JuliaLang · vchuravy · Dec 14, 2021 · Oct 12, 2021 · Oct 22, 2021 · Oct 22, 2021
diff --git a/doc/src/devdocs/llvm.md b/doc/src/devdocs/llvm.md
@@ -79,7 +79,7 @@ environment. In addition, it exposes the `-julia` meta-pass, which runs the
 entire Julia pass-pipeline over the IR. As an example, to generate a system
 image, one could do:
 ```
-opt -load libjulia-internal.so -julia -o opt.bc unopt.bc
+opt -enable-new-pm=0 -load libjulia-internal.so -julia -o opt.bc unopt.bc
 llc -o sys.o opt.bc
 cc -shared -o sys.so sys.o
 ```

diff --git a/src/abi_aarch64.cpp b/src/abi_aarch64.cpp
@@ -187,7 +187,7 @@ Type *isHFAorHVA(jl_datatype_t *dt, size_t &nele, LLVMContext &ctx) const
     return NULL;
 }
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     // B.2
     //   If the argument type is an HFA or an HVA, then the argument is used

diff --git a/src/abi_arm.cpp b/src/abi_arm.cpp
@@ -23,7 +23,7 @@
 
 struct ABI_ARMLayout : AbiLayout {
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &abi, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &abi, LLVMContext &ctx, Type *Ty) override
 {
     return false;
 }

diff --git a/src/abi_llvm.cpp b/src/abi_llvm.cpp
@@ -45,7 +45,7 @@ bool use_sret(jl_datatype_t *ty, LLVMContext &ctx) override
     return false;
 }
 
-bool needPassByRef(jl_datatype_t *ty, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *ty, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     return false;
 }

diff --git a/src/abi_ppc64le.cpp b/src/abi_ppc64le.cpp
@@ -101,12 +101,16 @@ bool use_sret(jl_datatype_t *dt, LLVMContext &ctx) override
     return false;
 }
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     jl_datatype_t *ty0 = NULL;
     bool hva = false;
     if (jl_datatype_size(dt) > 64 && isHFA(dt, &ty0, &hva) > 8) {
+#if JL_LLVM_VERSION < 120000
         ab.addAttribute(Attribute::ByVal);
+#else
+        ab.addByValAttr(Ty);
+#endif
         return true;
     }
     return false;

diff --git a/src/abi_win32.cpp b/src/abi_win32.cpp
@@ -49,11 +49,15 @@ bool use_sret(jl_datatype_t *dt, LLVMContext &ctx) override
     return true;
 }
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     // Use pass by reference for all structs
     if (dt->layout->nfields > 0) {
+#if JL_LLVM_VERSION < 120000
         ab.addAttribute(Attribute::ByVal);
+#else
+        ab.addByValAttr(Ty);
+#endif
         return true;
     }
     return false;

diff --git a/src/abi_win64.cpp b/src/abi_win64.cpp
@@ -56,14 +56,19 @@ bool use_sret(jl_datatype_t *dt, LLVMContext &ctx) override
     return true;
 }
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     nargs++;
     size_t size = jl_datatype_size(dt);
     if (win64_reg_size(size))
         return false;
-    if (nargs <= 4)
+    if (nargs <= 4) {
+#if JL_LLVM_VERSION < 120000
         ab.addAttribute(Attribute::ByVal);
+#else
+        ab.addByValAttr(Ty);
+#endif
+    }
     return true;
 }
 

diff --git a/src/abi_x86.cpp b/src/abi_x86.cpp
@@ -67,12 +67,16 @@ bool use_sret(jl_datatype_t *dt, LLVMContext &ctx) override
     return true;
 }
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     size_t size = jl_datatype_size(dt);
     if (is_complex64(dt) || is_complex128(dt) || (jl_is_primitivetype(dt) && size <= 8))
         return false;
-    ab.addAttribute(Attribute::ByVal);
+#if JL_LLVM_VERSION < 120000
+        ab.addAttribute(Attribute::ByVal);
+#else
+        ab.addByValAttr(Ty);
+#endif
     return true;
 }
 

diff --git a/src/abi_x86_64.cpp b/src/abi_x86_64.cpp
@@ -178,11 +178,15 @@ bool use_sret(jl_datatype_t *dt, LLVMContext &ctx) override
     return sret;
 }
 
-bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) override
+bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *Ty) override
 {
     Classification cl = classify(dt);
     if (cl.isMemory) {
+#if JL_LLVM_VERSION < 120000
         ab.addAttribute(Attribute::ByVal);
+#else
+        ab.addByValAttr(Ty);
+#endif
         return true;
     }
 
@@ -202,7 +206,12 @@ bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx) overrid
     else if (jl_is_structtype(dt)) {
         // spill to memory even though we would ordinarily pass
         // it in registers
+#if JL_LLVM_VERSION < 120000
         ab.addAttribute(Attribute::ByVal);
+#else
+        Type* Ty = preferred_llvm_type(dt, false, ctx);
+        ab.addByValAttr(Ty);
+#endif
         return true;
     }
     return false;

diff --git a/src/aotcompile.cpp b/src/aotcompile.cpp
@@ -550,6 +550,10 @@ void jl_dump_native_impl(void *native_code,
     std::unique_ptr<Module> sysimage(new Module("sysimage", Context));
     sysimage->setTargetTriple(data->M->getTargetTriple());
     sysimage->setDataLayout(data->M->getDataLayout());
+#if JL_LLVM_VERSION >= 130000
+    sysimage->setStackProtectorGuard(data->M->getStackProtectorGuard());
+    sysimage->setOverrideStackAlignment(data->M->getOverrideStackAlignment());
+#endif
     data->M.reset(); // free memory for data->M
 
     if (sysimg_data) {

diff --git a/src/ccall.cpp b/src/ccall.cpp
@@ -288,7 +288,7 @@ class AbiLayout {
 public:
     virtual ~AbiLayout() {}
     virtual bool use_sret(jl_datatype_t *ty, LLVMContext &ctx) = 0;
-    virtual bool needPassByRef(jl_datatype_t *ty, AttrBuilder&, LLVMContext &ctx) = 0;
+    virtual bool needPassByRef(jl_datatype_t *ty, AttrBuilder&, LLVMContext &ctx, Type* llvm_t) = 0;
     virtual Type *preferred_llvm_type(jl_datatype_t *ty, bool isret, LLVMContext &ctx) const = 0;
 };
 
@@ -891,6 +891,10 @@ static jl_cgval_t emit_llvmcall(jl_codectx_t &ctx, jl_value_t **args, size_t nar
     // copy module properties that should always match
     Mod->setTargetTriple(jl_Module->getTargetTriple());
     Mod->setDataLayout(jl_Module->getDataLayout());
+#if JL_LLVM_VERSION >= 130000
+    Mod->setStackProtectorGuard(jl_Module->getStackProtectorGuard());
+    Mod->setOverrideStackAlignment(jl_Module->getOverrideStackAlignment());
+#endif
 
     // verify the definition
     Function *def = Mod->getFunction(ir_name);
@@ -1086,7 +1090,7 @@ std::string generate_func_sig(const char *fname)
         }
 
         // Whether or not LLVM wants us to emit a pointer to the data
-        bool byRef = abi->needPassByRef((jl_datatype_t*)tti, ab, jl_LLVMContext);
+        bool byRef = abi->needPassByRef((jl_datatype_t*)tti, ab, jl_LLVMContext, t);
 
         if (jl_is_cpointer_type(tti)) {
             pat = t;

diff --git a/src/codegen.cpp b/src/codegen.cpp
@@ -144,7 +144,7 @@ extern JITEventListener *CreateJuliaJITEventListener();
 bool imaging_mode = false;
 
 // shared llvm state
-JL_DLLEXPORT LLVMContext &jl_LLVMContext = *(new LLVMContext());
+static LLVMContext &jl_LLVMContext = *(new LLVMContext());
 TargetMachine *jl_TargetMachine;
 static DataLayout &jl_data_layout = *(new DataLayout(""));
 #define jl_Module ctx.f->getParent()
@@ -1707,6 +1707,16 @@ static void jl_setup_module(Module *m, const jl_cgparams_t *params = &jl_default
             llvm::DEBUG_METADATA_VERSION);
     m->setDataLayout(jl_data_layout);
     m->setTargetTriple(jl_TargetMachine->getTargetTriple().str());
+
+#if defined(_OS_WINDOWS_) && !defined(_CPU_X86_64_) && JL_LLVM_VERSION >= 130000
+    // tell Win32 to assume the stack is always 16-byte aligned,
+    // and to ensure that it is 16-byte aligned for out-going calls,
+    // to ensure compatibility with GCC codes
+    m->setOverrideStackAlignment(16);
+#endif
+#if defined(JL_DEBUG_BUILD) && JL_LLVM_VERSION >= 130000
+    m->setStackProtectorGuard("global");
+#endif
 }
 
 Module *jl_create_llvm_module(StringRef name)
@@ -4780,7 +4790,7 @@ static Value *get_current_task(jl_codectx_t &ctx)
 // Get PTLS through current task.
 static Value *get_current_ptls(jl_codectx_t &ctx)
 {
-    return get_current_ptls_from_task(ctx.builder, get_current_task(ctx));
+    return get_current_ptls_from_task(ctx.builder, get_current_task(ctx), tbaa_gcframe);
 }
 
 // Store world age at the entry block of the function. This function should be
@@ -7679,7 +7689,7 @@ void jl_compile_workqueue(
 
 // --- initialization ---
 
-std::pair<MDNode*,MDNode*> tbaa_make_child(const char *name, MDNode *parent=nullptr, bool isConstant=false)
+static std::pair<MDNode*,MDNode*> tbaa_make_child(const char *name, MDNode *parent=nullptr, bool isConstant=false)
 {
     MDBuilder mbuilder(jl_LLVMContext);
     if (tbaa_root == nullptr) {
@@ -8033,13 +8043,14 @@ extern "C" void jl_init_llvm(void)
 
     TargetOptions options = TargetOptions();
     //options.PrintMachineCode = true; //Print machine code produced during JIT compiling
-#if defined(_OS_WINDOWS_) && !defined(_CPU_X86_64_)
+#if defined(_OS_WINDOWS_) && !defined(_CPU_X86_64_) && JL_LLVM_VERSION < 130000
     // tell Win32 to assume the stack is always 16-byte aligned,
     // and to ensure that it is 16-byte aligned for out-going calls,
     // to ensure compatibility with GCC codes
+    // In LLVM 13 and onwards this has turned into a module option
     options.StackAlignmentOverride = 16;
 #endif
-#ifdef JL_DEBUG_BUILD
+#if defined(JL_DEBUG_BUILD) && JL_LLVM_VERSION < 130000
     // LLVM defaults to tls stack guard, which causes issues with Julia's tls implementation
     options.StackProtectorGuard = StackProtectorGuards::Global;
 #endif

diff --git a/src/codegen_shared.h b/src/codegen_shared.h
@@ -99,9 +99,6 @@ static inline std::pair<llvm::MDNode*,llvm::MDNode*> tbaa_make_child_with_contex
     return std::make_pair(n, scalar);
 }
 
-static inline llvm::MDNode *get_tbaa_gcframe(llvm::LLVMContext &ctxt) {
-    return tbaa_make_child_with_context(ctxt, "jtbaa_gcframe").first;
-}
 static inline llvm::MDNode *get_tbaa_const(llvm::LLVMContext &ctxt) {
     return tbaa_make_child_with_context(ctxt, "jtbaa_const", nullptr, true).first;
 }
@@ -132,7 +129,7 @@ static inline llvm::Value *emit_bitcast_with_builder(llvm::IRBuilder<> &builder,
 }
 
 // Get PTLS through current task.
-static inline llvm::Value *get_current_ptls_from_task(llvm::IRBuilder<> &builder, llvm::Value *current_task)
+static inline llvm::Value *get_current_ptls_from_task(llvm::IRBuilder<> &builder, llvm::Value *current_task, llvm::MDNode *tbaa)
 {
     using namespace llvm;
     auto T_ppjlvalue = JuliaType::get_ppjlvalue_ty(builder.getContext());
@@ -145,9 +142,9 @@ static inline llvm::Value *get_current_ptls_from_task(llvm::IRBuilder<> &builder
     LoadInst *ptls_load = builder.CreateAlignedLoad(
         emit_bitcast_with_builder(builder, pptls, T_ppjlvalue), Align(sizeof(void *)), "ptls_load");
     // Note: Corresponding store (`t->ptls = ptls`) happens in `ctx_switch` of tasks.c.
-    tbaa_decorate(get_tbaa_gcframe(builder.getContext()), ptls_load);
+    tbaa_decorate(tbaa, ptls_load);
     // Using `CastInst::Create` to get an `Instruction*` without explicit cast:
     auto ptls = CastInst::Create(Instruction::BitCast, ptls_load, T_ppjlvalue, "ptls");
     builder.Insert(ptls);
     return ptls;
-}
+}
diff --git a/src/debuginfo.cpp b/src/debuginfo.cpp
@@ -177,13 +177,6 @@ struct revcomp {
     { return lhs>rhs; }
 };
 
-struct strrefcomp {
-    bool operator() (const StringRef& lhs, const StringRef& rhs) const
-    {
-        return lhs.compare(rhs) > 0;
-    }
-};
-
 class JuliaJITEventListener: public JITEventListener
 {
     std::map<size_t, ObjectInfo, revcomp> objectmap;
@@ -234,11 +227,13 @@ class JuliaJITEventListener: public JITEventListener
         SavedObject.second.release();
 
         object::section_iterator EndSection = debugObj.section_end();
-        std::map<StringRef, object::SectionRef, strrefcomp> loadedSections;
+        StringMap<object::SectionRef> loadedSections;
         for (const object::SectionRef &lSection: Object.sections()) {
             auto sName = lSection.getName();
-            if (sName)
-                loadedSections[*sName] = lSection;
+            if (sName) {
+                bool inserted = loadedSections.insert(std::make_pair(*sName, lSection)).second;
+                assert(inserted); (void)inserted;
+            }
         }
         auto getLoadAddress = [&] (const StringRef &sName) -> uint64_t {
             auto search = loadedSections.find(sName);

diff --git a/src/disasm.cpp b/src/disasm.cpp
@@ -860,21 +860,21 @@ static void jl_dump_asm_internal(
     std::unique_ptr<MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TheTriple.str()));
     assert(MRI && "Unable to create target register info!");
 
-    std::unique_ptr<MCObjectFileInfo> MOFI(new MCObjectFileInfo());
-#if JL_LLVM_VERSION >= 130000
-    MCSubtargetInfo *MSTI = TheTarget->createMCSubtargetInfo(TheTriple.str(), cpu, features);
-    assert(MSTI && "Unable to create subtarget info!");
+    std::unique_ptr<llvm::MCSubtargetInfo> STI(
+      TheTarget->createMCSubtargetInfo(TheTriple.str(), cpu, features));
+    assert(STI && "Unable to create subtarget info!");
 
-    MCContext Ctx(TheTriple, MAI.get(), MRI.get(), MSTI, &SrcMgr);
-    MOFI->initMCObjectFileInfo(Ctx, /* PIC */ false, /* LargeCodeModel */ false);
+#if JL_LLVM_VERSION >= 130000
+    MCContext Ctx(TheTriple, MAI.get(), MRI.get(), STI.get(), &SrcMgr);
+    std::unique_ptr<MCObjectFileInfo> MOFI(
+      TheTarget->createMCObjectFileInfo(Ctx, /*PIC=*/false, /*LargeCodeModel=*/ false));
+    Ctx.setObjectFileInfo(MOFI.get());
 #else
+    std::unique_ptr<MCObjectFileInfo> MOFI(new MCObjectFileInfo());
     MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &SrcMgr);
     MOFI->InitMCObjectFileInfo(TheTriple, /* PIC */ false, Ctx);
 #endif
 
-    // Set up Subtarget and Disassembler
-    std::unique_ptr<MCSubtargetInfo>
-        STI(TheTarget->createMCSubtargetInfo(TheTriple.str(), cpu, features));
     std::unique_ptr<MCDisassembler> DisAsm(TheTarget->createMCDisassembler(*STI, Ctx));
     if (!DisAsm) {
         rstream << "ERROR: no disassembler for target " << TheTriple.str();

diff --git a/src/jitlayers.cpp b/src/jitlayers.cpp
@@ -1045,7 +1045,7 @@ static void jl_add_to_ee(std::unique_ptr<Module> m)
             ConstantAggregateZero::get(atype), "__catchjmp") };
     gvs[0]->setSection(".text");
     gvs[1]->setSection(".text");
-    appendToUsed(*m, makeArrayRef((GlobalValue**)gvs, 2));
+    appendToCompilerUsed(*m, makeArrayRef((GlobalValue**)gvs, 2));
 #endif
     jl_jit_share_data(*m);
     assert(jl_ExecutionEngine);

diff --git a/src/llvm-late-gc-lowering.cpp b/src/llvm-late-gc-lowering.cpp
@@ -616,7 +616,7 @@ std::vector<Value*> LateLowerGCFrame::MaybeExtractVector(State &S, Value *BaseVe
     std::vector<Value*> V{Numbers.size()};
     Value *V_rnull = ConstantPointerNull::get(cast<PointerType>(T_prjlvalue));
     for (unsigned i = 0; i < V.size(); ++i) {
-        if (Numbers[i] >= 0)
+        if (Numbers[i] >= 0) // ignores undef and poison values
             V[i] = GetPtrForNumber(S, Numbers[i], InsertBefore);
         else
             V[i] = V_rnull;
@@ -860,8 +860,9 @@ std::vector<int> LateLowerGCFrame::NumberAllBase(State &S, Value *CurrentV) {
         std::vector<int> Numbers2 = NumberAll(S, SVI->getOperand(1));
         auto Mask = SVI->getShuffleMask();
         for (auto idx : Mask) {
-            assert(idx != -1 && "Undef tracked value is invalid");
-            if ((unsigned)idx < Numbers1.size()) {
+            if (idx == -1) {
+                Numbers.push_back(-1);
+            } else if ((unsigned)idx < Numbers1.size()) {
                 Numbers.push_back(Numbers1.at(idx));
             } else {
                 Numbers.push_back(Numbers2.at(idx - Numbers1.size()));
@@ -966,8 +967,9 @@ std::vector<int> LateLowerGCFrame::NumberAll(State &S, Value *V) {
             Number = Numbers[CurrentV.second]; // only needed a subset of the values
             Numbers.resize(tracked.count, Number);
         }
-        else
+        else {
             assert(!isa<PointerType>(V->getType()));
+        }
     }
     if (CurrentV.first != V) {
         if (isa<PointerType>(V->getType())) {
@@ -2296,7 +2298,7 @@ bool LateLowerGCFrame::CleanupIR(Function &F, State *S) {
                 // Create a call to the `julia.gc_alloc_bytes` intrinsic, which is like
                 // `julia.gc_alloc_obj` except it doesn't set the tag.
                 auto allocBytesIntrinsic = getOrDeclare(jl_intrinsics::GCAllocBytes);
-                auto ptlsLoad = get_current_ptls_from_task(builder, CI->getArgOperand(0));
+                auto ptlsLoad = get_current_ptls_from_task(builder, CI->getArgOperand(0), tbaa_gcframe);
                 auto ptls = builder.CreateBitCast(ptlsLoad, Type::getInt8PtrTy(builder.getContext()));
                 auto newI = builder.CreateCall(
                     allocBytesIntrinsic,