gcc4-pr24879.patch

   1 --- gcc/config/i386/i386.c.sse3 2005-11-15 17:33:02.000000000 -0800
   2 +++ gcc/config/i386/i386.c      2005-11-17 09:44:47.000000000 -0800
   3 @@ -14347,10 +14347,20 @@ ix86_init_mmx_sse_builtins (void)
   4    tree void_ftype_unsigned_unsigned
   5      = build_function_type_list (void_type_node, unsigned_type_node,
   6                                 unsigned_type_node, NULL_TREE);
   7 +  tree void_ftype_di_di
   8 +    = build_function_type_list (void_type_node,
   9 +                               long_long_unsigned_type_node,
  10 +                               long_long_unsigned_type_node,
  11 +                               NULL_TREE);
  12    tree void_ftype_pcvoid_unsigned_unsigned
  13      = build_function_type_list (void_type_node, const_ptr_type_node,
  14                                 unsigned_type_node, unsigned_type_node,
  15                                 NULL_TREE);
  16 +  tree void_ftype_pcvoid_di_di
  17 +    = build_function_type_list (void_type_node, const_ptr_type_node,
  18 +                               long_long_unsigned_type_node,
  19 +                               long_long_unsigned_type_node,
  20 +                               NULL_TREE);
  21    tree unsigned_ftype_void
  22      = build_function_type (unsigned_type_node, void_list_node);
  23    tree v2si_ftype_v4sf
  24 @@ -14811,12 +14821,24 @@ ix86_init_mmx_sse_builtins (void)
  25    def_builtin (MASK_SSE2, "__builtin_ia32_pmaddwd128", v4si_ftype_v8hi_v8hi, IX86_BUILTIN_PMADDWD128);
  26
  27    /* Prescott New Instructions.  */
  28 -  def_builtin (MASK_SSE3, "__builtin_ia32_monitor",
  29 -              void_ftype_pcvoid_unsigned_unsigned,
  30 -              IX86_BUILTIN_MONITOR);
  31 -  def_builtin (MASK_SSE3, "__builtin_ia32_mwait",
  32 -              void_ftype_unsigned_unsigned,
  33 -              IX86_BUILTIN_MWAIT);
  34 +  if (TARGET_64BIT)
  35 +    {
  36 +      def_builtin (MASK_SSE3, "__builtin_ia32_monitor",
  37 +                  void_ftype_pcvoid_di_di,
  38 +                  IX86_BUILTIN_MONITOR);
  39 +      def_builtin (MASK_SSE3, "__builtin_ia32_mwait",
  40 +                  void_ftype_di_di,
  41 +                  IX86_BUILTIN_MWAIT);
  42 +    }
  43 +  else
  44 +    {
  45 +      def_builtin (MASK_SSE3, "__builtin_ia32_monitor",
  46 +                  void_ftype_pcvoid_unsigned_unsigned,
  47 +                  IX86_BUILTIN_MONITOR);
  48 +      def_builtin (MASK_SSE3, "__builtin_ia32_mwait",
  49 +                  void_ftype_unsigned_unsigned,
  50 +                  IX86_BUILTIN_MWAIT);
  51 +    }
  52    def_builtin (MASK_SSE3, "__builtin_ia32_movshdup",
  53                v4sf_ftype_v4sf,
  54                IX86_BUILTIN_MOVSHDUP);
  55 @@ -15660,13 +15682,17 @@ ix86_expand_builtin (tree exp, rtx targe
  56        op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
  57        op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
  58        op2 = expand_expr (arg2, NULL_RTX, VOIDmode, 0);
  59 +      mode = TARGET_64BIT ? DImode : SImode;
  60        if (!REG_P (op0))
  61 -       op0 = copy_to_mode_reg (SImode, op0);
  62 +       op0 = copy_to_mode_reg (mode, op0);
  63        if (!REG_P (op1))
  64 -       op1 = copy_to_mode_reg (SImode, op1);
  65 +       op1 = copy_to_mode_reg (mode, op1);
  66        if (!REG_P (op2))
  67 -       op2 = copy_to_mode_reg (SImode, op2);
  68 -      emit_insn (gen_sse3_monitor (op0, op1, op2));
  69 +       op2 = copy_to_mode_reg (mode, op2);
  70 +      if (TARGET_64BIT)
  71 +       emit_insn (gen_sse3_monitor64 (op0, op1, op2));
  72 +      else
  73 +       emit_insn (gen_sse3_monitor (op0, op1, op2));
  74        return 0;
  75
  76      case IX86_BUILTIN_MWAIT:
  77 @@ -15674,11 +15700,15 @@ ix86_expand_builtin (tree exp, rtx targe
  78        arg1 = TREE_VALUE (TREE_CHAIN (arglist));
  79        op0 = expand_expr (arg0, NULL_RTX, VOIDmode, 0);
  80        op1 = expand_expr (arg1, NULL_RTX, VOIDmode, 0);
  81 +      mode = TARGET_64BIT ? DImode : SImode;
  82        if (!REG_P (op0))
  83 -       op0 = copy_to_mode_reg (SImode, op0);
  84 +       op0 = copy_to_mode_reg (mode, op0);
  85        if (!REG_P (op1))
  86 -       op1 = copy_to_mode_reg (SImode, op1);
  87 -      emit_insn (gen_sse3_mwait (op0, op1));
  88 +       op1 = copy_to_mode_reg (mode, op1);
  89 +      if (TARGET_64BIT)
  90 +       emit_insn (gen_sse3_mwait64 (op0, op1));
  91 +      else
  92 +       emit_insn (gen_sse3_mwait (op0, op1));
  93        return 0;
  94
  95      case IX86_BUILTIN_LDDQU:
  96 --- gcc/config/i386/pmmintrin.h.sse3    2005-11-04 14:13:48.000000000 -0800
  97 +++ gcc/config/i386/pmmintrin.h 2005-11-15 22:44:02.000000000 -0800
  98 @@ -110,7 +110,6 @@ _mm_lddqu_si128 (__m128i const *__P)
  99    return (__m128i) __builtin_ia32_lddqu ((char const *)__P);
 100  }
 101
 102 -#if 0
 103  static __inline void __attribute__((__always_inline__))
 104  _mm_monitor (void const * __P, unsigned int __E, unsigned int __H)
 105  {
 106 @@ -122,10 +121,6 @@ _mm_mwait (unsigned int __E, unsigned in
 107  {
 108    __builtin_ia32_mwait (__E, __H);
 109  }
 110 -#else
 111 -#define _mm_monitor(P, E, H)   __builtin_ia32_monitor ((P), (E), (H))
 112 -#define _mm_mwait(E, H)                __builtin_ia32_mwait ((E), (H))
 113 -#endif
 114
 115  #endif /* __SSE3__ */
 116
 117 --- gcc/config/i386/sse.md.sse3 2005-11-04 14:13:48.000000000 -0800
 118 +++ gcc/config/i386/sse.md      2005-11-17 09:25:31.000000000 -0800
 119 @@ -3890,15 +3890,36 @@
 120    [(unspec_volatile [(match_operand:SI 0 "register_operand" "a")
 121                      (match_operand:SI 1 "register_operand" "c")]
 122                     UNSPECV_MWAIT)]
 123 -  "TARGET_SSE3"
 124 +  "TARGET_SSE3 && !TARGET_64BIT"
 125    "mwait\t%0, %1"
 126    [(set_attr "length" "3")])
 127
 128 +(define_insn "sse3_mwait64"
 129 +  [(unspec_volatile [(match_operand:DI 0 "register_operand" "a")
 130 +                    (match_operand:DI 1 "register_operand" "c")]
 131 +                   UNSPECV_MWAIT)]
 132 +  "TARGET_SSE3 && TARGET_64BIT"
 133 +;; Older assembler doesn't support "mwait %rax,%rcx".
 134 +;;  "mwait\t%0, %1"
 135 +  "mwait"
 136 +  [(set_attr "length" "3")])
 137 +
 138  (define_insn "sse3_monitor"
 139    [(unspec_volatile [(match_operand:SI 0 "register_operand" "a")
 140                      (match_operand:SI 1 "register_operand" "c")
 141                      (match_operand:SI 2 "register_operand" "d")]
 142                     UNSPECV_MONITOR)]
 143 -  "TARGET_SSE3"
 144 +  "TARGET_SSE3 && !TARGET_64BIT"
 145    "monitor\t%0, %1, %2"
 146    [(set_attr "length" "3")])
 147 +
 148 +(define_insn "sse3_monitor64"
 149 +  [(unspec_volatile [(match_operand:DI 0 "register_operand" "a")
 150 +                    (match_operand:DI 1 "register_operand" "c")
 151 +                    (match_operand:DI 2 "register_operand" "d")]
 152 +                   UNSPECV_MONITOR)]
 153 +  "TARGET_SSE3 && TARGET_64BIT"
 154 +;; Older assembler doesn't support "monitor %rax,%rcx,%rdx".
 155 +;;  "monitor\t%0, %1, %2"
 156 +  "monitor"
 157 +  [(set_attr "length" "3")])
 158 --- gcc/testsuite/gcc.target/i386/monitor.c.sse3        2005-11-15 22:59:54.000000000 -0800
 159 +++ gcc/testsuite/gcc.target/i386/monitor.c     2005-11-15 23:25:23.000000000 -0800
 160 @@ -0,0 +1,27 @@
 161 +/* { dg-do compile { target i?86-*-* x86_64-*-* } } */
 162 +/* { dg-options "-O2 -msse3" } */
 163 +
 164 +/* Verify that they work in both 32bit and 64bit.  */
 165 +
 166 +#include <pmmintrin.h>
 167 +
 168 +void
 169 +foo (char *p, int x, int y, int z)
 170 +{
 171 +   _mm_monitor (p, y, x);
 172 +   _mm_mwait (z, y);
 173 +}
 174 +
 175 +void
 176 +bar (char *p, long x, long y, long z)
 177 +{
 178 +   _mm_monitor (p, y, x);
 179 +   _mm_mwait (z, y);
 180 +}
 181 +
 182 +void
 183 +foo1 (char *p)
 184 +{
 185 +   _mm_monitor (p, 0, 0);
 186 +   _mm_mwait (0, 0);
 187 +}
 188