[X86][MC] Support encoding/decoding for APX variant ADD/SUB/ADC/SBB/OR/XOR/NEG/NOT instructions

llvm/include/llvm/Support/X86DisassemblerDecoderCommon.h

@@ -70,7 +70,8 @@ enum attributeBits {
   ATTR_EVEXKZ = 0x1 << 11,
   ATTR_EVEXB = 0x1 << 12,
   ATTR_REX2 = 0x1 << 13,
-  ATTR_max = 0x1 << 14,
+  ATTR_EVEXNF = 0x1 << 14,
+  ATTR_max = 0x1 << 15,
 };
 
 // Combinations of the above attributes that are relevant to instruction
@@ -137,12 +138,15 @@ enum attributeBits {
   ENUM_ENTRY(IC_VEX_L_W_XD, 5, "requires VEX, L, W and XD prefix")             \
   ENUM_ENTRY(IC_VEX_L_W_OPSIZE, 5, "requires VEX, L, W and OpSize")            \
   ENUM_ENTRY(IC_EVEX, 1, "requires an EVEX prefix")                            \
+  ENUM_ENTRY(IC_EVEX_NF, 2, "requires EVEX and NF prefix")                     \
   ENUM_ENTRY(IC_EVEX_XS, 2, "requires EVEX and the XS prefix")                 \
   ENUM_ENTRY(IC_EVEX_XD, 2, "requires EVEX and the XD prefix")                 \
   ENUM_ENTRY(IC_EVEX_OPSIZE, 2, "requires EVEX and the OpSize prefix")         \
+  ENUM_ENTRY(IC_EVEX_OPSIZE_NF, 3, "requires EVEX, NF and the OpSize prefix")  \
   ENUM_ENTRY(IC_EVEX_OPSIZE_ADSIZE, 3,                                         \
              "requires EVEX, OPSIZE and the ADSIZE prefix")                    \
   ENUM_ENTRY(IC_EVEX_W, 3, "requires EVEX and the W prefix")                   \
+  ENUM_ENTRY(IC_EVEX_W_NF, 4, "requires EVEX, W and NF prefix")                \
   ENUM_ENTRY(IC_EVEX_W_XS, 4, "requires EVEX, W, and XS prefix")               \
   ENUM_ENTRY(IC_EVEX_W_XD, 4, "requires EVEX, W, and XD prefix")               \
   ENUM_ENTRY(IC_EVEX_W_OPSIZE, 4, "requires EVEX, W, and OpSize")              \
@@ -187,10 +191,13 @@ enum attributeBits {
   ENUM_ENTRY(IC_EVEX_L2_W_XD_K, 4, "requires EVEX_K, L2, W and XD prefix")     \
   ENUM_ENTRY(IC_EVEX_L2_W_OPSIZE_K, 4, "requires EVEX_K, L2, W and OpSize")    \
   ENUM_ENTRY(IC_EVEX_B, 1, "requires an EVEX_B prefix")                        \
+  ENUM_ENTRY(IC_EVEX_B_NF, 2, "requires EVEX_NF and EVEX_B prefix")            \
   ENUM_ENTRY(IC_EVEX_XS_B, 2, "requires EVEX_B and the XS prefix")             \
   ENUM_ENTRY(IC_EVEX_XD_B, 2, "requires EVEX_B and the XD prefix")             \
   ENUM_ENTRY(IC_EVEX_OPSIZE_B, 2, "requires EVEX_B and the OpSize prefix")     \
+  ENUM_ENTRY(IC_EVEX_OPSIZE_B_NF, 3, "requires EVEX_B, NF and Opsize prefix")  \
   ENUM_ENTRY(IC_EVEX_W_B, 3, "requires EVEX_B and the W prefix")               \
+  ENUM_ENTRY(IC_EVEX_W_B_NF, 4, "requires EVEX_NF, EVEX_B and the W prefix")   \
   ENUM_ENTRY(IC_EVEX_W_XS_B, 4, "requires EVEX_B, W, and XS prefix")           \
   ENUM_ENTRY(IC_EVEX_W_XD_B, 4, "requires EVEX_B, W, and XD prefix")           \
   ENUM_ENTRY(IC_EVEX_W_OPSIZE_B, 4, "requires EVEX_B, W, and OpSize")          \

llvm/lib/Target/X86/AsmParser/X86AsmParser.cpp

@@ -108,6 +108,8 @@ class X86AsmParser : public MCTargetAsmParser {
 
   // Does this instruction use apx extended register?
   bool UseApxExtendedReg = false;
+  // Is this instruction explicitly required not to update flags?
+  bool ForcedNoFlag = false;
 
 private:
   SMLoc consumeToken() {
@@ -3125,6 +3127,7 @@ bool X86AsmParser::ParseInstruction(ParseInstructionInfo &Info, StringRef Name,
   ForcedVEXEncoding = VEXEncoding_Default;
   ForcedDispEncoding = DispEncoding_Default;
   UseApxExtendedReg = false;
+  ForcedNoFlag = false;
 
   // Parse pseudo prefixes.
   while (true) {
@@ -3149,6 +3152,8 @@ bool X86AsmParser::ParseInstruction(ParseInstructionInfo &Info, StringRef Name,
         ForcedDispEncoding = DispEncoding_Disp8;
       else if (Prefix == "disp32")
         ForcedDispEncoding = DispEncoding_Disp32;
+      else if (Prefix == "nf")
+        ForcedNoFlag = true;
       else
         return Error(NameLoc, "unknown prefix");
 
@@ -3996,6 +4001,8 @@ unsigned X86AsmParser::checkTargetMatchPredicate(MCInst &Inst) {
 
   if (UseApxExtendedReg && !X86II::canUseApxExtendedReg(MCID))
     return Match_Unsupported;
+  if (ForcedNoFlag != !!(MCID.TSFlags & X86II::EVEX_NF))
+    return Match_Unsupported;
 
   if (ForcedVEXEncoding == VEXEncoding_EVEX &&
       (MCID.TSFlags & X86II::EncodingMask) != X86II::EVEX)

llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp

@@ -1169,7 +1169,11 @@ static int getInstructionID(struct InternalInstruction *insn,
         attrMask |= ATTR_EVEXKZ;
       if (bFromEVEX4of4(insn->vectorExtensionPrefix[3]))
         attrMask |= ATTR_EVEXB;
-      if (aaaFromEVEX4of4(insn->vectorExtensionPrefix[3]))
+      // nf bit is the MSB of aaa
+      if (nfFromEVEX4of4(insn->vectorExtensionPrefix[3]) &&
+          insn->opcodeType == MAP4)
+        attrMask |= ATTR_EVEXNF;
+      else if (aaaFromEVEX4of4(insn->vectorExtensionPrefix[3]))
         attrMask |= ATTR_EVEXK;
       if (lFromEVEX4of4(insn->vectorExtensionPrefix[3]))
         attrMask |= ATTR_VEXL;

llvm/lib/Target/X86/Disassembler/X86DisassemblerDecoder.h

@@ -103,6 +103,7 @@ namespace X86Disassembler {
 #define bFromEVEX4of4(evex) bitFromOffset4(evex)
 #define v2FromEVEX4of4(evex) invertedBitFromOffset3(evex)
 #define aaaFromEVEX4of4(evex) threeBitsFromOffset0(evex)
+#define nfFromEVEX4of4(evex) bitFromOffset2(evex)
 
 // These enums represent Intel registers for use by the decoder.
 #define REGS_8BIT                                                              \

llvm/lib/Target/X86/MCTargetDesc/X86BaseInfo.h

@@ -870,7 +870,10 @@ enum : uint64_t {
   ExplicitVEXPrefix = 2ULL << ExplicitOpPrefixShift,
   /// For instructions that are promoted to EVEX space for EGPR.
   ExplicitEVEXPrefix = 3ULL << ExplicitOpPrefixShift,
-  ExplicitOpPrefixMask = 3ULL << ExplicitOpPrefixShift
+  ExplicitOpPrefixMask = 3ULL << ExplicitOpPrefixShift,
+  /// EVEX_NF - Set if this instruction has EVEX.NF field set.
+  EVEX_NFShift = ExplicitOpPrefixShift + 2,
+  EVEX_NF = 1ULL << EVEX_NFShift
 };
 
 /// \returns true if the instruction with given opcode is a prefix.
@@ -992,6 +995,12 @@ inline unsigned getOperandBias(const MCInstrDesc &Desc) {
   }
 }
 
+/// \returns true if the instruction has a NDD (new data destination).
+inline bool hasNewDataDest(uint64_t TSFlags) {
+  return (TSFlags & X86II::OpMapMask) == X86II::T_MAP4 &&
+         (TSFlags & X86II::EVEX_B) && (TSFlags & X86II::VEX_4V);
+}
+
 /// \returns operand # for the first field of the memory operand or -1 if no
 /// memory operands.
 /// NOTE: This ignores tied operands.  If there is a tied register which is
@@ -1018,7 +1027,7 @@ inline int getMemoryOperandNo(uint64_t TSFlags) {
     return -1;
   case X86II::MRMDestMem:
   case X86II::MRMDestMemFSIB:
-    return 0;
+    return hasNewDataDest(TSFlags);
   case X86II::MRMSrcMem:
   case X86II::MRMSrcMemFSIB:
     // Start from 1, skip any registers encoded in VEX_VVVV or I8IMM, or a

llvm/lib/Target/X86/MCTargetDesc/X86InstPrinterCommon.cpp

@@ -369,6 +369,9 @@ void X86InstPrinterCommon::printInstFlags(const MCInst *MI, raw_ostream &O,
   else if (Flags & X86::IP_HAS_REPEAT)
     O << "\trep\t";
 
+  if (TSFlags & X86II::EVEX_NF)
+    O << "\t{nf}";
+
   // These all require a pseudo prefix
   if ((Flags & X86::IP_USE_VEX) ||
       (TSFlags & X86II::ExplicitOpPrefixMask) == X86II::ExplicitVEXPrefix)

llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp

@@ -251,6 +251,7 @@ class X86OpcodePrefixHelper {
   void setAAA(const MCInst &MI, unsigned OpNum) {
     EVEX_aaa = getRegEncoding(MI, OpNum);
   }
+  void setNF(bool V) { EVEX_aaa |= V << 2; }
 
   X86OpcodePrefixHelper(const MCRegisterInfo &MRI)
       : W(0), R(0), X(0), B(0), M(0), R2(0), X2(0), B2(0), VEX_4V(0), VEX_L(0),
@@ -987,9 +988,11 @@ X86MCCodeEmitter::emitVEXOpcodePrefix(int MemOperand, const MCInst &MI,
   }
 
   Prefix.setW(TSFlags & X86II::REX_W);
+  Prefix.setNF(TSFlags & X86II::EVEX_NF);
 
   bool HasEVEX_K = TSFlags & X86II::EVEX_K;
   bool HasVEX_4V = TSFlags & X86II::VEX_4V;
+  bool IsND = X86II::hasNewDataDest(TSFlags); // IsND implies HasVEX_4V
   bool HasEVEX_RC = TSFlags & X86II::EVEX_RC;
 
   switch (TSFlags & X86II::OpMapMask) {
@@ -1049,6 +1052,7 @@ X86MCCodeEmitter::emitVEXOpcodePrefix(int MemOperand, const MCInst &MI,
 
   bool EncodeRC = false;
   uint8_t EVEX_rc = 0;
+
   unsigned CurOp = X86II::getOperandBias(Desc);
 
   switch (TSFlags & X86II::FormMask) {
@@ -1073,16 +1077,21 @@ X86MCCodeEmitter::emitVEXOpcodePrefix(int MemOperand, const MCInst &MI,
     //  MemAddr, src1(VEX_4V), src2(ModR/M)
     //  MemAddr, src1(ModR/M), imm8
     //
+    // NDD:
+    //  dst(VEX_4V), MemAddr, src1(ModR/M)
     Prefix.setBB2(MI, MemOperand + X86::AddrBaseReg);
     Prefix.setXX2(MI, MemOperand + X86::AddrIndexReg);
     Prefix.setV2(MI, MemOperand + X86::AddrIndexReg, HasVEX_4V);
 
+    if (IsND)
+      Prefix.set4VV2(MI, CurOp++);
+
     CurOp += X86::AddrNumOperands;
 
     if (HasEVEX_K)
       Prefix.setAAA(MI, CurOp++);
 
-    if (HasVEX_4V)
+    if (!IsND && HasVEX_4V)
       Prefix.set4VV2(MI, CurOp++);
 
     Prefix.setRR2(MI, CurOp++);
@@ -1098,12 +1107,18 @@ X86MCCodeEmitter::emitVEXOpcodePrefix(int MemOperand, const MCInst &MI,
     //
     //  FMA4:
     //  dst(ModR/M.reg), src1(VEX_4V), src2(ModR/M), src3(Imm[7:4])
+    //
+    //  NDD:
+    //  dst(VEX_4V), src1(ModR/M), MemAddr
+    if (IsND)
+      Prefix.set4VV2(MI, CurOp++);
+
     Prefix.setRR2(MI, CurOp++);
 
     if (HasEVEX_K)
       Prefix.setAAA(MI, CurOp++);
 
-    if (HasVEX_4V)
+    if (!IsND && HasVEX_4V)
       Prefix.set4VV2(MI, CurOp++);
 
     Prefix.setBB2(MI, MemOperand + X86::AddrBaseReg);
@@ -1160,12 +1175,17 @@ X86MCCodeEmitter::emitVEXOpcodePrefix(int MemOperand, const MCInst &MI,
     //
     //  FMA4:
     //  dst(ModR/M.reg), src1(VEX_4V), src2(Imm[7:4]), src3(ModR/M),
+    //
+    //  NDD:
+    //  dst(VEX_4V), src1(ModR/M.reg), src2(ModR/M)
+    if (IsND)
+      Prefix.set4VV2(MI, CurOp++);
     Prefix.setRR2(MI, CurOp++);
 
     if (HasEVEX_K)
       Prefix.setAAA(MI, CurOp++);
 
-    if (HasVEX_4V)
+    if (!IsND && HasVEX_4V)
       Prefix.set4VV2(MI, CurOp++);
 
     Prefix.setBB2(MI, CurOp);
@@ -1209,14 +1229,19 @@ X86MCCodeEmitter::emitVEXOpcodePrefix(int MemOperand, const MCInst &MI,
     //  dst(ModR/M), src(ModR/M)
     //  dst(ModR/M), src(ModR/M), imm8
     //  dst(ModR/M), src1(VEX_4V), src2(ModR/M)
+    //
+    // NDD:
+    // dst(VEX_4V), src1(ModR/M), src2(ModR/M)
+    if (IsND)
+      Prefix.set4VV2(MI, CurOp++);
     Prefix.setBB2(MI, CurOp);
     Prefix.setX(MI, CurOp, 4);
     ++CurOp;
 
     if (HasEVEX_K)
       Prefix.setAAA(MI, CurOp++);
 
-    if (HasVEX_4V)
+    if (!IsND && HasVEX_4V)
       Prefix.set4VV2(MI, CurOp++);
 
     Prefix.setRR2(MI, CurOp++);
@@ -1508,6 +1533,8 @@ void X86MCCodeEmitter::encodeInstruction(const MCInst &MI,
 
   unsigned OpcodeOffset = 0;
 
+  bool IsND = X86II::hasNewDataDest(TSFlags);
+
   uint64_t Form = TSFlags & X86II::FormMask;
   switch (Form) {
   default:
@@ -1576,6 +1603,8 @@ void X86MCCodeEmitter::encodeInstruction(const MCInst &MI,
 
     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
       ++SrcRegNum;
+    if (IsND) // Skip the NDD operand encoded in EVEX_VVVV
+      ++CurOp;
 
     emitRegModRMByte(MI.getOperand(CurOp),
                      getX86RegNum(MI.getOperand(SrcRegNum)), CB);
@@ -1602,6 +1631,9 @@ void X86MCCodeEmitter::encodeInstruction(const MCInst &MI,
     if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV)
       ++SrcRegNum;
 
+    if (IsND) // Skip new data destination
+      ++CurOp;
+
     bool ForceSIB = (Form == X86II::MRMDestMemFSIB);
     emitMemModRMByte(MI, CurOp, getX86RegNum(MI.getOperand(SrcRegNum)), TSFlags,
                      Kind, StartByte, CB, Fixups, STI, ForceSIB);
@@ -1669,6 +1701,9 @@ void X86MCCodeEmitter::encodeInstruction(const MCInst &MI,
   case X86II::MRMSrcMem: {
     unsigned FirstMemOp = CurOp + 1;
 
+    if (IsND) // Skip new data destination
+      CurOp++;
+
     if (HasEVEX_K) // Skip writemask
       ++FirstMemOp;