Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1 | // |
| 2 | // The LLVM Compiler Infrastructure |
| 3 | // |
| 4 | // This file is distributed under the University of Illinois Open Source |
| 5 | // License. See LICENSE.TXT for details. |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file defines the interfaces that NVPTX uses to lower LLVM code into a |
| 10 | // selection DAG. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 14 | #include "NVPTXISelLowering.h" |
Chandler Carruth | ed0881b | 2012-12-03 16:50:05 +0000 | [diff] [blame] | 15 | #include "NVPTX.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 16 | #include "NVPTXTargetMachine.h" |
| 17 | #include "NVPTXTargetObjectFile.h" |
| 18 | #include "NVPTXUtilities.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 19 | #include "llvm/CodeGen/Analysis.h" |
| 20 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 21 | #include "llvm/CodeGen/MachineFunction.h" |
| 22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 23 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 24 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
Chandler Carruth | 219b89b | 2014-03-04 11:01:28 +0000 | [diff] [blame] | 25 | #include "llvm/IR/CallSite.h" |
Chandler Carruth | 9fb823b | 2013-01-02 11:36:10 +0000 | [diff] [blame] | 26 | #include "llvm/IR/DerivedTypes.h" |
| 27 | #include "llvm/IR/Function.h" |
| 28 | #include "llvm/IR/GlobalValue.h" |
| 29 | #include "llvm/IR/IntrinsicInst.h" |
| 30 | #include "llvm/IR/Intrinsics.h" |
| 31 | #include "llvm/IR/Module.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 32 | #include "llvm/MC/MCSectionELF.h" |
Chandler Carruth | ed0881b | 2012-12-03 16:50:05 +0000 | [diff] [blame] | 33 | #include "llvm/Support/CommandLine.h" |
| 34 | #include "llvm/Support/Debug.h" |
| 35 | #include "llvm/Support/ErrorHandling.h" |
| 36 | #include "llvm/Support/raw_ostream.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 37 | #include <sstream> |
| 38 | |
| 39 | #undef DEBUG_TYPE |
| 40 | #define DEBUG_TYPE "nvptx-lower" |
| 41 | |
| 42 | using namespace llvm; |
| 43 | |
| 44 | static unsigned int uniqueCallSite = 0; |
| 45 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 46 | static cl::opt<bool> sched4reg( |
| 47 | "nvptx-sched4reg", |
| 48 | cl::desc("NVPTX Specific: schedule for register pressue"), cl::init(false)); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 49 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 50 | static bool IsPTXVectorType(MVT VT) { |
| 51 | switch (VT.SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 52 | default: |
| 53 | return false; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 54 | case MVT::v2i1: |
| 55 | case MVT::v4i1: |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 56 | case MVT::v2i8: |
| 57 | case MVT::v4i8: |
| 58 | case MVT::v2i16: |
| 59 | case MVT::v4i16: |
| 60 | case MVT::v2i32: |
| 61 | case MVT::v4i32: |
| 62 | case MVT::v2i64: |
| 63 | case MVT::v2f32: |
| 64 | case MVT::v4f32: |
| 65 | case MVT::v2f64: |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 66 | return true; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 67 | } |
| 68 | } |
| 69 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 70 | /// ComputePTXValueVTs - For the given Type \p Ty, returns the set of primitive |
| 71 | /// EVTs that compose it. Unlike ComputeValueVTs, this will break apart vectors |
| 72 | /// into their primitive components. |
| 73 | /// NOTE: This is a band-aid for code that expects ComputeValueVTs to return the |
| 74 | /// same number of types as the Ins/Outs arrays in LowerFormalArguments, |
| 75 | /// LowerCall, and LowerReturn. |
| 76 | static void ComputePTXValueVTs(const TargetLowering &TLI, Type *Ty, |
| 77 | SmallVectorImpl<EVT> &ValueVTs, |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 78 | SmallVectorImpl<uint64_t> *Offsets = nullptr, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 79 | uint64_t StartingOffset = 0) { |
| 80 | SmallVector<EVT, 16> TempVTs; |
| 81 | SmallVector<uint64_t, 16> TempOffsets; |
| 82 | |
| 83 | ComputeValueVTs(TLI, Ty, TempVTs, &TempOffsets, StartingOffset); |
| 84 | for (unsigned i = 0, e = TempVTs.size(); i != e; ++i) { |
| 85 | EVT VT = TempVTs[i]; |
| 86 | uint64_t Off = TempOffsets[i]; |
| 87 | if (VT.isVector()) |
| 88 | for (unsigned j = 0, je = VT.getVectorNumElements(); j != je; ++j) { |
| 89 | ValueVTs.push_back(VT.getVectorElementType()); |
| 90 | if (Offsets) |
| 91 | Offsets->push_back(Off+j*VT.getVectorElementType().getStoreSize()); |
| 92 | } |
| 93 | else { |
| 94 | ValueVTs.push_back(VT); |
| 95 | if (Offsets) |
| 96 | Offsets->push_back(Off); |
| 97 | } |
| 98 | } |
| 99 | } |
| 100 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 101 | // NVPTXTargetLowering Constructor. |
| 102 | NVPTXTargetLowering::NVPTXTargetLowering(NVPTXTargetMachine &TM) |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 103 | : TargetLowering(TM, new NVPTXTargetObjectFile()), nvTM(&TM), |
| 104 | nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 105 | |
| 106 | // always lower memset, memcpy, and memmove intrinsics to load/store |
| 107 | // instructions, rather |
| 108 | // then generating calls to memset, mempcy or memmove. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 109 | MaxStoresPerMemset = (unsigned) 0xFFFFFFFF; |
| 110 | MaxStoresPerMemcpy = (unsigned) 0xFFFFFFFF; |
| 111 | MaxStoresPerMemmove = (unsigned) 0xFFFFFFFF; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 112 | |
| 113 | setBooleanContents(ZeroOrNegativeOneBooleanContent); |
| 114 | |
| 115 | // Jump is Expensive. Don't create extra control flow for 'and', 'or' |
| 116 | // condition branches. |
| 117 | setJumpIsExpensive(true); |
| 118 | |
| 119 | // By default, use the Source scheduling |
| 120 | if (sched4reg) |
| 121 | setSchedulingPreference(Sched::RegPressure); |
| 122 | else |
| 123 | setSchedulingPreference(Sched::Source); |
| 124 | |
| 125 | addRegisterClass(MVT::i1, &NVPTX::Int1RegsRegClass); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 126 | addRegisterClass(MVT::i16, &NVPTX::Int16RegsRegClass); |
| 127 | addRegisterClass(MVT::i32, &NVPTX::Int32RegsRegClass); |
| 128 | addRegisterClass(MVT::i64, &NVPTX::Int64RegsRegClass); |
| 129 | addRegisterClass(MVT::f32, &NVPTX::Float32RegsRegClass); |
| 130 | addRegisterClass(MVT::f64, &NVPTX::Float64RegsRegClass); |
| 131 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 132 | // Operations not directly supported by NVPTX. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 133 | setOperationAction(ISD::SELECT_CC, MVT::Other, Expand); |
| 134 | setOperationAction(ISD::BR_CC, MVT::f32, Expand); |
| 135 | setOperationAction(ISD::BR_CC, MVT::f64, Expand); |
| 136 | setOperationAction(ISD::BR_CC, MVT::i1, Expand); |
| 137 | setOperationAction(ISD::BR_CC, MVT::i8, Expand); |
| 138 | setOperationAction(ISD::BR_CC, MVT::i16, Expand); |
| 139 | setOperationAction(ISD::BR_CC, MVT::i32, Expand); |
| 140 | setOperationAction(ISD::BR_CC, MVT::i64, Expand); |
Justin Holewinski | 318c625 | 2013-07-01 12:58:56 +0000 | [diff] [blame] | 141 | // Some SIGN_EXTEND_INREG can be done using cvt instruction. |
| 142 | // For others we will expand to a SHL/SRA pair. |
| 143 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i64, Legal); |
| 144 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal); |
| 145 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal); |
| 146 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 147 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 148 | |
| 149 | if (nvptxSubtarget.hasROT64()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 150 | setOperationAction(ISD::ROTL, MVT::i64, Legal); |
| 151 | setOperationAction(ISD::ROTR, MVT::i64, Legal); |
| 152 | } else { |
| 153 | setOperationAction(ISD::ROTL, MVT::i64, Expand); |
| 154 | setOperationAction(ISD::ROTR, MVT::i64, Expand); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 155 | } |
| 156 | if (nvptxSubtarget.hasROT32()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 157 | setOperationAction(ISD::ROTL, MVT::i32, Legal); |
| 158 | setOperationAction(ISD::ROTR, MVT::i32, Legal); |
| 159 | } else { |
| 160 | setOperationAction(ISD::ROTL, MVT::i32, Expand); |
| 161 | setOperationAction(ISD::ROTR, MVT::i32, Expand); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 162 | } |
| 163 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 164 | setOperationAction(ISD::ROTL, MVT::i16, Expand); |
| 165 | setOperationAction(ISD::ROTR, MVT::i16, Expand); |
| 166 | setOperationAction(ISD::ROTL, MVT::i8, Expand); |
| 167 | setOperationAction(ISD::ROTR, MVT::i8, Expand); |
| 168 | setOperationAction(ISD::BSWAP, MVT::i16, Expand); |
| 169 | setOperationAction(ISD::BSWAP, MVT::i32, Expand); |
| 170 | setOperationAction(ISD::BSWAP, MVT::i64, Expand); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 171 | |
| 172 | // Indirect branch is not supported. |
| 173 | // This also disables Jump Table creation. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 174 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); |
| 175 | setOperationAction(ISD::BRIND, MVT::Other, Expand); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 176 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 177 | setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); |
| 178 | setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 179 | |
| 180 | // We want to legalize constant related memmove and memcopy |
| 181 | // intrinsics. |
| 182 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
| 183 | |
| 184 | // Turn FP extload into load/fextend |
| 185 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); |
| 186 | // Turn FP truncstore into trunc + store. |
| 187 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
| 188 | |
| 189 | // PTX does not support load / store predicate registers |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 190 | setOperationAction(ISD::LOAD, MVT::i1, Custom); |
| 191 | setOperationAction(ISD::STORE, MVT::i1, Custom); |
| 192 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 193 | setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); |
| 194 | setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 195 | setTruncStoreAction(MVT::i64, MVT::i1, Expand); |
| 196 | setTruncStoreAction(MVT::i32, MVT::i1, Expand); |
| 197 | setTruncStoreAction(MVT::i16, MVT::i1, Expand); |
| 198 | setTruncStoreAction(MVT::i8, MVT::i1, Expand); |
| 199 | |
| 200 | // This is legal in NVPTX |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 201 | setOperationAction(ISD::ConstantFP, MVT::f64, Legal); |
| 202 | setOperationAction(ISD::ConstantFP, MVT::f32, Legal); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 203 | |
| 204 | // TRAP can be lowered to PTX trap |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 205 | setOperationAction(ISD::TRAP, MVT::Other, Legal); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 206 | |
Justin Holewinski | 51cb134 | 2013-07-01 12:59:04 +0000 | [diff] [blame] | 207 | setOperationAction(ISD::ADDC, MVT::i64, Expand); |
| 208 | setOperationAction(ISD::ADDE, MVT::i64, Expand); |
| 209 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 210 | // Register custom handling for vector loads/stores |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 211 | for (int i = MVT::FIRST_VECTOR_VALUETYPE; i <= MVT::LAST_VECTOR_VALUETYPE; |
| 212 | ++i) { |
| 213 | MVT VT = (MVT::SimpleValueType) i; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 214 | if (IsPTXVectorType(VT)) { |
| 215 | setOperationAction(ISD::LOAD, VT, Custom); |
| 216 | setOperationAction(ISD::STORE, VT, Custom); |
| 217 | setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom); |
| 218 | } |
| 219 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 220 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 221 | // Custom handling for i8 intrinsics |
| 222 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom); |
| 223 | |
Justin Holewinski | dc372df | 2013-06-28 17:58:07 +0000 | [diff] [blame] | 224 | setOperationAction(ISD::CTLZ, MVT::i16, Legal); |
| 225 | setOperationAction(ISD::CTLZ, MVT::i32, Legal); |
| 226 | setOperationAction(ISD::CTLZ, MVT::i64, Legal); |
| 227 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Legal); |
| 228 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Legal); |
| 229 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Legal); |
| 230 | setOperationAction(ISD::CTTZ, MVT::i16, Expand); |
| 231 | setOperationAction(ISD::CTTZ, MVT::i32, Expand); |
| 232 | setOperationAction(ISD::CTTZ, MVT::i64, Expand); |
| 233 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Expand); |
| 234 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); |
| 235 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand); |
| 236 | setOperationAction(ISD::CTPOP, MVT::i16, Legal); |
| 237 | setOperationAction(ISD::CTPOP, MVT::i32, Legal); |
| 238 | setOperationAction(ISD::CTPOP, MVT::i64, Legal); |
| 239 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 240 | // Now deduce the information based on the above mentioned |
| 241 | // actions |
| 242 | computeRegisterProperties(); |
| 243 | } |
| 244 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 245 | const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const { |
| 246 | switch (Opcode) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 247 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 248 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 249 | case NVPTXISD::CALL: |
| 250 | return "NVPTXISD::CALL"; |
| 251 | case NVPTXISD::RET_FLAG: |
| 252 | return "NVPTXISD::RET_FLAG"; |
| 253 | case NVPTXISD::Wrapper: |
| 254 | return "NVPTXISD::Wrapper"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 255 | case NVPTXISD::DeclareParam: |
| 256 | return "NVPTXISD::DeclareParam"; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 257 | case NVPTXISD::DeclareScalarParam: |
| 258 | return "NVPTXISD::DeclareScalarParam"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 259 | case NVPTXISD::DeclareRet: |
| 260 | return "NVPTXISD::DeclareRet"; |
| 261 | case NVPTXISD::DeclareRetParam: |
| 262 | return "NVPTXISD::DeclareRetParam"; |
| 263 | case NVPTXISD::PrintCall: |
| 264 | return "NVPTXISD::PrintCall"; |
| 265 | case NVPTXISD::LoadParam: |
| 266 | return "NVPTXISD::LoadParam"; |
Justin Holewinski | fe44314f | 2013-06-28 17:57:51 +0000 | [diff] [blame] | 267 | case NVPTXISD::LoadParamV2: |
| 268 | return "NVPTXISD::LoadParamV2"; |
| 269 | case NVPTXISD::LoadParamV4: |
| 270 | return "NVPTXISD::LoadParamV4"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 271 | case NVPTXISD::StoreParam: |
| 272 | return "NVPTXISD::StoreParam"; |
Justin Holewinski | fe44314f | 2013-06-28 17:57:51 +0000 | [diff] [blame] | 273 | case NVPTXISD::StoreParamV2: |
| 274 | return "NVPTXISD::StoreParamV2"; |
| 275 | case NVPTXISD::StoreParamV4: |
| 276 | return "NVPTXISD::StoreParamV4"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 277 | case NVPTXISD::StoreParamS32: |
| 278 | return "NVPTXISD::StoreParamS32"; |
| 279 | case NVPTXISD::StoreParamU32: |
| 280 | return "NVPTXISD::StoreParamU32"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 281 | case NVPTXISD::CallArgBegin: |
| 282 | return "NVPTXISD::CallArgBegin"; |
| 283 | case NVPTXISD::CallArg: |
| 284 | return "NVPTXISD::CallArg"; |
| 285 | case NVPTXISD::LastCallArg: |
| 286 | return "NVPTXISD::LastCallArg"; |
| 287 | case NVPTXISD::CallArgEnd: |
| 288 | return "NVPTXISD::CallArgEnd"; |
| 289 | case NVPTXISD::CallVoid: |
| 290 | return "NVPTXISD::CallVoid"; |
| 291 | case NVPTXISD::CallVal: |
| 292 | return "NVPTXISD::CallVal"; |
| 293 | case NVPTXISD::CallSymbol: |
| 294 | return "NVPTXISD::CallSymbol"; |
| 295 | case NVPTXISD::Prototype: |
| 296 | return "NVPTXISD::Prototype"; |
| 297 | case NVPTXISD::MoveParam: |
| 298 | return "NVPTXISD::MoveParam"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 299 | case NVPTXISD::StoreRetval: |
| 300 | return "NVPTXISD::StoreRetval"; |
Justin Holewinski | fe44314f | 2013-06-28 17:57:51 +0000 | [diff] [blame] | 301 | case NVPTXISD::StoreRetvalV2: |
| 302 | return "NVPTXISD::StoreRetvalV2"; |
| 303 | case NVPTXISD::StoreRetvalV4: |
| 304 | return "NVPTXISD::StoreRetvalV4"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 305 | case NVPTXISD::PseudoUseParam: |
| 306 | return "NVPTXISD::PseudoUseParam"; |
| 307 | case NVPTXISD::RETURN: |
| 308 | return "NVPTXISD::RETURN"; |
| 309 | case NVPTXISD::CallSeqBegin: |
| 310 | return "NVPTXISD::CallSeqBegin"; |
| 311 | case NVPTXISD::CallSeqEnd: |
| 312 | return "NVPTXISD::CallSeqEnd"; |
Justin Holewinski | 3d49e5c | 2013-11-15 12:30:04 +0000 | [diff] [blame] | 313 | case NVPTXISD::CallPrototype: |
| 314 | return "NVPTXISD::CallPrototype"; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 315 | case NVPTXISD::LoadV2: |
| 316 | return "NVPTXISD::LoadV2"; |
| 317 | case NVPTXISD::LoadV4: |
| 318 | return "NVPTXISD::LoadV4"; |
| 319 | case NVPTXISD::LDGV2: |
| 320 | return "NVPTXISD::LDGV2"; |
| 321 | case NVPTXISD::LDGV4: |
| 322 | return "NVPTXISD::LDGV4"; |
| 323 | case NVPTXISD::LDUV2: |
| 324 | return "NVPTXISD::LDUV2"; |
| 325 | case NVPTXISD::LDUV4: |
| 326 | return "NVPTXISD::LDUV4"; |
| 327 | case NVPTXISD::StoreV2: |
| 328 | return "NVPTXISD::StoreV2"; |
| 329 | case NVPTXISD::StoreV4: |
| 330 | return "NVPTXISD::StoreV4"; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 331 | case NVPTXISD::Tex1DFloatI32: return "NVPTXISD::Tex1DFloatI32"; |
| 332 | case NVPTXISD::Tex1DFloatFloat: return "NVPTXISD::Tex1DFloatFloat"; |
| 333 | case NVPTXISD::Tex1DFloatFloatLevel: |
| 334 | return "NVPTXISD::Tex1DFloatFloatLevel"; |
| 335 | case NVPTXISD::Tex1DFloatFloatGrad: |
| 336 | return "NVPTXISD::Tex1DFloatFloatGrad"; |
| 337 | case NVPTXISD::Tex1DI32I32: return "NVPTXISD::Tex1DI32I32"; |
| 338 | case NVPTXISD::Tex1DI32Float: return "NVPTXISD::Tex1DI32Float"; |
| 339 | case NVPTXISD::Tex1DI32FloatLevel: |
| 340 | return "NVPTXISD::Tex1DI32FloatLevel"; |
| 341 | case NVPTXISD::Tex1DI32FloatGrad: |
| 342 | return "NVPTXISD::Tex1DI32FloatGrad"; |
| 343 | case NVPTXISD::Tex1DArrayFloatI32: return "NVPTXISD::Tex2DArrayFloatI32"; |
| 344 | case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat"; |
| 345 | case NVPTXISD::Tex1DArrayFloatFloatLevel: |
| 346 | return "NVPTXISD::Tex2DArrayFloatFloatLevel"; |
| 347 | case NVPTXISD::Tex1DArrayFloatFloatGrad: |
| 348 | return "NVPTXISD::Tex2DArrayFloatFloatGrad"; |
| 349 | case NVPTXISD::Tex1DArrayI32I32: return "NVPTXISD::Tex2DArrayI32I32"; |
| 350 | case NVPTXISD::Tex1DArrayI32Float: return "NVPTXISD::Tex2DArrayI32Float"; |
| 351 | case NVPTXISD::Tex1DArrayI32FloatLevel: |
| 352 | return "NVPTXISD::Tex2DArrayI32FloatLevel"; |
| 353 | case NVPTXISD::Tex1DArrayI32FloatGrad: |
| 354 | return "NVPTXISD::Tex2DArrayI32FloatGrad"; |
| 355 | case NVPTXISD::Tex2DFloatI32: return "NVPTXISD::Tex2DFloatI32"; |
| 356 | case NVPTXISD::Tex2DFloatFloat: return "NVPTXISD::Tex2DFloatFloat"; |
| 357 | case NVPTXISD::Tex2DFloatFloatLevel: |
| 358 | return "NVPTXISD::Tex2DFloatFloatLevel"; |
| 359 | case NVPTXISD::Tex2DFloatFloatGrad: |
| 360 | return "NVPTXISD::Tex2DFloatFloatGrad"; |
| 361 | case NVPTXISD::Tex2DI32I32: return "NVPTXISD::Tex2DI32I32"; |
| 362 | case NVPTXISD::Tex2DI32Float: return "NVPTXISD::Tex2DI32Float"; |
| 363 | case NVPTXISD::Tex2DI32FloatLevel: |
| 364 | return "NVPTXISD::Tex2DI32FloatLevel"; |
| 365 | case NVPTXISD::Tex2DI32FloatGrad: |
| 366 | return "NVPTXISD::Tex2DI32FloatGrad"; |
| 367 | case NVPTXISD::Tex2DArrayFloatI32: return "NVPTXISD::Tex2DArrayFloatI32"; |
| 368 | case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat"; |
| 369 | case NVPTXISD::Tex2DArrayFloatFloatLevel: |
| 370 | return "NVPTXISD::Tex2DArrayFloatFloatLevel"; |
| 371 | case NVPTXISD::Tex2DArrayFloatFloatGrad: |
| 372 | return "NVPTXISD::Tex2DArrayFloatFloatGrad"; |
| 373 | case NVPTXISD::Tex2DArrayI32I32: return "NVPTXISD::Tex2DArrayI32I32"; |
| 374 | case NVPTXISD::Tex2DArrayI32Float: return "NVPTXISD::Tex2DArrayI32Float"; |
| 375 | case NVPTXISD::Tex2DArrayI32FloatLevel: |
| 376 | return "NVPTXISD::Tex2DArrayI32FloatLevel"; |
| 377 | case NVPTXISD::Tex2DArrayI32FloatGrad: |
| 378 | return "NVPTXISD::Tex2DArrayI32FloatGrad"; |
| 379 | case NVPTXISD::Tex3DFloatI32: return "NVPTXISD::Tex3DFloatI32"; |
| 380 | case NVPTXISD::Tex3DFloatFloat: return "NVPTXISD::Tex3DFloatFloat"; |
| 381 | case NVPTXISD::Tex3DFloatFloatLevel: |
| 382 | return "NVPTXISD::Tex3DFloatFloatLevel"; |
| 383 | case NVPTXISD::Tex3DFloatFloatGrad: |
| 384 | return "NVPTXISD::Tex3DFloatFloatGrad"; |
| 385 | case NVPTXISD::Tex3DI32I32: return "NVPTXISD::Tex3DI32I32"; |
| 386 | case NVPTXISD::Tex3DI32Float: return "NVPTXISD::Tex3DI32Float"; |
| 387 | case NVPTXISD::Tex3DI32FloatLevel: |
| 388 | return "NVPTXISD::Tex3DI32FloatLevel"; |
| 389 | case NVPTXISD::Tex3DI32FloatGrad: |
| 390 | return "NVPTXISD::Tex3DI32FloatGrad"; |
| 391 | |
| 392 | case NVPTXISD::Suld1DI8Trap: return "NVPTXISD::Suld1DI8Trap"; |
| 393 | case NVPTXISD::Suld1DI16Trap: return "NVPTXISD::Suld1DI16Trap"; |
| 394 | case NVPTXISD::Suld1DI32Trap: return "NVPTXISD::Suld1DI32Trap"; |
| 395 | case NVPTXISD::Suld1DV2I8Trap: return "NVPTXISD::Suld1DV2I8Trap"; |
| 396 | case NVPTXISD::Suld1DV2I16Trap: return "NVPTXISD::Suld1DV2I16Trap"; |
| 397 | case NVPTXISD::Suld1DV2I32Trap: return "NVPTXISD::Suld1DV2I32Trap"; |
| 398 | case NVPTXISD::Suld1DV4I8Trap: return "NVPTXISD::Suld1DV4I8Trap"; |
| 399 | case NVPTXISD::Suld1DV4I16Trap: return "NVPTXISD::Suld1DV4I16Trap"; |
| 400 | case NVPTXISD::Suld1DV4I32Trap: return "NVPTXISD::Suld1DV4I32Trap"; |
| 401 | |
| 402 | case NVPTXISD::Suld1DArrayI8Trap: return "NVPTXISD::Suld1DArrayI8Trap"; |
| 403 | case NVPTXISD::Suld1DArrayI16Trap: return "NVPTXISD::Suld1DArrayI16Trap"; |
| 404 | case NVPTXISD::Suld1DArrayI32Trap: return "NVPTXISD::Suld1DArrayI32Trap"; |
| 405 | case NVPTXISD::Suld1DArrayV2I8Trap: return "NVPTXISD::Suld1DArrayV2I8Trap"; |
| 406 | case NVPTXISD::Suld1DArrayV2I16Trap: return "NVPTXISD::Suld1DArrayV2I16Trap"; |
| 407 | case NVPTXISD::Suld1DArrayV2I32Trap: return "NVPTXISD::Suld1DArrayV2I32Trap"; |
| 408 | case NVPTXISD::Suld1DArrayV4I8Trap: return "NVPTXISD::Suld1DArrayV4I8Trap"; |
| 409 | case NVPTXISD::Suld1DArrayV4I16Trap: return "NVPTXISD::Suld1DArrayV4I16Trap"; |
| 410 | case NVPTXISD::Suld1DArrayV4I32Trap: return "NVPTXISD::Suld1DArrayV4I32Trap"; |
| 411 | |
| 412 | case NVPTXISD::Suld2DI8Trap: return "NVPTXISD::Suld2DI8Trap"; |
| 413 | case NVPTXISD::Suld2DI16Trap: return "NVPTXISD::Suld2DI16Trap"; |
| 414 | case NVPTXISD::Suld2DI32Trap: return "NVPTXISD::Suld2DI32Trap"; |
| 415 | case NVPTXISD::Suld2DV2I8Trap: return "NVPTXISD::Suld2DV2I8Trap"; |
| 416 | case NVPTXISD::Suld2DV2I16Trap: return "NVPTXISD::Suld2DV2I16Trap"; |
| 417 | case NVPTXISD::Suld2DV2I32Trap: return "NVPTXISD::Suld2DV2I32Trap"; |
| 418 | case NVPTXISD::Suld2DV4I8Trap: return "NVPTXISD::Suld2DV4I8Trap"; |
| 419 | case NVPTXISD::Suld2DV4I16Trap: return "NVPTXISD::Suld2DV4I16Trap"; |
| 420 | case NVPTXISD::Suld2DV4I32Trap: return "NVPTXISD::Suld2DV4I32Trap"; |
| 421 | |
| 422 | case NVPTXISD::Suld2DArrayI8Trap: return "NVPTXISD::Suld2DArrayI8Trap"; |
| 423 | case NVPTXISD::Suld2DArrayI16Trap: return "NVPTXISD::Suld2DArrayI16Trap"; |
| 424 | case NVPTXISD::Suld2DArrayI32Trap: return "NVPTXISD::Suld2DArrayI32Trap"; |
| 425 | case NVPTXISD::Suld2DArrayV2I8Trap: return "NVPTXISD::Suld2DArrayV2I8Trap"; |
| 426 | case NVPTXISD::Suld2DArrayV2I16Trap: return "NVPTXISD::Suld2DArrayV2I16Trap"; |
| 427 | case NVPTXISD::Suld2DArrayV2I32Trap: return "NVPTXISD::Suld2DArrayV2I32Trap"; |
| 428 | case NVPTXISD::Suld2DArrayV4I8Trap: return "NVPTXISD::Suld2DArrayV4I8Trap"; |
| 429 | case NVPTXISD::Suld2DArrayV4I16Trap: return "NVPTXISD::Suld2DArrayV4I16Trap"; |
| 430 | case NVPTXISD::Suld2DArrayV4I32Trap: return "NVPTXISD::Suld2DArrayV4I32Trap"; |
| 431 | |
| 432 | case NVPTXISD::Suld3DI8Trap: return "NVPTXISD::Suld3DI8Trap"; |
| 433 | case NVPTXISD::Suld3DI16Trap: return "NVPTXISD::Suld3DI16Trap"; |
| 434 | case NVPTXISD::Suld3DI32Trap: return "NVPTXISD::Suld3DI32Trap"; |
| 435 | case NVPTXISD::Suld3DV2I8Trap: return "NVPTXISD::Suld3DV2I8Trap"; |
| 436 | case NVPTXISD::Suld3DV2I16Trap: return "NVPTXISD::Suld3DV2I16Trap"; |
| 437 | case NVPTXISD::Suld3DV2I32Trap: return "NVPTXISD::Suld3DV2I32Trap"; |
| 438 | case NVPTXISD::Suld3DV4I8Trap: return "NVPTXISD::Suld3DV4I8Trap"; |
| 439 | case NVPTXISD::Suld3DV4I16Trap: return "NVPTXISD::Suld3DV4I16Trap"; |
| 440 | case NVPTXISD::Suld3DV4I32Trap: return "NVPTXISD::Suld3DV4I32Trap"; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 441 | } |
| 442 | } |
| 443 | |
Matt Arsenault | f751d62 | 2014-03-31 20:54:58 +0000 | [diff] [blame] | 444 | bool NVPTXTargetLowering::shouldSplitVectorType(EVT VT) const { |
| 445 | return VT.getScalarType() == MVT::i1; |
Justin Holewinski | bc45119 | 2012-11-29 14:26:24 +0000 | [diff] [blame] | 446 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 447 | |
| 448 | SDValue |
| 449 | NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const { |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 450 | SDLoc dl(Op); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 451 | const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); |
| 452 | Op = DAG.getTargetGlobalAddress(GV, dl, getPointerTy()); |
| 453 | return DAG.getNode(NVPTXISD::Wrapper, dl, getPointerTy(), Op); |
| 454 | } |
| 455 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 456 | std::string |
| 457 | NVPTXTargetLowering::getPrototype(Type *retTy, const ArgListTy &Args, |
| 458 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 459 | unsigned retAlignment, |
| 460 | const ImmutableCallSite *CS) const { |
| 461 | |
| 462 | bool isABI = (nvptxSubtarget.getSmVersion() >= 20); |
| 463 | assert(isABI && "Non-ABI compilation is not supported"); |
| 464 | if (!isABI) |
| 465 | return ""; |
| 466 | |
| 467 | std::stringstream O; |
| 468 | O << "prototype_" << uniqueCallSite << " : .callprototype "; |
| 469 | |
| 470 | if (retTy->getTypeID() == Type::VoidTyID) { |
| 471 | O << "()"; |
| 472 | } else { |
| 473 | O << "("; |
Rafael Espindola | 0801334 | 2013-12-07 19:34:20 +0000 | [diff] [blame] | 474 | if (retTy->isFloatingPointTy() || retTy->isIntegerTy()) { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 475 | unsigned size = 0; |
| 476 | if (const IntegerType *ITy = dyn_cast<IntegerType>(retTy)) { |
| 477 | size = ITy->getBitWidth(); |
| 478 | if (size < 32) |
| 479 | size = 32; |
| 480 | } else { |
| 481 | assert(retTy->isFloatingPointTy() && |
| 482 | "Floating point type expected here"); |
| 483 | size = retTy->getPrimitiveSizeInBits(); |
| 484 | } |
| 485 | |
| 486 | O << ".param .b" << size << " _"; |
| 487 | } else if (isa<PointerType>(retTy)) { |
| 488 | O << ".param .b" << getPointerTy().getSizeInBits() << " _"; |
| 489 | } else { |
| 490 | if ((retTy->getTypeID() == Type::StructTyID) || isa<VectorType>(retTy)) { |
| 491 | SmallVector<EVT, 16> vtparts; |
| 492 | ComputeValueVTs(*this, retTy, vtparts); |
| 493 | unsigned totalsz = 0; |
| 494 | for (unsigned i = 0, e = vtparts.size(); i != e; ++i) { |
| 495 | unsigned elems = 1; |
| 496 | EVT elemtype = vtparts[i]; |
| 497 | if (vtparts[i].isVector()) { |
| 498 | elems = vtparts[i].getVectorNumElements(); |
| 499 | elemtype = vtparts[i].getVectorElementType(); |
| 500 | } |
| 501 | // TODO: no need to loop |
| 502 | for (unsigned j = 0, je = elems; j != je; ++j) { |
| 503 | unsigned sz = elemtype.getSizeInBits(); |
| 504 | if (elemtype.isInteger() && (sz < 8)) |
| 505 | sz = 8; |
| 506 | totalsz += sz / 8; |
| 507 | } |
| 508 | } |
| 509 | O << ".param .align " << retAlignment << " .b8 _[" << totalsz << "]"; |
| 510 | } else { |
| 511 | assert(false && "Unknown return type"); |
| 512 | } |
| 513 | } |
| 514 | O << ") "; |
| 515 | } |
| 516 | O << "_ ("; |
| 517 | |
| 518 | bool first = true; |
| 519 | MVT thePointerTy = getPointerTy(); |
| 520 | |
| 521 | unsigned OIdx = 0; |
| 522 | for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) { |
| 523 | Type *Ty = Args[i].Ty; |
| 524 | if (!first) { |
| 525 | O << ", "; |
| 526 | } |
| 527 | first = false; |
| 528 | |
| 529 | if (Outs[OIdx].Flags.isByVal() == false) { |
| 530 | if (Ty->isAggregateType() || Ty->isVectorTy()) { |
| 531 | unsigned align = 0; |
| 532 | const CallInst *CallI = cast<CallInst>(CS->getInstruction()); |
| 533 | const DataLayout *TD = getDataLayout(); |
| 534 | // +1 because index 0 is reserved for return type alignment |
| 535 | if (!llvm::getAlign(*CallI, i + 1, align)) |
| 536 | align = TD->getABITypeAlignment(Ty); |
| 537 | unsigned sz = TD->getTypeAllocSize(Ty); |
| 538 | O << ".param .align " << align << " .b8 "; |
| 539 | O << "_"; |
| 540 | O << "[" << sz << "]"; |
| 541 | // update the index for Outs |
| 542 | SmallVector<EVT, 16> vtparts; |
| 543 | ComputeValueVTs(*this, Ty, vtparts); |
| 544 | if (unsigned len = vtparts.size()) |
| 545 | OIdx += len - 1; |
| 546 | continue; |
| 547 | } |
Justin Holewinski | dff28d2 | 2013-07-01 12:59:01 +0000 | [diff] [blame] | 548 | // i8 types in IR will be i16 types in SDAG |
| 549 | assert((getValueType(Ty) == Outs[OIdx].VT || |
| 550 | (getValueType(Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) && |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 551 | "type mismatch between callee prototype and arguments"); |
| 552 | // scalar type |
| 553 | unsigned sz = 0; |
| 554 | if (isa<IntegerType>(Ty)) { |
| 555 | sz = cast<IntegerType>(Ty)->getBitWidth(); |
| 556 | if (sz < 32) |
| 557 | sz = 32; |
| 558 | } else if (isa<PointerType>(Ty)) |
| 559 | sz = thePointerTy.getSizeInBits(); |
| 560 | else |
| 561 | sz = Ty->getPrimitiveSizeInBits(); |
| 562 | O << ".param .b" << sz << " "; |
| 563 | O << "_"; |
| 564 | continue; |
| 565 | } |
| 566 | const PointerType *PTy = dyn_cast<PointerType>(Ty); |
| 567 | assert(PTy && "Param with byval attribute should be a pointer type"); |
| 568 | Type *ETy = PTy->getElementType(); |
| 569 | |
| 570 | unsigned align = Outs[OIdx].Flags.getByValAlign(); |
| 571 | unsigned sz = getDataLayout()->getTypeAllocSize(ETy); |
| 572 | O << ".param .align " << align << " .b8 "; |
| 573 | O << "_"; |
| 574 | O << "[" << sz << "]"; |
| 575 | } |
| 576 | O << ");"; |
| 577 | return O.str(); |
| 578 | } |
| 579 | |
| 580 | unsigned |
| 581 | NVPTXTargetLowering::getArgumentAlignment(SDValue Callee, |
| 582 | const ImmutableCallSite *CS, |
| 583 | Type *Ty, |
| 584 | unsigned Idx) const { |
| 585 | const DataLayout *TD = getDataLayout(); |
Justin Holewinski | 124e93d | 2013-11-11 19:28:19 +0000 | [diff] [blame] | 586 | unsigned Align = 0; |
| 587 | const Value *DirectCallee = CS->getCalledFunction(); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 588 | |
Justin Holewinski | 124e93d | 2013-11-11 19:28:19 +0000 | [diff] [blame] | 589 | if (!DirectCallee) { |
| 590 | // We don't have a direct function symbol, but that may be because of |
| 591 | // constant cast instructions in the call. |
| 592 | const Instruction *CalleeI = CS->getInstruction(); |
| 593 | assert(CalleeI && "Call target is not a function or derived value?"); |
| 594 | |
| 595 | // With bitcast'd call targets, the instruction will be the call |
| 596 | if (isa<CallInst>(CalleeI)) { |
| 597 | // Check if we have call alignment metadata |
| 598 | if (llvm::getAlign(*cast<CallInst>(CalleeI), Idx, Align)) |
| 599 | return Align; |
| 600 | |
| 601 | const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue(); |
| 602 | // Ignore any bitcast instructions |
| 603 | while(isa<ConstantExpr>(CalleeV)) { |
| 604 | const ConstantExpr *CE = cast<ConstantExpr>(CalleeV); |
| 605 | if (!CE->isCast()) |
| 606 | break; |
| 607 | // Look through the bitcast |
| 608 | CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0); |
| 609 | } |
| 610 | |
| 611 | // We have now looked past all of the bitcasts. Do we finally have a |
| 612 | // Function? |
| 613 | if (isa<Function>(CalleeV)) |
| 614 | DirectCallee = CalleeV; |
| 615 | } |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 616 | } |
| 617 | |
Justin Holewinski | 124e93d | 2013-11-11 19:28:19 +0000 | [diff] [blame] | 618 | // Check for function alignment information if we found that the |
| 619 | // ultimate target is a Function |
| 620 | if (DirectCallee) |
| 621 | if (llvm::getAlign(*cast<Function>(DirectCallee), Idx, Align)) |
| 622 | return Align; |
| 623 | |
| 624 | // Call is indirect or alignment information is not available, fall back to |
| 625 | // the ABI type alignment |
| 626 | return TD->getABITypeAlignment(Ty); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 627 | } |
| 628 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 629 | SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
| 630 | SmallVectorImpl<SDValue> &InVals) const { |
| 631 | SelectionDAG &DAG = CLI.DAG; |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 632 | SDLoc dl = CLI.DL; |
Craig Topper | b94011f | 2013-07-14 04:42:23 +0000 | [diff] [blame] | 633 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
| 634 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
| 635 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 636 | SDValue Chain = CLI.Chain; |
| 637 | SDValue Callee = CLI.Callee; |
| 638 | bool &isTailCall = CLI.IsTailCall; |
| 639 | ArgListTy &Args = CLI.Args; |
| 640 | Type *retTy = CLI.RetTy; |
| 641 | ImmutableCallSite *CS = CLI.CS; |
Justin Holewinski | aa58397 | 2012-05-25 16:35:28 +0000 | [diff] [blame] | 642 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 643 | bool isABI = (nvptxSubtarget.getSmVersion() >= 20); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 644 | assert(isABI && "Non-ABI compilation is not supported"); |
| 645 | if (!isABI) |
| 646 | return Chain; |
| 647 | const DataLayout *TD = getDataLayout(); |
| 648 | MachineFunction &MF = DAG.getMachineFunction(); |
| 649 | const Function *F = MF.getFunction(); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 650 | |
| 651 | SDValue tempChain = Chain; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 652 | Chain = |
| 653 | DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(uniqueCallSite, true), |
| 654 | dl); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 655 | SDValue InFlag = Chain.getValue(1); |
| 656 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 657 | unsigned paramCount = 0; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 658 | // Args.size() and Outs.size() need not match. |
| 659 | // Outs.size() will be larger |
| 660 | // * if there is an aggregate argument with multiple fields (each field |
| 661 | // showing up separately in Outs) |
| 662 | // * if there is a vector argument with more than typical vector-length |
| 663 | // elements (generally if more than 4) where each vector element is |
| 664 | // individually present in Outs. |
| 665 | // So a different index should be used for indexing into Outs/OutVals. |
| 666 | // See similar issue in LowerFormalArguments. |
| 667 | unsigned OIdx = 0; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 668 | // Declare the .params or .reg need to pass values |
| 669 | // to the function |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 670 | for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) { |
| 671 | EVT VT = Outs[OIdx].VT; |
| 672 | Type *Ty = Args[i].Ty; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 673 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 674 | if (Outs[OIdx].Flags.isByVal() == false) { |
| 675 | if (Ty->isAggregateType()) { |
| 676 | // aggregate |
| 677 | SmallVector<EVT, 16> vtparts; |
| 678 | ComputeValueVTs(*this, Ty, vtparts); |
| 679 | |
| 680 | unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1); |
| 681 | // declare .param .align <align> .b8 .param<n>[<size>]; |
| 682 | unsigned sz = TD->getTypeAllocSize(Ty); |
| 683 | SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 684 | SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32), |
| 685 | DAG.getConstant(paramCount, MVT::i32), |
| 686 | DAG.getConstant(sz, MVT::i32), InFlag }; |
| 687 | Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 688 | DeclareParamOps); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 689 | InFlag = Chain.getValue(1); |
| 690 | unsigned curOffset = 0; |
| 691 | for (unsigned j = 0, je = vtparts.size(); j != je; ++j) { |
| 692 | unsigned elems = 1; |
| 693 | EVT elemtype = vtparts[j]; |
| 694 | if (vtparts[j].isVector()) { |
| 695 | elems = vtparts[j].getVectorNumElements(); |
| 696 | elemtype = vtparts[j].getVectorElementType(); |
| 697 | } |
| 698 | for (unsigned k = 0, ke = elems; k != ke; ++k) { |
| 699 | unsigned sz = elemtype.getSizeInBits(); |
| 700 | if (elemtype.isInteger() && (sz < 8)) |
| 701 | sz = 8; |
| 702 | SDValue StVal = OutVals[OIdx]; |
| 703 | if (elemtype.getSizeInBits() < 16) { |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 704 | StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 705 | } |
| 706 | SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 707 | SDValue CopyParamOps[] = { Chain, |
| 708 | DAG.getConstant(paramCount, MVT::i32), |
| 709 | DAG.getConstant(curOffset, MVT::i32), |
| 710 | StVal, InFlag }; |
| 711 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 712 | CopyParamVTs, CopyParamOps, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 713 | elemtype, MachinePointerInfo()); |
| 714 | InFlag = Chain.getValue(1); |
| 715 | curOffset += sz / 8; |
| 716 | ++OIdx; |
| 717 | } |
| 718 | } |
| 719 | if (vtparts.size() > 0) |
| 720 | --OIdx; |
| 721 | ++paramCount; |
| 722 | continue; |
| 723 | } |
| 724 | if (Ty->isVectorTy()) { |
| 725 | EVT ObjectVT = getValueType(Ty); |
| 726 | unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1); |
| 727 | // declare .param .align <align> .b8 .param<n>[<size>]; |
| 728 | unsigned sz = TD->getTypeAllocSize(Ty); |
| 729 | SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 730 | SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32), |
| 731 | DAG.getConstant(paramCount, MVT::i32), |
| 732 | DAG.getConstant(sz, MVT::i32), InFlag }; |
| 733 | Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 734 | DeclareParamOps); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 735 | InFlag = Chain.getValue(1); |
| 736 | unsigned NumElts = ObjectVT.getVectorNumElements(); |
| 737 | EVT EltVT = ObjectVT.getVectorElementType(); |
| 738 | EVT MemVT = EltVT; |
| 739 | bool NeedExtend = false; |
| 740 | if (EltVT.getSizeInBits() < 16) { |
| 741 | NeedExtend = true; |
| 742 | EltVT = MVT::i16; |
| 743 | } |
| 744 | |
| 745 | // V1 store |
| 746 | if (NumElts == 1) { |
| 747 | SDValue Elt = OutVals[OIdx++]; |
| 748 | if (NeedExtend) |
| 749 | Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt); |
| 750 | |
| 751 | SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 752 | SDValue CopyParamOps[] = { Chain, |
| 753 | DAG.getConstant(paramCount, MVT::i32), |
| 754 | DAG.getConstant(0, MVT::i32), Elt, |
| 755 | InFlag }; |
| 756 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 757 | CopyParamVTs, CopyParamOps, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 758 | MemVT, MachinePointerInfo()); |
| 759 | InFlag = Chain.getValue(1); |
| 760 | } else if (NumElts == 2) { |
| 761 | SDValue Elt0 = OutVals[OIdx++]; |
| 762 | SDValue Elt1 = OutVals[OIdx++]; |
| 763 | if (NeedExtend) { |
| 764 | Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0); |
| 765 | Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1); |
| 766 | } |
| 767 | |
| 768 | SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 769 | SDValue CopyParamOps[] = { Chain, |
| 770 | DAG.getConstant(paramCount, MVT::i32), |
| 771 | DAG.getConstant(0, MVT::i32), Elt0, Elt1, |
| 772 | InFlag }; |
| 773 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 774 | CopyParamVTs, CopyParamOps, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 775 | MemVT, MachinePointerInfo()); |
| 776 | InFlag = Chain.getValue(1); |
| 777 | } else { |
| 778 | unsigned curOffset = 0; |
| 779 | // V4 stores |
| 780 | // We have at least 4 elements (<3 x Ty> expands to 4 elements) and |
| 781 | // the |
| 782 | // vector will be expanded to a power of 2 elements, so we know we can |
| 783 | // always round up to the next multiple of 4 when creating the vector |
| 784 | // stores. |
| 785 | // e.g. 4 elem => 1 st.v4 |
| 786 | // 6 elem => 2 st.v4 |
| 787 | // 8 elem => 2 st.v4 |
| 788 | // 11 elem => 3 st.v4 |
| 789 | unsigned VecSize = 4; |
| 790 | if (EltVT.getSizeInBits() == 64) |
| 791 | VecSize = 2; |
| 792 | |
| 793 | // This is potentially only part of a vector, so assume all elements |
| 794 | // are packed together. |
| 795 | unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize; |
| 796 | |
| 797 | for (unsigned i = 0; i < NumElts; i += VecSize) { |
| 798 | // Get values |
| 799 | SDValue StoreVal; |
| 800 | SmallVector<SDValue, 8> Ops; |
| 801 | Ops.push_back(Chain); |
| 802 | Ops.push_back(DAG.getConstant(paramCount, MVT::i32)); |
| 803 | Ops.push_back(DAG.getConstant(curOffset, MVT::i32)); |
| 804 | |
| 805 | unsigned Opc = NVPTXISD::StoreParamV2; |
| 806 | |
| 807 | StoreVal = OutVals[OIdx++]; |
| 808 | if (NeedExtend) |
| 809 | StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal); |
| 810 | Ops.push_back(StoreVal); |
| 811 | |
| 812 | if (i + 1 < NumElts) { |
| 813 | StoreVal = OutVals[OIdx++]; |
| 814 | if (NeedExtend) |
| 815 | StoreVal = |
| 816 | DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal); |
| 817 | } else { |
| 818 | StoreVal = DAG.getUNDEF(EltVT); |
| 819 | } |
| 820 | Ops.push_back(StoreVal); |
| 821 | |
| 822 | if (VecSize == 4) { |
| 823 | Opc = NVPTXISD::StoreParamV4; |
| 824 | if (i + 2 < NumElts) { |
| 825 | StoreVal = OutVals[OIdx++]; |
| 826 | if (NeedExtend) |
| 827 | StoreVal = |
| 828 | DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal); |
| 829 | } else { |
| 830 | StoreVal = DAG.getUNDEF(EltVT); |
| 831 | } |
| 832 | Ops.push_back(StoreVal); |
| 833 | |
| 834 | if (i + 3 < NumElts) { |
| 835 | StoreVal = OutVals[OIdx++]; |
| 836 | if (NeedExtend) |
| 837 | StoreVal = |
| 838 | DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal); |
| 839 | } else { |
| 840 | StoreVal = DAG.getUNDEF(EltVT); |
| 841 | } |
| 842 | Ops.push_back(StoreVal); |
| 843 | } |
| 844 | |
Justin Holewinski | dff28d2 | 2013-07-01 12:59:01 +0000 | [diff] [blame] | 845 | Ops.push_back(InFlag); |
| 846 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 847 | SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 848 | Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops, |
| 849 | MemVT, MachinePointerInfo()); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 850 | InFlag = Chain.getValue(1); |
| 851 | curOffset += PerStoreOffset; |
| 852 | } |
| 853 | } |
| 854 | ++paramCount; |
| 855 | --OIdx; |
| 856 | continue; |
| 857 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 858 | // Plain scalar |
| 859 | // for ABI, declare .param .b<size> .param<n>; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 860 | unsigned sz = VT.getSizeInBits(); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 861 | bool needExtend = false; |
| 862 | if (VT.isInteger()) { |
| 863 | if (sz < 16) |
| 864 | needExtend = true; |
| 865 | if (sz < 32) |
| 866 | sz = 32; |
| 867 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 868 | SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 869 | SDValue DeclareParamOps[] = { Chain, |
| 870 | DAG.getConstant(paramCount, MVT::i32), |
| 871 | DAG.getConstant(sz, MVT::i32), |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 872 | DAG.getConstant(0, MVT::i32), InFlag }; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 873 | Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 874 | DeclareParamOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 875 | InFlag = Chain.getValue(1); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 876 | SDValue OutV = OutVals[OIdx]; |
| 877 | if (needExtend) { |
| 878 | // zext/sext i1 to i16 |
| 879 | unsigned opc = ISD::ZERO_EXTEND; |
| 880 | if (Outs[OIdx].Flags.isSExt()) |
| 881 | opc = ISD::SIGN_EXTEND; |
| 882 | OutV = DAG.getNode(opc, dl, MVT::i16, OutV); |
| 883 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 884 | SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 885 | SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32), |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 886 | DAG.getConstant(0, MVT::i32), OutV, InFlag }; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 887 | |
| 888 | unsigned opcode = NVPTXISD::StoreParam; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 889 | if (Outs[OIdx].Flags.isZExt()) |
| 890 | opcode = NVPTXISD::StoreParamU32; |
| 891 | else if (Outs[OIdx].Flags.isSExt()) |
| 892 | opcode = NVPTXISD::StoreParamS32; |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 893 | Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 894 | VT, MachinePointerInfo()); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 895 | |
| 896 | InFlag = Chain.getValue(1); |
| 897 | ++paramCount; |
| 898 | continue; |
| 899 | } |
| 900 | // struct or vector |
| 901 | SmallVector<EVT, 16> vtparts; |
| 902 | const PointerType *PTy = dyn_cast<PointerType>(Args[i].Ty); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 903 | assert(PTy && "Type of a byval parameter should be pointer"); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 904 | ComputeValueVTs(*this, PTy->getElementType(), vtparts); |
| 905 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 906 | // declare .param .align <align> .b8 .param<n>[<size>]; |
| 907 | unsigned sz = Outs[OIdx].Flags.getByValSize(); |
| 908 | SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 909 | // The ByValAlign in the Outs[OIdx].Flags is alway set at this point, |
| 910 | // so we don't need to worry about natural alignment or not. |
| 911 | // See TargetLowering::LowerCallTo(). |
| 912 | SDValue DeclareParamOps[] = { |
| 913 | Chain, DAG.getConstant(Outs[OIdx].Flags.getByValAlign(), MVT::i32), |
| 914 | DAG.getConstant(paramCount, MVT::i32), DAG.getConstant(sz, MVT::i32), |
| 915 | InFlag |
| 916 | }; |
| 917 | Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 918 | DeclareParamOps); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 919 | InFlag = Chain.getValue(1); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 920 | unsigned curOffset = 0; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 921 | for (unsigned j = 0, je = vtparts.size(); j != je; ++j) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 922 | unsigned elems = 1; |
| 923 | EVT elemtype = vtparts[j]; |
| 924 | if (vtparts[j].isVector()) { |
| 925 | elems = vtparts[j].getVectorNumElements(); |
| 926 | elemtype = vtparts[j].getVectorElementType(); |
| 927 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 928 | for (unsigned k = 0, ke = elems; k != ke; ++k) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 929 | unsigned sz = elemtype.getSizeInBits(); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 930 | if (elemtype.isInteger() && (sz < 8)) |
| 931 | sz = 8; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 932 | SDValue srcAddr = |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 933 | DAG.getNode(ISD::ADD, dl, getPointerTy(), OutVals[OIdx], |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 934 | DAG.getConstant(curOffset, getPointerTy())); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 935 | SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr, |
| 936 | MachinePointerInfo(), false, false, false, |
| 937 | 0); |
| 938 | if (elemtype.getSizeInBits() < 16) { |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 939 | theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 940 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 941 | SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 942 | SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32), |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 943 | DAG.getConstant(curOffset, MVT::i32), theVal, |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 944 | InFlag }; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 945 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 946 | CopyParamOps, elemtype, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 947 | MachinePointerInfo()); |
| 948 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 949 | InFlag = Chain.getValue(1); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 950 | curOffset += sz / 8; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 951 | } |
| 952 | } |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 953 | ++paramCount; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 954 | } |
| 955 | |
| 956 | GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode()); |
| 957 | unsigned retAlignment = 0; |
| 958 | |
| 959 | // Handle Result |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 960 | if (Ins.size() > 0) { |
| 961 | SmallVector<EVT, 16> resvtparts; |
| 962 | ComputeValueVTs(*this, retTy, resvtparts); |
| 963 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 964 | // Declare |
| 965 | // .param .align 16 .b8 retval0[<size-in-bytes>], or |
| 966 | // .param .b<size-in-bits> retval0 |
| 967 | unsigned resultsz = TD->getTypeAllocSizeInBits(retTy); |
Rafael Espindola | 0801334 | 2013-12-07 19:34:20 +0000 | [diff] [blame] | 968 | if (retTy->isSingleValueType()) { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 969 | // Scalar needs to be at least 32bit wide |
| 970 | if (resultsz < 32) |
| 971 | resultsz = 32; |
| 972 | SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 973 | SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, MVT::i32), |
| 974 | DAG.getConstant(resultsz, MVT::i32), |
| 975 | DAG.getConstant(0, MVT::i32), InFlag }; |
| 976 | Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 977 | DeclareRetOps); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 978 | InFlag = Chain.getValue(1); |
| 979 | } else { |
| 980 | retAlignment = getArgumentAlignment(Callee, CS, retTy, 0); |
| 981 | SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 982 | SDValue DeclareRetOps[] = { Chain, |
| 983 | DAG.getConstant(retAlignment, MVT::i32), |
| 984 | DAG.getConstant(resultsz / 8, MVT::i32), |
| 985 | DAG.getConstant(0, MVT::i32), InFlag }; |
| 986 | Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 987 | DeclareRetOps); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 988 | InFlag = Chain.getValue(1); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 989 | } |
| 990 | } |
| 991 | |
| 992 | if (!Func) { |
| 993 | // This is indirect function call case : PTX requires a prototype of the |
| 994 | // form |
| 995 | // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _); |
| 996 | // to be emitted, and the label has to used as the last arg of call |
| 997 | // instruction. |
Justin Holewinski | 3d49e5c | 2013-11-15 12:30:04 +0000 | [diff] [blame] | 998 | // The prototype is embedded in a string and put as the operand for a |
| 999 | // CallPrototype SDNode which will print out to the value of the string. |
| 1000 | SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 1001 | std::string Proto = getPrototype(retTy, Args, Outs, retAlignment, CS); |
| 1002 | const char *ProtoStr = |
| 1003 | nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str(); |
| 1004 | SDValue ProtoOps[] = { |
| 1005 | Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag, |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1006 | }; |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1007 | Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1008 | InFlag = Chain.getValue(1); |
| 1009 | } |
| 1010 | // Op to just print "call" |
| 1011 | SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1012 | SDValue PrintCallOps[] = { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1013 | Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, MVT::i32), InFlag |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1014 | }; |
| 1015 | Chain = DAG.getNode(Func ? (NVPTXISD::PrintCallUni) : (NVPTXISD::PrintCall), |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1016 | dl, PrintCallVTs, PrintCallOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1017 | InFlag = Chain.getValue(1); |
| 1018 | |
| 1019 | // Ops to print out the function name |
| 1020 | SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 1021 | SDValue CallVoidOps[] = { Chain, Callee, InFlag }; |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1022 | Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1023 | InFlag = Chain.getValue(1); |
| 1024 | |
| 1025 | // Ops to print out the param list |
| 1026 | SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 1027 | SDValue CallArgBeginOps[] = { Chain, InFlag }; |
| 1028 | Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs, |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1029 | CallArgBeginOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1030 | InFlag = Chain.getValue(1); |
| 1031 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1032 | for (unsigned i = 0, e = paramCount; i != e; ++i) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1033 | unsigned opcode; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1034 | if (i == (e - 1)) |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1035 | opcode = NVPTXISD::LastCallArg; |
| 1036 | else |
| 1037 | opcode = NVPTXISD::CallArg; |
| 1038 | SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
| 1039 | SDValue CallArgOps[] = { Chain, DAG.getConstant(1, MVT::i32), |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1040 | DAG.getConstant(i, MVT::i32), InFlag }; |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1041 | Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1042 | InFlag = Chain.getValue(1); |
| 1043 | } |
| 1044 | SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1045 | SDValue CallArgEndOps[] = { Chain, DAG.getConstant(Func ? 1 : 0, MVT::i32), |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1046 | InFlag }; |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1047 | Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1048 | InFlag = Chain.getValue(1); |
| 1049 | |
| 1050 | if (!Func) { |
| 1051 | SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1052 | SDValue PrototypeOps[] = { Chain, DAG.getConstant(uniqueCallSite, MVT::i32), |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1053 | InFlag }; |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1054 | Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1055 | InFlag = Chain.getValue(1); |
| 1056 | } |
| 1057 | |
| 1058 | // Generate loads from param memory/moves from registers for result |
| 1059 | if (Ins.size() > 0) { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1060 | unsigned resoffset = 0; |
| 1061 | if (retTy && retTy->isVectorTy()) { |
| 1062 | EVT ObjectVT = getValueType(retTy); |
| 1063 | unsigned NumElts = ObjectVT.getVectorNumElements(); |
| 1064 | EVT EltVT = ObjectVT.getVectorElementType(); |
Benjamin Kramer | 3cc579a | 2013-06-29 22:51:12 +0000 | [diff] [blame] | 1065 | assert(nvTM->getTargetLowering()->getNumRegisters(F->getContext(), |
| 1066 | ObjectVT) == NumElts && |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1067 | "Vector was not scalarized"); |
| 1068 | unsigned sz = EltVT.getSizeInBits(); |
| 1069 | bool needTruncate = sz < 16 ? true : false; |
| 1070 | |
| 1071 | if (NumElts == 1) { |
| 1072 | // Just a simple load |
Craig Topper | 59f626d | 2014-04-26 19:29:47 +0000 | [diff] [blame] | 1073 | SmallVector<EVT, 4> LoadRetVTs; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1074 | if (needTruncate) { |
| 1075 | // If loading i1 result, generate |
| 1076 | // load i16 |
| 1077 | // trunc i16 to i1 |
| 1078 | LoadRetVTs.push_back(MVT::i16); |
| 1079 | } else |
| 1080 | LoadRetVTs.push_back(EltVT); |
| 1081 | LoadRetVTs.push_back(MVT::Other); |
| 1082 | LoadRetVTs.push_back(MVT::Glue); |
Craig Topper | 59f626d | 2014-04-26 19:29:47 +0000 | [diff] [blame] | 1083 | SmallVector<SDValue, 4> LoadRetOps; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1084 | LoadRetOps.push_back(Chain); |
| 1085 | LoadRetOps.push_back(DAG.getConstant(1, MVT::i32)); |
| 1086 | LoadRetOps.push_back(DAG.getConstant(0, MVT::i32)); |
| 1087 | LoadRetOps.push_back(InFlag); |
| 1088 | SDValue retval = DAG.getMemIntrinsicNode( |
| 1089 | NVPTXISD::LoadParam, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1090 | DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo()); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1091 | Chain = retval.getValue(1); |
| 1092 | InFlag = retval.getValue(2); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1093 | SDValue Ret0 = retval; |
| 1094 | if (needTruncate) |
| 1095 | Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0); |
| 1096 | InVals.push_back(Ret0); |
| 1097 | } else if (NumElts == 2) { |
| 1098 | // LoadV2 |
Craig Topper | 59f626d | 2014-04-26 19:29:47 +0000 | [diff] [blame] | 1099 | SmallVector<EVT, 4> LoadRetVTs; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1100 | if (needTruncate) { |
| 1101 | // If loading i1 result, generate |
| 1102 | // load i16 |
| 1103 | // trunc i16 to i1 |
| 1104 | LoadRetVTs.push_back(MVT::i16); |
| 1105 | LoadRetVTs.push_back(MVT::i16); |
| 1106 | } else { |
| 1107 | LoadRetVTs.push_back(EltVT); |
| 1108 | LoadRetVTs.push_back(EltVT); |
| 1109 | } |
| 1110 | LoadRetVTs.push_back(MVT::Other); |
| 1111 | LoadRetVTs.push_back(MVT::Glue); |
Craig Topper | 59f626d | 2014-04-26 19:29:47 +0000 | [diff] [blame] | 1112 | SmallVector<SDValue, 4> LoadRetOps; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1113 | LoadRetOps.push_back(Chain); |
| 1114 | LoadRetOps.push_back(DAG.getConstant(1, MVT::i32)); |
| 1115 | LoadRetOps.push_back(DAG.getConstant(0, MVT::i32)); |
| 1116 | LoadRetOps.push_back(InFlag); |
| 1117 | SDValue retval = DAG.getMemIntrinsicNode( |
| 1118 | NVPTXISD::LoadParamV2, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1119 | DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo()); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1120 | Chain = retval.getValue(2); |
| 1121 | InFlag = retval.getValue(3); |
| 1122 | SDValue Ret0 = retval.getValue(0); |
| 1123 | SDValue Ret1 = retval.getValue(1); |
| 1124 | if (needTruncate) { |
| 1125 | Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0); |
| 1126 | InVals.push_back(Ret0); |
| 1127 | Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1); |
| 1128 | InVals.push_back(Ret1); |
| 1129 | } else { |
| 1130 | InVals.push_back(Ret0); |
| 1131 | InVals.push_back(Ret1); |
| 1132 | } |
| 1133 | } else { |
| 1134 | // Split into N LoadV4 |
| 1135 | unsigned Ofst = 0; |
| 1136 | unsigned VecSize = 4; |
| 1137 | unsigned Opc = NVPTXISD::LoadParamV4; |
| 1138 | if (EltVT.getSizeInBits() == 64) { |
| 1139 | VecSize = 2; |
| 1140 | Opc = NVPTXISD::LoadParamV2; |
| 1141 | } |
| 1142 | EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize); |
| 1143 | for (unsigned i = 0; i < NumElts; i += VecSize) { |
| 1144 | SmallVector<EVT, 8> LoadRetVTs; |
| 1145 | if (needTruncate) { |
| 1146 | // If loading i1 result, generate |
| 1147 | // load i16 |
| 1148 | // trunc i16 to i1 |
| 1149 | for (unsigned j = 0; j < VecSize; ++j) |
| 1150 | LoadRetVTs.push_back(MVT::i16); |
| 1151 | } else { |
| 1152 | for (unsigned j = 0; j < VecSize; ++j) |
| 1153 | LoadRetVTs.push_back(EltVT); |
| 1154 | } |
| 1155 | LoadRetVTs.push_back(MVT::Other); |
| 1156 | LoadRetVTs.push_back(MVT::Glue); |
| 1157 | SmallVector<SDValue, 4> LoadRetOps; |
| 1158 | LoadRetOps.push_back(Chain); |
| 1159 | LoadRetOps.push_back(DAG.getConstant(1, MVT::i32)); |
| 1160 | LoadRetOps.push_back(DAG.getConstant(Ofst, MVT::i32)); |
| 1161 | LoadRetOps.push_back(InFlag); |
| 1162 | SDValue retval = DAG.getMemIntrinsicNode( |
Craig Topper | abb4ac7 | 2014-04-16 06:10:51 +0000 | [diff] [blame] | 1163 | Opc, dl, DAG.getVTList(LoadRetVTs), |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1164 | LoadRetOps, EltVT, MachinePointerInfo()); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1165 | if (VecSize == 2) { |
| 1166 | Chain = retval.getValue(2); |
| 1167 | InFlag = retval.getValue(3); |
| 1168 | } else { |
| 1169 | Chain = retval.getValue(4); |
| 1170 | InFlag = retval.getValue(5); |
| 1171 | } |
| 1172 | |
| 1173 | for (unsigned j = 0; j < VecSize; ++j) { |
| 1174 | if (i + j >= NumElts) |
| 1175 | break; |
| 1176 | SDValue Elt = retval.getValue(j); |
| 1177 | if (needTruncate) |
| 1178 | Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt); |
| 1179 | InVals.push_back(Elt); |
| 1180 | } |
| 1181 | Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext())); |
| 1182 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1183 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1184 | } else { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1185 | SmallVector<EVT, 16> VTs; |
| 1186 | ComputePTXValueVTs(*this, retTy, VTs); |
| 1187 | assert(VTs.size() == Ins.size() && "Bad value decomposition"); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1188 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1189 | unsigned sz = VTs[i].getSizeInBits(); |
| 1190 | bool needTruncate = sz < 8 ? true : false; |
| 1191 | if (VTs[i].isInteger() && (sz < 8)) |
| 1192 | sz = 8; |
| 1193 | |
| 1194 | SmallVector<EVT, 4> LoadRetVTs; |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1195 | EVT TheLoadType = VTs[i]; |
| 1196 | if (retTy->isIntegerTy() && |
| 1197 | TD->getTypeAllocSizeInBits(retTy) < 32) { |
| 1198 | // This is for integer types only, and specifically not for |
| 1199 | // aggregates. |
| 1200 | LoadRetVTs.push_back(MVT::i32); |
| 1201 | TheLoadType = MVT::i32; |
| 1202 | } else if (sz < 16) { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1203 | // If loading i1/i8 result, generate |
| 1204 | // load i8 (-> i16) |
| 1205 | // trunc i16 to i1/i8 |
| 1206 | LoadRetVTs.push_back(MVT::i16); |
| 1207 | } else |
| 1208 | LoadRetVTs.push_back(Ins[i].VT); |
| 1209 | LoadRetVTs.push_back(MVT::Other); |
| 1210 | LoadRetVTs.push_back(MVT::Glue); |
| 1211 | |
| 1212 | SmallVector<SDValue, 4> LoadRetOps; |
| 1213 | LoadRetOps.push_back(Chain); |
| 1214 | LoadRetOps.push_back(DAG.getConstant(1, MVT::i32)); |
| 1215 | LoadRetOps.push_back(DAG.getConstant(resoffset, MVT::i32)); |
| 1216 | LoadRetOps.push_back(InFlag); |
| 1217 | SDValue retval = DAG.getMemIntrinsicNode( |
| 1218 | NVPTXISD::LoadParam, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1219 | DAG.getVTList(LoadRetVTs), LoadRetOps, |
| 1220 | TheLoadType, MachinePointerInfo()); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1221 | Chain = retval.getValue(1); |
| 1222 | InFlag = retval.getValue(2); |
| 1223 | SDValue Ret0 = retval.getValue(0); |
| 1224 | if (needTruncate) |
| 1225 | Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0); |
| 1226 | InVals.push_back(Ret0); |
| 1227 | resoffset += sz / 8; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1228 | } |
| 1229 | } |
| 1230 | } |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1231 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1232 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(uniqueCallSite, true), |
| 1233 | DAG.getIntPtrConstant(uniqueCallSite + 1, true), |
Andrew Trick | ad6d08a | 2013-05-29 22:03:55 +0000 | [diff] [blame] | 1234 | InFlag, dl); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1235 | uniqueCallSite++; |
| 1236 | |
| 1237 | // set isTailCall to false for now, until we figure out how to express |
| 1238 | // tail call optimization in PTX |
| 1239 | isTailCall = false; |
| 1240 | return Chain; |
| 1241 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1242 | |
| 1243 | // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack() |
| 1244 | // (see LegalizeDAG.cpp). This is slow and uses local memory. |
| 1245 | // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5 |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1246 | SDValue |
| 1247 | NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1248 | SDNode *Node = Op.getNode(); |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 1249 | SDLoc dl(Node); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1250 | SmallVector<SDValue, 8> Ops; |
| 1251 | unsigned NumOperands = Node->getNumOperands(); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1252 | for (unsigned i = 0; i < NumOperands; ++i) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1253 | SDValue SubOp = Node->getOperand(i); |
| 1254 | EVT VVT = SubOp.getNode()->getValueType(0); |
| 1255 | EVT EltVT = VVT.getVectorElementType(); |
| 1256 | unsigned NumSubElem = VVT.getVectorNumElements(); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1257 | for (unsigned j = 0; j < NumSubElem; ++j) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1258 | Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp, |
| 1259 | DAG.getIntPtrConstant(j))); |
| 1260 | } |
| 1261 | } |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1262 | return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Ops); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1263 | } |
| 1264 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1265 | SDValue |
| 1266 | NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1267 | switch (Op.getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1268 | case ISD::RETURNADDR: |
| 1269 | return SDValue(); |
| 1270 | case ISD::FRAMEADDR: |
| 1271 | return SDValue(); |
| 1272 | case ISD::GlobalAddress: |
| 1273 | return LowerGlobalAddress(Op, DAG); |
| 1274 | case ISD::INTRINSIC_W_CHAIN: |
| 1275 | return Op; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1276 | case ISD::BUILD_VECTOR: |
| 1277 | case ISD::EXTRACT_SUBVECTOR: |
| 1278 | return Op; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1279 | case ISD::CONCAT_VECTORS: |
| 1280 | return LowerCONCAT_VECTORS(Op, DAG); |
| 1281 | case ISD::STORE: |
| 1282 | return LowerSTORE(Op, DAG); |
| 1283 | case ISD::LOAD: |
| 1284 | return LowerLOAD(Op, DAG); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1285 | default: |
David Blaikie | 891d0a3 | 2012-05-04 22:34:16 +0000 | [diff] [blame] | 1286 | llvm_unreachable("Custom lowering not defined for operation"); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1287 | } |
| 1288 | } |
| 1289 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1290 | SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
| 1291 | if (Op.getValueType() == MVT::i1) |
| 1292 | return LowerLOADi1(Op, DAG); |
| 1293 | else |
| 1294 | return SDValue(); |
| 1295 | } |
| 1296 | |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1297 | // v = ld i1* addr |
| 1298 | // => |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1299 | // v1 = ld i8* addr (-> i16) |
| 1300 | // v = trunc i16 to i1 |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1301 | SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const { |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1302 | SDNode *Node = Op.getNode(); |
| 1303 | LoadSDNode *LD = cast<LoadSDNode>(Node); |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 1304 | SDLoc dl(Node); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1305 | assert(LD->getExtensionType() == ISD::NON_EXTLOAD); |
NAKAMURA Takumi | 5bbe0e1 | 2012-11-14 23:46:15 +0000 | [diff] [blame] | 1306 | assert(Node->getValueType(0) == MVT::i1 && |
| 1307 | "Custom lowering for i1 load only"); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1308 | SDValue newLD = |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1309 | DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(), |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1310 | LD->getPointerInfo(), LD->isVolatile(), LD->isNonTemporal(), |
| 1311 | LD->isInvariant(), LD->getAlignment()); |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1312 | SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD); |
| 1313 | // The legalizer (the caller) is expecting two values from the legalized |
| 1314 | // load, so we build a MergeValues node for it. See ExpandUnalignedLoad() |
| 1315 | // in LegalizeDAG.cpp which also uses MergeValues. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1316 | SDValue Ops[] = { result, LD->getChain() }; |
Craig Topper | 64941d9 | 2014-04-27 19:20:57 +0000 | [diff] [blame^] | 1317 | return DAG.getMergeValues(Ops, dl); |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1318 | } |
| 1319 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1320 | SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
| 1321 | EVT ValVT = Op.getOperand(1).getValueType(); |
| 1322 | if (ValVT == MVT::i1) |
| 1323 | return LowerSTOREi1(Op, DAG); |
| 1324 | else if (ValVT.isVector()) |
| 1325 | return LowerSTOREVector(Op, DAG); |
| 1326 | else |
| 1327 | return SDValue(); |
| 1328 | } |
| 1329 | |
| 1330 | SDValue |
| 1331 | NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const { |
| 1332 | SDNode *N = Op.getNode(); |
| 1333 | SDValue Val = N->getOperand(1); |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 1334 | SDLoc DL(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1335 | EVT ValVT = Val.getValueType(); |
| 1336 | |
| 1337 | if (ValVT.isVector()) { |
| 1338 | // We only handle "native" vector sizes for now, e.g. <4 x double> is not |
| 1339 | // legal. We can (and should) split that into 2 stores of <2 x double> here |
| 1340 | // but I'm leaving that as a TODO for now. |
| 1341 | if (!ValVT.isSimple()) |
| 1342 | return SDValue(); |
| 1343 | switch (ValVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1344 | default: |
| 1345 | return SDValue(); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1346 | case MVT::v2i8: |
| 1347 | case MVT::v2i16: |
| 1348 | case MVT::v2i32: |
| 1349 | case MVT::v2i64: |
| 1350 | case MVT::v2f32: |
| 1351 | case MVT::v2f64: |
| 1352 | case MVT::v4i8: |
| 1353 | case MVT::v4i16: |
| 1354 | case MVT::v4i32: |
| 1355 | case MVT::v4f32: |
| 1356 | // This is a "native" vector type |
| 1357 | break; |
| 1358 | } |
| 1359 | |
| 1360 | unsigned Opcode = 0; |
| 1361 | EVT EltVT = ValVT.getVectorElementType(); |
| 1362 | unsigned NumElts = ValVT.getVectorNumElements(); |
| 1363 | |
| 1364 | // Since StoreV2 is a target node, we cannot rely on DAG type legalization. |
| 1365 | // Therefore, we must ensure the type is legal. For i1 and i8, we set the |
Alp Toker | cb40291 | 2014-01-24 17:20:08 +0000 | [diff] [blame] | 1366 | // stored type to i16 and propagate the "real" type as the memory type. |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1367 | bool NeedExt = false; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1368 | if (EltVT.getSizeInBits() < 16) |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1369 | NeedExt = true; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1370 | |
| 1371 | switch (NumElts) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1372 | default: |
| 1373 | return SDValue(); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1374 | case 2: |
| 1375 | Opcode = NVPTXISD::StoreV2; |
| 1376 | break; |
| 1377 | case 4: { |
| 1378 | Opcode = NVPTXISD::StoreV4; |
| 1379 | break; |
| 1380 | } |
| 1381 | } |
| 1382 | |
| 1383 | SmallVector<SDValue, 8> Ops; |
| 1384 | |
| 1385 | // First is the chain |
| 1386 | Ops.push_back(N->getOperand(0)); |
| 1387 | |
| 1388 | // Then the split values |
| 1389 | for (unsigned i = 0; i < NumElts; ++i) { |
| 1390 | SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val, |
| 1391 | DAG.getIntPtrConstant(i)); |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1392 | if (NeedExt) |
| 1393 | ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1394 | Ops.push_back(ExtVal); |
| 1395 | } |
| 1396 | |
| 1397 | // Then any remaining arguments |
| 1398 | for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) { |
| 1399 | Ops.push_back(N->getOperand(i)); |
| 1400 | } |
| 1401 | |
| 1402 | MemSDNode *MemSD = cast<MemSDNode>(N); |
| 1403 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1404 | SDValue NewSt = DAG.getMemIntrinsicNode( |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1405 | Opcode, DL, DAG.getVTList(MVT::Other), Ops, |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1406 | MemSD->getMemoryVT(), MemSD->getMemOperand()); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1407 | |
| 1408 | //return DCI.CombineTo(N, NewSt, true); |
| 1409 | return NewSt; |
| 1410 | } |
| 1411 | |
| 1412 | return SDValue(); |
| 1413 | } |
| 1414 | |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1415 | // st i1 v, addr |
| 1416 | // => |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1417 | // v1 = zxt v to i16 |
| 1418 | // st.u8 i16, addr |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1419 | SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const { |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1420 | SDNode *Node = Op.getNode(); |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 1421 | SDLoc dl(Node); |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1422 | StoreSDNode *ST = cast<StoreSDNode>(Node); |
| 1423 | SDValue Tmp1 = ST->getChain(); |
| 1424 | SDValue Tmp2 = ST->getBasePtr(); |
| 1425 | SDValue Tmp3 = ST->getValue(); |
NAKAMURA Takumi | 5bbe0e1 | 2012-11-14 23:46:15 +0000 | [diff] [blame] | 1426 | assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only"); |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1427 | unsigned Alignment = ST->getAlignment(); |
| 1428 | bool isVolatile = ST->isVolatile(); |
| 1429 | bool isNonTemporal = ST->isNonTemporal(); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1430 | Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3); |
| 1431 | SDValue Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, |
| 1432 | ST->getPointerInfo(), MVT::i8, isNonTemporal, |
| 1433 | isVolatile, Alignment); |
Justin Holewinski | c6462aa | 2012-11-14 19:19:16 +0000 | [diff] [blame] | 1434 | return Result; |
| 1435 | } |
| 1436 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1437 | SDValue NVPTXTargetLowering::getExtSymb(SelectionDAG &DAG, const char *inname, |
| 1438 | int idx, EVT v) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1439 | std::string *name = nvTM->getManagedStrPool()->getManagedString(inname); |
| 1440 | std::stringstream suffix; |
| 1441 | suffix << idx; |
| 1442 | *name += suffix.str(); |
| 1443 | return DAG.getTargetExternalSymbol(name->c_str(), v); |
| 1444 | } |
| 1445 | |
| 1446 | SDValue |
| 1447 | NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const { |
Justin Holewinski | a2a63d2 | 2013-08-06 14:13:27 +0000 | [diff] [blame] | 1448 | std::string ParamSym; |
| 1449 | raw_string_ostream ParamStr(ParamSym); |
| 1450 | |
| 1451 | ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx; |
| 1452 | ParamStr.flush(); |
| 1453 | |
| 1454 | std::string *SavedStr = |
| 1455 | nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str()); |
| 1456 | return DAG.getTargetExternalSymbol(SavedStr->c_str(), v); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1457 | } |
| 1458 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1459 | SDValue NVPTXTargetLowering::getParamHelpSymbol(SelectionDAG &DAG, int idx) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1460 | return getExtSymb(DAG, ".HLPPARAM", idx); |
| 1461 | } |
| 1462 | |
| 1463 | // Check to see if the kernel argument is image*_t or sampler_t |
| 1464 | |
| 1465 | bool llvm::isImageOrSamplerVal(const Value *arg, const Module *context) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1466 | static const char *const specialTypes[] = { "struct._image2d_t", |
| 1467 | "struct._image3d_t", |
| 1468 | "struct._sampler_t" }; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1469 | |
| 1470 | const Type *Ty = arg->getType(); |
| 1471 | const PointerType *PTy = dyn_cast<PointerType>(Ty); |
| 1472 | |
| 1473 | if (!PTy) |
| 1474 | return false; |
| 1475 | |
| 1476 | if (!context) |
| 1477 | return false; |
| 1478 | |
| 1479 | const StructType *STy = dyn_cast<StructType>(PTy->getElementType()); |
Justin Holewinski | fb71115 | 2012-12-05 20:50:28 +0000 | [diff] [blame] | 1480 | const std::string TypeName = STy && !STy->isLiteral() ? STy->getName() : ""; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1481 | |
Craig Topper | e4260f9 | 2012-05-24 04:22:05 +0000 | [diff] [blame] | 1482 | for (int i = 0, e = array_lengthof(specialTypes); i != e; ++i) |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1483 | if (TypeName == specialTypes[i]) |
| 1484 | return true; |
| 1485 | |
| 1486 | return false; |
| 1487 | } |
| 1488 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1489 | SDValue NVPTXTargetLowering::LowerFormalArguments( |
| 1490 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 1491 | const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG, |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1492 | SmallVectorImpl<SDValue> &InVals) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1493 | MachineFunction &MF = DAG.getMachineFunction(); |
Micah Villmow | cdfe20b | 2012-10-08 16:38:25 +0000 | [diff] [blame] | 1494 | const DataLayout *TD = getDataLayout(); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1495 | |
| 1496 | const Function *F = MF.getFunction(); |
Bill Wendling | e94d843 | 2012-12-07 23:16:57 +0000 | [diff] [blame] | 1497 | const AttributeSet &PAL = F->getAttributes(); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1498 | const TargetLowering *TLI = nvTM->getTargetLowering(); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1499 | |
| 1500 | SDValue Root = DAG.getRoot(); |
| 1501 | std::vector<SDValue> OutChains; |
| 1502 | |
| 1503 | bool isKernel = llvm::isKernelFunction(*F); |
| 1504 | bool isABI = (nvptxSubtarget.getSmVersion() >= 20); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1505 | assert(isABI && "Non-ABI compilation is not supported"); |
| 1506 | if (!isABI) |
| 1507 | return Chain; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1508 | |
| 1509 | std::vector<Type *> argTypes; |
| 1510 | std::vector<const Argument *> theArgs; |
| 1511 | for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1512 | I != E; ++I) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1513 | theArgs.push_back(I); |
| 1514 | argTypes.push_back(I->getType()); |
| 1515 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1516 | // argTypes.size() (or theArgs.size()) and Ins.size() need not match. |
| 1517 | // Ins.size() will be larger |
| 1518 | // * if there is an aggregate argument with multiple fields (each field |
| 1519 | // showing up separately in Ins) |
| 1520 | // * if there is a vector argument with more than typical vector-length |
| 1521 | // elements (generally if more than 4) where each vector element is |
| 1522 | // individually present in Ins. |
| 1523 | // So a different index should be used for indexing into Ins. |
| 1524 | // See similar issue in LowerCall. |
| 1525 | unsigned InsIdx = 0; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1526 | |
| 1527 | int idx = 0; |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1528 | for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1529 | Type *Ty = argTypes[i]; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1530 | |
| 1531 | // If the kernel argument is image*_t or sampler_t, convert it to |
| 1532 | // a i32 constant holding the parameter position. This can later |
| 1533 | // matched in the AsmPrinter to output the correct mangled name. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1534 | if (isImageOrSamplerVal( |
| 1535 | theArgs[i], |
| 1536 | (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent() |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1537 | : nullptr))) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1538 | assert(isKernel && "Only kernels can have image/sampler params"); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1539 | InVals.push_back(DAG.getConstant(i + 1, MVT::i32)); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1540 | continue; |
| 1541 | } |
| 1542 | |
| 1543 | if (theArgs[i]->use_empty()) { |
| 1544 | // argument is dead |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1545 | if (Ty->isAggregateType()) { |
| 1546 | SmallVector<EVT, 16> vtparts; |
| 1547 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1548 | ComputePTXValueVTs(*this, Ty, vtparts); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1549 | assert(vtparts.size() > 0 && "empty aggregate type not expected"); |
| 1550 | for (unsigned parti = 0, parte = vtparts.size(); parti != parte; |
| 1551 | ++parti) { |
| 1552 | EVT partVT = vtparts[parti]; |
| 1553 | InVals.push_back(DAG.getNode(ISD::UNDEF, dl, partVT)); |
| 1554 | ++InsIdx; |
Justin Holewinski | e988409 | 2013-03-24 21:17:47 +0000 | [diff] [blame] | 1555 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1556 | if (vtparts.size() > 0) |
| 1557 | --InsIdx; |
| 1558 | continue; |
Justin Holewinski | e988409 | 2013-03-24 21:17:47 +0000 | [diff] [blame] | 1559 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1560 | if (Ty->isVectorTy()) { |
| 1561 | EVT ObjectVT = getValueType(Ty); |
| 1562 | unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT); |
| 1563 | for (unsigned parti = 0; parti < NumRegs; ++parti) { |
| 1564 | InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT)); |
| 1565 | ++InsIdx; |
| 1566 | } |
| 1567 | if (NumRegs > 0) |
| 1568 | --InsIdx; |
| 1569 | continue; |
| 1570 | } |
| 1571 | InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT)); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1572 | continue; |
| 1573 | } |
| 1574 | |
| 1575 | // In the following cases, assign a node order of "idx+1" |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1576 | // to newly created nodes. The SDNodes for params have to |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1577 | // appear in the same order as their order of appearance |
| 1578 | // in the original function. "idx+1" holds that order. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1579 | if (PAL.hasAttribute(i + 1, Attribute::ByVal) == false) { |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1580 | if (Ty->isAggregateType()) { |
| 1581 | SmallVector<EVT, 16> vtparts; |
| 1582 | SmallVector<uint64_t, 16> offsets; |
| 1583 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1584 | // NOTE: Here, we lose the ability to issue vector loads for vectors |
| 1585 | // that are a part of a struct. This should be investigated in the |
| 1586 | // future. |
| 1587 | ComputePTXValueVTs(*this, Ty, vtparts, &offsets, 0); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1588 | assert(vtparts.size() > 0 && "empty aggregate type not expected"); |
| 1589 | bool aggregateIsPacked = false; |
| 1590 | if (StructType *STy = llvm::dyn_cast<StructType>(Ty)) |
| 1591 | aggregateIsPacked = STy->isPacked(); |
| 1592 | |
| 1593 | SDValue Arg = getParamSymbol(DAG, idx, getPointerTy()); |
| 1594 | for (unsigned parti = 0, parte = vtparts.size(); parti != parte; |
| 1595 | ++parti) { |
| 1596 | EVT partVT = vtparts[parti]; |
| 1597 | Value *srcValue = Constant::getNullValue( |
| 1598 | PointerType::get(partVT.getTypeForEVT(F->getContext()), |
| 1599 | llvm::ADDRESS_SPACE_PARAM)); |
| 1600 | SDValue srcAddr = |
| 1601 | DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, |
| 1602 | DAG.getConstant(offsets[parti], getPointerTy())); |
| 1603 | unsigned partAlign = |
| 1604 | aggregateIsPacked ? 1 |
| 1605 | : TD->getABITypeAlignment( |
| 1606 | partVT.getTypeForEVT(F->getContext())); |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1607 | SDValue p; |
| 1608 | if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) { |
| 1609 | ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ? |
| 1610 | ISD::SEXTLOAD : ISD::ZEXTLOAD; |
| 1611 | p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1612 | MachinePointerInfo(srcValue), partVT, false, |
| 1613 | false, partAlign); |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1614 | } else { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1615 | p = DAG.getLoad(partVT, dl, Root, srcAddr, |
| 1616 | MachinePointerInfo(srcValue), false, false, false, |
| 1617 | partAlign); |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1618 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1619 | if (p.getNode()) |
| 1620 | p.getNode()->setIROrder(idx + 1); |
| 1621 | InVals.push_back(p); |
| 1622 | ++InsIdx; |
Justin Holewinski | e988409 | 2013-03-24 21:17:47 +0000 | [diff] [blame] | 1623 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1624 | if (vtparts.size() > 0) |
| 1625 | --InsIdx; |
Justin Holewinski | e988409 | 2013-03-24 21:17:47 +0000 | [diff] [blame] | 1626 | continue; |
| 1627 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1628 | if (Ty->isVectorTy()) { |
| 1629 | EVT ObjectVT = getValueType(Ty); |
Justin Holewinski | aaaf289 | 2013-06-25 12:22:21 +0000 | [diff] [blame] | 1630 | SDValue Arg = getParamSymbol(DAG, idx, getPointerTy()); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1631 | unsigned NumElts = ObjectVT.getVectorNumElements(); |
| 1632 | assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts && |
| 1633 | "Vector was not scalarized"); |
| 1634 | unsigned Ofst = 0; |
| 1635 | EVT EltVT = ObjectVT.getVectorElementType(); |
| 1636 | |
| 1637 | // V1 load |
| 1638 | // f32 = load ... |
| 1639 | if (NumElts == 1) { |
| 1640 | // We only have one element, so just directly load it |
| 1641 | Value *SrcValue = Constant::getNullValue(PointerType::get( |
| 1642 | EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM)); |
| 1643 | SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, |
| 1644 | DAG.getConstant(Ofst, getPointerTy())); |
| 1645 | SDValue P = DAG.getLoad( |
| 1646 | EltVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false, |
| 1647 | false, true, |
| 1648 | TD->getABITypeAlignment(EltVT.getTypeForEVT(F->getContext()))); |
| 1649 | if (P.getNode()) |
| 1650 | P.getNode()->setIROrder(idx + 1); |
| 1651 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1652 | if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1653 | P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1654 | InVals.push_back(P); |
| 1655 | Ofst += TD->getTypeAllocSize(EltVT.getTypeForEVT(F->getContext())); |
| 1656 | ++InsIdx; |
| 1657 | } else if (NumElts == 2) { |
| 1658 | // V2 load |
| 1659 | // f32,f32 = load ... |
| 1660 | EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2); |
| 1661 | Value *SrcValue = Constant::getNullValue(PointerType::get( |
| 1662 | VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM)); |
| 1663 | SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, |
| 1664 | DAG.getConstant(Ofst, getPointerTy())); |
| 1665 | SDValue P = DAG.getLoad( |
| 1666 | VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false, |
| 1667 | false, true, |
| 1668 | TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext()))); |
| 1669 | if (P.getNode()) |
| 1670 | P.getNode()->setIROrder(idx + 1); |
| 1671 | |
| 1672 | SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P, |
| 1673 | DAG.getIntPtrConstant(0)); |
| 1674 | SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P, |
| 1675 | DAG.getIntPtrConstant(1)); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1676 | |
| 1677 | if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) { |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1678 | Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0); |
| 1679 | Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1680 | } |
| 1681 | |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1682 | InVals.push_back(Elt0); |
| 1683 | InVals.push_back(Elt1); |
| 1684 | Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext())); |
| 1685 | InsIdx += 2; |
| 1686 | } else { |
| 1687 | // V4 loads |
| 1688 | // We have at least 4 elements (<3 x Ty> expands to 4 elements) and |
| 1689 | // the |
| 1690 | // vector will be expanded to a power of 2 elements, so we know we can |
| 1691 | // always round up to the next multiple of 4 when creating the vector |
| 1692 | // loads. |
| 1693 | // e.g. 4 elem => 1 ld.v4 |
| 1694 | // 6 elem => 2 ld.v4 |
| 1695 | // 8 elem => 2 ld.v4 |
| 1696 | // 11 elem => 3 ld.v4 |
| 1697 | unsigned VecSize = 4; |
| 1698 | if (EltVT.getSizeInBits() == 64) { |
| 1699 | VecSize = 2; |
| 1700 | } |
| 1701 | EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize); |
| 1702 | for (unsigned i = 0; i < NumElts; i += VecSize) { |
| 1703 | Value *SrcValue = Constant::getNullValue( |
| 1704 | PointerType::get(VecVT.getTypeForEVT(F->getContext()), |
| 1705 | llvm::ADDRESS_SPACE_PARAM)); |
| 1706 | SDValue SrcAddr = |
| 1707 | DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, |
| 1708 | DAG.getConstant(Ofst, getPointerTy())); |
| 1709 | SDValue P = DAG.getLoad( |
| 1710 | VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false, |
| 1711 | false, true, |
| 1712 | TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext()))); |
| 1713 | if (P.getNode()) |
| 1714 | P.getNode()->setIROrder(idx + 1); |
| 1715 | |
| 1716 | for (unsigned j = 0; j < VecSize; ++j) { |
| 1717 | if (i + j >= NumElts) |
| 1718 | break; |
| 1719 | SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P, |
| 1720 | DAG.getIntPtrConstant(j)); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1721 | if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1722 | Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1723 | InVals.push_back(Elt); |
| 1724 | } |
| 1725 | Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext())); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1726 | } |
Justin Holewinski | 4f5bc9b | 2013-11-11 19:28:16 +0000 | [diff] [blame] | 1727 | InsIdx += NumElts; |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1728 | } |
| 1729 | |
| 1730 | if (NumElts > 0) |
| 1731 | --InsIdx; |
| 1732 | continue; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1733 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1734 | // A plain scalar. |
| 1735 | EVT ObjectVT = getValueType(Ty); |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1736 | // If ABI, load from the param symbol |
| 1737 | SDValue Arg = getParamSymbol(DAG, idx, getPointerTy()); |
| 1738 | Value *srcValue = Constant::getNullValue(PointerType::get( |
| 1739 | ObjectVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM)); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1740 | SDValue p; |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1741 | if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) { |
| 1742 | ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ? |
| 1743 | ISD::SEXTLOAD : ISD::ZEXTLOAD; |
| 1744 | p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, Arg, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1745 | MachinePointerInfo(srcValue), ObjectVT, false, false, |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1746 | TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext()))); |
| 1747 | } else { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1748 | p = DAG.getLoad(Ins[InsIdx].VT, dl, Root, Arg, |
| 1749 | MachinePointerInfo(srcValue), false, false, false, |
Justin Holewinski | a291128 | 2013-07-01 12:58:58 +0000 | [diff] [blame] | 1750 | TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext()))); |
| 1751 | } |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1752 | if (p.getNode()) |
| 1753 | p.getNode()->setIROrder(idx + 1); |
| 1754 | InVals.push_back(p); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1755 | continue; |
| 1756 | } |
| 1757 | |
| 1758 | // Param has ByVal attribute |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1759 | // Return MoveParam(param symbol). |
| 1760 | // Ideally, the param symbol can be returned directly, |
| 1761 | // but when SDNode builder decides to use it in a CopyToReg(), |
| 1762 | // machine instruction fails because TargetExternalSymbol |
| 1763 | // (not lowered) is target dependent, and CopyToReg assumes |
| 1764 | // the source is lowered. |
| 1765 | EVT ObjectVT = getValueType(Ty); |
| 1766 | assert(ObjectVT == Ins[InsIdx].VT && |
| 1767 | "Ins type did not match function type"); |
| 1768 | SDValue Arg = getParamSymbol(DAG, idx, getPointerTy()); |
| 1769 | SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg); |
| 1770 | if (p.getNode()) |
| 1771 | p.getNode()->setIROrder(idx + 1); |
| 1772 | if (isKernel) |
| 1773 | InVals.push_back(p); |
| 1774 | else { |
| 1775 | SDValue p2 = DAG.getNode( |
| 1776 | ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT, |
| 1777 | DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, MVT::i32), p); |
| 1778 | InVals.push_back(p2); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1779 | } |
| 1780 | } |
| 1781 | |
| 1782 | // Clang will check explicit VarArg and issue error if any. However, Clang |
| 1783 | // will let code with |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1784 | // implicit var arg like f() pass. See bug 617733. |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1785 | // We treat this case as if the arg list is empty. |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1786 | // if (F.isVarArg()) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1787 | // assert(0 && "VarArg not supported yet!"); |
| 1788 | //} |
| 1789 | |
| 1790 | if (!OutChains.empty()) |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 1791 | DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains)); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1792 | |
| 1793 | return Chain; |
| 1794 | } |
| 1795 | |
Justin Holewinski | 44f5c60 | 2013-06-28 17:57:53 +0000 | [diff] [blame] | 1796 | |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1797 | SDValue |
| 1798 | NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
| 1799 | bool isVarArg, |
| 1800 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 1801 | const SmallVectorImpl<SDValue> &OutVals, |
| 1802 | SDLoc dl, SelectionDAG &DAG) const { |
| 1803 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1804 | const Function *F = MF.getFunction(); |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1805 | Type *RetTy = F->getReturnType(); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1806 | const DataLayout *TD = getDataLayout(); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1807 | |
| 1808 | bool isABI = (nvptxSubtarget.getSmVersion() >= 20); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1809 | assert(isABI && "Non-ABI compilation is not supported"); |
| 1810 | if (!isABI) |
| 1811 | return Chain; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1812 | |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1813 | if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) { |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1814 | // If we have a vector type, the OutVals array will be the scalarized |
| 1815 | // components and we have combine them into 1 or more vector stores. |
| 1816 | unsigned NumElts = VTy->getNumElements(); |
| 1817 | assert(NumElts == Outs.size() && "Bad scalarization of return value"); |
| 1818 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1819 | // const_cast can be removed in later LLVM versions |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1820 | EVT EltVT = getValueType(RetTy).getVectorElementType(); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1821 | bool NeedExtend = false; |
| 1822 | if (EltVT.getSizeInBits() < 16) |
| 1823 | NeedExtend = true; |
| 1824 | |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1825 | // V1 store |
| 1826 | if (NumElts == 1) { |
| 1827 | SDValue StoreVal = OutVals[0]; |
| 1828 | // We only have one element, so just directly store it |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1829 | if (NeedExtend) |
| 1830 | StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal); |
| 1831 | SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal }; |
| 1832 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1833 | DAG.getVTList(MVT::Other), Ops, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1834 | EltVT, MachinePointerInfo()); |
| 1835 | |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1836 | } else if (NumElts == 2) { |
| 1837 | // V2 store |
| 1838 | SDValue StoreVal0 = OutVals[0]; |
| 1839 | SDValue StoreVal1 = OutVals[1]; |
| 1840 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1841 | if (NeedExtend) { |
| 1842 | StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0); |
| 1843 | StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1844 | } |
| 1845 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1846 | SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal0, |
| 1847 | StoreVal1 }; |
| 1848 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1849 | DAG.getVTList(MVT::Other), Ops, |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1850 | EltVT, MachinePointerInfo()); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1851 | } else { |
| 1852 | // V4 stores |
| 1853 | // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the |
| 1854 | // vector will be expanded to a power of 2 elements, so we know we can |
| 1855 | // always round up to the next multiple of 4 when creating the vector |
| 1856 | // stores. |
| 1857 | // e.g. 4 elem => 1 st.v4 |
| 1858 | // 6 elem => 2 st.v4 |
| 1859 | // 8 elem => 2 st.v4 |
| 1860 | // 11 elem => 3 st.v4 |
| 1861 | |
| 1862 | unsigned VecSize = 4; |
| 1863 | if (OutVals[0].getValueType().getSizeInBits() == 64) |
| 1864 | VecSize = 2; |
| 1865 | |
| 1866 | unsigned Offset = 0; |
| 1867 | |
| 1868 | EVT VecVT = |
| 1869 | EVT::getVectorVT(F->getContext(), OutVals[0].getValueType(), VecSize); |
| 1870 | unsigned PerStoreOffset = |
| 1871 | TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext())); |
| 1872 | |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1873 | for (unsigned i = 0; i < NumElts; i += VecSize) { |
| 1874 | // Get values |
| 1875 | SDValue StoreVal; |
| 1876 | SmallVector<SDValue, 8> Ops; |
| 1877 | Ops.push_back(Chain); |
| 1878 | Ops.push_back(DAG.getConstant(Offset, MVT::i32)); |
| 1879 | unsigned Opc = NVPTXISD::StoreRetvalV2; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1880 | EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType(); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1881 | |
| 1882 | StoreVal = OutVals[i]; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1883 | if (NeedExtend) |
| 1884 | StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1885 | Ops.push_back(StoreVal); |
| 1886 | |
| 1887 | if (i + 1 < NumElts) { |
| 1888 | StoreVal = OutVals[i + 1]; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1889 | if (NeedExtend) |
| 1890 | StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1891 | } else { |
| 1892 | StoreVal = DAG.getUNDEF(ExtendedVT); |
| 1893 | } |
| 1894 | Ops.push_back(StoreVal); |
| 1895 | |
| 1896 | if (VecSize == 4) { |
| 1897 | Opc = NVPTXISD::StoreRetvalV4; |
| 1898 | if (i + 2 < NumElts) { |
| 1899 | StoreVal = OutVals[i + 2]; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1900 | if (NeedExtend) |
| 1901 | StoreVal = |
| 1902 | DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1903 | } else { |
| 1904 | StoreVal = DAG.getUNDEF(ExtendedVT); |
| 1905 | } |
| 1906 | Ops.push_back(StoreVal); |
| 1907 | |
| 1908 | if (i + 3 < NumElts) { |
| 1909 | StoreVal = OutVals[i + 3]; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1910 | if (NeedExtend) |
| 1911 | StoreVal = |
| 1912 | DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1913 | } else { |
| 1914 | StoreVal = DAG.getUNDEF(ExtendedVT); |
| 1915 | } |
| 1916 | Ops.push_back(StoreVal); |
| 1917 | } |
| 1918 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1919 | // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size()); |
| 1920 | Chain = |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1921 | DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops, |
| 1922 | EltVT, MachinePointerInfo()); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1923 | Offset += PerStoreOffset; |
| 1924 | } |
| 1925 | } |
| 1926 | } else { |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1927 | SmallVector<EVT, 16> ValVTs; |
| 1928 | // const_cast is necessary since we are still using an LLVM version from |
| 1929 | // before the type system re-write. |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1930 | ComputePTXValueVTs(*this, RetTy, ValVTs); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1931 | assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition"); |
| 1932 | |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1933 | unsigned SizeSoFar = 0; |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1934 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
| 1935 | SDValue theVal = OutVals[i]; |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1936 | EVT TheValType = theVal.getValueType(); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1937 | unsigned numElems = 1; |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1938 | if (TheValType.isVector()) |
| 1939 | numElems = TheValType.getVectorNumElements(); |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1940 | for (unsigned j = 0, je = numElems; j != je; ++j) { |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1941 | SDValue TmpVal = theVal; |
| 1942 | if (TheValType.isVector()) |
| 1943 | TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, |
| 1944 | TheValType.getVectorElementType(), TmpVal, |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1945 | DAG.getIntPtrConstant(j)); |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1946 | EVT TheStoreType = ValVTs[i]; |
| 1947 | if (RetTy->isIntegerTy() && |
| 1948 | TD->getTypeAllocSizeInBits(RetTy) < 32) { |
| 1949 | // The following zero-extension is for integer types only, and |
| 1950 | // specifically not for aggregates. |
| 1951 | TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal); |
| 1952 | TheStoreType = MVT::i32; |
| 1953 | } |
| 1954 | else if (TmpVal.getValueType().getSizeInBits() < 16) |
| 1955 | TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal); |
| 1956 | |
| 1957 | SDValue Ops[] = { Chain, DAG.getConstant(SizeSoFar, MVT::i32), TmpVal }; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1958 | Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl, |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 1959 | DAG.getVTList(MVT::Other), Ops, |
| 1960 | TheStoreType, |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1961 | MachinePointerInfo()); |
| 1962 | if(TheValType.isVector()) |
| 1963 | SizeSoFar += |
| 1964 | TheStoreType.getVectorElementType().getStoreSizeInBits() / 8; |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1965 | else |
Justin Holewinski | e04e4bd | 2013-06-28 17:58:10 +0000 | [diff] [blame] | 1966 | SizeSoFar += TheStoreType.getStoreSizeInBits()/8; |
Justin Holewinski | 120baee | 2013-06-28 17:57:55 +0000 | [diff] [blame] | 1967 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1968 | } |
| 1969 | } |
| 1970 | |
| 1971 | return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain); |
| 1972 | } |
| 1973 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1974 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1975 | void NVPTXTargetLowering::LowerAsmOperandForConstraint( |
| 1976 | SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops, |
| 1977 | SelectionDAG &DAG) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1978 | if (Constraint.length() > 1) |
| 1979 | return; |
| 1980 | else |
| 1981 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
| 1982 | } |
| 1983 | |
| 1984 | // NVPTX suuport vector of legal types of any length in Intrinsics because the |
| 1985 | // NVPTX specific type legalizer |
| 1986 | // will legalize them to the PTX supported length. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1987 | bool NVPTXTargetLowering::isTypeSupportedInIntrinsic(MVT VT) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1988 | if (isTypeLegal(VT)) |
| 1989 | return true; |
| 1990 | if (VT.isVector()) { |
| 1991 | MVT eVT = VT.getVectorElementType(); |
| 1992 | if (isTypeLegal(eVT)) |
| 1993 | return true; |
| 1994 | } |
| 1995 | return false; |
| 1996 | } |
| 1997 | |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 1998 | static unsigned getOpcForTextureInstr(unsigned Intrinsic) { |
| 1999 | switch (Intrinsic) { |
| 2000 | default: |
| 2001 | return 0; |
| 2002 | |
| 2003 | case Intrinsic::nvvm_tex_1d_v4f32_i32: |
| 2004 | return NVPTXISD::Tex1DFloatI32; |
| 2005 | case Intrinsic::nvvm_tex_1d_v4f32_f32: |
| 2006 | return NVPTXISD::Tex1DFloatFloat; |
| 2007 | case Intrinsic::nvvm_tex_1d_level_v4f32_f32: |
| 2008 | return NVPTXISD::Tex1DFloatFloatLevel; |
| 2009 | case Intrinsic::nvvm_tex_1d_grad_v4f32_f32: |
| 2010 | return NVPTXISD::Tex1DFloatFloatGrad; |
| 2011 | case Intrinsic::nvvm_tex_1d_v4i32_i32: |
| 2012 | return NVPTXISD::Tex1DI32I32; |
| 2013 | case Intrinsic::nvvm_tex_1d_v4i32_f32: |
| 2014 | return NVPTXISD::Tex1DI32Float; |
| 2015 | case Intrinsic::nvvm_tex_1d_level_v4i32_f32: |
| 2016 | return NVPTXISD::Tex1DI32FloatLevel; |
| 2017 | case Intrinsic::nvvm_tex_1d_grad_v4i32_f32: |
| 2018 | return NVPTXISD::Tex1DI32FloatGrad; |
| 2019 | |
| 2020 | case Intrinsic::nvvm_tex_1d_array_v4f32_i32: |
| 2021 | return NVPTXISD::Tex1DArrayFloatI32; |
| 2022 | case Intrinsic::nvvm_tex_1d_array_v4f32_f32: |
| 2023 | return NVPTXISD::Tex1DArrayFloatFloat; |
| 2024 | case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32: |
| 2025 | return NVPTXISD::Tex1DArrayFloatFloatLevel; |
| 2026 | case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32: |
| 2027 | return NVPTXISD::Tex1DArrayFloatFloatGrad; |
| 2028 | case Intrinsic::nvvm_tex_1d_array_v4i32_i32: |
| 2029 | return NVPTXISD::Tex1DArrayI32I32; |
| 2030 | case Intrinsic::nvvm_tex_1d_array_v4i32_f32: |
| 2031 | return NVPTXISD::Tex1DArrayI32Float; |
| 2032 | case Intrinsic::nvvm_tex_1d_array_level_v4i32_f32: |
| 2033 | return NVPTXISD::Tex1DArrayI32FloatLevel; |
| 2034 | case Intrinsic::nvvm_tex_1d_array_grad_v4i32_f32: |
| 2035 | return NVPTXISD::Tex1DArrayI32FloatGrad; |
| 2036 | |
| 2037 | case Intrinsic::nvvm_tex_2d_v4f32_i32: |
| 2038 | return NVPTXISD::Tex2DFloatI32; |
| 2039 | case Intrinsic::nvvm_tex_2d_v4f32_f32: |
| 2040 | return NVPTXISD::Tex2DFloatFloat; |
| 2041 | case Intrinsic::nvvm_tex_2d_level_v4f32_f32: |
| 2042 | return NVPTXISD::Tex2DFloatFloatLevel; |
| 2043 | case Intrinsic::nvvm_tex_2d_grad_v4f32_f32: |
| 2044 | return NVPTXISD::Tex2DFloatFloatGrad; |
| 2045 | case Intrinsic::nvvm_tex_2d_v4i32_i32: |
| 2046 | return NVPTXISD::Tex2DI32I32; |
| 2047 | case Intrinsic::nvvm_tex_2d_v4i32_f32: |
| 2048 | return NVPTXISD::Tex2DI32Float; |
| 2049 | case Intrinsic::nvvm_tex_2d_level_v4i32_f32: |
| 2050 | return NVPTXISD::Tex2DI32FloatLevel; |
| 2051 | case Intrinsic::nvvm_tex_2d_grad_v4i32_f32: |
| 2052 | return NVPTXISD::Tex2DI32FloatGrad; |
| 2053 | |
| 2054 | case Intrinsic::nvvm_tex_2d_array_v4f32_i32: |
| 2055 | return NVPTXISD::Tex2DArrayFloatI32; |
| 2056 | case Intrinsic::nvvm_tex_2d_array_v4f32_f32: |
| 2057 | return NVPTXISD::Tex2DArrayFloatFloat; |
| 2058 | case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32: |
| 2059 | return NVPTXISD::Tex2DArrayFloatFloatLevel; |
| 2060 | case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32: |
| 2061 | return NVPTXISD::Tex2DArrayFloatFloatGrad; |
| 2062 | case Intrinsic::nvvm_tex_2d_array_v4i32_i32: |
| 2063 | return NVPTXISD::Tex2DArrayI32I32; |
| 2064 | case Intrinsic::nvvm_tex_2d_array_v4i32_f32: |
| 2065 | return NVPTXISD::Tex2DArrayI32Float; |
| 2066 | case Intrinsic::nvvm_tex_2d_array_level_v4i32_f32: |
| 2067 | return NVPTXISD::Tex2DArrayI32FloatLevel; |
| 2068 | case Intrinsic::nvvm_tex_2d_array_grad_v4i32_f32: |
| 2069 | return NVPTXISD::Tex2DArrayI32FloatGrad; |
| 2070 | |
| 2071 | case Intrinsic::nvvm_tex_3d_v4f32_i32: |
| 2072 | return NVPTXISD::Tex3DFloatI32; |
| 2073 | case Intrinsic::nvvm_tex_3d_v4f32_f32: |
| 2074 | return NVPTXISD::Tex3DFloatFloat; |
| 2075 | case Intrinsic::nvvm_tex_3d_level_v4f32_f32: |
| 2076 | return NVPTXISD::Tex3DFloatFloatLevel; |
| 2077 | case Intrinsic::nvvm_tex_3d_grad_v4f32_f32: |
| 2078 | return NVPTXISD::Tex3DFloatFloatGrad; |
| 2079 | case Intrinsic::nvvm_tex_3d_v4i32_i32: |
| 2080 | return NVPTXISD::Tex3DI32I32; |
| 2081 | case Intrinsic::nvvm_tex_3d_v4i32_f32: |
| 2082 | return NVPTXISD::Tex3DI32Float; |
| 2083 | case Intrinsic::nvvm_tex_3d_level_v4i32_f32: |
| 2084 | return NVPTXISD::Tex3DI32FloatLevel; |
| 2085 | case Intrinsic::nvvm_tex_3d_grad_v4i32_f32: |
| 2086 | return NVPTXISD::Tex3DI32FloatGrad; |
| 2087 | } |
| 2088 | } |
| 2089 | |
| 2090 | static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) { |
| 2091 | switch (Intrinsic) { |
| 2092 | default: |
| 2093 | return 0; |
| 2094 | case Intrinsic::nvvm_suld_1d_i8_trap: |
| 2095 | return NVPTXISD::Suld1DI8Trap; |
| 2096 | case Intrinsic::nvvm_suld_1d_i16_trap: |
| 2097 | return NVPTXISD::Suld1DI16Trap; |
| 2098 | case Intrinsic::nvvm_suld_1d_i32_trap: |
| 2099 | return NVPTXISD::Suld1DI32Trap; |
| 2100 | case Intrinsic::nvvm_suld_1d_v2i8_trap: |
| 2101 | return NVPTXISD::Suld1DV2I8Trap; |
| 2102 | case Intrinsic::nvvm_suld_1d_v2i16_trap: |
| 2103 | return NVPTXISD::Suld1DV2I16Trap; |
| 2104 | case Intrinsic::nvvm_suld_1d_v2i32_trap: |
| 2105 | return NVPTXISD::Suld1DV2I32Trap; |
| 2106 | case Intrinsic::nvvm_suld_1d_v4i8_trap: |
| 2107 | return NVPTXISD::Suld1DV4I8Trap; |
| 2108 | case Intrinsic::nvvm_suld_1d_v4i16_trap: |
| 2109 | return NVPTXISD::Suld1DV4I16Trap; |
| 2110 | case Intrinsic::nvvm_suld_1d_v4i32_trap: |
| 2111 | return NVPTXISD::Suld1DV4I32Trap; |
| 2112 | case Intrinsic::nvvm_suld_1d_array_i8_trap: |
| 2113 | return NVPTXISD::Suld1DArrayI8Trap; |
| 2114 | case Intrinsic::nvvm_suld_1d_array_i16_trap: |
| 2115 | return NVPTXISD::Suld1DArrayI16Trap; |
| 2116 | case Intrinsic::nvvm_suld_1d_array_i32_trap: |
| 2117 | return NVPTXISD::Suld1DArrayI32Trap; |
| 2118 | case Intrinsic::nvvm_suld_1d_array_v2i8_trap: |
| 2119 | return NVPTXISD::Suld1DArrayV2I8Trap; |
| 2120 | case Intrinsic::nvvm_suld_1d_array_v2i16_trap: |
| 2121 | return NVPTXISD::Suld1DArrayV2I16Trap; |
| 2122 | case Intrinsic::nvvm_suld_1d_array_v2i32_trap: |
| 2123 | return NVPTXISD::Suld1DArrayV2I32Trap; |
| 2124 | case Intrinsic::nvvm_suld_1d_array_v4i8_trap: |
| 2125 | return NVPTXISD::Suld1DArrayV4I8Trap; |
| 2126 | case Intrinsic::nvvm_suld_1d_array_v4i16_trap: |
| 2127 | return NVPTXISD::Suld1DArrayV4I16Trap; |
| 2128 | case Intrinsic::nvvm_suld_1d_array_v4i32_trap: |
| 2129 | return NVPTXISD::Suld1DArrayV4I32Trap; |
| 2130 | case Intrinsic::nvvm_suld_2d_i8_trap: |
| 2131 | return NVPTXISD::Suld2DI8Trap; |
| 2132 | case Intrinsic::nvvm_suld_2d_i16_trap: |
| 2133 | return NVPTXISD::Suld2DI16Trap; |
| 2134 | case Intrinsic::nvvm_suld_2d_i32_trap: |
| 2135 | return NVPTXISD::Suld2DI32Trap; |
| 2136 | case Intrinsic::nvvm_suld_2d_v2i8_trap: |
| 2137 | return NVPTXISD::Suld2DV2I8Trap; |
| 2138 | case Intrinsic::nvvm_suld_2d_v2i16_trap: |
| 2139 | return NVPTXISD::Suld2DV2I16Trap; |
| 2140 | case Intrinsic::nvvm_suld_2d_v2i32_trap: |
| 2141 | return NVPTXISD::Suld2DV2I32Trap; |
| 2142 | case Intrinsic::nvvm_suld_2d_v4i8_trap: |
| 2143 | return NVPTXISD::Suld2DV4I8Trap; |
| 2144 | case Intrinsic::nvvm_suld_2d_v4i16_trap: |
| 2145 | return NVPTXISD::Suld2DV4I16Trap; |
| 2146 | case Intrinsic::nvvm_suld_2d_v4i32_trap: |
| 2147 | return NVPTXISD::Suld2DV4I32Trap; |
| 2148 | case Intrinsic::nvvm_suld_2d_array_i8_trap: |
| 2149 | return NVPTXISD::Suld2DArrayI8Trap; |
| 2150 | case Intrinsic::nvvm_suld_2d_array_i16_trap: |
| 2151 | return NVPTXISD::Suld2DArrayI16Trap; |
| 2152 | case Intrinsic::nvvm_suld_2d_array_i32_trap: |
| 2153 | return NVPTXISD::Suld2DArrayI32Trap; |
| 2154 | case Intrinsic::nvvm_suld_2d_array_v2i8_trap: |
| 2155 | return NVPTXISD::Suld2DArrayV2I8Trap; |
| 2156 | case Intrinsic::nvvm_suld_2d_array_v2i16_trap: |
| 2157 | return NVPTXISD::Suld2DArrayV2I16Trap; |
| 2158 | case Intrinsic::nvvm_suld_2d_array_v2i32_trap: |
| 2159 | return NVPTXISD::Suld2DArrayV2I32Trap; |
| 2160 | case Intrinsic::nvvm_suld_2d_array_v4i8_trap: |
| 2161 | return NVPTXISD::Suld2DArrayV4I8Trap; |
| 2162 | case Intrinsic::nvvm_suld_2d_array_v4i16_trap: |
| 2163 | return NVPTXISD::Suld2DArrayV4I16Trap; |
| 2164 | case Intrinsic::nvvm_suld_2d_array_v4i32_trap: |
| 2165 | return NVPTXISD::Suld2DArrayV4I32Trap; |
| 2166 | case Intrinsic::nvvm_suld_3d_i8_trap: |
| 2167 | return NVPTXISD::Suld3DI8Trap; |
| 2168 | case Intrinsic::nvvm_suld_3d_i16_trap: |
| 2169 | return NVPTXISD::Suld3DI16Trap; |
| 2170 | case Intrinsic::nvvm_suld_3d_i32_trap: |
| 2171 | return NVPTXISD::Suld3DI32Trap; |
| 2172 | case Intrinsic::nvvm_suld_3d_v2i8_trap: |
| 2173 | return NVPTXISD::Suld3DV2I8Trap; |
| 2174 | case Intrinsic::nvvm_suld_3d_v2i16_trap: |
| 2175 | return NVPTXISD::Suld3DV2I16Trap; |
| 2176 | case Intrinsic::nvvm_suld_3d_v2i32_trap: |
| 2177 | return NVPTXISD::Suld3DV2I32Trap; |
| 2178 | case Intrinsic::nvvm_suld_3d_v4i8_trap: |
| 2179 | return NVPTXISD::Suld3DV4I8Trap; |
| 2180 | case Intrinsic::nvvm_suld_3d_v4i16_trap: |
| 2181 | return NVPTXISD::Suld3DV4I16Trap; |
| 2182 | case Intrinsic::nvvm_suld_3d_v4i32_trap: |
| 2183 | return NVPTXISD::Suld3DV4I32Trap; |
| 2184 | } |
| 2185 | } |
| 2186 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2187 | // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as |
| 2188 | // TgtMemIntrinsic |
| 2189 | // because we need the information that is only available in the "Value" type |
| 2190 | // of destination |
| 2191 | // pointer. In particular, the address space information. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2192 | bool NVPTXTargetLowering::getTgtMemIntrinsic( |
| 2193 | IntrinsicInfo &Info, const CallInst &I, unsigned Intrinsic) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2194 | switch (Intrinsic) { |
| 2195 | default: |
| 2196 | return false; |
| 2197 | |
| 2198 | case Intrinsic::nvvm_atomic_load_add_f32: |
| 2199 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 2200 | Info.memVT = MVT::f32; |
| 2201 | Info.ptrVal = I.getArgOperand(0); |
| 2202 | Info.offset = 0; |
| 2203 | Info.vol = 0; |
| 2204 | Info.readMem = true; |
| 2205 | Info.writeMem = true; |
| 2206 | Info.align = 0; |
| 2207 | return true; |
| 2208 | |
| 2209 | case Intrinsic::nvvm_atomic_load_inc_32: |
| 2210 | case Intrinsic::nvvm_atomic_load_dec_32: |
| 2211 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 2212 | Info.memVT = MVT::i32; |
| 2213 | Info.ptrVal = I.getArgOperand(0); |
| 2214 | Info.offset = 0; |
| 2215 | Info.vol = 0; |
| 2216 | Info.readMem = true; |
| 2217 | Info.writeMem = true; |
| 2218 | Info.align = 0; |
| 2219 | return true; |
| 2220 | |
| 2221 | case Intrinsic::nvvm_ldu_global_i: |
| 2222 | case Intrinsic::nvvm_ldu_global_f: |
| 2223 | case Intrinsic::nvvm_ldu_global_p: |
| 2224 | |
| 2225 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 2226 | if (Intrinsic == Intrinsic::nvvm_ldu_global_i) |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2227 | Info.memVT = getValueType(I.getType()); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2228 | else if (Intrinsic == Intrinsic::nvvm_ldu_global_p) |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2229 | Info.memVT = getValueType(I.getType()); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2230 | else |
| 2231 | Info.memVT = MVT::f32; |
| 2232 | Info.ptrVal = I.getArgOperand(0); |
| 2233 | Info.offset = 0; |
| 2234 | Info.vol = 0; |
| 2235 | Info.readMem = true; |
| 2236 | Info.writeMem = false; |
| 2237 | Info.align = 0; |
| 2238 | return true; |
| 2239 | |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 2240 | case Intrinsic::nvvm_tex_1d_v4f32_i32: |
| 2241 | case Intrinsic::nvvm_tex_1d_v4f32_f32: |
| 2242 | case Intrinsic::nvvm_tex_1d_level_v4f32_f32: |
| 2243 | case Intrinsic::nvvm_tex_1d_grad_v4f32_f32: |
| 2244 | case Intrinsic::nvvm_tex_1d_array_v4f32_i32: |
| 2245 | case Intrinsic::nvvm_tex_1d_array_v4f32_f32: |
| 2246 | case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32: |
| 2247 | case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32: |
| 2248 | case Intrinsic::nvvm_tex_2d_v4f32_i32: |
| 2249 | case Intrinsic::nvvm_tex_2d_v4f32_f32: |
| 2250 | case Intrinsic::nvvm_tex_2d_level_v4f32_f32: |
| 2251 | case Intrinsic::nvvm_tex_2d_grad_v4f32_f32: |
| 2252 | case Intrinsic::nvvm_tex_2d_array_v4f32_i32: |
| 2253 | case Intrinsic::nvvm_tex_2d_array_v4f32_f32: |
| 2254 | case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32: |
| 2255 | case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32: |
| 2256 | case Intrinsic::nvvm_tex_3d_v4f32_i32: |
| 2257 | case Intrinsic::nvvm_tex_3d_v4f32_f32: |
| 2258 | case Intrinsic::nvvm_tex_3d_level_v4f32_f32: |
| 2259 | case Intrinsic::nvvm_tex_3d_grad_v4f32_f32: { |
| 2260 | Info.opc = getOpcForTextureInstr(Intrinsic); |
| 2261 | Info.memVT = MVT::f32; |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2262 | Info.ptrVal = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 2263 | Info.offset = 0; |
| 2264 | Info.vol = 0; |
| 2265 | Info.readMem = true; |
| 2266 | Info.writeMem = false; |
| 2267 | Info.align = 16; |
| 2268 | return true; |
| 2269 | } |
| 2270 | case Intrinsic::nvvm_tex_1d_v4i32_i32: |
| 2271 | case Intrinsic::nvvm_tex_1d_v4i32_f32: |
| 2272 | case Intrinsic::nvvm_tex_1d_level_v4i32_f32: |
| 2273 | case Intrinsic::nvvm_tex_1d_grad_v4i32_f32: |
| 2274 | case Intrinsic::nvvm_tex_1d_array_v4i32_i32: |
| 2275 | case Intrinsic::nvvm_tex_1d_array_v4i32_f32: |
| 2276 | case Intrinsic::nvvm_tex_1d_array_level_v4i32_f32: |
| 2277 | case Intrinsic::nvvm_tex_1d_array_grad_v4i32_f32: |
| 2278 | case Intrinsic::nvvm_tex_2d_v4i32_i32: |
| 2279 | case Intrinsic::nvvm_tex_2d_v4i32_f32: |
| 2280 | case Intrinsic::nvvm_tex_2d_level_v4i32_f32: |
| 2281 | case Intrinsic::nvvm_tex_2d_grad_v4i32_f32: |
| 2282 | case Intrinsic::nvvm_tex_2d_array_v4i32_i32: |
| 2283 | case Intrinsic::nvvm_tex_2d_array_v4i32_f32: |
| 2284 | case Intrinsic::nvvm_tex_2d_array_level_v4i32_f32: |
| 2285 | case Intrinsic::nvvm_tex_2d_array_grad_v4i32_f32: |
| 2286 | case Intrinsic::nvvm_tex_3d_v4i32_i32: |
| 2287 | case Intrinsic::nvvm_tex_3d_v4i32_f32: |
| 2288 | case Intrinsic::nvvm_tex_3d_level_v4i32_f32: |
| 2289 | case Intrinsic::nvvm_tex_3d_grad_v4i32_f32: { |
| 2290 | Info.opc = getOpcForTextureInstr(Intrinsic); |
| 2291 | Info.memVT = MVT::i32; |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2292 | Info.ptrVal = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 2293 | Info.offset = 0; |
| 2294 | Info.vol = 0; |
| 2295 | Info.readMem = true; |
| 2296 | Info.writeMem = false; |
| 2297 | Info.align = 16; |
| 2298 | return true; |
| 2299 | } |
| 2300 | case Intrinsic::nvvm_suld_1d_i8_trap: |
| 2301 | case Intrinsic::nvvm_suld_1d_v2i8_trap: |
| 2302 | case Intrinsic::nvvm_suld_1d_v4i8_trap: |
| 2303 | case Intrinsic::nvvm_suld_1d_array_i8_trap: |
| 2304 | case Intrinsic::nvvm_suld_1d_array_v2i8_trap: |
| 2305 | case Intrinsic::nvvm_suld_1d_array_v4i8_trap: |
| 2306 | case Intrinsic::nvvm_suld_2d_i8_trap: |
| 2307 | case Intrinsic::nvvm_suld_2d_v2i8_trap: |
| 2308 | case Intrinsic::nvvm_suld_2d_v4i8_trap: |
| 2309 | case Intrinsic::nvvm_suld_2d_array_i8_trap: |
| 2310 | case Intrinsic::nvvm_suld_2d_array_v2i8_trap: |
| 2311 | case Intrinsic::nvvm_suld_2d_array_v4i8_trap: |
| 2312 | case Intrinsic::nvvm_suld_3d_i8_trap: |
| 2313 | case Intrinsic::nvvm_suld_3d_v2i8_trap: |
| 2314 | case Intrinsic::nvvm_suld_3d_v4i8_trap: { |
| 2315 | Info.opc = getOpcForSurfaceInstr(Intrinsic); |
| 2316 | Info.memVT = MVT::i8; |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2317 | Info.ptrVal = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 2318 | Info.offset = 0; |
| 2319 | Info.vol = 0; |
| 2320 | Info.readMem = true; |
| 2321 | Info.writeMem = false; |
| 2322 | Info.align = 16; |
| 2323 | return true; |
| 2324 | } |
| 2325 | case Intrinsic::nvvm_suld_1d_i16_trap: |
| 2326 | case Intrinsic::nvvm_suld_1d_v2i16_trap: |
| 2327 | case Intrinsic::nvvm_suld_1d_v4i16_trap: |
| 2328 | case Intrinsic::nvvm_suld_1d_array_i16_trap: |
| 2329 | case Intrinsic::nvvm_suld_1d_array_v2i16_trap: |
| 2330 | case Intrinsic::nvvm_suld_1d_array_v4i16_trap: |
| 2331 | case Intrinsic::nvvm_suld_2d_i16_trap: |
| 2332 | case Intrinsic::nvvm_suld_2d_v2i16_trap: |
| 2333 | case Intrinsic::nvvm_suld_2d_v4i16_trap: |
| 2334 | case Intrinsic::nvvm_suld_2d_array_i16_trap: |
| 2335 | case Intrinsic::nvvm_suld_2d_array_v2i16_trap: |
| 2336 | case Intrinsic::nvvm_suld_2d_array_v4i16_trap: |
| 2337 | case Intrinsic::nvvm_suld_3d_i16_trap: |
| 2338 | case Intrinsic::nvvm_suld_3d_v2i16_trap: |
| 2339 | case Intrinsic::nvvm_suld_3d_v4i16_trap: { |
| 2340 | Info.opc = getOpcForSurfaceInstr(Intrinsic); |
| 2341 | Info.memVT = MVT::i16; |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2342 | Info.ptrVal = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 2343 | Info.offset = 0; |
| 2344 | Info.vol = 0; |
| 2345 | Info.readMem = true; |
| 2346 | Info.writeMem = false; |
| 2347 | Info.align = 16; |
| 2348 | return true; |
| 2349 | } |
| 2350 | case Intrinsic::nvvm_suld_1d_i32_trap: |
| 2351 | case Intrinsic::nvvm_suld_1d_v2i32_trap: |
| 2352 | case Intrinsic::nvvm_suld_1d_v4i32_trap: |
| 2353 | case Intrinsic::nvvm_suld_1d_array_i32_trap: |
| 2354 | case Intrinsic::nvvm_suld_1d_array_v2i32_trap: |
| 2355 | case Intrinsic::nvvm_suld_1d_array_v4i32_trap: |
| 2356 | case Intrinsic::nvvm_suld_2d_i32_trap: |
| 2357 | case Intrinsic::nvvm_suld_2d_v2i32_trap: |
| 2358 | case Intrinsic::nvvm_suld_2d_v4i32_trap: |
| 2359 | case Intrinsic::nvvm_suld_2d_array_i32_trap: |
| 2360 | case Intrinsic::nvvm_suld_2d_array_v2i32_trap: |
| 2361 | case Intrinsic::nvvm_suld_2d_array_v4i32_trap: |
| 2362 | case Intrinsic::nvvm_suld_3d_i32_trap: |
| 2363 | case Intrinsic::nvvm_suld_3d_v2i32_trap: |
| 2364 | case Intrinsic::nvvm_suld_3d_v4i32_trap: { |
| 2365 | Info.opc = getOpcForSurfaceInstr(Intrinsic); |
| 2366 | Info.memVT = MVT::i32; |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2367 | Info.ptrVal = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 2368 | Info.offset = 0; |
| 2369 | Info.vol = 0; |
| 2370 | Info.readMem = true; |
| 2371 | Info.writeMem = false; |
| 2372 | Info.align = 16; |
| 2373 | return true; |
| 2374 | } |
| 2375 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2376 | } |
| 2377 | return false; |
| 2378 | } |
| 2379 | |
| 2380 | /// isLegalAddressingMode - Return true if the addressing mode represented |
| 2381 | /// by AM is legal for this target, for a load/store of the specified type. |
| 2382 | /// Used to guide target specific optimizations, like loop strength reduction |
| 2383 | /// (LoopStrengthReduce.cpp) and memory optimization for address mode |
| 2384 | /// (CodeGenPrepare.cpp) |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2385 | bool NVPTXTargetLowering::isLegalAddressingMode(const AddrMode &AM, |
| 2386 | Type *Ty) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2387 | |
| 2388 | // AddrMode - This represents an addressing mode of: |
| 2389 | // BaseGV + BaseOffs + BaseReg + Scale*ScaleReg |
| 2390 | // |
| 2391 | // The legal address modes are |
| 2392 | // - [avar] |
| 2393 | // - [areg] |
| 2394 | // - [areg+immoff] |
| 2395 | // - [immAddr] |
| 2396 | |
| 2397 | if (AM.BaseGV) { |
| 2398 | if (AM.BaseOffs || AM.HasBaseReg || AM.Scale) |
| 2399 | return false; |
| 2400 | return true; |
| 2401 | } |
| 2402 | |
| 2403 | switch (AM.Scale) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2404 | case 0: // "r", "r+i" or "i" is allowed |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2405 | break; |
| 2406 | case 1: |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2407 | if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed. |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2408 | return false; |
| 2409 | // Otherwise we have r+i. |
| 2410 | break; |
| 2411 | default: |
| 2412 | // No scale > 1 is allowed |
| 2413 | return false; |
| 2414 | } |
| 2415 | return true; |
| 2416 | } |
| 2417 | |
| 2418 | //===----------------------------------------------------------------------===// |
| 2419 | // NVPTX Inline Assembly Support |
| 2420 | //===----------------------------------------------------------------------===// |
| 2421 | |
| 2422 | /// getConstraintType - Given a constraint letter, return the type of |
| 2423 | /// constraint it is for this target. |
| 2424 | NVPTXTargetLowering::ConstraintType |
| 2425 | NVPTXTargetLowering::getConstraintType(const std::string &Constraint) const { |
| 2426 | if (Constraint.size() == 1) { |
| 2427 | switch (Constraint[0]) { |
| 2428 | default: |
| 2429 | break; |
| 2430 | case 'r': |
| 2431 | case 'h': |
| 2432 | case 'c': |
| 2433 | case 'l': |
| 2434 | case 'f': |
| 2435 | case 'd': |
| 2436 | case '0': |
| 2437 | case 'N': |
| 2438 | return C_RegisterClass; |
| 2439 | } |
| 2440 | } |
| 2441 | return TargetLowering::getConstraintType(Constraint); |
| 2442 | } |
| 2443 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2444 | std::pair<unsigned, const TargetRegisterClass *> |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2445 | NVPTXTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, |
Chad Rosier | 295bd43 | 2013-06-22 18:37:38 +0000 | [diff] [blame] | 2446 | MVT VT) const { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2447 | if (Constraint.size() == 1) { |
| 2448 | switch (Constraint[0]) { |
| 2449 | case 'c': |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2450 | return std::make_pair(0U, &NVPTX::Int16RegsRegClass); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2451 | case 'h': |
| 2452 | return std::make_pair(0U, &NVPTX::Int16RegsRegClass); |
| 2453 | case 'r': |
| 2454 | return std::make_pair(0U, &NVPTX::Int32RegsRegClass); |
| 2455 | case 'l': |
| 2456 | case 'N': |
| 2457 | return std::make_pair(0U, &NVPTX::Int64RegsRegClass); |
| 2458 | case 'f': |
| 2459 | return std::make_pair(0U, &NVPTX::Float32RegsRegClass); |
| 2460 | case 'd': |
| 2461 | return std::make_pair(0U, &NVPTX::Float64RegsRegClass); |
| 2462 | } |
| 2463 | } |
| 2464 | return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); |
| 2465 | } |
| 2466 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2467 | /// getFunctionAlignment - Return the Log2 alignment of this function. |
| 2468 | unsigned NVPTXTargetLowering::getFunctionAlignment(const Function *) const { |
| 2469 | return 4; |
| 2470 | } |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2471 | |
| 2472 | /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads. |
| 2473 | static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG, |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2474 | SmallVectorImpl<SDValue> &Results) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2475 | EVT ResVT = N->getValueType(0); |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 2476 | SDLoc DL(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2477 | |
| 2478 | assert(ResVT.isVector() && "Vector load must have vector type"); |
| 2479 | |
| 2480 | // We only handle "native" vector sizes for now, e.g. <4 x double> is not |
| 2481 | // legal. We can (and should) split that into 2 loads of <2 x double> here |
| 2482 | // but I'm leaving that as a TODO for now. |
| 2483 | assert(ResVT.isSimple() && "Can only handle simple types"); |
| 2484 | switch (ResVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2485 | default: |
| 2486 | return; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2487 | case MVT::v2i8: |
| 2488 | case MVT::v2i16: |
| 2489 | case MVT::v2i32: |
| 2490 | case MVT::v2i64: |
| 2491 | case MVT::v2f32: |
| 2492 | case MVT::v2f64: |
| 2493 | case MVT::v4i8: |
| 2494 | case MVT::v4i16: |
| 2495 | case MVT::v4i32: |
| 2496 | case MVT::v4f32: |
| 2497 | // This is a "native" vector type |
| 2498 | break; |
| 2499 | } |
| 2500 | |
| 2501 | EVT EltVT = ResVT.getVectorElementType(); |
| 2502 | unsigned NumElts = ResVT.getVectorNumElements(); |
| 2503 | |
| 2504 | // Since LoadV2 is a target node, we cannot rely on DAG type legalization. |
| 2505 | // Therefore, we must ensure the type is legal. For i1 and i8, we set the |
Alp Toker | cb40291 | 2014-01-24 17:20:08 +0000 | [diff] [blame] | 2506 | // loaded type to i16 and propagate the "real" type as the memory type. |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2507 | bool NeedTrunc = false; |
| 2508 | if (EltVT.getSizeInBits() < 16) { |
| 2509 | EltVT = MVT::i16; |
| 2510 | NeedTrunc = true; |
| 2511 | } |
| 2512 | |
| 2513 | unsigned Opcode = 0; |
| 2514 | SDVTList LdResVTs; |
| 2515 | |
| 2516 | switch (NumElts) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2517 | default: |
| 2518 | return; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2519 | case 2: |
| 2520 | Opcode = NVPTXISD::LoadV2; |
| 2521 | LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other); |
| 2522 | break; |
| 2523 | case 4: { |
| 2524 | Opcode = NVPTXISD::LoadV4; |
| 2525 | EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other }; |
Craig Topper | abb4ac7 | 2014-04-16 06:10:51 +0000 | [diff] [blame] | 2526 | LdResVTs = DAG.getVTList(ListVTs); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2527 | break; |
| 2528 | } |
| 2529 | } |
| 2530 | |
| 2531 | SmallVector<SDValue, 8> OtherOps; |
| 2532 | |
| 2533 | // Copy regular operands |
| 2534 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) |
| 2535 | OtherOps.push_back(N->getOperand(i)); |
| 2536 | |
| 2537 | LoadSDNode *LD = cast<LoadSDNode>(N); |
| 2538 | |
| 2539 | // The select routine does not have access to the LoadSDNode instance, so |
| 2540 | // pass along the extension information |
| 2541 | OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType())); |
| 2542 | |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 2543 | SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps, |
| 2544 | LD->getMemoryVT(), |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2545 | LD->getMemOperand()); |
| 2546 | |
| 2547 | SmallVector<SDValue, 4> ScalarRes; |
| 2548 | |
| 2549 | for (unsigned i = 0; i < NumElts; ++i) { |
| 2550 | SDValue Res = NewLD.getValue(i); |
| 2551 | if (NeedTrunc) |
| 2552 | Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res); |
| 2553 | ScalarRes.push_back(Res); |
| 2554 | } |
| 2555 | |
| 2556 | SDValue LoadChain = NewLD.getValue(NumElts); |
| 2557 | |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 2558 | SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2559 | |
| 2560 | Results.push_back(BuildVec); |
| 2561 | Results.push_back(LoadChain); |
| 2562 | } |
| 2563 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2564 | static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG, |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2565 | SmallVectorImpl<SDValue> &Results) { |
| 2566 | SDValue Chain = N->getOperand(0); |
| 2567 | SDValue Intrin = N->getOperand(1); |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 2568 | SDLoc DL(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2569 | |
| 2570 | // Get the intrinsic ID |
| 2571 | unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue(); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2572 | switch (IntrinNo) { |
| 2573 | default: |
| 2574 | return; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2575 | case Intrinsic::nvvm_ldg_global_i: |
| 2576 | case Intrinsic::nvvm_ldg_global_f: |
| 2577 | case Intrinsic::nvvm_ldg_global_p: |
| 2578 | case Intrinsic::nvvm_ldu_global_i: |
| 2579 | case Intrinsic::nvvm_ldu_global_f: |
| 2580 | case Intrinsic::nvvm_ldu_global_p: { |
| 2581 | EVT ResVT = N->getValueType(0); |
| 2582 | |
| 2583 | if (ResVT.isVector()) { |
| 2584 | // Vector LDG/LDU |
| 2585 | |
| 2586 | unsigned NumElts = ResVT.getVectorNumElements(); |
| 2587 | EVT EltVT = ResVT.getVectorElementType(); |
| 2588 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2589 | // Since LDU/LDG are target nodes, we cannot rely on DAG type |
| 2590 | // legalization. |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2591 | // Therefore, we must ensure the type is legal. For i1 and i8, we set the |
Alp Toker | cb40291 | 2014-01-24 17:20:08 +0000 | [diff] [blame] | 2592 | // loaded type to i16 and propagate the "real" type as the memory type. |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2593 | bool NeedTrunc = false; |
| 2594 | if (EltVT.getSizeInBits() < 16) { |
| 2595 | EltVT = MVT::i16; |
| 2596 | NeedTrunc = true; |
| 2597 | } |
| 2598 | |
| 2599 | unsigned Opcode = 0; |
| 2600 | SDVTList LdResVTs; |
| 2601 | |
| 2602 | switch (NumElts) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2603 | default: |
| 2604 | return; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2605 | case 2: |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2606 | switch (IntrinNo) { |
| 2607 | default: |
| 2608 | return; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2609 | case Intrinsic::nvvm_ldg_global_i: |
| 2610 | case Intrinsic::nvvm_ldg_global_f: |
| 2611 | case Intrinsic::nvvm_ldg_global_p: |
| 2612 | Opcode = NVPTXISD::LDGV2; |
| 2613 | break; |
| 2614 | case Intrinsic::nvvm_ldu_global_i: |
| 2615 | case Intrinsic::nvvm_ldu_global_f: |
| 2616 | case Intrinsic::nvvm_ldu_global_p: |
| 2617 | Opcode = NVPTXISD::LDUV2; |
| 2618 | break; |
| 2619 | } |
| 2620 | LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other); |
| 2621 | break; |
| 2622 | case 4: { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2623 | switch (IntrinNo) { |
| 2624 | default: |
| 2625 | return; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2626 | case Intrinsic::nvvm_ldg_global_i: |
| 2627 | case Intrinsic::nvvm_ldg_global_f: |
| 2628 | case Intrinsic::nvvm_ldg_global_p: |
| 2629 | Opcode = NVPTXISD::LDGV4; |
| 2630 | break; |
| 2631 | case Intrinsic::nvvm_ldu_global_i: |
| 2632 | case Intrinsic::nvvm_ldu_global_f: |
| 2633 | case Intrinsic::nvvm_ldu_global_p: |
| 2634 | Opcode = NVPTXISD::LDUV4; |
| 2635 | break; |
| 2636 | } |
| 2637 | EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other }; |
Craig Topper | abb4ac7 | 2014-04-16 06:10:51 +0000 | [diff] [blame] | 2638 | LdResVTs = DAG.getVTList(ListVTs); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2639 | break; |
| 2640 | } |
| 2641 | } |
| 2642 | |
| 2643 | SmallVector<SDValue, 8> OtherOps; |
| 2644 | |
| 2645 | // Copy regular operands |
| 2646 | |
| 2647 | OtherOps.push_back(Chain); // Chain |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2648 | // Skip operand 1 (intrinsic ID) |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2649 | // Others |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2650 | for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) |
| 2651 | OtherOps.push_back(N->getOperand(i)); |
| 2652 | |
| 2653 | MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N); |
| 2654 | |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 2655 | SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps, |
| 2656 | MemSD->getMemoryVT(), |
| 2657 | MemSD->getMemOperand()); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2658 | |
| 2659 | SmallVector<SDValue, 4> ScalarRes; |
| 2660 | |
| 2661 | for (unsigned i = 0; i < NumElts; ++i) { |
| 2662 | SDValue Res = NewLD.getValue(i); |
| 2663 | if (NeedTrunc) |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2664 | Res = |
| 2665 | DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2666 | ScalarRes.push_back(Res); |
| 2667 | } |
| 2668 | |
| 2669 | SDValue LoadChain = NewLD.getValue(NumElts); |
| 2670 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2671 | SDValue BuildVec = |
Craig Topper | 48d114b | 2014-04-26 18:35:24 +0000 | [diff] [blame] | 2672 | DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2673 | |
| 2674 | Results.push_back(BuildVec); |
| 2675 | Results.push_back(LoadChain); |
| 2676 | } else { |
| 2677 | // i8 LDG/LDU |
| 2678 | assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 && |
| 2679 | "Custom handling of non-i8 ldu/ldg?"); |
| 2680 | |
| 2681 | // Just copy all operands as-is |
| 2682 | SmallVector<SDValue, 4> Ops; |
| 2683 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) |
| 2684 | Ops.push_back(N->getOperand(i)); |
| 2685 | |
| 2686 | // Force output to i16 |
| 2687 | SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other); |
| 2688 | |
| 2689 | MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N); |
| 2690 | |
| 2691 | // We make sure the memory type is i8, which will be used during isel |
| 2692 | // to select the proper instruction. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2693 | SDValue NewLD = |
Craig Topper | 206fcd4 | 2014-04-26 19:29:41 +0000 | [diff] [blame] | 2694 | DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops, |
| 2695 | MVT::i8, MemSD->getMemOperand()); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2696 | |
Justin Holewinski | e8c93e3 | 2013-07-01 12:58:48 +0000 | [diff] [blame] | 2697 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, |
| 2698 | NewLD.getValue(0))); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2699 | Results.push_back(NewLD.getValue(1)); |
| 2700 | } |
| 2701 | } |
| 2702 | } |
| 2703 | } |
| 2704 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2705 | void NVPTXTargetLowering::ReplaceNodeResults( |
| 2706 | SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2707 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2708 | default: |
| 2709 | report_fatal_error("Unhandled custom legalization"); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2710 | case ISD::LOAD: |
| 2711 | ReplaceLoadVector(N, DAG, Results); |
| 2712 | return; |
| 2713 | case ISD::INTRINSIC_W_CHAIN: |
| 2714 | ReplaceINTRINSIC_W_CHAIN(N, DAG, Results); |
| 2715 | return; |
| 2716 | } |
| 2717 | } |
Juergen Ributzka | d12ccbd | 2013-11-19 00:57:56 +0000 | [diff] [blame] | 2718 | |
| 2719 | // Pin NVPTXSection's and NVPTXTargetObjectFile's vtables to this file. |
| 2720 | void NVPTXSection::anchor() {} |
| 2721 | |
| 2722 | NVPTXTargetObjectFile::~NVPTXTargetObjectFile() { |
| 2723 | delete TextSection; |
| 2724 | delete DataSection; |
| 2725 | delete BSSSection; |
| 2726 | delete ReadOnlySection; |
| 2727 | |
| 2728 | delete StaticCtorSection; |
| 2729 | delete StaticDtorSection; |
| 2730 | delete LSDASection; |
| 2731 | delete EHFrameSection; |
| 2732 | delete DwarfAbbrevSection; |
| 2733 | delete DwarfInfoSection; |
| 2734 | delete DwarfLineSection; |
| 2735 | delete DwarfFrameSection; |
| 2736 | delete DwarfPubTypesSection; |
| 2737 | delete DwarfDebugInlineSection; |
| 2738 | delete DwarfStrSection; |
| 2739 | delete DwarfLocSection; |
| 2740 | delete DwarfARangesSection; |
| 2741 | delete DwarfRangesSection; |
| 2742 | delete DwarfMacroInfoSection; |
| 2743 | } |