2 * Copyright 2009 Nicolai Hähnle <nhaehnle@gmail.com>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE. */
23 #include "radeon_compiler.h"
27 #include "../r300_reg.h"
29 #include "radeon_compiler_util.h"
30 #include "radeon_dataflow.h"
31 #include "radeon_program_alu.h"
32 #include "radeon_swizzle.h"
33 #include "radeon_emulate_branches.h"
34 #include "radeon_emulate_loops.h"
35 #include "radeon_remove_constants.h"
43 * Take an already-setup and valid source then swizzle it appropriately to
44 * obtain a constant ZERO or ONE source.
46 #define __CONST(x, y) \
47 (PVS_SRC_OPERAND(t_src_index(vp, &vpi->SrcReg[x]), \
52 t_src_class(vpi->SrcReg[x].File), \
53 RC_MASK_NONE) | (vpi->SrcReg[x].RelAddr << 4))
56 static unsigned long t_dst_mask(unsigned int mask)
58 /* RC_MASK_* is equivalent to VSF_FLAG_* */
59 return mask & RC_MASK_XYZW;
62 static unsigned long t_dst_class(rc_register_file file)
66 fprintf(stderr, "%s: Bad register file %i\n", __FUNCTION__, file);
68 case RC_FILE_TEMPORARY:
69 return PVS_DST_REG_TEMPORARY;
71 return PVS_DST_REG_OUT;
73 return PVS_DST_REG_A0;
77 static unsigned long t_dst_index(struct r300_vertex_program_code *vp,
78 struct rc_dst_register *dst)
80 if (dst->File == RC_FILE_OUTPUT)
81 return vp->outputs[dst->Index];
86 static unsigned long t_src_class(rc_register_file file)
90 fprintf(stderr, "%s: Bad register file %i\n", __FUNCTION__, file);
93 case RC_FILE_TEMPORARY:
94 return PVS_SRC_REG_TEMPORARY;
96 return PVS_SRC_REG_INPUT;
97 case RC_FILE_CONSTANT:
98 return PVS_SRC_REG_CONSTANT;
102 static int t_src_conflict(struct rc_src_register a, struct rc_src_register b)
104 unsigned long aclass = t_src_class(a.File);
105 unsigned long bclass = t_src_class(b.File);
107 if (aclass != bclass)
109 if (aclass == PVS_SRC_REG_TEMPORARY)
112 if (a.RelAddr || b.RelAddr)
114 if (a.Index != b.Index)
120 static inline unsigned long t_swizzle(unsigned int swizzle)
122 /* this is in fact a NOP as the Mesa RC_SWIZZLE_* are all identical to VSF_IN_COMPONENT_* */
126 static unsigned long t_src_index(struct r300_vertex_program_code *vp,
127 struct rc_src_register *src)
129 if (src->File == RC_FILE_INPUT) {
130 assert(vp->inputs[src->Index] != -1);
131 return vp->inputs[src->Index];
133 if (src->Index < 0) {
135 "negative offsets for indirect addressing do not work.\n");
142 /* these two functions should probably be merged... */
144 static unsigned long t_src(struct r300_vertex_program_code *vp,
145 struct rc_src_register *src)
147 /* src->Negate uses the RC_MASK_ flags from program_instruction.h,
148 * which equal our VSF_FLAGS_ values, so it's safe to just pass it here.
150 return PVS_SRC_OPERAND(t_src_index(vp, src),
151 t_swizzle(GET_SWZ(src->Swizzle, 0)),
152 t_swizzle(GET_SWZ(src->Swizzle, 1)),
153 t_swizzle(GET_SWZ(src->Swizzle, 2)),
154 t_swizzle(GET_SWZ(src->Swizzle, 3)),
155 t_src_class(src->File),
157 (src->RelAddr << 4) | (src->Abs << 3);
160 static unsigned long t_src_scalar(struct r300_vertex_program_code *vp,
161 struct rc_src_register *src)
163 /* src->Negate uses the RC_MASK_ flags from program_instruction.h,
164 * which equal our VSF_FLAGS_ values, so it's safe to just pass it here.
166 return PVS_SRC_OPERAND(t_src_index(vp, src),
167 t_swizzle(GET_SWZ(src->Swizzle, 0)),
168 t_swizzle(GET_SWZ(src->Swizzle, 0)),
169 t_swizzle(GET_SWZ(src->Swizzle, 0)),
170 t_swizzle(GET_SWZ(src->Swizzle, 0)),
171 t_src_class(src->File),
172 src->Negate ? RC_MASK_XYZW : RC_MASK_NONE) |
173 (src->RelAddr << 4) | (src->Abs << 3);
176 static int valid_dst(struct r300_vertex_program_code *vp,
177 struct rc_dst_register *dst)
179 if (dst->File == RC_FILE_OUTPUT && vp->outputs[dst->Index] == -1) {
181 } else if (dst->File == RC_FILE_ADDRESS) {
182 assert(dst->Index == 0);
188 static void ei_vector1(struct r300_vertex_program_code *vp,
189 unsigned int hw_opcode,
190 struct rc_sub_instruction *vpi,
193 inst[0] = PVS_OP_DST_OPERAND(hw_opcode,
196 t_dst_index(vp, &vpi->DstReg),
197 t_dst_mask(vpi->DstReg.WriteMask),
198 t_dst_class(vpi->DstReg.File));
199 inst[1] = t_src(vp, &vpi->SrcReg[0]);
200 inst[2] = __CONST(0, RC_SWIZZLE_ZERO);
201 inst[3] = __CONST(0, RC_SWIZZLE_ZERO);
204 static void ei_vector2(struct r300_vertex_program_code *vp,
205 unsigned int hw_opcode,
206 struct rc_sub_instruction *vpi,
209 inst[0] = PVS_OP_DST_OPERAND(hw_opcode,
212 t_dst_index(vp, &vpi->DstReg),
213 t_dst_mask(vpi->DstReg.WriteMask),
214 t_dst_class(vpi->DstReg.File));
215 inst[1] = t_src(vp, &vpi->SrcReg[0]);
216 inst[2] = t_src(vp, &vpi->SrcReg[1]);
217 inst[3] = __CONST(1, RC_SWIZZLE_ZERO);
220 static void ei_math1(struct r300_vertex_program_code *vp,
221 unsigned int hw_opcode,
222 struct rc_sub_instruction *vpi,
225 inst[0] = PVS_OP_DST_OPERAND(hw_opcode,
228 t_dst_index(vp, &vpi->DstReg),
229 t_dst_mask(vpi->DstReg.WriteMask),
230 t_dst_class(vpi->DstReg.File));
231 inst[1] = t_src_scalar(vp, &vpi->SrcReg[0]);
232 inst[2] = __CONST(0, RC_SWIZZLE_ZERO);
233 inst[3] = __CONST(0, RC_SWIZZLE_ZERO);
236 static void ei_lit(struct r300_vertex_program_code *vp,
237 struct rc_sub_instruction *vpi,
240 //LIT TMP 1.Y Z TMP 1{} {X W Z Y} TMP 1{} {Y W Z X} TMP 1{} {Y X Z W}
242 inst[0] = PVS_OP_DST_OPERAND(ME_LIGHT_COEFF_DX,
245 t_dst_index(vp, &vpi->DstReg),
246 t_dst_mask(vpi->DstReg.WriteMask),
247 t_dst_class(vpi->DstReg.File));
248 /* NOTE: Users swizzling might not work. */
249 inst[1] = PVS_SRC_OPERAND(t_src_index(vp, &vpi->SrcReg[0]), t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 0)), // X
250 t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 3)), // W
251 PVS_SRC_SELECT_FORCE_0, // Z
252 t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 1)), // Y
253 t_src_class(vpi->SrcReg[0].File),
254 vpi->SrcReg[0].Negate ? RC_MASK_XYZW : RC_MASK_NONE) |
255 (vpi->SrcReg[0].RelAddr << 4);
256 inst[2] = PVS_SRC_OPERAND(t_src_index(vp, &vpi->SrcReg[0]), t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 1)), // Y
257 t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 3)), // W
258 PVS_SRC_SELECT_FORCE_0, // Z
259 t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 0)), // X
260 t_src_class(vpi->SrcReg[0].File),
261 vpi->SrcReg[0].Negate ? RC_MASK_XYZW : RC_MASK_NONE) |
262 (vpi->SrcReg[0].RelAddr << 4);
263 inst[3] = PVS_SRC_OPERAND(t_src_index(vp, &vpi->SrcReg[0]), t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 1)), // Y
264 t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 0)), // X
265 PVS_SRC_SELECT_FORCE_0, // Z
266 t_swizzle(GET_SWZ(vpi->SrcReg[0].Swizzle, 3)), // W
267 t_src_class(vpi->SrcReg[0].File),
268 vpi->SrcReg[0].Negate ? RC_MASK_XYZW : RC_MASK_NONE) |
269 (vpi->SrcReg[0].RelAddr << 4);
272 static void ei_mad(struct r300_vertex_program_code *vp,
273 struct rc_sub_instruction *vpi,
277 /* Remarks about hardware limitations of MAD
278 * (please preserve this comment, as this information is _NOT_
279 * in the documentation provided by AMD).
281 * As described in the documentation, MAD with three unique temporary
282 * source registers requires the use of the macro version.
284 * However (and this is not mentioned in the documentation), apparently
285 * the macro version is _NOT_ a full superset of the normal version.
286 * In particular, the macro version does not always work when relative
287 * addressing is used in the source operands.
289 * This limitation caused incorrect rendering in Sauerbraten's OpenGL
290 * assembly shader path when using medium quality animations
291 * (i.e. animations with matrix blending instead of quaternion blending).
293 * Unfortunately, I (nha) have been unable to extract a Piglit regression
294 * test for this issue - for some reason, it is possible to have vertex
295 * programs whose prefix is *exactly* the same as the prefix of the
296 * offending program in Sauerbraten up to the offending instruction
297 * without causing any trouble.
299 * Bottom line: Only use the macro version only when really necessary;
300 * according to AMD docs, this should improve performance by one clock
301 * as a nice side bonus.
303 if (vpi->SrcReg[0].File == RC_FILE_TEMPORARY &&
304 vpi->SrcReg[1].File == RC_FILE_TEMPORARY &&
305 vpi->SrcReg[2].File == RC_FILE_TEMPORARY &&
306 vpi->SrcReg[0].Index != vpi->SrcReg[1].Index &&
307 vpi->SrcReg[0].Index != vpi->SrcReg[2].Index &&
308 vpi->SrcReg[1].Index != vpi->SrcReg[2].Index) {
309 inst[0] = PVS_OP_DST_OPERAND(PVS_MACRO_OP_2CLK_MADD,
312 t_dst_index(vp, &vpi->DstReg),
313 t_dst_mask(vpi->DstReg.WriteMask),
314 t_dst_class(vpi->DstReg.File));
316 inst[0] = PVS_OP_DST_OPERAND(VE_MULTIPLY_ADD,
319 t_dst_index(vp, &vpi->DstReg),
320 t_dst_mask(vpi->DstReg.WriteMask),
321 t_dst_class(vpi->DstReg.File));
323 /* Arguments with constant swizzles still count as a unique
324 * temporary, so we should make sure these arguments share a
325 * register index with one of the other arguments. */
326 for (i = 0; i < 3; i++) {
328 if (vpi->SrcReg[i].File != RC_FILE_NONE)
331 for (j = 0; j < 3; j++) {
333 vpi->SrcReg[i].Index =
334 vpi->SrcReg[j].Index;
340 inst[1] = t_src(vp, &vpi->SrcReg[0]);
341 inst[2] = t_src(vp, &vpi->SrcReg[1]);
342 inst[3] = t_src(vp, &vpi->SrcReg[2]);
345 static void ei_pow(struct r300_vertex_program_code *vp,
346 struct rc_sub_instruction *vpi,
349 inst[0] = PVS_OP_DST_OPERAND(ME_POWER_FUNC_FF,
352 t_dst_index(vp, &vpi->DstReg),
353 t_dst_mask(vpi->DstReg.WriteMask),
354 t_dst_class(vpi->DstReg.File));
355 inst[1] = t_src_scalar(vp, &vpi->SrcReg[0]);
356 inst[2] = __CONST(0, RC_SWIZZLE_ZERO);
357 inst[3] = t_src_scalar(vp, &vpi->SrcReg[1]);
360 static void mark_write(void * userdata, struct rc_instruction * inst,
361 rc_register_file file, unsigned int index, unsigned int mask)
363 unsigned int * writemasks = userdata;
365 if (file != RC_FILE_TEMPORARY)
368 if (index >= R300_VS_MAX_TEMPS)
371 writemasks[index] |= mask;
374 static unsigned long t_pred_src(struct r300_vertex_program_compiler * compiler)
376 return PVS_SRC_OPERAND(compiler->PredicateIndex,
377 t_swizzle(RC_SWIZZLE_ZERO),
378 t_swizzle(RC_SWIZZLE_ZERO),
379 t_swizzle(RC_SWIZZLE_ZERO),
380 t_swizzle(RC_SWIZZLE_W),
381 t_src_class(RC_FILE_TEMPORARY),
385 static unsigned long t_pred_dst(struct r300_vertex_program_compiler * compiler,
386 unsigned int hw_opcode, int is_math)
388 return PVS_OP_DST_OPERAND(hw_opcode,
391 compiler->PredicateIndex,
393 t_dst_class(RC_FILE_TEMPORARY));
397 static void ei_if(struct r300_vertex_program_compiler * compiler,
398 struct rc_instruction *rci,
400 unsigned int branch_depth)
402 unsigned int predicate_opcode;
405 if (!compiler->Base.is_r500) {
406 rc_error(&compiler->Base,"Opcode IF not supported\n");
410 /* Reserve a temporary to use as our predicate stack counter, if we
411 * don't already have one. */
412 if (!compiler->PredicateMask) {
413 unsigned int writemasks[RC_REGISTER_MAX_INDEX];
414 struct rc_instruction * inst;
416 memset(writemasks, 0, sizeof(writemasks));
417 for(inst = compiler->Base.Program.Instructions.Next;
418 inst != &compiler->Base.Program.Instructions;
420 rc_for_all_writes_mask(inst, mark_write, writemasks);
422 for(i = 0; i < compiler->Base.max_temp_regs; i++) {
423 unsigned int mask = ~writemasks[i] & RC_MASK_XYZW;
424 /* Only the W component can be used fo the predicate
426 if (mask & RC_MASK_W) {
427 compiler->PredicateMask = RC_MASK_W;
428 compiler->PredicateIndex = i;
432 if (i == compiler->Base.max_temp_regs) {
433 rc_error(&compiler->Base, "No free temporary to use for"
434 " predicate stack counter.\n");
439 branch_depth ? VE_PRED_SET_NEQ_PUSH : ME_PRED_SET_NEQ;
441 rci->U.I.SrcReg[0].Swizzle = RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(rci->U.I.SrcReg[0].Swizzle,0));
442 if (branch_depth == 0) {
444 predicate_opcode = ME_PRED_SET_NEQ;
445 inst[1] = t_src(compiler->code, &rci->U.I.SrcReg[0]);
448 predicate_opcode = VE_PRED_SET_NEQ_PUSH;
449 inst[1] = t_pred_src(compiler);
450 inst[2] = t_src(compiler->code, &rci->U.I.SrcReg[0]);
453 inst[0] = t_pred_dst(compiler, predicate_opcode, is_math);
458 static void ei_else(struct r300_vertex_program_compiler * compiler,
461 if (!compiler->Base.is_r500) {
462 rc_error(&compiler->Base,"Opcode ELSE not supported\n");
465 inst[0] = t_pred_dst(compiler, ME_PRED_SET_INV, 1);
466 inst[1] = t_pred_src(compiler);
471 static void ei_endif(struct r300_vertex_program_compiler *compiler,
474 if (!compiler->Base.is_r500) {
475 rc_error(&compiler->Base,"Opcode ENDIF not supported\n");
478 inst[0] = t_pred_dst(compiler, ME_PRED_SET_POP, 1);
479 inst[1] = t_pred_src(compiler);
484 static void translate_vertex_program(struct radeon_compiler *c, void *user)
486 struct r300_vertex_program_compiler *compiler = (struct r300_vertex_program_compiler*)c;
487 struct rc_instruction *rci;
489 struct loop * loops = NULL;
490 int current_loop_depth = 0;
491 int loops_reserved = 0;
493 unsigned int branch_depth = 0;
495 compiler->code->pos_end = 0; /* Not supported yet */
496 compiler->code->length = 0;
497 compiler->code->num_temporaries = 0;
499 compiler->SetHwInputOutput(compiler);
501 for(rci = compiler->Base.Program.Instructions.Next; rci != &compiler->Base.Program.Instructions; rci = rci->Next) {
502 struct rc_sub_instruction *vpi = &rci->U.I;
503 unsigned int *inst = compiler->code->body.d + compiler->code->length;
504 const struct rc_opcode_info *info = rc_get_opcode_info(vpi->Opcode);
506 /* Skip instructions writing to non-existing destination */
507 if (!valid_dst(compiler->code, &vpi->DstReg))
510 if (info->HasDstReg) {
511 /* Neither is Saturate. */
512 if (vpi->SaturateMode != RC_SATURATE_NONE) {
513 rc_error(&compiler->Base, "Vertex program does not support the Saturate "
514 "modifier (yet).\n");
518 if (compiler->code->length >= c->max_alu_insts * 4) {
519 rc_error(&compiler->Base, "Vertex program has too many instructions\n");
523 assert(compiler->Base.is_r500 ||
524 (vpi->Opcode != RC_OPCODE_SEQ &&
525 vpi->Opcode != RC_OPCODE_SNE));
527 switch (vpi->Opcode) {
528 case RC_OPCODE_ADD: ei_vector2(compiler->code, VE_ADD, vpi, inst); break;
529 case RC_OPCODE_ARL: ei_vector1(compiler->code, VE_FLT2FIX_DX, vpi, inst); break;
530 case RC_OPCODE_COS: ei_math1(compiler->code, ME_COS, vpi, inst); break;
531 case RC_OPCODE_DP4: ei_vector2(compiler->code, VE_DOT_PRODUCT, vpi, inst); break;
532 case RC_OPCODE_DST: ei_vector2(compiler->code, VE_DISTANCE_VECTOR, vpi, inst); break;
533 case RC_OPCODE_ELSE: ei_else(compiler, inst); break;
534 case RC_OPCODE_ENDIF: ei_endif(compiler, inst); branch_depth--; break;
535 case RC_OPCODE_EX2: ei_math1(compiler->code, ME_EXP_BASE2_FULL_DX, vpi, inst); break;
536 case RC_OPCODE_EXP: ei_math1(compiler->code, ME_EXP_BASE2_DX, vpi, inst); break;
537 case RC_OPCODE_FRC: ei_vector1(compiler->code, VE_FRACTION, vpi, inst); break;
538 case RC_OPCODE_IF: ei_if(compiler, rci, inst, branch_depth); branch_depth++; break;
539 case RC_OPCODE_LG2: ei_math1(compiler->code, ME_LOG_BASE2_FULL_DX, vpi, inst); break;
540 case RC_OPCODE_LIT: ei_lit(compiler->code, vpi, inst); break;
541 case RC_OPCODE_LOG: ei_math1(compiler->code, ME_LOG_BASE2_DX, vpi, inst); break;
542 case RC_OPCODE_MAD: ei_mad(compiler->code, vpi, inst); break;
543 case RC_OPCODE_MAX: ei_vector2(compiler->code, VE_MAXIMUM, vpi, inst); break;
544 case RC_OPCODE_MIN: ei_vector2(compiler->code, VE_MINIMUM, vpi, inst); break;
545 case RC_OPCODE_MOV: ei_vector1(compiler->code, VE_ADD, vpi, inst); break;
546 case RC_OPCODE_MUL: ei_vector2(compiler->code, VE_MULTIPLY, vpi, inst); break;
547 case RC_OPCODE_POW: ei_pow(compiler->code, vpi, inst); break;
548 case RC_OPCODE_RCP: ei_math1(compiler->code, ME_RECIP_DX, vpi, inst); break;
549 case RC_OPCODE_RSQ: ei_math1(compiler->code, ME_RECIP_SQRT_DX, vpi, inst); break;
550 case RC_OPCODE_SEQ: ei_vector2(compiler->code, VE_SET_EQUAL, vpi, inst); break;
551 case RC_OPCODE_SGE: ei_vector2(compiler->code, VE_SET_GREATER_THAN_EQUAL, vpi, inst); break;
552 case RC_OPCODE_SIN: ei_math1(compiler->code, ME_SIN, vpi, inst); break;
553 case RC_OPCODE_SLT: ei_vector2(compiler->code, VE_SET_LESS_THAN, vpi, inst); break;
554 case RC_OPCODE_SNE: ei_vector2(compiler->code, VE_SET_NOT_EQUAL, vpi, inst); break;
555 case RC_OPCODE_BGNLOOP:
559 if ((!compiler->Base.is_r500
560 && loops_reserved >= R300_VS_MAX_LOOP_DEPTH)
561 || loops_reserved >= R500_VS_MAX_FC_DEPTH) {
562 rc_error(&compiler->Base,
563 "Loops are nested too deep.");
566 memory_pool_array_reserve(&compiler->Base.Pool,
567 struct loop, loops, current_loop_depth,
569 l = &loops[current_loop_depth++];
570 memset(l , 0, sizeof(struct loop));
571 l->BgnLoop = (compiler->code->length / 4);
574 case RC_OPCODE_ENDLOOP:
577 unsigned int act_addr;
578 unsigned int last_addr;
579 unsigned int ret_addr;
582 l = &loops[current_loop_depth - 1];
583 act_addr = l->BgnLoop - 1;
584 last_addr = (compiler->code->length / 4) - 1;
585 ret_addr = l->BgnLoop;
587 if (loops_reserved >= R300_VS_MAX_FC_OPS) {
588 rc_error(&compiler->Base,
589 "Too many flow control instructions.");
592 if (compiler->Base.is_r500) {
593 compiler->code->fc_op_addrs.r500
594 [compiler->code->num_fc_ops].lw =
595 R500_PVS_FC_ACT_ADRS(act_addr)
596 | R500_PVS_FC_LOOP_CNT_JMP_INST(0xffff)
598 compiler->code->fc_op_addrs.r500
599 [compiler->code->num_fc_ops].uw =
600 R500_PVS_FC_LAST_INST(last_addr)
601 | R500_PVS_FC_RTN_INST(ret_addr)
604 compiler->code->fc_op_addrs.r300
605 [compiler->code->num_fc_ops] =
606 R300_PVS_FC_ACT_ADRS(act_addr)
607 | R300_PVS_FC_LOOP_CNT_JMP_INST(0xff)
608 | R300_PVS_FC_LAST_INST(last_addr)
609 | R300_PVS_FC_RTN_INST(ret_addr)
612 compiler->code->fc_loop_index[compiler->code->num_fc_ops] =
613 R300_PVS_FC_LOOP_INIT_VAL(0x0)
614 | R300_PVS_FC_LOOP_STEP_VAL(0x1)
616 compiler->code->fc_ops |= R300_VAP_PVS_FC_OPC_LOOP(
617 compiler->code->num_fc_ops);
618 compiler->code->num_fc_ops++;
619 current_loop_depth--;
624 rc_error(&compiler->Base, "Unknown opcode %s\n", info->Name);
628 /* Non-flow control instructions that are inside an if statement
629 * need to pay attention to the predicate bit. */
631 && vpi->Opcode != RC_OPCODE_IF
632 && vpi->Opcode != RC_OPCODE_ELSE
633 && vpi->Opcode != RC_OPCODE_ENDIF) {
635 inst[0] |= (PVS_DST_PRED_ENABLE_MASK
636 << PVS_DST_PRED_ENABLE_SHIFT);
637 inst[0] |= (PVS_DST_PRED_SENSE_MASK
638 << PVS_DST_PRED_SENSE_SHIFT);
641 /* Update the number of temporaries. */
642 if (info->HasDstReg && vpi->DstReg.File == RC_FILE_TEMPORARY &&
643 vpi->DstReg.Index >= compiler->code->num_temporaries)
644 compiler->code->num_temporaries = vpi->DstReg.Index + 1;
646 for (unsigned i = 0; i < info->NumSrcRegs; i++)
647 if (vpi->SrcReg[i].File == RC_FILE_TEMPORARY &&
648 vpi->SrcReg[i].Index >= compiler->code->num_temporaries)
649 compiler->code->num_temporaries = vpi->SrcReg[i].Index + 1;
651 if (compiler->PredicateMask)
652 if (compiler->PredicateIndex >= compiler->code->num_temporaries)
653 compiler->code->num_temporaries = compiler->PredicateIndex + 1;
655 if (compiler->code->num_temporaries > compiler->Base.max_temp_regs) {
656 rc_error(&compiler->Base, "Too many temporaries.\n");
660 compiler->code->length += 4;
662 if (compiler->Base.Error)
667 struct temporary_allocation {
668 unsigned int Allocated:1;
669 unsigned int HwTemp:15;
670 struct rc_instruction * LastRead;
673 static void allocate_temporary_registers(struct radeon_compiler *c, void *user)
675 struct r300_vertex_program_compiler *compiler = (struct r300_vertex_program_compiler*)c;
676 struct rc_instruction *inst;
677 struct rc_instruction *end_loop = NULL;
678 unsigned int num_orig_temps = 0;
679 char hwtemps[RC_REGISTER_MAX_INDEX];
680 struct temporary_allocation * ta;
683 memset(hwtemps, 0, sizeof(hwtemps));
687 /* Pass 1: Count original temporaries. */
688 for(inst = compiler->Base.Program.Instructions.Next; inst != &compiler->Base.Program.Instructions; inst = inst->Next) {
689 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
691 for (i = 0; i < opcode->NumSrcRegs; ++i) {
692 if (inst->U.I.SrcReg[i].File == RC_FILE_TEMPORARY) {
693 if (inst->U.I.SrcReg[i].Index >= num_orig_temps)
694 num_orig_temps = inst->U.I.SrcReg[i].Index + 1;
698 if (opcode->HasDstReg) {
699 if (inst->U.I.DstReg.File == RC_FILE_TEMPORARY) {
700 if (inst->U.I.DstReg.Index >= num_orig_temps)
701 num_orig_temps = inst->U.I.DstReg.Index + 1;
706 ta = (struct temporary_allocation*)memory_pool_malloc(&compiler->Base.Pool,
707 sizeof(struct temporary_allocation) * num_orig_temps);
708 memset(ta, 0, sizeof(struct temporary_allocation) * num_orig_temps);
710 /* Pass 2: Determine original temporary lifetimes */
711 for(inst = compiler->Base.Program.Instructions.Next; inst != &compiler->Base.Program.Instructions; inst = inst->Next) {
712 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
713 /* Instructions inside of loops need to use the ENDLOOP
714 * instruction as their LastRead. */
715 if (!end_loop && inst->U.I.Opcode == RC_OPCODE_BGNLOOP) {
717 struct rc_instruction * ptr;
718 for(ptr = inst->Next;
719 ptr != &compiler->Base.Program.Instructions;
721 if (ptr->U.I.Opcode == RC_OPCODE_BGNLOOP) {
723 } else if (ptr->U.I.Opcode == RC_OPCODE_ENDLOOP) {
733 if (inst == end_loop) {
738 for (i = 0; i < opcode->NumSrcRegs; ++i) {
739 if (inst->U.I.SrcReg[i].File == RC_FILE_TEMPORARY) {
740 ta[inst->U.I.SrcReg[i].Index].LastRead = end_loop ? end_loop : inst;
745 /* Pass 3: Register allocation */
746 for(inst = compiler->Base.Program.Instructions.Next; inst != &compiler->Base.Program.Instructions; inst = inst->Next) {
747 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
749 for (i = 0; i < opcode->NumSrcRegs; ++i) {
750 if (inst->U.I.SrcReg[i].File == RC_FILE_TEMPORARY) {
751 unsigned int orig = inst->U.I.SrcReg[i].Index;
752 inst->U.I.SrcReg[i].Index = ta[orig].HwTemp;
754 if (ta[orig].Allocated && inst == ta[orig].LastRead)
755 hwtemps[ta[orig].HwTemp] = 0;
759 if (opcode->HasDstReg) {
760 if (inst->U.I.DstReg.File == RC_FILE_TEMPORARY) {
761 unsigned int orig = inst->U.I.DstReg.Index;
763 if (!ta[orig].Allocated) {
764 for(j = 0; j < c->max_temp_regs; ++j) {
768 ta[orig].Allocated = 1;
770 hwtemps[ta[orig].HwTemp] = 1;
773 inst->U.I.DstReg.Index = ta[orig].HwTemp;
780 * R3xx-R4xx vertex engine does not support the Absolute source operand modifier
781 * and the Saturate opcode modifier. Only Absolute is currently transformed.
783 static int transform_nonnative_modifiers(
784 struct radeon_compiler *c,
785 struct rc_instruction *inst,
788 const struct rc_opcode_info *opcode = rc_get_opcode_info(inst->U.I.Opcode);
791 /* Transform ABS(a) to MAX(a, -a). */
792 for (i = 0; i < opcode->NumSrcRegs; i++) {
793 if (inst->U.I.SrcReg[i].Abs) {
794 struct rc_instruction *new_inst;
797 inst->U.I.SrcReg[i].Abs = 0;
799 temp = rc_find_free_temporary(c);
801 new_inst = rc_insert_new_instruction(c, inst->Prev);
802 new_inst->U.I.Opcode = RC_OPCODE_MAX;
803 new_inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
804 new_inst->U.I.DstReg.Index = temp;
805 new_inst->U.I.SrcReg[0] = inst->U.I.SrcReg[i];
806 new_inst->U.I.SrcReg[1] = inst->U.I.SrcReg[i];
807 new_inst->U.I.SrcReg[1].Negate ^= RC_MASK_XYZW;
809 memset(&inst->U.I.SrcReg[i], 0, sizeof(inst->U.I.SrcReg[i]));
810 inst->U.I.SrcReg[i].File = RC_FILE_TEMPORARY;
811 inst->U.I.SrcReg[i].Index = temp;
812 inst->U.I.SrcReg[i].Swizzle = RC_SWIZZLE_XYZW;
819 * Vertex engine cannot read two inputs or two constants at the same time.
820 * Introduce intermediate MOVs to temporary registers to account for this.
822 static int transform_source_conflicts(
823 struct radeon_compiler *c,
824 struct rc_instruction* inst,
827 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
829 if (opcode->NumSrcRegs == 3) {
830 if (t_src_conflict(inst->U.I.SrcReg[1], inst->U.I.SrcReg[2])
831 || t_src_conflict(inst->U.I.SrcReg[0], inst->U.I.SrcReg[2])) {
832 int tmpreg = rc_find_free_temporary(c);
833 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
834 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
835 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
836 inst_mov->U.I.DstReg.Index = tmpreg;
837 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[2];
839 reset_srcreg(&inst->U.I.SrcReg[2]);
840 inst->U.I.SrcReg[2].File = RC_FILE_TEMPORARY;
841 inst->U.I.SrcReg[2].Index = tmpreg;
845 if (opcode->NumSrcRegs >= 2) {
846 if (t_src_conflict(inst->U.I.SrcReg[1], inst->U.I.SrcReg[0])) {
847 int tmpreg = rc_find_free_temporary(c);
848 struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
849 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
850 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
851 inst_mov->U.I.DstReg.Index = tmpreg;
852 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[1];
854 reset_srcreg(&inst->U.I.SrcReg[1]);
855 inst->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
856 inst->U.I.SrcReg[1].Index = tmpreg;
863 static void rc_vs_add_artificial_outputs(struct radeon_compiler *c, void *user)
865 struct r300_vertex_program_compiler * compiler = (struct r300_vertex_program_compiler*)c;
868 for(i = 0; i < 32; ++i) {
869 if ((compiler->RequiredOutputs & (1 << i)) &&
870 !(compiler->Base.Program.OutputsWritten & (1 << i))) {
871 struct rc_instruction * inst = rc_insert_new_instruction(&compiler->Base, compiler->Base.Program.Instructions.Prev);
872 inst->U.I.Opcode = RC_OPCODE_MOV;
874 inst->U.I.DstReg.File = RC_FILE_OUTPUT;
875 inst->U.I.DstReg.Index = i;
876 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
878 inst->U.I.SrcReg[0].File = RC_FILE_CONSTANT;
879 inst->U.I.SrcReg[0].Index = 0;
880 inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;
882 compiler->Base.Program.OutputsWritten |= 1 << i;
887 static void dataflow_outputs_mark_used(void * userdata, void * data,
888 void (*callback)(void *, unsigned int, unsigned int))
890 struct r300_vertex_program_compiler * c = userdata;
893 for(i = 0; i < 32; ++i) {
894 if (c->RequiredOutputs & (1 << i))
895 callback(data, i, RC_MASK_XYZW);
899 static int swizzle_is_native(rc_opcode opcode, struct rc_src_register reg)
907 static void transform_negative_addressing(struct r300_vertex_program_compiler *c,
908 struct rc_instruction *arl,
909 struct rc_instruction *end,
912 struct rc_instruction *inst, *add;
913 unsigned const_swizzle;
916 add = rc_insert_new_instruction(&c->Base, arl->Prev);
917 add->U.I.Opcode = RC_OPCODE_ADD;
918 add->U.I.DstReg.File = RC_FILE_TEMPORARY;
919 add->U.I.DstReg.Index = rc_find_free_temporary(&c->Base);
920 add->U.I.DstReg.WriteMask = RC_MASK_X;
921 add->U.I.SrcReg[0] = arl->U.I.SrcReg[0];
922 add->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
923 add->U.I.SrcReg[1].Index = rc_constants_add_immediate_scalar(&c->Base.Program.Constants,
924 min_offset, &const_swizzle);
925 add->U.I.SrcReg[1].Swizzle = const_swizzle;
927 arl->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
928 arl->U.I.SrcReg[0].Index = add->U.I.DstReg.Index;
929 arl->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XXXX;
931 /* Rewrite offsets up to and excluding inst. */
932 for (inst = arl->Next; inst != end; inst = inst->Next) {
933 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
935 for (unsigned i = 0; i < opcode->NumSrcRegs; i++)
936 if (inst->U.I.SrcReg[i].RelAddr)
937 inst->U.I.SrcReg[i].Index -= min_offset;
941 static void rc_emulate_negative_addressing(struct radeon_compiler *compiler, void *user)
943 struct r300_vertex_program_compiler * c = (struct r300_vertex_program_compiler*)compiler;
944 struct rc_instruction *inst, *lastARL = NULL;
947 for (inst = c->Base.Program.Instructions.Next; inst != &c->Base.Program.Instructions; inst = inst->Next) {
948 const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
950 if (inst->U.I.Opcode == RC_OPCODE_ARL) {
951 if (lastARL != NULL && min_offset < 0)
952 transform_negative_addressing(c, lastARL, inst, min_offset);
959 for (unsigned i = 0; i < opcode->NumSrcRegs; i++) {
960 if (inst->U.I.SrcReg[i].RelAddr &&
961 inst->U.I.SrcReg[i].Index < 0) {
962 /* ARL must precede any indirect addressing. */
963 if (lastARL == NULL) {
964 rc_error(&c->Base, "Vertex shader: Found relative addressing without ARL.");
968 if (inst->U.I.SrcReg[i].Index < min_offset)
969 min_offset = inst->U.I.SrcReg[i].Index;
974 if (lastARL != NULL && min_offset < 0)
975 transform_negative_addressing(c, lastARL, inst, min_offset);
978 static struct rc_swizzle_caps r300_vertprog_swizzle_caps = {
979 .IsNative = &swizzle_is_native,
980 .Split = 0 /* should never be called */
983 void r3xx_compile_vertex_program(struct r300_vertex_program_compiler *c)
985 int is_r500 = c->Base.is_r500;
986 int opt = !c->Base.disable_optimizations;
988 /* Lists of instruction transformations. */
989 struct radeon_program_transformation alu_rewrite_r500[] = {
990 { &r300_transform_vertex_alu, 0 },
991 { &r300_transform_trig_scale_vertex, 0 },
995 struct radeon_program_transformation alu_rewrite_r300[] = {
996 { &r300_transform_vertex_alu, 0 },
997 { &r300_transform_trig_simple, 0 },
1001 /* Note: These passes have to be done seperately from ALU rewrite,
1002 * otherwise non-native ALU instructions with source conflits
1003 * or non-native modifiers will not be treated properly.
1005 struct radeon_program_transformation emulate_modifiers[] = {
1006 { &transform_nonnative_modifiers, 0 },
1010 struct radeon_program_transformation resolve_src_conflicts[] = {
1011 { &transform_source_conflicts, 0 },
1015 /* List of compiler passes. */
1016 struct radeon_compiler_pass vs_list[] = {
1017 /* NAME DUMP PREDICATE FUNCTION PARAM */
1018 {"add artificial outputs", 0, 1, rc_vs_add_artificial_outputs, NULL},
1019 {"transform loops", 1, 1, rc_transform_loops, NULL},
1020 {"emulate branches", 1, !is_r500, rc_emulate_branches, NULL},
1021 {"emulate negative addressing", 1, 1, rc_emulate_negative_addressing, NULL},
1022 {"native rewrite", 1, is_r500, rc_local_transform, alu_rewrite_r500},
1023 {"native rewrite", 1, !is_r500, rc_local_transform, alu_rewrite_r300},
1024 {"emulate modifiers", 1, !is_r500, rc_local_transform, emulate_modifiers},
1025 {"deadcode", 1, opt, rc_dataflow_deadcode, dataflow_outputs_mark_used},
1026 {"dataflow optimize", 1, opt, rc_optimize, NULL},
1027 /* This pass must be done after optimizations. */
1028 {"source conflict resolve", 1, 1, rc_local_transform, resolve_src_conflicts},
1029 {"register allocation", 1, opt, allocate_temporary_registers, NULL},
1030 {"dead constants", 1, 1, rc_remove_unused_constants, &c->code->constants_remap_table},
1031 {"final code validation", 0, 1, rc_validate_final_shader, NULL},
1032 {"machine code generation", 0, 1, translate_vertex_program, NULL},
1033 {"dump machine code", 0, c->Base.Debug & RC_DBG_LOG, r300_vertex_program_dump, NULL},
1034 {NULL, 0, 0, NULL, NULL}
1037 c->Base.type = RC_VERTEX_PROGRAM;
1038 c->Base.SwizzleCaps = &r300_vertprog_swizzle_caps;
1040 rc_run_compiler(&c->Base, vs_list);
1042 c->code->InputsRead = c->Base.Program.InputsRead;
1043 c->code->OutputsWritten = c->Base.Program.OutputsWritten;
1044 rc_constants_copy(&c->code->constants, &c->Base.Program.Constants);