r300/compiler: Don't pair output writes with GPR writes in the scheduler
[mesa.git] / src / gallium / drivers / r300 / compiler / r500_fragprog.c
1 /*
2 * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
3 *
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sublicense, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial
16 * portions of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 */
27
28 #include "r500_fragprog.h"
29
30 #include <stdio.h>
31
32 #include "radeon_compiler_util.h"
33 #include "radeon_list.h"
34 #include "radeon_variable.h"
35 #include "../r300_reg.h"
36
37 /**
38 * Rewrite IF instructions to use the ALU result special register.
39 */
40 int r500_transform_IF(
41 struct radeon_compiler * c,
42 struct rc_instruction * inst_if,
43 void *data)
44 {
45 struct rc_variable * writer;
46 struct rc_list * writer_list, * list_ptr;
47 struct rc_list * var_list = rc_get_variables(c);
48 unsigned int generic_if = 0;
49 unsigned int alu_chan;
50
51 if (inst_if->U.I.Opcode != RC_OPCODE_IF) {
52 return 0;
53 }
54
55 writer_list = rc_variable_list_get_writers(
56 var_list, inst_if->Type, &inst_if->U.I.SrcReg[0]);
57 if (!writer_list) {
58 generic_if = 1;
59 } else {
60
61 /* Make sure it is safe for the writers to write to
62 * ALU Result */
63 for (list_ptr = writer_list; list_ptr;
64 list_ptr = list_ptr->Next) {
65 struct rc_instruction * inst;
66 writer = list_ptr->Item;
67 /* We are going to modify the destination register
68 * of writer, so if it has a reader other than
69 * inst_if (aka ReaderCount > 1) we must fall back to
70 * our generic IF.
71 * If the writer has a lower IP than inst_if, this
72 * means that inst_if is above the writer in a loop.
73 * I'm not sure why this would ever happen, but
74 * if it does we want to make sure we fall back
75 * to our generic IF. */
76 if (writer->ReaderCount > 1 || writer->Inst->IP < inst_if->IP) {
77 generic_if = 1;
78 break;
79 }
80
81 /* The ALU Result is not preserved across IF
82 * instructions, so if there is another IF
83 * instruction between writer and inst_if, then
84 * we need to fall back to generic IF. */
85 for (inst = writer->Inst; inst != inst_if; inst = inst->Next) {
86 const struct rc_opcode_info * info =
87 rc_get_opcode_info(inst->U.I.Opcode);
88 if (info->IsFlowControl) {
89 generic_if = 1;
90 break;
91 }
92 }
93 if (generic_if) {
94 break;
95 }
96 }
97 }
98
99 if (GET_SWZ(inst_if->U.I.SrcReg[0].Swizzle, 0) == RC_SWIZZLE_X) {
100 alu_chan = RC_ALURESULT_X;
101 } else {
102 alu_chan = RC_ALURESULT_W;
103 }
104 if (generic_if) {
105 struct rc_instruction * inst_mov =
106 rc_insert_new_instruction(c, inst_if->Prev);
107
108 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
109 inst_mov->U.I.DstReg.WriteMask = 0;
110 inst_mov->U.I.DstReg.File = RC_FILE_NONE;
111 inst_mov->U.I.ALUResultCompare = RC_COMPARE_FUNC_NOTEQUAL;
112 inst_mov->U.I.WriteALUResult = alu_chan;
113 inst_mov->U.I.SrcReg[0] = inst_if->U.I.SrcReg[0];
114 if (alu_chan == RC_ALURESULT_X) {
115 inst_mov->U.I.SrcReg[0].Swizzle = combine_swizzles4(
116 inst_mov->U.I.SrcReg[0].Swizzle,
117 RC_SWIZZLE_X, RC_SWIZZLE_UNUSED,
118 RC_SWIZZLE_UNUSED, RC_SWIZZLE_UNUSED);
119 } else {
120 inst_mov->U.I.SrcReg[0].Swizzle = combine_swizzles4(
121 inst_mov->U.I.SrcReg[0].Swizzle,
122 RC_SWIZZLE_UNUSED, RC_SWIZZLE_UNUSED,
123 RC_SWIZZLE_UNUSED, RC_SWIZZLE_Z);
124 }
125 } else {
126 rc_compare_func compare_func = RC_COMPARE_FUNC_NEVER;
127 unsigned int reverse_srcs = 0;
128 unsigned int preserve_opcode = 0;
129 for (list_ptr = writer_list; list_ptr;
130 list_ptr = list_ptr->Next) {
131 writer = list_ptr->Item;
132 switch(writer->Inst->U.I.Opcode) {
133 case RC_OPCODE_SEQ:
134 compare_func = RC_COMPARE_FUNC_EQUAL;
135 break;
136 case RC_OPCODE_SNE:
137 compare_func = RC_COMPARE_FUNC_NOTEQUAL;
138 break;
139 case RC_OPCODE_SLE:
140 reverse_srcs = 1;
141 /* Fall through */
142 case RC_OPCODE_SGE:
143 compare_func = RC_COMPARE_FUNC_GEQUAL;
144 break;
145 case RC_OPCODE_SGT:
146 reverse_srcs = 1;
147 /* Fall through */
148 case RC_OPCODE_SLT:
149 compare_func = RC_COMPARE_FUNC_LESS;
150 break;
151 default:
152 compare_func = RC_COMPARE_FUNC_NOTEQUAL;
153 preserve_opcode = 1;
154 break;
155 }
156 if (!preserve_opcode) {
157 writer->Inst->U.I.Opcode = RC_OPCODE_SUB;
158 }
159 writer->Inst->U.I.DstReg.WriteMask = 0;
160 writer->Inst->U.I.DstReg.File = RC_FILE_NONE;
161 writer->Inst->U.I.WriteALUResult = alu_chan;
162 writer->Inst->U.I.ALUResultCompare = compare_func;
163 if (reverse_srcs) {
164 struct rc_src_register temp_src;
165 temp_src = writer->Inst->U.I.SrcReg[0];
166 writer->Inst->U.I.SrcReg[0] =
167 writer->Inst->U.I.SrcReg[1];
168 writer->Inst->U.I.SrcReg[1] = temp_src;
169 }
170 }
171 }
172
173 inst_if->U.I.SrcReg[0].File = RC_FILE_SPECIAL;
174 inst_if->U.I.SrcReg[0].Index = RC_SPECIAL_ALU_RESULT;
175 inst_if->U.I.SrcReg[0].Swizzle = RC_MAKE_SWIZZLE(
176 RC_SWIZZLE_X, RC_SWIZZLE_UNUSED,
177 RC_SWIZZLE_UNUSED, RC_SWIZZLE_UNUSED);
178 inst_if->U.I.SrcReg[0].Negate = 0;
179
180 return 1;
181 }
182
183 static int r500_swizzle_is_native(rc_opcode opcode, struct rc_src_register reg)
184 {
185 unsigned int relevant;
186 int i;
187
188 if (opcode == RC_OPCODE_TEX ||
189 opcode == RC_OPCODE_TXB ||
190 opcode == RC_OPCODE_TXP ||
191 opcode == RC_OPCODE_TXD ||
192 opcode == RC_OPCODE_TXL ||
193 opcode == RC_OPCODE_KIL) {
194 if (reg.Abs)
195 return 0;
196
197 if (opcode == RC_OPCODE_KIL && (reg.Swizzle != RC_SWIZZLE_XYZW || reg.Negate != RC_MASK_NONE))
198 return 0;
199
200 for(i = 0; i < 4; ++i) {
201 unsigned int swz = GET_SWZ(reg.Swizzle, i);
202 if (swz == RC_SWIZZLE_UNUSED) {
203 reg.Negate &= ~(1 << i);
204 continue;
205 }
206 if (swz >= 4)
207 return 0;
208 }
209
210 if (reg.Negate)
211 return 0;
212
213 return 1;
214 } else if (opcode == RC_OPCODE_DDX || opcode == RC_OPCODE_DDY) {
215 /* DDX/MDH and DDY/MDV explicitly ignore incoming swizzles;
216 * if it doesn't fit perfectly into a .xyzw case... */
217 if (reg.Swizzle == RC_SWIZZLE_XYZW && !reg.Abs && !reg.Negate)
218 return 1;
219
220 return 0;
221 } else {
222 /* ALU instructions support almost everything */
223 relevant = 0;
224 for(i = 0; i < 3; ++i) {
225 unsigned int swz = GET_SWZ(reg.Swizzle, i);
226 if (swz != RC_SWIZZLE_UNUSED && swz != RC_SWIZZLE_ZERO)
227 relevant |= 1 << i;
228 }
229 if ((reg.Negate & relevant) && ((reg.Negate & relevant) != relevant))
230 return 0;
231
232 return 1;
233 }
234 }
235
236 /**
237 * Split source register access.
238 *
239 * The only thing we *cannot* do in an ALU instruction is per-component
240 * negation.
241 */
242 static void r500_swizzle_split(struct rc_src_register src, unsigned int usemask,
243 struct rc_swizzle_split * split)
244 {
245 unsigned int negatebase[2] = { 0, 0 };
246 int i;
247
248 for(i = 0; i < 4; ++i) {
249 unsigned int swz = GET_SWZ(src.Swizzle, i);
250 if (swz == RC_SWIZZLE_UNUSED || !GET_BIT(usemask, i))
251 continue;
252 negatebase[GET_BIT(src.Negate, i)] |= 1 << i;
253 }
254
255 split->NumPhases = 0;
256
257 for(i = 0; i <= 1; ++i) {
258 if (!negatebase[i])
259 continue;
260
261 split->Phase[split->NumPhases++] = negatebase[i];
262 }
263 }
264
265 struct rc_swizzle_caps r500_swizzle_caps = {
266 .IsNative = r500_swizzle_is_native,
267 .Split = r500_swizzle_split
268 };
269
270 static char *toswiz(int swiz_val) {
271 switch(swiz_val) {
272 case 0: return "R";
273 case 1: return "G";
274 case 2: return "B";
275 case 3: return "A";
276 case 4: return "0";
277 case 5: return "H";
278 case 6: return "1";
279 case 7: return "U";
280 }
281 return NULL;
282 }
283
284 static char *toop(int op_val)
285 {
286 char *str = NULL;
287 switch (op_val) {
288 case 0: str = "MAD"; break;
289 case 1: str = "DP3"; break;
290 case 2: str = "DP4"; break;
291 case 3: str = "D2A"; break;
292 case 4: str = "MIN"; break;
293 case 5: str = "MAX"; break;
294 case 6: str = "Reserved"; break;
295 case 7: str = "CND"; break;
296 case 8: str = "CMP"; break;
297 case 9: str = "FRC"; break;
298 case 10: str = "SOP"; break;
299 case 11: str = "MDH"; break;
300 case 12: str = "MDV"; break;
301 }
302 return str;
303 }
304
305 static char *to_alpha_op(int op_val)
306 {
307 char *str = NULL;
308 switch (op_val) {
309 case 0: str = "MAD"; break;
310 case 1: str = "DP"; break;
311 case 2: str = "MIN"; break;
312 case 3: str = "MAX"; break;
313 case 4: str = "Reserved"; break;
314 case 5: str = "CND"; break;
315 case 6: str = "CMP"; break;
316 case 7: str = "FRC"; break;
317 case 8: str = "EX2"; break;
318 case 9: str = "LN2"; break;
319 case 10: str = "RCP"; break;
320 case 11: str = "RSQ"; break;
321 case 12: str = "SIN"; break;
322 case 13: str = "COS"; break;
323 case 14: str = "MDH"; break;
324 case 15: str = "MDV"; break;
325 }
326 return str;
327 }
328
329 static char *to_mask(int val)
330 {
331 char *str = NULL;
332 switch(val) {
333 case 0: str = "NONE"; break;
334 case 1: str = "R"; break;
335 case 2: str = "G"; break;
336 case 3: str = "RG"; break;
337 case 4: str = "B"; break;
338 case 5: str = "RB"; break;
339 case 6: str = "GB"; break;
340 case 7: str = "RGB"; break;
341 case 8: str = "A"; break;
342 case 9: str = "AR"; break;
343 case 10: str = "AG"; break;
344 case 11: str = "ARG"; break;
345 case 12: str = "AB"; break;
346 case 13: str = "ARB"; break;
347 case 14: str = "AGB"; break;
348 case 15: str = "ARGB"; break;
349 }
350 return str;
351 }
352
353 static char *to_texop(int val)
354 {
355 switch(val) {
356 case 0: return "NOP";
357 case 1: return "LD";
358 case 2: return "TEXKILL";
359 case 3: return "PROJ";
360 case 4: return "LODBIAS";
361 case 5: return "LOD";
362 case 6: return "DXDY";
363 }
364 return NULL;
365 }
366
367 void r500FragmentProgramDump(struct radeon_compiler *c, void *user)
368 {
369 struct r300_fragment_program_compiler *compiler = (struct r300_fragment_program_compiler*)c;
370 struct r500_fragment_program_code *code = &compiler->code->code.r500;
371 int n, i;
372 uint32_t inst;
373 uint32_t inst0;
374 char *str = NULL;
375 fprintf(stderr, "R500 Fragment Program:\n--------\n");
376
377 for (n = 0; n < code->inst_end+1; n++) {
378 inst0 = inst = code->inst[n].inst0;
379 fprintf(stderr,"%d\t0:CMN_INST 0x%08x:", n, inst);
380 switch(inst & 0x3) {
381 case R500_INST_TYPE_ALU: str = "ALU"; break;
382 case R500_INST_TYPE_OUT: str = "OUT"; break;
383 case R500_INST_TYPE_FC: str = "FC"; break;
384 case R500_INST_TYPE_TEX: str = "TEX"; break;
385 };
386 fprintf(stderr,"%s %s %s %s %s ", str,
387 inst & R500_INST_TEX_SEM_WAIT ? "TEX_WAIT" : "",
388 inst & R500_INST_LAST ? "LAST" : "",
389 inst & R500_INST_NOP ? "NOP" : "",
390 inst & R500_INST_ALU_WAIT ? "ALU WAIT" : "");
391 fprintf(stderr,"wmask: %s omask: %s\n", to_mask((inst >> 11) & 0xf),
392 to_mask((inst >> 15) & 0xf));
393
394 switch(inst0 & 0x3) {
395 case R500_INST_TYPE_ALU:
396 case R500_INST_TYPE_OUT:
397 fprintf(stderr,"\t1:RGB_ADDR 0x%08x:", code->inst[n].inst1);
398 inst = code->inst[n].inst1;
399
400 fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
401 inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
402 (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
403 (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
404 (inst >> 30));
405
406 fprintf(stderr,"\t2:ALPHA_ADDR 0x%08x:", code->inst[n].inst2);
407 inst = code->inst[n].inst2;
408 fprintf(stderr,"Addr0: %d%c, Addr1: %d%c, Addr2: %d%c, srcp:%d\n",
409 inst & 0xff, (inst & (1<<8)) ? 'c' : 't',
410 (inst >> 10) & 0xff, (inst & (1<<18)) ? 'c' : 't',
411 (inst >> 20) & 0xff, (inst & (1<<28)) ? 'c' : 't',
412 (inst >> 30));
413 fprintf(stderr,"\t3 RGB_INST: 0x%08x:", code->inst[n].inst3);
414 inst = code->inst[n].inst3;
415 fprintf(stderr,"rgb_A_src:%d %s/%s/%s %d rgb_B_src:%d %s/%s/%s %d targ: %d\n",
416 (inst) & 0x3, toswiz((inst >> 2) & 0x7), toswiz((inst >> 5) & 0x7), toswiz((inst >> 8) & 0x7),
417 (inst >> 11) & 0x3,
418 (inst >> 13) & 0x3, toswiz((inst >> 15) & 0x7), toswiz((inst >> 18) & 0x7), toswiz((inst >> 21) & 0x7),
419 (inst >> 24) & 0x3, (inst >> 29) & 0x3);
420
421
422 fprintf(stderr,"\t4 ALPHA_INST:0x%08x:", code->inst[n].inst4);
423 inst = code->inst[n].inst4;
424 fprintf(stderr,"%s dest:%d%s alp_A_src:%d %s %d alp_B_src:%d %s %d targ %d w:%d\n", to_alpha_op(inst & 0xf),
425 (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
426 (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), (inst >> 17) & 0x3,
427 (inst >> 19) & 0x3, toswiz((inst >> 21) & 0x7), (inst >> 24) & 0x3,
428 (inst >> 29) & 0x3,
429 (inst >> 31) & 0x1);
430
431 fprintf(stderr,"\t5 RGBA_INST: 0x%08x:", code->inst[n].inst5);
432 inst = code->inst[n].inst5;
433 fprintf(stderr,"%s dest:%d%s rgb_C_src:%d %s/%s/%s %d alp_C_src:%d %s %d\n", toop(inst & 0xf),
434 (inst >> 4) & 0x7f, inst & (1<<11) ? "(rel)":"",
435 (inst >> 12) & 0x3, toswiz((inst >> 14) & 0x7), toswiz((inst >> 17) & 0x7), toswiz((inst >> 20) & 0x7),
436 (inst >> 23) & 0x3,
437 (inst >> 25) & 0x3, toswiz((inst >> 27) & 0x7), (inst >> 30) & 0x3);
438 break;
439 case R500_INST_TYPE_FC:
440 fprintf(stderr, "\t2:FC_INST 0x%08x:", code->inst[n].inst2);
441 inst = code->inst[n].inst2;
442 /* JUMP_FUNC JUMP_ANY*/
443 fprintf(stderr, "0x%02x %1x ", inst >> 8 & 0xff,
444 (inst & R500_FC_JUMP_ANY) >> 5);
445
446 /* OP */
447 switch(inst & 0x7){
448 case R500_FC_OP_JUMP:
449 fprintf(stderr, "JUMP");
450 break;
451 case R500_FC_OP_LOOP:
452 fprintf(stderr, "LOOP");
453 break;
454 case R500_FC_OP_ENDLOOP:
455 fprintf(stderr, "ENDLOOP");
456 break;
457 case R500_FC_OP_REP:
458 fprintf(stderr, "REP");
459 break;
460 case R500_FC_OP_ENDREP:
461 fprintf(stderr, "ENDREP");
462 break;
463 case R500_FC_OP_BREAKLOOP:
464 fprintf(stderr, "BREAKLOOP");
465 break;
466 case R500_FC_OP_BREAKREP:
467 fprintf(stderr, "BREAKREP");
468 break;
469 case R500_FC_OP_CONTINUE:
470 fprintf(stderr, "CONTINUE");
471 break;
472 }
473 fprintf(stderr," ");
474 /* A_OP */
475 switch(inst & (0x3 << 6)){
476 case R500_FC_A_OP_NONE:
477 fprintf(stderr, "NONE");
478 break;
479 case R500_FC_A_OP_POP:
480 fprintf(stderr, "POP");
481 break;
482 case R500_FC_A_OP_PUSH:
483 fprintf(stderr, "PUSH");
484 break;
485 }
486 /* B_OP0 B_OP1 */
487 for(i=0; i<2; i++){
488 fprintf(stderr, " ");
489 switch(inst & (0x3 << (24 + (i * 2)))){
490 /* R500_FC_B_OP0_NONE
491 * R500_FC_B_OP1_NONE */
492 case 0:
493 fprintf(stderr, "NONE");
494 break;
495 case R500_FC_B_OP0_DECR:
496 case R500_FC_B_OP1_DECR:
497 fprintf(stderr, "DECR");
498 break;
499 case R500_FC_B_OP0_INCR:
500 case R500_FC_B_OP1_INCR:
501 fprintf(stderr, "INCR");
502 break;
503 }
504 }
505 /*POP_CNT B_ELSE */
506 fprintf(stderr, " %d %1x", (inst >> 16) & 0x1f, (inst & R500_FC_B_ELSE) >> 4);
507 inst = code->inst[n].inst3;
508 /* JUMP_ADDR */
509 fprintf(stderr, " %d", inst >> 16);
510
511 if(code->inst[n].inst2 & R500_FC_IGNORE_UNCOVERED){
512 fprintf(stderr, " IGN_UNC");
513 }
514 inst = code->inst[n].inst3;
515 fprintf(stderr, "\n\t3:FC_ADDR 0x%08x:", inst);
516 fprintf(stderr, "BOOL: 0x%02x, INT: 0x%02x, JUMP_ADDR: %d, JMP_GLBL: %1x\n",
517 inst & 0x1f, (inst >> 8) & 0x1f, (inst >> 16) & 0x1ff, inst >> 31);
518 break;
519 case R500_INST_TYPE_TEX:
520 inst = code->inst[n].inst1;
521 fprintf(stderr,"\t1:TEX_INST: 0x%08x: id: %d op:%s, %s, %s %s\n", inst, (inst >> 16) & 0xf,
522 to_texop((inst >> 22) & 0x7), (inst & (1<<25)) ? "ACQ" : "",
523 (inst & (1<<26)) ? "IGNUNC" : "", (inst & (1<<27)) ? "UNSCALED" : "SCALED");
524 inst = code->inst[n].inst2;
525 fprintf(stderr,"\t2:TEX_ADDR: 0x%08x: src: %d%s %s/%s/%s/%s dst: %d%s %s/%s/%s/%s\n", inst,
526 inst & 127, inst & (1<<7) ? "(rel)" : "",
527 toswiz((inst >> 8) & 0x3), toswiz((inst >> 10) & 0x3),
528 toswiz((inst >> 12) & 0x3), toswiz((inst >> 14) & 0x3),
529 (inst >> 16) & 127, inst & (1<<23) ? "(rel)" : "",
530 toswiz((inst >> 24) & 0x3), toswiz((inst >> 26) & 0x3),
531 toswiz((inst >> 28) & 0x3), toswiz((inst >> 30) & 0x3));
532
533 fprintf(stderr,"\t3:TEX_DXDY: 0x%08x\n", code->inst[n].inst3);
534 break;
535 }
536 fprintf(stderr,"\n");
537 }
538
539 }