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