Late cast expansion: remove the null check if possible (#97234)

src/coreclr/jit/assertionprop.cpp

@@ -5013,10 +5013,16 @@ GenTree* Compiler::optAssertionProp_Call(ASSERT_VALARG_TP assertions, GenTreeCal
                 return optAssertionProp_Update(arg1, call, stmt);
             }
 
-            // TODO-InlineCast: check optAssertionIsNonNull for the object argument and replace
-            // the helper with its nonnull version, e.g.:
-            // CORINFO_HELP_ISINSTANCEOFANY -> CORINFO_HELP_ISINSTANCEOFANY_NONNULL
-            // so then fgLateCastExpansion can skip the null check.
+            // Leave a hint for fgLateCastExpansion that obj is never null.
+            INDEBUG(AssertionIndex nonNullIdx = NO_ASSERTION_INDEX);
+            INDEBUG(bool vnBased = false);
+            // GTF_CALL_M_CAST_CAN_BE_EXPANDED check is to improve TP
+            if (((call->gtCallMoreFlags & GTF_CALL_M_CAST_CAN_BE_EXPANDED) != 0) &&
+                optAssertionIsNonNull(arg1, assertions DEBUGARG(&vnBased) DEBUGARG(&nonNullIdx)))
+            {
+                call->gtCallMoreFlags |= GTF_CALL_M_CAST_OBJ_NONNULL;
+                return optAssertionProp_Update(call, call, stmt);
+            }
         }
     }
 

src/coreclr/jit/compiler.cpp

@@ -10028,11 +10028,6 @@ JITDBGAPI void __cdecl cTreeFlags(Compiler* comp, GenTree* tree)
                         chars += printf("[CALL_GUARDED]");
                     }
 
-                    if (call->IsExpRuntimeLookup())
-                    {
-                        chars += printf("[CALL_EXP_RUNTIME_LOOKUP]");
-                    }
-
                     if (call->gtCallDebugFlags & GTF_CALL_MD_STRESS_TAILCALL)
                     {
                         chars += printf("[CALL_MD_STRESS_TAILCALL]");

src/coreclr/jit/gcinfo.cpp

@@ -240,7 +240,7 @@ GCInfo::WriteBarrierForm GCInfo::gcIsWriteBarrierCandidate(GenTreeStoreInd* stor
     }
 
     // Ignore any assignments of NULL.
-    GenTree* const data = store->Data()->gtSkipReloadOrCopy()->gtEffectiveVal();
+    GenTree* const data = store->Data()->gtSkipReloadOrCopy();
     if (data->IsIntegralConst(0))
     {
         return WBF_NoBarrier;

src/coreclr/jit/gentree.cpp

@@ -2492,6 +2492,34 @@ bool GenTreeCall::IsHelperCall(Compiler* compiler, unsigned helper) const
     return IsHelperCall(compiler->eeFindHelper(helper));
 }
 
+//-------------------------------------------------------------------------
+// IsRuntimeLookupHelperCall: Determine if this GT_CALL node represents a runtime lookup helper call.
+//
+// Arguments:
+//     compiler - the compiler instance so that we can call eeGetHelperNum
+//
+// Return Value:
+//     Returns true if this GT_CALL node represents a runtime lookup helper call.
+//
+bool GenTreeCall::IsRuntimeLookupHelperCall(Compiler* compiler) const
+{
+    if (!IsHelperCall())
+    {
+        return false;
+    }
+
+    switch (compiler->eeGetHelperNum(gtCallMethHnd))
+    {
+        case CORINFO_HELP_RUNTIMEHANDLE_METHOD:
+        case CORINFO_HELP_RUNTIMEHANDLE_CLASS:
+        case CORINFO_HELP_RUNTIMEHANDLE_METHOD_LOG:
+        case CORINFO_HELP_RUNTIMEHANDLE_CLASS_LOG:
+            return true;
+        default:
+            return false;
+    }
+}
+
 //-------------------------------------------------------------------------
 // IsSpecialIntrinsic: Determine if this GT_CALL node is a specific intrinsic.
 //
@@ -2550,6 +2578,12 @@ bool GenTreeCall::Equals(GenTreeCall* c1, GenTreeCall* c2)
             return false;
         }
 
+        if (c1->IsHelperCall() && ((c1->gtCallMoreFlags & GTF_CALL_M_CAST_OBJ_NONNULL) !=
+                                   (c2->gtCallMoreFlags & GTF_CALL_M_CAST_OBJ_NONNULL)))
+        {
+            return false;
+        }
+
 #ifdef FEATURE_READYTORUN
         if (c1->gtEntryPoint.addr != c2->gtEntryPoint.addr)
         {

src/coreclr/jit/gentree.h

@@ -4101,11 +4101,12 @@ enum GenTreeCallFlags : unsigned int
     GTF_CALL_M_GUARDED_DEVIRT_CHAIN    = 0x00080000, // this call is a candidate for chained guarded devirtualization
     GTF_CALL_M_ALLOC_SIDE_EFFECTS      = 0x00100000, // this is a call to an allocator with side effects
     GTF_CALL_M_SUPPRESS_GC_TRANSITION  = 0x00200000, // suppress the GC transition (i.e. during a pinvoke) but a separate GC safe point is required.
-    GTF_CALL_M_EXP_RUNTIME_LOOKUP      = 0x00400000, // [DEBUG only] this call needs to be transformed into CFG for the dynamic dictionary expansion feature.
     GTF_CALL_M_EXPANDED_EARLY          = 0x00800000, // the Virtual Call target address is expanded and placed in gtControlExpr in Morph rather than in Lower
     GTF_CALL_M_HAS_LATE_DEVIRT_INFO    = 0x01000000, // this call has late devirtualzation info
     GTF_CALL_M_LDVIRTFTN_INTERFACE     = 0x02000000, // ldvirtftn on an interface type
     GTF_CALL_M_CAST_CAN_BE_EXPANDED    = 0x04000000, // this cast (helper call) can be expanded if it's profitable. To be removed.
+    GTF_CALL_M_CAST_OBJ_NONNULL        = 0x08000000, // if we expand this specific cast we don't need to check the input object for null
+                                                     // NOTE: if needed, this flag can be removed, and we can introduce new _NONNUL cast helpers
 };
 
 inline constexpr GenTreeCallFlags operator ~(GenTreeCallFlags a)
@@ -4141,6 +4142,7 @@ enum GenTreeCallDebugFlags : unsigned int
     GTF_CALL_MD_DEVIRTUALIZED           = 0x00000002, // this call was devirtualized
     GTF_CALL_MD_UNBOXED                 = 0x00000004, // this call was optimized to use the unboxed entry point
     GTF_CALL_MD_GUARDED                 = 0x00000008, // this call was transformed by guarded devirtualization
+    GTF_CALL_MD_RUNTIME_LOOKUP_EXPANDED = 0x00000010, // this runtime lookup helper is expanded
 };
 
 inline constexpr GenTreeCallDebugFlags operator ~(GenTreeCallDebugFlags a)
@@ -5476,21 +5478,6 @@ struct GenTreeCall final : public GenTree
     }
 #endif
 
-    void SetExpRuntimeLookup()
-    {
-        gtCallMoreFlags |= GTF_CALL_M_EXP_RUNTIME_LOOKUP;
-    }
-
-    void ClearExpRuntimeLookup()
-    {
-        gtCallMoreFlags &= ~GTF_CALL_M_EXP_RUNTIME_LOOKUP;
-    }
-
-    bool IsExpRuntimeLookup() const
-    {
-        return (gtCallMoreFlags & GTF_CALL_M_EXP_RUNTIME_LOOKUP) != 0;
-    }
-
     void SetExpandedEarly()
     {
         gtCallMoreFlags |= GTF_CALL_M_EXPANDED_EARLY;
@@ -5684,6 +5671,8 @@ struct GenTreeCall final : public GenTree
 
     bool IsHelperCall(Compiler* compiler, unsigned helper) const;
 
+    bool IsRuntimeLookupHelperCall(Compiler* compiler) const;
+
     bool IsSpecialIntrinsic(Compiler* compiler, NamedIntrinsic ni) const;
 
     CorInfoHelpFunc GetHelperNum() const;

src/coreclr/jit/helperexpansion.cpp

@@ -94,7 +94,6 @@ GenTreeCall* Compiler::gtNewRuntimeLookupHelperCallNode(CORINFO_RUNTIME_LOOKUP*
     // Leave a note that this method has runtime lookups we might want to expand (nullchecks, size checks) later.
     // We can also consider marking current block as a runtime lookup holder to improve TP for Tier0
     impInlineRoot()->setMethodHasExpRuntimeLookup();
-    helperCall->SetExpRuntimeLookup();
     if (!impInlineRoot()->GetSignatureToLookupInfoMap()->Lookup(pRuntimeLookup->signature))
     {
         JITDUMP("Registering %p in SignatureToLookupInfoMap\n", pRuntimeLookup->signature)
@@ -152,13 +151,12 @@ bool Compiler::fgExpandRuntimeLookupsForCall(BasicBlock** pBlock, Statement* stm
 {
     BasicBlock* block = *pBlock;
 
-    if (!call->IsHelperCall() || !call->IsExpRuntimeLookup())
+    if (!call->IsRuntimeLookupHelperCall(this))
     {
         return false;
     }
 
-    // Clear ExpRuntimeLookup flag so we won't miss any runtime lookup that needs partial expansion
-    call->ClearExpRuntimeLookup();
+    INDEBUG(call->gtCallDebugFlags |= GTF_CALL_MD_RUNTIME_LOOKUP_EXPANDED);
 
     if (call->IsTailCall())
     {
@@ -2405,6 +2403,15 @@ bool Compiler::fgLateCastExpansionForCall(BasicBlock** pBlock, Statement* stmt,
     assert(BasicBlock::sameEHRegion(firstBb, fallbackBb));
     assert(BasicBlock::sameEHRegion(firstBb, lastTypeCheckBb));
 
+    // call guarantees that obj is never null, we can drop the nullcheck
+    // by converting it to a BBJ_ALWAYS to typeCheckBb.
+    if ((call->gtCallMoreFlags & GTF_CALL_M_CAST_OBJ_NONNULL) != 0)
+    {
+        fgRemoveStmt(nullcheckBb, nullcheckBb->lastStmt());
+        nullcheckBb->SetKindAndTarget(BBJ_ALWAYS, typeChecksBbs[0]);
+        fgRemoveRefPred(lastBb, nullcheckBb);
+    }
+
     // Bonus step: merge prevBb with nullcheckBb as they are likely to be mergeable
     if (fgCanCompactBlocks(firstBb, nullcheckBb))
     {

src/coreclr/jit/lower.cpp

@@ -2203,7 +2203,8 @@ GenTree* Lowering::LowerCall(GenTree* node)
     JITDUMP("\n");
 
     // All runtime lookups are expected to be expanded in fgExpandRuntimeLookups
-    assert(!call->IsExpRuntimeLookup());
+    assert(!call->IsRuntimeLookupHelperCall(comp) ||
+           (call->gtCallDebugFlags & GTF_CALL_MD_RUNTIME_LOOKUP_EXPANDED) != 0);
 
     // Also, always expand static cctor helper for NativeAOT, see
     // /~https://github.com/dotnet/runtime/issues/68278#issuecomment-1543322819