[mips] Add CCValAssign::[ASZ]ExtUpper and CCPromoteToUpperBitsInType and handle struct's correctly on big-endian N32/N64 return values.

[dsanders]

Summary

The N32/N64 ABI's require that structs passed in registers are laid out
such that spilling the register with 'sd' places the struct at the lowest
address. For little endian this is trivial but for big-endian it requires
that structs are shifted into the upper bits of the register.

We also require that structs passed in registers have the 'inreg'
attribute for big-endian N32/N64 to work correctly. This is because the
tablegen-erated calling convention implementation only has access to the
lowered form of struct arguments (one or more integers of up to 64-bits
each) and is unable to determine the original type.

[dsanders]

ping...

[vmedic]

LGTM

include/llvm/CodeGen/CallingConvLower.h

 class MVT;
 class TargetMachine;
 class TargetRegisterInfo;
 
 /// CCValAssign - Represent assignment of one arg/retval to a location.
 class CCValAssign {
 public:
   enum LocInfo {
-    Full,   // The value fills the full location.
-    SExt,   // The value is sign extended in the location.
-    ZExt,   // The value is zero extended in the location.
-    AExt,   // The value is extended with undefined upper bits.
-    BCvt,   // The value is bit-converted in the location.
-    VExt,   // The value is vector-widened in the location.
+    Full,      // The value fills the full location.
+    SExt,      // The value is sign extended in the location.
+    ZExt,      // The value is zero extended in the location.
+    AExt,      // The value is extended with undefined upper bits.
+    SExtUpper, // The value is in the upper bits of the location and should be
+               // sign extended when retrieved.
+    ZExtUpper, // The value is in the upper bits of the location and should be
+               // zero extended when retrieved.
+    AExtUpper, // The value is in the upper bits of the location and should be
+               // extended with undefined upper bits when retrieved.
+    BCvt,      // The value is bit-converted in the location.
+    VExt,      // The value is vector-widened in the location.
                // FIXME: Not implemented yet. Code that uses AExt to mean
                // vector-widen should be fixed to use VExt instead.
-    FPExt,  // The floating-point value is fp-extended in the location.
-    Indirect // The location contains pointer to the value.
+    FPExt,     // The floating-point value is fp-extended in the location.
+    Indirect   // The location contains pointer to the value.
     // TODO: a subset of the value is in the location.
   };
+
 private:
   /// ValNo - This is the value number begin assigned (e.g. an argument number).
   unsigned ValNo;
 
   /// Loc is either a stack offset or a register number.
   unsigned Loc;
 
   /// isMem - True if this is a memory loc, false if it is a register loc.
public:
   unsigned getLocMemOffset() const { assert(isMemLoc()); return Loc; }
   MVT getLocVT() const { return LocVT; }
 
   LocInfo getLocInfo() const { return HTP; }
   bool isExtInLoc() const {
     return (HTP == AExt || HTP == SExt || HTP == ZExt);
   }
 
+  bool isUpperBitsInLoc() const {
+    return HTP == AExtUpper || HTP == SExtUpper || HTP == ZExtUpper;
+  }
 };
 
 /// CCAssignFn - This function assigns a location for Val, updating State to
 /// reflect the change.  It returns 'true' if it failed to handle Val.
 typedef bool CCAssignFn(unsigned ValNo, MVT ValVT,
                         MVT LocVT, CCValAssign::LocInfo LocInfo,
                         ISD::ArgFlagsTy ArgFlags, CCState &State);
 

include/llvm/Target/TargetCallingConv.td

 }
 
 /// CCPromoteToType - If applied, this promotes the specified current value to
 /// the specified type.
 class CCPromoteToType<ValueType destTy> : CCAction {
   ValueType DestTy = destTy;
 }
 
+/// CCPromoteToUpperBitsInType - If applied, this promotes the specified current
+/// value to the specified type and shifts the value into the upper bits.
+class CCPromoteToUpperBitsInType<ValueType destTy> : CCAction {
+  ValueType DestTy = destTy;
+}
+
 /// CCBitConvertToType - If applied, this bitconverts the specified current
 /// value to the specified type.
 class CCBitConvertToType<ValueType destTy> : CCAction {
   ValueType DestTy = destTy;
 }
 
 /// CCPassIndirect - If applied, this stores the value to stack and passes the pointer
 /// as normal argument.

lib/Target/Mips/MipsCallingConv.td

CCIfType<[i64, f64], CCAssignToReg<[A0_64, A1_64, A2_64, A3_64,
                                       T0_64, T1_64, T2_64, T3_64]>>,
 
   // All stack parameter slots become 64-bit doublewords and are 8-byte aligned.
   CCIfType<[i32, f32], CCAssignToStack<4, 8>>,
   CCIfType<[i64, f64], CCAssignToStack<8, 8>>
 ]>;
 
 def RetCC_MipsN : CallingConv<[
+  // Aggregate returns are positioned at the lowest address in the slot for
+  // both little and big-endian targets. When passing in registers, this
+  // requires that big-endian targets shift the value into the upper bits.
+  CCIfSubtarget<"isLittle()",
+      CCIfType<[i8, i16, i32], CCIfInReg<CCPromoteToType<i64>>>>,
+  CCIfNotSubtarget<"isLittle()",
+      CCIfType<[i8, i16, i32], CCIfInReg<CCPromoteToUpperBitsInType<i64>>>>,
+
   // i32 are returned in registers V0, V1
   CCIfType<[i32], CCAssignToReg<[V0, V1]>>,
 
   // i64 are returned in registers V0_64, V1_64
   CCIfType<[i64], CCAssignToReg<[V0_64, V1_64]>>,
 
   // f32 are returned in registers F0, F2
   CCIfType<[f32], CCAssignToReg<[F0, F2]>>,

lib/Target/Mips/MipsISelLowering.cpp

MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
 
   if (originalTypeIsF128(RetTy, CallNode))
     CCInfo.AnalyzeCallResult(Ins, RetCC_F128);
   else
     CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
 
   // Copy all of the result registers out of their specified physreg.
   for (unsigned i = 0; i != RVLocs.size(); ++i) {
+    CCValAssign &VA = RVLocs[i];
+    assert(VA.isRegLoc() && "Can only return in registers!");
+
     SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
                                      RVLocs[i].getLocVT(), InFlag);
     Chain = Val.getValue(1);
     InFlag = Val.getValue(2);
 
-    if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
-      Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
+    if (VA.isUpperBitsInLoc()) {
+      unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits();
+      unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
+      unsigned Shift =
+          VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
+      Val = DAG.getNode(
+          Shift, DL, VA.getLocVT(), Val,
+          DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT()));
+    }
+
+    switch (VA.getLocInfo()) {
+    default:
+      llvm_unreachable("Unknown loc info!");
+    case CCValAssign::Full:
+      break;
+    case CCValAssign::BCvt:
+      Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
+      break;
+    case CCValAssign::AExt:
+    case CCValAssign::AExtUpper:
+      Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
+      break;
+    case CCValAssign::ZExt:
+    case CCValAssign::ZExtUpper:
+      Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
+                        DAG.getValueType(VA.getValVT()));
+      Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
+      break;
+    case CCValAssign::SExt:
+    case CCValAssign::SExtUpper:
+      Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
+                        DAG.getValueType(VA.getValVT()));
+      Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
+      break;
+    }
 
     InVals.push_back(Val);
   }
 
   return Chain;
 }
 
 //===----------------------------------------------------------------------===//
MipsTargetLowering::LowerReturn(SDValue Chain,
   SDValue Flag;
   SmallVector<SDValue, 4> RetOps(1, Chain);
 
   // Copy the result values into the output registers.
   for (unsigned i = 0; i != RVLocs.size(); ++i) {
     SDValue Val = OutVals[i];
     CCValAssign &VA = RVLocs[i];
     assert(VA.isRegLoc() && "Can only return in registers!");
+    bool UseUpperBits = false;
 
-    if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
-      Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
+    switch (VA.getLocInfo()) {
+    default:
+      llvm_unreachable("Unknown loc info!");
+    case CCValAssign::Full:
+      break;
+    case CCValAssign::BCvt:
+      Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val);
+      break;
+    case CCValAssign::AExtUpper:
+      UseUpperBits = true;
+      // Fallthrough
+    case CCValAssign::AExt:
+      Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val);
+      break;
+    case CCValAssign::ZExtUpper:
+      UseUpperBits = true;
+      // Fallthrough
+    case CCValAssign::ZExt:
+      Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val);
+      break;
+    case CCValAssign::SExtUpper:
+      UseUpperBits = true;
+      // Fallthrough
+    case CCValAssign::SExt:
+      Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val);
+      break;
+    }
+
+    if (UseUpperBits) {
+      unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
+      unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
+      Val = DAG.getNode(
+          ISD::SHL, DL, VA.getLocVT(), Val,
+          DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT()));
+    }
 
     Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
 
     // Guarantee that all emitted copies are stuck together with flags.
     Flag = Chain.getValue(1);
     RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
   }
 

test/CodeGen/Mips/cconv/return-struct.ll

+; RUN: llc -mtriple=mips-linux-gnu -relocation-model=static < %s | FileCheck --check-prefix=ALL --check-prefix=O32 --check-prefix=O32-BE %s
+; RUN: llc -mtriple=mipsel-linux-gnu -relocation-model=static < %s | FileCheck --check-prefix=ALL --check-prefix=O32 --check-prefix=O32-LE %s
+
+; RUN-TODO: llc -mtriple=mips64-linux-gnu -relocation-model=static -mattr=-n64,+o32 < %s | FileCheck --check-prefix=ALL --check-prefix=O32 %s
+; RUN-TODO: llc -mtriple=mips64el-linux-gnu -relocation-model=static -mattr=-n64,+o32 < %s | FileCheck --check-prefix=ALL --check-prefix=O32 %s
+
+; RUN: llc -mtriple=mips64-linux-gnu -relocation-model=static -mattr=-n64,+n32 < %s | FileCheck --check-prefix=ALL --check-prefix=N32 --check-prefix=N32-BE %s
+; RUN: llc -mtriple=mips64el-linux-gnu -relocation-model=static -mattr=-n64,+n32 < %s | FileCheck --check-prefix=ALL --check-prefix=N32 --check-prefix=N32-LE %s
+
+; RUN: llc -mtriple=mips64-linux-gnu -relocation-model=static -mattr=-n64,+n64 < %s | FileCheck --check-prefix=ALL --check-prefix=N64 --check-prefix=N64-BE %s
+; RUN: llc -mtriple=mips64el-linux-gnu -relocation-model=static -mattr=-n64,+n64 < %s | FileCheck --check-prefix=ALL --check-prefix=N64 --check-prefix=N64-LE %s
+
+; Test struct returns for all ABI's and byte orders.
+
+@struct_byte = global {i8} zeroinitializer
+@struct_2byte = global {i8,i8} zeroinitializer
+
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32, i1)
+
+define inreg {i8} @ret_struct_i8() nounwind {
+entry:
+        %0 = load volatile {i8}* @struct_byte
+        ret {i8} %0
+}
+
+; ALL-LABEL: ret_struct_i8:
+; O32-DAG:           lui [[R1:\$[0-9]+]], %hi(struct_byte)
+; O32-DAG:           lbu $2, %lo(struct_byte)([[R1]])
+
+; N32-LE-DAG:        lui [[R1:\$[0-9]+]], %hi(struct_byte)
+; N32-LE-DAG:        lb $2, %lo(struct_byte)([[R1]])
+
+; N32-BE-DAG:        lui [[R1:\$[0-9]+]], %hi(struct_byte)
+; N32-BE-DAG:        lb [[R2:\$[0-9]+]], %lo(struct_byte)([[R1]])
+; N32-BE-DAG:        dsll $2, [[R2]], 56
+
+; N64-LE-DAG:        ld  [[R1:\$[0-9]+]], %got_disp(struct_byte)($1)
+; N64-LE-DAG:        lb $2, 0([[R1]])
+
+; N64-BE-DAG:        ld  [[R1:\$[0-9]+]], %got_disp(struct_byte)($1)
+; N64-BE-DAG:        lb [[R2:\$[0-9]+]], 0([[R1]])
+; N64-BE-DAG:        dsll $2, [[R2]], 56
+
+; This test is based on the way clang currently lowers {i8,i8} to {i16}.
+; FIXME: It should probably work for without any lowering too but this doesn't
+;        work as expected. Each member gets mapped to a register rather than
+;        packed into a single register.
+define inreg {i16} @ret_struct_i16() nounwind {
+entry:
+        %retval = alloca {i8,i8}, align 1
+        %0 = bitcast {i8,i8}* %retval to i8*
+        call void @llvm.memcpy.p0i8.p0i8.i64(i8* %0, i8* getelementptr inbounds ({i8,i8}* @struct_2byte, i32 0, i32 0), i64 2, i32 1, i1 false)
+        %1 = bitcast {i8,i8}* %retval to {i16}*
+        %2 = load volatile {i16}* %1
+        ret {i16} %2
+}
+
+; ALL-LABEL: ret_struct_i16:
+; O32-DAG:           lui [[R1:\$[0-9]+]], %hi(struct_2byte)
+; O32-DAG:           lhu [[R2:\$[0-9]+]], %lo(struct_2byte)([[R1]])
+; O32-DAG:           sh  [[R2]], 0([[SP:\$sp]])
+; O32-DAG:           lhu $2, 0([[SP:\$sp]])
+
+; N32-LE-DAG:        lui [[R1:\$[0-9]+]], %hi(struct_2byte)
+; N32-LE-DAG:        lhu [[R2:\$[0-9]+]], %lo(struct_2byte)([[R1]])
+; N32-LE-DAG:        sh  [[R2]], 8([[SP:\$sp]])
+; N32-LE-DAG:        lh  $2, 8([[SP:\$sp]])
+
+; N32-BE-DAG:        lui [[R1:\$[0-9]+]], %hi(struct_2byte)
+; N32-BE-DAG:        lhu [[R2:\$[0-9]+]], %lo(struct_2byte)([[R1]])
+; N32-BE-DAG:        sh  [[R2]], 8([[SP:\$sp]])
+; N32-BE-DAG:        lh  [[R3:\$[0-9]+]], 8([[SP:\$sp]])
+; N32-BE-DAG:        dsll $2, [[R3]], 48
+
+; N64-LE-DAG:        ld  [[R1:\$[0-9]+]], %got_disp(struct_2byte)($1)
+; N64-LE-DAG:        lhu [[R2:\$[0-9]+]], 0([[R1]])
+; N64-LE-DAG:        sh  [[R2]], 8([[SP:\$sp]])
+; N64-LE-DAG:        lh  $2, 8([[SP:\$sp]])
+
+; N64-BE-DAG:        ld  [[R1:\$[0-9]+]], %got_disp(struct_2byte)($1)
+; N64-BE-DAG:        lhu [[R2:\$[0-9]+]], 0([[R1]])
+; N64-BE-DAG:        sh  [[R2]], 8([[SP:\$sp]])
+; N64-BE-DAG:        lh  [[R3:\$[0-9]+]], 8([[SP:\$sp]])
+; N64-BE-DAG:        dsll $2, [[R3]], 48

utils/TableGen/CallingConvEmitter.cpp

if (Action->isSubClassOf("CCDelegateTo")) {
       } else {
         O << IndentStr << "if (ArgFlags.isSExt())\n"
           << IndentStr << IndentStr << "LocInfo = CCValAssign::SExt;\n"
           << IndentStr << "else if (ArgFlags.isZExt())\n"
           << IndentStr << IndentStr << "LocInfo = CCValAssign::ZExt;\n"
           << IndentStr << "else\n"
           << IndentStr << IndentStr << "LocInfo = CCValAssign::AExt;\n";
       }
+    } else if (Action->isSubClassOf("CCPromoteToUpperBitsInType")) {
+      Record *DestTy = Action->getValueAsDef("DestTy");
+      MVT::SimpleValueType DestVT = getValueType(DestTy);
+      O << IndentStr << "LocVT = " << getEnumName(DestVT) << ";\n";
+      if (MVT(DestVT).isFloatingPoint()) {
+        PrintFatalError("CCPromoteToUpperBitsInType does not handle floating "
+                        "point");
+      } else {
+        O << IndentStr << "if (ArgFlags.isSExt())\n"
+          << IndentStr << IndentStr << "LocInfo = CCValAssign::SExtUpper;\n"
+          << IndentStr << "else if (ArgFlags.isZExt())\n"
+          << IndentStr << IndentStr << "LocInfo = CCValAssign::ZExtUpper;\n"
+          << IndentStr << "else\n"
+          << IndentStr << IndentStr << "LocInfo = CCValAssign::AExtUpper;\n";
+      }
     } else if (Action->isSubClassOf("CCBitConvertToType")) {
       Record *DestTy = Action->getValueAsDef("DestTy");
       O << IndentStr << "LocVT = " << getEnumName(getValueType(DestTy)) <<";\n";
       O << IndentStr << "LocInfo = CCValAssign::BCvt;\n";
     } else if (Action->isSubClassOf("CCPassIndirect")) {
       Record *DestTy = Action->getValueAsDef("DestTy");
       O << IndentStr << "LocVT = " << getEnumName(getValueType(DestTy)) <<";\n";
       O << IndentStr << "LocInfo = CCValAssign::Indirect;\n";