• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2011 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.h"
24 #include "codegen/nv50_ir_build_util.h"
25 
26 #include "codegen/nv50_ir_target_nvc0.h"
27 #include "codegen/nv50_ir_lowering_nvc0.h"
28 
29 #include <limits>
30 
31 namespace nv50_ir {
32 
33 #define QOP_ADD  0
34 #define QOP_SUBR 1
35 #define QOP_SUB  2
36 #define QOP_MOV2 3
37 
38 //             UL UR LL LR
39 #define QUADOP(q, r, s, t)                      \
40    ((QOP_##q << 6) | (QOP_##r << 4) |           \
41     (QOP_##s << 2) | (QOP_##t << 0))
42 
43 void
handleDIV(Instruction * i)44 NVC0LegalizeSSA::handleDIV(Instruction *i)
45 {
46    FlowInstruction *call;
47    int builtin;
48 
49    bld.setPosition(i, false);
50 
51    // Generate movs to the input regs for the call we want to generate
52    for (int s = 0; i->srcExists(s); ++s) {
53       Instruction *ld = i->getSrc(s)->getInsn();
54       // check if we are moving an immediate, propagate it in that case
55       if (!ld || ld->fixed || (ld->op != OP_LOAD && ld->op != OP_MOV) ||
56             !(ld->src(0).getFile() == FILE_IMMEDIATE))
57          bld.mkMovToReg(s, i->getSrc(s));
58       else {
59          assert(ld->getSrc(0) != NULL);
60          bld.mkMovToReg(s, ld->getSrc(0));
61          // Clear the src, to make code elimination possible here before we
62          // delete the instruction i later
63          i->setSrc(s, NULL);
64          if (ld->isDead())
65             delete_Instruction(prog, ld);
66       }
67    }
68 
69    switch (i->dType) {
70    case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break;
71    case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break;
72    default:
73       return;
74    }
75    call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
76    bld.mkMovFromReg(i->getDef(0), i->op == OP_DIV ? 0 : 1);
77    bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2);
78    bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0);
79 
80    call->fixed = 1;
81    call->absolute = call->builtin = 1;
82    call->target.builtin = builtin;
83    delete_Instruction(prog, i);
84 }
85 
86 void
handleRCPRSQLib(Instruction * i,Value * src[])87 NVC0LegalizeSSA::handleRCPRSQLib(Instruction *i, Value *src[])
88 {
89    FlowInstruction *call;
90    Value *def[2];
91    int builtin;
92 
93    def[0] = bld.mkMovToReg(0, src[0])->getDef(0);
94    def[1] = bld.mkMovToReg(1, src[1])->getDef(0);
95 
96    if (i->op == OP_RCP)
97       builtin = NVC0_BUILTIN_RCP_F64;
98    else
99       builtin = NVC0_BUILTIN_RSQ_F64;
100 
101    call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
102    def[0] = bld.getSSA();
103    def[1] = bld.getSSA();
104    bld.mkMovFromReg(def[0], 0);
105    bld.mkMovFromReg(def[1], 1);
106    bld.mkClobber(FILE_GPR, 0x3fc, 2);
107    bld.mkClobber(FILE_PREDICATE, i->op == OP_RSQ ? 0x3 : 0x1, 0);
108    bld.mkOp2(OP_MERGE, TYPE_U64, i->getDef(0), def[0], def[1]);
109 
110    call->fixed = 1;
111    call->absolute = call->builtin = 1;
112    call->target.builtin = builtin;
113    delete_Instruction(prog, i);
114 
115    prog->fp64 = true;
116 }
117 
118 void
handleRCPRSQ(Instruction * i)119 NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
120 {
121    assert(i->dType == TYPE_F64);
122    // There are instructions that will compute the high 32 bits of the 64-bit
123    // float. We will just stick 0 in the bottom 32 bits.
124 
125    bld.setPosition(i, false);
126 
127    // 1. Take the source and it up.
128    Value *src[2], *dst[2], *def = i->getDef(0);
129    bld.mkSplit(src, 4, i->getSrc(0));
130 
131    int chip = prog->getTarget()->getChipset();
132    if (chip >= NVISA_GK104_CHIPSET) {
133       handleRCPRSQLib(i, src);
134       return;
135    }
136 
137    // 2. We don't care about the low 32 bits of the destination. Stick a 0 in.
138    dst[0] = bld.loadImm(NULL, 0);
139    dst[1] = bld.getSSA();
140 
141    // 3. The new version of the instruction takes the high 32 bits of the
142    // source and outputs the high 32 bits of the destination.
143    i->setSrc(0, src[1]);
144    i->setDef(0, dst[1]);
145    i->setType(TYPE_F32);
146    i->subOp = NV50_IR_SUBOP_RCPRSQ_64H;
147 
148    // 4. Recombine the two dst pieces back into the original destination.
149    bld.setPosition(i, true);
150    bld.mkOp2(OP_MERGE, TYPE_U64, def, dst[0], dst[1]);
151 }
152 
153 void
handleFTZ(Instruction * i)154 NVC0LegalizeSSA::handleFTZ(Instruction *i)
155 {
156    // Only want to flush float inputs
157    assert(i->sType == TYPE_F32);
158 
159    // If we're already flushing denorms (and NaN's) to zero, no need for this.
160    if (i->dnz)
161       return;
162 
163    // Only certain classes of operations can flush
164    OpClass cls = prog->getTarget()->getOpClass(i->op);
165    if (cls != OPCLASS_ARITH && cls != OPCLASS_COMPARE &&
166        cls != OPCLASS_CONVERT)
167       return;
168 
169    i->ftz = true;
170 }
171 
172 void
handleTEXLOD(TexInstruction * i)173 NVC0LegalizeSSA::handleTEXLOD(TexInstruction *i)
174 {
175    if (i->tex.levelZero)
176       return;
177 
178    ImmediateValue lod;
179 
180    // The LOD argument comes right after the coordinates (before depth bias,
181    // offsets, etc).
182    int arg = i->tex.target.getArgCount();
183 
184    // SM30+ stores the indirect handle as a separate arg, which comes before
185    // the LOD.
186    if (prog->getTarget()->getChipset() >= NVISA_GK104_CHIPSET &&
187        i->tex.rIndirectSrc >= 0)
188       arg++;
189    // SM20 stores indirect handle combined with array coordinate
190    if (prog->getTarget()->getChipset() < NVISA_GK104_CHIPSET &&
191        !i->tex.target.isArray() &&
192        i->tex.rIndirectSrc >= 0)
193       arg++;
194 
195    if (!i->src(arg).getImmediate(lod) || !lod.isInteger(0))
196       return;
197 
198    if (i->op == OP_TXL)
199       i->op = OP_TEX;
200    i->tex.levelZero = true;
201    i->moveSources(arg + 1, -1);
202 }
203 
204 void
handleShift(Instruction * lo)205 NVC0LegalizeSSA::handleShift(Instruction *lo)
206 {
207    Value *shift = lo->getSrc(1);
208    Value *dst64 = lo->getDef(0);
209    Value *src[2], *dst[2];
210    operation op = lo->op;
211 
212    bld.setPosition(lo, false);
213 
214    bld.mkSplit(src, 4, lo->getSrc(0));
215 
216    // SM30 and prior don't have the fancy new SHF.L/R ops. So the logic has to
217    // be completely emulated. For SM35+, we can use the more directed SHF
218    // operations.
219    if (prog->getTarget()->getChipset() < NVISA_GK20A_CHIPSET) {
220       // The strategy here is to handle shifts >= 32 and less than 32 as
221       // separate parts.
222       //
223       // For SHL:
224       // If the shift is <= 32, then
225       //   (HI,LO) << x = (HI << x | (LO >> (32 - x)), LO << x)
226       // If the shift is > 32, then
227       //   (HI,LO) << x = (LO << (x - 32), 0)
228       //
229       // For SHR:
230       // If the shift is <= 32, then
231       //   (HI,LO) >> x = (HI >> x, (HI << (32 - x)) | LO >> x)
232       // If the shift is > 32, then
233       //   (HI,LO) >> x = (0, HI >> (x - 32))
234       //
235       // Note that on NVIDIA hardware, a shift > 32 yields a 0 value, which we
236       // can use to our advantage. Also note the structural similarities
237       // between the right/left cases. The main difference is swapping hi/lo
238       // on input and output.
239 
240       Value *x32_minus_shift, *pred, *hi1, *hi2;
241       DataType type = isSignedIntType(lo->dType) ? TYPE_S32 : TYPE_U32;
242       operation antiop = op == OP_SHR ? OP_SHL : OP_SHR;
243       if (op == OP_SHR)
244          std::swap(src[0], src[1]);
245       bld.mkOp2(OP_ADD, TYPE_U32, (x32_minus_shift = bld.getSSA()), shift, bld.mkImm(0x20))
246          ->src(0).mod = Modifier(NV50_IR_MOD_NEG);
247       bld.mkCmp(OP_SET, CC_LE, TYPE_U8, (pred = bld.getSSA(1, FILE_PREDICATE)),
248                 TYPE_U32, shift, bld.mkImm(32));
249       // Compute HI (shift <= 32)
250       bld.mkOp2(OP_OR, TYPE_U32, (hi1 = bld.getSSA()),
251                 bld.mkOp2v(op, TYPE_U32, bld.getSSA(), src[1], shift),
252                 bld.mkOp2v(antiop, TYPE_U32, bld.getSSA(), src[0], x32_minus_shift))
253          ->setPredicate(CC_P, pred);
254       // Compute LO (all shift values)
255       bld.mkOp2(op, type, (dst[0] = bld.getSSA()), src[0], shift);
256       // Compute HI (shift > 32)
257       bld.mkOp2(op, type, (hi2 = bld.getSSA()), src[0],
258                 bld.mkOp1v(OP_NEG, TYPE_S32, bld.getSSA(), x32_minus_shift))
259          ->setPredicate(CC_NOT_P, pred);
260       bld.mkOp2(OP_UNION, TYPE_U32, (dst[1] = bld.getSSA()), hi1, hi2);
261       if (op == OP_SHR)
262          std::swap(dst[0], dst[1]);
263       bld.mkOp2(OP_MERGE, TYPE_U64, dst64, dst[0], dst[1]);
264       delete_Instruction(prog, lo);
265       return;
266    }
267 
268    Instruction *hi = new_Instruction(func, op, TYPE_U32);
269    lo->bb->insertAfter(lo, hi);
270 
271    hi->sType = lo->sType;
272    lo->dType = TYPE_U32;
273 
274    hi->setDef(0, (dst[1] = bld.getSSA()));
275    if (lo->op == OP_SHR)
276       hi->subOp |= NV50_IR_SUBOP_SHIFT_HIGH;
277    lo->setDef(0, (dst[0] = bld.getSSA()));
278 
279    bld.setPosition(hi, true);
280 
281    if (lo->op == OP_SHL)
282       std::swap(hi, lo);
283 
284    hi->setSrc(0, new_ImmediateValue(prog, 0u));
285    hi->setSrc(1, shift);
286    hi->setSrc(2, lo->op == OP_SHL ? src[0] : src[1]);
287 
288    lo->setSrc(0, src[0]);
289    lo->setSrc(1, shift);
290    lo->setSrc(2, src[1]);
291 
292    bld.mkOp2(OP_MERGE, TYPE_U64, dst64, dst[0], dst[1]);
293 }
294 
295 void
handleSET(CmpInstruction * cmp)296 NVC0LegalizeSSA::handleSET(CmpInstruction *cmp)
297 {
298    DataType hTy = cmp->sType == TYPE_S64 ? TYPE_S32 : TYPE_U32;
299    Value *carry;
300    Value *src0[2], *src1[2];
301    bld.setPosition(cmp, false);
302 
303    bld.mkSplit(src0, 4, cmp->getSrc(0));
304    bld.mkSplit(src1, 4, cmp->getSrc(1));
305    bld.mkOp2(OP_SUB, hTy, NULL, src0[0], src1[0])
306       ->setFlagsDef(0, (carry = bld.getSSA(1, FILE_FLAGS)));
307    cmp->setFlagsSrc(cmp->srcCount(), carry);
308    cmp->setSrc(0, src0[1]);
309    cmp->setSrc(1, src1[1]);
310    cmp->sType = hTy;
311 }
312 
313 void
handleBREV(Instruction * i)314 NVC0LegalizeSSA::handleBREV(Instruction *i)
315 {
316    i->op = OP_EXTBF;
317    i->subOp = NV50_IR_SUBOP_EXTBF_REV;
318    i->setSrc(1, bld.mkImm(0x2000));
319 }
320 
321 bool
visit(Function * fn)322 NVC0LegalizeSSA::visit(Function *fn)
323 {
324    bld.setProgram(fn->getProgram());
325    return true;
326 }
327 
328 bool
visit(BasicBlock * bb)329 NVC0LegalizeSSA::visit(BasicBlock *bb)
330 {
331    Instruction *next;
332    for (Instruction *i = bb->getEntry(); i; i = next) {
333       next = i->next;
334 
335       if (i->sType == TYPE_F32 && prog->getType() != Program::TYPE_COMPUTE)
336          handleFTZ(i);
337 
338       switch (i->op) {
339       case OP_DIV:
340       case OP_MOD:
341          if (i->sType != TYPE_F32)
342             handleDIV(i);
343          break;
344       case OP_RCP:
345       case OP_RSQ:
346          if (i->dType == TYPE_F64)
347             handleRCPRSQ(i);
348          break;
349       case OP_TXL:
350       case OP_TXF:
351          handleTEXLOD(i->asTex());
352          break;
353       case OP_SHR:
354       case OP_SHL:
355          if (typeSizeof(i->sType) == 8)
356             handleShift(i);
357          break;
358       case OP_SET:
359       case OP_SET_AND:
360       case OP_SET_OR:
361       case OP_SET_XOR:
362          if (typeSizeof(i->sType) == 8 && i->sType != TYPE_F64)
363             handleSET(i->asCmp());
364          break;
365       case OP_BREV:
366          handleBREV(i);
367          break;
368       default:
369          break;
370       }
371    }
372    return true;
373 }
374 
NVC0LegalizePostRA(const Program * prog)375 NVC0LegalizePostRA::NVC0LegalizePostRA(const Program *prog)
376    : rZero(NULL),
377      carry(NULL),
378      pOne(NULL),
379      needTexBar(prog->getTarget()->getChipset() >= 0xe0 &&
380                 prog->getTarget()->getChipset() < 0x110)
381 {
382 }
383 
384 bool
insnDominatedBy(const Instruction * later,const Instruction * early) const385 NVC0LegalizePostRA::insnDominatedBy(const Instruction *later,
386                                     const Instruction *early) const
387 {
388    if (early->bb == later->bb)
389       return early->serial < later->serial;
390    return later->bb->dominatedBy(early->bb);
391 }
392 
393 void
addTexUse(std::list<TexUse> & uses,Instruction * usei,const Instruction * texi)394 NVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses,
395                               Instruction *usei, const Instruction *texi)
396 {
397    bool add = true;
398    bool dominated = insnDominatedBy(usei, texi);
399    // Uses before the tex have to all be included. Just because an earlier
400    // instruction dominates another instruction doesn't mean that there's no
401    // way to get from the tex to the later instruction. For example you could
402    // have nested loops, with the tex in the inner loop, and uses before it in
403    // both loops - even though the outer loop's instruction would dominate the
404    // inner's, we still want a texbar before the inner loop's instruction.
405    //
406    // However we can still use the eliding logic between uses dominated by the
407    // tex instruction, as that is unambiguously correct.
408    if (dominated) {
409       for (std::list<TexUse>::iterator it = uses.begin(); it != uses.end();) {
410          if (it->after) {
411             if (insnDominatedBy(usei, it->insn)) {
412                add = false;
413                break;
414             }
415             if (insnDominatedBy(it->insn, usei)) {
416                it = uses.erase(it);
417                continue;
418             }
419          }
420          ++it;
421       }
422    }
423    if (add)
424       uses.push_back(TexUse(usei, texi, dominated));
425 }
426 
427 // While it might be tempting to use the an algorithm that just looks at tex
428 // uses, not all texture results are guaranteed to be used on all paths. In
429 // the case where along some control flow path a texture result is never used,
430 // we might reuse that register for something else, creating a
431 // write-after-write hazard. So we have to manually look through all
432 // instructions looking for ones that reference the registers in question.
433 void
findFirstUses(Instruction * texi,std::list<TexUse> & uses)434 NVC0LegalizePostRA::findFirstUses(
435    Instruction *texi, std::list<TexUse> &uses)
436 {
437    int minGPR = texi->def(0).rep()->reg.data.id;
438    int maxGPR = minGPR + texi->def(0).rep()->reg.size / 4 - 1;
439 
440    unordered_set<const BasicBlock *> visited;
441    findFirstUsesBB(minGPR, maxGPR, texi->next, texi, uses, visited);
442 }
443 
444 void
findFirstUsesBB(int minGPR,int maxGPR,Instruction * start,const Instruction * texi,std::list<TexUse> & uses,unordered_set<const BasicBlock * > & visited)445 NVC0LegalizePostRA::findFirstUsesBB(
446    int minGPR, int maxGPR, Instruction *start,
447    const Instruction *texi, std::list<TexUse> &uses,
448    unordered_set<const BasicBlock *> &visited)
449 {
450    const BasicBlock *bb = start->bb;
451 
452    // We don't process the whole bb the first time around. This is correct,
453    // however we might be in a loop and hit this BB again, and need to process
454    // the full thing. So only mark a bb as visited if we processed it from the
455    // beginning.
456    if (start == bb->getEntry()) {
457       if (visited.find(bb) != visited.end())
458          return;
459       visited.insert(bb);
460    }
461 
462    for (Instruction *insn = start; insn != bb->getExit(); insn = insn->next) {
463       if (insn->isNop())
464          continue;
465 
466       for (int d = 0; insn->defExists(d); ++d) {
467          const Value *def = insn->def(d).rep();
468          if (insn->def(d).getFile() != FILE_GPR ||
469              def->reg.data.id + def->reg.size / 4 - 1 < minGPR ||
470              def->reg.data.id > maxGPR)
471             continue;
472          addTexUse(uses, insn, texi);
473          return;
474       }
475 
476       for (int s = 0; insn->srcExists(s); ++s) {
477          const Value *src = insn->src(s).rep();
478          if (insn->src(s).getFile() != FILE_GPR ||
479              src->reg.data.id + src->reg.size / 4 - 1 < minGPR ||
480              src->reg.data.id > maxGPR)
481             continue;
482          addTexUse(uses, insn, texi);
483          return;
484       }
485    }
486 
487    for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
488       findFirstUsesBB(minGPR, maxGPR, BasicBlock::get(ei.getNode())->getEntry(),
489                       texi, uses, visited);
490    }
491 }
492 
493 // Texture barriers:
494 // This pass is a bit long and ugly and can probably be optimized.
495 //
496 // 1. obtain a list of TEXes and their outputs' first use(s)
497 // 2. calculate the barrier level of each first use (minimal number of TEXes,
498 //    over all paths, between the TEX and the use in question)
499 // 3. for each barrier, if all paths from the source TEX to that barrier
500 //    contain a barrier of lesser level, it can be culled
501 bool
insertTextureBarriers(Function * fn)502 NVC0LegalizePostRA::insertTextureBarriers(Function *fn)
503 {
504    std::list<TexUse> *uses;
505    std::vector<Instruction *> texes;
506    std::vector<int> bbFirstTex;
507    std::vector<int> bbFirstUse;
508    std::vector<int> texCounts;
509    std::vector<TexUse> useVec;
510    ArrayList insns;
511 
512    fn->orderInstructions(insns);
513 
514    texCounts.resize(fn->allBBlocks.getSize(), 0);
515    bbFirstTex.resize(fn->allBBlocks.getSize(), insns.getSize());
516    bbFirstUse.resize(fn->allBBlocks.getSize(), insns.getSize());
517 
518    // tag BB CFG nodes by their id for later
519    for (ArrayList::Iterator i = fn->allBBlocks.iterator(); !i.end(); i.next()) {
520       BasicBlock *bb = reinterpret_cast<BasicBlock *>(i.get());
521       if (bb)
522          bb->cfg.tag = bb->getId();
523    }
524 
525    // gather the first uses for each TEX
526    for (int i = 0; i < insns.getSize(); ++i) {
527       Instruction *tex = reinterpret_cast<Instruction *>(insns.get(i));
528       if (isTextureOp(tex->op)) {
529          texes.push_back(tex);
530          if (!texCounts.at(tex->bb->getId()))
531             bbFirstTex[tex->bb->getId()] = texes.size() - 1;
532          texCounts[tex->bb->getId()]++;
533       }
534    }
535    insns.clear();
536    if (texes.empty())
537       return false;
538    uses = new std::list<TexUse>[texes.size()];
539    if (!uses)
540       return false;
541    for (size_t i = 0; i < texes.size(); ++i) {
542       findFirstUses(texes[i], uses[i]);
543    }
544 
545    // determine the barrier level at each use
546    for (size_t i = 0; i < texes.size(); ++i) {
547       for (std::list<TexUse>::iterator u = uses[i].begin(); u != uses[i].end();
548            ++u) {
549          BasicBlock *tb = texes[i]->bb;
550          BasicBlock *ub = u->insn->bb;
551          if (tb == ub) {
552             u->level = 0;
553             for (size_t j = i + 1; j < texes.size() &&
554                     texes[j]->bb == tb && texes[j]->serial < u->insn->serial;
555                  ++j)
556                u->level++;
557          } else {
558             u->level = fn->cfg.findLightestPathWeight(&tb->cfg,
559                                                       &ub->cfg, texCounts);
560             if (u->level < 0) {
561                WARN("Failed to find path TEX -> TEXBAR\n");
562                u->level = 0;
563                continue;
564             }
565             // this counted all TEXes in the origin block, correct that
566             u->level -= i - bbFirstTex.at(tb->getId()) + 1 /* this TEX */;
567             // and did not count the TEXes in the destination block, add those
568             for (size_t j = bbFirstTex.at(ub->getId()); j < texes.size() &&
569                     texes[j]->bb == ub && texes[j]->serial < u->insn->serial;
570                  ++j)
571                u->level++;
572          }
573          assert(u->level >= 0);
574          useVec.push_back(*u);
575       }
576    }
577    delete[] uses;
578 
579    // insert the barriers
580    for (size_t i = 0; i < useVec.size(); ++i) {
581       Instruction *prev = useVec[i].insn->prev;
582       if (useVec[i].level < 0)
583          continue;
584       if (prev && prev->op == OP_TEXBAR) {
585          if (prev->subOp > useVec[i].level)
586             prev->subOp = useVec[i].level;
587          prev->setSrc(prev->srcCount(), useVec[i].tex->getDef(0));
588       } else {
589          Instruction *bar = new_Instruction(func, OP_TEXBAR, TYPE_NONE);
590          bar->fixed = 1;
591          bar->subOp = useVec[i].level;
592          // make use explicit to ease latency calculation
593          bar->setSrc(bar->srcCount(), useVec[i].tex->getDef(0));
594          useVec[i].insn->bb->insertBefore(useVec[i].insn, bar);
595       }
596    }
597 
598    if (fn->getProgram()->optLevel < 3)
599       return true;
600 
601    std::vector<Limits> limitT, limitB, limitS; // entry, exit, single
602 
603    limitT.resize(fn->allBBlocks.getSize(), Limits(0, 0));
604    limitB.resize(fn->allBBlocks.getSize(), Limits(0, 0));
605    limitS.resize(fn->allBBlocks.getSize());
606 
607    // cull unneeded barriers (should do that earlier, but for simplicity)
608    IteratorRef bi = fn->cfg.iteratorCFG();
609    // first calculate min/max outstanding TEXes for each BB
610    for (bi->reset(); !bi->end(); bi->next()) {
611       Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
612       BasicBlock *bb = BasicBlock::get(n);
613       int min = 0;
614       int max = std::numeric_limits<int>::max();
615       for (Instruction *i = bb->getFirst(); i; i = i->next) {
616          if (isTextureOp(i->op)) {
617             min++;
618             if (max < std::numeric_limits<int>::max())
619                max++;
620          } else
621          if (i->op == OP_TEXBAR) {
622             min = MIN2(min, i->subOp);
623             max = MIN2(max, i->subOp);
624          }
625       }
626       // limits when looking at an isolated block
627       limitS[bb->getId()].min = min;
628       limitS[bb->getId()].max = max;
629    }
630    // propagate the min/max values
631    for (unsigned int l = 0; l <= fn->loopNestingBound; ++l) {
632       for (bi->reset(); !bi->end(); bi->next()) {
633          Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
634          BasicBlock *bb = BasicBlock::get(n);
635          const int bbId = bb->getId();
636          for (Graph::EdgeIterator ei = n->incident(); !ei.end(); ei.next()) {
637             BasicBlock *in = BasicBlock::get(ei.getNode());
638             const int inId = in->getId();
639             limitT[bbId].min = MAX2(limitT[bbId].min, limitB[inId].min);
640             limitT[bbId].max = MAX2(limitT[bbId].max, limitB[inId].max);
641          }
642          // I just hope this is correct ...
643          if (limitS[bbId].max == std::numeric_limits<int>::max()) {
644             // no barrier
645             limitB[bbId].min = limitT[bbId].min + limitS[bbId].min;
646             limitB[bbId].max = limitT[bbId].max + limitS[bbId].min;
647          } else {
648             // block contained a barrier
649             limitB[bbId].min = MIN2(limitS[bbId].max,
650                                     limitT[bbId].min + limitS[bbId].min);
651             limitB[bbId].max = MIN2(limitS[bbId].max,
652                                     limitT[bbId].max + limitS[bbId].min);
653          }
654       }
655    }
656    // finally delete unnecessary barriers
657    for (bi->reset(); !bi->end(); bi->next()) {
658       Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
659       BasicBlock *bb = BasicBlock::get(n);
660       Instruction *prev = NULL;
661       Instruction *next;
662       int max = limitT[bb->getId()].max;
663       for (Instruction *i = bb->getFirst(); i; i = next) {
664          next = i->next;
665          if (i->op == OP_TEXBAR) {
666             if (i->subOp >= max) {
667                delete_Instruction(prog, i);
668                i = NULL;
669             } else {
670                max = i->subOp;
671                if (prev && prev->op == OP_TEXBAR && prev->subOp >= max) {
672                   delete_Instruction(prog, prev);
673                   prev = NULL;
674                }
675             }
676          } else
677          if (isTextureOp(i->op)) {
678             max++;
679          }
680          if (i && !i->isNop())
681             prev = i;
682       }
683    }
684    return true;
685 }
686 
687 bool
visit(Function * fn)688 NVC0LegalizePostRA::visit(Function *fn)
689 {
690    if (needTexBar)
691       insertTextureBarriers(fn);
692 
693    rZero = new_LValue(fn, FILE_GPR);
694    pOne = new_LValue(fn, FILE_PREDICATE);
695    carry = new_LValue(fn, FILE_FLAGS);
696 
697    rZero->reg.data.id = (prog->getTarget()->getChipset() >= NVISA_GK20A_CHIPSET) ? 255 : 63;
698    carry->reg.data.id = 0;
699    pOne->reg.data.id = 7;
700 
701    return true;
702 }
703 
704 void
replaceZero(Instruction * i)705 NVC0LegalizePostRA::replaceZero(Instruction *i)
706 {
707    for (int s = 0; i->srcExists(s); ++s) {
708       if (s == 2 && i->op == OP_SUCLAMP)
709          continue;
710       if (s == 1 && i->op == OP_SHLADD)
711          continue;
712       ImmediateValue *imm = i->getSrc(s)->asImm();
713       if (imm) {
714          if (i->op == OP_SELP && s == 2) {
715             i->setSrc(s, pOne);
716             if (imm->reg.data.u64 == 0)
717                i->src(s).mod = i->src(s).mod ^ Modifier(NV50_IR_MOD_NOT);
718          } else if (imm->reg.data.u64 == 0) {
719             i->setSrc(s, rZero);
720          }
721       }
722    }
723 }
724 
725 // replace CONT with BRA for single unconditional continue
726 bool
tryReplaceContWithBra(BasicBlock * bb)727 NVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock *bb)
728 {
729    if (bb->cfg.incidentCount() != 2 || bb->getEntry()->op != OP_PRECONT)
730       return false;
731    Graph::EdgeIterator ei = bb->cfg.incident();
732    if (ei.getType() != Graph::Edge::BACK)
733       ei.next();
734    if (ei.getType() != Graph::Edge::BACK)
735       return false;
736    BasicBlock *contBB = BasicBlock::get(ei.getNode());
737 
738    if (!contBB->getExit() || contBB->getExit()->op != OP_CONT ||
739        contBB->getExit()->getPredicate())
740       return false;
741    contBB->getExit()->op = OP_BRA;
742    bb->remove(bb->getEntry()); // delete PRECONT
743 
744    ei.next();
745    assert(ei.end() || ei.getType() != Graph::Edge::BACK);
746    return true;
747 }
748 
749 // replace branches to join blocks with join ops
750 void
propagateJoin(BasicBlock * bb)751 NVC0LegalizePostRA::propagateJoin(BasicBlock *bb)
752 {
753    if (bb->getEntry()->op != OP_JOIN || bb->getEntry()->asFlow()->limit)
754       return;
755    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
756       BasicBlock *in = BasicBlock::get(ei.getNode());
757       Instruction *exit = in->getExit();
758       if (!exit) {
759          in->insertTail(new FlowInstruction(func, OP_JOIN, bb));
760          // there should always be a terminator instruction
761          WARN("inserted missing terminator in BB:%i\n", in->getId());
762       } else
763       if (exit->op == OP_BRA) {
764          exit->op = OP_JOIN;
765          exit->asFlow()->limit = 1; // must-not-propagate marker
766       }
767    }
768    bb->remove(bb->getEntry());
769 }
770 
771 // replaces instructions which would end up as f2f or i2i with faster
772 // alternatives:
773 //  - fabs(a)     -> fadd(0, abs a)
774 //  - fneg(a)     -> fadd(neg 0, neg a)
775 //  - ineg(a)     -> iadd(0, neg a)
776 //  - fneg(abs a) -> fadd(neg 0, neg abs a)
777 //  - sat(a)      -> sat add(0, a)
778 void
replaceCvt(Instruction * cvt)779 NVC0LegalizePostRA::replaceCvt(Instruction *cvt)
780 {
781    if (!isFloatType(cvt->sType) && typeSizeof(cvt->sType) != 4)
782       return;
783    if (cvt->sType != cvt->dType)
784       return;
785    // we could make it work, but in this case we have optimizations disabled
786    // and we don't really care either way.
787    if (cvt->src(0).getFile() != FILE_GPR &&
788        cvt->src(0).getFile() != FILE_MEMORY_CONST)
789       return;
790 
791    Modifier mod0, mod1;
792 
793    switch (cvt->op) {
794    case OP_ABS:
795       if (cvt->src(0).mod)
796          return;
797       if (!isFloatType(cvt->sType))
798          return;
799       mod0 = 0;
800       mod1 = NV50_IR_MOD_ABS;
801       break;
802    case OP_NEG:
803       if (!isFloatType(cvt->sType) && cvt->src(0).mod)
804          return;
805       if (isFloatType(cvt->sType) &&
806           (cvt->src(0).mod && cvt->src(0).mod != Modifier(NV50_IR_MOD_ABS)))
807          return;
808 
809       mod0 = isFloatType(cvt->sType) ? NV50_IR_MOD_NEG : 0;
810       mod1 = cvt->src(0).mod == Modifier(NV50_IR_MOD_ABS) ?
811          NV50_IR_MOD_NEG_ABS : NV50_IR_MOD_NEG;
812       break;
813    case OP_SAT:
814       if (!isFloatType(cvt->sType) && cvt->src(0).mod.abs())
815          return;
816       mod0 = 0;
817       mod1 = cvt->src(0).mod;
818       cvt->saturate = true;
819       break;
820    default:
821       return;
822    }
823 
824    cvt->op = OP_ADD;
825    cvt->moveSources(0, 1);
826    cvt->setSrc(0, rZero);
827    cvt->src(0).mod = mod0;
828    cvt->src(1).mod = mod1;
829 }
830 
831 bool
visit(BasicBlock * bb)832 NVC0LegalizePostRA::visit(BasicBlock *bb)
833 {
834    Instruction *i, *next;
835 
836    // remove pseudo operations and non-fixed no-ops, split 64 bit operations
837    for (i = bb->getFirst(); i; i = next) {
838       next = i->next;
839       if (i->op == OP_EMIT || i->op == OP_RESTART) {
840          if (!i->getDef(0)->refCount())
841             i->setDef(0, NULL);
842          if (i->src(0).getFile() == FILE_IMMEDIATE)
843             i->setSrc(0, rZero); // initial value must be 0
844          replaceZero(i);
845       } else
846       if (i->isNop()) {
847          bb->remove(i);
848       } else
849       if (i->op == OP_BAR && i->subOp == NV50_IR_SUBOP_BAR_SYNC &&
850           prog->getType() != Program::TYPE_COMPUTE) {
851          // It seems like barriers are never required for tessellation since
852          // the warp size is 32, and there are always at most 32 tcs threads.
853          bb->remove(i);
854       } else
855       if (i->op == OP_LOAD && i->subOp == NV50_IR_SUBOP_LDC_IS) {
856          int offset = i->src(0).get()->reg.data.offset;
857          if (abs(offset) >= 0x10000)
858             i->src(0).get()->reg.fileIndex += offset >> 16;
859          i->src(0).get()->reg.data.offset = (int)(short)offset;
860       } else {
861          // TODO: Move this to before register allocation for operations that
862          // need the $c register !
863          if (typeSizeof(i->sType) == 8 || typeSizeof(i->dType) == 8) {
864             Instruction *hi;
865             hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry);
866             if (hi)
867                next = hi;
868          }
869 
870          if (i->op != OP_MOV && i->op != OP_PFETCH)
871             replaceZero(i);
872 
873          if (i->op == OP_SAT || i->op == OP_NEG || i->op == OP_ABS)
874             replaceCvt(i);
875       }
876    }
877    if (!bb->getEntry())
878       return true;
879 
880    if (!tryReplaceContWithBra(bb))
881       propagateJoin(bb);
882 
883    return true;
884 }
885 
NVC0LoweringPass(Program * prog)886 NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
887 {
888    bld.setProgram(prog);
889 }
890 
891 bool
visit(Function * fn)892 NVC0LoweringPass::visit(Function *fn)
893 {
894    if (prog->getType() == Program::TYPE_GEOMETRY) {
895       assert(!strncmp(fn->getName(), "MAIN", 4));
896       // TODO: when we generate actual functions pass this value along somehow
897       bld.setPosition(BasicBlock::get(fn->cfg.getRoot()), false);
898       gpEmitAddress = bld.loadImm(NULL, 0)->asLValue();
899       if (fn->cfgExit) {
900          bld.setPosition(BasicBlock::get(fn->cfgExit)->getExit(), false);
901          if (prog->getTarget()->getChipset() >= NVISA_GV100_CHIPSET)
902             bld.mkOp1(OP_FINAL, TYPE_NONE, NULL, gpEmitAddress)->fixed = 1;
903          bld.mkMovToReg(0, gpEmitAddress);
904       }
905    }
906    return true;
907 }
908 
909 bool
visit(BasicBlock * bb)910 NVC0LoweringPass::visit(BasicBlock *bb)
911 {
912    return true;
913 }
914 
915 inline Value *
loadTexHandle(Value * ptr,unsigned int slot)916 NVC0LoweringPass::loadTexHandle(Value *ptr, unsigned int slot)
917 {
918    uint8_t b = prog->driver->io.auxCBSlot;
919    uint32_t off = prog->driver->io.texBindBase + slot * 4;
920 
921    if (ptr)
922       ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(2));
923 
924    return bld.
925       mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
926 }
927 
928 // move array source to first slot, convert to u16, add indirections
929 bool
handleTEX(TexInstruction * i)930 NVC0LoweringPass::handleTEX(TexInstruction *i)
931 {
932    const int dim = i->tex.target.getDim() + i->tex.target.isCube();
933    const int arg = i->tex.target.getArgCount();
934    const int lyr = arg - (i->tex.target.isMS() ? 2 : 1);
935    const int chipset = prog->getTarget()->getChipset();
936 
937    /* Only normalize in the non-explicit derivatives case. For explicit
938     * derivatives, this is handled in handleManualTXD.
939     */
940    if (i->tex.target.isCube() && i->dPdx[0].get() == NULL) {
941       Value *src[3], *val;
942       int c;
943       for (c = 0; c < 3; ++c)
944          src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), i->getSrc(c));
945       val = bld.getScratch();
946       bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
947       bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
948       bld.mkOp1(OP_RCP, TYPE_F32, val, val);
949       for (c = 0; c < 3; ++c) {
950          i->setSrc(c, bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(),
951                                  i->getSrc(c), val));
952       }
953    }
954 
955    // Arguments to the TEX instruction are a little insane. Even though the
956    // encoding is identical between SM20 and SM30, the arguments mean
957    // different things between Fermi and Kepler+. A lot of arguments are
958    // optional based on flags passed to the instruction. This summarizes the
959    // order of things.
960    //
961    // Fermi:
962    //  array/indirect
963    //  coords
964    //  sample
965    //  lod bias
966    //  depth compare
967    //  offsets:
968    //    - tg4: 8 bits each, either 2 (1 offset reg) or 8 (2 offset reg)
969    //    - other: 4 bits each, single reg
970    //
971    // Kepler+:
972    //  indirect handle
973    //  array (+ offsets for txd in upper 16 bits)
974    //  coords
975    //  sample
976    //  lod bias
977    //  depth compare
978    //  offsets (same as fermi, except txd which takes it with array)
979    //
980    // Maxwell (tex):
981    //  array
982    //  coords
983    //  indirect handle
984    //  sample
985    //  lod bias
986    //  depth compare
987    //  offsets
988    //
989    // Maxwell (txd):
990    //  indirect handle
991    //  coords
992    //  array + offsets
993    //  derivatives
994 
995    if (chipset >= NVISA_GK104_CHIPSET) {
996       if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
997          // XXX this ignores tsc, and assumes a 1:1 mapping
998          assert(i->tex.rIndirectSrc >= 0);
999          if (!i->tex.bindless) {
1000             Value *hnd = loadTexHandle(i->getIndirectR(), i->tex.r);
1001             i->tex.r = 0xff;
1002             i->tex.s = 0x1f;
1003             i->setIndirectR(hnd);
1004          }
1005          i->setIndirectS(NULL);
1006       } else if (i->tex.r == i->tex.s || i->op == OP_TXF) {
1007          if (i->tex.r == 0xffff)
1008             i->tex.r = prog->driver->io.fbtexBindBase / 4;
1009          else
1010             i->tex.r += prog->driver->io.texBindBase / 4;
1011          i->tex.s  = 0; // only a single cX[] value possible here
1012       } else {
1013          Value *hnd = bld.getScratch();
1014          Value *rHnd = loadTexHandle(NULL, i->tex.r);
1015          Value *sHnd = loadTexHandle(NULL, i->tex.s);
1016 
1017          bld.mkOp3(OP_INSBF, TYPE_U32, hnd, rHnd, bld.mkImm(0x1400), sHnd);
1018 
1019          i->tex.r = 0; // not used for indirect tex
1020          i->tex.s = 0;
1021          i->setIndirectR(hnd);
1022       }
1023       if (i->tex.target.isArray()) {
1024          LValue *layer = new_LValue(func, FILE_GPR);
1025          Value *src = i->getSrc(lyr);
1026          const int sat = (i->op == OP_TXF) ? 1 : 0;
1027          DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
1028          bld.mkCvt(OP_CVT, TYPE_U16, layer, sTy, src)->saturate = sat;
1029          if (i->op != OP_TXD || chipset < NVISA_GM107_CHIPSET) {
1030             for (int s = dim; s >= 1; --s)
1031                i->setSrc(s, i->getSrc(s - 1));
1032             i->setSrc(0, layer);
1033          } else {
1034             i->setSrc(dim, layer);
1035          }
1036       }
1037       // Move the indirect reference to the first place
1038       if (i->tex.rIndirectSrc >= 0 && (
1039                 i->op == OP_TXD || chipset < NVISA_GM107_CHIPSET)) {
1040          Value *hnd = i->getIndirectR();
1041 
1042          i->setIndirectR(NULL);
1043          i->moveSources(0, 1);
1044          i->setSrc(0, hnd);
1045          i->tex.rIndirectSrc = 0;
1046          i->tex.sIndirectSrc = -1;
1047       }
1048       // Move the indirect reference to right after the coords
1049       else if (i->tex.rIndirectSrc >= 0 && chipset >= NVISA_GM107_CHIPSET) {
1050          Value *hnd = i->getIndirectR();
1051 
1052          i->setIndirectR(NULL);
1053          i->moveSources(arg, 1);
1054          i->setSrc(arg, hnd);
1055          i->tex.rIndirectSrc = 0;
1056          i->tex.sIndirectSrc = -1;
1057       }
1058    } else
1059    // (nvc0) generate and move the tsc/tic/array source to the front
1060    if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
1061       LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
1062 
1063       Value *ticRel = i->getIndirectR();
1064       Value *tscRel = i->getIndirectS();
1065 
1066       if (i->tex.r == 0xffff) {
1067          i->tex.r = 0x20;
1068          i->tex.s = 0x10;
1069       }
1070 
1071       if (ticRel) {
1072          i->setSrc(i->tex.rIndirectSrc, NULL);
1073          if (i->tex.r)
1074             ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
1075                                 ticRel, bld.mkImm(i->tex.r));
1076       }
1077       if (tscRel) {
1078          i->setSrc(i->tex.sIndirectSrc, NULL);
1079          if (i->tex.s)
1080             tscRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
1081                                 tscRel, bld.mkImm(i->tex.s));
1082       }
1083 
1084       Value *arrayIndex = i->tex.target.isArray() ? i->getSrc(lyr) : NULL;
1085       if (arrayIndex) {
1086          for (int s = dim; s >= 1; --s)
1087             i->setSrc(s, i->getSrc(s - 1));
1088          i->setSrc(0, arrayIndex);
1089       } else {
1090          i->moveSources(0, 1);
1091       }
1092 
1093       if (arrayIndex) {
1094          int sat = (i->op == OP_TXF) ? 1 : 0;
1095          DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
1096          bld.mkCvt(OP_CVT, TYPE_U16, src, sTy, arrayIndex)->saturate = sat;
1097       } else {
1098          bld.loadImm(src, 0);
1099       }
1100 
1101       if (ticRel)
1102          bld.mkOp3(OP_INSBF, TYPE_U32, src, ticRel, bld.mkImm(0x0917), src);
1103       if (tscRel)
1104          bld.mkOp3(OP_INSBF, TYPE_U32, src, tscRel, bld.mkImm(0x0710), src);
1105 
1106       i->setSrc(0, src);
1107    }
1108 
1109    // For nvc0, the sample id has to be in the second operand, as the offset
1110    // does. Right now we don't know how to pass both in, and this case can't
1111    // happen with OpenGL. On nve0, the sample id is part of the texture
1112    // coordinate argument.
1113    assert(chipset >= NVISA_GK104_CHIPSET ||
1114           !i->tex.useOffsets || !i->tex.target.isMS());
1115 
1116    // offset is between lod and dc
1117    if (i->tex.useOffsets) {
1118       int n, c;
1119       int s = i->srcCount(0xff, true);
1120       if (i->op != OP_TXD || chipset < NVISA_GK104_CHIPSET) {
1121          if (i->tex.target.isShadow())
1122             s--;
1123          if (i->srcExists(s)) // move potential predicate out of the way
1124             i->moveSources(s, 1);
1125          if (i->tex.useOffsets == 4 && i->srcExists(s + 1))
1126             i->moveSources(s + 1, 1);
1127       }
1128       if (i->op == OP_TXG) {
1129          // Either there is 1 offset, which goes into the 2 low bytes of the
1130          // first source, or there are 4 offsets, which go into 2 sources (8
1131          // values, 1 byte each).
1132          Value *offs[2] = {NULL, NULL};
1133          for (n = 0; n < i->tex.useOffsets; n++) {
1134             for (c = 0; c < 2; ++c) {
1135                if ((n % 2) == 0 && c == 0)
1136                   bld.mkMov(offs[n / 2] = bld.getScratch(), i->offset[n][c].get());
1137                else
1138                   bld.mkOp3(OP_INSBF, TYPE_U32,
1139                             offs[n / 2],
1140                             i->offset[n][c].get(),
1141                             bld.mkImm(0x800 | ((n * 16 + c * 8) % 32)),
1142                             offs[n / 2]);
1143             }
1144          }
1145          i->setSrc(s, offs[0]);
1146          if (offs[1])
1147             i->setSrc(s + 1, offs[1]);
1148       } else {
1149          unsigned imm = 0;
1150          assert(i->tex.useOffsets == 1);
1151          for (c = 0; c < 3; ++c) {
1152             ImmediateValue val;
1153             if (!i->offset[0][c].getImmediate(val))
1154                assert(!"non-immediate offset passed to non-TXG");
1155             imm |= (val.reg.data.u32 & 0xf) << (c * 4);
1156          }
1157          if (i->op == OP_TXD && chipset >= NVISA_GK104_CHIPSET) {
1158             // The offset goes into the upper 16 bits of the array index. So
1159             // create it if it's not already there, and INSBF it if it already
1160             // is.
1161             s = (i->tex.rIndirectSrc >= 0) ? 1 : 0;
1162             if (chipset >= NVISA_GM107_CHIPSET)
1163                s += dim;
1164             if (i->tex.target.isArray()) {
1165                Value *offset = bld.getScratch();
1166                bld.mkOp3(OP_INSBF, TYPE_U32, offset,
1167                          bld.loadImm(NULL, imm), bld.mkImm(0xc10),
1168                          i->getSrc(s));
1169                i->setSrc(s, offset);
1170             } else {
1171                i->moveSources(s, 1);
1172                i->setSrc(s, bld.loadImm(NULL, imm << 16));
1173             }
1174          } else {
1175             i->setSrc(s, bld.loadImm(NULL, imm));
1176          }
1177       }
1178    }
1179 
1180    return true;
1181 }
1182 
1183 bool
handleManualTXD(TexInstruction * i)1184 NVC0LoweringPass::handleManualTXD(TexInstruction *i)
1185 {
1186    // Always done from the l0 perspective. This is the way that NVIDIA's
1187    // driver does it, and doing it from the "current" lane's perpsective
1188    // doesn't seem to always work for reasons that aren't altogether clear,
1189    // even in frag shaders.
1190    //
1191    // Note that we must move not only the coordinates into lane0, but also all
1192    // ancillary arguments, like array indices and depth compare as they may
1193    // differ between lanes. Offsets for TXD are supposed to be uniform, so we
1194    // leave them alone.
1195    static const uint8_t qOps[2] =
1196       { QUADOP(MOV2, ADD,  MOV2, ADD),  QUADOP(MOV2, MOV2, ADD,  ADD) };
1197 
1198    Value *def[4][4];
1199    Value *crd[3], *arr[2], *shadow;
1200    Instruction *tex;
1201    Value *zero = bld.loadImm(bld.getSSA(), 0);
1202    int l, c;
1203    const int dim = i->tex.target.getDim() + i->tex.target.isCube();
1204 
1205    // This function is invoked after handleTEX lowering, so we have to expect
1206    // the arguments in the order that the hw wants them. For Fermi, array and
1207    // indirect are both in the leading arg, while for Kepler, array and
1208    // indirect are separate (and both precede the coordinates). Maxwell is
1209    // handled in a separate function.
1210    int array;
1211    if (targ->getChipset() < NVISA_GK104_CHIPSET)
1212       array = i->tex.target.isArray() || i->tex.rIndirectSrc >= 0;
1213    else
1214       array = i->tex.target.isArray() + (i->tex.rIndirectSrc >= 0);
1215 
1216    i->op = OP_TEX; // no need to clone dPdx/dPdy later
1217 
1218    for (c = 0; c < dim; ++c)
1219       crd[c] = bld.getScratch();
1220    for (c = 0; c < array; ++c)
1221       arr[c] = bld.getScratch();
1222    shadow = bld.getScratch();
1223 
1224    for (l = 0; l < 4; ++l) {
1225       Value *src[3], *val;
1226 
1227       bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
1228       // we're using the texture result from lane 0 in all cases, so make sure
1229       // that lane 0 is pointing at the proper array index, indirect value,
1230       // and depth compare.
1231       if (l != 0) {
1232          for (c = 0; c < array; ++c)
1233             bld.mkQuadop(0x00, arr[c], l, i->getSrc(c), zero);
1234          if (i->tex.target.isShadow()) {
1235             // The next argument after coords is the depth compare
1236             bld.mkQuadop(0x00, shadow, l, i->getSrc(array + dim), zero);
1237          }
1238       }
1239       // mov position coordinates from lane l to all lanes
1240       for (c = 0; c < dim; ++c)
1241          bld.mkQuadop(0x00, crd[c], l, i->getSrc(c + array), zero);
1242       // add dPdx from lane l to lanes dx
1243       for (c = 0; c < dim; ++c)
1244          bld.mkQuadop(qOps[0], crd[c], l, i->dPdx[c].get(), crd[c]);
1245       // add dPdy from lane l to lanes dy
1246       for (c = 0; c < dim; ++c)
1247          bld.mkQuadop(qOps[1], crd[c], l, i->dPdy[c].get(), crd[c]);
1248       // normalize cube coordinates
1249       if (i->tex.target.isCube()) {
1250          for (c = 0; c < 3; ++c)
1251             src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), crd[c]);
1252          val = bld.getScratch();
1253          bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
1254          bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
1255          bld.mkOp1(OP_RCP, TYPE_F32, val, val);
1256          for (c = 0; c < 3; ++c)
1257             src[c] = bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(), crd[c], val);
1258       } else {
1259          for (c = 0; c < dim; ++c)
1260             src[c] = crd[c];
1261       }
1262       // texture
1263       bld.insert(tex = cloneForward(func, i));
1264       if (l != 0) {
1265          for (c = 0; c < array; ++c)
1266             tex->setSrc(c, arr[c]);
1267          if (i->tex.target.isShadow())
1268             tex->setSrc(array + dim, shadow);
1269       }
1270       for (c = 0; c < dim; ++c)
1271          tex->setSrc(c + array, src[c]);
1272       // broadcast results from lane 0 to all lanes so that the moves *into*
1273       // the target lane pick up the proper value.
1274       if (l != 0)
1275          for (c = 0; i->defExists(c); ++c)
1276             bld.mkQuadop(0x00, tex->getDef(c), 0, tex->getDef(c), zero);
1277       bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
1278 
1279       // save results
1280       for (c = 0; i->defExists(c); ++c) {
1281          Instruction *mov;
1282          def[c][l] = bld.getSSA();
1283          mov = bld.mkMov(def[c][l], tex->getDef(c));
1284          mov->fixed = 1;
1285          mov->lanes = 1 << l;
1286       }
1287    }
1288 
1289    for (c = 0; i->defExists(c); ++c) {
1290       Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c));
1291       for (l = 0; l < 4; ++l)
1292          u->setSrc(l, def[c][l]);
1293    }
1294 
1295    i->bb->remove(i);
1296    return true;
1297 }
1298 
1299 bool
handleTXD(TexInstruction * txd)1300 NVC0LoweringPass::handleTXD(TexInstruction *txd)
1301 {
1302    int dim = txd->tex.target.getDim() + txd->tex.target.isCube();
1303    unsigned arg = txd->tex.target.getArgCount();
1304    unsigned expected_args = arg;
1305    const int chipset = prog->getTarget()->getChipset();
1306 
1307    if (chipset >= NVISA_GK104_CHIPSET) {
1308       if (!txd->tex.target.isArray() && txd->tex.useOffsets)
1309          expected_args++;
1310       if (txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0)
1311          expected_args++;
1312    } else {
1313       if (txd->tex.useOffsets)
1314          expected_args++;
1315       if (!txd->tex.target.isArray() && (
1316                 txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0))
1317          expected_args++;
1318    }
1319 
1320    if (expected_args > 4 ||
1321        dim > 2 ||
1322        txd->tex.target.isShadow())
1323       txd->op = OP_TEX;
1324 
1325    handleTEX(txd);
1326    while (txd->srcExists(arg))
1327       ++arg;
1328 
1329    txd->tex.derivAll = true;
1330    if (txd->op == OP_TEX)
1331       return handleManualTXD(txd);
1332 
1333    assert(arg == expected_args);
1334    for (int c = 0; c < dim; ++c) {
1335       txd->setSrc(arg + c * 2 + 0, txd->dPdx[c]);
1336       txd->setSrc(arg + c * 2 + 1, txd->dPdy[c]);
1337       txd->dPdx[c].set(NULL);
1338       txd->dPdy[c].set(NULL);
1339    }
1340 
1341    // In this case we have fewer than 4 "real" arguments, which means that
1342    // handleTEX didn't apply any padding. However we have to make sure that
1343    // the second "group" of arguments still gets padded up to 4.
1344    if (chipset >= NVISA_GK104_CHIPSET) {
1345       int s = arg + 2 * dim;
1346       if (s >= 4 && s < 7) {
1347          if (txd->srcExists(s)) // move potential predicate out of the way
1348             txd->moveSources(s, 7 - s);
1349          while (s < 7)
1350             txd->setSrc(s++, bld.loadImm(NULL, 0));
1351       }
1352    }
1353 
1354    return true;
1355 }
1356 
1357 bool
handleTXQ(TexInstruction * txq)1358 NVC0LoweringPass::handleTXQ(TexInstruction *txq)
1359 {
1360    const int chipset = prog->getTarget()->getChipset();
1361    if (chipset >= NVISA_GK104_CHIPSET && txq->tex.rIndirectSrc < 0)
1362       txq->tex.r += prog->driver->io.texBindBase / 4;
1363 
1364    if (txq->tex.rIndirectSrc < 0)
1365       return true;
1366 
1367    Value *ticRel = txq->getIndirectR();
1368 
1369    txq->setIndirectS(NULL);
1370    txq->tex.sIndirectSrc = -1;
1371 
1372    assert(ticRel);
1373 
1374    if (chipset < NVISA_GK104_CHIPSET) {
1375       LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
1376 
1377       txq->setSrc(txq->tex.rIndirectSrc, NULL);
1378       if (txq->tex.r)
1379          ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
1380                              ticRel, bld.mkImm(txq->tex.r));
1381 
1382       bld.mkOp2(OP_SHL, TYPE_U32, src, ticRel, bld.mkImm(0x17));
1383 
1384       txq->moveSources(0, 1);
1385       txq->setSrc(0, src);
1386    } else {
1387       Value *hnd = loadTexHandle(txq->getIndirectR(), txq->tex.r);
1388       txq->tex.r = 0xff;
1389       txq->tex.s = 0x1f;
1390 
1391       txq->setIndirectR(NULL);
1392       txq->moveSources(0, 1);
1393       txq->setSrc(0, hnd);
1394       txq->tex.rIndirectSrc = 0;
1395    }
1396 
1397    return true;
1398 }
1399 
1400 bool
handleTXLQ(TexInstruction * i)1401 NVC0LoweringPass::handleTXLQ(TexInstruction *i)
1402 {
1403    /* The outputs are inverted compared to what the TGSI instruction
1404     * expects. Take that into account in the mask.
1405     */
1406    assert((i->tex.mask & ~3) == 0);
1407    if (i->tex.mask == 1)
1408       i->tex.mask = 2;
1409    else if (i->tex.mask == 2)
1410       i->tex.mask = 1;
1411    handleTEX(i);
1412    bld.setPosition(i, true);
1413 
1414    /* The returned values are not quite what we want:
1415     * (a) convert from s16/u16 to f32
1416     * (b) multiply by 1/256
1417     */
1418    for (int def = 0; def < 2; ++def) {
1419       if (!i->defExists(def))
1420          continue;
1421       enum DataType type = TYPE_S16;
1422       if (i->tex.mask == 2 || def > 0)
1423          type = TYPE_U16;
1424       bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(def), type, i->getDef(def));
1425       bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(def),
1426                 i->getDef(def), bld.loadImm(NULL, 1.0f / 256));
1427    }
1428    if (i->tex.mask == 3) {
1429       LValue *t = new_LValue(func, FILE_GPR);
1430       bld.mkMov(t, i->getDef(0));
1431       bld.mkMov(i->getDef(0), i->getDef(1));
1432       bld.mkMov(i->getDef(1), t);
1433    }
1434    return true;
1435 }
1436 
1437 bool
handleBUFQ(Instruction * bufq)1438 NVC0LoweringPass::handleBUFQ(Instruction *bufq)
1439 {
1440    bufq->op = OP_MOV;
1441    bufq->setSrc(0, loadBufLength32(bufq->getIndirect(0, 1),
1442                                    bufq->getSrc(0)->reg.fileIndex * 16));
1443    bufq->setIndirect(0, 0, NULL);
1444    bufq->setIndirect(0, 1, NULL);
1445    return true;
1446 }
1447 
1448 void
handleSharedATOMNVE4(Instruction * atom)1449 NVC0LoweringPass::handleSharedATOMNVE4(Instruction *atom)
1450 {
1451    assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
1452 
1453    BasicBlock *currBB = atom->bb;
1454    BasicBlock *tryLockBB = atom->bb->splitBefore(atom, false);
1455    BasicBlock *joinBB = atom->bb->splitAfter(atom);
1456    BasicBlock *setAndUnlockBB = new BasicBlock(func);
1457    BasicBlock *failLockBB = new BasicBlock(func);
1458 
1459    bld.setPosition(currBB, true);
1460    assert(!currBB->joinAt);
1461    currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
1462 
1463    CmpInstruction *pred =
1464       bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1465                 TYPE_U32, bld.mkImm(0), bld.mkImm(1));
1466 
1467    bld.mkFlow(OP_BRA, tryLockBB, CC_ALWAYS, NULL);
1468    currBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::TREE);
1469 
1470    bld.setPosition(tryLockBB, true);
1471 
1472    Instruction *ld =
1473       bld.mkLoad(TYPE_U32, atom->getDef(0), atom->getSrc(0)->asSym(),
1474                  atom->getIndirect(0, 0));
1475    ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
1476    ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;
1477 
1478    bld.mkFlow(OP_BRA, setAndUnlockBB, CC_P, ld->getDef(1));
1479    bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
1480    tryLockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::CROSS);
1481    tryLockBB->cfg.attach(&setAndUnlockBB->cfg, Graph::Edge::TREE);
1482 
1483    tryLockBB->cfg.detach(&joinBB->cfg);
1484    bld.remove(atom);
1485 
1486    bld.setPosition(setAndUnlockBB, true);
1487    Value *stVal;
1488    if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
1489       // Read the old value, and write the new one.
1490       stVal = atom->getSrc(1);
1491    } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1492       CmpInstruction *set =
1493          bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(),
1494                    TYPE_U32, ld->getDef(0), atom->getSrc(1));
1495 
1496       bld.mkCmp(OP_SLCT, CC_NE, TYPE_U32, (stVal = bld.getSSA()),
1497                 TYPE_U32, atom->getSrc(2), ld->getDef(0), set->getDef(0));
1498    } else {
1499       operation op;
1500 
1501       switch (atom->subOp) {
1502       case NV50_IR_SUBOP_ATOM_ADD:
1503          op = OP_ADD;
1504          break;
1505       case NV50_IR_SUBOP_ATOM_AND:
1506          op = OP_AND;
1507          break;
1508       case NV50_IR_SUBOP_ATOM_OR:
1509          op = OP_OR;
1510          break;
1511       case NV50_IR_SUBOP_ATOM_XOR:
1512          op = OP_XOR;
1513          break;
1514       case NV50_IR_SUBOP_ATOM_MIN:
1515          op = OP_MIN;
1516          break;
1517       case NV50_IR_SUBOP_ATOM_MAX:
1518          op = OP_MAX;
1519          break;
1520       default:
1521          assert(0);
1522          return;
1523       }
1524 
1525       stVal = bld.mkOp2v(op, atom->dType, bld.getSSA(), ld->getDef(0),
1526                          atom->getSrc(1));
1527    }
1528 
1529    Instruction *st =
1530       bld.mkStore(OP_STORE, TYPE_U32, atom->getSrc(0)->asSym(),
1531                   atom->getIndirect(0, 0), stVal);
1532    st->setDef(0, pred->getDef(0));
1533    st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;
1534 
1535    bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
1536    setAndUnlockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::TREE);
1537 
1538    // Lock until the store has not been performed.
1539    bld.setPosition(failLockBB, true);
1540    bld.mkFlow(OP_BRA, tryLockBB, CC_NOT_P, pred->getDef(0));
1541    bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
1542    failLockBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::BACK);
1543    failLockBB->cfg.attach(&joinBB->cfg, Graph::Edge::TREE);
1544 
1545    bld.setPosition(joinBB, false);
1546    bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1547 }
1548 
1549 void
handleSharedATOM(Instruction * atom)1550 NVC0LoweringPass::handleSharedATOM(Instruction *atom)
1551 {
1552    assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
1553 
1554    BasicBlock *currBB = atom->bb;
1555    BasicBlock *tryLockAndSetBB = atom->bb->splitBefore(atom, false);
1556    BasicBlock *joinBB = atom->bb->splitAfter(atom);
1557 
1558    bld.setPosition(currBB, true);
1559    assert(!currBB->joinAt);
1560    currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
1561 
1562    bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_ALWAYS, NULL);
1563    currBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::TREE);
1564 
1565    bld.setPosition(tryLockAndSetBB, true);
1566 
1567    Instruction *ld =
1568       bld.mkLoad(TYPE_U32, atom->getDef(0), atom->getSrc(0)->asSym(),
1569                  atom->getIndirect(0, 0));
1570    ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
1571    ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;
1572 
1573    Value *stVal;
1574    if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
1575       // Read the old value, and write the new one.
1576       stVal = atom->getSrc(1);
1577    } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1578       CmpInstruction *set =
1579          bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1580                    TYPE_U32, ld->getDef(0), atom->getSrc(1));
1581       set->setPredicate(CC_P, ld->getDef(1));
1582 
1583       Instruction *selp =
1584          bld.mkOp3(OP_SELP, TYPE_U32, bld.getSSA(), ld->getDef(0),
1585                    atom->getSrc(2), set->getDef(0));
1586       selp->src(2).mod = Modifier(NV50_IR_MOD_NOT);
1587       selp->setPredicate(CC_P, ld->getDef(1));
1588 
1589       stVal = selp->getDef(0);
1590    } else {
1591       operation op;
1592 
1593       switch (atom->subOp) {
1594       case NV50_IR_SUBOP_ATOM_ADD:
1595          op = OP_ADD;
1596          break;
1597       case NV50_IR_SUBOP_ATOM_AND:
1598          op = OP_AND;
1599          break;
1600       case NV50_IR_SUBOP_ATOM_OR:
1601          op = OP_OR;
1602          break;
1603       case NV50_IR_SUBOP_ATOM_XOR:
1604          op = OP_XOR;
1605          break;
1606       case NV50_IR_SUBOP_ATOM_MIN:
1607          op = OP_MIN;
1608          break;
1609       case NV50_IR_SUBOP_ATOM_MAX:
1610          op = OP_MAX;
1611          break;
1612       default:
1613          assert(0);
1614          return;
1615       }
1616 
1617       Instruction *i =
1618          bld.mkOp2(op, atom->dType, bld.getSSA(), ld->getDef(0),
1619                    atom->getSrc(1));
1620       i->setPredicate(CC_P, ld->getDef(1));
1621 
1622       stVal = i->getDef(0);
1623    }
1624 
1625    Instruction *st =
1626       bld.mkStore(OP_STORE, TYPE_U32, atom->getSrc(0)->asSym(),
1627                   atom->getIndirect(0, 0), stVal);
1628    st->setPredicate(CC_P, ld->getDef(1));
1629    st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;
1630 
1631    // Loop until the lock is acquired.
1632    bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_NOT_P, ld->getDef(1));
1633    tryLockAndSetBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::BACK);
1634    tryLockAndSetBB->cfg.attach(&joinBB->cfg, Graph::Edge::CROSS);
1635    bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
1636 
1637    bld.remove(atom);
1638 
1639    bld.setPosition(joinBB, false);
1640    bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1641 }
1642 
1643 bool
handleATOM(Instruction * atom)1644 NVC0LoweringPass::handleATOM(Instruction *atom)
1645 {
1646    SVSemantic sv;
1647    Value *ptr = atom->getIndirect(0, 0), *ind = atom->getIndirect(0, 1), *base;
1648 
1649    switch (atom->src(0).getFile()) {
1650    case FILE_MEMORY_LOCAL:
1651       sv = SV_LBASE;
1652       break;
1653    case FILE_MEMORY_SHARED:
1654       // For Fermi/Kepler, we have to use ld lock/st unlock to perform atomic
1655       // operations on shared memory. For Maxwell, ATOMS is enough.
1656       if (targ->getChipset() < NVISA_GK104_CHIPSET)
1657          handleSharedATOM(atom);
1658       else if (targ->getChipset() < NVISA_GM107_CHIPSET)
1659          handleSharedATOMNVE4(atom);
1660       return true;
1661    case FILE_MEMORY_GLOBAL:
1662       return true;
1663    default:
1664       assert(atom->src(0).getFile() == FILE_MEMORY_BUFFER);
1665       base = loadBufInfo64(ind, atom->getSrc(0)->reg.fileIndex * 16);
1666       assert(base->reg.size == 8);
1667       if (ptr)
1668          base = bld.mkOp2v(OP_ADD, TYPE_U64, base, base, ptr);
1669       assert(base->reg.size == 8);
1670       atom->setIndirect(0, 0, base);
1671       atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1672 
1673       // Harden against out-of-bounds accesses
1674       Value *offset = bld.loadImm(NULL, atom->getSrc(0)->reg.data.offset + typeSizeof(atom->sType));
1675       Value *length = loadBufLength32(ind, atom->getSrc(0)->reg.fileIndex * 16);
1676       Value *pred = new_LValue(func, FILE_PREDICATE);
1677       if (ptr)
1678          bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, ptr);
1679       bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
1680       atom->setPredicate(CC_NOT_P, pred);
1681       if (atom->defExists(0)) {
1682          Value *zero, *dst = atom->getDef(0);
1683          atom->setDef(0, bld.getSSA());
1684 
1685          bld.setPosition(atom, true);
1686          bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
1687             ->setPredicate(CC_P, pred);
1688          bld.mkOp2(OP_UNION, TYPE_U32, dst, atom->getDef(0), zero);
1689       }
1690 
1691       return true;
1692    }
1693    base =
1694       bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(), bld.mkSysVal(sv, 0));
1695 
1696    atom->setSrc(0, cloneShallow(func, atom->getSrc(0)));
1697    atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1698    if (ptr)
1699       base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr);
1700    atom->setIndirect(0, 1, NULL);
1701    atom->setIndirect(0, 0, base);
1702 
1703    return true;
1704 }
1705 
1706 bool
handleCasExch(Instruction * cas,bool needCctl)1707 NVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl)
1708 {
1709    if (targ->getChipset() < NVISA_GM107_CHIPSET) {
1710       if (cas->src(0).getFile() == FILE_MEMORY_SHARED) {
1711          // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM().
1712          return false;
1713       }
1714    }
1715 
1716    if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS &&
1717        cas->subOp != NV50_IR_SUBOP_ATOM_EXCH)
1718       return false;
1719    bld.setPosition(cas, true);
1720 
1721    if (needCctl) {
1722       Instruction *cctl = bld.mkOp1(OP_CCTL, TYPE_NONE, NULL, cas->getSrc(0));
1723       cctl->setIndirect(0, 0, cas->getIndirect(0, 0));
1724       cctl->fixed = 1;
1725       cctl->subOp = NV50_IR_SUBOP_CCTL_IV;
1726       if (cas->isPredicated())
1727          cctl->setPredicate(cas->cc, cas->getPredicate());
1728    }
1729 
1730    if (cas->subOp == NV50_IR_SUBOP_ATOM_CAS &&
1731        targ->getChipset() < NVISA_GV100_CHIPSET) {
1732       // CAS is crazy. It's 2nd source is a double reg, and the 3rd source
1733       // should be set to the high part of the double reg or bad things will
1734       // happen elsewhere in the universe.
1735       // Also, it sometimes returns the new value instead of the old one
1736       // under mysterious circumstances.
1737       DataType ty = typeOfSize(typeSizeof(cas->dType) * 2);
1738       Value *dreg = bld.getSSA(typeSizeof(ty));
1739       bld.setPosition(cas, false);
1740       bld.mkOp2(OP_MERGE, ty, dreg, cas->getSrc(1), cas->getSrc(2));
1741       cas->setSrc(1, dreg);
1742       cas->setSrc(2, dreg);
1743    }
1744 
1745    return true;
1746 }
1747 
1748 inline Value *
loadResInfo32(Value * ptr,uint32_t off,uint16_t base)1749 NVC0LoweringPass::loadResInfo32(Value *ptr, uint32_t off, uint16_t base)
1750 {
1751    uint8_t b = prog->driver->io.auxCBSlot;
1752    off += base;
1753 
1754    return bld.
1755       mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1756 }
1757 
1758 inline Value *
loadResInfo64(Value * ptr,uint32_t off,uint16_t base)1759 NVC0LoweringPass::loadResInfo64(Value *ptr, uint32_t off, uint16_t base)
1760 {
1761    uint8_t b = prog->driver->io.auxCBSlot;
1762    off += base;
1763 
1764    if (ptr)
1765       ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
1766 
1767    return bld.
1768       mkLoadv(TYPE_U64, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off), ptr);
1769 }
1770 
1771 inline Value *
loadResLength32(Value * ptr,uint32_t off,uint16_t base)1772 NVC0LoweringPass::loadResLength32(Value *ptr, uint32_t off, uint16_t base)
1773 {
1774    uint8_t b = prog->driver->io.auxCBSlot;
1775    off += base;
1776 
1777    if (ptr)
1778       ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
1779 
1780    return bld.
1781       mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr);
1782 }
1783 
1784 inline Value *
loadBufInfo64(Value * ptr,uint32_t off)1785 NVC0LoweringPass::loadBufInfo64(Value *ptr, uint32_t off)
1786 {
1787    return loadResInfo64(ptr, off, prog->driver->io.bufInfoBase);
1788 }
1789 
1790 inline Value *
loadBufLength32(Value * ptr,uint32_t off)1791 NVC0LoweringPass::loadBufLength32(Value *ptr, uint32_t off)
1792 {
1793    return loadResLength32(ptr, off, prog->driver->io.bufInfoBase);
1794 }
1795 
1796 inline Value *
loadUboInfo64(Value * ptr,uint32_t off)1797 NVC0LoweringPass::loadUboInfo64(Value *ptr, uint32_t off)
1798 {
1799    return loadResInfo64(ptr, off, prog->driver->io.uboInfoBase);
1800 }
1801 
1802 inline Value *
loadUboLength32(Value * ptr,uint32_t off)1803 NVC0LoweringPass::loadUboLength32(Value *ptr, uint32_t off)
1804 {
1805    return loadResLength32(ptr, off, prog->driver->io.uboInfoBase);
1806 }
1807 
1808 inline Value *
loadMsInfo32(Value * ptr,uint32_t off)1809 NVC0LoweringPass::loadMsInfo32(Value *ptr, uint32_t off)
1810 {
1811    uint8_t b = prog->driver->io.msInfoCBSlot;
1812    off += prog->driver->io.msInfoBase;
1813    return bld.
1814       mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1815 }
1816 
1817 inline Value *
loadSuInfo32(Value * ptr,int slot,uint32_t off,bool bindless)1818 NVC0LoweringPass::loadSuInfo32(Value *ptr, int slot, uint32_t off, bool bindless)
1819 {
1820    uint32_t base = slot * NVC0_SU_INFO__STRIDE;
1821 
1822    // We don't upload surface info for bindless for GM107+
1823    assert(!bindless || targ->getChipset() < NVISA_GM107_CHIPSET);
1824 
1825    if (ptr) {
1826       ptr = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(slot));
1827       if (bindless)
1828          ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(511));
1829       else
1830          ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(7));
1831       ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(6));
1832       base = 0;
1833    }
1834    off += base;
1835 
1836    return loadResInfo32(ptr, off, bindless ? prog->driver->io.bindlessBase :
1837                         prog->driver->io.suInfoBase);
1838 }
1839 
1840 Value *
loadMsAdjInfo32(TexInstruction::Target target,uint32_t index,int slot,Value * ind,bool bindless)1841 NVC0LoweringPass::loadMsAdjInfo32(TexInstruction::Target target, uint32_t index, int slot, Value *ind, bool bindless)
1842 {
1843    if (!bindless || targ->getChipset() < NVISA_GM107_CHIPSET)
1844       return loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(index), bindless);
1845 
1846    assert(bindless);
1847 
1848    Value *samples = bld.getSSA();
1849    // this shouldn't be lowered because it's being inserted before the current instruction
1850    TexInstruction *tex = new_TexInstruction(func, OP_TXQ);
1851    tex->tex.target = target;
1852    tex->tex.query = TXQ_TYPE;
1853    tex->tex.mask = 0x4;
1854    tex->tex.r = 0xff;
1855    tex->tex.s = 0x1f;
1856    tex->tex.rIndirectSrc = 0;
1857    tex->setDef(0, samples);
1858    tex->setSrc(0, ind);
1859    tex->setSrc(1, bld.loadImm(NULL, 0));
1860    bld.insert(tex);
1861 
1862    // doesn't work with sample counts other than 1/2/4/8 but they aren't supported
1863    switch (index) {
1864    case 0: {
1865       Value *tmp = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), samples, bld.mkImm(2));
1866       return bld.mkOp2v(OP_SHR, TYPE_U32, bld.getSSA(), tmp, bld.mkImm(2));
1867    }
1868    case 1: {
1869       Value *tmp = bld.mkCmp(OP_SET, CC_GT, TYPE_U32, bld.getSSA(), TYPE_U32, samples, bld.mkImm(2))->getDef(0);
1870       return bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), tmp, bld.mkImm(1));
1871    }
1872    default: {
1873       assert(false);
1874       return NULL;
1875    }
1876    }
1877 }
1878 
getSuClampSubOp(const TexInstruction * su,int c)1879 static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
1880 {
1881    switch (su->tex.target.getEnum()) {
1882    case TEX_TARGET_BUFFER:      return NV50_IR_SUBOP_SUCLAMP_PL(0, 1);
1883    case TEX_TARGET_RECT:        return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1884    case TEX_TARGET_1D:          return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1885    case TEX_TARGET_1D_ARRAY:    return (c == 1) ?
1886                                    NV50_IR_SUBOP_SUCLAMP_PL(0, 2) :
1887                                    NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1888    case TEX_TARGET_2D:          return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1889    case TEX_TARGET_2D_MS:       return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1890    case TEX_TARGET_2D_ARRAY:    return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1891    case TEX_TARGET_2D_MS_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1892    case TEX_TARGET_3D:          return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1893    case TEX_TARGET_CUBE:        return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1894    case TEX_TARGET_CUBE_ARRAY:  return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1895    default:
1896       assert(0);
1897       return 0;
1898    }
1899 }
1900 
1901 bool
handleSUQ(TexInstruction * suq)1902 NVC0LoweringPass::handleSUQ(TexInstruction *suq)
1903 {
1904    int mask = suq->tex.mask;
1905    int dim = suq->tex.target.getDim();
1906    int arg = dim + (suq->tex.target.isArray() || suq->tex.target.isCube());
1907    Value *ind = suq->getIndirectR();
1908    int slot = suq->tex.r;
1909    int c, d;
1910 
1911    for (c = 0, d = 0; c < 3; ++c, mask >>= 1) {
1912       if (c >= arg || !(mask & 1))
1913          continue;
1914 
1915       int offset;
1916 
1917       if (c == 1 && suq->tex.target == TEX_TARGET_1D_ARRAY) {
1918          offset = NVC0_SU_INFO_SIZE(2);
1919       } else {
1920          offset = NVC0_SU_INFO_SIZE(c);
1921       }
1922       bld.mkMov(suq->getDef(d++), loadSuInfo32(ind, slot, offset, suq->tex.bindless));
1923       if (c == 2 && suq->tex.target.isCube())
1924          bld.mkOp2(OP_DIV, TYPE_U32, suq->getDef(d - 1), suq->getDef(d - 1),
1925                    bld.loadImm(NULL, 6));
1926    }
1927 
1928    if (mask & 1) {
1929       if (suq->tex.target.isMS()) {
1930          Value *ms_x = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(0), suq->tex.bindless);
1931          Value *ms_y = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(1), suq->tex.bindless);
1932          Value *ms = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(), ms_x, ms_y);
1933          bld.mkOp2(OP_SHL, TYPE_U32, suq->getDef(d++), bld.loadImm(NULL, 1), ms);
1934       } else {
1935          bld.mkMov(suq->getDef(d++), bld.loadImm(NULL, 1));
1936       }
1937    }
1938 
1939    bld.remove(suq);
1940    return true;
1941 }
1942 
1943 void
adjustCoordinatesMS(TexInstruction * tex)1944 NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
1945 {
1946    const int arg = tex->tex.target.getArgCount();
1947    int slot = tex->tex.r;
1948 
1949    if (tex->tex.target == TEX_TARGET_2D_MS)
1950       tex->tex.target = TEX_TARGET_2D;
1951    else
1952    if (tex->tex.target == TEX_TARGET_2D_MS_ARRAY)
1953       tex->tex.target = TEX_TARGET_2D_ARRAY;
1954    else
1955       return;
1956 
1957    Value *x = tex->getSrc(0);
1958    Value *y = tex->getSrc(1);
1959    Value *s = tex->getSrc(arg - 1);
1960 
1961    Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
1962    Value *ind = tex->getIndirectR();
1963 
1964    Value *ms_x = loadMsAdjInfo32(tex->tex.target, 0, slot, ind, tex->tex.bindless);
1965    Value *ms_y = loadMsAdjInfo32(tex->tex.target, 1, slot, ind, tex->tex.bindless);
1966 
1967    bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
1968    bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);
1969 
1970    s = bld.mkOp2v(OP_AND, TYPE_U32, ts, s, bld.loadImm(NULL, 0x7));
1971    s = bld.mkOp2v(OP_SHL, TYPE_U32, ts, ts, bld.mkImm(3));
1972 
1973    Value *dx = loadMsInfo32(ts, 0x0);
1974    Value *dy = loadMsInfo32(ts, 0x4);
1975 
1976    bld.mkOp2(OP_ADD, TYPE_U32, tx, tx, dx);
1977    bld.mkOp2(OP_ADD, TYPE_U32, ty, ty, dy);
1978 
1979    tex->setSrc(0, tx);
1980    tex->setSrc(1, ty);
1981    tex->moveSources(arg, -1);
1982 }
1983 
1984 // Sets 64-bit "generic address", predicate and format sources for SULD/SUST.
1985 // They're computed from the coordinates using the surface info in c[] space.
1986 void
processSurfaceCoordsNVE4(TexInstruction * su)1987 NVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction *su)
1988 {
1989    Instruction *insn;
1990    const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP;
1991    const bool raw =
1992       su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB;
1993    const int slot = su->tex.r;
1994    const int dim = su->tex.target.getDim();
1995    const bool array = su->tex.target.isArray() || su->tex.target.isCube();
1996    const int arg = dim + array;
1997    int c;
1998    Value *zero = bld.mkImm(0);
1999    Value *p1 = NULL;
2000    Value *v;
2001    Value *src[3];
2002    Value *bf, *eau, *off;
2003    Value *addr, *pred;
2004    Value *ind = su->getIndirectR();
2005    Value *y, *z;
2006 
2007    off = bld.getScratch(4);
2008    bf = bld.getScratch(4);
2009    addr = bld.getSSA(8);
2010    pred = bld.getScratch(1, FILE_PREDICATE);
2011 
2012    bld.setPosition(su, false);
2013 
2014    adjustCoordinatesMS(su);
2015 
2016    // calculate clamped coordinates
2017    for (c = 0; c < arg; ++c) {
2018       int dimc = c;
2019 
2020       if (c == 1 && su->tex.target == TEX_TARGET_1D_ARRAY) {
2021          // The array index is stored in the Z component for 1D arrays.
2022          dimc = 2;
2023       }
2024 
2025       src[c] = bld.getScratch();
2026       if (c == 0 && raw)
2027          v = loadSuInfo32(ind, slot, NVC0_SU_INFO_RAW_X, su->tex.bindless);
2028       else
2029          v = loadSuInfo32(ind, slot, NVC0_SU_INFO_DIM(dimc), su->tex.bindless);
2030       bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero)
2031          ->subOp = getSuClampSubOp(su, dimc);
2032    }
2033    for (; c < 3; ++c)
2034       src[c] = zero;
2035 
2036    if (dim == 2 && !array) {
2037       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C, su->tex.bindless);
2038       src[2] = bld.mkOp2v(OP_SHR, TYPE_U32, bld.getSSA(),
2039                           v, bld.loadImm(NULL, 16));
2040 
2041       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_DIM(2), su->tex.bindless);
2042       bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[2], src[2], v, zero)
2043          ->subOp = NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
2044    }
2045 
2046    // set predicate output
2047    if (su->tex.target == TEX_TARGET_BUFFER) {
2048       src[0]->getInsn()->setFlagsDef(1, pred);
2049    } else
2050    if (array) {
2051       p1 = bld.getSSA(1, FILE_PREDICATE);
2052       src[dim]->getInsn()->setFlagsDef(1, p1);
2053    }
2054 
2055    // calculate pixel offset
2056    if (dim == 1) {
2057       y = z = zero;
2058       if (su->tex.target != TEX_TARGET_BUFFER)
2059          bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff));
2060    } else {
2061       y = src[1];
2062       z = src[2];
2063 
2064       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C, su->tex.bindless);
2065       bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1])
2066          ->subOp = NV50_IR_SUBOP_MADSP(4,4,8); // u16l u16l u16l
2067 
2068       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_PITCH, su->tex.bindless);
2069       bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0])
2070          ->subOp = array ?
2071          NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
2072    }
2073 
2074    // calculate effective address part 1
2075    if (su->tex.target == TEX_TARGET_BUFFER) {
2076       if (raw) {
2077          bf = src[0];
2078       } else {
2079          v = loadSuInfo32(ind, slot, NVC0_SU_INFO_FMT, su->tex.bindless);
2080          bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero)
2081             ->subOp = NV50_IR_SUBOP_V1(7,6,8|2);
2082       }
2083    } else {
2084       uint16_t subOp = 0;
2085 
2086       switch (dim) {
2087       case 1:
2088          break;
2089       case 2:
2090          if (array) {
2091             z = off;
2092          } else {
2093             subOp = NV50_IR_SUBOP_SUBFM_3D;
2094          }
2095          break;
2096       default:
2097          subOp = NV50_IR_SUBOP_SUBFM_3D;
2098          assert(dim == 3);
2099          break;
2100       }
2101       insn = bld.mkOp3(OP_SUBFM, TYPE_U32, bf, src[0], y, z);
2102       insn->subOp = subOp;
2103       insn->setFlagsDef(1, pred);
2104    }
2105 
2106    // part 2
2107    v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless);
2108 
2109    if (su->tex.target == TEX_TARGET_BUFFER) {
2110       eau = v;
2111    } else {
2112       eau = bld.mkOp3v(OP_SUEAU, TYPE_U32, bld.getScratch(4), off, bf, v);
2113    }
2114    // add array layer offset
2115    if (array) {
2116       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ARRAY, su->tex.bindless);
2117       if (dim == 1)
2118          bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau)
2119             ->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
2120       else
2121          bld.mkOp3(OP_MADSP, TYPE_U32, eau, v, src[2], eau)
2122             ->subOp = NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32
2123       // combine predicates
2124       assert(p1);
2125       bld.mkOp2(OP_OR, TYPE_U8, pred, pred, p1);
2126    }
2127 
2128    if (atom) {
2129       Value *lo = bf;
2130       if (su->tex.target == TEX_TARGET_BUFFER) {
2131          lo = zero;
2132          bld.mkMov(off, bf);
2133       }
2134       //  bf == g[] address & 0xff
2135       // eau == g[] address >> 8
2136       bld.mkOp3(OP_PERMT, TYPE_U32,  bf,   lo, bld.loadImm(NULL, 0x6540), eau);
2137       bld.mkOp3(OP_PERMT, TYPE_U32, eau, zero, bld.loadImm(NULL, 0x0007), eau);
2138    } else
2139    if (su->op == OP_SULDP && su->tex.target == TEX_TARGET_BUFFER) {
2140       // Convert from u32 to u8 address format, which is what the library code
2141       // doing SULDP currently uses.
2142       // XXX: can SUEAU do this ?
2143       // XXX: does it matter that we don't mask high bytes in bf ?
2144       // Grrr.
2145       bld.mkOp2(OP_SHR, TYPE_U32, off, bf, bld.mkImm(8));
2146       bld.mkOp2(OP_ADD, TYPE_U32, eau, eau, off);
2147    }
2148 
2149    bld.mkOp2(OP_MERGE, TYPE_U64, addr, bf, eau);
2150 
2151    if (atom && su->tex.target == TEX_TARGET_BUFFER)
2152       bld.mkOp2(OP_ADD, TYPE_U64, addr, addr, off);
2153 
2154    // let's just set it 0 for raw access and hope it works
2155    v = raw ?
2156       bld.mkImm(0) : loadSuInfo32(ind, slot, NVC0_SU_INFO_FMT, su->tex.bindless);
2157 
2158    // get rid of old coordinate sources, make space for fmt info and predicate
2159    su->moveSources(arg, 3 - arg);
2160    // set 64 bit address and 32-bit format sources
2161    su->setSrc(0, addr);
2162    su->setSrc(1, v);
2163    su->setSrc(2, pred);
2164    su->setIndirectR(NULL);
2165 
2166    // prevent read fault when the image is not actually bound
2167    CmpInstruction *pred1 =
2168       bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
2169                 TYPE_U32, bld.mkImm(0),
2170                 loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless));
2171 
2172    if (su->op != OP_SUSTP && su->tex.format) {
2173       const TexInstruction::ImgFormatDesc *format = su->tex.format;
2174       int blockwidth = format->bits[0] + format->bits[1] +
2175                        format->bits[2] + format->bits[3];
2176 
2177       // make sure that the format doesn't mismatch
2178       assert(format->components != 0);
2179       bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred1->getDef(0),
2180                 TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
2181                 loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless),
2182                 pred1->getDef(0));
2183    }
2184    su->setPredicate(CC_NOT_P, pred1->getDef(0));
2185 
2186    // TODO: initialize def values to 0 when the surface operation is not
2187    // performed (not needed for stores). Also, fix the "address bounds test"
2188    // subtests from arb_shader_image_load_store-invalid for buffers, because it
2189    // seems like that the predicate is not correctly set by suclamp.
2190 }
2191 
2192 static DataType
getSrcType(const TexInstruction::ImgFormatDesc * t,int c)2193 getSrcType(const TexInstruction::ImgFormatDesc *t, int c)
2194 {
2195    switch (t->type) {
2196    case FLOAT: return t->bits[c] == 16 ? TYPE_F16 : TYPE_F32;
2197    case UNORM: return t->bits[c] == 8 ? TYPE_U8 : TYPE_U16;
2198    case SNORM: return t->bits[c] == 8 ? TYPE_S8 : TYPE_S16;
2199    case UINT:
2200       return (t->bits[c] == 8 ? TYPE_U8 :
2201               (t->bits[c] == 16 ? TYPE_U16 : TYPE_U32));
2202    case SINT:
2203       return (t->bits[c] == 8 ? TYPE_S8 :
2204               (t->bits[c] == 16 ? TYPE_S16 : TYPE_S32));
2205    }
2206    return TYPE_NONE;
2207 }
2208 
2209 static DataType
getDestType(const ImgType type)2210 getDestType(const ImgType type) {
2211    switch (type) {
2212    case FLOAT:
2213    case UNORM:
2214    case SNORM:
2215       return TYPE_F32;
2216    case UINT:
2217       return TYPE_U32;
2218    case SINT:
2219       return TYPE_S32;
2220    default:
2221       assert(!"Impossible type");
2222       return TYPE_NONE;
2223    }
2224 }
2225 
2226 void
convertSurfaceFormat(TexInstruction * su,Instruction ** loaded)2227 NVC0LoweringPass::convertSurfaceFormat(TexInstruction *su, Instruction **loaded)
2228 {
2229    const TexInstruction::ImgFormatDesc *format = su->tex.format;
2230    int width = format->bits[0] + format->bits[1] +
2231       format->bits[2] + format->bits[3];
2232    Value *untypedDst[4] = {};
2233    Value *typedDst[4] = {};
2234 
2235    // We must convert this to a generic load.
2236    su->op = OP_SULDB;
2237 
2238    su->dType = typeOfSize(width / 8);
2239    su->sType = TYPE_U8;
2240 
2241    for (int i = 0; i < width / 32; i++)
2242       untypedDst[i] = bld.getSSA();
2243    if (width < 32)
2244       untypedDst[0] = bld.getSSA();
2245 
2246    if (loaded && loaded[0]) {
2247       for (int i = 0; i < 4; i++) {
2248          if (loaded[i])
2249             typedDst[i] = loaded[i]->getDef(0);
2250       }
2251    } else {
2252       for (int i = 0; i < 4; i++) {
2253          typedDst[i] = su->getDef(i);
2254       }
2255    }
2256 
2257    // Set the untyped dsts as the su's destinations
2258    if (loaded && loaded[0]) {
2259       for (int i = 0; i < 4; i++)
2260          if (loaded[i])
2261             loaded[i]->setDef(0, untypedDst[i]);
2262    } else {
2263       for (int i = 0; i < 4; i++)
2264          su->setDef(i, untypedDst[i]);
2265 
2266       bld.setPosition(su, true);
2267    }
2268 
2269    // Unpack each component into the typed dsts
2270    int bits = 0;
2271    for (int i = 0; i < 4; bits += format->bits[i], i++) {
2272       if (!typedDst[i])
2273          continue;
2274 
2275       if (loaded && loaded[0])
2276          bld.setPosition(loaded[i], true);
2277 
2278       if (i >= format->components) {
2279          if (format->type == FLOAT ||
2280              format->type == UNORM ||
2281              format->type == SNORM)
2282             bld.loadImm(typedDst[i], i == 3 ? 1.0f : 0.0f);
2283          else
2284             bld.loadImm(typedDst[i], i == 3 ? 1 : 0);
2285          continue;
2286       }
2287 
2288       // Get just that component's data into the relevant place
2289       if (format->bits[i] == 32)
2290          bld.mkMov(typedDst[i], untypedDst[i]);
2291       else if (format->bits[i] == 16)
2292          bld.mkCvt(OP_CVT, getDestType(format->type), typedDst[i],
2293                    getSrcType(format, i), untypedDst[i / 2])
2294          ->subOp = (i & 1) << (format->type == FLOAT ? 0 : 1);
2295       else if (format->bits[i] == 8)
2296          bld.mkCvt(OP_CVT, getDestType(format->type), typedDst[i],
2297                    getSrcType(format, i), untypedDst[0])->subOp = i;
2298       else {
2299          bld.mkOp2(OP_EXTBF, TYPE_U32, typedDst[i], untypedDst[bits / 32],
2300                    bld.mkImm((bits % 32) | (format->bits[i] << 8)));
2301          if (format->type == UNORM || format->type == SNORM)
2302             bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], getSrcType(format, i), typedDst[i]);
2303       }
2304 
2305       // Normalize / convert as necessary
2306       if (format->type == UNORM)
2307          bld.mkOp2(OP_MUL, TYPE_F32, typedDst[i], typedDst[i], bld.loadImm(NULL, 1.0f / ((1 << format->bits[i]) - 1)));
2308       else if (format->type == SNORM)
2309          bld.mkOp2(OP_MUL, TYPE_F32, typedDst[i], typedDst[i], bld.loadImm(NULL, 1.0f / ((1 << (format->bits[i] - 1)) - 1)));
2310       else if (format->type == FLOAT && format->bits[i] < 16) {
2311          bld.mkOp2(OP_SHL, TYPE_U32, typedDst[i], typedDst[i], bld.loadImm(NULL, 15 - format->bits[i]));
2312          bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], TYPE_F16, typedDst[i]);
2313       }
2314    }
2315 
2316    if (format->bgra) {
2317       std::swap(typedDst[0], typedDst[2]);
2318    }
2319 }
2320 
2321 void
insertOOBSurfaceOpResult(TexInstruction * su)2322 NVC0LoweringPass::insertOOBSurfaceOpResult(TexInstruction *su)
2323 {
2324    if (!su->getPredicate())
2325       return;
2326 
2327    bld.setPosition(su, true);
2328 
2329    for (unsigned i = 0; su->defExists(i); ++i) {
2330       ValueDef &def = su->def(i);
2331 
2332       Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2333       assert(su->cc == CC_NOT_P);
2334       mov->setPredicate(CC_P, su->getPredicate());
2335       Instruction *uni = bld.mkOp2(OP_UNION, TYPE_U32, bld.getSSA(), NULL, mov->getDef(0));
2336 
2337       def.replace(uni->getDef(0), false);
2338       uni->setSrc(0, def.get());
2339    }
2340 }
2341 
2342 void
handleSurfaceOpNVE4(TexInstruction * su)2343 NVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction *su)
2344 {
2345    processSurfaceCoordsNVE4(su);
2346 
2347    if (su->op == OP_SULDP) {
2348       convertSurfaceFormat(su, NULL);
2349       insertOOBSurfaceOpResult(su);
2350    }
2351 
2352    if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
2353       assert(su->getPredicate());
2354       Value *pred =
2355          bld.mkOp2v(OP_OR, TYPE_U8, bld.getScratch(1, FILE_PREDICATE),
2356                     su->getPredicate(), su->getSrc(2));
2357 
2358       Instruction *red = bld.mkOp(OP_ATOM, su->dType, bld.getSSA());
2359       red->subOp = su->subOp;
2360       red->setSrc(0, bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0));
2361       red->setSrc(1, su->getSrc(3));
2362       if (su->subOp == NV50_IR_SUBOP_ATOM_CAS)
2363          red->setSrc(2, su->getSrc(4));
2364       red->setIndirect(0, 0, su->getSrc(0));
2365 
2366       // make sure to initialize dst value when the atomic operation is not
2367       // performed
2368       Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2369 
2370       assert(su->cc == CC_NOT_P);
2371       red->setPredicate(su->cc, pred);
2372       mov->setPredicate(CC_P, pred);
2373 
2374       bld.mkOp2(OP_UNION, TYPE_U32, su->getDef(0),
2375                 red->getDef(0), mov->getDef(0));
2376 
2377       delete_Instruction(bld.getProgram(), su);
2378       handleCasExch(red, true);
2379    }
2380 
2381    if (su->op == OP_SUSTB || su->op == OP_SUSTP)
2382       su->sType = (su->tex.target == TEX_TARGET_BUFFER) ? TYPE_U32 : TYPE_U8;
2383 }
2384 
2385 void
processSurfaceCoordsNVC0(TexInstruction * su)2386 NVC0LoweringPass::processSurfaceCoordsNVC0(TexInstruction *su)
2387 {
2388    const int slot = su->tex.r;
2389    const int dim = su->tex.target.getDim();
2390    const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
2391    int c;
2392    Value *zero = bld.mkImm(0);
2393    Value *src[3];
2394    Value *v;
2395    Value *ind = su->getIndirectR();
2396 
2397    bld.setPosition(su, false);
2398 
2399    adjustCoordinatesMS(su);
2400 
2401    if (ind) {
2402       Value *ptr;
2403       ptr = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind, bld.mkImm(su->tex.r));
2404       ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(7));
2405       su->setIndirectR(ptr);
2406    }
2407 
2408    // get surface coordinates
2409    for (c = 0; c < arg; ++c)
2410       src[c] = su->getSrc(c);
2411    for (; c < 3; ++c)
2412       src[c] = zero;
2413 
2414    // calculate pixel offset
2415    if (su->op == OP_SULDP || su->op == OP_SUREDP) {
2416       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless);
2417       su->setSrc(0, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[0], v));
2418    }
2419 
2420    // add array layer offset
2421    if (su->tex.target.isArray() || su->tex.target.isCube()) {
2422       v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ARRAY, su->tex.bindless);
2423       assert(dim > 1);
2424       su->setSrc(2, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[2], v));
2425    }
2426 
2427    // prevent read fault when the image is not actually bound
2428    CmpInstruction *pred =
2429       bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
2430                 TYPE_U32, bld.mkImm(0),
2431                 loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless));
2432    if (su->op != OP_SUSTP && su->tex.format) {
2433       const TexInstruction::ImgFormatDesc *format = su->tex.format;
2434       int blockwidth = format->bits[0] + format->bits[1] +
2435                        format->bits[2] + format->bits[3];
2436 
2437       assert(format->components != 0);
2438       // make sure that the format doesn't mismatch when it's not FMT_NONE
2439       bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred->getDef(0),
2440                 TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
2441                 loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless),
2442                 pred->getDef(0));
2443    }
2444    su->setPredicate(CC_NOT_P, pred->getDef(0));
2445 }
2446 
2447 void
handleSurfaceOpNVC0(TexInstruction * su)2448 NVC0LoweringPass::handleSurfaceOpNVC0(TexInstruction *su)
2449 {
2450    if (su->tex.target == TEX_TARGET_1D_ARRAY) {
2451       /* As 1d arrays also need 3 coordinates, switching to TEX_TARGET_2D_ARRAY
2452        * will simplify the lowering pass and the texture constraints. */
2453       su->moveSources(1, 1);
2454       su->setSrc(1, bld.loadImm(NULL, 0));
2455       su->tex.target = TEX_TARGET_2D_ARRAY;
2456    }
2457 
2458    processSurfaceCoordsNVC0(su);
2459 
2460    if (su->op == OP_SULDP) {
2461       convertSurfaceFormat(su, NULL);
2462       insertOOBSurfaceOpResult(su);
2463    }
2464 
2465    if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
2466       const int dim = su->tex.target.getDim();
2467       const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
2468       LValue *addr = bld.getSSA(8);
2469       Value *def = su->getDef(0);
2470 
2471       su->op = OP_SULEA;
2472 
2473       // Set the destination to the address
2474       su->dType = TYPE_U64;
2475       su->setDef(0, addr);
2476       su->setDef(1, su->getPredicate());
2477 
2478       bld.setPosition(su, true);
2479 
2480       // Perform the atomic op
2481       Instruction *red = bld.mkOp(OP_ATOM, su->sType, bld.getSSA());
2482       red->subOp = su->subOp;
2483       red->setSrc(0, bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, su->sType, 0));
2484       red->setSrc(1, su->getSrc(arg));
2485       if (red->subOp == NV50_IR_SUBOP_ATOM_CAS)
2486          red->setSrc(2, su->getSrc(arg + 1));
2487       red->setIndirect(0, 0, addr);
2488 
2489       // make sure to initialize dst value when the atomic operation is not
2490       // performed
2491       Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2492 
2493       assert(su->cc == CC_NOT_P);
2494       red->setPredicate(su->cc, su->getPredicate());
2495       mov->setPredicate(CC_P, su->getPredicate());
2496 
2497       bld.mkOp2(OP_UNION, TYPE_U32, def, red->getDef(0), mov->getDef(0));
2498 
2499       handleCasExch(red, false);
2500    }
2501 }
2502 
2503 TexInstruction *
processSurfaceCoordsGM107(TexInstruction * su,Instruction * ret[4])2504 NVC0LoweringPass::processSurfaceCoordsGM107(TexInstruction *su, Instruction *ret[4])
2505 {
2506    const int slot = su->tex.r;
2507    const int dim = su->tex.target.getDim();
2508    const bool array = su->tex.target.isArray() || su->tex.target.isCube();
2509    const int arg = dim + array;
2510    Value *ind = su->getIndirectR();
2511    Value *handle;
2512    Instruction *pred = NULL, *pred2d = NULL;
2513    int pos = 0;
2514 
2515    bld.setPosition(su, false);
2516 
2517    adjustCoordinatesMS(su);
2518 
2519    // add texture handle
2520    switch (su->op) {
2521    case OP_SUSTP:
2522       pos = 4;
2523       break;
2524    case OP_SUREDP:
2525       pos = (su->subOp == NV50_IR_SUBOP_ATOM_CAS) ? 2 : 1;
2526       break;
2527    default:
2528       assert(pos == 0);
2529       break;
2530    }
2531 
2532    if (dim == 2 && !array) {
2533       // This might be a 2d slice of a 3d texture, try to load the z
2534       // coordinate in.
2535       Value *v;
2536       if (!su->tex.bindless)
2537          v = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C, su->tex.bindless);
2538       else
2539          v = bld.mkOp2v(OP_SHR, TYPE_U32, bld.getSSA(), ind, bld.mkImm(11));
2540       Value *is_3d = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), v, bld.mkImm(1));
2541       pred2d = bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
2542                          TYPE_U32, bld.mkImm(0), is_3d);
2543 
2544       bld.mkOp2(OP_SHR, TYPE_U32, v, v, bld.loadImm(NULL, 16));
2545       su->moveSources(dim, 1);
2546       su->setSrc(dim, v);
2547       su->tex.target = nv50_ir::TEX_TARGET_3D;
2548       pos++;
2549    }
2550 
2551    if (su->tex.bindless)
2552       handle = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ind, bld.mkImm(2047));
2553    else
2554       handle = loadTexHandle(ind, slot + 32);
2555 
2556    su->setSrc(arg + pos, handle);
2557 
2558    // The address check doesn't make sense here. The format check could make
2559    // sense but it's a bit of a pain.
2560    if (!su->tex.bindless) {
2561       // prevent read fault when the image is not actually bound
2562       pred =
2563          bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
2564                    TYPE_U32, bld.mkImm(0),
2565                    loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR, su->tex.bindless));
2566       if (su->op != OP_SUSTP && su->tex.format) {
2567          const TexInstruction::ImgFormatDesc *format = su->tex.format;
2568          int blockwidth = format->bits[0] + format->bits[1] +
2569             format->bits[2] + format->bits[3];
2570 
2571          assert(format->components != 0);
2572          // make sure that the format doesn't mismatch when it's not FMT_NONE
2573          bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred->getDef(0),
2574                    TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
2575                    loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE, su->tex.bindless),
2576                    pred->getDef(0));
2577       }
2578    }
2579 
2580    // Now we have "pred" which (optionally) contains whether to do the surface
2581    // op at all, and a "pred2d" which indicates that, in case of doing the
2582    // surface op, we have to create a 2d and 3d version, conditioned on pred2d.
2583    TexInstruction *su2d = NULL;
2584    if (pred2d) {
2585       su2d = cloneForward(func, su)->asTex();
2586       for (unsigned i = 0; su->defExists(i); ++i)
2587          su2d->setDef(i, bld.getSSA());
2588       su2d->moveSources(dim + 1, -1);
2589       su2d->tex.target = nv50_ir::TEX_TARGET_2D;
2590    }
2591    if (pred2d && pred) {
2592       Instruction *pred3d = bld.mkOp2(OP_AND, TYPE_U8,
2593                                       bld.getSSA(1, FILE_PREDICATE),
2594                                       pred->getDef(0), pred2d->getDef(0));
2595       pred3d->src(0).mod = Modifier(NV50_IR_MOD_NOT);
2596       pred3d->src(1).mod = Modifier(NV50_IR_MOD_NOT);
2597       su->setPredicate(CC_P, pred3d->getDef(0));
2598       pred2d = bld.mkOp2(OP_AND, TYPE_U8, bld.getSSA(1, FILE_PREDICATE),
2599                          pred->getDef(0), pred2d->getDef(0));
2600       pred2d->src(0).mod = Modifier(NV50_IR_MOD_NOT);
2601    } else if (pred) {
2602       su->setPredicate(CC_NOT_P, pred->getDef(0));
2603    } else if (pred2d) {
2604       su->setPredicate(CC_NOT_P, pred2d->getDef(0));
2605    }
2606    if (su2d) {
2607       su2d->setPredicate(CC_P, pred2d->getDef(0));
2608       bld.insert(su2d);
2609 
2610       // Create a UNION so that RA assigns the same registers
2611       bld.setPosition(su, true);
2612       for (unsigned i = 0; su->defExists(i); ++i) {
2613          assert(i < 4);
2614 
2615          ValueDef &def = su->def(i);
2616          ValueDef &def2 = su2d->def(i);
2617          Instruction *mov = NULL;
2618 
2619          if (pred) {
2620             mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2621             mov->setPredicate(CC_P, pred->getDef(0));
2622          }
2623 
2624          Instruction *uni = ret[i] = bld.mkOp2(OP_UNION, TYPE_U32,
2625                                       bld.getSSA(),
2626                                       NULL, def2.get());
2627          def.replace(uni->getDef(0), false);
2628          uni->setSrc(0, def.get());
2629          if (mov)
2630             uni->setSrc(2, mov->getDef(0));
2631       }
2632    } else if (pred) {
2633       // Create a UNION so that RA assigns the same registers
2634       bld.setPosition(su, true);
2635       for (unsigned i = 0; su->defExists(i); ++i) {
2636          assert(i < 4);
2637 
2638          ValueDef &def = su->def(i);
2639 
2640          Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));
2641          mov->setPredicate(CC_P, pred->getDef(0));
2642 
2643          Instruction *uni = ret[i] = bld.mkOp2(OP_UNION, TYPE_U32,
2644                                       bld.getSSA(),
2645                                       NULL, mov->getDef(0));
2646          def.replace(uni->getDef(0), false);
2647          uni->setSrc(0, def.get());
2648       }
2649    }
2650 
2651    return su2d;
2652 }
2653 
2654 void
handleSurfaceOpGM107(TexInstruction * su)2655 NVC0LoweringPass::handleSurfaceOpGM107(TexInstruction *su)
2656 {
2657    // processSurfaceCoords also takes care of fixing up the outputs and
2658    // union'ing them with 0 as necessary. Additionally it may create a second
2659    // surface which needs some of the similar fixups.
2660 
2661    Instruction *loaded[4] = {};
2662    TexInstruction *su2 = processSurfaceCoordsGM107(su, loaded);
2663 
2664    if (su->op == OP_SULDP) {
2665       convertSurfaceFormat(su, loaded);
2666    }
2667 
2668    if (su->op == OP_SUREDP) {
2669       su->op = OP_SUREDB;
2670    }
2671 
2672    // If we fixed up the type of the regular surface load instruction, we also
2673    // have to fix up the copy.
2674    if (su2) {
2675       su2->op = su->op;
2676       su2->dType = su->dType;
2677       su2->sType = su->sType;
2678    }
2679 }
2680 
2681 bool
handleWRSV(Instruction * i)2682 NVC0LoweringPass::handleWRSV(Instruction *i)
2683 {
2684    Instruction *st;
2685    Symbol *sym;
2686    uint32_t addr;
2687 
2688    // must replace, $sreg are not writeable
2689    addr = targ->getSVAddress(FILE_SHADER_OUTPUT, i->getSrc(0)->asSym());
2690    if (addr >= 0x400)
2691       return false;
2692    sym = bld.mkSymbol(FILE_SHADER_OUTPUT, 0, i->sType, addr);
2693 
2694    st = bld.mkStore(OP_EXPORT, i->dType, sym, i->getIndirect(0, 0),
2695                     i->getSrc(1));
2696    st->perPatch = i->perPatch;
2697 
2698    bld.getBB()->remove(i);
2699    return true;
2700 }
2701 
2702 void
handleLDST(Instruction * i)2703 NVC0LoweringPass::handleLDST(Instruction *i)
2704 {
2705    if (i->src(0).getFile() == FILE_SHADER_INPUT) {
2706       if (prog->getType() == Program::TYPE_COMPUTE) {
2707          i->getSrc(0)->reg.file = FILE_MEMORY_CONST;
2708          i->getSrc(0)->reg.fileIndex = 0;
2709       } else
2710       if (prog->getType() == Program::TYPE_GEOMETRY &&
2711           i->src(0).isIndirect(0)) {
2712          // XXX: this assumes vec4 units
2713          Value *ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
2714                                  i->getIndirect(0, 0), bld.mkImm(4));
2715          i->setIndirect(0, 0, ptr);
2716          i->op = OP_VFETCH;
2717       } else {
2718          i->op = OP_VFETCH;
2719          assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP
2720       }
2721    } else if (i->src(0).getFile() == FILE_MEMORY_CONST) {
2722       int8_t fileIndex = i->getSrc(0)->reg.fileIndex - 1;
2723       Value *ind = i->getIndirect(0, 1);
2724 
2725       if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
2726           prog->getType() == Program::TYPE_COMPUTE &&
2727           (fileIndex >= 6 || ind)) {
2728          // The launch descriptor only allows to set up 8 CBs, but OpenGL
2729          // requires at least 12 UBOs. To bypass this limitation, for constant
2730          // buffers 7+, we store the addrs into the driver constbuf and we
2731          // directly load from the global memory.
2732          if (ind) {
2733             // Clamp the UBO index when an indirect access is used to avoid
2734             // loading information from the wrong place in the driver cb.
2735             // TODO - synchronize the max with the driver.
2736             ind = bld.mkOp2v(OP_MIN, TYPE_U32, bld.getSSA(),
2737                              bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(),
2738                                         ind, bld.loadImm(NULL, fileIndex)),
2739                              bld.loadImm(NULL, 13));
2740             fileIndex = 0;
2741          }
2742 
2743          Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
2744          Value *ptr = loadUboInfo64(ind, fileIndex * 16);
2745          Value *length = loadUboLength32(ind, fileIndex * 16);
2746          Value *pred = new_LValue(func, FILE_PREDICATE);
2747          if (i->src(0).isIndirect(0)) {
2748             bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
2749             bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
2750          }
2751          i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
2752          i->setIndirect(0, 1, NULL);
2753          i->setIndirect(0, 0, ptr);
2754          bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
2755          i->setPredicate(CC_NOT_P, pred);
2756          Value *zero, *dst = i->getDef(0);
2757          i->setDef(0, bld.getSSA());
2758 
2759          bld.setPosition(i, true);
2760          bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
2761             ->setPredicate(CC_P, pred);
2762          bld.mkOp2(OP_UNION, TYPE_U32, dst, i->getDef(0), zero);
2763       } else if (i->src(0).isIndirect(1)) {
2764          Value *ptr;
2765          if (i->src(0).isIndirect(0))
2766             ptr = bld.mkOp3v(OP_INSBF, TYPE_U32, bld.getSSA(),
2767                              i->getIndirect(0, 1), bld.mkImm(0x1010),
2768                              i->getIndirect(0, 0));
2769          else
2770             ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
2771                              i->getIndirect(0, 1), bld.mkImm(16));
2772          i->setIndirect(0, 1, NULL);
2773          i->setIndirect(0, 0, ptr);
2774          i->subOp = NV50_IR_SUBOP_LDC_IS;
2775       }
2776    } else if (i->src(0).getFile() == FILE_SHADER_OUTPUT) {
2777       assert(prog->getType() == Program::TYPE_TESSELLATION_CONTROL);
2778       i->op = OP_VFETCH;
2779    } else if (i->src(0).getFile() == FILE_MEMORY_BUFFER) {
2780       Value *ind = i->getIndirect(0, 1);
2781       Value *ptr = loadBufInfo64(ind, i->getSrc(0)->reg.fileIndex * 16);
2782       // XXX come up with a way not to do this for EVERY little access but
2783       // rather to batch these up somehow. Unfortunately we've lost the
2784       // information about the field width by the time we get here.
2785       Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
2786       Value *length = loadBufLength32(ind, i->getSrc(0)->reg.fileIndex * 16);
2787       Value *pred = new_LValue(func, FILE_PREDICATE);
2788       if (i->src(0).isIndirect(0)) {
2789          bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
2790          bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
2791       }
2792       i->setIndirect(0, 1, NULL);
2793       i->setIndirect(0, 0, ptr);
2794       i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
2795       bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
2796       i->setPredicate(CC_NOT_P, pred);
2797       if (i->defExists(0)) {
2798          Value *zero, *dst = i->getDef(0);
2799          i->setDef(0, bld.getSSA());
2800 
2801          bld.setPosition(i, true);
2802          bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
2803             ->setPredicate(CC_P, pred);
2804          bld.mkOp2(OP_UNION, TYPE_U32, dst, i->getDef(0), zero);
2805       }
2806    }
2807 }
2808 
2809 void
readTessCoord(LValue * dst,int c)2810 NVC0LoweringPass::readTessCoord(LValue *dst, int c)
2811 {
2812    Value *laneid = bld.getSSA();
2813    Value *x, *y;
2814 
2815    bld.mkOp1(OP_RDSV, TYPE_U32, laneid, bld.mkSysVal(SV_LANEID, 0));
2816 
2817    if (c == 0) {
2818       x = dst;
2819       y = NULL;
2820    } else
2821    if (c == 1) {
2822       x = NULL;
2823       y = dst;
2824    } else {
2825       assert(c == 2);
2826       if (prog->driver_out->prop.tp.domain != PIPE_PRIM_TRIANGLES) {
2827          bld.mkMov(dst, bld.loadImm(NULL, 0));
2828          return;
2829       }
2830       x = bld.getSSA();
2831       y = bld.getSSA();
2832    }
2833    if (x)
2834       bld.mkFetch(x, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f0, NULL, laneid);
2835    if (y)
2836       bld.mkFetch(y, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f4, NULL, laneid);
2837 
2838    if (c == 2) {
2839       bld.mkOp2(OP_ADD, TYPE_F32, dst, x, y);
2840       bld.mkOp2(OP_SUB, TYPE_F32, dst, bld.loadImm(NULL, 1.0f), dst);
2841    }
2842 }
2843 
2844 bool
handleRDSV(Instruction * i)2845 NVC0LoweringPass::handleRDSV(Instruction *i)
2846 {
2847    Symbol *sym = i->getSrc(0)->asSym();
2848    const SVSemantic sv = sym->reg.data.sv.sv;
2849    Value *vtx = NULL;
2850    Instruction *ld;
2851    uint32_t addr = targ->getSVAddress(FILE_SHADER_INPUT, sym);
2852 
2853    if (addr >= 0x400) {
2854       // mov $sreg
2855       if (sym->reg.data.sv.index == 3) {
2856          // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
2857          i->op = OP_MOV;
2858          i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0));
2859       } else
2860       if (sv == SV_TID) {
2861          // Help CSE combine TID fetches
2862          Value *tid = bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(),
2863                                  bld.mkSysVal(SV_COMBINED_TID, 0));
2864          i->op = OP_EXTBF;
2865          i->setSrc(0, tid);
2866          switch (sym->reg.data.sv.index) {
2867          case 0: i->setSrc(1, bld.mkImm(0x1000)); break;
2868          case 1: i->setSrc(1, bld.mkImm(0x0a10)); break;
2869          case 2: i->setSrc(1, bld.mkImm(0x061a)); break;
2870          }
2871       }
2872       if (sv == SV_VERTEX_COUNT) {
2873          bld.setPosition(i, true);
2874          bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0), bld.mkImm(0x808));
2875       }
2876       return true;
2877    }
2878 
2879    switch (sv) {
2880    case SV_POSITION:
2881       assert(prog->getType() == Program::TYPE_FRAGMENT);
2882       if (i->srcExists(1)) {
2883          // Pass offset through to the interpolation logic
2884          ld = bld.mkInterp(NV50_IR_INTERP_LINEAR | NV50_IR_INTERP_OFFSET,
2885                            i->getDef(0), addr, NULL);
2886          ld->setSrc(1, i->getSrc(1));
2887       } else {
2888          bld.mkInterp(NV50_IR_INTERP_LINEAR, i->getDef(0), addr, NULL);
2889       }
2890       break;
2891    case SV_FACE:
2892    {
2893       Value *face = i->getDef(0);
2894       bld.mkInterp(NV50_IR_INTERP_FLAT, face, addr, NULL);
2895       if (i->dType == TYPE_F32) {
2896          bld.mkOp2(OP_OR, TYPE_U32, face, face, bld.mkImm(0x00000001));
2897          bld.mkOp1(OP_NEG, TYPE_S32, face, face);
2898          bld.mkCvt(OP_CVT, TYPE_F32, face, TYPE_S32, face);
2899       }
2900    }
2901       break;
2902    case SV_TESS_COORD:
2903       assert(prog->getType() == Program::TYPE_TESSELLATION_EVAL);
2904       readTessCoord(i->getDef(0)->asLValue(), i->getSrc(0)->reg.data.sv.index);
2905       break;
2906    case SV_NTID:
2907    case SV_NCTAID:
2908    case SV_GRIDID:
2909       assert(targ->getChipset() >= NVISA_GK104_CHIPSET); // mov $sreg otherwise
2910       if (sym->reg.data.sv.index == 3) {
2911          i->op = OP_MOV;
2912          i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1));
2913          return true;
2914       }
2915       // Fallthrough
2916    case SV_WORK_DIM:
2917       addr += prog->driver->prop.cp.gridInfoBase;
2918       bld.mkLoad(TYPE_U32, i->getDef(0),
2919                  bld.mkSymbol(FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2920                               TYPE_U32, addr), NULL);
2921       break;
2922    case SV_SAMPLE_INDEX:
2923       // TODO: Properly pass source as an address in the PIX address space
2924       // (which can be of the form [r0+offset]). But this is currently
2925       // unnecessary.
2926       ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2927       ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2928       break;
2929    case SV_SAMPLE_POS: {
2930       Value *sampleID = bld.getScratch();
2931       ld = bld.mkOp1(OP_PIXLD, TYPE_U32, sampleID, bld.mkImm(0));
2932       ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2933       Value *offset = calculateSampleOffset(sampleID);
2934 
2935       assert(prog->driver_out->prop.fp.readsSampleLocations);
2936 
2937       if (targ->getChipset() >= NVISA_GM200_CHIPSET) {
2938          bld.mkLoad(TYPE_F32,
2939                     i->getDef(0),
2940                     bld.mkSymbol(
2941                           FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2942                           TYPE_U32, prog->driver->io.sampleInfoBase),
2943                     offset);
2944          bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0),
2945                    bld.mkImm(0x040c + sym->reg.data.sv.index * 16));
2946          bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(0), TYPE_U32, i->getDef(0));
2947          bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(0), i->getDef(0), bld.mkImm(1.0f / 16.0f));
2948       } else {
2949          bld.mkLoad(TYPE_F32,
2950                     i->getDef(0),
2951                     bld.mkSymbol(
2952                           FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2953                           TYPE_U32, prog->driver->io.sampleInfoBase +
2954                           4 * sym->reg.data.sv.index),
2955                     offset);
2956       }
2957       break;
2958    }
2959    case SV_SAMPLE_MASK: {
2960       ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2961       ld->subOp = NV50_IR_SUBOP_PIXLD_COVMASK;
2962       Instruction *sampleid =
2963          bld.mkOp1(OP_PIXLD, TYPE_U32, bld.getSSA(), bld.mkImm(0));
2964       sampleid->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2965       Value *masked =
2966          bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ld->getDef(0),
2967                     bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
2968                                bld.loadImm(NULL, 1), sampleid->getDef(0)));
2969       if (prog->persampleInvocation) {
2970          bld.mkMov(i->getDef(0), masked);
2971       } else {
2972          bld.mkOp3(OP_SELP, TYPE_U32, i->getDef(0), ld->getDef(0), masked,
2973                    bld.mkImm(0))
2974             ->subOp = 1;
2975       }
2976       break;
2977    }
2978    case SV_BASEVERTEX:
2979    case SV_BASEINSTANCE:
2980    case SV_DRAWID:
2981       ld = bld.mkLoad(TYPE_U32, i->getDef(0),
2982                       bld.mkSymbol(FILE_MEMORY_CONST,
2983                                    prog->driver->io.auxCBSlot,
2984                                    TYPE_U32,
2985                                    prog->driver->io.drawInfoBase +
2986                                    4 * (sv - SV_BASEVERTEX)),
2987                       NULL);
2988       break;
2989    default:
2990       if (prog->getType() == Program::TYPE_TESSELLATION_EVAL && !i->perPatch)
2991          vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0));
2992       if (prog->getType() == Program::TYPE_FRAGMENT) {
2993          bld.mkInterp(NV50_IR_INTERP_FLAT, i->getDef(0), addr, NULL);
2994       } else {
2995          ld = bld.mkFetch(i->getDef(0), i->dType,
2996                           FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
2997          ld->perPatch = i->perPatch;
2998       }
2999       break;
3000    }
3001    bld.getBB()->remove(i);
3002    return true;
3003 }
3004 
3005 bool
handleDIV(Instruction * i)3006 NVC0LoweringPass::handleDIV(Instruction *i)
3007 {
3008    if (!isFloatType(i->dType))
3009       return true;
3010    bld.setPosition(i, false);
3011    Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(typeSizeof(i->dType)), i->getSrc(1));
3012    i->op = OP_MUL;
3013    i->setSrc(1, rcp->getDef(0));
3014    return true;
3015 }
3016 
3017 bool
handleMOD(Instruction * i)3018 NVC0LoweringPass::handleMOD(Instruction *i)
3019 {
3020    if (!isFloatType(i->dType))
3021       return true;
3022    LValue *value = bld.getScratch(typeSizeof(i->dType));
3023    bld.mkOp1(OP_RCP, i->dType, value, i->getSrc(1));
3024    bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(0), value);
3025    bld.mkOp1(OP_TRUNC, i->dType, value, value);
3026    bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(1), value);
3027    i->op = OP_SUB;
3028    i->setSrc(1, value);
3029    return true;
3030 }
3031 
3032 bool
handleSQRT(Instruction * i)3033 NVC0LoweringPass::handleSQRT(Instruction *i)
3034 {
3035    if (targ->isOpSupported(OP_SQRT, i->dType))
3036       return true;
3037 
3038    if (i->dType == TYPE_F64) {
3039       Value *pred = bld.getSSA(1, FILE_PREDICATE);
3040       Value *zero = bld.loadImm(NULL, 0.0);
3041       Value *dst = bld.getSSA(8);
3042       bld.mkOp1(OP_RSQ, i->dType, dst, i->getSrc(0));
3043       bld.mkCmp(OP_SET, CC_LE, i->dType, pred, i->dType, i->getSrc(0), zero);
3044       bld.mkOp3(OP_SELP, TYPE_U64, dst, zero, dst, pred);
3045       i->op = OP_MUL;
3046       i->setSrc(1, dst);
3047       // TODO: Handle this properly with a library function
3048    } else {
3049       bld.setPosition(i, true);
3050       i->op = OP_RSQ;
3051       bld.mkOp1(OP_RCP, i->dType, i->getDef(0), i->getDef(0));
3052    }
3053 
3054    return true;
3055 }
3056 
3057 bool
handlePOW(Instruction * i)3058 NVC0LoweringPass::handlePOW(Instruction *i)
3059 {
3060    LValue *val = bld.getScratch();
3061 
3062    bld.mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0));
3063    bld.mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1;
3064    bld.mkOp1(OP_PREEX2, TYPE_F32, val, val);
3065 
3066    i->op = OP_EX2;
3067    i->setSrc(0, val);
3068    i->setSrc(1, NULL);
3069 
3070    return true;
3071 }
3072 
3073 bool
handleEXPORT(Instruction * i)3074 NVC0LoweringPass::handleEXPORT(Instruction *i)
3075 {
3076    if (prog->getType() == Program::TYPE_FRAGMENT) {
3077       int id = i->getSrc(0)->reg.data.offset / 4;
3078 
3079       if (i->src(0).isIndirect(0)) // TODO, ugly
3080          return false;
3081       i->op = OP_MOV;
3082       i->subOp = NV50_IR_SUBOP_MOV_FINAL;
3083       i->src(0).set(i->src(1));
3084       i->setSrc(1, NULL);
3085       i->setDef(0, new_LValue(func, FILE_GPR));
3086       i->getDef(0)->reg.data.id = id;
3087 
3088       prog->maxGPR = MAX2(prog->maxGPR, id);
3089    } else
3090    if (prog->getType() == Program::TYPE_GEOMETRY) {
3091       i->setIndirect(0, 1, gpEmitAddress);
3092    }
3093    return true;
3094 }
3095 
3096 bool
handleOUT(Instruction * i)3097 NVC0LoweringPass::handleOUT(Instruction *i)
3098 {
3099    Instruction *prev = i->prev;
3100    ImmediateValue stream, prevStream;
3101 
3102    // Only merge if the stream ids match. Also, note that the previous
3103    // instruction would have already been lowered, so we take arg1 from it.
3104    if (i->op == OP_RESTART && prev && prev->op == OP_EMIT &&
3105        i->src(0).getImmediate(stream) &&
3106        prev->src(1).getImmediate(prevStream) &&
3107        stream.reg.data.u32 == prevStream.reg.data.u32) {
3108       i->prev->subOp = NV50_IR_SUBOP_EMIT_RESTART;
3109       delete_Instruction(prog, i);
3110    } else {
3111       assert(gpEmitAddress);
3112       i->setDef(0, gpEmitAddress);
3113       i->setSrc(1, i->getSrc(0));
3114       i->setSrc(0, gpEmitAddress);
3115    }
3116    return true;
3117 }
3118 
3119 Value *
calculateSampleOffset(Value * sampleID)3120 NVC0LoweringPass::calculateSampleOffset(Value *sampleID)
3121 {
3122    Value *offset = bld.getScratch();
3123    if (targ->getChipset() >= NVISA_GM200_CHIPSET) {
3124       // Sample location offsets (in bytes) are calculated like so:
3125       // offset = (SV_POSITION.y % 4 * 2) + (SV_POSITION.x % 2)
3126       // offset = offset * 32 + sampleID % 8 * 4;
3127       // which is equivalent to:
3128       // offset = (SV_POSITION.y & 0x3) << 6 + (SV_POSITION.x & 0x1) << 5;
3129       // offset += sampleID << 2
3130 
3131       // The second operand (src1) of the INSBF instructions are like so:
3132       // 0xssll where ss is the size and ll is the offset.
3133       // so: dest = src2 | (src0 & (1 << ss - 1)) << ll
3134 
3135       // Add sample ID (offset = (sampleID & 0x7) << 2)
3136       bld.mkOp3(OP_INSBF, TYPE_U32, offset, sampleID, bld.mkImm(0x0302), bld.mkImm(0x0));
3137 
3138       Symbol *xSym = bld.mkSysVal(SV_POSITION, 0);
3139       Symbol *ySym = bld.mkSysVal(SV_POSITION, 1);
3140       Value *coord = bld.getScratch();
3141 
3142       // Add X coordinate (offset |= (SV_POSITION.x & 0x1) << 5)
3143       bld.mkInterp(NV50_IR_INTERP_LINEAR, coord,
3144                    targ->getSVAddress(FILE_SHADER_INPUT, xSym), NULL);
3145       bld.mkCvt(OP_CVT, TYPE_U32, coord, TYPE_F32, coord)
3146          ->rnd = ROUND_ZI;
3147       bld.mkOp3(OP_INSBF, TYPE_U32, offset, coord, bld.mkImm(0x0105), offset);
3148 
3149       // Add Y coordinate (offset |= (SV_POSITION.y & 0x3) << 6)
3150       bld.mkInterp(NV50_IR_INTERP_LINEAR, coord,
3151                    targ->getSVAddress(FILE_SHADER_INPUT, ySym), NULL);
3152       bld.mkCvt(OP_CVT, TYPE_U32, coord, TYPE_F32, coord)
3153          ->rnd = ROUND_ZI;
3154       bld.mkOp3(OP_INSBF, TYPE_U32, offset, coord, bld.mkImm(0x0206), offset);
3155    } else {
3156       bld.mkOp2(OP_SHL, TYPE_U32, offset, sampleID, bld.mkImm(3));
3157    }
3158    return offset;
3159 }
3160 
3161 // Handle programmable sample locations for GM20x+
3162 void
handlePIXLD(Instruction * i)3163 NVC0LoweringPass::handlePIXLD(Instruction *i)
3164 {
3165    if (i->subOp != NV50_IR_SUBOP_PIXLD_OFFSET)
3166       return;
3167    if (targ->getChipset() < NVISA_GM200_CHIPSET)
3168       return;
3169 
3170    assert(prog->driver_out->prop.fp.readsSampleLocations);
3171 
3172    bld.mkLoad(TYPE_F32,
3173               i->getDef(0),
3174               bld.mkSymbol(
3175                     FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
3176                     TYPE_U32, prog->driver->io.sampleInfoBase),
3177               calculateSampleOffset(i->getSrc(0)));
3178 
3179    bld.getBB()->remove(i);
3180 }
3181 
3182 // Generate a binary predicate if an instruction is predicated by
3183 // e.g. an f32 value.
3184 void
checkPredicate(Instruction * insn)3185 NVC0LoweringPass::checkPredicate(Instruction *insn)
3186 {
3187    Value *pred = insn->getPredicate();
3188    Value *pdst;
3189 
3190    if (!pred || pred->reg.file == FILE_PREDICATE)
3191       return;
3192    pdst = new_LValue(func, FILE_PREDICATE);
3193 
3194    // CAUTION: don't use pdst->getInsn, the definition might not be unique,
3195    //  delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass
3196 
3197    bld.mkCmp(OP_SET, CC_NEU, insn->dType, pdst, insn->dType, bld.mkImm(0), pred);
3198 
3199    insn->setPredicate(insn->cc, pdst);
3200 }
3201 
3202 //
3203 // - add quadop dance for texturing
3204 // - put FP outputs in GPRs
3205 // - convert instruction sequences
3206 //
3207 bool
visit(Instruction * i)3208 NVC0LoweringPass::visit(Instruction *i)
3209 {
3210    bool ret = true;
3211    bld.setPosition(i, false);
3212 
3213    if (i->cc != CC_ALWAYS)
3214       checkPredicate(i);
3215 
3216    switch (i->op) {
3217    case OP_TEX:
3218    case OP_TXB:
3219    case OP_TXL:
3220    case OP_TXF:
3221    case OP_TXG:
3222       return handleTEX(i->asTex());
3223    case OP_TXD:
3224       return handleTXD(i->asTex());
3225    case OP_TXLQ:
3226       return handleTXLQ(i->asTex());
3227    case OP_TXQ:
3228      return handleTXQ(i->asTex());
3229    case OP_EX2:
3230       bld.mkOp1(OP_PREEX2, TYPE_F32, i->getDef(0), i->getSrc(0));
3231       i->setSrc(0, i->getDef(0));
3232       break;
3233    case OP_POW:
3234       return handlePOW(i);
3235    case OP_DIV:
3236       return handleDIV(i);
3237    case OP_MOD:
3238       return handleMOD(i);
3239    case OP_SQRT:
3240       return handleSQRT(i);
3241    case OP_EXPORT:
3242       ret = handleEXPORT(i);
3243       break;
3244    case OP_EMIT:
3245    case OP_RESTART:
3246       return handleOUT(i);
3247    case OP_RDSV:
3248       return handleRDSV(i);
3249    case OP_WRSV:
3250       return handleWRSV(i);
3251    case OP_STORE:
3252    case OP_LOAD:
3253       handleLDST(i);
3254       break;
3255    case OP_ATOM:
3256    {
3257       const bool cctl = i->src(0).getFile() == FILE_MEMORY_BUFFER;
3258       handleATOM(i);
3259       handleCasExch(i, cctl);
3260    }
3261       break;
3262    case OP_SULDB:
3263    case OP_SULDP:
3264    case OP_SUSTB:
3265    case OP_SUSTP:
3266    case OP_SUREDB:
3267    case OP_SUREDP:
3268       if (targ->getChipset() >= NVISA_GM107_CHIPSET)
3269          handleSurfaceOpGM107(i->asTex());
3270       else if (targ->getChipset() >= NVISA_GK104_CHIPSET)
3271          handleSurfaceOpNVE4(i->asTex());
3272       else
3273          handleSurfaceOpNVC0(i->asTex());
3274       break;
3275    case OP_SUQ:
3276       handleSUQ(i->asTex());
3277       break;
3278    case OP_BUFQ:
3279       handleBUFQ(i);
3280       break;
3281    case OP_PIXLD:
3282       handlePIXLD(i);
3283       break;
3284    default:
3285       break;
3286    }
3287 
3288    /* Kepler+ has a special opcode to compute a new base address to be used
3289     * for indirect loads.
3290     *
3291     * Maxwell+ has an additional similar requirement for indirect
3292     * interpolation ops in frag shaders.
3293     */
3294    bool doAfetch = false;
3295    if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
3296        !i->perPatch &&
3297        (i->op == OP_VFETCH || i->op == OP_EXPORT) &&
3298        i->src(0).isIndirect(0)) {
3299       doAfetch = true;
3300    }
3301    if (targ->getChipset() >= NVISA_GM107_CHIPSET &&
3302        (i->op == OP_LINTERP || i->op == OP_PINTERP) &&
3303        i->src(0).isIndirect(0)) {
3304       doAfetch = true;
3305    }
3306 
3307    if (doAfetch) {
3308       Value *addr = cloneShallow(func, i->getSrc(0));
3309       Instruction *afetch = bld.mkOp1(OP_AFETCH, TYPE_U32, bld.getSSA(),
3310                                       i->getSrc(0));
3311       afetch->setIndirect(0, 0, i->getIndirect(0, 0));
3312       addr->reg.data.offset = 0;
3313       i->setSrc(0, addr);
3314       i->setIndirect(0, 0, afetch->getDef(0));
3315    }
3316 
3317    return ret;
3318 }
3319 
3320 bool
runLegalizePass(Program * prog,CGStage stage) const3321 TargetNVC0::runLegalizePass(Program *prog, CGStage stage) const
3322 {
3323    if (stage == CG_STAGE_PRE_SSA) {
3324       NVC0LoweringPass pass(prog);
3325       return pass.run(prog, false, true);
3326    } else
3327    if (stage == CG_STAGE_POST_RA) {
3328       NVC0LegalizePostRA pass(prog);
3329       return pass.run(prog, false, true);
3330    } else
3331    if (stage == CG_STAGE_SSA) {
3332       NVC0LegalizeSSA pass;
3333       return pass.run(prog, false, true);
3334    }
3335    return false;
3336 }
3337 
3338 } // namespace nv50_ir
3339