Merge pull request #875 from atkurtul/master

Run an early memcpy pass regardless of the opt flag and return large structs by pointer on linux

src/build_settings.cpp

@@ -873,7 +873,7 @@ void init_build_context(TargetMetrics *cross_target) {
 		//   -memcpyopt: MemCpy optimization
 	}
 	if (bc->ODIN_DEBUG == false) {
-		opt_flags = gb_string_appendc(opt_flags, "-mem2reg -memcpyopt -die ");
+		opt_flags = gb_string_appendc(opt_flags, "-mem2reg -die ");
 	}
 
 

src/check_type.cpp

@@ -2423,7 +2423,7 @@ bool abi_compat_return_by_pointer(gbAllocator a, ProcCallingConvention cc, Type
 		}
 	}
 
-	if (build_context.ODIN_OS == "windows") {
+	if (build_context.ODIN_OS == "windows" || build_context.ODIN_OS == "linux" ) {
 		i64 size = 8*type_size_of(abi_return_type);
 		switch (size) {
 		case 0:

src/ir_print.cpp

@@ -339,8 +339,7 @@ void ir_print_proc_type_without_pointer(irFileBuffer *f, irModule *m, Type *t) {
 		// ir_fprintf(f, "* sret noalias ");
 		// ir_write_string(f, str_lit("* noalias "));
 		ir_write_string(f, str_lit("*"));
-		if (build_context.ODIN_OS == "darwin" ||
-		    build_context.ODIN_OS == "linux") {
+		if (build_context.ODIN_OS == "darwin") {
 			ir_fprintf(f, " byval");
 		}
 		if (param_count > 0 || t->Proc.calling_convention == ProcCC_Odin)  {
@@ -2456,8 +2455,7 @@ void ir_print_proc(irFileBuffer *f, irModule *m, irProcedure *proc) {
 	if (proc_type->return_by_pointer) {
 		ir_print_type(f, m, reduce_tuple_to_single_type(proc_type->results));
 		ir_write_str_lit(f, "* sret noalias ");
-		if (build_context.ODIN_OS == "darwin" ||
-		    build_context.ODIN_OS == "linux") {
+		if (build_context.ODIN_OS == "darwin") {
 			ir_fprintf(f, "byval ");
 		}
 		ir_write_str_lit(f, "%agg.result");

src/main.cpp

@@ -1477,7 +1477,7 @@ void show_timings(Checker *c, Timings *t) {
 void remove_temp_files(String output_base) {
 	if (build_context.keep_temp_files) return;
 
-	auto data = array_make<u8>(heap_allocator(), output_base.len + 10);
+	auto data = array_make<u8>(heap_allocator(), output_base.len + 30);
 	defer (array_free(&data));
 
 	isize n = output_base.len;
@@ -1488,6 +1488,7 @@ void remove_temp_files(String output_base) {
 	} while (0)
 	EXT_REMOVE(".ll");
 	EXT_REMOVE(".bc");
+	EXT_REMOVE("_memcpy_pass.bc");
 	if (build_context.build_mode != BuildMode_Object && !build_context.keep_object_files) {
 	#if defined(GB_SYSTEM_WINDOWS)
 		EXT_REMOVE(".obj");
@@ -1506,8 +1507,15 @@ void remove_temp_files(String output_base) {
 i32 exec_llvm_opt(String output_base) {
 #if defined(GB_SYSTEM_WINDOWS)
 	// For more passes arguments: http://llvm.org/docs/Passes.html
-	return system_exec_command_line_app("llvm-opt",
-		"\"%.*sbin/opt\" \"%.*s.ll\" -o \"%.*s.bc\" %.*s "
+
+  return system_exec_command_line_app("llvm-opt",
+		"\"%.*sbin/opt\" \"%.*s.ll\" -o \"%.*s_memcpy_pass.bc\" -memcpyopt"
+		"",
+		LIT(build_context.ODIN_ROOT),
+		LIT(output_base), LIT(output_base))
+
+  || system_exec_command_line_app("llvm-opt",
+		"\"%.*sbin/opt\" \"%.*s_memcpy_pass.bc\" -o \"%.*s.bc\" %.*s "
 		"",
 		LIT(build_context.ODIN_ROOT),
 		LIT(output_base), LIT(output_base),
@@ -1515,8 +1523,14 @@ i32 exec_llvm_opt(String output_base) {
 #else
 	// NOTE(zangent): This is separate because it seems that LLVM tools are packaged
 	//   with the Windows version, while they will be system-provided on MacOS and GNU/Linux
-	return system_exec_command_line_app("llvm-opt",
-		"opt \"%.*s.ll\" -o \"%.*s.bc\" %.*s "
+
+  return system_exec_command_line_app("llvm-opt",
+    "opt \"%.*s.ll\" -o \"%.*s_memcpy_pass.bc\" -memcpyopt"
+    "",
+    LIT(output_base), LIT(output_base))
+  
+	|| system_exec_command_line_app("llvm-opt",
+		"opt \"%.*s_memcpy_pass.bc\" -o \"%.*s.bc\" %.*s "
 		"",
 		LIT(output_base), LIT(output_base),
 		LIT(build_context.opt_flags));