Properly handle divides. As a bonus - implement memory versions of them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@76003 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 313 insertions, 66 deletions

diff --git a/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp b/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp
index e009eec48b..3df814e1bb 100644
--- a/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp
+++ b/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp
@@ -30,6 +30,10 @@
 #include "llvm/Support/Debug.h"
 using namespace llvm;
 
+static const unsigned subreg_32bit = 1;
+static const unsigned subreg_even = 1;
+static const unsigned subreg_odd = 2;
+
 namespace {
   /// SystemZRRIAddressMode - This corresponds to rriaddr, but uses SDValue's
   /// instead of register numbers for the leaves of the matched tree.
@@ -129,6 +133,10 @@ namespace {
                       SDValue &Base, SDValue &Disp, SDValue &Index);
 
     SDNode *Select(SDValue Op);
+
+    bool TryFoldLoad(SDValue P, SDValue N,
+                     SDValue &Base, SDValue &Disp, SDValue &Index);
+
     bool MatchAddress(SDValue N, SystemZRRIAddressMode &AM,
                       bool is12Bit, unsigned Depth = 0);
     bool MatchAddressBase(SDValue N, SystemZRRIAddressMode &AM);
@@ -573,6 +581,15 @@ bool SystemZDAGToDAGISel::SelectLAAddr(SDValue Op, SDValue Addr,
   return false;
 }
 
+bool SystemZDAGToDAGISel::TryFoldLoad(SDValue P, SDValue N,
+                                 SDValue &Base, SDValue &Disp, SDValue &Index) {
+  if (ISD::isNON_EXTLoad(N.getNode()) &&
+      N.hasOneUse() &&
+      IsLegalAndProfitableToFold(N.getNode(), P.getNode(), P.getNode()))
+    return SelectAddrRRI20(P, N.getOperand(1), Base, Disp, Index);
+  return false;
+}
+
 /// InstructionSelect - This callback is invoked by
 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
 void SystemZDAGToDAGISel::InstructionSelect() {
@@ -593,7 +610,9 @@ void SystemZDAGToDAGISel::InstructionSelect() {
 
 SDNode *SystemZDAGToDAGISel::Select(SDValue Op) {
   SDNode *Node = Op.getNode();
+  MVT NVT = Node->getValueType(0);
   DebugLoc dl = Op.getDebugLoc();
+  unsigned Opcode = Node->getOpcode();
 
   // Dump information about the Node being selected
   #ifndef NDEBUG
@@ -611,8 +630,195 @@ SDNode *SystemZDAGToDAGISel::Select(SDValue Op) {
     DOUT << "\n";
     Indent -= 2;
     #endif
+    return NULL; // Already selected.
+  }
+
+  switch (Opcode) {
+  default: break;
+  case ISD::SDIVREM: {
+    unsigned Opc, MOpc, ClrOpc = 0;
+    SDValue N0 = Node->getOperand(0);
+    SDValue N1 = Node->getOperand(1);
+
+    MVT ResVT;
+    switch (NVT.getSimpleVT()) {
+      default: assert(0 && "Unsupported VT!");
+      case MVT::i32:
+        Opc = SystemZ::SDIVREM32r; MOpc = SystemZ::SDIVREM32m;
+        ClrOpc = SystemZ::MOV32ri16;
+        ResVT = MVT::v2i32;
+        break;
+      case MVT::i64:
+        Opc = SystemZ::SDIVREM64r; MOpc = SystemZ::SDIVREM64m;
+        ResVT = MVT::v2i64;
+        break;
+    }
+
+    SDValue Tmp0, Tmp1, Tmp2;
+    bool foldedLoad = TryFoldLoad(Op, N1, Tmp0, Tmp1, Tmp2);
+
+    // Prepare the dividend
+    SDNode *Dividend = N0.getNode();
+
+    // Insert prepared dividend into suitable 'subreg'
+    SDNode *Tmp = CurDAG->getTargetNode(TargetInstrInfo::IMPLICIT_DEF,
+                                        dl, ResVT);
+    Dividend =
+      CurDAG->getTargetNode(TargetInstrInfo::INSERT_SUBREG, dl, ResVT,
+                            SDValue(Tmp, 0), SDValue(Dividend, 0),
+                            CurDAG->getTargetConstant(subreg_odd, MVT::i32));
+
+    // Zero out even subreg, if needed
+    if (ClrOpc) {
+      SDNode * ZeroHi = CurDAG->getTargetNode(SystemZ::MOV32ri16, dl, NVT,
+                                        CurDAG->getTargetConstant(0, MVT::i32));
+      Dividend =
+        CurDAG->getTargetNode(TargetInstrInfo::INSERT_SUBREG, dl, ResVT,
+                              SDValue(Dividend, 0),
+                              SDValue(ZeroHi, 0),
+                              CurDAG->getTargetConstant(subreg_even, MVT::i32));
+    }
+
+    SDNode *Result;
+    SDValue DivVal = SDValue(Dividend, 0);
+    if (foldedLoad) {
+      SDValue Ops[] = { DivVal, Tmp0, Tmp1, Tmp2, N1.getOperand(0) };
+      Result = CurDAG->getTargetNode(MOpc, dl, ResVT, Ops, array_lengthof(Ops));
+      // Update the chain.
+      ReplaceUses(N1.getValue(1), SDValue(Result, 0));
+    } else {
+      Result = CurDAG->getTargetNode(Opc, dl, ResVT, SDValue(Dividend, 0), N1);
+    }
+
+    // Copy the division (odd subreg) result, if it is needed.
+    if (!Op.getValue(0).use_empty()) {
+      SDNode *Div = CurDAG->getTargetNode(TargetInstrInfo::EXTRACT_SUBREG,
+                                          dl, NVT,
+                                          SDValue(Result, 0),
+                                          CurDAG->getTargetConstant(subreg_odd,
+                                                                    MVT::i32));
+      ReplaceUses(Op.getValue(0), SDValue(Div, 0));
+      #ifndef NDEBUG
+      DOUT << std::string(Indent-2, ' ') << "=> ";
+      DEBUG(Result->dump(CurDAG));
+      DOUT << "\n";
+      #endif
+    }
+
+    // Copy the remainder (even subreg) result, if it is needed.
+    if (!Op.getValue(1).use_empty()) {
+      SDNode *Rem = CurDAG->getTargetNode(TargetInstrInfo::EXTRACT_SUBREG,
+                                          dl, NVT,
+                                          SDValue(Result, 0),
+                                          CurDAG->getTargetConstant(subreg_even,
+                                                                    MVT::i32));
+      ReplaceUses(Op.getValue(1), SDValue(Rem, 0));
+      #ifndef NDEBUG
+      DOUT << std::string(Indent-2, ' ') << "=> ";
+      DEBUG(Result->dump(CurDAG));
+      DOUT << "\n";
+      #endif
+    }
+
+#ifndef NDEBUG
+    Indent -= 2;
+#endif
+
     return NULL;
   }
+  case ISD::UDIVREM: {
+    unsigned Opc, MOpc, ClrOpc;
+    SDValue N0 = Node->getOperand(0);
+    SDValue N1 = Node->getOperand(1);
+    MVT ResVT;
+
+    switch (NVT.getSimpleVT()) {
+      default: assert(0 && "Unsupported VT!");
+      case MVT::i32:
+        Opc = SystemZ::UDIVREM32r; MOpc = SystemZ::UDIVREM32m;
+        ClrOpc = SystemZ::MOV32ri16;
+        ResVT = MVT::v2i32;
+        break;
+      case MVT::i64:
+        Opc = SystemZ::UDIVREM64r; MOpc = SystemZ::UDIVREM64m;
+        ClrOpc = SystemZ::MOV64ri16;
+        ResVT = MVT::v2i64;
+        break;
+    }
+
+    SDValue Tmp0, Tmp1, Tmp2;
+    bool foldedLoad = TryFoldLoad(Op, N1, Tmp0, Tmp1, Tmp2);
+
+    // Prepare the dividend
+    SDNode *Dividend = N0.getNode();
+
+    // Insert prepared dividend into suitable 'subreg'
+    SDNode *Tmp = CurDAG->getTargetNode(TargetInstrInfo::IMPLICIT_DEF,
+                                        dl, ResVT);
+    Dividend =
+      CurDAG->getTargetNode(TargetInstrInfo::INSERT_SUBREG, dl, ResVT,
+                            SDValue(Tmp, 0), SDValue(Dividend, 0),
+                            CurDAG->getTargetConstant(subreg_odd, MVT::i32));
+
+    // Zero out even subreg, if needed
+    SDNode * ZeroHi = CurDAG->getTargetNode(ClrOpc, dl, NVT,
+                                            CurDAG->getTargetConstant(0,
+                                                                      MVT::i32));
+    Dividend =
+      CurDAG->getTargetNode(TargetInstrInfo::INSERT_SUBREG, dl, ResVT,
+                            SDValue(Dividend, 0),
+                            SDValue(ZeroHi, 0),
+                            CurDAG->getTargetConstant(subreg_even, MVT::i32));
+
+    SDValue DivVal = SDValue(Dividend, 0);
+    SDNode *Result;
+    if (foldedLoad) {
+      SDValue Ops[] = { DivVal, Tmp0, Tmp1, Tmp2, N1.getOperand(0) };
+      Result = CurDAG->getTargetNode(MOpc, dl,ResVT,
+                                     Ops, array_lengthof(Ops));
+      // Update the chain.
+      ReplaceUses(N1.getValue(1), SDValue(Result, 0));
+    } else {
+      Result = CurDAG->getTargetNode(Opc, dl, ResVT, DivVal, N1);
+    }
+
+    // Copy the division (odd subreg) result, if it is needed.
+    if (!Op.getValue(0).use_empty()) {
+      SDNode *Div = CurDAG->getTargetNode(TargetInstrInfo::EXTRACT_SUBREG,
+                                          dl, NVT,
+                                          SDValue(Result, 0),
+                                          CurDAG->getTargetConstant(subreg_odd,
+                                                                    MVT::i32));
+      ReplaceUses(Op.getValue(0), SDValue(Div, 0));
+      #ifndef NDEBUG
+      DOUT << std::string(Indent-2, ' ') << "=> ";
+      DEBUG(Result->dump(CurDAG));
+      DOUT << "\n";
+      #endif
+    }
+
+    // Copy the remainder (even subreg) result, if it is needed.
+    if (!Op.getValue(1).use_empty()) {
+      SDNode *Rem = CurDAG->getTargetNode(TargetInstrInfo::EXTRACT_SUBREG,
+                                          dl, NVT,
+                                          SDValue(Result, 0),
+                                          CurDAG->getTargetConstant(subreg_even,
+                                                                    MVT::i32));
+      ReplaceUses(Op.getValue(1), SDValue(Rem, 0));
+      #ifndef NDEBUG
+      DOUT << std::string(Indent-2, ' ') << "=> ";
+      DEBUG(Result->dump(CurDAG));
+      DOUT << "\n";
+      #endif
+    }
+
+#ifndef NDEBUG
+    Indent -= 2;
+#endif
+
+    return NULL;
+  }
+  }
 
   // Select the default instruction
   SDNode *ResNode = SelectCode(Op);
diff --git a/lib/Target/SystemZ/SystemZISelLowering.cpp b/lib/Target/SystemZ/SystemZISelLowering.cpp
index 096fe420cc..952f4edb63 100644
--- a/lib/Target/SystemZ/SystemZISelLowering.cpp
+++ b/lib/Target/SystemZ/SystemZISelLowering.cpp
@@ -43,7 +43,9 @@ SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) :
   // Set up the register classes.
   addRegisterClass(MVT::i32,  SystemZ::GR32RegisterClass);
   addRegisterClass(MVT::i64,  SystemZ::GR64RegisterClass);
+  addRegisterClass(MVT::v2i32,SystemZ::GR64PRegisterClass);
   addRegisterClass(MVT::i128, SystemZ::GR128RegisterClass);
+  addRegisterClass(MVT::v2i64,SystemZ::GR128RegisterClass);
 
   // Compute derived properties from the register classes
   computeRegisterProperties();
@@ -70,6 +72,15 @@ SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) :
   setOperationAction(ISD::JumpTable,        MVT::i64, Custom);
   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
 
+  setOperationAction(ISD::SDIV,             MVT::i32, Expand);
+  setOperationAction(ISD::UDIV,             MVT::i32, Expand);
+  setOperationAction(ISD::SDIV,             MVT::i64, Expand);
+  setOperationAction(ISD::UDIV,             MVT::i64, Expand);
+  setOperationAction(ISD::SREM,             MVT::i32, Expand);
+  setOperationAction(ISD::UREM,             MVT::i32, Expand);
+  setOperationAction(ISD::SREM,             MVT::i64, Expand);
+  setOperationAction(ISD::UREM,             MVT::i64, Expand);
+
   // FIXME: Can we lower these 2 efficiently?
   setOperationAction(ISD::SETCC,            MVT::i32, Expand);
   setOperationAction(ISD::SETCC,            MVT::i64, Expand);
diff --git a/lib/Target/SystemZ/SystemZInstrInfo.td b/lib/Target/SystemZ/SystemZInstrInfo.td
index 17bf19c6d9..5a34d95e67 100644
--- a/lib/Target/SystemZ/SystemZInstrInfo.td
+++ b/lib/Target/SystemZ/SystemZInstrInfo.td
@@ -580,22 +580,34 @@ def MULSX64rr32 : Pseudo<(outs GR64:$dst), (ins GR64:$src1, GR32:$src2),
                          "msgfr\t{$dst, $src2}",
                          [(set GR64:$dst, (mul GR64:$src1, (sext GR32:$src2)))]>;
 
-def SDIVREM64rrP  : Pseudo<(outs GR64P:$dst), (ins GR64P:$src1, GR32:$src2),
-                           "dr\t{$dst, $src2}",
-                           []>;
-
-def SDIVREM128rrP : Pseudo<(outs GR128:$dst), (ins GR128:$src1, GR64:$src2),
-                           "dsgr\t{$dst, $src2}",
-                           []>;
-
-def UDIVREM64rrP  : Pseudo<(outs GR64P:$dst), (ins GR64P:$src1, GR32:$src2),
-                           "dlr\t{$dst, $src2}",
-                           []>;
+def SDIVREM32r : Pseudo<(outs GR64P:$dst), (ins GR64P:$src1, GR32:$src2),
+                        "dr\t{$dst, $src2}",
+                        []>;
+def SDIVREM64r : Pseudo<(outs GR128:$dst), (ins GR128:$src1, GR64:$src2),
+                        "dsgr\t{$dst, $src2}",
+                        []>;
 
-def UDIVREM128rrP : Pseudo<(outs GR128:$dst), (ins GR128:$src1, GR64:$src2),
-                           "dlgr\t{$dst, $src2}",
-                           []>;
+def UDIVREM32r : Pseudo<(outs GR64P:$dst), (ins GR64P:$src1, GR32:$src2),
+                        "dlr\t{$dst, $src2}",
+                        []>;
+def UDIVREM64r : Pseudo<(outs GR128:$dst), (ins GR128:$src1, GR64:$src2),
+                        "dlgr\t{$dst, $src2}",
+                        []>;
+let mayLoad = 1 in {
+def SDIVREM32m : Pseudo<(outs GR64P:$dst), (ins GR64P:$src1, rriaddr:$src2),
+                        "d\t{$dst, $src2}",
+                        []>;
+def SDIVREM64m : Pseudo<(outs GR128:$dst), (ins GR128:$src1, rriaddr:$src2),
+                        "dsg\t{$dst, $src2}",
+                        []>;
 
+def UDIVREM32m : Pseudo<(outs GR64P:$dst), (ins GR64P:$src1, rriaddr:$src2),
+                        "dl\t{$dst, $src2}",
+                        []>;
+def UDIVREM64m : Pseudo<(outs GR128:$dst), (ins GR128:$src1, rriaddr:$src2),
+                        "dlg\t{$dst, $src2}",
+                        []>;
+} // mayLoad
 } // isTwoAddress = 1
 
 //===----------------------------------------------------------------------===//
@@ -790,51 +802,5 @@ def : Pat<(mulhu GR64:$src1, GR64:$src2),
                                       GR64:$src2),
                           subreg_even)>;
 
-// divs
-// FIXME: Add memory versions
-def : Pat<(sdiv GR32:$src1, GR32:$src2),
-          (EXTRACT_SUBREG (SDIVREM64rrP (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-                                                       GR32:$src1, subreg_odd),
-                                         GR32:$src2),
-                          subreg_odd)>;
-def : Pat<(sdiv GR64:$src1, GR64:$src2),
-          (EXTRACT_SUBREG (SDIVREM128rrP (INSERT_SUBREG (i128 (IMPLICIT_DEF)),
-                                                        GR64:$src1, subreg_odd),
-                                         GR64:$src2),
-                          subreg_odd)>;
-def : Pat<(udiv GR32:$src1, GR32:$src2),
-          (EXTRACT_SUBREG (UDIVREM64rrP (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-                                                       GR32:$src1, subreg_odd),
-                                         GR32:$src2),
-                          subreg_odd)>;
-def : Pat<(udiv GR64:$src1, GR64:$src2),
-          (EXTRACT_SUBREG (UDIVREM128rrP (INSERT_SUBREG (i128 (IMPLICIT_DEF)),
-                                                        GR64:$src1, subreg_odd),
-                                         GR64:$src2),
-                          subreg_odd)>;
-
-// rems
-// FIXME: Add memory versions
-def : Pat<(srem GR32:$src1, GR32:$src2),
-          (EXTRACT_SUBREG (SDIVREM64rrP (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-                                                       GR32:$src1, subreg_odd),
-                                         GR32:$src2),
-                          subreg_even)>;
-def : Pat<(srem GR64:$src1, GR64:$src2),
-          (EXTRACT_SUBREG (SDIVREM128rrP (INSERT_SUBREG (i128 (IMPLICIT_DEF)),
-                                                        GR64:$src1, subreg_odd),
-                                         GR64:$src2),
-                          subreg_even)>;
-def : Pat<(urem GR32:$src1, GR32:$src2),
-          (EXTRACT_SUBREG (UDIVREM64rrP (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-                                                       GR32:$src1, subreg_odd),
-                                         GR32:$src2),
-                          subreg_even)>;
-def : Pat<(urem GR64:$src1, GR64:$src2),
-          (EXTRACT_SUBREG (UDIVREM128rrP (INSERT_SUBREG (i128 (IMPLICIT_DEF)),
-                                                        GR64:$src1, subreg_odd),
-                                         GR64:$src2),
-                          subreg_even)>;
-
 def : Pat<(i32 imm:$src),
           (EXTRACT_SUBREG (MOV64ri32 (i64 imm:$src)), subreg_32bit)>;
diff --git a/lib/Target/SystemZ/SystemZRegisterInfo.td b/lib/Target/SystemZ/SystemZRegisterInfo.td
index d3e7159364..6fbb4cdb87 100644
--- a/lib/Target/SystemZ/SystemZRegisterInfo.td
+++ b/lib/Target/SystemZ/SystemZRegisterInfo.td
@@ -332,7 +332,7 @@ def ADDR64 : RegisterClass<"SystemZ", [i64], 64,
 }
 
 // Even-odd register pairs
-def GR64P : RegisterClass<"SystemZ", [i64], 64,
+def GR64P : RegisterClass<"SystemZ", [v2i32], 64,
   [R0P, R2P, R4P, R6P, R8P, R10P, R12P, R14P]>
 {
   let SubRegClassList = [GR32, GR32];
@@ -368,7 +368,7 @@ def GR64P : RegisterClass<"SystemZ", [i64], 64,
   }];
 }
 
-def GR128 : RegisterClass<"SystemZ", [i128], 128,
+def GR128 : RegisterClass<"SystemZ", [i128, v2i64], 128,
   [R0Q, R2Q, R4Q, R6Q, R8Q, R10Q, R12Q, R14Q]>
 {
   let SubRegClassList = [GR64, GR64];
diff --git a/test/CodeGen/SystemZ/08-DivRem.ll b/test/CodeGen/SystemZ/08-DivRem.ll
index ea9214f812..45cbfea0e0 100644
--- a/test/CodeGen/SystemZ/08-DivRem.ll
+++ b/test/CodeGen/SystemZ/08-DivRem.ll
@@ -1,7 +1,7 @@
-; RUN: llvm-as < %s | llc | grep dsgr | count 2
-; RUN: llvm-as < %s | llc | grep dr   | count 2
-; RUN: llvm-as < %s | llc | grep dlr  | count 2
-; RUN: llvm-as < %s | llc | grep dlgr | count 2
+; RUN: llvm-as < %s | llc | grep dsgr  | count 2
+; RUN: llvm-as < %s | llc | grep dr    | count 2
+; RUN: llvm-as < %s | llc | grep dlr   | count 2
+; RUN: llvm-as < %s | llc | grep dlgr  | count 2
 
 target datalayout = "E-p:64:64:64-i1:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128"
 target triple = "s390x-unknown-linux-gnu"
diff --git a/test/CodeGen/SystemZ/08-DivRemMemOp.ll b/test/CodeGen/SystemZ/08-DivRemMemOp.ll
new file mode 100644
index 0000000000..41eef36b2d
--- /dev/null
+++ b/test/CodeGen/SystemZ/08-DivRemMemOp.ll
@@ -0,0 +1,64 @@
+; RUN: llvm-as < %s | llc | grep {d.%}  | count 2
+; RUN: llvm-as < %s | llc | grep dsg    | count 2
+; RUN: llvm-as < %s | llc | grep {dl.%} | count 2
+; RUN: llvm-as < %s | llc | grep dlg    | count 2
+
+target datalayout = "E-p:64:64:64-i1:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128"
+target triple = "s390x-unknown-linux-gnu"
+
+define i64 @div(i64 %a, i64* %b) nounwind readnone {
+entry:
+	%b1 = load i64* %b
+	%div = sdiv i64 %a, %b1
+	ret i64 %div
+}
+
+define i64 @div1(i64 %a, i64* %b) nounwind readnone {
+entry:
+        %b1 = load i64* %b
+        %div = udiv i64 %a, %b1
+        ret i64 %div
+}
+
+define i64 @rem(i64 %a, i64* %b) nounwind readnone {
+entry:
+        %b1 = load i64* %b
+        %div = srem i64 %a, %b1
+        ret i64 %div
+}
+
+define i64 @rem1(i64 %a, i64* %b) nounwind readnone {
+entry:
+        %b1 = load i64* %b
+        %div = urem i64 %a, %b1
+        ret i64 %div
+}
+
+define i32 @div2(i32 %a, i32* %b) nounwind readnone {
+entry:
+        %b1 = load i32* %b
+        %div = sdiv i32 %a, %b1
+        ret i32 %div
+}
+
+define i32 @div3(i32 %a, i32* %b) nounwind readnone {
+entry:
+        %b1 = load i32* %b
+        %div = udiv i32 %a, %b1
+        ret i32 %div
+}
+
+define i32 @rem2(i32 %a, i32* %b) nounwind readnone {
+entry:
+        %b1 = load i32* %b
+        %div = srem i32 %a, %b1
+        ret i32 %div
+}
+
+define i32 @rem3(i32 %a, i32* %b) nounwind readnone {
+entry:
+        %b1 = load i32* %b
+        %div = urem i32 %a, %b1
+        ret i32 %div
+}
+