?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
comctl32: Use SetRect() instead of open coding it.
[wine.git] / dlls / d3dcompiler_43 / bytecodewriter.c
bloba3e6f63fd929420e03179ca9bf1d4ce4d4b7c535
1 /*
2 * Direct3D bytecode output functions
3 *
4 * Copyright 2008 Stefan Dösinger
5 * Copyright 2009 Matteo Bruni
6 *
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
20 *
21 */
22
23 #include "config.h"
24 #include "wine/port.h"
25 #include "wine/debug.h"
26
27 #include "d3d9types.h"
28 #include "d3dcompiler_private.h"
29
30 WINE_DEFAULT_DEBUG_CHANNEL(bytecodewriter);
31
32 /****************************************************************
33 * General assembler shader construction helper routines follow *
34 ****************************************************************/
35 /* struct instruction *alloc_instr
36 *
37 * Allocates a new instruction structure with srcs registers
38 *
39 * Parameters:
40 * srcs: Number of source registers to allocate
41 *
42 * Returns:
43 * A pointer to the allocated instruction structure
44 * NULL in case of an allocation failure
45 */
46 struct instruction *alloc_instr(unsigned int srcs) {
47 struct instruction *ret = d3dcompiler_alloc(sizeof(*ret));
48 if(!ret) {
49 ERR("Failed to allocate memory for an instruction structure\n");
50 return NULL;
51 }
52
53 if(srcs) {
54 ret->src = d3dcompiler_alloc(srcs * sizeof(*ret->src));
55 if(!ret->src) {
56 ERR("Failed to allocate memory for instruction registers\n");
57 d3dcompiler_free(ret);
58 return NULL;
59 }
60 ret->num_srcs = srcs;
61 }
62 return ret;
63 }
64
65 /* void add_instruction
66 *
67 * Adds a new instruction to the shader's instructions array and grows the instruction array
68 * if needed.
69 *
70 * The function does NOT copy the instruction structure. Make sure not to release the
71 * instruction or any of its substructures like registers.
72 *
73 * Parameters:
74 * shader: Shader to add the instruction to
75 * instr: Instruction to add to the shader
76 */
77 BOOL add_instruction(struct bwriter_shader *shader, struct instruction *instr) {
78 struct instruction **new_instructions;
79
80 if(!shader) return FALSE;
81
82 if(shader->instr_alloc_size == 0) {
83 shader->instr = d3dcompiler_alloc(sizeof(*shader->instr) * INSTRARRAY_INITIAL_SIZE);
84 if(!shader->instr) {
85 ERR("Failed to allocate the shader instruction array\n");
86 return FALSE;
87 }
88 shader->instr_alloc_size = INSTRARRAY_INITIAL_SIZE;
89 } else if(shader->instr_alloc_size == shader->num_instrs) {
90 new_instructions = d3dcompiler_realloc(shader->instr,
91 sizeof(*shader->instr) * (shader->instr_alloc_size) * 2);
92 if(!new_instructions) {
93 ERR("Failed to grow the shader instruction array\n");
94 return FALSE;
95 }
96 shader->instr = new_instructions;
97 shader->instr_alloc_size = shader->instr_alloc_size * 2;
98 } else if(shader->num_instrs > shader->instr_alloc_size) {
99 ERR("More instructions than allocated. This should not happen\n");
100 return FALSE;
101 }
102
103 shader->instr[shader->num_instrs] = instr;
104 shader->num_instrs++;
105 return TRUE;
106 }
107
108 BOOL add_constF(struct bwriter_shader *shader, DWORD reg, float x, float y, float z, float w) {
109 struct constant *newconst;
110
111 if(shader->num_cf) {
112 struct constant **newarray;
113 newarray = d3dcompiler_realloc(shader->constF,
114 sizeof(*shader->constF) * (shader->num_cf + 1));
115 if(!newarray) {
116 ERR("Failed to grow the constants array\n");
117 return FALSE;
118 }
119 shader->constF = newarray;
120 } else {
121 shader->constF = d3dcompiler_alloc(sizeof(*shader->constF));
122 if(!shader->constF) {
123 ERR("Failed to allocate the constants array\n");
124 return FALSE;
125 }
126 }
127
128 newconst = d3dcompiler_alloc(sizeof(*newconst));
129 if(!newconst) {
130 ERR("Failed to allocate a new constant\n");
131 return FALSE;
132 }
133 newconst->regnum = reg;
134 newconst->value[0].f = x;
135 newconst->value[1].f = y;
136 newconst->value[2].f = z;
137 newconst->value[3].f = w;
138 shader->constF[shader->num_cf] = newconst;
139
140 shader->num_cf++;
141 return TRUE;
142 }
143
144 BOOL add_constI(struct bwriter_shader *shader, DWORD reg, INT x, INT y, INT z, INT w) {
145 struct constant *newconst;
146
147 if(shader->num_ci) {
148 struct constant **newarray;
149 newarray = d3dcompiler_realloc(shader->constI,
150 sizeof(*shader->constI) * (shader->num_ci + 1));
151 if(!newarray) {
152 ERR("Failed to grow the constants array\n");
153 return FALSE;
154 }
155 shader->constI = newarray;
156 } else {
157 shader->constI = d3dcompiler_alloc(sizeof(*shader->constI));
158 if(!shader->constI) {
159 ERR("Failed to allocate the constants array\n");
160 return FALSE;
161 }
162 }
163
164 newconst = d3dcompiler_alloc(sizeof(*newconst));
165 if(!newconst) {
166 ERR("Failed to allocate a new constant\n");
167 return FALSE;
168 }
169 newconst->regnum = reg;
170 newconst->value[0].i = x;
171 newconst->value[1].i = y;
172 newconst->value[2].i = z;
173 newconst->value[3].i = w;
174 shader->constI[shader->num_ci] = newconst;
175
176 shader->num_ci++;
177 return TRUE;
178 }
179
180 BOOL add_constB(struct bwriter_shader *shader, DWORD reg, BOOL x) {
181 struct constant *newconst;
182
183 if(shader->num_cb) {
184 struct constant **newarray;
185 newarray = d3dcompiler_realloc(shader->constB,
186 sizeof(*shader->constB) * (shader->num_cb + 1));
187 if(!newarray) {
188 ERR("Failed to grow the constants array\n");
189 return FALSE;
190 }
191 shader->constB = newarray;
192 } else {
193 shader->constB = d3dcompiler_alloc(sizeof(*shader->constB));
194 if(!shader->constB) {
195 ERR("Failed to allocate the constants array\n");
196 return FALSE;
197 }
198 }
199
200 newconst = d3dcompiler_alloc(sizeof(*newconst));
201 if(!newconst) {
202 ERR("Failed to allocate a new constant\n");
203 return FALSE;
204 }
205 newconst->regnum = reg;
206 newconst->value[0].b = x;
207 shader->constB[shader->num_cb] = newconst;
208
209 shader->num_cb++;
210 return TRUE;
211 }
212
213 BOOL record_declaration(struct bwriter_shader *shader, DWORD usage,
214 DWORD usage_idx, DWORD mod, BOOL output,
215 DWORD regnum, DWORD writemask, BOOL builtin) {
216 unsigned int *num;
217 struct declaration **decl;
218 unsigned int i;
219
220 if(!shader) return FALSE;
221
222 if(output) {
223 num = &shader->num_outputs;
224 decl = &shader->outputs;
225 } else {
226 num = &shader->num_inputs;
227 decl = &shader->inputs;
228 }
229
230 if(*num == 0) {
231 *decl = d3dcompiler_alloc(sizeof(**decl));
232 if(!*decl) {
233 ERR("Error allocating declarations array\n");
234 return FALSE;
235 }
236 } else {
237 struct declaration *newdecl;
238 for(i = 0; i < *num; i++) {
239 if((*decl)[i].regnum == regnum && ((*decl)[i].writemask & writemask)) {
240 WARN("Declaration of register %u already exists, writemask match 0x%x\n",
241 regnum, (*decl)[i].writemask & writemask);
242 }
243 }
244
245 newdecl = d3dcompiler_realloc(*decl,
246 sizeof(**decl) * ((*num) + 1));
247 if(!newdecl) {
248 ERR("Error reallocating declarations array\n");
249 return FALSE;
250 }
251 *decl = newdecl;
252 }
253 (*decl)[*num].usage = usage;
254 (*decl)[*num].usage_idx = usage_idx;
255 (*decl)[*num].regnum = regnum;
256 (*decl)[*num].mod = mod;
257 (*decl)[*num].writemask = writemask;
258 (*decl)[*num].builtin = builtin;
259 (*num)++;
260
261 return TRUE;
262 }
263
264 BOOL record_sampler(struct bwriter_shader *shader, DWORD samptype, DWORD mod, DWORD regnum) {
265 unsigned int i;
266
267 if(!shader) return FALSE;
268
269 if(shader->num_samplers == 0) {
270 shader->samplers = d3dcompiler_alloc(sizeof(*shader->samplers));
271 if(!shader->samplers) {
272 ERR("Error allocating samplers array\n");
273 return FALSE;
274 }
275 } else {
276 struct samplerdecl *newarray;
277
278 for(i = 0; i < shader->num_samplers; i++) {
279 if(shader->samplers[i].regnum == regnum) {
280 WARN("Sampler %u already declared\n", regnum);
281 /* This is not an error as far as the assembler is concerned.
282 * Direct3D might refuse to load the compiled shader though
283 */
284 }
285 }
286
287 newarray = d3dcompiler_realloc(shader->samplers,
288 sizeof(*shader->samplers) * (shader->num_samplers + 1));
289 if(!newarray) {
290 ERR("Error reallocating samplers array\n");
291 return FALSE;
292 }
293 shader->samplers = newarray;
294 }
295
296 shader->samplers[shader->num_samplers].type = samptype;
297 shader->samplers[shader->num_samplers].mod = mod;
298 shader->samplers[shader->num_samplers].regnum = regnum;
299 shader->num_samplers++;
300 return TRUE;
301 }
302
303
304 /* shader bytecode buffer manipulation functions.
305 * allocate_buffer creates a new buffer structure, put_dword adds a new
306 * DWORD to the buffer. In the rare case of a memory allocation failure
307 * when trying to grow the buffer a flag is set in the buffer to mark it
308 * invalid. This avoids return value checking and passing in many places
309 */
310 static struct bytecode_buffer *allocate_buffer(void) {
311 struct bytecode_buffer *ret;
312
313 ret = d3dcompiler_alloc(sizeof(*ret));
314 if(!ret) return NULL;
315
316 ret->alloc_size = BYTECODEBUFFER_INITIAL_SIZE;
317 ret->data = d3dcompiler_alloc(sizeof(DWORD) * ret->alloc_size);
318 if(!ret->data) {
319 d3dcompiler_free(ret);
320 return NULL;
321 }
322 ret->state = S_OK;
323 return ret;
324 }
325
326 static void put_dword(struct bytecode_buffer *buffer, DWORD value) {
327 if(FAILED(buffer->state)) return;
328
329 if(buffer->alloc_size == buffer->size) {
330 DWORD *newarray;
331 buffer->alloc_size *= 2;
332 newarray = d3dcompiler_realloc(buffer->data,
333 sizeof(DWORD) * buffer->alloc_size);
334 if(!newarray) {
335 ERR("Failed to grow the buffer data memory\n");
336 buffer->state = E_OUTOFMEMORY;
337 return;
338 }
339 buffer->data = newarray;
340 }
341 buffer->data[buffer->size++] = value;
342 }
343
344 /* bwriter -> d3d9 conversion functions. */
345 static DWORD d3d9_swizzle(DWORD bwriter_swizzle)
346 {
347 /* Currently a NOP, but this allows changing the internal definitions
348 * without side effects. */
349 DWORD ret = 0;
350
351 if ((bwriter_swizzle & BWRITERVS_X_X) == BWRITERVS_X_X) ret |= D3DVS_X_X;
352 if ((bwriter_swizzle & BWRITERVS_X_Y) == BWRITERVS_X_Y) ret |= D3DVS_X_Y;
353 if ((bwriter_swizzle & BWRITERVS_X_Z) == BWRITERVS_X_Z) ret |= D3DVS_X_Z;
354 if ((bwriter_swizzle & BWRITERVS_X_W) == BWRITERVS_X_W) ret |= D3DVS_X_W;
355
356 if ((bwriter_swizzle & BWRITERVS_Y_X) == BWRITERVS_Y_X) ret |= D3DVS_Y_X;
357 if ((bwriter_swizzle & BWRITERVS_Y_Y) == BWRITERVS_Y_Y) ret |= D3DVS_Y_Y;
358 if ((bwriter_swizzle & BWRITERVS_Y_Z) == BWRITERVS_Y_Z) ret |= D3DVS_Y_Z;
359 if ((bwriter_swizzle & BWRITERVS_Y_W) == BWRITERVS_Y_W) ret |= D3DVS_Y_W;
360
361 if ((bwriter_swizzle & BWRITERVS_Z_X) == BWRITERVS_Z_X) ret |= D3DVS_Z_X;
362 if ((bwriter_swizzle & BWRITERVS_Z_Y) == BWRITERVS_Z_Y) ret |= D3DVS_Z_Y;
363 if ((bwriter_swizzle & BWRITERVS_Z_Z) == BWRITERVS_Z_Z) ret |= D3DVS_Z_Z;
364 if ((bwriter_swizzle & BWRITERVS_Z_W) == BWRITERVS_Z_W) ret |= D3DVS_Z_W;
365
366 if ((bwriter_swizzle & BWRITERVS_W_X) == BWRITERVS_W_X) ret |= D3DVS_W_X;
367 if ((bwriter_swizzle & BWRITERVS_W_Y) == BWRITERVS_W_Y) ret |= D3DVS_W_Y;
368 if ((bwriter_swizzle & BWRITERVS_W_Z) == BWRITERVS_W_Z) ret |= D3DVS_W_Z;
369 if ((bwriter_swizzle & BWRITERVS_W_W) == BWRITERVS_W_W) ret |= D3DVS_W_W;
370
371 return ret;
372 }
373
374 static DWORD d3d9_writemask(DWORD bwriter_writemask)
375 {
376 DWORD ret = 0;
377
378 if (bwriter_writemask & BWRITERSP_WRITEMASK_0) ret |= D3DSP_WRITEMASK_0;
379 if (bwriter_writemask & BWRITERSP_WRITEMASK_1) ret |= D3DSP_WRITEMASK_1;
380 if (bwriter_writemask & BWRITERSP_WRITEMASK_2) ret |= D3DSP_WRITEMASK_2;
381 if (bwriter_writemask & BWRITERSP_WRITEMASK_3) ret |= D3DSP_WRITEMASK_3;
382
383 return ret;
384 }
385
386 static DWORD d3d9_srcmod(DWORD bwriter_srcmod)
387 {
388 switch (bwriter_srcmod)
389 {
390 case BWRITERSPSM_NONE: return D3DSPSM_NONE;
391 case BWRITERSPSM_NEG: return D3DSPSM_NEG;
392 case BWRITERSPSM_BIAS: return D3DSPSM_BIAS;
393 case BWRITERSPSM_BIASNEG: return D3DSPSM_BIASNEG;
394 case BWRITERSPSM_SIGN: return D3DSPSM_SIGN;
395 case BWRITERSPSM_SIGNNEG: return D3DSPSM_SIGNNEG;
396 case BWRITERSPSM_COMP: return D3DSPSM_COMP;
397 case BWRITERSPSM_X2: return D3DSPSM_X2;
398 case BWRITERSPSM_X2NEG: return D3DSPSM_X2NEG;
399 case BWRITERSPSM_DZ: return D3DSPSM_DZ;
400 case BWRITERSPSM_DW: return D3DSPSM_DW;
401 case BWRITERSPSM_ABS: return D3DSPSM_ABS;
402 case BWRITERSPSM_ABSNEG: return D3DSPSM_ABSNEG;
403 case BWRITERSPSM_NOT: return D3DSPSM_NOT;
404 default:
405 FIXME("Unhandled BWRITERSPSM token %#x.\n", bwriter_srcmod);
406 return 0;
407 }
408 }
409
410 static DWORD d3d9_dstmod(DWORD bwriter_mod)
411 {
412 DWORD ret = 0;
413
414 if (bwriter_mod & BWRITERSPDM_SATURATE) ret |= D3DSPDM_SATURATE;
415 if (bwriter_mod & BWRITERSPDM_PARTIALPRECISION) ret |= D3DSPDM_PARTIALPRECISION;
416 if (bwriter_mod & BWRITERSPDM_MSAMPCENTROID) ret |= D3DSPDM_MSAMPCENTROID;
417
418 return ret;
419 }
420
421 static DWORD d3d9_comparetype(DWORD asmshader_comparetype)
422 {
423 switch (asmshader_comparetype)
424 {
425 case BWRITER_COMPARISON_GT: return D3DSPC_GT;
426 case BWRITER_COMPARISON_EQ: return D3DSPC_EQ;
427 case BWRITER_COMPARISON_GE: return D3DSPC_GE;
428 case BWRITER_COMPARISON_LT: return D3DSPC_LT;
429 case BWRITER_COMPARISON_NE: return D3DSPC_NE;
430 case BWRITER_COMPARISON_LE: return D3DSPC_LE;
431 default:
432 FIXME("Unexpected BWRITER_COMPARISON type %#x.\n", asmshader_comparetype);
433 return 0;
434 }
435 }
436
437 static DWORD d3d9_sampler(DWORD bwriter_sampler)
438 {
439 if (bwriter_sampler == BWRITERSTT_UNKNOWN) return D3DSTT_UNKNOWN;
440 if (bwriter_sampler == BWRITERSTT_1D) return D3DSTT_1D;
441 if (bwriter_sampler == BWRITERSTT_2D) return D3DSTT_2D;
442 if (bwriter_sampler == BWRITERSTT_CUBE) return D3DSTT_CUBE;
443 if (bwriter_sampler == BWRITERSTT_VOLUME) return D3DSTT_VOLUME;
444 FIXME("Unexpected BWRITERSAMPLER_TEXTURE_TYPE type %#x.\n", bwriter_sampler);
445
446 return 0;
447 }
448
449 static DWORD d3d9_register(DWORD bwriter_register)
450 {
451 if (bwriter_register == BWRITERSPR_TEMP) return D3DSPR_TEMP;
452 if (bwriter_register == BWRITERSPR_INPUT) return D3DSPR_INPUT;
453 if (bwriter_register == BWRITERSPR_CONST) return D3DSPR_CONST;
454 if (bwriter_register == BWRITERSPR_ADDR) return D3DSPR_ADDR;
455 if (bwriter_register == BWRITERSPR_TEXTURE) return D3DSPR_TEXTURE;
456 if (bwriter_register == BWRITERSPR_RASTOUT) return D3DSPR_RASTOUT;
457 if (bwriter_register == BWRITERSPR_ATTROUT) return D3DSPR_ATTROUT;
458 if (bwriter_register == BWRITERSPR_TEXCRDOUT) return D3DSPR_TEXCRDOUT;
459 if (bwriter_register == BWRITERSPR_OUTPUT) return D3DSPR_OUTPUT;
460 if (bwriter_register == BWRITERSPR_CONSTINT) return D3DSPR_CONSTINT;
461 if (bwriter_register == BWRITERSPR_COLOROUT) return D3DSPR_COLOROUT;
462 if (bwriter_register == BWRITERSPR_DEPTHOUT) return D3DSPR_DEPTHOUT;
463 if (bwriter_register == BWRITERSPR_SAMPLER) return D3DSPR_SAMPLER;
464 if (bwriter_register == BWRITERSPR_CONSTBOOL) return D3DSPR_CONSTBOOL;
465 if (bwriter_register == BWRITERSPR_LOOP) return D3DSPR_LOOP;
466 if (bwriter_register == BWRITERSPR_MISCTYPE) return D3DSPR_MISCTYPE;
467 if (bwriter_register == BWRITERSPR_LABEL) return D3DSPR_LABEL;
468 if (bwriter_register == BWRITERSPR_PREDICATE) return D3DSPR_PREDICATE;
469
470 FIXME("Unexpected BWRITERSPR %#x.\n", bwriter_register);
471 return ~0U;
472 }
473
474 static DWORD d3d9_opcode(DWORD bwriter_opcode)
475 {
476 switch (bwriter_opcode)
477 {
478 case BWRITERSIO_NOP: return D3DSIO_NOP;
479 case BWRITERSIO_MOV: return D3DSIO_MOV;
480 case BWRITERSIO_ADD: return D3DSIO_ADD;
481 case BWRITERSIO_SUB: return D3DSIO_SUB;
482 case BWRITERSIO_MAD: return D3DSIO_MAD;
483 case BWRITERSIO_MUL: return D3DSIO_MUL;
484 case BWRITERSIO_RCP: return D3DSIO_RCP;
485 case BWRITERSIO_RSQ: return D3DSIO_RSQ;
486 case BWRITERSIO_DP3: return D3DSIO_DP3;
487 case BWRITERSIO_DP4: return D3DSIO_DP4;
488 case BWRITERSIO_MIN: return D3DSIO_MIN;
489 case BWRITERSIO_MAX: return D3DSIO_MAX;
490 case BWRITERSIO_SLT: return D3DSIO_SLT;
491 case BWRITERSIO_SGE: return D3DSIO_SGE;
492 case BWRITERSIO_EXP: return D3DSIO_EXP;
493 case BWRITERSIO_LOG: return D3DSIO_LOG;
494 case BWRITERSIO_LIT: return D3DSIO_LIT;
495 case BWRITERSIO_DST: return D3DSIO_DST;
496 case BWRITERSIO_LRP: return D3DSIO_LRP;
497 case BWRITERSIO_FRC: return D3DSIO_FRC;
498 case BWRITERSIO_M4x4: return D3DSIO_M4x4;
499 case BWRITERSIO_M4x3: return D3DSIO_M4x3;
500 case BWRITERSIO_M3x4: return D3DSIO_M3x4;
501 case BWRITERSIO_M3x3: return D3DSIO_M3x3;
502 case BWRITERSIO_M3x2: return D3DSIO_M3x2;
503 case BWRITERSIO_CALL: return D3DSIO_CALL;
504 case BWRITERSIO_CALLNZ: return D3DSIO_CALLNZ;
505 case BWRITERSIO_LOOP: return D3DSIO_LOOP;
506 case BWRITERSIO_RET: return D3DSIO_RET;
507 case BWRITERSIO_ENDLOOP: return D3DSIO_ENDLOOP;
508 case BWRITERSIO_LABEL: return D3DSIO_LABEL;
509 case BWRITERSIO_DCL: return D3DSIO_DCL;
510 case BWRITERSIO_POW: return D3DSIO_POW;
511 case BWRITERSIO_CRS: return D3DSIO_CRS;
512 case BWRITERSIO_SGN: return D3DSIO_SGN;
513 case BWRITERSIO_ABS: return D3DSIO_ABS;
514 case BWRITERSIO_NRM: return D3DSIO_NRM;
515 case BWRITERSIO_SINCOS: return D3DSIO_SINCOS;
516 case BWRITERSIO_REP: return D3DSIO_REP;
517 case BWRITERSIO_ENDREP: return D3DSIO_ENDREP;
518 case BWRITERSIO_IF: return D3DSIO_IF;
519 case BWRITERSIO_IFC: return D3DSIO_IFC;
520 case BWRITERSIO_ELSE: return D3DSIO_ELSE;
521 case BWRITERSIO_ENDIF: return D3DSIO_ENDIF;
522 case BWRITERSIO_BREAK: return D3DSIO_BREAK;
523 case BWRITERSIO_BREAKC: return D3DSIO_BREAKC;
524 case BWRITERSIO_MOVA: return D3DSIO_MOVA;
525 case BWRITERSIO_DEFB: return D3DSIO_DEFB;
526 case BWRITERSIO_DEFI: return D3DSIO_DEFI;
527
528 case BWRITERSIO_TEXCOORD: return D3DSIO_TEXCOORD;
529 case BWRITERSIO_TEXKILL: return D3DSIO_TEXKILL;
530 case BWRITERSIO_TEX: return D3DSIO_TEX;
531 case BWRITERSIO_TEXBEM: return D3DSIO_TEXBEM;
532 case BWRITERSIO_TEXBEML: return D3DSIO_TEXBEML;
533 case BWRITERSIO_TEXREG2AR: return D3DSIO_TEXREG2AR;
534 case BWRITERSIO_TEXREG2GB: return D3DSIO_TEXREG2GB;
535 case BWRITERSIO_TEXM3x2PAD: return D3DSIO_TEXM3x2PAD;
536 case BWRITERSIO_TEXM3x2TEX: return D3DSIO_TEXM3x2TEX;
537 case BWRITERSIO_TEXM3x3PAD: return D3DSIO_TEXM3x3PAD;
538 case BWRITERSIO_TEXM3x3TEX: return D3DSIO_TEXM3x3TEX;
539 case BWRITERSIO_TEXM3x3SPEC: return D3DSIO_TEXM3x3SPEC;
540 case BWRITERSIO_TEXM3x3VSPEC:return D3DSIO_TEXM3x3VSPEC;
541 case BWRITERSIO_EXPP: return D3DSIO_EXPP;
542 case BWRITERSIO_LOGP: return D3DSIO_LOGP;
543 case BWRITERSIO_CND: return D3DSIO_CND;
544 case BWRITERSIO_DEF: return D3DSIO_DEF;
545 case BWRITERSIO_TEXREG2RGB: return D3DSIO_TEXREG2RGB;
546 case BWRITERSIO_TEXDP3TEX: return D3DSIO_TEXDP3TEX;
547 case BWRITERSIO_TEXM3x2DEPTH:return D3DSIO_TEXM3x2DEPTH;
548 case BWRITERSIO_TEXDP3: return D3DSIO_TEXDP3;
549 case BWRITERSIO_TEXM3x3: return D3DSIO_TEXM3x3;
550 case BWRITERSIO_TEXDEPTH: return D3DSIO_TEXDEPTH;
551 case BWRITERSIO_CMP: return D3DSIO_CMP;
552 case BWRITERSIO_BEM: return D3DSIO_BEM;
553 case BWRITERSIO_DP2ADD: return D3DSIO_DP2ADD;
554 case BWRITERSIO_DSX: return D3DSIO_DSX;
555 case BWRITERSIO_DSY: return D3DSIO_DSY;
556 case BWRITERSIO_TEXLDD: return D3DSIO_TEXLDD;
557 case BWRITERSIO_SETP: return D3DSIO_SETP;
558 case BWRITERSIO_TEXLDL: return D3DSIO_TEXLDL;
559 case BWRITERSIO_BREAKP: return D3DSIO_BREAKP;
560
561 case BWRITERSIO_PHASE: return D3DSIO_PHASE;
562 case BWRITERSIO_COMMENT: return D3DSIO_COMMENT;
563 case BWRITERSIO_END: return D3DSIO_END;
564
565 case BWRITERSIO_TEXLDP: return D3DSIO_TEX | D3DSI_TEXLD_PROJECT;
566 case BWRITERSIO_TEXLDB: return D3DSIO_TEX | D3DSI_TEXLD_BIAS;
567
568 default:
569 FIXME("Unhandled BWRITERSIO token %#x.\n", bwriter_opcode);
570 return ~0U;
571 }
572 }
573
574 static DWORD d3dsp_register( D3DSHADER_PARAM_REGISTER_TYPE type, DWORD num )
575 {
576 return ((type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK) |
577 ((type << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2) |
578 (num & D3DSP_REGNUM_MASK); /* No shift */
579 }
580
581 /******************************************************
582 * Implementation of the writer functions starts here *
583 ******************************************************/
584 static void write_declarations(struct bc_writer *This,
585 struct bytecode_buffer *buffer, BOOL len,
586 const struct declaration *decls, unsigned int num, DWORD type) {
587 DWORD i;
588 DWORD instr_dcl = D3DSIO_DCL;
589 DWORD token;
590 struct shader_reg reg;
591
592 ZeroMemory(&reg, sizeof(reg));
593
594 if(len) {
595 instr_dcl |= 2 << D3DSI_INSTLENGTH_SHIFT;
596 }
597
598 for(i = 0; i < num; i++) {
599 if(decls[i].builtin) continue;
600
601 /* Write the DCL instruction */
602 put_dword(buffer, instr_dcl);
603
604 /* Write the usage and index */
605 token = (1u << 31); /* Bit 31 of non-instruction opcodes is 1 */
606 token |= (decls[i].usage << D3DSP_DCL_USAGE_SHIFT) & D3DSP_DCL_USAGE_MASK;
607 token |= (decls[i].usage_idx << D3DSP_DCL_USAGEINDEX_SHIFT) & D3DSP_DCL_USAGEINDEX_MASK;
608 put_dword(buffer, token);
609
610 /* Write the dest register */
611 reg.type = type;
612 reg.regnum = decls[i].regnum;
613 reg.u.writemask = decls[i].writemask;
614 This->funcs->dstreg(This, &reg, buffer, 0, decls[i].mod);
615 }
616 }
617
618 static void write_const(struct constant **consts, int num, DWORD opcode, DWORD reg_type, struct bytecode_buffer *buffer, BOOL len) {
619 int i;
620 DWORD instr_def = opcode;
621 const DWORD reg = (1u << 31) | d3dsp_register( reg_type, 0 ) | D3DSP_WRITEMASK_ALL;
622
623 if(len) {
624 if(opcode == D3DSIO_DEFB)
625 instr_def |= 2 << D3DSI_INSTLENGTH_SHIFT;
626 else
627 instr_def |= 5 << D3DSI_INSTLENGTH_SHIFT;
628 }
629
630 for(i = 0; i < num; i++) {
631 /* Write the DEF instruction */
632 put_dword(buffer, instr_def);
633
634 put_dword(buffer, reg | (consts[i]->regnum & D3DSP_REGNUM_MASK));
635 put_dword(buffer, consts[i]->value[0].d);
636 if(opcode != D3DSIO_DEFB) {
637 put_dword(buffer, consts[i]->value[1].d);
638 put_dword(buffer, consts[i]->value[2].d);
639 put_dword(buffer, consts[i]->value[3].d);
640 }
641 }
642 }
643
644 static void write_constF(const struct bwriter_shader *shader, struct bytecode_buffer *buffer, BOOL len) {
645 write_const(shader->constF, shader->num_cf, D3DSIO_DEF, D3DSPR_CONST, buffer, len);
646 }
647
648 /* This function looks for VS 1/2 registers mapping to VS 3 output registers */
649 static HRESULT vs_find_builtin_varyings(struct bc_writer *This, const struct bwriter_shader *shader) {
650 DWORD i;
651 DWORD usage, usage_idx, writemask, regnum;
652
653 for(i = 0; i < shader->num_outputs; i++) {
654 if(!shader->outputs[i].builtin) continue;
655
656 usage = shader->outputs[i].usage;
657 usage_idx = shader->outputs[i].usage_idx;
658 writemask = shader->outputs[i].writemask;
659 regnum = shader->outputs[i].regnum;
660
661 switch(usage) {
662 case BWRITERDECLUSAGE_POSITION:
663 case BWRITERDECLUSAGE_POSITIONT:
664 if(usage_idx > 0) {
665 WARN("dcl_position%u not supported in sm 1/2 shaders\n", usage_idx);
666 return E_INVALIDARG;
667 }
668 TRACE("o%u is oPos\n", regnum);
669 This->oPos_regnum = regnum;
670 break;
671
672 case BWRITERDECLUSAGE_COLOR:
673 if(usage_idx > 1) {
674 WARN("dcl_color%u not supported in sm 1/2 shaders\n", usage_idx);
675 return E_INVALIDARG;
676 }
677 if(writemask != BWRITERSP_WRITEMASK_ALL) {
678 WARN("Only WRITEMASK_ALL is supported on color in sm 1/2\n");
679 return E_INVALIDARG;
680 }
681 TRACE("o%u is oD%u\n", regnum, usage_idx);
682 This->oD_regnum[usage_idx] = regnum;
683 break;
684
685 case BWRITERDECLUSAGE_TEXCOORD:
686 if(usage_idx >= 8) {
687 WARN("dcl_color%u not supported in sm 1/2 shaders\n", usage_idx);
688 return E_INVALIDARG;
689 }
690 if(writemask != (BWRITERSP_WRITEMASK_0) &&
691 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1) &&
692 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1 | BWRITERSP_WRITEMASK_2) &&
693 writemask != (BWRITERSP_WRITEMASK_ALL)) {
694 WARN("Partial writemasks not supported on texture coordinates in sm 1 and 2\n");
695 return E_INVALIDARG;
696 }
697 TRACE("o%u is oT%u\n", regnum, usage_idx);
698 This->oT_regnum[usage_idx] = regnum;
699 break;
700
701 case BWRITERDECLUSAGE_PSIZE:
702 if(usage_idx > 0) {
703 WARN("dcl_psize%u not supported in sm 1/2 shaders\n", usage_idx);
704 return E_INVALIDARG;
705 }
706 TRACE("o%u writemask 0x%08x is oPts\n", regnum, writemask);
707 This->oPts_regnum = regnum;
708 This->oPts_mask = writemask;
709 break;
710
711 case BWRITERDECLUSAGE_FOG:
712 if(usage_idx > 0) {
713 WARN("dcl_fog%u not supported in sm 1 shaders\n", usage_idx);
714 return E_INVALIDARG;
715 }
716 if(writemask != BWRITERSP_WRITEMASK_0 && writemask != BWRITERSP_WRITEMASK_1 &&
717 writemask != BWRITERSP_WRITEMASK_2 && writemask != BWRITERSP_WRITEMASK_3) {
718 WARN("Unsupported fog writemask\n");
719 return E_INVALIDARG;
720 }
721 TRACE("o%u writemask 0x%08x is oFog\n", regnum, writemask);
722 This->oFog_regnum = regnum;
723 This->oFog_mask = writemask;
724 break;
725
726 default:
727 WARN("Varying type %u is not supported in shader model 1.x\n", usage);
728 return E_INVALIDARG;
729 }
730 }
731
732 return S_OK;
733 }
734
735 static void vs_1_x_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
736 HRESULT hr;
737
738 if(shader->num_ci || shader->num_cb) {
739 WARN("Int and bool constants are not supported in shader model 1 shaders\n");
740 WARN("Got %u int and %u boolean constants\n", shader->num_ci, shader->num_cb);
741 This->state = E_INVALIDARG;
742 return;
743 }
744
745 hr = vs_find_builtin_varyings(This, shader);
746 if(FAILED(hr)) {
747 This->state = hr;
748 return;
749 }
750
751 write_declarations(This, buffer, FALSE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
752 write_constF(shader, buffer, FALSE);
753 }
754
755 static HRESULT find_ps_builtin_semantics(struct bc_writer *This,
756 const struct bwriter_shader *shader,
757 DWORD texcoords) {
758 DWORD i;
759 DWORD usage, usage_idx, writemask, regnum;
760
761 This->v_regnum[0] = -1; This->v_regnum[1] = -1;
762 for(i = 0; i < 8; i++) This->t_regnum[i] = -1;
763
764 for(i = 0; i < shader->num_inputs; i++) {
765 if(!shader->inputs[i].builtin) continue;
766
767 usage = shader->inputs[i].usage;
768 usage_idx = shader->inputs[i].usage_idx;
769 writemask = shader->inputs[i].writemask;
770 regnum = shader->inputs[i].regnum;
771
772 switch(usage) {
773 case BWRITERDECLUSAGE_COLOR:
774 if(usage_idx > 1) {
775 WARN("dcl_color%u not supported in sm 1 shaders\n", usage_idx);
776 return E_INVALIDARG;
777 }
778 if(writemask != BWRITERSP_WRITEMASK_ALL) {
779 WARN("Only WRITEMASK_ALL is supported on color in sm 1\n");
780 return E_INVALIDARG;
781 }
782 TRACE("v%u is v%u\n", regnum, usage_idx);
783 This->v_regnum[usage_idx] = regnum;
784 break;
785
786 case BWRITERDECLUSAGE_TEXCOORD:
787 if(usage_idx > texcoords) {
788 WARN("dcl_texcoord%u not supported in this shader version\n", usage_idx);
789 return E_INVALIDARG;
790 }
791 if(writemask != (BWRITERSP_WRITEMASK_0) &&
792 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1) &&
793 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1 | BWRITERSP_WRITEMASK_2) &&
794 writemask != (BWRITERSP_WRITEMASK_ALL)) {
795 WARN("Partial writemasks not supported on texture coordinates in sm 1 and 2\n");
796 } else {
797 writemask = BWRITERSP_WRITEMASK_ALL;
798 }
799 TRACE("v%u is t%u\n", regnum, usage_idx);
800 This->t_regnum[usage_idx] = regnum;
801 break;
802
803 default:
804 WARN("Varying type %u is not supported in shader model 1.x\n", usage);
805 return E_INVALIDARG;
806 }
807 }
808
809 return S_OK;
810 }
811
812 static void ps_1_x_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
813 HRESULT hr;
814
815 /* First check the constants and varyings, and complain if unsupported things are used */
816 if(shader->num_ci || shader->num_cb) {
817 WARN("Int and bool constants are not supported in shader model 1 shaders\n");
818 WARN("Got %u int and %u boolean constants\n", shader->num_ci, shader->num_cb);
819 This->state = E_INVALIDARG;
820 return;
821 }
822
823 hr = find_ps_builtin_semantics(This, shader, 4);
824 if(FAILED(hr)) {
825 This->state = hr;
826 return;
827 }
828
829 write_constF(shader, buffer, FALSE);
830 }
831
832 static void ps_1_4_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
833 HRESULT hr;
834
835 /* First check the constants and varyings, and complain if unsupported things are used */
836 if(shader->num_ci || shader->num_cb) {
837 WARN("Int and bool constants are not supported in shader model 1 shaders\n");
838 WARN("Got %u int and %u boolean constants\n", shader->num_ci, shader->num_cb);
839 This->state = E_INVALIDARG;
840 return;
841 }
842 hr = find_ps_builtin_semantics(This, shader, 6);
843 if(FAILED(hr)) {
844 This->state = hr;
845 return;
846 }
847
848 write_constF(shader, buffer, FALSE);
849 }
850
851 static void end(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
852 put_dword(buffer, D3DSIO_END);
853 }
854
855 static DWORD map_vs_output(struct bc_writer *This, DWORD regnum, DWORD mask, DWORD *has_components) {
856 DWORD i;
857
858 *has_components = TRUE;
859 if(regnum == This->oPos_regnum) {
860 return d3dsp_register( D3DSPR_RASTOUT, D3DSRO_POSITION );
861 }
862 if(regnum == This->oFog_regnum && mask == This->oFog_mask) {
863 *has_components = FALSE;
864 return d3dsp_register( D3DSPR_RASTOUT, D3DSRO_FOG ) | D3DSP_WRITEMASK_ALL;
865 }
866 if(regnum == This->oPts_regnum && mask == This->oPts_mask) {
867 *has_components = FALSE;
868 return d3dsp_register( D3DSPR_RASTOUT, D3DSRO_POINT_SIZE ) | D3DSP_WRITEMASK_ALL;
869 }
870 for(i = 0; i < 2; i++) {
871 if(regnum == This->oD_regnum[i]) {
872 return d3dsp_register( D3DSPR_ATTROUT, i );
873 }
874 }
875 for(i = 0; i < 8; i++) {
876 if(regnum == This->oT_regnum[i]) {
877 return d3dsp_register( D3DSPR_TEXCRDOUT, i );
878 }
879 }
880
881 /* The varying must be undeclared - if an unsupported varying was declared,
882 * the vs_find_builtin_varyings function would have caught it and this code
883 * would not run */
884 WARN("Undeclared varying %u\n", regnum);
885 This->state = E_INVALIDARG;
886 return -1;
887 }
888
889 static void vs_12_dstreg(struct bc_writer *This, const struct shader_reg *reg,
890 struct bytecode_buffer *buffer,
891 DWORD shift, DWORD mod) {
892 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
893 DWORD has_wmask;
894
895 if(reg->rel_reg) {
896 WARN("Relative addressing not supported for destination registers\n");
897 This->state = E_INVALIDARG;
898 return;
899 }
900
901 switch(reg->type) {
902 case BWRITERSPR_OUTPUT:
903 token |= map_vs_output(This, reg->regnum, reg->u.writemask, &has_wmask);
904 break;
905
906 case BWRITERSPR_RASTOUT:
907 case BWRITERSPR_ATTROUT:
908 /* These registers are mapped to input and output regs. They can be encoded in the bytecode,
909 * but are unexpected. If we hit this path it might be due to an error.
910 */
911 FIXME("Unexpected register type %u\n", reg->type);
912 /* drop through */
913 case BWRITERSPR_INPUT:
914 case BWRITERSPR_TEMP:
915 case BWRITERSPR_CONST:
916 token |= d3dsp_register( reg->type, reg->regnum );
917 has_wmask = TRUE;
918 break;
919
920 case BWRITERSPR_ADDR:
921 if(reg->regnum != 0) {
922 WARN("Only a0 exists\n");
923 This->state = E_INVALIDARG;
924 return;
925 }
926 token |= d3dsp_register( D3DSPR_ADDR, 0 );
927 has_wmask = TRUE;
928 break;
929
930 case BWRITERSPR_PREDICATE:
931 if(This->version != BWRITERVS_VERSION(2, 1)){
932 WARN("Predicate register is allowed only in vs_2_x\n");
933 This->state = E_INVALIDARG;
934 return;
935 }
936 if(reg->regnum != 0) {
937 WARN("Only predicate register p0 exists\n");
938 This->state = E_INVALIDARG;
939 return;
940 }
941 token |= d3dsp_register( D3DSPR_PREDICATE, 0 );
942 has_wmask = TRUE;
943 break;
944
945 default:
946 WARN("Invalid register type for 1.x-2.x vertex shader\n");
947 This->state = E_INVALIDARG;
948 return;
949 }
950
951 /* strictly speaking there are no modifiers in vs_2_0 and vs_1_x, but they can be written
952 * into the bytecode and since the compiler doesn't do such checks write them
953 * (the checks are done by the undocumented shader validator)
954 */
955 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
956 token |= d3d9_dstmod(mod);
957
958 if(has_wmask) {
959 token |= d3d9_writemask(reg->u.writemask);
960 }
961 put_dword(buffer, token);
962 }
963
964 static void vs_1_x_srcreg(struct bc_writer *This, const struct shader_reg *reg,
965 struct bytecode_buffer *buffer) {
966 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
967 DWORD has_swizzle;
968 DWORD component;
969
970 switch(reg->type) {
971 case BWRITERSPR_OUTPUT:
972 /* Map the swizzle to a writemask, the format expected
973 by map_vs_output
974 */
975 switch(reg->u.swizzle) {
976 case BWRITERVS_SWIZZLE_X:
977 component = BWRITERSP_WRITEMASK_0;
978 break;
979 case BWRITERVS_SWIZZLE_Y:
980 component = BWRITERSP_WRITEMASK_1;
981 break;
982 case BWRITERVS_SWIZZLE_Z:
983 component = BWRITERSP_WRITEMASK_2;
984 break;
985 case BWRITERVS_SWIZZLE_W:
986 component = BWRITERSP_WRITEMASK_3;
987 break;
988 default:
989 component = 0;
990 }
991 token |= map_vs_output(This, reg->regnum, component, &has_swizzle);
992 break;
993
994 case BWRITERSPR_RASTOUT:
995 case BWRITERSPR_ATTROUT:
996 /* These registers are mapped to input and output regs. They can be encoded in the bytecode,
997 * but are unexpected. If we hit this path it might be due to an error.
998 */
999 FIXME("Unexpected register type %u\n", reg->type);
1000 /* drop through */
1001 case BWRITERSPR_INPUT:
1002 case BWRITERSPR_TEMP:
1003 case BWRITERSPR_CONST:
1004 case BWRITERSPR_ADDR:
1005 token |= d3dsp_register( reg->type, reg->regnum );
1006 if(reg->rel_reg) {
1007 if(reg->rel_reg->type != BWRITERSPR_ADDR ||
1008 reg->rel_reg->regnum != 0 ||
1009 reg->rel_reg->u.swizzle != BWRITERVS_SWIZZLE_X) {
1010 WARN("Relative addressing in vs_1_x is only allowed with a0.x\n");
1011 This->state = E_INVALIDARG;
1012 return;
1013 }
1014 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
1015 }
1016 break;
1017
1018 default:
1019 WARN("Invalid register type for 1.x vshader\n");
1020 This->state = E_INVALIDARG;
1021 return;
1022 }
1023
1024 token |= d3d9_swizzle(reg->u.swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1025
1026 token |= d3d9_srcmod(reg->srcmod);
1027 put_dword(buffer, token);
1028 }
1029
1030 static void write_srcregs(struct bc_writer *This, const struct instruction *instr,
1031 struct bytecode_buffer *buffer){
1032 unsigned int i;
1033 if(instr->has_predicate){
1034 This->funcs->srcreg(This, &instr->predicate, buffer);
1035 }
1036 for(i = 0; i < instr->num_srcs; i++){
1037 This->funcs->srcreg(This, &instr->src[i], buffer);
1038 }
1039 }
1040
1041 static DWORD map_ps13_temp(struct bc_writer *This, const struct shader_reg *reg) {
1042 if(reg->regnum == T0_REG) {
1043 return d3dsp_register( D3DSPR_TEXTURE, 0 );
1044 } else if(reg->regnum == T1_REG) {
1045 return d3dsp_register( D3DSPR_TEXTURE, 1 );
1046 } else if(reg->regnum == T2_REG) {
1047 return d3dsp_register( D3DSPR_TEXTURE, 2 );
1048 } else if(reg->regnum == T3_REG) {
1049 return d3dsp_register( D3DSPR_TEXTURE, 3 );
1050 } else {
1051 return d3dsp_register( D3DSPR_TEMP, reg->regnum );
1052 }
1053 }
1054
1055 static DWORD map_ps_input(struct bc_writer *This,
1056 const struct shader_reg *reg) {
1057 DWORD i;
1058 /* Map color interpolators */
1059 for(i = 0; i < 2; i++) {
1060 if(reg->regnum == This->v_regnum[i]) {
1061 return d3dsp_register( D3DSPR_INPUT, i );
1062 }
1063 }
1064 for(i = 0; i < 8; i++) {
1065 if(reg->regnum == This->t_regnum[i]) {
1066 return d3dsp_register( D3DSPR_TEXTURE, i );
1067 }
1068 }
1069
1070 WARN("Invalid ps 1/2 varying\n");
1071 This->state = E_INVALIDARG;
1072 return 0;
1073 }
1074
1075 static void ps_1_0123_srcreg(struct bc_writer *This, const struct shader_reg *reg,
1076 struct bytecode_buffer *buffer) {
1077 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1078 if(reg->rel_reg) {
1079 WARN("Relative addressing not supported in <= ps_3_0\n");
1080 This->state = E_INVALIDARG;
1081 return;
1082 }
1083
1084 switch(reg->type) {
1085 case BWRITERSPR_INPUT:
1086 token |= map_ps_input(This, reg);
1087 break;
1088
1089 /* Take care about the texture temporaries. There's a problem: They aren't
1090 * declared anywhere, so we can only hardcode the values that are used
1091 * to map ps_1_3 shaders to the common shader structure
1092 */
1093 case BWRITERSPR_TEMP:
1094 token |= map_ps13_temp(This, reg);
1095 break;
1096
1097 case BWRITERSPR_CONST: /* Can be mapped 1:1 */
1098 token |= d3dsp_register( reg->type, reg->regnum );
1099 break;
1100
1101 default:
1102 WARN("Invalid register type for <= ps_1_3 shader\n");
1103 This->state = E_INVALIDARG;
1104 return;
1105 }
1106
1107 token |= d3d9_swizzle(reg->u.swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1108
1109 if(reg->srcmod == BWRITERSPSM_DZ || reg->srcmod == BWRITERSPSM_DW ||
1110 reg->srcmod == BWRITERSPSM_ABS || reg->srcmod == BWRITERSPSM_ABSNEG ||
1111 reg->srcmod == BWRITERSPSM_NOT) {
1112 WARN("Invalid source modifier %u for <= ps_1_3\n", reg->srcmod);
1113 This->state = E_INVALIDARG;
1114 return;
1115 }
1116 token |= d3d9_srcmod(reg->srcmod);
1117 put_dword(buffer, token);
1118 }
1119
1120 static void ps_1_0123_dstreg(struct bc_writer *This, const struct shader_reg *reg,
1121 struct bytecode_buffer *buffer,
1122 DWORD shift, DWORD mod) {
1123 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1124
1125 if(reg->rel_reg) {
1126 WARN("Relative addressing not supported for destination registers\n");
1127 This->state = E_INVALIDARG;
1128 return;
1129 }
1130
1131 switch(reg->type) {
1132 case BWRITERSPR_TEMP:
1133 token |= map_ps13_temp(This, reg);
1134 break;
1135
1136 /* texkill uses the input register as a destination parameter */
1137 case BWRITERSPR_INPUT:
1138 token |= map_ps_input(This, reg);
1139 break;
1140
1141 default:
1142 WARN("Invalid dest register type for 1.x pshader\n");
1143 This->state = E_INVALIDARG;
1144 return;
1145 }
1146
1147 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
1148 token |= d3d9_dstmod(mod);
1149
1150 token |= d3d9_writemask(reg->u.writemask);
1151 put_dword(buffer, token);
1152 }
1153
1154 /* The length of an instruction consists of the destination register (if any),
1155 * the number of source registers, the number of address registers used for
1156 * indirect addressing, and optionally the predicate register
1157 */
1158 static DWORD instrlen(const struct instruction *instr, unsigned int srcs, unsigned int dsts) {
1159 unsigned int i;
1160 DWORD ret = srcs + dsts + (instr->has_predicate ? 1 : 0);
1161
1162 if(dsts){
1163 if(instr->dst.rel_reg) ret++;
1164 }
1165 for(i = 0; i < srcs; i++) {
1166 if(instr->src[i].rel_reg) ret++;
1167 }
1168 return ret;
1169 }
1170
1171 static void sm_1_x_opcode(struct bc_writer *This,
1172 const struct instruction *instr,
1173 DWORD token, struct bytecode_buffer *buffer) {
1174 /* In sm_1_x instruction length isn't encoded */
1175 if(instr->coissue){
1176 token |= D3DSI_COISSUE;
1177 }
1178 put_dword(buffer, token);
1179 }
1180
1181 static void instr_handler(struct bc_writer *This,
1182 const struct instruction *instr,
1183 struct bytecode_buffer *buffer) {
1184 DWORD token = d3d9_opcode(instr->opcode);
1185
1186 This->funcs->opcode(This, instr, token, buffer);
1187 if(instr->has_dst) This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1188 write_srcregs(This, instr, buffer);
1189 }
1190
1191 static const struct instr_handler_table vs_1_x_handlers[] = {
1192 {BWRITERSIO_ADD, instr_handler},
1193 {BWRITERSIO_NOP, instr_handler},
1194 {BWRITERSIO_MOV, instr_handler},
1195 {BWRITERSIO_SUB, instr_handler},
1196 {BWRITERSIO_MAD, instr_handler},
1197 {BWRITERSIO_MUL, instr_handler},
1198 {BWRITERSIO_RCP, instr_handler},
1199 {BWRITERSIO_RSQ, instr_handler},
1200 {BWRITERSIO_DP3, instr_handler},
1201 {BWRITERSIO_DP4, instr_handler},
1202 {BWRITERSIO_MIN, instr_handler},
1203 {BWRITERSIO_MAX, instr_handler},
1204 {BWRITERSIO_SLT, instr_handler},
1205 {BWRITERSIO_SGE, instr_handler},
1206 {BWRITERSIO_EXP, instr_handler},
1207 {BWRITERSIO_LOG, instr_handler},
1208 {BWRITERSIO_EXPP, instr_handler},
1209 {BWRITERSIO_LOGP, instr_handler},
1210 {BWRITERSIO_DST, instr_handler},
1211 {BWRITERSIO_FRC, instr_handler},
1212 {BWRITERSIO_M4x4, instr_handler},
1213 {BWRITERSIO_M4x3, instr_handler},
1214 {BWRITERSIO_M3x4, instr_handler},
1215 {BWRITERSIO_M3x3, instr_handler},
1216 {BWRITERSIO_M3x2, instr_handler},
1217 {BWRITERSIO_LIT, instr_handler},
1218
1219 {BWRITERSIO_END, NULL}, /* Sentinel value, it signals
1220 the end of the list */
1221 };
1222
1223 static const struct bytecode_backend vs_1_x_backend = {
1224 vs_1_x_header,
1225 end,
1226 vs_1_x_srcreg,
1227 vs_12_dstreg,
1228 sm_1_x_opcode,
1229 vs_1_x_handlers
1230 };
1231
1232 static void instr_ps_1_0123_texld(struct bc_writer *This,
1233 const struct instruction *instr,
1234 struct bytecode_buffer *buffer) {
1235 DWORD idx;
1236 struct shader_reg reg;
1237 DWORD swizzlemask;
1238
1239 if(instr->src[1].type != BWRITERSPR_SAMPLER ||
1240 instr->src[1].regnum > 3) {
1241 WARN("Unsupported sampler type %u regnum %u\n",
1242 instr->src[1].type, instr->src[1].regnum);
1243 This->state = E_INVALIDARG;
1244 return;
1245 } else if(instr->dst.type != BWRITERSPR_TEMP) {
1246 WARN("Can only sample into a temp register\n");
1247 This->state = E_INVALIDARG;
1248 return;
1249 }
1250
1251 idx = instr->src[1].regnum;
1252 if((idx == 0 && instr->dst.regnum != T0_REG) ||
1253 (idx == 1 && instr->dst.regnum != T1_REG) ||
1254 (idx == 2 && instr->dst.regnum != T2_REG) ||
1255 (idx == 3 && instr->dst.regnum != T3_REG)) {
1256 WARN("Sampling from sampler s%u to register r%u is not possible in ps_1_x\n",
1257 idx, instr->dst.regnum);
1258 This->state = E_INVALIDARG;
1259 return;
1260 }
1261 if(instr->src[0].type == BWRITERSPR_INPUT) {
1262 /* A simple non-dependent read tex instruction */
1263 if(instr->src[0].regnum != This->t_regnum[idx]) {
1264 WARN("Cannot sample from s%u with texture address data from interpolator %u\n",
1265 idx, instr->src[0].regnum);
1266 This->state = E_INVALIDARG;
1267 return;
1268 }
1269 This->funcs->opcode(This, instr, D3DSIO_TEX & D3DSI_OPCODE_MASK, buffer);
1270
1271 /* map the temp dstreg to the ps_1_3 texture temporary register */
1272 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1273 } else if(instr->src[0].type == BWRITERSPR_TEMP) {
1274
1275 swizzlemask = (3 << BWRITERVS_SWIZZLE_SHIFT) |
1276 (3 << (BWRITERVS_SWIZZLE_SHIFT + 2)) |
1277 (3 << (BWRITERVS_SWIZZLE_SHIFT + 4));
1278 if((instr->src[0].u.swizzle & swizzlemask) == (BWRITERVS_X_X | BWRITERVS_Y_Y | BWRITERVS_Z_Z)) {
1279 TRACE("writing texreg2rgb\n");
1280 This->funcs->opcode(This, instr, D3DSIO_TEXREG2RGB & D3DSI_OPCODE_MASK, buffer);
1281 } else if(instr->src[0].u.swizzle == (BWRITERVS_X_W | BWRITERVS_Y_X | BWRITERVS_Z_X | BWRITERVS_W_X)) {
1282 TRACE("writing texreg2ar\n");
1283 This->funcs->opcode(This, instr, D3DSIO_TEXREG2AR & D3DSI_OPCODE_MASK, buffer);
1284 } else if(instr->src[0].u.swizzle == (BWRITERVS_X_Y | BWRITERVS_Y_Z | BWRITERVS_Z_Z | BWRITERVS_W_Z)) {
1285 TRACE("writing texreg2gb\n");
1286 This->funcs->opcode(This, instr, D3DSIO_TEXREG2GB & D3DSI_OPCODE_MASK, buffer);
1287 } else {
1288 WARN("Unsupported src addr swizzle in dependent texld: 0x%08x\n", instr->src[0].u.swizzle);
1289 This->state = E_INVALIDARG;
1290 return;
1291 }
1292
1293 /* Dst and src reg can be mapped normally. Both registers are temporary registers in the
1294 * source shader and have to be mapped to the temporary form of the texture registers. However,
1295 * the src reg doesn't have a swizzle
1296 */
1297 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1298 reg = instr->src[0];
1299 reg.u.swizzle = BWRITERVS_NOSWIZZLE;
1300 This->funcs->srcreg(This, &reg, buffer);
1301 } else {
1302 WARN("Invalid address data source register\n");
1303 This->state = E_INVALIDARG;
1304 return;
1305 }
1306 }
1307
1308 static void instr_ps_1_0123_mov(struct bc_writer *This,
1309 const struct instruction *instr,
1310 struct bytecode_buffer *buffer) {
1311 DWORD token = D3DSIO_MOV & D3DSI_OPCODE_MASK;
1312
1313 if(instr->dst.type == BWRITERSPR_TEMP && instr->src[0].type == BWRITERSPR_INPUT) {
1314 if((instr->dst.regnum == T0_REG && instr->src[0].regnum == This->t_regnum[0]) ||
1315 (instr->dst.regnum == T1_REG && instr->src[0].regnum == This->t_regnum[1]) ||
1316 (instr->dst.regnum == T2_REG && instr->src[0].regnum == This->t_regnum[2]) ||
1317 (instr->dst.regnum == T3_REG && instr->src[0].regnum == This->t_regnum[3])) {
1318 if(instr->dstmod & BWRITERSPDM_SATURATE) {
1319 This->funcs->opcode(This, instr, D3DSIO_TEXCOORD & D3DSI_OPCODE_MASK, buffer);
1320 /* Remove the SATURATE flag, it's implicit to the instruction */
1321 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod & (~BWRITERSPDM_SATURATE));
1322 return;
1323 } else {
1324 WARN("A varying -> temp copy is only supported with the SATURATE modifier in <=ps_1_3\n");
1325 This->state = E_INVALIDARG;
1326 return;
1327 }
1328 } else if(instr->src[0].regnum == This->v_regnum[0] ||
1329 instr->src[0].regnum == This->v_regnum[1]) {
1330 /* Handled by the normal mov below. Just drop out of the if condition */
1331 } else {
1332 WARN("Unsupported varying -> temp mov in <= ps_1_3\n");
1333 This->state = E_INVALIDARG;
1334 return;
1335 }
1336 }
1337
1338 This->funcs->opcode(This, instr, token, buffer);
1339 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1340 This->funcs->srcreg(This, &instr->src[0], buffer);
1341 }
1342
1343 static const struct instr_handler_table ps_1_0123_handlers[] = {
1344 {BWRITERSIO_ADD, instr_handler},
1345 {BWRITERSIO_NOP, instr_handler},
1346 {BWRITERSIO_MOV, instr_ps_1_0123_mov},
1347 {BWRITERSIO_SUB, instr_handler},
1348 {BWRITERSIO_MAD, instr_handler},
1349 {BWRITERSIO_MUL, instr_handler},
1350 {BWRITERSIO_DP3, instr_handler},
1351 {BWRITERSIO_DP4, instr_handler},
1352 {BWRITERSIO_LRP, instr_handler},
1353
1354 /* pshader instructions */
1355 {BWRITERSIO_CND, instr_handler},
1356 {BWRITERSIO_CMP, instr_handler},
1357 {BWRITERSIO_TEXKILL, instr_handler},
1358 {BWRITERSIO_TEX, instr_ps_1_0123_texld},
1359 {BWRITERSIO_TEXBEM, instr_handler},
1360 {BWRITERSIO_TEXBEML, instr_handler},
1361 {BWRITERSIO_TEXM3x2PAD, instr_handler},
1362 {BWRITERSIO_TEXM3x3PAD, instr_handler},
1363 {BWRITERSIO_TEXM3x3SPEC, instr_handler},
1364 {BWRITERSIO_TEXM3x3VSPEC, instr_handler},
1365 {BWRITERSIO_TEXM3x3TEX, instr_handler},
1366 {BWRITERSIO_TEXM3x3, instr_handler},
1367 {BWRITERSIO_TEXM3x2DEPTH, instr_handler},
1368 {BWRITERSIO_TEXM3x2TEX, instr_handler},
1369 {BWRITERSIO_TEXDP3, instr_handler},
1370 {BWRITERSIO_TEXDP3TEX, instr_handler},
1371 {BWRITERSIO_END, NULL},
1372 };
1373
1374 static const struct bytecode_backend ps_1_0123_backend = {
1375 ps_1_x_header,
1376 end,
1377 ps_1_0123_srcreg,
1378 ps_1_0123_dstreg,
1379 sm_1_x_opcode,
1380 ps_1_0123_handlers
1381 };
1382
1383 static void ps_1_4_srcreg(struct bc_writer *This, const struct shader_reg *reg,
1384 struct bytecode_buffer *buffer) {
1385 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1386 if(reg->rel_reg) {
1387 WARN("Relative addressing not supported in <= ps_3_0\n");
1388 This->state = E_INVALIDARG;
1389 return;
1390 }
1391
1392 switch(reg->type) {
1393 case BWRITERSPR_INPUT:
1394 token |= map_ps_input(This, reg);
1395 break;
1396
1397 /* Can be mapped 1:1 */
1398 case BWRITERSPR_TEMP:
1399 case BWRITERSPR_CONST:
1400 token |= d3dsp_register( reg->type, reg->regnum );
1401 break;
1402
1403 default:
1404 WARN("Invalid register type for ps_1_4 shader\n");
1405 This->state = E_INVALIDARG;
1406 return;
1407 }
1408
1409 token |= d3d9_swizzle(reg->u.swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1410
1411 if(reg->srcmod == BWRITERSPSM_ABS || reg->srcmod == BWRITERSPSM_ABSNEG ||
1412 reg->srcmod == BWRITERSPSM_NOT) {
1413 WARN("Invalid source modifier %u for ps_1_4\n", reg->srcmod);
1414 This->state = E_INVALIDARG;
1415 return;
1416 }
1417 token |= d3d9_srcmod(reg->srcmod);
1418 put_dword(buffer, token);
1419 }
1420
1421 static void ps_1_4_dstreg(struct bc_writer *This, const struct shader_reg *reg,
1422 struct bytecode_buffer *buffer,
1423 DWORD shift, DWORD mod) {
1424 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1425
1426 if(reg->rel_reg) {
1427 WARN("Relative addressing not supported for destination registers\n");
1428 This->state = E_INVALIDARG;
1429 return;
1430 }
1431
1432 switch(reg->type) {
1433 case BWRITERSPR_TEMP: /* 1:1 mapping */
1434 token |= d3dsp_register( reg->type, reg->regnum );
1435 break;
1436
1437 /* For texkill */
1438 case BWRITERSPR_INPUT:
1439 token |= map_ps_input(This, reg);
1440 break;
1441
1442 default:
1443 WARN("Invalid dest register type for 1.x pshader\n");
1444 This->state = E_INVALIDARG;
1445 return;
1446 }
1447
1448 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
1449 token |= d3d9_dstmod(mod);
1450
1451 token |= d3d9_writemask(reg->u.writemask);
1452 put_dword(buffer, token);
1453 }
1454
1455 static void instr_ps_1_4_mov(struct bc_writer *This,
1456 const struct instruction *instr,
1457 struct bytecode_buffer *buffer) {
1458 DWORD token = D3DSIO_MOV & D3DSI_OPCODE_MASK;
1459
1460 if(instr->dst.type == BWRITERSPR_TEMP && instr->src[0].type == BWRITERSPR_INPUT) {
1461 if(instr->src[0].regnum == This->t_regnum[0] ||
1462 instr->src[0].regnum == This->t_regnum[1] ||
1463 instr->src[0].regnum == This->t_regnum[2] ||
1464 instr->src[0].regnum == This->t_regnum[3] ||
1465 instr->src[0].regnum == This->t_regnum[4] ||
1466 instr->src[0].regnum == This->t_regnum[5]) {
1467 /* Similar to a regular mov, but a different opcode */
1468 token = D3DSIO_TEXCOORD & D3DSI_OPCODE_MASK;
1469 } else if(instr->src[0].regnum == This->v_regnum[0] ||
1470 instr->src[0].regnum == This->v_regnum[1]) {
1471 /* Handled by the normal mov below. Just drop out of the if condition */
1472 } else {
1473 WARN("Unsupported varying -> temp mov in ps_1_4\n");
1474 This->state = E_INVALIDARG;
1475 return;
1476 }
1477 }
1478
1479 This->funcs->opcode(This, instr, token, buffer);
1480 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1481 This->funcs->srcreg(This, &instr->src[0], buffer);
1482 }
1483
1484 static void instr_ps_1_4_texld(struct bc_writer *This,
1485 const struct instruction *instr,
1486 struct bytecode_buffer *buffer) {
1487 if(instr->src[1].type != BWRITERSPR_SAMPLER ||
1488 instr->src[1].regnum > 5) {
1489 WARN("Unsupported sampler type %u regnum %u\n",
1490 instr->src[1].type, instr->src[1].regnum);
1491 This->state = E_INVALIDARG;
1492 return;
1493 } else if(instr->dst.type != BWRITERSPR_TEMP) {
1494 WARN("Can only sample into a temp register\n");
1495 This->state = E_INVALIDARG;
1496 return;
1497 }
1498
1499 if(instr->src[1].regnum != instr->dst.regnum) {
1500 WARN("Sampling from sampler s%u to register r%u is not possible in ps_1_4\n",
1501 instr->src[1].regnum, instr->dst.regnum);
1502 This->state = E_INVALIDARG;
1503 return;
1504 }
1505
1506 This->funcs->opcode(This, instr, D3DSIO_TEX & D3DSI_OPCODE_MASK, buffer);
1507 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1508 This->funcs->srcreg(This, &instr->src[0], buffer);
1509 }
1510
1511 static const struct instr_handler_table ps_1_4_handlers[] = {
1512 {BWRITERSIO_ADD, instr_handler},
1513 {BWRITERSIO_NOP, instr_handler},
1514 {BWRITERSIO_MOV, instr_ps_1_4_mov},
1515 {BWRITERSIO_SUB, instr_handler},
1516 {BWRITERSIO_MAD, instr_handler},
1517 {BWRITERSIO_MUL, instr_handler},
1518 {BWRITERSIO_DP3, instr_handler},
1519 {BWRITERSIO_DP4, instr_handler},
1520 {BWRITERSIO_LRP, instr_handler},
1521
1522 /* pshader instructions */
1523 {BWRITERSIO_CND, instr_handler},
1524 {BWRITERSIO_CMP, instr_handler},
1525 {BWRITERSIO_TEXKILL, instr_handler},
1526 {BWRITERSIO_TEX, instr_ps_1_4_texld},
1527 {BWRITERSIO_TEXDEPTH, instr_handler},
1528 {BWRITERSIO_BEM, instr_handler},
1529
1530 {BWRITERSIO_PHASE, instr_handler},
1531 {BWRITERSIO_END, NULL},
1532 };
1533
1534 static const struct bytecode_backend ps_1_4_backend = {
1535 ps_1_4_header,
1536 end,
1537 ps_1_4_srcreg,
1538 ps_1_4_dstreg,
1539 sm_1_x_opcode,
1540 ps_1_4_handlers
1541 };
1542
1543 static void write_constB(const struct bwriter_shader *shader, struct bytecode_buffer *buffer, BOOL len) {
1544 write_const(shader->constB, shader->num_cb, D3DSIO_DEFB, D3DSPR_CONSTBOOL, buffer, len);
1545 }
1546
1547 static void write_constI(const struct bwriter_shader *shader, struct bytecode_buffer *buffer, BOOL len) {
1548 write_const(shader->constI, shader->num_ci, D3DSIO_DEFI, D3DSPR_CONSTINT, buffer, len);
1549 }
1550
1551 static void vs_2_header(struct bc_writer *This,
1552 const struct bwriter_shader *shader,
1553 struct bytecode_buffer *buffer) {
1554 HRESULT hr;
1555
1556 hr = vs_find_builtin_varyings(This, shader);
1557 if(FAILED(hr)) {
1558 This->state = hr;
1559 return;
1560 }
1561
1562 write_declarations(This, buffer, TRUE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
1563 write_constF(shader, buffer, TRUE);
1564 write_constB(shader, buffer, TRUE);
1565 write_constI(shader, buffer, TRUE);
1566 }
1567
1568 static void vs_2_srcreg(struct bc_writer *This,
1569 const struct shader_reg *reg,
1570 struct bytecode_buffer *buffer) {
1571 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1572 DWORD has_swizzle;
1573 DWORD component;
1574 DWORD d3d9reg;
1575
1576 switch(reg->type) {
1577 case BWRITERSPR_OUTPUT:
1578 /* Map the swizzle to a writemask, the format expected
1579 by map_vs_output
1580 */
1581 switch(reg->u.swizzle) {
1582 case BWRITERVS_SWIZZLE_X:
1583 component = BWRITERSP_WRITEMASK_0;
1584 break;
1585 case BWRITERVS_SWIZZLE_Y:
1586 component = BWRITERSP_WRITEMASK_1;
1587 break;
1588 case BWRITERVS_SWIZZLE_Z:
1589 component = BWRITERSP_WRITEMASK_2;
1590 break;
1591 case BWRITERVS_SWIZZLE_W:
1592 component = BWRITERSP_WRITEMASK_3;
1593 break;
1594 default:
1595 component = 0;
1596 }
1597 token |= map_vs_output(This, reg->regnum, component, &has_swizzle);
1598 break;
1599
1600 case BWRITERSPR_RASTOUT:
1601 case BWRITERSPR_ATTROUT:
1602 /* These registers are mapped to input and output regs. They can be encoded in the bytecode,
1603 * but are unexpected. If we hit this path it might be due to an error.
1604 */
1605 FIXME("Unexpected register type %u\n", reg->type);
1606 /* drop through */
1607 case BWRITERSPR_INPUT:
1608 case BWRITERSPR_TEMP:
1609 case BWRITERSPR_CONST:
1610 case BWRITERSPR_ADDR:
1611 case BWRITERSPR_CONSTINT:
1612 case BWRITERSPR_CONSTBOOL:
1613 case BWRITERSPR_LABEL:
1614 d3d9reg = d3d9_register(reg->type);
1615 token |= d3dsp_register( d3d9reg, reg->regnum );
1616 break;
1617
1618 case BWRITERSPR_LOOP:
1619 if(reg->regnum != 0) {
1620 WARN("Only regnum 0 is supported for the loop index register in vs_2_0\n");
1621 This->state = E_INVALIDARG;
1622 return;
1623 }
1624 token |= d3dsp_register( D3DSPR_LOOP, 0 );
1625 break;
1626
1627 case BWRITERSPR_PREDICATE:
1628 if(This->version != BWRITERVS_VERSION(2, 1)){
1629 WARN("Predicate register is allowed only in vs_2_x\n");
1630 This->state = E_INVALIDARG;
1631 return;
1632 }
1633 if(reg->regnum > 0) {
1634 WARN("Only predicate register 0 is supported\n");
1635 This->state = E_INVALIDARG;
1636 return;
1637 }
1638 token |= d3dsp_register( D3DSPR_PREDICATE, 0 );
1639 break;
1640
1641 default:
1642 WARN("Invalid register type for 2.0 vshader\n");
1643 This->state = E_INVALIDARG;
1644 return;
1645 }
1646
1647 token |= d3d9_swizzle(reg->u.swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1648
1649 token |= d3d9_srcmod(reg->srcmod);
1650
1651 if(reg->rel_reg)
1652 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
1653
1654 put_dword(buffer, token);
1655
1656 /* vs_2_0 and newer write the register containing the index explicitly in the
1657 * binary code
1658 */
1659 if(token & D3DVS_ADDRMODE_RELATIVE)
1660 vs_2_srcreg(This, reg->rel_reg, buffer);
1661 }
1662
1663 static void sm_2_opcode(struct bc_writer *This,
1664 const struct instruction *instr,
1665 DWORD token, struct bytecode_buffer *buffer) {
1666 /* From sm 2 onwards instruction length is encoded in the opcode field */
1667 int dsts = instr->has_dst ? 1 : 0;
1668 token |= instrlen(instr, instr->num_srcs, dsts) << D3DSI_INSTLENGTH_SHIFT;
1669 if(instr->comptype)
1670 token |= (d3d9_comparetype(instr->comptype) << 16) & (0xf << 16);
1671 if(instr->has_predicate)
1672 token |= D3DSHADER_INSTRUCTION_PREDICATED;
1673 put_dword(buffer,token);
1674 }
1675
1676 static const struct instr_handler_table vs_2_0_handlers[] = {
1677 {BWRITERSIO_ADD, instr_handler},
1678 {BWRITERSIO_NOP, instr_handler},
1679 {BWRITERSIO_MOV, instr_handler},
1680 {BWRITERSIO_SUB, instr_handler},
1681 {BWRITERSIO_MAD, instr_handler},
1682 {BWRITERSIO_MUL, instr_handler},
1683 {BWRITERSIO_RCP, instr_handler},
1684 {BWRITERSIO_RSQ, instr_handler},
1685 {BWRITERSIO_DP3, instr_handler},
1686 {BWRITERSIO_DP4, instr_handler},
1687 {BWRITERSIO_MIN, instr_handler},
1688 {BWRITERSIO_MAX, instr_handler},
1689 {BWRITERSIO_SLT, instr_handler},
1690 {BWRITERSIO_SGE, instr_handler},
1691 {BWRITERSIO_ABS, instr_handler},
1692 {BWRITERSIO_EXP, instr_handler},
1693 {BWRITERSIO_LOG, instr_handler},
1694 {BWRITERSIO_EXPP, instr_handler},
1695 {BWRITERSIO_LOGP, instr_handler},
1696 {BWRITERSIO_DST, instr_handler},
1697 {BWRITERSIO_LRP, instr_handler},
1698 {BWRITERSIO_FRC, instr_handler},
1699 {BWRITERSIO_CRS, instr_handler},
1700 {BWRITERSIO_SGN, instr_handler},
1701 {BWRITERSIO_NRM, instr_handler},
1702 {BWRITERSIO_SINCOS, instr_handler},
1703 {BWRITERSIO_M4x4, instr_handler},
1704 {BWRITERSIO_M4x3, instr_handler},
1705 {BWRITERSIO_M3x4, instr_handler},
1706 {BWRITERSIO_M3x3, instr_handler},
1707 {BWRITERSIO_M3x2, instr_handler},
1708 {BWRITERSIO_LIT, instr_handler},
1709 {BWRITERSIO_POW, instr_handler},
1710 {BWRITERSIO_MOVA, instr_handler},
1711
1712 {BWRITERSIO_CALL, instr_handler},
1713 {BWRITERSIO_CALLNZ, instr_handler},
1714 {BWRITERSIO_REP, instr_handler},
1715 {BWRITERSIO_ENDREP, instr_handler},
1716 {BWRITERSIO_IF, instr_handler},
1717 {BWRITERSIO_LABEL, instr_handler},
1718 {BWRITERSIO_ELSE, instr_handler},
1719 {BWRITERSIO_ENDIF, instr_handler},
1720 {BWRITERSIO_LOOP, instr_handler},
1721 {BWRITERSIO_RET, instr_handler},
1722 {BWRITERSIO_ENDLOOP, instr_handler},
1723
1724 {BWRITERSIO_END, NULL},
1725 };
1726
1727 static const struct bytecode_backend vs_2_0_backend = {
1728 vs_2_header,
1729 end,
1730 vs_2_srcreg,
1731 vs_12_dstreg,
1732 sm_2_opcode,
1733 vs_2_0_handlers
1734 };
1735
1736 static const struct instr_handler_table vs_2_x_handlers[] = {
1737 {BWRITERSIO_ADD, instr_handler},
1738 {BWRITERSIO_NOP, instr_handler},
1739 {BWRITERSIO_MOV, instr_handler},
1740 {BWRITERSIO_SUB, instr_handler},
1741 {BWRITERSIO_MAD, instr_handler},
1742 {BWRITERSIO_MUL, instr_handler},
1743 {BWRITERSIO_RCP, instr_handler},
1744 {BWRITERSIO_RSQ, instr_handler},
1745 {BWRITERSIO_DP3, instr_handler},
1746 {BWRITERSIO_DP4, instr_handler},
1747 {BWRITERSIO_MIN, instr_handler},
1748 {BWRITERSIO_MAX, instr_handler},
1749 {BWRITERSIO_SLT, instr_handler},
1750 {BWRITERSIO_SGE, instr_handler},
1751 {BWRITERSIO_ABS, instr_handler},
1752 {BWRITERSIO_EXP, instr_handler},
1753 {BWRITERSIO_LOG, instr_handler},
1754 {BWRITERSIO_EXPP, instr_handler},
1755 {BWRITERSIO_LOGP, instr_handler},
1756 {BWRITERSIO_DST, instr_handler},
1757 {BWRITERSIO_LRP, instr_handler},
1758 {BWRITERSIO_FRC, instr_handler},
1759 {BWRITERSIO_CRS, instr_handler},
1760 {BWRITERSIO_SGN, instr_handler},
1761 {BWRITERSIO_NRM, instr_handler},
1762 {BWRITERSIO_SINCOS, instr_handler},
1763 {BWRITERSIO_M4x4, instr_handler},
1764 {BWRITERSIO_M4x3, instr_handler},
1765 {BWRITERSIO_M3x4, instr_handler},
1766 {BWRITERSIO_M3x3, instr_handler},
1767 {BWRITERSIO_M3x2, instr_handler},
1768 {BWRITERSIO_LIT, instr_handler},
1769 {BWRITERSIO_POW, instr_handler},
1770 {BWRITERSIO_MOVA, instr_handler},
1771
1772 {BWRITERSIO_CALL, instr_handler},
1773 {BWRITERSIO_CALLNZ, instr_handler},
1774 {BWRITERSIO_REP, instr_handler},
1775 {BWRITERSIO_ENDREP, instr_handler},
1776 {BWRITERSIO_IF, instr_handler},
1777 {BWRITERSIO_LABEL, instr_handler},
1778 {BWRITERSIO_IFC, instr_handler},
1779 {BWRITERSIO_ELSE, instr_handler},
1780 {BWRITERSIO_ENDIF, instr_handler},
1781 {BWRITERSIO_BREAK, instr_handler},
1782 {BWRITERSIO_BREAKC, instr_handler},
1783 {BWRITERSIO_LOOP, instr_handler},
1784 {BWRITERSIO_RET, instr_handler},
1785 {BWRITERSIO_ENDLOOP, instr_handler},
1786
1787 {BWRITERSIO_SETP, instr_handler},
1788 {BWRITERSIO_BREAKP, instr_handler},
1789
1790 {BWRITERSIO_END, NULL},
1791 };
1792
1793 static const struct bytecode_backend vs_2_x_backend = {
1794 vs_2_header,
1795 end,
1796 vs_2_srcreg,
1797 vs_12_dstreg,
1798 sm_2_opcode,
1799 vs_2_x_handlers
1800 };
1801
1802 static void write_samplers(const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
1803 DWORD i;
1804 DWORD instr_dcl = D3DSIO_DCL | (2 << D3DSI_INSTLENGTH_SHIFT);
1805 DWORD token;
1806 const DWORD reg = (1u << 31) | d3dsp_register( D3DSPR_SAMPLER, 0 ) | D3DSP_WRITEMASK_ALL;
1807
1808 for(i = 0; i < shader->num_samplers; i++) {
1809 /* Write the DCL instruction */
1810 put_dword(buffer, instr_dcl);
1811 token = (1u << 31);
1812 /* Already shifted */
1813 token |= (d3d9_sampler(shader->samplers[i].type)) & D3DSP_TEXTURETYPE_MASK;
1814 put_dword(buffer, token);
1815 token = reg | (shader->samplers[i].regnum & D3DSP_REGNUM_MASK);
1816 token |= d3d9_dstmod(shader->samplers[i].mod);
1817 put_dword(buffer, token);
1818 }
1819 }
1820
1821 static void ps_2_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
1822 HRESULT hr = find_ps_builtin_semantics(This, shader, 8);
1823 if(FAILED(hr)) {
1824 This->state = hr;
1825 return;
1826 }
1827
1828 write_declarations(This, buffer, TRUE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
1829 write_samplers(shader, buffer);
1830 write_constF(shader, buffer, TRUE);
1831 write_constB(shader, buffer, TRUE);
1832 write_constI(shader, buffer, TRUE);
1833 }
1834
1835 static void ps_2_srcreg(struct bc_writer *This,
1836 const struct shader_reg *reg,
1837 struct bytecode_buffer *buffer) {
1838 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1839 DWORD d3d9reg;
1840 if(reg->rel_reg) {
1841 WARN("Relative addressing not supported in <= ps_3_0\n");
1842 This->state = E_INVALIDARG;
1843 return;
1844 }
1845
1846 switch(reg->type) {
1847 case BWRITERSPR_INPUT:
1848 token |= map_ps_input(This, reg);
1849 break;
1850
1851 /* Can be mapped 1:1 */
1852 case BWRITERSPR_TEMP:
1853 case BWRITERSPR_CONST:
1854 case BWRITERSPR_COLOROUT:
1855 case BWRITERSPR_CONSTBOOL:
1856 case BWRITERSPR_CONSTINT:
1857 case BWRITERSPR_SAMPLER:
1858 case BWRITERSPR_LABEL:
1859 case BWRITERSPR_DEPTHOUT:
1860 d3d9reg = d3d9_register(reg->type);
1861 token |= d3dsp_register( d3d9reg, reg->regnum );
1862 break;
1863
1864 case BWRITERSPR_PREDICATE:
1865 if(This->version != BWRITERPS_VERSION(2, 1)){
1866 WARN("Predicate register not supported in ps_2_0\n");
1867 This->state = E_INVALIDARG;
1868 }
1869 if(reg->regnum) {
1870 WARN("Predicate register with regnum %u not supported\n",
1871 reg->regnum);
1872 This->state = E_INVALIDARG;
1873 }
1874 token |= d3dsp_register( D3DSPR_PREDICATE, 0 );
1875 break;
1876
1877 default:
1878 WARN("Invalid register type for ps_2_0 shader\n");
1879 This->state = E_INVALIDARG;
1880 return;
1881 }
1882
1883 token |= d3d9_swizzle(reg->u.swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1884
1885 token |= d3d9_srcmod(reg->srcmod);
1886 put_dword(buffer, token);
1887 }
1888
1889 static void ps_2_0_dstreg(struct bc_writer *This,
1890 const struct shader_reg *reg,
1891 struct bytecode_buffer *buffer,
1892 DWORD shift, DWORD mod) {
1893 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
1894 DWORD d3d9reg;
1895
1896 if(reg->rel_reg) {
1897 WARN("Relative addressing not supported for destination registers\n");
1898 This->state = E_INVALIDARG;
1899 return;
1900 }
1901
1902 switch(reg->type) {
1903 case BWRITERSPR_TEMP: /* 1:1 mapping */
1904 case BWRITERSPR_COLOROUT:
1905 case BWRITERSPR_DEPTHOUT:
1906 d3d9reg = d3d9_register(reg->type);
1907 token |= d3dsp_register( d3d9reg, reg->regnum );
1908 break;
1909
1910 case BWRITERSPR_PREDICATE:
1911 if(This->version != BWRITERPS_VERSION(2, 1)){
1912 WARN("Predicate register not supported in ps_2_0\n");
1913 This->state = E_INVALIDARG;
1914 }
1915 token |= d3dsp_register( D3DSPR_PREDICATE, reg->regnum );
1916 break;
1917
1918 /* texkill uses the input register as a destination parameter */
1919 case BWRITERSPR_INPUT:
1920 token |= map_ps_input(This, reg);
1921 break;
1922
1923 default:
1924 WARN("Invalid dest register type for 2.x pshader\n");
1925 This->state = E_INVALIDARG;
1926 return;
1927 }
1928
1929 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
1930 token |= d3d9_dstmod(mod);
1931
1932 token |= d3d9_writemask(reg->u.writemask);
1933 put_dword(buffer, token);
1934 }
1935
1936 static const struct instr_handler_table ps_2_0_handlers[] = {
1937 {BWRITERSIO_ADD, instr_handler},
1938 {BWRITERSIO_NOP, instr_handler},
1939 {BWRITERSIO_MOV, instr_handler},
1940 {BWRITERSIO_SUB, instr_handler},
1941 {BWRITERSIO_MAD, instr_handler},
1942 {BWRITERSIO_MUL, instr_handler},
1943 {BWRITERSIO_RCP, instr_handler},
1944 {BWRITERSIO_RSQ, instr_handler},
1945 {BWRITERSIO_DP3, instr_handler},
1946 {BWRITERSIO_DP4, instr_handler},
1947 {BWRITERSIO_MIN, instr_handler},
1948 {BWRITERSIO_MAX, instr_handler},
1949 {BWRITERSIO_ABS, instr_handler},
1950 {BWRITERSIO_EXP, instr_handler},
1951 {BWRITERSIO_LOG, instr_handler},
1952 {BWRITERSIO_EXPP, instr_handler},
1953 {BWRITERSIO_LOGP, instr_handler},
1954 {BWRITERSIO_LRP, instr_handler},
1955 {BWRITERSIO_FRC, instr_handler},
1956 {BWRITERSIO_CRS, instr_handler},
1957 {BWRITERSIO_NRM, instr_handler},
1958 {BWRITERSIO_SINCOS, instr_handler},
1959 {BWRITERSIO_M4x4, instr_handler},
1960 {BWRITERSIO_M4x3, instr_handler},
1961 {BWRITERSIO_M3x4, instr_handler},
1962 {BWRITERSIO_M3x3, instr_handler},
1963 {BWRITERSIO_M3x2, instr_handler},
1964 {BWRITERSIO_POW, instr_handler},
1965 {BWRITERSIO_DP2ADD, instr_handler},
1966 {BWRITERSIO_CMP, instr_handler},
1967
1968 {BWRITERSIO_TEX, instr_handler},
1969 {BWRITERSIO_TEXLDP, instr_handler},
1970 {BWRITERSIO_TEXLDB, instr_handler},
1971 {BWRITERSIO_TEXKILL, instr_handler},
1972
1973 {BWRITERSIO_END, NULL},
1974 };
1975
1976 static const struct bytecode_backend ps_2_0_backend = {
1977 ps_2_header,
1978 end,
1979 ps_2_srcreg,
1980 ps_2_0_dstreg,
1981 sm_2_opcode,
1982 ps_2_0_handlers
1983 };
1984
1985 static const struct instr_handler_table ps_2_x_handlers[] = {
1986 {BWRITERSIO_ADD, instr_handler},
1987 {BWRITERSIO_NOP, instr_handler},
1988 {BWRITERSIO_MOV, instr_handler},
1989 {BWRITERSIO_SUB, instr_handler},
1990 {BWRITERSIO_MAD, instr_handler},
1991 {BWRITERSIO_MUL, instr_handler},
1992 {BWRITERSIO_RCP, instr_handler},
1993 {BWRITERSIO_RSQ, instr_handler},
1994 {BWRITERSIO_DP3, instr_handler},
1995 {BWRITERSIO_DP4, instr_handler},
1996 {BWRITERSIO_MIN, instr_handler},
1997 {BWRITERSIO_MAX, instr_handler},
1998 {BWRITERSIO_ABS, instr_handler},
1999 {BWRITERSIO_EXP, instr_handler},
2000 {BWRITERSIO_LOG, instr_handler},
2001 {BWRITERSIO_EXPP, instr_handler},
2002 {BWRITERSIO_LOGP, instr_handler},
2003 {BWRITERSIO_LRP, instr_handler},
2004 {BWRITERSIO_FRC, instr_handler},
2005 {BWRITERSIO_CRS, instr_handler},
2006 {BWRITERSIO_NRM, instr_handler},
2007 {BWRITERSIO_SINCOS, instr_handler},
2008 {BWRITERSIO_M4x4, instr_handler},
2009 {BWRITERSIO_M4x3, instr_handler},
2010 {BWRITERSIO_M3x4, instr_handler},
2011 {BWRITERSIO_M3x3, instr_handler},
2012 {BWRITERSIO_M3x2, instr_handler},
2013 {BWRITERSIO_POW, instr_handler},
2014 {BWRITERSIO_DP2ADD, instr_handler},
2015 {BWRITERSIO_CMP, instr_handler},
2016
2017 {BWRITERSIO_CALL, instr_handler},
2018 {BWRITERSIO_CALLNZ, instr_handler},
2019 {BWRITERSIO_REP, instr_handler},
2020 {BWRITERSIO_ENDREP, instr_handler},
2021 {BWRITERSIO_IF, instr_handler},
2022 {BWRITERSIO_LABEL, instr_handler},
2023 {BWRITERSIO_IFC, instr_handler},
2024 {BWRITERSIO_ELSE, instr_handler},
2025 {BWRITERSIO_ENDIF, instr_handler},
2026 {BWRITERSIO_BREAK, instr_handler},
2027 {BWRITERSIO_BREAKC, instr_handler},
2028 {BWRITERSIO_RET, instr_handler},
2029
2030 {BWRITERSIO_TEX, instr_handler},
2031 {BWRITERSIO_TEXLDP, instr_handler},
2032 {BWRITERSIO_TEXLDB, instr_handler},
2033 {BWRITERSIO_TEXKILL, instr_handler},
2034 {BWRITERSIO_DSX, instr_handler},
2035 {BWRITERSIO_DSY, instr_handler},
2036
2037 {BWRITERSIO_SETP, instr_handler},
2038 {BWRITERSIO_BREAKP, instr_handler},
2039
2040 {BWRITERSIO_TEXLDD, instr_handler},
2041
2042 {BWRITERSIO_END, NULL},
2043 };
2044
2045 static const struct bytecode_backend ps_2_x_backend = {
2046 ps_2_header,
2047 end,
2048 ps_2_srcreg,
2049 ps_2_0_dstreg,
2050 sm_2_opcode,
2051 ps_2_x_handlers
2052 };
2053
2054 static void sm_3_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
2055 write_declarations(This, buffer, TRUE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
2056 write_declarations(This, buffer, TRUE, shader->outputs, shader->num_outputs, BWRITERSPR_OUTPUT);
2057 write_constF(shader, buffer, TRUE);
2058 write_constB(shader, buffer, TRUE);
2059 write_constI(shader, buffer, TRUE);
2060 write_samplers(shader, buffer);
2061 }
2062
2063 static void sm_3_srcreg(struct bc_writer *This,
2064 const struct shader_reg *reg,
2065 struct bytecode_buffer *buffer) {
2066 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
2067 DWORD d3d9reg;
2068
2069 d3d9reg = d3d9_register(reg->type);
2070 token |= d3dsp_register( d3d9reg, reg->regnum );
2071 token |= d3d9_swizzle(reg->u.swizzle) & D3DVS_SWIZZLE_MASK;
2072 token |= d3d9_srcmod(reg->srcmod);
2073
2074 if(reg->rel_reg) {
2075 if(reg->type == BWRITERSPR_CONST && This->version == BWRITERPS_VERSION(3, 0)) {
2076 WARN("c%u[...] is unsupported in ps_3_0\n", reg->regnum);
2077 This->state = E_INVALIDARG;
2078 return;
2079 }
2080 if(((reg->rel_reg->type == BWRITERSPR_ADDR && This->version == BWRITERVS_VERSION(3, 0)) ||
2081 reg->rel_reg->type == BWRITERSPR_LOOP) &&
2082 reg->rel_reg->regnum == 0) {
2083 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
2084 } else {
2085 WARN("Unsupported relative addressing register\n");
2086 This->state = E_INVALIDARG;
2087 return;
2088 }
2089 }
2090
2091 put_dword(buffer, token);
2092
2093 /* vs_2_0 and newer write the register containing the index explicitly in the
2094 * binary code
2095 */
2096 if(token & D3DVS_ADDRMODE_RELATIVE) {
2097 sm_3_srcreg(This, reg->rel_reg, buffer);
2098 }
2099 }
2100
2101 static void sm_3_dstreg(struct bc_writer *This,
2102 const struct shader_reg *reg,
2103 struct bytecode_buffer *buffer,
2104 DWORD shift, DWORD mod) {
2105 DWORD token = (1u << 31); /* Bit 31 of registers is 1 */
2106 DWORD d3d9reg;
2107
2108 if(reg->rel_reg) {
2109 if(This->version == BWRITERVS_VERSION(3, 0) &&
2110 reg->type == BWRITERSPR_OUTPUT) {
2111 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
2112 } else {
2113 WARN("Relative addressing not supported for this shader type or register type\n");
2114 This->state = E_INVALIDARG;
2115 return;
2116 }
2117 }
2118
2119 d3d9reg = d3d9_register(reg->type);
2120 token |= d3dsp_register( d3d9reg, reg->regnum );
2121 token |= d3d9_dstmod(mod);
2122 token |= d3d9_writemask(reg->u.writemask);
2123 put_dword(buffer, token);
2124
2125 /* vs_2_0 and newer write the register containing the index explicitly in the
2126 * binary code
2127 */
2128 if(token & D3DVS_ADDRMODE_RELATIVE) {
2129 sm_3_srcreg(This, reg->rel_reg, buffer);
2130 }
2131 }
2132
2133 static const struct instr_handler_table vs_3_handlers[] = {
2134 {BWRITERSIO_ADD, instr_handler},
2135 {BWRITERSIO_NOP, instr_handler},
2136 {BWRITERSIO_MOV, instr_handler},
2137 {BWRITERSIO_SUB, instr_handler},
2138 {BWRITERSIO_MAD, instr_handler},
2139 {BWRITERSIO_MUL, instr_handler},
2140 {BWRITERSIO_RCP, instr_handler},
2141 {BWRITERSIO_RSQ, instr_handler},
2142 {BWRITERSIO_DP3, instr_handler},
2143 {BWRITERSIO_DP4, instr_handler},
2144 {BWRITERSIO_MIN, instr_handler},
2145 {BWRITERSIO_MAX, instr_handler},
2146 {BWRITERSIO_SLT, instr_handler},
2147 {BWRITERSIO_SGE, instr_handler},
2148 {BWRITERSIO_ABS, instr_handler},
2149 {BWRITERSIO_EXP, instr_handler},
2150 {BWRITERSIO_LOG, instr_handler},
2151 {BWRITERSIO_EXPP, instr_handler},
2152 {BWRITERSIO_LOGP, instr_handler},
2153 {BWRITERSIO_DST, instr_handler},
2154 {BWRITERSIO_LRP, instr_handler},
2155 {BWRITERSIO_FRC, instr_handler},
2156 {BWRITERSIO_CRS, instr_handler},
2157 {BWRITERSIO_SGN, instr_handler},
2158 {BWRITERSIO_NRM, instr_handler},
2159 {BWRITERSIO_SINCOS, instr_handler},
2160 {BWRITERSIO_M4x4, instr_handler},
2161 {BWRITERSIO_M4x3, instr_handler},
2162 {BWRITERSIO_M3x4, instr_handler},
2163 {BWRITERSIO_M3x3, instr_handler},
2164 {BWRITERSIO_M3x2, instr_handler},
2165 {BWRITERSIO_LIT, instr_handler},
2166 {BWRITERSIO_POW, instr_handler},
2167 {BWRITERSIO_MOVA, instr_handler},
2168
2169 {BWRITERSIO_CALL, instr_handler},
2170 {BWRITERSIO_CALLNZ, instr_handler},
2171 {BWRITERSIO_REP, instr_handler},
2172 {BWRITERSIO_ENDREP, instr_handler},
2173 {BWRITERSIO_IF, instr_handler},
2174 {BWRITERSIO_LABEL, instr_handler},
2175 {BWRITERSIO_IFC, instr_handler},
2176 {BWRITERSIO_ELSE, instr_handler},
2177 {BWRITERSIO_ENDIF, instr_handler},
2178 {BWRITERSIO_BREAK, instr_handler},
2179 {BWRITERSIO_BREAKC, instr_handler},
2180 {BWRITERSIO_LOOP, instr_handler},
2181 {BWRITERSIO_RET, instr_handler},
2182 {BWRITERSIO_ENDLOOP, instr_handler},
2183
2184 {BWRITERSIO_SETP, instr_handler},
2185 {BWRITERSIO_BREAKP, instr_handler},
2186 {BWRITERSIO_TEXLDL, instr_handler},
2187
2188 {BWRITERSIO_END, NULL},
2189 };
2190
2191 static const struct bytecode_backend vs_3_backend = {
2192 sm_3_header,
2193 end,
2194 sm_3_srcreg,
2195 sm_3_dstreg,
2196 sm_2_opcode,
2197 vs_3_handlers
2198 };
2199
2200 static const struct instr_handler_table ps_3_handlers[] = {
2201 {BWRITERSIO_ADD, instr_handler},
2202 {BWRITERSIO_NOP, instr_handler},
2203 {BWRITERSIO_MOV, instr_handler},
2204 {BWRITERSIO_SUB, instr_handler},
2205 {BWRITERSIO_MAD, instr_handler},
2206 {BWRITERSIO_MUL, instr_handler},
2207 {BWRITERSIO_RCP, instr_handler},
2208 {BWRITERSIO_RSQ, instr_handler},
2209 {BWRITERSIO_DP3, instr_handler},
2210 {BWRITERSIO_DP4, instr_handler},
2211 {BWRITERSIO_MIN, instr_handler},
2212 {BWRITERSIO_MAX, instr_handler},
2213 {BWRITERSIO_ABS, instr_handler},
2214 {BWRITERSIO_EXP, instr_handler},
2215 {BWRITERSIO_LOG, instr_handler},
2216 {BWRITERSIO_EXPP, instr_handler},
2217 {BWRITERSIO_LOGP, instr_handler},
2218 {BWRITERSIO_LRP, instr_handler},
2219 {BWRITERSIO_FRC, instr_handler},
2220 {BWRITERSIO_CRS, instr_handler},
2221 {BWRITERSIO_NRM, instr_handler},
2222 {BWRITERSIO_SINCOS, instr_handler},
2223 {BWRITERSIO_M4x4, instr_handler},
2224 {BWRITERSIO_M4x3, instr_handler},
2225 {BWRITERSIO_M3x4, instr_handler},
2226 {BWRITERSIO_M3x3, instr_handler},
2227 {BWRITERSIO_M3x2, instr_handler},
2228 {BWRITERSIO_POW, instr_handler},
2229 {BWRITERSIO_DP2ADD, instr_handler},
2230 {BWRITERSIO_CMP, instr_handler},
2231
2232 {BWRITERSIO_CALL, instr_handler},
2233 {BWRITERSIO_CALLNZ, instr_handler},
2234 {BWRITERSIO_REP, instr_handler},
2235 {BWRITERSIO_ENDREP, instr_handler},
2236 {BWRITERSIO_IF, instr_handler},
2237 {BWRITERSIO_LABEL, instr_handler},
2238 {BWRITERSIO_IFC, instr_handler},
2239 {BWRITERSIO_ELSE, instr_handler},
2240 {BWRITERSIO_ENDIF, instr_handler},
2241 {BWRITERSIO_BREAK, instr_handler},
2242 {BWRITERSIO_BREAKC, instr_handler},
2243 {BWRITERSIO_LOOP, instr_handler},
2244 {BWRITERSIO_RET, instr_handler},
2245 {BWRITERSIO_ENDLOOP, instr_handler},
2246
2247 {BWRITERSIO_SETP, instr_handler},
2248 {BWRITERSIO_BREAKP, instr_handler},
2249 {BWRITERSIO_TEXLDL, instr_handler},
2250
2251 {BWRITERSIO_TEX, instr_handler},
2252 {BWRITERSIO_TEXLDP, instr_handler},
2253 {BWRITERSIO_TEXLDB, instr_handler},
2254 {BWRITERSIO_TEXKILL, instr_handler},
2255 {BWRITERSIO_DSX, instr_handler},
2256 {BWRITERSIO_DSY, instr_handler},
2257 {BWRITERSIO_TEXLDD, instr_handler},
2258
2259 {BWRITERSIO_END, NULL},
2260 };
2261
2262 static const struct bytecode_backend ps_3_backend = {
2263 sm_3_header,
2264 end,
2265 sm_3_srcreg,
2266 sm_3_dstreg,
2267 sm_2_opcode,
2268 ps_3_handlers
2269 };
2270
2271 static void init_vs10_dx9_writer(struct bc_writer *writer) {
2272 TRACE("Creating DirectX9 vertex shader 1.0 writer\n");
2273 writer->funcs = &vs_1_x_backend;
2274 }
2275
2276 static void init_vs11_dx9_writer(struct bc_writer *writer) {
2277 TRACE("Creating DirectX9 vertex shader 1.1 writer\n");
2278 writer->funcs = &vs_1_x_backend;
2279 }
2280
2281 static void init_vs20_dx9_writer(struct bc_writer *writer) {
2282 TRACE("Creating DirectX9 vertex shader 2.0 writer\n");
2283 writer->funcs = &vs_2_0_backend;
2284 }
2285
2286 static void init_vs2x_dx9_writer(struct bc_writer *writer) {
2287 TRACE("Creating DirectX9 vertex shader 2.x writer\n");
2288 writer->funcs = &vs_2_x_backend;
2289 }
2290
2291 static void init_vs30_dx9_writer(struct bc_writer *writer) {
2292 TRACE("Creating DirectX9 vertex shader 3.0 writer\n");
2293 writer->funcs = &vs_3_backend;
2294 }
2295
2296 static void init_ps10_dx9_writer(struct bc_writer *writer) {
2297 TRACE("Creating DirectX9 pixel shader 1.0 writer\n");
2298 writer->funcs = &ps_1_0123_backend;
2299 }
2300
2301 static void init_ps11_dx9_writer(struct bc_writer *writer) {
2302 TRACE("Creating DirectX9 pixel shader 1.1 writer\n");
2303 writer->funcs = &ps_1_0123_backend;
2304 }
2305
2306 static void init_ps12_dx9_writer(struct bc_writer *writer) {
2307 TRACE("Creating DirectX9 pixel shader 1.2 writer\n");
2308 writer->funcs = &ps_1_0123_backend;
2309 }
2310
2311 static void init_ps13_dx9_writer(struct bc_writer *writer) {
2312 TRACE("Creating DirectX9 pixel shader 1.3 writer\n");
2313 writer->funcs = &ps_1_0123_backend;
2314 }
2315
2316 static void init_ps14_dx9_writer(struct bc_writer *writer) {
2317 TRACE("Creating DirectX9 pixel shader 1.4 writer\n");
2318 writer->funcs = &ps_1_4_backend;
2319 }
2320
2321 static void init_ps20_dx9_writer(struct bc_writer *writer) {
2322 TRACE("Creating DirectX9 pixel shader 2.0 writer\n");
2323 writer->funcs = &ps_2_0_backend;
2324 }
2325
2326 static void init_ps2x_dx9_writer(struct bc_writer *writer) {
2327 TRACE("Creating DirectX9 pixel shader 2.x writer\n");
2328 writer->funcs = &ps_2_x_backend;
2329 }
2330
2331 static void init_ps30_dx9_writer(struct bc_writer *writer) {
2332 TRACE("Creating DirectX9 pixel shader 3.0 writer\n");
2333 writer->funcs = &ps_3_backend;
2334 }
2335
2336 static struct bc_writer *create_writer(DWORD version, DWORD dxversion) {
2337 struct bc_writer *ret = d3dcompiler_alloc(sizeof(*ret));
2338
2339 if(!ret) {
2340 WARN("Failed to allocate a bytecode writer instance\n");
2341 return NULL;
2342 }
2343
2344 switch(version) {
2345 case BWRITERVS_VERSION(1, 0):
2346 if(dxversion != 9) {
2347 WARN("Unsupported dxversion for vertex shader 1.0 requested: %u\n", dxversion);
2348 goto fail;
2349 }
2350 init_vs10_dx9_writer(ret);
2351 break;
2352 case BWRITERVS_VERSION(1, 1):
2353 if(dxversion != 9) {
2354 WARN("Unsupported dxversion for vertex shader 1.1 requested: %u\n", dxversion);
2355 goto fail;
2356 }
2357 init_vs11_dx9_writer(ret);
2358 break;
2359 case BWRITERVS_VERSION(2, 0):
2360 if(dxversion != 9) {
2361 WARN("Unsupported dxversion for vertex shader 2.0 requested: %u\n", dxversion);
2362 goto fail;
2363 }
2364 init_vs20_dx9_writer(ret);
2365 break;
2366 case BWRITERVS_VERSION(2, 1):
2367 if(dxversion != 9) {
2368 WARN("Unsupported dxversion for vertex shader 2.x requested: %u\n", dxversion);
2369 goto fail;
2370 }
2371 init_vs2x_dx9_writer(ret);
2372 break;
2373 case BWRITERVS_VERSION(3, 0):
2374 if(dxversion != 9) {
2375 WARN("Unsupported dxversion for vertex shader 3.0 requested: %u\n", dxversion);
2376 goto fail;
2377 }
2378 init_vs30_dx9_writer(ret);
2379 break;
2380
2381 case BWRITERPS_VERSION(1, 0):
2382 if(dxversion != 9) {
2383 WARN("Unsupported dxversion for pixel shader 1.0 requested: %u\n", dxversion);
2384 goto fail;
2385 }
2386 init_ps10_dx9_writer(ret);
2387 break;
2388 case BWRITERPS_VERSION(1, 1):
2389 if(dxversion != 9) {
2390 WARN("Unsupported dxversion for pixel shader 1.1 requested: %u\n", dxversion);
2391 goto fail;
2392 }
2393 init_ps11_dx9_writer(ret);
2394 break;
2395 case BWRITERPS_VERSION(1, 2):
2396 if(dxversion != 9) {
2397 WARN("Unsupported dxversion for pixel shader 1.2 requested: %u\n", dxversion);
2398 goto fail;
2399 }
2400 init_ps12_dx9_writer(ret);
2401 break;
2402 case BWRITERPS_VERSION(1, 3):
2403 if(dxversion != 9) {
2404 WARN("Unsupported dxversion for pixel shader 1.3 requested: %u\n", dxversion);
2405 goto fail;
2406 }
2407 init_ps13_dx9_writer(ret);
2408 break;
2409 case BWRITERPS_VERSION(1, 4):
2410 if(dxversion != 9) {
2411 WARN("Unsupported dxversion for pixel shader 1.4 requested: %u\n", dxversion);
2412 goto fail;
2413 }
2414 init_ps14_dx9_writer(ret);
2415 break;
2416
2417 case BWRITERPS_VERSION(2, 0):
2418 if(dxversion != 9) {
2419 WARN("Unsupported dxversion for pixel shader 2.0 requested: %u\n", dxversion);
2420 goto fail;
2421 }
2422 init_ps20_dx9_writer(ret);
2423 break;
2424
2425 case BWRITERPS_VERSION(2, 1):
2426 if(dxversion != 9) {
2427 WARN("Unsupported dxversion for pixel shader 2.x requested: %u\n", dxversion);
2428 goto fail;
2429 }
2430 init_ps2x_dx9_writer(ret);
2431 break;
2432
2433 case BWRITERPS_VERSION(3, 0):
2434 if(dxversion != 9) {
2435 WARN("Unsupported dxversion for pixel shader 3.0 requested: %u\n", dxversion);
2436 goto fail;
2437 }
2438 init_ps30_dx9_writer(ret);
2439 break;
2440
2441 default:
2442 WARN("Unexpected shader version requested: %08x\n", version);
2443 goto fail;
2444 }
2445 ret->version = version;
2446 return ret;
2447
2448 fail:
2449 d3dcompiler_free(ret);
2450 return NULL;
2451 }
2452
2453 static HRESULT call_instr_handler(struct bc_writer *writer,
2454 const struct instruction *instr,
2455 struct bytecode_buffer *buffer) {
2456 DWORD i=0;
2457
2458 while(writer->funcs->instructions[i].opcode != BWRITERSIO_END) {
2459 if(instr->opcode == writer->funcs->instructions[i].opcode) {
2460 if(!writer->funcs->instructions[i].func) {
2461 WARN("Opcode %u not supported by this profile\n", instr->opcode);
2462 return E_INVALIDARG;
2463 }
2464 writer->funcs->instructions[i].func(writer, instr, buffer);
2465 return S_OK;
2466 }
2467 i++;
2468 }
2469
2470 FIXME("Unhandled instruction %u - %s\n", instr->opcode,
2471 debug_print_opcode(instr->opcode));
2472 return E_INVALIDARG;
2473 }
2474
2475 HRESULT SlWriteBytecode(const struct bwriter_shader *shader, int dxversion, DWORD **result, DWORD *size)
2476 {
2477 struct bc_writer *writer;
2478 struct bytecode_buffer *buffer = NULL;
2479 HRESULT hr;
2480 unsigned int i;
2481
2482 if(!shader){
2483 ERR("NULL shader structure, aborting\n");
2484 return E_FAIL;
2485 }
2486 writer = create_writer(shader->version, dxversion);
2487 *result = NULL;
2488
2489 if(!writer) {
2490 WARN("Could not create a bytecode writer instance. Either unsupported version\n");
2491 WARN("or out of memory\n");
2492 hr = E_FAIL;
2493 goto error;
2494 }
2495
2496 buffer = allocate_buffer();
2497 if(!buffer) {
2498 WARN("Failed to allocate a buffer for the shader bytecode\n");
2499 hr = E_FAIL;
2500 goto error;
2501 }
2502
2503 /* Write shader type and version */
2504 put_dword(buffer, shader->version);
2505
2506 writer->funcs->header(writer, shader, buffer);
2507 if(FAILED(writer->state)) {
2508 hr = writer->state;
2509 goto error;
2510 }
2511
2512 for(i = 0; i < shader->num_instrs; i++) {
2513 hr = call_instr_handler(writer, shader->instr[i], buffer);
2514 if(FAILED(hr)) {
2515 goto error;
2516 }
2517 }
2518
2519 if(FAILED(writer->state)) {
2520 hr = writer->state;
2521 goto error;
2522 }
2523
2524 writer->funcs->end(writer, shader, buffer);
2525
2526 if(FAILED(buffer->state)) {
2527 hr = buffer->state;
2528 goto error;
2529 }
2530
2531 *size = buffer->size * sizeof(DWORD);
2532 *result = buffer->data;
2533 buffer->data = NULL;
2534 hr = S_OK;
2535
2536 error:
2537 if(buffer) {
2538 d3dcompiler_free(buffer->data);
2539 d3dcompiler_free(buffer);
2540 }
2541 d3dcompiler_free(writer);
2542 return hr;
2543 }
2544
2545 void SlDeleteShader(struct bwriter_shader *shader) {
2546 unsigned int i, j;
2547
2548 TRACE("Deleting shader %p\n", shader);
2549
2550 for(i = 0; i < shader->num_cf; i++) {
2551 d3dcompiler_free(shader->constF[i]);
2552 }
2553 d3dcompiler_free(shader->constF);
2554 for(i = 0; i < shader->num_ci; i++) {
2555 d3dcompiler_free(shader->constI[i]);
2556 }
2557 d3dcompiler_free(shader->constI);
2558 for(i = 0; i < shader->num_cb; i++) {
2559 d3dcompiler_free(shader->constB[i]);
2560 }
2561 d3dcompiler_free(shader->constB);
2562
2563 d3dcompiler_free(shader->inputs);
2564 d3dcompiler_free(shader->outputs);
2565 d3dcompiler_free(shader->samplers);
2566
2567 for(i = 0; i < shader->num_instrs; i++) {
2568 for(j = 0; j < shader->instr[i]->num_srcs; j++) {
2569 d3dcompiler_free(shader->instr[i]->src[j].rel_reg);
2570 }
2571 d3dcompiler_free(shader->instr[i]->src);
2572 d3dcompiler_free(shader->instr[i]->dst.rel_reg);
2573 d3dcompiler_free(shader->instr[i]);
2574 }
2575 d3dcompiler_free(shader->instr);
2576
2577 d3dcompiler_free(shader);
2578 }
Windows API implementation