user32: Change the desktop colour and pattern to match win2k.
[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                         case GL_TEXTURE_2D:
353                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
354                             break;
355
356                         case GL_TEXTURE_3D:
357                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
358                             break;
359
360                         case GL_TEXTURE_CUBE_MAP_ARB:
361                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
362                             break;
363
364                         default:
365                             ERR("Unexpected gl texture type found: %d\n", texType);
366                             reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
367                     }
368                 }
369
370                 /* texbem is only valid with < 1.4 pixel shaders */
371                 if(WINED3DSIO_TEXBEM  == curOpcode->opcode ||
372                     WINED3DSIO_TEXBEML == curOpcode->opcode) {
373                     if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != sampler_code) {
374                         FIXME("Pixel shader uses texbem instruction on more than 1 sampler\n");
375                     } else {
376                         reg_maps->bumpmat = sampler_code;
377                         if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
378                             reg_maps->luminanceparams = sampler_code;
379                         }
380                     }
381                 }
382             }
383             if(WINED3DSIO_NRM  == curOpcode->opcode) {
384                 reg_maps->usesnrm = 1;
385             } else if(WINED3DSIO_BEM == curOpcode->opcode) {
386                 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
387                 if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != regnum) {
388                     FIXME("Pixel shader uses bem or texbem instruction on more than 1 sampler\n");
389                 } else {
390                     reg_maps->bumpmat = regnum;
391                 }
392             } else if(WINED3DSIO_DSY  == curOpcode->opcode) {
393                 reg_maps->usesdsy = 1;
394             }
395
396             /* This will loop over all the registers and try to
397              * make a bitmask of the ones we're interested in. 
398              *
399              * Relative addressing tokens are ignored, but that's 
400              * okay, since we'll catch any address registers when 
401              * they are initialized (required by spec) */
402
403             limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
404                 curOpcode->num_params + 1: curOpcode->num_params;
405
406             for (i = 0; i < limit; ++i) {
407
408                 DWORD param, addr_token, reg, regtype;
409                 pToken += shader_get_param(iface, pToken, &param, &addr_token);
410
411                 regtype = shader_get_regtype(param);
412                 reg = param & WINED3DSP_REGNUM_MASK;
413
414                 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
415
416                     if (pshader)
417                         reg_maps->texcoord[reg] = 1;
418                     else
419                         reg_maps->address[reg] = 1;
420                 }
421
422                 else if (WINED3DSPR_TEMP == regtype)
423                     reg_maps->temporary[reg] = 1;
424
425                 else if (WINED3DSPR_INPUT == regtype) {
426                     if( !pshader)
427                         reg_maps->attributes[reg] = 1;
428                     else {
429                         if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
430                             /* If relative addressing is used, we must assume that all registers
431                              * are used. Even if it is a construct like v3[aL], we can't assume
432                              * that v0, v1 and v2 aren't read because aL can be negative
433                              */
434                             unsigned int i;
435                             for(i = 0; i < MAX_REG_INPUT; i++) {
436                                 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[i] = TRUE;
437                             }
438                         } else {
439                             ((IWineD3DPixelShaderImpl *) This)->input_reg_used[reg] = TRUE;
440                         }
441                     }
442                 }
443
444                 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
445                     reg_maps->fog = 1;
446
447                 else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
448                     reg_maps->vpos = 1;
449
450                 else if(WINED3DSPR_CONST == regtype) {
451                     if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
452                         if(!pshader) {
453                             if(reg <= ((IWineD3DVertexShaderImpl *) This)->min_rel_offset) {
454                                 ((IWineD3DVertexShaderImpl *) This)->min_rel_offset = reg;
455                             } else if(reg >= ((IWineD3DVertexShaderImpl *) This)->max_rel_offset) {
456                                 ((IWineD3DVertexShaderImpl *) This)->max_rel_offset = reg;
457                             }
458                         }
459                         reg_maps->usesrelconstF = TRUE;
460                     }
461                 }
462             }
463         }
464     }
465     reg_maps->loop_depth = max_loop_depth;
466
467     return WINED3D_OK;
468 }
469
470 static void shader_dump_decl_usage(
471     IWineD3DBaseShaderImpl* This,
472     DWORD decl, 
473     DWORD param) {
474
475     DWORD regtype = shader_get_regtype(param);
476
477     TRACE("dcl");
478
479     if (regtype == WINED3DSPR_SAMPLER) {
480         DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
481
482         switch (ttype) {
483             case WINED3DSTT_2D: TRACE("_2d"); break;
484             case WINED3DSTT_CUBE: TRACE("_cube"); break;
485             case WINED3DSTT_VOLUME: TRACE("_volume"); break;
486             default: TRACE("_unknown_ttype(%08x)", ttype); 
487        }
488
489     } else { 
490
491         DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
492         DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
493
494         /* Pixel shaders 3.0 don't have usage semantics */
495         char pshader = shader_is_pshader_version(This->baseShader.hex_version);
496         if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
497             return;
498         else
499             TRACE("_");
500
501         switch(usage) {
502         case WINED3DDECLUSAGE_POSITION:
503             TRACE("position%d", idx);
504             break;
505         case WINED3DDECLUSAGE_BLENDINDICES:
506             TRACE("blend");
507             break;
508         case WINED3DDECLUSAGE_BLENDWEIGHT:
509             TRACE("weight");
510             break;
511         case WINED3DDECLUSAGE_NORMAL:
512             TRACE("normal%d", idx);
513             break;
514         case WINED3DDECLUSAGE_PSIZE:
515             TRACE("psize");
516             break;
517         case WINED3DDECLUSAGE_COLOR:
518             if(idx == 0)  {
519                 TRACE("color");
520             } else {
521                 TRACE("specular%d", (idx - 1));
522             }
523             break;
524         case WINED3DDECLUSAGE_TEXCOORD:
525             TRACE("texture%d", idx);
526             break;
527         case WINED3DDECLUSAGE_TANGENT:
528             TRACE("tangent");
529             break;
530         case WINED3DDECLUSAGE_BINORMAL:
531             TRACE("binormal");
532             break;
533         case WINED3DDECLUSAGE_TESSFACTOR:
534             TRACE("tessfactor");
535             break;
536         case WINED3DDECLUSAGE_POSITIONT:
537             TRACE("positionT%d", idx);
538             break;
539         case WINED3DDECLUSAGE_FOG:
540             TRACE("fog");
541             break;
542         case WINED3DDECLUSAGE_DEPTH:
543             TRACE("depth");
544             break;
545         case WINED3DDECLUSAGE_SAMPLE:
546             TRACE("sample");
547             break;
548         default:
549             FIXME("unknown_semantics(%08x)", usage);
550         }
551     }
552 }
553
554 static void shader_dump_arr_entry(
555     IWineD3DBaseShader *iface,
556     const DWORD param,
557     const DWORD addr_token,
558     unsigned int reg,
559     int input) {
560
561     char relative =
562         ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
563
564     if (relative) {
565         TRACE("[");
566         if (addr_token)
567             shader_dump_param(iface, addr_token, 0, input);
568         else
569             TRACE("a0.x");
570         TRACE(" + ");
571      }
572      TRACE("%u", reg);
573      if (relative)
574          TRACE("]");
575 }
576
577 void shader_dump_param(
578     IWineD3DBaseShader *iface,
579     const DWORD param, 
580     const DWORD addr_token,
581     int input) {
582
583     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
584     static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
585     static const char * const misctype_reg_names[] = { "vPos", "vFace"};
586     char swizzle_reg_chars[4];
587
588     DWORD reg = param & WINED3DSP_REGNUM_MASK;
589     DWORD regtype = shader_get_regtype(param);
590     DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
591
592     /* There are some minor differences between pixel and vertex shaders */
593     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
594
595     /* For one, we'd prefer color components to be shown for pshaders.
596      * FIXME: use the swizzle function for this */
597
598     swizzle_reg_chars[0] = pshader? 'r': 'x';
599     swizzle_reg_chars[1] = pshader? 'g': 'y';
600     swizzle_reg_chars[2] = pshader? 'b': 'z';
601     swizzle_reg_chars[3] = pshader? 'a': 'w';
602
603     if (input) {
604         if ( (modifier == WINED3DSPSM_NEG) ||
605              (modifier == WINED3DSPSM_BIASNEG) ||
606              (modifier == WINED3DSPSM_SIGNNEG) ||
607              (modifier == WINED3DSPSM_X2NEG) ||
608              (modifier == WINED3DSPSM_ABSNEG) )
609             TRACE("-");
610         else if (modifier == WINED3DSPSM_COMP)
611             TRACE("1-");
612         else if (modifier == WINED3DSPSM_NOT)
613             TRACE("!");
614
615         if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG) 
616             TRACE("abs(");
617     }
618
619     switch (regtype) {
620         case WINED3DSPR_TEMP:
621             TRACE("r%u", reg);
622             break;
623         case WINED3DSPR_INPUT:
624             TRACE("v");
625             shader_dump_arr_entry(iface, param, addr_token, reg, input);
626             break;
627         case WINED3DSPR_CONST:
628         case WINED3DSPR_CONST2:
629         case WINED3DSPR_CONST3:
630         case WINED3DSPR_CONST4:
631             TRACE("c");
632             shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
633             break;
634         case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
635             TRACE("%c%u", (pshader? 't':'a'), reg);
636             break;        
637         case WINED3DSPR_RASTOUT:
638             TRACE("%s", rastout_reg_names[reg]);
639             break;
640         case WINED3DSPR_COLOROUT:
641             TRACE("oC%u", reg);
642             break;
643         case WINED3DSPR_DEPTHOUT:
644             TRACE("oDepth");
645             break;
646         case WINED3DSPR_ATTROUT:
647             TRACE("oD%u", reg);
648             break;
649         case WINED3DSPR_TEXCRDOUT: 
650
651             /* Vertex shaders >= 3.0 use general purpose output registers
652              * (WINED3DSPR_OUTPUT), which can include an address token */
653
654             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
655                 TRACE("o");
656                 shader_dump_arr_entry(iface, param, addr_token, reg, input);
657             }
658             else 
659                TRACE("oT%u", reg);
660             break;
661         case WINED3DSPR_CONSTINT:
662             TRACE("i");
663             shader_dump_arr_entry(iface, param, addr_token, reg, input);
664             break;
665         case WINED3DSPR_CONSTBOOL:
666             TRACE("b");
667             shader_dump_arr_entry(iface, param, addr_token, reg, input);
668             break;
669         case WINED3DSPR_LABEL:
670             TRACE("l%u", reg);
671             break;
672         case WINED3DSPR_LOOP:
673             TRACE("aL");
674             break;
675         case WINED3DSPR_SAMPLER:
676             TRACE("s%u", reg);
677             break;
678         case WINED3DSPR_MISCTYPE:
679             if (reg > 1) {
680                 FIXME("Unhandled misctype register %d\n", reg);
681             } else {
682                 TRACE("%s", misctype_reg_names[reg]);
683             }
684             break;
685         case WINED3DSPR_PREDICATE:
686             TRACE("p%u", reg);
687             break;
688         default:
689             TRACE("unhandled_rtype(%#x)", regtype);
690             break;
691    }
692
693    if (!input) {
694        /* operand output (for modifiers and shift, see dump_ins_modifiers) */
695
696        if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
697            TRACE(".");
698            if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
699            if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
700            if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
701            if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
702        }
703
704    } else {
705         /** operand input */
706         DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
707         DWORD swizzle_r = swizzle & 0x03;
708         DWORD swizzle_g = (swizzle >> 2) & 0x03;
709         DWORD swizzle_b = (swizzle >> 4) & 0x03;
710         DWORD swizzle_a = (swizzle >> 6) & 0x03;
711
712         if (0 != modifier) {
713             switch (modifier) {
714                 case WINED3DSPSM_NONE:    break;
715                 case WINED3DSPSM_NEG:     break;
716                 case WINED3DSPSM_NOT:     break;
717                 case WINED3DSPSM_BIAS:    TRACE("_bias"); break;
718                 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
719                 case WINED3DSPSM_SIGN:    TRACE("_bx2"); break;
720                 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
721                 case WINED3DSPSM_COMP:    break;
722                 case WINED3DSPSM_X2:      TRACE("_x2"); break;
723                 case WINED3DSPSM_X2NEG:   TRACE("_x2"); break;
724                 case WINED3DSPSM_DZ:      TRACE("_dz"); break;
725                 case WINED3DSPSM_DW:      TRACE("_dw"); break;
726                 case WINED3DSPSM_ABSNEG:  TRACE(")"); break;
727                 case WINED3DSPSM_ABS:     TRACE(")"); break;
728                 default:
729                     TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
730             }
731         }
732
733         /**
734         * swizzle bits fields:
735         *  RRGGBBAA
736         */
737         if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
738             if (swizzle_r == swizzle_g &&
739                 swizzle_r == swizzle_b &&
740                 swizzle_r == swizzle_a) {
741                     TRACE(".%c", swizzle_reg_chars[swizzle_r]);
742             } else {
743                 TRACE(".%c%c%c%c",
744                 swizzle_reg_chars[swizzle_r],
745                 swizzle_reg_chars[swizzle_g],
746                 swizzle_reg_chars[swizzle_b],
747                 swizzle_reg_chars[swizzle_a]);
748             }
749         }
750     }
751 }
752
753 /** Shared code in order to generate the bulk of the shader string.
754     Use the shader_header_fct & shader_footer_fct to add strings
755     that are specific to pixel or vertex functions
756     NOTE: A description of how to parse tokens can be found at:
757           http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
758 void shader_generate_main(
759     IWineD3DBaseShader *iface,
760     SHADER_BUFFER* buffer,
761     shader_reg_maps* reg_maps,
762     CONST DWORD* pFunction) {
763
764     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
765     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
766     const DWORD *pToken = pFunction;
767     const SHADER_OPCODE *curOpcode = NULL;
768     SHADER_HANDLER hw_fct = NULL;
769     DWORD i;
770     SHADER_OPCODE_ARG hw_arg;
771
772     /* Initialize current parsing state */
773     hw_arg.shader = iface;
774     hw_arg.buffer = buffer;
775     hw_arg.reg_maps = reg_maps;
776     This->baseShader.parse_state.current_row = 0;
777
778     /* Second pass, process opcodes */
779     if (NULL != pToken) {
780         while (WINED3DPS_END() != *pToken) {
781
782             /* Skip version token */
783             if (shader_is_version_token(*pToken)) {
784                 ++pToken;
785                 continue;
786             }
787
788             /* Skip comment tokens */
789             if (shader_is_comment(*pToken)) {
790                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
791                 ++pToken;
792                 TRACE("#%s\n", (const char*)pToken);
793                 pToken += comment_len;
794                 continue;
795             }
796
797             /* Read opcode */
798             hw_arg.opcode_token = *pToken++;
799             curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
800
801             /* Select handler */
802             if (curOpcode == NULL)
803                 hw_fct = NULL;
804             else if (This->baseShader.shader_mode == SHADER_GLSL)
805                 hw_fct = curOpcode->hw_glsl_fct;
806             else if (This->baseShader.shader_mode == SHADER_ARB)
807                 hw_fct = curOpcode->hw_fct;
808
809             /* Unknown opcode and its parameters */
810             if (NULL == curOpcode) {
811                 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
812                 pToken += shader_skip_unrecognized(iface, pToken); 
813
814             /* Nothing to do */
815             } else if (WINED3DSIO_DCL == curOpcode->opcode ||
816                        WINED3DSIO_NOP == curOpcode->opcode ||
817                        WINED3DSIO_DEF == curOpcode->opcode ||
818                        WINED3DSIO_DEFI == curOpcode->opcode ||
819                        WINED3DSIO_DEFB == curOpcode->opcode ||
820                        WINED3DSIO_PHASE == curOpcode->opcode ||
821                        WINED3DSIO_RET == curOpcode->opcode) {
822
823                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
824
825             /* If a generator function is set for current shader target, use it */
826             } else if (hw_fct != NULL) {
827
828                 hw_arg.opcode = curOpcode;
829
830                 /* Destination token */
831                 if (curOpcode->dst_token) {
832
833                     DWORD param, addr_token = 0;
834                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
835                     hw_arg.dst = param;
836                     hw_arg.dst_addr = addr_token;
837                 }
838
839                 /* Predication token */
840                 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) 
841                     hw_arg.predicate = *pToken++;
842
843                 /* Other source tokens */
844                 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
845
846                     DWORD param, addr_token = 0; 
847                     pToken += shader_get_param(iface, pToken, &param, &addr_token);
848                     hw_arg.src[i] = param;
849                     hw_arg.src_addr[i] = addr_token;
850                 }
851
852                 /* Call appropriate function for output target */
853                 hw_fct(&hw_arg);
854
855                 /* Add color correction if needed */
856                 device->shader_backend->shader_color_correction(&hw_arg);
857
858                 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
859                 if (This->baseShader.shader_mode == SHADER_GLSL)
860                     shader_glsl_add_instruction_modifiers(&hw_arg);
861
862             /* Unhandled opcode */
863             } else {
864
865                 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
866                 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
867             }
868         }
869         /* TODO: What about result.depth? */
870
871     }
872 }
873
874 void shader_dump_ins_modifiers(const DWORD output) {
875
876     DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
877     DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
878
879     switch (shift) {
880         case 0: break;
881         case 13: TRACE("_d8"); break;
882         case 14: TRACE("_d4"); break;
883         case 15: TRACE("_d2"); break;
884         case 1: TRACE("_x2"); break;
885         case 2: TRACE("_x4"); break;
886         case 3: TRACE("_x8"); break;
887         default: TRACE("_unhandled_shift(%d)", shift); break;
888     }
889
890     if (mmask & WINED3DSPDM_SATURATE)         TRACE("_sat");
891     if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
892     if (mmask & WINED3DSPDM_MSAMPCENTROID)    TRACE("_centroid");
893
894     mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
895     if (mmask)
896         FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
897 }
898
899 /* First pass: trace shader, initialize length and version */
900 void shader_trace_init(
901     IWineD3DBaseShader *iface,
902     const DWORD* pFunction) {
903
904     IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
905
906     const DWORD* pToken = pFunction;
907     const SHADER_OPCODE* curOpcode = NULL;
908     DWORD opcode_token;
909     unsigned int len = 0;
910     DWORD i;
911
912     TRACE("(%p) : Parsing programme\n", This);
913
914     if (NULL != pToken) {
915         while (WINED3DVS_END() != *pToken) {
916             if (shader_is_version_token(*pToken)) { /** version */
917                 This->baseShader.hex_version = *pToken;
918                 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
919                     WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
920                     WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
921                 ++pToken;
922                 ++len;
923                 continue;
924             }
925             if (shader_is_comment(*pToken)) { /** comment */
926                 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
927                 ++pToken;
928                 TRACE("//%s\n", (const char*)pToken);
929                 pToken += comment_len;
930                 len += comment_len + 1;
931                 continue;
932             }
933             opcode_token = *pToken++;
934             curOpcode = shader_get_opcode(iface, opcode_token);
935             len++;
936
937             if (NULL == curOpcode) {
938                 int tokens_read;
939                 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
940                 tokens_read = shader_skip_unrecognized(iface, pToken);
941                 pToken += tokens_read;
942                 len += tokens_read;
943
944             } else {
945                 if (curOpcode->opcode == WINED3DSIO_DCL) {
946
947                     DWORD usage = *pToken;
948                     DWORD param = *(pToken + 1);
949
950                     shader_dump_decl_usage(This, usage, param);
951                     shader_dump_ins_modifiers(param);
952                     TRACE(" ");
953                     shader_dump_param(iface, param, 0, 0);
954                     pToken += 2;
955                     len += 2;
956
957                 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
958
959                         unsigned int offset = shader_get_float_offset(*pToken);
960
961                         TRACE("def c%u = %f, %f, %f, %f", offset,
962                             *(const float *)(pToken + 1),
963                             *(const float *)(pToken + 2),
964                             *(const float *)(pToken + 3),
965                             *(const float *)(pToken + 4));
966
967                         pToken += 5;
968                         len += 5;
969                 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
970
971                         TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
972                             *(pToken + 1),
973                             *(pToken + 2),
974                             *(pToken + 3),
975                             *(pToken + 4));
976
977                         pToken += 5;
978                         len += 5;
979
980                 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
981
982                         TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
983                             *(pToken + 1)? "true": "false");
984
985                         pToken += 2;
986                         len += 2;
987
988                 } else {
989
990                     DWORD param, addr_token;
991                     int tokens_read;
992
993                     /* Print out predication source token first - it follows
994                      * the destination token. */
995                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
996                         TRACE("(");
997                         shader_dump_param(iface, *(pToken + 2), 0, 1);
998                         TRACE(") ");
999                     }
1000                     if (opcode_token & WINED3DSI_COISSUE) {
1001                         /* PixWin marks instructions with the coissue flag with a '+' */
1002                         TRACE("+");
1003                     }
1004
1005                     TRACE("%s", curOpcode->name);
1006
1007                     if (curOpcode->opcode == WINED3DSIO_IFC ||
1008                         curOpcode->opcode == WINED3DSIO_BREAKC) {
1009
1010                         DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
1011                         switch (op) {
1012                             case COMPARISON_GT: TRACE("_gt"); break;
1013                             case COMPARISON_EQ: TRACE("_eq"); break;
1014                             case COMPARISON_GE: TRACE("_ge"); break;
1015                             case COMPARISON_LT: TRACE("_lt"); break;
1016                             case COMPARISON_NE: TRACE("_ne"); break;
1017                             case COMPARISON_LE: TRACE("_le"); break;
1018                             default:
1019                                 TRACE("_(%u)", op);
1020                         }
1021                     } else if (curOpcode->opcode == WINED3DSIO_TEX &&
1022                                This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
1023                         if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
1024                     }
1025
1026                     /* Destination token */
1027                     if (curOpcode->dst_token) {
1028
1029                         /* Destination token */
1030                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1031                         pToken += tokens_read;
1032                         len += tokens_read;
1033
1034                         shader_dump_ins_modifiers(param);
1035                         TRACE(" ");
1036                         shader_dump_param(iface, param, addr_token, 0);
1037                     }
1038
1039                     /* Predication token - already printed out, just skip it */
1040                     if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1041                         pToken++;
1042                         len++;
1043                     }
1044
1045                     /* Other source tokens */
1046                     for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
1047
1048                         tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1049                         pToken += tokens_read;
1050                         len += tokens_read;
1051
1052                         TRACE((i == 0)? " " : ", ");
1053                         shader_dump_param(iface, param, addr_token, 1);
1054                     }
1055                 }
1056                 TRACE("\n");
1057             }
1058         }
1059         This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1060     } else {
1061         This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1062     }
1063 }
1064
1065 static void shader_delete_constant_list(
1066     struct list* clist) {
1067
1068     struct list *ptr;
1069     struct local_constant* constant;
1070
1071     ptr = list_head(clist);
1072     while (ptr) {
1073         constant = LIST_ENTRY(ptr, struct local_constant, entry);
1074         ptr = list_next(clist, ptr);
1075         HeapFree(GetProcessHeap(), 0, constant);
1076     }
1077 }
1078
1079 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1080 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1081 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1082 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1083 static void shader_none_color_correction(SHADER_OPCODE_ARG* arg) {}
1084 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1085
1086 const shader_backend_t none_shader_backend = {
1087     &shader_none_select,
1088     &shader_none_select_depth_blt,
1089     &shader_none_load_constants,
1090     &shader_none_cleanup,
1091     &shader_none_color_correction,
1092     &shader_none_destroy
1093 };
1094
1095 /* *******************************************
1096    IWineD3DPixelShader IUnknown parts follow
1097    ******************************************* */
1098 HRESULT  WINAPI IWineD3DBaseShaderImpl_QueryInterface(IWineD3DBaseShader *iface, REFIID riid, LPVOID *ppobj)
1099 {
1100     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1101     TRACE("(%p)->(%s,%p)\n",This,debugstr_guid(riid),ppobj);
1102     if (IsEqualGUID(riid, &IID_IUnknown)
1103         || IsEqualGUID(riid, &IID_IWineD3DBase)
1104         || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1105         || IsEqualGUID(riid, &IID_IWineD3DPixelShader)) {
1106         IUnknown_AddRef(iface);
1107         *ppobj = This;
1108         return S_OK;
1109     }
1110     *ppobj = NULL;
1111     return E_NOINTERFACE;
1112 }
1113
1114 ULONG  WINAPI IWineD3DBaseShaderImpl_AddRef(IWineD3DBaseShader *iface) {
1115     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1116     TRACE("(%p) : AddRef increasing from %d\n", This, This->baseShader.ref);
1117     return InterlockedIncrement(&This->baseShader.ref);
1118 }
1119
1120 ULONG  WINAPI IWineD3DBaseShaderImpl_Release(IWineD3DBaseShader *iface) {
1121     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1122     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
1123     ULONG ref;
1124     TRACE("(%p) : Releasing from %d\n", This, This->baseShader.ref);
1125     ref = InterlockedDecrement(&This->baseShader.ref);
1126     if (ref == 0) {
1127         deviceImpl->shader_backend->shader_destroy(iface);
1128         HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1129         shader_delete_constant_list(&This->baseShader.constantsF);
1130         shader_delete_constant_list(&This->baseShader.constantsB);
1131         shader_delete_constant_list(&This->baseShader.constantsI);
1132         HeapFree(GetProcessHeap(), 0, This);
1133     }
1134     return ref;
1135 }