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 "AMDGPUHSAMetadataStreamer.h" |
21 | #include "AMDGPUMCResourceInfo.h" |
22 | #include "AMDGPUResourceUsageAnalysis.h" |
23 | #include "GCNSubtarget.h" |
24 | #include "MCTargetDesc/AMDGPUInstPrinter.h" |
25 | #include "MCTargetDesc/AMDGPUMCExpr.h" |
26 | #include "MCTargetDesc/AMDGPUMCKernelDescriptor.h" |
27 | #include "MCTargetDesc/AMDGPUTargetStreamer.h" |
28 | #include "R600AsmPrinter.h" |
29 | #include "SIMachineFunctionInfo.h" |
30 | #include "TargetInfo/AMDGPUTargetInfo.h" |
31 | #include "Utils/AMDGPUBaseInfo.h" |
32 | #include "Utils/AMDKernelCodeTUtils.h" |
33 | #include "Utils/SIDefinesUtils.h" |
34 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
35 | #include "llvm/BinaryFormat/ELF.h" |
36 | #include "llvm/CodeGen/MachineFrameInfo.h" |
37 | #include "llvm/CodeGen/MachineModuleInfo.h" |
38 | #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" |
39 | #include "llvm/IR/DiagnosticInfo.h" |
40 | #include "llvm/MC/MCAssembler.h" |
41 | #include "llvm/MC/MCContext.h" |
42 | #include "llvm/MC/MCSectionELF.h" |
43 | #include "llvm/MC/MCStreamer.h" |
44 | #include "llvm/MC/TargetRegistry.h" |
45 | #include "llvm/Support/AMDHSAKernelDescriptor.h" |
46 | #include "llvm/Support/Compiler.h" |
47 | #include "llvm/Target/TargetLoweringObjectFile.h" |
48 | #include "llvm/Target/TargetMachine.h" |
49 | #include "llvm/TargetParser/TargetParser.h" |
50 | |
51 | using namespace llvm; |
52 | using namespace llvm::AMDGPU; |
53 | |
54 | // This should get the default rounding mode from the kernel. We just set the |
55 | // default here, but this could change if the OpenCL rounding mode pragmas are |
56 | // used. |
57 | // |
58 | // The denormal mode here should match what is reported by the OpenCL runtime |
59 | // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but |
60 | // can also be override to flush with the -cl-denorms-are-zero compiler flag. |
61 | // |
62 | // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double |
63 | // precision, and leaves single precision to flush all and does not report |
64 | // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports |
65 | // CL_FP_DENORM for both. |
66 | // |
67 | // FIXME: It seems some instructions do not support single precision denormals |
68 | // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32, |
69 | // and sin_f32, cos_f32 on most parts). |
70 | |
71 | // We want to use these instructions, and using fp32 denormals also causes |
72 | // instructions to run at the double precision rate for the device so it's |
73 | // probably best to just report no single precision denormals. |
74 | static uint32_t getFPMode(SIModeRegisterDefaults Mode) { |
75 | return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) | |
76 | FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) | |
77 | FP_DENORM_MODE_SP(Mode.fpDenormModeSPValue()) | |
78 | FP_DENORM_MODE_DP(Mode.fpDenormModeDPValue()); |
79 | } |
80 | |
81 | static AsmPrinter * |
82 | createAMDGPUAsmPrinterPass(TargetMachine &tm, |
83 | std::unique_ptr<MCStreamer> &&Streamer) { |
84 | return new AMDGPUAsmPrinter(tm, std::move(Streamer)); |
85 | } |
86 | |
87 | extern "C" LLVM_ABI LLVM_EXTERNAL_VISIBILITY void |
88 | LLVMInitializeAMDGPUAsmPrinter() { |
89 | TargetRegistry::RegisterAsmPrinter(T&: getTheR600Target(), |
90 | Fn: llvm::createR600AsmPrinterPass); |
91 | TargetRegistry::RegisterAsmPrinter(T&: getTheGCNTarget(), |
92 | Fn: createAMDGPUAsmPrinterPass); |
93 | } |
94 | |
95 | AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM, |
96 | std::unique_ptr<MCStreamer> Streamer) |
97 | : AsmPrinter(TM, std::move(Streamer)) { |
98 | assert(OutStreamer && "AsmPrinter constructed without streamer" ); |
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 | IsTargetStreamerInitialized = false; |
117 | } |
118 | |
119 | void AMDGPUAsmPrinter::initTargetStreamer(Module &M) { |
120 | IsTargetStreamerInitialized = true; |
121 | |
122 | // TODO: Which one is called first, emitStartOfAsmFile or |
123 | // emitFunctionBodyStart? |
124 | if (getTargetStreamer() && !getTargetStreamer()->getTargetID()) |
125 | initializeTargetID(M); |
126 | |
127 | if (TM.getTargetTriple().getOS() != Triple::AMDHSA && |
128 | TM.getTargetTriple().getOS() != Triple::AMDPAL) |
129 | return; |
130 | |
131 | getTargetStreamer()->EmitDirectiveAMDGCNTarget(); |
132 | |
133 | if (TM.getTargetTriple().getOS() == Triple::AMDHSA) { |
134 | getTargetStreamer()->EmitDirectiveAMDHSACodeObjectVersion( |
135 | COV: CodeObjectVersion); |
136 | HSAMetadataStream->begin(Mod: M, TargetID: *getTargetStreamer()->getTargetID()); |
137 | } |
138 | |
139 | if (TM.getTargetTriple().getOS() == Triple::AMDPAL) |
140 | getTargetStreamer()->getPALMetadata()->readFromIR(M); |
141 | } |
142 | |
143 | void AMDGPUAsmPrinter::emitEndOfAsmFile(Module &M) { |
144 | // Init target streamer if it has not yet happened |
145 | if (!IsTargetStreamerInitialized) |
146 | initTargetStreamer(M); |
147 | |
148 | if (TM.getTargetTriple().getOS() != Triple::AMDHSA) |
149 | getTargetStreamer()->EmitISAVersion(); |
150 | |
151 | // Emit HSA Metadata (NT_AMD_AMDGPU_HSA_METADATA). |
152 | // Emit HSA Metadata (NT_AMD_HSA_METADATA). |
153 | if (TM.getTargetTriple().getOS() == Triple::AMDHSA) { |
154 | HSAMetadataStream->end(); |
155 | bool Success = HSAMetadataStream->emitTo(TargetStreamer&: *getTargetStreamer()); |
156 | (void)Success; |
157 | assert(Success && "Malformed HSA Metadata" ); |
158 | } |
159 | } |
160 | |
161 | void AMDGPUAsmPrinter::emitFunctionBodyStart() { |
162 | const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>(); |
163 | const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>(); |
164 | const Function &F = MF->getFunction(); |
165 | |
166 | // TODO: We're checking this late, would be nice to check it earlier. |
167 | if (STM.requiresCodeObjectV6() && CodeObjectVersion < AMDGPU::AMDHSA_COV6) { |
168 | reportFatalUsageError( |
169 | reason: STM.getCPU() + " is only available on code object version 6 or better" ); |
170 | } |
171 | |
172 | // TODO: Which one is called first, emitStartOfAsmFile or |
173 | // emitFunctionBodyStart? |
174 | if (!getTargetStreamer()->getTargetID()) |
175 | initializeTargetID(M: *F.getParent()); |
176 | |
177 | const auto &FunctionTargetID = STM.getTargetID(); |
178 | // Make sure function's xnack settings are compatible with module's |
179 | // xnack settings. |
180 | if (FunctionTargetID.isXnackSupported() && |
181 | FunctionTargetID.getXnackSetting() != IsaInfo::TargetIDSetting::Any && |
182 | FunctionTargetID.getXnackSetting() != getTargetStreamer()->getTargetID()->getXnackSetting()) { |
183 | OutContext.reportError(L: {}, Msg: "xnack setting of '" + Twine(MF->getName()) + |
184 | "' function does not match module xnack setting" ); |
185 | return; |
186 | } |
187 | // Make sure function's sramecc settings are compatible with module's |
188 | // sramecc settings. |
189 | if (FunctionTargetID.isSramEccSupported() && |
190 | FunctionTargetID.getSramEccSetting() != IsaInfo::TargetIDSetting::Any && |
191 | FunctionTargetID.getSramEccSetting() != getTargetStreamer()->getTargetID()->getSramEccSetting()) { |
192 | OutContext.reportError(L: {}, Msg: "sramecc setting of '" + Twine(MF->getName()) + |
193 | "' function does not match module sramecc setting" ); |
194 | return; |
195 | } |
196 | |
197 | if (!MFI.isEntryFunction()) |
198 | return; |
199 | |
200 | if (STM.isMesaKernel(F) && |
201 | (F.getCallingConv() == CallingConv::AMDGPU_KERNEL || |
202 | F.getCallingConv() == CallingConv::SPIR_KERNEL)) { |
203 | AMDGPUMCKernelCodeT KernelCode; |
204 | getAmdKernelCode(Out&: KernelCode, KernelInfo: CurrentProgramInfo, MF: *MF); |
205 | KernelCode.validate(STI: &STM, Ctx&: MF->getContext()); |
206 | getTargetStreamer()->EmitAMDKernelCodeT(Header&: KernelCode); |
207 | } |
208 | |
209 | if (STM.isAmdHsaOS()) |
210 | HSAMetadataStream->emitKernel(MF: *MF, ProgramInfo: CurrentProgramInfo); |
211 | } |
212 | |
213 | void AMDGPUAsmPrinter::emitFunctionBodyEnd() { |
214 | const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>(); |
215 | if (!MFI.isEntryFunction()) |
216 | return; |
217 | |
218 | if (TM.getTargetTriple().getOS() != Triple::AMDHSA) |
219 | return; |
220 | |
221 | auto &Streamer = getTargetStreamer()->getStreamer(); |
222 | auto &Context = Streamer.getContext(); |
223 | auto &ObjectFileInfo = *Context.getObjectFileInfo(); |
224 | auto &ReadOnlySection = *ObjectFileInfo.getReadOnlySection(); |
225 | |
226 | Streamer.pushSection(); |
227 | Streamer.switchSection(Section: &ReadOnlySection); |
228 | |
229 | // CP microcode requires the kernel descriptor to be allocated on 64 byte |
230 | // alignment. |
231 | Streamer.emitValueToAlignment(Alignment: Align(64), Value: 0, ValueSize: 1, MaxBytesToEmit: 0); |
232 | ReadOnlySection.ensureMinAlignment(MinAlignment: Align(64)); |
233 | |
234 | const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>(); |
235 | |
236 | SmallString<128> KernelName; |
237 | getNameWithPrefix(Name&: KernelName, GV: &MF->getFunction()); |
238 | getTargetStreamer()->EmitAmdhsaKernelDescriptor( |
239 | STI: STM, KernelName, KernelDescriptor: getAmdhsaKernelDescriptor(MF: *MF, PI: CurrentProgramInfo), |
240 | NextVGPR: CurrentProgramInfo.NumVGPRsForWavesPerEU, |
241 | NextSGPR: MCBinaryExpr::createSub( |
242 | LHS: CurrentProgramInfo.NumSGPRsForWavesPerEU, |
243 | RHS: AMDGPUMCExpr::createExtraSGPRs( |
244 | VCCUsed: CurrentProgramInfo.VCCUsed, FlatScrUsed: CurrentProgramInfo.FlatUsed, |
245 | XNACKUsed: getTargetStreamer()->getTargetID()->isXnackOnOrAny(), Ctx&: Context), |
246 | Ctx&: Context), |
247 | ReserveVCC: CurrentProgramInfo.VCCUsed, ReserveFlatScr: CurrentProgramInfo.FlatUsed); |
248 | |
249 | Streamer.popSection(); |
250 | } |
251 | |
252 | void AMDGPUAsmPrinter::emitImplicitDef(const MachineInstr *MI) const { |
253 | Register RegNo = MI->getOperand(i: 0).getReg(); |
254 | |
255 | SmallString<128> Str; |
256 | raw_svector_ostream OS(Str); |
257 | OS << "implicit-def: " |
258 | << printReg(Reg: RegNo, TRI: MF->getSubtarget().getRegisterInfo()); |
259 | |
260 | if (MI->getAsmPrinterFlags() & AMDGPU::SGPR_SPILL) |
261 | OS << " : SGPR spill to VGPR lane" ; |
262 | |
263 | OutStreamer->AddComment(T: OS.str()); |
264 | OutStreamer->addBlankLine(); |
265 | } |
266 | |
267 | void AMDGPUAsmPrinter::emitFunctionEntryLabel() { |
268 | if (TM.getTargetTriple().getOS() == Triple::AMDHSA) { |
269 | AsmPrinter::emitFunctionEntryLabel(); |
270 | return; |
271 | } |
272 | |
273 | const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
274 | const GCNSubtarget &STM = MF->getSubtarget<GCNSubtarget>(); |
275 | if (MFI->isEntryFunction() && STM.isAmdHsaOrMesa(F: MF->getFunction())) { |
276 | SmallString<128> SymbolName; |
277 | getNameWithPrefix(Name&: SymbolName, GV: &MF->getFunction()), |
278 | getTargetStreamer()->EmitAMDGPUSymbolType( |
279 | SymbolName, Type: ELF::STT_AMDGPU_HSA_KERNEL); |
280 | } |
281 | if (DumpCodeInstEmitter) { |
282 | // Disassemble function name label to text. |
283 | DisasmLines.push_back(x: MF->getName().str() + ":" ); |
284 | DisasmLineMaxLen = std::max(a: DisasmLineMaxLen, b: DisasmLines.back().size()); |
285 | HexLines.emplace_back(args: "" ); |
286 | } |
287 | |
288 | AsmPrinter::emitFunctionEntryLabel(); |
289 | } |
290 | |
291 | void AMDGPUAsmPrinter::emitBasicBlockStart(const MachineBasicBlock &MBB) { |
292 | if (DumpCodeInstEmitter && !isBlockOnlyReachableByFallthrough(MBB: &MBB)) { |
293 | // Write a line for the basic block label if it is not only fallthrough. |
294 | DisasmLines.push_back( |
295 | x: (Twine("BB" ) + Twine(getFunctionNumber()) |
296 | + "_" + Twine(MBB.getNumber()) + ":" ).str()); |
297 | DisasmLineMaxLen = std::max(a: DisasmLineMaxLen, b: DisasmLines.back().size()); |
298 | HexLines.emplace_back(args: "" ); |
299 | } |
300 | AsmPrinter::emitBasicBlockStart(MBB); |
301 | } |
302 | |
303 | void AMDGPUAsmPrinter::emitGlobalVariable(const GlobalVariable *GV) { |
304 | if (GV->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) { |
305 | if (GV->hasInitializer() && !isa<UndefValue>(Val: GV->getInitializer())) { |
306 | OutContext.reportError(L: {}, |
307 | Msg: Twine(GV->getName()) + |
308 | ": unsupported initializer for address space" ); |
309 | return; |
310 | } |
311 | |
312 | // LDS variables aren't emitted in HSA or PAL yet. |
313 | const Triple::OSType OS = TM.getTargetTriple().getOS(); |
314 | if (OS == Triple::AMDHSA || OS == Triple::AMDPAL) |
315 | return; |
316 | |
317 | MCSymbol *GVSym = getSymbol(GV); |
318 | |
319 | GVSym->redefineIfPossible(); |
320 | if (GVSym->isDefined() || GVSym->isVariable()) |
321 | report_fatal_error(reason: "symbol '" + Twine(GVSym->getName()) + |
322 | "' is already defined" ); |
323 | |
324 | const DataLayout &DL = GV->getDataLayout(); |
325 | uint64_t Size = DL.getTypeAllocSize(Ty: GV->getValueType()); |
326 | Align Alignment = GV->getAlign().value_or(u: Align(4)); |
327 | |
328 | emitVisibility(Sym: GVSym, Visibility: GV->getVisibility(), IsDefinition: !GV->isDeclaration()); |
329 | emitLinkage(GV, GVSym); |
330 | auto *TS = getTargetStreamer(); |
331 | TS->emitAMDGPULDS(Symbol: GVSym, Size, Alignment); |
332 | return; |
333 | } |
334 | |
335 | AsmPrinter::emitGlobalVariable(GV); |
336 | } |
337 | |
338 | bool AMDGPUAsmPrinter::doInitialization(Module &M) { |
339 | CodeObjectVersion = AMDGPU::getAMDHSACodeObjectVersion(M); |
340 | |
341 | if (TM.getTargetTriple().getOS() == Triple::AMDHSA) { |
342 | switch (CodeObjectVersion) { |
343 | case AMDGPU::AMDHSA_COV4: |
344 | HSAMetadataStream = std::make_unique<HSAMD::MetadataStreamerMsgPackV4>(); |
345 | break; |
346 | case AMDGPU::AMDHSA_COV5: |
347 | HSAMetadataStream = std::make_unique<HSAMD::MetadataStreamerMsgPackV5>(); |
348 | break; |
349 | case AMDGPU::AMDHSA_COV6: |
350 | HSAMetadataStream = std::make_unique<HSAMD::MetadataStreamerMsgPackV6>(); |
351 | break; |
352 | default: |
353 | reportFatalUsageError(reason: "unsupported code object version" ); |
354 | } |
355 | } |
356 | |
357 | return AsmPrinter::doInitialization(M); |
358 | } |
359 | |
360 | void AMDGPUAsmPrinter::validateMCResourceInfo(Function &F) { |
361 | if (F.isDeclaration() || !AMDGPU::isModuleEntryFunctionCC(CC: F.getCallingConv())) |
362 | return; |
363 | |
364 | using RIK = MCResourceInfo::ResourceInfoKind; |
365 | const GCNSubtarget &STM = TM.getSubtarget<GCNSubtarget>(F); |
366 | MCSymbol *FnSym = TM.getSymbol(GV: &F); |
367 | bool IsLocal = F.hasLocalLinkage(); |
368 | |
369 | auto TryGetMCExprValue = [](const MCExpr *Value, uint64_t &Res) -> bool { |
370 | int64_t Val; |
371 | if (Value->evaluateAsAbsolute(Res&: Val)) { |
372 | Res = Val; |
373 | return true; |
374 | } |
375 | return false; |
376 | }; |
377 | |
378 | const uint64_t MaxScratchPerWorkitem = |
379 | STM.getMaxWaveScratchSize() / STM.getWavefrontSize(); |
380 | MCSymbol *ScratchSizeSymbol = RI.getSymbol( |
381 | FuncName: FnSym->getName(), RIK: RIK::RIK_PrivateSegSize, OutContext, IsLocal); |
382 | uint64_t ScratchSize; |
383 | if (ScratchSizeSymbol->isVariable() && |
384 | TryGetMCExprValue(ScratchSizeSymbol->getVariableValue(), ScratchSize) && |
385 | ScratchSize > MaxScratchPerWorkitem) { |
386 | DiagnosticInfoStackSize DiagStackSize(F, ScratchSize, MaxScratchPerWorkitem, |
387 | DS_Error); |
388 | F.getContext().diagnose(DI: DiagStackSize); |
389 | } |
390 | |
391 | // Validate addressable scalar registers (i.e., prior to added implicit |
392 | // SGPRs). |
393 | MCSymbol *NumSGPRSymbol = |
394 | RI.getSymbol(FuncName: FnSym->getName(), RIK: RIK::RIK_NumSGPR, OutContext, IsLocal); |
395 | if (STM.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS && |
396 | !STM.hasSGPRInitBug()) { |
397 | unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs(); |
398 | uint64_t NumSgpr; |
399 | if (NumSGPRSymbol->isVariable() && |
400 | TryGetMCExprValue(NumSGPRSymbol->getVariableValue(), NumSgpr) && |
401 | NumSgpr > MaxAddressableNumSGPRs) { |
402 | DiagnosticInfoResourceLimit Diag(F, "addressable scalar registers" , |
403 | NumSgpr, MaxAddressableNumSGPRs, |
404 | DS_Error, DK_ResourceLimit); |
405 | F.getContext().diagnose(DI: Diag); |
406 | return; |
407 | } |
408 | } |
409 | |
410 | MCSymbol *VCCUsedSymbol = |
411 | RI.getSymbol(FuncName: FnSym->getName(), RIK: RIK::RIK_UsesVCC, OutContext, IsLocal); |
412 | MCSymbol *FlatUsedSymbol = RI.getSymbol( |
413 | FuncName: FnSym->getName(), RIK: RIK::RIK_UsesFlatScratch, OutContext, IsLocal); |
414 | uint64_t VCCUsed, FlatUsed, NumSgpr; |
415 | |
416 | if (NumSGPRSymbol->isVariable() && VCCUsedSymbol->isVariable() && |
417 | FlatUsedSymbol->isVariable() && |
418 | TryGetMCExprValue(NumSGPRSymbol->getVariableValue(), NumSgpr) && |
419 | TryGetMCExprValue(VCCUsedSymbol->getVariableValue(), VCCUsed) && |
420 | TryGetMCExprValue(FlatUsedSymbol->getVariableValue(), FlatUsed)) { |
421 | |
422 | // Recomputes NumSgprs + implicit SGPRs but all symbols should now be |
423 | // resolvable. |
424 | NumSgpr += IsaInfo::getNumExtraSGPRs( |
425 | STI: &STM, VCCUsed, FlatScrUsed: FlatUsed, |
426 | XNACKUsed: getTargetStreamer()->getTargetID()->isXnackOnOrAny()); |
427 | if (STM.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS || |
428 | STM.hasSGPRInitBug()) { |
429 | unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs(); |
430 | if (NumSgpr > MaxAddressableNumSGPRs) { |
431 | DiagnosticInfoResourceLimit Diag(F, "scalar registers" , NumSgpr, |
432 | MaxAddressableNumSGPRs, DS_Error, |
433 | DK_ResourceLimit); |
434 | F.getContext().diagnose(DI: Diag); |
435 | return; |
436 | } |
437 | } |
438 | |
439 | MCSymbol *NumVgprSymbol = |
440 | RI.getSymbol(FuncName: FnSym->getName(), RIK: RIK::RIK_NumVGPR, OutContext, IsLocal); |
441 | MCSymbol *NumAgprSymbol = |
442 | RI.getSymbol(FuncName: FnSym->getName(), RIK: RIK::RIK_NumAGPR, OutContext, IsLocal); |
443 | uint64_t NumVgpr, NumAgpr; |
444 | |
445 | MachineModuleInfo &MMI = |
446 | getAnalysis<MachineModuleInfoWrapperPass>().getMMI(); |
447 | MachineFunction *MF = MMI.getMachineFunction(F); |
448 | if (MF && NumVgprSymbol->isVariable() && NumAgprSymbol->isVariable() && |
449 | TryGetMCExprValue(NumVgprSymbol->getVariableValue(), NumVgpr) && |
450 | TryGetMCExprValue(NumAgprSymbol->getVariableValue(), NumAgpr)) { |
451 | const SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>(); |
452 | unsigned MaxWaves = MFI.getMaxWavesPerEU(); |
453 | uint64_t TotalNumVgpr = |
454 | getTotalNumVGPRs(has90AInsts: STM.hasGFX90AInsts(), ArgNumAGPR: NumAgpr, ArgNumVGPR: NumVgpr); |
455 | uint64_t NumVGPRsForWavesPerEU = |
456 | std::max(l: {TotalNumVgpr, (uint64_t)1, |
457 | (uint64_t)STM.getMinNumVGPRs( |
458 | WavesPerEU: MaxWaves, DynamicVGPRBlockSize: MFI.getDynamicVGPRBlockSize())}); |
459 | uint64_t NumSGPRsForWavesPerEU = std::max( |
460 | l: {NumSgpr, (uint64_t)1, (uint64_t)STM.getMinNumSGPRs(WavesPerEU: MaxWaves)}); |
461 | const MCExpr *OccupancyExpr = AMDGPUMCExpr::createOccupancy( |
462 | InitOcc: STM.getOccupancyWithWorkGroupSizes(MF: *MF).second, |
463 | NumSGPRs: MCConstantExpr::create(Value: NumSGPRsForWavesPerEU, Ctx&: OutContext), |
464 | NumVGPRs: MCConstantExpr::create(Value: NumVGPRsForWavesPerEU, Ctx&: OutContext), |
465 | DynamicVGPRBlockSize: MFI.getDynamicVGPRBlockSize(), STM, Ctx&: OutContext); |
466 | uint64_t Occupancy; |
467 | |
468 | const auto [MinWEU, MaxWEU] = AMDGPU::getIntegerPairAttribute( |
469 | F, Name: "amdgpu-waves-per-eu" , Default: {0, 0}, OnlyFirstRequired: true); |
470 | |
471 | if (TryGetMCExprValue(OccupancyExpr, Occupancy) && Occupancy < MinWEU) { |
472 | DiagnosticInfoOptimizationFailure Diag( |
473 | F, F.getSubprogram(), |
474 | "failed to meet occupancy target given by 'amdgpu-waves-per-eu' in " |
475 | "'" + |
476 | F.getName() + "': desired occupancy was " + Twine(MinWEU) + |
477 | ", final occupancy is " + Twine(Occupancy)); |
478 | F.getContext().diagnose(DI: Diag); |
479 | return; |
480 | } |
481 | } |
482 | } |
483 | } |
484 | |
485 | bool AMDGPUAsmPrinter::doFinalization(Module &M) { |
486 | // Pad with s_code_end to help tools and guard against instruction prefetch |
487 | // causing stale data in caches. Arguably this should be done by the linker, |
488 | // which is why this isn't done for Mesa. |
489 | const MCSubtargetInfo &STI = *getGlobalSTI(); |
490 | if ((AMDGPU::isGFX10Plus(STI) || AMDGPU::isGFX90A(STI)) && |
491 | (STI.getTargetTriple().getOS() == Triple::AMDHSA || |
492 | STI.getTargetTriple().getOS() == Triple::AMDPAL)) { |
493 | OutStreamer->switchSection(Section: getObjFileLowering().getTextSection()); |
494 | getTargetStreamer()->EmitCodeEnd(STI); |
495 | } |
496 | |
497 | // Assign expressions which can only be resolved when all other functions are |
498 | // known. |
499 | RI.finalize(OutContext); |
500 | |
501 | // Switch section and emit all GPR maximums within the processed module. |
502 | OutStreamer->pushSection(); |
503 | MCSectionELF *MaxGPRSection = |
504 | OutContext.getELFSection(Section: ".AMDGPU.gpr_maximums" , Type: ELF::SHT_PROGBITS, Flags: 0); |
505 | OutStreamer->switchSection(Section: MaxGPRSection); |
506 | getTargetStreamer()->EmitMCResourceMaximums(MaxVGPR: RI.getMaxVGPRSymbol(OutContext), |
507 | MaxAGPR: RI.getMaxAGPRSymbol(OutContext), |
508 | MaxSGPR: RI.getMaxSGPRSymbol(OutContext)); |
509 | OutStreamer->popSection(); |
510 | |
511 | for (Function &F : M.functions()) |
512 | validateMCResourceInfo(F); |
513 | |
514 | RI.reset(); |
515 | |
516 | return AsmPrinter::doFinalization(M); |
517 | } |
518 | |
519 | SmallString<128> AMDGPUAsmPrinter::getMCExprStr(const MCExpr *Value) { |
520 | SmallString<128> Str; |
521 | raw_svector_ostream OSS(Str); |
522 | auto &Streamer = getTargetStreamer()->getStreamer(); |
523 | auto &Context = Streamer.getContext(); |
524 | const MCExpr *New = foldAMDGPUMCExpr(Expr: Value, Ctx&: Context); |
525 | printAMDGPUMCExpr(Expr: New, OS&: OSS, MAI); |
526 | return Str; |
527 | } |
528 | |
529 | // Print comments that apply to both callable functions and entry points. |
530 | void AMDGPUAsmPrinter::( |
531 | const MCExpr *NumVGPR, const MCExpr *NumAGPR, const MCExpr *TotalNumVGPR, |
532 | const MCExpr *NumSGPR, const MCExpr *ScratchSize, uint64_t CodeSize, |
533 | const AMDGPUMachineFunction *MFI) { |
534 | OutStreamer->emitRawComment(T: " codeLenInByte = " + Twine(CodeSize), TabPrefix: false); |
535 | OutStreamer->emitRawComment(T: " TotalNumSgprs: " + getMCExprStr(Value: NumSGPR), |
536 | TabPrefix: false); |
537 | OutStreamer->emitRawComment(T: " NumVgprs: " + getMCExprStr(Value: NumVGPR), TabPrefix: false); |
538 | if (NumAGPR && TotalNumVGPR) { |
539 | OutStreamer->emitRawComment(T: " NumAgprs: " + getMCExprStr(Value: NumAGPR), TabPrefix: false); |
540 | OutStreamer->emitRawComment(T: " TotalNumVgprs: " + getMCExprStr(Value: TotalNumVGPR), |
541 | TabPrefix: false); |
542 | } |
543 | OutStreamer->emitRawComment(T: " ScratchSize: " + getMCExprStr(Value: ScratchSize), |
544 | TabPrefix: false); |
545 | OutStreamer->emitRawComment(T: " MemoryBound: " + Twine(MFI->isMemoryBound()), |
546 | TabPrefix: false); |
547 | } |
548 | |
549 | const MCExpr *AMDGPUAsmPrinter::getAmdhsaKernelCodeProperties( |
550 | const MachineFunction &MF) const { |
551 | const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>(); |
552 | MCContext &Ctx = MF.getContext(); |
553 | uint16_t KernelCodeProperties = 0; |
554 | const GCNUserSGPRUsageInfo &UserSGPRInfo = MFI.getUserSGPRInfo(); |
555 | |
556 | if (UserSGPRInfo.hasPrivateSegmentBuffer()) { |
557 | KernelCodeProperties |= |
558 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER; |
559 | } |
560 | if (UserSGPRInfo.hasDispatchPtr()) { |
561 | KernelCodeProperties |= |
562 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR; |
563 | } |
564 | if (UserSGPRInfo.hasQueuePtr()) { |
565 | KernelCodeProperties |= |
566 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR; |
567 | } |
568 | if (UserSGPRInfo.hasKernargSegmentPtr()) { |
569 | KernelCodeProperties |= |
570 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR; |
571 | } |
572 | if (UserSGPRInfo.hasDispatchID()) { |
573 | KernelCodeProperties |= |
574 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID; |
575 | } |
576 | if (UserSGPRInfo.hasFlatScratchInit()) { |
577 | KernelCodeProperties |= |
578 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT; |
579 | } |
580 | if (UserSGPRInfo.hasPrivateSegmentSize()) { |
581 | KernelCodeProperties |= |
582 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE; |
583 | } |
584 | if (MF.getSubtarget<GCNSubtarget>().isWave32()) { |
585 | KernelCodeProperties |= |
586 | amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32; |
587 | } |
588 | |
589 | // CurrentProgramInfo.DynamicCallStack is a MCExpr and could be |
590 | // un-evaluatable at this point so it cannot be conditionally checked here. |
591 | // Instead, we'll directly shift the possibly unknown MCExpr into its place |
592 | // and bitwise-or it into KernelCodeProperties. |
593 | const MCExpr *KernelCodePropExpr = |
594 | MCConstantExpr::create(Value: KernelCodeProperties, Ctx); |
595 | const MCExpr *OrValue = MCConstantExpr::create( |
596 | Value: amdhsa::KERNEL_CODE_PROPERTY_USES_DYNAMIC_STACK_SHIFT, Ctx); |
597 | OrValue = MCBinaryExpr::createShl(LHS: CurrentProgramInfo.DynamicCallStack, |
598 | RHS: OrValue, Ctx); |
599 | KernelCodePropExpr = MCBinaryExpr::createOr(LHS: KernelCodePropExpr, RHS: OrValue, Ctx); |
600 | |
601 | return KernelCodePropExpr; |
602 | } |
603 | |
604 | MCKernelDescriptor |
605 | AMDGPUAsmPrinter::getAmdhsaKernelDescriptor(const MachineFunction &MF, |
606 | const SIProgramInfo &PI) const { |
607 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
608 | const Function &F = MF.getFunction(); |
609 | const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); |
610 | MCContext &Ctx = MF.getContext(); |
611 | |
612 | MCKernelDescriptor KernelDescriptor; |
613 | |
614 | KernelDescriptor.group_segment_fixed_size = |
615 | MCConstantExpr::create(Value: PI.LDSSize, Ctx); |
616 | KernelDescriptor.private_segment_fixed_size = PI.ScratchSize; |
617 | |
618 | Align MaxKernArgAlign; |
619 | KernelDescriptor.kernarg_size = MCConstantExpr::create( |
620 | Value: STM.getKernArgSegmentSize(F, MaxAlign&: MaxKernArgAlign), Ctx); |
621 | |
622 | KernelDescriptor.compute_pgm_rsrc1 = PI.getComputePGMRSrc1(ST: STM, Ctx); |
623 | KernelDescriptor.compute_pgm_rsrc2 = PI.getComputePGMRSrc2(Ctx); |
624 | KernelDescriptor.kernel_code_properties = getAmdhsaKernelCodeProperties(MF); |
625 | |
626 | int64_t PGRM_Rsrc3 = 1; |
627 | bool EvaluatableRsrc3 = |
628 | CurrentProgramInfo.ComputePGMRSrc3->evaluateAsAbsolute(Res&: PGRM_Rsrc3); |
629 | (void)PGRM_Rsrc3; |
630 | (void)EvaluatableRsrc3; |
631 | assert(STM.getGeneration() >= AMDGPUSubtarget::GFX10 || |
632 | STM.hasGFX90AInsts() || !EvaluatableRsrc3 || |
633 | static_cast<uint64_t>(PGRM_Rsrc3) == 0); |
634 | KernelDescriptor.compute_pgm_rsrc3 = CurrentProgramInfo.ComputePGMRSrc3; |
635 | |
636 | KernelDescriptor.kernarg_preload = MCConstantExpr::create( |
637 | Value: AMDGPU::hasKernargPreload(STI: STM) ? Info->getNumKernargPreloadedSGPRs() : 0, |
638 | Ctx); |
639 | |
640 | return KernelDescriptor; |
641 | } |
642 | |
643 | bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) { |
644 | // Init target streamer lazily on the first function so that previous passes |
645 | // can set metadata. |
646 | if (!IsTargetStreamerInitialized) |
647 | initTargetStreamer(M&: *MF.getFunction().getParent()); |
648 | |
649 | ResourceUsage = &getAnalysis<AMDGPUResourceUsageAnalysis>(); |
650 | CurrentProgramInfo.reset(MF); |
651 | |
652 | const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>(); |
653 | MCContext &Ctx = MF.getContext(); |
654 | |
655 | // The starting address of all shader programs must be 256 bytes aligned. |
656 | // Regular functions just need the basic required instruction alignment. |
657 | MF.setAlignment(MFI->isEntryFunction() ? Align(256) : Align(4)); |
658 | |
659 | SetupMachineFunction(MF); |
660 | |
661 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
662 | MCContext &Context = getObjFileLowering().getContext(); |
663 | bool IsLocal = MF.getFunction().hasLocalLinkage(); |
664 | // FIXME: This should be an explicit check for Mesa. |
665 | if (!STM.isAmdHsaOS() && !STM.isAmdPalOS()) { |
666 | MCSectionELF *ConfigSection = |
667 | Context.getELFSection(Section: ".AMDGPU.config" , Type: ELF::SHT_PROGBITS, Flags: 0); |
668 | OutStreamer->switchSection(Section: ConfigSection); |
669 | } |
670 | |
671 | const AMDGPUResourceUsageAnalysis::SIFunctionResourceInfo &Info = |
672 | ResourceUsage->getResourceInfo(); |
673 | RI.gatherResourceInfo(MF, FRI: Info, OutContext); |
674 | |
675 | if (MFI->isModuleEntryFunction()) { |
676 | getSIProgramInfo(Out&: CurrentProgramInfo, MF); |
677 | } |
678 | |
679 | if (STM.isAmdPalOS()) { |
680 | if (MFI->isEntryFunction()) |
681 | EmitPALMetadata(MF, KernelInfo: CurrentProgramInfo); |
682 | else if (MFI->isModuleEntryFunction()) |
683 | emitPALFunctionMetadata(MF); |
684 | } else if (!STM.isAmdHsaOS()) { |
685 | EmitProgramInfoSI(MF, KernelInfo: CurrentProgramInfo); |
686 | } |
687 | |
688 | DumpCodeInstEmitter = nullptr; |
689 | if (STM.dumpCode()) { |
690 | // For -dumpcode, get the assembler out of the streamer. This only works |
691 | // with -filetype=obj. |
692 | MCAssembler *Assembler = OutStreamer->getAssemblerPtr(); |
693 | if (Assembler) |
694 | DumpCodeInstEmitter = Assembler->getEmitterPtr(); |
695 | } |
696 | |
697 | DisasmLines.clear(); |
698 | HexLines.clear(); |
699 | DisasmLineMaxLen = 0; |
700 | |
701 | emitFunctionBody(); |
702 | |
703 | emitResourceUsageRemarks(MF, CurrentProgramInfo, isModuleEntryFunction: MFI->isModuleEntryFunction(), |
704 | hasMAIInsts: STM.hasMAIInsts()); |
705 | |
706 | { |
707 | using RIK = MCResourceInfo::ResourceInfoKind; |
708 | getTargetStreamer()->EmitMCResourceInfo( |
709 | NumVGPR: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_NumVGPR, OutContext, |
710 | IsLocal), |
711 | NumAGPR: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_NumAGPR, OutContext, |
712 | IsLocal), |
713 | NumExplicitSGPR: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_NumSGPR, OutContext, |
714 | IsLocal), |
715 | PrivateSegmentSize: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_PrivateSegSize, |
716 | OutContext, IsLocal), |
717 | UsesVCC: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_UsesVCC, OutContext, |
718 | IsLocal), |
719 | UsesFlatScratch: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_UsesFlatScratch, |
720 | OutContext, IsLocal), |
721 | HasDynamicallySizedStack: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_HasDynSizedStack, |
722 | OutContext, IsLocal), |
723 | HasRecursion: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_HasRecursion, OutContext, |
724 | IsLocal), |
725 | HasIndirectCall: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_HasIndirectCall, |
726 | OutContext, IsLocal)); |
727 | } |
728 | |
729 | if (isVerbose()) { |
730 | MCSectionELF * = |
731 | Context.getELFSection(Section: ".AMDGPU.csdata" , Type: ELF::SHT_PROGBITS, Flags: 0); |
732 | OutStreamer->switchSection(Section: CommentSection); |
733 | |
734 | if (!MFI->isEntryFunction()) { |
735 | using RIK = MCResourceInfo::ResourceInfoKind; |
736 | OutStreamer->emitRawComment(T: " Function info:" , TabPrefix: false); |
737 | |
738 | emitCommonFunctionComments( |
739 | NumVGPR: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_NumVGPR, OutContext, |
740 | IsLocal) |
741 | ->getVariableValue(), |
742 | NumAGPR: STM.hasMAIInsts() |
743 | ? RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_NumAGPR, |
744 | OutContext, IsLocal) |
745 | ->getVariableValue() |
746 | : nullptr, |
747 | TotalNumVGPR: RI.createTotalNumVGPRs(MF, Ctx), |
748 | NumSGPR: RI.createTotalNumSGPRs( |
749 | MF, |
750 | hasXnack: MF.getSubtarget<GCNSubtarget>().getTargetID().isXnackOnOrAny(), |
751 | Ctx), |
752 | ScratchSize: RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK: RIK::RIK_PrivateSegSize, |
753 | OutContext, IsLocal) |
754 | ->getVariableValue(), |
755 | CodeSize: CurrentProgramInfo.getFunctionCodeSize(MF), MFI); |
756 | return false; |
757 | } |
758 | |
759 | OutStreamer->emitRawComment(T: " Kernel info:" , TabPrefix: false); |
760 | emitCommonFunctionComments( |
761 | NumVGPR: CurrentProgramInfo.NumArchVGPR, |
762 | NumAGPR: STM.hasMAIInsts() ? CurrentProgramInfo.NumAccVGPR : nullptr, |
763 | TotalNumVGPR: CurrentProgramInfo.NumVGPR, NumSGPR: CurrentProgramInfo.NumSGPR, |
764 | ScratchSize: CurrentProgramInfo.ScratchSize, |
765 | CodeSize: CurrentProgramInfo.getFunctionCodeSize(MF), MFI); |
766 | |
767 | OutStreamer->emitRawComment( |
768 | T: " FloatMode: " + Twine(CurrentProgramInfo.FloatMode), TabPrefix: false); |
769 | OutStreamer->emitRawComment( |
770 | T: " IeeeMode: " + Twine(CurrentProgramInfo.IEEEMode), TabPrefix: false); |
771 | OutStreamer->emitRawComment( |
772 | T: " LDSByteSize: " + Twine(CurrentProgramInfo.LDSSize) + |
773 | " bytes/workgroup (compile time only)" , TabPrefix: false); |
774 | |
775 | OutStreamer->emitRawComment( |
776 | T: " SGPRBlocks: " + getMCExprStr(Value: CurrentProgramInfo.SGPRBlocks), TabPrefix: false); |
777 | |
778 | OutStreamer->emitRawComment( |
779 | T: " VGPRBlocks: " + getMCExprStr(Value: CurrentProgramInfo.VGPRBlocks), TabPrefix: false); |
780 | |
781 | OutStreamer->emitRawComment( |
782 | T: " NumSGPRsForWavesPerEU: " + |
783 | getMCExprStr(Value: CurrentProgramInfo.NumSGPRsForWavesPerEU), |
784 | TabPrefix: false); |
785 | OutStreamer->emitRawComment( |
786 | T: " NumVGPRsForWavesPerEU: " + |
787 | getMCExprStr(Value: CurrentProgramInfo.NumVGPRsForWavesPerEU), |
788 | TabPrefix: false); |
789 | |
790 | if (STM.hasGFX90AInsts()) { |
791 | const MCExpr *AdjustedAccum = MCBinaryExpr::createAdd( |
792 | LHS: CurrentProgramInfo.AccumOffset, RHS: MCConstantExpr::create(Value: 1, Ctx), Ctx); |
793 | AdjustedAccum = MCBinaryExpr::createMul( |
794 | LHS: AdjustedAccum, RHS: MCConstantExpr::create(Value: 4, Ctx), Ctx); |
795 | OutStreamer->emitRawComment( |
796 | T: " AccumOffset: " + getMCExprStr(Value: AdjustedAccum), TabPrefix: false); |
797 | } |
798 | |
799 | OutStreamer->emitRawComment( |
800 | T: " Occupancy: " + getMCExprStr(Value: CurrentProgramInfo.Occupancy), TabPrefix: false); |
801 | |
802 | OutStreamer->emitRawComment( |
803 | T: " WaveLimiterHint : " + Twine(MFI->needsWaveLimiter()), TabPrefix: false); |
804 | |
805 | OutStreamer->emitRawComment( |
806 | T: " COMPUTE_PGM_RSRC2:SCRATCH_EN: " + |
807 | getMCExprStr(Value: CurrentProgramInfo.ScratchEnable), |
808 | TabPrefix: false); |
809 | OutStreamer->emitRawComment(T: " COMPUTE_PGM_RSRC2:USER_SGPR: " + |
810 | Twine(CurrentProgramInfo.UserSGPR), |
811 | TabPrefix: false); |
812 | OutStreamer->emitRawComment(T: " COMPUTE_PGM_RSRC2:TRAP_HANDLER: " + |
813 | Twine(CurrentProgramInfo.TrapHandlerEnable), |
814 | TabPrefix: false); |
815 | OutStreamer->emitRawComment(T: " COMPUTE_PGM_RSRC2:TGID_X_EN: " + |
816 | Twine(CurrentProgramInfo.TGIdXEnable), |
817 | TabPrefix: false); |
818 | OutStreamer->emitRawComment(T: " COMPUTE_PGM_RSRC2:TGID_Y_EN: " + |
819 | Twine(CurrentProgramInfo.TGIdYEnable), |
820 | TabPrefix: false); |
821 | OutStreamer->emitRawComment(T: " COMPUTE_PGM_RSRC2:TGID_Z_EN: " + |
822 | Twine(CurrentProgramInfo.TGIdZEnable), |
823 | TabPrefix: false); |
824 | OutStreamer->emitRawComment(T: " COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " + |
825 | Twine(CurrentProgramInfo.TIdIGCompCount), |
826 | TabPrefix: false); |
827 | |
828 | [[maybe_unused]] int64_t PGMRSrc3; |
829 | assert(STM.getGeneration() >= AMDGPUSubtarget::GFX10 || |
830 | STM.hasGFX90AInsts() || |
831 | (CurrentProgramInfo.ComputePGMRSrc3->evaluateAsAbsolute(PGMRSrc3) && |
832 | static_cast<uint64_t>(PGMRSrc3) == 0)); |
833 | if (STM.hasGFX90AInsts()) { |
834 | OutStreamer->emitRawComment( |
835 | T: " COMPUTE_PGM_RSRC3_GFX90A:ACCUM_OFFSET: " + |
836 | getMCExprStr(Value: MCKernelDescriptor::bits_get( |
837 | Src: CurrentProgramInfo.ComputePGMRSrc3, |
838 | Shift: amdhsa::COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET_SHIFT, |
839 | Mask: amdhsa::COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET, Ctx)), |
840 | TabPrefix: false); |
841 | OutStreamer->emitRawComment( |
842 | T: " COMPUTE_PGM_RSRC3_GFX90A:TG_SPLIT: " + |
843 | getMCExprStr(Value: MCKernelDescriptor::bits_get( |
844 | Src: CurrentProgramInfo.ComputePGMRSrc3, |
845 | Shift: amdhsa::COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT_SHIFT, |
846 | Mask: amdhsa::COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT, Ctx)), |
847 | TabPrefix: false); |
848 | } |
849 | } |
850 | |
851 | if (DumpCodeInstEmitter) { |
852 | |
853 | OutStreamer->switchSection( |
854 | Section: Context.getELFSection(Section: ".AMDGPU.disasm" , Type: ELF::SHT_PROGBITS, Flags: 0)); |
855 | |
856 | for (size_t i = 0; i < DisasmLines.size(); ++i) { |
857 | std::string = "\n" ; |
858 | if (!HexLines[i].empty()) { |
859 | Comment = std::string(DisasmLineMaxLen - DisasmLines[i].size(), ' '); |
860 | Comment += " ; " + HexLines[i] + "\n" ; |
861 | } |
862 | |
863 | OutStreamer->emitBytes(Data: StringRef(DisasmLines[i])); |
864 | OutStreamer->emitBytes(Data: StringRef(Comment)); |
865 | } |
866 | } |
867 | |
868 | return false; |
869 | } |
870 | |
871 | // TODO: Fold this into emitFunctionBodyStart. |
872 | void AMDGPUAsmPrinter::initializeTargetID(const Module &M) { |
873 | // In the beginning all features are either 'Any' or 'NotSupported', |
874 | // depending on global target features. This will cover empty modules. |
875 | getTargetStreamer()->initializeTargetID(STI: *getGlobalSTI(), |
876 | FeatureString: getGlobalSTI()->getFeatureString()); |
877 | |
878 | // If module is empty, we are done. |
879 | if (M.empty()) |
880 | return; |
881 | |
882 | // If module is not empty, need to find first 'Off' or 'On' feature |
883 | // setting per feature from functions in module. |
884 | for (auto &F : M) { |
885 | auto &TSTargetID = getTargetStreamer()->getTargetID(); |
886 | if ((!TSTargetID->isXnackSupported() || TSTargetID->isXnackOnOrOff()) && |
887 | (!TSTargetID->isSramEccSupported() || TSTargetID->isSramEccOnOrOff())) |
888 | break; |
889 | |
890 | const GCNSubtarget &STM = TM.getSubtarget<GCNSubtarget>(F); |
891 | const IsaInfo::AMDGPUTargetID &STMTargetID = STM.getTargetID(); |
892 | if (TSTargetID->isXnackSupported()) |
893 | if (TSTargetID->getXnackSetting() == IsaInfo::TargetIDSetting::Any) |
894 | TSTargetID->setXnackSetting(STMTargetID.getXnackSetting()); |
895 | if (TSTargetID->isSramEccSupported()) |
896 | if (TSTargetID->getSramEccSetting() == IsaInfo::TargetIDSetting::Any) |
897 | TSTargetID->setSramEccSetting(STMTargetID.getSramEccSetting()); |
898 | } |
899 | } |
900 | |
901 | // AccumOffset computed for the MCExpr equivalent of: |
902 | // alignTo(std::max(1, NumVGPR), 4) / 4 - 1; |
903 | static const MCExpr *computeAccumOffset(const MCExpr *NumVGPR, MCContext &Ctx) { |
904 | const MCExpr *ConstFour = MCConstantExpr::create(Value: 4, Ctx); |
905 | const MCExpr *ConstOne = MCConstantExpr::create(Value: 1, Ctx); |
906 | |
907 | // Can't be lower than 1 for subsequent alignTo. |
908 | const MCExpr *MaximumTaken = |
909 | AMDGPUMCExpr::createMax(Args: {ConstOne, NumVGPR}, Ctx); |
910 | |
911 | // Practically, it's computing divideCeil(MaximumTaken, 4). |
912 | const MCExpr *DivCeil = MCBinaryExpr::createDiv( |
913 | LHS: AMDGPUMCExpr::createAlignTo(Value: MaximumTaken, Align: ConstFour, Ctx), RHS: ConstFour, |
914 | Ctx); |
915 | |
916 | return MCBinaryExpr::createSub(LHS: DivCeil, RHS: ConstOne, Ctx); |
917 | } |
918 | |
919 | void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo, |
920 | const MachineFunction &MF) { |
921 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
922 | bool IsLocal = MF.getFunction().hasLocalLinkage(); |
923 | MCContext &Ctx = MF.getContext(); |
924 | |
925 | auto CreateExpr = [&Ctx](int64_t Value) { |
926 | return MCConstantExpr::create(Value, Ctx); |
927 | }; |
928 | |
929 | auto TryGetMCExprValue = [](const MCExpr *Value, uint64_t &Res) -> bool { |
930 | int64_t Val; |
931 | if (Value->evaluateAsAbsolute(Res&: Val)) { |
932 | Res = Val; |
933 | return true; |
934 | } |
935 | return false; |
936 | }; |
937 | |
938 | auto GetSymRefExpr = |
939 | [&](MCResourceInfo::ResourceInfoKind RIK) -> const MCExpr * { |
940 | MCSymbol *Sym = |
941 | RI.getSymbol(FuncName: CurrentFnSym->getName(), RIK, OutContext, IsLocal); |
942 | return MCSymbolRefExpr::create(Symbol: Sym, Ctx); |
943 | }; |
944 | |
945 | using RIK = MCResourceInfo::ResourceInfoKind; |
946 | ProgInfo.NumArchVGPR = GetSymRefExpr(RIK::RIK_NumVGPR); |
947 | ProgInfo.NumAccVGPR = GetSymRefExpr(RIK::RIK_NumAGPR); |
948 | ProgInfo.NumVGPR = AMDGPUMCExpr::createTotalNumVGPR( |
949 | NumAGPR: ProgInfo.NumAccVGPR, NumVGPR: ProgInfo.NumArchVGPR, Ctx); |
950 | |
951 | ProgInfo.AccumOffset = computeAccumOffset(NumVGPR: ProgInfo.NumArchVGPR, Ctx); |
952 | ProgInfo.TgSplit = STM.isTgSplitEnabled(); |
953 | ProgInfo.NumSGPR = GetSymRefExpr(RIK::RIK_NumSGPR); |
954 | ProgInfo.ScratchSize = GetSymRefExpr(RIK::RIK_PrivateSegSize); |
955 | ProgInfo.VCCUsed = GetSymRefExpr(RIK::RIK_UsesVCC); |
956 | ProgInfo.FlatUsed = GetSymRefExpr(RIK::RIK_UsesFlatScratch); |
957 | ProgInfo.DynamicCallStack = |
958 | MCBinaryExpr::createOr(LHS: GetSymRefExpr(RIK::RIK_HasDynSizedStack), |
959 | RHS: GetSymRefExpr(RIK::RIK_HasRecursion), Ctx); |
960 | |
961 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
962 | |
963 | // The calculations related to SGPR/VGPR blocks are |
964 | // duplicated in part in AMDGPUAsmParser::calculateGPRBlocks, and could be |
965 | // unified. |
966 | const MCExpr * = AMDGPUMCExpr::createExtraSGPRs( |
967 | VCCUsed: ProgInfo.VCCUsed, FlatScrUsed: ProgInfo.FlatUsed, |
968 | XNACKUsed: getTargetStreamer()->getTargetID()->isXnackOnOrAny(), Ctx); |
969 | |
970 | // Check the addressable register limit before we add ExtraSGPRs. |
971 | if (STM.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS && |
972 | !STM.hasSGPRInitBug()) { |
973 | unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs(); |
974 | uint64_t NumSgpr; |
975 | if (TryGetMCExprValue(ProgInfo.NumSGPR, NumSgpr) && |
976 | NumSgpr > MaxAddressableNumSGPRs) { |
977 | // This can happen due to a compiler bug or when using inline asm. |
978 | LLVMContext &Ctx = MF.getFunction().getContext(); |
979 | DiagnosticInfoResourceLimit Diag( |
980 | MF.getFunction(), "addressable scalar registers" , NumSgpr, |
981 | MaxAddressableNumSGPRs, DS_Error, DK_ResourceLimit); |
982 | Ctx.diagnose(DI: Diag); |
983 | ProgInfo.NumSGPR = CreateExpr(MaxAddressableNumSGPRs - 1); |
984 | } |
985 | } |
986 | |
987 | // Account for extra SGPRs and VGPRs reserved for debugger use. |
988 | ProgInfo.NumSGPR = MCBinaryExpr::createAdd(LHS: ProgInfo.NumSGPR, RHS: ExtraSGPRs, Ctx); |
989 | |
990 | const Function &F = MF.getFunction(); |
991 | |
992 | // Ensure there are enough SGPRs and VGPRs for wave dispatch, where wave |
993 | // dispatch registers are function args. |
994 | unsigned WaveDispatchNumSGPR = 0, WaveDispatchNumVGPR = 0; |
995 | |
996 | if (isShader(CC: F.getCallingConv())) { |
997 | bool IsPixelShader = |
998 | F.getCallingConv() == CallingConv::AMDGPU_PS && !STM.isAmdHsaOS(); |
999 | |
1000 | // Calculate the number of VGPR registers based on the SPI input registers |
1001 | uint32_t InputEna = 0; |
1002 | uint32_t InputAddr = 0; |
1003 | unsigned LastEna = 0; |
1004 | |
1005 | if (IsPixelShader) { |
1006 | // Note for IsPixelShader: |
1007 | // By this stage, all enabled inputs are tagged in InputAddr as well. |
1008 | // We will use InputAddr to determine whether the input counts against the |
1009 | // vgpr total and only use the InputEnable to determine the last input |
1010 | // that is relevant - if extra arguments are used, then we have to honour |
1011 | // the InputAddr for any intermediate non-enabled inputs. |
1012 | InputEna = MFI->getPSInputEnable(); |
1013 | InputAddr = MFI->getPSInputAddr(); |
1014 | |
1015 | // We only need to consider input args up to the last used arg. |
1016 | assert((InputEna || InputAddr) && |
1017 | "PSInputAddr and PSInputEnable should " |
1018 | "never both be 0 for AMDGPU_PS shaders" ); |
1019 | // There are some rare circumstances where InputAddr is non-zero and |
1020 | // InputEna can be set to 0. In this case we default to setting LastEna |
1021 | // to 1. |
1022 | LastEna = InputEna ? llvm::Log2_32(Value: InputEna) + 1 : 1; |
1023 | } |
1024 | |
1025 | // FIXME: We should be using the number of registers determined during |
1026 | // calling convention lowering to legalize the types. |
1027 | const DataLayout &DL = F.getDataLayout(); |
1028 | unsigned PSArgCount = 0; |
1029 | unsigned IntermediateVGPR = 0; |
1030 | for (auto &Arg : F.args()) { |
1031 | unsigned NumRegs = (DL.getTypeSizeInBits(Ty: Arg.getType()) + 31) / 32; |
1032 | if (Arg.hasAttribute(Kind: Attribute::InReg)) { |
1033 | WaveDispatchNumSGPR += NumRegs; |
1034 | } else { |
1035 | // If this is a PS shader and we're processing the PS Input args (first |
1036 | // 16 VGPR), use the InputEna and InputAddr bits to define how many |
1037 | // VGPRs are actually used. |
1038 | // Any extra VGPR arguments are handled as normal arguments (and |
1039 | // contribute to the VGPR count whether they're used or not). |
1040 | if (IsPixelShader && PSArgCount < 16) { |
1041 | if ((1 << PSArgCount) & InputAddr) { |
1042 | if (PSArgCount < LastEna) |
1043 | WaveDispatchNumVGPR += NumRegs; |
1044 | else |
1045 | IntermediateVGPR += NumRegs; |
1046 | } |
1047 | PSArgCount++; |
1048 | } else { |
1049 | // If there are extra arguments we have to include the allocation for |
1050 | // the non-used (but enabled with InputAddr) input arguments |
1051 | if (IntermediateVGPR) { |
1052 | WaveDispatchNumVGPR += IntermediateVGPR; |
1053 | IntermediateVGPR = 0; |
1054 | } |
1055 | WaveDispatchNumVGPR += NumRegs; |
1056 | } |
1057 | } |
1058 | } |
1059 | ProgInfo.NumSGPR = AMDGPUMCExpr::createMax( |
1060 | Args: {ProgInfo.NumSGPR, CreateExpr(WaveDispatchNumSGPR)}, Ctx); |
1061 | |
1062 | ProgInfo.NumArchVGPR = AMDGPUMCExpr::createMax( |
1063 | Args: {ProgInfo.NumVGPR, CreateExpr(WaveDispatchNumVGPR)}, Ctx); |
1064 | |
1065 | ProgInfo.NumVGPR = AMDGPUMCExpr::createTotalNumVGPR( |
1066 | NumAGPR: ProgInfo.NumAccVGPR, NumVGPR: ProgInfo.NumArchVGPR, Ctx); |
1067 | } else if (isKernel(CC: F.getCallingConv()) && |
1068 | MFI->getNumKernargPreloadedSGPRs()) { |
1069 | // Consider cases where the total number of UserSGPRs with trailing |
1070 | // allocated preload SGPRs, is greater than the number of explicitly |
1071 | // referenced SGPRs. |
1072 | const MCExpr * = MCBinaryExpr::createAdd( |
1073 | LHS: CreateExpr(MFI->getNumUserSGPRs()), RHS: ExtraSGPRs, Ctx); |
1074 | ProgInfo.NumSGPR = |
1075 | AMDGPUMCExpr::createMax(Args: {ProgInfo.NumSGPR, UserPlusExtraSGPRs}, Ctx); |
1076 | } |
1077 | |
1078 | // Adjust number of registers used to meet default/requested minimum/maximum |
1079 | // number of waves per execution unit request. |
1080 | unsigned MaxWaves = MFI->getMaxWavesPerEU(); |
1081 | ProgInfo.NumSGPRsForWavesPerEU = |
1082 | AMDGPUMCExpr::createMax(Args: {ProgInfo.NumSGPR, CreateExpr(1ul), |
1083 | CreateExpr(STM.getMinNumSGPRs(WavesPerEU: MaxWaves))}, |
1084 | Ctx); |
1085 | ProgInfo.NumVGPRsForWavesPerEU = |
1086 | AMDGPUMCExpr::createMax(Args: {ProgInfo.NumVGPR, CreateExpr(1ul), |
1087 | CreateExpr(STM.getMinNumVGPRs( |
1088 | WavesPerEU: MaxWaves, DynamicVGPRBlockSize: MFI->getDynamicVGPRBlockSize()))}, |
1089 | Ctx); |
1090 | |
1091 | if (STM.getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS || |
1092 | STM.hasSGPRInitBug()) { |
1093 | unsigned MaxAddressableNumSGPRs = STM.getAddressableNumSGPRs(); |
1094 | uint64_t NumSgpr; |
1095 | if (TryGetMCExprValue(ProgInfo.NumSGPR, NumSgpr) && |
1096 | NumSgpr > MaxAddressableNumSGPRs) { |
1097 | // This can happen due to a compiler bug or when using inline asm to use |
1098 | // the registers which are usually reserved for vcc etc. |
1099 | LLVMContext &Ctx = MF.getFunction().getContext(); |
1100 | DiagnosticInfoResourceLimit Diag(MF.getFunction(), "scalar registers" , |
1101 | NumSgpr, MaxAddressableNumSGPRs, |
1102 | DS_Error, DK_ResourceLimit); |
1103 | Ctx.diagnose(DI: Diag); |
1104 | ProgInfo.NumSGPR = CreateExpr(MaxAddressableNumSGPRs); |
1105 | ProgInfo.NumSGPRsForWavesPerEU = CreateExpr(MaxAddressableNumSGPRs); |
1106 | } |
1107 | } |
1108 | |
1109 | if (STM.hasSGPRInitBug()) { |
1110 | ProgInfo.NumSGPR = |
1111 | CreateExpr(AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG); |
1112 | ProgInfo.NumSGPRsForWavesPerEU = |
1113 | CreateExpr(AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG); |
1114 | } |
1115 | |
1116 | if (MFI->getNumUserSGPRs() > STM.getMaxNumUserSGPRs()) { |
1117 | LLVMContext &Ctx = MF.getFunction().getContext(); |
1118 | DiagnosticInfoResourceLimit Diag(MF.getFunction(), "user SGPRs" , |
1119 | MFI->getNumUserSGPRs(), |
1120 | STM.getMaxNumUserSGPRs(), DS_Error); |
1121 | Ctx.diagnose(DI: Diag); |
1122 | } |
1123 | |
1124 | if (MFI->getLDSSize() > |
1125 | static_cast<unsigned>(STM.getAddressableLocalMemorySize())) { |
1126 | LLVMContext &Ctx = MF.getFunction().getContext(); |
1127 | DiagnosticInfoResourceLimit Diag( |
1128 | MF.getFunction(), "local memory" , MFI->getLDSSize(), |
1129 | STM.getAddressableLocalMemorySize(), DS_Error); |
1130 | Ctx.diagnose(DI: Diag); |
1131 | } |
1132 | // The MCExpr equivalent of getNumSGPRBlocks/getNumVGPRBlocks: |
1133 | // (alignTo(max(1u, NumGPR), GPREncodingGranule) / GPREncodingGranule) - 1 |
1134 | auto GetNumGPRBlocks = [&CreateExpr, &Ctx](const MCExpr *NumGPR, |
1135 | unsigned Granule) { |
1136 | const MCExpr *OneConst = CreateExpr(1ul); |
1137 | const MCExpr *GranuleConst = CreateExpr(Granule); |
1138 | const MCExpr *MaxNumGPR = AMDGPUMCExpr::createMax(Args: {NumGPR, OneConst}, Ctx); |
1139 | const MCExpr *AlignToGPR = |
1140 | AMDGPUMCExpr::createAlignTo(Value: MaxNumGPR, Align: GranuleConst, Ctx); |
1141 | const MCExpr *DivGPR = |
1142 | MCBinaryExpr::createDiv(LHS: AlignToGPR, RHS: GranuleConst, Ctx); |
1143 | const MCExpr *SubGPR = MCBinaryExpr::createSub(LHS: DivGPR, RHS: OneConst, Ctx); |
1144 | return SubGPR; |
1145 | }; |
1146 | |
1147 | ProgInfo.SGPRBlocks = GetNumGPRBlocks(ProgInfo.NumSGPRsForWavesPerEU, |
1148 | IsaInfo::getSGPREncodingGranule(STI: &STM)); |
1149 | ProgInfo.VGPRBlocks = GetNumGPRBlocks(ProgInfo.NumVGPRsForWavesPerEU, |
1150 | IsaInfo::getVGPREncodingGranule(STI: &STM)); |
1151 | |
1152 | const SIModeRegisterDefaults Mode = MFI->getMode(); |
1153 | |
1154 | // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode |
1155 | // register. |
1156 | ProgInfo.FloatMode = getFPMode(Mode); |
1157 | |
1158 | ProgInfo.IEEEMode = Mode.IEEE; |
1159 | |
1160 | // Make clamp modifier on NaN input returns 0. |
1161 | ProgInfo.DX10Clamp = Mode.DX10Clamp; |
1162 | |
1163 | unsigned LDSAlignShift; |
1164 | if (STM.getFeatureBits().test(I: FeatureAddressableLocalMemorySize163840)) { |
1165 | // LDS is allocated in 320 dword blocks. |
1166 | LDSAlignShift = 11; |
1167 | } else if (STM.getFeatureBits().test( |
1168 | I: FeatureAddressableLocalMemorySize65536)) { |
1169 | // LDS is allocated in 128 dword blocks. |
1170 | LDSAlignShift = 9; |
1171 | } else { |
1172 | // LDS is allocated in 64 dword blocks. |
1173 | LDSAlignShift = 8; |
1174 | } |
1175 | |
1176 | ProgInfo.SGPRSpill = MFI->getNumSpilledSGPRs(); |
1177 | ProgInfo.VGPRSpill = MFI->getNumSpilledVGPRs(); |
1178 | |
1179 | ProgInfo.LDSSize = MFI->getLDSSize(); |
1180 | ProgInfo.LDSBlocks = |
1181 | alignTo(Value: ProgInfo.LDSSize, Align: 1ULL << LDSAlignShift) >> LDSAlignShift; |
1182 | |
1183 | // The MCExpr equivalent of divideCeil. |
1184 | auto DivideCeil = [&Ctx](const MCExpr *Numerator, const MCExpr *Denominator) { |
1185 | const MCExpr *Ceil = |
1186 | AMDGPUMCExpr::createAlignTo(Value: Numerator, Align: Denominator, Ctx); |
1187 | return MCBinaryExpr::createDiv(LHS: Ceil, RHS: Denominator, Ctx); |
1188 | }; |
1189 | |
1190 | // Scratch is allocated in 64-dword or 256-dword blocks. |
1191 | unsigned ScratchAlignShift = |
1192 | STM.getGeneration() >= AMDGPUSubtarget::GFX11 ? 8 : 10; |
1193 | // We need to program the hardware with the amount of scratch memory that |
1194 | // is used by the entire wave. ProgInfo.ScratchSize is the amount of |
1195 | // scratch memory used per thread. |
1196 | ProgInfo.ScratchBlocks = DivideCeil( |
1197 | MCBinaryExpr::createMul(LHS: ProgInfo.ScratchSize, |
1198 | RHS: CreateExpr(STM.getWavefrontSize()), Ctx), |
1199 | CreateExpr(1ULL << ScratchAlignShift)); |
1200 | |
1201 | if (getIsaVersion(GPU: getGlobalSTI()->getCPU()).Major >= 10) { |
1202 | ProgInfo.WgpMode = STM.isCuModeEnabled() ? 0 : 1; |
1203 | ProgInfo.MemOrdered = 1; |
1204 | ProgInfo.FwdProgress = 1; |
1205 | } |
1206 | |
1207 | // 0 = X, 1 = XY, 2 = XYZ |
1208 | unsigned TIDIGCompCnt = 0; |
1209 | if (MFI->hasWorkItemIDZ()) |
1210 | TIDIGCompCnt = 2; |
1211 | else if (MFI->hasWorkItemIDY()) |
1212 | TIDIGCompCnt = 1; |
1213 | |
1214 | // The private segment wave byte offset is the last of the system SGPRs. We |
1215 | // initially assumed it was allocated, and may have used it. It shouldn't harm |
1216 | // anything to disable it if we know the stack isn't used here. We may still |
1217 | // have emitted code reading it to initialize scratch, but if that's unused |
1218 | // reading garbage should be OK. |
1219 | ProgInfo.ScratchEnable = MCBinaryExpr::createLOr( |
1220 | LHS: MCBinaryExpr::createGT(LHS: ProgInfo.ScratchBlocks, |
1221 | RHS: MCConstantExpr::create(Value: 0, Ctx), Ctx), |
1222 | RHS: ProgInfo.DynamicCallStack, Ctx); |
1223 | |
1224 | ProgInfo.UserSGPR = MFI->getNumUserSGPRs(); |
1225 | // For AMDHSA, TRAP_HANDLER must be zero, as it is populated by the CP. |
1226 | ProgInfo.TrapHandlerEnable = |
1227 | STM.isAmdHsaOS() ? 0 : STM.isTrapHandlerEnabled(); |
1228 | ProgInfo.TGIdXEnable = MFI->hasWorkGroupIDX(); |
1229 | ProgInfo.TGIdYEnable = MFI->hasWorkGroupIDY(); |
1230 | ProgInfo.TGIdZEnable = MFI->hasWorkGroupIDZ(); |
1231 | ProgInfo.TGSizeEnable = MFI->hasWorkGroupInfo(); |
1232 | ProgInfo.TIdIGCompCount = TIDIGCompCnt; |
1233 | ProgInfo.EXCPEnMSB = 0; |
1234 | // For AMDHSA, LDS_SIZE must be zero, as it is populated by the CP. |
1235 | ProgInfo.LdsSize = STM.isAmdHsaOS() ? 0 : ProgInfo.LDSBlocks; |
1236 | ProgInfo.EXCPEnable = 0; |
1237 | |
1238 | // return ((Dst & ~Mask) | (Value << Shift)) |
1239 | auto SetBits = [&Ctx](const MCExpr *Dst, const MCExpr *Value, uint32_t Mask, |
1240 | uint32_t Shift) { |
1241 | const auto *Shft = MCConstantExpr::create(Value: Shift, Ctx); |
1242 | const auto *Msk = MCConstantExpr::create(Value: Mask, Ctx); |
1243 | Dst = MCBinaryExpr::createAnd(LHS: Dst, RHS: MCUnaryExpr::createNot(Expr: Msk, Ctx), Ctx); |
1244 | Dst = MCBinaryExpr::createOr(LHS: Dst, RHS: MCBinaryExpr::createShl(LHS: Value, RHS: Shft, Ctx), |
1245 | Ctx); |
1246 | return Dst; |
1247 | }; |
1248 | |
1249 | if (STM.hasGFX90AInsts()) { |
1250 | ProgInfo.ComputePGMRSrc3 = |
1251 | SetBits(ProgInfo.ComputePGMRSrc3, ProgInfo.AccumOffset, |
1252 | amdhsa::COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET, |
1253 | amdhsa::COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET_SHIFT); |
1254 | ProgInfo.ComputePGMRSrc3 = |
1255 | SetBits(ProgInfo.ComputePGMRSrc3, CreateExpr(ProgInfo.TgSplit), |
1256 | amdhsa::COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT, |
1257 | amdhsa::COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT_SHIFT); |
1258 | } |
1259 | |
1260 | ProgInfo.Occupancy = AMDGPUMCExpr::createOccupancy( |
1261 | InitOcc: STM.computeOccupancy(F, LDSSize: ProgInfo.LDSSize).second, |
1262 | NumSGPRs: ProgInfo.NumSGPRsForWavesPerEU, NumVGPRs: ProgInfo.NumVGPRsForWavesPerEU, |
1263 | DynamicVGPRBlockSize: MFI->getDynamicVGPRBlockSize(), STM, Ctx); |
1264 | |
1265 | const auto [MinWEU, MaxWEU] = |
1266 | AMDGPU::getIntegerPairAttribute(F, Name: "amdgpu-waves-per-eu" , Default: {0, 0}, OnlyFirstRequired: true); |
1267 | uint64_t Occupancy; |
1268 | if (TryGetMCExprValue(ProgInfo.Occupancy, Occupancy) && Occupancy < MinWEU) { |
1269 | DiagnosticInfoOptimizationFailure Diag( |
1270 | F, F.getSubprogram(), |
1271 | "failed to meet occupancy target given by 'amdgpu-waves-per-eu' in " |
1272 | "'" + |
1273 | F.getName() + "': desired occupancy was " + Twine(MinWEU) + |
1274 | ", final occupancy is " + Twine(Occupancy)); |
1275 | F.getContext().diagnose(DI: Diag); |
1276 | } |
1277 | |
1278 | if (isGFX11Plus(STI: STM)) { |
1279 | uint32_t CodeSizeInBytes = (uint32_t)std::min( |
1280 | a: ProgInfo.getFunctionCodeSize(MF, IsLowerBound: true /* IsLowerBound */), |
1281 | b: (uint64_t)std::numeric_limits<uint32_t>::max()); |
1282 | uint32_t CodeSizeInLines = divideCeil(Numerator: CodeSizeInBytes, Denominator: 128); |
1283 | uint32_t Field, Shift, Width; |
1284 | if (isGFX11(STI: STM)) { |
1285 | Field = amdhsa::COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE; |
1286 | Shift = amdhsa::COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE_SHIFT; |
1287 | Width = amdhsa::COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE_WIDTH; |
1288 | } else { |
1289 | Field = amdhsa::COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE; |
1290 | Shift = amdhsa::COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE_SHIFT; |
1291 | Width = amdhsa::COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE_WIDTH; |
1292 | } |
1293 | uint64_t InstPrefSize = std::min(a: CodeSizeInLines, b: (1u << Width) - 1); |
1294 | ProgInfo.ComputePGMRSrc3 = SetBits(ProgInfo.ComputePGMRSrc3, |
1295 | CreateExpr(InstPrefSize), Field, Shift); |
1296 | } |
1297 | } |
1298 | |
1299 | static unsigned getRsrcReg(CallingConv::ID CallConv) { |
1300 | switch (CallConv) { |
1301 | default: [[fallthrough]]; |
1302 | case CallingConv::AMDGPU_CS: return R_00B848_COMPUTE_PGM_RSRC1; |
1303 | case CallingConv::AMDGPU_LS: return R_00B528_SPI_SHADER_PGM_RSRC1_LS; |
1304 | case CallingConv::AMDGPU_HS: return R_00B428_SPI_SHADER_PGM_RSRC1_HS; |
1305 | case CallingConv::AMDGPU_ES: return R_00B328_SPI_SHADER_PGM_RSRC1_ES; |
1306 | case CallingConv::AMDGPU_GS: return R_00B228_SPI_SHADER_PGM_RSRC1_GS; |
1307 | case CallingConv::AMDGPU_VS: return R_00B128_SPI_SHADER_PGM_RSRC1_VS; |
1308 | case CallingConv::AMDGPU_PS: return R_00B028_SPI_SHADER_PGM_RSRC1_PS; |
1309 | } |
1310 | } |
1311 | |
1312 | void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF, |
1313 | const SIProgramInfo &CurrentProgramInfo) { |
1314 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
1315 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
1316 | unsigned RsrcReg = getRsrcReg(CallConv: MF.getFunction().getCallingConv()); |
1317 | MCContext &Ctx = MF.getContext(); |
1318 | |
1319 | // (((Value) & Mask) << Shift) |
1320 | auto SetBits = [&Ctx](const MCExpr *Value, uint32_t Mask, uint32_t Shift) { |
1321 | const MCExpr *msk = MCConstantExpr::create(Value: Mask, Ctx); |
1322 | const MCExpr *shft = MCConstantExpr::create(Value: Shift, Ctx); |
1323 | return MCBinaryExpr::createShl(LHS: MCBinaryExpr::createAnd(LHS: Value, RHS: msk, Ctx), |
1324 | RHS: shft, Ctx); |
1325 | }; |
1326 | |
1327 | auto EmitResolvedOrExpr = [this](const MCExpr *Value, unsigned Size) { |
1328 | int64_t Val; |
1329 | if (Value->evaluateAsAbsolute(Res&: Val)) |
1330 | OutStreamer->emitIntValue(Value: static_cast<uint64_t>(Val), Size); |
1331 | else |
1332 | OutStreamer->emitValue(Value, Size); |
1333 | }; |
1334 | |
1335 | if (AMDGPU::isCompute(CC: MF.getFunction().getCallingConv())) { |
1336 | OutStreamer->emitInt32(R_00B848_COMPUTE_PGM_RSRC1); |
1337 | |
1338 | EmitResolvedOrExpr(CurrentProgramInfo.getComputePGMRSrc1(ST: STM, Ctx), |
1339 | /*Size=*/4); |
1340 | |
1341 | OutStreamer->emitInt32(R_00B84C_COMPUTE_PGM_RSRC2); |
1342 | EmitResolvedOrExpr(CurrentProgramInfo.getComputePGMRSrc2(Ctx), /*Size=*/4); |
1343 | |
1344 | OutStreamer->emitInt32(R_00B860_COMPUTE_TMPRING_SIZE); |
1345 | |
1346 | // Sets bits according to S_0286E8_WAVESIZE_* mask and shift values for the |
1347 | // appropriate generation. |
1348 | if (STM.getGeneration() >= AMDGPUSubtarget::GFX12) { |
1349 | EmitResolvedOrExpr(SetBits(CurrentProgramInfo.ScratchBlocks, |
1350 | /*Mask=*/0x3FFFF, /*Shift=*/12), |
1351 | /*Size=*/4); |
1352 | } else if (STM.getGeneration() == AMDGPUSubtarget::GFX11) { |
1353 | EmitResolvedOrExpr(SetBits(CurrentProgramInfo.ScratchBlocks, |
1354 | /*Mask=*/0x7FFF, /*Shift=*/12), |
1355 | /*Size=*/4); |
1356 | } else { |
1357 | EmitResolvedOrExpr(SetBits(CurrentProgramInfo.ScratchBlocks, |
1358 | /*Mask=*/0x1FFF, /*Shift=*/12), |
1359 | /*Size=*/4); |
1360 | } |
1361 | |
1362 | // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 = |
1363 | // 0" comment but I don't see a corresponding field in the register spec. |
1364 | } else { |
1365 | OutStreamer->emitInt32(Value: RsrcReg); |
1366 | |
1367 | const MCExpr *GPRBlocks = MCBinaryExpr::createOr( |
1368 | LHS: SetBits(CurrentProgramInfo.VGPRBlocks, /*Mask=*/0x3F, /*Shift=*/0), |
1369 | RHS: SetBits(CurrentProgramInfo.SGPRBlocks, /*Mask=*/0x0F, /*Shift=*/6), |
1370 | Ctx&: MF.getContext()); |
1371 | EmitResolvedOrExpr(GPRBlocks, /*Size=*/4); |
1372 | OutStreamer->emitInt32(R_0286E8_SPI_TMPRING_SIZE); |
1373 | |
1374 | // Sets bits according to S_0286E8_WAVESIZE_* mask and shift values for the |
1375 | // appropriate generation. |
1376 | if (STM.getGeneration() >= AMDGPUSubtarget::GFX12) { |
1377 | EmitResolvedOrExpr(SetBits(CurrentProgramInfo.ScratchBlocks, |
1378 | /*Mask=*/0x3FFFF, /*Shift=*/12), |
1379 | /*Size=*/4); |
1380 | } else if (STM.getGeneration() == AMDGPUSubtarget::GFX11) { |
1381 | EmitResolvedOrExpr(SetBits(CurrentProgramInfo.ScratchBlocks, |
1382 | /*Mask=*/0x7FFF, /*Shift=*/12), |
1383 | /*Size=*/4); |
1384 | } else { |
1385 | EmitResolvedOrExpr(SetBits(CurrentProgramInfo.ScratchBlocks, |
1386 | /*Mask=*/0x1FFF, /*Shift=*/12), |
1387 | /*Size=*/4); |
1388 | } |
1389 | } |
1390 | |
1391 | if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS) { |
1392 | OutStreamer->emitInt32(R_00B02C_SPI_SHADER_PGM_RSRC2_PS); |
1393 | unsigned = STM.getGeneration() >= AMDGPUSubtarget::GFX11 |
1394 | ? divideCeil(Numerator: CurrentProgramInfo.LDSBlocks, Denominator: 2) |
1395 | : CurrentProgramInfo.LDSBlocks; |
1396 | OutStreamer->emitInt32(S_00B02C_EXTRA_LDS_SIZE(ExtraLDSSize)); |
1397 | OutStreamer->emitInt32(R_0286CC_SPI_PS_INPUT_ENA); |
1398 | OutStreamer->emitInt32(Value: MFI->getPSInputEnable()); |
1399 | OutStreamer->emitInt32(R_0286D0_SPI_PS_INPUT_ADDR); |
1400 | OutStreamer->emitInt32(Value: MFI->getPSInputAddr()); |
1401 | } |
1402 | |
1403 | OutStreamer->emitInt32(R_SPILLED_SGPRS); |
1404 | OutStreamer->emitInt32(Value: MFI->getNumSpilledSGPRs()); |
1405 | OutStreamer->emitInt32(R_SPILLED_VGPRS); |
1406 | OutStreamer->emitInt32(Value: MFI->getNumSpilledVGPRs()); |
1407 | } |
1408 | |
1409 | // Helper function to add common PAL Metadata 3.0+ |
1410 | static void EmitPALMetadataCommon(AMDGPUPALMetadata *MD, |
1411 | const SIProgramInfo &CurrentProgramInfo, |
1412 | CallingConv::ID CC, const GCNSubtarget &ST, |
1413 | unsigned DynamicVGPRBlockSize) { |
1414 | if (ST.hasIEEEMode()) |
1415 | MD->setHwStage(CC, field: ".ieee_mode" , Val: (bool)CurrentProgramInfo.IEEEMode); |
1416 | |
1417 | MD->setHwStage(CC, field: ".wgp_mode" , Val: (bool)CurrentProgramInfo.WgpMode); |
1418 | MD->setHwStage(CC, field: ".mem_ordered" , Val: (bool)CurrentProgramInfo.MemOrdered); |
1419 | |
1420 | if (AMDGPU::isCompute(CC)) { |
1421 | MD->setHwStage(CC, field: ".trap_present" , |
1422 | Val: (bool)CurrentProgramInfo.TrapHandlerEnable); |
1423 | MD->setHwStage(CC, field: ".excp_en" , Val: CurrentProgramInfo.EXCPEnable); |
1424 | |
1425 | if (DynamicVGPRBlockSize != 0) |
1426 | MD->setComputeRegisters(field: ".dynamic_vgpr_en" , Val: true); |
1427 | } |
1428 | |
1429 | MD->setHwStage(CC, field: ".lds_size" , |
1430 | Val: (unsigned)(CurrentProgramInfo.LdsSize * |
1431 | getLdsDwGranularity(ST) * sizeof(uint32_t))); |
1432 | } |
1433 | |
1434 | // This is the equivalent of EmitProgramInfoSI above, but for when the OS type |
1435 | // is AMDPAL. It stores each compute/SPI register setting and other PAL |
1436 | // metadata items into the PALMD::Metadata, combining with any provided by the |
1437 | // frontend as LLVM metadata. Once all functions are written, the PAL metadata |
1438 | // is then written as a single block in the .note section. |
1439 | void AMDGPUAsmPrinter::EmitPALMetadata(const MachineFunction &MF, |
1440 | const SIProgramInfo &CurrentProgramInfo) { |
1441 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
1442 | auto CC = MF.getFunction().getCallingConv(); |
1443 | auto *MD = getTargetStreamer()->getPALMetadata(); |
1444 | auto &Ctx = MF.getContext(); |
1445 | |
1446 | MD->setEntryPoint(CC, Name: MF.getFunction().getName()); |
1447 | MD->setNumUsedVgprs(CC, Val: CurrentProgramInfo.NumVGPRsForWavesPerEU, Ctx); |
1448 | |
1449 | // For targets that support dynamic VGPRs, set the number of saved dynamic |
1450 | // VGPRs (if any) in the PAL metadata. |
1451 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
1452 | if (MFI->isDynamicVGPREnabled() && |
1453 | MFI->getScratchReservedForDynamicVGPRs() > 0) |
1454 | MD->setHwStage(CC, field: ".dynamic_vgpr_saved_count" , |
1455 | Val: MFI->getScratchReservedForDynamicVGPRs() / 4); |
1456 | |
1457 | // Only set AGPRs for supported devices |
1458 | if (STM.hasMAIInsts()) { |
1459 | MD->setNumUsedAgprs(CC, Val: CurrentProgramInfo.NumAccVGPR); |
1460 | } |
1461 | |
1462 | MD->setNumUsedSgprs(CC, Val: CurrentProgramInfo.NumSGPRsForWavesPerEU, Ctx); |
1463 | if (MD->getPALMajorVersion() < 3) { |
1464 | MD->setRsrc1(CC, Val: CurrentProgramInfo.getPGMRSrc1(CC, ST: STM, Ctx), Ctx); |
1465 | if (AMDGPU::isCompute(CC)) { |
1466 | MD->setRsrc2(CC, Val: CurrentProgramInfo.getComputePGMRSrc2(Ctx), Ctx); |
1467 | } else { |
1468 | const MCExpr *HasScratchBlocks = |
1469 | MCBinaryExpr::createGT(LHS: CurrentProgramInfo.ScratchBlocks, |
1470 | RHS: MCConstantExpr::create(Value: 0, Ctx), Ctx); |
1471 | auto [Shift, Mask] = getShiftMask(C_00B84C_SCRATCH_EN); |
1472 | MD->setRsrc2(CC, Val: maskShiftSet(Val: HasScratchBlocks, Mask, Shift, Ctx), Ctx); |
1473 | } |
1474 | } else { |
1475 | MD->setHwStage(CC, field: ".debug_mode" , Val: (bool)CurrentProgramInfo.DebugMode); |
1476 | MD->setHwStage(CC, field: ".scratch_en" , Type: msgpack::Type::Boolean, |
1477 | Val: CurrentProgramInfo.ScratchEnable); |
1478 | EmitPALMetadataCommon(MD, CurrentProgramInfo, CC, ST: STM, |
1479 | DynamicVGPRBlockSize: MFI->getDynamicVGPRBlockSize()); |
1480 | } |
1481 | |
1482 | // ScratchSize is in bytes, 16 aligned. |
1483 | MD->setScratchSize( |
1484 | CC, |
1485 | Val: AMDGPUMCExpr::createAlignTo(Value: CurrentProgramInfo.ScratchSize, |
1486 | Align: MCConstantExpr::create(Value: 16, Ctx), Ctx), |
1487 | Ctx); |
1488 | |
1489 | if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS) { |
1490 | unsigned = STM.getGeneration() >= AMDGPUSubtarget::GFX11 |
1491 | ? divideCeil(Numerator: CurrentProgramInfo.LDSBlocks, Denominator: 2) |
1492 | : CurrentProgramInfo.LDSBlocks; |
1493 | if (MD->getPALMajorVersion() < 3) { |
1494 | MD->setRsrc2( |
1495 | CC, |
1496 | Val: MCConstantExpr::create(S_00B02C_EXTRA_LDS_SIZE(ExtraLDSSize), Ctx), |
1497 | Ctx); |
1498 | MD->setSpiPsInputEna(MFI->getPSInputEnable()); |
1499 | MD->setSpiPsInputAddr(MFI->getPSInputAddr()); |
1500 | } else { |
1501 | // Graphics registers |
1502 | const unsigned = |
1503 | STM.getGeneration() >= AMDGPUSubtarget::GFX11 ? 256 : 128; |
1504 | MD->setGraphicsRegisters( |
1505 | field: ".ps_extra_lds_size" , |
1506 | Val: (unsigned)(ExtraLDSSize * ExtraLdsDwGranularity * sizeof(uint32_t))); |
1507 | |
1508 | // Set PsInputEna and PsInputAddr .spi_ps_input_ena and .spi_ps_input_addr |
1509 | static StringLiteral const PsInputFields[] = { |
1510 | ".persp_sample_ena" , ".persp_center_ena" , |
1511 | ".persp_centroid_ena" , ".persp_pull_model_ena" , |
1512 | ".linear_sample_ena" , ".linear_center_ena" , |
1513 | ".linear_centroid_ena" , ".line_stipple_tex_ena" , |
1514 | ".pos_x_float_ena" , ".pos_y_float_ena" , |
1515 | ".pos_z_float_ena" , ".pos_w_float_ena" , |
1516 | ".front_face_ena" , ".ancillary_ena" , |
1517 | ".sample_coverage_ena" , ".pos_fixed_pt_ena" }; |
1518 | unsigned PSInputEna = MFI->getPSInputEnable(); |
1519 | unsigned PSInputAddr = MFI->getPSInputAddr(); |
1520 | for (auto [Idx, Field] : enumerate(First: PsInputFields)) { |
1521 | MD->setGraphicsRegisters(field1: ".spi_ps_input_ena" , field2: Field, |
1522 | Val: (bool)((PSInputEna >> Idx) & 1)); |
1523 | MD->setGraphicsRegisters(field1: ".spi_ps_input_addr" , field2: Field, |
1524 | Val: (bool)((PSInputAddr >> Idx) & 1)); |
1525 | } |
1526 | } |
1527 | } |
1528 | |
1529 | // For version 3 and above the wave front size is already set in the metadata |
1530 | if (MD->getPALMajorVersion() < 3 && STM.isWave32()) |
1531 | MD->setWave32(MF.getFunction().getCallingConv()); |
1532 | } |
1533 | |
1534 | void AMDGPUAsmPrinter::emitPALFunctionMetadata(const MachineFunction &MF) { |
1535 | auto *MD = getTargetStreamer()->getPALMetadata(); |
1536 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
1537 | StringRef FnName = MF.getFunction().getName(); |
1538 | MD->setFunctionScratchSize(FnName, Val: MFI.getStackSize()); |
1539 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
1540 | MCContext &Ctx = MF.getContext(); |
1541 | |
1542 | if (MD->getPALMajorVersion() < 3) { |
1543 | // Set compute registers |
1544 | MD->setRsrc1( |
1545 | CC: CallingConv::AMDGPU_CS, |
1546 | Val: CurrentProgramInfo.getPGMRSrc1(CC: CallingConv::AMDGPU_CS, ST, Ctx), Ctx); |
1547 | MD->setRsrc2(CC: CallingConv::AMDGPU_CS, |
1548 | Val: CurrentProgramInfo.getComputePGMRSrc2(Ctx), Ctx); |
1549 | } else { |
1550 | EmitPALMetadataCommon( |
1551 | MD, CurrentProgramInfo, CC: CallingConv::AMDGPU_CS, ST, |
1552 | DynamicVGPRBlockSize: MF.getInfo<SIMachineFunctionInfo>()->getDynamicVGPRBlockSize()); |
1553 | } |
1554 | |
1555 | // Set optional info |
1556 | MD->setFunctionLdsSize(FnName, Val: CurrentProgramInfo.LDSSize); |
1557 | MD->setFunctionNumUsedVgprs(FnName, Val: CurrentProgramInfo.NumVGPRsForWavesPerEU); |
1558 | MD->setFunctionNumUsedSgprs(FnName, Val: CurrentProgramInfo.NumSGPRsForWavesPerEU); |
1559 | } |
1560 | |
1561 | // This is supposed to be log2(Size) |
1562 | static amd_element_byte_size_t getElementByteSizeValue(unsigned Size) { |
1563 | switch (Size) { |
1564 | case 4: |
1565 | return AMD_ELEMENT_4_BYTES; |
1566 | case 8: |
1567 | return AMD_ELEMENT_8_BYTES; |
1568 | case 16: |
1569 | return AMD_ELEMENT_16_BYTES; |
1570 | default: |
1571 | llvm_unreachable("invalid private_element_size" ); |
1572 | } |
1573 | } |
1574 | |
1575 | void AMDGPUAsmPrinter::getAmdKernelCode(AMDGPUMCKernelCodeT &Out, |
1576 | const SIProgramInfo &CurrentProgramInfo, |
1577 | const MachineFunction &MF) const { |
1578 | const Function &F = MF.getFunction(); |
1579 | assert(F.getCallingConv() == CallingConv::AMDGPU_KERNEL || |
1580 | F.getCallingConv() == CallingConv::SPIR_KERNEL); |
1581 | |
1582 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
1583 | const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>(); |
1584 | MCContext &Ctx = MF.getContext(); |
1585 | |
1586 | Out.initDefault(STI: &STM, Ctx, /*InitMCExpr=*/false); |
1587 | |
1588 | Out.compute_pgm_resource1_registers = |
1589 | CurrentProgramInfo.getComputePGMRSrc1(ST: STM, Ctx); |
1590 | Out.compute_pgm_resource2_registers = |
1591 | CurrentProgramInfo.getComputePGMRSrc2(Ctx); |
1592 | Out.code_properties |= AMD_CODE_PROPERTY_IS_PTR64; |
1593 | |
1594 | Out.is_dynamic_callstack = CurrentProgramInfo.DynamicCallStack; |
1595 | |
1596 | AMD_HSA_BITS_SET(Out.code_properties, AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE, |
1597 | getElementByteSizeValue(STM.getMaxPrivateElementSize(true))); |
1598 | |
1599 | const GCNUserSGPRUsageInfo &UserSGPRInfo = MFI->getUserSGPRInfo(); |
1600 | if (UserSGPRInfo.hasPrivateSegmentBuffer()) { |
1601 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER; |
1602 | } |
1603 | |
1604 | if (UserSGPRInfo.hasDispatchPtr()) |
1605 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR; |
1606 | |
1607 | if (UserSGPRInfo.hasQueuePtr()) |
1608 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR; |
1609 | |
1610 | if (UserSGPRInfo.hasKernargSegmentPtr()) |
1611 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR; |
1612 | |
1613 | if (UserSGPRInfo.hasDispatchID()) |
1614 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID; |
1615 | |
1616 | if (UserSGPRInfo.hasFlatScratchInit()) |
1617 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT; |
1618 | |
1619 | if (UserSGPRInfo.hasPrivateSegmentSize()) |
1620 | Out.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE; |
1621 | |
1622 | if (STM.isXNACKEnabled()) |
1623 | Out.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED; |
1624 | |
1625 | Align MaxKernArgAlign; |
1626 | Out.kernarg_segment_byte_size = STM.getKernArgSegmentSize(F, MaxAlign&: MaxKernArgAlign); |
1627 | Out.wavefront_sgpr_count = CurrentProgramInfo.NumSGPR; |
1628 | Out.workitem_vgpr_count = CurrentProgramInfo.NumVGPR; |
1629 | Out.workitem_private_segment_byte_size = CurrentProgramInfo.ScratchSize; |
1630 | Out.workgroup_group_segment_byte_size = CurrentProgramInfo.LDSSize; |
1631 | |
1632 | // kernarg_segment_alignment is specified as log of the alignment. |
1633 | // The minimum alignment is 16. |
1634 | // FIXME: The metadata treats the minimum as 4? |
1635 | Out.kernarg_segment_alignment = Log2(A: std::max(a: Align(16), b: MaxKernArgAlign)); |
1636 | } |
1637 | |
1638 | bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo, |
1639 | const char *, raw_ostream &O) { |
1640 | // First try the generic code, which knows about modifiers like 'c' and 'n'. |
1641 | if (!AsmPrinter::PrintAsmOperand(MI, OpNo, ExtraCode, OS&: O)) |
1642 | return false; |
1643 | |
1644 | if (ExtraCode && ExtraCode[0]) { |
1645 | if (ExtraCode[1] != 0) |
1646 | return true; // Unknown modifier. |
1647 | |
1648 | switch (ExtraCode[0]) { |
1649 | case 'r': |
1650 | break; |
1651 | default: |
1652 | return true; |
1653 | } |
1654 | } |
1655 | |
1656 | // TODO: Should be able to support other operand types like globals. |
1657 | const MachineOperand &MO = MI->getOperand(i: OpNo); |
1658 | if (MO.isReg()) { |
1659 | AMDGPUInstPrinter::printRegOperand(Reg: MO.getReg(), O, |
1660 | MRI: *MF->getSubtarget().getRegisterInfo()); |
1661 | return false; |
1662 | } |
1663 | if (MO.isImm()) { |
1664 | int64_t Val = MO.getImm(); |
1665 | if (AMDGPU::isInlinableIntLiteral(Literal: Val)) { |
1666 | O << Val; |
1667 | } else if (isUInt<16>(x: Val)) { |
1668 | O << format(Fmt: "0x%" PRIx16, Vals: static_cast<uint16_t>(Val)); |
1669 | } else if (isUInt<32>(x: Val)) { |
1670 | O << format(Fmt: "0x%" PRIx32, Vals: static_cast<uint32_t>(Val)); |
1671 | } else { |
1672 | O << format(Fmt: "0x%" PRIx64, Vals: static_cast<uint64_t>(Val)); |
1673 | } |
1674 | return false; |
1675 | } |
1676 | return true; |
1677 | } |
1678 | |
1679 | void AMDGPUAsmPrinter::getAnalysisUsage(AnalysisUsage &AU) const { |
1680 | AU.addRequired<AMDGPUResourceUsageAnalysis>(); |
1681 | AU.addPreserved<AMDGPUResourceUsageAnalysis>(); |
1682 | AU.addRequired<MachineModuleInfoWrapperPass>(); |
1683 | AU.addPreserved<MachineModuleInfoWrapperPass>(); |
1684 | AsmPrinter::getAnalysisUsage(AU); |
1685 | } |
1686 | |
1687 | void AMDGPUAsmPrinter::( |
1688 | const MachineFunction &MF, const SIProgramInfo &CurrentProgramInfo, |
1689 | bool isModuleEntryFunction, bool hasMAIInsts) { |
1690 | if (!ORE) |
1691 | return; |
1692 | |
1693 | const char *Name = "kernel-resource-usage" ; |
1694 | const char *Indent = " " ; |
1695 | |
1696 | // If the remark is not specifically enabled, do not output to yaml |
1697 | LLVMContext &Ctx = MF.getFunction().getContext(); |
1698 | if (!Ctx.getDiagHandlerPtr()->isAnalysisRemarkEnabled(PassName: Name)) |
1699 | return; |
1700 | |
1701 | // Currently non-kernel functions have no resources to emit. |
1702 | if (!isEntryFunctionCC(CC: MF.getFunction().getCallingConv())) |
1703 | return; |
1704 | |
1705 | auto = [&](StringRef , |
1706 | StringRef , auto Argument) { |
1707 | // Add an indent for every line besides the line with the kernel name. This |
1708 | // makes it easier to tell which resource usage go with which kernel since |
1709 | // the kernel name will always be displayed first. |
1710 | std::string LabelStr = RemarkLabel.str() + ": " ; |
1711 | if (RemarkName != "FunctionName" ) |
1712 | LabelStr = Indent + LabelStr; |
1713 | |
1714 | ORE->emit([&]() { |
1715 | return MachineOptimizationRemarkAnalysis(Name, RemarkName, |
1716 | MF.getFunction().getSubprogram(), |
1717 | &MF.front()) |
1718 | << LabelStr << ore::NV(RemarkName, Argument); |
1719 | }); |
1720 | }; |
1721 | |
1722 | // FIXME: Formatting here is pretty nasty because clang does not accept |
1723 | // newlines from diagnostics. This forces us to emit multiple diagnostic |
1724 | // remarks to simulate newlines. If and when clang does accept newlines, this |
1725 | // formatting should be aggregated into one remark with newlines to avoid |
1726 | // printing multiple diagnostic location and diag opts. |
1727 | EmitResourceUsageRemark("FunctionName" , "Function Name" , |
1728 | MF.getFunction().getName()); |
1729 | EmitResourceUsageRemark("NumSGPR" , "TotalSGPRs" , |
1730 | getMCExprStr(Value: CurrentProgramInfo.NumSGPR)); |
1731 | EmitResourceUsageRemark("NumVGPR" , "VGPRs" , |
1732 | getMCExprStr(Value: CurrentProgramInfo.NumArchVGPR)); |
1733 | if (hasMAIInsts) { |
1734 | EmitResourceUsageRemark("NumAGPR" , "AGPRs" , |
1735 | getMCExprStr(Value: CurrentProgramInfo.NumAccVGPR)); |
1736 | } |
1737 | EmitResourceUsageRemark("ScratchSize" , "ScratchSize [bytes/lane]" , |
1738 | getMCExprStr(Value: CurrentProgramInfo.ScratchSize)); |
1739 | int64_t DynStack; |
1740 | bool DynStackEvaluatable = |
1741 | CurrentProgramInfo.DynamicCallStack->evaluateAsAbsolute(Res&: DynStack); |
1742 | StringRef DynamicStackStr = |
1743 | DynStackEvaluatable && DynStack ? "True" : "False" ; |
1744 | EmitResourceUsageRemark("DynamicStack" , "Dynamic Stack" , DynamicStackStr); |
1745 | EmitResourceUsageRemark("Occupancy" , "Occupancy [waves/SIMD]" , |
1746 | getMCExprStr(Value: CurrentProgramInfo.Occupancy)); |
1747 | EmitResourceUsageRemark("SGPRSpill" , "SGPRs Spill" , |
1748 | CurrentProgramInfo.SGPRSpill); |
1749 | EmitResourceUsageRemark("VGPRSpill" , "VGPRs Spill" , |
1750 | CurrentProgramInfo.VGPRSpill); |
1751 | if (isModuleEntryFunction) |
1752 | EmitResourceUsageRemark("BytesLDS" , "LDS Size [bytes/block]" , |
1753 | CurrentProgramInfo.LDSSize); |
1754 | } |
1755 | |
1756 | char AMDGPUAsmPrinter::ID = 0; |
1757 | |
1758 | INITIALIZE_PASS(AMDGPUAsmPrinter, "amdgpu-asm-printer" , |
1759 | "AMDGPU Assembly Printer" , false, false) |
1760 | |