Finally, this adds _CALL_ELF == 2 support. ELFv1 objects can't be

linked with ELFv2 objects, so this is one case where preprocessor
tests in ffi.c are fine.  Also, there is no need to define a new
FFI_ELFv2 or somesuch value in enum ffi_abi.  FFI_LINUX64 will happily
serve both ABIs.

ChangeLog

@@ -1,3 +1,32 @@
+2013-11-16  Alan Modra  <amodra@gmail.com>
+
+	* src/powerpc/ffi.c (ffi_prep_cif_machdep_core): Use #if _CALL_ELF
+	test to select parameter save sizing for ELFv2 vs. ELFv1.
+	* src/powerpc/ffitarget.h (FFI_V2_TYPE_FLOAT_HOMOG,
+	FFI_V2_TYPE_DOUBLE_HOMOG, FFI_V2_TYPE_SMALL_STRUCT): Define.
+	(FFI_TRAMPOLINE_SIZE): Define variant for ELFv2.
+	* src/powerpc/ffi.c (FLAG_ARG_NEEDS_PSAVE): Define.
+	(discover_homogeneous_aggregate): New function.
+	(ffi_prep_args64): Adjust start of param save area for ELFv2.
+	Handle homogenous floating point struct parms.
+	(ffi_prep_cif_machdep_core): Adjust space calculation for ELFv2.
+	Handle ELFv2 return values.  Set FLAG_ARG_NEEDS_PSAVE.  Handle
+	homogenous floating point structs.
+	(ffi_call): Increase size of smst_buffer for ELFv2.  Handle ELFv2.
+	(flush_icache): Compile for ELFv2.
+	(ffi_prep_closure_loc): Set up ELFv2 trampoline.
+	(ffi_closure_helper_LINUX64): Don't return all structs directly
+	to caller.  Handle homogenous floating point structs.  Handle
+	ELFv2 struct return values.
+	* src/powerpc/linux64.S (ffi_call_LINUX64): Set up r2 for
+	ELFv2.  Adjust toc save location.  Call function pointer using
+	r12.  Handle FLAG_RETURNS_SMST.  Don't predict branches.
+	* src/powerpc/linux64_closure.S (ffi_closure_LINUX64): Set up r2
+	for ELFv2.  Define ELFv2 versions of STACKFRAME, PARMSAVE, and
+	RETVAL.  Handle possibly missing parameter save area.  Handle
+	ELFv2 return values.
+	(.note.GNU-stack): Move inside outer #ifdef.
+
 2013-11-16  Alan Modra  <amodra@gmail.com>
 
 	* src/powerpc/ffi.c (ffi_prep_cif_machdep): Revert 2013-02-08

src/powerpc/ffi.c

@@ -49,6 +49,7 @@ enum {
   FLAG_RETURNS_128BITS  = 1 << (31-27), /* cr6  */
 
   FLAG_ARG_NEEDS_COPY   = 1 << (31- 7),
+  FLAG_ARG_NEEDS_PSAVE  = FLAG_ARG_NEEDS_COPY, /* Used by ELFv2 */
 #ifndef __NO_FPRS__
   FLAG_FP_ARGUMENTS     = 1 << (31- 6), /* cr1.eq; specified by ABI */
 #endif
@@ -389,6 +390,45 @@ enum {
 };
 enum { ASM_NEEDS_REGISTERS64 = 4 };
 
+#if _CALL_ELF == 2
+static unsigned int
+discover_homogeneous_aggregate (const ffi_type *t, unsigned int *elnum)
+{
+  switch (t->type)
+    {
+    case FFI_TYPE_FLOAT:
+    case FFI_TYPE_DOUBLE:
+      *elnum = 1;
+      return (int) t->type;
+
+    case FFI_TYPE_STRUCT:;
+      {
+	unsigned int base_elt = 0, total_elnum = 0;
+	ffi_type **el = t->elements;
+	while (*el)
+	  {
+	    unsigned int el_elt, el_elnum = 0;
+	    el_elt = discover_homogeneous_aggregate (*el, &el_elnum);
+	    if (el_elt == 0
+		|| (base_elt && base_elt != el_elt))
+	      return 0;
+	    base_elt = el_elt;
+	    total_elnum += el_elnum;
+	    if (total_elnum > 8)
+	      return 0;
+	    el++;
+	  }
+	*elnum = total_elnum;
+	return base_elt;
+      }
+
+    default:
+      return 0;
+    }
+}
+#endif
+
+
 /* ffi_prep_args64 is called by the assembly routine once stack space
    has been allocated for the function's arguments.
 
@@ -476,7 +516,11 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
   stacktop.c = (char *) stack + bytes;
   gpr_base.ul = stacktop.ul - ASM_NEEDS_REGISTERS64 - NUM_GPR_ARG_REGISTERS64;
   gpr_end.ul = gpr_base.ul + NUM_GPR_ARG_REGISTERS64;
+#if _CALL_ELF == 2
+  rest.ul = stack + 4 + NUM_GPR_ARG_REGISTERS64;
+#else
   rest.ul = stack + 6 + NUM_GPR_ARG_REGISTERS64;
+#endif
   fpr_base.d = gpr_base.d - NUM_FPR_ARG_REGISTERS64;
   fparg_count = 0;
   next_arg.ul = gpr_base.ul;
@@ -498,6 +542,8 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
        i < nargs;
        i++, ptr++, p_argv.v++)
     {
+      unsigned int elt, elnum;
+
       switch ((*ptr)->type)
 	{
 	case FFI_TYPE_FLOAT:
@@ -555,6 +601,56 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
 	  if (align > 1)
 	    next_arg.p = ALIGN (next_arg.p, align);
 #endif
+	  elt = 0;
+#if _CALL_ELF == 2
+	  elt = discover_homogeneous_aggregate (*ptr, &elnum);
+#endif
+	  if (elt)
+	    {
+	      union {
+		void *v;
+		float *f;
+		double *d;
+	      } arg;
+
+	      arg.v = *p_argv.v;
+	      if (elt == FFI_TYPE_FLOAT)
+		{
+		  do
+		    {
+		      double_tmp = *arg.f++;
+		      if (fparg_count < NUM_FPR_ARG_REGISTERS64
+			  && i < nfixedargs)
+			*fpr_base.d++ = double_tmp;
+		      else
+			*next_arg.f = (float) double_tmp;
+		      if (++next_arg.f == gpr_end.f)
+			next_arg.f = rest.f;
+		      fparg_count++;
+		    }
+		  while (--elnum != 0);
+		  if ((next_arg.p & 3) != 0)
+		    {
+		      if (++next_arg.f == gpr_end.f)
+			next_arg.f = rest.f;
+		    }
+		}
+	      else
+		do
+		  {
+		    double_tmp = *arg.d++;
+		    if (fparg_count < NUM_FPR_ARG_REGISTERS64 && i < nfixedargs)
+		      *fpr_base.d++ = double_tmp;
+		    else
+		      *next_arg.d = double_tmp;
+		    if (++next_arg.d == gpr_end.d)
+		      next_arg.d = rest.d;
+		    fparg_count++;
+		  }
+		while (--elnum != 0);
+	    }
+	  else
+	    {
 	      words = ((*ptr)->size + 7) / 8;
 	      if (next_arg.ul >= gpr_base.ul && next_arg.ul + words > gpr_end.ul)
 		{
@@ -578,6 +674,7 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
 		  if (next_arg.ul == gpr_end.ul)
 		    next_arg.ul = rest.ul;
 		}
+	    }
 	  break;
 
 	case FFI_TYPE_UINT8:
@@ -632,11 +729,10 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
   unsigned type = cif->rtype->type;
   unsigned size = cif->rtype->size;
 
+  /* The machine-independent calculation of cif->bytes doesn't work
+     for us.  Redo the calculation.  */
   if (cif->abi != FFI_LINUX64)
     {
-      /* All the machine-independent calculation of cif->bytes will be wrong.
-	 Redo the calculation for SYSV.  */
-
       /* Space for the frame pointer, callee's LR, and the asm's temp regs.  */
       bytes = (2 + ASM_NEEDS_REGISTERS) * sizeof (int);
 
@@ -646,13 +742,20 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
   else
     {
       /* 64-bit ABI.  */
+#if _CALL_ELF == 2
+      /* Space for backchain, CR, LR, TOC and the asm's temp regs.  */
+      bytes = (4 + ASM_NEEDS_REGISTERS64) * sizeof (long);
 
+      /* Space for the general registers.  */
+      bytes += NUM_GPR_ARG_REGISTERS64 * sizeof (long);
+#else
       /* Space for backchain, CR, LR, cc/ld doubleword, TOC and the asm's temp
 	 regs.  */
       bytes = (6 + ASM_NEEDS_REGISTERS64) * sizeof (long);
 
       /* Space for the mandatory parm save area and general registers.  */
       bytes += 2 * NUM_GPR_ARG_REGISTERS64 * sizeof (long);
+#endif
     }
 
   /* Return value handling.  The rules for SYSV are as follows:
@@ -730,6 +833,25 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
 	  flags |= FLAG_RETURNS_SMST;
 	  break;
 	}
+#if _CALL_ELF == 2
+      if (cif->abi == FFI_LINUX64)
+	{
+	  unsigned int elt, elnum;
+	  elt = discover_homogeneous_aggregate (cif->rtype, &elnum);
+	  if (elt)
+	    {
+	      if (elt == FFI_TYPE_DOUBLE)
+		flags |= FLAG_RETURNS_64BITS;
+	      flags |= FLAG_RETURNS_FP | FLAG_RETURNS_SMST;
+	      break;
+	    }
+	  if (size <= 16)
+	    {
+	      flags |= FLAG_RETURNS_SMST;
+	      break;
+	    }
+	}
+#endif
       intarg_count++;
       flags |= FLAG_RETVAL_REFERENCE;
       /* Fall through.  */
@@ -845,6 +967,7 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
   else
     for (ptr = cif->arg_types, i = cif->nargs; i > 0; i--, ptr++)
       {
+	unsigned int elt, elnum;
 #ifdef __STRUCT_PARM_ALIGN__
 	unsigned int align;
 #endif
@@ -855,12 +978,16 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
 	  case FFI_TYPE_LONGDOUBLE:
 	    fparg_count += 2;
 	    intarg_count += 2;
+	    if (fparg_count > NUM_FPR_ARG_REGISTERS)
+	      flags |= FLAG_ARG_NEEDS_PSAVE;
 	    break;
 #endif
 	  case FFI_TYPE_FLOAT:
 	  case FFI_TYPE_DOUBLE:
 	    fparg_count++;
 	    intarg_count++;
+	    if (fparg_count > NUM_FPR_ARG_REGISTERS)
+	      flags |= FLAG_ARG_NEEDS_PSAVE;
 	    break;
 
 	  case FFI_TYPE_STRUCT:
@@ -873,6 +1000,21 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
 	      intarg_count = ALIGN (intarg_count, align);
 #endif
 	    intarg_count += ((*ptr)->size + 7) / 8;
+	    elt = 0;
+#if _CALL_ELF == 2
+	    elt = discover_homogeneous_aggregate (*ptr, &elnum);
+#endif
+	    if (elt)
+	      {
+		fparg_count += elnum;
+		if (fparg_count > NUM_FPR_ARG_REGISTERS)
+		  flags |= FLAG_ARG_NEEDS_PSAVE;
+	      }
+	    else
+	      {
+		if (intarg_count > NUM_GPR_ARG_REGISTERS)
+		  flags |= FLAG_ARG_NEEDS_PSAVE;
+	      }
 	    break;
 
 	  case FFI_TYPE_POINTER:
@@ -888,6 +1030,8 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
 	    /* Everything else is passed as a 8-byte word in a GPR, either
 	       the object itself or a pointer to it.  */
 	    intarg_count++;
+	    if (intarg_count > NUM_GPR_ARG_REGISTERS)
+	      flags |= FLAG_ARG_NEEDS_PSAVE;
 	    break;
 	  default:
 	    FFI_ASSERT (0);
@@ -928,8 +1072,13 @@ ffi_prep_cif_machdep_core (ffi_cif *cif)
 #endif
 
       /* Stack space.  */
+#if _CALL_ELF == 2
+      if ((flags & FLAG_ARG_NEEDS_PSAVE) != 0)
+	bytes += intarg_count * sizeof (long);
+#else
       if (intarg_count > NUM_GPR_ARG_REGISTERS64)
 	bytes += (intarg_count - NUM_GPR_ARG_REGISTERS64) * sizeof (long);
+#endif
     }
 
   /* The stack space allocated needs to be a multiple of 16 bytes.  */
@@ -957,6 +1106,10 @@ ffi_prep_cif_machdep_var (ffi_cif *cif,
 			  unsigned int ntotalargs MAYBE_UNUSED)
 {
   cif->nfixedargs = nfixedargs;
+#if _CALL_ELF == 2
+  if (cif->abi == FFI_LINUX64)
+    cif->flags |= FLAG_ARG_NEEDS_PSAVE;
+#endif
   return ffi_prep_cif_machdep_core (cif);
 }
 
@@ -976,8 +1129,11 @@ ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
    *
    * We bounce-buffer SYSV small struct return values so that sysv.S
    * can write r3 and r4 to memory without worrying about struct size.
+   *
+   * For ELFv2 ABI, use a bounce buffer for homogeneous structs too,
+   * for similar reasons.
    */
-  unsigned int smst_buffer[2];
+  unsigned long smst_buffer[8];
   extended_cif ecif;
 
   ecif.cif = cif;
@@ -1019,10 +1175,11 @@ ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 #ifndef __LITTLE_ENDIAN__
       /* The SYSV ABI returns a structure of up to 4 bytes in size
 	 left-padded in r3.  */
-      if (rsize <= 4)
+      if (cif->abi == FFI_SYSV && rsize <= 4)
 	memcpy (rvalue, (char *) smst_buffer + 4 - rsize, rsize);
       /* The SYSV ABI returns a structure of up to 8 bytes in size
-	 left-padded in r3/r4.  */
+	 left-padded in r3/r4, and the ELFv2 ABI similarly returns a
+	 structure of up to 8 bytes in size left-padded in r3.  */
       else if (rsize <= 8)
 	memcpy (rvalue, (char *) smst_buffer + 8 - rsize, rsize);
       else
@@ -1032,7 +1189,7 @@ ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 }
 
 
-#ifndef POWERPC64
+#if !defined POWERPC64 || _CALL_ELF == 2
 #define MIN_CACHE_LINE_SIZE 8
 
 static void
@@ -1056,6 +1213,22 @@ ffi_prep_closure_loc (ffi_closure *closure,
 		      void *codeloc)
 {
 #ifdef POWERPC64
+# if _CALL_ELF == 2
+  unsigned int *tramp = (unsigned int *) &closure->tramp[0];
+
+  if (cif->abi != FFI_LINUX64)
+    return FFI_BAD_ABI;
+
+  tramp[0] = 0xe96c0018;	/* 0:	ld	11,2f-0b(12)	*/
+  tramp[1] = 0xe98c0010;	/*	ld	12,1f-0b(12)	*/
+  tramp[2] = 0x7d8903a6;	/*	mtctr	12		*/
+  tramp[3] = 0x4e800420;	/*	bctr			*/
+				/* 1:	.quad	function_addr	*/
+				/* 2:	.quad	context		*/
+  *(void **) &tramp[4] = (void *) ffi_closure_LINUX64;
+  *(void **) &tramp[6] = codeloc;
+  flush_icache ((char *)tramp, (char *)codeloc, FFI_TRAMPOLINE_SIZE);
+# else
   void **tramp = (void **) &closure->tramp[0];
 
   if (cif->abi != FFI_LINUX64)
@@ -1063,6 +1236,7 @@ ffi_prep_closure_loc (ffi_closure *closure,
   /* Copy function address and TOC from ffi_closure_LINUX64.  */
   memcpy (tramp, (char *) ffi_closure_LINUX64, 16);
   tramp[2] = codeloc;
+# endif
 #else
   unsigned int *tramp;
 
@@ -1419,9 +1593,10 @@ ffi_closure_helper_LINUX64 (ffi_closure *closure, void *rvalue,
   cif = closure->cif;
   avalue = alloca (cif->nargs * sizeof (void *));
 
-  /* Copy the caller's structure return value address so that the closure
+  /* Copy the caller's structure return value address so that the
-     returns the data directly to the caller.  */
+     closure returns the data directly to the caller.  */
-  if (cif->rtype->type == FFI_TYPE_STRUCT)
+  if (cif->rtype->type == FFI_TYPE_STRUCT
+      && (cif->flags & FLAG_RETURNS_SMST) == 0)
     {
       rvalue = (void *) *pst;
       pst++;
@@ -1435,6 +1610,8 @@ ffi_closure_helper_LINUX64 (ffi_closure *closure, void *rvalue,
   /* Grab the addresses of the arguments from the stack frame.  */
   while (i < avn)
     {
+      unsigned int elt, elnum;
+
       switch (arg_types[i]->type)
 	{
 	case FFI_TYPE_SINT8:
@@ -1476,6 +1653,66 @@ ffi_closure_helper_LINUX64 (ffi_closure *closure, void *rvalue,
 	  if (align > 1)
 	    pst = (unsigned long *) ALIGN ((size_t) pst, align);
 #endif
+	  elt = 0;
+#if _CALL_ELF == 2
+	  elt = discover_homogeneous_aggregate (arg_types[i], &elnum);
+#endif
+	  if (elt)
+	    {
+	      union {
+		void *v;
+		unsigned long *ul;
+		float *f;
+		double *d;
+		size_t p;
+	      } to, from;
+
+	      /* Repackage the aggregate from its parts.  The
+		 aggregate size is not greater than the space taken by
+		 the registers so store back to the register/parameter
+		 save arrays.  */
+	      if (pfr + elnum <= end_pfr)
+		to.v = pfr;
+	      else
+		to.v = pst;
+
+	      avalue[i] = to.v;
+	      from.ul = pst;
+	      if (elt == FFI_TYPE_FLOAT)
+		{
+		  do
+		    {
+		      if (pfr < end_pfr && i < nfixedargs)
+			{
+			  *to.f = (float) pfr->d;
+			  pfr++;
+			}
+		      else
+			*to.f = *from.f;
+		      to.f++;
+		      from.f++;
+		    }
+		  while (--elnum != 0);
+		}
+	      else
+		{
+		  do
+		    {
+		      if (pfr < end_pfr && i < nfixedargs)
+			{
+			  *to.d = pfr->d;
+			  pfr++;
+			}
+		      else
+			*to.d = *from.d;
+		      to.d++;
+		      from.d++;
+		    }
+		  while (--elnum != 0);
+		}
+	    }
+	  else
+	    {
 #ifndef __LITTLE_ENDIAN__
 	      /* Structures with size less than eight bytes are passed
 		 left-padded.  */
@@ -1484,6 +1721,7 @@ ffi_closure_helper_LINUX64 (ffi_closure *closure, void *rvalue,
 	      else
 #endif
 		avalue[i] = pst;
+	    }
 	  pst += (arg_types[i]->size + 7) / 8;
 	  break;
 
@@ -1554,5 +1792,14 @@ ffi_closure_helper_LINUX64 (ffi_closure *closure, void *rvalue,
   (closure->fun) (cif, rvalue, avalue, closure->user_data);
 
   /* Tell ffi_closure_LINUX64 how to perform return type promotions.  */
+  if ((cif->flags & FLAG_RETURNS_SMST) != 0)
+    {
+      if ((cif->flags & FLAG_RETURNS_FP) == 0)
+	return FFI_V2_TYPE_SMALL_STRUCT + cif->rtype->size - 1;
+      else if ((cif->flags & FLAG_RETURNS_64BITS) != 0)
+	return FFI_V2_TYPE_DOUBLE_HOMOG;
+      else
+	return FFI_V2_TYPE_FLOAT_HOMOG;
+    }
   return cif->rtype->type;
 }

src/powerpc/ffitarget.h

@@ -122,14 +122,23 @@ typedef enum ffi_abi {
    defined in ffi.c, to determine the exact return type and its size.  */
 #define FFI_SYSV_TYPE_SMALL_STRUCT (FFI_TYPE_LAST + 2)
 
-#if defined(POWERPC64) || defined(POWERPC_AIX)
+/* Used by ELFv2 for homogenous structure returns.  */
+#define FFI_V2_TYPE_FLOAT_HOMOG		(FFI_TYPE_LAST + 1)
+#define FFI_V2_TYPE_DOUBLE_HOMOG	(FFI_TYPE_LAST + 2)
+#define FFI_V2_TYPE_SMALL_STRUCT	(FFI_TYPE_LAST + 3)
+
+#if _CALL_ELF == 2
+# define FFI_TRAMPOLINE_SIZE 32
+#else
+# if defined(POWERPC64) || defined(POWERPC_AIX)
 #  if defined(POWERPC_DARWIN64)
 #    define FFI_TRAMPOLINE_SIZE 48
 #  else
 #    define FFI_TRAMPOLINE_SIZE 24
 #  endif
-#else /* POWERPC || POWERPC_AIX */
+# else /* POWERPC || POWERPC_AIX */
 #  define FFI_TRAMPOLINE_SIZE 40
+# endif
 #endif
 
 #ifndef LIBFFI_ASM

src/powerpc/linux64.S

@@ -32,15 +32,22 @@
 #ifdef __powerpc64__
 	.hidden	ffi_call_LINUX64
 	.globl	ffi_call_LINUX64
+# if _CALL_ELF == 2
+	.text
+ffi_call_LINUX64:
+	addis	%r2, %r12, .TOC.-ffi_call_LINUX64@ha
+	addi	%r2, %r2, .TOC.-ffi_call_LINUX64@l
+	.localentry ffi_call_LINUX64, . - ffi_call_LINUX64
+# else
 	.section	".opd","aw"
 	.align	3
 ffi_call_LINUX64:
-#ifdef _CALL_LINUX
+#  ifdef _CALL_LINUX
 	.quad	.L.ffi_call_LINUX64,.TOC.@tocbase,0
 	.type	ffi_call_LINUX64,@function
 	.text
 .L.ffi_call_LINUX64:
-#else
+#  else
 	.hidden	.ffi_call_LINUX64
 	.globl	.ffi_call_LINUX64
 	.quad	.ffi_call_LINUX64,.TOC.@tocbase,0
@@ -48,7 +55,8 @@ ffi_call_LINUX64:
 	.type	.ffi_call_LINUX64,@function
 	.text
 .ffi_call_LINUX64:
-#endif
+#  endif
+# endif
 .LFB1:
 	mflr	%r0
 	std	%r28, -32(%r1)
@@ -63,26 +71,35 @@ ffi_call_LINUX64:
 	mr	%r31, %r5	/* flags, */
 	mr	%r30, %r6	/* rvalue, */
 	mr	%r29, %r7	/* function address.  */
+/* Save toc pointer, not for the ffi_prep_args64 call, but for the later
+   bctrl function call.  */
+# if _CALL_ELF == 2
+	std	%r2, 24(%r1)
+# else
 	std	%r2, 40(%r1)
+# endif
 
 	/* Call ffi_prep_args64.  */
 	mr	%r4, %r1
-#ifdef _CALL_LINUX
+# if defined _CALL_LINUX || _CALL_ELF == 2
 	bl	ffi_prep_args64
-#else
+# else
 	bl	.ffi_prep_args64
-#endif
+# endif
 
-	ld	%r0, 0(%r29)
+# if _CALL_ELF == 2
+	mr	%r12, %r29
+# else
+	ld	%r12, 0(%r29)
 	ld	%r2, 8(%r29)
 	ld	%r11, 16(%r29)
-
+# endif
 	/* Now do the call.  */
 	/* Set up cr1 with bits 4-7 of the flags.  */
 	mtcrf	0x40, %r31
 
 	/* Get the address to call into CTR.  */
-	mtctr	%r0
+	mtctr	%r12
 	/* Load all those argument registers.  */
 	ld	%r3, -32-(8*8)(%r28)
 	ld	%r4, -32-(7*8)(%r28)
@@ -117,12 +134,17 @@ ffi_call_LINUX64:
 
 	/* This must follow the call immediately, the unwinder
 	   uses this to find out if r2 has been saved or not.  */
+# if _CALL_ELF == 2
+	ld	%r2, 24(%r1)
+# else
 	ld	%r2, 40(%r1)
+# endif
 
 	/* Now, deal with the return value.  */
 	mtcrf	0x01, %r31
-	bt-	30, .Ldone_return_value
+	bt	31, .Lstruct_return_value
-	bt-	29, .Lfp_return_value
+	bt	30, .Ldone_return_value
+	bt	29, .Lfp_return_value
 	std	%r3, 0(%r30)
 	/* Fall through...  */
 
@@ -147,14 +169,48 @@ ffi_call_LINUX64:
 .Lfloat_return_value:
 	stfs	%f1, 0(%r30)
 	b	.Ldone_return_value
+
+.Lstruct_return_value:
+	bf	29, .Lsmall_struct
+	bf	28, .Lfloat_homog_return_value
+	stfd	%f1, 0(%r30)
+	stfd	%f2, 8(%r30)
+	stfd	%f3, 16(%r30)
+	stfd	%f4, 24(%r30)
+	stfd	%f5, 32(%r30)
+	stfd	%f6, 40(%r30)
+	stfd	%f7, 48(%r30)
+	stfd	%f8, 56(%r30)
+	b	.Ldone_return_value
+
+.Lfloat_homog_return_value:
+	stfs	%f1, 0(%r30)
+	stfs	%f2, 4(%r30)
+	stfs	%f3, 8(%r30)
+	stfs	%f4, 12(%r30)
+	stfs	%f5, 16(%r30)
+	stfs	%f6, 20(%r30)
+	stfs	%f7, 24(%r30)
+	stfs	%f8, 28(%r30)
+	b	.Ldone_return_value
+
+.Lsmall_struct:
+	std	%r3, 0(%r30)
+	std	%r4, 8(%r30)
+	b	.Ldone_return_value
+
 .LFE1:
 	.long	0
 	.byte	0,12,0,1,128,4,0,0
-#ifdef _CALL_LINUX
+# if _CALL_ELF == 2
+	.size	ffi_call_LINUX64,.-ffi_call_LINUX64
+# else
+#  ifdef _CALL_LINUX
 	.size	ffi_call_LINUX64,.-.L.ffi_call_LINUX64
-#else
+#  else
 	.size	.ffi_call_LINUX64,.-.ffi_call_LINUX64
-#endif
+#  endif
+# endif
 
 	.section	.eh_frame,EH_FRAME_FLAGS,@progbits
 .Lframe1:

src/powerpc/linux64_closure.S

@@ -33,15 +33,22 @@
 #ifdef __powerpc64__
 	FFI_HIDDEN (ffi_closure_LINUX64)
 	.globl  ffi_closure_LINUX64
+# if _CALL_ELF == 2
+	.text
+ffi_closure_LINUX64:
+	addis	%r2, %r12, .TOC.-ffi_closure_LINUX64@ha
+	addi	%r2, %r2, .TOC.-ffi_closure_LINUX64@l
+	.localentry ffi_closure_LINUX64, . - ffi_closure_LINUX64
+# else
 	.section        ".opd","aw"
 	.align  3
 ffi_closure_LINUX64:
-#ifdef _CALL_LINUX
+#  ifdef _CALL_LINUX
 	.quad   .L.ffi_closure_LINUX64,.TOC.@tocbase,0
 	.type   ffi_closure_LINUX64,@function
 	.text
 .L.ffi_closure_LINUX64:
-#else
+#  else
 	FFI_HIDDEN (.ffi_closure_LINUX64)
 	.globl  .ffi_closure_LINUX64
 	.quad   .ffi_closure_LINUX64,.TOC.@tocbase,0
@@ -49,15 +56,52 @@ ffi_closure_LINUX64:
 	.type   .ffi_closure_LINUX64,@function
 	.text
 .ffi_closure_LINUX64:
-#endif
+#  endif
+# endif
 
+# if _CALL_ELF == 2
+#  32 byte special reg save area + 64 byte parm save area and retval
+#  + 13*8 fpr save area + round to 16
+#  define STACKFRAME 208
+#  define PARMSAVE 32
+#  No parameter save area is needed for the call to ffi_closure_helper_LINUX64,
+#  so return value can start there.
+#  define RETVAL PARMSAVE
+# else
 #  48 bytes special reg save area + 64 bytes parm save area
 #  + 16 bytes retval area + 13*8 bytes fpr save area + round to 16
 #  define STACKFRAME 240
 #  define PARMSAVE 48
 #  define RETVAL PARMSAVE+64
+# endif
 
 .LFB1:
+# if _CALL_ELF == 2
+	ld	%r12, FFI_TRAMPOLINE_SIZE(%r11)		# closure->cif
+	mflr	%r0
+	lwz	%r12, 28(%r12)				# cif->flags
+	mtcrf	0x40, %r12
+	addi	%r12, %r1, PARMSAVE
+	bt	7, .Lparmsave
+	# Our caller has not allocated a parameter save area.
+	# We need to allocate one here and use it to pass gprs to
+	# ffi_closure_helper_LINUX64.  The return value area will do.
+	addi	%r12, %r1, -STACKFRAME+RETVAL
+.Lparmsave:
+	std	%r0, 16(%r1)
+	# Save general regs into parm save area
+	std	%r3, 0(%r12)
+	std	%r4, 8(%r12)
+	std	%r5, 16(%r12)
+	std	%r6, 24(%r12)
+	std	%r7, 32(%r12)
+	std	%r8, 40(%r12)
+	std	%r9, 48(%r12)
+	std	%r10, 56(%r12)
+
+	# load up the pointer to the parm save area
+	mr	%r5, %r12
+# else
 	mflr	%r0
 	# Save general regs into parm save area
 	# This is the parameter save area set up by our caller.
@@ -74,6 +118,7 @@ ffi_closure_LINUX64:
 
 	# load up the pointer to the parm save area
 	addi	%r5, %r1, PARMSAVE
+# endif
 
 	# next save fpr 1 to fpr 13
 	stfd	%f1, -104+(0*8)(%r1)
@@ -103,11 +148,11 @@ ffi_closure_LINUX64:
 	mr	%r3, %r11
 
 	# make the call
-#ifdef _CALL_LINUX
+# if defined _CALL_LINUX || _CALL_ELF == 2
 	bl ffi_closure_helper_LINUX64
-#else
+# else
 	bl .ffi_closure_helper_LINUX64
-#endif
+# endif
 .Lret:
 
 	# now r3 contains the return type
@@ -116,10 +161,12 @@ ffi_closure_LINUX64:
 
 	# look up the proper starting point in table
 	# by using return type as offset
+	ld %r0, STACKFRAME+16(%r1)
+	cmpldi %r3, FFI_V2_TYPE_SMALL_STRUCT
+	bge .Lsmall
 	mflr %r4		# move address of .Lret to r4
 	sldi %r3, %r3, 4	# now multiply return type by 16
 	addi %r4, %r4, .Lret_type0 - .Lret
-	ld %r0, STACKFRAME+16(%r1)
 	add %r3, %r3, %r4	# add contents of table to table address
 	mtctr %r3
 	bctr			# jump to it
@@ -234,15 +281,73 @@ ffi_closure_LINUX64:
 	mtlr %r0
 	addi %r1, %r1, STACKFRAME
 	blr
-# esac
+# case FFI_V2_TYPE_FLOAT_HOMOG
+	lfs %f1, RETVAL+0(%r1)
+	lfs %f2, RETVAL+4(%r1)
+	lfs %f3, RETVAL+8(%r1)
+	b .Lmorefloat
+# case FFI_V2_TYPE_DOUBLE_HOMOG
+	lfd %f1, RETVAL+0(%r1)
+	lfd %f2, RETVAL+8(%r1)
+	lfd %f3, RETVAL+16(%r1)
+	lfd %f4, RETVAL+24(%r1)
+	mtlr %r0
+	lfd %f5, RETVAL+32(%r1)
+	lfd %f6, RETVAL+40(%r1)
+	lfd %f7, RETVAL+48(%r1)
+	lfd %f8, RETVAL+56(%r1)
+	addi %r1, %r1, STACKFRAME
+	blr
+.Lmorefloat:
+	lfs %f4, RETVAL+12(%r1)
+	mtlr %r0
+	lfs %f5, RETVAL+16(%r1)
+	lfs %f6, RETVAL+20(%r1)
+	lfs %f7, RETVAL+24(%r1)
+	lfs %f8, RETVAL+28(%r1)
+	addi %r1, %r1, STACKFRAME
+	blr
+.Lsmall:
+# ifdef __LITTLE_ENDIAN__
+	ld %r3,RETVAL+0(%r1)
+	mtlr %r0
+	ld %r4,RETVAL+8(%r1)
+	addi %r1, %r1, STACKFRAME
+	blr
+# else
+	# A struct smaller than a dword is returned in the low bits of r3
+	# ie. right justified.  Larger structs are passed left justified
+	# in r3 and r4.  The return value area on the stack will have
+	# the structs as they are usually stored in memory.
+	cmpldi %r3, FFI_V2_TYPE_SMALL_STRUCT + 7 # size 8 bytes?
+	neg %r5, %r3
+	ld %r3,RETVAL+0(%r1)
+	blt .Lsmalldown
+	mtlr %r0
+	ld %r4,RETVAL+8(%r1)
+	addi %r1, %r1, STACKFRAME
+	blr
+.Lsmalldown:
+	addi %r5, %r5, FFI_V2_TYPE_SMALL_STRUCT + 7
+	mtlr %r0
+	sldi %r5, %r5, 3
+	addi %r1, %r1, STACKFRAME
+	srd %r3, %r3, %r5
+	blr
+# endif
+
 .LFE1:
 	.long	0
 	.byte	0,12,0,1,128,0,0,0
-#ifdef _CALL_LINUX
+# if _CALL_ELF == 2
+	.size	ffi_closure_LINUX64,.-ffi_closure_LINUX64
+# else
+#  ifdef _CALL_LINUX
 	.size	ffi_closure_LINUX64,.-.L.ffi_closure_LINUX64
-#else
+#  else
 	.size	.ffi_closure_LINUX64,.-.ffi_closure_LINUX64
-#endif
+#  endif
+# endif
 
 	.section	.eh_frame,EH_FRAME_FLAGS,@progbits
 .Lframe1: