projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Move nv30, nv50 and nvc0 to nouveau.
[mesa.git] / src / gallium / drivers / nouveau / codegen / nv50_ir_emit_gk110.cpp
1 /*
2 * Copyright 2012 Christoph Bumiller
3 *
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23 #include "codegen/nv50_ir_target_nvc0.h"
24
25 // CodeEmitter for GK110 encoding of the Fermi/Kepler ISA.
26
27 namespace nv50_ir {
28
29 class CodeEmitterGK110 : public CodeEmitter
30 {
31 public:
32 CodeEmitterGK110(const TargetNVC0 *);
33
34 virtual bool emitInstruction(Instruction *);
35 virtual uint32_t getMinEncodingSize(const Instruction *) const;
36 virtual void prepareEmission(Function *);
37
38 inline void setProgramType(Program::Type pType) { progType = pType; }
39
40 private:
41 const TargetNVC0 *targNVC0;
42
43 Program::Type progType;
44
45 const bool writeIssueDelays;
46
47 private:
48 void emitForm_21(const Instruction *, uint32_t opc2, uint32_t opc1);
49 void emitForm_C(const Instruction *, uint32_t opc, uint8_t ctg);
50 void emitForm_L(const Instruction *, uint32_t opc, uint8_t ctg, Modifier);
51
52 void emitPredicate(const Instruction *);
53
54 void setCAddress14(const ValueRef&);
55 void setShortImmediate(const Instruction *, const int s);
56 void setImmediate32(const Instruction *, const int s, Modifier);
57
58 void modNegAbsF32_3b(const Instruction *, const int s);
59
60 void emitCondCode(CondCode cc, int pos, uint8_t mask);
61 void emitInterpMode(const Instruction *);
62 void emitLoadStoreType(DataType ty, const int pos);
63 void emitCachingMode(CacheMode c, const int pos);
64
65 inline uint8_t getSRegEncoding(const ValueRef&);
66
67 void emitRoundMode(RoundMode, const int pos, const int rintPos);
68 void emitRoundModeF(RoundMode, const int pos);
69 void emitRoundModeI(RoundMode, const int pos);
70
71 void emitNegAbs12(const Instruction *);
72
73 void emitNOP(const Instruction *);
74
75 void emitLOAD(const Instruction *);
76 void emitSTORE(const Instruction *);
77 void emitMOV(const Instruction *);
78
79 void emitINTERP(const Instruction *);
80 void emitPFETCH(const Instruction *);
81 void emitVFETCH(const Instruction *);
82 void emitEXPORT(const Instruction *);
83 void emitOUT(const Instruction *);
84
85 void emitUADD(const Instruction *);
86 void emitFADD(const Instruction *);
87 void emitIMUL(const Instruction *);
88 void emitFMUL(const Instruction *);
89 void emitIMAD(const Instruction *);
90 void emitISAD(const Instruction *);
91 void emitFMAD(const Instruction *);
92
93 void emitNOT(const Instruction *);
94 void emitLogicOp(const Instruction *, uint8_t subOp);
95 void emitPOPC(const Instruction *);
96 void emitINSBF(const Instruction *);
97 void emitShift(const Instruction *);
98
99 void emitSFnOp(const Instruction *, uint8_t subOp);
100
101 void emitCVT(const Instruction *);
102 void emitMINMAX(const Instruction *);
103 void emitPreOp(const Instruction *);
104
105 void emitSET(const CmpInstruction *);
106 void emitSLCT(const CmpInstruction *);
107 void emitSELP(const Instruction *);
108
109 void emitTEXBAR(const Instruction *);
110 void emitTEX(const TexInstruction *);
111 void emitTEXCSAA(const TexInstruction *);
112 void emitTXQ(const TexInstruction *);
113
114 void emitQUADOP(const Instruction *, uint8_t qOp, uint8_t laneMask);
115
116 void emitFlow(const Instruction *);
117
118 inline void defId(const ValueDef&, const int pos);
119 inline void srcId(const ValueRef&, const int pos);
120 inline void srcId(const ValueRef *, const int pos);
121 inline void srcId(const Instruction *, int s, const int pos);
122
123 inline void srcAddr32(const ValueRef&, const int pos); // address / 4
124
125 inline bool isLIMM(const ValueRef&, DataType ty, bool mod = false);
126 };
127
128 #define GK110_GPR_ZERO 255
129
130 #define NEG_(b, s) \
131 if (i->src(s).mod.neg()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
132 #define ABS_(b, s) \
133 if (i->src(s).mod.abs()) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
134
135 #define NOT_(b, s) if (i->src(s).mod & Modifier(NV50_IR_MOD_NOT)) \
136 code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
137
138 #define FTZ_(b) if (i->ftz) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
139
140 #define SAT_(b) if (i->saturate) code[(0x##b) / 32] |= 1 << ((0x##b) % 32)
141
142 #define RND_(b, t) emitRoundMode##t(i->rnd, 0x##b)
143
144 #define SDATA(a) ((a).rep()->reg.data)
145 #define DDATA(a) ((a).rep()->reg.data)
146
147 void CodeEmitterGK110::srcId(const ValueRef& src, const int pos)
148 {
149 code[pos / 32] |= (src.get() ? SDATA(src).id : GK110_GPR_ZERO) << (pos % 32);
150 }
151
152 void CodeEmitterGK110::srcId(const ValueRef *src, const int pos)
153 {
154 code[pos / 32] |= (src ? SDATA(*src).id : GK110_GPR_ZERO) << (pos % 32);
155 }
156
157 void CodeEmitterGK110::srcId(const Instruction *insn, int s, int pos)
158 {
159 int r = insn->srcExists(s) ? SDATA(insn->src(s)).id : GK110_GPR_ZERO;
160 code[pos / 32] |= r << (pos % 32);
161 }
162
163 void CodeEmitterGK110::srcAddr32(const ValueRef& src, const int pos)
164 {
165 code[pos / 32] |= (SDATA(src).offset >> 2) << (pos % 32);
166 }
167
168 void CodeEmitterGK110::defId(const ValueDef& def, const int pos)
169 {
170 code[pos / 32] |= (def.get() ? DDATA(def).id : GK110_GPR_ZERO) << (pos % 32);
171 }
172
173 bool CodeEmitterGK110::isLIMM(const ValueRef& ref, DataType ty, bool mod)
174 {
175 const ImmediateValue *imm = ref.get()->asImm();
176
177 return imm && (imm->reg.data.u32 & ((ty == TYPE_F32) ? 0xfff : 0xfff00000));
178 }
179
180 void
181 CodeEmitterGK110::emitRoundMode(RoundMode rnd, const int pos, const int rintPos)
182 {
183 bool rint = false;
184 uint8_t n;
185
186 switch (rnd) {
187 case ROUND_MI: rint = true; /* fall through */ case ROUND_M: n = 1; break;
188 case ROUND_PI: rint = true; /* fall through */ case ROUND_P: n = 2; break;
189 case ROUND_ZI: rint = true; /* fall through */ case ROUND_Z: n = 3; break;
190 default:
191 rint = rnd == ROUND_NI;
192 n = 0;
193 assert(rnd == ROUND_N || rnd == ROUND_NI);
194 break;
195 }
196 code[pos / 32] |= n << (pos % 32);
197 if (rint && rintPos >= 0)
198 code[rintPos / 32] |= 1 << (rintPos % 32);
199 }
200
201 void
202 CodeEmitterGK110::emitRoundModeF(RoundMode rnd, const int pos)
203 {
204 uint8_t n;
205
206 switch (rnd) {
207 case ROUND_M: n = 1; break;
208 case ROUND_P: n = 2; break;
209 case ROUND_Z: n = 3; break;
210 default:
211 n = 0;
212 assert(rnd == ROUND_N);
213 break;
214 }
215 code[pos / 32] |= n << (pos % 32);
216 }
217
218 void
219 CodeEmitterGK110::emitRoundModeI(RoundMode rnd, const int pos)
220 {
221 uint8_t n;
222
223 switch (rnd) {
224 case ROUND_MI: n = 1; break;
225 case ROUND_PI: n = 2; break;
226 case ROUND_ZI: n = 3; break;
227 default:
228 n = 0;
229 assert(rnd == ROUND_NI);
230 break;
231 }
232 code[pos / 32] |= n << (pos % 32);
233 }
234
235 void CodeEmitterGK110::emitCondCode(CondCode cc, int pos, uint8_t mask)
236 {
237 uint8_t n;
238
239 switch (cc) {
240 case CC_FL: n = 0x00; break;
241 case CC_LT: n = 0x01; break;
242 case CC_EQ: n = 0x02; break;
243 case CC_LE: n = 0x03; break;
244 case CC_GT: n = 0x04; break;
245 case CC_NE: n = 0x05; break;
246 case CC_GE: n = 0x06; break;
247 case CC_LTU: n = 0x09; break;
248 case CC_EQU: n = 0x0a; break;
249 case CC_LEU: n = 0x0b; break;
250 case CC_GTU: n = 0x0c; break;
251 case CC_NEU: n = 0x0d; break;
252 case CC_GEU: n = 0x0e; break;
253 case CC_TR: n = 0x0f; break;
254 case CC_NO: n = 0x10; break;
255 case CC_NC: n = 0x11; break;
256 case CC_NS: n = 0x12; break;
257 case CC_NA: n = 0x13; break;
258 case CC_A: n = 0x14; break;
259 case CC_S: n = 0x15; break;
260 case CC_C: n = 0x16; break;
261 case CC_O: n = 0x17; break;
262 default:
263 n = 0;
264 assert(!"invalid condition code");
265 break;
266 }
267 code[pos / 32] |= (n & mask) << (pos % 32);
268 }
269
270 void
271 CodeEmitterGK110::emitPredicate(const Instruction *i)
272 {
273 if (i->predSrc >= 0) {
274 srcId(i->src(i->predSrc), 18);
275 if (i->cc == CC_NOT_P)
276 code[0] |= 8 << 18; // negate
277 assert(i->getPredicate()->reg.file == FILE_PREDICATE);
278 } else {
279 code[0] |= 7 << 18;
280 }
281 }
282
283 void
284 CodeEmitterGK110::setCAddress14(const ValueRef& src)
285 {
286 const int32_t addr = src.get()->asSym()->reg.data.offset / 4;
287
288 code[0] |= (addr & 0x01ff) << 23;
289 code[1] |= (addr & 0x3e00) >> 9;
290 }
291
292 void
293 CodeEmitterGK110::setShortImmediate(const Instruction *i, const int s)
294 {
295 const uint32_t u32 = i->getSrc(s)->asImm()->reg.data.u32;
296 const uint64_t u64 = i->getSrc(s)->asImm()->reg.data.u64;
297
298 if (i->sType == TYPE_F32) {
299 assert(!(u32 & 0x00000fff));
300 code[0] |= ((u32 & 0x001ff000) >> 12) << 23;
301 code[1] |= ((u32 & 0x7fe00000) >> 21);
302 code[1] |= ((u32 & 0x80000000) >> 4);
303 } else
304 if (i->sType == TYPE_F64) {
305 assert(!(u64 & 0x00000fffffffffffULL));
306 code[0] |= ((u64 & 0x001ff00000000000ULL) >> 44) << 23;
307 code[1] |= ((u64 & 0x7fe0000000000000ULL) >> 53);
308 code[1] |= ((u64 & 0x8000000000000000ULL) >> 36);
309 } else {
310 assert((u32 & 0xfff00000) == 0 || (u32 & 0xfff00000) == 0xfff00000);
311 code[0] |= (u32 & 0x001ff) << 23;
312 code[1] |= (u32 & 0x7fe00) >> 9;
313 code[1] |= (u32 & 0x80000) << 8;
314 }
315 }
316
317 void
318 CodeEmitterGK110::setImmediate32(const Instruction *i, const int s,
319 Modifier mod)
320 {
321 uint32_t u32 = i->getSrc(s)->asImm()->reg.data.u32;
322
323 if (mod) {
324 ImmediateValue imm(i->getSrc(s)->asImm(), i->sType);
325 mod.applyTo(imm);
326 u32 = imm.reg.data.u32;
327 }
328
329 code[0] |= u32 << 23;
330 code[1] |= u32 >> 9;
331 }
332
333 void
334 CodeEmitterGK110::emitForm_L(const Instruction *i, uint32_t opc, uint8_t ctg,
335 Modifier mod)
336 {
337 code[0] = ctg;
338 code[1] = opc << 20;
339
340 emitPredicate(i);
341
342 defId(i->def(0), 2);
343
344 for (int s = 0; s < 3 && i->srcExists(s); ++s) {
345 switch (i->src(s).getFile()) {
346 case FILE_GPR:
347 srcId(i->src(s), s ? 42 : 10);
348 break;
349 case FILE_IMMEDIATE:
350 setImmediate32(i, s, mod);
351 break;
352 default:
353 break;
354 }
355 }
356 }
357
358
359 void
360 CodeEmitterGK110::emitForm_C(const Instruction *i, uint32_t opc, uint8_t ctg)
361 {
362 code[0] = ctg;
363 code[1] = opc << 20;
364
365 emitPredicate(i);
366
367 defId(i->def(0), 2);
368
369 switch (i->src(0).getFile()) {
370 case FILE_MEMORY_CONST:
371 code[1] |= 0x4 << 28;
372 setCAddress14(i->src(0));
373 break;
374 case FILE_GPR:
375 code[1] |= 0xc << 28;
376 srcId(i->src(0), 23);
377 break;
378 default:
379 assert(0);
380 break;
381 }
382 }
383
384 // 0x2 for GPR, c[] and 0x1 for short immediate
385 void
386 CodeEmitterGK110::emitForm_21(const Instruction *i, uint32_t opc2,
387 uint32_t opc1)
388 {
389 const bool imm = i->srcExists(1) && i->src(1).getFile() == FILE_IMMEDIATE;
390
391 int s1 = 23;
392 if (i->srcExists(2) && i->src(2).getFile() == FILE_MEMORY_CONST)
393 s1 = 42;
394
395 if (imm) {
396 code[0] = 0x1;
397 code[1] = opc1 << 20;
398 } else {
399 code[0] = 0x2;
400 code[1] = (0xc << 28) | (opc2 << 20);
401 }
402
403 emitPredicate(i);
404
405 defId(i->def(0), 2);
406
407 for (int s = 0; s < 3 && i->srcExists(s); ++s) {
408 switch (i->src(s).getFile()) {
409 case FILE_MEMORY_CONST:
410 code[1] &= (s == 2) ? ~(0x4 << 28) : ~(0x8 << 28);
411 setCAddress14(i->src(s));
412 code[1] |= i->getSrc(s)->reg.fileIndex << 5;
413 break;
414 case FILE_IMMEDIATE:
415 setShortImmediate(i, s);
416 break;
417 case FILE_GPR:
418 srcId(i->src(s), s ? ((s == 2) ? 42 : s1) : 10);
419 break;
420 default:
421 // ignore here, can be predicate or flags, but must not be address
422 break;
423 }
424 }
425 // 0x0 = invalid
426 // 0xc = rrr
427 // 0x8 = rrc
428 // 0x4 = rcr
429 assert(imm || (code[1] & (0xc << 28)));
430 }
431
432 inline void
433 CodeEmitterGK110::modNegAbsF32_3b(const Instruction *i, const int s)
434 {
435 if (i->src(s).mod.abs()) code[1] &= ~(1 << 27);
436 if (i->src(s).mod.neg()) code[1] ^= (1 << 27);
437 }
438
439 void
440 CodeEmitterGK110::emitNOP(const Instruction *i)
441 {
442 code[0] = 0x00003c02;
443 code[1] = 0x85800000;
444
445 if (i)
446 emitPredicate(i);
447 else
448 code[0] = 0x001c3c02;
449 }
450
451 void
452 CodeEmitterGK110::emitFMAD(const Instruction *i)
453 {
454 assert(!isLIMM(i->src(1), TYPE_F32));
455
456 emitForm_21(i, 0x0c0, 0x940);
457
458 NEG_(34, 2);
459 SAT_(35);
460 RND_(36, F);
461 FTZ_(38);
462
463 bool neg1 = (i->src(0).mod ^ i->src(1).mod).neg();
464
465 if (code[0] & 0x1) {
466 if (neg1)
467 code[1] ^= 1 << 27;
468 } else
469 if (neg1) {
470 code[1] |= 1 << 19;
471 }
472 }
473
474 void
475 CodeEmitterGK110::emitFMUL(const Instruction *i)
476 {
477 bool neg = (i->src(0).mod ^ i->src(1).mod).neg();
478
479 assert(i->postFactor >= -3 && i->postFactor <= 3);
480
481 if (isLIMM(i->src(1), TYPE_F32)) {
482 emitForm_L(i, 0x200, 0x2, Modifier(0));
483
484 FTZ_(38);
485 SAT_(3a);
486 if (neg)
487 code[1] ^= 1 << 22;
488
489 assert(i->postFactor == 0);
490 } else {
491 emitForm_21(i, 0x234, 0xc34);
492
493 RND_(2a, F);
494 FTZ_(2f);
495 SAT_(35);
496
497 if (code[0] & 0x1) {
498 if (neg)
499 code[1] ^= 1 << 27;
500 } else
501 if (neg) {
502 code[1] |= 1 << 19;
503 }
504 }
505 }
506
507 void
508 CodeEmitterGK110::emitIMUL(const Instruction *i)
509 {
510 assert(!i->src(0).mod.neg() && !i->src(1).mod.neg());
511 assert(!i->src(0).mod.abs() && !i->src(1).mod.abs());
512
513 if (isLIMM(i->src(1), TYPE_S32)) {
514 emitForm_L(i, 0x280, 2, Modifier(0));
515
516 assert(i->subOp != NV50_IR_SUBOP_MUL_HIGH);
517
518 if (i->sType == TYPE_S32)
519 code[1] |= 3 << 25;
520 } else {
521 emitForm_21(i, 0x21c, 0xc1c);
522
523 if (i->subOp == NV50_IR_SUBOP_MUL_HIGH)
524 code[1] |= 1 << 10;
525 if (i->sType == TYPE_S32)
526 code[1] |= 3 << 11;
527 }
528 }
529
530 void
531 CodeEmitterGK110::emitFADD(const Instruction *i)
532 {
533 if (isLIMM(i->src(1), TYPE_F32)) {
534 assert(i->rnd == ROUND_N);
535 assert(!i->saturate);
536
537 emitForm_L(i, 0x400, 0, i->src(1).mod);
538
539 FTZ_(3a);
540 NEG_(3b, 0);
541 ABS_(39, 0);
542 } else {
543 emitForm_21(i, 0x22c, 0xc2c);
544
545 FTZ_(2f);
546 RND_(2a, F);
547 ABS_(31, 0);
548 NEG_(33, 0);
549
550 if (code[0] & 0x1) {
551 modNegAbsF32_3b(i, 1);
552 } else {
553 ABS_(34, 1);
554 NEG_(30, 1);
555 }
556 }
557 }
558
559 void
560 CodeEmitterGK110::emitUADD(const Instruction *i)
561 {
562 uint8_t addOp = (i->src(0).mod.neg() << 1) | i->src(1).mod.neg();
563
564 if (i->op == OP_SUB)
565 addOp ^= 1;
566
567 assert(!i->src(0).mod.abs() && !i->src(1).mod.abs());
568
569 if (isLIMM(i->src(1), TYPE_S32)) {
570 emitForm_L(i, 0x400, 1, Modifier((addOp & 1) ? NV50_IR_MOD_NEG : 0));
571
572 if (addOp & 2)
573 code[1] |= 1 << 27;
574
575 assert(!i->defExists(1));
576 assert(i->flagsSrc < 0);
577
578 SAT_(39);
579 } else {
580 emitForm_21(i, 0x208, 0xc08);
581
582 assert(addOp != 3); // would be add-plus-one
583
584 code[1] |= addOp << 19;
585
586 if (i->defExists(1))
587 code[1] |= 1 << 18; // write carry
588 if (i->flagsSrc >= 0)
589 code[1] |= 1 << 14; // add carry
590
591 SAT_(35);
592 }
593 }
594
595 // TODO: shl-add
596 void
597 CodeEmitterGK110::emitIMAD(const Instruction *i)
598 {
599 uint8_t addOp =
600 (i->src(2).mod.neg() << 1) | (i->src(0).mod.neg() ^ i->src(1).mod.neg());
601
602 emitForm_21(i, 0x100, 0xa00);
603
604 assert(addOp != 3);
605 code[1] |= addOp << 26;
606
607 if (i->sType == TYPE_S32)
608 code[1] |= (1 << 19) | (1 << 24);
609
610 if (code[0] & 0x1) {
611 assert(!i->subOp);
612 SAT_(39);
613 } else {
614 if (i->subOp == NV50_IR_SUBOP_MUL_HIGH)
615 code[1] |= 1 << 25;
616 SAT_(35);
617 }
618 }
619
620 void
621 CodeEmitterGK110::emitISAD(const Instruction *i)
622 {
623 assert(i->dType == TYPE_S32 || i->dType == TYPE_U32);
624
625 emitForm_21(i, 0x1fc, 0xb74);
626
627 if (i->dType == TYPE_S32)
628 code[1] |= 1 << 19;
629 }
630
631 void
632 CodeEmitterGK110::emitNOT(const Instruction *i)
633 {
634 code[0] = 0x0003fc02; // logop(mov2) dst, 0, not src
635 code[1] = 0x22003800;
636
637 emitPredicate(i);
638
639 defId(i->def(0), 2);
640
641 switch (i->src(0).getFile()) {
642 case FILE_GPR:
643 code[1] |= 0xc << 28;
644 srcId(i->src(0), 23);
645 break;
646 case FILE_MEMORY_CONST:
647 code[1] |= 0x4 << 28;
648 setCAddress14(i->src(1));
649 break;
650 default:
651 assert(0);
652 break;
653 }
654 }
655
656 void
657 CodeEmitterGK110::emitLogicOp(const Instruction *i, uint8_t subOp)
658 {
659 assert(!(i->src(0).mod & Modifier(NV50_IR_MOD_NOT))); // XXX: find me
660
661 if (isLIMM(i->src(1), TYPE_S32)) {
662 emitForm_L(i, 0x200, 0, i->src(1).mod);
663 code[1] |= subOp << 24;
664 } else {
665 emitForm_21(i, 0x220, 0xc20);
666 code[1] |= subOp << 12;
667 NOT_(2b, 1);
668 }
669 assert(!(code[0] & 0x1) || !(i->src(1).mod & Modifier(NV50_IR_MOD_NOT)));
670 }
671
672 void
673 CodeEmitterGK110::emitPOPC(const Instruction *i)
674 {
675 assert(!isLIMM(i->src(1), TYPE_S32, true));
676
677 emitForm_21(i, 0x204, 0xc04);
678
679 NOT_(2a, 0);
680 if (!(code[0] & 0x1))
681 NOT_(2b, 1);
682 }
683
684 void
685 CodeEmitterGK110::emitINSBF(const Instruction *i)
686 {
687 emitForm_21(i, 0x1f8, 0xb78);
688 }
689
690 void
691 CodeEmitterGK110::emitShift(const Instruction *i)
692 {
693 const bool sar = i->op == OP_SHR && isSignedType(i->sType);
694
695 if (sar) {
696 emitForm_21(i, 0x214, 0x014);
697 code[1] |= 1 << 19;
698 } else
699 if (i->op == OP_SHR) {
700 // this is actually RSHF
701 emitForm_21(i, 0x27c, 0x87c);
702 code[1] |= GK110_GPR_ZERO << 10;
703 } else {
704 // this is actually LSHF
705 emitForm_21(i, 0x1fc, 0xb7c);
706 code[1] |= GK110_GPR_ZERO << 10;
707 }
708
709 if (i->subOp == NV50_IR_SUBOP_SHIFT_WRAP) {
710 if (!sar)
711 code[1] |= 1 << 21;
712 // XXX: find wrap modifier for SHR S32
713 }
714 }
715
716 void
717 CodeEmitterGK110::emitPreOp(const Instruction *i)
718 {
719 emitForm_21(i, 0x248, -1);
720
721 if (i->op == OP_PREEX2)
722 code[1] |= 1 << 10;
723
724 NEG_(30, 0);
725 ABS_(34, 0);
726 }
727
728 void
729 CodeEmitterGK110::emitSFnOp(const Instruction *i, uint8_t subOp)
730 {
731 code[0] = 0x00000002 | (subOp << 23);
732 code[1] = 0x84000000;
733
734 emitPredicate(i);
735
736 defId(i->def(0), 2);
737 srcId(i->src(0), 10);
738
739 NEG_(33, 0);
740 ABS_(31, 0);
741
742 // XXX: find saturate
743 }
744
745 void
746 CodeEmitterGK110::emitMINMAX(const Instruction *i)
747 {
748 uint32_t op2, op1;
749
750 switch (i->dType) {
751 case TYPE_U32:
752 case TYPE_S32:
753 op2 = 0x210;
754 op1 = 0xc10;
755 break;
756 case TYPE_F32:
757 op2 = 0x230;
758 op1 = 0xc30;
759 break;
760 case TYPE_F64:
761 op2 = 0x228;
762 op1 = 0xc28;
763 break;
764 default:
765 assert(0);
766 op2 = 0;
767 op1 = 0;
768 break;
769 }
770 emitForm_21(i, op2, op1);
771
772 if (i->dType == TYPE_S32)
773 code[1] |= 1 << 19;
774 code[1] |= (i->op == OP_MIN) ? 0x1c00 : 0x3c00; // [!]pt
775
776 FTZ_(2f);
777 ABS_(31, 0);
778 NEG_(33, 0);
779 if (code[0] & 0x1) {
780 modNegAbsF32_3b(i, 1);
781 } else {
782 ABS_(34, 1);
783 NEG_(30, 1);
784 }
785 }
786
787 void
788 CodeEmitterGK110::emitCVT(const Instruction *i)
789 {
790 const bool f2f = isFloatType(i->dType) && isFloatType(i->sType);
791 const bool f2i = !isFloatType(i->dType) && isFloatType(i->sType);
792 const bool i2f = isFloatType(i->dType) && !isFloatType(i->sType);
793
794 bool sat = i->saturate;
795 bool abs = i->src(0).mod.abs();
796 bool neg = i->src(0).mod.neg();
797
798 RoundMode rnd = i->rnd;
799
800 switch (i->op) {
801 case OP_CEIL: rnd = f2f ? ROUND_PI : ROUND_P; break;
802 case OP_FLOOR: rnd = f2f ? ROUND_MI : ROUND_M; break;
803 case OP_TRUNC: rnd = f2f ? ROUND_ZI : ROUND_Z; break;
804 case OP_SAT: sat = true; break;
805 case OP_NEG: neg = !neg; break;
806 case OP_ABS: abs = true; neg = false; break;
807 default:
808 break;
809 }
810
811 uint32_t op;
812
813 if (f2f) op = 0x254;
814 else if (f2i) op = 0x258;
815 else if (i2f) op = 0x25c;
816 else op = 0x260;
817
818 emitForm_C(i, op, 0x2);
819
820 FTZ_(2f);
821 if (neg) code[1] |= 1 << 16;
822 if (abs) code[1] |= 1 << 20;
823 if (sat) code[1] |= 1 << 21;
824
825 emitRoundMode(rnd, 32 + 10, f2f ? (32 + 13) : -1);
826
827 code[0] |= typeSizeofLog2(i->dType) << 10;
828 code[0] |= typeSizeofLog2(i->sType) << 12;
829
830 if (isSignedIntType(i->dType))
831 code[0] |= 0x4000;
832 if (isSignedIntType(i->sType))
833 code[0] |= 0x8000;
834 }
835
836 void
837 CodeEmitterGK110::emitSET(const CmpInstruction *i)
838 {
839 uint16_t op1, op2;
840
841 if (i->def(0).getFile() == FILE_PREDICATE) {
842 switch (i->sType) {
843 case TYPE_F32: op2 = 0x1d8; op1 = 0xb58; break;
844 case TYPE_F64: op2 = 0x1c0; op1 = 0xb40; break;
845 default:
846 op2 = 0x1b0;
847 op1 = 0xb30;
848 break;
849 }
850 emitForm_21(i, op2, op1);
851
852 NEG_(2e, 0);
853 ABS_(9, 0);
854 if (!(code[0] & 0x1)) {
855 NEG_(8, 1);
856 ABS_(2f, 1);
857 } else {
858 modNegAbsF32_3b(i, 1);
859 }
860 FTZ_(32);
861
862 // normal DST field is negated predicate result
863 code[0] = (code[0] & ~0xfc) | ((code[0] << 3) & 0xe0);
864 if (i->defExists(1))
865 defId(i->def(1), 2);
866 else
867 code[0] |= 0x1c;
868 } else {
869 switch (i->sType) {
870 case TYPE_F32: op2 = 0x000; op1 = 0x820; break;
871 case TYPE_F64: op2 = 0x080; op1 = 0x900; break;
872 default:
873 op2 = 0x1a8;
874 op1 = 0xb28;
875 break;
876 }
877 emitForm_21(i, op2, op1);
878
879 NEG_(2e, 0);
880 ABS_(39, 0);
881 if (!(code[0] & 0x1)) {
882 NEG_(38, 1);
883 ABS_(2f, 1);
884 } else {
885 modNegAbsF32_3b(i, 1);
886 }
887 FTZ_(3a);
888 }
889 if (i->sType == TYPE_S32)
890 code[1] |= 1 << 19;
891
892 if (i->op != OP_SET) {
893 switch (i->op) {
894 case OP_SET_AND: code[1] |= 0x0 << 16; break;
895 case OP_SET_OR: code[1] |= 0x1 << 16; break;
896 case OP_SET_XOR: code[1] |= 0x2 << 16; break;
897 default:
898 assert(0);
899 break;
900 }
901 srcId(i->src(2), 0x2a);
902 } else {
903 code[1] |= 0x7 << 10;
904 }
905 emitCondCode(i->setCond,
906 isFloatType(i->sType) ? 0x33 : 0x34,
907 isFloatType(i->sType) ? 0xf : 0x7);
908 }
909
910 void
911 CodeEmitterGK110::emitSLCT(const CmpInstruction *i)
912 {
913 CondCode cc = i->setCond;
914 if (i->src(2).mod.neg())
915 cc = reverseCondCode(cc);
916
917 if (i->dType == TYPE_F32) {
918 emitForm_21(i, 0x1d0, 0xb50);
919 FTZ_(32);
920 emitCondCode(cc, 0x33, 0xf);
921 } else {
922 emitForm_21(i, 0x1a4, 0xb20);
923 emitCondCode(cc, 0x34, 0x7);
924 }
925 }
926
927 void CodeEmitterGK110::emitSELP(const Instruction *i)
928 {
929 emitForm_21(i, 0x250, 0x050);
930
931 if ((i->cc == CC_NOT_P) ^ (bool)(i->src(2).mod & Modifier(NV50_IR_MOD_NOT)))
932 code[1] |= 1 << 13;
933 }
934
935 void CodeEmitterGK110::emitTEXBAR(const Instruction *i)
936 {
937 code[0] = 0x00000002 | (i->subOp << 23);
938 code[1] = 0x77000000;
939
940 emitPredicate(i);
941 }
942
943 void CodeEmitterGK110::emitTEXCSAA(const TexInstruction *i)
944 {
945 emitNOP(i); // TODO
946 }
947
948 static inline bool
949 isNextIndependentTex(const TexInstruction *i)
950 {
951 if (!i->next || !isTextureOp(i->next->op))
952 return false;
953 if (i->getDef(0)->interfers(i->next->getSrc(0)))
954 return false;
955 return !i->next->srcExists(1) || !i->getDef(0)->interfers(i->next->getSrc(1));
956 }
957
958 void
959 CodeEmitterGK110::emitTEX(const TexInstruction *i)
960 {
961 const bool ind = i->tex.rIndirectSrc >= 0;
962
963 if (ind) {
964 code[0] = 0x00000002;
965 switch (i->op) {
966 case OP_TXD:
967 code[1] = 0x7e000000;
968 break;
969 default:
970 code[1] = 0x7d800000;
971 break;
972 }
973 } else {
974 switch (i->op) {
975 case OP_TXD:
976 code[0] = 0x00000002;
977 code[1] = 0x76000000;
978 break;
979 default:
980 code[0] = 0x00000001;
981 code[1] = 0x60000000;
982 break;
983 }
984 code[1] |= i->tex.r << 15;
985 }
986
987 code[1] |= isNextIndependentTex(i) ? 0x1 : 0x2; // t : p mode
988
989 // if (i->tex.liveOnly)
990 // ?
991
992 switch (i->op) {
993 case OP_TEX: break;
994 case OP_TXB: code[1] |= 0x2000; break;
995 case OP_TXL: code[1] |= 0x3000; break;
996 case OP_TXF: break; // XXX
997 case OP_TXG: break; // XXX
998 case OP_TXD: break;
999 default:
1000 assert(!"invalid texture op");
1001 break;
1002 }
1003 /*
1004 if (i->op == OP_TXF) {
1005 if (!i->tex.levelZero)
1006 code[1] |= 0x02000000;
1007 } else */
1008 if (i->tex.levelZero) {
1009 code[1] |= 0x1000;
1010 }
1011
1012 // if (i->op != OP_TXD && i->tex.derivAll)
1013 // code[1] |= 1 << 13;
1014
1015 emitPredicate(i);
1016
1017 code[1] |= i->tex.mask << 2;
1018
1019 const int src1 = (i->predSrc == 1) ? 2 : 1; // if predSrc == 1, !srcExists(2)
1020
1021 defId(i->def(0), 2);
1022 srcId(i->src(0), 10);
1023 srcId(i, src1, 23);
1024
1025 // if (i->op == OP_TXG) code[0] |= i->tex.gatherComp << 5;
1026
1027 // texture target:
1028 code[1] |= (i->tex.target.isCube() ? 3 : (i->tex.target.getDim() - 1)) << 7;
1029 if (i->tex.target.isArray())
1030 code[1] |= 0x40;
1031 // if (i->tex.target.isShadow())
1032 // ?
1033 // if (i->tex.target == TEX_TARGET_2D_MS ||
1034 // i->tex.target == TEX_TARGET_2D_MS_ARRAY)
1035 // ?
1036
1037 if (i->srcExists(src1) && i->src(src1).getFile() == FILE_IMMEDIATE) {
1038 // ?
1039 }
1040
1041 // if (i->tex.useOffsets)
1042 // ?
1043 }
1044
1045 void
1046 CodeEmitterGK110::emitTXQ(const TexInstruction *i)
1047 {
1048 emitNOP(i); // TODO
1049 }
1050
1051 void
1052 CodeEmitterGK110::emitQUADOP(const Instruction *i, uint8_t qOp, uint8_t laneMask)
1053 {
1054 emitNOP(i); // TODO
1055 }
1056
1057 void
1058 CodeEmitterGK110::emitFlow(const Instruction *i)
1059 {
1060 const FlowInstruction *f = i->asFlow();
1061
1062 unsigned mask; // bit 0: predicate, bit 1: target
1063
1064 code[0] = 0x00000000;
1065
1066 switch (i->op) {
1067 case OP_BRA:
1068 code[1] = f->absolute ? 0x00000 : 0x12000000; // XXX
1069 // if (i->srcExists(0) && i->src(0).getFile() == FILE_MEMORY_CONST)
1070 // code[0] |= 0x4000;
1071 mask = 3;
1072 break;
1073 case OP_CALL:
1074 code[1] = f->absolute ? 0x00000 : 0x13000000; // XXX
1075 // if (i->srcExists(0) && i->src(0).getFile() == FILE_MEMORY_CONST)
1076 // code[0] |= 0x4000;
1077 mask = 2;
1078 break;
1079
1080 case OP_EXIT: code[1] = 0x18000000; mask = 1; break;
1081 case OP_RET: code[1] = 0x19000000; mask = 1; break;
1082 case OP_DISCARD: code[1] = 0x19800000; mask = 1; break; // XXX: guess
1083 case OP_BREAK: code[1] = 0x1a800000; mask = 1; break; // XXX: guess
1084 case OP_CONT: code[1] = 0x1b000000; mask = 1; break; // XXX: guess
1085
1086 case OP_JOINAT: code[1] = 0x14800000; mask = 2; break;
1087 case OP_PREBREAK: code[1] = 0x15000000; mask = 2; break; // XXX: guess
1088 case OP_PRECONT: code[1] = 0x15800000; mask = 2; break; // XXX: guess
1089 case OP_PRERET: code[1] = 0x16000000; mask = 2; break; // XXX: guess
1090
1091 case OP_QUADON: code[1] = 0x1c000000; mask = 0; break; // XXX: guess
1092 case OP_QUADPOP: code[1] = 0x1c800000; mask = 0; break; // XXX: guess
1093 case OP_BRKPT: code[1] = 0x1d000000; mask = 0; break; // XXX: guess
1094 default:
1095 assert(!"invalid flow operation");
1096 return;
1097 }
1098
1099 if (mask & 1) {
1100 emitPredicate(i);
1101 if (i->flagsSrc < 0)
1102 code[0] |= 0x3c;
1103 }
1104
1105 if (!f)
1106 return;
1107
1108 // TODO
1109 /*
1110 if (f->allWarp)
1111 code[0] |= 1 << 15;
1112 if (f->limit)
1113 code[0] |= 1 << 16;
1114 */
1115
1116 if (f->op == OP_CALL) {
1117 if (f->builtin) {
1118 assert(f->absolute);
1119 uint32_t pcAbs = targNVC0->getBuiltinOffset(f->target.builtin);
1120 addReloc(RelocEntry::TYPE_BUILTIN, 0, pcAbs, 0xff800000, 23);
1121 addReloc(RelocEntry::TYPE_BUILTIN, 1, pcAbs, 0x007fffff, -9);
1122 } else {
1123 assert(!f->absolute);
1124 int32_t pcRel = f->target.fn->binPos - (codeSize + 8);
1125 code[0] |= (pcRel & 0x1ff) << 23;
1126 code[1] |= (pcRel >> 9) & 0x7fff;
1127 }
1128 } else
1129 if (mask & 2) {
1130 int32_t pcRel = f->target.bb->binPos - (codeSize + 8);
1131 // currently we don't want absolute branches
1132 assert(!f->absolute);
1133 code[0] |= (pcRel & 0x1ff) << 23;
1134 code[1] |= (pcRel >> 9) & 0x7fff;
1135 }
1136 }
1137
1138 void
1139 CodeEmitterGK110::emitPFETCH(const Instruction *i)
1140 {
1141 emitNOP(i); // TODO
1142 }
1143
1144 void
1145 CodeEmitterGK110::emitVFETCH(const Instruction *i)
1146 {
1147 uint32_t offset = i->src(0).get()->reg.data.offset;
1148
1149 code[0] = 0x00000002 | (offset << 23);
1150 code[1] = 0x7ec00000 | (offset >> 9);
1151
1152 #if 0
1153 if (i->perPatch)
1154 code[0] |= 0x100;
1155 if (i->getSrc(0)->reg.file == FILE_SHADER_OUTPUT)
1156 code[0] |= 0x200; // yes, TCPs can read from *outputs* of other threads
1157 #endif
1158
1159 emitPredicate(i);
1160
1161 defId(i->def(0), 2);
1162 srcId(i->src(0).getIndirect(0), 10);
1163 srcId(i->src(0).getIndirect(1), 32 + 10); // vertex address
1164 }
1165
1166 void
1167 CodeEmitterGK110::emitEXPORT(const Instruction *i)
1168 {
1169 uint32_t offset = i->src(0).get()->reg.data.offset;
1170
1171 code[0] = 0x00000002 | (offset << 23);
1172 code[1] = 0x7f000000 | (offset >> 9);
1173
1174 #if 0
1175 if (i->perPatch)
1176 code[0] |= 0x100;
1177 #endif
1178
1179 emitPredicate(i);
1180
1181 assert(i->src(1).getFile() == FILE_GPR);
1182
1183 srcId(i->src(0).getIndirect(0), 10);
1184 srcId(i->src(0).getIndirect(1), 32 + 10); // vertex base address
1185 srcId(i->src(1), 2);
1186 }
1187
1188 void
1189 CodeEmitterGK110::emitOUT(const Instruction *i)
1190 {
1191 emitNOP(i); // TODO
1192 }
1193
1194 void
1195 CodeEmitterGK110::emitInterpMode(const Instruction *i)
1196 {
1197 code[1] |= i->ipa << 21; // TODO: INTERP_SAMPLEID
1198 }
1199
1200 void
1201 CodeEmitterGK110::emitINTERP(const Instruction *i)
1202 {
1203 const uint32_t base = i->getSrc(0)->reg.data.offset;
1204
1205 code[0] = 0x00000002 | (base << 31);
1206 code[1] = 0x74800000 | (base >> 1);
1207
1208 if (i->saturate)
1209 code[1] |= 1 << 18;
1210
1211 if (i->op == OP_PINTERP)
1212 srcId(i->src(1), 23);
1213 else
1214 code[0] |= 0xff << 23;
1215
1216 srcId(i->src(0).getIndirect(0), 10);
1217 emitInterpMode(i);
1218
1219 emitPredicate(i);
1220 defId(i->def(0), 2);
1221
1222 if (i->getSampleMode() == NV50_IR_INTERP_OFFSET)
1223 srcId(i->src(i->op == OP_PINTERP ? 2 : 1), 32 + 10);
1224 else
1225 code[1] |= 0xff << 10;
1226 }
1227
1228 void
1229 CodeEmitterGK110::emitLoadStoreType(DataType ty, const int pos)
1230 {
1231 uint8_t n;
1232
1233 switch (ty) {
1234 case TYPE_U8:
1235 n = 0;
1236 break;
1237 case TYPE_S8:
1238 n = 1;
1239 break;
1240 case TYPE_U16:
1241 n = 2;
1242 break;
1243 case TYPE_S16:
1244 n = 3;
1245 break;
1246 case TYPE_F32:
1247 case TYPE_U32:
1248 case TYPE_S32:
1249 n = 4;
1250 break;
1251 case TYPE_F64:
1252 case TYPE_U64:
1253 case TYPE_S64:
1254 n = 5;
1255 break;
1256 case TYPE_B128:
1257 n = 6;
1258 break;
1259 default:
1260 n = 0;
1261 assert(!"invalid ld/st type");
1262 break;
1263 }
1264 code[pos / 32] |= n << (pos % 32);
1265 }
1266
1267 void
1268 CodeEmitterGK110::emitCachingMode(CacheMode c, const int pos)
1269 {
1270 uint8_t n;
1271
1272 switch (c) {
1273 case CACHE_CA:
1274 // case CACHE_WB:
1275 n = 0;
1276 break;
1277 case CACHE_CG:
1278 n = 1;
1279 break;
1280 case CACHE_CS:
1281 n = 2;
1282 break;
1283 case CACHE_CV:
1284 // case CACHE_WT:
1285 n = 3;
1286 break;
1287 default:
1288 n = 0;
1289 assert(!"invalid caching mode");
1290 break;
1291 }
1292 code[pos / 32] |= n << (pos % 32);
1293 }
1294
1295 void
1296 CodeEmitterGK110::emitSTORE(const Instruction *i)
1297 {
1298 int32_t offset = SDATA(i->src(0)).offset;
1299
1300 switch (i->src(0).getFile()) {
1301 case FILE_MEMORY_GLOBAL: code[1] = 0xe0000000; code[0] = 0x00000000; break;
1302 case FILE_MEMORY_LOCAL: code[1] = 0x7a800000; code[0] = 0x00000002; break;
1303 case FILE_MEMORY_SHARED: code[1] = 0x7ac00000; code[0] = 0x00000002; break;
1304 default:
1305 assert(!"invalid memory file");
1306 break;
1307 }
1308
1309 if (i->src(0).getFile() != FILE_MEMORY_GLOBAL)
1310 offset &= 0xffffff;
1311
1312 if (code[0] & 0x2) {
1313 emitLoadStoreType(i->dType, 0x33);
1314 if (i->src(0).getFile() == FILE_MEMORY_LOCAL)
1315 emitCachingMode(i->cache, 0x2f);
1316 } else {
1317 emitLoadStoreType(i->dType, 0x38);
1318 emitCachingMode(i->cache, 0x3b);
1319 }
1320 code[0] |= offset << 23;
1321 code[1] |= offset >> 9;
1322
1323 emitPredicate(i);
1324
1325 srcId(i->src(1), 2);
1326 srcId(i->src(0).getIndirect(0), 10);
1327 }
1328
1329 void
1330 CodeEmitterGK110::emitLOAD(const Instruction *i)
1331 {
1332 int32_t offset = SDATA(i->src(0)).offset;
1333
1334 switch (i->src(0).getFile()) {
1335 case FILE_MEMORY_GLOBAL: code[1] = 0xc0000000; code[0] = 0x00000000; break;
1336 case FILE_MEMORY_LOCAL: code[1] = 0x7a000000; code[0] = 0x00000002; break;
1337 case FILE_MEMORY_SHARED: code[1] = 0x7ac00000; code[0] = 0x00000002; break;
1338 case FILE_MEMORY_CONST:
1339 if (!i->src(0).isIndirect(0) && typeSizeof(i->dType) == 4) {
1340 emitMOV(i);
1341 return;
1342 }
1343 offset &= 0xffff;
1344 code[0] = 0x00000002;
1345 code[1] = 0x7c800000 | (i->src(0).get()->reg.fileIndex << 7);
1346 break;
1347 default:
1348 assert(!"invalid memory file");
1349 break;
1350 }
1351
1352 if (code[0] & 0x2) {
1353 offset &= 0xffffff;
1354 emitLoadStoreType(i->dType, 0x33);
1355 if (i->src(0).getFile() == FILE_MEMORY_LOCAL)
1356 emitCachingMode(i->cache, 0x2f);
1357 } else {
1358 emitLoadStoreType(i->dType, 0x38);
1359 emitCachingMode(i->cache, 0x3b);
1360 }
1361 code[0] |= offset << 23;
1362 code[1] |= offset >> 9;
1363
1364 emitPredicate(i);
1365
1366 defId(i->def(0), 2);
1367 srcId(i->src(0).getIndirect(0), 10);
1368 }
1369
1370 uint8_t
1371 CodeEmitterGK110::getSRegEncoding(const ValueRef& ref)
1372 {
1373 switch (SDATA(ref).sv.sv) {
1374 case SV_LANEID: return 0x00;
1375 case SV_PHYSID: return 0x03;
1376 case SV_VERTEX_COUNT: return 0x10;
1377 case SV_INVOCATION_ID: return 0x11;
1378 case SV_YDIR: return 0x12;
1379 case SV_TID: return 0x21 + SDATA(ref).sv.index;
1380 case SV_CTAID: return 0x25 + SDATA(ref).sv.index;
1381 case SV_NTID: return 0x29 + SDATA(ref).sv.index;
1382 case SV_GRIDID: return 0x2c;
1383 case SV_NCTAID: return 0x2d + SDATA(ref).sv.index;
1384 case SV_LBASE: return 0x34;
1385 case SV_SBASE: return 0x30;
1386 case SV_CLOCK: return 0x50 + SDATA(ref).sv.index;
1387 default:
1388 assert(!"no sreg for system value");
1389 return 0;
1390 }
1391 }
1392
1393 void
1394 CodeEmitterGK110::emitMOV(const Instruction *i)
1395 {
1396 if (i->src(0).getFile() == FILE_SYSTEM_VALUE) {
1397 code[0] = 0x00000002 | (getSRegEncoding(i->src(0)) << 23);
1398 code[1] = 0x86400000;
1399 emitPredicate(i);
1400 defId(i->def(0), 2);
1401 } else
1402 if (i->src(0).getFile() == FILE_IMMEDIATE) {
1403 code[0] = 0x00000002 | (i->lanes << 14);
1404 code[1] = 0x74000000;
1405 emitPredicate(i);
1406 defId(i->def(0), 2);
1407 setImmediate32(i, 0, Modifier(0));
1408 } else
1409 if (i->src(0).getFile() == FILE_PREDICATE) {
1410 // TODO
1411 } else {
1412 emitForm_C(i, 0x24c, 2);
1413 code[1] |= i->lanes << 10;
1414 }
1415 }
1416
1417 bool
1418 CodeEmitterGK110::emitInstruction(Instruction *insn)
1419 {
1420 const unsigned int size = (writeIssueDelays && !(codeSize & 0x3f)) ? 16 : 8;
1421
1422 if (insn->encSize != 8) {
1423 ERROR("skipping unencodable instruction: ");
1424 insn->print();
1425 return false;
1426 } else
1427 if (codeSize + size > codeSizeLimit) {
1428 ERROR("code emitter output buffer too small\n");
1429 return false;
1430 }
1431
1432 if (writeIssueDelays) {
1433 int id = (codeSize & 0x3f) / 8 - 1;
1434 if (id < 0) {
1435 id += 1;
1436 code[0] = 0x00000000; // cf issue delay "instruction"
1437 code[1] = 0x08000000;
1438 code += 2;
1439 codeSize += 8;
1440 }
1441 uint32_t *data = code - (id * 2 + 2);
1442
1443 switch (id) {
1444 case 0: data[0] |= insn->sched << 2; break;
1445 case 1: data[0] |= insn->sched << 10; break;
1446 case 2: data[0] |= insn->sched << 18; break;
1447 case 3: data[0] |= insn->sched << 26; data[1] |= insn->sched >> 6; break;
1448 case 4: data[1] |= insn->sched << 2;
1449 case 5: data[1] |= insn->sched << 10; break;
1450 case 6: data[1] |= insn->sched << 18; break;
1451 default:
1452 assert(0);
1453 break;
1454 }
1455 }
1456
1457 // assert that instructions with multiple defs don't corrupt registers
1458 for (int d = 0; insn->defExists(d); ++d)
1459 assert(insn->asTex() || insn->def(d).rep()->reg.data.id >= 0);
1460
1461 switch (insn->op) {
1462 case OP_MOV:
1463 case OP_RDSV:
1464 emitMOV(insn);
1465 break;
1466 case OP_NOP:
1467 break;
1468 case OP_LOAD:
1469 emitLOAD(insn);
1470 break;
1471 case OP_STORE:
1472 emitSTORE(insn);
1473 break;
1474 case OP_LINTERP:
1475 case OP_PINTERP:
1476 emitINTERP(insn);
1477 break;
1478 case OP_VFETCH:
1479 emitVFETCH(insn);
1480 break;
1481 case OP_EXPORT:
1482 emitEXPORT(insn);
1483 break;
1484 case OP_PFETCH:
1485 emitPFETCH(insn);
1486 break;
1487 case OP_EMIT:
1488 case OP_RESTART:
1489 emitOUT(insn);
1490 break;
1491 case OP_ADD:
1492 case OP_SUB:
1493 if (isFloatType(insn->dType))
1494 emitFADD(insn);
1495 else
1496 emitUADD(insn);
1497 break;
1498 case OP_MUL:
1499 if (isFloatType(insn->dType))
1500 emitFMUL(insn);
1501 else
1502 emitIMUL(insn);
1503 break;
1504 case OP_MAD:
1505 case OP_FMA:
1506 if (isFloatType(insn->dType))
1507 emitFMAD(insn);
1508 else
1509 emitIMAD(insn);
1510 break;
1511 case OP_SAD:
1512 emitISAD(insn);
1513 break;
1514 case OP_NOT:
1515 emitNOT(insn);
1516 break;
1517 case OP_AND:
1518 emitLogicOp(insn, 0);
1519 break;
1520 case OP_OR:
1521 emitLogicOp(insn, 1);
1522 break;
1523 case OP_XOR:
1524 emitLogicOp(insn, 2);
1525 break;
1526 case OP_SHL:
1527 case OP_SHR:
1528 emitShift(insn);
1529 break;
1530 case OP_SET:
1531 case OP_SET_AND:
1532 case OP_SET_OR:
1533 case OP_SET_XOR:
1534 emitSET(insn->asCmp());
1535 break;
1536 case OP_SELP:
1537 emitSELP(insn);
1538 break;
1539 case OP_SLCT:
1540 emitSLCT(insn->asCmp());
1541 break;
1542 case OP_MIN:
1543 case OP_MAX:
1544 emitMINMAX(insn);
1545 break;
1546 case OP_ABS:
1547 case OP_NEG:
1548 case OP_CEIL:
1549 case OP_FLOOR:
1550 case OP_TRUNC:
1551 case OP_CVT:
1552 case OP_SAT:
1553 emitCVT(insn);
1554 break;
1555 case OP_RSQ:
1556 emitSFnOp(insn, 5);
1557 break;
1558 case OP_RCP:
1559 emitSFnOp(insn, 4);
1560 break;
1561 case OP_LG2:
1562 emitSFnOp(insn, 3);
1563 break;
1564 case OP_EX2:
1565 emitSFnOp(insn, 2);
1566 break;
1567 case OP_SIN:
1568 emitSFnOp(insn, 1);
1569 break;
1570 case OP_COS:
1571 emitSFnOp(insn, 0);
1572 break;
1573 case OP_PRESIN:
1574 case OP_PREEX2:
1575 emitPreOp(insn);
1576 break;
1577 case OP_TEX:
1578 case OP_TXB:
1579 case OP_TXL:
1580 case OP_TXD:
1581 case OP_TXF:
1582 emitTEX(insn->asTex());
1583 break;
1584 case OP_TXQ:
1585 emitTXQ(insn->asTex());
1586 break;
1587 case OP_TEXBAR:
1588 emitTEXBAR(insn);
1589 break;
1590 case OP_BRA:
1591 case OP_CALL:
1592 case OP_PRERET:
1593 case OP_RET:
1594 case OP_DISCARD:
1595 case OP_EXIT:
1596 case OP_PRECONT:
1597 case OP_CONT:
1598 case OP_PREBREAK:
1599 case OP_BREAK:
1600 case OP_JOINAT:
1601 case OP_BRKPT:
1602 case OP_QUADON:
1603 case OP_QUADPOP:
1604 emitFlow(insn);
1605 break;
1606 case OP_QUADOP:
1607 emitQUADOP(insn, insn->subOp, insn->lanes);
1608 break;
1609 case OP_DFDX:
1610 emitQUADOP(insn, insn->src(0).mod.neg() ? 0x66 : 0x99, 0x4);
1611 break;
1612 case OP_DFDY:
1613 emitQUADOP(insn, insn->src(0).mod.neg() ? 0x5a : 0xa5, 0x5);
1614 break;
1615 case OP_POPCNT:
1616 emitPOPC(insn);
1617 break;
1618 case OP_JOIN:
1619 emitNOP(insn);
1620 insn->join = 1;
1621 break;
1622 case OP_PHI:
1623 case OP_UNION:
1624 case OP_CONSTRAINT:
1625 ERROR("operation should have been eliminated");
1626 return false;
1627 case OP_EXP:
1628 case OP_LOG:
1629 case OP_SQRT:
1630 case OP_POW:
1631 ERROR("operation should have been lowered\n");
1632 return false;
1633 default:
1634 ERROR("unknow op\n");
1635 return false;
1636 }
1637
1638 if (insn->join)
1639 code[0] |= 1 << 22;
1640
1641 code += 2;
1642 codeSize += 8;
1643 return true;
1644 }
1645
1646 uint32_t
1647 CodeEmitterGK110::getMinEncodingSize(const Instruction *i) const
1648 {
1649 // No more short instruction encodings.
1650 return 8;
1651 }
1652
1653 void
1654 CodeEmitterGK110::prepareEmission(Function *func)
1655 {
1656 const Target *targ = func->getProgram()->getTarget();
1657
1658 CodeEmitter::prepareEmission(func);
1659
1660 if (targ->hasSWSched)
1661 calculateSchedDataNVC0(targ, func);
1662 }
1663
1664 CodeEmitterGK110::CodeEmitterGK110(const TargetNVC0 *target)
1665 : CodeEmitter(target),
1666 targNVC0(target),
1667 writeIssueDelays(target->hasSWSched)
1668 {
1669 code = NULL;
1670 codeSize = codeSizeLimit = 0;
1671 relocInfo = NULL;
1672 }
1673
1674 CodeEmitter *
1675 TargetNVC0::createCodeEmitterGK110(Program::Type type)
1676 {
1677 CodeEmitterGK110 *emit = new CodeEmitterGK110(this);
1678 emit->setProgramType(type);
1679 return emit;
1680 }
1681
1682 } // namespace nv50_ir