1 //===-- AMDGPUAsmPrinter.cpp - AMDGPU assembly printer --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 ///
11 /// The AMDGPUAsmPrinter is used to print both assembly string and also binary
12 /// code.  When passed an MCAsmStreamer it prints assembly and when passed
13 /// an MCObjectStreamer it outputs binary code.
14 //
15 //===----------------------------------------------------------------------===//
16 //
17 
18 #include "AMDGPUAsmPrinter.h"
19 #include "AMDGPU.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDGPUTargetMachine.h"
22 #include "MCTargetDesc/AMDGPUInstPrinter.h"
23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
24 #include "MCTargetDesc/AMDGPUTargetStreamer.h"
25 #include "R600AsmPrinter.h"
26 #include "R600Defines.h"
27 #include "R600MachineFunctionInfo.h"
28 #include "R600RegisterInfo.h"
29 #include "SIDefines.h"
30 #include "SIInstrInfo.h"
31 #include "SIMachineFunctionInfo.h"
32 #include "SIRegisterInfo.h"
33 #include "TargetInfo/AMDGPUTargetInfo.h"
34 #include "Utils/AMDGPUBaseInfo.h"
35 #include "llvm/BinaryFormat/ELF.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/IR/DiagnosticInfo.h"
38 #include "llvm/MC/MCAssembler.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCSectionELF.h"
41 #include "llvm/MC/MCStreamer.h"
42 #include "llvm/Support/AMDGPUMetadata.h"
43 #include "llvm/Support/MathExtras.h"
44 #include "llvm/Support/TargetParser.h"
45 #include "llvm/Support/TargetRegistry.h"
46 #include "llvm/Target/TargetLoweringObjectFile.h"
47 
48 using namespace llvm;
49 using namespace llvm::AMDGPU;
50 using namespace llvm::AMDGPU::HSAMD;
51 
52 // This should get the default rounding mode from the kernel. We just set the
53 // default here, but this could change if the OpenCL rounding mode pragmas are
54 // used.
55 //
56 // The denormal mode here should match what is reported by the OpenCL runtime
57 // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but
58 // can also be override to flush with the -cl-denorms-are-zero compiler flag.
59 //
60 // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double
61 // precision, and leaves single precision to flush all and does not report
62 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
63 // CL_FP_DENORM for both.
64 //
65 // FIXME: It seems some instructions do not support single precision denormals
66 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
67 // and sin_f32, cos_f32 on most parts).
68 
69 // We want to use these instructions, and using fp32 denormals also causes
70 // instructions to run at the double precision rate for the device so it's
71 // probably best to just report no single precision denormals.
72 static uint32_t getFPMode(AMDGPU::SIModeRegisterDefaults Mode) {
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(Mode.fpDenormModeSPValue()) |
76          FP_DENORM_MODE_DP(Mode.fpDenormModeDPValue());
77 }
78 
79 static AsmPrinter *
80 createAMDGPUAsmPrinterPass(TargetMachine &tm,
81                            std::unique_ptr<MCStreamer> &&Streamer) {
82   return new AMDGPUAsmPrinter(tm, std::move(Streamer));
83 }
84 
85 extern "C" void LLVM_EXTERNAL_VISIBILITY LLVMInitializeAMDGPUAsmPrinter() {
86   TargetRegistry::RegisterAsmPrinter(getTheAMDGPUTarget(),
87                                      llvm::createR600AsmPrinterPass);
88   TargetRegistry::RegisterAsmPrinter(getTheGCNTarget(),
89                                      createAMDGPUAsmPrinterPass);
90 }
91 
92 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM,
93                                    std::unique_ptr<MCStreamer> Streamer)
94   : AsmPrinter(TM, std::move(Streamer)) {
95     if (IsaInfo::hasCodeObjectV3(getGlobalSTI()))
96       HSAMetadataStream.reset(new MetadataStreamerV3());
97     else
98       HSAMetadataStream.reset(new MetadataStreamerV2());
99 }
100 
101 StringRef AMDGPUAsmPrinter::getPassName() const {
102   return "AMDGPU Assembly Printer";
103 }
104 
105 const MCSubtargetInfo *AMDGPUAsmPrinter::getGlobalSTI() const {
106   return TM.getMCSubtargetInfo();
107 }
108 
109 AMDGPUTargetStreamer* AMDGPUAsmPrinter::getTargetStreamer() const {
110   if (!OutStreamer)
111     return nullptr;
112   return static_cast<AMDGPUTargetStreamer*>(OutStreamer->getTargetStreamer());
113 }
114 
115 void AMDGPUAsmPrinter::emitStartOfAsmFile(Module &M) {
116   if (IsaInfo::hasCodeObjectV3(getGlobalSTI())) {
117     std::string ExpectedTarget;
118     raw_string_ostream ExpectedTargetOS(ExpectedTarget);
119     IsaInfo::streamIsaVersion(getGlobalSTI(), ExpectedTargetOS);
120 
121     getTargetStreamer()->EmitDirectiveAMDGCNTarget(ExpectedTarget);
122   }
123 
124   if (TM.getTargetTriple().getOS() != Triple::AMDHSA &&
125       TM.getTargetTriple().getOS() != Triple::AMDPAL)
126     return;
127 
128   if (TM.getTargetTriple().getOS() == Triple::AMDHSA)
129     HSAMetadataStream->begin(M);
130 
131   if (TM.getTargetTriple().getOS() == Triple::AMDPAL)
132     getTargetStreamer()->getPALMetadata()->readFromIR(M);
133 
134   if (IsaInfo::hasCodeObjectV3(getGlobalSTI()))
135     return;
136 
137   // HSA emits NT_AMDGPU_HSA_CODE_OBJECT_VERSION for code objects v2.
138   if (TM.getTargetTriple().getOS() == Triple::AMDHSA)
139     getTargetStreamer()->EmitDirectiveHSACodeObjectVersion(2, 1);
140 
141   // HSA and PAL emit NT_AMDGPU_HSA_ISA for code objects v2.
142   IsaVersion Version = getIsaVersion(getGlobalSTI()->getCPU());
143   getTargetStreamer()->EmitDirectiveHSACodeObjectISA(
144       Version.Major, Version.Minor, Version.Stepping, "AMD", "AMDGPU");
145 }
146 
147 void AMDGPUAsmPrinter::emitEndOfAsmFile(Module &M) {
148   // Following code requires TargetStreamer to be present.
149   if (!getTargetStreamer())
150     return;
151 
152   if (!IsaInfo::hasCodeObjectV3(getGlobalSTI())) {
153     // Emit ISA Version (NT_AMD_AMDGPU_ISA).
154     std::string ISAVersionString;
155     raw_string_ostream ISAVersionStream(ISAVersionString);
156     IsaInfo::streamIsaVersion(getGlobalSTI(), ISAVersionStream);
157     getTargetStreamer()->EmitISAVersion(ISAVersionStream.str());
158   }
159 
160   // Emit HSA Metadata (NT_AMD_AMDGPU_HSA_METADATA).
161   if (TM.getTargetTriple().getOS() == Triple::AMDHSA) {
162     HSAMetadataStream->end();
163     bool Success = HSAMetadataStream->emitTo(*getTargetStreamer());
164     (void)Success;
165     assert(Success && "Malformed HSA Metadata");
166   }
167 }
168 
169 bool AMDGPUAsmPrinter::isBlockOnlyReachableByFallthrough(
170   const MachineBasicBlock *MBB) const {
171   if (!AsmPrinter::isBlockOnlyReachableByFallthrough(MBB))
172     return false;
173 
174   if (MBB->empty())
175     return true;
176 
177   // If this is a block implementing a long branch, an expression relative to
178   // the start of the block is needed.  to the start of the block.
179   // XXX - Is there a smarter way to check this?
180   return (MBB->back().getOpcode() != AMDGPU::S_SETPC_B64);
181 }
182 
183 void AMDGPUAsmPrinter::emitFunctionBodyStart() {
184   const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>();
185   if (!MFI.isEntryFunction())
186     return;
187 
188   const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>();
189   const Function &F = MF->getFunction();
190   if (!STM.hasCodeObjectV3() && STM.isAmdHsaOrMesa(F) &&
191       (F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
192        F.getCallingConv() == CallingConv::SPIR_KERNEL)) {
193     amd_kernel_code_t KernelCode;
194     getAmdKernelCode(KernelCode, CurrentProgramInfo, *MF);
195     getTargetStreamer()->EmitAMDKernelCodeT(KernelCode);
196   }
197 
198   if (STM.isAmdHsaOS())
199     HSAMetadataStream->emitKernel(*MF, CurrentProgramInfo);
200 }
201 
202 void AMDGPUAsmPrinter::emitFunctionBodyEnd() {
203   const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>();
204   if (!MFI.isEntryFunction())
205     return;
206 
207   if (!IsaInfo::hasCodeObjectV3(getGlobalSTI()) ||
208       TM.getTargetTriple().getOS() != Triple::AMDHSA)
209     return;
210 
211   auto &Streamer = getTargetStreamer()->getStreamer();
212   auto &Context = Streamer.getContext();
213   auto &ObjectFileInfo = *Context.getObjectFileInfo();
214   auto &ReadOnlySection = *ObjectFileInfo.getReadOnlySection();
215 
216   Streamer.PushSection();
217   Streamer.SwitchSection(&ReadOnlySection);
218 
219   // CP microcode requires the kernel descriptor to be allocated on 64 byte
220   // alignment.
221   Streamer.emitValueToAlignment(64, 0, 1, 0);
222   if (ReadOnlySection.getAlignment() < 64)
223     ReadOnlySection.setAlignment(Align(64));
224 
225   const MCSubtargetInfo &STI = MF->getSubtarget();
226 
227   SmallString<128> KernelName;
228   getNameWithPrefix(KernelName, &MF->getFunction());
229   getTargetStreamer()->EmitAmdhsaKernelDescriptor(
230       STI, KernelName, getAmdhsaKernelDescriptor(*MF, CurrentProgramInfo),
231       CurrentProgramInfo.NumVGPRsForWavesPerEU,
232       CurrentProgramInfo.NumSGPRsForWavesPerEU -
233           IsaInfo::getNumExtraSGPRs(&STI,
234                                     CurrentProgramInfo.VCCUsed,
235                                     CurrentProgramInfo.FlatUsed),
236       CurrentProgramInfo.VCCUsed, CurrentProgramInfo.FlatUsed,
237       hasXNACK(STI));
238 
239   Streamer.PopSection();
240 }
241 
242 void AMDGPUAsmPrinter::emitFunctionEntryLabel() {
243   if (IsaInfo::hasCodeObjectV3(getGlobalSTI()) &&
244       TM.getTargetTriple().getOS() == Triple::AMDHSA) {
245     AsmPrinter::emitFunctionEntryLabel();
246     return;
247   }
248 
249   const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
250   const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>();
251   if (MFI->isEntryFunction() && STM.isAmdHsaOrMesa(MF->getFunction())) {
252     SmallString<128> SymbolName;
253     getNameWithPrefix(SymbolName, &MF->getFunction()),
254     getTargetStreamer()->EmitAMDGPUSymbolType(
255         SymbolName, ELF::STT_AMDGPU_HSA_KERNEL);
256   }
257   if (DumpCodeInstEmitter) {
258     // Disassemble function name label to text.
259     DisasmLines.push_back(MF->getName().str() + ":");
260     DisasmLineMaxLen = std::max(DisasmLineMaxLen, DisasmLines.back().size());
261     HexLines.push_back("");
262   }
263 
264   AsmPrinter::emitFunctionEntryLabel();
265 }
266 
267 void AMDGPUAsmPrinter::emitBasicBlockStart(const MachineBasicBlock &MBB) {
268   if (DumpCodeInstEmitter && !isBlockOnlyReachableByFallthrough(&MBB)) {
269     // Write a line for the basic block label if it is not only fallthrough.
270     DisasmLines.push_back(
271         (Twine("BB") + Twine(getFunctionNumber())
272          + "_" + Twine(MBB.getNumber()) + ":").str());
273     DisasmLineMaxLen = std::max(DisasmLineMaxLen, DisasmLines.back().size());
274     HexLines.push_back("");
275   }
276   AsmPrinter::emitBasicBlockStart(MBB);
277 }
278 
279 void AMDGPUAsmPrinter::emitGlobalVariable(const GlobalVariable *GV) {
280   if (GV->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
281     if (GV->hasInitializer() && !isa<UndefValue>(GV->getInitializer())) {
282       OutContext.reportError({},
283                              Twine(GV->getName()) +
284                                  ": unsupported initializer for address space");
285       return;
286     }
287 
288     // LDS variables aren't emitted in HSA or PAL yet.
289     const Triple::OSType OS = TM.getTargetTriple().getOS();
290     if (OS == Triple::AMDHSA || OS == Triple::AMDPAL)
291       return;
292 
293     MCSymbol *GVSym = getSymbol(GV);
294 
295     GVSym->redefineIfPossible();
296     if (GVSym->isDefined() || GVSym->isVariable())
297       report_fatal_error("symbol '" + Twine(GVSym->getName()) +
298                          "' is already defined");
299 
300     const DataLayout &DL = GV->getParent()->getDataLayout();
301     uint64_t Size = DL.getTypeAllocSize(GV->getValueType());
302     unsigned Align = GV->getAlignment();
303     if (!Align)
304       Align = 4;
305 
306     emitVisibility(GVSym, GV->getVisibility(), !GV->isDeclaration());
307     emitLinkage(GV, GVSym);
308     if (auto TS = getTargetStreamer())
309       TS->emitAMDGPULDS(GVSym, Size, Align);
310     return;
311   }
312 
313   AsmPrinter::emitGlobalVariable(GV);
314 }
315 
316 bool AMDGPUAsmPrinter::doFinalization(Module &M) {
317   CallGraphResourceInfo.clear();
318 
319   // Pad with s_code_end to help tools and guard against instruction prefetch
320   // causing stale data in caches. Arguably this should be done by the linker,
321   // which is why this isn't done for Mesa.
322   const MCSubtargetInfo &STI = *getGlobalSTI();
323   if (AMDGPU::isGFX10(STI) &&
324       (STI.getTargetTriple().getOS() == Triple::AMDHSA ||
325        STI.getTargetTriple().getOS() == Triple::AMDPAL)) {
326     OutStreamer->SwitchSection(getObjFileLowering().getTextSection());
327     getTargetStreamer()->EmitCodeEnd();
328   }
329 
330   return AsmPrinter::doFinalization(M);
331 }
332 
333 // Print comments that apply to both callable functions and entry points.
334 void AMDGPUAsmPrinter::emitCommonFunctionComments(
335   uint32_t NumVGPR,
336   Optional<uint32_t> NumAGPR,
337   uint32_t TotalNumVGPR,
338   uint32_t NumSGPR,
339   uint64_t ScratchSize,
340   uint64_t CodeSize,
341   const AMDGPUMachineFunction *MFI) {
342   OutStreamer->emitRawComment(" codeLenInByte = " + Twine(CodeSize), false);
343   OutStreamer->emitRawComment(" NumSgprs: " + Twine(NumSGPR), false);
344   OutStreamer->emitRawComment(" NumVgprs: " + Twine(NumVGPR), false);
345   if (NumAGPR) {
346     OutStreamer->emitRawComment(" NumAgprs: " + Twine(*NumAGPR), false);
347     OutStreamer->emitRawComment(" TotalNumVgprs: " + Twine(TotalNumVGPR),
348                                 false);
349   }
350   OutStreamer->emitRawComment(" ScratchSize: " + Twine(ScratchSize), false);
351   OutStreamer->emitRawComment(" MemoryBound: " + Twine(MFI->isMemoryBound()),
352                               false);
353 }
354 
355 uint16_t AMDGPUAsmPrinter::getAmdhsaKernelCodeProperties(
356     const MachineFunction &MF) const {
357   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
358   uint16_t KernelCodeProperties = 0;
359 
360   if (MFI.hasPrivateSegmentBuffer()) {
361     KernelCodeProperties |=
362         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
363   }
364   if (MFI.hasDispatchPtr()) {
365     KernelCodeProperties |=
366         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
367   }
368   if (MFI.hasQueuePtr()) {
369     KernelCodeProperties |=
370         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
371   }
372   if (MFI.hasKernargSegmentPtr()) {
373     KernelCodeProperties |=
374         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
375   }
376   if (MFI.hasDispatchID()) {
377     KernelCodeProperties |=
378         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
379   }
380   if (MFI.hasFlatScratchInit()) {
381     KernelCodeProperties |=
382         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
383   }
384   if (MF.getSubtarget<GCNSubtarget>().isWave32()) {
385     KernelCodeProperties |=
386         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32;
387   }
388 
389   return KernelCodeProperties;
390 }
391 
392 amdhsa::kernel_descriptor_t AMDGPUAsmPrinter::getAmdhsaKernelDescriptor(
393     const MachineFunction &MF,
394     const SIProgramInfo &PI) const {
395   amdhsa::kernel_descriptor_t KernelDescriptor;
396   memset(&KernelDescriptor, 0x0, sizeof(KernelDescriptor));
397 
398   assert(isUInt<32>(PI.ScratchSize));
399   assert(isUInt<32>(PI.ComputePGMRSrc1));
400   assert(isUInt<32>(PI.ComputePGMRSrc2));
401 
402   KernelDescriptor.group_segment_fixed_size = PI.LDSSize;
403   KernelDescriptor.private_segment_fixed_size = PI.ScratchSize;
404   KernelDescriptor.compute_pgm_rsrc1 = PI.ComputePGMRSrc1;
405   KernelDescriptor.compute_pgm_rsrc2 = PI.ComputePGMRSrc2;
406   KernelDescriptor.kernel_code_properties = getAmdhsaKernelCodeProperties(MF);
407 
408   return KernelDescriptor;
409 }
410 
411 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
412   CurrentProgramInfo = SIProgramInfo();
413 
414   const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
415 
416   // The starting address of all shader programs must be 256 bytes aligned.
417   // Regular functions just need the basic required instruction alignment.
418   MF.setAlignment(MFI->isEntryFunction() ? Align(256) : Align(4));
419 
420   SetupMachineFunction(MF);
421 
422   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
423   MCContext &Context = getObjFileLowering().getContext();
424   // FIXME: This should be an explicit check for Mesa.
425   if (!STM.isAmdHsaOS() && !STM.isAmdPalOS()) {
426     MCSectionELF *ConfigSection =
427         Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);
428     OutStreamer->SwitchSection(ConfigSection);
429   }
430 
431   if (MFI->isEntryFunction()) {
432     getSIProgramInfo(CurrentProgramInfo, MF);
433   } else {
434     auto I = CallGraphResourceInfo.insert(
435       std::make_pair(&MF.getFunction(), SIFunctionResourceInfo()));
436     SIFunctionResourceInfo &Info = I.first->second;
437     assert(I.second && "should only be called once per function");
438     Info = analyzeResourceUsage(MF);
439   }
440 
441   if (STM.isAmdPalOS())
442     EmitPALMetadata(MF, CurrentProgramInfo);
443   else if (!STM.isAmdHsaOS()) {
444     EmitProgramInfoSI(MF, CurrentProgramInfo);
445   }
446 
447   DumpCodeInstEmitter = nullptr;
448   if (STM.dumpCode()) {
449     // For -dumpcode, get the assembler out of the streamer, even if it does
450     // not really want to let us have it. This only works with -filetype=obj.
451     bool SaveFlag = OutStreamer->getUseAssemblerInfoForParsing();
452     OutStreamer->setUseAssemblerInfoForParsing(true);
453     MCAssembler *Assembler = OutStreamer->getAssemblerPtr();
454     OutStreamer->setUseAssemblerInfoForParsing(SaveFlag);
455     if (Assembler)
456       DumpCodeInstEmitter = Assembler->getEmitterPtr();
457   }
458 
459   DisasmLines.clear();
460   HexLines.clear();
461   DisasmLineMaxLen = 0;
462 
463   emitFunctionBody();
464 
465   if (isVerbose()) {
466     MCSectionELF *CommentSection =
467         Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0);
468     OutStreamer->SwitchSection(CommentSection);
469 
470     if (!MFI->isEntryFunction()) {
471       OutStreamer->emitRawComment(" Function info:", false);
472       SIFunctionResourceInfo &Info = CallGraphResourceInfo[&MF.getFunction()];
473       emitCommonFunctionComments(
474         Info.NumVGPR,
475         STM.hasMAIInsts() ? Info.NumAGPR : Optional<uint32_t>(),
476         Info.getTotalNumVGPRs(STM),
477         Info.getTotalNumSGPRs(MF.getSubtarget<GCNSubtarget>()),
478         Info.PrivateSegmentSize,
479         getFunctionCodeSize(MF), MFI);
480       return false;
481     }
482 
483     OutStreamer->emitRawComment(" Kernel info:", false);
484     emitCommonFunctionComments(CurrentProgramInfo.NumArchVGPR,
485                                STM.hasMAIInsts()
486                                  ? CurrentProgramInfo.NumAccVGPR
487                                  : Optional<uint32_t>(),
488                                CurrentProgramInfo.NumVGPR,
489                                CurrentProgramInfo.NumSGPR,
490                                CurrentProgramInfo.ScratchSize,
491                                getFunctionCodeSize(MF), MFI);
492 
493     OutStreamer->emitRawComment(
494       " FloatMode: " + Twine(CurrentProgramInfo.FloatMode), false);
495     OutStreamer->emitRawComment(
496       " IeeeMode: " + Twine(CurrentProgramInfo.IEEEMode), false);
497     OutStreamer->emitRawComment(
498       " LDSByteSize: " + Twine(CurrentProgramInfo.LDSSize) +
499       " bytes/workgroup (compile time only)", false);
500 
501     OutStreamer->emitRawComment(
502       " SGPRBlocks: " + Twine(CurrentProgramInfo.SGPRBlocks), false);
503     OutStreamer->emitRawComment(
504       " VGPRBlocks: " + Twine(CurrentProgramInfo.VGPRBlocks), false);
505 
506     OutStreamer->emitRawComment(
507       " NumSGPRsForWavesPerEU: " +
508       Twine(CurrentProgramInfo.NumSGPRsForWavesPerEU), false);
509     OutStreamer->emitRawComment(
510       " NumVGPRsForWavesPerEU: " +
511       Twine(CurrentProgramInfo.NumVGPRsForWavesPerEU), false);
512 
513     OutStreamer->emitRawComment(
514       " Occupancy: " +
515       Twine(CurrentProgramInfo.Occupancy), false);
516 
517     OutStreamer->emitRawComment(
518       " WaveLimiterHint : " + Twine(MFI->needsWaveLimiter()), false);
519 
520     OutStreamer->emitRawComment(
521       " COMPUTE_PGM_RSRC2:USER_SGPR: " +
522       Twine(G_00B84C_USER_SGPR(CurrentProgramInfo.ComputePGMRSrc2)), false);
523     OutStreamer->emitRawComment(
524       " COMPUTE_PGM_RSRC2:TRAP_HANDLER: " +
525       Twine(G_00B84C_TRAP_HANDLER(CurrentProgramInfo.ComputePGMRSrc2)), false);
526     OutStreamer->emitRawComment(
527       " COMPUTE_PGM_RSRC2:TGID_X_EN: " +
528       Twine(G_00B84C_TGID_X_EN(CurrentProgramInfo.ComputePGMRSrc2)), false);
529     OutStreamer->emitRawComment(
530       " COMPUTE_PGM_RSRC2:TGID_Y_EN: " +
531       Twine(G_00B84C_TGID_Y_EN(CurrentProgramInfo.ComputePGMRSrc2)), false);
532     OutStreamer->emitRawComment(
533       " COMPUTE_PGM_RSRC2:TGID_Z_EN: " +
534       Twine(G_00B84C_TGID_Z_EN(CurrentProgramInfo.ComputePGMRSrc2)), false);
535     OutStreamer->emitRawComment(
536       " COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " +
537       Twine(G_00B84C_TIDIG_COMP_CNT(CurrentProgramInfo.ComputePGMRSrc2)),
538       false);
539   }
540 
541   if (DumpCodeInstEmitter) {
542 
543     OutStreamer->SwitchSection(
544         Context.getELFSection(".AMDGPU.disasm", ELF::SHT_PROGBITS, 0));
545 
546     for (size_t i = 0; i < DisasmLines.size(); ++i) {
547       std::string Comment = "\n";
548       if (!HexLines[i].empty()) {
549         Comment = std::string(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
550         Comment += " ; " + HexLines[i] + "\n";
551       }
552 
553       OutStreamer->emitBytes(StringRef(DisasmLines[i]));
554       OutStreamer->emitBytes(StringRef(Comment));
555     }
556   }
557 
558   return false;
559 }
560 
561 uint64_t AMDGPUAsmPrinter::getFunctionCodeSize(const MachineFunction &MF) const {
562   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
563   const SIInstrInfo *TII = STM.getInstrInfo();
564 
565   uint64_t CodeSize = 0;
566 
567   for (const MachineBasicBlock &MBB : MF) {
568     for (const MachineInstr &MI : MBB) {
569       // TODO: CodeSize should account for multiple functions.
570 
571       // TODO: Should we count size of debug info?
572       if (MI.isDebugInstr())
573         continue;
574 
575       CodeSize += TII->getInstSizeInBytes(MI);
576     }
577   }
578 
579   return CodeSize;
580 }
581 
582 static bool hasAnyNonFlatUseOfReg(const MachineRegisterInfo &MRI,
583                                   const SIInstrInfo &TII,
584                                   unsigned Reg) {
585   for (const MachineOperand &UseOp : MRI.reg_operands(Reg)) {
586     if (!UseOp.isImplicit() || !TII.isFLAT(*UseOp.getParent()))
587       return true;
588   }
589 
590   return false;
591 }
592 
593 int32_t AMDGPUAsmPrinter::SIFunctionResourceInfo::getTotalNumSGPRs(
594   const GCNSubtarget &ST) const {
595   return NumExplicitSGPR + IsaInfo::getNumExtraSGPRs(&ST,
596                                                      UsesVCC, UsesFlatScratch);
597 }
598 
599 int32_t AMDGPUAsmPrinter::SIFunctionResourceInfo::getTotalNumVGPRs(
600   const GCNSubtarget &ST) const {
601   return std::max(NumVGPR, NumAGPR);
602 }
603 
604 static const Function *getCalleeFunction(const MachineOperand &Op) {
605   if (Op.isImm()) {
606     assert(Op.getImm() == 0);
607     return nullptr;
608   }
609 
610   return cast<Function>(Op.getGlobal());
611 }
612 
613 AMDGPUAsmPrinter::SIFunctionResourceInfo AMDGPUAsmPrinter::analyzeResourceUsage(
614   const MachineFunction &MF) const {
615   SIFunctionResourceInfo Info;
616 
617   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
618   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
619   const MachineFrameInfo &FrameInfo = MF.getFrameInfo();
620   const MachineRegisterInfo &MRI = MF.getRegInfo();
621   const SIInstrInfo *TII = ST.getInstrInfo();
622   const SIRegisterInfo &TRI = TII->getRegisterInfo();
623 
624   Info.UsesFlatScratch = MRI.isPhysRegUsed(AMDGPU::FLAT_SCR_LO) ||
625                          MRI.isPhysRegUsed(AMDGPU::FLAT_SCR_HI);
626 
627   // Even if FLAT_SCRATCH is implicitly used, it has no effect if flat
628   // instructions aren't used to access the scratch buffer. Inline assembly may
629   // need it though.
630   //
631   // If we only have implicit uses of flat_scr on flat instructions, it is not
632   // really needed.
633   if (Info.UsesFlatScratch && !MFI->hasFlatScratchInit() &&
634       (!hasAnyNonFlatUseOfReg(MRI, *TII, AMDGPU::FLAT_SCR) &&
635        !hasAnyNonFlatUseOfReg(MRI, *TII, AMDGPU::FLAT_SCR_LO) &&
636        !hasAnyNonFlatUseOfReg(MRI, *TII, AMDGPU::FLAT_SCR_HI))) {
637     Info.UsesFlatScratch = false;
638   }
639 
640   Info.HasDynamicallySizedStack = FrameInfo.hasVarSizedObjects();
641   Info.PrivateSegmentSize = FrameInfo.getStackSize();
642   if (MFI->isStackRealigned())
643     Info.PrivateSegmentSize += FrameInfo.getMaxAlign().value();
644 
645   Info.UsesVCC = MRI.isPhysRegUsed(AMDGPU::VCC_LO) ||
646                  MRI.isPhysRegUsed(AMDGPU::VCC_HI);
647 
648   // If there are no calls, MachineRegisterInfo can tell us the used register
649   // count easily.
650   // A tail call isn't considered a call for MachineFrameInfo's purposes.
651   if (!FrameInfo.hasCalls() && !FrameInfo.hasTailCall()) {
652     MCPhysReg HighestVGPRReg = AMDGPU::NoRegister;
653     for (MCPhysReg Reg : reverse(AMDGPU::VGPR_32RegClass.getRegisters())) {
654       if (MRI.isPhysRegUsed(Reg)) {
655         HighestVGPRReg = Reg;
656         break;
657       }
658     }
659 
660     if (ST.hasMAIInsts()) {
661       MCPhysReg HighestAGPRReg = AMDGPU::NoRegister;
662       for (MCPhysReg Reg : reverse(AMDGPU::AGPR_32RegClass.getRegisters())) {
663         if (MRI.isPhysRegUsed(Reg)) {
664           HighestAGPRReg = Reg;
665           break;
666         }
667       }
668       Info.NumAGPR = HighestAGPRReg == AMDGPU::NoRegister ? 0 :
669         TRI.getHWRegIndex(HighestAGPRReg) + 1;
670     }
671 
672     MCPhysReg HighestSGPRReg = AMDGPU::NoRegister;
673     for (MCPhysReg Reg : reverse(AMDGPU::SGPR_32RegClass.getRegisters())) {
674       if (MRI.isPhysRegUsed(Reg)) {
675         HighestSGPRReg = Reg;
676         break;
677       }
678     }
679 
680     // We found the maximum register index. They start at 0, so add one to get the
681     // number of registers.
682     Info.NumVGPR = HighestVGPRReg == AMDGPU::NoRegister ? 0 :
683       TRI.getHWRegIndex(HighestVGPRReg) + 1;
684     Info.NumExplicitSGPR = HighestSGPRReg == AMDGPU::NoRegister ? 0 :
685       TRI.getHWRegIndex(HighestSGPRReg) + 1;
686 
687     return Info;
688   }
689 
690   int32_t MaxVGPR = -1;
691   int32_t MaxAGPR = -1;
692   int32_t MaxSGPR = -1;
693   uint64_t CalleeFrameSize = 0;
694 
695   for (const MachineBasicBlock &MBB : MF) {
696     for (const MachineInstr &MI : MBB) {
697       // TODO: Check regmasks? Do they occur anywhere except calls?
698       for (const MachineOperand &MO : MI.operands()) {
699         unsigned Width = 0;
700         bool IsSGPR = false;
701         bool IsAGPR = false;
702 
703         if (!MO.isReg())
704           continue;
705 
706         Register Reg = MO.getReg();
707         switch (Reg) {
708         case AMDGPU::EXEC:
709         case AMDGPU::EXEC_LO:
710         case AMDGPU::EXEC_HI:
711         case AMDGPU::SCC:
712         case AMDGPU::M0:
713         case AMDGPU::SRC_SHARED_BASE:
714         case AMDGPU::SRC_SHARED_LIMIT:
715         case AMDGPU::SRC_PRIVATE_BASE:
716         case AMDGPU::SRC_PRIVATE_LIMIT:
717         case AMDGPU::SGPR_NULL:
718           continue;
719 
720         case AMDGPU::SRC_POPS_EXITING_WAVE_ID:
721           llvm_unreachable("src_pops_exiting_wave_id should not be used");
722 
723         case AMDGPU::NoRegister:
724           assert(MI.isDebugInstr());
725           continue;
726 
727         case AMDGPU::VCC:
728         case AMDGPU::VCC_LO:
729         case AMDGPU::VCC_HI:
730         case AMDGPU::VCC_LO_LO16:
731         case AMDGPU::VCC_LO_HI16:
732         case AMDGPU::VCC_HI_LO16:
733         case AMDGPU::VCC_HI_HI16:
734           Info.UsesVCC = true;
735           continue;
736 
737         case AMDGPU::FLAT_SCR:
738         case AMDGPU::FLAT_SCR_LO:
739         case AMDGPU::FLAT_SCR_HI:
740           continue;
741 
742         case AMDGPU::XNACK_MASK:
743         case AMDGPU::XNACK_MASK_LO:
744         case AMDGPU::XNACK_MASK_HI:
745           llvm_unreachable("xnack_mask registers should not be used");
746 
747         case AMDGPU::LDS_DIRECT:
748           llvm_unreachable("lds_direct register should not be used");
749 
750         case AMDGPU::TBA:
751         case AMDGPU::TBA_LO:
752         case AMDGPU::TBA_HI:
753         case AMDGPU::TMA:
754         case AMDGPU::TMA_LO:
755         case AMDGPU::TMA_HI:
756           llvm_unreachable("trap handler registers should not be used");
757 
758         case AMDGPU::SRC_VCCZ:
759           llvm_unreachable("src_vccz register should not be used");
760 
761         case AMDGPU::SRC_EXECZ:
762           llvm_unreachable("src_execz register should not be used");
763 
764         case AMDGPU::SRC_SCC:
765           llvm_unreachable("src_scc register should not be used");
766 
767         default:
768           break;
769         }
770 
771         if (AMDGPU::SReg_32RegClass.contains(Reg) ||
772             AMDGPU::SReg_LO16RegClass.contains(Reg) ||
773             AMDGPU::SGPR_HI16RegClass.contains(Reg)) {
774           assert(!AMDGPU::TTMP_32RegClass.contains(Reg) &&
775                  "trap handler registers should not be used");
776           IsSGPR = true;
777           Width = 1;
778         } else if (AMDGPU::VGPR_32RegClass.contains(Reg) ||
779                    AMDGPU::VGPR_LO16RegClass.contains(Reg) ||
780                    AMDGPU::VGPR_HI16RegClass.contains(Reg)) {
781           IsSGPR = false;
782           Width = 1;
783         } else if (AMDGPU::AGPR_32RegClass.contains(Reg) ||
784                    AMDGPU::AGPR_LO16RegClass.contains(Reg)) {
785           IsSGPR = false;
786           IsAGPR = true;
787           Width = 1;
788         } else if (AMDGPU::SReg_64RegClass.contains(Reg)) {
789           assert(!AMDGPU::TTMP_64RegClass.contains(Reg) &&
790                  "trap handler registers should not be used");
791           IsSGPR = true;
792           Width = 2;
793         } else if (AMDGPU::VReg_64RegClass.contains(Reg)) {
794           IsSGPR = false;
795           Width = 2;
796         } else if (AMDGPU::AReg_64RegClass.contains(Reg)) {
797           IsSGPR = false;
798           IsAGPR = true;
799           Width = 2;
800         } else if (AMDGPU::VReg_96RegClass.contains(Reg)) {
801           IsSGPR = false;
802           Width = 3;
803         } else if (AMDGPU::SReg_96RegClass.contains(Reg)) {
804           IsSGPR = true;
805           Width = 3;
806         } else if (AMDGPU::AReg_96RegClass.contains(Reg)) {
807           IsSGPR = false;
808           IsAGPR = true;
809           Width = 3;
810         } else if (AMDGPU::SReg_128RegClass.contains(Reg)) {
811           assert(!AMDGPU::TTMP_128RegClass.contains(Reg) &&
812             "trap handler registers should not be used");
813           IsSGPR = true;
814           Width = 4;
815         } else if (AMDGPU::VReg_128RegClass.contains(Reg)) {
816           IsSGPR = false;
817           Width = 4;
818         } else if (AMDGPU::AReg_128RegClass.contains(Reg)) {
819           IsSGPR = false;
820           IsAGPR = true;
821           Width = 4;
822         } else if (AMDGPU::VReg_160RegClass.contains(Reg)) {
823           IsSGPR = false;
824           Width = 5;
825         } else if (AMDGPU::SReg_160RegClass.contains(Reg)) {
826           IsSGPR = true;
827           Width = 5;
828         } else if (AMDGPU::AReg_160RegClass.contains(Reg)) {
829           IsSGPR = false;
830           IsAGPR = true;
831           Width = 5;
832         } else if (AMDGPU::VReg_192RegClass.contains(Reg)) {
833           IsSGPR = false;
834           Width = 6;
835         } else if (AMDGPU::SReg_192RegClass.contains(Reg)) {
836           IsSGPR = true;
837           Width = 6;
838         } else if (AMDGPU::AReg_192RegClass.contains(Reg)) {
839           IsSGPR = false;
840           IsAGPR = true;
841           Width = 6;
842         } else if (AMDGPU::SReg_256RegClass.contains(Reg)) {
843           assert(!AMDGPU::TTMP_256RegClass.contains(Reg) &&
844             "trap handler registers should not be used");
845           IsSGPR = true;
846           Width = 8;
847         } else if (AMDGPU::VReg_256RegClass.contains(Reg)) {
848           IsSGPR = false;
849           Width = 8;
850         } else if (AMDGPU::AReg_256RegClass.contains(Reg)) {
851           IsSGPR = false;
852           IsAGPR = true;
853           Width = 8;
854         } else if (AMDGPU::SReg_512RegClass.contains(Reg)) {
855           assert(!AMDGPU::TTMP_512RegClass.contains(Reg) &&
856             "trap handler registers should not be used");
857           IsSGPR = true;
858           Width = 16;
859         } else if (AMDGPU::VReg_512RegClass.contains(Reg)) {
860           IsSGPR = false;
861           Width = 16;
862         } else if (AMDGPU::AReg_512RegClass.contains(Reg)) {
863           IsSGPR = false;
864           IsAGPR = true;
865           Width = 16;
866         } else if (AMDGPU::SReg_1024RegClass.contains(Reg)) {
867           IsSGPR = true;
868           Width = 32;
869         } else if (AMDGPU::VReg_1024RegClass.contains(Reg)) {
870           IsSGPR = false;
871           Width = 32;
872         } else if (AMDGPU::AReg_1024RegClass.contains(Reg)) {
873           IsSGPR = false;
874           IsAGPR = true;
875           Width = 32;
876         } else {
877           llvm_unreachable("Unknown register class");
878         }
879         unsigned HWReg = TRI.getHWRegIndex(Reg);
880         int MaxUsed = HWReg + Width - 1;
881         if (IsSGPR) {
882           MaxSGPR = MaxUsed > MaxSGPR ? MaxUsed : MaxSGPR;
883         } else if (IsAGPR) {
884           MaxAGPR = MaxUsed > MaxAGPR ? MaxUsed : MaxAGPR;
885         } else {
886           MaxVGPR = MaxUsed > MaxVGPR ? MaxUsed : MaxVGPR;
887         }
888       }
889 
890       if (MI.isCall()) {
891         // Pseudo used just to encode the underlying global. Is there a better
892         // way to track this?
893 
894         const MachineOperand *CalleeOp
895           = TII->getNamedOperand(MI, AMDGPU::OpName::callee);
896 
897         const Function *Callee = getCalleeFunction(*CalleeOp);
898         if (!Callee || Callee->isDeclaration()) {
899           // If this is a call to an external function, we can't do much. Make
900           // conservative guesses.
901 
902           // 48 SGPRs - vcc, - flat_scr, -xnack
903           int MaxSGPRGuess =
904             47 - IsaInfo::getNumExtraSGPRs(&ST, true, ST.hasFlatAddressSpace());
905           MaxSGPR = std::max(MaxSGPR, MaxSGPRGuess);
906           MaxVGPR = std::max(MaxVGPR, 23);
907           MaxAGPR = std::max(MaxAGPR, 23);
908 
909           CalleeFrameSize = std::max(CalleeFrameSize, UINT64_C(16384));
910           Info.UsesVCC = true;
911           Info.UsesFlatScratch = ST.hasFlatAddressSpace();
912           Info.HasDynamicallySizedStack = true;
913         } else {
914           // We force CodeGen to run in SCC order, so the callee's register
915           // usage etc. should be the cumulative usage of all callees.
916 
917           auto I = CallGraphResourceInfo.find(Callee);
918           if (I == CallGraphResourceInfo.end()) {
919             // Avoid crashing on undefined behavior with an illegal call to a
920             // kernel. If a callsite's calling convention doesn't match the
921             // function's, it's undefined behavior. If the callsite calling
922             // convention does match, that would have errored earlier.
923             // FIXME: The verifier shouldn't allow this.
924             if (AMDGPU::isEntryFunctionCC(Callee->getCallingConv()))
925               report_fatal_error("invalid call to entry function");
926 
927             llvm_unreachable("callee should have been handled before caller");
928           }
929 
930           MaxSGPR = std::max(I->second.NumExplicitSGPR - 1, MaxSGPR);
931           MaxVGPR = std::max(I->second.NumVGPR - 1, MaxVGPR);
932           MaxAGPR = std::max(I->second.NumAGPR - 1, MaxAGPR);
933           CalleeFrameSize
934             = std::max(I->second.PrivateSegmentSize, CalleeFrameSize);
935           Info.UsesVCC |= I->second.UsesVCC;
936           Info.UsesFlatScratch |= I->second.UsesFlatScratch;
937           Info.HasDynamicallySizedStack |= I->second.HasDynamicallySizedStack;
938           Info.HasRecursion |= I->second.HasRecursion;
939         }
940 
941         // FIXME: Call site could have norecurse on it
942         if (!Callee || !Callee->doesNotRecurse())
943           Info.HasRecursion = true;
944       }
945     }
946   }
947 
948   Info.NumExplicitSGPR = MaxSGPR + 1;
949   Info.NumVGPR = MaxVGPR + 1;
950   Info.NumAGPR = MaxAGPR + 1;
951   Info.PrivateSegmentSize += CalleeFrameSize;
952 
953   return Info;
954 }
955 
956 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
957                                         const MachineFunction &MF) {
958   SIFunctionResourceInfo Info = analyzeResourceUsage(MF);
959   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
960 
961   ProgInfo.NumArchVGPR = Info.NumVGPR;
962   ProgInfo.NumAccVGPR = Info.NumAGPR;
963   ProgInfo.NumVGPR = Info.getTotalNumVGPRs(STM);
964   ProgInfo.NumSGPR = Info.NumExplicitSGPR;
965   ProgInfo.ScratchSize = Info.PrivateSegmentSize;
966   ProgInfo.VCCUsed = Info.UsesVCC;
967   ProgInfo.FlatUsed = Info.UsesFlatScratch;
968   ProgInfo.DynamicCallStack = Info.HasDynamicallySizedStack || Info.HasRecursion;
969 
970   if (!isUInt<32>(ProgInfo.ScratchSize)) {
971     DiagnosticInfoStackSize DiagStackSize(MF.getFunction(),
972                                           ProgInfo.ScratchSize, DS_Error);
973     MF.getFunction().getContext().diagnose(DiagStackSize);
974   }
975 
976   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
977 
978   // TODO(scott.linder): The calculations related to SGPR/VGPR blocks are
979   // duplicated in part in AMDGPUAsmParser::calculateGPRBlocks, and could be
980   // unified.
981   unsigned ExtraSGPRs = IsaInfo::getNumExtraSGPRs(
982       &STM, ProgInfo.VCCUsed, ProgInfo.FlatUsed);
983 
984   // Check the addressable register limit before we add ExtraSGPRs.
985   if (STM.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS &&
986       !STM.hasSGPRInitBug()) {
987     unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs();
988     if (ProgInfo.NumSGPR > MaxAddressableNumSGPRs) {
989       // This can happen due to a compiler bug or when using inline asm.
990       LLVMContext &Ctx = MF.getFunction().getContext();
991       DiagnosticInfoResourceLimit Diag(MF.getFunction(),
992                                        "addressable scalar registers",
993                                        ProgInfo.NumSGPR, DS_Error,
994                                        DK_ResourceLimit,
995                                        MaxAddressableNumSGPRs);
996       Ctx.diagnose(Diag);
997       ProgInfo.NumSGPR = MaxAddressableNumSGPRs - 1;
998     }
999   }
1000 
1001   // Account for extra SGPRs and VGPRs reserved for debugger use.
1002   ProgInfo.NumSGPR += ExtraSGPRs;
1003 
1004   // Ensure there are enough SGPRs and VGPRs for wave dispatch, where wave
1005   // dispatch registers are function args.
1006   unsigned WaveDispatchNumSGPR = 0, WaveDispatchNumVGPR = 0;
1007   for (auto &Arg : MF.getFunction().args()) {
1008     unsigned NumRegs = (Arg.getType()->getPrimitiveSizeInBits() + 31) / 32;
1009     if (Arg.hasAttribute(Attribute::InReg))
1010       WaveDispatchNumSGPR += NumRegs;
1011     else
1012       WaveDispatchNumVGPR += NumRegs;
1013   }
1014   ProgInfo.NumSGPR = std::max(ProgInfo.NumSGPR, WaveDispatchNumSGPR);
1015   ProgInfo.NumVGPR = std::max(ProgInfo.NumVGPR, WaveDispatchNumVGPR);
1016 
1017   // Adjust number of registers used to meet default/requested minimum/maximum
1018   // number of waves per execution unit request.
1019   ProgInfo.NumSGPRsForWavesPerEU = std::max(
1020     std::max(ProgInfo.NumSGPR, 1u), STM.getMinNumSGPRs(MFI->getMaxWavesPerEU()));
1021   ProgInfo.NumVGPRsForWavesPerEU = std::max(
1022     std::max(ProgInfo.NumVGPR, 1u), STM.getMinNumVGPRs(MFI->getMaxWavesPerEU()));
1023 
1024   if (STM.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS ||
1025       STM.hasSGPRInitBug()) {
1026     unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs();
1027     if (ProgInfo.NumSGPR > MaxAddressableNumSGPRs) {
1028       // This can happen due to a compiler bug or when using inline asm to use
1029       // the registers which are usually reserved for vcc etc.
1030       LLVMContext &Ctx = MF.getFunction().getContext();
1031       DiagnosticInfoResourceLimit Diag(MF.getFunction(),
1032                                        "scalar registers",
1033                                        ProgInfo.NumSGPR, DS_Error,
1034                                        DK_ResourceLimit,
1035                                        MaxAddressableNumSGPRs);
1036       Ctx.diagnose(Diag);
1037       ProgInfo.NumSGPR = MaxAddressableNumSGPRs;
1038       ProgInfo.NumSGPRsForWavesPerEU = MaxAddressableNumSGPRs;
1039     }
1040   }
1041 
1042   if (STM.hasSGPRInitBug()) {
1043     ProgInfo.NumSGPR =
1044         AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG;
1045     ProgInfo.NumSGPRsForWavesPerEU =
1046         AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG;
1047   }
1048 
1049   if (MFI->getNumUserSGPRs() > STM.getMaxNumUserSGPRs()) {
1050     LLVMContext &Ctx = MF.getFunction().getContext();
1051     DiagnosticInfoResourceLimit Diag(MF.getFunction(), "user SGPRs",
1052                                      MFI->getNumUserSGPRs(), DS_Error);
1053     Ctx.diagnose(Diag);
1054   }
1055 
1056   if (MFI->getLDSSize() > static_cast<unsigned>(STM.getLocalMemorySize())) {
1057     LLVMContext &Ctx = MF.getFunction().getContext();
1058     DiagnosticInfoResourceLimit Diag(MF.getFunction(), "local memory",
1059                                      MFI->getLDSSize(), DS_Error);
1060     Ctx.diagnose(Diag);
1061   }
1062 
1063   ProgInfo.SGPRBlocks = IsaInfo::getNumSGPRBlocks(
1064       &STM, ProgInfo.NumSGPRsForWavesPerEU);
1065   ProgInfo.VGPRBlocks = IsaInfo::getNumVGPRBlocks(
1066       &STM, ProgInfo.NumVGPRsForWavesPerEU);
1067 
1068   const SIModeRegisterDefaults Mode = MFI->getMode();
1069 
1070   // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
1071   // register.
1072   ProgInfo.FloatMode = getFPMode(Mode);
1073 
1074   ProgInfo.IEEEMode = Mode.IEEE;
1075 
1076   // Make clamp modifier on NaN input returns 0.
1077   ProgInfo.DX10Clamp = Mode.DX10Clamp;
1078 
1079   unsigned LDSAlignShift;
1080   if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
1081     // LDS is allocated in 64 dword blocks.
1082     LDSAlignShift = 8;
1083   } else {
1084     // LDS is allocated in 128 dword blocks.
1085     LDSAlignShift = 9;
1086   }
1087 
1088   unsigned LDSSpillSize =
1089     MFI->getLDSWaveSpillSize() * MFI->getMaxFlatWorkGroupSize();
1090 
1091   ProgInfo.LDSSize = MFI->getLDSSize() + LDSSpillSize;
1092   ProgInfo.LDSBlocks =
1093       alignTo(ProgInfo.LDSSize, 1ULL << LDSAlignShift) >> LDSAlignShift;
1094 
1095   // Scratch is allocated in 256 dword blocks.
1096   unsigned ScratchAlignShift = 10;
1097   // We need to program the hardware with the amount of scratch memory that
1098   // is used by the entire wave.  ProgInfo.ScratchSize is the amount of
1099   // scratch memory used per thread.
1100   ProgInfo.ScratchBlocks =
1101       alignTo(ProgInfo.ScratchSize * STM.getWavefrontSize(),
1102               1ULL << ScratchAlignShift) >>
1103       ScratchAlignShift;
1104 
1105   if (getIsaVersion(getGlobalSTI()->getCPU()).Major >= 10) {
1106     ProgInfo.WgpMode = STM.isCuModeEnabled() ? 0 : 1;
1107     ProgInfo.MemOrdered = 1;
1108   }
1109 
1110   ProgInfo.ComputePGMRSrc1 =
1111       S_00B848_VGPRS(ProgInfo.VGPRBlocks) |
1112       S_00B848_SGPRS(ProgInfo.SGPRBlocks) |
1113       S_00B848_PRIORITY(ProgInfo.Priority) |
1114       S_00B848_FLOAT_MODE(ProgInfo.FloatMode) |
1115       S_00B848_PRIV(ProgInfo.Priv) |
1116       S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) |
1117       S_00B848_DEBUG_MODE(ProgInfo.DebugMode) |
1118       S_00B848_IEEE_MODE(ProgInfo.IEEEMode) |
1119       S_00B848_WGP_MODE(ProgInfo.WgpMode) |
1120       S_00B848_MEM_ORDERED(ProgInfo.MemOrdered);
1121 
1122   // 0 = X, 1 = XY, 2 = XYZ
1123   unsigned TIDIGCompCnt = 0;
1124   if (MFI->hasWorkItemIDZ())
1125     TIDIGCompCnt = 2;
1126   else if (MFI->hasWorkItemIDY())
1127     TIDIGCompCnt = 1;
1128 
1129   ProgInfo.ComputePGMRSrc2 =
1130       S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
1131       S_00B84C_USER_SGPR(MFI->getNumUserSGPRs()) |
1132       // For AMDHSA, TRAP_HANDLER must be zero, as it is populated by the CP.
1133       S_00B84C_TRAP_HANDLER(STM.isAmdHsaOS() ? 0 : STM.isTrapHandlerEnabled()) |
1134       S_00B84C_TGID_X_EN(MFI->hasWorkGroupIDX()) |
1135       S_00B84C_TGID_Y_EN(MFI->hasWorkGroupIDY()) |
1136       S_00B84C_TGID_Z_EN(MFI->hasWorkGroupIDZ()) |
1137       S_00B84C_TG_SIZE_EN(MFI->hasWorkGroupInfo()) |
1138       S_00B84C_TIDIG_COMP_CNT(TIDIGCompCnt) |
1139       S_00B84C_EXCP_EN_MSB(0) |
1140       // For AMDHSA, LDS_SIZE must be zero, as it is populated by the CP.
1141       S_00B84C_LDS_SIZE(STM.isAmdHsaOS() ? 0 : ProgInfo.LDSBlocks) |
1142       S_00B84C_EXCP_EN(0);
1143 
1144   ProgInfo.Occupancy = STM.computeOccupancy(MF, ProgInfo.LDSSize,
1145                                             ProgInfo.NumSGPRsForWavesPerEU,
1146                                             ProgInfo.NumVGPRsForWavesPerEU);
1147 }
1148 
1149 static unsigned getRsrcReg(CallingConv::ID CallConv) {
1150   switch (CallConv) {
1151   default: LLVM_FALLTHROUGH;
1152   case CallingConv::AMDGPU_CS: return R_00B848_COMPUTE_PGM_RSRC1;
1153   case CallingConv::AMDGPU_LS: return R_00B528_SPI_SHADER_PGM_RSRC1_LS;
1154   case CallingConv::AMDGPU_HS: return R_00B428_SPI_SHADER_PGM_RSRC1_HS;
1155   case CallingConv::AMDGPU_ES: return R_00B328_SPI_SHADER_PGM_RSRC1_ES;
1156   case CallingConv::AMDGPU_GS: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
1157   case CallingConv::AMDGPU_VS: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
1158   case CallingConv::AMDGPU_PS: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
1159   }
1160 }
1161 
1162 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
1163                                          const SIProgramInfo &CurrentProgramInfo) {
1164   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1165   unsigned RsrcReg = getRsrcReg(MF.getFunction().getCallingConv());
1166 
1167   if (AMDGPU::isCompute(MF.getFunction().getCallingConv())) {
1168     OutStreamer->emitInt32(R_00B848_COMPUTE_PGM_RSRC1);
1169 
1170     OutStreamer->emitInt32(CurrentProgramInfo.ComputePGMRSrc1);
1171 
1172     OutStreamer->emitInt32(R_00B84C_COMPUTE_PGM_RSRC2);
1173     OutStreamer->emitInt32(CurrentProgramInfo.ComputePGMRSrc2);
1174 
1175     OutStreamer->emitInt32(R_00B860_COMPUTE_TMPRING_SIZE);
1176     OutStreamer->emitInt32(S_00B860_WAVESIZE(CurrentProgramInfo.ScratchBlocks));
1177 
1178     // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
1179     // 0" comment but I don't see a corresponding field in the register spec.
1180   } else {
1181     OutStreamer->emitInt32(RsrcReg);
1182     OutStreamer->emitIntValue(S_00B028_VGPRS(CurrentProgramInfo.VGPRBlocks) |
1183                               S_00B028_SGPRS(CurrentProgramInfo.SGPRBlocks), 4);
1184     OutStreamer->emitInt32(R_0286E8_SPI_TMPRING_SIZE);
1185     OutStreamer->emitIntValue(
1186         S_0286E8_WAVESIZE(CurrentProgramInfo.ScratchBlocks), 4);
1187   }
1188 
1189   if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS) {
1190     OutStreamer->emitInt32(R_00B02C_SPI_SHADER_PGM_RSRC2_PS);
1191     OutStreamer->emitInt32(
1192         S_00B02C_EXTRA_LDS_SIZE(CurrentProgramInfo.LDSBlocks));
1193     OutStreamer->emitInt32(R_0286CC_SPI_PS_INPUT_ENA);
1194     OutStreamer->emitInt32(MFI->getPSInputEnable());
1195     OutStreamer->emitInt32(R_0286D0_SPI_PS_INPUT_ADDR);
1196     OutStreamer->emitInt32(MFI->getPSInputAddr());
1197   }
1198 
1199   OutStreamer->emitInt32(R_SPILLED_SGPRS);
1200   OutStreamer->emitInt32(MFI->getNumSpilledSGPRs());
1201   OutStreamer->emitInt32(R_SPILLED_VGPRS);
1202   OutStreamer->emitInt32(MFI->getNumSpilledVGPRs());
1203 }
1204 
1205 // This is the equivalent of EmitProgramInfoSI above, but for when the OS type
1206 // is AMDPAL.  It stores each compute/SPI register setting and other PAL
1207 // metadata items into the PALMD::Metadata, combining with any provided by the
1208 // frontend as LLVM metadata. Once all functions are written, the PAL metadata
1209 // is then written as a single block in the .note section.
1210 void AMDGPUAsmPrinter::EmitPALMetadata(const MachineFunction &MF,
1211        const SIProgramInfo &CurrentProgramInfo) {
1212   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1213   auto CC = MF.getFunction().getCallingConv();
1214   auto MD = getTargetStreamer()->getPALMetadata();
1215 
1216   MD->setEntryPoint(CC, MF.getFunction().getName());
1217   MD->setNumUsedVgprs(CC, CurrentProgramInfo.NumVGPRsForWavesPerEU);
1218   MD->setNumUsedSgprs(CC, CurrentProgramInfo.NumSGPRsForWavesPerEU);
1219   if (AMDGPU::isCompute(MF.getFunction().getCallingConv())) {
1220     MD->setRsrc1(CC, CurrentProgramInfo.ComputePGMRSrc1);
1221     MD->setRsrc2(CC, CurrentProgramInfo.ComputePGMRSrc2);
1222   } else {
1223     MD->setRsrc1(CC, S_00B028_VGPRS(CurrentProgramInfo.VGPRBlocks) |
1224         S_00B028_SGPRS(CurrentProgramInfo.SGPRBlocks));
1225     if (CurrentProgramInfo.ScratchBlocks > 0)
1226       MD->setRsrc2(CC, S_00B84C_SCRATCH_EN(1));
1227   }
1228   // ScratchSize is in bytes, 16 aligned.
1229   MD->setScratchSize(CC, alignTo(CurrentProgramInfo.ScratchSize, 16));
1230   if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS) {
1231     MD->setRsrc2(CC, S_00B02C_EXTRA_LDS_SIZE(CurrentProgramInfo.LDSBlocks));
1232     MD->setSpiPsInputEna(MFI->getPSInputEnable());
1233     MD->setSpiPsInputAddr(MFI->getPSInputAddr());
1234   }
1235 
1236   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
1237   if (STM.isWave32())
1238     MD->setWave32(MF.getFunction().getCallingConv());
1239 }
1240 
1241 // This is supposed to be log2(Size)
1242 static amd_element_byte_size_t getElementByteSizeValue(unsigned Size) {
1243   switch (Size) {
1244   case 4:
1245     return AMD_ELEMENT_4_BYTES;
1246   case 8:
1247     return AMD_ELEMENT_8_BYTES;
1248   case 16:
1249     return AMD_ELEMENT_16_BYTES;
1250   default:
1251     llvm_unreachable("invalid private_element_size");
1252   }
1253 }
1254 
1255 void AMDGPUAsmPrinter::getAmdKernelCode(amd_kernel_code_t &Out,
1256                                         const SIProgramInfo &CurrentProgramInfo,
1257                                         const MachineFunction &MF) const {
1258   const Function &F = MF.getFunction();
1259   assert(F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
1260          F.getCallingConv() == CallingConv::SPIR_KERNEL);
1261 
1262   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1263   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
1264 
1265   AMDGPU::initDefaultAMDKernelCodeT(Out, &STM);
1266 
1267   Out.compute_pgm_resource_registers =
1268       CurrentProgramInfo.ComputePGMRSrc1 |
1269       (CurrentProgramInfo.ComputePGMRSrc2 << 32);
1270   Out.code_properties |= AMD_CODE_PROPERTY_IS_PTR64;
1271 
1272   if (CurrentProgramInfo.DynamicCallStack)
1273     Out.code_properties |= AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK;
1274 
1275   AMD_HSA_BITS_SET(Out.code_properties,
1276                    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE,
1277                    getElementByteSizeValue(STM.getMaxPrivateElementSize()));
1278 
1279   if (MFI->hasPrivateSegmentBuffer()) {
1280     Out.code_properties |=
1281       AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
1282   }
1283 
1284   if (MFI->hasDispatchPtr())
1285     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
1286 
1287   if (MFI->hasQueuePtr())
1288     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
1289 
1290   if (MFI->hasKernargSegmentPtr())
1291     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
1292 
1293   if (MFI->hasDispatchID())
1294     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
1295 
1296   if (MFI->hasFlatScratchInit())
1297     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
1298 
1299   if (MFI->hasDispatchPtr())
1300     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
1301 
1302   if (STM.isXNACKEnabled())
1303     Out.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED;
1304 
1305   Align MaxKernArgAlign;
1306   Out.kernarg_segment_byte_size = STM.getKernArgSegmentSize(F, MaxKernArgAlign);
1307   Out.wavefront_sgpr_count = CurrentProgramInfo.NumSGPR;
1308   Out.workitem_vgpr_count = CurrentProgramInfo.NumVGPR;
1309   Out.workitem_private_segment_byte_size = CurrentProgramInfo.ScratchSize;
1310   Out.workgroup_group_segment_byte_size = CurrentProgramInfo.LDSSize;
1311 
1312   // kernarg_segment_alignment is specified as log of the alignment.
1313   // The minimum alignment is 16.
1314   Out.kernarg_segment_alignment = Log2(std::max(Align(16), MaxKernArgAlign));
1315 }
1316 
1317 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
1318                                        const char *ExtraCode, raw_ostream &O) {
1319   // First try the generic code, which knows about modifiers like 'c' and 'n'.
1320   if (!AsmPrinter::PrintAsmOperand(MI, OpNo, ExtraCode, O))
1321     return false;
1322 
1323   if (ExtraCode && ExtraCode[0]) {
1324     if (ExtraCode[1] != 0)
1325       return true; // Unknown modifier.
1326 
1327     switch (ExtraCode[0]) {
1328     case 'r':
1329       break;
1330     default:
1331       return true;
1332     }
1333   }
1334 
1335   // TODO: Should be able to support other operand types like globals.
1336   const MachineOperand &MO = MI->getOperand(OpNo);
1337   if (MO.isReg()) {
1338     AMDGPUInstPrinter::printRegOperand(MO.getReg(), O,
1339                                        *MF->getSubtarget().getRegisterInfo());
1340     return false;
1341   }
1342 
1343   return true;
1344 }
1345