JIT: Enable inlining for late devirtualization

src/coreclr/jit/compiler.h

@@ -5109,6 +5109,8 @@ class Compiler
         SpillCliqueSucc
     };
 
+    friend class SubstitutePlaceholdersAndDevirtualizeWalker;
+
     // Abstract class for receiving a callback while walking a spill clique
     class SpillCliqueWalker
     {

src/coreclr/jit/fginline.cpp

@@ -204,7 +204,8 @@ bool Compiler::TypeInstantiationComplexityExceeds(CORINFO_CLASS_HANDLE handle, i
 
 class SubstitutePlaceholdersAndDevirtualizeWalker : public GenTreeVisitor<SubstitutePlaceholdersAndDevirtualizeWalker>
 {
-    bool m_madeChanges = false;
+    bool       m_madeChanges = false;
+    Statement* m_curStmt     = nullptr;
 
 public:
     enum
@@ -219,11 +220,28 @@ class SubstitutePlaceholdersAndDevirtualizeWalker : public GenTreeVisitor<Substi
     {
     }
 
-    bool MadeChanges()
+    bool MadeChanges() const
     {
         return m_madeChanges;
     }
 
+    // ------------------------------------------------------------------------
+    // WalkStatement: Walk the tree of a statement, and return the newly added
+    // statement if any, otherwise return the original statement.
+    //
+    // Arguments:
+    //    stmt - the statement to walk.
+    //
+    // Return Value:
+    //    The newly added statement if any, or the original statement.
+    //
+    Statement* WalkStatement(Statement* stmt)
+    {
+        m_curStmt = stmt;
+        WalkTree(m_curStmt->GetRootNodePointer(), nullptr);
+        return m_curStmt;
+    }
+
     fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
     {
         GenTree* tree = *use;
@@ -586,8 +604,52 @@ class SubstitutePlaceholdersAndDevirtualizeWalker : public GenTreeVisitor<Substi
                 }
 
                 CORINFO_CONTEXT_HANDLE contextInput = context;
+                context                             = nullptr;
                 m_compiler->impDevirtualizeCall(call, nullptr, &method, &methodFlags, &contextInput, &context,
                                                 isLateDevirtualization, explicitTailCall);
+                if (context != nullptr)
+                {
+#if defined(DEBUG)
+                    IL_OFFSET ilOffset = call->gtRawILOffset;
+#else
+                    IL_OFFSET ilOffset = m_curStmt->GetDebugInfo().GetLocation().GetOffset();
+#endif // defined(DEBUG)
+                    CORINFO_CALL_INFO callInfo = {};
+                    callInfo.hMethod           = method;
+                    callInfo.methodFlags       = methodFlags;
+                    m_compiler->impMarkInlineCandidate(call, context, false, &callInfo, ilOffset);
+
+                    if (call->IsInlineCandidate())
+                    {
+                        CallArg* retBuffer = call->gtArgs.GetRetBufferArg();
+
+                        if (parent != nullptr || genActualType(call->TypeGet()) != TYP_VOID || retBuffer != nullptr)
+                        {
+                            Statement* stmt = m_compiler->gtNewStmt(call);
+                            m_compiler->fgInsertStmtBefore(m_compiler->compCurBB, m_curStmt, stmt);
+                            GenTreeRetExpr* retExpr =
+                                m_compiler->gtNewInlineCandidateReturnExpr(call->AsCall(),
+                                                                           genActualType(call->TypeGet()));
+
+                            call->GetSingleInlineCandidateInfo()->retExpr = retExpr;
+
+                            m_curStmt = stmt;
+                            *pTree    = retExpr;
+                        }
+
+                        if (retBuffer != nullptr)
+                        {
+                            call->GetSingleInlineCandidateInfo()->preexistingSpillTemp =
+                                retBuffer->GetNode()->AsLclVarCommon()->GetLclNum();
+                        }
+
+                        call->GetSingleInlineCandidateInfo()->exactContextHandle = context;
+                        INDEBUG(call->GetSingleInlineCandidateInfo()->inlinersContext = call->gtInlineContext);
+
+                        JITDUMP("New inline candidate due to late devirtualization:\n");
+                        DISPTREE(call);
+                    }
+                }
                 m_madeChanges = true;
             }
         }
@@ -730,17 +792,10 @@ PhaseStatus Compiler::fgInline()
     do
     {
         // Make the current basic block address available globally
-        compCurBB = block;
-
-        for (Statement* const stmt : block->Statements())
+        compCurBB       = block;
+        Statement* stmt = block->firstStmt();
+        while (stmt != nullptr)
         {
-
-#if defined(DEBUG)
-            // In debug builds we want the inline tree to show all failed
-            // inlines. Some inlines may fail very early and never make it to
-            // candidate stage. So scan the tree looking for those early failures.
-            fgWalkTreePre(stmt->GetRootNodePointer(), fgFindNonInlineCandidate, stmt);
-#endif
             // See if we need to replace some return value place holders.
             // Also, see if this replacement enables further devirtualization.
             //
@@ -755,7 +810,7 @@ PhaseStatus Compiler::fgInline()
             // possible further optimization, as the (now complete) GT_RET_EXPR
             // replacement may have enabled optimizations by providing more
             // specific types for trees or variables.
-            walker.WalkTree(stmt->GetRootNodePointer(), nullptr);
+            stmt = walker.WalkStatement(stmt);
 
             GenTree* expr = stmt->GetRootNode();
 
@@ -805,6 +860,13 @@ PhaseStatus Compiler::fgInline()
                 madeChanges = true;
                 stmt->SetRootNode(expr->AsOp()->gtOp1);
             }
+
+#if defined(DEBUG)
+            // In debug builds we want the inline tree to show all failed
+            // inlines.
+            fgWalkTreePre(stmt->GetRootNodePointer(), fgFindNonInlineCandidate, stmt);
+#endif
+            stmt = stmt->GetNextStmt();
         }
 
         block = block->Next();

src/coreclr/jit/importercalls.cpp

@@ -329,7 +329,7 @@ var_types Compiler::impImportCall(OPCODE                  opcode,
                     assert((sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_VARARG &&
                            (sig->callConv & CORINFO_CALLCONV_MASK) != CORINFO_CALLCONV_NATIVEVARARG);
 
-                    call = gtNewIndCallNode(stubAddr, callRetTyp);
+                    call = gtNewIndCallNode(stubAddr, callRetTyp, di);
 
                     call->gtFlags |= GTF_EXCEPT | (stubAddr->gtFlags & GTF_GLOB_EFFECT);
                     call->gtFlags |= GTF_CALL_VIRT_STUB;
@@ -1027,7 +1027,7 @@ var_types Compiler::impImportCall(OPCODE                  opcode,
             if (!bIntrinsicImported)
             {
                 // Keep track of the raw IL offset of the call
-                INDEBUG(call->AsCall()->gtRawILOffset = rawILOffset);
+                call->AsCall()->gtRawILOffset = rawILOffset;
 
                 // Is it an inline candidate?
                 impMarkInlineCandidate(call, exactContextHnd, exactContextNeedsRuntimeLookup, callInfo, rawILOffset);
@@ -1237,7 +1237,7 @@ var_types Compiler::impImportCall(OPCODE                  opcode,
         }
 
         // Keep track of the raw IL offset of the call
-        INDEBUG(call->AsCall()->gtRawILOffset = rawILOffset);
+        call->AsCall()->gtRawILOffset = rawILOffset;
 
         // Is it an inline candidate?
         impMarkInlineCandidate(call, exactContextHnd, exactContextNeedsRuntimeLookup, callInfo, rawILOffset);
@@ -7695,6 +7695,19 @@ void Compiler::impMarkInlineCandidateHelper(GenTreeCall*           call,
 
     if (!(methAttr & CORINFO_FLG_FORCEINLINE))
     {
+        if (compCurBB->HasFlag(BBF_INTERNAL) && ehGetBlockHndDsc(compCurBB) != nullptr)
+        {
+#ifdef DEBUG
+            if (verbose)
+            {
+                printf("\nWill not inline blocks that have EH descriptor when IL offset may not be valid\n");
+            }
+
+#endif
+            inlineResult->NoteFatal(InlineObservation::CALLSITE_NOT_CANDIDATE);
+            return;
+        }
+
         /* Don't bother inline blocks that are in the filter region */
         if (bbInCatchHandlerILRange(compCurBB))
         {
@@ -9201,9 +9214,13 @@ void Compiler::impCheckCanInline(GenTreeCall*           call,
         CORINFO_CLASS_HANDLE clsHandle = compCompHnd->getMethodClass(ftn);
         unsigned const       clsAttr   = compCompHnd->getClassAttribs(clsHandle);
 
+        CallArg* pRetBuf = pParam->call->gtArgs.GetRetBufferArg();
+
         // Return type
         //
-        var_types const fncRetType = pParam->call->TypeGet();
+        var_types const fncRetType =
+            pRetBuf == nullptr ? pParam->call->TypeGet()
+                               : pParam->pThis->lvaGetRealType(pRetBuf->GetEarlyNode()->AsLclVarCommon()->GetLclNum());
 
 #ifdef DEBUG
         var_types fncRealRetType = JITtype2varType(methInfo.args.retType);