LLVM API Documentation

NVPTXISelLowering.cpp
Go to the documentation of this file.
00001 //
00002 //                     The LLVM Compiler Infrastructure
00003 //
00004 // This file is distributed under the University of Illinois Open Source
00005 // License. See LICENSE.TXT for details.
00006 //
00007 //===----------------------------------------------------------------------===//
00008 //
00009 // This file defines the interfaces that NVPTX uses to lower LLVM code into a
00010 // selection DAG.
00011 //
00012 //===----------------------------------------------------------------------===//
00013 
00014 #include "NVPTXISelLowering.h"
00015 #include "NVPTX.h"
00016 #include "NVPTXTargetMachine.h"
00017 #include "NVPTXTargetObjectFile.h"
00018 #include "NVPTXUtilities.h"
00019 #include "llvm/CodeGen/Analysis.h"
00020 #include "llvm/CodeGen/MachineFrameInfo.h"
00021 #include "llvm/CodeGen/MachineFunction.h"
00022 #include "llvm/CodeGen/MachineInstrBuilder.h"
00023 #include "llvm/CodeGen/MachineRegisterInfo.h"
00024 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
00025 #include "llvm/IR/CallSite.h"
00026 #include "llvm/IR/DerivedTypes.h"
00027 #include "llvm/IR/Function.h"
00028 #include "llvm/IR/GlobalValue.h"
00029 #include "llvm/IR/IntrinsicInst.h"
00030 #include "llvm/IR/Intrinsics.h"
00031 #include "llvm/IR/Module.h"
00032 #include "llvm/MC/MCSectionELF.h"
00033 #include "llvm/Support/CommandLine.h"
00034 #include "llvm/Support/Debug.h"
00035 #include "llvm/Support/ErrorHandling.h"
00036 #include "llvm/Support/MathExtras.h"
00037 #include "llvm/Support/raw_ostream.h"
00038 #include <sstream>
00039 
00040 #undef DEBUG_TYPE
00041 #define DEBUG_TYPE "nvptx-lower"
00042 
00043 using namespace llvm;
00044 
00045 static unsigned int uniqueCallSite = 0;
00046 
00047 static cl::opt<bool> sched4reg(
00048     "nvptx-sched4reg",
00049     cl::desc("NVPTX Specific: schedule for register pressue"), cl::init(false));
00050 
00051 static cl::opt<unsigned>
00052 FMAContractLevelOpt("nvptx-fma-level", cl::ZeroOrMore, cl::Hidden,
00053                     cl::desc("NVPTX Specific: FMA contraction (0: don't do it"
00054                              " 1: do it  2: do it aggressively"),
00055                     cl::init(2));
00056 
00057 static bool IsPTXVectorType(MVT VT) {
00058   switch (VT.SimpleTy) {
00059   default:
00060     return false;
00061   case MVT::v2i1:
00062   case MVT::v4i1:
00063   case MVT::v2i8:
00064   case MVT::v4i8:
00065   case MVT::v2i16:
00066   case MVT::v4i16:
00067   case MVT::v2i32:
00068   case MVT::v4i32:
00069   case MVT::v2i64:
00070   case MVT::v2f32:
00071   case MVT::v4f32:
00072   case MVT::v2f64:
00073     return true;
00074   }
00075 }
00076 
00077 /// ComputePTXValueVTs - For the given Type \p Ty, returns the set of primitive
00078 /// EVTs that compose it.  Unlike ComputeValueVTs, this will break apart vectors
00079 /// into their primitive components.
00080 /// NOTE: This is a band-aid for code that expects ComputeValueVTs to return the
00081 /// same number of types as the Ins/Outs arrays in LowerFormalArguments,
00082 /// LowerCall, and LowerReturn.
00083 static void ComputePTXValueVTs(const TargetLowering &TLI, Type *Ty,
00084                                SmallVectorImpl<EVT> &ValueVTs,
00085                                SmallVectorImpl<uint64_t> *Offsets = nullptr,
00086                                uint64_t StartingOffset = 0) {
00087   SmallVector<EVT, 16> TempVTs;
00088   SmallVector<uint64_t, 16> TempOffsets;
00089 
00090   ComputeValueVTs(TLI, Ty, TempVTs, &TempOffsets, StartingOffset);
00091   for (unsigned i = 0, e = TempVTs.size(); i != e; ++i) {
00092     EVT VT = TempVTs[i];
00093     uint64_t Off = TempOffsets[i];
00094     if (VT.isVector())
00095       for (unsigned j = 0, je = VT.getVectorNumElements(); j != je; ++j) {
00096         ValueVTs.push_back(VT.getVectorElementType());
00097         if (Offsets)
00098           Offsets->push_back(Off+j*VT.getVectorElementType().getStoreSize());
00099       }
00100     else {
00101       ValueVTs.push_back(VT);
00102       if (Offsets)
00103         Offsets->push_back(Off);
00104     }
00105   }
00106 }
00107 
00108 // NVPTXTargetLowering Constructor.
00109 NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM)
00110     : TargetLowering(TM, new NVPTXTargetObjectFile()), nvTM(&TM),
00111       nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
00112 
00113   // always lower memset, memcpy, and memmove intrinsics to load/store
00114   // instructions, rather
00115   // then generating calls to memset, mempcy or memmove.
00116   MaxStoresPerMemset = (unsigned) 0xFFFFFFFF;
00117   MaxStoresPerMemcpy = (unsigned) 0xFFFFFFFF;
00118   MaxStoresPerMemmove = (unsigned) 0xFFFFFFFF;
00119 
00120   setBooleanContents(ZeroOrNegativeOneBooleanContent);
00121   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
00122 
00123   // Jump is Expensive. Don't create extra control flow for 'and', 'or'
00124   // condition branches.
00125   setJumpIsExpensive(true);
00126 
00127   // By default, use the Source scheduling
00128   if (sched4reg)
00129     setSchedulingPreference(Sched::RegPressure);
00130   else
00131     setSchedulingPreference(Sched::Source);
00132 
00133   addRegisterClass(MVT::i1, &NVPTX::Int1RegsRegClass);
00134   addRegisterClass(MVT::i16, &NVPTX::Int16RegsRegClass);
00135   addRegisterClass(MVT::i32, &NVPTX::Int32RegsRegClass);
00136   addRegisterClass(MVT::i64, &NVPTX::Int64RegsRegClass);
00137   addRegisterClass(MVT::f32, &NVPTX::Float32RegsRegClass);
00138   addRegisterClass(MVT::f64, &NVPTX::Float64RegsRegClass);
00139 
00140   // Operations not directly supported by NVPTX.
00141   setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
00142   setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
00143   setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
00144   setOperationAction(ISD::SELECT_CC, MVT::i8, Expand);
00145   setOperationAction(ISD::SELECT_CC, MVT::i16, Expand);
00146   setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
00147   setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
00148   setOperationAction(ISD::BR_CC, MVT::f32, Expand);
00149   setOperationAction(ISD::BR_CC, MVT::f64, Expand);
00150   setOperationAction(ISD::BR_CC, MVT::i1, Expand);
00151   setOperationAction(ISD::BR_CC, MVT::i8, Expand);
00152   setOperationAction(ISD::BR_CC, MVT::i16, Expand);
00153   setOperationAction(ISD::BR_CC, MVT::i32, Expand);
00154   setOperationAction(ISD::BR_CC, MVT::i64, Expand);
00155   // Some SIGN_EXTEND_INREG can be done using cvt instruction.
00156   // For others we will expand to a SHL/SRA pair.
00157   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i64, Legal);
00158   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
00159   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
00160   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
00161   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
00162 
00163   setOperationAction(ISD::SHL_PARTS, MVT::i32  , Custom);
00164   setOperationAction(ISD::SRA_PARTS, MVT::i32  , Custom);
00165   setOperationAction(ISD::SRL_PARTS, MVT::i32  , Custom);
00166   setOperationAction(ISD::SHL_PARTS, MVT::i64  , Custom);
00167   setOperationAction(ISD::SRA_PARTS, MVT::i64  , Custom);
00168   setOperationAction(ISD::SRL_PARTS, MVT::i64  , Custom);
00169 
00170   if (nvptxSubtarget.hasROT64()) {
00171     setOperationAction(ISD::ROTL, MVT::i64, Legal);
00172     setOperationAction(ISD::ROTR, MVT::i64, Legal);
00173   } else {
00174     setOperationAction(ISD::ROTL, MVT::i64, Expand);
00175     setOperationAction(ISD::ROTR, MVT::i64, Expand);
00176   }
00177   if (nvptxSubtarget.hasROT32()) {
00178     setOperationAction(ISD::ROTL, MVT::i32, Legal);
00179     setOperationAction(ISD::ROTR, MVT::i32, Legal);
00180   } else {
00181     setOperationAction(ISD::ROTL, MVT::i32, Expand);
00182     setOperationAction(ISD::ROTR, MVT::i32, Expand);
00183   }
00184 
00185   setOperationAction(ISD::ROTL, MVT::i16, Expand);
00186   setOperationAction(ISD::ROTR, MVT::i16, Expand);
00187   setOperationAction(ISD::ROTL, MVT::i8, Expand);
00188   setOperationAction(ISD::ROTR, MVT::i8, Expand);
00189   setOperationAction(ISD::BSWAP, MVT::i16, Expand);
00190   setOperationAction(ISD::BSWAP, MVT::i32, Expand);
00191   setOperationAction(ISD::BSWAP, MVT::i64, Expand);
00192 
00193   // Indirect branch is not supported.
00194   // This also disables Jump Table creation.
00195   setOperationAction(ISD::BR_JT, MVT::Other, Expand);
00196   setOperationAction(ISD::BRIND, MVT::Other, Expand);
00197 
00198   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
00199   setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
00200 
00201   // We want to legalize constant related memmove and memcopy
00202   // intrinsics.
00203   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
00204 
00205   // Turn FP extload into load/fextend
00206   setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
00207   setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
00208   // Turn FP truncstore into trunc + store.
00209   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
00210   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
00211   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
00212 
00213   // PTX does not support load / store predicate registers
00214   setOperationAction(ISD::LOAD, MVT::i1, Custom);
00215   setOperationAction(ISD::STORE, MVT::i1, Custom);
00216 
00217   setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
00218   setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
00219   setTruncStoreAction(MVT::i64, MVT::i1, Expand);
00220   setTruncStoreAction(MVT::i32, MVT::i1, Expand);
00221   setTruncStoreAction(MVT::i16, MVT::i1, Expand);
00222   setTruncStoreAction(MVT::i8, MVT::i1, Expand);
00223 
00224   // This is legal in NVPTX
00225   setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
00226   setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
00227 
00228   // TRAP can be lowered to PTX trap
00229   setOperationAction(ISD::TRAP, MVT::Other, Legal);
00230 
00231   setOperationAction(ISD::ADDC, MVT::i64, Expand);
00232   setOperationAction(ISD::ADDE, MVT::i64, Expand);
00233 
00234   // Register custom handling for vector loads/stores
00235   for (int i = MVT::FIRST_VECTOR_VALUETYPE; i <= MVT::LAST_VECTOR_VALUETYPE;
00236        ++i) {
00237     MVT VT = (MVT::SimpleValueType) i;
00238     if (IsPTXVectorType(VT)) {
00239       setOperationAction(ISD::LOAD, VT, Custom);
00240       setOperationAction(ISD::STORE, VT, Custom);
00241       setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom);
00242     }
00243   }
00244 
00245   // Custom handling for i8 intrinsics
00246   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
00247 
00248   setOperationAction(ISD::CTLZ, MVT::i16, Legal);
00249   setOperationAction(ISD::CTLZ, MVT::i32, Legal);
00250   setOperationAction(ISD::CTLZ, MVT::i64, Legal);
00251   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Legal);
00252   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Legal);
00253   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Legal);
00254   setOperationAction(ISD::CTTZ, MVT::i16, Expand);
00255   setOperationAction(ISD::CTTZ, MVT::i32, Expand);
00256   setOperationAction(ISD::CTTZ, MVT::i64, Expand);
00257   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Expand);
00258   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
00259   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
00260   setOperationAction(ISD::CTPOP, MVT::i16, Legal);
00261   setOperationAction(ISD::CTPOP, MVT::i32, Legal);
00262   setOperationAction(ISD::CTPOP, MVT::i64, Legal);
00263 
00264   // We have some custom DAG combine patterns for these nodes
00265   setTargetDAGCombine(ISD::ADD);
00266   setTargetDAGCombine(ISD::AND);
00267   setTargetDAGCombine(ISD::FADD);
00268   setTargetDAGCombine(ISD::MUL);
00269   setTargetDAGCombine(ISD::SHL);
00270 
00271   // Now deduce the information based on the above mentioned
00272   // actions
00273   computeRegisterProperties();
00274 }
00275 
00276 const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
00277   switch (Opcode) {
00278   default:
00279     return nullptr;
00280   case NVPTXISD::CALL:
00281     return "NVPTXISD::CALL";
00282   case NVPTXISD::RET_FLAG:
00283     return "NVPTXISD::RET_FLAG";
00284   case NVPTXISD::Wrapper:
00285     return "NVPTXISD::Wrapper";
00286   case NVPTXISD::DeclareParam:
00287     return "NVPTXISD::DeclareParam";
00288   case NVPTXISD::DeclareScalarParam:
00289     return "NVPTXISD::DeclareScalarParam";
00290   case NVPTXISD::DeclareRet:
00291     return "NVPTXISD::DeclareRet";
00292   case NVPTXISD::DeclareRetParam:
00293     return "NVPTXISD::DeclareRetParam";
00294   case NVPTXISD::PrintCall:
00295     return "NVPTXISD::PrintCall";
00296   case NVPTXISD::LoadParam:
00297     return "NVPTXISD::LoadParam";
00298   case NVPTXISD::LoadParamV2:
00299     return "NVPTXISD::LoadParamV2";
00300   case NVPTXISD::LoadParamV4:
00301     return "NVPTXISD::LoadParamV4";
00302   case NVPTXISD::StoreParam:
00303     return "NVPTXISD::StoreParam";
00304   case NVPTXISD::StoreParamV2:
00305     return "NVPTXISD::StoreParamV2";
00306   case NVPTXISD::StoreParamV4:
00307     return "NVPTXISD::StoreParamV4";
00308   case NVPTXISD::StoreParamS32:
00309     return "NVPTXISD::StoreParamS32";
00310   case NVPTXISD::StoreParamU32:
00311     return "NVPTXISD::StoreParamU32";
00312   case NVPTXISD::CallArgBegin:
00313     return "NVPTXISD::CallArgBegin";
00314   case NVPTXISD::CallArg:
00315     return "NVPTXISD::CallArg";
00316   case NVPTXISD::LastCallArg:
00317     return "NVPTXISD::LastCallArg";
00318   case NVPTXISD::CallArgEnd:
00319     return "NVPTXISD::CallArgEnd";
00320   case NVPTXISD::CallVoid:
00321     return "NVPTXISD::CallVoid";
00322   case NVPTXISD::CallVal:
00323     return "NVPTXISD::CallVal";
00324   case NVPTXISD::CallSymbol:
00325     return "NVPTXISD::CallSymbol";
00326   case NVPTXISD::Prototype:
00327     return "NVPTXISD::Prototype";
00328   case NVPTXISD::MoveParam:
00329     return "NVPTXISD::MoveParam";
00330   case NVPTXISD::StoreRetval:
00331     return "NVPTXISD::StoreRetval";
00332   case NVPTXISD::StoreRetvalV2:
00333     return "NVPTXISD::StoreRetvalV2";
00334   case NVPTXISD::StoreRetvalV4:
00335     return "NVPTXISD::StoreRetvalV4";
00336   case NVPTXISD::PseudoUseParam:
00337     return "NVPTXISD::PseudoUseParam";
00338   case NVPTXISD::RETURN:
00339     return "NVPTXISD::RETURN";
00340   case NVPTXISD::CallSeqBegin:
00341     return "NVPTXISD::CallSeqBegin";
00342   case NVPTXISD::CallSeqEnd:
00343     return "NVPTXISD::CallSeqEnd";
00344   case NVPTXISD::CallPrototype:
00345     return "NVPTXISD::CallPrototype";
00346   case NVPTXISD::LoadV2:
00347     return "NVPTXISD::LoadV2";
00348   case NVPTXISD::LoadV4:
00349     return "NVPTXISD::LoadV4";
00350   case NVPTXISD::LDGV2:
00351     return "NVPTXISD::LDGV2";
00352   case NVPTXISD::LDGV4:
00353     return "NVPTXISD::LDGV4";
00354   case NVPTXISD::LDUV2:
00355     return "NVPTXISD::LDUV2";
00356   case NVPTXISD::LDUV4:
00357     return "NVPTXISD::LDUV4";
00358   case NVPTXISD::StoreV2:
00359     return "NVPTXISD::StoreV2";
00360   case NVPTXISD::StoreV4:
00361     return "NVPTXISD::StoreV4";
00362   case NVPTXISD::FUN_SHFL_CLAMP:
00363     return "NVPTXISD::FUN_SHFL_CLAMP";
00364   case NVPTXISD::FUN_SHFR_CLAMP:
00365     return "NVPTXISD::FUN_SHFR_CLAMP";
00366   case NVPTXISD::IMAD:
00367     return "NVPTXISD::IMAD";
00368   case NVPTXISD::MUL_WIDE_SIGNED:
00369     return "NVPTXISD::MUL_WIDE_SIGNED";
00370   case NVPTXISD::MUL_WIDE_UNSIGNED:
00371     return "NVPTXISD::MUL_WIDE_UNSIGNED";
00372   case NVPTXISD::Tex1DFloatS32:        return "NVPTXISD::Tex1DFloatS32";
00373   case NVPTXISD::Tex1DFloatFloat:      return "NVPTXISD::Tex1DFloatFloat";
00374   case NVPTXISD::Tex1DFloatFloatLevel:
00375     return "NVPTXISD::Tex1DFloatFloatLevel";
00376   case NVPTXISD::Tex1DFloatFloatGrad:
00377     return "NVPTXISD::Tex1DFloatFloatGrad";
00378   case NVPTXISD::Tex1DS32S32:          return "NVPTXISD::Tex1DS32S32";
00379   case NVPTXISD::Tex1DS32Float:        return "NVPTXISD::Tex1DS32Float";
00380   case NVPTXISD::Tex1DS32FloatLevel:
00381     return "NVPTXISD::Tex1DS32FloatLevel";
00382   case NVPTXISD::Tex1DS32FloatGrad:
00383     return "NVPTXISD::Tex1DS32FloatGrad";
00384   case NVPTXISD::Tex1DU32S32:          return "NVPTXISD::Tex1DU32S32";
00385   case NVPTXISD::Tex1DU32Float:        return "NVPTXISD::Tex1DU32Float";
00386   case NVPTXISD::Tex1DU32FloatLevel:
00387     return "NVPTXISD::Tex1DU32FloatLevel";
00388   case NVPTXISD::Tex1DU32FloatGrad:
00389     return "NVPTXISD::Tex1DU32FloatGrad";
00390   case NVPTXISD::Tex1DArrayFloatS32:   return "NVPTXISD::Tex1DArrayFloatS32";
00391   case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex1DArrayFloatFloat";
00392   case NVPTXISD::Tex1DArrayFloatFloatLevel:
00393     return "NVPTXISD::Tex1DArrayFloatFloatLevel";
00394   case NVPTXISD::Tex1DArrayFloatFloatGrad:
00395     return "NVPTXISD::Tex1DArrayFloatFloatGrad";
00396   case NVPTXISD::Tex1DArrayS32S32:     return "NVPTXISD::Tex1DArrayS32S32";
00397   case NVPTXISD::Tex1DArrayS32Float:   return "NVPTXISD::Tex1DArrayS32Float";
00398   case NVPTXISD::Tex1DArrayS32FloatLevel:
00399     return "NVPTXISD::Tex1DArrayS32FloatLevel";
00400   case NVPTXISD::Tex1DArrayS32FloatGrad:
00401     return "NVPTXISD::Tex1DArrayS32FloatGrad";
00402   case NVPTXISD::Tex1DArrayU32S32:     return "NVPTXISD::Tex1DArrayU32S32";
00403   case NVPTXISD::Tex1DArrayU32Float:   return "NVPTXISD::Tex1DArrayU32Float";
00404   case NVPTXISD::Tex1DArrayU32FloatLevel:
00405     return "NVPTXISD::Tex1DArrayU32FloatLevel";
00406   case NVPTXISD::Tex1DArrayU32FloatGrad:
00407     return "NVPTXISD::Tex1DArrayU32FloatGrad";
00408   case NVPTXISD::Tex2DFloatS32:        return "NVPTXISD::Tex2DFloatS32";
00409   case NVPTXISD::Tex2DFloatFloat:      return "NVPTXISD::Tex2DFloatFloat";
00410   case NVPTXISD::Tex2DFloatFloatLevel:
00411     return "NVPTXISD::Tex2DFloatFloatLevel";
00412   case NVPTXISD::Tex2DFloatFloatGrad:
00413     return "NVPTXISD::Tex2DFloatFloatGrad";
00414   case NVPTXISD::Tex2DS32S32:          return "NVPTXISD::Tex2DS32S32";
00415   case NVPTXISD::Tex2DS32Float:        return "NVPTXISD::Tex2DS32Float";
00416   case NVPTXISD::Tex2DS32FloatLevel:
00417     return "NVPTXISD::Tex2DS32FloatLevel";
00418   case NVPTXISD::Tex2DS32FloatGrad:
00419     return "NVPTXISD::Tex2DS32FloatGrad";
00420   case NVPTXISD::Tex2DU32S32:          return "NVPTXISD::Tex2DU32S32";
00421   case NVPTXISD::Tex2DU32Float:        return "NVPTXISD::Tex2DU32Float";
00422   case NVPTXISD::Tex2DU32FloatLevel:
00423     return "NVPTXISD::Tex2DU32FloatLevel";
00424   case NVPTXISD::Tex2DU32FloatGrad:
00425     return "NVPTXISD::Tex2DU32FloatGrad";
00426   case NVPTXISD::Tex2DArrayFloatS32:   return "NVPTXISD::Tex2DArrayFloatS32";
00427   case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat";
00428   case NVPTXISD::Tex2DArrayFloatFloatLevel:
00429     return "NVPTXISD::Tex2DArrayFloatFloatLevel";
00430   case NVPTXISD::Tex2DArrayFloatFloatGrad:
00431     return "NVPTXISD::Tex2DArrayFloatFloatGrad";
00432   case NVPTXISD::Tex2DArrayS32S32:     return "NVPTXISD::Tex2DArrayS32S32";
00433   case NVPTXISD::Tex2DArrayS32Float:   return "NVPTXISD::Tex2DArrayS32Float";
00434   case NVPTXISD::Tex2DArrayS32FloatLevel:
00435     return "NVPTXISD::Tex2DArrayS32FloatLevel";
00436   case NVPTXISD::Tex2DArrayS32FloatGrad:
00437     return "NVPTXISD::Tex2DArrayS32FloatGrad";
00438   case NVPTXISD::Tex2DArrayU32S32:     return "NVPTXISD::Tex2DArrayU32S32";
00439   case NVPTXISD::Tex2DArrayU32Float:   return "NVPTXISD::Tex2DArrayU32Float";
00440   case NVPTXISD::Tex2DArrayU32FloatLevel:
00441     return "NVPTXISD::Tex2DArrayU32FloatLevel";
00442   case NVPTXISD::Tex2DArrayU32FloatGrad:
00443     return "NVPTXISD::Tex2DArrayU32FloatGrad";
00444   case NVPTXISD::Tex3DFloatS32:        return "NVPTXISD::Tex3DFloatS32";
00445   case NVPTXISD::Tex3DFloatFloat:      return "NVPTXISD::Tex3DFloatFloat";
00446   case NVPTXISD::Tex3DFloatFloatLevel:
00447     return "NVPTXISD::Tex3DFloatFloatLevel";
00448   case NVPTXISD::Tex3DFloatFloatGrad:
00449     return "NVPTXISD::Tex3DFloatFloatGrad";
00450   case NVPTXISD::Tex3DS32S32:          return "NVPTXISD::Tex3DS32S32";
00451   case NVPTXISD::Tex3DS32Float:        return "NVPTXISD::Tex3DS32Float";
00452   case NVPTXISD::Tex3DS32FloatLevel:
00453     return "NVPTXISD::Tex3DS32FloatLevel";
00454   case NVPTXISD::Tex3DS32FloatGrad:
00455     return "NVPTXISD::Tex3DS32FloatGrad";
00456   case NVPTXISD::Tex3DU32S32:          return "NVPTXISD::Tex3DU32S32";
00457   case NVPTXISD::Tex3DU32Float:        return "NVPTXISD::Tex3DU32Float";
00458   case NVPTXISD::Tex3DU32FloatLevel:
00459     return "NVPTXISD::Tex3DU32FloatLevel";
00460   case NVPTXISD::Tex3DU32FloatGrad:
00461     return "NVPTXISD::Tex3DU32FloatGrad";
00462   case NVPTXISD::TexCubeFloatFloat:      return "NVPTXISD::TexCubeFloatFloat";
00463   case NVPTXISD::TexCubeFloatFloatLevel:
00464     return "NVPTXISD::TexCubeFloatFloatLevel";
00465   case NVPTXISD::TexCubeS32Float:        return "NVPTXISD::TexCubeS32Float";
00466   case NVPTXISD::TexCubeS32FloatLevel:
00467     return "NVPTXISD::TexCubeS32FloatLevel";
00468   case NVPTXISD::TexCubeU32Float:        return "NVPTXISD::TexCubeU32Float";
00469   case NVPTXISD::TexCubeU32FloatLevel:
00470     return "NVPTXISD::TexCubeU32FloatLevel";
00471   case NVPTXISD::TexCubeArrayFloatFloat:
00472     return "NVPTXISD::TexCubeArrayFloatFloat";
00473   case NVPTXISD::TexCubeArrayFloatFloatLevel:
00474     return "NVPTXISD::TexCubeArrayFloatFloatLevel";
00475   case NVPTXISD::TexCubeArrayS32Float:
00476     return "NVPTXISD::TexCubeArrayS32Float";
00477   case NVPTXISD::TexCubeArrayS32FloatLevel:
00478     return "NVPTXISD::TexCubeArrayS32FloatLevel";
00479   case NVPTXISD::TexCubeArrayU32Float:
00480     return "NVPTXISD::TexCubeArrayU32Float";
00481   case NVPTXISD::TexCubeArrayU32FloatLevel:
00482     return "NVPTXISD::TexCubeArrayU32FloatLevel";
00483   case NVPTXISD::Tld4R2DFloatFloat:
00484     return "NVPTXISD::Tld4R2DFloatFloat";
00485   case NVPTXISD::Tld4G2DFloatFloat:
00486     return "NVPTXISD::Tld4G2DFloatFloat";
00487   case NVPTXISD::Tld4B2DFloatFloat:
00488     return "NVPTXISD::Tld4B2DFloatFloat";
00489   case NVPTXISD::Tld4A2DFloatFloat:
00490     return "NVPTXISD::Tld4A2DFloatFloat";
00491   case NVPTXISD::Tld4R2DS64Float:
00492     return "NVPTXISD::Tld4R2DS64Float";
00493   case NVPTXISD::Tld4G2DS64Float:
00494     return "NVPTXISD::Tld4G2DS64Float";
00495   case NVPTXISD::Tld4B2DS64Float:
00496     return "NVPTXISD::Tld4B2DS64Float";
00497   case NVPTXISD::Tld4A2DS64Float:
00498     return "NVPTXISD::Tld4A2DS64Float";
00499   case NVPTXISD::Tld4R2DU64Float:
00500     return "NVPTXISD::Tld4R2DU64Float";
00501   case NVPTXISD::Tld4G2DU64Float:
00502     return "NVPTXISD::Tld4G2DU64Float";
00503   case NVPTXISD::Tld4B2DU64Float:
00504     return "NVPTXISD::Tld4B2DU64Float";
00505   case NVPTXISD::Tld4A2DU64Float:
00506     return "NVPTXISD::Tld4A2DU64Float";
00507 
00508   case NVPTXISD::TexUnified1DFloatS32:
00509     return "NVPTXISD::TexUnified1DFloatS32";
00510   case NVPTXISD::TexUnified1DFloatFloat:
00511     return "NVPTXISD::TexUnified1DFloatFloat";
00512   case NVPTXISD::TexUnified1DFloatFloatLevel:
00513     return "NVPTXISD::TexUnified1DFloatFloatLevel";
00514   case NVPTXISD::TexUnified1DFloatFloatGrad:
00515     return "NVPTXISD::TexUnified1DFloatFloatGrad";
00516   case NVPTXISD::TexUnified1DS32S32:
00517     return "NVPTXISD::TexUnified1DS32S32";
00518   case NVPTXISD::TexUnified1DS32Float:
00519     return "NVPTXISD::TexUnified1DS32Float";
00520   case NVPTXISD::TexUnified1DS32FloatLevel:
00521     return "NVPTXISD::TexUnified1DS32FloatLevel";
00522   case NVPTXISD::TexUnified1DS32FloatGrad:
00523     return "NVPTXISD::TexUnified1DS32FloatGrad";
00524   case NVPTXISD::TexUnified1DU32S32:
00525     return "NVPTXISD::TexUnified1DU32S32";
00526   case NVPTXISD::TexUnified1DU32Float:
00527     return "NVPTXISD::TexUnified1DU32Float";
00528   case NVPTXISD::TexUnified1DU32FloatLevel:
00529     return "NVPTXISD::TexUnified1DU32FloatLevel";
00530   case NVPTXISD::TexUnified1DU32FloatGrad:
00531     return "NVPTXISD::TexUnified1DU32FloatGrad";
00532   case NVPTXISD::TexUnified1DArrayFloatS32:
00533     return "NVPTXISD::TexUnified1DArrayFloatS32";
00534   case NVPTXISD::TexUnified1DArrayFloatFloat:
00535     return "NVPTXISD::TexUnified1DArrayFloatFloat";
00536   case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
00537     return "NVPTXISD::TexUnified1DArrayFloatFloatLevel";
00538   case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
00539     return "NVPTXISD::TexUnified1DArrayFloatFloatGrad";
00540   case NVPTXISD::TexUnified1DArrayS32S32:
00541     return "NVPTXISD::TexUnified1DArrayS32S32";
00542   case NVPTXISD::TexUnified1DArrayS32Float:
00543     return "NVPTXISD::TexUnified1DArrayS32Float";
00544   case NVPTXISD::TexUnified1DArrayS32FloatLevel:
00545     return "NVPTXISD::TexUnified1DArrayS32FloatLevel";
00546   case NVPTXISD::TexUnified1DArrayS32FloatGrad:
00547     return "NVPTXISD::TexUnified1DArrayS32FloatGrad";
00548   case NVPTXISD::TexUnified1DArrayU32S32:
00549     return "NVPTXISD::TexUnified1DArrayU32S32";
00550   case NVPTXISD::TexUnified1DArrayU32Float:
00551     return "NVPTXISD::TexUnified1DArrayU32Float";
00552   case NVPTXISD::TexUnified1DArrayU32FloatLevel:
00553     return "NVPTXISD::TexUnified1DArrayU32FloatLevel";
00554   case NVPTXISD::TexUnified1DArrayU32FloatGrad:
00555     return "NVPTXISD::TexUnified1DArrayU32FloatGrad";
00556   case NVPTXISD::TexUnified2DFloatS32:
00557     return "NVPTXISD::TexUnified2DFloatS32";
00558   case NVPTXISD::TexUnified2DFloatFloat:
00559     return "NVPTXISD::TexUnified2DFloatFloat";
00560   case NVPTXISD::TexUnified2DFloatFloatLevel:
00561     return "NVPTXISD::TexUnified2DFloatFloatLevel";
00562   case NVPTXISD::TexUnified2DFloatFloatGrad:
00563     return "NVPTXISD::TexUnified2DFloatFloatGrad";
00564   case NVPTXISD::TexUnified2DS32S32:
00565     return "NVPTXISD::TexUnified2DS32S32";
00566   case NVPTXISD::TexUnified2DS32Float:
00567     return "NVPTXISD::TexUnified2DS32Float";
00568   case NVPTXISD::TexUnified2DS32FloatLevel:
00569     return "NVPTXISD::TexUnified2DS32FloatLevel";
00570   case NVPTXISD::TexUnified2DS32FloatGrad:
00571     return "NVPTXISD::TexUnified2DS32FloatGrad";
00572   case NVPTXISD::TexUnified2DU32S32:
00573     return "NVPTXISD::TexUnified2DU32S32";
00574   case NVPTXISD::TexUnified2DU32Float:
00575     return "NVPTXISD::TexUnified2DU32Float";
00576   case NVPTXISD::TexUnified2DU32FloatLevel:
00577     return "NVPTXISD::TexUnified2DU32FloatLevel";
00578   case NVPTXISD::TexUnified2DU32FloatGrad:
00579     return "NVPTXISD::TexUnified2DU32FloatGrad";
00580   case NVPTXISD::TexUnified2DArrayFloatS32:
00581     return "NVPTXISD::TexUnified2DArrayFloatS32";
00582   case NVPTXISD::TexUnified2DArrayFloatFloat:
00583     return "NVPTXISD::TexUnified2DArrayFloatFloat";
00584   case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
00585     return "NVPTXISD::TexUnified2DArrayFloatFloatLevel";
00586   case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
00587     return "NVPTXISD::TexUnified2DArrayFloatFloatGrad";
00588   case NVPTXISD::TexUnified2DArrayS32S32:
00589     return "NVPTXISD::TexUnified2DArrayS32S32";
00590   case NVPTXISD::TexUnified2DArrayS32Float:
00591     return "NVPTXISD::TexUnified2DArrayS32Float";
00592   case NVPTXISD::TexUnified2DArrayS32FloatLevel:
00593     return "NVPTXISD::TexUnified2DArrayS32FloatLevel";
00594   case NVPTXISD::TexUnified2DArrayS32FloatGrad:
00595     return "NVPTXISD::TexUnified2DArrayS32FloatGrad";
00596   case NVPTXISD::TexUnified2DArrayU32S32:
00597     return "NVPTXISD::TexUnified2DArrayU32S32";
00598   case NVPTXISD::TexUnified2DArrayU32Float:
00599     return "NVPTXISD::TexUnified2DArrayU32Float";
00600   case NVPTXISD::TexUnified2DArrayU32FloatLevel:
00601     return "NVPTXISD::TexUnified2DArrayU32FloatLevel";
00602   case NVPTXISD::TexUnified2DArrayU32FloatGrad:
00603     return "NVPTXISD::TexUnified2DArrayU32FloatGrad";
00604   case NVPTXISD::TexUnified3DFloatS32:
00605     return "NVPTXISD::TexUnified3DFloatS32";
00606   case NVPTXISD::TexUnified3DFloatFloat:
00607     return "NVPTXISD::TexUnified3DFloatFloat";
00608   case NVPTXISD::TexUnified3DFloatFloatLevel:
00609     return "NVPTXISD::TexUnified3DFloatFloatLevel";
00610   case NVPTXISD::TexUnified3DFloatFloatGrad:
00611     return "NVPTXISD::TexUnified3DFloatFloatGrad";
00612   case NVPTXISD::TexUnified3DS32S32:
00613     return "NVPTXISD::TexUnified3DS32S32";
00614   case NVPTXISD::TexUnified3DS32Float:
00615     return "NVPTXISD::TexUnified3DS32Float";
00616   case NVPTXISD::TexUnified3DS32FloatLevel:
00617     return "NVPTXISD::TexUnified3DS32FloatLevel";
00618   case NVPTXISD::TexUnified3DS32FloatGrad:
00619     return "NVPTXISD::TexUnified3DS32FloatGrad";
00620   case NVPTXISD::TexUnified3DU32S32:
00621     return "NVPTXISD::TexUnified3DU32S32";
00622   case NVPTXISD::TexUnified3DU32Float:
00623     return "NVPTXISD::TexUnified3DU32Float";
00624   case NVPTXISD::TexUnified3DU32FloatLevel:
00625     return "NVPTXISD::TexUnified3DU32FloatLevel";
00626   case NVPTXISD::TexUnified3DU32FloatGrad:
00627     return "NVPTXISD::TexUnified3DU32FloatGrad";
00628   case NVPTXISD::TexUnifiedCubeFloatFloat:
00629     return "NVPTXISD::TexUnifiedCubeFloatFloat";
00630   case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
00631     return "NVPTXISD::TexUnifiedCubeFloatFloatLevel";
00632   case NVPTXISD::TexUnifiedCubeS32Float:
00633     return "NVPTXISD::TexUnifiedCubeS32Float";
00634   case NVPTXISD::TexUnifiedCubeS32FloatLevel:
00635     return "NVPTXISD::TexUnifiedCubeS32FloatLevel";
00636   case NVPTXISD::TexUnifiedCubeU32Float:
00637     return "NVPTXISD::TexUnifiedCubeU32Float";
00638   case NVPTXISD::TexUnifiedCubeU32FloatLevel:
00639     return "NVPTXISD::TexUnifiedCubeU32FloatLevel";
00640   case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
00641     return "NVPTXISD::TexUnifiedCubeArrayFloatFloat";
00642   case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
00643     return "NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel";
00644   case NVPTXISD::TexUnifiedCubeArrayS32Float:
00645     return "NVPTXISD::TexUnifiedCubeArrayS32Float";
00646   case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
00647     return "NVPTXISD::TexUnifiedCubeArrayS32FloatLevel";
00648   case NVPTXISD::TexUnifiedCubeArrayU32Float:
00649     return "NVPTXISD::TexUnifiedCubeArrayU32Float";
00650   case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
00651     return "NVPTXISD::TexUnifiedCubeArrayU32FloatLevel";
00652   case NVPTXISD::Tld4UnifiedR2DFloatFloat:
00653     return "NVPTXISD::Tld4UnifiedR2DFloatFloat";
00654   case NVPTXISD::Tld4UnifiedG2DFloatFloat:
00655     return "NVPTXISD::Tld4UnifiedG2DFloatFloat";
00656   case NVPTXISD::Tld4UnifiedB2DFloatFloat:
00657     return "NVPTXISD::Tld4UnifiedB2DFloatFloat";
00658   case NVPTXISD::Tld4UnifiedA2DFloatFloat:
00659     return "NVPTXISD::Tld4UnifiedA2DFloatFloat";
00660   case NVPTXISD::Tld4UnifiedR2DS64Float:
00661     return "NVPTXISD::Tld4UnifiedR2DS64Float";
00662   case NVPTXISD::Tld4UnifiedG2DS64Float:
00663     return "NVPTXISD::Tld4UnifiedG2DS64Float";
00664   case NVPTXISD::Tld4UnifiedB2DS64Float:
00665     return "NVPTXISD::Tld4UnifiedB2DS64Float";
00666   case NVPTXISD::Tld4UnifiedA2DS64Float:
00667     return "NVPTXISD::Tld4UnifiedA2DS64Float";
00668   case NVPTXISD::Tld4UnifiedR2DU64Float:
00669     return "NVPTXISD::Tld4UnifiedR2DU64Float";
00670   case NVPTXISD::Tld4UnifiedG2DU64Float:
00671     return "NVPTXISD::Tld4UnifiedG2DU64Float";
00672   case NVPTXISD::Tld4UnifiedB2DU64Float:
00673     return "NVPTXISD::Tld4UnifiedB2DU64Float";
00674   case NVPTXISD::Tld4UnifiedA2DU64Float:
00675     return "NVPTXISD::Tld4UnifiedA2DU64Float";
00676 
00677   case NVPTXISD::Suld1DI8Clamp:          return "NVPTXISD::Suld1DI8Clamp";
00678   case NVPTXISD::Suld1DI16Clamp:         return "NVPTXISD::Suld1DI16Clamp";
00679   case NVPTXISD::Suld1DI32Clamp:         return "NVPTXISD::Suld1DI32Clamp";
00680   case NVPTXISD::Suld1DI64Clamp:         return "NVPTXISD::Suld1DI64Clamp";
00681   case NVPTXISD::Suld1DV2I8Clamp:        return "NVPTXISD::Suld1DV2I8Clamp";
00682   case NVPTXISD::Suld1DV2I16Clamp:       return "NVPTXISD::Suld1DV2I16Clamp";
00683   case NVPTXISD::Suld1DV2I32Clamp:       return "NVPTXISD::Suld1DV2I32Clamp";
00684   case NVPTXISD::Suld1DV2I64Clamp:       return "NVPTXISD::Suld1DV2I64Clamp";
00685   case NVPTXISD::Suld1DV4I8Clamp:        return "NVPTXISD::Suld1DV4I8Clamp";
00686   case NVPTXISD::Suld1DV4I16Clamp:       return "NVPTXISD::Suld1DV4I16Clamp";
00687   case NVPTXISD::Suld1DV4I32Clamp:       return "NVPTXISD::Suld1DV4I32Clamp";
00688 
00689   case NVPTXISD::Suld1DArrayI8Clamp:   return "NVPTXISD::Suld1DArrayI8Clamp";
00690   case NVPTXISD::Suld1DArrayI16Clamp:  return "NVPTXISD::Suld1DArrayI16Clamp";
00691   case NVPTXISD::Suld1DArrayI32Clamp:  return "NVPTXISD::Suld1DArrayI32Clamp";
00692   case NVPTXISD::Suld1DArrayI64Clamp:  return "NVPTXISD::Suld1DArrayI64Clamp";
00693   case NVPTXISD::Suld1DArrayV2I8Clamp: return "NVPTXISD::Suld1DArrayV2I8Clamp";
00694   case NVPTXISD::Suld1DArrayV2I16Clamp:return "NVPTXISD::Suld1DArrayV2I16Clamp";
00695   case NVPTXISD::Suld1DArrayV2I32Clamp:return "NVPTXISD::Suld1DArrayV2I32Clamp";
00696   case NVPTXISD::Suld1DArrayV2I64Clamp:return "NVPTXISD::Suld1DArrayV2I64Clamp";
00697   case NVPTXISD::Suld1DArrayV4I8Clamp: return "NVPTXISD::Suld1DArrayV4I8Clamp";
00698   case NVPTXISD::Suld1DArrayV4I16Clamp:return "NVPTXISD::Suld1DArrayV4I16Clamp";
00699   case NVPTXISD::Suld1DArrayV4I32Clamp:return "NVPTXISD::Suld1DArrayV4I32Clamp";
00700 
00701   case NVPTXISD::Suld2DI8Clamp:          return "NVPTXISD::Suld2DI8Clamp";
00702   case NVPTXISD::Suld2DI16Clamp:         return "NVPTXISD::Suld2DI16Clamp";
00703   case NVPTXISD::Suld2DI32Clamp:         return "NVPTXISD::Suld2DI32Clamp";
00704   case NVPTXISD::Suld2DI64Clamp:         return "NVPTXISD::Suld2DI64Clamp";
00705   case NVPTXISD::Suld2DV2I8Clamp:        return "NVPTXISD::Suld2DV2I8Clamp";
00706   case NVPTXISD::Suld2DV2I16Clamp:       return "NVPTXISD::Suld2DV2I16Clamp";
00707   case NVPTXISD::Suld2DV2I32Clamp:       return "NVPTXISD::Suld2DV2I32Clamp";
00708   case NVPTXISD::Suld2DV2I64Clamp:       return "NVPTXISD::Suld2DV2I64Clamp";
00709   case NVPTXISD::Suld2DV4I8Clamp:        return "NVPTXISD::Suld2DV4I8Clamp";
00710   case NVPTXISD::Suld2DV4I16Clamp:       return "NVPTXISD::Suld2DV4I16Clamp";
00711   case NVPTXISD::Suld2DV4I32Clamp:       return "NVPTXISD::Suld2DV4I32Clamp";
00712 
00713   case NVPTXISD::Suld2DArrayI8Clamp:   return "NVPTXISD::Suld2DArrayI8Clamp";
00714   case NVPTXISD::Suld2DArrayI16Clamp:  return "NVPTXISD::Suld2DArrayI16Clamp";
00715   case NVPTXISD::Suld2DArrayI32Clamp:  return "NVPTXISD::Suld2DArrayI32Clamp";
00716   case NVPTXISD::Suld2DArrayI64Clamp:  return "NVPTXISD::Suld2DArrayI64Clamp";
00717   case NVPTXISD::Suld2DArrayV2I8Clamp: return "NVPTXISD::Suld2DArrayV2I8Clamp";
00718   case NVPTXISD::Suld2DArrayV2I16Clamp:return "NVPTXISD::Suld2DArrayV2I16Clamp";
00719   case NVPTXISD::Suld2DArrayV2I32Clamp:return "NVPTXISD::Suld2DArrayV2I32Clamp";
00720   case NVPTXISD::Suld2DArrayV2I64Clamp:return "NVPTXISD::Suld2DArrayV2I64Clamp";
00721   case NVPTXISD::Suld2DArrayV4I8Clamp: return "NVPTXISD::Suld2DArrayV4I8Clamp";
00722   case NVPTXISD::Suld2DArrayV4I16Clamp:return "NVPTXISD::Suld2DArrayV4I16Clamp";
00723   case NVPTXISD::Suld2DArrayV4I32Clamp:return "NVPTXISD::Suld2DArrayV4I32Clamp";
00724 
00725   case NVPTXISD::Suld3DI8Clamp:          return "NVPTXISD::Suld3DI8Clamp";
00726   case NVPTXISD::Suld3DI16Clamp:         return "NVPTXISD::Suld3DI16Clamp";
00727   case NVPTXISD::Suld3DI32Clamp:         return "NVPTXISD::Suld3DI32Clamp";
00728   case NVPTXISD::Suld3DI64Clamp:         return "NVPTXISD::Suld3DI64Clamp";
00729   case NVPTXISD::Suld3DV2I8Clamp:        return "NVPTXISD::Suld3DV2I8Clamp";
00730   case NVPTXISD::Suld3DV2I16Clamp:       return "NVPTXISD::Suld3DV2I16Clamp";
00731   case NVPTXISD::Suld3DV2I32Clamp:       return "NVPTXISD::Suld3DV2I32Clamp";
00732   case NVPTXISD::Suld3DV2I64Clamp:       return "NVPTXISD::Suld3DV2I64Clamp";
00733   case NVPTXISD::Suld3DV4I8Clamp:        return "NVPTXISD::Suld3DV4I8Clamp";
00734   case NVPTXISD::Suld3DV4I16Clamp:       return "NVPTXISD::Suld3DV4I16Clamp";
00735   case NVPTXISD::Suld3DV4I32Clamp:       return "NVPTXISD::Suld3DV4I32Clamp";
00736 
00737   case NVPTXISD::Suld1DI8Trap:          return "NVPTXISD::Suld1DI8Trap";
00738   case NVPTXISD::Suld1DI16Trap:         return "NVPTXISD::Suld1DI16Trap";
00739   case NVPTXISD::Suld1DI32Trap:         return "NVPTXISD::Suld1DI32Trap";
00740   case NVPTXISD::Suld1DI64Trap:         return "NVPTXISD::Suld1DI64Trap";
00741   case NVPTXISD::Suld1DV2I8Trap:        return "NVPTXISD::Suld1DV2I8Trap";
00742   case NVPTXISD::Suld1DV2I16Trap:       return "NVPTXISD::Suld1DV2I16Trap";
00743   case NVPTXISD::Suld1DV2I32Trap:       return "NVPTXISD::Suld1DV2I32Trap";
00744   case NVPTXISD::Suld1DV2I64Trap:       return "NVPTXISD::Suld1DV2I64Trap";
00745   case NVPTXISD::Suld1DV4I8Trap:        return "NVPTXISD::Suld1DV4I8Trap";
00746   case NVPTXISD::Suld1DV4I16Trap:       return "NVPTXISD::Suld1DV4I16Trap";
00747   case NVPTXISD::Suld1DV4I32Trap:       return "NVPTXISD::Suld1DV4I32Trap";
00748 
00749   case NVPTXISD::Suld1DArrayI8Trap:     return "NVPTXISD::Suld1DArrayI8Trap";
00750   case NVPTXISD::Suld1DArrayI16Trap:    return "NVPTXISD::Suld1DArrayI16Trap";
00751   case NVPTXISD::Suld1DArrayI32Trap:    return "NVPTXISD::Suld1DArrayI32Trap";
00752   case NVPTXISD::Suld1DArrayI64Trap:    return "NVPTXISD::Suld1DArrayI64Trap";
00753   case NVPTXISD::Suld1DArrayV2I8Trap:   return "NVPTXISD::Suld1DArrayV2I8Trap";
00754   case NVPTXISD::Suld1DArrayV2I16Trap:  return "NVPTXISD::Suld1DArrayV2I16Trap";
00755   case NVPTXISD::Suld1DArrayV2I32Trap:  return "NVPTXISD::Suld1DArrayV2I32Trap";
00756   case NVPTXISD::Suld1DArrayV2I64Trap:  return "NVPTXISD::Suld1DArrayV2I64Trap";
00757   case NVPTXISD::Suld1DArrayV4I8Trap:   return "NVPTXISD::Suld1DArrayV4I8Trap";
00758   case NVPTXISD::Suld1DArrayV4I16Trap:  return "NVPTXISD::Suld1DArrayV4I16Trap";
00759   case NVPTXISD::Suld1DArrayV4I32Trap:  return "NVPTXISD::Suld1DArrayV4I32Trap";
00760 
00761   case NVPTXISD::Suld2DI8Trap:          return "NVPTXISD::Suld2DI8Trap";
00762   case NVPTXISD::Suld2DI16Trap:         return "NVPTXISD::Suld2DI16Trap";
00763   case NVPTXISD::Suld2DI32Trap:         return "NVPTXISD::Suld2DI32Trap";
00764   case NVPTXISD::Suld2DI64Trap:         return "NVPTXISD::Suld2DI64Trap";
00765   case NVPTXISD::Suld2DV2I8Trap:        return "NVPTXISD::Suld2DV2I8Trap";
00766   case NVPTXISD::Suld2DV2I16Trap:       return "NVPTXISD::Suld2DV2I16Trap";
00767   case NVPTXISD::Suld2DV2I32Trap:       return "NVPTXISD::Suld2DV2I32Trap";
00768   case NVPTXISD::Suld2DV2I64Trap:       return "NVPTXISD::Suld2DV2I64Trap";
00769   case NVPTXISD::Suld2DV4I8Trap:        return "NVPTXISD::Suld2DV4I8Trap";
00770   case NVPTXISD::Suld2DV4I16Trap:       return "NVPTXISD::Suld2DV4I16Trap";
00771   case NVPTXISD::Suld2DV4I32Trap:       return "NVPTXISD::Suld2DV4I32Trap";
00772 
00773   case NVPTXISD::Suld2DArrayI8Trap:     return "NVPTXISD::Suld2DArrayI8Trap";
00774   case NVPTXISD::Suld2DArrayI16Trap:    return "NVPTXISD::Suld2DArrayI16Trap";
00775   case NVPTXISD::Suld2DArrayI32Trap:    return "NVPTXISD::Suld2DArrayI32Trap";
00776   case NVPTXISD::Suld2DArrayI64Trap:    return "NVPTXISD::Suld2DArrayI64Trap";
00777   case NVPTXISD::Suld2DArrayV2I8Trap:   return "NVPTXISD::Suld2DArrayV2I8Trap";
00778   case NVPTXISD::Suld2DArrayV2I16Trap:  return "NVPTXISD::Suld2DArrayV2I16Trap";
00779   case NVPTXISD::Suld2DArrayV2I32Trap:  return "NVPTXISD::Suld2DArrayV2I32Trap";
00780   case NVPTXISD::Suld2DArrayV2I64Trap:  return "NVPTXISD::Suld2DArrayV2I64Trap";
00781   case NVPTXISD::Suld2DArrayV4I8Trap:   return "NVPTXISD::Suld2DArrayV4I8Trap";
00782   case NVPTXISD::Suld2DArrayV4I16Trap:  return "NVPTXISD::Suld2DArrayV4I16Trap";
00783   case NVPTXISD::Suld2DArrayV4I32Trap:  return "NVPTXISD::Suld2DArrayV4I32Trap";
00784 
00785   case NVPTXISD::Suld3DI8Trap:          return "NVPTXISD::Suld3DI8Trap";
00786   case NVPTXISD::Suld3DI16Trap:         return "NVPTXISD::Suld3DI16Trap";
00787   case NVPTXISD::Suld3DI32Trap:         return "NVPTXISD::Suld3DI32Trap";
00788   case NVPTXISD::Suld3DI64Trap:         return "NVPTXISD::Suld3DI64Trap";
00789   case NVPTXISD::Suld3DV2I8Trap:        return "NVPTXISD::Suld3DV2I8Trap";
00790   case NVPTXISD::Suld3DV2I16Trap:       return "NVPTXISD::Suld3DV2I16Trap";
00791   case NVPTXISD::Suld3DV2I32Trap:       return "NVPTXISD::Suld3DV2I32Trap";
00792   case NVPTXISD::Suld3DV2I64Trap:       return "NVPTXISD::Suld3DV2I64Trap";
00793   case NVPTXISD::Suld3DV4I8Trap:        return "NVPTXISD::Suld3DV4I8Trap";
00794   case NVPTXISD::Suld3DV4I16Trap:       return "NVPTXISD::Suld3DV4I16Trap";
00795   case NVPTXISD::Suld3DV4I32Trap:       return "NVPTXISD::Suld3DV4I32Trap";
00796 
00797   case NVPTXISD::Suld1DI8Zero:          return "NVPTXISD::Suld1DI8Zero";
00798   case NVPTXISD::Suld1DI16Zero:         return "NVPTXISD::Suld1DI16Zero";
00799   case NVPTXISD::Suld1DI32Zero:         return "NVPTXISD::Suld1DI32Zero";
00800   case NVPTXISD::Suld1DI64Zero:         return "NVPTXISD::Suld1DI64Zero";
00801   case NVPTXISD::Suld1DV2I8Zero:        return "NVPTXISD::Suld1DV2I8Zero";
00802   case NVPTXISD::Suld1DV2I16Zero:       return "NVPTXISD::Suld1DV2I16Zero";
00803   case NVPTXISD::Suld1DV2I32Zero:       return "NVPTXISD::Suld1DV2I32Zero";
00804   case NVPTXISD::Suld1DV2I64Zero:       return "NVPTXISD::Suld1DV2I64Zero";
00805   case NVPTXISD::Suld1DV4I8Zero:        return "NVPTXISD::Suld1DV4I8Zero";
00806   case NVPTXISD::Suld1DV4I16Zero:       return "NVPTXISD::Suld1DV4I16Zero";
00807   case NVPTXISD::Suld1DV4I32Zero:       return "NVPTXISD::Suld1DV4I32Zero";
00808 
00809   case NVPTXISD::Suld1DArrayI8Zero:     return "NVPTXISD::Suld1DArrayI8Zero";
00810   case NVPTXISD::Suld1DArrayI16Zero:    return "NVPTXISD::Suld1DArrayI16Zero";
00811   case NVPTXISD::Suld1DArrayI32Zero:    return "NVPTXISD::Suld1DArrayI32Zero";
00812   case NVPTXISD::Suld1DArrayI64Zero:    return "NVPTXISD::Suld1DArrayI64Zero";
00813   case NVPTXISD::Suld1DArrayV2I8Zero:   return "NVPTXISD::Suld1DArrayV2I8Zero";
00814   case NVPTXISD::Suld1DArrayV2I16Zero:  return "NVPTXISD::Suld1DArrayV2I16Zero";
00815   case NVPTXISD::Suld1DArrayV2I32Zero:  return "NVPTXISD::Suld1DArrayV2I32Zero";
00816   case NVPTXISD::Suld1DArrayV2I64Zero:  return "NVPTXISD::Suld1DArrayV2I64Zero";
00817   case NVPTXISD::Suld1DArrayV4I8Zero:   return "NVPTXISD::Suld1DArrayV4I8Zero";
00818   case NVPTXISD::Suld1DArrayV4I16Zero:  return "NVPTXISD::Suld1DArrayV4I16Zero";
00819   case NVPTXISD::Suld1DArrayV4I32Zero:  return "NVPTXISD::Suld1DArrayV4I32Zero";
00820 
00821   case NVPTXISD::Suld2DI8Zero:          return "NVPTXISD::Suld2DI8Zero";
00822   case NVPTXISD::Suld2DI16Zero:         return "NVPTXISD::Suld2DI16Zero";
00823   case NVPTXISD::Suld2DI32Zero:         return "NVPTXISD::Suld2DI32Zero";
00824   case NVPTXISD::Suld2DI64Zero:         return "NVPTXISD::Suld2DI64Zero";
00825   case NVPTXISD::Suld2DV2I8Zero:        return "NVPTXISD::Suld2DV2I8Zero";
00826   case NVPTXISD::Suld2DV2I16Zero:       return "NVPTXISD::Suld2DV2I16Zero";
00827   case NVPTXISD::Suld2DV2I32Zero:       return "NVPTXISD::Suld2DV2I32Zero";
00828   case NVPTXISD::Suld2DV2I64Zero:       return "NVPTXISD::Suld2DV2I64Zero";
00829   case NVPTXISD::Suld2DV4I8Zero:        return "NVPTXISD::Suld2DV4I8Zero";
00830   case NVPTXISD::Suld2DV4I16Zero:       return "NVPTXISD::Suld2DV4I16Zero";
00831   case NVPTXISD::Suld2DV4I32Zero:       return "NVPTXISD::Suld2DV4I32Zero";
00832 
00833   case NVPTXISD::Suld2DArrayI8Zero:     return "NVPTXISD::Suld2DArrayI8Zero";
00834   case NVPTXISD::Suld2DArrayI16Zero:    return "NVPTXISD::Suld2DArrayI16Zero";
00835   case NVPTXISD::Suld2DArrayI32Zero:    return "NVPTXISD::Suld2DArrayI32Zero";
00836   case NVPTXISD::Suld2DArrayI64Zero:    return "NVPTXISD::Suld2DArrayI64Zero";
00837   case NVPTXISD::Suld2DArrayV2I8Zero:   return "NVPTXISD::Suld2DArrayV2I8Zero";
00838   case NVPTXISD::Suld2DArrayV2I16Zero:  return "NVPTXISD::Suld2DArrayV2I16Zero";
00839   case NVPTXISD::Suld2DArrayV2I32Zero:  return "NVPTXISD::Suld2DArrayV2I32Zero";
00840   case NVPTXISD::Suld2DArrayV2I64Zero:  return "NVPTXISD::Suld2DArrayV2I64Zero";
00841   case NVPTXISD::Suld2DArrayV4I8Zero:   return "NVPTXISD::Suld2DArrayV4I8Zero";
00842   case NVPTXISD::Suld2DArrayV4I16Zero:  return "NVPTXISD::Suld2DArrayV4I16Zero";
00843   case NVPTXISD::Suld2DArrayV4I32Zero:  return "NVPTXISD::Suld2DArrayV4I32Zero";
00844 
00845   case NVPTXISD::Suld3DI8Zero:          return "NVPTXISD::Suld3DI8Zero";
00846   case NVPTXISD::Suld3DI16Zero:         return "NVPTXISD::Suld3DI16Zero";
00847   case NVPTXISD::Suld3DI32Zero:         return "NVPTXISD::Suld3DI32Zero";
00848   case NVPTXISD::Suld3DI64Zero:         return "NVPTXISD::Suld3DI64Zero";
00849   case NVPTXISD::Suld3DV2I8Zero:        return "NVPTXISD::Suld3DV2I8Zero";
00850   case NVPTXISD::Suld3DV2I16Zero:       return "NVPTXISD::Suld3DV2I16Zero";
00851   case NVPTXISD::Suld3DV2I32Zero:       return "NVPTXISD::Suld3DV2I32Zero";
00852   case NVPTXISD::Suld3DV2I64Zero:       return "NVPTXISD::Suld3DV2I64Zero";
00853   case NVPTXISD::Suld3DV4I8Zero:        return "NVPTXISD::Suld3DV4I8Zero";
00854   case NVPTXISD::Suld3DV4I16Zero:       return "NVPTXISD::Suld3DV4I16Zero";
00855   case NVPTXISD::Suld3DV4I32Zero:       return "NVPTXISD::Suld3DV4I32Zero";
00856   }
00857 }
00858 
00859 TargetLoweringBase::LegalizeTypeAction
00860 NVPTXTargetLowering::getPreferredVectorAction(EVT VT) const {
00861   if (VT.getVectorNumElements() != 1 && VT.getScalarType() == MVT::i1)
00862     return TypeSplitVector;
00863 
00864   return TargetLoweringBase::getPreferredVectorAction(VT);
00865 }
00866 
00867 SDValue
00868 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
00869   SDLoc dl(Op);
00870   const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
00871   Op = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
00872   return DAG.getNode(NVPTXISD::Wrapper, dl, getPointerTy(), Op);
00873 }
00874 
00875 std::string
00876 NVPTXTargetLowering::getPrototype(Type *retTy, const ArgListTy &Args,
00877                                   const SmallVectorImpl<ISD::OutputArg> &Outs,
00878                                   unsigned retAlignment,
00879                                   const ImmutableCallSite *CS) const {
00880 
00881   bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
00882   assert(isABI && "Non-ABI compilation is not supported");
00883   if (!isABI)
00884     return "";
00885 
00886   std::stringstream O;
00887   O << "prototype_" << uniqueCallSite << " : .callprototype ";
00888 
00889   if (retTy->getTypeID() == Type::VoidTyID) {
00890     O << "()";
00891   } else {
00892     O << "(";
00893     if (retTy->isFloatingPointTy() || retTy->isIntegerTy()) {
00894       unsigned size = 0;
00895       if (const IntegerType *ITy = dyn_cast<IntegerType>(retTy)) {
00896         size = ITy->getBitWidth();
00897         if (size < 32)
00898           size = 32;
00899       } else {
00900         assert(retTy->isFloatingPointTy() &&
00901                "Floating point type expected here");
00902         size = retTy->getPrimitiveSizeInBits();
00903       }
00904 
00905       O << ".param .b" << size << " _";
00906     } else if (isa<PointerType>(retTy)) {
00907       O << ".param .b" << getPointerTy().getSizeInBits() << " _";
00908     } else {
00909       if((retTy->getTypeID() == Type::StructTyID) ||
00910          isa<VectorType>(retTy)) {
00911         O << ".param .align "
00912           << retAlignment
00913           << " .b8 _["
00914           << getDataLayout()->getTypeAllocSize(retTy) << "]";
00915       } else {
00916         assert(false && "Unknown return type");
00917       }
00918     }
00919     O << ") ";
00920   }
00921   O << "_ (";
00922 
00923   bool first = true;
00924   MVT thePointerTy = getPointerTy();
00925 
00926   unsigned OIdx = 0;
00927   for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
00928     Type *Ty = Args[i].Ty;
00929     if (!first) {
00930       O << ", ";
00931     }
00932     first = false;
00933 
00934     if (Outs[OIdx].Flags.isByVal() == false) {
00935       if (Ty->isAggregateType() || Ty->isVectorTy()) {
00936         unsigned align = 0;
00937         const CallInst *CallI = cast<CallInst>(CS->getInstruction());
00938         const DataLayout *TD = getDataLayout();
00939         // +1 because index 0 is reserved for return type alignment
00940         if (!llvm::getAlign(*CallI, i + 1, align))
00941           align = TD->getABITypeAlignment(Ty);
00942         unsigned sz = TD->getTypeAllocSize(Ty);
00943         O << ".param .align " << align << " .b8 ";
00944         O << "_";
00945         O << "[" << sz << "]";
00946         // update the index for Outs
00947         SmallVector<EVT, 16> vtparts;
00948         ComputeValueVTs(*this, Ty, vtparts);
00949         if (unsigned len = vtparts.size())
00950           OIdx += len - 1;
00951         continue;
00952       }
00953        // i8 types in IR will be i16 types in SDAG
00954       assert((getValueType(Ty) == Outs[OIdx].VT ||
00955              (getValueType(Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) &&
00956              "type mismatch between callee prototype and arguments");
00957       // scalar type
00958       unsigned sz = 0;
00959       if (isa<IntegerType>(Ty)) {
00960         sz = cast<IntegerType>(Ty)->getBitWidth();
00961         if (sz < 32)
00962           sz = 32;
00963       } else if (isa<PointerType>(Ty))
00964         sz = thePointerTy.getSizeInBits();
00965       else
00966         sz = Ty->getPrimitiveSizeInBits();
00967       O << ".param .b" << sz << " ";
00968       O << "_";
00969       continue;
00970     }
00971     const PointerType *PTy = dyn_cast<PointerType>(Ty);
00972     assert(PTy && "Param with byval attribute should be a pointer type");
00973     Type *ETy = PTy->getElementType();
00974 
00975     unsigned align = Outs[OIdx].Flags.getByValAlign();
00976     unsigned sz = getDataLayout()->getTypeAllocSize(ETy);
00977     O << ".param .align " << align << " .b8 ";
00978     O << "_";
00979     O << "[" << sz << "]";
00980   }
00981   O << ");";
00982   return O.str();
00983 }
00984 
00985 unsigned
00986 NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
00987                                           const ImmutableCallSite *CS,
00988                                           Type *Ty,
00989                                           unsigned Idx) const {
00990   const DataLayout *TD = getDataLayout();
00991   unsigned Align = 0;
00992   const Value *DirectCallee = CS->getCalledFunction();
00993 
00994   if (!DirectCallee) {
00995     // We don't have a direct function symbol, but that may be because of
00996     // constant cast instructions in the call.
00997     const Instruction *CalleeI = CS->getInstruction();
00998     assert(CalleeI && "Call target is not a function or derived value?");
00999 
01000     // With bitcast'd call targets, the instruction will be the call
01001     if (isa<CallInst>(CalleeI)) {
01002       // Check if we have call alignment metadata
01003       if (llvm::getAlign(*cast<CallInst>(CalleeI), Idx, Align))
01004         return Align;
01005 
01006       const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
01007       // Ignore any bitcast instructions
01008       while(isa<ConstantExpr>(CalleeV)) {
01009         const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
01010         if (!CE->isCast())
01011           break;
01012         // Look through the bitcast
01013         CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0);
01014       }
01015 
01016       // We have now looked past all of the bitcasts.  Do we finally have a
01017       // Function?
01018       if (isa<Function>(CalleeV))
01019         DirectCallee = CalleeV;
01020     }
01021   }
01022 
01023   // Check for function alignment information if we found that the
01024   // ultimate target is a Function
01025   if (DirectCallee)
01026     if (llvm::getAlign(*cast<Function>(DirectCallee), Idx, Align))
01027       return Align;
01028 
01029   // Call is indirect or alignment information is not available, fall back to
01030   // the ABI type alignment
01031   return TD->getABITypeAlignment(Ty);
01032 }
01033 
01034 SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
01035                                        SmallVectorImpl<SDValue> &InVals) const {
01036   SelectionDAG &DAG = CLI.DAG;
01037   SDLoc dl = CLI.DL;
01038   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
01039   SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
01040   SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
01041   SDValue Chain = CLI.Chain;
01042   SDValue Callee = CLI.Callee;
01043   bool &isTailCall = CLI.IsTailCall;
01044   ArgListTy &Args = CLI.getArgs();
01045   Type *retTy = CLI.RetTy;
01046   ImmutableCallSite *CS = CLI.CS;
01047 
01048   bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
01049   assert(isABI && "Non-ABI compilation is not supported");
01050   if (!isABI)
01051     return Chain;
01052   const DataLayout *TD = getDataLayout();
01053   MachineFunction &MF = DAG.getMachineFunction();
01054   const Function *F = MF.getFunction();
01055 
01056   SDValue tempChain = Chain;
01057   Chain =
01058       DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(uniqueCallSite, true),
01059                            dl);
01060   SDValue InFlag = Chain.getValue(1);
01061 
01062   unsigned paramCount = 0;
01063   // Args.size() and Outs.size() need not match.
01064   // Outs.size() will be larger
01065   //   * if there is an aggregate argument with multiple fields (each field
01066   //     showing up separately in Outs)
01067   //   * if there is a vector argument with more than typical vector-length
01068   //     elements (generally if more than 4) where each vector element is
01069   //     individually present in Outs.
01070   // So a different index should be used for indexing into Outs/OutVals.
01071   // See similar issue in LowerFormalArguments.
01072   unsigned OIdx = 0;
01073   // Declare the .params or .reg need to pass values
01074   // to the function
01075   for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
01076     EVT VT = Outs[OIdx].VT;
01077     Type *Ty = Args[i].Ty;
01078 
01079     if (Outs[OIdx].Flags.isByVal() == false) {
01080       if (Ty->isAggregateType()) {
01081         // aggregate
01082         SmallVector<EVT, 16> vtparts;
01083         SmallVector<uint64_t, 16> Offsets;
01084         ComputePTXValueVTs(*this, Ty, vtparts, &Offsets, 0);
01085 
01086         unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
01087         // declare .param .align <align> .b8 .param<n>[<size>];
01088         unsigned sz = TD->getTypeAllocSize(Ty);
01089         SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01090         SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32),
01091                                       DAG.getConstant(paramCount, MVT::i32),
01092                                       DAG.getConstant(sz, MVT::i32), InFlag };
01093         Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
01094                             DeclareParamOps);
01095         InFlag = Chain.getValue(1);
01096         for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
01097           EVT elemtype = vtparts[j];
01098           unsigned ArgAlign = GreatestCommonDivisor64(align, Offsets[j]);
01099           if (elemtype.isInteger() && (sz < 8))
01100             sz = 8;
01101           SDValue StVal = OutVals[OIdx];
01102           if (elemtype.getSizeInBits() < 16) {
01103             StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
01104           }
01105           SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01106           SDValue CopyParamOps[] = { Chain,
01107                                      DAG.getConstant(paramCount, MVT::i32),
01108                                      DAG.getConstant(Offsets[j], MVT::i32),
01109                                      StVal, InFlag };
01110           Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
01111                                           CopyParamVTs, CopyParamOps,
01112                                           elemtype, MachinePointerInfo(),
01113                                           ArgAlign);
01114           InFlag = Chain.getValue(1);
01115           ++OIdx;
01116         }
01117         if (vtparts.size() > 0)
01118           --OIdx;
01119         ++paramCount;
01120         continue;
01121       }
01122       if (Ty->isVectorTy()) {
01123         EVT ObjectVT = getValueType(Ty);
01124         unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
01125         // declare .param .align <align> .b8 .param<n>[<size>];
01126         unsigned sz = TD->getTypeAllocSize(Ty);
01127         SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01128         SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32),
01129                                       DAG.getConstant(paramCount, MVT::i32),
01130                                       DAG.getConstant(sz, MVT::i32), InFlag };
01131         Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
01132                             DeclareParamOps);
01133         InFlag = Chain.getValue(1);
01134         unsigned NumElts = ObjectVT.getVectorNumElements();
01135         EVT EltVT = ObjectVT.getVectorElementType();
01136         EVT MemVT = EltVT;
01137         bool NeedExtend = false;
01138         if (EltVT.getSizeInBits() < 16) {
01139           NeedExtend = true;
01140           EltVT = MVT::i16;
01141         }
01142 
01143         // V1 store
01144         if (NumElts == 1) {
01145           SDValue Elt = OutVals[OIdx++];
01146           if (NeedExtend)
01147             Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt);
01148 
01149           SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01150           SDValue CopyParamOps[] = { Chain,
01151                                      DAG.getConstant(paramCount, MVT::i32),
01152                                      DAG.getConstant(0, MVT::i32), Elt,
01153                                      InFlag };
01154           Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
01155                                           CopyParamVTs, CopyParamOps,
01156                                           MemVT, MachinePointerInfo());
01157           InFlag = Chain.getValue(1);
01158         } else if (NumElts == 2) {
01159           SDValue Elt0 = OutVals[OIdx++];
01160           SDValue Elt1 = OutVals[OIdx++];
01161           if (NeedExtend) {
01162             Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0);
01163             Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1);
01164           }
01165 
01166           SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01167           SDValue CopyParamOps[] = { Chain,
01168                                      DAG.getConstant(paramCount, MVT::i32),
01169                                      DAG.getConstant(0, MVT::i32), Elt0, Elt1,
01170                                      InFlag };
01171           Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl,
01172                                           CopyParamVTs, CopyParamOps,
01173                                           MemVT, MachinePointerInfo());
01174           InFlag = Chain.getValue(1);
01175         } else {
01176           unsigned curOffset = 0;
01177           // V4 stores
01178           // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
01179           // the
01180           // vector will be expanded to a power of 2 elements, so we know we can
01181           // always round up to the next multiple of 4 when creating the vector
01182           // stores.
01183           // e.g.  4 elem => 1 st.v4
01184           //       6 elem => 2 st.v4
01185           //       8 elem => 2 st.v4
01186           //      11 elem => 3 st.v4
01187           unsigned VecSize = 4;
01188           if (EltVT.getSizeInBits() == 64)
01189             VecSize = 2;
01190 
01191           // This is potentially only part of a vector, so assume all elements
01192           // are packed together.
01193           unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize;
01194 
01195           for (unsigned i = 0; i < NumElts; i += VecSize) {
01196             // Get values
01197             SDValue StoreVal;
01198             SmallVector<SDValue, 8> Ops;
01199             Ops.push_back(Chain);
01200             Ops.push_back(DAG.getConstant(paramCount, MVT::i32));
01201             Ops.push_back(DAG.getConstant(curOffset, MVT::i32));
01202 
01203             unsigned Opc = NVPTXISD::StoreParamV2;
01204 
01205             StoreVal = OutVals[OIdx++];
01206             if (NeedExtend)
01207               StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
01208             Ops.push_back(StoreVal);
01209 
01210             if (i + 1 < NumElts) {
01211               StoreVal = OutVals[OIdx++];
01212               if (NeedExtend)
01213                 StoreVal =
01214                     DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
01215             } else {
01216               StoreVal = DAG.getUNDEF(EltVT);
01217             }
01218             Ops.push_back(StoreVal);
01219 
01220             if (VecSize == 4) {
01221               Opc = NVPTXISD::StoreParamV4;
01222               if (i + 2 < NumElts) {
01223                 StoreVal = OutVals[OIdx++];
01224                 if (NeedExtend)
01225                   StoreVal =
01226                       DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
01227               } else {
01228                 StoreVal = DAG.getUNDEF(EltVT);
01229               }
01230               Ops.push_back(StoreVal);
01231 
01232               if (i + 3 < NumElts) {
01233                 StoreVal = OutVals[OIdx++];
01234                 if (NeedExtend)
01235                   StoreVal =
01236                       DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
01237               } else {
01238                 StoreVal = DAG.getUNDEF(EltVT);
01239               }
01240               Ops.push_back(StoreVal);
01241             }
01242 
01243             Ops.push_back(InFlag);
01244 
01245             SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01246             Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops,
01247                                             MemVT, MachinePointerInfo());
01248             InFlag = Chain.getValue(1);
01249             curOffset += PerStoreOffset;
01250           }
01251         }
01252         ++paramCount;
01253         --OIdx;
01254         continue;
01255       }
01256       // Plain scalar
01257       // for ABI,    declare .param .b<size> .param<n>;
01258       unsigned sz = VT.getSizeInBits();
01259       bool needExtend = false;
01260       if (VT.isInteger()) {
01261         if (sz < 16)
01262           needExtend = true;
01263         if (sz < 32)
01264           sz = 32;
01265       }
01266       SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01267       SDValue DeclareParamOps[] = { Chain,
01268                                     DAG.getConstant(paramCount, MVT::i32),
01269                                     DAG.getConstant(sz, MVT::i32),
01270                                     DAG.getConstant(0, MVT::i32), InFlag };
01271       Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
01272                           DeclareParamOps);
01273       InFlag = Chain.getValue(1);
01274       SDValue OutV = OutVals[OIdx];
01275       if (needExtend) {
01276         // zext/sext i1 to i16
01277         unsigned opc = ISD::ZERO_EXTEND;
01278         if (Outs[OIdx].Flags.isSExt())
01279           opc = ISD::SIGN_EXTEND;
01280         OutV = DAG.getNode(opc, dl, MVT::i16, OutV);
01281       }
01282       SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01283       SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
01284                                  DAG.getConstant(0, MVT::i32), OutV, InFlag };
01285 
01286       unsigned opcode = NVPTXISD::StoreParam;
01287       if (Outs[OIdx].Flags.isZExt())
01288         opcode = NVPTXISD::StoreParamU32;
01289       else if (Outs[OIdx].Flags.isSExt())
01290         opcode = NVPTXISD::StoreParamS32;
01291       Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps,
01292                                       VT, MachinePointerInfo());
01293 
01294       InFlag = Chain.getValue(1);
01295       ++paramCount;
01296       continue;
01297     }
01298     // struct or vector
01299     SmallVector<EVT, 16> vtparts;
01300     SmallVector<uint64_t, 16> Offsets;
01301     const PointerType *PTy = dyn_cast<PointerType>(Args[i].Ty);
01302     assert(PTy && "Type of a byval parameter should be pointer");
01303     ComputePTXValueVTs(*this, PTy->getElementType(), vtparts, &Offsets, 0);
01304 
01305     // declare .param .align <align> .b8 .param<n>[<size>];
01306     unsigned sz = Outs[OIdx].Flags.getByValSize();
01307     SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01308     unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
01309     // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
01310     // so we don't need to worry about natural alignment or not.
01311     // See TargetLowering::LowerCallTo().
01312     SDValue DeclareParamOps[] = {
01313       Chain, DAG.getConstant(Outs[OIdx].Flags.getByValAlign(), MVT::i32),
01314       DAG.getConstant(paramCount, MVT::i32), DAG.getConstant(sz, MVT::i32),
01315       InFlag
01316     };
01317     Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
01318                         DeclareParamOps);
01319     InFlag = Chain.getValue(1);
01320     for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
01321       EVT elemtype = vtparts[j];
01322       int curOffset = Offsets[j];
01323       unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
01324       SDValue srcAddr =
01325           DAG.getNode(ISD::ADD, dl, getPointerTy(), OutVals[OIdx],
01326                       DAG.getConstant(curOffset, getPointerTy()));
01327       SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
01328                                    MachinePointerInfo(), false, false, false,
01329                                    PartAlign);
01330       if (elemtype.getSizeInBits() < 16) {
01331         theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
01332       }
01333       SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01334       SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
01335                                  DAG.getConstant(curOffset, MVT::i32), theVal,
01336                                  InFlag };
01337       Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
01338                                       CopyParamOps, elemtype,
01339                                       MachinePointerInfo());
01340 
01341       InFlag = Chain.getValue(1);
01342     }
01343     ++paramCount;
01344   }
01345 
01346   GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
01347   unsigned retAlignment = 0;
01348 
01349   // Handle Result
01350   if (Ins.size() > 0) {
01351     SmallVector<EVT, 16> resvtparts;
01352     ComputeValueVTs(*this, retTy, resvtparts);
01353 
01354     // Declare
01355     //  .param .align 16 .b8 retval0[<size-in-bytes>], or
01356     //  .param .b<size-in-bits> retval0
01357     unsigned resultsz = TD->getTypeAllocSizeInBits(retTy);
01358     if (retTy->isSingleValueType()) {
01359       // Scalar needs to be at least 32bit wide
01360       if (resultsz < 32)
01361         resultsz = 32;
01362       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01363       SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, MVT::i32),
01364                                   DAG.getConstant(resultsz, MVT::i32),
01365                                   DAG.getConstant(0, MVT::i32), InFlag };
01366       Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
01367                           DeclareRetOps);
01368       InFlag = Chain.getValue(1);
01369     } else {
01370       retAlignment = getArgumentAlignment(Callee, CS, retTy, 0);
01371       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01372       SDValue DeclareRetOps[] = { Chain,
01373                                   DAG.getConstant(retAlignment, MVT::i32),
01374                                   DAG.getConstant(resultsz / 8, MVT::i32),
01375                                   DAG.getConstant(0, MVT::i32), InFlag };
01376       Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
01377                           DeclareRetOps);
01378       InFlag = Chain.getValue(1);
01379     }
01380   }
01381 
01382   if (!Func) {
01383     // This is indirect function call case : PTX requires a prototype of the
01384     // form
01385     // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
01386     // to be emitted, and the label has to used as the last arg of call
01387     // instruction.
01388     // The prototype is embedded in a string and put as the operand for a
01389     // CallPrototype SDNode which will print out to the value of the string.
01390     SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01391     std::string Proto = getPrototype(retTy, Args, Outs, retAlignment, CS);
01392     const char *ProtoStr =
01393       nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
01394     SDValue ProtoOps[] = {
01395       Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
01396     };
01397     Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
01398     InFlag = Chain.getValue(1);
01399   }
01400   // Op to just print "call"
01401   SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01402   SDValue PrintCallOps[] = {
01403     Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, MVT::i32), InFlag
01404   };
01405   Chain = DAG.getNode(Func ? (NVPTXISD::PrintCallUni) : (NVPTXISD::PrintCall),
01406                       dl, PrintCallVTs, PrintCallOps);
01407   InFlag = Chain.getValue(1);
01408 
01409   // Ops to print out the function name
01410   SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01411   SDValue CallVoidOps[] = { Chain, Callee, InFlag };
01412   Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps);
01413   InFlag = Chain.getValue(1);
01414 
01415   // Ops to print out the param list
01416   SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01417   SDValue CallArgBeginOps[] = { Chain, InFlag };
01418   Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
01419                       CallArgBeginOps);
01420   InFlag = Chain.getValue(1);
01421 
01422   for (unsigned i = 0, e = paramCount; i != e; ++i) {
01423     unsigned opcode;
01424     if (i == (e - 1))
01425       opcode = NVPTXISD::LastCallArg;
01426     else
01427       opcode = NVPTXISD::CallArg;
01428     SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01429     SDValue CallArgOps[] = { Chain, DAG.getConstant(1, MVT::i32),
01430                              DAG.getConstant(i, MVT::i32), InFlag };
01431     Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps);
01432     InFlag = Chain.getValue(1);
01433   }
01434   SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01435   SDValue CallArgEndOps[] = { Chain, DAG.getConstant(Func ? 1 : 0, MVT::i32),
01436                               InFlag };
01437   Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps);
01438   InFlag = Chain.getValue(1);
01439 
01440   if (!Func) {
01441     SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
01442     SDValue PrototypeOps[] = { Chain, DAG.getConstant(uniqueCallSite, MVT::i32),
01443                                InFlag };
01444     Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
01445     InFlag = Chain.getValue(1);
01446   }
01447 
01448   // Generate loads from param memory/moves from registers for result
01449   if (Ins.size() > 0) {
01450     if (retTy && retTy->isVectorTy()) {
01451       EVT ObjectVT = getValueType(retTy);
01452       unsigned NumElts = ObjectVT.getVectorNumElements();
01453       EVT EltVT = ObjectVT.getVectorElementType();
01454       assert(nvTM->getSubtargetImpl()->getTargetLowering()->getNumRegisters(
01455                  F->getContext(), ObjectVT) == NumElts &&
01456              "Vector was not scalarized");
01457       unsigned sz = EltVT.getSizeInBits();
01458       bool needTruncate = sz < 8 ? true : false;
01459 
01460       if (NumElts == 1) {
01461         // Just a simple load
01462         SmallVector<EVT, 4> LoadRetVTs;
01463         if (EltVT == MVT::i1 || EltVT == MVT::i8) {
01464           // If loading i1/i8 result, generate
01465           //   load.b8 i16
01466           //   if i1
01467           //   trunc i16 to i1
01468           LoadRetVTs.push_back(MVT::i16);
01469         } else
01470           LoadRetVTs.push_back(EltVT);
01471         LoadRetVTs.push_back(MVT::Other);
01472         LoadRetVTs.push_back(MVT::Glue);
01473         SmallVector<SDValue, 4> LoadRetOps;
01474         LoadRetOps.push_back(Chain);
01475         LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
01476         LoadRetOps.push_back(DAG.getConstant(0, MVT::i32));
01477         LoadRetOps.push_back(InFlag);
01478         SDValue retval = DAG.getMemIntrinsicNode(
01479             NVPTXISD::LoadParam, dl,
01480             DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
01481         Chain = retval.getValue(1);
01482         InFlag = retval.getValue(2);
01483         SDValue Ret0 = retval;
01484         if (needTruncate)
01485           Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0);
01486         InVals.push_back(Ret0);
01487       } else if (NumElts == 2) {
01488         // LoadV2
01489         SmallVector<EVT, 4> LoadRetVTs;
01490         if (EltVT == MVT::i1 || EltVT == MVT::i8) {
01491           // If loading i1/i8 result, generate
01492           //   load.b8 i16
01493           //   if i1
01494           //   trunc i16 to i1
01495           LoadRetVTs.push_back(MVT::i16);
01496           LoadRetVTs.push_back(MVT::i16);
01497         } else {
01498           LoadRetVTs.push_back(EltVT);
01499           LoadRetVTs.push_back(EltVT);
01500         }
01501         LoadRetVTs.push_back(MVT::Other);
01502         LoadRetVTs.push_back(MVT::Glue);
01503         SmallVector<SDValue, 4> LoadRetOps;
01504         LoadRetOps.push_back(Chain);
01505         LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
01506         LoadRetOps.push_back(DAG.getConstant(0, MVT::i32));
01507         LoadRetOps.push_back(InFlag);
01508         SDValue retval = DAG.getMemIntrinsicNode(
01509             NVPTXISD::LoadParamV2, dl,
01510             DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
01511         Chain = retval.getValue(2);
01512         InFlag = retval.getValue(3);
01513         SDValue Ret0 = retval.getValue(0);
01514         SDValue Ret1 = retval.getValue(1);
01515         if (needTruncate) {
01516           Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0);
01517           InVals.push_back(Ret0);
01518           Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1);
01519           InVals.push_back(Ret1);
01520         } else {
01521           InVals.push_back(Ret0);
01522           InVals.push_back(Ret1);
01523         }
01524       } else {
01525         // Split into N LoadV4
01526         unsigned Ofst = 0;
01527         unsigned VecSize = 4;
01528         unsigned Opc = NVPTXISD::LoadParamV4;
01529         if (EltVT.getSizeInBits() == 64) {
01530           VecSize = 2;
01531           Opc = NVPTXISD::LoadParamV2;
01532         }
01533         EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
01534         for (unsigned i = 0; i < NumElts; i += VecSize) {
01535           SmallVector<EVT, 8> LoadRetVTs;
01536           if (EltVT == MVT::i1 || EltVT == MVT::i8) {
01537             // If loading i1/i8 result, generate
01538             //   load.b8 i16
01539             //   if i1
01540             //   trunc i16 to i1
01541             for (unsigned j = 0; j < VecSize; ++j)
01542               LoadRetVTs.push_back(MVT::i16);
01543           } else {
01544             for (unsigned j = 0; j < VecSize; ++j)
01545               LoadRetVTs.push_back(EltVT);
01546           }
01547           LoadRetVTs.push_back(MVT::Other);
01548           LoadRetVTs.push_back(MVT::Glue);
01549           SmallVector<SDValue, 4> LoadRetOps;
01550           LoadRetOps.push_back(Chain);
01551           LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
01552           LoadRetOps.push_back(DAG.getConstant(Ofst, MVT::i32));
01553           LoadRetOps.push_back(InFlag);
01554           SDValue retval = DAG.getMemIntrinsicNode(
01555               Opc, dl, DAG.getVTList(LoadRetVTs),
01556               LoadRetOps, EltVT, MachinePointerInfo());
01557           if (VecSize == 2) {
01558             Chain = retval.getValue(2);
01559             InFlag = retval.getValue(3);
01560           } else {
01561             Chain = retval.getValue(4);
01562             InFlag = retval.getValue(5);
01563           }
01564 
01565           for (unsigned j = 0; j < VecSize; ++j) {
01566             if (i + j >= NumElts)
01567               break;
01568             SDValue Elt = retval.getValue(j);
01569             if (needTruncate)
01570               Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
01571             InVals.push_back(Elt);
01572           }
01573           Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
01574         }
01575       }
01576     } else {
01577       SmallVector<EVT, 16> VTs;
01578       SmallVector<uint64_t, 16> Offsets;
01579       ComputePTXValueVTs(*this, retTy, VTs, &Offsets, 0);
01580       assert(VTs.size() == Ins.size() && "Bad value decomposition");
01581       unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0);
01582       for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
01583         unsigned sz = VTs[i].getSizeInBits();
01584         unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);
01585         bool needTruncate = sz < 8 ? true : false;
01586         if (VTs[i].isInteger() && (sz < 8))
01587           sz = 8;
01588 
01589         SmallVector<EVT, 4> LoadRetVTs;
01590         EVT TheLoadType = VTs[i];
01591         if (retTy->isIntegerTy() &&
01592             TD->getTypeAllocSizeInBits(retTy) < 32) {
01593           // This is for integer types only, and specifically not for
01594           // aggregates.
01595           LoadRetVTs.push_back(MVT::i32);
01596           TheLoadType = MVT::i32;
01597         } else if (sz < 16) {
01598           // If loading i1/i8 result, generate
01599           //   load i8 (-> i16)
01600           //   trunc i16 to i1/i8
01601           LoadRetVTs.push_back(MVT::i16);
01602         } else
01603           LoadRetVTs.push_back(Ins[i].VT);
01604         LoadRetVTs.push_back(MVT::Other);
01605         LoadRetVTs.push_back(MVT::Glue);
01606 
01607         SmallVector<SDValue, 4> LoadRetOps;
01608         LoadRetOps.push_back(Chain);
01609         LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
01610         LoadRetOps.push_back(DAG.getConstant(Offsets[i], MVT::i32));
01611         LoadRetOps.push_back(InFlag);
01612         SDValue retval = DAG.getMemIntrinsicNode(
01613             NVPTXISD::LoadParam, dl,
01614             DAG.getVTList(LoadRetVTs), LoadRetOps,
01615             TheLoadType, MachinePointerInfo(), AlignI);
01616         Chain = retval.getValue(1);
01617         InFlag = retval.getValue(2);
01618         SDValue Ret0 = retval.getValue(0);
01619         if (needTruncate)
01620           Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0);
01621         InVals.push_back(Ret0);
01622       }
01623     }
01624   }
01625 
01626   Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(uniqueCallSite, true),
01627                              DAG.getIntPtrConstant(uniqueCallSite + 1, true),
01628                              InFlag, dl);
01629   uniqueCallSite++;
01630 
01631   // set isTailCall to false for now, until we figure out how to express
01632   // tail call optimization in PTX
01633   isTailCall = false;
01634   return Chain;
01635 }
01636 
01637 // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
01638 // (see LegalizeDAG.cpp). This is slow and uses local memory.
01639 // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
01640 SDValue
01641 NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
01642   SDNode *Node = Op.getNode();
01643   SDLoc dl(Node);
01644   SmallVector<SDValue, 8> Ops;
01645   unsigned NumOperands = Node->getNumOperands();
01646   for (unsigned i = 0; i < NumOperands; ++i) {
01647     SDValue SubOp = Node->getOperand(i);
01648     EVT VVT = SubOp.getNode()->getValueType(0);
01649     EVT EltVT = VVT.getVectorElementType();
01650     unsigned NumSubElem = VVT.getVectorNumElements();
01651     for (unsigned j = 0; j < NumSubElem; ++j) {
01652       Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
01653                                 DAG.getIntPtrConstant(j)));
01654     }
01655   }
01656   return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Ops);
01657 }
01658 
01659 /// LowerShiftRightParts - Lower SRL_PARTS, SRA_PARTS, which
01660 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
01661 ///    amount, or
01662 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
01663 ///    amount.
01664 SDValue NVPTXTargetLowering::LowerShiftRightParts(SDValue Op,
01665                                                   SelectionDAG &DAG) const {
01666   assert(Op.getNumOperands() == 3 && "Not a double-shift!");
01667   assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
01668 
01669   EVT VT = Op.getValueType();
01670   unsigned VTBits = VT.getSizeInBits();
01671   SDLoc dl(Op);
01672   SDValue ShOpLo = Op.getOperand(0);
01673   SDValue ShOpHi = Op.getOperand(1);
01674   SDValue ShAmt  = Op.getOperand(2);
01675   unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
01676 
01677   if (VTBits == 32 && nvptxSubtarget.getSmVersion() >= 35) {
01678 
01679     // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
01680     // {dHi, dLo} = {aHi, aLo} >> Amt
01681     //   dHi = aHi >> Amt
01682     //   dLo = shf.r.clamp aLo, aHi, Amt
01683 
01684     SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
01685     SDValue Lo = DAG.getNode(NVPTXISD::FUN_SHFR_CLAMP, dl, VT, ShOpLo, ShOpHi,
01686                              ShAmt);
01687 
01688     SDValue Ops[2] = { Lo, Hi };
01689     return DAG.getMergeValues(Ops, dl);
01690   }
01691   else {
01692 
01693     // {dHi, dLo} = {aHi, aLo} >> Amt
01694     // - if (Amt>=size) then
01695     //      dLo = aHi >> (Amt-size)
01696     //      dHi = aHi >> Amt (this is either all 0 or all 1)
01697     //   else
01698     //      dLo = (aLo >>logic Amt) | (aHi << (size-Amt))
01699     //      dHi = aHi >> Amt
01700 
01701     SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
01702                                    DAG.getConstant(VTBits, MVT::i32), ShAmt);
01703     SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
01704     SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
01705                                      DAG.getConstant(VTBits, MVT::i32));
01706     SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
01707     SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
01708     SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
01709 
01710     SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
01711                                DAG.getConstant(VTBits, MVT::i32), ISD::SETGE);
01712     SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
01713     SDValue Lo = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
01714 
01715     SDValue Ops[2] = { Lo, Hi };
01716     return DAG.getMergeValues(Ops, dl);
01717   }
01718 }
01719 
01720 /// LowerShiftLeftParts - Lower SHL_PARTS, which
01721 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
01722 ///    amount, or
01723 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
01724 ///    amount.
01725 SDValue NVPTXTargetLowering::LowerShiftLeftParts(SDValue Op,
01726                                                  SelectionDAG &DAG) const {
01727   assert(Op.getNumOperands() == 3 && "Not a double-shift!");
01728   assert(Op.getOpcode() == ISD::SHL_PARTS);
01729 
01730   EVT VT = Op.getValueType();
01731   unsigned VTBits = VT.getSizeInBits();
01732   SDLoc dl(Op);
01733   SDValue ShOpLo = Op.getOperand(0);
01734   SDValue ShOpHi = Op.getOperand(1);
01735   SDValue ShAmt  = Op.getOperand(2);
01736 
01737   if (VTBits == 32 && nvptxSubtarget.getSmVersion() >= 35) {
01738 
01739     // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
01740     // {dHi, dLo} = {aHi, aLo} << Amt
01741     //   dHi = shf.l.clamp aLo, aHi, Amt
01742     //   dLo = aLo << Amt
01743 
01744     SDValue Hi = DAG.getNode(NVPTXISD::FUN_SHFL_CLAMP, dl, VT, ShOpLo, ShOpHi,
01745                              ShAmt);
01746     SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
01747 
01748     SDValue Ops[2] = { Lo, Hi };
01749     return DAG.getMergeValues(Ops, dl);
01750   }
01751   else {
01752 
01753     // {dHi, dLo} = {aHi, aLo} << Amt
01754     // - if (Amt>=size) then
01755     //      dLo = aLo << Amt (all 0)
01756     //      dLo = aLo << (Amt-size)
01757     //   else
01758     //      dLo = aLo << Amt
01759     //      dHi = (aHi << Amt) | (aLo >> (size-Amt))
01760 
01761     SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
01762                                    DAG.getConstant(VTBits, MVT::i32), ShAmt);
01763     SDValue Tmp1 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
01764     SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
01765                                      DAG.getConstant(VTBits, MVT::i32));
01766     SDValue Tmp2 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
01767     SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
01768     SDValue TrueVal = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
01769 
01770     SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
01771                                DAG.getConstant(VTBits, MVT::i32), ISD::SETGE);
01772     SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
01773     SDValue Hi = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
01774 
01775     SDValue Ops[2] = { Lo, Hi };
01776     return DAG.getMergeValues(Ops, dl);
01777   }
01778 }
01779 
01780 SDValue
01781 NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
01782   switch (Op.getOpcode()) {
01783   case ISD::RETURNADDR:
01784     return SDValue();
01785   case ISD::FRAMEADDR:
01786     return SDValue();
01787   case ISD::GlobalAddress:
01788     return LowerGlobalAddress(Op, DAG);
01789   case ISD::INTRINSIC_W_CHAIN:
01790     return Op;
01791   case ISD::BUILD_VECTOR:
01792   case ISD::EXTRACT_SUBVECTOR:
01793     return Op;
01794   case ISD::CONCAT_VECTORS:
01795     return LowerCONCAT_VECTORS(Op, DAG);
01796   case ISD::STORE:
01797     return LowerSTORE(Op, DAG);
01798   case ISD::LOAD:
01799     return LowerLOAD(Op, DAG);
01800   case ISD::SHL_PARTS:
01801     return LowerShiftLeftParts(Op, DAG);
01802   case ISD::SRA_PARTS:
01803   case ISD::SRL_PARTS:
01804     return LowerShiftRightParts(Op, DAG);
01805   default:
01806     llvm_unreachable("Custom lowering not defined for operation");
01807   }
01808 }
01809 
01810 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
01811   if (Op.getValueType() == MVT::i1)
01812     return LowerLOADi1(Op, DAG);
01813   else
01814     return SDValue();
01815 }
01816 
01817 // v = ld i1* addr
01818 //   =>
01819 // v1 = ld i8* addr (-> i16)
01820 // v = trunc i16 to i1
01821 SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const {
01822   SDNode *Node = Op.getNode();
01823   LoadSDNode *LD = cast<LoadSDNode>(Node);
01824   SDLoc dl(Node);
01825   assert(LD->getExtensionType() == ISD::NON_EXTLOAD);
01826   assert(Node->getValueType(0) == MVT::i1 &&
01827          "Custom lowering for i1 load only");
01828   SDValue newLD =
01829       DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(),
01830                   LD->getPointerInfo(), LD->isVolatile(), LD->isNonTemporal(),
01831                   LD->isInvariant(), LD->getAlignment());
01832   SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD);
01833   // The legalizer (the caller) is expecting two values from the legalized
01834   // load, so we build a MergeValues node for it. See ExpandUnalignedLoad()
01835   // in LegalizeDAG.cpp which also uses MergeValues.
01836   SDValue Ops[] = { result, LD->getChain() };
01837   return DAG.getMergeValues(Ops, dl);
01838 }
01839 
01840 SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
01841   EVT ValVT = Op.getOperand(1).getValueType();
01842   if (ValVT == MVT::i1)
01843     return LowerSTOREi1(Op, DAG);
01844   else if (ValVT.isVector())
01845     return LowerSTOREVector(Op, DAG);
01846   else
01847     return SDValue();
01848 }
01849 
01850 SDValue
01851 NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const {
01852   SDNode *N = Op.getNode();
01853   SDValue Val = N->getOperand(1);
01854   SDLoc DL(N);
01855   EVT ValVT = Val.getValueType();
01856 
01857   if (ValVT.isVector()) {
01858     // We only handle "native" vector sizes for now, e.g. <4 x double> is not
01859     // legal.  We can (and should) split that into 2 stores of <2 x double> here
01860     // but I'm leaving that as a TODO for now.
01861     if (!ValVT.isSimple())
01862       return SDValue();
01863     switch (ValVT.getSimpleVT().SimpleTy) {
01864     default:
01865       return SDValue();
01866     case MVT::v2i8:
01867     case MVT::v2i16:
01868     case MVT::v2i32:
01869     case MVT::v2i64:
01870     case MVT::v2f32:
01871     case MVT::v2f64:
01872     case MVT::v4i8:
01873     case MVT::v4i16:
01874     case MVT::v4i32:
01875     case MVT::v4f32:
01876       // This is a "native" vector type
01877       break;
01878     }
01879 
01880     MemSDNode *MemSD = cast<MemSDNode>(N);
01881     const DataLayout *TD = getDataLayout();
01882 
01883     unsigned Align = MemSD->getAlignment();
01884     unsigned PrefAlign =
01885       TD->getPrefTypeAlignment(ValVT.getTypeForEVT(*DAG.getContext()));
01886     if (Align < PrefAlign) {
01887       // This store is not sufficiently aligned, so bail out and let this vector
01888       // store be scalarized.  Note that we may still be able to emit smaller
01889       // vector stores.  For example, if we are storing a <4 x float> with an
01890       // alignment of 8, this check will fail but the legalizer will try again
01891       // with 2 x <2 x float>, which will succeed with an alignment of 8.
01892       return SDValue();
01893     }
01894 
01895     unsigned Opcode = 0;
01896     EVT EltVT = ValVT.getVectorElementType();
01897     unsigned NumElts = ValVT.getVectorNumElements();
01898 
01899     // Since StoreV2 is a target node, we cannot rely on DAG type legalization.
01900     // Therefore, we must ensure the type is legal.  For i1 and i8, we set the
01901     // stored type to i16 and propagate the "real" type as the memory type.
01902     bool NeedExt = false;
01903     if (EltVT.getSizeInBits() < 16)
01904       NeedExt = true;
01905 
01906     switch (NumElts) {
01907     default:
01908       return SDValue();
01909     case 2:
01910       Opcode = NVPTXISD::StoreV2;
01911       break;
01912     case 4: {
01913       Opcode = NVPTXISD::StoreV4;
01914       break;
01915     }
01916     }
01917 
01918     SmallVector<SDValue, 8> Ops;
01919 
01920     // First is the chain
01921     Ops.push_back(N->getOperand(0));
01922 
01923     // Then the split values
01924     for (unsigned i = 0; i < NumElts; ++i) {
01925       SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val,
01926                                    DAG.getIntPtrConstant(i));
01927       if (NeedExt)
01928         ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal);
01929       Ops.push_back(ExtVal);
01930     }
01931 
01932     // Then any remaining arguments
01933     for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
01934       Ops.push_back(N->getOperand(i));
01935     }
01936 
01937     SDValue NewSt = DAG.getMemIntrinsicNode(
01938         Opcode, DL, DAG.getVTList(MVT::Other), Ops,
01939         MemSD->getMemoryVT(), MemSD->getMemOperand());
01940 
01941     //return DCI.CombineTo(N, NewSt, true);
01942     return NewSt;
01943   }
01944 
01945   return SDValue();
01946 }
01947 
01948 // st i1 v, addr
01949 //    =>
01950 // v1 = zxt v to i16
01951 // st.u8 i16, addr
01952 SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const {
01953   SDNode *Node = Op.getNode();
01954   SDLoc dl(Node);
01955   StoreSDNode *ST = cast<StoreSDNode>(Node);
01956   SDValue Tmp1 = ST->getChain();
01957   SDValue Tmp2 = ST->getBasePtr();
01958   SDValue Tmp3 = ST->getValue();
01959   assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only");
01960   unsigned Alignment = ST->getAlignment();
01961   bool isVolatile = ST->isVolatile();
01962   bool isNonTemporal = ST->isNonTemporal();
01963   Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3);
01964   SDValue Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
01965                                      ST->getPointerInfo(), MVT::i8, isNonTemporal,
01966                                      isVolatile, Alignment);
01967   return Result;
01968 }
01969 
01970 SDValue NVPTXTargetLowering::getExtSymb(SelectionDAG &DAG, const char *inname,
01971                                         int idx, EVT v) const {
01972   std::string *name = nvTM->getManagedStrPool()->getManagedString(inname);
01973   std::stringstream suffix;
01974   suffix << idx;
01975   *name += suffix.str();
01976   return DAG.getTargetExternalSymbol(name->c_str(), v);
01977 }
01978 
01979 SDValue
01980 NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const {
01981   std::string ParamSym;
01982   raw_string_ostream ParamStr(ParamSym);
01983 
01984   ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx;
01985   ParamStr.flush();
01986 
01987   std::string *SavedStr =
01988     nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str());
01989   return DAG.getTargetExternalSymbol(SavedStr->c_str(), v);
01990 }
01991 
01992 SDValue NVPTXTargetLowering::getParamHelpSymbol(SelectionDAG &DAG, int idx) {
01993   return getExtSymb(DAG, ".HLPPARAM", idx);
01994 }
01995 
01996 // Check to see if the kernel argument is image*_t or sampler_t
01997 
01998 bool llvm::isImageOrSamplerVal(const Value *arg, const Module *context) {
01999   static const char *const specialTypes[] = { "struct._image2d_t",
02000                                               "struct._image3d_t",
02001                                               "struct._sampler_t" };
02002 
02003   const Type *Ty = arg->getType();
02004   const PointerType *PTy = dyn_cast<PointerType>(Ty);
02005 
02006   if (!PTy)
02007     return false;
02008 
02009   if (!context)
02010     return false;
02011 
02012   const StructType *STy = dyn_cast<StructType>(PTy->getElementType());
02013   const std::string TypeName = STy && !STy->isLiteral() ? STy->getName() : "";
02014 
02015   for (int i = 0, e = array_lengthof(specialTypes); i != e; ++i)
02016     if (TypeName == specialTypes[i])
02017       return true;
02018 
02019   return false;
02020 }
02021 
02022 SDValue NVPTXTargetLowering::LowerFormalArguments(
02023     SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
02024     const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
02025     SmallVectorImpl<SDValue> &InVals) const {
02026   MachineFunction &MF = DAG.getMachineFunction();
02027   const DataLayout *TD = getDataLayout();
02028 
02029   const Function *F = MF.getFunction();
02030   const AttributeSet &PAL = F->getAttributes();
02031   const TargetLowering *TLI = DAG.getSubtarget().getTargetLowering();
02032 
02033   SDValue Root = DAG.getRoot();
02034   std::vector<SDValue> OutChains;
02035 
02036   bool isKernel = llvm::isKernelFunction(*F);
02037   bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
02038   assert(isABI && "Non-ABI compilation is not supported");
02039   if (!isABI)
02040     return Chain;
02041 
02042   std::vector<Type *> argTypes;
02043   std::vector<const Argument *> theArgs;
02044   for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
02045        I != E; ++I) {
02046     theArgs.push_back(I);
02047     argTypes.push_back(I->getType());
02048   }
02049   // argTypes.size() (or theArgs.size()) and Ins.size() need not match.
02050   // Ins.size() will be larger
02051   //   * if there is an aggregate argument with multiple fields (each field
02052   //     showing up separately in Ins)
02053   //   * if there is a vector argument with more than typical vector-length
02054   //     elements (generally if more than 4) where each vector element is
02055   //     individually present in Ins.
02056   // So a different index should be used for indexing into Ins.
02057   // See similar issue in LowerCall.
02058   unsigned InsIdx = 0;
02059 
02060   int idx = 0;
02061   for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) {
02062     Type *Ty = argTypes[i];
02063 
02064     // If the kernel argument is image*_t or sampler_t, convert it to
02065     // a i32 constant holding the parameter position. This can later
02066     // matched in the AsmPrinter to output the correct mangled name.
02067     if (isImageOrSamplerVal(
02068             theArgs[i],
02069             (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent()
02070                                      : nullptr))) {
02071       assert(isKernel && "Only kernels can have image/sampler params");
02072       InVals.push_back(DAG.getConstant(i + 1, MVT::i32));
02073       continue;
02074     }
02075 
02076     if (theArgs[i]->use_empty()) {
02077       // argument is dead
02078       if (Ty->isAggregateType()) {
02079         SmallVector<EVT, 16> vtparts;
02080 
02081         ComputePTXValueVTs(*this, Ty, vtparts);
02082         assert(vtparts.size() > 0 && "empty aggregate type not expected");
02083         for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
02084              ++parti) {
02085           InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
02086           ++InsIdx;
02087         }
02088         if (vtparts.size() > 0)
02089           --InsIdx;
02090         continue;
02091       }
02092       if (Ty->isVectorTy()) {
02093         EVT ObjectVT = getValueType(Ty);
02094         unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT);
02095         for (unsigned parti = 0; parti < NumRegs; ++parti) {
02096           InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
02097           ++InsIdx;
02098         }
02099         if (NumRegs > 0)
02100           --InsIdx;
02101         continue;
02102       }
02103       InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
02104       continue;
02105     }
02106 
02107     // In the following cases, assign a node order of "idx+1"
02108     // to newly created nodes. The SDNodes for params have to
02109     // appear in the same order as their order of appearance
02110     // in the original function. "idx+1" holds that order.
02111     if (PAL.hasAttribute(i + 1, Attribute::ByVal) == false) {
02112       if (Ty->isAggregateType()) {
02113         SmallVector<EVT, 16> vtparts;
02114         SmallVector<uint64_t, 16> offsets;
02115 
02116         // NOTE: Here, we lose the ability to issue vector loads for vectors
02117         // that are a part of a struct.  This should be investigated in the
02118         // future.
02119         ComputePTXValueVTs(*this, Ty, vtparts, &offsets, 0);
02120         assert(vtparts.size() > 0 && "empty aggregate type not expected");
02121         bool aggregateIsPacked = false;
02122         if (StructType *STy = llvm::dyn_cast<StructType>(Ty))
02123           aggregateIsPacked = STy->isPacked();
02124 
02125         SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
02126         for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
02127              ++parti) {
02128           EVT partVT = vtparts[parti];
02129           Value *srcValue = Constant::getNullValue(
02130               PointerType::get(partVT.getTypeForEVT(F->getContext()),
02131                                llvm::ADDRESS_SPACE_PARAM));
02132           SDValue srcAddr =
02133               DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
02134                           DAG.getConstant(offsets[parti], getPointerTy()));
02135           unsigned partAlign =
02136               aggregateIsPacked ? 1
02137                                 : TD->getABITypeAlignment(
02138                                       partVT.getTypeForEVT(F->getContext()));
02139           SDValue p;
02140           if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) {
02141             ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ? 
02142                                      ISD::SEXTLOAD : ISD::ZEXTLOAD;
02143             p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr,
02144                                MachinePointerInfo(srcValue), partVT, false,
02145                                false, false, partAlign);
02146           } else {
02147             p = DAG.getLoad(partVT, dl, Root, srcAddr,
02148                             MachinePointerInfo(srcValue), false, false, false,
02149                             partAlign);
02150           }
02151           if (p.getNode())
02152             p.getNode()->setIROrder(idx + 1);
02153           InVals.push_back(p);
02154           ++InsIdx;
02155         }
02156         if (vtparts.size() > 0)
02157           --InsIdx;
02158         continue;
02159       }
02160       if (Ty->isVectorTy()) {
02161         EVT ObjectVT = getValueType(Ty);
02162         SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
02163         unsigned NumElts = ObjectVT.getVectorNumElements();
02164         assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
02165                "Vector was not scalarized");
02166         unsigned Ofst = 0;
02167         EVT EltVT = ObjectVT.getVectorElementType();
02168 
02169         // V1 load
02170         // f32 = load ...
02171         if (NumElts == 1) {
02172           // We only have one element, so just directly load it
02173           Value *SrcValue = Constant::getNullValue(PointerType::get(
02174               EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
02175           SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
02176                                         DAG.getConstant(Ofst, getPointerTy()));
02177           SDValue P = DAG.getLoad(
02178               EltVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
02179               false, true,
02180               TD->getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())));
02181           if (P.getNode())
02182             P.getNode()->setIROrder(idx + 1);
02183 
02184           if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
02185             P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
02186           InVals.push_back(P);
02187           Ofst += TD->getTypeAllocSize(EltVT.getTypeForEVT(F->getContext()));
02188           ++InsIdx;
02189         } else if (NumElts == 2) {
02190           // V2 load
02191           // f32,f32 = load ...
02192           EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
02193           Value *SrcValue = Constant::getNullValue(PointerType::get(
02194               VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
02195           SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
02196                                         DAG.getConstant(Ofst, getPointerTy()));
02197           SDValue P = DAG.getLoad(
02198               VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
02199               false, true,
02200               TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
02201           if (P.getNode())
02202             P.getNode()->setIROrder(idx + 1);
02203 
02204           SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
02205                                      DAG.getIntPtrConstant(0));
02206           SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
02207                                      DAG.getIntPtrConstant(1));
02208 
02209           if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) {
02210             Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0);
02211             Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1);
02212           }
02213 
02214           InVals.push_back(Elt0);
02215           InVals.push_back(Elt1);
02216           Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
02217           InsIdx += 2;
02218         } else {
02219           // V4 loads
02220           // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
02221           // the
02222           // vector will be expanded to a power of 2 elements, so we know we can
02223           // always round up to the next multiple of 4 when creating the vector
02224           // loads.
02225           // e.g.  4 elem => 1 ld.v4
02226           //       6 elem => 2 ld.v4
02227           //       8 elem => 2 ld.v4
02228           //      11 elem => 3 ld.v4
02229           unsigned VecSize = 4;
02230           if (EltVT.getSizeInBits() == 64) {
02231             VecSize = 2;
02232           }
02233           EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
02234           for (unsigned i = 0; i < NumElts; i += VecSize) {
02235             Value *SrcValue = Constant::getNullValue(
02236                 PointerType::get(VecVT.getTypeForEVT(F->getContext()),
02237                                  llvm::ADDRESS_SPACE_PARAM));
02238             SDValue SrcAddr =
02239                 DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
02240                             DAG.getConstant(Ofst, getPointerTy()));
02241             SDValue P = DAG.getLoad(
02242                 VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
02243                 false, true,
02244                 TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
02245             if (P.getNode())
02246               P.getNode()->setIROrder(idx + 1);
02247 
02248             for (unsigned j = 0; j < VecSize; ++j) {
02249               if (i + j >= NumElts)
02250                 break;
02251               SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
02252                                         DAG.getIntPtrConstant(j));
02253               if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
02254                 Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt);
02255               InVals.push_back(Elt);
02256             }
02257             Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
02258           }
02259           InsIdx += NumElts;
02260         }
02261 
02262         if (NumElts > 0)
02263           --InsIdx;
02264         continue;
02265       }
02266       // A plain scalar.
02267       EVT ObjectVT = getValueType(Ty);
02268       // If ABI, load from the param symbol
02269       SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
02270       Value *srcValue = Constant::getNullValue(PointerType::get(
02271           ObjectVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
02272       SDValue p;
02273        if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) {
02274         ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ? 
02275                                        ISD::SEXTLOAD : ISD::ZEXTLOAD;
02276         p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, Arg,
02277                            MachinePointerInfo(srcValue), ObjectVT, false, false,
02278                            false,
02279         TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
02280       } else {
02281         p = DAG.getLoad(Ins[InsIdx].VT, dl, Root, Arg,
02282                         MachinePointerInfo(srcValue), false, false, false,
02283         TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
02284       }
02285       if (p.getNode())
02286         p.getNode()->setIROrder(idx + 1);
02287       InVals.push_back(p);
02288       continue;
02289     }
02290 
02291     // Param has ByVal attribute
02292     // Return MoveParam(param symbol).
02293     // Ideally, the param symbol can be returned directly,
02294     // but when SDNode builder decides to use it in a CopyToReg(),
02295     // machine instruction fails because TargetExternalSymbol
02296     // (not lowered) is target dependent, and CopyToReg assumes
02297     // the source is lowered.
02298     EVT ObjectVT = getValueType(Ty);
02299     assert(ObjectVT == Ins[InsIdx].VT &&
02300            "Ins type did not match function type");
02301     SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
02302     SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
02303     if (p.getNode())
02304       p.getNode()->setIROrder(idx + 1);
02305     if (isKernel)
02306       InVals.push_back(p);
02307     else {
02308       SDValue p2 = DAG.getNode(
02309           ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT,
02310           DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, MVT::i32), p);
02311       InVals.push_back(p2);
02312     }
02313   }
02314 
02315   // Clang will check explicit VarArg and issue error if any. However, Clang
02316   // will let code with
02317   // implicit var arg like f() pass. See bug 617733.
02318   // We treat this case as if the arg list is empty.
02319   // if (F.isVarArg()) {
02320   // assert(0 && "VarArg not supported yet!");
02321   //}
02322 
02323   if (!OutChains.empty())
02324     DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains));
02325 
02326   return Chain;
02327 }
02328 
02329 
02330 SDValue
02331 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
02332                                  bool isVarArg,
02333                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
02334                                  const SmallVectorImpl<SDValue> &OutVals,
02335                                  SDLoc dl, SelectionDAG &DAG) const {
02336   MachineFunction &MF = DAG.getMachineFunction();
02337   const Function *F = MF.getFunction();
02338   Type *RetTy = F->getReturnType();
02339   const DataLayout *TD = getDataLayout();
02340 
02341   bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
02342   assert(isABI && "Non-ABI compilation is not supported");
02343   if (!isABI)
02344     return Chain;
02345 
02346   if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) {
02347     // If we have a vector type, the OutVals array will be the scalarized
02348     // components and we have combine them into 1 or more vector stores.
02349     unsigned NumElts = VTy->getNumElements();
02350     assert(NumElts == Outs.size() && "Bad scalarization of return value");
02351 
02352     // const_cast can be removed in later LLVM versions
02353     EVT EltVT = getValueType(RetTy).getVectorElementType();
02354     bool NeedExtend = false;
02355     if (EltVT.getSizeInBits() < 16)
02356       NeedExtend = true;
02357 
02358     // V1 store
02359     if (NumElts == 1) {
02360       SDValue StoreVal = OutVals[0];
02361       // We only have one element, so just directly store it
02362       if (NeedExtend)
02363         StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
02364       SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal };
02365       Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
02366                                       DAG.getVTList(MVT::Other), Ops,
02367                                       EltVT, MachinePointerInfo());
02368 
02369     } else if (NumElts == 2) {
02370       // V2 store
02371       SDValue StoreVal0 = OutVals[0];
02372       SDValue StoreVal1 = OutVals[1];
02373 
02374       if (NeedExtend) {
02375         StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0);
02376         StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1);
02377       }
02378 
02379       SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal0,
02380                         StoreVal1 };
02381       Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl,
02382                                       DAG.getVTList(MVT::Other), Ops,
02383                                       EltVT, MachinePointerInfo());
02384     } else {
02385       // V4 stores
02386       // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the
02387       // vector will be expanded to a power of 2 elements, so we know we can
02388       // always round up to the next multiple of 4 when creating the vector
02389       // stores.
02390       // e.g.  4 elem => 1 st.v4
02391       //       6 elem => 2 st.v4
02392       //       8 elem => 2 st.v4
02393       //      11 elem => 3 st.v4
02394 
02395       unsigned VecSize = 4;
02396       if (OutVals[0].getValueType().getSizeInBits() == 64)
02397         VecSize = 2;
02398 
02399       unsigned Offset = 0;
02400 
02401       EVT VecVT =
02402           EVT::getVectorVT(F->getContext(), EltVT, VecSize);
02403       unsigned PerStoreOffset =
02404           TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
02405 
02406       for (unsigned i = 0; i < NumElts; i += VecSize) {
02407         // Get values
02408         SDValue StoreVal;
02409         SmallVector<SDValue, 8> Ops;
02410         Ops.push_back(Chain);
02411         Ops.push_back(DAG.getConstant(Offset, MVT::i32));
02412         unsigned Opc = NVPTXISD::StoreRetvalV2;
02413         EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType();
02414 
02415         StoreVal = OutVals[i];
02416         if (NeedExtend)
02417           StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
02418         Ops.push_back(StoreVal);
02419 
02420         if (i + 1 < NumElts) {
02421           StoreVal = OutVals[i + 1];
02422           if (NeedExtend)
02423             StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
02424         } else {
02425           StoreVal = DAG.getUNDEF(ExtendedVT);
02426         }
02427         Ops.push_back(StoreVal);
02428 
02429         if (VecSize == 4) {
02430           Opc = NVPTXISD::StoreRetvalV4;
02431           if (i + 2 < NumElts) {
02432             StoreVal = OutVals[i + 2];
02433             if (NeedExtend)
02434               StoreVal =
02435                   DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
02436           } else {
02437             StoreVal = DAG.getUNDEF(ExtendedVT);
02438           }
02439           Ops.push_back(StoreVal);
02440 
02441           if (i + 3 < NumElts) {
02442             StoreVal = OutVals[i + 3];
02443             if (NeedExtend)
02444               StoreVal =
02445                   DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
02446           } else {
02447             StoreVal = DAG.getUNDEF(ExtendedVT);
02448           }
02449           Ops.push_back(StoreVal);
02450         }
02451 
02452         // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size());
02453         Chain =
02454             DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops,
02455                                     EltVT, MachinePointerInfo());
02456         Offset += PerStoreOffset;
02457       }
02458     }
02459   } else {
02460     SmallVector<EVT, 16> ValVTs;
02461     SmallVector<uint64_t, 16> Offsets;
02462     ComputePTXValueVTs(*this, RetTy, ValVTs, &Offsets, 0);
02463     assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition");
02464 
02465     for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
02466       SDValue theVal = OutVals[i];
02467       EVT TheValType = theVal.getValueType();
02468       unsigned numElems = 1;
02469       if (TheValType.isVector())
02470         numElems = TheValType.getVectorNumElements();
02471       for (unsigned j = 0, je = numElems; j != je; ++j) {
02472         SDValue TmpVal = theVal;
02473         if (TheValType.isVector())
02474           TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
02475                                TheValType.getVectorElementType(), TmpVal,
02476                                DAG.getIntPtrConstant(j));
02477         EVT TheStoreType = ValVTs[i];
02478         if (RetTy->isIntegerTy() &&
02479             TD->getTypeAllocSizeInBits(RetTy) < 32) {
02480           // The following zero-extension is for integer types only, and
02481           // specifically not for aggregates.
02482           TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal);
02483           TheStoreType = MVT::i32;
02484         }
02485         else if (TmpVal.getValueType().getSizeInBits() < 16)
02486           TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal);
02487 
02488         SDValue Ops[] = {
02489           Chain,
02490           DAG.getConstant(Offsets[i], MVT::i32),
02491           TmpVal };
02492         Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
02493                                         DAG.getVTList(MVT::Other), Ops,
02494                                         TheStoreType,
02495                                         MachinePointerInfo());
02496       }
02497     }
02498   }
02499 
02500   return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
02501 }
02502 
02503 
02504 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
02505     SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
02506     SelectionDAG &DAG) const {
02507   if (Constraint.length() > 1)
02508     return;
02509   else
02510     TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
02511 }
02512 
02513 // NVPTX suuport vector of legal types of any length in Intrinsics because the
02514 // NVPTX specific type legalizer
02515 // will legalize them to the PTX supported length.
02516 bool NVPTXTargetLowering::isTypeSupportedInIntrinsic(MVT VT) const {
02517   if (isTypeLegal(VT))
02518     return true;
02519   if (VT.isVector()) {
02520     MVT eVT = VT.getVectorElementType();
02521     if (isTypeLegal(eVT))
02522       return true;
02523   }
02524   return false;
02525 }
02526 
02527 static unsigned getOpcForTextureInstr(unsigned Intrinsic) {
02528   switch (Intrinsic) {
02529   default:
02530     return 0;
02531 
02532   case Intrinsic::nvvm_tex_1d_v4f32_s32:
02533     return NVPTXISD::Tex1DFloatS32;
02534   case Intrinsic::nvvm_tex_1d_v4f32_f32:
02535     return NVPTXISD::Tex1DFloatFloat;
02536   case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
02537     return NVPTXISD::Tex1DFloatFloatLevel;
02538   case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
02539     return NVPTXISD::Tex1DFloatFloatGrad;
02540   case Intrinsic::nvvm_tex_1d_v4s32_s32:
02541     return NVPTXISD::Tex1DS32S32;
02542   case Intrinsic::nvvm_tex_1d_v4s32_f32:
02543     return NVPTXISD::Tex1DS32Float;
02544   case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
02545     return NVPTXISD::Tex1DS32FloatLevel;
02546   case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
02547     return NVPTXISD::Tex1DS32FloatGrad;
02548   case Intrinsic::nvvm_tex_1d_v4u32_s32:
02549     return NVPTXISD::Tex1DU32S32;
02550   case Intrinsic::nvvm_tex_1d_v4u32_f32:
02551     return NVPTXISD::Tex1DU32Float;
02552   case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
02553     return NVPTXISD::Tex1DU32FloatLevel;
02554   case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
02555     return NVPTXISD::Tex1DU32FloatGrad;
02556 
02557   case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
02558     return NVPTXISD::Tex1DArrayFloatS32;
02559   case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
02560     return NVPTXISD::Tex1DArrayFloatFloat;
02561   case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
02562     return NVPTXISD::Tex1DArrayFloatFloatLevel;
02563   case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
02564     return NVPTXISD::Tex1DArrayFloatFloatGrad;
02565   case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
02566     return NVPTXISD::Tex1DArrayS32S32;
02567   case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
02568     return NVPTXISD::Tex1DArrayS32Float;
02569   case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
02570     return NVPTXISD::Tex1DArrayS32FloatLevel;
02571   case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
02572     return NVPTXISD::Tex1DArrayS32FloatGrad;
02573   case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
02574     return NVPTXISD::Tex1DArrayU32S32;
02575   case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
02576     return NVPTXISD::Tex1DArrayU32Float;
02577   case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
02578     return NVPTXISD::Tex1DArrayU32FloatLevel;
02579   case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
02580     return NVPTXISD::Tex1DArrayU32FloatGrad;
02581 
02582   case Intrinsic::nvvm_tex_2d_v4f32_s32:
02583     return NVPTXISD::Tex2DFloatS32;
02584   case Intrinsic::nvvm_tex_2d_v4f32_f32:
02585     return NVPTXISD::Tex2DFloatFloat;
02586   case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
02587     return NVPTXISD::Tex2DFloatFloatLevel;
02588   case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
02589     return NVPTXISD::Tex2DFloatFloatGrad;
02590   case Intrinsic::nvvm_tex_2d_v4s32_s32:
02591     return NVPTXISD::Tex2DS32S32;
02592   case Intrinsic::nvvm_tex_2d_v4s32_f32:
02593     return NVPTXISD::Tex2DS32Float;
02594   case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
02595     return NVPTXISD::Tex2DS32FloatLevel;
02596   case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
02597     return NVPTXISD::Tex2DS32FloatGrad;
02598   case Intrinsic::nvvm_tex_2d_v4u32_s32:
02599     return NVPTXISD::Tex2DU32S32;
02600   case Intrinsic::nvvm_tex_2d_v4u32_f32:
02601     return NVPTXISD::Tex2DU32Float;
02602   case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
02603     return NVPTXISD::Tex2DU32FloatLevel;
02604   case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
02605     return NVPTXISD::Tex2DU32FloatGrad;
02606 
02607   case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
02608     return NVPTXISD::Tex2DArrayFloatS32;
02609   case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
02610     return NVPTXISD::Tex2DArrayFloatFloat;
02611   case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
02612     return NVPTXISD::Tex2DArrayFloatFloatLevel;
02613   case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
02614     return NVPTXISD::Tex2DArrayFloatFloatGrad;
02615   case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
02616     return NVPTXISD::Tex2DArrayS32S32;
02617   case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
02618     return NVPTXISD::Tex2DArrayS32Float;
02619   case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
02620     return NVPTXISD::Tex2DArrayS32FloatLevel;
02621   case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
02622     return NVPTXISD::Tex2DArrayS32FloatGrad;
02623   case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
02624     return NVPTXISD::Tex2DArrayU32S32;
02625   case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
02626     return NVPTXISD::Tex2DArrayU32Float;
02627   case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
02628     return NVPTXISD::Tex2DArrayU32FloatLevel;
02629   case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
02630     return NVPTXISD::Tex2DArrayU32FloatGrad;
02631 
02632   case Intrinsic::nvvm_tex_3d_v4f32_s32:
02633     return NVPTXISD::Tex3DFloatS32;
02634   case Intrinsic::nvvm_tex_3d_v4f32_f32:
02635     return NVPTXISD::Tex3DFloatFloat;
02636   case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
02637     return NVPTXISD::Tex3DFloatFloatLevel;
02638   case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
02639     return NVPTXISD::Tex3DFloatFloatGrad;
02640   case Intrinsic::nvvm_tex_3d_v4s32_s32:
02641     return NVPTXISD::Tex3DS32S32;
02642   case Intrinsic::nvvm_tex_3d_v4s32_f32:
02643     return NVPTXISD::Tex3DS32Float;
02644   case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
02645     return NVPTXISD::Tex3DS32FloatLevel;
02646   case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
02647     return NVPTXISD::Tex3DS32FloatGrad;
02648   case Intrinsic::nvvm_tex_3d_v4u32_s32:
02649     return NVPTXISD::Tex3DU32S32;
02650   case Intrinsic::nvvm_tex_3d_v4u32_f32:
02651     return NVPTXISD::Tex3DU32Float;
02652   case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
02653     return NVPTXISD::Tex3DU32FloatLevel;
02654   case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
02655     return NVPTXISD::Tex3DU32FloatGrad;
02656 
02657   case Intrinsic::nvvm_tex_cube_v4f32_f32:
02658     return NVPTXISD::TexCubeFloatFloat;
02659   case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
02660     return NVPTXISD::TexCubeFloatFloatLevel;
02661   case Intrinsic::nvvm_tex_cube_v4s32_f32:
02662     return NVPTXISD::TexCubeS32Float;
02663   case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
02664     return NVPTXISD::TexCubeS32FloatLevel;
02665   case Intrinsic::nvvm_tex_cube_v4u32_f32:
02666     return NVPTXISD::TexCubeU32Float;
02667   case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
02668     return NVPTXISD::TexCubeU32FloatLevel;
02669 
02670   case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
02671     return NVPTXISD::TexCubeArrayFloatFloat;
02672   case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
02673     return NVPTXISD::TexCubeArrayFloatFloatLevel;
02674   case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
02675     return NVPTXISD::TexCubeArrayS32Float;
02676   case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
02677     return NVPTXISD::TexCubeArrayS32FloatLevel;
02678   case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
02679     return NVPTXISD::TexCubeArrayU32Float;
02680   case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
02681     return NVPTXISD::TexCubeArrayU32FloatLevel;
02682 
02683   case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
02684     return NVPTXISD::Tld4R2DFloatFloat;
02685   case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
02686     return NVPTXISD::Tld4G2DFloatFloat;
02687   case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
02688     return NVPTXISD::Tld4B2DFloatFloat;
02689   case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
02690     return NVPTXISD::Tld4A2DFloatFloat;
02691   case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
02692     return NVPTXISD::Tld4R2DS64Float;
02693   case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
02694     return NVPTXISD::Tld4G2DS64Float;
02695   case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
02696     return NVPTXISD::Tld4B2DS64Float;
02697   case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
02698     return NVPTXISD::Tld4A2DS64Float;
02699   case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
02700     return NVPTXISD::Tld4R2DU64Float;
02701   case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
02702     return NVPTXISD::Tld4G2DU64Float;
02703   case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
02704     return NVPTXISD::Tld4B2DU64Float;
02705   case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
02706     return NVPTXISD::Tld4A2DU64Float;
02707 
02708   case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
02709     return NVPTXISD::TexUnified1DFloatS32;
02710   case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
02711     return NVPTXISD::TexUnified1DFloatFloat;
02712   case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
02713     return NVPTXISD::TexUnified1DFloatFloatLevel;
02714   case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
02715     return NVPTXISD::TexUnified1DFloatFloatGrad;
02716   case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
02717     return NVPTXISD::TexUnified1DS32S32;
02718   case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
02719     return NVPTXISD::TexUnified1DS32Float;
02720   case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
02721     return NVPTXISD::TexUnified1DS32FloatLevel;
02722   case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
02723     return NVPTXISD::TexUnified1DS32FloatGrad;
02724   case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
02725     return NVPTXISD::TexUnified1DU32S32;
02726   case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
02727     return NVPTXISD::TexUnified1DU32Float;
02728   case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
02729     return NVPTXISD::TexUnified1DU32FloatLevel;
02730   case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
02731     return NVPTXISD::TexUnified1DU32FloatGrad;
02732 
02733   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
02734     return NVPTXISD::TexUnified1DArrayFloatS32;
02735   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
02736     return NVPTXISD::TexUnified1DArrayFloatFloat;
02737   case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
02738     return NVPTXISD::TexUnified1DArrayFloatFloatLevel;
02739   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
02740     return NVPTXISD::TexUnified1DArrayFloatFloatGrad;
02741   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
02742     return NVPTXISD::TexUnified1DArrayS32S32;
02743   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
02744     return NVPTXISD::TexUnified1DArrayS32Float;
02745   case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
02746     return NVPTXISD::TexUnified1DArrayS32FloatLevel;
02747   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
02748     return NVPTXISD::TexUnified1DArrayS32FloatGrad;
02749   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
02750     return NVPTXISD::TexUnified1DArrayU32S32;
02751   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
02752     return NVPTXISD::TexUnified1DArrayU32Float;
02753   case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
02754     return NVPTXISD::TexUnified1DArrayU32FloatLevel;
02755   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
02756     return NVPTXISD::TexUnified1DArrayU32FloatGrad;
02757 
02758   case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
02759     return NVPTXISD::TexUnified2DFloatS32;
02760   case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
02761     return NVPTXISD::TexUnified2DFloatFloat;
02762   case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
02763     return NVPTXISD::TexUnified2DFloatFloatLevel;
02764   case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
02765     return NVPTXISD::TexUnified2DFloatFloatGrad;
02766   case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
02767     return NVPTXISD::TexUnified2DS32S32;
02768   case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
02769     return NVPTXISD::TexUnified2DS32Float;
02770   case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
02771     return NVPTXISD::TexUnified2DS32FloatLevel;
02772   case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
02773     return NVPTXISD::TexUnified2DS32FloatGrad;
02774   case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
02775     return NVPTXISD::TexUnified2DU32S32;
02776   case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
02777     return NVPTXISD::TexUnified2DU32Float;
02778   case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
02779     return NVPTXISD::TexUnified2DU32FloatLevel;
02780   case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
02781     return NVPTXISD::TexUnified2DU32FloatGrad;
02782 
02783   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
02784     return NVPTXISD::TexUnified2DArrayFloatS32;
02785   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
02786     return NVPTXISD::TexUnified2DArrayFloatFloat;
02787   case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
02788     return NVPTXISD::TexUnified2DArrayFloatFloatLevel;
02789   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
02790     return NVPTXISD::TexUnified2DArrayFloatFloatGrad;
02791   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
02792     return NVPTXISD::TexUnified2DArrayS32S32;
02793   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
02794     return NVPTXISD::TexUnified2DArrayS32Float;
02795   case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
02796     return NVPTXISD::TexUnified2DArrayS32FloatLevel;
02797   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
02798     return NVPTXISD::TexUnified2DArrayS32FloatGrad;
02799   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
02800     return NVPTXISD::TexUnified2DArrayU32S32;
02801   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
02802     return NVPTXISD::TexUnified2DArrayU32Float;
02803   case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
02804     return NVPTXISD::TexUnified2DArrayU32FloatLevel;
02805   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
02806     return NVPTXISD::TexUnified2DArrayU32FloatGrad;
02807 
02808   case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
02809     return NVPTXISD::TexUnified3DFloatS32;
02810   case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
02811     return NVPTXISD::TexUnified3DFloatFloat;
02812   case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
02813     return NVPTXISD::TexUnified3DFloatFloatLevel;
02814   case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
02815     return NVPTXISD::TexUnified3DFloatFloatGrad;
02816   case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
02817     return NVPTXISD::TexUnified3DS32S32;
02818   case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
02819     return NVPTXISD::TexUnified3DS32Float;
02820   case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
02821     return NVPTXISD::TexUnified3DS32FloatLevel;
02822   case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
02823     return NVPTXISD::TexUnified3DS32FloatGrad;
02824   case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
02825     return NVPTXISD::TexUnified3DU32S32;
02826   case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
02827     return NVPTXISD::TexUnified3DU32Float;
02828   case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
02829     return NVPTXISD::TexUnified3DU32FloatLevel;
02830   case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
02831     return NVPTXISD::TexUnified3DU32FloatGrad;
02832 
02833   case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
02834     return NVPTXISD::TexUnifiedCubeFloatFloat;
02835   case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
02836     return NVPTXISD::TexUnifiedCubeFloatFloatLevel;
02837   case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
02838     return NVPTXISD::TexUnifiedCubeS32Float;
02839   case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
02840     return NVPTXISD::TexUnifiedCubeS32FloatLevel;
02841   case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
02842     return NVPTXISD::TexUnifiedCubeU32Float;
02843   case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
02844     return NVPTXISD::TexUnifiedCubeU32FloatLevel;
02845 
02846   case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
02847     return NVPTXISD::TexUnifiedCubeArrayFloatFloat;
02848   case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
02849     return NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel;
02850   case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
02851     return NVPTXISD::TexUnifiedCubeArrayS32Float;
02852   case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
02853     return NVPTXISD::TexUnifiedCubeArrayS32FloatLevel;
02854   case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
02855     return NVPTXISD::TexUnifiedCubeArrayU32Float;
02856   case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
02857     return NVPTXISD::TexUnifiedCubeArrayU32FloatLevel;
02858 
02859   case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
02860     return NVPTXISD::Tld4UnifiedR2DFloatFloat;
02861   case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
02862     return NVPTXISD::Tld4UnifiedG2DFloatFloat;
02863   case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
02864     return NVPTXISD::Tld4UnifiedB2DFloatFloat;
02865   case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
02866     return NVPTXISD::Tld4UnifiedA2DFloatFloat;
02867   case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
02868     return NVPTXISD::Tld4UnifiedR2DS64Float;
02869   case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
02870     return NVPTXISD::Tld4UnifiedG2DS64Float;
02871   case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
02872     return NVPTXISD::Tld4UnifiedB2DS64Float;
02873   case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
02874     return NVPTXISD::Tld4UnifiedA2DS64Float;
02875   case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
02876     return NVPTXISD::Tld4UnifiedR2DU64Float;
02877   case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
02878     return NVPTXISD::Tld4UnifiedG2DU64Float;
02879   case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
02880     return NVPTXISD::Tld4UnifiedB2DU64Float;
02881   case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
02882     return NVPTXISD::Tld4UnifiedA2DU64Float;
02883   }
02884 }
02885 
02886 static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) {
02887   switch (Intrinsic) {
02888   default:
02889     return 0;
02890   case Intrinsic::nvvm_suld_1d_i8_clamp:
02891     return NVPTXISD::Suld1DI8Clamp;
02892   case Intrinsic::nvvm_suld_1d_i16_clamp:
02893     return NVPTXISD::Suld1DI16Clamp;
02894   case Intrinsic::nvvm_suld_1d_i32_clamp:
02895     return NVPTXISD::Suld1DI32Clamp;
02896   case Intrinsic::nvvm_suld_1d_i64_clamp:
02897     return NVPTXISD::Suld1DI64Clamp;
02898   case Intrinsic::nvvm_suld_1d_v2i8_clamp:
02899     return NVPTXISD::Suld1DV2I8Clamp;
02900   case Intrinsic::nvvm_suld_1d_v2i16_clamp:
02901     return NVPTXISD::Suld1DV2I16Clamp;
02902   case Intrinsic::nvvm_suld_1d_v2i32_clamp:
02903     return NVPTXISD::Suld1DV2I32Clamp;
02904   case Intrinsic::nvvm_suld_1d_v2i64_clamp:
02905     return NVPTXISD::Suld1DV2I64Clamp;
02906   case Intrinsic::nvvm_suld_1d_v4i8_clamp:
02907     return NVPTXISD::Suld1DV4I8Clamp;
02908   case Intrinsic::nvvm_suld_1d_v4i16_clamp:
02909     return NVPTXISD::Suld1DV4I16Clamp;
02910   case Intrinsic::nvvm_suld_1d_v4i32_clamp:
02911     return NVPTXISD::Suld1DV4I32Clamp;
02912   case Intrinsic::nvvm_suld_1d_array_i8_clamp:
02913     return NVPTXISD::Suld1DArrayI8Clamp;
02914   case Intrinsic::nvvm_suld_1d_array_i16_clamp:
02915     return NVPTXISD::Suld1DArrayI16Clamp;
02916   case Intrinsic::nvvm_suld_1d_array_i32_clamp:
02917     return NVPTXISD::Suld1DArrayI32Clamp;
02918   case Intrinsic::nvvm_suld_1d_array_i64_clamp:
02919     return NVPTXISD::Suld1DArrayI64Clamp;
02920   case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
02921     return NVPTXISD::Suld1DArrayV2I8Clamp;
02922   case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
02923     return NVPTXISD::Suld1DArrayV2I16Clamp;
02924   case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
02925     return NVPTXISD::Suld1DArrayV2I32Clamp;
02926   case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
02927     return NVPTXISD::Suld1DArrayV2I64Clamp;
02928   case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
02929     return NVPTXISD::Suld1DArrayV4I8Clamp;
02930   case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
02931     return NVPTXISD::Suld1DArrayV4I16Clamp;
02932   case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
02933     return NVPTXISD::Suld1DArrayV4I32Clamp;
02934   case Intrinsic::nvvm_suld_2d_i8_clamp:
02935     return NVPTXISD::Suld2DI8Clamp;
02936   case Intrinsic::nvvm_suld_2d_i16_clamp:
02937     return NVPTXISD::Suld2DI16Clamp;
02938   case Intrinsic::nvvm_suld_2d_i32_clamp:
02939     return NVPTXISD::Suld2DI32Clamp;
02940   case Intrinsic::nvvm_suld_2d_i64_clamp:
02941     return NVPTXISD::Suld2DI64Clamp;
02942   case Intrinsic::nvvm_suld_2d_v2i8_clamp:
02943     return NVPTXISD::Suld2DV2I8Clamp;
02944   case Intrinsic::nvvm_suld_2d_v2i16_clamp:
02945     return NVPTXISD::Suld2DV2I16Clamp;
02946   case Intrinsic::nvvm_suld_2d_v2i32_clamp:
02947     return NVPTXISD::Suld2DV2I32Clamp;
02948   case Intrinsic::nvvm_suld_2d_v2i64_clamp:
02949     return NVPTXISD::Suld2DV2I64Clamp;
02950   case Intrinsic::nvvm_suld_2d_v4i8_clamp:
02951     return NVPTXISD::Suld2DV4I8Clamp;
02952   case Intrinsic::nvvm_suld_2d_v4i16_clamp:
02953     return NVPTXISD::Suld2DV4I16Clamp;
02954   case Intrinsic::nvvm_suld_2d_v4i32_clamp:
02955     return NVPTXISD::Suld2DV4I32Clamp;
02956   case Intrinsic::nvvm_suld_2d_array_i8_clamp:
02957     return NVPTXISD::Suld2DArrayI8Clamp;
02958   case Intrinsic::nvvm_suld_2d_array_i16_clamp:
02959     return NVPTXISD::Suld2DArrayI16Clamp;
02960   case Intrinsic::nvvm_suld_2d_array_i32_clamp:
02961     return NVPTXISD::Suld2DArrayI32Clamp;
02962   case Intrinsic::nvvm_suld_2d_array_i64_clamp:
02963     return NVPTXISD::Suld2DArrayI64Clamp;
02964   case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
02965     return NVPTXISD::Suld2DArrayV2I8Clamp;
02966   case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
02967     return NVPTXISD::Suld2DArrayV2I16Clamp;
02968   case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
02969     return NVPTXISD::Suld2DArrayV2I32Clamp;
02970   case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
02971     return NVPTXISD::Suld2DArrayV2I64Clamp;
02972   case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
02973     return NVPTXISD::Suld2DArrayV4I8Clamp;
02974   case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
02975     return NVPTXISD::Suld2DArrayV4I16Clamp;
02976   case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
02977     return NVPTXISD::Suld2DArrayV4I32Clamp;
02978   case Intrinsic::nvvm_suld_3d_i8_clamp:
02979     return NVPTXISD::Suld3DI8Clamp;
02980   case Intrinsic::nvvm_suld_3d_i16_clamp:
02981     return NVPTXISD::Suld3DI16Clamp;
02982   case Intrinsic::nvvm_suld_3d_i32_clamp:
02983     return NVPTXISD::Suld3DI32Clamp;
02984   case Intrinsic::nvvm_suld_3d_i64_clamp:
02985     return NVPTXISD::Suld3DI64Clamp;
02986   case Intrinsic::nvvm_suld_3d_v2i8_clamp:
02987     return NVPTXISD::Suld3DV2I8Clamp;
02988   case Intrinsic::nvvm_suld_3d_v2i16_clamp:
02989     return NVPTXISD::Suld3DV2I16Clamp;
02990   case Intrinsic::nvvm_suld_3d_v2i32_clamp:
02991     return NVPTXISD::Suld3DV2I32Clamp;
02992   case Intrinsic::nvvm_suld_3d_v2i64_clamp:
02993     return NVPTXISD::Suld3DV2I64Clamp;
02994   case Intrinsic::nvvm_suld_3d_v4i8_clamp:
02995     return NVPTXISD::Suld3DV4I8Clamp;
02996   case Intrinsic::nvvm_suld_3d_v4i16_clamp:
02997     return NVPTXISD::Suld3DV4I16Clamp;
02998   case Intrinsic::nvvm_suld_3d_v4i32_clamp:
02999     return NVPTXISD::Suld3DV4I32Clamp;
03000   case Intrinsic::nvvm_suld_1d_i8_trap:
03001     return NVPTXISD::Suld1DI8Trap;
03002   case Intrinsic::nvvm_suld_1d_i16_trap:
03003     return NVPTXISD::Suld1DI16Trap;
03004   case Intrinsic::nvvm_suld_1d_i32_trap:
03005     return NVPTXISD::Suld1DI32Trap;
03006   case Intrinsic::nvvm_suld_1d_i64_trap:
03007     return NVPTXISD::Suld1DI64Trap;
03008   case Intrinsic::nvvm_suld_1d_v2i8_trap:
03009     return NVPTXISD::Suld1DV2I8Trap;
03010   case Intrinsic::nvvm_suld_1d_v2i16_trap:
03011     return NVPTXISD::Suld1DV2I16Trap;
03012   case Intrinsic::nvvm_suld_1d_v2i32_trap:
03013     return NVPTXISD::Suld1DV2I32Trap;
03014   case Intrinsic::nvvm_suld_1d_v2i64_trap:
03015     return NVPTXISD::Suld1DV2I64Trap;
03016   case Intrinsic::nvvm_suld_1d_v4i8_trap:
03017     return NVPTXISD::Suld1DV4I8Trap;
03018   case Intrinsic::nvvm_suld_1d_v4i16_trap:
03019     return NVPTXISD::Suld1DV4I16Trap;
03020   case Intrinsic::nvvm_suld_1d_v4i32_trap:
03021     return NVPTXISD::Suld1DV4I32Trap;
03022   case Intrinsic::nvvm_suld_1d_array_i8_trap:
03023     return NVPTXISD::Suld1DArrayI8Trap;
03024   case Intrinsic::nvvm_suld_1d_array_i16_trap:
03025     return NVPTXISD::Suld1DArrayI16Trap;
03026   case Intrinsic::nvvm_suld_1d_array_i32_trap:
03027     return NVPTXISD::Suld1DArrayI32Trap;
03028   case Intrinsic::nvvm_suld_1d_array_i64_trap:
03029     return NVPTXISD::Suld1DArrayI64Trap;
03030   case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
03031     return NVPTXISD::Suld1DArrayV2I8Trap;
03032   case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
03033     return NVPTXISD::Suld1DArrayV2I16Trap;
03034   case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
03035     return NVPTXISD::Suld1DArrayV2I32Trap;
03036   case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
03037     return NVPTXISD::Suld1DArrayV2I64Trap;
03038   case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
03039     return NVPTXISD::Suld1DArrayV4I8Trap;
03040   case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
03041     return NVPTXISD::Suld1DArrayV4I16Trap;
03042   case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
03043     return NVPTXISD::Suld1DArrayV4I32Trap;
03044   case Intrinsic::nvvm_suld_2d_i8_trap:
03045     return NVPTXISD::Suld2DI8Trap;
03046   case Intrinsic::nvvm_suld_2d_i16_trap:
03047     return NVPTXISD::Suld2DI16Trap;
03048   case Intrinsic::nvvm_suld_2d_i32_trap:
03049     return NVPTXISD::Suld2DI32Trap;
03050   case Intrinsic::nvvm_suld_2d_i64_trap:
03051     return NVPTXISD::Suld2DI64Trap;
03052   case Intrinsic::nvvm_suld_2d_v2i8_trap:
03053     return NVPTXISD::Suld2DV2I8Trap;
03054   case Intrinsic::nvvm_suld_2d_v2i16_trap:
03055     return NVPTXISD::Suld2DV2I16Trap;
03056   case Intrinsic::nvvm_suld_2d_v2i32_trap:
03057     return NVPTXISD::Suld2DV2I32Trap;
03058   case Intrinsic::nvvm_suld_2d_v2i64_trap:
03059     return NVPTXISD::Suld2DV2I64Trap;
03060   case Intrinsic::nvvm_suld_2d_v4i8_trap:
03061     return NVPTXISD::Suld2DV4I8Trap;
03062   case Intrinsic::nvvm_suld_2d_v4i16_trap:
03063     return NVPTXISD::Suld2DV4I16Trap;
03064   case Intrinsic::nvvm_suld_2d_v4i32_trap:
03065     return NVPTXISD::Suld2DV4I32Trap;
03066   case Intrinsic::nvvm_suld_2d_array_i8_trap:
03067     return NVPTXISD::Suld2DArrayI8Trap;
03068   case Intrinsic::nvvm_suld_2d_array_i16_trap:
03069     return NVPTXISD::Suld2DArrayI16Trap;
03070   case Intrinsic::nvvm_suld_2d_array_i32_trap:
03071     return NVPTXISD::Suld2DArrayI32Trap;
03072   case Intrinsic::nvvm_suld_2d_array_i64_trap:
03073     return NVPTXISD::Suld2DArrayI64Trap;
03074   case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
03075     return NVPTXISD::Suld2DArrayV2I8Trap;
03076   case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
03077     return NVPTXISD::Suld2DArrayV2I16Trap;
03078   case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
03079     return NVPTXISD::Suld2DArrayV2I32Trap;
03080   case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
03081     return NVPTXISD::Suld2DArrayV2I64Trap;
03082   case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
03083     return NVPTXISD::Suld2DArrayV4I8Trap;
03084   case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
03085     return NVPTXISD::Suld2DArrayV4I16Trap;
03086   case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
03087     return NVPTXISD::Suld2DArrayV4I32Trap;
03088   case Intrinsic::nvvm_suld_3d_i8_trap:
03089     return NVPTXISD::Suld3DI8Trap;
03090   case Intrinsic::nvvm_suld_3d_i16_trap:
03091     return NVPTXISD::Suld3DI16Trap;
03092   case Intrinsic::nvvm_suld_3d_i32_trap:
03093     return NVPTXISD::Suld3DI32Trap;
03094   case Intrinsic::nvvm_suld_3d_i64_trap:
03095     return NVPTXISD::Suld3DI64Trap;
03096   case Intrinsic::nvvm_suld_3d_v2i8_trap:
03097     return NVPTXISD::Suld3DV2I8Trap;
03098   case Intrinsic::nvvm_suld_3d_v2i16_trap:
03099     return NVPTXISD::Suld3DV2I16Trap;
03100   case Intrinsic::nvvm_suld_3d_v2i32_trap:
03101     return NVPTXISD::Suld3DV2I32Trap;
03102   case Intrinsic::nvvm_suld_3d_v2i64_trap:
03103     return NVPTXISD::Suld3DV2I64Trap;
03104   case Intrinsic::nvvm_suld_3d_v4i8_trap:
03105     return NVPTXISD::Suld3DV4I8Trap;
03106   case Intrinsic::nvvm_suld_3d_v4i16_trap:
03107     return NVPTXISD::Suld3DV4I16Trap;
03108   case Intrinsic::nvvm_suld_3d_v4i32_trap:
03109     return NVPTXISD::Suld3DV4I32Trap;
03110   case Intrinsic::nvvm_suld_1d_i8_zero:
03111     return NVPTXISD::Suld1DI8Zero;
03112   case Intrinsic::nvvm_suld_1d_i16_zero:
03113     return NVPTXISD::Suld1DI16Zero;
03114   case Intrinsic::nvvm_suld_1d_i32_zero:
03115     return NVPTXISD::Suld1DI32Zero;
03116   case Intrinsic::nvvm_suld_1d_i64_zero:
03117     return NVPTXISD::Suld1DI64Zero;
03118   case Intrinsic::nvvm_suld_1d_v2i8_zero:
03119     return NVPTXISD::Suld1DV2I8Zero;
03120   case Intrinsic::nvvm_suld_1d_v2i16_zero:
03121     return NVPTXISD::Suld1DV2I16Zero;
03122   case Intrinsic::nvvm_suld_1d_v2i32_zero:
03123     return NVPTXISD::Suld1DV2I32Zero;
03124   case Intrinsic::nvvm_suld_1d_v2i64_zero:
03125     return NVPTXISD::Suld1DV2I64Zero;
03126   case Intrinsic::nvvm_suld_1d_v4i8_zero:
03127     return NVPTXISD::Suld1DV4I8Zero;
03128   case Intrinsic::nvvm_suld_1d_v4i16_zero:
03129     return NVPTXISD::Suld1DV4I16Zero;
03130   case Intrinsic::nvvm_suld_1d_v4i32_zero:
03131     return NVPTXISD::Suld1DV4I32Zero;
03132   case Intrinsic::nvvm_suld_1d_array_i8_zero:
03133     return NVPTXISD::Suld1DArrayI8Zero;
03134   case Intrinsic::nvvm_suld_1d_array_i16_zero:
03135     return NVPTXISD::Suld1DArrayI16Zero;
03136   case Intrinsic::nvvm_suld_1d_array_i32_zero:
03137     return NVPTXISD::Suld1DArrayI32Zero;
03138   case Intrinsic::nvvm_suld_1d_array_i64_zero:
03139     return NVPTXISD::Suld1DArrayI64Zero;
03140   case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
03141     return NVPTXISD::Suld1DArrayV2I8Zero;
03142   case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
03143     return NVPTXISD::Suld1DArrayV2I16Zero;
03144   case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
03145     return NVPTXISD::Suld1DArrayV2I32Zero;
03146   case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
03147     return NVPTXISD::Suld1DArrayV2I64Zero;
03148   case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
03149     return NVPTXISD::Suld1DArrayV4I8Zero;
03150   case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
03151     return NVPTXISD::Suld1DArrayV4I16Zero;
03152   case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
03153     return NVPTXISD::Suld1DArrayV4I32Zero;
03154   case Intrinsic::nvvm_suld_2d_i8_zero:
03155     return NVPTXISD::Suld2DI8Zero;
03156   case Intrinsic::nvvm_suld_2d_i16_zero:
03157     return NVPTXISD::Suld2DI16Zero;
03158   case Intrinsic::nvvm_suld_2d_i32_zero:
03159     return NVPTXISD::Suld2DI32Zero;
03160   case Intrinsic::nvvm_suld_2d_i64_zero:
03161     return NVPTXISD::Suld2DI64Zero;
03162   case Intrinsic::nvvm_suld_2d_v2i8_zero:
03163     return NVPTXISD::Suld2DV2I8Zero;
03164   case Intrinsic::nvvm_suld_2d_v2i16_zero:
03165     return NVPTXISD::Suld2DV2I16Zero;
03166   case Intrinsic::nvvm_suld_2d_v2i32_zero:
03167     return NVPTXISD::Suld2DV2I32Zero;
03168   case Intrinsic::nvvm_suld_2d_v2i64_zero:
03169     return NVPTXISD::Suld2DV2I64Zero;
03170   case Intrinsic::nvvm_suld_2d_v4i8_zero:
03171     return NVPTXISD::Suld2DV4I8Zero;
03172   case Intrinsic::nvvm_suld_2d_v4i16_zero:
03173     return NVPTXISD::Suld2DV4I16Zero;
03174   case Intrinsic::nvvm_suld_2d_v4i32_zero:
03175     return NVPTXISD::Suld2DV4I32Zero;
03176   case Intrinsic::nvvm_suld_2d_array_i8_zero:
03177     return NVPTXISD::Suld2DArrayI8Zero;
03178   case Intrinsic::nvvm_suld_2d_array_i16_zero:
03179     return NVPTXISD::Suld2DArrayI16Zero;
03180   case Intrinsic::nvvm_suld_2d_array_i32_zero:
03181     return NVPTXISD::Suld2DArrayI32Zero;
03182   case Intrinsic::nvvm_suld_2d_array_i64_zero:
03183     return NVPTXISD::Suld2DArrayI64Zero;
03184   case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
03185     return NVPTXISD::Suld2DArrayV2I8Zero;
03186   case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
03187     return NVPTXISD::Suld2DArrayV2I16Zero;
03188   case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
03189     return NVPTXISD::Suld2DArrayV2I32Zero;
03190   case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
03191     return NVPTXISD::Suld2DArrayV2I64Zero;
03192   case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
03193     return NVPTXISD::Suld2DArrayV4I8Zero;
03194   case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
03195     return NVPTXISD::Suld2DArrayV4I16Zero;
03196   case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
03197     return NVPTXISD::Suld2DArrayV4I32Zero;
03198   case Intrinsic::nvvm_suld_3d_i8_zero:
03199     return NVPTXISD::Suld3DI8Zero;
03200   case Intrinsic::nvvm_suld_3d_i16_zero:
03201     return NVPTXISD::Suld3DI16Zero;
03202   case Intrinsic::nvvm_suld_3d_i32_zero:
03203     return NVPTXISD::Suld3DI32Zero;
03204   case Intrinsic::nvvm_suld_3d_i64_zero:
03205     return NVPTXISD::Suld3DI64Zero;
03206   case Intrinsic::nvvm_suld_3d_v2i8_zero:
03207     return NVPTXISD::Suld3DV2I8Zero;
03208   case Intrinsic::nvvm_suld_3d_v2i16_zero:
03209     return NVPTXISD::Suld3DV2I16Zero;
03210   case Intrinsic::nvvm_suld_3d_v2i32_zero:
03211     return NVPTXISD::Suld3DV2I32Zero;
03212   case Intrinsic::nvvm_suld_3d_v2i64_zero:
03213     return NVPTXISD::Suld3DV2I64Zero;
03214   case Intrinsic::nvvm_suld_3d_v4i8_zero:
03215     return NVPTXISD::Suld3DV4I8Zero;
03216   case Intrinsic::nvvm_suld_3d_v4i16_zero:
03217     return NVPTXISD::Suld3DV4I16Zero;
03218   case Intrinsic::nvvm_suld_3d_v4i32_zero:
03219     return NVPTXISD::Suld3DV4I32Zero;
03220   }
03221 }
03222 
03223 // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as
03224 // TgtMemIntrinsic
03225 // because we need the information that is only available in the "Value" type
03226 // of destination
03227 // pointer. In particular, the address space information.
03228 bool NVPTXTargetLowering::getTgtMemIntrinsic(
03229     IntrinsicInfo &Info, const CallInst &I, unsigned Intrinsic) const {
03230   switch (Intrinsic) {
03231   default:
03232     return false;
03233 
03234   case Intrinsic::nvvm_atomic_load_add_f32:
03235     Info.opc = ISD::INTRINSIC_W_CHAIN;
03236     Info.memVT = MVT::f32;
03237     Info.ptrVal = I.getArgOperand(0);
03238     Info.offset = 0;
03239     Info.vol = 0;
03240     Info.readMem = true;
03241     Info.writeMem = true;
03242     Info.align = 0;
03243     return true;
03244 
03245   case Intrinsic::nvvm_atomic_load_inc_32:
03246   case Intrinsic::nvvm_atomic_load_dec_32:
03247     Info.opc = ISD::INTRINSIC_W_CHAIN;
03248     Info.memVT = MVT::i32;
03249     Info.ptrVal = I.getArgOperand(0);
03250     Info.offset = 0;
03251     Info.vol = 0;
03252     Info.readMem = true;
03253     Info.writeMem = true;
03254     Info.align = 0;
03255     return true;
03256 
03257   case Intrinsic::nvvm_ldu_global_i:
03258   case Intrinsic::nvvm_ldu_global_f:
03259   case Intrinsic::nvvm_ldu_global_p: {
03260 
03261     Info.opc = ISD::INTRINSIC_W_CHAIN;
03262     if (Intrinsic == Intrinsic::nvvm_ldu_global_i)
03263       Info.memVT = getValueType(I.getType());
03264     else if(Intrinsic == Intrinsic::nvvm_ldu_global_p)
03265       Info.memVT = getPointerTy();
03266     else
03267       Info.memVT = getValueType(I.getType());
03268     Info.ptrVal = I.getArgOperand(0);
03269     Info.offset = 0;
03270     Info.vol = 0;
03271     Info.readMem = true;
03272     Info.writeMem = false;
03273 
03274     // alignment is available as metadata.
03275     // Grab it and set the alignment.
03276     assert(I.hasMetadataOtherThanDebugLoc() && "Must have alignment metadata");
03277     MDNode *AlignMD = I.getMetadata("align");
03278     assert(AlignMD && "Must have a non-null MDNode");
03279     assert(AlignMD->getNumOperands() == 1 && "Must have a single operand");
03280     Value *Align = AlignMD->getOperand(0);
03281     int64_t Alignment = cast<ConstantInt>(Align)->getZExtValue();
03282     Info.align = Alignment;
03283 
03284     return true;
03285   }
03286   case Intrinsic::nvvm_ldg_global_i:
03287   case Intrinsic::nvvm_ldg_global_f:
03288   case Intrinsic::nvvm_ldg_global_p: {
03289 
03290     Info.opc = ISD::INTRINSIC_W_CHAIN;
03291     if (Intrinsic == Intrinsic::nvvm_ldg_global_i)
03292       Info.memVT = getValueType(I.getType());
03293     else if(Intrinsic == Intrinsic::nvvm_ldg_global_p)
03294       Info.memVT = getPointerTy();
03295     else
03296       Info.memVT = getValueType(I.getType());
03297     Info.ptrVal = I.getArgOperand(0);
03298     Info.offset = 0;
03299     Info.vol = 0;
03300     Info.readMem = true;
03301     Info.writeMem = false;
03302 
03303     // alignment is available as metadata.
03304     // Grab it and set the alignment.
03305     assert(I.hasMetadataOtherThanDebugLoc() && "Must have alignment metadata");
03306     MDNode *AlignMD = I.getMetadata("align");
03307     assert(AlignMD && "Must have a non-null MDNode");
03308     assert(AlignMD->getNumOperands() == 1 && "Must have a single operand");
03309     Value *Align = AlignMD->getOperand(0);
03310     int64_t Alignment = cast<ConstantInt>(Align)->getZExtValue();
03311     Info.align = Alignment;
03312 
03313     return true;
03314   }
03315 
03316   case Intrinsic::nvvm_tex_1d_v4f32_s32:
03317   case Intrinsic::nvvm_tex_1d_v4f32_f32:
03318   case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
03319   case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
03320   case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
03321   case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
03322   case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
03323   case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
03324   case Intrinsic::nvvm_tex_2d_v4f32_s32:
03325   case Intrinsic::nvvm_tex_2d_v4f32_f32:
03326   case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
03327   case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
03328   case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
03329   case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
03330   case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
03331   case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
03332   case Intrinsic::nvvm_tex_3d_v4f32_s32:
03333   case Intrinsic::nvvm_tex_3d_v4f32_f32:
03334   case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
03335   case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
03336   case Intrinsic::nvvm_tex_cube_v4f32_f32:
03337   case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
03338   case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
03339   case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
03340   case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
03341   case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
03342   case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
03343   case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
03344   case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
03345   case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
03346   case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
03347   case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
03348   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
03349   case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
03350   case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
03351   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
03352   case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
03353   case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
03354   case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
03355   case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
03356   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
03357   case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
03358   case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
03359   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
03360   case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
03361   case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
03362   case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
03363   case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
03364   case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
03365   case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
03366   case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
03367   case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
03368   case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
03369   case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
03370   case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
03371   case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32: {
03372     Info.opc = getOpcForTextureInstr(Intrinsic);
03373     Info.memVT = MVT::v4f32;
03374     Info.ptrVal = nullptr;
03375     Info.offset = 0;
03376     Info.vol = 0;
03377     Info.readMem = true;
03378     Info.writeMem = false;
03379     Info.align = 16;
03380     return true;
03381   }
03382   case Intrinsic::nvvm_tex_1d_v4s32_s32:
03383   case Intrinsic::nvvm_tex_1d_v4s32_f32:
03384   case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
03385   case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
03386   case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
03387   case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
03388   case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
03389   case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
03390   case Intrinsic::nvvm_tex_2d_v4s32_s32:
03391   case Intrinsic::nvvm_tex_2d_v4s32_f32:
03392   case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
03393   case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
03394   case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
03395   case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
03396   case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
03397   case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
03398   case Intrinsic::nvvm_tex_3d_v4s32_s32:
03399   case Intrinsic::nvvm_tex_3d_v4s32_f32:
03400   case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
03401   case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
03402   case Intrinsic::nvvm_tex_cube_v4s32_f32:
03403   case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
03404   case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
03405   case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
03406   case Intrinsic::nvvm_tex_cube_v4u32_f32:
03407   case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
03408   case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
03409   case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
03410   case Intrinsic::nvvm_tex_1d_v4u32_s32:
03411   case Intrinsic::nvvm_tex_1d_v4u32_f32:
03412   case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
03413   case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
03414   case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
03415   case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
03416   case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
03417   case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
03418   case Intrinsic::nvvm_tex_2d_v4u32_s32:
03419   case Intrinsic::nvvm_tex_2d_v4u32_f32:
03420   case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
03421   case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
03422   case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
03423   case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
03424   case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
03425   case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
03426   case Intrinsic::nvvm_tex_3d_v4u32_s32:
03427   case Intrinsic::nvvm_tex_3d_v4u32_f32:
03428   case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
03429   case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
03430   case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
03431   case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
03432   case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
03433   case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
03434   case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
03435   case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
03436   case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
03437   case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
03438   case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
03439   case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
03440   case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
03441   case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
03442   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
03443   case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
03444   case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
03445   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
03446   case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
03447   case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
03448   case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
03449   case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
03450   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
03451   case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
03452   case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
03453   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
03454   case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
03455   case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
03456   case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
03457   case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
03458   case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
03459   case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
03460   case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
03461   case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
03462   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
03463   case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
03464   case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
03465   case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
03466   case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
03467   case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
03468   case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
03469   case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
03470   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
03471   case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
03472   case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
03473   case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
03474   case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
03475   case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
03476   case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
03477   case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
03478   case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
03479   case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
03480   case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
03481   case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
03482   case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
03483   case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
03484   case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
03485   case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
03486   case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
03487   case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
03488   case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
03489   case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
03490   case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
03491   case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
03492   case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
03493   case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32: {
03494     Info.opc = getOpcForTextureInstr(Intrinsic);
03495     Info.memVT = MVT::v4i32;
03496     Info.ptrVal = nullptr;
03497     Info.offset = 0;
03498     Info.vol = 0;
03499     Info.readMem = true;
03500     Info.writeMem = false;
03501     Info.align = 16;
03502     return true;
03503   }
03504   case Intrinsic::nvvm_suld_1d_i8_clamp:
03505   case Intrinsic::nvvm_suld_1d_v2i8_clamp:
03506   case Intrinsic::nvvm_suld_1d_v4i8_clamp:
03507   case Intrinsic::nvvm_suld_1d_array_i8_clamp:
03508   case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
03509   case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
03510   case Intrinsic::nvvm_suld_2d_i8_clamp:
03511   case Intrinsic::nvvm_suld_2d_v2i8_clamp:
03512   case Intrinsic::nvvm_suld_2d_v4i8_clamp:
03513   case Intrinsic::nvvm_suld_2d_array_i8_clamp:
03514   case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
03515   case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
03516   case Intrinsic::nvvm_suld_3d_i8_clamp:
03517   case Intrinsic::nvvm_suld_3d_v2i8_clamp:
03518   case Intrinsic::nvvm_suld_3d_v4i8_clamp:
03519   case Intrinsic::nvvm_suld_1d_i8_trap:
03520   case Intrinsic::nvvm_suld_1d_v2i8_trap:
03521   case Intrinsic::nvvm_suld_1d_v4i8_trap:
03522   case Intrinsic::nvvm_suld_1d_array_i8_trap:
03523   case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
03524   case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
03525   case Intrinsic::nvvm_suld_2d_i8_trap:
03526   case Intrinsic::nvvm_suld_2d_v2i8_trap:
03527   case Intrinsic::nvvm_suld_2d_v4i8_trap:
03528   case Intrinsic::nvvm_suld_2d_array_i8_trap:
03529   case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
03530   case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
03531   case Intrinsic::nvvm_suld_3d_i8_trap:
03532   case Intrinsic::nvvm_suld_3d_v2i8_trap:
03533   case Intrinsic::nvvm_suld_3d_v4i8_trap:
03534   case Intrinsic::nvvm_suld_1d_i8_zero:
03535   case Intrinsic::nvvm_suld_1d_v2i8_zero:
03536   case Intrinsic::nvvm_suld_1d_v4i8_zero:
03537   case Intrinsic::nvvm_suld_1d_array_i8_zero:
03538   case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
03539   case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
03540   case Intrinsic::nvvm_suld_2d_i8_zero:
03541   case Intrinsic::nvvm_suld_2d_v2i8_zero:
03542   case Intrinsic::nvvm_suld_2d_v4i8_zero:
03543   case Intrinsic::nvvm_suld_2d_array_i8_zero:
03544   case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
03545   case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
03546   case Intrinsic::nvvm_suld_3d_i8_zero:
03547   case Intrinsic::nvvm_suld_3d_v2i8_zero:
03548   case Intrinsic::nvvm_suld_3d_v4i8_zero: {
03549     Info.opc = getOpcForSurfaceInstr(Intrinsic);
03550     Info.memVT = MVT::i8;
03551     Info.ptrVal = nullptr;
03552     Info.offset = 0;
03553     Info.vol = 0;
03554     Info.readMem = true;
03555     Info.writeMem = false;
03556     Info.align = 16;
03557     return true;
03558   }
03559   case Intrinsic::nvvm_suld_1d_i16_clamp:
03560   case Intrinsic::nvvm_suld_1d_v2i16_clamp:
03561   case Intrinsic::nvvm_suld_1d_v4i16_clamp:
03562   case Intrinsic::nvvm_suld_1d_array_i16_clamp:
03563   case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
03564   case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
03565   case Intrinsic::nvvm_suld_2d_i16_clamp:
03566   case Intrinsic::nvvm_suld_2d_v2i16_clamp:
03567   case Intrinsic::nvvm_suld_2d_v4i16_clamp:
03568   case Intrinsic::nvvm_suld_2d_array_i16_clamp:
03569   case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
03570   case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
03571   case Intrinsic::nvvm_suld_3d_i16_clamp:
03572   case Intrinsic::nvvm_suld_3d_v2i16_clamp:
03573   case Intrinsic::nvvm_suld_3d_v4i16_clamp:
03574   case Intrinsic::nvvm_suld_1d_i16_trap:
03575   case Intrinsic::nvvm_suld_1d_v2i16_trap:
03576   case Intrinsic::nvvm_suld_1d_v4i16_trap:
03577   case Intrinsic::nvvm_suld_1d_array_i16_trap:
03578   case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
03579   case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
03580   case Intrinsic::nvvm_suld_2d_i16_trap:
03581   case Intrinsic::nvvm_suld_2d_v2i16_trap:
03582   case Intrinsic::nvvm_suld_2d_v4i16_trap:
03583   case Intrinsic::nvvm_suld_2d_array_i16_trap:
03584   case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
03585   case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
03586   case Intrinsic::nvvm_suld_3d_i16_trap:
03587   case Intrinsic::nvvm_suld_3d_v2i16_trap:
03588   case Intrinsic::nvvm_suld_3d_v4i16_trap:
03589   case Intrinsic::nvvm_suld_1d_i16_zero:
03590   case Intrinsic::nvvm_suld_1d_v2i16_zero:
03591   case Intrinsic::nvvm_suld_1d_v4i16_zero:
03592   case Intrinsic::nvvm_suld_1d_array_i16_zero:
03593   case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
03594   case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
03595   case Intrinsic::nvvm_suld_2d_i16_zero:
03596   case Intrinsic::nvvm_suld_2d_v2i16_zero:
03597   case Intrinsic::nvvm_suld_2d_v4i16_zero:
03598   case Intrinsic::nvvm_suld_2d_array_i16_zero:
03599   case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
03600   case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
03601   case Intrinsic::nvvm_suld_3d_i16_zero:
03602   case Intrinsic::nvvm_suld_3d_v2i16_zero:
03603   case Intrinsic::nvvm_suld_3d_v4i16_zero: {
03604     Info.opc = getOpcForSurfaceInstr(Intrinsic);
03605     Info.memVT = MVT::i16;
03606     Info.ptrVal = nullptr;
03607     Info.offset = 0;
03608     Info.vol = 0;
03609     Info.readMem = true;
03610     Info.writeMem = false;
03611     Info.align = 16;
03612     return true;
03613   }
03614   case Intrinsic::nvvm_suld_1d_i32_clamp:
03615   case Intrinsic::nvvm_suld_1d_v2i32_clamp:
03616   case Intrinsic::nvvm_suld_1d_v4i32_clamp:
03617   case Intrinsic::nvvm_suld_1d_array_i32_clamp:
03618   case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
03619   case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
03620   case Intrinsic::nvvm_suld_2d_i32_clamp:
03621   case Intrinsic::nvvm_suld_2d_v2i32_clamp:
03622   case Intrinsic::nvvm_suld_2d_v4i32_clamp:
03623   case Intrinsic::nvvm_suld_2d_array_i32_clamp:
03624   case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
03625   case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
03626   case Intrinsic::nvvm_suld_3d_i32_clamp:
03627   case Intrinsic::nvvm_suld_3d_v2i32_clamp:
03628   case Intrinsic::nvvm_suld_3d_v4i32_clamp:
03629   case Intrinsic::nvvm_suld_1d_i32_trap:
03630   case Intrinsic::nvvm_suld_1d_v2i32_trap:
03631   case Intrinsic::nvvm_suld_1d_v4i32_trap:
03632   case Intrinsic::nvvm_suld_1d_array_i32_trap:
03633   case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
03634   case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
03635   case Intrinsic::nvvm_suld_2d_i32_trap:
03636   case Intrinsic::nvvm_suld_2d_v2i32_trap:
03637   case Intrinsic::nvvm_suld_2d_v4i32_trap:
03638   case Intrinsic::nvvm_suld_2d_array_i32_trap:
03639   case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
03640   case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
03641   case Intrinsic::nvvm_suld_3d_i32_trap:
03642   case Intrinsic::nvvm_suld_3d_v2i32_trap:
03643   case Intrinsic::nvvm_suld_3d_v4i32_trap:
03644   case Intrinsic::nvvm_suld_1d_i32_zero:
03645   case Intrinsic::nvvm_suld_1d_v2i32_zero:
03646   case Intrinsic::nvvm_suld_1d_v4i32_zero:
03647   case Intrinsic::nvvm_suld_1d_array_i32_zero:
03648   case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
03649   case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
03650   case Intrinsic::nvvm_suld_2d_i32_zero:
03651   case Intrinsic::nvvm_suld_2d_v2i32_zero:
03652   case Intrinsic::nvvm_suld_2d_v4i32_zero:
03653   case Intrinsic::nvvm_suld_2d_array_i32_zero:
03654   case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
03655   case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
03656   case Intrinsic::nvvm_suld_3d_i32_zero:
03657   case Intrinsic::nvvm_suld_3d_v2i32_zero:
03658   case Intrinsic::nvvm_suld_3d_v4i32_zero: {
03659     Info.opc = getOpcForSurfaceInstr(Intrinsic);
03660     Info.memVT = MVT::i32;
03661     Info.ptrVal = nullptr;
03662     Info.offset = 0;
03663     Info.vol = 0;
03664     Info.readMem = true;
03665     Info.writeMem = false;
03666     Info.align = 16;
03667     return true;
03668   }
03669   case Intrinsic::nvvm_suld_1d_i64_clamp:
03670   case Intrinsic::nvvm_suld_1d_v2i64_clamp:
03671   case Intrinsic::nvvm_suld_1d_array_i64_clamp:
03672   case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
03673   case Intrinsic::nvvm_suld_2d_i64_clamp:
03674   case Intrinsic::nvvm_suld_2d_v2i64_clamp:
03675   case Intrinsic::nvvm_suld_2d_array_i64_clamp:
03676   case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
03677   case Intrinsic::nvvm_suld_3d_i64_clamp:
03678   case Intrinsic::nvvm_suld_3d_v2i64_clamp:
03679   case Intrinsic::nvvm_suld_1d_i64_trap:
03680   case Intrinsic::nvvm_suld_1d_v2i64_trap:
03681   case Intrinsic::nvvm_suld_1d_array_i64_trap:
03682   case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
03683   case Intrinsic::nvvm_suld_2d_i64_trap:
03684   case Intrinsic::nvvm_suld_2d_v2i64_trap:
03685   case Intrinsic::nvvm_suld_2d_array_i64_trap:
03686   case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
03687   case Intrinsic::nvvm_suld_3d_i64_trap:
03688   case Intrinsic::nvvm_suld_3d_v2i64_trap:
03689   case Intrinsic::nvvm_suld_1d_i64_zero:
03690   case Intrinsic::nvvm_suld_1d_v2i64_zero:
03691   case Intrinsic::nvvm_suld_1d_array_i64_zero:
03692   case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
03693   case Intrinsic::nvvm_suld_2d_i64_zero:
03694   case Intrinsic::nvvm_suld_2d_v2i64_zero:
03695   case Intrinsic::nvvm_suld_2d_array_i64_zero:
03696   case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
03697   case Intrinsic::nvvm_suld_3d_i64_zero:
03698   case Intrinsic::nvvm_suld_3d_v2i64_zero: {
03699     Info.opc = getOpcForSurfaceInstr(Intrinsic);
03700     Info.memVT = MVT::i64;
03701     Info.ptrVal = nullptr;
03702     Info.offset = 0;
03703     Info.vol = 0;
03704     Info.readMem = true;
03705     Info.writeMem = false;
03706     Info.align = 16;
03707     return true;
03708   }
03709   }
03710   return false;
03711 }
03712 
03713 /// isLegalAddressingMode - Return true if the addressing mode represented
03714 /// by AM is legal for this target, for a load/store of the specified type.
03715 /// Used to guide target specific optimizations, like loop strength reduction
03716 /// (LoopStrengthReduce.cpp) and memory optimization for address mode
03717 /// (CodeGenPrepare.cpp)
03718 bool NVPTXTargetLowering::isLegalAddressingMode(const AddrMode &AM,
03719                                                 Type *Ty) const {
03720 
03721   // AddrMode - This represents an addressing mode of:
03722   //    BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
03723   //
03724   // The legal address modes are
03725   // - [avar]
03726   // - [areg]
03727   // - [areg+immoff]
03728   // - [immAddr]
03729 
03730   if (AM.BaseGV) {
03731     if (AM.BaseOffs || AM.HasBaseReg || AM.Scale)
03732       return false;
03733     return true;
03734   }
03735 
03736   switch (AM.Scale) {
03737   case 0: // "r", "r+i" or "i" is allowed
03738     break;
03739   case 1:
03740     if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed.
03741       return false;
03742     // Otherwise we have r+i.
03743     break;
03744   default:
03745     // No scale > 1 is allowed
03746     return false;
03747   }
03748   return true;
03749 }
03750 
03751 //===----------------------------------------------------------------------===//
03752 //                         NVPTX Inline Assembly Support
03753 //===----------------------------------------------------------------------===//
03754 
03755 /// getConstraintType - Given a constraint letter, return the type of
03756 /// constraint it is for this target.
03757 NVPTXTargetLowering::ConstraintType
03758 NVPTXTargetLowering::getConstraintType(const std::string &Constraint) const {
03759   if (Constraint.size() == 1) {
03760     switch (Constraint[0]) {
03761     default:
03762       break;
03763     case 'b':
03764     case 'r':
03765     case 'h':
03766     case 'c':
03767     case 'l':
03768     case 'f':
03769     case 'd':
03770     case '0':
03771     case 'N':
03772       return C_RegisterClass;
03773     }
03774   }
03775   return TargetLowering::getConstraintType(Constraint);
03776 }
03777 
03778 std::pair<unsigned, const TargetRegisterClass *>
03779 NVPTXTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
03780                                                   MVT VT) const {
03781   if (Constraint.size() == 1) {
03782     switch (Constraint[0]) {
03783     case 'b':
03784       return std::make_pair(0U, &NVPTX::Int1RegsRegClass);
03785     case 'c':
03786       return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
03787     case 'h':
03788       return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
03789     case 'r':
03790       return std::make_pair(0U, &NVPTX::Int32RegsRegClass);
03791     case 'l':
03792     case 'N':
03793       return std::make_pair(0U, &NVPTX::Int64RegsRegClass);
03794     case 'f':
03795       return std::make_pair(0U, &NVPTX::Float32RegsRegClass);
03796     case 'd':
03797       return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
03798     }
03799   }
03800   return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
03801 }
03802 
03803 /// getFunctionAlignment - Return the Log2 alignment of this function.
03804 unsigned NVPTXTargetLowering::getFunctionAlignment(const Function *) const {
03805   return 4;
03806 }
03807 
03808 //===----------------------------------------------------------------------===//
03809 //                         NVPTX DAG Combining
03810 //===----------------------------------------------------------------------===//
03811 
03812 bool NVPTXTargetLowering::allowFMA(MachineFunction &MF,
03813                                    CodeGenOpt::Level OptLevel) const {
03814   const Function *F = MF.getFunction();
03815   const TargetOptions &TO = MF.getTarget().Options;
03816 
03817   // Always honor command-line argument
03818   if (FMAContractLevelOpt.getNumOccurrences() > 0) {
03819     return FMAContractLevelOpt > 0;
03820   } else if (OptLevel == 0) {
03821     // Do not contract if we're not optimizing the code
03822     return false;
03823   } else if (TO.AllowFPOpFusion == FPOpFusion::Fast || TO.UnsafeFPMath) {
03824     // Honor TargetOptions flags that explicitly say fusion is okay
03825     return true;
03826   } else if (F->hasFnAttribute("unsafe-fp-math")) {
03827     // Check for unsafe-fp-math=true coming from Clang
03828     Attribute Attr = F->getFnAttribute("unsafe-fp-math");
03829     StringRef Val = Attr.getValueAsString();
03830     if (Val == "true")
03831       return true;
03832   }
03833 
03834   // We did not have a clear indication that fusion is allowed, so assume not
03835   return false;
03836 }
03837 
03838 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
03839 /// operands N0 and N1.  This is a helper for PerformADDCombine that is
03840 /// called with the default operands, and if that fails, with commuted
03841 /// operands.
03842 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
03843                                            TargetLowering::DAGCombinerInfo &DCI,
03844                                              const NVPTXSubtarget &Subtarget,
03845                                              CodeGenOpt::Level OptLevel) {
03846   SelectionDAG  &DAG = DCI.DAG;
03847   // Skip non-integer, non-scalar case
03848   EVT VT=N0.getValueType();
03849   if (VT.isVector())
03850     return SDValue();
03851 
03852   // fold (add (mul a, b), c) -> (mad a, b, c)
03853   //
03854   if (N0.getOpcode() == ISD::MUL) {
03855     assert (VT.isInteger());
03856     // For integer:
03857     // Since integer multiply-add costs the same as integer multiply
03858     // but is more costly than integer add, do the fusion only when
03859     // the mul is only used in the add.
03860     if (OptLevel==CodeGenOpt::None || VT != MVT::i32 ||
03861         !N0.getNode()->hasOneUse())
03862       return SDValue();
03863 
03864     // Do the folding
03865     return DAG.getNode(NVPTXISD::IMAD, SDLoc(N), VT,
03866                        N0.getOperand(0), N0.getOperand(1), N1);
03867   }
03868   else if (N0.getOpcode() == ISD::FMUL) {
03869     if (VT == MVT::f32 || VT == MVT::f64) {
03870       const auto *TLI = static_cast<const NVPTXTargetLowering *>(
03871           &DAG.getTargetLoweringInfo());
03872       if (!TLI->allowFMA(DAG.getMachineFunction(), OptLevel))
03873         return SDValue();
03874 
03875       // For floating point:
03876       // Do the fusion only when the mul has less than 5 uses and all
03877       // are add.
03878       // The heuristic is that if a use is not an add, then that use
03879       // cannot be fused into fma, therefore mul is still needed anyway.
03880       // If there are more than 4 uses, even if they are all add, fusing
03881       // them will increase register pressue.
03882       //
03883       int numUses = 0;
03884       int nonAddCount = 0;
03885       for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
03886            UE = N0.getNode()->use_end();
03887            UI != UE; ++UI) {
03888         numUses++;
03889         SDNode *User = *UI;
03890         if (User->getOpcode() != ISD::FADD)
03891           ++nonAddCount;
03892       }
03893       if (numUses >= 5)
03894         return SDValue();
03895       if (nonAddCount) {
03896         int orderNo = N->getIROrder();
03897         int orderNo2 = N0.getNode()->getIROrder();
03898         // simple heuristics here for considering potential register
03899         // pressure, the logics here is that the differnce are used
03900         // to measure the distance between def and use, the longer distance
03901         // more likely cause register pressure.
03902         if (orderNo - orderNo2 < 500)
03903           return SDValue();
03904 
03905         // Now, check if at least one of the FMUL's operands is live beyond the node N,
03906         // which guarantees that the FMA will not increase register pressure at node N.
03907         bool opIsLive = false;
03908         const SDNode *left = N0.getOperand(0).getNode();
03909         const SDNode *right = N0.getOperand(1).getNode();
03910 
03911         if (dyn_cast<ConstantSDNode>(left) || dyn_cast<ConstantSDNode>(right))
03912           opIsLive = true;
03913 
03914         if (!opIsLive)
03915           for (SDNode::use_iterator UI = left->use_begin(), UE = left->use_end(); UI != UE; ++UI) {
03916             SDNode *User = *UI;
03917             int orderNo3 = User->getIROrder();
03918             if (orderNo3 > orderNo) {
03919               opIsLive = true;
03920               break;
03921             }
03922           }
03923 
03924         if (!opIsLive)
03925           for (SDNode::use_iterator UI = right->use_begin(), UE = right->use_end(); UI != UE; ++UI) {
03926             SDNode *User = *UI;
03927             int orderNo3 = User->getIROrder();
03928             if (orderNo3 > orderNo) {
03929               opIsLive = true;
03930               break;
03931             }
03932           }
03933 
03934         if (!opIsLive)
03935           return SDValue();
03936       }
03937 
03938       return DAG.getNode(ISD::FMA, SDLoc(N), VT,
03939                          N0.getOperand(0), N0.getOperand(1), N1);
03940     }
03941   }
03942 
03943   return SDValue();
03944 }
03945 
03946 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
03947 ///
03948 static SDValue PerformADDCombine(SDNode *N,
03949                                  TargetLowering::DAGCombinerInfo &DCI,
03950                                  const NVPTXSubtarget &Subtarget,
03951                                  CodeGenOpt::Level OptLevel) {
03952   SDValue N0 = N->getOperand(0);
03953   SDValue N1 = N->getOperand(1);
03954 
03955   // First try with the default operand order.
03956   SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget,
03957                                                  OptLevel);
03958   if (Result.getNode())
03959     return Result;
03960 
03961   // If that didn't work, try again with the operands commuted.
03962   return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget, OptLevel);
03963 }
03964 
03965 static SDValue PerformANDCombine(SDNode *N,
03966                                  TargetLowering::DAGCombinerInfo &DCI) {
03967   // The type legalizer turns a vector load of i8 values into a zextload to i16
03968   // registers, optionally ANY_EXTENDs it (if target type is integer),
03969   // and ANDs off the high 8 bits. Since we turn this load into a
03970   // target-specific DAG node, the DAG combiner fails to eliminate these AND
03971   // nodes. Do that here.
03972   SDValue Val = N->getOperand(0);
03973   SDValue Mask = N->getOperand(1);
03974 
03975   if (isa<ConstantSDNode>(Val)) {
03976     std::swap(Val, Mask);
03977   }
03978 
03979   SDValue AExt;
03980   // Generally, we will see zextload -> IMOV16rr -> ANY_EXTEND -> and
03981   if (Val.getOpcode() == ISD::ANY_EXTEND) {
03982     AExt = Val;
03983     Val = Val->getOperand(0);
03984   }
03985 
03986   if (Val->isMachineOpcode() && Val->getMachineOpcode() == NVPTX::IMOV16rr) {
03987     Val = Val->getOperand(0);
03988   }
03989 
03990   if (Val->getOpcode() == NVPTXISD::LoadV2 ||
03991       Val->getOpcode() == NVPTXISD::LoadV4) {
03992     ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(Mask);
03993     if (!MaskCnst) {
03994       // Not an AND with a constant
03995       return SDValue();
03996     }
03997 
03998     uint64_t MaskVal = MaskCnst->getZExtValue();
03999     if (MaskVal != 0xff) {
04000       // Not an AND that chops off top 8 bits
04001       return SDValue();
04002     }
04003 
04004     MemSDNode *Mem = dyn_cast<MemSDNode>(Val);
04005     if (!Mem) {
04006       // Not a MemSDNode?!?
04007       return SDValue();
04008     }
04009 
04010     EVT MemVT = Mem->getMemoryVT();
04011     if (MemVT != MVT::v2i8 && MemVT != MVT::v4i8) {
04012       // We only handle the i8 case
04013       return SDValue();
04014     }
04015 
04016     unsigned ExtType =
04017       cast<ConstantSDNode>(Val->getOperand(Val->getNumOperands()-1))->
04018         getZExtValue();
04019     if (ExtType == ISD::SEXTLOAD) {
04020       // If for some reason the load is a sextload, the and is needed to zero
04021       // out the high 8 bits
04022       return SDValue();
04023     }
04024 
04025     bool AddTo = false;
04026     if (AExt.getNode() != 0) {
04027       // Re-insert the ext as a zext.
04028       Val = DCI.DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
04029                             AExt.getValueType(), Val);
04030       AddTo = true;
04031     }
04032 
04033     // If we get here, the AND is unnecessary.  Just replace it with the load
04034     DCI.CombineTo(N, Val, AddTo);
04035   }
04036 
04037   return SDValue();
04038 }
04039 
04040 enum OperandSignedness {
04041   Signed = 0,
04042   Unsigned,
04043   Unknown
04044 };
04045 
04046 /// IsMulWideOperandDemotable - Checks if the provided DAG node is an operand
04047 /// that can be demoted to \p OptSize bits without loss of information. The
04048 /// signedness of the operand, if determinable, is placed in \p S.
04049 static bool IsMulWideOperandDemotable(SDValue Op,
04050                                       unsigned OptSize,
04051                                       OperandSignedness &S) {
04052   S = Unknown;
04053 
04054   if (Op.getOpcode() == ISD::SIGN_EXTEND ||
04055       Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
04056     EVT OrigVT = Op.getOperand(0).getValueType();
04057     if (OrigVT.getSizeInBits() <= OptSize) {
04058       S = Signed;
04059       return true;
04060     }
04061   } else if (Op.getOpcode() == ISD::ZERO_EXTEND) {
04062     EVT OrigVT = Op.getOperand(0).getValueType();
04063     if (OrigVT.getSizeInBits() <= OptSize) {
04064       S = Unsigned;
04065       return true;
04066     }
04067   }
04068 
04069   return false;
04070 }
04071 
04072 /// AreMulWideOperandsDemotable - Checks if the given LHS and RHS operands can
04073 /// be demoted to \p OptSize bits without loss of information. If the operands
04074 /// contain a constant, it should appear as the RHS operand. The signedness of
04075 /// the operands is placed in \p IsSigned.
04076 static bool AreMulWideOperandsDemotable(SDValue LHS, SDValue RHS,
04077                                         unsigned OptSize,
04078                                         bool &IsSigned) {
04079 
04080   OperandSignedness LHSSign;
04081 
04082   // The LHS operand must be a demotable op
04083   if (!IsMulWideOperandDemotable(LHS, OptSize, LHSSign))
04084     return false;
04085 
04086   // We should have been able to determine the signedness from the LHS
04087   if (LHSSign == Unknown)
04088     return false;
04089 
04090   IsSigned = (LHSSign == Signed);
04091 
04092   // The RHS can be a demotable op or a constant
04093   if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(RHS)) {
04094     APInt Val = CI->getAPIntValue();
04095     if (LHSSign == Unsigned) {
04096       if (Val.isIntN(OptSize)) {
04097         return true;
04098       }
04099       return false;
04100     } else {
04101       if (Val.isSignedIntN(OptSize)) {
04102         return true;
04103       }
04104       return false;
04105     }
04106   } else {
04107     OperandSignedness RHSSign;
04108     if (!IsMulWideOperandDemotable(RHS, OptSize, RHSSign))
04109       return false;
04110 
04111     if (LHSSign != RHSSign)
04112       return false;
04113 
04114     return true;
04115   }
04116 }
04117 
04118 /// TryMULWIDECombine - Attempt to replace a multiply of M bits with a multiply
04119 /// of M/2 bits that produces an M-bit result (i.e. mul.wide). This transform
04120 /// works on both multiply DAG nodes and SHL DAG nodes with a constant shift
04121 /// amount.
04122 static SDValue TryMULWIDECombine(SDNode *N,
04123                                  TargetLowering::DAGCombinerInfo &DCI) {
04124   EVT MulType = N->getValueType(0);
04125   if (MulType != MVT::i32 && MulType != MVT::i64) {
04126     return SDValue();
04127   }
04128 
04129   unsigned OptSize = MulType.getSizeInBits() >> 1;
04130   SDValue LHS = N->getOperand(0);
04131   SDValue RHS = N->getOperand(1);
04132 
04133   // Canonicalize the multiply so the constant (if any) is on the right
04134   if (N->getOpcode() == ISD::MUL) {
04135     if (isa<ConstantSDNode>(LHS)) {
04136       std::swap(LHS, RHS);
04137     }
04138   }
04139 
04140   // If we have a SHL, determine the actual multiply amount
04141   if (N->getOpcode() == ISD::SHL) {
04142     ConstantSDNode *ShlRHS = dyn_cast<ConstantSDNode>(RHS);
04143     if (!ShlRHS) {
04144       return SDValue();
04145     }
04146 
04147     APInt ShiftAmt = ShlRHS->getAPIntValue();
04148     unsigned BitWidth = MulType.getSizeInBits();
04149     if (ShiftAmt.sge(0) && ShiftAmt.slt(BitWidth)) {
04150       APInt MulVal = APInt(BitWidth, 1) << ShiftAmt;
04151       RHS = DCI.DAG.getConstant(MulVal, MulType);
04152     } else {
04153       return SDValue();
04154     }
04155   }
04156 
04157   bool Signed;
04158   // Verify that our operands are demotable
04159   if (!AreMulWideOperandsDemotable(LHS, RHS, OptSize, Signed)) {
04160     return SDValue();
04161   }
04162 
04163   EVT DemotedVT;
04164   if (MulType == MVT::i32) {
04165     DemotedVT = MVT::i16;
04166   } else {
04167     DemotedVT = MVT::i32;
04168   }
04169 
04170   // Truncate the operands to the correct size. Note that these are just for
04171   // type consistency and will (likely) be eliminated in later phases.
04172   SDValue TruncLHS =
04173     DCI.DAG.getNode(ISD::TRUNCATE, SDLoc(N), DemotedVT, LHS);
04174   SDValue TruncRHS =
04175     DCI.DAG.getNode(ISD::TRUNCATE, SDLoc(N), DemotedVT, RHS);
04176 
04177   unsigned Opc;
04178   if (Signed) {
04179     Opc = NVPTXISD::MUL_WIDE_SIGNED;
04180   } else {
04181     Opc = NVPTXISD::MUL_WIDE_UNSIGNED;
04182   }
04183 
04184   return DCI.DAG.getNode(Opc, SDLoc(N), MulType, TruncLHS, TruncRHS);
04185 }
04186 
04187 /// PerformMULCombine - Runs PTX-specific DAG combine patterns on MUL nodes.
04188 static SDValue PerformMULCombine(SDNode *N,
04189                                  TargetLowering::DAGCombinerInfo &DCI,
04190                                  CodeGenOpt::Level OptLevel) {
04191   if (OptLevel > 0) {
04192     // Try mul.wide combining at OptLevel > 0
04193     SDValue Ret = TryMULWIDECombine(N, DCI);
04194     if (Ret.getNode())
04195       return Ret;
04196   }
04197 
04198   return SDValue();
04199 }
04200 
04201 /// PerformSHLCombine - Runs PTX-specific DAG combine patterns on SHL nodes.
04202 static SDValue PerformSHLCombine(SDNode *N,
04203                                  TargetLowering::DAGCombinerInfo &DCI,
04204                                  CodeGenOpt::Level OptLevel) {
04205   if (OptLevel > 0) {
04206     // Try mul.wide combining at OptLevel > 0
04207     SDValue Ret = TryMULWIDECombine(N, DCI);
04208     if (Ret.getNode())
04209       return Ret;
04210   }
04211 
04212   return SDValue();
04213 }
04214 
04215 SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
04216                                                DAGCombinerInfo &DCI) const {
04217   CodeGenOpt::Level OptLevel = getTargetMachine().getOptLevel();
04218   switch (N->getOpcode()) {
04219     default: break;
04220     case ISD::ADD:
04221     case ISD::FADD:
04222       return PerformADDCombine(N, DCI, nvptxSubtarget, OptLevel);
04223     case ISD::MUL:
04224       return PerformMULCombine(N, DCI, OptLevel);
04225     case ISD::SHL:
04226       return PerformSHLCombine(N, DCI, OptLevel);
04227     case ISD::AND:
04228       return PerformANDCombine(N, DCI);
04229   }
04230   return SDValue();
04231 }
04232 
04233 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
04234 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
04235                               const DataLayout *TD,
04236                               SmallVectorImpl<SDValue> &Results) {
04237   EVT ResVT = N->getValueType(0);
04238   SDLoc DL(N);
04239 
04240   assert(ResVT.isVector() && "Vector load must have vector type");
04241 
04242   // We only handle "native" vector sizes for now, e.g. <4 x double> is not
04243   // legal.  We can (and should) split that into 2 loads of <2 x double> here
04244   // but I'm leaving that as a TODO for now.
04245   assert(ResVT.isSimple() && "Can only handle simple types");
04246   switch (ResVT.getSimpleVT().SimpleTy) {
04247   default:
04248     return;
04249   case MVT::v2i8:
04250   case MVT::v2i16:
04251   case MVT::v2i32:
04252   case MVT::v2i64:
04253   case MVT::v2f32:
04254   case MVT::v2f64:
04255   case MVT::v4i8:
04256   case MVT::v4i16:
04257   case MVT::v4i32:
04258   case MVT::v4f32:
04259     // This is a "native" vector type
04260     break;
04261   }
04262 
04263   LoadSDNode *LD = cast<LoadSDNode>(N);
04264 
04265   unsigned Align = LD->getAlignment();
04266   unsigned PrefAlign =
04267     TD->getPrefTypeAlignment(ResVT.getTypeForEVT(*DAG.getContext()));
04268   if (Align < PrefAlign) {
04269     // This load is not sufficiently aligned, so bail out and let this vector
04270     // load be scalarized.  Note that we may still be able to emit smaller
04271     // vector loads.  For example, if we are loading a <4 x float> with an
04272     // alignment of 8, this check will fail but the legalizer will try again
04273     // with 2 x <2 x float>, which will succeed with an alignment of 8.
04274     return;
04275   }
04276 
04277   EVT EltVT = ResVT.getVectorElementType();
04278   unsigned NumElts = ResVT.getVectorNumElements();
04279 
04280   // Since LoadV2 is a target node, we cannot rely on DAG type legalization.
04281   // Therefore, we must ensure the type is legal.  For i1 and i8, we set the
04282   // loaded type to i16 and propagate the "real" type as the memory type.
04283   bool NeedTrunc = false;
04284   if (EltVT.getSizeInBits() < 16) {
04285     EltVT = MVT::i16;
04286     NeedTrunc = true;
04287   }
04288 
04289   unsigned Opcode = 0;
04290   SDVTList LdResVTs;
04291 
04292   switch (NumElts) {
04293   default:
04294     return;
04295   case 2:
04296     Opcode = NVPTXISD::LoadV2;
04297     LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
04298     break;
04299   case 4: {
04300     Opcode = NVPTXISD::LoadV4;
04301     EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
04302     LdResVTs = DAG.getVTList(ListVTs);
04303     break;
04304   }
04305   }
04306 
04307   SmallVector<SDValue, 8> OtherOps;
04308 
04309   // Copy regular operands
04310   for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
04311     OtherOps.push_back(N->getOperand(i));
04312 
04313   // The select routine does not have access to the LoadSDNode instance, so
04314   // pass along the extension information
04315   OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType()));
04316 
04317   SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
04318                                           LD->getMemoryVT(),
04319                                           LD->getMemOperand());
04320 
04321   SmallVector<SDValue, 4> ScalarRes;
04322 
04323   for (unsigned i = 0; i < NumElts; ++i) {
04324     SDValue Res = NewLD.getValue(i);
04325     if (NeedTrunc)
04326       Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
04327     ScalarRes.push_back(Res);
04328   }
04329 
04330   SDValue LoadChain = NewLD.getValue(NumElts);
04331 
04332   SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
04333 
04334   Results.push_back(BuildVec);
04335   Results.push_back(LoadChain);
04336 }
04337 
04338 static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG,
04339                                      SmallVectorImpl<SDValue> &Results) {
04340   SDValue Chain = N->getOperand(0);
04341   SDValue Intrin = N->getOperand(1);
04342   SDLoc DL(N);
04343 
04344   // Get the intrinsic ID
04345   unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue();
04346   switch (IntrinNo) {
04347   default:
04348     return;
04349   case Intrinsic::nvvm_ldg_global_i:
04350   case Intrinsic::nvvm_ldg_global_f:
04351   case Intrinsic::nvvm_ldg_global_p:
04352   case Intrinsic::nvvm_ldu_global_i:
04353   case Intrinsic::nvvm_ldu_global_f:
04354   case Intrinsic::nvvm_ldu_global_p: {
04355     EVT ResVT = N->getValueType(0);
04356 
04357     if (ResVT.isVector()) {
04358       // Vector LDG/LDU
04359 
04360       unsigned NumElts = ResVT.getVectorNumElements();
04361       EVT EltVT = ResVT.getVectorElementType();
04362 
04363       // Since LDU/LDG are target nodes, we cannot rely on DAG type
04364       // legalization.
04365       // Therefore, we must ensure the type is legal.  For i1 and i8, we set the
04366       // loaded type to i16 and propagate the "real" type as the memory type.
04367       bool NeedTrunc = false;
04368       if (EltVT.getSizeInBits() < 16) {
04369         EltVT = MVT::i16;
04370         NeedTrunc = true;
04371       }
04372 
04373       unsigned Opcode = 0;
04374       SDVTList LdResVTs;
04375 
04376       switch (NumElts) {
04377       default:
04378         return;
04379       case 2:
04380         switch (IntrinNo) {
04381         default:
04382           return;
04383         case Intrinsic::nvvm_ldg_global_i:
04384         case Intrinsic::nvvm_ldg_global_f:
04385         case Intrinsic::nvvm_ldg_global_p:
04386           Opcode = NVPTXISD::LDGV2;
04387           break;
04388         case Intrinsic::nvvm_ldu_global_i:
04389         case Intrinsic::nvvm_ldu_global_f:
04390         case Intrinsic::nvvm_ldu_global_p:
04391           Opcode = NVPTXISD::LDUV2;
04392           break;
04393         }
04394         LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
04395         break;
04396       case 4: {
04397         switch (IntrinNo) {
04398         default:
04399           return;
04400         case Intrinsic::nvvm_ldg_global_i:
04401         case Intrinsic::nvvm_ldg_global_f:
04402         case Intrinsic::nvvm_ldg_global_p:
04403           Opcode = NVPTXISD::LDGV4;
04404           break;
04405         case Intrinsic::nvvm_ldu_global_i:
04406         case Intrinsic::nvvm_ldu_global_f:
04407         case Intrinsic::nvvm_ldu_global_p:
04408           Opcode = NVPTXISD::LDUV4;
04409           break;
04410         }
04411         EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
04412         LdResVTs = DAG.getVTList(ListVTs);
04413         break;
04414       }
04415       }
04416 
04417       SmallVector<SDValue, 8> OtherOps;
04418 
04419       // Copy regular operands
04420 
04421       OtherOps.push_back(Chain); // Chain
04422                                  // Skip operand 1 (intrinsic ID)
04423       // Others
04424       for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i)
04425         OtherOps.push_back(N->getOperand(i));
04426 
04427       MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
04428 
04429       SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
04430                                               MemSD->getMemoryVT(),
04431                                               MemSD->getMemOperand());
04432 
04433       SmallVector<SDValue, 4> ScalarRes;
04434 
04435       for (unsigned i = 0; i < NumElts; ++i) {
04436         SDValue Res = NewLD.getValue(i);
04437         if (NeedTrunc)
04438           Res =
04439               DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
04440         ScalarRes.push_back(Res);
04441       }
04442 
04443       SDValue LoadChain = NewLD.getValue(NumElts);
04444 
04445       SDValue BuildVec =
04446           DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
04447 
04448       Results.push_back(BuildVec);
04449       Results.push_back(LoadChain);
04450     } else {
04451       // i8 LDG/LDU
04452       assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 &&
04453              "Custom handling of non-i8 ldu/ldg?");
04454 
04455       // Just copy all operands as-is
04456       SmallVector<SDValue, 4> Ops;
04457       for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
04458         Ops.push_back(N->getOperand(i));
04459 
04460       // Force output to i16
04461       SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
04462 
04463       MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
04464 
04465       // We make sure the memory type is i8, which will be used during isel
04466       // to select the proper instruction.
04467       SDValue NewLD =
04468           DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops,
04469                                   MVT::i8, MemSD->getMemOperand());
04470 
04471       Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
04472                                     NewLD.getValue(0)));
04473       Results.push_back(NewLD.getValue(1));
04474     }
04475   }
04476   }
04477 }
04478 
04479 void NVPTXTargetLowering::ReplaceNodeResults(
04480     SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
04481   switch (N->getOpcode()) {
04482   default:
04483     report_fatal_error("Unhandled custom legalization");
04484   case ISD::LOAD:
04485     ReplaceLoadVector(N, DAG, getDataLayout(), Results);
04486     return;
04487   case ISD::INTRINSIC_W_CHAIN:
04488     ReplaceINTRINSIC_W_CHAIN(N, DAG, Results);
04489     return;
04490   }
04491 }
04492 
04493 // Pin NVPTXSection's and NVPTXTargetObjectFile's vtables to this file.
04494 void NVPTXSection::anchor() {}
04495 
04496 NVPTXTargetObjectFile::~NVPTXTargetObjectFile() {
04497   delete TextSection;
04498   delete DataSection;
04499   delete BSSSection;
04500   delete ReadOnlySection;
04501 
04502   delete StaticCtorSection;
04503   delete StaticDtorSection;
04504   delete LSDASection;
04505   delete EHFrameSection;
04506   delete DwarfAbbrevSection;
04507   delete DwarfInfoSection;
04508   delete DwarfLineSection;
04509   delete DwarfFrameSection;
04510   delete DwarfPubTypesSection;
04511   delete DwarfDebugInlineSection;
04512   delete DwarfStrSection;
04513   delete DwarfLocSection;
04514   delete DwarfARangesSection;
04515   delete DwarfRangesSection;
04516   delete DwarfMacroInfoSection;
04517 }