* @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 <stdbool.h>
#include "lc_opts.h"
#include "lc_opts_enum.h"
/**
* CPU architectures and features.
*/
-enum cpu_arch_features {
+typedef enum cpu_arch_features {
arch_generic32 = 0x00000001, /**< no specific architecture */
arch_i386 = 0x00000002, /**< i386 architecture */
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_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 */
-};
-
-#define FLAGS(x, f) (((x) & (f)) != 0)
-/**
- * CPU's.
- */
-typedef enum cpu_support {
- cpu_generic = arch_generic32,
+ 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_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_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 = 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_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_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 = 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 = arch_k10 | arch_3DNowE_insn | arch_feature_cmov | arch_feature_p6_insn | arch_feature_popcnt | arch_64bit_insn | arch_sse4a_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_c3_2 = arch_ppro | arch_feature_cmov | arch_feature_p6_insn | arch_sse1_insn, /* really no 3DNow! */
cpu_autodetect = 0,
-} cpu_support;
+} cpu_arch_features;
+ENUM_BITSET(cpu_arch_features)
-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 fpu_arch = 0;
-static int opt_cc = 1;
-static int opt_unsafe_floatconv = 0;
+static int opt_size = 0;
+static int emit_machcode = 0;
+static cpu_arch_features arch = cpu_generic;
+static cpu_arch_features opt_arch = cpu_generic;
+static int fpu_arch = 0;
+static int opt_cc = 1;
+static int opt_unsafe_floatconv = 0;
/* instruction set architectures. */
static const lc_opt_enum_int_items_t arch_items[] = {
{ "merom", cpu_core2 },
{ "core2", cpu_core2 },
{ "penryn", cpu_penryn },
+ { "atom", cpu_atom },
{ "k6", cpu_k6 },
{ "k6-2", cpu_k6_PLUS },
CPUID_FEAT_EDX_PBE = 1 << 31
};
-static cpu_support auto_detect_Intel(x86_cpu_info_t const *info)
+static cpu_arch_features auto_detect_Intel(x86_cpu_info_t const *info)
{
- cpu_support auto_arch = cpu_generic;
+ cpu_arch_features auto_arch = cpu_generic;
- unsigned family = (info->cpu_ext_family << 4) | info->cpu_family;
- unsigned model = (info->cpu_ext_model << 4) | info->cpu_model;
+ unsigned family = info->cpu_ext_family + info->cpu_family;
+ unsigned model = (info->cpu_ext_model << 4) | info->cpu_model;
switch (family) {
case 4:
- auto_arch = arch_i486;
+ auto_arch = cpu_i486;
break;
case 5:
- auto_arch = arch_pentium;
+ auto_arch = cpu_pentium;
break;
case 6:
switch (model) {
case 0x0A: /* Pentium III Model 0A */
case 0x0B: /* Pentium III Model 0B */
case 0x0D: /* Pentium M Model 0D */
- auto_arch = arch_ppro | arch_feature_p6_insn;
- break;
case 0x0E: /* Core Model 0E */
- auto_arch = arch_ppro | arch_feature_p6_insn;
+ 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 = arch_core2 | arch_feature_p6_insn;
+ auto_arch = cpu_core2_generic;
break;
default:
/* unknown */
case 0x03: /* Pentium 4 Model 03 */
case 0x04: /* Pentium 4 Model 04 */
case 0x06: /* Pentium 4 Model 06 */
- auto_arch = arch_netburst | arch_feature_p6_insn;
+ auto_arch = cpu_netburst_generic;
break;
case 0x1A: /* Core i7 */
- auto_arch = arch_core2 | arch_feature_p6_insn;
+ auto_arch = cpu_core2_generic;
break;
case 0x1C: /* Atom */
- auto_arch = arch_atom;
+ auto_arch = cpu_atom_generic;
break;
case 0x1D: /* Xeon MP */
- auto_arch = arch_core2 | arch_feature_p6_insn;
+ auto_arch = cpu_core2_generic;
break;
default:
/* unknown */
return auto_arch;
}
-static cpu_support auto_detect_AMD(x86_cpu_info_t const *info) {
- cpu_support auto_arch = cpu_generic;
+static cpu_arch_features auto_detect_AMD(x86_cpu_info_t const *info)
+{
+ cpu_arch_features 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;
+ family = info->cpu_ext_family + 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 = arch_i486;
+ auto_arch = cpu_i486;
break;
case 0x05:
switch (model) {
case 0x01: /* K5 Model 1 */
case 0x02: /* K5 Model 2 */
case 0x03: /* K5 Model 3 */
- auto_arch = arch_pentium;
+ 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 = arch_k6;
+ auto_arch = cpu_k6_generic;
break;
case 0x0A: /* Geode LX */
- auto_arch = arch_geode;
+ auto_arch = cpu_geode_generic;
break;
default:
/* unknown K6 */
- auto_arch = arch_k6;
+ auto_arch = cpu_k6_generic;
break;
}
break;
case 0x07: /* Mobile Duron Model 7 */
case 0x08: /* Athlon (TH/AP core) including Geode NX */
case 0x0A: /* Athlon (BT core) */
- auto_arch = arch_athlon | arch_feature_p6_insn;
- break;
- default:
- /* unknown K7 */
- auto_arch = arch_athlon | arch_feature_p6_insn;
+ default: /* unknown K7 */
+ auto_arch = cpu_athlon_generic;
break;
}
break;
case 0x0F:
- auto_arch = arch_k8 | arch_feature_p6_insn;
+ auto_arch = cpu_k8_generic;
break;
- case 0x1F:
- case 0x2F: /* AMD Family 11h */
- auto_arch = arch_k10 | arch_feature_p6_insn;
+ case 0x10:
+ case 0x11: /* AMD Family 11h */
+ case 0x12: /* AMD Family 12h */
+ case 0x14: /* AMD Family 14h */
+ case 0x15: /* AMD Family 15h */
+ auto_arch = cpu_k10_generic;
break;
default:
/* unknown */
#endif
}
-static int x86_toogle_cpuid(void)
+static bool x86_toogle_cpuid(void)
{
- unsigned eflags_before = 0, eflags_after = 0;
+ unsigned eflags_before = 0;
+ unsigned eflags_after = 0;
#if defined(__GNUC__)
#ifdef __i386__
: "=r" (eflags_before), "=r" (eflags_after) :: "cc"
);
#else
- /* cpuid always available on 64bit */
- return true;
+ eflags_after = 0x00200000;
#endif
#elif defined(_MSC_VER)
#if defined(_M_IX86)
mov eflags_after, eax
}
#else
- return true;
+ eflags_after = 0x00200000;
#endif
#endif
return (eflags_before ^ eflags_after) & 0x00200000;
static void autodetect_arch(void)
{
- cpu_support auto_arch = cpu_generic;
+ cpu_arch_features auto_arch = cpu_generic;
/* We use the cpuid instruction to detect the CPU features */
if (x86_toogle_cpuid()) {
} else if (0 == strcmp(vendorid, "AuthenticAMD")) {
auto_arch = auto_detect_AMD(&cpu_info);
} else if (0 == strcmp(vendorid, "Geode by NSC")) {
- auto_arch = arch_geode;
+ auto_arch = cpu_geode_generic;
}
if (cpu_info.edx_features & CPUID_FEAT_EDX_CMOV)
}
#endif /* NATIVE_X86 */
+static bool flags(cpu_arch_features features, cpu_arch_features flags)
+{
+ return (features & flags) != 0;
+}
+
void ia32_setup_cg_config(void)
{
ia32_code_gen_config_t *const c = &ia32_cg_config;
c->optimize_size = opt_size != 0;
/* 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;
+ 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 */
- c->use_incdec = !FLAGS(opt_arch, arch_netburst | arch_nocona | arch_core2 | arch_geode) || opt_size;
- 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);
+ c->use_incdec = !flags(opt_arch, arch_netburst | arch_nocona | arch_core2 | arch_geode) || opt_size;
+ c->use_softfloat = (fpu_arch & IA32_FPU_ARCH_SOFTFLOAT) != 0;
+ c->use_sse2 = (fpu_arch & IA32_FPU_ARCH_SSE2) != 0 && flags(arch, arch_feature_sse2);
+ c->use_ffreep = flags(opt_arch, arch_athlon_plus);
/* valgrind can't cope with femms yet and the usefulness of the optimization
* is questionable anyway */
#if 0
- c->use_femms = FLAGS(opt_arch, arch_athlon_plus) &&
- FLAGS(arch, arch_feature_mmx | arch_all_amd);
+ c->use_femms = flags(opt_arch, arch_athlon_plus) &&
+ flags(arch, arch_feature_mmx | arch_all_amd);
#else
c->use_femms = 0;
#endif
- c->use_fucomi = 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;
- c->use_sub_esp_4 = FLAGS(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro) && !opt_size;
- c->use_sub_esp_8 = FLAGS(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro | arch_i386 | arch_i486) && !opt_size;
- c->use_imul_mem_imm32 = !FLAGS(opt_arch, arch_k8 | arch_k10) || opt_size;
- 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) && !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_i486 = (arch & arch_mask) >= arch_i486;
+ c->use_fucomi = 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;
+ c->use_sub_esp_4 = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro) && !opt_size;
+ c->use_sub_esp_8 = flags(opt_arch, arch_generic32 | arch_athlon_plus | arch_netburst | arch_nocona | arch_core2 | arch_ppro | arch_i386 | arch_i486) && !opt_size;
+ c->use_imul_mem_imm32 = !flags(opt_arch, arch_k8 | arch_k10) || opt_size;
+ 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) && !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->label_alignment_max_skip = arch_costs->label_alignment_max_skip;
c->label_alignment_factor =
- FLAGS(opt_arch, arch_i386 | arch_i486) || opt_size ? 0 :
+ flags(opt_arch, arch_i386 | arch_i486) || opt_size ? 0 :
opt_arch & arch_all_amd ? 3 :
2;
}
projects / libfirm / blobdiff