1 /*
2 * Copyright (C) 2019, Alex Bennée <alex.bennee@linaro.org>
3 *
4 * How vectorised is this code?
5 *
6 * Attempt to measure the amount of vectorisation that has been done
7 * on some code by counting classes of instruction.
8 *
9 * License: GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
11 */
12 #include <inttypes.h>
13 #include <assert.h>
14 #include <stdlib.h>
15 #include <inttypes.h>
16 #include <string.h>
17 #include <unistd.h>
18 #include <stdio.h>
19 #include <glib.h>
20
21 #include <qemu-plugin.h>
22
23 QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION;
24
25 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
26
27 typedef enum {
28 COUNT_CLASS,
29 COUNT_INDIVIDUAL,
30 COUNT_NONE
31 } CountType;
32
33 static int limit = 50;
34 static bool do_inline;
35 static bool verbose;
36
37 static GMutex lock;
38 static GHashTable *insns;
39
40 typedef struct {
41 const char *class;
42 const char *opt;
43 uint32_t mask;
44 uint32_t pattern;
45 CountType what;
46 qemu_plugin_u64 count;
47 } InsnClassExecCount;
48
49 typedef struct {
50 char *insn;
51 uint32_t opcode;
52 qemu_plugin_u64 count;
53 InsnClassExecCount *class;
54 } InsnExecCount;
55
56 /*
57 * Matchers for classes of instructions, order is important.
58 *
59 * Your most precise match must be before looser matches. If no match
60 * is found in the table we can create an individual entry.
61 *
62 * 31..28 27..24 23..20 19..16 15..12 11..8 7..4 3..0
63 */
64 static InsnClassExecCount aarch64_insn_classes[] = {
65 /* "Reserved"" */
66 { " UDEF", "udef", 0xffff0000, 0x00000000, COUNT_NONE},
67 { " SVE", "sve", 0x1e000000, 0x04000000, COUNT_CLASS},
68 { "Reserved", "res", 0x1e000000, 0x00000000, COUNT_CLASS},
69 /* Data Processing Immediate */
70 { " PCrel addr", "pcrel", 0x1f000000, 0x10000000, COUNT_CLASS},
71 { " Add/Sub (imm,tags)", "asit", 0x1f800000, 0x11800000, COUNT_CLASS},
72 { " Add/Sub (imm)", "asi", 0x1f000000, 0x11000000, COUNT_CLASS},
73 { " Logical (imm)", "logi", 0x1f800000, 0x12000000, COUNT_CLASS},
74 { " Move Wide (imm)", "movwi", 0x1f800000, 0x12800000, COUNT_CLASS},
75 { " Bitfield", "bitf", 0x1f800000, 0x13000000, COUNT_CLASS},
76 { " Extract", "extr", 0x1f800000, 0x13800000, COUNT_CLASS},
77 { "Data Proc Imm", "dpri", 0x1c000000, 0x10000000, COUNT_CLASS},
78 /* Branches */
79 { " Cond Branch (imm)", "cndb", 0xfe000000, 0x54000000, COUNT_CLASS},
80 { " Exception Gen", "excp", 0xff000000, 0xd4000000, COUNT_CLASS},
81 { " NOP", "nop", 0xffffffff, 0xd503201f, COUNT_NONE},
82 { " Hints", "hint", 0xfffff000, 0xd5032000, COUNT_CLASS},
83 { " Barriers", "barr", 0xfffff000, 0xd5033000, COUNT_CLASS},
84 { " PSTATE", "psta", 0xfff8f000, 0xd5004000, COUNT_CLASS},
85 { " System Insn", "sins", 0xffd80000, 0xd5080000, COUNT_CLASS},
86 { " System Reg", "sreg", 0xffd00000, 0xd5100000, COUNT_CLASS},
87 { " Branch (reg)", "breg", 0xfe000000, 0xd6000000, COUNT_CLASS},
88 { " Branch (imm)", "bimm", 0x7c000000, 0x14000000, COUNT_CLASS},
89 { " Cmp & Branch", "cmpb", 0x7e000000, 0x34000000, COUNT_CLASS},
90 { " Tst & Branch", "tstb", 0x7e000000, 0x36000000, COUNT_CLASS},
91 { "Branches", "branch", 0x1c000000, 0x14000000, COUNT_CLASS},
92 /* Loads and Stores */
93 { " AdvSimd ldstmult", "advlsm", 0xbfbf0000, 0x0c000000, COUNT_CLASS},
94 { " AdvSimd ldstmult++", "advlsmp", 0xbfb00000, 0x0c800000, COUNT_CLASS},
95 { " AdvSimd ldst", "advlss", 0xbf9f0000, 0x0d000000, COUNT_CLASS},
96 { " AdvSimd ldst++", "advlssp", 0xbf800000, 0x0d800000, COUNT_CLASS},
97 { " ldst excl", "ldstx", 0x3f000000, 0x08000000, COUNT_CLASS},
98 { " Prefetch", "prfm", 0xff000000, 0xd8000000, COUNT_CLASS},
99 { " Load Reg (lit)", "ldlit", 0x1b000000, 0x18000000, COUNT_CLASS},
100 { " ldst noalloc pair", "ldstnap", 0x3b800000, 0x28000000, COUNT_CLASS},
101 { " ldst pair", "ldstp", 0x38000000, 0x28000000, COUNT_CLASS},
102 { " ldst reg", "ldstr", 0x3b200000, 0x38000000, COUNT_CLASS},
103 { " Atomic ldst", "atomic", 0x3b200c00, 0x38200000, COUNT_CLASS},
104 { " ldst reg (reg off)", "ldstro", 0x3b200b00, 0x38200800, COUNT_CLASS},
105 { " ldst reg (pac)", "ldstpa", 0x3b200200, 0x38200800, COUNT_CLASS},
106 { " ldst reg (imm)", "ldsti", 0x3b000000, 0x39000000, COUNT_CLASS},
107 { "Loads & Stores", "ldst", 0x0a000000, 0x08000000, COUNT_CLASS},
108 /* Data Processing Register */
109 { "Data Proc Reg", "dprr", 0x0e000000, 0x0a000000, COUNT_CLASS},
110 /* Scalar FP */
111 { "Scalar FP ", "fpsimd", 0x0e000000, 0x0e000000, COUNT_CLASS},
112 /* Unclassified */
113 { "Unclassified", "unclas", 0x00000000, 0x00000000, COUNT_CLASS},
114 };
115
116 static InsnClassExecCount sparc32_insn_classes[] = {
117 { "Call", "call", 0xc0000000, 0x40000000, COUNT_CLASS},
118 { "Branch ICond", "bcc", 0xc1c00000, 0x00800000, COUNT_CLASS},
119 { "Branch Fcond", "fbcc", 0xc1c00000, 0x01800000, COUNT_CLASS},
120 { "SetHi", "sethi", 0xc1c00000, 0x01000000, COUNT_CLASS},
121 { "FPU ALU", "fpu", 0xc1f00000, 0x81a00000, COUNT_CLASS},
122 { "ALU", "alu", 0xc0000000, 0x80000000, COUNT_CLASS},
123 { "Load/Store", "ldst", 0xc0000000, 0xc0000000, COUNT_CLASS},
124 /* Unclassified */
125 { "Unclassified", "unclas", 0x00000000, 0x00000000, COUNT_INDIVIDUAL},
126 };
127
128 static InsnClassExecCount sparc64_insn_classes[] = {
129 { "SetHi & Branches", "op0", 0xc0000000, 0x00000000, COUNT_CLASS},
130 { "Call", "op1", 0xc0000000, 0x40000000, COUNT_CLASS},
131 { "Arith/Logical/Move", "op2", 0xc0000000, 0x80000000, COUNT_CLASS},
132 { "Arith/Logical/Move", "op3", 0xc0000000, 0xc0000000, COUNT_CLASS},
133 /* Unclassified */
134 { "Unclassified", "unclas", 0x00000000, 0x00000000, COUNT_INDIVIDUAL},
135 };
136
137 /* Default matcher for currently unclassified architectures */
138 static InsnClassExecCount default_insn_classes[] = {
139 { "Unclassified", "unclas", 0x00000000, 0x00000000, COUNT_INDIVIDUAL},
140 };
141
142 typedef struct {
143 const char *qemu_target;
144 InsnClassExecCount *table;
145 int table_sz;
146 } ClassSelector;
147
148 static ClassSelector class_tables[] = {
149 { "aarch64", aarch64_insn_classes, ARRAY_SIZE(aarch64_insn_classes) },
150 { "sparc", sparc32_insn_classes, ARRAY_SIZE(sparc32_insn_classes) },
151 { "sparc64", sparc64_insn_classes, ARRAY_SIZE(sparc64_insn_classes) },
152 { NULL, default_insn_classes, ARRAY_SIZE(default_insn_classes) },
153 };
154
155 static InsnClassExecCount *class_table;
156 static int class_table_sz;
157
cmp_exec_count(gconstpointer a,gconstpointer b,gpointer d)158 static gint cmp_exec_count(gconstpointer a, gconstpointer b, gpointer d)
159 {
160 InsnExecCount *ea = (InsnExecCount *) a;
161 InsnExecCount *eb = (InsnExecCount *) b;
162 uint64_t count_a = qemu_plugin_u64_sum(ea->count);
163 uint64_t count_b = qemu_plugin_u64_sum(eb->count);
164 return count_a > count_b ? -1 : 1;
165 }
166
free_record(gpointer data)167 static void free_record(gpointer data)
168 {
169 InsnExecCount *rec = (InsnExecCount *) data;
170 qemu_plugin_scoreboard_free(rec->count.score);
171 g_free(rec->insn);
172 g_free(rec);
173 }
174
plugin_exit(qemu_plugin_id_t id,void * p)175 static void plugin_exit(qemu_plugin_id_t id, void *p)
176 {
177 g_autoptr(GString) report = g_string_new("Instruction Classes:\n");
178 int i;
179 uint64_t total_count;
180 GList *counts;
181 InsnClassExecCount *class = NULL;
182
183 for (i = 0; i < class_table_sz; i++) {
184 class = &class_table[i];
185 switch (class->what) {
186 case COUNT_CLASS:
187 total_count = qemu_plugin_u64_sum(class->count);
188 if (total_count || verbose) {
189 g_string_append_printf(report,
190 "Class: %-24s\t(%" PRId64 " hits)\n",
191 class->class,
192 total_count);
193 }
194 break;
195 case COUNT_INDIVIDUAL:
196 g_string_append_printf(report, "Class: %-24s\tcounted individually\n",
197 class->class);
198 break;
199 case COUNT_NONE:
200 g_string_append_printf(report, "Class: %-24s\tnot counted\n",
201 class->class);
202 break;
203 default:
204 break;
205 }
206 }
207
208 counts = g_hash_table_get_values(insns);
209 if (counts && g_list_next(counts)) {
210 g_string_append_printf(report, "Individual Instructions:\n");
211 counts = g_list_sort_with_data(counts, cmp_exec_count, NULL);
212
213 for (i = 0; i < limit && g_list_next(counts);
214 i++, counts = g_list_next(counts)) {
215 InsnExecCount *rec = (InsnExecCount *) counts->data;
216 g_string_append_printf(report,
217 "Instr: %-24s\t(%" PRId64 " hits)"
218 "\t(op=0x%08x/%s)\n",
219 rec->insn,
220 qemu_plugin_u64_sum(rec->count),
221 rec->opcode,
222 rec->class ?
223 rec->class->class : "un-categorised");
224 }
225 g_list_free(counts);
226 }
227
228 g_hash_table_destroy(insns);
229 for (i = 0; i < ARRAY_SIZE(class_tables); i++) {
230 for (int j = 0; j < class_tables[i].table_sz; ++j) {
231 qemu_plugin_scoreboard_free(class_tables[i].table[j].count.score);
232 }
233 }
234
235
236 qemu_plugin_outs(report->str);
237 }
238
plugin_init(void)239 static void plugin_init(void)
240 {
241 insns = g_hash_table_new_full(NULL, g_direct_equal, NULL, &free_record);
242 }
243
vcpu_insn_exec_before(unsigned int cpu_index,void * udata)244 static void vcpu_insn_exec_before(unsigned int cpu_index, void *udata)
245 {
246 struct qemu_plugin_scoreboard *score = udata;
247 qemu_plugin_u64_add(qemu_plugin_scoreboard_u64(score), cpu_index, 1);
248 }
249
find_counter(struct qemu_plugin_insn * insn)250 static struct qemu_plugin_scoreboard *find_counter(
251 struct qemu_plugin_insn *insn)
252 {
253 int i;
254 uint64_t *cnt = NULL;
255 uint32_t opcode = 0;
256 /* if opcode is greater than 32 bits, we should refactor insn hash table. */
257 G_STATIC_ASSERT(sizeof(opcode) == sizeof(uint32_t));
258 InsnClassExecCount *class = NULL;
259
260 /*
261 * We only match the first 32 bits of the instruction which is
262 * fine for most RISCs but a bit limiting for CISC architectures.
263 * They would probably benefit from a more tailored plugin.
264 * However we can fall back to individual instruction counting.
265 */
266 qemu_plugin_insn_data(insn, &opcode, sizeof(opcode));
267
268 for (i = 0; !cnt && i < class_table_sz; i++) {
269 class = &class_table[i];
270 uint32_t masked_bits = opcode & class->mask;
271 if (masked_bits == class->pattern) {
272 break;
273 }
274 }
275
276 g_assert(class);
277
278 switch (class->what) {
279 case COUNT_NONE:
280 return NULL;
281 case COUNT_CLASS:
282 return class->count.score;
283 case COUNT_INDIVIDUAL:
284 {
285 InsnExecCount *icount;
286
287 g_mutex_lock(&lock);
288 icount = (InsnExecCount *) g_hash_table_lookup(insns,
289 (gpointer)(intptr_t) opcode);
290
291 if (!icount) {
292 icount = g_new0(InsnExecCount, 1);
293 icount->opcode = opcode;
294 icount->insn = qemu_plugin_insn_disas(insn);
295 icount->class = class;
296 struct qemu_plugin_scoreboard *score =
297 qemu_plugin_scoreboard_new(sizeof(uint64_t));
298 icount->count = qemu_plugin_scoreboard_u64(score);
299
300 g_hash_table_insert(insns, (gpointer)(intptr_t) opcode, icount);
301 }
302 g_mutex_unlock(&lock);
303
304 return icount->count.score;
305 }
306 default:
307 g_assert_not_reached();
308 }
309
310 return NULL;
311 }
312
vcpu_tb_trans(qemu_plugin_id_t id,struct qemu_plugin_tb * tb)313 static void vcpu_tb_trans(qemu_plugin_id_t id, struct qemu_plugin_tb *tb)
314 {
315 size_t n = qemu_plugin_tb_n_insns(tb);
316 size_t i;
317
318 for (i = 0; i < n; i++) {
319 struct qemu_plugin_insn *insn = qemu_plugin_tb_get_insn(tb, i);
320 struct qemu_plugin_scoreboard *cnt = find_counter(insn);
321
322 if (cnt) {
323 if (do_inline) {
324 qemu_plugin_register_vcpu_insn_exec_inline_per_vcpu(
325 insn, QEMU_PLUGIN_INLINE_ADD_U64,
326 qemu_plugin_scoreboard_u64(cnt), 1);
327 } else {
328 qemu_plugin_register_vcpu_insn_exec_cb(
329 insn, vcpu_insn_exec_before, QEMU_PLUGIN_CB_NO_REGS, cnt);
330 }
331 }
332 }
333 }
334
qemu_plugin_install(qemu_plugin_id_t id,const qemu_info_t * info,int argc,char ** argv)335 QEMU_PLUGIN_EXPORT int qemu_plugin_install(qemu_plugin_id_t id,
336 const qemu_info_t *info,
337 int argc, char **argv)
338 {
339 int i;
340
341 for (i = 0; i < ARRAY_SIZE(class_tables); i++) {
342 for (int j = 0; j < class_tables[i].table_sz; ++j) {
343 struct qemu_plugin_scoreboard *score =
344 qemu_plugin_scoreboard_new(sizeof(uint64_t));
345 class_tables[i].table[j].count = qemu_plugin_scoreboard_u64(score);
346 }
347 }
348
349 /* Select a class table appropriate to the guest architecture */
350 for (i = 0; i < ARRAY_SIZE(class_tables); i++) {
351 ClassSelector *entry = &class_tables[i];
352 if (!entry->qemu_target ||
353 strcmp(entry->qemu_target, info->target_name) == 0) {
354 class_table = entry->table;
355 class_table_sz = entry->table_sz;
356 break;
357 }
358 }
359
360 for (i = 0; i < argc; i++) {
361 char *p = argv[i];
362 g_auto(GStrv) tokens = g_strsplit(p, "=", -1);
363 if (g_strcmp0(tokens[0], "inline") == 0) {
364 if (!qemu_plugin_bool_parse(tokens[0], tokens[1], &do_inline)) {
365 fprintf(stderr, "boolean argument parsing failed: %s\n", p);
366 return -1;
367 }
368 } else if (g_strcmp0(tokens[0], "verbose") == 0) {
369 if (!qemu_plugin_bool_parse(tokens[0], tokens[1], &verbose)) {
370 fprintf(stderr, "boolean argument parsing failed: %s\n", p);
371 return -1;
372 }
373 } else if (g_strcmp0(tokens[0], "count") == 0) {
374 char *value = tokens[1];
375 int j;
376 CountType type = COUNT_INDIVIDUAL;
377 if (*value == '!') {
378 type = COUNT_NONE;
379 value++;
380 }
381 for (j = 0; j < class_table_sz; j++) {
382 if (strcmp(value, class_table[j].opt) == 0) {
383 class_table[j].what = type;
384 break;
385 }
386 }
387 } else {
388 fprintf(stderr, "option parsing failed: %s\n", p);
389 return -1;
390 }
391 }
392
393 plugin_init();
394
395 qemu_plugin_register_vcpu_tb_trans_cb(id, vcpu_tb_trans);
396 qemu_plugin_register_atexit_cb(id, plugin_exit, NULL);
397 return 0;
398 }
399