Merge "add a kSlowSSSE3 feature for CPUInfo"

2 files changed, 41 insertions, 5 deletions

diff --git a/src/dsp/cpu.c b/src/dsp/cpu.c
index cbb08db9..b5583b6e 100644
--- a/src/dsp/cpu.c
+++ b/src/dsp/cpu.c
@@ -95,26 +95,62 @@ static WEBP_INLINE uint64_t xgetbv(void) {
 #endif
 
 #if defined(__i386__) || defined(__x86_64__) || defined(WEBP_MSC_SSE2)
+
+// helper function for run-time detection of slow SSSE3 platforms
+static int CheckSlowModel(int info) {
+  // Table listing display models with longer latencies for the bsr instruction
+  // (ie 2 cycles vs 10/16 cycles) and some SSSE3 instructions like pshufb.
+  // Refer to Intel 64 and IA-32 Architectures Optimization Reference Manual.
+  static const uint8_t kSlowModels[] = {
+    0x37, 0x4a, 0x4d,  // Silvermont Microarchitecture
+    0x1c, 0x26, 0x27   // Atom Microarchitecture
+  };
+  const uint32_t model = ((info & 0xf0000) >> 12) | ((info >> 4) & 0xf);
+  const uint32_t family = (info >> 8) & 0xf;
+  if (family == 0x06) {
+    size_t i;
+    for (i = 0; i < sizeof(kSlowModels) / sizeof(kSlowModels[0]); ++i) {
+      if (model == kSlowModels[i]) return 1;
+    }
+  }
+  return 0;
+}
+
 static int x86CPUInfo(CPUFeature feature) {
   int max_cpuid_value;
   int cpu_info[4];
+  int is_intel = 0;
 
   // get the highest feature value cpuid supports
   GetCPUInfo(cpu_info, 0);
   max_cpuid_value = cpu_info[0];
   if (max_cpuid_value < 1) {
     return 0;
+  } else {
+    const int VENDOR_ID_INTEL_EBX = 0x756e6547;  // uneG
+    const int VENDOR_ID_INTEL_EDX = 0x49656e69;  // Ieni
+    const int VENDOR_ID_INTEL_ECX = 0x6c65746e;  // letn
+    is_intel = (cpu_info[1] == VENDOR_ID_INTEL_EBX &&
+                cpu_info[2] == VENDOR_ID_INTEL_ECX &&
+                cpu_info[3] == VENDOR_ID_INTEL_EDX);    // genuine Intel?
   }
 
   GetCPUInfo(cpu_info, 1);
   if (feature == kSSE2) {
-    return 0 != (cpu_info[3] & 0x04000000);
+    return !!(cpu_info[3] & (1 << 26));
   }
   if (feature == kSSE3) {
-    return 0 != (cpu_info[2] & 0x00000001);
+    return !!(cpu_info[2] & (1 << 0));
+  }
+  if (feature == kSlowSSSE3) {
+    if (is_intel && (cpu_info[2] & (1 << 0))) {   // SSSE3?
+      return CheckSlowModel(cpu_info[0]);
+    }
+    return 0;
   }
+
   if (feature == kSSE4_1) {
-    return 0 != (cpu_info[2] & 0x00080000);
+    return !!(cpu_info[2] & (1 << 19));
   }
   if (feature == kAVX) {
     // bits 27 (OSXSAVE) & 28 (256-bit AVX)
@@ -126,7 +162,7 @@ static int x86CPUInfo(CPUFeature feature) {
   if (feature == kAVX2) {
     if (x86CPUInfo(kAVX) && max_cpuid_value >= 7) {
       GetCPUInfo(cpu_info, 7);
-      return ((cpu_info[1] & 0x00000020) == 0x00000020);
+      return !!(cpu_info[1] & (1 << 5));
     }
   }
   return 0;
@@ -184,4 +220,3 @@ VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo;
 #else
 VP8CPUInfo VP8GetCPUInfo = NULL;
 #endif
-
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index e0fb4e5d..d6d50e77 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -132,6 +132,7 @@ extern "C" {
 typedef enum {
   kSSE2,
   kSSE3,
+  kSlowSSSE3,  // special feature for slow SSSE3 architectures
   kSSE4_1,
   kAVX,
   kAVX2,