ffi 1.10.0 → 1.11.1

data/ext/ffi_c/Thread.h

@@ -66,7 +66,9 @@ typedef struct rbffi_frame {
     struct thread_data* td;
 #endif
     struct rbffi_frame* prev;
+#ifndef HAVE_RUBY_THREAD_HAS_GVL_P
     bool has_gvl;
+#endif
     VALUE exc;
 } rbffi_frame_t;
 
@@ -74,19 +76,6 @@ rbffi_frame_t* rbffi_frame_current(void);
 void rbffi_frame_push(rbffi_frame_t* frame);
 void rbffi_frame_pop(rbffi_frame_t* frame);
 
-#ifdef HAVE_RB_THREAD_CALL_WITHOUT_GVL
-# define rbffi_thread_blocking_region rb_thread_call_without_gvl
-
-#elif defined(HAVE_RB_THREAD_BLOCKING_REGION)
-# define rbffi_thread_blocking_region rb_thread_blocking_region
-
-#else
-
-VALUE rbffi_thread_blocking_region(VALUE (*func)(void *), void *data1, void (*ubf)(void *), void *data2);
-
-#endif
-
-
 #ifdef	__cplusplus
 }
 #endif

data/ext/ffi_c/Type.c

@@ -235,24 +235,6 @@ rbffi_Type_Lookup(VALUE name)
     return Qnil;
 }
 
-/**
- * rbffi_Type_Find() is like rbffi_Type_Lookup, but an error is raised if the
- * type is not found.
- */
-VALUE
-rbffi_Type_Find(VALUE name)
-{
-    VALUE rbType = rbffi_Type_Lookup(name);
-
-    if (!RTEST(rbType)) {
-        VALUE s = rb_inspect(name);
-        rb_raise(rb_eTypeError, "invalid type, %s", RSTRING_PTR(s));
-        RB_GC_GUARD(s);
-    }
-
-    return rbType;
-}
-
 void
 rbffi_Type_Init(VALUE moduleFFI)
 {

data/ext/ffi_c/Type.h

@@ -52,7 +52,6 @@ struct Type_ {
 
 extern VALUE rbffi_TypeClass;
 extern VALUE rbffi_Type_Lookup(VALUE type);
-extern VALUE rbffi_Type_Find(VALUE type);
 
 #ifdef	__cplusplus
 }

data/ext/ffi_c/Variadic.c

@@ -191,7 +191,7 @@ variadic_invoke(VALUE self, VALUE parameterTypes, VALUE parameterValues)
 
     for (i = 0; i < paramCount; ++i) {
         VALUE rbType = rb_ary_entry(parameterTypes, i);
-        
+
         if (!rb_obj_is_kind_of(rbType, rbffi_TypeClass)) {
             rb_raise(rb_eTypeError, "wrong type.  Expected (FFI::Type)");
         }
@@ -210,7 +210,7 @@ variadic_invoke(VALUE self, VALUE parameterTypes, VALUE parameterValues)
                 rbType = rb_const_get(rbffi_TypeClass, rb_intern("UINT32"));
                 Data_Get_Struct(rbType, Type, paramTypes[i]);
                 break;
-            
+
             case NATIVE_FLOAT32:
                 rbType = rb_const_get(rbffi_TypeClass, rb_intern("DOUBLE"));
                 Data_Get_Struct(rbType, Type, paramTypes[i]);
@@ -219,8 +219,8 @@ variadic_invoke(VALUE self, VALUE parameterTypes, VALUE parameterValues)
             default:
                 break;
         }
-        
-        
+
+
         ffiParamTypes[i] = paramTypes[i]->ffiType;
         if (ffiParamTypes[i] == NULL) {
             rb_raise(rb_eArgError, "Invalid parameter type #%x", paramTypes[i]->nativeType);
@@ -254,13 +254,9 @@ variadic_invoke(VALUE self, VALUE parameterTypes, VALUE parameterValues)
 
     rbffi_SetupCallParams(paramCount, argv, -1, paramTypes, params,
         ffiValues, NULL, 0, invoker->rbEnums);
-    
+
     rbffi_frame_push(&frame);
-#ifdef HAVE_RB_THREAD_CALL_WITHOUT_GVL
-    /* In Call.c, blocking: true is supported on older ruby variants
-     * without rb_thread_call_without_gvl by allocating on the heap instead
-     * of the stack. Since this functionality is being added later,
-     * we’re skipping support for old rubies here. */
+
     if(unlikely(invoker->blocking)) {
         rbffi_blocking_call_t* bc;
         bc = ALLOCA_N(rbffi_blocking_call_t, 1);
@@ -275,13 +271,11 @@ variadic_invoke(VALUE self, VALUE parameterTypes, VALUE parameterValues)
     } else {
         ffi_call(&cif, FFI_FN(invoker->function), retval, ffiValues);
     }
-#else
-    ffi_call(&cif, FFI_FN(invoker->function), retval, ffiValues);
-#endif
+
     rbffi_frame_pop(&frame);
-    
+
     rbffi_save_errno();
-    
+
     if (RTEST(frame.exc) && frame.exc != Qnil) {
         rb_exc_raise(frame.exc);
     }

data/ext/ffi_c/extconf.rb

@@ -3,6 +3,21 @@
 if !defined?(RUBY_ENGINE) || RUBY_ENGINE == 'ruby' || RUBY_ENGINE == 'rbx'
   require 'mkmf'
   require 'rbconfig'
+
+  def system_libffi_usable?
+    # We need pkg_config or ffi.h
+    libffi_ok = pkg_config("libffi") ||
+        have_header("ffi.h") ||
+        find_header("ffi.h", "/usr/local/include", "/usr/include/ffi")
+
+    # Ensure we can link to ffi_call
+    libffi_ok &&= have_library("ffi", "ffi_call", [ "ffi.h" ]) ||
+                  have_library("libffi", "ffi_call", [ "ffi.h" ])
+
+    # And we need a libffi version recent enough to provide ffi_closure_alloc
+    libffi_ok &&= have_func("ffi_closure_alloc")
+  end
+
   dir_config("ffi_c")
 
   # recent versions of ruby add restrictive ansi and warning flags on a whim - kill them all
@@ -12,42 +27,41 @@ if !defined?(RUBY_ENGINE) || RUBY_ENGINE == 'ruby' || RUBY_ENGINE == 'rbx'
   # solaris 10 needs -c99 for <stdbool.h>
   $CFLAGS << " -std=c99" if RbConfig::CONFIG['host_os'] =~ /solaris(!?2\.11)/
 
-  if ENV['RUBY_CC_VERSION'].nil? && (pkg_config("libffi") ||
-     have_header("ffi.h") ||
-     find_header("ffi.h", "/usr/local/include", "/usr/include/ffi"))
+  # Check whether we use system libffi
+  system_libffi = enable_config('system-libffi', :try)
 
-    # We need at least ffi_call and ffi_closure_alloc
-    libffi_ok = have_library("ffi", "ffi_call", [ "ffi.h" ]) ||
-                have_library("libffi", "ffi_call", [ "ffi.h" ])
-    libffi_ok &&= have_func("ffi_closure_alloc")
-
-    # Check if the raw api is available.
-    $defs << "-DHAVE_RAW_API" if have_func("ffi_raw_call") && have_func("ffi_prep_raw_closure")
+  if system_libffi == :try
+    system_libffi = ENV['RUBY_CC_VERSION'].nil? && system_libffi_usable?
+  elsif system_libffi
+    abort "system libffi is not usable" unless system_libffi_usable?
   end
 
   have_header('shlwapi.h')
-  have_header('ruby/thread.h') # for compat with ruby < 2.0
-  have_func('rb_thread_blocking_region')
-  have_func('rb_thread_call_with_gvl')
-  have_func('rb_thread_call_without_gvl')
+  have_func('rb_thread_call_without_gvl') || abort("Ruby C-API function `rb_thread_call_without_gvl` is missing")
+  have_func('ruby_native_thread_p')
+  if RUBY_VERSION >= "2.3.0"
+    # On OSX and Linux ruby_thread_has_gvl_p() is detected but fails at runtime for ruby < 2.3.0
+    have_func('ruby_thread_has_gvl_p')
+  end
 
-  if libffi_ok
+  if system_libffi
     have_func('ffi_prep_cif_var')
+    $defs << "-DHAVE_RAW_API" if have_func("ffi_raw_call") && have_func("ffi_prep_raw_closure")
   else
     $defs << "-DHAVE_FFI_PREP_CIF_VAR"
+    $defs << "-DUSE_INTERNAL_LIBFFI"
   end
 
   $defs << "-DHAVE_EXTCONF_H" if $defs.empty? # needed so create_header works
-  $defs << "-DUSE_INTERNAL_LIBFFI" unless libffi_ok
   $defs << "-DRUBY_1_9" if RUBY_VERSION >= "1.9.0"
   $defs << "-DFFI_BUILDING" if RbConfig::CONFIG['host_os'] =~ /mswin/ # for compatibility with newer libffi
 
   create_header
-  
-  $LOCAL_LIBS << " ./libffi/.libs/libffi_convenience.lib" if !libffi_ok && RbConfig::CONFIG['host_os'] =~ /mswin/
+
+  $LOCAL_LIBS << " ./libffi/.libs/libffi_convenience.lib" if !system_libffi && RbConfig::CONFIG['host_os'] =~ /mswin/
 
   create_makefile("ffi_c")
-  unless libffi_ok
+  unless system_libffi
     File.open("Makefile", "a") do |mf|
       mf.puts "LIBFFI_HOST=--host=#{RbConfig::CONFIG['host_alias']}" if RbConfig::CONFIG.has_key?("host_alias")
       if RbConfig::CONFIG['host_os'].downcase =~ /darwin/
@@ -63,7 +77,7 @@ if !defined?(RUBY_ENGINE) || RUBY_ENGINE == 'ruby' || RUBY_ENGINE == 'rbx'
       end
     end
   end
-  
+
 else
   File.open("Makefile", "w") do |mf|
     mf.puts "# Dummy makefile for non-mri rubies"

data/ext/ffi_c/ffi.c

@@ -40,7 +40,6 @@
 #include "MemoryPointer.h"
 #include "Struct.h"
 #include "StructByValue.h"
-#include "StructByReference.h"
 #include "DynamicLibrary.h"
 #include "Platform.h"
 #include "Types.h"
@@ -59,9 +58,9 @@ VALUE rbffi_FFIModule = Qnil;
 static VALUE moduleFFI = Qnil;
 
 void
-Init_ffi_c(void) 
+Init_ffi_c(void)
 {
-    /* 
+    /*
      * Document-module: FFI
      *
      * This module embbed type constants from {FFI::NativeType}.
@@ -70,12 +69,10 @@ Init_ffi_c(void)
     rb_global_variable(&rbffi_FFIModule);
 
     rbffi_Thread_Init(rbffi_FFIModule);
-    
+
     /* FFI::Type needs to be initialized before most other classes */
     rbffi_Type_Init(moduleFFI);
 
-    rbffi_DataConverter_Init(moduleFFI);
-
     rbffi_ArrayType_Init(moduleFFI);
     rbffi_LastError_Init(moduleFFI);
     rbffi_Call_Init(moduleFFI);
@@ -88,11 +85,9 @@ Init_ffi_c(void)
     rbffi_MemoryPointer_Init(moduleFFI);
     rbffi_Buffer_Init(moduleFFI);
     rbffi_StructByValue_Init(moduleFFI);
-    rbffi_StructByReference_Init(moduleFFI);
     rbffi_Struct_Init(moduleFFI);
     rbffi_DynamicLibrary_Init(moduleFFI);
     rbffi_Variadic_Init(moduleFFI);
     rbffi_Types_Init(moduleFFI);
     rbffi_MappedType_Init(moduleFFI);
 }
-

data/ext/ffi_c/libffi/configure.ac

@@ -270,7 +270,7 @@ if test "x$GCC" = "xyes"; then
   	echo 'int __attribute__ ((visibility ("hidden"))) foo (void) { return 1  ; }' > conftest.c
   	libffi_cv_hidden_visibility_attribute=no
   	if AC_TRY_COMMAND(${CC-cc} -Werror -S conftest.c -o conftest.s 1>&AS_MESSAGE_LOG_FD); then
-  	    if grep '\.hidden.*foo' conftest.s >/dev/null; then
+  	    if egrep '(\.hidden|\.private_extern).*foo' conftest.s >/dev/null; then
   		libffi_cv_hidden_visibility_attribute=yes
   	    fi
   	fi
@@ -292,7 +292,11 @@ AM_CONDITIONAL(BUILD_DOCS, [test x$enable_docs = xyes])
 AH_BOTTOM([
 #ifdef HAVE_HIDDEN_VISIBILITY_ATTRIBUTE
 #ifdef LIBFFI_ASM
+#ifdef __APPLE__
+#define FFI_HIDDEN(name) .private_extern name
+#else
 #define FFI_HIDDEN(name) .hidden name
+#endif
 #else
 #define FFI_HIDDEN __attribute__ ((visibility ("hidden")))
 #endif

data/ext/ffi_c/libffi/include/ffi_common.h

@@ -77,7 +77,7 @@ void ffi_type_test(ffi_type *a, char *file, int line);
 /* v cast to size_t and aligned up to a multiple of a */
 #define FFI_ALIGN(v, a)  (((((size_t) (v))-1) | ((a)-1))+1)
 /* v cast to size_t and aligned down to a multiple of a */
-#define ALIGN_DOWN(v, a) (((size_t) (v)) & -a)
+#define FFI_ALIGN_DOWN(v, a) (((size_t) (v)) & -a)
 
 /* Perform machine dependent cif processing */
 ffi_status ffi_prep_cif_machdep(ffi_cif *cif);

data/ext/ffi_c/libffi/src/aarch64/ffi.c

@@ -350,10 +350,10 @@ extend_hfa_type (void *dest, void *src, int h)
 "	b	1f\n"
 "	nop\n"
 "	ldp	q16, q17, [%3]\n"	/* Q4 */
-"	ldp	q18, q19, [%3, #16]\n"
+"	ldp	q18, q19, [%3, #32]\n"
 "	b	4f\n"
 "	ldp	q16, q17, [%3]\n"	/* Q3 */
-"	ldr	q18, [%3, #16]\n"
+"	ldr	q18, [%3, #32]\n"
 "	b	3f\n"
 "	ldp	q16, q17, [%3]\n"	/* Q2 */
 "	b	2f\n"

data/ext/ffi_c/libffi/src/frv/ffi.c

@@ -107,7 +107,7 @@ void *ffi_prep_args(char *stack, extended_cif *ecif)
       count += z;
     }
 
-  return (stack + ((count > 24) ? 24 : ALIGN_DOWN(count, 8)));
+  return (stack + ((count > 24) ? 24 : FFI_ALIGN_DOWN(count, 8)));
 }
 
 /* Perform machine dependent cif processing */

data/ext/ffi_c/libffi/src/metag/ffi.c

@@ -61,7 +61,7 @@ unsigned int ffi_prep_args(char *stack, extended_cif *ecif)
 		argp -= z;
 
 		/* Align if necessary */
-		argp = (char *) ALIGN_DOWN(ALIGN_DOWN(argp, (*p_arg)->alignment), 4);
+		argp = (char *) FFI_ALIGN_DOWN(FFI_ALIGN_DOWN(argp, (*p_arg)->alignment), 4);
 
 		if (z < sizeof(int)) {
 			z = sizeof(int);

data/ext/ffi_c/libffi/src/moxie/ffi.c

@@ -100,7 +100,7 @@ void *ffi_prep_args(char *stack, extended_cif *ecif)
       count += z;
     }
 
-  return (stack + ((count > 24) ? 24 : ALIGN_DOWN(count, 8)));
+  return (stack + ((count > 24) ? 24 : FFI_ALIGN_DOWN(count, 8)));
 }
 
 /* Perform machine dependent cif processing */

data/ext/ffi_c/libffi/src/riscv/ffi.c

@@ -324,9 +324,12 @@ ffi_status ffi_prep_cif_machdep_var(ffi_cif *cif, unsigned int nfixedargs, unsig
 }
 
 /* Low level routine for calling functions */
-extern void ffi_call_asm(void *stack, struct call_context *regs, void (*fn)(void)) FFI_HIDDEN;
+extern void ffi_call_asm (void *stack, struct call_context *regs,
+			  void (*fn) (void), void *closure) FFI_HIDDEN;
 
-void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+static void
+ffi_call_int (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue,
+	      void *closure)
 {
     /* this is a conservative estimate, assuming a complex return value and
        that all remaining arguments are long long / __int128 */
@@ -366,13 +369,26 @@ void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
     for (i = 0; i < cif->nargs; i++)
         marshal(&cb, cif->arg_types[i], i >= cif->riscv_nfixedargs, avalue[i]);
 
-    ffi_call_asm((void*)alloc_base, cb.aregs, fn);
+    ffi_call_asm ((void *) alloc_base, cb.aregs, fn, closure);
 
     cb.used_float = cb.used_integer = 0;
     if (!return_by_ref && rvalue)
         unmarshal(&cb, cif->rtype, 0, rvalue);
 }
 
+void
+ffi_call (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, NULL);
+}
+
+void
+ffi_call_go (ffi_cif *cif, void (*fn) (void), void *rvalue,
+	     void **avalue, void *closure)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, closure);
+}
+
 extern void ffi_closure_asm(void) FFI_HIDDEN;
 
 ffi_status ffi_prep_closure_loc(ffi_closure *closure, ffi_cif *cif, void (*fun)(ffi_cif*,void*,void**,void*), void *user_data, void *codeloc)
@@ -406,11 +422,31 @@ ffi_status ffi_prep_closure_loc(ffi_closure *closure, ffi_cif *cif, void (*fun)(
     return FFI_OK;
 }
 
+extern void ffi_go_closure_asm (void) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_go_closure (ffi_go_closure *closure, ffi_cif *cif,
+		     void (*fun) (ffi_cif *, void *, void **, void *))
+{
+  if (cif->abi <= FFI_FIRST_ABI || cif->abi >= FFI_LAST_ABI)
+    return FFI_BAD_ABI;
+
+  closure->tramp = (void *) ffi_go_closure_asm;
+  closure->cif = cif;
+  closure->fun = fun;
+
+  return FFI_OK;
+}
+
 /* Called by the assembly code with aregs pointing to saved argument registers
    and stack pointing to the stacked arguments.  Return values passed in
    registers will be reloaded from aregs. */
-void FFI_HIDDEN ffi_closure_inner(size_t *stack, call_context *aregs, ffi_closure *closure) {
-    ffi_cif *cif = closure->cif;
+void FFI_HIDDEN
+ffi_closure_inner (ffi_cif *cif,
+		   void (*fun) (ffi_cif *, void *, void **, void *),
+		   void *user_data,
+		   size_t *stack, call_context *aregs)
+{
     void **avalue = alloca(cif->nargs * sizeof(void*));
     /* storage for arguments which will be copied by unmarshal().  We could
        theoretically avoid the copies in many cases and use at most 128 bytes
@@ -436,7 +472,7 @@ void FFI_HIDDEN ffi_closure_inner(size_t *stack, call_context *aregs, ffi_closur
         avalue[i] = unmarshal(&cb, cif->arg_types[i],
             i >= cif->riscv_nfixedargs, astorage + i*MAXCOPYARG);
 
-    (closure->fun)(cif, rvalue, avalue, closure->user_data);
+    fun (cif, rvalue, avalue, user_data);
 
     if (!return_by_ref && cif->rtype->type != FFI_TYPE_VOID) {
         cb.used_integer = cb.used_float = 0;

data/ext/ffi_c/libffi/src/riscv/ffitarget.h

@@ -59,6 +59,7 @@ typedef enum ffi_abi {
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
+#define FFI_GO_CLOSURES 1
 #define FFI_TRAMPOLINE_SIZE 24
 #define FFI_NATIVE_RAW_API 0
 #define FFI_EXTRA_CIF_FIELDS unsigned riscv_nfixedargs; unsigned riscv_unused;

data/ext/ffi_c/libffi/src/riscv/sysv.S

@@ -67,8 +67,8 @@
       intreg pad[rv32 ? 2 : 0];
       intreg save_fp, save_ra;
   }
-  void ffi_call_asm(size_t *stackargs, struct call_context *regargs,
-      void (*fn)(void));
+  void ffi_call_asm (size_t *stackargs, struct call_context *regargs,
+                     void (*fn) (void), void *closure);
 */
 
 #define FRAME_LEN (8 * FLTS + 8 * PTRS + 16)
@@ -98,6 +98,7 @@ ffi_call_asm:
 
     # Load arguments
     mv      t1, a2
+    mv      t2, a3
 
 #if FLTS
     FLARG   fa0, -FRAME_LEN+0*FLTS(fp)
@@ -145,8 +146,10 @@ ffi_call_asm:
 
 /*
   ffi_closure_asm. Expects address of the passed-in ffi_closure in t1.
-  void ffi_closure_inner(size_t *stackargs, struct call_context *regargs,
-      ffi_closure *closure);
+  void ffi_closure_inner (ffi_cif *cif,
+		          void (*fun) (ffi_cif *, void *, void **, void *),
+		          void *user_data,
+		          size_t *stackargs, struct call_context *regargs)
 */
 
     .globl ffi_closure_asm
@@ -187,9 +190,11 @@ ffi_closure_asm:
     SARG    a7, 8*FLTS+7*PTRS(sp)
 
     /* enter C */
-    addi    a0, sp, FRAME_LEN
-    mv      a1, sp
-    mv      a2, t1
+    LARG    a0, FFI_TRAMPOLINE_SIZE+0*PTRS(t1)
+    LARG    a1, FFI_TRAMPOLINE_SIZE+1*PTRS(t1)
+    LARG    a2, FFI_TRAMPOLINE_SIZE+2*PTRS(t1)
+    addi    a3, sp, FRAME_LEN
+    mv      a4, sp
 
     call    ffi_closure_inner
 
@@ -212,3 +217,77 @@ ffi_closure_asm:
     ret
     .cfi_endproc
     .size ffi_closure_asm, .-ffi_closure_asm
+
+/*
+  ffi_go_closure_asm.  Expects address of the passed-in ffi_go_closure in t2.
+  void ffi_closure_inner (ffi_cif *cif,
+		          void (*fun) (ffi_cif *, void *, void **, void *),
+		          void *user_data,
+		          size_t *stackargs, struct call_context *regargs)
+*/
+
+    .globl ffi_go_closure_asm
+    .hidden ffi_go_closure_asm
+    .type ffi_go_closure_asm, @function
+ffi_go_closure_asm:
+    .cfi_startproc
+
+    addi    sp,  sp, -FRAME_LEN
+    .cfi_def_cfa_offset FRAME_LEN
+
+    /* make a frame */
+    SARG    fp, FRAME_LEN - 2*PTRS(sp)
+    .cfi_offset 8, -2*PTRS
+    SARG    ra, FRAME_LEN - 1*PTRS(sp)
+    .cfi_offset 1, -1*PTRS
+    addi    fp, sp, FRAME_LEN
+
+    /* save arguments */
+#if FLTS
+    FSARG   fa0, 0*FLTS(sp)
+    FSARG   fa1, 1*FLTS(sp)
+    FSARG   fa2, 2*FLTS(sp)
+    FSARG   fa3, 3*FLTS(sp)
+    FSARG   fa4, 4*FLTS(sp)
+    FSARG   fa5, 5*FLTS(sp)
+    FSARG   fa6, 6*FLTS(sp)
+    FSARG   fa7, 7*FLTS(sp)
+#endif
+
+    SARG    a0, 8*FLTS+0*PTRS(sp)
+    SARG    a1, 8*FLTS+1*PTRS(sp)
+    SARG    a2, 8*FLTS+2*PTRS(sp)
+    SARG    a3, 8*FLTS+3*PTRS(sp)
+    SARG    a4, 8*FLTS+4*PTRS(sp)
+    SARG    a5, 8*FLTS+5*PTRS(sp)
+    SARG    a6, 8*FLTS+6*PTRS(sp)
+    SARG    a7, 8*FLTS+7*PTRS(sp)
+
+    /* enter C */
+    LARG    a0, 1*PTRS(t2)
+    LARG    a1, 2*PTRS(t2)
+    mv      a2, t2
+    addi    a3, sp, FRAME_LEN
+    mv      a4, sp
+
+    call    ffi_closure_inner
+
+    /* return values */
+#if FLTS
+    FLARG   fa0, 0*FLTS(sp)
+    FLARG   fa1, 1*FLTS(sp)
+#endif
+
+    LARG    a0, 8*FLTS+0*PTRS(sp)
+    LARG    a1, 8*FLTS+1*PTRS(sp)
+
+    /* restore and return */
+    LARG    ra, FRAME_LEN-1*PTRS(sp)
+    .cfi_restore 1
+    LARG    fp, FRAME_LEN-2*PTRS(sp)
+    .cfi_restore 8
+    addi    sp, sp, FRAME_LEN
+    .cfi_def_cfa_offset 0
+    ret
+    .cfi_endproc
+    .size ffi_go_closure_asm, .-ffi_go_closure_asm