[X86] Add AVX-VNNI-INT16 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/D155145

Author: FreddyLeaf

clang/docs/ReleaseNotes.rst

@@ -824,6 +824,10 @@ X86 Support
 - Support ISA of ``SM4``.
   * Support intrinsic of ``_mm(256)_sm4key4_epi32``.
   * Support intrinsic of ``_mm(256)_sm4rnds4_epi32``.
+- Support ISA of ``AVX-VNNI-INT16``.
+  * Support intrinsic of ``_mm(256)_dpwsud(s)_epi32``.
+  * Support intrinsic of ``_mm(256)_dpwusd(s)_epi32``.
+  * Support intrinsic of ``_mm(256)_dpwuud(s)_epi32``.
 
 Arm and AArch64 Support
 ^^^^^^^^^^^^^^^^^^^^^^^

clang/include/clang/Basic/BuiltinsX86.def

@@ -2116,6 +2116,20 @@ TARGET_HEADER_BUILTIN(__readgsword,  "UsUNi", "nh", INTRIN_H, ALL_MS_LANGUAGES,
 TARGET_HEADER_BUILTIN(__readgsdword, "UNiUNi", "nh", INTRIN_H, ALL_MS_LANGUAGES, "")
 TARGET_HEADER_BUILTIN(__readgsqword, "ULLiUNi", "nh", INTRIN_H, ALL_MS_LANGUAGES, "")
 
+// AVX-VNNI-INT16
+TARGET_BUILTIN(__builtin_ia32_vpdpwsud128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwsud256, "V8iV8iV8iV8i", "nV:256:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwsuds128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwsuds256, "V8iV8iV8iV8i", "nV:256:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwusd128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwusd256, "V8iV8iV8iV8i", "nV:256:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwusds128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwusds256, "V8iV8iV8iV8i", "nV:256:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwuud128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwuud256, "V8iV8iV8iV8i", "nV:256:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwuuds128, "V4iV4iV4iV4i", "nV:128:", "avxvnniint16")
+TARGET_BUILTIN(__builtin_ia32_vpdpwuuds256, "V8iV8iV8iV8i", "nV:256:", "avxvnniint16")
+
 // AVX-NE-CONVERT
 TARGET_BUILTIN(__builtin_ia32_vbcstnebf162ps128, "V4fyC*", "nV:128:", "avxneconvert")
 TARGET_BUILTIN(__builtin_ia32_vbcstnebf162ps256, "V8fyC*", "nV:256:", "avxneconvert")

clang/include/clang/Driver/Options.td

@@ -4960,6 +4960,8 @@ def mavxifma : Flag<["-"], "mavxifma">, Group<m_x86_Features_Group>;
 def mno_avxifma : Flag<["-"], "mno-avxifma">, Group<m_x86_Features_Group>;
 def mavxneconvert : Flag<["-"], "mavxneconvert">, Group<m_x86_Features_Group>;
 def mno_avxneconvert : Flag<["-"], "mno-avxneconvert">, Group<m_x86_Features_Group>;
+def mavxvnniint16 : Flag<["-"], "mavxvnniint16">, Group<m_x86_Features_Group>;
+def mno_avxvnniint16 : Flag<["-"], "mno-avxvnniint16">, 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>;

clang/lib/Basic/Targets/X86.cpp

@@ -353,6 +353,8 @@ bool X86TargetInfo::handleTargetFeatures(std::vector<std::string> &Features,
       HasAVXNECONVERT= true;
     } else if (Feature == "+avxvnni") {
       HasAVXVNNI = true;
+    } else if (Feature == "+avxvnniint16") {
+      HasAVXVNNIINT16 = true;
     } else if (Feature == "+avxvnniint8") {
       HasAVXVNNIINT8 = true;
     } else if (Feature == "+serialize") {
@@ -836,6 +838,8 @@ void X86TargetInfo::getTargetDefines(const LangOptions &Opts,
     Builder.defineMacro("__AVXNECONVERT__");
   if (HasAVXVNNI)
     Builder.defineMacro("__AVXVNNI__");
+  if (HasAVXVNNIINT16)
+    Builder.defineMacro("__AVXVNNIINT16__");
   if (HasAVXVNNIINT8)
     Builder.defineMacro("__AVXVNNIINT8__");
   if (HasSERIALIZE)
@@ -964,6 +968,7 @@ bool X86TargetInfo::isValidFeatureName(StringRef Name) const {
       .Case("avxifma", true)
       .Case("avxneconvert", true)
       .Case("avxvnni", true)
+      .Case("avxvnniint16", true)
       .Case("avxvnniint8", true)
       .Case("bmi", true)
       .Case("bmi2", true)
@@ -1069,6 +1074,7 @@ bool X86TargetInfo::hasFeature(StringRef Feature) const {
       .Case("avxifma", HasAVXIFMA)
       .Case("avxneconvert", HasAVXNECONVERT)
       .Case("avxvnni", HasAVXVNNI)
+      .Case("avxvnniint16", HasAVXVNNIINT16)
       .Case("avxvnniint8", HasAVXVNNIINT8)
       .Case("bmi", HasBMI)
       .Case("bmi2", HasBMI2)

clang/lib/Basic/Targets/X86.h

@@ -145,6 +145,7 @@ class LLVM_LIBRARY_VISIBILITY X86TargetInfo : public TargetInfo {
   bool HasPTWRITE = false;
   bool HasINVPCID = false;
   bool HasENQCMD = false;
+  bool HasAVXVNNIINT16 = false;
   bool HasAMXFP16 = false;
   bool HasCMPCCXADD = false;
   bool HasRAOINT = false;

clang/lib/Headers/CMakeLists.txt

@@ -160,6 +160,7 @@ set(x86_files
   avxifmaintrin.h
   avxintrin.h
   avxneconvertintrin.h
+  avxvnniint16intrin.h
   avxvnniint8intrin.h
   avxvnniintrin.h
   bmi2intrin.h