-//===------- LegalizeVectorTypes.cpp - Legalization of vector types -------===//\r
-//\r
-// The LLVM Compiler Infrastructure\r
-//\r
-// This file is distributed under the University of Illinois Open Source\r
-// License. See LICENSE.TXT for details.\r
-//\r
-//===----------------------------------------------------------------------===//\r
-//\r
-// This file performs vector type splitting and scalarization for LegalizeTypes.\r
-// Scalarization is the act of changing a computation in an illegal one-element\r
-// vector type to be a computation in its scalar element type. For example,\r
-// implementing <1 x f32> arithmetic in a scalar f32 register. This is needed\r
-// as a base case when scalarizing vector arithmetic like <4 x f32>, which\r
-// eventually decomposes to scalars if the target doesn't support v4f32 or v2f32\r
-// types.\r
-// Splitting is the act of changing a computation in an invalid vector type to\r
-// be a computation in two vectors of half the size. For example, implementing\r
-// <128 x f32> operations in terms of two <64 x f32> operations.\r
-//\r
-//===----------------------------------------------------------------------===//\r
-\r
-#include "LegalizeTypes.h"\r
-#include "llvm/IR/DataLayout.h"\r
-#include "llvm/Support/ErrorHandling.h"\r
-#include "llvm/Support/raw_ostream.h"\r
-using namespace llvm;\r
-\r
-#define DEBUG_TYPE "legalize-types"\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Result Vector Scalarization: <1 x ty> -> ty.\r
-//===----------------------------------------------------------------------===//\r
-\r
-void DAGTypeLegalizer::ScalarizeVectorResult(SDNode *N, unsigned ResNo) {\r
- DEBUG(dbgs() << "Scalarize node result " << ResNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n");\r
- SDValue R = SDValue();\r
-\r
- switch (N->getOpcode()) {\r
- default:\r
-#ifndef NDEBUG\r
- dbgs() << "ScalarizeVectorResult #" << ResNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n";\r
-#endif\r
- report_fatal_error("Do not know how to scalarize the result of this "\r
- "operator!\n");\r
-\r
- case ISD::MERGE_VALUES: R = ScalarizeVecRes_MERGE_VALUES(N, ResNo);break;\r
- case ISD::BITCAST: R = ScalarizeVecRes_BITCAST(N); break;\r
- case ISD::BUILD_VECTOR: R = ScalarizeVecRes_BUILD_VECTOR(N); break;\r
- case ISD::CONVERT_RNDSAT: R = ScalarizeVecRes_CONVERT_RNDSAT(N); break;\r
- case ISD::EXTRACT_SUBVECTOR: R = ScalarizeVecRes_EXTRACT_SUBVECTOR(N); break;\r
- case ISD::FP_ROUND: R = ScalarizeVecRes_FP_ROUND(N); break;\r
- case ISD::FP_ROUND_INREG: R = ScalarizeVecRes_InregOp(N); break;\r
- case ISD::FPOWI: R = ScalarizeVecRes_FPOWI(N); break;\r
- case ISD::INSERT_VECTOR_ELT: R = ScalarizeVecRes_INSERT_VECTOR_ELT(N); break;\r
- case ISD::LOAD: R = ScalarizeVecRes_LOAD(cast<LoadSDNode>(N));break;\r
- case ISD::SCALAR_TO_VECTOR: R = ScalarizeVecRes_SCALAR_TO_VECTOR(N); break;\r
- case ISD::SIGN_EXTEND_INREG: R = ScalarizeVecRes_InregOp(N); break;\r
- case ISD::VSELECT: R = ScalarizeVecRes_VSELECT(N); break;\r
- case ISD::SELECT: R = ScalarizeVecRes_SELECT(N); break;\r
- case ISD::SELECT_CC: R = ScalarizeVecRes_SELECT_CC(N); break;\r
- case ISD::SETCC: R = ScalarizeVecRes_SETCC(N); break;\r
- case ISD::UNDEF: R = ScalarizeVecRes_UNDEF(N); break;\r
- case ISD::VECTOR_SHUFFLE: R = ScalarizeVecRes_VECTOR_SHUFFLE(N); break;\r
- case ISD::ANY_EXTEND:\r
- case ISD::BSWAP:\r
- case ISD::CTLZ:\r
- case ISD::CTLZ_ZERO_UNDEF:\r
- case ISD::CTPOP:\r
- case ISD::CTTZ:\r
- case ISD::CTTZ_ZERO_UNDEF:\r
- case ISD::FABS:\r
- case ISD::FCEIL:\r
- case ISD::FCOS:\r
- case ISD::FEXP:\r
- case ISD::FEXP2:\r
- case ISD::FFLOOR:\r
- case ISD::FLOG:\r
- case ISD::FLOG10:\r
- case ISD::FLOG2:\r
- case ISD::FNEARBYINT:\r
- case ISD::FNEG:\r
- case ISD::FP_EXTEND:\r
- case ISD::FP_TO_SINT:\r
- case ISD::FP_TO_UINT:\r
- case ISD::FRINT:\r
- case ISD::FROUND:\r
- case ISD::FSIN:\r
- case ISD::FSQRT:\r
- case ISD::FTRUNC:\r
- case ISD::SIGN_EXTEND:\r
- case ISD::SINT_TO_FP:\r
- case ISD::TRUNCATE:\r
- case ISD::UINT_TO_FP:\r
- case ISD::ZERO_EXTEND:\r
- R = ScalarizeVecRes_UnaryOp(N);\r
- break;\r
-\r
- case ISD::ADD:\r
- case ISD::AND:\r
- case ISD::FADD:\r
- case ISD::FCOPYSIGN:\r
- case ISD::FDIV:\r
- case ISD::FMUL:\r
- case ISD::FMINNUM:\r
- case ISD::FMAXNUM:\r
-\r
- case ISD::FPOW:\r
- case ISD::FREM:\r
- case ISD::FSUB:\r
- case ISD::MUL:\r
- case ISD::OR:\r
- case ISD::SDIV:\r
- case ISD::SREM:\r
- case ISD::SUB:\r
- case ISD::UDIV:\r
- case ISD::UREM:\r
- case ISD::XOR:\r
- case ISD::SHL:\r
- case ISD::SRA:\r
- case ISD::SRL:\r
- R = ScalarizeVecRes_BinOp(N);\r
- break;\r
- case ISD::FMA:\r
- R = ScalarizeVecRes_TernaryOp(N);\r
- break;\r
- }\r
-\r
- // If R is null, the sub-method took care of registering the result.\r
- if (R.getNode())\r
- SetScalarizedVector(SDValue(N, ResNo), R);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_BinOp(SDNode *N) {\r
- SDValue LHS = GetScalarizedVector(N->getOperand(0));\r
- SDValue RHS = GetScalarizedVector(N->getOperand(1));\r
- return DAG.getNode(N->getOpcode(), SDLoc(N),\r
- LHS.getValueType(), LHS, RHS);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_TernaryOp(SDNode *N) {\r
- SDValue Op0 = GetScalarizedVector(N->getOperand(0));\r
- SDValue Op1 = GetScalarizedVector(N->getOperand(1));\r
- SDValue Op2 = GetScalarizedVector(N->getOperand(2));\r
- return DAG.getNode(N->getOpcode(), SDLoc(N),\r
- Op0.getValueType(), Op0, Op1, Op2);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_MERGE_VALUES(SDNode *N,\r
- unsigned ResNo) {\r
- SDValue Op = DisintegrateMERGE_VALUES(N, ResNo);\r
- return GetScalarizedVector(Op);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_BITCAST(SDNode *N) {\r
- EVT NewVT = N->getValueType(0).getVectorElementType();\r
- return DAG.getNode(ISD::BITCAST, SDLoc(N),\r
- NewVT, N->getOperand(0));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_BUILD_VECTOR(SDNode *N) {\r
- EVT EltVT = N->getValueType(0).getVectorElementType();\r
- SDValue InOp = N->getOperand(0);\r
- // The BUILD_VECTOR operands may be of wider element types and\r
- // we may need to truncate them back to the requested return type.\r
- if (EltVT.isInteger())\r
- return DAG.getNode(ISD::TRUNCATE, SDLoc(N), EltVT, InOp);\r
- return InOp;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_CONVERT_RNDSAT(SDNode *N) {\r
- EVT NewVT = N->getValueType(0).getVectorElementType();\r
- SDValue Op0 = GetScalarizedVector(N->getOperand(0));\r
- return DAG.getConvertRndSat(NewVT, SDLoc(N),\r
- Op0, DAG.getValueType(NewVT),\r
- DAG.getValueType(Op0.getValueType()),\r
- N->getOperand(3),\r
- N->getOperand(4),\r
- cast<CvtRndSatSDNode>(N)->getCvtCode());\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_EXTRACT_SUBVECTOR(SDNode *N) {\r
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),\r
- N->getValueType(0).getVectorElementType(),\r
- N->getOperand(0), N->getOperand(1));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_FP_ROUND(SDNode *N) {\r
- EVT NewVT = N->getValueType(0).getVectorElementType();\r
- SDValue Op = GetScalarizedVector(N->getOperand(0));\r
- return DAG.getNode(ISD::FP_ROUND, SDLoc(N),\r
- NewVT, Op, N->getOperand(1));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_FPOWI(SDNode *N) {\r
- SDValue Op = GetScalarizedVector(N->getOperand(0));\r
- return DAG.getNode(ISD::FPOWI, SDLoc(N),\r
- Op.getValueType(), Op, N->getOperand(1));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_INSERT_VECTOR_ELT(SDNode *N) {\r
- // The value to insert may have a wider type than the vector element type,\r
- // so be sure to truncate it to the element type if necessary.\r
- SDValue Op = N->getOperand(1);\r
- EVT EltVT = N->getValueType(0).getVectorElementType();\r
- if (Op.getValueType() != EltVT)\r
- // FIXME: Can this happen for floating point types?\r
- Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), EltVT, Op);\r
- return Op;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_LOAD(LoadSDNode *N) {\r
- assert(N->isUnindexed() && "Indexed vector load?");\r
-\r
- SDValue Result = DAG.getLoad(ISD::UNINDEXED,\r
- N->getExtensionType(),\r
- N->getValueType(0).getVectorElementType(),\r
- SDLoc(N),\r
- N->getChain(), N->getBasePtr(),\r
- DAG.getUNDEF(N->getBasePtr().getValueType()),\r
- N->getPointerInfo(),\r
- N->getMemoryVT().getVectorElementType(),\r
- N->isVolatile(), N->isNonTemporal(),\r
- N->isInvariant(), N->getOriginalAlignment(),\r
- N->getAAInfo());\r
-\r
- // Legalized the chain result - switch anything that used the old chain to\r
- // use the new one.\r
- ReplaceValueWith(SDValue(N, 1), Result.getValue(1));\r
- return Result;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_UnaryOp(SDNode *N) {\r
- // Get the dest type - it doesn't always match the input type, e.g. int_to_fp.\r
- EVT DestVT = N->getValueType(0).getVectorElementType();\r
- SDValue Op = N->getOperand(0);\r
- EVT OpVT = Op.getValueType();\r
- SDLoc DL(N);\r
- // The result needs scalarizing, but it's not a given that the source does.\r
- // This is a workaround for targets where it's impossible to scalarize the\r
- // result of a conversion, because the source type is legal.\r
- // For instance, this happens on AArch64: v1i1 is illegal but v1i{8,16,32}\r
- // are widened to v8i8, v4i16, and v2i32, which is legal, because v1i64 is\r
- // legal and was not scalarized.\r
- // See the similar logic in ScalarizeVecRes_VSETCC\r
- if (getTypeAction(OpVT) == TargetLowering::TypeScalarizeVector) {\r
- Op = GetScalarizedVector(Op);\r
- } else {\r
- EVT VT = OpVT.getVectorElementType();\r
- Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- }\r
- return DAG.getNode(N->getOpcode(), SDLoc(N), DestVT, Op);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_InregOp(SDNode *N) {\r
- EVT EltVT = N->getValueType(0).getVectorElementType();\r
- EVT ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT().getVectorElementType();\r
- SDValue LHS = GetScalarizedVector(N->getOperand(0));\r
- return DAG.getNode(N->getOpcode(), SDLoc(N), EltVT,\r
- LHS, DAG.getValueType(ExtVT));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_SCALAR_TO_VECTOR(SDNode *N) {\r
- // If the operand is wider than the vector element type then it is implicitly\r
- // truncated. Make that explicit here.\r
- EVT EltVT = N->getValueType(0).getVectorElementType();\r
- SDValue InOp = N->getOperand(0);\r
- if (InOp.getValueType() != EltVT)\r
- return DAG.getNode(ISD::TRUNCATE, SDLoc(N), EltVT, InOp);\r
- return InOp;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_VSELECT(SDNode *N) {\r
- SDValue Cond = GetScalarizedVector(N->getOperand(0));\r
- SDValue LHS = GetScalarizedVector(N->getOperand(1));\r
- TargetLowering::BooleanContent ScalarBool =\r
- TLI.getBooleanContents(false, false);\r
- TargetLowering::BooleanContent VecBool = TLI.getBooleanContents(true, false);\r
-\r
- // If integer and float booleans have different contents then we can't\r
- // reliably optimize in all cases. There is a full explanation for this in\r
- // DAGCombiner::visitSELECT() where the same issue affects folding\r
- // (select C, 0, 1) to (xor C, 1).\r
- if (TLI.getBooleanContents(false, false) !=\r
- TLI.getBooleanContents(false, true)) {\r
- // At least try the common case where the boolean is generated by a\r
- // comparison.\r
- if (Cond->getOpcode() == ISD::SETCC) {\r
- EVT OpVT = Cond->getOperand(0)->getValueType(0);\r
- ScalarBool = TLI.getBooleanContents(OpVT.getScalarType());\r
- VecBool = TLI.getBooleanContents(OpVT);\r
- } else\r
- ScalarBool = TargetLowering::UndefinedBooleanContent;\r
- }\r
-\r
- if (ScalarBool != VecBool) {\r
- EVT CondVT = Cond.getValueType();\r
- switch (ScalarBool) {\r
- case TargetLowering::UndefinedBooleanContent:\r
- break;\r
- case TargetLowering::ZeroOrOneBooleanContent:\r
- assert(VecBool == TargetLowering::UndefinedBooleanContent ||\r
- VecBool == TargetLowering::ZeroOrNegativeOneBooleanContent);\r
- // Vector read from all ones, scalar expects a single 1 so mask.\r
- Cond = DAG.getNode(ISD::AND, SDLoc(N), CondVT,\r
- Cond, DAG.getConstant(1, CondVT));\r
- break;\r
- case TargetLowering::ZeroOrNegativeOneBooleanContent:\r
- assert(VecBool == TargetLowering::UndefinedBooleanContent ||\r
- VecBool == TargetLowering::ZeroOrOneBooleanContent);\r
- // Vector reads from a one, scalar from all ones so sign extend.\r
- Cond = DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), CondVT,\r
- Cond, DAG.getValueType(MVT::i1));\r
- break;\r
- }\r
- }\r
-\r
- return DAG.getSelect(SDLoc(N),\r
- LHS.getValueType(), Cond, LHS,\r
- GetScalarizedVector(N->getOperand(2)));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT(SDNode *N) {\r
- SDValue LHS = GetScalarizedVector(N->getOperand(1));\r
- return DAG.getSelect(SDLoc(N),\r
- LHS.getValueType(), N->getOperand(0), LHS,\r
- GetScalarizedVector(N->getOperand(2)));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT_CC(SDNode *N) {\r
- SDValue LHS = GetScalarizedVector(N->getOperand(2));\r
- return DAG.getNode(ISD::SELECT_CC, SDLoc(N), LHS.getValueType(),\r
- N->getOperand(0), N->getOperand(1),\r
- LHS, GetScalarizedVector(N->getOperand(3)),\r
- N->getOperand(4));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_SETCC(SDNode *N) {\r
- assert(N->getValueType(0).isVector() ==\r
- N->getOperand(0).getValueType().isVector() &&\r
- "Scalar/Vector type mismatch");\r
-\r
- if (N->getValueType(0).isVector()) return ScalarizeVecRes_VSETCC(N);\r
-\r
- SDValue LHS = GetScalarizedVector(N->getOperand(0));\r
- SDValue RHS = GetScalarizedVector(N->getOperand(1));\r
- SDLoc DL(N);\r
-\r
- // Turn it into a scalar SETCC.\r
- return DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, N->getOperand(2));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_UNDEF(SDNode *N) {\r
- return DAG.getUNDEF(N->getValueType(0).getVectorElementType());\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_VECTOR_SHUFFLE(SDNode *N) {\r
- // Figure out if the scalar is the LHS or RHS and return it.\r
- SDValue Arg = N->getOperand(2).getOperand(0);\r
- if (Arg.getOpcode() == ISD::UNDEF)\r
- return DAG.getUNDEF(N->getValueType(0).getVectorElementType());\r
- unsigned Op = !cast<ConstantSDNode>(Arg)->isNullValue();\r
- return GetScalarizedVector(N->getOperand(Op));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::ScalarizeVecRes_VSETCC(SDNode *N) {\r
- assert(N->getValueType(0).isVector() &&\r
- N->getOperand(0).getValueType().isVector() &&\r
- "Operand types must be vectors");\r
- SDValue LHS = N->getOperand(0);\r
- SDValue RHS = N->getOperand(1);\r
- EVT OpVT = LHS.getValueType();\r
- EVT NVT = N->getValueType(0).getVectorElementType();\r
- SDLoc DL(N);\r
-\r
- // The result needs scalarizing, but it's not a given that the source does.\r
- if (getTypeAction(OpVT) == TargetLowering::TypeScalarizeVector) {\r
- LHS = GetScalarizedVector(LHS);\r
- RHS = GetScalarizedVector(RHS);\r
- } else {\r
- EVT VT = OpVT.getVectorElementType();\r
- LHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, LHS,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- RHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, RHS,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- }\r
-\r
- // Turn it into a scalar SETCC.\r
- SDValue Res = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS,\r
- N->getOperand(2));\r
- // Vectors may have a different boolean contents to scalars. Promote the\r
- // value appropriately.\r
- ISD::NodeType ExtendCode =\r
- TargetLowering::getExtendForContent(TLI.getBooleanContents(OpVT));\r
- return DAG.getNode(ExtendCode, DL, NVT, Res);\r
-}\r
-\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Operand Vector Scalarization <1 x ty> -> ty.\r
-//===----------------------------------------------------------------------===//\r
-\r
-bool DAGTypeLegalizer::ScalarizeVectorOperand(SDNode *N, unsigned OpNo) {\r
- DEBUG(dbgs() << "Scalarize node operand " << OpNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n");\r
- SDValue Res = SDValue();\r
-\r
- if (!Res.getNode()) {\r
- switch (N->getOpcode()) {\r
- default:\r
-#ifndef NDEBUG\r
- dbgs() << "ScalarizeVectorOperand Op #" << OpNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n";\r
-#endif\r
- llvm_unreachable("Do not know how to scalarize this operator's operand!");\r
- case ISD::BITCAST:\r
- Res = ScalarizeVecOp_BITCAST(N);\r
- break;\r
- case ISD::ANY_EXTEND:\r
- case ISD::ZERO_EXTEND:\r
- case ISD::SIGN_EXTEND:\r
- case ISD::TRUNCATE:\r
- case ISD::FP_TO_SINT:\r
- case ISD::FP_TO_UINT:\r
- case ISD::SINT_TO_FP:\r
- case ISD::UINT_TO_FP:\r
- Res = ScalarizeVecOp_UnaryOp(N);\r
- break;\r
- case ISD::CONCAT_VECTORS:\r
- Res = ScalarizeVecOp_CONCAT_VECTORS(N);\r
- break;\r
- case ISD::EXTRACT_VECTOR_ELT:\r
- Res = ScalarizeVecOp_EXTRACT_VECTOR_ELT(N);\r
- break;\r
- case ISD::VSELECT:\r
- Res = ScalarizeVecOp_VSELECT(N);\r
- break;\r
- case ISD::STORE:\r
- Res = ScalarizeVecOp_STORE(cast<StoreSDNode>(N), OpNo);\r
- break;\r
- case ISD::FP_ROUND:\r
- Res = ScalarizeVecOp_FP_ROUND(N, OpNo);\r
- break;\r
- }\r
- }\r
-\r
- // If the result is null, the sub-method took care of registering results etc.\r
- if (!Res.getNode()) return false;\r
-\r
- // If the result is N, the sub-method updated N in place. Tell the legalizer\r
- // core about this.\r
- if (Res.getNode() == N)\r
- return true;\r
-\r
- assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&\r
- "Invalid operand expansion");\r
-\r
- ReplaceValueWith(SDValue(N, 0), Res);\r
- return false;\r
-}\r
-\r
-/// ScalarizeVecOp_BITCAST - If the value to convert is a vector that needs\r
-/// to be scalarized, it must be <1 x ty>. Convert the element instead.\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_BITCAST(SDNode *N) {\r
- SDValue Elt = GetScalarizedVector(N->getOperand(0));\r
- return DAG.getNode(ISD::BITCAST, SDLoc(N),\r
- N->getValueType(0), Elt);\r
-}\r
-\r
-/// ScalarizeVecOp_UnaryOp - If the input is a vector that needs to be\r
-/// scalarized, it must be <1 x ty>. Do the operation on the element instead.\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_UnaryOp(SDNode *N) {\r
- assert(N->getValueType(0).getVectorNumElements() == 1 &&\r
- "Unexpected vector type!");\r
- SDValue Elt = GetScalarizedVector(N->getOperand(0));\r
- SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N),\r
- N->getValueType(0).getScalarType(), Elt);\r
- // Revectorize the result so the types line up with what the uses of this\r
- // expression expect.\r
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N->getValueType(0), Op);\r
-}\r
-\r
-/// ScalarizeVecOp_CONCAT_VECTORS - The vectors to concatenate have length one -\r
-/// use a BUILD_VECTOR instead.\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_CONCAT_VECTORS(SDNode *N) {\r
- SmallVector<SDValue, 8> Ops(N->getNumOperands());\r
- for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)\r
- Ops[i] = GetScalarizedVector(N->getOperand(i));\r
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N->getValueType(0), Ops);\r
-}\r
-\r
-/// ScalarizeVecOp_EXTRACT_VECTOR_ELT - If the input is a vector that needs to\r
-/// be scalarized, it must be <1 x ty>, so just return the element, ignoring the\r
-/// index.\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_EXTRACT_VECTOR_ELT(SDNode *N) {\r
- SDValue Res = GetScalarizedVector(N->getOperand(0));\r
- if (Res.getValueType() != N->getValueType(0))\r
- Res = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), N->getValueType(0),\r
- Res);\r
- return Res;\r
-}\r
-\r
-\r
-/// ScalarizeVecOp_VSELECT - If the input condition is a vector that needs to be\r
-/// scalarized, it must be <1 x i1>, so just convert to a normal ISD::SELECT\r
-/// (still with vector output type since that was acceptable if we got here).\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_VSELECT(SDNode *N) {\r
- SDValue ScalarCond = GetScalarizedVector(N->getOperand(0));\r
- EVT VT = N->getValueType(0);\r
-\r
- return DAG.getNode(ISD::SELECT, SDLoc(N), VT, ScalarCond, N->getOperand(1),\r
- N->getOperand(2));\r
-}\r
-\r
-/// ScalarizeVecOp_STORE - If the value to store is a vector that needs to be\r
-/// scalarized, it must be <1 x ty>. Just store the element.\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_STORE(StoreSDNode *N, unsigned OpNo){\r
- assert(N->isUnindexed() && "Indexed store of one-element vector?");\r
- assert(OpNo == 1 && "Do not know how to scalarize this operand!");\r
- SDLoc dl(N);\r
-\r
- if (N->isTruncatingStore())\r
- return DAG.getTruncStore(N->getChain(), dl,\r
- GetScalarizedVector(N->getOperand(1)),\r
- N->getBasePtr(), N->getPointerInfo(),\r
- N->getMemoryVT().getVectorElementType(),\r
- N->isVolatile(), N->isNonTemporal(),\r
- N->getAlignment(), N->getAAInfo());\r
-\r
- return DAG.getStore(N->getChain(), dl, GetScalarizedVector(N->getOperand(1)),\r
- N->getBasePtr(), N->getPointerInfo(),\r
- N->isVolatile(), N->isNonTemporal(),\r
- N->getOriginalAlignment(), N->getAAInfo());\r
-}\r
-\r
-/// ScalarizeVecOp_FP_ROUND - If the value to round is a vector that needs\r
-/// to be scalarized, it must be <1 x ty>. Convert the element instead.\r
-SDValue DAGTypeLegalizer::ScalarizeVecOp_FP_ROUND(SDNode *N, unsigned OpNo) {\r
- SDValue Elt = GetScalarizedVector(N->getOperand(0));\r
- SDValue Res = DAG.getNode(ISD::FP_ROUND, SDLoc(N),\r
- N->getValueType(0).getVectorElementType(), Elt,\r
- N->getOperand(1));\r
- return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), N->getValueType(0), Res);\r
-}\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Result Vector Splitting\r
-//===----------------------------------------------------------------------===//\r
-\r
-/// SplitVectorResult - This method is called when the specified result of the\r
-/// specified node is found to need vector splitting. At this point, the node\r
-/// may also have invalid operands or may have other results that need\r
-/// legalization, we just know that (at least) one result needs vector\r
-/// splitting.\r
-void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) {\r
- DEBUG(dbgs() << "Split node result: ";\r
- N->dump(&DAG);\r
- dbgs() << "\n");\r
- SDValue Lo, Hi;\r
-\r
- // See if the target wants to custom expand this node.\r
- if (CustomLowerNode(N, N->getValueType(ResNo), true))\r
- return;\r
-\r
- switch (N->getOpcode()) {\r
- default:\r
-#ifndef NDEBUG\r
- dbgs() << "SplitVectorResult #" << ResNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n";\r
-#endif\r
- report_fatal_error("Do not know how to split the result of this "\r
- "operator!\n");\r
-\r
- case ISD::MERGE_VALUES: SplitRes_MERGE_VALUES(N, ResNo, Lo, Hi); break;\r
- case ISD::VSELECT:\r
- case ISD::SELECT: SplitRes_SELECT(N, Lo, Hi); break;\r
- case ISD::SELECT_CC: SplitRes_SELECT_CC(N, Lo, Hi); break;\r
- case ISD::UNDEF: SplitRes_UNDEF(N, Lo, Hi); break;\r
- case ISD::BITCAST: SplitVecRes_BITCAST(N, Lo, Hi); break;\r
- case ISD::BUILD_VECTOR: SplitVecRes_BUILD_VECTOR(N, Lo, Hi); break;\r
- case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break;\r
- case ISD::EXTRACT_SUBVECTOR: SplitVecRes_EXTRACT_SUBVECTOR(N, Lo, Hi); break;\r
- case ISD::INSERT_SUBVECTOR: SplitVecRes_INSERT_SUBVECTOR(N, Lo, Hi); break;\r
- case ISD::FP_ROUND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;\r
- case ISD::FPOWI: SplitVecRes_FPOWI(N, Lo, Hi); break;\r
- case ISD::INSERT_VECTOR_ELT: SplitVecRes_INSERT_VECTOR_ELT(N, Lo, Hi); break;\r
- case ISD::SCALAR_TO_VECTOR: SplitVecRes_SCALAR_TO_VECTOR(N, Lo, Hi); break;\r
- case ISD::SIGN_EXTEND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;\r
- case ISD::LOAD:\r
- SplitVecRes_LOAD(cast<LoadSDNode>(N), Lo, Hi);\r
- break;\r
- case ISD::MLOAD:\r
- SplitVecRes_MLOAD(cast<MaskedLoadSDNode>(N), Lo, Hi);\r
- break;\r
- case ISD::SETCC:\r
- SplitVecRes_SETCC(N, Lo, Hi);\r
- break;\r
- case ISD::VECTOR_SHUFFLE:\r
- SplitVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N), Lo, Hi);\r
- break;\r
-\r
- case ISD::BSWAP:\r
- case ISD::CONVERT_RNDSAT:\r
- case ISD::CTLZ:\r
- case ISD::CTTZ:\r
- case ISD::CTLZ_ZERO_UNDEF:\r
- case ISD::CTTZ_ZERO_UNDEF:\r
- case ISD::CTPOP:\r
- case ISD::FABS:\r
- case ISD::FCEIL:\r
- case ISD::FCOS:\r
- case ISD::FEXP:\r
- case ISD::FEXP2:\r
- case ISD::FFLOOR:\r
- case ISD::FLOG:\r
- case ISD::FLOG10:\r
- case ISD::FLOG2:\r
- case ISD::FNEARBYINT:\r
- case ISD::FNEG:\r
- case ISD::FP_EXTEND:\r
- case ISD::FP_ROUND:\r
- case ISD::FP_TO_SINT:\r
- case ISD::FP_TO_UINT:\r
- case ISD::FRINT:\r
- case ISD::FROUND:\r
- case ISD::FSIN:\r
- case ISD::FSQRT:\r
- case ISD::FTRUNC:\r
- case ISD::SINT_TO_FP:\r
- case ISD::TRUNCATE:\r
- case ISD::UINT_TO_FP:\r
- SplitVecRes_UnaryOp(N, Lo, Hi);\r
- break;\r
-\r
- case ISD::ANY_EXTEND:\r
- case ISD::SIGN_EXTEND:\r
- case ISD::ZERO_EXTEND:\r
- SplitVecRes_ExtendOp(N, Lo, Hi);\r
- break;\r
-\r
- case ISD::ADD:\r
- case ISD::SUB:\r
- case ISD::MUL:\r
- case ISD::FADD:\r
- case ISD::FCOPYSIGN:\r
- case ISD::FSUB:\r
- case ISD::FMUL:\r
- case ISD::FMINNUM:\r
- case ISD::FMAXNUM:\r
- case ISD::SDIV:\r
- case ISD::UDIV:\r
- case ISD::FDIV:\r
- case ISD::FPOW:\r
- case ISD::AND:\r
- case ISD::OR:\r
- case ISD::XOR:\r
- case ISD::SHL:\r
- case ISD::SRA:\r
- case ISD::SRL:\r
- case ISD::UREM:\r
- case ISD::SREM:\r
- case ISD::FREM:\r
- SplitVecRes_BinOp(N, Lo, Hi);\r
- break;\r
- case ISD::FMA:\r
- SplitVecRes_TernaryOp(N, Lo, Hi);\r
- break;\r
- }\r
-\r
- // If Lo/Hi is null, the sub-method took care of registering results etc.\r
- if (Lo.getNode())\r
- SetSplitVector(SDValue(N, ResNo), Lo, Hi);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_BinOp(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDValue LHSLo, LHSHi;\r
- GetSplitVector(N->getOperand(0), LHSLo, LHSHi);\r
- SDValue RHSLo, RHSHi;\r
- GetSplitVector(N->getOperand(1), RHSLo, RHSHi);\r
- SDLoc dl(N);\r
-\r
- Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo, RHSLo);\r
- Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi, RHSHi);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_TernaryOp(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDValue Op0Lo, Op0Hi;\r
- GetSplitVector(N->getOperand(0), Op0Lo, Op0Hi);\r
- SDValue Op1Lo, Op1Hi;\r
- GetSplitVector(N->getOperand(1), Op1Lo, Op1Hi);\r
- SDValue Op2Lo, Op2Hi;\r
- GetSplitVector(N->getOperand(2), Op2Lo, Op2Hi);\r
- SDLoc dl(N);\r
-\r
- Lo = DAG.getNode(N->getOpcode(), dl, Op0Lo.getValueType(),\r
- Op0Lo, Op1Lo, Op2Lo);\r
- Hi = DAG.getNode(N->getOpcode(), dl, Op0Hi.getValueType(),\r
- Op0Hi, Op1Hi, Op2Hi);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_BITCAST(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- // We know the result is a vector. The input may be either a vector or a\r
- // scalar value.\r
- EVT LoVT, HiVT;\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
- SDLoc dl(N);\r
-\r
- SDValue InOp = N->getOperand(0);\r
- EVT InVT = InOp.getValueType();\r
-\r
- // Handle some special cases efficiently.\r
- switch (getTypeAction(InVT)) {\r
- case TargetLowering::TypeLegal:\r
- case TargetLowering::TypePromoteInteger:\r
- case TargetLowering::TypeSoftenFloat:\r
- case TargetLowering::TypeScalarizeVector:\r
- case TargetLowering::TypeWidenVector:\r
- break;\r
- case TargetLowering::TypeExpandInteger:\r
- case TargetLowering::TypeExpandFloat:\r
- // A scalar to vector conversion, where the scalar needs expansion.\r
- // If the vector is being split in two then we can just convert the\r
- // expanded pieces.\r
- if (LoVT == HiVT) {\r
- GetExpandedOp(InOp, Lo, Hi);\r
- if (TLI.isBigEndian())\r
- std::swap(Lo, Hi);\r
- Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);\r
- Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);\r
- return;\r
- }\r
- break;\r
- case TargetLowering::TypeSplitVector:\r
- // If the input is a vector that needs to be split, convert each split\r
- // piece of the input now.\r
- GetSplitVector(InOp, Lo, Hi);\r
- Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);\r
- Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);\r
- return;\r
- }\r
-\r
- // In the general case, convert the input to an integer and split it by hand.\r
- EVT LoIntVT = EVT::getIntegerVT(*DAG.getContext(), LoVT.getSizeInBits());\r
- EVT HiIntVT = EVT::getIntegerVT(*DAG.getContext(), HiVT.getSizeInBits());\r
- if (TLI.isBigEndian())\r
- std::swap(LoIntVT, HiIntVT);\r
-\r
- SplitInteger(BitConvertToInteger(InOp), LoIntVT, HiIntVT, Lo, Hi);\r
-\r
- if (TLI.isBigEndian())\r
- std::swap(Lo, Hi);\r
- Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);\r
- Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_BUILD_VECTOR(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- EVT LoVT, HiVT;\r
- SDLoc dl(N);\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
- unsigned LoNumElts = LoVT.getVectorNumElements();\r
- SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+LoNumElts);\r
- Lo = DAG.getNode(ISD::BUILD_VECTOR, dl, LoVT, LoOps);\r
-\r
- SmallVector<SDValue, 8> HiOps(N->op_begin()+LoNumElts, N->op_end());\r
- Hi = DAG.getNode(ISD::BUILD_VECTOR, dl, HiVT, HiOps);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_CONCAT_VECTORS(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- assert(!(N->getNumOperands() & 1) && "Unsupported CONCAT_VECTORS");\r
- SDLoc dl(N);\r
- unsigned NumSubvectors = N->getNumOperands() / 2;\r
- if (NumSubvectors == 1) {\r
- Lo = N->getOperand(0);\r
- Hi = N->getOperand(1);\r
- return;\r
- }\r
-\r
- EVT LoVT, HiVT;\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
-\r
- SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+NumSubvectors);\r
- Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, LoVT, LoOps);\r
-\r
- SmallVector<SDValue, 8> HiOps(N->op_begin()+NumSubvectors, N->op_end());\r
- Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HiVT, HiOps);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_EXTRACT_SUBVECTOR(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDValue Vec = N->getOperand(0);\r
- SDValue Idx = N->getOperand(1);\r
- SDLoc dl(N);\r
-\r
- EVT LoVT, HiVT;\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
-\r
- Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, LoVT, Vec, Idx);\r
- uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();\r
- Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, HiVT, Vec,\r
- DAG.getConstant(IdxVal + LoVT.getVectorNumElements(),\r
- TLI.getVectorIdxTy()));\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_INSERT_SUBVECTOR(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDValue Vec = N->getOperand(0);\r
- SDValue SubVec = N->getOperand(1);\r
- SDValue Idx = N->getOperand(2);\r
- SDLoc dl(N);\r
- GetSplitVector(Vec, Lo, Hi);\r
-\r
- // Spill the vector to the stack.\r
- EVT VecVT = Vec.getValueType();\r
- EVT SubVecVT = VecVT.getVectorElementType();\r
- SDValue StackPtr = DAG.CreateStackTemporary(VecVT);\r
- SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,\r
- MachinePointerInfo(), false, false, 0);\r
-\r
- // Store the new subvector into the specified index.\r
- SDValue SubVecPtr = GetVectorElementPointer(StackPtr, SubVecVT, Idx);\r
- Type *VecType = VecVT.getTypeForEVT(*DAG.getContext());\r
- unsigned Alignment = TLI.getDataLayout()->getPrefTypeAlignment(VecType);\r
- Store = DAG.getStore(Store, dl, SubVec, SubVecPtr, MachinePointerInfo(),\r
- false, false, 0);\r
-\r
- // Load the Lo part from the stack slot.\r
- Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),\r
- false, false, false, 0);\r
-\r
- // Increment the pointer to the other part.\r
- unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8;\r
- StackPtr =\r
- DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,\r
- DAG.getConstant(IncrementSize, StackPtr.getValueType()));\r
-\r
- // Load the Hi part from the stack slot.\r
- Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),\r
- false, false, false, MinAlign(Alignment, IncrementSize));\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_FPOWI(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDLoc dl(N);\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
- Lo = DAG.getNode(ISD::FPOWI, dl, Lo.getValueType(), Lo, N->getOperand(1));\r
- Hi = DAG.getNode(ISD::FPOWI, dl, Hi.getValueType(), Hi, N->getOperand(1));\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_InregOp(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDValue LHSLo, LHSHi;\r
- GetSplitVector(N->getOperand(0), LHSLo, LHSHi);\r
- SDLoc dl(N);\r
-\r
- EVT LoVT, HiVT;\r
- std::tie(LoVT, HiVT) =\r
- DAG.GetSplitDestVTs(cast<VTSDNode>(N->getOperand(1))->getVT());\r
-\r
- Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo,\r
- DAG.getValueType(LoVT));\r
- Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi,\r
- DAG.getValueType(HiVT));\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDValue Vec = N->getOperand(0);\r
- SDValue Elt = N->getOperand(1);\r
- SDValue Idx = N->getOperand(2);\r
- SDLoc dl(N);\r
- GetSplitVector(Vec, Lo, Hi);\r
-\r
- if (ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx)) {\r
- unsigned IdxVal = CIdx->getZExtValue();\r
- unsigned LoNumElts = Lo.getValueType().getVectorNumElements();\r
- if (IdxVal < LoNumElts)\r
- Lo = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,\r
- Lo.getValueType(), Lo, Elt, Idx);\r
- else\r
- Hi = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, Hi.getValueType(), Hi, Elt,\r
- DAG.getConstant(IdxVal - LoNumElts,\r
- TLI.getVectorIdxTy()));\r
- return;\r
- }\r
-\r
- // See if the target wants to custom expand this node.\r
- if (CustomLowerNode(N, N->getValueType(0), true))\r
- return;\r
-\r
- // Spill the vector to the stack.\r
- EVT VecVT = Vec.getValueType();\r
- EVT EltVT = VecVT.getVectorElementType();\r
- SDValue StackPtr = DAG.CreateStackTemporary(VecVT);\r
- SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,\r
- MachinePointerInfo(), false, false, 0);\r
-\r
- // Store the new element. This may be larger than the vector element type,\r
- // so use a truncating store.\r
- SDValue EltPtr = GetVectorElementPointer(StackPtr, EltVT, Idx);\r
- Type *VecType = VecVT.getTypeForEVT(*DAG.getContext());\r
- unsigned Alignment =\r
- TLI.getDataLayout()->getPrefTypeAlignment(VecType);\r
- Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, MachinePointerInfo(), EltVT,\r
- false, false, 0);\r
-\r
- // Load the Lo part from the stack slot.\r
- Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),\r
- false, false, false, 0);\r
-\r
- // Increment the pointer to the other part.\r
- unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8;\r
- StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,\r
- DAG.getConstant(IncrementSize, StackPtr.getValueType()));\r
-\r
- // Load the Hi part from the stack slot.\r
- Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),\r
- false, false, false, MinAlign(Alignment, IncrementSize));\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_SCALAR_TO_VECTOR(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- EVT LoVT, HiVT;\r
- SDLoc dl(N);\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
- Lo = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoVT, N->getOperand(0));\r
- Hi = DAG.getUNDEF(HiVT);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_LOAD(LoadSDNode *LD, SDValue &Lo,\r
- SDValue &Hi) {\r
- assert(ISD::isUNINDEXEDLoad(LD) && "Indexed load during type legalization!");\r
- EVT LoVT, HiVT;\r
- SDLoc dl(LD);\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(LD->getValueType(0));\r
-\r
- ISD::LoadExtType ExtType = LD->getExtensionType();\r
- SDValue Ch = LD->getChain();\r
- SDValue Ptr = LD->getBasePtr();\r
- SDValue Offset = DAG.getUNDEF(Ptr.getValueType());\r
- EVT MemoryVT = LD->getMemoryVT();\r
- unsigned Alignment = LD->getOriginalAlignment();\r
- bool isVolatile = LD->isVolatile();\r
- bool isNonTemporal = LD->isNonTemporal();\r
- bool isInvariant = LD->isInvariant();\r
- AAMDNodes AAInfo = LD->getAAInfo();\r
-\r
- EVT LoMemVT, HiMemVT;\r
- std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);\r
-\r
- Lo = DAG.getLoad(ISD::UNINDEXED, ExtType, LoVT, dl, Ch, Ptr, Offset,\r
- LD->getPointerInfo(), LoMemVT, isVolatile, isNonTemporal,\r
- isInvariant, Alignment, AAInfo);\r
-\r
- unsigned IncrementSize = LoMemVT.getSizeInBits()/8;\r
- Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,\r
- DAG.getConstant(IncrementSize, Ptr.getValueType()));\r
- Hi = DAG.getLoad(ISD::UNINDEXED, ExtType, HiVT, dl, Ch, Ptr, Offset,\r
- LD->getPointerInfo().getWithOffset(IncrementSize),\r
- HiMemVT, isVolatile, isNonTemporal, isInvariant, Alignment,\r
- AAInfo);\r
-\r
- // Build a factor node to remember that this load is independent of the\r
- // other one.\r
- Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),\r
- Hi.getValue(1));\r
-\r
- // Legalized the chain result - switch anything that used the old chain to\r
- // use the new one.\r
- ReplaceValueWith(SDValue(LD, 1), Ch);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_MLOAD(MaskedLoadSDNode *MLD,\r
- SDValue &Lo, SDValue &Hi) {\r
- EVT LoVT, HiVT;\r
- SDLoc dl(MLD);\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));\r
-\r
- SDValue Ch = MLD->getChain();\r
- SDValue Ptr = MLD->getBasePtr();\r
- SDValue Mask = MLD->getMask();\r
- unsigned Alignment = MLD->getOriginalAlignment();\r
-\r
- // if Alignment is equal to the vector size,\r
- // take the half of it for the second part\r
- unsigned SecondHalfAlignment =\r
- (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?\r
- Alignment/2 : Alignment;\r
-\r
- SDValue MaskLo, MaskHi;\r
- std::tie(MaskLo, MaskHi) = DAG.SplitVector(Mask, dl);\r
-\r
- EVT MemoryVT = MLD->getMemoryVT();\r
- EVT LoMemVT, HiMemVT;\r
- std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);\r
-\r
- SDValue Src0 = MLD->getSrc0();\r
- SDValue Src0Lo, Src0Hi;\r
- std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, dl);\r
-\r
- MachineMemOperand *MMO = DAG.getMachineFunction().\r
- getMachineMemOperand(MLD->getPointerInfo(), \r
- MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),\r
- Alignment, MLD->getAAInfo(), MLD->getRanges());\r
-\r
- Lo = DAG.getMaskedLoad(LoVT, dl, Ch, Ptr, MaskLo, Src0Lo, MMO);\r
-\r
- unsigned IncrementSize = LoMemVT.getSizeInBits()/8;\r
- Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,\r
- DAG.getConstant(IncrementSize, Ptr.getValueType()));\r
-\r
- MMO = DAG.getMachineFunction().\r
- getMachineMemOperand(MLD->getPointerInfo(), \r
- MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),\r
- SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());\r
-\r
- Hi = DAG.getMaskedLoad(HiVT, dl, Ch, Ptr, MaskHi, Src0Hi, MMO);\r
-\r
-\r
- // Build a factor node to remember that this load is independent of the\r
- // other one.\r
- Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),\r
- Hi.getValue(1));\r
-\r
- // Legalized the chain result - switch anything that used the old chain to\r
- // use the new one.\r
- ReplaceValueWith(SDValue(MLD, 1), Ch);\r
-\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_SETCC(SDNode *N, SDValue &Lo, SDValue &Hi) {\r
- assert(N->getValueType(0).isVector() &&\r
- N->getOperand(0).getValueType().isVector() &&\r
- "Operand types must be vectors");\r
-\r
- EVT LoVT, HiVT;\r
- SDLoc DL(N);\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
-\r
- // Split the input.\r
- SDValue LL, LH, RL, RH;\r
- std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);\r
- std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);\r
-\r
- Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));\r
- Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_UnaryOp(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- // Get the dest types - they may not match the input types, e.g. int_to_fp.\r
- EVT LoVT, HiVT;\r
- SDLoc dl(N);\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));\r
-\r
- // If the input also splits, handle it directly for a compile time speedup.\r
- // Otherwise split it by hand.\r
- EVT InVT = N->getOperand(0).getValueType();\r
- if (getTypeAction(InVT) == TargetLowering::TypeSplitVector)\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
- else\r
- std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0);\r
-\r
- if (N->getOpcode() == ISD::FP_ROUND) {\r
- Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo, N->getOperand(1));\r
- Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi, N->getOperand(1));\r
- } else if (N->getOpcode() == ISD::CONVERT_RNDSAT) {\r
- SDValue DTyOpLo = DAG.getValueType(LoVT);\r
- SDValue DTyOpHi = DAG.getValueType(HiVT);\r
- SDValue STyOpLo = DAG.getValueType(Lo.getValueType());\r
- SDValue STyOpHi = DAG.getValueType(Hi.getValueType());\r
- SDValue RndOp = N->getOperand(3);\r
- SDValue SatOp = N->getOperand(4);\r
- ISD::CvtCode CvtCode = cast<CvtRndSatSDNode>(N)->getCvtCode();\r
- Lo = DAG.getConvertRndSat(LoVT, dl, Lo, DTyOpLo, STyOpLo, RndOp, SatOp,\r
- CvtCode);\r
- Hi = DAG.getConvertRndSat(HiVT, dl, Hi, DTyOpHi, STyOpHi, RndOp, SatOp,\r
- CvtCode);\r
- } else {\r
- Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo);\r
- Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi);\r
- }\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_ExtendOp(SDNode *N, SDValue &Lo,\r
- SDValue &Hi) {\r
- SDLoc dl(N);\r
- EVT SrcVT = N->getOperand(0).getValueType();\r
- EVT DestVT = N->getValueType(0);\r
- EVT LoVT, HiVT;\r
- std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(DestVT);\r
-\r
- // We can do better than a generic split operation if the extend is doing\r
- // more than just doubling the width of the elements and the following are\r
- // true:\r
- // - The number of vector elements is even,\r
- // - the source type is legal,\r
- // - the type of a split source is illegal,\r
- // - the type of an extended (by doubling element size) source is legal, and\r
- // - the type of that extended source when split is legal.\r
- //\r
- // This won't necessarily completely legalize the operation, but it will\r
- // more effectively move in the right direction and prevent falling down\r
- // to scalarization in many cases due to the input vector being split too\r
- // far.\r
- unsigned NumElements = SrcVT.getVectorNumElements();\r
- if ((NumElements & 1) == 0 &&\r
- SrcVT.getSizeInBits() * 2 < DestVT.getSizeInBits()) {\r
- LLVMContext &Ctx = *DAG.getContext();\r
- EVT NewSrcVT = EVT::getVectorVT(\r
- Ctx, EVT::getIntegerVT(\r
- Ctx, SrcVT.getVectorElementType().getSizeInBits() * 2),\r
- NumElements);\r
- EVT SplitSrcVT =\r
- EVT::getVectorVT(Ctx, SrcVT.getVectorElementType(), NumElements / 2);\r
- EVT SplitLoVT, SplitHiVT;\r
- std::tie(SplitLoVT, SplitHiVT) = DAG.GetSplitDestVTs(NewSrcVT);\r
- if (TLI.isTypeLegal(SrcVT) && !TLI.isTypeLegal(SplitSrcVT) &&\r
- TLI.isTypeLegal(NewSrcVT) && TLI.isTypeLegal(SplitLoVT)) {\r
- DEBUG(dbgs() << "Split vector extend via incremental extend:";\r
- N->dump(&DAG); dbgs() << "\n");\r
- // Extend the source vector by one step.\r
- SDValue NewSrc =\r
- DAG.getNode(N->getOpcode(), dl, NewSrcVT, N->getOperand(0));\r
- // Get the low and high halves of the new, extended one step, vector.\r
- std::tie(Lo, Hi) = DAG.SplitVector(NewSrc, dl);\r
- // Extend those vector halves the rest of the way.\r
- Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo);\r
- Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi);\r
- return;\r
- }\r
- }\r
- // Fall back to the generic unary operator splitting otherwise.\r
- SplitVecRes_UnaryOp(N, Lo, Hi);\r
-}\r
-\r
-void DAGTypeLegalizer::SplitVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N,\r
- SDValue &Lo, SDValue &Hi) {\r
- // The low and high parts of the original input give four input vectors.\r
- SDValue Inputs[4];\r
- SDLoc dl(N);\r
- GetSplitVector(N->getOperand(0), Inputs[0], Inputs[1]);\r
- GetSplitVector(N->getOperand(1), Inputs[2], Inputs[3]);\r
- EVT NewVT = Inputs[0].getValueType();\r
- unsigned NewElts = NewVT.getVectorNumElements();\r
-\r
- // If Lo or Hi uses elements from at most two of the four input vectors, then\r
- // express it as a vector shuffle of those two inputs. Otherwise extract the\r
- // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR.\r
- SmallVector<int, 16> Ops;\r
- for (unsigned High = 0; High < 2; ++High) {\r
- SDValue &Output = High ? Hi : Lo;\r
-\r
- // Build a shuffle mask for the output, discovering on the fly which\r
- // input vectors to use as shuffle operands (recorded in InputUsed).\r
- // If building a suitable shuffle vector proves too hard, then bail\r
- // out with useBuildVector set.\r
- unsigned InputUsed[2] = { -1U, -1U }; // Not yet discovered.\r
- unsigned FirstMaskIdx = High * NewElts;\r
- bool useBuildVector = false;\r
- for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) {\r
- // The mask element. This indexes into the input.\r
- int Idx = N->getMaskElt(FirstMaskIdx + MaskOffset);\r
-\r
- // The input vector this mask element indexes into.\r
- unsigned Input = (unsigned)Idx / NewElts;\r
-\r
- if (Input >= array_lengthof(Inputs)) {\r
- // The mask element does not index into any input vector.\r
- Ops.push_back(-1);\r
- continue;\r
- }\r
-\r
- // Turn the index into an offset from the start of the input vector.\r
- Idx -= Input * NewElts;\r
-\r
- // Find or create a shuffle vector operand to hold this input.\r
- unsigned OpNo;\r
- for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {\r
- if (InputUsed[OpNo] == Input) {\r
- // This input vector is already an operand.\r
- break;\r
- } else if (InputUsed[OpNo] == -1U) {\r
- // Create a new operand for this input vector.\r
- InputUsed[OpNo] = Input;\r
- break;\r
- }\r
- }\r
-\r
- if (OpNo >= array_lengthof(InputUsed)) {\r
- // More than two input vectors used! Give up on trying to create a\r
- // shuffle vector. Insert all elements into a BUILD_VECTOR instead.\r
- useBuildVector = true;\r
- break;\r
- }\r
-\r
- // Add the mask index for the new shuffle vector.\r
- Ops.push_back(Idx + OpNo * NewElts);\r
- }\r
-\r
- if (useBuildVector) {\r
- EVT EltVT = NewVT.getVectorElementType();\r
- SmallVector<SDValue, 16> SVOps;\r
-\r
- // Extract the input elements by hand.\r
- for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) {\r
- // The mask element. This indexes into the input.\r
- int Idx = N->getMaskElt(FirstMaskIdx + MaskOffset);\r
-\r
- // The input vector this mask element indexes into.\r
- unsigned Input = (unsigned)Idx / NewElts;\r
-\r
- if (Input >= array_lengthof(Inputs)) {\r
- // The mask element is "undef" or indexes off the end of the input.\r
- SVOps.push_back(DAG.getUNDEF(EltVT));\r
- continue;\r
- }\r
-\r
- // Turn the index into an offset from the start of the input vector.\r
- Idx -= Input * NewElts;\r
-\r
- // Extract the vector element by hand.\r
- SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,\r
- Inputs[Input], DAG.getConstant(Idx,\r
- TLI.getVectorIdxTy())));\r
- }\r
-\r
- // Construct the Lo/Hi output using a BUILD_VECTOR.\r
- Output = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, SVOps);\r
- } else if (InputUsed[0] == -1U) {\r
- // No input vectors were used! The result is undefined.\r
- Output = DAG.getUNDEF(NewVT);\r
- } else {\r
- SDValue Op0 = Inputs[InputUsed[0]];\r
- // If only one input was used, use an undefined vector for the other.\r
- SDValue Op1 = InputUsed[1] == -1U ?\r
- DAG.getUNDEF(NewVT) : Inputs[InputUsed[1]];\r
- // At least one input vector was used. Create a new shuffle vector.\r
- Output = DAG.getVectorShuffle(NewVT, dl, Op0, Op1, &Ops[0]);\r
- }\r
-\r
- Ops.clear();\r
- }\r
-}\r
-\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Operand Vector Splitting\r
-//===----------------------------------------------------------------------===//\r
-\r
-/// SplitVectorOperand - This method is called when the specified operand of the\r
-/// specified node is found to need vector splitting. At this point, all of the\r
-/// result types of the node are known to be legal, but other operands of the\r
-/// node may need legalization as well as the specified one.\r
-bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) {\r
- DEBUG(dbgs() << "Split node operand: ";\r
- N->dump(&DAG);\r
- dbgs() << "\n");\r
- SDValue Res = SDValue();\r
-\r
- // See if the target wants to custom split this node.\r
- if (CustomLowerNode(N, N->getOperand(OpNo).getValueType(), false))\r
- return false;\r
-\r
- if (!Res.getNode()) {\r
- switch (N->getOpcode()) {\r
- default:\r
-#ifndef NDEBUG\r
- dbgs() << "SplitVectorOperand Op #" << OpNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n";\r
-#endif\r
- report_fatal_error("Do not know how to split this operator's "\r
- "operand!\n");\r
-\r
- case ISD::SETCC: Res = SplitVecOp_VSETCC(N); break;\r
- case ISD::BITCAST: Res = SplitVecOp_BITCAST(N); break;\r
- case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break;\r
- case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break;\r
- case ISD::CONCAT_VECTORS: Res = SplitVecOp_CONCAT_VECTORS(N); break;\r
- case ISD::TRUNCATE: Res = SplitVecOp_TRUNCATE(N); break;\r
- case ISD::FP_ROUND: Res = SplitVecOp_FP_ROUND(N); break;\r
- case ISD::STORE:\r
- Res = SplitVecOp_STORE(cast<StoreSDNode>(N), OpNo);\r
- break;\r
- case ISD::MSTORE:\r
- Res = SplitVecOp_MSTORE(cast<MaskedStoreSDNode>(N), OpNo);\r
- break;\r
- case ISD::VSELECT:\r
- Res = SplitVecOp_VSELECT(N, OpNo);\r
- break;\r
- case ISD::CTTZ:\r
- case ISD::CTLZ:\r
- case ISD::CTPOP:\r
- case ISD::FP_EXTEND:\r
- case ISD::FP_TO_SINT:\r
- case ISD::FP_TO_UINT:\r
- case ISD::SINT_TO_FP:\r
- case ISD::UINT_TO_FP:\r
- case ISD::FTRUNC:\r
- case ISD::SIGN_EXTEND:\r
- case ISD::ZERO_EXTEND:\r
- case ISD::ANY_EXTEND:\r
- Res = SplitVecOp_UnaryOp(N);\r
- break;\r
- }\r
- }\r
-\r
- // If the result is null, the sub-method took care of registering results etc.\r
- if (!Res.getNode()) return false;\r
-\r
- // If the result is N, the sub-method updated N in place. Tell the legalizer\r
- // core about this.\r
- if (Res.getNode() == N)\r
- return true;\r
-\r
- assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&\r
- "Invalid operand expansion");\r
-\r
- ReplaceValueWith(SDValue(N, 0), Res);\r
- return false;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_VSELECT(SDNode *N, unsigned OpNo) {\r
- // The only possibility for an illegal operand is the mask, since result type\r
- // legalization would have handled this node already otherwise.\r
- assert(OpNo == 0 && "Illegal operand must be mask");\r
-\r
- SDValue Mask = N->getOperand(0);\r
- SDValue Src0 = N->getOperand(1);\r
- SDValue Src1 = N->getOperand(2);\r
- EVT Src0VT = Src0.getValueType();\r
- SDLoc DL(N);\r
- assert(Mask.getValueType().isVector() && "VSELECT without a vector mask?");\r
-\r
- SDValue Lo, Hi;\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
- assert(Lo.getValueType() == Hi.getValueType() &&\r
- "Lo and Hi have differing types");\r
-\r
- EVT LoOpVT, HiOpVT;\r
- std::tie(LoOpVT, HiOpVT) = DAG.GetSplitDestVTs(Src0VT);\r
- assert(LoOpVT == HiOpVT && "Asymmetric vector split?");\r
-\r
- SDValue LoOp0, HiOp0, LoOp1, HiOp1, LoMask, HiMask;\r
- std::tie(LoOp0, HiOp0) = DAG.SplitVector(Src0, DL);\r
- std::tie(LoOp1, HiOp1) = DAG.SplitVector(Src1, DL);\r
- std::tie(LoMask, HiMask) = DAG.SplitVector(Mask, DL);\r
-\r
- SDValue LoSelect =\r
- DAG.getNode(ISD::VSELECT, DL, LoOpVT, LoMask, LoOp0, LoOp1);\r
- SDValue HiSelect =\r
- DAG.getNode(ISD::VSELECT, DL, HiOpVT, HiMask, HiOp0, HiOp1);\r
-\r
- return DAG.getNode(ISD::CONCAT_VECTORS, DL, Src0VT, LoSelect, HiSelect);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_UnaryOp(SDNode *N) {\r
- // The result has a legal vector type, but the input needs splitting.\r
- EVT ResVT = N->getValueType(0);\r
- SDValue Lo, Hi;\r
- SDLoc dl(N);\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
- EVT InVT = Lo.getValueType();\r
-\r
- EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(),\r
- InVT.getVectorNumElements());\r
-\r
- Lo = DAG.getNode(N->getOpcode(), dl, OutVT, Lo);\r
- Hi = DAG.getNode(N->getOpcode(), dl, OutVT, Hi);\r
-\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_BITCAST(SDNode *N) {\r
- // For example, i64 = BITCAST v4i16 on alpha. Typically the vector will\r
- // end up being split all the way down to individual components. Convert the\r
- // split pieces into integers and reassemble.\r
- SDValue Lo, Hi;\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
- Lo = BitConvertToInteger(Lo);\r
- Hi = BitConvertToInteger(Hi);\r
-\r
- if (TLI.isBigEndian())\r
- std::swap(Lo, Hi);\r
-\r
- return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0),\r
- JoinIntegers(Lo, Hi));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_SUBVECTOR(SDNode *N) {\r
- // We know that the extracted result type is legal.\r
- EVT SubVT = N->getValueType(0);\r
- SDValue Idx = N->getOperand(1);\r
- SDLoc dl(N);\r
- SDValue Lo, Hi;\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
-\r
- uint64_t LoElts = Lo.getValueType().getVectorNumElements();\r
- uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();\r
-\r
- if (IdxVal < LoElts) {\r
- assert(IdxVal + SubVT.getVectorNumElements() <= LoElts &&\r
- "Extracted subvector crosses vector split!");\r
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, Lo, Idx);\r
- } else {\r
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, Hi,\r
- DAG.getConstant(IdxVal - LoElts, Idx.getValueType()));\r
- }\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N) {\r
- SDValue Vec = N->getOperand(0);\r
- SDValue Idx = N->getOperand(1);\r
- EVT VecVT = Vec.getValueType();\r
-\r
- if (isa<ConstantSDNode>(Idx)) {\r
- uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();\r
- assert(IdxVal < VecVT.getVectorNumElements() && "Invalid vector index!");\r
-\r
- SDValue Lo, Hi;\r
- GetSplitVector(Vec, Lo, Hi);\r
-\r
- uint64_t LoElts = Lo.getValueType().getVectorNumElements();\r
-\r
- if (IdxVal < LoElts)\r
- return SDValue(DAG.UpdateNodeOperands(N, Lo, Idx), 0);\r
- return SDValue(DAG.UpdateNodeOperands(N, Hi,\r
- DAG.getConstant(IdxVal - LoElts,\r
- Idx.getValueType())), 0);\r
- }\r
-\r
- // See if the target wants to custom expand this node.\r
- if (CustomLowerNode(N, N->getValueType(0), true))\r
- return SDValue();\r
-\r
- // Store the vector to the stack.\r
- EVT EltVT = VecVT.getVectorElementType();\r
- SDLoc dl(N);\r
- SDValue StackPtr = DAG.CreateStackTemporary(VecVT);\r
- SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,\r
- MachinePointerInfo(), false, false, 0);\r
-\r
- // Load back the required element.\r
- StackPtr = GetVectorElementPointer(StackPtr, EltVT, Idx);\r
- return DAG.getExtLoad(ISD::EXTLOAD, dl, N->getValueType(0), Store, StackPtr,\r
- MachinePointerInfo(), EltVT, false, false, false, 0);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_MSTORE(MaskedStoreSDNode *N,\r
- unsigned OpNo) {\r
- SDValue Ch = N->getChain();\r
- SDValue Ptr = N->getBasePtr();\r
- SDValue Mask = N->getMask();\r
- SDValue Data = N->getData();\r
- EVT MemoryVT = N->getMemoryVT();\r
- unsigned Alignment = N->getOriginalAlignment();\r
- SDLoc DL(N);\r
- \r
- EVT LoMemVT, HiMemVT;\r
- std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);\r
-\r
- SDValue DataLo, DataHi;\r
- GetSplitVector(Data, DataLo, DataHi);\r
- SDValue MaskLo, MaskHi;\r
- GetSplitVector(Mask, MaskLo, MaskHi);\r
-\r
- // if Alignment is equal to the vector size,\r
- // take the half of it for the second part\r
- unsigned SecondHalfAlignment =\r
- (Alignment == Data->getValueType(0).getSizeInBits()/8) ?\r
- Alignment/2 : Alignment;\r
-\r
- SDValue Lo, Hi;\r
- MachineMemOperand *MMO = DAG.getMachineFunction().\r
- getMachineMemOperand(N->getPointerInfo(), \r
- MachineMemOperand::MOStore, LoMemVT.getStoreSize(),\r
- Alignment, N->getAAInfo(), N->getRanges());\r
-\r
- Lo = DAG.getMaskedStore(Ch, DL, DataLo, Ptr, MaskLo, MMO);\r
-\r
- unsigned IncrementSize = LoMemVT.getSizeInBits()/8;\r
- Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,\r
- DAG.getConstant(IncrementSize, Ptr.getValueType()));\r
-\r
- MMO = DAG.getMachineFunction().\r
- getMachineMemOperand(N->getPointerInfo(), \r
- MachineMemOperand::MOStore, HiMemVT.getStoreSize(),\r
- SecondHalfAlignment, N->getAAInfo(), N->getRanges());\r
-\r
- Hi = DAG.getMaskedStore(Ch, DL, DataHi, Ptr, MaskHi, MMO);\r
-\r
-\r
- // Build a factor node to remember that this store is independent of the\r
- // other one.\r
- return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);\r
-\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) {\r
- assert(N->isUnindexed() && "Indexed store of vector?");\r
- assert(OpNo == 1 && "Can only split the stored value");\r
- SDLoc DL(N);\r
-\r
- bool isTruncating = N->isTruncatingStore();\r
- SDValue Ch = N->getChain();\r
- SDValue Ptr = N->getBasePtr();\r
- EVT MemoryVT = N->getMemoryVT();\r
- unsigned Alignment = N->getOriginalAlignment();\r
- bool isVol = N->isVolatile();\r
- bool isNT = N->isNonTemporal();\r
- AAMDNodes AAInfo = N->getAAInfo();\r
- SDValue Lo, Hi;\r
- GetSplitVector(N->getOperand(1), Lo, Hi);\r
-\r
- EVT LoMemVT, HiMemVT;\r
- std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);\r
-\r
- unsigned IncrementSize = LoMemVT.getSizeInBits()/8;\r
-\r
- if (isTruncating)\r
- Lo = DAG.getTruncStore(Ch, DL, Lo, Ptr, N->getPointerInfo(),\r
- LoMemVT, isVol, isNT, Alignment, AAInfo);\r
- else\r
- Lo = DAG.getStore(Ch, DL, Lo, Ptr, N->getPointerInfo(),\r
- isVol, isNT, Alignment, AAInfo);\r
-\r
- // Increment the pointer to the other half.\r
- Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,\r
- DAG.getConstant(IncrementSize, Ptr.getValueType()));\r
-\r
- if (isTruncating)\r
- Hi = DAG.getTruncStore(Ch, DL, Hi, Ptr,\r
- N->getPointerInfo().getWithOffset(IncrementSize),\r
- HiMemVT, isVol, isNT, Alignment, AAInfo);\r
- else\r
- Hi = DAG.getStore(Ch, DL, Hi, Ptr,\r
- N->getPointerInfo().getWithOffset(IncrementSize),\r
- isVol, isNT, Alignment, AAInfo);\r
-\r
- return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) {\r
- SDLoc DL(N);\r
-\r
- // The input operands all must have the same type, and we know the result\r
- // type is valid. Convert this to a buildvector which extracts all the\r
- // input elements.\r
- // TODO: If the input elements are power-two vectors, we could convert this to\r
- // a new CONCAT_VECTORS node with elements that are half-wide.\r
- SmallVector<SDValue, 32> Elts;\r
- EVT EltVT = N->getValueType(0).getVectorElementType();\r
- for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {\r
- SDValue Op = N->getOperand(op);\r
- for (unsigned i = 0, e = Op.getValueType().getVectorNumElements();\r
- i != e; ++i) {\r
- Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT,\r
- Op, DAG.getConstant(i, TLI.getVectorIdxTy())));\r
-\r
- }\r
- }\r
-\r
- return DAG.getNode(ISD::BUILD_VECTOR, DL, N->getValueType(0), Elts);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_TRUNCATE(SDNode *N) {\r
- // The result type is legal, but the input type is illegal. If splitting\r
- // ends up with the result type of each half still being legal, just\r
- // do that. If, however, that would result in an illegal result type,\r
- // we can try to get more clever with power-two vectors. Specifically,\r
- // split the input type, but also widen the result element size, then\r
- // concatenate the halves and truncate again. For example, consider a target\r
- // where v8i8 is legal and v8i32 is not (ARM, which doesn't have 256-bit\r
- // vectors). To perform a "%res = v8i8 trunc v8i32 %in" we do:\r
- // %inlo = v4i32 extract_subvector %in, 0\r
- // %inhi = v4i32 extract_subvector %in, 4\r
- // %lo16 = v4i16 trunc v4i32 %inlo\r
- // %hi16 = v4i16 trunc v4i32 %inhi\r
- // %in16 = v8i16 concat_vectors v4i16 %lo16, v4i16 %hi16\r
- // %res = v8i8 trunc v8i16 %in16\r
- //\r
- // Without this transform, the original truncate would end up being\r
- // scalarized, which is pretty much always a last resort.\r
- SDValue InVec = N->getOperand(0);\r
- EVT InVT = InVec->getValueType(0);\r
- EVT OutVT = N->getValueType(0);\r
- unsigned NumElements = OutVT.getVectorNumElements();\r
- // Widening should have already made sure this is a power-two vector\r
- // if we're trying to split it at all. assert() that's true, just in case.\r
- assert(!(NumElements & 1) && "Splitting vector, but not in half!");\r
-\r
- unsigned InElementSize = InVT.getVectorElementType().getSizeInBits();\r
- unsigned OutElementSize = OutVT.getVectorElementType().getSizeInBits();\r
-\r
- // If the input elements are only 1/2 the width of the result elements,\r
- // just use the normal splitting. Our trick only work if there's room\r
- // to split more than once.\r
- if (InElementSize <= OutElementSize * 2)\r
- return SplitVecOp_UnaryOp(N);\r
- SDLoc DL(N);\r
-\r
- // Extract the halves of the input via extract_subvector.\r
- SDValue InLoVec, InHiVec;\r
- std::tie(InLoVec, InHiVec) = DAG.SplitVector(InVec, DL);\r
- // Truncate them to 1/2 the element size.\r
- EVT HalfElementVT = EVT::getIntegerVT(*DAG.getContext(), InElementSize/2);\r
- EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), HalfElementVT,\r
- NumElements/2);\r
- SDValue HalfLo = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, InLoVec);\r
- SDValue HalfHi = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, InHiVec);\r
- // Concatenate them to get the full intermediate truncation result.\r
- EVT InterVT = EVT::getVectorVT(*DAG.getContext(), HalfElementVT, NumElements);\r
- SDValue InterVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, InterVT, HalfLo,\r
- HalfHi);\r
- // Now finish up by truncating all the way down to the original result\r
- // type. This should normally be something that ends up being legal directly,\r
- // but in theory if a target has very wide vectors and an annoyingly\r
- // restricted set of legal types, this split can chain to build things up.\r
- return DAG.getNode(ISD::TRUNCATE, DL, OutVT, InterVec);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {\r
- assert(N->getValueType(0).isVector() &&\r
- N->getOperand(0).getValueType().isVector() &&\r
- "Operand types must be vectors");\r
- // The result has a legal vector type, but the input needs splitting.\r
- SDValue Lo0, Hi0, Lo1, Hi1, LoRes, HiRes;\r
- SDLoc DL(N);\r
- GetSplitVector(N->getOperand(0), Lo0, Hi0);\r
- GetSplitVector(N->getOperand(1), Lo1, Hi1);\r
- unsigned PartElements = Lo0.getValueType().getVectorNumElements();\r
- EVT PartResVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, PartElements);\r
- EVT WideResVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, 2*PartElements);\r
-\r
- LoRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Lo0, Lo1, N->getOperand(2));\r
- HiRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Hi0, Hi1, N->getOperand(2));\r
- SDValue Con = DAG.getNode(ISD::CONCAT_VECTORS, DL, WideResVT, LoRes, HiRes);\r
- return PromoteTargetBoolean(Con, N->getValueType(0));\r
-}\r
-\r
-\r
-SDValue DAGTypeLegalizer::SplitVecOp_FP_ROUND(SDNode *N) {\r
- // The result has a legal vector type, but the input needs splitting.\r
- EVT ResVT = N->getValueType(0);\r
- SDValue Lo, Hi;\r
- SDLoc DL(N);\r
- GetSplitVector(N->getOperand(0), Lo, Hi);\r
- EVT InVT = Lo.getValueType();\r
-\r
- EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(),\r
- InVT.getVectorNumElements());\r
-\r
- Lo = DAG.getNode(ISD::FP_ROUND, DL, OutVT, Lo, N->getOperand(1));\r
- Hi = DAG.getNode(ISD::FP_ROUND, DL, OutVT, Hi, N->getOperand(1));\r
-\r
- return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);\r
-}\r
-\r
-\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Result Vector Widening\r
-//===----------------------------------------------------------------------===//\r
-\r
-void DAGTypeLegalizer::WidenVectorResult(SDNode *N, unsigned ResNo) {\r
- DEBUG(dbgs() << "Widen node result " << ResNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n");\r
-\r
- // See if the target wants to custom widen this node.\r
- if (CustomWidenLowerNode(N, N->getValueType(ResNo)))\r
- return;\r
-\r
- SDValue Res = SDValue();\r
- switch (N->getOpcode()) {\r
- default:\r
-#ifndef NDEBUG\r
- dbgs() << "WidenVectorResult #" << ResNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n";\r
-#endif\r
- llvm_unreachable("Do not know how to widen the result of this operator!");\r
-\r
- case ISD::MERGE_VALUES: Res = WidenVecRes_MERGE_VALUES(N, ResNo); break;\r
- case ISD::BITCAST: Res = WidenVecRes_BITCAST(N); break;\r
- case ISD::BUILD_VECTOR: Res = WidenVecRes_BUILD_VECTOR(N); break;\r
- case ISD::CONCAT_VECTORS: Res = WidenVecRes_CONCAT_VECTORS(N); break;\r
- case ISD::CONVERT_RNDSAT: Res = WidenVecRes_CONVERT_RNDSAT(N); break;\r
- case ISD::EXTRACT_SUBVECTOR: Res = WidenVecRes_EXTRACT_SUBVECTOR(N); break;\r
- case ISD::FP_ROUND_INREG: Res = WidenVecRes_InregOp(N); break;\r
- case ISD::INSERT_VECTOR_ELT: Res = WidenVecRes_INSERT_VECTOR_ELT(N); break;\r
- case ISD::LOAD: Res = WidenVecRes_LOAD(N); break;\r
- case ISD::SCALAR_TO_VECTOR: Res = WidenVecRes_SCALAR_TO_VECTOR(N); break;\r
- case ISD::SIGN_EXTEND_INREG: Res = WidenVecRes_InregOp(N); break;\r
- case ISD::VSELECT:\r
- case ISD::SELECT: Res = WidenVecRes_SELECT(N); break;\r
- case ISD::SELECT_CC: Res = WidenVecRes_SELECT_CC(N); break;\r
- case ISD::SETCC: Res = WidenVecRes_SETCC(N); break;\r
- case ISD::UNDEF: Res = WidenVecRes_UNDEF(N); break;\r
- case ISD::VECTOR_SHUFFLE:\r
- Res = WidenVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N));\r
- break;\r
-\r
- case ISD::ADD:\r
- case ISD::AND:\r
- case ISD::MUL:\r
- case ISD::MULHS:\r
- case ISD::MULHU:\r
- case ISD::OR:\r
- case ISD::SUB:\r
- case ISD::XOR:\r
- case ISD::FMINNUM:\r
- case ISD::FMAXNUM:\r
- Res = WidenVecRes_Binary(N);\r
- break;\r
-\r
- case ISD::FADD:\r
- case ISD::FCOPYSIGN:\r
- case ISD::FMUL:\r
- case ISD::FPOW:\r
- case ISD::FSUB:\r
- case ISD::FDIV:\r
- case ISD::FREM:\r
- case ISD::SDIV:\r
- case ISD::UDIV:\r
- case ISD::SREM:\r
- case ISD::UREM:\r
- Res = WidenVecRes_BinaryCanTrap(N);\r
- break;\r
-\r
- case ISD::FPOWI:\r
- Res = WidenVecRes_POWI(N);\r
- break;\r
-\r
- case ISD::SHL:\r
- case ISD::SRA:\r
- case ISD::SRL:\r
- Res = WidenVecRes_Shift(N);\r
- break;\r
-\r
- case ISD::ANY_EXTEND:\r
- case ISD::FP_EXTEND:\r
- case ISD::FP_ROUND:\r
- case ISD::FP_TO_SINT:\r
- case ISD::FP_TO_UINT:\r
- case ISD::SIGN_EXTEND:\r
- case ISD::SINT_TO_FP:\r
- case ISD::TRUNCATE:\r
- case ISD::UINT_TO_FP:\r
- case ISD::ZERO_EXTEND:\r
- Res = WidenVecRes_Convert(N);\r
- break;\r
-\r
- case ISD::BSWAP:\r
- case ISD::CTLZ:\r
- case ISD::CTPOP:\r
- case ISD::CTTZ:\r
- case ISD::FABS:\r
- case ISD::FCEIL:\r
- case ISD::FCOS:\r
- case ISD::FEXP:\r
- case ISD::FEXP2:\r
- case ISD::FFLOOR:\r
- case ISD::FLOG:\r
- case ISD::FLOG10:\r
- case ISD::FLOG2:\r
- case ISD::FNEARBYINT:\r
- case ISD::FNEG:\r
- case ISD::FRINT:\r
- case ISD::FROUND:\r
- case ISD::FSIN:\r
- case ISD::FSQRT:\r
- case ISD::FTRUNC:\r
- Res = WidenVecRes_Unary(N);\r
- break;\r
- case ISD::FMA:\r
- Res = WidenVecRes_Ternary(N);\r
- break;\r
- }\r
-\r
- // If Res is null, the sub-method took care of registering the result.\r
- if (Res.getNode())\r
- SetWidenedVector(SDValue(N, ResNo), Res);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_Ternary(SDNode *N) {\r
- // Ternary op widening.\r
- SDLoc dl(N);\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
- SDValue InOp3 = GetWidenedVector(N->getOperand(2));\r
- return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2, InOp3);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) {\r
- // Binary op widening.\r
- SDLoc dl(N);\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
- return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_BinaryCanTrap(SDNode *N) {\r
- // Binary op widening for operations that can trap.\r
- unsigned Opcode = N->getOpcode();\r
- SDLoc dl(N);\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- EVT WidenEltVT = WidenVT.getVectorElementType();\r
- EVT VT = WidenVT;\r
- unsigned NumElts = VT.getVectorNumElements();\r
- while (!TLI.isTypeLegal(VT) && NumElts != 1) {\r
- NumElts = NumElts / 2;\r
- VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts);\r
- }\r
-\r
- if (NumElts != 1 && !TLI.canOpTrap(N->getOpcode(), VT)) {\r
- // Operation doesn't trap so just widen as normal.\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
- return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2);\r
- }\r
-\r
- // No legal vector version so unroll the vector operation and then widen.\r
- if (NumElts == 1)\r
- return DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements());\r
-\r
- // Since the operation can trap, apply operation on the original vector.\r
- EVT MaxVT = VT;\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
- unsigned CurNumElts = N->getValueType(0).getVectorNumElements();\r
-\r
- SmallVector<SDValue, 16> ConcatOps(CurNumElts);\r
- unsigned ConcatEnd = 0; // Current ConcatOps index.\r
- int Idx = 0; // Current Idx into input vectors.\r
-\r
- // NumElts := greatest legal vector size (at most WidenVT)\r
- // while (orig. vector has unhandled elements) {\r
- // take munches of size NumElts from the beginning and add to ConcatOps\r
- // NumElts := next smaller supported vector size or 1\r
- // }\r
- while (CurNumElts != 0) {\r
- while (CurNumElts >= NumElts) {\r
- SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1,\r
- DAG.getConstant(Idx, TLI.getVectorIdxTy()));\r
- SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2,\r
- DAG.getConstant(Idx, TLI.getVectorIdxTy()));\r
- ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2);\r
- Idx += NumElts;\r
- CurNumElts -= NumElts;\r
- }\r
- do {\r
- NumElts = NumElts / 2;\r
- VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts);\r
- } while (!TLI.isTypeLegal(VT) && NumElts != 1);\r
-\r
- if (NumElts == 1) {\r
- for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) {\r
- SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT,\r
- InOp1, DAG.getConstant(Idx,\r
- TLI.getVectorIdxTy()));\r
- SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT,\r
- InOp2, DAG.getConstant(Idx,\r
- TLI.getVectorIdxTy()));\r
- ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, WidenEltVT,\r
- EOp1, EOp2);\r
- }\r
- CurNumElts = 0;\r
- }\r
- }\r
-\r
- // Check to see if we have a single operation with the widen type.\r
- if (ConcatEnd == 1) {\r
- VT = ConcatOps[0].getValueType();\r
- if (VT == WidenVT)\r
- return ConcatOps[0];\r
- }\r
-\r
- // while (Some element of ConcatOps is not of type MaxVT) {\r
- // From the end of ConcatOps, collect elements of the same type and put\r
- // them into an op of the next larger supported type\r
- // }\r
- while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) {\r
- Idx = ConcatEnd - 1;\r
- VT = ConcatOps[Idx--].getValueType();\r
- while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT)\r
- Idx--;\r
-\r
- int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1;\r
- EVT NextVT;\r
- do {\r
- NextSize *= 2;\r
- NextVT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NextSize);\r
- } while (!TLI.isTypeLegal(NextVT));\r
-\r
- if (!VT.isVector()) {\r
- // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT\r
- SDValue VecOp = DAG.getUNDEF(NextVT);\r
- unsigned NumToInsert = ConcatEnd - Idx - 1;\r
- for (unsigned i = 0, OpIdx = Idx+1; i < NumToInsert; i++, OpIdx++) {\r
- VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NextVT, VecOp,\r
- ConcatOps[OpIdx], DAG.getConstant(i,\r
- TLI.getVectorIdxTy()));\r
- }\r
- ConcatOps[Idx+1] = VecOp;\r
- ConcatEnd = Idx + 2;\r
- } else {\r
- // Vector type, create a CONCAT_VECTORS of type NextVT\r
- SDValue undefVec = DAG.getUNDEF(VT);\r
- unsigned OpsToConcat = NextSize/VT.getVectorNumElements();\r
- SmallVector<SDValue, 16> SubConcatOps(OpsToConcat);\r
- unsigned RealVals = ConcatEnd - Idx - 1;\r
- unsigned SubConcatEnd = 0;\r
- unsigned SubConcatIdx = Idx + 1;\r
- while (SubConcatEnd < RealVals)\r
- SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx];\r
- while (SubConcatEnd < OpsToConcat)\r
- SubConcatOps[SubConcatEnd++] = undefVec;\r
- ConcatOps[SubConcatIdx] = DAG.getNode(ISD::CONCAT_VECTORS, dl,\r
- NextVT, SubConcatOps);\r
- ConcatEnd = SubConcatIdx + 1;\r
- }\r
- }\r
-\r
- // Check to see if we have a single operation with the widen type.\r
- if (ConcatEnd == 1) {\r
- VT = ConcatOps[0].getValueType();\r
- if (VT == WidenVT)\r
- return ConcatOps[0];\r
- }\r
-\r
- // add undefs of size MaxVT until ConcatOps grows to length of WidenVT\r
- unsigned NumOps = WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements();\r
- if (NumOps != ConcatEnd ) {\r
- SDValue UndefVal = DAG.getUNDEF(MaxVT);\r
- for (unsigned j = ConcatEnd; j < NumOps; ++j)\r
- ConcatOps[j] = UndefVal;\r
- }\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT,\r
- makeArrayRef(ConcatOps.data(), NumOps));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) {\r
- SDValue InOp = N->getOperand(0);\r
- SDLoc DL(N);\r
-\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
-\r
- EVT InVT = InOp.getValueType();\r
- EVT InEltVT = InVT.getVectorElementType();\r
- EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts);\r
-\r
- unsigned Opcode = N->getOpcode();\r
- unsigned InVTNumElts = InVT.getVectorNumElements();\r
-\r
- if (getTypeAction(InVT) == TargetLowering::TypeWidenVector) {\r
- InOp = GetWidenedVector(N->getOperand(0));\r
- InVT = InOp.getValueType();\r
- InVTNumElts = InVT.getVectorNumElements();\r
- if (InVTNumElts == WidenNumElts) {\r
- if (N->getNumOperands() == 1)\r
- return DAG.getNode(Opcode, DL, WidenVT, InOp);\r
- return DAG.getNode(Opcode, DL, WidenVT, InOp, N->getOperand(1));\r
- }\r
- }\r
-\r
- if (TLI.isTypeLegal(InWidenVT)) {\r
- // Because the result and the input are different vector types, widening\r
- // the result could create a legal type but widening the input might make\r
- // it an illegal type that might lead to repeatedly splitting the input\r
- // and then widening it. To avoid this, we widen the input only if\r
- // it results in a legal type.\r
- if (WidenNumElts % InVTNumElts == 0) {\r
- // Widen the input and call convert on the widened input vector.\r
- unsigned NumConcat = WidenNumElts/InVTNumElts;\r
- SmallVector<SDValue, 16> Ops(NumConcat);\r
- Ops[0] = InOp;\r
- SDValue UndefVal = DAG.getUNDEF(InVT);\r
- for (unsigned i = 1; i != NumConcat; ++i)\r
- Ops[i] = UndefVal;\r
- SDValue InVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, InWidenVT, Ops);\r
- if (N->getNumOperands() == 1)\r
- return DAG.getNode(Opcode, DL, WidenVT, InVec);\r
- return DAG.getNode(Opcode, DL, WidenVT, InVec, N->getOperand(1));\r
- }\r
-\r
- if (InVTNumElts % WidenNumElts == 0) {\r
- SDValue InVal = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InWidenVT,\r
- InOp, DAG.getConstant(0,\r
- TLI.getVectorIdxTy()));\r
- // Extract the input and convert the shorten input vector.\r
- if (N->getNumOperands() == 1)\r
- return DAG.getNode(Opcode, DL, WidenVT, InVal);\r
- return DAG.getNode(Opcode, DL, WidenVT, InVal, N->getOperand(1));\r
- }\r
- }\r
-\r
- // Otherwise unroll into some nasty scalar code and rebuild the vector.\r
- SmallVector<SDValue, 16> Ops(WidenNumElts);\r
- EVT EltVT = WidenVT.getVectorElementType();\r
- unsigned MinElts = std::min(InVTNumElts, WidenNumElts);\r
- unsigned i;\r
- for (i=0; i < MinElts; ++i) {\r
- SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, InEltVT, InOp,\r
- DAG.getConstant(i, TLI.getVectorIdxTy()));\r
- if (N->getNumOperands() == 1)\r
- Ops[i] = DAG.getNode(Opcode, DL, EltVT, Val);\r
- else\r
- Ops[i] = DAG.getNode(Opcode, DL, EltVT, Val, N->getOperand(1));\r
- }\r
-\r
- SDValue UndefVal = DAG.getUNDEF(EltVT);\r
- for (; i < WidenNumElts; ++i)\r
- Ops[i] = UndefVal;\r
-\r
- return DAG.getNode(ISD::BUILD_VECTOR, DL, WidenVT, Ops);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_POWI(SDNode *N) {\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- SDValue ShOp = N->getOperand(1);\r
- return DAG.getNode(N->getOpcode(), SDLoc(N), WidenVT, InOp, ShOp);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_Shift(SDNode *N) {\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- SDValue ShOp = N->getOperand(1);\r
-\r
- EVT ShVT = ShOp.getValueType();\r
- if (getTypeAction(ShVT) == TargetLowering::TypeWidenVector) {\r
- ShOp = GetWidenedVector(ShOp);\r
- ShVT = ShOp.getValueType();\r
- }\r
- EVT ShWidenVT = EVT::getVectorVT(*DAG.getContext(),\r
- ShVT.getVectorElementType(),\r
- WidenVT.getVectorNumElements());\r
- if (ShVT != ShWidenVT)\r
- ShOp = ModifyToType(ShOp, ShWidenVT);\r
-\r
- return DAG.getNode(N->getOpcode(), SDLoc(N), WidenVT, InOp, ShOp);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_Unary(SDNode *N) {\r
- // Unary op widening.\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- return DAG.getNode(N->getOpcode(), SDLoc(N), WidenVT, InOp);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_InregOp(SDNode *N) {\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- EVT ExtVT = EVT::getVectorVT(*DAG.getContext(),\r
- cast<VTSDNode>(N->getOperand(1))->getVT()\r
- .getVectorElementType(),\r
- WidenVT.getVectorNumElements());\r
- SDValue WidenLHS = GetWidenedVector(N->getOperand(0));\r
- return DAG.getNode(N->getOpcode(), SDLoc(N),\r
- WidenVT, WidenLHS, DAG.getValueType(ExtVT));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_MERGE_VALUES(SDNode *N, unsigned ResNo) {\r
- SDValue WidenVec = DisintegrateMERGE_VALUES(N, ResNo);\r
- return GetWidenedVector(WidenVec);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_BITCAST(SDNode *N) {\r
- SDValue InOp = N->getOperand(0);\r
- EVT InVT = InOp.getValueType();\r
- EVT VT = N->getValueType(0);\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);\r
- SDLoc dl(N);\r
-\r
- switch (getTypeAction(InVT)) {\r
- case TargetLowering::TypeLegal:\r
- break;\r
- case TargetLowering::TypePromoteInteger:\r
- // If the incoming type is a vector that is being promoted, then\r
- // we know that the elements are arranged differently and that we\r
- // must perform the conversion using a stack slot.\r
- if (InVT.isVector())\r
- break;\r
-\r
- // If the InOp is promoted to the same size, convert it. Otherwise,\r
- // fall out of the switch and widen the promoted input.\r
- InOp = GetPromotedInteger(InOp);\r
- InVT = InOp.getValueType();\r
- if (WidenVT.bitsEq(InVT))\r
- return DAG.getNode(ISD::BITCAST, dl, WidenVT, InOp);\r
- break;\r
- case TargetLowering::TypeSoftenFloat:\r
- case TargetLowering::TypeExpandInteger:\r
- case TargetLowering::TypeExpandFloat:\r
- case TargetLowering::TypeScalarizeVector:\r
- case TargetLowering::TypeSplitVector:\r
- break;\r
- case TargetLowering::TypeWidenVector:\r
- // If the InOp is widened to the same size, convert it. Otherwise, fall\r
- // out of the switch and widen the widened input.\r
- InOp = GetWidenedVector(InOp);\r
- InVT = InOp.getValueType();\r
- if (WidenVT.bitsEq(InVT))\r
- // The input widens to the same size. Convert to the widen value.\r
- return DAG.getNode(ISD::BITCAST, dl, WidenVT, InOp);\r
- break;\r
- }\r
-\r
- unsigned WidenSize = WidenVT.getSizeInBits();\r
- unsigned InSize = InVT.getSizeInBits();\r
- // x86mmx is not an acceptable vector element type, so don't try.\r
- if (WidenSize % InSize == 0 && InVT != MVT::x86mmx) {\r
- // Determine new input vector type. The new input vector type will use\r
- // the same element type (if its a vector) or use the input type as a\r
- // vector. It is the same size as the type to widen to.\r
- EVT NewInVT;\r
- unsigned NewNumElts = WidenSize / InSize;\r
- if (InVT.isVector()) {\r
- EVT InEltVT = InVT.getVectorElementType();\r
- NewInVT = EVT::getVectorVT(*DAG.getContext(), InEltVT,\r
- WidenSize / InEltVT.getSizeInBits());\r
- } else {\r
- NewInVT = EVT::getVectorVT(*DAG.getContext(), InVT, NewNumElts);\r
- }\r
-\r
- if (TLI.isTypeLegal(NewInVT)) {\r
- // Because the result and the input are different vector types, widening\r
- // the result could create a legal type but widening the input might make\r
- // it an illegal type that might lead to repeatedly splitting the input\r
- // and then widening it. To avoid this, we widen the input only if\r
- // it results in a legal type.\r
- SmallVector<SDValue, 16> Ops(NewNumElts);\r
- SDValue UndefVal = DAG.getUNDEF(InVT);\r
- Ops[0] = InOp;\r
- for (unsigned i = 1; i < NewNumElts; ++i)\r
- Ops[i] = UndefVal;\r
-\r
- SDValue NewVec;\r
- if (InVT.isVector())\r
- NewVec = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewInVT, Ops);\r
- else\r
- NewVec = DAG.getNode(ISD::BUILD_VECTOR, dl, NewInVT, Ops);\r
- return DAG.getNode(ISD::BITCAST, dl, WidenVT, NewVec);\r
- }\r
- }\r
-\r
- return CreateStackStoreLoad(InOp, WidenVT);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_BUILD_VECTOR(SDNode *N) {\r
- SDLoc dl(N);\r
- // Build a vector with undefined for the new nodes.\r
- EVT VT = N->getValueType(0);\r
-\r
- // Integer BUILD_VECTOR operands may be larger than the node's vector element\r
- // type. The UNDEFs need to have the same type as the existing operands.\r
- EVT EltVT = N->getOperand(0).getValueType();\r
- unsigned NumElts = VT.getVectorNumElements();\r
-\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
-\r
- SmallVector<SDValue, 16> NewOps(N->op_begin(), N->op_end());\r
- assert(WidenNumElts >= NumElts && "Shrinking vector instead of widening!");\r
- NewOps.append(WidenNumElts - NumElts, DAG.getUNDEF(EltVT));\r
-\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, NewOps);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_CONCAT_VECTORS(SDNode *N) {\r
- EVT InVT = N->getOperand(0).getValueType();\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDLoc dl(N);\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
- unsigned NumInElts = InVT.getVectorNumElements();\r
- unsigned NumOperands = N->getNumOperands();\r
-\r
- bool InputWidened = false; // Indicates we need to widen the input.\r
- if (getTypeAction(InVT) != TargetLowering::TypeWidenVector) {\r
- if (WidenVT.getVectorNumElements() % InVT.getVectorNumElements() == 0) {\r
- // Add undef vectors to widen to correct length.\r
- unsigned NumConcat = WidenVT.getVectorNumElements() /\r
- InVT.getVectorNumElements();\r
- SDValue UndefVal = DAG.getUNDEF(InVT);\r
- SmallVector<SDValue, 16> Ops(NumConcat);\r
- for (unsigned i=0; i < NumOperands; ++i)\r
- Ops[i] = N->getOperand(i);\r
- for (unsigned i = NumOperands; i != NumConcat; ++i)\r
- Ops[i] = UndefVal;\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, Ops);\r
- }\r
- } else {\r
- InputWidened = true;\r
- if (WidenVT == TLI.getTypeToTransformTo(*DAG.getContext(), InVT)) {\r
- // The inputs and the result are widen to the same value.\r
- unsigned i;\r
- for (i=1; i < NumOperands; ++i)\r
- if (N->getOperand(i).getOpcode() != ISD::UNDEF)\r
- break;\r
-\r
- if (i == NumOperands)\r
- // Everything but the first operand is an UNDEF so just return the\r
- // widened first operand.\r
- return GetWidenedVector(N->getOperand(0));\r
-\r
- if (NumOperands == 2) {\r
- // Replace concat of two operands with a shuffle.\r
- SmallVector<int, 16> MaskOps(WidenNumElts, -1);\r
- for (unsigned i = 0; i < NumInElts; ++i) {\r
- MaskOps[i] = i;\r
- MaskOps[i + NumInElts] = i + WidenNumElts;\r
- }\r
- return DAG.getVectorShuffle(WidenVT, dl,\r
- GetWidenedVector(N->getOperand(0)),\r
- GetWidenedVector(N->getOperand(1)),\r
- &MaskOps[0]);\r
- }\r
- }\r
- }\r
-\r
- // Fall back to use extracts and build vector.\r
- EVT EltVT = WidenVT.getVectorElementType();\r
- SmallVector<SDValue, 16> Ops(WidenNumElts);\r
- unsigned Idx = 0;\r
- for (unsigned i=0; i < NumOperands; ++i) {\r
- SDValue InOp = N->getOperand(i);\r
- if (InputWidened)\r
- InOp = GetWidenedVector(InOp);\r
- for (unsigned j=0; j < NumInElts; ++j)\r
- Ops[Idx++] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,\r
- DAG.getConstant(j, TLI.getVectorIdxTy()));\r
- }\r
- SDValue UndefVal = DAG.getUNDEF(EltVT);\r
- for (; Idx < WidenNumElts; ++Idx)\r
- Ops[Idx] = UndefVal;\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_CONVERT_RNDSAT(SDNode *N) {\r
- SDLoc dl(N);\r
- SDValue InOp = N->getOperand(0);\r
- SDValue RndOp = N->getOperand(3);\r
- SDValue SatOp = N->getOperand(4);\r
-\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
-\r
- EVT InVT = InOp.getValueType();\r
- EVT InEltVT = InVT.getVectorElementType();\r
- EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts);\r
-\r
- SDValue DTyOp = DAG.getValueType(WidenVT);\r
- SDValue STyOp = DAG.getValueType(InWidenVT);\r
- ISD::CvtCode CvtCode = cast<CvtRndSatSDNode>(N)->getCvtCode();\r
-\r
- unsigned InVTNumElts = InVT.getVectorNumElements();\r
- if (getTypeAction(InVT) == TargetLowering::TypeWidenVector) {\r
- InOp = GetWidenedVector(InOp);\r
- InVT = InOp.getValueType();\r
- InVTNumElts = InVT.getVectorNumElements();\r
- if (InVTNumElts == WidenNumElts)\r
- return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp,\r
- SatOp, CvtCode);\r
- }\r
-\r
- if (TLI.isTypeLegal(InWidenVT)) {\r
- // Because the result and the input are different vector types, widening\r
- // the result could create a legal type but widening the input might make\r
- // it an illegal type that might lead to repeatedly splitting the input\r
- // and then widening it. To avoid this, we widen the input only if\r
- // it results in a legal type.\r
- if (WidenNumElts % InVTNumElts == 0) {\r
- // Widen the input and call convert on the widened input vector.\r
- unsigned NumConcat = WidenNumElts/InVTNumElts;\r
- SmallVector<SDValue, 16> Ops(NumConcat);\r
- Ops[0] = InOp;\r
- SDValue UndefVal = DAG.getUNDEF(InVT);\r
- for (unsigned i = 1; i != NumConcat; ++i)\r
- Ops[i] = UndefVal;\r
-\r
- InOp = DAG.getNode(ISD::CONCAT_VECTORS, dl, InWidenVT, Ops);\r
- return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp,\r
- SatOp, CvtCode);\r
- }\r
-\r
- if (InVTNumElts % WidenNumElts == 0) {\r
- // Extract the input and convert the shorten input vector.\r
- InOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InWidenVT, InOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp,\r
- SatOp, CvtCode);\r
- }\r
- }\r
-\r
- // Otherwise unroll into some nasty scalar code and rebuild the vector.\r
- SmallVector<SDValue, 16> Ops(WidenNumElts);\r
- EVT EltVT = WidenVT.getVectorElementType();\r
- DTyOp = DAG.getValueType(EltVT);\r
- STyOp = DAG.getValueType(InEltVT);\r
-\r
- unsigned MinElts = std::min(InVTNumElts, WidenNumElts);\r
- unsigned i;\r
- for (i=0; i < MinElts; ++i) {\r
- SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp,\r
- DAG.getConstant(i, TLI.getVectorIdxTy()));\r
- Ops[i] = DAG.getConvertRndSat(WidenVT, dl, ExtVal, DTyOp, STyOp, RndOp,\r
- SatOp, CvtCode);\r
- }\r
-\r
- SDValue UndefVal = DAG.getUNDEF(EltVT);\r
- for (; i < WidenNumElts; ++i)\r
- Ops[i] = UndefVal;\r
-\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_EXTRACT_SUBVECTOR(SDNode *N) {\r
- EVT VT = N->getValueType(0);\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
- SDValue InOp = N->getOperand(0);\r
- SDValue Idx = N->getOperand(1);\r
- SDLoc dl(N);\r
-\r
- if (getTypeAction(InOp.getValueType()) == TargetLowering::TypeWidenVector)\r
- InOp = GetWidenedVector(InOp);\r
-\r
- EVT InVT = InOp.getValueType();\r
-\r
- // Check if we can just return the input vector after widening.\r
- uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();\r
- if (IdxVal == 0 && InVT == WidenVT)\r
- return InOp;\r
-\r
- // Check if we can extract from the vector.\r
- unsigned InNumElts = InVT.getVectorNumElements();\r
- if (IdxVal % WidenNumElts == 0 && IdxVal + WidenNumElts < InNumElts)\r
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, WidenVT, InOp, Idx);\r
-\r
- // We could try widening the input to the right length but for now, extract\r
- // the original elements, fill the rest with undefs and build a vector.\r
- SmallVector<SDValue, 16> Ops(WidenNumElts);\r
- EVT EltVT = VT.getVectorElementType();\r
- unsigned NumElts = VT.getVectorNumElements();\r
- unsigned i;\r
- for (i=0; i < NumElts; ++i)\r
- Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,\r
- DAG.getConstant(IdxVal+i, TLI.getVectorIdxTy()));\r
-\r
- SDValue UndefVal = DAG.getUNDEF(EltVT);\r
- for (; i < WidenNumElts; ++i)\r
- Ops[i] = UndefVal;\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_INSERT_VECTOR_ELT(SDNode *N) {\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N),\r
- InOp.getValueType(), InOp,\r
- N->getOperand(1), N->getOperand(2));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_LOAD(SDNode *N) {\r
- LoadSDNode *LD = cast<LoadSDNode>(N);\r
- ISD::LoadExtType ExtType = LD->getExtensionType();\r
-\r
- SDValue Result;\r
- SmallVector<SDValue, 16> LdChain; // Chain for the series of load\r
- if (ExtType != ISD::NON_EXTLOAD)\r
- Result = GenWidenVectorExtLoads(LdChain, LD, ExtType);\r
- else\r
- Result = GenWidenVectorLoads(LdChain, LD);\r
-\r
- // If we generate a single load, we can use that for the chain. Otherwise,\r
- // build a factor node to remember the multiple loads are independent and\r
- // chain to that.\r
- SDValue NewChain;\r
- if (LdChain.size() == 1)\r
- NewChain = LdChain[0];\r
- else\r
- NewChain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other, LdChain);\r
-\r
- // Modified the chain - switch anything that used the old chain to use\r
- // the new one.\r
- ReplaceValueWith(SDValue(N, 1), NewChain);\r
-\r
- return Result;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_SCALAR_TO_VECTOR(SDNode *N) {\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N),\r
- WidenVT, N->getOperand(0));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_SELECT(SDNode *N) {\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
-\r
- SDValue Cond1 = N->getOperand(0);\r
- EVT CondVT = Cond1.getValueType();\r
- if (CondVT.isVector()) {\r
- EVT CondEltVT = CondVT.getVectorElementType();\r
- EVT CondWidenVT = EVT::getVectorVT(*DAG.getContext(),\r
- CondEltVT, WidenNumElts);\r
- if (getTypeAction(CondVT) == TargetLowering::TypeWidenVector)\r
- Cond1 = GetWidenedVector(Cond1);\r
-\r
- // If we have to split the condition there is no point in widening the\r
- // select. This would result in an cycle of widening the select ->\r
- // widening the condition operand -> splitting the condition operand ->\r
- // splitting the select -> widening the select. Instead split this select\r
- // further and widen the resulting type.\r
- if (getTypeAction(CondVT) == TargetLowering::TypeSplitVector) {\r
- SDValue SplitSelect = SplitVecOp_VSELECT(N, 0);\r
- SDValue Res = ModifyToType(SplitSelect, WidenVT);\r
- return Res;\r
- }\r
-\r
- if (Cond1.getValueType() != CondWidenVT)\r
- Cond1 = ModifyToType(Cond1, CondWidenVT);\r
- }\r
-\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(1));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(2));\r
- assert(InOp1.getValueType() == WidenVT && InOp2.getValueType() == WidenVT);\r
- return DAG.getNode(N->getOpcode(), SDLoc(N),\r
- WidenVT, Cond1, InOp1, InOp2);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_SELECT_CC(SDNode *N) {\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(2));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(3));\r
- return DAG.getNode(ISD::SELECT_CC, SDLoc(N),\r
- InOp1.getValueType(), N->getOperand(0),\r
- N->getOperand(1), InOp1, InOp2, N->getOperand(4));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_SETCC(SDNode *N) {\r
- assert(N->getValueType(0).isVector() ==\r
- N->getOperand(0).getValueType().isVector() &&\r
- "Scalar/Vector type mismatch");\r
- if (N->getValueType(0).isVector()) return WidenVecRes_VSETCC(N);\r
-\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
- return DAG.getNode(ISD::SETCC, SDLoc(N), WidenVT,\r
- InOp1, InOp2, N->getOperand(2));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_UNDEF(SDNode *N) {\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- return DAG.getUNDEF(WidenVT);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N) {\r
- EVT VT = N->getValueType(0);\r
- SDLoc dl(N);\r
-\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);\r
- unsigned NumElts = VT.getVectorNumElements();\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
-\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
-\r
- // Adjust mask based on new input vector length.\r
- SmallVector<int, 16> NewMask;\r
- for (unsigned i = 0; i != NumElts; ++i) {\r
- int Idx = N->getMaskElt(i);\r
- if (Idx < (int)NumElts)\r
- NewMask.push_back(Idx);\r
- else\r
- NewMask.push_back(Idx - NumElts + WidenNumElts);\r
- }\r
- for (unsigned i = NumElts; i != WidenNumElts; ++i)\r
- NewMask.push_back(-1);\r
- return DAG.getVectorShuffle(WidenVT, dl, InOp1, InOp2, &NewMask[0]);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecRes_VSETCC(SDNode *N) {\r
- assert(N->getValueType(0).isVector() &&\r
- N->getOperand(0).getValueType().isVector() &&\r
- "Operands must be vectors");\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
-\r
- SDValue InOp1 = N->getOperand(0);\r
- EVT InVT = InOp1.getValueType();\r
- assert(InVT.isVector() && "can not widen non-vector type");\r
- EVT WidenInVT = EVT::getVectorVT(*DAG.getContext(),\r
- InVT.getVectorElementType(), WidenNumElts);\r
- InOp1 = GetWidenedVector(InOp1);\r
- SDValue InOp2 = GetWidenedVector(N->getOperand(1));\r
-\r
- // Assume that the input and output will be widen appropriately. If not,\r
- // we will have to unroll it at some point.\r
- assert(InOp1.getValueType() == WidenInVT &&\r
- InOp2.getValueType() == WidenInVT &&\r
- "Input not widened to expected type!");\r
- (void)WidenInVT;\r
- return DAG.getNode(ISD::SETCC, SDLoc(N),\r
- WidenVT, InOp1, InOp2, N->getOperand(2));\r
-}\r
-\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Widen Vector Operand\r
-//===----------------------------------------------------------------------===//\r
-bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned OpNo) {\r
- DEBUG(dbgs() << "Widen node operand " << OpNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n");\r
- SDValue Res = SDValue();\r
-\r
- // See if the target wants to custom widen this node.\r
- if (CustomLowerNode(N, N->getOperand(OpNo).getValueType(), false))\r
- return false;\r
-\r
- switch (N->getOpcode()) {\r
- default:\r
-#ifndef NDEBUG\r
- dbgs() << "WidenVectorOperand op #" << OpNo << ": ";\r
- N->dump(&DAG);\r
- dbgs() << "\n";\r
-#endif\r
- llvm_unreachable("Do not know how to widen this operator's operand!");\r
-\r
- case ISD::BITCAST: Res = WidenVecOp_BITCAST(N); break;\r
- case ISD::CONCAT_VECTORS: Res = WidenVecOp_CONCAT_VECTORS(N); break;\r
- case ISD::EXTRACT_SUBVECTOR: Res = WidenVecOp_EXTRACT_SUBVECTOR(N); break;\r
- case ISD::EXTRACT_VECTOR_ELT: Res = WidenVecOp_EXTRACT_VECTOR_ELT(N); break;\r
- case ISD::STORE: Res = WidenVecOp_STORE(N); break;\r
- case ISD::SETCC: Res = WidenVecOp_SETCC(N); break;\r
-\r
- case ISD::ANY_EXTEND:\r
- case ISD::SIGN_EXTEND:\r
- case ISD::ZERO_EXTEND:\r
- Res = WidenVecOp_EXTEND(N);\r
- break;\r
-\r
- case ISD::FP_EXTEND:\r
- case ISD::FP_TO_SINT:\r
- case ISD::FP_TO_UINT:\r
- case ISD::SINT_TO_FP:\r
- case ISD::UINT_TO_FP:\r
- case ISD::TRUNCATE:\r
- Res = WidenVecOp_Convert(N);\r
- break;\r
- }\r
-\r
- // If Res is null, the sub-method took care of registering the result.\r
- if (!Res.getNode()) return false;\r
-\r
- // If the result is N, the sub-method updated N in place. Tell the legalizer\r
- // core about this.\r
- if (Res.getNode() == N)\r
- return true;\r
-\r
-\r
- assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&\r
- "Invalid operand expansion");\r
-\r
- ReplaceValueWith(SDValue(N, 0), Res);\r
- return false;\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_EXTEND(SDNode *N) {\r
- SDLoc DL(N);\r
- EVT VT = N->getValueType(0);\r
-\r
- SDValue InOp = N->getOperand(0);\r
- // If some legalization strategy other than widening is used on the operand,\r
- // we can't safely assume that just extending the low lanes is the correct\r
- // transformation.\r
- if (getTypeAction(InOp.getValueType()) != TargetLowering::TypeWidenVector)\r
- return WidenVecOp_Convert(N);\r
- InOp = GetWidenedVector(InOp);\r
- assert(VT.getVectorNumElements() <\r
- InOp.getValueType().getVectorNumElements() &&\r
- "Input wasn't widened!");\r
-\r
- // We may need to further widen the operand until it has the same total\r
- // vector size as the result.\r
- EVT InVT = InOp.getValueType();\r
- if (InVT.getSizeInBits() != VT.getSizeInBits()) {\r
- EVT InEltVT = InVT.getVectorElementType();\r
- for (int i = MVT::FIRST_VECTOR_VALUETYPE, e = MVT::LAST_VECTOR_VALUETYPE; i < e; ++i) {\r
- EVT FixedVT = (MVT::SimpleValueType)i;\r
- EVT FixedEltVT = FixedVT.getVectorElementType();\r
- if (TLI.isTypeLegal(FixedVT) &&\r
- FixedVT.getSizeInBits() == VT.getSizeInBits() &&\r
- FixedEltVT == InEltVT) {\r
- assert(FixedVT.getVectorNumElements() >= VT.getVectorNumElements() &&\r
- "Not enough elements in the fixed type for the operand!");\r
- assert(FixedVT.getVectorNumElements() != InVT.getVectorNumElements() &&\r
- "We can't have the same type as we started with!");\r
- if (FixedVT.getVectorNumElements() > InVT.getVectorNumElements())\r
- InOp = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, FixedVT,\r
- DAG.getUNDEF(FixedVT), InOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- else\r
- InOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, FixedVT, InOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- break;\r
- }\r
- }\r
- InVT = InOp.getValueType();\r
- if (InVT.getSizeInBits() != VT.getSizeInBits())\r
- // We couldn't find a legal vector type that was a widening of the input\r
- // and could be extended in-register to the result type, so we have to\r
- // scalarize.\r
- return WidenVecOp_Convert(N);\r
- }\r
-\r
- // Use special DAG nodes to represent the operation of extending the\r
- // low lanes.\r
- switch (N->getOpcode()) {\r
- default:\r
- llvm_unreachable("Extend legalization on on extend operation!");\r
- case ISD::ANY_EXTEND:\r
- return DAG.getAnyExtendVectorInReg(InOp, DL, VT);\r
- case ISD::SIGN_EXTEND:\r
- return DAG.getSignExtendVectorInReg(InOp, DL, VT);\r
- case ISD::ZERO_EXTEND:\r
- return DAG.getZeroExtendVectorInReg(InOp, DL, VT);\r
- }\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_Convert(SDNode *N) {\r
- // Since the result is legal and the input is illegal, it is unlikely\r
- // that we can fix the input to a legal type so unroll the convert\r
- // into some scalar code and create a nasty build vector.\r
- EVT VT = N->getValueType(0);\r
- EVT EltVT = VT.getVectorElementType();\r
- SDLoc dl(N);\r
- unsigned NumElts = VT.getVectorNumElements();\r
- SDValue InOp = N->getOperand(0);\r
- if (getTypeAction(InOp.getValueType()) == TargetLowering::TypeWidenVector)\r
- InOp = GetWidenedVector(InOp);\r
- EVT InVT = InOp.getValueType();\r
- EVT InEltVT = InVT.getVectorElementType();\r
-\r
- unsigned Opcode = N->getOpcode();\r
- SmallVector<SDValue, 16> Ops(NumElts);\r
- for (unsigned i=0; i < NumElts; ++i)\r
- Ops[i] = DAG.getNode(Opcode, dl, EltVT,\r
- DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp,\r
- DAG.getConstant(i, TLI.getVectorIdxTy())));\r
-\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_BITCAST(SDNode *N) {\r
- EVT VT = N->getValueType(0);\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- EVT InWidenVT = InOp.getValueType();\r
- SDLoc dl(N);\r
-\r
- // Check if we can convert between two legal vector types and extract.\r
- unsigned InWidenSize = InWidenVT.getSizeInBits();\r
- unsigned Size = VT.getSizeInBits();\r
- // x86mmx is not an acceptable vector element type, so don't try.\r
- if (InWidenSize % Size == 0 && !VT.isVector() && VT != MVT::x86mmx) {\r
- unsigned NewNumElts = InWidenSize / Size;\r
- EVT NewVT = EVT::getVectorVT(*DAG.getContext(), VT, NewNumElts);\r
- if (TLI.isTypeLegal(NewVT)) {\r
- SDValue BitOp = DAG.getNode(ISD::BITCAST, dl, NewVT, InOp);\r
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, BitOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- }\r
- }\r
-\r
- return CreateStackStoreLoad(InOp, VT);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_CONCAT_VECTORS(SDNode *N) {\r
- // If the input vector is not legal, it is likely that we will not find a\r
- // legal vector of the same size. Replace the concatenate vector with a\r
- // nasty build vector.\r
- EVT VT = N->getValueType(0);\r
- EVT EltVT = VT.getVectorElementType();\r
- SDLoc dl(N);\r
- unsigned NumElts = VT.getVectorNumElements();\r
- SmallVector<SDValue, 16> Ops(NumElts);\r
-\r
- EVT InVT = N->getOperand(0).getValueType();\r
- unsigned NumInElts = InVT.getVectorNumElements();\r
-\r
- unsigned Idx = 0;\r
- unsigned NumOperands = N->getNumOperands();\r
- for (unsigned i=0; i < NumOperands; ++i) {\r
- SDValue InOp = N->getOperand(i);\r
- if (getTypeAction(InOp.getValueType()) == TargetLowering::TypeWidenVector)\r
- InOp = GetWidenedVector(InOp);\r
- for (unsigned j=0; j < NumInElts; ++j)\r
- Ops[Idx++] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,\r
- DAG.getConstant(j, TLI.getVectorIdxTy()));\r
- }\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_EXTRACT_SUBVECTOR(SDNode *N) {\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N),\r
- N->getValueType(0), InOp, N->getOperand(1));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_EXTRACT_VECTOR_ELT(SDNode *N) {\r
- SDValue InOp = GetWidenedVector(N->getOperand(0));\r
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),\r
- N->getValueType(0), InOp, N->getOperand(1));\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_STORE(SDNode *N) {\r
- // We have to widen the value but we want only to store the original\r
- // vector type.\r
- StoreSDNode *ST = cast<StoreSDNode>(N);\r
-\r
- SmallVector<SDValue, 16> StChain;\r
- if (ST->isTruncatingStore())\r
- GenWidenVectorTruncStores(StChain, ST);\r
- else\r
- GenWidenVectorStores(StChain, ST);\r
-\r
- if (StChain.size() == 1)\r
- return StChain[0];\r
- else\r
- return DAG.getNode(ISD::TokenFactor, SDLoc(ST), MVT::Other, StChain);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::WidenVecOp_SETCC(SDNode *N) {\r
- SDValue InOp0 = GetWidenedVector(N->getOperand(0));\r
- SDValue InOp1 = GetWidenedVector(N->getOperand(1));\r
- SDLoc dl(N);\r
-\r
- // WARNING: In this code we widen the compare instruction with garbage.\r
- // This garbage may contain denormal floats which may be slow. Is this a real\r
- // concern ? Should we zero the unused lanes if this is a float compare ?\r
-\r
- // Get a new SETCC node to compare the newly widened operands.\r
- // Only some of the compared elements are legal.\r
- EVT SVT = TLI.getSetCCResultType(*DAG.getContext(), InOp0.getValueType());\r
- SDValue WideSETCC = DAG.getNode(ISD::SETCC, SDLoc(N),\r
- SVT, InOp0, InOp1, N->getOperand(2));\r
-\r
- // Extract the needed results from the result vector.\r
- EVT ResVT = EVT::getVectorVT(*DAG.getContext(),\r
- SVT.getVectorElementType(),\r
- N->getValueType(0).getVectorNumElements());\r
- SDValue CC = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl,\r
- ResVT, WideSETCC, DAG.getConstant(0,\r
- TLI.getVectorIdxTy()));\r
-\r
- return PromoteTargetBoolean(CC, N->getValueType(0));\r
-}\r
-\r
-\r
-//===----------------------------------------------------------------------===//\r
-// Vector Widening Utilities\r
-//===----------------------------------------------------------------------===//\r
-\r
-// Utility function to find the type to chop up a widen vector for load/store\r
-// TLI: Target lowering used to determine legal types.\r
-// Width: Width left need to load/store.\r
-// WidenVT: The widen vector type to load to/store from\r
-// Align: If 0, don't allow use of a wider type\r
-// WidenEx: If Align is not 0, the amount additional we can load/store from.\r
-\r
-static EVT FindMemType(SelectionDAG& DAG, const TargetLowering &TLI,\r
- unsigned Width, EVT WidenVT,\r
- unsigned Align = 0, unsigned WidenEx = 0) {\r
- EVT WidenEltVT = WidenVT.getVectorElementType();\r
- unsigned WidenWidth = WidenVT.getSizeInBits();\r
- unsigned WidenEltWidth = WidenEltVT.getSizeInBits();\r
- unsigned AlignInBits = Align*8;\r
-\r
- // If we have one element to load/store, return it.\r
- EVT RetVT = WidenEltVT;\r
- if (Width == WidenEltWidth)\r
- return RetVT;\r
-\r
- // See if there is larger legal integer than the element type to load/store\r
- unsigned VT;\r
- for (VT = (unsigned)MVT::LAST_INTEGER_VALUETYPE;\r
- VT >= (unsigned)MVT::FIRST_INTEGER_VALUETYPE; --VT) {\r
- EVT MemVT((MVT::SimpleValueType) VT);\r
- unsigned MemVTWidth = MemVT.getSizeInBits();\r
- if (MemVT.getSizeInBits() <= WidenEltWidth)\r
- break;\r
- if (TLI.isTypeLegal(MemVT) && (WidenWidth % MemVTWidth) == 0 &&\r
- isPowerOf2_32(WidenWidth / MemVTWidth) &&\r
- (MemVTWidth <= Width ||\r
- (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) {\r
- RetVT = MemVT;\r
- break;\r
- }\r
- }\r
-\r
- // See if there is a larger vector type to load/store that has the same vector\r
- // element type and is evenly divisible with the WidenVT.\r
- for (VT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;\r
- VT >= (unsigned)MVT::FIRST_VECTOR_VALUETYPE; --VT) {\r
- EVT MemVT = (MVT::SimpleValueType) VT;\r
- unsigned MemVTWidth = MemVT.getSizeInBits();\r
- if (TLI.isTypeLegal(MemVT) && WidenEltVT == MemVT.getVectorElementType() &&\r
- (WidenWidth % MemVTWidth) == 0 &&\r
- isPowerOf2_32(WidenWidth / MemVTWidth) &&\r
- (MemVTWidth <= Width ||\r
- (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) {\r
- if (RetVT.getSizeInBits() < MemVTWidth || MemVT == WidenVT)\r
- return MemVT;\r
- }\r
- }\r
-\r
- return RetVT;\r
-}\r
-\r
-// Builds a vector type from scalar loads\r
-// VecTy: Resulting Vector type\r
-// LDOps: Load operators to build a vector type\r
-// [Start,End) the list of loads to use.\r
-static SDValue BuildVectorFromScalar(SelectionDAG& DAG, EVT VecTy,\r
- SmallVectorImpl<SDValue> &LdOps,\r
- unsigned Start, unsigned End) {\r
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();\r
- SDLoc dl(LdOps[Start]);\r
- EVT LdTy = LdOps[Start].getValueType();\r
- unsigned Width = VecTy.getSizeInBits();\r
- unsigned NumElts = Width / LdTy.getSizeInBits();\r
- EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), LdTy, NumElts);\r
-\r
- unsigned Idx = 1;\r
- SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT,LdOps[Start]);\r
-\r
- for (unsigned i = Start + 1; i != End; ++i) {\r
- EVT NewLdTy = LdOps[i].getValueType();\r
- if (NewLdTy != LdTy) {\r
- NumElts = Width / NewLdTy.getSizeInBits();\r
- NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewLdTy, NumElts);\r
- VecOp = DAG.getNode(ISD::BITCAST, dl, NewVecVT, VecOp);\r
- // Readjust position and vector position based on new load type\r
- Idx = Idx * LdTy.getSizeInBits() / NewLdTy.getSizeInBits();\r
- LdTy = NewLdTy;\r
- }\r
- VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, VecOp, LdOps[i],\r
- DAG.getConstant(Idx++, TLI.getVectorIdxTy()));\r
- }\r
- return DAG.getNode(ISD::BITCAST, dl, VecTy, VecOp);\r
-}\r
-\r
-SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVectorImpl<SDValue> &LdChain,\r
- LoadSDNode *LD) {\r
- // The strategy assumes that we can efficiently load powers of two widths.\r
- // The routines chops the vector into the largest vector loads with the same\r
- // element type or scalar loads and then recombines it to the widen vector\r
- // type.\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0));\r
- unsigned WidenWidth = WidenVT.getSizeInBits();\r
- EVT LdVT = LD->getMemoryVT();\r
- SDLoc dl(LD);\r
- assert(LdVT.isVector() && WidenVT.isVector());\r
- assert(LdVT.getVectorElementType() == WidenVT.getVectorElementType());\r
-\r
- // Load information\r
- SDValue Chain = LD->getChain();\r
- SDValue BasePtr = LD->getBasePtr();\r
- unsigned Align = LD->getAlignment();\r
- bool isVolatile = LD->isVolatile();\r
- bool isNonTemporal = LD->isNonTemporal();\r
- bool isInvariant = LD->isInvariant();\r
- AAMDNodes AAInfo = LD->getAAInfo();\r
-\r
- int LdWidth = LdVT.getSizeInBits();\r
- int WidthDiff = WidenWidth - LdWidth; // Difference\r
- unsigned LdAlign = (isVolatile) ? 0 : Align; // Allow wider loads\r
-\r
- // Find the vector type that can load from.\r
- EVT NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff);\r
- int NewVTWidth = NewVT.getSizeInBits();\r
- SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, LD->getPointerInfo(),\r
- isVolatile, isNonTemporal, isInvariant, Align,\r
- AAInfo);\r
- LdChain.push_back(LdOp.getValue(1));\r
-\r
- // Check if we can load the element with one instruction\r
- if (LdWidth <= NewVTWidth) {\r
- if (!NewVT.isVector()) {\r
- unsigned NumElts = WidenWidth / NewVTWidth;\r
- EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts);\r
- SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT, LdOp);\r
- return DAG.getNode(ISD::BITCAST, dl, WidenVT, VecOp);\r
- }\r
- if (NewVT == WidenVT)\r
- return LdOp;\r
-\r
- assert(WidenWidth % NewVTWidth == 0);\r
- unsigned NumConcat = WidenWidth / NewVTWidth;\r
- SmallVector<SDValue, 16> ConcatOps(NumConcat);\r
- SDValue UndefVal = DAG.getUNDEF(NewVT);\r
- ConcatOps[0] = LdOp;\r
- for (unsigned i = 1; i != NumConcat; ++i)\r
- ConcatOps[i] = UndefVal;\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, ConcatOps);\r
- }\r
-\r
- // Load vector by using multiple loads from largest vector to scalar\r
- SmallVector<SDValue, 16> LdOps;\r
- LdOps.push_back(LdOp);\r
-\r
- LdWidth -= NewVTWidth;\r
- unsigned Offset = 0;\r
-\r
- while (LdWidth > 0) {\r
- unsigned Increment = NewVTWidth / 8;\r
- Offset += Increment;\r
- BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,\r
- DAG.getConstant(Increment, BasePtr.getValueType()));\r
-\r
- SDValue L;\r
- if (LdWidth < NewVTWidth) {\r
- // Our current type we are using is too large, find a better size\r
- NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff);\r
- NewVTWidth = NewVT.getSizeInBits();\r
- L = DAG.getLoad(NewVT, dl, Chain, BasePtr,\r
- LD->getPointerInfo().getWithOffset(Offset), isVolatile,\r
- isNonTemporal, isInvariant, MinAlign(Align, Increment),\r
- AAInfo);\r
- LdChain.push_back(L.getValue(1));\r
- if (L->getValueType(0).isVector()) {\r
- SmallVector<SDValue, 16> Loads;\r
- Loads.push_back(L);\r
- unsigned size = L->getValueSizeInBits(0);\r
- while (size < LdOp->getValueSizeInBits(0)) {\r
- Loads.push_back(DAG.getUNDEF(L->getValueType(0)));\r
- size += L->getValueSizeInBits(0);\r
- }\r
- L = DAG.getNode(ISD::CONCAT_VECTORS, dl, LdOp->getValueType(0), Loads);\r
- }\r
- } else {\r
- L = DAG.getLoad(NewVT, dl, Chain, BasePtr,\r
- LD->getPointerInfo().getWithOffset(Offset), isVolatile,\r
- isNonTemporal, isInvariant, MinAlign(Align, Increment),\r
- AAInfo);\r
- LdChain.push_back(L.getValue(1));\r
- }\r
-\r
- LdOps.push_back(L);\r
-\r
-\r
- LdWidth -= NewVTWidth;\r
- }\r
-\r
- // Build the vector from the loads operations\r
- unsigned End = LdOps.size();\r
- if (!LdOps[0].getValueType().isVector())\r
- // All the loads are scalar loads.\r
- return BuildVectorFromScalar(DAG, WidenVT, LdOps, 0, End);\r
-\r
- // If the load contains vectors, build the vector using concat vector.\r
- // All of the vectors used to loads are power of 2 and the scalars load\r
- // can be combined to make a power of 2 vector.\r
- SmallVector<SDValue, 16> ConcatOps(End);\r
- int i = End - 1;\r
- int Idx = End;\r
- EVT LdTy = LdOps[i].getValueType();\r
- // First combine the scalar loads to a vector\r
- if (!LdTy.isVector()) {\r
- for (--i; i >= 0; --i) {\r
- LdTy = LdOps[i].getValueType();\r
- if (LdTy.isVector())\r
- break;\r
- }\r
- ConcatOps[--Idx] = BuildVectorFromScalar(DAG, LdTy, LdOps, i+1, End);\r
- }\r
- ConcatOps[--Idx] = LdOps[i];\r
- for (--i; i >= 0; --i) {\r
- EVT NewLdTy = LdOps[i].getValueType();\r
- if (NewLdTy != LdTy) {\r
- // Create a larger vector\r
- ConcatOps[End-1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewLdTy,\r
- makeArrayRef(&ConcatOps[Idx], End - Idx));\r
- Idx = End - 1;\r
- LdTy = NewLdTy;\r
- }\r
- ConcatOps[--Idx] = LdOps[i];\r
- }\r
-\r
- if (WidenWidth == LdTy.getSizeInBits()*(End - Idx))\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT,\r
- makeArrayRef(&ConcatOps[Idx], End - Idx));\r
-\r
- // We need to fill the rest with undefs to build the vector\r
- unsigned NumOps = WidenWidth / LdTy.getSizeInBits();\r
- SmallVector<SDValue, 16> WidenOps(NumOps);\r
- SDValue UndefVal = DAG.getUNDEF(LdTy);\r
- {\r
- unsigned i = 0;\r
- for (; i != End-Idx; ++i)\r
- WidenOps[i] = ConcatOps[Idx+i];\r
- for (; i != NumOps; ++i)\r
- WidenOps[i] = UndefVal;\r
- }\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, WidenOps);\r
-}\r
-\r
-SDValue\r
-DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVectorImpl<SDValue> &LdChain,\r
- LoadSDNode *LD,\r
- ISD::LoadExtType ExtType) {\r
- // For extension loads, it may not be more efficient to chop up the vector\r
- // and then extended it. Instead, we unroll the load and build a new vector.\r
- EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0));\r
- EVT LdVT = LD->getMemoryVT();\r
- SDLoc dl(LD);\r
- assert(LdVT.isVector() && WidenVT.isVector());\r
-\r
- // Load information\r
- SDValue Chain = LD->getChain();\r
- SDValue BasePtr = LD->getBasePtr();\r
- unsigned Align = LD->getAlignment();\r
- bool isVolatile = LD->isVolatile();\r
- bool isNonTemporal = LD->isNonTemporal();\r
- bool isInvariant = LD->isInvariant();\r
- AAMDNodes AAInfo = LD->getAAInfo();\r
-\r
- EVT EltVT = WidenVT.getVectorElementType();\r
- EVT LdEltVT = LdVT.getVectorElementType();\r
- unsigned NumElts = LdVT.getVectorNumElements();\r
-\r
- // Load each element and widen\r
- unsigned WidenNumElts = WidenVT.getVectorNumElements();\r
- SmallVector<SDValue, 16> Ops(WidenNumElts);\r
- unsigned Increment = LdEltVT.getSizeInBits() / 8;\r
- Ops[0] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, BasePtr,\r
- LD->getPointerInfo(),\r
- LdEltVT, isVolatile, isNonTemporal, isInvariant,\r
- Align, AAInfo);\r
- LdChain.push_back(Ops[0].getValue(1));\r
- unsigned i = 0, Offset = Increment;\r
- for (i=1; i < NumElts; ++i, Offset += Increment) {\r
- SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),\r
- BasePtr,\r
- DAG.getConstant(Offset,\r
- BasePtr.getValueType()));\r
- Ops[i] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, NewBasePtr,\r
- LD->getPointerInfo().getWithOffset(Offset), LdEltVT,\r
- isVolatile, isNonTemporal, isInvariant, Align,\r
- AAInfo);\r
- LdChain.push_back(Ops[i].getValue(1));\r
- }\r
-\r
- // Fill the rest with undefs\r
- SDValue UndefVal = DAG.getUNDEF(EltVT);\r
- for (; i != WidenNumElts; ++i)\r
- Ops[i] = UndefVal;\r
-\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);\r
-}\r
-\r
-\r
-void DAGTypeLegalizer::GenWidenVectorStores(SmallVectorImpl<SDValue> &StChain,\r
- StoreSDNode *ST) {\r
- // The strategy assumes that we can efficiently store powers of two widths.\r
- // The routines chops the vector into the largest vector stores with the same\r
- // element type or scalar stores.\r
- SDValue Chain = ST->getChain();\r
- SDValue BasePtr = ST->getBasePtr();\r
- unsigned Align = ST->getAlignment();\r
- bool isVolatile = ST->isVolatile();\r
- bool isNonTemporal = ST->isNonTemporal();\r
- AAMDNodes AAInfo = ST->getAAInfo();\r
- SDValue ValOp = GetWidenedVector(ST->getValue());\r
- SDLoc dl(ST);\r
-\r
- EVT StVT = ST->getMemoryVT();\r
- unsigned StWidth = StVT.getSizeInBits();\r
- EVT ValVT = ValOp.getValueType();\r
- unsigned ValWidth = ValVT.getSizeInBits();\r
- EVT ValEltVT = ValVT.getVectorElementType();\r
- unsigned ValEltWidth = ValEltVT.getSizeInBits();\r
- assert(StVT.getVectorElementType() == ValEltVT);\r
-\r
- int Idx = 0; // current index to store\r
- unsigned Offset = 0; // offset from base to store\r
- while (StWidth != 0) {\r
- // Find the largest vector type we can store with\r
- EVT NewVT = FindMemType(DAG, TLI, StWidth, ValVT);\r
- unsigned NewVTWidth = NewVT.getSizeInBits();\r
- unsigned Increment = NewVTWidth / 8;\r
- if (NewVT.isVector()) {\r
- unsigned NumVTElts = NewVT.getVectorNumElements();\r
- do {\r
- SDValue EOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NewVT, ValOp,\r
- DAG.getConstant(Idx, TLI.getVectorIdxTy()));\r
- StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr,\r
- ST->getPointerInfo().getWithOffset(Offset),\r
- isVolatile, isNonTemporal,\r
- MinAlign(Align, Offset), AAInfo));\r
- StWidth -= NewVTWidth;\r
- Offset += Increment;\r
- Idx += NumVTElts;\r
- BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,\r
- DAG.getConstant(Increment, BasePtr.getValueType()));\r
- } while (StWidth != 0 && StWidth >= NewVTWidth);\r
- } else {\r
- // Cast the vector to the scalar type we can store\r
- unsigned NumElts = ValWidth / NewVTWidth;\r
- EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts);\r
- SDValue VecOp = DAG.getNode(ISD::BITCAST, dl, NewVecVT, ValOp);\r
- // Readjust index position based on new vector type\r
- Idx = Idx * ValEltWidth / NewVTWidth;\r
- do {\r
- SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewVT, VecOp,\r
- DAG.getConstant(Idx++, TLI.getVectorIdxTy()));\r
- StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr,\r
- ST->getPointerInfo().getWithOffset(Offset),\r
- isVolatile, isNonTemporal,\r
- MinAlign(Align, Offset), AAInfo));\r
- StWidth -= NewVTWidth;\r
- Offset += Increment;\r
- BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,\r
- DAG.getConstant(Increment, BasePtr.getValueType()));\r
- } while (StWidth != 0 && StWidth >= NewVTWidth);\r
- // Restore index back to be relative to the original widen element type\r
- Idx = Idx * NewVTWidth / ValEltWidth;\r
- }\r
- }\r
-}\r
-\r
-void\r
-DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVectorImpl<SDValue> &StChain,\r
- StoreSDNode *ST) {\r
- // For extension loads, it may not be more efficient to truncate the vector\r
- // and then store it. Instead, we extract each element and then store it.\r
- SDValue Chain = ST->getChain();\r
- SDValue BasePtr = ST->getBasePtr();\r
- unsigned Align = ST->getAlignment();\r
- bool isVolatile = ST->isVolatile();\r
- bool isNonTemporal = ST->isNonTemporal();\r
- AAMDNodes AAInfo = ST->getAAInfo();\r
- SDValue ValOp = GetWidenedVector(ST->getValue());\r
- SDLoc dl(ST);\r
-\r
- EVT StVT = ST->getMemoryVT();\r
- EVT ValVT = ValOp.getValueType();\r
-\r
- // It must be true that we the widen vector type is bigger than where\r
- // we need to store.\r
- assert(StVT.isVector() && ValOp.getValueType().isVector());\r
- assert(StVT.bitsLT(ValOp.getValueType()));\r
-\r
- // For truncating stores, we can not play the tricks of chopping legal\r
- // vector types and bit cast it to the right type. Instead, we unroll\r
- // the store.\r
- EVT StEltVT = StVT.getVectorElementType();\r
- EVT ValEltVT = ValVT.getVectorElementType();\r
- unsigned Increment = ValEltVT.getSizeInBits() / 8;\r
- unsigned NumElts = StVT.getVectorNumElements();\r
- SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, BasePtr,\r
- ST->getPointerInfo(), StEltVT,\r
- isVolatile, isNonTemporal, Align,\r
- AAInfo));\r
- unsigned Offset = Increment;\r
- for (unsigned i=1; i < NumElts; ++i, Offset += Increment) {\r
- SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),\r
- BasePtr, DAG.getConstant(Offset,\r
- BasePtr.getValueType()));\r
- SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
- StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, NewBasePtr,\r
- ST->getPointerInfo().getWithOffset(Offset),\r
- StEltVT, isVolatile, isNonTemporal,\r
- MinAlign(Align, Offset), AAInfo));\r
- }\r
-}\r
-\r
-/// Modifies a vector input (widen or narrows) to a vector of NVT. The\r
-/// input vector must have the same element type as NVT.\r
-SDValue DAGTypeLegalizer::ModifyToType(SDValue InOp, EVT NVT) {\r
- // Note that InOp might have been widened so it might already have\r
- // the right width or it might need be narrowed.\r
- EVT InVT = InOp.getValueType();\r
- assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&\r
- "input and widen element type must match");\r
- SDLoc dl(InOp);\r
-\r
- // Check if InOp already has the right width.\r
- if (InVT == NVT)\r
- return InOp;\r
-\r
- unsigned InNumElts = InVT.getVectorNumElements();\r
- unsigned WidenNumElts = NVT.getVectorNumElements();\r
- if (WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0) {\r
- unsigned NumConcat = WidenNumElts / InNumElts;\r
- SmallVector<SDValue, 16> Ops(NumConcat);\r
- SDValue UndefVal = DAG.getUNDEF(InVT);\r
- Ops[0] = InOp;\r
- for (unsigned i = 1; i != NumConcat; ++i)\r
- Ops[i] = UndefVal;\r
-\r
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, NVT, Ops);\r
- }\r
-\r
- if (WidenNumElts < InNumElts && InNumElts % WidenNumElts)\r
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT, InOp,\r
- DAG.getConstant(0, TLI.getVectorIdxTy()));\r
-\r
- // Fall back to extract and build.\r
- SmallVector<SDValue, 16> Ops(WidenNumElts);\r
- EVT EltVT = NVT.getVectorElementType();\r
- unsigned MinNumElts = std::min(WidenNumElts, InNumElts);\r
- unsigned Idx;\r
- for (Idx = 0; Idx < MinNumElts; ++Idx)\r
- Ops[Idx] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,\r
- DAG.getConstant(Idx, TLI.getVectorIdxTy()));\r
-\r
- SDValue UndefVal = DAG.getUNDEF(EltVT);\r
- for ( ; Idx < WidenNumElts; ++Idx)\r
- Ops[Idx] = UndefVal;\r
- return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);\r
-}\r
+//===------- LegalizeVectorTypes.cpp - Legalization of vector types -------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file performs vector type splitting and scalarization for LegalizeTypes.
+// Scalarization is the act of changing a computation in an illegal one-element
+// vector type to be a computation in its scalar element type. For example,
+// implementing <1 x f32> arithmetic in a scalar f32 register. This is needed
+// as a base case when scalarizing vector arithmetic like <4 x f32>, which
+// eventually decomposes to scalars if the target doesn't support v4f32 or v2f32
+// types.
+// Splitting is the act of changing a computation in an invalid vector type to
+// be a computation in two vectors of half the size. For example, implementing
+// <128 x f32> operations in terms of two <64 x f32> operations.
+//
+//===----------------------------------------------------------------------===//
+
+#include "LegalizeTypes.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+#define DEBUG_TYPE "legalize-types"
+
+//===----------------------------------------------------------------------===//
+// Result Vector Scalarization: <1 x ty> -> ty.
+//===----------------------------------------------------------------------===//
+
+void DAGTypeLegalizer::ScalarizeVectorResult(SDNode *N, unsigned ResNo) {
+ DEBUG(dbgs() << "Scalarize node result " << ResNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n");
+ SDValue R = SDValue();
+
+ switch (N->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "ScalarizeVectorResult #" << ResNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n";
+#endif
+ report_fatal_error("Do not know how to scalarize the result of this "
+ "operator!\n");
+
+ case ISD::MERGE_VALUES: R = ScalarizeVecRes_MERGE_VALUES(N, ResNo);break;
+ case ISD::BITCAST: R = ScalarizeVecRes_BITCAST(N); break;
+ case ISD::BUILD_VECTOR: R = ScalarizeVecRes_BUILD_VECTOR(N); break;
+ case ISD::CONVERT_RNDSAT: R = ScalarizeVecRes_CONVERT_RNDSAT(N); break;
+ case ISD::EXTRACT_SUBVECTOR: R = ScalarizeVecRes_EXTRACT_SUBVECTOR(N); break;
+ case ISD::FP_ROUND: R = ScalarizeVecRes_FP_ROUND(N); break;
+ case ISD::FP_ROUND_INREG: R = ScalarizeVecRes_InregOp(N); break;
+ case ISD::FPOWI: R = ScalarizeVecRes_FPOWI(N); break;
+ case ISD::INSERT_VECTOR_ELT: R = ScalarizeVecRes_INSERT_VECTOR_ELT(N); break;
+ case ISD::LOAD: R = ScalarizeVecRes_LOAD(cast<LoadSDNode>(N));break;
+ case ISD::SCALAR_TO_VECTOR: R = ScalarizeVecRes_SCALAR_TO_VECTOR(N); break;
+ case ISD::SIGN_EXTEND_INREG: R = ScalarizeVecRes_InregOp(N); break;
+ case ISD::VSELECT: R = ScalarizeVecRes_VSELECT(N); break;
+ case ISD::SELECT: R = ScalarizeVecRes_SELECT(N); break;
+ case ISD::SELECT_CC: R = ScalarizeVecRes_SELECT_CC(N); break;
+ case ISD::SETCC: R = ScalarizeVecRes_SETCC(N); break;
+ case ISD::UNDEF: R = ScalarizeVecRes_UNDEF(N); break;
+ case ISD::VECTOR_SHUFFLE: R = ScalarizeVecRes_VECTOR_SHUFFLE(N); break;
+ case ISD::ANY_EXTEND:
+ case ISD::BSWAP:
+ case ISD::CTLZ:
+ case ISD::CTLZ_ZERO_UNDEF:
+ case ISD::CTPOP:
+ case ISD::CTTZ:
+ case ISD::CTTZ_ZERO_UNDEF:
+ case ISD::FABS:
+ case ISD::FCEIL:
+ case ISD::FCOS:
+ case ISD::FEXP:
+ case ISD::FEXP2:
+ case ISD::FFLOOR:
+ case ISD::FLOG:
+ case ISD::FLOG10:
+ case ISD::FLOG2:
+ case ISD::FNEARBYINT:
+ case ISD::FNEG:
+ case ISD::FP_EXTEND:
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT:
+ case ISD::FRINT:
+ case ISD::FROUND:
+ case ISD::FSIN:
+ case ISD::FSQRT:
+ case ISD::FTRUNC:
+ case ISD::SIGN_EXTEND:
+ case ISD::SINT_TO_FP:
+ case ISD::TRUNCATE:
+ case ISD::UINT_TO_FP:
+ case ISD::ZERO_EXTEND:
+ R = ScalarizeVecRes_UnaryOp(N);
+ break;
+
+ case ISD::ADD:
+ case ISD::AND:
+ case ISD::FADD:
+ case ISD::FCOPYSIGN:
+ case ISD::FDIV:
+ case ISD::FMUL:
+ case ISD::FMINNUM:
+ case ISD::FMAXNUM:
+
+ case ISD::FPOW:
+ case ISD::FREM:
+ case ISD::FSUB:
+ case ISD::MUL:
+ case ISD::OR:
+ case ISD::SDIV:
+ case ISD::SREM:
+ case ISD::SUB:
+ case ISD::UDIV:
+ case ISD::UREM:
+ case ISD::XOR:
+ case ISD::SHL:
+ case ISD::SRA:
+ case ISD::SRL:
+ R = ScalarizeVecRes_BinOp(N);
+ break;
+ case ISD::FMA:
+ R = ScalarizeVecRes_TernaryOp(N);
+ break;
+ }
+
+ // If R is null, the sub-method took care of registering the result.
+ if (R.getNode())
+ SetScalarizedVector(SDValue(N, ResNo), R);
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_BinOp(SDNode *N) {
+ SDValue LHS = GetScalarizedVector(N->getOperand(0));
+ SDValue RHS = GetScalarizedVector(N->getOperand(1));
+ return DAG.getNode(N->getOpcode(), SDLoc(N),
+ LHS.getValueType(), LHS, RHS);
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_TernaryOp(SDNode *N) {
+ SDValue Op0 = GetScalarizedVector(N->getOperand(0));
+ SDValue Op1 = GetScalarizedVector(N->getOperand(1));
+ SDValue Op2 = GetScalarizedVector(N->getOperand(2));
+ return DAG.getNode(N->getOpcode(), SDLoc(N),
+ Op0.getValueType(), Op0, Op1, Op2);
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_MERGE_VALUES(SDNode *N,
+ unsigned ResNo) {
+ SDValue Op = DisintegrateMERGE_VALUES(N, ResNo);
+ return GetScalarizedVector(Op);
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_BITCAST(SDNode *N) {
+ EVT NewVT = N->getValueType(0).getVectorElementType();
+ return DAG.getNode(ISD::BITCAST, SDLoc(N),
+ NewVT, N->getOperand(0));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_BUILD_VECTOR(SDNode *N) {
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ SDValue InOp = N->getOperand(0);
+ // The BUILD_VECTOR operands may be of wider element types and
+ // we may need to truncate them back to the requested return type.
+ if (EltVT.isInteger())
+ return DAG.getNode(ISD::TRUNCATE, SDLoc(N), EltVT, InOp);
+ return InOp;
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_CONVERT_RNDSAT(SDNode *N) {
+ EVT NewVT = N->getValueType(0).getVectorElementType();
+ SDValue Op0 = GetScalarizedVector(N->getOperand(0));
+ return DAG.getConvertRndSat(NewVT, SDLoc(N),
+ Op0, DAG.getValueType(NewVT),
+ DAG.getValueType(Op0.getValueType()),
+ N->getOperand(3),
+ N->getOperand(4),
+ cast<CvtRndSatSDNode>(N)->getCvtCode());
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_EXTRACT_SUBVECTOR(SDNode *N) {
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
+ N->getValueType(0).getVectorElementType(),
+ N->getOperand(0), N->getOperand(1));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_FP_ROUND(SDNode *N) {
+ EVT NewVT = N->getValueType(0).getVectorElementType();
+ SDValue Op = GetScalarizedVector(N->getOperand(0));
+ return DAG.getNode(ISD::FP_ROUND, SDLoc(N),
+ NewVT, Op, N->getOperand(1));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_FPOWI(SDNode *N) {
+ SDValue Op = GetScalarizedVector(N->getOperand(0));
+ return DAG.getNode(ISD::FPOWI, SDLoc(N),
+ Op.getValueType(), Op, N->getOperand(1));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_INSERT_VECTOR_ELT(SDNode *N) {
+ // The value to insert may have a wider type than the vector element type,
+ // so be sure to truncate it to the element type if necessary.
+ SDValue Op = N->getOperand(1);
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ if (Op.getValueType() != EltVT)
+ // FIXME: Can this happen for floating point types?
+ Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), EltVT, Op);
+ return Op;
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_LOAD(LoadSDNode *N) {
+ assert(N->isUnindexed() && "Indexed vector load?");
+
+ SDValue Result = DAG.getLoad(ISD::UNINDEXED,
+ N->getExtensionType(),
+ N->getValueType(0).getVectorElementType(),
+ SDLoc(N),
+ N->getChain(), N->getBasePtr(),
+ DAG.getUNDEF(N->getBasePtr().getValueType()),
+ N->getPointerInfo(),
+ N->getMemoryVT().getVectorElementType(),
+ N->isVolatile(), N->isNonTemporal(),
+ N->isInvariant(), N->getOriginalAlignment(),
+ N->getAAInfo());
+
+ // Legalized the chain result - switch anything that used the old chain to
+ // use the new one.
+ ReplaceValueWith(SDValue(N, 1), Result.getValue(1));
+ return Result;
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_UnaryOp(SDNode *N) {
+ // Get the dest type - it doesn't always match the input type, e.g. int_to_fp.
+ EVT DestVT = N->getValueType(0).getVectorElementType();
+ SDValue Op = N->getOperand(0);
+ EVT OpVT = Op.getValueType();
+ SDLoc DL(N);
+ // The result needs scalarizing, but it's not a given that the source does.
+ // This is a workaround for targets where it's impossible to scalarize the
+ // result of a conversion, because the source type is legal.
+ // For instance, this happens on AArch64: v1i1 is illegal but v1i{8,16,32}
+ // are widened to v8i8, v4i16, and v2i32, which is legal, because v1i64 is
+ // legal and was not scalarized.
+ // See the similar logic in ScalarizeVecRes_VSETCC
+ if (getTypeAction(OpVT) == TargetLowering::TypeScalarizeVector) {
+ Op = GetScalarizedVector(Op);
+ } else {
+ EVT VT = OpVT.getVectorElementType();
+ Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ }
+ return DAG.getNode(N->getOpcode(), SDLoc(N), DestVT, Op);
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_InregOp(SDNode *N) {
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ EVT ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT().getVectorElementType();
+ SDValue LHS = GetScalarizedVector(N->getOperand(0));
+ return DAG.getNode(N->getOpcode(), SDLoc(N), EltVT,
+ LHS, DAG.getValueType(ExtVT));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_SCALAR_TO_VECTOR(SDNode *N) {
+ // If the operand is wider than the vector element type then it is implicitly
+ // truncated. Make that explicit here.
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ SDValue InOp = N->getOperand(0);
+ if (InOp.getValueType() != EltVT)
+ return DAG.getNode(ISD::TRUNCATE, SDLoc(N), EltVT, InOp);
+ return InOp;
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_VSELECT(SDNode *N) {
+ SDValue Cond = GetScalarizedVector(N->getOperand(0));
+ SDValue LHS = GetScalarizedVector(N->getOperand(1));
+ TargetLowering::BooleanContent ScalarBool =
+ TLI.getBooleanContents(false, false);
+ TargetLowering::BooleanContent VecBool = TLI.getBooleanContents(true, false);
+
+ // If integer and float booleans have different contents then we can't
+ // reliably optimize in all cases. There is a full explanation for this in
+ // DAGCombiner::visitSELECT() where the same issue affects folding
+ // (select C, 0, 1) to (xor C, 1).
+ if (TLI.getBooleanContents(false, false) !=
+ TLI.getBooleanContents(false, true)) {
+ // At least try the common case where the boolean is generated by a
+ // comparison.
+ if (Cond->getOpcode() == ISD::SETCC) {
+ EVT OpVT = Cond->getOperand(0)->getValueType(0);
+ ScalarBool = TLI.getBooleanContents(OpVT.getScalarType());
+ VecBool = TLI.getBooleanContents(OpVT);
+ } else
+ ScalarBool = TargetLowering::UndefinedBooleanContent;
+ }
+
+ if (ScalarBool != VecBool) {
+ EVT CondVT = Cond.getValueType();
+ switch (ScalarBool) {
+ case TargetLowering::UndefinedBooleanContent:
+ break;
+ case TargetLowering::ZeroOrOneBooleanContent:
+ assert(VecBool == TargetLowering::UndefinedBooleanContent ||
+ VecBool == TargetLowering::ZeroOrNegativeOneBooleanContent);
+ // Vector read from all ones, scalar expects a single 1 so mask.
+ Cond = DAG.getNode(ISD::AND, SDLoc(N), CondVT,
+ Cond, DAG.getConstant(1, CondVT));
+ break;
+ case TargetLowering::ZeroOrNegativeOneBooleanContent:
+ assert(VecBool == TargetLowering::UndefinedBooleanContent ||
+ VecBool == TargetLowering::ZeroOrOneBooleanContent);
+ // Vector reads from a one, scalar from all ones so sign extend.
+ Cond = DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), CondVT,
+ Cond, DAG.getValueType(MVT::i1));
+ break;
+ }
+ }
+
+ return DAG.getSelect(SDLoc(N),
+ LHS.getValueType(), Cond, LHS,
+ GetScalarizedVector(N->getOperand(2)));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT(SDNode *N) {
+ SDValue LHS = GetScalarizedVector(N->getOperand(1));
+ return DAG.getSelect(SDLoc(N),
+ LHS.getValueType(), N->getOperand(0), LHS,
+ GetScalarizedVector(N->getOperand(2)));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT_CC(SDNode *N) {
+ SDValue LHS = GetScalarizedVector(N->getOperand(2));
+ return DAG.getNode(ISD::SELECT_CC, SDLoc(N), LHS.getValueType(),
+ N->getOperand(0), N->getOperand(1),
+ LHS, GetScalarizedVector(N->getOperand(3)),
+ N->getOperand(4));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_SETCC(SDNode *N) {
+ assert(N->getValueType(0).isVector() ==
+ N->getOperand(0).getValueType().isVector() &&
+ "Scalar/Vector type mismatch");
+
+ if (N->getValueType(0).isVector()) return ScalarizeVecRes_VSETCC(N);
+
+ SDValue LHS = GetScalarizedVector(N->getOperand(0));
+ SDValue RHS = GetScalarizedVector(N->getOperand(1));
+ SDLoc DL(N);
+
+ // Turn it into a scalar SETCC.
+ return DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, N->getOperand(2));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_UNDEF(SDNode *N) {
+ return DAG.getUNDEF(N->getValueType(0).getVectorElementType());
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_VECTOR_SHUFFLE(SDNode *N) {
+ // Figure out if the scalar is the LHS or RHS and return it.
+ SDValue Arg = N->getOperand(2).getOperand(0);
+ if (Arg.getOpcode() == ISD::UNDEF)
+ return DAG.getUNDEF(N->getValueType(0).getVectorElementType());
+ unsigned Op = !cast<ConstantSDNode>(Arg)->isNullValue();
+ return GetScalarizedVector(N->getOperand(Op));
+}
+
+SDValue DAGTypeLegalizer::ScalarizeVecRes_VSETCC(SDNode *N) {
+ assert(N->getValueType(0).isVector() &&
+ N->getOperand(0).getValueType().isVector() &&
+ "Operand types must be vectors");
+ SDValue LHS = N->getOperand(0);
+ SDValue RHS = N->getOperand(1);
+ EVT OpVT = LHS.getValueType();
+ EVT NVT = N->getValueType(0).getVectorElementType();
+ SDLoc DL(N);
+
+ // The result needs scalarizing, but it's not a given that the source does.
+ if (getTypeAction(OpVT) == TargetLowering::TypeScalarizeVector) {
+ LHS = GetScalarizedVector(LHS);
+ RHS = GetScalarizedVector(RHS);
+ } else {
+ EVT VT = OpVT.getVectorElementType();
+ LHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, LHS,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ RHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, RHS,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ }
+
+ // Turn it into a scalar SETCC.
+ SDValue Res = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS,
+ N->getOperand(2));
+ // Vectors may have a different boolean contents to scalars. Promote the
+ // value appropriately.
+ ISD::NodeType ExtendCode =
+ TargetLowering::getExtendForContent(TLI.getBooleanContents(OpVT));
+ return DAG.getNode(ExtendCode, DL, NVT, Res);
+}
+
+
+//===----------------------------------------------------------------------===//
+// Operand Vector Scalarization <1 x ty> -> ty.
+//===----------------------------------------------------------------------===//
+
+bool DAGTypeLegalizer::ScalarizeVectorOperand(SDNode *N, unsigned OpNo) {
+ DEBUG(dbgs() << "Scalarize node operand " << OpNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n");
+ SDValue Res = SDValue();
+
+ if (!Res.getNode()) {
+ switch (N->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "ScalarizeVectorOperand Op #" << OpNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n";
+#endif
+ llvm_unreachable("Do not know how to scalarize this operator's operand!");
+ case ISD::BITCAST:
+ Res = ScalarizeVecOp_BITCAST(N);
+ break;
+ case ISD::ANY_EXTEND:
+ case ISD::ZERO_EXTEND:
+ case ISD::SIGN_EXTEND:
+ case ISD::TRUNCATE:
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT:
+ case ISD::SINT_TO_FP:
+ case ISD::UINT_TO_FP:
+ Res = ScalarizeVecOp_UnaryOp(N);
+ break;
+ case ISD::CONCAT_VECTORS:
+ Res = ScalarizeVecOp_CONCAT_VECTORS(N);
+ break;
+ case ISD::EXTRACT_VECTOR_ELT:
+ Res = ScalarizeVecOp_EXTRACT_VECTOR_ELT(N);
+ break;
+ case ISD::VSELECT:
+ Res = ScalarizeVecOp_VSELECT(N);
+ break;
+ case ISD::STORE:
+ Res = ScalarizeVecOp_STORE(cast<StoreSDNode>(N), OpNo);
+ break;
+ case ISD::FP_ROUND:
+ Res = ScalarizeVecOp_FP_ROUND(N, OpNo);
+ break;
+ }
+ }
+
+ // If the result is null, the sub-method took care of registering results etc.
+ if (!Res.getNode()) return false;
+
+ // If the result is N, the sub-method updated N in place. Tell the legalizer
+ // core about this.
+ if (Res.getNode() == N)
+ return true;
+
+ assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&
+ "Invalid operand expansion");
+
+ ReplaceValueWith(SDValue(N, 0), Res);
+ return false;
+}
+
+/// ScalarizeVecOp_BITCAST - If the value to convert is a vector that needs
+/// to be scalarized, it must be <1 x ty>. Convert the element instead.
+SDValue DAGTypeLegalizer::ScalarizeVecOp_BITCAST(SDNode *N) {
+ SDValue Elt = GetScalarizedVector(N->getOperand(0));
+ return DAG.getNode(ISD::BITCAST, SDLoc(N),
+ N->getValueType(0), Elt);
+}
+
+/// ScalarizeVecOp_UnaryOp - If the input is a vector that needs to be
+/// scalarized, it must be <1 x ty>. Do the operation on the element instead.
+SDValue DAGTypeLegalizer::ScalarizeVecOp_UnaryOp(SDNode *N) {
+ assert(N->getValueType(0).getVectorNumElements() == 1 &&
+ "Unexpected vector type!");
+ SDValue Elt = GetScalarizedVector(N->getOperand(0));
+ SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N),
+ N->getValueType(0).getScalarType(), Elt);
+ // Revectorize the result so the types line up with what the uses of this
+ // expression expect.
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N->getValueType(0), Op);
+}
+
+/// ScalarizeVecOp_CONCAT_VECTORS - The vectors to concatenate have length one -
+/// use a BUILD_VECTOR instead.
+SDValue DAGTypeLegalizer::ScalarizeVecOp_CONCAT_VECTORS(SDNode *N) {
+ SmallVector<SDValue, 8> Ops(N->getNumOperands());
+ for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
+ Ops[i] = GetScalarizedVector(N->getOperand(i));
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N->getValueType(0), Ops);
+}
+
+/// ScalarizeVecOp_EXTRACT_VECTOR_ELT - If the input is a vector that needs to
+/// be scalarized, it must be <1 x ty>, so just return the element, ignoring the
+/// index.
+SDValue DAGTypeLegalizer::ScalarizeVecOp_EXTRACT_VECTOR_ELT(SDNode *N) {
+ SDValue Res = GetScalarizedVector(N->getOperand(0));
+ if (Res.getValueType() != N->getValueType(0))
+ Res = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), N->getValueType(0),
+ Res);
+ return Res;
+}
+
+
+/// ScalarizeVecOp_VSELECT - If the input condition is a vector that needs to be
+/// scalarized, it must be <1 x i1>, so just convert to a normal ISD::SELECT
+/// (still with vector output type since that was acceptable if we got here).
+SDValue DAGTypeLegalizer::ScalarizeVecOp_VSELECT(SDNode *N) {
+ SDValue ScalarCond = GetScalarizedVector(N->getOperand(0));
+ EVT VT = N->getValueType(0);
+
+ return DAG.getNode(ISD::SELECT, SDLoc(N), VT, ScalarCond, N->getOperand(1),
+ N->getOperand(2));
+}
+
+/// ScalarizeVecOp_STORE - If the value to store is a vector that needs to be
+/// scalarized, it must be <1 x ty>. Just store the element.
+SDValue DAGTypeLegalizer::ScalarizeVecOp_STORE(StoreSDNode *N, unsigned OpNo){
+ assert(N->isUnindexed() && "Indexed store of one-element vector?");
+ assert(OpNo == 1 && "Do not know how to scalarize this operand!");
+ SDLoc dl(N);
+
+ if (N->isTruncatingStore())
+ return DAG.getTruncStore(N->getChain(), dl,
+ GetScalarizedVector(N->getOperand(1)),
+ N->getBasePtr(), N->getPointerInfo(),
+ N->getMemoryVT().getVectorElementType(),
+ N->isVolatile(), N->isNonTemporal(),
+ N->getAlignment(), N->getAAInfo());
+
+ return DAG.getStore(N->getChain(), dl, GetScalarizedVector(N->getOperand(1)),
+ N->getBasePtr(), N->getPointerInfo(),
+ N->isVolatile(), N->isNonTemporal(),
+ N->getOriginalAlignment(), N->getAAInfo());
+}
+
+/// ScalarizeVecOp_FP_ROUND - If the value to round is a vector that needs
+/// to be scalarized, it must be <1 x ty>. Convert the element instead.
+SDValue DAGTypeLegalizer::ScalarizeVecOp_FP_ROUND(SDNode *N, unsigned OpNo) {
+ SDValue Elt = GetScalarizedVector(N->getOperand(0));
+ SDValue Res = DAG.getNode(ISD::FP_ROUND, SDLoc(N),
+ N->getValueType(0).getVectorElementType(), Elt,
+ N->getOperand(1));
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), N->getValueType(0), Res);
+}
+
+//===----------------------------------------------------------------------===//
+// Result Vector Splitting
+//===----------------------------------------------------------------------===//
+
+/// SplitVectorResult - This method is called when the specified result of the
+/// specified node is found to need vector splitting. At this point, the node
+/// may also have invalid operands or may have other results that need
+/// legalization, we just know that (at least) one result needs vector
+/// splitting.
+void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) {
+ DEBUG(dbgs() << "Split node result: ";
+ N->dump(&DAG);
+ dbgs() << "\n");
+ SDValue Lo, Hi;
+
+ // See if the target wants to custom expand this node.
+ if (CustomLowerNode(N, N->getValueType(ResNo), true))
+ return;
+
+ switch (N->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "SplitVectorResult #" << ResNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n";
+#endif
+ report_fatal_error("Do not know how to split the result of this "
+ "operator!\n");
+
+ case ISD::MERGE_VALUES: SplitRes_MERGE_VALUES(N, ResNo, Lo, Hi); break;
+ case ISD::VSELECT:
+ case ISD::SELECT: SplitRes_SELECT(N, Lo, Hi); break;
+ case ISD::SELECT_CC: SplitRes_SELECT_CC(N, Lo, Hi); break;
+ case ISD::UNDEF: SplitRes_UNDEF(N, Lo, Hi); break;
+ case ISD::BITCAST: SplitVecRes_BITCAST(N, Lo, Hi); break;
+ case ISD::BUILD_VECTOR: SplitVecRes_BUILD_VECTOR(N, Lo, Hi); break;
+ case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break;
+ case ISD::EXTRACT_SUBVECTOR: SplitVecRes_EXTRACT_SUBVECTOR(N, Lo, Hi); break;
+ case ISD::INSERT_SUBVECTOR: SplitVecRes_INSERT_SUBVECTOR(N, Lo, Hi); break;
+ case ISD::FP_ROUND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;
+ case ISD::FPOWI: SplitVecRes_FPOWI(N, Lo, Hi); break;
+ case ISD::INSERT_VECTOR_ELT: SplitVecRes_INSERT_VECTOR_ELT(N, Lo, Hi); break;
+ case ISD::SCALAR_TO_VECTOR: SplitVecRes_SCALAR_TO_VECTOR(N, Lo, Hi); break;
+ case ISD::SIGN_EXTEND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break;
+ case ISD::LOAD:
+ SplitVecRes_LOAD(cast<LoadSDNode>(N), Lo, Hi);
+ break;
+ case ISD::MLOAD:
+ SplitVecRes_MLOAD(cast<MaskedLoadSDNode>(N), Lo, Hi);
+ break;
+ case ISD::SETCC:
+ SplitVecRes_SETCC(N, Lo, Hi);
+ break;
+ case ISD::VECTOR_SHUFFLE:
+ SplitVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N), Lo, Hi);
+ break;
+
+ case ISD::BSWAP:
+ case ISD::CONVERT_RNDSAT:
+ case ISD::CTLZ:
+ case ISD::CTTZ:
+ case ISD::CTLZ_ZERO_UNDEF:
+ case ISD::CTTZ_ZERO_UNDEF:
+ case ISD::CTPOP:
+ case ISD::FABS:
+ case ISD::FCEIL:
+ case ISD::FCOS:
+ case ISD::FEXP:
+ case ISD::FEXP2:
+ case ISD::FFLOOR:
+ case ISD::FLOG:
+ case ISD::FLOG10:
+ case ISD::FLOG2:
+ case ISD::FNEARBYINT:
+ case ISD::FNEG:
+ case ISD::FP_EXTEND:
+ case ISD::FP_ROUND:
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT:
+ case ISD::FRINT:
+ case ISD::FROUND:
+ case ISD::FSIN:
+ case ISD::FSQRT:
+ case ISD::FTRUNC:
+ case ISD::SINT_TO_FP:
+ case ISD::TRUNCATE:
+ case ISD::UINT_TO_FP:
+ SplitVecRes_UnaryOp(N, Lo, Hi);
+ break;
+
+ case ISD::ANY_EXTEND:
+ case ISD::SIGN_EXTEND:
+ case ISD::ZERO_EXTEND:
+ SplitVecRes_ExtendOp(N, Lo, Hi);
+ break;
+
+ case ISD::ADD:
+ case ISD::SUB:
+ case ISD::MUL:
+ case ISD::FADD:
+ case ISD::FCOPYSIGN:
+ case ISD::FSUB:
+ case ISD::FMUL:
+ case ISD::FMINNUM:
+ case ISD::FMAXNUM:
+ case ISD::SDIV:
+ case ISD::UDIV:
+ case ISD::FDIV:
+ case ISD::FPOW:
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR:
+ case ISD::SHL:
+ case ISD::SRA:
+ case ISD::SRL:
+ case ISD::UREM:
+ case ISD::SREM:
+ case ISD::FREM:
+ SplitVecRes_BinOp(N, Lo, Hi);
+ break;
+ case ISD::FMA:
+ SplitVecRes_TernaryOp(N, Lo, Hi);
+ break;
+ }
+
+ // If Lo/Hi is null, the sub-method took care of registering results etc.
+ if (Lo.getNode())
+ SetSplitVector(SDValue(N, ResNo), Lo, Hi);
+}
+
+void DAGTypeLegalizer::SplitVecRes_BinOp(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue LHSLo, LHSHi;
+ GetSplitVector(N->getOperand(0), LHSLo, LHSHi);
+ SDValue RHSLo, RHSHi;
+ GetSplitVector(N->getOperand(1), RHSLo, RHSHi);
+ SDLoc dl(N);
+
+ Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo, RHSLo);
+ Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi, RHSHi);
+}
+
+void DAGTypeLegalizer::SplitVecRes_TernaryOp(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue Op0Lo, Op0Hi;
+ GetSplitVector(N->getOperand(0), Op0Lo, Op0Hi);
+ SDValue Op1Lo, Op1Hi;
+ GetSplitVector(N->getOperand(1), Op1Lo, Op1Hi);
+ SDValue Op2Lo, Op2Hi;
+ GetSplitVector(N->getOperand(2), Op2Lo, Op2Hi);
+ SDLoc dl(N);
+
+ Lo = DAG.getNode(N->getOpcode(), dl, Op0Lo.getValueType(),
+ Op0Lo, Op1Lo, Op2Lo);
+ Hi = DAG.getNode(N->getOpcode(), dl, Op0Hi.getValueType(),
+ Op0Hi, Op1Hi, Op2Hi);
+}
+
+void DAGTypeLegalizer::SplitVecRes_BITCAST(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ // We know the result is a vector. The input may be either a vector or a
+ // scalar value.
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+ SDLoc dl(N);
+
+ SDValue InOp = N->getOperand(0);
+ EVT InVT = InOp.getValueType();
+
+ // Handle some special cases efficiently.
+ switch (getTypeAction(InVT)) {
+ case TargetLowering::TypeLegal:
+ case TargetLowering::TypePromoteInteger:
+ case TargetLowering::TypeSoftenFloat:
+ case TargetLowering::TypeScalarizeVector:
+ case TargetLowering::TypeWidenVector:
+ break;
+ case TargetLowering::TypeExpandInteger:
+ case TargetLowering::TypeExpandFloat:
+ // A scalar to vector conversion, where the scalar needs expansion.
+ // If the vector is being split in two then we can just convert the
+ // expanded pieces.
+ if (LoVT == HiVT) {
+ GetExpandedOp(InOp, Lo, Hi);
+ if (TLI.isBigEndian())
+ std::swap(Lo, Hi);
+ Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);
+ Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);
+ return;
+ }
+ break;
+ case TargetLowering::TypeSplitVector:
+ // If the input is a vector that needs to be split, convert each split
+ // piece of the input now.
+ GetSplitVector(InOp, Lo, Hi);
+ Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);
+ Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);
+ return;
+ }
+
+ // In the general case, convert the input to an integer and split it by hand.
+ EVT LoIntVT = EVT::getIntegerVT(*DAG.getContext(), LoVT.getSizeInBits());
+ EVT HiIntVT = EVT::getIntegerVT(*DAG.getContext(), HiVT.getSizeInBits());
+ if (TLI.isBigEndian())
+ std::swap(LoIntVT, HiIntVT);
+
+ SplitInteger(BitConvertToInteger(InOp), LoIntVT, HiIntVT, Lo, Hi);
+
+ if (TLI.isBigEndian())
+ std::swap(Lo, Hi);
+ Lo = DAG.getNode(ISD::BITCAST, dl, LoVT, Lo);
+ Hi = DAG.getNode(ISD::BITCAST, dl, HiVT, Hi);
+}
+
+void DAGTypeLegalizer::SplitVecRes_BUILD_VECTOR(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ EVT LoVT, HiVT;
+ SDLoc dl(N);
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+ unsigned LoNumElts = LoVT.getVectorNumElements();
+ SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+LoNumElts);
+ Lo = DAG.getNode(ISD::BUILD_VECTOR, dl, LoVT, LoOps);
+
+ SmallVector<SDValue, 8> HiOps(N->op_begin()+LoNumElts, N->op_end());
+ Hi = DAG.getNode(ISD::BUILD_VECTOR, dl, HiVT, HiOps);
+}
+
+void DAGTypeLegalizer::SplitVecRes_CONCAT_VECTORS(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ assert(!(N->getNumOperands() & 1) && "Unsupported CONCAT_VECTORS");
+ SDLoc dl(N);
+ unsigned NumSubvectors = N->getNumOperands() / 2;
+ if (NumSubvectors == 1) {
+ Lo = N->getOperand(0);
+ Hi = N->getOperand(1);
+ return;
+ }
+
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+
+ SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+NumSubvectors);
+ Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, LoVT, LoOps);
+
+ SmallVector<SDValue, 8> HiOps(N->op_begin()+NumSubvectors, N->op_end());
+ Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HiVT, HiOps);
+}
+
+void DAGTypeLegalizer::SplitVecRes_EXTRACT_SUBVECTOR(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue Vec = N->getOperand(0);
+ SDValue Idx = N->getOperand(1);
+ SDLoc dl(N);
+
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+
+ Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, LoVT, Vec, Idx);
+ uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+ Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, HiVT, Vec,
+ DAG.getConstant(IdxVal + LoVT.getVectorNumElements(),
+ TLI.getVectorIdxTy()));
+}
+
+void DAGTypeLegalizer::SplitVecRes_INSERT_SUBVECTOR(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue Vec = N->getOperand(0);
+ SDValue SubVec = N->getOperand(1);
+ SDValue Idx = N->getOperand(2);
+ SDLoc dl(N);
+ GetSplitVector(Vec, Lo, Hi);
+
+ // Spill the vector to the stack.
+ EVT VecVT = Vec.getValueType();
+ EVT SubVecVT = VecVT.getVectorElementType();
+ SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
+ SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
+ MachinePointerInfo(), false, false, 0);
+
+ // Store the new subvector into the specified index.
+ SDValue SubVecPtr = GetVectorElementPointer(StackPtr, SubVecVT, Idx);
+ Type *VecType = VecVT.getTypeForEVT(*DAG.getContext());
+ unsigned Alignment = TLI.getDataLayout()->getPrefTypeAlignment(VecType);
+ Store = DAG.getStore(Store, dl, SubVec, SubVecPtr, MachinePointerInfo(),
+ false, false, 0);
+
+ // Load the Lo part from the stack slot.
+ Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),
+ false, false, false, 0);
+
+ // Increment the pointer to the other part.
+ unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8;
+ StackPtr =
+ DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
+ DAG.getConstant(IncrementSize, StackPtr.getValueType()));
+
+ // Load the Hi part from the stack slot.
+ Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),
+ false, false, false, MinAlign(Alignment, IncrementSize));
+}
+
+void DAGTypeLegalizer::SplitVecRes_FPOWI(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDLoc dl(N);
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+ Lo = DAG.getNode(ISD::FPOWI, dl, Lo.getValueType(), Lo, N->getOperand(1));
+ Hi = DAG.getNode(ISD::FPOWI, dl, Hi.getValueType(), Hi, N->getOperand(1));
+}
+
+void DAGTypeLegalizer::SplitVecRes_InregOp(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue LHSLo, LHSHi;
+ GetSplitVector(N->getOperand(0), LHSLo, LHSHi);
+ SDLoc dl(N);
+
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) =
+ DAG.GetSplitDestVTs(cast<VTSDNode>(N->getOperand(1))->getVT());
+
+ Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo,
+ DAG.getValueType(LoVT));
+ Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi,
+ DAG.getValueType(HiVT));
+}
+
+void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDValue Vec = N->getOperand(0);
+ SDValue Elt = N->getOperand(1);
+ SDValue Idx = N->getOperand(2);
+ SDLoc dl(N);
+ GetSplitVector(Vec, Lo, Hi);
+
+ if (ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx)) {
+ unsigned IdxVal = CIdx->getZExtValue();
+ unsigned LoNumElts = Lo.getValueType().getVectorNumElements();
+ if (IdxVal < LoNumElts)
+ Lo = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
+ Lo.getValueType(), Lo, Elt, Idx);
+ else
+ Hi = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, Hi.getValueType(), Hi, Elt,
+ DAG.getConstant(IdxVal - LoNumElts,
+ TLI.getVectorIdxTy()));
+ return;
+ }
+
+ // See if the target wants to custom expand this node.
+ if (CustomLowerNode(N, N->getValueType(0), true))
+ return;
+
+ // Spill the vector to the stack.
+ EVT VecVT = Vec.getValueType();
+ EVT EltVT = VecVT.getVectorElementType();
+ SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
+ SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
+ MachinePointerInfo(), false, false, 0);
+
+ // Store the new element. This may be larger than the vector element type,
+ // so use a truncating store.
+ SDValue EltPtr = GetVectorElementPointer(StackPtr, EltVT, Idx);
+ Type *VecType = VecVT.getTypeForEVT(*DAG.getContext());
+ unsigned Alignment =
+ TLI.getDataLayout()->getPrefTypeAlignment(VecType);
+ Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, MachinePointerInfo(), EltVT,
+ false, false, 0);
+
+ // Load the Lo part from the stack slot.
+ Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),
+ false, false, false, 0);
+
+ // Increment the pointer to the other part.
+ unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8;
+ StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
+ DAG.getConstant(IncrementSize, StackPtr.getValueType()));
+
+ // Load the Hi part from the stack slot.
+ Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, MachinePointerInfo(),
+ false, false, false, MinAlign(Alignment, IncrementSize));
+}
+
+void DAGTypeLegalizer::SplitVecRes_SCALAR_TO_VECTOR(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ EVT LoVT, HiVT;
+ SDLoc dl(N);
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+ Lo = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoVT, N->getOperand(0));
+ Hi = DAG.getUNDEF(HiVT);
+}
+
+void DAGTypeLegalizer::SplitVecRes_LOAD(LoadSDNode *LD, SDValue &Lo,
+ SDValue &Hi) {
+ assert(ISD::isUNINDEXEDLoad(LD) && "Indexed load during type legalization!");
+ EVT LoVT, HiVT;
+ SDLoc dl(LD);
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(LD->getValueType(0));
+
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+ SDValue Ch = LD->getChain();
+ SDValue Ptr = LD->getBasePtr();
+ SDValue Offset = DAG.getUNDEF(Ptr.getValueType());
+ EVT MemoryVT = LD->getMemoryVT();
+ unsigned Alignment = LD->getOriginalAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+ bool isInvariant = LD->isInvariant();
+ AAMDNodes AAInfo = LD->getAAInfo();
+
+ EVT LoMemVT, HiMemVT;
+ std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
+
+ Lo = DAG.getLoad(ISD::UNINDEXED, ExtType, LoVT, dl, Ch, Ptr, Offset,
+ LD->getPointerInfo(), LoMemVT, isVolatile, isNonTemporal,
+ isInvariant, Alignment, AAInfo);
+
+ unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getConstant(IncrementSize, Ptr.getValueType()));
+ Hi = DAG.getLoad(ISD::UNINDEXED, ExtType, HiVT, dl, Ch, Ptr, Offset,
+ LD->getPointerInfo().getWithOffset(IncrementSize),
+ HiMemVT, isVolatile, isNonTemporal, isInvariant, Alignment,
+ AAInfo);
+
+ // Build a factor node to remember that this load is independent of the
+ // other one.
+ Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+ Hi.getValue(1));
+
+ // Legalized the chain result - switch anything that used the old chain to
+ // use the new one.
+ ReplaceValueWith(SDValue(LD, 1), Ch);
+}
+
+void DAGTypeLegalizer::SplitVecRes_MLOAD(MaskedLoadSDNode *MLD,
+ SDValue &Lo, SDValue &Hi) {
+ EVT LoVT, HiVT;
+ SDLoc dl(MLD);
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
+
+ SDValue Ch = MLD->getChain();
+ SDValue Ptr = MLD->getBasePtr();
+ SDValue Mask = MLD->getMask();
+ unsigned Alignment = MLD->getOriginalAlignment();
+
+ // if Alignment is equal to the vector size,
+ // take the half of it for the second part
+ unsigned SecondHalfAlignment =
+ (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
+ Alignment/2 : Alignment;
+
+ SDValue MaskLo, MaskHi;
+ std::tie(MaskLo, MaskHi) = DAG.SplitVector(Mask, dl);
+
+ EVT MemoryVT = MLD->getMemoryVT();
+ EVT LoMemVT, HiMemVT;
+ std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
+
+ SDValue Src0 = MLD->getSrc0();
+ SDValue Src0Lo, Src0Hi;
+ std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, dl);
+
+ MachineMemOperand *MMO = DAG.getMachineFunction().
+ getMachineMemOperand(MLD->getPointerInfo(),
+ MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
+ Alignment, MLD->getAAInfo(), MLD->getRanges());
+
+ Lo = DAG.getMaskedLoad(LoVT, dl, Ch, Ptr, MaskLo, Src0Lo, MMO);
+
+ unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
+ DAG.getConstant(IncrementSize, Ptr.getValueType()));
+
+ MMO = DAG.getMachineFunction().
+ getMachineMemOperand(MLD->getPointerInfo(),
+ MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
+ SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
+
+ Hi = DAG.getMaskedLoad(HiVT, dl, Ch, Ptr, MaskHi, Src0Hi, MMO);
+
+
+ // Build a factor node to remember that this load is independent of the
+ // other one.
+ Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
+ Hi.getValue(1));
+
+ // Legalized the chain result - switch anything that used the old chain to
+ // use the new one.
+ ReplaceValueWith(SDValue(MLD, 1), Ch);
+
+}
+
+void DAGTypeLegalizer::SplitVecRes_SETCC(SDNode *N, SDValue &Lo, SDValue &Hi) {
+ assert(N->getValueType(0).isVector() &&
+ N->getOperand(0).getValueType().isVector() &&
+ "Operand types must be vectors");
+
+ EVT LoVT, HiVT;
+ SDLoc DL(N);
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+
+ // Split the input.
+ SDValue LL, LH, RL, RH;
+ std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
+ std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
+
+ Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
+ Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
+}
+
+void DAGTypeLegalizer::SplitVecRes_UnaryOp(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ // Get the dest types - they may not match the input types, e.g. int_to_fp.
+ EVT LoVT, HiVT;
+ SDLoc dl(N);
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+
+ // If the input also splits, handle it directly for a compile time speedup.
+ // Otherwise split it by hand.
+ EVT InVT = N->getOperand(0).getValueType();
+ if (getTypeAction(InVT) == TargetLowering::TypeSplitVector)
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+ else
+ std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0);
+
+ if (N->getOpcode() == ISD::FP_ROUND) {
+ Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo, N->getOperand(1));
+ Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi, N->getOperand(1));
+ } else if (N->getOpcode() == ISD::CONVERT_RNDSAT) {
+ SDValue DTyOpLo = DAG.getValueType(LoVT);
+ SDValue DTyOpHi = DAG.getValueType(HiVT);
+ SDValue STyOpLo = DAG.getValueType(Lo.getValueType());
+ SDValue STyOpHi = DAG.getValueType(Hi.getValueType());
+ SDValue RndOp = N->getOperand(3);
+ SDValue SatOp = N->getOperand(4);
+ ISD::CvtCode CvtCode = cast<CvtRndSatSDNode>(N)->getCvtCode();
+ Lo = DAG.getConvertRndSat(LoVT, dl, Lo, DTyOpLo, STyOpLo, RndOp, SatOp,
+ CvtCode);
+ Hi = DAG.getConvertRndSat(HiVT, dl, Hi, DTyOpHi, STyOpHi, RndOp, SatOp,
+ CvtCode);
+ } else {
+ Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo);
+ Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi);
+ }
+}
+
+void DAGTypeLegalizer::SplitVecRes_ExtendOp(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ SDLoc dl(N);
+ EVT SrcVT = N->getOperand(0).getValueType();
+ EVT DestVT = N->getValueType(0);
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(DestVT);
+
+ // We can do better than a generic split operation if the extend is doing
+ // more than just doubling the width of the elements and the following are
+ // true:
+ // - The number of vector elements is even,
+ // - the source type is legal,
+ // - the type of a split source is illegal,
+ // - the type of an extended (by doubling element size) source is legal, and
+ // - the type of that extended source when split is legal.
+ //
+ // This won't necessarily completely legalize the operation, but it will
+ // more effectively move in the right direction and prevent falling down
+ // to scalarization in many cases due to the input vector being split too
+ // far.
+ unsigned NumElements = SrcVT.getVectorNumElements();
+ if ((NumElements & 1) == 0 &&
+ SrcVT.getSizeInBits() * 2 < DestVT.getSizeInBits()) {
+ LLVMContext &Ctx = *DAG.getContext();
+ EVT NewSrcVT = EVT::getVectorVT(
+ Ctx, EVT::getIntegerVT(
+ Ctx, SrcVT.getVectorElementType().getSizeInBits() * 2),
+ NumElements);
+ EVT SplitSrcVT =
+ EVT::getVectorVT(Ctx, SrcVT.getVectorElementType(), NumElements / 2);
+ EVT SplitLoVT, SplitHiVT;
+ std::tie(SplitLoVT, SplitHiVT) = DAG.GetSplitDestVTs(NewSrcVT);
+ if (TLI.isTypeLegal(SrcVT) && !TLI.isTypeLegal(SplitSrcVT) &&
+ TLI.isTypeLegal(NewSrcVT) && TLI.isTypeLegal(SplitLoVT)) {
+ DEBUG(dbgs() << "Split vector extend via incremental extend:";
+ N->dump(&DAG); dbgs() << "\n");
+ // Extend the source vector by one step.
+ SDValue NewSrc =
+ DAG.getNode(N->getOpcode(), dl, NewSrcVT, N->getOperand(0));
+ // Get the low and high halves of the new, extended one step, vector.
+ std::tie(Lo, Hi) = DAG.SplitVector(NewSrc, dl);
+ // Extend those vector halves the rest of the way.
+ Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo);
+ Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi);
+ return;
+ }
+ }
+ // Fall back to the generic unary operator splitting otherwise.
+ SplitVecRes_UnaryOp(N, Lo, Hi);
+}
+
+void DAGTypeLegalizer::SplitVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N,
+ SDValue &Lo, SDValue &Hi) {
+ // The low and high parts of the original input give four input vectors.
+ SDValue Inputs[4];
+ SDLoc dl(N);
+ GetSplitVector(N->getOperand(0), Inputs[0], Inputs[1]);
+ GetSplitVector(N->getOperand(1), Inputs[2], Inputs[3]);
+ EVT NewVT = Inputs[0].getValueType();
+ unsigned NewElts = NewVT.getVectorNumElements();
+
+ // If Lo or Hi uses elements from at most two of the four input vectors, then
+ // express it as a vector shuffle of those two inputs. Otherwise extract the
+ // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR.
+ SmallVector<int, 16> Ops;
+ for (unsigned High = 0; High < 2; ++High) {
+ SDValue &Output = High ? Hi : Lo;
+
+ // Build a shuffle mask for the output, discovering on the fly which
+ // input vectors to use as shuffle operands (recorded in InputUsed).
+ // If building a suitable shuffle vector proves too hard, then bail
+ // out with useBuildVector set.
+ unsigned InputUsed[2] = { -1U, -1U }; // Not yet discovered.
+ unsigned FirstMaskIdx = High * NewElts;
+ bool useBuildVector = false;
+ for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) {
+ // The mask element. This indexes into the input.
+ int Idx = N->getMaskElt(FirstMaskIdx + MaskOffset);
+
+ // The input vector this mask element indexes into.
+ unsigned Input = (unsigned)Idx / NewElts;
+
+ if (Input >= array_lengthof(Inputs)) {
+ // The mask element does not index into any input vector.
+ Ops.push_back(-1);
+ continue;
+ }
+
+ // Turn the index into an offset from the start of the input vector.
+ Idx -= Input * NewElts;
+
+ // Find or create a shuffle vector operand to hold this input.
+ unsigned OpNo;
+ for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
+ if (InputUsed[OpNo] == Input) {
+ // This input vector is already an operand.
+ break;
+ } else if (InputUsed[OpNo] == -1U) {
+ // Create a new operand for this input vector.
+ InputUsed[OpNo] = Input;
+ break;
+ }
+ }
+
+ if (OpNo >= array_lengthof(InputUsed)) {
+ // More than two input vectors used! Give up on trying to create a
+ // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
+ useBuildVector = true;
+ break;
+ }
+
+ // Add the mask index for the new shuffle vector.
+ Ops.push_back(Idx + OpNo * NewElts);
+ }
+
+ if (useBuildVector) {
+ EVT EltVT = NewVT.getVectorElementType();
+ SmallVector<SDValue, 16> SVOps;
+
+ // Extract the input elements by hand.
+ for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) {
+ // The mask element. This indexes into the input.
+ int Idx = N->getMaskElt(FirstMaskIdx + MaskOffset);
+
+ // The input vector this mask element indexes into.
+ unsigned Input = (unsigned)Idx / NewElts;
+
+ if (Input >= array_lengthof(Inputs)) {
+ // The mask element is "undef" or indexes off the end of the input.
+ SVOps.push_back(DAG.getUNDEF(EltVT));
+ continue;
+ }
+
+ // Turn the index into an offset from the start of the input vector.
+ Idx -= Input * NewElts;
+
+ // Extract the vector element by hand.
+ SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
+ Inputs[Input], DAG.getConstant(Idx,
+ TLI.getVectorIdxTy())));
+ }
+
+ // Construct the Lo/Hi output using a BUILD_VECTOR.
+ Output = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, SVOps);
+ } else if (InputUsed[0] == -1U) {
+ // No input vectors were used! The result is undefined.
+ Output = DAG.getUNDEF(NewVT);
+ } else {
+ SDValue Op0 = Inputs[InputUsed[0]];
+ // If only one input was used, use an undefined vector for the other.
+ SDValue Op1 = InputUsed[1] == -1U ?
+ DAG.getUNDEF(NewVT) : Inputs[InputUsed[1]];
+ // At least one input vector was used. Create a new shuffle vector.
+ Output = DAG.getVectorShuffle(NewVT, dl, Op0, Op1, &Ops[0]);
+ }
+
+ Ops.clear();
+ }
+}
+
+
+//===----------------------------------------------------------------------===//
+// Operand Vector Splitting
+//===----------------------------------------------------------------------===//
+
+/// SplitVectorOperand - This method is called when the specified operand of the
+/// specified node is found to need vector splitting. At this point, all of the
+/// result types of the node are known to be legal, but other operands of the
+/// node may need legalization as well as the specified one.
+bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) {
+ DEBUG(dbgs() << "Split node operand: ";
+ N->dump(&DAG);
+ dbgs() << "\n");
+ SDValue Res = SDValue();
+
+ // See if the target wants to custom split this node.
+ if (CustomLowerNode(N, N->getOperand(OpNo).getValueType(), false))
+ return false;
+
+ if (!Res.getNode()) {
+ switch (N->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "SplitVectorOperand Op #" << OpNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n";
+#endif
+ report_fatal_error("Do not know how to split this operator's "
+ "operand!\n");
+
+ case ISD::SETCC: Res = SplitVecOp_VSETCC(N); break;
+ case ISD::BITCAST: Res = SplitVecOp_BITCAST(N); break;
+ case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break;
+ case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break;
+ case ISD::CONCAT_VECTORS: Res = SplitVecOp_CONCAT_VECTORS(N); break;
+ case ISD::TRUNCATE: Res = SplitVecOp_TRUNCATE(N); break;
+ case ISD::FP_ROUND: Res = SplitVecOp_FP_ROUND(N); break;
+ case ISD::STORE:
+ Res = SplitVecOp_STORE(cast<StoreSDNode>(N), OpNo);
+ break;
+ case ISD::MSTORE:
+ Res = SplitVecOp_MSTORE(cast<MaskedStoreSDNode>(N), OpNo);
+ break;
+ case ISD::VSELECT:
+ Res = SplitVecOp_VSELECT(N, OpNo);
+ break;
+ case ISD::CTTZ:
+ case ISD::CTLZ:
+ case ISD::CTPOP:
+ case ISD::FP_EXTEND:
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT:
+ case ISD::SINT_TO_FP:
+ case ISD::UINT_TO_FP:
+ case ISD::FTRUNC:
+ case ISD::SIGN_EXTEND:
+ case ISD::ZERO_EXTEND:
+ case ISD::ANY_EXTEND:
+ Res = SplitVecOp_UnaryOp(N);
+ break;
+ }
+ }
+
+ // If the result is null, the sub-method took care of registering results etc.
+ if (!Res.getNode()) return false;
+
+ // If the result is N, the sub-method updated N in place. Tell the legalizer
+ // core about this.
+ if (Res.getNode() == N)
+ return true;
+
+ assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&
+ "Invalid operand expansion");
+
+ ReplaceValueWith(SDValue(N, 0), Res);
+ return false;
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_VSELECT(SDNode *N, unsigned OpNo) {
+ // The only possibility for an illegal operand is the mask, since result type
+ // legalization would have handled this node already otherwise.
+ assert(OpNo == 0 && "Illegal operand must be mask");
+
+ SDValue Mask = N->getOperand(0);
+ SDValue Src0 = N->getOperand(1);
+ SDValue Src1 = N->getOperand(2);
+ EVT Src0VT = Src0.getValueType();
+ SDLoc DL(N);
+ assert(Mask.getValueType().isVector() && "VSELECT without a vector mask?");
+
+ SDValue Lo, Hi;
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+ assert(Lo.getValueType() == Hi.getValueType() &&
+ "Lo and Hi have differing types");
+
+ EVT LoOpVT, HiOpVT;
+ std::tie(LoOpVT, HiOpVT) = DAG.GetSplitDestVTs(Src0VT);
+ assert(LoOpVT == HiOpVT && "Asymmetric vector split?");
+
+ SDValue LoOp0, HiOp0, LoOp1, HiOp1, LoMask, HiMask;
+ std::tie(LoOp0, HiOp0) = DAG.SplitVector(Src0, DL);
+ std::tie(LoOp1, HiOp1) = DAG.SplitVector(Src1, DL);
+ std::tie(LoMask, HiMask) = DAG.SplitVector(Mask, DL);
+
+ SDValue LoSelect =
+ DAG.getNode(ISD::VSELECT, DL, LoOpVT, LoMask, LoOp0, LoOp1);
+ SDValue HiSelect =
+ DAG.getNode(ISD::VSELECT, DL, HiOpVT, HiMask, HiOp0, HiOp1);
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, Src0VT, LoSelect, HiSelect);
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_UnaryOp(SDNode *N) {
+ // The result has a legal vector type, but the input needs splitting.
+ EVT ResVT = N->getValueType(0);
+ SDValue Lo, Hi;
+ SDLoc dl(N);
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+ EVT InVT = Lo.getValueType();
+
+ EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(),
+ InVT.getVectorNumElements());
+
+ Lo = DAG.getNode(N->getOpcode(), dl, OutVT, Lo);
+ Hi = DAG.getNode(N->getOpcode(), dl, OutVT, Hi);
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_BITCAST(SDNode *N) {
+ // For example, i64 = BITCAST v4i16 on alpha. Typically the vector will
+ // end up being split all the way down to individual components. Convert the
+ // split pieces into integers and reassemble.
+ SDValue Lo, Hi;
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+ Lo = BitConvertToInteger(Lo);
+ Hi = BitConvertToInteger(Hi);
+
+ if (TLI.isBigEndian())
+ std::swap(Lo, Hi);
+
+ return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0),
+ JoinIntegers(Lo, Hi));
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_SUBVECTOR(SDNode *N) {
+ // We know that the extracted result type is legal.
+ EVT SubVT = N->getValueType(0);
+ SDValue Idx = N->getOperand(1);
+ SDLoc dl(N);
+ SDValue Lo, Hi;
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+
+ uint64_t LoElts = Lo.getValueType().getVectorNumElements();
+ uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+
+ if (IdxVal < LoElts) {
+ assert(IdxVal + SubVT.getVectorNumElements() <= LoElts &&
+ "Extracted subvector crosses vector split!");
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, Lo, Idx);
+ } else {
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, Hi,
+ DAG.getConstant(IdxVal - LoElts, Idx.getValueType()));
+ }
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N) {
+ SDValue Vec = N->getOperand(0);
+ SDValue Idx = N->getOperand(1);
+ EVT VecVT = Vec.getValueType();
+
+ if (isa<ConstantSDNode>(Idx)) {
+ uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+ assert(IdxVal < VecVT.getVectorNumElements() && "Invalid vector index!");
+
+ SDValue Lo, Hi;
+ GetSplitVector(Vec, Lo, Hi);
+
+ uint64_t LoElts = Lo.getValueType().getVectorNumElements();
+
+ if (IdxVal < LoElts)
+ return SDValue(DAG.UpdateNodeOperands(N, Lo, Idx), 0);
+ return SDValue(DAG.UpdateNodeOperands(N, Hi,
+ DAG.getConstant(IdxVal - LoElts,
+ Idx.getValueType())), 0);
+ }
+
+ // See if the target wants to custom expand this node.
+ if (CustomLowerNode(N, N->getValueType(0), true))
+ return SDValue();
+
+ // Store the vector to the stack.
+ EVT EltVT = VecVT.getVectorElementType();
+ SDLoc dl(N);
+ SDValue StackPtr = DAG.CreateStackTemporary(VecVT);
+ SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
+ MachinePointerInfo(), false, false, 0);
+
+ // Load back the required element.
+ StackPtr = GetVectorElementPointer(StackPtr, EltVT, Idx);
+ return DAG.getExtLoad(ISD::EXTLOAD, dl, N->getValueType(0), Store, StackPtr,
+ MachinePointerInfo(), EltVT, false, false, false, 0);
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_MSTORE(MaskedStoreSDNode *N,
+ unsigned OpNo) {
+ SDValue Ch = N->getChain();
+ SDValue Ptr = N->getBasePtr();
+ SDValue Mask = N->getMask();
+ SDValue Data = N->getData();
+ EVT MemoryVT = N->getMemoryVT();
+ unsigned Alignment = N->getOriginalAlignment();
+ SDLoc DL(N);
+
+ EVT LoMemVT, HiMemVT;
+ std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
+
+ SDValue DataLo, DataHi;
+ GetSplitVector(Data, DataLo, DataHi);
+ SDValue MaskLo, MaskHi;
+ GetSplitVector(Mask, MaskLo, MaskHi);
+
+ // if Alignment is equal to the vector size,
+ // take the half of it for the second part
+ unsigned SecondHalfAlignment =
+ (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
+ Alignment/2 : Alignment;
+
+ SDValue Lo, Hi;
+ MachineMemOperand *MMO = DAG.getMachineFunction().
+ getMachineMemOperand(N->getPointerInfo(),
+ MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
+ Alignment, N->getAAInfo(), N->getRanges());
+
+ Lo = DAG.getMaskedStore(Ch, DL, DataLo, Ptr, MaskLo, MMO);
+
+ unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
+ Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
+ DAG.getConstant(IncrementSize, Ptr.getValueType()));
+
+ MMO = DAG.getMachineFunction().
+ getMachineMemOperand(N->getPointerInfo(),
+ MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
+ SecondHalfAlignment, N->getAAInfo(), N->getRanges());
+
+ Hi = DAG.getMaskedStore(Ch, DL, DataHi, Ptr, MaskHi, MMO);
+
+
+ // Build a factor node to remember that this store is independent of the
+ // other one.
+ return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
+
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) {
+ assert(N->isUnindexed() && "Indexed store of vector?");
+ assert(OpNo == 1 && "Can only split the stored value");
+ SDLoc DL(N);
+
+ bool isTruncating = N->isTruncatingStore();
+ SDValue Ch = N->getChain();
+ SDValue Ptr = N->getBasePtr();
+ EVT MemoryVT = N->getMemoryVT();
+ unsigned Alignment = N->getOriginalAlignment();
+ bool isVol = N->isVolatile();
+ bool isNT = N->isNonTemporal();
+ AAMDNodes AAInfo = N->getAAInfo();
+ SDValue Lo, Hi;
+ GetSplitVector(N->getOperand(1), Lo, Hi);
+
+ EVT LoMemVT, HiMemVT;
+ std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
+
+ unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
+
+ if (isTruncating)
+ Lo = DAG.getTruncStore(Ch, DL, Lo, Ptr, N->getPointerInfo(),
+ LoMemVT, isVol, isNT, Alignment, AAInfo);
+ else
+ Lo = DAG.getStore(Ch, DL, Lo, Ptr, N->getPointerInfo(),
+ isVol, isNT, Alignment, AAInfo);
+
+ // Increment the pointer to the other half.
+ Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
+ DAG.getConstant(IncrementSize, Ptr.getValueType()));
+
+ if (isTruncating)
+ Hi = DAG.getTruncStore(Ch, DL, Hi, Ptr,
+ N->getPointerInfo().getWithOffset(IncrementSize),
+ HiMemVT, isVol, isNT, Alignment, AAInfo);
+ else
+ Hi = DAG.getStore(Ch, DL, Hi, Ptr,
+ N->getPointerInfo().getWithOffset(IncrementSize),
+ isVol, isNT, Alignment, AAInfo);
+
+ return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) {
+ SDLoc DL(N);
+
+ // The input operands all must have the same type, and we know the result
+ // type is valid. Convert this to a buildvector which extracts all the
+ // input elements.
+ // TODO: If the input elements are power-two vectors, we could convert this to
+ // a new CONCAT_VECTORS node with elements that are half-wide.
+ SmallVector<SDValue, 32> Elts;
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) {
+ SDValue Op = N->getOperand(op);
+ for (unsigned i = 0, e = Op.getValueType().getVectorNumElements();
+ i != e; ++i) {
+ Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT,
+ Op, DAG.getConstant(i, TLI.getVectorIdxTy())));
+
+ }
+ }
+
+ return DAG.getNode(ISD::BUILD_VECTOR, DL, N->getValueType(0), Elts);
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_TRUNCATE(SDNode *N) {
+ // The result type is legal, but the input type is illegal. If splitting
+ // ends up with the result type of each half still being legal, just
+ // do that. If, however, that would result in an illegal result type,
+ // we can try to get more clever with power-two vectors. Specifically,
+ // split the input type, but also widen the result element size, then
+ // concatenate the halves and truncate again. For example, consider a target
+ // where v8i8 is legal and v8i32 is not (ARM, which doesn't have 256-bit
+ // vectors). To perform a "%res = v8i8 trunc v8i32 %in" we do:
+ // %inlo = v4i32 extract_subvector %in, 0
+ // %inhi = v4i32 extract_subvector %in, 4
+ // %lo16 = v4i16 trunc v4i32 %inlo
+ // %hi16 = v4i16 trunc v4i32 %inhi
+ // %in16 = v8i16 concat_vectors v4i16 %lo16, v4i16 %hi16
+ // %res = v8i8 trunc v8i16 %in16
+ //
+ // Without this transform, the original truncate would end up being
+ // scalarized, which is pretty much always a last resort.
+ SDValue InVec = N->getOperand(0);
+ EVT InVT = InVec->getValueType(0);
+ EVT OutVT = N->getValueType(0);
+ unsigned NumElements = OutVT.getVectorNumElements();
+ // Widening should have already made sure this is a power-two vector
+ // if we're trying to split it at all. assert() that's true, just in case.
+ assert(!(NumElements & 1) && "Splitting vector, but not in half!");
+
+ unsigned InElementSize = InVT.getVectorElementType().getSizeInBits();
+ unsigned OutElementSize = OutVT.getVectorElementType().getSizeInBits();
+
+ // If the input elements are only 1/2 the width of the result elements,
+ // just use the normal splitting. Our trick only work if there's room
+ // to split more than once.
+ if (InElementSize <= OutElementSize * 2)
+ return SplitVecOp_UnaryOp(N);
+ SDLoc DL(N);
+
+ // Extract the halves of the input via extract_subvector.
+ SDValue InLoVec, InHiVec;
+ std::tie(InLoVec, InHiVec) = DAG.SplitVector(InVec, DL);
+ // Truncate them to 1/2 the element size.
+ EVT HalfElementVT = EVT::getIntegerVT(*DAG.getContext(), InElementSize/2);
+ EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), HalfElementVT,
+ NumElements/2);
+ SDValue HalfLo = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, InLoVec);
+ SDValue HalfHi = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, InHiVec);
+ // Concatenate them to get the full intermediate truncation result.
+ EVT InterVT = EVT::getVectorVT(*DAG.getContext(), HalfElementVT, NumElements);
+ SDValue InterVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, InterVT, HalfLo,
+ HalfHi);
+ // Now finish up by truncating all the way down to the original result
+ // type. This should normally be something that ends up being legal directly,
+ // but in theory if a target has very wide vectors and an annoyingly
+ // restricted set of legal types, this split can chain to build things up.
+ return DAG.getNode(ISD::TRUNCATE, DL, OutVT, InterVec);
+}
+
+SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
+ assert(N->getValueType(0).isVector() &&
+ N->getOperand(0).getValueType().isVector() &&
+ "Operand types must be vectors");
+ // The result has a legal vector type, but the input needs splitting.
+ SDValue Lo0, Hi0, Lo1, Hi1, LoRes, HiRes;
+ SDLoc DL(N);
+ GetSplitVector(N->getOperand(0), Lo0, Hi0);
+ GetSplitVector(N->getOperand(1), Lo1, Hi1);
+ unsigned PartElements = Lo0.getValueType().getVectorNumElements();
+ EVT PartResVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, PartElements);
+ EVT WideResVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, 2*PartElements);
+
+ LoRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Lo0, Lo1, N->getOperand(2));
+ HiRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Hi0, Hi1, N->getOperand(2));
+ SDValue Con = DAG.getNode(ISD::CONCAT_VECTORS, DL, WideResVT, LoRes, HiRes);
+ return PromoteTargetBoolean(Con, N->getValueType(0));
+}
+
+
+SDValue DAGTypeLegalizer::SplitVecOp_FP_ROUND(SDNode *N) {
+ // The result has a legal vector type, but the input needs splitting.
+ EVT ResVT = N->getValueType(0);
+ SDValue Lo, Hi;
+ SDLoc DL(N);
+ GetSplitVector(N->getOperand(0), Lo, Hi);
+ EVT InVT = Lo.getValueType();
+
+ EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(),
+ InVT.getVectorNumElements());
+
+ Lo = DAG.getNode(ISD::FP_ROUND, DL, OutVT, Lo, N->getOperand(1));
+ Hi = DAG.getNode(ISD::FP_ROUND, DL, OutVT, Hi, N->getOperand(1));
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
+}
+
+
+
+//===----------------------------------------------------------------------===//
+// Result Vector Widening
+//===----------------------------------------------------------------------===//
+
+void DAGTypeLegalizer::WidenVectorResult(SDNode *N, unsigned ResNo) {
+ DEBUG(dbgs() << "Widen node result " << ResNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n");
+
+ // See if the target wants to custom widen this node.
+ if (CustomWidenLowerNode(N, N->getValueType(ResNo)))
+ return;
+
+ SDValue Res = SDValue();
+ switch (N->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "WidenVectorResult #" << ResNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n";
+#endif
+ llvm_unreachable("Do not know how to widen the result of this operator!");
+
+ case ISD::MERGE_VALUES: Res = WidenVecRes_MERGE_VALUES(N, ResNo); break;
+ case ISD::BITCAST: Res = WidenVecRes_BITCAST(N); break;
+ case ISD::BUILD_VECTOR: Res = WidenVecRes_BUILD_VECTOR(N); break;
+ case ISD::CONCAT_VECTORS: Res = WidenVecRes_CONCAT_VECTORS(N); break;
+ case ISD::CONVERT_RNDSAT: Res = WidenVecRes_CONVERT_RNDSAT(N); break;
+ case ISD::EXTRACT_SUBVECTOR: Res = WidenVecRes_EXTRACT_SUBVECTOR(N); break;
+ case ISD::FP_ROUND_INREG: Res = WidenVecRes_InregOp(N); break;
+ case ISD::INSERT_VECTOR_ELT: Res = WidenVecRes_INSERT_VECTOR_ELT(N); break;
+ case ISD::LOAD: Res = WidenVecRes_LOAD(N); break;
+ case ISD::SCALAR_TO_VECTOR: Res = WidenVecRes_SCALAR_TO_VECTOR(N); break;
+ case ISD::SIGN_EXTEND_INREG: Res = WidenVecRes_InregOp(N); break;
+ case ISD::VSELECT:
+ case ISD::SELECT: Res = WidenVecRes_SELECT(N); break;
+ case ISD::SELECT_CC: Res = WidenVecRes_SELECT_CC(N); break;
+ case ISD::SETCC: Res = WidenVecRes_SETCC(N); break;
+ case ISD::UNDEF: Res = WidenVecRes_UNDEF(N); break;
+ case ISD::VECTOR_SHUFFLE:
+ Res = WidenVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N));
+ break;
+
+ case ISD::ADD:
+ case ISD::AND:
+ case ISD::MUL:
+ case ISD::MULHS:
+ case ISD::MULHU:
+ case ISD::OR:
+ case ISD::SUB:
+ case ISD::XOR:
+ case ISD::FMINNUM:
+ case ISD::FMAXNUM:
+ Res = WidenVecRes_Binary(N);
+ break;
+
+ case ISD::FADD:
+ case ISD::FCOPYSIGN:
+ case ISD::FMUL:
+ case ISD::FPOW:
+ case ISD::FSUB:
+ case ISD::FDIV:
+ case ISD::FREM:
+ case ISD::SDIV:
+ case ISD::UDIV:
+ case ISD::SREM:
+ case ISD::UREM:
+ Res = WidenVecRes_BinaryCanTrap(N);
+ break;
+
+ case ISD::FPOWI:
+ Res = WidenVecRes_POWI(N);
+ break;
+
+ case ISD::SHL:
+ case ISD::SRA:
+ case ISD::SRL:
+ Res = WidenVecRes_Shift(N);
+ break;
+
+ case ISD::ANY_EXTEND:
+ case ISD::FP_EXTEND:
+ case ISD::FP_ROUND:
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT:
+ case ISD::SIGN_EXTEND:
+ case ISD::SINT_TO_FP:
+ case ISD::TRUNCATE:
+ case ISD::UINT_TO_FP:
+ case ISD::ZERO_EXTEND:
+ Res = WidenVecRes_Convert(N);
+ break;
+
+ case ISD::BSWAP:
+ case ISD::CTLZ:
+ case ISD::CTPOP:
+ case ISD::CTTZ:
+ case ISD::FABS:
+ case ISD::FCEIL:
+ case ISD::FCOS:
+ case ISD::FEXP:
+ case ISD::FEXP2:
+ case ISD::FFLOOR:
+ case ISD::FLOG:
+ case ISD::FLOG10:
+ case ISD::FLOG2:
+ case ISD::FNEARBYINT:
+ case ISD::FNEG:
+ case ISD::FRINT:
+ case ISD::FROUND:
+ case ISD::FSIN:
+ case ISD::FSQRT:
+ case ISD::FTRUNC:
+ Res = WidenVecRes_Unary(N);
+ break;
+ case ISD::FMA:
+ Res = WidenVecRes_Ternary(N);
+ break;
+ }
+
+ // If Res is null, the sub-method took care of registering the result.
+ if (Res.getNode())
+ SetWidenedVector(SDValue(N, ResNo), Res);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_Ternary(SDNode *N) {
+ // Ternary op widening.
+ SDLoc dl(N);
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ SDValue InOp3 = GetWidenedVector(N->getOperand(2));
+ return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2, InOp3);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) {
+ // Binary op widening.
+ SDLoc dl(N);
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_BinaryCanTrap(SDNode *N) {
+ // Binary op widening for operations that can trap.
+ unsigned Opcode = N->getOpcode();
+ SDLoc dl(N);
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ EVT WidenEltVT = WidenVT.getVectorElementType();
+ EVT VT = WidenVT;
+ unsigned NumElts = VT.getVectorNumElements();
+ while (!TLI.isTypeLegal(VT) && NumElts != 1) {
+ NumElts = NumElts / 2;
+ VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts);
+ }
+
+ if (NumElts != 1 && !TLI.canOpTrap(N->getOpcode(), VT)) {
+ // Operation doesn't trap so just widen as normal.
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2);
+ }
+
+ // No legal vector version so unroll the vector operation and then widen.
+ if (NumElts == 1)
+ return DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements());
+
+ // Since the operation can trap, apply operation on the original vector.
+ EVT MaxVT = VT;
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ unsigned CurNumElts = N->getValueType(0).getVectorNumElements();
+
+ SmallVector<SDValue, 16> ConcatOps(CurNumElts);
+ unsigned ConcatEnd = 0; // Current ConcatOps index.
+ int Idx = 0; // Current Idx into input vectors.
+
+ // NumElts := greatest legal vector size (at most WidenVT)
+ // while (orig. vector has unhandled elements) {
+ // take munches of size NumElts from the beginning and add to ConcatOps
+ // NumElts := next smaller supported vector size or 1
+ // }
+ while (CurNumElts != 0) {
+ while (CurNumElts >= NumElts) {
+ SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1,
+ DAG.getConstant(Idx, TLI.getVectorIdxTy()));
+ SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2,
+ DAG.getConstant(Idx, TLI.getVectorIdxTy()));
+ ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2);
+ Idx += NumElts;
+ CurNumElts -= NumElts;
+ }
+ do {
+ NumElts = NumElts / 2;
+ VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts);
+ } while (!TLI.isTypeLegal(VT) && NumElts != 1);
+
+ if (NumElts == 1) {
+ for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) {
+ SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT,
+ InOp1, DAG.getConstant(Idx,
+ TLI.getVectorIdxTy()));
+ SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT,
+ InOp2, DAG.getConstant(Idx,
+ TLI.getVectorIdxTy()));
+ ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, WidenEltVT,
+ EOp1, EOp2);
+ }
+ CurNumElts = 0;
+ }
+ }
+
+ // Check to see if we have a single operation with the widen type.
+ if (ConcatEnd == 1) {
+ VT = ConcatOps[0].getValueType();
+ if (VT == WidenVT)
+ return ConcatOps[0];
+ }
+
+ // while (Some element of ConcatOps is not of type MaxVT) {
+ // From the end of ConcatOps, collect elements of the same type and put
+ // them into an op of the next larger supported type
+ // }
+ while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) {
+ Idx = ConcatEnd - 1;
+ VT = ConcatOps[Idx--].getValueType();
+ while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT)
+ Idx--;
+
+ int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1;
+ EVT NextVT;
+ do {
+ NextSize *= 2;
+ NextVT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NextSize);
+ } while (!TLI.isTypeLegal(NextVT));
+
+ if (!VT.isVector()) {
+ // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT
+ SDValue VecOp = DAG.getUNDEF(NextVT);
+ unsigned NumToInsert = ConcatEnd - Idx - 1;
+ for (unsigned i = 0, OpIdx = Idx+1; i < NumToInsert; i++, OpIdx++) {
+ VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NextVT, VecOp,
+ ConcatOps[OpIdx], DAG.getConstant(i,
+ TLI.getVectorIdxTy()));
+ }
+ ConcatOps[Idx+1] = VecOp;
+ ConcatEnd = Idx + 2;
+ } else {
+ // Vector type, create a CONCAT_VECTORS of type NextVT
+ SDValue undefVec = DAG.getUNDEF(VT);
+ unsigned OpsToConcat = NextSize/VT.getVectorNumElements();
+ SmallVector<SDValue, 16> SubConcatOps(OpsToConcat);
+ unsigned RealVals = ConcatEnd - Idx - 1;
+ unsigned SubConcatEnd = 0;
+ unsigned SubConcatIdx = Idx + 1;
+ while (SubConcatEnd < RealVals)
+ SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx];
+ while (SubConcatEnd < OpsToConcat)
+ SubConcatOps[SubConcatEnd++] = undefVec;
+ ConcatOps[SubConcatIdx] = DAG.getNode(ISD::CONCAT_VECTORS, dl,
+ NextVT, SubConcatOps);
+ ConcatEnd = SubConcatIdx + 1;
+ }
+ }
+
+ // Check to see if we have a single operation with the widen type.
+ if (ConcatEnd == 1) {
+ VT = ConcatOps[0].getValueType();
+ if (VT == WidenVT)
+ return ConcatOps[0];
+ }
+
+ // add undefs of size MaxVT until ConcatOps grows to length of WidenVT
+ unsigned NumOps = WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements();
+ if (NumOps != ConcatEnd ) {
+ SDValue UndefVal = DAG.getUNDEF(MaxVT);
+ for (unsigned j = ConcatEnd; j < NumOps; ++j)
+ ConcatOps[j] = UndefVal;
+ }
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT,
+ makeArrayRef(ConcatOps.data(), NumOps));
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) {
+ SDValue InOp = N->getOperand(0);
+ SDLoc DL(N);
+
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+
+ EVT InVT = InOp.getValueType();
+ EVT InEltVT = InVT.getVectorElementType();
+ EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts);
+
+ unsigned Opcode = N->getOpcode();
+ unsigned InVTNumElts = InVT.getVectorNumElements();
+
+ if (getTypeAction(InVT) == TargetLowering::TypeWidenVector) {
+ InOp = GetWidenedVector(N->getOperand(0));
+ InVT = InOp.getValueType();
+ InVTNumElts = InVT.getVectorNumElements();
+ if (InVTNumElts == WidenNumElts) {
+ if (N->getNumOperands() == 1)
+ return DAG.getNode(Opcode, DL, WidenVT, InOp);
+ return DAG.getNode(Opcode, DL, WidenVT, InOp, N->getOperand(1));
+ }
+ }
+
+ if (TLI.isTypeLegal(InWidenVT)) {
+ // Because the result and the input are different vector types, widening
+ // the result could create a legal type but widening the input might make
+ // it an illegal type that might lead to repeatedly splitting the input
+ // and then widening it. To avoid this, we widen the input only if
+ // it results in a legal type.
+ if (WidenNumElts % InVTNumElts == 0) {
+ // Widen the input and call convert on the widened input vector.
+ unsigned NumConcat = WidenNumElts/InVTNumElts;
+ SmallVector<SDValue, 16> Ops(NumConcat);
+ Ops[0] = InOp;
+ SDValue UndefVal = DAG.getUNDEF(InVT);
+ for (unsigned i = 1; i != NumConcat; ++i)
+ Ops[i] = UndefVal;
+ SDValue InVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, InWidenVT, Ops);
+ if (N->getNumOperands() == 1)
+ return DAG.getNode(Opcode, DL, WidenVT, InVec);
+ return DAG.getNode(Opcode, DL, WidenVT, InVec, N->getOperand(1));
+ }
+
+ if (InVTNumElts % WidenNumElts == 0) {
+ SDValue InVal = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InWidenVT,
+ InOp, DAG.getConstant(0,
+ TLI.getVectorIdxTy()));
+ // Extract the input and convert the shorten input vector.
+ if (N->getNumOperands() == 1)
+ return DAG.getNode(Opcode, DL, WidenVT, InVal);
+ return DAG.getNode(Opcode, DL, WidenVT, InVal, N->getOperand(1));
+ }
+ }
+
+ // Otherwise unroll into some nasty scalar code and rebuild the vector.
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ EVT EltVT = WidenVT.getVectorElementType();
+ unsigned MinElts = std::min(InVTNumElts, WidenNumElts);
+ unsigned i;
+ for (i=0; i < MinElts; ++i) {
+ SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, InEltVT, InOp,
+ DAG.getConstant(i, TLI.getVectorIdxTy()));
+ if (N->getNumOperands() == 1)
+ Ops[i] = DAG.getNode(Opcode, DL, EltVT, Val);
+ else
+ Ops[i] = DAG.getNode(Opcode, DL, EltVT, Val, N->getOperand(1));
+ }
+
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for (; i < WidenNumElts; ++i)
+ Ops[i] = UndefVal;
+
+ return DAG.getNode(ISD::BUILD_VECTOR, DL, WidenVT, Ops);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_POWI(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ SDValue ShOp = N->getOperand(1);
+ return DAG.getNode(N->getOpcode(), SDLoc(N), WidenVT, InOp, ShOp);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_Shift(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ SDValue ShOp = N->getOperand(1);
+
+ EVT ShVT = ShOp.getValueType();
+ if (getTypeAction(ShVT) == TargetLowering::TypeWidenVector) {
+ ShOp = GetWidenedVector(ShOp);
+ ShVT = ShOp.getValueType();
+ }
+ EVT ShWidenVT = EVT::getVectorVT(*DAG.getContext(),
+ ShVT.getVectorElementType(),
+ WidenVT.getVectorNumElements());
+ if (ShVT != ShWidenVT)
+ ShOp = ModifyToType(ShOp, ShWidenVT);
+
+ return DAG.getNode(N->getOpcode(), SDLoc(N), WidenVT, InOp, ShOp);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_Unary(SDNode *N) {
+ // Unary op widening.
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ return DAG.getNode(N->getOpcode(), SDLoc(N), WidenVT, InOp);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_InregOp(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ EVT ExtVT = EVT::getVectorVT(*DAG.getContext(),
+ cast<VTSDNode>(N->getOperand(1))->getVT()
+ .getVectorElementType(),
+ WidenVT.getVectorNumElements());
+ SDValue WidenLHS = GetWidenedVector(N->getOperand(0));
+ return DAG.getNode(N->getOpcode(), SDLoc(N),
+ WidenVT, WidenLHS, DAG.getValueType(ExtVT));
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_MERGE_VALUES(SDNode *N, unsigned ResNo) {
+ SDValue WidenVec = DisintegrateMERGE_VALUES(N, ResNo);
+ return GetWidenedVector(WidenVec);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_BITCAST(SDNode *N) {
+ SDValue InOp = N->getOperand(0);
+ EVT InVT = InOp.getValueType();
+ EVT VT = N->getValueType(0);
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+ SDLoc dl(N);
+
+ switch (getTypeAction(InVT)) {
+ case TargetLowering::TypeLegal:
+ break;
+ case TargetLowering::TypePromoteInteger:
+ // If the incoming type is a vector that is being promoted, then
+ // we know that the elements are arranged differently and that we
+ // must perform the conversion using a stack slot.
+ if (InVT.isVector())
+ break;
+
+ // If the InOp is promoted to the same size, convert it. Otherwise,
+ // fall out of the switch and widen the promoted input.
+ InOp = GetPromotedInteger(InOp);
+ InVT = InOp.getValueType();
+ if (WidenVT.bitsEq(InVT))
+ return DAG.getNode(ISD::BITCAST, dl, WidenVT, InOp);
+ break;
+ case TargetLowering::TypeSoftenFloat:
+ case TargetLowering::TypeExpandInteger:
+ case TargetLowering::TypeExpandFloat:
+ case TargetLowering::TypeScalarizeVector:
+ case TargetLowering::TypeSplitVector:
+ break;
+ case TargetLowering::TypeWidenVector:
+ // If the InOp is widened to the same size, convert it. Otherwise, fall
+ // out of the switch and widen the widened input.
+ InOp = GetWidenedVector(InOp);
+ InVT = InOp.getValueType();
+ if (WidenVT.bitsEq(InVT))
+ // The input widens to the same size. Convert to the widen value.
+ return DAG.getNode(ISD::BITCAST, dl, WidenVT, InOp);
+ break;
+ }
+
+ unsigned WidenSize = WidenVT.getSizeInBits();
+ unsigned InSize = InVT.getSizeInBits();
+ // x86mmx is not an acceptable vector element type, so don't try.
+ if (WidenSize % InSize == 0 && InVT != MVT::x86mmx) {
+ // Determine new input vector type. The new input vector type will use
+ // the same element type (if its a vector) or use the input type as a
+ // vector. It is the same size as the type to widen to.
+ EVT NewInVT;
+ unsigned NewNumElts = WidenSize / InSize;
+ if (InVT.isVector()) {
+ EVT InEltVT = InVT.getVectorElementType();
+ NewInVT = EVT::getVectorVT(*DAG.getContext(), InEltVT,
+ WidenSize / InEltVT.getSizeInBits());
+ } else {
+ NewInVT = EVT::getVectorVT(*DAG.getContext(), InVT, NewNumElts);
+ }
+
+ if (TLI.isTypeLegal(NewInVT)) {
+ // Because the result and the input are different vector types, widening
+ // the result could create a legal type but widening the input might make
+ // it an illegal type that might lead to repeatedly splitting the input
+ // and then widening it. To avoid this, we widen the input only if
+ // it results in a legal type.
+ SmallVector<SDValue, 16> Ops(NewNumElts);
+ SDValue UndefVal = DAG.getUNDEF(InVT);
+ Ops[0] = InOp;
+ for (unsigned i = 1; i < NewNumElts; ++i)
+ Ops[i] = UndefVal;
+
+ SDValue NewVec;
+ if (InVT.isVector())
+ NewVec = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewInVT, Ops);
+ else
+ NewVec = DAG.getNode(ISD::BUILD_VECTOR, dl, NewInVT, Ops);
+ return DAG.getNode(ISD::BITCAST, dl, WidenVT, NewVec);
+ }
+ }
+
+ return CreateStackStoreLoad(InOp, WidenVT);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_BUILD_VECTOR(SDNode *N) {
+ SDLoc dl(N);
+ // Build a vector with undefined for the new nodes.
+ EVT VT = N->getValueType(0);
+
+ // Integer BUILD_VECTOR operands may be larger than the node's vector element
+ // type. The UNDEFs need to have the same type as the existing operands.
+ EVT EltVT = N->getOperand(0).getValueType();
+ unsigned NumElts = VT.getVectorNumElements();
+
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+
+ SmallVector<SDValue, 16> NewOps(N->op_begin(), N->op_end());
+ assert(WidenNumElts >= NumElts && "Shrinking vector instead of widening!");
+ NewOps.append(WidenNumElts - NumElts, DAG.getUNDEF(EltVT));
+
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, NewOps);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_CONCAT_VECTORS(SDNode *N) {
+ EVT InVT = N->getOperand(0).getValueType();
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDLoc dl(N);
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+ unsigned NumInElts = InVT.getVectorNumElements();
+ unsigned NumOperands = N->getNumOperands();
+
+ bool InputWidened = false; // Indicates we need to widen the input.
+ if (getTypeAction(InVT) != TargetLowering::TypeWidenVector) {
+ if (WidenVT.getVectorNumElements() % InVT.getVectorNumElements() == 0) {
+ // Add undef vectors to widen to correct length.
+ unsigned NumConcat = WidenVT.getVectorNumElements() /
+ InVT.getVectorNumElements();
+ SDValue UndefVal = DAG.getUNDEF(InVT);
+ SmallVector<SDValue, 16> Ops(NumConcat);
+ for (unsigned i=0; i < NumOperands; ++i)
+ Ops[i] = N->getOperand(i);
+ for (unsigned i = NumOperands; i != NumConcat; ++i)
+ Ops[i] = UndefVal;
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, Ops);
+ }
+ } else {
+ InputWidened = true;
+ if (WidenVT == TLI.getTypeToTransformTo(*DAG.getContext(), InVT)) {
+ // The inputs and the result are widen to the same value.
+ unsigned i;
+ for (i=1; i < NumOperands; ++i)
+ if (N->getOperand(i).getOpcode() != ISD::UNDEF)
+ break;
+
+ if (i == NumOperands)
+ // Everything but the first operand is an UNDEF so just return the
+ // widened first operand.
+ return GetWidenedVector(N->getOperand(0));
+
+ if (NumOperands == 2) {
+ // Replace concat of two operands with a shuffle.
+ SmallVector<int, 16> MaskOps(WidenNumElts, -1);
+ for (unsigned i = 0; i < NumInElts; ++i) {
+ MaskOps[i] = i;
+ MaskOps[i + NumInElts] = i + WidenNumElts;
+ }
+ return DAG.getVectorShuffle(WidenVT, dl,
+ GetWidenedVector(N->getOperand(0)),
+ GetWidenedVector(N->getOperand(1)),
+ &MaskOps[0]);
+ }
+ }
+ }
+
+ // Fall back to use extracts and build vector.
+ EVT EltVT = WidenVT.getVectorElementType();
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ unsigned Idx = 0;
+ for (unsigned i=0; i < NumOperands; ++i) {
+ SDValue InOp = N->getOperand(i);
+ if (InputWidened)
+ InOp = GetWidenedVector(InOp);
+ for (unsigned j=0; j < NumInElts; ++j)
+ Ops[Idx++] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,
+ DAG.getConstant(j, TLI.getVectorIdxTy()));
+ }
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for (; Idx < WidenNumElts; ++Idx)
+ Ops[Idx] = UndefVal;
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_CONVERT_RNDSAT(SDNode *N) {
+ SDLoc dl(N);
+ SDValue InOp = N->getOperand(0);
+ SDValue RndOp = N->getOperand(3);
+ SDValue SatOp = N->getOperand(4);
+
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+
+ EVT InVT = InOp.getValueType();
+ EVT InEltVT = InVT.getVectorElementType();
+ EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts);
+
+ SDValue DTyOp = DAG.getValueType(WidenVT);
+ SDValue STyOp = DAG.getValueType(InWidenVT);
+ ISD::CvtCode CvtCode = cast<CvtRndSatSDNode>(N)->getCvtCode();
+
+ unsigned InVTNumElts = InVT.getVectorNumElements();
+ if (getTypeAction(InVT) == TargetLowering::TypeWidenVector) {
+ InOp = GetWidenedVector(InOp);
+ InVT = InOp.getValueType();
+ InVTNumElts = InVT.getVectorNumElements();
+ if (InVTNumElts == WidenNumElts)
+ return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp,
+ SatOp, CvtCode);
+ }
+
+ if (TLI.isTypeLegal(InWidenVT)) {
+ // Because the result and the input are different vector types, widening
+ // the result could create a legal type but widening the input might make
+ // it an illegal type that might lead to repeatedly splitting the input
+ // and then widening it. To avoid this, we widen the input only if
+ // it results in a legal type.
+ if (WidenNumElts % InVTNumElts == 0) {
+ // Widen the input and call convert on the widened input vector.
+ unsigned NumConcat = WidenNumElts/InVTNumElts;
+ SmallVector<SDValue, 16> Ops(NumConcat);
+ Ops[0] = InOp;
+ SDValue UndefVal = DAG.getUNDEF(InVT);
+ for (unsigned i = 1; i != NumConcat; ++i)
+ Ops[i] = UndefVal;
+
+ InOp = DAG.getNode(ISD::CONCAT_VECTORS, dl, InWidenVT, Ops);
+ return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp,
+ SatOp, CvtCode);
+ }
+
+ if (InVTNumElts % WidenNumElts == 0) {
+ // Extract the input and convert the shorten input vector.
+ InOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InWidenVT, InOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp,
+ SatOp, CvtCode);
+ }
+ }
+
+ // Otherwise unroll into some nasty scalar code and rebuild the vector.
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ EVT EltVT = WidenVT.getVectorElementType();
+ DTyOp = DAG.getValueType(EltVT);
+ STyOp = DAG.getValueType(InEltVT);
+
+ unsigned MinElts = std::min(InVTNumElts, WidenNumElts);
+ unsigned i;
+ for (i=0; i < MinElts; ++i) {
+ SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp,
+ DAG.getConstant(i, TLI.getVectorIdxTy()));
+ Ops[i] = DAG.getConvertRndSat(WidenVT, dl, ExtVal, DTyOp, STyOp, RndOp,
+ SatOp, CvtCode);
+ }
+
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for (; i < WidenNumElts; ++i)
+ Ops[i] = UndefVal;
+
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_EXTRACT_SUBVECTOR(SDNode *N) {
+ EVT VT = N->getValueType(0);
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+ SDValue InOp = N->getOperand(0);
+ SDValue Idx = N->getOperand(1);
+ SDLoc dl(N);
+
+ if (getTypeAction(InOp.getValueType()) == TargetLowering::TypeWidenVector)
+ InOp = GetWidenedVector(InOp);
+
+ EVT InVT = InOp.getValueType();
+
+ // Check if we can just return the input vector after widening.
+ uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+ if (IdxVal == 0 && InVT == WidenVT)
+ return InOp;
+
+ // Check if we can extract from the vector.
+ unsigned InNumElts = InVT.getVectorNumElements();
+ if (IdxVal % WidenNumElts == 0 && IdxVal + WidenNumElts < InNumElts)
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, WidenVT, InOp, Idx);
+
+ // We could try widening the input to the right length but for now, extract
+ // the original elements, fill the rest with undefs and build a vector.
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ EVT EltVT = VT.getVectorElementType();
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned i;
+ for (i=0; i < NumElts; ++i)
+ Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,
+ DAG.getConstant(IdxVal+i, TLI.getVectorIdxTy()));
+
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for (; i < WidenNumElts; ++i)
+ Ops[i] = UndefVal;
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_INSERT_VECTOR_ELT(SDNode *N) {
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N),
+ InOp.getValueType(), InOp,
+ N->getOperand(1), N->getOperand(2));
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_LOAD(SDNode *N) {
+ LoadSDNode *LD = cast<LoadSDNode>(N);
+ ISD::LoadExtType ExtType = LD->getExtensionType();
+
+ SDValue Result;
+ SmallVector<SDValue, 16> LdChain; // Chain for the series of load
+ if (ExtType != ISD::NON_EXTLOAD)
+ Result = GenWidenVectorExtLoads(LdChain, LD, ExtType);
+ else
+ Result = GenWidenVectorLoads(LdChain, LD);
+
+ // If we generate a single load, we can use that for the chain. Otherwise,
+ // build a factor node to remember the multiple loads are independent and
+ // chain to that.
+ SDValue NewChain;
+ if (LdChain.size() == 1)
+ NewChain = LdChain[0];
+ else
+ NewChain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other, LdChain);
+
+ // Modified the chain - switch anything that used the old chain to use
+ // the new one.
+ ReplaceValueWith(SDValue(N, 1), NewChain);
+
+ return Result;
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_SCALAR_TO_VECTOR(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N),
+ WidenVT, N->getOperand(0));
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_SELECT(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+
+ SDValue Cond1 = N->getOperand(0);
+ EVT CondVT = Cond1.getValueType();
+ if (CondVT.isVector()) {
+ EVT CondEltVT = CondVT.getVectorElementType();
+ EVT CondWidenVT = EVT::getVectorVT(*DAG.getContext(),
+ CondEltVT, WidenNumElts);
+ if (getTypeAction(CondVT) == TargetLowering::TypeWidenVector)
+ Cond1 = GetWidenedVector(Cond1);
+
+ // If we have to split the condition there is no point in widening the
+ // select. This would result in an cycle of widening the select ->
+ // widening the condition operand -> splitting the condition operand ->
+ // splitting the select -> widening the select. Instead split this select
+ // further and widen the resulting type.
+ if (getTypeAction(CondVT) == TargetLowering::TypeSplitVector) {
+ SDValue SplitSelect = SplitVecOp_VSELECT(N, 0);
+ SDValue Res = ModifyToType(SplitSelect, WidenVT);
+ return Res;
+ }
+
+ if (Cond1.getValueType() != CondWidenVT)
+ Cond1 = ModifyToType(Cond1, CondWidenVT);
+ }
+
+ SDValue InOp1 = GetWidenedVector(N->getOperand(1));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(2));
+ assert(InOp1.getValueType() == WidenVT && InOp2.getValueType() == WidenVT);
+ return DAG.getNode(N->getOpcode(), SDLoc(N),
+ WidenVT, Cond1, InOp1, InOp2);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_SELECT_CC(SDNode *N) {
+ SDValue InOp1 = GetWidenedVector(N->getOperand(2));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(3));
+ return DAG.getNode(ISD::SELECT_CC, SDLoc(N),
+ InOp1.getValueType(), N->getOperand(0),
+ N->getOperand(1), InOp1, InOp2, N->getOperand(4));
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_SETCC(SDNode *N) {
+ assert(N->getValueType(0).isVector() ==
+ N->getOperand(0).getValueType().isVector() &&
+ "Scalar/Vector type mismatch");
+ if (N->getValueType(0).isVector()) return WidenVecRes_VSETCC(N);
+
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+ return DAG.getNode(ISD::SETCC, SDLoc(N), WidenVT,
+ InOp1, InOp2, N->getOperand(2));
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_UNDEF(SDNode *N) {
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ return DAG.getUNDEF(WidenVT);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N) {
+ EVT VT = N->getValueType(0);
+ SDLoc dl(N);
+
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+
+ SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+
+ // Adjust mask based on new input vector length.
+ SmallVector<int, 16> NewMask;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ int Idx = N->getMaskElt(i);
+ if (Idx < (int)NumElts)
+ NewMask.push_back(Idx);
+ else
+ NewMask.push_back(Idx - NumElts + WidenNumElts);
+ }
+ for (unsigned i = NumElts; i != WidenNumElts; ++i)
+ NewMask.push_back(-1);
+ return DAG.getVectorShuffle(WidenVT, dl, InOp1, InOp2, &NewMask[0]);
+}
+
+SDValue DAGTypeLegalizer::WidenVecRes_VSETCC(SDNode *N) {
+ assert(N->getValueType(0).isVector() &&
+ N->getOperand(0).getValueType().isVector() &&
+ "Operands must be vectors");
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+
+ SDValue InOp1 = N->getOperand(0);
+ EVT InVT = InOp1.getValueType();
+ assert(InVT.isVector() && "can not widen non-vector type");
+ EVT WidenInVT = EVT::getVectorVT(*DAG.getContext(),
+ InVT.getVectorElementType(), WidenNumElts);
+ InOp1 = GetWidenedVector(InOp1);
+ SDValue InOp2 = GetWidenedVector(N->getOperand(1));
+
+ // Assume that the input and output will be widen appropriately. If not,
+ // we will have to unroll it at some point.
+ assert(InOp1.getValueType() == WidenInVT &&
+ InOp2.getValueType() == WidenInVT &&
+ "Input not widened to expected type!");
+ (void)WidenInVT;
+ return DAG.getNode(ISD::SETCC, SDLoc(N),
+ WidenVT, InOp1, InOp2, N->getOperand(2));
+}
+
+
+//===----------------------------------------------------------------------===//
+// Widen Vector Operand
+//===----------------------------------------------------------------------===//
+bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned OpNo) {
+ DEBUG(dbgs() << "Widen node operand " << OpNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n");
+ SDValue Res = SDValue();
+
+ // See if the target wants to custom widen this node.
+ if (CustomLowerNode(N, N->getOperand(OpNo).getValueType(), false))
+ return false;
+
+ switch (N->getOpcode()) {
+ default:
+#ifndef NDEBUG
+ dbgs() << "WidenVectorOperand op #" << OpNo << ": ";
+ N->dump(&DAG);
+ dbgs() << "\n";
+#endif
+ llvm_unreachable("Do not know how to widen this operator's operand!");
+
+ case ISD::BITCAST: Res = WidenVecOp_BITCAST(N); break;
+ case ISD::CONCAT_VECTORS: Res = WidenVecOp_CONCAT_VECTORS(N); break;
+ case ISD::EXTRACT_SUBVECTOR: Res = WidenVecOp_EXTRACT_SUBVECTOR(N); break;
+ case ISD::EXTRACT_VECTOR_ELT: Res = WidenVecOp_EXTRACT_VECTOR_ELT(N); break;
+ case ISD::STORE: Res = WidenVecOp_STORE(N); break;
+ case ISD::SETCC: Res = WidenVecOp_SETCC(N); break;
+
+ case ISD::ANY_EXTEND:
+ case ISD::SIGN_EXTEND:
+ case ISD::ZERO_EXTEND:
+ Res = WidenVecOp_EXTEND(N);
+ break;
+
+ case ISD::FP_EXTEND:
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT:
+ case ISD::SINT_TO_FP:
+ case ISD::UINT_TO_FP:
+ case ISD::TRUNCATE:
+ Res = WidenVecOp_Convert(N);
+ break;
+ }
+
+ // If Res is null, the sub-method took care of registering the result.
+ if (!Res.getNode()) return false;
+
+ // If the result is N, the sub-method updated N in place. Tell the legalizer
+ // core about this.
+ if (Res.getNode() == N)
+ return true;
+
+
+ assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&
+ "Invalid operand expansion");
+
+ ReplaceValueWith(SDValue(N, 0), Res);
+ return false;
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_EXTEND(SDNode *N) {
+ SDLoc DL(N);
+ EVT VT = N->getValueType(0);
+
+ SDValue InOp = N->getOperand(0);
+ // If some legalization strategy other than widening is used on the operand,
+ // we can't safely assume that just extending the low lanes is the correct
+ // transformation.
+ if (getTypeAction(InOp.getValueType()) != TargetLowering::TypeWidenVector)
+ return WidenVecOp_Convert(N);
+ InOp = GetWidenedVector(InOp);
+ assert(VT.getVectorNumElements() <
+ InOp.getValueType().getVectorNumElements() &&
+ "Input wasn't widened!");
+
+ // We may need to further widen the operand until it has the same total
+ // vector size as the result.
+ EVT InVT = InOp.getValueType();
+ if (InVT.getSizeInBits() != VT.getSizeInBits()) {
+ EVT InEltVT = InVT.getVectorElementType();
+ for (int i = MVT::FIRST_VECTOR_VALUETYPE, e = MVT::LAST_VECTOR_VALUETYPE; i < e; ++i) {
+ EVT FixedVT = (MVT::SimpleValueType)i;
+ EVT FixedEltVT = FixedVT.getVectorElementType();
+ if (TLI.isTypeLegal(FixedVT) &&
+ FixedVT.getSizeInBits() == VT.getSizeInBits() &&
+ FixedEltVT == InEltVT) {
+ assert(FixedVT.getVectorNumElements() >= VT.getVectorNumElements() &&
+ "Not enough elements in the fixed type for the operand!");
+ assert(FixedVT.getVectorNumElements() != InVT.getVectorNumElements() &&
+ "We can't have the same type as we started with!");
+ if (FixedVT.getVectorNumElements() > InVT.getVectorNumElements())
+ InOp = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, FixedVT,
+ DAG.getUNDEF(FixedVT), InOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ else
+ InOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, FixedVT, InOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ break;
+ }
+ }
+ InVT = InOp.getValueType();
+ if (InVT.getSizeInBits() != VT.getSizeInBits())
+ // We couldn't find a legal vector type that was a widening of the input
+ // and could be extended in-register to the result type, so we have to
+ // scalarize.
+ return WidenVecOp_Convert(N);
+ }
+
+ // Use special DAG nodes to represent the operation of extending the
+ // low lanes.
+ switch (N->getOpcode()) {
+ default:
+ llvm_unreachable("Extend legalization on on extend operation!");
+ case ISD::ANY_EXTEND:
+ return DAG.getAnyExtendVectorInReg(InOp, DL, VT);
+ case ISD::SIGN_EXTEND:
+ return DAG.getSignExtendVectorInReg(InOp, DL, VT);
+ case ISD::ZERO_EXTEND:
+ return DAG.getZeroExtendVectorInReg(InOp, DL, VT);
+ }
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_Convert(SDNode *N) {
+ // Since the result is legal and the input is illegal, it is unlikely
+ // that we can fix the input to a legal type so unroll the convert
+ // into some scalar code and create a nasty build vector.
+ EVT VT = N->getValueType(0);
+ EVT EltVT = VT.getVectorElementType();
+ SDLoc dl(N);
+ unsigned NumElts = VT.getVectorNumElements();
+ SDValue InOp = N->getOperand(0);
+ if (getTypeAction(InOp.getValueType()) == TargetLowering::TypeWidenVector)
+ InOp = GetWidenedVector(InOp);
+ EVT InVT = InOp.getValueType();
+ EVT InEltVT = InVT.getVectorElementType();
+
+ unsigned Opcode = N->getOpcode();
+ SmallVector<SDValue, 16> Ops(NumElts);
+ for (unsigned i=0; i < NumElts; ++i)
+ Ops[i] = DAG.getNode(Opcode, dl, EltVT,
+ DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp,
+ DAG.getConstant(i, TLI.getVectorIdxTy())));
+
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_BITCAST(SDNode *N) {
+ EVT VT = N->getValueType(0);
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ EVT InWidenVT = InOp.getValueType();
+ SDLoc dl(N);
+
+ // Check if we can convert between two legal vector types and extract.
+ unsigned InWidenSize = InWidenVT.getSizeInBits();
+ unsigned Size = VT.getSizeInBits();
+ // x86mmx is not an acceptable vector element type, so don't try.
+ if (InWidenSize % Size == 0 && !VT.isVector() && VT != MVT::x86mmx) {
+ unsigned NewNumElts = InWidenSize / Size;
+ EVT NewVT = EVT::getVectorVT(*DAG.getContext(), VT, NewNumElts);
+ if (TLI.isTypeLegal(NewVT)) {
+ SDValue BitOp = DAG.getNode(ISD::BITCAST, dl, NewVT, InOp);
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, BitOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ }
+ }
+
+ return CreateStackStoreLoad(InOp, VT);
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_CONCAT_VECTORS(SDNode *N) {
+ // If the input vector is not legal, it is likely that we will not find a
+ // legal vector of the same size. Replace the concatenate vector with a
+ // nasty build vector.
+ EVT VT = N->getValueType(0);
+ EVT EltVT = VT.getVectorElementType();
+ SDLoc dl(N);
+ unsigned NumElts = VT.getVectorNumElements();
+ SmallVector<SDValue, 16> Ops(NumElts);
+
+ EVT InVT = N->getOperand(0).getValueType();
+ unsigned NumInElts = InVT.getVectorNumElements();
+
+ unsigned Idx = 0;
+ unsigned NumOperands = N->getNumOperands();
+ for (unsigned i=0; i < NumOperands; ++i) {
+ SDValue InOp = N->getOperand(i);
+ if (getTypeAction(InOp.getValueType()) == TargetLowering::TypeWidenVector)
+ InOp = GetWidenedVector(InOp);
+ for (unsigned j=0; j < NumInElts; ++j)
+ Ops[Idx++] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,
+ DAG.getConstant(j, TLI.getVectorIdxTy()));
+ }
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_EXTRACT_SUBVECTOR(SDNode *N) {
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N),
+ N->getValueType(0), InOp, N->getOperand(1));
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_EXTRACT_VECTOR_ELT(SDNode *N) {
+ SDValue InOp = GetWidenedVector(N->getOperand(0));
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
+ N->getValueType(0), InOp, N->getOperand(1));
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_STORE(SDNode *N) {
+ // We have to widen the value but we want only to store the original
+ // vector type.
+ StoreSDNode *ST = cast<StoreSDNode>(N);
+
+ SmallVector<SDValue, 16> StChain;
+ if (ST->isTruncatingStore())
+ GenWidenVectorTruncStores(StChain, ST);
+ else
+ GenWidenVectorStores(StChain, ST);
+
+ if (StChain.size() == 1)
+ return StChain[0];
+ else
+ return DAG.getNode(ISD::TokenFactor, SDLoc(ST), MVT::Other, StChain);
+}
+
+SDValue DAGTypeLegalizer::WidenVecOp_SETCC(SDNode *N) {
+ SDValue InOp0 = GetWidenedVector(N->getOperand(0));
+ SDValue InOp1 = GetWidenedVector(N->getOperand(1));
+ SDLoc dl(N);
+
+ // WARNING: In this code we widen the compare instruction with garbage.
+ // This garbage may contain denormal floats which may be slow. Is this a real
+ // concern ? Should we zero the unused lanes if this is a float compare ?
+
+ // Get a new SETCC node to compare the newly widened operands.
+ // Only some of the compared elements are legal.
+ EVT SVT = TLI.getSetCCResultType(*DAG.getContext(), InOp0.getValueType());
+ SDValue WideSETCC = DAG.getNode(ISD::SETCC, SDLoc(N),
+ SVT, InOp0, InOp1, N->getOperand(2));
+
+ // Extract the needed results from the result vector.
+ EVT ResVT = EVT::getVectorVT(*DAG.getContext(),
+ SVT.getVectorElementType(),
+ N->getValueType(0).getVectorNumElements());
+ SDValue CC = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl,
+ ResVT, WideSETCC, DAG.getConstant(0,
+ TLI.getVectorIdxTy()));
+
+ return PromoteTargetBoolean(CC, N->getValueType(0));
+}
+
+
+//===----------------------------------------------------------------------===//
+// Vector Widening Utilities
+//===----------------------------------------------------------------------===//
+
+// Utility function to find the type to chop up a widen vector for load/store
+// TLI: Target lowering used to determine legal types.
+// Width: Width left need to load/store.
+// WidenVT: The widen vector type to load to/store from
+// Align: If 0, don't allow use of a wider type
+// WidenEx: If Align is not 0, the amount additional we can load/store from.
+
+static EVT FindMemType(SelectionDAG& DAG, const TargetLowering &TLI,
+ unsigned Width, EVT WidenVT,
+ unsigned Align = 0, unsigned WidenEx = 0) {
+ EVT WidenEltVT = WidenVT.getVectorElementType();
+ unsigned WidenWidth = WidenVT.getSizeInBits();
+ unsigned WidenEltWidth = WidenEltVT.getSizeInBits();
+ unsigned AlignInBits = Align*8;
+
+ // If we have one element to load/store, return it.
+ EVT RetVT = WidenEltVT;
+ if (Width == WidenEltWidth)
+ return RetVT;
+
+ // See if there is larger legal integer than the element type to load/store
+ unsigned VT;
+ for (VT = (unsigned)MVT::LAST_INTEGER_VALUETYPE;
+ VT >= (unsigned)MVT::FIRST_INTEGER_VALUETYPE; --VT) {
+ EVT MemVT((MVT::SimpleValueType) VT);
+ unsigned MemVTWidth = MemVT.getSizeInBits();
+ if (MemVT.getSizeInBits() <= WidenEltWidth)
+ break;
+ if (TLI.isTypeLegal(MemVT) && (WidenWidth % MemVTWidth) == 0 &&
+ isPowerOf2_32(WidenWidth / MemVTWidth) &&
+ (MemVTWidth <= Width ||
+ (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) {
+ RetVT = MemVT;
+ break;
+ }
+ }
+
+ // See if there is a larger vector type to load/store that has the same vector
+ // element type and is evenly divisible with the WidenVT.
+ for (VT = (unsigned)MVT::LAST_VECTOR_VALUETYPE;
+ VT >= (unsigned)MVT::FIRST_VECTOR_VALUETYPE; --VT) {
+ EVT MemVT = (MVT::SimpleValueType) VT;
+ unsigned MemVTWidth = MemVT.getSizeInBits();
+ if (TLI.isTypeLegal(MemVT) && WidenEltVT == MemVT.getVectorElementType() &&
+ (WidenWidth % MemVTWidth) == 0 &&
+ isPowerOf2_32(WidenWidth / MemVTWidth) &&
+ (MemVTWidth <= Width ||
+ (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) {
+ if (RetVT.getSizeInBits() < MemVTWidth || MemVT == WidenVT)
+ return MemVT;
+ }
+ }
+
+ return RetVT;
+}
+
+// Builds a vector type from scalar loads
+// VecTy: Resulting Vector type
+// LDOps: Load operators to build a vector type
+// [Start,End) the list of loads to use.
+static SDValue BuildVectorFromScalar(SelectionDAG& DAG, EVT VecTy,
+ SmallVectorImpl<SDValue> &LdOps,
+ unsigned Start, unsigned End) {
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ SDLoc dl(LdOps[Start]);
+ EVT LdTy = LdOps[Start].getValueType();
+ unsigned Width = VecTy.getSizeInBits();
+ unsigned NumElts = Width / LdTy.getSizeInBits();
+ EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), LdTy, NumElts);
+
+ unsigned Idx = 1;
+ SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT,LdOps[Start]);
+
+ for (unsigned i = Start + 1; i != End; ++i) {
+ EVT NewLdTy = LdOps[i].getValueType();
+ if (NewLdTy != LdTy) {
+ NumElts = Width / NewLdTy.getSizeInBits();
+ NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewLdTy, NumElts);
+ VecOp = DAG.getNode(ISD::BITCAST, dl, NewVecVT, VecOp);
+ // Readjust position and vector position based on new load type
+ Idx = Idx * LdTy.getSizeInBits() / NewLdTy.getSizeInBits();
+ LdTy = NewLdTy;
+ }
+ VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, VecOp, LdOps[i],
+ DAG.getConstant(Idx++, TLI.getVectorIdxTy()));
+ }
+ return DAG.getNode(ISD::BITCAST, dl, VecTy, VecOp);
+}
+
+SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVectorImpl<SDValue> &LdChain,
+ LoadSDNode *LD) {
+ // The strategy assumes that we can efficiently load powers of two widths.
+ // The routines chops the vector into the largest vector loads with the same
+ // element type or scalar loads and then recombines it to the widen vector
+ // type.
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0));
+ unsigned WidenWidth = WidenVT.getSizeInBits();
+ EVT LdVT = LD->getMemoryVT();
+ SDLoc dl(LD);
+ assert(LdVT.isVector() && WidenVT.isVector());
+ assert(LdVT.getVectorElementType() == WidenVT.getVectorElementType());
+
+ // Load information
+ SDValue Chain = LD->getChain();
+ SDValue BasePtr = LD->getBasePtr();
+ unsigned Align = LD->getAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+ bool isInvariant = LD->isInvariant();
+ AAMDNodes AAInfo = LD->getAAInfo();
+
+ int LdWidth = LdVT.getSizeInBits();
+ int WidthDiff = WidenWidth - LdWidth; // Difference
+ unsigned LdAlign = (isVolatile) ? 0 : Align; // Allow wider loads
+
+ // Find the vector type that can load from.
+ EVT NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff);
+ int NewVTWidth = NewVT.getSizeInBits();
+ SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, LD->getPointerInfo(),
+ isVolatile, isNonTemporal, isInvariant, Align,
+ AAInfo);
+ LdChain.push_back(LdOp.getValue(1));
+
+ // Check if we can load the element with one instruction
+ if (LdWidth <= NewVTWidth) {
+ if (!NewVT.isVector()) {
+ unsigned NumElts = WidenWidth / NewVTWidth;
+ EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts);
+ SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT, LdOp);
+ return DAG.getNode(ISD::BITCAST, dl, WidenVT, VecOp);
+ }
+ if (NewVT == WidenVT)
+ return LdOp;
+
+ assert(WidenWidth % NewVTWidth == 0);
+ unsigned NumConcat = WidenWidth / NewVTWidth;
+ SmallVector<SDValue, 16> ConcatOps(NumConcat);
+ SDValue UndefVal = DAG.getUNDEF(NewVT);
+ ConcatOps[0] = LdOp;
+ for (unsigned i = 1; i != NumConcat; ++i)
+ ConcatOps[i] = UndefVal;
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, ConcatOps);
+ }
+
+ // Load vector by using multiple loads from largest vector to scalar
+ SmallVector<SDValue, 16> LdOps;
+ LdOps.push_back(LdOp);
+
+ LdWidth -= NewVTWidth;
+ unsigned Offset = 0;
+
+ while (LdWidth > 0) {
+ unsigned Increment = NewVTWidth / 8;
+ Offset += Increment;
+ BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
+ DAG.getConstant(Increment, BasePtr.getValueType()));
+
+ SDValue L;
+ if (LdWidth < NewVTWidth) {
+ // Our current type we are using is too large, find a better size
+ NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff);
+ NewVTWidth = NewVT.getSizeInBits();
+ L = DAG.getLoad(NewVT, dl, Chain, BasePtr,
+ LD->getPointerInfo().getWithOffset(Offset), isVolatile,
+ isNonTemporal, isInvariant, MinAlign(Align, Increment),
+ AAInfo);
+ LdChain.push_back(L.getValue(1));
+ if (L->getValueType(0).isVector()) {
+ SmallVector<SDValue, 16> Loads;
+ Loads.push_back(L);
+ unsigned size = L->getValueSizeInBits(0);
+ while (size < LdOp->getValueSizeInBits(0)) {
+ Loads.push_back(DAG.getUNDEF(L->getValueType(0)));
+ size += L->getValueSizeInBits(0);
+ }
+ L = DAG.getNode(ISD::CONCAT_VECTORS, dl, LdOp->getValueType(0), Loads);
+ }
+ } else {
+ L = DAG.getLoad(NewVT, dl, Chain, BasePtr,
+ LD->getPointerInfo().getWithOffset(Offset), isVolatile,
+ isNonTemporal, isInvariant, MinAlign(Align, Increment),
+ AAInfo);
+ LdChain.push_back(L.getValue(1));
+ }
+
+ LdOps.push_back(L);
+
+
+ LdWidth -= NewVTWidth;
+ }
+
+ // Build the vector from the loads operations
+ unsigned End = LdOps.size();
+ if (!LdOps[0].getValueType().isVector())
+ // All the loads are scalar loads.
+ return BuildVectorFromScalar(DAG, WidenVT, LdOps, 0, End);
+
+ // If the load contains vectors, build the vector using concat vector.
+ // All of the vectors used to loads are power of 2 and the scalars load
+ // can be combined to make a power of 2 vector.
+ SmallVector<SDValue, 16> ConcatOps(End);
+ int i = End - 1;
+ int Idx = End;
+ EVT LdTy = LdOps[i].getValueType();
+ // First combine the scalar loads to a vector
+ if (!LdTy.isVector()) {
+ for (--i; i >= 0; --i) {
+ LdTy = LdOps[i].getValueType();
+ if (LdTy.isVector())
+ break;
+ }
+ ConcatOps[--Idx] = BuildVectorFromScalar(DAG, LdTy, LdOps, i+1, End);
+ }
+ ConcatOps[--Idx] = LdOps[i];
+ for (--i; i >= 0; --i) {
+ EVT NewLdTy = LdOps[i].getValueType();
+ if (NewLdTy != LdTy) {
+ // Create a larger vector
+ ConcatOps[End-1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewLdTy,
+ makeArrayRef(&ConcatOps[Idx], End - Idx));
+ Idx = End - 1;
+ LdTy = NewLdTy;
+ }
+ ConcatOps[--Idx] = LdOps[i];
+ }
+
+ if (WidenWidth == LdTy.getSizeInBits()*(End - Idx))
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT,
+ makeArrayRef(&ConcatOps[Idx], End - Idx));
+
+ // We need to fill the rest with undefs to build the vector
+ unsigned NumOps = WidenWidth / LdTy.getSizeInBits();
+ SmallVector<SDValue, 16> WidenOps(NumOps);
+ SDValue UndefVal = DAG.getUNDEF(LdTy);
+ {
+ unsigned i = 0;
+ for (; i != End-Idx; ++i)
+ WidenOps[i] = ConcatOps[Idx+i];
+ for (; i != NumOps; ++i)
+ WidenOps[i] = UndefVal;
+ }
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, WidenOps);
+}
+
+SDValue
+DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVectorImpl<SDValue> &LdChain,
+ LoadSDNode *LD,
+ ISD::LoadExtType ExtType) {
+ // For extension loads, it may not be more efficient to chop up the vector
+ // and then extended it. Instead, we unroll the load and build a new vector.
+ EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0));
+ EVT LdVT = LD->getMemoryVT();
+ SDLoc dl(LD);
+ assert(LdVT.isVector() && WidenVT.isVector());
+
+ // Load information
+ SDValue Chain = LD->getChain();
+ SDValue BasePtr = LD->getBasePtr();
+ unsigned Align = LD->getAlignment();
+ bool isVolatile = LD->isVolatile();
+ bool isNonTemporal = LD->isNonTemporal();
+ bool isInvariant = LD->isInvariant();
+ AAMDNodes AAInfo = LD->getAAInfo();
+
+ EVT EltVT = WidenVT.getVectorElementType();
+ EVT LdEltVT = LdVT.getVectorElementType();
+ unsigned NumElts = LdVT.getVectorNumElements();
+
+ // Load each element and widen
+ unsigned WidenNumElts = WidenVT.getVectorNumElements();
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ unsigned Increment = LdEltVT.getSizeInBits() / 8;
+ Ops[0] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, BasePtr,
+ LD->getPointerInfo(),
+ LdEltVT, isVolatile, isNonTemporal, isInvariant,
+ Align, AAInfo);
+ LdChain.push_back(Ops[0].getValue(1));
+ unsigned i = 0, Offset = Increment;
+ for (i=1; i < NumElts; ++i, Offset += Increment) {
+ SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
+ BasePtr,
+ DAG.getConstant(Offset,
+ BasePtr.getValueType()));
+ Ops[i] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, NewBasePtr,
+ LD->getPointerInfo().getWithOffset(Offset), LdEltVT,
+ isVolatile, isNonTemporal, isInvariant, Align,
+ AAInfo);
+ LdChain.push_back(Ops[i].getValue(1));
+ }
+
+ // Fill the rest with undefs
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for (; i != WidenNumElts; ++i)
+ Ops[i] = UndefVal;
+
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, Ops);
+}
+
+
+void DAGTypeLegalizer::GenWidenVectorStores(SmallVectorImpl<SDValue> &StChain,
+ StoreSDNode *ST) {
+ // The strategy assumes that we can efficiently store powers of two widths.
+ // The routines chops the vector into the largest vector stores with the same
+ // element type or scalar stores.
+ SDValue Chain = ST->getChain();
+ SDValue BasePtr = ST->getBasePtr();
+ unsigned Align = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+ AAMDNodes AAInfo = ST->getAAInfo();
+ SDValue ValOp = GetWidenedVector(ST->getValue());
+ SDLoc dl(ST);
+
+ EVT StVT = ST->getMemoryVT();
+ unsigned StWidth = StVT.getSizeInBits();
+ EVT ValVT = ValOp.getValueType();
+ unsigned ValWidth = ValVT.getSizeInBits();
+ EVT ValEltVT = ValVT.getVectorElementType();
+ unsigned ValEltWidth = ValEltVT.getSizeInBits();
+ assert(StVT.getVectorElementType() == ValEltVT);
+
+ int Idx = 0; // current index to store
+ unsigned Offset = 0; // offset from base to store
+ while (StWidth != 0) {
+ // Find the largest vector type we can store with
+ EVT NewVT = FindMemType(DAG, TLI, StWidth, ValVT);
+ unsigned NewVTWidth = NewVT.getSizeInBits();
+ unsigned Increment = NewVTWidth / 8;
+ if (NewVT.isVector()) {
+ unsigned NumVTElts = NewVT.getVectorNumElements();
+ do {
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NewVT, ValOp,
+ DAG.getConstant(Idx, TLI.getVectorIdxTy()));
+ StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr,
+ ST->getPointerInfo().getWithOffset(Offset),
+ isVolatile, isNonTemporal,
+ MinAlign(Align, Offset), AAInfo));
+ StWidth -= NewVTWidth;
+ Offset += Increment;
+ Idx += NumVTElts;
+ BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
+ DAG.getConstant(Increment, BasePtr.getValueType()));
+ } while (StWidth != 0 && StWidth >= NewVTWidth);
+ } else {
+ // Cast the vector to the scalar type we can store
+ unsigned NumElts = ValWidth / NewVTWidth;
+ EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts);
+ SDValue VecOp = DAG.getNode(ISD::BITCAST, dl, NewVecVT, ValOp);
+ // Readjust index position based on new vector type
+ Idx = Idx * ValEltWidth / NewVTWidth;
+ do {
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewVT, VecOp,
+ DAG.getConstant(Idx++, TLI.getVectorIdxTy()));
+ StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr,
+ ST->getPointerInfo().getWithOffset(Offset),
+ isVolatile, isNonTemporal,
+ MinAlign(Align, Offset), AAInfo));
+ StWidth -= NewVTWidth;
+ Offset += Increment;
+ BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
+ DAG.getConstant(Increment, BasePtr.getValueType()));
+ } while (StWidth != 0 && StWidth >= NewVTWidth);
+ // Restore index back to be relative to the original widen element type
+ Idx = Idx * NewVTWidth / ValEltWidth;
+ }
+ }
+}
+
+void
+DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVectorImpl<SDValue> &StChain,
+ StoreSDNode *ST) {
+ // For extension loads, it may not be more efficient to truncate the vector
+ // and then store it. Instead, we extract each element and then store it.
+ SDValue Chain = ST->getChain();
+ SDValue BasePtr = ST->getBasePtr();
+ unsigned Align = ST->getAlignment();
+ bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
+ AAMDNodes AAInfo = ST->getAAInfo();
+ SDValue ValOp = GetWidenedVector(ST->getValue());
+ SDLoc dl(ST);
+
+ EVT StVT = ST->getMemoryVT();
+ EVT ValVT = ValOp.getValueType();
+
+ // It must be true that we the widen vector type is bigger than where
+ // we need to store.
+ assert(StVT.isVector() && ValOp.getValueType().isVector());
+ assert(StVT.bitsLT(ValOp.getValueType()));
+
+ // For truncating stores, we can not play the tricks of chopping legal
+ // vector types and bit cast it to the right type. Instead, we unroll
+ // the store.
+ EVT StEltVT = StVT.getVectorElementType();
+ EVT ValEltVT = ValVT.getVectorElementType();
+ unsigned Increment = ValEltVT.getSizeInBits() / 8;
+ unsigned NumElts = StVT.getVectorNumElements();
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, BasePtr,
+ ST->getPointerInfo(), StEltVT,
+ isVolatile, isNonTemporal, Align,
+ AAInfo));
+ unsigned Offset = Increment;
+ for (unsigned i=1; i < NumElts; ++i, Offset += Increment) {
+ SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
+ BasePtr, DAG.getConstant(Offset,
+ BasePtr.getValueType()));
+ SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+ StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, NewBasePtr,
+ ST->getPointerInfo().getWithOffset(Offset),
+ StEltVT, isVolatile, isNonTemporal,
+ MinAlign(Align, Offset), AAInfo));
+ }
+}
+
+/// Modifies a vector input (widen or narrows) to a vector of NVT. The
+/// input vector must have the same element type as NVT.
+SDValue DAGTypeLegalizer::ModifyToType(SDValue InOp, EVT NVT) {
+ // Note that InOp might have been widened so it might already have
+ // the right width or it might need be narrowed.
+ EVT InVT = InOp.getValueType();
+ assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&
+ "input and widen element type must match");
+ SDLoc dl(InOp);
+
+ // Check if InOp already has the right width.
+ if (InVT == NVT)
+ return InOp;
+
+ unsigned InNumElts = InVT.getVectorNumElements();
+ unsigned WidenNumElts = NVT.getVectorNumElements();
+ if (WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0) {
+ unsigned NumConcat = WidenNumElts / InNumElts;
+ SmallVector<SDValue, 16> Ops(NumConcat);
+ SDValue UndefVal = DAG.getUNDEF(InVT);
+ Ops[0] = InOp;
+ for (unsigned i = 1; i != NumConcat; ++i)
+ Ops[i] = UndefVal;
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, NVT, Ops);
+ }
+
+ if (WidenNumElts < InNumElts && InNumElts % WidenNumElts)
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT, InOp,
+ DAG.getConstant(0, TLI.getVectorIdxTy()));
+
+ // Fall back to extract and build.
+ SmallVector<SDValue, 16> Ops(WidenNumElts);
+ EVT EltVT = NVT.getVectorElementType();
+ unsigned MinNumElts = std::min(WidenNumElts, InNumElts);
+ unsigned Idx;
+ for (Idx = 0; Idx < MinNumElts; ++Idx)
+ Ops[Idx] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp,
+ DAG.getConstant(Idx, TLI.getVectorIdxTy()));
+
+ SDValue UndefVal = DAG.getUNDEF(EltVT);
+ for ( ; Idx < WidenNumElts; ++Idx)
+ Ops[Idx] = UndefVal;
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);
+}