Merge pull request #4548 from stenzek/gcc-sse

Support SSSE3 texture decoders and CRC32 hashing on non-native builds (gcc)

Source/Core/Common/Hash.cpp

@@ -238,9 +238,11 @@ u64 GetMurmurHash3(const u8* src, u32 len, u32 samples)
 }
 
 // CRC32 hash using the SSE4.2 instruction
+#if defined(_M_X86_64)
+
+FUNCTION_TARGET_SSE42
 u64 GetCRC32(const u8* src, u32 len, u32 samples)
 {
-#if _M_SSE >= 0x402 || defined(_M_ARM_64)
   u64 h[4] = {len, 0, 0, 0};
   u32 Step = (len / 8);
   const u64* data = (const u64*)src;
@@ -250,9 +252,7 @@ u64 GetCRC32(const u8* src, u32 len, u32 samples)
   Step = Step / samples;
   if (Step < 1)
     Step = 1;
-#endif
 
-#if _M_SSE >= 0x402
   while (data < end - Step * 3)
   {
     h[0] = _mm_crc32_u64(h[0], data[Step * 0]);
@@ -274,7 +274,25 @@ u64 GetCRC32(const u8* src, u32 len, u32 samples)
     memcpy(&temp, end, len & 7);
     h[0] = _mm_crc32_u64(h[0], temp);
   }
+
+  // FIXME: is there a better way to combine these partial hashes?
+  return h[0] + (h[1] << 10) + (h[2] << 21) + (h[3] << 32);
+}
+
 #elif defined(_M_ARM_64)
+
+u64 GetCRC32(const u8* src, u32 len, u32 samples)
+{
+  u64 h[4] = {len, 0, 0, 0};
+  u32 Step = (len / 8);
+  const u64* data = (const u64*)src;
+  const u64* end = data + Step;
+  if (samples == 0)
+    samples = std::max(Step, 1u);
+  Step = Step / samples;
+  if (Step < 1)
+    Step = 1;
+
   // We should be able to use intrinsics for this
   // Too bad the intrinsics for this instruction was added in GCC 4.9.1
   // The Android NDK (as of r10e) only has GCC 4.9
@@ -317,16 +335,20 @@ u64 GetCRC32(const u8* src, u32 len, u32 samples)
         : [res] "=r"(h[0])
         : [two] "r"(h[0]), [three] "r"(temp));
   }
-#endif
 
-#if _M_SSE >= 0x402 || defined(_M_ARM_64)
   // FIXME: is there a better way to combine these partial hashes?
   return h[0] + (h[1] << 10) + (h[2] << 21) + (h[3] << 32);
-#else
-  return 0;
-#endif
 }
 
+#else
+
+u64 GetCRC32(const u8* src, u32 len, u32 samples)
+{
+  return 0;
+}
+
+#endif
+
 /*
  * NOTE: This hash function is used for custom texture loading/dumping, so
  * it should not be changed, which would require all custom textures to be
@@ -386,10 +408,13 @@ u64 GetHashHiresTexture(const u8* src, u32 len, u32 samples)
   return h;
 }
 #else
+
 // CRC32 hash using the SSE4.2 instruction
+#if defined(_M_X86)
+
+FUNCTION_TARGET_SSE42
 u64 GetCRC32(const u8* src, u32 len, u32 samples)
 {
-#if _M_SSE >= 0x402
   u32 h = len;
   u32 Step = (len / 4);
   const u32* data = (const u32*)src;
@@ -407,11 +432,17 @@ u64 GetCRC32(const u8* src, u32 len, u32 samples)
 
   const u8* data2 = (const u8*)end;
   return (u64)_mm_crc32_u32(h, u32(data2[0]));
-#else
-  return 0;
-#endif
 }
 
+#else
+
+u64 GetCRC32(const u8* src, u32 len, u32 samples)
+{
+  return 0;
+}
+
+#endif
+
 //-----------------------------------------------------------------------------
 // Block read - if your platform needs to do endian-swapping or can only
 // handle aligned reads, do the conversion here
@@ -606,7 +637,7 @@ u64 GetHash64(const u8* src, u32 len, u32 samples)
 // sets the hash function used for the texture cache
 void SetHash64Function()
 {
-#if _M_SSE >= 0x402
+#if defined(_M_X86_64) || defined(_M_X86)
   if (cpu_info.bSSE4_2)  // sse crc32 version
   {
     ptrHashFunction = &GetCRC32;

Source/Core/Common/Intrinsics.h

@@ -4,28 +4,69 @@
 
 #pragma once
 
-#ifdef _M_X86
+#if defined(_M_X86)
+
+/**
+ * It is assumed that all compilers used to build Dolphin support intrinsics up to and including
+ * SSE 4.2 on x86/x64.
+ */
+
+#if defined(__GNUC__) || defined(__clang__)
+
+/**
+ * Due to limitations in GCC, SSE intrinsics are only available when compiling with the
+ * corresponding instruction set enabled. However, using the target attribute, we can compile
+ * single functions with a different target instruction set, while still creating a generic build.
+ *
+ * Since this instruction set is enabled per-function, any callers should verify that the
+ * instruction set is supported at runtime before calling it, and provide a fallback implementation
+ * when not supported.
+ *
+ * When building with -march=native, or enabling the instruction sets in the compile flags, permit
+ * usage of the instrinsics without any function attributes. If the command-line architecture does
+ * not support this instruction set, enable it via function targeting.
+*/
 
-#ifdef _MSC_VER
-#include <intrin.h>
-#else
 #include <x86intrin.h>
+#ifndef __SSE4_2__
+#define FUNCTION_TARGET_SSE42 [[gnu::target("sse4.2")]]
+#endif
+#ifndef __SSE4_1__
+#define FUNCTION_TARGET_SSR41 [[gnu::target("sse4.1")]]
+#endif
+#ifndef __SSSE3__
+#define FUNCTION_TARGET_SSSE3 [[gnu::target("ssse3")]]
+#endif
+#ifndef __SSE3__
+#define FUNCTION_TARGET_SSE3 [[gnu::target("sse3")]]
 #endif
 
-#if defined _M_GENERIC
-#define _M_SSE 0
-#elif _MSC_VER || __INTEL_COMPILER
-#define _M_SSE 0x402
-#elif defined __GNUC__
-#if defined __SSE4_2__
-#define _M_SSE 0x402
-#elif defined __SSE4_1__
-#define _M_SSE 0x401
-#elif defined __SSSE3__
-#define _M_SSE 0x301
-#elif defined __SSE3__
-#define _M_SSE 0x300
-#endif
-#endif
+#elif defined(_MSC_VER) || defined(__INTEL_COMPILER)
+
+/**
+ * MSVC and ICC support intrinsics for any instruction set without any function attributes.
+ */
+#include <intrin.h>
+
+#endif  // defined(_MSC_VER) || defined(__INTEL_COMPILER)
 
 #endif  // _M_X86
+
+/**
+ * Define the FUNCTION_TARGET macros to nothing if they are not needed, or not on an X86 platform.
+ * This way when a function is defined with FUNCTION_TARGET you don't need to define a second
+ * version without the macro around a #ifdef guard. Be careful when using intrinsics, as all use
+ * should still be placed around a #ifdef _M_X86 if the file is compiled on all architectures.
+ */
+#ifndef FUNCTION_TARGET_SSE42
+#define FUNCTION_TARGET_SSE42
+#endif
+#ifndef FUNCTION_TARGET_SSR41
+#define FUNCTION_TARGET_SSR41
+#endif
+#ifndef FUNCTION_TARGET_SSSE3
+#define FUNCTION_TARGET_SSSE3
+#endif
+#ifndef FUNCTION_TARGET_SSE3
+#define FUNCTION_TARGET_SSE3
+#endif