aco: shorten disassembly for repeated instructions
[mesa.git] / src / amd / compiler / aco_print_asm.cpp
1 #include <array>
2 #include <iomanip>
3 #include "aco_ir.h"
4 #include "llvm-c/Disassembler.h"
5 #include "ac_llvm_util.h"
6
7 #include <llvm/ADT/StringRef.h>
8 #if LLVM_VERSION_MAJOR >= 11
9 #include <llvm/MC/MCDisassembler/MCDisassembler.h>
10 #endif
11
12 namespace aco {
13
14 /* LLVM disassembler only supports GFX8+, try to disassemble with CLRXdisasm
15 * for GFX6-GFX7 if found on the system, this is better than nothing.
16 */
17 void print_asm_gfx6_gfx7(Program *program, std::vector<uint32_t>& binary,
18 std::ostream& out)
19 {
20 char path[] = "/tmp/fileXXXXXX";
21 char line[2048], command[128];
22 const char *gpu_type;
23 FILE *p;
24 int fd;
25
26 /* Dump the binary into a temporary file. */
27 fd = mkstemp(path);
28 if (fd < 0)
29 return;
30
31 for (uint32_t w : binary)
32 {
33 if (write(fd, &w, sizeof(w)) == -1)
34 goto fail;
35 }
36
37 /* Determine the GPU type for CLRXdisasm. Use the family for GFX6 chips
38 * because it doesn't allow to use gfx600 directly.
39 */
40 switch (program->chip_class) {
41 case GFX6:
42 switch (program->family) {
43 case CHIP_TAHITI:
44 gpu_type = "tahiti";
45 break;
46 case CHIP_PITCAIRN:
47 gpu_type = "pitcairn";
48 break;
49 case CHIP_VERDE:
50 gpu_type = "capeverde";
51 break;
52 case CHIP_OLAND:
53 gpu_type = "oland";
54 break;
55 case CHIP_HAINAN:
56 gpu_type = "hainan";
57 break;
58 default:
59 unreachable("Invalid GFX6 family!");
60 }
61 break;
62 case GFX7:
63 gpu_type = "gfx700";
64 break;
65 default:
66 unreachable("Invalid chip class!");
67 }
68
69 sprintf(command, "clrxdisasm --gpuType=%s -r %s", gpu_type, path);
70
71 p = popen(command, "r");
72 if (p) {
73 if (!fgets(line, sizeof(line), p)) {
74 out << "clrxdisasm not found\n";
75 pclose(p);
76 goto fail;
77 }
78
79 do {
80 out << line;
81 } while (fgets(line, sizeof(line), p));
82
83 pclose(p);
84 }
85
86 fail:
87 close(fd);
88 unlink(path);
89 }
90
91 void print_asm(Program *program, std::vector<uint32_t>& binary,
92 unsigned exec_size, std::ostream& out)
93 {
94 if (program->chip_class <= GFX7) {
95 print_asm_gfx6_gfx7(program, binary, out);
96 return;
97 }
98
99 std::vector<bool> referenced_blocks(program->blocks.size());
100 referenced_blocks[0] = true;
101 for (Block& block : program->blocks) {
102 for (unsigned succ : block.linear_succs)
103 referenced_blocks[succ] = true;
104 }
105
106 #if LLVM_VERSION_MAJOR >= 11
107 std::vector<llvm::SymbolInfoTy> symbols;
108 #else
109 std::vector<std::tuple<uint64_t, llvm::StringRef, uint8_t>> symbols;
110 #endif
111 std::vector<std::array<char,16>> block_names;
112 block_names.reserve(program->blocks.size());
113 for (Block& block : program->blocks) {
114 if (!referenced_blocks[block.index])
115 continue;
116 std::array<char, 16> name;
117 sprintf(name.data(), "BB%u", block.index);
118 block_names.push_back(name);
119 symbols.emplace_back(block.offset * 4, llvm::StringRef(block_names[block_names.size() - 1].data()), 0);
120 }
121
122 const char *features = "";
123 if (program->chip_class >= GFX10 && program->wave_size == 64) {
124 features = "+wavefrontsize64";
125 }
126
127 LLVMDisasmContextRef disasm = LLVMCreateDisasmCPUFeatures("amdgcn-mesa-mesa3d",
128 ac_get_llvm_processor_name(program->family),
129 features,
130 &symbols, 0, NULL, NULL);
131
132 char outline[1024];
133 size_t pos = 0;
134 bool invalid = false;
135 unsigned next_block = 0;
136 while (pos < exec_size) {
137 while (next_block < program->blocks.size() && pos == program->blocks[next_block].offset) {
138 if (referenced_blocks[next_block])
139 out << "BB" << std::dec << next_block << ":" << std::endl;
140 next_block++;
141 }
142
143 /* mask out src2 on v_writelane_b32 */
144 if (((program->chip_class == GFX8 || program->chip_class == GFX9) && (binary[pos] & 0xffff8000) == 0xd28a0000) ||
145 (program->chip_class >= GFX10 && (binary[pos] & 0xffff8000) == 0xd7610000)) {
146 binary[pos+1] = binary[pos+1] & 0xF803FFFF;
147 }
148
149 const int align_width = 60;
150 out << std::left << std::setw(align_width) << std::setfill(' ');
151
152 size_t l = LLVMDisasmInstruction(disasm, (uint8_t *) &binary[pos],
153 (exec_size - pos) * sizeof(uint32_t), pos * 4,
154 outline, sizeof(outline));
155
156 size_t new_pos;
157 if (!l &&
158 ((program->chip_class >= GFX9 && (binary[pos] & 0xffff8000) == 0xd1348000) || /* v_add_u32_e64 + clamp */
159 (program->chip_class >= GFX10 && (binary[pos] & 0xffff8000) == 0xd7038000) || /* v_add_u16_e64 + clamp */
160 (program->chip_class <= GFX9 && (binary[pos] & 0xffff8000) == 0xd1268000)) /* v_add_u16_e64 + clamp */) {
161 out << "\tinteger addition + clamp";
162 bool has_literal = program->chip_class >= GFX10 &&
163 (((binary[pos+1] & 0x1ff) == 0xff) || (((binary[pos+1] >> 9) & 0x1ff) == 0xff));
164 new_pos = pos + 2 + has_literal;
165 } else if (program->chip_class >= GFX10 && l == 4 && ((binary[pos] & 0xfe0001ff) == 0x020000f9)) {
166 out << "\tv_cndmask_b32 + sdwa";
167 new_pos = pos + 2;
168 } else if (!l) {
169 out << "(invalid instruction)";
170 new_pos = pos + 1;
171 invalid = true;
172 } else {
173 out << outline;
174 assert(l % 4 == 0);
175 new_pos = pos + l / 4;
176 }
177 size_t size = new_pos - pos;
178 out << std::right;
179
180 out << " ;";
181 for (unsigned i = 0; i < size; i++)
182 out << " " << std::setfill('0') << std::setw(8) << std::hex << binary[pos + i];
183 out << std::endl;
184
185 size_t original_pos = pos;
186 pos += size;
187
188 unsigned repeat_count = 0;
189 while (pos + size <= exec_size && memcmp(&binary[pos], &binary[original_pos], size * 4) == 0) {
190 repeat_count++;
191 pos += size;
192 }
193 if (repeat_count)
194 out << std::left << std::setw(0) << std::dec << std::setfill(' ') << "\t(then repeated " << repeat_count << " times)" << std::endl;
195 }
196 out << std::setfill(' ') << std::setw(0) << std::dec;
197 assert(next_block == program->blocks.size());
198
199 LLVMDisasmDispose(disasm);
200
201 if (program->constant_data.size()) {
202 out << std::endl << "/* constant data */" << std::endl;
203 for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
204 out << '[' << std::setw(6) << std::setfill('0') << std::dec << i << ']';
205 unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
206 for (unsigned j = 0; j < line_size; j += 4) {
207 unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
208 uint32_t v = 0;
209 memcpy(&v, &program->constant_data[i + j], size);
210 out << " " << std::setw(8) << std::setfill('0') << std::hex << v;
211 }
212 out << std::endl;
213 }
214 }
215
216 out << std::setfill(' ') << std::setw(0) << std::dec;
217
218 if (invalid) {
219 /* Invalid instructions usually lead to GPU hangs, which can make
220 * getting the actual invalid instruction hard. Abort here so that we
221 * can find the problem.
222 */
223 abort();
224 }
225 }
226
227 }