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"
21 #include "AMDGPUSubtarget.h"
22 #include "R600Defines.h"
23 #include "R600MachineFunctionInfo.h"
24 #include "R600RegisterInfo.h"
25 #include "SIDefines.h"
26 #include "SIMachineFunctionInfo.h"
27 #include "SIRegisterInfo.h"
28 #include "llvm/MC/MCContext.h"
29 #include "llvm/MC/MCSectionELF.h"
30 #include "llvm/MC/MCStreamer.h"
31 #include "llvm/Support/ELF.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/TargetRegistry.h"
34 #include "llvm/Target/TargetLoweringObjectFile.h"
38 // TODO: This should get the default rounding mode from the kernel. We just set
39 // the default here, but this could change if the OpenCL rounding mode pragmas
42 // The denormal mode here should match what is reported by the OpenCL runtime
43 // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but
44 // can also be override to flush with the -cl-denorms-are-zero compiler flag.
46 // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double
47 // precision, and leaves single precision to flush all and does not report
48 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
49 // CL_FP_DENORM for both.
51 // FIXME: It seems some instructions do not support single precision denormals
52 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
53 // and sin_f32, cos_f32 on most parts).
55 // We want to use these instructions, and using fp32 denormals also causes
56 // instructions to run at the double precision rate for the device so it's
57 // probably best to just report no single precision denormals.
58 static uint32_t getFPMode(const MachineFunction &) {
59 return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
60 FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
61 FP_DENORM_MODE_SP(FP_DENORM_FLUSH_IN_FLUSH_OUT) |
62 FP_DENORM_MODE_DP(FP_DENORM_FLUSH_NONE);
65 static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,
66 MCStreamer &Streamer) {
67 return new AMDGPUAsmPrinter(tm, Streamer);
70 extern "C" void LLVMInitializeR600AsmPrinter() {
71 TargetRegistry::RegisterAsmPrinter(TheAMDGPUTarget, createAMDGPUAsmPrinterPass);
74 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
75 : AsmPrinter(TM, Streamer) {
76 DisasmEnabled = TM.getSubtarget<AMDGPUSubtarget>().dumpCode();
79 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
80 SetupMachineFunction(MF);
82 OutStreamer.emitRawComment(Twine('@') + MF.getName() + Twine(':'));
84 MCContext &Context = getObjFileLowering().getContext();
85 const MCSectionELF *ConfigSection = Context.getELFSection(".AMDGPU.config",
87 SectionKind::getReadOnly());
88 OutStreamer.SwitchSection(ConfigSection);
90 const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
91 SIProgramInfo KernelInfo;
92 if (STM.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
93 getSIProgramInfo(KernelInfo, MF);
94 EmitProgramInfoSI(MF, KernelInfo);
96 EmitProgramInfoR600(MF);
101 DisasmLineMaxLen = 0;
103 OutStreamer.SwitchSection(getObjFileLowering().getTextSection());
107 const MCSectionELF *CommentSection
108 = Context.getELFSection(".AMDGPU.csdata",
109 ELF::SHT_PROGBITS, 0,
110 SectionKind::getReadOnly());
111 OutStreamer.SwitchSection(CommentSection);
113 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
114 OutStreamer.emitRawComment(" Kernel info:", false);
115 OutStreamer.emitRawComment(" codeLenInByte = " + Twine(KernelInfo.CodeLen),
117 OutStreamer.emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR),
119 OutStreamer.emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
121 OutStreamer.emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode),
123 OutStreamer.emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
126 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
127 OutStreamer.emitRawComment(
128 Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->StackSize)));
132 if (STM.dumpCode()) {
133 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
138 OutStreamer.SwitchSection(Context.getELFSection(".AMDGPU.disasm",
140 SectionKind::getReadOnly()));
142 for (size_t i = 0; i < DisasmLines.size(); ++i) {
143 std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
144 Comment += " ; " + HexLines[i] + "\n";
146 OutStreamer.EmitBytes(StringRef(DisasmLines[i]));
147 OutStreamer.EmitBytes(StringRef(Comment));
155 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
157 bool killPixel = false;
158 const R600RegisterInfo *RI
159 = static_cast<const R600RegisterInfo*>(TM.getRegisterInfo());
160 const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
161 const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
163 for (const MachineBasicBlock &MBB : MF) {
164 for (const MachineInstr &MI : MBB) {
165 if (MI.getOpcode() == AMDGPU::KILLGT)
167 unsigned numOperands = MI.getNumOperands();
168 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
169 const MachineOperand &MO = MI.getOperand(op_idx);
172 unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff;
174 // Register with value > 127 aren't GPR
177 MaxGPR = std::max(MaxGPR, HWReg);
183 if (STM.getGeneration() >= AMDGPUSubtarget::EVERGREEN) {
184 // Evergreen / Northern Islands
185 switch (MFI->getShaderType()) {
186 default: // Fall through
187 case ShaderType::COMPUTE: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
188 case ShaderType::GEOMETRY: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
189 case ShaderType::PIXEL: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
190 case ShaderType::VERTEX: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
194 switch (MFI->getShaderType()) {
195 default: // Fall through
196 case ShaderType::GEOMETRY: // Fall through
197 case ShaderType::COMPUTE: // Fall through
198 case ShaderType::VERTEX: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
199 case ShaderType::PIXEL: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
203 OutStreamer.EmitIntValue(RsrcReg, 4);
204 OutStreamer.EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
205 S_STACK_SIZE(MFI->StackSize), 4);
206 OutStreamer.EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
207 OutStreamer.EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
209 if (MFI->getShaderType() == ShaderType::COMPUTE) {
210 OutStreamer.EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4);
211 OutStreamer.EmitIntValue(RoundUpToAlignment(MFI->LDSSize, 4) >> 2, 4);
215 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
216 const MachineFunction &MF) const {
217 uint64_t CodeSize = 0;
218 unsigned MaxSGPR = 0;
219 unsigned MaxVGPR = 0;
220 bool VCCUsed = false;
221 const SIRegisterInfo *RI
222 = static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
224 for (const MachineBasicBlock &MBB : MF) {
225 for (const MachineInstr &MI : MBB) {
226 // TODO: CodeSize should account for multiple functions.
227 CodeSize += MI.getDesc().Size;
229 unsigned numOperands = MI.getNumOperands();
230 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
231 const MachineOperand &MO = MI.getOperand(op_idx);
238 unsigned reg = MO.getReg();
239 if (reg == AMDGPU::VCC || reg == AMDGPU::VCC_LO ||
240 reg == AMDGPU::VCC_HI) {
253 if (AMDGPU::SReg_32RegClass.contains(reg)) {
256 } else if (AMDGPU::VReg_32RegClass.contains(reg)) {
259 } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
262 } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
265 } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
268 } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
271 } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
274 } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
277 } else if (AMDGPU::VReg_256RegClass.contains(reg)) {
280 } else if (AMDGPU::SReg_512RegClass.contains(reg)) {
283 } else if (AMDGPU::VReg_512RegClass.contains(reg)) {
287 llvm_unreachable("Unknown register class");
289 unsigned hwReg = RI->getEncodingValue(reg) & 0xff;
290 unsigned maxUsed = hwReg + width - 1;
292 MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
294 MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
303 ProgInfo.NumVGPR = MaxVGPR;
304 ProgInfo.NumSGPR = MaxSGPR;
306 // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
308 ProgInfo.FloatMode = getFPMode(MF);
310 // XXX: Not quite sure what this does, but sc seems to unset this.
311 ProgInfo.IEEEMode = 0;
313 // Do not clamp NAN to 0.
314 ProgInfo.DX10Clamp = 0;
316 ProgInfo.CodeLen = CodeSize;
319 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
320 const SIProgramInfo &KernelInfo) {
321 const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
322 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
325 switch (MFI->getShaderType()) {
326 default: // Fall through
327 case ShaderType::COMPUTE: RsrcReg = R_00B848_COMPUTE_PGM_RSRC1; break;
328 case ShaderType::GEOMETRY: RsrcReg = R_00B228_SPI_SHADER_PGM_RSRC1_GS; break;
329 case ShaderType::PIXEL: RsrcReg = R_00B028_SPI_SHADER_PGM_RSRC1_PS; break;
330 case ShaderType::VERTEX: RsrcReg = R_00B128_SPI_SHADER_PGM_RSRC1_VS; break;
333 unsigned LDSAlignShift;
334 if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
335 // LDS is allocated in 64 dword blocks.
338 // LDS is allocated in 128 dword blocks.
343 RoundUpToAlignment(MFI->LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
345 if (MFI->getShaderType() == ShaderType::COMPUTE) {
346 OutStreamer.EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
348 const uint32_t ComputePGMRSrc1 =
349 S_00B848_VGPRS(KernelInfo.NumVGPR / 4) |
350 S_00B848_SGPRS(KernelInfo.NumSGPR / 8) |
351 S_00B848_PRIORITY(KernelInfo.Priority) |
352 S_00B848_FLOAT_MODE(KernelInfo.FloatMode) |
353 S_00B848_PRIV(KernelInfo.Priv) |
354 S_00B848_DX10_CLAMP(KernelInfo.DX10Clamp) |
355 S_00B848_IEEE_MODE(KernelInfo.DebugMode) |
356 S_00B848_IEEE_MODE(KernelInfo.IEEEMode);
358 OutStreamer.EmitIntValue(ComputePGMRSrc1, 4);
360 OutStreamer.EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
361 OutStreamer.EmitIntValue(S_00B84C_LDS_SIZE(LDSBlocks), 4);
363 OutStreamer.EmitIntValue(RsrcReg, 4);
364 OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.NumVGPR / 4) |
365 S_00B028_SGPRS(KernelInfo.NumSGPR / 8), 4);
368 if (MFI->getShaderType() == ShaderType::PIXEL) {
369 OutStreamer.EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
370 OutStreamer.EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(LDSBlocks), 4);
371 OutStreamer.EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
372 OutStreamer.EmitIntValue(MFI->PSInputAddr, 4);