1 //===-- AMDGPUAsmPrinter.cpp - AMDGPU Assebly printer  --------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 ///
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.
15 //
16 //===----------------------------------------------------------------------===//
17 //
18 
19 #include "AMDGPUAsmPrinter.h"
20 #include "MCTargetDesc/AMDGPUTargetStreamer.h"
21 #include "InstPrinter/AMDGPUInstPrinter.h"
22 #include "Utils/AMDGPUBaseInfo.h"
23 #include "AMDGPU.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 "SIInstrInfo.h"
32 #include "SIRegisterInfo.h"
33 #include "llvm/CodeGen/MachineFrameInfo.h"
34 #include "llvm/IR/DiagnosticInfo.h"
35 #include "llvm/MC/MCContext.h"
36 #include "llvm/MC/MCSectionELF.h"
37 #include "llvm/MC/MCStreamer.h"
38 #include "llvm/Support/ELF.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/TargetRegistry.h"
41 #include "llvm/Target/TargetLoweringObjectFile.h"
42 
43 using namespace llvm;
44 
45 // TODO: This should get the default rounding mode from the kernel. We just set
46 // the default here, but this could change if the OpenCL rounding mode pragmas
47 // are used.
48 //
49 // The denormal mode here should match what is reported by the OpenCL runtime
50 // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but
51 // can also be override to flush with the -cl-denorms-are-zero compiler flag.
52 //
53 // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double
54 // precision, and leaves single precision to flush all and does not report
55 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
56 // CL_FP_DENORM for both.
57 //
58 // FIXME: It seems some instructions do not support single precision denormals
59 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
60 // and sin_f32, cos_f32 on most parts).
61 
62 // We want to use these instructions, and using fp32 denormals also causes
63 // instructions to run at the double precision rate for the device so it's
64 // probably best to just report no single precision denormals.
65 static uint32_t getFPMode(const MachineFunction &F) {
66   const SISubtarget& ST = F.getSubtarget<SISubtarget>();
67   // TODO: Is there any real use for the flush in only / flush out only modes?
68 
69   uint32_t FP32Denormals =
70     ST.hasFP32Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
71 
72   uint32_t FP64Denormals =
73     ST.hasFP64Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
74 
75   return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
76          FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
77          FP_DENORM_MODE_SP(FP32Denormals) |
78          FP_DENORM_MODE_DP(FP64Denormals);
79 }
80 
81 static AsmPrinter *
82 createAMDGPUAsmPrinterPass(TargetMachine &tm,
83                            std::unique_ptr<MCStreamer> &&Streamer) {
84   return new AMDGPUAsmPrinter(tm, std::move(Streamer));
85 }
86 
87 extern "C" void LLVMInitializeAMDGPUAsmPrinter() {
88   TargetRegistry::RegisterAsmPrinter(getTheAMDGPUTarget(),
89                                      createAMDGPUAsmPrinterPass);
90   TargetRegistry::RegisterAsmPrinter(getTheGCNTarget(),
91                                      createAMDGPUAsmPrinterPass);
92 }
93 
94 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM,
95                                    std::unique_ptr<MCStreamer> Streamer)
96   : AsmPrinter(TM, std::move(Streamer)) {}
97 
98 StringRef AMDGPUAsmPrinter::getPassName() const {
99   return "AMDGPU Assembly Printer";
100 }
101 
102 void AMDGPUAsmPrinter::EmitStartOfAsmFile(Module &M) {
103   if (TM.getTargetTriple().getOS() != Triple::AMDHSA)
104     return;
105 
106   // Need to construct an MCSubtargetInfo here in case we have no functions
107   // in the module.
108   std::unique_ptr<MCSubtargetInfo> STI(TM.getTarget().createMCSubtargetInfo(
109         TM.getTargetTriple().str(), TM.getTargetCPU(),
110         TM.getTargetFeatureString()));
111 
112   AMDGPUTargetStreamer *TS =
113       static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
114 
115   TS->EmitDirectiveHSACodeObjectVersion(2, 1);
116 
117   AMDGPU::IsaVersion ISA = AMDGPU::getIsaVersion(STI->getFeatureBits());
118   TS->EmitDirectiveHSACodeObjectISA(ISA.Major, ISA.Minor, ISA.Stepping,
119                                     "AMD", "AMDGPU");
120 
121   // Emit runtime metadata.
122   TS->emitRuntimeMetadata(M);
123 }
124 
125 bool AMDGPUAsmPrinter::isBlockOnlyReachableByFallthrough(
126   const MachineBasicBlock *MBB) const {
127   if (!AsmPrinter::isBlockOnlyReachableByFallthrough(MBB))
128     return false;
129 
130   if (MBB->empty())
131     return true;
132 
133   // If this is a block implementing a long branch, an expression relative to
134   // the start of the block is needed.  to the start of the block.
135   // XXX - Is there a smarter way to check this?
136   return (MBB->back().getOpcode() != AMDGPU::S_SETPC_B64);
137 }
138 
139 
140 void AMDGPUAsmPrinter::EmitFunctionBodyStart() {
141   const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
142   SIProgramInfo KernelInfo;
143   if (STM.isAmdCodeObjectV2()) {
144     getSIProgramInfo(KernelInfo, *MF);
145     EmitAmdKernelCodeT(*MF, KernelInfo);
146   }
147 }
148 
149 void AMDGPUAsmPrinter::EmitFunctionEntryLabel() {
150   const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
151   const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>();
152   if (MFI->isKernel() && STM.isAmdCodeObjectV2()) {
153     AMDGPUTargetStreamer *TS =
154         static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
155     SmallString<128> SymbolName;
156     getNameWithPrefix(SymbolName, MF->getFunction()),
157     TS->EmitAMDGPUSymbolType(SymbolName, ELF::STT_AMDGPU_HSA_KERNEL);
158   }
159 
160   AsmPrinter::EmitFunctionEntryLabel();
161 }
162 
163 void AMDGPUAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
164 
165   // Group segment variables aren't emitted in HSA.
166   if (AMDGPU::isGroupSegment(GV))
167     return;
168 
169   AsmPrinter::EmitGlobalVariable(GV);
170 }
171 
172 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
173 
174   // The starting address of all shader programs must be 256 bytes aligned.
175   MF.setAlignment(8);
176 
177   SetupMachineFunction(MF);
178 
179   MCContext &Context = getObjFileLowering().getContext();
180   MCSectionELF *ConfigSection =
181       Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);
182   OutStreamer->SwitchSection(ConfigSection);
183 
184   const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>();
185   SIProgramInfo KernelInfo;
186   if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
187     getSIProgramInfo(KernelInfo, MF);
188     if (!STM.isAmdHsaOS()) {
189       EmitProgramInfoSI(MF, KernelInfo);
190     }
191   } else {
192     EmitProgramInfoR600(MF);
193   }
194 
195   DisasmLines.clear();
196   HexLines.clear();
197   DisasmLineMaxLen = 0;
198 
199   EmitFunctionBody();
200 
201   if (isVerbose()) {
202     MCSectionELF *CommentSection =
203         Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0);
204     OutStreamer->SwitchSection(CommentSection);
205 
206     if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
207       OutStreamer->emitRawComment(" Kernel info:", false);
208       OutStreamer->emitRawComment(" codeLenInByte = " + Twine(KernelInfo.CodeLen),
209                                   false);
210       OutStreamer->emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR),
211                                   false);
212       OutStreamer->emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
213                                   false);
214       OutStreamer->emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode),
215                                   false);
216       OutStreamer->emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
217                                   false);
218       OutStreamer->emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize),
219                                   false);
220       OutStreamer->emitRawComment(" LDSByteSize: " + Twine(KernelInfo.LDSSize) +
221                                   " bytes/workgroup (compile time only)", false);
222 
223       OutStreamer->emitRawComment(" SGPRBlocks: " +
224                                   Twine(KernelInfo.SGPRBlocks), false);
225       OutStreamer->emitRawComment(" VGPRBlocks: " +
226                                   Twine(KernelInfo.VGPRBlocks), false);
227 
228       OutStreamer->emitRawComment(" NumSGPRsForWavesPerEU: " +
229                                   Twine(KernelInfo.NumSGPRsForWavesPerEU), false);
230       OutStreamer->emitRawComment(" NumVGPRsForWavesPerEU: " +
231                                   Twine(KernelInfo.NumVGPRsForWavesPerEU), false);
232 
233       OutStreamer->emitRawComment(" ReservedVGPRFirst: " + Twine(KernelInfo.ReservedVGPRFirst),
234                                   false);
235       OutStreamer->emitRawComment(" ReservedVGPRCount: " + Twine(KernelInfo.ReservedVGPRCount),
236                                   false);
237 
238       if (MF.getSubtarget<SISubtarget>().debuggerEmitPrologue()) {
239         OutStreamer->emitRawComment(" DebuggerWavefrontPrivateSegmentOffsetSGPR: s" +
240                                     Twine(KernelInfo.DebuggerWavefrontPrivateSegmentOffsetSGPR), false);
241         OutStreamer->emitRawComment(" DebuggerPrivateSegmentBufferSGPR: s" +
242                                     Twine(KernelInfo.DebuggerPrivateSegmentBufferSGPR), false);
243       }
244 
245       OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:USER_SGPR: " +
246                                   Twine(G_00B84C_USER_SGPR(KernelInfo.ComputePGMRSrc2)),
247                                   false);
248       OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_X_EN: " +
249                                   Twine(G_00B84C_TGID_X_EN(KernelInfo.ComputePGMRSrc2)),
250                                   false);
251       OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Y_EN: " +
252                                   Twine(G_00B84C_TGID_Y_EN(KernelInfo.ComputePGMRSrc2)),
253                                   false);
254       OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Z_EN: " +
255                                   Twine(G_00B84C_TGID_Z_EN(KernelInfo.ComputePGMRSrc2)),
256                                   false);
257       OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " +
258                                   Twine(G_00B84C_TIDIG_COMP_CNT(KernelInfo.ComputePGMRSrc2)),
259                                   false);
260 
261     } else {
262       R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
263       OutStreamer->emitRawComment(
264         Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->CFStackSize)));
265     }
266   }
267 
268   if (STM.dumpCode()) {
269 
270     OutStreamer->SwitchSection(
271         Context.getELFSection(".AMDGPU.disasm", ELF::SHT_NOTE, 0));
272 
273     for (size_t i = 0; i < DisasmLines.size(); ++i) {
274       std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
275       Comment += " ; " + HexLines[i] + "\n";
276 
277       OutStreamer->EmitBytes(StringRef(DisasmLines[i]));
278       OutStreamer->EmitBytes(StringRef(Comment));
279     }
280   }
281 
282   return false;
283 }
284 
285 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
286   unsigned MaxGPR = 0;
287   bool killPixel = false;
288   const R600Subtarget &STM = MF.getSubtarget<R600Subtarget>();
289   const R600RegisterInfo *RI = STM.getRegisterInfo();
290   const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
291 
292   for (const MachineBasicBlock &MBB : MF) {
293     for (const MachineInstr &MI : MBB) {
294       if (MI.getOpcode() == AMDGPU::KILLGT)
295         killPixel = true;
296       unsigned numOperands = MI.getNumOperands();
297       for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
298         const MachineOperand &MO = MI.getOperand(op_idx);
299         if (!MO.isReg())
300           continue;
301         unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff;
302 
303         // Register with value > 127 aren't GPR
304         if (HWReg > 127)
305           continue;
306         MaxGPR = std::max(MaxGPR, HWReg);
307       }
308     }
309   }
310 
311   unsigned RsrcReg;
312   if (STM.getGeneration() >= R600Subtarget::EVERGREEN) {
313     // Evergreen / Northern Islands
314     switch (MF.getFunction()->getCallingConv()) {
315     default: LLVM_FALLTHROUGH;
316     case CallingConv::AMDGPU_CS: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
317     case CallingConv::AMDGPU_GS: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
318     case CallingConv::AMDGPU_PS: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
319     case CallingConv::AMDGPU_VS: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
320     }
321   } else {
322     // R600 / R700
323     switch (MF.getFunction()->getCallingConv()) {
324     default: LLVM_FALLTHROUGH;
325     case CallingConv::AMDGPU_GS: LLVM_FALLTHROUGH;
326     case CallingConv::AMDGPU_CS: LLVM_FALLTHROUGH;
327     case CallingConv::AMDGPU_VS: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
328     case CallingConv::AMDGPU_PS: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
329     }
330   }
331 
332   OutStreamer->EmitIntValue(RsrcReg, 4);
333   OutStreamer->EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
334                            S_STACK_SIZE(MFI->CFStackSize), 4);
335   OutStreamer->EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
336   OutStreamer->EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
337 
338   if (AMDGPU::isCompute(MF.getFunction()->getCallingConv())) {
339     OutStreamer->EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4);
340     OutStreamer->EmitIntValue(alignTo(MFI->getLDSSize(), 4) >> 2, 4);
341   }
342 }
343 
344 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
345                                         const MachineFunction &MF) const {
346   const SISubtarget &STM = MF.getSubtarget<SISubtarget>();
347   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
348   uint64_t CodeSize = 0;
349   unsigned MaxSGPR = 0;
350   unsigned MaxVGPR = 0;
351   bool VCCUsed = false;
352   bool FlatUsed = false;
353   const SIRegisterInfo *RI = STM.getRegisterInfo();
354   const SIInstrInfo *TII = STM.getInstrInfo();
355 
356   for (const MachineBasicBlock &MBB : MF) {
357     for (const MachineInstr &MI : MBB) {
358       // TODO: CodeSize should account for multiple functions.
359 
360       // TODO: Should we count size of debug info?
361       if (MI.isDebugValue())
362         continue;
363 
364       if (isVerbose())
365         CodeSize += TII->getInstSizeInBytes(MI);
366 
367       unsigned numOperands = MI.getNumOperands();
368       for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
369         const MachineOperand &MO = MI.getOperand(op_idx);
370         unsigned width = 0;
371         bool isSGPR = false;
372 
373         if (!MO.isReg())
374           continue;
375 
376         unsigned reg = MO.getReg();
377         switch (reg) {
378         case AMDGPU::EXEC:
379         case AMDGPU::EXEC_LO:
380         case AMDGPU::EXEC_HI:
381         case AMDGPU::SCC:
382         case AMDGPU::M0:
383           continue;
384 
385         case AMDGPU::VCC:
386         case AMDGPU::VCC_LO:
387         case AMDGPU::VCC_HI:
388           VCCUsed = true;
389           continue;
390 
391         case AMDGPU::FLAT_SCR:
392         case AMDGPU::FLAT_SCR_LO:
393         case AMDGPU::FLAT_SCR_HI:
394           // Even if FLAT_SCRATCH is implicitly used, it has no effect if flat
395           // instructions aren't used to access the scratch buffer.
396           if (MFI->hasFlatScratchInit())
397             FlatUsed = true;
398           continue;
399 
400         case AMDGPU::TBA:
401         case AMDGPU::TBA_LO:
402         case AMDGPU::TBA_HI:
403         case AMDGPU::TMA:
404         case AMDGPU::TMA_LO:
405         case AMDGPU::TMA_HI:
406           llvm_unreachable("trap handler registers should not be used");
407 
408         default:
409           break;
410         }
411 
412         if (AMDGPU::SReg_32RegClass.contains(reg)) {
413           assert(!AMDGPU::TTMP_32RegClass.contains(reg) &&
414                  "trap handler registers should not be used");
415           isSGPR = true;
416           width = 1;
417         } else if (AMDGPU::VGPR_32RegClass.contains(reg)) {
418           isSGPR = false;
419           width = 1;
420         } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
421           assert(!AMDGPU::TTMP_64RegClass.contains(reg) &&
422                  "trap handler registers should not be used");
423           isSGPR = true;
424           width = 2;
425         } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
426           isSGPR = false;
427           width = 2;
428         } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
429           isSGPR = false;
430           width = 3;
431         } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
432           isSGPR = true;
433           width = 4;
434         } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
435           isSGPR = false;
436           width = 4;
437         } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
438           isSGPR = true;
439           width = 8;
440         } else if (AMDGPU::VReg_256RegClass.contains(reg)) {
441           isSGPR = false;
442           width = 8;
443         } else if (AMDGPU::SReg_512RegClass.contains(reg)) {
444           isSGPR = true;
445           width = 16;
446         } else if (AMDGPU::VReg_512RegClass.contains(reg)) {
447           isSGPR = false;
448           width = 16;
449         } else {
450           llvm_unreachable("Unknown register class");
451         }
452         unsigned hwReg = RI->getEncodingValue(reg) & 0xff;
453         unsigned maxUsed = hwReg + width - 1;
454         if (isSGPR) {
455           MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
456         } else {
457           MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
458         }
459       }
460     }
461   }
462 
463   unsigned ExtraSGPRs = 0;
464 
465   if (VCCUsed)
466     ExtraSGPRs = 2;
467 
468   if (STM.getGeneration() < SISubtarget::VOLCANIC_ISLANDS) {
469     if (FlatUsed)
470       ExtraSGPRs = 4;
471   } else {
472     if (STM.isXNACKEnabled())
473       ExtraSGPRs = 4;
474 
475     if (FlatUsed)
476       ExtraSGPRs = 6;
477   }
478 
479   // Record first reserved register and reserved register count fields, and
480   // update max register counts if "amdgpu-debugger-reserve-regs" attribute was
481   // requested.
482   ProgInfo.ReservedVGPRFirst = STM.debuggerReserveRegs() ? MaxVGPR + 1 : 0;
483   ProgInfo.ReservedVGPRCount = RI->getNumDebuggerReservedVGPRs(STM);
484 
485   // Update DebuggerWavefrontPrivateSegmentOffsetSGPR and
486   // DebuggerPrivateSegmentBufferSGPR fields if "amdgpu-debugger-emit-prologue"
487   // attribute was requested.
488   if (STM.debuggerEmitPrologue()) {
489     ProgInfo.DebuggerWavefrontPrivateSegmentOffsetSGPR =
490       RI->getHWRegIndex(MFI->getScratchWaveOffsetReg());
491     ProgInfo.DebuggerPrivateSegmentBufferSGPR =
492       RI->getHWRegIndex(MFI->getScratchRSrcReg());
493   }
494 
495   // Check the addressable register limit before we add ExtraSGPRs.
496   if (STM.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS &&
497       !STM.hasSGPRInitBug()) {
498     unsigned MaxAddressableNumSGPRs = STM.getMaxNumSGPRs();
499     if (MaxSGPR + 1 > MaxAddressableNumSGPRs) {
500       // This can happen due to a compiler bug or when using inline asm.
501       LLVMContext &Ctx = MF.getFunction()->getContext();
502       DiagnosticInfoResourceLimit Diag(*MF.getFunction(),
503                                        "addressable scalar registers",
504                                        MaxSGPR + 1, DS_Error,
505                                        DK_ResourceLimit, MaxAddressableNumSGPRs);
506       Ctx.diagnose(Diag);
507       MaxSGPR = MaxAddressableNumSGPRs - 1;
508     }
509   }
510 
511   // Account for extra SGPRs and VGPRs reserved for debugger use.
512   MaxSGPR += ExtraSGPRs;
513   MaxVGPR += RI->getNumDebuggerReservedVGPRs(STM);
514 
515   // We found the maximum register index. They start at 0, so add one to get the
516   // number of registers.
517   ProgInfo.NumVGPR = MaxVGPR + 1;
518   ProgInfo.NumSGPR = MaxSGPR + 1;
519 
520   // Adjust number of registers used to meet default/requested minimum/maximum
521   // number of waves per execution unit request.
522   ProgInfo.NumSGPRsForWavesPerEU = std::max(
523     ProgInfo.NumSGPR, RI->getMinNumSGPRs(STM, MFI->getMaxWavesPerEU()));
524   ProgInfo.NumVGPRsForWavesPerEU = std::max(
525     ProgInfo.NumVGPR, RI->getMinNumVGPRs(MFI->getMaxWavesPerEU()));
526 
527   if (STM.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS ||
528       STM.hasSGPRInitBug()) {
529     unsigned MaxNumSGPRs = STM.getMaxNumSGPRs();
530     if (ProgInfo.NumSGPR > MaxNumSGPRs) {
531       // This can happen due to a compiler bug or when using inline asm to use the
532       // registers which are usually reserved for vcc etc.
533 
534       LLVMContext &Ctx = MF.getFunction()->getContext();
535       DiagnosticInfoResourceLimit Diag(*MF.getFunction(),
536                                        "scalar registers",
537                                        ProgInfo.NumSGPR, DS_Error,
538                                        DK_ResourceLimit, MaxNumSGPRs);
539       Ctx.diagnose(Diag);
540       ProgInfo.NumSGPR = MaxNumSGPRs;
541       ProgInfo.NumSGPRsForWavesPerEU = MaxNumSGPRs;
542     }
543   }
544 
545   if (STM.hasSGPRInitBug()) {
546     ProgInfo.NumSGPR = SISubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG;
547     ProgInfo.NumSGPRsForWavesPerEU = SISubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG;
548   }
549 
550   if (MFI->NumUserSGPRs > STM.getMaxNumUserSGPRs()) {
551     LLVMContext &Ctx = MF.getFunction()->getContext();
552     DiagnosticInfoResourceLimit Diag(*MF.getFunction(), "user SGPRs",
553                                      MFI->NumUserSGPRs, DS_Error);
554     Ctx.diagnose(Diag);
555   }
556 
557   if (MFI->getLDSSize() > static_cast<unsigned>(STM.getLocalMemorySize())) {
558     LLVMContext &Ctx = MF.getFunction()->getContext();
559     DiagnosticInfoResourceLimit Diag(*MF.getFunction(), "local memory",
560                                      MFI->getLDSSize(), DS_Error);
561     Ctx.diagnose(Diag);
562   }
563 
564   // SGPRBlocks is actual number of SGPR blocks minus 1.
565   ProgInfo.SGPRBlocks = alignTo(ProgInfo.NumSGPRsForWavesPerEU,
566                                 RI->getSGPRAllocGranule());
567   ProgInfo.SGPRBlocks = ProgInfo.SGPRBlocks / RI->getSGPRAllocGranule() - 1;
568 
569   // VGPRBlocks is actual number of VGPR blocks minus 1.
570   ProgInfo.VGPRBlocks = alignTo(ProgInfo.NumVGPRsForWavesPerEU,
571                                 RI->getVGPRAllocGranule());
572   ProgInfo.VGPRBlocks = ProgInfo.VGPRBlocks / RI->getVGPRAllocGranule() - 1;
573 
574   // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
575   // register.
576   ProgInfo.FloatMode = getFPMode(MF);
577 
578   ProgInfo.IEEEMode = STM.enableIEEEBit(MF);
579 
580   // Make clamp modifier on NaN input returns 0.
581   ProgInfo.DX10Clamp = 1;
582 
583   const MachineFrameInfo &FrameInfo = MF.getFrameInfo();
584   ProgInfo.ScratchSize = FrameInfo.getStackSize();
585 
586   ProgInfo.FlatUsed = FlatUsed;
587   ProgInfo.VCCUsed = VCCUsed;
588   ProgInfo.CodeLen = CodeSize;
589 
590   unsigned LDSAlignShift;
591   if (STM.getGeneration() < SISubtarget::SEA_ISLANDS) {
592     // LDS is allocated in 64 dword blocks.
593     LDSAlignShift = 8;
594   } else {
595     // LDS is allocated in 128 dword blocks.
596     LDSAlignShift = 9;
597   }
598 
599   unsigned LDSSpillSize =
600     MFI->LDSWaveSpillSize * MFI->getMaxFlatWorkGroupSize();
601 
602   ProgInfo.LDSSize = MFI->getLDSSize() + LDSSpillSize;
603   ProgInfo.LDSBlocks =
604       alignTo(ProgInfo.LDSSize, 1ULL << LDSAlignShift) >> LDSAlignShift;
605 
606   // Scratch is allocated in 256 dword blocks.
607   unsigned ScratchAlignShift = 10;
608   // We need to program the hardware with the amount of scratch memory that
609   // is used by the entire wave.  ProgInfo.ScratchSize is the amount of
610   // scratch memory used per thread.
611   ProgInfo.ScratchBlocks =
612       alignTo(ProgInfo.ScratchSize * STM.getWavefrontSize(),
613               1ULL << ScratchAlignShift) >>
614       ScratchAlignShift;
615 
616   ProgInfo.ComputePGMRSrc1 =
617       S_00B848_VGPRS(ProgInfo.VGPRBlocks) |
618       S_00B848_SGPRS(ProgInfo.SGPRBlocks) |
619       S_00B848_PRIORITY(ProgInfo.Priority) |
620       S_00B848_FLOAT_MODE(ProgInfo.FloatMode) |
621       S_00B848_PRIV(ProgInfo.Priv) |
622       S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) |
623       S_00B848_DEBUG_MODE(ProgInfo.DebugMode) |
624       S_00B848_IEEE_MODE(ProgInfo.IEEEMode);
625 
626   // 0 = X, 1 = XY, 2 = XYZ
627   unsigned TIDIGCompCnt = 0;
628   if (MFI->hasWorkItemIDZ())
629     TIDIGCompCnt = 2;
630   else if (MFI->hasWorkItemIDY())
631     TIDIGCompCnt = 1;
632 
633   ProgInfo.ComputePGMRSrc2 =
634       S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
635       S_00B84C_USER_SGPR(MFI->getNumUserSGPRs()) |
636       S_00B84C_TGID_X_EN(MFI->hasWorkGroupIDX()) |
637       S_00B84C_TGID_Y_EN(MFI->hasWorkGroupIDY()) |
638       S_00B84C_TGID_Z_EN(MFI->hasWorkGroupIDZ()) |
639       S_00B84C_TG_SIZE_EN(MFI->hasWorkGroupInfo()) |
640       S_00B84C_TIDIG_COMP_CNT(TIDIGCompCnt) |
641       S_00B84C_EXCP_EN_MSB(0) |
642       S_00B84C_LDS_SIZE(ProgInfo.LDSBlocks) |
643       S_00B84C_EXCP_EN(0);
644 }
645 
646 static unsigned getRsrcReg(CallingConv::ID CallConv) {
647   switch (CallConv) {
648   default: LLVM_FALLTHROUGH;
649   case CallingConv::AMDGPU_CS: return R_00B848_COMPUTE_PGM_RSRC1;
650   case CallingConv::AMDGPU_GS: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
651   case CallingConv::AMDGPU_PS: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
652   case CallingConv::AMDGPU_VS: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
653   }
654 }
655 
656 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
657                                          const SIProgramInfo &KernelInfo) {
658   const SISubtarget &STM = MF.getSubtarget<SISubtarget>();
659   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
660   unsigned RsrcReg = getRsrcReg(MF.getFunction()->getCallingConv());
661 
662   if (AMDGPU::isCompute(MF.getFunction()->getCallingConv())) {
663     OutStreamer->EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
664 
665     OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc1, 4);
666 
667     OutStreamer->EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
668     OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc2, 4);
669 
670     OutStreamer->EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4);
671     OutStreamer->EmitIntValue(S_00B860_WAVESIZE(KernelInfo.ScratchBlocks), 4);
672 
673     // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
674     // 0" comment but I don't see a corresponding field in the register spec.
675   } else {
676     OutStreamer->EmitIntValue(RsrcReg, 4);
677     OutStreamer->EmitIntValue(S_00B028_VGPRS(KernelInfo.VGPRBlocks) |
678                               S_00B028_SGPRS(KernelInfo.SGPRBlocks), 4);
679     if (STM.isVGPRSpillingEnabled(*MF.getFunction())) {
680       OutStreamer->EmitIntValue(R_0286E8_SPI_TMPRING_SIZE, 4);
681       OutStreamer->EmitIntValue(S_0286E8_WAVESIZE(KernelInfo.ScratchBlocks), 4);
682     }
683   }
684 
685   if (MF.getFunction()->getCallingConv() == CallingConv::AMDGPU_PS) {
686     OutStreamer->EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
687     OutStreamer->EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(KernelInfo.LDSBlocks), 4);
688     OutStreamer->EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
689     OutStreamer->EmitIntValue(MFI->PSInputEna, 4);
690     OutStreamer->EmitIntValue(R_0286D0_SPI_PS_INPUT_ADDR, 4);
691     OutStreamer->EmitIntValue(MFI->getPSInputAddr(), 4);
692   }
693 
694   OutStreamer->EmitIntValue(R_SPILLED_SGPRS, 4);
695   OutStreamer->EmitIntValue(MFI->getNumSpilledSGPRs(), 4);
696   OutStreamer->EmitIntValue(R_SPILLED_VGPRS, 4);
697   OutStreamer->EmitIntValue(MFI->getNumSpilledVGPRs(), 4);
698 }
699 
700 // This is supposed to be log2(Size)
701 static amd_element_byte_size_t getElementByteSizeValue(unsigned Size) {
702   switch (Size) {
703   case 4:
704     return AMD_ELEMENT_4_BYTES;
705   case 8:
706     return AMD_ELEMENT_8_BYTES;
707   case 16:
708     return AMD_ELEMENT_16_BYTES;
709   default:
710     llvm_unreachable("invalid private_element_size");
711   }
712 }
713 
714 void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF,
715                                          const SIProgramInfo &KernelInfo) const {
716   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
717   const SISubtarget &STM = MF.getSubtarget<SISubtarget>();
718   amd_kernel_code_t header;
719 
720   AMDGPU::initDefaultAMDKernelCodeT(header, STM.getFeatureBits());
721 
722   header.compute_pgm_resource_registers =
723       KernelInfo.ComputePGMRSrc1 |
724       (KernelInfo.ComputePGMRSrc2 << 32);
725   header.code_properties = AMD_CODE_PROPERTY_IS_PTR64;
726 
727 
728   AMD_HSA_BITS_SET(header.code_properties,
729                    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE,
730                    getElementByteSizeValue(STM.getMaxPrivateElementSize()));
731 
732   if (MFI->hasPrivateSegmentBuffer()) {
733     header.code_properties |=
734       AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
735   }
736 
737   if (MFI->hasDispatchPtr())
738     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
739 
740   if (MFI->hasQueuePtr())
741     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
742 
743   if (MFI->hasKernargSegmentPtr())
744     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
745 
746   if (MFI->hasDispatchID())
747     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
748 
749   if (MFI->hasFlatScratchInit())
750     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
751 
752   // TODO: Private segment size
753 
754   if (MFI->hasGridWorkgroupCountX()) {
755     header.code_properties |=
756       AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X;
757   }
758 
759   if (MFI->hasGridWorkgroupCountY()) {
760     header.code_properties |=
761       AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y;
762   }
763 
764   if (MFI->hasGridWorkgroupCountZ()) {
765     header.code_properties |=
766       AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z;
767   }
768 
769   if (MFI->hasDispatchPtr())
770     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
771 
772   if (STM.debuggerSupported())
773     header.code_properties |= AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED;
774 
775   if (STM.isXNACKEnabled())
776     header.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED;
777 
778   // FIXME: Should use getKernArgSize
779   header.kernarg_segment_byte_size =
780       STM.getKernArgSegmentSize(MFI->getABIArgOffset());
781   header.wavefront_sgpr_count = KernelInfo.NumSGPR;
782   header.workitem_vgpr_count = KernelInfo.NumVGPR;
783   header.workitem_private_segment_byte_size = KernelInfo.ScratchSize;
784   header.workgroup_group_segment_byte_size = KernelInfo.LDSSize;
785   header.reserved_vgpr_first = KernelInfo.ReservedVGPRFirst;
786   header.reserved_vgpr_count = KernelInfo.ReservedVGPRCount;
787 
788   // These alignment values are specified in powers of two, so alignment =
789   // 2^n.  The minimum alignment is 2^4 = 16.
790   header.kernarg_segment_alignment = std::max((size_t)4,
791       countTrailingZeros(MFI->getMaxKernArgAlign()));
792 
793   if (STM.debuggerEmitPrologue()) {
794     header.debug_wavefront_private_segment_offset_sgpr =
795       KernelInfo.DebuggerWavefrontPrivateSegmentOffsetSGPR;
796     header.debug_private_segment_buffer_sgpr =
797       KernelInfo.DebuggerPrivateSegmentBufferSGPR;
798   }
799 
800   AMDGPUTargetStreamer *TS =
801       static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer());
802 
803   OutStreamer->SwitchSection(getObjFileLowering().getTextSection());
804   TS->EmitAMDKernelCodeT(header);
805 }
806 
807 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
808                                        unsigned AsmVariant,
809                                        const char *ExtraCode, raw_ostream &O) {
810   if (ExtraCode && ExtraCode[0]) {
811     if (ExtraCode[1] != 0)
812       return true; // Unknown modifier.
813 
814     switch (ExtraCode[0]) {
815     default:
816       // See if this is a generic print operand
817       return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
818     case 'r':
819       break;
820     }
821   }
822 
823   AMDGPUInstPrinter::printRegOperand(MI->getOperand(OpNo).getReg(), O,
824                    *TM.getSubtargetImpl(*MF->getFunction())->getRegisterInfo());
825   return false;
826 }
827