Add support for the rep movs[bwd] instructions, and emit them when code

generating the llvm.memcpy intrinsic.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@11351 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 138 insertions, 15 deletions

diff --git a/lib/Target/X86/InstSelectSimple.cpp b/lib/Target/X86/InstSelectSimple.cpp
index 9df7697344..67849e2edb 100644
--- a/lib/Target/X86/InstSelectSimple.cpp
+++ b/lib/Target/X86/InstSelectSimple.cpp
@@ -1157,6 +1157,7 @@ void ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) {
           case Intrinsic::va_start:
           case Intrinsic::va_copy:
           case Intrinsic::va_end:
+          case Intrinsic::memcpy:
             // We directly implement these intrinsics
             break;
           default:
@@ -1188,6 +1189,58 @@ void ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) {
     return;
   case Intrinsic::va_end: return;   // Noop on X86
 
+  case Intrinsic::memcpy: {
+    assert(CI.getNumOperands() == 5 && "Illegal llvm.memcpy call!");
+    unsigned Align = 1;
+    if (ConstantInt *AlignC = dyn_cast<ConstantInt>(CI.getOperand(4))) {
+      Align = AlignC->getRawValue();
+      if (Align == 0) Align = 1;
+    }
+
+    // Turn the byte code into # iterations
+    unsigned ByteReg = getReg(CI.getOperand(3));
+    unsigned CountReg;
+    
+    switch (Align & 3) {
+    case 2:   // WORD aligned
+      CountReg = makeAnotherReg(Type::IntTy);
+      BuildMI(BB, X86::SHRir32, 2, CountReg).addReg(ByteReg).addZImm(1);
+      break;
+    case 0:   // DWORD aligned
+      CountReg = makeAnotherReg(Type::IntTy);
+      BuildMI(BB, X86::SHRir32, 2, CountReg).addReg(ByteReg).addZImm(2);
+      break;
+    case 1:   // BYTE aligned
+    case 3:   // BYTE aligned
+      CountReg = ByteReg;
+      break;
+    }
+
+    // No matter what the alignment is, we put the source in ESI, the
+    // destination in EDI, and the count in ECX.
+    TmpReg1 = getReg(CI.getOperand(1));
+    TmpReg2 = getReg(CI.getOperand(2));
+    BuildMI(BB, X86::MOVrr32, 1, X86::ECX).addReg(CountReg);
+    BuildMI(BB, X86::MOVrr32, 1, X86::EDI).addReg(TmpReg1);
+    BuildMI(BB, X86::MOVrr32, 1, X86::ESI).addReg(TmpReg2);
+
+    unsigned Bytes = getReg(CI.getOperand(3));
+    switch (Align & 3) {
+    case 1:   // BYTE aligned
+    case 3:   // BYTE aligned
+      BuildMI(BB, X86::REP_MOVSB, 0);
+      break;
+    case 2:   // WORD aligned
+      BuildMI(BB, X86::REP_MOVSW, 0);
+      break;
+    case 0:   // DWORD aligned
+      BuildMI(BB, X86::REP_MOVSD, 0);
+      break;
+    }
+
+    return;
+  }
+
   default: assert(0 && "Error: unknown intrinsics should have been lowered!");
   }
 }
diff --git a/lib/Target/X86/X86CodeEmitter.cpp b/lib/Target/X86/X86CodeEmitter.cpp
index be6319cd0c..83e5e102d5 100644
--- a/lib/Target/X86/X86CodeEmitter.cpp
+++ b/lib/Target/X86/X86CodeEmitter.cpp
@@ -470,6 +470,9 @@ void Emitter::emitInstruction(MachineInstr &MI) {
   unsigned Opcode = MI.getOpcode();
   const TargetInstrDescriptor &Desc = II->get(Opcode);
 
+  // Emit the repeat opcode prefix as needed.
+  if ((Desc.TSFlags & X86II::Op0Mask) == X86II::REP) MCE.emitByte(0xF3);
+
   // Emit instruction prefixes if necessary
   if (Desc.TSFlags & X86II::OpSize) MCE.emitByte(0x66);// Operand size...
 
@@ -477,6 +480,7 @@ void Emitter::emitInstruction(MachineInstr &MI) {
   case X86II::TB:
     MCE.emitByte(0x0F);   // Two-byte opcode prefix
     break;
+  case X86II::REP: break; // already handled.
   case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
   case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
     MCE.emitByte(0xD8+
diff --git a/lib/Target/X86/X86ISelSimple.cpp b/lib/Target/X86/X86ISelSimple.cpp
index 9df7697344..67849e2edb 100644
--- a/lib/Target/X86/X86ISelSimple.cpp
+++ b/lib/Target/X86/X86ISelSimple.cpp
@@ -1157,6 +1157,7 @@ void ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) {
           case Intrinsic::va_start:
           case Intrinsic::va_copy:
           case Intrinsic::va_end:
+          case Intrinsic::memcpy:
             // We directly implement these intrinsics
             break;
           default:
@@ -1188,6 +1189,58 @@ void ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) {
     return;
   case Intrinsic::va_end: return;   // Noop on X86
 
+  case Intrinsic::memcpy: {
+    assert(CI.getNumOperands() == 5 && "Illegal llvm.memcpy call!");
+    unsigned Align = 1;
+    if (ConstantInt *AlignC = dyn_cast<ConstantInt>(CI.getOperand(4))) {
+      Align = AlignC->getRawValue();
+      if (Align == 0) Align = 1;
+    }
+
+    // Turn the byte code into # iterations
+    unsigned ByteReg = getReg(CI.getOperand(3));
+    unsigned CountReg;
+    
+    switch (Align & 3) {
+    case 2:   // WORD aligned
+      CountReg = makeAnotherReg(Type::IntTy);
+      BuildMI(BB, X86::SHRir32, 2, CountReg).addReg(ByteReg).addZImm(1);
+      break;
+    case 0:   // DWORD aligned
+      CountReg = makeAnotherReg(Type::IntTy);
+      BuildMI(BB, X86::SHRir32, 2, CountReg).addReg(ByteReg).addZImm(2);
+      break;
+    case 1:   // BYTE aligned
+    case 3:   // BYTE aligned
+      CountReg = ByteReg;
+      break;
+    }
+
+    // No matter what the alignment is, we put the source in ESI, the
+    // destination in EDI, and the count in ECX.
+    TmpReg1 = getReg(CI.getOperand(1));
+    TmpReg2 = getReg(CI.getOperand(2));
+    BuildMI(BB, X86::MOVrr32, 1, X86::ECX).addReg(CountReg);
+    BuildMI(BB, X86::MOVrr32, 1, X86::EDI).addReg(TmpReg1);
+    BuildMI(BB, X86::MOVrr32, 1, X86::ESI).addReg(TmpReg2);
+
+    unsigned Bytes = getReg(CI.getOperand(3));
+    switch (Align & 3) {
+    case 1:   // BYTE aligned
+    case 3:   // BYTE aligned
+      BuildMI(BB, X86::REP_MOVSB, 0);
+      break;
+    case 2:   // WORD aligned
+      BuildMI(BB, X86::REP_MOVSW, 0);
+      break;
+    case 0:   // DWORD aligned
+      BuildMI(BB, X86::REP_MOVSD, 0);
+      break;
+    }
+
+    return;
+  }
+
   default: assert(0 && "Error: unknown intrinsics should have been lowered!");
   }
 }
diff --git a/lib/Target/X86/X86InstrInfo.h b/lib/Target/X86/X86InstrInfo.h
index 98f9fe68d7..c6e3b76176 100644
--- a/lib/Target/X86/X86InstrInfo.h
+++ b/lib/Target/X86/X86InstrInfo.h
@@ -86,9 +86,9 @@ namespace X86II {
     OpSize      = 1 << 5,
 
     // Op0Mask - There are several prefix bytes that are used to form two byte
-    // opcodes.  These are currently 0x0F, and 0xD8-0xDF.  This mask is used to
-    // obtain the setting of this field.  If no bits in this field is set, there
-    // is no prefix byte for obtaining a multibyte opcode.
+    // opcodes.  These are currently 0x0F, 0xF3, and 0xD8-0xDF.  This mask is
+    // used to obtain the setting of this field.  If no bits in this field is
+    // set, there is no prefix byte for obtaining a multibyte opcode.
     //
     Op0Shift    = 6,
     Op0Mask     = 0xF << Op0Shift,
@@ -97,12 +97,16 @@ namespace X86II {
     // starts with a 0x0F byte before the real opcode.
     TB          = 1 << Op0Shift,
 
+    // REP - The 0xF3 prefix byte indicating repetition of the following
+    // instruction.
+    REP         = 2 << Op0Shift,
+
     // D8-DF - These escape opcodes are used by the floating point unit.  These
     // values must remain sequential.
-    D8 = 2 << Op0Shift,   D9 = 3 << Op0Shift,
-    DA = 4 << Op0Shift,   DB = 5 << Op0Shift,
-    DC = 6 << Op0Shift,   DD = 7 << Op0Shift,
-    DE = 8 << Op0Shift,   DF = 9 << Op0Shift,
+    D8 = 3 << Op0Shift,   D9 = 4 << Op0Shift,
+    DA = 5 << Op0Shift,   DB = 6 << Op0Shift,
+    DC = 7 << Op0Shift,   DD = 8 << Op0Shift,
+    DE = 9 << Op0Shift,   DF = 10 << Op0Shift,
 
     //===------------------------------------------------------------------===//
     // This three-bit field describes the size of a memory operand.  Zero is
diff --git a/lib/Target/X86/X86InstrInfo.td b/lib/Target/X86/X86InstrInfo.td
index 92f193f985..4bb1a9550a 100644
--- a/lib/Target/X86/X86InstrInfo.td
+++ b/lib/Target/X86/X86InstrInfo.td
@@ -92,14 +92,15 @@ class Pattern<dag P> {
 // emitter that various prefix bytes are required.
 class OpSize { bit hasOpSizePrefix = 1; }
 class TB     { bits<4> Prefix = 1; }
-class D8     { bits<4> Prefix = 2; }
-class D9     { bits<4> Prefix = 3; }
-class DA     { bits<4> Prefix = 4; }
-class DB     { bits<4> Prefix = 5; }
-class DC     { bits<4> Prefix = 6; }
-class DD     { bits<4> Prefix = 7; }
-class DE     { bits<4> Prefix = 8; }
-class DF     { bits<4> Prefix = 9; }
+class REP    { bits<4> Prefix = 2; }
+class D8     { bits<4> Prefix = 3; }
+class D9     { bits<4> Prefix = 4; }
+class DA     { bits<4> Prefix = 5; }
+class DB     { bits<4> Prefix = 6; }
+class DC     { bits<4> Prefix = 7; }
+class DD     { bits<4> Prefix = 8; }
+class DE     { bits<4> Prefix = 9; }
+class DF     { bits<4> Prefix = 10; }
 
 
 
@@ -172,6 +173,14 @@ def XCHGrr32 : X86Inst<"xchg", 0x87, MRMDestReg, Arg32>;        // xchg R32, R32
 def LEAr16 : X86Inst<"lea", 0x8D, MRMSrcMem, Arg16>, OpSize; // R16 = lea [mem]
 def LEAr32 : X86Inst<"lea", 0x8D, MRMSrcMem, Arg32>;         // R32 = lea [mem]
 
+
+def REP_MOVSB : X86Inst<"rep movsb", 0xA4, RawFrm, NoArg>, REP,
+                Imp<[ECX,EDI,ESI], [ECX,EDI,ESI]>;
+def REP_MOVSW : X86Inst<"rep movsw", 0xA5, RawFrm, NoArg>, REP, OpSize,
+                Imp<[ECX,EDI,ESI], [ECX,EDI,ESI]>;
+def REP_MOVSD : X86Inst<"rep movsd", 0xA5, RawFrm, NoArg>, REP,
+                Imp<[ECX,EDI,ESI], [ECX,EDI,ESI]>;
+
 //===----------------------------------------------------------------------===//
 //  Move Instructions...
 //