* config/spu/spu.c (spu_immediate): Remove trailing comma.

	(reloc_diagnostic): Call warning when -mwarn-reloc is specified.
	* config/spu/spu.md: (zero_extendhisi2): Expand instead of split for
	better optimization.
	(floatv4siv4sf2): New.
	(fix_truncv4sfv4si2): New.
	(floatunsv4siv4sf2): New.
	(fixuns_truncv4sfv4si2): New.
	(addv16qi3): New.
	(subv16qi3): New.
	(negv16qi2): New.
	(mulv8hi3): New.
	(mulsi3): Remove.
	(mul<mode>3): New.
	(_mulv4si3): New.
	(cmp<mode>): Don't accept constant arguments for DI, TI and SF.
	* config/spu/spu_internals.h: Handle overloaded intrinsics in C++ with
	spu_resolve_overloaded_builtin instead of static inline functions.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@119424 138bc75d-0d04-0410-961f-82ee72b054a4

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index f322b66..5f0da7b 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,24 @@
+2006-12-01  Trevor Smigiel  <trevor_smigiel@playstation.sony.com>
+
+       * config/spu/spu.c (spu_immediate): Remove trailing comma.
+       (reloc_diagnostic): Call warning when -mwarn-reloc is specified.
+       * config/spu/spu.md: (zero_extendhisi2): Expand instead of split for
+       better optimization.
+       (floatv4siv4sf2): New.
+       (fix_truncv4sfv4si2): New.
+       (floatunsv4siv4sf2): New.
+       (fixuns_truncv4sfv4si2): New.
+       (addv16qi3): New.
+       (subv16qi3): New.
+       (negv16qi2): New.
+       (mulv8hi3): New.
+       (mulsi3): Remove.
+       (mul<mode>3): New.
+       (_mulv4si3): New.
+       (cmp<mode>): Don't accept constant arguments for DI, TI and SF.
+       * config/spu/spu_internals.h: Handle overloaded intrinsics in C++ with
+       spu_resolve_overloaded_builtin instead of static inline functions.
+
 2006-12-01  Eric Botcazou  <ebotcazou@adacore.com>
 
        * fold-const.c (fold_binary) <LT_EXPR>: Use the precision of the

diff --git a/gcc/config/spu/spu.c b/gcc/config/spu/spu.c
index e623908..c3916c7 100644 (file)
--- a/gcc/config/spu/spu.c
+++ b/gcc/config/spu/spu.c
@@ -141,7 +141,7 @@ enum spu_immediate {
   SPU_ORI,
   SPU_ORHI,
   SPU_ORBI,
-  SPU_IOHL,
+  SPU_IOHL
 };
 
 static enum spu_immediate which_immediate_load (HOST_WIDE_INT val);
@@ -3670,10 +3670,10 @@ reloc_diagnostic (rtx x)
   else
     msg = "creating run-time relocation";
 
-  if (TARGET_ERROR_RELOC) /** default : error reloc **/
-    error (msg, loc_decl, decl);
-  else
+  if (TARGET_WARN_RELOC)
     warning (0, msg, loc_decl, decl);
+  else
+    error (msg, loc_decl, decl);
 }
 
 /* Hook into assemble_integer so we can generate an error for run-time

diff --git a/gcc/config/spu/spu.md b/gcc/config/spu/spu.md
index ef47f81..1cdc20a 100644 (file)
--- a/gcc/config/spu/spu.md
+++ b/gcc/config/spu/spu.md
@@ -425,20 +425,19 @@
   ""
   "andi\t%0,%1,0x00ff")
 
-(define_insn_and_split "zero_extendhisi2"
+(define_expand "zero_extendhisi2"
   [(set (match_operand:SI 0 "spu_reg_operand" "=r")
        (zero_extend:SI (match_operand:HI 1 "spu_reg_operand" "r")))
    (clobber (match_scratch:SI 2 "=&r"))]
   ""
-  "#"
-  "reload_completed"
-  [(set (match_dup:SI 2)
-        (const_int 65535))
-   (set (match_dup:SI 0)
-        (and:SI (match_dup:SI 3)
-                (match_dup:SI 2)))]
-  "operands[3] = gen_rtx_REG (SImode, REGNO (operands[1]));")
-
+  {
+    rtx mask = gen_reg_rtx (SImode);
+    rtx op1 = simplify_gen_subreg (SImode, operands[1], HImode, 0);
+    emit_move_insn (mask, GEN_INT (0xffff));
+    emit_insn (gen_andsi3(operands[0], op1, mask));
+    DONE;
+  })
+  
 (define_insn "zero_extendsidi2"
   [(set (match_operand:DI 0 "spu_reg_operand" "=r")
        (zero_extend:DI (match_operand:SI 1 "spu_reg_operand" "r")))]
@@ -522,6 +521,13 @@
   "csflt\t%0,%1,0"
   [(set_attr "type" "fp7")])
 
+(define_insn "floatv4siv4sf2"
+  [(set (match_operand:V4SF 0 "spu_reg_operand" "=r")
+       (float:V4SF (match_operand:V4SI 1 "spu_reg_operand" "r")))]
+  ""
+  "csflt\t%0,%1,0"
+  [(set_attr "type" "fp7")])
+
 (define_insn "fix_truncsfsi2"
   [(set (match_operand:SI 0 "spu_reg_operand" "=r")
        (fix:SI (match_operand:SF 1 "spu_reg_operand" "r")))]
@@ -529,6 +535,13 @@
   "cflts\t%0,%1,0"
   [(set_attr "type" "fp7")])
 
+(define_insn "fix_truncv4sfv4si2"
+  [(set (match_operand:V4SI 0 "spu_reg_operand" "=r")
+       (fix:V4SI (match_operand:V4SF 1 "spu_reg_operand" "r")))]
+  ""
+  "cflts\t%0,%1,0"
+  [(set_attr "type" "fp7")])
+
 (define_insn "floatunssisf2"
   [(set (match_operand:SF 0 "spu_reg_operand" "=r")
        (unsigned_float:SF (match_operand:SI 1 "spu_reg_operand" "r")))]
@@ -536,6 +549,13 @@
   "cuflt\t%0,%1,0"
   [(set_attr "type" "fp7")])
 
+(define_insn "floatunsv4siv4sf2"
+  [(set (match_operand:V4SF 0 "spu_reg_operand" "=r")
+       (unsigned_float:V4SF (match_operand:V4SI 1 "spu_reg_operand" "r")))]
+  ""
+  "cuflt\t%0,%1,0"
+  [(set_attr "type" "fp7")])
+
 (define_insn "fixuns_truncsfsi2"
   [(set (match_operand:SI 0 "spu_reg_operand" "=r")
        (unsigned_fix:SI (match_operand:SF 1 "spu_reg_operand" "r")))]
@@ -543,6 +563,13 @@
   "cfltu\t%0,%1,0"
   [(set_attr "type" "fp7")])
 
+(define_insn "fixuns_truncv4sfv4si2"
+  [(set (match_operand:V4SI 0 "spu_reg_operand" "=r")
+       (unsigned_fix:V4SI (match_operand:V4SF 1 "spu_reg_operand" "r")))]
+  ""
+  "cfltu\t%0,%1,0"
+  [(set_attr "type" "fp7")])
+
 (define_insn "extendsfdf2"
   [(set (match_operand:DF 0 "spu_reg_operand" "=r")
        (float_extend:DF (match_operand:SF 1 "spu_reg_operand" "r")))]
@@ -627,6 +654,28 @@
 \f
 ;; add
 
+(define_expand "addv16qi3"
+  [(set (match_operand:V16QI 0 "spu_reg_operand" "=r")
+       (plus:V16QI (match_operand:V16QI 1 "spu_reg_operand" "r")
+                   (match_operand:V16QI 2 "spu_reg_operand" "r")))]
+  ""
+  "{
+    rtx res_short = simplify_gen_subreg (V8HImode, operands[0], V16QImode, 0);
+    rtx lhs_short = simplify_gen_subreg (V8HImode, operands[1], V16QImode, 0);
+    rtx rhs_short = simplify_gen_subreg (V8HImode, operands[2], V16QImode, 0);
+    rtx rhs_and = gen_reg_rtx (V8HImode);
+    rtx hi_char = gen_reg_rtx (V8HImode);
+    rtx lo_char = gen_reg_rtx (V8HImode);
+    rtx mask = gen_reg_rtx (V8HImode);
+
+    emit_move_insn (mask, spu_const (V8HImode, 0x00ff));
+    emit_insn (gen_andv8hi3 (rhs_and, rhs_short, spu_const (V8HImode, 0xff00)));
+    emit_insn (gen_addv8hi3 (hi_char, lhs_short, rhs_and));
+    emit_insn (gen_addv8hi3 (lo_char, lhs_short, rhs_short));
+    emit_insn (gen_selb (res_short, hi_char, lo_char, mask));
+    DONE;
+   }")
+
 (define_insn "add<mode>3"
   [(set (match_operand:VHSI 0 "spu_reg_operand" "=r,r")
        (plus:VHSI (match_operand:VHSI 1 "spu_reg_operand" "r,r")
@@ -728,6 +777,28 @@
 \f
 ;; sub
 
+(define_expand "subv16qi3"
+  [(set (match_operand:V16QI 0 "spu_reg_operand" "=r")
+       (minus:V16QI (match_operand:V16QI 1 "spu_reg_operand" "r")
+                    (match_operand:V16QI 2 "spu_reg_operand" "r")))]
+  ""
+  "{
+    rtx res_short = simplify_gen_subreg (V8HImode, operands[0], V16QImode, 0);
+    rtx lhs_short = simplify_gen_subreg (V8HImode, operands[1], V16QImode, 0);
+    rtx rhs_short = simplify_gen_subreg (V8HImode, operands[2], V16QImode, 0);
+    rtx rhs_and = gen_reg_rtx (V8HImode);
+    rtx hi_char = gen_reg_rtx (V8HImode);
+    rtx lo_char = gen_reg_rtx (V8HImode);
+    rtx mask = gen_reg_rtx (V8HImode);
+
+    emit_move_insn (mask, spu_const (V8HImode, 0x00ff));
+    emit_insn (gen_andv8hi3 (rhs_and, rhs_short, spu_const (V8HImode, 0xff00)));
+    emit_insn (gen_subv8hi3 (hi_char, lhs_short, rhs_and));
+    emit_insn (gen_subv8hi3 (lo_char, lhs_short, rhs_short));
+    emit_insn (gen_selb (res_short, hi_char, lo_char, mask));
+    DONE;
+   }")
+
 (define_insn "sub<mode>3"
   [(set (match_operand:VHSI 0 "spu_reg_operand" "=r,r")
        (minus:VHSI (match_operand:VHSI 1 "spu_arith_operand" "r,B")
@@ -825,6 +896,17 @@
 \f
 ;; neg
 
+(define_expand "negv16qi2"
+  [(set (match_operand:V16QI 0 "spu_reg_operand" "=r")
+       (neg:V16QI (match_operand:V16QI 1 "spu_reg_operand" "r")))]
+  ""
+  "{
+    rtx zero = gen_reg_rtx (V16QImode);
+    emit_move_insn (zero, CONST0_RTX (V16QImode));
+    emit_insn (gen_subv16qi3 (operands[0], zero, operands[1]));
+    DONE;
+   }")
+
 (define_insn "neg<mode>2"
   [(set (match_operand:VHSI 0 "spu_reg_operand" "=r")
        (neg:VHSI (match_operand:VHSI 1 "spu_reg_operand" "r")))]
@@ -935,27 +1017,47 @@
   mpyi\t%0,%1,%2"
   [(set_attr "type" "fp7")])
 
-(define_expand "mulsi3"
+(define_expand "mulv8hi3"
+  [(set (match_operand:V8HI 0 "spu_reg_operand" "")
+       (mult:V8HI (match_operand:V8HI 1 "spu_reg_operand" "")
+                  (match_operand:V8HI 2 "spu_reg_operand" "")))]
+  ""
+  "{
+    rtx result = simplify_gen_subreg (V4SImode, operands[0], V8HImode, 0);
+    rtx low = gen_reg_rtx (V4SImode);
+    rtx high = gen_reg_rtx (V4SImode);
+    rtx shift = gen_reg_rtx (V4SImode);
+    rtx mask = gen_reg_rtx (V4SImode);
+
+    emit_move_insn (mask, spu_const (V4SImode, 0x0000ffff));
+    emit_insn (gen_spu_mpyhh (high, operands[1], operands[2]));
+    emit_insn (gen_spu_mpy (low, operands[1], operands[2]));
+    emit_insn (gen_ashlv4si3 (shift, high, spu_const(V4SImode, 16)));
+    emit_insn (gen_selb (result, shift, low, mask));
+    DONE;
+   }")
+
+(define_expand "mul<mode>3"
   [(parallel
-    [(set (match_operand:SI 0 "spu_reg_operand" "")
-         (mult:SI (match_operand:SI 1 "spu_reg_operand" "")
-                  (match_operand:SI 2 "spu_reg_operand" "")))
-     (clobber (match_dup:SI 3))
-     (clobber (match_dup:SI 4))
-     (clobber (match_dup:SI 5))
-     (clobber (match_dup:SI 6))])]
+    [(set (match_operand:VSI 0 "spu_reg_operand" "")
+         (mult:VSI (match_operand:VSI 1 "spu_reg_operand" "")
+                   (match_operand:VSI 2 "spu_reg_operand" "")))
+     (clobber (match_dup:VSI 3))
+     (clobber (match_dup:VSI 4))
+     (clobber (match_dup:VSI 5))
+     (clobber (match_dup:VSI 6))])]
   ""
   {
-    operands[3] = gen_reg_rtx(SImode);
-    operands[4] = gen_reg_rtx(SImode);
-    operands[5] = gen_reg_rtx(SImode);
-    operands[6] = gen_reg_rtx(SImode);
+    operands[3] = gen_reg_rtx(<MODE>mode);
+    operands[4] = gen_reg_rtx(<MODE>mode);
+    operands[5] = gen_reg_rtx(<MODE>mode);
+    operands[6] = gen_reg_rtx(<MODE>mode);
   })
 
 (define_insn_and_split "_mulsi3"
   [(set (match_operand:SI 0 "spu_reg_operand" "=r")
        (mult:SI (match_operand:SI 1 "spu_reg_operand" "r")
-                (match_operand:SI 2 "spu_nonmem_operand" "ri")))
+                (match_operand:SI 2 "spu_arith_operand" "rK")))
    (clobber (match_operand:SI 3 "spu_reg_operand" "=&r"))
    (clobber (match_operand:SI 4 "spu_reg_operand" "=&r"))
    (clobber (match_operand:SI 5 "spu_reg_operand" "=&r"))
@@ -1000,6 +1102,37 @@
     DONE;
    })
 
+(define_insn_and_split "_mulv4si3"
+  [(set (match_operand:V4SI 0 "spu_reg_operand" "=r")
+       (mult:V4SI (match_operand:V4SI 1 "spu_reg_operand" "r")
+                  (match_operand:V4SI 2 "spu_reg_operand" "r")))
+   (clobber (match_operand:V4SI 3 "spu_reg_operand" "=&r"))
+   (clobber (match_operand:V4SI 4 "spu_reg_operand" "=&r"))
+   (clobber (match_operand:V4SI 5 "spu_reg_operand" "=&r"))
+   (clobber (match_operand:V4SI 6 "spu_reg_operand" "=&r"))]
+  ""
+  "#"
+  ""
+  [(set (match_dup:V4SI 0)
+       (mult:V4SI (match_dup:V4SI 1)
+                  (match_dup:V4SI 2)))]
+  {
+    HOST_WIDE_INT val = 0;
+    rtx a = operands[3];
+    rtx b = operands[4];
+    rtx c = operands[5];
+    rtx d = operands[6];
+    rtx op1 = simplify_gen_subreg (V8HImode, operands[1], V4SImode, 0);
+    rtx op2 = simplify_gen_subreg (V8HImode, operands[2], V4SImode, 0);
+    rtx op3 = simplify_gen_subreg (V8HImode, operands[3], V4SImode, 0);
+    emit_insn(gen_spu_mpyh(a, op1, op2));
+    emit_insn(gen_spu_mpyh(b, op2, op1));
+    emit_insn(gen_spu_mpyu(c, op1, op2));
+    emit_insn(gen_addv4si3(d, a, b));
+    emit_insn(gen_addv4si3(operands[0], d, c));
+    DONE;
+   })
+
 (define_insn "mulhisi3"
   [(set (match_operand:SI 0 "spu_reg_operand" "=r")
        (mult:SI (sign_extend:SI (match_operand:HI 1 "spu_reg_operand" "r"))
@@ -2532,8 +2665,19 @@ selb\t%0,%4,%0,%3"
 
 (define_expand "cmp<mode>"
   [(set (cc0)
-       (compare (match_operand:VINT 0 "spu_reg_operand" "")
-                (match_operand:VINT 1 "spu_nonmem_operand" "")))]
+       (compare (match_operand:VQHSI 0 "spu_reg_operand" "")
+                (match_operand:VQHSI 1 "spu_nonmem_operand" "")))]
+  ""
+  {
+    spu_compare_op0 = operands[0];
+    spu_compare_op1 = operands[1];
+    DONE;
+  })
+
+(define_expand "cmp<mode>"
+  [(set (cc0)
+       (compare (match_operand:DTI 0 "spu_reg_operand" "")
+                (match_operand:DTI 1 "spu_reg_operand" "")))]
   ""
   {
     spu_compare_op0 = operands[0];
@@ -2544,7 +2688,7 @@ selb\t%0,%4,%0,%3"
 (define_expand "cmp<mode>"
   [(set (cc0)
        (compare (match_operand:VSF 0 "spu_reg_operand" "")
-                (match_operand:VSF 1 "spu_nonmem_operand" "")))]
+                (match_operand:VSF 1 "spu_reg_operand" "")))]
   ""
   {
     spu_compare_op0 = operands[0];

diff --git a/gcc/config/spu/spu_internals.h b/gcc/config/spu/spu_internals.h
index 752ddb6..98e289d 100644 (file)
--- a/gcc/config/spu/spu_internals.h
+++ b/gcc/config/spu/spu_internals.h
@@ -256,8 +256,6 @@
 
 #define __align_hint(ptr,base,offset) __builtin_spu_align_hint(ptr,base,offset)
 
-#ifndef __cplusplus
-
 /* generic spu_* intrinisics */ 
 
 #define spu_splats(scalar)        __builtin_spu_splats(scalar) 
@@ -330,2444 +328,6 @@
 #define spu_insert(scalar,ra,pos) __builtin_spu_insert(scalar,ra,pos) 
 #define spu_promote(scalar,pos)   __builtin_spu_promote(scalar,pos) 
 
-#else /* __cplusplus */
-
-/* A bit of a hack...  Float conversion needs an immediate operand.
- * always_inline doesn't help because the compiler generates an error
- * before inlining happens. */
-static inline vec_float4 __hack_spu_convtf (vec_int4, vec_float4, vec_float4) __attribute__((__always_inline__));
-static inline vec_float4 __hack_spu_convtf (vec_uint4, vec_float4, vec_float4) __attribute__((__always_inline__));
-static inline vec_float4
-__hack_spu_convtf (vec_int4 ra, vec_float4 from_signed, vec_float4 from_unsigned)
-{
-  (void)ra;
-  (void)from_unsigned;
-  return from_signed;
-}
-static inline vec_float4
-__hack_spu_convtf (vec_uint4 ra, vec_float4 from_signed, vec_float4 from_unsigned)
-{
-  (void)ra;
-  (void)from_signed;
-  return from_unsigned;
-}
-#define spu_convtf(ra,imm) \
-  __hack_spu_convtf((ra), \
-                   __builtin_spu_convtf_1((vec_int4)(ra), (imm)), \
-                   __builtin_spu_convtf_0((vec_uint4)(ra), (imm)))
-
-/* The following defines and functions were created automatically from
- * spu_builtins.def. */
-#define spu_convts(a, b)       __builtin_spu_convts (a, b)
-#define spu_convtu(a, b)       __builtin_spu_convtu (a, b)
-#define spu_roundtf(a)         __builtin_spu_roundtf (a)
-#define spu_mulh(a, b)         __builtin_spu_mulh (a, b)
-#define spu_mulsr(a, b)                __builtin_spu_mulsr (a, b)
-#define spu_frest(a)           __builtin_spu_frest (a)
-#define spu_frsqest(a)         __builtin_spu_frsqest (a)
-#define spu_nmadd(a, b, c)     __builtin_spu_nmadd (a, b, c)
-#define spu_absd(a, b)         __builtin_spu_absd (a, b)
-#define spu_avg(a, b)          __builtin_spu_avg (a, b)
-#define spu_sumb(a, b)         __builtin_spu_sumb (a, b)
-#define spu_bisled(a)          __builtin_spu_bisled (a, 0)
-#define spu_bisled_d(a)                __builtin_spu_bisled_d (a, 0)
-#define spu_bisled_e(a)                __builtin_spu_bisled_e (a, 0)
-#define spu_cmpabseq(a, b)     __builtin_spu_cmpabseq (a, b)
-#define spu_cmpabsgt(a, b)     __builtin_spu_cmpabsgt (a, b)
-
-static inline vec_short8 spu_extend (vec_char16 a) __attribute__((__always_inline__));
-static inline vec_int4 spu_extend (vec_short8 a) __attribute__((__always_inline__));
-static inline vec_llong2 spu_extend (vec_int4 a) __attribute__((__always_inline__));
-static inline vec_double2 spu_extend (vec_float4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_add (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_add (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_add (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_add (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_add (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_add (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_add (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_short8 spu_add (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_add (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_add (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_addx (vec_int4 a, vec_int4 b, vec_int4 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_addx (vec_uint4 a, vec_uint4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_genc (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_genc (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_gencx (vec_int4 a, vec_int4 b, vec_int4 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gencx (vec_uint4 a, vec_uint4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_madd (vec_short8 a, vec_short8 b, vec_int4 c) __attribute__((__always_inline__));
-static inline vec_float4 spu_madd (vec_float4 a, vec_float4 b, vec_float4 c) __attribute__((__always_inline__));
-static inline vec_double2 spu_madd (vec_double2 a, vec_double2 b, vec_double2 c) __attribute__((__always_inline__));
-static inline vec_float4 spu_msub (vec_float4 a, vec_float4 b, vec_float4 c) __attribute__((__always_inline__));
-static inline vec_double2 spu_msub (vec_double2 a, vec_double2 b, vec_double2 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_mhhadd (vec_ushort8 a, vec_ushort8 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_mhhadd (vec_short8 a, vec_short8 b, vec_int4 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_mule (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_mule (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_mul (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_mul (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_mulo (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_mulo (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_mulo (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_mulo (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_float4 spu_nmsub (vec_float4 a, vec_float4 b, vec_float4 c) __attribute__((__always_inline__));
-static inline vec_double2 spu_nmsub (vec_double2 a, vec_double2 b, vec_double2 c) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_sub (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_sub (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_sub (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_sub (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_sub (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_sub (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_sub (unsigned short a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_sub (short a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_sub (unsigned int a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_sub (int a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_subx (vec_uint4 a, vec_uint4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_subx (vec_int4 a, vec_int4 b, vec_int4 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_genb (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_genb (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_genbx (vec_uint4 a, vec_uint4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_genbx (vec_int4 a, vec_int4 b, vec_int4 c) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpeq (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpeq (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpeq (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpeq (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpeq (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpeq (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpeq (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpeq (vec_uchar16 a, unsigned char b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpeq (vec_char16 a, signed char b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpeq (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpeq (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpeq (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpeq (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpgt (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpgt (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpgt (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpgt (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpgt (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpgt (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpgt (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpgt (vec_uchar16 a, unsigned char b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cmpgt (vec_char16 a, signed char b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpgt (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_cmpgt (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpgt (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cmpgt (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline void spu_hcmpeq (int a, int b) __attribute__((__always_inline__));
-static inline void spu_hcmpeq (unsigned int a, unsigned int b) __attribute__((__always_inline__));
-static inline void spu_hcmpgt (int a, int b) __attribute__((__always_inline__));
-static inline void spu_hcmpgt (unsigned int a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cntb (vec_char16 a) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_cntb (vec_uchar16 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cntlz (vec_int4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cntlz (vec_uint4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_cntlz (vec_float4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_int4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_uint4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_short8 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_ushort8 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_char16 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_uchar16 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_gather (vec_float4 a) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_maskb (unsigned short a) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_maskb (short a) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_maskb (unsigned int a) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_maskb (int a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (unsigned char a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (signed char a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (char a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (unsigned short a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (short a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (unsigned int a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_maskh (int a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (unsigned char a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (signed char a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (char a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (unsigned short a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (short a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (unsigned int a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_maskw (int a) __attribute__((__always_inline__));
-static inline vec_llong2 spu_sel (vec_llong2 a, vec_llong2 b, vec_ullong2 c) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_sel (vec_ullong2 a, vec_ullong2 b, vec_ullong2 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_sel (vec_int4 a, vec_int4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_sel (vec_uint4 a, vec_uint4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_short8 spu_sel (vec_short8 a, vec_short8 b, vec_ushort8 c) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_sel (vec_ushort8 a, vec_ushort8 b, vec_ushort8 c) __attribute__((__always_inline__));
-static inline vec_char16 spu_sel (vec_char16 a, vec_char16 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_sel (vec_uchar16 a, vec_uchar16 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_float4 spu_sel (vec_float4 a, vec_float4 b, vec_uint4 c) __attribute__((__always_inline__));
-static inline vec_double2 spu_sel (vec_double2 a, vec_double2 b, vec_ullong2 c) __attribute__((__always_inline__));
-static inline vec_llong2 spu_sel (vec_llong2 a, vec_llong2 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_sel (vec_ullong2 a, vec_ullong2 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_sel (vec_int4 a, vec_int4 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_sel (vec_uint4 a, vec_uint4 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_short8 spu_sel (vec_short8 a, vec_short8 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_sel (vec_ushort8 a, vec_ushort8 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_float4 spu_sel (vec_float4 a, vec_float4 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_double2 spu_sel (vec_double2 a, vec_double2 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_shuffle (vec_uchar16 a, vec_uchar16 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_char16 spu_shuffle (vec_char16 a, vec_char16 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_shuffle (vec_ushort8 a, vec_ushort8 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_short8 spu_shuffle (vec_short8 a, vec_short8 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_shuffle (vec_uint4 a, vec_uint4 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_int4 spu_shuffle (vec_int4 a, vec_int4 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_shuffle (vec_ullong2 a, vec_ullong2 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_llong2 spu_shuffle (vec_llong2 a, vec_llong2 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_float4 spu_shuffle (vec_float4 a, vec_float4 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_double2 spu_shuffle (vec_double2 a, vec_double2 b, vec_uchar16 c) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_and (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_and (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_and (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_and (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_and (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_and (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_and (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_and (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_and (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_and (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_and (vec_uchar16 a, unsigned char b) __attribute__((__always_inline__));
-static inline vec_char16 spu_and (vec_char16 a, signed char b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_and (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_short8 spu_and (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_and (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_and (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_andc (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_andc (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_andc (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_andc (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_andc (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_andc (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_andc (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_andc (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_andc (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_andc (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_eqv (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_eqv (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_eqv (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_eqv (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_eqv (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_eqv (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_eqv (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_eqv (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_eqv (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_eqv (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_nand (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_nand (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_nand (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_nand (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_nand (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_nand (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_nand (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_nand (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_nand (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_nand (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_nor (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_nor (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_nor (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_nor (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_nor (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_nor (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_nor (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_nor (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_nor (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_nor (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_or (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_or (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_or (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_or (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_or (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_or (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_or (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_or (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_or (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_or (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_or (vec_uchar16 a, unsigned char b) __attribute__((__always_inline__));
-static inline vec_char16 spu_or (vec_char16 a, signed char b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_or (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_short8 spu_or (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_or (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_or (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_orc (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_orc (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_orc (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_orc (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_orc (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_orc (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_orc (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_orc (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_orc (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_orc (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_orx (vec_int4 a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_orx (vec_uint4 a) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_xor (vec_uchar16 a, vec_uchar16 b) __attribute__((__always_inline__));
-static inline vec_char16 spu_xor (vec_char16 a, vec_char16 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_xor (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_xor (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_xor (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_xor (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_xor (vec_ullong2 a, vec_ullong2 b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_xor (vec_llong2 a, vec_llong2 b) __attribute__((__always_inline__));
-static inline vec_float4 spu_xor (vec_float4 a, vec_float4 b) __attribute__((__always_inline__));
-static inline vec_double2 spu_xor (vec_double2 a, vec_double2 b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_xor (vec_uchar16 a, unsigned char b) __attribute__((__always_inline__));
-static inline vec_char16 spu_xor (vec_char16 a, signed char b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_xor (vec_ushort8 a, unsigned short b) __attribute__((__always_inline__));
-static inline vec_short8 spu_xor (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_xor (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_xor (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rl (vec_ushort8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rl (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rl (vec_uint4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rl (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rl (vec_ushort8 a, short b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rl (vec_short8 a, short b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rl (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rl (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_rlqw (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_rlqw (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlqw (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlqw (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlqw (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlqw (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_rlqw (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_rlqw (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_rlqw (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_rlqw (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_rlqwbyte (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_rlqwbyte (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlqwbyte (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlqwbyte (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlqwbyte (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlqwbyte (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_rlqwbyte (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_rlqwbyte (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_rlqwbyte (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_rlqwbyte (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_rlqwbytebc (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_rlqwbytebc (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlqwbytebc (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlqwbytebc (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlqwbytebc (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlqwbytebc (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_rlqwbytebc (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_rlqwbytebc (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_rlqwbytebc (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_rlqwbytebc (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmask (vec_ushort8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmask (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmask (vec_uint4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmask (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmask (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmask (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmask (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmask (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmaska (vec_ushort8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmaska (vec_short8 a, vec_short8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmaska (vec_uint4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmaska (vec_int4 a, vec_int4 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmaska (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmaska (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmaska (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmaska (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_rlmaskqw (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_rlmaskqw (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmaskqw (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmaskqw (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmaskqw (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmaskqw (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_rlmaskqw (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_rlmaskqw (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_rlmaskqw (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_rlmaskqw (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_rlmaskqwbyte (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_rlmaskqwbyte (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmaskqwbyte (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmaskqwbyte (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmaskqwbyte (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmaskqwbyte (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_rlmaskqwbyte (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_rlmaskqwbyte (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_rlmaskqwbyte (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_rlmaskqwbyte (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_rlmaskqwbytebc (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_rlmaskqwbytebc (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_rlmaskqwbytebc (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_rlmaskqwbytebc (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_rlmaskqwbytebc (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_rlmaskqwbytebc (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_rlmaskqwbytebc (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_rlmaskqwbytebc (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_rlmaskqwbytebc (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_rlmaskqwbytebc (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_sl (vec_ushort8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_short8 spu_sl (vec_short8 a, vec_ushort8 b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_sl (vec_uint4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_int4 spu_sl (vec_int4 a, vec_uint4 b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_sl (vec_ushort8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_sl (vec_short8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_sl (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_sl (vec_int4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_slqw (vec_llong2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_slqw (vec_ullong2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_slqw (vec_int4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_slqw (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_slqw (vec_short8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_slqw (vec_ushort8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_slqw (vec_char16 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_slqw (vec_uchar16 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_slqw (vec_float4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_slqw (vec_double2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_slqwbyte (vec_llong2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_slqwbyte (vec_ullong2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_slqwbyte (vec_int4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_slqwbyte (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_slqwbyte (vec_short8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_slqwbyte (vec_ushort8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_slqwbyte (vec_char16 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_slqwbyte (vec_uchar16 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_slqwbyte (vec_float4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_slqwbyte (vec_double2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_slqwbytebc (vec_llong2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_slqwbytebc (vec_ullong2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_slqwbytebc (vec_int4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_slqwbytebc (vec_uint4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_slqwbytebc (vec_short8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_slqwbytebc (vec_ushort8 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_slqwbytebc (vec_char16 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_slqwbytebc (vec_uchar16 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_slqwbytebc (vec_float4 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_slqwbytebc (vec_double2 a, unsigned int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_splats (unsigned char a) __attribute__((__always_inline__));
-static inline vec_char16 spu_splats (signed char a) __attribute__((__always_inline__));
-static inline vec_char16 spu_splats (char a) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_splats (unsigned short a) __attribute__((__always_inline__));
-static inline vec_short8 spu_splats (short a) __attribute__((__always_inline__));
-static inline vec_uint4 spu_splats (unsigned int a) __attribute__((__always_inline__));
-static inline vec_int4 spu_splats (int a) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_splats (unsigned long long a) __attribute__((__always_inline__));
-static inline vec_llong2 spu_splats (long long a) __attribute__((__always_inline__));
-static inline vec_float4 spu_splats (float a) __attribute__((__always_inline__));
-static inline vec_double2 spu_splats (double a) __attribute__((__always_inline__));
-static inline unsigned char spu_extract (vec_uchar16 a, int b) __attribute__((__always_inline__));
-static inline signed char spu_extract (vec_char16 a, int b) __attribute__((__always_inline__));
-static inline unsigned short spu_extract (vec_ushort8 a, int b) __attribute__((__always_inline__));
-static inline short spu_extract (vec_short8 a, int b) __attribute__((__always_inline__));
-static inline unsigned int spu_extract (vec_uint4 a, int b) __attribute__((__always_inline__));
-static inline int spu_extract (vec_int4 a, int b) __attribute__((__always_inline__));
-static inline unsigned long long spu_extract (vec_ullong2 a, int b) __attribute__((__always_inline__));
-static inline long long spu_extract (vec_llong2 a, int b) __attribute__((__always_inline__));
-static inline float spu_extract (vec_float4 a, int b) __attribute__((__always_inline__));
-static inline double spu_extract (vec_double2 a, int b) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_insert (unsigned char a, vec_uchar16 b, int c) __attribute__((__always_inline__));
-static inline vec_char16 spu_insert (signed char a, vec_char16 b, int c) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_insert (unsigned short a, vec_ushort8 b, int c) __attribute__((__always_inline__));
-static inline vec_short8 spu_insert (short a, vec_short8 b, int c) __attribute__((__always_inline__));
-static inline vec_uint4 spu_insert (unsigned int a, vec_uint4 b, int c) __attribute__((__always_inline__));
-static inline vec_int4 spu_insert (int a, vec_int4 b, int c) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_insert (unsigned long long a, vec_ullong2 b, int c) __attribute__((__always_inline__));
-static inline vec_llong2 spu_insert (long long a, vec_llong2 b, int c) __attribute__((__always_inline__));
-static inline vec_float4 spu_insert (float a, vec_float4 b, int c) __attribute__((__always_inline__));
-static inline vec_double2 spu_insert (double a, vec_double2 b, int c) __attribute__((__always_inline__));
-static inline vec_uchar16 spu_promote (unsigned char a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_promote (signed char a, int b) __attribute__((__always_inline__));
-static inline vec_char16 spu_promote (char a, int b) __attribute__((__always_inline__));
-static inline vec_ushort8 spu_promote (unsigned short a, int b) __attribute__((__always_inline__));
-static inline vec_short8 spu_promote (short a, int b) __attribute__((__always_inline__));
-static inline vec_uint4 spu_promote (unsigned int a, int b) __attribute__((__always_inline__));
-static inline vec_int4 spu_promote (int a, int b) __attribute__((__always_inline__));
-static inline vec_ullong2 spu_promote (unsigned long long a, int b) __attribute__((__always_inline__));
-static inline vec_llong2 spu_promote (long long a, int b) __attribute__((__always_inline__));
-static inline vec_float4 spu_promote (float a, int b) __attribute__((__always_inline__));
-static inline vec_double2 spu_promote (double a, int b) __attribute__((__always_inline__));
-
-static inline vec_short8
-spu_extend (vec_char16 a)
-{
-  return __builtin_spu_extend_0 (a);
-}
-static inline vec_int4
-spu_extend (vec_short8 a)
-{
-  return __builtin_spu_extend_1 (a);
-}
-static inline vec_llong2
-spu_extend (vec_int4 a)
-{
-  return __builtin_spu_extend_2 (a);
-}
-static inline vec_double2
-spu_extend (vec_float4 a)
-{
-  return __builtin_spu_extend_3 (a);
-}
-static inline vec_uint4
-spu_add (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_add_0 (a, b);
-}
-static inline vec_int4
-spu_add (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_add_1 (a, b);
-}
-static inline vec_ushort8
-spu_add (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_add_2 (a, b);
-}
-static inline vec_short8
-spu_add (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_add_3 (a, b);
-}
-static inline vec_float4
-spu_add (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_add_4 (a, b);
-}
-static inline vec_double2
-spu_add (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_add_5 (a, b);
-}
-static inline vec_ushort8
-spu_add (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_add_6 (a, b);
-}
-static inline vec_short8
-spu_add (vec_short8 a, short b)
-{
-  return __builtin_spu_add_7 (a, b);
-}
-static inline vec_uint4
-spu_add (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_add_8 (a, b);
-}
-static inline vec_int4
-spu_add (vec_int4 a, int b)
-{
-  return __builtin_spu_add_9 (a, b);
-}
-static inline vec_int4
-spu_addx (vec_int4 a, vec_int4 b, vec_int4 c)
-{
-  return __builtin_spu_addx_0 (a, b, c);
-}
-static inline vec_uint4
-spu_addx (vec_uint4 a, vec_uint4 b, vec_uint4 c)
-{
-  return __builtin_spu_addx_1 (a, b, c);
-}
-static inline vec_int4
-spu_genc (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_genc_0 (a, b);
-}
-static inline vec_uint4
-spu_genc (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_genc_1 (a, b);
-}
-static inline vec_int4
-spu_gencx (vec_int4 a, vec_int4 b, vec_int4 c)
-{
-  return __builtin_spu_gencx_0 (a, b, c);
-}
-static inline vec_uint4
-spu_gencx (vec_uint4 a, vec_uint4 b, vec_uint4 c)
-{
-  return __builtin_spu_gencx_1 (a, b, c);
-}
-static inline vec_int4
-spu_madd (vec_short8 a, vec_short8 b, vec_int4 c)
-{
-  return __builtin_spu_madd_0 (a, b, c);
-}
-static inline vec_float4
-spu_madd (vec_float4 a, vec_float4 b, vec_float4 c)
-{
-  return __builtin_spu_madd_1 (a, b, c);
-}
-static inline vec_double2
-spu_madd (vec_double2 a, vec_double2 b, vec_double2 c)
-{
-  return __builtin_spu_madd_2 (a, b, c);
-}
-static inline vec_float4
-spu_msub (vec_float4 a, vec_float4 b, vec_float4 c)
-{
-  return __builtin_spu_msub_0 (a, b, c);
-}
-static inline vec_double2
-spu_msub (vec_double2 a, vec_double2 b, vec_double2 c)
-{
-  return __builtin_spu_msub_1 (a, b, c);
-}
-static inline vec_uint4
-spu_mhhadd (vec_ushort8 a, vec_ushort8 b, vec_uint4 c)
-{
-  return __builtin_spu_mhhadd_0 (a, b, c);
-}
-static inline vec_int4
-spu_mhhadd (vec_short8 a, vec_short8 b, vec_int4 c)
-{
-  return __builtin_spu_mhhadd_1 (a, b, c);
-}
-static inline vec_uint4
-spu_mule (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_mule_0 (a, b);
-}
-static inline vec_int4
-spu_mule (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_mule_1 (a, b);
-}
-static inline vec_float4
-spu_mul (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_mul_0 (a, b);
-}
-static inline vec_double2
-spu_mul (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_mul_1 (a, b);
-}
-static inline vec_int4
-spu_mulo (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_mulo_0 (a, b);
-}
-static inline vec_uint4
-spu_mulo (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_mulo_1 (a, b);
-}
-static inline vec_int4
-spu_mulo (vec_short8 a, short b)
-{
-  return __builtin_spu_mulo_2 (a, b);
-}
-static inline vec_uint4
-spu_mulo (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_mulo_3 (a, b);
-}
-static inline vec_float4
-spu_nmsub (vec_float4 a, vec_float4 b, vec_float4 c)
-{
-  return __builtin_spu_nmsub_0 (a, b, c);
-}
-static inline vec_double2
-spu_nmsub (vec_double2 a, vec_double2 b, vec_double2 c)
-{
-  return __builtin_spu_nmsub_1 (a, b, c);
-}
-static inline vec_ushort8
-spu_sub (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_sub_0 (a, b);
-}
-static inline vec_short8
-spu_sub (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_sub_1 (a, b);
-}
-static inline vec_uint4
-spu_sub (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_sub_2 (a, b);
-}
-static inline vec_int4
-spu_sub (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_sub_3 (a, b);
-}
-static inline vec_float4
-spu_sub (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_sub_4 (a, b);
-}
-static inline vec_double2
-spu_sub (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_sub_5 (a, b);
-}
-static inline vec_ushort8
-spu_sub (unsigned short a, vec_ushort8 b)
-{
-  return __builtin_spu_sub_6 (a, b);
-}
-static inline vec_short8
-spu_sub (short a, vec_short8 b)
-{
-  return __builtin_spu_sub_7 (a, b);
-}
-static inline vec_uint4
-spu_sub (unsigned int a, vec_uint4 b)
-{
-  return __builtin_spu_sub_8 (a, b);
-}
-static inline vec_int4
-spu_sub (int a, vec_int4 b)
-{
-  return __builtin_spu_sub_9 (a, b);
-}
-static inline vec_uint4
-spu_subx (vec_uint4 a, vec_uint4 b, vec_uint4 c)
-{
-  return __builtin_spu_subx_0 (a, b, c);
-}
-static inline vec_int4
-spu_subx (vec_int4 a, vec_int4 b, vec_int4 c)
-{
-  return __builtin_spu_subx_1 (a, b, c);
-}
-static inline vec_uint4
-spu_genb (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_genb_0 (a, b);
-}
-static inline vec_int4
-spu_genb (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_genb_1 (a, b);
-}
-static inline vec_uint4
-spu_genbx (vec_uint4 a, vec_uint4 b, vec_uint4 c)
-{
-  return __builtin_spu_genbx_0 (a, b, c);
-}
-static inline vec_int4
-spu_genbx (vec_int4 a, vec_int4 b, vec_int4 c)
-{
-  return __builtin_spu_genbx_1 (a, b, c);
-}
-static inline vec_uchar16
-spu_cmpeq (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_cmpeq_0 (a, b);
-}
-static inline vec_uchar16
-spu_cmpeq (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_cmpeq_1 (a, b);
-}
-static inline vec_ushort8
-spu_cmpeq (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_cmpeq_2 (a, b);
-}
-static inline vec_ushort8
-spu_cmpeq (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_cmpeq_3 (a, b);
-}
-static inline vec_uint4
-spu_cmpeq (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_cmpeq_4 (a, b);
-}
-static inline vec_uint4
-spu_cmpeq (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_cmpeq_5 (a, b);
-}
-static inline vec_uint4
-spu_cmpeq (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_cmpeq_6 (a, b);
-}
-static inline vec_uchar16
-spu_cmpeq (vec_uchar16 a, unsigned char b)
-{
-  return __builtin_spu_cmpeq_7 (a, b);
-}
-static inline vec_uchar16
-spu_cmpeq (vec_char16 a, signed char b)
-{
-  return __builtin_spu_cmpeq_8 (a, b);
-}
-static inline vec_ushort8
-spu_cmpeq (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_cmpeq_9 (a, b);
-}
-static inline vec_ushort8
-spu_cmpeq (vec_short8 a, short b)
-{
-  return __builtin_spu_cmpeq_10 (a, b);
-}
-static inline vec_uint4
-spu_cmpeq (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_cmpeq_11 (a, b);
-}
-static inline vec_uint4
-spu_cmpeq (vec_int4 a, int b)
-{
-  return __builtin_spu_cmpeq_12 (a, b);
-}
-static inline vec_uchar16
-spu_cmpgt (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_cmpgt_0 (a, b);
-}
-static inline vec_uchar16
-spu_cmpgt (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_cmpgt_1 (a, b);
-}
-static inline vec_ushort8
-spu_cmpgt (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_cmpgt_2 (a, b);
-}
-static inline vec_ushort8
-spu_cmpgt (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_cmpgt_3 (a, b);
-}
-static inline vec_uint4
-spu_cmpgt (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_cmpgt_4 (a, b);
-}
-static inline vec_uint4
-spu_cmpgt (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_cmpgt_5 (a, b);
-}
-static inline vec_uint4
-spu_cmpgt (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_cmpgt_6 (a, b);
-}
-static inline vec_uchar16
-spu_cmpgt (vec_uchar16 a, unsigned char b)
-{
-  return __builtin_spu_cmpgt_7 (a, b);
-}
-static inline vec_uchar16
-spu_cmpgt (vec_char16 a, signed char b)
-{
-  return __builtin_spu_cmpgt_8 (a, b);
-}
-static inline vec_ushort8
-spu_cmpgt (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_cmpgt_9 (a, b);
-}
-static inline vec_ushort8
-spu_cmpgt (vec_short8 a, short b)
-{
-  return __builtin_spu_cmpgt_10 (a, b);
-}
-static inline vec_uint4
-spu_cmpgt (vec_int4 a, int b)
-{
-  return __builtin_spu_cmpgt_11 (a, b);
-}
-static inline vec_uint4
-spu_cmpgt (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_cmpgt_12 (a, b);
-}
-static inline void
-spu_hcmpeq (int a, int b)
-{
-  return __builtin_spu_hcmpeq_0 (a, b);
-}
-static inline void
-spu_hcmpeq (unsigned int a, unsigned int b)
-{
-  return __builtin_spu_hcmpeq_1 (a, b);
-}
-static inline void
-spu_hcmpgt (int a, int b)
-{
-  return __builtin_spu_hcmpgt_0 (a, b);
-}
-static inline void
-spu_hcmpgt (unsigned int a, unsigned int b)
-{
-  return __builtin_spu_hcmpgt_1 (a, b);
-}
-static inline vec_uchar16
-spu_cntb (vec_char16 a)
-{
-  return __builtin_spu_cntb_0 (a);
-}
-static inline vec_uchar16
-spu_cntb (vec_uchar16 a)
-{
-  return __builtin_spu_cntb_1 (a);
-}
-static inline vec_uint4
-spu_cntlz (vec_int4 a)
-{
-  return __builtin_spu_cntlz_0 (a);
-}
-static inline vec_uint4
-spu_cntlz (vec_uint4 a)
-{
-  return __builtin_spu_cntlz_1 (a);
-}
-static inline vec_uint4
-spu_cntlz (vec_float4 a)
-{
-  return __builtin_spu_cntlz_2 (a);
-}
-static inline vec_uint4
-spu_gather (vec_int4 a)
-{
-  return __builtin_spu_gather_0 (a);
-}
-static inline vec_uint4
-spu_gather (vec_uint4 a)
-{
-  return __builtin_spu_gather_1 (a);
-}
-static inline vec_uint4
-spu_gather (vec_short8 a)
-{
-  return __builtin_spu_gather_2 (a);
-}
-static inline vec_uint4
-spu_gather (vec_ushort8 a)
-{
-  return __builtin_spu_gather_3 (a);
-}
-static inline vec_uint4
-spu_gather (vec_char16 a)
-{
-  return __builtin_spu_gather_4 (a);
-}
-static inline vec_uint4
-spu_gather (vec_uchar16 a)
-{
-  return __builtin_spu_gather_5 (a);
-}
-static inline vec_uint4
-spu_gather (vec_float4 a)
-{
-  return __builtin_spu_gather_6 (a);
-}
-static inline vec_uchar16
-spu_maskb (unsigned short a)
-{
-  return __builtin_spu_maskb_0 (a);
-}
-static inline vec_uchar16
-spu_maskb (short a)
-{
-  return __builtin_spu_maskb_1 (a);
-}
-static inline vec_uchar16
-spu_maskb (unsigned int a)
-{
-  return __builtin_spu_maskb_2 (a);
-}
-static inline vec_uchar16
-spu_maskb (int a)
-{
-  return __builtin_spu_maskb_3 (a);
-}
-static inline vec_ushort8
-spu_maskh (unsigned char a)
-{
-  return __builtin_spu_maskh_0 (a);
-}
-static inline vec_ushort8
-spu_maskh (signed char a)
-{
-  return __builtin_spu_maskh_1 (a);
-}
-static inline vec_ushort8
-spu_maskh (char a)
-{
-  return __builtin_spu_maskh_1 (a);
-}
-static inline vec_ushort8
-spu_maskh (unsigned short a)
-{
-  return __builtin_spu_maskh_2 (a);
-}
-static inline vec_ushort8
-spu_maskh (short a)
-{
-  return __builtin_spu_maskh_3 (a);
-}
-static inline vec_ushort8
-spu_maskh (unsigned int a)
-{
-  return __builtin_spu_maskh_4 (a);
-}
-static inline vec_ushort8
-spu_maskh (int a)
-{
-  return __builtin_spu_maskh_5 (a);
-}
-static inline vec_uint4
-spu_maskw (unsigned char a)
-{
-  return __builtin_spu_maskw_0 (a);
-}
-static inline vec_uint4
-spu_maskw (signed char a)
-{
-  return __builtin_spu_maskw_1 (a);
-}
-static inline vec_uint4
-spu_maskw (char a)
-{
-  return __builtin_spu_maskw_1 (a);
-}
-static inline vec_uint4
-spu_maskw (unsigned short a)
-{
-  return __builtin_spu_maskw_2 (a);
-}
-static inline vec_uint4
-spu_maskw (short a)
-{
-  return __builtin_spu_maskw_3 (a);
-}
-static inline vec_uint4
-spu_maskw (unsigned int a)
-{
-  return __builtin_spu_maskw_4 (a);
-}
-static inline vec_uint4
-spu_maskw (int a)
-{
-  return __builtin_spu_maskw_5 (a);
-}
-static inline vec_llong2
-spu_sel (vec_llong2 a, vec_llong2 b, vec_ullong2 c)
-{
-  return __builtin_spu_sel_0 (a, b, c);
-}
-static inline vec_ullong2
-spu_sel (vec_ullong2 a, vec_ullong2 b, vec_ullong2 c)
-{
-  return __builtin_spu_sel_1 (a, b, c);
-}
-static inline vec_int4
-spu_sel (vec_int4 a, vec_int4 b, vec_uint4 c)
-{
-  return __builtin_spu_sel_2 (a, b, c);
-}
-static inline vec_uint4
-spu_sel (vec_uint4 a, vec_uint4 b, vec_uint4 c)
-{
-  return __builtin_spu_sel_3 (a, b, c);
-}
-static inline vec_short8
-spu_sel (vec_short8 a, vec_short8 b, vec_ushort8 c)
-{
-  return __builtin_spu_sel_4 (a, b, c);
-}
-static inline vec_ushort8
-spu_sel (vec_ushort8 a, vec_ushort8 b, vec_ushort8 c)
-{
-  return __builtin_spu_sel_5 (a, b, c);
-}
-static inline vec_char16
-spu_sel (vec_char16 a, vec_char16 b, vec_uchar16 c)
-{
-  return __builtin_spu_sel_6 (a, b, c);
-}
-static inline vec_uchar16
-spu_sel (vec_uchar16 a, vec_uchar16 b, vec_uchar16 c)
-{
-  return __builtin_spu_sel_7 (a, b, c);
-}
-static inline vec_float4
-spu_sel (vec_float4 a, vec_float4 b, vec_uint4 c)
-{
-  return __builtin_spu_sel_8 (a, b, c);
-}
-static inline vec_double2
-spu_sel (vec_double2 a, vec_double2 b, vec_ullong2 c)
-{
-  return __builtin_spu_sel_9 (a, b, c);
-}
-static inline vec_uchar16
-spu_shuffle (vec_uchar16 a, vec_uchar16 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_0 (a, b, c);
-}
-static inline vec_char16
-spu_shuffle (vec_char16 a, vec_char16 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_1 (a, b, c);
-}
-static inline vec_ushort8
-spu_shuffle (vec_ushort8 a, vec_ushort8 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_2 (a, b, c);
-}
-static inline vec_short8
-spu_shuffle (vec_short8 a, vec_short8 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_3 (a, b, c);
-}
-static inline vec_uint4
-spu_shuffle (vec_uint4 a, vec_uint4 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_4 (a, b, c);
-}
-static inline vec_int4
-spu_shuffle (vec_int4 a, vec_int4 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_5 (a, b, c);
-}
-static inline vec_ullong2
-spu_shuffle (vec_ullong2 a, vec_ullong2 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_6 (a, b, c);
-}
-static inline vec_llong2
-spu_shuffle (vec_llong2 a, vec_llong2 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_7 (a, b, c);
-}
-static inline vec_float4
-spu_shuffle (vec_float4 a, vec_float4 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_8 (a, b, c);
-}
-static inline vec_double2
-spu_shuffle (vec_double2 a, vec_double2 b, vec_uchar16 c)
-{
-  return __builtin_spu_shuffle_9 (a, b, c);
-}
-static inline vec_uchar16
-spu_and (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_and_0 (a, b);
-}
-static inline vec_char16
-spu_and (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_and_1 (a, b);
-}
-static inline vec_ushort8
-spu_and (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_and_2 (a, b);
-}
-static inline vec_short8
-spu_and (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_and_3 (a, b);
-}
-static inline vec_uint4
-spu_and (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_and_4 (a, b);
-}
-static inline vec_int4
-spu_and (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_and_5 (a, b);
-}
-static inline vec_ullong2
-spu_and (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_and_6 (a, b);
-}
-static inline vec_llong2
-spu_and (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_and_7 (a, b);
-}
-static inline vec_float4
-spu_and (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_and_8 (a, b);
-}
-static inline vec_double2
-spu_and (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_and_9 (a, b);
-}
-static inline vec_uchar16
-spu_and (vec_uchar16 a, unsigned char b)
-{
-  return __builtin_spu_and_10 (a, b);
-}
-static inline vec_char16
-spu_and (vec_char16 a, signed char b)
-{
-  return __builtin_spu_and_11 (a, b);
-}
-static inline vec_ushort8
-spu_and (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_and_12 (a, b);
-}
-static inline vec_short8
-spu_and (vec_short8 a, short b)
-{
-  return __builtin_spu_and_13 (a, b);
-}
-static inline vec_uint4
-spu_and (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_and_14 (a, b);
-}
-static inline vec_int4
-spu_and (vec_int4 a, int b)
-{
-  return __builtin_spu_and_15 (a, b);
-}
-static inline vec_llong2
-spu_andc (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_andc_0 (a, b);
-}
-static inline vec_ullong2
-spu_andc (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_andc_1 (a, b);
-}
-static inline vec_int4
-spu_andc (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_andc_2 (a, b);
-}
-static inline vec_uint4
-spu_andc (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_andc_3 (a, b);
-}
-static inline vec_short8
-spu_andc (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_andc_4 (a, b);
-}
-static inline vec_ushort8
-spu_andc (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_andc_5 (a, b);
-}
-static inline vec_char16
-spu_andc (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_andc_6 (a, b);
-}
-static inline vec_uchar16
-spu_andc (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_andc_7 (a, b);
-}
-static inline vec_float4
-spu_andc (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_andc_8 (a, b);
-}
-static inline vec_double2
-spu_andc (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_andc_9 (a, b);
-}
-static inline vec_llong2
-spu_eqv (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_eqv_0 (a, b);
-}
-static inline vec_ullong2
-spu_eqv (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_eqv_1 (a, b);
-}
-static inline vec_int4
-spu_eqv (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_eqv_2 (a, b);
-}
-static inline vec_uint4
-spu_eqv (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_eqv_3 (a, b);
-}
-static inline vec_short8
-spu_eqv (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_eqv_4 (a, b);
-}
-static inline vec_ushort8
-spu_eqv (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_eqv_5 (a, b);
-}
-static inline vec_char16
-spu_eqv (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_eqv_6 (a, b);
-}
-static inline vec_uchar16
-spu_eqv (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_eqv_7 (a, b);
-}
-static inline vec_float4
-spu_eqv (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_eqv_8 (a, b);
-}
-static inline vec_double2
-spu_eqv (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_eqv_9 (a, b);
-}
-static inline vec_llong2
-spu_nand (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_nand_0 (a, b);
-}
-static inline vec_ullong2
-spu_nand (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_nand_1 (a, b);
-}
-static inline vec_int4
-spu_nand (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_nand_2 (a, b);
-}
-static inline vec_uint4
-spu_nand (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_nand_3 (a, b);
-}
-static inline vec_short8
-spu_nand (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_nand_4 (a, b);
-}
-static inline vec_ushort8
-spu_nand (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_nand_5 (a, b);
-}
-static inline vec_char16
-spu_nand (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_nand_6 (a, b);
-}
-static inline vec_uchar16
-spu_nand (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_nand_7 (a, b);
-}
-static inline vec_float4
-spu_nand (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_nand_8 (a, b);
-}
-static inline vec_double2
-spu_nand (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_nand_9 (a, b);
-}
-static inline vec_llong2
-spu_nor (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_nor_0 (a, b);
-}
-static inline vec_ullong2
-spu_nor (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_nor_1 (a, b);
-}
-static inline vec_int4
-spu_nor (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_nor_2 (a, b);
-}
-static inline vec_uint4
-spu_nor (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_nor_3 (a, b);
-}
-static inline vec_short8
-spu_nor (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_nor_4 (a, b);
-}
-static inline vec_ushort8
-spu_nor (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_nor_5 (a, b);
-}
-static inline vec_char16
-spu_nor (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_nor_6 (a, b);
-}
-static inline vec_uchar16
-spu_nor (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_nor_7 (a, b);
-}
-static inline vec_float4
-spu_nor (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_nor_8 (a, b);
-}
-static inline vec_double2
-spu_nor (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_nor_9 (a, b);
-}
-static inline vec_uchar16
-spu_or (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_or_0 (a, b);
-}
-static inline vec_char16
-spu_or (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_or_1 (a, b);
-}
-static inline vec_ushort8
-spu_or (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_or_2 (a, b);
-}
-static inline vec_short8
-spu_or (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_or_3 (a, b);
-}
-static inline vec_uint4
-spu_or (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_or_4 (a, b);
-}
-static inline vec_int4
-spu_or (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_or_5 (a, b);
-}
-static inline vec_ullong2
-spu_or (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_or_6 (a, b);
-}
-static inline vec_llong2
-spu_or (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_or_7 (a, b);
-}
-static inline vec_float4
-spu_or (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_or_8 (a, b);
-}
-static inline vec_double2
-spu_or (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_or_9 (a, b);
-}
-static inline vec_uchar16
-spu_or (vec_uchar16 a, unsigned char b)
-{
-  return __builtin_spu_or_10 (a, b);
-}
-static inline vec_char16
-spu_or (vec_char16 a, signed char b)
-{
-  return __builtin_spu_or_11 (a, b);
-}
-static inline vec_ushort8
-spu_or (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_or_12 (a, b);
-}
-static inline vec_short8
-spu_or (vec_short8 a, short b)
-{
-  return __builtin_spu_or_13 (a, b);
-}
-static inline vec_uint4
-spu_or (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_or_14 (a, b);
-}
-static inline vec_int4
-spu_or (vec_int4 a, int b)
-{
-  return __builtin_spu_or_15 (a, b);
-}
-static inline vec_llong2
-spu_orc (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_orc_0 (a, b);
-}
-static inline vec_ullong2
-spu_orc (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_orc_1 (a, b);
-}
-static inline vec_int4
-spu_orc (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_orc_2 (a, b);
-}
-static inline vec_uint4
-spu_orc (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_orc_3 (a, b);
-}
-static inline vec_short8
-spu_orc (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_orc_4 (a, b);
-}
-static inline vec_ushort8
-spu_orc (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_orc_5 (a, b);
-}
-static inline vec_char16
-spu_orc (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_orc_6 (a, b);
-}
-static inline vec_uchar16
-spu_orc (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_orc_7 (a, b);
-}
-static inline vec_float4
-spu_orc (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_orc_8 (a, b);
-}
-static inline vec_double2
-spu_orc (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_orc_9 (a, b);
-}
-static inline vec_int4
-spu_orx (vec_int4 a)
-{
-  return __builtin_spu_orx_0 (a);
-}
-static inline vec_uint4
-spu_orx (vec_uint4 a)
-{
-  return __builtin_spu_orx_1 (a);
-}
-static inline vec_uchar16
-spu_xor (vec_uchar16 a, vec_uchar16 b)
-{
-  return __builtin_spu_xor_0 (a, b);
-}
-static inline vec_char16
-spu_xor (vec_char16 a, vec_char16 b)
-{
-  return __builtin_spu_xor_1 (a, b);
-}
-static inline vec_ushort8
-spu_xor (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_xor_2 (a, b);
-}
-static inline vec_short8
-spu_xor (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_xor_3 (a, b);
-}
-static inline vec_uint4
-spu_xor (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_xor_4 (a, b);
-}
-static inline vec_int4
-spu_xor (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_xor_5 (a, b);
-}
-static inline vec_ullong2
-spu_xor (vec_ullong2 a, vec_ullong2 b)
-{
-  return __builtin_spu_xor_6 (a, b);
-}
-static inline vec_llong2
-spu_xor (vec_llong2 a, vec_llong2 b)
-{
-  return __builtin_spu_xor_7 (a, b);
-}
-static inline vec_float4
-spu_xor (vec_float4 a, vec_float4 b)
-{
-  return __builtin_spu_xor_8 (a, b);
-}
-static inline vec_double2
-spu_xor (vec_double2 a, vec_double2 b)
-{
-  return __builtin_spu_xor_9 (a, b);
-}
-static inline vec_uchar16
-spu_xor (vec_uchar16 a, unsigned char b)
-{
-  return __builtin_spu_xor_10 (a, b);
-}
-static inline vec_char16
-spu_xor (vec_char16 a, signed char b)
-{
-  return __builtin_spu_xor_11 (a, b);
-}
-static inline vec_ushort8
-spu_xor (vec_ushort8 a, unsigned short b)
-{
-  return __builtin_spu_xor_12 (a, b);
-}
-static inline vec_short8
-spu_xor (vec_short8 a, short b)
-{
-  return __builtin_spu_xor_13 (a, b);
-}
-static inline vec_uint4
-spu_xor (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_xor_14 (a, b);
-}
-static inline vec_int4
-spu_xor (vec_int4 a, int b)
-{
-  return __builtin_spu_xor_15 (a, b);
-}
-static inline vec_ushort8
-spu_rl (vec_ushort8 a, vec_short8 b)
-{
-  return __builtin_spu_rl_0 (a, b);
-}
-static inline vec_short8
-spu_rl (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_rl_1 (a, b);
-}
-static inline vec_uint4
-spu_rl (vec_uint4 a, vec_int4 b)
-{
-  return __builtin_spu_rl_2 (a, b);
-}
-static inline vec_int4
-spu_rl (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_rl_3 (a, b);
-}
-static inline vec_ushort8
-spu_rl (vec_ushort8 a, short b)
-{
-  return __builtin_spu_rl_4 (a, b);
-}
-static inline vec_short8
-spu_rl (vec_short8 a, short b)
-{
-  return __builtin_spu_rl_5 (a, b);
-}
-static inline vec_uint4
-spu_rl (vec_uint4 a, int b)
-{
-  return __builtin_spu_rl_6 (a, b);
-}
-static inline vec_int4
-spu_rl (vec_int4 a, int b)
-{
-  return __builtin_spu_rl_7 (a, b);
-}
-static inline vec_uchar16
-spu_rlqw (vec_uchar16 a, int b)
-{
-  return __builtin_spu_rlqw_0 (a, b);
-}
-static inline vec_char16
-spu_rlqw (vec_char16 a, int b)
-{
-  return __builtin_spu_rlqw_1 (a, b);
-}
-static inline vec_ushort8
-spu_rlqw (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlqw_2 (a, b);
-}
-static inline vec_short8
-spu_rlqw (vec_short8 a, int b)
-{
-  return __builtin_spu_rlqw_3 (a, b);
-}
-static inline vec_uint4
-spu_rlqw (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlqw_4 (a, b);
-}
-static inline vec_int4
-spu_rlqw (vec_int4 a, int b)
-{
-  return __builtin_spu_rlqw_5 (a, b);
-}
-static inline vec_ullong2
-spu_rlqw (vec_ullong2 a, int b)
-{
-  return __builtin_spu_rlqw_6 (a, b);
-}
-static inline vec_llong2
-spu_rlqw (vec_llong2 a, int b)
-{
-  return __builtin_spu_rlqw_7 (a, b);
-}
-static inline vec_float4
-spu_rlqw (vec_float4 a, int b)
-{
-  return __builtin_spu_rlqw_8 (a, b);
-}
-static inline vec_double2
-spu_rlqw (vec_double2 a, int b)
-{
-  return __builtin_spu_rlqw_9 (a, b);
-}
-static inline vec_uchar16
-spu_rlqwbyte (vec_uchar16 a, int b)
-{
-  return __builtin_spu_rlqwbyte_0 (a, b);
-}
-static inline vec_char16
-spu_rlqwbyte (vec_char16 a, int b)
-{
-  return __builtin_spu_rlqwbyte_1 (a, b);
-}
-static inline vec_ushort8
-spu_rlqwbyte (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlqwbyte_2 (a, b);
-}
-static inline vec_short8
-spu_rlqwbyte (vec_short8 a, int b)
-{
-  return __builtin_spu_rlqwbyte_3 (a, b);
-}
-static inline vec_uint4
-spu_rlqwbyte (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlqwbyte_4 (a, b);
-}
-static inline vec_int4
-spu_rlqwbyte (vec_int4 a, int b)
-{
-  return __builtin_spu_rlqwbyte_5 (a, b);
-}
-static inline vec_ullong2
-spu_rlqwbyte (vec_ullong2 a, int b)
-{
-  return __builtin_spu_rlqwbyte_6 (a, b);
-}
-static inline vec_llong2
-spu_rlqwbyte (vec_llong2 a, int b)
-{
-  return __builtin_spu_rlqwbyte_7 (a, b);
-}
-static inline vec_float4
-spu_rlqwbyte (vec_float4 a, int b)
-{
-  return __builtin_spu_rlqwbyte_8 (a, b);
-}
-static inline vec_double2
-spu_rlqwbyte (vec_double2 a, int b)
-{
-  return __builtin_spu_rlqwbyte_9 (a, b);
-}
-static inline vec_uchar16
-spu_rlqwbytebc (vec_uchar16 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_0 (a, b);
-}
-static inline vec_char16
-spu_rlqwbytebc (vec_char16 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_1 (a, b);
-}
-static inline vec_ushort8
-spu_rlqwbytebc (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_2 (a, b);
-}
-static inline vec_short8
-spu_rlqwbytebc (vec_short8 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_3 (a, b);
-}
-static inline vec_uint4
-spu_rlqwbytebc (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_4 (a, b);
-}
-static inline vec_int4
-spu_rlqwbytebc (vec_int4 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_5 (a, b);
-}
-static inline vec_ullong2
-spu_rlqwbytebc (vec_ullong2 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_6 (a, b);
-}
-static inline vec_llong2
-spu_rlqwbytebc (vec_llong2 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_7 (a, b);
-}
-static inline vec_float4
-spu_rlqwbytebc (vec_float4 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_8 (a, b);
-}
-static inline vec_double2
-spu_rlqwbytebc (vec_double2 a, int b)
-{
-  return __builtin_spu_rlqwbytebc_9 (a, b);
-}
-static inline vec_ushort8
-spu_rlmask (vec_ushort8 a, vec_short8 b)
-{
-  return __builtin_spu_rlmask_0 (a, b);
-}
-static inline vec_short8
-spu_rlmask (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_rlmask_1 (a, b);
-}
-static inline vec_uint4
-spu_rlmask (vec_uint4 a, vec_int4 b)
-{
-  return __builtin_spu_rlmask_2 (a, b);
-}
-static inline vec_int4
-spu_rlmask (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_rlmask_3 (a, b);
-}
-static inline vec_ushort8
-spu_rlmask (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlmask_4 (a, b);
-}
-static inline vec_short8
-spu_rlmask (vec_short8 a, int b)
-{
-  return __builtin_spu_rlmask_5 (a, b);
-}
-static inline vec_uint4
-spu_rlmask (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlmask_6 (a, b);
-}
-static inline vec_int4
-spu_rlmask (vec_int4 a, int b)
-{
-  return __builtin_spu_rlmask_7 (a, b);
-}
-static inline vec_ushort8
-spu_rlmaska (vec_ushort8 a, vec_short8 b)
-{
-  return __builtin_spu_rlmaska_0 (a, b);
-}
-static inline vec_short8
-spu_rlmaska (vec_short8 a, vec_short8 b)
-{
-  return __builtin_spu_rlmaska_1 (a, b);
-}
-static inline vec_uint4
-spu_rlmaska (vec_uint4 a, vec_int4 b)
-{
-  return __builtin_spu_rlmaska_2 (a, b);
-}
-static inline vec_int4
-spu_rlmaska (vec_int4 a, vec_int4 b)
-{
-  return __builtin_spu_rlmaska_3 (a, b);
-}
-static inline vec_ushort8
-spu_rlmaska (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlmaska_4 (a, b);
-}
-static inline vec_short8
-spu_rlmaska (vec_short8 a, int b)
-{
-  return __builtin_spu_rlmaska_5 (a, b);
-}
-static inline vec_uint4
-spu_rlmaska (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlmaska_6 (a, b);
-}
-static inline vec_int4
-spu_rlmaska (vec_int4 a, int b)
-{
-  return __builtin_spu_rlmaska_7 (a, b);
-}
-static inline vec_uchar16
-spu_rlmaskqw (vec_uchar16 a, int b)
-{
-  return __builtin_spu_rlmaskqw_0 (a, b);
-}
-static inline vec_char16
-spu_rlmaskqw (vec_char16 a, int b)
-{
-  return __builtin_spu_rlmaskqw_1 (a, b);
-}
-static inline vec_ushort8
-spu_rlmaskqw (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlmaskqw_2 (a, b);
-}
-static inline vec_short8
-spu_rlmaskqw (vec_short8 a, int b)
-{
-  return __builtin_spu_rlmaskqw_3 (a, b);
-}
-static inline vec_uint4
-spu_rlmaskqw (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlmaskqw_4 (a, b);
-}
-static inline vec_int4
-spu_rlmaskqw (vec_int4 a, int b)
-{
-  return __builtin_spu_rlmaskqw_5 (a, b);
-}
-static inline vec_ullong2
-spu_rlmaskqw (vec_ullong2 a, int b)
-{
-  return __builtin_spu_rlmaskqw_6 (a, b);
-}
-static inline vec_llong2
-spu_rlmaskqw (vec_llong2 a, int b)
-{
-  return __builtin_spu_rlmaskqw_7 (a, b);
-}
-static inline vec_float4
-spu_rlmaskqw (vec_float4 a, int b)
-{
-  return __builtin_spu_rlmaskqw_8 (a, b);
-}
-static inline vec_double2
-spu_rlmaskqw (vec_double2 a, int b)
-{
-  return __builtin_spu_rlmaskqw_9 (a, b);
-}
-static inline vec_uchar16
-spu_rlmaskqwbyte (vec_uchar16 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_0 (a, b);
-}
-static inline vec_char16
-spu_rlmaskqwbyte (vec_char16 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_1 (a, b);
-}
-static inline vec_ushort8
-spu_rlmaskqwbyte (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_2 (a, b);
-}
-static inline vec_short8
-spu_rlmaskqwbyte (vec_short8 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_3 (a, b);
-}
-static inline vec_uint4
-spu_rlmaskqwbyte (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_4 (a, b);
-}
-static inline vec_int4
-spu_rlmaskqwbyte (vec_int4 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_5 (a, b);
-}
-static inline vec_ullong2
-spu_rlmaskqwbyte (vec_ullong2 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_6 (a, b);
-}
-static inline vec_llong2
-spu_rlmaskqwbyte (vec_llong2 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_7 (a, b);
-}
-static inline vec_float4
-spu_rlmaskqwbyte (vec_float4 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_8 (a, b);
-}
-static inline vec_double2
-spu_rlmaskqwbyte (vec_double2 a, int b)
-{
-  return __builtin_spu_rlmaskqwbyte_9 (a, b);
-}
-static inline vec_uchar16
-spu_rlmaskqwbytebc (vec_uchar16 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_0 (a, b);
-}
-static inline vec_char16
-spu_rlmaskqwbytebc (vec_char16 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_1 (a, b);
-}
-static inline vec_ushort8
-spu_rlmaskqwbytebc (vec_ushort8 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_2 (a, b);
-}
-static inline vec_short8
-spu_rlmaskqwbytebc (vec_short8 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_3 (a, b);
-}
-static inline vec_uint4
-spu_rlmaskqwbytebc (vec_uint4 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_4 (a, b);
-}
-static inline vec_int4
-spu_rlmaskqwbytebc (vec_int4 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_5 (a, b);
-}
-static inline vec_ullong2
-spu_rlmaskqwbytebc (vec_ullong2 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_6 (a, b);
-}
-static inline vec_llong2
-spu_rlmaskqwbytebc (vec_llong2 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_7 (a, b);
-}
-static inline vec_float4
-spu_rlmaskqwbytebc (vec_float4 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_8 (a, b);
-}
-static inline vec_double2
-spu_rlmaskqwbytebc (vec_double2 a, int b)
-{
-  return __builtin_spu_rlmaskqwbytebc_9 (a, b);
-}
-static inline vec_ushort8
-spu_sl (vec_ushort8 a, vec_ushort8 b)
-{
-  return __builtin_spu_sl_0 (a, b);
-}
-static inline vec_short8
-spu_sl (vec_short8 a, vec_ushort8 b)
-{
-  return __builtin_spu_sl_1 (a, b);
-}
-static inline vec_uint4
-spu_sl (vec_uint4 a, vec_uint4 b)
-{
-  return __builtin_spu_sl_2 (a, b);
-}
-static inline vec_int4
-spu_sl (vec_int4 a, vec_uint4 b)
-{
-  return __builtin_spu_sl_3 (a, b);
-}
-static inline vec_ushort8
-spu_sl (vec_ushort8 a, unsigned int b)
-{
-  return __builtin_spu_sl_4 (a, b);
-}
-static inline vec_short8
-spu_sl (vec_short8 a, unsigned int b)
-{
-  return __builtin_spu_sl_5 (a, b);
-}
-static inline vec_uint4
-spu_sl (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_sl_6 (a, b);
-}
-static inline vec_int4
-spu_sl (vec_int4 a, unsigned int b)
-{
-  return __builtin_spu_sl_7 (a, b);
-}
-static inline vec_llong2
-spu_slqw (vec_llong2 a, unsigned int b)
-{
-  return __builtin_spu_slqw_0 (a, b);
-}
-static inline vec_ullong2
-spu_slqw (vec_ullong2 a, unsigned int b)
-{
-  return __builtin_spu_slqw_1 (a, b);
-}
-static inline vec_int4
-spu_slqw (vec_int4 a, unsigned int b)
-{
-  return __builtin_spu_slqw_2 (a, b);
-}
-static inline vec_uint4
-spu_slqw (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_slqw_3 (a, b);
-}
-static inline vec_short8
-spu_slqw (vec_short8 a, unsigned int b)
-{
-  return __builtin_spu_slqw_4 (a, b);
-}
-static inline vec_ushort8
-spu_slqw (vec_ushort8 a, unsigned int b)
-{
-  return __builtin_spu_slqw_5 (a, b);
-}
-static inline vec_char16
-spu_slqw (vec_char16 a, unsigned int b)
-{
-  return __builtin_spu_slqw_6 (a, b);
-}
-static inline vec_uchar16
-spu_slqw (vec_uchar16 a, unsigned int b)
-{
-  return __builtin_spu_slqw_7 (a, b);
-}
-static inline vec_float4
-spu_slqw (vec_float4 a, unsigned int b)
-{
-  return __builtin_spu_slqw_8 (a, b);
-}
-static inline vec_double2
-spu_slqw (vec_double2 a, unsigned int b)
-{
-  return __builtin_spu_slqw_9 (a, b);
-}
-static inline vec_llong2
-spu_slqwbyte (vec_llong2 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_0 (a, b);
-}
-static inline vec_ullong2
-spu_slqwbyte (vec_ullong2 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_1 (a, b);
-}
-static inline vec_int4
-spu_slqwbyte (vec_int4 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_2 (a, b);
-}
-static inline vec_uint4
-spu_slqwbyte (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_3 (a, b);
-}
-static inline vec_short8
-spu_slqwbyte (vec_short8 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_4 (a, b);
-}
-static inline vec_ushort8
-spu_slqwbyte (vec_ushort8 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_5 (a, b);
-}
-static inline vec_char16
-spu_slqwbyte (vec_char16 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_6 (a, b);
-}
-static inline vec_uchar16
-spu_slqwbyte (vec_uchar16 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_7 (a, b);
-}
-static inline vec_float4
-spu_slqwbyte (vec_float4 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_8 (a, b);
-}
-static inline vec_double2
-spu_slqwbyte (vec_double2 a, unsigned int b)
-{
-  return __builtin_spu_slqwbyte_9 (a, b);
-}
-static inline vec_llong2
-spu_slqwbytebc (vec_llong2 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_0 (a, b);
-}
-static inline vec_ullong2
-spu_slqwbytebc (vec_ullong2 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_1 (a, b);
-}
-static inline vec_int4
-spu_slqwbytebc (vec_int4 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_2 (a, b);
-}
-static inline vec_uint4
-spu_slqwbytebc (vec_uint4 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_3 (a, b);
-}
-static inline vec_short8
-spu_slqwbytebc (vec_short8 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_4 (a, b);
-}
-static inline vec_ushort8
-spu_slqwbytebc (vec_ushort8 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_5 (a, b);
-}
-static inline vec_char16
-spu_slqwbytebc (vec_char16 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_6 (a, b);
-}
-static inline vec_uchar16
-spu_slqwbytebc (vec_uchar16 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_7 (a, b);
-}
-static inline vec_float4
-spu_slqwbytebc (vec_float4 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_8 (a, b);
-}
-static inline vec_double2
-spu_slqwbytebc (vec_double2 a, unsigned int b)
-{
-  return __builtin_spu_slqwbytebc_9 (a, b);
-}
-static inline vec_uchar16
-spu_splats (unsigned char a)
-{
-  return __builtin_spu_splats_0 (a);
-}
-static inline vec_char16
-spu_splats (signed char a)
-{
-  return __builtin_spu_splats_1 (a);
-}
-static inline vec_char16
-spu_splats (char a)
-{
-  return __builtin_spu_splats_1 (a);
-}
-static inline vec_ushort8
-spu_splats (unsigned short a)
-{
-  return __builtin_spu_splats_2 (a);
-}
-static inline vec_short8
-spu_splats (short a)
-{
-  return __builtin_spu_splats_3 (a);
-}
-static inline vec_uint4
-spu_splats (unsigned int a)
-{
-  return __builtin_spu_splats_4 (a);
-}
-static inline vec_int4
-spu_splats (int a)
-{
-  return __builtin_spu_splats_5 (a);
-}
-static inline vec_ullong2
-spu_splats (unsigned long long a)
-{
-  return __builtin_spu_splats_6 (a);
-}
-static inline vec_llong2
-spu_splats (long long a)
-{
-  return __builtin_spu_splats_7 (a);
-}
-static inline vec_float4
-spu_splats (float a)
-{
-  return __builtin_spu_splats_8 (a);
-}
-static inline vec_double2
-spu_splats (double a)
-{
-  return __builtin_spu_splats_9 (a);
-}
-static inline unsigned char
-spu_extract (vec_uchar16 a, int b)
-{
-  return __builtin_spu_extract_0 (a, b);
-}
-static inline signed char
-spu_extract (vec_char16 a, int b)
-{
-  return __builtin_spu_extract_1 (a, b);
-}
-static inline unsigned short
-spu_extract (vec_ushort8 a, int b)
-{
-  return __builtin_spu_extract_2 (a, b);
-}
-static inline short
-spu_extract (vec_short8 a, int b)
-{
-  return __builtin_spu_extract_3 (a, b);
-}
-static inline unsigned int
-spu_extract (vec_uint4 a, int b)
-{
-  return __builtin_spu_extract_4 (a, b);
-}
-static inline int
-spu_extract (vec_int4 a, int b)
-{
-  return __builtin_spu_extract_5 (a, b);
-}
-static inline unsigned long long
-spu_extract (vec_ullong2 a, int b)
-{
-  return __builtin_spu_extract_6 (a, b);
-}
-static inline long long
-spu_extract (vec_llong2 a, int b)
-{
-  return __builtin_spu_extract_7 (a, b);
-}
-static inline float
-spu_extract (vec_float4 a, int b)
-{
-  return __builtin_spu_extract_8 (a, b);
-}
-static inline double
-spu_extract (vec_double2 a, int b)
-{
-  return __builtin_spu_extract_9 (a, b);
-}
-static inline vec_uchar16
-spu_insert (unsigned char a, vec_uchar16 b, int c)
-{
-  return __builtin_spu_insert_0 (a, b, c);
-}
-static inline vec_char16
-spu_insert (signed char a, vec_char16 b, int c)
-{
-  return __builtin_spu_insert_1 (a, b, c);
-}
-static inline vec_ushort8
-spu_insert (unsigned short a, vec_ushort8 b, int c)
-{
-  return __builtin_spu_insert_2 (a, b, c);
-}
-static inline vec_short8
-spu_insert (short a, vec_short8 b, int c)
-{
-  return __builtin_spu_insert_3 (a, b, c);
-}
-static inline vec_uint4
-spu_insert (unsigned int a, vec_uint4 b, int c)
-{
-  return __builtin_spu_insert_4 (a, b, c);
-}
-static inline vec_int4
-spu_insert (int a, vec_int4 b, int c)
-{
-  return __builtin_spu_insert_5 (a, b, c);
-}
-static inline vec_ullong2
-spu_insert (unsigned long long a, vec_ullong2 b, int c)
-{
-  return __builtin_spu_insert_6 (a, b, c);
-}
-static inline vec_llong2
-spu_insert (long long a, vec_llong2 b, int c)
-{
-  return __builtin_spu_insert_7 (a, b, c);
-}
-static inline vec_float4
-spu_insert (float a, vec_float4 b, int c)
-{
-  return __builtin_spu_insert_8 (a, b, c);
-}
-static inline vec_double2
-spu_insert (double a, vec_double2 b, int c)
-{
-  return __builtin_spu_insert_9 (a, b, c);
-}
-static inline vec_uchar16
-spu_promote (unsigned char a, int b)
-{
-  return __builtin_spu_promote_0 (a, b);
-}
-static inline vec_char16
-spu_promote (signed char a, int b)
-{
-  return __builtin_spu_promote_1 (a, b);
-}
-static inline vec_char16
-spu_promote (char a, int b)
-{
-  return __builtin_spu_promote_1 (a, b);
-}
-static inline vec_ushort8
-spu_promote (unsigned short a, int b)
-{
-  return __builtin_spu_promote_2 (a, b);
-}
-static inline vec_short8
-spu_promote (short a, int b)
-{
-  return __builtin_spu_promote_3 (a, b);
-}
-static inline vec_uint4
-spu_promote (unsigned int a, int b)
-{
-  return __builtin_spu_promote_4 (a, b);
-}
-static inline vec_int4
-spu_promote (int a, int b)
-{
-  return __builtin_spu_promote_5 (a, b);
-}
-static inline vec_ullong2
-spu_promote (unsigned long long a, int b)
-{
-  return __builtin_spu_promote_6 (a, b);
-}
-static inline vec_llong2
-spu_promote (long long a, int b)
-{
-  return __builtin_spu_promote_7 (a, b);
-}
-static inline vec_float4
-spu_promote (float a, int b)
-{
-  return __builtin_spu_promote_8 (a, b);
-}
-static inline vec_double2
-spu_promote (double a, int b)
-{
-  return __builtin_spu_promote_9 (a, b);
-}
-#endif  /* __cplusplus */
-
 #ifdef __cplusplus
 extern "C" {
 #endif