ffi 1.11.1

data/ext/ffi_c/libffi/src/sparc/v9.S

@@ -0,0 +1,440 @@
+/* -----------------------------------------------------------------------
+   v9.S - Copyright (c) 2000, 2003, 2004, 2008 Red Hat, Inc.
+   
+   SPARC 64-bit Foreign Function Interface 
+
+   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>
+#include "internal.h"
+
+#ifdef SPARC64
+
+#define C2(X, Y)  X ## Y
+#define C1(X, Y)  C2(X, Y)
+
+#ifdef __USER_LABEL_PREFIX__
+# define C(Y)	C1(__USER_LABEL_PREFIX__, Y)
+#else
+# define C(Y)	Y
+#endif
+#define L(Y)	C1(.L, Y)
+
+#if defined(__sun__) && defined(__svr4__)
+# define E(INDEX)	.align 16
+#else
+# define E(INDEX)	.align 16; .org 2b + INDEX * 16
+#endif
+
+#define STACK_BIAS 2047
+
+	.text
+        .align 8
+	.globl	C(ffi_call_v9)
+	.type	C(ffi_call_v9),#function
+	FFI_HIDDEN(C(ffi_call_v9))
+
+C(ffi_call_v9):
+.LUW0:
+	save	%sp, %o4, %sp
+.LUW1:
+	mov	%i0, %o0			! copy cif
+	add	%sp, STACK_BIAS+128+48, %o1	! load args area
+	mov	%i2, %o2			! copy rvalue
+	call	C(ffi_prep_args_v9)
+	 mov	%i3, %o3			! copy avalue
+
+	andcc	%o0, SPARC_FLAG_FP_ARGS, %g0	! need fp regs?
+	add	%sp, 48, %sp			! deallocate prep frame
+	be,pt	%xcc, 1f
+	 mov	%o0, %l0			! save flags
+
+	ldd	[%sp+STACK_BIAS+128], %f0	! load all fp arg regs
+	ldd	[%sp+STACK_BIAS+128+8], %f2
+	ldd	[%sp+STACK_BIAS+128+16], %f4
+	ldd	[%sp+STACK_BIAS+128+24], %f6
+	ldd	[%sp+STACK_BIAS+128+32], %f8
+	ldd	[%sp+STACK_BIAS+128+40], %f10
+	ldd	[%sp+STACK_BIAS+128+48], %f12
+	ldd	[%sp+STACK_BIAS+128+56], %f14
+	ldd	[%sp+STACK_BIAS+128+64], %f16
+	ldd	[%sp+STACK_BIAS+128+72], %f18
+	ldd	[%sp+STACK_BIAS+128+80], %f20
+	ldd	[%sp+STACK_BIAS+128+88], %f22
+	ldd	[%sp+STACK_BIAS+128+96], %f24
+	ldd	[%sp+STACK_BIAS+128+104], %f26
+	ldd	[%sp+STACK_BIAS+128+112], %f28
+	ldd	[%sp+STACK_BIAS+128+120], %f30
+
+1:	ldx	[%sp+STACK_BIAS+128], %o0	! load all int arg regs
+	ldx	[%sp+STACK_BIAS+128+8], %o1
+	ldx	[%sp+STACK_BIAS+128+16], %o2
+	ldx	[%sp+STACK_BIAS+128+24], %o3
+	ldx	[%sp+STACK_BIAS+128+32], %o4
+	ldx	[%sp+STACK_BIAS+128+40], %o5
+	call	%i1
+	 mov	%i5, %g5			! load static chain
+
+0:	call	1f		! load pc in %o7
+	 and	%l0, SPARC_FLAG_RET_MASK, %l1
+1:	sll	%l1, 4, %l1
+	add	%o7, %l1, %o7	! o7 = 0b + ret_type*16
+	jmp	%o7+(2f-0b)
+	 nop
+
+	.align	16
+2:
+E(SPARC_RET_VOID)
+	return	%i7+8
+	 nop
+E(SPARC_RET_STRUCT)
+	add	%sp, STACK_BIAS-64+128+48, %l2
+	sub	%sp, 64, %sp
+	b	8f
+	 stx	%o0, [%l2]
+E(SPARC_RET_UINT8)
+	and	%o0, 0xff, %i0
+	return	%i7+8
+	  stx	%o0, [%o2]
+E(SPARC_RET_SINT8)
+	sll	%o0, 24, %o0
+	sra	%o0, 24, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E(SPARC_RET_UINT16)
+	sll	%o0, 16, %o0
+	srl	%o0, 16, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E(SPARC_RET_SINT16)
+	sll	%o0, 16, %o0
+	sra	%o0, 16, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E(SPARC_RET_UINT32)
+	srl	%o0, 0, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E(SP_V9_RET_SINT32)
+	sra	%o0, 0, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E(SPARC_RET_INT64)
+	stx	%o0, [%i2]
+	return	%i7+8
+	 nop
+E(SPARC_RET_INT128)
+	stx	%o0, [%i2]
+	stx	%o1, [%i2+8]
+	return	%i7+8
+	 nop
+E(SPARC_RET_F_8)
+	st	%f7, [%i2+7*4]
+	nop
+	st	%f6, [%i2+6*4]
+	nop
+E(SPARC_RET_F_6)
+	st	%f5, [%i2+5*4]
+	nop
+	st	%f4, [%i2+4*4]
+	nop
+E(SPARC_RET_F_4)
+	std	%f2, [%i2+2*4]
+	return	%i7+8
+	 std	%f0, [%o2]
+E(SPARC_RET_F_2)
+	return	%i7+8
+	 std	%f0, [%o2]
+E(SP_V9_RET_F_3)
+	st	%f2, [%i2+2*4]
+	nop
+	st	%f1, [%i2+1*4]
+	nop
+E(SPARC_RET_F_1)
+	return	%i7+8
+	 st	%f0, [%o2]
+
+	! Finish the SPARC_RET_STRUCT sequence.
+	.align	8
+8:	stx	%o1, [%l2+8]
+	stx	%o2, [%l2+16]
+	stx	%o3, [%l2+24]
+	std	%f0, [%l2+32]
+	std	%f2, [%l2+40]
+	std	%f4, [%l2+48]
+	std	%f6, [%l2+56]
+
+	! Copy the structure into place.
+	srl	%l0, SPARC_SIZEMASK_SHIFT, %o0	! load size_mask
+	mov	%i2, %o1			! load dst
+	mov	%l2, %o2			! load src_gp
+	call	C(ffi_struct_float_copy)
+	 add	%l2, 32, %o3			! load src_fp
+
+	return	%i7+8
+	 nop
+
+.LUW2:
+	.size	C(ffi_call_v9), . - C(ffi_call_v9)
+
+
+#undef STACKFRAME
+#define	STACKFRAME	 336	/* 16*8 register window +
+				   6*8 args backing store +
+				   20*8 locals */
+#define	FP		%fp+STACK_BIAS
+
+/* ffi_closure_v9(...)
+
+   Receives the closure argument in %g1.   */
+
+	.align 8
+	.globl	C(ffi_go_closure_v9)
+	.type	C(ffi_go_closure_v9),#function
+	FFI_HIDDEN(C(ffi_go_closure_v9))
+
+C(ffi_go_closure_v9):
+.LUW3:
+	save	%sp, -STACKFRAME, %sp
+.LUW4:
+	ldx	[%g5+8], %o0
+	ldx	[%g5+16], %o1
+	b	0f
+	 mov	%g5, %o2
+
+.LUW5:
+	.size	C(ffi_go_closure_v9), . - C(ffi_go_closure_v9)
+
+	.align 8
+	.globl	C(ffi_closure_v9)
+	.type	C(ffi_closure_v9),#function
+	FFI_HIDDEN(C(ffi_closure_v9))
+
+C(ffi_closure_v9):
+.LUW6:
+	save	%sp, -STACKFRAME, %sp
+.LUW7:
+	ldx	[%g1+FFI_TRAMPOLINE_SIZE], %o0
+	ldx	[%g1+FFI_TRAMPOLINE_SIZE+8], %o1
+	ldx	[%g1+FFI_TRAMPOLINE_SIZE+16], %o2
+0:
+	! Store all of the potential argument registers in va_list format.
+	stx	%i0, [FP+128+0]
+	stx	%i1, [FP+128+8]
+	stx	%i2, [FP+128+16]
+	stx	%i3, [FP+128+24]
+	stx	%i4, [FP+128+32]
+	stx	%i5, [FP+128+40]
+
+	! Store possible floating point argument registers too.
+	std	%f0,  [FP-128]
+	std	%f2,  [FP-120]
+	std	%f4,  [FP-112]
+	std	%f6,  [FP-104]
+	std	%f8,  [FP-96]
+	std	%f10, [FP-88]
+	std     %f12, [FP-80]
+	std     %f14, [FP-72]
+	std     %f16, [FP-64]
+	std     %f18, [FP-56]
+	std     %f20, [FP-48]
+	std     %f22, [FP-40]
+	std     %f24, [FP-32]
+	std     %f26, [FP-24]
+	std     %f28, [FP-16]
+	std     %f30, [FP-8]
+
+	! Call ffi_closure_sparc_inner to do the bulk of the work.
+	add	%fp, STACK_BIAS-160, %o3
+	add	%fp, STACK_BIAS+128, %o4
+	call	C(ffi_closure_sparc_inner_v9)
+	 add	%fp, STACK_BIAS-128, %o5
+
+0:	call	1f		! load pc in %o7
+	 and	%o0, SPARC_FLAG_RET_MASK, %o0
+1:	sll	%o0, 4, %o0	! o2 = i2 * 16
+	add	%o7, %o0, %o7	! o7 = 0b + i2*16
+	jmp	%o7+(2f-0b)
+	 nop
+
+	! Note that we cannot load the data in the delay slot of
+	! the return insn because the data is in the stack frame
+	! that is deallocated by the return.
+	.align	16
+2:
+E(SPARC_RET_VOID)
+	return	%i7+8
+	 nop
+E(SPARC_RET_STRUCT)
+	ldx	[FP-160], %i0
+	ldd	[FP-160], %f0
+	b	8f
+	 ldx	[FP-152], %i1
+E(SPARC_RET_UINT8)
+	ldub	[FP-160+7], %i0
+	return	%i7+8
+	 nop
+E(SPARC_RET_SINT8)
+	ldsb	[FP-160+7], %i0
+	return	%i7+8
+	 nop
+E(SPARC_RET_UINT16)
+	lduh	[FP-160+6], %i0
+	return	%i7+8
+	 nop
+E(SPARC_RET_SINT16)
+	ldsh	[FP-160+6], %i0
+	return	%i7+8
+	 nop
+E(SPARC_RET_UINT32)
+	lduw	[FP-160+4], %i0
+	return	%i7+8
+	 nop
+E(SP_V9_RET_SINT32)
+	ldsw	[FP-160+4], %i0
+	return	%i7+8
+	 nop
+E(SPARC_RET_INT64)
+	ldx	[FP-160], %i0
+	return	%i7+8
+	 nop
+E(SPARC_RET_INT128)
+	ldx	[FP-160], %i0
+	ldx	[FP-160+8], %i1
+	return	%i7+8
+	 nop
+E(SPARC_RET_F_8)
+	ld	[FP-160+7*4], %f7
+	nop
+	ld	[FP-160+6*4], %f6
+	nop
+E(SPARC_RET_F_6)
+	ld	[FP-160+5*4], %f5
+	nop
+	ld	[FP-160+4*4], %f4
+	nop
+E(SPARC_RET_F_4)
+	ldd	[FP-160], %f0
+	ldd	[FP-160+8], %f2
+	return	%i7+8
+	 nop
+E(SPARC_RET_F_2)
+	ldd	[FP-160], %f0
+	return	%i7+8
+	 nop
+E(SP_V9_RET_F_3)
+	ld	[FP-160+2*4], %f2
+	nop
+	ld	[FP-160+1*4], %f1
+	nop
+E(SPARC_RET_F_1)
+	ld	[FP-160], %f0
+	return	%i7+8
+	 nop
+
+	! Finish the SPARC_RET_STRUCT sequence.
+	.align	8
+8:	ldd	[FP-152], %f2
+	ldx	[FP-144], %i2
+	ldd	[FP-144], %f4
+	ldx	[FP-136], %i3
+	ldd	[FP-136], %f6
+	return	%i7+8
+	 nop
+
+.LUW8:
+	.size	C(ffi_closure_v9), . - C(ffi_closure_v9)
+
+#ifdef HAVE_RO_EH_FRAME
+        .section        ".eh_frame",#alloc
+#else
+        .section        ".eh_frame",#alloc,#write
+#endif
+
+#ifdef HAVE_AS_SPARC_UA_PCREL
+# define FDE_RANGE(B, E)  .long %r_disp32(B), E - B
+#else
+# define FDE_RANGE(B, E)  .align 8; .xword B, E - B
+#endif
+
+	.align 8
+.LCIE:
+	.long	.LECIE - .LSCIE		! CIE Length
+.LSCIE:
+	.long	0			! CIE Identifier Tag
+	.byte	1			! CIE Version
+	.ascii	"zR\0"			! CIE Augmentation
+	.byte	4			! CIE Code Alignment Factor
+	.byte	0x78			! CIE Data Alignment Factor
+	.byte	15			! CIE RA Column
+	.byte	1			! Augmentation size
+#ifdef HAVE_AS_SPARC_UA_PCREL
+	.byte	0x1b			! FDE Encoding (pcrel sdata4)
+#else
+	.byte	0x50			! FDE Encoding (aligned absolute)
+#endif
+	.byte	0xc, 14, 0xff, 0xf	! DW_CFA_def_cfa, %o6, offset 0x7ff
+	.align	8
+.LECIE:
+
+	.long	.LEFDE1 - .LSFDE1	! FDE Length
+.LSFDE1:
+	.long	.LSFDE1 - .LCIE		! FDE CIE offset
+	FDE_RANGE(.LUW0, .LUW2)
+	.byte	0			! Augmentation size
+	.byte	0x40+1			! DW_CFA_advance_loc 4
+	.byte	0xd, 30			! DW_CFA_def_cfa_register, %i6
+	.byte	0x2d			! DW_CFA_GNU_window_save
+	.byte	0x9, 15, 31		! DW_CFA_register, %o7, %i7
+	.align	8
+.LEFDE1:
+
+	.long	.LEFDE2 - .LSFDE2	! FDE Length
+.LSFDE2:
+	.long	.LSFDE2 - .LCIE		! FDE CIE offset
+	FDE_RANGE(.LUW3, .LUW5)
+	.byte	0			! Augmentation size
+	.byte	0x40+1			! DW_CFA_advance_loc 4
+	.byte	0xd, 30			! DW_CFA_def_cfa_register, %i6
+	.byte	0x2d			! DW_CFA_GNU_window_save
+	.byte	0x9, 15, 31		! DW_CFA_register, %o7, %i7
+	.align	8
+.LEFDE2:
+
+	.long	.LEFDE3 - .LSFDE3	! FDE Length
+.LSFDE3:
+	.long	.LSFDE3 - .LCIE		! FDE CIE offset
+	FDE_RANGE(.LUW6, .LUW8)
+	.byte	0			! Augmentation size
+	.byte	0x40+1			! DW_CFA_advance_loc 4
+	.byte	0xd, 30			! DW_CFA_def_cfa_register, %i6
+	.byte	0x2d			! DW_CFA_GNU_window_save
+	.byte	0x9, 15, 31		! DW_CFA_register, %o7, %i7
+	.align	8
+.LEFDE3:
+
+#endif /* SPARC64 */
+#ifdef __linux__
+	.section	.note.GNU-stack,"",@progbits
+#endif

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

@@ -0,0 +1,355 @@
+/* -----------------------------------------------------------------------
+   ffi.c - Copyright (c) 2012 Tilera Corp.
+
+   TILE Foreign Function Interface
+
+   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 <ffi.h>
+#include <ffi_common.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <unistd.h>
+#include <arch/abi.h>
+#include <arch/icache.h>
+#include <arch/opcode.h>
+
+
+/* The first 10 registers are used to pass arguments and return values. */
+#define NUM_ARG_REGS 10
+
+/* Performs a raw function call with the given NUM_ARG_REGS register arguments
+   and the specified additional stack arguments (if any). */
+extern void ffi_call_tile(ffi_sarg reg_args[NUM_ARG_REGS],
+                          const ffi_sarg *stack_args,
+                          size_t stack_args_bytes,
+                          void (*fnaddr)(void))
+  FFI_HIDDEN;
+
+/* This handles the raw call from the closure stub, cleaning up the
+   parameters and delegating to ffi_closure_tile_inner. */
+extern void ffi_closure_tile(void) FFI_HIDDEN;
+
+
+ffi_status
+ffi_prep_cif_machdep(ffi_cif *cif)
+{
+  /* We always allocate room for all registers. Even if we don't
+     use them as parameters, they get returned in the same array
+     as struct return values so we need to make room. */
+  if (cif->bytes < NUM_ARG_REGS * FFI_SIZEOF_ARG)
+    cif->bytes = NUM_ARG_REGS * FFI_SIZEOF_ARG;
+
+  if (cif->rtype->size > NUM_ARG_REGS * FFI_SIZEOF_ARG)
+    cif->flags = FFI_TYPE_STRUCT;
+  else
+    cif->flags = FFI_TYPE_INT;
+
+  /* Nothing to do. */
+  return FFI_OK;
+}
+
+
+static long
+assign_to_ffi_arg(ffi_sarg *out, void *in, const ffi_type *type,
+                  int write_to_reg)
+{
+  switch (type->type)
+    {
+    case FFI_TYPE_SINT8:
+      *out = *(SINT8 *)in;
+      return 1;
+
+    case FFI_TYPE_UINT8:
+      *out = *(UINT8 *)in;
+      return 1;
+
+    case FFI_TYPE_SINT16:
+      *out = *(SINT16 *)in;
+      return 1;
+
+    case FFI_TYPE_UINT16:
+      *out = *(UINT16 *)in;
+      return 1;
+
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_UINT32:
+#ifndef __LP64__
+    case FFI_TYPE_POINTER:
+#endif
+      /* Note that even unsigned 32-bit quantities are sign extended
+         on tilegx when stored in a register.  */
+      *out = *(SINT32 *)in;
+      return 1;
+
+    case FFI_TYPE_FLOAT:
+#ifdef __tilegx__
+      if (write_to_reg)
+        {
+          /* Properly sign extend the value.  */
+          union { float f; SINT32 s32; } val;
+          val.f = *(float *)in;
+          *out = val.s32;
+        }
+      else
+#endif
+        {
+          *(float *)out = *(float *)in;
+        }
+      return 1;
+
+    case FFI_TYPE_SINT64:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_DOUBLE:
+#ifdef __LP64__
+    case FFI_TYPE_POINTER:
+#endif
+      *(UINT64 *)out = *(UINT64 *)in;
+      return sizeof(UINT64) / FFI_SIZEOF_ARG;
+
+    case FFI_TYPE_STRUCT:
+      memcpy(out, in, type->size);
+      return (type->size + FFI_SIZEOF_ARG - 1) / FFI_SIZEOF_ARG;
+
+    case FFI_TYPE_VOID:
+      /* Must be a return type. Nothing to do. */
+      return 0;
+
+    default:
+      FFI_ASSERT(0);
+      return -1;
+    }
+}
+
+
+void
+ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  ffi_sarg * const arg_mem = alloca(cif->bytes);
+  ffi_sarg * const reg_args = arg_mem;
+  ffi_sarg * const stack_args = &reg_args[NUM_ARG_REGS];
+  ffi_sarg *argp = arg_mem;
+  ffi_type ** const arg_types = cif->arg_types;
+  const long num_args = cif->nargs;
+  long i;
+
+  if (cif->flags == FFI_TYPE_STRUCT)
+    {
+      /* Pass a hidden pointer to the return value. We make sure there
+         is scratch space for the callee to store the return value even if
+         our caller doesn't care about it. */
+      *argp++ = (intptr_t)(rvalue ? rvalue : alloca(cif->rtype->size));
+
+      /* No more work needed to return anything. */
+      rvalue = NULL;
+    }
+
+  for (i = 0; i < num_args; i++)
+    {
+      ffi_type *type = arg_types[i];
+      void * const arg_in = avalue[i];
+      ptrdiff_t arg_word = argp - arg_mem;
+
+#ifndef __tilegx__
+      /* Doubleword-aligned values are always in an even-number register
+         pair, or doubleword-aligned stack slot if out of registers. */
+      long align = arg_word & (type->alignment > FFI_SIZEOF_ARG);
+      argp += align;
+      arg_word += align;
+#endif
+
+      if (type->type == FFI_TYPE_STRUCT)
+        {
+          const size_t arg_size_in_words =
+            (type->size + FFI_SIZEOF_ARG - 1) / FFI_SIZEOF_ARG;
+
+          if (arg_word < NUM_ARG_REGS &&
+              arg_word + arg_size_in_words > NUM_ARG_REGS)
+            {
+              /* Args are not allowed to span registers and the stack. */
+              argp = stack_args;
+            }
+
+          memcpy(argp, arg_in, type->size);
+          argp += arg_size_in_words;
+        }
+      else
+        {
+          argp += assign_to_ffi_arg(argp, arg_in, arg_types[i], 1);
+        }
+    }
+
+  /* Actually do the call. */
+  ffi_call_tile(reg_args, stack_args,
+                cif->bytes - (NUM_ARG_REGS * FFI_SIZEOF_ARG), fn);
+
+  if (rvalue != NULL)
+    assign_to_ffi_arg(rvalue, reg_args, cif->rtype, 0);
+}
+
+
+/* Template code for closure. */
+extern const UINT64 ffi_template_tramp_tile[] FFI_HIDDEN;
+
+
+ffi_status
+ffi_prep_closure_loc (ffi_closure *closure,
+                      ffi_cif *cif,
+                      void (*fun)(ffi_cif*, void*, void**, void*),
+                      void *user_data,
+                      void *codeloc)
+{
+#ifdef __tilegx__
+  /* TILE-Gx */
+  SINT64 c;
+  SINT64 h;
+  int s;
+  UINT64 *out;
+
+  if (cif->abi != FFI_UNIX)
+    return FFI_BAD_ABI;
+
+  out = (UINT64 *)closure->tramp;
+
+  c = (intptr_t)closure;
+  h = (intptr_t)ffi_closure_tile;
+  s = 0;
+
+  /* Find the smallest shift count that doesn't lose information
+     (i.e. no need to explicitly insert high bits of the address that
+     are just the sign extension of the low bits). */
+  while ((c >> s) != (SINT16)(c >> s) || (h >> s) != (SINT16)(h >> s))
+    s += 16;
+
+#define OPS(a, b, shift) \
+  (create_Imm16_X0((a) >> (shift)) | create_Imm16_X1((b) >> (shift)))
+
+  /* Emit the moveli. */
+  *out++ = ffi_template_tramp_tile[0] | OPS(c, h, s);
+  for (s -= 16; s >= 0; s -= 16)
+    *out++ = ffi_template_tramp_tile[1] | OPS(c, h, s);
+
+#undef OPS
+
+  *out++ = ffi_template_tramp_tile[2];
+
+#else
+  /* TILEPro */
+  UINT64 *out;
+  intptr_t delta;
+
+  if (cif->abi != FFI_UNIX)
+    return FFI_BAD_ABI;
+
+  out = (UINT64 *)closure->tramp;
+  delta = (intptr_t)ffi_closure_tile - (intptr_t)codeloc;
+
+  *out++ = ffi_template_tramp_tile[0] | create_JOffLong_X1(delta >> 3);
+#endif
+
+  closure->cif = cif;
+  closure->fun = fun;
+  closure->user_data = user_data;
+
+  invalidate_icache(closure->tramp, (char *)out - closure->tramp,
+                    getpagesize());
+
+  return FFI_OK;
+}
+
+
+/* This is called by the assembly wrapper for closures. This does
+   all of the work. On entry reg_args[0] holds the values the registers
+   had when the closure was invoked. On return reg_args[1] holds the register
+   values to be returned to the caller (many of which may be garbage). */
+void FFI_HIDDEN
+ffi_closure_tile_inner(ffi_closure *closure,
+                       ffi_sarg reg_args[2][NUM_ARG_REGS],
+                       ffi_sarg *stack_args)
+{
+  ffi_cif * const cif = closure->cif;
+  void ** const avalue = alloca(cif->nargs * sizeof(void *));
+  void *rvalue;
+  ffi_type ** const arg_types = cif->arg_types;
+  ffi_sarg * const reg_args_in = reg_args[0];
+  ffi_sarg * const reg_args_out = reg_args[1];
+  ffi_sarg * argp;
+  long i, arg_word, nargs = cif->nargs;
+  /* Use a union to guarantee proper alignment for double. */
+  union { ffi_sarg arg[NUM_ARG_REGS]; double d; UINT64 u64; } closure_ret;
+
+  /* Start out reading register arguments. */
+  argp = reg_args_in;
+
+  /* Copy the caller's structure return address to that the closure
+     returns the data directly to the caller.  */
+  if (cif->flags == FFI_TYPE_STRUCT)
+    {
+      /* Return by reference via hidden pointer. */
+      rvalue = (void *)(intptr_t)*argp++;
+      arg_word = 1;
+    }
+  else
+    {
+      /* Return the value in registers. */
+      rvalue = &closure_ret;
+      arg_word = 0;
+    }
+
+  /* Grab the addresses of the arguments. */
+  for (i = 0; i < nargs; i++)
+    {
+      ffi_type * const type = arg_types[i];
+      const size_t arg_size_in_words =
+        (type->size + FFI_SIZEOF_ARG - 1) / FFI_SIZEOF_ARG;
+
+#ifndef __tilegx__
+      /* Doubleword-aligned values are always in an even-number register
+         pair, or doubleword-aligned stack slot if out of registers. */
+      long align = arg_word & (type->alignment > FFI_SIZEOF_ARG);
+      argp += align;
+      arg_word += align;
+#endif
+
+      if (arg_word == NUM_ARG_REGS ||
+          (arg_word < NUM_ARG_REGS &&
+           arg_word + arg_size_in_words > NUM_ARG_REGS))
+        {
+          /* Switch to reading arguments from the stack. */
+          argp = stack_args;
+          arg_word = NUM_ARG_REGS;
+        }
+
+      avalue[i] = argp;
+      argp += arg_size_in_words;
+      arg_word += arg_size_in_words;
+    }
+
+  /* Invoke the closure.  */
+  closure->fun(cif, rvalue, avalue, closure->user_data);
+
+  if (cif->flags != FFI_TYPE_STRUCT)
+    {
+      /* Canonicalize for register representation. */
+      assign_to_ffi_arg(reg_args_out, &closure_ret, cif->rtype, 1);
+    }
+}