compile: make more efficient by discarding internal names (#56702)

These are not user-visible, so this makes the compiler faster and more
efficient with no effort on our part, and avoids duplicating the
debug_level parameter.

src/aotcompile.cpp

@@ -593,9 +593,10 @@ void *jl_create_native_impl(jl_array_t *methods, LLVMOrcThreadSafeModuleRef llvm
     });
     egal_set method_roots;
     jl_codegen_params_t params(ctxt, std::move(target_info.first), std::move(target_info.second));
+    if (!llvmmod)
+        params.getContext().setDiscardValueNames(true);
     params.params = cgparams;
     params.imaging_mode = imaging;
-    params.debug_level = cgparams->debug_info_level;
     params.external_linkage = _external_linkage;
     params.temporary_roots = jl_alloc_array_1d(jl_array_any_type, 0);
     JL_GC_PUSH3(&params.temporary_roots, &method_roots.list, &method_roots.keyset);
@@ -1719,6 +1720,7 @@ static SmallVector<AOTOutputs, 16> add_output(Module &M, TargetMachine &TM, Stri
         for (unsigned i = 0; i < threads; i++) {
             std::function<void()> func = [&, i]() {
                 LLVMContext ctx;
+                ctx.setDiscardValueNames(true);
                 #if JL_LLVM_VERSION < 170000
                 SetOpaquePointer(ctx);
                 #endif
@@ -1930,6 +1932,7 @@ void jl_dump_native_impl(void *native_code,
     if (z) {
         JL_TIMING(NATIVE_AOT, NATIVE_Sysimg);
         LLVMContext Context;
+        Context.setDiscardValueNames(true);
         #if JL_LLVM_VERSION < 170000
         SetOpaquePointer(Context);
         #endif
@@ -2077,6 +2080,7 @@ void jl_dump_native_impl(void *native_code,
     {
         JL_TIMING(NATIVE_AOT, NATIVE_Metadata);
         LLVMContext Context;
+        Context.setDiscardValueNames(true);
         #if JL_LLVM_VERSION < 170000
         SetOpaquePointer(Context);
         #endif
@@ -2278,7 +2282,6 @@ void jl_get_llvmf_defn_impl(jl_llvmf_dump_t* dump, jl_method_instance_t *mi, jl_
         // output.imaging = true;
         // This would also be nice, but it seems to cause OOMs on the windows32 builder
         // To get correct names in the IR this needs to be at least 2
-        output.debug_level = params.debug_info_level;
         output.temporary_roots = jl_alloc_array_1d(jl_array_any_type, 0);
         JL_GC_PUSH1(&output.temporary_roots);
         auto decls = jl_emit_code(m, mi, src, output);

src/ccall.cpp

@@ -851,6 +851,7 @@ static jl_cgval_t emit_llvmcall(jl_codectx_t &ctx, jl_value_t **args, size_t nar
 
     // generate a temporary module that contains our IR
     std::unique_ptr<Module> Mod;
+    bool shouldDiscardValueNames = ctx.builder.getContext().shouldDiscardValueNames();
     Function *f;
     if (entry == NULL) {
         // we only have function IR, which we should put in a function
@@ -878,7 +879,9 @@ static jl_cgval_t emit_llvmcall(jl_codectx_t &ctx, jl_value_t **args, size_t nar
                   << jl_string_data(ir) << "\n}";
 
         SMDiagnostic Err = SMDiagnostic();
+        ctx.builder.getContext().setDiscardValueNames(false);
         Mod = parseAssemblyString(ir_stream.str(), Err, ctx.builder.getContext());
+        ctx.builder.getContext().setDiscardValueNames(shouldDiscardValueNames);
 
         // backwards compatibility: support for IR with integer pointers
         if (!Mod) {
@@ -911,8 +914,9 @@ static jl_cgval_t emit_llvmcall(jl_codectx_t &ctx, jl_value_t **args, size_t nar
                              << jl_string_data(ir) << "\n}";
 
             SMDiagnostic Err = SMDiagnostic();
-            Mod =
-                parseAssemblyString(compat_ir_stream.str(), Err, ctx.builder.getContext());
+            ctx.builder.getContext().setDiscardValueNames(false);
+            Mod = parseAssemblyString(compat_ir_stream.str(), Err, ctx.builder.getContext());
+            ctx.builder.getContext().setDiscardValueNames(shouldDiscardValueNames);
         }
 
         if (!Mod) {
@@ -932,7 +936,9 @@ static jl_cgval_t emit_llvmcall(jl_codectx_t &ctx, jl_value_t **args, size_t nar
 
         if (jl_is_string(ir)) {
             SMDiagnostic Err = SMDiagnostic();
+            ctx.builder.getContext().setDiscardValueNames(false);
             Mod = parseAssemblyString(jl_string_data(ir), Err, ctx.builder.getContext());
+            ctx.builder.getContext().setDiscardValueNames(shouldDiscardValueNames);
             if (!Mod) {
                 std::string message = "Failed to parse LLVM assembly: \n";
                 raw_string_ostream stream(message);

src/codegen.cpp

@@ -167,31 +167,29 @@ void setName(jl_codegen_params_t &params, Value *V, const Twine &Name)
     // is not checking that setName is only called for non-folded instructions (e.g. folded bitcasts
     // and 0-byte geps), which can result in information loss on the renamed instruction.
     assert((isa<Constant>(V) || isa<Instruction>(V)) && "Should only set names on instructions!");
-    if (params.debug_level >= 2 && !isa<Constant>(V)) {
+    if (!isa<Constant>(V)) {
         V->setName(Name);
     }
 }
 
 void maybeSetName(jl_codegen_params_t &params, Value *V, const Twine &Name)
 {
     // To be used when we may get an Instruction or something that is not an instruction i.e Constants/Arguments
-    if (params.debug_level >= 2 && isa<Instruction>(V)) {
+    if (isa<Instruction>(V))
         V->setName(Name);
-    }
 }
 
 void setName(jl_codegen_params_t &params, Value *V, std::function<std::string()> GetName)
 {
     assert((isa<Constant>(V) || isa<Instruction>(V)) && "Should only set names on instructions!");
-    if (params.debug_level >= 2 && !isa<Constant>(V)) {
+    if (!params.getContext().shouldDiscardValueNames() && !isa<Constant>(V))
         V->setName(Twine(GetName()));
-    }
 }
 
 void setNameWithField(jl_codegen_params_t &params, Value *V, std::function<StringRef()> GetObjName, jl_datatype_t *jt, unsigned idx, const Twine &suffix)
 {
     assert((isa<Constant>(V) || isa<Instruction>(V)) && "Should only set names on instructions!");
-    if (params.debug_level >= 2 && !isa<Constant>(V)) {
+    if (!params.getContext().shouldDiscardValueNames() && !isa<Constant>(V)) {
         if (jl_is_tuple_type(jt)){
             V->setName(Twine(GetObjName()) + "[" + Twine(idx + 1) + "]"+ suffix);
             return;
@@ -8327,7 +8325,7 @@ static jl_returninfo_t get_specsig_function(jl_codectx_t &ctx, Module *M, Value
         if (f == NULL) {
             f = Function::Create(ftype, GlobalVariable::ExternalLinkage, name, M);
             jl_init_function(f, ctx.emission_context.TargetTriple);
-            if (ctx.emission_context.debug_level >= 2) {
+            if (ctx.emission_context.params->debug_info_level >= 2) {
                 ios_t sigbuf;
                 ios_mem(&sigbuf, 0);
                 jl_static_show_func_sig((JL_STREAM*) &sigbuf, sig);
@@ -8435,7 +8433,7 @@ static jl_llvm_functions_t
     std::map<int, BasicBlock*> labels;
     ctx.module = jl_is_method(lam->def.method) ? lam->def.method->module : lam->def.module;
     ctx.linfo = lam;
-    ctx.name = TSM.getModuleUnlocked()->getModuleIdentifier().data();
+    ctx.name = name_from_method_instance(lam);
     size_t nreq = src->nargs;
     int va = src->isva;
     ctx.nargs = nreq;
@@ -8488,7 +8486,7 @@ static jl_llvm_functions_t
     // jl_printf(JL_STDERR, "\n*** compiling %s at %s:%d\n\n",
     //           jl_symbol_name(ctx.name), ctx.file.str().c_str(), toplineno);
 
-    bool debug_enabled = ctx.emission_context.debug_level != 0;
+    bool debug_enabled = ctx.emission_context.params->debug_info_level != 0;
     if (dbgFuncName.empty()) // Should never happen anymore?
         debug_enabled = false;
 
@@ -8564,15 +8562,14 @@ static jl_llvm_functions_t
     // allocate Function declarations and wrapper objects
     //Safe because params holds ctx lock
     Module *M = TSM.getModuleUnlocked();
-    M->addModuleFlag(Module::Warning, "julia.debug_level", ctx.emission_context.debug_level);
     jl_debugcache_t debugcache;
     debugcache.initialize(M);
     jl_returninfo_t returninfo = {};
     Function *f = NULL;
     bool has_sret = false;
     if (specsig) { // assumes !va and !needsparams
         SmallVector<const char*,0> ArgNames(0);
-        if (ctx.emission_context.debug_level >= 2) {
+        if (!M->getContext().shouldDiscardValueNames()) {
             ArgNames.resize(ctx.nargs, "");
             for (int i = 0; i < ctx.nargs; i++) {
                 jl_sym_t *argname = slot_symbol(ctx, i);
@@ -8639,7 +8636,7 @@ static jl_llvm_functions_t
         declarations.functionObject = needsparams ? "jl_fptr_sparam" : "jl_fptr_args";
     }
 
-    if (ctx.emission_context.debug_level >= 2 && lam->def.method && jl_is_method(lam->def.method) && lam->specTypes != (jl_value_t*)jl_emptytuple_type) {
+    if (!params.getContext().shouldDiscardValueNames() && ctx.emission_context.params->debug_info_level >= 2 && lam->def.method && jl_is_method(lam->def.method) && lam->specTypes != (jl_value_t*)jl_emptytuple_type) {
         ios_t sigbuf;
         ios_mem(&sigbuf, 0);
         jl_static_show_func_sig((JL_STREAM*) &sigbuf, (jl_value_t*)lam->specTypes);
@@ -8694,7 +8691,7 @@ static jl_llvm_functions_t
     if (debug_enabled) {
         topfile = dbuilder.createFile(ctx.file, ".");
         DISubroutineType *subrty;
-        if (ctx.emission_context.debug_level <= 1)
+        if (ctx.emission_context.params->debug_info_level <= 1)
             subrty = debugcache.jl_di_func_null_sig;
         else if (!specsig)
             subrty = debugcache.jl_di_func_sig;
@@ -8715,7 +8712,7 @@ static jl_llvm_functions_t
                                      );
         topdebugloc = DILocation::get(ctx.builder.getContext(), toplineno, 0, SP, NULL);
         f->setSubprogram(SP);
-        if (ctx.emission_context.debug_level >= 2) {
+        if (ctx.emission_context.params->debug_info_level >= 2) {
             const bool AlwaysPreserve = true;
             // Go over all arguments and local variables and initialize their debug information
             for (i = 0; i < nreq; i++) {
@@ -10161,7 +10158,7 @@ jl_llvm_functions_t jl_emit_codeinst(
             if (// keep code when keeping everything
                 !(JL_DELETE_NON_INLINEABLE) ||
                 // aggressively keep code when debugging level >= 2
-                // note that this uses the global jl_options.debug_level, not the local emission_ctx.debug_level
+                // note that this uses the global jl_options.debug_level, not the local emission_ctx.debug_info_level
                 jl_options.debug_level > 1) {
                 // update the stored code
                 if (inferred != (jl_value_t*)src) {

src/jitlayers.cpp

@@ -690,9 +690,9 @@ static void jl_emit_codeinst_to_jit(
     JL_TIMING(CODEINST_COMPILE, CODEINST_COMPILE);
     // emit the code in LLVM IR form to the new context
     jl_codegen_params_t params(std::make_unique<LLVMContext>(), jl_ExecutionEngine->getDataLayout(), jl_ExecutionEngine->getTargetTriple()); // Locks the context
+    params.getContext().setDiscardValueNames(true);
     params.cache = true;
     params.imaging_mode = imaging_default();
-    params.debug_level = jl_options.debug_level;
     orc::ThreadSafeModule result_m =
         jl_create_ts_module(name_from_method_instance(codeinst->def), params.tsctx, params.DL, params.TargetTriple);
     params.temporary_roots = jl_alloc_array_1d(jl_array_any_type, 0);
@@ -795,9 +795,10 @@ int jl_compile_extern_c_impl(LLVMOrcThreadSafeModuleRef llvmmod, void *p, void *
         Module *M = into->getModuleUnlocked();
         jl_codegen_params_t params(into->getContext(), M->getDataLayout(), Triple(M->getTargetTriple()));
         params.imaging_mode = imaging_default();
-        params.debug_level = jl_options.debug_level;
-        if (pparams == NULL)
+        if (pparams == NULL) {
+            M->getContext().setDiscardValueNames(true);
             pparams = &params;
+        }
         assert(pparams->tsctx.getContext() == into->getContext().getContext());
         const char *name = jl_generate_ccallable(wrap(into), sysimg, declrt, sigt, *pparams);
         if (!sysimg) {

src/jitlayers.h

@@ -259,7 +259,6 @@ struct jl_codegen_params_t {
     bool cache = false;
     bool external_linkage = false;
     bool imaging_mode;
-    int debug_level;
     bool use_swiftcc = true;
     jl_codegen_params_t(orc::ThreadSafeContext ctx, DataLayout DL, Triple triple)
       : tsctx(std::move(ctx)),

src/llvm-late-gc-lowering.cpp

@@ -1944,36 +1944,27 @@ void LateLowerGCFrame::CleanupWriteBarriers(Function &F, State *S, const SmallVe
         if (CFGModified) {
             *CFGModified = true;
         }
-        auto DebugInfoMeta = F.getParent()->getModuleFlag("julia.debug_level");
-        int debug_info = 1;
-        if (DebugInfoMeta != nullptr) {
-            debug_info = cast<ConstantInt>(cast<ConstantAsMetadata>(DebugInfoMeta)->getValue())->getZExtValue();
-        }
 
         IRBuilder<> builder(CI);
         builder.SetCurrentDebugLocation(CI->getDebugLoc());
-        auto parBits = builder.CreateAnd(EmitLoadTag(builder, T_size, parent), GC_OLD_MARKED);
-        setName(parBits, "parent_bits", debug_info);
-        auto parOldMarked = builder.CreateICmpEQ(parBits, ConstantInt::get(T_size, GC_OLD_MARKED));
-        setName(parOldMarked, "parent_old_marked", debug_info);
+        auto parBits = builder.CreateAnd(EmitLoadTag(builder, T_size, parent), GC_OLD_MARKED, "parent_bits");
+        auto parOldMarked = builder.CreateICmpEQ(parBits, ConstantInt::get(T_size, GC_OLD_MARKED), "parent_old_marked");
         auto mayTrigTerm = SplitBlockAndInsertIfThen(parOldMarked, CI, false);
         builder.SetInsertPoint(mayTrigTerm);
-        setName(mayTrigTerm->getParent(), "may_trigger_wb", debug_info);
+        mayTrigTerm->getParent()->setName("may_trigger_wb");
         Value *anyChldNotMarked = NULL;
         for (unsigned i = 1; i < CI->arg_size(); i++) {
             Value *child = CI->getArgOperand(i);
-            Value *chldBit = builder.CreateAnd(EmitLoadTag(builder, T_size, child), GC_MARKED);
-            setName(chldBit, "child_bit", debug_info);
-            Value *chldNotMarked = builder.CreateICmpEQ(chldBit, ConstantInt::get(T_size, 0),"child_not_marked");
-            setName(chldNotMarked, "child_not_marked", debug_info);
+            Value *chldBit = builder.CreateAnd(EmitLoadTag(builder, T_size, child), GC_MARKED, "child_bit");
+            Value *chldNotMarked = builder.CreateICmpEQ(chldBit, ConstantInt::get(T_size, 0), "child_not_marked");
             anyChldNotMarked = anyChldNotMarked ? builder.CreateOr(anyChldNotMarked, chldNotMarked) : chldNotMarked;
         }
         assert(anyChldNotMarked); // handled by all_of test above
         MDBuilder MDB(parent->getContext());
         SmallVector<uint32_t, 2> Weights{1, 9};
         auto trigTerm = SplitBlockAndInsertIfThen(anyChldNotMarked, mayTrigTerm, false,
                                                   MDB.createBranchWeights(Weights));
-        setName(trigTerm->getParent(), "trigger_wb", debug_info);
+        trigTerm->getParent()->setName("trigger_wb");
         builder.SetInsertPoint(trigTerm);
         if (CI->getCalledOperand() == write_barrier_func) {
             builder.CreateCall(getOrDeclare(jl_intrinsics::queueGCRoot), parent);

src/llvm-pass-helpers.cpp

@@ -335,10 +335,3 @@ namespace jl_well_known {
             return addGCAllocAttributes(allocTypedFunc);
         });
 }
-
-void setName(llvm::Value *V, const llvm::Twine &Name, int debug_info)
-{
-    if (debug_info >= 2 && !llvm::isa<llvm::Constant>(V)) {
-        V->setName(Name);
-    }
-}