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::EmitEndOfAsmFile(Module &M) {
105 // This label is used to mark the end of the .text section.
106 const TargetLoweringObjectFile &TLOF = getObjFileLowering();
107 OutStreamer->SwitchSection(TLOF.getTextSection());
108 MCSymbol *EndOfTextLabel =
109 OutContext.getOrCreateSymbol(StringRef(END_OF_TEXT_LABEL_NAME));
110 OutStreamer->EmitLabel(EndOfTextLabel);
113 void AMDGPUAsmPrinter::EmitFunctionEntryLabel() {
114 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
115 const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
116 if (MFI->isKernel() && STM.isAmdHsaOS()) {
117 AMDGPUTargetStreamer *TS =
118 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
119 TS->EmitAMDGPUSymbolType(CurrentFnSym->getName(),
120 ELF::STT_AMDGPU_HSA_KERNEL);
123 AsmPrinter::EmitFunctionEntryLabel();
126 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
128 // The starting address of all shader programs must be 256 bytes aligned.
131 SetupMachineFunction(MF);
133 MCContext &Context = getObjFileLowering().getContext();
134 MCSectionELF *ConfigSection =
135 Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);
136 OutStreamer->SwitchSection(ConfigSection);
138 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
139 SIProgramInfo KernelInfo;
140 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
141 getSIProgramInfo(KernelInfo, MF);
142 if (!STM.isAmdHsaOS()) {
143 EmitProgramInfoSI(MF, KernelInfo);
146 AMDGPUTargetStreamer *TS =
147 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
148 TS->EmitDirectiveHSACodeObjectVersion(1, 0);
149 AMDGPU::IsaVersion ISA = STM.getIsaVersion();
150 TS->EmitDirectiveHSACodeObjectISA(ISA.Major, ISA.Minor, ISA.Stepping,
153 EmitProgramInfoR600(MF);
158 DisasmLineMaxLen = 0;
163 MCSectionELF *CommentSection =
164 Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0);
165 OutStreamer->SwitchSection(CommentSection);
167 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
168 OutStreamer->emitRawComment(" Kernel info:", false);
169 OutStreamer->emitRawComment(" codeLenInByte = " + Twine(KernelInfo.CodeLen),
171 OutStreamer->emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR),
173 OutStreamer->emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
175 OutStreamer->emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode),
177 OutStreamer->emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
179 OutStreamer->emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize),
182 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:USER_SGPR: " +
183 Twine(G_00B84C_USER_SGPR(KernelInfo.ComputePGMRSrc2)),
185 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_X_EN: " +
186 Twine(G_00B84C_TGID_X_EN(KernelInfo.ComputePGMRSrc2)),
188 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Y_EN: " +
189 Twine(G_00B84C_TGID_Y_EN(KernelInfo.ComputePGMRSrc2)),
191 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Z_EN: " +
192 Twine(G_00B84C_TGID_Z_EN(KernelInfo.ComputePGMRSrc2)),
194 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " +
195 Twine(G_00B84C_TIDIG_COMP_CNT(KernelInfo.ComputePGMRSrc2)),
199 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
200 OutStreamer->emitRawComment(
201 Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->StackSize)));
205 if (STM.dumpCode()) {
207 OutStreamer->SwitchSection(
208 Context.getELFSection(".AMDGPU.disasm", ELF::SHT_NOTE, 0));
210 for (size_t i = 0; i < DisasmLines.size(); ++i) {
211 std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
212 Comment += " ; " + HexLines[i] + "\n";
214 OutStreamer->EmitBytes(StringRef(DisasmLines[i]));
215 OutStreamer->EmitBytes(StringRef(Comment));
222 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
224 bool killPixel = false;
225 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
226 const R600RegisterInfo *RI =
227 static_cast<const R600RegisterInfo *>(STM.getRegisterInfo());
228 const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
230 for (const MachineBasicBlock &MBB : MF) {
231 for (const MachineInstr &MI : MBB) {
232 if (MI.getOpcode() == AMDGPU::KILLGT)
234 unsigned numOperands = MI.getNumOperands();
235 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
236 const MachineOperand &MO = MI.getOperand(op_idx);
239 unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff;
241 // Register with value > 127 aren't GPR
244 MaxGPR = std::max(MaxGPR, HWReg);
250 if (STM.getGeneration() >= AMDGPUSubtarget::EVERGREEN) {
251 // Evergreen / Northern Islands
252 switch (MFI->getShaderType()) {
253 default: // Fall through
254 case ShaderType::COMPUTE: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
255 case ShaderType::GEOMETRY: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
256 case ShaderType::PIXEL: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
257 case ShaderType::VERTEX: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
261 switch (MFI->getShaderType()) {
262 default: // Fall through
263 case ShaderType::GEOMETRY: // Fall through
264 case ShaderType::COMPUTE: // Fall through
265 case ShaderType::VERTEX: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
266 case ShaderType::PIXEL: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
270 OutStreamer->EmitIntValue(RsrcReg, 4);
271 OutStreamer->EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
272 S_STACK_SIZE(MFI->StackSize), 4);
273 OutStreamer->EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
274 OutStreamer->EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
276 if (MFI->getShaderType() == ShaderType::COMPUTE) {
277 OutStreamer->EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4);
278 OutStreamer->EmitIntValue(RoundUpToAlignment(MFI->LDSSize, 4) >> 2, 4);
282 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
283 const MachineFunction &MF) const {
284 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
285 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
286 uint64_t CodeSize = 0;
287 unsigned MaxSGPR = 0;
288 unsigned MaxVGPR = 0;
289 bool VCCUsed = false;
290 bool FlatUsed = false;
291 const SIRegisterInfo *RI =
292 static_cast<const SIRegisterInfo *>(STM.getRegisterInfo());
294 for (const MachineBasicBlock &MBB : MF) {
295 for (const MachineInstr &MI : MBB) {
296 // TODO: CodeSize should account for multiple functions.
298 // TODO: Should we count size of debug info?
299 if (MI.isDebugValue())
302 // FIXME: This is reporting 0 for many instructions.
303 CodeSize += MI.getDesc().Size;
305 unsigned numOperands = MI.getNumOperands();
306 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
307 const MachineOperand &MO = MI.getOperand(op_idx);
314 unsigned reg = MO.getReg();
327 case AMDGPU::FLAT_SCR:
328 case AMDGPU::FLAT_SCR_LO:
329 case AMDGPU::FLAT_SCR_HI:
337 if (AMDGPU::SReg_32RegClass.contains(reg)) {
340 } else if (AMDGPU::VGPR_32RegClass.contains(reg)) {
343 } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
346 } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
349 } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
352 } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
355 } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
358 } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
361 } else if (AMDGPU::VReg_256RegClass.contains(reg)) {
364 } else if (AMDGPU::SReg_512RegClass.contains(reg)) {
367 } else if (AMDGPU::VReg_512RegClass.contains(reg)) {
371 llvm_unreachable("Unknown register class");
373 unsigned hwReg = RI->getEncodingValue(reg) & 0xff;
374 unsigned maxUsed = hwReg + width - 1;
376 MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
378 MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
390 // We found the maximum register index. They start at 0, so add one to get the
391 // number of registers.
392 ProgInfo.NumVGPR = MaxVGPR + 1;
393 ProgInfo.NumSGPR = MaxSGPR + 1;
395 if (STM.hasSGPRInitBug()) {
396 if (ProgInfo.NumSGPR > AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG) {
397 LLVMContext &Ctx = MF.getFunction()->getContext();
398 Ctx.emitError("too many SGPRs used with the SGPR init bug");
401 ProgInfo.NumSGPR = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG;
404 ProgInfo.VGPRBlocks = (ProgInfo.NumVGPR - 1) / 4;
405 ProgInfo.SGPRBlocks = (ProgInfo.NumSGPR - 1) / 8;
406 // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
408 ProgInfo.FloatMode = getFPMode(MF);
410 // XXX: Not quite sure what this does, but sc seems to unset this.
411 ProgInfo.IEEEMode = 0;
413 // Do not clamp NAN to 0.
414 ProgInfo.DX10Clamp = 0;
416 const MachineFrameInfo *FrameInfo = MF.getFrameInfo();
417 ProgInfo.ScratchSize = FrameInfo->estimateStackSize(MF);
419 ProgInfo.FlatUsed = FlatUsed;
420 ProgInfo.VCCUsed = VCCUsed;
421 ProgInfo.CodeLen = CodeSize;
423 unsigned LDSAlignShift;
424 if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
425 // LDS is allocated in 64 dword blocks.
428 // LDS is allocated in 128 dword blocks.
432 unsigned LDSSpillSize = MFI->LDSWaveSpillSize *
433 MFI->getMaximumWorkGroupSize(MF);
435 ProgInfo.LDSSize = MFI->LDSSize + LDSSpillSize;
437 RoundUpToAlignment(ProgInfo.LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
439 // Scratch is allocated in 256 dword blocks.
440 unsigned ScratchAlignShift = 10;
441 // We need to program the hardware with the amount of scratch memory that
442 // is used by the entire wave. ProgInfo.ScratchSize is the amount of
443 // scratch memory used per thread.
444 ProgInfo.ScratchBlocks =
445 RoundUpToAlignment(ProgInfo.ScratchSize * STM.getWavefrontSize(),
446 1 << ScratchAlignShift) >> ScratchAlignShift;
448 ProgInfo.ComputePGMRSrc1 =
449 S_00B848_VGPRS(ProgInfo.VGPRBlocks) |
450 S_00B848_SGPRS(ProgInfo.SGPRBlocks) |
451 S_00B848_PRIORITY(ProgInfo.Priority) |
452 S_00B848_FLOAT_MODE(ProgInfo.FloatMode) |
453 S_00B848_PRIV(ProgInfo.Priv) |
454 S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) |
455 S_00B848_IEEE_MODE(ProgInfo.DebugMode) |
456 S_00B848_IEEE_MODE(ProgInfo.IEEEMode);
458 ProgInfo.ComputePGMRSrc2 =
459 S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
460 S_00B84C_USER_SGPR(MFI->NumUserSGPRs) |
461 S_00B84C_TGID_X_EN(1) |
462 S_00B84C_TGID_Y_EN(1) |
463 S_00B84C_TGID_Z_EN(1) |
464 S_00B84C_TG_SIZE_EN(1) |
465 S_00B84C_TIDIG_COMP_CNT(2) |
466 S_00B84C_LDS_SIZE(ProgInfo.LDSBlocks);
469 static unsigned getRsrcReg(unsigned ShaderType) {
470 switch (ShaderType) {
471 default: // Fall through
472 case ShaderType::COMPUTE: return R_00B848_COMPUTE_PGM_RSRC1;
473 case ShaderType::GEOMETRY: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
474 case ShaderType::PIXEL: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
475 case ShaderType::VERTEX: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
479 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
480 const SIProgramInfo &KernelInfo) {
481 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
482 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
483 unsigned RsrcReg = getRsrcReg(MFI->getShaderType());
485 if (MFI->getShaderType() == ShaderType::COMPUTE) {
486 OutStreamer->EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
488 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc1, 4);
490 OutStreamer->EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
491 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc2, 4);
493 OutStreamer->EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4);
494 OutStreamer->EmitIntValue(S_00B860_WAVESIZE(KernelInfo.ScratchBlocks), 4);
496 // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
497 // 0" comment but I don't see a corresponding field in the register spec.
499 OutStreamer->EmitIntValue(RsrcReg, 4);
500 OutStreamer->EmitIntValue(S_00B028_VGPRS(KernelInfo.VGPRBlocks) |
501 S_00B028_SGPRS(KernelInfo.SGPRBlocks), 4);
502 if (STM.isVGPRSpillingEnabled(MFI)) {
503 OutStreamer->EmitIntValue(R_0286E8_SPI_TMPRING_SIZE, 4);
504 OutStreamer->EmitIntValue(S_0286E8_WAVESIZE(KernelInfo.ScratchBlocks), 4);
508 if (MFI->getShaderType() == ShaderType::PIXEL) {
509 OutStreamer->EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
510 OutStreamer->EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(KernelInfo.LDSBlocks), 4);
511 OutStreamer->EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
512 OutStreamer->EmitIntValue(MFI->PSInputAddr, 4);
516 void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF,
517 const SIProgramInfo &KernelInfo) const {
518 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
519 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
520 amd_kernel_code_t header;
522 AMDGPU::initDefaultAMDKernelCodeT(header, STM.getFeatureBits());
524 header.compute_pgm_resource_registers =
525 KernelInfo.ComputePGMRSrc1 |
526 (KernelInfo.ComputePGMRSrc2 << 32);
527 header.code_properties =
528 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR |
529 AMD_CODE_PROPERTY_IS_PTR64;
531 if (MFI->hasDispatchPtr())
532 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
534 header.kernarg_segment_byte_size = MFI->ABIArgOffset;
535 header.wavefront_sgpr_count = KernelInfo.NumSGPR;
536 header.workitem_vgpr_count = KernelInfo.NumVGPR;
538 AMDGPUTargetStreamer *TS =
539 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
540 TS->EmitAMDKernelCodeT(header);
543 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
545 const char *ExtraCode, raw_ostream &O) {
546 if (ExtraCode && ExtraCode[0]) {
547 if (ExtraCode[1] != 0)
548 return true; // Unknown modifier.
550 switch (ExtraCode[0]) {
552 // See if this is a generic print operand
553 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
559 AMDGPUInstPrinter::printRegOperand(MI->getOperand(OpNo).getReg(), O,
560 *TM.getSubtargetImpl(*MF->getFunction())->getRegisterInfo());