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 // We need to tell the runtime some amount ahead of time if we don't know the
53 // true stack size. Assume a smaller number if this is only due to dynamic /
54 // non-entry block allocas.
55 static cl::opt<uint32_t> AssumedStackSizeForExternalCall(
56   "amdgpu-assume-external-call-stack-size",
57   cl::desc("Assumed stack use of any external call (in bytes)"),
58   cl::Hidden,
59   cl::init(16384));
60 
61 static cl::opt<uint32_t> AssumedStackSizeForDynamicSizeObjects(
62   "amdgpu-assume-dynamic-stack-object-size",
63   cl::desc("Assumed extra stack use if there are any "
64            "variable sized objects (in bytes)"),
65   cl::Hidden,
66   cl::init(4096));
67 
68 // This should get the default rounding mode from the kernel. We just set the
69 // default here, but this could change if the OpenCL rounding mode pragmas are
70 // used.
71 //
72 // The denormal mode here should match what is reported by the OpenCL runtime
73 // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but
74 // can also be override to flush with the -cl-denorms-are-zero compiler flag.
75 //
76 // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double
77 // precision, and leaves single precision to flush all and does not report
78 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
79 // CL_FP_DENORM for both.
80 //
81 // FIXME: It seems some instructions do not support single precision denormals
82 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
83 // and sin_f32, cos_f32 on most parts).
84 
85 // We want to use these instructions, and using fp32 denormals also causes
86 // instructions to run at the double precision rate for the device so it's
87 // probably best to just report no single precision denormals.
getFPMode(AMDGPU::SIModeRegisterDefaults Mode)88 static uint32_t getFPMode(AMDGPU::SIModeRegisterDefaults Mode) {
89   return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
90          FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
91          FP_DENORM_MODE_SP(Mode.fpDenormModeSPValue()) |
92          FP_DENORM_MODE_DP(Mode.fpDenormModeDPValue());
93 }
94 
95 static AsmPrinter *
createAMDGPUAsmPrinterPass(TargetMachine & tm,std::unique_ptr<MCStreamer> && Streamer)96 createAMDGPUAsmPrinterPass(TargetMachine &tm,
97                            std::unique_ptr<MCStreamer> &&Streamer) {
98   return new AMDGPUAsmPrinter(tm, std::move(Streamer));
99 }
100 
LLVMInitializeAMDGPUAsmPrinter()101 extern "C" void LLVM_EXTERNAL_VISIBILITY LLVMInitializeAMDGPUAsmPrinter() {
102   TargetRegistry::RegisterAsmPrinter(getTheAMDGPUTarget(),
103                                      llvm::createR600AsmPrinterPass);
104   TargetRegistry::RegisterAsmPrinter(getTheGCNTarget(),
105                                      createAMDGPUAsmPrinterPass);
106 }
107 
AMDGPUAsmPrinter(TargetMachine & TM,std::unique_ptr<MCStreamer> Streamer)108 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM,
109                                    std::unique_ptr<MCStreamer> Streamer)
110   : AsmPrinter(TM, std::move(Streamer)) {
111   if (TM.getTargetTriple().getOS() == Triple::AMDHSA) {
112     if (isHsaAbiVersion2(getGlobalSTI())) {
113       HSAMetadataStream.reset(new MetadataStreamerV2());
114     } else {
115       HSAMetadataStream.reset(new MetadataStreamerV3());
116     }
117   }
118 }
119 
getPassName() const120 StringRef AMDGPUAsmPrinter::getPassName() const {
121   return "AMDGPU Assembly Printer";
122 }
123 
getGlobalSTI() const124 const MCSubtargetInfo *AMDGPUAsmPrinter::getGlobalSTI() const {
125   return TM.getMCSubtargetInfo();
126 }
127 
getTargetStreamer() const128 AMDGPUTargetStreamer* AMDGPUAsmPrinter::getTargetStreamer() const {
129   if (!OutStreamer)
130     return nullptr;
131   return static_cast<AMDGPUTargetStreamer*>(OutStreamer->getTargetStreamer());
132 }
133 
emitStartOfAsmFile(Module & M)134 void AMDGPUAsmPrinter::emitStartOfAsmFile(Module &M) {
135   if (isHsaAbiVersion3(getGlobalSTI())) {
136     std::string ExpectedTarget;
137     raw_string_ostream ExpectedTargetOS(ExpectedTarget);
138     IsaInfo::streamIsaVersion(getGlobalSTI(), ExpectedTargetOS);
139 
140     getTargetStreamer()->EmitDirectiveAMDGCNTarget(ExpectedTarget);
141   }
142 
143   if (TM.getTargetTriple().getOS() != Triple::AMDHSA &&
144       TM.getTargetTriple().getOS() != Triple::AMDPAL)
145     return;
146 
147   if (TM.getTargetTriple().getOS() == Triple::AMDHSA)
148     HSAMetadataStream->begin(M);
149 
150   if (TM.getTargetTriple().getOS() == Triple::AMDPAL)
151     getTargetStreamer()->getPALMetadata()->readFromIR(M);
152 
153   if (isHsaAbiVersion3(getGlobalSTI()))
154     return;
155 
156   // HSA emits NT_AMDGPU_HSA_CODE_OBJECT_VERSION for code objects v2.
157   if (TM.getTargetTriple().getOS() == Triple::AMDHSA)
158     getTargetStreamer()->EmitDirectiveHSACodeObjectVersion(2, 1);
159 
160   // HSA and PAL emit NT_AMDGPU_HSA_ISA for code objects v2.
161   IsaVersion Version = getIsaVersion(getGlobalSTI()->getCPU());
162   getTargetStreamer()->EmitDirectiveHSACodeObjectISA(
163       Version.Major, Version.Minor, Version.Stepping, "AMD", "AMDGPU");
164 }
165 
emitEndOfAsmFile(Module & M)166 void AMDGPUAsmPrinter::emitEndOfAsmFile(Module &M) {
167   // Following code requires TargetStreamer to be present.
168   if (!getTargetStreamer())
169     return;
170 
171   if (TM.getTargetTriple().getOS() != Triple::AMDHSA ||
172       isHsaAbiVersion2(getGlobalSTI())) {
173     // Emit ISA Version (NT_AMD_AMDGPU_ISA).
174     std::string ISAVersionString;
175     raw_string_ostream ISAVersionStream(ISAVersionString);
176     IsaInfo::streamIsaVersion(getGlobalSTI(), ISAVersionStream);
177     getTargetStreamer()->EmitISAVersion(ISAVersionStream.str());
178   }
179 
180   // Emit HSA Metadata (NT_AMD_AMDGPU_HSA_METADATA).
181   if (TM.getTargetTriple().getOS() == Triple::AMDHSA) {
182     HSAMetadataStream->end();
183     bool Success = HSAMetadataStream->emitTo(*getTargetStreamer());
184     (void)Success;
185     assert(Success && "Malformed HSA Metadata");
186   }
187 }
188 
isBlockOnlyReachableByFallthrough(const MachineBasicBlock * MBB) const189 bool AMDGPUAsmPrinter::isBlockOnlyReachableByFallthrough(
190   const MachineBasicBlock *MBB) const {
191   if (!AsmPrinter::isBlockOnlyReachableByFallthrough(MBB))
192     return false;
193 
194   if (MBB->empty())
195     return true;
196 
197   // If this is a block implementing a long branch, an expression relative to
198   // the start of the block is needed.  to the start of the block.
199   // XXX - Is there a smarter way to check this?
200   return (MBB->back().getOpcode() != AMDGPU::S_SETPC_B64);
201 }
202 
emitFunctionBodyStart()203 void AMDGPUAsmPrinter::emitFunctionBodyStart() {
204   const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>();
205   if (!MFI.isEntryFunction())
206     return;
207 
208   const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>();
209   const Function &F = MF->getFunction();
210   if ((STM.isMesaKernel(F) || isHsaAbiVersion2(getGlobalSTI())) &&
211       (F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
212        F.getCallingConv() == CallingConv::SPIR_KERNEL)) {
213     amd_kernel_code_t KernelCode;
214     getAmdKernelCode(KernelCode, CurrentProgramInfo, *MF);
215     getTargetStreamer()->EmitAMDKernelCodeT(KernelCode);
216   }
217 
218   if (STM.isAmdHsaOS())
219     HSAMetadataStream->emitKernel(*MF, CurrentProgramInfo);
220 }
221 
emitFunctionBodyEnd()222 void AMDGPUAsmPrinter::emitFunctionBodyEnd() {
223   const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>();
224   if (!MFI.isEntryFunction())
225     return;
226 
227   if (TM.getTargetTriple().getOS() != Triple::AMDHSA ||
228       isHsaAbiVersion2(getGlobalSTI()))
229     return;
230 
231   auto &Streamer = getTargetStreamer()->getStreamer();
232   auto &Context = Streamer.getContext();
233   auto &ObjectFileInfo = *Context.getObjectFileInfo();
234   auto &ReadOnlySection = *ObjectFileInfo.getReadOnlySection();
235 
236   Streamer.PushSection();
237   Streamer.SwitchSection(&ReadOnlySection);
238 
239   // CP microcode requires the kernel descriptor to be allocated on 64 byte
240   // alignment.
241   Streamer.emitValueToAlignment(64, 0, 1, 0);
242   if (ReadOnlySection.getAlignment() < 64)
243     ReadOnlySection.setAlignment(Align(64));
244 
245   const MCSubtargetInfo &STI = MF->getSubtarget();
246 
247   SmallString<128> KernelName;
248   getNameWithPrefix(KernelName, &MF->getFunction());
249   getTargetStreamer()->EmitAmdhsaKernelDescriptor(
250       STI, KernelName, getAmdhsaKernelDescriptor(*MF, CurrentProgramInfo),
251       CurrentProgramInfo.NumVGPRsForWavesPerEU,
252       CurrentProgramInfo.NumSGPRsForWavesPerEU -
253           IsaInfo::getNumExtraSGPRs(&STI,
254                                     CurrentProgramInfo.VCCUsed,
255                                     CurrentProgramInfo.FlatUsed),
256       CurrentProgramInfo.VCCUsed, CurrentProgramInfo.FlatUsed,
257       hasXNACK(STI));
258 
259   Streamer.PopSection();
260 }
261 
emitFunctionEntryLabel()262 void AMDGPUAsmPrinter::emitFunctionEntryLabel() {
263   if (TM.getTargetTriple().getOS() == Triple::AMDHSA &&
264       isHsaAbiVersion3(getGlobalSTI())) {
265     AsmPrinter::emitFunctionEntryLabel();
266     return;
267   }
268 
269   const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
270   const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>();
271   if (MFI->isEntryFunction() && STM.isAmdHsaOrMesa(MF->getFunction())) {
272     SmallString<128> SymbolName;
273     getNameWithPrefix(SymbolName, &MF->getFunction()),
274     getTargetStreamer()->EmitAMDGPUSymbolType(
275         SymbolName, ELF::STT_AMDGPU_HSA_KERNEL);
276   }
277   if (DumpCodeInstEmitter) {
278     // Disassemble function name label to text.
279     DisasmLines.push_back(MF->getName().str() + ":");
280     DisasmLineMaxLen = std::max(DisasmLineMaxLen, DisasmLines.back().size());
281     HexLines.push_back("");
282   }
283 
284   AsmPrinter::emitFunctionEntryLabel();
285 }
286 
emitBasicBlockStart(const MachineBasicBlock & MBB)287 void AMDGPUAsmPrinter::emitBasicBlockStart(const MachineBasicBlock &MBB) {
288   if (DumpCodeInstEmitter && !isBlockOnlyReachableByFallthrough(&MBB)) {
289     // Write a line for the basic block label if it is not only fallthrough.
290     DisasmLines.push_back(
291         (Twine("BB") + Twine(getFunctionNumber())
292          + "_" + Twine(MBB.getNumber()) + ":").str());
293     DisasmLineMaxLen = std::max(DisasmLineMaxLen, DisasmLines.back().size());
294     HexLines.push_back("");
295   }
296   AsmPrinter::emitBasicBlockStart(MBB);
297 }
298 
emitGlobalVariable(const GlobalVariable * GV)299 void AMDGPUAsmPrinter::emitGlobalVariable(const GlobalVariable *GV) {
300   if (GV->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
301     if (GV->hasInitializer() && !isa<UndefValue>(GV->getInitializer())) {
302       OutContext.reportError({},
303                              Twine(GV->getName()) +
304                                  ": unsupported initializer for address space");
305       return;
306     }
307 
308     // LDS variables aren't emitted in HSA or PAL yet.
309     const Triple::OSType OS = TM.getTargetTriple().getOS();
310     if (OS == Triple::AMDHSA || OS == Triple::AMDPAL)
311       return;
312 
313     MCSymbol *GVSym = getSymbol(GV);
314 
315     GVSym->redefineIfPossible();
316     if (GVSym->isDefined() || GVSym->isVariable())
317       report_fatal_error("symbol '" + Twine(GVSym->getName()) +
318                          "' is already defined");
319 
320     const DataLayout &DL = GV->getParent()->getDataLayout();
321     uint64_t Size = DL.getTypeAllocSize(GV->getValueType());
322     Align Alignment = GV->getAlign().getValueOr(Align(4));
323 
324     emitVisibility(GVSym, GV->getVisibility(), !GV->isDeclaration());
325     emitLinkage(GV, GVSym);
326     if (auto TS = getTargetStreamer())
327       TS->emitAMDGPULDS(GVSym, Size, Alignment);
328     return;
329   }
330 
331   AsmPrinter::emitGlobalVariable(GV);
332 }
333 
doFinalization(Module & M)334 bool AMDGPUAsmPrinter::doFinalization(Module &M) {
335   CallGraphResourceInfo.clear();
336 
337   // Pad with s_code_end to help tools and guard against instruction prefetch
338   // causing stale data in caches. Arguably this should be done by the linker,
339   // which is why this isn't done for Mesa.
340   const MCSubtargetInfo &STI = *getGlobalSTI();
341   if (AMDGPU::isGFX10(STI) &&
342       (STI.getTargetTriple().getOS() == Triple::AMDHSA ||
343        STI.getTargetTriple().getOS() == Triple::AMDPAL)) {
344     OutStreamer->SwitchSection(getObjFileLowering().getTextSection());
345     getTargetStreamer()->EmitCodeEnd();
346   }
347 
348   return AsmPrinter::doFinalization(M);
349 }
350 
351 // Print comments that apply to both callable functions and entry points.
emitCommonFunctionComments(uint32_t NumVGPR,Optional<uint32_t> NumAGPR,uint32_t TotalNumVGPR,uint32_t NumSGPR,uint64_t ScratchSize,uint64_t CodeSize,const AMDGPUMachineFunction * MFI)352 void AMDGPUAsmPrinter::emitCommonFunctionComments(
353   uint32_t NumVGPR,
354   Optional<uint32_t> NumAGPR,
355   uint32_t TotalNumVGPR,
356   uint32_t NumSGPR,
357   uint64_t ScratchSize,
358   uint64_t CodeSize,
359   const AMDGPUMachineFunction *MFI) {
360   OutStreamer->emitRawComment(" codeLenInByte = " + Twine(CodeSize), false);
361   OutStreamer->emitRawComment(" NumSgprs: " + Twine(NumSGPR), false);
362   OutStreamer->emitRawComment(" NumVgprs: " + Twine(NumVGPR), false);
363   if (NumAGPR) {
364     OutStreamer->emitRawComment(" NumAgprs: " + Twine(*NumAGPR), false);
365     OutStreamer->emitRawComment(" TotalNumVgprs: " + Twine(TotalNumVGPR),
366                                 false);
367   }
368   OutStreamer->emitRawComment(" ScratchSize: " + Twine(ScratchSize), false);
369   OutStreamer->emitRawComment(" MemoryBound: " + Twine(MFI->isMemoryBound()),
370                               false);
371 }
372 
getAmdhsaKernelCodeProperties(const MachineFunction & MF) const373 uint16_t AMDGPUAsmPrinter::getAmdhsaKernelCodeProperties(
374     const MachineFunction &MF) const {
375   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
376   uint16_t KernelCodeProperties = 0;
377 
378   if (MFI.hasPrivateSegmentBuffer()) {
379     KernelCodeProperties |=
380         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
381   }
382   if (MFI.hasDispatchPtr()) {
383     KernelCodeProperties |=
384         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
385   }
386   if (MFI.hasQueuePtr()) {
387     KernelCodeProperties |=
388         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
389   }
390   if (MFI.hasKernargSegmentPtr()) {
391     KernelCodeProperties |=
392         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
393   }
394   if (MFI.hasDispatchID()) {
395     KernelCodeProperties |=
396         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
397   }
398   if (MFI.hasFlatScratchInit()) {
399     KernelCodeProperties |=
400         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
401   }
402   if (MF.getSubtarget<GCNSubtarget>().isWave32()) {
403     KernelCodeProperties |=
404         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32;
405   }
406 
407   return KernelCodeProperties;
408 }
409 
getAmdhsaKernelDescriptor(const MachineFunction & MF,const SIProgramInfo & PI) const410 amdhsa::kernel_descriptor_t AMDGPUAsmPrinter::getAmdhsaKernelDescriptor(
411     const MachineFunction &MF,
412     const SIProgramInfo &PI) const {
413   amdhsa::kernel_descriptor_t KernelDescriptor;
414   memset(&KernelDescriptor, 0x0, sizeof(KernelDescriptor));
415 
416   assert(isUInt<32>(PI.ScratchSize));
417   assert(isUInt<32>(PI.getComputePGMRSrc1()));
418   assert(isUInt<32>(PI.ComputePGMRSrc2));
419 
420   KernelDescriptor.group_segment_fixed_size = PI.LDSSize;
421   KernelDescriptor.private_segment_fixed_size = PI.ScratchSize;
422   KernelDescriptor.compute_pgm_rsrc1 = PI.getComputePGMRSrc1();
423   KernelDescriptor.compute_pgm_rsrc2 = PI.ComputePGMRSrc2;
424   KernelDescriptor.kernel_code_properties = getAmdhsaKernelCodeProperties(MF);
425 
426   return KernelDescriptor;
427 }
428 
runOnMachineFunction(MachineFunction & MF)429 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
430   CurrentProgramInfo = SIProgramInfo();
431 
432   const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
433 
434   // The starting address of all shader programs must be 256 bytes aligned.
435   // Regular functions just need the basic required instruction alignment.
436   MF.setAlignment(MFI->isEntryFunction() ? Align(256) : Align(4));
437 
438   SetupMachineFunction(MF);
439 
440   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
441   MCContext &Context = getObjFileLowering().getContext();
442   // FIXME: This should be an explicit check for Mesa.
443   if (!STM.isAmdHsaOS() && !STM.isAmdPalOS()) {
444     MCSectionELF *ConfigSection =
445         Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0);
446     OutStreamer->SwitchSection(ConfigSection);
447   }
448 
449   if (MFI->isEntryFunction()) {
450     getSIProgramInfo(CurrentProgramInfo, MF);
451   } else {
452     auto I = CallGraphResourceInfo.insert(
453       std::make_pair(&MF.getFunction(), SIFunctionResourceInfo()));
454     SIFunctionResourceInfo &Info = I.first->second;
455     assert(I.second && "should only be called once per function");
456     Info = analyzeResourceUsage(MF);
457   }
458 
459   if (STM.isAmdPalOS() && MFI->isEntryFunction())
460     EmitPALMetadata(MF, CurrentProgramInfo);
461   else if (!STM.isAmdHsaOS()) {
462     EmitProgramInfoSI(MF, CurrentProgramInfo);
463   }
464 
465   DumpCodeInstEmitter = nullptr;
466   if (STM.dumpCode()) {
467     // For -dumpcode, get the assembler out of the streamer, even if it does
468     // not really want to let us have it. This only works with -filetype=obj.
469     bool SaveFlag = OutStreamer->getUseAssemblerInfoForParsing();
470     OutStreamer->setUseAssemblerInfoForParsing(true);
471     MCAssembler *Assembler = OutStreamer->getAssemblerPtr();
472     OutStreamer->setUseAssemblerInfoForParsing(SaveFlag);
473     if (Assembler)
474       DumpCodeInstEmitter = Assembler->getEmitterPtr();
475   }
476 
477   DisasmLines.clear();
478   HexLines.clear();
479   DisasmLineMaxLen = 0;
480 
481   emitFunctionBody();
482 
483   if (isVerbose()) {
484     MCSectionELF *CommentSection =
485         Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0);
486     OutStreamer->SwitchSection(CommentSection);
487 
488     if (!MFI->isEntryFunction()) {
489       OutStreamer->emitRawComment(" Function info:", false);
490       SIFunctionResourceInfo &Info = CallGraphResourceInfo[&MF.getFunction()];
491       emitCommonFunctionComments(
492         Info.NumVGPR,
493         STM.hasMAIInsts() ? Info.NumAGPR : Optional<uint32_t>(),
494         Info.getTotalNumVGPRs(STM),
495         Info.getTotalNumSGPRs(MF.getSubtarget<GCNSubtarget>()),
496         Info.PrivateSegmentSize,
497         getFunctionCodeSize(MF), MFI);
498       return false;
499     }
500 
501     OutStreamer->emitRawComment(" Kernel info:", false);
502     emitCommonFunctionComments(CurrentProgramInfo.NumArchVGPR,
503                                STM.hasMAIInsts()
504                                  ? CurrentProgramInfo.NumAccVGPR
505                                  : Optional<uint32_t>(),
506                                CurrentProgramInfo.NumVGPR,
507                                CurrentProgramInfo.NumSGPR,
508                                CurrentProgramInfo.ScratchSize,
509                                getFunctionCodeSize(MF), MFI);
510 
511     OutStreamer->emitRawComment(
512       " FloatMode: " + Twine(CurrentProgramInfo.FloatMode), false);
513     OutStreamer->emitRawComment(
514       " IeeeMode: " + Twine(CurrentProgramInfo.IEEEMode), false);
515     OutStreamer->emitRawComment(
516       " LDSByteSize: " + Twine(CurrentProgramInfo.LDSSize) +
517       " bytes/workgroup (compile time only)", false);
518 
519     OutStreamer->emitRawComment(
520       " SGPRBlocks: " + Twine(CurrentProgramInfo.SGPRBlocks), false);
521     OutStreamer->emitRawComment(
522       " VGPRBlocks: " + Twine(CurrentProgramInfo.VGPRBlocks), false);
523 
524     OutStreamer->emitRawComment(
525       " NumSGPRsForWavesPerEU: " +
526       Twine(CurrentProgramInfo.NumSGPRsForWavesPerEU), false);
527     OutStreamer->emitRawComment(
528       " NumVGPRsForWavesPerEU: " +
529       Twine(CurrentProgramInfo.NumVGPRsForWavesPerEU), false);
530 
531     OutStreamer->emitRawComment(
532       " Occupancy: " +
533       Twine(CurrentProgramInfo.Occupancy), false);
534 
535     OutStreamer->emitRawComment(
536       " WaveLimiterHint : " + Twine(MFI->needsWaveLimiter()), false);
537 
538     OutStreamer->emitRawComment(
539       " COMPUTE_PGM_RSRC2:USER_SGPR: " +
540       Twine(G_00B84C_USER_SGPR(CurrentProgramInfo.ComputePGMRSrc2)), false);
541     OutStreamer->emitRawComment(
542       " COMPUTE_PGM_RSRC2:TRAP_HANDLER: " +
543       Twine(G_00B84C_TRAP_HANDLER(CurrentProgramInfo.ComputePGMRSrc2)), false);
544     OutStreamer->emitRawComment(
545       " COMPUTE_PGM_RSRC2:TGID_X_EN: " +
546       Twine(G_00B84C_TGID_X_EN(CurrentProgramInfo.ComputePGMRSrc2)), false);
547     OutStreamer->emitRawComment(
548       " COMPUTE_PGM_RSRC2:TGID_Y_EN: " +
549       Twine(G_00B84C_TGID_Y_EN(CurrentProgramInfo.ComputePGMRSrc2)), false);
550     OutStreamer->emitRawComment(
551       " COMPUTE_PGM_RSRC2:TGID_Z_EN: " +
552       Twine(G_00B84C_TGID_Z_EN(CurrentProgramInfo.ComputePGMRSrc2)), false);
553     OutStreamer->emitRawComment(
554       " COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " +
555       Twine(G_00B84C_TIDIG_COMP_CNT(CurrentProgramInfo.ComputePGMRSrc2)),
556       false);
557   }
558 
559   if (DumpCodeInstEmitter) {
560 
561     OutStreamer->SwitchSection(
562         Context.getELFSection(".AMDGPU.disasm", ELF::SHT_PROGBITS, 0));
563 
564     for (size_t i = 0; i < DisasmLines.size(); ++i) {
565       std::string Comment = "\n";
566       if (!HexLines[i].empty()) {
567         Comment = std::string(DisasmLineMaxLen - DisasmLines[i].size(), ' ');
568         Comment += " ; " + HexLines[i] + "\n";
569       }
570 
571       OutStreamer->emitBytes(StringRef(DisasmLines[i]));
572       OutStreamer->emitBytes(StringRef(Comment));
573     }
574   }
575 
576   return false;
577 }
578 
getFunctionCodeSize(const MachineFunction & MF) const579 uint64_t AMDGPUAsmPrinter::getFunctionCodeSize(const MachineFunction &MF) const {
580   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
581   const SIInstrInfo *TII = STM.getInstrInfo();
582 
583   uint64_t CodeSize = 0;
584 
585   for (const MachineBasicBlock &MBB : MF) {
586     for (const MachineInstr &MI : MBB) {
587       // TODO: CodeSize should account for multiple functions.
588 
589       // TODO: Should we count size of debug info?
590       if (MI.isDebugInstr())
591         continue;
592 
593       CodeSize += TII->getInstSizeInBytes(MI);
594     }
595   }
596 
597   return CodeSize;
598 }
599 
hasAnyNonFlatUseOfReg(const MachineRegisterInfo & MRI,const SIInstrInfo & TII,unsigned Reg)600 static bool hasAnyNonFlatUseOfReg(const MachineRegisterInfo &MRI,
601                                   const SIInstrInfo &TII,
602                                   unsigned Reg) {
603   for (const MachineOperand &UseOp : MRI.reg_operands(Reg)) {
604     if (!UseOp.isImplicit() || !TII.isFLAT(*UseOp.getParent()))
605       return true;
606   }
607 
608   return false;
609 }
610 
getTotalNumSGPRs(const GCNSubtarget & ST) const611 int32_t AMDGPUAsmPrinter::SIFunctionResourceInfo::getTotalNumSGPRs(
612   const GCNSubtarget &ST) const {
613   return NumExplicitSGPR + IsaInfo::getNumExtraSGPRs(&ST,
614                                                      UsesVCC, UsesFlatScratch);
615 }
616 
getTotalNumVGPRs(const GCNSubtarget & ST) const617 int32_t AMDGPUAsmPrinter::SIFunctionResourceInfo::getTotalNumVGPRs(
618   const GCNSubtarget &ST) const {
619   return std::max(NumVGPR, NumAGPR);
620 }
621 
getCalleeFunction(const MachineOperand & Op)622 static const Function *getCalleeFunction(const MachineOperand &Op) {
623   if (Op.isImm()) {
624     assert(Op.getImm() == 0);
625     return nullptr;
626   }
627 
628   return cast<Function>(Op.getGlobal());
629 }
630 
analyzeResourceUsage(const MachineFunction & MF) const631 AMDGPUAsmPrinter::SIFunctionResourceInfo AMDGPUAsmPrinter::analyzeResourceUsage(
632   const MachineFunction &MF) const {
633   SIFunctionResourceInfo Info;
634 
635   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
636   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
637   const MachineFrameInfo &FrameInfo = MF.getFrameInfo();
638   const MachineRegisterInfo &MRI = MF.getRegInfo();
639   const SIInstrInfo *TII = ST.getInstrInfo();
640   const SIRegisterInfo &TRI = TII->getRegisterInfo();
641 
642   Info.UsesFlatScratch = MRI.isPhysRegUsed(AMDGPU::FLAT_SCR_LO) ||
643                          MRI.isPhysRegUsed(AMDGPU::FLAT_SCR_HI);
644 
645   // Even if FLAT_SCRATCH is implicitly used, it has no effect if flat
646   // instructions aren't used to access the scratch buffer. Inline assembly may
647   // need it though.
648   //
649   // If we only have implicit uses of flat_scr on flat instructions, it is not
650   // really needed.
651   if (Info.UsesFlatScratch && !MFI->hasFlatScratchInit() &&
652       (!hasAnyNonFlatUseOfReg(MRI, *TII, AMDGPU::FLAT_SCR) &&
653        !hasAnyNonFlatUseOfReg(MRI, *TII, AMDGPU::FLAT_SCR_LO) &&
654        !hasAnyNonFlatUseOfReg(MRI, *TII, AMDGPU::FLAT_SCR_HI))) {
655     Info.UsesFlatScratch = false;
656   }
657 
658   Info.PrivateSegmentSize = FrameInfo.getStackSize();
659 
660   // Assume a big number if there are any unknown sized objects.
661   Info.HasDynamicallySizedStack = FrameInfo.hasVarSizedObjects();
662   if (Info.HasDynamicallySizedStack)
663     Info.PrivateSegmentSize += AssumedStackSizeForDynamicSizeObjects;
664 
665   if (MFI->isStackRealigned())
666     Info.PrivateSegmentSize += FrameInfo.getMaxAlign().value();
667 
668   Info.UsesVCC = MRI.isPhysRegUsed(AMDGPU::VCC_LO) ||
669                  MRI.isPhysRegUsed(AMDGPU::VCC_HI);
670 
671   // If there are no calls, MachineRegisterInfo can tell us the used register
672   // count easily.
673   // A tail call isn't considered a call for MachineFrameInfo's purposes.
674   if (!FrameInfo.hasCalls() && !FrameInfo.hasTailCall()) {
675     MCPhysReg HighestVGPRReg = AMDGPU::NoRegister;
676     for (MCPhysReg Reg : reverse(AMDGPU::VGPR_32RegClass.getRegisters())) {
677       if (MRI.isPhysRegUsed(Reg)) {
678         HighestVGPRReg = Reg;
679         break;
680       }
681     }
682 
683     if (ST.hasMAIInsts()) {
684       MCPhysReg HighestAGPRReg = AMDGPU::NoRegister;
685       for (MCPhysReg Reg : reverse(AMDGPU::AGPR_32RegClass.getRegisters())) {
686         if (MRI.isPhysRegUsed(Reg)) {
687           HighestAGPRReg = Reg;
688           break;
689         }
690       }
691       Info.NumAGPR = HighestAGPRReg == AMDGPU::NoRegister ? 0 :
692         TRI.getHWRegIndex(HighestAGPRReg) + 1;
693     }
694 
695     MCPhysReg HighestSGPRReg = AMDGPU::NoRegister;
696     for (MCPhysReg Reg : reverse(AMDGPU::SGPR_32RegClass.getRegisters())) {
697       if (MRI.isPhysRegUsed(Reg)) {
698         HighestSGPRReg = Reg;
699         break;
700       }
701     }
702 
703     // We found the maximum register index. They start at 0, so add one to get the
704     // number of registers.
705     Info.NumVGPR = HighestVGPRReg == AMDGPU::NoRegister ? 0 :
706       TRI.getHWRegIndex(HighestVGPRReg) + 1;
707     Info.NumExplicitSGPR = HighestSGPRReg == AMDGPU::NoRegister ? 0 :
708       TRI.getHWRegIndex(HighestSGPRReg) + 1;
709 
710     return Info;
711   }
712 
713   int32_t MaxVGPR = -1;
714   int32_t MaxAGPR = -1;
715   int32_t MaxSGPR = -1;
716   uint64_t CalleeFrameSize = 0;
717 
718   for (const MachineBasicBlock &MBB : MF) {
719     for (const MachineInstr &MI : MBB) {
720       // TODO: Check regmasks? Do they occur anywhere except calls?
721       for (const MachineOperand &MO : MI.operands()) {
722         unsigned Width = 0;
723         bool IsSGPR = false;
724         bool IsAGPR = false;
725 
726         if (!MO.isReg())
727           continue;
728 
729         Register Reg = MO.getReg();
730         switch (Reg) {
731         case AMDGPU::EXEC:
732         case AMDGPU::EXEC_LO:
733         case AMDGPU::EXEC_HI:
734         case AMDGPU::SCC:
735         case AMDGPU::M0:
736         case AMDGPU::SRC_SHARED_BASE:
737         case AMDGPU::SRC_SHARED_LIMIT:
738         case AMDGPU::SRC_PRIVATE_BASE:
739         case AMDGPU::SRC_PRIVATE_LIMIT:
740         case AMDGPU::SGPR_NULL:
741         case AMDGPU::MODE:
742           continue;
743 
744         case AMDGPU::SRC_POPS_EXITING_WAVE_ID:
745           llvm_unreachable("src_pops_exiting_wave_id should not be used");
746 
747         case AMDGPU::NoRegister:
748           assert(MI.isDebugInstr() && "Instruction uses invalid noreg register");
749           continue;
750 
751         case AMDGPU::VCC:
752         case AMDGPU::VCC_LO:
753         case AMDGPU::VCC_HI:
754         case AMDGPU::VCC_LO_LO16:
755         case AMDGPU::VCC_LO_HI16:
756         case AMDGPU::VCC_HI_LO16:
757         case AMDGPU::VCC_HI_HI16:
758           Info.UsesVCC = true;
759           continue;
760 
761         case AMDGPU::FLAT_SCR:
762         case AMDGPU::FLAT_SCR_LO:
763         case AMDGPU::FLAT_SCR_HI:
764           continue;
765 
766         case AMDGPU::XNACK_MASK:
767         case AMDGPU::XNACK_MASK_LO:
768         case AMDGPU::XNACK_MASK_HI:
769           llvm_unreachable("xnack_mask registers should not be used");
770 
771         case AMDGPU::LDS_DIRECT:
772           llvm_unreachable("lds_direct register should not be used");
773 
774         case AMDGPU::TBA:
775         case AMDGPU::TBA_LO:
776         case AMDGPU::TBA_HI:
777         case AMDGPU::TMA:
778         case AMDGPU::TMA_LO:
779         case AMDGPU::TMA_HI:
780           llvm_unreachable("trap handler registers should not be used");
781 
782         case AMDGPU::SRC_VCCZ:
783           llvm_unreachable("src_vccz register should not be used");
784 
785         case AMDGPU::SRC_EXECZ:
786           llvm_unreachable("src_execz register should not be used");
787 
788         case AMDGPU::SRC_SCC:
789           llvm_unreachable("src_scc register should not be used");
790 
791         default:
792           break;
793         }
794 
795         if (AMDGPU::SReg_32RegClass.contains(Reg) ||
796             AMDGPU::SReg_LO16RegClass.contains(Reg) ||
797             AMDGPU::SGPR_HI16RegClass.contains(Reg)) {
798           assert(!AMDGPU::TTMP_32RegClass.contains(Reg) &&
799                  "trap handler registers should not be used");
800           IsSGPR = true;
801           Width = 1;
802         } else if (AMDGPU::VGPR_32RegClass.contains(Reg) ||
803                    AMDGPU::VGPR_LO16RegClass.contains(Reg) ||
804                    AMDGPU::VGPR_HI16RegClass.contains(Reg)) {
805           IsSGPR = false;
806           Width = 1;
807         } else if (AMDGPU::AGPR_32RegClass.contains(Reg) ||
808                    AMDGPU::AGPR_LO16RegClass.contains(Reg)) {
809           IsSGPR = false;
810           IsAGPR = true;
811           Width = 1;
812         } else if (AMDGPU::SReg_64RegClass.contains(Reg)) {
813           assert(!AMDGPU::TTMP_64RegClass.contains(Reg) &&
814                  "trap handler registers should not be used");
815           IsSGPR = true;
816           Width = 2;
817         } else if (AMDGPU::VReg_64RegClass.contains(Reg)) {
818           IsSGPR = false;
819           Width = 2;
820         } else if (AMDGPU::AReg_64RegClass.contains(Reg)) {
821           IsSGPR = false;
822           IsAGPR = true;
823           Width = 2;
824         } else if (AMDGPU::VReg_96RegClass.contains(Reg)) {
825           IsSGPR = false;
826           Width = 3;
827         } else if (AMDGPU::SReg_96RegClass.contains(Reg)) {
828           IsSGPR = true;
829           Width = 3;
830         } else if (AMDGPU::AReg_96RegClass.contains(Reg)) {
831           IsSGPR = false;
832           IsAGPR = true;
833           Width = 3;
834         } else if (AMDGPU::SReg_128RegClass.contains(Reg)) {
835           assert(!AMDGPU::TTMP_128RegClass.contains(Reg) &&
836             "trap handler registers should not be used");
837           IsSGPR = true;
838           Width = 4;
839         } else if (AMDGPU::VReg_128RegClass.contains(Reg)) {
840           IsSGPR = false;
841           Width = 4;
842         } else if (AMDGPU::AReg_128RegClass.contains(Reg)) {
843           IsSGPR = false;
844           IsAGPR = true;
845           Width = 4;
846         } else if (AMDGPU::VReg_160RegClass.contains(Reg)) {
847           IsSGPR = false;
848           Width = 5;
849         } else if (AMDGPU::SReg_160RegClass.contains(Reg)) {
850           IsSGPR = true;
851           Width = 5;
852         } else if (AMDGPU::AReg_160RegClass.contains(Reg)) {
853           IsSGPR = false;
854           IsAGPR = true;
855           Width = 5;
856         } else if (AMDGPU::VReg_192RegClass.contains(Reg)) {
857           IsSGPR = false;
858           Width = 6;
859         } else if (AMDGPU::SReg_192RegClass.contains(Reg)) {
860           IsSGPR = true;
861           Width = 6;
862         } else if (AMDGPU::AReg_192RegClass.contains(Reg)) {
863           IsSGPR = false;
864           IsAGPR = true;
865           Width = 6;
866         } else if (AMDGPU::SReg_256RegClass.contains(Reg)) {
867           assert(!AMDGPU::TTMP_256RegClass.contains(Reg) &&
868             "trap handler registers should not be used");
869           IsSGPR = true;
870           Width = 8;
871         } else if (AMDGPU::VReg_256RegClass.contains(Reg)) {
872           IsSGPR = false;
873           Width = 8;
874         } else if (AMDGPU::AReg_256RegClass.contains(Reg)) {
875           IsSGPR = false;
876           IsAGPR = true;
877           Width = 8;
878         } else if (AMDGPU::SReg_512RegClass.contains(Reg)) {
879           assert(!AMDGPU::TTMP_512RegClass.contains(Reg) &&
880             "trap handler registers should not be used");
881           IsSGPR = true;
882           Width = 16;
883         } else if (AMDGPU::VReg_512RegClass.contains(Reg)) {
884           IsSGPR = false;
885           Width = 16;
886         } else if (AMDGPU::AReg_512RegClass.contains(Reg)) {
887           IsSGPR = false;
888           IsAGPR = true;
889           Width = 16;
890         } else if (AMDGPU::SReg_1024RegClass.contains(Reg)) {
891           IsSGPR = true;
892           Width = 32;
893         } else if (AMDGPU::VReg_1024RegClass.contains(Reg)) {
894           IsSGPR = false;
895           Width = 32;
896         } else if (AMDGPU::AReg_1024RegClass.contains(Reg)) {
897           IsSGPR = false;
898           IsAGPR = true;
899           Width = 32;
900         } else {
901           llvm_unreachable("Unknown register class");
902         }
903         unsigned HWReg = TRI.getHWRegIndex(Reg);
904         int MaxUsed = HWReg + Width - 1;
905         if (IsSGPR) {
906           MaxSGPR = MaxUsed > MaxSGPR ? MaxUsed : MaxSGPR;
907         } else if (IsAGPR) {
908           MaxAGPR = MaxUsed > MaxAGPR ? MaxUsed : MaxAGPR;
909         } else {
910           MaxVGPR = MaxUsed > MaxVGPR ? MaxUsed : MaxVGPR;
911         }
912       }
913 
914       if (MI.isCall()) {
915         // Pseudo used just to encode the underlying global. Is there a better
916         // way to track this?
917 
918         const MachineOperand *CalleeOp
919           = TII->getNamedOperand(MI, AMDGPU::OpName::callee);
920 
921         const Function *Callee = getCalleeFunction(*CalleeOp);
922         if (!Callee || Callee->isDeclaration()) {
923           // If this is a call to an external function, we can't do much. Make
924           // conservative guesses.
925 
926           // 48 SGPRs - vcc, - flat_scr, -xnack
927           int MaxSGPRGuess =
928             47 - IsaInfo::getNumExtraSGPRs(&ST, true, ST.hasFlatAddressSpace());
929           MaxSGPR = std::max(MaxSGPR, MaxSGPRGuess);
930           MaxVGPR = std::max(MaxVGPR, 23);
931           MaxAGPR = std::max(MaxAGPR, 23);
932 
933           CalleeFrameSize = std::max(CalleeFrameSize,
934             static_cast<uint64_t>(AssumedStackSizeForExternalCall));
935 
936           Info.UsesVCC = true;
937           Info.UsesFlatScratch = ST.hasFlatAddressSpace();
938           Info.HasDynamicallySizedStack = true;
939         } else {
940           // We force CodeGen to run in SCC order, so the callee's register
941           // usage etc. should be the cumulative usage of all callees.
942 
943           auto I = CallGraphResourceInfo.find(Callee);
944           if (I == CallGraphResourceInfo.end()) {
945             // Avoid crashing on undefined behavior with an illegal call to a
946             // kernel. If a callsite's calling convention doesn't match the
947             // function's, it's undefined behavior. If the callsite calling
948             // convention does match, that would have errored earlier.
949             // FIXME: The verifier shouldn't allow this.
950             if (AMDGPU::isEntryFunctionCC(Callee->getCallingConv()))
951               report_fatal_error("invalid call to entry function");
952 
953             llvm_unreachable("callee should have been handled before caller");
954           }
955 
956           MaxSGPR = std::max(I->second.NumExplicitSGPR - 1, MaxSGPR);
957           MaxVGPR = std::max(I->second.NumVGPR - 1, MaxVGPR);
958           MaxAGPR = std::max(I->second.NumAGPR - 1, MaxAGPR);
959           CalleeFrameSize
960             = std::max(I->second.PrivateSegmentSize, CalleeFrameSize);
961           Info.UsesVCC |= I->second.UsesVCC;
962           Info.UsesFlatScratch |= I->second.UsesFlatScratch;
963           Info.HasDynamicallySizedStack |= I->second.HasDynamicallySizedStack;
964           Info.HasRecursion |= I->second.HasRecursion;
965         }
966 
967         // FIXME: Call site could have norecurse on it
968         if (!Callee || !Callee->doesNotRecurse())
969           Info.HasRecursion = true;
970       }
971     }
972   }
973 
974   Info.NumExplicitSGPR = MaxSGPR + 1;
975   Info.NumVGPR = MaxVGPR + 1;
976   Info.NumAGPR = MaxAGPR + 1;
977   Info.PrivateSegmentSize += CalleeFrameSize;
978 
979   return Info;
980 }
981 
getSIProgramInfo(SIProgramInfo & ProgInfo,const MachineFunction & MF)982 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
983                                         const MachineFunction &MF) {
984   SIFunctionResourceInfo Info = analyzeResourceUsage(MF);
985   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
986 
987   ProgInfo.NumArchVGPR = Info.NumVGPR;
988   ProgInfo.NumAccVGPR = Info.NumAGPR;
989   ProgInfo.NumVGPR = Info.getTotalNumVGPRs(STM);
990   ProgInfo.NumSGPR = Info.NumExplicitSGPR;
991   ProgInfo.ScratchSize = Info.PrivateSegmentSize;
992   ProgInfo.VCCUsed = Info.UsesVCC;
993   ProgInfo.FlatUsed = Info.UsesFlatScratch;
994   ProgInfo.DynamicCallStack = Info.HasDynamicallySizedStack || Info.HasRecursion;
995 
996   const uint64_t MaxScratchPerWorkitem =
997       GCNSubtarget::MaxWaveScratchSize / STM.getWavefrontSize();
998   if (ProgInfo.ScratchSize > MaxScratchPerWorkitem) {
999     DiagnosticInfoStackSize DiagStackSize(MF.getFunction(),
1000                                           ProgInfo.ScratchSize, DS_Error);
1001     MF.getFunction().getContext().diagnose(DiagStackSize);
1002   }
1003 
1004   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1005 
1006   // TODO(scott.linder): The calculations related to SGPR/VGPR blocks are
1007   // duplicated in part in AMDGPUAsmParser::calculateGPRBlocks, and could be
1008   // unified.
1009   unsigned ExtraSGPRs = IsaInfo::getNumExtraSGPRs(
1010       &STM, ProgInfo.VCCUsed, ProgInfo.FlatUsed);
1011 
1012   // Check the addressable register limit before we add ExtraSGPRs.
1013   if (STM.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS &&
1014       !STM.hasSGPRInitBug()) {
1015     unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs();
1016     if (ProgInfo.NumSGPR > MaxAddressableNumSGPRs) {
1017       // This can happen due to a compiler bug or when using inline asm.
1018       LLVMContext &Ctx = MF.getFunction().getContext();
1019       DiagnosticInfoResourceLimit Diag(MF.getFunction(),
1020                                        "addressable scalar registers",
1021                                        ProgInfo.NumSGPR, DS_Error,
1022                                        DK_ResourceLimit,
1023                                        MaxAddressableNumSGPRs);
1024       Ctx.diagnose(Diag);
1025       ProgInfo.NumSGPR = MaxAddressableNumSGPRs - 1;
1026     }
1027   }
1028 
1029   // Account for extra SGPRs and VGPRs reserved for debugger use.
1030   ProgInfo.NumSGPR += ExtraSGPRs;
1031 
1032   // Ensure there are enough SGPRs and VGPRs for wave dispatch, where wave
1033   // dispatch registers are function args.
1034   unsigned WaveDispatchNumSGPR = 0, WaveDispatchNumVGPR = 0;
1035   for (auto &Arg : MF.getFunction().args()) {
1036     unsigned NumRegs = (Arg.getType()->getPrimitiveSizeInBits() + 31) / 32;
1037     if (Arg.hasAttribute(Attribute::InReg))
1038       WaveDispatchNumSGPR += NumRegs;
1039     else
1040       WaveDispatchNumVGPR += NumRegs;
1041   }
1042   ProgInfo.NumSGPR = std::max(ProgInfo.NumSGPR, WaveDispatchNumSGPR);
1043   ProgInfo.NumVGPR = std::max(ProgInfo.NumVGPR, WaveDispatchNumVGPR);
1044 
1045   // Adjust number of registers used to meet default/requested minimum/maximum
1046   // number of waves per execution unit request.
1047   ProgInfo.NumSGPRsForWavesPerEU = std::max(
1048     std::max(ProgInfo.NumSGPR, 1u), STM.getMinNumSGPRs(MFI->getMaxWavesPerEU()));
1049   ProgInfo.NumVGPRsForWavesPerEU = std::max(
1050     std::max(ProgInfo.NumVGPR, 1u), STM.getMinNumVGPRs(MFI->getMaxWavesPerEU()));
1051 
1052   if (STM.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS ||
1053       STM.hasSGPRInitBug()) {
1054     unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs();
1055     if (ProgInfo.NumSGPR > MaxAddressableNumSGPRs) {
1056       // This can happen due to a compiler bug or when using inline asm to use
1057       // the registers which are usually reserved for vcc etc.
1058       LLVMContext &Ctx = MF.getFunction().getContext();
1059       DiagnosticInfoResourceLimit Diag(MF.getFunction(),
1060                                        "scalar registers",
1061                                        ProgInfo.NumSGPR, DS_Error,
1062                                        DK_ResourceLimit,
1063                                        MaxAddressableNumSGPRs);
1064       Ctx.diagnose(Diag);
1065       ProgInfo.NumSGPR = MaxAddressableNumSGPRs;
1066       ProgInfo.NumSGPRsForWavesPerEU = MaxAddressableNumSGPRs;
1067     }
1068   }
1069 
1070   if (STM.hasSGPRInitBug()) {
1071     ProgInfo.NumSGPR =
1072         AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG;
1073     ProgInfo.NumSGPRsForWavesPerEU =
1074         AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG;
1075   }
1076 
1077   if (MFI->getNumUserSGPRs() > STM.getMaxNumUserSGPRs()) {
1078     LLVMContext &Ctx = MF.getFunction().getContext();
1079     DiagnosticInfoResourceLimit Diag(MF.getFunction(), "user SGPRs",
1080                                      MFI->getNumUserSGPRs(), DS_Error);
1081     Ctx.diagnose(Diag);
1082   }
1083 
1084   if (MFI->getLDSSize() > static_cast<unsigned>(STM.getLocalMemorySize())) {
1085     LLVMContext &Ctx = MF.getFunction().getContext();
1086     DiagnosticInfoResourceLimit Diag(MF.getFunction(), "local memory",
1087                                      MFI->getLDSSize(), DS_Error);
1088     Ctx.diagnose(Diag);
1089   }
1090 
1091   ProgInfo.SGPRBlocks = IsaInfo::getNumSGPRBlocks(
1092       &STM, ProgInfo.NumSGPRsForWavesPerEU);
1093   ProgInfo.VGPRBlocks = IsaInfo::getNumVGPRBlocks(
1094       &STM, ProgInfo.NumVGPRsForWavesPerEU);
1095 
1096   const SIModeRegisterDefaults Mode = MFI->getMode();
1097 
1098   // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
1099   // register.
1100   ProgInfo.FloatMode = getFPMode(Mode);
1101 
1102   ProgInfo.IEEEMode = Mode.IEEE;
1103 
1104   // Make clamp modifier on NaN input returns 0.
1105   ProgInfo.DX10Clamp = Mode.DX10Clamp;
1106 
1107   unsigned LDSAlignShift;
1108   if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
1109     // LDS is allocated in 64 dword blocks.
1110     LDSAlignShift = 8;
1111   } else {
1112     // LDS is allocated in 128 dword blocks.
1113     LDSAlignShift = 9;
1114   }
1115 
1116   unsigned LDSSpillSize =
1117     MFI->getLDSWaveSpillSize() * MFI->getMaxFlatWorkGroupSize();
1118 
1119   ProgInfo.LDSSize = MFI->getLDSSize() + LDSSpillSize;
1120   ProgInfo.LDSBlocks =
1121       alignTo(ProgInfo.LDSSize, 1ULL << LDSAlignShift) >> LDSAlignShift;
1122 
1123   // Scratch is allocated in 256 dword blocks.
1124   unsigned ScratchAlignShift = 10;
1125   // We need to program the hardware with the amount of scratch memory that
1126   // is used by the entire wave.  ProgInfo.ScratchSize is the amount of
1127   // scratch memory used per thread.
1128   ProgInfo.ScratchBlocks =
1129       alignTo(ProgInfo.ScratchSize * STM.getWavefrontSize(),
1130               1ULL << ScratchAlignShift) >>
1131       ScratchAlignShift;
1132 
1133   if (getIsaVersion(getGlobalSTI()->getCPU()).Major >= 10) {
1134     ProgInfo.WgpMode = STM.isCuModeEnabled() ? 0 : 1;
1135     ProgInfo.MemOrdered = 1;
1136   }
1137 
1138   // 0 = X, 1 = XY, 2 = XYZ
1139   unsigned TIDIGCompCnt = 0;
1140   if (MFI->hasWorkItemIDZ())
1141     TIDIGCompCnt = 2;
1142   else if (MFI->hasWorkItemIDY())
1143     TIDIGCompCnt = 1;
1144 
1145   ProgInfo.ComputePGMRSrc2 =
1146       S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
1147       S_00B84C_USER_SGPR(MFI->getNumUserSGPRs()) |
1148       // For AMDHSA, TRAP_HANDLER must be zero, as it is populated by the CP.
1149       S_00B84C_TRAP_HANDLER(STM.isAmdHsaOS() ? 0 : STM.isTrapHandlerEnabled()) |
1150       S_00B84C_TGID_X_EN(MFI->hasWorkGroupIDX()) |
1151       S_00B84C_TGID_Y_EN(MFI->hasWorkGroupIDY()) |
1152       S_00B84C_TGID_Z_EN(MFI->hasWorkGroupIDZ()) |
1153       S_00B84C_TG_SIZE_EN(MFI->hasWorkGroupInfo()) |
1154       S_00B84C_TIDIG_COMP_CNT(TIDIGCompCnt) |
1155       S_00B84C_EXCP_EN_MSB(0) |
1156       // For AMDHSA, LDS_SIZE must be zero, as it is populated by the CP.
1157       S_00B84C_LDS_SIZE(STM.isAmdHsaOS() ? 0 : ProgInfo.LDSBlocks) |
1158       S_00B84C_EXCP_EN(0);
1159 
1160   ProgInfo.Occupancy = STM.computeOccupancy(MF.getFunction(), ProgInfo.LDSSize,
1161                                             ProgInfo.NumSGPRsForWavesPerEU,
1162                                             ProgInfo.NumVGPRsForWavesPerEU);
1163 }
1164 
getRsrcReg(CallingConv::ID CallConv)1165 static unsigned getRsrcReg(CallingConv::ID CallConv) {
1166   switch (CallConv) {
1167   default: LLVM_FALLTHROUGH;
1168   case CallingConv::AMDGPU_CS: return R_00B848_COMPUTE_PGM_RSRC1;
1169   case CallingConv::AMDGPU_LS: return R_00B528_SPI_SHADER_PGM_RSRC1_LS;
1170   case CallingConv::AMDGPU_HS: return R_00B428_SPI_SHADER_PGM_RSRC1_HS;
1171   case CallingConv::AMDGPU_ES: return R_00B328_SPI_SHADER_PGM_RSRC1_ES;
1172   case CallingConv::AMDGPU_GS: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
1173   case CallingConv::AMDGPU_VS: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
1174   case CallingConv::AMDGPU_PS: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
1175   }
1176 }
1177 
EmitProgramInfoSI(const MachineFunction & MF,const SIProgramInfo & CurrentProgramInfo)1178 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
1179                                          const SIProgramInfo &CurrentProgramInfo) {
1180   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1181   unsigned RsrcReg = getRsrcReg(MF.getFunction().getCallingConv());
1182 
1183   if (AMDGPU::isCompute(MF.getFunction().getCallingConv())) {
1184     OutStreamer->emitInt32(R_00B848_COMPUTE_PGM_RSRC1);
1185 
1186     OutStreamer->emitInt32(CurrentProgramInfo.getComputePGMRSrc1());
1187 
1188     OutStreamer->emitInt32(R_00B84C_COMPUTE_PGM_RSRC2);
1189     OutStreamer->emitInt32(CurrentProgramInfo.ComputePGMRSrc2);
1190 
1191     OutStreamer->emitInt32(R_00B860_COMPUTE_TMPRING_SIZE);
1192     OutStreamer->emitInt32(S_00B860_WAVESIZE(CurrentProgramInfo.ScratchBlocks));
1193 
1194     // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
1195     // 0" comment but I don't see a corresponding field in the register spec.
1196   } else {
1197     OutStreamer->emitInt32(RsrcReg);
1198     OutStreamer->emitIntValue(S_00B028_VGPRS(CurrentProgramInfo.VGPRBlocks) |
1199                               S_00B028_SGPRS(CurrentProgramInfo.SGPRBlocks), 4);
1200     OutStreamer->emitInt32(R_0286E8_SPI_TMPRING_SIZE);
1201     OutStreamer->emitIntValue(
1202         S_0286E8_WAVESIZE(CurrentProgramInfo.ScratchBlocks), 4);
1203   }
1204 
1205   if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS) {
1206     OutStreamer->emitInt32(R_00B02C_SPI_SHADER_PGM_RSRC2_PS);
1207     OutStreamer->emitInt32(
1208         S_00B02C_EXTRA_LDS_SIZE(CurrentProgramInfo.LDSBlocks));
1209     OutStreamer->emitInt32(R_0286CC_SPI_PS_INPUT_ENA);
1210     OutStreamer->emitInt32(MFI->getPSInputEnable());
1211     OutStreamer->emitInt32(R_0286D0_SPI_PS_INPUT_ADDR);
1212     OutStreamer->emitInt32(MFI->getPSInputAddr());
1213   }
1214 
1215   OutStreamer->emitInt32(R_SPILLED_SGPRS);
1216   OutStreamer->emitInt32(MFI->getNumSpilledSGPRs());
1217   OutStreamer->emitInt32(R_SPILLED_VGPRS);
1218   OutStreamer->emitInt32(MFI->getNumSpilledVGPRs());
1219 }
1220 
1221 // This is the equivalent of EmitProgramInfoSI above, but for when the OS type
1222 // is AMDPAL.  It stores each compute/SPI register setting and other PAL
1223 // metadata items into the PALMD::Metadata, combining with any provided by the
1224 // frontend as LLVM metadata. Once all functions are written, the PAL metadata
1225 // is then written as a single block in the .note section.
EmitPALMetadata(const MachineFunction & MF,const SIProgramInfo & CurrentProgramInfo)1226 void AMDGPUAsmPrinter::EmitPALMetadata(const MachineFunction &MF,
1227        const SIProgramInfo &CurrentProgramInfo) {
1228   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1229   auto CC = MF.getFunction().getCallingConv();
1230   auto MD = getTargetStreamer()->getPALMetadata();
1231 
1232   MD->setEntryPoint(CC, MF.getFunction().getName());
1233   MD->setNumUsedVgprs(CC, CurrentProgramInfo.NumVGPRsForWavesPerEU);
1234   MD->setNumUsedSgprs(CC, CurrentProgramInfo.NumSGPRsForWavesPerEU);
1235   MD->setRsrc1(CC, CurrentProgramInfo.getPGMRSrc1(CC));
1236   if (AMDGPU::isCompute(CC)) {
1237     MD->setRsrc2(CC, CurrentProgramInfo.ComputePGMRSrc2);
1238   } else {
1239     if (CurrentProgramInfo.ScratchBlocks > 0)
1240       MD->setRsrc2(CC, S_00B84C_SCRATCH_EN(1));
1241   }
1242   // ScratchSize is in bytes, 16 aligned.
1243   MD->setScratchSize(CC, alignTo(CurrentProgramInfo.ScratchSize, 16));
1244   if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS) {
1245     MD->setRsrc2(CC, S_00B02C_EXTRA_LDS_SIZE(CurrentProgramInfo.LDSBlocks));
1246     MD->setSpiPsInputEna(MFI->getPSInputEnable());
1247     MD->setSpiPsInputAddr(MFI->getPSInputAddr());
1248   }
1249 
1250   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
1251   if (STM.isWave32())
1252     MD->setWave32(MF.getFunction().getCallingConv());
1253 }
1254 
1255 // This is supposed to be log2(Size)
getElementByteSizeValue(unsigned Size)1256 static amd_element_byte_size_t getElementByteSizeValue(unsigned Size) {
1257   switch (Size) {
1258   case 4:
1259     return AMD_ELEMENT_4_BYTES;
1260   case 8:
1261     return AMD_ELEMENT_8_BYTES;
1262   case 16:
1263     return AMD_ELEMENT_16_BYTES;
1264   default:
1265     llvm_unreachable("invalid private_element_size");
1266   }
1267 }
1268 
getAmdKernelCode(amd_kernel_code_t & Out,const SIProgramInfo & CurrentProgramInfo,const MachineFunction & MF) const1269 void AMDGPUAsmPrinter::getAmdKernelCode(amd_kernel_code_t &Out,
1270                                         const SIProgramInfo &CurrentProgramInfo,
1271                                         const MachineFunction &MF) const {
1272   const Function &F = MF.getFunction();
1273   assert(F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
1274          F.getCallingConv() == CallingConv::SPIR_KERNEL);
1275 
1276   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1277   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
1278 
1279   AMDGPU::initDefaultAMDKernelCodeT(Out, &STM);
1280 
1281   Out.compute_pgm_resource_registers =
1282       CurrentProgramInfo.getComputePGMRSrc1() |
1283       (CurrentProgramInfo.ComputePGMRSrc2 << 32);
1284   Out.code_properties |= AMD_CODE_PROPERTY_IS_PTR64;
1285 
1286   if (CurrentProgramInfo.DynamicCallStack)
1287     Out.code_properties |= AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK;
1288 
1289   AMD_HSA_BITS_SET(Out.code_properties,
1290                    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE,
1291                    getElementByteSizeValue(STM.getMaxPrivateElementSize()));
1292 
1293   if (MFI->hasPrivateSegmentBuffer()) {
1294     Out.code_properties |=
1295       AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER;
1296   }
1297 
1298   if (MFI->hasDispatchPtr())
1299     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
1300 
1301   if (MFI->hasQueuePtr())
1302     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR;
1303 
1304   if (MFI->hasKernargSegmentPtr())
1305     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR;
1306 
1307   if (MFI->hasDispatchID())
1308     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID;
1309 
1310   if (MFI->hasFlatScratchInit())
1311     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT;
1312 
1313   if (MFI->hasDispatchPtr())
1314     Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
1315 
1316   if (STM.isXNACKEnabled())
1317     Out.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED;
1318 
1319   Align MaxKernArgAlign;
1320   Out.kernarg_segment_byte_size = STM.getKernArgSegmentSize(F, MaxKernArgAlign);
1321   Out.wavefront_sgpr_count = CurrentProgramInfo.NumSGPR;
1322   Out.workitem_vgpr_count = CurrentProgramInfo.NumVGPR;
1323   Out.workitem_private_segment_byte_size = CurrentProgramInfo.ScratchSize;
1324   Out.workgroup_group_segment_byte_size = CurrentProgramInfo.LDSSize;
1325 
1326   // kernarg_segment_alignment is specified as log of the alignment.
1327   // The minimum alignment is 16.
1328   Out.kernarg_segment_alignment = Log2(std::max(Align(16), MaxKernArgAlign));
1329 }
1330 
PrintAsmOperand(const MachineInstr * MI,unsigned OpNo,const char * ExtraCode,raw_ostream & O)1331 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
1332                                        const char *ExtraCode, raw_ostream &O) {
1333   // First try the generic code, which knows about modifiers like 'c' and 'n'.
1334   if (!AsmPrinter::PrintAsmOperand(MI, OpNo, ExtraCode, O))
1335     return false;
1336 
1337   if (ExtraCode && ExtraCode[0]) {
1338     if (ExtraCode[1] != 0)
1339       return true; // Unknown modifier.
1340 
1341     switch (ExtraCode[0]) {
1342     case 'r':
1343       break;
1344     default:
1345       return true;
1346     }
1347   }
1348 
1349   // TODO: Should be able to support other operand types like globals.
1350   const MachineOperand &MO = MI->getOperand(OpNo);
1351   if (MO.isReg()) {
1352     AMDGPUInstPrinter::printRegOperand(MO.getReg(), O,
1353                                        *MF->getSubtarget().getRegisterInfo());
1354     return false;
1355   } else if (MO.isImm()) {
1356     int64_t Val = MO.getImm();
1357     if (AMDGPU::isInlinableIntLiteral(Val)) {
1358       O << Val;
1359     } else if (isUInt<16>(Val)) {
1360       O << format("0x%" PRIx16, static_cast<uint16_t>(Val));
1361     } else if (isUInt<32>(Val)) {
1362       O << format("0x%" PRIx32, static_cast<uint32_t>(Val));
1363     } else {
1364       O << format("0x%" PRIx64, static_cast<uint64_t>(Val));
1365     }
1366     return false;
1367   }
1368   return true;
1369 }
1370