Add CPU feature detection for X86.

Makefile

@@ -23,7 +23,8 @@ OBJ = retroarch.o \
 		gfx/scaler/scaler.o \
 		gfx/scaler/pixconv.o \
 		gfx/scaler/scaler_int.o \
-		gfx/scaler/filter.o
+		gfx/scaler/filter.o \
+		performance.o
 
 JOYCONFIG_OBJ = tools/retroarch-joyconfig.o \
 	conf/config_file.o \
@@ -56,7 +57,7 @@ endif
 
 ifneq ($(findstring Linux,$(OS)),)
    LIBS += -lrt
-   OBJ += input/linuxraw_input.o input/linuxraw_joypad.o performance/performance_linux.o
+   OBJ += input/linuxraw_input.o input/linuxraw_joypad.o
    JOYCONFIG_OBJ += input/linuxraw_joypad.o
 endif
 
@@ -143,7 +144,6 @@ endif
 
 ifeq ($(PERF_TEST), 1)
    DEFINES += -DPERF_TEST
-   OBJ += performance.o
 endif
 
 ifeq ($(HAVE_SDL), 1)

performance.c

@@ -16,6 +16,8 @@
 
 #include "performance.h"
 
+#ifdef PERF_TEST
+
 #if defined(__CELLOS_LV2__) || defined(GEKKO)
 #ifndef _PPU_INTRINSICS_H
 #include <ppu_intrinsics.h>
@@ -46,8 +48,6 @@ void rarch_perf_log(void)
       RARCH_PERFORMANCE_LOG(perf_counters[i]->ident, *perf_counters[i]);
 }
 
-
-
 rarch_perf_tick_t rarch_get_perf_counter(void)
 {
    rarch_perf_tick_t time = 0;
@@ -69,7 +69,7 @@ rarch_perf_tick_t rarch_get_perf_counter(void)
 
 #elif defined(__GNUC__) && !defined(RARCH_CONSOLE)
 
-#if defined(__i386__) || defined(__i486__)
+#if defined(__i386__) || defined(__i486__) || defined(__i686__)
    asm volatile ("rdtsc" : "=A" (time));
 #elif defined(__x86_64__)
    unsigned a, d;
@@ -85,3 +85,57 @@ rarch_perf_tick_t rarch_get_perf_counter(void)
 
    return time;
 }
+#endif
+
+#if defined(__x86_64__) || defined(__i386__) || defined(__i486__) || defined(__i686__)
+#define CPU_X86
+#endif
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#ifdef CPU_X86
+static void x86_cpuid(int func, int flags[4])
+{
+#ifdef __GNUC__
+   asm volatile("cpuid\n" :
+         "=a"(flags[0]),
+         "=b"(flags[1]),
+         "=c"(flags[2]),
+         "=d"(flags[3]) : "a"(func));
+#elif defined(_MSC_VER)
+   __cpuid(flags, func);
+#endif
+}
+#endif
+
+void rarch_get_cpu_features(struct rarch_cpu_features *cpu)
+{
+   memset(cpu, 0, sizeof(*cpu));
+
+#ifdef CPU_X86
+   int flags[4];
+   x86_cpuid(0, flags);
+
+   char vendor[13] = {0};
+   const int vendor_shuffle[3] = { flags[1], flags[3], flags[2] };
+   memcpy(vendor, vendor_shuffle, sizeof(vendor_shuffle));
+   RARCH_LOG("[CPUID]: Vendor: %s\n", vendor);
+
+   if (flags[0] < 1) // Does CPUID not support func = 1? (unlikely ...)
+      return;
+
+   x86_cpuid(1, flags);
+   cpu->sse  = flags[3] & (1 << 25);
+   cpu->sse2 = flags[3] & (1 << 26);
+
+   int avx_flags = (1 << 27) | (1 << 28);
+   cpu->avx = (flags[2] & avx_flags) == avx_flags; // Is this enough?
+
+   RARCH_LOG("[CPUID]: SSE:  %d\n", cpu->sse);
+   RARCH_LOG("[CPUID]: SSE2: %d\n", cpu->sse2);
+   RARCH_LOG("[CPUID]: AVX:  %d\n", cpu->avx);
+#endif
+}
+

performance.h

@@ -23,10 +23,6 @@
 #include "config.h"
 #endif
 
-#ifdef __linux
-#include "performance/performance_linux.h"
-#endif
-
 #include <stdint.h>
 typedef unsigned long long rarch_perf_tick_t;
 
@@ -44,6 +40,17 @@ rarch_perf_tick_t rarch_get_perf_counter(void);
 void rarch_perf_register(struct rarch_perf_counter *perf);
 void rarch_perf_log(void);
 
+struct rarch_cpu_features
+{
+   bool sse;
+   bool sse2;
+   bool vmx;
+   bool avx;
+   bool neon;
+};
+
+void rarch_get_cpu_features(struct rarch_cpu_features *cpu);
+
 #ifdef PERF_TEST
 
 #define RARCH_PERFORMANCE_INIT(X)  static rarch_perf_counter_t X = {#X}; \

retroarch.c

@@ -2539,6 +2539,30 @@ static void verify_api_version(void)
       RARCH_WARN("RetroArch is compiled against a different version of libretro than this libretro implementation.\n");
 }
 
+// Make sure we haven't compiled for something we cannot run.
+// Ideally, code would get swapped out depending on CPU support, but this will do for now.
+static void validate_cpu_features(void)
+{
+   struct rarch_cpu_features cpu;
+   rarch_get_cpu_features(&cpu);
+
+#ifdef __SSE2__
+   if (!cpu.sse2)
+   {
+      RARCH_ERR("SSE2 code is compiled in, but CPU does not support this feature. Cannot continue.\n");
+      rarch_fail(1, "validate_cpu_features()");
+   }
+#endif
+
+#ifdef __AVX__
+   if (!cpu.avx)
+   {
+      RARCH_ERR("AVX code is compiled in, but CPU does not support this feature. Cannot continue.\n");
+      rarch_fail(1, "validate_cpu_features()");
+   }
+#endif
+}
+
 int rarch_main_init(int argc, char *argv[])
 {
    init_state();
@@ -2550,7 +2574,6 @@ int rarch_main_init(int argc, char *argv[])
       return sjlj_ret;
    }
    g_extern.error_in_init = true;
-
    parse_input(argc, argv);
 
    if (g_extern.verbose)
@@ -2560,6 +2583,7 @@ int rarch_main_init(int argc, char *argv[])
       RARCH_LOG_OUTPUT("=================================================\n");
    }
 
+   validate_cpu_features();
    config_load();
 
    init_libretro_sym();
@@ -2792,9 +2816,6 @@ int rarch_main(int argc, char *argv[])
 // Consoles use the higher level API.
 int main(int argc, char *argv[])
 {
-#if 0
-   rarch_perf_get_cpu_features();
-#endif
    return rarch_main(argc, argv);
 }
 #endif

retroarch_logger.h

@@ -56,7 +56,7 @@
 #define RARCH_LOG_OUTPUT(...) do { \
       if (g_extern.verbose) \
       { \
-         fprintf(stderr, "stderr: " __VA_ARGS__); \
+         fprintf(stderr, __VA_ARGS__); \
          fflush(stderr); \
       } \
    } while (0)