glsl: simplify IR storage for samplers
[mesa.git] / src / mesa / shader / slang / slang_emit.c
1 /*
2 * Mesa 3-D graphics library
3 *
4 * Copyright (C) 2005-2008 Brian Paul All Rights Reserved.
5 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
21 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
22 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 /**
26 * \file slang_emit.c
27 * Emit program instructions (PI code) from IR trees.
28 * \author Brian Paul
29 */
30
31 /***
32 *** NOTES
33 ***
34 *** To emit GPU instructions, we basically just do an in-order traversal
35 *** of the IR tree.
36 ***/
37
38
39 #include "main/imports.h"
40 #include "main/context.h"
41 #include "main/macros.h"
42 #include "shader/program.h"
43 #include "shader/prog_instruction.h"
44 #include "shader/prog_parameter.h"
45 #include "shader/prog_print.h"
46 #include "slang_builtin.h"
47 #include "slang_emit.h"
48 #include "slang_mem.h"
49
50
51 #define PEEPHOLE_OPTIMIZATIONS 1
52 #define ANNOTATE 0
53
54
55 typedef struct
56 {
57 slang_info_log *log;
58 slang_var_table *vt;
59 struct gl_program *prog;
60 struct gl_program **Subroutines;
61 GLuint NumSubroutines;
62
63 GLuint MaxInstructions; /**< size of prog->Instructions[] buffer */
64
65 /* code-gen options */
66 GLboolean EmitHighLevelInstructions;
67 GLboolean EmitCondCodes;
68 GLboolean EmitComments;
69 GLboolean EmitBeginEndSub; /* XXX TEMPORARY */
70 } slang_emit_info;
71
72
73
74 static struct gl_program *
75 new_subroutine(slang_emit_info *emitInfo, GLuint *id)
76 {
77 GET_CURRENT_CONTEXT(ctx);
78 const GLuint n = emitInfo->NumSubroutines;
79
80 emitInfo->Subroutines = (struct gl_program **)
81 _mesa_realloc(emitInfo->Subroutines,
82 n * sizeof(struct gl_program),
83 (n + 1) * sizeof(struct gl_program));
84 emitInfo->Subroutines[n] = ctx->Driver.NewProgram(ctx, emitInfo->prog->Target, 0);
85 emitInfo->Subroutines[n]->Parameters = emitInfo->prog->Parameters;
86 emitInfo->NumSubroutines++;
87 *id = n;
88 return emitInfo->Subroutines[n];
89 }
90
91
92 /**
93 * Convert a writemask to a swizzle. Used for testing cond codes because
94 * we only want to test the cond code component(s) that was set by the
95 * previous instruction.
96 */
97 static GLuint
98 writemask_to_swizzle(GLuint writemask)
99 {
100 if (writemask == WRITEMASK_X)
101 return SWIZZLE_XXXX;
102 if (writemask == WRITEMASK_Y)
103 return SWIZZLE_YYYY;
104 if (writemask == WRITEMASK_Z)
105 return SWIZZLE_ZZZZ;
106 if (writemask == WRITEMASK_W)
107 return SWIZZLE_WWWW;
108 return SWIZZLE_XYZW; /* shouldn't be hit */
109 }
110
111
112 /**
113 * Convert a swizzle mask to a writemask.
114 * Note that the slang_ir_storage->Swizzle field can represent either a
115 * swizzle mask or a writemask, depending on how it's used. For example,
116 * when we parse "direction.yz" alone, we don't know whether .yz is a
117 * writemask or a swizzle. In this case, we encode ".yz" in store->Swizzle
118 * as a swizzle mask (.yz?? actually). Later, if direction.yz is used as
119 * an R-value, we use store->Swizzle as-is. Otherwise, if direction.yz is
120 * used as an L-value, we convert it to a writemask.
121 */
122 static GLuint
123 swizzle_to_writemask(GLuint swizzle)
124 {
125 GLuint i, writemask = 0x0;
126 for (i = 0; i < 4; i++) {
127 GLuint swz = GET_SWZ(swizzle, i);
128 if (swz <= SWIZZLE_W) {
129 writemask |= (1 << swz);
130 }
131 }
132 return writemask;
133 }
134
135
136 /**
137 * Swizzle a swizzle (function composition).
138 * That is, return swz2(swz1), or said another way: swz1.szw2
139 * Example: swizzle_swizzle(".zwxx", ".xxyw") yields ".zzwx"
140 */
141 GLuint
142 _slang_swizzle_swizzle(GLuint swz1, GLuint swz2)
143 {
144 GLuint i, swz, s[4];
145 for (i = 0; i < 4; i++) {
146 GLuint c = GET_SWZ(swz2, i);
147 if (c <= SWIZZLE_W)
148 s[i] = GET_SWZ(swz1, c);
149 else
150 s[i] = c;
151 }
152 swz = MAKE_SWIZZLE4(s[0], s[1], s[2], s[3]);
153 return swz;
154 }
155
156
157 /**
158 * Return the default swizzle mask for accessing a variable of the
159 * given size (in floats). If size = 1, comp is used to identify
160 * which component [0..3] of the register holds the variable.
161 */
162 GLuint
163 _slang_var_swizzle(GLint size, GLint comp)
164 {
165 switch (size) {
166 case 1:
167 return MAKE_SWIZZLE4(comp, comp, comp, comp);
168 case 2:
169 return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_NIL, SWIZZLE_NIL);
170 case 3:
171 return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_NIL);
172 default:
173 return SWIZZLE_XYZW;
174 }
175 }
176
177
178
179 /**
180 * Allocate storage for the given node (if it hasn't already been allocated).
181 *
182 * Typically this is temporary storage for an intermediate result (such as
183 * for a multiply or add, etc).
184 *
185 * If n->Store does not exist it will be created and will be of the size
186 * specified by defaultSize.
187 */
188 static GLboolean
189 alloc_node_storage(slang_emit_info *emitInfo, slang_ir_node *n,
190 GLint defaultSize)
191 {
192 assert(!n->Var);
193 if (!n->Store) {
194 assert(defaultSize > 0);
195 n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, defaultSize);
196 }
197
198 /* now allocate actual register(s). I.e. set n->Store->Index >= 0 */
199 if (n->Store->Index < 0) {
200 if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
201 slang_info_log_error(emitInfo->log,
202 "Ran out of registers, too many temporaries");
203 _slang_free(n->Store);
204 n->Store = NULL;
205 return GL_FALSE;
206 }
207 }
208 return GL_TRUE;
209 }
210
211
212 /**
213 * Free temporary storage, if n->Store is, in fact, temp storage.
214 * Otherwise, no-op.
215 */
216 static void
217 free_node_storage(slang_var_table *vt, slang_ir_node *n)
218 {
219 if (n->Store->File == PROGRAM_TEMPORARY &&
220 n->Store->Index >= 0 &&
221 n->Opcode != IR_SWIZZLE) {
222 if (_slang_is_temp(vt, n->Store)) {
223 _slang_free_temp(vt, n->Store);
224 n->Store->Index = -1;
225 n->Store = NULL; /* XXX this may not be needed */
226 }
227 }
228 }
229
230
231 /**
232 * Helper function to allocate a short-term temporary.
233 * Free it with _slang_free_temp().
234 */
235 static GLboolean
236 alloc_local_temp(slang_emit_info *emitInfo, slang_ir_storage *temp, GLint size)
237 {
238 assert(size >= 1);
239 assert(size <= 4);
240 _mesa_bzero(temp, sizeof(*temp));
241 temp->Size = size;
242 temp->File = PROGRAM_TEMPORARY;
243 temp->Index = -1;
244 return _slang_alloc_temp(emitInfo->vt, temp);
245 }
246
247
248 /**
249 * Remove any SWIZZLE_NIL terms from given swizzle mask.
250 * For a swizzle like .z??? generate .zzzz (replicate single component).
251 * Else, for .wx?? generate .wxzw (insert default component for the position).
252 */
253 static GLuint
254 fix_swizzle(GLuint swizzle)
255 {
256 GLuint c0 = GET_SWZ(swizzle, 0),
257 c1 = GET_SWZ(swizzle, 1),
258 c2 = GET_SWZ(swizzle, 2),
259 c3 = GET_SWZ(swizzle, 3);
260 if (c1 == SWIZZLE_NIL && c2 == SWIZZLE_NIL && c3 == SWIZZLE_NIL) {
261 /* smear first component across all positions */
262 c1 = c2 = c3 = c0;
263 }
264 else {
265 /* insert default swizzle components */
266 if (c0 == SWIZZLE_NIL)
267 c0 = SWIZZLE_X;
268 if (c1 == SWIZZLE_NIL)
269 c1 = SWIZZLE_Y;
270 if (c2 == SWIZZLE_NIL)
271 c2 = SWIZZLE_Z;
272 if (c3 == SWIZZLE_NIL)
273 c3 = SWIZZLE_W;
274 }
275 return MAKE_SWIZZLE4(c0, c1, c2, c3);
276 }
277
278
279
280 /**
281 * Convert IR storage to an instruction dst register.
282 */
283 static void
284 storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st)
285 {
286 const GLboolean relAddr = st->RelAddr;
287 const GLint size = st->Size;
288 GLint index = st->Index;
289 GLuint swizzle = st->Swizzle;
290
291 assert(index >= 0);
292 /* if this is storage relative to some parent storage, walk up the tree */
293 while (st->Parent) {
294 st = st->Parent;
295 assert(st->Index >= 0);
296 index += st->Index;
297 swizzle = _slang_swizzle_swizzle(st->Swizzle, swizzle);
298 }
299
300 assert(st->File != PROGRAM_UNDEFINED);
301 dst->File = st->File;
302
303 assert(index >= 0);
304 dst->Index = index;
305
306 assert(size >= 1);
307 assert(size <= 4);
308
309 if (swizzle != SWIZZLE_XYZW) {
310 dst->WriteMask = swizzle_to_writemask(swizzle);
311 }
312 else {
313 switch (size) {
314 case 1:
315 dst->WriteMask = WRITEMASK_X << GET_SWZ(st->Swizzle, 0);
316 break;
317 case 2:
318 dst->WriteMask = WRITEMASK_XY;
319 break;
320 case 3:
321 dst->WriteMask = WRITEMASK_XYZ;
322 break;
323 case 4:
324 dst->WriteMask = WRITEMASK_XYZW;
325 break;
326 default:
327 ; /* error would have been caught above */
328 }
329 }
330
331 dst->RelAddr = relAddr;
332 }
333
334
335 /**
336 * Convert IR storage to an instruction src register.
337 */
338 static void
339 storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st)
340 {
341 const GLboolean relAddr = st->RelAddr;
342 GLint index = st->Index;
343 GLuint swizzle = st->Swizzle;
344
345 /* if this is storage relative to some parent storage, walk up the tree */
346 assert(index >= 0);
347 while (st->Parent) {
348 st = st->Parent;
349 if (st->Index < 0) {
350 /* an error should have been reported already */
351 return;
352 }
353 assert(st->Index >= 0);
354 index += st->Index;
355 swizzle = _slang_swizzle_swizzle(fix_swizzle(st->Swizzle), swizzle);
356 }
357
358 assert(st->File >= 0);
359 #if 1 /* XXX temporary */
360 if (st->File == PROGRAM_UNDEFINED) {
361 slang_ir_storage *st0 = (slang_ir_storage *) st;
362 st0->File = PROGRAM_TEMPORARY;
363 }
364 #endif
365 assert(st->File < PROGRAM_UNDEFINED);
366 src->File = st->File;
367
368 assert(index >= 0);
369 src->Index = index;
370
371 swizzle = fix_swizzle(swizzle);
372 assert(GET_SWZ(swizzle, 0) <= SWIZZLE_W);
373 assert(GET_SWZ(swizzle, 1) <= SWIZZLE_W);
374 assert(GET_SWZ(swizzle, 2) <= SWIZZLE_W);
375 assert(GET_SWZ(swizzle, 3) <= SWIZZLE_W);
376 src->Swizzle = swizzle;
377
378 src->RelAddr = relAddr;
379 }
380
381
382 /*
383 * Setup storage pointing to a scalar constant/literal.
384 */
385 static void
386 constant_to_storage(slang_emit_info *emitInfo,
387 GLfloat val,
388 slang_ir_storage *store)
389 {
390 GLuint swizzle;
391 GLint reg;
392 GLfloat value[4];
393
394 value[0] = val;
395 reg = _mesa_add_unnamed_constant(emitInfo->prog->Parameters,
396 value, 1, &swizzle);
397
398 memset(store, 0, sizeof(*store));
399 store->File = PROGRAM_CONSTANT;
400 store->Index = reg;
401 store->Swizzle = swizzle;
402 }
403
404
405 /**
406 * Add new instruction at end of given program.
407 * \param prog the program to append instruction onto
408 * \param opcode opcode for the new instruction
409 * \return pointer to the new instruction
410 */
411 static struct prog_instruction *
412 new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode)
413 {
414 struct gl_program *prog = emitInfo->prog;
415 struct prog_instruction *inst;
416
417 #if 0
418 /* print prev inst */
419 if (prog->NumInstructions > 0) {
420 _mesa_print_instruction(prog->Instructions + prog->NumInstructions - 1);
421 }
422 #endif
423 assert(prog->NumInstructions <= emitInfo->MaxInstructions);
424
425 if (prog->NumInstructions == emitInfo->MaxInstructions) {
426 /* grow the instruction buffer */
427 emitInfo->MaxInstructions += 20;
428 prog->Instructions =
429 _mesa_realloc_instructions(prog->Instructions,
430 prog->NumInstructions,
431 emitInfo->MaxInstructions);
432 }
433
434 inst = prog->Instructions + prog->NumInstructions;
435 prog->NumInstructions++;
436 _mesa_init_instructions(inst, 1);
437 inst->Opcode = opcode;
438 inst->BranchTarget = -1; /* invalid */
439 /*
440 printf("New inst %d: %p %s\n", prog->NumInstructions-1,(void*)inst,
441 _mesa_opcode_string(inst->Opcode));
442 */
443 return inst;
444 }
445
446
447 static struct prog_instruction *
448 emit_arl_load(slang_emit_info *emitInfo,
449 enum register_file file, GLint index, GLuint swizzle)
450 {
451 struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_ARL);
452 inst->SrcReg[0].File = file;
453 inst->SrcReg[0].Index = index;
454 inst->SrcReg[0].Swizzle = swizzle;
455 inst->DstReg.File = PROGRAM_ADDRESS;
456 inst->DstReg.Index = 0;
457 inst->DstReg.WriteMask = WRITEMASK_X;
458 return inst;
459 }
460
461
462 /**
463 * Emit a new instruction with given opcode, operands.
464 * At this point the instruction may have multiple indirect register
465 * loads/stores. We convert those into ARL loads and address-relative
466 * operands. See comments inside.
467 * At some point in the future we could directly emit indirectly addressed
468 * registers in Mesa GPU instructions.
469 */
470 static struct prog_instruction *
471 emit_instruction(slang_emit_info *emitInfo,
472 gl_inst_opcode opcode,
473 const slang_ir_storage *dst,
474 const slang_ir_storage *src0,
475 const slang_ir_storage *src1,
476 const slang_ir_storage *src2)
477 {
478 struct prog_instruction *inst;
479 GLuint numIndirect = 0;
480 const slang_ir_storage *src[3];
481 slang_ir_storage newSrc[3], newDst;
482 GLuint i;
483 GLboolean isTemp[3];
484
485 isTemp[0] = isTemp[1] = isTemp[2] = GL_FALSE;
486
487 src[0] = src0;
488 src[1] = src1;
489 src[2] = src2;
490
491 /* count up how many operands are indirect loads */
492 for (i = 0; i < 3; i++) {
493 if (src[i] && src[i]->IsIndirect)
494 numIndirect++;
495 }
496 if (dst && dst->IsIndirect)
497 numIndirect++;
498
499 /* Take special steps for indirect register loads.
500 * If we had multiple address registers this would be simpler.
501 * For example, this GLSL code:
502 * x[i] = y[j] + z[k];
503 * would translate into something like:
504 * ARL ADDR.x, i;
505 * ARL ADDR.y, j;
506 * ARL ADDR.z, k;
507 * ADD TEMP[ADDR.x+5], TEMP[ADDR.y+9], TEMP[ADDR.z+4];
508 * But since we currently only have one address register we have to do this:
509 * ARL ADDR.x, i;
510 * MOV t1, TEMP[ADDR.x+9];
511 * ARL ADDR.x, j;
512 * MOV t2, TEMP[ADDR.x+4];
513 * ARL ADDR.x, k;
514 * ADD TEMP[ADDR.x+5], t1, t2;
515 * The code here figures this out...
516 */
517 if (numIndirect > 0) {
518 for (i = 0; i < 3; i++) {
519 if (src[i] && src[i]->IsIndirect) {
520 /* load the ARL register with the indirect register */
521 emit_arl_load(emitInfo,
522 src[i]->IndirectFile,
523 src[i]->IndirectIndex,
524 src[i]->IndirectSwizzle);
525
526 if (numIndirect > 1) {
527 /* Need to load src[i] into a temporary register */
528 slang_ir_storage srcRelAddr;
529 alloc_local_temp(emitInfo, &newSrc[i], src[i]->Size);
530 isTemp[i] = GL_TRUE;
531
532 /* set RelAddr flag on src register */
533 srcRelAddr = *src[i];
534 srcRelAddr.RelAddr = GL_TRUE;
535 srcRelAddr.IsIndirect = GL_FALSE; /* not really needed */
536
537 /* MOV newSrc, srcRelAddr; */
538 inst = emit_instruction(emitInfo,
539 OPCODE_MOV,
540 &newSrc[i],
541 &srcRelAddr,
542 NULL,
543 NULL);
544
545 src[i] = &newSrc[i];
546 }
547 else {
548 /* just rewrite the src[i] storage to be ARL-relative */
549 newSrc[i] = *src[i];
550 newSrc[i].RelAddr = GL_TRUE;
551 newSrc[i].IsIndirect = GL_FALSE; /* not really needed */
552 src[i] = &newSrc[i];
553 }
554 }
555 }
556 }
557
558 /* Take special steps for indirect dest register write */
559 if (dst && dst->IsIndirect) {
560 /* load the ARL register with the indirect register */
561 emit_arl_load(emitInfo,
562 dst->IndirectFile,
563 dst->IndirectIndex,
564 dst->IndirectSwizzle);
565 newDst = *dst;
566 newDst.RelAddr = GL_TRUE;
567 newDst.IsIndirect = GL_FALSE;
568 dst = &newDst;
569 }
570
571 /* OK, emit the instruction and its dst, src regs */
572 inst = new_instruction(emitInfo, opcode);
573 if (!inst)
574 return NULL;
575
576 if (dst)
577 storage_to_dst_reg(&inst->DstReg, dst);
578
579 for (i = 0; i < 3; i++) {
580 if (src[i])
581 storage_to_src_reg(&inst->SrcReg[i], src[i]);
582 }
583
584 /* Free any temp registers that we allocated above */
585 for (i = 0; i < 3; i++) {
586 if (isTemp[i])
587 _slang_free_temp(emitInfo->vt, &newSrc[i]);
588 }
589
590 return inst;
591 }
592
593
594
595 /**
596 * Put a comment on the given instruction.
597 */
598 static void
599 inst_comment(struct prog_instruction *inst, const char *comment)
600 {
601 if (inst)
602 inst->Comment = _mesa_strdup(comment);
603 }
604
605
606
607 /**
608 * Return pointer to last instruction in program.
609 */
610 static struct prog_instruction *
611 prev_instruction(slang_emit_info *emitInfo)
612 {
613 struct gl_program *prog = emitInfo->prog;
614 if (prog->NumInstructions == 0)
615 return NULL;
616 else
617 return prog->Instructions + prog->NumInstructions - 1;
618 }
619
620
621 static struct prog_instruction *
622 emit(slang_emit_info *emitInfo, slang_ir_node *n);
623
624
625 /**
626 * Return an annotation string for given node's storage.
627 */
628 static char *
629 storage_annotation(const slang_ir_node *n, const struct gl_program *prog)
630 {
631 #if ANNOTATE
632 const slang_ir_storage *st = n->Store;
633 static char s[100] = "";
634
635 if (!st)
636 return _mesa_strdup("");
637
638 switch (st->File) {
639 case PROGRAM_CONSTANT:
640 if (st->Index >= 0) {
641 const GLfloat *val = prog->Parameters->ParameterValues[st->Index];
642 if (st->Swizzle == SWIZZLE_NOOP)
643 sprintf(s, "{%g, %g, %g, %g}", val[0], val[1], val[2], val[3]);
644 else {
645 sprintf(s, "%g", val[GET_SWZ(st->Swizzle, 0)]);
646 }
647 }
648 break;
649 case PROGRAM_TEMPORARY:
650 if (n->Var)
651 sprintf(s, "%s", (char *) n->Var->a_name);
652 else
653 sprintf(s, "t[%d]", st->Index);
654 break;
655 case PROGRAM_STATE_VAR:
656 case PROGRAM_UNIFORM:
657 sprintf(s, "%s", prog->Parameters->Parameters[st->Index].Name);
658 break;
659 case PROGRAM_VARYING:
660 sprintf(s, "%s", prog->Varying->Parameters[st->Index].Name);
661 break;
662 case PROGRAM_INPUT:
663 sprintf(s, "input[%d]", st->Index);
664 break;
665 case PROGRAM_OUTPUT:
666 sprintf(s, "output[%d]", st->Index);
667 break;
668 default:
669 s[0] = 0;
670 }
671 return _mesa_strdup(s);
672 #else
673 return NULL;
674 #endif
675 }
676
677
678 /**
679 * Return an annotation string for an instruction.
680 */
681 static char *
682 instruction_annotation(gl_inst_opcode opcode, char *dstAnnot,
683 char *srcAnnot0, char *srcAnnot1, char *srcAnnot2)
684 {
685 #if ANNOTATE
686 const char *operator;
687 char *s;
688 int len = 50;
689
690 if (dstAnnot)
691 len += strlen(dstAnnot);
692 else
693 dstAnnot = _mesa_strdup("");
694
695 if (srcAnnot0)
696 len += strlen(srcAnnot0);
697 else
698 srcAnnot0 = _mesa_strdup("");
699
700 if (srcAnnot1)
701 len += strlen(srcAnnot1);
702 else
703 srcAnnot1 = _mesa_strdup("");
704
705 if (srcAnnot2)
706 len += strlen(srcAnnot2);
707 else
708 srcAnnot2 = _mesa_strdup("");
709
710 switch (opcode) {
711 case OPCODE_ADD:
712 operator = "+";
713 break;
714 case OPCODE_SUB:
715 operator = "-";
716 break;
717 case OPCODE_MUL:
718 operator = "*";
719 break;
720 case OPCODE_DP2:
721 operator = "DP2";
722 break;
723 case OPCODE_DP3:
724 operator = "DP3";
725 break;
726 case OPCODE_DP4:
727 operator = "DP4";
728 break;
729 case OPCODE_XPD:
730 operator = "XPD";
731 break;
732 case OPCODE_RSQ:
733 operator = "RSQ";
734 break;
735 case OPCODE_SGT:
736 operator = ">";
737 break;
738 default:
739 operator = ",";
740 }
741
742 s = (char *) malloc(len);
743 sprintf(s, "%s = %s %s %s %s", dstAnnot,
744 srcAnnot0, operator, srcAnnot1, srcAnnot2);
745 assert(_mesa_strlen(s) < len);
746
747 free(dstAnnot);
748 free(srcAnnot0);
749 free(srcAnnot1);
750 free(srcAnnot2);
751
752 return s;
753 #else
754 return NULL;
755 #endif
756 }
757
758
759 /**
760 * Emit an instruction that's just a comment.
761 */
762 static struct prog_instruction *
763 emit_comment(slang_emit_info *emitInfo, const char *comment)
764 {
765 struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_NOP);
766 inst_comment(inst, comment);
767 return inst;
768 }
769
770
771 /**
772 * Generate code for a simple arithmetic instruction.
773 * Either 1, 2 or 3 operands.
774 */
775 static struct prog_instruction *
776 emit_arith(slang_emit_info *emitInfo, slang_ir_node *n)
777 {
778 const slang_ir_info *info = _slang_ir_info(n->Opcode);
779 struct prog_instruction *inst;
780 GLuint i;
781
782 assert(info);
783 assert(info->InstOpcode != OPCODE_NOP);
784
785 #if PEEPHOLE_OPTIMIZATIONS
786 /* Look for MAD opportunity */
787 if (info->NumParams == 2 &&
788 n->Opcode == IR_ADD && n->Children[0]->Opcode == IR_MUL) {
789 /* found pattern IR_ADD(IR_MUL(A, B), C) */
790 emit(emitInfo, n->Children[0]->Children[0]); /* A */
791 emit(emitInfo, n->Children[0]->Children[1]); /* B */
792 emit(emitInfo, n->Children[1]); /* C */
793 alloc_node_storage(emitInfo, n, -1); /* dest */
794
795 inst = emit_instruction(emitInfo,
796 OPCODE_MAD,
797 n->Store,
798 n->Children[0]->Children[0]->Store,
799 n->Children[0]->Children[1]->Store,
800 n->Children[1]->Store);
801
802 free_node_storage(emitInfo->vt, n->Children[0]->Children[0]);
803 free_node_storage(emitInfo->vt, n->Children[0]->Children[1]);
804 free_node_storage(emitInfo->vt, n->Children[1]);
805 return inst;
806 }
807
808 if (info->NumParams == 2 &&
809 n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) {
810 /* found pattern IR_ADD(A, IR_MUL(B, C)) */
811 emit(emitInfo, n->Children[0]); /* A */
812 emit(emitInfo, n->Children[1]->Children[0]); /* B */
813 emit(emitInfo, n->Children[1]->Children[1]); /* C */
814 alloc_node_storage(emitInfo, n, -1); /* dest */
815
816 inst = emit_instruction(emitInfo,
817 OPCODE_MAD,
818 n->Store,
819 n->Children[1]->Children[0]->Store,
820 n->Children[1]->Children[1]->Store,
821 n->Children[0]->Store);
822
823 free_node_storage(emitInfo->vt, n->Children[1]->Children[0]);
824 free_node_storage(emitInfo->vt, n->Children[1]->Children[1]);
825 free_node_storage(emitInfo->vt, n->Children[0]);
826 return inst;
827 }
828 #endif
829
830 /* gen code for children, may involve temp allocation */
831 for (i = 0; i < info->NumParams; i++) {
832 emit(emitInfo, n->Children[i]);
833 if (!n->Children[i] || !n->Children[i]->Store) {
834 /* error recovery */
835 return NULL;
836 }
837 }
838
839 /* result storage */
840 alloc_node_storage(emitInfo, n, -1);
841
842 inst = emit_instruction(emitInfo,
843 info->InstOpcode,
844 n->Store, /* dest */
845 (info->NumParams > 0 ? n->Children[0]->Store : NULL),
846 (info->NumParams > 1 ? n->Children[1]->Store : NULL),
847 (info->NumParams > 2 ? n->Children[2]->Store : NULL)
848 );
849
850 /* free temps */
851 for (i = 0; i < info->NumParams; i++)
852 free_node_storage(emitInfo->vt, n->Children[i]);
853
854 return inst;
855 }
856
857
858 /**
859 * Emit code for == and != operators. These could normally be handled
860 * by emit_arith() except we need to be able to handle structure comparisons.
861 */
862 static struct prog_instruction *
863 emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
864 {
865 struct prog_instruction *inst = NULL;
866 GLint size;
867
868 assert(n->Opcode == IR_EQUAL || n->Opcode == IR_NOTEQUAL);
869
870 /* gen code for children */
871 emit(emitInfo, n->Children[0]);
872 emit(emitInfo, n->Children[1]);
873
874 if (n->Children[0]->Store->Size != n->Children[1]->Store->Size) {
875 slang_info_log_error(emitInfo->log, "invalid operands to == or !=");
876 return NULL;
877 }
878
879 /* final result is 1 bool */
880 if (!alloc_node_storage(emitInfo, n, 1))
881 return NULL;
882
883 size = n->Children[0]->Store->Size;
884
885 if (size == 1) {
886 gl_inst_opcode opcode = n->Opcode == IR_EQUAL ? OPCODE_SEQ : OPCODE_SNE;
887 inst = emit_instruction(emitInfo,
888 opcode,
889 n->Store, /* dest */
890 n->Children[0]->Store,
891 n->Children[1]->Store,
892 NULL);
893 }
894 else if (size <= 4) {
895 /* compare two vectors.
896 * Unfortunately, there's no instruction to compare vectors and
897 * return a scalar result. Do it with some compare and dot product
898 * instructions...
899 */
900 GLuint swizzle;
901 gl_inst_opcode dotOp;
902 slang_ir_storage tempStore;
903
904 if (!alloc_local_temp(emitInfo, &tempStore, 4)) {
905 return NULL;
906 /* out of temps */
907 }
908
909 if (size == 4) {
910 dotOp = OPCODE_DP4;
911 swizzle = SWIZZLE_XYZW;
912 }
913 else if (size == 3) {
914 dotOp = OPCODE_DP3;
915 swizzle = SWIZZLE_XYZW;
916 }
917 else {
918 assert(size == 2);
919 dotOp = OPCODE_DP3; /* XXX use OPCODE_DP2 eventually */
920 swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y);
921 }
922
923 /* Compute inequality (temp = (A != B)) */
924 inst = emit_instruction(emitInfo,
925 OPCODE_SNE,
926 &tempStore,
927 n->Children[0]->Store,
928 n->Children[1]->Store,
929 NULL);
930 inst_comment(inst, "Compare values");
931
932 /* Compute val = DOT(temp, temp) (reduction) */
933 inst = emit_instruction(emitInfo,
934 dotOp,
935 n->Store,
936 &tempStore,
937 &tempStore,
938 NULL);
939 inst->SrcReg[0].Swizzle = inst->SrcReg[1].Swizzle = swizzle; /*override*/
940 inst_comment(inst, "Reduce vec to bool");
941
942 _slang_free_temp(emitInfo->vt, &tempStore); /* free temp */
943
944 if (n->Opcode == IR_EQUAL) {
945 /* compute val = !val.x with SEQ val, val, 0; */
946 slang_ir_storage zero;
947 constant_to_storage(emitInfo, 0.0, &zero);
948 inst = emit_instruction(emitInfo,
949 OPCODE_SEQ,
950 n->Store, /* dest */
951 n->Store,
952 &zero,
953 NULL);
954 inst_comment(inst, "Invert true/false");
955 }
956 }
957 else {
958 /* size > 4, struct or array compare.
959 * XXX this won't work reliably for structs with padding!!
960 */
961 GLint i, num = (n->Children[0]->Store->Size + 3) / 4;
962 slang_ir_storage accTemp, sneTemp;
963
964 if (!alloc_local_temp(emitInfo, &accTemp, 4))
965 return NULL;
966
967 if (!alloc_local_temp(emitInfo, &sneTemp, 4))
968 return NULL;
969
970 for (i = 0; i < num; i++) {
971 slang_ir_storage srcStore0 = *n->Children[0]->Store;
972 slang_ir_storage srcStore1 = *n->Children[1]->Store;
973 srcStore0.Index += i;
974 srcStore1.Index += i;
975
976 if (i == 0) {
977 /* SNE accTemp, left[i], right[i] */
978 inst = emit_instruction(emitInfo, OPCODE_SNE,
979 &accTemp, /* dest */
980 &srcStore0,
981 &srcStore1,
982 NULL);
983 inst_comment(inst, "Begin struct/array comparison");
984 }
985 else {
986 /* SNE sneTemp, left[i], right[i] */
987 inst = emit_instruction(emitInfo, OPCODE_SNE,
988 &sneTemp, /* dest */
989 &srcStore0,
990 &srcStore1,
991 NULL);
992 /* ADD accTemp, accTemp, sneTemp; # like logical-OR */
993 inst = emit_instruction(emitInfo, OPCODE_ADD,
994 &accTemp, /* dest */
995 &accTemp,
996 &sneTemp,
997 NULL);
998 }
999 }
1000
1001 /* compute accTemp.x || accTemp.y || accTemp.z || accTemp.w with DOT4 */
1002 inst = emit_instruction(emitInfo, OPCODE_DP4,
1003 n->Store,
1004 &accTemp,
1005 &accTemp,
1006 NULL);
1007 inst_comment(inst, "End struct/array comparison");
1008
1009 if (n->Opcode == IR_EQUAL) {
1010 /* compute tmp.x = !tmp.x via tmp.x = (tmp.x == 0) */
1011 slang_ir_storage zero;
1012 constant_to_storage(emitInfo, 0.0, &zero);
1013 inst = emit_instruction(emitInfo, OPCODE_SEQ,
1014 n->Store, /* dest */
1015 n->Store,
1016 &zero,
1017 NULL);
1018 inst_comment(inst, "Invert true/false");
1019 }
1020
1021 _slang_free_temp(emitInfo->vt, &accTemp);
1022 _slang_free_temp(emitInfo->vt, &sneTemp);
1023 }
1024
1025 /* free temps */
1026 free_node_storage(emitInfo->vt, n->Children[0]);
1027 free_node_storage(emitInfo->vt, n->Children[1]);
1028
1029 return inst;
1030 }
1031
1032
1033
1034 /**
1035 * Generate code for an IR_CLAMP instruction.
1036 */
1037 static struct prog_instruction *
1038 emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n)
1039 {
1040 struct prog_instruction *inst;
1041 slang_ir_node tmpNode;
1042
1043 assert(n->Opcode == IR_CLAMP);
1044 /* ch[0] = value
1045 * ch[1] = min limit
1046 * ch[2] = max limit
1047 */
1048
1049 inst = emit(emitInfo, n->Children[0]);
1050
1051 /* If lower limit == 0.0 and upper limit == 1.0,
1052 * set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
1053 * Else,
1054 * emit OPCODE_MIN, OPCODE_MAX sequence.
1055 */
1056 #if 0
1057 /* XXX this isn't quite finished yet */
1058 if (n->Children[1]->Opcode == IR_FLOAT &&
1059 n->Children[1]->Value[0] == 0.0 &&
1060 n->Children[1]->Value[1] == 0.0 &&
1061 n->Children[1]->Value[2] == 0.0 &&
1062 n->Children[1]->Value[3] == 0.0 &&
1063 n->Children[2]->Opcode == IR_FLOAT &&
1064 n->Children[2]->Value[0] == 1.0 &&
1065 n->Children[2]->Value[1] == 1.0 &&
1066 n->Children[2]->Value[2] == 1.0 &&
1067 n->Children[2]->Value[3] == 1.0) {
1068 if (!inst) {
1069 inst = prev_instruction(prog);
1070 }
1071 if (inst && inst->Opcode != OPCODE_NOP) {
1072 /* and prev instruction's DstReg matches n->Children[0]->Store */
1073 inst->SaturateMode = SATURATE_ZERO_ONE;
1074 n->Store = n->Children[0]->Store;
1075 return inst;
1076 }
1077 }
1078 #endif
1079
1080 if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size))
1081 return NULL;
1082
1083 emit(emitInfo, n->Children[1]);
1084 emit(emitInfo, n->Children[2]);
1085
1086 /* Some GPUs don't allow reading from output registers. So if the
1087 * dest for this clamp() is an output reg, we can't use that reg for
1088 * the intermediate result. Use a temp register instead.
1089 */
1090 _mesa_bzero(&tmpNode, sizeof(tmpNode));
1091 alloc_node_storage(emitInfo, &tmpNode, n->Store->Size);
1092
1093 /* tmp = max(ch[0], ch[1]) */
1094 inst = emit_instruction(emitInfo, OPCODE_MAX,
1095 tmpNode.Store, /* dest */
1096 n->Children[0]->Store,
1097 n->Children[1]->Store,
1098 NULL);
1099
1100 /* n->dest = min(tmp, ch[2]) */
1101 inst = emit_instruction(emitInfo, OPCODE_MIN,
1102 n->Store, /* dest */
1103 tmpNode.Store,
1104 n->Children[2]->Store,
1105 NULL);
1106
1107 free_node_storage(emitInfo->vt, &tmpNode);
1108
1109 return inst;
1110 }
1111
1112
1113 static struct prog_instruction *
1114 emit_negation(slang_emit_info *emitInfo, slang_ir_node *n)
1115 {
1116 /* Implement as MOV dst, -src; */
1117 /* XXX we could look at the previous instruction and in some circumstances
1118 * modify it to accomplish the negation.
1119 */
1120 struct prog_instruction *inst;
1121
1122 emit(emitInfo, n->Children[0]);
1123
1124 if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size))
1125 return NULL;
1126
1127 inst = emit_instruction(emitInfo,
1128 OPCODE_MOV,
1129 n->Store, /* dest */
1130 n->Children[0]->Store,
1131 NULL,
1132 NULL);
1133 inst->SrcReg[0].NegateBase = NEGATE_XYZW;
1134 return inst;
1135 }
1136
1137
1138 static struct prog_instruction *
1139 emit_label(slang_emit_info *emitInfo, const slang_ir_node *n)
1140 {
1141 assert(n->Label);
1142 #if 0
1143 /* XXX this fails in loop tail code - investigate someday */
1144 assert(_slang_label_get_location(n->Label) < 0);
1145 _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
1146 emitInfo->prog);
1147 #else
1148 if (_slang_label_get_location(n->Label) < 0)
1149 _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
1150 emitInfo->prog);
1151 #endif
1152 return NULL;
1153 }
1154
1155
1156 /**
1157 * Emit code for a function call.
1158 * Note that for each time a function is called, we emit the function's
1159 * body code again because the set of available registers may be different.
1160 */
1161 static struct prog_instruction *
1162 emit_fcall(slang_emit_info *emitInfo, slang_ir_node *n)
1163 {
1164 struct gl_program *progSave;
1165 struct prog_instruction *inst;
1166 GLuint subroutineId;
1167 GLuint maxInstSave;
1168
1169 assert(n->Opcode == IR_CALL);
1170 assert(n->Label);
1171
1172 /* save/push cur program */
1173 maxInstSave = emitInfo->MaxInstructions;
1174 progSave = emitInfo->prog;
1175
1176 emitInfo->prog = new_subroutine(emitInfo, &subroutineId);
1177 emitInfo->MaxInstructions = emitInfo->prog->NumInstructions;
1178
1179 _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
1180 emitInfo->prog);
1181
1182 if (emitInfo->EmitBeginEndSub) {
1183 /* BGNSUB isn't a real instruction.
1184 * We require a label (i.e. "foobar:") though, if we're going to
1185 * print the program in the NV format. The BNGSUB instruction is
1186 * really just a NOP to attach the label to.
1187 */
1188 inst = new_instruction(emitInfo, OPCODE_BGNSUB);
1189 inst_comment(inst, n->Label->Name);
1190 }
1191
1192 /* body of function: */
1193 emit(emitInfo, n->Children[0]);
1194 n->Store = n->Children[0]->Store;
1195
1196 /* add RET instruction now, if needed */
1197 inst = prev_instruction(emitInfo);
1198 if (inst && inst->Opcode != OPCODE_RET) {
1199 inst = new_instruction(emitInfo, OPCODE_RET);
1200 }
1201
1202 if (emitInfo->EmitBeginEndSub) {
1203 inst = new_instruction(emitInfo, OPCODE_ENDSUB);
1204 inst_comment(inst, n->Label->Name);
1205 }
1206
1207 /* pop/restore cur program */
1208 emitInfo->prog = progSave;
1209 emitInfo->MaxInstructions = maxInstSave;
1210
1211 /* emit the function call */
1212 inst = new_instruction(emitInfo, OPCODE_CAL);
1213 /* The branch target is just the subroutine number (changed later) */
1214 inst->BranchTarget = subroutineId;
1215 inst_comment(inst, n->Label->Name);
1216 assert(inst->BranchTarget >= 0);
1217
1218 return inst;
1219 }
1220
1221
1222 /**
1223 * Emit code for a 'return' statement.
1224 */
1225 static struct prog_instruction *
1226 emit_return(slang_emit_info *emitInfo, slang_ir_node *n)
1227 {
1228 struct prog_instruction *inst;
1229 assert(n);
1230 assert(n->Opcode == IR_RETURN);
1231 assert(n->Label);
1232 inst = new_instruction(emitInfo, OPCODE_RET);
1233 inst->DstReg.CondMask = COND_TR; /* always return */
1234 return inst;
1235 }
1236
1237
1238 static struct prog_instruction *
1239 emit_kill(slang_emit_info *emitInfo)
1240 {
1241 struct gl_fragment_program *fp;
1242 struct prog_instruction *inst;
1243 /* NV-KILL - discard fragment depending on condition code.
1244 * Note that ARB-KILL depends on sign of vector operand.
1245 */
1246 inst = new_instruction(emitInfo, OPCODE_KIL_NV);
1247 inst->DstReg.CondMask = COND_TR; /* always kill */
1248
1249 assert(emitInfo->prog->Target == GL_FRAGMENT_PROGRAM_ARB);
1250 fp = (struct gl_fragment_program *) emitInfo->prog;
1251 fp->UsesKill = GL_TRUE;
1252
1253 return inst;
1254 }
1255
1256
1257 static struct prog_instruction *
1258 emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
1259 {
1260 struct prog_instruction *inst;
1261 gl_inst_opcode opcode;
1262
1263 if (n->Opcode == IR_TEX) {
1264 opcode = OPCODE_TEX;
1265 }
1266 else if (n->Opcode == IR_TEXB) {
1267 opcode = OPCODE_TXB;
1268 }
1269 else {
1270 assert(n->Opcode == IR_TEXP);
1271 opcode = OPCODE_TXP;
1272 }
1273
1274 if (n->Children[0]->Opcode == IR_ELEMENT) {
1275 /* array is the sampler (a uniform which'll indicate the texture unit) */
1276 assert(n->Children[0]->Children[0]->Store);
1277 assert(n->Children[0]->Children[0]->Store->File == PROGRAM_SAMPLER);
1278
1279 emit(emitInfo, n->Children[0]);
1280
1281 n->Children[0]->Var = n->Children[0]->Children[0]->Var;
1282 } else {
1283 /* this is the sampler (a uniform which'll indicate the texture unit) */
1284 assert(n->Children[0]->Store);
1285 assert(n->Children[0]->Store->File == PROGRAM_SAMPLER);
1286 }
1287
1288 /* emit code for the texcoord operand */
1289 (void) emit(emitInfo, n->Children[1]);
1290
1291 /* alloc storage for result of texture fetch */
1292 if (!alloc_node_storage(emitInfo, n, 4))
1293 return NULL;
1294
1295 /* emit TEX instruction; Child[1] is the texcoord */
1296 inst = emit_instruction(emitInfo,
1297 opcode,
1298 n->Store,
1299 n->Children[1]->Store,
1300 NULL,
1301 NULL);
1302
1303 /* Store->Index is the uniform/sampler index */
1304 assert(n->Children[0]->Store->Index >= 0);
1305 inst->TexSrcUnit = n->Children[0]->Store->Index;
1306 inst->TexSrcTarget = n->Children[0]->Store->TexTarget;
1307
1308 /* mark the sampler as being used */
1309 _mesa_use_uniform(emitInfo->prog->Parameters,
1310 (char *) n->Children[0]->Var->a_name);
1311
1312 return inst;
1313 }
1314
1315
1316 /**
1317 * Assignment/copy
1318 */
1319 static struct prog_instruction *
1320 emit_copy(slang_emit_info *emitInfo, slang_ir_node *n)
1321 {
1322 struct prog_instruction *inst;
1323
1324 assert(n->Opcode == IR_COPY);
1325
1326 /* lhs */
1327 emit(emitInfo, n->Children[0]);
1328 if (!n->Children[0]->Store || n->Children[0]->Store->Index < 0) {
1329 /* an error should have been already recorded */
1330 return NULL;
1331 }
1332
1333 /* rhs */
1334 assert(n->Children[1]);
1335 inst = emit(emitInfo, n->Children[1]);
1336
1337 if (!n->Children[1]->Store || n->Children[1]->Store->Index < 0) {
1338 if (!emitInfo->log->text) {
1339 slang_info_log_error(emitInfo->log, "invalid assignment");
1340 }
1341 return NULL;
1342 }
1343
1344 assert(n->Children[1]->Store->Index >= 0);
1345
1346 /*assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size);*/
1347
1348 n->Store = n->Children[0]->Store;
1349
1350 if (n->Store->File == PROGRAM_SAMPLER) {
1351 /* no code generated for sampler assignments,
1352 * just copy the sampler index at compile time.
1353 */
1354 n->Store->Index = n->Children[1]->Store->Index;
1355 return NULL;
1356 }
1357
1358 #if PEEPHOLE_OPTIMIZATIONS
1359 if (inst &&
1360 _slang_is_temp(emitInfo->vt, n->Children[1]->Store) &&
1361 (inst->DstReg.File == n->Children[1]->Store->File) &&
1362 (inst->DstReg.Index == n->Children[1]->Store->Index) &&
1363 !n->Children[0]->Store->IsIndirect &&
1364 n->Children[0]->Store->Size <= 4) {
1365 /* Peephole optimization:
1366 * The Right-Hand-Side has its results in a temporary place.
1367 * Modify the RHS (and the prev instruction) to store its results
1368 * in the destination specified by n->Children[0].
1369 * Then, this MOVE is a no-op.
1370 * Ex:
1371 * MUL tmp, x, y;
1372 * MOV a, tmp;
1373 * becomes:
1374 * MUL a, x, y;
1375 */
1376 if (n->Children[1]->Opcode != IR_SWIZZLE)
1377 _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
1378 *n->Children[1]->Store = *n->Children[0]->Store;
1379
1380 /* fixup the previous instruction (which stored the RHS result) */
1381 assert(n->Children[0]->Store->Index >= 0);
1382
1383 storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store);
1384 return inst;
1385 }
1386 else
1387 #endif
1388 {
1389 if (n->Children[0]->Store->Size > 4) {
1390 /* move matrix/struct etc (block of registers) */
1391 slang_ir_storage dstStore = *n->Children[0]->Store;
1392 slang_ir_storage srcStore = *n->Children[1]->Store;
1393 GLint size = srcStore.Size;
1394 ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP);
1395 dstStore.Size = 4;
1396 srcStore.Size = 4;
1397 while (size >= 4) {
1398 inst = emit_instruction(emitInfo, OPCODE_MOV,
1399 &dstStore,
1400 &srcStore,
1401 NULL,
1402 NULL);
1403 inst_comment(inst, "IR_COPY block");
1404 srcStore.Index++;
1405 dstStore.Index++;
1406 size -= 4;
1407 }
1408 }
1409 else {
1410 /* single register move */
1411 char *srcAnnot, *dstAnnot;
1412 assert(n->Children[0]->Store->Index >= 0);
1413 inst = emit_instruction(emitInfo, OPCODE_MOV,
1414 n->Children[0]->Store, /* dest */
1415 n->Children[1]->Store,
1416 NULL,
1417 NULL);
1418 dstAnnot = storage_annotation(n->Children[0], emitInfo->prog);
1419 srcAnnot = storage_annotation(n->Children[1], emitInfo->prog);
1420 inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
1421 srcAnnot, NULL, NULL);
1422 }
1423 free_node_storage(emitInfo->vt, n->Children[1]);
1424 return inst;
1425 }
1426 }
1427
1428
1429 /**
1430 * An IR_COND node wraps a boolean expression which is used by an
1431 * IF or WHILE test. This is where we'll set condition codes, if needed.
1432 */
1433 static struct prog_instruction *
1434 emit_cond(slang_emit_info *emitInfo, slang_ir_node *n)
1435 {
1436 struct prog_instruction *inst;
1437
1438 assert(n->Opcode == IR_COND);
1439
1440 if (!n->Children[0])
1441 return NULL;
1442
1443 /* emit code for the expression */
1444 inst = emit(emitInfo, n->Children[0]);
1445
1446 if (!n->Children[0]->Store) {
1447 /* error recovery */
1448 return NULL;
1449 }
1450
1451 assert(n->Children[0]->Store);
1452 /*assert(n->Children[0]->Store->Size == 1);*/
1453
1454 if (emitInfo->EmitCondCodes) {
1455 if (inst &&
1456 n->Children[0]->Store &&
1457 inst->DstReg.File == n->Children[0]->Store->File &&
1458 inst->DstReg.Index == n->Children[0]->Store->Index) {
1459 /* The previous instruction wrote to the register who's value
1460 * we're testing. Just fix that instruction so that the
1461 * condition codes are computed.
1462 */
1463 inst->CondUpdate = GL_TRUE;
1464 n->Store = n->Children[0]->Store;
1465 return inst;
1466 }
1467 else {
1468 /* This'll happen for things like "if (i) ..." where no code
1469 * is normally generated for the expression "i".
1470 * Generate a move instruction just to set condition codes.
1471 */
1472 if (!alloc_node_storage(emitInfo, n, 1))
1473 return NULL;
1474 inst = emit_instruction(emitInfo, OPCODE_MOV,
1475 n->Store, /* dest */
1476 n->Children[0]->Store,
1477 NULL,
1478 NULL);
1479 inst->CondUpdate = GL_TRUE;
1480 inst_comment(inst, "COND expr");
1481 _slang_free_temp(emitInfo->vt, n->Store);
1482 return inst;
1483 }
1484 }
1485 else {
1486 /* No-op: the boolean result of the expression is in a regular reg */
1487 n->Store = n->Children[0]->Store;
1488 return inst;
1489 }
1490 }
1491
1492
1493 /**
1494 * Logical-NOT
1495 */
1496 static struct prog_instruction *
1497 emit_not(slang_emit_info *emitInfo, slang_ir_node *n)
1498 {
1499 static const struct {
1500 gl_inst_opcode op, opNot;
1501 } operators[] = {
1502 { OPCODE_SLT, OPCODE_SGE },
1503 { OPCODE_SLE, OPCODE_SGT },
1504 { OPCODE_SGT, OPCODE_SLE },
1505 { OPCODE_SGE, OPCODE_SLT },
1506 { OPCODE_SEQ, OPCODE_SNE },
1507 { OPCODE_SNE, OPCODE_SEQ },
1508 { 0, 0 }
1509 };
1510 struct prog_instruction *inst;
1511 slang_ir_storage zero;
1512 GLuint i;
1513
1514 /* child expr */
1515 inst = emit(emitInfo, n->Children[0]);
1516
1517 #if PEEPHOLE_OPTIMIZATIONS
1518 if (inst) {
1519 /* if the prev instruction was a comparison instruction, invert it */
1520 for (i = 0; operators[i].op; i++) {
1521 if (inst->Opcode == operators[i].op) {
1522 inst->Opcode = operators[i].opNot;
1523 n->Store = n->Children[0]->Store;
1524 return inst;
1525 }
1526 }
1527 }
1528 #endif
1529
1530 /* else, invert using SEQ (v = v == 0) */
1531 if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size))
1532 return NULL;
1533
1534 constant_to_storage(emitInfo, 0.0, &zero);
1535 inst = emit_instruction(emitInfo,
1536 OPCODE_SEQ,
1537 n->Store,
1538 n->Children[0]->Store,
1539 &zero,
1540 NULL);
1541 inst_comment(inst, "NOT");
1542
1543 free_node_storage(emitInfo->vt, n->Children[0]);
1544
1545 return inst;
1546 }
1547
1548
1549 static struct prog_instruction *
1550 emit_if(slang_emit_info *emitInfo, slang_ir_node *n)
1551 {
1552 struct gl_program *prog = emitInfo->prog;
1553 GLuint ifInstLoc, elseInstLoc = 0;
1554 GLuint condWritemask = 0;
1555
1556 /* emit condition expression code */
1557 {
1558 struct prog_instruction *inst;
1559 inst = emit(emitInfo, n->Children[0]);
1560 if (emitInfo->EmitCondCodes) {
1561 if (!inst) {
1562 /* error recovery */
1563 return NULL;
1564 }
1565 condWritemask = inst->DstReg.WriteMask;
1566 }
1567 }
1568
1569 if (!n->Children[0]->Store)
1570 return NULL;
1571
1572 #if 0
1573 assert(n->Children[0]->Store->Size == 1); /* a bool! */
1574 #endif
1575
1576 ifInstLoc = prog->NumInstructions;
1577 if (emitInfo->EmitHighLevelInstructions) {
1578 if (emitInfo->EmitCondCodes) {
1579 /* IF condcode THEN ... */
1580 struct prog_instruction *ifInst;
1581 ifInst = new_instruction(emitInfo, OPCODE_IF);
1582 ifInst->DstReg.CondMask = COND_NE; /* if cond is non-zero */
1583 /* only test the cond code (1 of 4) that was updated by the
1584 * previous instruction.
1585 */
1586 ifInst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
1587 }
1588 else {
1589 /* IF src[0] THEN ... */
1590 emit_instruction(emitInfo, OPCODE_IF,
1591 NULL, /* dst */
1592 n->Children[0]->Store, /* op0 */
1593 NULL,
1594 NULL);
1595 }
1596 }
1597 else {
1598 /* conditional jump to else, or endif */
1599 struct prog_instruction *ifInst = new_instruction(emitInfo, OPCODE_BRA);
1600 ifInst->DstReg.CondMask = COND_EQ; /* BRA if cond is zero */
1601 inst_comment(ifInst, "if zero");
1602 ifInst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
1603 }
1604
1605 /* if body */
1606 emit(emitInfo, n->Children[1]);
1607
1608 if (n->Children[2]) {
1609 /* have else body */
1610 elseInstLoc = prog->NumInstructions;
1611 if (emitInfo->EmitHighLevelInstructions) {
1612 (void) new_instruction(emitInfo, OPCODE_ELSE);
1613 }
1614 else {
1615 /* jump to endif instruction */
1616 struct prog_instruction *inst;
1617 inst = new_instruction(emitInfo, OPCODE_BRA);
1618 inst_comment(inst, "else");
1619 inst->DstReg.CondMask = COND_TR; /* always branch */
1620 }
1621 prog->Instructions[ifInstLoc].BranchTarget = prog->NumInstructions;
1622 emit(emitInfo, n->Children[2]);
1623 }
1624 else {
1625 /* no else body */
1626 prog->Instructions[ifInstLoc].BranchTarget = prog->NumInstructions;
1627 }
1628
1629 if (emitInfo->EmitHighLevelInstructions) {
1630 (void) new_instruction(emitInfo, OPCODE_ENDIF);
1631 }
1632
1633 if (n->Children[2]) {
1634 prog->Instructions[elseInstLoc].BranchTarget = prog->NumInstructions;
1635 }
1636 return NULL;
1637 }
1638
1639
1640 static struct prog_instruction *
1641 emit_loop(slang_emit_info *emitInfo, slang_ir_node *n)
1642 {
1643 struct gl_program *prog = emitInfo->prog;
1644 struct prog_instruction *endInst;
1645 GLuint beginInstLoc, tailInstLoc, endInstLoc;
1646 slang_ir_node *ir;
1647
1648 /* emit OPCODE_BGNLOOP */
1649 beginInstLoc = prog->NumInstructions;
1650 if (emitInfo->EmitHighLevelInstructions) {
1651 (void) new_instruction(emitInfo, OPCODE_BGNLOOP);
1652 }
1653
1654 /* body */
1655 emit(emitInfo, n->Children[0]);
1656
1657 /* tail */
1658 tailInstLoc = prog->NumInstructions;
1659 if (n->Children[1]) {
1660 if (emitInfo->EmitComments)
1661 emit_comment(emitInfo, "Loop tail code:");
1662 emit(emitInfo, n->Children[1]);
1663 }
1664
1665 endInstLoc = prog->NumInstructions;
1666 if (emitInfo->EmitHighLevelInstructions) {
1667 /* emit OPCODE_ENDLOOP */
1668 endInst = new_instruction(emitInfo, OPCODE_ENDLOOP);
1669 }
1670 else {
1671 /* emit unconditional BRA-nch */
1672 endInst = new_instruction(emitInfo, OPCODE_BRA);
1673 endInst->DstReg.CondMask = COND_TR; /* always true */
1674 }
1675 /* ENDLOOP's BranchTarget points to the BGNLOOP inst */
1676 endInst->BranchTarget = beginInstLoc;
1677
1678 if (emitInfo->EmitHighLevelInstructions) {
1679 /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1680 prog->Instructions[beginInstLoc].BranchTarget = prog->NumInstructions -1;
1681 }
1682
1683 /* Done emitting loop code. Now walk over the loop's linked list of
1684 * BREAK and CONT nodes, filling in their BranchTarget fields (which
1685 * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1686 */
1687 for (ir = n->List; ir; ir = ir->List) {
1688 struct prog_instruction *inst = prog->Instructions + ir->InstLocation;
1689 assert(inst->BranchTarget < 0);
1690 if (ir->Opcode == IR_BREAK ||
1691 ir->Opcode == IR_BREAK_IF_TRUE) {
1692 assert(inst->Opcode == OPCODE_BRK ||
1693 inst->Opcode == OPCODE_BRA);
1694 /* go to instruction after end of loop */
1695 inst->BranchTarget = endInstLoc + 1;
1696 }
1697 else {
1698 assert(ir->Opcode == IR_CONT ||
1699 ir->Opcode == IR_CONT_IF_TRUE);
1700 assert(inst->Opcode == OPCODE_CONT ||
1701 inst->Opcode == OPCODE_BRA);
1702 /* go to instruction at tail of loop */
1703 inst->BranchTarget = endInstLoc;
1704 }
1705 }
1706 return NULL;
1707 }
1708
1709
1710 /**
1711 * Unconditional "continue" or "break" statement.
1712 * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1713 */
1714 static struct prog_instruction *
1715 emit_cont_break(slang_emit_info *emitInfo, slang_ir_node *n)
1716 {
1717 gl_inst_opcode opcode;
1718 struct prog_instruction *inst;
1719
1720 if (n->Opcode == IR_CONT) {
1721 /* we need to execute the loop's tail code before doing CONT */
1722 assert(n->Parent);
1723 assert(n->Parent->Opcode == IR_LOOP);
1724 if (n->Parent->Children[1]) {
1725 /* emit tail code */
1726 if (emitInfo->EmitComments) {
1727 emit_comment(emitInfo, "continue - tail code:");
1728 }
1729 emit(emitInfo, n->Parent->Children[1]);
1730 }
1731 }
1732
1733 /* opcode selection */
1734 if (emitInfo->EmitHighLevelInstructions) {
1735 opcode = (n->Opcode == IR_CONT) ? OPCODE_CONT : OPCODE_BRK;
1736 }
1737 else {
1738 opcode = OPCODE_BRA;
1739 }
1740 n->InstLocation = emitInfo->prog->NumInstructions;
1741 inst = new_instruction(emitInfo, opcode);
1742 inst->DstReg.CondMask = COND_TR; /* always true */
1743 return inst;
1744 }
1745
1746
1747 /**
1748 * Conditional "continue" or "break" statement.
1749 * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1750 */
1751 static struct prog_instruction *
1752 emit_cont_break_if_true(slang_emit_info *emitInfo, slang_ir_node *n)
1753 {
1754 struct prog_instruction *inst;
1755
1756 assert(n->Opcode == IR_CONT_IF_TRUE ||
1757 n->Opcode == IR_BREAK_IF_TRUE);
1758
1759 /* evaluate condition expr, setting cond codes */
1760 inst = emit(emitInfo, n->Children[0]);
1761 if (emitInfo->EmitCondCodes) {
1762 assert(inst);
1763 inst->CondUpdate = GL_TRUE;
1764 }
1765
1766 n->InstLocation = emitInfo->prog->NumInstructions;
1767
1768 /* opcode selection */
1769 if (emitInfo->EmitHighLevelInstructions) {
1770 const gl_inst_opcode opcode
1771 = (n->Opcode == IR_CONT_IF_TRUE) ? OPCODE_CONT : OPCODE_BRK;
1772 if (emitInfo->EmitCondCodes) {
1773 /* Get the writemask from the previous instruction which set
1774 * the condcodes. Use that writemask as the CondSwizzle.
1775 */
1776 const GLuint condWritemask = inst->DstReg.WriteMask;
1777 inst = new_instruction(emitInfo, opcode);
1778 inst->DstReg.CondMask = COND_NE;
1779 inst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
1780 return inst;
1781 }
1782 else {
1783 /* IF reg
1784 * BRK/CONT;
1785 * ENDIF
1786 */
1787 GLint ifInstLoc;
1788 ifInstLoc = emitInfo->prog->NumInstructions;
1789 inst = emit_instruction(emitInfo, OPCODE_IF,
1790 NULL, /* dest */
1791 n->Children[0]->Store,
1792 NULL,
1793 NULL);
1794 n->InstLocation = emitInfo->prog->NumInstructions;
1795
1796 inst = new_instruction(emitInfo, opcode);
1797 inst = new_instruction(emitInfo, OPCODE_ENDIF);
1798
1799 emitInfo->prog->Instructions[ifInstLoc].BranchTarget
1800 = emitInfo->prog->NumInstructions;
1801 return inst;
1802 }
1803 }
1804 else {
1805 const GLuint condWritemask = inst->DstReg.WriteMask;
1806 assert(emitInfo->EmitCondCodes);
1807 inst = new_instruction(emitInfo, OPCODE_BRA);
1808 inst->DstReg.CondMask = COND_NE;
1809 inst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
1810 return inst;
1811 }
1812 }
1813
1814
1815 static struct prog_instruction *
1816 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
1817 {
1818 struct prog_instruction *inst;
1819
1820 inst = emit(emitInfo, n->Children[0]);
1821
1822 #if 0
1823 assert(n->Store->Parent);
1824 /* Apply this node's swizzle to parent's storage */
1825 GLuint swizzle = n->Store->Swizzle;
1826 _slang_copy_ir_storage(n->Store, n->Store->Parent);
1827 n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
1828 assert(!n->Store->Parent);
1829 #endif
1830 return inst;
1831 }
1832
1833
1834 /**
1835 * Dereference array element: element == array[index]
1836 * This basically involves emitting code for computing the array index
1837 * and updating the node/element's storage info.
1838 */
1839 static struct prog_instruction *
1840 emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
1841 {
1842 slang_ir_storage *arrayStore, *indexStore;
1843 const int elemSize = n->Store->Size; /* number of floats */
1844 const GLint elemSizeVec = (elemSize + 3) / 4; /* number of vec4 */
1845 struct prog_instruction *inst;
1846
1847 assert(n->Opcode == IR_ELEMENT);
1848 assert(elemSize > 0);
1849
1850 /* special case for built-in state variables, like light state */
1851 {
1852 slang_ir_storage *root = n->Store;
1853 assert(!root->Parent);
1854 while (root->Parent)
1855 root = root->Parent;
1856
1857 if (root->File == PROGRAM_STATE_VAR) {
1858 GLboolean direct;
1859 GLint index =
1860 _slang_alloc_statevar(n, emitInfo->prog->Parameters, &direct);
1861 if (index < 0) {
1862 /* error */
1863 return NULL;
1864 }
1865 if (direct) {
1866 n->Store->Index = index;
1867 return NULL; /* all done */
1868 }
1869 }
1870 }
1871
1872 /* do codegen for array itself */
1873 emit(emitInfo, n->Children[0]);
1874 arrayStore = n->Children[0]->Store;
1875
1876 /* The initial array element storage is the array's storage,
1877 * then modified below.
1878 */
1879 _slang_copy_ir_storage(n->Store, arrayStore);
1880
1881
1882 if (n->Children[1]->Opcode == IR_FLOAT) {
1883 /* Constant array index */
1884 const GLint element = (GLint) n->Children[1]->Value[0];
1885
1886 /* this element's storage is the array's storage, plus constant offset */
1887 n->Store->Index += elemSizeVec * element;
1888 }
1889 else {
1890 /* Variable array index */
1891
1892 /* do codegen for array index expression */
1893 emit(emitInfo, n->Children[1]);
1894 indexStore = n->Children[1]->Store;
1895
1896 if (indexStore->IsIndirect) {
1897 /* need to put the array index into a temporary since we can't
1898 * directly support a[b[i]] constructs.
1899 */
1900
1901
1902 /*indexStore = tempstore();*/
1903 }
1904
1905
1906 if (elemSize > 4) {
1907 /* need to multiply array index by array element size */
1908 struct prog_instruction *inst;
1909 slang_ir_storage *indexTemp;
1910 slang_ir_storage elemSizeStore;
1911
1912 /* allocate 1 float indexTemp */
1913 indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1);
1914 _slang_alloc_temp(emitInfo->vt, indexTemp);
1915
1916 /* allocate a constant containing the element size */
1917 constant_to_storage(emitInfo, (float) elemSizeVec, &elemSizeStore);
1918
1919 /* multiply array index by element size */
1920 inst = emit_instruction(emitInfo,
1921 OPCODE_MUL,
1922 indexTemp, /* dest */
1923 indexStore, /* the index */
1924 &elemSizeStore,
1925 NULL);
1926
1927 indexStore = indexTemp;
1928 }
1929
1930 if (arrayStore->IsIndirect) {
1931 /* ex: in a[i][j], a[i] (the arrayStore) is indirect */
1932 /* Need to add indexStore to arrayStore->Indirect store */
1933 slang_ir_storage indirectArray;
1934 slang_ir_storage *indexTemp;
1935
1936 _slang_init_ir_storage(&indirectArray,
1937 arrayStore->IndirectFile,
1938 arrayStore->IndirectIndex,
1939 1,
1940 arrayStore->IndirectSwizzle);
1941
1942 /* allocate 1 float indexTemp */
1943 indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1);
1944 _slang_alloc_temp(emitInfo->vt, indexTemp);
1945
1946 inst = emit_instruction(emitInfo,
1947 OPCODE_ADD,
1948 indexTemp, /* dest */
1949 indexStore, /* the index */
1950 &indirectArray, /* indirect array base */
1951 NULL);
1952
1953 indexStore = indexTemp;
1954 }
1955
1956 /* update the array element storage info */
1957 n->Store->IsIndirect = GL_TRUE;
1958 n->Store->IndirectFile = indexStore->File;
1959 n->Store->IndirectIndex = indexStore->Index;
1960 n->Store->IndirectSwizzle = indexStore->Swizzle;
1961 }
1962
1963 n->Store->Size = elemSize;
1964 n->Store->Swizzle = _slang_var_swizzle(elemSize, 0);
1965
1966 return NULL; /* no instruction */
1967 }
1968
1969
1970 /**
1971 * Resolve storage for accessing a structure field.
1972 */
1973 static struct prog_instruction *
1974 emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
1975 {
1976 slang_ir_storage *root = n->Store;
1977 GLint fieldOffset, fieldSize;
1978
1979 assert(n->Opcode == IR_FIELD);
1980
1981 assert(!root->Parent);
1982 while (root->Parent)
1983 root = root->Parent;
1984
1985 /* If this is the field of a state var, allocate constant/uniform
1986 * storage for it now if we haven't already.
1987 * Note that we allocate storage (uniform/constant slots) for state
1988 * variables here rather than at declaration time so we only allocate
1989 * space for the ones that we actually use!
1990 */
1991 if (root->File == PROGRAM_STATE_VAR) {
1992 GLboolean direct;
1993 GLint index = _slang_alloc_statevar(n, emitInfo->prog->Parameters, &direct);
1994 if (index < 0) {
1995 slang_info_log_error(emitInfo->log, "Error parsing state variable");
1996 return NULL;
1997 }
1998 if (direct) {
1999 root->Index = index;
2000 return NULL; /* all done */
2001 }
2002 }
2003
2004 /* do codegen for struct */
2005 emit(emitInfo, n->Children[0]);
2006 assert(n->Children[0]->Store->Index >= 0);
2007
2008
2009 fieldOffset = n->Store->Index;
2010 fieldSize = n->Store->Size;
2011
2012 _slang_copy_ir_storage(n->Store, n->Children[0]->Store);
2013
2014 n->Store->Index = n->Children[0]->Store->Index + fieldOffset / 4;
2015 n->Store->Size = fieldSize;
2016
2017 switch (fieldSize) {
2018 case 1:
2019 {
2020 GLint swz = fieldOffset % 4;
2021 n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz);
2022 }
2023 break;
2024 case 2:
2025 n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
2026 SWIZZLE_NIL, SWIZZLE_NIL);
2027 break;
2028 case 3:
2029 n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
2030 SWIZZLE_Z, SWIZZLE_NIL);
2031 break;
2032 default:
2033 n->Store->Swizzle = SWIZZLE_XYZW;
2034 }
2035
2036 assert(n->Store->Index >= 0);
2037
2038 return NULL; /* no instruction */
2039 }
2040
2041
2042 /**
2043 * Emit code for a variable declaration.
2044 * This usually doesn't result in any code generation, but just
2045 * memory allocation.
2046 */
2047 static struct prog_instruction *
2048 emit_var_decl(slang_emit_info *emitInfo, slang_ir_node *n)
2049 {
2050 assert(n->Store);
2051 assert(n->Store->File != PROGRAM_UNDEFINED);
2052 assert(n->Store->Size > 0);
2053 /*assert(n->Store->Index < 0);*/
2054
2055 if (!n->Var || n->Var->isTemp) {
2056 /* a nameless/temporary variable, will be freed after first use */
2057 /*NEW*/
2058 if (n->Store->Index < 0 && !_slang_alloc_temp(emitInfo->vt, n->Store)) {
2059 slang_info_log_error(emitInfo->log,
2060 "Ran out of registers, too many temporaries");
2061 return NULL;
2062 }
2063 }
2064 else {
2065 /* a regular variable */
2066 _slang_add_variable(emitInfo->vt, n->Var);
2067 if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
2068 slang_info_log_error(emitInfo->log,
2069 "Ran out of registers, too many variables");
2070 return NULL;
2071 }
2072 /*
2073 printf("IR_VAR_DECL %s %d store %p\n",
2074 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
2075 */
2076 assert(n->Var->store == n->Store);
2077 }
2078 if (emitInfo->EmitComments) {
2079 /* emit NOP with comment describing the variable's storage location */
2080 char s[1000];
2081 sprintf(s, "TEMP[%d]%s = variable %s (size %d)",
2082 n->Store->Index,
2083 _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
2084 (n->Var ? (char *) n->Var->a_name : "anonymous"),
2085 n->Store->Size);
2086 emit_comment(emitInfo, s);
2087 }
2088 return NULL;
2089 }
2090
2091
2092 /**
2093 * Emit code for a reference to a variable.
2094 * Actually, no code is generated but we may do some memory allocation.
2095 * In particular, state vars (uniforms) are allocated on an as-needed basis.
2096 */
2097 static struct prog_instruction *
2098 emit_var_ref(slang_emit_info *emitInfo, slang_ir_node *n)
2099 {
2100 assert(n->Store);
2101 assert(n->Store->File != PROGRAM_UNDEFINED);
2102
2103 if (n->Store->File == PROGRAM_STATE_VAR && n->Store->Index < 0) {
2104 GLboolean direct;
2105 GLint index = _slang_alloc_statevar(n, emitInfo->prog->Parameters, &direct);
2106 if (index < 0) {
2107 /* error */
2108 char s[100];
2109 _mesa_snprintf(s, sizeof(s), "Undefined variable '%s'",
2110 (char *) n->Var->a_name);
2111 slang_info_log_error(emitInfo->log, s);
2112 return NULL;
2113 }
2114
2115 n->Store->Index = index;
2116 }
2117 else if (n->Store->File == PROGRAM_UNIFORM ||
2118 n->Store->File == PROGRAM_SAMPLER) {
2119 /* mark var as used */
2120 _mesa_use_uniform(emitInfo->prog->Parameters, (char *) n->Var->a_name);
2121 }
2122
2123 if (n->Store->Index < 0) {
2124 /* probably ran out of registers */
2125 return NULL;
2126 }
2127 assert(n->Store->Size > 0);
2128
2129 return NULL;
2130 }
2131
2132
2133 static struct prog_instruction *
2134 emit(slang_emit_info *emitInfo, slang_ir_node *n)
2135 {
2136 struct prog_instruction *inst;
2137 if (!n)
2138 return NULL;
2139
2140 if (emitInfo->log->error_flag) {
2141 return NULL;
2142 }
2143
2144 switch (n->Opcode) {
2145 case IR_SEQ:
2146 /* sequence of two sub-trees */
2147 assert(n->Children[0]);
2148 assert(n->Children[1]);
2149 emit(emitInfo, n->Children[0]);
2150 if (emitInfo->log->error_flag)
2151 return NULL;
2152 inst = emit(emitInfo, n->Children[1]);
2153 #if 0
2154 assert(!n->Store);
2155 #endif
2156 n->Store = n->Children[1]->Store;
2157 return inst;
2158
2159 case IR_SCOPE:
2160 /* new variable scope */
2161 _slang_push_var_table(emitInfo->vt);
2162 inst = emit(emitInfo, n->Children[0]);
2163 _slang_pop_var_table(emitInfo->vt);
2164 return inst;
2165
2166 case IR_VAR_DECL:
2167 /* Variable declaration - allocate a register for it */
2168 inst = emit_var_decl(emitInfo, n);
2169 return inst;
2170
2171 case IR_VAR:
2172 /* Reference to a variable
2173 * Storage should have already been resolved/allocated.
2174 */
2175 return emit_var_ref(emitInfo, n);
2176
2177 case IR_ELEMENT:
2178 return emit_array_element(emitInfo, n);
2179 case IR_FIELD:
2180 return emit_struct_field(emitInfo, n);
2181 case IR_SWIZZLE:
2182 return emit_swizzle(emitInfo, n);
2183
2184 /* Simple arithmetic */
2185 /* unary */
2186 case IR_MOVE:
2187 case IR_RSQ:
2188 case IR_RCP:
2189 case IR_FLOOR:
2190 case IR_FRAC:
2191 case IR_F_TO_I:
2192 case IR_I_TO_F:
2193 case IR_ABS:
2194 case IR_SIN:
2195 case IR_COS:
2196 case IR_DDX:
2197 case IR_DDY:
2198 case IR_EXP:
2199 case IR_EXP2:
2200 case IR_LOG2:
2201 case IR_NOISE1:
2202 case IR_NOISE2:
2203 case IR_NOISE3:
2204 case IR_NOISE4:
2205 case IR_NRM4:
2206 case IR_NRM3:
2207 /* binary */
2208 case IR_ADD:
2209 case IR_SUB:
2210 case IR_MUL:
2211 case IR_DOT4:
2212 case IR_DOT3:
2213 case IR_DOT2:
2214 case IR_CROSS:
2215 case IR_MIN:
2216 case IR_MAX:
2217 case IR_SEQUAL:
2218 case IR_SNEQUAL:
2219 case IR_SGE:
2220 case IR_SGT:
2221 case IR_SLE:
2222 case IR_SLT:
2223 case IR_POW:
2224 /* trinary operators */
2225 case IR_LRP:
2226 return emit_arith(emitInfo, n);
2227
2228 case IR_EQUAL:
2229 case IR_NOTEQUAL:
2230 return emit_compare(emitInfo, n);
2231
2232 case IR_CLAMP:
2233 return emit_clamp(emitInfo, n);
2234 case IR_TEX:
2235 case IR_TEXB:
2236 case IR_TEXP:
2237 return emit_tex(emitInfo, n);
2238 case IR_NEG:
2239 return emit_negation(emitInfo, n);
2240 case IR_FLOAT:
2241 /* find storage location for this float constant */
2242 n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters,
2243 n->Value,
2244 n->Store->Size,
2245 &n->Store->Swizzle);
2246 if (n->Store->Index < 0) {
2247 slang_info_log_error(emitInfo->log, "Ran out of space for constants");
2248 return NULL;
2249 }
2250 return NULL;
2251
2252 case IR_COPY:
2253 return emit_copy(emitInfo, n);
2254
2255 case IR_COND:
2256 return emit_cond(emitInfo, n);
2257
2258 case IR_NOT:
2259 return emit_not(emitInfo, n);
2260
2261 case IR_LABEL:
2262 return emit_label(emitInfo, n);
2263
2264 case IR_KILL:
2265 return emit_kill(emitInfo);
2266
2267 case IR_CALL:
2268 /* new variable scope for subroutines/function calls */
2269 _slang_push_var_table(emitInfo->vt);
2270 inst = emit_fcall(emitInfo, n);
2271 _slang_pop_var_table(emitInfo->vt);
2272 return inst;
2273
2274 case IR_IF:
2275 return emit_if(emitInfo, n);
2276
2277 case IR_LOOP:
2278 return emit_loop(emitInfo, n);
2279 case IR_BREAK_IF_TRUE:
2280 case IR_CONT_IF_TRUE:
2281 return emit_cont_break_if_true(emitInfo, n);
2282 case IR_BREAK:
2283 /* fall-through */
2284 case IR_CONT:
2285 return emit_cont_break(emitInfo, n);
2286
2287 case IR_BEGIN_SUB:
2288 return new_instruction(emitInfo, OPCODE_BGNSUB);
2289 case IR_END_SUB:
2290 return new_instruction(emitInfo, OPCODE_ENDSUB);
2291 case IR_RETURN:
2292 return emit_return(emitInfo, n);
2293
2294 case IR_NOP:
2295 return NULL;
2296
2297 default:
2298 _mesa_problem(NULL, "Unexpected IR opcode in emit()\n");
2299 }
2300 return NULL;
2301 }
2302
2303
2304 /**
2305 * After code generation, any subroutines will be in separate program
2306 * objects. This function appends all the subroutines onto the main
2307 * program and resolves the linking of all the branch/call instructions.
2308 * XXX this logic should really be part of the linking process...
2309 */
2310 static void
2311 _slang_resolve_subroutines(slang_emit_info *emitInfo)
2312 {
2313 GET_CURRENT_CONTEXT(ctx);
2314 struct gl_program *mainP = emitInfo->prog;
2315 GLuint *subroutineLoc, i, total;
2316
2317 subroutineLoc
2318 = (GLuint *) _mesa_malloc(emitInfo->NumSubroutines * sizeof(GLuint));
2319
2320 /* total number of instructions */
2321 total = mainP->NumInstructions;
2322 for (i = 0; i < emitInfo->NumSubroutines; i++) {
2323 subroutineLoc[i] = total;
2324 total += emitInfo->Subroutines[i]->NumInstructions;
2325 }
2326
2327 /* adjust BranchTargets within the functions */
2328 for (i = 0; i < emitInfo->NumSubroutines; i++) {
2329 struct gl_program *sub = emitInfo->Subroutines[i];
2330 GLuint j;
2331 for (j = 0; j < sub->NumInstructions; j++) {
2332 struct prog_instruction *inst = sub->Instructions + j;
2333 if (inst->Opcode != OPCODE_CAL && inst->BranchTarget >= 0) {
2334 inst->BranchTarget += subroutineLoc[i];
2335 }
2336 }
2337 }
2338
2339 /* append subroutines' instructions after main's instructions */
2340 mainP->Instructions = _mesa_realloc_instructions(mainP->Instructions,
2341 mainP->NumInstructions,
2342 total);
2343 mainP->NumInstructions = total;
2344 for (i = 0; i < emitInfo->NumSubroutines; i++) {
2345 struct gl_program *sub = emitInfo->Subroutines[i];
2346 _mesa_copy_instructions(mainP->Instructions + subroutineLoc[i],
2347 sub->Instructions,
2348 sub->NumInstructions);
2349 /* delete subroutine code */
2350 sub->Parameters = NULL; /* prevent double-free */
2351 _mesa_reference_program(ctx, &emitInfo->Subroutines[i], NULL);
2352 }
2353
2354 /* free subroutine list */
2355 if (emitInfo->Subroutines) {
2356 _mesa_free(emitInfo->Subroutines);
2357 emitInfo->Subroutines = NULL;
2358 }
2359 emitInfo->NumSubroutines = 0;
2360
2361 /* Examine CAL instructions.
2362 * At this point, the BranchTarget field of the CAL instruction is
2363 * the number/id of the subroutine to call (an index into the
2364 * emitInfo->Subroutines list).
2365 * Translate that into an actual instruction location now.
2366 */
2367 for (i = 0; i < mainP->NumInstructions; i++) {
2368 struct prog_instruction *inst = mainP->Instructions + i;
2369 if (inst->Opcode == OPCODE_CAL) {
2370 const GLuint f = inst->BranchTarget;
2371 inst->BranchTarget = subroutineLoc[f];
2372 }
2373 }
2374
2375 _mesa_free(subroutineLoc);
2376 }
2377
2378
2379
2380
2381 GLboolean
2382 _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
2383 struct gl_program *prog, GLboolean withEnd,
2384 slang_info_log *log)
2385 {
2386 GET_CURRENT_CONTEXT(ctx);
2387 GLboolean success;
2388 slang_emit_info emitInfo;
2389 GLuint maxUniforms;
2390
2391 emitInfo.log = log;
2392 emitInfo.vt = vt;
2393 emitInfo.prog = prog;
2394 emitInfo.Subroutines = NULL;
2395 emitInfo.NumSubroutines = 0;
2396 emitInfo.MaxInstructions = prog->NumInstructions;
2397
2398 emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
2399 emitInfo.EmitCondCodes = ctx->Shader.EmitCondCodes;
2400 emitInfo.EmitComments = ctx->Shader.EmitComments;
2401 emitInfo.EmitBeginEndSub = GL_TRUE;
2402
2403 if (!emitInfo.EmitCondCodes) {
2404 emitInfo.EmitHighLevelInstructions = GL_TRUE;
2405 }
2406
2407 /* Check uniform/constant limits */
2408 if (prog->Target == GL_FRAGMENT_PROGRAM_ARB) {
2409 maxUniforms = ctx->Const.FragmentProgram.MaxUniformComponents / 4;
2410 }
2411 else {
2412 assert(prog->Target == GL_VERTEX_PROGRAM_ARB);
2413 maxUniforms = ctx->Const.VertexProgram.MaxUniformComponents / 4;
2414 }
2415 if (prog->Parameters->NumParameters > maxUniforms) {
2416 slang_info_log_error(log, "Constant/uniform register limit exceeded");
2417 return GL_FALSE;
2418 }
2419
2420 (void) emit(&emitInfo, n);
2421
2422 /* finish up by adding the END opcode to program */
2423 if (withEnd) {
2424 struct prog_instruction *inst;
2425 inst = new_instruction(&emitInfo, OPCODE_END);
2426 }
2427
2428 _slang_resolve_subroutines(&emitInfo);
2429
2430 success = GL_TRUE;
2431
2432 #if 0
2433 printf("*********** End emit code (%u inst):\n", prog->NumInstructions);
2434 _mesa_print_program(prog);
2435 _mesa_print_program_parameters(ctx,prog);
2436 #endif
2437
2438 return success;
2439 }