rewrite catchjmp asm to use normal relocations instead of manual editing

Author: vtjnash

src/cgmemmgr.cpp

@@ -833,28 +833,6 @@ class RTDyldMemoryManagerJL : public SectionMemoryManager {
         mapAddresses(Dyld, ro_alloc);
         mapAddresses(Dyld, exe_alloc);
     }
-#ifdef _OS_WINDOWS_
-    template <typename Alloc>
-    void *lookupWriteAddressFor(void *rt_addr, Alloc &&allocator)
-    {
-        for (auto &alloc: allocator->allocations) {
-            if (alloc.rt_addr == rt_addr) {
-                return alloc.wr_addr;
-            }
-        }
-        return nullptr;
-    }
-    void *lookupWriteAddressFor(void *rt_addr)
-    {
-        if (!ro_alloc)
-            return rt_addr;
-        if (void *ptr = lookupWriteAddressFor(rt_addr, ro_alloc))
-            return ptr;
-        if (void *ptr = lookupWriteAddressFor(rt_addr, exe_alloc))
-            return ptr;
-        return rt_addr;
-    }
-#endif // _OS_WINDOWS_
 };
 
 uint8_t *RTDyldMemoryManagerJL::allocateCodeSection(uintptr_t Size,
@@ -947,13 +925,6 @@ void RTDyldMemoryManagerJL::deregisterEHFrames(uint8_t *Addr,
 
 }
 
-#ifdef _OS_WINDOWS_
-void *lookupWriteAddressFor(RTDyldMemoryManager *memmgr, void *rt_addr)
-{
-    return ((RTDyldMemoryManagerJL*)memmgr)->lookupWriteAddressFor(rt_addr);
-}
-#endif
-
 RTDyldMemoryManager* createRTDyldMemoryManager()
 {
     return new RTDyldMemoryManagerJL();

src/codegen.cpp

@@ -10371,6 +10371,7 @@ static void init_jit_functions(void)
 
 #ifdef _OS_WINDOWS_
 #if defined(_CPU_X86_64_)
+    add_named_global("__julia_personality", &__julia_personality);
 #if defined(_COMPILER_GCC_)
     add_named_global("___chkstk_ms", &___chkstk_ms);
 #else

src/debug-registry.h

@@ -145,8 +145,7 @@ class JITDebugInfoRegistry
     void add_code_in_flight(llvm::StringRef name, jl_code_instance_t *codeinst, const llvm::DataLayout &DL) JL_NOTSAFEPOINT;
     jl_method_instance_t *lookupLinfo(size_t pointer) JL_NOTSAFEPOINT;
     void registerJITObject(const llvm::object::ObjectFile &Object,
-                        std::function<uint64_t(const llvm::StringRef &)> getLoadAddress,
-                        std::function<void*(void*)> lookupWriteAddress);
+                        std::function<uint64_t(const llvm::StringRef &)> getLoadAddress);
     objectmap_t& getObjectMap() JL_NOTSAFEPOINT;
     void add_image_info(image_info_t info) JL_NOTSAFEPOINT;
     bool get_image_info(uint64_t base, image_info_t *info) const JL_NOTSAFEPOINT;

src/debuginfo.cpp

@@ -223,11 +223,21 @@ static void create_PRUNTIME_FUNCTION(uint8_t *Code, size_t Size, StringRef fnnam
 #endif
 
 void JITDebugInfoRegistry::registerJITObject(const object::ObjectFile &Object,
-                        std::function<uint64_t(const StringRef &)> getLoadAddress,
-                        std::function<void*(void*)> lookupWriteAddress)
+                        std::function<uint64_t(const StringRef &)> getLoadAddress)
 {
     object::section_iterator EndSection = Object.section_end();
 
+    bool anyfunctions = false;
+    for (const object::SymbolRef &sym_iter : Object.symbols()) {
+        object::SymbolRef::Type SymbolType = cantFail(sym_iter.getType());
+        if (SymbolType != object::SymbolRef::ST_Function)
+            continue;
+        anyfunctions = true;
+        break;
+    }
+    if (!anyfunctions)
+        return;
+
 #ifdef _CPU_ARM_
     // ARM does not have/use .eh_frame
     uint64_t arm_exidx_addr = 0;
@@ -281,14 +291,13 @@ void JITDebugInfoRegistry::registerJITObject(const object::ObjectFile &Object,
 #if defined(_OS_WINDOWS_)
     uint64_t SectionAddrCheck = 0;
     uint64_t SectionLoadCheck = 0; (void)SectionLoadCheck;
-    uint64_t SectionWriteCheck = 0; (void)SectionWriteCheck;
     uint8_t *UnwindData = NULL;
 #if defined(_CPU_X86_64_)
     uint8_t *catchjmp = NULL;
     for (const object::SymbolRef &sym_iter : Object.symbols()) {
         StringRef sName = cantFail(sym_iter.getName());
         if (sName.equals("__UnwindData") || sName.equals("__catchjmp")) {
-            uint64_t Addr = cantFail(sym_iter.getAddress());
+            uint64_t Addr = cantFail(sym_iter.getAddress()); // offset into object (including section offset)
             auto Section = cantFail(sym_iter.getSection());
             assert(Section != EndSection && Section->isText());
             uint64_t SectionAddr = Section->getAddress();
@@ -300,10 +309,7 @@ void JITDebugInfoRegistry::registerJITObject(const object::ObjectFile &Object,
                         SectionLoadCheck == SectionLoadAddr);
             SectionAddrCheck = SectionAddr;
             SectionLoadCheck = SectionLoadAddr;
-            SectionWriteCheck = SectionLoadAddr;
-            if (lookupWriteAddress)
-                SectionWriteCheck = (uintptr_t)lookupWriteAddress((void*)SectionLoadAddr);
-            Addr += SectionWriteCheck - SectionLoadCheck;
+            Addr += SectionLoadAddr - SectionAddr;
             if (sName.equals("__UnwindData")) {
                 UnwindData = (uint8_t*)Addr;
             }
@@ -314,25 +320,7 @@ void JITDebugInfoRegistry::registerJITObject(const object::ObjectFile &Object,
     }
     assert(catchjmp);
     assert(UnwindData);
-    assert(SectionAddrCheck);
     assert(SectionLoadCheck);
-    assert(!memcmp(catchjmp, "\0\0\0\0\0\0\0\0\0\0\0\0", 12) &&
-            !memcmp(UnwindData, "\0\0\0\0\0\0\0\0\0\0\0\0", 12));
-    catchjmp[0] = 0x48;
-    catchjmp[1] = 0xb8; // mov RAX, QWORD PTR [&__julia_personality]
-    *(uint64_t*)(&catchjmp[2]) = (uint64_t)&__julia_personality;
-    catchjmp[10] = 0xff;
-    catchjmp[11] = 0xe0; // jmp RAX
-    UnwindData[0] = 0x09; // version info, UNW_FLAG_EHANDLER
-    UnwindData[1] = 4;    // size of prolog (bytes)
-    UnwindData[2] = 2;    // count of unwind codes (slots)
-    UnwindData[3] = 0x05; // frame register (rbp) = rsp
-    UnwindData[4] = 4;    // second instruction
-    UnwindData[5] = 0x03; // mov RBP, RSP
-    UnwindData[6] = 1;    // first instruction
-    UnwindData[7] = 0x50; // push RBP
-    *(DWORD*)&UnwindData[8] = (DWORD)(catchjmp - (uint8_t*)SectionWriteCheck); // relative location of catchjmp
-    UnwindData -= SectionWriteCheck - SectionLoadCheck;
 #endif // defined(_OS_X86_64_)
 #endif // defined(_OS_WINDOWS_)
 
@@ -353,7 +341,7 @@ void JITDebugInfoRegistry::registerJITObject(const object::ObjectFile &Object,
         uint64_t SectionAddr = Section->getAddress();
         StringRef secName = cantFail(Section->getName());
         uint64_t SectionLoadAddr = getLoadAddress(secName);
-        Addr -= SectionAddr - SectionLoadAddr;
+        Addr += SectionLoadAddr - SectionAddr;
         StringRef sName = cantFail(sym_iter.getName());
         uint64_t SectionSize = Section->getSize();
         size_t Size = sym_size.second;
@@ -404,10 +392,9 @@ void JITDebugInfoRegistry::registerJITObject(const object::ObjectFile &Object,
 }
 
 void jl_register_jit_object(const object::ObjectFile &Object,
-                            std::function<uint64_t(const StringRef &)> getLoadAddress,
-                            std::function<void *(void *)> lookupWriteAddress)
+                            std::function<uint64_t(const StringRef &)> getLoadAddress)
 {
-    getJITDebugRegistry().registerJITObject(Object, getLoadAddress, lookupWriteAddress);
+    getJITDebugRegistry().registerJITObject(Object, getLoadAddress);
 }
 
 // TODO: convert the safe names from aotcomile.cpp:makeSafeName back into symbols

src/jitlayers.cpp

@@ -665,8 +665,7 @@ static Expected<orc::ThreadSafeModule> selectOptLevel(orc::ThreadSafeModule TSM,
 }
 
 void jl_register_jit_object(const object::ObjectFile &debugObj,
-                            std::function<uint64_t(const StringRef &)> getLoadAddress,
-                            std::function<void *(void *)> lookupWriteAddress);
+                            std::function<uint64_t(const StringRef &)> getLoadAddress);
 
 namespace {
 
@@ -726,7 +725,7 @@ class JLDebuginfoPlugin : public ObjectLinkingLayer::Plugin {
                 return result->second;
             };
 
-            jl_register_jit_object(*NewInfo->Object, getLoadAddress, nullptr);
+            jl_register_jit_object(*NewInfo->Object, getLoadAddress);
             PendingObjs.erase(&MR);
         }
 
@@ -957,10 +956,6 @@ class ForwardingMemoryManager : public RuntimeDyld::MemoryManager {
 };
 
 
-#if defined(_OS_WINDOWS_) && defined(_CPU_X86_64_)
-void *lookupWriteAddressFor(RTDyldMemoryManager *MemMgr, void *rt_addr);
-#endif
-
 void registerRTDyldJITObject(const object::ObjectFile &Object,
                              const RuntimeDyld::LoadedObjectInfo &L,
                              const std::shared_ptr<RTDyldMemoryManager> &MemMgr)
@@ -983,14 +978,7 @@ void registerRTDyldJITObject(const object::ObjectFile &Object,
     };
 
     auto DebugObject = L.getObjectForDebug(Object); // ELF requires us to make a copy to mutate the header with the section load addresses. On other platforms this is a no-op.
-    jl_register_jit_object(DebugObject.getBinary() ? *DebugObject.getBinary() : Object,
-        getLoadAddress,
-#if defined(_OS_WINDOWS_) && defined(_CPU_X86_64_)
-        [MemMgr](void *p) { return lookupWriteAddressFor(MemMgr.get(), p); }
-#else
-        nullptr
-#endif
-    );
+    jl_register_jit_object(DebugObject.getBinary() ? *DebugObject.getBinary() : Object, getLoadAddress);
 }
 namespace {
     static std::unique_ptr<TargetMachine> createTargetMachine() JL_NOTSAFEPOINT {
@@ -1281,8 +1269,9 @@ namespace {
                 }
 
                 // Windows needs some inline asm to help
-                // build unwind tables
-                jl_decorate_module(M);
+                // build unwind tables, if they have any functions to decorate
+                if (!M.functions().empty())
+                    jl_decorate_module(M);
             });
             return std::move(TSM);
         }
@@ -2255,30 +2244,64 @@ static void jl_decorate_module(Module &M) {
     if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) {
         // Add special values used by debuginfo to build the UnwindData table registration for Win64
         // This used to be GV, but with https://reviews.llvm.org/D100944 we no longer can emit GV into `.text`
-        // TODO: The data is set in debuginfo.cpp but it should be okay to actually emit it here.
-        std::string inline_asm = "\
-    .section ";
-        inline_asm +=
+        // and with JITLink it became difficult to change the content afterwards, but we
+        // would prefer that this simple content wasn't recompiled in every single module,
+        // so we emit the necessary PLT trampoline as inline assembly.
+        // This is somewhat duplicated with the .pdata section, but we haven't been able to
+        // use that yet due to relocation issues.
+#define ASM_USES_ELF // use ELF or COFF syntax based on FORCE_ELF
+        StringRef inline_asm(
+    ".section"
 #if JL_LLVM_VERSION >= 180000
-    ".ltext,\"ax\",@progbits";
+        " .ltext,\"ax\",@progbits\n"
 #else
-    ".text";
+        " .text\n"
 #endif
-        inline_asm +=              "\n\
-    .type   __UnwindData,@object    \n\
-    .p2align        2, 0x90         \n\
-    __UnwindData:                   \n\
-        .zero   12                  \n\
-        .size   __UnwindData, 12    \n\
-                                    \n\
-        .type   __catchjmp,@object  \n\
-        .p2align        2, 0x90     \n\
-    __catchjmp:                     \n\
-        .zero   12                  \n\
-        .size   __catchjmp, 12";
-
+    ".globl __julia_personality\n"
+    "\n"
+#ifdef ASM_USES_ELF
+    ".type __UnwindData,@object\n"
+#else
+    ".def __UnwindData\n"
+    ".scl 2\n"
+    ".type 0\n"
+    ".endef\n"
+#endif
+    ".p2align        2, 0x90\n"
+    "__UnwindData:\n"
+    "  .byte 0x09;\n" // version info, UNW_FLAG_EHANDLER
+    "  .byte 4;\n"    // size of prolog (bytes)
+    "  .byte 2;\n"    // count of unwind codes (slots)
+    "  .byte 0x05;\n" // frame register (rbp) = rsp
+    "  .byte 4;\n"    // second instruction
+    "  .byte 0x03;\n" // mov RBP, RSP
+    "  .byte 1;\n"    // first instruction
+    "  .byte 0x50;\n" // push RBP
+    "  .int __catchjmp - "
+#if JL_LLVM_VERSION >= 180000
+    ".ltext;\n" // Section-relative offset (if using COFF and JITLink, this can be relative to __ImageBase instead, though then we could possibly use pdata/xdata directly then)
+#else
+    ".text;\n"
+#endif
+    ".size __UnwindData, 12\n"
+    "\n"
+#ifdef ASM_USES_ELF
+    ".type __catchjmp,@function\n"
+#else
+    ".def __catchjmp\n"
+    ".scl 2\n"
+    ".type 32\n"
+    ".endef\n"
+#endif
+    ".p2align        2, 0x90\n"
+    "__catchjmp:\n"
+    "  movabsq $__julia_personality, %rax\n"
+    "  jmpq *%rax\n"
+    ".size __catchjmp, . - __catchjmp\n"
+    "\n");
         M.appendModuleInlineAsm(inline_asm);
     }
+#undef ASM_USES_ELF
 }
 
 #ifndef JL_USE_JITLINK