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 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
115 // The starting address of all shader programs must be 256 bytes aligned.
118 SetupMachineFunction(MF);
120 MCContext &Context = getObjFileLowering().getContext();
121 MCSectionELF *ConfigSection =
122 Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);
123 OutStreamer->SwitchSection(ConfigSection);
125 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
126 SIProgramInfo KernelInfo;
127 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
128 getSIProgramInfo(KernelInfo, MF);
129 if (!STM.isAmdHsaOS()) {
130 EmitProgramInfoSI(MF, KernelInfo);
133 AMDGPUTargetStreamer *TS =
134 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
135 TS->EmitDirectiveHSACodeObjectVersion(1, 0);
136 AMDGPU::IsaVersion ISA = STM.getIsaVersion();
137 TS->EmitDirectiveHSACodeObjectISA(ISA.Major, ISA.Minor, ISA.Stepping,
140 EmitProgramInfoR600(MF);
145 DisasmLineMaxLen = 0;
150 MCSectionELF *CommentSection =
151 Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0);
152 OutStreamer->SwitchSection(CommentSection);
154 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
155 OutStreamer->emitRawComment(" Kernel info:", false);
156 OutStreamer->emitRawComment(" codeLenInByte = " + Twine(KernelInfo.CodeLen),
158 OutStreamer->emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR),
160 OutStreamer->emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
162 OutStreamer->emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode),
164 OutStreamer->emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
166 OutStreamer->emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize),
169 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
170 OutStreamer->emitRawComment(
171 Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->StackSize)));
175 if (STM.dumpCode()) {
177 OutStreamer->SwitchSection(
178 Context.getELFSection(".AMDGPU.disasm", ELF::SHT_NOTE, 0));
180 for (size_t i = 0; i < DisasmLines.size(); ++i) {
181 std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
182 Comment += " ; " + HexLines[i] + "\n";
184 OutStreamer->EmitBytes(StringRef(DisasmLines[i]));
185 OutStreamer->EmitBytes(StringRef(Comment));
192 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
194 bool killPixel = false;
195 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
196 const R600RegisterInfo *RI =
197 static_cast<const R600RegisterInfo *>(STM.getRegisterInfo());
198 const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
200 for (const MachineBasicBlock &MBB : MF) {
201 for (const MachineInstr &MI : MBB) {
202 if (MI.getOpcode() == AMDGPU::KILLGT)
204 unsigned numOperands = MI.getNumOperands();
205 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
206 const MachineOperand &MO = MI.getOperand(op_idx);
209 unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff;
211 // Register with value > 127 aren't GPR
214 MaxGPR = std::max(MaxGPR, HWReg);
220 if (STM.getGeneration() >= AMDGPUSubtarget::EVERGREEN) {
221 // Evergreen / Northern Islands
222 switch (MFI->getShaderType()) {
223 default: // Fall through
224 case ShaderType::COMPUTE: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
225 case ShaderType::GEOMETRY: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
226 case ShaderType::PIXEL: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
227 case ShaderType::VERTEX: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
231 switch (MFI->getShaderType()) {
232 default: // Fall through
233 case ShaderType::GEOMETRY: // Fall through
234 case ShaderType::COMPUTE: // Fall through
235 case ShaderType::VERTEX: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
236 case ShaderType::PIXEL: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
240 OutStreamer->EmitIntValue(RsrcReg, 4);
241 OutStreamer->EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
242 S_STACK_SIZE(MFI->StackSize), 4);
243 OutStreamer->EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
244 OutStreamer->EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
246 if (MFI->getShaderType() == ShaderType::COMPUTE) {
247 OutStreamer->EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4);
248 OutStreamer->EmitIntValue(RoundUpToAlignment(MFI->LDSSize, 4) >> 2, 4);
252 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
253 const MachineFunction &MF) const {
254 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
255 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
256 uint64_t CodeSize = 0;
257 unsigned MaxSGPR = 0;
258 unsigned MaxVGPR = 0;
259 bool VCCUsed = false;
260 bool FlatUsed = false;
261 const SIRegisterInfo *RI =
262 static_cast<const SIRegisterInfo *>(STM.getRegisterInfo());
264 for (const MachineBasicBlock &MBB : MF) {
265 for (const MachineInstr &MI : MBB) {
266 // TODO: CodeSize should account for multiple functions.
268 // TODO: Should we count size of debug info?
269 if (MI.isDebugValue())
272 // FIXME: This is reporting 0 for many instructions.
273 CodeSize += MI.getDesc().Size;
275 unsigned numOperands = MI.getNumOperands();
276 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
277 const MachineOperand &MO = MI.getOperand(op_idx);
284 unsigned reg = MO.getReg();
297 case AMDGPU::FLAT_SCR:
298 case AMDGPU::FLAT_SCR_LO:
299 case AMDGPU::FLAT_SCR_HI:
307 if (AMDGPU::SReg_32RegClass.contains(reg)) {
310 } else if (AMDGPU::VGPR_32RegClass.contains(reg)) {
313 } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
316 } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
319 } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
322 } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
325 } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
328 } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
331 } else if (AMDGPU::VReg_256RegClass.contains(reg)) {
334 } else if (AMDGPU::SReg_512RegClass.contains(reg)) {
337 } else if (AMDGPU::VReg_512RegClass.contains(reg)) {
341 llvm_unreachable("Unknown register class");
343 unsigned hwReg = RI->getEncodingValue(reg) & 0xff;
344 unsigned maxUsed = hwReg + width - 1;
346 MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
348 MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
360 // We found the maximum register index. They start at 0, so add one to get the
361 // number of registers.
362 ProgInfo.NumVGPR = MaxVGPR + 1;
363 ProgInfo.NumSGPR = MaxSGPR + 1;
365 if (STM.hasSGPRInitBug()) {
366 if (ProgInfo.NumSGPR > AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG) {
367 LLVMContext &Ctx = MF.getFunction()->getContext();
368 Ctx.emitError("too many SGPRs used with the SGPR init bug");
371 ProgInfo.NumSGPR = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG;
374 ProgInfo.VGPRBlocks = (ProgInfo.NumVGPR - 1) / 4;
375 ProgInfo.SGPRBlocks = (ProgInfo.NumSGPR - 1) / 8;
376 // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
378 ProgInfo.FloatMode = getFPMode(MF);
380 // XXX: Not quite sure what this does, but sc seems to unset this.
381 ProgInfo.IEEEMode = 0;
383 // Do not clamp NAN to 0.
384 ProgInfo.DX10Clamp = 0;
386 const MachineFrameInfo *FrameInfo = MF.getFrameInfo();
387 ProgInfo.ScratchSize = FrameInfo->estimateStackSize(MF);
389 ProgInfo.FlatUsed = FlatUsed;
390 ProgInfo.VCCUsed = VCCUsed;
391 ProgInfo.CodeLen = CodeSize;
393 unsigned LDSAlignShift;
394 if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
395 // LDS is allocated in 64 dword blocks.
398 // LDS is allocated in 128 dword blocks.
402 unsigned LDSSpillSize = MFI->LDSWaveSpillSize *
403 MFI->getMaximumWorkGroupSize(MF);
405 ProgInfo.LDSSize = MFI->LDSSize + LDSSpillSize;
407 RoundUpToAlignment(ProgInfo.LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
409 // Scratch is allocated in 256 dword blocks.
410 unsigned ScratchAlignShift = 10;
411 // We need to program the hardware with the amount of scratch memory that
412 // is used by the entire wave. ProgInfo.ScratchSize is the amount of
413 // scratch memory used per thread.
414 ProgInfo.ScratchBlocks =
415 RoundUpToAlignment(ProgInfo.ScratchSize * STM.getWavefrontSize(),
416 1 << ScratchAlignShift) >> ScratchAlignShift;
418 ProgInfo.ComputePGMRSrc1 =
419 S_00B848_VGPRS(ProgInfo.VGPRBlocks) |
420 S_00B848_SGPRS(ProgInfo.SGPRBlocks) |
421 S_00B848_PRIORITY(ProgInfo.Priority) |
422 S_00B848_FLOAT_MODE(ProgInfo.FloatMode) |
423 S_00B848_PRIV(ProgInfo.Priv) |
424 S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) |
425 S_00B848_IEEE_MODE(ProgInfo.DebugMode) |
426 S_00B848_IEEE_MODE(ProgInfo.IEEEMode);
428 ProgInfo.ComputePGMRSrc2 =
429 S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
430 S_00B84C_USER_SGPR(MFI->NumUserSGPRs) |
431 S_00B84C_TGID_X_EN(1) |
432 S_00B84C_TGID_Y_EN(1) |
433 S_00B84C_TGID_Z_EN(1) |
434 S_00B84C_TG_SIZE_EN(1) |
435 S_00B84C_TIDIG_COMP_CNT(2) |
436 S_00B84C_LDS_SIZE(ProgInfo.LDSBlocks);
439 static unsigned getRsrcReg(unsigned ShaderType) {
440 switch (ShaderType) {
441 default: // Fall through
442 case ShaderType::COMPUTE: return R_00B848_COMPUTE_PGM_RSRC1;
443 case ShaderType::GEOMETRY: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
444 case ShaderType::PIXEL: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
445 case ShaderType::VERTEX: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
449 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
450 const SIProgramInfo &KernelInfo) {
451 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
452 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
453 unsigned RsrcReg = getRsrcReg(MFI->getShaderType());
455 if (MFI->getShaderType() == ShaderType::COMPUTE) {
456 OutStreamer->EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
458 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc1, 4);
460 OutStreamer->EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
461 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc2, 4);
463 OutStreamer->EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4);
464 OutStreamer->EmitIntValue(S_00B860_WAVESIZE(KernelInfo.ScratchBlocks), 4);
466 // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
467 // 0" comment but I don't see a corresponding field in the register spec.
469 OutStreamer->EmitIntValue(RsrcReg, 4);
470 OutStreamer->EmitIntValue(S_00B028_VGPRS(KernelInfo.VGPRBlocks) |
471 S_00B028_SGPRS(KernelInfo.SGPRBlocks), 4);
472 if (STM.isVGPRSpillingEnabled(MFI)) {
473 OutStreamer->EmitIntValue(R_0286E8_SPI_TMPRING_SIZE, 4);
474 OutStreamer->EmitIntValue(S_0286E8_WAVESIZE(KernelInfo.ScratchBlocks), 4);
478 if (MFI->getShaderType() == ShaderType::PIXEL) {
479 OutStreamer->EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
480 OutStreamer->EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(KernelInfo.LDSBlocks), 4);
481 OutStreamer->EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
482 OutStreamer->EmitIntValue(MFI->PSInputAddr, 4);
486 void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF,
487 const SIProgramInfo &KernelInfo) const {
488 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
489 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
490 amd_kernel_code_t header;
492 AMDGPU::initDefaultAMDKernelCodeT(header, STM.getFeatureBits());
494 header.compute_pgm_resource_registers =
495 KernelInfo.ComputePGMRSrc1 |
496 (KernelInfo.ComputePGMRSrc2 << 32);
497 header.code_properties =
498 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR |
499 AMD_CODE_PROPERTY_IS_PTR64;
501 header.kernarg_segment_byte_size = MFI->ABIArgOffset;
502 header.wavefront_sgpr_count = KernelInfo.NumSGPR;
503 header.workitem_vgpr_count = KernelInfo.NumVGPR;
505 AMDGPUTargetStreamer *TS =
506 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
507 TS->EmitAMDKernelCodeT(header);
510 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
512 const char *ExtraCode, raw_ostream &O) {
513 if (ExtraCode && ExtraCode[0]) {
514 if (ExtraCode[1] != 0)
515 return true; // Unknown modifier.
517 switch (ExtraCode[0]) {
519 // See if this is a generic print operand
520 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
526 AMDGPUInstPrinter::printRegOperand(MI->getOperand(OpNo).getReg(), O,
527 *TM.getSubtargetImpl(*MF->getFunction())->getRegisterInfo());