2 * Copyright 2011 Christoph Bumiller
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include "codegen/nv50_ir.h"
24 #include "codegen/nv50_ir_build_util.h"
26 #include "codegen/nv50_ir_target_nvc0.h"
38 #define QUADOP(q, r, s, t) \
39 ((QOP_##q << 6) | (QOP_##r << 4) | \
40 (QOP_##s << 2) | (QOP_##t << 0))
42 class NVC0LegalizeSSA : public Pass
45 virtual bool visit(BasicBlock *);
46 virtual bool visit(Function *);
48 // we want to insert calls to the builtin library only after optimization
49 void handleDIV(Instruction *); // integer division, modulus
50 void handleRCPRSQ(Instruction *); // double precision float recip/rsqrt
57 NVC0LegalizeSSA::handleDIV(Instruction *i)
59 FlowInstruction *call;
63 bld.setPosition(i, false);
64 def[0] = bld.mkMovToReg(0, i->getSrc(0))->getDef(0);
65 def[1] = bld.mkMovToReg(1, i->getSrc(1))->getDef(0);
67 case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break;
68 case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break;
72 call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
73 bld.mkMov(i->getDef(0), def[(i->op == OP_DIV) ? 0 : 1]);
74 bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2);
75 bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0);
78 call->absolute = call->builtin = 1;
79 call->target.builtin = builtin;
80 delete_Instruction(prog, i);
84 NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
90 NVC0LegalizeSSA::visit(Function *fn)
92 bld.setProgram(fn->getProgram());
97 NVC0LegalizeSSA::visit(BasicBlock *bb)
100 for (Instruction *i = bb->getEntry(); i; i = next) {
102 if (i->dType == TYPE_F32)
111 if (i->dType == TYPE_F64)
121 class NVC0LegalizePostRA : public Pass
124 NVC0LegalizePostRA(const Program *);
127 virtual bool visit(Function *);
128 virtual bool visit(BasicBlock *);
130 void replaceZero(Instruction *);
131 bool tryReplaceContWithBra(BasicBlock *);
132 void propagateJoin(BasicBlock *);
136 TexUse(Instruction *use, const Instruction *tex)
137 : insn(use), tex(tex), level(-1) { }
139 const Instruction *tex; // or split / mov
145 Limits(int min, int max) : min(min), max(max) { }
148 bool insertTextureBarriers(Function *);
149 inline bool insnDominatedBy(const Instruction *, const Instruction *) const;
150 void findFirstUses(const Instruction *tex, const Instruction *def,
152 void findOverwritingDefs(const Instruction *tex, Instruction *insn,
153 const BasicBlock *term,
155 void addTexUse(std::list<TexUse>&, Instruction *, const Instruction *);
156 const Instruction *recurseDef(const Instruction *);
161 const bool needTexBar;
164 NVC0LegalizePostRA::NVC0LegalizePostRA(const Program *prog)
167 needTexBar(prog->getTarget()->getChipset() >= 0xe0)
172 NVC0LegalizePostRA::insnDominatedBy(const Instruction *later,
173 const Instruction *early) const
175 if (early->bb == later->bb)
176 return early->serial < later->serial;
177 return later->bb->dominatedBy(early->bb);
181 NVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses,
182 Instruction *usei, const Instruction *insn)
185 for (std::list<TexUse>::iterator it = uses.begin();
187 if (insnDominatedBy(usei, it->insn)) {
191 if (insnDominatedBy(it->insn, usei))
197 uses.push_back(TexUse(usei, insn));
201 NVC0LegalizePostRA::findOverwritingDefs(const Instruction *texi,
203 const BasicBlock *term,
204 std::list<TexUse> &uses)
206 while (insn->op == OP_MOV && insn->getDef(0)->equals(insn->getSrc(0)))
207 insn = insn->getSrc(0)->getUniqueInsn();
209 if (!insn || !insn->bb->reachableBy(texi->bb, term))
213 /* Values not connected to the tex's definition through any of these should
214 * not be conflicting.
221 for (int s = 0; insn->srcExists(s); ++s)
222 findOverwritingDefs(texi, insn->getSrc(s)->getUniqueInsn(), term,
226 // if (!isTextureOp(insn->op)) // TODO: are TEXes always ordered ?
227 addTexUse(uses, insn, texi);
233 NVC0LegalizePostRA::findFirstUses(const Instruction *texi,
234 const Instruction *insn,
235 std::list<TexUse> &uses)
237 for (int d = 0; insn->defExists(d); ++d) {
238 Value *v = insn->getDef(d);
239 for (Value::UseIterator u = v->uses.begin(); u != v->uses.end(); ++u) {
240 Instruction *usei = (*u)->getInsn();
242 if (usei->op == OP_PHI || usei->op == OP_UNION) {
243 // need a barrier before WAW cases
244 for (int s = 0; usei->srcExists(s); ++s) {
245 Instruction *defi = usei->getSrc(s)->getUniqueInsn();
246 if (defi && &usei->src(s) != *u)
247 findOverwritingDefs(texi, defi, usei->bb, uses);
251 if (usei->op == OP_SPLIT ||
252 usei->op == OP_MERGE ||
253 usei->op == OP_PHI ||
254 usei->op == OP_UNION) {
255 // these uses don't manifest in the machine code
256 findFirstUses(texi, usei, uses);
258 if (usei->op == OP_MOV && usei->getDef(0)->equals(usei->getSrc(0)) &&
259 usei->subOp != NV50_IR_SUBOP_MOV_FINAL) {
260 findFirstUses(texi, usei, uses);
262 addTexUse(uses, usei, insn);
269 // This pass is a bit long and ugly and can probably be optimized.
271 // 1. obtain a list of TEXes and their outputs' first use(s)
272 // 2. calculate the barrier level of each first use (minimal number of TEXes,
273 // over all paths, between the TEX and the use in question)
274 // 3. for each barrier, if all paths from the source TEX to that barrier
275 // contain a barrier of lesser level, it can be culled
277 NVC0LegalizePostRA::insertTextureBarriers(Function *fn)
279 std::list<TexUse> *uses;
280 std::vector<Instruction *> texes;
281 std::vector<int> bbFirstTex;
282 std::vector<int> bbFirstUse;
283 std::vector<int> texCounts;
284 std::vector<TexUse> useVec;
287 fn->orderInstructions(insns);
289 texCounts.resize(fn->allBBlocks.getSize(), 0);
290 bbFirstTex.resize(fn->allBBlocks.getSize(), insns.getSize());
291 bbFirstUse.resize(fn->allBBlocks.getSize(), insns.getSize());
293 // tag BB CFG nodes by their id for later
294 for (ArrayList::Iterator i = fn->allBBlocks.iterator(); !i.end(); i.next()) {
295 BasicBlock *bb = reinterpret_cast<BasicBlock *>(i.get());
297 bb->cfg.tag = bb->getId();
300 // gather the first uses for each TEX
301 for (int i = 0; i < insns.getSize(); ++i) {
302 Instruction *tex = reinterpret_cast<Instruction *>(insns.get(i));
303 if (isTextureOp(tex->op)) {
304 texes.push_back(tex);
305 if (!texCounts.at(tex->bb->getId()))
306 bbFirstTex[tex->bb->getId()] = texes.size() - 1;
307 texCounts[tex->bb->getId()]++;
313 uses = new std::list<TexUse>[texes.size()];
316 for (size_t i = 0; i < texes.size(); ++i)
317 findFirstUses(texes[i], texes[i], uses[i]);
319 // determine the barrier level at each use
320 for (size_t i = 0; i < texes.size(); ++i) {
321 for (std::list<TexUse>::iterator u = uses[i].begin(); u != uses[i].end();
323 BasicBlock *tb = texes[i]->bb;
324 BasicBlock *ub = u->insn->bb;
327 for (size_t j = i + 1; j < texes.size() &&
328 texes[j]->bb == tb && texes[j]->serial < u->insn->serial;
332 u->level = fn->cfg.findLightestPathWeight(&tb->cfg,
333 &ub->cfg, texCounts);
335 WARN("Failed to find path TEX -> TEXBAR\n");
339 // this counted all TEXes in the origin block, correct that
340 u->level -= i - bbFirstTex.at(tb->getId()) + 1 /* this TEX */;
341 // and did not count the TEXes in the destination block, add those
342 for (size_t j = bbFirstTex.at(ub->getId()); j < texes.size() &&
343 texes[j]->bb == ub && texes[j]->serial < u->insn->serial;
347 assert(u->level >= 0);
348 useVec.push_back(*u);
354 // insert the barriers
355 for (size_t i = 0; i < useVec.size(); ++i) {
356 Instruction *prev = useVec[i].insn->prev;
357 if (useVec[i].level < 0)
359 if (prev && prev->op == OP_TEXBAR) {
360 if (prev->subOp > useVec[i].level)
361 prev->subOp = useVec[i].level;
362 prev->setSrc(prev->srcCount(), useVec[i].tex->getDef(0));
364 Instruction *bar = new_Instruction(func, OP_TEXBAR, TYPE_NONE);
366 bar->subOp = useVec[i].level;
367 // make use explicit to ease latency calculation
368 bar->setSrc(bar->srcCount(), useVec[i].tex->getDef(0));
369 useVec[i].insn->bb->insertBefore(useVec[i].insn, bar);
373 if (fn->getProgram()->optLevel < 3) {
379 std::vector<Limits> limitT, limitB, limitS; // entry, exit, single
381 limitT.resize(fn->allBBlocks.getSize(), Limits(0, 0));
382 limitB.resize(fn->allBBlocks.getSize(), Limits(0, 0));
383 limitS.resize(fn->allBBlocks.getSize());
385 // cull unneeded barriers (should do that earlier, but for simplicity)
386 IteratorRef bi = fn->cfg.iteratorCFG();
387 // first calculate min/max outstanding TEXes for each BB
388 for (bi->reset(); !bi->end(); bi->next()) {
389 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
390 BasicBlock *bb = BasicBlock::get(n);
392 int max = std::numeric_limits<int>::max();
393 for (Instruction *i = bb->getFirst(); i; i = i->next) {
394 if (isTextureOp(i->op)) {
396 if (max < std::numeric_limits<int>::max())
399 if (i->op == OP_TEXBAR) {
400 min = MIN2(min, i->subOp);
401 max = MIN2(max, i->subOp);
404 // limits when looking at an isolated block
405 limitS[bb->getId()].min = min;
406 limitS[bb->getId()].max = max;
408 // propagate the min/max values
409 for (unsigned int l = 0; l <= fn->loopNestingBound; ++l) {
410 for (bi->reset(); !bi->end(); bi->next()) {
411 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
412 BasicBlock *bb = BasicBlock::get(n);
413 const int bbId = bb->getId();
414 for (Graph::EdgeIterator ei = n->incident(); !ei.end(); ei.next()) {
415 BasicBlock *in = BasicBlock::get(ei.getNode());
416 const int inId = in->getId();
417 limitT[bbId].min = MAX2(limitT[bbId].min, limitB[inId].min);
418 limitT[bbId].max = MAX2(limitT[bbId].max, limitB[inId].max);
420 // I just hope this is correct ...
421 if (limitS[bbId].max == std::numeric_limits<int>::max()) {
423 limitB[bbId].min = limitT[bbId].min + limitS[bbId].min;
424 limitB[bbId].max = limitT[bbId].max + limitS[bbId].min;
426 // block contained a barrier
427 limitB[bbId].min = MIN2(limitS[bbId].max,
428 limitT[bbId].min + limitS[bbId].min);
429 limitB[bbId].max = MIN2(limitS[bbId].max,
430 limitT[bbId].max + limitS[bbId].min);
434 // finally delete unnecessary barriers
435 for (bi->reset(); !bi->end(); bi->next()) {
436 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
437 BasicBlock *bb = BasicBlock::get(n);
438 Instruction *prev = NULL;
440 int max = limitT[bb->getId()].max;
441 for (Instruction *i = bb->getFirst(); i; i = next) {
443 if (i->op == OP_TEXBAR) {
444 if (i->subOp >= max) {
445 delete_Instruction(prog, i);
449 if (prev && prev->op == OP_TEXBAR && prev->subOp >= max) {
450 delete_Instruction(prog, prev);
455 if (isTextureOp(i->op)) {
458 if (i && !i->isNop())
468 NVC0LegalizePostRA::visit(Function *fn)
471 insertTextureBarriers(fn);
473 rZero = new_LValue(fn, FILE_GPR);
474 carry = new_LValue(fn, FILE_FLAGS);
476 rZero->reg.data.id = prog->getTarget()->getFileSize(FILE_GPR);
477 carry->reg.data.id = 0;
483 NVC0LegalizePostRA::replaceZero(Instruction *i)
485 for (int s = 0; i->srcExists(s); ++s) {
486 if (s == 2 && i->op == OP_SUCLAMP)
488 ImmediateValue *imm = i->getSrc(s)->asImm();
489 if (imm && imm->reg.data.u64 == 0)
494 // replace CONT with BRA for single unconditional continue
496 NVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock *bb)
498 if (bb->cfg.incidentCount() != 2 || bb->getEntry()->op != OP_PRECONT)
500 Graph::EdgeIterator ei = bb->cfg.incident();
501 if (ei.getType() != Graph::Edge::BACK)
503 if (ei.getType() != Graph::Edge::BACK)
505 BasicBlock *contBB = BasicBlock::get(ei.getNode());
507 if (!contBB->getExit() || contBB->getExit()->op != OP_CONT ||
508 contBB->getExit()->getPredicate())
510 contBB->getExit()->op = OP_BRA;
511 bb->remove(bb->getEntry()); // delete PRECONT
514 assert(ei.end() || ei.getType() != Graph::Edge::BACK);
518 // replace branches to join blocks with join ops
520 NVC0LegalizePostRA::propagateJoin(BasicBlock *bb)
522 if (bb->getEntry()->op != OP_JOIN || bb->getEntry()->asFlow()->limit)
524 for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
525 BasicBlock *in = BasicBlock::get(ei.getNode());
526 Instruction *exit = in->getExit();
528 in->insertTail(new FlowInstruction(func, OP_JOIN, bb));
529 // there should always be a terminator instruction
530 WARN("inserted missing terminator in BB:%i\n", in->getId());
532 if (exit->op == OP_BRA) {
534 exit->asFlow()->limit = 1; // must-not-propagate marker
537 bb->remove(bb->getEntry());
541 NVC0LegalizePostRA::visit(BasicBlock *bb)
543 Instruction *i, *next;
545 // remove pseudo operations and non-fixed no-ops, split 64 bit operations
546 for (i = bb->getFirst(); i; i = next) {
548 if (i->op == OP_EMIT || i->op == OP_RESTART) {
549 if (!i->getDef(0)->refCount())
551 if (i->src(0).getFile() == FILE_IMMEDIATE)
552 i->setSrc(0, rZero); // initial value must be 0
557 // TODO: Move this to before register allocation for operations that
558 // need the $c register !
559 if (typeSizeof(i->dType) == 8) {
561 hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry);
566 if (i->op != OP_MOV && i->op != OP_PFETCH)
573 if (!tryReplaceContWithBra(bb))
579 class NVC0LoweringPass : public Pass
582 NVC0LoweringPass(Program *);
585 virtual bool visit(Function *);
586 virtual bool visit(BasicBlock *);
587 virtual bool visit(Instruction *);
589 bool handleRDSV(Instruction *);
590 bool handleWRSV(Instruction *);
591 bool handleEXPORT(Instruction *);
592 bool handleOUT(Instruction *);
593 bool handleDIV(Instruction *);
594 bool handleMOD(Instruction *);
595 bool handleSQRT(Instruction *);
596 bool handlePOW(Instruction *);
597 bool handleTEX(TexInstruction *);
598 bool handleTXD(TexInstruction *);
599 bool handleTXQ(TexInstruction *);
600 bool handleManualTXD(TexInstruction *);
601 bool handleTXLQ(TexInstruction *);
602 bool handleATOM(Instruction *);
603 bool handleCasExch(Instruction *, bool needCctl);
604 void handleSurfaceOpNVE4(TexInstruction *);
606 void checkPredicate(Instruction *);
608 void readTessCoord(LValue *dst, int c);
610 Value *loadResInfo32(Value *ptr, uint32_t off);
611 Value *loadMsInfo32(Value *ptr, uint32_t off);
612 Value *loadTexHandle(Value *ptr, unsigned int slot);
614 void adjustCoordinatesMS(TexInstruction *);
615 void processSurfaceCoordsNVE4(TexInstruction *);
618 const Target *const targ;
623 LValue *gpEmitAddress;
626 NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
628 bld.setProgram(prog);
633 NVC0LoweringPass::visit(Function *fn)
635 if (prog->getType() == Program::TYPE_GEOMETRY) {
636 assert(!strncmp(fn->getName(), "MAIN", 4));
637 // TODO: when we generate actual functions pass this value along somehow
638 bld.setPosition(BasicBlock::get(fn->cfg.getRoot()), false);
639 gpEmitAddress = bld.loadImm(NULL, 0)->asLValue();
641 bld.setPosition(BasicBlock::get(fn->cfgExit)->getExit(), false);
642 bld.mkMovToReg(0, gpEmitAddress);
649 NVC0LoweringPass::visit(BasicBlock *bb)
655 NVC0LoweringPass::loadTexHandle(Value *ptr, unsigned int slot)
657 uint8_t b = prog->driver->io.resInfoCBSlot;
658 uint32_t off = prog->driver->io.texBindBase + slot * 4;
660 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
663 // move array source to first slot, convert to u16, add indirections
665 NVC0LoweringPass::handleTEX(TexInstruction *i)
667 const int dim = i->tex.target.getDim() + i->tex.target.isCube();
668 const int arg = i->tex.target.getArgCount();
669 const int lyr = arg - (i->tex.target.isMS() ? 2 : 1);
670 const int chipset = prog->getTarget()->getChipset();
672 if (chipset >= NVISA_GK104_CHIPSET) {
673 if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
674 WARN("indirect TEX not implemented\n");
676 if (i->tex.r == i->tex.s) {
677 i->tex.r += prog->driver->io.texBindBase / 4;
678 i->tex.s = 0; // only a single cX[] value possible here
680 Value *hnd = bld.getScratch();
681 Value *rHnd = loadTexHandle(NULL, i->tex.r);
682 Value *sHnd = loadTexHandle(NULL, i->tex.s);
684 bld.mkOp3(OP_INSBF, TYPE_U32, hnd, rHnd, bld.mkImm(0x1400), sHnd);
686 i->tex.r = 0; // not used for indirect tex
688 i->setIndirectR(hnd);
690 if (i->tex.target.isArray()) {
691 LValue *layer = new_LValue(func, FILE_GPR);
692 Value *src = i->getSrc(lyr);
693 const int sat = (i->op == OP_TXF) ? 1 : 0;
694 DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
695 bld.mkCvt(OP_CVT, TYPE_U16, layer, sTy, src)->saturate = sat;
696 for (int s = dim; s >= 1; --s)
697 i->setSrc(s, i->getSrc(s - 1));
701 // (nvc0) generate and move the tsc/tic/array source to the front
702 if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
703 LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
705 Value *arrayIndex = i->tex.target.isArray() ? i->getSrc(lyr) : NULL;
706 for (int s = dim; s >= 1; --s)
707 i->setSrc(s, i->getSrc(s - 1));
708 i->setSrc(0, arrayIndex);
710 Value *ticRel = i->getIndirectR();
711 Value *tscRel = i->getIndirectS();
714 int sat = (i->op == OP_TXF) ? 1 : 0;
715 DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
716 bld.mkCvt(OP_CVT, TYPE_U16, src, sTy, arrayIndex)->saturate = sat;
722 i->setSrc(i->tex.rIndirectSrc, NULL);
723 bld.mkOp3(OP_INSBF, TYPE_U32, src, ticRel, bld.mkImm(0x0917), src);
726 i->setSrc(i->tex.sIndirectSrc, NULL);
727 bld.mkOp3(OP_INSBF, TYPE_U32, src, tscRel, bld.mkImm(0x0710), src);
733 // For nvc0, the sample id has to be in the second operand, as the offset
734 // does. Right now we don't know how to pass both in, and this case can't
735 // happen with OpenGL. On nve0, the sample id is part of the texture
736 // coordinate argument.
737 assert(chipset >= NVISA_GK104_CHIPSET ||
738 !i->tex.useOffsets || !i->tex.target.isMS());
740 // offset is between lod and dc
741 if (i->tex.useOffsets) {
743 int s = i->srcCount(0xff, true);
744 if (i->tex.target.isShadow())
746 if (i->srcExists(s)) // move potential predicate out of the way
747 i->moveSources(s, 1);
748 if (i->tex.useOffsets == 4 && i->srcExists(s + 1))
749 i->moveSources(s + 1, 1);
750 if (i->op == OP_TXG) {
751 // Either there is 1 offset, which goes into the 2 low bytes of the
752 // first source, or there are 4 offsets, which go into 2 sources (8
753 // values, 1 byte each).
754 Value *offs[2] = {NULL, NULL};
755 for (n = 0; n < i->tex.useOffsets; n++) {
756 for (c = 0; c < 2; ++c) {
757 if ((n % 2) == 0 && c == 0)
758 offs[n / 2] = i->offset[n][c].get();
760 bld.mkOp3(OP_INSBF, TYPE_U32,
762 i->offset[n][c].get(),
763 bld.mkImm(0x800 | ((n * 16 + c * 8) % 32)),
767 i->setSrc(s, offs[0]);
769 i->setSrc(s + 1, offs[1]);
772 assert(i->tex.useOffsets == 1);
773 for (c = 0; c < 3; ++c) {
775 assert(i->offset[0][c].getImmediate(val));
776 imm |= (val.reg.data.u32 & 0xf) << (c * 4);
778 i->setSrc(s, bld.loadImm(NULL, imm));
782 if (chipset >= NVISA_GK104_CHIPSET) {
784 // If TEX requires more than 4 sources, the 2nd register tuple must be
785 // aligned to 4, even if it consists of just a single 4-byte register.
787 // XXX HACK: We insert 0 sources to avoid the 5 or 6 regs case.
789 int s = i->srcCount(0xff, true);
790 if (s > 4 && s < 7) {
791 if (i->srcExists(s)) // move potential predicate out of the way
792 i->moveSources(s, 7 - s);
794 i->setSrc(s++, bld.loadImm(NULL, 0));
802 NVC0LoweringPass::handleManualTXD(TexInstruction *i)
804 static const uint8_t qOps[4][2] =
806 { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(MOV2, MOV2, ADD, ADD) }, // l0
807 { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD, ADD) }, // l1
808 { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2
809 { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3
814 Value *zero = bld.loadImm(bld.getSSA(), 0);
816 const int dim = i->tex.target.getDim();
818 i->op = OP_TEX; // no need to clone dPdx/dPdy later
820 for (c = 0; c < dim; ++c)
821 crd[c] = bld.getScratch();
823 bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
824 for (l = 0; l < 4; ++l) {
825 // mov coordinates from lane l to all lanes
826 for (c = 0; c < dim; ++c)
827 bld.mkQuadop(0x00, crd[c], l, i->getSrc(c), zero);
828 // add dPdx from lane l to lanes dx
829 for (c = 0; c < dim; ++c)
830 bld.mkQuadop(qOps[l][0], crd[c], l, i->dPdx[c].get(), crd[c]);
831 // add dPdy from lane l to lanes dy
832 for (c = 0; c < dim; ++c)
833 bld.mkQuadop(qOps[l][1], crd[c], l, i->dPdy[c].get(), crd[c]);
835 bld.insert(tex = cloneForward(func, i));
836 for (c = 0; c < dim; ++c)
837 tex->setSrc(c, crd[c]);
839 for (c = 0; i->defExists(c); ++c) {
841 def[c][l] = bld.getSSA();
842 mov = bld.mkMov(def[c][l], tex->getDef(c));
847 bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
849 for (c = 0; i->defExists(c); ++c) {
850 Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c));
851 for (l = 0; l < 4; ++l)
852 u->setSrc(l, def[c][l]);
860 NVC0LoweringPass::handleTXD(TexInstruction *txd)
862 int dim = txd->tex.target.getDim();
863 int arg = txd->tex.target.getArgCount();
866 while (txd->srcExists(arg))
869 txd->tex.derivAll = true;
871 txd->tex.target.isCube() ||
873 txd->tex.target.isShadow())
874 return handleManualTXD(txd);
876 for (int c = 0; c < dim; ++c) {
877 txd->setSrc(arg + c * 2 + 0, txd->dPdx[c]);
878 txd->setSrc(arg + c * 2 + 1, txd->dPdy[c]);
879 txd->dPdx[c].set(NULL);
880 txd->dPdy[c].set(NULL);
886 NVC0LoweringPass::handleTXQ(TexInstruction *txq)
888 // TODO: indirect resource/sampler index
893 NVC0LoweringPass::handleTXLQ(TexInstruction *i)
895 /* The outputs are inverted compared to what the TGSI instruction
896 * expects. Take that into account in the mask.
898 assert((i->tex.mask & ~3) == 0);
899 if (i->tex.mask == 1)
901 else if (i->tex.mask == 2)
904 bld.setPosition(i, true);
906 /* The returned values are not quite what we want:
907 * (a) convert from s16/u16 to f32
908 * (b) multiply by 1/256
910 for (int def = 0; def < 2; ++def) {
911 if (!i->defExists(def))
913 enum DataType type = TYPE_S16;
914 if (i->tex.mask == 2 || def > 0)
916 bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(def), type, i->getDef(def));
917 bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(def),
918 i->getDef(def), bld.loadImm(NULL, 1.0f / 256));
920 if (i->tex.mask == 3) {
921 LValue *t = new_LValue(func, FILE_GPR);
922 bld.mkMov(t, i->getDef(0));
923 bld.mkMov(i->getDef(0), i->getDef(1));
924 bld.mkMov(i->getDef(1), t);
931 NVC0LoweringPass::handleATOM(Instruction *atom)
935 switch (atom->src(0).getFile()) {
936 case FILE_MEMORY_LOCAL:
939 case FILE_MEMORY_SHARED:
943 assert(atom->src(0).getFile() == FILE_MEMORY_GLOBAL);
947 bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(), bld.mkSysVal(sv, 0));
948 Value *ptr = atom->getIndirect(0, 0);
950 atom->setSrc(0, cloneShallow(func, atom->getSrc(0)));
951 atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
953 base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr);
954 atom->setIndirect(0, 0, base);
960 NVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl)
962 if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS &&
963 cas->subOp != NV50_IR_SUBOP_ATOM_EXCH)
965 bld.setPosition(cas, true);
968 Instruction *cctl = bld.mkOp1(OP_CCTL, TYPE_NONE, NULL, cas->getSrc(0));
969 cctl->setIndirect(0, 0, cas->getIndirect(0, 0));
971 cctl->subOp = NV50_IR_SUBOP_CCTL_IV;
972 if (cas->isPredicated())
973 cctl->setPredicate(cas->cc, cas->getPredicate());
976 if (cas->defExists(0) && cas->subOp == NV50_IR_SUBOP_ATOM_CAS) {
977 // CAS is crazy. It's 2nd source is a double reg, and the 3rd source
978 // should be set to the high part of the double reg or bad things will
979 // happen elsewhere in the universe.
980 // Also, it sometimes returns the new value instead of the old one
981 // under mysterious circumstances.
982 Value *dreg = bld.getSSA(8);
983 bld.setPosition(cas, false);
984 bld.mkOp2(OP_MERGE, TYPE_U64, dreg, cas->getSrc(1), cas->getSrc(2));
985 cas->setSrc(1, dreg);
992 NVC0LoweringPass::loadResInfo32(Value *ptr, uint32_t off)
994 uint8_t b = prog->driver->io.resInfoCBSlot;
995 off += prog->driver->io.suInfoBase;
997 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1001 NVC0LoweringPass::loadMsInfo32(Value *ptr, uint32_t off)
1003 uint8_t b = prog->driver->io.msInfoCBSlot;
1004 off += prog->driver->io.msInfoBase;
1006 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1009 /* On nvc0, surface info is obtained via the surface binding points passed
1010 * to the SULD/SUST instructions.
1011 * On nve4, surface info is stored in c[] and is used by various special
1012 * instructions, e.g. for clamping coordiantes or generating an address.
1013 * They couldn't just have added an equivalent to TIC now, couldn't they ?
1015 #define NVE4_SU_INFO_ADDR 0x00
1016 #define NVE4_SU_INFO_FMT 0x04
1017 #define NVE4_SU_INFO_DIM_X 0x08
1018 #define NVE4_SU_INFO_PITCH 0x0c
1019 #define NVE4_SU_INFO_DIM_Y 0x10
1020 #define NVE4_SU_INFO_ARRAY 0x14
1021 #define NVE4_SU_INFO_DIM_Z 0x18
1022 #define NVE4_SU_INFO_UNK1C 0x1c
1023 #define NVE4_SU_INFO_WIDTH 0x20
1024 #define NVE4_SU_INFO_HEIGHT 0x24
1025 #define NVE4_SU_INFO_DEPTH 0x28
1026 #define NVE4_SU_INFO_TARGET 0x2c
1027 #define NVE4_SU_INFO_CALL 0x30
1028 #define NVE4_SU_INFO_RAW_X 0x34
1029 #define NVE4_SU_INFO_MS_X 0x38
1030 #define NVE4_SU_INFO_MS_Y 0x3c
1032 #define NVE4_SU_INFO__STRIDE 0x40
1034 #define NVE4_SU_INFO_DIM(i) (0x08 + (i) * 8)
1035 #define NVE4_SU_INFO_SIZE(i) (0x20 + (i) * 4)
1036 #define NVE4_SU_INFO_MS(i) (0x38 + (i) * 4)
1038 static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
1040 switch (su->tex.target.getEnum()) {
1041 case TEX_TARGET_BUFFER: return NV50_IR_SUBOP_SUCLAMP_PL(0, 1);
1042 case TEX_TARGET_RECT: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1043 case TEX_TARGET_1D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1044 case TEX_TARGET_1D_ARRAY: return (c == 1) ?
1045 NV50_IR_SUBOP_SUCLAMP_PL(0, 2) :
1046 NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1047 case TEX_TARGET_2D: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1048 case TEX_TARGET_2D_MS: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1049 case TEX_TARGET_2D_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1050 case TEX_TARGET_2D_MS_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1051 case TEX_TARGET_3D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1052 case TEX_TARGET_CUBE: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1053 case TEX_TARGET_CUBE_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1061 NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
1063 const uint16_t base = tex->tex.r * NVE4_SU_INFO__STRIDE;
1064 const int arg = tex->tex.target.getArgCount();
1066 if (tex->tex.target == TEX_TARGET_2D_MS)
1067 tex->tex.target = TEX_TARGET_2D;
1069 if (tex->tex.target == TEX_TARGET_2D_MS_ARRAY)
1070 tex->tex.target = TEX_TARGET_2D_ARRAY;
1074 Value *x = tex->getSrc(0);
1075 Value *y = tex->getSrc(1);
1076 Value *s = tex->getSrc(arg - 1);
1078 Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
1080 Value *ms_x = loadResInfo32(NULL, base + NVE4_SU_INFO_MS(0));
1081 Value *ms_y = loadResInfo32(NULL, base + NVE4_SU_INFO_MS(1));
1083 bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
1084 bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);
1086 s = bld.mkOp2v(OP_AND, TYPE_U32, ts, s, bld.loadImm(NULL, 0x7));
1087 s = bld.mkOp2v(OP_SHL, TYPE_U32, ts, ts, bld.mkImm(3));
1089 Value *dx = loadMsInfo32(ts, 0x0);
1090 Value *dy = loadMsInfo32(ts, 0x4);
1092 bld.mkOp2(OP_ADD, TYPE_U32, tx, tx, dx);
1093 bld.mkOp2(OP_ADD, TYPE_U32, ty, ty, dy);
1097 tex->moveSources(arg, -1);
1100 // Sets 64-bit "generic address", predicate and format sources for SULD/SUST.
1101 // They're computed from the coordinates using the surface info in c[] space.
1103 NVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction *su)
1106 const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP;
1108 su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB;
1109 const int idx = su->tex.r;
1110 const int dim = su->tex.target.getDim();
1111 const int arg = dim + (su->tex.target.isArray() ? 1 : 0);
1112 const uint16_t base = idx * NVE4_SU_INFO__STRIDE;
1114 Value *zero = bld.mkImm(0);
1118 Value *bf, *eau, *off;
1121 off = bld.getScratch(4);
1122 bf = bld.getScratch(4);
1123 addr = bld.getSSA(8);
1124 pred = bld.getScratch(1, FILE_PREDICATE);
1126 bld.setPosition(su, false);
1128 adjustCoordinatesMS(su);
1130 // calculate clamped coordinates
1131 for (c = 0; c < arg; ++c) {
1132 src[c] = bld.getScratch();
1134 v = loadResInfo32(NULL, base + NVE4_SU_INFO_RAW_X);
1136 v = loadResInfo32(NULL, base + NVE4_SU_INFO_DIM(c));
1137 bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero)
1138 ->subOp = getSuClampSubOp(su, c);
1143 // set predicate output
1144 if (su->tex.target == TEX_TARGET_BUFFER) {
1145 src[0]->getInsn()->setFlagsDef(1, pred);
1147 if (su->tex.target.isArray()) {
1148 p1 = bld.getSSA(1, FILE_PREDICATE);
1149 src[dim]->getInsn()->setFlagsDef(1, p1);
1152 // calculate pixel offset
1154 if (su->tex.target != TEX_TARGET_BUFFER)
1155 bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff));
1158 v = loadResInfo32(NULL, base + NVE4_SU_INFO_UNK1C);
1159 bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1])
1160 ->subOp = NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1162 v = loadResInfo32(NULL, base + NVE4_SU_INFO_PITCH);
1163 bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0])
1164 ->subOp = NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
1167 v = loadResInfo32(NULL, base + NVE4_SU_INFO_PITCH);
1168 bld.mkOp3(OP_MADSP, TYPE_U32, off, src[1], v, src[0])
1169 ->subOp = su->tex.target.isArray() ?
1170 NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1173 // calculate effective address part 1
1174 if (su->tex.target == TEX_TARGET_BUFFER) {
1178 v = loadResInfo32(NULL, base + NVE4_SU_INFO_FMT);
1179 bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero)
1180 ->subOp = NV50_IR_SUBOP_V1(7,6,8|2);
1194 if (!su->tex.target.isArray()) {
1195 z = loadResInfo32(NULL, base + NVE4_SU_INFO_UNK1C);
1196 subOp = NV50_IR_SUBOP_SUBFM_3D;
1200 subOp = NV50_IR_SUBOP_SUBFM_3D;
1204 insn = bld.mkOp3(OP_SUBFM, TYPE_U32, bf, src[0], y, z);
1205 insn->subOp = subOp;
1206 insn->setFlagsDef(1, pred);
1210 v = loadResInfo32(NULL, base + NVE4_SU_INFO_ADDR);
1212 if (su->tex.target == TEX_TARGET_BUFFER) {
1215 eau = bld.mkOp3v(OP_SUEAU, TYPE_U32, bld.getScratch(4), off, bf, v);
1217 // add array layer offset
1218 if (su->tex.target.isArray()) {
1219 v = loadResInfo32(NULL, base + NVE4_SU_INFO_ARRAY);
1221 bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau)
1222 ->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
1224 bld.mkOp3(OP_MADSP, TYPE_U32, eau, v, src[2], eau)
1225 ->subOp = NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32
1226 // combine predicates
1228 bld.mkOp2(OP_OR, TYPE_U8, pred, pred, p1);
1233 if (su->tex.target == TEX_TARGET_BUFFER) {
1237 // bf == g[] address & 0xff
1238 // eau == g[] address >> 8
1239 bld.mkOp3(OP_PERMT, TYPE_U32, bf, lo, bld.loadImm(NULL, 0x6540), eau);
1240 bld.mkOp3(OP_PERMT, TYPE_U32, eau, zero, bld.loadImm(NULL, 0x0007), eau);
1242 if (su->op == OP_SULDP && su->tex.target == TEX_TARGET_BUFFER) {
1243 // Convert from u32 to u8 address format, which is what the library code
1244 // doing SULDP currently uses.
1245 // XXX: can SUEAU do this ?
1246 // XXX: does it matter that we don't mask high bytes in bf ?
1248 bld.mkOp2(OP_SHR, TYPE_U32, off, bf, bld.mkImm(8));
1249 bld.mkOp2(OP_ADD, TYPE_U32, eau, eau, off);
1252 bld.mkOp2(OP_MERGE, TYPE_U64, addr, bf, eau);
1254 if (atom && su->tex.target == TEX_TARGET_BUFFER)
1255 bld.mkOp2(OP_ADD, TYPE_U64, addr, addr, off);
1257 // let's just set it 0 for raw access and hope it works
1259 bld.mkImm(0) : loadResInfo32(NULL, base + NVE4_SU_INFO_FMT);
1261 // get rid of old coordinate sources, make space for fmt info and predicate
1262 su->moveSources(arg, 3 - arg);
1263 // set 64 bit address and 32-bit format sources
1264 su->setSrc(0, addr);
1266 su->setSrc(2, pred);
1270 NVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction *su)
1272 processSurfaceCoordsNVE4(su);
1274 // Who do we hate more ? The person who decided that nvc0's SULD doesn't
1275 // have to support conversion or the person who decided that, in OpenCL,
1276 // you don't have to specify the format here like you do in OpenGL ?
1278 if (su->op == OP_SULDP) {
1279 // We don't patch shaders. Ever.
1280 // You get an indirect call to our library blob here.
1281 // But at least it's uniform.
1282 FlowInstruction *call;
1285 uint16_t base = su->tex.r * NVE4_SU_INFO__STRIDE + NVE4_SU_INFO_CALL;
1287 for (int i = 0; i < 4; ++i)
1288 (r[i] = bld.getScratch(4, FILE_GPR))->reg.data.id = i;
1289 for (int i = 0; i < 3; ++i)
1290 (p[i] = bld.getScratch(1, FILE_PREDICATE))->reg.data.id = i;
1291 (r[4] = bld.getScratch(8, FILE_GPR))->reg.data.id = 4;
1293 bld.mkMov(p[1], bld.mkImm((su->cache == CACHE_CA) ? 1 : 0), TYPE_U8);
1294 bld.mkMov(p[2], bld.mkImm((su->cache == CACHE_CG) ? 1 : 0), TYPE_U8);
1295 bld.mkMov(p[0], su->getSrc(2), TYPE_U8);
1296 bld.mkMov(r[4], su->getSrc(0), TYPE_U64);
1297 bld.mkMov(r[2], su->getSrc(1), TYPE_U32);
1299 call = bld.mkFlow(OP_CALL, NULL, su->cc, su->getPredicate());
1303 call->setSrc(0, bld.mkSymbol(FILE_MEMORY_CONST,
1304 prog->driver->io.resInfoCBSlot, TYPE_U32,
1305 prog->driver->io.suInfoBase + base));
1306 call->setSrc(1, r[2]);
1307 call->setSrc(2, r[4]);
1308 for (int i = 0; i < 3; ++i)
1309 call->setSrc(3 + i, p[i]);
1310 for (int i = 0; i < 4; ++i) {
1311 call->setDef(i, r[i]);
1312 bld.mkMov(su->getDef(i), r[i]);
1314 call->setDef(4, p[1]);
1315 delete_Instruction(bld.getProgram(), su);
1318 if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
1319 // FIXME: for out of bounds access, destination value will be undefined !
1320 Value *pred = su->getSrc(2);
1321 CondCode cc = CC_NOT_P;
1322 if (su->getPredicate()) {
1323 pred = bld.getScratch(1, FILE_PREDICATE);
1325 if (cc == CC_NOT_P) {
1326 bld.mkOp2(OP_OR, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
1328 bld.mkOp2(OP_AND, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
1329 pred->getInsn()->src(1).mod = Modifier(NV50_IR_MOD_NOT);
1332 Instruction *red = bld.mkOp(OP_ATOM, su->dType, su->getDef(0));
1333 red->subOp = su->subOp;
1335 gMemBase = bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0);
1336 red->setSrc(0, gMemBase);
1337 red->setSrc(1, su->getSrc(3));
1338 if (su->subOp == NV50_IR_SUBOP_ATOM_CAS)
1339 red->setSrc(2, su->getSrc(4));
1340 red->setIndirect(0, 0, su->getSrc(0));
1341 red->setPredicate(cc, pred);
1342 delete_Instruction(bld.getProgram(), su);
1343 handleCasExch(red, true);
1345 su->sType = (su->tex.target == TEX_TARGET_BUFFER) ? TYPE_U32 : TYPE_U8;
1350 NVC0LoweringPass::handleWRSV(Instruction *i)
1356 // must replace, $sreg are not writeable
1357 addr = targ->getSVAddress(FILE_SHADER_OUTPUT, i->getSrc(0)->asSym());
1360 sym = bld.mkSymbol(FILE_SHADER_OUTPUT, 0, i->sType, addr);
1362 st = bld.mkStore(OP_EXPORT, i->dType, sym, i->getIndirect(0, 0),
1364 st->perPatch = i->perPatch;
1366 bld.getBB()->remove(i);
1371 NVC0LoweringPass::readTessCoord(LValue *dst, int c)
1373 Value *laneid = bld.getSSA();
1376 bld.mkOp1(OP_RDSV, TYPE_U32, laneid, bld.mkSysVal(SV_LANEID, 0));
1391 bld.mkFetch(x, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f0, NULL, laneid);
1393 bld.mkFetch(y, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f4, NULL, laneid);
1396 bld.mkOp2(OP_ADD, TYPE_F32, dst, x, y);
1397 bld.mkOp2(OP_SUB, TYPE_F32, dst, bld.loadImm(NULL, 1.0f), dst);
1402 NVC0LoweringPass::handleRDSV(Instruction *i)
1404 Symbol *sym = i->getSrc(0)->asSym();
1405 const SVSemantic sv = sym->reg.data.sv.sv;
1408 uint32_t addr = targ->getSVAddress(FILE_SHADER_INPUT, sym);
1410 if (addr >= 0x400) {
1412 if (sym->reg.data.sv.index == 3) {
1413 // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
1415 i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0));
1422 assert(prog->getType() == Program::TYPE_FRAGMENT);
1423 bld.mkInterp(NV50_IR_INTERP_LINEAR, i->getDef(0), addr, NULL);
1427 Value *face = i->getDef(0);
1428 bld.mkInterp(NV50_IR_INTERP_FLAT, face, addr, NULL);
1429 if (i->dType == TYPE_F32) {
1430 bld.mkOp2(OP_AND, TYPE_U32, face, face, bld.mkImm(0x80000000));
1431 bld.mkOp2(OP_XOR, TYPE_U32, face, face, bld.mkImm(0xbf800000));
1436 assert(prog->getType() == Program::TYPE_TESSELLATION_EVAL);
1437 readTessCoord(i->getDef(0)->asLValue(), i->getSrc(0)->reg.data.sv.index);
1442 assert(targ->getChipset() >= NVISA_GK104_CHIPSET); // mov $sreg otherwise
1443 if (sym->reg.data.sv.index == 3) {
1445 i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1));
1448 addr += prog->driver->prop.cp.gridInfoBase;
1449 bld.mkLoad(TYPE_U32, i->getDef(0),
1450 bld.mkSymbol(FILE_MEMORY_CONST, 0, TYPE_U32, addr), NULL);
1452 case SV_SAMPLE_INDEX:
1453 // TODO: Properly pass source as an address in the PIX address space
1454 // (which can be of the form [r0+offset]). But this is currently
1456 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
1457 ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
1459 case SV_SAMPLE_POS: {
1460 Value *off = new_LValue(func, FILE_GPR);
1461 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
1462 ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
1463 bld.mkOp2(OP_SHL, TYPE_U32, off, i->getDef(0), bld.mkImm(3));
1464 bld.mkLoad(TYPE_F32,
1467 FILE_MEMORY_CONST, prog->driver->io.resInfoCBSlot,
1468 TYPE_U32, prog->driver->io.sampleInfoBase +
1469 4 * sym->reg.data.sv.index),
1473 case SV_SAMPLE_MASK:
1474 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
1475 ld->subOp = NV50_IR_SUBOP_PIXLD_COVMASK;
1478 if (prog->getType() == Program::TYPE_TESSELLATION_EVAL)
1479 vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0));
1480 ld = bld.mkFetch(i->getDef(0), i->dType,
1481 FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
1482 ld->perPatch = i->perPatch;
1485 bld.getBB()->remove(i);
1490 NVC0LoweringPass::handleDIV(Instruction *i)
1492 if (!isFloatType(i->dType))
1494 bld.setPosition(i, false);
1495 Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(), i->getSrc(1));
1497 i->setSrc(1, rcp->getDef(0));
1502 NVC0LoweringPass::handleMOD(Instruction *i)
1504 if (i->dType != TYPE_F32)
1506 LValue *value = bld.getScratch();
1507 bld.mkOp1(OP_RCP, TYPE_F32, value, i->getSrc(1));
1508 bld.mkOp2(OP_MUL, TYPE_F32, value, i->getSrc(0), value);
1509 bld.mkOp1(OP_TRUNC, TYPE_F32, value, value);
1510 bld.mkOp2(OP_MUL, TYPE_F32, value, i->getSrc(1), value);
1512 i->setSrc(1, value);
1517 NVC0LoweringPass::handleSQRT(Instruction *i)
1519 Instruction *rsq = bld.mkOp1(OP_RSQ, TYPE_F32,
1520 bld.getSSA(), i->getSrc(0));
1522 i->setSrc(1, rsq->getDef(0));
1528 NVC0LoweringPass::handlePOW(Instruction *i)
1530 LValue *val = bld.getScratch();
1532 bld.mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0));
1533 bld.mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1;
1534 bld.mkOp1(OP_PREEX2, TYPE_F32, val, val);
1544 NVC0LoweringPass::handleEXPORT(Instruction *i)
1546 if (prog->getType() == Program::TYPE_FRAGMENT) {
1547 int id = i->getSrc(0)->reg.data.offset / 4;
1549 if (i->src(0).isIndirect(0)) // TODO, ugly
1552 i->subOp = NV50_IR_SUBOP_MOV_FINAL;
1553 i->src(0).set(i->src(1));
1555 i->setDef(0, new_LValue(func, FILE_GPR));
1556 i->getDef(0)->reg.data.id = id;
1558 prog->maxGPR = MAX2(prog->maxGPR, id);
1560 if (prog->getType() == Program::TYPE_GEOMETRY) {
1561 i->setIndirect(0, 1, gpEmitAddress);
1567 NVC0LoweringPass::handleOUT(Instruction *i)
1569 if (i->op == OP_RESTART && i->prev && i->prev->op == OP_EMIT) {
1570 i->prev->subOp = NV50_IR_SUBOP_EMIT_RESTART;
1571 delete_Instruction(prog, i);
1573 assert(gpEmitAddress);
1574 i->setDef(0, gpEmitAddress);
1575 if (i->srcExists(0))
1576 i->setSrc(1, i->getSrc(0));
1577 i->setSrc(0, gpEmitAddress);
1582 // Generate a binary predicate if an instruction is predicated by
1583 // e.g. an f32 value.
1585 NVC0LoweringPass::checkPredicate(Instruction *insn)
1587 Value *pred = insn->getPredicate();
1590 if (!pred || pred->reg.file == FILE_PREDICATE)
1592 pdst = new_LValue(func, FILE_PREDICATE);
1594 // CAUTION: don't use pdst->getInsn, the definition might not be unique,
1595 // delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass
1597 bld.mkCmp(OP_SET, CC_NEU, insn->dType, pdst, insn->dType, bld.mkImm(0), pred);
1599 insn->setPredicate(insn->cc, pdst);
1603 // - add quadop dance for texturing
1604 // - put FP outputs in GPRs
1605 // - convert instruction sequences
1608 NVC0LoweringPass::visit(Instruction *i)
1610 bld.setPosition(i, false);
1612 if (i->cc != CC_ALWAYS)
1621 return handleTEX(i->asTex());
1623 return handleTXD(i->asTex());
1625 return handleTXLQ(i->asTex());
1627 return handleTXQ(i->asTex());
1629 bld.mkOp1(OP_PREEX2, TYPE_F32, i->getDef(0), i->getSrc(0));
1630 i->setSrc(0, i->getDef(0));
1633 return handlePOW(i);
1635 return handleDIV(i);
1637 return handleMOD(i);
1639 return handleSQRT(i);
1641 return handleEXPORT(i);
1644 return handleOUT(i);
1646 return handleRDSV(i);
1648 return handleWRSV(i);
1650 if (i->src(0).getFile() == FILE_SHADER_INPUT) {
1651 if (prog->getType() == Program::TYPE_COMPUTE) {
1652 i->getSrc(0)->reg.file = FILE_MEMORY_CONST;
1653 i->getSrc(0)->reg.fileIndex = 0;
1655 if (prog->getType() == Program::TYPE_GEOMETRY &&
1656 i->src(0).isIndirect(0)) {
1657 // XXX: this assumes vec4 units
1658 Value *ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
1659 i->getIndirect(0, 0), bld.mkImm(4));
1660 i->setIndirect(0, 0, ptr);
1663 assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP
1669 const bool cctl = i->src(0).getFile() == FILE_MEMORY_GLOBAL;
1671 handleCasExch(i, cctl);
1680 if (targ->getChipset() >= NVISA_GK104_CHIPSET)
1681 handleSurfaceOpNVE4(i->asTex());
1690 TargetNVC0::runLegalizePass(Program *prog, CGStage stage) const
1692 if (stage == CG_STAGE_PRE_SSA) {
1693 NVC0LoweringPass pass(prog);
1694 return pass.run(prog, false, true);
1696 if (stage == CG_STAGE_POST_RA) {
1697 NVC0LegalizePostRA pass(prog);
1698 return pass.run(prog, false, true);
1700 if (stage == CG_STAGE_SSA) {
1701 NVC0LegalizeSSA pass;
1702 return pass.run(prog, false, true);
1707 } // namespace nv50_ir