![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to profiledsp.c CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [HATARI the Atari ST Emulator] / hatari / src / debug|
Return to profiledsp.c CVS log | Up to [HATARI the Atari ST Emulator] / hatari / src / debug |
1.1 root 1: /*
2: * Hatari - profiledsp.c
3: *
1.1.1.3 root 4: * Copyright (C) 2010-2015 by Eero Tamminen
1.1 root 5: *
6: * This file is distributed under the GNU General Public License, version 2
7: * or at your option any later version. Read the file gpl.txt for details.
8: *
9: * profiledsp.c - functions for profiling DSP and showing the results.
10: */
11: const char Profiledsp_fileid[] = "Hatari profiledsp.c : " __DATE__ " " __TIME__;
12:
13: #include <stdio.h>
14: #include <inttypes.h>
1.1.1.4 ! root 15: #include <limits.h>
1.1 root 16: #include <assert.h>
17: #include "main.h"
18: #include "configuration.h"
19: #include "clocks_timings.h"
20: #include "dsp.h"
1.1.1.4 ! root 21: #include "symbols.h"
1.1 root 22: #include "profile.h"
23: #include "profile_priv.h"
1.1.1.4 ! root 24: #include "debug_priv.h"
1.1.1.2 root 25: /* for VBL info */
26: #include "screen.h"
27: #include "video.h"
1.1 root 28:
29: static callinfo_t dsp_callinfo;
30:
31: #define DSP_PROFILE_ARR_SIZE 0x10000
32: #define MAX_DSP_PROFILE_VALUE 0xFFFFFFFFFFFFFFFFLL
33:
34: typedef struct {
35: Uint64 count; /* how many times this address is used */
36: Uint64 cycles; /* how many DSP cycles was taken at this address */
37: Uint16 min_cycle;
38: Uint16 max_cycle;
39: } dsp_profile_item_t;
40:
41: static struct {
42: dsp_profile_item_t *data; /* profile data */
43: profile_area_t ram; /* statistics for whole memory */
44: Uint16 *sort_arr; /* data indexes used for sorting */
45: Uint16 prev_pc; /* previous PC for which the cycles are for */
1.1.1.2 root 46: Uint16 loop_start; /* address of last loop start */
47: Uint16 loop_end; /* address of last loop end */
48: Uint32 loop_count; /* how many times it was looped */
1.1 root 49: Uint32 disasm_addr; /* 'dspaddresses' command start address */
50: bool processed; /* true when data is already processed */
51: bool enabled; /* true when profiling enabled */
52: } dsp_profile;
53:
54:
55: /* ------------------ DSP profile results ----------------- */
56:
57: /**
58: * Get DSP cycles, count and count percentage for given address.
59: * Return true if data was available and non-zero, false otherwise.
60: */
61: bool Profile_DspAddressData(Uint16 addr, float *percentage, Uint64 *count, Uint64 *cycles, Uint16 *cycle_diff)
62: {
63: dsp_profile_item_t *item;
64: if (!dsp_profile.data) {
65: return false;
66: }
67: item = dsp_profile.data + addr;
68:
69: *cycles = item->cycles;
70: *count = item->count;
71: if (item->max_cycle) {
72: *cycle_diff = item->max_cycle - item->min_cycle;
73: } else {
74: *cycle_diff = 0;
75: }
76: if (dsp_profile.ram.counters.count) {
77: *percentage = 100.0*(*count)/dsp_profile.ram.counters.count;
78: } else {
79: *percentage = 0.0;
80: }
81: return (*count > 0);
82: }
83:
84: /**
85: * show DSP specific profile statistics.
86: */
87: void Profile_DspShowStats(void)
88: {
89: profile_area_t *area = &dsp_profile.ram;
90: fprintf(stderr, "DSP profile statistics (0x0-0xFFFF):\n");
91: if (!area->active) {
92: fprintf(stderr, "- no activity\n");
93: return;
94: }
95: fprintf(stderr, "- active address range:\n 0x%04x-0x%04x\n",
96: area->lowest, area->highest);
97: fprintf(stderr, "- active instruction addresses:\n %d\n",
98: area->active);
99: fprintf(stderr, "- executed instructions:\n %"PRIu64"\n",
100: area->counters.count);
101: /* indicates either instruction(s) that address different memory areas
102: * (they can have different access costs), or more significantly,
103: * DSP code that has changed during profiling.
104: */
105: fprintf(stderr, "- sum of per instruction cycle changes\n"
106: " (can indicate code change during profiling):\n %"PRIu64"\n",
1.1.1.3 root 107: area->counters.cycles_diffs);
1.1 root 108:
109: fprintf(stderr, "- used cycles:\n %"PRIu64"\n",
110: area->counters.cycles);
111: if (area->overflow) {
112: fprintf(stderr, " *** COUNTERS OVERFLOW! ***\n");
113: }
114: fprintf(stderr, "\n= %.5fs\n", (double)(area->counters.cycles) / MachineClocks.DSP_Freq);
115: }
116:
117: /**
118: * Show DSP instructions which execution was profiled, in the address order,
119: * starting from the given address. Return next disassembly address.
120: */
1.1.1.4 ! root 121: Uint16 Profile_DspShowAddresses(Uint32 addr, Uint32 upper, FILE *out, paging_t use_paging)
1.1 root 122: {
1.1.1.4 ! root 123: int show, shown, addrs, active;
1.1 root 124: dsp_profile_item_t *data;
125: Uint16 nextpc;
126: Uint32 end;
127: const char *symbol;
128:
129: data = dsp_profile.data;
130: if (!data) {
131: fprintf(stderr, "ERROR: no DSP profiling data available!\n");
132: return 0;
133: }
134:
135: end = DSP_PROFILE_ARR_SIZE;
136: active = dsp_profile.ram.active;
137: if (upper) {
138: if (upper < end) {
139: end = upper;
140: }
141: show = active;
142: } else {
1.1.1.4 ! root 143: show = DebugUI_GetPageLines(ConfigureParams.Debugger.nDisasmLines, 0);
1.1 root 144: if (!show || show > active) {
145: show = active;
146: }
147: }
1.1.1.4 ! root 148: if (use_paging == PAGING_DISABLED) {
! 149: show = INT_MAX;
! 150: }
1.1 root 151:
152: fputs("# disassembly with profile data: <instructions percentage>% (<sum of instructions>, <sum of cycles>, <max cycle difference>)\n", out);
1.1.1.4 ! root 153: shown = 2; /* first and last printf */
1.1 root 154:
1.1.1.4 ! root 155: addrs = nextpc = 0;
! 156: for (; shown < show && addr < end; addr++) {
1.1 root 157: if (!data[addr].count) {
158: continue;
159: }
160: if (addr != nextpc && nextpc) {
161: fputs("[...]\n", out);
1.1.1.4 ! root 162: shown++;
1.1 root 163: }
1.1.1.4 ! root 164: symbol = Symbols_GetByDspAddress(addr, SYMTYPE_TEXT);
1.1 root 165: if (symbol) {
166: fprintf(out, "%s:\n", symbol);
1.1.1.4 ! root 167: shown++;
1.1 root 168: }
169: nextpc = DSP_DisasmAddress(out, addr, addr);
1.1.1.4 ! root 170: addrs++;
1.1 root 171: shown++;
172: }
1.1.1.4 ! root 173: printf("Disassembled %d (of active %d) DSP addresses.\n", addrs, active);
1.1 root 174: return nextpc;
175: }
176:
177: /**
178: * compare function for qsort() to sort DSP profile data by descdending
179: * address cycles counts.
180: */
181: static int cmp_dsp_cycles(const void *p1, const void *p2)
182: {
183: Uint64 count1 = dsp_profile.data[*(const Uint16*)p1].cycles;
184: Uint64 count2 = dsp_profile.data[*(const Uint16*)p2].cycles;
185: if (count1 > count2) {
186: return -1;
187: }
188: if (count1 < count2) {
189: return 1;
190: }
191: return 0;
192: }
193:
194: /**
195: * Sort DSP profile data addresses by cycle counts and show the results.
196: */
197: void Profile_DspShowCycles(int show)
198: {
199: int active;
200: Uint16 *sort_arr, *end, addr;
201: dsp_profile_item_t *data = dsp_profile.data;
202: float percentage;
203: Uint64 count;
204:
205: if (!data) {
206: fprintf(stderr, "ERROR: no DSP profiling data available!\n");
207: return;
208: }
209:
210: active = dsp_profile.ram.active;
211: sort_arr = dsp_profile.sort_arr;
212: qsort(sort_arr, active, sizeof(*sort_arr), cmp_dsp_cycles);
213:
214: printf("addr:\tcycles:\n");
215: show = (show < active ? show : active);
216: for (end = sort_arr + show; sort_arr < end; sort_arr++) {
217: addr = *sort_arr;
218: count = data[addr].cycles;
219: percentage = 100.0*count/dsp_profile.ram.counters.cycles;
220: printf("0x%04x\t%5.2f%%\t%"PRIu64"%s\n", addr, percentage, count,
221: count == MAX_DSP_PROFILE_VALUE ? " (OVERFLOW)" : "");
222: }
223: printf("%d DSP addresses listed.\n", show);
224: }
225:
226:
227: /**
228: * compare function for qsort() to sort DSP profile data by descdending
229: * address access counts.
230: */
231: static int cmp_dsp_count(const void *p1, const void *p2)
232: {
233: Uint64 count1 = dsp_profile.data[*(const Uint16*)p1].count;
234: Uint64 count2 = dsp_profile.data[*(const Uint16*)p2].count;
235: if (count1 > count2) {
236: return -1;
237: }
238: if (count1 < count2) {
239: return 1;
240: }
241: return 0;
242: }
243:
244: /**
245: * Sort DSP profile data addresses by call counts and show the results.
246: * If symbols are requested and symbols are loaded, show (only) addresses
247: * matching a symbol.
248: */
249: void Profile_DspShowCounts(int show, bool only_symbols)
250: {
251: dsp_profile_item_t *data = dsp_profile.data;
252: int symbols, matched, active;
253: Uint16 *sort_arr, *end, addr;
254: const char *name;
255: float percentage;
256: Uint64 count;
257:
258: if (!data) {
259: fprintf(stderr, "ERROR: no DSP profiling data available!\n");
260: return;
261: }
262: active = dsp_profile.ram.active;
263: show = (show < active ? show : active);
264:
265: sort_arr = dsp_profile.sort_arr;
266: qsort(sort_arr, active, sizeof(*sort_arr), cmp_dsp_count);
267:
268: if (!only_symbols) {
269: printf("addr:\tcount:\n");
270: for (end = sort_arr + show; sort_arr < end; sort_arr++) {
271: addr = *sort_arr;
272: count = data[addr].count;
273: percentage = 100.0*count/dsp_profile.ram.counters.count;
274: printf("0x%04x\t%5.2f%%\t%"PRIu64"%s\n",
275: addr, percentage, count,
276: count == MAX_DSP_PROFILE_VALUE ? " (OVERFLOW)" : "");
277: }
278: printf("%d DSP addresses listed.\n", show);
279: return;
280: }
281:
1.1.1.4 ! root 282: symbols = Symbols_DspCodeCount();
1.1 root 283: if (!symbols) {
284: fprintf(stderr, "ERROR: no DSP symbols loaded!\n");
285: return;
286: }
287: matched = 0;
288:
289: printf("addr:\tcount:\t\tsymbol:\n");
290: for (end = sort_arr + active; sort_arr < end; sort_arr++) {
291:
292: addr = *sort_arr;
1.1.1.4 ! root 293: name = Symbols_GetByDspAddress(addr, SYMTYPE_TEXT);
1.1 root 294: if (!name) {
295: continue;
296: }
297: count = data[addr].count;
298: percentage = 100.0*count/dsp_profile.ram.counters.count;
299: printf("0x%04x\t%.2f%%\t%"PRIu64"\t%s%s\n",
300: addr, percentage, count, name,
301: count == MAX_DSP_PROFILE_VALUE ? " (OVERFLOW)" : "");
302:
303: matched++;
304: if (matched >= show || matched >= symbols) {
305: break;
306: }
307: }
308: printf("%d DSP symbols listed.\n", matched);
309: }
310:
311:
312: static const char * addr2name(Uint32 addr, Uint64 *total)
313: {
314: *total = dsp_profile.data[addr].count;
1.1.1.4 ! root 315: return Symbols_GetByDspAddress(addr, SYMTYPE_TEXT);
1.1 root 316: }
317:
318: /**
319: * Output DSP callers info to given file.
320: */
321: void Profile_DspShowCallers(FILE *fp)
322: {
323: Profile_ShowCallers(fp, dsp_callinfo.sites, dsp_callinfo.site, addr2name);
324: }
325:
326: /**
327: * Save DSP profile information to given file.
328: */
329: void Profile_DspSave(FILE *out)
330: {
331: /* Comma separated descriptions for the profile disassembly data fields.
332: * Instructions and cycles need to be first two fields!
333: */
334: fputs("Field names:\tExecuted instructions, Used cycles, Largest cycle differences (= code changes during profiling)\n", out);
335: /* (Python) pegexp that matches address and all describled fields from disassembly:
336: * <space>:<address> <opcodes> (<instr cycles>) <instr> <count>% (<count>, <cycles>)
337: * p:0202 0aa980 000200 (07 cyc) jclr #0,x:$ffe9,p:$0200 0.00% (6, 42)
338: */
339: fputs("Field regexp:\t^p:([0-9a-f]+) .*% \\((.*)\\)$\n", out);
1.1.1.4 ! root 340: Profile_DspShowAddresses(0, DSP_PROFILE_ARR_SIZE, out, PAGING_DISABLED);
1.1 root 341: Profile_DspShowCallers(out);
342: }
343:
344: /* ------------------ DSP profile control ----------------- */
345:
346: /**
347: * Initialize DSP profiling when necessary. Return true if profiling.
348: */
349: bool Profile_DspStart(void)
350: {
351: dsp_profile_item_t *item;
352: int i;
353:
354: Profile_FreeCallinfo(&(dsp_callinfo));
355: if (dsp_profile.sort_arr) {
356: /* remove previous results */
357: free(dsp_profile.sort_arr);
358: free(dsp_profile.data);
359: dsp_profile.sort_arr = NULL;
360: dsp_profile.data = NULL;
361: printf("Freed previous DSP profile buffers.\n");
362: }
363: if (!dsp_profile.enabled) {
364: return false;
365: }
366: /* zero everything */
367: memset(&dsp_profile, 0, sizeof(dsp_profile));
368:
369: dsp_profile.data = calloc(DSP_PROFILE_ARR_SIZE, sizeof(*dsp_profile.data));
370: if (!dsp_profile.data) {
371: perror("ERROR, new DSP profile buffer alloc failed");
372: return false;
373: }
374: printf("Allocated DSP profile buffer (%d KB).\n",
375: (int)sizeof(*dsp_profile.data)*DSP_PROFILE_ARR_SIZE/1024);
376:
1.1.1.4 ! root 377: Profile_AllocCallinfo(&(dsp_callinfo), Symbols_DspCodeCount(), "DSP");
1.1 root 378:
379: item = dsp_profile.data;
380: for (i = 0; i < DSP_PROFILE_ARR_SIZE; i++, item++) {
381: item->min_cycle = 0xFFFF;
382: }
383: dsp_profile.prev_pc = DSP_GetPC();
384:
1.1.1.2 root 385: dsp_profile.loop_start = 0xFFFF;
386: dsp_profile.loop_end = 0xFFFF;
387: dsp_profile.loop_count = 0;
388: Profile_LoopReset();
389:
1.1 root 390: dsp_profile.disasm_addr = 0;
391: dsp_profile.processed = false;
392: dsp_profile.enabled = true;
393: return dsp_profile.enabled;
394: }
395:
396: /* return true if pc is next instruction for previous pc */
397: static bool is_prev_instr(Uint16 prev_pc, Uint16 pc)
398: {
399: /* just moved to next instruction (1-2 words)? */
400: if (prev_pc < pc && (pc - prev_pc) <= 4) {
401: return true;
402: }
403: return false;
404: }
405:
406: /* return branch type based on caller instruction type */
407: static calltype_t dsp_opcode_type(Uint16 prev_pc, Uint16 pc)
408: {
409: const char *dummy;
410: Uint32 opcode;
411:
412: /* 24-bit instruction opcode */
413: opcode = DSP_ReadMemory(prev_pc, 'P', &dummy) & 0xFFFFFF;
414:
415: /* subroutine returns */
416: if (opcode == 0xC) { /* (just) RTS */
417: return CALL_SUBRETURN;
418: }
419: /* unconditional subroutine calls */
420: if ((opcode & 0xFFF000) == 0xD0000 || /* JSR 00001101 0000aaaa aaaaaaaa */
421: (opcode & 0xFFC0FF) == 0xBC080) { /* JSR 00001011 11MMMRRR 10000000 */
422: return CALL_SUBROUTINE;
423: }
424: /* conditional subroutine calls */
425: if ((opcode & 0xFF0000) == 0xF0000 || /* JSCC 00001111 CCCCaaaa aaaaaaaa */
426: (opcode & 0xFFC0F0) == 0xBC0A0 || /* JSCC 00001011 11MMMRRR 1010CCCC */
427: (opcode & 0xFFC0A0) == 0xB4080 || /* JSCLR 00001011 01MMMRRR 1S0bbbbb */
428: (opcode & 0xFFC0A0) == 0xB0080 || /* JSCLR 00001011 00aaaaaa 1S0bbbbb */
429: (opcode & 0xFFC0A0) == 0xB8080 || /* JSCLR 00001011 10pppppp 1S0bbbbb */
430: (opcode & 0xFFC0E0) == 0xBC000 || /* JSCLR 00001011 11DDDDDD 000bbbbb */
431: (opcode & 0xFFC0A0) == 0xB40A0 || /* JSSET 00001011 01MMMRRR 1S1bbbbb */
432: (opcode & 0xFFC0A0) == 0xB00A0 || /* JSSET 00001011 00aaaaaa 1S1bbbbb */
433: (opcode & 0xFFC0A0) == 0xB80A0 || /* JSSET 00001011 10pppppp 1S1bbbbb */
434: (opcode & 0xFFC0E0) == 0xBC020) { /* JSSET 00001011 11DDDDDD 001bbbbb */
435: /* hopefully fairly safe heuristic:
436: * if previously executed instruction
437: * was one before current one, no
438: * subroutine call was made to next
439: * instruction, the condition just
440: * wasn't met.
441: */
442: if (is_prev_instr(prev_pc, pc)) {
443: return CALL_NEXT;
444: }
445: return CALL_SUBROUTINE;
446: }
447: /* exception handler returns */
448: if (opcode == 0x4) { /* (just) RTI */
449: return CALL_EXCRETURN;
450: }
451:
452: /* Besides CALL_UNKNOWN, rest isn't used by subroutine call
453: * cost collection. However, it's useful info when debugging
454: * code or reading full callgraphs (because optimized code uses
455: * also jumps/branches for subroutine calls).
456: */
457:
458: /* TODO: exception invocation.
459: * Could be detected by PC going through low interrupt vector adresses,
460: * but fast-calls using JSR/RTS would need separate handling.
461: */
462: if (0) { /* TODO */
463: return CALL_EXCEPTION;
464: }
465: /* branches */
466: if ((opcode & 0xFFF000) == 0xC0000 || /* JMP 00001100 0000aaaa aaaaaaaa */
467: (opcode & 0xFFC0FF) == 0xAC080 || /* JMP 00001010 11MMMRRR 10000000 */
468: (opcode & 0xFF0000) == 0xE0000 || /* JCC 00001110 CCCCaaaa aaaaaaaa */
469: (opcode & 0xFFC0F0) == 0xAC0A0 || /* JCC 00001010 11MMMRRR 1010CCCC */
470: (opcode & 0xFFC0A0) == 0xA8080 || /* JCLR 00001010 10pppppp 1S0bbbbb */
471: (opcode & 0xFFC0A0) == 0xA4080 || /* JCLR 00001010 01MMMRRR 1S0bbbbb */
472: (opcode & 0xFFC0A0) == 0xA0080 || /* JCLR 00001010 00aaaaaa 1S0bbbbb */
473: (opcode & 0xFFC0E0) == 0xAC000 || /* JCLR 00001010 11dddddd 000bbbbb */
474: (opcode & 0xFFC0A0) == 0xA80A0 || /* JSET 00001010 10pppppp 1S1bbbbb */
475: (opcode & 0xFFC0A0) == 0xA40A0 || /* JSET 00001010 01MMMRRR 1S1bbbbb */
476: (opcode & 0xFFC0A0) == 0xA00A0 || /* JSET 00001010 00aaaaaa 1S1bbbbb */
477: (opcode & 0xFFC0E0) == 0xAC020 || /* JSET 00001010 11dddddd 001bbbbb */
478: (opcode & 0xFF00F0) == 0x600A0 || /* REP 00000110 iiiiiiii 1010hhhh */
479: (opcode & 0xFFC0FF) == 0x6C020 || /* REP 00000110 11dddddd 00100000 */
480: (opcode & 0xFFC0BF) == 0x64020 || /* REP 00000110 01MMMRRR 0s100000 */
481: (opcode & 0xFFC0BF) == 0x60020 || /* REP 00000110 00aaaaaa 0s100000 */
482: (opcode & 0xFF00F0) == 0x60080 || /* DO/ENDO 00000110 iiiiiiii 1000hhhh */
483: (opcode & 0xFFC0FF) == 0x6C000 || /* DO/ENDO 00000110 11DDDDDD 00000000 */
484: (opcode & 0xFFC0BF) == 0x64000 || /* DO/ENDO 00000110 01MMMRRR 0S000000 */
485: (opcode & 0xFFC0BF) == 0x60000) { /* DO/ENDO 00000110 00aaaaaa 0S000000 */
486: return CALL_BRANCH;
487: }
488: if (is_prev_instr(prev_pc, pc)) {
489: return CALL_NEXT;
490: }
491: return CALL_UNKNOWN;
492: }
493:
494: /**
495: * If call tracking is enabled (there are symbols), collect
496: * information about subroutine and other calls, and their costs.
497: *
498: * Like with profile data, caller info checks need to be for previous
499: * instruction, that's why "pc" argument for this function actually
500: * needs to be previous PC.
501: */
502: static void collect_calls(Uint16 pc, counters_t *counters)
503: {
504: calltype_t flag;
505: Uint16 prev_pc;
506: Uint32 caller_pc;
507: int idx;
508:
509: prev_pc = dsp_callinfo.prev_pc;
510: dsp_callinfo.prev_pc = pc;
511: caller_pc = PC_UNDEFINED;
512:
513: /* address is return address for last subroutine call? */
514: if (unlikely(pc == dsp_callinfo.return_pc) && likely(dsp_callinfo.depth)) {
515:
516: flag = dsp_opcode_type(prev_pc, pc);
517: /* return address is entered either by subroutine return,
518: * or by returning from exception that interrupted
519: * the instruction at return address.
520: */
521: if (likely(flag == CALL_SUBRETURN || flag == CALL_EXCRETURN)) {
522: caller_pc = Profile_CallEnd(&dsp_callinfo, counters);
523: }
524: }
525:
526: /* address is one which we're tracking? */
1.1.1.4 ! root 527: idx = Symbols_GetDspCodeIndex(pc);
1.1 root 528: if (unlikely(idx >= 0)) {
529:
530: flag = dsp_opcode_type(prev_pc, pc);
531: if (flag == CALL_SUBROUTINE) {
532: dsp_callinfo.return_pc = DSP_GetNextPC(prev_pc); /* slow! */
533: } else if (caller_pc != PC_UNDEFINED) {
534: /* returned from function, change return
535: * instruction address to address of
536: * what did the returned call.
537: */
538: prev_pc = caller_pc;
539: assert(is_prev_instr(prev_pc, pc));
540: flag = CALL_NEXT;
541: }
542: Profile_CallStart(idx, &dsp_callinfo, prev_pc, flag, pc, counters);
543:
544: }
545: }
546:
547: /**
1.1.1.2 root 548: * log last loop info, if there's suitable data for one
549: */
550: static void log_last_loop(void)
551: {
552: unsigned len = dsp_profile.loop_end - dsp_profile.loop_start;
553: if (dsp_profile.loop_count > 1 && (len < profile_loop.dsp_limit || !profile_loop.dsp_limit)) {
554: fprintf(profile_loop.fp, "DSP %d 0x%04x %d %d\n", nVBLs,
555: dsp_profile.loop_start, len, dsp_profile.loop_count);
556: fflush(profile_loop.fp);
557: }
558: }
559:
560: /**
1.1 root 561: * Update DSP cycle and count statistics for PC address.
562: *
563: * This is called after instruction is executed and PC points
564: * to next instruction i.e. info is for previous PC address.
565: */
566: void Profile_DspUpdate(void)
567: {
568: dsp_profile_item_t *prev;
569: Uint16 pc, prev_pc, cycles;
570: counters_t *counters;
571:
572: prev_pc = dsp_profile.prev_pc;
573: dsp_profile.prev_pc = pc = DSP_GetPC();
1.1.1.2 root 574:
575: if (unlikely(profile_loop.fp)) {
576: if (pc < prev_pc) {
577: if (pc == dsp_profile.loop_start && prev_pc == dsp_profile.loop_end) {
578: dsp_profile.loop_count++;
579: } else {
580: dsp_profile.loop_start = pc;
581: dsp_profile.loop_end = prev_pc;
582: dsp_profile.loop_count = 1;
583: }
584: } else {
585: if (pc > dsp_profile.loop_end) {
586: log_last_loop();
587: dsp_profile.loop_end = 0xFFFF;
588: dsp_profile.loop_count = 0;
589: }
590: }
591: }
592:
1.1 root 593: prev = dsp_profile.data + prev_pc;
594:
595: if (likely(prev->count < MAX_DSP_PROFILE_VALUE)) {
596: prev->count++;
597: }
598:
599: cycles = DSP_GetInstrCycles();
600: if (likely(prev->cycles < MAX_DSP_PROFILE_VALUE - cycles)) {
601: prev->cycles += cycles;
602: } else {
603: prev->cycles = MAX_DSP_PROFILE_VALUE;
604: }
605:
606: if (unlikely(cycles < prev->min_cycle)) {
607: prev->min_cycle = cycles;
608: }
609: if (unlikely(cycles > prev->max_cycle)) {
610: prev->max_cycle = cycles;
611: }
612:
613: counters = &(dsp_profile.ram.counters);
614: if (dsp_callinfo.sites) {
615: collect_calls(prev_pc, counters);
616: }
617: /* counters are increased after caller info is processed,
618: * otherwise cost for the instruction calling the callee
619: * doesn't get accounted to caller (but callee).
620: */
621: counters->cycles += cycles;
622: counters->count++;
623: }
624:
625: /**
626: * Helper for collecting DSP profile area statistics.
627: */
628: static void update_area_item(profile_area_t *area, Uint16 addr, dsp_profile_item_t *item)
629: {
630: Uint64 cycles = item->cycles;
631: Uint64 count = item->count;
632: Uint16 diff;
633:
634: if (!count) {
635: return;
636: }
637: if (cycles == MAX_DSP_PROFILE_VALUE) {
638: area->overflow = true;
639: }
640: if (item->max_cycle) {
641: diff = item->max_cycle - item->min_cycle;
642: } else {
643: diff = 0;
644: }
645:
646: area->counters.count += count;
647: area->counters.cycles += cycles;
1.1.1.3 root 648: area->counters.cycles_diffs += diff;
1.1 root 649:
650: if (addr < area->lowest) {
651: area->lowest = addr;
652: }
653: area->highest = addr;
654:
655: area->active++;
656: }
657:
658: /**
659: * Stop and process the DSP profiling data; collect stats and
660: * prepare for more optimal sorting.
661: */
662: void Profile_DspStop(void)
663: {
664: dsp_profile_item_t *item;
665: profile_area_t *area;
666: Uint16 *sort_arr;
667: Uint32 addr;
668:
669: if (dsp_profile.processed || !dsp_profile.enabled) {
670: return;
671: }
672:
1.1.1.2 root 673: log_last_loop();
674: if (profile_loop.fp) {
675: fflush(profile_loop.fp);
676: }
677:
1.1 root 678: Profile_FinalizeCalls(&(dsp_callinfo), &(dsp_profile.ram.counters), Symbols_GetByDspAddress);
679:
680: /* find lowest and highest addresses executed */
681: area = &dsp_profile.ram;
682: memset(area, 0, sizeof(profile_area_t));
683: area->lowest = DSP_PROFILE_ARR_SIZE;
684:
685: item = dsp_profile.data;
686: for (addr = 0; addr < DSP_PROFILE_ARR_SIZE; addr++, item++) {
687: update_area_item(area, addr, item);
688: }
689:
690: /* allocate address array for sorting */
691: sort_arr = calloc(dsp_profile.ram.active, sizeof(*sort_arr));
692:
693: if (!sort_arr) {
694: perror("ERROR: allocating DSP profile address data");
695: free(dsp_profile.data);
696: dsp_profile.data = NULL;
697: return;
698: }
699: printf("Allocated DSP profile address buffer (%d KB).\n",
700: (int)sizeof(*sort_arr)*(dsp_profile.ram.active+512)/1024);
701: dsp_profile.sort_arr = sort_arr;
702:
703: /* ...and fill addresses for used instructions... */
704: area = &dsp_profile.ram;
705: item = &(dsp_profile.data[area->lowest]);
706: for (addr = area->lowest; addr <= area->highest; addr++, item++) {
707: if (item->count) {
708: *sort_arr++ = addr;
709: }
710: }
711: //printf("%d/%d/%d\n", area->active, sort_arr-dsp_profile.sort_arr, active);
712:
713: Profile_DspShowStats();
714: dsp_profile.processed = true;
715: }
716:
717: /**
718: * Get pointers to DSP profile enabling and disasm address variables
719: * for updating them (in parser).
720: */
721: void Profile_DspGetPointers(bool **enabled, Uint32 **disasm_addr)
722: {
723: *disasm_addr = &dsp_profile.disasm_addr;
724: *enabled = &dsp_profile.enabled;
725: }
726:
727: /**
728: * Get callinfo & symbol search pointers for stack walking.
729: */
1.1.1.4 ! root 730: void Profile_DspGetCallinfo(callinfo_t **callinfo, const char* (**get_symbol)(Uint32, symtype_t))
1.1 root 731: {
732: *callinfo = &(dsp_callinfo);
733: *get_symbol = Symbols_GetByDspAddress;
734: }
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.