?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
configure: Print a warning when prelink isn't found.
[wine/hramrach.git] / dlls / d3dx9_36 / bytecodewriter.c
blobe7bd8a55a0dfe25dad815c92d7293d560588bb9c
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 "d3dx9_36_private.h"
28
29 WINE_DEFAULT_DEBUG_CHANNEL(asmshader);
30
31 /****************************************************************
32 * General assembler shader construction helper routines follow *
33 ****************************************************************/
34 /* struct instruction *alloc_instr
35 *
36 * Allocates a new instruction structure with srcs registers
37 *
38 * Parameters:
39 * srcs: Number of source registers to allocate
40 *
41 * Returns:
42 * A pointer to the allocated instruction structure
43 * NULL in case of an allocation failure
44 */
45 struct instruction *alloc_instr(unsigned int srcs) {
46 struct instruction *ret = asm_alloc(sizeof(*ret));
47 if(!ret) {
48 ERR("Failed to allocate memory for an instruction structure\n");
49 return NULL;
50 }
51
52 if(srcs) {
53 ret->src = asm_alloc(srcs * sizeof(*ret->src));
54 if(!ret->src) {
55 ERR("Failed to allocate memory for instruction registers\n");
56 asm_free(ret);
57 return NULL;
58 }
59 ret->num_srcs = srcs;
60 }
61 return ret;
62 }
63
64 /* void add_instruction
65 *
66 * Adds a new instruction to the shader's instructions array and grows the instruction array
67 * if needed.
68 *
69 * The function does NOT copy the instruction structure. Make sure not to release the
70 * instruction or any of its substructures like registers.
71 *
72 * Parameters:
73 * shader: Shader to add the instruction to
74 * instr: Instruction to add to the shader
75 */
76 BOOL add_instruction(struct bwriter_shader *shader, struct instruction *instr) {
77 struct instruction **new_instructions;
78
79 if(!shader) return FALSE;
80
81 if(shader->instr_alloc_size == 0) {
82 shader->instr = asm_alloc(sizeof(*shader->instr) * INSTRARRAY_INITIAL_SIZE);
83 if(!shader->instr) {
84 ERR("Failed to allocate the shader instruction array\n");
85 return FALSE;
86 }
87 shader->instr_alloc_size = INSTRARRAY_INITIAL_SIZE;
88 } else if(shader->instr_alloc_size == shader->num_instrs) {
89 new_instructions = asm_realloc(shader->instr,
90 sizeof(*shader->instr) * (shader->instr_alloc_size) * 2);
91 if(!new_instructions) {
92 ERR("Failed to grow the shader instruction array\n");
93 return FALSE;
94 }
95 shader->instr = new_instructions;
96 shader->instr_alloc_size = shader->instr_alloc_size * 2;
97 } else if(shader->num_instrs > shader->instr_alloc_size) {
98 ERR("More instructions than allocated. This should not happen\n");
99 return FALSE;
100 }
101
102 shader->instr[shader->num_instrs] = instr;
103 shader->num_instrs++;
104 return TRUE;
105 }
106
107 BOOL add_constF(struct bwriter_shader *shader, DWORD reg, float x, float y, float z, float w) {
108 struct constant *newconst;
109
110 if(shader->num_cf) {
111 struct constant **newarray;
112 newarray = asm_realloc(shader->constF,
113 sizeof(*shader->constF) * (shader->num_cf + 1));
114 if(!newarray) {
115 ERR("Failed to grow the constants array\n");
116 return FALSE;
117 }
118 shader->constF = newarray;
119 } else {
120 shader->constF = asm_alloc(sizeof(*shader->constF));
121 if(!shader->constF) {
122 ERR("Failed to allocate the constants array\n");
123 return FALSE;
124 }
125 }
126
127 newconst = asm_alloc(sizeof(*newconst));
128 if(!newconst) {
129 ERR("Failed to allocate a new constant\n");
130 return FALSE;
131 }
132 newconst->regnum = reg;
133 newconst->value[0].f = x;
134 newconst->value[1].f = y;
135 newconst->value[2].f = z;
136 newconst->value[3].f = w;
137 shader->constF[shader->num_cf] = newconst;
138
139 shader->num_cf++;
140 return TRUE;
141 }
142
143 BOOL add_constI(struct bwriter_shader *shader, DWORD reg, INT x, INT y, INT z, INT w) {
144 struct constant *newconst;
145
146 if(shader->num_ci) {
147 struct constant **newarray;
148 newarray = asm_realloc(shader->constI,
149 sizeof(*shader->constI) * (shader->num_ci + 1));
150 if(!newarray) {
151 ERR("Failed to grow the constants array\n");
152 return FALSE;
153 }
154 shader->constI = newarray;
155 } else {
156 shader->constI = asm_alloc(sizeof(*shader->constI));
157 if(!shader->constI) {
158 ERR("Failed to allocate the constants array\n");
159 return FALSE;
160 }
161 }
162
163 newconst = asm_alloc(sizeof(*newconst));
164 if(!newconst) {
165 ERR("Failed to allocate a new constant\n");
166 return FALSE;
167 }
168 newconst->regnum = reg;
169 newconst->value[0].i = x;
170 newconst->value[1].i = y;
171 newconst->value[2].i = z;
172 newconst->value[3].i = w;
173 shader->constI[shader->num_ci] = newconst;
174
175 shader->num_ci++;
176 return TRUE;
177 }
178
179 BOOL add_constB(struct bwriter_shader *shader, DWORD reg, BOOL x) {
180 struct constant *newconst;
181
182 if(shader->num_cb) {
183 struct constant **newarray;
184 newarray = asm_realloc(shader->constB,
185 sizeof(*shader->constB) * (shader->num_cb + 1));
186 if(!newarray) {
187 ERR("Failed to grow the constants array\n");
188 return FALSE;
189 }
190 shader->constB = newarray;
191 } else {
192 shader->constB = asm_alloc(sizeof(*shader->constB));
193 if(!shader->constB) {
194 ERR("Failed to allocate the constants array\n");
195 return FALSE;
196 }
197 }
198
199 newconst = asm_alloc(sizeof(*newconst));
200 if(!newconst) {
201 ERR("Failed to allocate a new constant\n");
202 return FALSE;
203 }
204 newconst->regnum = reg;
205 newconst->value[0].b = x;
206 shader->constB[shader->num_cb] = newconst;
207
208 shader->num_cb++;
209 return TRUE;
210 }
211
212 BOOL record_declaration(struct bwriter_shader *shader, DWORD usage,
213 DWORD usage_idx, DWORD mod, BOOL output,
214 DWORD regnum, DWORD writemask, BOOL builtin) {
215 unsigned int *num;
216 struct declaration **decl;
217 unsigned int i;
218
219 if(!shader) return FALSE;
220
221 if(output) {
222 num = &shader->num_outputs;
223 decl = &shader->outputs;
224 } else {
225 num = &shader->num_inputs;
226 decl = &shader->inputs;
227 }
228
229 if(*num == 0) {
230 *decl = asm_alloc(sizeof(**decl));
231 if(!*decl) {
232 ERR("Error allocating declarations array\n");
233 return FALSE;
234 }
235 } else {
236 struct declaration *newdecl;
237 for(i = 0; i < *num; i++) {
238 if((*decl)[i].regnum == regnum && ((*decl)[i].writemask & writemask)) {
239 WARN("Declaration of register %u already exists, writemask match 0x%x\n",
240 regnum, (*decl)[i].writemask & writemask);
241 }
242 }
243
244 newdecl = asm_realloc(*decl,
245 sizeof(**decl) * ((*num) + 1));
246 if(!newdecl) {
247 ERR("Error reallocating declarations array\n");
248 return FALSE;
249 }
250 *decl = newdecl;
251 }
252 (*decl)[*num].usage = usage;
253 (*decl)[*num].usage_idx = usage_idx;
254 (*decl)[*num].regnum = regnum;
255 (*decl)[*num].mod = mod;
256 (*decl)[*num].writemask = writemask;
257 (*decl)[*num].builtin = builtin;
258 (*num)++;
259
260 return TRUE;
261 }
262
263 BOOL record_sampler(struct bwriter_shader *shader, DWORD samptype, DWORD mod, DWORD regnum) {
264 unsigned int i;
265
266 if(!shader) return FALSE;
267
268 if(shader->num_samplers == 0) {
269 shader->samplers = asm_alloc(sizeof(*shader->samplers));
270 if(!shader->samplers) {
271 ERR("Error allocating samplers array\n");
272 return FALSE;
273 }
274 } else {
275 struct samplerdecl *newarray;
276
277 for(i = 0; i < shader->num_samplers; i++) {
278 if(shader->samplers[i].regnum == regnum) {
279 WARN("Sampler %u already declared\n", regnum);
280 /* This is not an error as far as the assembler is concerned.
281 * Direct3D might refuse to load the compiled shader though
282 */
283 }
284 }
285
286 newarray = asm_realloc(shader->samplers,
287 sizeof(*shader->samplers) * (shader->num_samplers + 1));
288 if(!newarray) {
289 ERR("Error reallocating samplers array\n");
290 return FALSE;
291 }
292 shader->samplers = newarray;
293 }
294
295 shader->samplers[shader->num_samplers].type = samptype;
296 shader->samplers[shader->num_samplers].mod = mod;
297 shader->samplers[shader->num_samplers].regnum = regnum;
298 shader->num_samplers++;
299 return TRUE;
300 }
301
302
303 /* shader bytecode buffer manipulation functions.
304 * allocate_buffer creates a new buffer structure, put_dword adds a new
305 * DWORD to the buffer. In the rare case of a memory allocation failure
306 * when trying to grow the buffer a flag is set in the buffer to mark it
307 * invalid. This avoids return value checking and passing in many places
308 */
309 static struct bytecode_buffer *allocate_buffer(void) {
310 struct bytecode_buffer *ret;
311
312 ret = asm_alloc(sizeof(*ret));
313 if(!ret) return NULL;
314
315 ret->alloc_size = BYTECODEBUFFER_INITIAL_SIZE;
316 ret->data = asm_alloc(sizeof(DWORD) * ret->alloc_size);
317 if(!ret->data) {
318 asm_free(ret);
319 return NULL;
320 }
321 ret->state = S_OK;
322 return ret;
323 }
324
325 static void put_dword(struct bytecode_buffer *buffer, DWORD value) {
326 if(FAILED(buffer->state)) return;
327
328 if(buffer->alloc_size == buffer->size) {
329 DWORD *newarray;
330 buffer->alloc_size *= 2;
331 newarray = asm_realloc(buffer->data,
332 sizeof(DWORD) * buffer->alloc_size);
333 if(!newarray) {
334 ERR("Failed to grow the buffer data memory\n");
335 buffer->state = E_OUTOFMEMORY;
336 return;
337 }
338 buffer->data = newarray;
339 }
340 buffer->data[buffer->size++] = value;
341 }
342
343 /******************************************************
344 * Implementation of the writer functions starts here *
345 ******************************************************/
346 static void write_declarations(struct bc_writer *This,
347 struct bytecode_buffer *buffer, BOOL len,
348 const struct declaration *decls, unsigned int num, DWORD type) {
349 DWORD i;
350 DWORD instr_dcl = D3DSIO_DCL;
351 DWORD token;
352 struct shader_reg reg;
353
354 ZeroMemory(&reg, sizeof(reg));
355
356 if(len) {
357 instr_dcl |= 2 << D3DSI_INSTLENGTH_SHIFT;
358 }
359
360 for(i = 0; i < num; i++) {
361 if(decls[i].builtin) continue;
362
363 /* Write the DCL instruction */
364 put_dword(buffer, instr_dcl);
365
366 /* Write the usage and index */
367 token = (1 << 31); /* Bit 31 of non-instruction opcodes is 1 */
368 token |= (decls[i].usage << D3DSP_DCL_USAGE_SHIFT) & D3DSP_DCL_USAGE_MASK;
369 token |= (decls[i].usage_idx << D3DSP_DCL_USAGEINDEX_SHIFT) & D3DSP_DCL_USAGEINDEX_MASK;
370 put_dword(buffer, token);
371
372 /* Write the dest register */
373 reg.type = type;
374 reg.regnum = decls[i].regnum;
375 reg.writemask = decls[i].writemask;
376 This->funcs->dstreg(This, &reg, buffer, 0, decls[i].mod);
377 }
378 }
379
380 static void write_const(struct constant **consts, int num, DWORD opcode, DWORD reg_type, struct bytecode_buffer *buffer, BOOL len) {
381 DWORD i;
382 DWORD instr_def = opcode;
383 const DWORD reg = (1<<31) |
384 ((reg_type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK) |
385 ((reg_type << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2) |
386 D3DSP_WRITEMASK_ALL;
387
388 if(len) {
389 if(opcode == D3DSIO_DEFB)
390 instr_def |= 2 << D3DSI_INSTLENGTH_SHIFT;
391 else
392 instr_def |= 5 << D3DSI_INSTLENGTH_SHIFT;
393 }
394
395 for(i = 0; i < num; i++) {
396 /* Write the DEF instruction */
397 put_dword(buffer, instr_def);
398
399 put_dword(buffer, reg | (consts[i]->regnum & D3DSP_REGNUM_MASK));
400 put_dword(buffer, consts[i]->value[0].d);
401 if(opcode != D3DSIO_DEFB) {
402 put_dword(buffer, consts[i]->value[1].d);
403 put_dword(buffer, consts[i]->value[2].d);
404 put_dword(buffer, consts[i]->value[3].d);
405 }
406 }
407 }
408
409 static void write_constF(const struct bwriter_shader *shader, struct bytecode_buffer *buffer, BOOL len) {
410 write_const(shader->constF, shader->num_cf, D3DSIO_DEF, D3DSPR_CONST, buffer, len);
411 }
412
413 /* This function looks for VS 1/2 registers mapping to VS 3 output registers */
414 static HRESULT vs_find_builtin_varyings(struct bc_writer *This, const struct bwriter_shader *shader) {
415 DWORD i;
416 DWORD usage, usage_idx, writemask, regnum;
417
418 for(i = 0; i < shader->num_outputs; i++) {
419 if(!shader->outputs[i].builtin) continue;
420
421 usage = shader->outputs[i].usage;
422 usage_idx = shader->outputs[i].usage_idx;
423 writemask = shader->outputs[i].writemask;
424 regnum = shader->outputs[i].regnum;
425
426 switch(usage) {
427 case BWRITERDECLUSAGE_POSITION:
428 case BWRITERDECLUSAGE_POSITIONT:
429 if(usage_idx > 0) {
430 WARN("dcl_position%u not supported in sm 1/2 shaders\n", usage_idx);
431 return E_INVALIDARG;
432 }
433 TRACE("o%u is oPos\n", regnum);
434 This->oPos_regnum = regnum;
435 break;
436
437 case BWRITERDECLUSAGE_COLOR:
438 if(usage_idx > 1) {
439 WARN("dcl_color%u not supported in sm 1/2 shaders\n", usage_idx);
440 return E_INVALIDARG;
441 }
442 if(writemask != BWRITERSP_WRITEMASK_ALL) {
443 WARN("Only WRITEMASK_ALL is supported on color in sm 1/2\n");
444 return E_INVALIDARG;
445 }
446 TRACE("o%u is oD%u\n", regnum, usage_idx);
447 This->oD_regnum[usage_idx] = regnum;
448 break;
449
450 case BWRITERDECLUSAGE_TEXCOORD:
451 if(usage_idx >= 8) {
452 WARN("dcl_color%u not supported in sm 1/2 shaders\n", usage_idx);
453 return E_INVALIDARG;
454 }
455 if(writemask != (BWRITERSP_WRITEMASK_0) &&
456 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1) &&
457 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1 | BWRITERSP_WRITEMASK_2) &&
458 writemask != (BWRITERSP_WRITEMASK_ALL)) {
459 WARN("Partial writemasks not supported on texture coordinates in sm 1 and 2\n");
460 return E_INVALIDARG;
461 }
462 TRACE("o%u is oT%u\n", regnum, usage_idx);
463 This->oT_regnum[usage_idx] = regnum;
464 break;
465
466 case BWRITERDECLUSAGE_PSIZE:
467 if(usage_idx > 0) {
468 WARN("dcl_psize%u not supported in sm 1/2 shaders\n", usage_idx);
469 return E_INVALIDARG;
470 }
471 TRACE("o%u writemask 0x%08x is oPts\n", regnum, writemask);
472 This->oPts_regnum = regnum;
473 This->oPts_mask = writemask;
474 break;
475
476 case BWRITERDECLUSAGE_FOG:
477 if(usage_idx > 0) {
478 WARN("dcl_fog%u not supported in sm 1 shaders\n", usage_idx);
479 return E_INVALIDARG;
480 }
481 if(writemask != BWRITERSP_WRITEMASK_0 && writemask != BWRITERSP_WRITEMASK_1 &&
482 writemask != BWRITERSP_WRITEMASK_2 && writemask != BWRITERSP_WRITEMASK_3) {
483 WARN("Unsupported fog writemask\n");
484 return E_INVALIDARG;
485 }
486 TRACE("o%u writemask 0x%08x is oFog\n", regnum, writemask);
487 This->oFog_regnum = regnum;
488 This->oFog_mask = writemask;
489 break;
490
491 default:
492 WARN("Varying type %u is not supported in shader model 1.x\n", usage);
493 return E_INVALIDARG;
494 }
495 }
496
497 return S_OK;
498 }
499
500 static void vs_1_x_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
501 HRESULT hr;
502
503 if(shader->num_ci || shader->num_cb) {
504 WARN("Int and bool constants are not supported in shader model 1 shaders\n");
505 WARN("Got %u int and %u boolean constants\n", shader->num_ci, shader->num_cb);
506 This->state = E_INVALIDARG;
507 return;
508 }
509
510 hr = vs_find_builtin_varyings(This, shader);
511 if(FAILED(hr)) {
512 This->state = hr;
513 return;
514 }
515
516 /* Declare the shader type and version */
517 put_dword(buffer, This->version);
518
519 write_declarations(This, buffer, FALSE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
520 write_constF(shader, buffer, FALSE);
521 return;
522 }
523
524 static HRESULT find_ps_builtin_semantics(struct bc_writer *This,
525 const struct bwriter_shader *shader,
526 DWORD texcoords) {
527 DWORD i;
528 DWORD usage, usage_idx, writemask, regnum;
529
530 This->v_regnum[0] = -1; This->v_regnum[1] = -1;
531 for(i = 0; i < 8; i++) This->t_regnum[i] = -1;
532
533 for(i = 0; i < shader->num_inputs; i++) {
534 if(!shader->inputs[i].builtin) continue;
535
536 usage = shader->inputs[i].usage;
537 usage_idx = shader->inputs[i].usage_idx;
538 writemask = shader->inputs[i].writemask;
539 regnum = shader->inputs[i].regnum;
540
541 switch(usage) {
542 case BWRITERDECLUSAGE_COLOR:
543 if(usage_idx > 1) {
544 WARN("dcl_color%u not supported in sm 1 shaders\n", usage_idx);
545 return E_INVALIDARG;
546 }
547 if(writemask != BWRITERSP_WRITEMASK_ALL) {
548 WARN("Only WRITEMASK_ALL is supported on color in sm 1\n");
549 return E_INVALIDARG;
550 }
551 TRACE("v%u is v%u\n", regnum, usage_idx);
552 This->v_regnum[usage_idx] = regnum;
553 break;
554
555 case BWRITERDECLUSAGE_TEXCOORD:
556 if(usage_idx > texcoords) {
557 WARN("dcl_texcoord%u not supported in this shader version\n", usage_idx);
558 return E_INVALIDARG;
559 }
560 if(writemask != (BWRITERSP_WRITEMASK_0) &&
561 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1) &&
562 writemask != (BWRITERSP_WRITEMASK_0 | BWRITERSP_WRITEMASK_1 | BWRITERSP_WRITEMASK_2) &&
563 writemask != (BWRITERSP_WRITEMASK_ALL)) {
564 WARN("Partial writemasks not supported on texture coordinates in sm 1 and 2\n");
565 } else {
566 writemask = BWRITERSP_WRITEMASK_ALL;
567 }
568 TRACE("v%u is t%u\n", regnum, usage_idx);
569 This->t_regnum[usage_idx] = regnum;
570 break;
571
572 default:
573 WARN("Varying type %u is not supported in shader model 1.x\n", usage);
574 return E_INVALIDARG;
575 }
576 }
577
578 return S_OK;
579 }
580
581 static void ps_1_x_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
582 HRESULT hr;
583
584 /* First check the constants and varyings, and complain if unsupported things are used */
585 if(shader->num_ci || shader->num_cb) {
586 WARN("Int and bool constants are not supported in shader model 1 shaders\n");
587 WARN("Got %u int and %u boolean constants\n", shader->num_ci, shader->num_cb);
588 This->state = E_INVALIDARG;
589 return;
590 }
591
592 hr = find_ps_builtin_semantics(This, shader, 4);
593 if(FAILED(hr)) {
594 This->state = hr;
595 return;
596 }
597
598 /* Declare the shader type and version */
599 put_dword(buffer, This->version);
600 write_constF(shader, buffer, TRUE);
601 }
602
603 static void ps_1_4_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
604 HRESULT hr;
605
606 /* First check the constants and varyings, and complain if unsupported things are used */
607 if(shader->num_ci || shader->num_cb) {
608 WARN("Int and bool constants are not supported in shader model 1 shaders\n");
609 WARN("Got %u int and %u boolean constants\n", shader->num_ci, shader->num_cb);
610 This->state = E_INVALIDARG;
611 return;
612 }
613 hr = find_ps_builtin_semantics(This, shader, 6);
614 if(FAILED(hr)) {
615 This->state = hr;
616 return;
617 }
618
619 /* Declare the shader type and version */
620 put_dword(buffer, This->version);
621 write_constF(shader, buffer, TRUE);
622 }
623
624 static void end(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
625 put_dword(buffer, D3DSIO_END);
626 }
627
628 static DWORD map_vs_output(struct bc_writer *This, DWORD regnum, DWORD mask, DWORD *has_components) {
629 DWORD token = 0;
630 DWORD i;
631
632 *has_components = TRUE;
633 if(regnum == This->oPos_regnum) {
634 token |= (D3DSPR_RASTOUT << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
635 token |= D3DSRO_POSITION & D3DSP_REGNUM_MASK; /* No shift */
636 return token;
637 }
638 if(regnum == This->oFog_regnum && mask == This->oFog_mask) {
639 token |= (D3DSPR_RASTOUT << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
640 token |= D3DSRO_FOG & D3DSP_REGNUM_MASK; /* No shift */
641 token |= D3DSP_WRITEMASK_ALL;
642 *has_components = FALSE;
643 return token;
644 }
645 if(regnum == This->oPts_regnum && mask == This->oPts_mask) {
646 token |= (D3DSPR_RASTOUT << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
647 token |= D3DSRO_POINT_SIZE & D3DSP_REGNUM_MASK; /* No shift */
648 token |= D3DSP_WRITEMASK_ALL;
649 *has_components = FALSE;
650 return token;
651 }
652 for(i = 0; i < 2; i++) {
653 if(regnum == This->oD_regnum[i]) {
654 token |= (D3DSPR_ATTROUT << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
655 token |= i & D3DSP_REGNUM_MASK; /* No shift */
656 return token;
657 }
658 }
659 for(i = 0; i < 8; i++) {
660 if(regnum == This->oT_regnum[i]) {
661 token |= (D3DSPR_TEXCRDOUT << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
662 token |= i & D3DSP_REGNUM_MASK; /* No shift */
663 return token;
664 }
665 }
666
667 /* The varying must be undeclared - if an unsupported varying was declared,
668 * the vs_find_builtin_varyings function would have caught it and this code
669 * would not run */
670 WARN("Undeclared varying %u\n", regnum);
671 This->state = E_INVALIDARG;
672 return -1;
673 }
674
675 static void vs_12_dstreg(struct bc_writer *This, const struct shader_reg *reg,
676 struct bytecode_buffer *buffer,
677 DWORD shift, DWORD mod) {
678 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
679 DWORD has_wmask;
680
681 if(reg->rel_reg) {
682 WARN("Relative addressing not supported for destination registers\n");
683 This->state = E_INVALIDARG;
684 return;
685 }
686
687 switch(reg->type) {
688 case BWRITERSPR_OUTPUT:
689 token |= map_vs_output(This, reg->regnum, reg->writemask, &has_wmask);
690 break;
691
692 case BWRITERSPR_RASTOUT:
693 case BWRITERSPR_ATTROUT:
694 /* These registers are mapped to input and output regs. They can be encoded in the bytecode,
695 * but are unexpected. If we hit this path it might be due to an error.
696 */
697 FIXME("Unexpected register type %u\n", reg->type);
698 /* drop through */
699 case BWRITERSPR_INPUT:
700 case BWRITERSPR_TEMP:
701 case BWRITERSPR_CONST:
702 token |= (reg->type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
703 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
704 has_wmask = TRUE;
705 break;
706
707 case BWRITERSPR_ADDR:
708 if(reg->regnum != 0) {
709 WARN("Only a0 exists\n");
710 This->state = E_INVALIDARG;
711 return;
712 }
713 token |= (D3DSPR_ADDR << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
714 token |= 0 & D3DSP_REGNUM_MASK; /* No shift */
715 has_wmask = TRUE;
716 break;
717
718 case BWRITERSPR_PREDICATE:
719 if(This->version != BWRITERVS_VERSION(2, 1)){
720 WARN("Predicate register is allowed only in vs_2_x\n");
721 This->state = E_INVALIDARG;
722 return;
723 }
724 if(reg->regnum != 0) {
725 WARN("Only predicate register p0 exists\n");
726 This->state = E_INVALIDARG;
727 return;
728 }
729 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
730 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
731 token |= 0 & D3DSP_REGNUM_MASK; /* No shift */
732 has_wmask = TRUE;
733 break;
734
735 default:
736 WARN("Invalid register type for 1.x-2.x vertex shader\n");
737 This->state = E_INVALIDARG;
738 return;
739 }
740
741 /* strictly speaking there are no modifiers in vs_2_0 and vs_1_x, but they can be written
742 * into the bytecode and since the compiler doesn't do such checks write them
743 * (the checks are done by the undocumented shader validator)
744 */
745 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
746 token |= d3d9_dstmod(mod);
747
748 if(has_wmask) {
749 token |= d3d9_writemask(reg->writemask);
750 }
751 put_dword(buffer, token);
752 }
753
754 static void vs_1_x_srcreg(struct bc_writer *This, const struct shader_reg *reg,
755 struct bytecode_buffer *buffer) {
756 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
757 DWORD has_swizzle;
758 DWORD component;
759
760 switch(reg->type) {
761 case BWRITERSPR_OUTPUT:
762 /* Map the swizzle to a writemask, the format expected
763 by map_vs_output
764 */
765 switch(reg->swizzle) {
766 case BWRITERVS_SWIZZLE_X:
767 component = BWRITERSP_WRITEMASK_0;
768 break;
769 case BWRITERVS_SWIZZLE_Y:
770 component = BWRITERSP_WRITEMASK_1;
771 break;
772 case BWRITERVS_SWIZZLE_Z:
773 component = BWRITERSP_WRITEMASK_2;
774 break;
775 case BWRITERVS_SWIZZLE_W:
776 component = BWRITERSP_WRITEMASK_3;
777 break;
778 default:
779 component = 0;
780 }
781 token |= map_vs_output(This, reg->regnum, component, &has_swizzle);
782 break;
783
784 case BWRITERSPR_RASTOUT:
785 case BWRITERSPR_ATTROUT:
786 /* These registers are mapped to input and output regs. They can be encoded in the bytecode,
787 * but are unexpected. If we hit this path it might be due to an error.
788 */
789 FIXME("Unexpected register type %u\n", reg->type);
790 /* drop through */
791 case BWRITERSPR_INPUT:
792 case BWRITERSPR_TEMP:
793 case BWRITERSPR_CONST:
794 case BWRITERSPR_ADDR:
795 token |= (reg->type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
796 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
797 if(reg->rel_reg) {
798 if(reg->rel_reg->type != BWRITERSPR_ADDR ||
799 reg->rel_reg->regnum != 0 ||
800 reg->rel_reg->swizzle != BWRITERVS_SWIZZLE_X) {
801 WARN("Relative addressing in vs_1_x is only allowed with a0.x\n");
802 This->state = E_INVALIDARG;
803 return;
804 }
805 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
806 }
807 break;
808
809 default:
810 WARN("Invalid register type for 1.x vshader\n");
811 This->state = E_INVALIDARG;
812 return;
813 }
814
815 token |= d3d9_swizzle(reg->swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
816
817 token |= d3d9_srcmod(reg->srcmod);
818 put_dword(buffer, token);
819 }
820
821 static void write_srcregs(struct bc_writer *This, const struct instruction *instr,
822 struct bytecode_buffer *buffer){
823 unsigned int i;
824 if(instr->has_predicate){
825 This->funcs->srcreg(This, &instr->predicate, buffer);
826 }
827 for(i = 0; i < instr->num_srcs; i++){
828 This->funcs->srcreg(This, &instr->src[i], buffer);
829 }
830 }
831
832 static DWORD map_ps13_temp(struct bc_writer *This, const struct shader_reg *reg) {
833 DWORD token = 0;
834 if(reg->regnum == T0_REG) {
835 token |= (D3DSPR_TEXTURE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
836 token |= 0 & D3DSP_REGNUM_MASK; /* No shift */
837 } else if(reg->regnum == T1_REG) {
838 token |= (D3DSPR_TEXTURE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
839 token |= 1 & D3DSP_REGNUM_MASK; /* No shift */
840 } else if(reg->regnum == T2_REG) {
841 token |= (D3DSPR_TEXTURE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
842 token |= 2 & D3DSP_REGNUM_MASK; /* No shift */
843 } else if(reg->regnum == T3_REG) {
844 token |= (D3DSPR_TEXTURE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
845 token |= 3 & D3DSP_REGNUM_MASK; /* No shift */
846 } else {
847 token |= (D3DSPR_TEMP << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
848 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
849 }
850 return token;
851 }
852
853 static DWORD map_ps_input(struct bc_writer *This,
854 const struct shader_reg *reg) {
855 DWORD i, token = 0;
856 /* Map color interpolators */
857 for(i = 0; i < 2; i++) {
858 if(reg->regnum == This->v_regnum[i]) {
859 token |= (D3DSPR_INPUT << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
860 token |= i & D3DSP_REGNUM_MASK; /* No shift */
861 return token;
862 }
863 }
864 for(i = 0; i < 8; i++) {
865 if(reg->regnum == This->t_regnum[i]) {
866 token |= (D3DSPR_TEXTURE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
867 token |= i & D3DSP_REGNUM_MASK; /* No shift */
868 return token;
869 }
870 }
871
872 WARN("Invalid ps 1/2 varying\n");
873 This->state = E_INVALIDARG;
874 return token;
875 }
876
877 static void ps_1_0123_srcreg(struct bc_writer *This, const struct shader_reg *reg,
878 struct bytecode_buffer *buffer) {
879 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
880 if(reg->rel_reg) {
881 WARN("Relative addressing not supported in <= ps_3_0\n");
882 This->state = E_INVALIDARG;
883 return;
884 }
885
886 switch(reg->type) {
887 case BWRITERSPR_INPUT:
888 token |= map_ps_input(This, reg);
889 break;
890
891 /* Take care about the texture temporaries. There's a problem: They aren't
892 * declared anywhere, so we can only hardcode the values that are used
893 * to map ps_1_3 shaders to the common shader structure
894 */
895 case BWRITERSPR_TEMP:
896 token |= map_ps13_temp(This, reg);
897 break;
898
899 case BWRITERSPR_CONST: /* Can be mapped 1:1 */
900 token |= (reg->type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
901 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
902 break;
903
904 default:
905 WARN("Invalid register type for <= ps_1_3 shader\n");
906 This->state = E_INVALIDARG;
907 return;
908 }
909
910 token |= d3d9_swizzle(reg->swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
911
912 if(reg->srcmod == BWRITERSPSM_DZ || reg->srcmod == BWRITERSPSM_DW ||
913 reg->srcmod == BWRITERSPSM_ABS || reg->srcmod == BWRITERSPSM_ABSNEG ||
914 reg->srcmod == BWRITERSPSM_NOT) {
915 WARN("Invalid source modifier %u for <= ps_1_3\n", reg->srcmod);
916 This->state = E_INVALIDARG;
917 return;
918 }
919 token |= d3d9_srcmod(reg->srcmod);
920 put_dword(buffer, token);
921 }
922
923 static void ps_1_0123_dstreg(struct bc_writer *This, const struct shader_reg *reg,
924 struct bytecode_buffer *buffer,
925 DWORD shift, DWORD mod) {
926 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
927
928 if(reg->rel_reg) {
929 WARN("Relative addressing not supported for destination registers\n");
930 This->state = E_INVALIDARG;
931 return;
932 }
933
934 switch(reg->type) {
935 case BWRITERSPR_TEMP:
936 token |= map_ps13_temp(This, reg);
937 break;
938
939 /* texkill uses the input register as a destination parameter */
940 case BWRITERSPR_INPUT:
941 token |= map_ps_input(This, reg);
942 break;
943
944 default:
945 WARN("Invalid dest register type for 1.x pshader\n");
946 This->state = E_INVALIDARG;
947 return;
948 }
949
950 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
951 token |= d3d9_dstmod(mod);
952
953 token |= d3d9_writemask(reg->writemask);
954 put_dword(buffer, token);
955 }
956
957 /* The length of an instruction consists of the destination register (if any),
958 * the number of source registers, the number of address registers used for
959 * indirect addressing, and optionally the predicate register
960 */
961 static DWORD instrlen(const struct instruction *instr, unsigned int srcs, unsigned int dsts) {
962 unsigned int i;
963 DWORD ret = srcs + dsts + (instr->has_predicate ? 1 : 0);
964
965 if(dsts){
966 if(instr->dst.rel_reg) ret++;
967 }
968 for(i = 0; i < srcs; i++) {
969 if(instr->src[i].rel_reg) ret++;
970 }
971 return ret;
972 }
973
974 static void sm_1_x_opcode(struct bc_writer *This,
975 const struct instruction *instr,
976 DWORD token, struct bytecode_buffer *buffer) {
977 /* In sm_1_x instruction length isn't encoded */
978 if(instr->coissue){
979 token |= D3DSI_COISSUE;
980 }
981 put_dword(buffer, token);
982 }
983
984 static void instr_handler(struct bc_writer *This,
985 const struct instruction *instr,
986 struct bytecode_buffer *buffer) {
987 DWORD token = d3d9_opcode(instr->opcode);
988
989 This->funcs->opcode(This, instr, token, buffer);
990 if(instr->has_dst) This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
991 write_srcregs(This, instr, buffer);
992 }
993
994 static const struct instr_handler_table vs_1_x_handlers[] = {
995 {BWRITERSIO_ADD, instr_handler},
996 {BWRITERSIO_NOP, instr_handler},
997 {BWRITERSIO_MOV, instr_handler},
998 {BWRITERSIO_SUB, instr_handler},
999 {BWRITERSIO_MAD, instr_handler},
1000 {BWRITERSIO_MUL, instr_handler},
1001 {BWRITERSIO_RCP, instr_handler},
1002 {BWRITERSIO_RSQ, instr_handler},
1003 {BWRITERSIO_DP3, instr_handler},
1004 {BWRITERSIO_DP4, instr_handler},
1005 {BWRITERSIO_MIN, instr_handler},
1006 {BWRITERSIO_MAX, instr_handler},
1007 {BWRITERSIO_SLT, instr_handler},
1008 {BWRITERSIO_SGE, instr_handler},
1009 {BWRITERSIO_EXP, instr_handler},
1010 {BWRITERSIO_LOG, instr_handler},
1011 {BWRITERSIO_EXPP, instr_handler},
1012 {BWRITERSIO_LOGP, instr_handler},
1013 {BWRITERSIO_DST, instr_handler},
1014 {BWRITERSIO_FRC, instr_handler},
1015 {BWRITERSIO_M4x4, instr_handler},
1016 {BWRITERSIO_M4x3, instr_handler},
1017 {BWRITERSIO_M3x4, instr_handler},
1018 {BWRITERSIO_M3x3, instr_handler},
1019 {BWRITERSIO_M3x2, instr_handler},
1020 {BWRITERSIO_LIT, instr_handler},
1021
1022 {BWRITERSIO_END, NULL}, /* Sentinel value, it signals
1023 the end of the list */
1024 };
1025
1026 static const struct bytecode_backend vs_1_x_backend = {
1027 vs_1_x_header,
1028 end,
1029 vs_1_x_srcreg,
1030 vs_12_dstreg,
1031 sm_1_x_opcode,
1032 vs_1_x_handlers
1033 };
1034
1035 static void instr_ps_1_0123_texld(struct bc_writer *This,
1036 const struct instruction *instr,
1037 struct bytecode_buffer *buffer) {
1038 DWORD idx;
1039 struct shader_reg reg;
1040 DWORD swizzlemask;
1041
1042 if(instr->src[1].type != BWRITERSPR_SAMPLER ||
1043 instr->src[1].regnum > 3) {
1044 WARN("Unsupported sampler type %u regnum %u\n",
1045 instr->src[1].type, instr->src[1].regnum);
1046 This->state = E_INVALIDARG;
1047 return;
1048 } else if(instr->dst.type != BWRITERSPR_TEMP) {
1049 WARN("Can only sample into a temp register\n");
1050 This->state = E_INVALIDARG;
1051 return;
1052 }
1053
1054 idx = instr->src[1].regnum;
1055 if((idx == 0 && instr->dst.regnum != T0_REG) ||
1056 (idx == 1 && instr->dst.regnum != T1_REG) ||
1057 (idx == 2 && instr->dst.regnum != T2_REG) ||
1058 (idx == 3 && instr->dst.regnum != T3_REG)) {
1059 WARN("Sampling from sampler s%u to register r%u is not possible in ps_1_x\n",
1060 idx, instr->dst.regnum);
1061 This->state = E_INVALIDARG;
1062 return;
1063 }
1064 if(instr->src[0].type == BWRITERSPR_INPUT) {
1065 /* A simple non-dependent read tex instruction */
1066 if(instr->src[0].regnum != This->t_regnum[idx]) {
1067 WARN("Cannot sample from s%u with texture address data from interpolator %u\n",
1068 idx, instr->src[0].regnum);
1069 This->state = E_INVALIDARG;
1070 return;
1071 }
1072 This->funcs->opcode(This, instr, D3DSIO_TEX & D3DSI_OPCODE_MASK, buffer);
1073
1074 /* map the temp dstreg to the ps_1_3 texture temporary register */
1075 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1076 } else if(instr->src[0].type == BWRITERSPR_TEMP) {
1077
1078 swizzlemask = (3 << BWRITERVS_SWIZZLE_SHIFT) |
1079 (3 << (BWRITERVS_SWIZZLE_SHIFT + 2)) |
1080 (3 << (BWRITERVS_SWIZZLE_SHIFT + 4));
1081 if((instr->src[0].swizzle & swizzlemask) == (BWRITERVS_X_X | BWRITERVS_Y_Y | BWRITERVS_Z_Z)) {
1082 TRACE("writing texreg2rgb\n");
1083 This->funcs->opcode(This, instr, D3DSIO_TEXREG2RGB & D3DSI_OPCODE_MASK, buffer);
1084 } else if(instr->src[0].swizzle == (BWRITERVS_X_W | BWRITERVS_Y_X | BWRITERVS_Z_X | BWRITERVS_W_X)) {
1085 TRACE("writing texreg2ar\n");
1086 This->funcs->opcode(This, instr, D3DSIO_TEXREG2AR & D3DSI_OPCODE_MASK, buffer);
1087 } else if(instr->src[0].swizzle == (BWRITERVS_X_Y | BWRITERVS_Y_Z | BWRITERVS_Z_Z | BWRITERVS_W_Z)) {
1088 TRACE("writing texreg2gb\n");
1089 This->funcs->opcode(This, instr, D3DSIO_TEXREG2GB & D3DSI_OPCODE_MASK, buffer);
1090 } else {
1091 WARN("Unsupported src addr swizzle in dependent texld: 0x%08x\n", instr->src[0].swizzle);
1092 This->state = E_INVALIDARG;
1093 return;
1094 }
1095
1096 /* Dst and src reg can be mapped normally. Both registers are temporary registers in the
1097 * source shader and have to be mapped to the temporary form of the texture registers. However,
1098 * the src reg doesn't have a swizzle
1099 */
1100 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1101 reg = instr->src[0];
1102 reg.swizzle = BWRITERVS_NOSWIZZLE;
1103 This->funcs->srcreg(This, &reg, buffer);
1104 } else {
1105 WARN("Invalid address data source register\n");
1106 This->state = E_INVALIDARG;
1107 return;
1108 }
1109 }
1110
1111 static void instr_ps_1_0123_mov(struct bc_writer *This,
1112 const struct instruction *instr,
1113 struct bytecode_buffer *buffer) {
1114 DWORD token = D3DSIO_MOV & D3DSI_OPCODE_MASK;
1115
1116 if(instr->dst.type == BWRITERSPR_TEMP && instr->src[0].type == BWRITERSPR_INPUT) {
1117 if((instr->dst.regnum == T0_REG && instr->src[0].regnum == This->t_regnum[0]) ||
1118 (instr->dst.regnum == T1_REG && instr->src[0].regnum == This->t_regnum[1]) ||
1119 (instr->dst.regnum == T2_REG && instr->src[0].regnum == This->t_regnum[2]) ||
1120 (instr->dst.regnum == T3_REG && instr->src[0].regnum == This->t_regnum[3])) {
1121 if(instr->dstmod & BWRITERSPDM_SATURATE) {
1122 This->funcs->opcode(This, instr, D3DSIO_TEXCOORD & D3DSI_OPCODE_MASK, buffer);
1123 /* Remove the SATURATE flag, it's implicit to the instruction */
1124 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod & (~BWRITERSPDM_SATURATE));
1125 return;
1126 } else {
1127 WARN("A varying -> temp copy is only supported with the SATURATE modifier in <=ps_1_3\n");
1128 This->state = E_INVALIDARG;
1129 return;
1130 }
1131 } else if(instr->src[0].regnum == This->v_regnum[0] ||
1132 instr->src[0].regnum == This->v_regnum[1]) {
1133 /* Handled by the normal mov below. Just drop out of the if condition */
1134 } else {
1135 WARN("Unsupported varying -> temp mov in <= ps_1_3\n");
1136 This->state = E_INVALIDARG;
1137 return;
1138 }
1139 }
1140
1141 This->funcs->opcode(This, instr, token, buffer);
1142 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1143 This->funcs->srcreg(This, &instr->src[0], buffer);
1144 }
1145
1146 static const struct instr_handler_table ps_1_0123_handlers[] = {
1147 {BWRITERSIO_ADD, instr_handler},
1148 {BWRITERSIO_NOP, instr_handler},
1149 {BWRITERSIO_MOV, instr_ps_1_0123_mov},
1150 {BWRITERSIO_SUB, instr_handler},
1151 {BWRITERSIO_MAD, instr_handler},
1152 {BWRITERSIO_MUL, instr_handler},
1153 {BWRITERSIO_DP3, instr_handler},
1154 {BWRITERSIO_DP4, instr_handler},
1155 {BWRITERSIO_LRP, instr_handler},
1156
1157 /* pshader instructions */
1158 {BWRITERSIO_CND, instr_handler},
1159 {BWRITERSIO_CMP, instr_handler},
1160 {BWRITERSIO_TEXKILL, instr_handler},
1161 {BWRITERSIO_TEX, instr_ps_1_0123_texld},
1162 {BWRITERSIO_TEXBEM, instr_handler},
1163 {BWRITERSIO_TEXBEML, instr_handler},
1164 {BWRITERSIO_TEXM3x2PAD, instr_handler},
1165 {BWRITERSIO_TEXM3x3PAD, instr_handler},
1166 {BWRITERSIO_TEXM3x3SPEC, instr_handler},
1167 {BWRITERSIO_TEXM3x3VSPEC, instr_handler},
1168 {BWRITERSIO_TEXM3x3TEX, instr_handler},
1169 {BWRITERSIO_TEXM3x3, instr_handler},
1170 {BWRITERSIO_TEXM3x2DEPTH, instr_handler},
1171 {BWRITERSIO_TEXM3x2TEX, instr_handler},
1172 {BWRITERSIO_TEXDP3, instr_handler},
1173 {BWRITERSIO_TEXDP3TEX, instr_handler},
1174 {BWRITERSIO_END, NULL},
1175 };
1176
1177 static const struct bytecode_backend ps_1_0123_backend = {
1178 ps_1_x_header,
1179 end,
1180 ps_1_0123_srcreg,
1181 ps_1_0123_dstreg,
1182 sm_1_x_opcode,
1183 ps_1_0123_handlers
1184 };
1185
1186 static void ps_1_4_srcreg(struct bc_writer *This, const struct shader_reg *reg,
1187 struct bytecode_buffer *buffer) {
1188 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1189 if(reg->rel_reg) {
1190 WARN("Relative addressing not supported in <= ps_3_0\n");
1191 This->state = E_INVALIDARG;
1192 return;
1193 }
1194
1195 switch(reg->type) {
1196 case BWRITERSPR_INPUT:
1197 token |= map_ps_input(This, reg);
1198 break;
1199
1200 /* Can be mapped 1:1 */
1201 case BWRITERSPR_TEMP:
1202 case BWRITERSPR_CONST:
1203 token |= (reg->type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1204 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1205 break;
1206
1207 default:
1208 WARN("Invalid register type for ps_1_4 shader\n");
1209 This->state = E_INVALIDARG;
1210 return;
1211 }
1212
1213 token |= d3d9_swizzle(reg->swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1214
1215 if(reg->srcmod == BWRITERSPSM_ABS || reg->srcmod == BWRITERSPSM_ABSNEG ||
1216 reg->srcmod == BWRITERSPSM_NOT) {
1217 WARN("Invalid source modifier %u for ps_1_4\n", reg->srcmod);
1218 This->state = E_INVALIDARG;
1219 return;
1220 }
1221 token |= d3d9_srcmod(reg->srcmod);
1222 put_dword(buffer, token);
1223 }
1224
1225 static void ps_1_4_dstreg(struct bc_writer *This, const struct shader_reg *reg,
1226 struct bytecode_buffer *buffer,
1227 DWORD shift, DWORD mod) {
1228 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1229
1230 if(reg->rel_reg) {
1231 WARN("Relative addressing not supported for destination registers\n");
1232 This->state = E_INVALIDARG;
1233 return;
1234 }
1235
1236 switch(reg->type) {
1237 case BWRITERSPR_TEMP: /* 1:1 mapping */
1238 token |= (reg->type << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1239 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1240 break;
1241
1242 /* For texkill */
1243 case BWRITERSPR_INPUT:
1244 token |= map_ps_input(This, reg);
1245 break;
1246
1247 default:
1248 WARN("Invalid dest register type for 1.x pshader\n");
1249 This->state = E_INVALIDARG;
1250 return;
1251 }
1252
1253 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
1254 token |= d3d9_dstmod(mod);
1255
1256 token |= d3d9_writemask(reg->writemask);
1257 put_dword(buffer, token);
1258 }
1259
1260 static void instr_ps_1_4_mov(struct bc_writer *This,
1261 const struct instruction *instr,
1262 struct bytecode_buffer *buffer) {
1263 DWORD token = D3DSIO_MOV & D3DSI_OPCODE_MASK;
1264
1265 if(instr->dst.type == BWRITERSPR_TEMP && instr->src[0].type == BWRITERSPR_INPUT) {
1266 if(instr->src[0].regnum == This->t_regnum[0] ||
1267 instr->src[0].regnum == This->t_regnum[1] ||
1268 instr->src[0].regnum == This->t_regnum[2] ||
1269 instr->src[0].regnum == This->t_regnum[3] ||
1270 instr->src[0].regnum == This->t_regnum[4] ||
1271 instr->src[0].regnum == This->t_regnum[5]) {
1272 /* Similar to a regular mov, but a different opcode */
1273 token = D3DSIO_TEXCOORD & D3DSI_OPCODE_MASK;
1274 } else if(instr->src[0].regnum == This->v_regnum[0] ||
1275 instr->src[0].regnum == This->v_regnum[1]) {
1276 /* Handled by the normal mov below. Just drop out of the if condition */
1277 } else {
1278 WARN("Unsupported varying -> temp mov in ps_1_4\n");
1279 This->state = E_INVALIDARG;
1280 return;
1281 }
1282 }
1283
1284 This->funcs->opcode(This, instr, token, buffer);
1285 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1286 This->funcs->srcreg(This, &instr->src[0], buffer);
1287 }
1288
1289 static void instr_ps_1_4_texld(struct bc_writer *This,
1290 const struct instruction *instr,
1291 struct bytecode_buffer *buffer) {
1292 if(instr->src[1].type != BWRITERSPR_SAMPLER ||
1293 instr->src[1].regnum > 5) {
1294 WARN("Unsupported sampler type %u regnum %u\n",
1295 instr->src[1].type, instr->src[1].regnum);
1296 This->state = E_INVALIDARG;
1297 return;
1298 } else if(instr->dst.type != BWRITERSPR_TEMP) {
1299 WARN("Can only sample into a temp register\n");
1300 This->state = E_INVALIDARG;
1301 return;
1302 }
1303
1304 if(instr->src[1].regnum != instr->dst.regnum) {
1305 WARN("Sampling from sampler s%u to register r%u is not possible in ps_1_4\n",
1306 instr->src[1].regnum, instr->dst.regnum);
1307 This->state = E_INVALIDARG;
1308 return;
1309 }
1310
1311 This->funcs->opcode(This, instr, D3DSIO_TEX & D3DSI_OPCODE_MASK, buffer);
1312 This->funcs->dstreg(This, &instr->dst, buffer, instr->shift, instr->dstmod);
1313 This->funcs->srcreg(This, &instr->src[0], buffer);
1314 }
1315
1316 static const struct instr_handler_table ps_1_4_handlers[] = {
1317 {BWRITERSIO_ADD, instr_handler},
1318 {BWRITERSIO_NOP, instr_handler},
1319 {BWRITERSIO_MOV, instr_ps_1_4_mov},
1320 {BWRITERSIO_SUB, instr_handler},
1321 {BWRITERSIO_MAD, instr_handler},
1322 {BWRITERSIO_MUL, instr_handler},
1323 {BWRITERSIO_DP3, instr_handler},
1324 {BWRITERSIO_DP4, instr_handler},
1325 {BWRITERSIO_LRP, instr_handler},
1326
1327 /* pshader instructions */
1328 {BWRITERSIO_CND, instr_handler},
1329 {BWRITERSIO_CMP, instr_handler},
1330 {BWRITERSIO_TEXKILL, instr_handler},
1331 {BWRITERSIO_TEX, instr_ps_1_4_texld},
1332 {BWRITERSIO_TEXDEPTH, instr_handler},
1333 {BWRITERSIO_BEM, instr_handler},
1334
1335 {BWRITERSIO_PHASE, instr_handler},
1336 {BWRITERSIO_END, NULL},
1337 };
1338
1339 static const struct bytecode_backend ps_1_4_backend = {
1340 ps_1_4_header,
1341 end,
1342 ps_1_4_srcreg,
1343 ps_1_4_dstreg,
1344 sm_1_x_opcode,
1345 ps_1_4_handlers
1346 };
1347
1348 static void write_constB(const struct bwriter_shader *shader, struct bytecode_buffer *buffer, BOOL len) {
1349 write_const(shader->constB, shader->num_cb, D3DSIO_DEFB, D3DSPR_CONSTBOOL, buffer, len);
1350 }
1351
1352 static void write_constI(const struct bwriter_shader *shader, struct bytecode_buffer *buffer, BOOL len) {
1353 write_const(shader->constI, shader->num_ci, D3DSIO_DEFI, D3DSPR_CONSTINT, buffer, len);
1354 }
1355
1356 static void vs_2_header(struct bc_writer *This,
1357 const struct bwriter_shader *shader,
1358 struct bytecode_buffer *buffer) {
1359 HRESULT hr;
1360
1361 hr = vs_find_builtin_varyings(This, shader);
1362 if(FAILED(hr)) {
1363 This->state = hr;
1364 return;
1365 }
1366
1367 /* Declare the shader type and version */
1368 put_dword(buffer, This->version);
1369
1370 write_declarations(This, buffer, TRUE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
1371 write_constF(shader, buffer, TRUE);
1372 write_constB(shader, buffer, TRUE);
1373 write_constI(shader, buffer, TRUE);
1374 return;
1375 }
1376
1377 static void vs_2_srcreg(struct bc_writer *This,
1378 const struct shader_reg *reg,
1379 struct bytecode_buffer *buffer) {
1380 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1381 DWORD has_swizzle;
1382 DWORD component;
1383 DWORD d3d9reg;
1384
1385 switch(reg->type) {
1386 case BWRITERSPR_OUTPUT:
1387 /* Map the swizzle to a writemask, the format expected
1388 by map_vs_output
1389 */
1390 switch(reg->swizzle) {
1391 case BWRITERVS_SWIZZLE_X:
1392 component = BWRITERSP_WRITEMASK_0;
1393 break;
1394 case BWRITERVS_SWIZZLE_Y:
1395 component = BWRITERSP_WRITEMASK_1;
1396 break;
1397 case BWRITERVS_SWIZZLE_Z:
1398 component = BWRITERSP_WRITEMASK_2;
1399 break;
1400 case BWRITERVS_SWIZZLE_W:
1401 component = BWRITERSP_WRITEMASK_3;
1402 break;
1403 default:
1404 component = 0;
1405 }
1406 token |= map_vs_output(This, reg->regnum, component, &has_swizzle);
1407 break;
1408
1409 case BWRITERSPR_RASTOUT:
1410 case BWRITERSPR_ATTROUT:
1411 /* These registers are mapped to input and output regs. They can be encoded in the bytecode,
1412 * but are unexpected. If we hit this path it might be due to an error.
1413 */
1414 FIXME("Unexpected register type %u\n", reg->type);
1415 /* drop through */
1416 case BWRITERSPR_INPUT:
1417 case BWRITERSPR_TEMP:
1418 case BWRITERSPR_CONST:
1419 case BWRITERSPR_ADDR:
1420 case BWRITERSPR_CONSTINT:
1421 case BWRITERSPR_CONSTBOOL:
1422 case BWRITERSPR_LABEL:
1423 d3d9reg = d3d9_register(reg->type);
1424 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1425 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1426 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1427 break;
1428
1429 case BWRITERSPR_LOOP:
1430 if(reg->regnum != 0) {
1431 WARN("Only regnum 0 is supported for the loop index register in vs_2_0\n");
1432 This->state = E_INVALIDARG;
1433 return;
1434 }
1435 token |= (D3DSPR_LOOP << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1436 token |= (D3DSPR_LOOP << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1437 token |= 0 & D3DSP_REGNUM_MASK; /* No shift */
1438 break;
1439
1440 case BWRITERSPR_PREDICATE:
1441 if(This->version != BWRITERVS_VERSION(2, 1)){
1442 WARN("Predicate register is allowed only in vs_2_x\n");
1443 This->state = E_INVALIDARG;
1444 return;
1445 }
1446 if(reg->regnum > 0) {
1447 WARN("Only predicate register 0 is supported\n");
1448 This->state = E_INVALIDARG;
1449 return;
1450 }
1451 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1452 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1453 token |= 0 & D3DSP_REGNUM_MASK; /* No shift */
1454
1455 break;
1456
1457 default:
1458 WARN("Invalid register type for 2.0 vshader\n");
1459 This->state = E_INVALIDARG;
1460 return;
1461 }
1462
1463 token |= d3d9_swizzle(reg->swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1464
1465 token |= d3d9_srcmod(reg->srcmod);
1466
1467 if(reg->rel_reg)
1468 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
1469
1470 put_dword(buffer, token);
1471
1472 /* vs_2_0 and newer write the register containing the index explicitly in the
1473 * binary code
1474 */
1475 if(token & D3DVS_ADDRMODE_RELATIVE)
1476 vs_2_srcreg(This, reg->rel_reg, buffer);
1477 }
1478
1479 static void sm_2_opcode(struct bc_writer *This,
1480 const struct instruction *instr,
1481 DWORD token, struct bytecode_buffer *buffer) {
1482 /* From sm 2 onwards instruction length is encoded in the opcode field */
1483 int dsts = instr->has_dst ? 1 : 0;
1484 token |= instrlen(instr, instr->num_srcs, dsts) << D3DSI_INSTLENGTH_SHIFT;
1485 if(instr->comptype)
1486 token |= (d3d9_comparetype(instr->comptype) << 16) & (0xf << 16);
1487 if(instr->has_predicate)
1488 token |= D3DSHADER_INSTRUCTION_PREDICATED;
1489 put_dword(buffer,token);
1490 }
1491
1492 static const struct instr_handler_table vs_2_0_handlers[] = {
1493 {BWRITERSIO_ADD, instr_handler},
1494 {BWRITERSIO_NOP, instr_handler},
1495 {BWRITERSIO_MOV, instr_handler},
1496 {BWRITERSIO_SUB, instr_handler},
1497 {BWRITERSIO_MAD, instr_handler},
1498 {BWRITERSIO_MUL, instr_handler},
1499 {BWRITERSIO_RCP, instr_handler},
1500 {BWRITERSIO_RSQ, instr_handler},
1501 {BWRITERSIO_DP3, instr_handler},
1502 {BWRITERSIO_DP4, instr_handler},
1503 {BWRITERSIO_MIN, instr_handler},
1504 {BWRITERSIO_MAX, instr_handler},
1505 {BWRITERSIO_SLT, instr_handler},
1506 {BWRITERSIO_SGE, instr_handler},
1507 {BWRITERSIO_ABS, instr_handler},
1508 {BWRITERSIO_EXP, instr_handler},
1509 {BWRITERSIO_LOG, instr_handler},
1510 {BWRITERSIO_EXPP, instr_handler},
1511 {BWRITERSIO_LOGP, instr_handler},
1512 {BWRITERSIO_DST, instr_handler},
1513 {BWRITERSIO_LRP, instr_handler},
1514 {BWRITERSIO_FRC, instr_handler},
1515 {BWRITERSIO_CRS, instr_handler},
1516 {BWRITERSIO_SGN, instr_handler},
1517 {BWRITERSIO_NRM, instr_handler},
1518 {BWRITERSIO_SINCOS, instr_handler},
1519 {BWRITERSIO_M4x4, instr_handler},
1520 {BWRITERSIO_M4x3, instr_handler},
1521 {BWRITERSIO_M3x4, instr_handler},
1522 {BWRITERSIO_M3x3, instr_handler},
1523 {BWRITERSIO_M3x2, instr_handler},
1524 {BWRITERSIO_LIT, instr_handler},
1525 {BWRITERSIO_POW, instr_handler},
1526 {BWRITERSIO_MOVA, instr_handler},
1527
1528 {BWRITERSIO_CALL, instr_handler},
1529 {BWRITERSIO_CALLNZ, instr_handler},
1530 {BWRITERSIO_REP, instr_handler},
1531 {BWRITERSIO_ENDREP, instr_handler},
1532 {BWRITERSIO_IF, instr_handler},
1533 {BWRITERSIO_LABEL, instr_handler},
1534 {BWRITERSIO_ELSE, instr_handler},
1535 {BWRITERSIO_ENDIF, instr_handler},
1536 {BWRITERSIO_LOOP, instr_handler},
1537 {BWRITERSIO_RET, instr_handler},
1538 {BWRITERSIO_ENDLOOP, instr_handler},
1539
1540 {BWRITERSIO_END, NULL},
1541 };
1542
1543 static const struct bytecode_backend vs_2_0_backend = {
1544 vs_2_header,
1545 end,
1546 vs_2_srcreg,
1547 vs_12_dstreg,
1548 sm_2_opcode,
1549 vs_2_0_handlers
1550 };
1551
1552 static const struct instr_handler_table vs_2_x_handlers[] = {
1553 {BWRITERSIO_ADD, instr_handler},
1554 {BWRITERSIO_NOP, instr_handler},
1555 {BWRITERSIO_MOV, instr_handler},
1556 {BWRITERSIO_SUB, instr_handler},
1557 {BWRITERSIO_MAD, instr_handler},
1558 {BWRITERSIO_MUL, instr_handler},
1559 {BWRITERSIO_RCP, instr_handler},
1560 {BWRITERSIO_RSQ, instr_handler},
1561 {BWRITERSIO_DP3, instr_handler},
1562 {BWRITERSIO_DP4, instr_handler},
1563 {BWRITERSIO_MIN, instr_handler},
1564 {BWRITERSIO_MAX, instr_handler},
1565 {BWRITERSIO_SLT, instr_handler},
1566 {BWRITERSIO_SGE, instr_handler},
1567 {BWRITERSIO_ABS, instr_handler},
1568 {BWRITERSIO_EXP, instr_handler},
1569 {BWRITERSIO_LOG, instr_handler},
1570 {BWRITERSIO_EXPP, instr_handler},
1571 {BWRITERSIO_LOGP, instr_handler},
1572 {BWRITERSIO_DST, instr_handler},
1573 {BWRITERSIO_LRP, instr_handler},
1574 {BWRITERSIO_FRC, instr_handler},
1575 {BWRITERSIO_CRS, instr_handler},
1576 {BWRITERSIO_SGN, instr_handler},
1577 {BWRITERSIO_NRM, instr_handler},
1578 {BWRITERSIO_SINCOS, instr_handler},
1579 {BWRITERSIO_M4x4, instr_handler},
1580 {BWRITERSIO_M4x3, instr_handler},
1581 {BWRITERSIO_M3x4, instr_handler},
1582 {BWRITERSIO_M3x3, instr_handler},
1583 {BWRITERSIO_M3x2, instr_handler},
1584 {BWRITERSIO_LIT, instr_handler},
1585 {BWRITERSIO_POW, instr_handler},
1586 {BWRITERSIO_MOVA, instr_handler},
1587
1588 {BWRITERSIO_CALL, instr_handler},
1589 {BWRITERSIO_CALLNZ, instr_handler},
1590 {BWRITERSIO_REP, instr_handler},
1591 {BWRITERSIO_ENDREP, instr_handler},
1592 {BWRITERSIO_IF, instr_handler},
1593 {BWRITERSIO_LABEL, instr_handler},
1594 {BWRITERSIO_IFC, instr_handler},
1595 {BWRITERSIO_ELSE, instr_handler},
1596 {BWRITERSIO_ENDIF, instr_handler},
1597 {BWRITERSIO_BREAK, instr_handler},
1598 {BWRITERSIO_BREAKC, instr_handler},
1599 {BWRITERSIO_LOOP, instr_handler},
1600 {BWRITERSIO_RET, instr_handler},
1601 {BWRITERSIO_ENDLOOP, instr_handler},
1602
1603 {BWRITERSIO_SETP, instr_handler},
1604 {BWRITERSIO_BREAKP, instr_handler},
1605
1606 {BWRITERSIO_END, NULL},
1607 };
1608
1609 static const struct bytecode_backend vs_2_x_backend = {
1610 vs_2_header,
1611 end,
1612 vs_2_srcreg,
1613 vs_12_dstreg,
1614 sm_2_opcode,
1615 vs_2_x_handlers
1616 };
1617
1618 static void write_samplers(const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
1619 DWORD i;
1620 DWORD instr_dcl = D3DSIO_DCL | (2 << D3DSI_INSTLENGTH_SHIFT);
1621 DWORD token;
1622 const DWORD reg = (1<<31) |
1623 ((D3DSPR_SAMPLER << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK) |
1624 ((D3DSPR_SAMPLER << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2) |
1625 D3DSP_WRITEMASK_ALL;
1626
1627 for(i = 0; i < shader->num_samplers; i++) {
1628 /* Write the DCL instruction */
1629 put_dword(buffer, instr_dcl);
1630 token = (1<<31);
1631 /* Already shifted */
1632 token |= (d3d9_sampler(shader->samplers[i].type)) & D3DSP_TEXTURETYPE_MASK;
1633 put_dword(buffer, token);
1634 token = reg | (shader->samplers[i].regnum & D3DSP_REGNUM_MASK);
1635 token |= d3d9_dstmod(shader->samplers[i].mod);
1636 put_dword(buffer, token);
1637 }
1638 }
1639
1640 static void ps_2_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
1641 HRESULT hr = find_ps_builtin_semantics(This, shader, 8);
1642 if(FAILED(hr)) {
1643 This->state = hr;
1644 return;
1645 }
1646
1647 /* Declare the shader type and version */
1648 put_dword(buffer, This->version);
1649 write_declarations(This, buffer, TRUE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
1650 write_samplers(shader, buffer);
1651 write_constF(shader, buffer, TRUE);
1652 write_constB(shader, buffer, TRUE);
1653 write_constI(shader, buffer, TRUE);
1654 }
1655
1656 static void ps_2_srcreg(struct bc_writer *This,
1657 const struct shader_reg *reg,
1658 struct bytecode_buffer *buffer) {
1659 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1660 DWORD d3d9reg;
1661 if(reg->rel_reg) {
1662 WARN("Relative addressing not supported in <= ps_3_0\n");
1663 This->state = E_INVALIDARG;
1664 return;
1665 }
1666
1667 switch(reg->type) {
1668 case BWRITERSPR_INPUT:
1669 token |= map_ps_input(This, reg);
1670 break;
1671
1672 /* Can be mapped 1:1 */
1673 case BWRITERSPR_TEMP:
1674 case BWRITERSPR_CONST:
1675 case BWRITERSPR_COLOROUT:
1676 case BWRITERSPR_CONSTBOOL:
1677 case BWRITERSPR_CONSTINT:
1678 case BWRITERSPR_SAMPLER:
1679 case BWRITERSPR_LABEL:
1680 case BWRITERSPR_DEPTHOUT:
1681 d3d9reg = d3d9_register(reg->type);
1682 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1683 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1684 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1685 break;
1686
1687 case BWRITERSPR_PREDICATE:
1688 if(This->version != BWRITERPS_VERSION(2, 1)){
1689 WARN("Predicate register not supported in ps_2_0\n");
1690 This->state = E_INVALIDARG;
1691 }
1692 if(reg->regnum) {
1693 WARN("Predicate register with regnum %u not supported\n",
1694 reg->regnum);
1695 This->state = E_INVALIDARG;
1696 }
1697 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1698 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1699 token |= 0 & D3DSP_REGNUM_MASK; /* No shift */
1700 break;
1701
1702 default:
1703 WARN("Invalid register type for ps_2_0 shader\n");
1704 This->state = E_INVALIDARG;
1705 return;
1706 }
1707
1708 token |= d3d9_swizzle(reg->swizzle) & D3DVS_SWIZZLE_MASK; /* already shifted */
1709
1710 token |= d3d9_srcmod(reg->srcmod);
1711 put_dword(buffer, token);
1712 }
1713
1714 static void ps_2_0_dstreg(struct bc_writer *This,
1715 const struct shader_reg *reg,
1716 struct bytecode_buffer *buffer,
1717 DWORD shift, DWORD mod) {
1718 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1719 DWORD d3d9reg;
1720
1721 if(reg->rel_reg) {
1722 WARN("Relative addressing not supported for destination registers\n");
1723 This->state = E_INVALIDARG;
1724 return;
1725 }
1726
1727 switch(reg->type) {
1728 case BWRITERSPR_TEMP: /* 1:1 mapping */
1729 case BWRITERSPR_COLOROUT:
1730 case BWRITERSPR_DEPTHOUT:
1731 d3d9reg = d3d9_register(reg->type);
1732 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1733 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1734 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1735 break;
1736
1737 case BWRITERSPR_PREDICATE:
1738 if(This->version != BWRITERPS_VERSION(2, 1)){
1739 WARN("Predicate register not supported in ps_2_0\n");
1740 This->state = E_INVALIDARG;
1741 }
1742 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1743 token |= (D3DSPR_PREDICATE << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1744 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1745 break;
1746
1747 /* texkill uses the input register as a destination parameter */
1748 case BWRITERSPR_INPUT:
1749 token |= map_ps_input(This, reg);
1750 break;
1751
1752 default:
1753 WARN("Invalid dest register type for 2.x pshader\n");
1754 This->state = E_INVALIDARG;
1755 return;
1756 }
1757
1758 token |= (shift << D3DSP_DSTSHIFT_SHIFT) & D3DSP_DSTSHIFT_MASK;
1759 token |= d3d9_dstmod(mod);
1760
1761 token |= d3d9_writemask(reg->writemask);
1762 put_dword(buffer, token);
1763 }
1764
1765 static const struct instr_handler_table ps_2_0_handlers[] = {
1766 {BWRITERSIO_ADD, instr_handler},
1767 {BWRITERSIO_NOP, instr_handler},
1768 {BWRITERSIO_MOV, instr_handler},
1769 {BWRITERSIO_SUB, instr_handler},
1770 {BWRITERSIO_MAD, instr_handler},
1771 {BWRITERSIO_MUL, instr_handler},
1772 {BWRITERSIO_RCP, instr_handler},
1773 {BWRITERSIO_RSQ, instr_handler},
1774 {BWRITERSIO_DP3, instr_handler},
1775 {BWRITERSIO_DP4, instr_handler},
1776 {BWRITERSIO_MIN, instr_handler},
1777 {BWRITERSIO_MAX, instr_handler},
1778 {BWRITERSIO_ABS, instr_handler},
1779 {BWRITERSIO_EXP, instr_handler},
1780 {BWRITERSIO_LOG, instr_handler},
1781 {BWRITERSIO_EXPP, instr_handler},
1782 {BWRITERSIO_LOGP, instr_handler},
1783 {BWRITERSIO_LRP, instr_handler},
1784 {BWRITERSIO_FRC, instr_handler},
1785 {BWRITERSIO_CRS, instr_handler},
1786 {BWRITERSIO_NRM, instr_handler},
1787 {BWRITERSIO_SINCOS, instr_handler},
1788 {BWRITERSIO_M4x4, instr_handler},
1789 {BWRITERSIO_M4x3, instr_handler},
1790 {BWRITERSIO_M3x4, instr_handler},
1791 {BWRITERSIO_M3x3, instr_handler},
1792 {BWRITERSIO_M3x2, instr_handler},
1793 {BWRITERSIO_POW, instr_handler},
1794 {BWRITERSIO_DP2ADD, instr_handler},
1795 {BWRITERSIO_CMP, instr_handler},
1796
1797 {BWRITERSIO_TEX, instr_handler},
1798 {BWRITERSIO_TEXLDP, instr_handler},
1799 {BWRITERSIO_TEXLDB, instr_handler},
1800 {BWRITERSIO_TEXKILL, instr_handler},
1801
1802 {BWRITERSIO_END, NULL},
1803 };
1804
1805 static const struct bytecode_backend ps_2_0_backend = {
1806 ps_2_header,
1807 end,
1808 ps_2_srcreg,
1809 ps_2_0_dstreg,
1810 sm_2_opcode,
1811 ps_2_0_handlers
1812 };
1813
1814 static const struct instr_handler_table ps_2_x_handlers[] = {
1815 {BWRITERSIO_ADD, instr_handler},
1816 {BWRITERSIO_NOP, instr_handler},
1817 {BWRITERSIO_MOV, instr_handler},
1818 {BWRITERSIO_SUB, instr_handler},
1819 {BWRITERSIO_MAD, instr_handler},
1820 {BWRITERSIO_MUL, instr_handler},
1821 {BWRITERSIO_RCP, instr_handler},
1822 {BWRITERSIO_RSQ, instr_handler},
1823 {BWRITERSIO_DP3, instr_handler},
1824 {BWRITERSIO_DP4, instr_handler},
1825 {BWRITERSIO_MIN, instr_handler},
1826 {BWRITERSIO_MAX, instr_handler},
1827 {BWRITERSIO_ABS, instr_handler},
1828 {BWRITERSIO_EXP, instr_handler},
1829 {BWRITERSIO_LOG, instr_handler},
1830 {BWRITERSIO_EXPP, instr_handler},
1831 {BWRITERSIO_LOGP, instr_handler},
1832 {BWRITERSIO_LRP, instr_handler},
1833 {BWRITERSIO_FRC, instr_handler},
1834 {BWRITERSIO_CRS, instr_handler},
1835 {BWRITERSIO_NRM, instr_handler},
1836 {BWRITERSIO_SINCOS, instr_handler},
1837 {BWRITERSIO_M4x4, instr_handler},
1838 {BWRITERSIO_M4x3, instr_handler},
1839 {BWRITERSIO_M3x4, instr_handler},
1840 {BWRITERSIO_M3x3, instr_handler},
1841 {BWRITERSIO_M3x2, instr_handler},
1842 {BWRITERSIO_POW, instr_handler},
1843 {BWRITERSIO_DP2ADD, instr_handler},
1844 {BWRITERSIO_CMP, instr_handler},
1845
1846 {BWRITERSIO_CALL, instr_handler},
1847 {BWRITERSIO_CALLNZ, instr_handler},
1848 {BWRITERSIO_REP, instr_handler},
1849 {BWRITERSIO_ENDREP, instr_handler},
1850 {BWRITERSIO_IF, instr_handler},
1851 {BWRITERSIO_LABEL, instr_handler},
1852 {BWRITERSIO_IFC, instr_handler},
1853 {BWRITERSIO_ELSE, instr_handler},
1854 {BWRITERSIO_ENDIF, instr_handler},
1855 {BWRITERSIO_BREAK, instr_handler},
1856 {BWRITERSIO_BREAKC, instr_handler},
1857 {BWRITERSIO_RET, instr_handler},
1858
1859 {BWRITERSIO_TEX, instr_handler},
1860 {BWRITERSIO_TEXLDP, instr_handler},
1861 {BWRITERSIO_TEXLDB, instr_handler},
1862 {BWRITERSIO_TEXKILL, instr_handler},
1863 {BWRITERSIO_DSX, instr_handler},
1864 {BWRITERSIO_DSY, instr_handler},
1865
1866 {BWRITERSIO_SETP, instr_handler},
1867 {BWRITERSIO_BREAKP, instr_handler},
1868
1869 {BWRITERSIO_TEXLDD, instr_handler},
1870
1871 {BWRITERSIO_END, NULL},
1872 };
1873
1874 static const struct bytecode_backend ps_2_x_backend = {
1875 ps_2_header,
1876 end,
1877 ps_2_srcreg,
1878 ps_2_0_dstreg,
1879 sm_2_opcode,
1880 ps_2_x_handlers
1881 };
1882
1883 static void sm_3_header(struct bc_writer *This, const struct bwriter_shader *shader, struct bytecode_buffer *buffer) {
1884 /* Declare the shader type and version */
1885 put_dword(buffer, This->version);
1886
1887 write_declarations(This, buffer, TRUE, shader->inputs, shader->num_inputs, BWRITERSPR_INPUT);
1888 write_declarations(This, buffer, TRUE, shader->outputs, shader->num_outputs, BWRITERSPR_OUTPUT);
1889 write_constF(shader, buffer, TRUE);
1890 write_constB(shader, buffer, TRUE);
1891 write_constI(shader, buffer, TRUE);
1892 write_samplers(shader, buffer);
1893 return;
1894 }
1895
1896 static void sm_3_srcreg(struct bc_writer *This,
1897 const struct shader_reg *reg,
1898 struct bytecode_buffer *buffer) {
1899 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1900 DWORD d3d9reg;
1901
1902 d3d9reg = d3d9_register(reg->type);
1903 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1904 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1905 token |= reg->regnum & D3DSP_REGNUM_MASK;
1906
1907 token |= d3d9_swizzle(reg->swizzle) & D3DVS_SWIZZLE_MASK;
1908 token |= d3d9_srcmod(reg->srcmod);
1909
1910 if(reg->rel_reg) {
1911 if(reg->type == BWRITERSPR_CONST && This->version == BWRITERPS_VERSION(3, 0)) {
1912 WARN("c%u[...] is unsupported in ps_3_0\n", reg->regnum);
1913 This->state = E_INVALIDARG;
1914 return;
1915 }
1916 if(((reg->rel_reg->type == BWRITERSPR_ADDR && This->version == BWRITERVS_VERSION(3, 0)) ||
1917 reg->rel_reg->type == BWRITERSPR_LOOP) &&
1918 reg->rel_reg->regnum == 0) {
1919 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
1920 } else {
1921 WARN("Unsupported relative addressing register\n");
1922 This->state = E_INVALIDARG;
1923 return;
1924 }
1925 }
1926
1927 put_dword(buffer, token);
1928
1929 /* vs_2_0 and newer write the register containing the index explicitly in the
1930 * binary code
1931 */
1932 if(token & D3DVS_ADDRMODE_RELATIVE) {
1933 sm_3_srcreg(This, reg->rel_reg, buffer);
1934 }
1935 }
1936
1937 static void sm_3_dstreg(struct bc_writer *This,
1938 const struct shader_reg *reg,
1939 struct bytecode_buffer *buffer,
1940 DWORD shift, DWORD mod) {
1941 DWORD token = (1 << 31); /* Bit 31 of registers is 1 */
1942 DWORD d3d9reg;
1943
1944 if(reg->rel_reg) {
1945 if(This->version == BWRITERVS_VERSION(3, 0) &&
1946 reg->type == BWRITERSPR_OUTPUT) {
1947 token |= D3DVS_ADDRMODE_RELATIVE & D3DVS_ADDRESSMODE_MASK;
1948 } else {
1949 WARN("Relative addressing not supported for this shader type or register type\n");
1950 This->state = E_INVALIDARG;
1951 return;
1952 }
1953 }
1954
1955 d3d9reg = d3d9_register(reg->type);
1956 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT) & D3DSP_REGTYPE_MASK;
1957 token |= (d3d9reg << D3DSP_REGTYPE_SHIFT2) & D3DSP_REGTYPE_MASK2;
1958 token |= reg->regnum & D3DSP_REGNUM_MASK; /* No shift */
1959
1960 token |= d3d9_dstmod(mod);
1961
1962 token |= d3d9_writemask(reg->writemask);
1963 put_dword(buffer, token);
1964
1965 /* vs_2_0 and newer write the register containing the index explicitly in the
1966 * binary code
1967 */
1968 if(token & D3DVS_ADDRMODE_RELATIVE) {
1969 sm_3_srcreg(This, reg->rel_reg, buffer);
1970 }
1971 }
1972
1973 static const struct instr_handler_table vs_3_handlers[] = {
1974 {BWRITERSIO_ADD, instr_handler},
1975 {BWRITERSIO_NOP, instr_handler},
1976 {BWRITERSIO_MOV, instr_handler},
1977 {BWRITERSIO_SUB, instr_handler},
1978 {BWRITERSIO_MAD, instr_handler},
1979 {BWRITERSIO_MUL, instr_handler},
1980 {BWRITERSIO_RCP, instr_handler},
1981 {BWRITERSIO_RSQ, instr_handler},
1982 {BWRITERSIO_DP3, instr_handler},
1983 {BWRITERSIO_DP4, instr_handler},
1984 {BWRITERSIO_MIN, instr_handler},
1985 {BWRITERSIO_MAX, instr_handler},
1986 {BWRITERSIO_SLT, instr_handler},
1987 {BWRITERSIO_SGE, instr_handler},
1988 {BWRITERSIO_ABS, instr_handler},
1989 {BWRITERSIO_EXP, instr_handler},
1990 {BWRITERSIO_LOG, instr_handler},
1991 {BWRITERSIO_EXPP, instr_handler},
1992 {BWRITERSIO_LOGP, instr_handler},
1993 {BWRITERSIO_DST, instr_handler},
1994 {BWRITERSIO_LRP, instr_handler},
1995 {BWRITERSIO_FRC, instr_handler},
1996 {BWRITERSIO_CRS, instr_handler},
1997 {BWRITERSIO_SGN, instr_handler},
1998 {BWRITERSIO_NRM, instr_handler},
1999 {BWRITERSIO_SINCOS, instr_handler},
2000 {BWRITERSIO_M4x4, instr_handler},
2001 {BWRITERSIO_M4x3, instr_handler},
2002 {BWRITERSIO_M3x4, instr_handler},
2003 {BWRITERSIO_M3x3, instr_handler},
2004 {BWRITERSIO_M3x2, instr_handler},
2005 {BWRITERSIO_LIT, instr_handler},
2006 {BWRITERSIO_POW, instr_handler},
2007 {BWRITERSIO_MOVA, instr_handler},
2008
2009 {BWRITERSIO_CALL, instr_handler},
2010 {BWRITERSIO_CALLNZ, instr_handler},
2011 {BWRITERSIO_REP, instr_handler},
2012 {BWRITERSIO_ENDREP, instr_handler},
2013 {BWRITERSIO_IF, instr_handler},
2014 {BWRITERSIO_LABEL, instr_handler},
2015 {BWRITERSIO_IFC, instr_handler},
2016 {BWRITERSIO_ELSE, instr_handler},
2017 {BWRITERSIO_ENDIF, instr_handler},
2018 {BWRITERSIO_BREAK, instr_handler},
2019 {BWRITERSIO_BREAKC, instr_handler},
2020 {BWRITERSIO_LOOP, instr_handler},
2021 {BWRITERSIO_RET, instr_handler},
2022 {BWRITERSIO_ENDLOOP, instr_handler},
2023
2024 {BWRITERSIO_SETP, instr_handler},
2025 {BWRITERSIO_BREAKP, instr_handler},
2026 {BWRITERSIO_TEXLDL, instr_handler},
2027
2028 {BWRITERSIO_END, NULL},
2029 };
2030
2031 static const struct bytecode_backend vs_3_backend = {
2032 sm_3_header,
2033 end,
2034 sm_3_srcreg,
2035 sm_3_dstreg,
2036 sm_2_opcode,
2037 vs_3_handlers
2038 };
2039
2040 static const struct instr_handler_table ps_3_handlers[] = {
2041 {BWRITERSIO_ADD, instr_handler},
2042 {BWRITERSIO_NOP, instr_handler},
2043 {BWRITERSIO_MOV, instr_handler},
2044 {BWRITERSIO_SUB, instr_handler},
2045 {BWRITERSIO_MAD, instr_handler},
2046 {BWRITERSIO_MUL, instr_handler},
2047 {BWRITERSIO_RCP, instr_handler},
2048 {BWRITERSIO_RSQ, instr_handler},
2049 {BWRITERSIO_DP3, instr_handler},
2050 {BWRITERSIO_DP4, instr_handler},
2051 {BWRITERSIO_MIN, instr_handler},
2052 {BWRITERSIO_MAX, instr_handler},
2053 {BWRITERSIO_ABS, instr_handler},
2054 {BWRITERSIO_EXP, instr_handler},
2055 {BWRITERSIO_LOG, instr_handler},
2056 {BWRITERSIO_EXPP, instr_handler},
2057 {BWRITERSIO_LOGP, instr_handler},
2058 {BWRITERSIO_LRP, instr_handler},
2059 {BWRITERSIO_FRC, instr_handler},
2060 {BWRITERSIO_CRS, instr_handler},
2061 {BWRITERSIO_NRM, instr_handler},
2062 {BWRITERSIO_SINCOS, instr_handler},
2063 {BWRITERSIO_M4x4, instr_handler},
2064 {BWRITERSIO_M4x3, instr_handler},
2065 {BWRITERSIO_M3x4, instr_handler},
2066 {BWRITERSIO_M3x3, instr_handler},
2067 {BWRITERSIO_M3x2, instr_handler},
2068 {BWRITERSIO_POW, instr_handler},
2069 {BWRITERSIO_DP2ADD, instr_handler},
2070 {BWRITERSIO_CMP, instr_handler},
2071
2072 {BWRITERSIO_CALL, instr_handler},
2073 {BWRITERSIO_CALLNZ, instr_handler},
2074 {BWRITERSIO_REP, instr_handler},
2075 {BWRITERSIO_ENDREP, instr_handler},
2076 {BWRITERSIO_IF, instr_handler},
2077 {BWRITERSIO_LABEL, instr_handler},
2078 {BWRITERSIO_IFC, instr_handler},
2079 {BWRITERSIO_ELSE, instr_handler},
2080 {BWRITERSIO_ENDIF, instr_handler},
2081 {BWRITERSIO_BREAK, instr_handler},
2082 {BWRITERSIO_BREAKC, instr_handler},
2083 {BWRITERSIO_LOOP, instr_handler},
2084 {BWRITERSIO_RET, instr_handler},
2085 {BWRITERSIO_ENDLOOP, instr_handler},
2086
2087 {BWRITERSIO_SETP, instr_handler},
2088 {BWRITERSIO_BREAKP, instr_handler},
2089 {BWRITERSIO_TEXLDL, instr_handler},
2090
2091 {BWRITERSIO_TEX, instr_handler},
2092 {BWRITERSIO_TEXLDP, instr_handler},
2093 {BWRITERSIO_TEXLDB, instr_handler},
2094 {BWRITERSIO_TEXKILL, instr_handler},
2095 {BWRITERSIO_DSX, instr_handler},
2096 {BWRITERSIO_DSY, instr_handler},
2097 {BWRITERSIO_TEXLDD, instr_handler},
2098
2099 {BWRITERSIO_END, NULL},
2100 };
2101
2102 static const struct bytecode_backend ps_3_backend = {
2103 sm_3_header,
2104 end,
2105 sm_3_srcreg,
2106 sm_3_dstreg,
2107 sm_2_opcode,
2108 ps_3_handlers
2109 };
2110
2111 static void init_vs10_dx9_writer(struct bc_writer *writer) {
2112 TRACE("Creating DirectX9 vertex shader 1.0 writer\n");
2113 writer->funcs = &vs_1_x_backend;
2114 }
2115
2116 static void init_vs11_dx9_writer(struct bc_writer *writer) {
2117 TRACE("Creating DirectX9 vertex shader 1.1 writer\n");
2118 writer->funcs = &vs_1_x_backend;
2119 }
2120
2121 static void init_vs20_dx9_writer(struct bc_writer *writer) {
2122 TRACE("Creating DirectX9 vertex shader 2.0 writer\n");
2123 writer->funcs = &vs_2_0_backend;
2124 }
2125
2126 static void init_vs2x_dx9_writer(struct bc_writer *writer) {
2127 TRACE("Creating DirectX9 vertex shader 2.x writer\n");
2128 writer->funcs = &vs_2_x_backend;
2129 }
2130
2131 static void init_vs30_dx9_writer(struct bc_writer *writer) {
2132 TRACE("Creating DirectX9 vertex shader 3.0 writer\n");
2133 writer->funcs = &vs_3_backend;
2134 }
2135
2136 static void init_ps10_dx9_writer(struct bc_writer *writer) {
2137 TRACE("Creating DirectX9 pixel shader 1.0 writer\n");
2138 writer->funcs = &ps_1_0123_backend;
2139 }
2140
2141 static void init_ps11_dx9_writer(struct bc_writer *writer) {
2142 TRACE("Creating DirectX9 pixel shader 1.1 writer\n");
2143 writer->funcs = &ps_1_0123_backend;
2144 }
2145
2146 static void init_ps12_dx9_writer(struct bc_writer *writer) {
2147 TRACE("Creating DirectX9 pixel shader 1.2 writer\n");
2148 writer->funcs = &ps_1_0123_backend;
2149 }
2150
2151 static void init_ps13_dx9_writer(struct bc_writer *writer) {
2152 TRACE("Creating DirectX9 pixel shader 1.3 writer\n");
2153 writer->funcs = &ps_1_0123_backend;
2154 }
2155
2156 static void init_ps14_dx9_writer(struct bc_writer *writer) {
2157 TRACE("Creating DirectX9 pixel shader 1.4 writer\n");
2158 writer->funcs = &ps_1_4_backend;
2159 }
2160
2161 static void init_ps20_dx9_writer(struct bc_writer *writer) {
2162 TRACE("Creating DirectX9 pixel shader 2.0 writer\n");
2163 writer->funcs = &ps_2_0_backend;
2164 }
2165
2166 static void init_ps2x_dx9_writer(struct bc_writer *writer) {
2167 TRACE("Creating DirectX9 pixel shader 2.x writer\n");
2168 writer->funcs = &ps_2_x_backend;
2169 }
2170
2171 static void init_ps30_dx9_writer(struct bc_writer *writer) {
2172 TRACE("Creating DirectX9 pixel shader 3.0 writer\n");
2173 writer->funcs = &ps_3_backend;
2174 }
2175
2176 static struct bc_writer *create_writer(DWORD version, DWORD dxversion) {
2177 struct bc_writer *ret = asm_alloc(sizeof(*ret));
2178
2179 if(!ret) {
2180 WARN("Failed to allocate a bytecode writer instance\n");
2181 return NULL;
2182 }
2183
2184 switch(version) {
2185 case BWRITERVS_VERSION(1, 0):
2186 if(dxversion != 9) {
2187 WARN("Unsupported dxversion for vertex shader 1.0 requested: %u\n", dxversion);
2188 goto fail;
2189 }
2190 init_vs10_dx9_writer(ret);
2191 break;
2192 case BWRITERVS_VERSION(1, 1):
2193 if(dxversion != 9) {
2194 WARN("Unsupported dxversion for vertex shader 1.1 requested: %u\n", dxversion);
2195 goto fail;
2196 }
2197 init_vs11_dx9_writer(ret);
2198 break;
2199 case BWRITERVS_VERSION(2, 0):
2200 if(dxversion != 9) {
2201 WARN("Unsupported dxversion for vertex shader 2.0 requested: %u\n", dxversion);
2202 goto fail;
2203 }
2204 init_vs20_dx9_writer(ret);
2205 break;
2206 case BWRITERVS_VERSION(2, 1):
2207 if(dxversion != 9) {
2208 WARN("Unsupported dxversion for vertex shader 2.x requested: %u\n", dxversion);
2209 goto fail;
2210 }
2211 init_vs2x_dx9_writer(ret);
2212 break;
2213 case BWRITERVS_VERSION(3, 0):
2214 if(dxversion != 9) {
2215 WARN("Unsupported dxversion for vertex shader 3.0 requested: %u\n", dxversion);
2216 goto fail;
2217 }
2218 init_vs30_dx9_writer(ret);
2219 break;
2220
2221 case BWRITERPS_VERSION(1, 0):
2222 if(dxversion != 9) {
2223 WARN("Unsupported dxversion for pixel shader 1.0 requested: %u\n", dxversion);
2224 goto fail;
2225 }
2226 init_ps10_dx9_writer(ret);
2227 break;
2228 case BWRITERPS_VERSION(1, 1):
2229 if(dxversion != 9) {
2230 WARN("Unsupported dxversion for pixel shader 1.1 requested: %u\n", dxversion);
2231 goto fail;
2232 }
2233 init_ps11_dx9_writer(ret);
2234 break;
2235 case BWRITERPS_VERSION(1, 2):
2236 if(dxversion != 9) {
2237 WARN("Unsupported dxversion for pixel shader 1.2 requested: %u\n", dxversion);
2238 goto fail;
2239 }
2240 init_ps12_dx9_writer(ret);
2241 break;
2242 case BWRITERPS_VERSION(1, 3):
2243 if(dxversion != 9) {
2244 WARN("Unsupported dxversion for pixel shader 1.3 requested: %u\n", dxversion);
2245 goto fail;
2246 }
2247 init_ps13_dx9_writer(ret);
2248 break;
2249 case BWRITERPS_VERSION(1, 4):
2250 if(dxversion != 9) {
2251 WARN("Unsupported dxversion for pixel shader 1.4 requested: %u\n", dxversion);
2252 goto fail;
2253 }
2254 init_ps14_dx9_writer(ret);
2255 break;
2256
2257 case BWRITERPS_VERSION(2, 0):
2258 if(dxversion != 9) {
2259 WARN("Unsupported dxversion for pixel shader 2.0 requested: %u\n", dxversion);
2260 goto fail;
2261 }
2262 init_ps20_dx9_writer(ret);
2263 break;
2264
2265 case BWRITERPS_VERSION(2, 1):
2266 if(dxversion != 9) {
2267 WARN("Unsupported dxversion for pixel shader 2.x requested: %u\n", dxversion);
2268 goto fail;
2269 }
2270 init_ps2x_dx9_writer(ret);
2271 break;
2272
2273 case BWRITERPS_VERSION(3, 0):
2274 if(dxversion != 9) {
2275 WARN("Unsupported dxversion for pixel shader 3.0 requested: %u\n", dxversion);
2276 goto fail;
2277 }
2278 init_ps30_dx9_writer(ret);
2279 break;
2280
2281 default:
2282 WARN("Unexpected shader version requested: %08x\n", version);
2283 goto fail;
2284 }
2285 ret->version = version;
2286 return ret;
2287
2288 fail:
2289 asm_free(ret);
2290 return NULL;
2291 }
2292
2293 static HRESULT call_instr_handler(struct bc_writer *writer,
2294 const struct instruction *instr,
2295 struct bytecode_buffer *buffer) {
2296 DWORD i=0;
2297
2298 while(writer->funcs->instructions[i].opcode != BWRITERSIO_END) {
2299 if(instr->opcode == writer->funcs->instructions[i].opcode) {
2300 if(!writer->funcs->instructions[i].func) {
2301 WARN("Opcode %u not supported by this profile\n", instr->opcode);
2302 return E_INVALIDARG;
2303 }
2304 writer->funcs->instructions[i].func(writer, instr, buffer);
2305 return S_OK;
2306 }
2307 i++;
2308 }
2309
2310 FIXME("Unhandled instruction %u - %s\n", instr->opcode,
2311 debug_print_opcode(instr->opcode));
2312 return E_INVALIDARG;
2313 }
2314
2315 /* SlWriteBytecode (wineshader.@)
2316 *
2317 * Writes shader version specific bytecode from the shader passed in.
2318 * The returned bytecode can be passed to the Direct3D runtime like
2319 * IDirect3DDevice9::Create*Shader.
2320 *
2321 * Parameters:
2322 * shader: Shader to translate into bytecode
2323 * version: Shader version to generate(d3d version token)
2324 * dxversion: DirectX version the code targets
2325 * result: the resulting shader bytecode
2326 *
2327 * Return values:
2328 * S_OK on success
2329 */
2330 DWORD SlWriteBytecode(const struct bwriter_shader *shader, int dxversion, DWORD **result) {
2331 struct bc_writer *writer;
2332 struct bytecode_buffer *buffer = NULL;
2333 HRESULT hr;
2334 unsigned int i;
2335
2336 if(!shader){
2337 ERR("NULL shader structure, aborting\n");
2338 return E_FAIL;
2339 }
2340 writer = create_writer(shader->version, dxversion);
2341 *result = NULL;
2342
2343 if(!writer) {
2344 WARN("Could not create a bytecode writer instance. Either unsupported version\n");
2345 WARN("or out of memory\n");
2346 hr = E_FAIL;
2347 goto error;
2348 }
2349
2350 buffer = allocate_buffer();
2351 if(!buffer) {
2352 WARN("Failed to allocate a buffer for the shader bytecode\n");
2353 hr = E_FAIL;
2354 goto error;
2355 }
2356
2357 writer->funcs->header(writer, shader, buffer);
2358 if(FAILED(writer->state)) {
2359 hr = writer->state;
2360 goto error;
2361 }
2362
2363 for(i = 0; i < shader->num_instrs; i++) {
2364 hr = call_instr_handler(writer, shader->instr[i], buffer);
2365 if(FAILED(hr)) {
2366 goto error;
2367 }
2368 }
2369
2370 if(FAILED(writer->state)) {
2371 hr = writer->state;
2372 goto error;
2373 }
2374
2375 writer->funcs->end(writer, shader, buffer);
2376
2377 if(FAILED(buffer->state)) {
2378 hr = buffer->state;
2379 goto error;
2380 }
2381
2382 /* Cut off unneeded memory from the result buffer */
2383 *result = asm_realloc(buffer->data,
2384 sizeof(DWORD) * buffer->size);
2385 if(!*result) {
2386 *result = buffer->data;
2387 }
2388 buffer->data = NULL;
2389 hr = S_OK;
2390
2391 error:
2392 if(buffer) {
2393 asm_free(buffer->data);
2394 asm_free(buffer);
2395 }
2396 asm_free(writer);
2397 return hr;
2398 }
2399
2400 void SlDeleteShader(struct bwriter_shader *shader) {
2401 unsigned int i, j;
2402
2403 TRACE("Deleting shader %p\n", shader);
2404
2405 for(i = 0; i < shader->num_cf; i++) {
2406 asm_free(shader->constF[i]);
2407 }
2408 asm_free(shader->constF);
2409 for(i = 0; i < shader->num_ci; i++) {
2410 asm_free(shader->constI[i]);
2411 }
2412 asm_free(shader->constI);
2413 for(i = 0; i < shader->num_cb; i++) {
2414 asm_free(shader->constB[i]);
2415 }
2416 asm_free(shader->constB);
2417
2418 asm_free(shader->inputs);
2419 asm_free(shader->outputs);
2420 asm_free(shader->samplers);
2421
2422 for(i = 0; i < shader->num_instrs; i++) {
2423 for(j = 0; j < shader->instr[i]->num_srcs; j++) {
2424 asm_free(shader->instr[i]->src[j].rel_reg);
2425 }
2426 asm_free(shader->instr[i]->src);
2427 asm_free(shader->instr[i]);
2428 }
2429 asm_free(shader->instr);
2430
2431 asm_free(shader);
2432 }
an attempt at megring old patches into wine