i386.c (builtin_description): Add IX86_BUILTIN_PUNPCKHQDQ128.

	* i386.c (builtin_description): Add IX86_BUILTIN_PUNPCKHQDQ128.
	(ix86_expand_builtin): Fix MASKMOVDQU expasion.
	* i386.h (ix86_builtins): Add IX86_BUILTIN_PUNPCKHQDQ128.
	* i386.md (mmx_punpck?dq): Simplify.
	(sse2_pubpcklqdq): Fix.
	(sse2_pubpckhqdq): New.
	* xmmintrin.h (_mm_unpackhi_epi32): New.

	* xmmintrin.h (_mm_cvt*, _mm_stream_pd): Fix prototypes.
	(_mm_shufflehi_epi16, _mm_shufflelo_epi16): Fix typo.

From-SVN: r58412

5 files changed, 61 insertions, 30 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 2d85ff9..c2f7fbd 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,16 @@
+Tue Oct 22 19:07:03 CEST 2002  Jan Hubicka  <jh@suse.cz>
+
+	* i386.c (builtin_description): Add IX86_BUILTIN_PUNPCKHQDQ128.
+	(ix86_expand_builtin): Fix MASKMOVDQU expasion.
+	* i386.h (ix86_builtins): Add IX86_BUILTIN_PUNPCKHQDQ128.
+	* i386.md (mmx_punpck?dq): Simplify.
+	(sse2_pubpcklqdq): Fix.
+	(sse2_pubpckhqdq): New.
+	* xmmintrin.h (_mm_unpackhi_epi32): New.
+
+	* xmmintrin.h (_mm_cvt*, _mm_stream_pd): Fix prototypes.
+	(_mm_shufflehi_epi16, _mm_shufflelo_epi16): Fix typo.
+
 2002-10-22  Nathan Sidwell  <nathan@codesourcery.com>
 
 	PR c++/7209
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 2b857bf..537aae8 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -12115,6 +12115,7 @@ static const struct builtin_description bdesc_2arg[] =
   { MASK_SSE2, CODE_FOR_sse2_punpckhbw, "__builtin_ia32_punpckhbw128", IX86_BUILTIN_PUNPCKHBW128, 0, 0 },
   { MASK_SSE2, CODE_FOR_sse2_punpckhwd, "__builtin_ia32_punpckhwd128", IX86_BUILTIN_PUNPCKHWD128, 0, 0 },
   { MASK_SSE2, CODE_FOR_sse2_punpckhdq, "__builtin_ia32_punpckhdq128", IX86_BUILTIN_PUNPCKHDQ128, 0, 0 },
+  { MASK_SSE2, CODE_FOR_sse2_punpckhqdq, "__builtin_ia32_punpckhqdq128", IX86_BUILTIN_PUNPCKHQDQ128, 0, 0 },
   { MASK_SSE2, CODE_FOR_sse2_punpcklbw, "__builtin_ia32_punpcklbw128", IX86_BUILTIN_PUNPCKLBW128, 0, 0 },
   { MASK_SSE2, CODE_FOR_sse2_punpcklwd, "__builtin_ia32_punpcklwd128", IX86_BUILTIN_PUNPCKLWD128, 0, 0 },
   { MASK_SSE2, CODE_FOR_sse2_punpckldq, "__builtin_ia32_punpckldq128", IX86_BUILTIN_PUNPCKLDQ128, 0, 0 },
@@ -13145,6 +13146,7 @@ ix86_expand_builtin (exp, target, subtarget, mode, ignore)
       return target;
 
     case IX86_BUILTIN_MASKMOVQ:
+    case IX86_BUILTIN_MASKMOVDQU:
       icode = (fcode == IX86_BUILTIN_MASKMOVQ
 	       ? (TARGET_64BIT ? CODE_FOR_mmx_maskmovq_rex : CODE_FOR_mmx_maskmovq)
 	       : CODE_FOR_sse2_maskmovdqu);
diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h
index f832e10..aabf3e5 100644
--- a/gcc/config/i386/i386.h
+++ b/gcc/config/i386/i386.h
@@ -2427,6 +2427,7 @@ enum ix86_builtins
   IX86_BUILTIN_PUNPCKHBW128,
   IX86_BUILTIN_PUNPCKHWD128,
   IX86_BUILTIN_PUNPCKHDQ128,
+  IX86_BUILTIN_PUNPCKHQDQ128,
   IX86_BUILTIN_PUNPCKLBW128,
   IX86_BUILTIN_PUNPCKLWD128,
   IX86_BUILTIN_PUNPCKLDQ128,
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index a499131..13a8c21 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -19694,9 +19694,7 @@
 (define_insn "mmx_punpckhdq"
   [(set (match_operand:V2SI 0 "register_operand" "=y")
 	(vec_merge:V2SI
-	 (vec_select:V2SI (match_operand:V2SI 1 "register_operand" "0")
-			  (parallel [(const_int 0)
-				     (const_int 1)]))
+	 (match_operand:V2SI 1 "register_operand" "0")
 	 (vec_select:V2SI (match_operand:V2SI 2 "register_operand" "y")
 			  (parallel [(const_int 1)
 				     (const_int 0)]))
@@ -19758,9 +19756,7 @@
 	 (vec_select:V2SI (match_operand:V2SI 1 "register_operand" "0")
 			   (parallel [(const_int 1)
 				      (const_int 0)]))
-	 (vec_select:V2SI (match_operand:V2SI 2 "register_operand" "y")
-			  (parallel [(const_int 0)
-				     (const_int 1)]))
+	 (match_operand:V2SI 2 "register_operand" "y")
 	 (const_int 1)))]
   "TARGET_MMX"
   "punpckldq\t{%2, %0|%0, %2}"
@@ -21548,16 +21544,29 @@
 (define_insn "sse2_punpcklqdq"
   [(set (match_operand:V2DI 0 "register_operand" "=x")
 	(vec_merge:V2DI
-	 (match_operand:V2DI 1 "register_operand" "0")
 	 (vec_select:V2DI (match_operand:V2DI 2 "register_operand" "x")
 			  (parallel [(const_int 1)
 				     (const_int 0)]))
+	 (match_operand:V2DI 1 "register_operand" "0")
 	 (const_int 1)))]
   "TARGET_SSE2"
   "punpcklqdq\t{%2, %0|%0, %2}"
   [(set_attr "type" "ssecvt")
    (set_attr "mode" "TI")])
 
+(define_insn "sse2_punpckhqdq"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+	(vec_merge:V2DI
+	 (match_operand:V2DI 1 "register_operand" "0")
+	 (vec_select:V2DI (match_operand:V2DI 2 "register_operand" "x")
+			  (parallel [(const_int 1)
+				     (const_int 0)]))
+	 (const_int 1)))]
+  "TARGET_SSE2"
+  "punpckhqdq\t{%2, %0|%0, %2}"
+  [(set_attr "type" "ssecvt")
+   (set_attr "mode" "TI")])
+
 ;; SSE2 moves
 
 (define_insn "sse2_movapd"
diff --git a/gcc/config/i386/xmmintrin.h b/gcc/config/i386/xmmintrin.h
index 08ef76a..14fdcd5 100644
--- a/gcc/config/i386/xmmintrin.h
+++ b/gcc/config/i386/xmmintrin.h
@@ -1827,16 +1827,16 @@ _mm_cvtepi32_pd (__m128i __A)
   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 }
 
-static __inline __m128d
+static __inline __m128
 _mm_cvtepi32_ps (__m128i __A)
 {
-  return (__m128d)__builtin_ia32_cvtdq2ps ((__v4si) __A);
+  return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 }
 
-static __inline __m128d
+static __inline __m128i
 _mm_cvtpd_epi32 (__m128d __A)
 {
-  return (__m128d)__builtin_ia32_cvtpd2dq ((__v2df) __A);
+  return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 }
 
 static __inline __m64
@@ -1845,16 +1845,16 @@ _mm_cvtpd_pi32 (__m128d __A)
   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 }
 
-static __inline __m128d
+static __inline __m128
 _mm_cvtpd_ps (__m128d __A)
 {
-  return (__m128d)__builtin_ia32_cvtpd2ps ((__v2df) __A);
+  return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 }
 
-static __inline __m128d
+static __inline __m128i
 _mm_cvttpd_epi32 (__m128d __A)
 {
-  return (__m128d)__builtin_ia32_cvttpd2dq ((__v2df) __A);
+  return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 }
 
 static __inline __m64
@@ -1869,20 +1869,20 @@ _mm_cvtpi32_pd (__m64 __A)
   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 }
 
-static __inline __m128d
-_mm_cvtps_epi32 (__m128d __A)
+static __inline __m128i
+_mm_cvtps_epi32 (__m128 __A)
 {
-  return (__m128d)__builtin_ia32_cvtps2dq ((__v4sf) __A);
+  return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 }
 
-static __inline __m128d
-_mm_cvttps_epi32 (__m128d __A)
+static __inline __m128i
+_mm_cvttps_epi32 (__m128 __A)
 {
-  return (__m128d)__builtin_ia32_cvttps2dq ((__v4sf) __A);
+  return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 }
 
 static __inline __m128d
-_mm_cvtps_pd (__m128d __A)
+_mm_cvtps_pd (__m128 __A)
 {
   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 }
@@ -1899,10 +1899,10 @@ _mm_cvttsd_si32 (__m128d __A)
   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 }
 
-static __inline __m128d
-_mm_cvtsd_ss (__m128d __A, __m128d __B)
+static __inline __m128
+_mm_cvtsd_ss (__m128 __A, __m128d __B)
 {
-  return (__m128d)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
+  return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 }
 
 static __inline __m128d
@@ -1912,7 +1912,7 @@ _mm_cvtsi32_sd (__m128d __A, int __B)
 }
 
 static __inline __m128d
-_mm_cvtss_sd (__m128d __A, __m128d __B)
+_mm_cvtss_sd (__m128d __A, __m128 __B)
 {
   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 }
@@ -1998,6 +1998,12 @@ _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 }
 
 static __inline __m128i
+_mm_unpackhi_epi64 (__m128i __A, __m128i __B)
+{
+  return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
+}
+
+static __inline __m128i
 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 {
   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
@@ -2359,8 +2365,8 @@ _mm_mulhi_epu16 (__m128i __A, __m128i __B)
   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
 }
 
-#define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw128 ((__v8hi)__A, __B))
-#define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw128 ((__v8hi)__A, __B))
+#define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))
+#define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))
 #define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))
 
 static __inline void
@@ -2400,9 +2406,9 @@ _mm_stream_si128 (__m128i *__A, __m128i __B)
 }
 
 static __inline void
-_mm_stream_pd (__m128d *__A, __m128d __B)
+_mm_stream_pd (double *__A, __m128d __B)
 {
-  __builtin_ia32_movntpd (__A, (__v2df)__B);
+  __builtin_ia32_movntpd ((__m128d *)__A, (__v2df)__B);
 }
 
 static __inline __m128i