kc3-lang/libffi

diff --git a/Makefile.am b/Makefile.am
index a4a4887..fde7ec2 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -81,7 +81,7 @@ EXTRA_libffi_la_SOURCES = src/aarch64/ffi.c src/aarch64/sysv.S		\
 	src/mips/n32.S src/moxie/ffi.c src/moxie/eabi.S			\
 	src/nios2/ffi.c src/nios2/sysv.S src/or1k/ffi.c			\
 	src/or1k/sysv.S src/pa/ffi.c src/pa/linux.S src/pa/hpux32.S	\
-	src/powerpc/ffi.c src/powerpc/ffi_sysv.c			\
+	src/pa/hpux64.S src/powerpc/ffi.c src/powerpc/ffi_sysv.c	\
 	src/powerpc/ffi_linux64.c src/powerpc/sysv.S			\
 	src/powerpc/linux64.S src/powerpc/linux64_closure.S		\
 	src/powerpc/ppc_closure.S src/powerpc/aix.S			\
diff --git a/README.md b/README.md
index a5b60e8..ddad877 100644
--- a/README.md
+++ b/README.md
@@ -59,6 +59,7 @@ tested:
 | Blackfin        | uClinux          | GCC                     |
 | CSKY            | Linux            | GCC                     |
 | HPPA            | HPUX             | GCC                     |
+| HPPA64          | HPUX             | GCC                     |
 | KVX             | Linux            | GCC                     |
 | IA-64           | Linux            | GCC                     |
 | LoongArch64     | Linux            | GCC                     |
@@ -197,6 +198,10 @@ History
 
 See the git log for details at http://github.com/libffi/libffi.
 
+    TBD - TBD
+        Add support for HPPA64.
+        Many x86 Darwin fixes.
+
     3.4.4 Oct-23-2022
         Important aarch64 fixes, including support for linux builds
           with Link Time Optimization (-flto).
@@ -474,6 +479,7 @@ developers:
     nios ii             Sandra Loosemore
     openrisc            Sebastian Macke
     pa                  Randolph Chung, Dave Anglin, Andreas Tobler
+    pa64                Dave Anglin
     powerpc             Geoffrey Keating, Andreas Tobler,
                         David Edelsohn, John Hornkvist
     powerpc64           Jakub Jelinek
diff --git a/configure.host b/configure.host
index f23716f..a4ca511 100644
--- a/configure.host
+++ b/configure.host
@@ -73,6 +73,7 @@ case "${host}" in
 	;;
   hppa*64-*-hpux*)
 	TARGET=PA64_HPUX; TARGETDIR=pa
+	SOURCES="ffi64.c hpux64.S"
 	;;
   hppa*-*-hpux*)
 	TARGET=PA_HPUX; TARGETDIR=pa
diff --git a/src/pa/ffi64.c b/src/pa/ffi64.c
new file mode 100644
index 0000000..08807c8
--- /dev/null
+++ b/src/pa/ffi64.c
@@ -0,0 +1,614 @@
+/* -----------------------------------------------------------------------
+   ffi64.c - (c) 2022 John David Anglin <dave.anglin@bell.net>
+           
+   HPPA Foreign Function Interface
+   PA 64-Bit ABI support 
+
+   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 FIRST_ARG_SLOT  0
+#define DEBUG_LEVEL   0
+
+#define fldw(addr, fpreg) \
+  __asm__ volatile ("fldw 4(%0), %%" #fpreg "R" : : "r"(addr) : #fpreg)
+#define fstw(fpreg, addr) \
+  __asm__ volatile ("fstw %%" #fpreg "R, 4(%0)" : : "r"(addr))
+#define fldd(addr, fpreg) \
+  __asm__ volatile ("fldd 0(%0), %%" #fpreg "L" : : "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)
+{
+  int sz = t->size;
+
+  /* Small structure results are returned in registers 28 and 29,
+     larger ones are in a buffer allocated by the callee.  The
+     address of the buffer is passed in r28.  The buffer is supposed
+     to be aligned on a 16-byte boundary.   Register return values are
+     padded on the right.  The pad bits on the right are undefined.  */
+
+  if (sz <= 16)
+    return -sz;
+  else
+    return FFI_TYPE_STRUCT;
+}
+
+/* PA has a downward growing stack, which looks like this.  Stack
+   arguments are offset from the argument ponter (AP) in r29.
+
+   Offset
+	[ Fixed args ]
+   AP-64                arg word 0 (r26, fr4)
+   AP-56                arg word 1 (r25, fr5)
+   AP-48                arg word 2 (r24, fr6)
+   AP-40                arg word 3 (r23, fr7)
+   AP-32                arg word 4 (r22, fr8)
+   AP-24                arg word 5 (r21, fr9)
+   AP-16                arg word 6 (r20, fr10)
+   AP-8	                arg word 7 (r19, fr11)
+	[ Variable args; AP = SP-16 if there are no variable args ]
+   AP			stack arg 0
+   AP+8			stack arg 1
+   ...
+	[ Frame marker ]
+   SP-16                RP
+   SP-8                 previous SP
+
+   The first eight 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.  Single and double FP arguments are passed
+   in both general and floating registers when using libffi.
+
+   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 0 -> gr26 or fr4L or fr4R
+   arg word 1 -> gr25 or fr5L or fr5R
+   arg word 2 -> gr24 or fr6L or fr6R
+   arg word 3 -> gr23 or fr7L or fr7R
+   ...
+
+   Single Single-precision floating-point parameters, when passed in
+   floating-point registers, are passed in the right halves of the
+   floating point registers; the left halves are unused.
+
+   Quad-precision floating-point parameters within the first 64 bytes of
+   the parameter list are always passed in general registers.
+
+   The rest of the arguments are passed on the stack starting at AP.
+
+   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_pa64(UINT64 *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;
+  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;
+
+      len = (*p_arg)->size;
+
+      switch (type)
+	{
+	case FFI_TYPE_SINT8:
+	  *(SINT64 *)(stack + slot) = *(SINT8 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT8:
+	  *(UINT64 *)(stack + slot) = *(UINT8 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_SINT16:
+	  *(SINT64 *)(stack + slot) = *(SINT16 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT16:
+	  *(UINT64 *)(stack + slot) = *(UINT16 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_SINT32:
+	  *(SINT64 *)(stack + slot) = *(SINT32 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT32:
+	  *(UINT64 *)(stack + slot) = *(UINT32 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_POINTER:
+	  debug(3, "Storing UINT64 %lu in slot %u\n", *(UINT64 *)(*p_argv),
+		slot);
+	  *(UINT64 *)(stack + slot) = *(UINT64 *)(*p_argv);
+	  break;
+
+	case FFI_TYPE_FLOAT:
+	  /* First 8 args go in fr4L - fr11L.  */
+	  debug(3, "Storing UINT32(float) in slot %u\n", slot);
+	  *(UINT64 *)(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;
+	    case 4: fldw(stack + slot, fr8); break;
+	    case 5: fldw(stack + slot, fr9); break;
+	    case 6: fldw(stack + slot, fr10); break;
+	    case 7: fldw(stack + slot, fr11); break;
+	    }
+	  break;
+
+	case FFI_TYPE_DOUBLE:
+	  debug(3, "Storing UINT64(double) at slot %u\n", slot);
+	  *(UINT64 *)(stack + slot) = *(UINT64 *)(*p_argv);
+	  switch (slot - FIRST_ARG_SLOT)
+	    {
+	    /* First 8 args go in fr4 to fr11.  */
+	    case 0: fldd(stack + slot, fr4); break;
+	    case 1: fldd(stack + slot, fr5); break;
+	    case 2: fldd(stack + slot, fr6); break;
+	    case 3: fldd(stack + slot, fr7); break;
+	    case 4: fldd(stack + slot, fr8); break;
+	    case 5: fldd(stack + slot, fr9); break;
+	    case 6: fldd(stack + slot, fr10); break;
+	    case 7: fldd(stack + slot, fr11); break;
+	    }
+	  break;
+
+#ifdef PA64_HPUX
+	case FFI_TYPE_LONGDOUBLE:
+	  /* Align slot to a 16-byte boundary.  */
+	  slot += (slot & 1);
+	  *(UINT64 *)(stack + slot) = *(UINT64 *)(*p_argv);
+	  *(UINT64 *)(stack + slot + 1) = *(UINT64 *)(*p_argv + 8);
+	  break;
+#endif
+
+	case FFI_TYPE_STRUCT:
+	  /* Structs larger than 8 bytes are aligned on a 16-byte boundary. */
+	  if (len > 8)
+	    slot += (slot & 1);
+	  memcpy((char *)(stack + slot), (char *)*p_argv, len);
+	  break;
+
+	default:
+	  FFI_ASSERT(0);
+	}
+
+      slot += ROUND_UP (len, 8) >> 3;
+      p_arg++;
+      p_argv++;
+    }
+
+  FFI_ASSERT(slot * 8 <= bytes);
+
+  return;
+}
+
+static void ffi_size_stack_pa64(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;
+      int size = (*ptr)->size;
+
+      switch (type)
+	{
+#ifdef PA64_HPUX
+	case FFI_TYPE_LONGDOUBLE:
+	  z += 2 + (z & 1);
+	  break;
+#endif
+
+	case FFI_TYPE_STRUCT:
+	  if (size > 8)
+	    z += (z & 1);
+	  z += ROUND_UP (size, 8) >> 3;
+	  break;
+
+	default: /* 64-bit values */
+	  z++;
+	}
+    }
+
+  /* We need a minimum of 8 argument slots.  Stack must be 16-byte
+     aligned.  */
+  if (z <= 8)
+    z = 8;
+  else
+    z += (z & 1);
+
+  /* Add 16 bytes for frame marker.  */
+  cif->bytes = z * 8 + 64;
+  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 for jump table.  */
+  switch (cif->rtype->type)
+    {
+    case FFI_TYPE_COMPLEX:
+    case FFI_TYPE_STRUCT:
+      /* For the return type we have to check the size of the structures.
+	 If the size is smaller or equal 8 bytes, the result is given back
+	 in one register. If the size is smaller or equal 16 bytes than we
+	 return the result in two registers. If the size is bigger than
+	 16 bytes, the return is in a buffer allocated by the caller.  */
+      cif->flags = ffi_struct_type(cif->rtype);
+      break;
+
+    default:
+      cif->flags = (unsigned) cif->rtype->type;
+      break;
+    }
+
+  /* Lucky us, because of the unique PA ABI we get to do our
+     own stack sizing.  */
+  switch (cif->abi)
+    {
+    case FFI_PA64:
+      ffi_size_stack_pa64(cif);
+      break;
+
+    default:
+      FFI_ASSERT(0);
+      break;
+    }
+
+  return FFI_OK;
+}
+
+extern void ffi_call_pa64(void (*)(UINT64 *, 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
+      && (cif->rtype->type == FFI_TYPE_STRUCT
+	  || cif->rtype->type == FFI_TYPE_COMPLEX)
+      && cif->rtype->size > 16)
+    ecif.rvalue = alloca(ROUND_UP (cif->rtype->size, 16));
+  else
+    ecif.rvalue = rvalue;
+
+
+  switch (cif->abi)
+    {
+    case FFI_PA64:
+      debug(3, "Calling ffi_call_pa64: ecif=%p, bytes=%u, flags=%u, rvalue=%p, fn=%p\n", &ecif, cif->bytes, cif->flags, ecif.rvalue, (void *)fn);
+      ffi_call_pa64(ffi_prep_args_pa64, &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_pa64(ffi_closure *closure, UINT64 *stack)
+{
+  ffi_cif *cif;
+  void **avalue;
+  void *rvalue;
+  /* Functions can return up to 128-bits in registers.  Return address
+     must be double word aligned.  */
+  union { long double rld; UINT64 ret[2]; } u;
+  ffi_type **p_arg;
+  char *tmp;
+  int i, avn;
+  unsigned int slot = FIRST_ARG_SLOT;
+  register UINT64 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 = &u;
+
+  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:
+	  avalue[i] = (void *)(stack + slot) + 7;
+	  break;
+
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT16:
+	  avalue[i] = (void *)(stack + slot) + 6;
+	  break;
+
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_UINT32:
+	  avalue[i] = (void *)(stack + slot) + 4;
+	  break;
+
+	case FFI_TYPE_POINTER:
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_UINT64:
+	  avalue[i] = (void *)(stack + slot);
+	  break;
+
+	case FFI_TYPE_FLOAT:
+	  /* 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;
+	    case 4: fstw(fr8, (void *)(stack + slot)); break;
+	    case 5: fstw(fr9, (void *)(stack + slot)); break;
+	    case 6: fstw(fr10, (void *)(stack + slot)); break;
+	    case 7: fstw(fr11, (void *)(stack + slot)); break;
+	    }
+	  avalue[i] = (void *)(stack + slot) + 4;
+	  break;
+
+	case FFI_TYPE_DOUBLE:
+	  /* See previous comment for FFI_TYPE_FLOAT.  */
+	  switch (slot + FIRST_ARG_SLOT)
+	    {
+	    case 0: fstd(fr4, (void *)(stack + slot)); break;
+	    case 1: fstd(fr5, (void *)(stack + slot)); break;
+	    case 2: fstd(fr6, (void *)(stack + slot)); break;
+	    case 3: fstd(fr7, (void *)(stack + slot)); break;
+	    case 4: fstd(fr8, (void *)(stack + slot)); break;
+	    case 5: fstd(fr9, (void *)(stack + slot)); break;
+	    case 6: fstd(fr10, (void *)(stack + slot)); break;
+	    case 7: fstd(fr11, (void *)(stack + slot)); break;
+	    }
+	  avalue[i] = (void *)(stack + slot);
+	  break;
+
+#ifdef PA64_HPUX
+	case FFI_TYPE_LONGDOUBLE:
+	  /* Long doubles are treated like a big structure.  */
+	  slot += (slot & 1);
+	  avalue[i] = (void *)(stack + slot);
+	  break;
+#endif
+
+	case FFI_TYPE_STRUCT:
+	  /* All structs are passed in registers.  Structs larger
+	     than 8 bytes are aligned on a 16-byte boundary.  */
+	  if((*p_arg)->size > 8)
+	    slot += (slot & 1);
+	  avalue[i] = (void *) (stack + slot);
+	  break;
+
+	default:
+	  FFI_ASSERT(0);
+	}
+
+      slot += (ROUND_UP ((*p_arg)->size, 8) >> 3);
+      p_arg++;
+    }
+
+  /* Invoke the closure.  */
+  (closure->fun) (cif, rvalue, avalue, closure->user_data);
+
+  debug(3, "after calling function, ret[0] = %16lx, ret[1] = %16lx\n", u.ret[0],
+	u.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)u.ret[0];
+      break;
+    case FFI_TYPE_SINT8:
+      *(stack + FIRST_ARG_SLOT) = (SINT8)u.ret[0];
+      break;
+    case FFI_TYPE_UINT16:
+      *(stack + FIRST_ARG_SLOT) = (UINT16)u.ret[0];
+      break;
+    case FFI_TYPE_SINT16:
+      *(stack + FIRST_ARG_SLOT) = (SINT16)u.ret[0];
+      break;
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+      *(stack + FIRST_ARG_SLOT) = (SINT32)u.ret[0];
+      break;
+    case FFI_TYPE_UINT32:
+      *(stack - FIRST_ARG_SLOT) = (UINT32)u.ret[0];
+      break;
+    case FFI_TYPE_SINT64:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_POINTER:
+      *(stack - FIRST_ARG_SLOT) = u.ret[0];
+      break;
+
+    case FFI_TYPE_LONGDOUBLE:
+      *(stack + FIRST_ARG_SLOT) = u.ret[0];
+      *(stack + FIRST_ARG_SLOT + 1) = u.ret[1];
+      break;
+
+    case FFI_TYPE_DOUBLE:
+      fldd(rvalue, fr4);
+      break;
+
+    case FFI_TYPE_FLOAT:
+      /* Adjust for address adjustment in fldw macro.  */
+      fldw(rvalue - 4, fr4);
+      break;
+
+    case FFI_TYPE_STRUCT:
+      /* Don't need a return value, done by caller.  */
+      break;
+
+    case -1:
+    case -2:
+    case -3:
+    case -4:
+    case -5:
+    case -6:
+    case -7:
+    case -8:
+    case -9:
+    case -10:
+    case -11:
+    case -12:
+    case -13:
+    case -14:
+    case -15:
+    case -16:
+      tmp = (void*)(stack + FIRST_ARG_SLOT);
+      memcpy((void*)tmp, &u, cif->rtype->size);
+      break;
+
+    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_pa64(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)
+{
+  /* The layout of a function descriptor.  */
+  struct pa64_fd
+  {
+    UINT64 tmp1;
+    UINT64 tmp2;
+    UINT64 code_pointer;
+    UINT64 gp;
+  };
+
+  struct ffi_pa64_trampoline_struct
+  {
+    UINT64 real_gp;             /* Real gp value.  */
+    UINT64 tmp2;
+    UINT64 code_pointer;        /* Pointer to ffi_closure_unix.  */
+    UINT64 fake_gp;             /* Pointer to closure, installed as gp.  */
+  };
+
+  struct ffi_pa64_trampoline_struct *tramp;
+  struct pa64_fd *fd;
+
+  if (cif->abi != FFI_PA64)
+    return FFI_BAD_ABI;
+
+  /* Get function descriptor address for ffi_closure_pa64.  */
+  fd = (struct pa64_fd *)((UINT64)ffi_closure_pa64);
+
+  /* Setup trampoline.  */
+  tramp = (struct ffi_pa64_trampoline_struct *)closure->tramp;
+  tramp->code_pointer = fd->code_pointer;
+  tramp->fake_gp = (UINT64)codeloc;
+  tramp->real_gp = fd->gp;
+
+  closure->cif  = cif;
+  closure->user_data = user_data;
+  closure->fun  = fun;
+
+  return FFI_OK;
+}
+#endif
diff --git a/src/pa/ffitarget.h b/src/pa/ffitarget.h
index df1209e..6a2c5dc 100644
--- a/src/pa/ffitarget.h
+++ b/src/pa/ffitarget.h
@@ -54,7 +54,6 @@ typedef enum ffi_abi {
 #endif
 
 #ifdef PA64_HPUX
-#error "PA64_HPUX FFI is not yet implemented"
   FFI_PA64,
   FFI_LAST_ABI,
   FFI_DEFAULT_ABI = FFI_PA64
@@ -68,7 +67,11 @@ typedef enum ffi_abi {
 
 #define FFI_CLOSURES 1
 #define FFI_NATIVE_RAW_API 0
+#if defined(PA64_HPUX)
+#define FFI_TRAMPOLINE_SIZE 32
+#else
 #define FFI_TRAMPOLINE_SIZE 12
+#endif
 
 #define FFI_TYPE_SMALL_STRUCT2 -1
 #define FFI_TYPE_SMALL_STRUCT3 -2
diff --git a/src/pa/hpux64.S b/src/pa/hpux64.S
new file mode 100644
index 0000000..6a82b57
--- /dev/null
+++ b/src/pa/hpux64.S
@@ -0,0 +1,681 @@
+/* -----------------------------------------------------------------------
+   hpux64.S - (c) 2005-2022 John David Anglin <dave.anglin@bell.net>
+
+   HPUX PA 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 AUTHOR 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>
+
+	.LEVEL 2.0w
+	.text
+	.align	4
+
+	/* void ffi_call_pa64(void (*)(char *, extended_cif *),
+			       extended_cif *ecif,
+			       unsigned bytes,
+			       unsigned flags,
+			       unsigned *rvalue,
+			       void (*fn)());
+	 */
+
+	.export	ffi_call_pa64,code
+	.import	ffi_prep_args_pa64,code
+
+	.align	4
+
+L$FB1
+ffi_call_pa64
+	.proc
+	.callinfo	FRAME=48,CALLS,SAVE_RP,ENTRY_GR=4
+	.entry
+	std	%rp, -16(%sp)
+	copy	%r3, %r1
+L$CFI11
+	copy	%sp, %r3
+L$CFI12
+	std,ma	%r1, 48(%sp)
+
+	/* Setup the stack for calling prep_args...
+	   We want the stack to look like this:
+
+	   [ Previous stack                            ] <- %r3
+
+	   [ 48-byte register save area                ]
+
+	   [ Stack space for call arguments            ] <- %r4
+
+	   [ 16-byte rame marker                       ]
+
+	   [ 128-byte stack for calling prep_args      ] <- %sp
+	 */
+
+	std	%r4, 8(%r3)		; save r4
+L$CFI13
+	std	%r23, 16(%r3)		; save flags we need it later
+	std	%r22, 24(%r3)		; save rvalue
+	std	%r21, 32(%r3)		; save fn pointer
+
+	copy	%sp, %r4
+	copy	%r4, %r26		; argument stack pointer
+	addl	%r24, %sp, %sp		; allocate argument space
+
+	ldo	112(%sp), %r29		; arg pointer for prep args
+
+	/* Call prep_args:
+	   %arg0(stack) -- set up above to point to call arguments
+	   %arg1(ecif) -- same as incoming param
+	   %arg2(bytes) -- same as incoming param */
+	bl	ffi_prep_args_pa64,%r2
+	ldo	128(%sp), %sp
+	ldo	-128(%sp), %sp
+
+	/* Load the arguments that should be passed in registers
+	   The fp args were loaded by the prep_args function.  */
+	ldd	0(%r4), %r26
+	ldd	8(%r4), %r25
+	ldd	16(%r4), %r24
+	ldd	24(%r4), %r23
+	ldd     32(%r4), %r22
+	ldd     40(%r4), %r21
+	ldd     48(%r4), %r20
+	ldd     56(%r4), %r19
+
+	ldd	24(%r3), %ret0		; %ret0 <- rvalue
+
+	ldd	32(%r3), %r1		; %r1 <- function pointer
+	ldd	16(%r1), %rp		; fn address
+	ldd	24(%r1), %dp		; New gp
+	bve,l	(%rp), %r2		; Call the user function
+	ldo	64(%r4), %r29		; Argument pointer
+
+	/* Prepare to store the result; recover flags and rvalue.  */
+	ldd	16(%r3), %r21		; r21 <- flags
+	extrd,s	%r21, 63, 32, %r21	; sign extend flags for blr
+
+	/* Adjust flags range from [-16, 15] to  [0, 31].  */
+	addi	16, %r21, %r21
+
+	blr	%r21, %r0
+	ldd	24(%r3), %r20		; r20 <- rvalue
+
+	/* Giant jump table */
+	/* 16-byte small struct */
+	b,n	L$smst16
+	nop
+	/* 15-byte small struct */
+	b,n	L$smst15
+	nop
+	/* 14-byte small struct */
+	b,n	L$smst14
+	nop
+	/* 13-byte small struct */
+	b,n	L$smst13
+	nop
+	/* 12-byte small struct */
+	b,n	L$smst12
+	nop
+	/* 11-byte small struct */
+	b,n	L$smst11
+	nop
+	/* 10-byte small struct */
+	b,n	L$smst10
+	nop
+	/* 9-byte small struct */
+	b,n	L$smst9
+	nop
+	/* 8-byte small struct */
+	b,n	L$smst8
+	nop
+	/* 7-byte small struct */
+	b,n	L$smst7
+	nop
+	/* 6-byte small struct */
+	b,n	L$smst6
+	nop
+	/* 5-byte small struct */
+	b,n	L$smst5
+	nop
+	/* 4-byte small struct */
+	b,n	L$smst4
+	nop
+	/* 3-byte small struct */
+	b,n	L$smst3
+	nop
+	/* 2-byte small struct */
+	b,n	L$smst2
+	nop
+	/* 1-byte small struct */
+	b,n	L$smst1
+	nop
+	/* void */
+	b,n	L$done
+	nop
+	/* int */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* float */
+	b	L$done
+	fstw	%fr4R, 0(%r20)
+	/* double */
+	b	L$done
+	fstd	%fr4, 0(%r20)
+	/* long double */
+	b,n	L$longdouble
+	nop
+	/* unsigned int8 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* signed int8 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* unsigned int16 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* signed int16 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* unsigned int32 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* signed int32 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* unsigned int64 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* signed int64 */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* large struct */
+	b,n	L$done
+	nop
+	/* pointer */
+	b	L$done
+	std	%ret0, 0(%r20)
+	/* complex */
+	b,n	L$done
+	nop
+
+L$longdouble
+	std	%ret0, 0(%r20)
+	b	L$done
+	std	%ret1, 8(%r20)
+
+	/* We need to copy byte-by-byte the exact number bytes
+	   in the struct to avoid clobbering other data.  */
+L$smst1
+	extrd,u	%ret0, 7, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst2
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst3
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst4
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst5
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst6
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst7
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst8
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	b	L$done
+	stb	%ret0, 0(%r20)
+
+L$smst9
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst10
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst11
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 23, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst12
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 31, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst13
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 39, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst14
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 47, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst15
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 55, 8, %r22
+	b	L$done
+	stb	%r22, 0(%r20)
+
+L$smst16
+	extrd,u	%ret0, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret0, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb,ma	%ret0, 1(%r20)
+	extrd,u	%ret1, 7, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 15, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 23, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 31, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 39, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 47, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	extrd,u	%ret1, 55, 8, %r22
+	stb,ma	%r22, 1(%r20)
+	stb	%ret1, 0(%r20)
+
+L$done
+	/* all done, restore registers and return */
+	copy	%r4, %sp
+	ldd	8(%r3), %r4
+	ldd	-16(%r3), %rp
+	bve	(%rp)
+	ldd,mb	-48(%sp), %r3
+	.exit
+	.procend
+L$FE1
+	.size	ffi_call_pa64, .-ffi_call_pa64
+
+	/* void ffi_closure_pa64(void);
+	   Called with closure argument in %r21 */
+
+	.export ffi_closure_pa64,code
+	.import ffi_closure_inner_pa64,code
+	.align 4
+L$FB2
+ffi_closure_pa64
+	.proc
+	.callinfo FRAME=128,CALLS,SAVE_RP,ENTRY_GR=3
+	.entry
+
+	std	%rp, -16(%sp)
+	copy	%r3, %r1
+L$CFI21
+	copy	%sp, %r3
+L$CFI22
+	std,ma	%r1, 128(%sp)
+L$CFI23
+
+	/* Put arguments onto the stack and call ffi_closure_inner.  */
+	std	%r26, -64(%r29)
+	std	%r25, -56(%r29)
+	std	%r24, -48(%r29)
+	std	%r23, -40(%r29)
+	std	%r22, -32(%r29)
+	std	%r21, -24(%r29)
+	std	%r20, -16(%r29)
+	std	%r19, -8(%r29)
+
+	/* Load and save start of argument stack.  */
+	ldo	-64(%r29), %r25
+	std	%r25, 8(%r3)
+
+	/* Setup arg pointer.  */
+	ldo	-16(%sp), %ret1
+
+	/* Retrieve closure pointer and real gp.  */
+	copy	%dp, %r26
+	bl	ffi_closure_inner_pa64, %r2
+	ldd	0(%dp), %dp
+
+	/* Retrieve start of argument stack.  */
+	ldd	8(%r3), %r1
+
+	/* Restore r3 and op stack.  */
+	ldd,mb	-128(%sp), %r3
+
+	/* Load return address.  */
+	ldd	-16(%sp), %rp
+
+	/* Load return values from first and second stack slots.  */
+	ldd	0(%r1), %ret0
+	bve	(%rp)
+	ldd	8(%r1), %ret1
+
+	.exit
+	.procend
+	.end
+L$FE2:
+	.size	ffi_closure_pa64, .-ffi_closure_pa64
+
+	.section	.eh_frame,"aw",@progbits
+L$frame1:
+	.word   L$ECIE1-L$SCIE1 ;# Length of Common Information Entry
+L$SCIE1:
+	.word   0x0     ;# CIE Identifier Tag
+	.byte   0x3     ;# CIE Version
+	.stringz ""     ;# CIE Augmentation
+	.uleb128 0x1    ;# CIE Code Alignment Factor
+	.sleb128 8      ;# CIE Data Alignment Factor
+	.byte   0x2     ;# CIE RA Column
+	.byte   0xc     ;# DW_CFA_def_cfa
+	.uleb128 0x1e
+	.uleb128 0x0
+	.align 8
+L$ECIE1:
+L$SFDE1:
+	.word   L$EFDE1-L$ASFDE1        ;# FDE Length
+L$ASFDE1:
+	.word   L$ASFDE1-L$frame1       ;# FDE CIE offset
+	.dword  L$FB1	;# FDE initial location
+	.dword  L$FE1-L$FB1	;# FDE address range
+
+	.byte   0x4	;# DW_CFA_advance_loc4
+	.word   L$CFI11-L$FB1
+	.byte	0x9	;# DW_CFA_register: r3 in r1
+	.uleb128 0x3
+	.uleb128 0x1
+	.byte   0x11    ;# DW_CFA_offset_extended_sf: r2 at cfa-16
+	.uleb128 0x2
+	.sleb128 -2
+	.byte   0x4     ;# DW_CFA_advance_loc4
+	.word   L$CFI12-L$CFI11
+	.byte   0xd     ;# DW_CFA_def_cfa_register: r3
+	.uleb128 0x3
+
+	.byte   0x4     ;# DW_CFA_advance_loc4
+	.word   L$CFI13-L$CFI12
+	.byte	0x83	;# DW_CFA_offset: r3 at cfa+0
+	.uleb128 0
+	.byte	0x84	;# DW_CFA_offset: r4 at cfa+8
+	.uleb128 1
+
+	.align 8
+L$EFDE1:
+
+L$SFDE2:
+	.word   L$EFDE2-L$ASFDE2        ;# FDE Length
+L$ASFDE2:
+	.word   L$ASFDE2-L$frame1       ;# FDE CIE offset
+	.dword   L$FB2	;# FDE initial location
+	.dword   L$FE2-L$FB2     ;# FDE address range
+	.byte   0x4	;# DW_CFA_advance_loc4
+	.word   L$CFI21-L$FB2
+	.byte   0x9	;# DW_CFA_register: r3 in r1
+	.uleb128 0x3
+	.uleb128 0x1
+	.byte   0x11    ;# DW_CFA_offset_extended_sf: r2 at cfa-16
+	.uleb128 0x2
+	.sleb128 -2
+
+	.byte   0x4     ;# DW_CFA_advance_loc4
+	.word   L$CFI22-L$CFI21
+	.byte   0xd     ;# DW_CFA_def_cfa_register: r3
+	.uleb128 0x3
+
+	.byte   0x4     ;# DW_CFA_advance_loc4
+	.word   L$CFI23-L$CFI22
+	.byte	0x83	;# DW_CFA_offset: r3 at cfa+0
+	.uleb128 0
+
+	.align 8
+L$EFDE2: