intel/intel_decode.c

   1 /*
   2  * Copyright © 2009-2011 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25 #include <stdint.h>
  26 #include <stdlib.h>
  27 #include <stdio.h>
  28 #include <stdbool.h>
  29 #include <stdarg.h>
  30 #include <string.h>
  31
  32 #include "intel_chipset.h"
  33 #include "intel_bufmgr.h"
  34
  35 /* Struct for tracking drm_intel_decode state. */
  36 struct drm_intel_decode {
  37         /** stdio file where the output should land.  Defaults to stdout. */
  38         FILE *out;
  39
  40         /** PCI device ID. */
  41         uint32_t devid;
  42
  43         /**
  44          * Shorthand device identifier: 3 is 915, 4 is 965, 5 is
  45          * Ironlake, etc.
  46          */
  47         int gen;
  48
  49         /** GPU address of the start of the current packet. */
  50         uint32_t hw_offset;
  51         /** CPU virtual address of the start of the current packet. */
  52         uint32_t *data;
  53         /** DWORDs of remaining batchbuffer data starting from the packet. */
  54         uint32_t count;
  55
  56         /** GPU address of the start of the batchbuffer data. */
  57         uint32_t base_hw_offset;
  58         /** CPU Virtual address of the start of the batchbuffer data. */
  59         uint32_t *base_data;
  60         /** Number of DWORDs of batchbuffer data. */
  61         uint32_t base_count;
  62
  63         /** @{
  64          * GPU head and tail pointers, which will be noted in the dump, or ~0.
  65          */
  66         uint32_t head, tail;
  67         /** @} */
  68
  69         /**
  70          * Whether to dump the dwords after MI_BATCHBUFFER_END.
  71          *
  72          * This sometimes provides clues in corrupted batchbuffers,
  73          * and is used by the intel-gpu-tools.
  74          */
  75         bool dump_past_end;
  76
  77         bool overflowed;
  78 };
  79
  80 static FILE *out;
  81 static uint32_t saved_s2 = 0, saved_s4 = 0;
  82 static char saved_s2_set = 0, saved_s4_set = 0;
  83 static uint32_t head_offset = 0xffffffff;       /* undefined */
  84 static uint32_t tail_offset = 0xffffffff;       /* undefined */
  85
  86 #ifndef ARRAY_SIZE
  87 #define ARRAY_SIZE(A) (sizeof(A)/sizeof(A[0]))
  88 #endif
  89
  90 #define BUFFER_FAIL(_count, _len, _name) do {                   \
  91     fprintf(out, "Buffer size too small in %s (%d < %d)\n",     \
  92             (_name), (_count), (_len));                         \
  93     return _count;                                              \
  94 } while (0)
  95
  96 static float int_as_float(uint32_t intval)
  97 {
  98         union intfloat {
  99                 uint32_t i;
 100                 float f;
 101         } uval;
 102
 103         uval.i = intval;
 104         return uval.f;
 105 }
 106
 107 static void
 108 instr_out(struct drm_intel_decode *ctx, unsigned int index,
 109           const char *fmt, ...) __attribute__((format(__printf__, 3, 4)));
 110
 111 static void
 112 instr_out(struct drm_intel_decode *ctx, unsigned int index,
 113           const char *fmt, ...)
 114 {
 115         va_list va;
 116         const char *parseinfo;
 117         uint32_t offset = ctx->hw_offset + index * 4;
 118
 119         if (index > ctx->count) {
 120                 if (!ctx->overflowed) {
 121                         fprintf(out, "ERROR: Decode attempted to continue beyond end of batchbuffer\n");
 122                         ctx->overflowed = true;
 123                 }
 124                 return;
 125         }
 126
 127         if (offset == head_offset)
 128                 parseinfo = "HEAD";
 129         else if (offset == tail_offset)
 130                 parseinfo = "TAIL";
 131         else
 132                 parseinfo = "    ";
 133
 134         fprintf(out, "0x%08x: %s 0x%08x: %s", offset, parseinfo,
 135                 ctx->data[index], index == 0 ? "" : "   ");
 136         va_start(va, fmt);
 137         vfprintf(out, fmt, va);
 138         va_end(va);
 139 }
 140
 141 static int
 142 decode_MI_WAIT_FOR_EVENT(struct drm_intel_decode *ctx)
 143 {
 144         const char *cc_wait;
 145         int cc_shift = 0;
 146         uint32_t data = ctx->data[0];
 147
 148         if (ctx->gen <= 5)
 149                 cc_shift = 9;
 150         else
 151                 cc_shift = 16;
 152
 153         switch ((data >> cc_shift) & 0x1f) {
 154         case 1:
 155                 cc_wait = ", cc wait 1";
 156                 break;
 157         case 2:
 158                 cc_wait = ", cc wait 2";
 159                 break;
 160         case 3:
 161                 cc_wait = ", cc wait 3";
 162                 break;
 163         case 4:
 164                 cc_wait = ", cc wait 4";
 165                 break;
 166         case 5:
 167                 cc_wait = ", cc wait 4";
 168                 break;
 169         default:
 170                 cc_wait = "";
 171                 break;
 172         }
 173
 174         if (ctx->gen <= 5) {
 175                 instr_out(ctx, 0, "MI_WAIT_FOR_EVENT%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
 176                           data & (1<<18)? ", pipe B start vblank wait": "",
 177                           data & (1<<17)? ", pipe A start vblank wait": "",
 178                           data & (1<<16)? ", overlay flip pending wait": "",
 179                           data & (1<<14)? ", pipe B hblank wait": "",
 180                           data & (1<<13)? ", pipe A hblank wait": "",
 181                           cc_wait,
 182                           data & (1<<8)? ", plane C pending flip wait": "",
 183                           data & (1<<7)? ", pipe B vblank wait": "",
 184                           data & (1<<6)? ", plane B pending flip wait": "",
 185                           data & (1<<5)? ", pipe B scan line wait": "",
 186                           data & (1<<4)? ", fbc idle wait": "",
 187                           data & (1<<3)? ", pipe A vblank wait": "",
 188                           data & (1<<2)? ", plane A pending flip wait": "",
 189                           data & (1<<1)? ", plane A scan line wait": "");
 190         } else {
 191                 instr_out(ctx, 0, "MI_WAIT_FOR_EVENT%s%s%s%s%s%s%s%s%s%s%s%s\n",
 192                           data & (1<<20)? ", sprite C pending flip wait": "", /* ivb */
 193                           cc_wait,
 194                           data & (1<<13)? ", pipe B hblank wait": "",
 195                           data & (1<<11)? ", pipe B vblank wait": "",
 196                           data & (1<<10)? ", sprite B pending flip wait": "",
 197                           data & (1<<9)? ", plane B pending flip wait": "",
 198                           data & (1<<8)? ", plane B scan line wait": "",
 199                           data & (1<<5)? ", pipe A hblank wait": "",
 200                           data & (1<<3)? ", pipe A vblank wait": "",
 201                           data & (1<<2)? ", sprite A pending flip wait": "",
 202                           data & (1<<1)? ", plane A pending flip wait": "",
 203                           data & (1<<0)? ", plane A scan line wait": "");
 204         }
 205
 206         return 1;
 207 }
 208
 209 static int
 210 decode_mi(struct drm_intel_decode *ctx)
 211 {
 212         unsigned int opcode, len = -1;
 213         const char *post_sync_op = "";
 214         uint32_t *data = ctx->data;
 215
 216         struct {
 217                 uint32_t opcode;
 218                 int len_mask;
 219                 unsigned int min_len;
 220                 unsigned int max_len;
 221                 const char *name;
 222                 int (*func)(struct drm_intel_decode *ctx);
 223         } opcodes_mi[] = {
 224                 { 0x08, 0, 1, 1, "MI_ARB_ON_OFF" },
 225                 { 0x0a, 0, 1, 1, "MI_BATCH_BUFFER_END" },
 226                 { 0x30, 0x3f, 3, 3, "MI_BATCH_BUFFER" },
 227                 { 0x31, 0x3f, 2, 2, "MI_BATCH_BUFFER_START" },
 228                 { 0x14, 0x3f, 3, 3, "MI_DISPLAY_BUFFER_INFO" },
 229                 { 0x04, 0, 1, 1, "MI_FLUSH" },
 230                 { 0x22, 0x1f, 3, 3, "MI_LOAD_REGISTER_IMM" },
 231                 { 0x13, 0x3f, 2, 2, "MI_LOAD_SCAN_LINES_EXCL" },
 232                 { 0x12, 0x3f, 2, 2, "MI_LOAD_SCAN_LINES_INCL" },
 233                 { 0x00, 0, 1, 1, "MI_NOOP" },
 234                 { 0x11, 0x3f, 2, 2, "MI_OVERLAY_FLIP" },
 235                 { 0x07, 0, 1, 1, "MI_REPORT_HEAD" },
 236                 { 0x18, 0x3f, 2, 2, "MI_SET_CONTEXT" },
 237                 { 0x20, 0x3f, 3, 4, "MI_STORE_DATA_IMM" },
 238                 { 0x21, 0x3f, 3, 4, "MI_STORE_DATA_INDEX" },
 239                 { 0x24, 0x3f, 3, 3, "MI_STORE_REGISTER_MEM" },
 240                 { 0x02, 0, 1, 1, "MI_USER_INTERRUPT" },
 241                 { 0x03, 0, 1, 1, "MI_WAIT_FOR_EVENT", decode_MI_WAIT_FOR_EVENT },
 242                 { 0x16, 0x7f, 3, 3, "MI_SEMAPHORE_MBOX" },
 243                 { 0x26, 0x1f, 3, 4, "MI_FLUSH_DW" },
 244                 { 0x0b, 0, 1, 1, "MI_SUSPEND_FLUSH"},
 245         }, *opcode_mi = NULL;
 246
 247         /* check instruction length */
 248         for (opcode = 0; opcode < sizeof(opcodes_mi) / sizeof(opcodes_mi[0]);
 249              opcode++) {
 250                 if ((data[0] & 0x1f800000) >> 23 == opcodes_mi[opcode].opcode) {
 251                         len = 1;
 252                         if (opcodes_mi[opcode].max_len > 1) {
 253                                 len =
 254                                     (data[0] & opcodes_mi[opcode].len_mask) + 2;
 255                                 if (len < opcodes_mi[opcode].min_len
 256                                     || len > opcodes_mi[opcode].max_len) {
 257                                         fprintf(out,
 258                                                 "Bad length (%d) in %s, [%d, %d]\n",
 259                                                 len, opcodes_mi[opcode].name,
 260                                                 opcodes_mi[opcode].min_len,
 261                                                 opcodes_mi[opcode].max_len);
 262                                 }
 263                         }
 264                         opcode_mi = &opcodes_mi[opcode];
 265                         break;
 266                 }
 267         }
 268
 269         if (opcode_mi && opcode_mi->func)
 270                 return opcode_mi->func(ctx);
 271
 272         switch ((data[0] & 0x1f800000) >> 23) {
 273         case 0x0a:
 274                 instr_out(ctx, 0, "MI_BATCH_BUFFER_END\n");
 275                 return -1;
 276         case 0x16:
 277                 instr_out(ctx, 0, "MI_SEMAPHORE_MBOX%s%s%s%s %u\n",
 278                           data[0] & (1 << 22) ? " global gtt," : "",
 279                           data[0] & (1 << 21) ? " update semaphore," : "",
 280                           data[0] & (1 << 20) ? " compare semaphore," : "",
 281                           data[0] & (1 << 18) ? " use compare reg" : "",
 282                           (data[0] & (0x3 << 16)) >> 16);
 283                 instr_out(ctx, 1, "value\n");
 284                 instr_out(ctx, 2, "address\n");
 285                 return len;
 286         case 0x21:
 287                 instr_out(ctx, 0, "MI_STORE_DATA_INDEX%s\n",
 288                           data[0] & (1 << 21) ? " use per-process HWS," : "");
 289                 instr_out(ctx, 1, "index\n");
 290                 instr_out(ctx, 2, "dword\n");
 291                 if (len == 4)
 292                         instr_out(ctx, 3, "upper dword\n");
 293                 return len;
 294         case 0x00:
 295                 if (data[0] & (1 << 22))
 296                         instr_out(ctx, 0,
 297                                   "MI_NOOP write NOPID reg, val=0x%x\n",
 298                                   data[0] & ((1 << 22) - 1));
 299                 else
 300                         instr_out(ctx, 0, "MI_NOOP\n");
 301                 return len;
 302         case 0x26:
 303                 switch (data[0] & (0x3 << 14)) {
 304                 case (0 << 14):
 305                         post_sync_op = "no write";
 306                         break;
 307                 case (1 << 14):
 308                         post_sync_op = "write data";
 309                         break;
 310                 case (2 << 14):
 311                         post_sync_op = "reserved";
 312                         break;
 313                 case (3 << 14):
 314                         post_sync_op = "write TIMESTAMP";
 315                         break;
 316                 }
 317                 instr_out(ctx, 0,
 318                           "MI_FLUSH_DW%s%s%s%s post_sync_op='%s' %s%s\n",
 319                           data[0] & (1 << 22) ?
 320                           " enable protected mem (BCS-only)," : "",
 321                           data[0] & (1 << 21) ? " store in hws," : "",
 322                           data[0] & (1 << 18) ? " invalidate tlb," : "",
 323                           data[0] & (1 << 17) ? " flush gfdt," : "",
 324                           post_sync_op,
 325                           data[0] & (1 << 8) ? " enable notify interrupt," : "",
 326                           data[0] & (1 << 7) ?
 327                           " invalidate video state (BCS-only)," : "");
 328                 if (data[0] & (1 << 21))
 329                         instr_out(ctx, 1, "hws index\n");
 330                 else
 331                         instr_out(ctx, 1, "address\n");
 332                 instr_out(ctx, 2, "dword\n");
 333                 if (len == 4)
 334                         instr_out(ctx, 3, "upper dword\n");
 335                 return len;
 336         }
 337
 338         for (opcode = 0; opcode < sizeof(opcodes_mi) / sizeof(opcodes_mi[0]);
 339              opcode++) {
 340                 if ((data[0] & 0x1f800000) >> 23 == opcodes_mi[opcode].opcode) {
 341                         unsigned int i;
 342
 343                         instr_out(ctx, 0, "%s\n",
 344                                   opcodes_mi[opcode].name);
 345                         for (i = 1; i < len; i++) {
 346                                 instr_out(ctx, i, "dword %d\n", i);
 347                         }
 348
 349                         return len;
 350                 }
 351         }
 352
 353         instr_out(ctx, 0, "MI UNKNOWN\n");
 354         return 1;
 355 }
 356
 357 static void
 358 decode_2d_br00(struct drm_intel_decode *ctx, const char *cmd)
 359 {
 360         instr_out(ctx, 0,
 361                   "%s (rgb %sabled, alpha %sabled, src tile %d, dst tile %d)\n",
 362                   cmd,
 363                   (ctx->data[0] & (1 << 20)) ? "en" : "dis",
 364                   (ctx->data[0] & (1 << 21)) ? "en" : "dis",
 365                   (ctx->data[0] >> 15) & 1,
 366                   (ctx->data[0] >> 11) & 1);
 367 }
 368
 369 static void
 370 decode_2d_br01(struct drm_intel_decode *ctx)
 371 {
 372         const char *format;
 373         switch ((ctx->data[1] >> 24) & 0x3) {
 374         case 0:
 375                 format = "8";
 376                 break;
 377         case 1:
 378                 format = "565";
 379                 break;
 380         case 2:
 381                 format = "1555";
 382                 break;
 383         case 3:
 384                 format = "8888";
 385                 break;
 386         }
 387
 388         instr_out(ctx, 1,
 389                   "format %s, pitch %d, rop 0x%02x, "
 390                   "clipping %sabled, %s%s \n",
 391                   format,
 392                   (short)(ctx->data[1] & 0xffff),
 393                   (ctx->data[1] >> 16) & 0xff,
 394                   ctx->data[1] & (1 << 30) ? "en" : "dis",
 395                   ctx->data[1] & (1 << 31) ? "solid pattern enabled, " : "",
 396                   ctx->data[1] & (1 << 31) ?
 397                   "mono pattern transparency enabled, " : "");
 398
 399 }
 400
 401 static int
 402 decode_2d(struct drm_intel_decode *ctx)
 403 {
 404         unsigned int opcode, len;
 405         uint32_t *data = ctx->data;
 406
 407         struct {
 408                 uint32_t opcode;
 409                 unsigned int min_len;
 410                 unsigned int max_len;
 411                 const char *name;
 412         } opcodes_2d[] = {
 413                 { 0x40, 5, 5, "COLOR_BLT" },
 414                 { 0x43, 6, 6, "SRC_COPY_BLT" },
 415                 { 0x01, 8, 8, "XY_SETUP_BLT" },
 416                 { 0x11, 9, 9, "XY_SETUP_MONO_PATTERN_SL_BLT" },
 417                 { 0x03, 3, 3, "XY_SETUP_CLIP_BLT" },
 418                 { 0x24, 2, 2, "XY_PIXEL_BLT" },
 419                 { 0x25, 3, 3, "XY_SCANLINES_BLT" },
 420                 { 0x26, 4, 4, "Y_TEXT_BLT" },
 421                 { 0x31, 5, 134, "XY_TEXT_IMMEDIATE_BLT" },
 422                 { 0x50, 6, 6, "XY_COLOR_BLT" },
 423                 { 0x51, 6, 6, "XY_PAT_BLT" },
 424                 { 0x76, 8, 8, "XY_PAT_CHROMA_BLT" },
 425                 { 0x72, 7, 135, "XY_PAT_BLT_IMMEDIATE" },
 426                 { 0x77, 9, 137, "XY_PAT_CHROMA_BLT_IMMEDIATE" },
 427                 { 0x52, 9, 9, "XY_MONO_PAT_BLT" },
 428                 { 0x59, 7, 7, "XY_MONO_PAT_FIXED_BLT" },
 429                 { 0x53, 8, 8, "XY_SRC_COPY_BLT" },
 430                 { 0x54, 8, 8, "XY_MONO_SRC_COPY_BLT" },
 431                 { 0x71, 9, 137, "XY_MONO_SRC_COPY_IMMEDIATE_BLT" },
 432                 { 0x55, 9, 9, "XY_FULL_BLT" },
 433                 { 0x55, 9, 137, "XY_FULL_IMMEDIATE_PATTERN_BLT" },
 434                 { 0x56, 9, 9, "XY_FULL_MONO_SRC_BLT" },
 435                 { 0x75, 10, 138, "XY_FULL_MONO_SRC_IMMEDIATE_PATTERN_BLT" },
 436                 { 0x57, 12, 12, "XY_FULL_MONO_PATTERN_BLT" },
 437                 { 0x58, 12, 12, "XY_FULL_MONO_PATTERN_MONO_SRC_BLT"},
 438         };
 439
 440         switch ((data[0] & 0x1fc00000) >> 22) {
 441         case 0x25:
 442                 instr_out(ctx, 0,
 443                           "XY_SCANLINES_BLT (pattern seed (%d, %d), dst tile %d)\n",
 444                           (data[0] >> 12) & 0x8,
 445                           (data[0] >> 8) & 0x8, (data[0] >> 11) & 1);
 446
 447                 len = (data[0] & 0x000000ff) + 2;
 448                 if (len != 3)
 449                         fprintf(out, "Bad count in XY_SCANLINES_BLT\n");
 450
 451                 instr_out(ctx, 1, "dest (%d,%d)\n",
 452                           data[1] & 0xffff, data[1] >> 16);
 453                 instr_out(ctx, 2, "dest (%d,%d)\n",
 454                           data[2] & 0xffff, data[2] >> 16);
 455                 return len;
 456         case 0x01:
 457                 decode_2d_br00(ctx, "XY_SETUP_BLT");
 458
 459                 len = (data[0] & 0x000000ff) + 2;
 460                 if (len != 8)
 461                         fprintf(out, "Bad count in XY_SETUP_BLT\n");
 462
 463                 decode_2d_br01(ctx);
 464                 instr_out(ctx, 2, "cliprect (%d,%d)\n",
 465                           data[2] & 0xffff, data[2] >> 16);
 466                 instr_out(ctx, 3, "cliprect (%d,%d)\n",
 467                           data[3] & 0xffff, data[3] >> 16);
 468                 instr_out(ctx, 4, "setup dst offset 0x%08x\n",
 469                           data[4]);
 470                 instr_out(ctx, 5, "setup background color\n");
 471                 instr_out(ctx, 6, "setup foreground color\n");
 472                 instr_out(ctx, 7, "color pattern offset\n");
 473                 return len;
 474         case 0x03:
 475                 decode_2d_br00(ctx, "XY_SETUP_CLIP_BLT");
 476
 477                 len = (data[0] & 0x000000ff) + 2;
 478                 if (len != 3)
 479                         fprintf(out, "Bad count in XY_SETUP_CLIP_BLT\n");
 480
 481                 instr_out(ctx, 1, "cliprect (%d,%d)\n",
 482                           data[1] & 0xffff, data[2] >> 16);
 483                 instr_out(ctx, 2, "cliprect (%d,%d)\n",
 484                           data[2] & 0xffff, data[3] >> 16);
 485                 return len;
 486         case 0x11:
 487                 decode_2d_br00(ctx, "XY_SETUP_MONO_PATTERN_SL_BLT");
 488
 489                 len = (data[0] & 0x000000ff) + 2;
 490                 if (len != 9)
 491                         fprintf(out,
 492                                 "Bad count in XY_SETUP_MONO_PATTERN_SL_BLT\n");
 493
 494                 decode_2d_br01(ctx);
 495                 instr_out(ctx, 2, "cliprect (%d,%d)\n",
 496                           data[2] & 0xffff, data[2] >> 16);
 497                 instr_out(ctx, 3, "cliprect (%d,%d)\n",
 498                           data[3] & 0xffff, data[3] >> 16);
 499                 instr_out(ctx, 4, "setup dst offset 0x%08x\n",
 500                           data[4]);
 501                 instr_out(ctx, 5, "setup background color\n");
 502                 instr_out(ctx, 6, "setup foreground color\n");
 503                 instr_out(ctx, 7, "mono pattern dw0\n");
 504                 instr_out(ctx, 8, "mono pattern dw1\n");
 505                 return len;
 506         case 0x50:
 507                 decode_2d_br00(ctx, "XY_COLOR_BLT");
 508
 509                 len = (data[0] & 0x000000ff) + 2;
 510                 if (len != 6)
 511                         fprintf(out, "Bad count in XY_COLOR_BLT\n");
 512
 513                 decode_2d_br01(ctx);
 514                 instr_out(ctx, 2, "(%d,%d)\n",
 515                           data[2] & 0xffff, data[2] >> 16);
 516                 instr_out(ctx, 3, "(%d,%d)\n",
 517                           data[3] & 0xffff, data[3] >> 16);
 518                 instr_out(ctx, 4, "offset 0x%08x\n", data[4]);
 519                 instr_out(ctx, 5, "color\n");
 520                 return len;
 521         case 0x53:
 522                 decode_2d_br00(ctx, "XY_SRC_COPY_BLT");
 523
 524                 len = (data[0] & 0x000000ff) + 2;
 525                 if (len != 8)
 526                         fprintf(out, "Bad count in XY_SRC_COPY_BLT\n");
 527
 528                 decode_2d_br01(ctx);
 529                 instr_out(ctx, 2, "dst (%d,%d)\n",
 530                           data[2] & 0xffff, data[2] >> 16);
 531                 instr_out(ctx, 3, "dst (%d,%d)\n",
 532                           data[3] & 0xffff, data[3] >> 16);
 533                 instr_out(ctx, 4, "dst offset 0x%08x\n", data[4]);
 534                 instr_out(ctx, 5, "src (%d,%d)\n",
 535                           data[5] & 0xffff, data[5] >> 16);
 536                 instr_out(ctx, 6, "src pitch %d\n",
 537                           (short)(data[6] & 0xffff));
 538                 instr_out(ctx, 7, "src offset 0x%08x\n", data[7]);
 539                 return len;
 540         }
 541
 542         for (opcode = 0; opcode < sizeof(opcodes_2d) / sizeof(opcodes_2d[0]);
 543              opcode++) {
 544                 if ((data[0] & 0x1fc00000) >> 22 == opcodes_2d[opcode].opcode) {
 545                         unsigned int i;
 546
 547                         len = 1;
 548                         instr_out(ctx, 0, "%s\n",
 549                                   opcodes_2d[opcode].name);
 550                         if (opcodes_2d[opcode].max_len > 1) {
 551                                 len = (data[0] & 0x000000ff) + 2;
 552                                 if (len < opcodes_2d[opcode].min_len ||
 553                                     len > opcodes_2d[opcode].max_len) {
 554                                         fprintf(out, "Bad count in %s\n",
 555                                                 opcodes_2d[opcode].name);
 556                                 }
 557                         }
 558
 559                         for (i = 1; i < len; i++) {
 560                                 instr_out(ctx, i, "dword %d\n", i);
 561                         }
 562
 563                         return len;
 564                 }
 565         }
 566
 567         instr_out(ctx, 0, "2D UNKNOWN\n");
 568         return 1;
 569 }
 570
 571 static int
 572 decode_3d_1c(struct drm_intel_decode *ctx)
 573 {
 574         uint32_t *data = ctx->data;
 575         uint32_t opcode;
 576
 577         opcode = (data[0] & 0x00f80000) >> 19;
 578
 579         switch (opcode) {
 580         case 0x11:
 581                 instr_out(ctx, 0,
 582                           "3DSTATE_DEPTH_SUBRECTANGLE_DISABLE\n");
 583                 return 1;
 584         case 0x10:
 585                 instr_out(ctx, 0, "3DSTATE_SCISSOR_ENABLE %s\n",
 586                           data[0] & 1 ? "enabled" : "disabled");
 587                 return 1;
 588         case 0x01:
 589                 instr_out(ctx, 0, "3DSTATE_MAP_COORD_SET_I830\n");
 590                 return 1;
 591         case 0x0a:
 592                 instr_out(ctx, 0, "3DSTATE_MAP_CUBE_I830\n");
 593                 return 1;
 594         case 0x05:
 595                 instr_out(ctx, 0, "3DSTATE_MAP_TEX_STREAM_I830\n");
 596                 return 1;
 597         }
 598
 599         instr_out(ctx, 0, "3D UNKNOWN: 3d_1c opcode = 0x%x\n",
 600                   opcode);
 601         return 1;
 602 }
 603
 604 /** Sets the string dstname to describe the destination of the PS instruction */
 605 static void
 606 i915_get_instruction_dst(uint32_t *data, int i, char *dstname, int do_mask)
 607 {
 608         uint32_t a0 = data[i];
 609         int dst_nr = (a0 >> 14) & 0xf;
 610         char dstmask[8];
 611         const char *sat;
 612
 613         if (do_mask) {
 614                 if (((a0 >> 10) & 0xf) == 0xf) {
 615                         dstmask[0] = 0;
 616                 } else {
 617                         int dstmask_index = 0;
 618
 619                         dstmask[dstmask_index++] = '.';
 620                         if (a0 & (1 << 10))
 621                                 dstmask[dstmask_index++] = 'x';
 622                         if (a0 & (1 << 11))
 623                                 dstmask[dstmask_index++] = 'y';
 624                         if (a0 & (1 << 12))
 625                                 dstmask[dstmask_index++] = 'z';
 626                         if (a0 & (1 << 13))
 627                                 dstmask[dstmask_index++] = 'w';
 628                         dstmask[dstmask_index++] = 0;
 629                 }
 630
 631                 if (a0 & (1 << 22))
 632                         sat = ".sat";
 633                 else
 634                         sat = "";
 635         } else {
 636                 dstmask[0] = 0;
 637                 sat = "";
 638         }
 639
 640         switch ((a0 >> 19) & 0x7) {
 641         case 0:
 642                 if (dst_nr > 15)
 643                         fprintf(out, "bad destination reg R%d\n", dst_nr);
 644                 sprintf(dstname, "R%d%s%s", dst_nr, dstmask, sat);
 645                 break;
 646         case 4:
 647                 if (dst_nr > 0)
 648                         fprintf(out, "bad destination reg oC%d\n", dst_nr);
 649                 sprintf(dstname, "oC%s%s", dstmask, sat);
 650                 break;
 651         case 5:
 652                 if (dst_nr > 0)
 653                         fprintf(out, "bad destination reg oD%d\n", dst_nr);
 654                 sprintf(dstname, "oD%s%s", dstmask, sat);
 655                 break;
 656         case 6:
 657                 if (dst_nr > 3)
 658                         fprintf(out, "bad destination reg U%d\n", dst_nr);
 659                 sprintf(dstname, "U%d%s%s", dst_nr, dstmask, sat);
 660                 break;
 661         default:
 662                 sprintf(dstname, "RESERVED");
 663                 break;
 664         }
 665 }
 666
 667 static const char *
 668 i915_get_channel_swizzle(uint32_t select)
 669 {
 670         switch (select & 0x7) {
 671         case 0:
 672                 return (select & 8) ? "-x" : "x";
 673         case 1:
 674                 return (select & 8) ? "-y" : "y";
 675         case 2:
 676                 return (select & 8) ? "-z" : "z";
 677         case 3:
 678                 return (select & 8) ? "-w" : "w";
 679         case 4:
 680                 return (select & 8) ? "-0" : "0";
 681         case 5:
 682                 return (select & 8) ? "-1" : "1";
 683         default:
 684                 return (select & 8) ? "-bad" : "bad";
 685         }
 686 }
 687
 688 static void
 689 i915_get_instruction_src_name(uint32_t src_type, uint32_t src_nr, char *name)
 690 {
 691         switch (src_type) {
 692         case 0:
 693                 sprintf(name, "R%d", src_nr);
 694                 if (src_nr > 15)
 695                         fprintf(out, "bad src reg %s\n", name);
 696                 break;
 697         case 1:
 698                 if (src_nr < 8)
 699                         sprintf(name, "T%d", src_nr);
 700                 else if (src_nr == 8)
 701                         sprintf(name, "DIFFUSE");
 702                 else if (src_nr == 9)
 703                         sprintf(name, "SPECULAR");
 704                 else if (src_nr == 10)
 705                         sprintf(name, "FOG");
 706                 else {
 707                         fprintf(out, "bad src reg T%d\n", src_nr);
 708                         sprintf(name, "RESERVED");
 709                 }
 710                 break;
 711         case 2:
 712                 sprintf(name, "C%d", src_nr);
 713                 if (src_nr > 31)
 714                         fprintf(out, "bad src reg %s\n", name);
 715                 break;
 716         case 4:
 717                 sprintf(name, "oC");
 718                 if (src_nr > 0)
 719                         fprintf(out, "bad src reg oC%d\n", src_nr);
 720                 break;
 721         case 5:
 722                 sprintf(name, "oD");
 723                 if (src_nr > 0)
 724                         fprintf(out, "bad src reg oD%d\n", src_nr);
 725                 break;
 726         case 6:
 727                 sprintf(name, "U%d", src_nr);
 728                 if (src_nr > 3)
 729                         fprintf(out, "bad src reg %s\n", name);
 730                 break;
 731         default:
 732                 fprintf(out, "bad src reg type %d\n", src_type);
 733                 sprintf(name, "RESERVED");
 734                 break;
 735         }
 736 }
 737
 738 static void i915_get_instruction_src0(uint32_t *data, int i, char *srcname)
 739 {
 740         uint32_t a0 = data[i];
 741         uint32_t a1 = data[i + 1];
 742         int src_nr = (a0 >> 2) & 0x1f;
 743         const char *swizzle_x = i915_get_channel_swizzle((a1 >> 28) & 0xf);
 744         const char *swizzle_y = i915_get_channel_swizzle((a1 >> 24) & 0xf);
 745         const char *swizzle_z = i915_get_channel_swizzle((a1 >> 20) & 0xf);
 746         const char *swizzle_w = i915_get_channel_swizzle((a1 >> 16) & 0xf);
 747         char swizzle[100];
 748
 749         i915_get_instruction_src_name((a0 >> 7) & 0x7, src_nr, srcname);
 750         sprintf(swizzle, ".%s%s%s%s", swizzle_x, swizzle_y, swizzle_z,
 751                 swizzle_w);
 752         if (strcmp(swizzle, ".xyzw") != 0)
 753                 strcat(srcname, swizzle);
 754 }
 755
 756 static void i915_get_instruction_src1(uint32_t *data, int i, char *srcname)
 757 {
 758         uint32_t a1 = data[i + 1];
 759         uint32_t a2 = data[i + 2];
 760         int src_nr = (a1 >> 8) & 0x1f;
 761         const char *swizzle_x = i915_get_channel_swizzle((a1 >> 4) & 0xf);
 762         const char *swizzle_y = i915_get_channel_swizzle((a1 >> 0) & 0xf);
 763         const char *swizzle_z = i915_get_channel_swizzle((a2 >> 28) & 0xf);
 764         const char *swizzle_w = i915_get_channel_swizzle((a2 >> 24) & 0xf);
 765         char swizzle[100];
 766
 767         i915_get_instruction_src_name((a1 >> 13) & 0x7, src_nr, srcname);
 768         sprintf(swizzle, ".%s%s%s%s", swizzle_x, swizzle_y, swizzle_z,
 769                 swizzle_w);
 770         if (strcmp(swizzle, ".xyzw") != 0)
 771                 strcat(srcname, swizzle);
 772 }
 773
 774 static void i915_get_instruction_src2(uint32_t *data, int i, char *srcname)
 775 {
 776         uint32_t a2 = data[i + 2];
 777         int src_nr = (a2 >> 16) & 0x1f;
 778         const char *swizzle_x = i915_get_channel_swizzle((a2 >> 12) & 0xf);
 779         const char *swizzle_y = i915_get_channel_swizzle((a2 >> 8) & 0xf);
 780         const char *swizzle_z = i915_get_channel_swizzle((a2 >> 4) & 0xf);
 781         const char *swizzle_w = i915_get_channel_swizzle((a2 >> 0) & 0xf);
 782         char swizzle[100];
 783
 784         i915_get_instruction_src_name((a2 >> 21) & 0x7, src_nr, srcname);
 785         sprintf(swizzle, ".%s%s%s%s", swizzle_x, swizzle_y, swizzle_z,
 786                 swizzle_w);
 787         if (strcmp(swizzle, ".xyzw") != 0)
 788                 strcat(srcname, swizzle);
 789 }
 790
 791 static void
 792 i915_get_instruction_addr(uint32_t src_type, uint32_t src_nr, char *name)
 793 {
 794         switch (src_type) {
 795         case 0:
 796                 sprintf(name, "R%d", src_nr);
 797                 if (src_nr > 15)
 798                         fprintf(out, "bad src reg %s\n", name);
 799                 break;
 800         case 1:
 801                 if (src_nr < 8)
 802                         sprintf(name, "T%d", src_nr);
 803                 else if (src_nr == 8)
 804                         sprintf(name, "DIFFUSE");
 805                 else if (src_nr == 9)
 806                         sprintf(name, "SPECULAR");
 807                 else if (src_nr == 10)
 808                         sprintf(name, "FOG");
 809                 else {
 810                         fprintf(out, "bad src reg T%d\n", src_nr);
 811                         sprintf(name, "RESERVED");
 812                 }
 813                 break;
 814         case 4:
 815                 sprintf(name, "oC");
 816                 if (src_nr > 0)
 817                         fprintf(out, "bad src reg oC%d\n", src_nr);
 818                 break;
 819         case 5:
 820                 sprintf(name, "oD");
 821                 if (src_nr > 0)
 822                         fprintf(out, "bad src reg oD%d\n", src_nr);
 823                 break;
 824         default:
 825                 fprintf(out, "bad src reg type %d\n", src_type);
 826                 sprintf(name, "RESERVED");
 827                 break;
 828         }
 829 }
 830
 831 static void
 832 i915_decode_alu1(struct drm_intel_decode *ctx,
 833                  int i, char *instr_prefix, const char *op_name)
 834 {
 835         char dst[100], src0[100];
 836
 837         i915_get_instruction_dst(ctx->data, i, dst, 1);
 838         i915_get_instruction_src0(ctx->data, i, src0);
 839
 840         instr_out(ctx, i++, "%s: %s %s, %s\n", instr_prefix,
 841                   op_name, dst, src0);
 842         instr_out(ctx, i++, "%s\n", instr_prefix);
 843         instr_out(ctx, i++, "%s\n", instr_prefix);
 844 }
 845
 846 static void
 847 i915_decode_alu2(struct drm_intel_decode *ctx,
 848                  int i, char *instr_prefix, const char *op_name)
 849 {
 850         char dst[100], src0[100], src1[100];
 851
 852         i915_get_instruction_dst(ctx->data, i, dst, 1);
 853         i915_get_instruction_src0(ctx->data, i, src0);
 854         i915_get_instruction_src1(ctx->data, i, src1);
 855
 856         instr_out(ctx, i++, "%s: %s %s, %s, %s\n", instr_prefix,
 857                   op_name, dst, src0, src1);
 858         instr_out(ctx, i++, "%s\n", instr_prefix);
 859         instr_out(ctx, i++, "%s\n", instr_prefix);
 860 }
 861
 862 static void
 863 i915_decode_alu3(struct drm_intel_decode *ctx,
 864                  int i, char *instr_prefix, const char *op_name)
 865 {
 866         char dst[100], src0[100], src1[100], src2[100];
 867
 868         i915_get_instruction_dst(ctx->data, i, dst, 1);
 869         i915_get_instruction_src0(ctx->data, i, src0);
 870         i915_get_instruction_src1(ctx->data, i, src1);
 871         i915_get_instruction_src2(ctx->data, i, src2);
 872
 873         instr_out(ctx, i++, "%s: %s %s, %s, %s, %s\n", instr_prefix,
 874                   op_name, dst, src0, src1, src2);
 875         instr_out(ctx, i++, "%s\n", instr_prefix);
 876         instr_out(ctx, i++, "%s\n", instr_prefix);
 877 }
 878
 879 static void
 880 i915_decode_tex(struct drm_intel_decode *ctx, int i,
 881                 const char *instr_prefix, const char *tex_name)
 882 {
 883         uint32_t t0 = ctx->data[i];
 884         uint32_t t1 = ctx->data[i + 1];
 885         char dst_name[100];
 886         char addr_name[100];
 887         int sampler_nr;
 888
 889         i915_get_instruction_dst(ctx->data, i, dst_name, 0);
 890         i915_get_instruction_addr((t1 >> 24) & 0x7,
 891                                   (t1 >> 17) & 0xf, addr_name);
 892         sampler_nr = t0 & 0xf;
 893
 894         instr_out(ctx, i++, "%s: %s %s, S%d, %s\n", instr_prefix,
 895                   tex_name, dst_name, sampler_nr, addr_name);
 896         instr_out(ctx, i++, "%s\n", instr_prefix);
 897         instr_out(ctx, i++, "%s\n", instr_prefix);
 898 }
 899
 900 static void
 901 i915_decode_dcl(struct drm_intel_decode *ctx, int i, char *instr_prefix)
 902 {
 903         uint32_t d0 = ctx->data[i];
 904         const char *sampletype;
 905         int dcl_nr = (d0 >> 14) & 0xf;
 906         const char *dcl_x = d0 & (1 << 10) ? "x" : "";
 907         const char *dcl_y = d0 & (1 << 11) ? "y" : "";
 908         const char *dcl_z = d0 & (1 << 12) ? "z" : "";
 909         const char *dcl_w = d0 & (1 << 13) ? "w" : "";
 910         char dcl_mask[10];
 911
 912         switch ((d0 >> 19) & 0x3) {
 913         case 1:
 914                 sprintf(dcl_mask, ".%s%s%s%s", dcl_x, dcl_y, dcl_z, dcl_w);
 915                 if (strcmp(dcl_mask, ".") == 0)
 916                         fprintf(out, "bad (empty) dcl mask\n");
 917
 918                 if (dcl_nr > 10)
 919                         fprintf(out, "bad T%d dcl register number\n", dcl_nr);
 920                 if (dcl_nr < 8) {
 921                         if (strcmp(dcl_mask, ".x") != 0 &&
 922                             strcmp(dcl_mask, ".xy") != 0 &&
 923                             strcmp(dcl_mask, ".xz") != 0 &&
 924                             strcmp(dcl_mask, ".w") != 0 &&
 925                             strcmp(dcl_mask, ".xyzw") != 0) {
 926                                 fprintf(out, "bad T%d.%s dcl mask\n", dcl_nr,
 927                                         dcl_mask);
 928                         }
 929                         instr_out(ctx, i++, "%s: DCL T%d%s\n",
 930                                   instr_prefix, dcl_nr, dcl_mask);
 931                 } else {
 932                         if (strcmp(dcl_mask, ".xz") == 0)
 933                                 fprintf(out, "errataed bad dcl mask %s\n",
 934                                         dcl_mask);
 935                         else if (strcmp(dcl_mask, ".xw") == 0)
 936                                 fprintf(out, "errataed bad dcl mask %s\n",
 937                                         dcl_mask);
 938                         else if (strcmp(dcl_mask, ".xzw") == 0)
 939                                 fprintf(out, "errataed bad dcl mask %s\n",
 940                                         dcl_mask);
 941
 942                         if (dcl_nr == 8) {
 943                                 instr_out(ctx, i++,
 944                                           "%s: DCL DIFFUSE%s\n", instr_prefix,
 945                                           dcl_mask);
 946                         } else if (dcl_nr == 9) {
 947                                 instr_out(ctx, i++,
 948                                           "%s: DCL SPECULAR%s\n", instr_prefix,
 949                                           dcl_mask);
 950                         } else if (dcl_nr == 10) {
 951                                 instr_out(ctx, i++,
 952                                           "%s: DCL FOG%s\n", instr_prefix,
 953                                           dcl_mask);
 954                         }
 955                 }
 956                 instr_out(ctx, i++, "%s\n", instr_prefix);
 957                 instr_out(ctx, i++, "%s\n", instr_prefix);
 958                 break;
 959         case 3:
 960                 switch ((d0 >> 22) & 0x3) {
 961                 case 0:
 962                         sampletype = "2D";
 963                         break;
 964                 case 1:
 965                         sampletype = "CUBE";
 966                         break;
 967                 case 2:
 968                         sampletype = "3D";
 969                         break;
 970                 default:
 971                         sampletype = "RESERVED";
 972                         break;
 973                 }
 974                 if (dcl_nr > 15)
 975                         fprintf(out, "bad S%d dcl register number\n", dcl_nr);
 976                 instr_out(ctx, i++, "%s: DCL S%d %s\n",
 977                           instr_prefix, dcl_nr, sampletype);
 978                 instr_out(ctx, i++, "%s\n", instr_prefix);
 979                 instr_out(ctx, i++, "%s\n", instr_prefix);
 980                 break;
 981         default:
 982                 instr_out(ctx, i++, "%s: DCL RESERVED%d\n",
 983                           instr_prefix, dcl_nr);
 984                 instr_out(ctx, i++, "%s\n", instr_prefix);
 985                 instr_out(ctx, i++, "%s\n", instr_prefix);
 986         }
 987 }
 988
 989 static void
 990 i915_decode_instruction(struct drm_intel_decode *ctx,
 991                         int i, char *instr_prefix)
 992 {
 993         switch ((ctx->data[i] >> 24) & 0x1f) {
 994         case 0x0:
 995                 instr_out(ctx, i++, "%s: NOP\n", instr_prefix);
 996                 instr_out(ctx, i++, "%s\n", instr_prefix);
 997                 instr_out(ctx, i++, "%s\n", instr_prefix);
 998                 break;
 999         case 0x01:
1000                 i915_decode_alu2(ctx, i, instr_prefix, "ADD");
1001                 break;
1002         case 0x02:
1003                 i915_decode_alu1(ctx, i, instr_prefix, "MOV");
1004                 break;
1005         case 0x03:
1006                 i915_decode_alu2(ctx, i, instr_prefix, "MUL");
1007                 break;
1008         case 0x04:
1009                 i915_decode_alu3(ctx, i, instr_prefix, "MAD");
1010                 break;
1011         case 0x05:
1012                 i915_decode_alu3(ctx, i, instr_prefix, "DP2ADD");
1013                 break;
1014         case 0x06:
1015                 i915_decode_alu2(ctx, i, instr_prefix, "DP3");
1016                 break;
1017         case 0x07:
1018                 i915_decode_alu2(ctx, i, instr_prefix, "DP4");
1019                 break;
1020         case 0x08:
1021                 i915_decode_alu1(ctx, i, instr_prefix, "FRC");
1022                 break;
1023         case 0x09:
1024                 i915_decode_alu1(ctx, i, instr_prefix, "RCP");
1025                 break;
1026         case 0x0a:
1027                 i915_decode_alu1(ctx, i, instr_prefix, "RSQ");
1028                 break;
1029         case 0x0b:
1030                 i915_decode_alu1(ctx, i, instr_prefix, "EXP");
1031                 break;
1032         case 0x0c:
1033                 i915_decode_alu1(ctx, i, instr_prefix, "LOG");
1034                 break;
1035         case 0x0d:
1036                 i915_decode_alu2(ctx, i, instr_prefix, "CMP");
1037                 break;
1038         case 0x0e:
1039                 i915_decode_alu2(ctx, i, instr_prefix, "MIN");
1040                 break;
1041         case 0x0f:
1042                 i915_decode_alu2(ctx, i, instr_prefix, "MAX");
1043                 break;
1044         case 0x10:
1045                 i915_decode_alu1(ctx, i, instr_prefix, "FLR");
1046                 break;
1047         case 0x11:
1048                 i915_decode_alu1(ctx, i, instr_prefix, "MOD");
1049                 break;
1050         case 0x12:
1051                 i915_decode_alu1(ctx, i, instr_prefix, "TRC");
1052                 break;
1053         case 0x13:
1054                 i915_decode_alu2(ctx, i, instr_prefix, "SGE");
1055                 break;
1056         case 0x14:
1057                 i915_decode_alu2(ctx, i, instr_prefix, "SLT");
1058                 break;
1059         case 0x15:
1060                 i915_decode_tex(ctx, i, instr_prefix, "TEXLD");
1061                 break;
1062         case 0x16:
1063                 i915_decode_tex(ctx, i, instr_prefix, "TEXLDP");
1064                 break;
1065         case 0x17:
1066                 i915_decode_tex(ctx, i, instr_prefix, "TEXLDB");
1067                 break;
1068         case 0x19:
1069                 i915_decode_dcl(ctx, i, instr_prefix);
1070                 break;
1071         default:
1072                 instr_out(ctx, i++, "%s: unknown\n", instr_prefix);
1073                 instr_out(ctx, i++, "%s\n", instr_prefix);
1074                 instr_out(ctx, i++, "%s\n", instr_prefix);
1075                 break;
1076         }
1077 }
1078
1079 static const char *
1080 decode_compare_func(uint32_t op)
1081 {
1082         switch (op & 0x7) {
1083         case 0:
1084                 return "always";
1085         case 1:
1086                 return "never";
1087         case 2:
1088                 return "less";
1089         case 3:
1090                 return "equal";
1091         case 4:
1092                 return "lequal";
1093         case 5:
1094                 return "greater";
1095         case 6:
1096                 return "notequal";
1097         case 7:
1098                 return "gequal";
1099         }
1100         return "";
1101 }
1102
1103 static const char *
1104 decode_stencil_op(uint32_t op)
1105 {
1106         switch (op & 0x7) {
1107         case 0:
1108                 return "keep";
1109         case 1:
1110                 return "zero";
1111         case 2:
1112                 return "replace";
1113         case 3:
1114                 return "incr_sat";
1115         case 4:
1116                 return "decr_sat";
1117         case 5:
1118                 return "greater";
1119         case 6:
1120                 return "incr";
1121         case 7:
1122                 return "decr";
1123         }
1124         return "";
1125 }
1126
1127 #if 0
1128 static const char *
1129 decode_logic_op(uint32_t op)
1130 {
1131         switch (op & 0xf) {
1132         case 0:
1133                 return "clear";
1134         case 1:
1135                 return "nor";
1136         case 2:
1137                 return "and_inv";
1138         case 3:
1139                 return "copy_inv";
1140         case 4:
1141                 return "and_rvrse";
1142         case 5:
1143                 return "inv";
1144         case 6:
1145                 return "xor";
1146         case 7:
1147                 return "nand";
1148         case 8:
1149                 return "and";
1150         case 9:
1151                 return "equiv";
1152         case 10:
1153                 return "noop";
1154         case 11:
1155                 return "or_inv";
1156         case 12:
1157                 return "copy";
1158         case 13:
1159                 return "or_rvrse";
1160         case 14:
1161                 return "or";
1162         case 15:
1163                 return "set";
1164         }
1165         return "";
1166 }
1167 #endif
1168
1169 static const char *
1170 decode_blend_fact(uint32_t op)
1171 {
1172         switch (op & 0xf) {
1173         case 1:
1174                 return "zero";
1175         case 2:
1176                 return "one";
1177         case 3:
1178                 return "src_colr";
1179         case 4:
1180                 return "inv_src_colr";
1181         case 5:
1182                 return "src_alpha";
1183         case 6:
1184                 return "inv_src_alpha";
1185         case 7:
1186                 return "dst_alpha";
1187         case 8:
1188                 return "inv_dst_alpha";
1189         case 9:
1190                 return "dst_colr";
1191         case 10:
1192                 return "inv_dst_colr";
1193         case 11:
1194                 return "src_alpha_sat";
1195         case 12:
1196                 return "cnst_colr";
1197         case 13:
1198                 return "inv_cnst_colr";
1199         case 14:
1200                 return "cnst_alpha";
1201         case 15:
1202                 return "inv_const_alpha";
1203         }
1204         return "";
1205 }
1206
1207 static const char *
1208 decode_tex_coord_mode(uint32_t mode)
1209 {
1210         switch (mode & 0x7) {
1211         case 0:
1212                 return "wrap";
1213         case 1:
1214                 return "mirror";
1215         case 2:
1216                 return "clamp_edge";
1217         case 3:
1218                 return "cube";
1219         case 4:
1220                 return "clamp_border";
1221         case 5:
1222                 return "mirror_once";
1223         }
1224         return "";
1225 }
1226
1227 static const char *
1228 decode_sample_filter(uint32_t mode)
1229 {
1230         switch (mode & 0x7) {
1231         case 0:
1232                 return "nearest";
1233         case 1:
1234                 return "linear";
1235         case 2:
1236                 return "anisotropic";
1237         case 3:
1238                 return "4x4_1";
1239         case 4:
1240                 return "4x4_2";
1241         case 5:
1242                 return "4x4_flat";
1243         case 6:
1244                 return "6x5_mono";
1245         }
1246         return "";
1247 }
1248
1249 static int
1250 decode_3d_1d(struct drm_intel_decode *ctx)
1251 {
1252         unsigned int len, i, c, idx, word, map, sampler, instr;
1253         const char *format, *zformat, *type;
1254         uint32_t opcode;
1255         uint32_t *data = ctx->data;
1256         uint32_t devid = ctx->devid;
1257
1258         struct {
1259                 uint32_t opcode;
1260                 int i830_only;
1261                 unsigned int min_len;
1262                 unsigned int max_len;
1263                 const char *name;
1264         } opcodes_3d_1d[] = {
1265                 { 0x86, 0, 4, 4, "3DSTATE_CHROMA_KEY" },
1266                 { 0x88, 0, 2, 2, "3DSTATE_CONSTANT_BLEND_COLOR" },
1267                 { 0x99, 0, 2, 2, "3DSTATE_DEFAULT_DIFFUSE" },
1268                 { 0x9a, 0, 2, 2, "3DSTATE_DEFAULT_SPECULAR" },
1269                 { 0x98, 0, 2, 2, "3DSTATE_DEFAULT_Z" },
1270                 { 0x97, 0, 2, 2, "3DSTATE_DEPTH_OFFSET_SCALE" },
1271                 { 0x9d, 0, 65, 65, "3DSTATE_FILTER_COEFFICIENTS_4X4" },
1272                 { 0x9e, 0, 4, 4, "3DSTATE_MONO_FILTER" },
1273                 { 0x89, 0, 4, 4, "3DSTATE_FOG_MODE" },
1274                 { 0x8f, 0, 2, 16, "3DSTATE_MAP_PALLETE_LOAD_32" },
1275                 { 0x83, 0, 2, 2, "3DSTATE_SPAN_STIPPLE" },
1276                 { 0x8c, 1, 2, 2, "3DSTATE_MAP_COORD_TRANSFORM_I830" },
1277                 { 0x8b, 1, 2, 2, "3DSTATE_MAP_VERTEX_TRANSFORM_I830" },
1278                 { 0x8d, 1, 3, 3, "3DSTATE_W_STATE_I830" },
1279                 { 0x01, 1, 2, 2, "3DSTATE_COLOR_FACTOR_I830" },
1280                 { 0x02, 1, 2, 2, "3DSTATE_MAP_COORD_SETBIND_I830"},
1281         }, *opcode_3d_1d;
1282
1283         opcode = (data[0] & 0x00ff0000) >> 16;
1284
1285         switch (opcode) {
1286         case 0x07:
1287                 /* This instruction is unusual.  A 0 length means just
1288                  * 1 DWORD instead of 2.  The 0 length is specified in
1289                  * one place to be unsupported, but stated to be
1290                  * required in another, and 0 length LOAD_INDIRECTs
1291                  * appear to cause no harm at least.
1292                  */
1293                 instr_out(ctx, 0, "3DSTATE_LOAD_INDIRECT\n");
1294                 len = (data[0] & 0x000000ff) + 1;
1295                 i = 1;
1296                 if (data[0] & (0x01 << 8)) {
1297                         instr_out(ctx, i++, "SIS.0\n");
1298                         instr_out(ctx, i++, "SIS.1\n");
1299                 }
1300                 if (data[0] & (0x02 << 8)) {
1301                         instr_out(ctx, i++, "DIS.0\n");
1302                 }
1303                 if (data[0] & (0x04 << 8)) {
1304                         instr_out(ctx, i++, "SSB.0\n");
1305                         instr_out(ctx, i++, "SSB.1\n");
1306                 }
1307                 if (data[0] & (0x08 << 8)) {
1308                         instr_out(ctx, i++, "MSB.0\n");
1309                         instr_out(ctx, i++, "MSB.1\n");
1310                 }
1311                 if (data[0] & (0x10 << 8)) {
1312                         instr_out(ctx, i++, "PSP.0\n");
1313                         instr_out(ctx, i++, "PSP.1\n");
1314                 }
1315                 if (data[0] & (0x20 << 8)) {
1316                         instr_out(ctx, i++, "PSC.0\n");
1317                         instr_out(ctx, i++, "PSC.1\n");
1318                 }
1319                 if (len != i) {
1320                         fprintf(out, "Bad count in 3DSTATE_LOAD_INDIRECT\n");
1321                         return len;
1322                 }
1323                 return len;
1324         case 0x04:
1325                 instr_out(ctx, 0,
1326                           "3DSTATE_LOAD_STATE_IMMEDIATE_1\n");
1327                 len = (data[0] & 0x0000000f) + 2;
1328                 i = 1;
1329                 for (word = 0; word <= 8; word++) {
1330                         if (data[0] & (1 << (4 + word))) {
1331                                 /* save vertex state for decode */
1332                                 if (!IS_GEN2(devid)) {
1333                                         int tex_num;
1334
1335                                         if (word == 2) {
1336                                                 saved_s2_set = 1;
1337                                                 saved_s2 = data[i];
1338                                         }
1339                                         if (word == 4) {
1340                                                 saved_s4_set = 1;
1341                                                 saved_s4 = data[i];
1342                                         }
1343
1344                                         switch (word) {
1345                                         case 0:
1346                                                 instr_out(ctx, i,
1347                                                           "S0: vbo offset: 0x%08x%s\n",
1348                                                           data[i] & (~1),
1349                                                           data[i] & 1 ?
1350                                                           ", auto cache invalidate disabled"
1351                                                           : "");
1352                                                 break;
1353                                         case 1:
1354                                                 instr_out(ctx, i,
1355                                                           "S1: vertex width: %i, vertex pitch: %i\n",
1356                                                           (data[i] >> 24) &
1357                                                           0x3f,
1358                                                           (data[i] >> 16) &
1359                                                           0x3f);
1360                                                 break;
1361                                         case 2:
1362                                                 instr_out(ctx, i,
1363                                                           "S2: texcoord formats: ");
1364                                                 for (tex_num = 0;
1365                                                      tex_num < 8; tex_num++) {
1366                                                         switch ((data[i] >>
1367                                                                  tex_num *
1368                                                                  4) & 0xf) {
1369                                                         case 0:
1370                                                                 fprintf(out,
1371                                                                         "%i=2D ",
1372                                                                         tex_num);
1373                                                                 break;
1374                                                         case 1:
1375                                                                 fprintf(out,
1376                                                                         "%i=3D ",
1377                                                                         tex_num);
1378                                                                 break;
1379                                                         case 2:
1380                                                                 fprintf(out,
1381                                                                         "%i=4D ",
1382                                                                         tex_num);
1383                                                                 break;
1384                                                         case 3:
1385                                                                 fprintf(out,
1386                                                                         "%i=1D ",
1387                                                                         tex_num);
1388                                                                 break;
1389                                                         case 4:
1390                                                                 fprintf(out,
1391                                                                         "%i=2D_16 ",
1392                                                                         tex_num);
1393                                                                 break;
1394                                                         case 5:
1395                                                                 fprintf(out,
1396                                                                         "%i=4D_16 ",
1397                                                                         tex_num);
1398                                                                 break;
1399                                                         case 0xf:
1400                                                                 fprintf(out,
1401                                                                         "%i=NP ",
1402                                                                         tex_num);
1403                                                                 break;
1404                                                         }
1405                                                 }
1406                                                 fprintf(out, "\n");
1407
1408                                                 break;
1409                                         case 3:
1410                                                 instr_out(ctx, i,
1411                                                           "S3: not documented\n");
1412                                                 break;
1413                                         case 4:
1414                                                 {
1415                                                         const char *cullmode = "";
1416                                                         const char *vfmt_xyzw = "";
1417                                                         switch ((data[i] >> 13)
1418                                                                 & 0x3) {
1419                                                         case 0:
1420                                                                 cullmode =
1421                                                                     "both";
1422                                                                 break;
1423                                                         case 1:
1424                                                                 cullmode =
1425                                                                     "none";
1426                                                                 break;
1427                                                         case 2:
1428                                                                 cullmode = "cw";
1429                                                                 break;
1430                                                         case 3:
1431                                                                 cullmode =
1432                                                                     "ccw";
1433                                                                 break;
1434                                                         }
1435                                                         switch (data[i] &
1436                                                                 (7 << 6 | 1 <<
1437                                                                  2)) {
1438                                                         case 1 << 6:
1439                                                                 vfmt_xyzw =
1440                                                                     "XYZ,";
1441                                                                 break;
1442                                                         case 2 << 6:
1443                                                                 vfmt_xyzw =
1444                                                                     "XYZW,";
1445                                                                 break;
1446                                                         case 3 << 6:
1447                                                                 vfmt_xyzw =
1448                                                                     "XY,";
1449                                                                 break;
1450                                                         case 4 << 6:
1451                                                                 vfmt_xyzw =
1452                                                                     "XYW,";
1453                                                                 break;
1454                                                         case 1 << 6 | 1 << 2:
1455                                                                 vfmt_xyzw =
1456                                                                     "XYZF,";
1457                                                                 break;
1458                                                         case 2 << 6 | 1 << 2:
1459                                                                 vfmt_xyzw =
1460                                                                     "XYZWF,";
1461                                                                 break;
1462                                                         case 3 << 6 | 1 << 2:
1463                                                                 vfmt_xyzw =
1464                                                                     "XYF,";
1465                                                                 break;
1466                                                         case 4 << 6 | 1 << 2:
1467                                                                 vfmt_xyzw =
1468                                                                     "XYWF,";
1469                                                                 break;
1470                                                         }
1471                                                         instr_out(ctx, i,
1472                                                                   "S4: point_width=%i, line_width=%.1f,"
1473                                                                   "%s%s%s%s%s cullmode=%s, vfmt=%s%s%s%s%s%s "
1474                                                                   "%s%s%s%s%s\n",
1475                                                                   (data[i] >>
1476                                                                    23) & 0x1ff,
1477                                                                   ((data[i] >>
1478                                                                     19) & 0xf) /
1479                                                                   2.0,
1480                                                                   data[i] & (0xf
1481                                                                              <<
1482                                                                              15)
1483                                                                   ?
1484                                                                   " flatshade="
1485                                                                   : "",
1486                                                                   data[i] & (1
1487                                                                              <<
1488                                                                              18)
1489                                                                   ? "Alpha," :
1490                                                                   "",
1491                                                                   data[i] & (1
1492                                                                              <<
1493                                                                              17)
1494                                                                   ? "Fog," : "",
1495                                                                   data[i] & (1
1496                                                                              <<
1497                                                                              16)
1498                                                                   ? "Specular,"
1499                                                                   : "",
1500                                                                   data[i] & (1
1501                                                                              <<
1502                                                                              15)
1503                                                                   ? "Color," :
1504                                                                   "", cullmode,
1505                                                                   data[i] & (1
1506                                                                              <<
1507                                                                              12)
1508                                                                   ?
1509                                                                   "PointWidth,"
1510                                                                   : "",
1511                                                                   data[i] & (1
1512                                                                              <<
1513                                                                              11)
1514                                                                   ? "SpecFog," :
1515                                                                   "",
1516                                                                   data[i] & (1
1517                                                                              <<
1518                                                                              10)
1519                                                                   ? "Color," :
1520                                                                   "",
1521                                                                   data[i] & (1
1522                                                                              <<
1523                                                                              9)
1524                                                                   ? "DepthOfs,"
1525                                                                   : "",
1526                                                                   vfmt_xyzw,
1527                                                                   data[i] & (1
1528                                                                              <<
1529                                                                              9)
1530                                                                   ? "FogParam,"
1531                                                                   : "",
1532                                                                   data[i] & (1
1533                                                                              <<
1534                                                                              5)
1535                                                                   ?
1536                                                                   "force default diffuse, "
1537                                                                   : "",
1538                                                                   data[i] & (1
1539                                                                              <<
1540                                                                              4)
1541                                                                   ?
1542                                                                   "force default specular, "
1543                                                                   : "",
1544                                                                   data[i] & (1
1545                                                                              <<
1546                                                                              3)
1547                                                                   ?
1548                                                                   "local depth ofs enable, "
1549                                                                   : "",
1550                                                                   data[i] & (1
1551                                                                              <<
1552                                                                              1)
1553                                                                   ?
1554                                                                   "point sprite enable, "
1555                                                                   : "",
1556                                                                   data[i] & (1
1557                                                                              <<
1558                                                                              0)
1559                                                                   ?
1560                                                                   "line AA enable, "
1561                                                                   : "");
1562                                                         break;
1563                                                 }
1564                                         case 5:
1565                                                 {
1566                                                         instr_out(ctx, i,
1567                                                                   "S5:%s%s%s%s%s"
1568                                                                   "%s%s%s%s stencil_ref=0x%x, stencil_test=%s, "
1569                                                                   "stencil_fail=%s, stencil_pass_z_fail=%s, "
1570                                                                   "stencil_pass_z_pass=%s, %s%s%s%s\n",
1571                                                                   data[i] & (0xf
1572                                                                              <<
1573                                                                              28)
1574                                                                   ?
1575                                                                   " write_disable="
1576                                                                   : "",
1577                                                                   data[i] & (1
1578                                                                              <<
1579                                                                              31)
1580                                                                   ? "Alpha," :
1581                                                                   "",
1582                                                                   data[i] & (1
1583                                                                              <<
1584                                                                              30)
1585                                                                   ? "Red," : "",
1586                                                                   data[i] & (1
1587                                                                              <<
1588                                                                              29)
1589                                                                   ? "Green," :
1590                                                                   "",
1591                                                                   data[i] & (1
1592                                                                              <<
1593                                                                              28)
1594                                                                   ? "Blue," :
1595                                                                   "",
1596                                                                   data[i] & (1
1597                                                                              <<
1598                                                                              27)
1599                                                                   ?
1600                                                                   " force default point size,"
1601                                                                   : "",
1602                                                                   data[i] & (1
1603                                                                              <<
1604                                                                              26)
1605                                                                   ?
1606                                                                   " last pixel enable,"
1607                                                                   : "",
1608                                                                   data[i] & (1
1609                                                                              <<
1610                                                                              25)
1611                                                                   ?
1612                                                                   " global depth ofs enable,"
1613                                                                   : "",
1614                                                                   data[i] & (1
1615                                                                              <<
1616                                                                              24)
1617                                                                   ?
1618                                                                   " fog enable,"
1619                                                                   : "",
1620                                                                   (data[i] >>
1621                                                                    16) & 0xff,
1622                                                                   decode_compare_func
1623                                                                   (data[i] >>
1624                                                                    13),
1625                                                                   decode_stencil_op
1626                                                                   (data[i] >>
1627                                                                    10),
1628                                                                   decode_stencil_op
1629                                                                   (data[i] >>
1630                                                                    7),
1631                                                                   decode_stencil_op
1632                                                                   (data[i] >>
1633                                                                    4),
1634                                                                   data[i] & (1
1635                                                                              <<
1636                                                                              3)
1637                                                                   ?
1638                                                                   "stencil write enable, "
1639                                                                   : "",
1640                                                                   data[i] & (1
1641                                                                              <<
1642                                                                              2)
1643                                                                   ?
1644                                                                   "stencil test enable, "
1645                                                                   : "",
1646                                                                   data[i] & (1
1647                                                                              <<
1648                                                                              1)
1649                                                                   ?
1650                                                                   "color dither enable, "
1651                                                                   : "",
1652                                                                   data[i] & (1
1653                                                                              <<
1654                                                                              0)
1655                                                                   ?
1656                                                                   "logicop enable, "
1657                                                                   : "");
1658                                                 }
1659                                                 break;
1660                                         case 6:
1661                                                 instr_out(ctx, i,
1662                                                           "S6: %salpha_test=%s, alpha_ref=0x%x, "
1663                                                           "depth_test=%s, %ssrc_blnd_fct=%s, dst_blnd_fct=%s, "
1664                                                           "%s%stristrip_provoking_vertex=%i\n",
1665                                                           data[i] & (1 << 31) ?
1666                                                           "alpha test enable, "
1667                                                           : "",
1668                                                           decode_compare_func
1669                                                           (data[i] >> 28),
1670                                                           data[i] & (0xff <<
1671                                                                      20),
1672                                                           decode_compare_func
1673                                                           (data[i] >> 16),
1674                                                           data[i] & (1 << 15) ?
1675                                                           "cbuf blend enable, "
1676                                                           : "",
1677                                                           decode_blend_fact(data
1678                                                                             [i]
1679                                                                             >>
1680                                                                             8),
1681                                                           decode_blend_fact(data
1682                                                                             [i]
1683                                                                             >>
1684                                                                             4),
1685                                                           data[i] & (1 << 3) ?
1686                                                           "depth write enable, "
1687                                                           : "",
1688                                                           data[i] & (1 << 2) ?
1689                                                           "cbuf write enable, "
1690                                                           : "",
1691                                                           data[i] & (0x3));
1692                                                 break;
1693                                         case 7:
1694                                                 instr_out(ctx, i,
1695                                                           "S7: depth offset constant: 0x%08x\n",
1696                                                           data[i]);
1697                                                 break;
1698                                         }
1699                                 } else {
1700                                         instr_out(ctx, i,
1701                                                   "S%d: 0x%08x\n", i, data[i]);
1702                                 }
1703                                 i++;
1704                         }
1705                 }
1706                 if (len != i) {
1707                         fprintf(out,
1708                                 "Bad count in 3DSTATE_LOAD_STATE_IMMEDIATE_1\n");
1709                 }
1710                 return len;
1711         case 0x03:
1712                 instr_out(ctx, 0,
1713                           "3DSTATE_LOAD_STATE_IMMEDIATE_2\n");
1714                 len = (data[0] & 0x0000000f) + 2;
1715                 i = 1;
1716                 for (word = 6; word <= 14; word++) {
1717                         if (data[0] & (1 << word)) {
1718                                 if (word == 6)
1719                                         instr_out(ctx, i++,
1720                                                   "TBCF\n");
1721                                 else if (word >= 7 && word <= 10) {
1722                                         instr_out(ctx, i++,
1723                                                   "TB%dC\n", word - 7);
1724                                         instr_out(ctx, i++,
1725                                                   "TB%dA\n", word - 7);
1726                                 } else if (word >= 11 && word <= 14) {
1727                                         instr_out(ctx, i,
1728                                                   "TM%dS0: offset=0x%08x, %s\n",
1729                                                   word - 11,
1730                                                   data[i] & 0xfffffffe,
1731                                                   data[i] & 1 ? "use fence" :
1732                                                   "");
1733                                         i++;
1734                                         instr_out(ctx, i,
1735                                                   "TM%dS1: height=%i, width=%i, %s\n",
1736                                                   word - 11, data[i] >> 21,
1737                                                   (data[i] >> 10) & 0x3ff,
1738                                                   data[i] & 2 ? (data[i] & 1 ?
1739                                                                  "y-tiled" :
1740                                                                  "x-tiled") :
1741                                                   "");
1742                                         i++;
1743                                         instr_out(ctx, i,
1744                                                   "TM%dS2: pitch=%i, \n",
1745                                                   word - 11,
1746                                                   ((data[i] >> 21) + 1) * 4);
1747                                         i++;
1748                                         instr_out(ctx, i++,
1749                                                   "TM%dS3\n", word - 11);
1750                                         instr_out(ctx, i++,
1751                                                   "TM%dS4: dflt color\n",
1752                                                   word - 11);
1753                                 }
1754                         }
1755                 }
1756                 if (len != i) {
1757                         fprintf(out,
1758                                 "Bad count in 3DSTATE_LOAD_STATE_IMMEDIATE_2\n");
1759                 }
1760                 return len;
1761         case 0x00:
1762                 instr_out(ctx, 0, "3DSTATE_MAP_STATE\n");
1763                 len = (data[0] & 0x0000003f) + 2;
1764                 instr_out(ctx, 1, "mask\n");
1765
1766                 i = 2;
1767                 for (map = 0; map <= 15; map++) {
1768                         if (data[1] & (1 << map)) {
1769                                 int width, height, pitch, dword;
1770                                 const char *tiling;
1771
1772                                 dword = data[i];
1773                                 instr_out(ctx, i++,
1774                                           "map %d MS2 %s%s%s\n", map,
1775                                           dword & (1 << 31) ?
1776                                           "untrusted surface, " : "",
1777                                           dword & (1 << 1) ?
1778                                           "vertical line stride enable, " : "",
1779                                           dword & (1 << 0) ?
1780                                           "vertical ofs enable, " : "");
1781
1782                                 dword = data[i];
1783                                 width = ((dword >> 10) & ((1 << 11) - 1)) + 1;
1784                                 height = ((dword >> 21) & ((1 << 11) - 1)) + 1;
1785
1786                                 tiling = "none";
1787                                 if (dword & (1 << 2))
1788                                         tiling = "fenced";
1789                                 else if (dword & (1 << 1))
1790                                         tiling = dword & (1 << 0) ? "Y" : "X";
1791                                 type = " BAD";
1792                                 format = "BAD";
1793                                 switch ((dword >> 7) & 0x7) {
1794                                 case 1:
1795                                         type = "8b";
1796                                         switch ((dword >> 3) & 0xf) {
1797                                         case 0:
1798                                                 format = "I";
1799                                                 break;
1800                                         case 1:
1801                                                 format = "L";
1802                                                 break;
1803                                         case 4:
1804                                                 format = "A";
1805                                                 break;
1806                                         case 5:
1807                                                 format = " mono";
1808                                                 break;
1809                                         }
1810                                         break;
1811                                 case 2:
1812                                         type = "16b";
1813                                         switch ((dword >> 3) & 0xf) {
1814                                         case 0:
1815                                                 format = " rgb565";
1816                                                 break;
1817                                         case 1:
1818                                                 format = " argb1555";
1819                                                 break;
1820                                         case 2:
1821                                                 format = " argb4444";
1822                                                 break;
1823                                         case 5:
1824                                                 format = " ay88";
1825                                                 break;
1826                                         case 6:
1827                                                 format = " bump655";
1828                                                 break;
1829                                         case 7:
1830                                                 format = "I";
1831                                                 break;
1832                                         case 8:
1833                                                 format = "L";
1834                                                 break;
1835                                         case 9:
1836                                                 format = "A";
1837                                                 break;
1838                                         }
1839                                         break;
1840                                 case 3:
1841                                         type = "32b";
1842                                         switch ((dword >> 3) & 0xf) {
1843                                         case 0:
1844                                                 format = " argb8888";
1845                                                 break;
1846                                         case 1:
1847                                                 format = " abgr8888";
1848                                                 break;
1849                                         case 2:
1850                                                 format = " xrgb8888";
1851                                                 break;
1852                                         case 3:
1853                                                 format = " xbgr8888";
1854                                                 break;
1855                                         case 4:
1856                                                 format = " qwvu8888";
1857                                                 break;
1858                                         case 5:
1859                                                 format = " axvu8888";
1860                                                 break;
1861                                         case 6:
1862                                                 format = " lxvu8888";
1863                                                 break;
1864                                         case 7:
1865                                                 format = " xlvu8888";
1866                                                 break;
1867                                         case 8:
1868                                                 format = " argb2101010";
1869                                                 break;
1870                                         case 9:
1871                                                 format = " abgr2101010";
1872                                                 break;
1873                                         case 10:
1874                                                 format = " awvu2101010";
1875                                                 break;
1876                                         case 11:
1877                                                 format = " gr1616";
1878                                                 break;
1879                                         case 12:
1880                                                 format = " vu1616";
1881                                                 break;
1882                                         case 13:
1883                                                 format = " xI824";
1884                                                 break;
1885                                         case 14:
1886                                                 format = " xA824";
1887                                                 break;
1888                                         case 15:
1889                                                 format = " xL824";
1890                                                 break;
1891                                         }
1892                                         break;
1893                                 case 5:
1894                                         type = "422";
1895                                         switch ((dword >> 3) & 0xf) {
1896                                         case 0:
1897                                                 format = " yuv_swapy";
1898                                                 break;
1899                                         case 1:
1900                                                 format = " yuv";
1901                                                 break;
1902                                         case 2:
1903                                                 format = " yuv_swapuv";
1904                                                 break;
1905                                         case 3:
1906                                                 format = " yuv_swapuvy";
1907                                                 break;
1908                                         }
1909                                         break;
1910                                 case 6:
1911                                         type = "compressed";
1912                                         switch ((dword >> 3) & 0x7) {
1913                                         case 0:
1914                                                 format = " dxt1";
1915                                                 break;
1916                                         case 1:
1917                                                 format = " dxt2_3";
1918                                                 break;
1919                                         case 2:
1920                                                 format = " dxt4_5";
1921                                                 break;
1922                                         case 3:
1923                                                 format = " fxt1";
1924                                                 break;
1925                                         case 4:
1926                                                 format = " dxt1_rb";
1927                                                 break;
1928                                         }
1929                                         break;
1930                                 case 7:
1931                                         type = "4b indexed";
1932                                         switch ((dword >> 3) & 0xf) {
1933                                         case 7:
1934                                                 format = " argb8888";
1935                                                 break;
1936                                         }
1937                                         break;
1938                                 }
1939                                 dword = data[i];
1940                                 instr_out(ctx, i++,
1941                                           "map %d MS3 [width=%d, height=%d, format=%s%s, tiling=%s%s]\n",
1942                                           map, width, height, type, format,
1943                                           tiling,
1944                                           dword & (1 << 9) ? " palette select" :
1945                                           "");
1946
1947                                 dword = data[i];
1948                                 pitch =
1949                                     4 * (((dword >> 21) & ((1 << 11) - 1)) + 1);
1950                                 instr_out(ctx, i++,
1951                                           "map %d MS4 [pitch=%d, max_lod=%i, vol_depth=%i, cube_face_ena=%x, %s]\n",
1952                                           map, pitch, (dword >> 9) & 0x3f,
1953                                           dword & 0xff, (dword >> 15) & 0x3f,
1954                                           dword & (1 << 8) ? "miplayout legacy"
1955                                           : "miplayout right");
1956                         }
1957                 }
1958                 if (len != i) {
1959                         fprintf(out, "Bad count in 3DSTATE_MAP_STATE\n");
1960                         return len;
1961                 }
1962                 return len;
1963         case 0x06:
1964                 instr_out(ctx, 0,
1965                           "3DSTATE_PIXEL_SHADER_CONSTANTS\n");
1966                 len = (data[0] & 0x000000ff) + 2;
1967
1968                 i = 2;
1969                 for (c = 0; c <= 31; c++) {
1970                         if (data[1] & (1 << c)) {
1971                                 instr_out(ctx, i, "C%d.X = %f\n", c,
1972                                           int_as_float(data[i]));
1973                                 i++;
1974                                 instr_out(ctx, i, "C%d.Y = %f\n",
1975                                           c, int_as_float(data[i]));
1976                                 i++;
1977                                 instr_out(ctx, i, "C%d.Z = %f\n",
1978                                           c, int_as_float(data[i]));
1979                                 i++;
1980                                 instr_out(ctx, i, "C%d.W = %f\n",
1981                                           c, int_as_float(data[i]));
1982                                 i++;
1983                         }
1984                 }
1985                 if (len != i) {
1986                         fprintf(out,
1987                                 "Bad count in 3DSTATE_PIXEL_SHADER_CONSTANTS\n");
1988                 }
1989                 return len;
1990         case 0x05:
1991                 instr_out(ctx, 0, "3DSTATE_PIXEL_SHADER_PROGRAM\n");
1992                 len = (data[0] & 0x000000ff) + 2;
1993                 if ((len - 1) % 3 != 0 || len > 370) {
1994                         fprintf(out,
1995                                 "Bad count in 3DSTATE_PIXEL_SHADER_PROGRAM\n");
1996                 }
1997                 i = 1;
1998                 for (instr = 0; instr < (len - 1) / 3; instr++) {
1999                         char instr_prefix[10];
2000
2001                         sprintf(instr_prefix, "PS%03d", instr);
2002                         i915_decode_instruction(ctx, i,
2003                                                 instr_prefix);
2004                         i += 3;
2005                 }
2006                 return len;
2007         case 0x01:
2008                 if (IS_GEN2(devid))
2009                         break;
2010                 instr_out(ctx, 0, "3DSTATE_SAMPLER_STATE\n");
2011                 instr_out(ctx, 1, "mask\n");
2012                 len = (data[0] & 0x0000003f) + 2;
2013                 i = 2;
2014                 for (sampler = 0; sampler <= 15; sampler++) {
2015                         if (data[1] & (1 << sampler)) {
2016                                 uint32_t dword;
2017                                 const char *mip_filter = "";
2018
2019                                 dword = data[i];
2020                                 switch ((dword >> 20) & 0x3) {
2021                                 case 0:
2022                                         mip_filter = "none";
2023                                         break;
2024                                 case 1:
2025                                         mip_filter = "nearest";
2026                                         break;
2027                                 case 3:
2028                                         mip_filter = "linear";
2029                                         break;
2030                                 }
2031                                 instr_out(ctx, i++,
2032                                           "sampler %d SS2:%s%s%s "
2033                                           "base_mip_level=%i, mip_filter=%s, mag_filter=%s, min_filter=%s "
2034                                           "lod_bias=%.2f,%s max_aniso=%i, shadow_func=%s\n",
2035                                           sampler,
2036                                           dword & (1 << 31) ? " reverse gamma,"
2037                                           : "",
2038                                           dword & (1 << 30) ? " packed2planar,"
2039                                           : "",
2040                                           dword & (1 << 29) ?
2041                                           " colorspace conversion," : "",
2042                                           (dword >> 22) & 0x1f, mip_filter,
2043                                           decode_sample_filter(dword >> 17),
2044                                           decode_sample_filter(dword >> 14),
2045                                           ((dword >> 5) & 0x1ff) / (0x10 * 1.0),
2046                                           dword & (1 << 4) ? " shadow," : "",
2047                                           dword & (1 << 3) ? 4 : 2,
2048                                           decode_compare_func(dword));
2049                                 dword = data[i];
2050                                 instr_out(ctx, i++,
2051                                           "sampler %d SS3: min_lod=%.2f,%s "
2052                                           "tcmode_x=%s, tcmode_y=%s, tcmode_z=%s,%s texmap_idx=%i,%s\n",
2053                                           sampler,
2054                                           ((dword >> 24) & 0xff) / (0x10 * 1.0),
2055                                           dword & (1 << 17) ?
2056                                           " kill pixel enable," : "",
2057                                           decode_tex_coord_mode(dword >> 12),
2058                                           decode_tex_coord_mode(dword >> 9),
2059                                           decode_tex_coord_mode(dword >> 6),
2060                                           dword & (1 << 5) ?
2061                                           " normalized coords," : "",
2062                                           (dword >> 1) & 0xf,
2063                                           dword & (1 << 0) ? " deinterlacer," :
2064                                           "");
2065                                 dword = data[i];
2066                                 instr_out(ctx, i++,
2067                                           "sampler %d SS4: border color\n",
2068                                           sampler);
2069                         }
2070                 }
2071                 if (len != i) {
2072                         fprintf(out, "Bad count in 3DSTATE_SAMPLER_STATE\n");
2073                 }
2074                 return len;
2075         case 0x85:
2076                 len = (data[0] & 0x0000000f) + 2;
2077
2078                 if (len != 2)
2079                         fprintf(out,
2080                                 "Bad count in 3DSTATE_DEST_BUFFER_VARIABLES\n");
2081
2082                 instr_out(ctx, 0,
2083                           "3DSTATE_DEST_BUFFER_VARIABLES\n");
2084
2085                 switch ((data[1] >> 8) & 0xf) {
2086                 case 0x0:
2087                         format = "g8";
2088                         break;
2089                 case 0x1:
2090                         format = "x1r5g5b5";
2091                         break;
2092                 case 0x2:
2093                         format = "r5g6b5";
2094                         break;
2095                 case 0x3:
2096                         format = "a8r8g8b8";
2097                         break;
2098                 case 0x4:
2099                         format = "ycrcb_swapy";
2100                         break;
2101                 case 0x5:
2102                         format = "ycrcb_normal";
2103                         break;
2104                 case 0x6:
2105                         format = "ycrcb_swapuv";
2106                         break;
2107                 case 0x7:
2108                         format = "ycrcb_swapuvy";
2109                         break;
2110                 case 0x8:
2111                         format = "a4r4g4b4";
2112                         break;
2113                 case 0x9:
2114                         format = "a1r5g5b5";
2115                         break;
2116                 case 0xa:
2117                         format = "a2r10g10b10";
2118                         break;
2119                 default:
2120                         format = "BAD";
2121                         break;
2122                 }
2123                 switch ((data[1] >> 2) & 0x3) {
2124                 case 0x0:
2125                         zformat = "u16";
2126                         break;
2127                 case 0x1:
2128                         zformat = "f16";
2129                         break;
2130                 case 0x2:
2131                         zformat = "u24x8";
2132                         break;
2133                 default:
2134                         zformat = "BAD";
2135                         break;
2136                 }
2137                 instr_out(ctx, 1,
2138                           "%s format, %s depth format, early Z %sabled\n",
2139                           format, zformat,
2140                           (data[1] & (1 << 31)) ? "en" : "dis");
2141                 return len;
2142
2143         case 0x8e:
2144                 {
2145                         const char *name, *tiling;
2146
2147                         len = (data[0] & 0x0000000f) + 2;
2148                         if (len != 3)
2149                                 fprintf(out,
2150                                         "Bad count in 3DSTATE_BUFFER_INFO\n");
2151
2152                         switch ((data[1] >> 24) & 0x7) {
2153                         case 0x3:
2154                                 name = "color";
2155                                 break;
2156                         case 0x7:
2157                                 name = "depth";
2158                                 break;
2159                         default:
2160                                 name = "unknown";
2161                                 break;
2162                         }
2163
2164                         tiling = "none";
2165                         if (data[1] & (1 << 23))
2166                                 tiling = "fenced";
2167                         else if (data[1] & (1 << 22))
2168                                 tiling = data[1] & (1 << 21) ? "Y" : "X";
2169
2170                         instr_out(ctx, 0, "3DSTATE_BUFFER_INFO\n");
2171                         instr_out(ctx, 1,
2172                                   "%s, tiling = %s, pitch=%d\n", name, tiling,
2173                                   data[1] & 0xffff);
2174
2175                         instr_out(ctx, 2, "address\n");
2176                         return len;
2177                 }
2178         case 0x81:
2179                 len = (data[0] & 0x0000000f) + 2;
2180
2181                 if (len != 3)
2182                         fprintf(out,
2183                                 "Bad count in 3DSTATE_SCISSOR_RECTANGLE\n");
2184
2185                 instr_out(ctx, 0, "3DSTATE_SCISSOR_RECTANGLE\n");
2186                 instr_out(ctx, 1, "(%d,%d)\n",
2187                           data[1] & 0xffff, data[1] >> 16);
2188                 instr_out(ctx, 2, "(%d,%d)\n",
2189                           data[2] & 0xffff, data[2] >> 16);
2190
2191                 return len;
2192         case 0x80:
2193                 len = (data[0] & 0x0000000f) + 2;
2194
2195                 if (len != 5)
2196                         fprintf(out,
2197                                 "Bad count in 3DSTATE_DRAWING_RECTANGLE\n");
2198
2199                 instr_out(ctx, 0, "3DSTATE_DRAWING_RECTANGLE\n");
2200                 instr_out(ctx, 1, "%s\n",
2201                           data[1] & (1 << 30) ? "depth ofs disabled " : "");
2202                 instr_out(ctx, 2, "(%d,%d)\n",
2203                           data[2] & 0xffff, data[2] >> 16);
2204                 instr_out(ctx, 3, "(%d,%d)\n",
2205                           data[3] & 0xffff, data[3] >> 16);
2206                 instr_out(ctx, 4, "(%d,%d)\n",
2207                           data[4] & 0xffff, data[4] >> 16);
2208
2209                 return len;
2210         case 0x9c:
2211                 len = (data[0] & 0x0000000f) + 2;
2212
2213                 if (len != 7)
2214                         fprintf(out, "Bad count in 3DSTATE_CLEAR_PARAMETERS\n");
2215
2216                 instr_out(ctx, 0, "3DSTATE_CLEAR_PARAMETERS\n");
2217                 instr_out(ctx, 1, "prim_type=%s, clear=%s%s%s\n",
2218                           data[1] & (1 << 16) ? "CLEAR_RECT" : "ZONE_INIT",
2219                           data[1] & (1 << 2) ? "color," : "",
2220                           data[1] & (1 << 1) ? "depth," : "",
2221                           data[1] & (1 << 0) ? "stencil," : "");
2222                 instr_out(ctx, 2, "clear color\n");
2223                 instr_out(ctx, 3, "clear depth/stencil\n");
2224                 instr_out(ctx, 4, "color value (rgba8888)\n");
2225                 instr_out(ctx, 5, "depth value %f\n",
2226                           int_as_float(data[5]));
2227                 instr_out(ctx, 6, "clear stencil\n");
2228                 return len;
2229         }
2230
2231         for (idx = 0; idx < ARRAY_SIZE(opcodes_3d_1d); idx++) {
2232                 opcode_3d_1d = &opcodes_3d_1d[idx];
2233                 if (opcode_3d_1d->i830_only && !IS_GEN2(devid))
2234                         continue;
2235
2236                 if (((data[0] & 0x00ff0000) >> 16) == opcode_3d_1d->opcode) {
2237                         len = 1;
2238
2239                         instr_out(ctx, 0, "%s\n",
2240                                   opcode_3d_1d->name);
2241                         if (opcode_3d_1d->max_len > 1) {
2242                                 len = (data[0] & 0x0000ffff) + 2;
2243                                 if (len < opcode_3d_1d->min_len ||
2244                                     len > opcode_3d_1d->max_len) {
2245                                         fprintf(out, "Bad count in %s\n",
2246                                                 opcode_3d_1d->name);
2247                                 }
2248                         }
2249
2250                         for (i = 1; i < len; i++) {
2251                                 instr_out(ctx, i, "dword %d\n", i);
2252                         }
2253
2254                         return len;
2255                 }
2256         }
2257
2258         instr_out(ctx, 0, "3D UNKNOWN: 3d_1d opcode = 0x%x\n",
2259                   opcode);
2260         return 1;
2261 }
2262
2263 static int
2264 decode_3d_primitive(struct drm_intel_decode *ctx)
2265 {
2266         uint32_t *data = ctx->data;
2267         uint32_t count = ctx->count;
2268         char immediate = (data[0] & (1 << 23)) == 0;
2269         unsigned int len, i, j, ret;
2270         const char *primtype;
2271         int original_s2 = saved_s2;
2272         int original_s4 = saved_s4;
2273
2274         switch ((data[0] >> 18) & 0xf) {
2275         case 0x0:
2276                 primtype = "TRILIST";
2277                 break;
2278         case 0x1:
2279                 primtype = "TRISTRIP";
2280                 break;
2281         case 0x2:
2282                 primtype = "TRISTRIP_REVERSE";
2283                 break;
2284         case 0x3:
2285                 primtype = "TRIFAN";
2286                 break;
2287         case 0x4:
2288                 primtype = "POLYGON";
2289                 break;
2290         case 0x5:
2291                 primtype = "LINELIST";
2292                 break;
2293         case 0x6:
2294                 primtype = "LINESTRIP";
2295                 break;
2296         case 0x7:
2297                 primtype = "RECTLIST";
2298                 break;
2299         case 0x8:
2300                 primtype = "POINTLIST";
2301                 break;
2302         case 0x9:
2303                 primtype = "DIB";
2304                 break;
2305         case 0xa:
2306                 primtype = "CLEAR_RECT";
2307                 saved_s4 = 3 << 6;
2308                 saved_s2 = ~0;
2309                 break;
2310         default:
2311                 primtype = "unknown";
2312                 break;
2313         }
2314
2315         /* XXX: 3DPRIM_DIB not supported */
2316         if (immediate) {
2317                 len = (data[0] & 0x0003ffff) + 2;
2318                 instr_out(ctx, 0, "3DPRIMITIVE inline %s\n",
2319                           primtype);
2320                 if (count < len)
2321                         BUFFER_FAIL(count, len, "3DPRIMITIVE inline");
2322                 if (!saved_s2_set || !saved_s4_set) {
2323                         fprintf(out, "unknown vertex format\n");
2324                         for (i = 1; i < len; i++) {
2325                                 instr_out(ctx, i,
2326                                           "           vertex data (%f float)\n",
2327                                           int_as_float(data[i]));
2328                         }
2329                 } else {
2330                         unsigned int vertex = 0;
2331                         for (i = 1; i < len;) {
2332                                 unsigned int tc;
2333
2334 #define VERTEX_OUT(fmt, ...) do {                                       \
2335     if (i < len)                                                        \
2336         instr_out(ctx, i, " V%d."fmt"\n", vertex, __VA_ARGS__); \
2337     else                                                                \
2338         fprintf(out, " missing data in V%d\n", vertex);                 \
2339     i++;                                                                \
2340 } while (0)
2341
2342                                 VERTEX_OUT("X = %f", int_as_float(data[i]));
2343                                 VERTEX_OUT("Y = %f", int_as_float(data[i]));
2344                                 switch (saved_s4 >> 6 & 0x7) {
2345                                 case 0x1:
2346                                         VERTEX_OUT("Z = %f",
2347                                                    int_as_float(data[i]));
2348                                         break;
2349                                 case 0x2:
2350                                         VERTEX_OUT("Z = %f",
2351                                                    int_as_float(data[i]));
2352                                         VERTEX_OUT("W = %f",
2353                                                    int_as_float(data[i]));
2354                                         break;
2355                                 case 0x3:
2356                                         break;
2357                                 case 0x4:
2358                                         VERTEX_OUT("W = %f",
2359                                                    int_as_float(data[i]));
2360                                         break;
2361                                 default:
2362                                         fprintf(out, "bad S4 position mask\n");
2363                                 }
2364
2365                                 if (saved_s4 & (1 << 10)) {
2366                                         VERTEX_OUT
2367                                             ("color = (A=0x%02x, R=0x%02x, G=0x%02x, "
2368                                              "B=0x%02x)", data[i] >> 24,
2369                                              (data[i] >> 16) & 0xff,
2370                                              (data[i] >> 8) & 0xff,
2371                                              data[i] & 0xff);
2372                                 }
2373                                 if (saved_s4 & (1 << 11)) {
2374                                         VERTEX_OUT
2375                                             ("spec = (A=0x%02x, R=0x%02x, G=0x%02x, "
2376                                              "B=0x%02x)", data[i] >> 24,
2377                                              (data[i] >> 16) & 0xff,
2378                                              (data[i] >> 8) & 0xff,
2379                                              data[i] & 0xff);
2380                                 }
2381                                 if (saved_s4 & (1 << 12))
2382                                         VERTEX_OUT("width = 0x%08x)", data[i]);
2383
2384                                 for (tc = 0; tc <= 7; tc++) {
2385                                         switch ((saved_s2 >> (tc * 4)) & 0xf) {
2386                                         case 0x0:
2387                                                 VERTEX_OUT("T%d.X = %f", tc,
2388                                                            int_as_float(data
2389                                                                         [i]));
2390                                                 VERTEX_OUT("T%d.Y = %f", tc,
2391                                                            int_as_float(data
2392                                                                         [i]));
2393                                                 break;
2394                                         case 0x1:
2395                                                 VERTEX_OUT("T%d.X = %f", tc,
2396                                                            int_as_float(data
2397                                                                         [i]));
2398                                                 VERTEX_OUT("T%d.Y = %f", tc,
2399                                                            int_as_float(data
2400                                                                         [i]));
2401                                                 VERTEX_OUT("T%d.Z = %f", tc,
2402                                                            int_as_float(data
2403                                                                         [i]));
2404                                                 break;
2405                                         case 0x2:
2406                                                 VERTEX_OUT("T%d.X = %f", tc,
2407                                                            int_as_float(data
2408                                                                         [i]));
2409                                                 VERTEX_OUT("T%d.Y = %f", tc,
2410                                                            int_as_float(data
2411                                                                         [i]));
2412                                                 VERTEX_OUT("T%d.Z = %f", tc,
2413                                                            int_as_float(data
2414                                                                         [i]));
2415                                                 VERTEX_OUT("T%d.W = %f", tc,
2416                                                            int_as_float(data
2417                                                                         [i]));
2418                                                 break;
2419                                         case 0x3:
2420                                                 VERTEX_OUT("T%d.X = %f", tc,
2421                                                            int_as_float(data
2422                                                                         [i]));
2423                                                 break;
2424                                         case 0x4:
2425                                                 VERTEX_OUT
2426                                                     ("T%d.XY = 0x%08x half-float",
2427                                                      tc, data[i]);
2428                                                 break;
2429                                         case 0x5:
2430                                                 VERTEX_OUT
2431                                                     ("T%d.XY = 0x%08x half-float",
2432                                                      tc, data[i]);
2433                                                 VERTEX_OUT
2434                                                     ("T%d.ZW = 0x%08x half-float",
2435                                                      tc, data[i]);
2436                                                 break;
2437                                         case 0xf:
2438                                                 break;
2439                                         default:
2440                                                 fprintf(out,
2441                                                         "bad S2.T%d format\n",
2442                                                         tc);
2443                                         }
2444                                 }
2445                                 vertex++;
2446                         }
2447                 }
2448
2449                 ret = len;
2450         } else {
2451                 /* indirect vertices */
2452                 len = data[0] & 0x0000ffff;     /* index count */
2453                 if (data[0] & (1 << 17)) {
2454                         /* random vertex access */
2455                         if (count < (len + 1) / 2 + 1) {
2456                                 BUFFER_FAIL(count, (len + 1) / 2 + 1,
2457                                             "3DPRIMITIVE random indirect");
2458                         }
2459                         instr_out(ctx, 0,
2460                                   "3DPRIMITIVE random indirect %s (%d)\n",
2461                                   primtype, len);
2462                         if (len == 0) {
2463                                 /* vertex indices continue until 0xffff is
2464                                  * found
2465                                  */
2466                                 for (i = 1; i < count; i++) {
2467                                         if ((data[i] & 0xffff) == 0xffff) {
2468                                                 instr_out(ctx, i,
2469                                                           "    indices: (terminator)\n");
2470                                                 ret = i;
2471                                                 goto out;
2472                                         } else if ((data[i] >> 16) == 0xffff) {
2473                                                 instr_out(ctx, i,
2474                                                           "    indices: 0x%04x, (terminator)\n",
2475                                                           data[i] & 0xffff);
2476                                                 ret = i;
2477                                                 goto out;
2478                                         } else {
2479                                                 instr_out(ctx, i,
2480                                                           "    indices: 0x%04x, 0x%04x\n",
2481                                                           data[i] & 0xffff,
2482                                                           data[i] >> 16);
2483                                         }
2484                                 }
2485                                 fprintf(out,
2486                                         "3DPRIMITIVE: no terminator found in index buffer\n");
2487                                 ret = count;
2488                                 goto out;
2489                         } else {
2490                                 /* fixed size vertex index buffer */
2491                                 for (j = 1, i = 0; i < len; i += 2, j++) {
2492                                         if (i * 2 == len - 1) {
2493                                                 instr_out(ctx, j,
2494                                                           "    indices: 0x%04x\n",
2495                                                           data[j] & 0xffff);
2496                                         } else {
2497                                                 instr_out(ctx, j,
2498                                                           "    indices: 0x%04x, 0x%04x\n",
2499                                                           data[j] & 0xffff,
2500                                                           data[j] >> 16);
2501                                         }
2502                                 }
2503                         }
2504                         ret = (len + 1) / 2 + 1;
2505                         goto out;
2506                 } else {
2507                         /* sequential vertex access */
2508                         instr_out(ctx, 0,
2509                                   "3DPRIMITIVE sequential indirect %s, %d starting from "
2510                                   "%d\n", primtype, len, data[1] & 0xffff);
2511                         instr_out(ctx, 1, "           start\n");
2512                         ret = 2;
2513                         goto out;
2514                 }
2515         }
2516
2517 out:
2518         saved_s2 = original_s2;
2519         saved_s4 = original_s4;
2520         return ret;
2521 }
2522
2523 static int
2524 decode_3d(struct drm_intel_decode *ctx)
2525 {
2526         uint32_t opcode;
2527         unsigned int idx;
2528         uint32_t *data = ctx->data;
2529
2530         struct {
2531                 uint32_t opcode;
2532                 unsigned int min_len;
2533                 unsigned int max_len;
2534                 const char *name;
2535         } opcodes_3d[] = {
2536                 { 0x06, 1, 1, "3DSTATE_ANTI_ALIASING" },
2537                 { 0x08, 1, 1, "3DSTATE_BACKFACE_STENCIL_OPS" },
2538                 { 0x09, 1, 1, "3DSTATE_BACKFACE_STENCIL_MASKS" },
2539                 { 0x16, 1, 1, "3DSTATE_COORD_SET_BINDINGS" },
2540                 { 0x15, 1, 1, "3DSTATE_FOG_COLOR" },
2541                 { 0x0b, 1, 1, "3DSTATE_INDEPENDENT_ALPHA_BLEND" },
2542                 { 0x0d, 1, 1, "3DSTATE_MODES_4" },
2543                 { 0x0c, 1, 1, "3DSTATE_MODES_5" },
2544                 { 0x07, 1, 1, "3DSTATE_RASTERIZATION_RULES"},
2545         }, *opcode_3d;
2546
2547         opcode = (data[0] & 0x1f000000) >> 24;
2548
2549         switch (opcode) {
2550         case 0x1f:
2551                 return decode_3d_primitive(ctx);
2552         case 0x1d:
2553                 return decode_3d_1d(ctx);
2554         case 0x1c:
2555                 return decode_3d_1c(ctx);
2556         }
2557
2558         for (idx = 0; idx < ARRAY_SIZE(opcodes_3d); idx++) {
2559                 opcode_3d = &opcodes_3d[idx];
2560                 if (opcode == opcode_3d->opcode) {
2561                         unsigned int len = 1, i;
2562
2563                         instr_out(ctx, 0, "%s\n", opcode_3d->name);
2564                         if (opcode_3d->max_len > 1) {
2565                                 len = (data[0] & 0xff) + 2;
2566                                 if (len < opcode_3d->min_len ||
2567                                     len > opcode_3d->max_len) {
2568                                         fprintf(out, "Bad count in %s\n",
2569                                                 opcode_3d->name);
2570                                 }
2571                         }
2572
2573                         for (i = 1; i < len; i++) {
2574                                 instr_out(ctx, i, "dword %d\n", i);
2575                         }
2576                         return len;
2577                 }
2578         }
2579
2580         instr_out(ctx, 0, "3D UNKNOWN: 3d opcode = 0x%x\n", opcode);
2581         return 1;
2582 }
2583
2584 static const char *get_965_surfacetype(unsigned int surfacetype)
2585 {
2586         switch (surfacetype) {
2587         case 0:
2588                 return "1D";
2589         case 1:
2590                 return "2D";
2591         case 2:
2592                 return "3D";
2593         case 3:
2594                 return "CUBE";
2595         case 4:
2596                 return "BUFFER";
2597         case 7:
2598                 return "NULL";
2599         default:
2600                 return "unknown";
2601         }
2602 }
2603
2604 static const char *get_965_depthformat(unsigned int depthformat)
2605 {
2606         switch (depthformat) {
2607         case 0:
2608                 return "s8_z24float";
2609         case 1:
2610                 return "z32float";
2611         case 2:
2612                 return "z24s8";
2613         case 5:
2614                 return "z16";
2615         default:
2616                 return "unknown";
2617         }
2618 }
2619
2620 static const char *get_965_element_component(uint32_t data, int component)
2621 {
2622         uint32_t component_control = (data >> (16 + (3 - component) * 4)) & 0x7;
2623
2624         switch (component_control) {
2625         case 0:
2626                 return "nostore";
2627         case 1:
2628                 switch (component) {
2629                 case 0:
2630                         return "X";
2631                 case 1:
2632                         return "Y";
2633                 case 2:
2634                         return "Z";
2635                 case 3:
2636                         return "W";
2637                 default:
2638                         return "fail";
2639                 }
2640         case 2:
2641                 return "0.0";
2642         case 3:
2643                 return "1.0";
2644         case 4:
2645                 return "0x1";
2646         case 5:
2647                 return "VID";
2648         default:
2649                 return "fail";
2650         }
2651 }
2652
2653 static const char *get_965_prim_type(uint32_t primtype)
2654 {
2655         switch (primtype) {
2656         case 0x01:
2657                 return "point list";
2658         case 0x02:
2659                 return "line list";
2660         case 0x03:
2661                 return "line strip";
2662         case 0x04:
2663                 return "tri list";
2664         case 0x05:
2665                 return "tri strip";
2666         case 0x06:
2667                 return "tri fan";
2668         case 0x07:
2669                 return "quad list";
2670         case 0x08:
2671                 return "quad strip";
2672         case 0x09:
2673                 return "line list adj";
2674         case 0x0a:
2675                 return "line strip adj";
2676         case 0x0b:
2677                 return "tri list adj";
2678         case 0x0c:
2679                 return "tri strip adj";
2680         case 0x0d:
2681                 return "tri strip reverse";
2682         case 0x0e:
2683                 return "polygon";
2684         case 0x0f:
2685                 return "rect list";
2686         case 0x10:
2687                 return "line loop";
2688         case 0x11:
2689                 return "point list bf";
2690         case 0x12:
2691                 return "line strip cont";
2692         case 0x13:
2693                 return "line strip bf";
2694         case 0x14:
2695                 return "line strip cont bf";
2696         case 0x15:
2697                 return "tri fan no stipple";
2698         default:
2699                 return "fail";
2700         }
2701 }
2702
2703 static int
2704 i965_decode_urb_fence(struct drm_intel_decode *ctx, int len)
2705 {
2706         uint32_t vs_fence, clip_fence, gs_fence, sf_fence, vfe_fence, cs_fence;
2707         uint32_t *data = ctx->data;
2708
2709         if (len != 3)
2710                 fprintf(out, "Bad count in URB_FENCE\n");
2711
2712         vs_fence = data[1] & 0x3ff;
2713         gs_fence = (data[1] >> 10) & 0x3ff;
2714         clip_fence = (data[1] >> 20) & 0x3ff;
2715         sf_fence = data[2] & 0x3ff;
2716         vfe_fence = (data[2] >> 10) & 0x3ff;
2717         cs_fence = (data[2] >> 20) & 0x7ff;
2718
2719         instr_out(ctx, 0, "URB_FENCE: %s%s%s%s%s%s\n",
2720                   (data[0] >> 13) & 1 ? "cs " : "",
2721                   (data[0] >> 12) & 1 ? "vfe " : "",
2722                   (data[0] >> 11) & 1 ? "sf " : "",
2723                   (data[0] >> 10) & 1 ? "clip " : "",
2724                   (data[0] >> 9) & 1 ? "gs " : "",
2725                   (data[0] >> 8) & 1 ? "vs " : "");
2726         instr_out(ctx, 1,
2727                   "vs fence: %d, clip_fence: %d, gs_fence: %d\n",
2728                   vs_fence, clip_fence, gs_fence);
2729         instr_out(ctx, 2,
2730                   "sf fence: %d, vfe_fence: %d, cs_fence: %d\n",
2731                   sf_fence, vfe_fence, cs_fence);
2732         if (gs_fence < vs_fence)
2733                 fprintf(out, "gs fence < vs fence!\n");
2734         if (clip_fence < gs_fence)
2735                 fprintf(out, "clip fence < gs fence!\n");
2736         if (sf_fence < clip_fence)
2737                 fprintf(out, "sf fence < clip fence!\n");
2738         if (cs_fence < sf_fence)
2739                 fprintf(out, "cs fence < sf fence!\n");
2740
2741         return len;
2742 }
2743
2744 static void
2745 state_base_out(struct drm_intel_decode *ctx, unsigned int index,
2746                const char *name)
2747 {
2748         if (ctx->data[index] & 1) {
2749                 instr_out(ctx, index,
2750                           "%s state base address 0x%08x\n", name,
2751                           ctx->data[index] & ~1);
2752         } else {
2753                 instr_out(ctx, index, "%s state base not updated\n",
2754                           name);
2755         }
2756 }
2757
2758 static void
2759 state_max_out(struct drm_intel_decode *ctx, unsigned int index,
2760               const char *name)
2761 {
2762         if (ctx->data[index] & 1) {
2763                 if (ctx->data[index] == 1) {
2764                         instr_out(ctx, index,
2765                                   "%s state upper bound disabled\n", name);
2766                 } else {
2767                         instr_out(ctx, index,
2768                                   "%s state upper bound 0x%08x\n", name,
2769                                   ctx->data[index] & ~1);
2770                 }
2771         } else {
2772                 instr_out(ctx, index,
2773                           "%s state upper bound not updated\n", name);
2774         }
2775 }
2776
2777 static int
2778 gen7_3DSTATE_VIEWPORT_STATE_POINTERS_CC(struct drm_intel_decode *ctx)
2779 {
2780         instr_out(ctx, 0, "3DSTATE_VIEWPORT_STATE_POINTERS_CC\n");
2781         instr_out(ctx, 1, "pointer to CC viewport\n");
2782
2783         return 2;
2784 }
2785
2786 static int
2787 gen7_3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP(struct drm_intel_decode *ctx)
2788 {
2789         instr_out(ctx, 0, "3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP\n");
2790         instr_out(ctx, 1, "pointer to SF_CLIP viewport\n");
2791
2792         return 2;
2793 }
2794
2795 static int
2796 gen7_3DSTATE_BLEND_STATE_POINTERS(struct drm_intel_decode *ctx)
2797 {
2798         instr_out(ctx, 0, "3DSTATE_BLEND_STATE_POINTERS\n");
2799         instr_out(ctx, 1, "pointer to BLEND_STATE at 0x%08x (%s)\n",
2800                   ctx->data[1] & ~1,
2801                   (ctx->data[1] & 1) ? "changed" : "unchanged");
2802
2803         return 2;
2804 }
2805
2806 static int
2807 gen7_3DSTATE_DEPTH_STENCIL_STATE_POINTERS(struct drm_intel_decode *ctx)
2808 {
2809         instr_out(ctx, 0, "3DSTATE_DEPTH_STENCIL_STATE_POINTERS\n");
2810         instr_out(ctx, 1,
2811                   "pointer to DEPTH_STENCIL_STATE at 0x%08x (%s)\n",
2812                   ctx->data[1] & ~1,
2813                   (ctx->data[1] & 1) ? "changed" : "unchanged");
2814
2815         return 2;
2816 }
2817
2818 static int
2819 gen7_3DSTATE_HIER_DEPTH_BUFFER(struct drm_intel_decode *ctx)
2820 {
2821         instr_out(ctx, 0, "3DSTATE_HIER_DEPTH_BUFFER\n");
2822         instr_out(ctx, 1, "pitch %db\n",
2823                   (ctx->data[1] & 0x1ffff) + 1);
2824         instr_out(ctx, 2, "pointer to HiZ buffer\n");
2825
2826         return 3;
2827 }
2828
2829 static int
2830 gen6_3DSTATE_CC_STATE_POINTERS(struct drm_intel_decode *ctx)
2831 {
2832         instr_out(ctx, 0, "3DSTATE_CC_STATE_POINTERS\n");
2833         instr_out(ctx, 1, "blend change %d\n", ctx->data[1] & 1);
2834         instr_out(ctx, 2, "depth stencil change %d\n",
2835                   ctx->data[2] & 1);
2836         instr_out(ctx, 3, "cc change %d\n", ctx->data[3] & 1);
2837
2838         return 4;
2839 }
2840
2841 static int
2842 gen7_3DSTATE_CC_STATE_POINTERS(struct drm_intel_decode *ctx)
2843 {
2844         instr_out(ctx, 0, "3DSTATE_CC_STATE_POINTERS\n");
2845         instr_out(ctx, 1, "pointer to COLOR_CALC_STATE at 0x%08x "
2846                   "(%s)\n",
2847                   ctx->data[1] & ~1,
2848                   (ctx->data[1] & 1) ? "changed" : "unchanged");
2849
2850         return 2;
2851 }
2852
2853 static int
2854 gen7_3DSTATE_URB_unit(struct drm_intel_decode *ctx, const char *unit)
2855 {
2856     int start_kb = ((ctx->data[1] >> 25) & 0x3f) * 8;
2857     /* the field is # of 512-bit rows - 1, we print bytes */
2858     int entry_size = (((ctx->data[1] >> 16) & 0x1ff) + 1);
2859     int nr_entries = ctx->data[1] & 0xffff;
2860
2861     instr_out(ctx, 0, "3DSTATE_URB_%s\n", unit);
2862     instr_out(ctx, 1,
2863               "%dKB start, size=%d 64B rows, nr_entries=%d, total size %dB\n",
2864               start_kb, entry_size, nr_entries, nr_entries * 64 * entry_size);
2865
2866     return 2;
2867 }
2868
2869 static int
2870 gen7_3DSTATE_URB_VS(struct drm_intel_decode *ctx)
2871 {
2872         return gen7_3DSTATE_URB_unit(ctx, "VS");
2873 }
2874
2875 static int
2876 gen7_3DSTATE_URB_HS(struct drm_intel_decode *ctx)
2877 {
2878         return gen7_3DSTATE_URB_unit(ctx, "HS");
2879 }
2880
2881 static int
2882 gen7_3DSTATE_URB_DS(struct drm_intel_decode *ctx)
2883 {
2884         return gen7_3DSTATE_URB_unit(ctx, "DS");
2885 }
2886
2887 static int
2888 gen7_3DSTATE_URB_GS(struct drm_intel_decode *ctx)
2889 {
2890         return gen7_3DSTATE_URB_unit(ctx, "GS");
2891 }
2892
2893 static int
2894 gen7_3DSTATE_CONSTANT(struct drm_intel_decode *ctx, const char *unit)
2895 {
2896         int rlen[4];
2897
2898         rlen[0] = (ctx->data[1] >> 0) & 0xffff;
2899         rlen[1] = (ctx->data[1] >> 16) & 0xffff;
2900         rlen[2] = (ctx->data[2] >> 0) & 0xffff;
2901         rlen[3] = (ctx->data[2] >> 16) & 0xffff;
2902
2903         instr_out(ctx, 0, "3DSTATE_CONSTANT_%s\n", unit);
2904         instr_out(ctx, 1, "len 0 = %d, len 1 = %d\n", rlen[0], rlen[1]);
2905         instr_out(ctx, 2, "len 2 = %d, len 3 = %d\n", rlen[2], rlen[3]);
2906         instr_out(ctx, 3, "pointer to constbuf 0\n");
2907         instr_out(ctx, 4, "pointer to constbuf 1\n");
2908         instr_out(ctx, 5, "pointer to constbuf 2\n");
2909         instr_out(ctx, 6, "pointer to constbuf 3\n");
2910
2911         return 7;
2912 }
2913
2914 static int
2915 gen7_3DSTATE_CONSTANT_VS(struct drm_intel_decode *ctx)
2916 {
2917         return gen7_3DSTATE_CONSTANT(ctx, "VS");
2918 }
2919
2920 static int
2921 gen7_3DSTATE_CONSTANT_GS(struct drm_intel_decode *ctx)
2922 {
2923         return gen7_3DSTATE_CONSTANT(ctx, "GS");
2924 }
2925
2926 static int
2927 gen7_3DSTATE_CONSTANT_PS(struct drm_intel_decode *ctx)
2928 {
2929         return gen7_3DSTATE_CONSTANT(ctx, "PS");
2930 }
2931
2932 static int
2933 gen7_3DSTATE_CONSTANT_DS(struct drm_intel_decode *ctx)
2934 {
2935         return gen7_3DSTATE_CONSTANT(ctx, "DS");
2936 }
2937
2938 static int
2939 gen7_3DSTATE_CONSTANT_HS(struct drm_intel_decode *ctx)
2940 {
2941         return gen7_3DSTATE_CONSTANT(ctx, "HS");
2942 }
2943
2944
2945 static int
2946 gen6_3DSTATE_WM(struct drm_intel_decode *ctx)
2947 {
2948         instr_out(ctx, 0, "3DSTATE_WM\n");
2949         instr_out(ctx, 1, "kernel start pointer 0\n");
2950         instr_out(ctx, 2,
2951                   "SPF=%d, VME=%d, Sampler Count %d, "
2952                   "Binding table count %d\n",
2953                   (ctx->data[2] >> 31) & 1,
2954                   (ctx->data[2] >> 30) & 1,
2955                   (ctx->data[2] >> 27) & 7,
2956                   (ctx->data[2] >> 18) & 0xff);
2957         instr_out(ctx, 3, "scratch offset\n");
2958         instr_out(ctx, 4,
2959                   "Depth Clear %d, Depth Resolve %d, HiZ Resolve %d, "
2960                   "Dispatch GRF start[0] %d, start[1] %d, start[2] %d\n",
2961                   (ctx->data[4] & (1 << 30)) != 0,
2962                   (ctx->data[4] & (1 << 28)) != 0,
2963                   (ctx->data[4] & (1 << 27)) != 0,
2964                   (ctx->data[4] >> 16) & 0x7f,
2965                   (ctx->data[4] >> 8) & 0x7f,
2966                   (ctx->data[4] & 0x7f));
2967         instr_out(ctx, 5,
2968                   "MaxThreads %d, PS KillPixel %d, PS computed Z %d, "
2969                   "PS use sourceZ %d, Thread Dispatch %d, PS use sourceW %d, "
2970                   "Dispatch32 %d, Dispatch16 %d, Dispatch8 %d\n",
2971                   ((ctx->data[5] >> 25) & 0x7f) + 1,
2972                   (ctx->data[5] & (1 << 22)) != 0,
2973                   (ctx->data[5] & (1 << 21)) != 0,
2974                   (ctx->data[5] & (1 << 20)) != 0,
2975                   (ctx->data[5] & (1 << 19)) != 0,
2976                   (ctx->data[5] & (1 << 8)) != 0,
2977                   (ctx->data[5] & (1 << 2)) != 0,
2978                   (ctx->data[5] & (1 << 1)) != 0,
2979                   (ctx->data[5] & (1 << 0)) != 0);
2980         instr_out(ctx, 6,
2981                   "Num SF output %d, Pos XY offset %d, ZW interp mode %d , "
2982                   "Barycentric interp mode 0x%x, Point raster rule %d, "
2983                   "Multisample mode %d, "
2984                   "Multisample Dispatch mode %d\n",
2985                   (ctx->data[6] >> 20) & 0x3f,
2986                   (ctx->data[6] >> 18) & 3,
2987                   (ctx->data[6] >> 16) & 3,
2988                   (ctx->data[6] >> 10) & 0x3f,
2989                   (ctx->data[6] & (1 << 9)) != 0,
2990                   (ctx->data[6] >> 1) & 3,
2991                   (ctx->data[6] & 1));
2992         instr_out(ctx, 7, "kernel start pointer 1\n");
2993         instr_out(ctx, 8, "kernel start pointer 2\n");
2994
2995         return 9;
2996 }
2997
2998 static int
2999 gen7_3DSTATE_WM(struct drm_intel_decode *ctx)
3000 {
3001         const char *computed_depth = "";
3002         const char *early_depth = "";
3003         const char *zw_interp = "";
3004
3005         switch ((ctx->data[1] >> 23) & 0x3) {
3006         case 0:
3007                 computed_depth = "";
3008                 break;
3009         case 1:
3010                 computed_depth = "computed depth";
3011                 break;
3012         case 2:
3013                 computed_depth = "computed depth >=";
3014                 break;
3015         case 3:
3016                 computed_depth = "computed depth <=";
3017                 break;
3018         }
3019
3020         switch ((ctx->data[1] >> 21) & 0x3) {
3021         case 0:
3022                 early_depth = "";
3023                 break;
3024         case 1:
3025                 early_depth = ", EDSC_PSEXEC";
3026                 break;
3027         case 2:
3028                 early_depth = ", EDSC_PREPS";
3029                 break;
3030         case 3:
3031                 early_depth = ", BAD EDSC";
3032                 break;
3033         }
3034
3035         switch ((ctx->data[1] >> 17) & 0x3) {
3036         case 0:
3037                 early_depth = "";
3038                 break;
3039         case 1:
3040                 early_depth = ", BAD ZW interp";
3041                 break;
3042         case 2:
3043                 early_depth = ", ZW centroid";
3044                 break;
3045         case 3:
3046                 early_depth = ", ZW sample";
3047                 break;
3048         }
3049
3050         instr_out(ctx, 0, "3DSTATE_WM\n");
3051         instr_out(ctx, 1, "(%s%s%s%s%s%s)%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
3052                   (ctx->data[1] & (1 << 11)) ? "PP " : "",
3053                   (ctx->data[1] & (1 << 12)) ? "PC " : "",
3054                   (ctx->data[1] & (1 << 13)) ? "PS " : "",
3055                   (ctx->data[1] & (1 << 14)) ? "NPP " : "",
3056                   (ctx->data[1] & (1 << 15)) ? "NPC " : "",
3057                   (ctx->data[1] & (1 << 16)) ? "NPS " : "",
3058                   (ctx->data[1] & (1 << 30)) ? ", depth clear" : "",
3059                   (ctx->data[1] & (1 << 29)) ? "" : ", disabled",
3060                   (ctx->data[1] & (1 << 28)) ? ", depth resolve" : "",
3061                   (ctx->data[1] & (1 << 27)) ? ", hiz resolve" : "",
3062                   (ctx->data[1] & (1 << 25)) ? ", kill" : "",
3063                   computed_depth,
3064                   early_depth,
3065                   zw_interp,
3066                   (ctx->data[1] & (1 << 20)) ? ", source depth" : "",
3067                   (ctx->data[1] & (1 << 19)) ? ", source W" : "",
3068                   (ctx->data[1] & (1 << 10)) ? ", coverage" : "",
3069                   (ctx->data[1] & (1 << 4)) ? ", poly stipple" : "",
3070                   (ctx->data[1] & (1 << 3)) ? ", line stipple" : "",
3071                   (ctx->data[1] & (1 << 2)) ? ", point UL" : ", point UR"
3072                   );
3073         instr_out(ctx, 2, "MS\n");
3074
3075         return 3;
3076 }
3077
3078 static int
3079 gen4_3DPRIMITIVE(struct drm_intel_decode *ctx)
3080 {
3081         instr_out(ctx, 0,
3082                   "3DPRIMITIVE: %s %s\n",
3083                   get_965_prim_type((ctx->data[0] >> 10) & 0x1f),
3084                   (ctx->data[0] & (1 << 15)) ? "random" : "sequential");
3085         instr_out(ctx, 1, "vertex count\n");
3086         instr_out(ctx, 2, "start vertex\n");
3087         instr_out(ctx, 3, "instance count\n");
3088         instr_out(ctx, 4, "start instance\n");
3089         instr_out(ctx, 5, "index bias\n");
3090
3091         return 6;
3092 }
3093
3094 static int
3095 gen7_3DPRIMITIVE(struct drm_intel_decode *ctx)
3096 {
3097         bool indirect = !!(ctx->data[0] & (1 << 10));
3098
3099         instr_out(ctx, 0,
3100                   "3DPRIMITIVE: %s%s\n",
3101                   indirect ? " indirect" : "",
3102                   (ctx->data[0] & (1 << 8)) ? " predicated" : "");
3103         instr_out(ctx, 1, "%s %s\n",
3104                   get_965_prim_type(ctx->data[1] & 0x3f),
3105                   (ctx->data[1] & (1 << 8)) ? "random" : "sequential");
3106         instr_out(ctx, 2, indirect ? "ignored" : "vertex count\n");
3107         instr_out(ctx, 3, indirect ? "ignored" : "start vertex\n");
3108         instr_out(ctx, 4, indirect ? "ignored" : "instance count\n");
3109         instr_out(ctx, 5, indirect ? "ignored" : "start instance\n");
3110         instr_out(ctx, 6, indirect ? "ignored" : "index bias\n");
3111
3112         return 7;
3113 }
3114
3115 static int
3116 decode_3d_965(struct drm_intel_decode *ctx)
3117 {
3118         uint32_t opcode;
3119         unsigned int len;
3120         unsigned int i, j, sba_len;
3121         const char *desc1 = NULL;
3122         uint32_t *data = ctx->data;
3123         uint32_t devid = ctx->devid;
3124
3125         struct {
3126                 uint32_t opcode;
3127                 uint32_t len_mask;
3128                 int unsigned min_len;
3129                 int unsigned max_len;
3130                 const char *name;
3131                 int gen;
3132                 int (*func)(struct drm_intel_decode *ctx);
3133         } opcodes_3d[] = {
3134                 { 0x6000, 0x00ff, 3, 3, "URB_FENCE" },
3135                 { 0x6001, 0xffff, 2, 2, "CS_URB_STATE" },
3136                 { 0x6002, 0x00ff, 2, 2, "CONSTANT_BUFFER" },
3137                 { 0x6101, 0xffff, 6, 10, "STATE_BASE_ADDRESS" },
3138                 { 0x6102, 0xffff, 2, 2, "STATE_SIP" },
3139                 { 0x6104, 0xffff, 1, 1, "3DSTATE_PIPELINE_SELECT" },
3140                 { 0x680b, 0xffff, 1, 1, "3DSTATE_VF_STATISTICS" },
3141                 { 0x6904, 0xffff, 1, 1, "3DSTATE_PIPELINE_SELECT" },
3142                 { 0x7800, 0xffff, 7, 7, "3DSTATE_PIPELINED_POINTERS" },
3143                 { 0x7801, 0x00ff, 4, 6, "3DSTATE_BINDING_TABLE_POINTERS" },
3144                 { 0x7802, 0x00ff, 4, 4, "3DSTATE_SAMPLER_STATE_POINTERS" },
3145                 { 0x7805, 0x00ff, 7, 7, "3DSTATE_DEPTH_BUFFER", 7 },
3146                 { 0x7805, 0x00ff, 3, 3, "3DSTATE_URB" },
3147                 { 0x7804, 0x00ff, 3, 3, "3DSTATE_CLEAR_PARAMS" },
3148                 { 0x7806, 0x00ff, 3, 3, "3DSTATE_STENCIL_BUFFER" },
3149                 { 0x7807, 0x00ff, 4, 4, "3DSTATE_HIER_DEPTH_BUFFER", 6 },
3150                 { 0x7807, 0x00ff, 3, 3, "3DSTATE_HIER_DEPTH_BUFFER", 7, gen7_3DSTATE_HIER_DEPTH_BUFFER },
3151                 { 0x7808, 0x00ff, 5, 257, "3DSTATE_VERTEX_BUFFERS" },
3152                 { 0x7809, 0x00ff, 3, 256, "3DSTATE_VERTEX_ELEMENTS" },
3153                 { 0x780a, 0x00ff, 3, 3, "3DSTATE_INDEX_BUFFER" },
3154                 { 0x780b, 0xffff, 1, 1, "3DSTATE_VF_STATISTICS" },
3155                 { 0x780d, 0x00ff, 4, 4, "3DSTATE_VIEWPORT_STATE_POINTERS" },
3156                 { 0x780e, 0xffff, 4, 4, NULL, 6, gen6_3DSTATE_CC_STATE_POINTERS },
3157                 { 0x780e, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_CC_STATE_POINTERS },
3158                 { 0x780f, 0x00ff, 2, 2, "3DSTATE_SCISSOR_POINTERS" },
3159                 { 0x7810, 0x00ff, 6, 6, "3DSTATE_VS" },
3160                 { 0x7811, 0x00ff, 7, 7, "3DSTATE_GS" },
3161                 { 0x7812, 0x00ff, 4, 4, "3DSTATE_CLIP" },
3162                 { 0x7813, 0x00ff, 20, 20, "3DSTATE_SF", 6 },
3163                 { 0x7813, 0x00ff, 7, 7, "3DSTATE_SF", 7 },
3164                 { 0x7814, 0x00ff, 3, 3, "3DSTATE_WM", 7, gen7_3DSTATE_WM },
3165                 { 0x7814, 0x00ff, 9, 9, "3DSTATE_WM", 6, gen6_3DSTATE_WM },
3166                 { 0x7815, 0x00ff, 5, 5, "3DSTATE_CONSTANT_VS_STATE", 6 },
3167                 { 0x7815, 0x00ff, 7, 7, "3DSTATE_CONSTANT_VS", 7, gen7_3DSTATE_CONSTANT_VS },
3168                 { 0x7816, 0x00ff, 5, 5, "3DSTATE_CONSTANT_GS_STATE", 6 },
3169                 { 0x7816, 0x00ff, 7, 7, "3DSTATE_CONSTANT_GS", 7, gen7_3DSTATE_CONSTANT_GS },
3170                 { 0x7817, 0x00ff, 5, 5, "3DSTATE_CONSTANT_PS_STATE", 6 },
3171                 { 0x7817, 0x00ff, 7, 7, "3DSTATE_CONSTANT_PS", 7, gen7_3DSTATE_CONSTANT_PS },
3172                 { 0x7818, 0xffff, 2, 2, "3DSTATE_SAMPLE_MASK" },
3173                 { 0x7819, 0x00ff, 7, 7, "3DSTATE_CONSTANT_HS", 7, gen7_3DSTATE_CONSTANT_HS },
3174                 { 0x781a, 0x00ff, 7, 7, "3DSTATE_CONSTANT_DS", 7, gen7_3DSTATE_CONSTANT_DS },
3175                 { 0x781b, 0x00ff, 7, 7, "3DSTATE_HS" },
3176                 { 0x781c, 0x00ff, 4, 4, "3DSTATE_TE" },
3177                 { 0x781d, 0x00ff, 6, 6, "3DSTATE_DS" },
3178                 { 0x781e, 0x00ff, 3, 3, "3DSTATE_STREAMOUT" },
3179                 { 0x781f, 0x00ff, 14, 14, "3DSTATE_SBE" },
3180                 { 0x7820, 0x00ff, 8, 8, "3DSTATE_PS" },
3181                 { 0x7821, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP },
3182                 { 0x7823, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_VIEWPORT_STATE_POINTERS_CC },
3183                 { 0x7824, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_BLEND_STATE_POINTERS },
3184                 { 0x7825, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_DEPTH_STENCIL_STATE_POINTERS },
3185                 { 0x7826, 0x00ff, 2, 2, "3DSTATE_BINDING_TABLE_POINTERS_VS" },
3186                 { 0x7827, 0x00ff, 2, 2, "3DSTATE_BINDING_TABLE_POINTERS_HS" },
3187                 { 0x7828, 0x00ff, 2, 2, "3DSTATE_BINDING_TABLE_POINTERS_DS" },
3188                 { 0x7829, 0x00ff, 2, 2, "3DSTATE_BINDING_TABLE_POINTERS_GS" },
3189                 { 0x782a, 0x00ff, 2, 2, "3DSTATE_BINDING_TABLE_POINTERS_PS" },
3190                 { 0x782b, 0x00ff, 2, 2, "3DSTATE_SAMPLER_STATE_POINTERS_VS" },
3191                 { 0x782c, 0x00ff, 2, 2, "3DSTATE_SAMPLER_STATE_POINTERS_HS" },
3192                 { 0x782d, 0x00ff, 2, 2, "3DSTATE_SAMPLER_STATE_POINTERS_DS" },
3193                 { 0x782e, 0x00ff, 2, 2, "3DSTATE_SAMPLER_STATE_POINTERS_GS" },
3194                 { 0x782f, 0x00ff, 2, 2, "3DSTATE_SAMPLER_STATE_POINTERS_PS" },
3195                 { 0x7830, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_URB_VS },
3196                 { 0x7831, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_URB_HS },
3197                 { 0x7832, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_URB_DS },
3198                 { 0x7833, 0x00ff, 2, 2, NULL, 7, gen7_3DSTATE_URB_GS },
3199                 { 0x7900, 0xffff, 4, 4, "3DSTATE_DRAWING_RECTANGLE" },
3200                 { 0x7901, 0xffff, 5, 5, "3DSTATE_CONSTANT_COLOR" },
3201                 { 0x7905, 0xffff, 5, 7, "3DSTATE_DEPTH_BUFFER" },
3202                 { 0x7906, 0xffff, 2, 2, "3DSTATE_POLY_STIPPLE_OFFSET" },
3203                 { 0x7907, 0xffff, 33, 33, "3DSTATE_POLY_STIPPLE_PATTERN" },
3204                 { 0x7908, 0xffff, 3, 3, "3DSTATE_LINE_STIPPLE" },
3205                 { 0x7909, 0xffff, 2, 2, "3DSTATE_GLOBAL_DEPTH_OFFSET_CLAMP" },
3206                 { 0x7909, 0xffff, 2, 2, "3DSTATE_CLEAR_PARAMS" },
3207                 { 0x790a, 0xffff, 3, 3, "3DSTATE_AA_LINE_PARAMETERS" },
3208                 { 0x790b, 0xffff, 4, 4, "3DSTATE_GS_SVB_INDEX" },
3209                 { 0x790d, 0xffff, 3, 3, "3DSTATE_MULTISAMPLE", 6 },
3210                 { 0x790d, 0xffff, 4, 4, "3DSTATE_MULTISAMPLE", 7 },
3211                 { 0x7910, 0xffff, 2, 2, "3DSTATE_CLEAR_PARAMS" },
3212                 { 0x7912, 0x00ff, 2, 2, "3DSTATE_PUSH_CONSTANT_ALLOC_VS" },
3213                 { 0x7913, 0x00ff, 2, 2, "3DSTATE_PUSH_CONSTANT_ALLOC_HS" },
3214                 { 0x7914, 0x00ff, 2, 2, "3DSTATE_PUSH_CONSTANT_ALLOC_DS" },
3215                 { 0x7915, 0x00ff, 2, 2, "3DSTATE_PUSH_CONSTANT_ALLOC_GS" },
3216                 { 0x7916, 0x00ff, 2, 2, "3DSTATE_PUSH_CONSTANT_ALLOC_PS" },
3217                 { 0x7917, 0x00ff, 2, 2+128*2, "3DSTATE_SO_DECL_LIST" },
3218                 { 0x7918, 0x00ff, 4, 4, "3DSTATE_SO_BUFFER" },
3219                 { 0x7a00, 0x00ff, 4, 6, "PIPE_CONTROL" },
3220                 { 0x7b00, 0x00ff, 7, 7, NULL, 7, gen7_3DPRIMITIVE },
3221                 { 0x7b00, 0x00ff, 6, 6, NULL, 0, gen4_3DPRIMITIVE },
3222         }, *opcode_3d = NULL;
3223
3224         opcode = (data[0] & 0xffff0000) >> 16;
3225
3226         for (i = 0; i < ARRAY_SIZE(opcodes_3d); i++) {
3227                 if (opcode != opcodes_3d[i].opcode)
3228                         continue;
3229
3230                 /* If it's marked as not our gen, skip. */
3231                 if (opcodes_3d[i].gen && opcodes_3d[i].gen != ctx->gen)
3232                         continue;
3233
3234                 opcode_3d = &opcodes_3d[i];
3235                 break;
3236         }
3237
3238         if (opcode_3d) {
3239                 if (opcode_3d->max_len == 1)
3240                         len = 1;
3241                 else
3242                         len = (data[0] & opcode_3d->len_mask) + 2;
3243
3244                 if (len < opcode_3d->min_len ||
3245                     len > opcode_3d->max_len) {
3246                         fprintf(out, "Bad length %d in %s, expected %d-%d\n",
3247                                 len, opcode_3d->name,
3248                                 opcode_3d->min_len, opcode_3d->max_len);
3249                 }
3250         } else {
3251                 len = (data[0] & 0x0000ffff) + 2;
3252         }
3253
3254         switch (opcode) {
3255         case 0x6000:
3256                 return i965_decode_urb_fence(ctx, len);
3257         case 0x6001:
3258                 instr_out(ctx, 0, "CS_URB_STATE\n");
3259                 instr_out(ctx, 1,
3260                           "entry_size: %d [%d bytes], n_entries: %d\n",
3261                           (data[1] >> 4) & 0x1f,
3262                           (((data[1] >> 4) & 0x1f) + 1) * 64, data[1] & 0x7);
3263                 return len;
3264         case 0x6002:
3265                 instr_out(ctx, 0, "CONSTANT_BUFFER: %s\n",
3266                           (data[0] >> 8) & 1 ? "valid" : "invalid");
3267                 instr_out(ctx, 1,
3268                           "offset: 0x%08x, length: %d bytes\n", data[1] & ~0x3f,
3269                           ((data[1] & 0x3f) + 1) * 64);
3270                 return len;
3271         case 0x6101:
3272                 i = 0;
3273                 instr_out(ctx, 0, "STATE_BASE_ADDRESS\n");
3274                 i++;
3275
3276                 if (IS_GEN6(devid) || IS_GEN7(devid))
3277                         sba_len = 10;
3278                 else if (IS_GEN5(devid))
3279                         sba_len = 8;
3280                 else
3281                         sba_len = 6;
3282                 if (len != sba_len)
3283                         fprintf(out, "Bad count in STATE_BASE_ADDRESS\n");
3284
3285                 state_base_out(ctx, i++, "general");
3286                 state_base_out(ctx, i++, "surface");
3287                 if (IS_GEN6(devid) || IS_GEN7(devid))
3288                         state_base_out(ctx, i++, "dynamic");
3289                 state_base_out(ctx, i++, "indirect");
3290                 if (IS_GEN5(devid) || IS_GEN6(devid) || IS_GEN7(devid))
3291                         state_base_out(ctx, i++, "instruction");
3292
3293                 state_max_out(ctx, i++, "general");
3294                 if (IS_GEN6(devid) || IS_GEN7(devid))
3295                         state_max_out(ctx, i++, "dynamic");
3296                 state_max_out(ctx, i++, "indirect");
3297                 if (IS_GEN5(devid) || IS_GEN6(devid) || IS_GEN7(devid))
3298                         state_max_out(ctx, i++, "instruction");
3299
3300                 return len;
3301         case 0x7800:
3302                 instr_out(ctx, 0, "3DSTATE_PIPELINED_POINTERS\n");
3303                 instr_out(ctx, 1, "VS state\n");
3304                 instr_out(ctx, 2, "GS state\n");
3305                 instr_out(ctx, 3, "Clip state\n");
3306                 instr_out(ctx, 4, "SF state\n");
3307                 instr_out(ctx, 5, "WM state\n");
3308                 instr_out(ctx, 6, "CC state\n");
3309                 return len;
3310         case 0x7801:
3311                 if (len != 6 && len != 4)
3312                         fprintf(out,
3313                                 "Bad count in 3DSTATE_BINDING_TABLE_POINTERS\n");
3314                 if (len == 6) {
3315                         instr_out(ctx, 0,
3316                                   "3DSTATE_BINDING_TABLE_POINTERS\n");
3317                         instr_out(ctx, 1, "VS binding table\n");
3318                         instr_out(ctx, 2, "GS binding table\n");
3319                         instr_out(ctx, 3, "Clip binding table\n");
3320                         instr_out(ctx, 4, "SF binding table\n");
3321                         instr_out(ctx, 5, "WM binding table\n");
3322                 } else {
3323                         instr_out(ctx, 0,
3324                                   "3DSTATE_BINDING_TABLE_POINTERS: VS mod %d, "
3325                                   "GS mod %d, PS mod %d\n",
3326                                   (data[0] & (1 << 8)) != 0,
3327                                   (data[0] & (1 << 9)) != 0,
3328                                   (data[0] & (1 << 12)) != 0);
3329                         instr_out(ctx, 1, "VS binding table\n");
3330                         instr_out(ctx, 2, "GS binding table\n");
3331                         instr_out(ctx, 3, "WM binding table\n");
3332                 }
3333
3334                 return len;
3335         case 0x7802:
3336                 instr_out(ctx, 0,
3337                           "3DSTATE_SAMPLER_STATE_POINTERS: VS mod %d, "
3338                           "GS mod %d, PS mod %d\n", (data[0] & (1 << 8)) != 0,
3339                           (data[0] & (1 << 9)) != 0,
3340                           (data[0] & (1 << 12)) != 0);
3341                 instr_out(ctx, 1, "VS sampler state\n");
3342                 instr_out(ctx, 2, "GS sampler state\n");
3343                 instr_out(ctx, 3, "WM sampler state\n");
3344                 return len;
3345         case 0x7805:
3346                 /* Actually 3DSTATE_DEPTH_BUFFER on gen7. */
3347                 if (ctx->gen == 7)
3348                         break;
3349
3350                 instr_out(ctx, 0, "3DSTATE_URB\n");
3351                 instr_out(ctx, 1,
3352                           "VS entries %d, alloc size %d (1024bit row)\n",
3353                           data[1] & 0xffff, ((data[1] >> 16) & 0x07f) + 1);
3354                 instr_out(ctx, 2,
3355                           "GS entries %d, alloc size %d (1024bit row)\n",
3356                           (data[2] >> 8) & 0x3ff, (data[2] & 7) + 1);
3357                 return len;
3358
3359         case 0x7808:
3360                 if ((len - 1) % 4 != 0)
3361                         fprintf(out, "Bad count in 3DSTATE_VERTEX_BUFFERS\n");
3362                 instr_out(ctx, 0, "3DSTATE_VERTEX_BUFFERS\n");
3363
3364                 for (i = 1; i < len;) {
3365                         int idx, access;
3366                         if (IS_GEN6(devid)) {
3367                                 idx = 26;
3368                                 access = 20;
3369                         } else {
3370                                 idx = 27;
3371                                 access = 26;
3372                         }
3373                         instr_out(ctx, i,
3374                                   "buffer %d: %s, pitch %db\n", data[i] >> idx,
3375                                   data[i] & (1 << access) ? "random" :
3376                                   "sequential", data[i] & 0x07ff);
3377                         i++;
3378                         instr_out(ctx, i++, "buffer address\n");
3379                         instr_out(ctx, i++, "max index\n");
3380                         instr_out(ctx, i++, "mbz\n");
3381                 }
3382                 return len;
3383
3384         case 0x7809:
3385                 if ((len + 1) % 2 != 0)
3386                         fprintf(out, "Bad count in 3DSTATE_VERTEX_ELEMENTS\n");
3387                 instr_out(ctx, 0, "3DSTATE_VERTEX_ELEMENTS\n");
3388
3389                 for (i = 1; i < len;) {
3390                         instr_out(ctx, i,
3391                                   "buffer %d: %svalid, type 0x%04x, "
3392                                   "src offset 0x%04x bytes\n",
3393                                   data[i] >> (IS_GEN6(devid) ? 26 : 27),
3394                                   data[i] & (1 << (IS_GEN6(devid) ? 25 : 26)) ?
3395                                   "" : "in", (data[i] >> 16) & 0x1ff,
3396                                   data[i] & 0x07ff);
3397                         i++;
3398                         instr_out(ctx, i, "(%s, %s, %s, %s), "
3399                                   "dst offset 0x%02x bytes\n",
3400                                   get_965_element_component(data[i], 0),
3401                                   get_965_element_component(data[i], 1),
3402                                   get_965_element_component(data[i], 2),
3403                                   get_965_element_component(data[i], 3),
3404                                   (data[i] & 0xff) * 4);
3405                         i++;
3406                 }
3407                 return len;
3408
3409         case 0x780d:
3410                 instr_out(ctx, 0,
3411                           "3DSTATE_VIEWPORT_STATE_POINTERS\n");
3412                 instr_out(ctx, 1, "clip\n");
3413                 instr_out(ctx, 2, "sf\n");
3414                 instr_out(ctx, 3, "cc\n");
3415                 return len;
3416
3417         case 0x780a:
3418                 instr_out(ctx, 0, "3DSTATE_INDEX_BUFFER\n");
3419                 instr_out(ctx, 1, "beginning buffer address\n");
3420                 instr_out(ctx, 2, "ending buffer address\n");
3421                 return len;
3422
3423         case 0x780f:
3424                 instr_out(ctx, 0, "3DSTATE_SCISSOR_POINTERS\n");
3425                 instr_out(ctx, 1, "scissor rect offset\n");
3426                 return len;
3427
3428         case 0x7810:
3429                 instr_out(ctx, 0, "3DSTATE_VS\n");
3430                 instr_out(ctx, 1, "kernel pointer\n");
3431                 instr_out(ctx, 2,
3432                           "SPF=%d, VME=%d, Sampler Count %d, "
3433                           "Binding table count %d\n", (data[2] >> 31) & 1,
3434                           (data[2] >> 30) & 1, (data[2] >> 27) & 7,
3435                           (data[2] >> 18) & 0xff);
3436                 instr_out(ctx, 3, "scratch offset\n");
3437                 instr_out(ctx, 4,
3438                           "Dispatch GRF start %d, VUE read length %d, "
3439                           "VUE read offset %d\n", (data[4] >> 20) & 0x1f,
3440                           (data[4] >> 11) & 0x3f, (data[4] >> 4) & 0x3f);
3441                 instr_out(ctx, 5,
3442                           "Max Threads %d, Vertex Cache %sable, "
3443                           "VS func %sable\n", ((data[5] >> 25) & 0x7f) + 1,
3444                           (data[5] & (1 << 1)) != 0 ? "dis" : "en",
3445                           (data[5] & 1) != 0 ? "en" : "dis");
3446                 return len;
3447
3448         case 0x7811:
3449                 instr_out(ctx, 0, "3DSTATE_GS\n");
3450                 instr_out(ctx, 1, "kernel pointer\n");
3451                 instr_out(ctx, 2,
3452                           "SPF=%d, VME=%d, Sampler Count %d, "
3453                           "Binding table count %d\n", (data[2] >> 31) & 1,
3454                           (data[2] >> 30) & 1, (data[2] >> 27) & 7,
3455                           (data[2] >> 18) & 0xff);
3456                 instr_out(ctx, 3, "scratch offset\n");
3457                 instr_out(ctx, 4,
3458                           "Dispatch GRF start %d, VUE read length %d, "
3459                           "VUE read offset %d\n", (data[4] & 0xf),
3460                           (data[4] >> 11) & 0x3f, (data[4] >> 4) & 0x3f);
3461                 instr_out(ctx, 5,
3462                           "Max Threads %d, Rendering %sable\n",
3463                           ((data[5] >> 25) & 0x7f) + 1,
3464                           (data[5] & (1 << 8)) != 0 ? "en" : "dis");
3465                 instr_out(ctx, 6,
3466                           "Reorder %sable, Discard Adjaceny %sable, "
3467                           "GS %sable\n",
3468                           (data[6] & (1 << 30)) != 0 ? "en" : "dis",
3469                           (data[6] & (1 << 29)) != 0 ? "en" : "dis",
3470                           (data[6] & (1 << 15)) != 0 ? "en" : "dis");
3471                 return len;
3472
3473         case 0x7812:
3474                 instr_out(ctx, 0, "3DSTATE_CLIP\n");
3475                 instr_out(ctx, 1,
3476                           "UserClip distance cull test mask 0x%x\n",
3477                           data[1] & 0xff);
3478                 instr_out(ctx, 2,
3479                           "Clip %sable, API mode %s, Viewport XY test %sable, "
3480                           "Viewport Z test %sable, Guardband test %sable, Clip mode %d, "
3481                           "Perspective Divide %sable, Non-Perspective Barycentric %sable, "
3482                           "Tri Provoking %d, Line Provoking %d, Trifan Provoking %d\n",
3483                           (data[2] & (1 << 31)) != 0 ? "en" : "dis",
3484                           (data[2] & (1 << 30)) != 0 ? "D3D" : "OGL",
3485                           (data[2] & (1 << 28)) != 0 ? "en" : "dis",
3486                           (data[2] & (1 << 27)) != 0 ? "en" : "dis",
3487                           (data[2] & (1 << 26)) != 0 ? "en" : "dis",
3488                           (data[2] >> 13) & 7,
3489                           (data[2] & (1 << 9)) != 0 ? "dis" : "en",
3490                           (data[2] & (1 << 8)) != 0 ? "en" : "dis",
3491                           (data[2] >> 4) & 3, (data[2] >> 2) & 3,
3492                           (data[2] & 3));
3493                 instr_out(ctx, 3,
3494                           "Min PointWidth %d, Max PointWidth %d, "
3495                           "Force Zero RTAIndex %sable, Max VPIndex %d\n",
3496                           (data[3] >> 17) & 0x7ff, (data[3] >> 6) & 0x7ff,
3497                           (data[3] & (1 << 5)) != 0 ? "en" : "dis",
3498                           (data[3] & 0xf));
3499                 return len;
3500
3501         case 0x7813:
3502                 if (ctx->gen == 7)
3503                         break;
3504
3505                 instr_out(ctx, 0, "3DSTATE_SF\n");
3506                 instr_out(ctx, 1,
3507                           "Attrib Out %d, Attrib Swizzle %sable, VUE read length %d, "
3508                           "VUE read offset %d\n", (data[1] >> 22) & 0x3f,
3509                           (data[1] & (1 << 21)) != 0 ? "en" : "dis",
3510                           (data[1] >> 11) & 0x1f, (data[1] >> 4) & 0x3f);
3511                 instr_out(ctx, 2,
3512                           "Legacy Global DepthBias %sable, FrontFace fill %d, BF fill %d, "
3513                           "VP transform %sable, FrontWinding_%s\n",
3514                           (data[2] & (1 << 11)) != 0 ? "en" : "dis",
3515                           (data[2] >> 5) & 3, (data[2] >> 3) & 3,
3516                           (data[2] & (1 << 1)) != 0 ? "en" : "dis",
3517                           (data[2] & 1) != 0 ? "CCW" : "CW");
3518                 instr_out(ctx, 3,
3519                           "AA %sable, CullMode %d, Scissor %sable, Multisample m ode %d\n",
3520                           (data[3] & (1 << 31)) != 0 ? "en" : "dis",
3521                           (data[3] >> 29) & 3,
3522                           (data[3] & (1 << 11)) != 0 ? "en" : "dis",
3523                           (data[3] >> 8) & 3);
3524                 instr_out(ctx, 4,
3525                           "Last Pixel %sable, SubPixel Precision %d, Use PixelWidth %d\n",
3526                           (data[4] & (1 << 31)) != 0 ? "en" : "dis",
3527                           (data[4] & (1 << 12)) != 0 ? 4 : 8,
3528                           (data[4] & (1 << 11)) != 0);
3529                 instr_out(ctx, 5,
3530                           "Global Depth Offset Constant %f\n",
3531                           *(float *)(&data[5]));
3532                 instr_out(ctx, 6, "Global Depth Offset Scale %f\n",
3533                           *(float *)(&data[6]));
3534                 instr_out(ctx, 7, "Global Depth Offset Clamp %f\n",
3535                           *(float *)(&data[7]));
3536
3537                 for (i = 0, j = 0; i < 8; i++, j += 2)
3538                         instr_out(ctx, i + 8,
3539                                   "Attrib %d (Override %s%s%s%s, Const Source %d, Swizzle Select %d, "
3540                                   "Source %d); Attrib %d (Override %s%s%s%s, Const Source %d, Swizzle Select %d, Source %d)\n",
3541                                   j + 1,
3542                                   (data[8 + i] & (1 << 31)) != 0 ? "W" : "",
3543                                   (data[8 + i] & (1 << 30)) != 0 ? "Z" : "",
3544                                   (data[8 + i] & (1 << 29)) != 0 ? "Y" : "",
3545                                   (data[8 + i] & (1 << 28)) != 0 ? "X" : "",
3546                                   (data[8 + i] >> 25) & 3,
3547                                   (data[8 + i] >> 22) & 3,
3548                                   (data[8 + i] >> 16) & 0x1f, j,
3549                                   (data[8 + i] & (1 << 15)) != 0 ? "W" : "",
3550                                   (data[8 + i] & (1 << 14)) != 0 ? "Z" : "",
3551                                   (data[8 + i] & (1 << 13)) != 0 ? "Y" : "",
3552                                   (data[8 + i] & (1 << 12)) != 0 ? "X" : "",
3553                                   (data[8 + i] >> 9) & 3,
3554                                   (data[8 + i] >> 6) & 3, (data[8 + i] & 0x1f));
3555                 instr_out(ctx, 16,
3556                           "Point Sprite TexCoord Enable\n");
3557                 instr_out(ctx, 17, "Const Interp Enable\n");
3558                 instr_out(ctx, 18,
3559                           "Attrib 7-0 WrapShortest Enable\n");
3560                 instr_out(ctx, 19,
3561                           "Attrib 15-8 WrapShortest Enable\n");
3562
3563                 return len;
3564
3565         case 0x7900:
3566                 instr_out(ctx, 0, "3DSTATE_DRAWING_RECTANGLE\n");
3567                 instr_out(ctx, 1, "top left: %d,%d\n",
3568                           data[1] & 0xffff, (data[1] >> 16) & 0xffff);
3569                 instr_out(ctx, 2, "bottom right: %d,%d\n",
3570                           data[2] & 0xffff, (data[2] >> 16) & 0xffff);
3571                 instr_out(ctx, 3, "origin: %d,%d\n",
3572                           (int)data[3] & 0xffff, ((int)data[3] >> 16) & 0xffff);
3573
3574                 return len;
3575
3576         case 0x7905:
3577                 instr_out(ctx, 0, "3DSTATE_DEPTH_BUFFER\n");
3578                 if (IS_GEN5(devid) || IS_GEN6(devid))
3579                         instr_out(ctx, 1,
3580                                   "%s, %s, pitch = %d bytes, %stiled, HiZ %d, Seperate Stencil %d\n",
3581                                   get_965_surfacetype(data[1] >> 29),
3582                                   get_965_depthformat((data[1] >> 18) & 0x7),
3583                                   (data[1] & 0x0001ffff) + 1,
3584                                   data[1] & (1 << 27) ? "" : "not ",
3585                                   (data[1] & (1 << 22)) != 0,
3586                                   (data[1] & (1 << 21)) != 0);
3587                 else
3588                         instr_out(ctx, 1,
3589                                   "%s, %s, pitch = %d bytes, %stiled\n",
3590                                   get_965_surfacetype(data[1] >> 29),
3591                                   get_965_depthformat((data[1] >> 18) & 0x7),
3592                                   (data[1] & 0x0001ffff) + 1,
3593                                   data[1] & (1 << 27) ? "" : "not ");
3594                 instr_out(ctx, 2, "depth offset\n");
3595                 instr_out(ctx, 3, "%dx%d\n",
3596                           ((data[3] & 0x0007ffc0) >> 6) + 1,
3597                           ((data[3] & 0xfff80000) >> 19) + 1);
3598                 instr_out(ctx, 4, "volume depth\n");
3599                 if (len >= 6)
3600                         instr_out(ctx, 5, "\n");
3601                 if (len >= 7) {
3602                         if (IS_GEN6(devid))
3603                                 instr_out(ctx, 6, "\n");
3604                         else
3605                                 instr_out(ctx, 6,
3606                                           "render target view extent\n");
3607                 }
3608
3609                 return len;
3610
3611         case 0x7a00:
3612                 if (IS_GEN6(devid) || IS_GEN7(devid)) {
3613                         unsigned int i;
3614                         if (len != 4 && len != 5)
3615                                 fprintf(out, "Bad count in PIPE_CONTROL\n");
3616
3617                         switch ((data[1] >> 14) & 0x3) {
3618                         case 0:
3619                                 desc1 = "no write";
3620                                 break;
3621                         case 1:
3622                                 desc1 = "qword write";
3623                                 break;
3624                         case 2:
3625                                 desc1 = "PS_DEPTH_COUNT write";
3626                                 break;
3627                         case 3:
3628                                 desc1 = "TIMESTAMP write";
3629                                 break;
3630                         }
3631                         instr_out(ctx, 0, "PIPE_CONTROL\n");
3632                         instr_out(ctx, 1,
3633                                   "%s, %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
3634                                   desc1,
3635                                   data[1] & (1 << 20) ? "cs stall, " : "",
3636                                   data[1] & (1 << 19) ?
3637                                   "global snapshot count reset, " : "",
3638                                   data[1] & (1 << 18) ? "tlb invalidate, " : "",
3639                                   data[1] & (1 << 17) ? "gfdt flush, " : "",
3640                                   data[1] & (1 << 17) ? "media state clear, " :
3641                                   "",
3642                                   data[1] & (1 << 13) ? "depth stall, " : "",
3643                                   data[1] & (1 << 12) ?
3644                                   "render target cache flush, " : "",
3645                                   data[1] & (1 << 11) ?
3646                                   "instruction cache invalidate, " : "",
3647                                   data[1] & (1 << 10) ?
3648                                   "texture cache invalidate, " : "",
3649                                   data[1] & (1 << 9) ?
3650                                   "indirect state invalidate, " : "",
3651                                   data[1] & (1 << 8) ? "notify irq, " : "",
3652                                   data[1] & (1 << 7) ? "PIPE_CONTROL flush, " :
3653                                   "",
3654                                   data[1] & (1 << 6) ? "protect mem app_id, " :
3655                                   "", data[1] & (1 << 5) ? "DC flush, " : "",
3656                                   data[1] & (1 << 4) ? "vf fetch invalidate, " :
3657                                   "",
3658                                   data[1] & (1 << 3) ?
3659                                   "constant cache invalidate, " : "",
3660                                   data[1] & (1 << 2) ?
3661                                   "state cache invalidate, " : "",
3662                                   data[1] & (1 << 1) ? "stall at scoreboard, " :
3663                                   "",
3664                                   data[1] & (1 << 0) ? "depth cache flush, " :
3665                                   "");
3666                         if (len == 5) {
3667                                 instr_out(ctx, 2,
3668                                           "destination address\n");
3669                                 instr_out(ctx, 3,
3670                                           "immediate dword low\n");
3671                                 instr_out(ctx, 4,
3672                                           "immediate dword high\n");
3673                         } else {
3674                                 for (i = 2; i < len; i++) {
3675                                         instr_out(ctx, i, "\n");
3676                                 }
3677                         }
3678                         return len;
3679                 } else {
3680                         if (len != 4)
3681                                 fprintf(out, "Bad count in PIPE_CONTROL\n");
3682
3683                         switch ((data[0] >> 14) & 0x3) {
3684                         case 0:
3685                                 desc1 = "no write";
3686                                 break;
3687                         case 1:
3688                                 desc1 = "qword write";
3689                                 break;
3690                         case 2:
3691                                 desc1 = "PS_DEPTH_COUNT write";
3692                                 break;
3693                         case 3:
3694                                 desc1 = "TIMESTAMP write";
3695                                 break;
3696                         }
3697                         instr_out(ctx, 0,
3698                                   "PIPE_CONTROL: %s, %sdepth stall, %sRC write flush, "
3699                                   "%sinst flush\n",
3700                                   desc1,
3701                                   data[0] & (1 << 13) ? "" : "no ",
3702                                   data[0] & (1 << 12) ? "" : "no ",
3703                                   data[0] & (1 << 11) ? "" : "no ");
3704                         instr_out(ctx, 1, "destination address\n");
3705                         instr_out(ctx, 2, "immediate dword low\n");
3706                         instr_out(ctx, 3, "immediate dword high\n");
3707                         return len;
3708                 }
3709         }
3710
3711         if (opcode_3d) {
3712                 if (opcode_3d->func) {
3713                         return opcode_3d->func(ctx);
3714                 } else {
3715                         unsigned int i;
3716
3717                         instr_out(ctx, 0, "%s\n", opcode_3d->name);
3718
3719                         for (i = 1; i < len; i++) {
3720                                 instr_out(ctx, i, "dword %d\n", i);
3721                         }
3722                         return len;
3723                 }
3724         }
3725
3726         instr_out(ctx, 0, "3D UNKNOWN: 3d_965 opcode = 0x%x\n",
3727                   opcode);
3728         return 1;
3729 }
3730
3731 static int
3732 decode_3d_i830(struct drm_intel_decode *ctx)
3733 {
3734         unsigned int idx;
3735         uint32_t opcode;
3736         uint32_t *data = ctx->data;
3737
3738         struct {
3739                 uint32_t opcode;
3740                 unsigned int min_len;
3741                 unsigned int max_len;
3742                 const char *name;
3743         } opcodes_3d[] = {
3744                 { 0x02, 1, 1, "3DSTATE_MODES_3" },
3745                 { 0x03, 1, 1, "3DSTATE_ENABLES_1" },
3746                 { 0x04, 1, 1, "3DSTATE_ENABLES_2" },
3747                 { 0x05, 1, 1, "3DSTATE_VFT0" },
3748                 { 0x06, 1, 1, "3DSTATE_AA" },
3749                 { 0x07, 1, 1, "3DSTATE_RASTERIZATION_RULES" },
3750                 { 0x08, 1, 1, "3DSTATE_MODES_1" },
3751                 { 0x09, 1, 1, "3DSTATE_STENCIL_TEST" },
3752                 { 0x0a, 1, 1, "3DSTATE_VFT1" },
3753                 { 0x0b, 1, 1, "3DSTATE_INDPT_ALPHA_BLEND" },
3754                 { 0x0c, 1, 1, "3DSTATE_MODES_5" },
3755                 { 0x0d, 1, 1, "3DSTATE_MAP_BLEND_OP" },
3756                 { 0x0e, 1, 1, "3DSTATE_MAP_BLEND_ARG" },
3757                 { 0x0f, 1, 1, "3DSTATE_MODES_2" },
3758                 { 0x15, 1, 1, "3DSTATE_FOG_COLOR" },
3759                 { 0x16, 1, 1, "3DSTATE_MODES_4"},
3760         }, *opcode_3d;
3761
3762         opcode = (data[0] & 0x1f000000) >> 24;
3763
3764         switch (opcode) {
3765         case 0x1f:
3766                 return decode_3d_primitive(ctx);
3767         case 0x1d:
3768                 return decode_3d_1d(ctx);
3769         case 0x1c:
3770                 return decode_3d_1c(ctx);
3771         }
3772
3773         for (idx = 0; idx < ARRAY_SIZE(opcodes_3d); idx++) {
3774                 opcode_3d = &opcodes_3d[idx];
3775                 if ((data[0] & 0x1f000000) >> 24 == opcode_3d->opcode) {
3776                         unsigned int len = 1, i;
3777
3778                         instr_out(ctx, 0, "%s\n", opcode_3d->name);
3779                         if (opcode_3d->max_len > 1) {
3780                                 len = (data[0] & 0xff) + 2;
3781                                 if (len < opcode_3d->min_len ||
3782                                     len > opcode_3d->max_len) {
3783                                         fprintf(out, "Bad count in %s\n",
3784                                                 opcode_3d->name);
3785                                 }
3786                         }
3787
3788                         for (i = 1; i < len; i++) {
3789                                 instr_out(ctx, i, "dword %d\n", i);
3790                         }
3791                         return len;
3792                 }
3793         }
3794
3795         instr_out(ctx, 0, "3D UNKNOWN: 3d_i830 opcode = 0x%x\n",
3796                   opcode);
3797         return 1;
3798 }
3799
3800 struct drm_intel_decode *
3801 drm_intel_decode_context_alloc(uint32_t devid)
3802 {
3803         struct drm_intel_decode *ctx;
3804
3805         ctx = calloc(1, sizeof(struct drm_intel_decode));
3806         if (!ctx)
3807                 return NULL;
3808
3809         ctx->devid = devid;
3810         ctx->out = stdout;
3811
3812         if (IS_GEN7(devid))
3813                 ctx->gen = 7;
3814         else if (IS_GEN6(devid))
3815                 ctx->gen = 6;
3816         else if (IS_GEN5(devid))
3817                 ctx->gen = 5;
3818         else if (IS_GEN4(devid))
3819                 ctx->gen = 4;
3820         else if (IS_9XX(devid))
3821                 ctx->gen = 3;
3822         else {
3823                 assert(IS_GEN2(devid));
3824                 ctx->gen = 2;
3825         }
3826
3827         return ctx;
3828 }
3829
3830 void
3831 drm_intel_decode_context_free(struct drm_intel_decode *ctx)
3832 {
3833         free(ctx);
3834 }
3835
3836 void
3837 drm_intel_decode_set_dump_past_end(struct drm_intel_decode *ctx,
3838                                    int dump_past_end)
3839 {
3840         ctx->dump_past_end = !!dump_past_end;
3841 }
3842
3843 void
3844 drm_intel_decode_set_batch_pointer(struct drm_intel_decode *ctx,
3845                                    void *data, uint32_t hw_offset, int count)
3846 {
3847         ctx->base_data = data;
3848         ctx->base_hw_offset = hw_offset;
3849         ctx->base_count = count;
3850 }
3851
3852 void
3853 drm_intel_decode_set_head_tail(struct drm_intel_decode *ctx,
3854                                uint32_t head, uint32_t tail)
3855 {
3856         ctx->head = head;
3857         ctx->tail = tail;
3858 }
3859
3860 void
3861 drm_intel_decode_set_output_file(struct drm_intel_decode *ctx,
3862                                  FILE *out)
3863 {
3864         ctx->out = out;
3865 }
3866
3867 /**
3868  * Decodes an i830-i915 batch buffer, writing the output to stdout.
3869  *
3870  * \param data batch buffer contents
3871  * \param count number of DWORDs to decode in the batch buffer
3872  * \param hw_offset hardware address for the buffer
3873  */
3874 void
3875 drm_intel_decode(struct drm_intel_decode *ctx)
3876 {
3877         int ret;
3878         unsigned int index = 0;
3879         uint32_t devid;
3880         int size = ctx->base_count * 4;
3881         void *temp;
3882
3883         if (!ctx)
3884                 return;
3885
3886         /* Put a scratch page full of obviously undefined data after
3887          * the batchbuffer.  This lets us avoid a bunch of length
3888          * checking in statically sized packets.
3889          */
3890         temp = malloc(size + 4096);
3891         memcpy(temp, ctx->base_data, size);
3892         memset((char *)temp + size, 0xd0, 4096);
3893         ctx->data = temp;
3894
3895         ctx->hw_offset = ctx->base_hw_offset;
3896         ctx->count = ctx->base_count;
3897
3898         devid = ctx->devid;
3899         head_offset = ctx->head;
3900         tail_offset = ctx->tail;
3901         out = ctx->out;
3902
3903         saved_s2_set = 0;
3904         saved_s4_set = 1;
3905
3906         while (ctx->count > 0) {
3907                 index = 0;
3908
3909                 switch ((ctx->data[index] & 0xe0000000) >> 29) {
3910                 case 0x0:
3911                         ret = decode_mi(ctx);
3912
3913                         /* If MI_BATCHBUFFER_END happened, then dump
3914                          * the rest of the output in case we some day
3915                          * want it in debugging, but don't decode it
3916                          * since it'll just confuse in the common
3917                          * case.
3918                          */
3919                         if (ret == -1) {
3920                                 if (ctx->dump_past_end) {
3921                                         index++;
3922                                 } else {
3923                                         for (index = index + 1; index < ctx->count;
3924                                              index++) {
3925                                                 instr_out(ctx, index, "\n");
3926                                         }
3927                                 }
3928                         } else
3929                                 index += ret;
3930                         break;
3931                 case 0x2:
3932                         index += decode_2d(ctx);
3933                         break;
3934                 case 0x3:
3935                         if (IS_9XX(devid) && !IS_GEN3(devid)) {
3936                                 index +=
3937                                     decode_3d_965(ctx);
3938                         } else if (IS_GEN3(devid)) {
3939                                 index += decode_3d(ctx);
3940                         } else {
3941                                 index +=
3942                                     decode_3d_i830(ctx);
3943                         }
3944                         break;
3945                 default:
3946                         instr_out(ctx, index, "UNKNOWN\n");
3947                         index++;
3948                         break;
3949                 }
3950                 fflush(out);
3951
3952                 if (ctx->count < index)
3953                         break;
3954
3955                 ctx->count -= index;
3956                 ctx->data += index;
3957                 ctx->hw_offset += 4 * index;
3958         }
3959
3960         free(temp);
3961 }