[X86] Add AVX-VNNI-INT8 instructions.

For more details about these instructions, please refer to the latest ISE document: https://www.intel.com/content/www/us/en/develop/download/intel-architecture-instruction-set-extensions-programming-reference.html

Reviewed By: pengfei, skan

Differential Revision: https://reviews.llvm.org/D135938

Author: FreddyLeaf

clang/docs/ReleaseNotes.rst

@@ -648,6 +648,10 @@ X86 Support in Clang
 - Support ISA of ``AVX-IFMA``.
   * Support intrinsic of ``_mm(256)_madd52hi_avx_epu64``.
   * Support intrinsic of ``_mm(256)_madd52lo_avx_epu64``.
+- Support ISA of ``AVX-VNNI-INT8``.
+  * Support intrinsic of ``_mm(256)_dpbssd(s)_epi32``.
+  * Support intrinsic of ``_mm(256)_dpbsud(s)_epi32``.
+  * Support intrinsic of ``_mm(256)_dpbuud(s)_epi32``.
 
 WebAssembly Support in Clang
 ----------------------------

clang/include/clang/Basic/BuiltinsX86.def

@@ -906,6 +906,7 @@ TARGET_BUILTIN(__builtin_ia32_alignq256, "V4OiV4OiV4OiIi", "ncV:256:", "avx512vl
 TARGET_BUILTIN(__builtin_ia32_extractf64x4_mask, "V4dV8dIiV4dUc", "ncV:512:", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_extractf32x4_mask, "V4fV16fIiV4fUc", "ncV:512:", "avx512f")
 
+// AVX-VNNI and AVX512-VNNI
 TARGET_BUILTIN(__builtin_ia32_vpdpbusd128, "V4iV4iV4iV4i", "ncV:128:", "avx512vl,avx512vnni|avxvnni")
 TARGET_BUILTIN(__builtin_ia32_vpdpbusd256, "V8iV8iV8iV8i", "ncV:256:", "avx512vl,avx512vnni|avxvnni")
 TARGET_BUILTIN(__builtin_ia32_vpdpbusd512, "V16iV16iV16iV16i", "ncV:512:", "avx512vnni")
@@ -919,6 +920,20 @@ TARGET_BUILTIN(__builtin_ia32_vpdpwssds128, "V4iV4iV4iV4i", "ncV:128:", "avx512v
 TARGET_BUILTIN(__builtin_ia32_vpdpwssds256, "V8iV8iV8iV8i", "ncV:256:", "avx512vl,avx512vnni|avxvnni")
 TARGET_BUILTIN(__builtin_ia32_vpdpwssds512, "V16iV16iV16iV16i", "ncV:512:", "avx512vnni")
 
+// AVX-VNNI-INT8
+TARGET_BUILTIN(__builtin_ia32_vpdpbssd128, "V4iV4iV4iV4i", "ncV:128:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbssd256, "V8iV8iV8iV8i", "ncV:256:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbssds128, "V4iV4iV4iV4i", "ncV:128:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbssds256, "V8iV8iV8iV8i", "ncV:256:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbsud128, "V4iV4iV4iV4i", "ncV:128:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbsud256, "V8iV8iV8iV8i", "ncV:256:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbsuds128, "V4iV4iV4iV4i", "ncV:128:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbsuds256, "V8iV8iV8iV8i", "ncV:256:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbuud128, "V4iV4iV4iV4i", "ncV:128:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbuud256, "V8iV8iV8iV8i", "ncV:256:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbuuds128, "V4iV4iV4iV4i", "ncV:128:", "avxvnniint8")
+TARGET_BUILTIN(__builtin_ia32_vpdpbuuds256, "V8iV8iV8iV8i", "ncV:256:", "avxvnniint8")
+
 TARGET_BUILTIN(__builtin_ia32_gather3div2df, "V2dV2dvC*V2OiUcIi", "nV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_gather3div2di, "V2OiV2OivC*V2OiUcIi", "nV:128:", "avx512vl")
 TARGET_BUILTIN(__builtin_ia32_gather3div4df, "V4dV4dvC*V4OiUcIi", "nV:256:", "avx512vl")

clang/include/clang/Driver/Options.td

@@ -4590,6 +4590,8 @@ def mavx512vp2intersect : Flag<["-"], "mavx512vp2intersect">, Group<m_x86_Featur
 def mno_avx512vp2intersect : Flag<["-"], "mno-avx512vp2intersect">, Group<m_x86_Features_Group>;
 def mavxifma : Flag<["-"], "mavxifma">, Group<m_x86_Features_Group>;
 def mno_avxifma : Flag<["-"], "mno-avxifma">, Group<m_x86_Features_Group>;
+def mavxvnniint8 : Flag<["-"], "mavxvnniint8">, Group<m_x86_Features_Group>;
+def mno_avxvnniint8 : Flag<["-"], "mno-avxvnniint8">, Group<m_x86_Features_Group>;
 def mavxvnni : Flag<["-"], "mavxvnni">, Group<m_x86_Features_Group>;
 def mno_avxvnni : Flag<["-"], "mno-avxvnni">, Group<m_x86_Features_Group>;
 def madx : Flag<["-"], "madx">, Group<m_x86_Features_Group>;

clang/lib/Basic/Targets/X86.cpp

@@ -342,6 +342,8 @@ bool X86TargetInfo::handleTargetFeatures(std::vector<std::string> &Features,
       HasAVXIFMA = true;
     } else if (Feature == "+avxvnni") {
       HasAVXVNNI = true;
+    } else if (Feature == "+avxvnniint8") {
+      HasAVXVNNIINT8 = true;
     } else if (Feature == "+serialize") {
       HasSERIALIZE = true;
     } else if (Feature == "+tsxldtrk") {
@@ -796,6 +798,8 @@ void X86TargetInfo::getTargetDefines(const LangOptions &Opts,
     Builder.defineMacro("__AVXIFMA__");
   if (HasAVXVNNI)
     Builder.defineMacro("__AVXVNNI__");
+  if (HasAVXVNNIINT8)
+    Builder.defineMacro("__AVXVNNIINT8__");
   if (HasSERIALIZE)
     Builder.defineMacro("__SERIALIZE__");
   if (HasTSXLDTRK)
@@ -920,6 +924,7 @@ bool X86TargetInfo::isValidFeatureName(StringRef Name) const {
       .Case("avx512vp2intersect", true)
       .Case("avxifma", true)
       .Case("avxvnni", true)
+      .Case("avxvnniint8", true)
       .Case("bmi", true)
       .Case("bmi2", true)
       .Case("cldemote", true)
@@ -1019,6 +1024,8 @@ bool X86TargetInfo::hasFeature(StringRef Feature) const {
       .Case("avx512vp2intersect", HasAVX512VP2INTERSECT)
       .Case("avxifma", HasAVXIFMA)
       .Case("avxvnni", HasAVXVNNI)
+      .Case("avxvnni", HasAVXVNNI)
+      .Case("avxvnniint8", HasAVXVNNIINT8)
       .Case("bmi", HasBMI)
       .Case("bmi2", HasBMI2)
       .Case("cldemote", HasCLDEMOTE)

clang/lib/Basic/Targets/X86.h

@@ -141,6 +141,7 @@ class LLVM_LIBRARY_VISIBILITY X86TargetInfo : public TargetInfo {
   bool HasAMXFP16 = false;
   bool HasCMPCCXADD = false;
   bool HasRAOINT = false;
+  bool HasAVXVNNIINT8 = false;
   bool HasKL = false;      // For key locker
   bool HasWIDEKL = false; // For wide key locker
   bool HasHRESET = false;

clang/lib/Headers/CMakeLists.txt

@@ -144,6 +144,7 @@ set(x86_files
   avx512vpopcntdqvlintrin.h
   avxifmaintrin.h
   avxintrin.h
+  avxvnniint8intrin.h
   avxvnniintrin.h
   bmi2intrin.h
   bmiintrin.h