[NativeAOT-LLVM] Avoid unnecessary zeroing in LSSA (#3125)

* Avoid zeroing more than necessary in LSSA

* Add a few tests

---------

Co-authored-by: yowl <scott.waye@hubse.com>

Author: SingleAccretion

src/coreclr/jit/compiler.cpp

@@ -2680,6 +2680,10 @@ void Compiler::compInitOptions(JitFlags* jitFlags)
             {
                 verboseDump = true;
             }
+
+#ifdef TARGET_WASM
+            verboseDump |= m_llvm->EnableVerboseDump();
+#endif // TARGET_WASM
         }
     }
 

src/coreclr/jit/jitconfigvalues.h

@@ -850,6 +850,7 @@ RELEASE_CONFIG_INTEGER(JitCheckLlvmIR, "JitCheckLlvmIR", 0)
 RELEASE_CONFIG_INTEGER(JitRunLssaTests, "JitRunLssaTests", 0)
 RELEASE_CONFIG_INTEGER(JitGcStress, "JitGcStress", 0)
 
+CONFIG_STRING(JitDumpSymbol, "JitDumpSymbol")
 CONFIG_STRING(JitEnableLssaRange, "JitEnableLssaRange")
 #endif // TARGET_WASM
 

src/coreclr/jit/llvm.cpp

@@ -82,6 +82,9 @@ Llvm::Llvm(Compiler* compiler)
     , m_phiPairs(compiler->getAllocator(CMK_Codegen))
     , m_ehModel(GetExceptionHandlingModel())
     , m_debugVariablesMap(compiler->getAllocator(CMK_Codegen))
+#ifdef DEBUG
+    , m_optimizationRequirements(compiler->getAllocator(CMK_DebugOnly))
+#endif // DEBUG
 {
 }
 
@@ -108,6 +111,20 @@ Llvm::Llvm(Compiler* compiler)
 #endif // HOST_WINDOWS
 }
 
+bool Llvm::EnableVerboseDump()
+{
+#ifdef DEBUG
+    const char* dumpSymbolName = JitConfig.JitDumpSymbol();
+    const char* compilingSymbolName = GetMangledMethodName(m_info->compMethodHnd);
+    if ((dumpSymbolName != nullptr) && (strcmp(dumpSymbolName, compilingSymbolName) == 0))
+    {
+        return true;
+    }
+#endif // DEBUG
+
+    return false;
+}
+
 var_types Llvm::GetArgTypeForStructWasm(CORINFO_CLASS_HANDLE structHnd, structPassingKind* pPassKind)
 {
     // Note the managed and unmanaged ABIs are the same in terms of values, but do differ w.r.t by-ref
@@ -806,6 +823,44 @@ void Llvm::GetJitTestInfo(CorInfoLlvmJitTestKind kind, CORINFO_LLVM_JIT_TEST_INF
     CallEEApi<EEAI_GetJitTestInfo, CORINFO_GENERIC_HANDLE>(m_pEECorInfo, kind, pInfo);
 }
 
+void Llvm::ImposeOptimizationRequirement(GenTree* node, NodeOptimizationRequirement requirement)
+{
+#ifdef DEBUG
+    assert(node != nullptr);
+    *m_optimizationRequirements.LookupPointerOrAdd(node, NOR_NONE) |= requirement;
+#endif // DEBUG
+}
+
+void Llvm::SatisfyOptimizationRequirement(GenTree* node, NodeOptimizationRequirement requirement)
+{
+#ifdef DEBUG
+    assert(node != nullptr);
+    NodeOptimizationRequirement* pRequired = m_optimizationRequirements.LookupPointer(node);
+    if (pRequired != nullptr)
+    {
+        *pRequired &= ~requirement;
+        if (*pRequired == NOR_NONE)
+        {
+            m_optimizationRequirements.Remove(node);
+        }
+    }
+#endif // DEBUG
+}
+
+void Llvm::VerifyAllOptimizationRequirementsSatisfied()
+{
+#ifdef DEBUG
+    if (m_optimizationRequirements.GetCount() != 0)
+    {
+        for (GenTree* node : decltype(m_optimizationRequirements)::KeyIteration(&m_optimizationRequirements))
+        {
+            _compiler->gtDispTree(node, nullptr, nullptr, true);
+        }
+        assert(!"Found nodes with unsatisfied optimization requirements");
+    }
+#endif // DEBUG
+}
+
 static SingleThreadedCompilationContext* StartSingleThreadedCompilation(
     CorInfoLlvmSingleThreadedCompilationContextFlags flags,
     const char* path,

src/coreclr/jit/llvm.h

@@ -14,9 +14,6 @@
 #include "JitEEApi.Shared.cspp"
 #include <new>
 
-// this breaks StringMap.h
-#undef NumItems
-
 // TODO-LLVM-Upstream: figure out how to fix these warnings in LLVM headers.
 #pragma warning(push)
 #pragma warning(disable : 4146)
@@ -215,6 +212,27 @@ inline CallSiteFacts& operator |=(CallSiteFacts& a, CallSiteFacts b)
     return a = static_cast<CallSiteFacts>(static_cast<int>(a) | static_cast<int>(b));
 }
 
+enum NodeOptimizationRequirement
+{
+    NOR_NONE = 0,
+    NOR_SHADOW_TAILCALL = 1
+};
+
+inline NodeOptimizationRequirement& operator |=(NodeOptimizationRequirement& a, NodeOptimizationRequirement b)
+{
+    return a = static_cast<NodeOptimizationRequirement>(static_cast<int>(a) | static_cast<int>(b));
+}
+
+inline NodeOptimizationRequirement& operator &=(NodeOptimizationRequirement& a, NodeOptimizationRequirement b)
+{
+    return a = static_cast<NodeOptimizationRequirement>(static_cast<int>(a) & static_cast<int>(b));
+}
+
+inline NodeOptimizationRequirement operator ~(NodeOptimizationRequirement a)
+{
+    return static_cast<NodeOptimizationRequirement>(~static_cast<int>(a));
+}
+
 struct ThrowHelperKey
 {
     unsigned ThrowIndex;
@@ -231,6 +249,14 @@ struct ThrowHelperKey
     }
 };
 
+struct ThrowHelperCode
+{
+    llvm::BasicBlock* LlvmBlock;
+#ifdef DEBUG
+    bool ShadowTailCalledThrowHelper;
+#endif // DEBUG
+};
+
 struct PhiPair
 {
     GenTreeLclVar* StoreNode;
@@ -330,7 +356,7 @@ class Llvm
     llvm::IRBuilder<> _builder;
     JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, Value*> _sdsuMap;
     JitHashTable<SSAName, SSAName, Value*> _localsMap;
-    JitHashTable<ThrowHelperKey, ThrowHelperKey, llvm::BasicBlock*> m_throwHelperBlocksMap;
+    JitHashTable<ThrowHelperKey, ThrowHelperKey, ThrowHelperCode> m_throwHelperBlocksMap;
     jitstd::vector<PhiPair> m_phiPairs;
     FunctionInfo* m_functions;
 
@@ -361,6 +387,10 @@ class Llvm
     unsigned m_preciseVirtualUnwindFrameLclNum = BAD_VAR_NUM;
     unsigned _llvmArgCount = 0;
 
+#ifdef DEBUG
+    JitHashTable<GenTree*, JitPtrKeyFuncs<GenTree>, NodeOptimizationRequirement> m_optimizationRequirements;
+#endif // DEBUG
+
     // ================================================================================================================
     // |                                                   General                                                    |
     // ================================================================================================================
@@ -369,6 +399,7 @@ class Llvm
     Llvm(Compiler* compiler);
 
     static void ConfigureDiagnosticOutput();
+    bool EnableVerboseDump();
 
     var_types GetArgTypeForStructWasm(CORINFO_CLASS_HANDLE structHnd, structPassingKind* pPassKind);
     var_types GetReturnTypeForStructWasm(CORINFO_CLASS_HANDLE structHnd, structPassingKind* pPassKind);
@@ -425,6 +456,10 @@ class Llvm
     bool IsVirtualUnwindFrameVisible();
     void GetJitTestInfo(CorInfoLlvmJitTestKind kind, CORINFO_LLVM_JIT_TEST_INFO* pInfo);
 
+    void ImposeOptimizationRequirement(GenTree* node, NodeOptimizationRequirement requirement);
+    void SatisfyOptimizationRequirement(GenTree* node, NodeOptimizationRequirement requirement);
+    void VerifyAllOptimizationRequirementsSatisfied();
+
     // ================================================================================================================
     // |                                                 Type system                                                  |
     // ================================================================================================================
@@ -495,9 +530,8 @@ class Llvm
     void lowerDissolveDependentlyPromotedLocals();
     void dissolvePromotedLocal(unsigned lclNum);
 
-    void lowerCanonicalizeFirstBlock();
     bool isFirstBlockCanonical();
-    void lowerAndInsertIntoFirstBlock(LIR::Range& range, GenTree* insertAfter = nullptr);
+    GenTree* lowerAndInsertIntoFirstBlock(LIR::Range& range, GenTree* insertAfter = nullptr);
 
 public:
     PhaseStatus AddVirtualUnwindFrame();
@@ -543,7 +577,8 @@ class Llvm
 
     unsigned getShadowFrameSize(unsigned funcIdx) const;
     unsigned getCalleeShadowStackOffset(unsigned funcIdx, bool isTailCall) const;
-    bool canEmitCallAsShadowTailCall(bool callIsInTry, bool callIsInFilter DEBUGARG(const char** pReasonWhyNot)) const;
+    bool canEmitCallAsShadowTailCall(
+        bool callIsInTry, bool callIsInFilter DEBUGARG(const char** pReasonWhyNot = nullptr)) const;
     bool isPotentialGcSafePoint(GenTree* node) const;
     bool isShadowFrameLocal(LclVarDsc* varDsc) const;
     bool isShadowStackLocal(unsigned lclNum) const;
@@ -621,19 +656,21 @@ class Llvm
     void buildCallFinally(BasicBlock* block);
 
     Value* consumeAddressAndEmitNullCheck(GenTreeIndir* indir);
-    void emitNullCheckForAddress(GenTree* addr, Value* addrValue);
-    void emitAlignmentCheckForAddress(GenTree* addr, Value* addrValue, unsigned alignment);
+    void emitNullCheckForAddress(GenTree* addr, Value* addrValue DEBUGARG(GenTree* indir));
+    void emitAlignmentCheckForAddress(GenTree* addr, Value* addrValue, unsigned alignment DEBUGARG(GenTree* indir));
     bool isAddressAligned(GenTree* addr, unsigned alignment);
 
     Value* consumeInitVal(GenTree* initVal);
     void storeObjAtAddress(Value* baseAddress, Value* data, StructDesc* structDesc);
     unsigned buildMemCpy(Value* baseAddress, unsigned startOffset, unsigned endOffset, Value* srcAddress);
 
-    void emitJumpToThrowHelper(Value* jumpCondValue, CorInfoHelpFunc helperFunc);
-    Value* emitCheckedArithmeticOperation(llvm::Intrinsic::ID intrinsicId, Value* op1Value, Value* op2Value);
+    void emitJumpToThrowHelper(Value* jumpCondValue, CorInfoHelpFunc helperFunc DEBUGARG(GenTree* nodeThrowing));
+    Value* emitCheckedArithmeticOperation(
+        llvm::Intrinsic::ID intrinsicId, Value* op1Value, Value* op2Value DEBUGARG(GenTree* opNode));
 
     llvm::CallBase* emitGcStressCall(GenTreeCall* call, llvm::CallBase* callValue);
-    llvm::CallBase* emitHelperCall(CorInfoHelpFunc helperFunc, ArrayRef<Value*> sigArgs = {});
+    llvm::CallBase* emitHelperCall(
+        CorInfoHelpFunc helperFunc, ArrayRef<Value*> sigArgs = {} DEBUGARG(bool* pIsShadowTailCall = nullptr));
     bool canEmitHelperCallAsShadowTailCall(CorInfoHelpFunc helperFunc);
     llvm::CallBase* emitCallOrInvoke(llvm::FunctionCallee callee, ArrayRef<Value*> args, CallSiteFacts facts);