Enregister structs on win x64. (#55045)

* Enreg structs x64 windows.

* try to get zero diffs on other platforms.

* fix comment

src/coreclr/jit/codegencommon.cpp

@@ -4355,16 +4355,10 @@ void CodeGen::genFnPrologCalleeRegArgs(regNumber xtraReg, bool* pXtraRegClobbere
 #endif // defined(UNIX_AMD64_ABI)
 
             noway_assert(varDsc->lvIsParam && varDsc->lvIsRegArg);
-#ifndef TARGET_64BIT
-#ifndef TARGET_ARM
-            // Right now we think that incoming arguments are not pointer sized.  When we eventually
-            // understand the calling convention, this still won't be true. But maybe we'll have a better
-            // idea of how to ignore it.
-
-            // On Arm, a long can be passed in register
-            noway_assert(genTypeSize(genActualType(varDsc->TypeGet())) == TARGET_POINTER_SIZE);
-#endif
-#endif // TARGET_64BIT
+#ifdef TARGET_X86
+            // On x86 we don't enregister args that are not pointer sized.
+            noway_assert(genTypeSize(varDsc->GetActualRegisterType()) == TARGET_POINTER_SIZE);
+#endif // TARGET_X86
 
             noway_assert(varDsc->lvIsInReg() && !regArgTab[argNum].circular);
 

src/coreclr/jit/codegenlinear.cpp

@@ -1191,12 +1191,12 @@ void CodeGen::genUnspillRegIfNeeded(GenTree* tree)
             assert(spillType != TYP_UNDEF);
 
 // TODO-Cleanup: The following code could probably be further merged and cleaned up.
-#ifdef TARGET_XARCH
+#if defined(TARGET_XARCH) || defined(TARGET_ARM64)
             // Load local variable from its home location.
             // In most cases the tree type will indicate the correct type to use for the load.
             // However, if it is NOT a normalizeOnLoad lclVar (i.e. NOT a small int that always gets
-            // widened when loaded into a register), and its size is not the same as genActualType of
-            // the type of the lclVar, then we need to change the type of the tree node when loading.
+            // widened when loaded into a register), and its size is not the same as the actual register type
+            // of the lclVar, then we need to change the type of the tree node when loading.
             // This situation happens due to "optimizations" that avoid a cast and
             // simply retype the node when using long type lclVar as an int.
             // While loading the int in that case would work for this use of the lclVar, if it is
@@ -1210,13 +1210,6 @@ void CodeGen::genUnspillRegIfNeeded(GenTree* tree)
                 assert(!varTypeIsGC(varDsc));
                 spillType = lclActualType;
             }
-#elif defined(TARGET_ARM64)
-            var_types targetType = unspillTree->gtType;
-            if (spillType != genActualType(varDsc->lvType) && !varTypeIsGC(spillType) && !varDsc->lvNormalizeOnLoad())
-            {
-                assert(!varTypeIsGC(varDsc));
-                spillType = genActualType(varDsc->lvType);
-            }
 #elif defined(TARGET_ARM)
 // No normalizing for ARM
 #else
@@ -1465,7 +1458,8 @@ regNumber CodeGen::genConsumeReg(GenTree* tree)
         LclVarDsc*           varDsc = &compiler->lvaTable[lcl->GetLclNum()];
         if (varDsc->GetRegNum() != REG_STK)
         {
-            inst_Mov(tree->TypeGet(), tree->GetRegNum(), varDsc->GetRegNum(), /* canSkip */ true);
+            var_types regType = varDsc->GetRegisterType(lcl);
+            inst_Mov(regType, tree->GetRegNum(), varDsc->GetRegNum(), /* canSkip */ true);
         }
     }
 

src/coreclr/jit/compiler.h

@@ -1016,11 +1016,20 @@ class LclVarDsc
 
     var_types GetActualRegisterType() const;
 
-    bool IsEnregisterable() const
+    bool IsEnregisterableType() const
     {
         return GetRegisterType() != TYP_UNDEF;
     }
 
+    bool IsEnregisterableLcl() const
+    {
+        if (lvDoNotEnregister)
+        {
+            return false;
+        }
+        return IsEnregisterableType();
+    }
+
     bool CanBeReplacedWithItsField(Compiler* comp) const;
 
 #ifdef DEBUG

src/coreclr/jit/jitconfigvalues.h

@@ -552,8 +552,12 @@ CONFIG_INTEGER(JitSaveFpLrWithCalleeSavedRegisters, W("JitSaveFpLrWithCalleeSave
 #endif // defined(TARGET_ARM64)
 #endif // DEBUG
 
-// Allow to enregister locals with struct type.
-CONFIG_INTEGER(JitEnregStructLocals, W("JitEnregStructLocals"), 0)
+#if defined(TARGET_AMD64) && defined(TARGET_WINDOWS)
+CONFIG_INTEGER(JitEnregStructLocals, W("JitEnregStructLocals"), 1) // Allow to enregister locals with struct type.
+#else
+CONFIG_INTEGER(JitEnregStructLocals, W("JitEnregStructLocals"), 0) // Don't allow to enregister locals with struct type
+                                                                   // yet.
+#endif
 
 #undef CONFIG_INTEGER
 #undef CONFIG_STRING

src/coreclr/jit/lclvars.cpp

@@ -3493,7 +3493,7 @@ void Compiler::lvaSortByRefCount()
             {
                 varDsc->lvTracked = 0;
             }
-            else if ((varDsc->lvType == TYP_STRUCT) && !varDsc->lvRegStruct)
+            else if ((varDsc->lvType == TYP_STRUCT) && !varDsc->lvRegStruct && !compEnregStructLocals())
             {
                 lvaSetVarDoNotEnregister(lclNum DEBUGARG(DNER_IsStruct));
             }

src/coreclr/jit/lower.cpp

@@ -3160,7 +3160,7 @@ void Lowering::LowerStoreLocCommon(GenTreeLclVarCommon* lclStore)
 #endif // !WINDOWS_AMD64_ABI
             convertToStoreObj = false;
         }
-        else if (!varDsc->IsEnregisterable())
+        else if (!varDsc->IsEnregisterableType())
         {
             convertToStoreObj = true;
         }
@@ -3418,11 +3418,10 @@ void Lowering::LowerRetSingleRegStructLclVar(GenTreeUnOp* ret)
         lclNum = fieldLclNum;
         varDsc = comp->lvaGetDesc(lclNum);
     }
-    else if (!varDsc->lvRegStruct && !varTypeIsEnregisterable(varDsc))
-
+    else if (varDsc->lvPromoted)
     {
-        // TODO-1stClassStructs: We can no longer promote or enregister this struct,
-        // since it is referenced as a whole.
+        // TODO-1stClassStructs: We can no longer independently promote
+        // or enregister this struct, since it is referenced as a whole.
         comp->lvaSetVarDoNotEnregister(lclNum DEBUGARG(Compiler::DNER_BlockOp));
     }
 
@@ -3434,9 +3433,10 @@ void Lowering::LowerRetSingleRegStructLclVar(GenTreeUnOp* ret)
     }
     else
     {
-        var_types lclVarType = varDsc->GetRegisterType(lclVar);
+        const var_types lclVarType = varDsc->GetRegisterType(lclVar);
         assert(lclVarType != TYP_UNDEF);
-        lclVar->ChangeType(lclVarType);
+        const var_types actualType = genActualType(lclVarType);
+        lclVar->ChangeType(actualType);
 
         if (varTypeUsesFloatReg(ret) != varTypeUsesFloatReg(lclVarType))
         {
@@ -4039,12 +4039,12 @@ void Lowering::InsertPInvokeMethodProlog()
     GenTree* call = comp->gtNewHelperCallNode(CORINFO_HELP_INIT_PINVOKE_FRAME, TYP_I_IMPL, argList);
 
     // some sanity checks on the frame list root vardsc
-    LclVarDsc* varDsc = &comp->lvaTable[comp->info.compLvFrameListRoot];
+    const unsigned   lclNum = comp->info.compLvFrameListRoot;
+    const LclVarDsc* varDsc = comp->lvaGetDesc(lclNum);
     noway_assert(!varDsc->lvIsParam);
     noway_assert(varDsc->lvType == TYP_I_IMPL);
 
-    GenTree* store =
-        new (comp, GT_STORE_LCL_VAR) GenTreeLclVar(GT_STORE_LCL_VAR, TYP_I_IMPL, comp->info.compLvFrameListRoot);
+    GenTree* store       = new (comp, GT_STORE_LCL_VAR) GenTreeLclVar(GT_STORE_LCL_VAR, TYP_I_IMPL, lclNum);
     store->AsOp()->gtOp1 = call;
     store->gtFlags |= GTF_VAR_DEF;
 
@@ -4065,6 +4065,7 @@ void Lowering::InsertPInvokeMethodProlog()
         GenTreeLclFld(GT_STORE_LCL_FLD, TYP_I_IMPL, comp->lvaInlinedPInvokeFrameVar, callFrameInfo.offsetOfCallSiteSP);
     storeSP->gtOp1 = PhysReg(REG_SPBASE);
     storeSP->gtFlags |= GTF_VAR_DEF;
+    comp->lvaSetVarDoNotEnregister(comp->lvaInlinedPInvokeFrameVar DEBUGARG(Compiler::DNER_LocalField));
 
     firstBlockRange.InsertBefore(insertionPoint, LIR::SeqTree(comp, storeSP));
     DISPTREERANGE(firstBlockRange, storeSP);
@@ -6534,7 +6535,7 @@ void Lowering::ContainCheckRet(GenTreeUnOp* ret)
             assert(varDsc->lvIsMultiRegRet || (varDsc->lvIsHfa() && varTypeIsValidHfaType(varDsc->lvType)));
 
             // Mark var as contained if not enregisterable.
-            if (!varTypeIsEnregisterable(op1))
+            if (!varDsc->IsEnregisterableLcl())
             {
                 if (!op1->IsMultiRegLclVar())
                 {

src/coreclr/jit/lowerxarch.cpp

@@ -269,6 +269,12 @@ void Lowering::LowerBlockStore(GenTreeBlk* blkNode)
             // address, not knowing that GT_IND is part of a block op that has containment restrictions.
             src->AsIndir()->Addr()->ClearContained();
         }
+        else if (src->OperIs(GT_LCL_VAR))
+        {
+            // TODO-1stClassStructs: for now we can't work with STORE_BLOCK source in register.
+            const unsigned srcLclNum = src->AsLclVar()->GetLclNum();
+            comp->lvaSetVarDoNotEnregister(srcLclNum DEBUGARG(Compiler::DNER_BlockOp));
+        }
 
         if (blkNode->OperIs(GT_STORE_OBJ))
         {

src/coreclr/jit/lsra.cpp

@@ -1490,7 +1490,7 @@ bool LinearScan::isRegCandidate(LclVarDsc* varDsc)
     // or enregistered, on x86 -- it is believed that we can enregister pinned (more properly, "pinning")
     // references when using the general GC encoding.
     unsigned lclNum = (unsigned)(varDsc - compiler->lvaTable);
-    if (varDsc->lvAddrExposed || !varDsc->IsEnregisterable() ||
+    if (varDsc->lvAddrExposed || !varDsc->IsEnregisterableType() ||
         (!compiler->compEnregStructLocals() && (varDsc->lvType == TYP_STRUCT)))
     {
 #ifdef DEBUG