yencode 1.0.8 → 1.1.2

package/binding.gyp

@@ -1,40 +1,293 @@
-{
-  "targets": [
-    {
-      "target_name": "yencode",
-      "dependencies": ["crcutil"],
-      "sources": ["yencode.cc"],
-      "conditions": [
-        ['OS=="win"', {
-          "msvs_settings": {"VCCLCompilerTool": {"EnableEnhancedInstructionSet": "2"}}
-        }, {
-          "cflags": ["-march=native", "-O2"],
-          "cxxflags": ["-march=native", "-O2"]
-        }]
-      ],
-      "include_dirs": ["crcutil-1.0/code","crcutil-1.0/examples"]
-    },
-    {
-      "target_name": "crcutil",
-      "type": "static_library",
-      "sources": [
-        "crcutil-1.0/code/crc32c_sse4.cc",
-        "crcutil-1.0/code/multiword_64_64_cl_i386_mmx.cc",
-        "crcutil-1.0/code/multiword_64_64_gcc_amd64_asm.cc",
-        "crcutil-1.0/code/multiword_64_64_gcc_i386_mmx.cc",
-        "crcutil-1.0/code/multiword_64_64_intrinsic_i386_mmx.cc",
-        "crcutil-1.0/code/multiword_128_64_gcc_amd64_sse2.cc",
-        "crcutil-1.0/examples/interface.cc"
-      ],
-      "conditions": [
-        ['OS=="win"', {
-          "msvs_settings": {"VCCLCompilerTool": {"EnableEnhancedInstructionSet": "2"}}
-        }, {
-          "cxxflags": ["-march=native", "-O3", "-fomit-frame-pointer"]
-        }]
-      ],
-      "include_dirs": ["crcutil-1.0/code", "crcutil-1.0/tests"],
-      "defines": ["CRCUTIL_USE_MM_CRC32=0"]
-    }
-  ]
+{
+  "variables": {
+    "enable_native_tuning%": 1,
+    "disable_avx256%": 0
+  },
+  "target_defaults": {
+    "conditions": [
+      ['target_arch=="ia32"', {
+        "msvs_settings": {"VCCLCompilerTool": {"EnableEnhancedInstructionSet": "2"}}
+      }],
+      ['OS!="win" and enable_native_tuning!=0', {
+        "variables": {
+          "supports_native%": "<!(<!(echo ${CXX_target:-${CXX:-c++}}) -MM -E src/common.h -march=native 2>/dev/null || true)"
+        },
+        "conditions": [
+          ['supports_native!=""', {
+            "cflags": ["-march=native"],
+            "cxxflags": ["-march=native"],
+            "xcode_settings": {
+              "OTHER_CFLAGS": ["-march=native"],
+              "OTHER_CXXFLAGS": ["-march=native"],
+            }
+          }]
+        ]
+      }],
+      ['OS!="win" and target_arch in "ia32 x64" and enable_native_tuning==0', {
+        "variables": {
+          "supports_avx2_nosplit%": "<!(<!(echo ${CXX_target:-${CXX:-c++}}) -MM -E src/common.h -mavx2 -mno-avx256-split-unaligned-load -mno-avx256-split-unaligned-store 2>/dev/null || true)"
+        },
+        "conditions": [
+          ['supports_avx2_nosplit!=""', {
+            "cflags": ["-mno-avx256-split-unaligned-load", "-mno-avx256-split-unaligned-store"],
+            "cxxflags": ["-mno-avx256-split-unaligned-load", "-mno-avx256-split-unaligned-store"],
+            "xcode_settings": {
+              "OTHER_CFLAGS": ["-mno-avx256-split-unaligned-load", "-mno-avx256-split-unaligned-store"],
+              "OTHER_CXXFLAGS": ["-mno-avx256-split-unaligned-load", "-mno-avx256-split-unaligned-store"],
+            }
+          }]
+        ]
+      }],
+      ['disable_avx256!=0', {
+        "defines": ["YENC_DISABLE_AVX256=1"]
+      }],
+      ['OS!="win" and enable_native_tuning!=0', {
+        "defines": ["YENC_BUILD_NATIVE=1"]
+      }]
+    ],
+    "cflags": ["-Wno-unused-function"],
+    "cxxflags": ["-Wno-unused-function"],
+    "xcode_settings": {
+      "OTHER_CFLAGS": ["-Wno-unused-function"],
+      "OTHER_CXXFLAGS": ["-Wno-unused-function"]
+    },
+    "msvs_settings": {"VCCLCompilerTool": {"Optimization": "MaxSpeed"}},
+    "configurations": {"Release": {
+      "cflags": ["-fomit-frame-pointer"],
+      "cxxflags": ["-fomit-frame-pointer"],
+      "xcode_settings": {
+        "OTHER_CFLAGS": ["-fomit-frame-pointer"],
+        "OTHER_CXXFLAGS": ["-fomit-frame-pointer"]
+      }
+    }}
+  },
+  "targets": [
+    {
+      "target_name": "yencode",
+      "dependencies": ["crcutil", "yencode_sse2", "yencode_ssse3", "yencode_clmul", "yencode_avx", "yencode_avx2", "yencode_neon", "yencode_armcrc"],
+      "sources": [
+        "src/yencode.cc",
+        "src/platform.cc",
+        "src/encoder.cc",
+        "src/decoder.cc",
+        "src/crc.cc"
+      ],
+      "include_dirs": ["crcutil-1.0/code","crcutil-1.0/examples"]
+    },
+    {
+      "target_name": "yencode_sse2",
+      "type": "static_library",
+      "sources": [
+        "src/encoder_sse2.cc",
+        "src/decoder_sse2.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch in "ia32 x64"', {
+          "cflags": ["-msse2"],
+          "cxxflags": ["-msse2"],
+          "xcode_settings": {
+            "OTHER_CFLAGS": ["-msse2"],
+            "OTHER_CXXFLAGS": ["-msse2"],
+          }
+        }]
+      ]
+    },
+    {
+      "target_name": "yencode_ssse3",
+      "type": "static_library",
+      "sources": [
+        "src/encoder_ssse3.cc",
+        "src/decoder_ssse3.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch in "ia32 x64"', {
+          "cflags": ["-mssse3"],
+          "cxxflags": ["-mssse3"],
+          "xcode_settings": {
+            "OTHER_CFLAGS": ["-mssse3"],
+            "OTHER_CXXFLAGS": ["-mssse3"],
+          }
+        }]
+      ]
+    },
+    {
+      "target_name": "yencode_clmul",
+      "type": "static_library",
+      "sources": [
+        "src/crc_folding.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch in "ia32 x64"', {
+          "cflags": ["-mssse3", "-msse4.1", "-mpclmul"],
+          "cxxflags": ["-mssse3", "-msse4.1", "-mpclmul"],
+          "xcode_settings": {
+            "OTHER_CFLAGS": ["-mssse3", "-msse4.1", "-mpclmul"],
+            "OTHER_CXXFLAGS": ["-mssse3", "-msse4.1", "-mpclmul"],
+          }
+        }]
+      ]
+    },
+    {
+      "target_name": "yencode_avx",
+      "type": "static_library",
+      "sources": [
+        "src/encoder_avx.cc",
+        "src/decoder_avx.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch in "ia32 x64"', {
+          "msvs_settings": {"VCCLCompilerTool": {"EnableEnhancedInstructionSet": "3"}},
+          "cflags": ["-mavx", "-mpopcnt"],
+          "cxxflags": ["-mavx", "-mpopcnt"],
+          "xcode_settings": {
+            "OTHER_CFLAGS": ["-mavx", "-mpopcnt"],
+            "OTHER_CXXFLAGS": ["-mavx", "-mpopcnt"],
+          }
+        }]
+      ]
+    },
+    {
+      "target_name": "yencode_avx2",
+      "type": "static_library",
+      "sources": [
+        "src/decoder_avx2.cc", "src/encoder_avx2.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch in "ia32 x64" and OS!="win"', {
+          "variables": {"supports_avx2%": "<!(<!(echo ${CXX_target:-${CXX:-c++}}) -MM -E src/decoder_avx2.cc -mavx2 2>/dev/null || true)"},
+          "conditions": [
+            ['supports_avx2!=""', {
+              "cflags": ["-mavx2", "-mpopcnt", "-mbmi", "-mbmi2", "-mlzcnt"],
+              "cxxflags": ["-mavx2", "-mpopcnt", "-mbmi", "-mbmi2", "-mlzcnt"],
+              "xcode_settings": {
+                "OTHER_CFLAGS": ["-mavx2", "-mpopcnt", "-mbmi", "-mbmi2", "-mlzcnt"],
+                "OTHER_CXXFLAGS": ["-mavx2", "-mpopcnt", "-mbmi", "-mbmi2", "-mlzcnt"],
+              }
+            }]
+          ]
+        }],
+        ['target_arch in "ia32 x64" and OS=="win"', {
+          "msvs_settings": {"VCCLCompilerTool": {"EnableEnhancedInstructionSet": "3"}}
+        }]
+      ]
+    },
+    {
+      "target_name": "yencode_neon",
+      "type": "static_library",
+      "sources": [
+        "src/encoder_neon.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch=="arm"', {
+          "cflags": ["-mfpu=neon"],
+          "cxxflags": ["-mfpu=neon"],
+          "xcode_settings": {
+            "OTHER_CFLAGS": ["-mfpu=neon"],
+            "OTHER_CXXFLAGS": ["-mfpu=neon"],
+          }
+        }],
+        ['target_arch=="arm64"', {
+          "sources": ["src/decoder_neon64.cc"]
+        }, {
+          "sources": ["src/decoder_neon.cc"]
+        }]
+      ]
+    },
+    {
+      "target_name": "yencode_armcrc",
+      "type": "static_library",
+      "sources": [
+        "src/crc_arm.cc"
+      ],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "conditions": [
+        ['target_arch in "arm arm64"', {
+          "cflags!": ["-march=native"],
+          "cxxflags!": ["-march=native"],
+          "cflags": ["-march=armv8-a+crc"],
+          "cxxflags": ["-march=armv8-a+crc"],
+          "xcode_settings": {
+            "OTHER_CFLAGS!": ["-march=native"],
+            "OTHER_CXXFLAGS!": ["-march=native"],
+            "OTHER_CFLAGS": ["-march=armv8-a+crc"],
+            "OTHER_CXXFLAGS": ["-march=armv8-a+crc"],
+          }
+        }]
+      ]
+    },
+    {
+      "target_name": "crcutil",
+      "type": "static_library",
+      "sources": [
+        "crcutil-1.0/code/crc32c_sse4.cc",
+        "crcutil-1.0/code/multiword_64_64_cl_i386_mmx.cc",
+        "crcutil-1.0/code/multiword_64_64_gcc_amd64_asm.cc",
+        "crcutil-1.0/code/multiword_64_64_gcc_i386_mmx.cc",
+        "crcutil-1.0/code/multiword_64_64_intrinsic_i386_mmx.cc",
+        "crcutil-1.0/code/multiword_128_64_gcc_amd64_sse2.cc",
+        "crcutil-1.0/examples/interface.cc"
+      ],
+      "cflags": ["-fomit-frame-pointer", "-Wno-expansion-to-defined"],
+      "cxxflags": ["-fomit-frame-pointer", "-Wno-expansion-to-defined"],
+      "cflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "cxxflags!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+      "xcode_settings": {
+        "OTHER_CFLAGS": ["-fomit-frame-pointer", "-Wno-expansion-to-defined"],
+        "OTHER_CXXFLAGS": ["-fomit-frame-pointer", "-Wno-expansion-to-defined"],
+        "OTHER_CFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"],
+        "OTHER_CXXFLAGS!": ["-fno-omit-frame-pointer", "-fno-tree-vrp", "-fno-strict-aliasing"]
+      },
+      "msvs_settings": {"VCCLCompilerTool": {"BufferSecurityCheck": "false"}},
+      "include_dirs": ["crcutil-1.0/code", "crcutil-1.0/tests"],
+      "defines": ["CRCUTIL_USE_MM_CRC32=0"]
+    }
+  ]
 }

package/crcutil-1.0/code/multiword_64_64_gcc_amd64_asm.cc

@@ -142,7 +142,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordGccAmd64(
   uint64 crc3;
 
   asm(
-    "sub $2*4*8 - 1, %[end]\n"
+    "subq $2*4*8 - 1, %[end]\n"
     "cmpq  %[src], %[end]\n"
     "jbe 2f\n"
     "xorq %[crc1], %[crc1]\n"
@@ -157,7 +157,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordGccAmd64(
 #if HAVE_SSE && CRCUTIL_PREFETCH_WIDTH > 0
     "prefetcht0 " TO_STRING(CRCUTIL_PREFETCH_WIDTH) "(%[src])\n"
 #endif  // HAVE_SSE
-    "add $4*8, %[src]\n"
+    "addq $4*8, %[src]\n"
 
     // Set buffer data.
     "xorq %[crc0], " BUF0 "\n"
@@ -237,22 +237,22 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordGccAmd64(
     "movq " BUF3 ", " BUF0 "\n"
     CRC_WORD_ASM()
 
-    "add $4*8, %[src]\n"
+    "addq $4*8, %[src]\n"
 
     "2:\n"
-    "add $2*4*8 - 8, %[end]\n"
+    "addq $2*4*8 - 8, %[end]\n"
     "cmpq %[src], %[end]\n"
     "jbe 4f\n"
 
     "3:\n"
     "movq (%[src]), " BUF0 "\n"
-    "add $8, %[src]\n"
+    "addq $8, %[src]\n"
     CRC_WORD_ASM()
     "cmpq %[src], %[end]\n"
     "ja 3b\n"
 
     "4:\n"
-    "add $7, %[end]\n"
+    "addq $7, %[end]\n"
 
     "cmpq %[src], %[end]\n"
     "jbe 6f\n"
@@ -262,7 +262,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordGccAmd64(
     "movzbq %b[crc0], " TMP0 "\n"
     "shrq  $8, %[crc0]\n"
     "xorq " BUF0 ", " TMP0 "\n"
-    "add $1, %[src]\n"
+    "addq $1, %[src]\n"
     "xorq 7*256*8(%[table_word], " TMP0 ", 8), %[crc0]\n"
     "cmpq %[src], %[end]\n"
     "ja 5b\n"

package/crcutil-1.0/code/multiword_64_64_gcc_i386_mmx.cc

@@ -16,7 +16,7 @@
 
 #include "generic_crc.h"
 
-#if defined(__GNUC__) && CRCUTIL_USE_ASM && HAVE_I386 && HAVE_MMX
+#if defined(__GNUC__) && !defined(__clang__) && CRCUTIL_USE_ASM && HAVE_I386 && HAVE_MMX
 
 #if defined(__PIC__) && __GNUC__ < 5
 /* workaround for issue with PIC reserving ebx: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54232 */
@@ -111,7 +111,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     #define TMP3 "%[tmp3]"
     #define TMP3b "%b[tmp3]"
 #endif
-    "sub $2*4*8 - 1, %[end]\n"
+    "subl $2*4*8 - 1, %[end]\n"
     "cmpl  %[src], %[end]\n"
     "jbe 2f\n"
 
@@ -140,7 +140,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "psrlq $32, %[buf1]\n"
     "movd %[buf2], %[tmp2]\n"
     "psrlq $32, %[buf2]\n"
-    "movd %[buf3], "TMP3"\n"
+    "movd %[buf3], " TMP3 "\n"
     "psrlq $32, %[buf3]\n"
 
     "movzbl %b[tmp0], %[temp]\n"
@@ -152,8 +152,8 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "movzbl %b[tmp2], %[temp]\n"
     "shrl $8, %[tmp2]\n"
     "movq (%[table], %[temp], 8), %[crc2]\n"
-    "movzbl "TMP3b", %[temp]\n"
-    "shrl $8, "TMP3"\n"
+    "movzbl " TMP3b ", %[temp]\n"
+    "shrl $8, " TMP3 "\n"
     "movq (%[table], %[temp], 8), %[crc3]\n"
 
 #define XOR(byte) \
@@ -166,8 +166,8 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "movzbl %b[tmp2], %[temp]\n" \
     "shrl $8, %[tmp2]\n" \
     "pxor " #byte "*256*8(%[table], %[temp], 8), %[crc2]\n" \
-    "movzbl "TMP3b", %[temp]\n" \
-    "shrl $8, "TMP3"\n" \
+    "movzbl " TMP3b ", %[temp]\n" \
+    "shrl $8, " TMP3 "\n" \
     "pxor " #byte "*256*8(%[table], %[temp], 8), %[crc3]\n"
 
     XOR(1)
@@ -179,8 +179,8 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "movd %[buf1], %[tmp1]\n"
     "pxor 3*256*8(%[table], %[tmp2], 8), %[crc2]\n"
     "movd %[buf2], %[tmp2]\n"
-    "pxor 3*256*8(%[table], "TMP3", 8), %[crc3]\n"
-    "movd %[buf3], "TMP3"\n"
+    "pxor 3*256*8(%[table], " TMP3 ", 8), %[crc3]\n"
+    "movd %[buf3], " TMP3 "\n"
 
     XOR(4)
     XOR(5)
@@ -192,7 +192,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "movq 1*8(%[src]), %[buf1]\n"
     "pxor 7*256*8(%[table], %[tmp2], 8), %[crc2]\n"
     "movq 2*8(%[src]), %[buf2]\n"
-    "pxor 7*256*8(%[table], "TMP3", 8), %[crc3]\n"
+    "pxor 7*256*8(%[table], " TMP3 ", 8), %[crc3]\n"
     "movq 3*8(%[src]), %[buf3]\n"
     "cmpl %[src], %[end]\n"
     "ja 1b\n"
@@ -213,11 +213,11 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "movq %[buf3], %[buf0]\n"
     CRC_WORD_MMX()
 
-    "add $4*8, %[src]\n"
+    "addl $4*8, %[src]\n"
     "2:\n"
     "movl %[table_word], %[table]\n"
 
-    "add $2*4*8 - 8, %[end]\n"
+    "addl $2*4*8 - 8, %[end]\n"
     "cmpl %[src], %[end]\n"
     "jbe 4f\n"
     "3:\n"
@@ -227,7 +227,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "cmpl %[src], %[end]\n"
     "ja 3b\n"
     "4:\n"
-    "add $7, %[end]\n"
+    "addl $7, %[end]\n"
 
     "cmpl %[src], %[end]\n"
     "jbe 6f\n"
@@ -238,7 +238,7 @@ template<> uint64 GenericCrc<uint64, uint64, uint64, 4>::CrcMultiwordI386Mmx(
     "movzbl %b[tmp0], %[tmp0]\n"
     "psrlq $8, %[crc0]\n"
     "xorl %[tmp0], %[temp]\n"
-    "add $1, %[src]\n"
+    "addl $1, %[src]\n"
     "pxor 7*256*8(%[table], %[temp], 8), %[crc0]\n"
     "cmpl %[src], %[end]\n"
     "ja 5b\n"

package/crcutil-1.0/code/multiword_64_64_intrinsic_i386_mmx.cc

@@ -16,7 +16,7 @@
 
 #include "generic_crc.h"
 
-#if CRCUTIL_USE_ASM && HAVE_I386 && HAVE_MMX
+#if CRCUTIL_USE_ASM && HAVE_I386 && HAVE_MMX && !defined(_MSC_VER) && !(defined(__GNUC__) && !defined(__clang__))
 
 namespace crcutil {
 

package/crcutil-1.0/code/uint128_sse2.h

@@ -70,6 +70,8 @@ class uint128_sse2 {
   }
   __forceinline uint128_sse2(const __m128i x) : x_(x) {
   }
+  __forceinline uint128_sse2(const uint128_sse2& x) : x_(x.x_) {
+  }
   __forceinline operator __m128i() const {
     return x_;
   }