ia32: fix and optimize bswap sequence on i386
[libfirm] / ir / be / ia32 / ia32_architecture.c
index ecffa92..6fdfb60 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2010 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -21,7 +21,6 @@
  * @file
  * @brief       ia32 architecture variants
  * @author      Michael Beck, Matthias Braun
- * @version     $Id: bearch_ia32_t.h 16363 2007-10-25 23:27:07Z beck $
  */
 #include "config.h"
 
 #include "bearch_ia32_t.h"
 #include "ia32_architecture.h"
 
+#undef NATIVE_X86
+
+#ifdef _MSC_VER
+#if defined(_M_IX86) || defined(_M_X64)
+#include <intrin.h>
+#define NATIVE_X86
+#endif
+#else
+#if defined(__i386__) || defined(__x86_64__)
+#define NATIVE_X86
+#endif
+#endif
+
 ia32_code_gen_config_t  ia32_cg_config;
 
 /**
@@ -62,22 +74,30 @@ enum cpu_arch_features {
        arch_all_amd          = arch_k6 | arch_geode | arch_athlon_plus,
 
        arch_feature_mmx      = 0x00004000, /**< MMX instructions */
-       arch_feature_p6_insn  = 0x00008000, /**< PentiumPro instructions */
-       arch_feature_sse1     = 0x00010000, /**< SSE1 instructions */
-       arch_feature_sse2     = 0x00020000, /**< SSE2 instructions */
-       arch_feature_sse3     = 0x00040000, /**< SSE3 instructions */
-       arch_feature_ssse3    = 0x00080000, /**< SSSE3 instructions */
-       arch_feature_3DNow    = 0x00100000, /**< 3DNow! instructions */
-       arch_feature_3DNowE   = 0x00200000, /**< Enhanced 3DNow! instructions */
-       arch_feature_64bit    = 0x00400000, /**< x86_64 support */
+       arch_feature_cmov     = 0x00008000, /**< cmov instructions */
+       arch_feature_p6_insn  = 0x00010000, /**< PentiumPro instructions */
+       arch_feature_sse1     = 0x00020000, /**< SSE1 instructions */
+       arch_feature_sse2     = 0x00040000, /**< SSE2 instructions */
+       arch_feature_sse3     = 0x00080000, /**< SSE3 instructions */
+       arch_feature_ssse3    = 0x00100000, /**< SSSE3 instructions */
+       arch_feature_3DNow    = 0x00200000, /**< 3DNow! instructions */
+       arch_feature_3DNowE   = 0x00400000, /**< Enhanced 3DNow! instructions */
+       arch_feature_64bit    = 0x00800000, /**< x86_64 support */
+       arch_feature_sse4_1   = 0x01000000, /**< SSE4.1 instructions */
+       arch_feature_sse4_2   = 0x02000000, /**< SSE4.2 instructions */
+       arch_feature_sse4a    = 0x04000000, /**< SSE4a instructions */
+       arch_feature_popcnt   = 0x08000000, /**< popcnt instruction */
 
        arch_mmx_insn     = arch_feature_mmx,                         /**< MMX instructions */
-       arch_sse1_insn    = arch_feature_sse1  | arch_mmx_insn,       /**< SSE1 instructions, include MMX */
-       arch_sse2_insn    = arch_feature_sse2  | arch_sse1_insn,      /**< SSE2 instructions, include SSE1 */
-       arch_sse3_insn    = arch_feature_sse3  | arch_sse2_insn,      /**< SSE3 instructions, include SSE2 */
-       arch_ssse3_insn   = arch_feature_ssse3 | arch_sse3_insn,      /**< SSSE3 instructions, include SSE3 */
-
-       arch_3DNow_insn   = arch_feature_3DNow | arch_feature_mmx,    /**< 3DNow! instructions, including MMX */
+       arch_sse1_insn    = arch_feature_sse1   | arch_mmx_insn,      /**< SSE1 instructions, include MMX */
+       arch_sse2_insn    = arch_feature_sse2   | arch_sse1_insn,     /**< SSE2 instructions, include SSE1 */
+       arch_sse3_insn    = arch_feature_sse3   | arch_sse2_insn,     /**< SSE3 instructions, include SSE2 */
+       arch_ssse3_insn   = arch_feature_ssse3  | arch_sse3_insn,     /**< SSSE3 instructions, include SSE3 */
+       arch_sse4_1_insn  = arch_feature_sse4_1 | arch_ssse3_insn,    /**< SSE4.1 instructions, include SSSE3 */
+       arch_sse4_2_insn  = arch_feature_sse4_2 | arch_sse4_1_insn,   /**< SSE4.2 instructions, include SSE4.1 */
+       arch_sse4a_insn   = arch_feature_sse4a  | arch_ssse3_insn,    /**< SSE4a instructions, include SSSE3 */
+
+       arch_3DNow_insn   = arch_feature_3DNow  | arch_feature_mmx,   /**< 3DNow! instructions, including MMX */
        arch_3DNowE_insn  = arch_feature_3DNowE | arch_3DNow_insn,    /**< Enhanced 3DNow! instructions */
        arch_64bit_insn   = arch_feature_64bit  | arch_sse2_insn,     /**< x86_64 support, includes SSE2 */
 };
@@ -87,44 +107,59 @@ enum cpu_arch_features {
 /**
  * CPU's.
  */
-enum cpu_support {
-       cpu_generic     = arch_generic32,
+typedef enum cpu_support {
+       cpu_generic             = arch_generic32,
 
        /* intel CPUs */
-       cpu_i386        = arch_i386,
-       cpu_i486        = arch_i486,
-       cpu_pentium     = arch_pentium,
-       cpu_pentium_mmx = arch_pentium | arch_mmx_insn,
-       cpu_pentium_pro = arch_ppro | arch_feature_p6_insn,
-       cpu_pentium_2   = arch_ppro | arch_feature_p6_insn | arch_mmx_insn,
-       cpu_pentium_3   = arch_ppro | arch_feature_p6_insn | arch_sse1_insn,
-       cpu_pentium_m   = arch_ppro | arch_feature_p6_insn | arch_sse2_insn,
-       cpu_pentium_4   = arch_netburst | arch_feature_p6_insn | arch_sse2_insn,
-       cpu_prescott    = arch_nocona | arch_feature_p6_insn | arch_sse3_insn,
-       cpu_nocona      = arch_nocona | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
-       cpu_core2       = arch_core2 | arch_feature_p6_insn | arch_64bit_insn | arch_ssse3_insn,
+       cpu_i386                = arch_i386,
+       cpu_i486                = arch_i486,
+       cpu_pentium             = arch_pentium,
+       cpu_pentium_mmx         = arch_pentium | arch_mmx_insn,
+       cpu_pentium_pro_generic = arch_ppro | arch_feature_p6_insn,
+       cpu_pentium_pro         = arch_ppro | arch_feature_cmov | arch_feature_p6_insn,
+       cpu_pentium_2           = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_mmx_insn,
+       cpu_pentium_3           = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse1_insn,
+       cpu_pentium_m           = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse2_insn,
+       cpu_netburst_generic    = arch_netburst | arch_feature_p6_insn,
+       cpu_pentium_4           = arch_netburst | arch_feature_cmov | arch_feature_p6_insn | arch_sse2_insn,
+       cpu_prescott            = arch_nocona | arch_feature_cmov | arch_feature_p6_insn | arch_sse3_insn,
+       cpu_nocona              = arch_nocona | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
+       cpu_core2_generic       = arch_core2 | arch_feature_p6_insn,
+       cpu_core2               = arch_core2 | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_ssse3_insn,
+       cpu_penryn              = arch_core2 | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_sse4_1_insn,
+       cpu_atom_generic        = arch_atom | arch_feature_p6_insn,
+       cpu_atom                = arch_atom | arch_feature_cmov | arch_feature_p6_insn | arch_ssse3_insn,
 
        /* AMD CPUs */
-       cpu_k6          = arch_k6 | arch_mmx_insn,
-       cpu_k6_PLUS     = arch_k6 | arch_3DNow_insn,
-       cpu_geode       = arch_geode  | arch_sse1_insn | arch_3DNowE_insn,
-       cpu_athlon      = arch_athlon | arch_sse1_insn | arch_3DNowE_insn | arch_feature_p6_insn,
-       cpu_athlon64    = arch_athlon | arch_sse2_insn | arch_3DNowE_insn | arch_feature_p6_insn | arch_64bit_insn,
-       cpu_k8          = arch_k8  | arch_3DNowE_insn | arch_feature_p6_insn | arch_64bit_insn,
-       cpu_k8_sse3     = arch_k8  | arch_3DNowE_insn | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
-       cpu_k10         = arch_k10 | arch_3DNowE_insn | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
+       cpu_k6_generic     = arch_k6,
+       cpu_k6             = arch_k6 | arch_mmx_insn,
+       cpu_k6_PLUS        = arch_k6 | arch_3DNow_insn,
+       cpu_geode_generic  = arch_geode,
+       cpu_geode          = arch_geode  | arch_sse1_insn | arch_3DNowE_insn,
+       cpu_athlon_generic = arch_athlon | arch_feature_p6_insn,
+       cpu_athlon_old     = arch_athlon | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn,
+       cpu_athlon         = arch_athlon | arch_sse1_insn | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn,
+       cpu_athlon64       = arch_athlon | arch_sse2_insn | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn,
+       cpu_k8_generic     = arch_k8  | arch_feature_p6_insn,
+       cpu_k8             = arch_k8  | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn,
+       cpu_k8_sse3        = arch_k8  | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_64bit_insn | arch_sse3_insn,
+       cpu_k10_generic    = arch_k10 | arch_feature_p6_insn,
+       cpu_k10            = arch_k10 | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_feature_popcnt | arch_64bit_insn | arch_sse4a_insn,
 
        /* other CPUs */
        cpu_winchip_c6  = arch_i486 | arch_feature_mmx,
        cpu_winchip2    = arch_i486 | arch_feature_mmx | arch_feature_3DNow,
        cpu_c3          = arch_i486 | arch_feature_mmx | arch_feature_3DNow,
-       cpu_c3_2        = arch_ppro | arch_feature_p6_insn | arch_sse1_insn, /* really no 3DNow! */
-};
+       cpu_c3_2        = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse1_insn, /* really no 3DNow! */
+
+       cpu_autodetect  = 0,
+} cpu_support;
 
 static int         opt_size             = 0;
+static int         emit_machcode        = 0;
 static cpu_support arch                 = cpu_generic;
 static cpu_support opt_arch             = cpu_generic;
-static int         use_sse2             = 0;
+static int         fpu_arch             = 0;
 static int         opt_cc               = 1;
 static int         opt_unsafe_floatconv = 0;
 
@@ -151,12 +186,14 @@ static const lc_opt_enum_int_items_t arch_items[] = {
        { "nocona",       cpu_nocona },
        { "merom",        cpu_core2 },
        { "core2",        cpu_core2 },
+       { "penryn",       cpu_penryn },
+       { "atom",         cpu_atom },
 
        { "k6",           cpu_k6 },
        { "k6-2",         cpu_k6_PLUS },
        { "k6-3",         cpu_k6_PLUS },
        { "geode",        cpu_geode },
-       { "athlon",       cpu_athlon },
+       { "athlon",       cpu_athlon_old },
        { "athlon-tbird", cpu_athlon },
        { "athlon-4",     cpu_athlon },
        { "athlon-xp",    cpu_athlon },
@@ -178,6 +215,11 @@ static const lc_opt_enum_int_items_t arch_items[] = {
 
        { "generic",      cpu_generic },
        { "generic32",    cpu_generic },
+
+#ifdef NATIVE_X86
+       { "native",       cpu_autodetect },
+#endif
+
        { NULL,           0 }
 };
 
@@ -190,27 +232,24 @@ static lc_opt_enum_int_var_t opt_arch_var = {
 };
 
 static const lc_opt_enum_int_items_t fp_unit_items[] = {
-       { "x87" ,    0 },
-       { "sse2",    1 },
-       { NULL,      0 }
+       { "x87" ,      IA32_FPU_ARCH_X87 },
+       { "sse2",      IA32_FPU_ARCH_SSE2 },
+       { "softfloat", IA32_FPU_ARCH_SOFTFLOAT },
+       { NULL,        IA32_FPU_ARCH_NONE }
 };
 
 static lc_opt_enum_int_var_t fp_unit_var = {
-       &use_sse2, fp_unit_items
+       &fpu_arch, fp_unit_items
 };
 
 static const lc_opt_table_entry_t ia32_architecture_options[] = {
-       LC_OPT_ENT_BOOL("size",            "optimize for size", &opt_size),
-       LC_OPT_ENT_ENUM_INT("arch",        "select the instruction architecture",
-                           &arch_var),
-       LC_OPT_ENT_ENUM_INT("opt",         "optimize for instruction architecture",
-                           &opt_arch_var),
-       LC_OPT_ENT_ENUM_INT("fpunit",      "select the floating point unit",
-                           &fp_unit_var),
-       LC_OPT_ENT_NEGBIT("nooptcc",       "do not optimize calling convention",
-                         &opt_cc, 1),
-       LC_OPT_ENT_BIT("unsafe_floatconv", "do unsafe floating point controlword "
-                      "optimisations", &opt_unsafe_floatconv, 1),
+       LC_OPT_ENT_BOOL    ("size",             "optimize for size",                                  &opt_size),
+       LC_OPT_ENT_ENUM_INT("arch",             "select the instruction architecture",                &arch_var),
+       LC_OPT_ENT_ENUM_INT("opt",              "optimize for instruction architecture",              &opt_arch_var),
+       LC_OPT_ENT_ENUM_INT("fpunit",           "select the floating point unit",                     &fp_unit_var),
+       LC_OPT_ENT_NEGBOOL ("nooptcc",          "do not optimize calling convention",                 &opt_cc),
+       LC_OPT_ENT_BOOL    ("unsafe_floatconv", "do unsafe floating point controlword optimisations", &opt_unsafe_floatconv),
+       LC_OPT_ENT_BOOL    ("machcode",         "output machine code instead of assembler",           &emit_machcode),
        LC_OPT_LAST
 };
 
@@ -426,12 +465,13 @@ static void set_arch_costs(void)
 }
 
 /* Evaluate the costs of an instruction. */
-int ia32_evaluate_insn(insn_kind kind, tarval *tv) {
+int ia32_evaluate_insn(insn_kind kind, const ir_mode *mode, ir_tarval *tv)
+{
        int cost;
 
        switch (kind) {
        case MUL:
-               cost =  arch_costs->cost_mul_start;
+               cost = arch_costs->cost_mul_start;
                if (arch_costs->cost_mul_bit > 0) {
                        char *bitstr = get_tarval_bitpattern(tv);
                        int i;
@@ -443,14 +483,27 @@ int ia32_evaluate_insn(insn_kind kind, tarval *tv) {
                        }
                        free(bitstr);
                }
-               return cost;
+               if (get_mode_size_bits(mode) <= 32)
+                       return cost;
+               /* 64bit mul supported, approx 4times of a 32bit mul*/
+               return 4 * cost;
        case LEA:
-               return arch_costs->lea_cost;
+               /* lea is only supported for 32 bit */
+               if (get_mode_size_bits(mode) <= 32)
+                       return arch_costs->lea_cost;
+               /* in 64bit mode, the Lea cost are at wort 2 shifts and one add */
+               return 2 * arch_costs->add_cost + 2 * (2 * arch_costs->const_shf_cost);
        case ADD:
        case SUB:
-               return arch_costs->add_cost;
+               if (get_mode_size_bits(mode) <= 32)
+                       return arch_costs->add_cost;
+               /* 64bit add/sub supported, double the cost */
+               return 2 * arch_costs->add_cost;
        case SHIFT:
-               return arch_costs->const_shf_cost;
+               if (get_mode_size_bits(mode) <= 32)
+                       return arch_costs->const_shf_cost;
+               /* 64bit shift supported, double the cost */
+               return 2 * arch_costs->const_shf_cost;
        case ZERO:
                return arch_costs->add_cost;
        default:
@@ -458,6 +511,363 @@ int ia32_evaluate_insn(insn_kind kind, tarval *tv) {
        }
 }
 
+/* auto detection code only works if we're on an x86 cpu obviously */
+#ifdef NATIVE_X86
+typedef struct x86_cpu_info_t {
+       unsigned char cpu_stepping;
+       unsigned char cpu_model;
+       unsigned char cpu_family;
+       unsigned char cpu_type;
+       unsigned char cpu_ext_model;
+       unsigned char cpu_ext_family;
+       unsigned      edx_features;
+       unsigned      ecx_features;
+       unsigned      add_features;
+} x86_cpu_info_t;
+
+enum {
+       CPUID_FEAT_ECX_SSE3      = 1 << 0,
+       CPUID_FEAT_ECX_PCLMUL    = 1 << 1,
+       CPUID_FEAT_ECX_DTES64    = 1 << 2,
+       CPUID_FEAT_ECX_MONITOR   = 1 << 3,
+       CPUID_FEAT_ECX_DS_CPL    = 1 << 4,
+       CPUID_FEAT_ECX_VMX       = 1 << 5,
+       CPUID_FEAT_ECX_SMX       = 1 << 6,
+       CPUID_FEAT_ECX_EST       = 1 << 7,
+       CPUID_FEAT_ECX_TM2       = 1 << 8,
+       CPUID_FEAT_ECX_SSSE3     = 1 << 9,
+       CPUID_FEAT_ECX_CID       = 1 << 10,
+       CPUID_FEAT_ECX_FMA       = 1 << 12,
+       CPUID_FEAT_ECX_CX16      = 1 << 13,
+       CPUID_FEAT_ECX_ETPRD     = 1 << 14,
+       CPUID_FEAT_ECX_PDCM      = 1 << 15,
+       CPUID_FEAT_ECX_DCA       = 1 << 18,
+       CPUID_FEAT_ECX_SSE4_1    = 1 << 19,
+       CPUID_FEAT_ECX_SSE4_2    = 1 << 20,
+       CPUID_FEAT_ECX_x2APIC    = 1 << 21,
+       CPUID_FEAT_ECX_MOVBE     = 1 << 22,
+       CPUID_FEAT_ECX_POPCNT    = 1 << 23,
+       CPUID_FEAT_ECX_AES       = 1 << 25,
+       CPUID_FEAT_ECX_XSAVE     = 1 << 26,
+       CPUID_FEAT_ECX_OSXSAVE   = 1 << 27,
+       CPUID_FEAT_ECX_AVX       = 1 << 28,
+
+       CPUID_FEAT_EDX_FPU       = 1 << 0,
+       CPUID_FEAT_EDX_VME       = 1 << 1,
+       CPUID_FEAT_EDX_DE        = 1 << 2,
+       CPUID_FEAT_EDX_PSE       = 1 << 3,
+       CPUID_FEAT_EDX_TSC       = 1 << 4,
+       CPUID_FEAT_EDX_MSR       = 1 << 5,
+       CPUID_FEAT_EDX_PAE       = 1 << 6,
+       CPUID_FEAT_EDX_MCE       = 1 << 7,
+       CPUID_FEAT_EDX_CX8       = 1 << 8,
+       CPUID_FEAT_EDX_APIC      = 1 << 9,
+       CPUID_FEAT_EDX_SEP       = 1 << 11,
+       CPUID_FEAT_EDX_MTRR      = 1 << 12,
+       CPUID_FEAT_EDX_PGE       = 1 << 13,
+       CPUID_FEAT_EDX_MCA       = 1 << 14,
+       CPUID_FEAT_EDX_CMOV      = 1 << 15,
+       CPUID_FEAT_EDX_PAT       = 1 << 16,
+       CPUID_FEAT_EDX_PSE36     = 1 << 17,
+       CPUID_FEAT_EDX_PSN       = 1 << 18,
+       CPUID_FEAT_EDX_CLF       = 1 << 19,
+       CPUID_FEAT_EDX_DTES      = 1 << 21,
+       CPUID_FEAT_EDX_ACPI      = 1 << 22,
+       CPUID_FEAT_EDX_MMX       = 1 << 23,
+       CPUID_FEAT_EDX_FXSR      = 1 << 24,
+       CPUID_FEAT_EDX_SSE       = 1 << 25,
+       CPUID_FEAT_EDX_SSE2      = 1 << 26,
+       CPUID_FEAT_EDX_SS        = 1 << 27,
+       CPUID_FEAT_EDX_HTT       = 1 << 28,
+       CPUID_FEAT_EDX_TM1       = 1 << 29,
+       CPUID_FEAT_EDX_IA64      = 1 << 30,
+       CPUID_FEAT_EDX_PBE       = 1 << 31
+};
+
+static cpu_support auto_detect_Intel(x86_cpu_info_t const *info)
+{
+       cpu_support auto_arch = cpu_generic;
+
+       unsigned family = (info->cpu_ext_family << 4) | info->cpu_family;
+       unsigned model  = (info->cpu_ext_model  << 4) | info->cpu_model;
+
+       switch (family) {
+       case 4:
+               auto_arch = cpu_i486;
+               break;
+       case 5:
+               auto_arch = cpu_pentium;
+               break;
+       case 6:
+               switch (model) {
+               case 0x01: /* PentiumPro */
+               case 0x03: /* Pentium II Model 3 */
+               case 0x05: /* Pentium II Model 5 */
+               case 0x06: /* Celeron Model 6 */
+               case 0x07: /* Pentium III Model 7 */
+               case 0x08: /* Pentium III Model 8 */
+               case 0x09: /* Pentium M Model 9 */
+               case 0x0A: /* Pentium III Model 0A */
+               case 0x0B: /* Pentium III Model 0B */
+               case 0x0D: /* Pentium M Model 0D */
+               case 0x0E: /* Core Model 0E */
+                       auto_arch = cpu_pentium_pro_generic;
+                       break;
+               case 0x0F: /* Core2 Model 0F */
+               case 0x15: /* Intel EP80579 */
+               case 0x16: /* Celeron Model 16 */
+               case 0x17: /* Core2 Model 17 */
+                       auto_arch = cpu_core2_generic;
+                       break;
+               default:
+                       /* unknown */
+                       break;
+               }
+               break;
+       case 15:
+               switch (model) {
+               case 0x00: /* Pentium 4 Model 00 */
+               case 0x01: /* Pentium 4 Model 01 */
+               case 0x02: /* Pentium 4 Model 02 */
+               case 0x03: /* Pentium 4 Model 03 */
+               case 0x04: /* Pentium 4 Model 04 */
+               case 0x06: /* Pentium 4 Model 06 */
+                       auto_arch = cpu_netburst_generic;
+                       break;
+               case 0x1A: /* Core i7 */
+                       auto_arch = cpu_core2_generic;
+                       break;
+               case 0x1C: /* Atom */
+                       auto_arch = cpu_atom_generic;
+                       break;
+               case 0x1D: /* Xeon MP */
+                       auto_arch = cpu_core2_generic;
+                       break;
+               default:
+                       /* unknown */
+                       break;
+               }
+               break;
+       default:
+               /* unknown */
+               break;
+       }
+
+       return auto_arch;
+}
+
+static cpu_support auto_detect_AMD(x86_cpu_info_t const *info) {
+       cpu_support auto_arch = cpu_generic;
+
+       unsigned family, model;
+
+       if (info->cpu_family == 0x0F) {
+               family = (info->cpu_ext_family << 4) | info->cpu_family;
+               model  = (info->cpu_ext_model  << 4) | info->cpu_model;
+       } else {
+               family = info->cpu_family;
+               model  = info->cpu_model;
+       }
+
+       switch (family) {
+       case 0x04:
+               auto_arch = cpu_i486;
+               break;
+       case 0x05:
+               switch (model) {
+               case 0x00: /* K5 Model 0 */
+               case 0x01: /* K5 Model 1 */
+               case 0x02: /* K5 Model 2 */
+               case 0x03: /* K5 Model 3 */
+                       auto_arch = cpu_pentium;
+                       break;
+               case 0x06: /* K6 Model 6 */
+               case 0x07: /* K6 Model 7 */
+               case 0x08: /* K6-2 Model 8 */
+               case 0x09: /* K6-III Model 9 */
+               case 0x0D: /* K6-2+ or K6-III+ */
+                       auto_arch = cpu_k6_generic;
+                       break;
+               case 0x0A: /* Geode LX */
+                       auto_arch = cpu_geode_generic;
+                       break;
+               default:
+                       /* unknown K6 */
+                       auto_arch = cpu_k6_generic;
+                       break;
+               }
+               break;
+       case 0x06:
+               switch (model) {
+               case 0x01: /* Athlon Model 1 */
+               case 0x02: /* Athlon Model 2 */
+               case 0x03: /* Duron Model 3 */
+               case 0x04: /* Athlon Model 4 */
+               case 0x06: /* Athlon MP/Mobile Athlon Model 6 */
+               case 0x07: /* Mobile Duron Model 7 */
+               case 0x08: /* Athlon (TH/AP core) including Geode NX */
+               case 0x0A: /* Athlon (BT core) */
+               default:   /* unknown K7 */
+                       auto_arch = cpu_athlon_generic;
+                       break;
+               }
+               break;
+       case 0x0F:
+               auto_arch = cpu_k8_generic;
+               break;
+       case 0x1F:
+       case 0x2F: /* AMD Family 11h */
+               auto_arch = cpu_k10_generic;
+               break;
+       default:
+               /* unknown */
+               break;
+       }
+
+       return auto_arch;
+}
+
+typedef union {
+       struct {
+        unsigned eax;
+        unsigned ebx;
+        unsigned ecx;
+        unsigned edx;
+       } r;
+       int bulk[4];
+} cpuid_registers;
+
+static void x86_cpuid(cpuid_registers *regs, unsigned level)
+{
+#if defined(__GNUC__)
+#      if defined(__PIC__) && !defined(__amd64) // GCC cannot handle EBX in PIC
+       __asm (
+               "pushl %%ebx\n\t"
+               "cpuid\n\t"
+               "movl %%ebx, %1\n\t"
+               "popl %%ebx"
+       : "=a" (regs->r.eax), "=r" (regs->r.ebx), "=c" (regs->r.ecx), "=d" (regs->r.edx)
+       : "a" (level)
+       );
+#      else
+       __asm ("cpuid\n\t"
+       : "=a" (regs->r.eax), "=b" (regs->r.ebx), "=c" (regs->r.ecx), "=d" (regs->r.edx)
+       : "a" (level)
+       );
+#      endif
+#elif defined(_MSC_VER)
+       __cpuid(regs->bulk, level);
+#else
+#      error CPUID is missing
+#endif
+}
+
+static int x86_toogle_cpuid(void)
+{
+       unsigned eflags_before = 0, eflags_after = 0;
+
+#if defined(__GNUC__)
+#ifdef __i386__
+       /* If bit 21 of the EFLAGS register can be changed, the cpuid instruction is available */
+       __asm__(
+               "pushf\n\t"
+               "popl %0\n\t"
+               "movl %0, %1\n\t"
+               "xorl $0x00200000, %1\n\t"
+               "pushl %1\n\t"
+               "popf\n\t"
+               "pushf\n\t"
+               "popl %1"
+               : "=r" (eflags_before), "=r" (eflags_after) :: "cc"
+               );
+#else
+       /* cpuid always available on 64bit */
+       return true;
+#endif
+#elif defined(_MSC_VER)
+#if defined(_M_IX86)
+       __asm {
+               pushfd
+               pop eax
+               mov eflags_before, eax
+               xor eax, 0x00200000
+               push eax
+               popfd
+               pushfd
+               pop eax
+               mov eflags_after, eax
+       }
+#else
+       return true;
+#endif
+#endif
+       return (eflags_before ^ eflags_after) & 0x00200000;
+}
+
+static void autodetect_arch(void)
+{
+       cpu_support auto_arch = cpu_generic;
+
+       /* We use the cpuid instruction to detect the CPU features */
+       if (x86_toogle_cpuid()) {
+               cpuid_registers   regs;
+               char              vendorid[13];
+               x86_cpu_info_t    cpu_info;
+
+               /* get vendor ID */
+               x86_cpuid(&regs, 0);
+               memcpy(&vendorid[0], &regs.r.ebx, 4);
+               memcpy(&vendorid[4], &regs.r.edx, 4);
+               memcpy(&vendorid[8], &regs.r.ecx, 4);
+               vendorid[12] = '\0';
+
+               /* get processor info and feature bits */
+               x86_cpuid(&regs, 1);
+
+               cpu_info.cpu_stepping   = (regs.r.eax >>  0) & 0x0F;
+               cpu_info.cpu_model      = (regs.r.eax >>  4) & 0x0F;
+               cpu_info.cpu_family     = (regs.r.eax >>  8) & 0x0F;
+               cpu_info.cpu_type       = (regs.r.eax >> 12) & 0x03;
+               cpu_info.cpu_ext_model  = (regs.r.eax >> 16) & 0x0F;
+               cpu_info.cpu_ext_family = (regs.r.eax >> 20) & 0xFF;
+               cpu_info.edx_features   = regs.r.edx;
+               cpu_info.ecx_features   = regs.r.ecx;
+               cpu_info.add_features   = regs.r.ebx;
+
+               if        (0 == strcmp(vendorid, "GenuineIntel")) {
+                       auto_arch = auto_detect_Intel(&cpu_info);
+               } else if (0 == strcmp(vendorid, "AuthenticAMD")) {
+                       auto_arch = auto_detect_AMD(&cpu_info);
+               } else if (0 == strcmp(vendorid, "Geode by NSC")) {
+                       auto_arch = cpu_geode_generic;
+               }
+
+               if (cpu_info.edx_features & CPUID_FEAT_EDX_CMOV)
+                       auto_arch |= arch_feature_cmov;
+               if (cpu_info.edx_features & CPUID_FEAT_EDX_MMX)
+                       auto_arch |= arch_feature_mmx;
+               if (cpu_info.edx_features & CPUID_FEAT_EDX_SSE)
+                       auto_arch |= arch_feature_sse1;
+               if (cpu_info.edx_features & CPUID_FEAT_EDX_SSE2)
+                       auto_arch |= arch_feature_sse2;
+
+               if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSE3)
+                       auto_arch |= arch_feature_sse3;
+               if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSSE3)
+                       auto_arch |= arch_feature_ssse3;
+               if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSE4_1)
+                       auto_arch |= arch_feature_sse4_1;
+               if (cpu_info.ecx_features & CPUID_FEAT_ECX_SSE4_2)
+                       auto_arch |= arch_feature_sse4_2;
+               if (cpu_info.ecx_features & CPUID_FEAT_ECX_POPCNT)
+                       auto_arch |= arch_feature_popcnt;
+       }
+
+       arch     = auto_arch;
+       opt_arch = auto_arch;
+}
+#endif  /* NATIVE_X86 */
+
 void ia32_setup_cg_config(void)
 {
        ia32_code_gen_config_t *const c = &ia32_cg_config;
@@ -465,14 +875,20 @@ void ia32_setup_cg_config(void)
 
        set_arch_costs();
 
+#ifdef NATIVE_X86
+       if (arch == cpu_autodetect)
+               autodetect_arch();
+#endif
+
        c->optimize_size        = opt_size != 0;
-       /* on newer intel cpus mov, pop is often faster then leave although it has a
+       /* on newer intel cpus mov, pop is often faster than leave although it has a
         * longer opcode */
        c->use_leave            = FLAGS(opt_arch, arch_i386 | arch_all_amd | arch_core2) || opt_size;
        /* P4s don't like inc/decs because they only partially write the flags
-          register which produces false dependencies */
+        * register which produces false dependencies */
        c->use_incdec           = !FLAGS(opt_arch, arch_netburst | arch_nocona | arch_core2 | arch_geode) || opt_size;
-       c->use_sse2             = use_sse2 && FLAGS(arch, arch_feature_sse2);
+       c->use_softfloat        = FLAGS(fpu_arch, IA32_FPU_ARCH_SOFTFLOAT);
+       c->use_sse2             = FLAGS(fpu_arch, IA32_FPU_ARCH_SSE2) && FLAGS(arch, arch_feature_sse2);
        c->use_ffreep           = FLAGS(opt_arch, arch_athlon_plus);
        c->use_ftst             = !FLAGS(arch, arch_feature_p6_insn);
        /* valgrind can't cope with femms yet and the usefulness of the optimization
@@ -484,7 +900,7 @@ void ia32_setup_cg_config(void)
        c->use_femms            = 0;
 #endif
        c->use_fucomi           = FLAGS(arch, arch_feature_p6_insn);
-       c->use_cmov             = FLAGS(arch, arch_feature_p6_insn);
+       c->use_cmov             = FLAGS(arch, arch_feature_cmov);
        c->use_modeD_moves      = FLAGS(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro | arch_geode);
        c->use_add_esp_4        = FLAGS(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 |             arch_geode)                         && !opt_size;
        c->use_add_esp_8        = FLAGS(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro | arch_geode | arch_i386 | arch_i486) && !opt_size;
@@ -494,11 +910,16 @@ void ia32_setup_cg_config(void)
        c->use_pxor             = FLAGS(opt_arch, arch_netburst);
        c->use_mov_0            = FLAGS(opt_arch, arch_k6) && !opt_size;
        c->use_short_sex_eax    = !FLAGS(opt_arch, arch_k6) && !opt_size;
-       c->use_pad_return       = FLAGS(opt_arch, arch_athlon_plus | arch_core2 | arch_generic32) && !opt_size;
+       c->use_pad_return       = FLAGS(opt_arch, arch_athlon_plus) && !opt_size;
        c->use_bt               = FLAGS(opt_arch, arch_core2 | arch_athlon_plus) || opt_size;
        c->use_fisttp           = FLAGS(opt_arch & arch, arch_feature_sse3);
+       c->use_sse_prefetch     = FLAGS(arch, (arch_feature_3DNowE | arch_feature_sse1));
+       c->use_3dnow_prefetch   = FLAGS(arch, arch_feature_3DNow);
+       c->use_popcnt           = FLAGS(arch, arch_feature_popcnt);
+       c->use_bswap            = (arch & arch_mask) >= arch_i486;
        c->optimize_cc          = opt_cc;
        c->use_unsafe_floatconv = opt_unsafe_floatconv;
+       c->emit_machcode        = emit_machcode;
 
        c->function_alignment       = arch_costs->function_alignment;
        c->label_alignment          = arch_costs->label_alignment;