mshtml: Added IHTMLStyleSheet::get_rules implementation.
[wine] / dlls / wined3d / baseshader.c
1 /*
2  * shaders implementation
3  *
4  * Copyright 2002-2003 Jason Edmeades
5  * Copyright 2002-2003 Raphael Junqueira
6  * Copyright 2004 Christian Costa
7  * Copyright 2005 Oliver Stieber
8  * Copyright 2006 Ivan Gyurdiev
9  *
10  * This library is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * This library is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with this library; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23  */
24
25 #include "config.h"
26 #include <string.h>
27 #include <stdio.h>
28 #include "wined3d_private.h"
29
30 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
31
32 #define GLNAME_REQUIRE_GLSL  ((const char *)1)
33
34 static inline BOOL shader_is_version_token(DWORD token) {
35     return shader_is_pshader_version(token) ||
36            shader_is_vshader_version(token);
37 }
38
39 int shader_addline(
40     SHADER_BUFFER* buffer,  
41     const char *format, ...) {
42
43     char* base = buffer->buffer + buffer->bsize;
44     int rc;
45
46     va_list args;
47     va_start(args, format);
48     rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
49     va_end(args);
50
51     if (rc < 0 ||                                   /* C89 */ 
52         rc > SHADER_PGMSIZE - 1 - buffer->bsize) {  /* C99 */
53
54         ERR("The buffer allocated for the shader program string "
55             "is too small at %d bytes.\n", SHADER_PGMSIZE);
56         buffer->bsize = SHADER_PGMSIZE - 1;
57         return -1;
58     }
59
60     if (buffer->newline) {
61         TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
62         buffer->newline = FALSE;
63     } else {
64         TRACE("%s", base);
65     }
66
67     buffer->bsize += rc;
68     if (buffer->buffer[buffer->bsize-1] == '\n') {
69         buffer->lineNo++;
70         buffer->newline = TRUE;
71     }
72     return 0;
73 }
74
75 const SHADER_OPCODE* shader_get_opcode(
76     IWineD3DBaseShader *iface, const DWORD code) {
77
78     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl*) iface;
79
80     DWORD i = 0;
81     DWORD hex_version = This->baseShader.hex_version;
82     const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
83
84     /** TODO: use dichotomic search */
85     while (NULL != shader_ins[i].name) {
86         if (((code & WINED3DSI_OPCODE_MASK) == shader_ins[i].opcode) &&
87             (((hex_version >= shader_ins[i].min_version) && (hex_version <= shader_ins[i].max_version)) ||
88             ((shader_ins[i].min_version == 0) && (shader_ins[i].max_version == 0)))) {
89             return &shader_ins[i];
90         }
91         ++i;
92     }
93     FIXME("Unsupported opcode %#x(%d) masked %#x, shader version %#x\n", 
94        code, code, code & WINED3DSI_OPCODE_MASK, hex_version);
95     return NULL;
96 }
97
98 /* Read a parameter opcode from the input stream,
99  * and possibly a relative addressing token.
100  * Return the number of tokens read */
101 int shader_get_param(
102     IWineD3DBaseShader* iface,
103     const DWORD* pToken,
104     DWORD* param,
105     DWORD* addr_token) {
106
107     /* PS >= 3.0 have relative addressing (with token)
108      * VS >= 2.0 have relative addressing (with token)
109      * VS >= 1.0 < 2.0 have relative addressing (without token)
110      * The version check below should work in general */
111
112     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
113     char rel_token = WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
114         ((*pToken & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
115
116     *param = *pToken;
117     *addr_token = rel_token? *(pToken + 1): 0;
118     return rel_token? 2:1;
119 }
120
121 /* Return the number of parameters to skip for an opcode */
122 static inline int shader_skip_opcode(
123     IWineD3DBaseShaderImpl* This,
124     const SHADER_OPCODE* curOpcode,
125     DWORD opcode_token) {
126
127    /* Shaders >= 2.0 may contain address tokens, but fortunately they
128     * have a useful legnth mask - use it here. Shaders 1.0 contain no such tokens */
129
130     return (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
131         ((opcode_token & WINED3DSI_INSTLENGTH_MASK) >> WINED3DSI_INSTLENGTH_SHIFT):
132         curOpcode->num_params;
133 }
134
135 /* Read the parameters of an unrecognized opcode from the input stream
136  * Return the number of tokens read. 
137  * 
138  * Note: This function assumes source or destination token format.
139  * It will not work with specially-formatted tokens like DEF or DCL, 
140  * but hopefully those would be recognized */
141
142 int shader_skip_unrecognized(
143     IWineD3DBaseShader* iface,
144     const DWORD* pToken) {
145
146     int tokens_read = 0;
147     int i = 0;
148
149     /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
150     while (*pToken & 0x80000000) {
151
152         DWORD param, addr_token;
153         tokens_read += shader_get_param(iface, pToken, &param, &addr_token);
154         pToken += tokens_read;
155
156         FIXME("Unrecognized opcode param: token=%08x "
157             "addr_token=%08x name=", param, addr_token);
158         shader_dump_param(iface, param, addr_token, i);
159         FIXME("\n");
160         ++i;
161     }
162     return tokens_read;
163 }
164
165 /* Convert floating point offset relative
166  * to a register file to an absolute offset for float constants */
167
168 unsigned int shader_get_float_offset(const DWORD reg) {
169
170      unsigned int regnum = reg & WINED3DSP_REGNUM_MASK;
171      int regtype = shader_get_regtype(reg);
172
173      switch (regtype) {
174         case WINED3DSPR_CONST: return regnum;
175         case WINED3DSPR_CONST2: return 2048 + regnum;
176         case WINED3DSPR_CONST3: return 4096 + regnum;
177         case WINED3DSPR_CONST4: return 6144 + regnum;
178         default:
179             FIXME("Unsupported register type: %d\n", regtype);
180             return regnum;
181      }
182 }
183
184 /* Note that this does not count the loop register
185  * as an address register. */
186
187 HRESULT shader_get_registers_used(
188     IWineD3DBaseShader *iface,
189     shader_reg_maps* reg_maps,
190     semantic* semantics_in,
191     semantic* semantics_out,
192     CONST DWORD* pToken,
193     IWineD3DStateBlockImpl *stateBlock) {
194
195     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
196     unsigned int cur_loop_depth = 0, max_loop_depth = 0;
197
198     /* There are some minor differences between pixel and vertex shaders */
199     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
200
201     reg_maps->bumpmat = -1;
202     reg_maps->luminanceparams = -1;
203
204     if (pToken == NULL)
205         return WINED3D_OK;
206
207     while (WINED3DVS_END() != *pToken) {
208         CONST SHADER_OPCODE* curOpcode;
209         DWORD opcode_token;
210
211         /* Skip version */
212         if (shader_is_version_token(*pToken)) {
213              ++pToken;
214              continue;
215
216         /* Skip comments */
217         } else if (shader_is_comment(*pToken)) {
218              DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
219              ++pToken;
220              pToken += comment_len;
221              continue;
222         }
223
224         /* Fetch opcode */
225         opcode_token = *pToken++;
226         curOpcode = shader_get_opcode(iface, opcode_token);
227
228         /* Unhandled opcode, and its parameters */
229         if (NULL == curOpcode) {
230            while (*pToken & 0x80000000)
231                ++pToken;
232
233         /* Handle declarations */
234         } else if (WINED3DSIO_DCL == curOpcode->opcode) {
235
236             DWORD usage = *pToken++;
237             DWORD param = *pToken++;
238             DWORD regtype = shader_get_regtype(param);
239             unsigned int regnum = param & WINED3DSP_REGNUM_MASK;
240
241             /* Vshader: mark attributes used
242                Pshader: mark 3.0 input registers used, save token */
243             if (WINED3DSPR_INPUT == regtype) {
244
245                 if (!pshader)
246                     reg_maps->attributes[regnum] = 1;
247                 else
248                     reg_maps->packed_input[regnum] = 1;
249
250                 semantics_in[regnum].usage = usage;
251                 semantics_in[regnum].reg = param;
252
253             /* Vshader: mark 3.0 output registers used, save token */
254             } else if (WINED3DSPR_OUTPUT == regtype) {
255                 reg_maps->packed_output[regnum] = 1;
256                 semantics_out[regnum].usage = usage;
257                 semantics_out[regnum].reg = param;
258                 if (usage & (WINED3DDECLUSAGE_FOG << WINED3DSP_DCL_USAGE_SHIFT))
259                     reg_maps->fog = 1;
260
261             /* Save sampler usage token */
262             } else if (WINED3DSPR_SAMPLER == regtype)
263                 reg_maps->samplers[regnum] = usage;
264
265         } else if (WINED3DSIO_DEF == curOpcode->opcode) {
266
267             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
268             if (!lconst) return E_OUTOFMEMORY;
269             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
270             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
271
272             /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
273             if(WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 && pshader) {
274                 float *value = (float *) lconst->value;
275                 if(value[0] < -1.0) value[0] = -1.0;
276                 else if(value[0] >  1.0) value[0] =  1.0;
277                 if(value[1] < -1.0) value[1] = -1.0;
278                 else if(value[1] >  1.0) value[1] =  1.0;
279                 if(value[2] < -1.0) value[2] = -1.0;
280                 else if(value[2] >  1.0) value[2] =  1.0;
281                 if(value[3] < -1.0) value[3] = -1.0;
282                 else if(value[3] >  1.0) value[3] =  1.0;
283             }
284
285             list_add_head(&This->baseShader.constantsF, &lconst->entry);
286             pToken += curOpcode->num_params;
287
288         } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
289
290             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
291             if (!lconst) return E_OUTOFMEMORY;
292             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
293             memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
294             list_add_head(&This->baseShader.constantsI, &lconst->entry);
295             pToken += curOpcode->num_params;
296
297         } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
298
299             local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
300             if (!lconst) return E_OUTOFMEMORY;
301             lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
302             memcpy(&lconst->value, pToken + 1, 1 * sizeof(DWORD));
303             list_add_head(&This->baseShader.constantsB, &lconst->entry);
304             pToken += curOpcode->num_params;
305
306         /* If there's a loop in the shader */
307         } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
308                    WINED3DSIO_REP == curOpcode->opcode) {
309             cur_loop_depth++;
310             if(cur_loop_depth > max_loop_depth)
311                 max_loop_depth = cur_loop_depth;
312             pToken += curOpcode->num_params;
313
314         } else if (WINED3DSIO_ENDLOOP == curOpcode->opcode ||
315                    WINED3DSIO_ENDREP == curOpcode->opcode) {
316             cur_loop_depth--;
317
318         /* For subroutine prototypes */
319         } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
320
321             DWORD snum = *pToken & WINED3DSP_REGNUM_MASK; 
322             reg_maps->labels[snum] = 1;
323             pToken += curOpcode->num_params;
324
325         /* Set texture, address, temporary registers */
326         } else {
327             int i, limit;
328
329             /* Declare 1.X samplers implicitly, based on the destination reg. number */
330             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 && 
331                 (WINED3DSIO_TEX == curOpcode->opcode ||
332                  WINED3DSIO_TEXBEM == curOpcode->opcode ||
333                  WINED3DSIO_TEXBEML == curOpcode->opcode ||
334                  WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
335                  WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
336                  WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
337                  WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
338                  WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
339                  WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
340                  WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
341                  WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
342
343                 /* Fake sampler usage, only set reserved bit and ttype */
344                 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
345
346                 if(!stateBlock->textures[sampler_code]) {
347                     ERR("No texture bound to sampler %d\n", sampler_code);
348                     reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
349                 } else {
350                     int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
351                     switch(texType) {
352                         /* We have to select between texture rectangles and 2D textures later because 2.0 and
353                          * 3.0 shaders only have WINED3DSTT_2D as well
354                          */
355                         case GL_TEXTURE_RECTANGLE_ARB:
356                         case GL_TEXTURE_2D:
357                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
358                             break;
359
360                         case GL_TEXTURE_3D:
361                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
362                             break;
363
364                         case GL_TEXTURE_CUBE_MAP_ARB:
365                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
366                             break;
367
368                         default:
369                             ERR("Unexpected gl texture type found: %d\n", texType);
370                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
371                     }
372                 }
373
374                 /* texbem is only valid with < 1.4 pixel shaders */
375                 if(WINED3DSIO_TEXBEM  == curOpcode->opcode ||
376                     WINED3DSIO_TEXBEML == curOpcode->opcode) {
377                     if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != sampler_code) {
378                         FIXME("Pixel shader uses texbem instruction on more than 1 sampler\n");
379                     } else {
380                         reg_maps->bumpmat = sampler_code;
381                         if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
382                             reg_maps->luminanceparams = sampler_code;
383                         }
384                     }
385                 }
386             }
387             if(WINED3DSIO_NRM  == curOpcode->opcode) {
388                 reg_maps->usesnrm = 1;
389             } else if(WINED3DSIO_BEM == curOpcode->opcode) {
390                 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
391                 if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != regnum) {
392                     FIXME("Pixel shader uses bem or texbem instruction on more than 1 sampler\n");
393                 } else {
394                     reg_maps->bumpmat = regnum;
395                 }
396             } else if(WINED3DSIO_DSY  == curOpcode->opcode) {
397                 reg_maps->usesdsy = 1;
398             }
399
400             /* This will loop over all the registers and try to
401              * make a bitmask of the ones we're interested in. 
402              *
403              * Relative addressing tokens are ignored, but that's 
404              * okay, since we'll catch any address registers when 
405              * they are initialized (required by spec) */
406
407             limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
408                 curOpcode->num_params + 1: curOpcode->num_params;
409
410             for (i = 0; i < limit; ++i) {
411
412                 DWORD param, addr_token, reg, regtype;
413                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
414
415                 regtype = shader_get_regtype(param);
416                 reg = param & WINED3DSP_REGNUM_MASK;
417
418                 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
419
420                     if (pshader)
421                         reg_maps->texcoord[reg] = 1;
422                     else
423                         reg_maps->address[reg] = 1;
424                 }
425
426                 else if (WINED3DSPR_TEMP == regtype)
427                     reg_maps->temporary[reg] = 1;
428
429                 else if (WINED3DSPR_INPUT == regtype) {
430                     if( !pshader)
431                         reg_maps->attributes[reg] = 1;
432                     else {
433                         if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
434                             /* If relative addressing is used, we must assume that all registers
435                              * are used. Even if it is a construct like v3[aL], we can't assume
436                              * that v0, v1 and v2 aren't read because aL can be negative
437                              */
438                             unsigned int i;
439                             for(i = 0; i < MAX_REG_INPUT; i++) {
440                                 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[i] = TRUE;
441                             }
442                         } else {
443                             ((IWineD3DPixelShaderImpl *) This)->input_reg_used[reg] = TRUE;
444                         }
445                     }
446                 }
447
448                 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
449                     reg_maps->fog = 1;
450
451                 else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
452                     reg_maps->vpos = 1;
453
454                 else if(WINED3DSPR_CONST == regtype) {
455                     if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
456                         if(!pshader) {
457                             if(reg <= ((IWineD3DVertexShaderImpl *) This)->min_rel_offset) {
458                                 ((IWineD3DVertexShaderImpl *) This)->min_rel_offset = reg;
459                             } else if(reg >= ((IWineD3DVertexShaderImpl *) This)->max_rel_offset) {
460                                 ((IWineD3DVertexShaderImpl *) This)->max_rel_offset = reg;
461                             }
462                         }
463                         reg_maps->usesrelconstF = TRUE;
464                     }
465                 }
466             }
467         }
468     }
469     reg_maps->loop_depth = max_loop_depth;
470
471     return WINED3D_OK;
472 }
473
474 static void shader_dump_decl_usage(
475     IWineD3DBaseShaderImpl* This,
476     DWORD decl, 
477     DWORD param) {
478
479     DWORD regtype = shader_get_regtype(param);
480
481     TRACE("dcl");
482
483     if (regtype == WINED3DSPR_SAMPLER) {
484         DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
485
486         switch (ttype) {
487             case WINED3DSTT_2D: TRACE("_2d"); break;
488             case WINED3DSTT_CUBE: TRACE("_cube"); break;
489             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
490             default: TRACE("_unknown_ttype(%08x)", ttype); 
491        }
492
493     } else { 
494
495         DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
496         DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
497
498         /* Pixel shaders 3.0 don't have usage semantics */
499         char pshader = shader_is_pshader_version(This->baseShader.hex_version);
500         if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
501             return;
502         else
503             TRACE("_");
504
505         switch(usage) {
506         case WINED3DDECLUSAGE_POSITION:
507             TRACE("position%d", idx);
508             break;
509         case WINED3DDECLUSAGE_BLENDINDICES:
510             TRACE("blend");
511             break;
512         case WINED3DDECLUSAGE_BLENDWEIGHT:
513             TRACE("weight");
514             break;
515         case WINED3DDECLUSAGE_NORMAL:
516             TRACE("normal%d", idx);
517             break;
518         case WINED3DDECLUSAGE_PSIZE:
519             TRACE("psize");
520             break;
521         case WINED3DDECLUSAGE_COLOR:
522             if(idx == 0)  {
523                 TRACE("color");
524             } else {
525                 TRACE("specular%d", (idx - 1));
526             }
527             break;
528         case WINED3DDECLUSAGE_TEXCOORD:
529             TRACE("texture%d", idx);
530             break;
531         case WINED3DDECLUSAGE_TANGENT:
532             TRACE("tangent");
533             break;
534         case WINED3DDECLUSAGE_BINORMAL:
535             TRACE("binormal");
536             break;
537         case WINED3DDECLUSAGE_TESSFACTOR:
538             TRACE("tessfactor");
539             break;
540         case WINED3DDECLUSAGE_POSITIONT:
541             TRACE("positionT%d", idx);
542             break;
543         case WINED3DDECLUSAGE_FOG:
544             TRACE("fog");
545             break;
546         case WINED3DDECLUSAGE_DEPTH:
547             TRACE("depth");
548             break;
549         case WINED3DDECLUSAGE_SAMPLE:
550             TRACE("sample");
551             break;
552         default:
553             FIXME("unknown_semantics(%08x)", usage);
554         }
555     }
556 }
557
558 static void shader_dump_arr_entry(
559     IWineD3DBaseShader *iface,
560     const DWORD param,
561     const DWORD addr_token,
562     unsigned int reg,
563     int input) {
564
565     char relative =
566         ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
567
568     if (relative) {
569         TRACE("[");
570         if (addr_token)
571             shader_dump_param(iface, addr_token, 0, input);
572         else
573             TRACE("a0.x");
574         TRACE(" + ");
575      }
576      TRACE("%u", reg);
577      if (relative)
578          TRACE("]");
579 }
580
581 void shader_dump_param(
582     IWineD3DBaseShader *iface,
583     const DWORD param, 
584     const DWORD addr_token,
585     int input) {
586
587     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
588     static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
589     static const char * const misctype_reg_names[] = { "vPos", "vFace"};
590     char swizzle_reg_chars[4];
591
592     DWORD reg = param & WINED3DSP_REGNUM_MASK;
593     DWORD regtype = shader_get_regtype(param);
594     DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
595
596     /* There are some minor differences between pixel and vertex shaders */
597     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
598
599     /* For one, we'd prefer color components to be shown for pshaders.
600      * FIXME: use the swizzle function for this */
601
602     swizzle_reg_chars[0] = pshader? 'r': 'x';
603     swizzle_reg_chars[1] = pshader? 'g': 'y';
604     swizzle_reg_chars[2] = pshader? 'b': 'z';
605     swizzle_reg_chars[3] = pshader? 'a': 'w';
606
607     if (input) {
608         if ( (modifier == WINED3DSPSM_NEG) ||
609              (modifier == WINED3DSPSM_BIASNEG) ||
610              (modifier == WINED3DSPSM_SIGNNEG) ||
611              (modifier == WINED3DSPSM_X2NEG) ||
612              (modifier == WINED3DSPSM_ABSNEG) )
613             TRACE("-");
614         else if (modifier == WINED3DSPSM_COMP)
615             TRACE("1-");
616         else if (modifier == WINED3DSPSM_NOT)
617             TRACE("!");
618
619         if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG) 
620             TRACE("abs(");
621     }
622
623     switch (regtype) {
624         case WINED3DSPR_TEMP:
625             TRACE("r%u", reg);
626             break;
627         case WINED3DSPR_INPUT:
628             TRACE("v");
629             shader_dump_arr_entry(iface, param, addr_token, reg, input);
630             break;
631         case WINED3DSPR_CONST:
632         case WINED3DSPR_CONST2:
633         case WINED3DSPR_CONST3:
634         case WINED3DSPR_CONST4:
635             TRACE("c");
636             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
637             break;
638         case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
639             TRACE("%c%u", (pshader? 't':'a'), reg);
640             break;        
641         case WINED3DSPR_RASTOUT:
642             TRACE("%s", rastout_reg_names[reg]);
643             break;
644         case WINED3DSPR_COLOROUT:
645             TRACE("oC%u", reg);
646             break;
647         case WINED3DSPR_DEPTHOUT:
648             TRACE("oDepth");
649             break;
650         case WINED3DSPR_ATTROUT:
651             TRACE("oD%u", reg);
652             break;
653         case WINED3DSPR_TEXCRDOUT: 
654
655             /* Vertex shaders >= 3.0 use general purpose output registers
656              * (WINED3DSPR_OUTPUT), which can include an address token */
657
658             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
659                 TRACE("o");
660                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
661             }
662             else 
663                TRACE("oT%u", reg);
664             break;
665         case WINED3DSPR_CONSTINT:
666             TRACE("i");
667             shader_dump_arr_entry(iface, param, addr_token, reg, input);
668             break;
669         case WINED3DSPR_CONSTBOOL:
670             TRACE("b");
671             shader_dump_arr_entry(iface, param, addr_token, reg, input);
672             break;
673         case WINED3DSPR_LABEL:
674             TRACE("l%u", reg);
675             break;
676         case WINED3DSPR_LOOP:
677             TRACE("aL");
678             break;
679         case WINED3DSPR_SAMPLER:
680             TRACE("s%u", reg);
681             break;
682         case WINED3DSPR_MISCTYPE:
683             if (reg > 1) {
684                 FIXME("Unhandled misctype register %d\n", reg);
685             } else {
686                 TRACE("%s", misctype_reg_names[reg]);
687             }
688             break;
689         case WINED3DSPR_PREDICATE:
690             TRACE("p%u", reg);
691             break;
692         default:
693             TRACE("unhandled_rtype(%#x)", regtype);
694             break;
695    }
696
697    if (!input) {
698        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
699
700        if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
701            TRACE(".");
702            if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
703            if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
704            if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
705            if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
706        }
707
708    } else {
709         /** operand input */
710         DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
711         DWORD swizzle_r = swizzle & 0x03;
712         DWORD swizzle_g = (swizzle >> 2) & 0x03;
713         DWORD swizzle_b = (swizzle >> 4) & 0x03;
714         DWORD swizzle_a = (swizzle >> 6) & 0x03;
715
716         if (0 != modifier) {
717             switch (modifier) {
718                 case WINED3DSPSM_NONE:    break;
719                 case WINED3DSPSM_NEG:     break;
720                 case WINED3DSPSM_NOT:     break;
721                 case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
722                 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
723                 case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
724                 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
725                 case WINED3DSPSM_COMP:    break;
726                 case WINED3DSPSM_X2:      TRACE("_x2"); break;
727                 case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
728                 case WINED3DSPSM_DZ:      TRACE("_dz"); break;
729                 case WINED3DSPSM_DW:      TRACE("_dw"); break;
730                 case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
731                 case WINED3DSPSM_ABS:     TRACE(")"); break;
732                 default:
733                     TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
734             }
735         }
736
737         /**
738         * swizzle bits fields:
739         *  RRGGBBAA
740         */
741         if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
742             if (swizzle_r == swizzle_g &&
743                 swizzle_r == swizzle_b &&
744                 swizzle_r == swizzle_a) {
745                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
746             } else {
747                 TRACE(".%c%c%c%c",
748                 swizzle_reg_chars[swizzle_r],
749                 swizzle_reg_chars[swizzle_g],
750                 swizzle_reg_chars[swizzle_b],
751                 swizzle_reg_chars[swizzle_a]);
752             }
753         }
754     }
755 }
756
757 /** Shared code in order to generate the bulk of the shader string.
758     Use the shader_header_fct & shader_footer_fct to add strings
759     that are specific to pixel or vertex functions
760     NOTE: A description of how to parse tokens can be found at:
761           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
762 void shader_generate_main(
763     IWineD3DBaseShader *iface,
764     SHADER_BUFFER* buffer,
765     shader_reg_maps* reg_maps,
766     CONST DWORD* pFunction) {
767
768     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
769     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
770     const DWORD *pToken = pFunction;
771     const SHADER_OPCODE *curOpcode = NULL;
772     SHADER_HANDLER hw_fct = NULL;
773     DWORD i;
774     SHADER_OPCODE_ARG hw_arg;
775
776     /* Initialize current parsing state */
777     hw_arg.shader = iface;
778     hw_arg.buffer = buffer;
779     hw_arg.reg_maps = reg_maps;
780     This->baseShader.parse_state.current_row = 0;
781
782     /* Second pass, process opcodes */
783     if (NULL != pToken) {
784         while (WINED3DPS_END() != *pToken) {
785
786             /* Skip version token */
787             if (shader_is_version_token(*pToken)) {
788                 ++pToken;
789                 continue;
790             }
791
792             /* Skip comment tokens */
793             if (shader_is_comment(*pToken)) {
794                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
795                 ++pToken;
796                 TRACE("#%s\n", (const char*)pToken);
797                 pToken += comment_len;
798                 continue;
799             }
800
801             /* Read opcode */
802             hw_arg.opcode_token = *pToken++;
803             curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
804
805             /* Select handler */
806             if (curOpcode == NULL)
807                 hw_fct = NULL;
808             else if (This->baseShader.shader_mode == SHADER_GLSL)
809                 hw_fct = curOpcode->hw_glsl_fct;
810             else if (This->baseShader.shader_mode == SHADER_ARB)
811                 hw_fct = curOpcode->hw_fct;
812
813             /* Unknown opcode and its parameters */
814             if (NULL == curOpcode) {
815                 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
816                 pToken += shader_skip_unrecognized(iface, pToken); 
817
818             /* Nothing to do */
819             } else if (WINED3DSIO_DCL == curOpcode->opcode ||
820                        WINED3DSIO_NOP == curOpcode->opcode ||
821                        WINED3DSIO_DEF == curOpcode->opcode ||
822                        WINED3DSIO_DEFI == curOpcode->opcode ||
823                        WINED3DSIO_DEFB == curOpcode->opcode ||
824                        WINED3DSIO_PHASE == curOpcode->opcode ||
825                        WINED3DSIO_RET == curOpcode->opcode) {
826
827                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
828
829             /* If a generator function is set for current shader target, use it */
830             } else if (hw_fct != NULL) {
831
832                 hw_arg.opcode = curOpcode;
833
834                 /* Destination token */
835                 if (curOpcode->dst_token) {
836
837                     DWORD param, addr_token = 0;
838                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
839                     hw_arg.dst = param;
840                     hw_arg.dst_addr = addr_token;
841                 }
842
843                 /* Predication token */
844                 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) 
845                     hw_arg.predicate = *pToken++;
846
847                 /* Other source tokens */
848                 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
849
850                     DWORD param, addr_token = 0; 
851                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
852                     hw_arg.src[i] = param;
853                     hw_arg.src_addr[i] = addr_token;
854                 }
855
856                 /* Call appropriate function for output target */
857                 hw_fct(&hw_arg);
858
859                 /* Add color correction if needed */
860                 device->shader_backend->shader_color_correction(&hw_arg);
861
862                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
863                 if (This->baseShader.shader_mode == SHADER_GLSL)
864                     shader_glsl_add_instruction_modifiers(&hw_arg);
865
866             /* Unhandled opcode */
867             } else {
868
869                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
870                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
871             }
872         }
873         /* TODO: What about result.depth? */
874
875     }
876 }
877
878 void shader_dump_ins_modifiers(const DWORD output) {
879
880     DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
881     DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
882
883     switch (shift) {
884         case 0: break;
885         case 13: TRACE("_d8"); break;
886         case 14: TRACE("_d4"); break;
887         case 15: TRACE("_d2"); break;
888         case 1: TRACE("_x2"); break;
889         case 2: TRACE("_x4"); break;
890         case 3: TRACE("_x8"); break;
891         default: TRACE("_unhandled_shift(%d)", shift); break;
892     }
893
894     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
895     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
896     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
897
898     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
899     if (mmask)
900         FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
901 }
902
903 /* First pass: trace shader, initialize length and version */
904 void shader_trace_init(
905     IWineD3DBaseShader *iface,
906     const DWORD* pFunction) {
907
908     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
909
910     const DWORD* pToken = pFunction;
911     const SHADER_OPCODE* curOpcode = NULL;
912     DWORD opcode_token;
913     unsigned int len = 0;
914     DWORD i;
915
916     TRACE("(%p) : Parsing programme\n", This);
917
918     if (NULL != pToken) {
919         while (WINED3DVS_END() != *pToken) {
920             if (shader_is_version_token(*pToken)) { /** version */
921                 This->baseShader.hex_version = *pToken;
922                 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
923                     WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
924                     WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
925                 ++pToken;
926                 ++len;
927                 continue;
928             }
929             if (shader_is_comment(*pToken)) { /** comment */
930                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
931                 ++pToken;
932                 TRACE("//%s\n", (const char*)pToken);
933                 pToken += comment_len;
934                 len += comment_len + 1;
935                 continue;
936             }
937             opcode_token = *pToken++;
938             curOpcode = shader_get_opcode(iface, opcode_token);
939             len++;
940
941             if (NULL == curOpcode) {
942                 int tokens_read;
943                 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
944                 tokens_read = shader_skip_unrecognized(iface, pToken);
945                 pToken += tokens_read;
946                 len += tokens_read;
947
948             } else {
949                 if (curOpcode->opcode == WINED3DSIO_DCL) {
950
951                     DWORD usage = *pToken;
952                     DWORD param = *(pToken + 1);
953
954                     shader_dump_decl_usage(This, usage, param);
955                     shader_dump_ins_modifiers(param);
956                     TRACE(" ");
957                     shader_dump_param(iface, param, 0, 0);
958                     pToken += 2;
959                     len += 2;
960
961                 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
962
963                         unsigned int offset = shader_get_float_offset(*pToken);
964
965                         TRACE("def c%u = %f, %f, %f, %f", offset,
966                             *(const float *)(pToken + 1),
967                             *(const float *)(pToken + 2),
968                             *(const float *)(pToken + 3),
969                             *(const float *)(pToken + 4));
970
971                         pToken += 5;
972                         len += 5;
973                 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
974
975                         TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
976                             *(pToken + 1),
977                             *(pToken + 2),
978                             *(pToken + 3),
979                             *(pToken + 4));
980
981                         pToken += 5;
982                         len += 5;
983
984                 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
985
986                         TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
987                             *(pToken + 1)? "true": "false");
988
989                         pToken += 2;
990                         len += 2;
991
992                 } else {
993
994                     DWORD param, addr_token;
995                     int tokens_read;
996
997                     /* Print out predication source token first - it follows
998                      * the destination token. */
999                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1000                         TRACE("(");
1001                         shader_dump_param(iface, *(pToken + 2), 0, 1);
1002                         TRACE(") ");
1003                     }
1004                     if (opcode_token & WINED3DSI_COISSUE) {
1005                         /* PixWin marks instructions with the coissue flag with a '+' */
1006                         TRACE("+");
1007                     }
1008
1009                     TRACE("%s", curOpcode->name);
1010
1011                     if (curOpcode->opcode == WINED3DSIO_IFC ||
1012                         curOpcode->opcode == WINED3DSIO_BREAKC) {
1013
1014                         DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
1015                         switch (op) {
1016                             case COMPARISON_GT: TRACE("_gt"); break;
1017                             case COMPARISON_EQ: TRACE("_eq"); break;
1018                             case COMPARISON_GE: TRACE("_ge"); break;
1019                             case COMPARISON_LT: TRACE("_lt"); break;
1020                             case COMPARISON_NE: TRACE("_ne"); break;
1021                             case COMPARISON_LE: TRACE("_le"); break;
1022                             default:
1023                                 TRACE("_(%u)", op);
1024                         }
1025                     } else if (curOpcode->opcode == WINED3DSIO_TEX &&
1026                                This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
1027                         if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
1028                     }
1029
1030                     /* Destination token */
1031                     if (curOpcode->dst_token) {
1032
1033                         /* Destination token */
1034                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1035                         pToken += tokens_read;
1036                         len += tokens_read;
1037
1038                         shader_dump_ins_modifiers(param);
1039                         TRACE(" ");
1040                         shader_dump_param(iface, param, addr_token, 0);
1041                     }
1042
1043                     /* Predication token - already printed out, just skip it */
1044                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1045                         pToken++;
1046                         len++;
1047                     }
1048
1049                     /* Other source tokens */
1050                     for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
1051
1052                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1053                         pToken += tokens_read;
1054                         len += tokens_read;
1055
1056                         TRACE((i == 0)? " " : ", ");
1057                         shader_dump_param(iface, param, addr_token, 1);
1058                     }
1059                 }
1060                 TRACE("\n");
1061             }
1062         }
1063         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1064     } else {
1065         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1066     }
1067 }
1068
1069 static void shader_delete_constant_list(
1070     struct list* clist) {
1071
1072     struct list *ptr;
1073     struct local_constant* constant;
1074
1075     ptr = list_head(clist);
1076     while (ptr) {
1077         constant = LIST_ENTRY(ptr, struct local_constant, entry);
1078         ptr = list_next(clist, ptr);
1079         HeapFree(GetProcessHeap(), 0, constant);
1080     }
1081 }
1082
1083 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1084 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1085 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1086 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1087 static void shader_none_color_correction(SHADER_OPCODE_ARG* arg) {}
1088 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1089
1090 const shader_backend_t none_shader_backend = {
1091     &shader_none_select,
1092     &shader_none_select_depth_blt,
1093     &shader_none_load_constants,
1094     &shader_none_cleanup,
1095     &shader_none_color_correction,
1096     &shader_none_destroy
1097 };
1098
1099 /* *******************************************
1100    IWineD3DPixelShader IUnknown parts follow
1101    ******************************************* */
1102 HRESULT  WINAPI IWineD3DBaseShaderImpl_QueryInterface(IWineD3DBaseShader *iface, REFIID riid, LPVOID *ppobj)
1103 {
1104     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1105     TRACE("(%p)->(%s,%p)\n",This,debugstr_guid(riid),ppobj);
1106     if (IsEqualGUID(riid, &IID_IUnknown)
1107         || IsEqualGUID(riid, &IID_IWineD3DBase)
1108         || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1109         || IsEqualGUID(riid, &IID_IWineD3DPixelShader)) {
1110         IUnknown_AddRef(iface);
1111         *ppobj = This;
1112         return S_OK;
1113     }
1114     *ppobj = NULL;
1115     return E_NOINTERFACE;
1116 }
1117
1118 ULONG  WINAPI IWineD3DBaseShaderImpl_AddRef(IWineD3DBaseShader *iface) {
1119     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1120     TRACE("(%p) : AddRef increasing from %d\n", This, This->baseShader.ref);
1121     return InterlockedIncrement(&This->baseShader.ref);
1122 }
1123
1124 ULONG  WINAPI IWineD3DBaseShaderImpl_Release(IWineD3DBaseShader *iface) {
1125     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1126     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
1127     ULONG ref;
1128     TRACE("(%p) : Releasing from %d\n", This, This->baseShader.ref);
1129     ref = InterlockedDecrement(&This->baseShader.ref);
1130     if (ref == 0) {
1131         deviceImpl->shader_backend->shader_destroy(iface);
1132         HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1133         shader_delete_constant_list(&This->baseShader.constantsF);
1134         shader_delete_constant_list(&This->baseShader.constantsB);
1135         shader_delete_constant_list(&This->baseShader.constantsI);
1136         HeapFree(GetProcessHeap(), 0, This);
1137     }
1138     return ref;
1139 }