xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDGPUPALMetadata.cpp (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
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 "AMDGPU.h"
19 #include "AMDGPUAsmPrinter.h"
20 #include "MCTargetDesc/AMDGPUTargetStreamer.h"
21 #include "SIDefines.h"
22 #include "llvm/BinaryFormat/ELF.h"
23 #include "llvm/IR/CallingConv.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Support/AMDGPUMetadata.h"
27 #include "llvm/Support/EndianStream.h"
28 
29 using namespace llvm;
30 using namespace llvm::AMDGPU;
31 
32 // Read the PAL metadata from IR metadata, where it was put by the frontend.
33 void AMDGPUPALMetadata::readFromIR(Module &M) {
34   auto NamedMD = M.getNamedMetadata("amdgpu.pal.metadata.msgpack");
35   if (NamedMD && NamedMD->getNumOperands()) {
36     // This is the new msgpack format for metadata. It is a NamedMD containing
37     // an MDTuple containing an MDString containing the msgpack data.
38     BlobType = ELF::NT_AMDGPU_METADATA;
39     auto MDN = dyn_cast<MDTuple>(NamedMD->getOperand(0));
40     if (MDN && MDN->getNumOperands()) {
41       if (auto MDS = dyn_cast<MDString>(MDN->getOperand(0)))
42         setFromMsgPackBlob(MDS->getString());
43     }
44     return;
45   }
46   BlobType = ELF::NT_AMD_AMDGPU_PAL_METADATA;
47   NamedMD = M.getNamedMetadata("amdgpu.pal.metadata");
48   if (!NamedMD || !NamedMD->getNumOperands())
49     return;
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_AMDGPU_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   msgpack::Reader Reader(Blob);
87   return MsgPackDoc.readFromBlob(Blob, /*Multi=*/false);
88 }
89 
90 // Given the calling convention, calculate the register number for rsrc1. In
91 // principle the register number could change in future hardware, but we know
92 // it is the same for gfx6-9 (except that LS and ES don't exist on gfx9), so
93 // we can use fixed values.
94 static unsigned getRsrc1Reg(CallingConv::ID CC) {
95   switch (CC) {
96   default:
97     return PALMD::R_2E12_COMPUTE_PGM_RSRC1;
98   case CallingConv::AMDGPU_LS:
99     return PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS;
100   case CallingConv::AMDGPU_HS:
101     return PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS;
102   case CallingConv::AMDGPU_ES:
103     return PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES;
104   case CallingConv::AMDGPU_GS:
105     return PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS;
106   case CallingConv::AMDGPU_VS:
107     return PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS;
108   case CallingConv::AMDGPU_PS:
109     return PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS;
110   }
111 }
112 
113 // Calculate the PAL metadata key for *S_SCRATCH_SIZE. It can be used
114 // with a constant offset to access any non-register shader-specific PAL
115 // metadata key.
116 static unsigned getScratchSizeKey(CallingConv::ID CC) {
117   switch (CC) {
118   case CallingConv::AMDGPU_PS:
119     return PALMD::Key::PS_SCRATCH_SIZE;
120   case CallingConv::AMDGPU_VS:
121     return PALMD::Key::VS_SCRATCH_SIZE;
122   case CallingConv::AMDGPU_GS:
123     return PALMD::Key::GS_SCRATCH_SIZE;
124   case CallingConv::AMDGPU_ES:
125     return PALMD::Key::ES_SCRATCH_SIZE;
126   case CallingConv::AMDGPU_HS:
127     return PALMD::Key::HS_SCRATCH_SIZE;
128   case CallingConv::AMDGPU_LS:
129     return PALMD::Key::LS_SCRATCH_SIZE;
130   default:
131     return PALMD::Key::CS_SCRATCH_SIZE;
132   }
133 }
134 
135 // Set the rsrc1 register in the metadata for a particular shader stage.
136 // In fact this ORs the value into any previous setting of the register.
137 void AMDGPUPALMetadata::setRsrc1(CallingConv::ID CC, unsigned Val) {
138   setRegister(getRsrc1Reg(CC), Val);
139 }
140 
141 // Set the rsrc2 register in the metadata for a particular shader stage.
142 // In fact this ORs the value into any previous setting of the register.
143 void AMDGPUPALMetadata::setRsrc2(CallingConv::ID CC, unsigned Val) {
144   setRegister(getRsrc1Reg(CC) + 1, Val);
145 }
146 
147 // Set the SPI_PS_INPUT_ENA register in the metadata.
148 // In fact this ORs the value into any previous setting of the register.
149 void AMDGPUPALMetadata::setSpiPsInputEna(unsigned Val) {
150   setRegister(PALMD::R_A1B3_SPI_PS_INPUT_ENA, Val);
151 }
152 
153 // Set the SPI_PS_INPUT_ADDR register in the metadata.
154 // In fact this ORs the value into any previous setting of the register.
155 void AMDGPUPALMetadata::setSpiPsInputAddr(unsigned Val) {
156   setRegister(PALMD::R_A1B4_SPI_PS_INPUT_ADDR, Val);
157 }
158 
159 // Get a register from the metadata, or 0 if not currently set.
160 unsigned AMDGPUPALMetadata::getRegister(unsigned Reg) {
161   auto Regs = getRegisters();
162   auto It = Regs.find(MsgPackDoc.getNode(Reg));
163   if (It == Regs.end())
164     return 0;
165   auto N = It->second;
166   if (N.getKind() != msgpack::Type::UInt)
167     return 0;
168   return N.getUInt();
169 }
170 
171 // Set a register in the metadata.
172 // In fact this ORs the value into any previous setting of the register.
173 void AMDGPUPALMetadata::setRegister(unsigned Reg, unsigned Val) {
174   if (!isLegacy()) {
175     // In the new MsgPack format, ignore register numbered >= 0x10000000. It
176     // is a PAL ABI pseudo-register in the old non-MsgPack format.
177     if (Reg >= 0x10000000)
178       return;
179   }
180   auto &N = getRegisters()[MsgPackDoc.getNode(Reg)];
181   if (N.getKind() == msgpack::Type::UInt)
182     Val |= N.getUInt();
183   N = N.getDocument()->getNode(Val);
184 }
185 
186 // Set the entry point name for one shader.
187 void AMDGPUPALMetadata::setEntryPoint(unsigned CC, StringRef Name) {
188   if (isLegacy())
189     return;
190   // Msgpack format.
191   getHwStage(CC)[".entry_point"] = MsgPackDoc.getNode(Name, /*Copy=*/true);
192 }
193 
194 // Set the number of used vgprs in the metadata. This is an optional
195 // advisory record for logging etc; wave dispatch actually uses the rsrc1
196 // register for the shader stage to determine the number of vgprs to
197 // allocate.
198 void AMDGPUPALMetadata::setNumUsedVgprs(CallingConv::ID CC, unsigned Val) {
199   if (isLegacy()) {
200     // Old non-msgpack format.
201     unsigned NumUsedVgprsKey = getScratchSizeKey(CC) +
202                                PALMD::Key::VS_NUM_USED_VGPRS -
203                                PALMD::Key::VS_SCRATCH_SIZE;
204     setRegister(NumUsedVgprsKey, Val);
205     return;
206   }
207   // Msgpack format.
208   getHwStage(CC)[".vgpr_count"] = MsgPackDoc.getNode(Val);
209 }
210 
211 // Set the number of used sgprs in the metadata. This is an optional advisory
212 // record for logging etc; wave dispatch actually uses the rsrc1 register for
213 // the shader stage to determine the number of sgprs to allocate.
214 void AMDGPUPALMetadata::setNumUsedSgprs(CallingConv::ID CC, unsigned Val) {
215   if (isLegacy()) {
216     // Old non-msgpack format.
217     unsigned NumUsedSgprsKey = getScratchSizeKey(CC) +
218                                PALMD::Key::VS_NUM_USED_SGPRS -
219                                PALMD::Key::VS_SCRATCH_SIZE;
220     setRegister(NumUsedSgprsKey, Val);
221     return;
222   }
223   // Msgpack format.
224   getHwStage(CC)[".sgpr_count"] = MsgPackDoc.getNode(Val);
225 }
226 
227 // Set the scratch size in the metadata.
228 void AMDGPUPALMetadata::setScratchSize(CallingConv::ID CC, unsigned Val) {
229   if (isLegacy()) {
230     // Old non-msgpack format.
231     setRegister(getScratchSizeKey(CC), Val);
232     return;
233   }
234   // Msgpack format.
235   getHwStage(CC)[".scratch_memory_size"] = MsgPackDoc.getNode(Val);
236 }
237 
238 // Set the hardware register bit in PAL metadata to enable wave32 on the
239 // shader of the given calling convention.
240 void AMDGPUPALMetadata::setWave32(unsigned CC) {
241   switch (CC) {
242   case CallingConv::AMDGPU_HS:
243     setRegister(PALMD::R_A2D5_VGT_SHADER_STAGES_EN, S_028B54_HS_W32_EN(1));
244     break;
245   case CallingConv::AMDGPU_GS:
246     setRegister(PALMD::R_A2D5_VGT_SHADER_STAGES_EN, S_028B54_GS_W32_EN(1));
247     break;
248   case CallingConv::AMDGPU_VS:
249     setRegister(PALMD::R_A2D5_VGT_SHADER_STAGES_EN, S_028B54_VS_W32_EN(1));
250     break;
251   case CallingConv::AMDGPU_PS:
252     setRegister(PALMD::R_A1B6_SPI_PS_IN_CONTROL, S_0286D8_PS_W32_EN(1));
253     break;
254   case CallingConv::AMDGPU_CS:
255     setRegister(PALMD::R_2E00_COMPUTE_DISPATCH_INITIATOR,
256                 S_00B800_CS_W32_EN(1));
257     break;
258   }
259 }
260 
261 // Convert a register number to name, for display by toString().
262 // Returns nullptr if none.
263 static const char *getRegisterName(unsigned RegNum) {
264   // Table of registers.
265   static const struct RegInfo {
266     unsigned Num;
267     const char *Name;
268   } RegInfoTable[] = {
269       // Registers that code generation sets/modifies metadata for.
270       {PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS, "SPI_SHADER_PGM_RSRC1_VS"},
271       {PALMD::R_2C4A_SPI_SHADER_PGM_RSRC1_VS + 1, "SPI_SHADER_PGM_RSRC2_VS"},
272       {PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS, "SPI_SHADER_PGM_RSRC1_LS"},
273       {PALMD::R_2D4A_SPI_SHADER_PGM_RSRC1_LS + 1, "SPI_SHADER_PGM_RSRC2_LS"},
274       {PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS, "SPI_SHADER_PGM_RSRC1_HS"},
275       {PALMD::R_2D0A_SPI_SHADER_PGM_RSRC1_HS + 1, "SPI_SHADER_PGM_RSRC2_HS"},
276       {PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES, "SPI_SHADER_PGM_RSRC1_ES"},
277       {PALMD::R_2CCA_SPI_SHADER_PGM_RSRC1_ES + 1, "SPI_SHADER_PGM_RSRC2_ES"},
278       {PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS, "SPI_SHADER_PGM_RSRC1_GS"},
279       {PALMD::R_2C8A_SPI_SHADER_PGM_RSRC1_GS + 1, "SPI_SHADER_PGM_RSRC2_GS"},
280       {PALMD::R_2E00_COMPUTE_DISPATCH_INITIATOR, "COMPUTE_DISPATCH_INITIATOR"},
281       {PALMD::R_2E12_COMPUTE_PGM_RSRC1, "COMPUTE_PGM_RSRC1"},
282       {PALMD::R_2E12_COMPUTE_PGM_RSRC1 + 1, "COMPUTE_PGM_RSRC2"},
283       {PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS, "SPI_SHADER_PGM_RSRC1_PS"},
284       {PALMD::R_2C0A_SPI_SHADER_PGM_RSRC1_PS + 1, "SPI_SHADER_PGM_RSRC2_PS"},
285       {PALMD::R_A1B3_SPI_PS_INPUT_ENA, "SPI_PS_INPUT_ENA"},
286       {PALMD::R_A1B4_SPI_PS_INPUT_ADDR, "SPI_PS_INPUT_ADDR"},
287       {PALMD::R_A1B6_SPI_PS_IN_CONTROL, "SPI_PS_IN_CONTROL"},
288       {PALMD::R_A2D5_VGT_SHADER_STAGES_EN, "VGT_SHADER_STAGES_EN"},
289 
290       // Registers not known to code generation.
291       {0x2c07, "SPI_SHADER_PGM_RSRC3_PS"},
292       {0x2c46, "SPI_SHADER_PGM_RSRC3_VS"},
293       {0x2c87, "SPI_SHADER_PGM_RSRC3_GS"},
294       {0x2cc7, "SPI_SHADER_PGM_RSRC3_ES"},
295       {0x2d07, "SPI_SHADER_PGM_RSRC3_HS"},
296       {0x2d47, "SPI_SHADER_PGM_RSRC3_LS"},
297 
298       {0xa1c3, "SPI_SHADER_POS_FORMAT"},
299       {0xa1b1, "SPI_VS_OUT_CONFIG"},
300       {0xa207, "PA_CL_VS_OUT_CNTL"},
301       {0xa204, "PA_CL_CLIP_CNTL"},
302       {0xa206, "PA_CL_VTE_CNTL"},
303       {0xa2f9, "PA_SU_VTX_CNTL"},
304       {0xa293, "PA_SC_MODE_CNTL_1"},
305       {0xa2a1, "VGT_PRIMITIVEID_EN"},
306       {0x2c81, "SPI_SHADER_PGM_RSRC4_GS"},
307       {0x2e18, "COMPUTE_TMPRING_SIZE"},
308       {0xa1b5, "SPI_INTERP_CONTROL_0"},
309       {0xa1ba, "SPI_TMPRING_SIZE"},
310       {0xa1c4, "SPI_SHADER_Z_FORMAT"},
311       {0xa1c5, "SPI_SHADER_COL_FORMAT"},
312       {0xa203, "DB_SHADER_CONTROL"},
313       {0xa08f, "CB_SHADER_MASK"},
314       {0xa191, "SPI_PS_INPUT_CNTL_0"},
315       {0xa192, "SPI_PS_INPUT_CNTL_1"},
316       {0xa193, "SPI_PS_INPUT_CNTL_2"},
317       {0xa194, "SPI_PS_INPUT_CNTL_3"},
318       {0xa195, "SPI_PS_INPUT_CNTL_4"},
319       {0xa196, "SPI_PS_INPUT_CNTL_5"},
320       {0xa197, "SPI_PS_INPUT_CNTL_6"},
321       {0xa198, "SPI_PS_INPUT_CNTL_7"},
322       {0xa199, "SPI_PS_INPUT_CNTL_8"},
323       {0xa19a, "SPI_PS_INPUT_CNTL_9"},
324       {0xa19b, "SPI_PS_INPUT_CNTL_10"},
325       {0xa19c, "SPI_PS_INPUT_CNTL_11"},
326       {0xa19d, "SPI_PS_INPUT_CNTL_12"},
327       {0xa19e, "SPI_PS_INPUT_CNTL_13"},
328       {0xa19f, "SPI_PS_INPUT_CNTL_14"},
329       {0xa1a0, "SPI_PS_INPUT_CNTL_15"},
330       {0xa1a1, "SPI_PS_INPUT_CNTL_16"},
331       {0xa1a2, "SPI_PS_INPUT_CNTL_17"},
332       {0xa1a3, "SPI_PS_INPUT_CNTL_18"},
333       {0xa1a4, "SPI_PS_INPUT_CNTL_19"},
334       {0xa1a5, "SPI_PS_INPUT_CNTL_20"},
335       {0xa1a6, "SPI_PS_INPUT_CNTL_21"},
336       {0xa1a7, "SPI_PS_INPUT_CNTL_22"},
337       {0xa1a8, "SPI_PS_INPUT_CNTL_23"},
338       {0xa1a9, "SPI_PS_INPUT_CNTL_24"},
339       {0xa1aa, "SPI_PS_INPUT_CNTL_25"},
340       {0xa1ab, "SPI_PS_INPUT_CNTL_26"},
341       {0xa1ac, "SPI_PS_INPUT_CNTL_27"},
342       {0xa1ad, "SPI_PS_INPUT_CNTL_28"},
343       {0xa1ae, "SPI_PS_INPUT_CNTL_29"},
344       {0xa1af, "SPI_PS_INPUT_CNTL_30"},
345       {0xa1b0, "SPI_PS_INPUT_CNTL_31"},
346 
347       {0xa2ce, "VGT_GS_MAX_VERT_OUT"},
348       {0xa2ab, "VGT_ESGS_RING_ITEMSIZE"},
349       {0xa290, "VGT_GS_MODE"},
350       {0xa291, "VGT_GS_ONCHIP_CNTL"},
351       {0xa2d7, "VGT_GS_VERT_ITEMSIZE"},
352       {0xa2d8, "VGT_GS_VERT_ITEMSIZE_1"},
353       {0xa2d9, "VGT_GS_VERT_ITEMSIZE_2"},
354       {0xa2da, "VGT_GS_VERT_ITEMSIZE_3"},
355       {0xa298, "VGT_GSVS_RING_OFFSET_1"},
356       {0xa299, "VGT_GSVS_RING_OFFSET_2"},
357       {0xa29a, "VGT_GSVS_RING_OFFSET_3"},
358 
359       {0xa2e4, "VGT_GS_INSTANCE_CNT"},
360       {0xa297, "VGT_GS_PER_VS"},
361       {0xa29b, "VGT_GS_OUT_PRIM_TYPE"},
362       {0xa2ac, "VGT_GSVS_RING_ITEMSIZE"},
363 
364       {0xa2ad, "VGT_REUSE_OFF"},
365       {0xa1b8, "SPI_BARYC_CNTL"},
366 
367       {0x2c4c, "SPI_SHADER_USER_DATA_VS_0"},
368       {0x2c4d, "SPI_SHADER_USER_DATA_VS_1"},
369       {0x2c4e, "SPI_SHADER_USER_DATA_VS_2"},
370       {0x2c4f, "SPI_SHADER_USER_DATA_VS_3"},
371       {0x2c50, "SPI_SHADER_USER_DATA_VS_4"},
372       {0x2c51, "SPI_SHADER_USER_DATA_VS_5"},
373       {0x2c52, "SPI_SHADER_USER_DATA_VS_6"},
374       {0x2c53, "SPI_SHADER_USER_DATA_VS_7"},
375       {0x2c54, "SPI_SHADER_USER_DATA_VS_8"},
376       {0x2c55, "SPI_SHADER_USER_DATA_VS_9"},
377       {0x2c56, "SPI_SHADER_USER_DATA_VS_10"},
378       {0x2c57, "SPI_SHADER_USER_DATA_VS_11"},
379       {0x2c58, "SPI_SHADER_USER_DATA_VS_12"},
380       {0x2c59, "SPI_SHADER_USER_DATA_VS_13"},
381       {0x2c5a, "SPI_SHADER_USER_DATA_VS_14"},
382       {0x2c5b, "SPI_SHADER_USER_DATA_VS_15"},
383       {0x2c5c, "SPI_SHADER_USER_DATA_VS_16"},
384       {0x2c5d, "SPI_SHADER_USER_DATA_VS_17"},
385       {0x2c5e, "SPI_SHADER_USER_DATA_VS_18"},
386       {0x2c5f, "SPI_SHADER_USER_DATA_VS_19"},
387       {0x2c60, "SPI_SHADER_USER_DATA_VS_20"},
388       {0x2c61, "SPI_SHADER_USER_DATA_VS_21"},
389       {0x2c62, "SPI_SHADER_USER_DATA_VS_22"},
390       {0x2c63, "SPI_SHADER_USER_DATA_VS_23"},
391       {0x2c64, "SPI_SHADER_USER_DATA_VS_24"},
392       {0x2c65, "SPI_SHADER_USER_DATA_VS_25"},
393       {0x2c66, "SPI_SHADER_USER_DATA_VS_26"},
394       {0x2c67, "SPI_SHADER_USER_DATA_VS_27"},
395       {0x2c68, "SPI_SHADER_USER_DATA_VS_28"},
396       {0x2c69, "SPI_SHADER_USER_DATA_VS_29"},
397       {0x2c6a, "SPI_SHADER_USER_DATA_VS_30"},
398       {0x2c6b, "SPI_SHADER_USER_DATA_VS_31"},
399 
400       {0x2ccc, "SPI_SHADER_USER_DATA_ES_0"},
401       {0x2ccd, "SPI_SHADER_USER_DATA_ES_1"},
402       {0x2cce, "SPI_SHADER_USER_DATA_ES_2"},
403       {0x2ccf, "SPI_SHADER_USER_DATA_ES_3"},
404       {0x2cd0, "SPI_SHADER_USER_DATA_ES_4"},
405       {0x2cd1, "SPI_SHADER_USER_DATA_ES_5"},
406       {0x2cd2, "SPI_SHADER_USER_DATA_ES_6"},
407       {0x2cd3, "SPI_SHADER_USER_DATA_ES_7"},
408       {0x2cd4, "SPI_SHADER_USER_DATA_ES_8"},
409       {0x2cd5, "SPI_SHADER_USER_DATA_ES_9"},
410       {0x2cd6, "SPI_SHADER_USER_DATA_ES_10"},
411       {0x2cd7, "SPI_SHADER_USER_DATA_ES_11"},
412       {0x2cd8, "SPI_SHADER_USER_DATA_ES_12"},
413       {0x2cd9, "SPI_SHADER_USER_DATA_ES_13"},
414       {0x2cda, "SPI_SHADER_USER_DATA_ES_14"},
415       {0x2cdb, "SPI_SHADER_USER_DATA_ES_15"},
416       {0x2cdc, "SPI_SHADER_USER_DATA_ES_16"},
417       {0x2cdd, "SPI_SHADER_USER_DATA_ES_17"},
418       {0x2cde, "SPI_SHADER_USER_DATA_ES_18"},
419       {0x2cdf, "SPI_SHADER_USER_DATA_ES_19"},
420       {0x2ce0, "SPI_SHADER_USER_DATA_ES_20"},
421       {0x2ce1, "SPI_SHADER_USER_DATA_ES_21"},
422       {0x2ce2, "SPI_SHADER_USER_DATA_ES_22"},
423       {0x2ce3, "SPI_SHADER_USER_DATA_ES_23"},
424       {0x2ce4, "SPI_SHADER_USER_DATA_ES_24"},
425       {0x2ce5, "SPI_SHADER_USER_DATA_ES_25"},
426       {0x2ce6, "SPI_SHADER_USER_DATA_ES_26"},
427       {0x2ce7, "SPI_SHADER_USER_DATA_ES_27"},
428       {0x2ce8, "SPI_SHADER_USER_DATA_ES_28"},
429       {0x2ce9, "SPI_SHADER_USER_DATA_ES_29"},
430       {0x2cea, "SPI_SHADER_USER_DATA_ES_30"},
431       {0x2ceb, "SPI_SHADER_USER_DATA_ES_31"},
432 
433       {0x2c0c, "SPI_SHADER_USER_DATA_PS_0"},
434       {0x2c0d, "SPI_SHADER_USER_DATA_PS_1"},
435       {0x2c0e, "SPI_SHADER_USER_DATA_PS_2"},
436       {0x2c0f, "SPI_SHADER_USER_DATA_PS_3"},
437       {0x2c10, "SPI_SHADER_USER_DATA_PS_4"},
438       {0x2c11, "SPI_SHADER_USER_DATA_PS_5"},
439       {0x2c12, "SPI_SHADER_USER_DATA_PS_6"},
440       {0x2c13, "SPI_SHADER_USER_DATA_PS_7"},
441       {0x2c14, "SPI_SHADER_USER_DATA_PS_8"},
442       {0x2c15, "SPI_SHADER_USER_DATA_PS_9"},
443       {0x2c16, "SPI_SHADER_USER_DATA_PS_10"},
444       {0x2c17, "SPI_SHADER_USER_DATA_PS_11"},
445       {0x2c18, "SPI_SHADER_USER_DATA_PS_12"},
446       {0x2c19, "SPI_SHADER_USER_DATA_PS_13"},
447       {0x2c1a, "SPI_SHADER_USER_DATA_PS_14"},
448       {0x2c1b, "SPI_SHADER_USER_DATA_PS_15"},
449       {0x2c1c, "SPI_SHADER_USER_DATA_PS_16"},
450       {0x2c1d, "SPI_SHADER_USER_DATA_PS_17"},
451       {0x2c1e, "SPI_SHADER_USER_DATA_PS_18"},
452       {0x2c1f, "SPI_SHADER_USER_DATA_PS_19"},
453       {0x2c20, "SPI_SHADER_USER_DATA_PS_20"},
454       {0x2c21, "SPI_SHADER_USER_DATA_PS_21"},
455       {0x2c22, "SPI_SHADER_USER_DATA_PS_22"},
456       {0x2c23, "SPI_SHADER_USER_DATA_PS_23"},
457       {0x2c24, "SPI_SHADER_USER_DATA_PS_24"},
458       {0x2c25, "SPI_SHADER_USER_DATA_PS_25"},
459       {0x2c26, "SPI_SHADER_USER_DATA_PS_26"},
460       {0x2c27, "SPI_SHADER_USER_DATA_PS_27"},
461       {0x2c28, "SPI_SHADER_USER_DATA_PS_28"},
462       {0x2c29, "SPI_SHADER_USER_DATA_PS_29"},
463       {0x2c2a, "SPI_SHADER_USER_DATA_PS_30"},
464       {0x2c2b, "SPI_SHADER_USER_DATA_PS_31"},
465 
466       {0x2e40, "COMPUTE_USER_DATA_0"},
467       {0x2e41, "COMPUTE_USER_DATA_1"},
468       {0x2e42, "COMPUTE_USER_DATA_2"},
469       {0x2e43, "COMPUTE_USER_DATA_3"},
470       {0x2e44, "COMPUTE_USER_DATA_4"},
471       {0x2e45, "COMPUTE_USER_DATA_5"},
472       {0x2e46, "COMPUTE_USER_DATA_6"},
473       {0x2e47, "COMPUTE_USER_DATA_7"},
474       {0x2e48, "COMPUTE_USER_DATA_8"},
475       {0x2e49, "COMPUTE_USER_DATA_9"},
476       {0x2e4a, "COMPUTE_USER_DATA_10"},
477       {0x2e4b, "COMPUTE_USER_DATA_11"},
478       {0x2e4c, "COMPUTE_USER_DATA_12"},
479       {0x2e4d, "COMPUTE_USER_DATA_13"},
480       {0x2e4e, "COMPUTE_USER_DATA_14"},
481       {0x2e4f, "COMPUTE_USER_DATA_15"},
482 
483       {0x2e07, "COMPUTE_NUM_THREAD_X"},
484       {0x2e08, "COMPUTE_NUM_THREAD_Y"},
485       {0x2e09, "COMPUTE_NUM_THREAD_Z"},
486       {0xa2db, "VGT_TF_PARAM"},
487       {0xa2d6, "VGT_LS_HS_CONFIG"},
488       {0xa287, "VGT_HOS_MIN_TESS_LEVEL"},
489       {0xa286, "VGT_HOS_MAX_TESS_LEVEL"},
490       {0xa2f8, "PA_SC_AA_CONFIG"},
491       {0xa310, "PA_SC_SHADER_CONTROL"},
492       {0xa313, "PA_SC_CONSERVATIVE_RASTERIZATION_CNTL"},
493 
494       {0x2d0c, "SPI_SHADER_USER_DATA_LS_0"},
495       {0x2d0d, "SPI_SHADER_USER_DATA_LS_1"},
496       {0x2d0e, "SPI_SHADER_USER_DATA_LS_2"},
497       {0x2d0f, "SPI_SHADER_USER_DATA_LS_3"},
498       {0x2d10, "SPI_SHADER_USER_DATA_LS_4"},
499       {0x2d11, "SPI_SHADER_USER_DATA_LS_5"},
500       {0x2d12, "SPI_SHADER_USER_DATA_LS_6"},
501       {0x2d13, "SPI_SHADER_USER_DATA_LS_7"},
502       {0x2d14, "SPI_SHADER_USER_DATA_LS_8"},
503       {0x2d15, "SPI_SHADER_USER_DATA_LS_9"},
504       {0x2d16, "SPI_SHADER_USER_DATA_LS_10"},
505       {0x2d17, "SPI_SHADER_USER_DATA_LS_11"},
506       {0x2d18, "SPI_SHADER_USER_DATA_LS_12"},
507       {0x2d19, "SPI_SHADER_USER_DATA_LS_13"},
508       {0x2d1a, "SPI_SHADER_USER_DATA_LS_14"},
509       {0x2d1b, "SPI_SHADER_USER_DATA_LS_15"},
510       {0x2d1c, "SPI_SHADER_USER_DATA_LS_16"},
511       {0x2d1d, "SPI_SHADER_USER_DATA_LS_17"},
512       {0x2d1e, "SPI_SHADER_USER_DATA_LS_18"},
513       {0x2d1f, "SPI_SHADER_USER_DATA_LS_19"},
514       {0x2d20, "SPI_SHADER_USER_DATA_LS_20"},
515       {0x2d21, "SPI_SHADER_USER_DATA_LS_21"},
516       {0x2d22, "SPI_SHADER_USER_DATA_LS_22"},
517       {0x2d23, "SPI_SHADER_USER_DATA_LS_23"},
518       {0x2d24, "SPI_SHADER_USER_DATA_LS_24"},
519       {0x2d25, "SPI_SHADER_USER_DATA_LS_25"},
520       {0x2d26, "SPI_SHADER_USER_DATA_LS_26"},
521       {0x2d27, "SPI_SHADER_USER_DATA_LS_27"},
522       {0x2d28, "SPI_SHADER_USER_DATA_LS_28"},
523       {0x2d29, "SPI_SHADER_USER_DATA_LS_29"},
524       {0x2d2a, "SPI_SHADER_USER_DATA_LS_30"},
525       {0x2d2b, "SPI_SHADER_USER_DATA_LS_31"},
526 
527       {0xa2aa, "IA_MULTI_VGT_PARAM"},
528       {0xa2a5, "VGT_GS_MAX_PRIMS_PER_SUBGROUP"},
529       {0xa2e6, "VGT_STRMOUT_BUFFER_CONFIG"},
530       {0xa2e5, "VGT_STRMOUT_CONFIG"},
531       {0xa2b5, "VGT_STRMOUT_VTX_STRIDE_0"},
532       {0xa2b9, "VGT_STRMOUT_VTX_STRIDE_1"},
533       {0xa2bd, "VGT_STRMOUT_VTX_STRIDE_2"},
534       {0xa2c1, "VGT_STRMOUT_VTX_STRIDE_3"},
535       {0xa316, "VGT_VERTEX_REUSE_BLOCK_CNTL"},
536 
537       {0, nullptr}};
538   auto Entry = RegInfoTable;
539   for (; Entry->Num && Entry->Num != RegNum; ++Entry)
540     ;
541   return Entry->Name;
542 }
543 
544 // Convert the accumulated PAL metadata into an asm directive.
545 void AMDGPUPALMetadata::toString(std::string &String) {
546   String.clear();
547   if (!BlobType)
548     return;
549   raw_string_ostream Stream(String);
550   if (isLegacy()) {
551     if (MsgPackDoc.getRoot().getKind() == msgpack::Type::Nil)
552       return;
553     // Old linear reg=val format.
554     Stream << '\t' << AMDGPU::PALMD::AssemblerDirective << ' ';
555     auto Regs = getRegisters();
556     for (auto I = Regs.begin(), E = Regs.end(); I != E; ++I) {
557       if (I != Regs.begin())
558         Stream << ',';
559       unsigned Reg = I->first.getUInt();
560       unsigned Val = I->second.getUInt();
561       Stream << "0x" << Twine::utohexstr(Reg) << ",0x" << Twine::utohexstr(Val);
562     }
563     Stream << '\n';
564     return;
565   }
566 
567   // New msgpack-based format -- output as YAML (with unsigned numbers in hex),
568   // but first change the registers map to use names.
569   MsgPackDoc.setHexMode();
570   auto &RegsObj = refRegisters();
571   auto OrigRegs = RegsObj.getMap();
572   RegsObj = MsgPackDoc.getMapNode();
573   for (auto I : OrigRegs) {
574     auto Key = I.first;
575     if (const char *RegName = getRegisterName(Key.getUInt())) {
576       std::string KeyName = Key.toString();
577       KeyName += " (";
578       KeyName += RegName;
579       KeyName += ')';
580       Key = MsgPackDoc.getNode(KeyName, /*Copy=*/true);
581     }
582     RegsObj.getMap()[Key] = I.second;
583   }
584 
585   // Output as YAML.
586   Stream << '\t' << AMDGPU::PALMD::AssemblerDirectiveBegin << '\n';
587   MsgPackDoc.toYAML(Stream);
588   Stream << '\t' << AMDGPU::PALMD::AssemblerDirectiveEnd << '\n';
589 
590   // Restore original registers map.
591   RegsObj = OrigRegs;
592 }
593 
594 // Convert the accumulated PAL metadata into a binary blob for writing as
595 // a .note record of the specified AMD type. Returns an empty blob if
596 // there is no PAL metadata,
597 void AMDGPUPALMetadata::toBlob(unsigned Type, std::string &Blob) {
598   if (Type == ELF::NT_AMD_AMDGPU_PAL_METADATA)
599     toLegacyBlob(Blob);
600   else if (Type)
601     toMsgPackBlob(Blob);
602 }
603 
604 void AMDGPUPALMetadata::toLegacyBlob(std::string &Blob) {
605   Blob.clear();
606   auto Registers = getRegisters();
607   if (Registers.getMap().empty())
608     return;
609   raw_string_ostream OS(Blob);
610   support::endian::Writer EW(OS, support::endianness::little);
611   for (auto I : Registers.getMap()) {
612     EW.write(uint32_t(I.first.getUInt()));
613     EW.write(uint32_t(I.second.getUInt()));
614   }
615 }
616 
617 void AMDGPUPALMetadata::toMsgPackBlob(std::string &Blob) {
618   Blob.clear();
619   MsgPackDoc.writeToBlob(Blob);
620 }
621 
622 // Set PAL metadata from YAML text. Returns false if failed.
623 bool AMDGPUPALMetadata::setFromString(StringRef S) {
624   BlobType = ELF::NT_AMDGPU_METADATA;
625   if (!MsgPackDoc.fromYAML(S))
626     return false;
627 
628   // In the registers map, some keys may be of the form "0xa191
629   // (SPI_PS_INPUT_CNTL_0)", in which case the YAML input code made it a
630   // string. We need to turn it into a number.
631   auto &RegsObj = refRegisters();
632   auto OrigRegs = RegsObj;
633   RegsObj = MsgPackDoc.getMapNode();
634   Registers = RegsObj.getMap();
635   bool Ok = true;
636   for (auto I : OrigRegs.getMap()) {
637     auto Key = I.first;
638     if (Key.getKind() == msgpack::Type::String) {
639       StringRef S = Key.getString();
640       uint64_t Val;
641       if (S.consumeInteger(0, Val)) {
642         Ok = false;
643         errs() << "Unrecognized PAL metadata register key '" << S << "'\n";
644         continue;
645       }
646       Key = MsgPackDoc.getNode(uint64_t(Val));
647     }
648     Registers.getMap()[Key] = I.second;
649   }
650   return Ok;
651 }
652 
653 // Reference (create if necessary) the node for the registers map.
654 msgpack::DocNode &AMDGPUPALMetadata::refRegisters() {
655   auto &N =
656       MsgPackDoc.getRoot()
657           .getMap(/*Convert=*/true)[MsgPackDoc.getNode("amdpal.pipelines")]
658           .getArray(/*Convert=*/true)[0]
659           .getMap(/*Convert=*/true)[MsgPackDoc.getNode(".registers")];
660   N.getMap(/*Convert=*/true);
661   return N;
662 }
663 
664 // Get (create if necessary) the registers map.
665 msgpack::MapDocNode AMDGPUPALMetadata::getRegisters() {
666   if (Registers.isEmpty())
667     Registers = refRegisters();
668   return Registers.getMap();
669 }
670 
671 // Return the PAL metadata hardware shader stage name.
672 static const char *getStageName(CallingConv::ID CC) {
673   switch (CC) {
674   case CallingConv::AMDGPU_PS:
675     return ".ps";
676   case CallingConv::AMDGPU_VS:
677     return ".vs";
678   case CallingConv::AMDGPU_GS:
679     return ".gs";
680   case CallingConv::AMDGPU_ES:
681     return ".es";
682   case CallingConv::AMDGPU_HS:
683     return ".hs";
684   case CallingConv::AMDGPU_LS:
685     return ".ls";
686   default:
687     return ".cs";
688   }
689 }
690 
691 // Get (create if necessary) the .hardware_stages entry for the given calling
692 // convention.
693 msgpack::MapDocNode AMDGPUPALMetadata::getHwStage(unsigned CC) {
694   if (HwStages.isEmpty())
695     HwStages = MsgPackDoc.getRoot()
696                    .getMap(/*Convert=*/true)["amdpal.pipelines"]
697                    .getArray(/*Convert=*/true)[0]
698                    .getMap(/*Convert=*/true)[".hardware_stages"]
699                    .getMap(/*Convert=*/true);
700   return HwStages.getMap()[getStageName(CC)].getMap(/*Convert=*/true);
701 }
702 
703 // Get .note record vendor name of metadata blob to be emitted.
704 const char *AMDGPUPALMetadata::getVendor() const {
705   return isLegacy() ? ElfNote::NoteNameV2 : ElfNote::NoteNameV3;
706 }
707 
708 // Get .note record type of metadata blob to be emitted:
709 // ELF::NT_AMD_AMDGPU_PAL_METADATA (legacy key=val format), or
710 // ELF::NT_AMDGPU_METADATA (MsgPack format), or
711 // 0 (no PAL metadata).
712 unsigned AMDGPUPALMetadata::getType() const {
713   return BlobType;
714 }
715 
716 // Return whether the blob type is legacy PAL metadata.
717 bool AMDGPUPALMetadata::isLegacy() const {
718   return BlobType == ELF::NT_AMD_AMDGPU_PAL_METADATA;
719 }
720 
721 // Set legacy PAL metadata format.
722 void AMDGPUPALMetadata::setLegacy() {
723   BlobType = ELF::NT_AMD_AMDGPU_PAL_METADATA;
724 }
725 
726