Update ARC's libffi port (#771)

* Add support for ARC and ARC64

Add support for ARC/ARC32/ARC64

* Implementation of GO Closure for ARC/ARC32/ARC64 Architectures

---------

Co-authored-by: Nuno Cardoso <cardoso@synopsys.com>
Co-authored-by: Luis Silva <luiss@synopsys.com>

3 files changed, 508 insertions, 242 deletions

diff --git a/src/arc/arcompact.S b/src/arc/arcompact.S
index 03715fde..1d7f1a1f 100644
--- a/src/arc/arcompact.S
+++ b/src/arc/arcompact.S
@@ -33,103 +33,178 @@
 #define ENTRY(x) .globl CNAME(x)` .type CNAME(x),%function` CNAME(x):
 #endif
 
+#if __SIZEOF_POINTER__ == 8
+#define PTRS 8
+#define FLTS 8
+#define LARG ldl
+#define SARG stl
+#define ADDPTR addl
+#define MOVPTR movl_s
+#else
+#define PTRS 4
+#define FLTS 4
+#define LARG ld
+#define SARG st
+#define ADDPTR add
+#define MOVPTR mov_s
+#endif
+
+#define FRAME_LEN (8 * PTRS + 16)
+
 .text
 
-        /* R0:   ffi_prep_args */
-        /* R1:   &ecif */
-        /* R2:   cif->bytes */
-        /* R3:   fig->flags */
-        /* R4:   ecif.rvalue */
-        /* R5:   fn */
-ENTRY(ffi_call_ARCompact)
+ENTRY(ffi_call_asm)
+        .cfi_startproc
+
         /* Save registers.  */
-        st.a       fp, [sp, -4]        /* fp + 20, fp */
-        push_s     blink               /* fp + 16, blink */
-        st.a       r4, [sp, -4]        /* fp + 12, ecif.rvalue */
-        push_s     r3                  /* fp +  8, fig->flags */
-        st.a       r5, [sp, -4]        /* fp +  4, fn */
-        push_s     r2                  /* fp +  0, cif->bytes */
-        mov        fp, sp
-
-        /* Make room for all of the new args.  */
-        sub        sp, sp, r2
-
-        /* Place all of the ffi_prep_args in position.  */
-        /* ffi_prep_args(char *stack, extended_cif *ecif) */
-        /* R1 already set.  */
-
-        /* And call.  */
-        jl_s.d     [r0]
-        mov_s      r0, sp
-
-        ld.ab      r12, [fp, 4]        /* cif->bytes */
-        ld.ab      r11, [fp, 4]        /* fn */
-
-        /* Move first 8 parameters in registers...  */
-        ld_s       r0, [sp]
-        ld_s       r1, [sp, 4]
-        ld_s       r2, [sp, 8]
-        ld_s       r3, [sp, 12]
-        ld         r4, [sp, 16]
-        ld         r5, [sp, 20]
-        ld         r6, [sp, 24]
-        ld         r7, [sp, 28]
-
-        /* ...and adjust the stack.  */
-        min        r12, r12, 32
+        .cfi_def_cfa r1, FRAME_LEN
+        SARG    fp, [r1, FRAME_LEN - 2*PTRS]
+        .cfi_offset fp, -2*PTRS
+        SARG    blink, [r1, FRAME_LEN - 1*PTRS]
+        .cfi_offset blink, -1*PTRS
+        ADDPTR  fp, r1, FRAME_LEN
+        MOVPTR  sp, r0
+        .cfi_def_cfa fp, 0
+
+        /* Load arguments.  */
+        MOVPTR  r11, r2      /* fn */
+        MOVPTR  r12, r3      /* closure */
+
+        /* Save arguments.  */
+        LARG    r0, [fp, -FRAME_LEN+0*PTRS]
+        LARG    r1, [fp, -FRAME_LEN+1*PTRS]
+        LARG    r2, [fp, -FRAME_LEN+2*PTRS]
+        LARG    r3, [fp, -FRAME_LEN+3*PTRS]
+        LARG    r4, [fp, -FRAME_LEN+4*PTRS]
+        LARG    r5, [fp, -FRAME_LEN+5*PTRS]
+        LARG    r6, [fp, -FRAME_LEN+6*PTRS]
+        LARG    r7, [fp, -FRAME_LEN+7*PTRS]
 
         /* Call the function.  */
-        jl.d       [r11]
-        add        sp, sp, r12 
-
-        mov        sp, fp        
-        pop_s      r3        /* fig->flags, return type */
-        pop_s      r2        /* ecif.rvalue, pointer for return value */
-
-        /* If the return value pointer is NULL, assume no return value.  */
-        breq.d     r2, 0, epilogue
-        pop_s      blink
-
-        /* Return INT.  */
-        brne       r3, FFI_TYPE_INT, return_double
-        b.d        epilogue
-        st_s       r0, [r2]        
-
-return_double:
-        brne       r3, FFI_TYPE_DOUBLE, epilogue
-        st_s       r0, [r2]        
-        st_s       r1, [r2,4]
-
-epilogue:
-        j_s.d      [blink]
-        ld.ab      fp, [sp, 4]
-
-ENTRY(ffi_closure_ARCompact)
-        st.a       r0, [sp, -32]
-        st_s       r1, [sp, 4]
-        st_s       r2, [sp, 8]
-        st_s       r3, [sp, 12]
-        st         r4, [sp, 16]
-        st         r5, [sp, 20]
-        st         r6, [sp, 24]
-        st         r7, [sp, 28]
-
-        /* pointer to arguments */
-        mov_s      r2, sp
-
-        /* return value goes here */
-        sub        sp, sp, 8
-        mov_s      r1, sp
-
-        push_s     blink
+        jl         [r11] 
+
+        /* Save return value (r0/r1)  */
+        SARG    r0, [fp, -FRAME_LEN+0*PTRS]
+        SARG    r1, [fp, -FRAME_LEN+1*PTRS]
+
+        /* Restore and return.  */
+        add     sp, fp, -FRAME_LEN
+        .cfi_def_cfa sp, FRAME_LEN
+        LARG    blink, [fp, -1*PTRS]
+        .cfi_restore blink
+        LARG    fp, [fp, -2*PTRS]
+        .cfi_restore fp
+        j_s     [blink]
+        .cfi_endproc
+        .size   ffi_call_asm, .-ffi_call_asm
+
+/*
+  ffi_closure_asm. Expects address of the passed-in ffi_closure in r8.
+  void ffi_closure_inner (ffi_cif *cif,
+		          void (*fun) (ffi_cif *, void *, void **, void *),
+		          void *user_data,
+		          size_t *stackargs, struct call_context *regargs)
+*/
+
+ENTRY(ffi_closure_asm)
+        .cfi_startproc
+
+        ADDPTR  sp, sp, -FRAME_LEN
+        .cfi_def_cfa_offset FRAME_LEN
+
+        /* Make a frame.  */
+        SARG    fp, [sp, FRAME_LEN-2*PTRS]
+        .cfi_offset fp, -2*PTRS
+        SARG    blink, [sp, FRAME_LEN-1*PTRS]
+        .cfi_offset blink, -1*PTRS
+        ADDPTR  fp, sp, FRAME_LEN
+
+        /* Save arguments.  */
+        SARG    r0, [sp, 0*PTRS]
+        SARG    r1, [sp, 1*PTRS]
+        SARG    r2, [sp, 2*PTRS]
+        SARG    r3, [sp, 3*PTRS]
+        SARG    r4, [sp, 4*PTRS]
+        SARG    r5, [sp, 5*PTRS]
+        SARG    r6, [sp, 6*PTRS]
+        SARG    r7, [sp, 7*PTRS]
+
+        /* Enter C.  */
+        LARG    r0, [r8, FFI_TRAMPOLINE_SIZE+0*PTRS]
+        LARG    r1, [r8, FFI_TRAMPOLINE_SIZE+1*PTRS]
+        LARG    r2, [r8, FFI_TRAMPOLINE_SIZE+2*PTRS]
+        ADDPTR  r3, sp, FRAME_LEN
+        MOVPTR  r4, sp
+
+        /* Call the C code.  */
+        bl      ffi_closure_inner
         
-        bl.d       ffi_closure_inner_ARCompact
-        mov_s      r0, r8                /* codeloc, set by trampoline */
-
-        pop_s      blink
-
-        /* set return value to r1:r0 */
-        pop_s      r0
-        pop_s      r1
-        j_s.d      [blink]
-        add_s      sp, sp, 32
+        /* Return values.  */
+        LARG    r0, [sp, 0*PTRS]
+        LARG    r1, [sp, 1*PTRS]
+
+        /* Restore and return.  */
+        LARG    blink, [sp, FRAME_LEN-1*PTRS]
+        .cfi_restore blink
+        LARG    fp, [sp, FRAME_LEN-2*PTRS]
+        .cfi_restore fp
+        ADDPTR  sp, sp, FRAME_LEN
+        .cfi_def_cfa_offset 0
+        j_s     [blink]
+        .cfi_endproc
+        .size ffi_closure_asm, .-ffi_closure_asm
+
+/*
+  ffi_go_closure_asm.  Expects address of the passed-in ffi_go_closure in r12.
+  void ffi_closure_inner (ffi_cif *cif,
+		          void (*fun) (ffi_cif *, void *, void **, void *),
+		          void *user_data,
+		          size_t *stackargs, struct call_context *regargs)
+*/
+
+ENTRY(ffi_go_closure_asm)
+        .cfi_startproc
+
+        ADDPTR  sp, sp, -FRAME_LEN
+        .cfi_def_cfa_offset FRAME_LEN
+
+        /* make a frame */
+        SARG    fp, [sp, FRAME_LEN-2*PTRS]
+        .cfi_offset fp, -2*PTRS
+        SARG    blink, [sp, FRAME_LEN-1*PTRS]
+        .cfi_offset blink, -1*PTRS
+        ADDPTR  fp, sp, FRAME_LEN
+
+        /* save arguments */
+        SARG    r0, [sp, 0*PTRS]
+        SARG    r1, [sp, 1*PTRS]
+        SARG    r2, [sp, 2*PTRS]
+        SARG    r3, [sp, 3*PTRS]
+        SARG    r4, [sp, 4*PTRS]
+        SARG    r5, [sp, 5*PTRS]
+        SARG    r6, [sp, 6*PTRS]
+        SARG    r7, [sp, 7*PTRS]
+
+        /* enter C */
+        LARG    r0, [r12, 1*PTRS]
+        LARG    r1, [r12, 2*PTRS]
+        MOVPTR  r2, r12
+        ADDPTR  r3, sp, FRAME_LEN
+        MOVPTR  r4, sp
+
+        bl      ffi_closure_inner
+
+        /* Return values. */
+        LARG    r0, [sp, 0*PTRS]
+        LARG    r1, [sp, 1*PTRS]
+
+
+        LARG    blink, [sp, FRAME_LEN-1*PTRS]
+        .cfi_restore blink
+        LARG    fp, [sp, FRAME_LEN-2*PTRS]
+        .cfi_restore fp
+        ADDPTR  sp, sp, FRAME_LEN
+        .cfi_def_cfa_offset 0
+        j_s     [blink]
+        .cfi_endproc
+        .size ffi_go_closure_asm, .-ffi_go_closure_asm
diff --git a/src/arc/ffi.c b/src/arc/ffi.c
index 4d10b21a..d7292748 100644
--- a/src/arc/ffi.c
+++ b/src/arc/ffi.c
@@ -31,6 +31,34 @@
 
 #include <sys/cachectl.h>
 
+#define NARGREG 8
+#define STKALIGN 4
+#define MAXCOPYARG (2 * sizeof(double))
+
+typedef struct call_context
+{
+  size_t r[8];
+  /* used by the assembly code to in-place construct its own stack frame */
+  char frame[16];
+} call_context;
+
+typedef struct call_builder
+{
+  call_context *aregs;
+  int used_integer;
+  //int used_float;
+  size_t *used_stack;
+  void *struct_stack;
+} call_builder;
+
+/* integer (not pointer) less than ABI XLEN */
+/* FFI_TYPE_INT does not appear to be used */
+#if defined(__ARC64_ARCH64__)
+#define IS_INT(type) ((type) >= FFI_TYPE_UINT8 && (type) <= FFI_TYPE_SINT64)
+#else
+#define IS_INT(type) ((type) >= FFI_TYPE_UINT8 && (type) <= FFI_TYPE_SINT32)
+#endif
+
 /* for little endian ARC, the code is in fact stored as mixed endian for
    performance reasons */
 #if __BIG_ENDIAN__
@@ -39,94 +67,6 @@
 #define CODE_ENDIAN(x) ( (((uint32_t) (x)) << 16) | (((uint32_t) (x)) >> 16))
 #endif
 
-/* ffi_prep_args is called by the assembly routine once stack
-   space has been allocated for the function's arguments.  */
-
-void
-ffi_prep_args (char *stack, extended_cif * ecif)
-{
-  unsigned int i;
-  void **p_argv;
-  char *argp;
-  ffi_type **p_arg;
-
-  argp = stack;
-
-  if (ecif->cif->rtype->type == FFI_TYPE_STRUCT)
-    {
-      *(void **) argp = ecif->rvalue;
-      argp += 4;
-    }
-
-  p_argv = ecif->avalue;
-
-  for (i = ecif->cif->nargs, p_arg = ecif->cif->arg_types;
-       (i != 0); i--, p_arg++)
-    {
-      size_t z;
-      int alignment;
-
-      /* align alignment to 4 */
-      alignment = (((*p_arg)->alignment - 1) | 3) + 1;
-
-      /* Align if necessary.  */
-      if ((alignment - 1) & (unsigned) argp)
-	argp = (char *) FFI_ALIGN (argp, alignment);
-
-      z = (*p_arg)->size;
-      if (z < sizeof (int))
-	{
-	  z = sizeof (int);
-
-	  switch ((*p_arg)->type)
-	    {
-	    case FFI_TYPE_SINT8:
-	      *(signed int *) argp = (signed int) *(SINT8 *) (*p_argv);
-	      break;
-
-	    case FFI_TYPE_UINT8:
-	      *(unsigned int *) argp = (unsigned int) *(UINT8 *) (*p_argv);
-	      break;
-
-	    case FFI_TYPE_SINT16:
-	      *(signed int *) argp = (signed int) *(SINT16 *) (*p_argv);
-	      break;
-
-	    case FFI_TYPE_UINT16:
-	      *(unsigned int *) argp = (unsigned int) *(UINT16 *) (*p_argv);
-	      break;
-
-	    case FFI_TYPE_STRUCT:
-	      memcpy (argp, *p_argv, (*p_arg)->size);
-	      break;
-
-	    default:
-	      FFI_ASSERT (0);
-	    }
-	}
-      else if (z == sizeof (int))
-	{
-	  *(unsigned int *) argp = (unsigned int) *(UINT32 *) (*p_argv);
-	}
-      else
-	{
-	  if ((*p_arg)->type == FFI_TYPE_STRUCT)
-	    {
-	      memcpy (argp, *p_argv, z);
-	    }
-	  else
-	    {
-	      /* Double or long long 64bit.  */
-	      memcpy (argp, *p_argv, z);
-	    }
-	}
-      p_argv++;
-      argp += z;
-    }
-
-  return;
-}
-
 /* Perform machine dependent cif processing.  */
 ffi_status
 ffi_prep_cif_machdep (ffi_cif * cif)
@@ -157,85 +97,250 @@ ffi_prep_cif_machdep (ffi_cif * cif)
   return FFI_OK;
 }
 
-extern void ffi_call_ARCompact (void (*)(char *, extended_cif *),
-				extended_cif *, unsigned, unsigned,
-				unsigned *, void (*fn) (void));
-
-void
-ffi_call (ffi_cif * cif, void (*fn) (void), void *rvalue, void **avalue)
-{
-  extended_cif ecif;
-
-  ecif.cif = cif;
-  ecif.avalue = avalue;
+/* allocates a single register, float register, or XLEN-sized stack slot to a datum */
+static void marshal_atom(call_builder *cb, int type, void *data) {
+  size_t value = 0;
+  switch (type) {
+    case FFI_TYPE_UINT8: value = *(uint8_t *)data; break;
+    case FFI_TYPE_SINT8: value = *(int8_t *)data; break;
+    case FFI_TYPE_UINT16: value = *(uint16_t *)data; break;
+    case FFI_TYPE_SINT16: value = *(int16_t *)data; break;
+    /* 32-bit quantities are always sign-extended in the ABI */
+    case FFI_TYPE_UINT32: value = *(int32_t *)data; break;
+    case FFI_TYPE_SINT32: value = *(int32_t *)data; break;
+#if defined(__ARC64_ARCH64__)
+    case FFI_TYPE_UINT64: value = *(uint64_t *)data; break;
+    case FFI_TYPE_SINT64: value = *(int64_t *)data; break;
+#endif
+    case FFI_TYPE_POINTER: value = *(size_t *)data; break;
+    default: FFI_ASSERT(0); break;
+  }
+
+  if (cb->used_integer == NARGREG) {
+    *cb->used_stack++ = value;
+  } else {
+    cb->aregs->r[cb->used_integer++] = value;
+  }
+}
 
-  /* If the return value is a struct and we don't have
-     a return value address then we need to make one.  */
-  if ((rvalue == NULL) && (cif->rtype->type == FFI_TYPE_STRUCT))
-    {
-      ecif.rvalue = alloca (cif->rtype->size);
+/* adds an argument to a call, or a not by reference return value */
+static void marshal(call_builder *cb, ffi_type *type, int var, void *data) {
+  size_t realign[2];
+
+#if (defined(__ARC64_ARCH64__) || defined(__ARC64_ARCH32__))
+  if (type->size > 2 * __SIZEOF_POINTER__) {
+    if (var) {
+      marshal_atom(cb, FFI_TYPE_POINTER, &data);
+    } else {
+      /* copy to stack and pass by reference */
+      data = memcpy (cb->struct_stack, data, type->size);
+      cb->struct_stack = (size_t *) FFI_ALIGN ((char *) cb->struct_stack + type->size, __SIZEOF_POINTER__);
+      marshal_atom(cb, FFI_TYPE_POINTER, &data);
     }
-  else
-    ecif.rvalue = rvalue;
+  }
+#else
+  if (type->type == FFI_TYPE_STRUCT) {
+    if (var) {
+      if (type->size > 0)
+        marshal_atom(cb, FFI_TYPE_POINTER, data);
+    } else {
+      int i;
+      
+      for (i = 0; i < type->size; i += sizeof(size_t)) {
+        marshal_atom(cb, FFI_TYPE_POINTER, data);
+        data += sizeof(size_t);
+      }
+    }
+  }
+#endif
+  else if (IS_INT(type->type) || type->type == FFI_TYPE_POINTER) {
+    marshal_atom(cb, type->type, data);
+  } else {
+      memcpy(realign, data, type->size);
+      if (type->size > 0)
+        marshal_atom(cb, FFI_TYPE_POINTER, realign);
+      if (type->size > __SIZEOF_POINTER__)
+        marshal_atom(cb, FFI_TYPE_POINTER, realign + 1);
+  }
+}
 
-  switch (cif->abi)
-    {
-    case FFI_ARCOMPACT:
-      ffi_call_ARCompact (ffi_prep_args, &ecif, cif->bytes,
-			  cif->flags, ecif.rvalue, fn);
-      break;
+static void unmarshal_atom(call_builder *cb, int type, void *data) {
+  size_t value;
+
+  if (cb->used_integer == NARGREG) {
+    value = *cb->used_stack++;
+  } else {
+    value = cb->aregs->r[cb->used_integer++];
+  }
+
+  switch (type) {
+    case FFI_TYPE_UINT8: *(uint8_t *)data = value; break;
+    case FFI_TYPE_SINT8: *(uint8_t *)data = value; break;
+    case FFI_TYPE_UINT16: *(uint16_t *)data = value; break;
+    case FFI_TYPE_SINT16: *(uint16_t *)data = value; break;
+    case FFI_TYPE_UINT32: *(uint32_t *)data = value; break;
+    case FFI_TYPE_SINT32: *(uint32_t *)data = value; break;
+#if defined(__ARC64_ARCH64__)
+    case FFI_TYPE_UINT64: *(uint64_t *)data = value; break;
+    case FFI_TYPE_SINT64: *(uint64_t *)data = value; break;
+#endif
+    case FFI_TYPE_POINTER: *(size_t *)data = value; break;
+    default: FFI_ASSERT(0); break;
+  }
+}
 
-    default:
-      FFI_ASSERT (0);
-      break;
+/* for arguments passed by reference returns the pointer, otherwise the arg is copied (up to MAXCOPYARG bytes) */
+static void *unmarshal(call_builder *cb, ffi_type *type, int var, void *data) {
+  size_t realign[2];
+  void *pointer;
+
+#if defined(__ARC64_ARCH64__)
+  if (type->size > 2 * __SIZEOF_POINTER__) {
+        /* pass by reference */
+        unmarshal_atom(cb, FFI_TYPE_POINTER, (char*)&pointer);
+        return pointer;
     }
+#elif defined(__ARC64_ARCH32__)
+  if (type->type == FFI_TYPE_STRUCT) {
+    if (type->size > 2 * __SIZEOF_POINTER__) {
+      unmarshal_atom(cb, FFI_TYPE_POINTER, &realign[0]);
+      memcpy(data, (const void*)realign[0], type->size);
+      return data;
+    } else {
+      int i;
+      void *pdata = data;
+
+      for (i = 0; i < type->size; i += sizeof(size_t)) {
+        unmarshal_atom(cb, FFI_TYPE_POINTER, pdata);
+        pdata += sizeof(size_t);
+      }
+      return data;
+    }
+  }
+#else
+  if (type->type == FFI_TYPE_STRUCT) {
+
+      if (var) {
+        int i;
+        void *pdata = data;
+
+        for (i = 0; i < type->size; i += sizeof(size_t)) {
+          unmarshal_atom(cb, FFI_TYPE_POINTER, pdata);
+          pdata += sizeof(size_t);
+        }
+        return data;
+      } else {
+        if (type->size > 0)
+          unmarshal_atom(cb, FFI_TYPE_POINTER, &realign[0]);
+        memcpy(data, (const void*)realign[0], type->size);
+        return data;
+      }
+  }
+#endif
+  else if (IS_INT(type->type) || type->type == FFI_TYPE_POINTER) {
+    unmarshal_atom(cb, type->type, data);
+    return data;
+  } else {
+      if (type->size > 0)
+        unmarshal_atom(cb, FFI_TYPE_POINTER, realign);
+      if (type->size > __SIZEOF_POINTER__)
+        unmarshal_atom(cb, FFI_TYPE_POINTER, realign + 1);
+      memcpy(data, realign, type->size);
+      return data;
+  }
 }
 
-int
-ffi_closure_inner_ARCompact (ffi_closure * closure, void *rvalue,
-			     ffi_arg * args)
+static int passed_by_ref(ffi_type *type, int var) {
+  if (type->type == FFI_TYPE_STRUCT)
+	  return 1;
+  
+  return type->size > 2 * __SIZEOF_POINTER__;
+}
+
+/* Low level routine for calling functions */
+extern void ffi_call_asm (void *stack, struct call_context *regs,
+			  void (*fn) (void), void *closure) FFI_HIDDEN;
+
+static void
+ffi_call_int (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue,
+	      void *closure)
 {
-  void **arg_area, **p_argv;
-  ffi_cif *cif = closure->cif;
-  char *argp = (char *) args;
-  ffi_type **p_argt;
-  int i;
+  int return_by_ref = passed_by_ref(cif->rtype, 0);
 
-  arg_area = (void **) alloca (cif->nargs * sizeof (void *));
+  /* Allocate space for stack arg parameters.  */
+  size_t arg_bytes = FFI_ALIGN(2 * sizeof(size_t) * cif->nargs, STKALIGN);
+  /* Allocate space for copies of big structures.  */
+  size_t struct_bytes = FFI_ALIGN(cif->bytes, STKALIGN);
+  // size_t rval_bytes = 0;
+  // if (rvalue == NULL && cif->rtype->size > 2*__SIZEOF_POINTER__)
+  //   rval_bytes = FFI_ALIGN(cif->rtype->size, STKALIGN);
+  size_t alloc_size = arg_bytes + /*rval_bytes +*/ struct_bytes + sizeof(call_context);
+  size_t alloc_base = (size_t)alloca(alloc_size);
 
-  /* handle hidden argument */
-  if (cif->flags == FFI_TYPE_STRUCT)
-    {
-      rvalue = *(void **) argp;
-      argp += 4;
-    }
+  // if (rval_bytes)
+  //   rvalue = (void*)(alloc_base + arg_bytes);
 
-  p_argv = arg_area;
+  call_builder cb;
+  cb.used_integer = 0;
+  cb.aregs = (call_context*)(alloc_base + arg_bytes /*+ rval_bytes*/ + struct_bytes);
+  cb.used_stack = (void*)alloc_base;
+  cb.struct_stack = (void *)(alloc_base + arg_bytes /*+ rval_bytes*/);
 
-  for (i = 0, p_argt = cif->arg_types; i < cif->nargs;
-       i++, p_argt++, p_argv++)
-    {
-      size_t z;
-      int alignment;
+  // if (cif->rtype->type == FFI_TYPE_STRUCT)
+  //   marshal(&cb, &ffi_type_pointer, 0, &rvalue);
+
+  if (return_by_ref)
+	  marshal(&cb, &ffi_type_pointer, 0, &rvalue);
 
-      /* align alignment to 4 */
-      alignment = (((*p_argt)->alignment - 1) | 3) + 1;
+  int i;
+  for (i = 0; i < cif->nargs; i++)
+    marshal(&cb, cif->arg_types[i], 0, avalue[i]);
 
-      /* Align if necessary.  */
-      if ((alignment - 1) & (unsigned) argp)
-	argp = (char *) FFI_ALIGN (argp, alignment);
+  ffi_call_asm ((void *) alloc_base, cb.aregs, fn, closure);
 
-      z = (*p_argt)->size;
-      *p_argv = (void *) argp;
-      argp += z;
+  cb.used_integer = 0;
+  if (!return_by_ref && rvalue)
+    {
+	    if (IS_INT(cif->rtype->type)
+	        && cif->rtype->size < sizeof (ffi_arg))
+	    {
+        /* Integer types smaller than ffi_arg need to be extended.  */
+	    switch (cif->rtype->type) {
+	      case FFI_TYPE_SINT8:
+	      case FFI_TYPE_SINT16:
+	      case FFI_TYPE_SINT32:
+		      unmarshal_atom (&cb, (sizeof (ffi_arg) > 4
+                          ? FFI_TYPE_SINT64 : FFI_TYPE_SINT32),
+				                  rvalue);
+		      break;
+	      case FFI_TYPE_UINT8:
+	      case FFI_TYPE_UINT16:
+	      case FFI_TYPE_UINT32:
+		      unmarshal_atom (&cb, (sizeof (ffi_arg) > 4
+				                  ? FFI_TYPE_UINT64 : FFI_TYPE_UINT32),
+				                  rvalue);
+		      break;
+	      }
+	    }
+	    else
+	      unmarshal(&cb, cif->rtype, 0, rvalue);
     }
+}
 
-  (closure->fun) (cif, rvalue, arg_area, closure->user_data);
+void
+ffi_call (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, NULL);
+}
 
-  return cif->flags;
+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_ARCompact (void);
+extern void ffi_closure_asm(void) FFI_HIDDEN;
 
 ffi_status
 ffi_prep_closure_loc (ffi_closure * closure, ffi_cif * cif,
@@ -243,15 +348,28 @@ ffi_prep_closure_loc (ffi_closure * closure, ffi_cif * cif,
 		      void *user_data, void *codeloc)
 {
   uint32_t *tramp = (uint32_t *) & (closure->tramp[0]);
+  size_t address_ffi_closure = (size_t) ffi_closure_asm;
 
   switch (cif->abi)
     {
+#if defined(__ARC64_ARCH64__)
+    case FFI_ARC64:
+      FFI_ASSERT (tramp == codeloc);
+      tramp[0] = CODE_ENDIAN (0x580a1fc0);	/* movl r8, pcl  */
+      tramp[1] = CODE_ENDIAN (0x5c0b1f80);	/* movhl r12, limm */
+      tramp[2] = CODE_ENDIAN ((uint32_t)(address_ffi_closure >> 32));
+      tramp[3] = CODE_ENDIAN (0x5c051f8c);	/* orl r12, r12, limm */
+      tramp[4] = CODE_ENDIAN ((uint32_t)(address_ffi_closure & 0xffffffff));
+      tramp[5] = CODE_ENDIAN (0x20200300);	/* j [r12] */
+      break;
+#else
     case FFI_ARCOMPACT:
       FFI_ASSERT (tramp == codeloc);
       tramp[0] = CODE_ENDIAN (0x200a1fc0);	/* mov r8, pcl  */
       tramp[1] = CODE_ENDIAN (0x20200f80);	/* j [long imm] */
-      tramp[2] = CODE_ENDIAN (ffi_closure_ARCompact);
+      tramp[2] = CODE_ENDIAN (ffi_closure_asm);
       break;
+#endif
 
     default:
       return FFI_BAD_ABI;
@@ -264,3 +382,62 @@ ffi_prep_closure_loc (ffi_closure * closure, ffi_cif * cif,
 
   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 (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
+       of memory, but allocating disjoint storage for each argument is
+       simpler. */
+    char *astorage = alloca(cif->bytes);
+    char *ptr = astorage;
+    void *rvalue;
+    call_builder cb;
+    int i;
+
+    cb.aregs = aregs;
+    cb.used_integer = 0;
+    cb.used_stack = stack;
+    
+    /* handle hidden argument */
+    if (cif->flags == FFI_TYPE_STRUCT)
+      unmarshal(&cb, &ffi_type_pointer, 0, &rvalue);
+    else
+      rvalue = alloca(cif->rtype->size);
+
+    for (i = 0; i < cif->nargs; i++) {
+      avalue[i] = unmarshal(&cb, cif->arg_types[i], 1, ptr);
+      ptr += cif->arg_types[i]->size;
+    }
+
+    fun (cif, rvalue, avalue, user_data);
+
+    if (cif->rtype->type != FFI_TYPE_VOID) {
+        cb.used_integer = 0;
+        marshal(&cb, cif->rtype, 1, rvalue);
+    }
+}
diff --git a/src/arc/ffitarget.h b/src/arc/ffitarget.h
index bf8311bc..5b36902e 100644
--- a/src/arc/ffitarget.h
+++ b/src/arc/ffitarget.h
@@ -40,14 +40,28 @@ typedef signed long ffi_sarg;
 typedef enum ffi_abi
 {
   FFI_FIRST_ABI = 0,
+#if __SIZEOF_POINTER__ == 8
+  FFI_ARC64,
+#else
   FFI_ARCOMPACT,
+#endif
   FFI_LAST_ABI,
+#if __SIZEOF_POINTER__ == 8
+  FFI_DEFAULT_ABI = FFI_ARC64
+#else
   FFI_DEFAULT_ABI = FFI_ARCOMPACT
+#endif
 } ffi_abi;
 #endif
 
 #define FFI_CLOSURES 		1
+#define FFI_GO_CLOSURES 1
+#if __SIZEOF_POINTER__ == 8
+#define FFI_TRAMPOLINE_SIZE	24
+#else
 #define FFI_TRAMPOLINE_SIZE	12
+#endif
+
 #define FFI_NATIVE_RAW_API 	0
 
 #endif