Add LLVM level allocation optimization pass

This can obtain escape information with much higher precision than what we can currently do
in typeinf. However, it does not replace the alloc_elim_pass! in type inference either since
this cannot handle objects with reference fields.

Fix #20452

Author: yuyichao

src/Makefile

@@ -50,7 +50,7 @@ endif
 LLVMLINK :=
 
 ifeq ($(JULIACODEGEN),LLVM)
-SRCS += codegen jitlayers disasm debuginfo llvm-simdloop llvm-ptls llvm-late-gc-lowering llvm-lower-handlers llvm-gc-invariant-verifier llvm-propagate-addrspaces cgmemmgr
+SRCS += codegen jitlayers disasm debuginfo llvm-simdloop llvm-ptls llvm-late-gc-lowering llvm-lower-handlers llvm-gc-invariant-verifier llvm-propagate-addrspaces llvm-alloc-opt cgmemmgr
 FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --includedir)
 LLVM_LIBS := all
 ifeq ($(USE_POLLY),1)

src/ccall.cpp

@@ -2143,7 +2143,7 @@ jl_cgval_t function_sig_t::emit_a_ccall(
                 size_t rtsz = jl_datatype_size(rt);
                 assert(rtsz > 0);
                 Value *strct = emit_allocobj(ctx, rtsz, runtime_bt);
-                int boxalign = jl_gc_alignment(rtsz);
+                int boxalign = jl_datatype_align(rt);
 #ifndef JL_NDEBUG
 #if JL_LLVM_VERSION >= 40000
                 const DataLayout &DL = jl_data_layout;

src/cgutils.cpp

@@ -2149,25 +2149,10 @@ static Value *emit_allocobj(jl_codectx_t &ctx, size_t static_size, Value *jt)
 {
     JL_FEAT_REQUIRE(ctx, dynamic_alloc);
     JL_FEAT_REQUIRE(ctx, runtime);
-
-    int osize;
-    int offset = jl_gc_classify_pools(static_size, &osize);
     Value *ptls_ptr = emit_bitcast(ctx, ctx.ptlsStates, T_pint8);
-    Value *v;
-    if (offset < 0) {
-        Value *args[] = {ptls_ptr,
-                         ConstantInt::get(T_size, static_size + sizeof(void*))};
-        v = ctx.builder.CreateCall(prepare_call(jlalloc_big_func),
-                               ArrayRef<Value*>(args, 2));
-    }
-    else {
-        Value *pool_offs = ConstantInt::get(T_int32, offset);
-        Value *args[] = {ptls_ptr, pool_offs, ConstantInt::get(T_int32, osize)};
-        v = ctx.builder.CreateCall(prepare_call(jlalloc_pool_func),
-                               ArrayRef<Value*>(args, 3));
-    }
-    tbaa_decorate(tbaa_tag, ctx.builder.CreateStore(maybe_decay_untracked(jt), emit_typeptr_addr(ctx, v)));
-    return v;
+    return ctx.builder.CreateCall(prepare_call(jl_alloc_obj_func),
+                                  {ptls_ptr, ConstantInt::get(T_size, static_size),
+                                          maybe_decay_untracked(jt)});
 }
 
 // if ptr is NULL this emits a write barrier _back_

src/codegen.cpp

@@ -315,8 +315,7 @@ static Function *jlgenericfunction_func;
 static Function *jlenter_func;
 static Function *jlleave_func;
 static Function *jlegal_func;
-static Function *jlalloc_pool_func;
-static Function *jlalloc_big_func;
+static Function *jl_alloc_obj_func;
 static Function *jlisa_func;
 static Function *jlsubtype_func;
 static Function *jlapplytype_func;
@@ -6635,24 +6634,19 @@ static void init_julia_llvm_env(Module *m)
                          "jl_instantiate_type_in_env", m);
     add_named_global(jlapplytype_func, &jl_instantiate_type_in_env);
 
-    std::vector<Type*> alloc_pool_args(0);
-    alloc_pool_args.push_back(T_pint8);
-    alloc_pool_args.push_back(T_int32);
-    alloc_pool_args.push_back(T_int32);
-    jlalloc_pool_func =
-        Function::Create(FunctionType::get(T_prjlvalue, alloc_pool_args, false),
-                         Function::ExternalLinkage,
-                         "jl_gc_pool_alloc", m);
-    add_named_global(jlalloc_pool_func, &jl_gc_pool_alloc);
-
-    std::vector<Type*> alloc_big_args(0);
-    alloc_big_args.push_back(T_pint8);
-    alloc_big_args.push_back(T_size);
-    jlalloc_big_func =
-        Function::Create(FunctionType::get(T_prjlvalue, alloc_big_args, false),
-                         Function::ExternalLinkage,
-                         "jl_gc_big_alloc", m);
-    add_named_global(jlalloc_big_func, &jl_gc_big_alloc);
+    std::vector<Type*> gc_alloc_args(0);
+    gc_alloc_args.push_back(T_pint8);
+    gc_alloc_args.push_back(T_size);
+    gc_alloc_args.push_back(T_prjlvalue);
+    jl_alloc_obj_func = Function::Create(FunctionType::get(T_prjlvalue, gc_alloc_args, false),
+                                         Function::ExternalLinkage,
+                                         "julia.gc_alloc_obj");
+#if JL_LLVM_VERSION >= 50000
+    jl_alloc_obj_func->addAttribute(AttributeList::ReturnIndex, Attribute::NoAlias);
+#else
+    jl_alloc_obj_func->addAttribute(AttributeSet::ReturnIndex, Attribute::NoAlias);
+#endif
+    add_named_global(jl_alloc_obj_func, (void*)NULL, /*dllimport*/false);
 
     std::vector<Type *> dlsym_args(0);
     dlsym_args.push_back(T_pint8);

src/intrinsics.cpp

@@ -325,11 +325,7 @@ static Value *emit_unbox(jl_codectx_t &ctx, Type *to, const jl_cgval_t &x, jl_va
     }
 
     int alignment;
-    if (x.isboxed) {
-        // julia's gc gives 16-byte aligned addresses
-        alignment = 16;
-    }
-    else if (jt) {
+    if (jt) {
         alignment = julia_alignment(p, jt, 0);
     }
     else {

src/jitlayers.cpp

@@ -108,6 +108,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level)
 #endif
     if (opt_level == 0) {
         PM->add(createCFGSimplificationPass()); // Clean up disgusting code
+        PM->add(createAllocOptPass(false));
 #if JL_LLVM_VERSION < 50000
         PM->add(createBarrierNoopPass());
         PM->add(createLowerExcHandlersPass());
@@ -147,6 +148,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level)
     // effectiveness of the optimization, but should retain correctness.
 #if JL_LLVM_VERSION < 50000
     PM->add(createLowerExcHandlersPass());
+    PM->add(createAllocOptPass(true));
     PM->add(createLateLowerGCFramePass());
     // Remove dead use of ptls
     PM->add(createDeadCodeEliminationPass());
@@ -161,6 +163,12 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level)
     PM->add(createAlwaysInlinerPass()); // Respect always_inline
 #endif
 
+#if JL_LLVM_VERSION >= 50000
+    // Running `memcpyopt` between this and `sroa` seems to give `sroa` a hard time
+    // merging the `alloca` for the unboxed data and the `alloca` created by the `alloc_opt`
+    // pass.
+    PM->add(createAllocOptPass(true));
+#endif
     PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
     PM->add(createSROAPass());                 // Break up aggregate allocas
     PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.

src/jitlayers.h

@@ -206,6 +206,7 @@ Pass *createLateLowerGCFramePass();
 Pass *createLowerExcHandlersPass();
 Pass *createGCInvariantVerifierPass(bool Strong);
 Pass *createPropagateJuliaAddrspaces();
+Pass *createAllocOptPass(bool);
 // Whether the Function is an llvm or julia intrinsic.
 static inline bool isIntrinsicFunction(Function *F)
 {