When both x/y and x%y are needed (x and y both scalar integer), compute

both results with a single div or idiv instruction. This uses new X86ISD
nodes for DIV and IDIV which are introduced during the legalize phase
so that the SelectionDAG's CSE can automatically eliminate redundant
computations.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@42308 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 146 insertions, 25 deletions

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index b33ced8ba6..117d2734c5 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -1162,12 +1162,9 @@ SDNode *X86DAGToDAGISel::Select(SDOperand N) {
       return NULL;
     }
       
-    case ISD::SDIV:
-    case ISD::UDIV:
-    case ISD::SREM:
-    case ISD::UREM: {
-      bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
-      bool isDiv    = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
+    case X86ISD::DIV:
+    case X86ISD::IDIV: {
+      bool isSigned = Opcode == X86ISD::IDIV;
       if (!isSigned)
         switch (NVT) {
         default: assert(0 && "Unsupported VT!");
@@ -1275,31 +1272,49 @@ SDNode *X86DAGToDAGISel::Select(SDOperand N) {
           SDOperand(CurDAG->getTargetNode(Opc, MVT::Flag, N1, InFlag), 0);
       }
 
-      unsigned Reg = isDiv ? LoReg : HiReg;
-      SDOperand Result;
-      if (Reg == X86::AH && Subtarget->is64Bit()) {
-        // Prevent use of AH in a REX instruction by referencing AX instead.
-        // Shift it down 8 bits.
-        Result = CurDAG->getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
-        Chain = Result.getValue(1);
-        Result = SDOperand(CurDAG->getTargetNode(X86::SHR16ri, MVT::i16, Result,
-                                     CurDAG->getTargetConstant(8, MVT::i8)), 0);
-        // Then truncate it down to i8.
-        SDOperand SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
-        Result = SDOperand(CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
-                                                 MVT::i8, Result, SRIdx), 0);
-      } else {
-        Result = CurDAG->getCopyFromReg(Chain, Reg, NVT, InFlag);
+      // Copy the division (low) result, if it is needed.
+      if (!N.getValue(0).use_empty()) {
+        SDOperand Result = CurDAG->getCopyFromReg(Chain, LoReg, NVT, InFlag);
         Chain = Result.getValue(1);
+        InFlag = Result.getValue(2);
+        ReplaceUses(N.getValue(0), Result);
+#ifndef NDEBUG
+        DOUT << std::string(Indent-2, ' ') << "=> ";
+        DEBUG(Result.Val->dump(CurDAG));
+        DOUT << "\n";
+#endif
+      }
+      // Copy the remainder (high) result, if it is needed.
+      if (!N.getValue(1).use_empty()) {
+        SDOperand Result;
+        if (HiReg == X86::AH && Subtarget->is64Bit()) {
+          // Prevent use of AH in a REX instruction by referencing AX instead.
+          // Shift it down 8 bits.
+          Result = CurDAG->getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
+          Chain = Result.getValue(1);
+          InFlag = Result.getValue(2);
+          Result = SDOperand(CurDAG->getTargetNode(X86::SHR16ri, MVT::i16, Result,
+                                       CurDAG->getTargetConstant(8, MVT::i8)), 0);
+          // Then truncate it down to i8.
+          SDOperand SRIdx = CurDAG->getTargetConstant(1, MVT::i32); // SubRegSet 1
+          Result = SDOperand(CurDAG->getTargetNode(X86::EXTRACT_SUBREG,
+                                                   MVT::i8, Result, SRIdx), 0);
+        } else {
+          Result = CurDAG->getCopyFromReg(Chain, HiReg, NVT, InFlag);
+          Chain = Result.getValue(1);
+          InFlag = Result.getValue(2);
+        }
+        ReplaceUses(N.getValue(1), Result);
+#ifndef NDEBUG
+        DOUT << std::string(Indent-2, ' ') << "=> ";
+        DEBUG(Result.Val->dump(CurDAG));
+        DOUT << "\n";
+#endif
       }
-      ReplaceUses(N.getValue(0), Result);
       if (foldedLoad)
         ReplaceUses(N1.getValue(1), Chain);
 
 #ifndef NDEBUG
-      DOUT << std::string(Indent-2, ' ') << "=> ";
-      DEBUG(Result.Val->dump(CurDAG));
-      DOUT << "\n";
       Indent -= 2;
 #endif
 
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index a67e77f2ce..1a8089688b 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -155,6 +155,27 @@ X86TargetLowering::X86TargetLowering(TargetMachine &TM)
     setOperationAction(ISD::BIT_CONVERT      , MVT::i32  , Expand);
   }
 
+  // Divide and remainder are lowered to use div or idiv in legalize in
+  // order to expose the intermediate computations to trivial CSE. This is
+  // most noticeable when both x/y and x%y are being computed; they can be
+  // done with a single div or idiv.
+  setOperationAction(ISD::SDIV            , MVT::i8    , Custom);
+  setOperationAction(ISD::UDIV            , MVT::i8    , Custom);
+  setOperationAction(ISD::SREM            , MVT::i8    , Custom);
+  setOperationAction(ISD::UREM            , MVT::i8    , Custom);
+  setOperationAction(ISD::SDIV            , MVT::i16   , Custom);
+  setOperationAction(ISD::UDIV            , MVT::i16   , Custom);
+  setOperationAction(ISD::SREM            , MVT::i16   , Custom);
+  setOperationAction(ISD::UREM            , MVT::i16   , Custom);
+  setOperationAction(ISD::SDIV            , MVT::i32   , Custom);
+  setOperationAction(ISD::UDIV            , MVT::i32   , Custom);
+  setOperationAction(ISD::SREM            , MVT::i32   , Custom);
+  setOperationAction(ISD::UREM            , MVT::i32   , Custom);
+  setOperationAction(ISD::SDIV            , MVT::i64   , Custom);
+  setOperationAction(ISD::UDIV            , MVT::i64   , Custom);
+  setOperationAction(ISD::SREM            , MVT::i64   , Custom);
+  setOperationAction(ISD::UREM            , MVT::i64   , Custom);
+
   setOperationAction(ISD::BR_JT            , MVT::Other, Expand);
   setOperationAction(ISD::BRCOND           , MVT::Other, Custom);
   setOperationAction(ISD::BR_CC            , MVT::Other, Expand);
@@ -3393,6 +3414,22 @@ SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
     return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
 }
 
+SDOperand X86TargetLowering::LowerIntegerDivOrRem(SDOperand Op, SelectionDAG &DAG) {
+  unsigned Opcode = Op.getOpcode();
+  MVT::ValueType NVT = Op.getValueType();
+  bool isSigned = Opcode == ISD::SDIV || Opcode == ISD::SREM;
+  bool isDiv    = Opcode == ISD::SDIV || Opcode == ISD::UDIV;
+  unsigned Opc = isSigned ? X86ISD::IDIV : X86ISD::DIV;
+
+  SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
+  SDOperand DR = DAG.getNode(Opc, DAG.getVTList(NVT, NVT), Ops, 2);
+
+  if (isDiv)
+    return DR;
+
+  return SDOperand(DR.Val, 1);
+}
+
 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
   assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
          Op.getOperand(0).getValueType() >= MVT::i16 &&
@@ -4668,6 +4705,10 @@ SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
   case ISD::SHL_PARTS:
   case ISD::SRA_PARTS:
   case ISD::SRL_PARTS:          return LowerShift(Op, DAG);
+  case ISD::SDIV:
+  case ISD::UDIV:
+  case ISD::SREM:
+  case ISD::UREM:               return LowerIntegerDivOrRem(Op, DAG);
   case ISD::SINT_TO_FP:         return LowerSINT_TO_FP(Op, DAG);
   case ISD::FP_TO_SINT:         return LowerFP_TO_SINT(Op, DAG);
   case ISD::FABS:               return LowerFABS(Op, DAG);
@@ -4751,6 +4792,8 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::TLSADDR:            return "X86ISD::TLSADDR";
   case X86ISD::THREAD_POINTER:     return "X86ISD::THREAD_POINTER";
   case X86ISD::EH_RETURN:          return "X86ISD::EH_RETURN";
+  case X86ISD::DIV:                return "X86ISD::DIV";
+  case X86ISD::IDIV:               return "X86ISD::IDIV";
   }
 }
 
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 04279e835c..10172d95a9 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -181,6 +181,10 @@ namespace llvm {
       /// in order to obtain suitable precision.
       FRSQRT, FRCP,
 
+      /// DIV, IDIV - Unsigned and signed integer division and reciprocal.
+      ///
+      DIV, IDIV,
+
       // Thread Local Storage
       TLSADDR, THREAD_POINTER,
 
@@ -420,6 +424,7 @@ namespace llvm {
     SDOperand LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerShift(SDOperand Op, SelectionDAG &DAG);
+    SDOperand LowerIntegerDivOrRem(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerFABS(SDOperand Op, SelectionDAG &DAG);
diff --git a/test/CodeGen/X86/divrem.ll b/test/CodeGen/X86/divrem.ll
new file mode 100644
index 0000000000..a611eddc76
--- /dev/null
+++ b/test/CodeGen/X86/divrem.ll
@@ -0,0 +1,58 @@
+; RUN: llvm-as < %s | llc -march=x86-64 | grep div | count 8
+
+define void @si64(i64 %x, i64 %y, i64* %p, i64* %q) {
+	%r = sdiv i64 %x, %y
+	%t = srem i64 %x, %y
+	store i64 %r, i64* %p
+	store i64 %t, i64* %q
+	ret void
+}
+define void @si32(i32 %x, i32 %y, i32* %p, i32* %q) {
+	%r = sdiv i32 %x, %y
+	%t = srem i32 %x, %y
+	store i32 %r, i32* %p
+	store i32 %t, i32* %q
+	ret void
+}
+define void @si16(i16 %x, i16 %y, i16* %p, i16* %q) {
+	%r = sdiv i16 %x, %y
+	%t = srem i16 %x, %y
+	store i16 %r, i16* %p
+	store i16 %t, i16* %q
+	ret void
+}
+define void @si8(i8 %x, i8 %y, i8* %p, i8* %q) {
+	%r = sdiv i8 %x, %y
+	%t = srem i8 %x, %y
+	store i8 %r, i8* %p
+	store i8 %t, i8* %q
+	ret void
+}
+define void @ui64(i64 %x, i64 %y, i64* %p, i64* %q) {
+	%r = udiv i64 %x, %y
+	%t = urem i64 %x, %y
+	store i64 %r, i64* %p
+	store i64 %t, i64* %q
+	ret void
+}
+define void @ui32(i32 %x, i32 %y, i32* %p, i32* %q) {
+	%r = udiv i32 %x, %y
+	%t = urem i32 %x, %y
+	store i32 %r, i32* %p
+	store i32 %t, i32* %q
+	ret void
+}
+define void @ui16(i16 %x, i16 %y, i16* %p, i16* %q) {
+	%r = udiv i16 %x, %y
+	%t = urem i16 %x, %y
+	store i16 %r, i16* %p
+	store i16 %t, i16* %q
+	ret void
+}
+define void @ui8(i8 %x, i8 %y, i8* %p, i8* %q) {
+	%r = udiv i8 %x, %y
+	%t = urem i8 %x, %y
+	store i8 %r, i8* %p
+	store i8 %t, i8* %q
+	ret void
+}