1 //===-- AMDGPUAsmPrinter.h - Print AMDGPU assembly code ---------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief AMDGPU Assembly printer class.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUASMPRINTER_H
16 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUASMPRINTER_H
17 
18 #include "AMDGPU.h"
19 #include "AMDKernelCodeT.h"
20 #include "MCTargetDesc/AMDGPUHSAMetadataStreamer.h"
21 #include "llvm/ADT/StringRef.h"
22 #include "llvm/CodeGen/AsmPrinter.h"
23 #include <cstddef>
24 #include <cstdint>
25 #include <limits>
26 #include <memory>
27 #include <string>
28 #include <vector>
29 
30 namespace llvm {
31 
32 class AMDGPUTargetStreamer;
33 class MCOperand;
34 class SISubtarget;
35 
36 class AMDGPUAsmPrinter final : public AsmPrinter {
37 private:
38   // Track resource usage for callee functions.
39   struct SIFunctionResourceInfo {
40     // Track the number of explicitly used VGPRs. Special registers reserved at
41     // the end are tracked separately.
42     int32_t NumVGPR = 0;
43     int32_t NumExplicitSGPR = 0;
44     uint32_t PrivateSegmentSize = 0;
45     bool UsesVCC = false;
46     bool UsesFlatScratch = false;
47     bool HasDynamicallySizedStack = false;
48     bool HasRecursion = false;
49 
50     int32_t getTotalNumSGPRs(const SISubtarget &ST) const;
51   };
52 
53   // Track resource usage for kernels / entry functions.
54   struct SIProgramInfo {
55     // Fields set in PGM_RSRC1 pm4 packet.
56     uint32_t VGPRBlocks = 0;
57     uint32_t SGPRBlocks = 0;
58     uint32_t Priority = 0;
59     uint32_t FloatMode = 0;
60     uint32_t Priv = 0;
61     uint32_t DX10Clamp = 0;
62     uint32_t DebugMode = 0;
63     uint32_t IEEEMode = 0;
64     uint32_t ScratchSize = 0;
65 
66     uint64_t ComputePGMRSrc1 = 0;
67 
68     // Fields set in PGM_RSRC2 pm4 packet.
69     uint32_t LDSBlocks = 0;
70     uint32_t ScratchBlocks = 0;
71 
72     uint64_t ComputePGMRSrc2 = 0;
73 
74     uint32_t NumVGPR = 0;
75     uint32_t NumSGPR = 0;
76     uint32_t LDSSize = 0;
77     bool FlatUsed = false;
78 
79     // Number of SGPRs that meets number of waves per execution unit request.
80     uint32_t NumSGPRsForWavesPerEU = 0;
81 
82     // Number of VGPRs that meets number of waves per execution unit request.
83     uint32_t NumVGPRsForWavesPerEU = 0;
84 
85     // If ReservedVGPRCount is 0 then must be 0. Otherwise, this is the first
86     // fixed VGPR number reserved.
87     uint16_t ReservedVGPRFirst = 0;
88 
89     // The number of consecutive VGPRs reserved.
90     uint16_t ReservedVGPRCount = 0;
91 
92     // Fixed SGPR number used to hold wave scratch offset for entire kernel
93     // execution, or std::numeric_limits<uint16_t>::max() if the register is not
94     // used or not known.
95     uint16_t DebuggerWavefrontPrivateSegmentOffsetSGPR =
96         std::numeric_limits<uint16_t>::max();
97 
98     // Fixed SGPR number of the first 4 SGPRs used to hold scratch V# for entire
99     // kernel execution, or std::numeric_limits<uint16_t>::max() if the register
100     // is not used or not known.
101     uint16_t DebuggerPrivateSegmentBufferSGPR =
102         std::numeric_limits<uint16_t>::max();
103 
104     // Whether there is recursion, dynamic allocas, indirect calls or some other
105     // reason there may be statically unknown stack usage.
106     bool DynamicCallStack = false;
107 
108     // Bonus information for debugging.
109     bool VCCUsed = false;
110 
111     SIProgramInfo() = default;
112   };
113 
114   SIProgramInfo CurrentProgramInfo;
115   DenseMap<const Function *, SIFunctionResourceInfo> CallGraphResourceInfo;
116 
117   AMDGPU::HSAMD::MetadataStreamer HSAMetadataStream;
118   std::map<uint32_t, uint32_t> PALMetadataMap;
119 
120   uint64_t getFunctionCodeSize(const MachineFunction &MF) const;
121   SIFunctionResourceInfo analyzeResourceUsage(const MachineFunction &MF) const;
122 
123   void readPALMetadata(Module &M);
124   void getSIProgramInfo(SIProgramInfo &Out, const MachineFunction &MF);
125   void getAmdKernelCode(amd_kernel_code_t &Out, const SIProgramInfo &KernelInfo,
126                         const MachineFunction &MF) const;
127   void findNumUsedRegistersSI(const MachineFunction &MF,
128                               unsigned &NumSGPR,
129                               unsigned &NumVGPR) const;
130 
131   /// \brief Emit register usage information so that the GPU driver
132   /// can correctly setup the GPU state.
133   void EmitProgramInfoR600(const MachineFunction &MF);
134   void EmitProgramInfoSI(const MachineFunction &MF,
135                          const SIProgramInfo &KernelInfo);
136   void EmitPALMetadata(const MachineFunction &MF,
137                        const SIProgramInfo &KernelInfo);
138   void emitCommonFunctionComments(uint32_t NumVGPR,
139                                   uint32_t NumSGPR,
140                                   uint32_t ScratchSize,
141                                   uint64_t CodeSize);
142 
143 public:
144   explicit AMDGPUAsmPrinter(TargetMachine &TM,
145                             std::unique_ptr<MCStreamer> Streamer);
146 
147   StringRef getPassName() const override;
148 
149   const MCSubtargetInfo* getSTI() const;
150 
151   AMDGPUTargetStreamer& getTargetStreamer() const;
152 
153   bool doFinalization(Module &M) override;
154   bool runOnMachineFunction(MachineFunction &MF) override;
155 
156   /// \brief Wrapper for MCInstLowering.lowerOperand() for the tblgen'erated
157   /// pseudo lowering.
158   bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp) const;
159 
160   /// \brief Lower the specified LLVM Constant to an MCExpr.
161   /// The AsmPrinter::lowerConstantof does not know how to lower
162   /// addrspacecast, therefore they should be lowered by this function.
163   const MCExpr *lowerConstant(const Constant *CV) override;
164 
165   /// \brief tblgen'erated driver function for lowering simple MI->MC pseudo
166   /// instructions.
167   bool emitPseudoExpansionLowering(MCStreamer &OutStreamer,
168                                    const MachineInstr *MI);
169 
170   /// Implemented in AMDGPUMCInstLower.cpp
171   void EmitInstruction(const MachineInstr *MI) override;
172 
173   void EmitFunctionBodyStart() override;
174 
175   void EmitFunctionEntryLabel() override;
176 
177   void EmitGlobalVariable(const GlobalVariable *GV) override;
178 
179   void EmitStartOfAsmFile(Module &M) override;
180 
181   void EmitEndOfAsmFile(Module &M) override;
182 
183   bool isBlockOnlyReachableByFallthrough(
184     const MachineBasicBlock *MBB) const override;
185 
186   bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
187                        unsigned AsmVariant, const char *ExtraCode,
188                        raw_ostream &O) override;
189 
190 protected:
191   std::vector<std::string> DisasmLines, HexLines;
192   size_t DisasmLineMaxLen;
193   AMDGPUAS AMDGPUASI;
194 };
195 
196 } // end namespace llvm
197 
198 #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUASMPRINTER_H
199