Merge pull request #7391 from kalsprite/codegen_worker_exit

codegen worker must not exit process with running siblings

src/llvm_backend.cpp

@@ -2003,7 +2003,8 @@ gb_internal void lb_verify_function(lbModule *m, lbProcedure *p, bool dump_ll=fa
 			}
 		}
 		LLVMVerifyFunction(p->value, LLVMPrintMessageAction);
-		exit_with_errors();
+		lb_record_worker_failure();
+		return;
 	}
 }
 
@@ -2023,11 +2024,11 @@ gb_internal WORKER_TASK_PROC(lb_llvm_module_verification_worker_proc) {
 			String filepath_ll = lb_filepath_ll_for_module(m);
 			if (LLVMPrintModuleToFile(m->mod, cast(char const *)filepath_ll.text, &llvm_error)) {
 				gb_printf_err("LLVM Error: %s\n", llvm_error);
-				exit_with_errors();
-				return false;
+				lb_record_worker_failure();
+				return 1;
 			}
 		}
-		exit_with_errors();
+		lb_record_worker_failure();
 		return 1;
 	}
 	return 0;
@@ -2390,11 +2391,13 @@ gb_internal WORKER_TASK_PROC(lb_llvm_emit_worker_proc) {
 	if (build_context.lto_kind != LTO_None) {
 		if (LLVMWriteBitcodeToFile(wd->m->mod, cast(char *)wd->filepath_obj.text)) {
 			gb_printf_err("Failed to write bitcode file: %.*s\n", LIT(wd->filepath_obj));
-			exit_with_errors();
+			lb_record_worker_failure();
+			return 1;
 		}
 	} else if (LLVMTargetMachineEmitToFile(wd->target_machine, wd->m->mod, cast(char *)wd->filepath_obj.text, wd->code_gen_file_type, &llvm_error)) {
 		gb_printf_err("LLVM Error: %s\n", llvm_error);
-		exit_with_errors();
+		lb_record_worker_failure();
+		return 1;
 	}
 	debugf("Generated File: %.*s\n", LIT(wd->filepath_obj));
 	return 0;
@@ -2561,7 +2564,7 @@ gb_internal WORKER_TASK_PROC(lb_llvm_module_pass_worker_proc) {
 				gb_printf_err("LLVM Error: %s\n", llvm_error);
 			}
 		}
-		exit_with_errors();
+		lb_record_worker_failure();
 		return 1;
 	}
 
@@ -2602,6 +2605,8 @@ gb_internal void lb_generate_procedures(lbGenerator *gen, bool do_threading) {
 			lb_generate_procedures_worker_proc(m);
 		}
 	}
+
+	lb_exit_if_worker_failed();
 }
 
 gb_internal WORKER_TASK_PROC(lb_generate_missing_procedures_to_check_worker_proc) {
@@ -2674,6 +2679,8 @@ gb_internal void lb_llvm_function_passes(lbGenerator *gen, bool do_threading) {
 			lb_llvm_function_pass_per_module(m);
 		}
 	}
+
+	lb_exit_if_worker_failed();
 }
 
 
@@ -2699,6 +2706,8 @@ gb_internal void lb_llvm_module_passes_and_verification(lbGenerator *gen, bool d
 			lb_llvm_module_pass_worker_proc(wd);
 		}
 	}
+
+	lb_exit_if_worker_failed();
 }
 
 gb_internal String lb_filepath_ll_for_module(lbModule *m) {
@@ -2829,6 +2838,7 @@ gb_internal bool lb_llvm_object_generation(lbGenerator *gen, bool do_threading)
 		}
 
 		thread_pool_wait(&global_thread_pool);
+		lb_exit_if_worker_failed();
 	} else {
 		for (auto const &entry : gen->modules) {
 			lbModule *m = entry.value;

src/llvm_backend_general.cpp

@@ -16,6 +16,41 @@ gb_global isize lb_global_type_info_member_offsets_index = 0;
 gb_global isize lb_global_type_info_member_usings_index  = 0;
 gb_global isize lb_global_type_info_member_tags_index    = 0;
 
+// A backend worker must not end the process: its siblings are still inside LLVM, and tearing the
+// process down under them is what turns a reported error into a crash. A failing worker records the
+// failure and returns; the driver exits once the pool has drained. Work stealing can run a task on
+// the main thread, so this must not depend on which thread is executing
+gb_global std::atomic<bool> lb_worker_failure;
+
+gb_internal void lb_record_worker_failure(void) {
+	lb_worker_failure.store(true, std::memory_order_release);
+}
+
+gb_internal void lb_exit_if_worker_failed(void) {
+	if (lb_worker_failure.load(std::memory_order_acquire)) {
+		exit_with_errors();
+	}
+}
+
+// Without a handler installed, LLVM prints an error of its own and calls exit(1) from whichever
+// thread it is on. 
+gb_internal void lb_llvm_diagnostic_handler(LLVMDiagnosticInfoRef di, void *) {
+	char *description = LLVMGetDiagInfoDescription(di);
+	defer (LLVMDisposeMessage(description));
+
+	switch (LLVMGetDiagInfoSeverity(di)) {
+	case LLVMDSError:
+		gb_printf_err("LLVM Error: %s\n", description);
+		lb_record_worker_failure();
+		break;
+	case LLVMDSWarning:
+		gb_printf_err("LLVM Warning: %s\n", description);
+		break;
+	default:
+		break;
+	}
+}
+
 gb_internal WORKER_TASK_PROC(lb_init_module_worker_proc) {
 	lbModule *m = cast(lbModule *)data;
 	Checker *c = m->checker;
@@ -58,6 +93,7 @@ gb_internal WORKER_TASK_PROC(lb_init_module_worker_proc) {
 
 	m->module_name = module_name;
 	m->ctx = LLVMContextCreate();
+	LLVMContextSetDiagnosticHandler(m->ctx, lb_llvm_diagnostic_handler, nullptr);
 	m->mod = LLVMModuleCreateWithNameInContext(m->module_name, m->ctx);
 	// m->debug_builder = nullptr;
 	if (build_context.no_plt) {