1 //===-- AMDGPUAsmPrinter.cpp - AMDGPU Assebly printer --------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
12 /// The AMDGPUAsmPrinter is used to print both assembly string and also binary
13 /// code. When passed an MCAsmStreamer it prints assembly and when passed
14 /// an MCObjectStreamer it outputs binary code.
16 //===----------------------------------------------------------------------===//
19 #include "AMDGPUAsmPrinter.h"
20 #include "MCTargetDesc/AMDGPUTargetStreamer.h"
21 #include "InstPrinter/AMDGPUInstPrinter.h"
22 #include "Utils/AMDGPUBaseInfo.h"
24 #include "AMDKernelCodeT.h"
25 #include "AMDGPUSubtarget.h"
26 #include "R600Defines.h"
27 #include "R600MachineFunctionInfo.h"
28 #include "R600RegisterInfo.h"
29 #include "SIDefines.h"
30 #include "SIMachineFunctionInfo.h"
31 #include "SIRegisterInfo.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/MC/MCContext.h"
34 #include "llvm/MC/MCSectionELF.h"
35 #include "llvm/MC/MCStreamer.h"
36 #include "llvm/Support/ELF.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/TargetRegistry.h"
39 #include "llvm/Target/TargetLoweringObjectFile.h"
43 // TODO: This should get the default rounding mode from the kernel. We just set
44 // the default here, but this could change if the OpenCL rounding mode pragmas
47 // The denormal mode here should match what is reported by the OpenCL runtime
48 // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but
49 // can also be override to flush with the -cl-denorms-are-zero compiler flag.
51 // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double
52 // precision, and leaves single precision to flush all and does not report
53 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
54 // CL_FP_DENORM for both.
56 // FIXME: It seems some instructions do not support single precision denormals
57 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
58 // and sin_f32, cos_f32 on most parts).
60 // We want to use these instructions, and using fp32 denormals also causes
61 // instructions to run at the double precision rate for the device so it's
62 // probably best to just report no single precision denormals.
63 static uint32_t getFPMode(const MachineFunction &F) {
64 const AMDGPUSubtarget& ST = F.getSubtarget<AMDGPUSubtarget>();
65 // TODO: Is there any real use for the flush in only / flush out only modes?
67 uint32_t FP32Denormals =
68 ST.hasFP32Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
70 uint32_t FP64Denormals =
71 ST.hasFP64Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
73 return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
74 FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
75 FP_DENORM_MODE_SP(FP32Denormals) |
76 FP_DENORM_MODE_DP(FP64Denormals);
80 createAMDGPUAsmPrinterPass(TargetMachine &tm,
81 std::unique_ptr<MCStreamer> &&Streamer) {
82 return new AMDGPUAsmPrinter(tm, std::move(Streamer));
85 extern "C" void LLVMInitializeAMDGPUAsmPrinter() {
86 TargetRegistry::RegisterAsmPrinter(TheAMDGPUTarget, createAMDGPUAsmPrinterPass);
87 TargetRegistry::RegisterAsmPrinter(TheGCNTarget, createAMDGPUAsmPrinterPass);
90 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM,
91 std::unique_ptr<MCStreamer> Streamer)
92 : AsmPrinter(TM, std::move(Streamer)) {}
94 void AMDGPUAsmPrinter::EmitFunctionBodyStart() {
95 const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
96 SIProgramInfo KernelInfo;
97 if (STM.isAmdHsaOS()) {
98 getSIProgramInfo(KernelInfo, *MF);
99 EmitAmdKernelCodeT(*MF, KernelInfo);
103 void AMDGPUAsmPrinter::EmitFunctionEntryLabel() {
104 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
105 const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
106 if (MFI->isKernel() && STM.isAmdHsaOS()) {
107 AMDGPUTargetStreamer *TS =
108 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
109 TS->EmitAMDGPUSymbolType(CurrentFnSym->getName(),
110 ELF::STT_AMDGPU_HSA_KERNEL);
113 AsmPrinter::EmitFunctionEntryLabel();
116 static bool isModuleLinkage(const GlobalValue *GV) {
117 switch (GV->getLinkage()) {
118 case GlobalValue::InternalLinkage:
119 case GlobalValue::CommonLinkage:
121 case GlobalValue::ExternalLinkage:
123 default: llvm_unreachable("unknown linkage type");
127 void AMDGPUAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
129 if (TM.getTargetTriple().getOS() != Triple::AMDHSA ||
130 GV->isDeclaration()) {
131 AsmPrinter::EmitGlobalVariable(GV);
135 // Group segment variables aren't emitted in HSA.
136 if (AMDGPU::isGroupSegment(GV))
139 AMDGPUTargetStreamer *TS =
140 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
141 if (isModuleLinkage(GV)) {
142 TS->EmitAMDGPUHsaModuleScopeGlobal(GV->getName());
144 TS->EmitAMDGPUHsaProgramScopeGlobal(GV->getName());
147 const DataLayout &DL = getDataLayout();
148 OutStreamer->PushSection();
149 OutStreamer->SwitchSection(
150 getObjFileLowering().SectionForGlobal(GV, *Mang, TM));
151 MCSymbol *GVSym = getSymbol(GV);
152 const Constant *C = GV->getInitializer();
153 OutStreamer->EmitLabel(GVSym);
154 EmitGlobalConstant(DL, C);
155 OutStreamer->PopSection();
158 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
160 // The starting address of all shader programs must be 256 bytes aligned.
163 SetupMachineFunction(MF);
165 MCContext &Context = getObjFileLowering().getContext();
166 MCSectionELF *ConfigSection =
167 Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);
168 OutStreamer->SwitchSection(ConfigSection);
170 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
171 SIProgramInfo KernelInfo;
172 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
173 getSIProgramInfo(KernelInfo, MF);
174 if (!STM.isAmdHsaOS()) {
175 EmitProgramInfoSI(MF, KernelInfo);
178 AMDGPUTargetStreamer *TS =
179 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
180 TS->EmitDirectiveHSACodeObjectVersion(1, 0);
181 AMDGPU::IsaVersion ISA = STM.getIsaVersion();
182 TS->EmitDirectiveHSACodeObjectISA(ISA.Major, ISA.Minor, ISA.Stepping,
185 EmitProgramInfoR600(MF);
190 DisasmLineMaxLen = 0;
195 MCSectionELF *CommentSection =
196 Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0);
197 OutStreamer->SwitchSection(CommentSection);
199 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
200 OutStreamer->emitRawComment(" Kernel info:", false);
201 OutStreamer->emitRawComment(" codeLenInByte = " + Twine(KernelInfo.CodeLen),
203 OutStreamer->emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR),
205 OutStreamer->emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
207 OutStreamer->emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode),
209 OutStreamer->emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
211 OutStreamer->emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize),
214 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:USER_SGPR: " +
215 Twine(G_00B84C_USER_SGPR(KernelInfo.ComputePGMRSrc2)),
217 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_X_EN: " +
218 Twine(G_00B84C_TGID_X_EN(KernelInfo.ComputePGMRSrc2)),
220 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Y_EN: " +
221 Twine(G_00B84C_TGID_Y_EN(KernelInfo.ComputePGMRSrc2)),
223 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Z_EN: " +
224 Twine(G_00B84C_TGID_Z_EN(KernelInfo.ComputePGMRSrc2)),
226 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " +
227 Twine(G_00B84C_TIDIG_COMP_CNT(KernelInfo.ComputePGMRSrc2)),
231 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
232 OutStreamer->emitRawComment(
233 Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->StackSize)));
237 if (STM.dumpCode()) {
239 OutStreamer->SwitchSection(
240 Context.getELFSection(".AMDGPU.disasm", ELF::SHT_NOTE, 0));
242 for (size_t i = 0; i < DisasmLines.size(); ++i) {
243 std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
244 Comment += " ; " + HexLines[i] + "\n";
246 OutStreamer->EmitBytes(StringRef(DisasmLines[i]));
247 OutStreamer->EmitBytes(StringRef(Comment));
254 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
256 bool killPixel = false;
257 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
258 const R600RegisterInfo *RI =
259 static_cast<const R600RegisterInfo *>(STM.getRegisterInfo());
260 const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
262 for (const MachineBasicBlock &MBB : MF) {
263 for (const MachineInstr &MI : MBB) {
264 if (MI.getOpcode() == AMDGPU::KILLGT)
266 unsigned numOperands = MI.getNumOperands();
267 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
268 const MachineOperand &MO = MI.getOperand(op_idx);
271 unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff;
273 // Register with value > 127 aren't GPR
276 MaxGPR = std::max(MaxGPR, HWReg);
282 if (STM.getGeneration() >= AMDGPUSubtarget::EVERGREEN) {
283 // Evergreen / Northern Islands
284 switch (MFI->getShaderType()) {
285 default: // Fall through
286 case ShaderType::COMPUTE: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
287 case ShaderType::GEOMETRY: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
288 case ShaderType::PIXEL: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
289 case ShaderType::VERTEX: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
293 switch (MFI->getShaderType()) {
294 default: // Fall through
295 case ShaderType::GEOMETRY: // Fall through
296 case ShaderType::COMPUTE: // Fall through
297 case ShaderType::VERTEX: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
298 case ShaderType::PIXEL: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
302 OutStreamer->EmitIntValue(RsrcReg, 4);
303 OutStreamer->EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
304 S_STACK_SIZE(MFI->StackSize), 4);
305 OutStreamer->EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
306 OutStreamer->EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
308 if (MFI->getShaderType() == ShaderType::COMPUTE) {
309 OutStreamer->EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4);
310 OutStreamer->EmitIntValue(RoundUpToAlignment(MFI->LDSSize, 4) >> 2, 4);
314 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
315 const MachineFunction &MF) const {
316 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
317 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
318 uint64_t CodeSize = 0;
319 unsigned MaxSGPR = 0;
320 unsigned MaxVGPR = 0;
321 bool VCCUsed = false;
322 bool FlatUsed = false;
323 const SIRegisterInfo *RI =
324 static_cast<const SIRegisterInfo *>(STM.getRegisterInfo());
326 for (const MachineBasicBlock &MBB : MF) {
327 for (const MachineInstr &MI : MBB) {
328 // TODO: CodeSize should account for multiple functions.
330 // TODO: Should we count size of debug info?
331 if (MI.isDebugValue())
334 // FIXME: This is reporting 0 for many instructions.
335 CodeSize += MI.getDesc().Size;
337 unsigned numOperands = MI.getNumOperands();
338 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
339 const MachineOperand &MO = MI.getOperand(op_idx);
346 unsigned reg = MO.getReg();
359 case AMDGPU::FLAT_SCR:
360 case AMDGPU::FLAT_SCR_LO:
361 case AMDGPU::FLAT_SCR_HI:
369 if (AMDGPU::SReg_32RegClass.contains(reg)) {
372 } else if (AMDGPU::VGPR_32RegClass.contains(reg)) {
375 } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
378 } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
381 } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
384 } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
387 } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
390 } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
393 } else if (AMDGPU::VReg_256RegClass.contains(reg)) {
396 } else if (AMDGPU::SReg_512RegClass.contains(reg)) {
399 } else if (AMDGPU::VReg_512RegClass.contains(reg)) {
403 llvm_unreachable("Unknown register class");
405 unsigned hwReg = RI->getEncodingValue(reg) & 0xff;
406 unsigned maxUsed = hwReg + width - 1;
408 MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
410 MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
422 // We found the maximum register index. They start at 0, so add one to get the
423 // number of registers.
424 ProgInfo.NumVGPR = MaxVGPR + 1;
425 ProgInfo.NumSGPR = MaxSGPR + 1;
427 if (STM.hasSGPRInitBug()) {
428 if (ProgInfo.NumSGPR > AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG) {
429 LLVMContext &Ctx = MF.getFunction()->getContext();
430 Ctx.emitError("too many SGPRs used with the SGPR init bug");
433 ProgInfo.NumSGPR = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG;
436 if (MFI->NumUserSGPRs > STM.getMaxNumUserSGPRs()) {
437 LLVMContext &Ctx = MF.getFunction()->getContext();
438 Ctx.emitError("too many user SGPRs used");
441 ProgInfo.VGPRBlocks = (ProgInfo.NumVGPR - 1) / 4;
442 ProgInfo.SGPRBlocks = (ProgInfo.NumSGPR - 1) / 8;
443 // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
445 ProgInfo.FloatMode = getFPMode(MF);
447 // XXX: Not quite sure what this does, but sc seems to unset this.
448 ProgInfo.IEEEMode = 0;
450 // Do not clamp NAN to 0.
451 ProgInfo.DX10Clamp = 0;
453 const MachineFrameInfo *FrameInfo = MF.getFrameInfo();
454 ProgInfo.ScratchSize = FrameInfo->estimateStackSize(MF);
456 ProgInfo.FlatUsed = FlatUsed;
457 ProgInfo.VCCUsed = VCCUsed;
458 ProgInfo.CodeLen = CodeSize;
460 unsigned LDSAlignShift;
461 if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
462 // LDS is allocated in 64 dword blocks.
465 // LDS is allocated in 128 dword blocks.
469 unsigned LDSSpillSize = MFI->LDSWaveSpillSize *
470 MFI->getMaximumWorkGroupSize(MF);
472 ProgInfo.LDSSize = MFI->LDSSize + LDSSpillSize;
474 RoundUpToAlignment(ProgInfo.LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
476 // Scratch is allocated in 256 dword blocks.
477 unsigned ScratchAlignShift = 10;
478 // We need to program the hardware with the amount of scratch memory that
479 // is used by the entire wave. ProgInfo.ScratchSize is the amount of
480 // scratch memory used per thread.
481 ProgInfo.ScratchBlocks =
482 RoundUpToAlignment(ProgInfo.ScratchSize * STM.getWavefrontSize(),
483 1 << ScratchAlignShift) >> ScratchAlignShift;
485 ProgInfo.ComputePGMRSrc1 =
486 S_00B848_VGPRS(ProgInfo.VGPRBlocks) |
487 S_00B848_SGPRS(ProgInfo.SGPRBlocks) |
488 S_00B848_PRIORITY(ProgInfo.Priority) |
489 S_00B848_FLOAT_MODE(ProgInfo.FloatMode) |
490 S_00B848_PRIV(ProgInfo.Priv) |
491 S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) |
492 S_00B848_DEBUG_MODE(ProgInfo.DebugMode) |
493 S_00B848_IEEE_MODE(ProgInfo.IEEEMode);
495 // 0 = X, 1 = XY, 2 = XYZ
496 unsigned TIDIGCompCnt = 0;
497 if (MFI->hasWorkItemIDZ())
499 else if (MFI->hasWorkItemIDY())
502 ProgInfo.ComputePGMRSrc2 =
503 S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
504 S_00B84C_USER_SGPR(MFI->getNumUserSGPRs()) |
505 S_00B84C_TGID_X_EN(MFI->hasWorkGroupIDX()) |
506 S_00B84C_TGID_Y_EN(MFI->hasWorkGroupIDY()) |
507 S_00B84C_TGID_Z_EN(MFI->hasWorkGroupIDZ()) |
508 S_00B84C_TG_SIZE_EN(MFI->hasWorkGroupInfo()) |
509 S_00B84C_TIDIG_COMP_CNT(TIDIGCompCnt) |
510 S_00B84C_EXCP_EN_MSB(0) |
511 S_00B84C_LDS_SIZE(ProgInfo.LDSBlocks) |
515 static unsigned getRsrcReg(unsigned ShaderType) {
516 switch (ShaderType) {
517 default: // Fall through
518 case ShaderType::COMPUTE: return R_00B848_COMPUTE_PGM_RSRC1;
519 case ShaderType::GEOMETRY: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
520 case ShaderType::PIXEL: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
521 case ShaderType::VERTEX: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
525 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
526 const SIProgramInfo &KernelInfo) {
527 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
528 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
529 unsigned RsrcReg = getRsrcReg(MFI->getShaderType());
531 if (MFI->getShaderType() == ShaderType::COMPUTE) {
532 OutStreamer->EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
534 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc1, 4);
536 OutStreamer->EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
537 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc2, 4);
539 OutStreamer->EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4);
540 OutStreamer->EmitIntValue(S_00B860_WAVESIZE(KernelInfo.ScratchBlocks), 4);
542 // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
543 // 0" comment but I don't see a corresponding field in the register spec.
545 OutStreamer->EmitIntValue(RsrcReg, 4);
546 OutStreamer->EmitIntValue(S_00B028_VGPRS(KernelInfo.VGPRBlocks) |
547 S_00B028_SGPRS(KernelInfo.SGPRBlocks), 4);
548 if (STM.isVGPRSpillingEnabled(MFI)) {
549 OutStreamer->EmitIntValue(R_0286E8_SPI_TMPRING_SIZE, 4);
550 OutStreamer->EmitIntValue(S_0286E8_WAVESIZE(KernelInfo.ScratchBlocks), 4);
554 if (MFI->getShaderType() == ShaderType::PIXEL) {
555 OutStreamer->EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
556 OutStreamer->EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(KernelInfo.LDSBlocks), 4);
557 OutStreamer->EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
558 OutStreamer->EmitIntValue(MFI->PSInputAddr, 4);
562 void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF,
563 const SIProgramInfo &KernelInfo) const {
564 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
565 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
566 amd_kernel_code_t header;
568 AMDGPU::initDefaultAMDKernelCodeT(header, STM.getFeatureBits());
570 header.compute_pgm_resource_registers =
571 KernelInfo.ComputePGMRSrc1 |
572 (KernelInfo.ComputePGMRSrc2 << 32);
573 header.code_properties = AMD_CODE_PROPERTY_IS_PTR64;
575 if (MFI->hasPrivateSegmentBuffer()) {
576 header.code_properties |=
577 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
580 if (MFI->hasDispatchPtr())
581 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
583 if (MFI->hasQueuePtr())
584 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
586 if (MFI->hasKernargSegmentPtr())
587 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
589 if (MFI->hasDispatchID())
590 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
592 if (MFI->hasFlatScratchInit())
593 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
595 // TODO: Private segment size
597 if (MFI->hasGridWorkgroupCountX()) {
598 header.code_properties |=
599 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X;
602 if (MFI->hasGridWorkgroupCountY()) {
603 header.code_properties |=
604 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y;
607 if (MFI->hasGridWorkgroupCountZ()) {
608 header.code_properties |=
609 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z;
612 if (MFI->hasDispatchPtr())
613 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
615 header.kernarg_segment_byte_size = MFI->ABIArgOffset;
616 header.wavefront_sgpr_count = KernelInfo.NumSGPR;
617 header.workitem_vgpr_count = KernelInfo.NumVGPR;
619 AMDGPUTargetStreamer *TS =
620 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
621 TS->EmitAMDKernelCodeT(header);
624 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
626 const char *ExtraCode, raw_ostream &O) {
627 if (ExtraCode && ExtraCode[0]) {
628 if (ExtraCode[1] != 0)
629 return true; // Unknown modifier.
631 switch (ExtraCode[0]) {
633 // See if this is a generic print operand
634 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
640 AMDGPUInstPrinter::printRegOperand(MI->getOperand(OpNo).getReg(), O,
641 *TM.getSubtargetImpl(*MF->getFunction())->getRegisterInfo());