Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 1 | //===-- NVPTXISelDAGToDAG.cpp - A dag to dag inst selector for NVPTX ------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This file defines an instruction selector for the NVPTX target. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 14 | #include "NVPTXISelDAGToDAG.h" |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 15 | #include "llvm/Analysis/ValueTracking.h" |
Chandler Carruth | 9fb823b | 2013-01-02 11:36:10 +0000 | [diff] [blame] | 16 | #include "llvm/IR/GlobalValue.h" |
| 17 | #include "llvm/IR/Instructions.h" |
Chandler Carruth | ed0881b | 2012-12-03 16:50:05 +0000 | [diff] [blame] | 18 | #include "llvm/Support/CommandLine.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 19 | #include "llvm/Support/Debug.h" |
| 20 | #include "llvm/Support/ErrorHandling.h" |
Chandler Carruth | ed0881b | 2012-12-03 16:50:05 +0000 | [diff] [blame] | 21 | #include "llvm/Support/raw_ostream.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 22 | #include "llvm/Target/TargetIntrinsicInfo.h" |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 23 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 24 | using namespace llvm; |
| 25 | |
Chandler Carruth | 84e68b2 | 2014-04-22 02:41:26 +0000 | [diff] [blame] | 26 | #define DEBUG_TYPE "nvptx-isel" |
| 27 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 28 | static cl::opt<int> UsePrecDivF32( |
Nadav Rotem | 7f27e0b | 2013-10-18 23:38:13 +0000 | [diff] [blame] | 29 | "nvptx-prec-divf32", cl::ZeroOrMore, cl::Hidden, |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 30 | cl::desc("NVPTX Specifies: 0 use div.approx, 1 use div.full, 2 use" |
Sylvestre Ledru | 469de19 | 2014-08-11 18:04:46 +0000 | [diff] [blame] | 31 | " IEEE Compliant F32 div.rnd if available."), |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 32 | cl::init(2)); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 33 | |
Justin Holewinski | 48f4ad3 | 2013-05-21 16:51:30 +0000 | [diff] [blame] | 34 | static cl::opt<bool> |
Nadav Rotem | 7f27e0b | 2013-10-18 23:38:13 +0000 | [diff] [blame] | 35 | UsePrecSqrtF32("nvptx-prec-sqrtf32", cl::Hidden, |
Justin Holewinski | 48f4ad3 | 2013-05-21 16:51:30 +0000 | [diff] [blame] | 36 | cl::desc("NVPTX Specific: 0 use sqrt.approx, 1 use sqrt.rn."), |
| 37 | cl::init(true)); |
| 38 | |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 39 | static cl::opt<bool> |
Nadav Rotem | 7f27e0b | 2013-10-18 23:38:13 +0000 | [diff] [blame] | 40 | FtzEnabled("nvptx-f32ftz", cl::ZeroOrMore, cl::Hidden, |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 41 | cl::desc("NVPTX Specific: Flush f32 subnormals to sign-preserving zero."), |
| 42 | cl::init(false)); |
| 43 | |
| 44 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 45 | /// createNVPTXISelDag - This pass converts a legalized DAG into a |
| 46 | /// NVPTX-specific DAG, ready for instruction scheduling. |
| 47 | FunctionPass *llvm::createNVPTXISelDag(NVPTXTargetMachine &TM, |
| 48 | llvm::CodeGenOpt::Level OptLevel) { |
| 49 | return new NVPTXDAGToDAGISel(TM, OptLevel); |
| 50 | } |
| 51 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 52 | NVPTXDAGToDAGISel::NVPTXDAGToDAGISel(NVPTXTargetMachine &tm, |
| 53 | CodeGenOpt::Level OptLevel) |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 54 | : SelectionDAGISel(tm, OptLevel), TM(tm) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 55 | doMulWide = (OptLevel > 0); |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 56 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 57 | |
Eric Christopher | 147bba2 | 2015-01-30 01:40:59 +0000 | [diff] [blame] | 58 | bool NVPTXDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) { |
| 59 | Subtarget = &static_cast<const NVPTXSubtarget &>(MF.getSubtarget()); |
| 60 | return SelectionDAGISel::runOnMachineFunction(MF); |
| 61 | } |
| 62 | |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 63 | int NVPTXDAGToDAGISel::getDivF32Level() const { |
| 64 | if (UsePrecDivF32.getNumOccurrences() > 0) { |
| 65 | // If nvptx-prec-div32=N is used on the command-line, always honor it |
| 66 | return UsePrecDivF32; |
| 67 | } else { |
| 68 | // Otherwise, use div.approx if fast math is enabled |
| 69 | if (TM.Options.UnsafeFPMath) |
| 70 | return 0; |
| 71 | else |
| 72 | return 2; |
| 73 | } |
| 74 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 75 | |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 76 | bool NVPTXDAGToDAGISel::usePrecSqrtF32() const { |
| 77 | if (UsePrecSqrtF32.getNumOccurrences() > 0) { |
| 78 | // If nvptx-prec-sqrtf32 is used on the command-line, always honor it |
| 79 | return UsePrecSqrtF32; |
| 80 | } else { |
| 81 | // Otherwise, use sqrt.approx if fast math is enabled |
Eli Bendersky | 3e84019 | 2015-03-23 16:26:23 +0000 | [diff] [blame] | 82 | return !TM.Options.UnsafeFPMath; |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 83 | } |
| 84 | } |
| 85 | |
| 86 | bool NVPTXDAGToDAGISel::useF32FTZ() const { |
| 87 | if (FtzEnabled.getNumOccurrences() > 0) { |
| 88 | // If nvptx-f32ftz is used on the command-line, always honor it |
| 89 | return FtzEnabled; |
| 90 | } else { |
| 91 | const Function *F = MF->getFunction(); |
| 92 | // Otherwise, check for an nvptx-f32ftz attribute on the function |
| 93 | if (F->hasFnAttribute("nvptx-f32ftz")) |
Duncan P. N. Exon Smith | b505433 | 2015-02-14 15:35:43 +0000 | [diff] [blame] | 94 | return F->getFnAttribute("nvptx-f32ftz").getValueAsString() == "true"; |
Justin Holewinski | cd069e6 | 2013-07-22 12:18:04 +0000 | [diff] [blame] | 95 | else |
| 96 | return false; |
| 97 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 98 | } |
| 99 | |
Justin Holewinski | 428cf0e | 2014-07-17 18:10:09 +0000 | [diff] [blame] | 100 | bool NVPTXDAGToDAGISel::allowFMA() const { |
Eric Christopher | 147bba2 | 2015-01-30 01:40:59 +0000 | [diff] [blame] | 101 | const NVPTXTargetLowering *TL = Subtarget->getTargetLowering(); |
Justin Holewinski | 428cf0e | 2014-07-17 18:10:09 +0000 | [diff] [blame] | 102 | return TL->allowFMA(*MF, OptLevel); |
| 103 | } |
| 104 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 105 | /// Select - Select instructions not customized! Used for |
| 106 | /// expanded, promoted and normal instructions. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 107 | SDNode *NVPTXDAGToDAGISel::Select(SDNode *N) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 108 | |
Tim Northover | 31d093c | 2013-09-22 08:21:56 +0000 | [diff] [blame] | 109 | if (N->isMachineOpcode()) { |
| 110 | N->setNodeId(-1); |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 111 | return nullptr; // Already selected. |
Tim Northover | 31d093c | 2013-09-22 08:21:56 +0000 | [diff] [blame] | 112 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 113 | |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 114 | SDNode *ResNode = nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 115 | switch (N->getOpcode()) { |
| 116 | case ISD::LOAD: |
| 117 | ResNode = SelectLoad(N); |
| 118 | break; |
| 119 | case ISD::STORE: |
| 120 | ResNode = SelectStore(N); |
| 121 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 122 | case NVPTXISD::LoadV2: |
| 123 | case NVPTXISD::LoadV4: |
| 124 | ResNode = SelectLoadVector(N); |
| 125 | break; |
| 126 | case NVPTXISD::LDGV2: |
| 127 | case NVPTXISD::LDGV4: |
| 128 | case NVPTXISD::LDUV2: |
| 129 | case NVPTXISD::LDUV4: |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 130 | ResNode = SelectLDGLDU(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 131 | break; |
| 132 | case NVPTXISD::StoreV2: |
| 133 | case NVPTXISD::StoreV4: |
| 134 | ResNode = SelectStoreVector(N); |
| 135 | break; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 136 | case NVPTXISD::LoadParam: |
| 137 | case NVPTXISD::LoadParamV2: |
| 138 | case NVPTXISD::LoadParamV4: |
| 139 | ResNode = SelectLoadParam(N); |
| 140 | break; |
| 141 | case NVPTXISD::StoreRetval: |
| 142 | case NVPTXISD::StoreRetvalV2: |
| 143 | case NVPTXISD::StoreRetvalV4: |
| 144 | ResNode = SelectStoreRetval(N); |
| 145 | break; |
| 146 | case NVPTXISD::StoreParam: |
| 147 | case NVPTXISD::StoreParamV2: |
| 148 | case NVPTXISD::StoreParamV4: |
| 149 | case NVPTXISD::StoreParamS32: |
| 150 | case NVPTXISD::StoreParamU32: |
| 151 | ResNode = SelectStoreParam(N); |
| 152 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 153 | case ISD::INTRINSIC_WO_CHAIN: |
| 154 | ResNode = SelectIntrinsicNoChain(N); |
| 155 | break; |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 156 | case ISD::INTRINSIC_W_CHAIN: |
| 157 | ResNode = SelectIntrinsicChain(N); |
| 158 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 159 | case NVPTXISD::Tex1DFloatS32: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 160 | case NVPTXISD::Tex1DFloatFloat: |
| 161 | case NVPTXISD::Tex1DFloatFloatLevel: |
| 162 | case NVPTXISD::Tex1DFloatFloatGrad: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 163 | case NVPTXISD::Tex1DS32S32: |
| 164 | case NVPTXISD::Tex1DS32Float: |
| 165 | case NVPTXISD::Tex1DS32FloatLevel: |
| 166 | case NVPTXISD::Tex1DS32FloatGrad: |
| 167 | case NVPTXISD::Tex1DU32S32: |
| 168 | case NVPTXISD::Tex1DU32Float: |
| 169 | case NVPTXISD::Tex1DU32FloatLevel: |
| 170 | case NVPTXISD::Tex1DU32FloatGrad: |
| 171 | case NVPTXISD::Tex1DArrayFloatS32: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 172 | case NVPTXISD::Tex1DArrayFloatFloat: |
| 173 | case NVPTXISD::Tex1DArrayFloatFloatLevel: |
| 174 | case NVPTXISD::Tex1DArrayFloatFloatGrad: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 175 | case NVPTXISD::Tex1DArrayS32S32: |
| 176 | case NVPTXISD::Tex1DArrayS32Float: |
| 177 | case NVPTXISD::Tex1DArrayS32FloatLevel: |
| 178 | case NVPTXISD::Tex1DArrayS32FloatGrad: |
| 179 | case NVPTXISD::Tex1DArrayU32S32: |
| 180 | case NVPTXISD::Tex1DArrayU32Float: |
| 181 | case NVPTXISD::Tex1DArrayU32FloatLevel: |
| 182 | case NVPTXISD::Tex1DArrayU32FloatGrad: |
| 183 | case NVPTXISD::Tex2DFloatS32: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 184 | case NVPTXISD::Tex2DFloatFloat: |
| 185 | case NVPTXISD::Tex2DFloatFloatLevel: |
| 186 | case NVPTXISD::Tex2DFloatFloatGrad: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 187 | case NVPTXISD::Tex2DS32S32: |
| 188 | case NVPTXISD::Tex2DS32Float: |
| 189 | case NVPTXISD::Tex2DS32FloatLevel: |
| 190 | case NVPTXISD::Tex2DS32FloatGrad: |
| 191 | case NVPTXISD::Tex2DU32S32: |
| 192 | case NVPTXISD::Tex2DU32Float: |
| 193 | case NVPTXISD::Tex2DU32FloatLevel: |
| 194 | case NVPTXISD::Tex2DU32FloatGrad: |
| 195 | case NVPTXISD::Tex2DArrayFloatS32: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 196 | case NVPTXISD::Tex2DArrayFloatFloat: |
| 197 | case NVPTXISD::Tex2DArrayFloatFloatLevel: |
| 198 | case NVPTXISD::Tex2DArrayFloatFloatGrad: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 199 | case NVPTXISD::Tex2DArrayS32S32: |
| 200 | case NVPTXISD::Tex2DArrayS32Float: |
| 201 | case NVPTXISD::Tex2DArrayS32FloatLevel: |
| 202 | case NVPTXISD::Tex2DArrayS32FloatGrad: |
| 203 | case NVPTXISD::Tex2DArrayU32S32: |
| 204 | case NVPTXISD::Tex2DArrayU32Float: |
| 205 | case NVPTXISD::Tex2DArrayU32FloatLevel: |
| 206 | case NVPTXISD::Tex2DArrayU32FloatGrad: |
| 207 | case NVPTXISD::Tex3DFloatS32: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 208 | case NVPTXISD::Tex3DFloatFloat: |
| 209 | case NVPTXISD::Tex3DFloatFloatLevel: |
| 210 | case NVPTXISD::Tex3DFloatFloatGrad: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 211 | case NVPTXISD::Tex3DS32S32: |
| 212 | case NVPTXISD::Tex3DS32Float: |
| 213 | case NVPTXISD::Tex3DS32FloatLevel: |
| 214 | case NVPTXISD::Tex3DS32FloatGrad: |
| 215 | case NVPTXISD::Tex3DU32S32: |
| 216 | case NVPTXISD::Tex3DU32Float: |
| 217 | case NVPTXISD::Tex3DU32FloatLevel: |
| 218 | case NVPTXISD::Tex3DU32FloatGrad: |
| 219 | case NVPTXISD::TexCubeFloatFloat: |
| 220 | case NVPTXISD::TexCubeFloatFloatLevel: |
| 221 | case NVPTXISD::TexCubeS32Float: |
| 222 | case NVPTXISD::TexCubeS32FloatLevel: |
| 223 | case NVPTXISD::TexCubeU32Float: |
| 224 | case NVPTXISD::TexCubeU32FloatLevel: |
| 225 | case NVPTXISD::TexCubeArrayFloatFloat: |
| 226 | case NVPTXISD::TexCubeArrayFloatFloatLevel: |
| 227 | case NVPTXISD::TexCubeArrayS32Float: |
| 228 | case NVPTXISD::TexCubeArrayS32FloatLevel: |
| 229 | case NVPTXISD::TexCubeArrayU32Float: |
| 230 | case NVPTXISD::TexCubeArrayU32FloatLevel: |
| 231 | case NVPTXISD::Tld4R2DFloatFloat: |
| 232 | case NVPTXISD::Tld4G2DFloatFloat: |
| 233 | case NVPTXISD::Tld4B2DFloatFloat: |
| 234 | case NVPTXISD::Tld4A2DFloatFloat: |
| 235 | case NVPTXISD::Tld4R2DS64Float: |
| 236 | case NVPTXISD::Tld4G2DS64Float: |
| 237 | case NVPTXISD::Tld4B2DS64Float: |
| 238 | case NVPTXISD::Tld4A2DS64Float: |
| 239 | case NVPTXISD::Tld4R2DU64Float: |
| 240 | case NVPTXISD::Tld4G2DU64Float: |
| 241 | case NVPTXISD::Tld4B2DU64Float: |
| 242 | case NVPTXISD::Tld4A2DU64Float: |
| 243 | case NVPTXISD::TexUnified1DFloatS32: |
| 244 | case NVPTXISD::TexUnified1DFloatFloat: |
| 245 | case NVPTXISD::TexUnified1DFloatFloatLevel: |
| 246 | case NVPTXISD::TexUnified1DFloatFloatGrad: |
| 247 | case NVPTXISD::TexUnified1DS32S32: |
| 248 | case NVPTXISD::TexUnified1DS32Float: |
| 249 | case NVPTXISD::TexUnified1DS32FloatLevel: |
| 250 | case NVPTXISD::TexUnified1DS32FloatGrad: |
| 251 | case NVPTXISD::TexUnified1DU32S32: |
| 252 | case NVPTXISD::TexUnified1DU32Float: |
| 253 | case NVPTXISD::TexUnified1DU32FloatLevel: |
| 254 | case NVPTXISD::TexUnified1DU32FloatGrad: |
| 255 | case NVPTXISD::TexUnified1DArrayFloatS32: |
| 256 | case NVPTXISD::TexUnified1DArrayFloatFloat: |
| 257 | case NVPTXISD::TexUnified1DArrayFloatFloatLevel: |
| 258 | case NVPTXISD::TexUnified1DArrayFloatFloatGrad: |
| 259 | case NVPTXISD::TexUnified1DArrayS32S32: |
| 260 | case NVPTXISD::TexUnified1DArrayS32Float: |
| 261 | case NVPTXISD::TexUnified1DArrayS32FloatLevel: |
| 262 | case NVPTXISD::TexUnified1DArrayS32FloatGrad: |
| 263 | case NVPTXISD::TexUnified1DArrayU32S32: |
| 264 | case NVPTXISD::TexUnified1DArrayU32Float: |
| 265 | case NVPTXISD::TexUnified1DArrayU32FloatLevel: |
| 266 | case NVPTXISD::TexUnified1DArrayU32FloatGrad: |
| 267 | case NVPTXISD::TexUnified2DFloatS32: |
| 268 | case NVPTXISD::TexUnified2DFloatFloat: |
| 269 | case NVPTXISD::TexUnified2DFloatFloatLevel: |
| 270 | case NVPTXISD::TexUnified2DFloatFloatGrad: |
| 271 | case NVPTXISD::TexUnified2DS32S32: |
| 272 | case NVPTXISD::TexUnified2DS32Float: |
| 273 | case NVPTXISD::TexUnified2DS32FloatLevel: |
| 274 | case NVPTXISD::TexUnified2DS32FloatGrad: |
| 275 | case NVPTXISD::TexUnified2DU32S32: |
| 276 | case NVPTXISD::TexUnified2DU32Float: |
| 277 | case NVPTXISD::TexUnified2DU32FloatLevel: |
| 278 | case NVPTXISD::TexUnified2DU32FloatGrad: |
| 279 | case NVPTXISD::TexUnified2DArrayFloatS32: |
| 280 | case NVPTXISD::TexUnified2DArrayFloatFloat: |
| 281 | case NVPTXISD::TexUnified2DArrayFloatFloatLevel: |
| 282 | case NVPTXISD::TexUnified2DArrayFloatFloatGrad: |
| 283 | case NVPTXISD::TexUnified2DArrayS32S32: |
| 284 | case NVPTXISD::TexUnified2DArrayS32Float: |
| 285 | case NVPTXISD::TexUnified2DArrayS32FloatLevel: |
| 286 | case NVPTXISD::TexUnified2DArrayS32FloatGrad: |
| 287 | case NVPTXISD::TexUnified2DArrayU32S32: |
| 288 | case NVPTXISD::TexUnified2DArrayU32Float: |
| 289 | case NVPTXISD::TexUnified2DArrayU32FloatLevel: |
| 290 | case NVPTXISD::TexUnified2DArrayU32FloatGrad: |
| 291 | case NVPTXISD::TexUnified3DFloatS32: |
| 292 | case NVPTXISD::TexUnified3DFloatFloat: |
| 293 | case NVPTXISD::TexUnified3DFloatFloatLevel: |
| 294 | case NVPTXISD::TexUnified3DFloatFloatGrad: |
| 295 | case NVPTXISD::TexUnified3DS32S32: |
| 296 | case NVPTXISD::TexUnified3DS32Float: |
| 297 | case NVPTXISD::TexUnified3DS32FloatLevel: |
| 298 | case NVPTXISD::TexUnified3DS32FloatGrad: |
| 299 | case NVPTXISD::TexUnified3DU32S32: |
| 300 | case NVPTXISD::TexUnified3DU32Float: |
| 301 | case NVPTXISD::TexUnified3DU32FloatLevel: |
| 302 | case NVPTXISD::TexUnified3DU32FloatGrad: |
| 303 | case NVPTXISD::TexUnifiedCubeFloatFloat: |
| 304 | case NVPTXISD::TexUnifiedCubeFloatFloatLevel: |
| 305 | case NVPTXISD::TexUnifiedCubeS32Float: |
| 306 | case NVPTXISD::TexUnifiedCubeS32FloatLevel: |
| 307 | case NVPTXISD::TexUnifiedCubeU32Float: |
| 308 | case NVPTXISD::TexUnifiedCubeU32FloatLevel: |
| 309 | case NVPTXISD::TexUnifiedCubeArrayFloatFloat: |
| 310 | case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel: |
| 311 | case NVPTXISD::TexUnifiedCubeArrayS32Float: |
| 312 | case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel: |
| 313 | case NVPTXISD::TexUnifiedCubeArrayU32Float: |
| 314 | case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel: |
| 315 | case NVPTXISD::Tld4UnifiedR2DFloatFloat: |
| 316 | case NVPTXISD::Tld4UnifiedG2DFloatFloat: |
| 317 | case NVPTXISD::Tld4UnifiedB2DFloatFloat: |
| 318 | case NVPTXISD::Tld4UnifiedA2DFloatFloat: |
| 319 | case NVPTXISD::Tld4UnifiedR2DS64Float: |
| 320 | case NVPTXISD::Tld4UnifiedG2DS64Float: |
| 321 | case NVPTXISD::Tld4UnifiedB2DS64Float: |
| 322 | case NVPTXISD::Tld4UnifiedA2DS64Float: |
| 323 | case NVPTXISD::Tld4UnifiedR2DU64Float: |
| 324 | case NVPTXISD::Tld4UnifiedG2DU64Float: |
| 325 | case NVPTXISD::Tld4UnifiedB2DU64Float: |
| 326 | case NVPTXISD::Tld4UnifiedA2DU64Float: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 327 | ResNode = SelectTextureIntrinsic(N); |
| 328 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 329 | case NVPTXISD::Suld1DI8Clamp: |
| 330 | case NVPTXISD::Suld1DI16Clamp: |
| 331 | case NVPTXISD::Suld1DI32Clamp: |
| 332 | case NVPTXISD::Suld1DI64Clamp: |
| 333 | case NVPTXISD::Suld1DV2I8Clamp: |
| 334 | case NVPTXISD::Suld1DV2I16Clamp: |
| 335 | case NVPTXISD::Suld1DV2I32Clamp: |
| 336 | case NVPTXISD::Suld1DV2I64Clamp: |
| 337 | case NVPTXISD::Suld1DV4I8Clamp: |
| 338 | case NVPTXISD::Suld1DV4I16Clamp: |
| 339 | case NVPTXISD::Suld1DV4I32Clamp: |
| 340 | case NVPTXISD::Suld1DArrayI8Clamp: |
| 341 | case NVPTXISD::Suld1DArrayI16Clamp: |
| 342 | case NVPTXISD::Suld1DArrayI32Clamp: |
| 343 | case NVPTXISD::Suld1DArrayI64Clamp: |
| 344 | case NVPTXISD::Suld1DArrayV2I8Clamp: |
| 345 | case NVPTXISD::Suld1DArrayV2I16Clamp: |
| 346 | case NVPTXISD::Suld1DArrayV2I32Clamp: |
| 347 | case NVPTXISD::Suld1DArrayV2I64Clamp: |
| 348 | case NVPTXISD::Suld1DArrayV4I8Clamp: |
| 349 | case NVPTXISD::Suld1DArrayV4I16Clamp: |
| 350 | case NVPTXISD::Suld1DArrayV4I32Clamp: |
| 351 | case NVPTXISD::Suld2DI8Clamp: |
| 352 | case NVPTXISD::Suld2DI16Clamp: |
| 353 | case NVPTXISD::Suld2DI32Clamp: |
| 354 | case NVPTXISD::Suld2DI64Clamp: |
| 355 | case NVPTXISD::Suld2DV2I8Clamp: |
| 356 | case NVPTXISD::Suld2DV2I16Clamp: |
| 357 | case NVPTXISD::Suld2DV2I32Clamp: |
| 358 | case NVPTXISD::Suld2DV2I64Clamp: |
| 359 | case NVPTXISD::Suld2DV4I8Clamp: |
| 360 | case NVPTXISD::Suld2DV4I16Clamp: |
| 361 | case NVPTXISD::Suld2DV4I32Clamp: |
| 362 | case NVPTXISD::Suld2DArrayI8Clamp: |
| 363 | case NVPTXISD::Suld2DArrayI16Clamp: |
| 364 | case NVPTXISD::Suld2DArrayI32Clamp: |
| 365 | case NVPTXISD::Suld2DArrayI64Clamp: |
| 366 | case NVPTXISD::Suld2DArrayV2I8Clamp: |
| 367 | case NVPTXISD::Suld2DArrayV2I16Clamp: |
| 368 | case NVPTXISD::Suld2DArrayV2I32Clamp: |
| 369 | case NVPTXISD::Suld2DArrayV2I64Clamp: |
| 370 | case NVPTXISD::Suld2DArrayV4I8Clamp: |
| 371 | case NVPTXISD::Suld2DArrayV4I16Clamp: |
| 372 | case NVPTXISD::Suld2DArrayV4I32Clamp: |
| 373 | case NVPTXISD::Suld3DI8Clamp: |
| 374 | case NVPTXISD::Suld3DI16Clamp: |
| 375 | case NVPTXISD::Suld3DI32Clamp: |
| 376 | case NVPTXISD::Suld3DI64Clamp: |
| 377 | case NVPTXISD::Suld3DV2I8Clamp: |
| 378 | case NVPTXISD::Suld3DV2I16Clamp: |
| 379 | case NVPTXISD::Suld3DV2I32Clamp: |
| 380 | case NVPTXISD::Suld3DV2I64Clamp: |
| 381 | case NVPTXISD::Suld3DV4I8Clamp: |
| 382 | case NVPTXISD::Suld3DV4I16Clamp: |
| 383 | case NVPTXISD::Suld3DV4I32Clamp: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 384 | case NVPTXISD::Suld1DI8Trap: |
| 385 | case NVPTXISD::Suld1DI16Trap: |
| 386 | case NVPTXISD::Suld1DI32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 387 | case NVPTXISD::Suld1DI64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 388 | case NVPTXISD::Suld1DV2I8Trap: |
| 389 | case NVPTXISD::Suld1DV2I16Trap: |
| 390 | case NVPTXISD::Suld1DV2I32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 391 | case NVPTXISD::Suld1DV2I64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 392 | case NVPTXISD::Suld1DV4I8Trap: |
| 393 | case NVPTXISD::Suld1DV4I16Trap: |
| 394 | case NVPTXISD::Suld1DV4I32Trap: |
| 395 | case NVPTXISD::Suld1DArrayI8Trap: |
| 396 | case NVPTXISD::Suld1DArrayI16Trap: |
| 397 | case NVPTXISD::Suld1DArrayI32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 398 | case NVPTXISD::Suld1DArrayI64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 399 | case NVPTXISD::Suld1DArrayV2I8Trap: |
| 400 | case NVPTXISD::Suld1DArrayV2I16Trap: |
| 401 | case NVPTXISD::Suld1DArrayV2I32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 402 | case NVPTXISD::Suld1DArrayV2I64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 403 | case NVPTXISD::Suld1DArrayV4I8Trap: |
| 404 | case NVPTXISD::Suld1DArrayV4I16Trap: |
| 405 | case NVPTXISD::Suld1DArrayV4I32Trap: |
| 406 | case NVPTXISD::Suld2DI8Trap: |
| 407 | case NVPTXISD::Suld2DI16Trap: |
| 408 | case NVPTXISD::Suld2DI32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 409 | case NVPTXISD::Suld2DI64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 410 | case NVPTXISD::Suld2DV2I8Trap: |
| 411 | case NVPTXISD::Suld2DV2I16Trap: |
| 412 | case NVPTXISD::Suld2DV2I32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 413 | case NVPTXISD::Suld2DV2I64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 414 | case NVPTXISD::Suld2DV4I8Trap: |
| 415 | case NVPTXISD::Suld2DV4I16Trap: |
| 416 | case NVPTXISD::Suld2DV4I32Trap: |
| 417 | case NVPTXISD::Suld2DArrayI8Trap: |
| 418 | case NVPTXISD::Suld2DArrayI16Trap: |
| 419 | case NVPTXISD::Suld2DArrayI32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 420 | case NVPTXISD::Suld2DArrayI64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 421 | case NVPTXISD::Suld2DArrayV2I8Trap: |
| 422 | case NVPTXISD::Suld2DArrayV2I16Trap: |
| 423 | case NVPTXISD::Suld2DArrayV2I32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 424 | case NVPTXISD::Suld2DArrayV2I64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 425 | case NVPTXISD::Suld2DArrayV4I8Trap: |
| 426 | case NVPTXISD::Suld2DArrayV4I16Trap: |
| 427 | case NVPTXISD::Suld2DArrayV4I32Trap: |
| 428 | case NVPTXISD::Suld3DI8Trap: |
| 429 | case NVPTXISD::Suld3DI16Trap: |
| 430 | case NVPTXISD::Suld3DI32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 431 | case NVPTXISD::Suld3DI64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 432 | case NVPTXISD::Suld3DV2I8Trap: |
| 433 | case NVPTXISD::Suld3DV2I16Trap: |
| 434 | case NVPTXISD::Suld3DV2I32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 435 | case NVPTXISD::Suld3DV2I64Trap: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 436 | case NVPTXISD::Suld3DV4I8Trap: |
| 437 | case NVPTXISD::Suld3DV4I16Trap: |
| 438 | case NVPTXISD::Suld3DV4I32Trap: |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 439 | case NVPTXISD::Suld1DI8Zero: |
| 440 | case NVPTXISD::Suld1DI16Zero: |
| 441 | case NVPTXISD::Suld1DI32Zero: |
| 442 | case NVPTXISD::Suld1DI64Zero: |
| 443 | case NVPTXISD::Suld1DV2I8Zero: |
| 444 | case NVPTXISD::Suld1DV2I16Zero: |
| 445 | case NVPTXISD::Suld1DV2I32Zero: |
| 446 | case NVPTXISD::Suld1DV2I64Zero: |
| 447 | case NVPTXISD::Suld1DV4I8Zero: |
| 448 | case NVPTXISD::Suld1DV4I16Zero: |
| 449 | case NVPTXISD::Suld1DV4I32Zero: |
| 450 | case NVPTXISD::Suld1DArrayI8Zero: |
| 451 | case NVPTXISD::Suld1DArrayI16Zero: |
| 452 | case NVPTXISD::Suld1DArrayI32Zero: |
| 453 | case NVPTXISD::Suld1DArrayI64Zero: |
| 454 | case NVPTXISD::Suld1DArrayV2I8Zero: |
| 455 | case NVPTXISD::Suld1DArrayV2I16Zero: |
| 456 | case NVPTXISD::Suld1DArrayV2I32Zero: |
| 457 | case NVPTXISD::Suld1DArrayV2I64Zero: |
| 458 | case NVPTXISD::Suld1DArrayV4I8Zero: |
| 459 | case NVPTXISD::Suld1DArrayV4I16Zero: |
| 460 | case NVPTXISD::Suld1DArrayV4I32Zero: |
| 461 | case NVPTXISD::Suld2DI8Zero: |
| 462 | case NVPTXISD::Suld2DI16Zero: |
| 463 | case NVPTXISD::Suld2DI32Zero: |
| 464 | case NVPTXISD::Suld2DI64Zero: |
| 465 | case NVPTXISD::Suld2DV2I8Zero: |
| 466 | case NVPTXISD::Suld2DV2I16Zero: |
| 467 | case NVPTXISD::Suld2DV2I32Zero: |
| 468 | case NVPTXISD::Suld2DV2I64Zero: |
| 469 | case NVPTXISD::Suld2DV4I8Zero: |
| 470 | case NVPTXISD::Suld2DV4I16Zero: |
| 471 | case NVPTXISD::Suld2DV4I32Zero: |
| 472 | case NVPTXISD::Suld2DArrayI8Zero: |
| 473 | case NVPTXISD::Suld2DArrayI16Zero: |
| 474 | case NVPTXISD::Suld2DArrayI32Zero: |
| 475 | case NVPTXISD::Suld2DArrayI64Zero: |
| 476 | case NVPTXISD::Suld2DArrayV2I8Zero: |
| 477 | case NVPTXISD::Suld2DArrayV2I16Zero: |
| 478 | case NVPTXISD::Suld2DArrayV2I32Zero: |
| 479 | case NVPTXISD::Suld2DArrayV2I64Zero: |
| 480 | case NVPTXISD::Suld2DArrayV4I8Zero: |
| 481 | case NVPTXISD::Suld2DArrayV4I16Zero: |
| 482 | case NVPTXISD::Suld2DArrayV4I32Zero: |
| 483 | case NVPTXISD::Suld3DI8Zero: |
| 484 | case NVPTXISD::Suld3DI16Zero: |
| 485 | case NVPTXISD::Suld3DI32Zero: |
| 486 | case NVPTXISD::Suld3DI64Zero: |
| 487 | case NVPTXISD::Suld3DV2I8Zero: |
| 488 | case NVPTXISD::Suld3DV2I16Zero: |
| 489 | case NVPTXISD::Suld3DV2I32Zero: |
| 490 | case NVPTXISD::Suld3DV2I64Zero: |
| 491 | case NVPTXISD::Suld3DV4I8Zero: |
| 492 | case NVPTXISD::Suld3DV4I16Zero: |
| 493 | case NVPTXISD::Suld3DV4I32Zero: |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 494 | ResNode = SelectSurfaceIntrinsic(N); |
| 495 | break; |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 496 | case ISD::AND: |
| 497 | case ISD::SRA: |
| 498 | case ISD::SRL: |
| 499 | // Try to select BFE |
| 500 | ResNode = SelectBFE(N); |
| 501 | break; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 502 | case ISD::ADDRSPACECAST: |
| 503 | ResNode = SelectAddrSpaceCast(N); |
| 504 | break; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 505 | default: |
| 506 | break; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 507 | } |
| 508 | if (ResNode) |
| 509 | return ResNode; |
| 510 | return SelectCode(N); |
| 511 | } |
| 512 | |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 513 | SDNode *NVPTXDAGToDAGISel::SelectIntrinsicChain(SDNode *N) { |
| 514 | unsigned IID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); |
| 515 | switch (IID) { |
| 516 | default: |
| 517 | return NULL; |
| 518 | case Intrinsic::nvvm_ldg_global_f: |
| 519 | case Intrinsic::nvvm_ldg_global_i: |
| 520 | case Intrinsic::nvvm_ldg_global_p: |
| 521 | case Intrinsic::nvvm_ldu_global_f: |
| 522 | case Intrinsic::nvvm_ldu_global_i: |
| 523 | case Intrinsic::nvvm_ldu_global_p: |
| 524 | return SelectLDGLDU(N); |
| 525 | } |
| 526 | } |
| 527 | |
Eric Christopher | 9745b3a | 2015-01-30 01:41:01 +0000 | [diff] [blame] | 528 | static unsigned int getCodeAddrSpace(MemSDNode *N) { |
Nick Lewycky | aad475b | 2014-04-15 07:22:52 +0000 | [diff] [blame] | 529 | const Value *Src = N->getMemOperand()->getValue(); |
Justin Holewinski | b96d139 | 2013-06-10 13:29:47 +0000 | [diff] [blame] | 530 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 531 | if (!Src) |
Justin Holewinski | b96d139 | 2013-06-10 13:29:47 +0000 | [diff] [blame] | 532 | return NVPTX::PTXLdStInstCode::GENERIC; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 533 | |
Craig Topper | e3dcce9 | 2015-08-01 22:20:21 +0000 | [diff] [blame^] | 534 | if (auto *PT = dyn_cast<PointerType>(Src->getType())) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 535 | switch (PT->getAddressSpace()) { |
Justin Holewinski | b96d139 | 2013-06-10 13:29:47 +0000 | [diff] [blame] | 536 | case llvm::ADDRESS_SPACE_LOCAL: return NVPTX::PTXLdStInstCode::LOCAL; |
| 537 | case llvm::ADDRESS_SPACE_GLOBAL: return NVPTX::PTXLdStInstCode::GLOBAL; |
| 538 | case llvm::ADDRESS_SPACE_SHARED: return NVPTX::PTXLdStInstCode::SHARED; |
| 539 | case llvm::ADDRESS_SPACE_GENERIC: return NVPTX::PTXLdStInstCode::GENERIC; |
| 540 | case llvm::ADDRESS_SPACE_PARAM: return NVPTX::PTXLdStInstCode::PARAM; |
| 541 | case llvm::ADDRESS_SPACE_CONST: return NVPTX::PTXLdStInstCode::CONSTANT; |
| 542 | default: break; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 543 | } |
| 544 | } |
Justin Holewinski | b96d139 | 2013-06-10 13:29:47 +0000 | [diff] [blame] | 545 | return NVPTX::PTXLdStInstCode::GENERIC; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 546 | } |
| 547 | |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 548 | static bool canLowerToLDG(MemSDNode *N, const NVPTXSubtarget &Subtarget, |
| 549 | unsigned codeAddrSpace, const DataLayout &DL) { |
| 550 | if (!Subtarget.hasLDG() || codeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL) { |
| 551 | return false; |
| 552 | } |
| 553 | |
| 554 | // Check whether load operates on a readonly argument. |
| 555 | bool canUseLDG = false; |
| 556 | if (const Argument *A = dyn_cast<const Argument>( |
| 557 | GetUnderlyingObject(N->getMemOperand()->getValue(), DL))) |
| 558 | canUseLDG = A->onlyReadsMemory() && A->hasNoAliasAttr(); |
| 559 | |
| 560 | return canUseLDG; |
| 561 | } |
| 562 | |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 563 | SDNode *NVPTXDAGToDAGISel::SelectIntrinsicNoChain(SDNode *N) { |
| 564 | unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); |
| 565 | switch (IID) { |
| 566 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 567 | return nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 568 | case Intrinsic::nvvm_texsurf_handle_internal: |
| 569 | return SelectTexSurfHandle(N); |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | SDNode *NVPTXDAGToDAGISel::SelectTexSurfHandle(SDNode *N) { |
| 574 | // Op 0 is the intrinsic ID |
| 575 | SDValue Wrapper = N->getOperand(1); |
| 576 | SDValue GlobalVal = Wrapper.getOperand(0); |
| 577 | return CurDAG->getMachineNode(NVPTX::texsurf_handles, SDLoc(N), MVT::i64, |
| 578 | GlobalVal); |
| 579 | } |
| 580 | |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 581 | SDNode *NVPTXDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) { |
| 582 | SDValue Src = N->getOperand(0); |
| 583 | AddrSpaceCastSDNode *CastN = cast<AddrSpaceCastSDNode>(N); |
| 584 | unsigned SrcAddrSpace = CastN->getSrcAddressSpace(); |
| 585 | unsigned DstAddrSpace = CastN->getDestAddressSpace(); |
| 586 | |
| 587 | assert(SrcAddrSpace != DstAddrSpace && |
| 588 | "addrspacecast must be between different address spaces"); |
| 589 | |
| 590 | if (DstAddrSpace == ADDRESS_SPACE_GENERIC) { |
| 591 | // Specific to generic |
| 592 | unsigned Opc; |
| 593 | switch (SrcAddrSpace) { |
| 594 | default: report_fatal_error("Bad address space in addrspacecast"); |
| 595 | case ADDRESS_SPACE_GLOBAL: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 596 | Opc = TM.is64Bit() ? NVPTX::cvta_global_yes_64 : NVPTX::cvta_global_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 597 | break; |
| 598 | case ADDRESS_SPACE_SHARED: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 599 | Opc = TM.is64Bit() ? NVPTX::cvta_shared_yes_64 : NVPTX::cvta_shared_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 600 | break; |
| 601 | case ADDRESS_SPACE_CONST: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 602 | Opc = TM.is64Bit() ? NVPTX::cvta_const_yes_64 : NVPTX::cvta_const_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 603 | break; |
| 604 | case ADDRESS_SPACE_LOCAL: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 605 | Opc = TM.is64Bit() ? NVPTX::cvta_local_yes_64 : NVPTX::cvta_local_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 606 | break; |
| 607 | } |
| 608 | return CurDAG->getMachineNode(Opc, SDLoc(N), N->getValueType(0), Src); |
| 609 | } else { |
| 610 | // Generic to specific |
| 611 | if (SrcAddrSpace != 0) |
| 612 | report_fatal_error("Cannot cast between two non-generic address spaces"); |
| 613 | unsigned Opc; |
| 614 | switch (DstAddrSpace) { |
| 615 | default: report_fatal_error("Bad address space in addrspacecast"); |
| 616 | case ADDRESS_SPACE_GLOBAL: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 617 | Opc = TM.is64Bit() ? NVPTX::cvta_to_global_yes_64 |
| 618 | : NVPTX::cvta_to_global_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 619 | break; |
| 620 | case ADDRESS_SPACE_SHARED: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 621 | Opc = TM.is64Bit() ? NVPTX::cvta_to_shared_yes_64 |
| 622 | : NVPTX::cvta_to_shared_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 623 | break; |
| 624 | case ADDRESS_SPACE_CONST: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 625 | Opc = |
| 626 | TM.is64Bit() ? NVPTX::cvta_to_const_yes_64 : NVPTX::cvta_to_const_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 627 | break; |
| 628 | case ADDRESS_SPACE_LOCAL: |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 629 | Opc = |
| 630 | TM.is64Bit() ? NVPTX::cvta_to_local_yes_64 : NVPTX::cvta_to_local_yes; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 631 | break; |
Jingyue Wu | a2f6027 | 2015-06-04 21:28:26 +0000 | [diff] [blame] | 632 | case ADDRESS_SPACE_PARAM: |
| 633 | Opc = TM.is64Bit() ? NVPTX::nvvm_ptr_gen_to_param_64 |
| 634 | : NVPTX::nvvm_ptr_gen_to_param; |
| 635 | break; |
Justin Holewinski | ba2fa6d | 2014-03-24 11:17:53 +0000 | [diff] [blame] | 636 | } |
| 637 | return CurDAG->getMachineNode(Opc, SDLoc(N), N->getValueType(0), Src); |
| 638 | } |
| 639 | } |
| 640 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 641 | SDNode *NVPTXDAGToDAGISel::SelectLoad(SDNode *N) { |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 642 | SDLoc dl(N); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 643 | LoadSDNode *LD = cast<LoadSDNode>(N); |
| 644 | EVT LoadedVT = LD->getMemoryVT(); |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 645 | SDNode *NVPTXLD = nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 646 | |
| 647 | // do not support pre/post inc/dec |
| 648 | if (LD->isIndexed()) |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 649 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 650 | |
| 651 | if (!LoadedVT.isSimple()) |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 652 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 653 | |
| 654 | // Address Space Setting |
Eric Christopher | 9745b3a | 2015-01-30 01:41:01 +0000 | [diff] [blame] | 655 | unsigned int codeAddrSpace = getCodeAddrSpace(LD); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 656 | |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 657 | if (canLowerToLDG(LD, *Subtarget, codeAddrSpace, CurDAG->getDataLayout())) { |
| 658 | return SelectLDGLDU(N); |
| 659 | } |
| 660 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 661 | // Volatile Setting |
| 662 | // - .volatile is only availalble for .global and .shared |
| 663 | bool isVolatile = LD->isVolatile(); |
| 664 | if (codeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL && |
| 665 | codeAddrSpace != NVPTX::PTXLdStInstCode::SHARED && |
| 666 | codeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC) |
| 667 | isVolatile = false; |
| 668 | |
| 669 | // Vector Setting |
| 670 | MVT SimpleVT = LoadedVT.getSimpleVT(); |
| 671 | unsigned vecType = NVPTX::PTXLdStInstCode::Scalar; |
| 672 | if (SimpleVT.isVector()) { |
| 673 | unsigned num = SimpleVT.getVectorNumElements(); |
| 674 | if (num == 2) |
| 675 | vecType = NVPTX::PTXLdStInstCode::V2; |
| 676 | else if (num == 4) |
| 677 | vecType = NVPTX::PTXLdStInstCode::V4; |
| 678 | else |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 679 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 680 | } |
| 681 | |
| 682 | // Type Setting: fromType + fromTypeWidth |
| 683 | // |
| 684 | // Sign : ISD::SEXTLOAD |
| 685 | // Unsign : ISD::ZEXTLOAD, ISD::NON_EXTLOAD or ISD::EXTLOAD and the |
| 686 | // type is integer |
| 687 | // Float : ISD::NON_EXTLOAD or ISD::EXTLOAD and the type is float |
| 688 | MVT ScalarVT = SimpleVT.getScalarType(); |
Justin Holewinski | 994d66a | 2013-05-30 12:22:39 +0000 | [diff] [blame] | 689 | // Read at least 8 bits (predicates are stored as 8-bit values) |
| 690 | unsigned fromTypeWidth = std::max(8U, ScalarVT.getSizeInBits()); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 691 | unsigned int fromType; |
| 692 | if ((LD->getExtensionType() == ISD::SEXTLOAD)) |
| 693 | fromType = NVPTX::PTXLdStInstCode::Signed; |
| 694 | else if (ScalarVT.isFloatingPoint()) |
| 695 | fromType = NVPTX::PTXLdStInstCode::Float; |
| 696 | else |
| 697 | fromType = NVPTX::PTXLdStInstCode::Unsigned; |
| 698 | |
| 699 | // Create the machine instruction DAG |
| 700 | SDValue Chain = N->getOperand(0); |
| 701 | SDValue N1 = N->getOperand(1); |
| 702 | SDValue Addr; |
| 703 | SDValue Offset, Base; |
| 704 | unsigned Opcode; |
Craig Topper | d9c2783 | 2013-08-15 02:44:19 +0000 | [diff] [blame] | 705 | MVT::SimpleValueType TargetVT = LD->getSimpleValueType(0).SimpleTy; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 706 | |
| 707 | if (SelectDirectAddr(N1, Addr)) { |
| 708 | switch (TargetVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 709 | case MVT::i8: |
| 710 | Opcode = NVPTX::LD_i8_avar; |
| 711 | break; |
| 712 | case MVT::i16: |
| 713 | Opcode = NVPTX::LD_i16_avar; |
| 714 | break; |
| 715 | case MVT::i32: |
| 716 | Opcode = NVPTX::LD_i32_avar; |
| 717 | break; |
| 718 | case MVT::i64: |
| 719 | Opcode = NVPTX::LD_i64_avar; |
| 720 | break; |
| 721 | case MVT::f32: |
| 722 | Opcode = NVPTX::LD_f32_avar; |
| 723 | break; |
| 724 | case MVT::f64: |
| 725 | Opcode = NVPTX::LD_f64_avar; |
| 726 | break; |
| 727 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 728 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 729 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 730 | SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(codeAddrSpace, dl), |
| 731 | getI32Imm(vecType, dl), getI32Imm(fromType, dl), |
| 732 | getI32Imm(fromTypeWidth, dl), Addr, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 733 | NVPTXLD = CurDAG->getMachineNode(Opcode, dl, TargetVT, MVT::Other, Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 734 | } else if (TM.is64Bit() ? SelectADDRsi64(N1.getNode(), N1, Base, Offset) |
| 735 | : SelectADDRsi(N1.getNode(), N1, Base, Offset)) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 736 | switch (TargetVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 737 | case MVT::i8: |
| 738 | Opcode = NVPTX::LD_i8_asi; |
| 739 | break; |
| 740 | case MVT::i16: |
| 741 | Opcode = NVPTX::LD_i16_asi; |
| 742 | break; |
| 743 | case MVT::i32: |
| 744 | Opcode = NVPTX::LD_i32_asi; |
| 745 | break; |
| 746 | case MVT::i64: |
| 747 | Opcode = NVPTX::LD_i64_asi; |
| 748 | break; |
| 749 | case MVT::f32: |
| 750 | Opcode = NVPTX::LD_f32_asi; |
| 751 | break; |
| 752 | case MVT::f64: |
| 753 | Opcode = NVPTX::LD_f64_asi; |
| 754 | break; |
| 755 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 756 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 757 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 758 | SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(codeAddrSpace, dl), |
| 759 | getI32Imm(vecType, dl), getI32Imm(fromType, dl), |
| 760 | getI32Imm(fromTypeWidth, dl), Base, Offset, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 761 | NVPTXLD = CurDAG->getMachineNode(Opcode, dl, TargetVT, MVT::Other, Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 762 | } else if (TM.is64Bit() ? SelectADDRri64(N1.getNode(), N1, Base, Offset) |
| 763 | : SelectADDRri(N1.getNode(), N1, Base, Offset)) { |
| 764 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 765 | switch (TargetVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 766 | case MVT::i8: |
| 767 | Opcode = NVPTX::LD_i8_ari_64; |
| 768 | break; |
| 769 | case MVT::i16: |
| 770 | Opcode = NVPTX::LD_i16_ari_64; |
| 771 | break; |
| 772 | case MVT::i32: |
| 773 | Opcode = NVPTX::LD_i32_ari_64; |
| 774 | break; |
| 775 | case MVT::i64: |
| 776 | Opcode = NVPTX::LD_i64_ari_64; |
| 777 | break; |
| 778 | case MVT::f32: |
| 779 | Opcode = NVPTX::LD_f32_ari_64; |
| 780 | break; |
| 781 | case MVT::f64: |
| 782 | Opcode = NVPTX::LD_f64_ari_64; |
| 783 | break; |
| 784 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 785 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 786 | } |
| 787 | } else { |
| 788 | switch (TargetVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 789 | case MVT::i8: |
| 790 | Opcode = NVPTX::LD_i8_ari; |
| 791 | break; |
| 792 | case MVT::i16: |
| 793 | Opcode = NVPTX::LD_i16_ari; |
| 794 | break; |
| 795 | case MVT::i32: |
| 796 | Opcode = NVPTX::LD_i32_ari; |
| 797 | break; |
| 798 | case MVT::i64: |
| 799 | Opcode = NVPTX::LD_i64_ari; |
| 800 | break; |
| 801 | case MVT::f32: |
| 802 | Opcode = NVPTX::LD_f32_ari; |
| 803 | break; |
| 804 | case MVT::f64: |
| 805 | Opcode = NVPTX::LD_f64_ari; |
| 806 | break; |
| 807 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 808 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 809 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 810 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 811 | SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(codeAddrSpace, dl), |
| 812 | getI32Imm(vecType, dl), getI32Imm(fromType, dl), |
| 813 | getI32Imm(fromTypeWidth, dl), Base, Offset, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 814 | NVPTXLD = CurDAG->getMachineNode(Opcode, dl, TargetVT, MVT::Other, Ops); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 815 | } else { |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 816 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 817 | switch (TargetVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 818 | case MVT::i8: |
| 819 | Opcode = NVPTX::LD_i8_areg_64; |
| 820 | break; |
| 821 | case MVT::i16: |
| 822 | Opcode = NVPTX::LD_i16_areg_64; |
| 823 | break; |
| 824 | case MVT::i32: |
| 825 | Opcode = NVPTX::LD_i32_areg_64; |
| 826 | break; |
| 827 | case MVT::i64: |
| 828 | Opcode = NVPTX::LD_i64_areg_64; |
| 829 | break; |
| 830 | case MVT::f32: |
| 831 | Opcode = NVPTX::LD_f32_areg_64; |
| 832 | break; |
| 833 | case MVT::f64: |
| 834 | Opcode = NVPTX::LD_f64_areg_64; |
| 835 | break; |
| 836 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 837 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 838 | } |
| 839 | } else { |
| 840 | switch (TargetVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 841 | case MVT::i8: |
| 842 | Opcode = NVPTX::LD_i8_areg; |
| 843 | break; |
| 844 | case MVT::i16: |
| 845 | Opcode = NVPTX::LD_i16_areg; |
| 846 | break; |
| 847 | case MVT::i32: |
| 848 | Opcode = NVPTX::LD_i32_areg; |
| 849 | break; |
| 850 | case MVT::i64: |
| 851 | Opcode = NVPTX::LD_i64_areg; |
| 852 | break; |
| 853 | case MVT::f32: |
| 854 | Opcode = NVPTX::LD_f32_areg; |
| 855 | break; |
| 856 | case MVT::f64: |
| 857 | Opcode = NVPTX::LD_f64_areg; |
| 858 | break; |
| 859 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 860 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 861 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 862 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 863 | SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(codeAddrSpace, dl), |
| 864 | getI32Imm(vecType, dl), getI32Imm(fromType, dl), |
| 865 | getI32Imm(fromTypeWidth, dl), N1, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 866 | NVPTXLD = CurDAG->getMachineNode(Opcode, dl, TargetVT, MVT::Other, Ops); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 867 | } |
| 868 | |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 869 | if (NVPTXLD) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 870 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
| 871 | MemRefs0[0] = cast<MemSDNode>(N)->getMemOperand(); |
| 872 | cast<MachineSDNode>(NVPTXLD)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 873 | } |
| 874 | |
| 875 | return NVPTXLD; |
| 876 | } |
| 877 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 878 | SDNode *NVPTXDAGToDAGISel::SelectLoadVector(SDNode *N) { |
| 879 | |
| 880 | SDValue Chain = N->getOperand(0); |
| 881 | SDValue Op1 = N->getOperand(1); |
| 882 | SDValue Addr, Offset, Base; |
| 883 | unsigned Opcode; |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 884 | SDLoc DL(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 885 | SDNode *LD; |
| 886 | MemSDNode *MemSD = cast<MemSDNode>(N); |
| 887 | EVT LoadedVT = MemSD->getMemoryVT(); |
| 888 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 889 | if (!LoadedVT.isSimple()) |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 890 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 891 | |
| 892 | // Address Space Setting |
Eric Christopher | 9745b3a | 2015-01-30 01:41:01 +0000 | [diff] [blame] | 893 | unsigned int CodeAddrSpace = getCodeAddrSpace(MemSD); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 894 | |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 895 | if (canLowerToLDG(MemSD, *Subtarget, CodeAddrSpace, CurDAG->getDataLayout())) { |
| 896 | return SelectLDGLDU(N); |
| 897 | } |
| 898 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 899 | // Volatile Setting |
| 900 | // - .volatile is only availalble for .global and .shared |
| 901 | bool IsVolatile = MemSD->isVolatile(); |
| 902 | if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL && |
| 903 | CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED && |
| 904 | CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC) |
| 905 | IsVolatile = false; |
| 906 | |
| 907 | // Vector Setting |
| 908 | MVT SimpleVT = LoadedVT.getSimpleVT(); |
| 909 | |
| 910 | // Type Setting: fromType + fromTypeWidth |
| 911 | // |
| 912 | // Sign : ISD::SEXTLOAD |
| 913 | // Unsign : ISD::ZEXTLOAD, ISD::NON_EXTLOAD or ISD::EXTLOAD and the |
| 914 | // type is integer |
| 915 | // Float : ISD::NON_EXTLOAD or ISD::EXTLOAD and the type is float |
| 916 | MVT ScalarVT = SimpleVT.getScalarType(); |
Justin Holewinski | 994d66a | 2013-05-30 12:22:39 +0000 | [diff] [blame] | 917 | // Read at least 8 bits (predicates are stored as 8-bit values) |
| 918 | unsigned FromTypeWidth = std::max(8U, ScalarVT.getSizeInBits()); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 919 | unsigned int FromType; |
| 920 | // The last operand holds the original LoadSDNode::getExtensionType() value |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 921 | unsigned ExtensionType = cast<ConstantSDNode>( |
| 922 | N->getOperand(N->getNumOperands() - 1))->getZExtValue(); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 923 | if (ExtensionType == ISD::SEXTLOAD) |
| 924 | FromType = NVPTX::PTXLdStInstCode::Signed; |
| 925 | else if (ScalarVT.isFloatingPoint()) |
| 926 | FromType = NVPTX::PTXLdStInstCode::Float; |
| 927 | else |
| 928 | FromType = NVPTX::PTXLdStInstCode::Unsigned; |
| 929 | |
| 930 | unsigned VecType; |
| 931 | |
| 932 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 933 | case NVPTXISD::LoadV2: |
| 934 | VecType = NVPTX::PTXLdStInstCode::V2; |
| 935 | break; |
| 936 | case NVPTXISD::LoadV4: |
| 937 | VecType = NVPTX::PTXLdStInstCode::V4; |
| 938 | break; |
| 939 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 940 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 941 | } |
| 942 | |
| 943 | EVT EltVT = N->getValueType(0); |
| 944 | |
| 945 | if (SelectDirectAddr(Op1, Addr)) { |
| 946 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 947 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 948 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 949 | case NVPTXISD::LoadV2: |
| 950 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 951 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 952 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 953 | case MVT::i8: |
| 954 | Opcode = NVPTX::LDV_i8_v2_avar; |
| 955 | break; |
| 956 | case MVT::i16: |
| 957 | Opcode = NVPTX::LDV_i16_v2_avar; |
| 958 | break; |
| 959 | case MVT::i32: |
| 960 | Opcode = NVPTX::LDV_i32_v2_avar; |
| 961 | break; |
| 962 | case MVT::i64: |
| 963 | Opcode = NVPTX::LDV_i64_v2_avar; |
| 964 | break; |
| 965 | case MVT::f32: |
| 966 | Opcode = NVPTX::LDV_f32_v2_avar; |
| 967 | break; |
| 968 | case MVT::f64: |
| 969 | Opcode = NVPTX::LDV_f64_v2_avar; |
| 970 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 971 | } |
| 972 | break; |
| 973 | case NVPTXISD::LoadV4: |
| 974 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 975 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 976 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 977 | case MVT::i8: |
| 978 | Opcode = NVPTX::LDV_i8_v4_avar; |
| 979 | break; |
| 980 | case MVT::i16: |
| 981 | Opcode = NVPTX::LDV_i16_v4_avar; |
| 982 | break; |
| 983 | case MVT::i32: |
| 984 | Opcode = NVPTX::LDV_i32_v4_avar; |
| 985 | break; |
| 986 | case MVT::f32: |
| 987 | Opcode = NVPTX::LDV_f32_v4_avar; |
| 988 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 989 | } |
| 990 | break; |
| 991 | } |
| 992 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 993 | SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL), |
| 994 | getI32Imm(VecType, DL), getI32Imm(FromType, DL), |
| 995 | getI32Imm(FromTypeWidth, DL), Addr, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 996 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 997 | } else if (TM.is64Bit() ? SelectADDRsi64(Op1.getNode(), Op1, Base, Offset) |
| 998 | : SelectADDRsi(Op1.getNode(), Op1, Base, Offset)) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 999 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1000 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1001 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1002 | case NVPTXISD::LoadV2: |
| 1003 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1004 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1005 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1006 | case MVT::i8: |
| 1007 | Opcode = NVPTX::LDV_i8_v2_asi; |
| 1008 | break; |
| 1009 | case MVT::i16: |
| 1010 | Opcode = NVPTX::LDV_i16_v2_asi; |
| 1011 | break; |
| 1012 | case MVT::i32: |
| 1013 | Opcode = NVPTX::LDV_i32_v2_asi; |
| 1014 | break; |
| 1015 | case MVT::i64: |
| 1016 | Opcode = NVPTX::LDV_i64_v2_asi; |
| 1017 | break; |
| 1018 | case MVT::f32: |
| 1019 | Opcode = NVPTX::LDV_f32_v2_asi; |
| 1020 | break; |
| 1021 | case MVT::f64: |
| 1022 | Opcode = NVPTX::LDV_f64_v2_asi; |
| 1023 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1024 | } |
| 1025 | break; |
| 1026 | case NVPTXISD::LoadV4: |
| 1027 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1028 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1029 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1030 | case MVT::i8: |
| 1031 | Opcode = NVPTX::LDV_i8_v4_asi; |
| 1032 | break; |
| 1033 | case MVT::i16: |
| 1034 | Opcode = NVPTX::LDV_i16_v4_asi; |
| 1035 | break; |
| 1036 | case MVT::i32: |
| 1037 | Opcode = NVPTX::LDV_i32_v4_asi; |
| 1038 | break; |
| 1039 | case MVT::f32: |
| 1040 | Opcode = NVPTX::LDV_f32_v4_asi; |
| 1041 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1042 | } |
| 1043 | break; |
| 1044 | } |
| 1045 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 1046 | SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL), |
| 1047 | getI32Imm(VecType, DL), getI32Imm(FromType, DL), |
| 1048 | getI32Imm(FromTypeWidth, DL), Base, Offset, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 1049 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 1050 | } else if (TM.is64Bit() ? SelectADDRri64(Op1.getNode(), Op1, Base, Offset) |
| 1051 | : SelectADDRri(Op1.getNode(), Op1, Base, Offset)) { |
| 1052 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1053 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1054 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1055 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1056 | case NVPTXISD::LoadV2: |
| 1057 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1058 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1059 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1060 | case MVT::i8: |
| 1061 | Opcode = NVPTX::LDV_i8_v2_ari_64; |
| 1062 | break; |
| 1063 | case MVT::i16: |
| 1064 | Opcode = NVPTX::LDV_i16_v2_ari_64; |
| 1065 | break; |
| 1066 | case MVT::i32: |
| 1067 | Opcode = NVPTX::LDV_i32_v2_ari_64; |
| 1068 | break; |
| 1069 | case MVT::i64: |
| 1070 | Opcode = NVPTX::LDV_i64_v2_ari_64; |
| 1071 | break; |
| 1072 | case MVT::f32: |
| 1073 | Opcode = NVPTX::LDV_f32_v2_ari_64; |
| 1074 | break; |
| 1075 | case MVT::f64: |
| 1076 | Opcode = NVPTX::LDV_f64_v2_ari_64; |
| 1077 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1078 | } |
| 1079 | break; |
| 1080 | case NVPTXISD::LoadV4: |
| 1081 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1082 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1083 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1084 | case MVT::i8: |
| 1085 | Opcode = NVPTX::LDV_i8_v4_ari_64; |
| 1086 | break; |
| 1087 | case MVT::i16: |
| 1088 | Opcode = NVPTX::LDV_i16_v4_ari_64; |
| 1089 | break; |
| 1090 | case MVT::i32: |
| 1091 | Opcode = NVPTX::LDV_i32_v4_ari_64; |
| 1092 | break; |
| 1093 | case MVT::f32: |
| 1094 | Opcode = NVPTX::LDV_f32_v4_ari_64; |
| 1095 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1096 | } |
| 1097 | break; |
| 1098 | } |
| 1099 | } else { |
| 1100 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1101 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1102 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1103 | case NVPTXISD::LoadV2: |
| 1104 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1105 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1106 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1107 | case MVT::i8: |
| 1108 | Opcode = NVPTX::LDV_i8_v2_ari; |
| 1109 | break; |
| 1110 | case MVT::i16: |
| 1111 | Opcode = NVPTX::LDV_i16_v2_ari; |
| 1112 | break; |
| 1113 | case MVT::i32: |
| 1114 | Opcode = NVPTX::LDV_i32_v2_ari; |
| 1115 | break; |
| 1116 | case MVT::i64: |
| 1117 | Opcode = NVPTX::LDV_i64_v2_ari; |
| 1118 | break; |
| 1119 | case MVT::f32: |
| 1120 | Opcode = NVPTX::LDV_f32_v2_ari; |
| 1121 | break; |
| 1122 | case MVT::f64: |
| 1123 | Opcode = NVPTX::LDV_f64_v2_ari; |
| 1124 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1125 | } |
| 1126 | break; |
| 1127 | case NVPTXISD::LoadV4: |
| 1128 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1129 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1130 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1131 | case MVT::i8: |
| 1132 | Opcode = NVPTX::LDV_i8_v4_ari; |
| 1133 | break; |
| 1134 | case MVT::i16: |
| 1135 | Opcode = NVPTX::LDV_i16_v4_ari; |
| 1136 | break; |
| 1137 | case MVT::i32: |
| 1138 | Opcode = NVPTX::LDV_i32_v4_ari; |
| 1139 | break; |
| 1140 | case MVT::f32: |
| 1141 | Opcode = NVPTX::LDV_f32_v4_ari; |
| 1142 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1143 | } |
| 1144 | break; |
| 1145 | } |
| 1146 | } |
| 1147 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 1148 | SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL), |
| 1149 | getI32Imm(VecType, DL), getI32Imm(FromType, DL), |
| 1150 | getI32Imm(FromTypeWidth, DL), Base, Offset, Chain }; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1151 | |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 1152 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1153 | } else { |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 1154 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1155 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1156 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1157 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1158 | case NVPTXISD::LoadV2: |
| 1159 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1160 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1161 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1162 | case MVT::i8: |
| 1163 | Opcode = NVPTX::LDV_i8_v2_areg_64; |
| 1164 | break; |
| 1165 | case MVT::i16: |
| 1166 | Opcode = NVPTX::LDV_i16_v2_areg_64; |
| 1167 | break; |
| 1168 | case MVT::i32: |
| 1169 | Opcode = NVPTX::LDV_i32_v2_areg_64; |
| 1170 | break; |
| 1171 | case MVT::i64: |
| 1172 | Opcode = NVPTX::LDV_i64_v2_areg_64; |
| 1173 | break; |
| 1174 | case MVT::f32: |
| 1175 | Opcode = NVPTX::LDV_f32_v2_areg_64; |
| 1176 | break; |
| 1177 | case MVT::f64: |
| 1178 | Opcode = NVPTX::LDV_f64_v2_areg_64; |
| 1179 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1180 | } |
| 1181 | break; |
| 1182 | case NVPTXISD::LoadV4: |
| 1183 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1184 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1185 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1186 | case MVT::i8: |
| 1187 | Opcode = NVPTX::LDV_i8_v4_areg_64; |
| 1188 | break; |
| 1189 | case MVT::i16: |
| 1190 | Opcode = NVPTX::LDV_i16_v4_areg_64; |
| 1191 | break; |
| 1192 | case MVT::i32: |
| 1193 | Opcode = NVPTX::LDV_i32_v4_areg_64; |
| 1194 | break; |
| 1195 | case MVT::f32: |
| 1196 | Opcode = NVPTX::LDV_f32_v4_areg_64; |
| 1197 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1198 | } |
| 1199 | break; |
| 1200 | } |
| 1201 | } else { |
| 1202 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1203 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1204 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1205 | case NVPTXISD::LoadV2: |
| 1206 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1207 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1208 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1209 | case MVT::i8: |
| 1210 | Opcode = NVPTX::LDV_i8_v2_areg; |
| 1211 | break; |
| 1212 | case MVT::i16: |
| 1213 | Opcode = NVPTX::LDV_i16_v2_areg; |
| 1214 | break; |
| 1215 | case MVT::i32: |
| 1216 | Opcode = NVPTX::LDV_i32_v2_areg; |
| 1217 | break; |
| 1218 | case MVT::i64: |
| 1219 | Opcode = NVPTX::LDV_i64_v2_areg; |
| 1220 | break; |
| 1221 | case MVT::f32: |
| 1222 | Opcode = NVPTX::LDV_f32_v2_areg; |
| 1223 | break; |
| 1224 | case MVT::f64: |
| 1225 | Opcode = NVPTX::LDV_f64_v2_areg; |
| 1226 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1227 | } |
| 1228 | break; |
| 1229 | case NVPTXISD::LoadV4: |
| 1230 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1231 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1232 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1233 | case MVT::i8: |
| 1234 | Opcode = NVPTX::LDV_i8_v4_areg; |
| 1235 | break; |
| 1236 | case MVT::i16: |
| 1237 | Opcode = NVPTX::LDV_i16_v4_areg; |
| 1238 | break; |
| 1239 | case MVT::i32: |
| 1240 | Opcode = NVPTX::LDV_i32_v4_areg; |
| 1241 | break; |
| 1242 | case MVT::f32: |
| 1243 | Opcode = NVPTX::LDV_f32_v4_areg; |
| 1244 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1245 | } |
| 1246 | break; |
| 1247 | } |
| 1248 | } |
| 1249 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 1250 | SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL), |
| 1251 | getI32Imm(VecType, DL), getI32Imm(FromType, DL), |
| 1252 | getI32Imm(FromTypeWidth, DL), Op1, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 1253 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1254 | } |
| 1255 | |
| 1256 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
| 1257 | MemRefs0[0] = cast<MemSDNode>(N)->getMemOperand(); |
| 1258 | cast<MachineSDNode>(LD)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 1259 | |
| 1260 | return LD; |
| 1261 | } |
| 1262 | |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1263 | SDNode *NVPTXDAGToDAGISel::SelectLDGLDU(SDNode *N) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1264 | |
| 1265 | SDValue Chain = N->getOperand(0); |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1266 | SDValue Op1; |
| 1267 | MemSDNode *Mem; |
| 1268 | bool IsLDG = true; |
| 1269 | |
| 1270 | // If this is an LDG intrinsic, the address is the third operand. Its its an |
| 1271 | // LDG/LDU SD node (from custom vector handling), then its the second operand |
| 1272 | if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) { |
| 1273 | Op1 = N->getOperand(2); |
| 1274 | Mem = cast<MemIntrinsicSDNode>(N); |
| 1275 | unsigned IID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); |
| 1276 | switch (IID) { |
| 1277 | default: |
| 1278 | return NULL; |
| 1279 | case Intrinsic::nvvm_ldg_global_f: |
| 1280 | case Intrinsic::nvvm_ldg_global_i: |
| 1281 | case Intrinsic::nvvm_ldg_global_p: |
| 1282 | IsLDG = true; |
| 1283 | break; |
| 1284 | case Intrinsic::nvvm_ldu_global_f: |
| 1285 | case Intrinsic::nvvm_ldu_global_i: |
| 1286 | case Intrinsic::nvvm_ldu_global_p: |
| 1287 | IsLDG = false; |
| 1288 | break; |
| 1289 | } |
| 1290 | } else { |
| 1291 | Op1 = N->getOperand(1); |
| 1292 | Mem = cast<MemSDNode>(N); |
| 1293 | } |
| 1294 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1295 | unsigned Opcode; |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 1296 | SDLoc DL(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1297 | SDNode *LD; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1298 | SDValue Base, Offset, Addr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 1299 | |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1300 | EVT EltVT = Mem->getMemoryVT(); |
| 1301 | if (EltVT.isVector()) { |
| 1302 | EltVT = EltVT.getVectorElementType(); |
| 1303 | } |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1304 | |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1305 | if (SelectDirectAddr(Op1, Addr)) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1306 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1307 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1308 | return nullptr; |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1309 | case ISD::INTRINSIC_W_CHAIN: |
| 1310 | if (IsLDG) { |
| 1311 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1312 | default: |
| 1313 | return nullptr; |
| 1314 | case MVT::i8: |
| 1315 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i8avar; |
| 1316 | break; |
| 1317 | case MVT::i16: |
| 1318 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i16avar; |
| 1319 | break; |
| 1320 | case MVT::i32: |
| 1321 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i32avar; |
| 1322 | break; |
| 1323 | case MVT::i64: |
| 1324 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i64avar; |
| 1325 | break; |
| 1326 | case MVT::f32: |
| 1327 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f32avar; |
| 1328 | break; |
| 1329 | case MVT::f64: |
| 1330 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f64avar; |
| 1331 | break; |
| 1332 | } |
| 1333 | } else { |
| 1334 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1335 | default: |
| 1336 | return nullptr; |
| 1337 | case MVT::i8: |
| 1338 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i8avar; |
| 1339 | break; |
| 1340 | case MVT::i16: |
| 1341 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i16avar; |
| 1342 | break; |
| 1343 | case MVT::i32: |
| 1344 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i32avar; |
| 1345 | break; |
| 1346 | case MVT::i64: |
| 1347 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i64avar; |
| 1348 | break; |
| 1349 | case MVT::f32: |
| 1350 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f32avar; |
| 1351 | break; |
| 1352 | case MVT::f64: |
| 1353 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f64avar; |
| 1354 | break; |
| 1355 | } |
| 1356 | } |
| 1357 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1358 | case NVPTXISD::LDGV2: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1359 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1360 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1361 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1362 | case MVT::i8: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1363 | Opcode = NVPTX::INT_PTX_LDG_G_v2i8_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1364 | break; |
| 1365 | case MVT::i16: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1366 | Opcode = NVPTX::INT_PTX_LDG_G_v2i16_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1367 | break; |
| 1368 | case MVT::i32: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1369 | Opcode = NVPTX::INT_PTX_LDG_G_v2i32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1370 | break; |
| 1371 | case MVT::i64: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1372 | Opcode = NVPTX::INT_PTX_LDG_G_v2i64_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1373 | break; |
| 1374 | case MVT::f32: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1375 | Opcode = NVPTX::INT_PTX_LDG_G_v2f32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1376 | break; |
| 1377 | case MVT::f64: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1378 | Opcode = NVPTX::INT_PTX_LDG_G_v2f64_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1379 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1380 | } |
| 1381 | break; |
| 1382 | case NVPTXISD::LDUV2: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1383 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1384 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1385 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1386 | case MVT::i8: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1387 | Opcode = NVPTX::INT_PTX_LDU_G_v2i8_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1388 | break; |
| 1389 | case MVT::i16: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1390 | Opcode = NVPTX::INT_PTX_LDU_G_v2i16_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1391 | break; |
| 1392 | case MVT::i32: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1393 | Opcode = NVPTX::INT_PTX_LDU_G_v2i32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1394 | break; |
| 1395 | case MVT::i64: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1396 | Opcode = NVPTX::INT_PTX_LDU_G_v2i64_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1397 | break; |
| 1398 | case MVT::f32: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1399 | Opcode = NVPTX::INT_PTX_LDU_G_v2f32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1400 | break; |
| 1401 | case MVT::f64: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1402 | Opcode = NVPTX::INT_PTX_LDU_G_v2f64_ELE_avar; |
| 1403 | break; |
| 1404 | } |
| 1405 | break; |
| 1406 | case NVPTXISD::LDGV4: |
| 1407 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1408 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1409 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1410 | case MVT::i8: |
| 1411 | Opcode = NVPTX::INT_PTX_LDG_G_v4i8_ELE_avar; |
| 1412 | break; |
| 1413 | case MVT::i16: |
| 1414 | Opcode = NVPTX::INT_PTX_LDG_G_v4i16_ELE_avar; |
| 1415 | break; |
| 1416 | case MVT::i32: |
| 1417 | Opcode = NVPTX::INT_PTX_LDG_G_v4i32_ELE_avar; |
| 1418 | break; |
| 1419 | case MVT::f32: |
| 1420 | Opcode = NVPTX::INT_PTX_LDG_G_v4f32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1421 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1422 | } |
| 1423 | break; |
| 1424 | case NVPTXISD::LDUV4: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1425 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1426 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1427 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1428 | case MVT::i8: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1429 | Opcode = NVPTX::INT_PTX_LDU_G_v4i8_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1430 | break; |
| 1431 | case MVT::i16: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1432 | Opcode = NVPTX::INT_PTX_LDU_G_v4i16_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1433 | break; |
| 1434 | case MVT::i32: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1435 | Opcode = NVPTX::INT_PTX_LDU_G_v4i32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1436 | break; |
| 1437 | case MVT::f32: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1438 | Opcode = NVPTX::INT_PTX_LDU_G_v4f32_ELE_avar; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1439 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1440 | } |
| 1441 | break; |
| 1442 | } |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1443 | |
| 1444 | SDValue Ops[] = { Addr, Chain }; |
Craig Topper | 2d2aa0c | 2014-04-30 07:17:30 +0000 | [diff] [blame] | 1445 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 1446 | } else if (TM.is64Bit() ? SelectADDRri64(Op1.getNode(), Op1, Base, Offset) |
| 1447 | : SelectADDRri(Op1.getNode(), Op1, Base, Offset)) { |
| 1448 | if (TM.is64Bit()) { |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1449 | switch (N->getOpcode()) { |
| 1450 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1451 | return nullptr; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1452 | case ISD::LOAD: |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1453 | case ISD::INTRINSIC_W_CHAIN: |
| 1454 | if (IsLDG) { |
| 1455 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1456 | default: |
| 1457 | return nullptr; |
| 1458 | case MVT::i8: |
| 1459 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i8ari64; |
| 1460 | break; |
| 1461 | case MVT::i16: |
| 1462 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i16ari64; |
| 1463 | break; |
| 1464 | case MVT::i32: |
| 1465 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i32ari64; |
| 1466 | break; |
| 1467 | case MVT::i64: |
| 1468 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i64ari64; |
| 1469 | break; |
| 1470 | case MVT::f32: |
| 1471 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f32ari64; |
| 1472 | break; |
| 1473 | case MVT::f64: |
| 1474 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f64ari64; |
| 1475 | break; |
| 1476 | } |
| 1477 | } else { |
| 1478 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1479 | default: |
| 1480 | return nullptr; |
| 1481 | case MVT::i8: |
| 1482 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i8ari64; |
| 1483 | break; |
| 1484 | case MVT::i16: |
| 1485 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i16ari64; |
| 1486 | break; |
| 1487 | case MVT::i32: |
| 1488 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i32ari64; |
| 1489 | break; |
| 1490 | case MVT::i64: |
| 1491 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i64ari64; |
| 1492 | break; |
| 1493 | case MVT::f32: |
| 1494 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f32ari64; |
| 1495 | break; |
| 1496 | case MVT::f64: |
| 1497 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f64ari64; |
| 1498 | break; |
| 1499 | } |
| 1500 | } |
| 1501 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1502 | case NVPTXISD::LoadV2: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1503 | case NVPTXISD::LDGV2: |
| 1504 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1505 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1506 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1507 | case MVT::i8: |
| 1508 | Opcode = NVPTX::INT_PTX_LDG_G_v2i8_ELE_ari64; |
| 1509 | break; |
| 1510 | case MVT::i16: |
| 1511 | Opcode = NVPTX::INT_PTX_LDG_G_v2i16_ELE_ari64; |
| 1512 | break; |
| 1513 | case MVT::i32: |
| 1514 | Opcode = NVPTX::INT_PTX_LDG_G_v2i32_ELE_ari64; |
| 1515 | break; |
| 1516 | case MVT::i64: |
| 1517 | Opcode = NVPTX::INT_PTX_LDG_G_v2i64_ELE_ari64; |
| 1518 | break; |
| 1519 | case MVT::f32: |
| 1520 | Opcode = NVPTX::INT_PTX_LDG_G_v2f32_ELE_ari64; |
| 1521 | break; |
| 1522 | case MVT::f64: |
| 1523 | Opcode = NVPTX::INT_PTX_LDG_G_v2f64_ELE_ari64; |
| 1524 | break; |
| 1525 | } |
| 1526 | break; |
| 1527 | case NVPTXISD::LDUV2: |
| 1528 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1529 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1530 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1531 | case MVT::i8: |
| 1532 | Opcode = NVPTX::INT_PTX_LDU_G_v2i8_ELE_ari64; |
| 1533 | break; |
| 1534 | case MVT::i16: |
| 1535 | Opcode = NVPTX::INT_PTX_LDU_G_v2i16_ELE_ari64; |
| 1536 | break; |
| 1537 | case MVT::i32: |
| 1538 | Opcode = NVPTX::INT_PTX_LDU_G_v2i32_ELE_ari64; |
| 1539 | break; |
| 1540 | case MVT::i64: |
| 1541 | Opcode = NVPTX::INT_PTX_LDU_G_v2i64_ELE_ari64; |
| 1542 | break; |
| 1543 | case MVT::f32: |
| 1544 | Opcode = NVPTX::INT_PTX_LDU_G_v2f32_ELE_ari64; |
| 1545 | break; |
| 1546 | case MVT::f64: |
| 1547 | Opcode = NVPTX::INT_PTX_LDU_G_v2f64_ELE_ari64; |
| 1548 | break; |
| 1549 | } |
| 1550 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1551 | case NVPTXISD::LoadV4: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1552 | case NVPTXISD::LDGV4: |
| 1553 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1554 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1555 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1556 | case MVT::i8: |
| 1557 | Opcode = NVPTX::INT_PTX_LDG_G_v4i8_ELE_ari64; |
| 1558 | break; |
| 1559 | case MVT::i16: |
| 1560 | Opcode = NVPTX::INT_PTX_LDG_G_v4i16_ELE_ari64; |
| 1561 | break; |
| 1562 | case MVT::i32: |
| 1563 | Opcode = NVPTX::INT_PTX_LDG_G_v4i32_ELE_ari64; |
| 1564 | break; |
| 1565 | case MVT::f32: |
| 1566 | Opcode = NVPTX::INT_PTX_LDG_G_v4f32_ELE_ari64; |
| 1567 | break; |
| 1568 | } |
| 1569 | break; |
| 1570 | case NVPTXISD::LDUV4: |
| 1571 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1572 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1573 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1574 | case MVT::i8: |
| 1575 | Opcode = NVPTX::INT_PTX_LDU_G_v4i8_ELE_ari64; |
| 1576 | break; |
| 1577 | case MVT::i16: |
| 1578 | Opcode = NVPTX::INT_PTX_LDU_G_v4i16_ELE_ari64; |
| 1579 | break; |
| 1580 | case MVT::i32: |
| 1581 | Opcode = NVPTX::INT_PTX_LDU_G_v4i32_ELE_ari64; |
| 1582 | break; |
| 1583 | case MVT::f32: |
| 1584 | Opcode = NVPTX::INT_PTX_LDU_G_v4f32_ELE_ari64; |
| 1585 | break; |
| 1586 | } |
| 1587 | break; |
| 1588 | } |
| 1589 | } else { |
| 1590 | switch (N->getOpcode()) { |
| 1591 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1592 | return nullptr; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1593 | case ISD::LOAD: |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1594 | case ISD::INTRINSIC_W_CHAIN: |
| 1595 | if (IsLDG) { |
| 1596 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1597 | default: |
| 1598 | return nullptr; |
| 1599 | case MVT::i8: |
| 1600 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i8ari; |
| 1601 | break; |
| 1602 | case MVT::i16: |
| 1603 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i16ari; |
| 1604 | break; |
| 1605 | case MVT::i32: |
| 1606 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i32ari; |
| 1607 | break; |
| 1608 | case MVT::i64: |
| 1609 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i64ari; |
| 1610 | break; |
| 1611 | case MVT::f32: |
| 1612 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f32ari; |
| 1613 | break; |
| 1614 | case MVT::f64: |
| 1615 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f64ari; |
| 1616 | break; |
| 1617 | } |
| 1618 | } else { |
| 1619 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1620 | default: |
| 1621 | return nullptr; |
| 1622 | case MVT::i8: |
| 1623 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i8ari; |
| 1624 | break; |
| 1625 | case MVT::i16: |
| 1626 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i16ari; |
| 1627 | break; |
| 1628 | case MVT::i32: |
| 1629 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i32ari; |
| 1630 | break; |
| 1631 | case MVT::i64: |
| 1632 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i64ari; |
| 1633 | break; |
| 1634 | case MVT::f32: |
| 1635 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f32ari; |
| 1636 | break; |
| 1637 | case MVT::f64: |
| 1638 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f64ari; |
| 1639 | break; |
| 1640 | } |
| 1641 | } |
| 1642 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1643 | case NVPTXISD::LoadV2: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1644 | case NVPTXISD::LDGV2: |
| 1645 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1646 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1647 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1648 | case MVT::i8: |
| 1649 | Opcode = NVPTX::INT_PTX_LDG_G_v2i8_ELE_ari32; |
| 1650 | break; |
| 1651 | case MVT::i16: |
| 1652 | Opcode = NVPTX::INT_PTX_LDG_G_v2i16_ELE_ari32; |
| 1653 | break; |
| 1654 | case MVT::i32: |
| 1655 | Opcode = NVPTX::INT_PTX_LDG_G_v2i32_ELE_ari32; |
| 1656 | break; |
| 1657 | case MVT::i64: |
| 1658 | Opcode = NVPTX::INT_PTX_LDG_G_v2i64_ELE_ari32; |
| 1659 | break; |
| 1660 | case MVT::f32: |
| 1661 | Opcode = NVPTX::INT_PTX_LDG_G_v2f32_ELE_ari32; |
| 1662 | break; |
| 1663 | case MVT::f64: |
| 1664 | Opcode = NVPTX::INT_PTX_LDG_G_v2f64_ELE_ari32; |
| 1665 | break; |
| 1666 | } |
| 1667 | break; |
| 1668 | case NVPTXISD::LDUV2: |
| 1669 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1670 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1671 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1672 | case MVT::i8: |
| 1673 | Opcode = NVPTX::INT_PTX_LDU_G_v2i8_ELE_ari32; |
| 1674 | break; |
| 1675 | case MVT::i16: |
| 1676 | Opcode = NVPTX::INT_PTX_LDU_G_v2i16_ELE_ari32; |
| 1677 | break; |
| 1678 | case MVT::i32: |
| 1679 | Opcode = NVPTX::INT_PTX_LDU_G_v2i32_ELE_ari32; |
| 1680 | break; |
| 1681 | case MVT::i64: |
| 1682 | Opcode = NVPTX::INT_PTX_LDU_G_v2i64_ELE_ari32; |
| 1683 | break; |
| 1684 | case MVT::f32: |
| 1685 | Opcode = NVPTX::INT_PTX_LDU_G_v2f32_ELE_ari32; |
| 1686 | break; |
| 1687 | case MVT::f64: |
| 1688 | Opcode = NVPTX::INT_PTX_LDU_G_v2f64_ELE_ari32; |
| 1689 | break; |
| 1690 | } |
| 1691 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1692 | case NVPTXISD::LoadV4: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1693 | case NVPTXISD::LDGV4: |
| 1694 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1695 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1696 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1697 | case MVT::i8: |
| 1698 | Opcode = NVPTX::INT_PTX_LDG_G_v4i8_ELE_ari32; |
| 1699 | break; |
| 1700 | case MVT::i16: |
| 1701 | Opcode = NVPTX::INT_PTX_LDG_G_v4i16_ELE_ari32; |
| 1702 | break; |
| 1703 | case MVT::i32: |
| 1704 | Opcode = NVPTX::INT_PTX_LDG_G_v4i32_ELE_ari32; |
| 1705 | break; |
| 1706 | case MVT::f32: |
| 1707 | Opcode = NVPTX::INT_PTX_LDG_G_v4f32_ELE_ari32; |
| 1708 | break; |
| 1709 | } |
| 1710 | break; |
| 1711 | case NVPTXISD::LDUV4: |
| 1712 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1713 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1714 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1715 | case MVT::i8: |
| 1716 | Opcode = NVPTX::INT_PTX_LDU_G_v4i8_ELE_ari32; |
| 1717 | break; |
| 1718 | case MVT::i16: |
| 1719 | Opcode = NVPTX::INT_PTX_LDU_G_v4i16_ELE_ari32; |
| 1720 | break; |
| 1721 | case MVT::i32: |
| 1722 | Opcode = NVPTX::INT_PTX_LDU_G_v4i32_ELE_ari32; |
| 1723 | break; |
| 1724 | case MVT::f32: |
| 1725 | Opcode = NVPTX::INT_PTX_LDU_G_v4f32_ELE_ari32; |
| 1726 | break; |
| 1727 | } |
| 1728 | break; |
| 1729 | } |
| 1730 | } |
| 1731 | |
| 1732 | SDValue Ops[] = { Base, Offset, Chain }; |
| 1733 | |
Craig Topper | 2d2aa0c | 2014-04-30 07:17:30 +0000 | [diff] [blame] | 1734 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1735 | } else { |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 1736 | if (TM.is64Bit()) { |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1737 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1738 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1739 | return nullptr; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1740 | case ISD::LOAD: |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1741 | case ISD::INTRINSIC_W_CHAIN: |
| 1742 | if (IsLDG) { |
| 1743 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1744 | default: |
| 1745 | return nullptr; |
| 1746 | case MVT::i8: |
| 1747 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i8areg64; |
| 1748 | break; |
| 1749 | case MVT::i16: |
| 1750 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i16areg64; |
| 1751 | break; |
| 1752 | case MVT::i32: |
| 1753 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i32areg64; |
| 1754 | break; |
| 1755 | case MVT::i64: |
| 1756 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i64areg64; |
| 1757 | break; |
| 1758 | case MVT::f32: |
| 1759 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f32areg64; |
| 1760 | break; |
| 1761 | case MVT::f64: |
| 1762 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f64areg64; |
| 1763 | break; |
| 1764 | } |
| 1765 | } else { |
| 1766 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1767 | default: |
| 1768 | return nullptr; |
| 1769 | case MVT::i8: |
| 1770 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i8areg64; |
| 1771 | break; |
| 1772 | case MVT::i16: |
| 1773 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i16areg64; |
| 1774 | break; |
| 1775 | case MVT::i32: |
| 1776 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i32areg64; |
| 1777 | break; |
| 1778 | case MVT::i64: |
| 1779 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i64areg64; |
| 1780 | break; |
| 1781 | case MVT::f32: |
| 1782 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f32areg64; |
| 1783 | break; |
| 1784 | case MVT::f64: |
| 1785 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f64areg64; |
| 1786 | break; |
| 1787 | } |
| 1788 | } |
| 1789 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1790 | case NVPTXISD::LoadV2: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1791 | case NVPTXISD::LDGV2: |
| 1792 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1793 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1794 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1795 | case MVT::i8: |
| 1796 | Opcode = NVPTX::INT_PTX_LDG_G_v2i8_ELE_areg64; |
| 1797 | break; |
| 1798 | case MVT::i16: |
| 1799 | Opcode = NVPTX::INT_PTX_LDG_G_v2i16_ELE_areg64; |
| 1800 | break; |
| 1801 | case MVT::i32: |
| 1802 | Opcode = NVPTX::INT_PTX_LDG_G_v2i32_ELE_areg64; |
| 1803 | break; |
| 1804 | case MVT::i64: |
| 1805 | Opcode = NVPTX::INT_PTX_LDG_G_v2i64_ELE_areg64; |
| 1806 | break; |
| 1807 | case MVT::f32: |
| 1808 | Opcode = NVPTX::INT_PTX_LDG_G_v2f32_ELE_areg64; |
| 1809 | break; |
| 1810 | case MVT::f64: |
| 1811 | Opcode = NVPTX::INT_PTX_LDG_G_v2f64_ELE_areg64; |
| 1812 | break; |
| 1813 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1814 | break; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1815 | case NVPTXISD::LDUV2: |
| 1816 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1817 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1818 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1819 | case MVT::i8: |
| 1820 | Opcode = NVPTX::INT_PTX_LDU_G_v2i8_ELE_areg64; |
| 1821 | break; |
| 1822 | case MVT::i16: |
| 1823 | Opcode = NVPTX::INT_PTX_LDU_G_v2i16_ELE_areg64; |
| 1824 | break; |
| 1825 | case MVT::i32: |
| 1826 | Opcode = NVPTX::INT_PTX_LDU_G_v2i32_ELE_areg64; |
| 1827 | break; |
| 1828 | case MVT::i64: |
| 1829 | Opcode = NVPTX::INT_PTX_LDU_G_v2i64_ELE_areg64; |
| 1830 | break; |
| 1831 | case MVT::f32: |
| 1832 | Opcode = NVPTX::INT_PTX_LDU_G_v2f32_ELE_areg64; |
| 1833 | break; |
| 1834 | case MVT::f64: |
| 1835 | Opcode = NVPTX::INT_PTX_LDU_G_v2f64_ELE_areg64; |
| 1836 | break; |
| 1837 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1838 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1839 | case NVPTXISD::LoadV4: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1840 | case NVPTXISD::LDGV4: |
| 1841 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1842 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1843 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1844 | case MVT::i8: |
| 1845 | Opcode = NVPTX::INT_PTX_LDG_G_v4i8_ELE_areg64; |
| 1846 | break; |
| 1847 | case MVT::i16: |
| 1848 | Opcode = NVPTX::INT_PTX_LDG_G_v4i16_ELE_areg64; |
| 1849 | break; |
| 1850 | case MVT::i32: |
| 1851 | Opcode = NVPTX::INT_PTX_LDG_G_v4i32_ELE_areg64; |
| 1852 | break; |
| 1853 | case MVT::f32: |
| 1854 | Opcode = NVPTX::INT_PTX_LDG_G_v4f32_ELE_areg64; |
| 1855 | break; |
| 1856 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1857 | break; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1858 | case NVPTXISD::LDUV4: |
| 1859 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1860 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1861 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1862 | case MVT::i8: |
| 1863 | Opcode = NVPTX::INT_PTX_LDU_G_v4i8_ELE_areg64; |
| 1864 | break; |
| 1865 | case MVT::i16: |
| 1866 | Opcode = NVPTX::INT_PTX_LDU_G_v4i16_ELE_areg64; |
| 1867 | break; |
| 1868 | case MVT::i32: |
| 1869 | Opcode = NVPTX::INT_PTX_LDU_G_v4i32_ELE_areg64; |
| 1870 | break; |
| 1871 | case MVT::f32: |
| 1872 | Opcode = NVPTX::INT_PTX_LDU_G_v4f32_ELE_areg64; |
| 1873 | break; |
| 1874 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1875 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 1876 | } |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1877 | } else { |
| 1878 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1879 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1880 | return nullptr; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1881 | case ISD::LOAD: |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 1882 | case ISD::INTRINSIC_W_CHAIN: |
| 1883 | if (IsLDG) { |
| 1884 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1885 | default: |
| 1886 | return nullptr; |
| 1887 | case MVT::i8: |
| 1888 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i8areg; |
| 1889 | break; |
| 1890 | case MVT::i16: |
| 1891 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i16areg; |
| 1892 | break; |
| 1893 | case MVT::i32: |
| 1894 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i32areg; |
| 1895 | break; |
| 1896 | case MVT::i64: |
| 1897 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_i64areg; |
| 1898 | break; |
| 1899 | case MVT::f32: |
| 1900 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f32areg; |
| 1901 | break; |
| 1902 | case MVT::f64: |
| 1903 | Opcode = NVPTX::INT_PTX_LDG_GLOBAL_f64areg; |
| 1904 | break; |
| 1905 | } |
| 1906 | } else { |
| 1907 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1908 | default: |
| 1909 | return nullptr; |
| 1910 | case MVT::i8: |
| 1911 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i8areg; |
| 1912 | break; |
| 1913 | case MVT::i16: |
| 1914 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i16areg; |
| 1915 | break; |
| 1916 | case MVT::i32: |
| 1917 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i32areg; |
| 1918 | break; |
| 1919 | case MVT::i64: |
| 1920 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_i64areg; |
| 1921 | break; |
| 1922 | case MVT::f32: |
| 1923 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f32areg; |
| 1924 | break; |
| 1925 | case MVT::f64: |
| 1926 | Opcode = NVPTX::INT_PTX_LDU_GLOBAL_f64areg; |
| 1927 | break; |
| 1928 | } |
| 1929 | } |
| 1930 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1931 | case NVPTXISD::LoadV2: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1932 | case NVPTXISD::LDGV2: |
| 1933 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1934 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1935 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1936 | case MVT::i8: |
| 1937 | Opcode = NVPTX::INT_PTX_LDG_G_v2i8_ELE_areg32; |
| 1938 | break; |
| 1939 | case MVT::i16: |
| 1940 | Opcode = NVPTX::INT_PTX_LDG_G_v2i16_ELE_areg32; |
| 1941 | break; |
| 1942 | case MVT::i32: |
| 1943 | Opcode = NVPTX::INT_PTX_LDG_G_v2i32_ELE_areg32; |
| 1944 | break; |
| 1945 | case MVT::i64: |
| 1946 | Opcode = NVPTX::INT_PTX_LDG_G_v2i64_ELE_areg32; |
| 1947 | break; |
| 1948 | case MVT::f32: |
| 1949 | Opcode = NVPTX::INT_PTX_LDG_G_v2f32_ELE_areg32; |
| 1950 | break; |
| 1951 | case MVT::f64: |
| 1952 | Opcode = NVPTX::INT_PTX_LDG_G_v2f64_ELE_areg32; |
| 1953 | break; |
| 1954 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1955 | break; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1956 | case NVPTXISD::LDUV2: |
| 1957 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1958 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1959 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1960 | case MVT::i8: |
| 1961 | Opcode = NVPTX::INT_PTX_LDU_G_v2i8_ELE_areg32; |
| 1962 | break; |
| 1963 | case MVT::i16: |
| 1964 | Opcode = NVPTX::INT_PTX_LDU_G_v2i16_ELE_areg32; |
| 1965 | break; |
| 1966 | case MVT::i32: |
| 1967 | Opcode = NVPTX::INT_PTX_LDU_G_v2i32_ELE_areg32; |
| 1968 | break; |
| 1969 | case MVT::i64: |
| 1970 | Opcode = NVPTX::INT_PTX_LDU_G_v2i64_ELE_areg32; |
| 1971 | break; |
| 1972 | case MVT::f32: |
| 1973 | Opcode = NVPTX::INT_PTX_LDU_G_v2f32_ELE_areg32; |
| 1974 | break; |
| 1975 | case MVT::f64: |
| 1976 | Opcode = NVPTX::INT_PTX_LDU_G_v2f64_ELE_areg32; |
| 1977 | break; |
| 1978 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1979 | break; |
Jingyue Wu | 48a9bdc | 2015-07-20 21:28:54 +0000 | [diff] [blame] | 1980 | case NVPTXISD::LoadV4: |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1981 | case NVPTXISD::LDGV4: |
| 1982 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 1983 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 1984 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1985 | case MVT::i8: |
| 1986 | Opcode = NVPTX::INT_PTX_LDG_G_v4i8_ELE_areg32; |
| 1987 | break; |
| 1988 | case MVT::i16: |
| 1989 | Opcode = NVPTX::INT_PTX_LDG_G_v4i16_ELE_areg32; |
| 1990 | break; |
| 1991 | case MVT::i32: |
| 1992 | Opcode = NVPTX::INT_PTX_LDG_G_v4i32_ELE_areg32; |
| 1993 | break; |
| 1994 | case MVT::f32: |
| 1995 | Opcode = NVPTX::INT_PTX_LDG_G_v4f32_ELE_areg32; |
| 1996 | break; |
| 1997 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 1998 | break; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 1999 | case NVPTXISD::LDUV4: |
| 2000 | switch (EltVT.getSimpleVT().SimpleTy) { |
| 2001 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2002 | return nullptr; |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 2003 | case MVT::i8: |
| 2004 | Opcode = NVPTX::INT_PTX_LDU_G_v4i8_ELE_areg32; |
| 2005 | break; |
| 2006 | case MVT::i16: |
| 2007 | Opcode = NVPTX::INT_PTX_LDU_G_v4i16_ELE_areg32; |
| 2008 | break; |
| 2009 | case MVT::i32: |
| 2010 | Opcode = NVPTX::INT_PTX_LDU_G_v4i32_ELE_areg32; |
| 2011 | break; |
| 2012 | case MVT::f32: |
| 2013 | Opcode = NVPTX::INT_PTX_LDU_G_v4f32_ELE_areg32; |
| 2014 | break; |
| 2015 | } |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2016 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2017 | } |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2018 | } |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2019 | |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 2020 | SDValue Ops[] = { Op1, Chain }; |
Craig Topper | 2d2aa0c | 2014-04-30 07:17:30 +0000 | [diff] [blame] | 2021 | LD = CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops); |
Justin Holewinski | e40e929 | 2013-07-01 12:58:52 +0000 | [diff] [blame] | 2022 | } |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2023 | |
| 2024 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
Justin Holewinski | b926d9d | 2014-06-27 18:35:51 +0000 | [diff] [blame] | 2025 | MemRefs0[0] = Mem->getMemOperand(); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2026 | cast<MachineSDNode>(LD)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 2027 | |
| 2028 | return LD; |
| 2029 | } |
| 2030 | |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2031 | SDNode *NVPTXDAGToDAGISel::SelectStore(SDNode *N) { |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 2032 | SDLoc dl(N); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2033 | StoreSDNode *ST = cast<StoreSDNode>(N); |
| 2034 | EVT StoreVT = ST->getMemoryVT(); |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2035 | SDNode *NVPTXST = nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2036 | |
| 2037 | // do not support pre/post inc/dec |
| 2038 | if (ST->isIndexed()) |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2039 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2040 | |
| 2041 | if (!StoreVT.isSimple()) |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2042 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2043 | |
| 2044 | // Address Space Setting |
Eric Christopher | 9745b3a | 2015-01-30 01:41:01 +0000 | [diff] [blame] | 2045 | unsigned int codeAddrSpace = getCodeAddrSpace(ST); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2046 | |
| 2047 | // Volatile Setting |
| 2048 | // - .volatile is only availalble for .global and .shared |
| 2049 | bool isVolatile = ST->isVolatile(); |
| 2050 | if (codeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL && |
| 2051 | codeAddrSpace != NVPTX::PTXLdStInstCode::SHARED && |
| 2052 | codeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC) |
| 2053 | isVolatile = false; |
| 2054 | |
| 2055 | // Vector Setting |
| 2056 | MVT SimpleVT = StoreVT.getSimpleVT(); |
| 2057 | unsigned vecType = NVPTX::PTXLdStInstCode::Scalar; |
| 2058 | if (SimpleVT.isVector()) { |
| 2059 | unsigned num = SimpleVT.getVectorNumElements(); |
| 2060 | if (num == 2) |
| 2061 | vecType = NVPTX::PTXLdStInstCode::V2; |
| 2062 | else if (num == 4) |
| 2063 | vecType = NVPTX::PTXLdStInstCode::V4; |
| 2064 | else |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2065 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2066 | } |
| 2067 | |
| 2068 | // Type Setting: toType + toTypeWidth |
| 2069 | // - for integer type, always use 'u' |
| 2070 | // |
| 2071 | MVT ScalarVT = SimpleVT.getScalarType(); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2072 | unsigned toTypeWidth = ScalarVT.getSizeInBits(); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2073 | unsigned int toType; |
| 2074 | if (ScalarVT.isFloatingPoint()) |
| 2075 | toType = NVPTX::PTXLdStInstCode::Float; |
| 2076 | else |
| 2077 | toType = NVPTX::PTXLdStInstCode::Unsigned; |
| 2078 | |
| 2079 | // Create the machine instruction DAG |
| 2080 | SDValue Chain = N->getOperand(0); |
| 2081 | SDValue N1 = N->getOperand(1); |
| 2082 | SDValue N2 = N->getOperand(2); |
| 2083 | SDValue Addr; |
| 2084 | SDValue Offset, Base; |
| 2085 | unsigned Opcode; |
Craig Topper | d9c2783 | 2013-08-15 02:44:19 +0000 | [diff] [blame] | 2086 | MVT::SimpleValueType SourceVT = N1.getNode()->getSimpleValueType(0).SimpleTy; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2087 | |
| 2088 | if (SelectDirectAddr(N2, Addr)) { |
| 2089 | switch (SourceVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2090 | case MVT::i8: |
| 2091 | Opcode = NVPTX::ST_i8_avar; |
| 2092 | break; |
| 2093 | case MVT::i16: |
| 2094 | Opcode = NVPTX::ST_i16_avar; |
| 2095 | break; |
| 2096 | case MVT::i32: |
| 2097 | Opcode = NVPTX::ST_i32_avar; |
| 2098 | break; |
| 2099 | case MVT::i64: |
| 2100 | Opcode = NVPTX::ST_i64_avar; |
| 2101 | break; |
| 2102 | case MVT::f32: |
| 2103 | Opcode = NVPTX::ST_f32_avar; |
| 2104 | break; |
| 2105 | case MVT::f64: |
| 2106 | Opcode = NVPTX::ST_f64_avar; |
| 2107 | break; |
| 2108 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2109 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2110 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2111 | SDValue Ops[] = { N1, getI32Imm(isVolatile, dl), |
| 2112 | getI32Imm(codeAddrSpace, dl), getI32Imm(vecType, dl), |
| 2113 | getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), Addr, |
| 2114 | Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 2115 | NVPTXST = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 2116 | } else if (TM.is64Bit() ? SelectADDRsi64(N2.getNode(), N2, Base, Offset) |
| 2117 | : SelectADDRsi(N2.getNode(), N2, Base, Offset)) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2118 | switch (SourceVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2119 | case MVT::i8: |
| 2120 | Opcode = NVPTX::ST_i8_asi; |
| 2121 | break; |
| 2122 | case MVT::i16: |
| 2123 | Opcode = NVPTX::ST_i16_asi; |
| 2124 | break; |
| 2125 | case MVT::i32: |
| 2126 | Opcode = NVPTX::ST_i32_asi; |
| 2127 | break; |
| 2128 | case MVT::i64: |
| 2129 | Opcode = NVPTX::ST_i64_asi; |
| 2130 | break; |
| 2131 | case MVT::f32: |
| 2132 | Opcode = NVPTX::ST_f32_asi; |
| 2133 | break; |
| 2134 | case MVT::f64: |
| 2135 | Opcode = NVPTX::ST_f64_asi; |
| 2136 | break; |
| 2137 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2138 | return nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2139 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2140 | SDValue Ops[] = { N1, getI32Imm(isVolatile, dl), |
| 2141 | getI32Imm(codeAddrSpace, dl), getI32Imm(vecType, dl), |
| 2142 | getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), Base, |
| 2143 | Offset, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 2144 | NVPTXST = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 2145 | } else if (TM.is64Bit() ? SelectADDRri64(N2.getNode(), N2, Base, Offset) |
| 2146 | : SelectADDRri(N2.getNode(), N2, Base, Offset)) { |
| 2147 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2148 | switch (SourceVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2149 | case MVT::i8: |
| 2150 | Opcode = NVPTX::ST_i8_ari_64; |
| 2151 | break; |
| 2152 | case MVT::i16: |
| 2153 | Opcode = NVPTX::ST_i16_ari_64; |
| 2154 | break; |
| 2155 | case MVT::i32: |
| 2156 | Opcode = NVPTX::ST_i32_ari_64; |
| 2157 | break; |
| 2158 | case MVT::i64: |
| 2159 | Opcode = NVPTX::ST_i64_ari_64; |
| 2160 | break; |
| 2161 | case MVT::f32: |
| 2162 | Opcode = NVPTX::ST_f32_ari_64; |
| 2163 | break; |
| 2164 | case MVT::f64: |
| 2165 | Opcode = NVPTX::ST_f64_ari_64; |
| 2166 | break; |
| 2167 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2168 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2169 | } |
| 2170 | } else { |
| 2171 | switch (SourceVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2172 | case MVT::i8: |
| 2173 | Opcode = NVPTX::ST_i8_ari; |
| 2174 | break; |
| 2175 | case MVT::i16: |
| 2176 | Opcode = NVPTX::ST_i16_ari; |
| 2177 | break; |
| 2178 | case MVT::i32: |
| 2179 | Opcode = NVPTX::ST_i32_ari; |
| 2180 | break; |
| 2181 | case MVT::i64: |
| 2182 | Opcode = NVPTX::ST_i64_ari; |
| 2183 | break; |
| 2184 | case MVT::f32: |
| 2185 | Opcode = NVPTX::ST_f32_ari; |
| 2186 | break; |
| 2187 | case MVT::f64: |
| 2188 | Opcode = NVPTX::ST_f64_ari; |
| 2189 | break; |
| 2190 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2191 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2192 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2193 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2194 | SDValue Ops[] = { N1, getI32Imm(isVolatile, dl), |
| 2195 | getI32Imm(codeAddrSpace, dl), getI32Imm(vecType, dl), |
| 2196 | getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), Base, |
| 2197 | Offset, Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 2198 | NVPTXST = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2199 | } else { |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 2200 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2201 | switch (SourceVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2202 | case MVT::i8: |
| 2203 | Opcode = NVPTX::ST_i8_areg_64; |
| 2204 | break; |
| 2205 | case MVT::i16: |
| 2206 | Opcode = NVPTX::ST_i16_areg_64; |
| 2207 | break; |
| 2208 | case MVT::i32: |
| 2209 | Opcode = NVPTX::ST_i32_areg_64; |
| 2210 | break; |
| 2211 | case MVT::i64: |
| 2212 | Opcode = NVPTX::ST_i64_areg_64; |
| 2213 | break; |
| 2214 | case MVT::f32: |
| 2215 | Opcode = NVPTX::ST_f32_areg_64; |
| 2216 | break; |
| 2217 | case MVT::f64: |
| 2218 | Opcode = NVPTX::ST_f64_areg_64; |
| 2219 | break; |
| 2220 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2221 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2222 | } |
| 2223 | } else { |
| 2224 | switch (SourceVT) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2225 | case MVT::i8: |
| 2226 | Opcode = NVPTX::ST_i8_areg; |
| 2227 | break; |
| 2228 | case MVT::i16: |
| 2229 | Opcode = NVPTX::ST_i16_areg; |
| 2230 | break; |
| 2231 | case MVT::i32: |
| 2232 | Opcode = NVPTX::ST_i32_areg; |
| 2233 | break; |
| 2234 | case MVT::i64: |
| 2235 | Opcode = NVPTX::ST_i64_areg; |
| 2236 | break; |
| 2237 | case MVT::f32: |
| 2238 | Opcode = NVPTX::ST_f32_areg; |
| 2239 | break; |
| 2240 | case MVT::f64: |
| 2241 | Opcode = NVPTX::ST_f64_areg; |
| 2242 | break; |
| 2243 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2244 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2245 | } |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2246 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2247 | SDValue Ops[] = { N1, getI32Imm(isVolatile, dl), |
| 2248 | getI32Imm(codeAddrSpace, dl), getI32Imm(vecType, dl), |
| 2249 | getI32Imm(toType, dl), getI32Imm(toTypeWidth, dl), N2, |
| 2250 | Chain }; |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 2251 | NVPTXST = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2252 | } |
| 2253 | |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2254 | if (NVPTXST) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 2255 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
| 2256 | MemRefs0[0] = cast<MemSDNode>(N)->getMemOperand(); |
| 2257 | cast<MachineSDNode>(NVPTXST)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 2258 | } |
| 2259 | |
| 2260 | return NVPTXST; |
| 2261 | } |
| 2262 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2263 | SDNode *NVPTXDAGToDAGISel::SelectStoreVector(SDNode *N) { |
| 2264 | SDValue Chain = N->getOperand(0); |
| 2265 | SDValue Op1 = N->getOperand(1); |
| 2266 | SDValue Addr, Offset, Base; |
| 2267 | unsigned Opcode; |
Andrew Trick | ef9de2a | 2013-05-25 02:42:55 +0000 | [diff] [blame] | 2268 | SDLoc DL(N); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2269 | SDNode *ST; |
| 2270 | EVT EltVT = Op1.getValueType(); |
| 2271 | MemSDNode *MemSD = cast<MemSDNode>(N); |
| 2272 | EVT StoreVT = MemSD->getMemoryVT(); |
| 2273 | |
| 2274 | // Address Space Setting |
Eric Christopher | 9745b3a | 2015-01-30 01:41:01 +0000 | [diff] [blame] | 2275 | unsigned CodeAddrSpace = getCodeAddrSpace(MemSD); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2276 | |
| 2277 | if (CodeAddrSpace == NVPTX::PTXLdStInstCode::CONSTANT) { |
| 2278 | report_fatal_error("Cannot store to pointer that points to constant " |
| 2279 | "memory space"); |
| 2280 | } |
| 2281 | |
| 2282 | // Volatile Setting |
| 2283 | // - .volatile is only availalble for .global and .shared |
| 2284 | bool IsVolatile = MemSD->isVolatile(); |
| 2285 | if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL && |
| 2286 | CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED && |
| 2287 | CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC) |
| 2288 | IsVolatile = false; |
| 2289 | |
| 2290 | // Type Setting: toType + toTypeWidth |
| 2291 | // - for integer type, always use 'u' |
| 2292 | assert(StoreVT.isSimple() && "Store value is not simple"); |
| 2293 | MVT ScalarVT = StoreVT.getSimpleVT().getScalarType(); |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2294 | unsigned ToTypeWidth = ScalarVT.getSizeInBits(); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2295 | unsigned ToType; |
| 2296 | if (ScalarVT.isFloatingPoint()) |
| 2297 | ToType = NVPTX::PTXLdStInstCode::Float; |
| 2298 | else |
| 2299 | ToType = NVPTX::PTXLdStInstCode::Unsigned; |
| 2300 | |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2301 | SmallVector<SDValue, 12> StOps; |
| 2302 | SDValue N2; |
| 2303 | unsigned VecType; |
| 2304 | |
| 2305 | switch (N->getOpcode()) { |
| 2306 | case NVPTXISD::StoreV2: |
| 2307 | VecType = NVPTX::PTXLdStInstCode::V2; |
| 2308 | StOps.push_back(N->getOperand(1)); |
| 2309 | StOps.push_back(N->getOperand(2)); |
| 2310 | N2 = N->getOperand(3); |
| 2311 | break; |
| 2312 | case NVPTXISD::StoreV4: |
| 2313 | VecType = NVPTX::PTXLdStInstCode::V4; |
| 2314 | StOps.push_back(N->getOperand(1)); |
| 2315 | StOps.push_back(N->getOperand(2)); |
| 2316 | StOps.push_back(N->getOperand(3)); |
| 2317 | StOps.push_back(N->getOperand(4)); |
| 2318 | N2 = N->getOperand(5); |
| 2319 | break; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2320 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2321 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2322 | } |
| 2323 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2324 | StOps.push_back(getI32Imm(IsVolatile, DL)); |
| 2325 | StOps.push_back(getI32Imm(CodeAddrSpace, DL)); |
| 2326 | StOps.push_back(getI32Imm(VecType, DL)); |
| 2327 | StOps.push_back(getI32Imm(ToType, DL)); |
| 2328 | StOps.push_back(getI32Imm(ToTypeWidth, DL)); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2329 | |
| 2330 | if (SelectDirectAddr(N2, Addr)) { |
| 2331 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2332 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2333 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2334 | case NVPTXISD::StoreV2: |
| 2335 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2336 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2337 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2338 | case MVT::i8: |
| 2339 | Opcode = NVPTX::STV_i8_v2_avar; |
| 2340 | break; |
| 2341 | case MVT::i16: |
| 2342 | Opcode = NVPTX::STV_i16_v2_avar; |
| 2343 | break; |
| 2344 | case MVT::i32: |
| 2345 | Opcode = NVPTX::STV_i32_v2_avar; |
| 2346 | break; |
| 2347 | case MVT::i64: |
| 2348 | Opcode = NVPTX::STV_i64_v2_avar; |
| 2349 | break; |
| 2350 | case MVT::f32: |
| 2351 | Opcode = NVPTX::STV_f32_v2_avar; |
| 2352 | break; |
| 2353 | case MVT::f64: |
| 2354 | Opcode = NVPTX::STV_f64_v2_avar; |
| 2355 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2356 | } |
| 2357 | break; |
| 2358 | case NVPTXISD::StoreV4: |
| 2359 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2360 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2361 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2362 | case MVT::i8: |
| 2363 | Opcode = NVPTX::STV_i8_v4_avar; |
| 2364 | break; |
| 2365 | case MVT::i16: |
| 2366 | Opcode = NVPTX::STV_i16_v4_avar; |
| 2367 | break; |
| 2368 | case MVT::i32: |
| 2369 | Opcode = NVPTX::STV_i32_v4_avar; |
| 2370 | break; |
| 2371 | case MVT::f32: |
| 2372 | Opcode = NVPTX::STV_f32_v4_avar; |
| 2373 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2374 | } |
| 2375 | break; |
| 2376 | } |
| 2377 | StOps.push_back(Addr); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 2378 | } else if (TM.is64Bit() ? SelectADDRsi64(N2.getNode(), N2, Base, Offset) |
| 2379 | : SelectADDRsi(N2.getNode(), N2, Base, Offset)) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2380 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2381 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2382 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2383 | case NVPTXISD::StoreV2: |
| 2384 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2385 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2386 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2387 | case MVT::i8: |
| 2388 | Opcode = NVPTX::STV_i8_v2_asi; |
| 2389 | break; |
| 2390 | case MVT::i16: |
| 2391 | Opcode = NVPTX::STV_i16_v2_asi; |
| 2392 | break; |
| 2393 | case MVT::i32: |
| 2394 | Opcode = NVPTX::STV_i32_v2_asi; |
| 2395 | break; |
| 2396 | case MVT::i64: |
| 2397 | Opcode = NVPTX::STV_i64_v2_asi; |
| 2398 | break; |
| 2399 | case MVT::f32: |
| 2400 | Opcode = NVPTX::STV_f32_v2_asi; |
| 2401 | break; |
| 2402 | case MVT::f64: |
| 2403 | Opcode = NVPTX::STV_f64_v2_asi; |
| 2404 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2405 | } |
| 2406 | break; |
| 2407 | case NVPTXISD::StoreV4: |
| 2408 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2409 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2410 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2411 | case MVT::i8: |
| 2412 | Opcode = NVPTX::STV_i8_v4_asi; |
| 2413 | break; |
| 2414 | case MVT::i16: |
| 2415 | Opcode = NVPTX::STV_i16_v4_asi; |
| 2416 | break; |
| 2417 | case MVT::i32: |
| 2418 | Opcode = NVPTX::STV_i32_v4_asi; |
| 2419 | break; |
| 2420 | case MVT::f32: |
| 2421 | Opcode = NVPTX::STV_f32_v4_asi; |
| 2422 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2423 | } |
| 2424 | break; |
| 2425 | } |
| 2426 | StOps.push_back(Base); |
| 2427 | StOps.push_back(Offset); |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 2428 | } else if (TM.is64Bit() ? SelectADDRri64(N2.getNode(), N2, Base, Offset) |
| 2429 | : SelectADDRri(N2.getNode(), N2, Base, Offset)) { |
| 2430 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2431 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2432 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2433 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2434 | case NVPTXISD::StoreV2: |
| 2435 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2436 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2437 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2438 | case MVT::i8: |
| 2439 | Opcode = NVPTX::STV_i8_v2_ari_64; |
| 2440 | break; |
| 2441 | case MVT::i16: |
| 2442 | Opcode = NVPTX::STV_i16_v2_ari_64; |
| 2443 | break; |
| 2444 | case MVT::i32: |
| 2445 | Opcode = NVPTX::STV_i32_v2_ari_64; |
| 2446 | break; |
| 2447 | case MVT::i64: |
| 2448 | Opcode = NVPTX::STV_i64_v2_ari_64; |
| 2449 | break; |
| 2450 | case MVT::f32: |
| 2451 | Opcode = NVPTX::STV_f32_v2_ari_64; |
| 2452 | break; |
| 2453 | case MVT::f64: |
| 2454 | Opcode = NVPTX::STV_f64_v2_ari_64; |
| 2455 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2456 | } |
| 2457 | break; |
| 2458 | case NVPTXISD::StoreV4: |
| 2459 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2460 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2461 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2462 | case MVT::i8: |
| 2463 | Opcode = NVPTX::STV_i8_v4_ari_64; |
| 2464 | break; |
| 2465 | case MVT::i16: |
| 2466 | Opcode = NVPTX::STV_i16_v4_ari_64; |
| 2467 | break; |
| 2468 | case MVT::i32: |
| 2469 | Opcode = NVPTX::STV_i32_v4_ari_64; |
| 2470 | break; |
| 2471 | case MVT::f32: |
| 2472 | Opcode = NVPTX::STV_f32_v4_ari_64; |
| 2473 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2474 | } |
| 2475 | break; |
| 2476 | } |
| 2477 | } else { |
| 2478 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2479 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2480 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2481 | case NVPTXISD::StoreV2: |
| 2482 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2483 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2484 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2485 | case MVT::i8: |
| 2486 | Opcode = NVPTX::STV_i8_v2_ari; |
| 2487 | break; |
| 2488 | case MVT::i16: |
| 2489 | Opcode = NVPTX::STV_i16_v2_ari; |
| 2490 | break; |
| 2491 | case MVT::i32: |
| 2492 | Opcode = NVPTX::STV_i32_v2_ari; |
| 2493 | break; |
| 2494 | case MVT::i64: |
| 2495 | Opcode = NVPTX::STV_i64_v2_ari; |
| 2496 | break; |
| 2497 | case MVT::f32: |
| 2498 | Opcode = NVPTX::STV_f32_v2_ari; |
| 2499 | break; |
| 2500 | case MVT::f64: |
| 2501 | Opcode = NVPTX::STV_f64_v2_ari; |
| 2502 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2503 | } |
| 2504 | break; |
| 2505 | case NVPTXISD::StoreV4: |
| 2506 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2507 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2508 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2509 | case MVT::i8: |
| 2510 | Opcode = NVPTX::STV_i8_v4_ari; |
| 2511 | break; |
| 2512 | case MVT::i16: |
| 2513 | Opcode = NVPTX::STV_i16_v4_ari; |
| 2514 | break; |
| 2515 | case MVT::i32: |
| 2516 | Opcode = NVPTX::STV_i32_v4_ari; |
| 2517 | break; |
| 2518 | case MVT::f32: |
| 2519 | Opcode = NVPTX::STV_f32_v4_ari; |
| 2520 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2521 | } |
| 2522 | break; |
| 2523 | } |
| 2524 | } |
| 2525 | StOps.push_back(Base); |
| 2526 | StOps.push_back(Offset); |
| 2527 | } else { |
Eric Christopher | 02389e3 | 2015-02-19 00:08:27 +0000 | [diff] [blame] | 2528 | if (TM.is64Bit()) { |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2529 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2530 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2531 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2532 | case NVPTXISD::StoreV2: |
| 2533 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2534 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2535 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2536 | case MVT::i8: |
| 2537 | Opcode = NVPTX::STV_i8_v2_areg_64; |
| 2538 | break; |
| 2539 | case MVT::i16: |
| 2540 | Opcode = NVPTX::STV_i16_v2_areg_64; |
| 2541 | break; |
| 2542 | case MVT::i32: |
| 2543 | Opcode = NVPTX::STV_i32_v2_areg_64; |
| 2544 | break; |
| 2545 | case MVT::i64: |
| 2546 | Opcode = NVPTX::STV_i64_v2_areg_64; |
| 2547 | break; |
| 2548 | case MVT::f32: |
| 2549 | Opcode = NVPTX::STV_f32_v2_areg_64; |
| 2550 | break; |
| 2551 | case MVT::f64: |
| 2552 | Opcode = NVPTX::STV_f64_v2_areg_64; |
| 2553 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2554 | } |
| 2555 | break; |
| 2556 | case NVPTXISD::StoreV4: |
| 2557 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2558 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2559 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2560 | case MVT::i8: |
| 2561 | Opcode = NVPTX::STV_i8_v4_areg_64; |
| 2562 | break; |
| 2563 | case MVT::i16: |
| 2564 | Opcode = NVPTX::STV_i16_v4_areg_64; |
| 2565 | break; |
| 2566 | case MVT::i32: |
| 2567 | Opcode = NVPTX::STV_i32_v4_areg_64; |
| 2568 | break; |
| 2569 | case MVT::f32: |
| 2570 | Opcode = NVPTX::STV_f32_v4_areg_64; |
| 2571 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2572 | } |
| 2573 | break; |
| 2574 | } |
| 2575 | } else { |
| 2576 | switch (N->getOpcode()) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2577 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2578 | return nullptr; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2579 | case NVPTXISD::StoreV2: |
| 2580 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2581 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2582 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2583 | case MVT::i8: |
| 2584 | Opcode = NVPTX::STV_i8_v2_areg; |
| 2585 | break; |
| 2586 | case MVT::i16: |
| 2587 | Opcode = NVPTX::STV_i16_v2_areg; |
| 2588 | break; |
| 2589 | case MVT::i32: |
| 2590 | Opcode = NVPTX::STV_i32_v2_areg; |
| 2591 | break; |
| 2592 | case MVT::i64: |
| 2593 | Opcode = NVPTX::STV_i64_v2_areg; |
| 2594 | break; |
| 2595 | case MVT::f32: |
| 2596 | Opcode = NVPTX::STV_f32_v2_areg; |
| 2597 | break; |
| 2598 | case MVT::f64: |
| 2599 | Opcode = NVPTX::STV_f64_v2_areg; |
| 2600 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2601 | } |
| 2602 | break; |
| 2603 | case NVPTXISD::StoreV4: |
| 2604 | switch (EltVT.getSimpleVT().SimpleTy) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2605 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2606 | return nullptr; |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 2607 | case MVT::i8: |
| 2608 | Opcode = NVPTX::STV_i8_v4_areg; |
| 2609 | break; |
| 2610 | case MVT::i16: |
| 2611 | Opcode = NVPTX::STV_i16_v4_areg; |
| 2612 | break; |
| 2613 | case MVT::i32: |
| 2614 | Opcode = NVPTX::STV_i32_v4_areg; |
| 2615 | break; |
| 2616 | case MVT::f32: |
| 2617 | Opcode = NVPTX::STV_f32_v4_areg; |
| 2618 | break; |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2619 | } |
| 2620 | break; |
| 2621 | } |
| 2622 | } |
| 2623 | StOps.push_back(N2); |
| 2624 | } |
| 2625 | |
| 2626 | StOps.push_back(Chain); |
| 2627 | |
Michael Liao | b53d896 | 2013-04-19 22:22:57 +0000 | [diff] [blame] | 2628 | ST = CurDAG->getMachineNode(Opcode, DL, MVT::Other, StOps); |
Justin Holewinski | be8dc64 | 2013-02-12 14:18:49 +0000 | [diff] [blame] | 2629 | |
| 2630 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
| 2631 | MemRefs0[0] = cast<MemSDNode>(N)->getMemOperand(); |
| 2632 | cast<MachineSDNode>(ST)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 2633 | |
| 2634 | return ST; |
| 2635 | } |
| 2636 | |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2637 | SDNode *NVPTXDAGToDAGISel::SelectLoadParam(SDNode *Node) { |
| 2638 | SDValue Chain = Node->getOperand(0); |
| 2639 | SDValue Offset = Node->getOperand(2); |
| 2640 | SDValue Flag = Node->getOperand(3); |
| 2641 | SDLoc DL(Node); |
| 2642 | MemSDNode *Mem = cast<MemSDNode>(Node); |
| 2643 | |
| 2644 | unsigned VecSize; |
| 2645 | switch (Node->getOpcode()) { |
| 2646 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2647 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2648 | case NVPTXISD::LoadParam: |
| 2649 | VecSize = 1; |
| 2650 | break; |
| 2651 | case NVPTXISD::LoadParamV2: |
| 2652 | VecSize = 2; |
| 2653 | break; |
| 2654 | case NVPTXISD::LoadParamV4: |
| 2655 | VecSize = 4; |
| 2656 | break; |
| 2657 | } |
| 2658 | |
| 2659 | EVT EltVT = Node->getValueType(0); |
| 2660 | EVT MemVT = Mem->getMemoryVT(); |
| 2661 | |
| 2662 | unsigned Opc = 0; |
| 2663 | |
| 2664 | switch (VecSize) { |
| 2665 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2666 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2667 | case 1: |
| 2668 | switch (MemVT.getSimpleVT().SimpleTy) { |
| 2669 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2670 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2671 | case MVT::i1: |
| 2672 | Opc = NVPTX::LoadParamMemI8; |
| 2673 | break; |
| 2674 | case MVT::i8: |
| 2675 | Opc = NVPTX::LoadParamMemI8; |
| 2676 | break; |
| 2677 | case MVT::i16: |
| 2678 | Opc = NVPTX::LoadParamMemI16; |
| 2679 | break; |
| 2680 | case MVT::i32: |
| 2681 | Opc = NVPTX::LoadParamMemI32; |
| 2682 | break; |
| 2683 | case MVT::i64: |
| 2684 | Opc = NVPTX::LoadParamMemI64; |
| 2685 | break; |
| 2686 | case MVT::f32: |
| 2687 | Opc = NVPTX::LoadParamMemF32; |
| 2688 | break; |
| 2689 | case MVT::f64: |
| 2690 | Opc = NVPTX::LoadParamMemF64; |
| 2691 | break; |
| 2692 | } |
| 2693 | break; |
| 2694 | case 2: |
| 2695 | switch (MemVT.getSimpleVT().SimpleTy) { |
| 2696 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2697 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2698 | case MVT::i1: |
| 2699 | Opc = NVPTX::LoadParamMemV2I8; |
| 2700 | break; |
| 2701 | case MVT::i8: |
| 2702 | Opc = NVPTX::LoadParamMemV2I8; |
| 2703 | break; |
| 2704 | case MVT::i16: |
| 2705 | Opc = NVPTX::LoadParamMemV2I16; |
| 2706 | break; |
| 2707 | case MVT::i32: |
| 2708 | Opc = NVPTX::LoadParamMemV2I32; |
| 2709 | break; |
| 2710 | case MVT::i64: |
| 2711 | Opc = NVPTX::LoadParamMemV2I64; |
| 2712 | break; |
| 2713 | case MVT::f32: |
| 2714 | Opc = NVPTX::LoadParamMemV2F32; |
| 2715 | break; |
| 2716 | case MVT::f64: |
| 2717 | Opc = NVPTX::LoadParamMemV2F64; |
| 2718 | break; |
| 2719 | } |
| 2720 | break; |
| 2721 | case 4: |
| 2722 | switch (MemVT.getSimpleVT().SimpleTy) { |
| 2723 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2724 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2725 | case MVT::i1: |
| 2726 | Opc = NVPTX::LoadParamMemV4I8; |
| 2727 | break; |
| 2728 | case MVT::i8: |
| 2729 | Opc = NVPTX::LoadParamMemV4I8; |
| 2730 | break; |
| 2731 | case MVT::i16: |
| 2732 | Opc = NVPTX::LoadParamMemV4I16; |
| 2733 | break; |
| 2734 | case MVT::i32: |
| 2735 | Opc = NVPTX::LoadParamMemV4I32; |
| 2736 | break; |
| 2737 | case MVT::f32: |
| 2738 | Opc = NVPTX::LoadParamMemV4F32; |
| 2739 | break; |
| 2740 | } |
| 2741 | break; |
| 2742 | } |
| 2743 | |
| 2744 | SDVTList VTs; |
| 2745 | if (VecSize == 1) { |
| 2746 | VTs = CurDAG->getVTList(EltVT, MVT::Other, MVT::Glue); |
| 2747 | } else if (VecSize == 2) { |
| 2748 | VTs = CurDAG->getVTList(EltVT, EltVT, MVT::Other, MVT::Glue); |
| 2749 | } else { |
| 2750 | EVT EVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other, MVT::Glue }; |
Craig Topper | abb4ac7 | 2014-04-16 06:10:51 +0000 | [diff] [blame] | 2751 | VTs = CurDAG->getVTList(EVTs); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2752 | } |
| 2753 | |
| 2754 | unsigned OffsetVal = cast<ConstantSDNode>(Offset)->getZExtValue(); |
| 2755 | |
| 2756 | SmallVector<SDValue, 2> Ops; |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2757 | Ops.push_back(CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32)); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2758 | Ops.push_back(Chain); |
| 2759 | Ops.push_back(Flag); |
| 2760 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2761 | return CurDAG->getMachineNode(Opc, DL, VTs, Ops); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2762 | } |
| 2763 | |
| 2764 | SDNode *NVPTXDAGToDAGISel::SelectStoreRetval(SDNode *N) { |
| 2765 | SDLoc DL(N); |
| 2766 | SDValue Chain = N->getOperand(0); |
| 2767 | SDValue Offset = N->getOperand(1); |
| 2768 | unsigned OffsetVal = cast<ConstantSDNode>(Offset)->getZExtValue(); |
| 2769 | MemSDNode *Mem = cast<MemSDNode>(N); |
| 2770 | |
| 2771 | // How many elements do we have? |
| 2772 | unsigned NumElts = 1; |
| 2773 | switch (N->getOpcode()) { |
| 2774 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2775 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2776 | case NVPTXISD::StoreRetval: |
| 2777 | NumElts = 1; |
| 2778 | break; |
| 2779 | case NVPTXISD::StoreRetvalV2: |
| 2780 | NumElts = 2; |
| 2781 | break; |
| 2782 | case NVPTXISD::StoreRetvalV4: |
| 2783 | NumElts = 4; |
| 2784 | break; |
| 2785 | } |
| 2786 | |
| 2787 | // Build vector of operands |
| 2788 | SmallVector<SDValue, 6> Ops; |
| 2789 | for (unsigned i = 0; i < NumElts; ++i) |
| 2790 | Ops.push_back(N->getOperand(i + 2)); |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2791 | Ops.push_back(CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32)); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2792 | Ops.push_back(Chain); |
| 2793 | |
| 2794 | // Determine target opcode |
| 2795 | // If we have an i1, use an 8-bit store. The lowering code in |
| 2796 | // NVPTXISelLowering will have already emitted an upcast. |
| 2797 | unsigned Opcode = 0; |
| 2798 | switch (NumElts) { |
| 2799 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2800 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2801 | case 1: |
| 2802 | switch (Mem->getMemoryVT().getSimpleVT().SimpleTy) { |
| 2803 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2804 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2805 | case MVT::i1: |
| 2806 | Opcode = NVPTX::StoreRetvalI8; |
| 2807 | break; |
| 2808 | case MVT::i8: |
| 2809 | Opcode = NVPTX::StoreRetvalI8; |
| 2810 | break; |
| 2811 | case MVT::i16: |
| 2812 | Opcode = NVPTX::StoreRetvalI16; |
| 2813 | break; |
| 2814 | case MVT::i32: |
| 2815 | Opcode = NVPTX::StoreRetvalI32; |
| 2816 | break; |
| 2817 | case MVT::i64: |
| 2818 | Opcode = NVPTX::StoreRetvalI64; |
| 2819 | break; |
| 2820 | case MVT::f32: |
| 2821 | Opcode = NVPTX::StoreRetvalF32; |
| 2822 | break; |
| 2823 | case MVT::f64: |
| 2824 | Opcode = NVPTX::StoreRetvalF64; |
| 2825 | break; |
| 2826 | } |
| 2827 | break; |
| 2828 | case 2: |
| 2829 | switch (Mem->getMemoryVT().getSimpleVT().SimpleTy) { |
| 2830 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2831 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2832 | case MVT::i1: |
| 2833 | Opcode = NVPTX::StoreRetvalV2I8; |
| 2834 | break; |
| 2835 | case MVT::i8: |
| 2836 | Opcode = NVPTX::StoreRetvalV2I8; |
| 2837 | break; |
| 2838 | case MVT::i16: |
| 2839 | Opcode = NVPTX::StoreRetvalV2I16; |
| 2840 | break; |
| 2841 | case MVT::i32: |
| 2842 | Opcode = NVPTX::StoreRetvalV2I32; |
| 2843 | break; |
| 2844 | case MVT::i64: |
| 2845 | Opcode = NVPTX::StoreRetvalV2I64; |
| 2846 | break; |
| 2847 | case MVT::f32: |
| 2848 | Opcode = NVPTX::StoreRetvalV2F32; |
| 2849 | break; |
| 2850 | case MVT::f64: |
| 2851 | Opcode = NVPTX::StoreRetvalV2F64; |
| 2852 | break; |
| 2853 | } |
| 2854 | break; |
| 2855 | case 4: |
| 2856 | switch (Mem->getMemoryVT().getSimpleVT().SimpleTy) { |
| 2857 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2858 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2859 | case MVT::i1: |
| 2860 | Opcode = NVPTX::StoreRetvalV4I8; |
| 2861 | break; |
| 2862 | case MVT::i8: |
| 2863 | Opcode = NVPTX::StoreRetvalV4I8; |
| 2864 | break; |
| 2865 | case MVT::i16: |
| 2866 | Opcode = NVPTX::StoreRetvalV4I16; |
| 2867 | break; |
| 2868 | case MVT::i32: |
| 2869 | Opcode = NVPTX::StoreRetvalV4I32; |
| 2870 | break; |
| 2871 | case MVT::f32: |
| 2872 | Opcode = NVPTX::StoreRetvalV4F32; |
| 2873 | break; |
| 2874 | } |
| 2875 | break; |
| 2876 | } |
| 2877 | |
| 2878 | SDNode *Ret = |
| 2879 | CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops); |
| 2880 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
| 2881 | MemRefs0[0] = cast<MemSDNode>(N)->getMemOperand(); |
| 2882 | cast<MachineSDNode>(Ret)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 2883 | |
| 2884 | return Ret; |
| 2885 | } |
| 2886 | |
| 2887 | SDNode *NVPTXDAGToDAGISel::SelectStoreParam(SDNode *N) { |
| 2888 | SDLoc DL(N); |
| 2889 | SDValue Chain = N->getOperand(0); |
| 2890 | SDValue Param = N->getOperand(1); |
| 2891 | unsigned ParamVal = cast<ConstantSDNode>(Param)->getZExtValue(); |
| 2892 | SDValue Offset = N->getOperand(2); |
| 2893 | unsigned OffsetVal = cast<ConstantSDNode>(Offset)->getZExtValue(); |
| 2894 | MemSDNode *Mem = cast<MemSDNode>(N); |
| 2895 | SDValue Flag = N->getOperand(N->getNumOperands() - 1); |
| 2896 | |
| 2897 | // How many elements do we have? |
| 2898 | unsigned NumElts = 1; |
| 2899 | switch (N->getOpcode()) { |
| 2900 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2901 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2902 | case NVPTXISD::StoreParamU32: |
| 2903 | case NVPTXISD::StoreParamS32: |
| 2904 | case NVPTXISD::StoreParam: |
| 2905 | NumElts = 1; |
| 2906 | break; |
| 2907 | case NVPTXISD::StoreParamV2: |
| 2908 | NumElts = 2; |
| 2909 | break; |
| 2910 | case NVPTXISD::StoreParamV4: |
| 2911 | NumElts = 4; |
| 2912 | break; |
| 2913 | } |
| 2914 | |
| 2915 | // Build vector of operands |
| 2916 | SmallVector<SDValue, 8> Ops; |
| 2917 | for (unsigned i = 0; i < NumElts; ++i) |
| 2918 | Ops.push_back(N->getOperand(i + 3)); |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 2919 | Ops.push_back(CurDAG->getTargetConstant(ParamVal, DL, MVT::i32)); |
| 2920 | Ops.push_back(CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32)); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2921 | Ops.push_back(Chain); |
| 2922 | Ops.push_back(Flag); |
| 2923 | |
| 2924 | // Determine target opcode |
| 2925 | // If we have an i1, use an 8-bit store. The lowering code in |
| 2926 | // NVPTXISelLowering will have already emitted an upcast. |
| 2927 | unsigned Opcode = 0; |
| 2928 | switch (N->getOpcode()) { |
| 2929 | default: |
| 2930 | switch (NumElts) { |
| 2931 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2932 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2933 | case 1: |
| 2934 | switch (Mem->getMemoryVT().getSimpleVT().SimpleTy) { |
| 2935 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2936 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2937 | case MVT::i1: |
| 2938 | Opcode = NVPTX::StoreParamI8; |
| 2939 | break; |
| 2940 | case MVT::i8: |
| 2941 | Opcode = NVPTX::StoreParamI8; |
| 2942 | break; |
| 2943 | case MVT::i16: |
| 2944 | Opcode = NVPTX::StoreParamI16; |
| 2945 | break; |
| 2946 | case MVT::i32: |
| 2947 | Opcode = NVPTX::StoreParamI32; |
| 2948 | break; |
| 2949 | case MVT::i64: |
| 2950 | Opcode = NVPTX::StoreParamI64; |
| 2951 | break; |
| 2952 | case MVT::f32: |
| 2953 | Opcode = NVPTX::StoreParamF32; |
| 2954 | break; |
| 2955 | case MVT::f64: |
| 2956 | Opcode = NVPTX::StoreParamF64; |
| 2957 | break; |
| 2958 | } |
| 2959 | break; |
| 2960 | case 2: |
| 2961 | switch (Mem->getMemoryVT().getSimpleVT().SimpleTy) { |
| 2962 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2963 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2964 | case MVT::i1: |
| 2965 | Opcode = NVPTX::StoreParamV2I8; |
| 2966 | break; |
| 2967 | case MVT::i8: |
| 2968 | Opcode = NVPTX::StoreParamV2I8; |
| 2969 | break; |
| 2970 | case MVT::i16: |
| 2971 | Opcode = NVPTX::StoreParamV2I16; |
| 2972 | break; |
| 2973 | case MVT::i32: |
| 2974 | Opcode = NVPTX::StoreParamV2I32; |
| 2975 | break; |
| 2976 | case MVT::i64: |
| 2977 | Opcode = NVPTX::StoreParamV2I64; |
| 2978 | break; |
| 2979 | case MVT::f32: |
| 2980 | Opcode = NVPTX::StoreParamV2F32; |
| 2981 | break; |
| 2982 | case MVT::f64: |
| 2983 | Opcode = NVPTX::StoreParamV2F64; |
| 2984 | break; |
| 2985 | } |
| 2986 | break; |
| 2987 | case 4: |
| 2988 | switch (Mem->getMemoryVT().getSimpleVT().SimpleTy) { |
| 2989 | default: |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 2990 | return nullptr; |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 2991 | case MVT::i1: |
| 2992 | Opcode = NVPTX::StoreParamV4I8; |
| 2993 | break; |
| 2994 | case MVT::i8: |
| 2995 | Opcode = NVPTX::StoreParamV4I8; |
| 2996 | break; |
| 2997 | case MVT::i16: |
| 2998 | Opcode = NVPTX::StoreParamV4I16; |
| 2999 | break; |
| 3000 | case MVT::i32: |
| 3001 | Opcode = NVPTX::StoreParamV4I32; |
| 3002 | break; |
| 3003 | case MVT::f32: |
| 3004 | Opcode = NVPTX::StoreParamV4F32; |
| 3005 | break; |
| 3006 | } |
| 3007 | break; |
| 3008 | } |
| 3009 | break; |
Justin Holewinski | dc5e3b6 | 2013-06-28 17:58:04 +0000 | [diff] [blame] | 3010 | // Special case: if we have a sign-extend/zero-extend node, insert the |
| 3011 | // conversion instruction first, and use that as the value operand to |
| 3012 | // the selected StoreParam node. |
| 3013 | case NVPTXISD::StoreParamU32: { |
| 3014 | Opcode = NVPTX::StoreParamI32; |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 3015 | SDValue CvtNone = CurDAG->getTargetConstant(NVPTX::PTXCvtMode::NONE, DL, |
Justin Holewinski | dc5e3b6 | 2013-06-28 17:58:04 +0000 | [diff] [blame] | 3016 | MVT::i32); |
| 3017 | SDNode *Cvt = CurDAG->getMachineNode(NVPTX::CVT_u32_u16, DL, |
| 3018 | MVT::i32, Ops[0], CvtNone); |
| 3019 | Ops[0] = SDValue(Cvt, 0); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 3020 | break; |
Justin Holewinski | dc5e3b6 | 2013-06-28 17:58:04 +0000 | [diff] [blame] | 3021 | } |
| 3022 | case NVPTXISD::StoreParamS32: { |
| 3023 | Opcode = NVPTX::StoreParamI32; |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 3024 | SDValue CvtNone = CurDAG->getTargetConstant(NVPTX::PTXCvtMode::NONE, DL, |
Justin Holewinski | dc5e3b6 | 2013-06-28 17:58:04 +0000 | [diff] [blame] | 3025 | MVT::i32); |
| 3026 | SDNode *Cvt = CurDAG->getMachineNode(NVPTX::CVT_s32_s16, DL, |
| 3027 | MVT::i32, Ops[0], CvtNone); |
| 3028 | Ops[0] = SDValue(Cvt, 0); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 3029 | break; |
| 3030 | } |
Justin Holewinski | dc5e3b6 | 2013-06-28 17:58:04 +0000 | [diff] [blame] | 3031 | } |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 3032 | |
Justin Holewinski | dff28d2 | 2013-07-01 12:59:01 +0000 | [diff] [blame] | 3033 | SDVTList RetVTs = CurDAG->getVTList(MVT::Other, MVT::Glue); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 3034 | SDNode *Ret = |
Justin Holewinski | dff28d2 | 2013-07-01 12:59:01 +0000 | [diff] [blame] | 3035 | CurDAG->getMachineNode(Opcode, DL, RetVTs, Ops); |
Justin Holewinski | f8f7091 | 2013-06-28 17:57:59 +0000 | [diff] [blame] | 3036 | MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); |
| 3037 | MemRefs0[0] = cast<MemSDNode>(N)->getMemOperand(); |
| 3038 | cast<MachineSDNode>(Ret)->setMemRefs(MemRefs0, MemRefs0 + 1); |
| 3039 | |
| 3040 | return Ret; |
| 3041 | } |
| 3042 | |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3043 | SDNode *NVPTXDAGToDAGISel::SelectTextureIntrinsic(SDNode *N) { |
| 3044 | SDValue Chain = N->getOperand(0); |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 3045 | SDNode *Ret = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3046 | unsigned Opc = 0; |
| 3047 | SmallVector<SDValue, 8> Ops; |
| 3048 | |
| 3049 | switch (N->getOpcode()) { |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 3050 | default: return nullptr; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3051 | case NVPTXISD::Tex1DFloatS32: |
| 3052 | Opc = NVPTX::TEX_1D_F32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3053 | break; |
| 3054 | case NVPTXISD::Tex1DFloatFloat: |
| 3055 | Opc = NVPTX::TEX_1D_F32_F32; |
| 3056 | break; |
| 3057 | case NVPTXISD::Tex1DFloatFloatLevel: |
| 3058 | Opc = NVPTX::TEX_1D_F32_F32_LEVEL; |
| 3059 | break; |
| 3060 | case NVPTXISD::Tex1DFloatFloatGrad: |
| 3061 | Opc = NVPTX::TEX_1D_F32_F32_GRAD; |
| 3062 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3063 | case NVPTXISD::Tex1DS32S32: |
| 3064 | Opc = NVPTX::TEX_1D_S32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3065 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3066 | case NVPTXISD::Tex1DS32Float: |
| 3067 | Opc = NVPTX::TEX_1D_S32_F32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3068 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3069 | case NVPTXISD::Tex1DS32FloatLevel: |
| 3070 | Opc = NVPTX::TEX_1D_S32_F32_LEVEL; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3071 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3072 | case NVPTXISD::Tex1DS32FloatGrad: |
| 3073 | Opc = NVPTX::TEX_1D_S32_F32_GRAD; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3074 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3075 | case NVPTXISD::Tex1DU32S32: |
| 3076 | Opc = NVPTX::TEX_1D_U32_S32; |
| 3077 | break; |
| 3078 | case NVPTXISD::Tex1DU32Float: |
| 3079 | Opc = NVPTX::TEX_1D_U32_F32; |
| 3080 | break; |
| 3081 | case NVPTXISD::Tex1DU32FloatLevel: |
| 3082 | Opc = NVPTX::TEX_1D_U32_F32_LEVEL; |
| 3083 | break; |
| 3084 | case NVPTXISD::Tex1DU32FloatGrad: |
| 3085 | Opc = NVPTX::TEX_1D_U32_F32_GRAD; |
| 3086 | break; |
| 3087 | case NVPTXISD::Tex1DArrayFloatS32: |
| 3088 | Opc = NVPTX::TEX_1D_ARRAY_F32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3089 | break; |
| 3090 | case NVPTXISD::Tex1DArrayFloatFloat: |
| 3091 | Opc = NVPTX::TEX_1D_ARRAY_F32_F32; |
| 3092 | break; |
| 3093 | case NVPTXISD::Tex1DArrayFloatFloatLevel: |
| 3094 | Opc = NVPTX::TEX_1D_ARRAY_F32_F32_LEVEL; |
| 3095 | break; |
| 3096 | case NVPTXISD::Tex1DArrayFloatFloatGrad: |
| 3097 | Opc = NVPTX::TEX_1D_ARRAY_F32_F32_GRAD; |
| 3098 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3099 | case NVPTXISD::Tex1DArrayS32S32: |
| 3100 | Opc = NVPTX::TEX_1D_ARRAY_S32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3101 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3102 | case NVPTXISD::Tex1DArrayS32Float: |
| 3103 | Opc = NVPTX::TEX_1D_ARRAY_S32_F32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3104 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3105 | case NVPTXISD::Tex1DArrayS32FloatLevel: |
| 3106 | Opc = NVPTX::TEX_1D_ARRAY_S32_F32_LEVEL; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3107 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3108 | case NVPTXISD::Tex1DArrayS32FloatGrad: |
| 3109 | Opc = NVPTX::TEX_1D_ARRAY_S32_F32_GRAD; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3110 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3111 | case NVPTXISD::Tex1DArrayU32S32: |
| 3112 | Opc = NVPTX::TEX_1D_ARRAY_U32_S32; |
| 3113 | break; |
| 3114 | case NVPTXISD::Tex1DArrayU32Float: |
| 3115 | Opc = NVPTX::TEX_1D_ARRAY_U32_F32; |
| 3116 | break; |
| 3117 | case NVPTXISD::Tex1DArrayU32FloatLevel: |
| 3118 | Opc = NVPTX::TEX_1D_ARRAY_U32_F32_LEVEL; |
| 3119 | break; |
| 3120 | case NVPTXISD::Tex1DArrayU32FloatGrad: |
| 3121 | Opc = NVPTX::TEX_1D_ARRAY_U32_F32_GRAD; |
| 3122 | break; |
| 3123 | case NVPTXISD::Tex2DFloatS32: |
| 3124 | Opc = NVPTX::TEX_2D_F32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3125 | break; |
| 3126 | case NVPTXISD::Tex2DFloatFloat: |
| 3127 | Opc = NVPTX::TEX_2D_F32_F32; |
| 3128 | break; |
| 3129 | case NVPTXISD::Tex2DFloatFloatLevel: |
| 3130 | Opc = NVPTX::TEX_2D_F32_F32_LEVEL; |
| 3131 | break; |
| 3132 | case NVPTXISD::Tex2DFloatFloatGrad: |
| 3133 | Opc = NVPTX::TEX_2D_F32_F32_GRAD; |
| 3134 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3135 | case NVPTXISD::Tex2DS32S32: |
| 3136 | Opc = NVPTX::TEX_2D_S32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3137 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3138 | case NVPTXISD::Tex2DS32Float: |
| 3139 | Opc = NVPTX::TEX_2D_S32_F32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3140 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3141 | case NVPTXISD::Tex2DS32FloatLevel: |
| 3142 | Opc = NVPTX::TEX_2D_S32_F32_LEVEL; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3143 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3144 | case NVPTXISD::Tex2DS32FloatGrad: |
| 3145 | Opc = NVPTX::TEX_2D_S32_F32_GRAD; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3146 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3147 | case NVPTXISD::Tex2DU32S32: |
| 3148 | Opc = NVPTX::TEX_2D_U32_S32; |
| 3149 | break; |
| 3150 | case NVPTXISD::Tex2DU32Float: |
| 3151 | Opc = NVPTX::TEX_2D_U32_F32; |
| 3152 | break; |
| 3153 | case NVPTXISD::Tex2DU32FloatLevel: |
| 3154 | Opc = NVPTX::TEX_2D_U32_F32_LEVEL; |
| 3155 | break; |
| 3156 | case NVPTXISD::Tex2DU32FloatGrad: |
| 3157 | Opc = NVPTX::TEX_2D_U32_F32_GRAD; |
| 3158 | break; |
| 3159 | case NVPTXISD::Tex2DArrayFloatS32: |
| 3160 | Opc = NVPTX::TEX_2D_ARRAY_F32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3161 | break; |
| 3162 | case NVPTXISD::Tex2DArrayFloatFloat: |
| 3163 | Opc = NVPTX::TEX_2D_ARRAY_F32_F32; |
| 3164 | break; |
| 3165 | case NVPTXISD::Tex2DArrayFloatFloatLevel: |
| 3166 | Opc = NVPTX::TEX_2D_ARRAY_F32_F32_LEVEL; |
| 3167 | break; |
| 3168 | case NVPTXISD::Tex2DArrayFloatFloatGrad: |
| 3169 | Opc = NVPTX::TEX_2D_ARRAY_F32_F32_GRAD; |
| 3170 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3171 | case NVPTXISD::Tex2DArrayS32S32: |
| 3172 | Opc = NVPTX::TEX_2D_ARRAY_S32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3173 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3174 | case NVPTXISD::Tex2DArrayS32Float: |
| 3175 | Opc = NVPTX::TEX_2D_ARRAY_S32_F32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3176 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3177 | case NVPTXISD::Tex2DArrayS32FloatLevel: |
| 3178 | Opc = NVPTX::TEX_2D_ARRAY_S32_F32_LEVEL; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3179 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3180 | case NVPTXISD::Tex2DArrayS32FloatGrad: |
| 3181 | Opc = NVPTX::TEX_2D_ARRAY_S32_F32_GRAD; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3182 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3183 | case NVPTXISD::Tex2DArrayU32S32: |
| 3184 | Opc = NVPTX::TEX_2D_ARRAY_U32_S32; |
| 3185 | break; |
| 3186 | case NVPTXISD::Tex2DArrayU32Float: |
| 3187 | Opc = NVPTX::TEX_2D_ARRAY_U32_F32; |
| 3188 | break; |
| 3189 | case NVPTXISD::Tex2DArrayU32FloatLevel: |
| 3190 | Opc = NVPTX::TEX_2D_ARRAY_U32_F32_LEVEL; |
| 3191 | break; |
| 3192 | case NVPTXISD::Tex2DArrayU32FloatGrad: |
| 3193 | Opc = NVPTX::TEX_2D_ARRAY_U32_F32_GRAD; |
| 3194 | break; |
| 3195 | case NVPTXISD::Tex3DFloatS32: |
| 3196 | Opc = NVPTX::TEX_3D_F32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3197 | break; |
| 3198 | case NVPTXISD::Tex3DFloatFloat: |
| 3199 | Opc = NVPTX::TEX_3D_F32_F32; |
| 3200 | break; |
| 3201 | case NVPTXISD::Tex3DFloatFloatLevel: |
| 3202 | Opc = NVPTX::TEX_3D_F32_F32_LEVEL; |
| 3203 | break; |
| 3204 | case NVPTXISD::Tex3DFloatFloatGrad: |
| 3205 | Opc = NVPTX::TEX_3D_F32_F32_GRAD; |
| 3206 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3207 | case NVPTXISD::Tex3DS32S32: |
| 3208 | Opc = NVPTX::TEX_3D_S32_S32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3209 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3210 | case NVPTXISD::Tex3DS32Float: |
| 3211 | Opc = NVPTX::TEX_3D_S32_F32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3212 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3213 | case NVPTXISD::Tex3DS32FloatLevel: |
| 3214 | Opc = NVPTX::TEX_3D_S32_F32_LEVEL; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3215 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3216 | case NVPTXISD::Tex3DS32FloatGrad: |
| 3217 | Opc = NVPTX::TEX_3D_S32_F32_GRAD; |
| 3218 | break; |
| 3219 | case NVPTXISD::Tex3DU32S32: |
| 3220 | Opc = NVPTX::TEX_3D_U32_S32; |
| 3221 | break; |
| 3222 | case NVPTXISD::Tex3DU32Float: |
| 3223 | Opc = NVPTX::TEX_3D_U32_F32; |
| 3224 | break; |
| 3225 | case NVPTXISD::Tex3DU32FloatLevel: |
| 3226 | Opc = NVPTX::TEX_3D_U32_F32_LEVEL; |
| 3227 | break; |
| 3228 | case NVPTXISD::Tex3DU32FloatGrad: |
| 3229 | Opc = NVPTX::TEX_3D_U32_F32_GRAD; |
| 3230 | break; |
| 3231 | case NVPTXISD::TexCubeFloatFloat: |
| 3232 | Opc = NVPTX::TEX_CUBE_F32_F32; |
| 3233 | break; |
| 3234 | case NVPTXISD::TexCubeFloatFloatLevel: |
| 3235 | Opc = NVPTX::TEX_CUBE_F32_F32_LEVEL; |
| 3236 | break; |
| 3237 | case NVPTXISD::TexCubeS32Float: |
| 3238 | Opc = NVPTX::TEX_CUBE_S32_F32; |
| 3239 | break; |
| 3240 | case NVPTXISD::TexCubeS32FloatLevel: |
| 3241 | Opc = NVPTX::TEX_CUBE_S32_F32_LEVEL; |
| 3242 | break; |
| 3243 | case NVPTXISD::TexCubeU32Float: |
| 3244 | Opc = NVPTX::TEX_CUBE_U32_F32; |
| 3245 | break; |
| 3246 | case NVPTXISD::TexCubeU32FloatLevel: |
| 3247 | Opc = NVPTX::TEX_CUBE_U32_F32_LEVEL; |
| 3248 | break; |
| 3249 | case NVPTXISD::TexCubeArrayFloatFloat: |
| 3250 | Opc = NVPTX::TEX_CUBE_ARRAY_F32_F32; |
| 3251 | break; |
| 3252 | case NVPTXISD::TexCubeArrayFloatFloatLevel: |
| 3253 | Opc = NVPTX::TEX_CUBE_ARRAY_F32_F32_LEVEL; |
| 3254 | break; |
| 3255 | case NVPTXISD::TexCubeArrayS32Float: |
| 3256 | Opc = NVPTX::TEX_CUBE_ARRAY_S32_F32; |
| 3257 | break; |
| 3258 | case NVPTXISD::TexCubeArrayS32FloatLevel: |
| 3259 | Opc = NVPTX::TEX_CUBE_ARRAY_S32_F32_LEVEL; |
| 3260 | break; |
| 3261 | case NVPTXISD::TexCubeArrayU32Float: |
| 3262 | Opc = NVPTX::TEX_CUBE_ARRAY_U32_F32; |
| 3263 | break; |
| 3264 | case NVPTXISD::TexCubeArrayU32FloatLevel: |
| 3265 | Opc = NVPTX::TEX_CUBE_ARRAY_U32_F32_LEVEL; |
| 3266 | break; |
| 3267 | case NVPTXISD::Tld4R2DFloatFloat: |
| 3268 | Opc = NVPTX::TLD4_R_2D_F32_F32; |
| 3269 | break; |
| 3270 | case NVPTXISD::Tld4G2DFloatFloat: |
| 3271 | Opc = NVPTX::TLD4_G_2D_F32_F32; |
| 3272 | break; |
| 3273 | case NVPTXISD::Tld4B2DFloatFloat: |
| 3274 | Opc = NVPTX::TLD4_B_2D_F32_F32; |
| 3275 | break; |
| 3276 | case NVPTXISD::Tld4A2DFloatFloat: |
| 3277 | Opc = NVPTX::TLD4_A_2D_F32_F32; |
| 3278 | break; |
| 3279 | case NVPTXISD::Tld4R2DS64Float: |
| 3280 | Opc = NVPTX::TLD4_R_2D_S32_F32; |
| 3281 | break; |
| 3282 | case NVPTXISD::Tld4G2DS64Float: |
| 3283 | Opc = NVPTX::TLD4_G_2D_S32_F32; |
| 3284 | break; |
| 3285 | case NVPTXISD::Tld4B2DS64Float: |
| 3286 | Opc = NVPTX::TLD4_B_2D_S32_F32; |
| 3287 | break; |
| 3288 | case NVPTXISD::Tld4A2DS64Float: |
| 3289 | Opc = NVPTX::TLD4_A_2D_S32_F32; |
| 3290 | break; |
| 3291 | case NVPTXISD::Tld4R2DU64Float: |
| 3292 | Opc = NVPTX::TLD4_R_2D_U32_F32; |
| 3293 | break; |
| 3294 | case NVPTXISD::Tld4G2DU64Float: |
| 3295 | Opc = NVPTX::TLD4_G_2D_U32_F32; |
| 3296 | break; |
| 3297 | case NVPTXISD::Tld4B2DU64Float: |
| 3298 | Opc = NVPTX::TLD4_B_2D_U32_F32; |
| 3299 | break; |
| 3300 | case NVPTXISD::Tld4A2DU64Float: |
| 3301 | Opc = NVPTX::TLD4_A_2D_U32_F32; |
| 3302 | break; |
| 3303 | case NVPTXISD::TexUnified1DFloatS32: |
| 3304 | Opc = NVPTX::TEX_UNIFIED_1D_F32_S32; |
| 3305 | break; |
| 3306 | case NVPTXISD::TexUnified1DFloatFloat: |
| 3307 | Opc = NVPTX::TEX_UNIFIED_1D_F32_F32; |
| 3308 | break; |
| 3309 | case NVPTXISD::TexUnified1DFloatFloatLevel: |
| 3310 | Opc = NVPTX::TEX_UNIFIED_1D_F32_F32_LEVEL; |
| 3311 | break; |
| 3312 | case NVPTXISD::TexUnified1DFloatFloatGrad: |
| 3313 | Opc = NVPTX::TEX_UNIFIED_1D_F32_F32_GRAD; |
| 3314 | break; |
| 3315 | case NVPTXISD::TexUnified1DS32S32: |
| 3316 | Opc = NVPTX::TEX_UNIFIED_1D_S32_S32; |
| 3317 | break; |
| 3318 | case NVPTXISD::TexUnified1DS32Float: |
| 3319 | Opc = NVPTX::TEX_UNIFIED_1D_S32_F32; |
| 3320 | break; |
| 3321 | case NVPTXISD::TexUnified1DS32FloatLevel: |
| 3322 | Opc = NVPTX::TEX_UNIFIED_1D_S32_F32_LEVEL; |
| 3323 | break; |
| 3324 | case NVPTXISD::TexUnified1DS32FloatGrad: |
| 3325 | Opc = NVPTX::TEX_UNIFIED_1D_S32_F32_GRAD; |
| 3326 | break; |
| 3327 | case NVPTXISD::TexUnified1DU32S32: |
| 3328 | Opc = NVPTX::TEX_UNIFIED_1D_U32_S32; |
| 3329 | break; |
| 3330 | case NVPTXISD::TexUnified1DU32Float: |
| 3331 | Opc = NVPTX::TEX_UNIFIED_1D_U32_F32; |
| 3332 | break; |
| 3333 | case NVPTXISD::TexUnified1DU32FloatLevel: |
| 3334 | Opc = NVPTX::TEX_UNIFIED_1D_U32_F32_LEVEL; |
| 3335 | break; |
| 3336 | case NVPTXISD::TexUnified1DU32FloatGrad: |
| 3337 | Opc = NVPTX::TEX_UNIFIED_1D_U32_F32_GRAD; |
| 3338 | break; |
| 3339 | case NVPTXISD::TexUnified1DArrayFloatS32: |
| 3340 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_S32; |
| 3341 | break; |
| 3342 | case NVPTXISD::TexUnified1DArrayFloatFloat: |
| 3343 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_F32; |
| 3344 | break; |
| 3345 | case NVPTXISD::TexUnified1DArrayFloatFloatLevel: |
| 3346 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_F32_LEVEL; |
| 3347 | break; |
| 3348 | case NVPTXISD::TexUnified1DArrayFloatFloatGrad: |
| 3349 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_F32_GRAD; |
| 3350 | break; |
| 3351 | case NVPTXISD::TexUnified1DArrayS32S32: |
| 3352 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_S32; |
| 3353 | break; |
| 3354 | case NVPTXISD::TexUnified1DArrayS32Float: |
| 3355 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_F32; |
| 3356 | break; |
| 3357 | case NVPTXISD::TexUnified1DArrayS32FloatLevel: |
| 3358 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_F32_LEVEL; |
| 3359 | break; |
| 3360 | case NVPTXISD::TexUnified1DArrayS32FloatGrad: |
| 3361 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_F32_GRAD; |
| 3362 | break; |
| 3363 | case NVPTXISD::TexUnified1DArrayU32S32: |
| 3364 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_S32; |
| 3365 | break; |
| 3366 | case NVPTXISD::TexUnified1DArrayU32Float: |
| 3367 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_F32; |
| 3368 | break; |
| 3369 | case NVPTXISD::TexUnified1DArrayU32FloatLevel: |
| 3370 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_F32_LEVEL; |
| 3371 | break; |
| 3372 | case NVPTXISD::TexUnified1DArrayU32FloatGrad: |
| 3373 | Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_F32_GRAD; |
| 3374 | break; |
| 3375 | case NVPTXISD::TexUnified2DFloatS32: |
| 3376 | Opc = NVPTX::TEX_UNIFIED_2D_F32_S32; |
| 3377 | break; |
| 3378 | case NVPTXISD::TexUnified2DFloatFloat: |
| 3379 | Opc = NVPTX::TEX_UNIFIED_2D_F32_F32; |
| 3380 | break; |
| 3381 | case NVPTXISD::TexUnified2DFloatFloatLevel: |
| 3382 | Opc = NVPTX::TEX_UNIFIED_2D_F32_F32_LEVEL; |
| 3383 | break; |
| 3384 | case NVPTXISD::TexUnified2DFloatFloatGrad: |
| 3385 | Opc = NVPTX::TEX_UNIFIED_2D_F32_F32_GRAD; |
| 3386 | break; |
| 3387 | case NVPTXISD::TexUnified2DS32S32: |
| 3388 | Opc = NVPTX::TEX_UNIFIED_2D_S32_S32; |
| 3389 | break; |
| 3390 | case NVPTXISD::TexUnified2DS32Float: |
| 3391 | Opc = NVPTX::TEX_UNIFIED_2D_S32_F32; |
| 3392 | break; |
| 3393 | case NVPTXISD::TexUnified2DS32FloatLevel: |
| 3394 | Opc = NVPTX::TEX_UNIFIED_2D_S32_F32_LEVEL; |
| 3395 | break; |
| 3396 | case NVPTXISD::TexUnified2DS32FloatGrad: |
| 3397 | Opc = NVPTX::TEX_UNIFIED_2D_S32_F32_GRAD; |
| 3398 | break; |
| 3399 | case NVPTXISD::TexUnified2DU32S32: |
| 3400 | Opc = NVPTX::TEX_UNIFIED_2D_U32_S32; |
| 3401 | break; |
| 3402 | case NVPTXISD::TexUnified2DU32Float: |
| 3403 | Opc = NVPTX::TEX_UNIFIED_2D_U32_F32; |
| 3404 | break; |
| 3405 | case NVPTXISD::TexUnified2DU32FloatLevel: |
| 3406 | Opc = NVPTX::TEX_UNIFIED_2D_U32_F32_LEVEL; |
| 3407 | break; |
| 3408 | case NVPTXISD::TexUnified2DU32FloatGrad: |
| 3409 | Opc = NVPTX::TEX_UNIFIED_2D_U32_F32_GRAD; |
| 3410 | break; |
| 3411 | case NVPTXISD::TexUnified2DArrayFloatS32: |
| 3412 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_S32; |
| 3413 | break; |
| 3414 | case NVPTXISD::TexUnified2DArrayFloatFloat: |
| 3415 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_F32; |
| 3416 | break; |
| 3417 | case NVPTXISD::TexUnified2DArrayFloatFloatLevel: |
| 3418 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_F32_LEVEL; |
| 3419 | break; |
| 3420 | case NVPTXISD::TexUnified2DArrayFloatFloatGrad: |
| 3421 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_F32_GRAD; |
| 3422 | break; |
| 3423 | case NVPTXISD::TexUnified2DArrayS32S32: |
| 3424 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_S32; |
| 3425 | break; |
| 3426 | case NVPTXISD::TexUnified2DArrayS32Float: |
| 3427 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_F32; |
| 3428 | break; |
| 3429 | case NVPTXISD::TexUnified2DArrayS32FloatLevel: |
| 3430 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_F32_LEVEL; |
| 3431 | break; |
| 3432 | case NVPTXISD::TexUnified2DArrayS32FloatGrad: |
| 3433 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_F32_GRAD; |
| 3434 | break; |
| 3435 | case NVPTXISD::TexUnified2DArrayU32S32: |
| 3436 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_S32; |
| 3437 | break; |
| 3438 | case NVPTXISD::TexUnified2DArrayU32Float: |
| 3439 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_F32; |
| 3440 | break; |
| 3441 | case NVPTXISD::TexUnified2DArrayU32FloatLevel: |
| 3442 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_F32_LEVEL; |
| 3443 | break; |
| 3444 | case NVPTXISD::TexUnified2DArrayU32FloatGrad: |
| 3445 | Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_F32_GRAD; |
| 3446 | break; |
| 3447 | case NVPTXISD::TexUnified3DFloatS32: |
| 3448 | Opc = NVPTX::TEX_UNIFIED_3D_F32_S32; |
| 3449 | break; |
| 3450 | case NVPTXISD::TexUnified3DFloatFloat: |
| 3451 | Opc = NVPTX::TEX_UNIFIED_3D_F32_F32; |
| 3452 | break; |
| 3453 | case NVPTXISD::TexUnified3DFloatFloatLevel: |
| 3454 | Opc = NVPTX::TEX_UNIFIED_3D_F32_F32_LEVEL; |
| 3455 | break; |
| 3456 | case NVPTXISD::TexUnified3DFloatFloatGrad: |
| 3457 | Opc = NVPTX::TEX_UNIFIED_3D_F32_F32_GRAD; |
| 3458 | break; |
| 3459 | case NVPTXISD::TexUnified3DS32S32: |
| 3460 | Opc = NVPTX::TEX_UNIFIED_3D_S32_S32; |
| 3461 | break; |
| 3462 | case NVPTXISD::TexUnified3DS32Float: |
| 3463 | Opc = NVPTX::TEX_UNIFIED_3D_S32_F32; |
| 3464 | break; |
| 3465 | case NVPTXISD::TexUnified3DS32FloatLevel: |
| 3466 | Opc = NVPTX::TEX_UNIFIED_3D_S32_F32_LEVEL; |
| 3467 | break; |
| 3468 | case NVPTXISD::TexUnified3DS32FloatGrad: |
| 3469 | Opc = NVPTX::TEX_UNIFIED_3D_S32_F32_GRAD; |
| 3470 | break; |
| 3471 | case NVPTXISD::TexUnified3DU32S32: |
| 3472 | Opc = NVPTX::TEX_UNIFIED_3D_U32_S32; |
| 3473 | break; |
| 3474 | case NVPTXISD::TexUnified3DU32Float: |
| 3475 | Opc = NVPTX::TEX_UNIFIED_3D_U32_F32; |
| 3476 | break; |
| 3477 | case NVPTXISD::TexUnified3DU32FloatLevel: |
| 3478 | Opc = NVPTX::TEX_UNIFIED_3D_U32_F32_LEVEL; |
| 3479 | break; |
| 3480 | case NVPTXISD::TexUnified3DU32FloatGrad: |
| 3481 | Opc = NVPTX::TEX_UNIFIED_3D_U32_F32_GRAD; |
| 3482 | break; |
| 3483 | case NVPTXISD::TexUnifiedCubeFloatFloat: |
| 3484 | Opc = NVPTX::TEX_UNIFIED_CUBE_F32_F32; |
| 3485 | break; |
| 3486 | case NVPTXISD::TexUnifiedCubeFloatFloatLevel: |
| 3487 | Opc = NVPTX::TEX_UNIFIED_CUBE_F32_F32_LEVEL; |
| 3488 | break; |
| 3489 | case NVPTXISD::TexUnifiedCubeS32Float: |
| 3490 | Opc = NVPTX::TEX_UNIFIED_CUBE_S32_F32; |
| 3491 | break; |
| 3492 | case NVPTXISD::TexUnifiedCubeS32FloatLevel: |
| 3493 | Opc = NVPTX::TEX_UNIFIED_CUBE_S32_F32_LEVEL; |
| 3494 | break; |
| 3495 | case NVPTXISD::TexUnifiedCubeU32Float: |
| 3496 | Opc = NVPTX::TEX_UNIFIED_CUBE_U32_F32; |
| 3497 | break; |
| 3498 | case NVPTXISD::TexUnifiedCubeU32FloatLevel: |
| 3499 | Opc = NVPTX::TEX_UNIFIED_CUBE_U32_F32_LEVEL; |
| 3500 | break; |
| 3501 | case NVPTXISD::TexUnifiedCubeArrayFloatFloat: |
| 3502 | Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_F32_F32; |
| 3503 | break; |
| 3504 | case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel: |
| 3505 | Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_F32_F32_LEVEL; |
| 3506 | break; |
| 3507 | case NVPTXISD::TexUnifiedCubeArrayS32Float: |
| 3508 | Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_S32_F32; |
| 3509 | break; |
| 3510 | case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel: |
| 3511 | Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_S32_F32_LEVEL; |
| 3512 | break; |
| 3513 | case NVPTXISD::TexUnifiedCubeArrayU32Float: |
| 3514 | Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_U32_F32; |
| 3515 | break; |
| 3516 | case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel: |
| 3517 | Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_U32_F32_LEVEL; |
| 3518 | break; |
| 3519 | case NVPTXISD::Tld4UnifiedR2DFloatFloat: |
| 3520 | Opc = NVPTX::TLD4_UNIFIED_R_2D_F32_F32; |
| 3521 | break; |
| 3522 | case NVPTXISD::Tld4UnifiedG2DFloatFloat: |
| 3523 | Opc = NVPTX::TLD4_UNIFIED_G_2D_F32_F32; |
| 3524 | break; |
| 3525 | case NVPTXISD::Tld4UnifiedB2DFloatFloat: |
| 3526 | Opc = NVPTX::TLD4_UNIFIED_B_2D_F32_F32; |
| 3527 | break; |
| 3528 | case NVPTXISD::Tld4UnifiedA2DFloatFloat: |
| 3529 | Opc = NVPTX::TLD4_UNIFIED_A_2D_F32_F32; |
| 3530 | break; |
| 3531 | case NVPTXISD::Tld4UnifiedR2DS64Float: |
| 3532 | Opc = NVPTX::TLD4_UNIFIED_R_2D_S32_F32; |
| 3533 | break; |
| 3534 | case NVPTXISD::Tld4UnifiedG2DS64Float: |
| 3535 | Opc = NVPTX::TLD4_UNIFIED_G_2D_S32_F32; |
| 3536 | break; |
| 3537 | case NVPTXISD::Tld4UnifiedB2DS64Float: |
| 3538 | Opc = NVPTX::TLD4_UNIFIED_B_2D_S32_F32; |
| 3539 | break; |
| 3540 | case NVPTXISD::Tld4UnifiedA2DS64Float: |
| 3541 | Opc = NVPTX::TLD4_UNIFIED_A_2D_S32_F32; |
| 3542 | break; |
| 3543 | case NVPTXISD::Tld4UnifiedR2DU64Float: |
| 3544 | Opc = NVPTX::TLD4_UNIFIED_R_2D_U32_F32; |
| 3545 | break; |
| 3546 | case NVPTXISD::Tld4UnifiedG2DU64Float: |
| 3547 | Opc = NVPTX::TLD4_UNIFIED_G_2D_U32_F32; |
| 3548 | break; |
| 3549 | case NVPTXISD::Tld4UnifiedB2DU64Float: |
| 3550 | Opc = NVPTX::TLD4_UNIFIED_B_2D_U32_F32; |
| 3551 | break; |
| 3552 | case NVPTXISD::Tld4UnifiedA2DU64Float: |
| 3553 | Opc = NVPTX::TLD4_UNIFIED_A_2D_U32_F32; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3554 | break; |
| 3555 | } |
| 3556 | |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3557 | // Copy over operands |
| 3558 | for (unsigned i = 1; i < N->getNumOperands(); ++i) { |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3559 | Ops.push_back(N->getOperand(i)); |
| 3560 | } |
| 3561 | |
| 3562 | Ops.push_back(Chain); |
| 3563 | Ret = CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops); |
| 3564 | return Ret; |
| 3565 | } |
| 3566 | |
| 3567 | SDNode *NVPTXDAGToDAGISel::SelectSurfaceIntrinsic(SDNode *N) { |
| 3568 | SDValue Chain = N->getOperand(0); |
| 3569 | SDValue TexHandle = N->getOperand(1); |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 3570 | SDNode *Ret = nullptr; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3571 | unsigned Opc = 0; |
| 3572 | SmallVector<SDValue, 8> Ops; |
| 3573 | switch (N->getOpcode()) { |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 3574 | default: return nullptr; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3575 | case NVPTXISD::Suld1DI8Clamp: |
| 3576 | Opc = NVPTX::SULD_1D_I8_CLAMP; |
| 3577 | Ops.push_back(TexHandle); |
| 3578 | Ops.push_back(N->getOperand(2)); |
| 3579 | Ops.push_back(Chain); |
| 3580 | break; |
| 3581 | case NVPTXISD::Suld1DI16Clamp: |
| 3582 | Opc = NVPTX::SULD_1D_I16_CLAMP; |
| 3583 | Ops.push_back(TexHandle); |
| 3584 | Ops.push_back(N->getOperand(2)); |
| 3585 | Ops.push_back(Chain); |
| 3586 | break; |
| 3587 | case NVPTXISD::Suld1DI32Clamp: |
| 3588 | Opc = NVPTX::SULD_1D_I32_CLAMP; |
| 3589 | Ops.push_back(TexHandle); |
| 3590 | Ops.push_back(N->getOperand(2)); |
| 3591 | Ops.push_back(Chain); |
| 3592 | break; |
| 3593 | case NVPTXISD::Suld1DI64Clamp: |
| 3594 | Opc = NVPTX::SULD_1D_I64_CLAMP; |
| 3595 | Ops.push_back(TexHandle); |
| 3596 | Ops.push_back(N->getOperand(2)); |
| 3597 | Ops.push_back(Chain); |
| 3598 | break; |
| 3599 | case NVPTXISD::Suld1DV2I8Clamp: |
| 3600 | Opc = NVPTX::SULD_1D_V2I8_CLAMP; |
| 3601 | Ops.push_back(TexHandle); |
| 3602 | Ops.push_back(N->getOperand(2)); |
| 3603 | Ops.push_back(Chain); |
| 3604 | break; |
| 3605 | case NVPTXISD::Suld1DV2I16Clamp: |
| 3606 | Opc = NVPTX::SULD_1D_V2I16_CLAMP; |
| 3607 | Ops.push_back(TexHandle); |
| 3608 | Ops.push_back(N->getOperand(2)); |
| 3609 | Ops.push_back(Chain); |
| 3610 | break; |
| 3611 | case NVPTXISD::Suld1DV2I32Clamp: |
| 3612 | Opc = NVPTX::SULD_1D_V2I32_CLAMP; |
| 3613 | Ops.push_back(TexHandle); |
| 3614 | Ops.push_back(N->getOperand(2)); |
| 3615 | Ops.push_back(Chain); |
| 3616 | break; |
| 3617 | case NVPTXISD::Suld1DV2I64Clamp: |
| 3618 | Opc = NVPTX::SULD_1D_V2I64_CLAMP; |
| 3619 | Ops.push_back(TexHandle); |
| 3620 | Ops.push_back(N->getOperand(2)); |
| 3621 | Ops.push_back(Chain); |
| 3622 | break; |
| 3623 | case NVPTXISD::Suld1DV4I8Clamp: |
| 3624 | Opc = NVPTX::SULD_1D_V4I8_CLAMP; |
| 3625 | Ops.push_back(TexHandle); |
| 3626 | Ops.push_back(N->getOperand(2)); |
| 3627 | Ops.push_back(Chain); |
| 3628 | break; |
| 3629 | case NVPTXISD::Suld1DV4I16Clamp: |
| 3630 | Opc = NVPTX::SULD_1D_V4I16_CLAMP; |
| 3631 | Ops.push_back(TexHandle); |
| 3632 | Ops.push_back(N->getOperand(2)); |
| 3633 | Ops.push_back(Chain); |
| 3634 | break; |
| 3635 | case NVPTXISD::Suld1DV4I32Clamp: |
| 3636 | Opc = NVPTX::SULD_1D_V4I32_CLAMP; |
| 3637 | Ops.push_back(TexHandle); |
| 3638 | Ops.push_back(N->getOperand(2)); |
| 3639 | Ops.push_back(Chain); |
| 3640 | break; |
| 3641 | case NVPTXISD::Suld1DArrayI8Clamp: |
| 3642 | Opc = NVPTX::SULD_1D_ARRAY_I8_CLAMP; |
| 3643 | Ops.push_back(TexHandle); |
| 3644 | Ops.push_back(N->getOperand(2)); |
| 3645 | Ops.push_back(N->getOperand(3)); |
| 3646 | Ops.push_back(Chain); |
| 3647 | break; |
| 3648 | case NVPTXISD::Suld1DArrayI16Clamp: |
| 3649 | Opc = NVPTX::SULD_1D_ARRAY_I16_CLAMP; |
| 3650 | Ops.push_back(TexHandle); |
| 3651 | Ops.push_back(N->getOperand(2)); |
| 3652 | Ops.push_back(N->getOperand(3)); |
| 3653 | Ops.push_back(Chain); |
| 3654 | break; |
| 3655 | case NVPTXISD::Suld1DArrayI32Clamp: |
| 3656 | Opc = NVPTX::SULD_1D_ARRAY_I32_CLAMP; |
| 3657 | Ops.push_back(TexHandle); |
| 3658 | Ops.push_back(N->getOperand(2)); |
| 3659 | Ops.push_back(N->getOperand(3)); |
| 3660 | Ops.push_back(Chain); |
| 3661 | break; |
| 3662 | case NVPTXISD::Suld1DArrayI64Clamp: |
| 3663 | Opc = NVPTX::SULD_1D_ARRAY_I64_CLAMP; |
| 3664 | Ops.push_back(TexHandle); |
| 3665 | Ops.push_back(N->getOperand(2)); |
| 3666 | Ops.push_back(N->getOperand(3)); |
| 3667 | Ops.push_back(Chain); |
| 3668 | break; |
| 3669 | case NVPTXISD::Suld1DArrayV2I8Clamp: |
| 3670 | Opc = NVPTX::SULD_1D_ARRAY_V2I8_CLAMP; |
| 3671 | Ops.push_back(TexHandle); |
| 3672 | Ops.push_back(N->getOperand(2)); |
| 3673 | Ops.push_back(N->getOperand(3)); |
| 3674 | Ops.push_back(Chain); |
| 3675 | break; |
| 3676 | case NVPTXISD::Suld1DArrayV2I16Clamp: |
| 3677 | Opc = NVPTX::SULD_1D_ARRAY_V2I16_CLAMP; |
| 3678 | Ops.push_back(TexHandle); |
| 3679 | Ops.push_back(N->getOperand(2)); |
| 3680 | Ops.push_back(N->getOperand(3)); |
| 3681 | Ops.push_back(Chain); |
| 3682 | break; |
| 3683 | case NVPTXISD::Suld1DArrayV2I32Clamp: |
| 3684 | Opc = NVPTX::SULD_1D_ARRAY_V2I32_CLAMP; |
| 3685 | Ops.push_back(TexHandle); |
| 3686 | Ops.push_back(N->getOperand(2)); |
| 3687 | Ops.push_back(N->getOperand(3)); |
| 3688 | Ops.push_back(Chain); |
| 3689 | break; |
| 3690 | case NVPTXISD::Suld1DArrayV2I64Clamp: |
| 3691 | Opc = NVPTX::SULD_1D_ARRAY_V2I64_CLAMP; |
| 3692 | Ops.push_back(TexHandle); |
| 3693 | Ops.push_back(N->getOperand(2)); |
| 3694 | Ops.push_back(N->getOperand(3)); |
| 3695 | Ops.push_back(Chain); |
| 3696 | break; |
| 3697 | case NVPTXISD::Suld1DArrayV4I8Clamp: |
| 3698 | Opc = NVPTX::SULD_1D_ARRAY_V4I8_CLAMP; |
| 3699 | Ops.push_back(TexHandle); |
| 3700 | Ops.push_back(N->getOperand(2)); |
| 3701 | Ops.push_back(N->getOperand(3)); |
| 3702 | Ops.push_back(Chain); |
| 3703 | break; |
| 3704 | case NVPTXISD::Suld1DArrayV4I16Clamp: |
| 3705 | Opc = NVPTX::SULD_1D_ARRAY_V4I16_CLAMP; |
| 3706 | Ops.push_back(TexHandle); |
| 3707 | Ops.push_back(N->getOperand(2)); |
| 3708 | Ops.push_back(N->getOperand(3)); |
| 3709 | Ops.push_back(Chain); |
| 3710 | break; |
| 3711 | case NVPTXISD::Suld1DArrayV4I32Clamp: |
| 3712 | Opc = NVPTX::SULD_1D_ARRAY_V4I32_CLAMP; |
| 3713 | Ops.push_back(TexHandle); |
| 3714 | Ops.push_back(N->getOperand(2)); |
| 3715 | Ops.push_back(N->getOperand(3)); |
| 3716 | Ops.push_back(Chain); |
| 3717 | break; |
| 3718 | case NVPTXISD::Suld2DI8Clamp: |
| 3719 | Opc = NVPTX::SULD_2D_I8_CLAMP; |
| 3720 | Ops.push_back(TexHandle); |
| 3721 | Ops.push_back(N->getOperand(2)); |
| 3722 | Ops.push_back(N->getOperand(3)); |
| 3723 | Ops.push_back(Chain); |
| 3724 | break; |
| 3725 | case NVPTXISD::Suld2DI16Clamp: |
| 3726 | Opc = NVPTX::SULD_2D_I16_CLAMP; |
| 3727 | Ops.push_back(TexHandle); |
| 3728 | Ops.push_back(N->getOperand(2)); |
| 3729 | Ops.push_back(N->getOperand(3)); |
| 3730 | Ops.push_back(Chain); |
| 3731 | break; |
| 3732 | case NVPTXISD::Suld2DI32Clamp: |
| 3733 | Opc = NVPTX::SULD_2D_I32_CLAMP; |
| 3734 | Ops.push_back(TexHandle); |
| 3735 | Ops.push_back(N->getOperand(2)); |
| 3736 | Ops.push_back(N->getOperand(3)); |
| 3737 | Ops.push_back(Chain); |
| 3738 | break; |
| 3739 | case NVPTXISD::Suld2DI64Clamp: |
| 3740 | Opc = NVPTX::SULD_2D_I64_CLAMP; |
| 3741 | Ops.push_back(TexHandle); |
| 3742 | Ops.push_back(N->getOperand(2)); |
| 3743 | Ops.push_back(N->getOperand(3)); |
| 3744 | Ops.push_back(Chain); |
| 3745 | break; |
| 3746 | case NVPTXISD::Suld2DV2I8Clamp: |
| 3747 | Opc = NVPTX::SULD_2D_V2I8_CLAMP; |
| 3748 | Ops.push_back(TexHandle); |
| 3749 | Ops.push_back(N->getOperand(2)); |
| 3750 | Ops.push_back(N->getOperand(3)); |
| 3751 | Ops.push_back(Chain); |
| 3752 | break; |
| 3753 | case NVPTXISD::Suld2DV2I16Clamp: |
| 3754 | Opc = NVPTX::SULD_2D_V2I16_CLAMP; |
| 3755 | Ops.push_back(TexHandle); |
| 3756 | Ops.push_back(N->getOperand(2)); |
| 3757 | Ops.push_back(N->getOperand(3)); |
| 3758 | Ops.push_back(Chain); |
| 3759 | break; |
| 3760 | case NVPTXISD::Suld2DV2I32Clamp: |
| 3761 | Opc = NVPTX::SULD_2D_V2I32_CLAMP; |
| 3762 | Ops.push_back(TexHandle); |
| 3763 | Ops.push_back(N->getOperand(2)); |
| 3764 | Ops.push_back(N->getOperand(3)); |
| 3765 | Ops.push_back(Chain); |
| 3766 | break; |
| 3767 | case NVPTXISD::Suld2DV2I64Clamp: |
| 3768 | Opc = NVPTX::SULD_2D_V2I64_CLAMP; |
| 3769 | Ops.push_back(TexHandle); |
| 3770 | Ops.push_back(N->getOperand(2)); |
| 3771 | Ops.push_back(N->getOperand(3)); |
| 3772 | Ops.push_back(Chain); |
| 3773 | break; |
| 3774 | case NVPTXISD::Suld2DV4I8Clamp: |
| 3775 | Opc = NVPTX::SULD_2D_V4I8_CLAMP; |
| 3776 | Ops.push_back(TexHandle); |
| 3777 | Ops.push_back(N->getOperand(2)); |
| 3778 | Ops.push_back(N->getOperand(3)); |
| 3779 | Ops.push_back(Chain); |
| 3780 | break; |
| 3781 | case NVPTXISD::Suld2DV4I16Clamp: |
| 3782 | Opc = NVPTX::SULD_2D_V4I16_CLAMP; |
| 3783 | Ops.push_back(TexHandle); |
| 3784 | Ops.push_back(N->getOperand(2)); |
| 3785 | Ops.push_back(N->getOperand(3)); |
| 3786 | Ops.push_back(Chain); |
| 3787 | break; |
| 3788 | case NVPTXISD::Suld2DV4I32Clamp: |
| 3789 | Opc = NVPTX::SULD_2D_V4I32_CLAMP; |
| 3790 | Ops.push_back(TexHandle); |
| 3791 | Ops.push_back(N->getOperand(2)); |
| 3792 | Ops.push_back(N->getOperand(3)); |
| 3793 | Ops.push_back(Chain); |
| 3794 | break; |
| 3795 | case NVPTXISD::Suld2DArrayI8Clamp: |
| 3796 | Opc = NVPTX::SULD_2D_ARRAY_I8_CLAMP; |
| 3797 | Ops.push_back(TexHandle); |
| 3798 | Ops.push_back(N->getOperand(2)); |
| 3799 | Ops.push_back(N->getOperand(3)); |
| 3800 | Ops.push_back(N->getOperand(4)); |
| 3801 | Ops.push_back(Chain); |
| 3802 | break; |
| 3803 | case NVPTXISD::Suld2DArrayI16Clamp: |
| 3804 | Opc = NVPTX::SULD_2D_ARRAY_I16_CLAMP; |
| 3805 | Ops.push_back(TexHandle); |
| 3806 | Ops.push_back(N->getOperand(2)); |
| 3807 | Ops.push_back(N->getOperand(3)); |
| 3808 | Ops.push_back(N->getOperand(4)); |
| 3809 | Ops.push_back(Chain); |
| 3810 | break; |
| 3811 | case NVPTXISD::Suld2DArrayI32Clamp: |
| 3812 | Opc = NVPTX::SULD_2D_ARRAY_I32_CLAMP; |
| 3813 | Ops.push_back(TexHandle); |
| 3814 | Ops.push_back(N->getOperand(2)); |
| 3815 | Ops.push_back(N->getOperand(3)); |
| 3816 | Ops.push_back(N->getOperand(4)); |
| 3817 | Ops.push_back(Chain); |
| 3818 | break; |
| 3819 | case NVPTXISD::Suld2DArrayI64Clamp: |
| 3820 | Opc = NVPTX::SULD_2D_ARRAY_I64_CLAMP; |
| 3821 | Ops.push_back(TexHandle); |
| 3822 | Ops.push_back(N->getOperand(2)); |
| 3823 | Ops.push_back(N->getOperand(3)); |
| 3824 | Ops.push_back(N->getOperand(4)); |
| 3825 | Ops.push_back(Chain); |
| 3826 | break; |
| 3827 | case NVPTXISD::Suld2DArrayV2I8Clamp: |
| 3828 | Opc = NVPTX::SULD_2D_ARRAY_V2I8_CLAMP; |
| 3829 | Ops.push_back(TexHandle); |
| 3830 | Ops.push_back(N->getOperand(2)); |
| 3831 | Ops.push_back(N->getOperand(3)); |
| 3832 | Ops.push_back(N->getOperand(4)); |
| 3833 | Ops.push_back(Chain); |
| 3834 | break; |
| 3835 | case NVPTXISD::Suld2DArrayV2I16Clamp: |
| 3836 | Opc = NVPTX::SULD_2D_ARRAY_V2I16_CLAMP; |
| 3837 | Ops.push_back(TexHandle); |
| 3838 | Ops.push_back(N->getOperand(2)); |
| 3839 | Ops.push_back(N->getOperand(3)); |
| 3840 | Ops.push_back(N->getOperand(4)); |
| 3841 | Ops.push_back(Chain); |
| 3842 | break; |
| 3843 | case NVPTXISD::Suld2DArrayV2I32Clamp: |
| 3844 | Opc = NVPTX::SULD_2D_ARRAY_V2I32_CLAMP; |
| 3845 | Ops.push_back(TexHandle); |
| 3846 | Ops.push_back(N->getOperand(2)); |
| 3847 | Ops.push_back(N->getOperand(3)); |
| 3848 | Ops.push_back(N->getOperand(4)); |
| 3849 | Ops.push_back(Chain); |
| 3850 | break; |
| 3851 | case NVPTXISD::Suld2DArrayV2I64Clamp: |
| 3852 | Opc = NVPTX::SULD_2D_ARRAY_V2I64_CLAMP; |
| 3853 | Ops.push_back(TexHandle); |
| 3854 | Ops.push_back(N->getOperand(2)); |
| 3855 | Ops.push_back(N->getOperand(3)); |
| 3856 | Ops.push_back(N->getOperand(4)); |
| 3857 | Ops.push_back(Chain); |
| 3858 | break; |
| 3859 | case NVPTXISD::Suld2DArrayV4I8Clamp: |
| 3860 | Opc = NVPTX::SULD_2D_ARRAY_V4I8_CLAMP; |
| 3861 | Ops.push_back(TexHandle); |
| 3862 | Ops.push_back(N->getOperand(2)); |
| 3863 | Ops.push_back(N->getOperand(3)); |
| 3864 | Ops.push_back(N->getOperand(4)); |
| 3865 | Ops.push_back(Chain); |
| 3866 | break; |
| 3867 | case NVPTXISD::Suld2DArrayV4I16Clamp: |
| 3868 | Opc = NVPTX::SULD_2D_ARRAY_V4I16_CLAMP; |
| 3869 | Ops.push_back(TexHandle); |
| 3870 | Ops.push_back(N->getOperand(2)); |
| 3871 | Ops.push_back(N->getOperand(3)); |
| 3872 | Ops.push_back(N->getOperand(4)); |
| 3873 | Ops.push_back(Chain); |
| 3874 | break; |
| 3875 | case NVPTXISD::Suld2DArrayV4I32Clamp: |
| 3876 | Opc = NVPTX::SULD_2D_ARRAY_V4I32_CLAMP; |
| 3877 | Ops.push_back(TexHandle); |
| 3878 | Ops.push_back(N->getOperand(2)); |
| 3879 | Ops.push_back(N->getOperand(3)); |
| 3880 | Ops.push_back(N->getOperand(4)); |
| 3881 | Ops.push_back(Chain); |
| 3882 | break; |
| 3883 | case NVPTXISD::Suld3DI8Clamp: |
| 3884 | Opc = NVPTX::SULD_3D_I8_CLAMP; |
| 3885 | Ops.push_back(TexHandle); |
| 3886 | Ops.push_back(N->getOperand(2)); |
| 3887 | Ops.push_back(N->getOperand(3)); |
| 3888 | Ops.push_back(N->getOperand(4)); |
| 3889 | Ops.push_back(Chain); |
| 3890 | break; |
| 3891 | case NVPTXISD::Suld3DI16Clamp: |
| 3892 | Opc = NVPTX::SULD_3D_I16_CLAMP; |
| 3893 | Ops.push_back(TexHandle); |
| 3894 | Ops.push_back(N->getOperand(2)); |
| 3895 | Ops.push_back(N->getOperand(3)); |
| 3896 | Ops.push_back(N->getOperand(4)); |
| 3897 | Ops.push_back(Chain); |
| 3898 | break; |
| 3899 | case NVPTXISD::Suld3DI32Clamp: |
| 3900 | Opc = NVPTX::SULD_3D_I32_CLAMP; |
| 3901 | Ops.push_back(TexHandle); |
| 3902 | Ops.push_back(N->getOperand(2)); |
| 3903 | Ops.push_back(N->getOperand(3)); |
| 3904 | Ops.push_back(N->getOperand(4)); |
| 3905 | Ops.push_back(Chain); |
| 3906 | break; |
| 3907 | case NVPTXISD::Suld3DI64Clamp: |
| 3908 | Opc = NVPTX::SULD_3D_I64_CLAMP; |
| 3909 | Ops.push_back(TexHandle); |
| 3910 | Ops.push_back(N->getOperand(2)); |
| 3911 | Ops.push_back(N->getOperand(3)); |
| 3912 | Ops.push_back(N->getOperand(4)); |
| 3913 | Ops.push_back(Chain); |
| 3914 | break; |
| 3915 | case NVPTXISD::Suld3DV2I8Clamp: |
| 3916 | Opc = NVPTX::SULD_3D_V2I8_CLAMP; |
| 3917 | Ops.push_back(TexHandle); |
| 3918 | Ops.push_back(N->getOperand(2)); |
| 3919 | Ops.push_back(N->getOperand(3)); |
| 3920 | Ops.push_back(N->getOperand(4)); |
| 3921 | Ops.push_back(Chain); |
| 3922 | break; |
| 3923 | case NVPTXISD::Suld3DV2I16Clamp: |
| 3924 | Opc = NVPTX::SULD_3D_V2I16_CLAMP; |
| 3925 | Ops.push_back(TexHandle); |
| 3926 | Ops.push_back(N->getOperand(2)); |
| 3927 | Ops.push_back(N->getOperand(3)); |
| 3928 | Ops.push_back(N->getOperand(4)); |
| 3929 | Ops.push_back(Chain); |
| 3930 | break; |
| 3931 | case NVPTXISD::Suld3DV2I32Clamp: |
| 3932 | Opc = NVPTX::SULD_3D_V2I32_CLAMP; |
| 3933 | Ops.push_back(TexHandle); |
| 3934 | Ops.push_back(N->getOperand(2)); |
| 3935 | Ops.push_back(N->getOperand(3)); |
| 3936 | Ops.push_back(N->getOperand(4)); |
| 3937 | Ops.push_back(Chain); |
| 3938 | break; |
| 3939 | case NVPTXISD::Suld3DV2I64Clamp: |
| 3940 | Opc = NVPTX::SULD_3D_V2I64_CLAMP; |
| 3941 | Ops.push_back(TexHandle); |
| 3942 | Ops.push_back(N->getOperand(2)); |
| 3943 | Ops.push_back(N->getOperand(3)); |
| 3944 | Ops.push_back(N->getOperand(4)); |
| 3945 | Ops.push_back(Chain); |
| 3946 | break; |
| 3947 | case NVPTXISD::Suld3DV4I8Clamp: |
| 3948 | Opc = NVPTX::SULD_3D_V4I8_CLAMP; |
| 3949 | Ops.push_back(TexHandle); |
| 3950 | Ops.push_back(N->getOperand(2)); |
| 3951 | Ops.push_back(N->getOperand(3)); |
| 3952 | Ops.push_back(N->getOperand(4)); |
| 3953 | Ops.push_back(Chain); |
| 3954 | break; |
| 3955 | case NVPTXISD::Suld3DV4I16Clamp: |
| 3956 | Opc = NVPTX::SULD_3D_V4I16_CLAMP; |
| 3957 | Ops.push_back(TexHandle); |
| 3958 | Ops.push_back(N->getOperand(2)); |
| 3959 | Ops.push_back(N->getOperand(3)); |
| 3960 | Ops.push_back(N->getOperand(4)); |
| 3961 | Ops.push_back(Chain); |
| 3962 | break; |
| 3963 | case NVPTXISD::Suld3DV4I32Clamp: |
| 3964 | Opc = NVPTX::SULD_3D_V4I32_CLAMP; |
| 3965 | Ops.push_back(TexHandle); |
| 3966 | Ops.push_back(N->getOperand(2)); |
| 3967 | Ops.push_back(N->getOperand(3)); |
| 3968 | Ops.push_back(N->getOperand(4)); |
| 3969 | Ops.push_back(Chain); |
| 3970 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3971 | case NVPTXISD::Suld1DI8Trap: |
| 3972 | Opc = NVPTX::SULD_1D_I8_TRAP; |
| 3973 | Ops.push_back(TexHandle); |
| 3974 | Ops.push_back(N->getOperand(2)); |
| 3975 | Ops.push_back(Chain); |
| 3976 | break; |
| 3977 | case NVPTXISD::Suld1DI16Trap: |
| 3978 | Opc = NVPTX::SULD_1D_I16_TRAP; |
| 3979 | Ops.push_back(TexHandle); |
| 3980 | Ops.push_back(N->getOperand(2)); |
| 3981 | Ops.push_back(Chain); |
| 3982 | break; |
| 3983 | case NVPTXISD::Suld1DI32Trap: |
| 3984 | Opc = NVPTX::SULD_1D_I32_TRAP; |
| 3985 | Ops.push_back(TexHandle); |
| 3986 | Ops.push_back(N->getOperand(2)); |
| 3987 | Ops.push_back(Chain); |
| 3988 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 3989 | case NVPTXISD::Suld1DI64Trap: |
| 3990 | Opc = NVPTX::SULD_1D_I64_TRAP; |
| 3991 | Ops.push_back(TexHandle); |
| 3992 | Ops.push_back(N->getOperand(2)); |
| 3993 | Ops.push_back(Chain); |
| 3994 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 3995 | case NVPTXISD::Suld1DV2I8Trap: |
| 3996 | Opc = NVPTX::SULD_1D_V2I8_TRAP; |
| 3997 | Ops.push_back(TexHandle); |
| 3998 | Ops.push_back(N->getOperand(2)); |
| 3999 | Ops.push_back(Chain); |
| 4000 | break; |
| 4001 | case NVPTXISD::Suld1DV2I16Trap: |
| 4002 | Opc = NVPTX::SULD_1D_V2I16_TRAP; |
| 4003 | Ops.push_back(TexHandle); |
| 4004 | Ops.push_back(N->getOperand(2)); |
| 4005 | Ops.push_back(Chain); |
| 4006 | break; |
| 4007 | case NVPTXISD::Suld1DV2I32Trap: |
| 4008 | Opc = NVPTX::SULD_1D_V2I32_TRAP; |
| 4009 | Ops.push_back(TexHandle); |
| 4010 | Ops.push_back(N->getOperand(2)); |
| 4011 | Ops.push_back(Chain); |
| 4012 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4013 | case NVPTXISD::Suld1DV2I64Trap: |
| 4014 | Opc = NVPTX::SULD_1D_V2I64_TRAP; |
| 4015 | Ops.push_back(TexHandle); |
| 4016 | Ops.push_back(N->getOperand(2)); |
| 4017 | Ops.push_back(Chain); |
| 4018 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4019 | case NVPTXISD::Suld1DV4I8Trap: |
| 4020 | Opc = NVPTX::SULD_1D_V4I8_TRAP; |
| 4021 | Ops.push_back(TexHandle); |
| 4022 | Ops.push_back(N->getOperand(2)); |
| 4023 | Ops.push_back(Chain); |
| 4024 | break; |
| 4025 | case NVPTXISD::Suld1DV4I16Trap: |
| 4026 | Opc = NVPTX::SULD_1D_V4I16_TRAP; |
| 4027 | Ops.push_back(TexHandle); |
| 4028 | Ops.push_back(N->getOperand(2)); |
| 4029 | Ops.push_back(Chain); |
| 4030 | break; |
| 4031 | case NVPTXISD::Suld1DV4I32Trap: |
| 4032 | Opc = NVPTX::SULD_1D_V4I32_TRAP; |
| 4033 | Ops.push_back(TexHandle); |
| 4034 | Ops.push_back(N->getOperand(2)); |
| 4035 | Ops.push_back(Chain); |
| 4036 | break; |
| 4037 | case NVPTXISD::Suld1DArrayI8Trap: |
| 4038 | Opc = NVPTX::SULD_1D_ARRAY_I8_TRAP; |
| 4039 | Ops.push_back(TexHandle); |
| 4040 | Ops.push_back(N->getOperand(2)); |
| 4041 | Ops.push_back(N->getOperand(3)); |
| 4042 | Ops.push_back(Chain); |
| 4043 | break; |
| 4044 | case NVPTXISD::Suld1DArrayI16Trap: |
| 4045 | Opc = NVPTX::SULD_1D_ARRAY_I16_TRAP; |
| 4046 | Ops.push_back(TexHandle); |
| 4047 | Ops.push_back(N->getOperand(2)); |
| 4048 | Ops.push_back(N->getOperand(3)); |
| 4049 | Ops.push_back(Chain); |
| 4050 | break; |
| 4051 | case NVPTXISD::Suld1DArrayI32Trap: |
| 4052 | Opc = NVPTX::SULD_1D_ARRAY_I32_TRAP; |
| 4053 | Ops.push_back(TexHandle); |
| 4054 | Ops.push_back(N->getOperand(2)); |
| 4055 | Ops.push_back(N->getOperand(3)); |
| 4056 | Ops.push_back(Chain); |
| 4057 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4058 | case NVPTXISD::Suld1DArrayI64Trap: |
| 4059 | Opc = NVPTX::SULD_1D_ARRAY_I64_TRAP; |
| 4060 | Ops.push_back(TexHandle); |
| 4061 | Ops.push_back(N->getOperand(2)); |
| 4062 | Ops.push_back(N->getOperand(3)); |
| 4063 | Ops.push_back(Chain); |
| 4064 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4065 | case NVPTXISD::Suld1DArrayV2I8Trap: |
| 4066 | Opc = NVPTX::SULD_1D_ARRAY_V2I8_TRAP; |
| 4067 | Ops.push_back(TexHandle); |
| 4068 | Ops.push_back(N->getOperand(2)); |
| 4069 | Ops.push_back(N->getOperand(3)); |
| 4070 | Ops.push_back(Chain); |
| 4071 | break; |
| 4072 | case NVPTXISD::Suld1DArrayV2I16Trap: |
| 4073 | Opc = NVPTX::SULD_1D_ARRAY_V2I16_TRAP; |
| 4074 | Ops.push_back(TexHandle); |
| 4075 | Ops.push_back(N->getOperand(2)); |
| 4076 | Ops.push_back(N->getOperand(3)); |
| 4077 | Ops.push_back(Chain); |
| 4078 | break; |
| 4079 | case NVPTXISD::Suld1DArrayV2I32Trap: |
| 4080 | Opc = NVPTX::SULD_1D_ARRAY_V2I32_TRAP; |
| 4081 | Ops.push_back(TexHandle); |
| 4082 | Ops.push_back(N->getOperand(2)); |
| 4083 | Ops.push_back(N->getOperand(3)); |
| 4084 | Ops.push_back(Chain); |
| 4085 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4086 | case NVPTXISD::Suld1DArrayV2I64Trap: |
| 4087 | Opc = NVPTX::SULD_1D_ARRAY_V2I64_TRAP; |
| 4088 | Ops.push_back(TexHandle); |
| 4089 | Ops.push_back(N->getOperand(2)); |
| 4090 | Ops.push_back(N->getOperand(3)); |
| 4091 | Ops.push_back(Chain); |
| 4092 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4093 | case NVPTXISD::Suld1DArrayV4I8Trap: |
| 4094 | Opc = NVPTX::SULD_1D_ARRAY_V4I8_TRAP; |
| 4095 | Ops.push_back(TexHandle); |
| 4096 | Ops.push_back(N->getOperand(2)); |
| 4097 | Ops.push_back(N->getOperand(3)); |
| 4098 | Ops.push_back(Chain); |
| 4099 | break; |
| 4100 | case NVPTXISD::Suld1DArrayV4I16Trap: |
| 4101 | Opc = NVPTX::SULD_1D_ARRAY_V4I16_TRAP; |
| 4102 | Ops.push_back(TexHandle); |
| 4103 | Ops.push_back(N->getOperand(2)); |
| 4104 | Ops.push_back(N->getOperand(3)); |
| 4105 | Ops.push_back(Chain); |
| 4106 | break; |
| 4107 | case NVPTXISD::Suld1DArrayV4I32Trap: |
| 4108 | Opc = NVPTX::SULD_1D_ARRAY_V4I32_TRAP; |
| 4109 | Ops.push_back(TexHandle); |
| 4110 | Ops.push_back(N->getOperand(2)); |
| 4111 | Ops.push_back(N->getOperand(3)); |
| 4112 | Ops.push_back(Chain); |
| 4113 | break; |
| 4114 | case NVPTXISD::Suld2DI8Trap: |
| 4115 | Opc = NVPTX::SULD_2D_I8_TRAP; |
| 4116 | Ops.push_back(TexHandle); |
| 4117 | Ops.push_back(N->getOperand(2)); |
| 4118 | Ops.push_back(N->getOperand(3)); |
| 4119 | Ops.push_back(Chain); |
| 4120 | break; |
| 4121 | case NVPTXISD::Suld2DI16Trap: |
| 4122 | Opc = NVPTX::SULD_2D_I16_TRAP; |
| 4123 | Ops.push_back(TexHandle); |
| 4124 | Ops.push_back(N->getOperand(2)); |
| 4125 | Ops.push_back(N->getOperand(3)); |
| 4126 | Ops.push_back(Chain); |
| 4127 | break; |
| 4128 | case NVPTXISD::Suld2DI32Trap: |
| 4129 | Opc = NVPTX::SULD_2D_I32_TRAP; |
| 4130 | Ops.push_back(TexHandle); |
| 4131 | Ops.push_back(N->getOperand(2)); |
| 4132 | Ops.push_back(N->getOperand(3)); |
| 4133 | Ops.push_back(Chain); |
| 4134 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4135 | case NVPTXISD::Suld2DI64Trap: |
| 4136 | Opc = NVPTX::SULD_2D_I64_TRAP; |
| 4137 | Ops.push_back(TexHandle); |
| 4138 | Ops.push_back(N->getOperand(2)); |
| 4139 | Ops.push_back(N->getOperand(3)); |
| 4140 | Ops.push_back(Chain); |
| 4141 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4142 | case NVPTXISD::Suld2DV2I8Trap: |
| 4143 | Opc = NVPTX::SULD_2D_V2I8_TRAP; |
| 4144 | Ops.push_back(TexHandle); |
| 4145 | Ops.push_back(N->getOperand(2)); |
| 4146 | Ops.push_back(N->getOperand(3)); |
| 4147 | Ops.push_back(Chain); |
| 4148 | break; |
| 4149 | case NVPTXISD::Suld2DV2I16Trap: |
| 4150 | Opc = NVPTX::SULD_2D_V2I16_TRAP; |
| 4151 | Ops.push_back(TexHandle); |
| 4152 | Ops.push_back(N->getOperand(2)); |
| 4153 | Ops.push_back(N->getOperand(3)); |
| 4154 | Ops.push_back(Chain); |
| 4155 | break; |
| 4156 | case NVPTXISD::Suld2DV2I32Trap: |
| 4157 | Opc = NVPTX::SULD_2D_V2I32_TRAP; |
| 4158 | Ops.push_back(TexHandle); |
| 4159 | Ops.push_back(N->getOperand(2)); |
| 4160 | Ops.push_back(N->getOperand(3)); |
| 4161 | Ops.push_back(Chain); |
| 4162 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4163 | case NVPTXISD::Suld2DV2I64Trap: |
| 4164 | Opc = NVPTX::SULD_2D_V2I64_TRAP; |
| 4165 | Ops.push_back(TexHandle); |
| 4166 | Ops.push_back(N->getOperand(2)); |
| 4167 | Ops.push_back(N->getOperand(3)); |
| 4168 | Ops.push_back(Chain); |
| 4169 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4170 | case NVPTXISD::Suld2DV4I8Trap: |
| 4171 | Opc = NVPTX::SULD_2D_V4I8_TRAP; |
| 4172 | Ops.push_back(TexHandle); |
| 4173 | Ops.push_back(N->getOperand(2)); |
| 4174 | Ops.push_back(N->getOperand(3)); |
| 4175 | Ops.push_back(Chain); |
| 4176 | break; |
| 4177 | case NVPTXISD::Suld2DV4I16Trap: |
| 4178 | Opc = NVPTX::SULD_2D_V4I16_TRAP; |
| 4179 | Ops.push_back(TexHandle); |
| 4180 | Ops.push_back(N->getOperand(2)); |
| 4181 | Ops.push_back(N->getOperand(3)); |
| 4182 | Ops.push_back(Chain); |
| 4183 | break; |
| 4184 | case NVPTXISD::Suld2DV4I32Trap: |
| 4185 | Opc = NVPTX::SULD_2D_V4I32_TRAP; |
| 4186 | Ops.push_back(TexHandle); |
| 4187 | Ops.push_back(N->getOperand(2)); |
| 4188 | Ops.push_back(N->getOperand(3)); |
| 4189 | Ops.push_back(Chain); |
| 4190 | break; |
| 4191 | case NVPTXISD::Suld2DArrayI8Trap: |
| 4192 | Opc = NVPTX::SULD_2D_ARRAY_I8_TRAP; |
| 4193 | Ops.push_back(TexHandle); |
| 4194 | Ops.push_back(N->getOperand(2)); |
| 4195 | Ops.push_back(N->getOperand(3)); |
| 4196 | Ops.push_back(N->getOperand(4)); |
| 4197 | Ops.push_back(Chain); |
| 4198 | break; |
| 4199 | case NVPTXISD::Suld2DArrayI16Trap: |
| 4200 | Opc = NVPTX::SULD_2D_ARRAY_I16_TRAP; |
| 4201 | Ops.push_back(TexHandle); |
| 4202 | Ops.push_back(N->getOperand(2)); |
| 4203 | Ops.push_back(N->getOperand(3)); |
| 4204 | Ops.push_back(N->getOperand(4)); |
| 4205 | Ops.push_back(Chain); |
| 4206 | break; |
| 4207 | case NVPTXISD::Suld2DArrayI32Trap: |
| 4208 | Opc = NVPTX::SULD_2D_ARRAY_I32_TRAP; |
| 4209 | Ops.push_back(TexHandle); |
| 4210 | Ops.push_back(N->getOperand(2)); |
| 4211 | Ops.push_back(N->getOperand(3)); |
| 4212 | Ops.push_back(N->getOperand(4)); |
| 4213 | Ops.push_back(Chain); |
| 4214 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4215 | case NVPTXISD::Suld2DArrayI64Trap: |
| 4216 | Opc = NVPTX::SULD_2D_ARRAY_I64_TRAP; |
| 4217 | Ops.push_back(TexHandle); |
| 4218 | Ops.push_back(N->getOperand(2)); |
| 4219 | Ops.push_back(N->getOperand(3)); |
| 4220 | Ops.push_back(N->getOperand(4)); |
| 4221 | Ops.push_back(Chain); |
| 4222 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4223 | case NVPTXISD::Suld2DArrayV2I8Trap: |
| 4224 | Opc = NVPTX::SULD_2D_ARRAY_V2I8_TRAP; |
| 4225 | Ops.push_back(TexHandle); |
| 4226 | Ops.push_back(N->getOperand(2)); |
| 4227 | Ops.push_back(N->getOperand(3)); |
| 4228 | Ops.push_back(N->getOperand(4)); |
| 4229 | Ops.push_back(Chain); |
| 4230 | break; |
| 4231 | case NVPTXISD::Suld2DArrayV2I16Trap: |
| 4232 | Opc = NVPTX::SULD_2D_ARRAY_V2I16_TRAP; |
| 4233 | Ops.push_back(TexHandle); |
| 4234 | Ops.push_back(N->getOperand(2)); |
| 4235 | Ops.push_back(N->getOperand(3)); |
| 4236 | Ops.push_back(N->getOperand(4)); |
| 4237 | Ops.push_back(Chain); |
| 4238 | break; |
| 4239 | case NVPTXISD::Suld2DArrayV2I32Trap: |
| 4240 | Opc = NVPTX::SULD_2D_ARRAY_V2I32_TRAP; |
| 4241 | Ops.push_back(TexHandle); |
| 4242 | Ops.push_back(N->getOperand(2)); |
| 4243 | Ops.push_back(N->getOperand(3)); |
| 4244 | Ops.push_back(N->getOperand(4)); |
| 4245 | Ops.push_back(Chain); |
| 4246 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4247 | case NVPTXISD::Suld2DArrayV2I64Trap: |
| 4248 | Opc = NVPTX::SULD_2D_ARRAY_V2I64_TRAP; |
| 4249 | Ops.push_back(TexHandle); |
| 4250 | Ops.push_back(N->getOperand(2)); |
| 4251 | Ops.push_back(N->getOperand(3)); |
| 4252 | Ops.push_back(N->getOperand(4)); |
| 4253 | Ops.push_back(Chain); |
| 4254 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4255 | case NVPTXISD::Suld2DArrayV4I8Trap: |
| 4256 | Opc = NVPTX::SULD_2D_ARRAY_V4I8_TRAP; |
| 4257 | Ops.push_back(TexHandle); |
| 4258 | Ops.push_back(N->getOperand(2)); |
| 4259 | Ops.push_back(N->getOperand(3)); |
| 4260 | Ops.push_back(N->getOperand(4)); |
| 4261 | Ops.push_back(Chain); |
| 4262 | break; |
| 4263 | case NVPTXISD::Suld2DArrayV4I16Trap: |
| 4264 | Opc = NVPTX::SULD_2D_ARRAY_V4I16_TRAP; |
| 4265 | Ops.push_back(TexHandle); |
| 4266 | Ops.push_back(N->getOperand(2)); |
| 4267 | Ops.push_back(N->getOperand(3)); |
| 4268 | Ops.push_back(N->getOperand(4)); |
| 4269 | Ops.push_back(Chain); |
| 4270 | break; |
| 4271 | case NVPTXISD::Suld2DArrayV4I32Trap: |
| 4272 | Opc = NVPTX::SULD_2D_ARRAY_V4I32_TRAP; |
| 4273 | Ops.push_back(TexHandle); |
| 4274 | Ops.push_back(N->getOperand(2)); |
| 4275 | Ops.push_back(N->getOperand(3)); |
| 4276 | Ops.push_back(N->getOperand(4)); |
| 4277 | Ops.push_back(Chain); |
| 4278 | break; |
| 4279 | case NVPTXISD::Suld3DI8Trap: |
| 4280 | Opc = NVPTX::SULD_3D_I8_TRAP; |
| 4281 | Ops.push_back(TexHandle); |
| 4282 | Ops.push_back(N->getOperand(2)); |
| 4283 | Ops.push_back(N->getOperand(3)); |
| 4284 | Ops.push_back(N->getOperand(4)); |
| 4285 | Ops.push_back(Chain); |
| 4286 | break; |
| 4287 | case NVPTXISD::Suld3DI16Trap: |
| 4288 | Opc = NVPTX::SULD_3D_I16_TRAP; |
| 4289 | Ops.push_back(TexHandle); |
| 4290 | Ops.push_back(N->getOperand(2)); |
| 4291 | Ops.push_back(N->getOperand(3)); |
| 4292 | Ops.push_back(N->getOperand(4)); |
| 4293 | Ops.push_back(Chain); |
| 4294 | break; |
| 4295 | case NVPTXISD::Suld3DI32Trap: |
| 4296 | Opc = NVPTX::SULD_3D_I32_TRAP; |
| 4297 | Ops.push_back(TexHandle); |
| 4298 | Ops.push_back(N->getOperand(2)); |
| 4299 | Ops.push_back(N->getOperand(3)); |
| 4300 | Ops.push_back(N->getOperand(4)); |
| 4301 | Ops.push_back(Chain); |
| 4302 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4303 | case NVPTXISD::Suld3DI64Trap: |
| 4304 | Opc = NVPTX::SULD_3D_I64_TRAP; |
| 4305 | Ops.push_back(TexHandle); |
| 4306 | Ops.push_back(N->getOperand(2)); |
| 4307 | Ops.push_back(N->getOperand(3)); |
| 4308 | Ops.push_back(N->getOperand(4)); |
| 4309 | Ops.push_back(Chain); |
| 4310 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4311 | case NVPTXISD::Suld3DV2I8Trap: |
| 4312 | Opc = NVPTX::SULD_3D_V2I8_TRAP; |
| 4313 | Ops.push_back(TexHandle); |
| 4314 | Ops.push_back(N->getOperand(2)); |
| 4315 | Ops.push_back(N->getOperand(3)); |
| 4316 | Ops.push_back(N->getOperand(4)); |
| 4317 | Ops.push_back(Chain); |
| 4318 | break; |
| 4319 | case NVPTXISD::Suld3DV2I16Trap: |
| 4320 | Opc = NVPTX::SULD_3D_V2I16_TRAP; |
| 4321 | Ops.push_back(TexHandle); |
| 4322 | Ops.push_back(N->getOperand(2)); |
| 4323 | Ops.push_back(N->getOperand(3)); |
| 4324 | Ops.push_back(N->getOperand(4)); |
| 4325 | Ops.push_back(Chain); |
| 4326 | break; |
| 4327 | case NVPTXISD::Suld3DV2I32Trap: |
| 4328 | Opc = NVPTX::SULD_3D_V2I32_TRAP; |
| 4329 | Ops.push_back(TexHandle); |
| 4330 | Ops.push_back(N->getOperand(2)); |
| 4331 | Ops.push_back(N->getOperand(3)); |
| 4332 | Ops.push_back(N->getOperand(4)); |
| 4333 | Ops.push_back(Chain); |
| 4334 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4335 | case NVPTXISD::Suld3DV2I64Trap: |
| 4336 | Opc = NVPTX::SULD_3D_V2I64_TRAP; |
| 4337 | Ops.push_back(TexHandle); |
| 4338 | Ops.push_back(N->getOperand(2)); |
| 4339 | Ops.push_back(N->getOperand(3)); |
| 4340 | Ops.push_back(N->getOperand(4)); |
| 4341 | Ops.push_back(Chain); |
| 4342 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4343 | case NVPTXISD::Suld3DV4I8Trap: |
| 4344 | Opc = NVPTX::SULD_3D_V4I8_TRAP; |
| 4345 | Ops.push_back(TexHandle); |
| 4346 | Ops.push_back(N->getOperand(2)); |
| 4347 | Ops.push_back(N->getOperand(3)); |
| 4348 | Ops.push_back(N->getOperand(4)); |
| 4349 | Ops.push_back(Chain); |
| 4350 | break; |
| 4351 | case NVPTXISD::Suld3DV4I16Trap: |
| 4352 | Opc = NVPTX::SULD_3D_V4I16_TRAP; |
| 4353 | Ops.push_back(TexHandle); |
| 4354 | Ops.push_back(N->getOperand(2)); |
| 4355 | Ops.push_back(N->getOperand(3)); |
| 4356 | Ops.push_back(N->getOperand(4)); |
| 4357 | Ops.push_back(Chain); |
| 4358 | break; |
| 4359 | case NVPTXISD::Suld3DV4I32Trap: |
| 4360 | Opc = NVPTX::SULD_3D_V4I32_TRAP; |
| 4361 | Ops.push_back(TexHandle); |
| 4362 | Ops.push_back(N->getOperand(2)); |
| 4363 | Ops.push_back(N->getOperand(3)); |
| 4364 | Ops.push_back(N->getOperand(4)); |
| 4365 | Ops.push_back(Chain); |
| 4366 | break; |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4367 | case NVPTXISD::Suld1DI8Zero: |
| 4368 | Opc = NVPTX::SULD_1D_I8_ZERO; |
| 4369 | Ops.push_back(TexHandle); |
| 4370 | Ops.push_back(N->getOperand(2)); |
| 4371 | Ops.push_back(Chain); |
| 4372 | break; |
| 4373 | case NVPTXISD::Suld1DI16Zero: |
| 4374 | Opc = NVPTX::SULD_1D_I16_ZERO; |
| 4375 | Ops.push_back(TexHandle); |
| 4376 | Ops.push_back(N->getOperand(2)); |
| 4377 | Ops.push_back(Chain); |
| 4378 | break; |
| 4379 | case NVPTXISD::Suld1DI32Zero: |
| 4380 | Opc = NVPTX::SULD_1D_I32_ZERO; |
| 4381 | Ops.push_back(TexHandle); |
| 4382 | Ops.push_back(N->getOperand(2)); |
| 4383 | Ops.push_back(Chain); |
| 4384 | break; |
| 4385 | case NVPTXISD::Suld1DI64Zero: |
| 4386 | Opc = NVPTX::SULD_1D_I64_ZERO; |
| 4387 | Ops.push_back(TexHandle); |
| 4388 | Ops.push_back(N->getOperand(2)); |
| 4389 | Ops.push_back(Chain); |
| 4390 | break; |
| 4391 | case NVPTXISD::Suld1DV2I8Zero: |
| 4392 | Opc = NVPTX::SULD_1D_V2I8_ZERO; |
| 4393 | Ops.push_back(TexHandle); |
| 4394 | Ops.push_back(N->getOperand(2)); |
| 4395 | Ops.push_back(Chain); |
| 4396 | break; |
| 4397 | case NVPTXISD::Suld1DV2I16Zero: |
| 4398 | Opc = NVPTX::SULD_1D_V2I16_ZERO; |
| 4399 | Ops.push_back(TexHandle); |
| 4400 | Ops.push_back(N->getOperand(2)); |
| 4401 | Ops.push_back(Chain); |
| 4402 | break; |
| 4403 | case NVPTXISD::Suld1DV2I32Zero: |
| 4404 | Opc = NVPTX::SULD_1D_V2I32_ZERO; |
| 4405 | Ops.push_back(TexHandle); |
| 4406 | Ops.push_back(N->getOperand(2)); |
| 4407 | Ops.push_back(Chain); |
| 4408 | break; |
| 4409 | case NVPTXISD::Suld1DV2I64Zero: |
| 4410 | Opc = NVPTX::SULD_1D_V2I64_ZERO; |
| 4411 | Ops.push_back(TexHandle); |
| 4412 | Ops.push_back(N->getOperand(2)); |
| 4413 | Ops.push_back(Chain); |
| 4414 | break; |
| 4415 | case NVPTXISD::Suld1DV4I8Zero: |
| 4416 | Opc = NVPTX::SULD_1D_V4I8_ZERO; |
| 4417 | Ops.push_back(TexHandle); |
| 4418 | Ops.push_back(N->getOperand(2)); |
| 4419 | Ops.push_back(Chain); |
| 4420 | break; |
| 4421 | case NVPTXISD::Suld1DV4I16Zero: |
| 4422 | Opc = NVPTX::SULD_1D_V4I16_ZERO; |
| 4423 | Ops.push_back(TexHandle); |
| 4424 | Ops.push_back(N->getOperand(2)); |
| 4425 | Ops.push_back(Chain); |
| 4426 | break; |
| 4427 | case NVPTXISD::Suld1DV4I32Zero: |
| 4428 | Opc = NVPTX::SULD_1D_V4I32_ZERO; |
| 4429 | Ops.push_back(TexHandle); |
| 4430 | Ops.push_back(N->getOperand(2)); |
| 4431 | Ops.push_back(Chain); |
| 4432 | break; |
| 4433 | case NVPTXISD::Suld1DArrayI8Zero: |
| 4434 | Opc = NVPTX::SULD_1D_ARRAY_I8_ZERO; |
| 4435 | Ops.push_back(TexHandle); |
| 4436 | Ops.push_back(N->getOperand(2)); |
| 4437 | Ops.push_back(N->getOperand(3)); |
| 4438 | Ops.push_back(Chain); |
| 4439 | break; |
| 4440 | case NVPTXISD::Suld1DArrayI16Zero: |
| 4441 | Opc = NVPTX::SULD_1D_ARRAY_I16_ZERO; |
| 4442 | Ops.push_back(TexHandle); |
| 4443 | Ops.push_back(N->getOperand(2)); |
| 4444 | Ops.push_back(N->getOperand(3)); |
| 4445 | Ops.push_back(Chain); |
| 4446 | break; |
| 4447 | case NVPTXISD::Suld1DArrayI32Zero: |
| 4448 | Opc = NVPTX::SULD_1D_ARRAY_I32_ZERO; |
| 4449 | Ops.push_back(TexHandle); |
| 4450 | Ops.push_back(N->getOperand(2)); |
| 4451 | Ops.push_back(N->getOperand(3)); |
| 4452 | Ops.push_back(Chain); |
| 4453 | break; |
| 4454 | case NVPTXISD::Suld1DArrayI64Zero: |
| 4455 | Opc = NVPTX::SULD_1D_ARRAY_I64_ZERO; |
| 4456 | Ops.push_back(TexHandle); |
| 4457 | Ops.push_back(N->getOperand(2)); |
| 4458 | Ops.push_back(N->getOperand(3)); |
| 4459 | Ops.push_back(Chain); |
| 4460 | break; |
| 4461 | case NVPTXISD::Suld1DArrayV2I8Zero: |
| 4462 | Opc = NVPTX::SULD_1D_ARRAY_V2I8_ZERO; |
| 4463 | Ops.push_back(TexHandle); |
| 4464 | Ops.push_back(N->getOperand(2)); |
| 4465 | Ops.push_back(N->getOperand(3)); |
| 4466 | Ops.push_back(Chain); |
| 4467 | break; |
| 4468 | case NVPTXISD::Suld1DArrayV2I16Zero: |
| 4469 | Opc = NVPTX::SULD_1D_ARRAY_V2I16_ZERO; |
| 4470 | Ops.push_back(TexHandle); |
| 4471 | Ops.push_back(N->getOperand(2)); |
| 4472 | Ops.push_back(N->getOperand(3)); |
| 4473 | Ops.push_back(Chain); |
| 4474 | break; |
| 4475 | case NVPTXISD::Suld1DArrayV2I32Zero: |
| 4476 | Opc = NVPTX::SULD_1D_ARRAY_V2I32_ZERO; |
| 4477 | Ops.push_back(TexHandle); |
| 4478 | Ops.push_back(N->getOperand(2)); |
| 4479 | Ops.push_back(N->getOperand(3)); |
| 4480 | Ops.push_back(Chain); |
| 4481 | break; |
| 4482 | case NVPTXISD::Suld1DArrayV2I64Zero: |
| 4483 | Opc = NVPTX::SULD_1D_ARRAY_V2I64_ZERO; |
| 4484 | Ops.push_back(TexHandle); |
| 4485 | Ops.push_back(N->getOperand(2)); |
| 4486 | Ops.push_back(N->getOperand(3)); |
| 4487 | Ops.push_back(Chain); |
| 4488 | break; |
| 4489 | case NVPTXISD::Suld1DArrayV4I8Zero: |
| 4490 | Opc = NVPTX::SULD_1D_ARRAY_V4I8_ZERO; |
| 4491 | Ops.push_back(TexHandle); |
| 4492 | Ops.push_back(N->getOperand(2)); |
| 4493 | Ops.push_back(N->getOperand(3)); |
| 4494 | Ops.push_back(Chain); |
| 4495 | break; |
| 4496 | case NVPTXISD::Suld1DArrayV4I16Zero: |
| 4497 | Opc = NVPTX::SULD_1D_ARRAY_V4I16_ZERO; |
| 4498 | Ops.push_back(TexHandle); |
| 4499 | Ops.push_back(N->getOperand(2)); |
| 4500 | Ops.push_back(N->getOperand(3)); |
| 4501 | Ops.push_back(Chain); |
| 4502 | break; |
| 4503 | case NVPTXISD::Suld1DArrayV4I32Zero: |
| 4504 | Opc = NVPTX::SULD_1D_ARRAY_V4I32_ZERO; |
| 4505 | Ops.push_back(TexHandle); |
| 4506 | Ops.push_back(N->getOperand(2)); |
| 4507 | Ops.push_back(N->getOperand(3)); |
| 4508 | Ops.push_back(Chain); |
| 4509 | break; |
| 4510 | case NVPTXISD::Suld2DI8Zero: |
| 4511 | Opc = NVPTX::SULD_2D_I8_ZERO; |
| 4512 | Ops.push_back(TexHandle); |
| 4513 | Ops.push_back(N->getOperand(2)); |
| 4514 | Ops.push_back(N->getOperand(3)); |
| 4515 | Ops.push_back(Chain); |
| 4516 | break; |
| 4517 | case NVPTXISD::Suld2DI16Zero: |
| 4518 | Opc = NVPTX::SULD_2D_I16_ZERO; |
| 4519 | Ops.push_back(TexHandle); |
| 4520 | Ops.push_back(N->getOperand(2)); |
| 4521 | Ops.push_back(N->getOperand(3)); |
| 4522 | Ops.push_back(Chain); |
| 4523 | break; |
| 4524 | case NVPTXISD::Suld2DI32Zero: |
| 4525 | Opc = NVPTX::SULD_2D_I32_ZERO; |
| 4526 | Ops.push_back(TexHandle); |
| 4527 | Ops.push_back(N->getOperand(2)); |
| 4528 | Ops.push_back(N->getOperand(3)); |
| 4529 | Ops.push_back(Chain); |
| 4530 | break; |
| 4531 | case NVPTXISD::Suld2DI64Zero: |
| 4532 | Opc = NVPTX::SULD_2D_I64_ZERO; |
| 4533 | Ops.push_back(TexHandle); |
| 4534 | Ops.push_back(N->getOperand(2)); |
| 4535 | Ops.push_back(N->getOperand(3)); |
| 4536 | Ops.push_back(Chain); |
| 4537 | break; |
| 4538 | case NVPTXISD::Suld2DV2I8Zero: |
| 4539 | Opc = NVPTX::SULD_2D_V2I8_ZERO; |
| 4540 | Ops.push_back(TexHandle); |
| 4541 | Ops.push_back(N->getOperand(2)); |
| 4542 | Ops.push_back(N->getOperand(3)); |
| 4543 | Ops.push_back(Chain); |
| 4544 | break; |
| 4545 | case NVPTXISD::Suld2DV2I16Zero: |
| 4546 | Opc = NVPTX::SULD_2D_V2I16_ZERO; |
| 4547 | Ops.push_back(TexHandle); |
| 4548 | Ops.push_back(N->getOperand(2)); |
| 4549 | Ops.push_back(N->getOperand(3)); |
| 4550 | Ops.push_back(Chain); |
| 4551 | break; |
| 4552 | case NVPTXISD::Suld2DV2I32Zero: |
| 4553 | Opc = NVPTX::SULD_2D_V2I32_ZERO; |
| 4554 | Ops.push_back(TexHandle); |
| 4555 | Ops.push_back(N->getOperand(2)); |
| 4556 | Ops.push_back(N->getOperand(3)); |
| 4557 | Ops.push_back(Chain); |
| 4558 | break; |
| 4559 | case NVPTXISD::Suld2DV2I64Zero: |
| 4560 | Opc = NVPTX::SULD_2D_V2I64_ZERO; |
| 4561 | Ops.push_back(TexHandle); |
| 4562 | Ops.push_back(N->getOperand(2)); |
| 4563 | Ops.push_back(N->getOperand(3)); |
| 4564 | Ops.push_back(Chain); |
| 4565 | break; |
| 4566 | case NVPTXISD::Suld2DV4I8Zero: |
| 4567 | Opc = NVPTX::SULD_2D_V4I8_ZERO; |
| 4568 | Ops.push_back(TexHandle); |
| 4569 | Ops.push_back(N->getOperand(2)); |
| 4570 | Ops.push_back(N->getOperand(3)); |
| 4571 | Ops.push_back(Chain); |
| 4572 | break; |
| 4573 | case NVPTXISD::Suld2DV4I16Zero: |
| 4574 | Opc = NVPTX::SULD_2D_V4I16_ZERO; |
| 4575 | Ops.push_back(TexHandle); |
| 4576 | Ops.push_back(N->getOperand(2)); |
| 4577 | Ops.push_back(N->getOperand(3)); |
| 4578 | Ops.push_back(Chain); |
| 4579 | break; |
| 4580 | case NVPTXISD::Suld2DV4I32Zero: |
| 4581 | Opc = NVPTX::SULD_2D_V4I32_ZERO; |
| 4582 | Ops.push_back(TexHandle); |
| 4583 | Ops.push_back(N->getOperand(2)); |
| 4584 | Ops.push_back(N->getOperand(3)); |
| 4585 | Ops.push_back(Chain); |
| 4586 | break; |
| 4587 | case NVPTXISD::Suld2DArrayI8Zero: |
| 4588 | Opc = NVPTX::SULD_2D_ARRAY_I8_ZERO; |
| 4589 | Ops.push_back(TexHandle); |
| 4590 | Ops.push_back(N->getOperand(2)); |
| 4591 | Ops.push_back(N->getOperand(3)); |
| 4592 | Ops.push_back(N->getOperand(4)); |
| 4593 | Ops.push_back(Chain); |
| 4594 | break; |
| 4595 | case NVPTXISD::Suld2DArrayI16Zero: |
| 4596 | Opc = NVPTX::SULD_2D_ARRAY_I16_ZERO; |
| 4597 | Ops.push_back(TexHandle); |
| 4598 | Ops.push_back(N->getOperand(2)); |
| 4599 | Ops.push_back(N->getOperand(3)); |
| 4600 | Ops.push_back(N->getOperand(4)); |
| 4601 | Ops.push_back(Chain); |
| 4602 | break; |
| 4603 | case NVPTXISD::Suld2DArrayI32Zero: |
| 4604 | Opc = NVPTX::SULD_2D_ARRAY_I32_ZERO; |
| 4605 | Ops.push_back(TexHandle); |
| 4606 | Ops.push_back(N->getOperand(2)); |
| 4607 | Ops.push_back(N->getOperand(3)); |
| 4608 | Ops.push_back(N->getOperand(4)); |
| 4609 | Ops.push_back(Chain); |
| 4610 | break; |
| 4611 | case NVPTXISD::Suld2DArrayI64Zero: |
| 4612 | Opc = NVPTX::SULD_2D_ARRAY_I64_ZERO; |
| 4613 | Ops.push_back(TexHandle); |
| 4614 | Ops.push_back(N->getOperand(2)); |
| 4615 | Ops.push_back(N->getOperand(3)); |
| 4616 | Ops.push_back(N->getOperand(4)); |
| 4617 | Ops.push_back(Chain); |
| 4618 | break; |
| 4619 | case NVPTXISD::Suld2DArrayV2I8Zero: |
| 4620 | Opc = NVPTX::SULD_2D_ARRAY_V2I8_ZERO; |
| 4621 | Ops.push_back(TexHandle); |
| 4622 | Ops.push_back(N->getOperand(2)); |
| 4623 | Ops.push_back(N->getOperand(3)); |
| 4624 | Ops.push_back(N->getOperand(4)); |
| 4625 | Ops.push_back(Chain); |
| 4626 | break; |
| 4627 | case NVPTXISD::Suld2DArrayV2I16Zero: |
| 4628 | Opc = NVPTX::SULD_2D_ARRAY_V2I16_ZERO; |
| 4629 | Ops.push_back(TexHandle); |
| 4630 | Ops.push_back(N->getOperand(2)); |
| 4631 | Ops.push_back(N->getOperand(3)); |
| 4632 | Ops.push_back(N->getOperand(4)); |
| 4633 | Ops.push_back(Chain); |
| 4634 | break; |
| 4635 | case NVPTXISD::Suld2DArrayV2I32Zero: |
| 4636 | Opc = NVPTX::SULD_2D_ARRAY_V2I32_ZERO; |
| 4637 | Ops.push_back(TexHandle); |
| 4638 | Ops.push_back(N->getOperand(2)); |
| 4639 | Ops.push_back(N->getOperand(3)); |
| 4640 | Ops.push_back(N->getOperand(4)); |
| 4641 | Ops.push_back(Chain); |
| 4642 | break; |
| 4643 | case NVPTXISD::Suld2DArrayV2I64Zero: |
| 4644 | Opc = NVPTX::SULD_2D_ARRAY_V2I64_ZERO; |
| 4645 | Ops.push_back(TexHandle); |
| 4646 | Ops.push_back(N->getOperand(2)); |
| 4647 | Ops.push_back(N->getOperand(3)); |
| 4648 | Ops.push_back(N->getOperand(4)); |
| 4649 | Ops.push_back(Chain); |
| 4650 | break; |
| 4651 | case NVPTXISD::Suld2DArrayV4I8Zero: |
| 4652 | Opc = NVPTX::SULD_2D_ARRAY_V4I8_ZERO; |
| 4653 | Ops.push_back(TexHandle); |
| 4654 | Ops.push_back(N->getOperand(2)); |
| 4655 | Ops.push_back(N->getOperand(3)); |
| 4656 | Ops.push_back(N->getOperand(4)); |
| 4657 | Ops.push_back(Chain); |
| 4658 | break; |
| 4659 | case NVPTXISD::Suld2DArrayV4I16Zero: |
| 4660 | Opc = NVPTX::SULD_2D_ARRAY_V4I16_ZERO; |
| 4661 | Ops.push_back(TexHandle); |
| 4662 | Ops.push_back(N->getOperand(2)); |
| 4663 | Ops.push_back(N->getOperand(3)); |
| 4664 | Ops.push_back(N->getOperand(4)); |
| 4665 | Ops.push_back(Chain); |
| 4666 | break; |
| 4667 | case NVPTXISD::Suld2DArrayV4I32Zero: |
| 4668 | Opc = NVPTX::SULD_2D_ARRAY_V4I32_ZERO; |
| 4669 | Ops.push_back(TexHandle); |
| 4670 | Ops.push_back(N->getOperand(2)); |
| 4671 | Ops.push_back(N->getOperand(3)); |
| 4672 | Ops.push_back(N->getOperand(4)); |
| 4673 | Ops.push_back(Chain); |
| 4674 | break; |
| 4675 | case NVPTXISD::Suld3DI8Zero: |
| 4676 | Opc = NVPTX::SULD_3D_I8_ZERO; |
| 4677 | Ops.push_back(TexHandle); |
| 4678 | Ops.push_back(N->getOperand(2)); |
| 4679 | Ops.push_back(N->getOperand(3)); |
| 4680 | Ops.push_back(N->getOperand(4)); |
| 4681 | Ops.push_back(Chain); |
| 4682 | break; |
| 4683 | case NVPTXISD::Suld3DI16Zero: |
| 4684 | Opc = NVPTX::SULD_3D_I16_ZERO; |
| 4685 | Ops.push_back(TexHandle); |
| 4686 | Ops.push_back(N->getOperand(2)); |
| 4687 | Ops.push_back(N->getOperand(3)); |
| 4688 | Ops.push_back(N->getOperand(4)); |
| 4689 | Ops.push_back(Chain); |
| 4690 | break; |
| 4691 | case NVPTXISD::Suld3DI32Zero: |
| 4692 | Opc = NVPTX::SULD_3D_I32_ZERO; |
| 4693 | Ops.push_back(TexHandle); |
| 4694 | Ops.push_back(N->getOperand(2)); |
| 4695 | Ops.push_back(N->getOperand(3)); |
| 4696 | Ops.push_back(N->getOperand(4)); |
| 4697 | Ops.push_back(Chain); |
| 4698 | break; |
| 4699 | case NVPTXISD::Suld3DI64Zero: |
| 4700 | Opc = NVPTX::SULD_3D_I64_ZERO; |
| 4701 | Ops.push_back(TexHandle); |
| 4702 | Ops.push_back(N->getOperand(2)); |
| 4703 | Ops.push_back(N->getOperand(3)); |
| 4704 | Ops.push_back(N->getOperand(4)); |
| 4705 | Ops.push_back(Chain); |
| 4706 | break; |
| 4707 | case NVPTXISD::Suld3DV2I8Zero: |
| 4708 | Opc = NVPTX::SULD_3D_V2I8_ZERO; |
| 4709 | Ops.push_back(TexHandle); |
| 4710 | Ops.push_back(N->getOperand(2)); |
| 4711 | Ops.push_back(N->getOperand(3)); |
| 4712 | Ops.push_back(N->getOperand(4)); |
| 4713 | Ops.push_back(Chain); |
| 4714 | break; |
| 4715 | case NVPTXISD::Suld3DV2I16Zero: |
| 4716 | Opc = NVPTX::SULD_3D_V2I16_ZERO; |
| 4717 | Ops.push_back(TexHandle); |
| 4718 | Ops.push_back(N->getOperand(2)); |
| 4719 | Ops.push_back(N->getOperand(3)); |
| 4720 | Ops.push_back(N->getOperand(4)); |
| 4721 | Ops.push_back(Chain); |
| 4722 | break; |
| 4723 | case NVPTXISD::Suld3DV2I32Zero: |
| 4724 | Opc = NVPTX::SULD_3D_V2I32_ZERO; |
| 4725 | Ops.push_back(TexHandle); |
| 4726 | Ops.push_back(N->getOperand(2)); |
| 4727 | Ops.push_back(N->getOperand(3)); |
| 4728 | Ops.push_back(N->getOperand(4)); |
| 4729 | Ops.push_back(Chain); |
| 4730 | break; |
| 4731 | case NVPTXISD::Suld3DV2I64Zero: |
| 4732 | Opc = NVPTX::SULD_3D_V2I64_ZERO; |
| 4733 | Ops.push_back(TexHandle); |
| 4734 | Ops.push_back(N->getOperand(2)); |
| 4735 | Ops.push_back(N->getOperand(3)); |
| 4736 | Ops.push_back(N->getOperand(4)); |
| 4737 | Ops.push_back(Chain); |
| 4738 | break; |
| 4739 | case NVPTXISD::Suld3DV4I8Zero: |
| 4740 | Opc = NVPTX::SULD_3D_V4I8_ZERO; |
| 4741 | Ops.push_back(TexHandle); |
| 4742 | Ops.push_back(N->getOperand(2)); |
| 4743 | Ops.push_back(N->getOperand(3)); |
| 4744 | Ops.push_back(N->getOperand(4)); |
| 4745 | Ops.push_back(Chain); |
| 4746 | break; |
| 4747 | case NVPTXISD::Suld3DV4I16Zero: |
| 4748 | Opc = NVPTX::SULD_3D_V4I16_ZERO; |
| 4749 | Ops.push_back(TexHandle); |
| 4750 | Ops.push_back(N->getOperand(2)); |
| 4751 | Ops.push_back(N->getOperand(3)); |
| 4752 | Ops.push_back(N->getOperand(4)); |
| 4753 | Ops.push_back(Chain); |
| 4754 | break; |
| 4755 | case NVPTXISD::Suld3DV4I32Zero: |
| 4756 | Opc = NVPTX::SULD_3D_V4I32_ZERO; |
| 4757 | Ops.push_back(TexHandle); |
| 4758 | Ops.push_back(N->getOperand(2)); |
| 4759 | Ops.push_back(N->getOperand(3)); |
| 4760 | Ops.push_back(N->getOperand(4)); |
| 4761 | Ops.push_back(Chain); |
| 4762 | break; |
Justin Holewinski | 30d56a7 | 2014-04-09 15:39:15 +0000 | [diff] [blame] | 4763 | } |
| 4764 | Ret = CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops); |
| 4765 | return Ret; |
| 4766 | } |
| 4767 | |
Justin Holewinski | 9a2350e | 2014-07-17 11:59:04 +0000 | [diff] [blame] | 4768 | |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4769 | /// SelectBFE - Look for instruction sequences that can be made more efficient |
| 4770 | /// by using the 'bfe' (bit-field extract) PTX instruction |
| 4771 | SDNode *NVPTXDAGToDAGISel::SelectBFE(SDNode *N) { |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4772 | SDLoc DL(N); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4773 | SDValue LHS = N->getOperand(0); |
| 4774 | SDValue RHS = N->getOperand(1); |
| 4775 | SDValue Len; |
| 4776 | SDValue Start; |
| 4777 | SDValue Val; |
| 4778 | bool IsSigned = false; |
| 4779 | |
| 4780 | if (N->getOpcode() == ISD::AND) { |
| 4781 | // Canonicalize the operands |
| 4782 | // We want 'and %val, %mask' |
| 4783 | if (isa<ConstantSDNode>(LHS) && !isa<ConstantSDNode>(RHS)) { |
| 4784 | std::swap(LHS, RHS); |
| 4785 | } |
| 4786 | |
| 4787 | ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(RHS); |
| 4788 | if (!Mask) { |
| 4789 | // We need a constant mask on the RHS of the AND |
| 4790 | return NULL; |
| 4791 | } |
| 4792 | |
| 4793 | // Extract the mask bits |
| 4794 | uint64_t MaskVal = Mask->getZExtValue(); |
| 4795 | if (!isMask_64(MaskVal)) { |
| 4796 | // We *could* handle shifted masks here, but doing so would require an |
| 4797 | // 'and' operation to fix up the low-order bits so we would trade |
| 4798 | // shr+and for bfe+and, which has the same throughput |
| 4799 | return NULL; |
| 4800 | } |
| 4801 | |
| 4802 | // How many bits are in our mask? |
Benjamin Kramer | 5f6a907 | 2015-02-12 15:35:40 +0000 | [diff] [blame] | 4803 | uint64_t NumBits = countTrailingOnes(MaskVal); |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4804 | Len = CurDAG->getTargetConstant(NumBits, DL, MVT::i32); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4805 | |
| 4806 | if (LHS.getOpcode() == ISD::SRL || LHS.getOpcode() == ISD::SRA) { |
| 4807 | // We have a 'srl/and' pair, extract the effective start bit and length |
| 4808 | Val = LHS.getNode()->getOperand(0); |
| 4809 | Start = LHS.getNode()->getOperand(1); |
| 4810 | ConstantSDNode *StartConst = dyn_cast<ConstantSDNode>(Start); |
| 4811 | if (StartConst) { |
| 4812 | uint64_t StartVal = StartConst->getZExtValue(); |
| 4813 | // How many "good" bits do we have left? "good" is defined here as bits |
| 4814 | // that exist in the original value, not shifted in. |
| 4815 | uint64_t GoodBits = Start.getValueType().getSizeInBits() - StartVal; |
| 4816 | if (NumBits > GoodBits) { |
| 4817 | // Do not handle the case where bits have been shifted in. In theory |
| 4818 | // we could handle this, but the cost is likely higher than just |
| 4819 | // emitting the srl/and pair. |
| 4820 | return NULL; |
| 4821 | } |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4822 | Start = CurDAG->getTargetConstant(StartVal, DL, MVT::i32); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4823 | } else { |
| 4824 | // Do not handle the case where the shift amount (can be zero if no srl |
| 4825 | // was found) is not constant. We could handle this case, but it would |
| 4826 | // require run-time logic that would be more expensive than just |
| 4827 | // emitting the srl/and pair. |
| 4828 | return NULL; |
| 4829 | } |
| 4830 | } else { |
| 4831 | // Do not handle the case where the LHS of the and is not a shift. While |
| 4832 | // it would be trivial to handle this case, it would just transform |
| 4833 | // 'and' -> 'bfe', but 'and' has higher-throughput. |
| 4834 | return NULL; |
| 4835 | } |
| 4836 | } else if (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) { |
| 4837 | if (LHS->getOpcode() == ISD::AND) { |
| 4838 | ConstantSDNode *ShiftCnst = dyn_cast<ConstantSDNode>(RHS); |
| 4839 | if (!ShiftCnst) { |
| 4840 | // Shift amount must be constant |
| 4841 | return NULL; |
| 4842 | } |
| 4843 | |
| 4844 | uint64_t ShiftAmt = ShiftCnst->getZExtValue(); |
| 4845 | |
| 4846 | SDValue AndLHS = LHS->getOperand(0); |
| 4847 | SDValue AndRHS = LHS->getOperand(1); |
| 4848 | |
| 4849 | // Canonicalize the AND to have the mask on the RHS |
| 4850 | if (isa<ConstantSDNode>(AndLHS)) { |
| 4851 | std::swap(AndLHS, AndRHS); |
| 4852 | } |
| 4853 | |
| 4854 | ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(AndRHS); |
| 4855 | if (!MaskCnst) { |
| 4856 | // Mask must be constant |
| 4857 | return NULL; |
| 4858 | } |
| 4859 | |
| 4860 | uint64_t MaskVal = MaskCnst->getZExtValue(); |
| 4861 | uint64_t NumZeros; |
| 4862 | uint64_t NumBits; |
| 4863 | if (isMask_64(MaskVal)) { |
| 4864 | NumZeros = 0; |
| 4865 | // The number of bits in the result bitfield will be the number of |
| 4866 | // trailing ones (the AND) minus the number of bits we shift off |
Benjamin Kramer | 5f6a907 | 2015-02-12 15:35:40 +0000 | [diff] [blame] | 4867 | NumBits = countTrailingOnes(MaskVal) - ShiftAmt; |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4868 | } else if (isShiftedMask_64(MaskVal)) { |
| 4869 | NumZeros = countTrailingZeros(MaskVal); |
Benjamin Kramer | 5f6a907 | 2015-02-12 15:35:40 +0000 | [diff] [blame] | 4870 | unsigned NumOnes = countTrailingOnes(MaskVal >> NumZeros); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4871 | // The number of bits in the result bitfield will be the number of |
| 4872 | // trailing zeros plus the number of set bits in the mask minus the |
| 4873 | // number of bits we shift off |
| 4874 | NumBits = NumZeros + NumOnes - ShiftAmt; |
| 4875 | } else { |
| 4876 | // This is not a mask we can handle |
| 4877 | return NULL; |
| 4878 | } |
| 4879 | |
| 4880 | if (ShiftAmt < NumZeros) { |
| 4881 | // Handling this case would require extra logic that would make this |
| 4882 | // transformation non-profitable |
| 4883 | return NULL; |
| 4884 | } |
| 4885 | |
| 4886 | Val = AndLHS; |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4887 | Start = CurDAG->getTargetConstant(ShiftAmt, DL, MVT::i32); |
| 4888 | Len = CurDAG->getTargetConstant(NumBits, DL, MVT::i32); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4889 | } else if (LHS->getOpcode() == ISD::SHL) { |
| 4890 | // Here, we have a pattern like: |
| 4891 | // |
| 4892 | // (sra (shl val, NN), MM) |
| 4893 | // or |
| 4894 | // (srl (shl val, NN), MM) |
| 4895 | // |
| 4896 | // If MM >= NN, we can efficiently optimize this with bfe |
| 4897 | Val = LHS->getOperand(0); |
| 4898 | |
| 4899 | SDValue ShlRHS = LHS->getOperand(1); |
| 4900 | ConstantSDNode *ShlCnst = dyn_cast<ConstantSDNode>(ShlRHS); |
| 4901 | if (!ShlCnst) { |
| 4902 | // Shift amount must be constant |
| 4903 | return NULL; |
| 4904 | } |
| 4905 | uint64_t InnerShiftAmt = ShlCnst->getZExtValue(); |
| 4906 | |
| 4907 | SDValue ShrRHS = RHS; |
| 4908 | ConstantSDNode *ShrCnst = dyn_cast<ConstantSDNode>(ShrRHS); |
| 4909 | if (!ShrCnst) { |
| 4910 | // Shift amount must be constant |
| 4911 | return NULL; |
| 4912 | } |
| 4913 | uint64_t OuterShiftAmt = ShrCnst->getZExtValue(); |
| 4914 | |
| 4915 | // To avoid extra codegen and be profitable, we need Outer >= Inner |
| 4916 | if (OuterShiftAmt < InnerShiftAmt) { |
| 4917 | return NULL; |
| 4918 | } |
| 4919 | |
| 4920 | // If the outer shift is more than the type size, we have no bitfield to |
| 4921 | // extract (since we also check that the inner shift is <= the outer shift |
| 4922 | // then this also implies that the inner shift is < the type size) |
| 4923 | if (OuterShiftAmt >= Val.getValueType().getSizeInBits()) { |
| 4924 | return NULL; |
| 4925 | } |
| 4926 | |
| 4927 | Start = |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4928 | CurDAG->getTargetConstant(OuterShiftAmt - InnerShiftAmt, DL, MVT::i32); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4929 | Len = |
| 4930 | CurDAG->getTargetConstant(Val.getValueType().getSizeInBits() - |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4931 | OuterShiftAmt, DL, MVT::i32); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4932 | |
| 4933 | if (N->getOpcode() == ISD::SRA) { |
| 4934 | // If we have a arithmetic right shift, we need to use the signed bfe |
| 4935 | // variant |
| 4936 | IsSigned = true; |
| 4937 | } |
| 4938 | } else { |
| 4939 | // No can do... |
| 4940 | return NULL; |
| 4941 | } |
| 4942 | } else { |
| 4943 | // No can do... |
| 4944 | return NULL; |
| 4945 | } |
| 4946 | |
| 4947 | |
| 4948 | unsigned Opc; |
| 4949 | // For the BFE operations we form here from "and" and "srl", always use the |
| 4950 | // unsigned variants. |
| 4951 | if (Val.getValueType() == MVT::i32) { |
| 4952 | if (IsSigned) { |
| 4953 | Opc = NVPTX::BFE_S32rii; |
| 4954 | } else { |
| 4955 | Opc = NVPTX::BFE_U32rii; |
| 4956 | } |
| 4957 | } else if (Val.getValueType() == MVT::i64) { |
| 4958 | if (IsSigned) { |
| 4959 | Opc = NVPTX::BFE_S64rii; |
| 4960 | } else { |
| 4961 | Opc = NVPTX::BFE_U64rii; |
| 4962 | } |
| 4963 | } else { |
| 4964 | // We cannot handle this type |
| 4965 | return NULL; |
| 4966 | } |
| 4967 | |
| 4968 | SDValue Ops[] = { |
| 4969 | Val, Start, Len |
| 4970 | }; |
| 4971 | |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 4972 | return CurDAG->getMachineNode(Opc, DL, N->getVTList(), Ops); |
Justin Holewinski | ca7a4f1 | 2014-06-27 18:35:27 +0000 | [diff] [blame] | 4973 | } |
| 4974 | |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 4975 | // SelectDirectAddr - Match a direct address for DAG. |
| 4976 | // A direct address could be a globaladdress or externalsymbol. |
| 4977 | bool NVPTXDAGToDAGISel::SelectDirectAddr(SDValue N, SDValue &Address) { |
| 4978 | // Return true if TGA or ES. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 4979 | if (N.getOpcode() == ISD::TargetGlobalAddress || |
| 4980 | N.getOpcode() == ISD::TargetExternalSymbol) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 4981 | Address = N; |
| 4982 | return true; |
| 4983 | } |
| 4984 | if (N.getOpcode() == NVPTXISD::Wrapper) { |
| 4985 | Address = N.getOperand(0); |
| 4986 | return true; |
| 4987 | } |
| 4988 | if (N.getOpcode() == ISD::INTRINSIC_WO_CHAIN) { |
| 4989 | unsigned IID = cast<ConstantSDNode>(N.getOperand(0))->getZExtValue(); |
| 4990 | if (IID == Intrinsic::nvvm_ptr_gen_to_param) |
| 4991 | if (N.getOperand(1).getOpcode() == NVPTXISD::MoveParam) |
| 4992 | return (SelectDirectAddr(N.getOperand(1).getOperand(0), Address)); |
| 4993 | } |
| 4994 | return false; |
| 4995 | } |
| 4996 | |
| 4997 | // symbol+offset |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 4998 | bool NVPTXDAGToDAGISel::SelectADDRsi_imp( |
| 4999 | SDNode *OpNode, SDValue Addr, SDValue &Base, SDValue &Offset, MVT mvt) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5000 | if (Addr.getOpcode() == ISD::ADD) { |
| 5001 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 5002 | SDValue base = Addr.getOperand(0); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5003 | if (SelectDirectAddr(base, Base)) { |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 5004 | Offset = CurDAG->getTargetConstant(CN->getZExtValue(), SDLoc(OpNode), |
| 5005 | mvt); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5006 | return true; |
| 5007 | } |
| 5008 | } |
| 5009 | } |
| 5010 | return false; |
| 5011 | } |
| 5012 | |
| 5013 | // symbol+offset |
| 5014 | bool NVPTXDAGToDAGISel::SelectADDRsi(SDNode *OpNode, SDValue Addr, |
| 5015 | SDValue &Base, SDValue &Offset) { |
| 5016 | return SelectADDRsi_imp(OpNode, Addr, Base, Offset, MVT::i32); |
| 5017 | } |
| 5018 | |
| 5019 | // symbol+offset |
| 5020 | bool NVPTXDAGToDAGISel::SelectADDRsi64(SDNode *OpNode, SDValue Addr, |
| 5021 | SDValue &Base, SDValue &Offset) { |
| 5022 | return SelectADDRsi_imp(OpNode, Addr, Base, Offset, MVT::i64); |
| 5023 | } |
| 5024 | |
| 5025 | // register+offset |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 5026 | bool NVPTXDAGToDAGISel::SelectADDRri_imp( |
| 5027 | SDNode *OpNode, SDValue Addr, SDValue &Base, SDValue &Offset, MVT mvt) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5028 | if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) { |
| 5029 | Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), mvt); |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 5030 | Offset = CurDAG->getTargetConstant(0, SDLoc(OpNode), mvt); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5031 | return true; |
| 5032 | } |
| 5033 | if (Addr.getOpcode() == ISD::TargetExternalSymbol || |
| 5034 | Addr.getOpcode() == ISD::TargetGlobalAddress) |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 5035 | return false; // direct calls. |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5036 | |
| 5037 | if (Addr.getOpcode() == ISD::ADD) { |
| 5038 | if (SelectDirectAddr(Addr.getOperand(0), Addr)) { |
| 5039 | return false; |
| 5040 | } |
| 5041 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) { |
| 5042 | if (FrameIndexSDNode *FIN = |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 5043 | dyn_cast<FrameIndexSDNode>(Addr.getOperand(0))) |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5044 | // Constant offset from frame ref. |
| 5045 | Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), mvt); |
| 5046 | else |
| 5047 | Base = Addr.getOperand(0); |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 5048 | Offset = CurDAG->getTargetConstant(CN->getZExtValue(), SDLoc(OpNode), |
| 5049 | mvt); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5050 | return true; |
| 5051 | } |
| 5052 | } |
| 5053 | return false; |
| 5054 | } |
| 5055 | |
| 5056 | // register+offset |
| 5057 | bool NVPTXDAGToDAGISel::SelectADDRri(SDNode *OpNode, SDValue Addr, |
| 5058 | SDValue &Base, SDValue &Offset) { |
| 5059 | return SelectADDRri_imp(OpNode, Addr, Base, Offset, MVT::i32); |
| 5060 | } |
| 5061 | |
| 5062 | // register+offset |
| 5063 | bool NVPTXDAGToDAGISel::SelectADDRri64(SDNode *OpNode, SDValue Addr, |
| 5064 | SDValue &Base, SDValue &Offset) { |
| 5065 | return SelectADDRri_imp(OpNode, Addr, Base, Offset, MVT::i64); |
| 5066 | } |
| 5067 | |
| 5068 | bool NVPTXDAGToDAGISel::ChkMemSDNodeAddressSpace(SDNode *N, |
| 5069 | unsigned int spN) const { |
Craig Topper | 062a2ba | 2014-04-25 05:30:21 +0000 | [diff] [blame] | 5070 | const Value *Src = nullptr; |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5071 | if (MemSDNode *mN = dyn_cast<MemSDNode>(N)) { |
Nick Lewycky | aad475b | 2014-04-15 07:22:52 +0000 | [diff] [blame] | 5072 | if (spN == 0 && mN->getMemOperand()->getPseudoValue()) |
| 5073 | return true; |
| 5074 | Src = mN->getMemOperand()->getValue(); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5075 | } |
| 5076 | if (!Src) |
| 5077 | return false; |
Craig Topper | e3dcce9 | 2015-08-01 22:20:21 +0000 | [diff] [blame^] | 5078 | if (auto *PT = dyn_cast<PointerType>(Src->getType())) |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5079 | return (PT->getAddressSpace() == spN); |
| 5080 | return false; |
| 5081 | } |
| 5082 | |
| 5083 | /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for |
| 5084 | /// inline asm expressions. |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 5085 | bool NVPTXDAGToDAGISel::SelectInlineAsmMemoryOperand( |
Daniel Sanders | 60f1db0 | 2015-03-13 12:45:09 +0000 | [diff] [blame] | 5086 | const SDValue &Op, unsigned ConstraintID, std::vector<SDValue> &OutOps) { |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5087 | SDValue Op0, Op1; |
Daniel Sanders | 60f1db0 | 2015-03-13 12:45:09 +0000 | [diff] [blame] | 5088 | switch (ConstraintID) { |
Justin Holewinski | 0497ab1 | 2013-03-30 14:29:21 +0000 | [diff] [blame] | 5089 | default: |
| 5090 | return true; |
Daniel Sanders | 60f1db0 | 2015-03-13 12:45:09 +0000 | [diff] [blame] | 5091 | case InlineAsm::Constraint_m: // memory |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5092 | if (SelectDirectAddr(Op, Op0)) { |
| 5093 | OutOps.push_back(Op0); |
Sergey Dmitrouk | 842a51b | 2015-04-28 14:05:47 +0000 | [diff] [blame] | 5094 | OutOps.push_back(CurDAG->getTargetConstant(0, SDLoc(Op), MVT::i32)); |
Justin Holewinski | ae556d3 | 2012-05-04 20:18:50 +0000 | [diff] [blame] | 5095 | return false; |
| 5096 | } |
| 5097 | if (SelectADDRri(Op.getNode(), Op, Op0, Op1)) { |
| 5098 | OutOps.push_back(Op0); |
| 5099 | OutOps.push_back(Op1); |
| 5100 | return false; |
| 5101 | } |
| 5102 | break; |
| 5103 | } |
| 5104 | return true; |
| 5105 | } |