARM_
 
@@ -1065,7 +1058,6 @@ void Compiler::lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo)
 #if FEATURE_MULTIREG_ARGS
             varDsc->lvOtherArgReg = REG_NA;
 #endif
-            varDsc->setPrefReg(varDsc->lvArgReg, this);
             varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame
 
             varDscInfo->intRegArgNum++;
@@ -1132,7 +1124,6 @@ void Compiler::lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo)
 #if FEATURE_MULTIREG_ARGS
             varDsc->lvOtherArgReg = REG_NA;
 #endif
-            varDsc->setPrefReg(varDsc->lvArgReg, this);
             varDsc->lvOnFrame = true; // The final home for this incoming register might be our local stack frame
 #ifdef _TARGET_ARM_
             // This has to be spilled right in front of the real arguments and we have
@@ -2011,7 +2002,6 @@ void Compiler::lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* Stru
         {
             fieldVarDsc->lvIsRegArg = true;
             fieldVarDsc->lvArgReg   = varDsc->lvArgReg;
-            fieldVarDsc->setPrefReg(varDsc->lvArgReg, this); // Set the preferred register
 #if FEATURE_MULTIREG_ARGS && defined(FEATURE_SIMD)
             if (varTypeIsSIMD(fieldVarDsc) && !lvaIsImplicitByRefLocal(lclNum))
             {
@@ -2024,7 +2014,6 @@ void Compiler::lvaPromoteStructVar(unsigned lclNum, lvaStructPromotionInfo* Stru
             }
 #endif // FEATURE_MULTIREG_ARGS && defined(FEATURE_SIMD)
 
-            lvaMarkRefsWeight = BB_UNITY_WEIGHT; // incRefCnts can use this compiler global variable
             fieldVarDsc->incRefCnts(BB_UNITY_WEIGHT, this,
                                     RCS_EARLY); // increment the ref count for prolog initialization
         }
@@ -2989,7 +2978,6 @@ void Compiler::lvaIncRefCnts(GenTree* tree)
  *    Return positive if dsc2 has a higher ref count
  *    Return negative if dsc1 has a higher ref count
  *    Return zero     if the ref counts are the same
- *    lvPrefReg is only used to break ties
  */
 
 /* static */
@@ -3049,24 +3037,17 @@ int __cdecl Compiler::RefCntCmp(const void* op1, const void* op2)
     /* We have equal ref counts and weighted ref counts */
 
     /* Break the tie by: */
-    /* Increasing the weight by 2   if we have exactly one bit set in lvPrefReg   */
-    /* Increasing the weight by 1   if we have more than one bit set in lvPrefReg */
+    /* Increasing the weight by 2   if we are a register arg */
     /* Increasing the weight by 0.5 if we are a GC type */
     /* Increasing the weight by 0.5 if we were enregistered in the previous pass  */
 
     if (weight1)
     {
-        if (dsc1->lvPrefReg)
+        if (dsc1->lvIsRegArg)
         {
-            if ((dsc1->lvPrefReg & ~RBM_BYTE_REG_FLAG) && genMaxOneBit((unsigned)dsc1->lvPrefReg))
-            {
-                weight1 += 2 * BB_UNITY_WEIGHT;
-            }
-            else
-            {
-                weight1 += 1 * BB_UNITY_WEIGHT;
-            }
+            weight2 += 2 * BB_UNITY_WEIGHT;
         }
+
         if (varTypeIsGC(dsc1->TypeGet()))
         {
             weight1 += BB_UNITY_WEIGHT / 2;
@@ -3080,17 +3061,11 @@ int __cdecl Compiler::RefCntCmp(const void* op1, const void* op2)
 
     if (weight2)
     {
-        if (dsc2->lvPrefReg)
+        if (dsc2->lvIsRegArg)
         {
-            if ((dsc2->lvPrefReg & ~RBM_BYTE_REG_FLAG) && genMaxOneBit((unsigned)dsc2->lvPrefReg))
-            {
-                weight2 += 2 * BB_UNITY_WEIGHT;
-            }
-            else
-            {
-                weight2 += 1 * BB_UNITY_WEIGHT;
-            }
+            weight2 += 2 * BB_UNITY_WEIGHT;
         }
+
         if (varTypeIsGC(dsc2->TypeGet()))
         {
             weight2 += BB_UNITY_WEIGHT / 2;
@@ -3170,31 +3145,14 @@ int __cdecl Compiler::WtdRefCntCmp(const void* op1, const void* op2)
     }
 #endif
 
-    /* Increase the weight by 2 if we have exactly one bit set in lvPrefReg */
-    /* Increase the weight by 1 if we have more than one bit set in lvPrefReg */
-
-    if (weight1 && dsc1->lvPrefReg)
+    if (weight1 && dsc1->lvIsRegArg)
     {
-        if ((dsc1->lvPrefReg & ~RBM_BYTE_REG_FLAG) && genMaxOneBit((unsigned)dsc1->lvPrefReg))
-        {
-            weight1 += 2 * BB_UNITY_WEIGHT;
-        }
-        else
-        {
-            weight1 += 1 * BB_UNITY_WEIGHT;
-        }
+        weight1 += 2 * BB_UNITY_WEIGHT;
     }
 
-    if (weight2 && dsc2->lvPrefReg)
+    if (weight2 && dsc2->lvIsRegArg)
     {
-        if ((dsc2->lvPrefReg & ~RBM_BYTE_REG_FLAG) && genMaxOneBit((unsigned)dsc2->lvPrefReg))
-        {
-            weight2 += 2 * BB_UNITY_WEIGHT;
-        }
-        else
-        {
-            weight2 += 1 * BB_UNITY_WEIGHT;
-        }
+        weight2 += 2 * BB_UNITY_WEIGHT;
     }
 
     if (weight2 > weight1)
@@ -3299,13 +3257,6 @@ void Compiler::lvaDumpRefCounts()
             gtDispLclVar((unsigned)(lvaRefSorted[lclNum] - lvaTable));
             printf(" [%6s]: refCnt = %4u, refCntWtd = %6s", varTypeName(lvaRefSorted[lclNum]->TypeGet()), refCnt,
                    refCntWtd2str(refCntWtd));
-
-            regMaskSmall pref = lvaRefSorted[lclNum]->lvPrefReg;
-            if (pref)
-            {
-                printf(" pref ");
-                dspRegMask(pref);
-            }
             printf("\n");
         }
 
@@ -3670,13 +3621,21 @@ var_types LclVarDsc::lvaArgType()
     return type;
 }
 
-/*****************************************************************************
- *
- *  This is called by lvaMarkLclRefsCallback() to do variable ref marking
- */
+//------------------------------------------------------------------------
+// lvaMarkLclRefs: increment local var references counts and more
+//
+// Arguments:
+//     tree - some node in a tree
+//     block - block that the tree node belongs to
+//     stmt - stmt that the tree node belongs to
+//
+// Notes:
+//     Invoked via the MarkLocalVarsVisitor
 
-void Compiler::lvaMarkLclRefs(GenTree* tree)
+void Compiler::lvaMarkLclRefs(GenTree* tree, BasicBlock* block, GenTreeStmt* stmt)
 {
+    const BasicBlock::weight_t weight = block->getBBWeight(this);
+
     /* Is this a call to unmanaged code ? */
     if (tree->gtOper == GT_CALL && tree->gtFlags & GTF_CALL_UNMANAGED)
     {
@@ -3691,8 +3650,8 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
             LclVarDsc* varDsc = lvaTable + lclNum;
 
             /* Increment the ref counts twice */
-            varDsc->incRefCnts(lvaMarkRefsWeight, this);
-            varDsc->incRefCnts(lvaMarkRefsWeight, this);
+            varDsc->incRefCnts(weight, this);
+            varDsc->incRefCnts(weight, this);
         }
     }
 
@@ -3703,25 +3662,6 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
         GenTree* op1 = tree->gtOp.gtOp1;
         GenTree* op2 = tree->gtOp.gtOp2;
 
-        /* Set target register for RHS local if assignment is of a "small" type */
-
-        if (varTypeIsByte(tree->gtType))
-        {
-            unsigned   lclNum;
-            LclVarDsc* varDsc = nullptr;
-
-            if (op2->gtOper == GT_LCL_VAR)
-            {
-                lclNum = op2->gtLclVarCommon.gtLclNum;
-                noway_assert(lclNum < lvaCount);
-                varDsc = &lvaTable[lclNum];
-            }
-#if CPU_HAS_BYTE_REGS
-            if (varDsc)
-                varDsc->addPrefReg(RBM_BYTE_REG_FLAG, this);
-#endif
-        }
-
 #if OPT_BOOL_OPS
 
         /* Is this an assignment to a local variable? */
@@ -3774,27 +3714,6 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
 #endif
     }
 
-#ifdef _TARGET_XARCH_
-    /* Special case: integer shift node by a variable amount */
-
-    if (tree->OperIsShiftOrRotate())
-    {
-        if (tree->gtType == TYP_INT)
-        {
-            GenTree* op2 = tree->gtOp.gtOp2;
-
-            if (op2->gtOper == GT_LCL_VAR)
-            {
-                unsigned lclNum = op2->gtLclVarCommon.gtLclNum;
-                assert(lclNum < lvaCount);
-                lvaTable[lclNum].setPrefReg(REG_ECX, this);
-            }
-        }
-
-        return;
-    }
-#endif
-
     if ((tree->gtOper != GT_LCL_VAR) && (tree->gtOper != GT_LCL_FLD))
     {
         return;
@@ -3810,7 +3729,7 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
 
     /* Increment the reference counts */
 
-    varDsc->incRefCnts(lvaMarkRefsWeight, this);
+    varDsc->incRefCnts(weight, this);
 
     if (lvaVarAddrExposed(lclNum))
     {
@@ -3828,7 +3747,7 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
     }
 
 #if ASSERTION_PROP
-    if (fgDomsComputed && IsDominatedByExceptionalEntry(lvaMarkRefsCurBlock))
+    if (fgDomsComputed && IsDominatedByExceptionalEntry(block))
     {
         SetVolatileHint(varDsc);
     }
@@ -3857,7 +3776,7 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
             else
             {
                 varDsc->lvSingleDef = true;
-                varDsc->lvDefStmt   = lvaMarkRefsCurStmt;
+                varDsc->lvDefStmt   = stmt;
             }
         }
         else // otherwise this is a ref of our variable
@@ -3867,7 +3786,7 @@ void Compiler::lvaMarkLclRefs(GenTree* tree)
                 // Lazy initialization
                 BlockSetOps::AssignNoCopy(this, varDsc->lvRefBlks, BlockSetOps::MakeEmpty(this));
             }
-            BlockSetOps::AddElemD(this, varDsc->lvRefBlks, lvaMarkRefsCurBlock->bbNum);
+            BlockSetOps::AddElemD(this, varDsc->lvRefBlks, block->bbNum);
         }
     }
 #endif // ASSERTION_PROP
@@ -3947,53 +3866,36 @@ void Compiler::lvaMarkLocalVars(BasicBlock* block)
 {
     class MarkLocalVarsVisitor final : public GenTreeVisitor<MarkLocalVarsVisitor>
     {
+    private:
+        BasicBlock*  m_block;
+        GenTreeStmt* m_stmt;
+
     public:
         enum
         {
             DoPreOrder = true,
         };
 
-        MarkLocalVarsVisitor(Compiler* compiler) : GenTreeVisitor<MarkLocalVarsVisitor>(compiler)
+        MarkLocalVarsVisitor(Compiler* compiler, BasicBlock* block, GenTreeStmt* stmt)
+            : GenTreeVisitor<MarkLocalVarsVisitor>(compiler), m_block(block), m_stmt(stmt)
         {
         }
 
         Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
         {
-            m_compiler->lvaMarkLclRefs(*use);
+            m_compiler->lvaMarkLclRefs(*use, m_block, m_stmt);
             return WALK_CONTINUE;
         }
     };
 
-#if ASSERTION_PROP
-    lvaMarkRefsCurBlock = block;
-#endif
-    lvaMarkRefsWeight = block->getBBWeight(this);
-
-#ifdef DEBUG
-    if (verbose)
-    {
-        printf("\n*** marking local variables in block BB%02u (weight=%s)\n", block->bbNum,
-               refCntWtd2str(lvaMarkRefsWeight));
-    }
-#endif
+    JITDUMP("\n*** marking local variables in block BB%02u (weight=%s)\n", block->bbNum,
+            refCntWtd2str(block->getBBWeight(this)));
 
-    MarkLocalVarsVisitor visitor(this);
-    for (GenTree* tree = block->FirstNonPhiDef(); tree; tree = tree->gtNext)
+    for (GenTreeStmt* stmt = block->FirstNonPhiDef(); stmt != nullptr; stmt = stmt->getNextStmt())
     {
-        assert(tree->gtOper == GT_STMT);
-
-#if ASSERTION_PROP
-        lvaMarkRefsCurStmt = tree;
-#endif
-
-#ifdef DEBUG
-        if (verbose)
-        {
-            gtDispTree(tree);
-        }
-#endif
-
-        visitor.WalkTree(&tree->gtStmt.gtStmtExpr, nullptr);
+        MarkLocalVarsVisitor visitor(this, block, stmt);
+        DISPTREE(stmt);
+        visitor.WalkTree(&stmt->gtStmtExpr, nullptr);
     }
 }
 
diff --git a/src/jit/morph.cpp b/src/jit/morph.cpp
index 44034e7336..500bad4269 100644
--- a/src/jit/morph.cpp
+++ b/src/jit/morph.cpp
@@ -17735,7 +17735,6 @@ void Compiler::fgRetypeImplicitByRefArgs()
 #if FEATURE_MULTIREG_ARGS
                     fieldVarDsc->lvOtherArgReg = REG_NA;
 #endif
-                    fieldVarDsc->lvPrefReg = 0;
                 }
 
                 // Hijack lvFieldLclStart to record the new temp number.
diff --git a/src/jit/target.h b/src/jit/target.h
index 3cbdca0f56..d3520d0220 100644
--- a/src/jit/target.h
+++ b/src/jit/target.h
@@ -1588,12 +1588,9 @@ C_ASSERT((FEATURE_TAILCALL_OPT == 0) || (FEATURE_FASTTAILCALL == 1));
 #if CPU_HAS_BYTE_REGS
   #define RBM_BYTE_REGS           (RBM_EAX|RBM_ECX|RBM_EDX|RBM_EBX)
   #define RBM_NON_BYTE_REGS       (RBM_ESI|RBM_EDI)
-  // We reuse the ESP register as a flag for byteable registers in lvPrefReg
-  #define RBM_BYTE_REG_FLAG        RBM_ESP
 #else
   #define RBM_BYTE_REGS            RBM_ALLINT
   #define RBM_NON_BYTE_REGS        RBM_NONE
-  #define RBM_BYTE_REG_FLAG        RBM_NONE
 #endif
 // clang-format on
 
diff --git a/src/jit/utils.cpp b/src/jit/utils.cpp
index 2b120dd347..b73f79f447 100644
--- a/src/jit/utils.cpp
+++ b/src/jit/utils.cpp
@@ -380,15 +380,6 @@ void dspRegMask(regMaskTP regMask, size_t minSiz)
         regPrev = regNum;
     }
 
-#if CPU_HAS_BYTE_REGS
-    if (regMask & RBM_BYTE_REG_FLAG)
-    {
-        const char* nam = "BYTE";
-        printf("%s%s", sep, nam);
-        minSiz -= (strlen(sep) + strlen(nam));
-    }
-#endif
-
     if (strlen(sep) > 0)
     {
         // We've already printed something.