OSDN Git Service

intel: Import intel_decode.c from intel-gpu-tools.
[android-x86/external-libdrm.git] / 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 <stdint.h>
25 #include <stdio.h>
26 #include <stdarg.h>
27 #include <string.h>
28
29 #include "intel_decode.h"
30 #include "intel_chipset.h"
31 #include "intel_gpu_tools.h"
32
33 static FILE *out;
34 static uint32_t saved_s2 = 0, saved_s4 = 0;
35 static char saved_s2_set = 0, saved_s4_set = 0;
36 static uint32_t head_offset = 0xffffffff; /* undefined */
37 static uint32_t tail_offset = 0xffffffff; /* undefined */
38
39 #ifndef ARRAY_SIZE
40 #define ARRAY_SIZE(A) (sizeof(A)/sizeof(A[0]))
41 #endif
42
43 #define BUFFER_FAIL(_count, _len, _name) do {                   \
44     fprintf(out, "Buffer size too small in %s (%d < %d)\n",     \
45             (_name), (_count), (_len));                         \
46     (*failures)++;                                              \
47     return count;                                               \
48 } while (0)
49
50
51 static float
52 int_as_float(uint32_t intval)
53 {
54     union intfloat {
55         uint32_t i;
56         float f;
57     } uval;
58
59     uval.i = intval;
60     return uval.f;
61 }
62
63 static void
64 instr_out(uint32_t *data, uint32_t hw_offset, unsigned int index,
65           char *fmt, ...)
66 {
67     va_list va;
68     char *parseinfo;
69     uint32_t offset = hw_offset + index * 4;
70
71     if (offset == head_offset)
72         parseinfo = "HEAD";
73     else if (offset == tail_offset)
74         parseinfo = "TAIL";
75     else
76         parseinfo = "    ";
77
78     fprintf(out, "0x%08x: %s 0x%08x: %s", offset, parseinfo,
79             data[index],
80             index == 0 ? "" : "   ");
81     va_start(va, fmt);
82     vfprintf(out, fmt, va);
83     va_end(va);
84 }
85
86 static int
87 decode_mi(uint32_t *data, int count, uint32_t hw_offset, int *failures)
88 {
89     unsigned int opcode, len = -1;
90     char *post_sync_op = "";
91
92     struct {
93         uint32_t opcode;
94         int len_mask;
95         int min_len;
96         int max_len;
97         char *name;
98     } opcodes_mi[] = {
99         { 0x08, 0, 1, 1, "MI_ARB_ON_OFF" },
100         { 0x0a, 0, 1, 1, "MI_BATCH_BUFFER_END" },
101         { 0x30, 0x3f, 3, 3, "MI_BATCH_BUFFER" },
102         { 0x31, 0x3f, 2, 2, "MI_BATCH_BUFFER_START" },
103         { 0x14, 0x3f, 3, 3, "MI_DISPLAY_BUFFER_INFO" },
104         { 0x04, 0, 1, 1, "MI_FLUSH" },
105         { 0x22, 0x1f, 3, 3, "MI_LOAD_REGISTER_IMM" },
106         { 0x13, 0x3f, 2, 2, "MI_LOAD_SCAN_LINES_EXCL" },
107         { 0x12, 0x3f, 2, 2, "MI_LOAD_SCAN_LINES_INCL" },
108         { 0x00, 0, 1, 1, "MI_NOOP" },
109         { 0x11, 0x3f, 2, 2, "MI_OVERLAY_FLIP" },
110         { 0x07, 0, 1, 1, "MI_REPORT_HEAD" },
111         { 0x18, 0x3f, 2, 2, "MI_SET_CONTEXT" },
112         { 0x20, 0x3f, 3, 4, "MI_STORE_DATA_IMM" },
113         { 0x21, 0x3f, 3, 4, "MI_STORE_DATA_INDEX" },
114         { 0x24, 0x3f, 3, 3, "MI_STORE_REGISTER_MEM" },
115         { 0x02, 0, 1, 1, "MI_USER_INTERRUPT" },
116         { 0x03, 0, 1, 1, "MI_WAIT_FOR_EVENT" },
117         { 0x16, 0x7f, 3, 3, "MI_SEMAPHORE_MBOX" },
118         { 0x26, 0x1f, 3, 4, "MI_FLUSH_DW" },
119         { 0x0b, 0, 1, 1, "MI_SUSPEND_FLUSH" },
120     };
121
122     /* check instruction length */
123     for (opcode = 0; opcode < sizeof(opcodes_mi) / sizeof(opcodes_mi[0]);
124          opcode++) {
125         if ((data[0] & 0x1f800000) >> 23 == opcodes_mi[opcode].opcode) {
126             len = 1;
127             if (opcodes_mi[opcode].max_len > 1) {
128                 len = (data[0] & opcodes_mi[opcode].len_mask) + 2;
129                 if (len < opcodes_mi[opcode].min_len ||
130                     len > opcodes_mi[opcode].max_len)
131                 {
132                     fprintf(out, "Bad length (%d) in %s, [%d, %d]\n",
133                             len, opcodes_mi[opcode].name,
134                             opcodes_mi[opcode].min_len,
135                             opcodes_mi[opcode].max_len);
136                 }
137             }
138             break;
139         }
140     }
141
142     switch ((data[0] & 0x1f800000) >> 23) {
143     case 0x0a:
144         instr_out(data, hw_offset, 0, "MI_BATCH_BUFFER_END\n");
145         return -1;
146     case 0x16:
147         instr_out(data, hw_offset, 0, "MI_SEMAPHORE_MBOX%s%s%s%s %u\n",
148                   data[0] & (1<<22) ? " global gtt," : "",
149                   data[0] & (1<<21) ? " update semaphore," : "",
150                   data[0] & (1<<20) ? " compare semaphore," : "",
151                   data[0] & (1<<18) ? " use compare reg" : "",
152                   (data[0] & (0x3<<16)) >> 16);
153         instr_out(data, hw_offset, 1, "value\n");
154         instr_out(data, hw_offset, 2, "address\n");
155         return len;
156     case 0x21:
157         instr_out(data, hw_offset, 0, "MI_STORE_DATA_INDEX%s\n",
158                   data[0] & (1<<21) ? " use per-process HWS," : "");
159         instr_out(data, hw_offset, 1, "index\n");
160         instr_out(data, hw_offset, 2, "dword\n");
161         if (len == 4)
162                 instr_out(data, hw_offset, 3, "upper dword\n");
163         return len;
164     case 0x00:
165         if (data[0] & (1<<22))
166                 instr_out(data, hw_offset, 0, "MI_NOOP write NOPID reg, val=0x%x\n",
167                           data[0] & ((1<<22) - 1));
168         else
169                 instr_out(data, hw_offset, 0, "MI_NOOP\n");
170         return len;
171     case 0x26:
172         switch (data[0] & (0x3<<14)) {
173         case (0<<14): post_sync_op = "no write"; break;
174         case (1<<14): post_sync_op = "write data"; break;
175         case (2<<14): post_sync_op = "reserved"; break;
176         case (3<<14): post_sync_op = "write TIMESTAMP"; break;
177         }
178         instr_out(data, hw_offset, 0, "MI_FLUSH_DW%s%s%s%s post_sync_op='%s' %s%s\n",
179                   data[0] & (1<<22) ? " enable protected mem (BCS-only)," : "",
180                   data[0] & (1<<21) ? " store in hws," : "",
181                   data[0] & (1<<18) ? " invalidate tlb," : "",
182                   data[0] & (1<<17) ? " flush gfdt," : "",
183                   post_sync_op,
184                   data[0] & (1<<8) ? " enable notify interrupt," : "",
185                   data[0] & (1<<7) ? " invalidate video state (BCS-only)," : "");
186         if (data[0] & (1<<21))
187                 instr_out(data, hw_offset, 1, "hws index\n");
188         else
189                 instr_out(data, hw_offset, 1, "address\n");
190         instr_out(data, hw_offset, 2, "dword\n");
191         if (len == 4)
192                 instr_out(data, hw_offset, 3, "upper dword\n");
193         return len;
194     }
195
196     for (opcode = 0; opcode < sizeof(opcodes_mi) / sizeof(opcodes_mi[0]);
197          opcode++) {
198         if ((data[0] & 0x1f800000) >> 23 == opcodes_mi[opcode].opcode) {
199
200             instr_out(data, hw_offset, 0, "%s\n", opcodes_mi[opcode].name);
201             for (int i = 1; i < len; i++) {
202                 if (i >= count)
203                     BUFFER_FAIL(count, len, opcodes_mi[opcode].name);
204                 instr_out(data, hw_offset, i, "dword %d\n", i);
205             }
206
207             return len;
208         }
209     }
210
211     instr_out(data, hw_offset, 0, "MI UNKNOWN\n");
212     (*failures)++;
213     return 1;
214 }
215
216 static void
217 decode_2d_br00(uint32_t *data, int count, uint32_t hw_offset, char *cmd)
218 {
219     instr_out(data, hw_offset, 0,
220               "%s (rgb %sabled, alpha %sabled, src tile %d, dst tile %d)\n",
221               cmd,
222               (data[count] & (1 << 20)) ? "en" : "dis",
223               (data[count] & (1 << 21)) ? "en" : "dis",
224               (data[count] >> 15) & 1,
225               (data[count] >> 11) & 1);
226 }
227
228 static void
229 decode_2d_br01(uint32_t *data, int count, uint32_t hw_offset)
230 {
231     char *format;
232     switch ((data[count] >> 24) & 0x3) {
233     case 0:
234         format="8";
235         break;
236     case 1:
237         format="565";
238         break;
239     case 2:
240         format="1555";
241         break;
242     case 3:
243         format="8888";
244         break;
245     }
246
247     instr_out(data, hw_offset, count, "format %s, pitch %d, rop 0x%02x, "
248               "clipping %sabled, %s%s \n",
249
250               format,
251               (short)(data[count] & 0xffff),
252               (data[count] >> 16) &0xff,
253               data[count] & (1 << 30) ? "en" : "dis",
254               data[count] & (1 << 31) ? "solid pattern enabled, " : "",
255               data[count] & (1 << 31) ? "mono pattern transparency enabled, " : "");
256
257 }
258
259 static int
260 decode_2d(uint32_t *data, int count, uint32_t hw_offset, int *failures)
261 {
262     unsigned int opcode, len;
263
264     struct {
265         uint32_t opcode;
266         int min_len;
267         int max_len;
268         char *name;
269     } opcodes_2d[] = {
270         { 0x40, 5, 5, "COLOR_BLT" },
271         { 0x43, 6, 6, "SRC_COPY_BLT" },
272         { 0x01, 8, 8, "XY_SETUP_BLT" },
273         { 0x11, 9, 9, "XY_SETUP_MONO_PATTERN_SL_BLT" },
274         { 0x03, 3, 3, "XY_SETUP_CLIP_BLT" },
275         { 0x24, 2, 2, "XY_PIXEL_BLT" },
276         { 0x25, 3, 3, "XY_SCANLINES_BLT" },
277         { 0x26, 4, 4, "Y_TEXT_BLT" },
278         { 0x31, 5, 134, "XY_TEXT_IMMEDIATE_BLT" },
279         { 0x50, 6, 6, "XY_COLOR_BLT" },
280         { 0x51, 6, 6, "XY_PAT_BLT" },
281         { 0x76, 8, 8, "XY_PAT_CHROMA_BLT" },
282         { 0x72, 7, 135, "XY_PAT_BLT_IMMEDIATE" },
283         { 0x77, 9, 137, "XY_PAT_CHROMA_BLT_IMMEDIATE" },
284         { 0x52, 9, 9, "XY_MONO_PAT_BLT" },
285         { 0x59, 7, 7, "XY_MONO_PAT_FIXED_BLT" },
286         { 0x53, 8, 8, "XY_SRC_COPY_BLT" },
287         { 0x54, 8, 8, "XY_MONO_SRC_COPY_BLT" },
288         { 0x71, 9, 137, "XY_MONO_SRC_COPY_IMMEDIATE_BLT" },
289         { 0x55, 9, 9, "XY_FULL_BLT" },
290         { 0x55, 9, 137, "XY_FULL_IMMEDIATE_PATTERN_BLT" },
291         { 0x56, 9, 9, "XY_FULL_MONO_SRC_BLT" },
292         { 0x75, 10, 138, "XY_FULL_MONO_SRC_IMMEDIATE_PATTERN_BLT" },
293         { 0x57, 12, 12, "XY_FULL_MONO_PATTERN_BLT" },
294         { 0x58, 12, 12, "XY_FULL_MONO_PATTERN_MONO_SRC_BLT" },
295     };
296
297     switch ((data[0] & 0x1fc00000) >> 22) {
298     case 0x25:
299         instr_out(data, hw_offset, 0,
300                   "XY_SCANLINES_BLT (pattern seed (%d, %d), dst tile %d)\n",
301                   (data[0] >> 12) &0x8,
302                   (data[0] >> 8) &0x8,
303                   (data[0] >> 11) & 1);
304
305         len = (data[0] & 0x000000ff) + 2;
306         if (len != 3)
307             fprintf(out, "Bad count in XY_SCANLINES_BLT\n");
308         if (count < 3)
309             BUFFER_FAIL(count, len, "XY_SCANLINES_BLT");
310
311         instr_out(data, hw_offset, 1, "dest (%d,%d)\n",
312                   data[1] & 0xffff, data[1] >> 16);
313         instr_out(data, hw_offset, 2, "dest (%d,%d)\n",
314                   data[2] & 0xffff, data[2] >> 16);
315         return len;
316     case 0x01:
317         decode_2d_br00(data, 0, hw_offset, "XY_SETUP_BLT");
318
319         len = (data[0] & 0x000000ff) + 2;
320         if (len != 8)
321             fprintf(out, "Bad count in XY_SETUP_BLT\n");
322         if (count < 8)
323             BUFFER_FAIL(count, len, "XY_SETUP_BLT");
324
325         decode_2d_br01(data, 1, hw_offset);
326         instr_out(data, hw_offset, 2, "cliprect (%d,%d)\n",
327                   data[2] & 0xffff, data[2] >> 16);
328         instr_out(data, hw_offset, 3, "cliprect (%d,%d)\n",
329                   data[3] & 0xffff, data[3] >> 16);
330         instr_out(data, hw_offset, 4, "setup dst offset 0x%08x\n", data[4]);
331         instr_out(data, hw_offset, 5, "setup background color\n");
332         instr_out(data, hw_offset, 6, "setup foreground color\n");
333         instr_out(data, hw_offset, 7, "color pattern offset\n");
334         return len;
335     case 0x03:
336         decode_2d_br00(data, 0, hw_offset, "XY_SETUP_CLIP_BLT");
337
338         len = (data[0] & 0x000000ff) + 2;
339         if (len != 3)
340             fprintf(out, "Bad count in XY_SETUP_CLIP_BLT\n");
341         if (count < 3)
342             BUFFER_FAIL(count, len, "XY_SETUP_CLIP_BLT");
343
344         instr_out(data, hw_offset, 1, "cliprect (%d,%d)\n",
345                   data[1] & 0xffff, data[2] >> 16);
346         instr_out(data, hw_offset, 2, "cliprect (%d,%d)\n",
347                   data[2] & 0xffff, data[3] >> 16);
348         return len;
349     case 0x11:
350         decode_2d_br00(data, 0, hw_offset, "XY_SETUP_MONO_PATTERN_SL_BLT");
351
352         len = (data[0] & 0x000000ff) + 2;
353         if (len != 9)
354             fprintf(out, "Bad count in XY_SETUP_MONO_PATTERN_SL_BLT\n");
355         if (count < 9)
356             BUFFER_FAIL(count, len, "XY_SETUP_MONO_PATTERN_SL_BLT");
357
358         decode_2d_br01(data, 1, hw_offset);
359         instr_out(data, hw_offset, 2, "cliprect (%d,%d)\n",
360                   data[2] & 0xffff, data[2] >> 16);
361         instr_out(data, hw_offset, 3, "cliprect (%d,%d)\n",
362                   data[3] & 0xffff, data[3] >> 16);
363         instr_out(data, hw_offset, 4, "setup dst offset 0x%08x\n", data[4]);
364         instr_out(data, hw_offset, 5, "setup background color\n");
365         instr_out(data, hw_offset, 6, "setup foreground color\n");
366         instr_out(data, hw_offset, 7, "mono pattern dw0\n");
367         instr_out(data, hw_offset, 8, "mono pattern dw1\n");
368         return len;
369     case 0x50:
370         decode_2d_br00(data, 0, hw_offset, "XY_COLOR_BLT");
371
372         len = (data[0] & 0x000000ff) + 2;
373         if (len != 6)
374             fprintf(out, "Bad count in XY_COLOR_BLT\n");
375         if (count < 6)
376             BUFFER_FAIL(count, len, "XY_COLOR_BLT");
377
378         decode_2d_br01(data, 1, hw_offset);
379         instr_out(data, hw_offset, 2, "(%d,%d)\n",
380                   data[2] & 0xffff, data[2] >> 16);
381         instr_out(data, hw_offset, 3, "(%d,%d)\n",
382                   data[3] & 0xffff, data[3] >> 16);
383         instr_out(data, hw_offset, 4, "offset 0x%08x\n", data[4]);
384         instr_out(data, hw_offset, 5, "color\n");
385         return len;
386     case 0x53:
387         decode_2d_br00(data, 0, hw_offset, "XY_SRC_COPY_BLT");
388
389         len = (data[0] & 0x000000ff) + 2;
390         if (len != 8)
391             fprintf(out, "Bad count in XY_SRC_COPY_BLT\n");
392         if (count < 8)
393             BUFFER_FAIL(count, len, "XY_SRC_COPY_BLT");
394
395         decode_2d_br01(data, 1, hw_offset);
396         instr_out(data, hw_offset, 2, "dst (%d,%d)\n",
397                   data[2] & 0xffff, data[2] >> 16);
398         instr_out(data, hw_offset, 3, "dst (%d,%d)\n",
399                   data[3] & 0xffff, data[3] >> 16);
400         instr_out(data, hw_offset, 4, "dst offset 0x%08x\n", data[4]);
401         instr_out(data, hw_offset, 5, "src (%d,%d)\n",
402                   data[5] & 0xffff, data[5] >> 16);
403         instr_out(data, hw_offset, 6, "src pitch %d\n",
404                   (short)(data[6] & 0xffff));
405         instr_out(data, hw_offset, 7, "src offset 0x%08x\n", data[7]);
406         return len;
407     }
408
409     for (opcode = 0; opcode < sizeof(opcodes_2d) / sizeof(opcodes_2d[0]);
410          opcode++) {
411         if ((data[0] & 0x1fc00000) >> 22 == opcodes_2d[opcode].opcode) {
412             unsigned int i;
413
414             len = 1;
415             instr_out(data, hw_offset, 0, "%s\n", opcodes_2d[opcode].name);
416             if (opcodes_2d[opcode].max_len > 1) {
417                 len = (data[0] & 0x000000ff) + 2;
418                 if (len < opcodes_2d[opcode].min_len ||
419                     len > opcodes_2d[opcode].max_len)
420                 {
421                     fprintf(out, "Bad count in %s\n", opcodes_2d[opcode].name);
422                 }
423             }
424
425             for (i = 1; i < len; i++) {
426                 if (i >= count)
427                     BUFFER_FAIL(count, len, opcodes_2d[opcode].name);
428                 instr_out(data, hw_offset, i, "dword %d\n", i);
429             }
430
431             return len;
432         }
433     }
434
435     instr_out(data, hw_offset, 0, "2D UNKNOWN\n");
436     (*failures)++;
437     return 1;
438 }
439
440 static int
441 decode_3d_1c(uint32_t *data, int count, uint32_t hw_offset, int *failures)
442 {
443     uint32_t opcode;
444
445     opcode = (data[0] & 0x00f80000) >> 19;
446
447     switch (opcode) {
448     case 0x11:
449         instr_out(data, hw_offset, 0, "3DSTATE_DEPTH_SUBRECTANGLE_DISABLE\n");
450         return 1;
451     case 0x10:
452         instr_out(data, hw_offset, 0, "3DSTATE_SCISSOR_ENABLE %s\n",
453                 data[0]&1?"enabled":"disabled");
454         return 1;
455     case 0x01:
456         instr_out(data, hw_offset, 0, "3DSTATE_MAP_COORD_SET_I830\n");
457         return 1;
458     case 0x0a:
459         instr_out(data, hw_offset, 0, "3DSTATE_MAP_CUBE_I830\n");
460         return 1;
461     case 0x05:
462         instr_out(data, hw_offset, 0, "3DSTATE_MAP_TEX_STREAM_I830\n");
463         return 1;
464     }
465
466     instr_out(data, hw_offset, 0, "3D UNKNOWN: 3d_1c opcode = 0x%x\n",
467               opcode);
468     (*failures)++;
469     return 1;
470 }
471
472 /** Sets the string dstname to describe the destination of the PS instruction */
473 static void
474 i915_get_instruction_dst(uint32_t *data, int i, char *dstname, int do_mask)
475 {
476     uint32_t a0 = data[i];
477     int dst_nr = (a0 >> 14) & 0xf;
478     char dstmask[8];
479     char *sat;
480
481     if (do_mask) {
482         if (((a0 >> 10) & 0xf) == 0xf) {
483             dstmask[0] = 0;
484         } else {
485             int dstmask_index = 0;
486
487             dstmask[dstmask_index++] = '.';
488             if (a0 & (1 << 10))
489                 dstmask[dstmask_index++] = 'x';
490             if (a0 & (1 << 11))
491                 dstmask[dstmask_index++] = 'y';
492             if (a0 & (1 << 12))
493                 dstmask[dstmask_index++] = 'z';
494             if (a0 & (1 << 13))
495                 dstmask[dstmask_index++] = 'w';
496             dstmask[dstmask_index++] = 0;
497         }
498
499         if (a0 & (1 << 22))
500             sat = ".sat";
501         else
502             sat = "";
503     } else {
504         dstmask[0] = 0;
505         sat = "";
506     }
507
508     switch ((a0 >> 19) & 0x7) {
509     case 0:
510         if (dst_nr > 15)
511             fprintf(out, "bad destination reg R%d\n", dst_nr);
512         sprintf(dstname, "R%d%s%s", dst_nr, dstmask, sat);
513         break;
514     case 4:
515         if (dst_nr > 0)
516             fprintf(out, "bad destination reg oC%d\n", dst_nr);
517         sprintf(dstname, "oC%s%s", dstmask, sat);
518         break;
519     case 5:
520         if (dst_nr > 0)
521             fprintf(out, "bad destination reg oD%d\n", dst_nr);
522         sprintf(dstname, "oD%s%s",  dstmask, sat);
523         break;
524     case 6:
525         if (dst_nr > 3)
526             fprintf(out, "bad destination reg U%d\n", dst_nr);
527         sprintf(dstname, "U%d%s%s", dst_nr, dstmask, sat);
528         break;
529     default:
530         sprintf(dstname, "RESERVED");
531         break;
532     }
533 }
534
535 static char *
536 i915_get_channel_swizzle(uint32_t select)
537 {
538     switch (select & 0x7) {
539     case 0:
540         return (select & 8) ? "-x" : "x";
541     case 1:
542         return (select & 8) ? "-y" : "y";
543     case 2:
544         return (select & 8) ? "-z" : "z";
545     case 3:
546         return (select & 8) ? "-w" : "w";
547     case 4:
548         return (select & 8) ? "-0" : "0";
549     case 5:
550         return (select & 8) ? "-1" : "1";
551     default:
552         return (select & 8) ? "-bad" : "bad";
553     }
554 }
555
556 static void
557 i915_get_instruction_src_name(uint32_t src_type, uint32_t src_nr, char *name)
558 {
559     switch (src_type) {
560     case 0:
561         sprintf(name, "R%d", src_nr);
562         if (src_nr > 15)
563             fprintf(out, "bad src reg %s\n", name);
564         break;
565     case 1:
566         if (src_nr < 8)
567             sprintf(name, "T%d", src_nr);
568         else if (src_nr == 8)
569             sprintf(name, "DIFFUSE");
570         else if (src_nr == 9)
571             sprintf(name, "SPECULAR");
572         else if (src_nr == 10)
573             sprintf(name, "FOG");
574         else {
575             fprintf(out, "bad src reg T%d\n", src_nr);
576             sprintf(name, "RESERVED");
577         }
578         break;
579     case 2:
580         sprintf(name, "C%d", src_nr);
581         if (src_nr > 31)
582             fprintf(out, "bad src reg %s\n", name);
583         break;
584     case 4:
585         sprintf(name, "oC");
586         if (src_nr > 0)
587             fprintf(out, "bad src reg oC%d\n", src_nr);
588         break;
589     case 5:
590         sprintf(name, "oD");
591         if (src_nr > 0)
592             fprintf(out, "bad src reg oD%d\n", src_nr);
593         break;
594     case 6:
595         sprintf(name, "U%d", src_nr);
596         if (src_nr > 3)
597             fprintf(out, "bad src reg %s\n", name);
598         break;
599     default:
600         fprintf(out, "bad src reg type %d\n", src_type);
601         sprintf(name, "RESERVED");
602         break;
603     }
604 }
605
606 static void
607 i915_get_instruction_src0(uint32_t *data, int i, char *srcname)
608 {
609     uint32_t a0 = data[i];
610     uint32_t a1 = data[i + 1];
611     int src_nr = (a0 >> 2) & 0x1f;
612     char *swizzle_x = i915_get_channel_swizzle((a1 >> 28) & 0xf);
613     char *swizzle_y = i915_get_channel_swizzle((a1 >> 24) & 0xf);
614     char *swizzle_z = i915_get_channel_swizzle((a1 >> 20) & 0xf);
615     char *swizzle_w = i915_get_channel_swizzle((a1 >> 16) & 0xf);
616     char swizzle[100];
617
618     i915_get_instruction_src_name((a0 >> 7) & 0x7, src_nr, srcname);
619     sprintf(swizzle, ".%s%s%s%s", swizzle_x, swizzle_y, swizzle_z, swizzle_w);
620     if (strcmp(swizzle, ".xyzw") != 0)
621         strcat(srcname, swizzle);
622 }
623
624 static void
625 i915_get_instruction_src1(uint32_t *data, int i, char *srcname)
626 {
627     uint32_t a1 = data[i + 1];
628     uint32_t a2 = data[i + 2];
629     int src_nr = (a1 >> 8) & 0x1f;
630     char *swizzle_x = i915_get_channel_swizzle((a1 >> 4) & 0xf);
631     char *swizzle_y = i915_get_channel_swizzle((a1 >> 0) & 0xf);
632     char *swizzle_z = i915_get_channel_swizzle((a2 >> 28) & 0xf);
633     char *swizzle_w = i915_get_channel_swizzle((a2 >> 24) & 0xf);
634     char swizzle[100];
635
636     i915_get_instruction_src_name((a1 >> 13) & 0x7, src_nr, srcname);
637     sprintf(swizzle, ".%s%s%s%s", swizzle_x, swizzle_y, swizzle_z, swizzle_w);
638     if (strcmp(swizzle, ".xyzw") != 0)
639         strcat(srcname, swizzle);
640 }
641
642 static void
643 i915_get_instruction_src2(uint32_t *data, int i, char *srcname)
644 {
645     uint32_t a2 = data[i + 2];
646     int src_nr = (a2 >> 16) & 0x1f;
647     char *swizzle_x = i915_get_channel_swizzle((a2 >> 12) & 0xf);
648     char *swizzle_y = i915_get_channel_swizzle((a2 >> 8) & 0xf);
649     char *swizzle_z = i915_get_channel_swizzle((a2 >> 4) & 0xf);
650     char *swizzle_w = i915_get_channel_swizzle((a2 >> 0) & 0xf);
651     char swizzle[100];
652
653     i915_get_instruction_src_name((a2 >> 21) & 0x7, src_nr, srcname);
654     sprintf(swizzle, ".%s%s%s%s", swizzle_x, swizzle_y, swizzle_z, swizzle_w);
655     if (strcmp(swizzle, ".xyzw") != 0)
656         strcat(srcname, swizzle);
657 }
658
659 static void
660 i915_get_instruction_addr(uint32_t src_type, uint32_t src_nr, char *name)
661 {
662     switch (src_type) {
663     case 0:
664         sprintf(name, "R%d", src_nr);
665         if (src_nr > 15)
666             fprintf(out, "bad src reg %s\n", name);
667         break;
668     case 1:
669         if (src_nr < 8)
670             sprintf(name, "T%d", src_nr);
671         else if (src_nr == 8)
672             sprintf(name, "DIFFUSE");
673         else if (src_nr == 9)
674             sprintf(name, "SPECULAR");
675         else if (src_nr == 10)
676             sprintf(name, "FOG");
677         else {
678             fprintf(out, "bad src reg T%d\n", src_nr);
679             sprintf(name, "RESERVED");
680         }
681         break;
682     case 4:
683         sprintf(name, "oC");
684         if (src_nr > 0)
685             fprintf(out, "bad src reg oC%d\n", src_nr);
686         break;
687     case 5:
688         sprintf(name, "oD");
689         if (src_nr > 0)
690             fprintf(out, "bad src reg oD%d\n", src_nr);
691         break;
692     default:
693         fprintf(out, "bad src reg type %d\n", src_type);
694         sprintf(name, "RESERVED");
695         break;
696     }
697 }
698
699 static void
700 i915_decode_alu1(uint32_t *data, uint32_t hw_offset,
701                  int i, char *instr_prefix, char *op_name)
702 {
703     char dst[100], src0[100];
704
705     i915_get_instruction_dst(data, i, dst, 1);
706     i915_get_instruction_src0(data, i, src0);
707
708     instr_out(data, hw_offset, i++, "%s: %s %s, %s\n", instr_prefix,
709               op_name, dst, src0);
710     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
711     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
712 }
713
714 static void
715 i915_decode_alu2(uint32_t *data, uint32_t hw_offset,
716                  int i, char *instr_prefix, char *op_name)
717 {
718     char dst[100], src0[100], src1[100];
719
720     i915_get_instruction_dst(data, i, dst, 1);
721     i915_get_instruction_src0(data, i, src0);
722     i915_get_instruction_src1(data, i, src1);
723
724     instr_out(data, hw_offset, i++, "%s: %s %s, %s, %s\n", instr_prefix,
725               op_name, dst, src0, src1);
726     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
727     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
728 }
729
730 static void
731 i915_decode_alu3(uint32_t *data, uint32_t hw_offset,
732                  int i, char *instr_prefix, char *op_name)
733 {
734     char dst[100], src0[100], src1[100], src2[100];
735
736     i915_get_instruction_dst(data, i, dst, 1);
737     i915_get_instruction_src0(data, i, src0);
738     i915_get_instruction_src1(data, i, src1);
739     i915_get_instruction_src2(data, i, src2);
740
741     instr_out(data, hw_offset, i++, "%s: %s %s, %s, %s, %s\n", instr_prefix,
742               op_name, dst, src0, src1, src2);
743     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
744     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
745 }
746
747 static void
748 i915_decode_tex(uint32_t *data, uint32_t hw_offset, int i, char *instr_prefix,
749                 char *tex_name)
750 {
751     uint32_t t0 = data[i];
752     uint32_t t1 = data[i + 1];
753     char dst_name[100];
754     char addr_name[100];
755     int sampler_nr;
756
757     i915_get_instruction_dst(data, i, dst_name, 0);
758     i915_get_instruction_addr((t1 >> 24) & 0x7,
759                               (t1 >> 17) & 0xf,
760                               addr_name);
761     sampler_nr = t0 & 0xf;
762
763     instr_out(data, hw_offset, i++, "%s: %s %s, S%d, %s\n", instr_prefix,
764               tex_name, dst_name, sampler_nr, addr_name);
765     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
766     instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
767 }
768
769 static void
770 i915_decode_dcl(uint32_t *data, uint32_t hw_offset, int i, char *instr_prefix)
771 {
772     uint32_t d0 = data[i];
773     char *sampletype;
774     int dcl_nr = (d0 >> 14) & 0xf;
775     char *dcl_x = d0 & (1 << 10) ? "x" : "";
776     char *dcl_y = d0 & (1 << 11) ? "y" : "";
777     char *dcl_z = d0 & (1 << 12) ? "z" : "";
778     char *dcl_w = d0 & (1 << 13) ? "w" : "";
779     char dcl_mask[10];
780
781     switch ((d0 >> 19) & 0x3) {
782     case 1:
783         sprintf(dcl_mask, ".%s%s%s%s", dcl_x, dcl_y, dcl_z, dcl_w);
784         if (strcmp(dcl_mask, ".") == 0)
785             fprintf(out, "bad (empty) dcl mask\n");
786
787         if (dcl_nr > 10)
788             fprintf(out, "bad T%d dcl register number\n", dcl_nr);
789         if (dcl_nr < 8) {
790             if (strcmp(dcl_mask, ".x") != 0 &&
791                 strcmp(dcl_mask, ".xy") != 0 &&
792                 strcmp(dcl_mask, ".xz") != 0 &&
793                 strcmp(dcl_mask, ".w") != 0 &&
794                 strcmp(dcl_mask, ".xyzw") != 0) {
795                 fprintf(out, "bad T%d.%s dcl mask\n", dcl_nr, dcl_mask);
796             }
797             instr_out(data, hw_offset, i++, "%s: DCL T%d%s\n", instr_prefix,
798                       dcl_nr, dcl_mask);
799         } else {
800             if (strcmp(dcl_mask, ".xz") == 0)
801                 fprintf(out, "errataed bad dcl mask %s\n", dcl_mask);
802             else if (strcmp(dcl_mask, ".xw") == 0)
803                 fprintf(out, "errataed bad dcl mask %s\n", dcl_mask);
804             else if (strcmp(dcl_mask, ".xzw") == 0)
805                 fprintf(out, "errataed bad dcl mask %s\n", dcl_mask);
806
807             if (dcl_nr == 8) {
808                 instr_out(data, hw_offset, i++, "%s: DCL DIFFUSE%s\n", instr_prefix,
809                           dcl_mask);
810             } else if (dcl_nr == 9) {
811                 instr_out(data, hw_offset, i++, "%s: DCL SPECULAR%s\n", instr_prefix,
812                           dcl_mask);
813             } else if (dcl_nr == 10) {
814                 instr_out(data, hw_offset, i++, "%s: DCL FOG%s\n", instr_prefix,
815                           dcl_mask);
816             }
817         }
818         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
819         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
820         break;
821     case 3:
822         switch ((d0 >> 22) & 0x3) {
823         case 0:
824             sampletype = "2D";
825             break;
826         case 1:
827             sampletype = "CUBE";
828             break;
829         case 2:
830             sampletype = "3D";
831             break;
832         default:
833             sampletype = "RESERVED";
834             break;
835         }
836         if (dcl_nr > 15)
837             fprintf(out, "bad S%d dcl register number\n", dcl_nr);
838         instr_out(data, hw_offset, i++, "%s: DCL S%d %s\n", instr_prefix,
839                   dcl_nr, sampletype);
840         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
841         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
842         break;
843     default:
844         instr_out(data, hw_offset, i++, "%s: DCL RESERVED%d\n", instr_prefix, dcl_nr);
845         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
846         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
847     }
848 }
849
850 static void
851 i915_decode_instruction(uint32_t *data, uint32_t hw_offset,
852                         int i, char *instr_prefix)
853 {
854     switch ((data[i] >> 24) & 0x1f) {
855     case 0x0:
856         instr_out(data, hw_offset, i++, "%s: NOP\n", instr_prefix);
857         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
858         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
859         break;
860     case 0x01:
861         i915_decode_alu2(data, hw_offset, i, instr_prefix, "ADD");
862         break;
863     case 0x02:
864         i915_decode_alu1(data, hw_offset, i, instr_prefix, "MOV");
865         break;
866     case 0x03:
867         i915_decode_alu2(data, hw_offset, i, instr_prefix, "MUL");
868         break;
869     case 0x04:
870         i915_decode_alu3(data, hw_offset, i, instr_prefix, "MAD");
871         break;
872     case 0x05:
873         i915_decode_alu3(data, hw_offset, i, instr_prefix, "DP2ADD");
874         break;
875     case 0x06:
876         i915_decode_alu2(data, hw_offset, i, instr_prefix, "DP3");
877         break;
878     case 0x07:
879         i915_decode_alu2(data, hw_offset, i, instr_prefix, "DP4");
880         break;
881     case 0x08:
882         i915_decode_alu1(data, hw_offset, i, instr_prefix, "FRC");
883         break;
884     case 0x09:
885         i915_decode_alu1(data, hw_offset, i, instr_prefix, "RCP");
886         break;
887     case 0x0a:
888         i915_decode_alu1(data, hw_offset, i, instr_prefix, "RSQ");
889         break;
890     case 0x0b:
891         i915_decode_alu1(data, hw_offset, i, instr_prefix, "EXP");
892         break;
893     case 0x0c:
894         i915_decode_alu1(data, hw_offset, i, instr_prefix, "LOG");
895         break;
896     case 0x0d:
897         i915_decode_alu2(data, hw_offset, i, instr_prefix, "CMP");
898         break;
899     case 0x0e:
900         i915_decode_alu2(data, hw_offset, i, instr_prefix, "MIN");
901         break;
902     case 0x0f:
903         i915_decode_alu2(data, hw_offset, i, instr_prefix, "MAX");
904         break;
905     case 0x10:
906         i915_decode_alu1(data, hw_offset, i, instr_prefix, "FLR");
907         break;
908     case 0x11:
909         i915_decode_alu1(data, hw_offset, i, instr_prefix, "MOD");
910         break;
911     case 0x12:
912         i915_decode_alu1(data, hw_offset, i, instr_prefix, "TRC");
913         break;
914     case 0x13:
915         i915_decode_alu2(data, hw_offset, i, instr_prefix, "SGE");
916         break;
917     case 0x14:
918         i915_decode_alu2(data, hw_offset, i, instr_prefix, "SLT");
919         break;
920     case 0x15:
921         i915_decode_tex(data, hw_offset, i, instr_prefix, "TEXLD");
922         break;
923     case 0x16:
924         i915_decode_tex(data, hw_offset, i, instr_prefix, "TEXLDP");
925         break;
926     case 0x17:
927         i915_decode_tex(data, hw_offset, i, instr_prefix, "TEXLDB");
928         break;
929     case 0x19:
930         i915_decode_dcl(data, hw_offset, i, instr_prefix);
931         break;
932     default:
933         instr_out(data, hw_offset, i++, "%s: unknown\n", instr_prefix);
934         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
935         instr_out(data, hw_offset, i++, "%s\n", instr_prefix);
936         break;
937     }
938 }
939
940 static char *
941 decode_compare_func(uint32_t op)
942 {
943     switch (op&0x7) {
944     case 0: return "always";
945     case 1: return "never";
946     case 2: return "less";
947     case 3: return "equal";
948     case 4: return "lequal";
949     case 5: return "greater";
950     case 6: return "notequal";
951     case 7: return "gequal";
952     }
953     return "";
954 }
955
956 static char *
957 decode_stencil_op(uint32_t op)
958 {
959     switch (op&0x7) {
960     case 0: return "keep";
961     case 1: return "zero";
962     case 2: return "replace";
963     case 3: return "incr_sat";
964     case 4: return "decr_sat";
965     case 5: return "greater";
966     case 6: return "incr";
967     case 7: return "decr";
968     }
969     return "";
970 }
971
972 static char *
973 decode_logic_op(uint32_t op)
974 {
975     switch (op&0xf) {
976     case 0: return "clear";
977     case 1: return "nor";
978     case 2: return "and_inv";
979     case 3: return "copy_inv";
980     case 4: return "and_rvrse";
981     case 5: return "inv";
982     case 6: return "xor";
983     case 7: return "nand";
984     case 8: return "and";
985     case 9: return "equiv";
986     case 10: return "noop";
987     case 11: return "or_inv";
988     case 12: return "copy";
989     case 13: return "or_rvrse";
990     case 14: return "or";
991     case 15: return "set";
992     }
993     return "";
994 }
995
996 static char *
997 decode_blend_fact(uint32_t op)
998 {
999     switch (op&0xf) {
1000     case 1: return "zero";
1001     case 2: return "one";
1002     case 3: return "src_colr";
1003     case 4: return "inv_src_colr";
1004     case 5: return "src_alpha";
1005     case 6: return "inv_src_alpha";
1006     case 7: return "dst_alpha";
1007     case 8: return "inv_dst_alpha";
1008     case 9: return "dst_colr";
1009     case 10: return "inv_dst_colr";
1010     case 11: return "src_alpha_sat";
1011     case 12: return "cnst_colr";
1012     case 13: return "inv_cnst_colr";
1013     case 14: return "cnst_alpha";
1014     case 15: return "inv_const_alpha";
1015     }
1016     return "";
1017 }
1018
1019 static char *
1020 decode_tex_coord_mode(uint32_t mode)
1021 {
1022     switch (mode&0x7) {
1023     case 0: return "wrap";
1024     case 1: return "mirror";
1025     case 2: return "clamp_edge";
1026     case 3: return "cube";
1027     case 4: return "clamp_border";
1028     case 5: return "mirror_once";
1029     }
1030     return "";
1031 }
1032
1033 static char *
1034 decode_sample_filter(uint32_t mode)
1035 {
1036     switch (mode&0x7) {
1037     case 0: return "nearest";
1038     case 1: return "linear";
1039     case 2: return "anisotropic";
1040     case 3: return "4x4_1";
1041     case 4: return "4x4_2";
1042     case 5: return "4x4_flat";
1043     case 6: return "6x5_mono";
1044     }
1045     return "";
1046 }
1047
1048 static int
1049 decode_3d_1d(uint32_t *data, int count,
1050              uint32_t hw_offset,
1051              uint32_t devid,
1052              int *failures)
1053 {
1054     unsigned int len, i, c, idx, word, map, sampler, instr;
1055     char *format, *zformat, *type;
1056     uint32_t opcode;
1057
1058     struct {
1059         uint32_t opcode;
1060         int i830_only;
1061         int min_len;
1062         int max_len;
1063         char *name;
1064     } opcodes_3d_1d[] = {
1065         { 0x86, 0, 4, 4, "3DSTATE_CHROMA_KEY" },
1066         { 0x88, 0, 2, 2, "3DSTATE_CONSTANT_BLEND_COLOR" },
1067         { 0x99, 0, 2, 2, "3DSTATE_DEFAULT_DIFFUSE" },
1068         { 0x9a, 0, 2, 2, "3DSTATE_DEFAULT_SPECULAR" },
1069         { 0x98, 0, 2, 2, "3DSTATE_DEFAULT_Z" },
1070         { 0x97, 0, 2, 2, "3DSTATE_DEPTH_OFFSET_SCALE" },
1071         { 0x9d, 0, 65, 65, "3DSTATE_FILTER_COEFFICIENTS_4X4" },
1072         { 0x9e, 0, 4, 4, "3DSTATE_MONO_FILTER" },
1073         { 0x89, 0, 4, 4, "3DSTATE_FOG_MODE" },
1074         { 0x8f, 0, 2, 16, "3DSTATE_MAP_PALLETE_LOAD_32" },
1075         { 0x83, 0, 2, 2, "3DSTATE_SPAN_STIPPLE" },
1076         { 0x8c, 1, 2, 2, "3DSTATE_MAP_COORD_TRANSFORM_I830" },
1077         { 0x8b, 1, 2, 2, "3DSTATE_MAP_VERTEX_TRANSFORM_I830" },
1078         { 0x8d, 1, 3, 3, "3DSTATE_W_STATE_I830" },
1079         { 0x01, 1, 2, 2, "3DSTATE_COLOR_FACTOR_I830" },
1080         { 0x02, 1, 2, 2, "3DSTATE_MAP_COORD_SETBIND_I830" },
1081     }, *opcode_3d_1d;
1082
1083     opcode = (data[0] & 0x00ff0000) >> 16;
1084
1085     switch (opcode) {
1086     case 0x07:
1087         /* This instruction is unusual.  A 0 length means just 1 DWORD instead of
1088          * 2.  The 0 length is specified in one place to be unsupported, but
1089          * stated to be required in another, and 0 length LOAD_INDIRECTs appear
1090          * to cause no harm at least.
1091          */
1092         instr_out(data, hw_offset, 0, "3DSTATE_LOAD_INDIRECT\n");
1093         len = (data[0] & 0x000000ff) + 1;
1094         i = 1;
1095         if (data[0] & (0x01 << 8)) {
1096             if (i + 2 >= count)
1097                 BUFFER_FAIL(count, len, "3DSTATE_LOAD_INDIRECT");
1098             instr_out(data, hw_offset, i++, "SIS.0\n");
1099             instr_out(data, hw_offset, i++, "SIS.1\n");
1100         }
1101         if (data[0] & (0x02 << 8)) {
1102             if (i + 1 >= count)
1103                 BUFFER_FAIL(count, len, "3DSTATE_LOAD_INDIRECT");
1104             instr_out(data, hw_offset, i++, "DIS.0\n");
1105         }
1106         if (data[0] & (0x04 << 8)) {
1107             if (i + 2 >= count)
1108                 BUFFER_FAIL(count, len, "3DSTATE_LOAD_INDIRECT");
1109             instr_out(data, hw_offset, i++, "SSB.0\n");
1110             instr_out(data, hw_offset, i++, "SSB.1\n");
1111         }
1112         if (data[0] & (0x08 << 8)) {
1113             if (i + 2 >= count)
1114                 BUFFER_FAIL(count, len, "3DSTATE_LOAD_INDIRECT");
1115             instr_out(data, hw_offset, i++, "MSB.0\n");
1116             instr_out(data, hw_offset, i++, "MSB.1\n");
1117         }
1118         if (data[0] & (0x10 << 8)) {
1119             if (i + 2 >= count)
1120                 BUFFER_FAIL(count, len, "3DSTATE_LOAD_INDIRECT");
1121             instr_out(data, hw_offset, i++, "PSP.0\n");
1122             instr_out(data, hw_offset, i++, "PSP.1\n");
1123         }
1124         if (data[0] & (0x20 << 8)) {
1125             if (i + 2 >= count)
1126                 BUFFER_FAIL(count, len, "3DSTATE_LOAD_INDIRECT");
1127             instr_out(data, hw_offset, i++, "PSC.0\n");
1128             instr_out(data, hw_offset, i++, "PSC.1\n");
1129         }
1130         if (len != i) {
1131             fprintf(out, "Bad count in 3DSTATE_LOAD_INDIRECT\n");
1132             (*failures)++;
1133             return len;
1134         }
1135         return len;
1136     case 0x04:
1137         instr_out(data, hw_offset, 0, "3DSTATE_LOAD_STATE_IMMEDIATE_1\n");
1138         len = (data[0] & 0x0000000f) + 2;
1139         i = 1;
1140         for (word = 0; word <= 8; word++) {
1141             if (data[0] & (1 << (4 + word))) {
1142                 if (i >= count)
1143                     BUFFER_FAIL(count, len, "3DSTATE_LOAD_STATE_IMMEDIATE_1");
1144
1145                 /* save vertex state for decode */
1146                 if (!IS_GEN2(devid)) {
1147                         if (word == 2) {
1148                                 saved_s2_set = 1;
1149                                 saved_s2 = data[i];
1150                         }
1151                         if (word == 4) {
1152                                 saved_s4_set = 1;
1153                                 saved_s4 = data[i];
1154                         }
1155
1156                         switch (word) {
1157                         case 0:
1158                                 instr_out(data, hw_offset, i, "S0: vbo offset: 0x%08x%s\n",
1159                                           data[i]&(~1),data[i]&1?", auto cache invalidate disabled":"");
1160                                 break;
1161                         case 1:
1162                                 instr_out(data, hw_offset, i, "S1: vertex width: %i, vertex pitch: %i\n",
1163                                           (data[i]>>24)&0x3f,(data[i]>>16)&0x3f);
1164                                 break;
1165                         case 2:
1166                                 instr_out(data, hw_offset, i, "S2: texcoord formats: ");
1167                                 for (int tex_num = 0; tex_num < 8; tex_num++) {
1168                                         switch((data[i]>>tex_num*4)&0xf) {
1169                                         case 0: fprintf(out, "%i=2D ", tex_num); break;
1170                                         case 1: fprintf(out, "%i=3D ", tex_num); break;
1171                                         case 2: fprintf(out, "%i=4D ", tex_num); break;
1172                                         case 3: fprintf(out, "%i=1D ", tex_num); break;
1173                                         case 4: fprintf(out, "%i=2D_16 ", tex_num); break;
1174                                         case 5: fprintf(out, "%i=4D_16 ", tex_num); break;
1175                                         case 0xf: fprintf(out, "%i=NP ", tex_num); break;
1176                                         }
1177                                 }
1178                                 fprintf(out, "\n");
1179
1180                                 break;
1181                         case 3:
1182                                 instr_out(data, hw_offset, i, "S3: not documented\n", word);
1183                                 break;
1184                         case 4:
1185                                 {
1186                                         char *cullmode = "";
1187                                         char *vfmt_xyzw = "";
1188                                         switch((data[i]>>13)&0x3) {
1189                                         case 0: cullmode = "both"; break;
1190                                         case 1: cullmode = "none"; break;
1191                                         case 2: cullmode = "cw"; break;
1192                                         case 3: cullmode = "ccw"; break;
1193                                         }
1194                                         switch(data[i] & (7<<6 | 1<<2)) {
1195                                         case 1<<6: vfmt_xyzw = "XYZ,"; break;
1196                                         case 2<<6: vfmt_xyzw = "XYZW,"; break;
1197                                         case 3<<6: vfmt_xyzw = "XY,"; break;
1198                                         case 4<<6: vfmt_xyzw = "XYW,"; break;
1199                                         case 1<<6 | 1<<2: vfmt_xyzw = "XYZF,"; break;
1200                                         case 2<<6 | 1<<2: vfmt_xyzw = "XYZWF,"; break;
1201                                         case 3<<6 | 1<<2: vfmt_xyzw = "XYF,"; break;
1202                                         case 4<<6 | 1<<2: vfmt_xyzw = "XYWF,"; break;
1203                                         }
1204                                         instr_out(data, hw_offset, i, "S4: point_width=%i, line_width=%.1f,"
1205                                                   "%s%s%s%s%s cullmode=%s, vfmt=%s%s%s%s%s%s "
1206                                                   "%s%s\n",
1207                                                   (data[i]>>23)&0x1ff,
1208                                                   ((data[i]>>19)&0xf) / 2.0,
1209                                                   data[i]&(0xf<<15)?" flatshade=":"",
1210                                                   data[i]&(1<<18)?"Alpha,":"",
1211                                                   data[i]&(1<<17)?"Fog,":"",
1212                                                   data[i]&(1<<16)?"Specular,":"",
1213                                                   data[i]&(1<<15)?"Color,":"",
1214                                                   cullmode,
1215                                                   data[i]&(1<<12)?"PointWidth,":"",
1216                                                   data[i]&(1<<11)?"SpecFog,":"",
1217                                                   data[i]&(1<<10)?"Color,":"",
1218                                                   data[i]&(1<<9)?"DepthOfs,":"",
1219                                                   vfmt_xyzw,
1220                                                   data[i]&(1<<9)?"FogParam,":"",
1221                                                   data[i]&(1<<5)?"force default diffuse, ":"",
1222                                                   data[i]&(1<<4)?"force default specular, ":"",
1223                                                   data[i]&(1<<3)?"local depth ofs enable, ":"",
1224                                                   data[i]&(1<<1)?"point sprite enable, ":"",
1225                                                   data[i]&(1<<0)?"line AA enable, ":"");
1226                                         break;
1227                                 }
1228                         case 5:
1229                                 {
1230                                         instr_out(data, hw_offset, i, "S5:%s%s%s%s%s"
1231                                                   "%s%s%s%s stencil_ref=0x%x, stencil_test=%s, "
1232                                                   "stencil_fail=%s, stencil_pass_z_fail=%s, "
1233                                                   "stencil_pass_z_pass=%s, %s%s%s%s\n",
1234                                                   data[i]&(0xf<<28)?" write_disable=":"",
1235                                                   data[i]&(1<<31)?"Alpha,":"",
1236                                                   data[i]&(1<<30)?"Red,":"",
1237                                                   data[i]&(1<<29)?"Green,":"",
1238                                                   data[i]&(1<<28)?"Blue,":"",
1239                                                   data[i]&(1<<27)?" force default point size,":"",
1240                                                   data[i]&(1<<26)?" last pixel enable,":"",
1241                                                   data[i]&(1<<25)?" global depth ofs enable,":"",
1242                                                   data[i]&(1<<24)?" fog enable,":"",
1243                                                   (data[i]>>16)&0xff,
1244                                                   decode_compare_func(data[i]>>13),
1245                                                   decode_stencil_op(data[i]>>10),
1246                                                   decode_stencil_op(data[i]>>7),
1247                                                   decode_stencil_op(data[i]>>4),
1248                                                   data[i]&(1<<3)?"stencil write enable, ":"",
1249                                                   data[i]&(1<<2)?"stencil test enable, ":"",
1250                                                   data[i]&(1<<1)?"color dither enable, ":"",
1251                                                   data[i]&(1<<0)?"logicop enable, ":"");
1252                                 }
1253                                 break;
1254                         case 6:
1255                                 instr_out(data, hw_offset, i, "S6: %salpha_test=%s, alpha_ref=0x%x, "
1256                                           "depth_test=%s, %ssrc_blnd_fct=%s, dst_blnd_fct=%s, "
1257                                           "%s%stristrip_provoking_vertex=%i\n",
1258                                           data[i]&(1<<31)?"alpha test enable, ":"",
1259                                           decode_compare_func(data[i]>>28),
1260                                           data[i]&(0xff<<20),
1261                                           decode_compare_func(data[i]>>16),
1262                                           data[i]&(1<<15)?"cbuf blend enable, ":"",
1263                                           decode_blend_fact(data[i]>>8),
1264                                           decode_blend_fact(data[i]>>4),
1265                                           data[i]&(1<<3)?"depth write enable, ":"",
1266                                           data[i]&(1<<2)?"cbuf write enable, ":"",
1267                                           data[i]&(0x3));
1268                                 break;
1269                         case 7:
1270                                 instr_out(data, hw_offset, i, "S7: depth offset constant: 0x%08x\n", data[i]);
1271                                 break;
1272                         }
1273                 } else {
1274                         instr_out(data, hw_offset, i, "S%d: 0x%08x\n", i, data[i]);
1275                 }
1276                 i++;
1277             }
1278         }
1279         if (len != i) {
1280             fprintf(out, "Bad count in 3DSTATE_LOAD_STATE_IMMEDIATE_1\n");
1281             (*failures)++;
1282         }
1283         return len;
1284     case 0x03:
1285         instr_out(data, hw_offset, 0, "3DSTATE_LOAD_STATE_IMMEDIATE_2\n");
1286         len = (data[0] & 0x0000000f) + 2;
1287         i = 1;
1288         for (word = 6; word <= 14; word++) {
1289             if (data[0] & (1 << word)) {
1290                 if (i >= count)
1291                     BUFFER_FAIL(count, len, "3DSTATE_LOAD_STATE_IMMEDIATE_2");
1292
1293                 if (word == 6)
1294                     instr_out(data, hw_offset, i++, "TBCF\n");
1295                 else if (word >= 7 && word <= 10) {
1296                     instr_out(data, hw_offset, i++, "TB%dC\n", word - 7);
1297                     instr_out(data, hw_offset, i++, "TB%dA\n", word - 7);
1298                 } else if (word >= 11 && word <= 14) {
1299                     instr_out(data, hw_offset, i, "TM%dS0: offset=0x%08x, %s\n",
1300                               word - 11,
1301                               data[i]&0xfffffffe,
1302                               data[i]&1?"use fence":"");
1303                     i++;
1304                     instr_out(data, hw_offset, i, "TM%dS1: height=%i, width=%i, %s\n",
1305                               word - 11,
1306                               data[i]>>21, (data[i]>>10)&0x3ff,
1307                               data[i]&2?(data[i]&1?"y-tiled":"x-tiled"):"");
1308                     i++;
1309                     instr_out(data, hw_offset, i, "TM%dS2: pitch=%i, \n",
1310                               word - 11,
1311                               ((data[i]>>21) + 1)*4);
1312                     i++;
1313                     instr_out(data, hw_offset, i++, "TM%dS3\n", word - 11);
1314                     instr_out(data, hw_offset, i++, "TM%dS4: dflt color\n", word - 11);
1315                 }
1316             }
1317         }
1318         if (len != i) {
1319             fprintf(out, "Bad count in 3DSTATE_LOAD_STATE_IMMEDIATE_2\n");
1320             (*failures)++;
1321         }
1322         return len;
1323     case 0x00:
1324         instr_out(data, hw_offset, 0, "3DSTATE_MAP_STATE\n");
1325         len = (data[0] & 0x0000003f) + 2;
1326         instr_out(data, hw_offset, 1, "mask\n");
1327
1328         i = 2;
1329         for (map = 0; map <= 15; map++) {
1330             if (data[1] & (1 << map)) {
1331                 int width, height, pitch, dword;
1332                 const char *tiling;
1333
1334                 if (i + 3 >= count)
1335                     BUFFER_FAIL(count, len, "3DSTATE_MAP_STATE");
1336
1337                 dword = data[i];
1338                 instr_out(data, hw_offset, i++, "map %d MS2 %s%s%s\n", map,
1339                           dword&(1<<31)?"untrusted surface, ":"",
1340                           dword&(1<<1)?"vertical line stride enable, ":"",
1341                           dword&(1<<0)?"vertical ofs enable, ":"");
1342
1343                 dword = data[i];
1344                 width = ((dword >> 10) & ((1 << 11) - 1))+1;
1345                 height = ((dword >> 21) & ((1 << 11) - 1))+1;
1346
1347                 tiling = "none";
1348                 if (dword & (1 << 2))
1349                         tiling = "fenced";
1350                 else if (dword & (1 << 1))
1351                         tiling = dword & (1 << 0) ? "Y" : "X";
1352                 type = " BAD";
1353                 format = "BAD";
1354                 switch ((dword>>7) & 0x7) {
1355                 case 1:
1356                     type = "8b";
1357                     switch ((dword>>3) & 0xf) {
1358                     case 0: format = "I"; break;
1359                     case 1: format = "L"; break;
1360                     case 4: format = "A"; break;
1361                     case 5: format = " mono"; break; }
1362                     break;
1363                 case 2:
1364                     type = "16b";
1365                     switch ((dword>>3) & 0xf) {
1366                     case 0: format = " rgb565"; break;
1367                     case 1: format = " argb1555"; break;
1368                     case 2: format = " argb4444"; break;
1369                     case 5: format = " ay88"; break;
1370                     case 6: format = " bump655"; break;
1371                     case 7: format = "I"; break;
1372                     case 8: format = "L"; break;
1373                     case 9: format = "A"; break; }
1374                     break;
1375                 case 3:
1376                     type = "32b";
1377                     switch ((dword>>3) & 0xf) {
1378                     case 0: format = " argb8888"; break;
1379                     case 1: format = " abgr8888"; break;
1380                     case 2: format = " xrgb8888"; break;
1381                     case 3: format = " xbgr8888"; break;
1382                     case 4: format = " qwvu8888"; break;
1383                     case 5: format = " axvu8888"; break;
1384                     case 6: format = " lxvu8888"; break;
1385                     case 7: format = " xlvu8888"; break;
1386                     case 8: format = " argb2101010"; break;
1387                     case 9: format = " abgr2101010"; break;
1388                     case 10: format = " awvu2101010"; break;
1389                     case 11: format = " gr1616"; break;
1390                     case 12: format = " vu1616"; break;
1391                     case 13: format = " xI824"; break;
1392                     case 14: format = " xA824"; break;
1393                     case 15: format = " xL824"; break; }
1394                     break;
1395                 case 5:
1396                     type = "422";
1397                     switch ((dword>>3) & 0xf) {
1398                     case 0: format = " yuv_swapy"; break;
1399                     case 1: format = " yuv"; break;
1400                     case 2: format = " yuv_swapuv"; break;
1401                     case 3: format = " yuv_swapuvy"; break; }
1402                     break;
1403                 case 6:
1404                     type = "compressed";
1405                     switch ((dword>>3) & 0x7) {
1406                     case 0: format = " dxt1"; break;
1407                     case 1: format = " dxt2_3"; break;
1408                     case 2: format = " dxt4_5"; break;
1409                     case 3: format = " fxt1"; break;
1410                     case 4: format = " dxt1_rb"; break; }
1411                     break;
1412                 case 7:
1413                     type = "4b indexed";
1414                     switch ((dword>>3) & 0xf) {
1415                     case 7: format = " argb8888"; break; }
1416                     break;
1417                 }
1418                 dword = data[i];
1419                 instr_out(data, hw_offset, i++, "map %d MS3 [width=%d, height=%d, format=%s%s, tiling=%s%s]\n",
1420                           map, width, height, type, format, tiling,
1421                           dword&(1<<9)?" palette select":"");
1422
1423                 dword = data[i];
1424                 pitch = 4*(((dword >> 21) & ((1 << 11) - 1))+1);
1425                 instr_out(data, hw_offset, i++, "map %d MS4 [pitch=%d, max_lod=%i, vol_depth=%i, cube_face_ena=%x, %s]\n",
1426                           map, pitch,
1427                           (dword>>9)&0x3f, dword&0xff, (dword>>15)&0x3f,
1428                           dword&(1<<8)?"miplayout legacy":"miplayout right");
1429             }
1430         }
1431         if (len != i) {
1432             fprintf(out, "Bad count in 3DSTATE_MAP_STATE\n");
1433             (*failures)++;
1434             return len;
1435         }
1436         return len;
1437     case 0x06:
1438         instr_out(data, hw_offset, 0, "3DSTATE_PIXEL_SHADER_CONSTANTS\n");
1439         len = (data[0] & 0x000000ff) + 2;
1440
1441         i = 2;
1442         for (c = 0; c <= 31; c++) {
1443             if (data[1] & (1 << c)) {
1444                 if (i + 4 >= count)
1445                     BUFFER_FAIL(count, len, "3DSTATE_PIXEL_SHADER_CONSTANTS");
1446                 instr_out(data, hw_offset, i, "C%d.X = %f\n",
1447                           c, int_as_float(data[i]));
1448                 i++;
1449                 instr_out(data, hw_offset, i, "C%d.Y = %f\n",
1450                           c, int_as_float(data[i]));
1451                 i++;
1452                 instr_out(data, hw_offset, i, "C%d.Z = %f\n",
1453                           c, int_as_float(data[i]));
1454                 i++;
1455                 instr_out(data, hw_offset, i, "C%d.W = %f\n",
1456                           c, int_as_float(data[i]));
1457                 i++;
1458             }
1459         }
1460         if (len != i) {
1461             fprintf(out, "Bad count in 3DSTATE_PIXEL_SHADER_CONSTANTS\n");
1462             (*failures)++;
1463         }
1464         return len;
1465     case 0x05:
1466         instr_out(data, hw_offset, 0, "3DSTATE_PIXEL_SHADER_PROGRAM\n");
1467         len = (data[0] & 0x000000ff) + 2;
1468         if ((len - 1) % 3 != 0 || len > 370) {
1469             fprintf(out, "Bad count in 3DSTATE_PIXEL_SHADER_PROGRAM\n");
1470             (*failures)++;
1471         }
1472         i = 1;
1473         for (instr = 0; instr < (len - 1) / 3; instr++) {
1474             char instr_prefix[10];
1475
1476             if (i + 3 >= count)
1477                 BUFFER_FAIL(count, len, "3DSTATE_PIXEL_SHADER_PROGRAM");
1478             sprintf(instr_prefix, "PS%03d", instr);
1479             i915_decode_instruction(data, hw_offset, i, instr_prefix);
1480             i += 3;
1481         }
1482         return len;
1483     case 0x01:
1484         if (IS_GEN2(devid))
1485                 break;
1486         instr_out(data, hw_offset, 0, "3DSTATE_SAMPLER_STATE\n");
1487         instr_out(data, hw_offset, 1, "mask\n");
1488         len = (data[0] & 0x0000003f) + 2;
1489         i = 2;
1490         for (sampler = 0; sampler <= 15; sampler++) {
1491             if (data[1] & (1 << sampler)) {
1492                 uint32_t dword;
1493                 char *mip_filter = "";
1494                 if (i + 3 >= count)
1495                     BUFFER_FAIL(count, len, "3DSTATE_SAMPLER_STATE");
1496                 dword = data[i];
1497                 switch ((dword>>20)&0x3) {
1498                 case 0: mip_filter = "none"; break;
1499                 case 1: mip_filter = "nearest"; break;
1500                 case 3: mip_filter = "linear"; break;
1501                 }
1502                 instr_out(data, hw_offset, i++, "sampler %d SS2:%s%s%s "
1503                           "base_mip_level=%i, mip_filter=%s, mag_filter=%s, min_filter=%s "
1504                           "lod_bias=%.2f,%s max_aniso=%i, shadow_func=%s\n", sampler,
1505                           dword&(1<<31)?" reverse gamma,":"",
1506                           dword&(1<<30)?" packed2planar,":"",
1507                           dword&(1<<29)?" colorspace conversion,":"",
1508                           (dword>>22)&0x1f,
1509                           mip_filter,
1510                           decode_sample_filter(dword>>17),
1511                           decode_sample_filter(dword>>14),
1512                           ((dword>>5)&0x1ff)/(0x10*1.0),
1513                           dword&(1<<4)?" shadow,":"",
1514                           dword&(1<<3)?4:2,
1515                           decode_compare_func(dword));
1516                 dword = data[i];
1517                 instr_out(data, hw_offset, i++, "sampler %d SS3: min_lod=%.2f,%s "
1518                           "tcmode_x=%s, tcmode_y=%s, tcmode_z=%s,%s texmap_idx=%i,%s\n",
1519                           sampler, ((dword>>24)&0xff)/(0x10*1.0),
1520                           dword&(1<<17)?" kill pixel enable,":"",
1521                           decode_tex_coord_mode(dword>>12),
1522                           decode_tex_coord_mode(dword>>9),
1523                           decode_tex_coord_mode(dword>>6),
1524                           dword&(1<<5)?" normalized coords,":"",
1525                           (dword>>1)&0xf,
1526                           dword&(1<<0)?" deinterlacer,":"");
1527                 dword = data[i];
1528                 instr_out(data, hw_offset, i++, "sampler %d SS4: border color\n",
1529                           sampler, ((dword>>24)&0xff)/(0x10*1.0),
1530                           dword);
1531             }
1532         }
1533         if (len != i) {
1534             fprintf(out, "Bad count in 3DSTATE_SAMPLER_STATE\n");
1535             (*failures)++;
1536         }
1537         return len;
1538     case 0x85:
1539         len = (data[0] & 0x0000000f) + 2;
1540
1541         if (len != 2)
1542             fprintf(out, "Bad count in 3DSTATE_DEST_BUFFER_VARIABLES\n");
1543         if (count < 2)
1544             BUFFER_FAIL(count, len, "3DSTATE_DEST_BUFFER_VARIABLES");
1545
1546         instr_out(data, hw_offset, 0,
1547                   "3DSTATE_DEST_BUFFER_VARIABLES\n");
1548
1549         switch ((data[1] >> 8) & 0xf) {
1550         case 0x0: format = "g8"; break;
1551         case 0x1: format = "x1r5g5b5"; break;
1552         case 0x2: format = "r5g6b5"; break;
1553         case 0x3: format = "a8r8g8b8"; break;
1554         case 0x4: format = "ycrcb_swapy"; break;
1555         case 0x5: format = "ycrcb_normal"; break;
1556         case 0x6: format = "ycrcb_swapuv"; break;
1557         case 0x7: format = "ycrcb_swapuvy"; break;
1558         case 0x8: format = "a4r4g4b4"; break;
1559         case 0x9: format = "a1r5g5b5"; break;
1560         case 0xa: format = "a2r10g10b10"; break;
1561         default: format = "BAD"; break;
1562         }
1563         switch ((data[1] >> 2) & 0x3) {
1564         case 0x0: zformat = "u16"; break;
1565         case 0x1: zformat = "f16"; break;
1566         case 0x2: zformat = "u24x8"; break;
1567         default: zformat = "BAD"; break;
1568         }
1569         instr_out(data, hw_offset, 1, "%s format, %s depth format, early Z %sabled\n",
1570                   format, zformat,
1571                   (data[1] & (1 << 31)) ? "en" : "dis");
1572         return len;
1573
1574     case 0x8e:
1575         {
1576             const char *name, *tiling;
1577
1578             len = (data[0] & 0x0000000f) + 2;
1579             if (len != 3)
1580                 fprintf(out, "Bad count in 3DSTATE_BUFFER_INFO\n");
1581             if (count < 3)
1582                 BUFFER_FAIL(count, len, "3DSTATE_BUFFER_INFO");
1583
1584             switch((data[1] >> 24) & 0x7) {
1585             case 0x3: name = "color"; break;
1586             case 0x7: name = "depth"; break;
1587             default: name = "unknown"; break;
1588             }
1589
1590             tiling = "none";
1591             if (data[1] & (1 << 23))
1592                 tiling = "fenced";
1593             else if (data[1] & (1 << 22))
1594                 tiling = data[1] & (1 << 21) ? "Y" : "X";
1595
1596             instr_out(data, hw_offset, 0, "3DSTATE_BUFFER_INFO\n");
1597             instr_out(data, hw_offset, 1, "%s, tiling = %s, pitch=%d\n", name, tiling, data[1]&0xffff);
1598
1599             instr_out(data, hw_offset, 2, "address\n");
1600             return len;
1601         }
1602     case 0x81:
1603         len = (data[0] & 0x0000000f) + 2;
1604
1605         if (len != 3)
1606             fprintf(out, "Bad count in 3DSTATE_SCISSOR_RECTANGLE\n");
1607         if (count < 3)
1608             BUFFER_FAIL(count, len, "3DSTATE_SCISSOR_RECTANGLE");
1609
1610         instr_out(data, hw_offset, 0,
1611                   "3DSTATE_SCISSOR_RECTANGLE\n");
1612         instr_out(data, hw_offset, 1, "(%d,%d)\n",
1613                   data[1] & 0xffff, data[1] >> 16);
1614         instr_out(data, hw_offset, 2, "(%d,%d)\n",
1615                   data[2] & 0xffff, data[2] >> 16);
1616
1617         return len;
1618     case 0x80:
1619         len = (data[0] & 0x0000000f) + 2;
1620
1621         if (len != 5)
1622             fprintf(out, "Bad count in 3DSTATE_DRAWING_RECTANGLE\n");
1623         if (count < 5)
1624             BUFFER_FAIL(count, len, "3DSTATE_DRAWING_RECTANGLE");
1625
1626         instr_out(data, hw_offset, 0,
1627                   "3DSTATE_DRAWING_RECTANGLE\n");
1628         instr_out(data, hw_offset, 1, "%s\n",
1629                   data[1]&(1<<30)?"depth ofs disabled ":"");
1630         instr_out(data, hw_offset, 2, "(%d,%d)\n",
1631                   data[2] & 0xffff, data[2] >> 16);
1632         instr_out(data, hw_offset, 3, "(%d,%d)\n",
1633                   data[3] & 0xffff, data[3] >> 16);
1634         instr_out(data, hw_offset, 4, "(%d,%d)\n",
1635                   data[4] & 0xffff, data[4] >> 16);
1636
1637         return len;
1638     case 0x9c:
1639         len = (data[0] & 0x0000000f) + 2;
1640
1641         if (len != 7)
1642             fprintf(out, "Bad count in 3DSTATE_CLEAR_PARAMETERS\n");
1643         if (count < 7)
1644             BUFFER_FAIL(count, len, "3DSTATE_CLEAR_PARAMETERS");
1645
1646         instr_out(data, hw_offset, 0,
1647                   "3DSTATE_CLEAR_PARAMETERS\n");
1648         instr_out(data, hw_offset, 1, "prim_type=%s, clear=%s%s%s\n",
1649                   data[1]&(1<<16)?"CLEAR_RECT":"ZONE_INIT",
1650                   data[1]&(1<<2)?"color,":"",
1651                   data[1]&(1<<1)?"depth,":"",
1652                   data[1]&(1<<0)?"stencil,":"");
1653         instr_out(data, hw_offset, 2, "clear color\n");
1654         instr_out(data, hw_offset, 3, "clear depth/stencil\n");
1655         instr_out(data, hw_offset, 4, "color value (rgba8888)\n");
1656         instr_out(data, hw_offset, 5, "depth value %f\n",
1657                   int_as_float(data[5]));
1658         instr_out(data, hw_offset, 6, "clear stencil\n");
1659         return len;
1660     }
1661
1662     for (idx = 0; idx < ARRAY_SIZE(opcodes_3d_1d); idx++)
1663     {
1664         opcode_3d_1d = &opcodes_3d_1d[idx];
1665         if (opcode_3d_1d->i830_only && !IS_GEN2(devid))
1666             continue;
1667
1668         if (((data[0] & 0x00ff0000) >> 16) == opcode_3d_1d->opcode) {
1669             len = 1;
1670
1671             instr_out(data, hw_offset, 0, "%s\n", opcode_3d_1d->name);
1672             if (opcode_3d_1d->max_len > 1) {
1673                 len = (data[0] & 0x0000ffff) + 2;
1674                 if (len < opcode_3d_1d->min_len ||
1675                     len > opcode_3d_1d->max_len)
1676                 {
1677                     fprintf(out, "Bad count in %s\n",
1678                             opcode_3d_1d->name);
1679                     (*failures)++;
1680                 }
1681             }
1682
1683             for (i = 1; i < len; i++) {
1684                 if (i >= count)
1685                     BUFFER_FAIL(count, len,  opcode_3d_1d->name);
1686                 instr_out(data, hw_offset, i, "dword %d\n", i);
1687             }
1688
1689             return len;
1690         }
1691     }
1692
1693     instr_out(data, hw_offset, 0, "3D UNKNOWN: 3d_1d opcode = 0x%x\n", opcode);
1694     (*failures)++;
1695     return 1;
1696 }
1697
1698 static int
1699 decode_3d_primitive(uint32_t *data, int count, uint32_t hw_offset,
1700                     int *failures)
1701 {
1702     char immediate = (data[0] & (1 << 23)) == 0;
1703     unsigned int len, i, j, ret;
1704     char *primtype;
1705     int original_s2 = saved_s2;
1706     int original_s4 = saved_s4;
1707
1708     switch ((data[0] >> 18) & 0xf) {
1709     case 0x0: primtype = "TRILIST"; break;
1710     case 0x1: primtype = "TRISTRIP"; break;
1711     case 0x2: primtype = "TRISTRIP_REVERSE"; break;
1712     case 0x3: primtype = "TRIFAN"; break;
1713     case 0x4: primtype = "POLYGON"; break;
1714     case 0x5: primtype = "LINELIST"; break;
1715     case 0x6: primtype = "LINESTRIP"; break;
1716     case 0x7: primtype = "RECTLIST"; break;
1717     case 0x8: primtype = "POINTLIST"; break;
1718     case 0x9: primtype = "DIB"; break;
1719     case 0xa: primtype = "CLEAR_RECT"; saved_s4 = 3 << 6; saved_s2 = ~0; break;
1720     default: primtype = "unknown"; break;
1721     }
1722
1723     /* XXX: 3DPRIM_DIB not supported */
1724     if (immediate) {
1725         len = (data[0] & 0x0003ffff) + 2;
1726         instr_out(data, hw_offset, 0, "3DPRIMITIVE inline %s\n", primtype);
1727         if (count < len)
1728             BUFFER_FAIL(count, len,  "3DPRIMITIVE inline");
1729         if (!saved_s2_set || !saved_s4_set) {
1730             fprintf(out, "unknown vertex format\n");
1731             for (i = 1; i < len; i++) {
1732                 instr_out(data, hw_offset, i,
1733                           "           vertex data (%f float)\n",
1734                           int_as_float(data[i]));
1735             }
1736         } else {
1737             unsigned int vertex = 0;
1738             for (i = 1; i < len;) {
1739                 unsigned int tc;
1740
1741 #define VERTEX_OUT(fmt, ...) do {                                       \
1742     if (i < len)                                                        \
1743         instr_out(data, hw_offset, i, " V%d."fmt"\n", vertex, __VA_ARGS__); \
1744     else                                                                \
1745         fprintf(out, " missing data in V%d\n", vertex);                 \
1746     i++;                                                                \
1747 } while (0)
1748
1749                 VERTEX_OUT("X = %f", int_as_float(data[i]));
1750                 VERTEX_OUT("Y = %f", int_as_float(data[i]));
1751                 switch (saved_s4 >> 6 & 0x7) {
1752                 case 0x1:
1753                     VERTEX_OUT("Z = %f", int_as_float(data[i]));
1754                     break;
1755                 case 0x2:
1756                     VERTEX_OUT("Z = %f", int_as_float(data[i]));
1757                     VERTEX_OUT("W = %f", int_as_float(data[i]));
1758                     break;
1759                 case 0x3:
1760                     break;
1761                 case 0x4:
1762                     VERTEX_OUT("W = %f", int_as_float(data[i]));
1763                     break;
1764                 default:
1765                     fprintf(out, "bad S4 position mask\n");
1766                 }
1767
1768                 if (saved_s4 & (1 << 10)) {
1769                     VERTEX_OUT("color = (A=0x%02x, R=0x%02x, G=0x%02x, "
1770                                "B=0x%02x)",
1771                                data[i] >> 24,
1772                                (data[i] >> 16) & 0xff,
1773                                (data[i] >> 8) & 0xff,
1774                                data[i] & 0xff);
1775                 }
1776                 if (saved_s4 & (1 << 11)) {
1777                     VERTEX_OUT("spec = (A=0x%02x, R=0x%02x, G=0x%02x, "
1778                                "B=0x%02x)",
1779                                data[i] >> 24,
1780                                (data[i] >> 16) & 0xff,
1781                                (data[i] >> 8) & 0xff,
1782                                data[i] & 0xff);
1783                 }
1784                 if (saved_s4 & (1 << 12))
1785                     VERTEX_OUT("width = 0x%08x)", data[i]);
1786
1787                 for (tc = 0; tc <= 7; tc++) {
1788                     switch ((saved_s2 >> (tc * 4)) & 0xf) {
1789                     case 0x0:
1790                         VERTEX_OUT("T%d.X = %f", tc, int_as_float(data[i]));
1791                         VERTEX_OUT("T%d.Y = %f", tc, int_as_float(data[i]));
1792                         break;
1793                     case 0x1:
1794                         VERTEX_OUT("T%d.X = %f", tc, int_as_float(data[i]));
1795                         VERTEX_OUT("T%d.Y = %f", tc, int_as_float(data[i]));
1796                         VERTEX_OUT("T%d.Z = %f", tc, int_as_float(data[i]));
1797                         break;
1798                     case 0x2:
1799                         VERTEX_OUT("T%d.X = %f", tc, int_as_float(data[i]));
1800                         VERTEX_OUT("T%d.Y = %f", tc, int_as_float(data[i]));
1801                         VERTEX_OUT("T%d.Z = %f", tc, int_as_float(data[i]));
1802                         VERTEX_OUT("T%d.W = %f", tc, int_as_float(data[i]));
1803                         break;
1804                     case 0x3:
1805                         VERTEX_OUT("T%d.X = %f", tc, int_as_float(data[i]));
1806                         break;
1807                     case 0x4:
1808                         VERTEX_OUT("T%d.XY = 0x%08x half-float", tc, data[i]);
1809                         break;
1810                     case 0x5:
1811                         VERTEX_OUT("T%d.XY = 0x%08x half-float", tc, data[i]);
1812                         VERTEX_OUT("T%d.ZW = 0x%08x half-float", tc, data[i]);
1813                         break;
1814                     case 0xf:
1815                         break;
1816                     default:
1817                         fprintf(out, "bad S2.T%d format\n", tc);
1818                     }
1819                 }
1820                 vertex++;
1821             }
1822         }
1823
1824         ret = len;
1825     } else {
1826         /* indirect vertices */
1827         len = data[0] & 0x0000ffff; /* index count */
1828         if (data[0] & (1 << 17)) {
1829             /* random vertex access */
1830             if (count < (len + 1) / 2 + 1) {
1831                 BUFFER_FAIL(count, (len + 1) / 2 + 1,
1832                             "3DPRIMITIVE random indirect");
1833             }
1834             instr_out(data, hw_offset, 0,
1835                       "3DPRIMITIVE random indirect %s (%d)\n", primtype, len);
1836             if (len == 0) {
1837                 /* vertex indices continue until 0xffff is found */
1838                 for (i = 1; i < count; i++) {
1839                     if ((data[i] & 0xffff) == 0xffff) {
1840                         instr_out(data, hw_offset, i,
1841                                   "    indices: (terminator)\n");
1842                         ret = i;
1843                         goto out;
1844                     } else if ((data[i] >> 16) == 0xffff) {
1845                         instr_out(data, hw_offset, i,
1846                                   "    indices: 0x%04x, (terminator)\n",
1847                                   data[i] & 0xffff);
1848                         ret = i;
1849                         goto out;
1850                     } else {
1851                         instr_out(data, hw_offset, i,
1852                                   "    indices: 0x%04x, 0x%04x\n",
1853                                   data[i] & 0xffff, data[i] >> 16);
1854                     }
1855                 }
1856                 fprintf(out,
1857                         "3DPRIMITIVE: no terminator found in index buffer\n");
1858                 (*failures)++;
1859                 ret = count;
1860                 goto out;
1861             } else {
1862                 /* fixed size vertex index buffer */
1863                 for (j = 1, i = 0; i < len; i += 2, j++) {
1864                     if (i * 2 == len - 1) {
1865                         instr_out(data, hw_offset, j,
1866                                   "    indices: 0x%04x\n",
1867                                   data[j] & 0xffff);
1868                     } else {
1869                         instr_out(data, hw_offset, j,
1870                                   "    indices: 0x%04x, 0x%04x\n",
1871                                   data[j] & 0xffff, data[j] >> 16);
1872                     }
1873                 }
1874             }
1875             ret = (len + 1) / 2 + 1;
1876             goto out;
1877         } else {
1878             /* sequential vertex access */
1879             if (count < 2)
1880                 BUFFER_FAIL(count, 2, "3DPRIMITIVE seq indirect");
1881             instr_out(data, hw_offset, 0,
1882                       "3DPRIMITIVE sequential indirect %s, %d starting from "
1883                       "%d\n", primtype, len, data[1] & 0xffff);
1884             instr_out(data, hw_offset, 1, "           start\n");
1885             ret = 2;
1886             goto out;
1887         }
1888     }
1889
1890 out:
1891     saved_s2 = original_s2;
1892     saved_s4 = original_s4;
1893     return ret;
1894 }
1895
1896 static int
1897 decode_3d(uint32_t *data, int count, uint32_t hw_offset, uint32_t devid, int *failures)
1898 {
1899     uint32_t opcode;
1900     unsigned int idx;
1901
1902     struct {
1903         uint32_t opcode;
1904         int min_len;
1905         int max_len;
1906         char *name;
1907     } opcodes_3d[] = {
1908         { 0x06, 1, 1, "3DSTATE_ANTI_ALIASING" },
1909         { 0x08, 1, 1, "3DSTATE_BACKFACE_STENCIL_OPS" },
1910         { 0x09, 1, 1, "3DSTATE_BACKFACE_STENCIL_MASKS" },
1911         { 0x16, 1, 1, "3DSTATE_COORD_SET_BINDINGS" },
1912         { 0x15, 1, 1, "3DSTATE_FOG_COLOR" },
1913         { 0x0b, 1, 1, "3DSTATE_INDEPENDENT_ALPHA_BLEND" },
1914         { 0x0d, 1, 1, "3DSTATE_MODES_4" },
1915         { 0x0c, 1, 1, "3DSTATE_MODES_5" },
1916         { 0x07, 1, 1, "3DSTATE_RASTERIZATION_RULES" },
1917     }, *opcode_3d;
1918
1919     opcode = (data[0] & 0x1f000000) >> 24;
1920
1921     switch (opcode) {
1922     case 0x1f:
1923         return decode_3d_primitive(data, count, hw_offset, failures);
1924     case 0x1d:
1925         return decode_3d_1d(data, count, hw_offset, devid, failures);
1926     case 0x1c:
1927         return decode_3d_1c(data, count, hw_offset, failures);
1928     }
1929
1930     for (idx = 0; idx < ARRAY_SIZE(opcodes_3d); idx++) {
1931         opcode_3d = &opcodes_3d[idx];
1932         if (opcode == opcode_3d->opcode) {
1933             unsigned int len = 1, i;
1934
1935             instr_out(data, hw_offset, 0, "%s\n", opcode_3d->name);
1936             if (opcode_3d->max_len > 1) {
1937                 len = (data[0] & 0xff) + 2;
1938                 if (len < opcode_3d->min_len ||
1939                     len > opcode_3d->max_len)
1940                 {
1941                     fprintf(out, "Bad count in %s\n", opcode_3d->name);
1942                 }
1943             }
1944
1945             for (i = 1; i < len; i++) {
1946                 if (i >= count)
1947                     BUFFER_FAIL(count, len, opcode_3d->name);
1948                 instr_out(data, hw_offset, i, "dword %d\n", i);
1949             }
1950             return len;
1951         }
1952     }
1953
1954     instr_out(data, hw_offset, 0, "3D UNKNOWN: 3d opcode = 0x%x\n", opcode);
1955     (*failures)++;
1956     return 1;
1957 }
1958
1959 static const char *
1960 get_965_surfacetype(unsigned int surfacetype)
1961 {
1962     switch (surfacetype) {
1963     case 0: return "1D";
1964     case 1: return "2D";
1965     case 2: return "3D";
1966     case 3: return "CUBE";
1967     case 4: return "BUFFER";
1968     case 7: return "NULL";
1969     default: return "unknown";
1970     }
1971 }
1972
1973 static const char *
1974 get_965_depthformat(unsigned int depthformat)
1975 {
1976     switch (depthformat) {
1977     case 0: return "s8_z24float";
1978     case 1: return "z32float";
1979     case 2: return "z24s8";
1980     case 5: return "z16";
1981     default: return "unknown";
1982     }
1983 }
1984
1985 static const char *
1986 get_965_element_component(uint32_t data, int component)
1987 {
1988     uint32_t component_control = (data >> (16 + (3 - component) * 4)) & 0x7;
1989
1990     switch (component_control) {
1991     case 0:
1992         return "nostore";
1993     case 1:
1994         switch (component) {
1995         case 0: return "X";
1996         case 1: return "Y";
1997         case 2: return "Z";
1998         case 3: return "W";
1999         default: return "fail";
2000         }
2001     case 2:
2002         return "0.0";
2003     case 3:
2004         return "1.0";
2005     case 4:
2006         return "0x1";
2007     case 5:
2008         return "VID";
2009     default:
2010         return "fail";
2011     }
2012 }
2013
2014 static const char *
2015 get_965_prim_type(uint32_t data)
2016 {
2017     uint32_t primtype = (data >> 10) & 0x1f;
2018
2019     switch (primtype) {
2020     case 0x01: return "point list";
2021     case 0x02: return "line list";
2022     case 0x03: return "line strip";
2023     case 0x04: return "tri list";
2024     case 0x05: return "tri strip";
2025     case 0x06: return "tri fan";
2026     case 0x07: return "quad list";
2027     case 0x08: return "quad strip";
2028     case 0x09: return "line list adj";
2029     case 0x0a: return "line strip adj";
2030     case 0x0b: return "tri list adj";
2031     case 0x0c: return "tri strip adj";
2032     case 0x0d: return "tri strip reverse";
2033     case 0x0e: return "polygon";
2034     case 0x0f: return "rect list";
2035     case 0x10: return "line loop";
2036     case 0x11: return "point list bf";
2037     case 0x12: return "line strip cont";
2038     case 0x13: return "line strip bf";
2039     case 0x14: return "line strip cont bf";
2040     case 0x15: return "tri fan no stipple";
2041     default: return "fail";
2042     }
2043 }
2044
2045 static int
2046 i965_decode_urb_fence(uint32_t *data, uint32_t hw_offset, int len, int count,
2047                       int *failures)
2048 {
2049         uint32_t vs_fence, clip_fence, gs_fence, sf_fence, vfe_fence, cs_fence;
2050
2051         if (len != 3)
2052             fprintf(out, "Bad count in URB_FENCE\n");
2053         if (count < 3)
2054             BUFFER_FAIL(count, len, "URB_FENCE");
2055
2056         vs_fence = data[1] & 0x3ff;
2057         gs_fence = (data[1] >> 10) & 0x3ff;
2058         clip_fence = (data[1] >> 20) & 0x3ff;
2059         sf_fence = data[2] & 0x3ff;
2060         vfe_fence = (data[2] >> 10) & 0x3ff;
2061         cs_fence = (data[2] >> 20) & 0x7ff;
2062
2063         instr_out(data, hw_offset, 0, "URB_FENCE: %s%s%s%s%s%s\n",
2064                         (data[0] >> 13) & 1 ? "cs " : "",
2065                         (data[0] >> 12) & 1 ? "vfe " : "",
2066                         (data[0] >> 11) & 1 ? "sf " : "",
2067                         (data[0] >> 10) & 1 ? "clip " : "",
2068                         (data[0] >> 9)  & 1 ? "gs " : "",
2069                         (data[0] >> 8)  & 1 ? "vs " : "");
2070         instr_out(data, hw_offset, 1,
2071                   "vs fence: %d, clip_fence: %d, gs_fence: %d\n",
2072                   vs_fence, clip_fence, gs_fence);
2073         instr_out(data, hw_offset, 2,
2074                   "sf fence: %d, vfe_fence: %d, cs_fence: %d\n",
2075                   sf_fence, vfe_fence, cs_fence);
2076         if (gs_fence < vs_fence)
2077             fprintf(out, "gs fence < vs fence!\n");
2078         if (clip_fence < gs_fence)
2079             fprintf(out, "clip fence < gs fence!\n");
2080         if (sf_fence < clip_fence)
2081             fprintf(out, "sf fence < clip fence!\n");
2082         if (cs_fence < sf_fence)
2083             fprintf(out, "cs fence < sf fence!\n");
2084
2085         return len;
2086 }
2087
2088 static void
2089 state_base_out(uint32_t *data, uint32_t hw_offset, unsigned int index,
2090                char *name)
2091 {
2092     if (data[index] & 1) {
2093         instr_out(data, hw_offset, index, "%s state base address 0x%08x\n",
2094                   name, data[index] & ~1);
2095     } else {
2096         instr_out(data, hw_offset, index, "%s state base not updated\n",
2097                   name);
2098     }
2099 }
2100
2101 static void
2102 state_max_out(uint32_t *data, uint32_t hw_offset, unsigned int index,
2103               char *name)
2104 {
2105     if (data[index] & 1) {
2106         if (data[index] == 1) {
2107             instr_out(data, hw_offset, index,
2108                       "%s state upper bound disabled\n", name);
2109         } else {
2110             instr_out(data, hw_offset, index, "%s state upper bound 0x%08x\n",
2111                       name, data[index] & ~1);
2112         }
2113     } else {
2114         instr_out(data, hw_offset, index, "%s state upper bound not updated\n",
2115                   name);
2116     }
2117 }
2118
2119 static int
2120 decode_3d_965(uint32_t *data, int count, uint32_t hw_offset, uint32_t devid, int *failures)
2121 {
2122     uint32_t opcode;
2123     unsigned int idx, len;
2124     int i, sba_len;
2125     char *desc1 = NULL;
2126
2127     struct {
2128         uint32_t opcode;
2129         int min_len;
2130         int max_len;
2131         char *name;
2132     } opcodes_3d[] = {
2133         { 0x6000, 3, 3, "URB_FENCE" },
2134         { 0x6001, 2, 2, "CS_URB_STATE" },
2135         { 0x6002, 2, 2, "CONSTANT_BUFFER" },
2136         { 0x6101, 6, 6, "STATE_BASE_ADDRESS" },
2137         { 0x6102, 2, 2 , "STATE_SIP" },
2138         { 0x6104, 1, 1, "3DSTATE_PIPELINE_SELECT" },
2139         { 0x680b, 1, 1, "3DSTATE_VF_STATISTICS" },
2140         { 0x6904, 1, 1, "3DSTATE_PIPELINE_SELECT" },
2141         { 0x7800, 7, 7, "3DSTATE_PIPELINED_POINTERS" },
2142         { 0x7801, 6, 6, "3DSTATE_BINDING_TABLE_POINTERS" },
2143         { 0x7808, 5, 257, "3DSTATE_VERTEX_BUFFERS" },
2144         { 0x7809, 3, 256, "3DSTATE_VERTEX_ELEMENTS" },
2145         { 0x780a, 3, 3, "3DSTATE_INDEX_BUFFER" },
2146         { 0x780b, 1, 1, "3DSTATE_VF_STATISTICS" },
2147         { 0x7900, 4, 4, "3DSTATE_DRAWING_RECTANGLE" },
2148         { 0x7901, 5, 5, "3DSTATE_CONSTANT_COLOR" },
2149         { 0x7905, 5, 7, "3DSTATE_DEPTH_BUFFER" },
2150         { 0x7906, 2, 2, "3DSTATE_POLY_STIPPLE_OFFSET" },
2151         { 0x7907, 33, 33, "3DSTATE_POLY_STIPPLE_PATTERN" },
2152         { 0x7908, 3, 3, "3DSTATE_LINE_STIPPLE" },
2153         { 0x7909, 2, 2, "3DSTATE_GLOBAL_DEPTH_OFFSET_CLAMP" },
2154         { 0x7909, 2, 2, "3DSTATE_CLEAR_PARAMS" },
2155         { 0x790a, 3, 3, "3DSTATE_AA_LINE_PARAMETERS" },
2156         { 0x790b, 4, 4, "3DSTATE_GS_SVB_INDEX" },
2157         { 0x790d, 3, 3, "3DSTATE_MULTISAMPLE" },
2158         { 0x7910, 2, 2, "3DSTATE_CLEAR_PARAMS" },
2159         { 0x7b00, 6, 6, "3DPRIMITIVE" },
2160         { 0x7802, 4, 4, "3DSTATE_SAMPLER_STATE_POINTERS" },
2161         { 0x7805, 3, 3, "3DSTATE_URB" },
2162         { 0x780d, 4, 4, "3DSTATE_VIEWPORT_STATE_POINTERS" },
2163         { 0x780e, 4, 4, "3DSTATE_CC_STATE_POINTERS" },
2164         { 0x780f, 2, 2, "3DSTATE_SCISSOR_STATE_POINTERS" },
2165         { 0x7810, 6, 6, "3DSTATE_VS_STATE" },
2166         { 0x7811, 7, 7, "3DSTATE_GS_STATE" },
2167         { 0x7812, 4, 4, "3DSTATE_CLIP_STATE" },
2168         { 0x7813, 20, 20, "3DSTATE_SF_STATE" },
2169         { 0x7814, 9, 9, "3DSTATE_WM_STATE" },
2170         { 0x7815, 5, 5, "3DSTATE_CONSTANT_VS_STATE" },
2171         { 0x7816, 5, 5, "3DSTATE_CONSTANT_GS_STATE" },
2172         { 0x7817, 5, 5, "3DSTATE_CONSTANT_PS_STATE" },
2173         { 0x7818, 2, 2, "3DSTATE_SAMPLE_MASK" },
2174    }, *opcode_3d;
2175
2176     len = (data[0] & 0x0000ffff) + 2;
2177
2178     opcode = (data[0] & 0xffff0000) >> 16;
2179     switch (opcode) {
2180     case 0x6000:
2181         len = (data[0] & 0x000000ff) + 2;
2182         return i965_decode_urb_fence(data, hw_offset, len, count, failures);
2183     case 0x6001:
2184         instr_out(data, hw_offset, 0, "CS_URB_STATE\n");
2185         instr_out(data, hw_offset, 1, "entry_size: %d [%d bytes], n_entries: %d\n",
2186                         (data[1] >> 4) & 0x1f,
2187                         (((data[1] >> 4) & 0x1f) + 1) * 64,
2188                         data[1] & 0x7);
2189         return len;
2190     case 0x6002:
2191         len = (data[0] & 0x000000ff) + 2;
2192         instr_out(data, hw_offset, 0, "CONSTANT_BUFFER: %s\n",
2193                         (data[0] >> 8) & 1 ? "valid" : "invalid");
2194         instr_out(data, hw_offset, 1, "offset: 0x%08x, length: %d bytes\n",
2195                         data[1] & ~0x3f, ((data[1] & 0x3f) + 1) * 64);
2196         return len;
2197     case 0x6101:
2198         i = 0;
2199         instr_out(data, hw_offset, 0,
2200                   "STATE_BASE_ADDRESS\n");
2201         i++;
2202
2203         if (IS_GEN6(devid) || IS_GEN7(devid))
2204             sba_len = 10;
2205         else if (IS_GEN5(devid))
2206             sba_len = 8;
2207         else
2208             sba_len = 6;
2209         if (len != sba_len)
2210             fprintf(out, "Bad count in STATE_BASE_ADDRESS\n");
2211         if (len != sba_len)
2212             BUFFER_FAIL(count, len, "STATE_BASE_ADDRESS");
2213
2214         state_base_out(data, hw_offset, i++, "general");
2215         state_base_out(data, hw_offset, i++, "surface");
2216         if (IS_GEN6(devid) || IS_GEN7(devid))
2217             state_base_out(data, hw_offset, i++, "dynamic");
2218         state_base_out(data, hw_offset, i++, "indirect");
2219         if (IS_GEN5(devid) || IS_GEN6(devid) || IS_GEN7(devid))
2220             state_base_out(data, hw_offset, i++, "instruction");
2221
2222         state_max_out(data, hw_offset, i++, "general");
2223         if (IS_GEN6(devid) || IS_GEN7(devid))
2224             state_max_out(data, hw_offset, i++, "dynamic");
2225         state_max_out(data, hw_offset, i++, "indirect");
2226         if (IS_GEN5(devid) || IS_GEN6(devid) || IS_GEN7(devid))
2227             state_max_out(data, hw_offset, i++, "instruction");
2228
2229         return len;
2230     case 0x7800:
2231         if (len != 7)
2232             fprintf(out, "Bad count in 3DSTATE_PIPELINED_POINTERS\n");
2233         if (count < 7)
2234             BUFFER_FAIL(count, len, "3DSTATE_PIPELINED_POINTERS");
2235
2236         instr_out(data, hw_offset, 0,
2237                   "3DSTATE_PIPELINED_POINTERS\n");
2238         instr_out(data, hw_offset, 1, "VS state\n");
2239         instr_out(data, hw_offset, 2, "GS state\n");
2240         instr_out(data, hw_offset, 3, "Clip state\n");
2241         instr_out(data, hw_offset, 4, "SF state\n");
2242         instr_out(data, hw_offset, 5, "WM state\n");
2243         instr_out(data, hw_offset, 6, "CC state\n");
2244         return len;
2245     case 0x7801:
2246         len = (data[0] & 0x000000ff) + 2;
2247         if (len != 6 && len != 4)
2248             fprintf(out, "Bad count in 3DSTATE_BINDING_TABLE_POINTERS\n");
2249         if (len == 6) {
2250             if (count < 6)
2251                 BUFFER_FAIL(count, len, "3DSTATE_BINDING_TABLE_POINTERS");
2252             instr_out(data, hw_offset, 0,
2253                       "3DSTATE_BINDING_TABLE_POINTERS\n");
2254             instr_out(data, hw_offset, 1, "VS binding table\n");
2255             instr_out(data, hw_offset, 2, "GS binding table\n");
2256             instr_out(data, hw_offset, 3, "Clip binding table\n");
2257             instr_out(data, hw_offset, 4, "SF binding table\n");
2258             instr_out(data, hw_offset, 5, "WM binding table\n");
2259         } else {
2260             if (count < 4)
2261                 BUFFER_FAIL(count, len, "3DSTATE_BINDING_TABLE_POINTERS");
2262
2263             instr_out(data, hw_offset, 0,
2264                       "3DSTATE_BINDING_TABLE_POINTERS: VS mod %d, "
2265                       "GS mod %d, PS mod %d\n",
2266                       (data[0] & (1 << 8)) != 0,
2267                       (data[0] & (1 << 9)) != 0,
2268                       (data[0] & (1 << 12)) != 0);
2269             instr_out(data, hw_offset, 1, "VS binding table\n");
2270             instr_out(data, hw_offset, 2, "GS binding table\n");
2271             instr_out(data, hw_offset, 3, "WM binding table\n");
2272         }
2273
2274         return len;
2275     case 0x7802:
2276         len = (data[0] & 0xff) + 2;
2277         if (len != 4)
2278             fprintf(out, "Bad count in 3DSTATE_SAMPLER_STATE_POINTERS\n");
2279         if (count < 4)
2280             BUFFER_FAIL(count, len, "3DSTATE_SAMPLER_STATE_POINTERS");
2281         instr_out(data, hw_offset, 0, "3DSTATE_SAMPLER_STATE_POINTERS: VS mod %d, "
2282                   "GS mod %d, PS mod %d\n",
2283                   (data[0] & (1 << 8)) != 0,
2284                   (data[0] & (1 << 9)) != 0,
2285                   (data[0] & (1 << 12)) != 0);
2286         instr_out(data, hw_offset, 1, "VS sampler state\n");
2287         instr_out(data, hw_offset, 2, "GS sampler state\n");
2288         instr_out(data, hw_offset, 3, "WM sampler state\n");
2289         return len;
2290     case 0x7805:
2291         len = (data[0] & 0xff) + 2;
2292         if (len != 3)
2293             fprintf(out, "Bad count in 3DSTATE_URB\n");
2294         if (count < 3)
2295             BUFFER_FAIL(count, len, "3DSTATE_URB");
2296         instr_out(data, hw_offset, 0, "3DSTATE_URB\n");
2297         instr_out(data, hw_offset, 1, "VS entries %d, alloc size %d (1024bit row)\n",
2298                         data[1] & 0xffff, ((data[1] >> 16) & 0x07f) + 1);
2299         instr_out(data, hw_offset, 2, "GS entries %d, alloc size %d (1024bit row)\n",
2300                         (data[2] >> 8) & 0x3ff, (data[2] & 7) + 1);
2301         return len;
2302
2303     case 0x7808:
2304         len = (data[0] & 0xff) + 2;
2305         if ((len - 1) % 4 != 0)
2306             fprintf(out, "Bad count in 3DSTATE_VERTEX_BUFFERS\n");
2307         if (count < len)
2308             BUFFER_FAIL(count, len, "3DSTATE_VERTEX_BUFFERS");
2309         instr_out(data, hw_offset, 0, "3DSTATE_VERTEX_BUFFERS\n");
2310
2311         for (i = 1; i < len;) {
2312             int idx, access;
2313             if (IS_GEN6(devid)) {
2314                 idx = 26;
2315                 access = 20;
2316             } else {
2317                 idx = 27;
2318                 access = 26;
2319             }
2320             instr_out(data, hw_offset, i, "buffer %d: %s, pitch %db\n",
2321                       data[i] >> idx,
2322                       data[i] & (1 << access) ? "random" : "sequential",
2323                       data[i] & 0x07ff);
2324             i++;
2325             instr_out(data, hw_offset, i++, "buffer address\n");
2326             instr_out(data, hw_offset, i++, "max index\n");
2327             instr_out(data, hw_offset, i++, "mbz\n");
2328         }
2329         return len;
2330
2331     case 0x7809:
2332         len = (data[0] & 0xff) + 2;
2333         if ((len + 1) % 2 != 0)
2334             fprintf(out, "Bad count in 3DSTATE_VERTEX_ELEMENTS\n");
2335         if (count < len)
2336             BUFFER_FAIL(count, len, "3DSTATE_VERTEX_ELEMENTS");
2337         instr_out(data, hw_offset, 0, "3DSTATE_VERTEX_ELEMENTS\n");
2338
2339         for (i = 1; i < len;) {
2340             instr_out(data, hw_offset, i, "buffer %d: %svalid, type 0x%04x, "
2341                       "src offset 0x%04x bytes\n",
2342                       data[i] >> (IS_GEN6(devid) ? 26 : 27),
2343                       data[i] & (1 << (IS_GEN6(devid) ? 25 : 26)) ? "" : "in",
2344                       (data[i] >> 16) & 0x1ff,
2345                       data[i] & 0x07ff);
2346             i++;
2347             instr_out(data, hw_offset, i, "(%s, %s, %s, %s), "
2348                       "dst offset 0x%02x bytes\n",
2349                       get_965_element_component(data[i], 0),
2350                       get_965_element_component(data[i], 1),
2351                       get_965_element_component(data[i], 2),
2352                       get_965_element_component(data[i], 3),
2353                       (data[i] & 0xff) * 4);
2354             i++;
2355         }
2356         return len;
2357
2358     case 0x780d:
2359         len = (data[0] & 0xff) + 2;
2360         if (len != 4)
2361             fprintf(out, "Bad count in 3DSTATE_VIEWPORT_STATE_POINTERS\n");
2362         if (count < len)
2363             BUFFER_FAIL(count, len, "3DSTATE_VIEWPORT_STATE_POINTERS");
2364         instr_out(data, hw_offset, 0, "3DSTATE_VIEWPORT_STATE_POINTERS\n");
2365         instr_out(data, hw_offset, 1, "clip\n");
2366         instr_out(data, hw_offset, 2, "sf\n");
2367         instr_out(data, hw_offset, 3, "cc\n");
2368         return len;
2369
2370     case 0x780a:
2371         len = (data[0] & 0xff) + 2;
2372         if (len != 3)
2373             fprintf(out, "Bad count in 3DSTATE_INDEX_BUFFER\n");
2374         if (count < len)
2375             BUFFER_FAIL(count, len, "3DSTATE_INDEX_BUFFER");
2376         instr_out(data, hw_offset, 0, "3DSTATE_INDEX_BUFFER\n");
2377         instr_out(data, hw_offset, 1, "beginning buffer address\n");
2378         instr_out(data, hw_offset, 2, "ending buffer address\n");
2379         return len;
2380
2381     case 0x780e:
2382         len = (data[0] & 0xff) + 2;
2383         if (len != 4)
2384             fprintf(out, "Bad count in 3DSTATE_CC_STATE_POINTERS\n");
2385         if (count < 4)
2386             BUFFER_FAIL(count, len, "3DSTATE_CC_STATE_POINTERS");
2387         instr_out(data, hw_offset, 0, "3DSTATE_CC_STATE_POINTERS\n");
2388         instr_out(data, hw_offset, 1, "blend change %d\n", data[1] & 1);
2389         instr_out(data, hw_offset, 2, "depth stencil change %d\n", data[2] & 1);
2390         instr_out(data, hw_offset, 3, "cc change %d\n", data[3] & 1);
2391         return len;
2392
2393     case 0x780f:
2394         len = (data[0] & 0xff) + 2;
2395         if (len != 2)
2396             fprintf(out, "Bad count in 3DSTATE_SCISSOR_POINTERS\n");
2397         if (count < 2)
2398             BUFFER_FAIL(count, len, "3DSTATE_SCISSOR_POINTERS");
2399         instr_out(data, hw_offset, 0, "3DSTATE_SCISSOR_POINTERS\n");
2400         instr_out(data, hw_offset, 1, "scissor rect offset\n");
2401         return len;
2402
2403     case 0x7810:
2404         len = (data[0] & 0xff) + 2;
2405         if (len != 6)
2406             fprintf(out, "Bad count in 3DSTATE_VS\n");
2407         if (count < 6)
2408             BUFFER_FAIL(count, len, "3DSTATE_VS");
2409         instr_out(data, hw_offset, 0, "3DSTATE_VS\n");
2410         instr_out(data, hw_offset, 1, "kernel pointer\n");
2411         instr_out(data, hw_offset, 2, "SPF=%d, VME=%d, Sampler Count %d, "
2412                   "Binding table count %d\n",
2413                   (data[2] >> 31) & 1,
2414                   (data[2] >> 30) & 1,
2415                   (data[2] >> 27) & 7,
2416                   (data[2] >> 18) & 0xff);
2417         instr_out(data, hw_offset, 3, "scratch offset\n");
2418         instr_out(data, hw_offset, 4, "Dispatch GRF start %d, VUE read length %d, "
2419                   "VUE read offset %d\n",
2420                   (data[4] >> 20) & 0x1f,
2421                   (data[4] >> 11) & 0x3f,
2422                   (data[4] >> 4) & 0x3f);
2423         instr_out(data, hw_offset, 5, "Max Threads %d, Vertex Cache %sable, "
2424                   "VS func %sable\n",
2425                   ((data[5] >> 25) & 0x7f) + 1,
2426                   (data[5] & (1 << 1)) != 0 ? "dis" : "en",
2427                   (data[5] & 1) != 0 ? "en" : "dis");
2428         return len;
2429
2430     case 0x7811:
2431         len = (data[0] & 0xff) + 2;
2432         if (len != 7)
2433             fprintf(out, "Bad count in 3DSTATE_GS\n");
2434         if (count < 7)
2435             BUFFER_FAIL(count, len, "3DSTATE_GS");
2436         instr_out(data, hw_offset, 0, "3DSTATE_GS\n");
2437         instr_out(data, hw_offset, 1, "kernel pointer\n");
2438         instr_out(data, hw_offset, 2, "SPF=%d, VME=%d, Sampler Count %d, "
2439                   "Binding table count %d\n",
2440                   (data[2] >> 31) & 1,
2441                   (data[2] >> 30) & 1,
2442                   (data[2] >> 27) & 7,
2443                   (data[2] >> 18) & 0xff);
2444         instr_out(data, hw_offset, 3, "scratch offset\n");
2445         instr_out(data, hw_offset, 4, "Dispatch GRF start %d, VUE read length %d, "
2446                   "VUE read offset %d\n",
2447                   (data[4] & 0xf),
2448                   (data[4] >> 11) & 0x3f,
2449                   (data[4] >> 4) & 0x3f);
2450         instr_out(data, hw_offset, 5, "Max Threads %d, Rendering %sable\n",
2451                   ((data[5] >> 25) & 0x7f) + 1,
2452                   (data[5] & (1 << 8)) != 0 ? "en" : "dis");
2453         instr_out(data, hw_offset, 6, "Reorder %sable, Discard Adjaceny %sable, "
2454                   "GS %sable\n",
2455                   (data[6] & (1 << 30)) != 0 ? "en" : "dis",
2456                   (data[6] & (1 << 29)) != 0 ? "en" : "dis",
2457                   (data[6] & (1 << 15)) != 0 ? "en" : "dis");
2458         return len;
2459
2460     case 0x7812:
2461         len = (data[0] & 0xff) + 2;
2462         if (len != 4)
2463             fprintf(out, "Bad count in 3DSTATE_CLIP\n");
2464         if (count < 4)
2465             BUFFER_FAIL(count, len, "3DSTATE_CLIP");
2466         instr_out(data, hw_offset, 0, "3DSTATE_CLIP\n");
2467         instr_out(data, hw_offset, 1, "UserClip distance cull test mask 0x%x\n",
2468                   data[1] & 0xff);
2469         instr_out(data, hw_offset, 2, "Clip %sable, API mode %s, Viewport XY test %sable, "
2470                   "Viewport Z test %sable, Guardband test %sable, Clip mode %d, "
2471                   "Perspective Divide %sable, Non-Perspective Barycentric %sable, "
2472                   "Tri Provoking %d, Line Provoking %d, Trifan Provoking %d\n",
2473                   (data[2] & (1 << 31)) != 0 ? "en" : "dis",
2474                   (data[2] & (1 << 30)) != 0 ? "D3D" : "OGL",
2475                   (data[2] & (1 << 28)) != 0 ? "en" : "dis",
2476                   (data[2] & (1 << 27)) != 0 ? "en" : "dis",
2477                   (data[2] & (1 << 26)) != 0 ? "en" : "dis",
2478                   (data[2] >> 13) & 7,
2479                   (data[2] & (1 << 9)) != 0 ? "dis" : "en",
2480                   (data[2] & (1 << 8)) != 0 ? "en" : "dis",
2481                   (data[2] >> 4) & 3,
2482                   (data[2] >> 2) & 3,
2483                   (data[2] & 3));
2484         instr_out(data, hw_offset, 3, "Min PointWidth %d, Max PointWidth %d, "
2485                   "Force Zero RTAIndex %sable, Max VPIndex %d\n",
2486                   (data[3] >> 17) & 0x7ff,
2487                   (data[3] >> 6) & 0x7ff,
2488                   (data[3] & (1 << 5)) != 0 ? "en" : "dis",
2489                   (data[3] & 0xf));
2490         return len;
2491
2492     case 0x7813:
2493         len = (data[0] & 0xff) + 2;
2494         if (len != 20)
2495             fprintf(out, "Bad count in 3DSTATE_SF\n");
2496         if (count < 20)
2497             BUFFER_FAIL(count, len, "3DSTATE_SF");
2498         instr_out(data, hw_offset, 0, "3DSTATE_SF\n");
2499         instr_out(data, hw_offset, 1, "Attrib Out %d, Attrib Swizzle %sable, VUE read length %d, "
2500                   "VUE read offset %d\n",
2501                   (data[1] >> 22) & 0x3f,
2502                   (data[1] & (1 << 21)) != 0 ? "en" : "dis",
2503                   (data[1] >> 11) & 0x1f,
2504                   (data[1] >> 4) & 0x3f);
2505         instr_out(data, hw_offset, 2, "Legacy Global DepthBias %sable, FrontFace fill %d, BF fill %d, "
2506                   "VP transform %sable, FrontWinding_%s\n",
2507                   (data[2] & (1 << 11)) != 0 ? "en" : "dis",
2508                   (data[2] >> 5) & 3,
2509                   (data[2] >> 3) & 3,
2510                   (data[2] & (1 << 1)) != 0 ? "en" : "dis",
2511                   (data[2] & 1) != 0 ? "CCW" : "CW");
2512         instr_out(data, hw_offset, 3, "AA %sable, CullMode %d, Scissor %sable, Multisample m ode %d\n",
2513                   (data[3] & (1 << 31)) != 0 ? "en" : "dis",
2514                   (data[3] >> 29) & 3,
2515                   (data[3] & (1 << 11)) != 0 ? "en" : "dis",
2516                   (data[3] >> 8) & 3);
2517         instr_out(data, hw_offset, 4, "Last Pixel %sable, SubPixel Precision %d, Use PixelWidth %d\n",
2518                   (data[4] & (1 << 31)) != 0 ? "en" : "dis",
2519                   (data[4] & (1 << 12)) != 0 ? 4 : 8,
2520                   (data[4] & (1 << 11)) != 0);
2521         instr_out(data, hw_offset, 5, "Global Depth Offset Constant %f\n", data[5]);
2522         instr_out(data, hw_offset, 6, "Global Depth Offset Scale %f\n", data[6]);
2523         instr_out(data, hw_offset, 7, "Global Depth Offset Clamp %f\n", data[7]);
2524         int i, j;
2525         for (i = 0, j = 0; i < 8; i++, j+=2)
2526             instr_out(data, hw_offset, i+8, "Attrib %d (Override %s%s%s%s, Const Source %d, Swizzle Select %d, "
2527                   "Source %d); Attrib %d (Override %s%s%s%s, Const Source %d, Swizzle Select %d, Source %d)\n",
2528                   j+1,
2529                   (data[8+i] & (1 << 31)) != 0 ? "W":"",
2530                   (data[8+i] & (1 << 30)) != 0 ? "Z":"",
2531                   (data[8+i] & (1 << 29)) != 0 ? "Y":"",
2532                   (data[8+i] & (1 << 28)) != 0 ? "X":"",
2533                   (data[8+i] >> 25) & 3, (data[8+i] >> 22) & 3,
2534                   (data[8+i] >> 16) & 0x1f,
2535                   j,
2536                   (data[8+i] & (1 << 15)) != 0 ? "W":"",
2537                   (data[8+i] & (1 << 14)) != 0 ? "Z":"",
2538                   (data[8+i] & (1 << 13)) != 0 ? "Y":"",
2539                   (data[8+i] & (1 << 12)) != 0 ? "X":"",
2540                   (data[8+i] >> 9) & 3, (data[8+i] >> 6) & 3,
2541                   (data[8+i] & 0x1f));
2542         instr_out(data, hw_offset, 16, "Point Sprite TexCoord Enable\n");
2543         instr_out(data, hw_offset, 17, "Const Interp Enable\n");
2544         instr_out(data, hw_offset, 18, "Attrib 7-0 WrapShortest Enable\n");
2545         instr_out(data, hw_offset, 19, "Attrib 15-8 WrapShortest Enable\n");
2546
2547         return len;
2548
2549     case 0x7814:
2550         len = (data[0] & 0xff) + 2;
2551         if (len != 9)
2552             fprintf(out, "Bad count in 3DSTATE_WM\n");
2553         if (count < 9)
2554             BUFFER_FAIL(count, len, "3DSTATE_WM");
2555         instr_out(data, hw_offset, 0, "3DSTATE_WM\n");
2556         instr_out(data, hw_offset, 1, "kernel start pointer 0\n");
2557         instr_out(data, hw_offset, 2, "SPF=%d, VME=%d, Sampler Count %d, "
2558                   "Binding table count %d\n",
2559                   (data[2] >> 31) & 1,
2560                   (data[2] >> 30) & 1,
2561                   (data[2] >> 27) & 7,
2562                   (data[2] >> 18) & 0xff);
2563         instr_out(data, hw_offset, 3, "scratch offset\n");
2564         instr_out(data, hw_offset, 4, "Depth Clear %d, Depth Resolve %d, HiZ Resolve %d, "
2565                   "Dispatch GRF start[0] %d, start[1] %d, start[2] %d\n",
2566                   (data[4] & (1 << 30)) != 0,
2567                   (data[4] & (1 << 28)) != 0,
2568                   (data[4] & (1 << 27)) != 0,
2569                   (data[4] >> 16) & 0x7f,
2570                   (data[4] >> 8) & 0x7f,
2571                   (data[4] & 0x7f));
2572         instr_out(data, hw_offset, 5, "MaxThreads %d, PS KillPixel %d, PS computed Z %d, "
2573                   "PS use sourceZ %d, Thread Dispatch %d, PS use sourceW %d, Dispatch32 %d, "
2574                   "Dispatch16 %d, Dispatch8 %d\n",
2575                   ((data[5] >> 25) & 0x7f) + 1,
2576                   (data[5] & (1 << 22)) != 0,
2577                   (data[5] & (1 << 21)) != 0,
2578                   (data[5] & (1 << 20)) != 0,
2579                   (data[5] & (1 << 19)) != 0,
2580                   (data[5] & (1 << 8)) != 0,
2581                   (data[5] & (1 << 2)) != 0,
2582                   (data[5] & (1 << 1)) != 0,
2583                   (data[5] & (1 << 0)) != 0);
2584         instr_out(data, hw_offset, 6, "Num SF output %d, Pos XY offset %d, ZW interp mode %d , "
2585                   "Barycentric interp mode 0x%x, Point raster rule %d, Multisample mode %d, "
2586                   "Multisample Dispatch mode %d\n",
2587                   (data[6] >> 20) & 0x3f,
2588                   (data[6] >> 18) & 3,
2589                   (data[6] >> 16) & 3,
2590                   (data[6] >> 10) & 0x3f,
2591                   (data[6] & (1 << 9)) != 0,
2592                   (data[6] >> 1) & 3,
2593                   (data[6] & 1));
2594         instr_out(data, hw_offset, 7, "kernel start pointer 1\n");
2595         instr_out(data, hw_offset, 8, "kernel start pointer 2\n");
2596
2597         return len;
2598
2599     case 0x7900:
2600         if (len != 4)
2601             fprintf(out, "Bad count in 3DSTATE_DRAWING_RECTANGLE\n");
2602         if (count < 4)
2603             BUFFER_FAIL(count, len, "3DSTATE_DRAWING_RECTANGLE");
2604
2605         instr_out(data, hw_offset, 0,
2606                   "3DSTATE_DRAWING_RECTANGLE\n");
2607         instr_out(data, hw_offset, 1, "top left: %d,%d\n",
2608                   data[1] & 0xffff,
2609                   (data[1] >> 16) & 0xffff);
2610         instr_out(data, hw_offset, 2, "bottom right: %d,%d\n",
2611                   data[2] & 0xffff,
2612                   (data[2] >> 16) & 0xffff);
2613         instr_out(data, hw_offset, 3, "origin: %d,%d\n",
2614                   (int)data[3] & 0xffff,
2615                   ((int)data[3] >> 16) & 0xffff);
2616
2617         return len;
2618
2619     case 0x7905:
2620         if (len < 5 || len > 7)
2621             fprintf(out, "Bad count in 3DSTATE_DEPTH_BUFFER\n");
2622         if (count < len)
2623             BUFFER_FAIL(count, len, "3DSTATE_DEPTH_BUFFER");
2624
2625         instr_out(data, hw_offset, 0,
2626                   "3DSTATE_DEPTH_BUFFER\n");
2627         if (IS_GEN5(devid) || IS_GEN6(devid))
2628             instr_out(data, hw_offset, 1, "%s, %s, pitch = %d bytes, %stiled, HiZ %d, Seperate Stencil %d\n",
2629                     get_965_surfacetype(data[1] >> 29),
2630                     get_965_depthformat((data[1] >> 18) & 0x7),
2631                     (data[1] & 0x0001ffff) + 1,
2632                     data[1] & (1 << 27) ? "" : "not ",
2633                     (data[1] & (1 << 22)) != 0,
2634                     (data[1] & (1 << 21)) != 0);
2635         else
2636             instr_out(data, hw_offset, 1, "%s, %s, pitch = %d bytes, %stiled\n",
2637                     get_965_surfacetype(data[1] >> 29),
2638                     get_965_depthformat((data[1] >> 18) & 0x7),
2639                     (data[1] & 0x0001ffff) + 1,
2640                     data[1] & (1 << 27) ? "" : "not ");
2641         instr_out(data, hw_offset, 2, "depth offset\n");
2642         instr_out(data, hw_offset, 3, "%dx%d\n",
2643                   ((data[3] & 0x0007ffc0) >> 6) + 1,
2644                   ((data[3] & 0xfff80000) >> 19) + 1);
2645         instr_out(data, hw_offset, 4, "volume depth\n");
2646         if (len >= 6)
2647             instr_out(data, hw_offset, 5, "\n");
2648         if (len >= 7) {
2649             if (IS_GEN6(devid))
2650                 instr_out(data, hw_offset, 6, "\n");
2651             else
2652                 instr_out(data, hw_offset, 6, "render target view extent\n");
2653         }
2654
2655
2656         return len;
2657
2658     case 0x7a00:
2659         if (intel_gen(devid) >= 6) {
2660                 int i;
2661                 len = (data[0] & 0xff) + 2;
2662                 if (len != 4 && len != 5)
2663                         fprintf(out, "Bad count in PIPE_CONTROL\n");
2664                 if (count < len)
2665                         BUFFER_FAIL(count, len, "PIPE_CONTROL");
2666
2667                 switch ((data[1] >> 14) & 0x3) {
2668                 case 0: desc1 = "no write"; break;
2669                 case 1: desc1 = "qword write"; break;
2670                 case 2: desc1 = "PS_DEPTH_COUNT write"; break;
2671                 case 3: desc1 = "TIMESTAMP write"; break;
2672                 }
2673                 instr_out(data, hw_offset, 0, "PIPE_CONTROL\n");
2674                 instr_out(data, hw_offset, 1,
2675                           "%s, %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
2676                           desc1,
2677                           data[1] & (1 << 20) ? "cs stall, " : "",
2678                           data[1] & (1 << 19) ? "global snapshot count reset, " : "",
2679                           data[1] & (1 << 18) ? "tlb invalidate, " : "",
2680                           data[1] & (1 << 17) ? "gfdt flush, " : "",
2681                           data[1] & (1 << 17) ? "media state clear, " : "",
2682                           data[1] & (1 << 13) ? "depth stall, " : "",
2683                           data[1] & (1 << 12) ? "render target cache flush, " : "",
2684                           data[1] & (1 << 11) ? "instruction cache invalidate, " : "",
2685                           data[1] & (1 << 10) ? "texture cache invalidate, " : "",
2686                           data[1] & (1 << 9) ? "indirect state invalidate, " : "",
2687                           data[1] & (1 << 8) ? "notify irq, " : "",
2688                           data[1] & (1 << 7) ? "PIPE_CONTROL flush, " : "",
2689                           data[1] & (1 << 6) ? "protect mem app_id, " : "",
2690                           data[1] & (1 << 5) ? "DC flush, " : "",
2691                           data[1] & (1 << 4) ? "vf fetch invalidate, " : "",
2692                           data[1] & (1 << 3) ? "constant cache invalidate, " : "",
2693                           data[1] & (1 << 2) ? "state cache invalidate, " : "",
2694                           data[1] & (1 << 1) ? "stall at scoreboard, " : "",
2695                           data[1] & (1 << 0) ? "depth cache flush, " : "");
2696                 if (len == 5) {
2697                     instr_out(data, hw_offset, 2, "destination address\n");
2698                     instr_out(data, hw_offset, 3, "immediate dword low\n");
2699                     instr_out(data, hw_offset, 4, "immediate dword high\n");
2700                 } else {
2701                     for (i = 2; i < len; i++) {
2702                         instr_out(data, hw_offset, i, "\n");
2703                     }
2704                 }
2705                 return len;
2706         } else {
2707                 len = (data[0] & 0xff) + 2;
2708                 if (len != 4)
2709                         fprintf(out, "Bad count in PIPE_CONTROL\n");
2710                 if (count < len)
2711                         BUFFER_FAIL(count, len, "PIPE_CONTROL");
2712
2713                 switch ((data[0] >> 14) & 0x3) {
2714                 case 0: desc1 = "no write"; break;
2715                 case 1: desc1 = "qword write"; break;
2716                 case 2: desc1 = "PS_DEPTH_COUNT write"; break;
2717                 case 3: desc1 = "TIMESTAMP write"; break;
2718                 }
2719                 instr_out(data, hw_offset, 0,
2720                           "PIPE_CONTROL: %s, %sdepth stall, %sRC write flush, "
2721                           "%sinst flush\n",
2722                           desc1,
2723                           data[0] & (1 << 13) ? "" : "no ",
2724                           data[0] & (1 << 12) ? "" : "no ",
2725                           data[0] & (1 << 11) ? "" : "no ");
2726                 instr_out(data, hw_offset, 1, "destination address\n");
2727                 instr_out(data, hw_offset, 2, "immediate dword low\n");
2728                 instr_out(data, hw_offset, 3, "immediate dword high\n");
2729                 return len;
2730         }
2731     case 0x7b00:
2732         len = (data[0] & 0xff) + 2;
2733         if (len != 6)
2734             fprintf(out, "Bad count in 3DPRIMITIVE\n");
2735         if (count < len)
2736             BUFFER_FAIL(count, len, "3DPRIMITIVE");
2737
2738         instr_out(data, hw_offset, 0,
2739                   "3DPRIMITIVE: %s %s\n",
2740                   get_965_prim_type(data[0]),
2741                   (data[0] & (1 << 15)) ? "random" : "sequential");
2742         instr_out(data, hw_offset, 1, "vertex count\n");
2743         instr_out(data, hw_offset, 2, "start vertex\n");
2744         instr_out(data, hw_offset, 3, "instance count\n");
2745         instr_out(data, hw_offset, 4, "start instance\n");
2746         instr_out(data, hw_offset, 5, "index bias\n");
2747         return len;
2748     }
2749
2750     for (idx = 0; idx < ARRAY_SIZE(opcodes_3d); idx++) {
2751         opcode_3d = &opcodes_3d[idx];
2752         if ((data[0] & 0xffff0000) >> 16 == opcode_3d->opcode) {
2753             unsigned int i;
2754             len = 1;
2755
2756             instr_out(data, hw_offset, 0, "%s\n", opcode_3d->name);
2757             if (opcode_3d->max_len > 1) {
2758                 len = (data[0] & 0xff) + 2;
2759                 if (len < opcode_3d->min_len ||
2760                     len > opcode_3d->max_len)
2761                 {
2762                     fprintf(out, "Bad count in %s\n", opcode_3d->name);
2763                 }
2764             }
2765
2766             for (i = 1; i < len; i++) {
2767                 if (i >= count)
2768                     BUFFER_FAIL(count, len, opcode_3d->name);
2769                 instr_out(data, hw_offset, i, "dword %d\n", i);
2770             }
2771             return len;
2772         }
2773     }
2774
2775     instr_out(data, hw_offset, 0, "3D UNKNOWN: 3d_965 opcode = 0x%x\n", opcode);
2776     (*failures)++;
2777     return 1;
2778 }
2779
2780 static int
2781 decode_3d_i830(uint32_t *data, int count, uint32_t hw_offset, uint32_t devid, int *failures)
2782 {
2783     unsigned int idx;
2784     uint32_t opcode;
2785
2786     struct {
2787         uint32_t opcode;
2788         int min_len;
2789         int max_len;
2790         char *name;
2791     } opcodes_3d[] = {
2792         { 0x02, 1, 1, "3DSTATE_MODES_3" },
2793         { 0x03, 1, 1, "3DSTATE_ENABLES_1"},
2794         { 0x04, 1, 1, "3DSTATE_ENABLES_2"},
2795         { 0x05, 1, 1, "3DSTATE_VFT0"},
2796         { 0x06, 1, 1, "3DSTATE_AA"},
2797         { 0x07, 1, 1, "3DSTATE_RASTERIZATION_RULES" },
2798         { 0x08, 1, 1, "3DSTATE_MODES_1" },
2799         { 0x09, 1, 1, "3DSTATE_STENCIL_TEST" },
2800         { 0x0a, 1, 1, "3DSTATE_VFT1"},
2801         { 0x0b, 1, 1, "3DSTATE_INDPT_ALPHA_BLEND" },
2802         { 0x0c, 1, 1, "3DSTATE_MODES_5" },
2803         { 0x0d, 1, 1, "3DSTATE_MAP_BLEND_OP" },
2804         { 0x0e, 1, 1, "3DSTATE_MAP_BLEND_ARG" },
2805         { 0x0f, 1, 1, "3DSTATE_MODES_2" },
2806         { 0x15, 1, 1, "3DSTATE_FOG_COLOR" },
2807         { 0x16, 1, 1, "3DSTATE_MODES_4" },
2808     }, *opcode_3d;
2809
2810     opcode = (data[0] & 0x1f000000) >> 24;
2811
2812     switch (opcode) {
2813     case 0x1f:
2814         return decode_3d_primitive(data, count, hw_offset, failures);
2815     case 0x1d:
2816         return decode_3d_1d(data, count, hw_offset, devid, failures);
2817     case 0x1c:
2818         return decode_3d_1c(data, count, hw_offset, failures);
2819     }
2820
2821     for (idx = 0; idx < ARRAY_SIZE(opcodes_3d); idx++) {
2822         opcode_3d = &opcodes_3d[idx];
2823         if ((data[0] & 0x1f000000) >> 24 == opcode_3d->opcode) {
2824             unsigned int len = 1, i;
2825
2826             instr_out(data, hw_offset, 0, "%s\n", opcode_3d->name);
2827             if (opcode_3d->max_len > 1) {
2828                 len = (data[0] & 0xff) + 2;
2829                 if (len < opcode_3d->min_len ||
2830                     len > opcode_3d->max_len)
2831                 {
2832                     fprintf(out, "Bad count in %s\n", opcode_3d->name);
2833                 }
2834             }
2835
2836             for (i = 1; i < len; i++) {
2837                 if (i >= count)
2838                     BUFFER_FAIL(count, len, opcode_3d->name);
2839                 instr_out(data, hw_offset, i, "dword %d\n", i);
2840             }
2841             return len;
2842         }
2843     }
2844
2845     instr_out(data, hw_offset, 0, "3D UNKNOWN: 3d_i830 opcode = 0x%x\n", opcode);
2846     (*failures)++;
2847     return 1;
2848 }
2849
2850 /**
2851  * Decodes an i830-i915 batch buffer, writing the output to stdout.
2852  *
2853  * \param data batch buffer contents
2854  * \param count number of DWORDs to decode in the batch buffer
2855  * \param hw_offset hardware address for the buffer
2856  */
2857 int
2858 intel_decode(uint32_t *data, int count,
2859              uint32_t hw_offset,
2860              uint32_t devid,
2861              uint32_t ignore_end_of_batchbuffer)
2862 {
2863     int ret;
2864     int index = 0;
2865     int failures = 0;
2866
2867     out = stdout;
2868
2869     while (index < count) {
2870         switch ((data[index] & 0xe0000000) >> 29) {
2871         case 0x0:
2872             ret = decode_mi(data + index, count - index,
2873                                hw_offset + index * 4, &failures);
2874
2875             /* If MI_BATCHBUFFER_END happened, then dump the rest of the
2876              * output in case we some day want it in debugging, but don't
2877              * decode it since it'll just confuse in the common case.
2878              */
2879             if (ret == -1) {
2880                 if (ignore_end_of_batchbuffer) {
2881                     index++;
2882                 } else {
2883                     for (index = index + 1; index < count; index++) {
2884                         instr_out(data, hw_offset, index, "\n");
2885                     }
2886                 }
2887             } else
2888                 index += ret;
2889             break;
2890         case 0x2:
2891             index += decode_2d(data + index, count - index,
2892                                hw_offset + index * 4, &failures);
2893             break;
2894         case 0x3:
2895             if (IS_965(devid)) {
2896                 index += decode_3d_965(data + index, count - index,
2897                                        hw_offset + index * 4,
2898                                        devid, &failures);
2899             } else if (IS_GEN3(devid)) {
2900                 index += decode_3d(data + index, count - index,
2901                                    hw_offset + index * 4,
2902                                    devid, &failures);
2903             } else {
2904                 index += decode_3d_i830(data + index, count - index,
2905                                         hw_offset + index * 4,
2906                                         devid, &failures);
2907             }
2908             break;
2909         default:
2910             instr_out(data, hw_offset, index, "UNKNOWN\n");
2911             failures++;
2912             index++;
2913             break;
2914         }
2915         fflush(out);
2916     }
2917
2918     return failures;
2919 }
2920
2921 void intel_decode_context_reset(void)
2922 {
2923     saved_s2_set = 0;
2924     saved_s4_set = 1;
2925 }
2926
2927 void intel_decode_context_set_head_tail(uint32_t head, uint32_t tail)
2928 {
2929         head_offset = head;
2930         tail_offset = tail;
2931 }