Towards being able to run GC frame lowering after LLVM optimizations

* Make the pass more robust against dead code elimination
* Remove module level setup code (so that the the GC frame lowering can be
  a `FunctionPass`)
* Create PTLS lowering (module) pass since it needs to be run after the GC frame
  lowering
* Set DLLImport attribute for `jl_tls_states` on windows,
  since it is not handled by module merger anymore

src/Makefile

@@ -40,7 +40,7 @@ SRCS := \
  alloc dlload sys init task array dump toplevel jl_uv \
  simplevector APInt-C runtime_intrinsics runtime_ccall \
  threadgroup threading stackwalk gc gc-debug gc-pages \
-  jlapi signal-handling safepoint jloptions timing
+ jlapi signal-handling safepoint jloptions timing
 
 ifeq ($(USEMSVC), 1)
 SRCS += getopt
@@ -49,7 +49,7 @@ endif
 LLVMLINK :=
 
 ifeq ($(JULIACODEGEN),LLVM)
-SRCS += codegen jitlayers disasm debuginfo llvm-simdloop llvm-gcroot cgmemmgr
+SRCS += codegen jitlayers disasm debuginfo llvm-simdloop llvm-ptls llvm-gcroot cgmemmgr
 FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --includedir)
 LLVM_LIBS := all
 ifeq ($(USE_POLLY),1)

src/codegen.cpp

@@ -85,8 +85,7 @@
 #include <llvm/Support/PrettyStackTrace.h>
 #include <llvm/Support/CommandLine.h>
 
-#if defined(_CPU_ARM_) || defined(_CPU_AARCH64_) || \
- (defined(LLVM37) && defined(JULIA_ENABLE_THREADING))
+#if defined(_CPU_ARM_) || defined(_CPU_AARCH64_)
 # include <llvm/IR/InlineAsm.h>
 #endif
 #if defined(USE_POLLY)
@@ -257,7 +256,7 @@ static MDNode *tbaa_arrayptr; // The pointer inside a jl_array_t
 static MDNode *tbaa_arraysize; // A size in a jl_array_t
 static MDNode *tbaa_arraylen; // The len in a jl_array_t
 static MDNode *tbaa_arrayflags; // The flags in a jl_array_t
-static MDNode *tbaa_const; // Memory that is immutable by the time LLVM can see it
+MDNode *tbaa_const;  // Memory that is immutable by the time LLVM can see it
 
 // Basic DITypes
 #ifdef LLVM37
@@ -388,7 +387,6 @@ static Function *jlarray_data_owner_func;
 
 // placeholder functions
 static Function *gcroot_func;
-static Function *gcstore_func;
 static Function *gckill_func;
 static Function *jlcall_frame_func;
 static Function *gcroot_flush_func;
@@ -3399,87 +3397,15 @@ static void allocate_gc_frame(BasicBlock *b0, jl_codectx_t *ctx)
  PointerType::get(T_psize, 0));
 }
 
-void jl_codegen_finalize_temp_arg(CallInst *ptlsStates, Type *T_pjlvalue,
- MDNode *tbaa_gcframe);
-static void finalize_gc_frame(Function *F)
-{
- Module *M = F->getParent();
- M->getOrInsertFunction(gcroot_func->getName(), gcroot_func->getFunctionType());
- M->getOrInsertFunction(gckill_func->getName(), gckill_func->getFunctionType());
- M->getOrInsertFunction(gcstore_func->getName(), gcstore_func->getFunctionType());
- M->getOrInsertFunction(jlcall_frame_func->getName(), jlcall_frame_func->getFunctionType());
- M->getOrInsertFunction(gcroot_flush_func->getName(), gcroot_flush_func->getFunctionType());
- Function *jl_get_ptls_states = M->getFunction("jl_get_ptls_states");
-
- CallInst *ptlsStates = NULL;
- for (BasicBlock::iterator i = F->getEntryBlock().begin(), e = F->getEntryBlock().end(); i != e; ++i) {
- if (CallInst *callInst = dyn_cast<CallInst>(&*i)) {
- if (callInst->getCalledFunction() == jl_get_ptls_states) {
- ptlsStates = callInst;
- break;
- }
- }
- }
- if (!ptlsStates)
- return;
-
- jl_codegen_finalize_temp_arg(ptlsStates, T_pjlvalue, tbaa_gcframe);
-
-#ifdef JULIA_ENABLE_THREADING
- if (imaging_mode) {
- GlobalVariable *GV = prepare_global(jltls_states_func_ptr, M);
- Value *getter = tbaa_decorate(tbaa_const,
- new LoadInst(GV, "", ptlsStates));
- ptlsStates->setCalledFunction(getter);
- ptlsStates->setAttributes(jltls_states_func->getAttributes());
- }
- else if (jl_tls_offset != -1) {
-#ifdef LLVM37
- // Replace the function call with inline assembly if we know
- // how to generate it.
- const char *asm_str = nullptr;
-# if defined(_CPU_X86_64_)
- asm_str = "movq %fs:0, $0";
-# elif defined(_CPU_X86_)
- asm_str = "movl %gs:0, $0";
-# elif defined(_CPU_AARCH64_)
- asm_str = "mrs $0, tpidr_el0";
-# endif
- assert(asm_str && "Cannot emit thread pointer for this architecture.");
- static auto offset = ConstantInt::getSigned(T_size, jl_tls_offset);
- static auto tp = InlineAsm::get(FunctionType::get(T_pint8, false),
- asm_str, "=r", false);
- Value *tls = CallInst::Create(tp, "thread_ptr", ptlsStates);
- tls = GetElementPtrInst::Create(T_int8, tls, {offset},
- "ptls_i8", ptlsStates);
- tls = new BitCastInst(tls, PointerType::get(T_ppjlvalue, 0),
- "ptls", ptlsStates);
- ptlsStates->replaceAllUsesWith(tls);
- ptlsStates->eraseFromParent();
-#endif
- }
-#else
- ptlsStates->replaceAllUsesWith(prepare_global(jltls_states_var, M));
- ptlsStates->eraseFromParent();
-#endif
-}
-
+void jl_codegen_finalize_temp_arg(Function *F, MDNode *tbaa_gcframe);
 void finalize_gc_frame(Module *m)
 {
  for (Module::iterator I = m->begin(), E = m->end(); I != E; ++I) {
  Function *F = &*I;
  if (F->isDeclaration())
  continue;
- finalize_gc_frame(F);
+ jl_codegen_finalize_temp_arg(F, tbaa_gcframe);
  }
-#ifndef JULIA_ENABLE_THREADING
- m->getFunction("jl_get_ptls_states")->eraseFromParent();
-#endif
- m->getFunction("julia.gc_root_decl")->eraseFromParent();
- m->getFunction("julia.gc_root_kill")->eraseFromParent();
- m->getFunction("julia.gc_store")->eraseFromParent();
- m->getFunction("julia.jlcall_frame_decl")->eraseFromParent();
- m->getFunction("julia.gcroot_flush")->eraseFromParent();
 }
 
 static Function *gen_cfun_wrapper(jl_function_t *ff, jl_value_t *jlrettype, jl_tupletype_t *argt,
@@ -5186,12 +5112,6 @@ static void init_julia_llvm_env(Module *m)
  jltls_states_func = Function::Create(FunctionType::get(PointerType::get(T_ppjlvalue, 0), false),
  Function::ExternalLinkage,
  "jl_get_ptls_states", m);
- jltls_states_func->setAttributes(
- jltls_states_func->getAttributes()
- .addAttribute(jltls_states_func->getContext(),
- AttributeSet::FunctionIndex, Attribute::ReadNone)
- .addAttribute(jltls_states_func->getContext(),
- AttributeSet::FunctionIndex, Attribute::NoUnwind));
  add_named_global(jltls_states_func, jl_get_ptls_states_getter());
  if (imaging_mode) {
  PointerType *pfunctype = jltls_states_func->getFunctionType()->getPointerTo();
@@ -5608,13 +5528,6 @@ static void init_julia_llvm_env(Module *m)
  "julia.gc_root_kill", m);
  add_named_global(gckill_func, (void*)NULL, /*dllimport*/false);
 
- Type* gc_store_args[2] = { T_ppjlvalue, T_pjlvalue }; // [1] <= [2]
- gcstore_func =
- Function::Create(FunctionType::get(T_void, makeArrayRef(gc_store_args), false),
- Function::ExternalLinkage,
- "julia.gc_store", m);
- add_named_global(gcstore_func, (void*)NULL, /*dllimport*/false);
-
  jlcall_frame_func =
  Function::Create(FunctionType::get(T_ppjlvalue, ArrayRef<Type*>(T_int32), false),
  Function::ExternalLinkage,

src/jitlayers.cpp

@@ -114,6 +114,7 @@ void addOptimizationPasses(PassManager *PM)
 # endif
 #endif
  if (jl_options.opt_level == 0) {
+ PM->add(createLowerPTLSPass(imaging_mode, tbaa_const));
  return;
  }
 #ifdef LLVM37
@@ -140,6 +141,9 @@ void addOptimizationPasses(PassManager *PM)
 #ifndef INSTCOMBINE_BUG
  PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
 #endif
+ // Let the InstCombine pass remove the unnecessary load of
+ // safepoint address first
+ PM->add(createLowerPTLSPass(imaging_mode, tbaa_const));
  PM->add(createSROAPass()); // Break up aggregate allocas
 #ifndef INSTCOMBINE_BUG
  PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.

src/jitlayers.h

@@ -218,3 +218,7 @@ JL_DLLEXPORT extern LLVMContext jl_LLVMContext;
 #else
 JL_DLLEXPORT extern LLVMContext &jl_LLVMContext;
 #endif
+
+extern MDNode *tbaa_const;
+
+Pass *createLowerPTLSPass(bool imaging_mode, MDNode *tbaa_const);