[X86][MC,Driver] Support -msse2avx to encode SSE instruction with VEX prefix

clang/include/clang/Basic/CodeGenOptions.def

@@ -38,6 +38,7 @@ VALUE_CODEGENOPT(Name, Bits, Default)
 CODEGENOPT(DisableIntegratedAS, 1, 0) ///< -no-integrated-as
 CODEGENOPT(Crel, 1, 0) ///< -Wa,--crel
 CODEGENOPT(RelaxELFRelocations, 1, 1) ///< -Wa,-mrelax-relocations={yes,no}
+CODEGENOPT(SSE2AVX          , 1, 0)   ///< -msse2avx
 CODEGENOPT(AsmVerbose        , 1, 0) ///< -dA, -fverbose-asm.
 CODEGENOPT(PreserveAsmComments, 1, 1) ///< -dA, -fno-preserve-as-comments.
 CODEGENOPT(AssumeSaneOperatorNew , 1, 1) ///< implicit __attribute__((malloc)) operator new

clang/include/clang/Driver/Options.td

@@ -5179,6 +5179,13 @@ def mvx : Flag<["-"], "mvx">, Group<m_Group>;
 def mno_vx : Flag<["-"], "mno-vx">, Group<m_Group>;
 } // let Flags = [TargetSpecific]
 
+let Flags = [TargetSpecific] in {
+def msse2avx : Flag<["-"], "msse2avx">, Group<m_Group>,
+  Visibility<[ClangOption, CC1Option, CC1AsOption]>,
+  HelpText<"Specify that the assembler should encode SSE instructions with VEX prefix">,
+  MarshallingInfoFlag<CodeGenOpts<"SSE2AVX">>;
+} // let Flags = [TargetSpecific]
+
 defm zvector : BoolFOption<"zvector",
   LangOpts<"ZVector">, DefaultFalse,
   PosFlag<SetTrue, [], [ClangOption, CC1Option],

clang/lib/Driver/ToolChains/Clang.cpp

@@ -2545,6 +2545,13 @@ static void CollectArgsForIntegratedAssembler(Compilation &C,
       switch (C.getDefaultToolChain().getArch()) {
       default:
         break;
+      case llvm::Triple::x86:
+      case llvm::Triple::x86_64:
+        if (Value == "-msse2avx") {
+          CmdArgs.push_back("-msse2avx");
+          continue;
+        }
+        break;
       case llvm::Triple::wasm32:
       case llvm::Triple::wasm64:
         if (Value == "--no-type-check") {

clang/test/Driver/msse2avx.c

@@ -0,0 +1,7 @@
+// RUN: %clang -### -c -march=x86-64 -Xassembler -msse2avx %s 2>&1 | FileCheck  %s
+// RUN: %clang -### -c -march=x86-64 -x assembler -Xassembler -msse2avx %s 2>&1 | FileCheck %s
+
+// CHECK: "-msse2avx"
+
+// RUN: not %clang -### -c --target=aarch64 -march=armv8a -msse2avx %s 2>&1 | FileCheck --check-prefix=ERR %s
+// ERR:   error: unsupported option '-msse2avx' for target 'aarch64'

clang/tools/driver/cc1as_main.cpp

@@ -98,6 +98,8 @@ struct AssemblerInvocation {
   LLVM_PREFERRED_TYPE(bool)
   unsigned RelaxELFRelocations : 1;
   LLVM_PREFERRED_TYPE(bool)
+  unsigned SSE2AVX : 1;
+  LLVM_PREFERRED_TYPE(bool)
   unsigned Dwarf64 : 1;
   unsigned DwarfVersion;
   std::string DwarfDebugFlags;
@@ -197,6 +199,7 @@ struct AssemblerInvocation {
     ShowInst = 0;
     ShowEncoding = 0;
     RelaxAll = 0;
+    SSE2AVX = 0;
     NoExecStack = 0;
     FatalWarnings = 0;
     NoWarn = 0;
@@ -288,6 +291,7 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
   }
 
   Opts.RelaxELFRelocations = !Args.hasArg(OPT_mrelax_relocations_no);
+  Opts.SSE2AVX = Args.hasArg(OPT_msse2avx);
   if (auto *DwarfFormatArg = Args.getLastArg(OPT_gdwarf64, OPT_gdwarf32))
     Opts.Dwarf64 = DwarfFormatArg->getOption().matches(OPT_gdwarf64);
   Opts.DwarfVersion = getLastArgIntValue(Args, OPT_dwarf_version_EQ, 2, Diags);
@@ -437,6 +441,7 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
   MCOptions.MCSaveTempLabels = Opts.SaveTemporaryLabels;
   MCOptions.Crel = Opts.Crel;
   MCOptions.X86RelaxRelocations = Opts.RelaxELFRelocations;
+  MCOptions.X86Sse2Avx = Opts.SSE2AVX;
   MCOptions.CompressDebugSections = Opts.CompressDebugSections;
   MCOptions.AsSecureLogFile = Opts.AsSecureLogFile;
 

llvm/include/llvm/MC/MCTargetOptions.h

@@ -68,6 +68,8 @@ class MCTargetOptions {
   // ELF.
   bool X86RelaxRelocations = true;
 
+  bool X86Sse2Avx = false;
+
   EmitDwarfUnwindType EmitDwarfUnwind;
 
   int DwarfVersion = 0;

llvm/include/llvm/MC/MCTargetOptionsCommandFlags.h

@@ -55,6 +55,8 @@ bool getCrel();
 
 bool getX86RelaxRelocations();
 
+bool getX86Sse2Avx();
+
 std::string getABIName();
 
 std::string getAsSecureLogFile();

llvm/lib/MC/MCTargetOptionsCommandFlags.cpp

@@ -49,6 +49,7 @@ MCOPT(bool, NoTypeCheck)
 MCOPT(bool, SaveTempLabels)
 MCOPT(bool, Crel)
 MCOPT(bool, X86RelaxRelocations)
+MCOPT(bool, X86Sse2Avx)
 MCOPT(std::string, ABIName)
 MCOPT(std::string, AsSecureLogFile)
 
@@ -140,6 +141,11 @@ llvm::mc::RegisterMCTargetOptionsFlags::RegisterMCTargetOptionsFlags() {
       cl::init(true));
   MCBINDOPT(X86RelaxRelocations);
 
+  static cl::opt<bool> X86Sse2Avx(
+      "x86-sse2avx", cl::desc("Specify that the assembler should encode SSE "
+                              "instructions with VEX prefix"));
+  MCBINDOPT(X86Sse2Avx);
+
   static cl::opt<std::string> ABIName(
       "target-abi", cl::Hidden,
       cl::desc("The name of the ABI to be targeted from the backend."),
@@ -169,6 +175,7 @@ MCTargetOptions llvm::mc::InitMCTargetOptionsFromFlags() {
   Options.MCSaveTempLabels = getSaveTempLabels();
   Options.Crel = getCrel();
   Options.X86RelaxRelocations = getX86RelaxRelocations();
+  Options.X86Sse2Avx = getX86Sse2Avx();
   Options.EmitDwarfUnwind = getEmitDwarfUnwind();
   Options.EmitCompactUnwindNonCanonical = getEmitCompactUnwindNonCanonical();
   Options.AsSecureLogFile = getAsSecureLogFile();

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

@@ -58,6 +58,10 @@ static bool checkScale(unsigned Scale, StringRef &ErrMsg) {
 
 namespace {
 
+// Including the generated SSE2AVX compression tables.
+#define GET_X86_SSE2AVX_TABLE
+#include "X86GenInstrMapping.inc"
+
 static const char OpPrecedence[] = {
     0,  // IC_OR
     1,  // IC_XOR
@@ -3745,7 +3749,27 @@ bool X86AsmParser::ParseInstruction(ParseInstructionInfo &Info, StringRef Name,
   return false;
 }
 
+static bool convertSSEToAVX(MCInst &Inst) {
+  ArrayRef<X86TableEntry> Table{X86SSE2AVXTable};
+  unsigned Opcode = Inst.getOpcode();
+  const auto I = llvm::lower_bound(Table, Opcode);
+  if (I == Table.end() || I->OldOpc != Opcode)
+    return false;
+
+  Inst.setOpcode(I->NewOpc);
+  // AVX variant of BLENDVPD/BLENDVPS/PBLENDVB instructions has more
+  // operand compare to SSE variant, which is added below
+  if (X86::isBLENDVPD(Opcode) || X86::isBLENDVPS(Opcode) ||
+      X86::isPBLENDVB(Opcode))
+    Inst.addOperand(Inst.getOperand(2));
+
+  return true;
+}
+
 bool X86AsmParser::processInstruction(MCInst &Inst, const OperandVector &Ops) {
+  if (MCOptions.X86Sse2Avx && convertSSEToAVX(Inst))
+    return true;
+
   if (ForcedOpcodePrefix != OpcodePrefix_VEX3 &&
       X86::optimizeInstFromVEX3ToVEX2(Inst, MII.get(Inst.getOpcode())))
     return true;

llvm/test/MC/AsmParser/sse2avx-att.s

@@ -0,0 +1,89 @@
+# RUN: llvm-mc -triple x86_64 -x86-sse2avx %s | FileCheck %s
+# RUN: llvm-mc -triple=x86_64 -output-asm-variant=1 %s | llvm-mc -triple=x86_64 -x86-asm-syntax=intel -x86-sse2avx
+	.text
+# CHECK: vmovsd  -352(%rbp), %xmm0
+	movsd	-352(%rbp), %xmm0               # xmm0 = mem[0],zero
+# CHECK-NEXT: vunpcklpd       %xmm1, %xmm0, %xmm0     # xmm0 = xmm0[0],xmm1[0]
+	unpcklpd	%xmm1, %xmm0                    # xmm0 = xmm0[0],xmm1[0]
+# CHECK-NEXT: vmovapd %xmm0, -368(%rbp)
+	movapd	%xmm0, -368(%rbp)
+# CHECK-NEXT: vmovapd -368(%rbp), %xmm0
+	movapd	-368(%rbp), %xmm0
+# CHECK-NEXT: vmovsd  -376(%rbp), %xmm1
+	movsd	-376(%rbp), %xmm1               # xmm1 = mem[0],zero
+# CHECK-NEXT: vmovsd  -384(%rbp), %xmm0
+	movsd	-384(%rbp), %xmm0               # xmm0 = mem[0],zero
+# CHECK-NEXT: vunpcklpd       %xmm1, %xmm0, %xmm0     # xmm0 = xmm0[0],xmm1[0]
+	unpcklpd	%xmm1, %xmm0                    # xmm0 = xmm0[0],xmm1[0]
+# CHECK-NEXT: vaddpd  %xmm1, %xmm0, %xmm0
+	addpd	%xmm1, %xmm0
+# CHECK-NEXT: vmovapd %xmm0, -464(%rbp)
+	movapd	%xmm0, -464(%rbp)
+# CHECK-NEXT: vmovaps -304(%rbp), %xmm1
+	movaps	-304(%rbp), %xmm1
+# CHECK-NEXT: vpandn  %xmm1, %xmm0, %xmm0
+	pandn	%xmm1, %xmm0
+# CHECK-NEXT: vmovaps %xmm0, -480(%rbp)
+	movaps	%xmm0, -480(%rbp)
+# CHECK-NEXT: vmovss  -220(%rbp), %xmm1
+	movss	-220(%rbp), %xmm1               # xmm1 = mem[0],zero,zero,zero
+# CHECK-NEXT: vinsertps       $16, %xmm1, %xmm0, %xmm0 # xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+	insertps	$16, %xmm1, %xmm0               # xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+# CHECK-NEXT: vmovaps %xmm0, -496(%rbp)
+	movaps	%xmm0, -496(%rbp)
+# CHECK-NEXT: vmovss  -256(%rbp), %xmm0
+	movss	-256(%rbp), %xmm0               # xmm0 = mem[0],zero,zero,zero
+# CHECK-NEXT: vmovaps -192(%rbp), %xmm0
+	movaps	-192(%rbp), %xmm0
+# CHECK-NEXT: vdivss  %xmm1, %xmm0, %xmm0
+	divss	%xmm1, %xmm0
+# CHECK-NEXT: vmovaps %xmm0, -192(%rbp)
+	movaps	%xmm0, -192(%rbp)
+# CHECK-NEXT: vmovd   -128(%rbp), %xmm0
+	movd	-128(%rbp), %xmm0               # xmm0 = mem[0],zero,zero,zero
+# CHECK-NEXT: vpinsrd $1, %edx, %xmm0, %xmm0
+	pinsrd	$1, %edx, %xmm0
+# CHECK-NEXT: vmovaps %xmm0, -144(%rbp)
+	movaps	%xmm0, -144(%rbp)
+# CHECK-NEXT: vmovd   -160(%rbp), %xmm0
+	movd	-160(%rbp), %xmm0               # xmm0 = mem[0],zero,zero,zero
+# CHECK-NEXT: vpblendw        $170, %xmm1, %xmm0, %xmm0       # xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
+	pblendw	$170, %xmm1, %xmm0              # xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
+# CHECK-NEXT: vmovdqa %xmm0, -576(%rbp)
+	movdqa	%xmm0, -576(%rbp)
+# CHECK-NEXT: vphsubw %xmm1, %xmm0, %xmm0
+	phsubw	%xmm1, %xmm0
+# CHECK-NEXT: vmovdqa %xmm0, -592(%rbp)
+	movdqa	%xmm0, -592(%rbp)
+# CHECK-NEXT: vmovaps -496(%rbp), %xmm0
+	movaps	-496(%rbp), %xmm0
+# CHECK-NEXT: vroundps        $8, %xmm0, %xmm0
+	roundps	$8, %xmm0, %xmm0
+# CHECK-NEXT: vmovaps %xmm0, -608(%rbp)
+	movaps	%xmm0, -608(%rbp)
+# CHECK-NEXT: vmovapd -432(%rbp), %xmm0
+	movapd	-432(%rbp), %xmm0
+# CHECK-NEXT: vpxor   %xmm1, %xmm0, %xmm0
+	pxor	%xmm1, %xmm0
+# CHECK-NEXT: vmovaps %xmm0, -640(%rbp)
+	movaps	%xmm0, -640(%rbp)
+# CHECK-NEXT: vmovapd -32(%rbp), %xmm0
+	movapd	-32(%rbp), %xmm0
+# CHECK-NEXT: vmovupd %xmm0, (%rax)
+	movupd	%xmm0, (%rax)
+# CHECK-NEXT: vmovsd  -656(%rbp), %xmm0
+	movsd	-656(%rbp), %xmm0               # xmm0 = mem[0],zero
+# CHECK-NEXT: extrq   $16, $8, %xmm0                  # xmm0 = xmm0[2],zero,zero,zero,zero,zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
+        extrq   $16, $8, %xmm0
+# CHECK-NEXT: insertq $16, $8, %xmm1, %xmm0           # xmm0 = xmm0[0,1],xmm1[0],xmm0[3,4,5,6,7,u,u,u,u,u,u,u,u]
+	insertq $16, $8, %xmm1, %xmm0
+# CHECK-NEXT: pshufw  $1, %mm0, %mm2                  # mm2 = mm0[1,0,0,0]
+	pshufw  $1, %mm0, %mm2
+# CHECK-NEXT: vpblendvb       %xmm2, %xmm2, %xmm1, %xmm1
+	pblendvb   %xmm0, %xmm2, %xmm1
+# CHECK-NEXT: vblendvps       %xmm0, %xmm0, %xmm2, %xmm2
+	blendvps   %xmm0, %xmm0, %xmm2
+# CHECK-NEXT: vblendvpd       %xmm0, %xmm0, %xmm2, %xmm2
+	blendvpd   %xmm0, %xmm0, %xmm2
+# CHECK-NEXT: vblendvpd       %xmm0, %xmm0, %xmm2, %xmm2
+	blendvpd   %xmm0, %xmm2

llvm/utils/TableGen/X86InstrMappingEmitter.cpp

@@ -56,6 +56,8 @@ class X86InstrMappingEmitter {
                             raw_ostream &OS);
   void emitND2NonNDTable(ArrayRef<const CodeGenInstruction *> Insts,
                          raw_ostream &OS);
+  void emitSSE2AVXTable(ArrayRef<const CodeGenInstruction *> Insts,
+                        raw_ostream &OS);
 
   // Prints the definition of class X86TableEntry.
   void printClassDef(raw_ostream &OS);
@@ -335,6 +337,38 @@ void X86InstrMappingEmitter::emitND2NonNDTable(
   printTable(Table, "X86ND2NonNDTable", "GET_X86_ND2NONND_TABLE", OS);
 }
 
+void X86InstrMappingEmitter::emitSSE2AVXTable(
+    ArrayRef<const CodeGenInstruction *> Insts, raw_ostream &OS) {
+
+  const std::map<StringRef, StringRef> ManualMap = {
+#define ENTRY_SSE2AVX(OLD, NEW) {#OLD, #NEW},
+#include "X86ManualInstrMapping.def"
+  };
+
+  std::vector<Entry> Table;
+  for (const CodeGenInstruction *Inst : Insts) {
+    const Record *Rec = Inst->TheDef;
+    StringRef Name = Rec->getName();
+    if (!isInteresting(Rec))
+      continue;
+    if (ManualMap.find(Name) != ManualMap.end()) {
+      auto *NewRec = Records.getDef(ManualMap.at(Rec->getName()));
+      assert(NewRec && "Instruction not found!");
+      auto &NewInst = Target.getInstruction(NewRec);
+      Table.push_back(std::pair(Inst, &NewInst));
+      continue;
+    }
+
+    std::string NewName = ("V" + Name).str();
+    auto *AVXRec = Records.getDef(NewName);
+    if (!AVXRec)
+      continue;
+    auto &AVXInst = Target.getInstruction(AVXRec);
+    Table.push_back(std::pair(Inst, &AVXInst));
+  }
+  printTable(Table, "X86SSE2AVXTable", "GET_X86_SSE2AVX_TABLE", OS);
+}
+
 void X86InstrMappingEmitter::run(raw_ostream &OS) {
   emitSourceFileHeader("X86 instruction mapping", OS);
 
@@ -344,6 +378,7 @@ void X86InstrMappingEmitter::run(raw_ostream &OS) {
   emitCompressEVEXTable(Insts, OS);
   emitNFTransformTable(Insts, OS);
   emitND2NonNDTable(Insts, OS);
+  emitSSE2AVXTable(Insts, OS);
 }
 } // namespace
 

llvm/utils/TableGen/X86ManualInstrMapping.def

@@ -349,3 +349,14 @@ NOCOMP_ND(CFCMOV64rr_ND)
 ENTRY_ND(MOVBE32rr, BSWAP32r)
 ENTRY_ND(MOVBE64rr, BSWAP64r)
 #undef ENTRY_ND
+
+#ifndef ENTRY_SSE2AVX
+#define ENTRY_SSE2AVX(OLD, NEW)
+#endif
+ENTRY_SSE2AVX(BLENDVPDrm0, VBLENDVPDrmr)
+ENTRY_SSE2AVX(BLENDVPDrr0, VBLENDVPDrrr)
+ENTRY_SSE2AVX(BLENDVPSrm0, VBLENDVPSrmr)
+ENTRY_SSE2AVX(BLENDVPSrr0, VBLENDVPSrrr)
+ENTRY_SSE2AVX(PBLENDVBrm0, VPBLENDVBrmr)
+ENTRY_SSE2AVX(PBLENDVBrr0, VPBLENDVBrrr)
+#undef ENTRY_SSE2AVX