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 //===----------------------------------------------------------------------===//
20 #include "AMDGPUAsmPrinter.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"
39 static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,
40 MCStreamer &Streamer) {
41 return new AMDGPUAsmPrinter(tm, Streamer);
44 extern "C" void LLVMInitializeR600AsmPrinter() {
45 TargetRegistry::RegisterAsmPrinter(TheAMDGPUTarget, createAMDGPUAsmPrinterPass);
48 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
49 : AsmPrinter(TM, Streamer) {
50 DisasmEnabled = TM.getSubtarget<AMDGPUSubtarget>().dumpCode() &&
51 ! Streamer.hasRawTextSupport();
54 /// We need to override this function so we can avoid
55 /// the call to EmitFunctionHeader(), which the MCPureStreamer can't handle.
56 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
57 SetupMachineFunction(MF);
59 OutStreamer.emitRawComment(Twine('@') + MF.getName() + Twine(':'));
61 MCContext &Context = getObjFileLowering().getContext();
62 const MCSectionELF *ConfigSection = Context.getELFSection(".AMDGPU.config",
64 SectionKind::getReadOnly());
65 OutStreamer.SwitchSection(ConfigSection);
67 const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
68 SIProgramInfo KernelInfo;
69 if (STM.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
70 findNumUsedRegistersSI(MF, KernelInfo.NumSGPR, KernelInfo.NumVGPR);
71 EmitProgramInfoSI(MF, KernelInfo);
73 EmitProgramInfoR600(MF);
80 OutStreamer.SwitchSection(getObjFileLowering().getTextSection());
84 const MCSectionELF *CommentSection
85 = Context.getELFSection(".AMDGPU.csdata",
87 SectionKind::getReadOnly());
88 OutStreamer.SwitchSection(CommentSection);
90 if (STM.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
91 OutStreamer.emitRawComment(" Kernel info:", false);
92 OutStreamer.emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR),
94 OutStreamer.emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
97 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
98 OutStreamer.emitRawComment(
99 Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->StackSize)));
103 if (STM.dumpCode()) {
104 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
109 OutStreamer.SwitchSection(Context.getELFSection(".AMDGPU.disasm",
111 SectionKind::getReadOnly()));
113 for (size_t i = 0; i < DisasmLines.size(); ++i) {
114 std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
115 Comment += " ; " + HexLines[i] + "\n";
117 OutStreamer.EmitBytes(StringRef(DisasmLines[i]));
118 OutStreamer.EmitBytes(StringRef(Comment));
126 void AMDGPUAsmPrinter::EmitProgramInfoR600(MachineFunction &MF) {
128 bool killPixel = false;
129 const R600RegisterInfo * RI =
130 static_cast<const R600RegisterInfo*>(TM.getRegisterInfo());
131 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
132 const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
134 for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
136 MachineBasicBlock &MBB = *BB;
137 for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
139 MachineInstr &MI = *I;
140 if (MI.getOpcode() == AMDGPU::KILLGT)
142 unsigned numOperands = MI.getNumOperands();
143 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
144 MachineOperand & MO = MI.getOperand(op_idx);
147 unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff;
149 // Register with value > 127 aren't GPR
152 MaxGPR = std::max(MaxGPR, HWReg);
158 if (STM.getGeneration() >= AMDGPUSubtarget::EVERGREEN) {
159 // Evergreen / Northern Islands
160 switch (MFI->ShaderType) {
161 default: // Fall through
162 case ShaderType::COMPUTE: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
163 case ShaderType::GEOMETRY: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
164 case ShaderType::PIXEL: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
165 case ShaderType::VERTEX: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
169 switch (MFI->ShaderType) {
170 default: // Fall through
171 case ShaderType::GEOMETRY: // Fall through
172 case ShaderType::COMPUTE: // Fall through
173 case ShaderType::VERTEX: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
174 case ShaderType::PIXEL: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
178 OutStreamer.EmitIntValue(RsrcReg, 4);
179 OutStreamer.EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
180 S_STACK_SIZE(MFI->StackSize), 4);
181 OutStreamer.EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
182 OutStreamer.EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
184 if (MFI->ShaderType == ShaderType::COMPUTE) {
185 OutStreamer.EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4);
186 OutStreamer.EmitIntValue(RoundUpToAlignment(MFI->LDSSize, 4) >> 2, 4);
190 void AMDGPUAsmPrinter::findNumUsedRegistersSI(MachineFunction &MF,
192 unsigned &NumVGPR) const {
193 unsigned MaxSGPR = 0;
194 unsigned MaxVGPR = 0;
195 bool VCCUsed = false;
196 const SIRegisterInfo * RI =
197 static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
199 for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
201 MachineBasicBlock &MBB = *BB;
202 for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
204 MachineInstr &MI = *I;
206 unsigned numOperands = MI.getNumOperands();
207 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
208 MachineOperand &MO = MI.getOperand(op_idx);
215 unsigned reg = MO.getReg();
216 if (reg == AMDGPU::VCC) {
229 if (AMDGPU::SReg_32RegClass.contains(reg)) {
232 } else if (AMDGPU::VReg_32RegClass.contains(reg)) {
235 } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
238 } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
241 } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
244 } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
247 } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
250 } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
253 } else if (AMDGPU::VReg_256RegClass.contains(reg)) {
256 } else if (AMDGPU::SReg_512RegClass.contains(reg)) {
259 } else if (AMDGPU::VReg_512RegClass.contains(reg)) {
263 llvm_unreachable("Unknown register class");
265 unsigned hwReg = RI->getEncodingValue(reg) & 0xff;
266 unsigned maxUsed = hwReg + width - 1;
268 MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
270 MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
283 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &Out,
284 MachineFunction &MF) const {
285 findNumUsedRegistersSI(MF, Out.NumSGPR, Out.NumVGPR);
288 void AMDGPUAsmPrinter::EmitProgramInfoSI(MachineFunction &MF,
289 const SIProgramInfo &KernelInfo) {
290 const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
292 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
294 switch (MFI->ShaderType) {
295 default: // Fall through
296 case ShaderType::COMPUTE: RsrcReg = R_00B848_COMPUTE_PGM_RSRC1; break;
297 case ShaderType::GEOMETRY: RsrcReg = R_00B228_SPI_SHADER_PGM_RSRC1_GS; break;
298 case ShaderType::PIXEL: RsrcReg = R_00B028_SPI_SHADER_PGM_RSRC1_PS; break;
299 case ShaderType::VERTEX: RsrcReg = R_00B128_SPI_SHADER_PGM_RSRC1_VS; break;
302 OutStreamer.EmitIntValue(RsrcReg, 4);
303 OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.NumVGPR / 4) |
304 S_00B028_SGPRS(KernelInfo.NumSGPR / 8), 4);
306 unsigned LDSAlignShift;
307 if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
308 // LDS is allocated in 64 dword blocks
311 // LDS is allocated in 128 dword blocks
315 RoundUpToAlignment(MFI->LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
317 if (MFI->ShaderType == ShaderType::COMPUTE) {
318 OutStreamer.EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
319 OutStreamer.EmitIntValue(S_00B84C_LDS_SIZE(LDSBlocks), 4);
321 if (MFI->ShaderType == ShaderType::PIXEL) {
322 OutStreamer.EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
323 OutStreamer.EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(LDSBlocks), 4);
324 OutStreamer.EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
325 OutStreamer.EmitIntValue(MFI->PSInputAddr, 4);