ia64.c (rtx_needs_barrier): Handle CONST_VECTOR and VEC_SELECT.

        * config/ia64/ia64.c (rtx_needs_barrier): Handle CONST_VECTOR
        and VEC_SELECT.
        * config/ia64/vect.md (mulv8qi3): Re-implement with mix_[rl].
        (mulv4hi3): Set itanium_class mmmul.
        (fpack_sfxf, fpack_xfsf, fpack_xfxf): Remove.

From-SVN: r92987

4 files changed, 83 insertions, 74 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 078a19e..48586a2 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,5 +1,14 @@
 2005-01-05  Richard Henderson  <rth@redhat.com>
 
+	* config/ia64/ia64.c (rtx_needs_barrier): Handle CONST_VECTOR
+	and VEC_SELECT.
+	* config/ia64/vect.md (mulv8qi3): Re-implement with mix_[rl].
+	(mulv4hi3): Set itanium_class mmmul.
+	(fpack_sfxf, fpack_xfsf, fpack_xfxf): Remove.
+	(fpack): Rename from fpack_sfsf.
+
+2005-01-05  Richard Henderson  <rth@redhat.com>
+
 	PR rtl-opt/10692
 	* reload1.c (do_input_reload): Restrict the optimization deleteing
 	a previous output reload to RELOAD_FOR_INPUT.
diff --git a/gcc/config/ia64/ia64.c b/gcc/config/ia64/ia64.c
index a6197c0..ce5d8aa 100644
--- a/gcc/config/ia64/ia64.c
+++ b/gcc/config/ia64/ia64.c
@@ -5198,17 +5198,27 @@ rtx_needs_barrier (rtx x, struct reg_flags flags, int pred)
       for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
 	{
 	  rtx pat = XVECEXP (x, 0, i);
-	  if (GET_CODE (pat) == SET)
+	  switch (GET_CODE (pat))
 	    {
+	    case SET:
 	      update_set_flags (pat, &new_flags, &pred, &cond);
-	      need_barrier |= set_src_needs_barrier (pat, new_flags, pred, cond);
+	      need_barrier |= set_src_needs_barrier (pat, new_flags,
+						     pred, cond);
+	      break;
+
+	    case USE:
+	    case CALL:
+	    case ASM_OPERANDS:
+	      need_barrier |= rtx_needs_barrier (pat, flags, pred);
+	      break;
+
+	    case CLOBBER:
+	    case RETURN:
+	      break;
+
+	    default:
+	      gcc_unreachable ();
 	    }
-	  else if (GET_CODE (pat) == USE
-		   || GET_CODE (pat) == CALL
-		   || GET_CODE (pat) == ASM_OPERANDS)
-	    need_barrier |= rtx_needs_barrier (pat, flags, pred);
-	  else if (GET_CODE (pat) != CLOBBER && GET_CODE (pat) != RETURN)
-	    abort ();
 	}
       for (i = XVECLEN (x, 0) - 1; i >= 0; --i)
 	{
@@ -5246,7 +5256,7 @@ rtx_needs_barrier (rtx x, struct reg_flags flags, int pred)
       need_barrier = rtx_needs_barrier (XEXP (x, 0), new_flags, pred);
       break;
 
-    case CONST_INT:   case CONST_DOUBLE:
+    case CONST_INT:   case CONST_DOUBLE:  case CONST_VECTOR:
     case SYMBOL_REF:  case LABEL_REF:     case CONST:
       break;
 
@@ -5290,6 +5300,14 @@ rtx_needs_barrier (rtx x, struct reg_flags flags, int pred)
       need_barrier = rtx_needs_barrier (XEXP (x, 0), flags, pred);
       break;
 
+    case VEC_SELECT:
+      /* VEC_SELECT's second argument is a PARALLEL with integers that
+	 describe the elements selected.  On ia64, those integers are
+	 always constants.  Avoid walking the PARALLEL so that we don't
+	 get confused with "normal" parallels and abort.  */
+      need_barrier = rtx_needs_barrier (XEXP (x, 0), flags, pred);
+      break;
+
     case UNSPEC:
       switch (XINT (x, 1))
 	{
diff --git a/gcc/config/ia64/vect.md b/gcc/config/ia64/vect.md
index a42ad2a..12729a9 100644
--- a/gcc/config/ia64/vect.md
+++ b/gcc/config/ia64/vect.md
@@ -172,43 +172,35 @@
 		   (match_operand:V8QI 2 "gr_register_operand" "r")))]
   ""
 {
-  rtx l1, h1, l2, h2, lm, hm, lz, hz;
+  rtx r1, l1, r2, l2, rm, lm;
 
+  r1 = gen_reg_rtx (V4HImode);
   l1 = gen_reg_rtx (V4HImode);
-  h1 = gen_reg_rtx (V4HImode);
+  r2 = gen_reg_rtx (V4HImode);
   l2 = gen_reg_rtx (V4HImode);
-  h2 = gen_reg_rtx (V4HImode);
-
-  /* Zero-extend the QImode elements into two words of HImode elements.  */
-  emit_insn (gen_unpack1_l (gen_lowpart (V8QImode, l1),
-			    operands[1], CONST0_RTX (V8QImode)));
-  emit_insn (gen_unpack1_l (gen_lowpart (V8QImode, l2),
-			    operands[2], CONST0_RTX (V8QImode)));
-  emit_insn (gen_unpack1_h (gen_lowpart (V8QImode, h1),
-			    operands[1], CONST0_RTX (V8QImode)));
-  emit_insn (gen_unpack1_h (gen_lowpart (V8QImode, h2),
-			    operands[2], CONST0_RTX (V8QImode)));
+
+  /* Zero-extend the QImode elements into two words of HImode elements
+     by interleaving them with zero bytes.  */
+  emit_insn (gen_mix1_r (gen_lowpart (V8QImode, r1),
+                         operands[1], CONST0_RTX (V8QImode)));
+  emit_insn (gen_mix1_r (gen_lowpart (V8QImode, r2),
+                         operands[2], CONST0_RTX (V8QImode)));
+  emit_insn (gen_mix1_l (gen_lowpart (V8QImode, l1),
+                         operands[1], CONST0_RTX (V8QImode)));
+  emit_insn (gen_mix1_l (gen_lowpart (V8QImode, l2),
+                         operands[2], CONST0_RTX (V8QImode)));
 
   /* Multiply.  */
+  rm = gen_reg_rtx (V4HImode);
   lm = gen_reg_rtx (V4HImode);
-  hm = gen_reg_rtx (V4HImode);
+  emit_insn (gen_mulv4hi3 (rm, r1, r2));
   emit_insn (gen_mulv4hi3 (lm, l1, l2));
-  emit_insn (gen_mulv4hi3 (hm, h1, h2));
-
-  /* Zap the high order bytes of the HImode elements.  There are several
-     ways that this could be done.  On Itanium2, there's 1 cycle latency
-     moving between the ALU units and the PALU units, so using AND would
-     be 3 cycles latency into the eventual pack insn, whereas using MIX
-     is only 2 cycles.  */
-  lz = gen_reg_rtx (V4HImode);
-  hz = gen_reg_rtx (V4HImode);
-  emit_insn (gen_mix1_r (gen_lowpart (V8QImode, lz),
-			 gen_lowpart (V8QImode, lm), CONST0_RTX (V8QImode)));
-  emit_insn (gen_mix1_r (gen_lowpart (V8QImode, lz),
-			 gen_lowpart (V8QImode, lm), CONST0_RTX (V8QImode)));
-
-  /* Repack the HImode elements as QImode elements.  */
-  emit_insn (gen_pack2_sss (operands[0], lz, hz));
+
+  /* Zap the high order bytes of the HImode elements by overwriting those
+     in one part with the low order bytes of the other.  */
+  emit_insn (gen_mix1_r (operands[0],
+                         gen_lowpart (V8QImode, rm),
+                         gen_lowpart (V8QImode, lm)));
   DONE;
 })
 
@@ -218,7 +210,7 @@
 		   (match_operand:V4HI 2 "gr_register_operand" "r")))]
   ""
   "pmpyshr2 %0 = %1, %2, 0"
-  [(set_attr "itanium_class" "mmalua")])
+  [(set_attr "itanium_class" "mmmul")])
 
 (define_expand "umax<mode>3"
   [(set (match_operand:VECINT 0 "gr_register_operand" "")
@@ -450,7 +442,7 @@
   "mix1.r %0 = %r2, %r1"
   [(set_attr "itanium_class" "mmshf")])
 
-(define_insn "*mix1_l"
+(define_insn "mix1_l"
   [(set (match_operand:V8QI 0 "gr_register_operand" "=r")
 	(vec_select:V8QI
 	  (vec_concat:V16QI
@@ -948,7 +940,7 @@
   DONE;
 })
 
-(define_insn "*fpack_sfsf"
+(define_insn "*fpack"
   [(set (match_operand:V2SF 0 "fr_register_operand" "=f")
 	(vec_concat:V2SF
 	  (match_operand:SF 1 "fr_reg_or_fp01_operand" "fG")
@@ -957,38 +949,6 @@
   "fpack %0 = %F2, %F1"
   [(set_attr "itanium_class" "fmisc")])
 
-(define_insn "*fpack_sfxf"
-  [(set (match_operand:V2SF 0 "fr_register_operand" "=f")
-	(vec_concat:V2SF
-	  (match_operand:SF 1 "fr_reg_or_fp01_operand" "fG")
-	  (float_truncate:SF
-	    (match_operand 2 "fr_register_operand" "f"))))]
-  "GET_MODE (operands[2]) == DFmode || GET_MODE (operands[2]) == XFmode"
-  "fpack %0 = %2, %F1"
-  [(set_attr "itanium_class" "fmisc")])
-
-(define_insn "*fpack_xfsf"
-  [(set (match_operand:V2SF 0 "fr_register_operand" "=f")
-	(vec_concat:V2SF
-	  (float_truncate:SF
-	    (match_operand 1 "fr_register_operand" "f"))
-	  (match_operand:SF 2 "fr_reg_or_fp01_operand" "fG")))]
-  "GET_MODE (operands[1]) == DFmode || GET_MODE (operands[1]) == XFmode"
-  "fpack %0 = %F2, %1"
-  [(set_attr "itanium_class" "fmisc")])
-
-(define_insn "*fpack_xfxf"
-  [(set (match_operand:V2SF 0 "fr_register_operand" "=f")
-	(vec_concat:V2SF
-	  (float_truncate:SF
-	    (match_operand 1 "fr_register_operand" "f"))
-	  (float_truncate:SF
-	    (match_operand 2 "fr_register_operand" "f"))))]
-  "(GET_MODE (operands[1]) == DFmode || GET_MODE (operands[1]) == XFmode)
-   && (GET_MODE (operands[2]) == DFmode || GET_MODE (operands[2]) == XFmode)"
-  "fpack %0 = %2, %1"
-  [(set_attr "itanium_class" "fmisc")])
-
 ;; Missing operations
 ;; fprcpa
 ;; fpsqrta
diff --git a/gcc/testsuite/gcc.c-torture/execute/simd-6.c b/gcc/testsuite/gcc.c-torture/execute/simd-6.c
new file mode 100644
index 0000000..ca4aabc
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/execute/simd-6.c
@@ -0,0 +1,22 @@
+extern void abort (void);
+extern int memcmp (const void *, const void *, __SIZE_TYPE__);
+
+typedef unsigned char v8qi __attribute__((vector_size(8)));
+
+v8qi foo(v8qi x, v8qi y)
+{
+  return x * y;
+}
+
+int main()
+{
+  v8qi a = { 1, 2, 3, 4, 5, 6, 7, 8 };
+  v8qi b = { 3, 3, 3, 3, 3, 3, 3, 3 };
+  v8qi c = { 3, 6, 9, 12, 15, 18, 21, 24 };
+  v8qi r;
+
+  r = foo (a, b);
+  if (memcmp (&r, &c, 8) != 0)
+    abort ();
+  return 0;
+}