| 1 |
/*========================== begin_copyright_notice ============================ |
| 2 |
|
| 3 |
Copyright (C) 2017-2021 Intel Corporation |
| 4 |
|
| 5 |
SPDX-License-Identifier: MIT |
| 6 |
|
| 7 |
============================= end_copyright_notice ===========================*/ |
| 8 |
|
| 9 |
#include "Encoder.hpp" |
| 10 |
#include "../../Frontend/IRToString.hpp" |
| 11 |
#include "../../IR/Kernel.hpp" |
| 12 |
#include "../../IR/SWSBSetter.hpp" |
| 13 |
#include "../../Models/Models.hpp" |
| 14 |
#include "../../Timer/Timer.hpp" |
| 15 |
#include "../../bits.hpp" |
| 16 |
#include "../../strings.hpp" |
| 17 |
#include "IGAToGEDTranslation.hpp" |
| 18 |
|
| 19 |
#include <cstring> |
| 20 |
|
| 21 |
using namespace iga; |
| 22 |
|
| 23 |
|
| 24 |
static const char *gedReturnValueToString(GED_RETURN_VALUE rv) { |
| 25 |
switch (rv) { |
| 26 |
case GED_RETURN_VALUE_SUCCESS: |
| 27 |
return "success"; |
| 28 |
case GED_RETURN_VALUE_CYCLIC_DEPENDENCY: |
| 29 |
return "cyclic dependency"; |
| 30 |
case GED_RETURN_VALUE_NULL_POINTER: |
| 31 |
return "null pointer"; |
| 32 |
case GED_RETURN_VALUE_OPCODE_NOT_SUPPORTED: |
| 33 |
return "unsupported opcode"; |
| 34 |
case GED_RETURN_VALUE_NO_COMPACT_FORM: |
| 35 |
return "no compact form"; |
| 36 |
case GED_RETURN_VALUE_INVALID_FIELD: |
| 37 |
return "invalid field"; |
| 38 |
case GED_RETURN_VALUE_INVALID_VALUE: |
| 39 |
return "invalid value"; |
| 40 |
case GED_RETURN_VALUE_INVALID_INTERPRETER: |
| 41 |
return "invalid interpreter"; |
| 42 |
default: |
| 43 |
return "other error"; |
| 44 |
} |
| 45 |
} |
| 46 |
void Encoder::handleGedError(int line, const char *setter, |
| 47 |
GED_RETURN_VALUE status) { |
| 48 |
errorT("IGALibrary/GED/Encoder.cpp:", line, ": GED_Set", setter, ": ", |
| 49 |
gedReturnValueToString(status)); |
| 50 |
} |
| 51 |
|
| 52 |
Encoder::Encoder(const Model &model, ErrorHandler &errHandler, |
| 53 |
const EncoderOpts &opts) |
| 54 |
: GEDBitProcessor(model, errHandler), m_opts(opts), |
| 55 |
m_numberInstructionsEncoded(0), m_mem(nullptr) { |
| 56 |
// derive the swsb encoding mode from platform if not set |
| 57 |
if (opts.swsbEncodeMode == SWSB_ENCODE_MODE::SWSBInvalidMode) { |
| 58 |
m_opts.swsbEncodeMode = model.getSWSBEncodeMode(); |
| 59 |
} |
| 60 |
} |
| 61 |
|
| 62 |
void Encoder::encodeKernelPreProcess(Kernel &k) { doEncodeKernelPreProcess(k); } |
| 63 |
|
| 64 |
void Encoder::doEncodeKernelPreProcess(Kernel &k) { |
| 65 |
if (m_opts.autoDepSet && platform() >= Platform::XE) { |
| 66 |
SWSBAnalyzer swsbAnalyzer(k, errorHandler(), m_opts.swsbEncodeMode, |
| 67 |
m_opts.sbidCount, m_opts.autoSBIDCounter); |
| 68 |
swsbAnalyzer.run(); |
| 69 |
} |
| 70 |
} |
| 71 |
|
| 72 |
double Encoder::getElapsedTimeMicros(unsigned int idx) { |
| 73 |
return getIGATimerUS(idx); |
| 74 |
} |
| 75 |
|
| 76 |
int64_t Encoder::getElapsedTimeTicks(unsigned int idx) { |
| 77 |
return getIGATimerTicks(idx); |
| 78 |
} |
| 79 |
|
| 80 |
std::string Encoder::getTimerName(unsigned int idx) { |
| 81 |
return getIGATimerNames(idx); |
| 82 |
} |
| 83 |
|
| 84 |
size_t Encoder::getNumInstructionsEncoded() const { |
| 85 |
return m_numberInstructionsEncoded; |
| 86 |
} |
| 87 |
|
| 88 |
void Encoder::encodeKernel(Kernel &k, MemManager &mem, void *&bits, |
| 89 |
uint32_t &bitsLen) { |
| 90 |
#ifndef IGA_DISABLE_ENCODER_EXCEPTIONS |
| 91 |
try { |
| 92 |
#endif |
| 93 |
initIGATimer(); |
| 94 |
setIGAKernelName("test"); |
| 95 |
IGA_ASSERT(k.getModel().platform == platform(), |
| 96 |
"kernel/encoder model mismatch"); |
| 97 |
|
| 98 |
encodeKernelPreProcess(k); |
| 99 |
m_needToPatch.clear(); |
| 100 |
m_mem = &mem; |
| 101 |
m_numberInstructionsEncoded = k.getInstructionCount(); |
| 102 |
size_t allocLen = m_numberInstructionsEncoded * UNCOMPACTED_SIZE; |
| 103 |
if (allocLen == 0) // for empty kernel case |
| 104 |
allocLen = 4; |
| 105 |
m_instBuf = (uint8_t *)mem.alloc(allocLen); |
| 106 |
if (!m_instBuf) { |
| 107 |
fatalAtT(0, "failed to allocate memory for kernel binary"); |
| 108 |
return; |
| 109 |
} |
| 110 |
|
| 111 |
for (auto blk : k.getBlockList()) { |
| 112 |
START_ENCODER_TIMER(); |
| 113 |
encodeBlock(k, blk); |
| 114 |
STOP_ENCODER_TIMER(); |
| 115 |
if (hasFatalError()) { |
| 116 |
return; |
| 117 |
} |
| 118 |
} |
| 119 |
START_ENCODER_TIMER(); |
| 120 |
patchJumpOffsets(); |
| 121 |
STOP_ENCODER_TIMER(); |
| 122 |
|
| 123 |
// setting actual size |
| 124 |
bitsLen = currentPc(); |
| 125 |
bits = m_instBuf; |
| 126 |
|
| 127 |
applyGedWorkarounds(k, currentPc()); |
| 128 |
|
| 129 |
// clear any padding |
| 130 |
memset(m_instBuf + bitsLen, 0, allocLen - bitsLen); |
| 131 |
#ifndef IGA_DISABLE_ENCODER_EXCEPTIONS |
| 132 |
} catch (const iga::FatalError &) { |
| 133 |
// error is already reported |
| 134 |
} |
| 135 |
#endif |
| 136 |
} |
| 137 |
|
| 138 |
void Encoder::encodeInlineBinaryInst(Instruction &inst) { |
| 139 |
setEncodedPC(&inst, currentPc()); |
| 140 |
for (auto i = 0, j = 3; i < 4; ++i, --j) { |
| 141 |
uint8_t *src = (uint8_t *)&inst.getInlineBinary().at(j); |
| 142 |
uint8_t *dst = m_instBuf + currentPc() + i * 4; |
| 143 |
memcpy_s(dst, 4, src, 4); |
| 144 |
} |
| 145 |
|
| 146 |
// Inline-binary-instruction must have size 16 |
| 147 |
advancePc(16); |
| 148 |
} |
| 149 |
|
| 150 |
void Encoder::encodeBlock(Kernel &k, Block *blk) { |
| 151 |
m_blockToOffsetMap[blk] = currentPc(); |
| 152 |
|
| 153 |
InstList &instList = blk->getInstList(); |
| 154 |
const auto instListEnd = instList.end(); |
| 155 |
for (auto instIter = instList.begin(); instIter != instListEnd; ++instIter) { |
| 156 |
Instruction *inst = *instIter; |
| 157 |
|
| 158 |
if (inst->hasInstOpt(InstOpt::CACHELINEALIGN)) { |
| 159 |
while (currentPc() / 64 != (currentPc() + 31) / 64) { |
| 160 |
SWSB swsb(SWSB::DistType::NO_DIST, SWSB::TokenType::NOTOKEN, 0, 0); |
| 161 |
Instruction *syncInst = k.createSyncNopInstruction(swsb); |
| 162 |
setCurrInst(syncInst); |
| 163 |
encodeInstruction(*syncInst); |
| 164 |
if (hasFatalError()) { |
| 165 |
return; |
| 166 |
} |
| 167 |
setEncodedPC(syncInst, currentPc()); |
| 168 |
GED_RETURN_VALUE status = GED_RETURN_VALUE_SIZE; |
| 169 |
status = GED_EncodeIns(&m_gedInst, GED_INS_TYPE_NATIVE, |
| 170 |
m_instBuf + currentPc()); |
| 171 |
if (status != GED_RETURN_VALUE_SUCCESS) { |
| 172 |
errorAtT(inst->getLoc(), "GED unable to encode instruction: ", |
| 173 |
gedReturnValueToString(status)); |
| 174 |
} |
| 175 |
blk->insertInstBefore(instIter, syncInst); |
| 176 |
advancePc(16); |
| 177 |
} |
| 178 |
} |
| 179 |
|
| 180 |
setCurrInst(inst); |
| 181 |
|
| 182 |
if (inst->isInlineBinaryInstruction()) { |
| 183 |
encodeInlineBinaryInst(*inst); |
| 184 |
continue; |
| 185 |
} |
| 186 |
|
| 187 |
encodeInstruction(*inst); |
| 188 |
if (hasFatalError()) { |
| 189 |
return; |
| 190 |
} |
| 191 |
setEncodedPC(inst, currentPc()); |
| 192 |
|
| 193 |
GED_RETURN_VALUE status = GED_RETURN_VALUE_SIZE; |
| 194 |
|
| 195 |
// If -Xforce-no-compact is set, do not compact any insruction |
| 196 |
// Otherwise, if {NoCompact} is set, do not compact the instruction |
| 197 |
// Otherwise, if {Compacted} is set on the instruction, try to compact it |
| 198 |
// and report error on fail Otherwise, if no compaction setting on the |
| 199 |
// instruction, try to compact the instruction if -Xauto-compact Otherwise, |
| 200 |
// do not compact the instruction |
| 201 |
bool mustCompact = inst->hasInstOpt(InstOpt::COMPACTED); |
| 202 |
bool mustNotCompact = inst->hasInstOpt(InstOpt::NOCOMPACT); |
| 203 |
if (m_opts.forceNoCompact) { |
| 204 |
mustCompact = false; |
| 205 |
mustNotCompact = true; |
| 206 |
} |
| 207 |
|
| 208 |
int32_t iLen = 16; |
| 209 |
if (mustCompact || (!mustNotCompact && m_opts.autoCompact)) { |
| 210 |
// try compact first |
| 211 |
status = GED_EncodeIns(&m_gedInst, GED_INS_TYPE_COMPACT, |
| 212 |
m_instBuf + currentPc()); |
| 213 |
if (status == GED_RETURN_VALUE_SUCCESS) { |
| 214 |
// If auto compation is turned on, in case we need to patch later. |
| 215 |
inst->addInstOpt(InstOpt::COMPACTED); |
| 216 |
iLen = 8; |
| 217 |
} else if (status == GED_RETURN_VALUE_NO_COMPACT_FORM) { |
| 218 |
if (mustCompact) { |
| 219 |
if (m_opts.explicitCompactMissIsWarning) { |
| 220 |
warningAtT(inst->getLoc(), "GED unable to compact instruction"); |
| 221 |
} else { |
| 222 |
errorAtT(inst->getLoc(), "GED unable to compact instruction"); |
| 223 |
} |
| 224 |
} |
| 225 |
} // else: some other error (unreachable?) |
| 226 |
} |
| 227 |
|
| 228 |
// try native encoding if compaction failed |
| 229 |
if (status != GED_RETURN_VALUE_SUCCESS) { |
| 230 |
inst->removeInstOpt(InstOpt::COMPACTED); |
| 231 |
status = GED_EncodeIns(&m_gedInst, GED_INS_TYPE_NATIVE, |
| 232 |
m_instBuf + currentPc()); |
| 233 |
if (status != GED_RETURN_VALUE_SUCCESS) { |
| 234 |
errorAtT(inst->getLoc(), "GED unable to encode instruction: ", |
| 235 |
gedReturnValueToString(status)); |
| 236 |
} |
| 237 |
} |
| 238 |
advancePc(iLen); |
| 239 |
} |
| 240 |
} |
| 241 |
|
| 242 |
bool Encoder::getBlockOffset(const Block *b, uint32_t &pc) { |
| 243 |
auto iter = m_blockToOffsetMap.find(b); |
| 244 |
if (iter != m_blockToOffsetMap.end()) { |
| 245 |
pc = iter->second; |
| 246 |
return true; |
| 247 |
} |
| 248 |
return false; |
| 249 |
} |
| 250 |
|
| 251 |
void Encoder::setEncodedPC(Instruction *inst, int32_t encodedPC) { |
| 252 |
#if 0 |
| 253 |
auto iter = m_instPcs.find(inst); |
| 254 |
IGA_ASSERT(iter == m_instPcs.end(), "resetting encode PC"); |
| 255 |
((void)iter); // dummy use where ASSERT_USER compiles out |
| 256 |
m_instPcs[inst] = encodedPC; |
| 257 |
#else |
| 258 |
inst->setPC(encodedPC); |
| 259 |
#endif |
| 260 |
} |
| 261 |
|
| 262 |
int32_t Encoder::getEncodedPC(const Instruction *inst) const { |
| 263 |
#if 0 |
| 264 |
auto iter = m_instPcs.find(inst); |
| 265 |
if (iter == m_instPcs.end()) { |
| 266 |
IGA_ASSERT_FALSE("inst PC not found"); |
| 267 |
return 0; |
| 268 |
} |
| 269 |
return iter->second; |
| 270 |
#else |
| 271 |
return inst->getPC(); |
| 272 |
#endif |
| 273 |
} |
| 274 |
|
| 275 |
void Encoder::encodeFC(const Instruction &i) { |
| 276 |
const OpSpec &os = i.getOpSpec(); |
| 277 |
|
| 278 |
if (os.is(Op::MATH)) { |
| 279 |
GED_MATH_FC mfc = lowerMathFC(i.getMathFc()); |
| 280 |
GED_ENCODE(MathFC, mfc); |
| 281 |
} else if (os.is(Op::BFN)) { |
| 282 |
GED_ENCODE(BfnFC, i.getBfnFc().value); |
| 283 |
} else if (os.isDpasFormat()) { |
| 284 |
auto sf = i.getDpasFc(); |
| 285 |
auto sdepth = GetDpasSystolicDepth(sf); |
| 286 |
auto rc = GetDpasRepeatCount(sf); |
| 287 |
if (os.is(Op::BDPAS)) { |
| 288 |
// in bdpas, repeat count is fixed to 8 |
| 289 |
// there is no instruction encoding bits for repeat count |
| 290 |
// Note that bdpas's enum class for systolic depth is different from |
| 291 |
// dpas'. Hence GED has different API name (GED_GetSystolicDepthBlkScl). |
| 292 |
GED_ENCODE(SystolicDepthBlkScl, sdepth); |
| 293 |
} else { |
| 294 |
GED_ENCODE(SystolicDepth, sdepth); |
| 295 |
GED_ENCODE(RepeatCount, rc); |
| 296 |
} |
| 297 |
} else if (os.isAnySendFormat()) { |
| 298 |
if (platform() >= Platform::XE) { |
| 299 |
// on earlier platforms this is stowed in ExDesc |
| 300 |
auto sfid = lowerSFID(i.getSendFc()); |
| 301 |
GED_ENCODE(SFID, sfid); |
| 302 |
} |
| 303 |
} else if (os.is(Op::SYNC)) { |
| 304 |
GED_SYNC_FC wfc = lowerSyncFC(i.getSyncFc()); |
| 305 |
GED_ENCODE(SyncFC, wfc); |
| 306 |
} else if (os.is(Op::SHFL)) { |
| 307 |
GED_SHUFFLE_FC sfc = lowerShuffleFC(i.getShuffleFc()); |
| 308 |
GED_ENCODE(ShuffleFC, sfc); |
| 309 |
} else if (os.is(Op::LFSR)) { |
| 310 |
GED_LFSR_FC fc = lowerLfsrFC(i.getLfsrFc()); |
| 311 |
GED_ENCODE(LfsrFC, fc); |
| 312 |
} else if (os.is(Op::DNSCL)) { |
| 313 |
auto fc = i.getDnsclFc(); |
| 314 |
GED_ENCODE(ConvSrcDataType, lowerConvSrcDataType(fc.convSrcTy)); |
| 315 |
GED_ENCODE(ConvDstDataType, lowerConvDstDataType(fc.convDstTy)); |
| 316 |
GED_ENCODE(DnsclMode, lowerDnsclMode(fc.dnsclMode)); |
| 317 |
GED_ENCODE(RoundingMode, lowerRoundingMode(fc.roundingMode)); |
| 318 |
} else if (os.supportsBranchCtrl()) { |
| 319 |
GED_ENCODE(BranchCtrl, lowerBranchCntrl(i.getBranchCtrl())); |
| 320 |
} else if (os.supportsSubfunction()) { |
| 321 |
IGA_ASSERT_FALSE("encoder needs to encode subfunction for this op"); |
| 322 |
} |
| 323 |
} |
| 324 |
|
| 325 |
void Encoder::encodeInstruction(Instruction &inst) { |
| 326 |
m_opcode = inst.getOp(); |
| 327 |
const auto gedPlat = lowerPlatform(platform()); |
| 328 |
const auto gedOp = lowerOpcode(m_opcode); |
| 329 |
if (gedOp == GED_OPCODE_INVALID) { |
| 330 |
fatalAtT(inst.getLoc(), "OpSpec failed to map to a valid GED opcode"); |
| 331 |
return; |
| 332 |
} |
| 333 |
GED_RETURN_VALUE status = GED_InitEmptyIns(gedPlat, &m_gedInst, gedOp); |
| 334 |
if (status != GED_RETURN_VALUE_SUCCESS) { |
| 335 |
fatalAtT(inst.getLoc(), "GED failed to create instruction template"); |
| 336 |
return; |
| 337 |
} |
| 338 |
|
| 339 |
if (m_opcode == Op::ILLEGAL || m_opcode == Op::THRYLD) { |
| 340 |
// GED does all the work for this instruction |
| 341 |
return; |
| 342 |
} else if (m_opcode == Op::NOP) { |
| 343 |
// nop supports {Breakpoint} |
| 344 |
encodeOptions(inst); |
| 345 |
return; |
| 346 |
} |
| 347 |
const OpSpec &os = inst.getOpSpec(); |
| 348 |
|
| 349 |
// Dwindling cases where we must use Align16 |
| 350 |
// Pre-GEN10 ternary ops are all align16 |
| 351 |
bool isTernary = platform() < Platform::GEN10 && os.isTernary(); |
| 352 |
bool contextSaveRestoreNeedsAlign16 = |
| 353 |
isAlign16MathMacroRegisterCsrOperand(inst.getDestination()) || |
| 354 |
isAlign16MathMacroRegisterCsrOperand(inst.getSource(0)); |
| 355 |
// IEEE macro instructions (math.invm and math.rsqrtm) |
| 356 |
bool align16MacroInst = m_model.supportsAlign16MacroInst() && inst.isMacro(); |
| 357 |
m_encodeAlign16 = |
| 358 |
isTernary || contextSaveRestoreNeedsAlign16 || align16MacroInst; |
| 359 |
GED_ACCESS_MODE accessMode = |
| 360 |
m_encodeAlign16 ? GED_ACCESS_MODE_Align16 : GED_ACCESS_MODE_Align1; |
| 361 |
if (m_model.supportsAccessMode()) { |
| 362 |
GED_ENCODE(AccessMode, accessMode); |
| 363 |
} // else GED will crash given this call (even given Align1) |
| 364 |
|
| 365 |
ExecSize execSize = inst.getExecSize(); |
| 366 |
if (os.isTernary() && m_model.supportsAlign16Ternary() && |
| 367 |
inst.getExecSize() == ExecSize::SIMD1) { |
| 368 |
// scalar ternary workaround for Align16 |
| 369 |
// (c.f. Encoder::encodeTernaryDestinationAlign16) |
| 370 |
execSize = inst.getDestination().getType() == Type::DF ? ExecSize::SIMD2 |
| 371 |
: ExecSize::SIMD4; |
| 372 |
} |
| 373 |
GED_ENCODE(ExecSize, lowerExecSize(execSize)); |
| 374 |
|
| 375 |
encodeFC(inst); |
| 376 |
|
| 377 |
if (os.supportsQtrCtrl()) { |
| 378 |
// use ExecSize from above since it may have been modified |
| 379 |
GED_CHANNEL_OFFSET qtrCtrl = lowerQtrCtrl(inst.getChannelOffset()); |
| 380 |
GED_ENCODE(ChannelOffset, qtrCtrl); |
| 381 |
} |
| 382 |
|
| 383 |
GED_ENCODE(MaskCtrl, lowerEmask(inst.getMaskCtrl())); |
| 384 |
|
| 385 |
// Predicate |
| 386 |
const Predication &pred = inst.getPredication(); |
| 387 |
if (os.supportsPredication()) { |
| 388 |
GED_ENCODE(PredCtrl, lowerPredCtrl(pred.function)); |
| 389 |
} else { |
| 390 |
GED_ENCODE(PredCtrl, GED_PRED_CTRL_Normal); |
| 391 |
} |
| 392 |
|
| 393 |
bool isImm64Src0Overlap = |
| 394 |
platform() >= Platform::XE && |
| 395 |
inst.getSource(0).getKind() == Operand::Kind::IMMEDIATE && |
| 396 |
TypeIs64b(inst.getSource(0).getType()); |
| 397 |
|
| 398 |
if (!isImm64Src0Overlap && inst.getOpSpec().supportsFlagModifier()) { |
| 399 |
if (os.op == Op::BFN) { |
| 400 |
switch (inst.getFlagModifier()) { |
| 401 |
case FlagModifier::NONE: |
| 402 |
case FlagModifier::EQ: |
| 403 |
case FlagModifier::GT: |
| 404 |
case FlagModifier::LT: |
| 405 |
// GED does the special mapping to CondMod2 |
| 406 |
// only a subset of cond modifiers are supported on this op |
| 407 |
GED_ENCODE(CondModifier, lowerCondModifier(inst.getFlagModifier())); |
| 408 |
break; |
| 409 |
default: |
| 410 |
errorT("this instruction format only supports " |
| 411 |
"(eq), (gt), and (lt) conditional modifiers"); |
| 412 |
} |
| 413 |
} else { |
| 414 |
GED_ENCODE(CondModifier, lowerCondModifier(inst.getFlagModifier())); |
| 415 |
} |
| 416 |
} |
| 417 |
|
| 418 |
bool hasFlagRegField = true; |
| 419 |
// For >= XE_HPC, Some fields only exist when having CondCtrl or PredCtrl: |
| 420 |
// PredInv, FlagRegNum, FlagSubRegNum |
| 421 |
// In GED, either CondCtrl or PredCtrl have to be set to non-zero before |
| 422 |
// these fields can be set |
| 423 |
if (platform() >= Platform::XE_HPC) { |
| 424 |
hasFlagRegField = (inst.getFlagModifier() != FlagModifier::NONE) || |
| 425 |
(pred.function != PredCtrl::NONE) || inst.isBranching(); |
| 426 |
} |
| 427 |
|
| 428 |
if (os.supportsPredication() && hasFlagRegField) |
| 429 |
GED_ENCODE(PredInv, |
| 430 |
pred.inverse ? GED_PRED_INV_Invert : GED_PRED_INV_Normal); |
| 431 |
|
| 432 |
// GED_ExecutionDataType |
| 433 |
RegRef flagReg = inst.getFlagReg(); |
| 434 |
if (hasFlagRegField && (flagReg != REGREF_INVALID)) { |
| 435 |
GED_ENCODE(FlagRegNum, static_cast<uint32_t>(inst.getFlagReg().regNum)); |
| 436 |
GED_ENCODE(FlagSubRegNum, inst.getFlagReg().subRegNum); |
| 437 |
} |
| 438 |
|
| 439 |
// set AccWrEn where supported |
| 440 |
if (inst.hasInstOpt(InstOpt::ACCWREN)) { |
| 441 |
GED_ENCODE(AccWrCtrl, GED_ACC_WR_CTRL_AccWrEn); |
| 442 |
} |
| 443 |
|
| 444 |
if (os.isBranching()) { |
| 445 |
if (m_model.supportsSimplifiedBranches()) { |
| 446 |
encodeBranchingInstructionSimplified(inst); |
| 447 |
} else { |
| 448 |
encodeBranchingInstruction(inst); |
| 449 |
} |
| 450 |
// options encoded internally |
| 451 |
} else if (os.isTernary()) { |
| 452 |
encodeTernaryInstruction(inst, accessMode); |
| 453 |
} else if (os.isQuinary()) { |
| 454 |
encodeQuinaryInstruction(inst); |
| 455 |
} else if (os.isAnySendFormat()) { |
| 456 |
encodeSendInstruction(inst); |
| 457 |
} else if (os.is(Op::SYNC)) { |
| 458 |
encodeSyncInstruction(inst); |
| 459 |
} else { |
| 460 |
encodeBasicInstruction(inst, accessMode); |
| 461 |
} |
| 462 |
|
| 463 |
if (!hasFatalError()) { |
| 464 |
encodeOptions(inst); |
| 465 |
|
| 466 |
// setup for back patching on branching ops |
| 467 |
if (os.isBranching() || inst.isMovWithLabel()) { |
| 468 |
bool src0IsLabel = inst.getSource(0).isImm(); |
| 469 |
bool src1IsLabel = inst.getSourceCount() > 1 && inst.getSource(1).isImm(); |
| 470 |
if (src0IsLabel || src1IsLabel) { |
| 471 |
m_needToPatch.emplace_back(&inst, m_gedInst, m_instBuf + currentPc()); |
| 472 |
// Force not to compact label instructions to avoid the compaction error |
| 473 |
// when auto-compaction is enabled. We could set this inst to |
| 474 |
// compactable during Encoder::encodeBlock when the value is unknown |
| 475 |
// (and assume to be 0). But we can only compact imm values use up to 12 |
| 476 |
// bits. |
| 477 |
inst.addInstOpt(InstOpt::NOCOMPACT); |
| 478 |
} |
| 479 |
} |
| 480 |
} |
| 481 |
} |
| 482 |
|
| 483 |
void Encoder::encodeBasicInstruction(const Instruction &inst, |
| 484 |
GED_ACCESS_MODE accessMode) { |
| 485 |
const OpSpec &os = inst.getOpSpec(); |
| 486 |
|
| 487 |
if (os.isBinaryWithExecDataType()) |
| 488 |
GED_ENCODE(ExecutionDataType, |
| 489 |
lowerExecDataType(inst.getDestination().getType())); |
| 490 |
|
| 491 |
switch (inst.getSourceCount()) { |
| 492 |
case 2: |
| 493 |
encodeBasicSource<SourceIndex::SRC1>(inst, inst.getSource(1), accessMode); |
| 494 |
// vvvv fall through vvvv |
| 495 |
case 1: |
| 496 |
encodeBasicSource<SourceIndex::SRC0>(inst, inst.getSource(0), accessMode); |
| 497 |
} |
| 498 |
|
| 499 |
if (os.supportsDestination()) { |
| 500 |
encodeBasicDestination(inst, inst.getDestination(), accessMode); |
| 501 |
if (os.op == Op::BDPAS) { |
| 502 |
GED_ENCODE(ExecutionDataType, |
| 503 |
lowerExecDataType(inst.getDestination().getType())); |
| 504 |
} |
| 505 |
} else if (os.op == Op::WAIT) { |
| 506 |
// wait has an implicit destination (same as first source) |
| 507 |
// but with dst region of <1> |
| 508 |
Operand copy(inst.getSource(0)); |
| 509 |
copy.setRegion(Region::DST1); |
| 510 |
encodeBasicDestination(inst, copy); |
| 511 |
} |
| 512 |
} |
| 513 |
|
| 514 |
void Encoder::encodeTernaryDestinationAlign1(const Instruction &inst) { |
| 515 |
const Operand &dst = inst.getDestination(); |
| 516 |
|
| 517 |
if (inst.getOpSpec().supportsSaturation()) { |
| 518 |
GED_ENCODE(Saturate, lowerSaturate(dst.getDstModifier())); |
| 519 |
} |
| 520 |
GED_ENCODE(DstDataType, lowerDataType(dst.getType())); |
| 521 |
GED_ENCODE(DstRegFile, lowerRegFile(dst.getDirRegName())); |
| 522 |
encodeDstReg(dst.getDirRegName(), dst.getDirRegRef().regNum); |
| 523 |
|
| 524 |
if (inst.isMacro()) { |
| 525 |
GED_ENCODE(DstMathMacroExt, lowerMathMacroReg(dst.getMathMacroExt())); |
| 526 |
// GED_ENCODE(DstHorzStride, 1); |
| 527 |
} else { |
| 528 |
GED_ENCODE(DstSubRegNum, |
| 529 |
SubRegToBinaryOffset(dst.getDirRegRef().subRegNum, |
| 530 |
dst.getDirRegName(), dst.getType(), |
| 531 |
m_model.platform)); |
| 532 |
if (inst.getOpSpec().hasDstHorzStride()) |
| 533 |
GED_ENCODE(DstHorzStride, static_cast<int>(dst.getRegion().getHz())); |
| 534 |
} |
| 535 |
} |
| 536 |
template <SourceIndex S> |
| 537 |
void Encoder::encodeQuinarySource(const Instruction &inst) { |
| 538 |
assert(platform() > Platform::XE3 && |
| 539 |
"Quinary sources not supported on platform"); |
| 540 |
assert(inst.getOpSpec().isBDPAS() && |
| 541 |
"Quinary source not supported for instruction; only supported for bdpas"); |
| 542 |
|
| 543 |
const Operand &src = inst.getSource(S); |
| 544 |
Type srcType = src.getType(); |
| 545 |
|
| 546 |
auto getSubRegOff = [&]() { |
| 547 |
return SubRegToBinaryOffset(src.getDirRegRef().subRegNum, |
| 548 |
src.getDirRegName(), srcType, |
| 549 |
m_model.platform); |
| 550 |
}; |
| 551 |
if (S == SourceIndex::SRC0) { |
| 552 |
GED_ENCODE(Src0DataType, lowerDataType(srcType)); |
| 553 |
encodeSrcSubRegNum<S>(getSubRegOff()); |
| 554 |
} else if (S == SourceIndex::SRC1) { |
| 555 |
GED_ENCODE(Src1Precision, lowerSubBytePrecision(srcType)); |
| 556 |
} else if (S == SourceIndex::SRC2) { |
| 557 |
GED_ENCODE(Src2Precision, lowerSubBytePrecision(srcType)); |
| 558 |
} else if (S == SourceIndex::SRC3) { |
| 559 |
encodeSrcSubRegNum<S>(getSubRegOff()); |
| 560 |
} else if (S == SourceIndex::SRC4) { |
| 561 |
encodeSrcSubRegNum<S>(getSubRegOff()); |
| 562 |
} |
| 563 |
encodeSrcRegFile<S>(lowerRegFile(src.getDirRegName())); |
| 564 |
encodeSrcReg<S>(src.getDirRegName(), src.getDirRegRef().regNum); |
| 565 |
} |
| 566 |
|
| 567 |
template <SourceIndex S> |
| 568 |
void Encoder::encodeTernarySourceAlign1(const Instruction &inst) { |
| 569 |
// CNL+ align1 ternary |
| 570 |
if (platform() < Platform::GEN10) { |
| 571 |
fatalT("src", (int)S, |
| 572 |
": align1 ternary is not supported on this " |
| 573 |
"platform"); |
| 574 |
return; |
| 575 |
} |
| 576 |
|
| 577 |
const Operand &src = inst.getSource(S); |
| 578 |
Type srcType = src.getType(); |
| 579 |
// DPAS |
| 580 |
if (inst.getOpSpec().isDpasFormat()) { |
| 581 |
// src0's type is the type for all sources |
| 582 |
|
| 583 |
if (S == SourceIndex::SRC0) { |
| 584 |
GED_ENCODE(Src0DataType, lowerDataType(srcType)); |
| 585 |
// GED: src0 HS = 0, VS=3 |
| 586 |
} else if (S == SourceIndex::SRC1) { |
| 587 |
GED_ENCODE(Src1Precision, lowerSubBytePrecision(srcType)); |
| 588 |
// GED sets both the type and the precision at the same time for us |
| 589 |
// GED: src1 HS = 1, VS=3 |
| 590 |
// via this higher-level API |
| 591 |
} else if (S == SourceIndex::SRC2) { |
| 592 |
GED_ENCODE(Src2Precision, lowerSubBytePrecision(srcType)); |
| 593 |
// GED: src2 HS = 3 |
| 594 |
// GED sets both the type and the precision at the same time for us |
| 595 |
// via this higher-level API |
| 596 |
} |
| 597 |
encodeSrcRegFile<S>(lowerRegFile(src.getDirRegName())); |
| 598 |
encodeSrcReg<S>(src.getDirRegName(), src.getDirRegRef().regNum); |
| 599 |
encodeSrcSubRegNum<S>(SubRegToBinaryOffset(src.getDirRegRef().subRegNum, |
| 600 |
src.getDirRegName(), srcType, |
| 601 |
m_model.platform)); |
| 602 |
|
| 603 |
return; |
| 604 |
} |
| 605 |
|
| 606 |
// GED will catch any mismatch between float and int (illegal mixed mode) |
| 607 |
encodeSrcType<S>(srcType); // GED dependency requires type before reg file |
| 608 |
|
| 609 |
switch (src.getKind()) { |
| 610 |
case Operand::Kind::DIRECT: |
| 611 |
case Operand::Kind::MACRO: { |
| 612 |
encodeSrcRegFile<S>(lowerRegFile(src.getDirRegName())); |
| 613 |
|
| 614 |
if (platform() <= Platform::GEN11) { |
| 615 |
encodeSrcAddrMode<S>(GED_ADDR_MODE_Direct); |
| 616 |
} |
| 617 |
|
| 618 |
// source modifiers |
| 619 |
if (inst.getOpSpec().supportsSourceModifiers()) { |
| 620 |
encodeSrcModifier<S>(src.getSrcModifier()); |
| 621 |
} |
| 622 |
|
| 623 |
// regioning |
| 624 |
// |
| 625 |
// ternary align1 puts SpcAcc into subreg, so regions may be set |
| 626 |
// in all cases |
| 627 |
// * madm doesn't have a region in GEN9 ... |
| 628 |
// it does in GEN10+, but we haven't supported it in syntax yet |
| 629 |
// and leave it to GED to set it |
| 630 |
// * src0 and src1 only has <w;h>, src2 only has <h> |
| 631 |
bool hasSrcRgnHz = !inst.isMacro(); |
| 632 |
bool hasSrcRgnVt = !inst.isMacro() && S < SourceIndex::SRC2; |
| 633 |
bool hasReducedRegion = |
| 634 |
m_model.srcHasReducedRegion(static_cast<uint32_t>(S)); |
| 635 |
auto rgn = src.getRegion(); |
| 636 |
if (hasReducedRegion) { |
| 637 |
encodeSrcReducedRegion<S>(rgn); |
| 638 |
} else { |
| 639 |
if (hasSrcRgnHz) |
| 640 |
encodeSrcRegionHorz<S>(rgn.getHz()); |
| 641 |
if (hasSrcRgnVt) |
| 642 |
encodeTernarySrcRegionVert(S, rgn.getVt()); |
| 643 |
} |
| 644 |
// register and subregister |
| 645 |
encodeSrcReg<S>(src.getDirRegName(), src.getDirRegRef().regNum); |
| 646 |
if (inst.isMacro()) { |
| 647 |
if (platform() < Platform::GEN11) { |
| 648 |
fatalT("src", (int)S, ": math macro operands require Align16"); |
| 649 |
return; |
| 650 |
} |
| 651 |
encodeSrcMathMacroReg<S>(src.getMathMacroExt()); |
| 652 |
if (!hasReducedRegion) { |
| 653 |
if (S != SourceIndex::SRC2) { |
| 654 |
encodeTernarySrcRegionVert(S, Region::Vert::VT_4); |
| 655 |
} |
| 656 |
encodeSrcRegionHorz<S>(Region::Horz::HZ_1); |
| 657 |
} |
| 658 |
} else { |
| 659 |
auto subReg = SubRegToBinaryOffset(src.getDirRegRef().subRegNum, |
| 660 |
src.getDirRegName(), src.getType(), |
| 661 |
m_model.platform); |
| 662 |
encodeSrcSubRegNum<S>(subReg); |
| 663 |
} |
| 664 |
break; |
| 665 |
} |
| 666 |
case Operand::Kind::IMMEDIATE: |
| 667 |
if (S == SourceIndex::SRC1) { |
| 668 |
fatalT("src1: immediate operand in ternary align1 must be " |
| 669 |
"src0 or src2"); |
| 670 |
return; |
| 671 |
} |
| 672 |
encodeSrcRegFile<S>(GED_REG_FILE_IMM); |
| 673 |
if (platform() < Platform::GEN10) { |
| 674 |
encodeImmVal(src.getImmediateValue(), src.getType()); |
| 675 |
} else { |
| 676 |
encodeTernaryImmVal<S>(src.getImmediateValue(), src.getType()); |
| 677 |
} |
| 678 |
break; |
| 679 |
default: |
| 680 |
fatalT("src", (int)S, ": invalid operand kind"); |
| 681 |
return; |
| 682 |
} |
| 683 |
} |
| 684 |
void Encoder::encodeQuinaryInstruction(const Instruction &inst) { |
| 685 |
// execution data type set based on dst |
| 686 |
Type dstType = inst.getDestination().getType(); |
| 687 |
GED_EXECUTION_DATA_TYPE execDataType; |
| 688 |
|
| 689 |
// use ternary instruction function support for dst, src0, src1 and src2 |
| 690 |
if (isTernaryAlign1Floating(dstType)) { |
| 691 |
execDataType = GED_EXECUTION_DATA_TYPE_Float; |
| 692 |
} else if (isTernaryAlign1Integral(dstType)) { |
| 693 |
execDataType = GED_EXECUTION_DATA_TYPE_Integer; |
| 694 |
} else { |
| 695 |
fatalT("unsupported execution data type"); |
| 696 |
return; |
| 697 |
} |
| 698 |
GED_ENCODE(ExecutionDataType, execDataType); |
| 699 |
|
| 700 |
if (inst.getOpSpec().supportsDestination()) { |
| 701 |
encodeTernaryDestinationAlign1(inst); |
| 702 |
} |
| 703 |
encodeQuinarySource<SourceIndex::SRC0>(inst); |
| 704 |
encodeQuinarySource<SourceIndex::SRC1>(inst); |
| 705 |
encodeQuinarySource<SourceIndex::SRC2>(inst); |
| 706 |
encodeQuinarySource<SourceIndex::SRC3>(inst); |
| 707 |
encodeQuinarySource<SourceIndex::SRC4>(inst); |
| 708 |
} |
| 709 |
|
| 710 |
void Encoder::encodeTernaryInstruction(const Instruction &inst, |
| 711 |
GED_ACCESS_MODE accessMode) { |
| 712 |
if (accessMode == GED_ACCESS_MODE_Align1) { |
| 713 |
encodeTernaryAlign1Instruction(inst); |
| 714 |
} else { |
| 715 |
encodeTernaryAlign16Instruction(inst); |
| 716 |
} |
| 717 |
} |
| 718 |
void Encoder::encodeTernaryAlign16Instruction(const Instruction &inst) { |
| 719 |
if (inst.getOpSpec().supportsDestination()) { |
| 720 |
encodeTernaryDestinationAlign16(inst); |
| 721 |
} |
| 722 |
encodeTernarySourceAlign16<SourceIndex::SRC0>(inst); |
| 723 |
encodeTernarySourceAlign16<SourceIndex::SRC1>(inst); |
| 724 |
encodeTernarySourceAlign16<SourceIndex::SRC2>(inst); |
| 725 |
} |
| 726 |
void Encoder::encodeTernaryAlign1Instruction(const Instruction &inst) { |
| 727 |
// set ExecutionDataType (integral or floating) |
| 728 |
// the operands must be part of the same type set |
| 729 |
Type src0Type = inst.getSource(0).getType(); |
| 730 |
GED_EXECUTION_DATA_TYPE execDataType; |
| 731 |
if (isTernaryAlign1Floating(src0Type)) { |
| 732 |
execDataType = GED_EXECUTION_DATA_TYPE_Float; |
| 733 |
} else if (isTernaryAlign1Integral(src0Type)) { |
| 734 |
execDataType = GED_EXECUTION_DATA_TYPE_Integer; |
| 735 |
} else { |
| 736 |
fatalT("src0: unsupported type for ternary align1 encoding"); |
| 737 |
return; |
| 738 |
} |
| 739 |
GED_ENCODE(ExecutionDataType, execDataType); |
| 740 |
|
| 741 |
if (inst.getOpSpec().supportsDestination()) { |
| 742 |
encodeTernaryDestinationAlign1(inst); |
| 743 |
} |
| 744 |
encodeTernarySourceAlign1<SourceIndex::SRC0>(inst); |
| 745 |
encodeTernarySourceAlign1<SourceIndex::SRC1>(inst); |
| 746 |
encodeTernarySourceAlign1<SourceIndex::SRC2>(inst); |
| 747 |
} |
| 748 |
|
| 749 |
void Encoder::encodeBranchingInstruction(const Instruction &inst) { |
| 750 |
// the destination stride is always 1 for all control flow |
| 751 |
GED_ENCODE(DstHorzStride, 1); |
| 752 |
|
| 753 |
// control flow instructions require patching later if any operand is a label |
| 754 |
bool src0IsLabel = inst.getSource(0).getKind() == Operand::Kind::LABEL; |
| 755 |
|
| 756 |
// break up instructions into various classes |
| 757 |
// - stuff with implicit operands: jmpi |
| 758 |
// - stuff that can take register operands: call, calla, return |
| 759 |
// - everything else: if, else, while, ..., goto, join, ... |
| 760 |
if (m_opcode == Op::JMPI) { |
| 761 |
// jmpi encodes the syntax |
| 762 |
// jmpi (1) LABEL |
| 763 |
// jmpi (1) reg32 |
| 764 |
// as |
| 765 |
// jmpi (1) ip ip LABEL |
| 766 |
// jmpi (1) ip ip reg32 |
| 767 |
// |
| 768 |
// "Restriction: The index data type must be D (Signed DWord Integer)." |
| 769 |
// |
| 770 |
// implicit IP ... |
| 771 |
encodeBasicDestination(inst, Operand::DST_REG_IP_UD); |
| 772 |
encodeBasicSource<SourceIndex::SRC0>(inst, Operand::SRC_REG_IP_UD); |
| 773 |
GED_ENCODE(Src1DataType, GED_DATA_TYPE_d); |
| 774 |
if (src0IsLabel) { |
| 775 |
// jmpi (1) LABEL (encodes into Src1) |
| 776 |
GED_ENCODE(Src1RegFile, GED_REG_FILE_IMM); |
| 777 |
} else { |
| 778 |
// jmpi (1) reg32 (encodes into Src1) |
| 779 |
encodeBasicSource<SourceIndex::SRC1>(inst, inst.getSource(0)); |
| 780 |
} |
| 781 |
} else if (m_opcode == Op::CALL || m_opcode == Op::CALLA || |
| 782 |
m_opcode == Op::RET) { |
| 783 |
// e.g. call, calla, ret |
| 784 |
// call (..) imm32 |
| 785 |
// call (..) reg32 |
| 786 |
// calla (..) imm32 |
| 787 |
// ret (...) reg32 => encodes as ret (...) null reg |
| 788 |
// |
| 789 |
if (m_opcode == Op::CALL || m_opcode == Op::CALLA) { |
| 790 |
encodeBasicDestination(inst, inst.getDestination()); |
| 791 |
} else if (m_opcode == Op::RET) { |
| 792 |
encodeBasicDestination(inst, Operand::DST_REG_NULL_UD); |
| 793 |
encodeBasicSource<SourceIndex::SRC0>(inst, inst.getSource(0)); |
| 794 |
} |
| 795 |
|
| 796 |
if (m_opcode == Op::CALL || m_opcode == Op::CALLA) { |
| 797 |
if (src0IsLabel) { |
| 798 |
// op == CALL (since it's a label), hence we have |
| 799 |
// call (..) imm32 => which uses src1 |
| 800 |
GED_ENCODE(Src1RegFile, GED_REG_FILE_IMM); |
| 801 |
GED_ENCODE(Src1DataType, GED_DATA_TYPE_d); |
| 802 |
} else { |
| 803 |
// call (..) reg32 |
| 804 |
// calla (..) imm32 |
| 805 |
encodeBasicSource<SourceIndex::SRC1>(inst, inst.getSource(0)); |
| 806 |
} |
| 807 |
|
| 808 |
// <2;2,1> restriction for CALL and CALLA restriction is only for |
| 809 |
// IVB+HSW, but simulator has it until CNL. So we have to support it |
| 810 |
// until we get CNL HW validation moves to it |
| 811 |
if (callNeedsSrc0Region221(inst)) { |
| 812 |
GED_ENCODE(Src0VertStride, 2); |
| 813 |
GED_ENCODE(Src0Width, 2); |
| 814 |
GED_ENCODE(Src0HorzStride, 1); |
| 815 |
} |
| 816 |
// though it's not stated in the spec, ICL requires src0 region be set to |
| 817 |
// <2;4,1> |
| 818 |
else if (callNeedsSrc0Region241(inst)) { |
| 819 |
GED_ENCODE(Src0VertStride, 2); |
| 820 |
GED_ENCODE(Src0Width, 4); |
| 821 |
GED_ENCODE(Src0HorzStride, 1); |
| 822 |
} |
| 823 |
} |
| 824 |
} else if (m_opcode == Op::BRD || m_opcode == Op::BRC) { |
| 825 |
// [brd/brc]: The ip register must be used (for example, by the assembler) |
| 826 |
// as dst. |
| 827 |
encodeBasicDestination(inst, Operand::DST_REG_IP_D); |
| 828 |
if (!src0IsLabel) { |
| 829 |
encodeBasicSource<SourceIndex::SRC0>(inst, inst.getSource(0)); |
| 830 |
} |
| 831 |
GED_DATA_TYPE ty = |
| 832 |
platform() < Platform::GEN8 ? GED_DATA_TYPE_w : GED_DATA_TYPE_d; |
| 833 |
GED_ENCODE(Src0RegFile, src0IsLabel ? GED_REG_FILE_IMM : GED_REG_FILE_GRF); |
| 834 |
GED_ENCODE(Src0DataType, ty); |
| 835 |
// GED automatically sets? |
| 836 |
// if (m_opcode == Op::BRC && src0IsLabel) { |
| 837 |
// GED_ENCODE(Src1DataType, ty); |
| 838 |
// } |
| 839 |
// if (m_opcode == Op::BRD && m_inst->getOpSpec().hasImplicitSrcRegion(0)) { |
| 840 |
// encodeSrcRegion(SourceIndex::SRC0, |
| 841 |
// m_inst->getOpSpec().implicitSrcRegion(0)); |
| 842 |
// } |
| 843 |
} else { |
| 844 |
// regular control flow that only accepts immediate values |
| 845 |
// e.g. if, else, endif, while, cont, break, goto, join, halt |
| 846 |
|
| 847 |
// Apparently, the implicit destination on these instructions |
| 848 |
// is null instead of ip (unlike jmpi etc) |
| 849 |
// destination is ip<1>:ud |
| 850 |
// |
| 851 |
// encodeDestination(&Operand::DST_IP); |
| 852 |
encodeBasicDestination(inst, Operand::DST_REG_NULL_UD); |
| 853 |
|
| 854 |
// UIP |
| 855 |
if (m_opcode != Op::ENDIF && m_opcode != Op::WHILE && |
| 856 |
m_opcode != Op::JOIN) { |
| 857 |
// if/else/halt/brk/cont.... all require :d on operands |
| 858 |
GED_DATA_TYPE ty = |
| 859 |
platform() < Platform::GEN8 ? GED_DATA_TYPE_w : GED_DATA_TYPE_d; |
| 860 |
GED_ENCODE(Src0RegFile, GED_REG_FILE_IMM); |
| 861 |
GED_ENCODE(Src0DataType, ty); |
| 862 |
} |
| 863 |
// before XE don't need to set JIP for control flow instructions that have |
| 864 |
// UIP JIP |
| 865 |
if (m_opcode == Op::WHILE || m_opcode == Op::ENDIF || |
| 866 |
m_opcode == Op::JOIN) { |
| 867 |
GED_ENCODE(Src1RegFile, GED_REG_FILE_IMM); |
| 868 |
GED_ENCODE(Src1DataType, GED_DATA_TYPE_d); |
| 869 |
} |
| 870 |
} |
| 871 |
} |
| 872 |
|
| 873 |
void Encoder::encodeBranchingInstructionSimplified(const Instruction &inst) { |
| 874 |
const OpSpec &os = inst.getOpSpec(); |
| 875 |
|
| 876 |
// set branch control |
| 877 |
if (os.supportsBranchCtrl()) { |
| 878 |
GED_ENCODE(BranchCtrl, lowerBranchCntrl(inst.getBranchCtrl())); |
| 879 |
} |
| 880 |
|
| 881 |
// control flow instructions require patching later if any operand is a label |
| 882 |
const Operand &src0 = inst.getSource(0); |
| 883 |
bool src0IsLabel = src0.getKind() == Operand::Kind::LABEL; |
| 884 |
|
| 885 |
// for jmpi HW will take care of IP so don't need to encode it for dst/src0 |
| 886 |
if (inst.getOpSpec().supportsDestination()) { |
| 887 |
encodeBranchDestination(inst.getDestination()); |
| 888 |
} else { |
| 889 |
encodeBranchDestination(Operand::DST_REG_NULL_UD); |
| 890 |
} |
| 891 |
// regualar control flow that only accepts immediate values |
| 892 |
// e.g. if, else, endif, while, cont, break, goto, join, halt |
| 893 |
|
| 894 |
// encoding JIP |
| 895 |
if (src0IsLabel) { |
| 896 |
GED_ENCODE(Src0RegFile, GED_REG_FILE_IMM); |
| 897 |
// if (src0.getTargetBlock() == nullptr) { |
| 898 |
// // the input value is immediate; use m_immVal as the value |
| 899 |
// encodeBranchSource(src0); |
| 900 |
// } |
| 901 |
} else { |
| 902 |
// jmpi, call, brc, ... |
| 903 |
if (src0.getKind() == Operand::Kind::INDIRECT) |
| 904 |
errorT("branch instructions forbid indirect register mode"); |
| 905 |
encodeBranchSource(src0); |
| 906 |
} |
| 907 |
|
| 908 |
if (inst.getSourceCount() == 2) { |
| 909 |
// encoding UIP always IMM except for brc with a register argument |
| 910 |
if (inst.getOp() != Op::BRC || src0.isImm()) { |
| 911 |
GED_ENCODE(Src1RegFile, GED_REG_FILE_IMM); |
| 912 |
} |
| 913 |
} |
| 914 |
} |
| 915 |
|
| 916 |
void Encoder::encodeSendInstruction(const Instruction &i) { |
| 917 |
//////////////////////////////////////////// |
| 918 |
// send operands |
| 919 |
const OpSpec &os = i.getOpSpec(); |
| 920 |
encodeSendDestination(i.getDestination()); |
| 921 |
if (os.isSendsFormat()) { |
| 922 |
// old sends sources get special treatment |
| 923 |
encodeSendsSource0(i.getSource(0)); |
| 924 |
encodeSendsSource1(i.getSource(1)); |
| 925 |
} else { |
| 926 |
encodeSendSource0(i.getSource(0)); |
| 927 |
if (m_model.supportsXeSend()) { |
| 928 |
encodeSendsSource1(i.getSource(1)); |
| 929 |
} |
| 930 |
} |
| 931 |
|
| 932 |
if (os.isSendgFormat()) { |
| 933 |
GED_ENCODE(Src0Length, (uint32_t)i.getSrc0Length()); |
| 934 |
} |
| 935 |
|
| 936 |
//////////////////////////////////////////// |
| 937 |
// send descriptors and other gunk |
| 938 |
encodeSendDescs(i); |
| 939 |
|
| 940 |
//////////////////////////////////////////// |
| 941 |
// send options |
| 942 |
|
| 943 |
// FusionCtrl is removed from XeHPC+ |
| 944 |
bool hasFusion = platform() >= Platform::XE && platform() < Platform::XE_HPC; |
| 945 |
if (hasFusion) { |
| 946 |
GED_ENCODE(FusionCtrl, i.hasInstOpt(InstOpt::SERIALIZE) |
| 947 |
? GED_FUSION_CTRL_Serialized |
| 948 |
: GED_FUSION_CTRL_Normal); |
| 949 |
} |
| 950 |
|
| 951 |
if (i.hasInstOpt(InstOpt::EOT)) { |
| 952 |
GED_ENCODE(EOT, GED_EOT_EOT); |
| 953 |
} |
| 954 |
} // end: encodeSendInstruction |
| 955 |
|
| 956 |
void Encoder::encodeSendDescs(const Instruction &i) { |
| 957 |
if (platform() < Platform::XE) { |
| 958 |
encodeSendDescsPreXe(i); |
| 959 |
} else if (platform() == Platform::XE) { |
| 960 |
encodeSendDescsXe(i); |
| 961 |
} else if (platform() == Platform::XE_HP) { |
| 962 |
encodeSendDescsXeHP(i); |
| 963 |
} else if (platform() == Platform::XE_HPG || platform() == Platform::XE_HPC |
| 964 |
) { |
| 965 |
encodeSendDescsXeHPG(i); |
| 966 |
} else if (i.getOpSpec().isSendgFormat()) { |
| 967 |
encodeSendgDescsXe3(i); |
| 968 |
} else if (platform() >= Platform::XE2) { |
| 969 |
encodeSendDescsXe2(i); |
| 970 |
} else { |
| 971 |
errorT("unsupported platform"); |
| 972 |
} |
| 973 |
|
| 974 |
bool noEOTinExDesc = m_model.supportsXeSend(); |
| 975 |
noEOTinExDesc &= platform() <= Platform::XE3; // efficient 64b send desc |
| 976 |
if (noEOTinExDesc && i.getExtMsgDescriptor().isImm() && |
| 977 |
(i.getExtMsgDescriptor().imm & 1 << 5)) |
| 978 |
errorT("Encoder: Send exDesc[5] must not be set (the legacy EOT bit)"); |
| 979 |
} |
| 980 |
|
| 981 |
void Encoder::encodeSendDescsPreXe(const Instruction &i) { |
| 982 |
SendDesc exDesc = i.getExtMsgDescriptor(); |
| 983 |
const OpSpec &os = i.getOpSpec(); |
| 984 |
if (exDesc.isReg()) { |
| 985 |
if (os.isSendFormat()) { |
| 986 |
errorT("unary send forbids register ExDesc"); |
| 987 |
} |
| 988 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_ARF); |
| 989 |
GED_ENCODE(ExDescAddrSubRegNum, 2 * exDesc.reg.subRegNum); |
| 990 |
} else { |
| 991 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_IMM); |
| 992 |
GED_ENCODE(ExMsgDescImm, exDesc.imm); |
| 993 |
} |
| 994 |
|
| 995 |
SendDesc desc = i.getMsgDescriptor(); |
| 996 |
if (desc.isReg()) { |
| 997 |
if (platform() == Platform::GEN9) { |
| 998 |
uint32_t msgDescriptor = 0; |
| 999 |
// There is a HW bug on SKL where HW will only copy bits 0-28 from |
| 1000 |
// the address register (descriptor register) and will miss bit 30 |
| 1001 |
// of the descriptor. Hence, even in the case of an register |
| 1002 |
// descriptor we must program bit 30 as immediate (it will be |
| 1003 |
// taken from the encoding and OR'd in correctly) |
| 1004 |
// |
| 1005 |
// E.g. (old syntax) |
| 1006 |
// sends (8) r74:hf r16 r73 0x42:ud a0.0 {Align1, Q1, NoMask} |
| 1007 |
// // sampler, resLen=3, msgLen=1, extMsgLen=1 |
| 1008 |
// On SKL, HW will copy bits 29-31 from the actual immediate |
| 1009 |
// descriptor bits. Hence, we must set immediate descriptor |
| 1010 |
// bit 30 even in the case of a register descriptor. (For SKL). |
| 1011 |
// |
| 1012 |
// For 3D sampler bit 30 indicates HF/F return format. |
| 1013 |
// For render target write bit 30 indicates HF/F input... |
| 1014 |
// Thankfully for SKL the 3D sampler doesn't support HF input. |
| 1015 |
// For CNL it does, and that will be bit 29. |
| 1016 |
// But this bug should be fixed in CNL. |
| 1017 |
if (platform() == Platform::GEN9 && desc.isReg()) { |
| 1018 |
if (i.getDestination().getType() == Type::HF || |
| 1019 |
i.getSource(0).getType() == Type::HF) { |
| 1020 |
msgDescriptor |= (1 << 30); |
| 1021 |
} |
| 1022 |
} |
| 1023 |
GED_ENCODE(DescRegFile, GED_REG_FILE_IMM); |
| 1024 |
GED_ENCODE(MsgDesc, msgDescriptor); |
| 1025 |
} |
| 1026 |
GED_ENCODE(DescRegFile, GED_REG_FILE_ARF); |
| 1027 |
const RegInfo *ri = m_model.lookupRegInfoByRegName(RegName::ARF_A); |
| 1028 |
IGA_ASSERT(ri, "failed to find a0 register"); |
| 1029 |
GED_ENCODE(DescRegNum, ri->encodeARFRegNum((int)desc.reg.regNum)); |
| 1030 |
} else if (desc.isImm()) { |
| 1031 |
GED_ENCODE(DescRegFile, GED_REG_FILE_IMM); |
| 1032 |
GED_ENCODE(MsgDesc, desc.imm); |
| 1033 |
} |
| 1034 |
} |
| 1035 |
void Encoder::encodeSendDescsXe(const Instruction &i) { |
| 1036 |
SendDesc exDesc = i.getExtMsgDescriptor(); |
| 1037 |
if (exDesc.isReg()) { |
| 1038 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_ARF); |
| 1039 |
GED_ENCODE(ExDescAddrSubRegNum, 2 * exDesc.reg.subRegNum); |
| 1040 |
} else { |
| 1041 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_IMM); |
| 1042 |
GED_ENCODE(ExMsgDescImm, exDesc.imm); |
| 1043 |
} |
| 1044 |
|
| 1045 |
SendDesc desc = i.getMsgDescriptor(); |
| 1046 |
if (desc.isReg()) { |
| 1047 |
GED_ENCODE(DescRegFile, GED_REG_FILE_ARF); |
| 1048 |
// a0.0 is implied (there's no field) |
| 1049 |
if (desc.reg.subRegNum != 0) { |
| 1050 |
errorT("send with reg desc must be a0.0"); |
| 1051 |
} |
| 1052 |
} else { |
| 1053 |
GED_ENCODE(DescRegFile, GED_REG_FILE_IMM); |
| 1054 |
GED_ENCODE(MsgDesc, desc.imm); |
| 1055 |
} |
| 1056 |
} |
| 1057 |
|
| 1058 |
// A bit harder than Xe |
| 1059 |
// * If ExBSO is set then Src1Length holds xlen |
| 1060 |
// * CPS has it's own field (ExDesc[11]) only if ExDesc.IsReg |
| 1061 |
void Encoder::encodeSendDescsXeHP(const Instruction &i) { |
| 1062 |
SendDesc exDesc = i.getExtMsgDescriptor(); |
| 1063 |
if (exDesc.isReg()) { |
| 1064 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_ARF); |
| 1065 |
GED_ENCODE(ExDescAddrSubRegNum, 2 * exDesc.reg.subRegNum); |
| 1066 |
GED_ENCODE(ExBSO, i.hasInstOpt(InstOpt::EXBSO) ? 1 : 0); |
| 1067 |
if (i.hasInstOpt(InstOpt::EXBSO)) { |
| 1068 |
GED_ENCODE(CPS, i.hasInstOpt(InstOpt::CPS) ? 1 : 0); |
| 1069 |
GED_ENCODE(Src1Length, (uint32_t)i.getSrc1Length()); |
| 1070 |
} else if (i.hasInstOpt(InstOpt::CPS)) { |
| 1071 |
errorT("{CPS} requires {ExBSO}"); |
| 1072 |
} |
| 1073 |
} else { |
| 1074 |
if (i.hasInstOpt(InstOpt::CPS)) { |
| 1075 |
warningT("when ExDesc is immediate use ExDesc[11] rather than {CPS}"); |
| 1076 |
exDesc.imm |= 1 << 11; |
| 1077 |
} |
| 1078 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_IMM); |
| 1079 |
GED_ENCODE(ExMsgDescImm, exDesc.imm); |
| 1080 |
} |
| 1081 |
|
| 1082 |
SendDesc desc = i.getMsgDescriptor(); |
| 1083 |
if (desc.isReg()) { |
| 1084 |
GED_ENCODE(DescRegFile, GED_REG_FILE_ARF); |
| 1085 |
if (desc.reg.subRegNum != 0) { // a0.0 is implied (there's no field) |
| 1086 |
errorT("send with reg desc must be a0.0"); |
| 1087 |
} |
| 1088 |
} else { |
| 1089 |
GED_ENCODE(DescRegFile, GED_REG_FILE_IMM); |
| 1090 |
GED_ENCODE(MsgDesc, desc.imm); |
| 1091 |
} |
| 1092 |
} |
| 1093 |
|
| 1094 |
// Similar to XeHP, except |
| 1095 |
// * ExDesc.IsImm implies use of Src1Length (Src.Length is in EU bits) |
| 1096 |
void Encoder::encodeSendDescsXeHPG(const Instruction &i) { |
| 1097 |
SendDesc exDesc = i.getExtMsgDescriptor(); |
| 1098 |
if (exDesc.isReg()) { |
| 1099 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_ARF); |
| 1100 |
GED_ENCODE(ExDescAddrSubRegNum, 2 * exDesc.reg.subRegNum); |
| 1101 |
GED_ENCODE(ExBSO, i.hasInstOpt(InstOpt::EXBSO) ? 1 : 0); |
| 1102 |
if (i.hasInstOpt(InstOpt::EXBSO)) { |
| 1103 |
GED_ENCODE(CPS, i.hasInstOpt(InstOpt::CPS) ? 1 : 0); |
| 1104 |
GED_ENCODE(Src1Length, (uint32_t)i.getSrc1Length()); |
| 1105 |
} else if (i.hasInstOpt(InstOpt::CPS)) { |
| 1106 |
errorT("{CPS} requires {ExBSO}"); |
| 1107 |
} |
| 1108 |
} else { |
| 1109 |
if (i.hasInstOpt(InstOpt::CPS)) { |
| 1110 |
warningT("when ExDesc is immediate use ExDesc[11] rather than {CPS}"); |
| 1111 |
exDesc.imm |= 1 << 11; |
| 1112 |
} |
| 1113 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_IMM); |
| 1114 |
GED_ENCODE(ExMsgDescImm, exDesc.imm); |
| 1115 |
GED_ENCODE(Src1Length, (uint32_t)i.getSrc1Length()); |
| 1116 |
} |
| 1117 |
|
| 1118 |
SendDesc desc = i.getMsgDescriptor(); |
| 1119 |
if (desc.isReg()) { |
| 1120 |
GED_ENCODE(DescRegFile, GED_REG_FILE_ARF); |
| 1121 |
if (desc.reg.subRegNum != 0) { // a0.0 is implied (there's no field) |
| 1122 |
errorT("send with reg desc must be a0.0"); |
| 1123 |
} |
| 1124 |
} else { |
| 1125 |
GED_ENCODE(DescRegFile, GED_REG_FILE_IMM); |
| 1126 |
GED_ENCODE(MsgDesc, desc.imm); |
| 1127 |
} |
| 1128 |
} |
| 1129 |
|
| 1130 |
// Similar to XeHPG/XeHPC except: |
| 1131 |
// * ExBSO doesn't exist for UGM (high bit of ExImmOffDesc) |
| 1132 |
// * if ExBSO field exists, it denotes if Src1Len in EU field or in ExDesc |
| 1133 |
// * CPS is gone (use ExImmOff[11]) |
| 1134 |
// * If ExDesc.IsReg, then we also set ExImmOffDesc. And if it's UGM, |
| 1135 |
// ExDescImm[15] is the same bit as ExBSO (ExBSO not) |
| 1136 |
void Encoder::encodeSendDescsXe2(const Instruction &i) { |
| 1137 |
SendDesc exDesc = i.getExtMsgDescriptor(); |
| 1138 |
bool hasSrc1Length = true; |
| 1139 |
// Gather Send lacks Src1.Len |
| 1140 |
// (the same field [103:99] encoding field becomes Src0.SubRegNum) |
| 1141 |
if (i.isGatherSend()) |
| 1142 |
hasSrc1Length = false; |
| 1143 |
if (exDesc.isReg()) { |
| 1144 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_ARF); |
| 1145 |
// set ExBSO for non-UGM |
| 1146 |
// (UGM uses the ExBSO bit for ExDescImm[15]) |
| 1147 |
uint32_t exImmOffDesc = i.getExtImmOffDescriptor(); |
| 1148 |
if (i.getSendFc() != SFID::UGM) { |
| 1149 |
// ExDescImm[15] must not be set for non-UGM which the bit is ExBSO |
| 1150 |
if (exImmOffDesc & (1 << 15)) { |
| 1151 |
exImmOffDesc &= ~(1 << 15); |
| 1152 |
errorT("ExDescImm[15] overlaps ExBSO for this SFID and must be 0"); |
| 1153 |
} |
| 1154 |
} |
| 1155 |
if (exImmOffDesc & (0x7 << 16)) { |
| 1156 |
exImmOffDesc &= ~(0x7 << 16); |
| 1157 |
errorT("ExDescImm[18:16] overlaps ExDesc.Reg and must be 0"); |
| 1158 |
} |
| 1159 |
if (exImmOffDesc & 0x7FF) { |
| 1160 |
exImmOffDesc &= ~0x7FF; |
| 1161 |
errorT("ExDescImm[10:0] are unmapped and must be 0"); |
| 1162 |
} |
| 1163 |
GED_ENCODE(ExMsgDescImm, exImmOffDesc); |
| 1164 |
|
| 1165 |
if (i.getSendFc() != SFID::UGM) { |
| 1166 |
// must be done *after* ExMsgDescImm since this will stomp |
| 1167 |
// ExDescImm[15] (we could flip it around and just set ExDescImm[15]) |
| 1168 |
GED_ENCODE(ExBSO, i.hasInstOpt(InstOpt::EXBSO) ? 1 : 0); |
| 1169 |
if (i.hasInstOpt(InstOpt::EXBSO)) { |
| 1170 |
GED_ENCODE(Src1Length, (uint32_t)i.getSrc1Length()); |
| 1171 |
} |
| 1172 |
} else { |
| 1173 |
if (i.hasInstOpt(InstOpt::EXBSO)) |
| 1174 |
warningAtT(i.getLoc(), |
| 1175 |
"{ExBSO} does not exist for send.ugm on this platform"); |
| 1176 |
// for UGM, src1.len must be in EU field |
| 1177 |
if (hasSrc1Length) |
| 1178 |
GED_ENCODE(Src1Length, (uint32_t)i.getSrc1Length()); |
| 1179 |
} |
| 1180 |
|
| 1181 |
if (i.hasInstOpt(InstOpt::CPS)) { |
| 1182 |
// this is ExDesc[11] |
| 1183 |
errorT("{CPS} does not exist on this platform"); |
| 1184 |
} |
| 1185 |
|
| 1186 |
// has to be set after ExMsgDescImm due to overlap |
| 1187 |
GED_ENCODE(ExDescAddrSubRegNum, 2 * exDesc.reg.subRegNum); |
| 1188 |
} else { |
| 1189 |
GED_ENCODE(ExDescRegFile, GED_REG_FILE_IMM); |
| 1190 |
GED_ENCODE(ExMsgDescImm, exDesc.imm); |
| 1191 |
if (hasSrc1Length) |
| 1192 |
GED_ENCODE(Src1Length, (uint32_t)i.getSrc1Length()); |
| 1193 |
} |
| 1194 |
|
| 1195 |
SendDesc desc = i.getMsgDescriptor(); |
| 1196 |
if (desc.isReg()) { |
| 1197 |
GED_ENCODE(DescRegFile, GED_REG_FILE_ARF); |
| 1198 |
if (desc.reg.subRegNum != 0) { // a0.0 is implied (there's no field) |
| 1199 |
errorT("send with reg desc must be a0.0"); |
| 1200 |
} |
| 1201 |
} else { |
| 1202 |
GED_ENCODE(DescRegFile, GED_REG_FILE_IMM); |
| 1203 |
GED_ENCODE(MsgDesc, desc.imm); |
| 1204 |
} |
| 1205 |
} |
| 1206 |
|
| 1207 |
// Completely new encoding. |
| 1208 |
// |
| 1209 |
void Encoder::encodeSendgDescsXe3(const Instruction &i) { |
| 1210 |
int src1Len = i.getSrc1Length(); |
| 1211 |
if (src1Len > 0) { |
| 1212 |
bool encoded = false; |
| 1213 |
if (!encoded) |
| 1214 |
GED_ENCODE(Src1Length, (uint32_t)src1Len); |
| 1215 |
} |
| 1216 |
auto id0 = i.getSendgIndDesc0Reg(); |
| 1217 |
if (id0 != REGREF_INVALID) { |
| 1218 |
GED_ENCODE(IndMsgDesc0IsPresent, 1); |
| 1219 |
if (id0.subRegNum % 8) { |
| 1220 |
errorT("ID0 must be a 8-byte-aligned offset"); |
| 1221 |
} |
| 1222 |
GED_ENCODE(IndMsgDesc0Addr, id0.subRegNum / 8); |
| 1223 |
} |
| 1224 |
auto id1 = i.getSendgIndDesc1Reg(); |
| 1225 |
if (id1 != REGREF_INVALID) { |
| 1226 |
if (id1.subRegNum % 8) { |
| 1227 |
errorT("ID1 must be a 8-byte-aligned offset"); |
| 1228 |
} |
| 1229 |
GED_ENCODE(IndMsgDesc1IsPresent, 1); |
| 1230 |
GED_ENCODE(IndMsgDesc1Addr, id1.subRegNum / 8); |
| 1231 |
} |
| 1232 |
auto desc = i.getSendgDesc(); |
| 1233 |
if (id1 != REGREF_INVALID && (desc & ~((1ull << 42) - 1))) { |
| 1234 |
errorT("Desc[46:42] overlaps ID1 (those bits must be zero in desc)"); |
| 1235 |
} |
| 1236 |
GED_ENCODE(MsgDesc64, desc); |
| 1237 |
} |
| 1238 |
|
| 1239 |
void Encoder::encodeSyncInstruction(const Instruction &inst) { |
| 1240 |
// Set the Dst.HorStride to 1 so that "sync.bar null" can be compacted |
| 1241 |
GED_ENCODE(DstHorzStride, 1); |
| 1242 |
|
| 1243 |
const Operand &src = inst.getSource(0); |
| 1244 |
if (src.getKind() == Operand::Kind::IMMEDIATE) { |
| 1245 |
encodeSrcRegFile<SourceIndex::SRC0>(GED_REG_FILE_IMM); |
| 1246 |
encodeSrcType<SourceIndex::SRC0>(src.getType()); |
| 1247 |
encodeImmVal(src.getImmediateValue(), src.getType()); |
| 1248 |
} else { |
| 1249 |
if (platform() <= Platform::XE_HPG) { |
| 1250 |
encodeSrcRegFile<SourceIndex::SRC0>(GED_REG_FILE_ARF); |
| 1251 |
} else { |
| 1252 |
// XeHPC+ supports sync with reg32. For earlier platforms encode it to the |
| 1253 |
// null reg anyway. If not doing so we'll encounter some weird behavior on |
| 1254 |
// validation. Suspect it's becuase on some previous platforms' testcase |
| 1255 |
// there are reg32 those are not valid, but IGA workaround (set it to |
| 1256 |
// NULL) them |
| 1257 |
if (src.isNull()) { |
| 1258 |
encodeSrcRegFile<SourceIndex::SRC0>(GED_REG_FILE_ARF); |
| 1259 |
} else { |
| 1260 |
// currently only flag register is supported in sync.bar |
| 1261 |
encodeSrcRegFile<SourceIndex::SRC0>(lowerRegFile(src.getDirRegName())); |
| 1262 |
encodeSrcReg<SourceIndex::SRC0>(src.getDirRegName(), |
| 1263 |
src.getDirRegRef().regNum); |
| 1264 |
encodeSrcType<SourceIndex::SRC0>(src.getType()); |
| 1265 |
// must be flag register (otherwise GED will return error), encode the |
| 1266 |
// subreg directly. |
| 1267 |
GED_ENCODE(Src0SubRegNum, |
| 1268 |
SubRegToBinaryOffset(src.getDirRegRef().subRegNum, |
| 1269 |
src.getDirRegName(), src.getType(), |
| 1270 |
m_model.platform)); |
| 1271 |
} |
| 1272 |
} |
| 1273 |
} |
| 1274 |
} |
| 1275 |
|
| 1276 |
void Encoder::encodeBranchDestination(const Operand &dst) { |
| 1277 |
GED_ENCODE(DstRegFile, lowerRegFile(dst.getDirRegName())); |
| 1278 |
encodeDstReg(dst.getDirRegName(), dst.getDirRegRef().regNum); |
| 1279 |
GED_ENCODE(DstSubRegNum, |
| 1280 |
SubRegToBinaryOffset(dst.getDirRegRef().subRegNum, |
| 1281 |
dst.getDirRegName(), dst.getType(), |
| 1282 |
m_model.platform)); |
| 1283 |
} |
| 1284 |
|
| 1285 |
void Encoder::encodeBasicDestination(const Instruction &inst, |
| 1286 |
const Operand &dst, |
| 1287 |
GED_ACCESS_MODE accessMode) { |
| 1288 |
IGA_ASSERT(accessMode != GED_ACCESS_MODE_Align16 || m_model.supportsAlign16(), |
| 1289 |
"Align16 not supported on this platform."); |
| 1290 |
|
| 1291 |
GED_ENCODE(DstRegFile, lowerRegFile(dst.getDirRegName())); |
| 1292 |
switch (dst.getKind()) { |
| 1293 |
case Operand::Kind::DIRECT: |
| 1294 |
case Operand::Kind::MACRO: |
| 1295 |
GED_ENCODE(DstAddrMode, GED_ADDR_MODE_Direct); |
| 1296 |
GED_ENCODE(DstDataType, lowerDataType(dst.getType())); |
| 1297 |
if (inst.getOpSpec().supportsSaturation()) { |
| 1298 |
GED_ENCODE(Saturate, lowerSaturate(dst.getDstModifier())); |
| 1299 |
} |
| 1300 |
// VVVVV fallthrough VVVVV |
| 1301 |
default: |
| 1302 |
break; |
| 1303 |
} |
| 1304 |
|
| 1305 |
switch (dst.getKind()) { |
| 1306 |
case Operand::Kind::DIRECT: |
| 1307 |
if (accessMode == GED_ACCESS_MODE_Align16) { |
| 1308 |
if (dst.getRegion() != Region::DST1) { |
| 1309 |
fatalT("dst has inconvertible region for Align16 encoding"); |
| 1310 |
return; |
| 1311 |
} |
| 1312 |
if (isAlign16MathMacroRegisterCsrOperand(dst)) { |
| 1313 |
// acc2.XXXX on BDW .. SKL is context save and restore |
| 1314 |
// This is really mme0 |
| 1315 |
encodeDstReg(RegName::ARF_MME, 0); |
| 1316 |
// on GEN8 and GEN9 all encode as acc2, but the mux varies |
| 1317 |
// to distinguish which acc it really is. |
| 1318 |
GED_DST_CHAN_EN chEn; |
| 1319 |
switch (dst.getDirRegRef().regNum) { |
| 1320 |
/// case 0: ... acc2 actually uses Align1! |
| 1321 |
// old-style for acc2 would be: |
| 1322 |
// mov(8) r113:ud acc2:ud {NoMask} // acc2 |
| 1323 |
// |
| 1324 |
// acc3-9 are Align16 |
| 1325 |
case 1: |
| 1326 |
chEn = GED_DST_CHAN_EN_x; |
| 1327 |
break; // mme0/acc3 -> acc2.x (0001b) |
| 1328 |
case 2: |
| 1329 |
chEn = GED_DST_CHAN_EN_y; |
| 1330 |
break; // mme1/acc4 -> acc2.y (0010b) |
| 1331 |
case 3: |
| 1332 |
chEn = GED_DST_CHAN_EN_xy; |
| 1333 |
break; |
| 1334 |
case 4: |
| 1335 |
chEn = GED_DST_CHAN_EN_z; |
| 1336 |
break; |
| 1337 |
case 5: |
| 1338 |
chEn = GED_DST_CHAN_EN_xz; |
| 1339 |
break; |
| 1340 |
case 6: |
| 1341 |
chEn = GED_DST_CHAN_EN_yz; |
| 1342 |
break; |
| 1343 |
case 7: |
| 1344 |
chEn = GED_DST_CHAN_EN_xyzw; |
| 1345 |
break; // mme7/acc9 -> acc2.xyzw (0111b) |
| 1346 |
default: |
| 1347 |
IGA_ASSERT_FALSE("unreachable"); |
| 1348 |
chEn = GED_DST_CHAN_EN_x; |
| 1349 |
} |
| 1350 |
GED_ENCODE(DstChanEn, chEn); |
| 1351 |
} else { |
| 1352 |
// normal align16 destination (this still might be a |
| 1353 |
// CSR work around op if the src is "acc2") |
| 1354 |
encodeDstReg(dst.getDirRegName(), dst.getDirRegRef().regNum); |
| 1355 |
GED_ENCODE(DstChanEn, GED_DST_CHAN_EN_xyzw); |
| 1356 |
} |
| 1357 |
GED_ENCODE(DstSubRegNum, |
| 1358 |
SubRegToBinaryOffset(dst.getDirRegRef().subRegNum, |
| 1359 |
dst.getDirRegName(), dst.getType(), |
| 1360 |
m_model.platform)); |
| 1361 |
} else { // Align1 |
| 1362 |
encodeDstReg(dst.getDirRegName(), dst.getDirRegRef().regNum); |
| 1363 |
GED_ENCODE(DstSubRegNum, |
| 1364 |
SubRegToBinaryOffset(dst.getDirRegRef().subRegNum, |
| 1365 |
dst.getDirRegName(), dst.getType(), |
| 1366 |
m_model.platform)); |
| 1367 |
} |
| 1368 |
break; |
| 1369 |
case Operand::Kind::MACRO: |
| 1370 |
encodeDstReg(dst.getDirRegName(), dst.getDirRegRef().regNum); |
| 1371 |
GED_ENCODE(DstMathMacroExt, lowerSpecialAcc(dst.getMathMacroExt())); |
| 1372 |
if (accessMode == GED_ACCESS_MODE_Align1 && |
| 1373 |
m_model.supportsAlign16ImplicitAcc()) { |
| 1374 |
fatalT("Align1 dst math macro unsupported on this platform."); |
| 1375 |
return; |
| 1376 |
} |
| 1377 |
break; |
| 1378 |
case Operand::Kind::INDIRECT: |
| 1379 |
GED_ENCODE(DstAddrMode, GED_ADDR_MODE_Indirect); |
| 1380 |
GED_ENCODE(DstDataType, lowerDataType(dst.getType())); |
| 1381 |
if (inst.getOpSpec().supportsSaturation()) { |
| 1382 |
GED_ENCODE(Saturate, lowerSaturate(dst.getDstModifier())); |
| 1383 |
} |
| 1384 |
|
| 1385 |
GED_ENCODE(DstAddrImm, dst.getIndImmAddr()); |
| 1386 |
GED_ENCODE(DstAddrSubRegNum, dst.getIndAddrReg().subRegNum); |
| 1387 |
break; |
| 1388 |
default: |
| 1389 |
IGA_ASSERT_FALSE("unsupported operand kind"); |
| 1390 |
break; |
| 1391 |
} |
| 1392 |
|
| 1393 |
if (accessMode == GED_ACCESS_MODE_Align1) { |
| 1394 |
auto dstRgn = dst.getRegion(); |
| 1395 |
if (inst.getOpSpec().hasImplicitDstRegion(inst.isMacro())) { |
| 1396 |
auto dstRgnImpl = inst.getOpSpec().implicitDstRegion(inst.isMacro()); |
| 1397 |
if (dstRgn != dstRgnImpl) { |
| 1398 |
warningT("dst region should be ", ToSyntax(dstRgnImpl)); |
| 1399 |
} |
| 1400 |
} |
| 1401 |
GED_ENCODE(DstHorzStride, lowerRegionHorz(dstRgn.getHz())); |
| 1402 |
} |
| 1403 |
} |
| 1404 |
|
| 1405 |
static void createChSelForCtxSavRst(GED_SWIZZLE *chSel, GED_SWIZZLE x, |
| 1406 |
GED_SWIZZLE y) { |
| 1407 |
// following IsaAsm rules here |
| 1408 |
// reg.ab expands to reg.abbb |
| 1409 |
chSel[0] = x; |
| 1410 |
chSel[1] = chSel[2] = chSel[3] = y; |
| 1411 |
} |
| 1412 |
|
| 1413 |
void Encoder::encodeBranchSource(const Operand &src) { |
| 1414 |
encodeSrcRegFile<SourceIndex::SRC0>(lowerRegFile(src.getDirRegName())); |
| 1415 |
encodeSrcReg<SourceIndex::SRC0>(src.getDirRegName(), |
| 1416 |
src.getDirRegRef().regNum); |
| 1417 |
auto subReg = |
| 1418 |
SubRegToBinaryOffset(src.getDirRegRef().subRegNum, src.getDirRegName(), |
| 1419 |
Type::D, m_model.platform); |
| 1420 |
encodeSrcSubRegNum<SourceIndex::SRC0>(subReg); |
| 1421 |
} |
| 1422 |
|
| 1423 |
template <SourceIndex S> |
| 1424 |
void Encoder::encodeBasicSource(const Instruction &inst, |
| 1425 |
const Operand &src, |
| 1426 |
GED_ACCESS_MODE accessMode) { |
| 1427 |
// setting the reg file must precede must precede setting the type in GED |
| 1428 |
switch (src.getKind()) { |
| 1429 |
case Operand::Kind::DIRECT: |
| 1430 |
case Operand::Kind::MACRO: |
| 1431 |
case Operand::Kind::INDIRECT: |
| 1432 |
encodeSrcRegFile<S>(lowerRegFile(src.getDirRegName())); |
| 1433 |
if (inst.getOpSpec().supportsSourceModifiers()) { |
| 1434 |
encodeSrcModifier<S>(src.getSrcModifier()); |
| 1435 |
} else if (src.getSrcModifier() != SrcModifier::NONE) { |
| 1436 |
// better be invalid in the IR if unsupported |
| 1437 |
errorT("src", (int)S, " source modifier not supported (invalid IR)"); |
| 1438 |
} |
| 1439 |
break; |
| 1440 |
case Operand::Kind::IMMEDIATE: |
| 1441 |
encodeSrcRegFile<S>(GED_REG_FILE_IMM); |
| 1442 |
break; |
| 1443 |
default: |
| 1444 |
break; |
| 1445 |
} |
| 1446 |
|
| 1447 |
encodeSrcType<S>(src.getType()); |
| 1448 |
|
| 1449 |
switch (src.getKind()) { |
| 1450 |
case Operand::Kind::DIRECT: |
| 1451 |
case Operand::Kind::MACRO: { |
| 1452 |
encodeSrcAddrMode<S>(GED_ADDR_MODE_Direct); |
| 1453 |
if (src.getKind() == Operand::Kind::DIRECT) { |
| 1454 |
if (isAlign16MathMacroRegisterCsrOperand(src)) { |
| 1455 |
// BDW..SKL context save and restore of acc3...acc9 |
| 1456 |
// encode as acc2.####, ChSel will be changed in regioning code |
| 1457 |
// recall acc2 is remapped to mme0 |
| 1458 |
encodeSrcReg<S>(RegName::ARF_MME, 0); |
| 1459 |
} else { |
| 1460 |
encodeSrcReg<S>(src.getDirRegName(), src.getDirRegRef().regNum); |
| 1461 |
auto subReg = SubRegToBinaryOffset(src.getDirRegRef().subRegNum, |
| 1462 |
src.getDirRegName(), src.getType(), |
| 1463 |
m_model.platform); |
| 1464 |
encodeSrcSubRegNum<S>(subReg); |
| 1465 |
} |
| 1466 |
} else { // (src.getKind() == Operand::Kind::MACRO) |
| 1467 |
encodeSrcReg<S>(src.getDirRegName(), src.getDirRegRef().regNum); |
| 1468 |
encodeSrcMathMacroReg<S>(src.getMathMacroExt()); |
| 1469 |
if (accessMode == GED_ACCESS_MODE_Align16) { |
| 1470 |
// vertical stride has to be halved for 8B types |
| 1471 |
if (src.getType() == Type::DF) { |
| 1472 |
encodeSrcRegionVert<S>(Region::Vert::VT_2); |
| 1473 |
} else { |
| 1474 |
encodeSrcRegionVert<S>(Region::Vert::VT_4); |
| 1475 |
} |
| 1476 |
} |
| 1477 |
} |
| 1478 |
break; |
| 1479 |
} |
| 1480 |
case Operand::Kind::INDIRECT: |
| 1481 |
encodeSrcAddrMode<S>(GED_ADDR_MODE_Indirect); |
| 1482 |
encodeSrcAddrImm<S>(src.getIndImmAddr()); |
| 1483 |
encodeSrcAddrSubRegNum<S>(src.getIndAddrReg().subRegNum); |
| 1484 |
break; |
| 1485 |
case Operand::Kind::IMMEDIATE: |
| 1486 |
encodeImmVal(src.getImmediateValue(), src.getType()); |
| 1487 |
break; |
| 1488 |
default: |
| 1489 |
// support mov label |
| 1490 |
if (static_cast<int>(S) == 0 && inst.isMovWithLabel()) { |
| 1491 |
GED_ENCODE(Src0RegFile, GED_REG_FILE_IMM); |
| 1492 |
} else { |
| 1493 |
fatalT("src", (int)S, |
| 1494 |
": unsupported source operand kind " |
| 1495 |
"(malformed IR)"); |
| 1496 |
return; |
| 1497 |
} |
| 1498 |
break; |
| 1499 |
} |
| 1500 |
|
| 1501 |
// sets stuff found in all register accesses (not macros) |
| 1502 |
// - region |
| 1503 |
switch (src.getKind()) { |
| 1504 |
case Operand::Kind::DIRECT: |
| 1505 |
case Operand::Kind::INDIRECT: |
| 1506 |
if (accessMode == GED_ACCESS_MODE_Align16) { |
| 1507 |
// r13.0<4>.xyzw is the only supported ChEn |
| 1508 |
// ^^^ |
| 1509 |
encodeSrcRegionVert<S>(Region::Vert::VT_4); |
| 1510 |
GED_SWIZZLE chSel[4] = {GED_SWIZZLE_x, GED_SWIZZLE_y, GED_SWIZZLE_z, |
| 1511 |
GED_SWIZZLE_w}; |
| 1512 |
if (isAlign16MathMacroRegisterCsrOperand(src)) { |
| 1513 |
// context save and restore workaround on GEN8 and GEN9 |
| 1514 |
switch (src.getDirRegRef().regNum) { |
| 1515 |
case 1: // acc2.yx = mme1 (acc3) |
| 1516 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_y, GED_SWIZZLE_x); |
| 1517 |
break; |
| 1518 |
case 2: // acc2.zx = mme2 (acc4) |
| 1519 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_z, GED_SWIZZLE_x); |
| 1520 |
break; |
| 1521 |
case 3: // acc2.wx = mme3 (acc5) |
| 1522 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_w, GED_SWIZZLE_x); |
| 1523 |
break; |
| 1524 |
case 4: // acc2.xy = mme4 (acc6) |
| 1525 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_x, GED_SWIZZLE_y); |
| 1526 |
break; |
| 1527 |
case 5: // acc2.yy = mme5 (acc7) |
| 1528 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_y, GED_SWIZZLE_y); |
| 1529 |
break; |
| 1530 |
case 6: // acc2.zy = mme6 (acc8) |
| 1531 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_z, GED_SWIZZLE_y); |
| 1532 |
break; |
| 1533 |
case 7: // acc2.wy = mme7 (acc9) |
| 1534 |
createChSelForCtxSavRst(chSel, GED_SWIZZLE_w, GED_SWIZZLE_y); |
| 1535 |
break; |
| 1536 |
} |
| 1537 |
} else { |
| 1538 |
// normal Align16 that we are converting to Align1 |
| 1539 |
if (src.getRegion() != Region::SRC110 && |
| 1540 |
// supports legacy bits that may use <K;K,1> for "block" |
| 1541 |
// access; this allows us to assemble/reassemble similar bits |
| 1542 |
src.getRegion() != Region::SRC221 && |
| 1543 |
src.getRegion() != Region::SRC441 && |
| 1544 |
src.getRegion() != Region::SRC881 && |
| 1545 |
src.getRegion() != Region::SRCFF1) { |
| 1546 |
fatalT("src", (int)S, |
| 1547 |
": unsupported region for " |
| 1548 |
"translation to align16 encoding"); |
| 1549 |
return; |
| 1550 |
} |
| 1551 |
// TODO: we could permit SIMD4 with .x to mean broadcast read |
| 1552 |
// of subreg 0, but I don't think any System Routine code uses |
| 1553 |
// this. |
| 1554 |
// |
| 1555 |
// NOTE: technically we could convert |
| 1556 |
// r13.0<0>.xxxx to r13.0<0;1,0> |
| 1557 |
// r13.0<0>.yyyy to r13.1<0;1,0> |
| 1558 |
// r13.0<0>.zzzz to r13.2<0;1,0> |
| 1559 |
// r13.0<0>.wwww to r13.3<0;1,0> |
| 1560 |
// Also be sure to handle stuff like: |
| 1561 |
// r13.4<0>.zzzz (would be r13.7<0;1,0>) |
| 1562 |
// |
| 1563 |
// Let's wait until we need this though. |
| 1564 |
} |
| 1565 |
encodeSrcChanSel<S>(chSel[0], chSel[1], chSel[2], chSel[3]); |
| 1566 |
} else { // Align1 |
| 1567 |
if (m_model.srcHasReducedRegion(static_cast<uint32_t>(S))) |
| 1568 |
encodeSrcReducedRegion<S>(src.getRegion()); |
| 1569 |
else |
| 1570 |
encodeSrcRegion<S>(src.getRegion(), true); |
| 1571 |
} |
| 1572 |
break; |
| 1573 |
case Operand::Kind::MACRO: |
| 1574 |
if (accessMode == GED_ACCESS_MODE_Align1) { |
| 1575 |
if (m_model.srcHasReducedRegion(static_cast<uint32_t>(S))) |
| 1576 |
encodeSrcReducedRegion<S>(src.getRegion()); |
| 1577 |
else |
| 1578 |
encodeSrcRegion<S>(src.getRegion()); |
| 1579 |
} // else {align16 macros use the regioning bits, don't clobber them} |
| 1580 |
break; |
| 1581 |
default: |
| 1582 |
break; |
| 1583 |
} |
| 1584 |
} |
| 1585 |
|
| 1586 |
void Encoder::encodeSendDirectDestination(const Operand &dst) { |
| 1587 |
if (platform() >= Platform::XE) { |
| 1588 |
// auto t = dst.getType() == Type::INVALID ? Type::UD : dst.getType(); |
| 1589 |
// GED_ENCODE(DstDataType, lowerDataType(t)); |
| 1590 |
GED_ENCODE(DstRegNum, dst.getDirRegRef().regNum); |
| 1591 |
} else { |
| 1592 |
auto t = dst.getType() == Type::INVALID ? Type::UD : dst.getType(); |
| 1593 |
GED_ENCODE(DstDataType, lowerDataType(t)); |
| 1594 |
|
| 1595 |
// GED_ENCODE(Saturate, lowerSaturate(dst->getDstModifier())); |
| 1596 |
if (m_opcode != Op::SENDS && m_opcode != Op::SENDSC) { |
| 1597 |
GED_ENCODE( |
| 1598 |
DstHorzStride, |
| 1599 |
static_cast<uint32_t>(dst.getRegion().getHz())); // not used for sends |
| 1600 |
} |
| 1601 |
|
| 1602 |
GED_ENCODE(DstRegNum, dst.getDirRegRef().regNum); |
| 1603 |
// GED_ENCODE(DstSubRegNum, |
| 1604 |
// SubRegToBinaryOffset(dst.getDirRegRef().subRegNum, RegName::GRF_R, |
| 1605 |
// dst.getType(), m_model.platform)); |
| 1606 |
} |
| 1607 |
} |
| 1608 |
|
| 1609 |
void Encoder::encodeSendDestinationDataType(const Operand &dst) { |
| 1610 |
if (platform() >= Platform::XE) |
| 1611 |
return; |
| 1612 |
|
| 1613 |
auto t = dst.getType() == Type::INVALID ? Type::UD : dst.getType(); |
| 1614 |
GED_ENCODE(DstDataType, lowerDataType(t)); |
| 1615 |
} |
| 1616 |
|
| 1617 |
void Encoder::encodeSendDestination(const Operand &dst) { |
| 1618 |
if (platform() > Platform::XE3) { |
| 1619 |
if (m_opcode == Op::SENDGX || m_opcode == Op::SENDGXC) { |
| 1620 |
encodeSendgxDestinationXe3p(dst); |
| 1621 |
return; |
| 1622 |
} |
| 1623 |
encodeSendgDestinationXe3p(dst); |
| 1624 |
return; |
| 1625 |
} |
| 1626 |
if (m_model.supportsUnarySend()) { |
| 1627 |
switch (dst.getKind()) { |
| 1628 |
case Operand::Kind::DIRECT: |
| 1629 |
GED_ENCODE(DstAddrMode, GED_ADDR_MODE_Direct); |
| 1630 |
break; |
| 1631 |
case Operand::Kind::INDIRECT: |
| 1632 |
GED_ENCODE(DstAddrMode, GED_ADDR_MODE_Indirect); |
| 1633 |
break; |
| 1634 |
default: |
| 1635 |
fatalT("dst: unsupported destination operand kind/addrMode " |
| 1636 |
"(malformed IR)"); |
| 1637 |
return; |
| 1638 |
} |
| 1639 |
} |
| 1640 |
|
| 1641 |
GED_ENCODE(DstRegFile, lowerRegFile(dst.getDirRegName())); |
| 1642 |
|
| 1643 |
if (dst.getKind() == Operand::Kind::DIRECT) { |
| 1644 |
encodeSendDirectDestination(dst); |
| 1645 |
} else if (dst.getKind() == Operand::Kind::INDIRECT) { |
| 1646 |
encodeSendDestinationDataType(dst); |
| 1647 |
if (m_opcode != Op::SENDS && m_opcode != Op::SENDSC) { |
| 1648 |
GED_ENCODE( |
| 1649 |
DstHorzStride, |
| 1650 |
static_cast<uint32_t>(dst.getRegion().getHz())); // not used for sends |
| 1651 |
} |
| 1652 |
GED_ENCODE(DstAddrImm, dst.getIndImmAddr()); |
| 1653 |
GED_ENCODE(DstAddrSubRegNum, dst.getIndAddrReg().subRegNum); |
| 1654 |
} |
| 1655 |
} |
| 1656 |
void Encoder::encodeSendgxDestinationXe3p(const Operand& dst) { |
| 1657 |
if (dst.getKind() != Operand::Kind::DIRECT) { |
| 1658 |
errorT("invalid dst operand kind"); |
| 1659 |
} |
| 1660 |
if (dst.getDirRegName() == RegName::ARF_NULL) { |
| 1661 |
GED_ENCODE(DstRegNum, (uint16_t)511); |
| 1662 |
} else { |
| 1663 |
GED_ENCODE(DstRegNum, dst.getDirRegRef().regNum); |
| 1664 |
} |
| 1665 |
} |
| 1666 |
void Encoder::encodeSendgDestinationXe3p(const Operand &dst) { |
| 1667 |
if (dst.getKind() != Operand::Kind::DIRECT) { |
| 1668 |
errorT("invalid dst operand kind"); |
| 1669 |
} |
| 1670 |
if (dst.isNull()) { |
| 1671 |
GED_ENCODE(DstRegFile, GED_REG_FILE_ARF); |
| 1672 |
} else { |
| 1673 |
GED_ENCODE(DstRegFile, GED_REG_FILE_GRF); |
| 1674 |
GED_ENCODE(DstRegNum, dst.getDirRegRef().regNum); |
| 1675 |
} |
| 1676 |
} |
| 1677 |
void Encoder::encodeSendgxSource1Xe3p(const Operand& src) { |
| 1678 |
if (src.getDirRegName() == RegName::ARF_NULL) { |
| 1679 |
GED_ENCODE(Src1RegNum, (uint16_t)511); |
| 1680 |
} |
| 1681 |
else { |
| 1682 |
GED_ENCODE(Src1RegNum, src.getDirRegRef().regNum); |
| 1683 |
} |
| 1684 |
} |
| 1685 |
|
| 1686 |
void Encoder::encodeSendgxSource0Xe3p(const Operand& src) { |
| 1687 |
if (src.getDirRegName() == RegName::ARF_NULL) { |
| 1688 |
GED_ENCODE(Src0RegNum, (uint16_t)511); |
| 1689 |
} |
| 1690 |
else { |
| 1691 |
GED_ENCODE(Src0RegNum, src.getDirRegRef().regNum); |
| 1692 |
} |
| 1693 |
} |
| 1694 |
|
| 1695 |
void Encoder::encodeSendSource0(const Operand &src) { |
| 1696 |
if (m_model.supportsUnarySend()) { |
| 1697 |
switch (src.getKind()) { |
| 1698 |
case Operand::Kind::DIRECT: |
| 1699 |
GED_ENCODE(Src0AddrMode, GED_ADDR_MODE_Direct); |
| 1700 |
break; |
| 1701 |
case Operand::Kind::INDIRECT: |
| 1702 |
GED_ENCODE(Src0AddrMode, GED_ADDR_MODE_Indirect); |
| 1703 |
break; |
| 1704 |
default: |
| 1705 |
fatalT("src0: unsupported source operand kind/addrMode " |
| 1706 |
"(malformed IR)"); |
| 1707 |
return; |
| 1708 |
} |
| 1709 |
} |
| 1710 |
|
| 1711 |
GED_REG_FILE gedRegFile = lowerRegFile(src.getDirRegName()); |
| 1712 |
// for sendgx, src0 reg file is repurposed for storing high bit of reg num |
| 1713 |
if (m_opcode == Op::SENDGX || m_opcode == Op::SENDGXC) { |
| 1714 |
return encodeSendgxSource0Xe3p(src); |
| 1715 |
} |
| 1716 |
GED_ENCODE(Src0RegFile, gedRegFile); |
| 1717 |
|
| 1718 |
auto t = src.getType() == Type::INVALID ? Type::UD : src.getType(); |
| 1719 |
|
| 1720 |
if (src.getKind() == Operand::Kind::DIRECT) { |
| 1721 |
if (m_model.supportsXeSend()) { |
| 1722 |
GED_ENCODE(Src0RegNum, src.getDirRegRef().regNum); |
| 1723 |
} else { |
| 1724 |
GED_ENCODE(Src0DataType, lowerDataType(t)); |
| 1725 |
GED_ENCODE(Src0RegNum, src.getDirRegRef().regNum); |
| 1726 |
GED_ENCODE(Src0SubRegNum, src.getDirRegRef().subRegNum); |
| 1727 |
} |
| 1728 |
} else if (src.getKind() == Operand::Kind::INDIRECT) { |
| 1729 |
// Gather Send src0 operand |
| 1730 |
if (src.getDirRegName() == RegName::ARF_S) { |
| 1731 |
// Though src0 is indirect register access, its reg number and subreg |
| 1732 |
// number can be get from "DirRegRef" since it is not a0 access |
| 1733 |
encodeSrcReg<SourceIndex::SRC0>(src.getDirRegName(), |
| 1734 |
src.getDirRegRef().regNum); |
| 1735 |
auto subReg = src.getDirRegRef().subRegNum; |
| 1736 |
if (subReg % 2) |
| 1737 |
errorT("src0 subreg must be 16b aligned"); |
| 1738 |
GED_ENCODE(Src0SubRegNum, subReg); |
| 1739 |
} else { |
| 1740 |
// legacy send indirect src operand |
| 1741 |
GED_ENCODE(Src0DataType, lowerDataType(t)); |
| 1742 |
GED_ENCODE(Src0AddrSubRegNum, src.getIndAddrReg().subRegNum); |
| 1743 |
// For platform >= XeHPC, the ImmAddr is represented in Word Offset in |
| 1744 |
// binary, |
| 1745 |
// platform < XeHPC, the ImmAddr is represented in Byte Offset in |
| 1746 |
// binary |
| 1747 |
// And for all platforms, the ImmAddr is represented in Byte Offset in |
| 1748 |
// assembly syntax |
| 1749 |
if (platform() >= Platform::XE_HPC) { |
| 1750 |
GED_ENCODE(Src0AddrImm, src.getIndImmAddr() / 2); |
| 1751 |
} else { |
| 1752 |
GED_ENCODE(Src0AddrImm, src.getIndImmAddr()); |
| 1753 |
} |
| 1754 |
} |
| 1755 |
} |
| 1756 |
} |
| 1757 |
|
| 1758 |
// The sends opCode exists on gen9+. There is no sends opcode on pre-gen9. |
| 1759 |
// Starting from XE, send opcode can have two sources, so the sends opcode |
| 1760 |
// is not needed. |
| 1761 |
|
| 1762 |
void Encoder::encodeSendsSource0(const Operand &src) { |
| 1763 |
// "...for sends/sendsc instructions Src0.SrcMod, ... and Src0.SrcType are not |
| 1764 |
// used." "Src0.RegFile[1], Src1.RegFile[1] are implicitly set to 0, |
| 1765 |
// and Src0.RegFile[0] is implicitly set as 1 for sends/sendsc instructions." |
| 1766 |
switch (src.getKind()) { |
| 1767 |
case Operand::Kind::DIRECT: |
| 1768 |
GED_ENCODE(Src0AddrMode, GED_ADDR_MODE_Direct); |
| 1769 |
break; |
| 1770 |
case Operand::Kind::INDIRECT: |
| 1771 |
GED_ENCODE(Src0AddrMode, GED_ADDR_MODE_Indirect); |
| 1772 |
break; |
| 1773 |
default: |
| 1774 |
fatalT("src0: unsupported source operand kind/addrMode (malformed IR)"); |
| 1775 |
return; |
| 1776 |
break; |
| 1777 |
} |
| 1778 |
|
| 1779 |
if (src.getKind() == Operand::Kind::DIRECT) { |
| 1780 |
GED_ENCODE(Src0RegNum, src.getDirRegRef().regNum); |
| 1781 |
GED_ENCODE(Src0SubRegNum, src.getDirRegRef().subRegNum); |
| 1782 |
} else if (src.getKind() == Operand::Kind::INDIRECT) { |
| 1783 |
auto immAddr = src.getIndImmAddr(); |
| 1784 |
// For platforms >= XeHPC, ImmAddr is encoded as words, |
| 1785 |
// platforms < XeHPC, ImmAddr is encoded as bytes |
| 1786 |
// For all platforms, ImmAddr is represented in Byte Offset in syntax |
| 1787 |
if (platform() >= Platform::XE_HPC) { |
| 1788 |
immAddr /= 2; |
| 1789 |
} |
| 1790 |
GED_ENCODE(Src0AddrImm, immAddr); |
| 1791 |
GED_ENCODE(Src0AddrSubRegNum, src.getIndAddrReg().subRegNum); |
| 1792 |
} |
| 1793 |
} |
| 1794 |
|
| 1795 |
void Encoder::encodeSendsSource1(const Operand &src) { |
| 1796 |
// GED_ENCODE(Src1AddrMode, GED_ADDR_MODE_Direct); |
| 1797 |
GED_REG_FILE gedRegFile = lowerRegFile(src.getDirRegName()); |
| 1798 |
if (m_opcode == Op::SENDGX || m_opcode == Op::SENDGXC) { |
| 1799 |
// for sendgx, src1 reg file is repurposed for storing high bit of reg num |
| 1800 |
return encodeSendgxSource1Xe3p(src); |
| 1801 |
} |
| 1802 |
GED_ENCODE(Src1RegFile, gedRegFile); |
| 1803 |
if (gedRegFile == GED_REG_FILE_GRF) |
| 1804 |
GED_ENCODE(Src1RegNum, src.getDirRegRef().regNum); |
| 1805 |
} |
| 1806 |
|
| 1807 |
void Encoder::encodeSendsDestination(const Operand &dst) { |
| 1808 |
GED_ENCODE(DstAddrMode, GED_ADDR_MODE_Direct); |
| 1809 |
GED_ENCODE(DstRegFile, lowerRegFile(dst.getDirRegName())); |
| 1810 |
// send types use :ud where possible |
| 1811 |
auto t = dst.getType() == Type::INVALID ? Type::UD : dst.getType(); |
| 1812 |
GED_ENCODE(DstDataType, lowerDataType(t)); |
| 1813 |
|
| 1814 |
// GED_ENCODE(Saturate, lowerSaturate(dst->getDstModifier())); |
| 1815 |
// GED_ENCODE(DstHorzStride, static_cast<uint32_t>(dst->getHz())); |
| 1816 |
|
| 1817 |
GED_ENCODE(DstRegNum, dst.getDirRegRef().regNum); |
| 1818 |
// TODO: set correct regType |
| 1819 |
GED_ENCODE(DstSubRegNum, |
| 1820 |
SubRegToBinaryOffset(dst.getDirRegRef().subRegNum, RegName::GRF_R, |
| 1821 |
dst.getType(), m_model.platform)); |
| 1822 |
} |
| 1823 |
|
| 1824 |
template <SourceIndex S> |
| 1825 |
void Encoder::encodeTernarySourceAlign16(const Instruction &inst) { |
| 1826 |
// PreCNL Align16 |
| 1827 |
// GRF-only |
| 1828 |
encodeSrcAddrMode<S>(GED_ADDR_MODE_Direct); |
| 1829 |
|
| 1830 |
const Operand &src = inst.getSource(S); |
| 1831 |
|
| 1832 |
if (inst.getOpSpec().supportsSourceModifiers()) { |
| 1833 |
encodeSrcModifier<S>(src.getSrcModifier()); |
| 1834 |
} |
| 1835 |
|
| 1836 |
// set the data type |
| 1837 |
GED_DATA_TYPE gedType = lowerDataType(src.getType()); |
| 1838 |
if (S == SourceIndex::SRC0) { |
| 1839 |
GED_ENCODE(SrcDataType, gedType); |
| 1840 |
} else { |
| 1841 |
const Operand &src0 = inst.getSource(SourceIndex::SRC0); |
| 1842 |
bool src0IsFloating = |
| 1843 |
src0.getType() == Type::F || src0.getType() == Type::HF; |
| 1844 |
if (platform() >= Platform::GEN8LP && src0IsFloating) { |
| 1845 |
bool srcNIsFloating = |
| 1846 |
src.getType() == Type::F || src.getType() == Type::HF; |
| 1847 |
if (src0IsFloating && srcNIsFloating) { |
| 1848 |
encodeSrcType<S>(src.getType()); |
| 1849 |
} else { |
| 1850 |
fatalT("src", (int)S, |
| 1851 |
": mixed types require :f and :hf " |
| 1852 |
"(or vice versa)"); |
| 1853 |
return; |
| 1854 |
} |
| 1855 |
} |
| 1856 |
} |
| 1857 |
|
| 1858 |
if (!inst.isMacro()) { |
| 1859 |
const Region &rgn = src.getRegion(); |
| 1860 |
const RegRef ® = src.getDirRegRef(); |
| 1861 |
// Adjusting sub register when going from align1 to align16 representation. |
| 1862 |
// in align 16 subregister is always 16 byte alligned, but we can play |
| 1863 |
// with swizzle to access none aligned sub register |
| 1864 |
uint16_t subRegNumber = reg.subRegNum; |
| 1865 |
// mad (8) r46.0.xyzw:df r46.0.xyzw:df r50.0.xyzw:df r48.0.xyzw:df {Align16, |
| 1866 |
// Q1} mad (2) r5.0.xy:df r5.0.xyxy:df r92.2.xyxy:df r93.0.xyxy:df {Align16, |
| 1867 |
// Q1, NoMask} // BDW,SKL |
| 1868 |
if (S != SourceIndex::SRC2) { |
| 1869 |
if (rgn == Region::SRC8X1 || rgn == Region::SRC4X1 || |
| 1870 |
rgn == Region::SRC2X1) { |
| 1871 |
encodeSrcRepCtrl<S>(GED_REP_CTRL_NoRep); |
| 1872 |
encodeSrcChanSel<S>(GED_SWIZZLE_x, GED_SWIZZLE_y, GED_SWIZZLE_z, |
| 1873 |
GED_SWIZZLE_w); |
| 1874 |
} else if (rgn == Region::SRC0X0) { |
| 1875 |
if (src.getType() == Type::DF) { |
| 1876 |
if (reg.subRegNum % 2 == 0) { |
| 1877 |
encodeSrcChanSel<S>(GED_SWIZZLE_x, GED_SWIZZLE_y, GED_SWIZZLE_x, |
| 1878 |
GED_SWIZZLE_y); |
| 1879 |
} else { |
| 1880 |
encodeSrcChanSel<S>(GED_SWIZZLE_z, GED_SWIZZLE_w, GED_SWIZZLE_z, |
| 1881 |
GED_SWIZZLE_w); |
| 1882 |
subRegNumber -= 1; |
| 1883 |
} |
| 1884 |
} else { |
| 1885 |
encodeSrcRepCtrl<S>(GED_REP_CTRL_Rep); |
| 1886 |
} |
| 1887 |
} else { |
| 1888 |
fatalT("src", (int)S, ": unsupported region for Align16 encoding"); |
| 1889 |
return; |
| 1890 |
} |
| 1891 |
} else { |
| 1892 |
if (rgn == Region::SRCXX1) { |
| 1893 |
encodeSrcRepCtrl<S>(GED_REP_CTRL_NoRep); |
| 1894 |
encodeSrcChanSel<S>(GED_SWIZZLE_x, GED_SWIZZLE_y, GED_SWIZZLE_z, |
| 1895 |
GED_SWIZZLE_w); |
| 1896 |
} else if (rgn == Region::SRCXX0) { |
| 1897 |
if (src.getType() == Type::DF) { |
| 1898 |
if (src.getDirRegRef().subRegNum % 2 == 0) { |
| 1899 |
encodeSrcChanSel<S>(GED_SWIZZLE_x, GED_SWIZZLE_y, GED_SWIZZLE_x, |
| 1900 |
GED_SWIZZLE_y); |
| 1901 |
} else { |
| 1902 |
encodeSrcChanSel<S>(GED_SWIZZLE_z, GED_SWIZZLE_w, GED_SWIZZLE_z, |
| 1903 |
GED_SWIZZLE_w); |
| 1904 |
subRegNumber -= 1; |
| 1905 |
} |
| 1906 |
} else { |
| 1907 |
encodeSrcRepCtrl<S>(GED_REP_CTRL_Rep); |
| 1908 |
} |
| 1909 |
} else if (rgn == Region::SRC0X0 && src.getType() == Type::DF) { |
| 1910 |
encodeSrcChanSel<S>(GED_SWIZZLE_x, GED_SWIZZLE_y, GED_SWIZZLE_x, |
| 1911 |
GED_SWIZZLE_y); |
| 1912 |
} else { |
| 1913 |
fatalT("src", (int)S, ": unsupported region for Align16 encoding"); |
| 1914 |
return; |
| 1915 |
} |
| 1916 |
} |
| 1917 |
uint32_t regNum = reg.regNum; |
| 1918 |
encodeSrcReg<S>(RegName::GRF_R, (uint16_t)regNum); |
| 1919 |
auto subReg = SubRegToBinaryOffset(subRegNumber, src.getDirRegName(), |
| 1920 |
src.getType(), m_model.platform); |
| 1921 |
encodeSrcSubRegNum<S>(subReg); |
| 1922 |
} else { |
| 1923 |
// implicit operand accumulator |
| 1924 |
// e.g. madm (4) ... -r14.acc3 |
| 1925 |
encodeSrcReg<S>(RegName::GRF_R, src.getDirRegRef().regNum); |
| 1926 |
encodeSrcMathMacroReg<S>(src.getMathMacroExt()); |
| 1927 |
} |
| 1928 |
} |
| 1929 |
|
| 1930 |
void Encoder::encodeTernaryDestinationAlign16(const Instruction &inst) { |
| 1931 |
const Operand &dst = inst.getDestination(); |
| 1932 |
if (inst.getOpSpec().supportsSaturation()) { |
| 1933 |
GED_ENCODE(Saturate, lowerSaturate(dst.getDstModifier())); |
| 1934 |
} |
| 1935 |
GED_ENCODE(DstDataType, lowerDataType(dst.getType())); |
| 1936 |
if (dst.getDirRegName() != RegName::GRF_R) { |
| 1937 |
fatalT("align16 ternary dst must be to GRF"); |
| 1938 |
return; |
| 1939 |
} |
| 1940 |
|
| 1941 |
// register / info (must be GRF) |
| 1942 |
GED_ENCODE(DstRegFile, lowerRegFile(dst.getDirRegName())); |
| 1943 |
uint32_t regNum = dst.getDirRegRef().regNum; |
| 1944 |
GED_ENCODE(DstRegNum, regNum); |
| 1945 |
if (inst.isMacro()) { |
| 1946 |
// macro only |
| 1947 |
GED_DST_CHAN_EN chanEn = mathMacroRegToChEn(dst.getMathMacroExt()); |
| 1948 |
GED_ENCODE(DstChanEn, chanEn); |
| 1949 |
} else { |
| 1950 |
// Align16 instruction (we must convert from Align1) |
| 1951 |
// |
| 1952 |
// As long as the Align1 sequences are packed (.xyzw), this is |
| 1953 |
// straightforward. However, "scalar" (braoadcast) sequences are |
| 1954 |
// a bit harder as we must carefully choose the ChEn based on the |
| 1955 |
// subregister that would be used in Align1 |
| 1956 |
// (See also Decoder::decodeDestinationTernaryAlign16) |
| 1957 |
GED_DST_CHAN_EN chanEn = GED_DST_CHAN_EN_xyzw; |
| 1958 |
auto reg = dst.getDirRegRef(); |
| 1959 |
if (inst.getExecSize() == ExecSize::SIMD1) { |
| 1960 |
// SIMD1 MAD is not allowed, so MDF (and IGC) are generating use |
| 1961 |
// SIMD4 and SIMD2 with specific channel masks to selectively |
| 1962 |
// enable just the bottom channel. |
| 1963 |
if (dst.getType() == Type::DF) { |
| 1964 |
// For 64-bit types we use a mad (2) ... |
| 1965 |
// Note, only :df is needed since :q and :uq are not supported |
| 1966 |
// |
| 1967 |
// e.g. mad (2) r5.0.xy:df ... {Align16, Q1, NoMask} // |
| 1968 |
if (dst.getDirRegRef().subRegNum % 2 == 0) { |
| 1969 |
chanEn = GED_DST_CHAN_EN_xy; |
| 1970 |
} else { |
| 1971 |
// e.g. mad (1) r5.1<1>:df |
| 1972 |
// encodes as |
| 1973 |
// mad (2) r5.0.zw:df |
| 1974 |
// ^ SIMD2 and .zw (~= .1) |
| 1975 |
chanEn = GED_DST_CHAN_EN_zw; |
| 1976 |
reg.subRegNum -= 1; |
| 1977 |
} |
| 1978 |
} else { |
| 1979 |
// 32-bit or 16-bit type (:hf). We use a SIMD4 |
| 1980 |
// |
| 1981 |
// one channel enabled. E.g. we'll parse |
| 1982 |
// mad (1|M0) r53.6<1>:f ... |
| 1983 |
// and encode it as |
| 1984 |
// mad (4) r53.4.z:f |
| 1985 |
// ^ SIMD4 ^^^ aligned subreg .4.z == subreg .6:f Align1 |
| 1986 |
switch (reg.subRegNum % 4) { |
| 1987 |
case 0: |
| 1988 |
chanEn = GED_DST_CHAN_EN_x; |
| 1989 |
break; |
| 1990 |
case 1: |
| 1991 |
chanEn = GED_DST_CHAN_EN_y; |
| 1992 |
break; |
| 1993 |
case 2: |
| 1994 |
chanEn = GED_DST_CHAN_EN_z; |
| 1995 |
break; |
| 1996 |
case 3: |
| 1997 |
chanEn = GED_DST_CHAN_EN_w; |
| 1998 |
break; |
| 1999 |
} |
| 2000 |
// align the subregister |
| 2001 |
reg.subRegNum -= (reg.subRegNum % 4); |
| 2002 |
} |
| 2003 |
} |
| 2004 |
GED_ENCODE(DstChanEn, chanEn); |
| 2005 |
GED_ENCODE(DstSubRegNum, |
| 2006 |
SubRegToBinaryOffset(reg.subRegNum, dst.getDirRegName(), |
| 2007 |
dst.getType(), m_model.platform)); |
| 2008 |
} |
| 2009 |
} |
| 2010 |
|
| 2011 |
void Encoder::encodeDstReg(RegName regName, uint16_t regNum) { |
| 2012 |
// encodes ARF or GRF |
| 2013 |
uint32_t gedBits = translateRegNum(-1, regName, regNum); |
| 2014 |
GED_ENCODE(DstRegNum, gedBits); |
| 2015 |
} |
| 2016 |
|
| 2017 |
void Encoder::encodeImmVal(const ImmVal &val, Type type) { |
| 2018 |
GED_ENCODE(Imm, typeConvesionHelper(val, type)); |
| 2019 |
} |
| 2020 |
|
| 2021 |
template <SourceIndex S> void Encoder::encodeSrcRepCtrl(GED_REP_CTRL rep) { |
| 2022 |
if (S == SourceIndex::SRC0) { |
| 2023 |
GED_ENCODE(Src0RepCtrl, rep); |
| 2024 |
} else if (S == SourceIndex::SRC1) { |
| 2025 |
GED_ENCODE(Src1RepCtrl, rep); |
| 2026 |
} else { |
| 2027 |
GED_ENCODE(Src2RepCtrl, rep); |
| 2028 |
} |
| 2029 |
} |
| 2030 |
|
| 2031 |
template <SourceIndex S> |
| 2032 |
void Encoder::encodeSrcChanSel(GED_SWIZZLE chSelX, GED_SWIZZLE chSelY, |
| 2033 |
GED_SWIZZLE chSelZ, GED_SWIZZLE chSelW) { |
| 2034 |
uint32_t chSelBits = createChanSel(chSelX, chSelY, chSelZ, chSelW); |
| 2035 |
if (S == SourceIndex::SRC0) { |
| 2036 |
GED_ENCODE(Src0ChanSel, chSelBits); |
| 2037 |
} else if (S == SourceIndex::SRC1) { |
| 2038 |
GED_ENCODE(Src1ChanSel, chSelBits); |
| 2039 |
} else { |
| 2040 |
GED_ENCODE(Src2ChanSel, chSelBits); |
| 2041 |
} |
| 2042 |
} |
| 2043 |
|
| 2044 |
uint32_t Encoder::translateRegNum(int opIx, RegName regName, uint16_t regNum) { |
| 2045 |
const char *whichOp = opIx == 0 ? "src0" |
| 2046 |
: opIx == 1 ? "src1" |
| 2047 |
: opIx == 2 ? "src2" |
| 2048 |
: "dst"; |
| 2049 |
|
| 2050 |
const RegInfo *ri = m_model.lookupRegInfoByRegName(regName); |
| 2051 |
if (ri == nullptr) { |
| 2052 |
errorT(whichOp, ": invalid register name for this platform"); |
| 2053 |
} else if (!ri->isRegNumberValid((int)regNum)) { |
| 2054 |
errorT(whichOp, ": ", ri->syntax, regNum, " number out of range"); |
| 2055 |
} else if (regName != RegName::GRF_R) { |
| 2056 |
return ri->encodeARFRegNum((int)regNum); |
| 2057 |
} |
| 2058 |
return regNum; // widen for GED |
| 2059 |
} |
| 2060 |
|
| 2061 |
uint32_t Encoder::mathMacroRegToBits(int src, MathMacroExt implAcc) { |
| 2062 |
uint32_t bits = 8; // NOACC |
| 2063 |
switch (implAcc) { |
| 2064 |
/// or 00000000b (GEN11) |
| 2065 |
case MathMacroExt::MME0: |
| 2066 |
bits = 0; |
| 2067 |
break; // 0000b |
| 2068 |
case MathMacroExt::MME1: |
| 2069 |
bits = 1; |
| 2070 |
break; |
| 2071 |
case MathMacroExt::MME2: |
| 2072 |
bits = 2; |
| 2073 |
break; |
| 2074 |
case MathMacroExt::MME3: |
| 2075 |
bits = 3; |
| 2076 |
break; |
| 2077 |
case MathMacroExt::MME4: |
| 2078 |
bits = 4; |
| 2079 |
break; |
| 2080 |
case MathMacroExt::MME5: |
| 2081 |
bits = 5; |
| 2082 |
break; |
| 2083 |
case MathMacroExt::MME6: |
| 2084 |
bits = 6; |
| 2085 |
break; |
| 2086 |
case MathMacroExt::MME7: |
| 2087 |
bits = 7; |
| 2088 |
break; |
| 2089 |
/// or 00008000b (GEN11) |
| 2090 |
case MathMacroExt::NOMME: |
| 2091 |
bits = 8; |
| 2092 |
break; // 1000b |
| 2093 |
default: |
| 2094 |
if (src < 0) { |
| 2095 |
fatalT("dst operand has invalid math macro register"); |
| 2096 |
} else { |
| 2097 |
fatalT("src", src, " operand has invalid math macro register"); |
| 2098 |
} |
| 2099 |
return bits; |
| 2100 |
} |
| 2101 |
return bits; |
| 2102 |
} |
| 2103 |
GED_DST_CHAN_EN Encoder::mathMacroRegToChEn(MathMacroExt implAcc) { |
| 2104 |
GED_DST_CHAN_EN bits = GED_DST_CHAN_EN_w; // NOACC |
| 2105 |
switch (implAcc) { |
| 2106 |
case MathMacroExt::MME0: |
| 2107 |
bits = GED_DST_CHAN_EN_None; |
| 2108 |
break; // 0000b |
| 2109 |
case MathMacroExt::MME1: |
| 2110 |
bits = GED_DST_CHAN_EN_x; |
| 2111 |
break; |
| 2112 |
case MathMacroExt::MME2: |
| 2113 |
bits = GED_DST_CHAN_EN_y; |
| 2114 |
break; |
| 2115 |
case MathMacroExt::MME3: |
| 2116 |
bits = GED_DST_CHAN_EN_xy; |
| 2117 |
break; |
| 2118 |
case MathMacroExt::MME4: |
| 2119 |
bits = GED_DST_CHAN_EN_z; |
| 2120 |
break; // 0100b |
| 2121 |
case MathMacroExt::MME5: |
| 2122 |
bits = GED_DST_CHAN_EN_xz; |
| 2123 |
break; |
| 2124 |
case MathMacroExt::MME6: |
| 2125 |
bits = GED_DST_CHAN_EN_yz; |
| 2126 |
break; |
| 2127 |
case MathMacroExt::MME7: |
| 2128 |
bits = GED_DST_CHAN_EN_xyz; |
| 2129 |
break; |
| 2130 |
case MathMacroExt::NOMME: |
| 2131 |
bits = GED_DST_CHAN_EN_w; |
| 2132 |
break; // 1000b |
| 2133 |
default: |
| 2134 |
fatalT("operand has invalid math macro register"); |
| 2135 |
} |
| 2136 |
return bits; |
| 2137 |
} |
| 2138 |
|
| 2139 |
void Encoder::encodeOptionsThreadControl(const Instruction &inst) { |
| 2140 |
if (inst.hasInstOpt(InstOpt::NOPREEMPT)) { |
| 2141 |
if (m_model.supportsNoPreempt()) { |
| 2142 |
GED_ENCODE(ThreadCtrl, GED_THREAD_CTRL_NoPreempt); |
| 2143 |
} else { |
| 2144 |
warningT("NoPreempt not supported on this platform (dropping)"); |
| 2145 |
} |
| 2146 |
} |
| 2147 |
} |
| 2148 |
|
| 2149 |
void Encoder::encodeOptions(const Instruction &inst) { |
| 2150 |
GED_ENCODE(DebugCtrl, inst.hasInstOpt(InstOpt::BREAKPOINT) |
| 2151 |
? GED_DEBUG_CTRL_Breakpoint |
| 2152 |
: GED_DEBUG_CTRL_Normal); |
| 2153 |
|
| 2154 |
auto &os = inst.getOpSpec(); |
| 2155 |
if (os.supportsDepCtrl()) { |
| 2156 |
if (inst.hasInstOpt(InstOpt::NODDCHK) && |
| 2157 |
!inst.hasInstOpt(InstOpt::NODDCLR)) { |
| 2158 |
GED_ENCODE(DepCtrl, GED_DEP_CTRL_NoDDChk); |
| 2159 |
} else if (!inst.hasInstOpt(InstOpt::NODDCHK) && |
| 2160 |
inst.hasInstOpt(InstOpt::NODDCLR)) { |
| 2161 |
GED_ENCODE(DepCtrl, GED_DEP_CTRL_NoDDClr); |
| 2162 |
} else if (inst.hasInstOpt(InstOpt::NODDCHK) && |
| 2163 |
inst.hasInstOpt(InstOpt::NODDCLR)) { |
| 2164 |
GED_ENCODE(DepCtrl, GED_DEP_CTRL_NoDDClr_NoDDChk); |
| 2165 |
} else if (!inst.getOpSpec().isAnySendFormat() && inst.getOp() != Op::NOP) { |
| 2166 |
GED_ENCODE(DepCtrl, GED_DEP_CTRL_Normal); |
| 2167 |
} |
| 2168 |
} |
| 2169 |
|
| 2170 |
if (inst.hasInstOpt(InstOpt::ATOMIC)) { |
| 2171 |
// this bit overlaps |
| 2172 |
if (os.isAnySendFormat() && platform() >= Platform::XE3) { |
| 2173 |
errorT("{Atomic} not available on sends for this platform"); |
| 2174 |
} |
| 2175 |
GED_ENCODE(ThreadCtrl, GED_THREAD_CTRL_Atomic); |
| 2176 |
} |
| 2177 |
|
| 2178 |
if (inst.hasInstOpt(InstOpt::FWD)) { |
| 2179 |
if (!os.supportsFwdCtrl()) |
| 2180 |
errorT("Invalid option {Fwd}"); |
| 2181 |
GED_ENCODE(FwdCtrl, 1); |
| 2182 |
} |
| 2183 |
|
| 2184 |
if (inst.hasInstOpt(InstOpt::SWITCH) && m_model.supportsHwDeps()) { |
| 2185 |
if (inst.getOp() == Op::NOP) { |
| 2186 |
warningT("nop doesn't support Switch option (dropping)"); |
| 2187 |
} else { |
| 2188 |
GED_ENCODE(ThreadCtrl, GED_THREAD_CTRL_Switch); |
| 2189 |
} |
| 2190 |
} |
| 2191 |
encodeOptionsThreadControl(inst); |
| 2192 |
|
| 2193 |
if (!inst.hasInstOpt(InstOpt::ATOMIC) && !inst.hasInstOpt(InstOpt::SWITCH) && |
| 2194 |
!inst.hasInstOpt(InstOpt::NOPREEMPT) && |
| 2195 |
!inst.getOpSpec().isAnySendFormat() && inst.getOp() != Op::NOP) { |
| 2196 |
GED_ENCODE(ThreadCtrl, GED_THREAD_CTRL_Normal); |
| 2197 |
} |
| 2198 |
|
| 2199 |
if (inst.hasInstOpt(InstOpt::NOSRCDEPSET)) { |
| 2200 |
GED_ENCODE(NoSrcDepSet, GED_NO_SRC_DEP_SET_NoSrcDepSet); |
| 2201 |
} else if (inst.getOpSpec().isAnySendFormat() && |
| 2202 |
m_model.supportNoSrcDepSet()) { |
| 2203 |
GED_ENCODE(NoSrcDepSet, GED_NO_SRC_DEP_SET_Normal); |
| 2204 |
} |
| 2205 |
|
| 2206 |
if (platform() >= Platform::XE && m_opcode != Op::ILLEGAL && |
| 2207 |
m_opcode != Op::THRYLD) { |
| 2208 |
SWSB::InstType inst_type = inst.getSWSBInstType(m_opts.swsbEncodeMode); |
| 2209 |
uint32_t swsbBinary = |
| 2210 |
inst.getSWSB().encode(m_opts.swsbEncodeMode, inst_type); |
| 2211 |
IGA_ASSERT(inst.getSWSB().verify(m_opts.swsbEncodeMode, inst_type), |
| 2212 |
"INTERNAL ERROR: invalid SWSB (parser/IR-creator should have " |
| 2213 |
"prevented this)"); |
| 2214 |
|
| 2215 |
GED_ENCODE(SWSB, swsbBinary); |
| 2216 |
} |
| 2217 |
} |
| 2218 |
|
| 2219 |
void Encoder::patchJumpOffsets() { |
| 2220 |
for (JumpPatch &jp : m_needToPatch) { |
| 2221 |
const Instruction *inst = jp.inst; |
| 2222 |
IGA_ASSERT(inst->getOpSpec().isBranching() || inst->isMovWithLabel(), |
| 2223 |
"patching non-control-flow/non-mov instruction"); |
| 2224 |
|
| 2225 |
// on some platforms jmpi os post-increment |
| 2226 |
uint32_t jmpiExtraOffset = 0; |
| 2227 |
bool isPostIncrementJmpi = |
| 2228 |
inst->getOp() == Op::JMPI && !m_model.supportsSimplifiedBranches(); |
| 2229 |
if (isPostIncrementJmpi) { |
| 2230 |
// jmpi is relative to the incremented PC, hence we must add |
| 2231 |
// the size of the instruction here. jmpi probably will never |
| 2232 |
// compact, but we'll be careful here |
| 2233 |
jmpiExtraOffset = inst->hasInstOpt(InstOpt::COMPACTED) ? 8 : 16; |
| 2234 |
IGA_ASSERT(inst->getSource(0).getKind() == Operand::Kind::LABEL, |
| 2235 |
"patching non label op"); |
| 2236 |
// skip registers |
| 2237 |
} |
| 2238 |
|
| 2239 |
// calla and mov is an absolute offset |
| 2240 |
uint32_t encodePC = |
| 2241 |
(inst->getOpSpec().isJipAbsolute()) || (inst->getOp() == Op::MOV) |
| 2242 |
? 0 |
| 2243 |
: getEncodedPC(inst); |
| 2244 |
|
| 2245 |
uint32_t jumpPC = 0; |
| 2246 |
const Block *jipBlk = inst->getJIP(); |
| 2247 |
if (jipBlk == nullptr) { |
| 2248 |
// immediate offset: we have to treat this as a relative offset |
| 2249 |
jumpPC = inst->getSource(0).getImmediateValue().s32 + encodePC; |
| 2250 |
} else if (!getBlockOffset(jipBlk, jumpPC)) { |
| 2251 |
// For call, its target symbol may not be resolvable until in the |
| 2252 |
// link stage when other kernels are available. |
| 2253 |
if (inst->getOp() != Op::CALL && inst->getOp() != Op::CALLA) { |
| 2254 |
fatalAtT(inst->getLoc(), "jip label invalid"); |
| 2255 |
} |
| 2256 |
} |
| 2257 |
|
| 2258 |
int32_t jip = jumpPC - encodePC - jmpiExtraOffset; |
| 2259 |
// JIP and UIP are in QWORDS for most ops on PreBDW |
| 2260 |
int32_t pcUnscale = arePcsInQWords(inst->getOpSpec()) ? 8 : 1; |
| 2261 |
|
| 2262 |
if (inst->isMovWithLabel()) { |
| 2263 |
// encode mov label |
| 2264 |
GED_DATA_TYPE src0_ty = lowerDataType(inst->getSource(0).getType()); |
| 2265 |
GED_ENCODE_TO(Src0DataType, src0_ty, &jp.gedInst); |
| 2266 |
GED_ENCODE_TO(Imm, jip, &jp.gedInst); |
| 2267 |
} else { |
| 2268 |
// encode other branch instructions |
| 2269 |
GED_ENCODE_TO(JIP, jip / pcUnscale, &jp.gedInst); |
| 2270 |
} |
| 2271 |
|
| 2272 |
if (inst->getSourceCount() == 2 && |
| 2273 |
(inst->getOp() != Op::BRC || inst->getSource(1).isImm())) { |
| 2274 |
// No need to set src1 regFile and type, |
| 2275 |
// it will be over written by UIP |
| 2276 |
const Block *uipBlk = inst->getUIP(); |
| 2277 |
if (uipBlk == nullptr) { |
| 2278 |
jumpPC = inst->getSource(1).getImmediateValue().s32 + encodePC; |
| 2279 |
} else if (!getBlockOffset(uipBlk, jumpPC)) { |
| 2280 |
fatalAtT(inst->getLoc(), "uip label invalid"); |
| 2281 |
} |
| 2282 |
encodePC = getEncodedPC(inst); |
| 2283 |
int32_t uip = jumpPC - encodePC; |
| 2284 |
GED_ENCODE_TO(UIP, uip / pcUnscale, &jp.gedInst); |
| 2285 |
} |
| 2286 |
|
| 2287 |
// re-encode branch |
| 2288 |
START_GED_TIMER(); |
| 2289 |
GED_RETURN_VALUE status = GED_EncodeIns(&jp.gedInst, |
| 2290 |
inst->hasInstOpt(InstOpt::COMPACTED) |
| 2291 |
? GED_INS_TYPE_COMPACT |
| 2292 |
: GED_INS_TYPE_NATIVE, |
| 2293 |
jp.bits); |
| 2294 |
STOP_GED_TIMER(); |
| 2295 |
if (status != GED_RETURN_VALUE_SUCCESS) { |
| 2296 |
fatalAtT(inst->getLoc(), |
| 2297 |
"GED_EncodeIns failed: ", gedReturnValueToString(status)); |
| 2298 |
} |
| 2299 |
} |
| 2300 |
} |
| 2301 |
|
| 2302 |
bool Encoder::arePcsInQWords(const OpSpec &os) const { |
| 2303 |
// everything is in bytes except: |
| 2304 |
// HSW calla, call, and jmpi |
| 2305 |
return platform() < Platform::GEN8 && os.op != Op::JMPI && |
| 2306 |
os.op != Op::CALL && os.op != Op::CALLA; |
| 2307 |
} |
| 2308 |
|
| 2309 |
bool Encoder::callNeedsSrc0Region221(const Instruction &inst) const { |
| 2310 |
// [call]: "Restriction: The src0 regioning control must be <2;2,1>" |
| 2311 |
// [calla]: "Restriction: The src0 regioning control must be <2;2,1>" |
| 2312 |
return (inst.getOp() == Op::CALL && platform() < Platform::GEN8) || |
| 2313 |
(inst.getOp() == Op::CALL && platform() == Platform::GEN9) || |
| 2314 |
(inst.getOp() == Op::CALLA && platform() <= Platform::GEN10); |
| 2315 |
} |
| 2316 |
|
| 2317 |
bool Encoder::callNeedsSrc0Region241(const Instruction &inst) const { |
| 2318 |
return (inst.getOp() == Op::CALL && platform() == Platform::GEN11); |
| 2319 |
} |
| 2320 |
|
| 2321 |
void Encoder::encodeTernarySrcRegionVert(SourceIndex S, Region::Vert v) { |
| 2322 |
if (S == SourceIndex::SRC0) { |
| 2323 |
GED_ENCODE(Src0VertStride, lowerRegionVert(v)); |
| 2324 |
} else { // (S == SourceIndex::SRC1) |
| 2325 |
GED_ENCODE(Src1VertStride, lowerRegionVert(v)); |
| 2326 |
} // S != SRC2 since ternary Align1 doesn't have bits for that |
| 2327 |
} |
| 2328 |
|
| 2329 |
// fixes stuff where GED just ignores or where it refuses to allow us to |
| 2330 |
// set bits. This should be empty unless GED fixes are in flight. |
| 2331 |
void Encoder::applyGedWorkarounds(const Kernel &, size_t) { |
| 2332 |
// Also consider |
| 2333 |
// GED_RETURN_VALUE GED_SetRawBits(ged_ins_t* ins, uint8_t low, uint8_t high, |
| 2334 |
// const uint64_t value) |
| 2335 |
} |