xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.cpp (revision 357378bbdedf24ce2b90e9bd831af4a9db3ec70a)
1 //===-- AMDGPUPALMetadata.cpp - Accumulate and print AMDGPU PAL metadata  -===//
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 /// This class has methods called by AMDGPUAsmPrinter to accumulate and print
12 /// the PAL metadata.
13 //
14 //===----------------------------------------------------------------------===//
15 //
16 
17 #include "AMDGPUPALMetadata.h"
18 #include "AMDGPUPTNote.h"
19 #include "SIDefines.h"
20 #include "llvm/BinaryFormat/ELF.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/Module.h"
23 #include "llvm/Support/AMDGPUMetadata.h"
24 #include "llvm/Support/EndianStream.h"
25 
26 using namespace llvm;
27 using namespace llvm::AMDGPU;
28 
29 // Read the PAL metadata from IR metadata, where it was put by the frontend.
30 void AMDGPUPALMetadata::readFromIR(Module &M) {
31   auto NamedMD = M.getNamedMetadata("amdgpu.pal.metadata.msgpack");
32   if (NamedMD && NamedMD->getNumOperands()) {
33     // This is the new msgpack format for metadata. It is a NamedMD containing
34     // an MDTuple containing an MDString containing the msgpack data.
35     BlobType = ELF::NT_AMDGPU_METADATA;
36     auto MDN = dyn_cast<MDTuple>(NamedMD->getOperand(0));
37     if (MDN && MDN->getNumOperands()) {
38       if (auto MDS = dyn_cast<MDString>(MDN->getOperand(0)))
39         setFromMsgPackBlob(MDS->getString());
40     }
41     return;
42   }
43   BlobType = ELF::NT_AMD_PAL_METADATA;
44   NamedMD = M.getNamedMetadata("amdgpu.pal.metadata");
45   if (!NamedMD || !NamedMD->getNumOperands()) {
46     // Emit msgpack metadata by default
47     BlobType = ELF::NT_AMDGPU_METADATA;
48     return;
49   }
50   // This is the old reg=value pair format for metadata. It is a NamedMD
51   // containing an MDTuple containing a number of MDNodes each of which is an
52   // integer value, and each two integer values forms a key=value pair that we
53   // store as Registers[key]=value in the map.
54   auto Tuple = dyn_cast<MDTuple>(NamedMD->getOperand(0));
55   if (!Tuple)
56     return;
57   for (unsigned I = 0, E = Tuple->getNumOperands() & -2; I != E; I += 2) {
58     auto Key = mdconst::dyn_extract<ConstantInt>(Tuple->getOperand(I));
59     auto Val = mdconst::dyn_extract<ConstantInt>(Tuple->getOperand(I + 1));
60     if (!Key || !Val)
61       continue;
62     setRegister(Key->getZExtValue(), Val->getZExtValue());
63   }
64 }
65 
66 // Set PAL metadata from a binary blob from the applicable .note record.
67 // Returns false if bad format.  Blob must remain valid for the lifetime of the
68 // Metadata.
69 bool AMDGPUPALMetadata::setFromBlob(unsigned Type, StringRef Blob) {
70   BlobType = Type;
71   if (Type == ELF::NT_AMD_PAL_METADATA)
72     return setFromLegacyBlob(Blob);
73   return setFromMsgPackBlob(Blob);
74 }
75 
76 // Set PAL metadata from legacy (array of key=value pairs) blob.
77 bool AMDGPUPALMetadata::setFromLegacyBlob(StringRef Blob) {
78   auto Data = reinterpret_cast<const uint32_t *>(Blob.data());
79   for (unsigned I = 0; I != Blob.size() / sizeof(uint32_t) / 2; ++I)
80     setRegister(Data[I * 2], Data[I * 2 + 1]);
81   return true;
82 }
83 
84 // Set PAL metadata from msgpack blob.
85 bool AMDGPUPALMetadata::setFromMsgPackBlob(StringRef Blob) {
86   return MsgPackDoc.readFromBlob(Blob, /*Multi=*/false);
87 }
88 
89 // Given the calling convention, calculate the register number for rsrc1. In
90 // principle the register number could change in future hardware, but we know
91 // it is the same for gfx6-9 (except that LS and ES don't exist on gfx9), so
92 // we can use fixed values.
93 static unsigned getRsrc1Reg(CallingConv::ID CC) {
94   switch (CC) {
95   default:
96     return PALMD::R_2E12_COMPUTE_PGM_RSRC1;
97   case CallingConv::AMDGPU_LS:
98     return PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS;
99   case CallingConv::AMDGPU_HS:
100     return PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS;
101   case CallingConv::AMDGPU_ES:
102     return PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES;
103   case CallingConv::AMDGPU_GS:
104     return PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS;
105   case CallingConv::AMDGPU_VS:
106     return PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS;
107   case CallingConv::AMDGPU_PS:
108     return PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS;
109   }
110 }
111 
112 // Calculate the PAL metadata key for *S_SCRATCH_SIZE. It can be used
113 // with a constant offset to access any non-register shader-specific PAL
114 // metadata key.
115 static unsigned getScratchSizeKey(CallingConv::ID CC) {
116   switch (CC) {
117   case CallingConv::AMDGPU_PS:
118     return PALMD::Key::PS_SCRATCH_SIZE;
119   case CallingConv::AMDGPU_VS:
120     return PALMD::Key::VS_SCRATCH_SIZE;
121   case CallingConv::AMDGPU_GS:
122     return PALMD::Key::GS_SCRATCH_SIZE;
123   case CallingConv::AMDGPU_ES:
124     return PALMD::Key::ES_SCRATCH_SIZE;
125   case CallingConv::AMDGPU_HS:
126     return PALMD::Key::HS_SCRATCH_SIZE;
127   case CallingConv::AMDGPU_LS:
128     return PALMD::Key::LS_SCRATCH_SIZE;
129   default:
130     return PALMD::Key::CS_SCRATCH_SIZE;
131   }
132 }
133 
134 // Set the rsrc1 register in the metadata for a particular shader stage.
135 // In fact this ORs the value into any previous setting of the register.
136 void AMDGPUPALMetadata::setRsrc1(CallingConv::ID CC, unsigned Val) {
137   setRegister(getRsrc1Reg(CC), Val);
138 }
139 
140 // Set the rsrc2 register in the metadata for a particular shader stage.
141 // In fact this ORs the value into any previous setting of the register.
142 void AMDGPUPALMetadata::setRsrc2(CallingConv::ID CC, unsigned Val) {
143   setRegister(getRsrc1Reg(CC) + 1, Val);
144 }
145 
146 // Set the SPI_PS_INPUT_ENA register in the metadata.
147 // In fact this ORs the value into any previous setting of the register.
148 void AMDGPUPALMetadata::setSpiPsInputEna(unsigned Val) {
149   setRegister(PALMD::R_A1B3_SPI_PS_INPUT_ENA, Val);
150 }
151 
152 // Set the SPI_PS_INPUT_ADDR register in the metadata.
153 // In fact this ORs the value into any previous setting of the register.
154 void AMDGPUPALMetadata::setSpiPsInputAddr(unsigned Val) {
155   setRegister(PALMD::R_A1B4_SPI_PS_INPUT_ADDR, Val);
156 }
157 
158 // Get a register from the metadata, or 0 if not currently set.
159 unsigned AMDGPUPALMetadata::getRegister(unsigned Reg) {
160   auto Regs = getRegisters();
161   auto It = Regs.find(MsgPackDoc.getNode(Reg));
162   if (It == Regs.end())
163     return 0;
164   auto N = It->second;
165   if (N.getKind() != msgpack::Type::UInt)
166     return 0;
167   return N.getUInt();
168 }
169 
170 // Set a register in the metadata.
171 // In fact this ORs the value into any previous setting of the register.
172 void AMDGPUPALMetadata::setRegister(unsigned Reg, unsigned Val) {
173   if (!isLegacy()) {
174     // In the new MsgPack format, ignore register numbered >= 0x10000000. It
175     // is a PAL ABI pseudo-register in the old non-MsgPack format.
176     if (Reg >= 0x10000000)
177       return;
178   }
179   auto &N = getRegisters()[MsgPackDoc.getNode(Reg)];
180   if (N.getKind() == msgpack::Type::UInt)
181     Val |= N.getUInt();
182   N = N.getDocument()->getNode(Val);
183 }
184 
185 // Set the entry point name for one shader.
186 void AMDGPUPALMetadata::setEntryPoint(unsigned CC, StringRef Name) {
187   if (isLegacy())
188     return;
189   // Msgpack format.
190   getHwStage(CC)[".entry_point"] = MsgPackDoc.getNode(Name, /*Copy=*/true);
191 }
192 
193 // Set the number of used vgprs in the metadata. This is an optional
194 // advisory record for logging etc; wave dispatch actually uses the rsrc1
195 // register for the shader stage to determine the number of vgprs to
196 // allocate.
197 void AMDGPUPALMetadata::setNumUsedVgprs(CallingConv::ID CC, unsigned Val) {
198   if (isLegacy()) {
199     // Old non-msgpack format.
200     unsigned NumUsedVgprsKey = getScratchSizeKey(CC) +
201                                PALMD::Key::VS_NUM_USED_VGPRS -
202                                PALMD::Key::VS_SCRATCH_SIZE;
203     setRegister(NumUsedVgprsKey, Val);
204     return;
205   }
206   // Msgpack format.
207   getHwStage(CC)[".vgpr_count"] = MsgPackDoc.getNode(Val);
208 }
209 
210 // Set the number of used agprs in the metadata.
211 void AMDGPUPALMetadata::setNumUsedAgprs(CallingConv::ID CC, unsigned Val) {
212   getHwStage(CC)[".agpr_count"] = Val;
213 }
214 
215 // Set the number of used sgprs in the metadata. This is an optional advisory
216 // record for logging etc; wave dispatch actually uses the rsrc1 register for
217 // the shader stage to determine the number of sgprs to allocate.
218 void AMDGPUPALMetadata::setNumUsedSgprs(CallingConv::ID CC, unsigned Val) {
219   if (isLegacy()) {
220     // Old non-msgpack format.
221     unsigned NumUsedSgprsKey = getScratchSizeKey(CC) +
222                                PALMD::Key::VS_NUM_USED_SGPRS -
223                                PALMD::Key::VS_SCRATCH_SIZE;
224     setRegister(NumUsedSgprsKey, Val);
225     return;
226   }
227   // Msgpack format.
228   getHwStage(CC)[".sgpr_count"] = MsgPackDoc.getNode(Val);
229 }
230 
231 // Set the scratch size in the metadata.
232 void AMDGPUPALMetadata::setScratchSize(CallingConv::ID CC, unsigned Val) {
233   if (isLegacy()) {
234     // Old non-msgpack format.
235     setRegister(getScratchSizeKey(CC), Val);
236     return;
237   }
238   // Msgpack format.
239   getHwStage(CC)[".scratch_memory_size"] = MsgPackDoc.getNode(Val);
240 }
241 
242 // Set the stack frame size of a function in the metadata.
243 void AMDGPUPALMetadata::setFunctionScratchSize(StringRef FnName, unsigned Val) {
244   auto Node = getShaderFunction(FnName);
245   Node[".stack_frame_size_in_bytes"] = MsgPackDoc.getNode(Val);
246   Node[".backend_stack_size"] = MsgPackDoc.getNode(Val);
247 }
248 
249 // Set the amount of LDS used in bytes in the metadata.
250 void AMDGPUPALMetadata::setFunctionLdsSize(StringRef FnName, unsigned Val) {
251   auto Node = getShaderFunction(FnName);
252   Node[".lds_size"] = MsgPackDoc.getNode(Val);
253 }
254 
255 // Set the number of used vgprs in the metadata.
256 void AMDGPUPALMetadata::setFunctionNumUsedVgprs(StringRef FnName,
257                                                 unsigned Val) {
258   auto Node = getShaderFunction(FnName);
259   Node[".vgpr_count"] = MsgPackDoc.getNode(Val);
260 }
261 
262 // Set the number of used vgprs in the metadata.
263 void AMDGPUPALMetadata::setFunctionNumUsedSgprs(StringRef FnName,
264                                                 unsigned Val) {
265   auto Node = getShaderFunction(FnName);
266   Node[".sgpr_count"] = MsgPackDoc.getNode(Val);
267 }
268 
269 // Set the hardware register bit in PAL metadata to enable wave32 on the
270 // shader of the given calling convention.
271 void AMDGPUPALMetadata::setWave32(unsigned CC) {
272   switch (CC) {
273   case CallingConv::AMDGPU_HS:
274     setRegister(PALMD::R_A2D5_VGT_SHADER_STAGES_EN, S_028B54_HS_W32_EN(1));
275     break;
276   case CallingConv::AMDGPU_GS:
277     setRegister(PALMD::R_A2D5_VGT_SHADER_STAGES_EN, S_028B54_GS_W32_EN(1));
278     break;
279   case CallingConv::AMDGPU_VS:
280     setRegister(PALMD::R_A2D5_VGT_SHADER_STAGES_EN, S_028B54_VS_W32_EN(1));
281     break;
282   case CallingConv::AMDGPU_PS:
283     setRegister(PALMD::R_A1B6_SPI_PS_IN_CONTROL, S_0286D8_PS_W32_EN(1));
284     break;
285   case CallingConv::AMDGPU_CS:
286     setRegister(PALMD::R_2E00_COMPUTE_DISPATCH_INITIATOR,
287                 S_00B800_CS_W32_EN(1));
288     break;
289   }
290 }
291 
292 // Convert a register number to name, for display by toString().
293 // Returns nullptr if none.
294 static const char *getRegisterName(unsigned RegNum) {
295   // Table of registers.
296   static const struct RegInfo {
297     unsigned Num;
298     const char *Name;
299   } RegInfoTable[] = {
300       // Registers that code generation sets/modifies metadata for.
301       {PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS, "SPI_SHADER_PGM_RSRC1_VS"},
302       {PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS + 1, "SPI_SHADER_PGM_RSRC2_VS"},
303       {PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS, "SPI_SHADER_PGM_RSRC1_LS"},
304       {PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS + 1, "SPI_SHADER_PGM_RSRC2_LS"},
305       {PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS, "SPI_SHADER_PGM_RSRC1_HS"},
306       {PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS + 1, "SPI_SHADER_PGM_RSRC2_HS"},
307       {PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES, "SPI_SHADER_PGM_RSRC1_ES"},
308       {PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES + 1, "SPI_SHADER_PGM_RSRC2_ES"},
309       {PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS, "SPI_SHADER_PGM_RSRC1_GS"},
310       {PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS + 1, "SPI_SHADER_PGM_RSRC2_GS"},
311       {PALMD::R_2E00_COMPUTE_DISPATCH_INITIATOR, "COMPUTE_DISPATCH_INITIATOR"},
312       {PALMD::R_2E12_COMPUTE_PGM_RSRC1, "COMPUTE_PGM_RSRC1"},
313       {PALMD::R_2E12_COMPUTE_PGM_RSRC1 + 1, "COMPUTE_PGM_RSRC2"},
314       {PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS, "SPI_SHADER_PGM_RSRC1_PS"},
315       {PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS + 1, "SPI_SHADER_PGM_RSRC2_PS"},
316       {PALMD::R_A1B3_SPI_PS_INPUT_ENA, "SPI_PS_INPUT_ENA"},
317       {PALMD::R_A1B4_SPI_PS_INPUT_ADDR, "SPI_PS_INPUT_ADDR"},
318       {PALMD::R_A1B6_SPI_PS_IN_CONTROL, "SPI_PS_IN_CONTROL"},
319       {PALMD::R_A2D5_VGT_SHADER_STAGES_EN, "VGT_SHADER_STAGES_EN"},
320 
321       // Registers not known to code generation.
322       {0x2c07, "SPI_SHADER_PGM_RSRC3_PS"},
323       {0x2c46, "SPI_SHADER_PGM_RSRC3_VS"},
324       {0x2c87, "SPI_SHADER_PGM_RSRC3_GS"},
325       {0x2cc7, "SPI_SHADER_PGM_RSRC3_ES"},
326       {0x2d07, "SPI_SHADER_PGM_RSRC3_HS"},
327       {0x2d47, "SPI_SHADER_PGM_RSRC3_LS"},
328 
329       {0xa1c3, "SPI_SHADER_POS_FORMAT"},
330       {0xa1b1, "SPI_VS_OUT_CONFIG"},
331       {0xa207, "PA_CL_VS_OUT_CNTL"},
332       {0xa204, "PA_CL_CLIP_CNTL"},
333       {0xa206, "PA_CL_VTE_CNTL"},
334       {0xa2f9, "PA_SU_VTX_CNTL"},
335       {0xa293, "PA_SC_MODE_CNTL_1"},
336       {0xa2a1, "VGT_PRIMITIVEID_EN"},
337       {0x2c81, "SPI_SHADER_PGM_RSRC4_GS"},
338       {0x2e18, "COMPUTE_TMPRING_SIZE"},
339       {0xa1b5, "SPI_INTERP_CONTROL_0"},
340       {0xa1ba, "SPI_TMPRING_SIZE"},
341       {0xa1c4, "SPI_SHADER_Z_FORMAT"},
342       {0xa1c5, "SPI_SHADER_COL_FORMAT"},
343       {0xa203, "DB_SHADER_CONTROL"},
344       {0xa08f, "CB_SHADER_MASK"},
345       {0xa191, "SPI_PS_INPUT_CNTL_0"},
346       {0xa192, "SPI_PS_INPUT_CNTL_1"},
347       {0xa193, "SPI_PS_INPUT_CNTL_2"},
348       {0xa194, "SPI_PS_INPUT_CNTL_3"},
349       {0xa195, "SPI_PS_INPUT_CNTL_4"},
350       {0xa196, "SPI_PS_INPUT_CNTL_5"},
351       {0xa197, "SPI_PS_INPUT_CNTL_6"},
352       {0xa198, "SPI_PS_INPUT_CNTL_7"},
353       {0xa199, "SPI_PS_INPUT_CNTL_8"},
354       {0xa19a, "SPI_PS_INPUT_CNTL_9"},
355       {0xa19b, "SPI_PS_INPUT_CNTL_10"},
356       {0xa19c, "SPI_PS_INPUT_CNTL_11"},
357       {0xa19d, "SPI_PS_INPUT_CNTL_12"},
358       {0xa19e, "SPI_PS_INPUT_CNTL_13"},
359       {0xa19f, "SPI_PS_INPUT_CNTL_14"},
360       {0xa1a0, "SPI_PS_INPUT_CNTL_15"},
361       {0xa1a1, "SPI_PS_INPUT_CNTL_16"},
362       {0xa1a2, "SPI_PS_INPUT_CNTL_17"},
363       {0xa1a3, "SPI_PS_INPUT_CNTL_18"},
364       {0xa1a4, "SPI_PS_INPUT_CNTL_19"},
365       {0xa1a5, "SPI_PS_INPUT_CNTL_20"},
366       {0xa1a6, "SPI_PS_INPUT_CNTL_21"},
367       {0xa1a7, "SPI_PS_INPUT_CNTL_22"},
368       {0xa1a8, "SPI_PS_INPUT_CNTL_23"},
369       {0xa1a9, "SPI_PS_INPUT_CNTL_24"},
370       {0xa1aa, "SPI_PS_INPUT_CNTL_25"},
371       {0xa1ab, "SPI_PS_INPUT_CNTL_26"},
372       {0xa1ac, "SPI_PS_INPUT_CNTL_27"},
373       {0xa1ad, "SPI_PS_INPUT_CNTL_28"},
374       {0xa1ae, "SPI_PS_INPUT_CNTL_29"},
375       {0xa1af, "SPI_PS_INPUT_CNTL_30"},
376       {0xa1b0, "SPI_PS_INPUT_CNTL_31"},
377 
378       {0xa2ce, "VGT_GS_MAX_VERT_OUT"},
379       {0xa2ab, "VGT_ESGS_RING_ITEMSIZE"},
380       {0xa290, "VGT_GS_MODE"},
381       {0xa291, "VGT_GS_ONCHIP_CNTL"},
382       {0xa2d7, "VGT_GS_VERT_ITEMSIZE"},
383       {0xa2d8, "VGT_GS_VERT_ITEMSIZE_1"},
384       {0xa2d9, "VGT_GS_VERT_ITEMSIZE_2"},
385       {0xa2da, "VGT_GS_VERT_ITEMSIZE_3"},
386       {0xa298, "VGT_GSVS_RING_OFFSET_1"},
387       {0xa299, "VGT_GSVS_RING_OFFSET_2"},
388       {0xa29a, "VGT_GSVS_RING_OFFSET_3"},
389 
390       {0xa2e4, "VGT_GS_INSTANCE_CNT"},
391       {0xa297, "VGT_GS_PER_VS"},
392       {0xa29b, "VGT_GS_OUT_PRIM_TYPE"},
393       {0xa2ac, "VGT_GSVS_RING_ITEMSIZE"},
394 
395       {0xa2ad, "VGT_REUSE_OFF"},
396       {0xa1b8, "SPI_BARYC_CNTL"},
397 
398       {0x2c4c, "SPI_SHADER_USER_DATA_VS_0"},
399       {0x2c4d, "SPI_SHADER_USER_DATA_VS_1"},
400       {0x2c4e, "SPI_SHADER_USER_DATA_VS_2"},
401       {0x2c4f, "SPI_SHADER_USER_DATA_VS_3"},
402       {0x2c50, "SPI_SHADER_USER_DATA_VS_4"},
403       {0x2c51, "SPI_SHADER_USER_DATA_VS_5"},
404       {0x2c52, "SPI_SHADER_USER_DATA_VS_6"},
405       {0x2c53, "SPI_SHADER_USER_DATA_VS_7"},
406       {0x2c54, "SPI_SHADER_USER_DATA_VS_8"},
407       {0x2c55, "SPI_SHADER_USER_DATA_VS_9"},
408       {0x2c56, "SPI_SHADER_USER_DATA_VS_10"},
409       {0x2c57, "SPI_SHADER_USER_DATA_VS_11"},
410       {0x2c58, "SPI_SHADER_USER_DATA_VS_12"},
411       {0x2c59, "SPI_SHADER_USER_DATA_VS_13"},
412       {0x2c5a, "SPI_SHADER_USER_DATA_VS_14"},
413       {0x2c5b, "SPI_SHADER_USER_DATA_VS_15"},
414       {0x2c5c, "SPI_SHADER_USER_DATA_VS_16"},
415       {0x2c5d, "SPI_SHADER_USER_DATA_VS_17"},
416       {0x2c5e, "SPI_SHADER_USER_DATA_VS_18"},
417       {0x2c5f, "SPI_SHADER_USER_DATA_VS_19"},
418       {0x2c60, "SPI_SHADER_USER_DATA_VS_20"},
419       {0x2c61, "SPI_SHADER_USER_DATA_VS_21"},
420       {0x2c62, "SPI_SHADER_USER_DATA_VS_22"},
421       {0x2c63, "SPI_SHADER_USER_DATA_VS_23"},
422       {0x2c64, "SPI_SHADER_USER_DATA_VS_24"},
423       {0x2c65, "SPI_SHADER_USER_DATA_VS_25"},
424       {0x2c66, "SPI_SHADER_USER_DATA_VS_26"},
425       {0x2c67, "SPI_SHADER_USER_DATA_VS_27"},
426       {0x2c68, "SPI_SHADER_USER_DATA_VS_28"},
427       {0x2c69, "SPI_SHADER_USER_DATA_VS_29"},
428       {0x2c6a, "SPI_SHADER_USER_DATA_VS_30"},
429       {0x2c6b, "SPI_SHADER_USER_DATA_VS_31"},
430 
431       {0x2c8c, "SPI_SHADER_USER_DATA_GS_0"},
432       {0x2c8d, "SPI_SHADER_USER_DATA_GS_1"},
433       {0x2c8e, "SPI_SHADER_USER_DATA_GS_2"},
434       {0x2c8f, "SPI_SHADER_USER_DATA_GS_3"},
435       {0x2c90, "SPI_SHADER_USER_DATA_GS_4"},
436       {0x2c91, "SPI_SHADER_USER_DATA_GS_5"},
437       {0x2c92, "SPI_SHADER_USER_DATA_GS_6"},
438       {0x2c93, "SPI_SHADER_USER_DATA_GS_7"},
439       {0x2c94, "SPI_SHADER_USER_DATA_GS_8"},
440       {0x2c95, "SPI_SHADER_USER_DATA_GS_9"},
441       {0x2c96, "SPI_SHADER_USER_DATA_GS_10"},
442       {0x2c97, "SPI_SHADER_USER_DATA_GS_11"},
443       {0x2c98, "SPI_SHADER_USER_DATA_GS_12"},
444       {0x2c99, "SPI_SHADER_USER_DATA_GS_13"},
445       {0x2c9a, "SPI_SHADER_USER_DATA_GS_14"},
446       {0x2c9b, "SPI_SHADER_USER_DATA_GS_15"},
447       {0x2c9c, "SPI_SHADER_USER_DATA_GS_16"},
448       {0x2c9d, "SPI_SHADER_USER_DATA_GS_17"},
449       {0x2c9e, "SPI_SHADER_USER_DATA_GS_18"},
450       {0x2c9f, "SPI_SHADER_USER_DATA_GS_19"},
451       {0x2ca0, "SPI_SHADER_USER_DATA_GS_20"},
452       {0x2ca1, "SPI_SHADER_USER_DATA_GS_21"},
453       {0x2ca2, "SPI_SHADER_USER_DATA_GS_22"},
454       {0x2ca3, "SPI_SHADER_USER_DATA_GS_23"},
455       {0x2ca4, "SPI_SHADER_USER_DATA_GS_24"},
456       {0x2ca5, "SPI_SHADER_USER_DATA_GS_25"},
457       {0x2ca6, "SPI_SHADER_USER_DATA_GS_26"},
458       {0x2ca7, "SPI_SHADER_USER_DATA_GS_27"},
459       {0x2ca8, "SPI_SHADER_USER_DATA_GS_28"},
460       {0x2ca9, "SPI_SHADER_USER_DATA_GS_29"},
461       {0x2caa, "SPI_SHADER_USER_DATA_GS_30"},
462       {0x2cab, "SPI_SHADER_USER_DATA_GS_31"},
463 
464       {0x2ccc, "SPI_SHADER_USER_DATA_ES_0"},
465       {0x2ccd, "SPI_SHADER_USER_DATA_ES_1"},
466       {0x2cce, "SPI_SHADER_USER_DATA_ES_2"},
467       {0x2ccf, "SPI_SHADER_USER_DATA_ES_3"},
468       {0x2cd0, "SPI_SHADER_USER_DATA_ES_4"},
469       {0x2cd1, "SPI_SHADER_USER_DATA_ES_5"},
470       {0x2cd2, "SPI_SHADER_USER_DATA_ES_6"},
471       {0x2cd3, "SPI_SHADER_USER_DATA_ES_7"},
472       {0x2cd4, "SPI_SHADER_USER_DATA_ES_8"},
473       {0x2cd5, "SPI_SHADER_USER_DATA_ES_9"},
474       {0x2cd6, "SPI_SHADER_USER_DATA_ES_10"},
475       {0x2cd7, "SPI_SHADER_USER_DATA_ES_11"},
476       {0x2cd8, "SPI_SHADER_USER_DATA_ES_12"},
477       {0x2cd9, "SPI_SHADER_USER_DATA_ES_13"},
478       {0x2cda, "SPI_SHADER_USER_DATA_ES_14"},
479       {0x2cdb, "SPI_SHADER_USER_DATA_ES_15"},
480       {0x2cdc, "SPI_SHADER_USER_DATA_ES_16"},
481       {0x2cdd, "SPI_SHADER_USER_DATA_ES_17"},
482       {0x2cde, "SPI_SHADER_USER_DATA_ES_18"},
483       {0x2cdf, "SPI_SHADER_USER_DATA_ES_19"},
484       {0x2ce0, "SPI_SHADER_USER_DATA_ES_20"},
485       {0x2ce1, "SPI_SHADER_USER_DATA_ES_21"},
486       {0x2ce2, "SPI_SHADER_USER_DATA_ES_22"},
487       {0x2ce3, "SPI_SHADER_USER_DATA_ES_23"},
488       {0x2ce4, "SPI_SHADER_USER_DATA_ES_24"},
489       {0x2ce5, "SPI_SHADER_USER_DATA_ES_25"},
490       {0x2ce6, "SPI_SHADER_USER_DATA_ES_26"},
491       {0x2ce7, "SPI_SHADER_USER_DATA_ES_27"},
492       {0x2ce8, "SPI_SHADER_USER_DATA_ES_28"},
493       {0x2ce9, "SPI_SHADER_USER_DATA_ES_29"},
494       {0x2cea, "SPI_SHADER_USER_DATA_ES_30"},
495       {0x2ceb, "SPI_SHADER_USER_DATA_ES_31"},
496 
497       {0x2c0c, "SPI_SHADER_USER_DATA_PS_0"},
498       {0x2c0d, "SPI_SHADER_USER_DATA_PS_1"},
499       {0x2c0e, "SPI_SHADER_USER_DATA_PS_2"},
500       {0x2c0f, "SPI_SHADER_USER_DATA_PS_3"},
501       {0x2c10, "SPI_SHADER_USER_DATA_PS_4"},
502       {0x2c11, "SPI_SHADER_USER_DATA_PS_5"},
503       {0x2c12, "SPI_SHADER_USER_DATA_PS_6"},
504       {0x2c13, "SPI_SHADER_USER_DATA_PS_7"},
505       {0x2c14, "SPI_SHADER_USER_DATA_PS_8"},
506       {0x2c15, "SPI_SHADER_USER_DATA_PS_9"},
507       {0x2c16, "SPI_SHADER_USER_DATA_PS_10"},
508       {0x2c17, "SPI_SHADER_USER_DATA_PS_11"},
509       {0x2c18, "SPI_SHADER_USER_DATA_PS_12"},
510       {0x2c19, "SPI_SHADER_USER_DATA_PS_13"},
511       {0x2c1a, "SPI_SHADER_USER_DATA_PS_14"},
512       {0x2c1b, "SPI_SHADER_USER_DATA_PS_15"},
513       {0x2c1c, "SPI_SHADER_USER_DATA_PS_16"},
514       {0x2c1d, "SPI_SHADER_USER_DATA_PS_17"},
515       {0x2c1e, "SPI_SHADER_USER_DATA_PS_18"},
516       {0x2c1f, "SPI_SHADER_USER_DATA_PS_19"},
517       {0x2c20, "SPI_SHADER_USER_DATA_PS_20"},
518       {0x2c21, "SPI_SHADER_USER_DATA_PS_21"},
519       {0x2c22, "SPI_SHADER_USER_DATA_PS_22"},
520       {0x2c23, "SPI_SHADER_USER_DATA_PS_23"},
521       {0x2c24, "SPI_SHADER_USER_DATA_PS_24"},
522       {0x2c25, "SPI_SHADER_USER_DATA_PS_25"},
523       {0x2c26, "SPI_SHADER_USER_DATA_PS_26"},
524       {0x2c27, "SPI_SHADER_USER_DATA_PS_27"},
525       {0x2c28, "SPI_SHADER_USER_DATA_PS_28"},
526       {0x2c29, "SPI_SHADER_USER_DATA_PS_29"},
527       {0x2c2a, "SPI_SHADER_USER_DATA_PS_30"},
528       {0x2c2b, "SPI_SHADER_USER_DATA_PS_31"},
529 
530       {0x2e40, "COMPUTE_USER_DATA_0"},
531       {0x2e41, "COMPUTE_USER_DATA_1"},
532       {0x2e42, "COMPUTE_USER_DATA_2"},
533       {0x2e43, "COMPUTE_USER_DATA_3"},
534       {0x2e44, "COMPUTE_USER_DATA_4"},
535       {0x2e45, "COMPUTE_USER_DATA_5"},
536       {0x2e46, "COMPUTE_USER_DATA_6"},
537       {0x2e47, "COMPUTE_USER_DATA_7"},
538       {0x2e48, "COMPUTE_USER_DATA_8"},
539       {0x2e49, "COMPUTE_USER_DATA_9"},
540       {0x2e4a, "COMPUTE_USER_DATA_10"},
541       {0x2e4b, "COMPUTE_USER_DATA_11"},
542       {0x2e4c, "COMPUTE_USER_DATA_12"},
543       {0x2e4d, "COMPUTE_USER_DATA_13"},
544       {0x2e4e, "COMPUTE_USER_DATA_14"},
545       {0x2e4f, "COMPUTE_USER_DATA_15"},
546 
547       {0x2e07, "COMPUTE_NUM_THREAD_X"},
548       {0x2e08, "COMPUTE_NUM_THREAD_Y"},
549       {0x2e09, "COMPUTE_NUM_THREAD_Z"},
550       {0xa2db, "VGT_TF_PARAM"},
551       {0xa2d6, "VGT_LS_HS_CONFIG"},
552       {0xa287, "VGT_HOS_MIN_TESS_LEVEL"},
553       {0xa286, "VGT_HOS_MAX_TESS_LEVEL"},
554       {0xa2f8, "PA_SC_AA_CONFIG"},
555       {0xa310, "PA_SC_SHADER_CONTROL"},
556       {0xa313, "PA_SC_CONSERVATIVE_RASTERIZATION_CNTL"},
557 
558       {0x2d0c, "SPI_SHADER_USER_DATA_HS_0"},
559       {0x2d0d, "SPI_SHADER_USER_DATA_HS_1"},
560       {0x2d0e, "SPI_SHADER_USER_DATA_HS_2"},
561       {0x2d0f, "SPI_SHADER_USER_DATA_HS_3"},
562       {0x2d10, "SPI_SHADER_USER_DATA_HS_4"},
563       {0x2d11, "SPI_SHADER_USER_DATA_HS_5"},
564       {0x2d12, "SPI_SHADER_USER_DATA_HS_6"},
565       {0x2d13, "SPI_SHADER_USER_DATA_HS_7"},
566       {0x2d14, "SPI_SHADER_USER_DATA_HS_8"},
567       {0x2d15, "SPI_SHADER_USER_DATA_HS_9"},
568       {0x2d16, "SPI_SHADER_USER_DATA_HS_10"},
569       {0x2d17, "SPI_SHADER_USER_DATA_HS_11"},
570       {0x2d18, "SPI_SHADER_USER_DATA_HS_12"},
571       {0x2d19, "SPI_SHADER_USER_DATA_HS_13"},
572       {0x2d1a, "SPI_SHADER_USER_DATA_HS_14"},
573       {0x2d1b, "SPI_SHADER_USER_DATA_HS_15"},
574       {0x2d1c, "SPI_SHADER_USER_DATA_HS_16"},
575       {0x2d1d, "SPI_SHADER_USER_DATA_HS_17"},
576       {0x2d1e, "SPI_SHADER_USER_DATA_HS_18"},
577       {0x2d1f, "SPI_SHADER_USER_DATA_HS_19"},
578       {0x2d20, "SPI_SHADER_USER_DATA_HS_20"},
579       {0x2d21, "SPI_SHADER_USER_DATA_HS_21"},
580       {0x2d22, "SPI_SHADER_USER_DATA_HS_22"},
581       {0x2d23, "SPI_SHADER_USER_DATA_HS_23"},
582       {0x2d24, "SPI_SHADER_USER_DATA_HS_24"},
583       {0x2d25, "SPI_SHADER_USER_DATA_HS_25"},
584       {0x2d26, "SPI_SHADER_USER_DATA_HS_26"},
585       {0x2d27, "SPI_SHADER_USER_DATA_HS_27"},
586       {0x2d28, "SPI_SHADER_USER_DATA_HS_28"},
587       {0x2d29, "SPI_SHADER_USER_DATA_HS_29"},
588       {0x2d2a, "SPI_SHADER_USER_DATA_HS_30"},
589       {0x2d2b, "SPI_SHADER_USER_DATA_HS_31"},
590 
591       {0x2d4c, "SPI_SHADER_USER_DATA_LS_0"},
592       {0x2d4d, "SPI_SHADER_USER_DATA_LS_1"},
593       {0x2d4e, "SPI_SHADER_USER_DATA_LS_2"},
594       {0x2d4f, "SPI_SHADER_USER_DATA_LS_3"},
595       {0x2d50, "SPI_SHADER_USER_DATA_LS_4"},
596       {0x2d51, "SPI_SHADER_USER_DATA_LS_5"},
597       {0x2d52, "SPI_SHADER_USER_DATA_LS_6"},
598       {0x2d53, "SPI_SHADER_USER_DATA_LS_7"},
599       {0x2d54, "SPI_SHADER_USER_DATA_LS_8"},
600       {0x2d55, "SPI_SHADER_USER_DATA_LS_9"},
601       {0x2d56, "SPI_SHADER_USER_DATA_LS_10"},
602       {0x2d57, "SPI_SHADER_USER_DATA_LS_11"},
603       {0x2d58, "SPI_SHADER_USER_DATA_LS_12"},
604       {0x2d59, "SPI_SHADER_USER_DATA_LS_13"},
605       {0x2d5a, "SPI_SHADER_USER_DATA_LS_14"},
606       {0x2d5b, "SPI_SHADER_USER_DATA_LS_15"},
607 
608       {0xa2aa, "IA_MULTI_VGT_PARAM"},
609       {0xa2a5, "VGT_GS_MAX_PRIMS_PER_SUBGROUP"},
610       {0xa2e6, "VGT_STRMOUT_BUFFER_CONFIG"},
611       {0xa2e5, "VGT_STRMOUT_CONFIG"},
612       {0xa2b5, "VGT_STRMOUT_VTX_STRIDE_0"},
613       {0xa2b9, "VGT_STRMOUT_VTX_STRIDE_1"},
614       {0xa2bd, "VGT_STRMOUT_VTX_STRIDE_2"},
615       {0xa2c1, "VGT_STRMOUT_VTX_STRIDE_3"},
616       {0xa316, "VGT_VERTEX_REUSE_BLOCK_CNTL"},
617 
618       {0x2e28, "COMPUTE_PGM_RSRC3"},
619       {0x2e2a, "COMPUTE_SHADER_CHKSUM"},
620       {0x2e24, "COMPUTE_USER_ACCUM_0"},
621       {0x2e25, "COMPUTE_USER_ACCUM_1"},
622       {0x2e26, "COMPUTE_USER_ACCUM_2"},
623       {0x2e27, "COMPUTE_USER_ACCUM_3"},
624       {0xa1ff, "GE_MAX_OUTPUT_PER_SUBGROUP"},
625       {0xa2d3, "GE_NGG_SUBGRP_CNTL"},
626       {0xc25f, "GE_STEREO_CNTL"},
627       {0xc262, "GE_USER_VGPR_EN"},
628       {0xc258, "IA_MULTI_VGT_PARAM_PIPED"},
629       {0xa210, "PA_STEREO_CNTL"},
630       {0xa1c2, "SPI_SHADER_IDX_FORMAT"},
631       {0x2c80, "SPI_SHADER_PGM_CHKSUM_GS"},
632       {0x2d00, "SPI_SHADER_PGM_CHKSUM_HS"},
633       {0x2c06, "SPI_SHADER_PGM_CHKSUM_PS"},
634       {0x2c45, "SPI_SHADER_PGM_CHKSUM_VS"},
635       {0x2c88, "SPI_SHADER_PGM_LO_GS"},
636       {0x2cb2, "SPI_SHADER_USER_ACCUM_ESGS_0"},
637       {0x2cb3, "SPI_SHADER_USER_ACCUM_ESGS_1"},
638       {0x2cb4, "SPI_SHADER_USER_ACCUM_ESGS_2"},
639       {0x2cb5, "SPI_SHADER_USER_ACCUM_ESGS_3"},
640       {0x2d32, "SPI_SHADER_USER_ACCUM_LSHS_0"},
641       {0x2d33, "SPI_SHADER_USER_ACCUM_LSHS_1"},
642       {0x2d34, "SPI_SHADER_USER_ACCUM_LSHS_2"},
643       {0x2d35, "SPI_SHADER_USER_ACCUM_LSHS_3"},
644       {0x2c32, "SPI_SHADER_USER_ACCUM_PS_0"},
645       {0x2c33, "SPI_SHADER_USER_ACCUM_PS_1"},
646       {0x2c34, "SPI_SHADER_USER_ACCUM_PS_2"},
647       {0x2c35, "SPI_SHADER_USER_ACCUM_PS_3"},
648       {0x2c72, "SPI_SHADER_USER_ACCUM_VS_0"},
649       {0x2c73, "SPI_SHADER_USER_ACCUM_VS_1"},
650       {0x2c74, "SPI_SHADER_USER_ACCUM_VS_2"},
651       {0x2c75, "SPI_SHADER_USER_ACCUM_VS_3"},
652 
653       {0, nullptr}};
654   auto Entry = RegInfoTable;
655   for (; Entry->Num && Entry->Num != RegNum; ++Entry)
656     ;
657   return Entry->Name;
658 }
659 
660 // Convert the accumulated PAL metadata into an asm directive.
661 void AMDGPUPALMetadata::toString(std::string &String) {
662   String.clear();
663   if (!BlobType)
664     return;
665   raw_string_ostream Stream(String);
666   if (isLegacy()) {
667     if (MsgPackDoc.getRoot().getKind() == msgpack::Type::Nil)
668       return;
669     // Old linear reg=val format.
670     Stream << '\t' << AMDGPU::PALMD::AssemblerDirective << ' ';
671     auto Regs = getRegisters();
672     for (auto I = Regs.begin(), E = Regs.end(); I != E; ++I) {
673       if (I != Regs.begin())
674         Stream << ',';
675       unsigned Reg = I->first.getUInt();
676       unsigned Val = I->second.getUInt();
677       Stream << "0x" << Twine::utohexstr(Reg) << ",0x" << Twine::utohexstr(Val);
678     }
679     Stream << '\n';
680     return;
681   }
682 
683   // New msgpack-based format -- output as YAML (with unsigned numbers in hex),
684   // but first change the registers map to use names.
685   MsgPackDoc.setHexMode();
686   auto &RegsObj = refRegisters();
687   auto OrigRegs = RegsObj.getMap();
688   RegsObj = MsgPackDoc.getMapNode();
689   for (auto I : OrigRegs) {
690     auto Key = I.first;
691     if (const char *RegName = getRegisterName(Key.getUInt())) {
692       std::string KeyName = Key.toString();
693       KeyName += " (";
694       KeyName += RegName;
695       KeyName += ')';
696       Key = MsgPackDoc.getNode(KeyName, /*Copy=*/true);
697     }
698     RegsObj.getMap()[Key] = I.second;
699   }
700 
701   // Output as YAML.
702   Stream << '\t' << AMDGPU::PALMD::AssemblerDirectiveBegin << '\n';
703   MsgPackDoc.toYAML(Stream);
704   Stream << '\t' << AMDGPU::PALMD::AssemblerDirectiveEnd << '\n';
705 
706   // Restore original registers map.
707   RegsObj = OrigRegs;
708 }
709 
710 // Convert the accumulated PAL metadata into a binary blob for writing as
711 // a .note record of the specified AMD type. Returns an empty blob if
712 // there is no PAL metadata,
713 void AMDGPUPALMetadata::toBlob(unsigned Type, std::string &Blob) {
714   if (Type == ELF::NT_AMD_PAL_METADATA)
715     toLegacyBlob(Blob);
716   else if (Type)
717     toMsgPackBlob(Blob);
718 }
719 
720 void AMDGPUPALMetadata::toLegacyBlob(std::string &Blob) {
721   Blob.clear();
722   auto Registers = getRegisters();
723   if (Registers.getMap().empty())
724     return;
725   raw_string_ostream OS(Blob);
726   support::endian::Writer EW(OS, llvm::endianness::little);
727   for (auto I : Registers.getMap()) {
728     EW.write(uint32_t(I.first.getUInt()));
729     EW.write(uint32_t(I.second.getUInt()));
730   }
731 }
732 
733 void AMDGPUPALMetadata::toMsgPackBlob(std::string &Blob) {
734   Blob.clear();
735   MsgPackDoc.writeToBlob(Blob);
736 }
737 
738 // Set PAL metadata from YAML text. Returns false if failed.
739 bool AMDGPUPALMetadata::setFromString(StringRef S) {
740   BlobType = ELF::NT_AMDGPU_METADATA;
741   if (!MsgPackDoc.fromYAML(S))
742     return false;
743 
744   // In the registers map, some keys may be of the form "0xa191
745   // (SPI_PS_INPUT_CNTL_0)", in which case the YAML input code made it a
746   // string. We need to turn it into a number.
747   auto &RegsObj = refRegisters();
748   auto OrigRegs = RegsObj;
749   RegsObj = MsgPackDoc.getMapNode();
750   Registers = RegsObj.getMap();
751   bool Ok = true;
752   for (auto I : OrigRegs.getMap()) {
753     auto Key = I.first;
754     if (Key.getKind() == msgpack::Type::String) {
755       StringRef S = Key.getString();
756       uint64_t Val;
757       if (S.consumeInteger(0, Val)) {
758         Ok = false;
759         errs() << "Unrecognized PAL metadata register key '" << S << "'\n";
760         continue;
761       }
762       Key = MsgPackDoc.getNode(uint64_t(Val));
763     }
764     Registers.getMap()[Key] = I.second;
765   }
766   return Ok;
767 }
768 
769 // Reference (create if necessary) the node for the registers map.
770 msgpack::DocNode &AMDGPUPALMetadata::refRegisters() {
771   auto &N =
772       MsgPackDoc.getRoot()
773           .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
774           .getArray(/*Convert=*/true)[0]
775           .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".registers")];
776   N.getMap(/*Convert=*/true);
777   return N;
778 }
779 
780 // Get (create if necessary) the registers map.
781 msgpack::MapDocNode AMDGPUPALMetadata::getRegisters() {
782   if (Registers.isEmpty())
783     Registers = refRegisters();
784   return Registers.getMap();
785 }
786 
787 // Reference (create if necessary) the node for the shader functions map.
788 msgpack::DocNode &AMDGPUPALMetadata::refShaderFunctions() {
789   auto &N =
790       MsgPackDoc.getRoot()
791           .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
792           .getArray(/*Convert=*/true)[0]
793           .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".shader_functions")];
794   N.getMap(/*Convert=*/true);
795   return N;
796 }
797 
798 // Get (create if necessary) the shader functions map.
799 msgpack::MapDocNode AMDGPUPALMetadata::getShaderFunctions() {
800   if (ShaderFunctions.isEmpty())
801     ShaderFunctions = refShaderFunctions();
802   return ShaderFunctions.getMap();
803 }
804 
805 // Get (create if necessary) a function in the shader functions map.
806 msgpack::MapDocNode AMDGPUPALMetadata::getShaderFunction(StringRef Name) {
807   auto Functions = getShaderFunctions();
808   return Functions[Name].getMap(/*Convert=*/true);
809 }
810 
811 msgpack::DocNode &AMDGPUPALMetadata::refComputeRegisters() {
812   auto &N =
813       MsgPackDoc.getRoot()
814           .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
815           .getArray(/*Convert=*/true)[0]
816           .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".compute_registers")];
817   N.getMap(/*Convert=*/true);
818   return N;
819 }
820 
821 msgpack::MapDocNode AMDGPUPALMetadata::getComputeRegisters() {
822   if (ComputeRegisters.isEmpty())
823     ComputeRegisters = refComputeRegisters();
824   return ComputeRegisters.getMap();
825 }
826 
827 msgpack::DocNode &AMDGPUPALMetadata::refGraphicsRegisters() {
828   auto &N =
829       MsgPackDoc.getRoot()
830           .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
831           .getArray(/*Convert=*/true)[0]
832           .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".graphics_registers")];
833   N.getMap(/*Convert=*/true);
834   return N;
835 }
836 
837 msgpack::MapDocNode AMDGPUPALMetadata::getGraphicsRegisters() {
838   if (GraphicsRegisters.isEmpty())
839     GraphicsRegisters = refGraphicsRegisters();
840   return GraphicsRegisters.getMap();
841 }
842 
843 // Return the PAL metadata hardware shader stage name.
844 static const char *getStageName(CallingConv::ID CC) {
845   switch (CC) {
846   case CallingConv::AMDGPU_PS:
847     return ".ps";
848   case CallingConv::AMDGPU_VS:
849     return ".vs";
850   case CallingConv::AMDGPU_GS:
851     return ".gs";
852   case CallingConv::AMDGPU_ES:
853     return ".es";
854   case CallingConv::AMDGPU_HS:
855     return ".hs";
856   case CallingConv::AMDGPU_LS:
857     return ".ls";
858   case CallingConv::AMDGPU_Gfx:
859     llvm_unreachable("Callable shader has no hardware stage");
860   default:
861     return ".cs";
862   }
863 }
864 
865 msgpack::DocNode &AMDGPUPALMetadata::refHwStage() {
866   auto &N =
867       MsgPackDoc.getRoot()
868           .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
869           .getArray(/*Convert=*/true)[0]
870           .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".hardware_stages")];
871   N.getMap(/*Convert=*/true);
872   return N;
873 }
874 
875 // Get (create if necessary) the .hardware_stages entry for the given calling
876 // convention.
877 msgpack::MapDocNode AMDGPUPALMetadata::getHwStage(unsigned CC) {
878   if (HwStages.isEmpty())
879     HwStages = refHwStage();
880   return HwStages.getMap()[getStageName(CC)].getMap(/*Convert=*/true);
881 }
882 
883 // Get .note record vendor name of metadata blob to be emitted.
884 const char *AMDGPUPALMetadata::getVendor() const {
885   return isLegacy() ? ElfNote::NoteNameV2 : ElfNote::NoteNameV3;
886 }
887 
888 // Get .note record type of metadata blob to be emitted:
889 // ELF::NT_AMD_PAL_METADATA (legacy key=val format), or
890 // ELF::NT_AMDGPU_METADATA (MsgPack format), or
891 // 0 (no PAL metadata).
892 unsigned AMDGPUPALMetadata::getType() const {
893   return BlobType;
894 }
895 
896 // Return whether the blob type is legacy PAL metadata.
897 bool AMDGPUPALMetadata::isLegacy() const {
898   return BlobType == ELF::NT_AMD_PAL_METADATA;
899 }
900 
901 // Set legacy PAL metadata format.
902 void AMDGPUPALMetadata::setLegacy() {
903   BlobType = ELF::NT_AMD_PAL_METADATA;
904 }
905 
906 // Erase all PAL metadata.
907 void AMDGPUPALMetadata::reset() {
908   MsgPackDoc.clear();
909   Registers = MsgPackDoc.getEmptyNode();
910   HwStages = MsgPackDoc.getEmptyNode();
911   ShaderFunctions = MsgPackDoc.getEmptyNode();
912 }
913 
914 unsigned AMDGPUPALMetadata::getPALVersion(unsigned idx) {
915   assert(idx < 2 &&
916          "illegal index to PAL version - should be 0 (major) or 1 (minor)");
917   if (!VersionChecked) {
918     if (Version.isEmpty()) {
919       auto &M = MsgPackDoc.getRoot().getMap(/*Convert=*/true);
920       auto I = M.find(MsgPackDoc.getNode("amdpal.version"));
921       if (I != M.end())
922         Version = I->second;
923     }
924     VersionChecked = true;
925   }
926   if (Version.isEmpty())
927     // Default to 2.6 if there's no version info
928     return idx ? 6 : 2;
929   return Version.getArray()[idx].getUInt();
930 }
931 
932 unsigned AMDGPUPALMetadata::getPALMajorVersion() { return getPALVersion(0); }
933 
934 unsigned AMDGPUPALMetadata::getPALMinorVersion() { return getPALVersion(1); }
935 
936 // Set the field in a given .hardware_stages entry
937 void AMDGPUPALMetadata::setHwStage(unsigned CC, StringRef field, unsigned Val) {
938   getHwStage(CC)[field] = Val;
939 }
940 
941 void AMDGPUPALMetadata::setHwStage(unsigned CC, StringRef field, bool Val) {
942   getHwStage(CC)[field] = Val;
943 }
944 
945 void AMDGPUPALMetadata::setComputeRegisters(StringRef field, unsigned Val) {
946   getComputeRegisters()[field] = Val;
947 }
948 
949 void AMDGPUPALMetadata::setComputeRegisters(StringRef field, bool Val) {
950   getComputeRegisters()[field] = Val;
951 }
952 
953 msgpack::DocNode *AMDGPUPALMetadata::refComputeRegister(StringRef field) {
954   auto M = getComputeRegisters();
955   auto I = M.find(field);
956   return I == M.end() ? nullptr : &I->second;
957 }
958 
959 bool AMDGPUPALMetadata::checkComputeRegisters(StringRef field, unsigned Val) {
960   if (auto N = refComputeRegister(field))
961     return N->getUInt() == Val;
962   return false;
963 }
964 
965 bool AMDGPUPALMetadata::checkComputeRegisters(StringRef field, bool Val) {
966   if (auto N = refComputeRegister(field))
967     return N->getBool() == Val;
968   return false;
969 }
970 
971 void AMDGPUPALMetadata::setGraphicsRegisters(StringRef field, unsigned Val) {
972   getGraphicsRegisters()[field] = Val;
973 }
974 
975 void AMDGPUPALMetadata::setGraphicsRegisters(StringRef field, bool Val) {
976   getGraphicsRegisters()[field] = Val;
977 }
978 
979 void AMDGPUPALMetadata::setGraphicsRegisters(StringRef field1, StringRef field2,
980                                              unsigned Val) {
981   getGraphicsRegisters()[field1].getMap(true)[field2] = Val;
982 }
983 
984 void AMDGPUPALMetadata::setGraphicsRegisters(StringRef field1, StringRef field2,
985                                              bool Val) {
986   getGraphicsRegisters()[field1].getMap(true)[field2] = Val;
987 }
988