IEEE754 binary128 long double support for PowerPC64 (#526)

* powerpc: Adjust flags to make room for vector types

* powerpc64 ELFv2 IEEE128 long double support

src/powerpc/ffi.c

@@ -85,8 +85,9 @@ ffi_call_int (ffi_cif *cif,
      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.  */
+     for similar reasons. This bounce buffer must be aligned to 16
-  unsigned long smst_buffer[8];
+     bytes for use with homogeneous structs of vectors (float128).  */
+  float128 smst_buffer[8];
   extended_cif ecif;
 
   ecif.cif = cif;

src/powerpc/ffi_linux64.c

@@ -38,7 +38,8 @@
 /* About the LINUX64 ABI.  */
 enum {
   NUM_GPR_ARG_REGISTERS64 = 8,
-  NUM_FPR_ARG_REGISTERS64 = 13
+  NUM_FPR_ARG_REGISTERS64 = 13,
+  NUM_VEC_ARG_REGISTERS64 = 12,
 };
 enum { ASM_NEEDS_REGISTERS64 = 4 };
 
@@ -81,11 +82,12 @@ discover_homogeneous_aggregate (ffi_abi abi,
          of FPRs, but according to the ABI must be considered
          distinct from doubles. They are also limited to a
          maximum of four members in a homogeneous aggregate. */
-      else
+      else if ((abi & FFI_LINUX_LONG_DOUBLE_IEEE128) == 0)
         {
           *elnum = 2;
           return FFI_TYPE_LONGDOUBLE;
         }
+      /* Fall through. */
 #endif
     case FFI_TYPE_FLOAT:
     case FFI_TYPE_DOUBLE:
@@ -130,13 +132,23 @@ ffi_prep_cif_linux64_core (ffi_cif *cif)
 {
   ffi_type **ptr;
   unsigned bytes;
-  unsigned i, fparg_count = 0, intarg_count = 0;
+  unsigned i, fparg_count = 0, intarg_count = 0, vecarg_count = 0;
   unsigned flags = cif->flags;
   unsigned elt, elnum, rtype;
 
 #if FFI_TYPE_LONGDOUBLE == FFI_TYPE_DOUBLE
-  /* If compiled without long double support..  */
+  /* If compiled without long double support... */
-  if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0)
+  if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0 ||
+      (cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+    return FFI_BAD_ABI;
+#elif !defined(__VEC__)
+  /* If compiled without vector register support (used by assembly)... */
+  if ((cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+    return FFI_BAD_ABI;
+#else
+  /* If the IEEE128 flag is set, but long double is only 64 bits wide... */
+  if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) == 0 &&
+      (cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
     return FFI_BAD_ABI;
 #endif
 
@@ -166,6 +178,11 @@ homogeneous:
     {
 #if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
     case FFI_TYPE_LONGDOUBLE:
+      if ((cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+        {
+          flags |= FLAG_RETURNS_VEC;
+          break;
+        }
       if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0)
 	flags |= FLAG_RETURNS_128BITS;
       /* Fall through.  */
@@ -221,6 +238,15 @@ homogeneous:
 	{
 #if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
 	case FFI_TYPE_LONGDOUBLE:
+          if ((cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+            {
+              vecarg_count++;
+              /* Align to 16 bytes, plus the 16-byte argument. */
+              intarg_count = (intarg_count + 3) & ~0x1;
+              if (vecarg_count > NUM_VEC_ARG_REGISTERS64)
+                flags |= FLAG_ARG_NEEDS_PSAVE;
+              break;
+            }
 	  if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0)
 	    {
 	      fparg_count++;
@@ -248,6 +274,17 @@ homogeneous:
 	    }
 	  intarg_count += ((*ptr)->size + 7) / 8;
 	  elt = discover_homogeneous_aggregate (cif->abi, *ptr, &elnum);
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+          if (elt == FFI_TYPE_LONGDOUBLE &&
+              (cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+            {
+              vecarg_count += elnum;
+              if (vecarg_count > NUM_VEC_ARG_REGISTERS64)
+                flags |= FLAG_ARG_NEEDS_PSAVE;
+              break;
+            }
+	  else
+#endif
 	  if (elt)
 	    {
 	      fparg_count += elnum;
@@ -286,10 +323,17 @@ homogeneous:
     flags |= FLAG_FP_ARGUMENTS;
   if (intarg_count > 4)
     flags |= FLAG_4_GPR_ARGUMENTS;
+  if (vecarg_count != 0)
+    flags |= FLAG_VEC_ARGUMENTS;
 
   /* Space for the FPR registers, if needed.  */
   if (fparg_count != 0)
     bytes += NUM_FPR_ARG_REGISTERS64 * sizeof (double);
+  /* Space for the vector registers, if needed, aligned to 16 bytes. */
+  if (vecarg_count != 0) {
+    bytes = (bytes + 15) & ~0xF;
+    bytes += NUM_VEC_ARG_REGISTERS64 * sizeof (float128);
+  }
 
   /* Stack space.  */
 #if _CALL_ELF == 2
@@ -372,6 +416,8 @@ ffi_prep_cif_linux64_var (ffi_cif *cif,
    |--------------------------------------------| |
    |   FPR registers f1-f13 (optional)	13*8	| |
    |--------------------------------------------| |
+   |   VEC registers v2-v13 (optional)  12*16   | |
+   |--------------------------------------------| |
    |   Parameter save area		        | |
    |--------------------------------------------| |
    |   TOC save area			8	| |
@@ -401,6 +447,7 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
     unsigned long *ul;
     float *f;
     double *d;
+    float128 *f128;
     size_t p;
   } valp;
 
@@ -419,6 +466,11 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
   valp fpr_base;
   unsigned int fparg_count;
 
+  /* 'vec_base' points at the space for v2, and grows upwards as
+     we use vector registers.  */
+  valp vec_base;
+  unsigned int vecarg_count;
+
   unsigned int i, words, nargs, nfixedargs;
   ffi_type **ptr;
   double double_tmp;
@@ -435,6 +487,7 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
     unsigned long **ul;
     float **f;
     double **d;
+    float128 **f128;
   } p_argv;
   unsigned long gprvalue;
   unsigned long align;
@@ -449,11 +502,21 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
 #endif
   fpr_base.d = gpr_base.d - NUM_FPR_ARG_REGISTERS64;
   fparg_count = 0;
+  /* Place the vector args below the FPRs, if used, else the GPRs. */
+  if (ecif->cif->flags & FLAG_FP_ARGUMENTS)
+    vec_base.p = fpr_base.p & ~0xF;
+  else
+    vec_base.p = gpr_base.p;
+  vec_base.f128 -= NUM_VEC_ARG_REGISTERS64;
+  vecarg_count = 0;
   next_arg.ul = gpr_base.ul;
 
   /* Check that everything starts aligned properly.  */
   FFI_ASSERT (((unsigned long) (char *) stack & 0xF) == 0);
   FFI_ASSERT (((unsigned long) stacktop.c & 0xF) == 0);
+  FFI_ASSERT (((unsigned long) gpr_base.c & 0xF) == 0);
+  FFI_ASSERT (((unsigned long) gpr_end.c  & 0xF) == 0);
+  FFI_ASSERT (((unsigned long) vec_base.c & 0xF) == 0);
   FFI_ASSERT ((bytes & 0xF) == 0);
 
   /* Deal with return values that are actually pass-by-reference.  */
@@ -478,6 +541,22 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
 	{
 #if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
 	case FFI_TYPE_LONGDOUBLE:
+          if ((ecif->cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+            {
+              next_arg.p = FFI_ALIGN (next_arg.p, 16);
+              if (next_arg.ul == gpr_end.ul)
+                next_arg.ul = rest.ul;
+              if (vecarg_count < NUM_VEC_ARG_REGISTERS64 && i < nfixedargs)
+                *vec_base.f128++ = **p_argv.f128;
+              else
+                *next_arg.f128 = **p_argv.f128;
+              if (++next_arg.f128 == gpr_end.f128)
+                next_arg.f128 = rest.f128;
+              vecarg_count++;
+              FFI_ASSERT (__LDBL_MANT_DIG__ == 113);
+              FFI_ASSERT (flags & FLAG_VEC_ARGUMENTS);
+              break;
+            }
 	  if ((ecif->cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0)
 	    {
 	      double_tmp = (*p_argv.d)[0];
@@ -589,9 +668,29 @@ ffi_prep_args64 (extended_cif *ecif, unsigned long *const stack)
 		void *v;
 		float *f;
 		double *d;
+		float128 *f128;
 	      } arg;
 
 	      arg.v = *p_argv.v;
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+              if (elt == FFI_TYPE_LONGDOUBLE &&
+                  (ecif->cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+                {
+                  do
+                    {
+                      if (vecarg_count < NUM_VEC_ARG_REGISTERS64
+                          && i < nfixedargs)
+                        *vec_base.f128++ = *arg.f128++;
+                      else
+                        *next_arg.f128 = *arg.f128++;
+                      if (++next_arg.f128 == gpr_end.f128)
+                        next_arg.f128 = rest.f128;
+                      vecarg_count++;
+                    }
+                  while (--elnum != 0);
+                }
+              else
+#endif
 	      if (elt == FFI_TYPE_FLOAT)
 		{
 		  do
@@ -762,17 +861,20 @@ ffi_closure_helper_LINUX64 (ffi_cif *cif,
 			    void *user_data,
 			    void *rvalue,
 			    unsigned long *pst,
-			    ffi_dblfl *pfr)
+                            ffi_dblfl *pfr,
+                            float128 *pvec)
 {
   /* rvalue is the pointer to space for return value in closure assembly */
   /* pst is the pointer to parameter save area
      (r3-r10 are stored into its first 8 slots by ffi_closure_LINUX64) */
   /* pfr is the pointer to where f1-f13 are stored in ffi_closure_LINUX64 */
+  /* pvec is the pointer to where v2-v13 are stored in ffi_closure_LINUX64 */
 
   void **avalue;
   ffi_type **arg_types;
   unsigned long i, avn, nfixedargs;
   ffi_dblfl *end_pfr = pfr + NUM_FPR_ARG_REGISTERS64;
+  float128 *end_pvec = pvec + NUM_VEC_ARG_REGISTERS64;
   unsigned long align;
 
   avalue = alloca (cif->nargs * sizeof (void *));
@@ -851,6 +953,7 @@ ffi_closure_helper_LINUX64 (ffi_cif *cif,
 		unsigned long *ul;
 		float *f;
 		double *d;
+		float128 *f128;
 		size_t p;
 	      } to, from;
 
@@ -858,6 +961,17 @@ ffi_closure_helper_LINUX64 (ffi_cif *cif,
 		 aggregate size is not greater than the space taken by
 		 the registers so store back to the register/parameter
 		 save arrays.  */
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+              if (elt == FFI_TYPE_LONGDOUBLE &&
+                  (cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+                {
+                  if (pvec + elnum <= end_pvec)
+                    to.v = pvec;
+                  else
+                    to.v = pst;
+                }
+              else
+#endif
 	      if (pfr + elnum <= end_pfr)
 		to.v = pfr;
 	      else
@@ -865,6 +979,23 @@ ffi_closure_helper_LINUX64 (ffi_cif *cif,
 
 	      avalue[i] = to.v;
 	      from.ul = pst;
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+              if (elt == FFI_TYPE_LONGDOUBLE &&
+                  (cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+                {
+                  do
+                    {
+                      if (pvec < end_pvec && i < nfixedargs)
+                        *to.f128 = *pvec++;
+                      else
+                        *to.f128 = *from.f128;
+                      to.f128++;
+                      from.f128++;
+                    }
+                  while (--elnum != 0);
+                }
+              else
+#endif
 	      if (elt == FFI_TYPE_FLOAT)
 		{
 		  do
@@ -920,7 +1051,18 @@ ffi_closure_helper_LINUX64 (ffi_cif *cif,
 
 #if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
 	case FFI_TYPE_LONGDOUBLE:
-	  if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0)
+          if ((cif->abi & FFI_LINUX_LONG_DOUBLE_IEEE128) != 0)
+            {
+              if (((unsigned long) pst & 0xF) != 0)
+                ++pst;
+              if (pvec < end_pvec && i < nfixedargs)
+                avalue[i] = pvec++;
+              else
+                avalue[i] = pst;
+              pst += 2;
+              break;
+            }
+          else if ((cif->abi & FFI_LINUX_LONG_DOUBLE_128) != 0)
 	    {
 	      if (pfr + 1 < end_pfr && i + 1 < nfixedargs)
 		{
@@ -995,13 +1137,17 @@ ffi_closure_helper_LINUX64 (ffi_cif *cif,
   /* Tell ffi_closure_LINUX64 how to perform return type promotions.  */
   if ((cif->flags & FLAG_RETURNS_SMST) != 0)
     {
-      if ((cif->flags & FLAG_RETURNS_FP) == 0)
+      if ((cif->flags & (FLAG_RETURNS_FP | FLAG_RETURNS_VEC)) == 0)
 	return FFI_V2_TYPE_SMALL_STRUCT + cif->rtype->size - 1;
+      else if ((cif->flags & FLAG_RETURNS_VEC) != 0)
+        return FFI_V2_TYPE_VECTOR_HOMOG;
       else if ((cif->flags & FLAG_RETURNS_64BITS) != 0)
 	return FFI_V2_TYPE_DOUBLE_HOMOG;
       else
 	return FFI_V2_TYPE_FLOAT_HOMOG;
     }
+  if ((cif->flags & FLAG_RETURNS_VEC) != 0)
+    return FFI_V2_TYPE_VECTOR;
   return cif->rtype->type;
 }
 #endif

src/powerpc/ffi_powerpc.h

@@ -31,22 +31,24 @@
 enum {
   /* The assembly depends on these exact flags.  */
   /* These go in cr7 */
-  FLAG_RETURNS_SMST	= 1 << (31-31), /* Used for FFI_SYSV small structs.  */
+  FLAG_RETURNS_SMST     = 1 << (31-31), /* Used for FFI_SYSV small structs.  */
   FLAG_RETURNS_NOTHING  = 1 << (31-30),
   FLAG_RETURNS_FP       = 1 << (31-29),
-  FLAG_RETURNS_64BITS   = 1 << (31-28),
+  FLAG_RETURNS_VEC      = 1 << (31-28),
 
-  /* This goes in cr6 */
+  /* These go in cr6 */
-  FLAG_RETURNS_128BITS  = 1 << (31-27),
+  FLAG_RETURNS_64BITS   = 1 << (31-27),
+  FLAG_RETURNS_128BITS  = 1 << (31-26),
 
-  FLAG_COMPAT		= 1 << (31- 8), /* Not used by assembly */
+  FLAG_COMPAT           = 1 << (31- 8), /* Not used by assembly */
 
   /* These go in cr1 */
   FLAG_ARG_NEEDS_COPY   = 1 << (31- 7), /* Used by sysv code */
   FLAG_ARG_NEEDS_PSAVE  = FLAG_ARG_NEEDS_COPY, /* Used by linux64 code */
   FLAG_FP_ARGUMENTS     = 1 << (31- 6), /* cr1.eq; specified by ABI */
   FLAG_4_GPR_ARGUMENTS  = 1 << (31- 5),
-  FLAG_RETVAL_REFERENCE = 1 << (31- 4)
+  FLAG_RETVAL_REFERENCE = 1 << (31- 4),
+  FLAG_VEC_ARGUMENTS    = 1 << (31- 3),
 };
 
 typedef union
@@ -55,6 +57,14 @@ typedef union
   double d;
 } ffi_dblfl;
 
+#if defined(__FLOAT128_TYPE__)
+typedef _Float128 float128;
+#elif defined(__FLOAT128__)
+typedef __float128 float128;
+#else
+typedef __int128 float128;
+#endif
+
 void FFI_HIDDEN ffi_closure_SYSV (void);
 void FFI_HIDDEN ffi_go_closure_sysv (void);
 void FFI_HIDDEN ffi_call_SYSV(extended_cif *, void (*)(void), void *,
@@ -91,4 +101,5 @@ int FFI_HIDDEN ffi_closure_helper_LINUX64 (ffi_cif *,
 					   void (*) (ffi_cif *, void *,
 						     void **, void *),
 					   void *, void *,
-					   unsigned long *, ffi_dblfl *);
+					   unsigned long *, ffi_dblfl *,
+					   float128 *);

src/powerpc/ffitarget.h

@@ -91,15 +91,19 @@ typedef enum ffi_abi {
   /* This and following bits can reuse FFI_COMPAT values.  */
   FFI_LINUX_STRUCT_ALIGN = 1,
   FFI_LINUX_LONG_DOUBLE_128 = 2,
+  FFI_LINUX_LONG_DOUBLE_IEEE128 = 4,
   FFI_DEFAULT_ABI = (FFI_LINUX
 #  ifdef __STRUCT_PARM_ALIGN__
 		     | FFI_LINUX_STRUCT_ALIGN
 #  endif
 #  ifdef __LONG_DOUBLE_128__
 		     | FFI_LINUX_LONG_DOUBLE_128
+#   ifdef __LONG_DOUBLE_IEEE128__
+		     | FFI_LINUX_LONG_DOUBLE_IEEE128
+#   endif
 #  endif
 		     ),
-  FFI_LAST_ABI = 12
+  FFI_LAST_ABI = 16
 
 # else
   /* This bit, always set in new code, must not be set in any of the
@@ -167,9 +171,11 @@ typedef enum ffi_abi {
 #define FFI_SYSV_TYPE_SMALL_STRUCT (FFI_PPC_TYPE_LAST + 2)
 
 /* Used by ELFv2 for homogenous structure returns.  */
-#define FFI_V2_TYPE_FLOAT_HOMOG		(FFI_PPC_TYPE_LAST + 1)
+#define FFI_V2_TYPE_VECTOR		(FFI_PPC_TYPE_LAST + 1)
-#define FFI_V2_TYPE_DOUBLE_HOMOG	(FFI_PPC_TYPE_LAST + 2)
+#define FFI_V2_TYPE_VECTOR_HOMOG	(FFI_PPC_TYPE_LAST + 2)
-#define FFI_V2_TYPE_SMALL_STRUCT	(FFI_PPC_TYPE_LAST + 3)
+#define FFI_V2_TYPE_FLOAT_HOMOG		(FFI_PPC_TYPE_LAST + 3)
+#define FFI_V2_TYPE_DOUBLE_HOMOG	(FFI_PPC_TYPE_LAST + 4)
+#define FFI_V2_TYPE_SMALL_STRUCT	(FFI_PPC_TYPE_LAST + 5)
 
 #if _CALL_ELF == 2
 # define FFI_TRAMPOLINE_SIZE 32

src/powerpc/linux64.S

@@ -101,40 +101,70 @@ ffi_call_LINUX64:
 	ld	%r2, 8(%r29)
 # endif
 	/* Now do the call.  */
-	/* Set up cr1 with bits 4-7 of the flags.  */
+	/* Set up cr1 with bits 3-7 of the flags.  */
-	mtcrf	0x40, %r31
+	mtcrf	0xc0, %r31
 
 	/* Get the address to call into CTR.  */
 	mtctr	%r12
 	/* Load all those argument registers.  */
-	ld	%r3, -32-(8*8)(%r28)
+	addi	%r29, %r28, -32-(8*8)
-	ld	%r4, -32-(7*8)(%r28)
+	ld	%r3,  (0*8)(%r29)
-	ld	%r5, -32-(6*8)(%r28)
+	ld	%r4,  (1*8)(%r29)
-	ld	%r6, -32-(5*8)(%r28)
+	ld	%r5,  (2*8)(%r29)
+	ld	%r6,  (3*8)(%r29)
 	bf-	5, 1f
-	ld	%r7, -32-(4*8)(%r28)
+	ld	%r7,  (4*8)(%r29)
-	ld	%r8, -32-(3*8)(%r28)
+	ld	%r8,  (5*8)(%r29)
-	ld	%r9, -32-(2*8)(%r28)
+	ld	%r9,  (6*8)(%r29)
-	ld	%r10, -32-(1*8)(%r28)
+	ld	%r10, (7*8)(%r29)
 1:
 
 	/* Load all the FP registers.  */
 	bf-	6, 2f
-	lfd	%f1, -32-(21*8)(%r28)
+	addi	%r29, %r29, -(14*8)
-	lfd	%f2, -32-(20*8)(%r28)
+	lfd	%f1,  ( 1*8)(%r29)
-	lfd	%f3, -32-(19*8)(%r28)
+	lfd	%f2,  ( 2*8)(%r29)
-	lfd	%f4, -32-(18*8)(%r28)
+	lfd	%f3,  ( 3*8)(%r29)
-	lfd	%f5, -32-(17*8)(%r28)
+	lfd	%f4,  ( 4*8)(%r29)
-	lfd	%f6, -32-(16*8)(%r28)
+	lfd	%f5,  ( 5*8)(%r29)
-	lfd	%f7, -32-(15*8)(%r28)
+	lfd	%f6,  ( 6*8)(%r29)
-	lfd	%f8, -32-(14*8)(%r28)
+	lfd	%f7,  ( 7*8)(%r29)
-	lfd	%f9, -32-(13*8)(%r28)
+	lfd	%f8,  ( 8*8)(%r29)
-	lfd	%f10, -32-(12*8)(%r28)
+	lfd	%f9,  ( 9*8)(%r29)
-	lfd	%f11, -32-(11*8)(%r28)
+	lfd	%f10, (10*8)(%r29)
-	lfd	%f12, -32-(10*8)(%r28)
+	lfd	%f11, (11*8)(%r29)
-	lfd	%f13, -32-(9*8)(%r28)
+	lfd	%f12, (12*8)(%r29)
+	lfd	%f13, (13*8)(%r29)
 2:
 
+	/* Load all the vector registers.  */
+	bf-	3, 3f
+	addi	%r29, %r29, -16
+	lvx	%v13, 0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v12, 0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v11, 0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v10, 0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v9,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v8,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v7,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v6,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v5,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v4,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v3,  0, %r29
+	addi	%r29, %r29, -16
+	lvx	%v2,  0, %r29
+3:
+
 	/* Make the call.  */
 	ld	%r11, 8(%r28)
 	bctrl
@@ -152,6 +182,7 @@ ffi_call_LINUX64:
 	bt	31, .Lstruct_return_value
 	bt	30, .Ldone_return_value
 	bt	29, .Lfp_return_value
+	bt	28, .Lvec_return_value
 	std	%r3, 0(%r30)
 	/* Fall through...  */
 
@@ -167,12 +198,16 @@ ffi_call_LINUX64:
 	ld	%r31, -8(%r1)
 	blr
 
+.Lvec_return_value:
+	stvx	%v2, 0, %r30
+	b	.Ldone_return_value
+
 .Lfp_return_value:
 	.cfi_def_cfa_register 28
-	bf	28, .Lfloat_return_value
-	stfd	%f1, 0(%r30)
 	mtcrf	0x02, %r31 /* cr6  */
-	bf	27, .Ldone_return_value
+	bf	27, .Lfloat_return_value
+	stfd	%f1, 0(%r30)
+	bf	26, .Ldone_return_value
 	stfd	%f2, 8(%r30)
 	b	.Ldone_return_value
 .Lfloat_return_value:
@@ -180,8 +215,9 @@ ffi_call_LINUX64:
 	b	.Ldone_return_value
 
 .Lstruct_return_value:
-	bf	29, .Lsmall_struct
+	bf	29, .Lvec_homog_or_small_struct
-	bf	28, .Lfloat_homog_return_value
+	mtcrf	0x02, %r31 /* cr6  */
+	bf	27, .Lfloat_homog_return_value
 	stfd	%f1, 0(%r30)
 	stfd	%f2, 8(%r30)
 	stfd	%f3, 16(%r30)
@@ -203,6 +239,25 @@ ffi_call_LINUX64:
 	stfs	%f8, 28(%r30)
 	b	.Ldone_return_value
 
+.Lvec_homog_or_small_struct:
+	bf	28, .Lsmall_struct
+	stvx	%v2, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v3, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v4, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v5, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v6, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v7, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v8, 0, %r30
+	addi	%r30, %r30, 16
+	stvx	%v9, 0, %r30
+	b	.Ldone_return_value
+
 .Lsmall_struct:
 	std	%r3, 0(%r30)
 	std	%r4, 8(%r30)

src/powerpc/linux64_closure.S

@@ -61,9 +61,15 @@ ffi_closure_LINUX64:
 # endif
 
 # if _CALL_ELF == 2
-#  32 byte special reg save area + 64 byte parm save area
+#  ifdef __VEC__
-#  + 64 byte retval area + 13*8 fpr save area + round to 16
+#   32 byte special reg save area + 64 byte parm save area
-#  define STACKFRAME 272
+#   + 128 byte retval area + 13*8 fpr save area + 12*16 vec save area + round to 16
+#   define STACKFRAME 528
+#  else
+#   32 byte special reg save area + 64 byte parm save area
+#   + 64 byte retval area + 13*8 fpr save area + round to 16
+#   define STACKFRAME 272
+#  endif
 #  define PARMSAVE 32
 #  define RETVAL PARMSAVE+64
 # else
@@ -146,6 +152,35 @@ ffi_closure_LINUX64:
 	# load up the pointer to the saved fpr registers
 	addi	%r8, %r1, -104
 
+# ifdef __VEC__
+	# load up the pointer to the saved vector registers
+	# 8 bytes padding for 16-byte alignment at -112(%r1)
+	addi	%r9, %r8, -24
+	stvx	%v13, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v12, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v11, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v10, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v9, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v8, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v7, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v6, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v5, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v4, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v3, 0, %r9
+	addi	%r9, %r9, -16
+	stvx	%v2, 0, %r9
+# endif
+
 	# load up the pointer to the result storage
 	addi	%r6, %r1, -STACKFRAME+RETVAL
 
@@ -313,6 +348,16 @@ ffi_closure_LINUX64:
 	.cfi_def_cfa_offset 0
 	blr
 	.cfi_def_cfa_offset STACKFRAME
+# case FFI_V2_TYPE_VECTOR
+	addi %r3, %r1, RETVAL
+	lvx %v2, 0, %r3
+	mtlr %r0
+	b .Lfinish
+# case FFI_V2_TYPE_VECTOR_HOMOG
+	addi %r3, %r1, RETVAL
+	lvx %v2, 0, %r3
+	addi %r3, %r3, 16
+	b .Lmorevector
 # case FFI_V2_TYPE_FLOAT_HOMOG
 	lfs %f1, RETVAL+0(%r1)
 	lfs %f2, RETVAL+4(%r1)
@@ -332,6 +377,25 @@ ffi_closure_LINUX64:
 	.cfi_def_cfa_offset 0
 	blr
 	.cfi_def_cfa_offset STACKFRAME
+.Lmorevector:
+	lvx %v3, 0, %r3
+	addi %r3, %r3, 16
+	lvx %v4, 0, %r3
+	addi %r3, %r3, 16
+	lvx %v5, 0, %r3
+	mtlr %r0
+	addi %r3, %r3, 16
+	lvx %v6, 0, %r3
+	addi %r3, %r3, 16
+	lvx %v7, 0, %r3
+	addi %r3, %r3, 16
+	lvx %v8, 0, %r3
+	addi %r3, %r3, 16
+	lvx %v9, 0, %r3
+	addi %r1, %r1, STACKFRAME
+	.cfi_def_cfa_offset 0
+	blr
+	.cfi_def_cfa_offset STACKFRAME
 .Lmorefloat:
 	lfs %f4, RETVAL+12(%r1)
 	mtlr %r0