[PATCH] New port for ARM AArch64

[PATCH] New port for ARM AArch64

[Marcus Shawcroft]

[-- Attachment #1: Type: text/plain, Size: 1186 bytes --]


ARM would like to contribute a libffi port for the ARM AArch64
architecture.  The port passes the test suite cleanly.  The proposed
ChangeLog and patches are included below.


/Marcus


2012-09-18  James Greenhalgh  <james.greenhalgh at arm.com>
	    Marcus Shawcroft  <marcus.shawcroft at arm.com>

	* README: Add details of aarch64 port.
	* src/aarch64/ffi.c: New.
	* src/aarch64/ffitarget.h: Likewise.
	* src/aarch64/sysv.S: Likewise.

2012-09-18  James Greenhalgh  <james.greenhalgh at arm.com>
	    Marcus Shawcroft  <marcus.shawcroft at arm.com>

	* testsuite/lib/libffi.exp: Add support for aarch64.
	* testsuite/libffi.call/cls_struct_va1.c: New.
	* testsuite/libffi.call/cls_uchar_va.c: Likewise.
	* testsuite/libffi.call/cls_uint_va.c: Likewise.
	* testsuite/libffi.call/cls_ulong_va.c: Liekwise.
	* testsuite/libffi.call/cls_ushort_va.c: Likewise.
	* testsuite/libffi.call/nested_struct11.c: Likewise.
	* testsuite/libffi.call/uninitialized.c: Likewise.
	* testsuite/libffi.call/va_1.c: Likewise.
	* testsuite/libffi.call/va_struct1.c: Likewise.
	* testsuite/libffi.call/va_struct2.c: Likewise.
	* testsuite/libffi.call/va_struct3.c: Likewise.

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: aarch64-libffi.diff --]
[-- Type: text/x-patch; name=aarch64-libffi.diff, Size: 43747 bytes --]

diff --git a/README b/README
index 0cf0720..8fc473f 100644
--- a/README
+++ b/README
@@ -51,6 +51,7 @@ tested:
 |--------------+------------------|
 | Architecture | Operating System |
 |--------------+------------------|
+| AArch64      | Linux            |
 | Alpha        | Linux            |
 | Alpha        | Tru64            |
 | ARM          | Linux            |
@@ -319,6 +320,7 @@ Thorup.
 Major processor architecture ports were contributed by the following
 developers:
 
+aarch64		Marcus Shawcroft, James Greenhalgh
 alpha		Richard Henderson
 arm		Raffaele Sena
 cris		Simon Posnjak, Hans-Peter Nilsson
diff --git a/src/aarch64/ffi.c b/src/aarch64/ffi.c
new file mode 100644
index 0000000..1405665
--- /dev/null
+++ b/src/aarch64/ffi.c
@@ -0,0 +1,1076 @@
+/* Copyright (c) 2009, 2010, 2011, 2012 ARM Ltd.
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+``Software''), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+#include <stdio.h>
+
+#include <ffi.h>
+#include <ffi_common.h>
+
+#include <stdlib.h>
+
+/* Stack alignment requirement in bytes */
+#define AARCH64_STACK_ALIGN 16
+
+#define N_X_ARG_REG 8
+#define N_V_ARG_REG 8
+
+#define AARCH64_FFI_WITH_V (1 << AARCH64_FFI_WITH_V_BIT)
+
+union _d
+{
+  UINT64 d;
+  UINT32 s[2];
+};
+
+struct call_context
+{
+  UINT64 x [AARCH64_N_XREG];
+  struct
+  {
+    union _d d[2];
+  } v [AARCH64_N_VREG];
+};
+
+static void *
+get_x_addr (struct call_context *context, unsigned n)
+{
+  return &context->x[n];
+}
+
+static void *
+get_s_addr (struct call_context *context, unsigned n)
+{
+#if defined __AARCH64EB__
+  return &context->v[n].d[1].s[1];
+#else
+  return &context->v[n].d[0].s[0];
+#endif
+}
+
+static void *
+get_d_addr (struct call_context *context, unsigned n)
+{
+#if defined __AARCH64EB__
+  return &context->v[n].d[1];
+#else
+  return &context->v[n].d[0];
+#endif
+}
+
+static void *
+get_v_addr (struct call_context *context, unsigned n)
+{
+  return &context->v[n];
+}
+
+/* Return the memory location at which a basic type would reside
+   were it to have been stored in register n.  */
+
+static void *
+get_basic_type_addr (unsigned short type, struct call_context *context,
+		     unsigned n)
+{
+  switch (type)
+    {
+    case FFI_TYPE_FLOAT:
+      return get_s_addr (context, n);
+    case FFI_TYPE_DOUBLE:
+      return get_d_addr (context, n);
+    case FFI_TYPE_LONGDOUBLE:
+      return get_v_addr (context, n);
+    case FFI_TYPE_UINT8:
+    case FFI_TYPE_SINT8:
+    case FFI_TYPE_UINT16:
+    case FFI_TYPE_SINT16:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_INT:
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+      return get_x_addr (context, n);
+    default:
+      FFI_ASSERT (0);
+      return NULL;
+    }
+}
+
+/* Return the alignment width for each of the basic types.  */
+
+static size_t
+get_basic_type_alignment (unsigned short type)
+{
+  switch (type)
+    {
+    case FFI_TYPE_FLOAT:
+    case FFI_TYPE_DOUBLE:
+      return sizeof (UINT64);
+    case FFI_TYPE_LONGDOUBLE:
+      return sizeof (long double);
+    case FFI_TYPE_UINT8:
+    case FFI_TYPE_SINT8:
+    case FFI_TYPE_UINT16:
+    case FFI_TYPE_SINT16:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+      return sizeof (UINT64);
+
+    default:
+      FFI_ASSERT (0);
+      return 0;
+    }
+}
+
+/* Return the size in bytes for each of the basic types.  */
+
+static size_t
+get_basic_type_size (unsigned short type)
+{
+  switch (type)
+    {
+    case FFI_TYPE_FLOAT:
+      return sizeof (UINT32);
+    case FFI_TYPE_DOUBLE:
+      return sizeof (UINT64);
+    case FFI_TYPE_LONGDOUBLE:
+      return sizeof (long double);
+    case FFI_TYPE_UINT8:
+      return sizeof (UINT8);
+    case FFI_TYPE_SINT8:
+      return sizeof (SINT8);
+    case FFI_TYPE_UINT16:
+      return sizeof (UINT16);
+    case FFI_TYPE_SINT16:
+      return sizeof (SINT16);
+    case FFI_TYPE_UINT32:
+      return sizeof (UINT32);
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+      return sizeof (SINT32);
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_UINT64:
+      return sizeof (UINT64);
+    case FFI_TYPE_SINT64:
+      return sizeof (SINT64);
+
+    default:
+      FFI_ASSERT (0);
+      return 0;
+    }
+}
+
+extern void
+ffi_call_SYSV (unsigned (*)(struct call_context *context, unsigned char *,
+			    extended_cif *),
+               struct call_context *context,
+               extended_cif *,
+               unsigned,
+               void (*fn)(void));
+
+extern void
+ffi_closure_SYSV (ffi_closure *);
+
+/* Test for an FFI floating point representation.  */
+
+static unsigned
+is_floating_type (unsigned short type)
+{
+  return (type == FFI_TYPE_FLOAT || type == FFI_TYPE_DOUBLE
+	  || type == FFI_TYPE_LONGDOUBLE);
+}
+
+/* Test for a homogeneous structure.  */
+
+static unsigned short
+get_homogeneous_type (ffi_type *ty)
+{
+  if (ty->type == FFI_TYPE_STRUCT && ty->elements)
+    {
+      unsigned i;
+      unsigned short candidate_type
+	= get_homogeneous_type (ty->elements[0]);
+      for (i =1; ty->elements[i]; i++)
+	{
+	  unsigned short iteration_type = 0;
+	  /* If we have a nested struct, we must find its homogeneous type.
+	     If that fits with our candidate type, we are still
+	     homogeneous.  */
+	  if (ty->elements[i]->type == FFI_TYPE_STRUCT
+	      && ty->elements[i]->elements)
+	    {
+	      iteration_type = get_homogeneous_type (ty->elements[i]);
+	    }
+	  else
+	    {
+	      iteration_type = ty->elements[i]->type;
+	    }
+
+	  /* If we are not homogeneous, return FFI_TYPE_STRUCT.  */
+	  if (candidate_type != iteration_type)
+	    return FFI_TYPE_STRUCT;
+	}
+      return candidate_type;
+    }
+
+  /* Base case, we have no more levels of nesting, so we
+     are a basic type, and so, trivially homogeneous in that type.  */
+  return ty->type;
+}
+
+/* Determine the number of elements within a STRUCT.
+
+   Note, we must handle nested structs.
+
+   If ty is not a STRUCT this function will return 0.  */
+
+static unsigned
+element_count (ffi_type *ty)
+{
+  if (ty->type == FFI_TYPE_STRUCT && ty->elements)
+    {
+      unsigned n;
+      unsigned elems = 0;
+      for (n = 0; ty->elements[n]; n++)
+	{
+	  if (ty->elements[n]->type == FFI_TYPE_STRUCT
+	      && ty->elements[n]->elements)
+	    elems += element_count (ty->elements[n]);
+	  else
+	    elems++;
+	}
+      return elems;
+    }
+  return 0;
+}
+
+/* Test for a homogeneous floating point aggregate.
+
+   A homogeneous floating point aggregate is a homogeneous aggregate of
+   a half- single- or double- precision floating point type with one
+   to four elements.  Note that this includes nested structs of the
+   basic type.  */
+
+static int
+is_hfa (ffi_type *ty)
+{
+  if (ty->type == FFI_TYPE_STRUCT
+      && ty->elements[0]
+      && is_floating_type (get_homogeneous_type (ty)))
+    {
+      unsigned n = element_count (ty);
+      return n >= 1 && n <= 4;
+    }
+  return 0;
+}
+
+/* Test if an ffi_type is a candidate for passing in a register.
+
+   This test does not check that sufficient registers of the
+   appropriate class are actually available, merely that IFF
+   sufficient registers are available then the argument will be passed
+   in register(s).
+
+   Note that an ffi_type that is deemed to be a register candidate
+   will always be returned in registers.
+
+   Returns 1 if a register candidate else 0.  */
+
+static int
+is_register_candidate (ffi_type *ty)
+{
+  switch (ty->type)
+    {
+    case FFI_TYPE_VOID:
+    case FFI_TYPE_FLOAT:
+    case FFI_TYPE_DOUBLE:
+    case FFI_TYPE_LONGDOUBLE:
+    case FFI_TYPE_UINT8:
+    case FFI_TYPE_UINT16:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_SINT8:
+    case FFI_TYPE_SINT16:
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT64:
+      return 1;
+
+    case FFI_TYPE_STRUCT:
+      if (is_hfa (ty))
+        {
+          return 1;
+        }
+      else if (ty->size > 16)
+        {
+          /* Too large. Will be replaced with a pointer to memory. The
+             pointer MAY be passed in a register, but the value will
+             not. This test specifically fails since the argument will
+             never be passed by value in registers. */
+          return 0;
+        }
+      else
+        {
+          /* Might be passed in registers depending on the number of
+             registers required. */
+          return (ty->size + 7) / 8 < N_X_ARG_REG;
+        }
+      break;
+
+    default:
+      FFI_ASSERT (0);
+      break;
+    }
+
+  return 0;
+}
+
+/* Test if an ffi_type argument or result is a candidate for a vector
+   register.  */
+
+static int
+is_v_register_candidate (ffi_type *ty)
+{
+  return is_floating_type (ty->type)
+	   || (ty->type == FFI_TYPE_STRUCT && is_hfa (ty));
+}
+
+/* Representation of the procedure call argument marshalling
+   state.
+
+   The terse state variable names match the names used in the AARCH64
+   PCS. */
+
+struct arg_state
+{
+  unsigned ngrn;                /* Next general-purpose register number. */
+  unsigned nsrn;                /* Next vector register number. */
+  unsigned nsaa;                /* Next stack offset. */
+};
+
+/* Initialize a procedure call argument marshalling state.  */
+static void
+arg_init (struct arg_state *state, unsigned call_frame_size)
+{
+  state->ngrn = 0;
+  state->nsrn = 0;
+  state->nsaa = 0;
+}
+
+/* Return the number of available consecutive core argument
+   registers.  */
+
+static unsigned
+available_x (struct arg_state *state)
+{
+  return N_X_ARG_REG - state->ngrn;
+}
+
+/* Return the number of available consecutive vector argument
+   registers.  */
+
+static unsigned
+available_v (struct arg_state *state)
+{
+  return N_V_ARG_REG - state->nsrn;
+}
+
+static void *
+allocate_to_x (struct call_context *context, struct arg_state *state)
+{
+  FFI_ASSERT (state->ngrn < N_X_ARG_REG)
+  return get_x_addr (context, (state->ngrn)++);
+}
+
+static void *
+allocate_to_s (struct call_context *context, struct arg_state *state)
+{
+  FFI_ASSERT (state->nsrn < N_V_ARG_REG)
+  return get_s_addr (context, (state->nsrn)++);
+}
+
+static void *
+allocate_to_d (struct call_context *context, struct arg_state *state)
+{
+  FFI_ASSERT (state->nsrn < N_V_ARG_REG)
+  return get_d_addr (context, (state->nsrn)++);
+}
+
+static void *
+allocate_to_v (struct call_context *context, struct arg_state *state)
+{
+  FFI_ASSERT (state->nsrn < N_V_ARG_REG)
+  return get_v_addr (context, (state->nsrn)++);
+}
+
+/* Allocate an aligned slot on the stack and return a pointer to it.  */
+static void *
+allocate_to_stack (struct arg_state *state, void *stack, unsigned alignment,
+		   unsigned size)
+{
+  void *allocation;
+
+  /* Round up the NSAA to the larger of 8 or the natural
+     alignment of the argument's type.  */
+  state->nsaa = ALIGN (state->nsaa, alignment);
+  state->nsaa = ALIGN (state->nsaa, alignment);
+  state->nsaa = ALIGN (state->nsaa, 8);
+
+  allocation = stack + state->nsaa;
+
+  state->nsaa += size;
+  return allocation;
+}
+
+static void
+copy_basic_type (void *dest, void *source, unsigned short type)
+{
+  /* This is neccessary to ensure that basic types are copied
+     sign extended to 64-bits as libffi expects.  */
+  switch (type)
+    {
+    case FFI_TYPE_FLOAT:
+      *(float *) dest = *(float *) source;
+      break;
+    case FFI_TYPE_DOUBLE:
+      *(double *) dest = *(double *) source;
+      break;
+    case FFI_TYPE_LONGDOUBLE:
+      *(long double *) dest = *(long double *) source;
+      break;
+    case FFI_TYPE_UINT8:
+      *(ffi_arg *) dest = *(UINT8 *) source;
+      break;
+    case FFI_TYPE_SINT8:
+      *(ffi_sarg *) dest = *(SINT8 *) source;
+      break;
+    case FFI_TYPE_UINT16:
+      *(ffi_arg *) dest = *(UINT16 *) source;
+      break;
+    case FFI_TYPE_SINT16:
+      *(ffi_sarg *) dest = *(SINT16 *) source;
+      break;
+    case FFI_TYPE_UINT32:
+      *(ffi_arg *) dest = *(UINT32 *) source;
+      break;
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+      *(ffi_sarg *) dest = *(SINT32 *) source;
+      break;
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_UINT64:
+      *(ffi_arg *) dest = *(UINT64 *) source;
+      break;
+    case FFI_TYPE_SINT64:
+      *(ffi_sarg *) dest = *(SINT64 *) source;
+      break;
+
+    default:
+      FFI_ASSERT (0);
+    }
+}
+
+static void
+copy_hfa_to_reg_or_stack (void *memory,
+			  ffi_type *ty,
+			  struct call_context *context,
+			  unsigned char *stack,
+			  struct arg_state *state)
+{
+  unsigned elems = element_count (ty);
+  if (available_v (state) < elems)
+    {
+      /* There are insufficient V registers. Further V register allocations
+	 are prevented, the NSAA is adjusted (by allocate_to_stack ())
+	 and the argument is copied to memory at the adjusted NSAA.  */
+      state->nsrn = N_V_ARG_REG;
+      memcpy (allocate_to_stack (state, stack, ty->alignment, ty->size),
+	      memory,
+	      ty->size);
+    }
+  else
+    {
+      int i;
+      unsigned short type = get_homogeneous_type (ty);
+      unsigned elems = element_count (ty);
+      for (i = 0; i < elems; i++)
+	{
+	  void *reg = allocate_to_v (context, state);
+	  copy_basic_type (reg, memory, type);
+	  memory += get_basic_type_size (type);
+	}
+    }
+}
+
+/* Either allocate an appropriate register for the argument type, or if
+   none are available, allocate a stack slot and return a pointer
+   to the allocated space.  */
+
+static void *
+allocate_to_register_or_stack (struct call_context *context,
+			       unsigned char *stack,
+			       struct arg_state *state,
+			       unsigned short type)
+{
+  size_t alignment = get_basic_type_alignment (type);
+  size_t size = alignment;
+  switch (type)
+    {
+    case FFI_TYPE_FLOAT:
+      /* This is the only case for which the allocated stack size
+	 should not match the alignment of the type.  */
+      size = sizeof (UINT32);
+      /* Fall through.  */
+    case FFI_TYPE_DOUBLE:
+      if (state->nsrn < N_V_ARG_REG)
+	return allocate_to_d (context, state);
+      state->nsrn = N_V_ARG_REG;
+      break;
+    case FFI_TYPE_LONGDOUBLE:
+      if (state->nsrn < N_V_ARG_REG)
+	return allocate_to_v (context, state);
+      state->nsrn = N_V_ARG_REG;
+      break;
+    case FFI_TYPE_UINT8:
+    case FFI_TYPE_SINT8:
+    case FFI_TYPE_UINT16:
+    case FFI_TYPE_SINT16:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_INT:
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+      if (state->ngrn < N_X_ARG_REG)
+	return allocate_to_x (context, state);
+      state->ngrn = N_X_ARG_REG;
+      break;
+    default:
+      FFI_ASSERT (0);
+    }
+
+    return allocate_to_stack (state, stack, alignment, size);
+}
+
+/* Copy a value to an appropriate register, or if none are
+   available, to the stack.  */
+
+static void
+copy_to_register_or_stack (struct call_context *context,
+			   unsigned char *stack,
+			   struct arg_state *state,
+			   void *value,
+			   unsigned short type)
+{
+  copy_basic_type (
+	  allocate_to_register_or_stack (context, stack, state, type),
+	  value,
+	  type);
+}
+
+/* Marshall the arguments from FFI representation to procedure call
+   context and stack.  */
+
+static unsigned
+aarch64_prep_args (struct call_context *context, unsigned char *stack,
+		   extended_cif *ecif)
+{
+  int i;
+  struct arg_state state;
+
+  arg_init (&state, ALIGN(ecif->cif->bytes, 16));
+
+  for (i = 0; i < ecif->cif->nargs; i++)
+    {
+      ffi_type *ty = ecif->cif->arg_types[i];
+      switch (ty->type)
+	{
+	case FFI_TYPE_VOID:
+	  FFI_ASSERT (0);
+	  break;
+
+	/* If the argument is a basic type the argument is allocated to an
+	   appropriate register, or if none are available, to the stack.  */
+	case FFI_TYPE_FLOAT:
+	case FFI_TYPE_DOUBLE:
+	case FFI_TYPE_LONGDOUBLE:
+	case FFI_TYPE_UINT8:
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT16:
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_INT:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_POINTER:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	  copy_to_register_or_stack (context, stack, &state,
+				     ecif->avalue[i], ty->type);
+	  break;
+
+	case FFI_TYPE_STRUCT:
+	  if (is_hfa (ty))
+	    {
+	      copy_hfa_to_reg_or_stack (ecif->avalue[i], ty, context,
+					stack, &state);
+	    }
+	  else if (ty->size > 16)
+	    {
+	      /* If the argument is a composite type that is larger than 16
+		 bytes, then the argument has been copied to memory, and
+		 the argument is replaced by a pointer to the copy.  */
+
+	      copy_to_register_or_stack (context, stack, &state,
+					 &(ecif->avalue[i]), FFI_TYPE_POINTER);
+	    }
+	  else if (available_x (&state) >= (ty->size + 7) / 8)
+	    {
+	      /* If the argument is a composite type and the size in
+		 double-words is not more than the number of available
+		 X registers, then the argument is copied into consecutive
+		 X registers.  */
+	      int j;
+	      for (j = 0; j < (ty->size + 7) / 8; j++)
+		{
+		  memcpy (allocate_to_x (context, &state),
+			  &(((UINT64 *) ecif->avalue[i])[j]),
+			  sizeof (UINT64));
+		}
+	    }
+	  else
+	    {
+	      /* Otherwise, there are insufficient X registers. Further X
+		 register allocations are prevented, the NSAA is adjusted
+		 (by allocate_to_stack ()) and the argument is copied to
+		 memory at the adjusted NSAA.  */
+	      state.ngrn = N_X_ARG_REG;
+
+	      memcpy (allocate_to_stack (&state, stack, ty->alignment,
+					 ty->size), ecif->avalue + i, ty->size);
+	    }
+	  break;
+
+	default:
+	  FFI_ASSERT (0);
+	  break;
+	}
+    }
+
+  return ecif->cif->aarch64_flags;
+}
+
+ffi_status
+ffi_prep_cif_machdep (ffi_cif *cif)
+{
+  /* Round the stack up to a multiple of the stack alignment requirement. */
+  cif->bytes =
+    (cif->bytes + (AARCH64_STACK_ALIGN - 1)) & ~ (AARCH64_STACK_ALIGN - 1);
+
+  /* Initialize our flags. We are interested if this CIF will touch a
+     vector register, if so we will enable context save and load to
+     those registers, otherwise not. This is intended to be friendly
+     to lazy float context switching in the kernel.  */
+  cif->aarch64_flags = 0;
+
+  if (is_v_register_candidate (cif->rtype))
+    {
+      cif->aarch64_flags |= AARCH64_FFI_WITH_V;
+    }
+  else
+    {
+      int i;
+      for (i = 0; i < cif->nargs; i++)
+        if (is_v_register_candidate (cif->arg_types[i]))
+          {
+            cif->aarch64_flags |= AARCH64_FFI_WITH_V;
+            break;
+          }
+    }
+
+  return FFI_OK;
+}
+
+/* Call a function with the provided arguments and capture the return
+   value.  */
+void
+ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  extended_cif ecif;
+
+  ecif.cif = cif;
+  ecif.avalue = avalue;
+  ecif.rvalue = rvalue;
+
+  switch (cif->abi)
+    {
+    case FFI_SYSV:
+      {
+        struct call_context context;
+	unsigned stack_bytes;
+
+	/* Figure out the total amount of stack space we need, the
+	   above call frame space needs to be 16 bytes aligned to
+	   ensure correct alignment of the first object inserted in
+	   that space hence the ALIGN applied to cif->bytes.*/
+	stack_bytes = ALIGN(cif->bytes, 16);
+
+	memset (&context, 0, sizeof (context));
+        if (is_register_candidate (cif->rtype))
+          {
+            ffi_call_SYSV (aarch64_prep_args, &context, &ecif, stack_bytes, fn);
+            switch (cif->rtype->type)
+              {
+              case FFI_TYPE_VOID:
+              case FFI_TYPE_FLOAT:
+              case FFI_TYPE_DOUBLE:
+              case FFI_TYPE_LONGDOUBLE:
+              case FFI_TYPE_UINT8:
+              case FFI_TYPE_SINT8:
+              case FFI_TYPE_UINT16:
+              case FFI_TYPE_SINT16:
+              case FFI_TYPE_UINT32:
+              case FFI_TYPE_SINT32:
+              case FFI_TYPE_POINTER:
+              case FFI_TYPE_UINT64:
+              case FFI_TYPE_INT:
+              case FFI_TYPE_SINT64:
+		{
+		  void *addr = get_basic_type_addr (cif->rtype->type,
+						    &context, 0);
+		  copy_basic_type (rvalue, addr, cif->rtype->type);
+		  break;
+		}
+
+              case FFI_TYPE_STRUCT:
+                if (is_hfa (cif->rtype))
+		  {
+		    int j;
+		    unsigned short type = get_homogeneous_type (cif->rtype);
+		    unsigned elems = element_count (cif->rtype);
+		    for (j = 0; j < elems; j++)
+		      {
+			void *reg = get_basic_type_addr (type, &context, j);
+			copy_basic_type (rvalue, reg, type);
+			rvalue += get_basic_type_size (type);
+		      }
+		  }
+                else if ((cif->rtype->size + 7) / 8 < N_X_ARG_REG)
+                  {
+                    unsigned size = ALIGN (cif->rtype->size, sizeof (UINT64));
+                    memcpy (rvalue, get_x_addr (&context, 0), size);
+                  }
+                else
+                  {
+                    FFI_ASSERT (0);
+                  }
+                break;
+
+              default:
+                FFI_ASSERT (0);
+                break;
+              }
+          }
+        else
+          {
+            memcpy (get_x_addr (&context, 8), &rvalue, sizeof (UINT64));
+            ffi_call_SYSV (aarch64_prep_args, &context, &ecif,
+			   stack_bytes, fn);
+          }
+        break;
+      }
+
+    default:
+      FFI_ASSERT (0);
+      break;
+    }
+}
+
+static unsigned char trampoline [] =
+{ 0x70, 0x00, 0x00, 0x58,	/* ldr	x16, 1f	*/
+  0x91, 0x00, 0x00, 0x10,	/* adr	x17, 2f	*/
+  0x00, 0x02, 0x1f, 0xd6	/* br	x16	*/
+};
+
+/* Build a trampoline.  */
+
+#define FFI_INIT_TRAMPOLINE(TRAMP,FUN,CTX,FLAGS)			\
+  ({unsigned char *__tramp = (unsigned char*)(TRAMP);			\
+    UINT64  __fun = (UINT64)(FUN);					\
+    UINT64  __ctx = (UINT64)(CTX);					\
+    UINT64  __flags = (UINT64)(FLAGS);					\
+    memcpy (__tramp, trampoline, sizeof (trampoline));			\
+    memcpy (__tramp + 12, &__fun, sizeof (__fun));			\
+    memcpy (__tramp + 20, &__ctx, sizeof (__ctx));			\
+    memcpy (__tramp + 28, &__flags, sizeof (__flags));			\
+    __clear_cache(__tramp, __tramp + FFI_TRAMPOLINE_SIZE);		\
+  })
+
+ffi_status
+ffi_prep_closure_loc (ffi_closure* closure,
+                      ffi_cif* cif,
+                      void (*fun)(ffi_cif*,void*,void**,void*),
+                      void *user_data,
+                      void *codeloc)
+{
+  if (cif->abi != FFI_SYSV)
+    return FFI_BAD_ABI;
+
+  FFI_INIT_TRAMPOLINE (&closure->tramp[0], &ffi_closure_SYSV, codeloc,
+		       cif->aarch64_flags);
+
+  closure->cif  = cif;
+  closure->user_data = user_data;
+  closure->fun  = fun;
+
+  return FFI_OK;
+}
+
+/* Primary handler to setup and invoke a function within a closure.
+
+   A closure when invoked enters via the assembler wrapper
+   ffi_closure_SYSV(). The wrapper allocates a call context on the
+   stack, saves the interesting registers (from the perspective of
+   the calling convention) into the context then passes control to
+   ffi_closure_SYSV_inner() passing the saved context and a pointer to
+   the stack at the point ffi_closure_SYSV() was invoked.
+
+   On the return path the assembler wrapper will reload call context
+   regsiters.
+
+   ffi_closure_SYSV_inner() marshalls the call context into ffi value
+   desriptors, invokes the wrapped function, then marshalls the return
+   value back into the call context.  */
+
+void
+ffi_closure_SYSV_inner (ffi_closure *closure, struct call_context *context,
+			void *stack)
+{
+  ffi_cif *cif = closure->cif;
+  void **avalue = (void**) alloca (cif->nargs * sizeof (void*));
+  void *rvalue = NULL;
+  int i;
+  struct arg_state state;
+
+  arg_init (&state, ALIGN(cif->bytes, 16));
+
+  for (i = 0; i < cif->nargs; i++)
+    {
+      ffi_type *ty = cif->arg_types[i];
+
+      switch (ty->type)
+	{
+	case FFI_TYPE_VOID:
+	  FFI_ASSERT (0);
+	  break;
+
+	case FFI_TYPE_UINT8:
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT16:
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_INT:
+	case FFI_TYPE_POINTER:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	case  FFI_TYPE_FLOAT:
+	case  FFI_TYPE_DOUBLE:
+	case  FFI_TYPE_LONGDOUBLE:
+	  avalue[i] = allocate_to_register_or_stack (context, stack,
+						     &state, ty->type);
+	  break;
+
+	case FFI_TYPE_STRUCT:
+	  if (is_hfa (ty))
+	    {
+	      unsigned n = element_count (ty);
+	      if (available_v (&state) < n)
+		{
+		  state.nsrn = N_V_ARG_REG;
+		  avalue[i] = allocate_to_stack (&state, stack, ty->alignment,
+						 ty->size);
+		}
+	      else
+		{
+		  switch (get_homogeneous_type (ty))
+		    {
+		    case FFI_TYPE_FLOAT:
+		      {
+			/* Eeek! We need a pointer to the structure,
+			   however the homogeneous float elements are
+			   being passed in individual S registers,
+			   therefore the structure is not represented as
+			   a contiguous sequence of bytes in our saved
+			   register context. We need to fake up a copy
+			   of the structure layed out in memory
+			   correctly. The fake can be tossed once the
+			   closure function has returned hence alloca()
+			   is sufficient. */
+			int j;
+			UINT32 *p = avalue[i] = alloca (ty->size);
+			for (j = 0; j < element_count (ty); j++)
+			  memcpy (&p[j],
+				  allocate_to_s (context, &state),
+				  sizeof (*p));
+			break;
+		      }
+
+		    case FFI_TYPE_DOUBLE:
+		      {
+			/* Eeek! We need a pointer to the structure,
+			   however the homogeneous float elements are
+			   being passed in individual S registers,
+			   therefore the structure is not represented as
+			   a contiguous sequence of bytes in our saved
+			   register context. We need to fake up a copy
+			   of the structure layed out in memory
+			   correctly. The fake can be tossed once the
+			   closure function has returned hence alloca()
+			   is sufficient. */
+			int j;
+			UINT64 *p = avalue[i] = alloca (ty->size);
+			for (j = 0; j < element_count (ty); j++)
+			  memcpy (&p[j],
+				  allocate_to_d (context, &state),
+				  sizeof (*p));
+			break;
+		      }
+
+		    case FFI_TYPE_LONGDOUBLE:
+			  memcpy (&avalue[i],
+				  allocate_to_v (context, &state),
+				  sizeof (*avalue));
+		      break;
+
+		    default:
+		      FFI_ASSERT (0);
+		      break;
+		    }
+		}
+	    }
+	  else if (ty->size > 16)
+	    {
+	      /* Replace Composite type of size greater than 16 with a
+		 pointer.  */
+	      memcpy (&avalue[i],
+		      allocate_to_register_or_stack (context, stack,
+						     &state, FFI_TYPE_POINTER),
+		      sizeof (avalue[i]));
+	    }
+	  else if (available_x (&state) >= (ty->size + 7) / 8)
+	    {
+	      avalue[i] = get_x_addr (context, state.ngrn);
+	      state.ngrn += (ty->size + 7) / 8;
+	    }
+	  else
+	    {
+	      state.ngrn = N_X_ARG_REG;
+
+	      avalue[i] = allocate_to_stack (&state, stack, ty->alignment,
+					     ty->size);
+	    }
+	  break;
+
+	default:
+	  FFI_ASSERT (0);
+	  break;
+	}
+    }
+
+  /* Figure out where the return value will be passed, either in
+     registers or in a memory block allocated by the caller and passed
+     in x8.  */
+
+  if (is_register_candidate (cif->rtype))
+    {
+      /* Register candidates are *always* returned in registers. */
+
+      /* Allocate a scratchpad for the return value, we will let the
+         callee scrible the result into the scratch pad then move the
+         contents into the appropriate return value location for the
+         call convention.  */
+      rvalue = alloca (cif->rtype->size);
+      (closure->fun) (cif, rvalue, avalue, closure->user_data);
+
+      /* Copy the return value into the call context so that it is returned
+         as expected to our caller.  */
+      switch (cif->rtype->type)
+        {
+        case FFI_TYPE_VOID:
+          break;
+
+        case FFI_TYPE_UINT8:
+        case FFI_TYPE_UINT16:
+        case FFI_TYPE_UINT32:
+        case FFI_TYPE_POINTER:
+        case FFI_TYPE_UINT64:
+        case FFI_TYPE_SINT8:
+        case FFI_TYPE_SINT16:
+        case FFI_TYPE_INT:
+        case FFI_TYPE_SINT32:
+        case FFI_TYPE_SINT64:
+        case FFI_TYPE_FLOAT:
+        case FFI_TYPE_DOUBLE:
+        case FFI_TYPE_LONGDOUBLE:
+	  {
+	    void *addr = get_basic_type_addr (cif->rtype->type, context, 0);
+	    copy_basic_type (addr, rvalue, cif->rtype->type);
+            break;
+	  }
+        case FFI_TYPE_STRUCT:
+          if (is_hfa (cif->rtype))
+	    {
+	      int i;
+	      unsigned short type = get_homogeneous_type (cif->rtype);
+	      unsigned elems = element_count (cif->rtype);
+	      for (i = 0; i < elems; i++)
+		{
+		  void *reg = get_basic_type_addr (type, context, i);
+		  copy_basic_type (reg, rvalue, type);
+		  rvalue += get_basic_type_size (type);
+		}
+	    }
+          else if ((cif->rtype->size + 7) / 8 < N_X_ARG_REG)
+            {
+              unsigned size = ALIGN (cif->rtype->size, sizeof (UINT64)) ;
+              memcpy (get_x_addr (context, 0), rvalue, size);
+            }
+          else
+            {
+              FFI_ASSERT (0);
+            }
+          break;
+        default:
+          FFI_ASSERT (0);
+          break;
+        }
+    }
+  else
+    {
+      memcpy (&rvalue, get_x_addr (context, 8), sizeof (UINT64));
+      (closure->fun) (cif, rvalue, avalue, closure->user_data);
+    }
+}
+
diff --git a/src/aarch64/ffitarget.h b/src/aarch64/ffitarget.h
new file mode 100644
index 0000000..6f1a348
--- /dev/null
+++ b/src/aarch64/ffitarget.h
@@ -0,0 +1,59 @@
+/* Copyright (c) 2009, 2010, 2011, 2012 ARM Ltd.
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+``Software''), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+#ifndef LIBFFI_TARGET_H
+#define LIBFFI_TARGET_H
+
+#ifndef LIBFFI_H
+#error "Please do not include ffitarget.h directly into your source.  Use ffi.h instead."
+#endif
+
+#ifndef LIBFFI_ASM
+typedef unsigned long ffi_arg;
+typedef signed long ffi_sarg;
+
+typedef enum ffi_abi
+  {
+    FFI_FIRST_ABI = 0,
+    FFI_SYSV,
+    FFI_LAST_ABI,
+    FFI_DEFAULT_ABI = FFI_SYSV
+  } ffi_abi;
+#endif
+
+/* ---- Definitions for closures ----------------------------------------- */
+
+#define FFI_CLOSURES 1
+#define FFI_TRAMPOLINE_SIZE 36
+#define FFI_NATIVE_RAW_API 0
+
+/* ---- Internal ---- */
+
+
+#define FFI_EXTRA_CIF_FIELDS unsigned aarch64_flags
+
+#define AARCH64_FFI_WITH_V_BIT 0
+
+#define AARCH64_N_XREG 32
+#define AARCH64_N_VREG 32
+#define AARCH64_CALL_CONTEXT_SIZE (AARCH64_N_XREG * 8 + AARCH64_N_VREG * 16)
+
+#endif
diff --git a/src/aarch64/sysv.S b/src/aarch64/sysv.S
new file mode 100644
index 0000000..b8cd421
--- /dev/null
+++ b/src/aarch64/sysv.S
@@ -0,0 +1,307 @@
+/* Copyright (c) 2009, 2010, 2011, 2012 ARM Ltd.
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+``Software''), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+#define LIBFFI_ASM
+#include <fficonfig.h>
+#include <ffi.h>
+
+#define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#define cfi_restore(reg)		.cfi_restore reg
+#define cfi_def_cfa_register(reg)	.cfi_def_cfa_register reg
+
+        .text
+        .globl ffi_call_SYSV
+        .type ffi_call_SYSV, #function
+
+/* ffi_call_SYSV()
+
+   Create a stack frame, setup an argument context, call the callee
+   and extract the result.
+
+   The maximum required argument stack size is provided,
+   ffi_call_SYSV() allocates that stack space then calls the
+   prepare_fn to populate register context and stack.  The
+   argument passing registers are loaded from the register
+   context and the callee called, on return the register passing
+   register are saved back to the context.  Our caller will
+   extract the return value from the final state of the saved
+   register context.
+
+   Prototype:
+
+   extern unsigned
+   ffi_call_SYSV (void (*)(struct call_context *context, unsigned char *,
+			   extended_cif *),
+                  struct call_context *context,
+                  extended_cif *,
+                  unsigned required_stack_size,
+                  void (*fn)(void));
+
+   Therefore on entry we have:
+
+   x0 prepare_fn
+   x1 &context
+   x2 &ecif
+   x3 bytes
+   x4 fn
+
+   This function uses the following stack frame layout:
+
+   ==
+                saved x30(lr)
+   x29(fp)->    saved x29(fp)
+                saved x24
+                saved x23
+                saved x22
+   sp'    ->    saved x21
+                ...
+   sp     ->    (constructed callee stack arguments)
+   ==
+
+   Voila! */
+
+#define ffi_call_SYSV_FS (8 * 4)
+
+        .cfi_startproc
+ffi_call_SYSV:
+        stp     x29, x30, [sp, #-16]!
+	cfi_adjust_cfa_offset (16)
+        cfi_rel_offset (x29, 0)
+        cfi_rel_offset (x30, 8)
+
+        mov     x29, sp
+	cfi_def_cfa_register (x29)
+        sub     sp, sp, #ffi_call_SYSV_FS
+
+        stp     x21, x22, [sp, 0]
+        cfi_rel_offset (x21, 0 - ffi_call_SYSV_FS)
+        cfi_rel_offset (x22, 8 - ffi_call_SYSV_FS)
+
+        stp     x23, x24, [sp, 16]
+        cfi_rel_offset (x23, 16 - ffi_call_SYSV_FS)
+        cfi_rel_offset (x24, 24 - ffi_call_SYSV_FS)
+
+        mov     x21, x1
+        mov     x22, x2
+        mov     x24, x4
+
+        /* Allocate the stack space for the actual arguments, many
+           arguments will be passed in registers, but we assume
+           worst case and allocate sufficient stack for ALL of
+           the arguments.  */
+        sub     sp, sp, x3
+
+        /* unsigned (*prepare_fn) (struct call_context *context,
+				   unsigned char *stack, extended_cif *ecif);
+	 */
+        mov     x23, x0
+        mov     x0, x1
+        mov     x1, sp
+        /* x2 already in place */
+        blr     x23
+
+        /* Preserve the flags returned.  */
+        mov     x23, x0
+
+        /* Figure out if we should touch the vector registers.  */
+        tbz     x23, #AARCH64_FFI_WITH_V_BIT, 1f
+
+        /* Load the vector argument passing registers.  */
+        ldp     q0, q1, [x21, #8*32 +  0]
+        ldp     q2, q3, [x21, #8*32 + 32]
+        ldp     q4, q5, [x21, #8*32 + 64]
+        ldp     q6, q7, [x21, #8*32 + 96]
+1:
+        /* Load the core argument passing registers.  */
+        ldp     x0, x1, [x21,  #0]
+        ldp     x2, x3, [x21, #16]
+        ldp     x4, x5, [x21, #32]
+        ldp     x6, x7, [x21, #48]
+
+        /* Don't forget x8 which may be holding the address of a return buffer.
+	 */
+        ldr     x8,     [x21, #8*8]
+
+        blr     x24
+
+        /* Save the core argument passing registers.  */
+        stp     x0, x1, [x21,  #0]
+        stp     x2, x3, [x21, #16]
+        stp     x4, x5, [x21, #32]
+        stp     x6, x7, [x21, #48]
+
+        /* Note nothing useful ever comes back in x8!  */
+
+        /* Figure out if we should touch the vector registers.  */
+        tbz     x23, #AARCH64_FFI_WITH_V_BIT, 1f
+
+        /* Save the vector argument passing registers.  */
+        stp     q0, q1, [x21, #8*32 + 0]
+        stp     q2, q3, [x21, #8*32 + 32]
+        stp     q4, q5, [x21, #8*32 + 64]
+        stp     q6, q7, [x21, #8*32 + 96]
+1:
+        /* All done, unwind our stack frame.  */
+        ldp     x21, x22, [x29,  # - ffi_call_SYSV_FS]
+        cfi_restore (x21)
+        cfi_restore (x22)
+
+        ldp     x23, x24, [x29,  # - ffi_call_SYSV_FS + 16]
+        cfi_restore (x23)
+        cfi_restore (x24)
+
+        mov     sp, x29
+	cfi_def_cfa_register (sp)
+
+        ldp     x29, x30, [sp], #16
+	cfi_adjust_cfa_offset (-16)
+        cfi_restore (x29)
+        cfi_restore (x30)
+
+        ret
+
+        .cfi_endproc
+        .size ffi_call_SYSV, .-ffi_call_SYSV
+
+#define ffi_closure_SYSV_FS (8 * 2 + AARCH64_CALL_CONTEXT_SIZE)
+
+/* ffi_closure_SYSV
+
+   Closure invocation glue. This is the low level code invoked directly by
+   the closure trampoline to setup and call a closure.
+
+   On entry x17 points to a struct trampoline_data, x16 has been clobbered
+   all other registers are preserved.
+
+   We allocate a call context and save the argument passing registers,
+   then invoked the generic C ffi_closure_SYSV_inner() function to do all
+   the real work, on return we load the result passing registers back from
+   the call context.
+
+   On entry
+
+   extern void
+   ffi_closure_SYSV (struct trampoline_data *);
+
+   struct trampoline_data
+   {
+        UINT64 *ffi_closure;
+        UINT64 flags;
+   };
+
+   This function uses the following stack frame layout:
+
+   ==
+                saved x30(lr)
+   x29(fp)->    saved x29(fp)
+                saved x22
+                saved x21
+                ...
+   sp     ->    call_context
+   ==
+
+   Voila!  */
+
+        .text
+        .globl ffi_closure_SYSV
+        .cfi_startproc
+ffi_closure_SYSV:
+        stp     x29, x30, [sp, #-16]!
+	cfi_adjust_cfa_offset (16)
+        cfi_rel_offset (x29, 0)
+        cfi_rel_offset (x30, 8)
+
+        mov     x29, sp
+
+        sub     sp, sp, #ffi_closure_SYSV_FS
+	cfi_adjust_cfa_offset (ffi_closure_SYSV_FS)
+
+        stp     x21, x22, [x29, #-16]
+        cfi_rel_offset (x21, 0)
+        cfi_rel_offset (x22, 8)
+
+        /* Load x21 with &call_context.  */
+        mov     x21, sp
+        /* Preserve our struct trampoline_data *  */
+        mov     x22, x17
+
+        /* Save the rest of the argument passing registers.  */
+        stp     x0, x1, [x21, #0]
+        stp     x2, x3, [x21, #16]
+        stp     x4, x5, [x21, #32]
+        stp     x6, x7, [x21, #48]
+        /* Don't forget we may have been given a result scratch pad address.
+	 */
+        str     x8,     [x21, #64]
+
+        /* Figure out if we should touch the vector registers.  */
+        ldr     x0, [x22, #8]
+        tbz     x0, #AARCH64_FFI_WITH_V_BIT, 1f
+
+        /* Save the argument passing vector registers.  */
+        stp     q0, q1, [x21, #8*32 + 0]
+        stp     q2, q3, [x21, #8*32 + 32]
+        stp     q4, q5, [x21, #8*32 + 64]
+        stp     q6, q7, [x21, #8*32 + 96]
+1:
+        /* Load &ffi_closure..  */
+        ldr     x0, [x22, #0]
+        mov     x1, x21
+        /* Compute the location of the stack at the point that the
+           trampoline was called.  */
+        add     x2, x29, #16
+
+        bl      ffi_closure_SYSV_inner
+
+        /* Figure out if we should touch the vector registers.  */
+        ldr     x0, [x22, #8]
+        tbz     x0, #AARCH64_FFI_WITH_V_BIT, 1f
+
+        /* Load the result passing vector registers.  */
+        ldp     q0, q1, [x21, #8*32 + 0]
+        ldp     q2, q3, [x21, #8*32 + 32]
+        ldp     q4, q5, [x21, #8*32 + 64]
+        ldp     q6, q7, [x21, #8*32 + 96]
+1:
+        /* Load the result passing core registers.  */
+        ldp     x0, x1, [x21,  #0]
+        ldp     x2, x3, [x21, #16]
+        ldp     x4, x5, [x21, #32]
+        ldp     x6, x7, [x21, #48]
+        /* Note nothing usefull is returned in x8.  */
+
+        /* We are done, unwind our frame.  */
+        ldp     x21, x22, [x29,  #-16]
+        cfi_restore (x21)
+        cfi_restore (x22)
+
+        mov     sp, x29
+	cfi_adjust_cfa_offset (-ffi_closure_SYSV_FS)
+
+        ldp     x29, x30, [sp], #16
+	cfi_adjust_cfa_offset (-16)
+        cfi_restore (x29)
+        cfi_restore (x30)
+
+        ret
+        .cfi_endproc
+        .size ffi_closure_SYSV, .-ffi_closure_SYSV

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #3: aarch64-libffi-testsuite.diff --]
[-- Type: text/x-patch; name=aarch64-libffi-testsuite.diff, Size: 26429 bytes --]

diff --git a/testsuite/lib/libffi.exp b/testsuite/lib/libffi.exp
index 4a65ed1..8ee3f15 100644
--- a/testsuite/lib/libffi.exp
+++ b/testsuite/lib/libffi.exp
@@ -203,6 +203,10 @@ proc libffi_target_compile { source dest type options } {
 
     lappend options "libs= -lffi"
 
+    if { [string match "aarch64*-*-linux*" $target_triplet] } {
+	lappend options "libs= -lpthread"
+    }
+
     verbose "options: $options"
     return [target_compile $source $dest $type $options]
 }
diff --git a/testsuite/libffi.call/cls_struct_va1.c b/testsuite/libffi.call/cls_struct_va1.c
new file mode 100644
index 0000000..91772bd
--- /dev/null
+++ b/testsuite/libffi.call/cls_struct_va1.c
@@ -0,0 +1,114 @@
+/* Area:		ffi_call, closure_call
+   Purpose:		Test doubles passed in variable argument lists.
+   Limitations:	none.
+   PR:			none.
+   Originator:	Blake Chaffin 6/6/2007	 */
+
+/* { dg-do run } */
+/* { dg-output "" { xfail avr32*-*-* } } */
+#include "ffitest.h"
+
+struct small_tag
+{
+  unsigned char a;
+  unsigned char b;
+};
+
+struct large_tag
+{
+  unsigned a;
+  unsigned b;
+  unsigned c;
+  unsigned d;
+  unsigned e;
+};
+
+static void
+test_fn (ffi_cif* cif __UNUSED__, void* resp,
+	 void** args, void* userdata __UNUSED__)
+{
+  int n = *(int*)args[0];
+  struct small_tag s1 = * (struct small_tag *) args[1];
+  struct large_tag l1 = * (struct large_tag *) args[2];
+  struct small_tag s2 = * (struct small_tag *) args[3];
+
+  printf ("%d %d %d %d %d %d %d %d %d %d\n", n, s1.a, s1.b,
+	  l1.a, l1.b, l1.c, l1.d, l1.e,
+	  s2.a, s2.b);
+  * (int*) resp = 42;
+}
+
+int
+main (void)
+{
+  ffi_cif cif;
+  void *code;
+  ffi_closure *pcl = ffi_closure_alloc (sizeof (ffi_closure), &code);
+  ffi_type* arg_types[5];
+
+  ffi_arg res = 0;
+
+  ffi_type s_type;
+  ffi_type *s_type_elements[3];
+
+  ffi_type l_type;
+  ffi_type *l_type_elements[6];
+
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l1;
+
+  int si;
+
+  s_type.size = 0;
+  s_type.alignment = 0;
+  s_type.type = FFI_TYPE_STRUCT;
+  s_type.elements = s_type_elements;
+
+  s_type_elements[0] = &ffi_type_uchar;
+  s_type_elements[1] = &ffi_type_uchar;
+  s_type_elements[2] = NULL;
+
+  l_type.size = 0;
+  l_type.alignment = 0;
+  l_type.type = FFI_TYPE_STRUCT;
+  l_type.elements = l_type_elements;
+
+  l_type_elements[0] = &ffi_type_uint;
+  l_type_elements[1] = &ffi_type_uint;
+  l_type_elements[2] = &ffi_type_uint;
+  l_type_elements[3] = &ffi_type_uint;
+  l_type_elements[4] = &ffi_type_uint;
+  l_type_elements[5] = NULL;
+
+  arg_types[0] = &ffi_type_sint;
+  arg_types[1] = &s_type;
+  arg_types[2] = &l_type;
+  arg_types[3] = &s_type;
+  arg_types[4] = NULL;
+
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 4, &ffi_type_sint,
+			 arg_types) == FFI_OK);
+
+  si = 4;
+  s1.a = 5;
+  s1.b = 6;
+
+  s2.a = 20;
+  s2.b = 21;
+
+  l1.a = 10;
+  l1.b = 11;
+  l1.c = 12;
+  l1.d = 13;
+  l1.e = 14;
+
+  CHECK(ffi_prep_closure_loc(pcl, &cif, test_fn, NULL, code) == FFI_OK);
+
+  res = ((int (*)(int, ...))(code))(si, s1, l1, s2);
+  // { dg-output "4 5 6 10 11 12 13 14 20 21" }
+  printf("res: %d\n", (int) res);
+  // { dg-output "\nres: 42" }
+
+  exit(0);
+}
diff --git a/testsuite/libffi.call/cls_uchar_va.c b/testsuite/libffi.call/cls_uchar_va.c
new file mode 100644
index 0000000..19cd4f3
--- /dev/null
+++ b/testsuite/libffi.call/cls_uchar_va.c
@@ -0,0 +1,44 @@
+/* Area:	closure_call
+   Purpose:	Test anonymous unsigned char argument.
+   Limitations:	none.
+   PR:		none.
+   Originator:	ARM Ltd. */
+
+/* { dg-do run } */
+#include "ffitest.h"
+
+typedef unsigned char T;
+
+static void cls_ret_T_fn(ffi_cif* cif __UNUSED__, void* resp, void** args,
+			 void* userdata __UNUSED__)
+ {
+   *(T *)resp = *(T *)args[0];
+
+   printf("%d: %d %d\n", *(T *)resp, *(T *)args[0], *(T *)args[1]);
+ }
+
+typedef T (*cls_ret_T)(T, ...);
+
+int main (void)
+{
+  ffi_cif cif;
+  void *code;
+  ffi_closure *pcl = ffi_closure_alloc(sizeof(ffi_closure), &code);
+  ffi_type * cl_arg_types[3];
+  T res;
+
+  cl_arg_types[0] = &ffi_type_uchar;
+  cl_arg_types[1] = &ffi_type_uchar;
+  cl_arg_types[2] = NULL;
+
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 2,
+			 &ffi_type_uchar, cl_arg_types) == FFI_OK);
+
+  CHECK(ffi_prep_closure_loc(pcl, &cif, cls_ret_T_fn, NULL, code)  == FFI_OK);
+  res = ((((cls_ret_T)code)(67, 4)));
+  /* { dg-output "67: 67 4" } */
+  printf("res: %d\n", res);
+  /* { dg-output "\nres: 67" } */
+  exit(0);
+}
diff --git a/testsuite/libffi.call/cls_uint_va.c b/testsuite/libffi.call/cls_uint_va.c
new file mode 100644
index 0000000..150fddd
--- /dev/null
+++ b/testsuite/libffi.call/cls_uint_va.c
@@ -0,0 +1,45 @@
+/* Area:	closure_call
+   Purpose:	Test anonymous unsigned int argument.
+   Limitations:	none.
+   PR:		none.
+   Originator:	ARM Ltd. */
+
+/* { dg-do run } */
+
+#include "ffitest.h"
+
+typedef unsigned int T;
+
+static void cls_ret_T_fn(ffi_cif* cif __UNUSED__, void* resp, void** args,
+			 void* userdata __UNUSED__)
+ {
+   *(T *)resp = *(T *)args[0];
+
+   printf("%d: %d %d\n", *(T *)resp, *(T *)args[0], *(T *)args[1]);
+ }
+
+typedef T (*cls_ret_T)(T, ...);
+
+int main (void)
+{
+  ffi_cif cif;
+  void *code;
+  ffi_closure *pcl = ffi_closure_alloc(sizeof(ffi_closure), &code);
+  ffi_type * cl_arg_types[3];
+  T res;
+
+  cl_arg_types[0] = &ffi_type_uint;
+  cl_arg_types[1] = &ffi_type_uint;
+  cl_arg_types[2] = NULL;
+
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 2,
+			 &ffi_type_uint, cl_arg_types) == FFI_OK);
+
+  CHECK(ffi_prep_closure_loc(pcl, &cif, cls_ret_T_fn, NULL, code)  == FFI_OK);
+  res = ((((cls_ret_T)code)(67, 4)));
+  /* { dg-output "67: 67 4" } */
+  printf("res: %d\n", res);
+  /* { dg-output "\nres: 67" } */
+  exit(0);
+}
diff --git a/testsuite/libffi.call/cls_ulong_va.c b/testsuite/libffi.call/cls_ulong_va.c
new file mode 100644
index 0000000..0315082
--- /dev/null
+++ b/testsuite/libffi.call/cls_ulong_va.c
@@ -0,0 +1,45 @@
+/* Area:	closure_call
+   Purpose:	Test anonymous unsigned long argument.
+   Limitations:	none.
+   PR:		none.
+   Originator:	ARM Ltd. */
+
+/* { dg-do run } */
+
+#include "ffitest.h"
+
+typedef unsigned long T;
+
+static void cls_ret_T_fn(ffi_cif* cif __UNUSED__, void* resp, void** args,
+			 void* userdata __UNUSED__)
+ {
+   *(T *)resp = *(T *)args[0];
+
+   printf("%ld: %ld %ld\n", *(T *)resp, *(T *)args[0], *(T *)args[1]);
+ }
+
+typedef T (*cls_ret_T)(T, ...);
+
+int main (void)
+{
+  ffi_cif cif;
+  void *code;
+  ffi_closure *pcl = ffi_closure_alloc(sizeof(ffi_closure), &code);
+  ffi_type * cl_arg_types[3];
+  T res;
+
+  cl_arg_types[0] = &ffi_type_ulong;
+  cl_arg_types[1] = &ffi_type_ulong;
+  cl_arg_types[2] = NULL;
+
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 2,
+			 &ffi_type_ulong, cl_arg_types) == FFI_OK);
+
+  CHECK(ffi_prep_closure_loc(pcl, &cif, cls_ret_T_fn, NULL, code)  == FFI_OK);
+  res = ((((cls_ret_T)code)(67, 4)));
+  /* { dg-output "67: 67 4" } */
+  printf("res: %ld\n", res);
+  /* { dg-output "\nres: 67" } */
+  exit(0);
+}
diff --git a/testsuite/libffi.call/cls_ushort_va.c b/testsuite/libffi.call/cls_ushort_va.c
new file mode 100644
index 0000000..b2b5a3b
--- /dev/null
+++ b/testsuite/libffi.call/cls_ushort_va.c
@@ -0,0 +1,44 @@
+/* Area:	closure_call
+   Purpose:	Test anonymous unsigned short argument.
+   Limitations:	none.
+   PR:		none.
+   Originator:	ARM Ltd. */
+
+/* { dg-do run } */
+#include "ffitest.h"
+
+typedef unsigned short T;
+
+static void cls_ret_T_fn(ffi_cif* cif __UNUSED__, void* resp, void** args,
+			 void* userdata __UNUSED__)
+ {
+   *(T *)resp = *(T *)args[0];
+
+   printf("%d: %d %d\n", *(T *)resp, *(T *)args[0], *(T *)args[1]);
+ }
+
+typedef T (*cls_ret_T)(T, ...);
+
+int main (void)
+{
+  ffi_cif cif;
+  void *code;
+  ffi_closure *pcl = ffi_closure_alloc(sizeof(ffi_closure), &code);
+  ffi_type * cl_arg_types[3];
+  T res;
+
+  cl_arg_types[0] = &ffi_type_ushort;
+  cl_arg_types[1] = &ffi_type_ushort;
+  cl_arg_types[2] = NULL;
+
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 2,
+			 &ffi_type_ushort, cl_arg_types) == FFI_OK);
+
+  CHECK(ffi_prep_closure_loc(pcl, &cif, cls_ret_T_fn, NULL, code)  == FFI_OK);
+  res = ((((cls_ret_T)code)(67, 4)));
+  /* { dg-output "67: 67 4" } */
+  printf("res: %d\n", res);
+  /* { dg-output "\nres: 67" } */
+  exit(0);
+}
diff --git a/testsuite/libffi.call/nested_struct11.c b/testsuite/libffi.call/nested_struct11.c
new file mode 100644
index 0000000..fce6948
--- /dev/null
+++ b/testsuite/libffi.call/nested_struct11.c
@@ -0,0 +1,121 @@
+/* Area:	ffi_call, closure_call
+   Purpose:	Check parameter passing with nested structs
+		of a single type.  This tests the special cases
+		for homogenous floating-point aggregates in the
+		AArch64 PCS.
+   Limitations:	none.
+   PR:		none.
+   Originator:  ARM Ltd.  */
+
+/* { dg-do run } */
+#include "ffitest.h"
+
+typedef struct A {
+  float a_x;
+  float a_y;
+} A;
+
+typedef struct B {
+  float b_x;
+  float b_y;
+} B;
+
+typedef struct C {
+  A a;
+  B b;
+} C;
+
+static C C_fn (int x, int y, int z, C source, int i, int j, int k)
+{
+  C result;
+  result.a.a_x = source.a.a_x;
+  result.a.a_y = source.a.a_y;
+  result.b.b_x = source.b.b_x;
+  result.b.b_y = source.b.b_y;
+
+  printf ("%d, %d, %d, %d, %d, %d\n", x, y, z, i, j, k);
+
+  printf ("%.1f, %.1f, %.1f, %.1f, "
+	  "%.1f, %.1f, %.1f, %.1f\n",
+	  source.a.a_x, source.a.a_y,
+	  source.b.b_x, source.b.b_y,
+	  result.a.a_x, result.a.a_y,
+	  result.b.b_x, result.b.b_y);
+
+  return result;
+}
+
+int main (void)
+{
+  ffi_cif cif;
+
+  ffi_type* struct_fields_source_a[3];
+  ffi_type* struct_fields_source_b[3];
+  ffi_type* struct_fields_source_c[3];
+  ffi_type* arg_types[8];
+
+  ffi_type struct_type_a, struct_type_b, struct_type_c;
+
+  struct A source_fld_a = {1.0, 2.0};
+  struct B source_fld_b = {4.0, 8.0};
+  int k = 1;
+
+  struct C result;
+  struct C source = {source_fld_a, source_fld_b};
+
+  struct_type_a.size = 0;
+  struct_type_a.alignment = 0;
+  struct_type_a.type = FFI_TYPE_STRUCT;
+  struct_type_a.elements = struct_fields_source_a;
+
+  struct_type_b.size = 0;
+  struct_type_b.alignment = 0;
+  struct_type_b.type = FFI_TYPE_STRUCT;
+  struct_type_b.elements = struct_fields_source_b;
+
+  struct_type_c.size = 0;
+  struct_type_c.alignment = 0;
+  struct_type_c.type = FFI_TYPE_STRUCT;
+  struct_type_c.elements = struct_fields_source_c;
+
+  struct_fields_source_a[0] = &ffi_type_float;
+  struct_fields_source_a[1] = &ffi_type_float;
+  struct_fields_source_a[2] = NULL;
+
+  struct_fields_source_b[0] = &ffi_type_float;
+  struct_fields_source_b[1] = &ffi_type_float;
+  struct_fields_source_b[2] = NULL;
+
+  struct_fields_source_c[0] = &struct_type_a;
+  struct_fields_source_c[1] = &struct_type_b;
+  struct_fields_source_c[2] = NULL;
+
+  arg_types[0] = &ffi_type_sint32;
+  arg_types[1] = &ffi_type_sint32;
+  arg_types[2] = &ffi_type_sint32;
+  arg_types[3] = &struct_type_c;
+  arg_types[4] = &ffi_type_sint32;
+  arg_types[5] = &ffi_type_sint32;
+  arg_types[6] = &ffi_type_sint32;
+  arg_types[7] = NULL;
+
+  void *args[7];
+  args[0] = &k;
+  args[1] = &k;
+  args[2] = &k;
+  args[3] = &source;
+  args[4] = &k;
+  args[5] = &k;
+  args[6] = &k;
+  CHECK (ffi_prep_cif (&cif, FFI_DEFAULT_ABI, 7, &struct_type_c,
+		       arg_types) == FFI_OK);
+
+  ffi_call (&cif, FFI_FN (C_fn), &result, args);
+  /* { dg-output "1, 1, 1, 1, 1, 1\n" } */
+  /* { dg-output "1.0, 2.0, 4.0, 8.0, 1.0, 2.0, 4.0, 8.0" } */
+  CHECK (result.a.a_x == source.a.a_x);
+  CHECK (result.a.a_y == source.a.a_y);
+  CHECK (result.b.b_x == source.b.b_x);
+  CHECK (result.b.b_y == source.b.b_y);
+  exit (0);
+}
diff --git a/testsuite/libffi.call/uninitialized.c b/testsuite/libffi.call/uninitialized.c
new file mode 100644
index 0000000..f00d830
--- /dev/null
+++ b/testsuite/libffi.call/uninitialized.c
@@ -0,0 +1,61 @@
+/* { dg-do run } */
+#include "ffitest.h"
+
+typedef struct
+{
+  unsigned char uc;
+  double d;
+  unsigned int ui;
+} test_structure_1;
+
+static test_structure_1 struct1(test_structure_1 ts)
+{
+  ts.uc++;
+  ts.d--;
+  ts.ui++;
+
+  return ts;
+}
+
+int main (void)
+{
+  ffi_cif cif;
+  ffi_type *args[MAX_ARGS];
+  void *values[MAX_ARGS];
+  ffi_type ts1_type;
+  ffi_type *ts1_type_elements[4];
+
+  memset(&cif, 1, sizeof(cif));
+  ts1_type.size = 0;
+  ts1_type.alignment = 0;
+  ts1_type.type = FFI_TYPE_STRUCT;
+  ts1_type.elements = ts1_type_elements;
+  ts1_type_elements[0] = &ffi_type_uchar;
+  ts1_type_elements[1] = &ffi_type_double;
+  ts1_type_elements[2] = &ffi_type_uint;
+  ts1_type_elements[3] = NULL;
+
+  test_structure_1 ts1_arg;
+  /* This is a hack to get a properly aligned result buffer */
+  test_structure_1 *ts1_result =
+    (test_structure_1 *) malloc (sizeof(test_structure_1));
+
+  args[0] = &ts1_type;
+  values[0] = &ts1_arg;
+
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif(&cif, FFI_DEFAULT_ABI, 1,
+		     &ts1_type, args) == FFI_OK);
+
+  ts1_arg.uc = '\x01';
+  ts1_arg.d = 3.14159;
+  ts1_arg.ui = 555;
+
+  ffi_call(&cif, FFI_FN(struct1), ts1_result, values);
+
+  CHECK(ts1_result->ui == 556);
+  CHECK(ts1_result->d == 3.14159 - 1);
+
+  free (ts1_result);
+  exit(0);
+}
diff --git a/testsuite/libffi.call/va_1.c b/testsuite/libffi.call/va_1.c
new file mode 100644
index 0000000..5c7cce9
--- /dev/null
+++ b/testsuite/libffi.call/va_1.c
@@ -0,0 +1,196 @@
+/* Area:		ffi_call
+   Purpose:		Test passing struct in variable argument lists.
+   Limitations:	none.
+   PR:			none.
+   Originator:	        ARM Ltd. */
+
+/* { dg-do run } */
+/* { dg-output "" { xfail avr32*-*-* x86_64-*-*-* } } */
+
+#include "ffitest.h"
+#include <stdarg.h>
+
+struct small_tag
+{
+  unsigned char a;
+  unsigned char b;
+};
+
+struct large_tag
+{
+  unsigned a;
+  unsigned b;
+  unsigned c;
+  unsigned d;
+  unsigned e;
+};
+
+static int
+test_fn (int n, ...)
+{
+  va_list ap;
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l;
+  unsigned char uc;
+  signed char sc;
+  unsigned short us;
+  signed short ss;
+  unsigned int ui;
+  signed int si;
+  unsigned long ul;
+  signed long sl;
+  float f;
+  double d;
+
+  va_start (ap, n);
+  s1 = va_arg (ap, struct small_tag);
+  l = va_arg (ap, struct large_tag);
+  s2 = va_arg (ap, struct small_tag);
+
+  uc = va_arg (ap, unsigned);
+  sc = va_arg (ap, signed);
+
+  us = va_arg (ap, unsigned);
+  ss = va_arg (ap, signed);
+
+  ui = va_arg (ap, unsigned int);
+  si = va_arg (ap, signed int);
+
+  ul = va_arg (ap, unsigned long);
+  sl = va_arg (ap, signed long);
+
+  f = va_arg (ap, double);	/* C standard promotes float->double
+				   when anonymous */
+  d = va_arg (ap, double);
+
+  printf ("%u %u %u %u %u %u %u %u %u uc=%u sc=%d %u %d %u %d %lu %ld %f %f\n",
+	  s1.a, s1.b, l.a, l.b, l.c, l.d, l.e,
+	  s2.a, s2.b,
+	  uc, sc,
+	  us, ss,
+	  ui, si,
+	  ul, sl,
+	  f, d);
+  va_end (ap);
+  return n + 1;
+}
+
+int
+main (void)
+{
+  ffi_cif cif;
+  void* args[15];
+  ffi_type* arg_types[15];
+
+  ffi_type s_type;
+  ffi_type *s_type_elements[3];
+
+  ffi_type l_type;
+  ffi_type *l_type_elements[6];
+
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l1;
+
+  int n;
+  int res;
+
+  unsigned char uc;
+  signed char sc;
+  unsigned short us;
+  signed short ss;
+  unsigned int ui;
+  signed int si;
+  unsigned long ul;
+  signed long sl;
+  double d1;
+  double f1;
+
+  s_type.size = 0;
+  s_type.alignment = 0;
+  s_type.type = FFI_TYPE_STRUCT;
+  s_type.elements = s_type_elements;
+
+  s_type_elements[0] = &ffi_type_uchar;
+  s_type_elements[1] = &ffi_type_uchar;
+  s_type_elements[2] = NULL;
+
+  l_type.size = 0;
+  l_type.alignment = 0;
+  l_type.type = FFI_TYPE_STRUCT;
+  l_type.elements = l_type_elements;
+
+  l_type_elements[0] = &ffi_type_uint;
+  l_type_elements[1] = &ffi_type_uint;
+  l_type_elements[2] = &ffi_type_uint;
+  l_type_elements[3] = &ffi_type_uint;
+  l_type_elements[4] = &ffi_type_uint;
+  l_type_elements[5] = NULL;
+
+  arg_types[0] = &ffi_type_sint;
+  arg_types[1] = &s_type;
+  arg_types[2] = &l_type;
+  arg_types[3] = &s_type;
+  arg_types[4] = &ffi_type_uint;
+  arg_types[5] = &ffi_type_sint;
+  arg_types[6] = &ffi_type_uint;
+  arg_types[7] = &ffi_type_sint;
+  arg_types[8] = &ffi_type_uint;
+  arg_types[9] = &ffi_type_sint;
+  arg_types[10] = &ffi_type_ulong;
+  arg_types[11] = &ffi_type_slong;
+  arg_types[12] = &ffi_type_double;
+  arg_types[13] = &ffi_type_double;
+  arg_types[14] = NULL;
+
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 14, &ffi_type_sint, arg_types) == FFI_OK);
+
+  s1.a = 5;
+  s1.b = 6;
+
+  l1.a = 10;
+  l1.b = 11;
+  l1.c = 12;
+  l1.d = 13;
+  l1.e = 14;
+
+  s2.a = 7;
+  s2.b = 8;
+
+  n = 41;
+
+  uc = 9;
+  sc = 10;
+  us = 11;
+  ss = 12;
+  ui = 13;
+  si = 14;
+  ul = 15;
+  sl = 16;
+  f1 = 2.12;
+  d1 = 3.13;
+
+  args[0] = &n;
+  args[1] = &s1;
+  args[2] = &l1;
+  args[3] = &s2;
+  args[4] = &uc;
+  args[5] = &sc;
+  args[6] = &us;
+  args[7] = &ss;
+  args[8] = &ui;
+  args[9] = &si;
+  args[10] = &ul;
+  args[11] = &sl;
+  args[12] = &f1;
+  args[13] = &d1;
+  args[14] = NULL;
+
+  ffi_call(&cif, FFI_FN(test_fn), &res, args);
+  /* { dg-output "5 6 10 11 12 13 14 7 8 uc=9 sc=10 11 12 13 14 15 16 2.120000 3.130000" } */
+  printf("res: %d\n", (int) res);
+  /* { dg-output "\nres: 42" } */
+
+  return 0;
+}
diff --git a/testsuite/libffi.call/va_struct1.c b/testsuite/libffi.call/va_struct1.c
new file mode 100644
index 0000000..11d1f10
--- /dev/null
+++ b/testsuite/libffi.call/va_struct1.c
@@ -0,0 +1,121 @@
+/* Area:		ffi_call
+   Purpose:		Test passing struct in variable argument lists.
+   Limitations:	none.
+   PR:			none.
+   Originator: ARM Ltd. */
+
+/* { dg-do run } */
+/* { dg-output "" { xfail avr32*-*-* } } */
+
+#include "ffitest.h"
+#include <stdarg.h>
+
+struct small_tag
+{
+  unsigned char a;
+  unsigned char b;
+};
+
+struct large_tag
+{
+  unsigned a;
+  unsigned b;
+  unsigned c;
+  unsigned d;
+  unsigned e;
+};
+
+static int
+test_fn (int n, ...)
+{
+  va_list ap;
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l;
+
+  va_start (ap, n);
+  s1 = va_arg (ap, struct small_tag);
+  l = va_arg (ap, struct large_tag);
+  s2 = va_arg (ap, struct small_tag);
+  printf ("%u %u %u %u %u %u %u %u %u\n", s1.a, s1.b, l.a, l.b, l.c, l.d, l.e,
+	  s2.a, s2.b);
+  va_end (ap);
+  return n + 1;
+}
+
+int
+main (void)
+{
+  ffi_cif cif;
+  void* args[5];
+  ffi_type* arg_types[5];
+
+  ffi_type s_type;
+  ffi_type *s_type_elements[3];
+
+  ffi_type l_type;
+  ffi_type *l_type_elements[6];
+
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l1;
+
+  int n;
+  int res;
+
+  s_type.size = 0;
+  s_type.alignment = 0;
+  s_type.type = FFI_TYPE_STRUCT;
+  s_type.elements = s_type_elements;
+
+  s_type_elements[0] = &ffi_type_uchar;
+  s_type_elements[1] = &ffi_type_uchar;
+  s_type_elements[2] = NULL;
+
+  l_type.size = 0;
+  l_type.alignment = 0;
+  l_type.type = FFI_TYPE_STRUCT;
+  l_type.elements = l_type_elements;
+
+  l_type_elements[0] = &ffi_type_uint;
+  l_type_elements[1] = &ffi_type_uint;
+  l_type_elements[2] = &ffi_type_uint;
+  l_type_elements[3] = &ffi_type_uint;
+  l_type_elements[4] = &ffi_type_uint;
+  l_type_elements[5] = NULL;
+
+  arg_types[0] = &ffi_type_sint;
+  arg_types[1] = &s_type;
+  arg_types[2] = &l_type;
+  arg_types[3] = &s_type;
+  arg_types[4] = NULL;
+
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 4, &ffi_type_sint, arg_types) == FFI_OK);
+
+  s1.a = 5;
+  s1.b = 6;
+
+  l1.a = 10;
+  l1.b = 11;
+  l1.c = 12;
+  l1.d = 13;
+  l1.e = 14;
+
+  s2.a = 7;
+  s2.b = 8;
+
+  n = 41;
+
+  args[0] = &n;
+  args[1] = &s1;
+  args[2] = &l1;
+  args[3] = &s2;
+  args[4] = NULL;
+
+  ffi_call(&cif, FFI_FN(test_fn), &res, args);
+  /* { dg-output "5 6 10 11 12 13 14 7 8" } */
+  printf("res: %d\n", (int) res);
+  /* { dg-output "\nres: 42" } */
+
+  return 0;
+}
diff --git a/testsuite/libffi.call/va_struct2.c b/testsuite/libffi.call/va_struct2.c
new file mode 100644
index 0000000..56f5b9c
--- /dev/null
+++ b/testsuite/libffi.call/va_struct2.c
@@ -0,0 +1,123 @@
+/* Area:		ffi_call
+   Purpose:		Test passing struct in variable argument lists.
+   Limitations:	none.
+   PR:			none.
+   Originator: ARM Ltd. */
+
+/* { dg-do run } */
+/* { dg-output "" { xfail avr32*-*-* } } */
+
+#include "ffitest.h"
+#include <stdarg.h>
+
+struct small_tag
+{
+  unsigned char a;
+  unsigned char b;
+};
+
+struct large_tag
+{
+  unsigned a;
+  unsigned b;
+  unsigned c;
+  unsigned d;
+  unsigned e;
+};
+
+static struct small_tag
+test_fn (int n, ...)
+{
+  va_list ap;
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l;
+
+  va_start (ap, n);
+  s1 = va_arg (ap, struct small_tag);
+  l = va_arg (ap, struct large_tag);
+  s2 = va_arg (ap, struct small_tag);
+  printf ("%u %u %u %u %u %u %u %u %u\n", s1.a, s1.b, l.a, l.b, l.c, l.d, l.e,
+	  s2.a, s2.b);
+  va_end (ap);
+  s1.a += s2.a;
+  s1.b += s2.b;
+  return s1;
+}
+
+int
+main (void)
+{
+  ffi_cif cif;
+  void* args[5];
+  ffi_type* arg_types[5];
+
+  ffi_type s_type;
+  ffi_type *s_type_elements[3];
+
+  ffi_type l_type;
+  ffi_type *l_type_elements[6];
+
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l1;
+
+  int n;
+  struct small_tag res;
+
+  s_type.size = 0;
+  s_type.alignment = 0;
+  s_type.type = FFI_TYPE_STRUCT;
+  s_type.elements = s_type_elements;
+
+  s_type_elements[0] = &ffi_type_uchar;
+  s_type_elements[1] = &ffi_type_uchar;
+  s_type_elements[2] = NULL;
+
+  l_type.size = 0;
+  l_type.alignment = 0;
+  l_type.type = FFI_TYPE_STRUCT;
+  l_type.elements = l_type_elements;
+
+  l_type_elements[0] = &ffi_type_uint;
+  l_type_elements[1] = &ffi_type_uint;
+  l_type_elements[2] = &ffi_type_uint;
+  l_type_elements[3] = &ffi_type_uint;
+  l_type_elements[4] = &ffi_type_uint;
+  l_type_elements[5] = NULL;
+
+  arg_types[0] = &ffi_type_sint;
+  arg_types[1] = &s_type;
+  arg_types[2] = &l_type;
+  arg_types[3] = &s_type;
+  arg_types[4] = NULL;
+
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 4, &s_type, arg_types) == FFI_OK);
+
+  s1.a = 5;
+  s1.b = 6;
+
+  l1.a = 10;
+  l1.b = 11;
+  l1.c = 12;
+  l1.d = 13;
+  l1.e = 14;
+
+  s2.a = 7;
+  s2.b = 8;
+
+  n = 41;
+
+  args[0] = &n;
+  args[1] = &s1;
+  args[2] = &l1;
+  args[3] = &s2;
+  args[4] = NULL;
+
+  ffi_call(&cif, FFI_FN(test_fn), &res, args);
+  /* { dg-output "5 6 10 11 12 13 14 7 8" } */
+  printf("res: %d %d\n", res.a, res.b);
+  /* { dg-output "\nres: 12 14" } */
+
+  return 0;
+}
diff --git a/testsuite/libffi.call/va_struct3.c b/testsuite/libffi.call/va_struct3.c
new file mode 100644
index 0000000..9a27e7f
--- /dev/null
+++ b/testsuite/libffi.call/va_struct3.c
@@ -0,0 +1,125 @@
+/* Area:		ffi_call
+   Purpose:		Test passing struct in variable argument lists.
+   Limitations:	none.
+   PR:			none.
+   Originator:	ARM Ltd. */
+
+/* { dg-do run } */
+/* { dg-output "" { xfail avr32*-*-* } } */
+
+#include "ffitest.h"
+#include <stdarg.h>
+
+struct small_tag
+{
+  unsigned char a;
+  unsigned char b;
+};
+
+struct large_tag
+{
+  unsigned a;
+  unsigned b;
+  unsigned c;
+  unsigned d;
+  unsigned e;
+};
+
+static struct large_tag
+test_fn (int n, ...)
+{
+  va_list ap;
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l;
+
+  va_start (ap, n);
+  s1 = va_arg (ap, struct small_tag);
+  l = va_arg (ap, struct large_tag);
+  s2 = va_arg (ap, struct small_tag);
+  printf ("%u %u %u %u %u %u %u %u %u\n", s1.a, s1.b, l.a, l.b, l.c, l.d, l.e,
+	  s2.a, s2.b);
+  va_end (ap);
+  l.a += s1.a;
+  l.b += s1.b;
+  l.c += s2.a;
+  l.d += s2.b;
+  return l;
+}
+
+int
+main (void)
+{
+  ffi_cif cif;
+  void* args[5];
+  ffi_type* arg_types[5];
+
+  ffi_type s_type;
+  ffi_type *s_type_elements[3];
+
+  ffi_type l_type;
+  ffi_type *l_type_elements[6];
+
+  struct small_tag s1;
+  struct small_tag s2;
+  struct large_tag l1;
+
+  int n;
+  struct large_tag res;
+
+  s_type.size = 0;
+  s_type.alignment = 0;
+  s_type.type = FFI_TYPE_STRUCT;
+  s_type.elements = s_type_elements;
+
+  s_type_elements[0] = &ffi_type_uchar;
+  s_type_elements[1] = &ffi_type_uchar;
+  s_type_elements[2] = NULL;
+
+  l_type.size = 0;
+  l_type.alignment = 0;
+  l_type.type = FFI_TYPE_STRUCT;
+  l_type.elements = l_type_elements;
+
+  l_type_elements[0] = &ffi_type_uint;
+  l_type_elements[1] = &ffi_type_uint;
+  l_type_elements[2] = &ffi_type_uint;
+  l_type_elements[3] = &ffi_type_uint;
+  l_type_elements[4] = &ffi_type_uint;
+  l_type_elements[5] = NULL;
+
+  arg_types[0] = &ffi_type_sint;
+  arg_types[1] = &s_type;
+  arg_types[2] = &l_type;
+  arg_types[3] = &s_type;
+  arg_types[4] = NULL;
+
+  CHECK(ffi_prep_cif_var(&cif, FFI_DEFAULT_ABI, 1, 4, &l_type, arg_types) == FFI_OK);
+
+  s1.a = 5;
+  s1.b = 6;
+
+  l1.a = 10;
+  l1.b = 11;
+  l1.c = 12;
+  l1.d = 13;
+  l1.e = 14;
+
+  s2.a = 7;
+  s2.b = 8;
+
+  n = 41;
+
+  args[0] = &n;
+  args[1] = &s1;
+  args[2] = &l1;
+  args[3] = &s2;
+  args[4] = NULL;
+
+  ffi_call(&cif, FFI_FN(test_fn), &res, args);
+  /* { dg-output "5 6 10 11 12 13 14 7 8" } */
+  printf("res: %d %d %d %d %d\n", res.a, res.b, res.c, res.d, res.e);
+  /* { dg-output "\nres: 15 17 19 21 14" } */
+
+  return 0;
+}

Re: [PATCH] New port for ARM AArch64

[Marcus Shawcroft]

On 28/09/12 17:28, Marcus Shawcroft wrote:
>
> ARM would like to contribute a libffi port for the ARM AArch64
> architecture.  The port passes the test suite cleanly.  The proposed
> ChangeLog and patches are included below.
>
>
> /Marcus

Ping!

Thanks
/Marcus

Re: [PATCH] New port for ARM AArch64

[Andrew Pinski]

On Wed, Oct 17, 2012 at 9:37 AM, Marcus Shawcroft
<marcus.shawcroft@arm.com> wrote:
> On 28/09/12 17:28, Marcus Shawcroft wrote:
>>
>>
>> ARM would like to contribute a libffi port for the ARM AArch64
>> architecture.  The port passes the test suite cleanly.  The proposed
>> ChangeLog and patches are included below.
>>
>>
>> /Marcus
>
>
> Ping!

Normally Anthony Green handles approvals here.  If he does not
respond, then you can ask for approval on the GCC side from the target
maintainer (which you are anyways) and Anthony will then merge it into
the libffi repo.

Thanks,
Andrew Pinski

Re: [PATCH] New port for ARM AArch64

[Anthony Green]

I've slowly been merging ports into libffi (Tile-GX/TilePRO last
week).  AArch64 would be next.  Let me have a quick look at it
tonight.

Thanks for your patience!

AG

On Wed, Oct 17, 2012 at 1:36 PM, Andrew Pinski <pinskia@gcc.gnu.org> wrote:
> On Wed, Oct 17, 2012 at 9:37 AM, Marcus Shawcroft
> <marcus.shawcroft@arm.com> wrote:
>> On 28/09/12 17:28, Marcus Shawcroft wrote:
>>>
>>>
>>> ARM would like to contribute a libffi port for the ARM AArch64
>>> architecture.  The port passes the test suite cleanly.  The proposed
>>> ChangeLog and patches are included below.
>>>
>>>
>>> /Marcus
>>
>>
>> Ping!
>
> Normally Anthony Green handles approvals here.  If he does not
> respond, then you can ask for approval on the GCC side from the target
> maintainer (which you are anyways) and Anthony will then merge it into
> the libffi repo.
>
> Thanks,
> Andrew Pinski

Re: [PATCH] New port for ARM AArch64

[Anthony Green]

Thanks for your patience, Marcus!  I think this patch is missing
configury bits.  Could you please send me a patch for just the missing
configure.ac changes.    The rest pf it has already been committed.

Thanks,

Anthony Green


On Fri, Sep 28, 2012 at 12:28 PM, Marcus Shawcroft
<marcus.shawcroft@arm.com> wrote:
>
> ARM would like to contribute a libffi port for the ARM AArch64
> architecture.  The port passes the test suite cleanly.  The proposed
> ChangeLog and patches are included below.
>
>
> /Marcus
>
>
> 2012-09-18  James Greenhalgh  <james.greenhalgh at arm.com>
>             Marcus Shawcroft  <marcus.shawcroft at arm.com>
>
>         * README: Add details of aarch64 port.
>         * src/aarch64/ffi.c: New.
>         * src/aarch64/ffitarget.h: Likewise.
>         * src/aarch64/sysv.S: Likewise.
>
> 2012-09-18  James Greenhalgh  <james.greenhalgh at arm.com>
>             Marcus Shawcroft  <marcus.shawcroft at arm.com>
>
>         * testsuite/lib/libffi.exp: Add support for aarch64.
>         * testsuite/libffi.call/cls_struct_va1.c: New.
>         * testsuite/libffi.call/cls_uchar_va.c: Likewise.
>         * testsuite/libffi.call/cls_uint_va.c: Likewise.
>         * testsuite/libffi.call/cls_ulong_va.c: Liekwise.
>         * testsuite/libffi.call/cls_ushort_va.c: Likewise.
>         * testsuite/libffi.call/nested_struct11.c: Likewise.
>         * testsuite/libffi.call/uninitialized.c: Likewise.
>         * testsuite/libffi.call/va_1.c: Likewise.
>         * testsuite/libffi.call/va_struct1.c: Likewise.
>         * testsuite/libffi.call/va_struct2.c: Likewise.
>         * testsuite/libffi.call/va_struct3.c: Likewise.

Re: [PATCH] New port for ARM AArch64

[Marcus Shawcroft]

[-- Attachment #1: Type: text/plain, Size: 487 bytes --]

On 30/10/12 11:10, Anthony Green wrote:
> Thanks for your patience, Marcus!  I think this patch is missing
> configury bits.  Could you please send me a patch for just the missing
> configure.ac changes.    The rest pf it has already been committed.


Anthony, Sorry about the delay. Patch attached includes the missing 
Makefile.am and configure.ac changes.

ChangeLog:

         * Makefile.am: Support aarch64.
         * configure.ac: Support aarch64.

Cheers
/Marcus

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: libffi-configury.diff --]
[-- Type: text/x-patch; name=libffi-configury.diff, Size: 1712 bytes --]

diff --git a/Makefile.am b/Makefile.am
index 16f32a6..bd4d5c4 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -5,6 +5,7 @@ AUTOMAKE_OPTIONS = foreign subdir-objects
 SUBDIRS = include testsuite man
 
 EXTRA_DIST = LICENSE ChangeLog.v1 ChangeLog.libgcj configure.host \
+	src/aarch64/ffi.c src/aarch64/ffitarget.h \
 	src/alpha/ffi.c src/alpha/osf.S src/alpha/ffitarget.h \
 	src/arm/ffi.c src/arm/sysv.S src/arm/ffitarget.h \
 	src/avr32/ffi.c src/avr32/sysv.S src/avr32/ffitarget.h \
@@ -151,6 +152,9 @@ endif
 if POWERPC_FREEBSD
 nodist_libffi_la_SOURCES += src/powerpc/ffi.c src/powerpc/sysv.S src/powerpc/ppc_closure.S
 endif
+if AARCH64
+nodist_libffi_la_SOURCES += src/aarch64/sysv.S src/aarch64/ffi.c
+endif
 if ARM
 nodist_libffi_la_SOURCES += src/arm/sysv.S src/arm/ffi.c
 if FFI_EXEC_TRAMPOLINE_TABLE
diff --git a/configure.ac b/configure.ac
index 9b946a2..8c3f40c 100644
--- a/configure.ac
+++ b/configure.ac
@@ -53,6 +53,10 @@ AM_CONDITIONAL(TESTSUBDIR, test -d $srcdir/testsuite)
 
 TARGETDIR="unknown"
 case "$host" in
+  aarch64*-*-*)
+	TARGET=AARCH64; TARGETDIR=aarch64
+	;;
+
   alpha*-*-*)
 	TARGET=ALPHA; TARGETDIR=alpha;
 	# Support 128-bit long double, changeable via command-line switch.
@@ -233,6 +237,7 @@ AM_CONDITIONAL(POWERPC, test x$TARGET = xPOWERPC)
 AM_CONDITIONAL(POWERPC_AIX, test x$TARGET = xPOWERPC_AIX)
 AM_CONDITIONAL(POWERPC_DARWIN, test x$TARGET = xPOWERPC_DARWIN)
 AM_CONDITIONAL(POWERPC_FREEBSD, test x$TARGET = xPOWERPC_FREEBSD)
+AM_CONDITIONAL(AARCH64, test x$TARGET = xAARCH64)
 AM_CONDITIONAL(ARM, test x$TARGET = xARM)
 AM_CONDITIONAL(AVR32, test x$TARGET = xAVR32)
 AM_CONDITIONAL(LIBFFI_CRIS, test x$TARGET = xLIBFFI_CRIS)

Re: [PATCH] New port for ARM AArch64

[Anthony Green]

Thanks.  I've checked this in.

AG


On Tue, Nov 6, 2012 at 10:32 AM, Marcus Shawcroft
<marcus.shawcroft@arm.com> wrote:
> On 30/10/12 11:10, Anthony Green wrote:
>>
>> Thanks for your patience, Marcus!  I think this patch is missing
>> configury bits.  Could you please send me a patch for just the missing
>> configure.ac changes.    The rest pf it has already been committed.
>
>
>
> Anthony, Sorry about the delay. Patch attached includes the missing
> Makefile.am and configure.ac changes.
>
> ChangeLog:
>
>         * Makefile.am: Support aarch64.
>         * configure.ac: Support aarch64.
>
> Cheers
> /Marcus