remogatto-ffi 0.5.0

data/ext/ffi_c/libffi/src/mips/o32.S

@@ -0,0 +1,381 @@
+/* -----------------------------------------------------------------------
+   o32.S - Copyright (c) 1996, 1998, 2005  Red Hat, Inc.
+   
+   MIPS 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>
+
+/* Only build this code if we are compiling for o32 */	
+
+#if defined(FFI_MIPS_O32)
+	
+#define callback a0
+#define bytes	 a2
+#define flags	 a3
+		
+#define SIZEOF_FRAME	(4 * FFI_SIZEOF_ARG + 2 * FFI_SIZEOF_ARG)
+#define A3_OFF		(SIZEOF_FRAME + 3 * FFI_SIZEOF_ARG)
+#define FP_OFF		(SIZEOF_FRAME - 2 * FFI_SIZEOF_ARG)
+#define RA_OFF		(SIZEOF_FRAME - 1 * FFI_SIZEOF_ARG)
+
+	.abicalls
+	.text
+	.align	2
+	.globl	ffi_call_O32
+	.ent	ffi_call_O32
+ffi_call_O32:	
+$LFB0:
+	# Prologue
+	SUBU	$sp, SIZEOF_FRAME	# Frame size
+$LCFI0:
+	REG_S	$fp, FP_OFF($sp)	# Save frame pointer
+$LCFI1:
+	REG_S	ra, RA_OFF($sp)		# Save return address
+$LCFI2:
+	move	$fp, $sp
+
+$LCFI3:
+	move	t9, callback		# callback function pointer
+	REG_S	flags, A3_OFF($fp)	# flags
+
+	# Allocate at least 4 words in the argstack
+	LI	v0, 4 * FFI_SIZEOF_ARG
+	blt	bytes, v0, sixteen
+
+	ADDU	v0, bytes, 7	# make sure it is aligned 
+	and	v0, -8		# to an 8 byte boundry
+
+sixteen:
+	SUBU	$sp, v0		# move the stack pointer to reflect the
+				# arg space
+
+	ADDU	a0, $sp, 4 * FFI_SIZEOF_ARG
+
+	jalr	t9
+	
+	REG_L	t0, A3_OFF($fp)		# load the flags word
+	SRL	t2, t0, 4		# shift our arg info
+	and     t0, ((1<<4)-1)          # mask out the return type
+		
+	ADDU	$sp, 4 * FFI_SIZEOF_ARG		# adjust $sp to new args
+
+	bnez	t0, pass_d			# make it quick for int
+	REG_L	a0, 0*FFI_SIZEOF_ARG($sp)	# just go ahead and load the
+	REG_L	a1, 1*FFI_SIZEOF_ARG($sp)	# four regs.
+	REG_L	a2, 2*FFI_SIZEOF_ARG($sp)
+	REG_L	a3, 3*FFI_SIZEOF_ARG($sp)
+	b	call_it
+
+pass_d:
+	bne	t0, FFI_ARGS_D, pass_f
+	l.d	$f12, 0*FFI_SIZEOF_ARG($sp)	# load $fp regs from args
+	REG_L	a2,   2*FFI_SIZEOF_ARG($sp)	# passing a double
+	REG_L	a3,   3*FFI_SIZEOF_ARG($sp)
+	b	call_it
+
+pass_f:	
+	bne	t0, FFI_ARGS_F, pass_d_d
+	l.s	$f12, 0*FFI_SIZEOF_ARG($sp)	# load $fp regs from args
+	REG_L	a1,   1*FFI_SIZEOF_ARG($sp)	# passing a float
+	REG_L	a2,   2*FFI_SIZEOF_ARG($sp)
+	REG_L	a3,   3*FFI_SIZEOF_ARG($sp)
+	b	call_it		
+
+pass_d_d:		
+	bne	t0, FFI_ARGS_DD, pass_f_f
+	l.d	$f12, 0*FFI_SIZEOF_ARG($sp)	# load $fp regs from args
+	l.d	$f14, 2*FFI_SIZEOF_ARG($sp)	# passing two doubles
+	b	call_it
+
+pass_f_f:	
+	bne	t0, FFI_ARGS_FF, pass_d_f
+	l.s	$f12, 0*FFI_SIZEOF_ARG($sp)	# load $fp regs from args
+	l.s	$f14, 1*FFI_SIZEOF_ARG($sp)	# passing two floats
+	REG_L	a2,   2*FFI_SIZEOF_ARG($sp)
+	REG_L	a3,   3*FFI_SIZEOF_ARG($sp)
+	b	call_it
+
+pass_d_f:		
+	bne	t0, FFI_ARGS_DF, pass_f_d
+	l.d	$f12, 0*FFI_SIZEOF_ARG($sp)	# load $fp regs from args
+	l.s	$f14, 2*FFI_SIZEOF_ARG($sp)	# passing double and float
+	REG_L	a3,   3*FFI_SIZEOF_ARG($sp)
+	b	call_it
+
+pass_f_d:		
+ # assume that the only other combination must be float then double
+ #	bne	t0, FFI_ARGS_F_D, call_it
+	l.s	$f12, 0*FFI_SIZEOF_ARG($sp)	# load $fp regs from args
+	l.d	$f14, 2*FFI_SIZEOF_ARG($sp)	# passing double and float
+
+call_it:	
+	# Load the function pointer
+	REG_L	t9, SIZEOF_FRAME + 5*FFI_SIZEOF_ARG($fp)
+
+	# If the return value pointer is NULL, assume no return value.
+	REG_L	t1, SIZEOF_FRAME + 4*FFI_SIZEOF_ARG($fp)
+	beqz	t1, noretval
+
+	bne     t2, FFI_TYPE_INT, retlonglong
+	jalr	t9
+	REG_L	t0, SIZEOF_FRAME + 4*FFI_SIZEOF_ARG($fp)
+	REG_S	v0, 0(t0)
+	b	epilogue
+
+retlonglong:
+	# Really any 64-bit int, signed or not.
+	bne	t2, FFI_TYPE_UINT64, retfloat
+	jalr	t9
+	REG_L	t0, SIZEOF_FRAME + 4*FFI_SIZEOF_ARG($fp)
+	REG_S	v1, 4(t0)
+	REG_S	v0, 0(t0)
+	b	epilogue
+
+retfloat:
+	bne     t2, FFI_TYPE_FLOAT, retdouble
+	jalr	t9
+	REG_L	t0, SIZEOF_FRAME + 4*FFI_SIZEOF_ARG($fp)
+	s.s	$f0, 0(t0)
+	b	epilogue
+
+retdouble:	
+	bne	t2, FFI_TYPE_DOUBLE, noretval
+	jalr	t9
+	REG_L	t0, SIZEOF_FRAME + 4*FFI_SIZEOF_ARG($fp)
+	s.d	$f0, 0(t0)
+	b	epilogue
+	
+noretval:	
+	jalr	t9
+	
+	# Epilogue
+epilogue:	
+	move	$sp, $fp	
+	REG_L	$fp, FP_OFF($sp)	# Restore frame pointer
+	REG_L	ra, RA_OFF($sp)		# Restore return address
+	ADDU	$sp, SIZEOF_FRAME	# Fix stack pointer
+	j	ra
+
+$LFE0:
+	.end	ffi_call_O32
+
+
+/* ffi_closure_O32. Expects address of the passed-in ffi_closure
+	in t4 ($12). Stores any arguments passed in registers onto the
+	stack, then calls ffi_closure_mips_inner_O32, which
+	then decodes them.
+	
+	Stack layout:
+
+	 3 - a3 save
+	 2 - a2 save
+	 1 - a1 save
+	 0 - a0 save, original sp
+	-1 - ra save
+	-2 - fp save
+	-3 - $16 (s0) save
+	-4 - cprestore
+	-5 - return value high (v1)
+	-6 - return value low (v0)
+	-7 - f14 (le high, be low)
+	-8 - f14 (le low, be high)
+	-9 - f12 (le high, be low)
+       -10 - f12 (le low, be high)
+       -11 - Called function a3 save
+       -12 - Called function a2 save
+       -13 - Called function a1 save
+       -14 - Called function a0 save, our sp and fp point here
+	 */
+	
+#define SIZEOF_FRAME2	(14 * FFI_SIZEOF_ARG)
+#define A3_OFF2		(SIZEOF_FRAME2 + 3 * FFI_SIZEOF_ARG)
+#define A2_OFF2		(SIZEOF_FRAME2 + 2 * FFI_SIZEOF_ARG)
+#define A1_OFF2		(SIZEOF_FRAME2 + 1 * FFI_SIZEOF_ARG)
+#define A0_OFF2		(SIZEOF_FRAME2 + 0 * FFI_SIZEOF_ARG)
+#define RA_OFF2		(SIZEOF_FRAME2 - 1 * FFI_SIZEOF_ARG)
+#define FP_OFF2		(SIZEOF_FRAME2 - 2 * FFI_SIZEOF_ARG)
+#define S0_OFF2		(SIZEOF_FRAME2 - 3 * FFI_SIZEOF_ARG)
+#define GP_OFF2		(SIZEOF_FRAME2 - 4 * FFI_SIZEOF_ARG)
+#define V1_OFF2		(SIZEOF_FRAME2 - 5 * FFI_SIZEOF_ARG)
+#define V0_OFF2		(SIZEOF_FRAME2 - 6 * FFI_SIZEOF_ARG)
+#define FA_1_1_OFF2	(SIZEOF_FRAME2 - 7 * FFI_SIZEOF_ARG)
+#define FA_1_0_OFF2	(SIZEOF_FRAME2 - 8 * FFI_SIZEOF_ARG)
+#define FA_0_1_OFF2	(SIZEOF_FRAME2 - 9 * FFI_SIZEOF_ARG)
+#define FA_0_0_OFF2	(SIZEOF_FRAME2 - 10 * FFI_SIZEOF_ARG)
+
+	.text
+	.align	2
+	.globl	ffi_closure_O32
+	.ent	ffi_closure_O32
+ffi_closure_O32:
+$LFB1:
+	# Prologue
+	.frame	$fp, SIZEOF_FRAME2, ra
+	.set	noreorder
+	.cpload	t9
+	.set	reorder
+	SUBU	$sp, SIZEOF_FRAME2
+	.cprestore GP_OFF2
+$LCFI4:
+	REG_S	$16, S0_OFF2($sp)	 # Save s0
+	REG_S	$fp, FP_OFF2($sp)	 # Save frame pointer
+	REG_S	ra, RA_OFF2($sp)	 # Save return address
+$LCFI6:
+	move	$fp, $sp
+
+$LCFI7:
+	# Store all possible argument registers. If there are more than
+	# four arguments, then they are stored above where we put a3.
+	REG_S	a0, A0_OFF2($fp)
+	REG_S	a1, A1_OFF2($fp)
+	REG_S	a2, A2_OFF2($fp)
+	REG_S	a3, A3_OFF2($fp)
+
+	# Load ABI enum to s0
+	REG_L	$16, 20($12)	# cif pointer follows tramp.
+	REG_L	$16, 0($16)	# abi is first member.
+
+	li	$13, 1		# FFI_O32
+	bne	$16, $13, 1f	# Skip fp save if FFI_O32_SOFT_FLOAT
+	
+	# Store all possible float/double registers.
+	s.d	$f12, FA_0_0_OFF2($fp)
+	s.d	$f14, FA_1_0_OFF2($fp)
+1:	
+	# Call ffi_closure_mips_inner_O32 to do the work.
+	la	t9, ffi_closure_mips_inner_O32
+	move	a0, $12	 # Pointer to the ffi_closure
+	addu	a1, $fp, V0_OFF2
+	addu	a2, $fp, A0_OFF2
+	addu	a3, $fp, FA_0_0_OFF2
+	jalr	t9
+
+	# Load the return value into the appropriate register.
+	move	$8, $2
+	li	$9, FFI_TYPE_VOID
+	beq	$8, $9, closure_done
+
+	li	$13, 1		# FFI_O32
+	bne	$16, $13, 1f	# Skip fp restore if FFI_O32_SOFT_FLOAT
+
+	li	$9, FFI_TYPE_FLOAT
+	l.s	$f0, V0_OFF2($fp)
+	beq	$8, $9, closure_done
+
+	li	$9, FFI_TYPE_DOUBLE
+	l.d	$f0, V0_OFF2($fp)
+	beq	$8, $9, closure_done
+1:	
+	REG_L	$3, V1_OFF2($fp)
+	REG_L	$2, V0_OFF2($fp)
+
+closure_done:
+	# Epilogue
+	move	$sp, $fp
+	REG_L	$16, S0_OFF2($sp)	 # Restore s0
+	REG_L	$fp, FP_OFF2($sp)	 # Restore frame pointer
+	REG_L	ra,  RA_OFF2($sp)	 # Restore return address
+	ADDU	$sp, SIZEOF_FRAME2
+	j	ra
+$LFE1:
+	.end	ffi_closure_O32
+
+/* DWARF-2 unwind info. */
+
+	.section	.eh_frame,"a",@progbits
+$Lframe0:
+	.4byte	$LECIE0-$LSCIE0	 # Length of Common Information Entry
+$LSCIE0:
+	.4byte	0x0	 # CIE Identifier Tag
+	.byte	0x1	 # CIE Version
+	.ascii "zR\0"	 # CIE Augmentation
+	.uleb128 0x1	 # CIE Code Alignment Factor
+	.sleb128 4	 # CIE Data Alignment Factor
+	.byte	0x1f	 # CIE RA Column
+	.uleb128 0x1	 # Augmentation size
+	.byte	0x00	 # FDE Encoding (absptr)
+	.byte	0xc	 # DW_CFA_def_cfa
+	.uleb128 0x1d
+	.uleb128 0x0
+	.align	2
+$LECIE0:
+$LSFDE0:
+	.4byte	$LEFDE0-$LASFDE0	 # FDE Length
+$LASFDE0:
+	.4byte	$LASFDE0-$Lframe0	 # FDE CIE offset
+	.4byte	$LFB0	 # FDE initial location
+	.4byte	$LFE0-$LFB0	 # FDE address range
+	.uleb128 0x0	 # Augmentation size
+	.byte	0x4	 # DW_CFA_advance_loc4
+	.4byte	$LCFI0-$LFB0
+	.byte	0xe	 # DW_CFA_def_cfa_offset
+	.uleb128 0x18
+	.byte	0x4	 # DW_CFA_advance_loc4
+	.4byte	$LCFI2-$LCFI0
+	.byte	0x11	 # DW_CFA_offset_extended_sf
+	.uleb128 0x1e	 # $fp
+	.sleb128 -2	 # SIZEOF_FRAME2 - 2*FFI_SIZEOF_ARG($sp)
+	.byte	0x11	 # DW_CFA_offset_extended_sf
+	.uleb128 0x1f	 # $ra
+	.sleb128 -1	 # SIZEOF_FRAME2 - 1*FFI_SIZEOF_ARG($sp)
+	.byte	0x4	 # DW_CFA_advance_loc4
+	.4byte	$LCFI3-$LCFI2
+	.byte	0xc	 # DW_CFA_def_cfa
+	.uleb128 0x1e
+	.uleb128 0x18
+	.align	2
+$LEFDE0:
+$LSFDE1:
+	.4byte	$LEFDE1-$LASFDE1	 # FDE Length
+$LASFDE1:
+	.4byte	$LASFDE1-$Lframe0	 # FDE CIE offset
+	.4byte	$LFB1	 # FDE initial location
+	.4byte	$LFE1-$LFB1	 # FDE address range
+	.uleb128 0x0	 # Augmentation size
+	.byte	0x4	 # DW_CFA_advance_loc4
+	.4byte	$LCFI4-$LFB1
+	.byte	0xe	 # DW_CFA_def_cfa_offset
+	.uleb128 0x38
+	.byte	0x4	 # DW_CFA_advance_loc4
+	.4byte	$LCFI6-$LCFI4
+	.byte	0x11	 # DW_CFA_offset_extended_sf
+	.uleb128 0x10	 # $16
+	.sleb128 -3	 # SIZEOF_FRAME2 - 3*FFI_SIZEOF_ARG($sp)
+	.byte	0x11	 # DW_CFA_offset_extended_sf
+	.uleb128 0x1e	 # $fp
+	.sleb128 -2	 # SIZEOF_FRAME2 - 2*FFI_SIZEOF_ARG($sp)
+	.byte	0x11	 # DW_CFA_offset_extended_sf
+	.uleb128 0x1f	 # $ra
+	.sleb128 -1	 # SIZEOF_FRAME2 - 1*FFI_SIZEOF_ARG($sp)
+	.byte	0x4	 # DW_CFA_advance_loc4
+	.4byte	$LCFI7-$LCFI6
+	.byte	0xc	 # DW_CFA_def_cfa
+	.uleb128 0x1e
+	.uleb128 0x38
+	.align	2
+$LEFDE1:
+
+#endif

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

@@ -0,0 +1,709 @@
+/* -----------------------------------------------------------------------
+   ffi.c - (c) 2003-2004 Randolph Chung <tausq@debian.org>
+           (c) 2008 Red Hat, Inc.
+
+   HPPA Foreign Function Interface
+   HP-UX PA ABI support (c) 2006 Free Software Foundation, Inc.
+
+   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 <stdio.h>
+
+#define ROUND_UP(v, a)  (((size_t)(v) + (a) - 1) & ~((a) - 1))
+
+#define MIN_STACK_SIZE  64
+#define FIRST_ARG_SLOT  9
+#define DEBUG_LEVEL   0
+
+#define fldw(addr, fpreg) \
+  __asm__ volatile ("fldw 0(%0), %%" #fpreg "L" : : "r"(addr) : #fpreg)
+#define fstw(fpreg, addr) \
+  __asm__ volatile ("fstw %%" #fpreg "L, 0(%0)" : : "r"(addr))
+#define fldd(addr, fpreg) \
+  __asm__ volatile ("fldd 0(%0), %%" #fpreg : : "r"(addr) : #fpreg)
+#define fstd(fpreg, addr) \
+  __asm__ volatile ("fstd %%" #fpreg "L, 0(%0)" : : "r"(addr))
+
+#define debug(lvl, x...) do { if (lvl <= DEBUG_LEVEL) { printf(x); } } while (0)
+
+static inline int ffi_struct_type(ffi_type *t)
+{
+  size_t sz = t->size;
+
+  /* Small structure results are passed in registers,
+     larger ones are passed by pointer.  Note that
+     small structures of size 2, 4 and 8 differ from
+     the corresponding integer types in that they have
+     different alignment requirements.  */
+
+  if (sz <= 1)
+    return FFI_TYPE_UINT8;
+  else if (sz == 2)
+    return FFI_TYPE_SMALL_STRUCT2;
+  else if (sz == 3)
+    return FFI_TYPE_SMALL_STRUCT3;
+  else if (sz == 4)
+    return FFI_TYPE_SMALL_STRUCT4;
+  else if (sz == 5)
+    return FFI_TYPE_SMALL_STRUCT5;
+  else if (sz == 6)
+    return FFI_TYPE_SMALL_STRUCT6;
+  else if (sz == 7)
+    return FFI_TYPE_SMALL_STRUCT7;
+  else if (sz <= 8)
+    return FFI_TYPE_SMALL_STRUCT8;
+  else
+    return FFI_TYPE_STRUCT; /* else, we pass it by pointer.  */
+}
+
+/* PA has a downward growing stack, which looks like this:
+
+   Offset
+	[ Variable args ]
+   SP = (4*(n+9))       arg word N
+   ...
+   SP-52                arg word 4
+	[ Fixed args ]
+   SP-48                arg word 3
+   SP-44                arg word 2
+   SP-40                arg word 1
+   SP-36                arg word 0
+	[ Frame marker ]
+   ...
+   SP-20                RP
+   SP-4                 previous SP
+
+   The first four argument words on the stack are reserved for use by
+   the callee.  Instead, the general and floating registers replace
+   the first four argument slots.  Non FP arguments are passed solely
+   in the general registers.  FP arguments are passed in both general
+   and floating registers when using libffi.
+
+   Non-FP 32-bit args are passed in gr26, gr25, gr24 and gr23.
+   Non-FP 64-bit args are passed in register pairs, starting
+   on an odd numbered register (i.e. r25+r26 and r23+r24).
+   FP 32-bit arguments are passed in fr4L, fr5L, fr6L and fr7L.
+   FP 64-bit arguments are passed in fr5 and fr7.
+
+   The registers are allocated in the same manner as stack slots.
+   This allows the callee to save its arguments on the stack if
+   necessary:
+
+   arg word 3 -> gr23 or fr7L
+   arg word 2 -> gr24 or fr6L or fr7R
+   arg word 1 -> gr25 or fr5L
+   arg word 0 -> gr26 or fr4L or fr5R
+
+   Note that fr4R and fr6R are never used for arguments (i.e.,
+   doubles are not passed in fr4 or fr6).
+
+   The rest of the arguments are passed on the stack starting at SP-52,
+   but 64-bit arguments need to be aligned to an 8-byte boundary
+
+   This means we can have holes either in the register allocation,
+   or in the stack.  */
+
+/* ffi_prep_args is called by the assembly routine once stack space
+   has been allocated for the function's arguments
+
+   The following code will put everything into the stack frame
+   (which was allocated by the asm routine), and on return
+   the asm routine will load the arguments that should be
+   passed by register into the appropriate registers
+
+   NOTE: We load floating point args in this function... that means we
+   assume gcc will not mess with fp regs in here.  */
+
+void ffi_prep_args_pa32(UINT32 *stack, extended_cif *ecif, unsigned bytes)
+{
+  register unsigned int i;
+  register ffi_type **p_arg;
+  register void **p_argv;
+  unsigned int slot = FIRST_ARG_SLOT;
+  char *dest_cpy;
+  size_t len;
+
+  debug(1, "%s: stack = %p, ecif = %p, bytes = %u\n", __FUNCTION__, stack,
+	ecif, bytes);
+
+  p_arg = ecif->cif->arg_types;
+  p_argv = ecif->avalue;
+
+  for (i = 0; i < ecif->cif->nargs; i++)
+    {
+      int type = (*p_arg)->type;
+
+      switch (type)
+	{
+	case FFI_TYPE_SINT8:
+	  *(SINT32 *)(stack - slot) = *(SINT8 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT8:
+	  *(UINT32 *)(stack - slot) = *(UINT8 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_SINT16:
+	  *(SINT32 *)(stack - slot) = *(SINT16 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT16:
+	  *(UINT32 *)(stack - slot) = *(UINT16 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_POINTER:
+	  debug(3, "Storing UINT32 %u in slot %u\n", *(UINT32 *)(*p_argv),
+		slot);
+	  *(UINT32 *)(stack - slot) = *(UINT32 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	  /* Align slot for 64-bit type.  */
+	  slot += (slot & 1) ? 1 : 2;
+	  *(UINT64 *)(stack - slot) = *(UINT64 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_FLOAT:
+	  /* First 4 args go in fr4L - fr7L.  */
+	  debug(3, "Storing UINT32(float) in slot %u\n", slot);
+	  *(UINT32 *)(stack - slot) = *(UINT32 *)(*p_argv);
+	  switch (slot - FIRST_ARG_SLOT)
+	    {
+	    /* First 4 args go in fr4L - fr7L.  */
+	    case 0: fldw(stack - slot, fr4); break;
+	    case 1: fldw(stack - slot, fr5); break;
+	    case 2: fldw(stack - slot, fr6); break;
+	    case 3: fldw(stack - slot, fr7); break;
+	    }
+	  break;
+
+	case FFI_TYPE_DOUBLE:
+	  /* Align slot for 64-bit type.  */
+	  slot += (slot & 1) ? 1 : 2;
+	  debug(3, "Storing UINT64(double) at slot %u\n", slot);
+	  *(UINT64 *)(stack - slot) = *(UINT64 *)(*p_argv);
+	  switch (slot - FIRST_ARG_SLOT)
+	    {
+	      /* First 2 args go in fr5, fr7.  */
+	      case 1: fldd(stack - slot, fr5); break;
+	      case 3: fldd(stack - slot, fr7); break;
+	    }
+	  break;
+
+#ifdef PA_HPUX
+	case FFI_TYPE_LONGDOUBLE:
+	  /* Long doubles are passed in the same manner as structures
+	     larger than 8 bytes.  */
+	  *(UINT32 *)(stack - slot) = (UINT32)(*p_argv);
+	  break;
+#endif
+
+	case FFI_TYPE_STRUCT:
+
+	  /* Structs smaller or equal than 4 bytes are passed in one
+	     register. Structs smaller or equal 8 bytes are passed in two
+	     registers. Larger structures are passed by pointer.  */
+
+	  len = (*p_arg)->size;
+	  if (len <= 4)
+	    {
+	      dest_cpy = (char *)(stack - slot) + 4 - len;
+	      memcpy(dest_cpy, (char *)*p_argv, len);
+	    }
+	  else if (len <= 8)
+	    {
+	      slot += (slot & 1) ? 1 : 2;
+	      dest_cpy = (char *)(stack - slot) + 8 - len;
+	      memcpy(dest_cpy, (char *)*p_argv, len);
+	    }
+	  else
+	    *(UINT32 *)(stack - slot) = (UINT32)(*p_argv);
+	  break;
+
+	default:
+	  FFI_ASSERT(0);
+	}
+
+      slot++;
+      p_arg++;
+      p_argv++;
+    }
+
+  /* Make sure we didn't mess up and scribble on the stack.  */
+  {
+    unsigned int n;
+
+    debug(5, "Stack setup:\n");
+    for (n = 0; n < (bytes + 3) / 4; n++)
+      {
+	if ((n%4) == 0) { debug(5, "\n%08x: ", (unsigned int)(stack - n)); }
+	debug(5, "%08x ", *(stack - n));
+      }
+    debug(5, "\n");
+  }
+
+  FFI_ASSERT(slot * 4 <= bytes);
+
+  return;
+}
+
+static void ffi_size_stack_pa32(ffi_cif *cif)
+{
+  ffi_type **ptr;
+  int i;
+  int z = 0; /* # stack slots */
+
+  for (ptr = cif->arg_types, i = 0; i < cif->nargs; ptr++, i++)
+    {
+      int type = (*ptr)->type;
+
+      switch (type)
+	{
+	case FFI_TYPE_DOUBLE:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	  z += 2 + (z & 1); /* must start on even regs, so we may waste one */
+	  break;
+
+#ifdef PA_HPUX
+	case FFI_TYPE_LONGDOUBLE:
+#endif
+	case FFI_TYPE_STRUCT:
+	  z += 1; /* pass by ptr, callee will copy */
+	  break;
+
+	default: /* <= 32-bit values */
+	  z++;
+	}
+    }
+
+  /* We can fit up to 6 args in the default 64-byte stack frame,
+     if we need more, we need more stack.  */
+  if (z <= 6)
+    cif->bytes = MIN_STACK_SIZE; /* min stack size */
+  else
+    cif->bytes = 64 + ROUND_UP((z - 6) * sizeof(UINT32), MIN_STACK_SIZE);
+
+  debug(3, "Calculated stack size is %u bytes\n", cif->bytes);
+}
+
+/* Perform machine dependent cif processing.  */
+ffi_status ffi_prep_cif_machdep(ffi_cif *cif)
+{
+  /* Set the return type flag */
+  switch (cif->rtype->type)
+    {
+    case FFI_TYPE_VOID:
+    case FFI_TYPE_FLOAT:
+    case FFI_TYPE_DOUBLE:
+      cif->flags = (unsigned) cif->rtype->type;
+      break;
+
+#ifdef PA_HPUX
+    case FFI_TYPE_LONGDOUBLE:
+      /* Long doubles are treated like a structure.  */
+      cif->flags = FFI_TYPE_STRUCT;
+      break;
+#endif
+
+    case FFI_TYPE_STRUCT:
+      /* For the return type we have to check the size of the structures.
+	 If the size is smaller or equal 4 bytes, the result is given back
+	 in one register. If the size is smaller or equal 8 bytes than we
+	 return the result in two registers. But if the size is bigger than
+	 8 bytes, we work with pointers.  */
+      cif->flags = ffi_struct_type(cif->rtype);
+      break;
+
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+      cif->flags = FFI_TYPE_UINT64;
+      break;
+
+    default:
+      cif->flags = FFI_TYPE_INT;
+      break;
+    }
+
+  /* Lucky us, because of the unique PA ABI we get to do our
+     own stack sizing.  */
+  switch (cif->abi)
+    {
+    case FFI_PA32:
+      ffi_size_stack_pa32(cif);
+      break;
+
+    default:
+      FFI_ASSERT(0);
+      break;
+    }
+
+  return FFI_OK;
+}
+
+extern void ffi_call_pa32(void (*)(UINT32 *, extended_cif *, unsigned),
+			  extended_cif *, unsigned, unsigned, unsigned *,
+			  void (*fn)(void));
+
+void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  extended_cif ecif;
+
+  ecif.cif = cif;
+  ecif.avalue = avalue;
+
+  /* If the return value is a struct and we don't have a return
+     value address then we need to make one.  */
+
+  if (rvalue == NULL
+#ifdef PA_HPUX
+      && (cif->rtype->type == FFI_TYPE_STRUCT
+	  || cif->rtype->type == FFI_TYPE_LONGDOUBLE))
+#else
+      && cif->rtype->type == FFI_TYPE_STRUCT)
+#endif
+    {
+      ecif.rvalue = alloca(cif->rtype->size);
+    }
+  else
+    ecif.rvalue = rvalue;
+
+
+  switch (cif->abi)
+    {
+    case FFI_PA32:
+      debug(3, "Calling ffi_call_pa32: ecif=%p, bytes=%u, flags=%u, rvalue=%p, fn=%p\n", &ecif, cif->bytes, cif->flags, ecif.rvalue, (void *)fn);
+      ffi_call_pa32(ffi_prep_args_pa32, &ecif, cif->bytes,
+		     cif->flags, ecif.rvalue, fn);
+      break;
+
+    default:
+      FFI_ASSERT(0);
+      break;
+    }
+}
+
+#if FFI_CLOSURES
+/* This is more-or-less an inverse of ffi_call -- we have arguments on
+   the stack, and we need to fill them into a cif structure and invoke
+   the user function. This really ought to be in asm to make sure
+   the compiler doesn't do things we don't expect.  */
+ffi_status ffi_closure_inner_pa32(ffi_closure *closure, UINT32 *stack)
+{
+  ffi_cif *cif;
+  void **avalue;
+  void *rvalue;
+  UINT32 ret[2]; /* function can return up to 64-bits in registers */
+  ffi_type **p_arg;
+  char *tmp;
+  int i, avn;
+  unsigned int slot = FIRST_ARG_SLOT;
+  register UINT32 r28 asm("r28");
+
+  cif = closure->cif;
+
+  /* If returning via structure, callee will write to our pointer.  */
+  if (cif->flags == FFI_TYPE_STRUCT)
+    rvalue = (void *)r28;
+  else
+    rvalue = &ret[0];
+
+  avalue = (void **)alloca(cif->nargs * FFI_SIZEOF_ARG);
+  avn = cif->nargs;
+  p_arg = cif->arg_types;
+
+  for (i = 0; i < avn; i++)
+    {
+      int type = (*p_arg)->type;
+
+      switch (type)
+	{
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT8:
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT16:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_POINTER:
+	  avalue[i] = (char *)(stack - slot) + sizeof(UINT32) - (*p_arg)->size;
+	  break;
+
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_UINT64:
+	  slot += (slot & 1) ? 1 : 2;
+	  avalue[i] = (void *)(stack - slot);
+	  break;
+
+	case FFI_TYPE_FLOAT:
+#ifdef PA_LINUX
+	  /* The closure call is indirect.  In Linux, floating point
+	     arguments in indirect calls with a prototype are passed
+	     in the floating point registers instead of the general
+	     registers.  So, we need to replace what was previously
+	     stored in the current slot with the value in the
+	     corresponding floating point register.  */
+	  switch (slot - FIRST_ARG_SLOT)
+	    {
+	    case 0: fstw(fr4, (void *)(stack - slot)); break;
+	    case 1: fstw(fr5, (void *)(stack - slot)); break;
+	    case 2: fstw(fr6, (void *)(stack - slot)); break;
+	    case 3: fstw(fr7, (void *)(stack - slot)); break;
+	    }
+#endif
+	  avalue[i] = (void *)(stack - slot);
+	  break;
+
+	case FFI_TYPE_DOUBLE:
+	  slot += (slot & 1) ? 1 : 2;
+#ifdef PA_LINUX
+	  /* See previous comment for FFI_TYPE_FLOAT.  */
+	  switch (slot - FIRST_ARG_SLOT)
+	    {
+	    case 1: fstd(fr5, (void *)(stack - slot)); break;
+	    case 3: fstd(fr7, (void *)(stack - slot)); break;
+	    }
+#endif
+	  avalue[i] = (void *)(stack - slot);
+	  break;
+
+	case FFI_TYPE_STRUCT:
+	  /* Structs smaller or equal than 4 bytes are passed in one
+	     register. Structs smaller or equal 8 bytes are passed in two
+	     registers. Larger structures are passed by pointer.  */
+	  if((*p_arg)->size <= 4)
+	    {
+	      avalue[i] = (void *)(stack - slot) + sizeof(UINT32) -
+		(*p_arg)->size;
+	    }
+	  else if ((*p_arg)->size <= 8)
+	    {
+	      slot += (slot & 1) ? 1 : 2;
+	      avalue[i] = (void *)(stack - slot) + sizeof(UINT64) -
+		(*p_arg)->size;
+	    }
+	  else
+	    avalue[i] = (void *) *(stack - slot);
+	  break;
+
+	default:
+	  FFI_ASSERT(0);
+	}
+
+      slot++;
+      p_arg++;
+    }
+
+  /* Invoke the closure.  */
+  (closure->fun) (cif, rvalue, avalue, closure->user_data);
+
+  debug(3, "after calling function, ret[0] = %08x, ret[1] = %08x\n", ret[0],
+	ret[1]);
+
+  /* Store the result using the lower 2 bytes of the flags.  */
+  switch (cif->flags)
+    {
+    case FFI_TYPE_UINT8:
+      *(stack - FIRST_ARG_SLOT) = (UINT8)(ret[0] >> 24);
+      break;
+    case FFI_TYPE_SINT8:
+      *(stack - FIRST_ARG_SLOT) = (SINT8)(ret[0] >> 24);
+      break;
+    case FFI_TYPE_UINT16:
+      *(stack - FIRST_ARG_SLOT) = (UINT16)(ret[0] >> 16);
+      break;
+    case FFI_TYPE_SINT16:
+      *(stack - FIRST_ARG_SLOT) = (SINT16)(ret[0] >> 16);
+      break;
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_UINT32:
+      *(stack - FIRST_ARG_SLOT) = ret[0];
+      break;
+    case FFI_TYPE_SINT64:
+    case FFI_TYPE_UINT64:
+      *(stack - FIRST_ARG_SLOT) = ret[0];
+      *(stack - FIRST_ARG_SLOT - 1) = ret[1];
+      break;
+
+    case FFI_TYPE_DOUBLE:
+      fldd(rvalue, fr4);
+      break;
+
+    case FFI_TYPE_FLOAT:
+      fldw(rvalue, fr4);
+      break;
+
+    case FFI_TYPE_STRUCT:
+      /* Don't need a return value, done by caller.  */
+      break;
+
+    case FFI_TYPE_SMALL_STRUCT2:
+    case FFI_TYPE_SMALL_STRUCT3:
+    case FFI_TYPE_SMALL_STRUCT4:
+      tmp = (void*)(stack -  FIRST_ARG_SLOT);
+      tmp += 4 - cif->rtype->size;
+      memcpy((void*)tmp, &ret[0], cif->rtype->size);
+      break;
+
+    case FFI_TYPE_SMALL_STRUCT5:
+    case FFI_TYPE_SMALL_STRUCT6:
+    case FFI_TYPE_SMALL_STRUCT7:
+    case FFI_TYPE_SMALL_STRUCT8:
+      {
+	unsigned int ret2[2];
+	int off;
+
+	/* Right justify ret[0] and ret[1] */
+	switch (cif->flags)
+	  {
+	    case FFI_TYPE_SMALL_STRUCT5: off = 3; break;
+	    case FFI_TYPE_SMALL_STRUCT6: off = 2; break;
+	    case FFI_TYPE_SMALL_STRUCT7: off = 1; break;
+	    default: off = 0; break;
+	  }
+
+	memset (ret2, 0, sizeof (ret2));
+	memcpy ((char *)ret2 + off, ret, 8 - off);
+
+	*(stack - FIRST_ARG_SLOT) = ret2[0];
+	*(stack - FIRST_ARG_SLOT - 1) = ret2[1];
+      }
+      break;
+
+    case FFI_TYPE_POINTER:
+    case FFI_TYPE_VOID:
+      break;
+
+    default:
+      debug(0, "assert with cif->flags: %d\n",cif->flags);
+      FFI_ASSERT(0);
+      break;
+    }
+  return FFI_OK;
+}
+
+/* Fill in a closure to refer to the specified fun and user_data.
+   cif specifies the argument and result types for fun.
+   The cif must already be prep'ed.  */
+
+extern void ffi_closure_pa32(void);
+
+ffi_status
+ffi_prep_closure_loc (ffi_closure* closure,
+		      ffi_cif* cif,
+		      void (*fun)(ffi_cif*,void*,void**,void*),
+		      void *user_data,
+		      void *codeloc)
+{
+  UINT32 *tramp = (UINT32 *)(closure->tramp);
+#ifdef PA_HPUX
+  UINT32 *tmp;
+#endif
+
+  FFI_ASSERT (cif->abi == FFI_PA32);
+
+  /* Make a small trampoline that will branch to our
+     handler function. Use PC-relative addressing.  */
+
+#ifdef PA_LINUX
+  tramp[0] = 0xeaa00000; /* b,l .+8,%r21        ; %r21 <- pc+8 */
+  tramp[1] = 0xd6a01c1e; /* depi 0,31,2,%r21    ; mask priv bits */
+  tramp[2] = 0x4aa10028; /* ldw 20(%r21),%r1    ; load plabel */
+  tramp[3] = 0x36b53ff1; /* ldo -8(%r21),%r21   ; get closure addr */
+  tramp[4] = 0x0c201096; /* ldw 0(%r1),%r22     ; address of handler */
+  tramp[5] = 0xeac0c000; /* bv%r0(%r22)         ; branch to handler */
+  tramp[6] = 0x0c281093; /* ldw 4(%r1),%r19     ; GP of handler */
+  tramp[7] = ((UINT32)(ffi_closure_pa32) & ~2);
+
+  /* Flush d/icache -- have to flush up 2 two lines because of
+     alignment.  */
+  __asm__ volatile(
+		   "fdc 0(%0)\n\t"
+		   "fdc %1(%0)\n\t"
+		   "fic 0(%%sr4, %0)\n\t"
+		   "fic %1(%%sr4, %0)\n\t"
+		   "sync\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n"
+		   :
+		   : "r"((unsigned long)tramp & ~31),
+		     "r"(32 /* stride */)
+		   : "memory");
+#endif
+
+#ifdef PA_HPUX
+  tramp[0] = 0xeaa00000; /* b,l .+8,%r21        ; %r21 <- pc+8  */
+  tramp[1] = 0xd6a01c1e; /* depi 0,31,2,%r21    ; mask priv bits  */
+  tramp[2] = 0x4aa10038; /* ldw 28(%r21),%r1    ; load plabel  */
+  tramp[3] = 0x36b53ff1; /* ldo -8(%r21),%r21   ; get closure addr  */
+  tramp[4] = 0x0c201096; /* ldw 0(%r1),%r22     ; address of handler  */
+  tramp[5] = 0x02c010b4; /* ldsid (%r22),%r20   ; load space id  */
+  tramp[6] = 0x00141820; /* mtsp %r20,%sr0      ; into %sr0  */
+  tramp[7] = 0xe2c00000; /* be 0(%sr0,%r22)     ; branch to handler  */
+  tramp[8] = 0x0c281093; /* ldw 4(%r1),%r19     ; GP of handler  */
+  tramp[9] = ((UINT32)(ffi_closure_pa32) & ~2);
+
+  /* Flush d/icache -- have to flush three lines because of alignment.  */
+  __asm__ volatile(
+		   "copy %1,%0\n\t"
+		   "fdc,m %2(%0)\n\t"
+		   "fdc,m %2(%0)\n\t"
+		   "fdc,m %2(%0)\n\t"
+		   "ldsid (%1),%0\n\t"
+		   "mtsp %0,%%sr0\n\t"
+		   "copy %1,%0\n\t"
+		   "fic,m %2(%%sr0,%0)\n\t"
+		   "fic,m %2(%%sr0,%0)\n\t"
+		   "fic,m %2(%%sr0,%0)\n\t"
+		   "sync\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n\t"
+		   "nop\n"
+		   : "=&r" ((unsigned long)tmp)
+		   : "r" ((unsigned long)tramp & ~31),
+		     "r" (32/* stride */)
+		   : "memory");
+#endif
+
+  closure->cif  = cif;
+  closure->user_data = user_data;
+  closure->fun  = fun;
+
+  return FFI_OK;
+}
+#endif