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"
27 #include "codegen/nv50_ir_lowering_nvc0.h"
39 #define QUADOP(q, r, s, t) \
40 ((QOP_##q << 6) | (QOP_##r << 4) | \
41 (QOP_##s << 2) | (QOP_##t << 0))
44 NVC0LegalizeSSA::handleDIV(Instruction *i)
46 FlowInstruction *call;
50 bld.setPosition(i, false);
51 def[0] = bld.mkMovToReg(0, i->getSrc(0))->getDef(0);
52 def[1] = bld.mkMovToReg(1, i->getSrc(1))->getDef(0);
54 case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break;
55 case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break;
59 call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
60 bld.mkMov(i->getDef(0), def[(i->op == OP_DIV) ? 0 : 1]);
61 bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2);
62 bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0);
65 call->absolute = call->builtin = 1;
66 call->target.builtin = builtin;
67 delete_Instruction(prog, i);
71 NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
73 assert(i->dType == TYPE_F64);
74 // There are instructions that will compute the high 32 bits of the 64-bit
75 // float. We will just stick 0 in the bottom 32 bits.
77 bld.setPosition(i, false);
79 // 1. Take the source and it up.
80 Value *src[2], *dst[2], *def = i->getDef(0);
81 bld.mkSplit(src, 4, i->getSrc(0));
83 // 2. We don't care about the low 32 bits of the destination. Stick a 0 in.
84 dst[0] = bld.loadImm(NULL, 0);
85 dst[1] = bld.getSSA();
87 // 3. The new version of the instruction takes the high 32 bits of the
88 // source and outputs the high 32 bits of the destination.
92 i->subOp = NV50_IR_SUBOP_RCPRSQ_64H;
94 // 4. Recombine the two dst pieces back into the original destination.
95 bld.setPosition(i, true);
96 bld.mkOp2(OP_MERGE, TYPE_U64, def, dst[0], dst[1]);
100 NVC0LegalizeSSA::handleFTZ(Instruction *i)
102 // Only want to flush float inputs
103 assert(i->sType == TYPE_F32);
105 // If we're already flushing denorms (and NaN's) to zero, no need for this.
109 // Only certain classes of operations can flush
110 OpClass cls = prog->getTarget()->getOpClass(i->op);
111 if (cls != OPCLASS_ARITH && cls != OPCLASS_COMPARE &&
112 cls != OPCLASS_CONVERT)
119 NVC0LegalizeSSA::visit(Function *fn)
121 bld.setProgram(fn->getProgram());
126 NVC0LegalizeSSA::visit(BasicBlock *bb)
129 for (Instruction *i = bb->getEntry(); i; i = next) {
131 if (i->sType == TYPE_F32) {
132 if (prog->getType() != Program::TYPE_COMPUTE)
143 if (i->dType == TYPE_F64)
153 NVC0LegalizePostRA::NVC0LegalizePostRA(const Program *prog)
156 needTexBar(prog->getTarget()->getChipset() >= 0xe0)
161 NVC0LegalizePostRA::insnDominatedBy(const Instruction *later,
162 const Instruction *early) const
164 if (early->bb == later->bb)
165 return early->serial < later->serial;
166 return later->bb->dominatedBy(early->bb);
170 NVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses,
171 Instruction *usei, const Instruction *texi)
174 for (std::list<TexUse>::iterator it = uses.begin();
176 if (insnDominatedBy(usei, it->insn)) {
180 if (insnDominatedBy(it->insn, usei))
186 uses.push_back(TexUse(usei, texi));
189 // While it might be tempting to use the an algorithm that just looks at tex
190 // uses, not all texture results are guaranteed to be used on all paths. In
191 // the case where along some control flow path a texture result is never used,
192 // we might reuse that register for something else, creating a
193 // write-after-write hazard. So we have to manually look through all
194 // instructions looking for ones that reference the registers in question.
196 NVC0LegalizePostRA::findFirstUses(
197 Instruction *texi, std::list<TexUse> &uses)
199 int minGPR = texi->def(0).rep()->reg.data.id;
200 int maxGPR = minGPR + texi->def(0).rep()->reg.size / 4 - 1;
202 unordered_set<const BasicBlock *> visited;
203 findFirstUsesBB(minGPR, maxGPR, texi->next, texi, uses, visited);
207 NVC0LegalizePostRA::findFirstUsesBB(
208 int minGPR, int maxGPR, Instruction *start,
209 const Instruction *texi, std::list<TexUse> &uses,
210 unordered_set<const BasicBlock *> &visited)
212 const BasicBlock *bb = start->bb;
214 // We don't process the whole bb the first time around. This is correct,
215 // however we might be in a loop and hit this BB again, and need to process
216 // the full thing. So only mark a bb as visited if we processed it from the
218 if (start == bb->getEntry()) {
219 if (visited.find(bb) != visited.end())
224 for (Instruction *insn = start; insn != bb->getExit(); insn = insn->next) {
228 for (int d = 0; insn->defExists(d); ++d) {
229 if (insn->def(d).getFile() != FILE_GPR ||
230 insn->def(d).rep()->reg.data.id < minGPR ||
231 insn->def(d).rep()->reg.data.id > maxGPR)
233 addTexUse(uses, insn, texi);
237 for (int s = 0; insn->srcExists(s); ++s) {
238 if (insn->src(s).getFile() != FILE_GPR ||
239 insn->src(s).rep()->reg.data.id < minGPR ||
240 insn->src(s).rep()->reg.data.id > maxGPR)
242 addTexUse(uses, insn, texi);
247 for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
248 findFirstUsesBB(minGPR, maxGPR, BasicBlock::get(ei.getNode())->getEntry(),
249 texi, uses, visited);
254 // This pass is a bit long and ugly and can probably be optimized.
256 // 1. obtain a list of TEXes and their outputs' first use(s)
257 // 2. calculate the barrier level of each first use (minimal number of TEXes,
258 // over all paths, between the TEX and the use in question)
259 // 3. for each barrier, if all paths from the source TEX to that barrier
260 // contain a barrier of lesser level, it can be culled
262 NVC0LegalizePostRA::insertTextureBarriers(Function *fn)
264 std::list<TexUse> *uses;
265 std::vector<Instruction *> texes;
266 std::vector<int> bbFirstTex;
267 std::vector<int> bbFirstUse;
268 std::vector<int> texCounts;
269 std::vector<TexUse> useVec;
272 fn->orderInstructions(insns);
274 texCounts.resize(fn->allBBlocks.getSize(), 0);
275 bbFirstTex.resize(fn->allBBlocks.getSize(), insns.getSize());
276 bbFirstUse.resize(fn->allBBlocks.getSize(), insns.getSize());
278 // tag BB CFG nodes by their id for later
279 for (ArrayList::Iterator i = fn->allBBlocks.iterator(); !i.end(); i.next()) {
280 BasicBlock *bb = reinterpret_cast<BasicBlock *>(i.get());
282 bb->cfg.tag = bb->getId();
285 // gather the first uses for each TEX
286 for (int i = 0; i < insns.getSize(); ++i) {
287 Instruction *tex = reinterpret_cast<Instruction *>(insns.get(i));
288 if (isTextureOp(tex->op)) {
289 texes.push_back(tex);
290 if (!texCounts.at(tex->bb->getId()))
291 bbFirstTex[tex->bb->getId()] = texes.size() - 1;
292 texCounts[tex->bb->getId()]++;
298 uses = new std::list<TexUse>[texes.size()];
301 for (size_t i = 0; i < texes.size(); ++i) {
302 findFirstUses(texes[i], uses[i]);
305 // determine the barrier level at each use
306 for (size_t i = 0; i < texes.size(); ++i) {
307 for (std::list<TexUse>::iterator u = uses[i].begin(); u != uses[i].end();
309 BasicBlock *tb = texes[i]->bb;
310 BasicBlock *ub = u->insn->bb;
313 for (size_t j = i + 1; j < texes.size() &&
314 texes[j]->bb == tb && texes[j]->serial < u->insn->serial;
318 u->level = fn->cfg.findLightestPathWeight(&tb->cfg,
319 &ub->cfg, texCounts);
321 WARN("Failed to find path TEX -> TEXBAR\n");
325 // this counted all TEXes in the origin block, correct that
326 u->level -= i - bbFirstTex.at(tb->getId()) + 1 /* this TEX */;
327 // and did not count the TEXes in the destination block, add those
328 for (size_t j = bbFirstTex.at(ub->getId()); j < texes.size() &&
329 texes[j]->bb == ub && texes[j]->serial < u->insn->serial;
333 assert(u->level >= 0);
334 useVec.push_back(*u);
339 // insert the barriers
340 for (size_t i = 0; i < useVec.size(); ++i) {
341 Instruction *prev = useVec[i].insn->prev;
342 if (useVec[i].level < 0)
344 if (prev && prev->op == OP_TEXBAR) {
345 if (prev->subOp > useVec[i].level)
346 prev->subOp = useVec[i].level;
347 prev->setSrc(prev->srcCount(), useVec[i].tex->getDef(0));
349 Instruction *bar = new_Instruction(func, OP_TEXBAR, TYPE_NONE);
351 bar->subOp = useVec[i].level;
352 // make use explicit to ease latency calculation
353 bar->setSrc(bar->srcCount(), useVec[i].tex->getDef(0));
354 useVec[i].insn->bb->insertBefore(useVec[i].insn, bar);
358 if (fn->getProgram()->optLevel < 3)
361 std::vector<Limits> limitT, limitB, limitS; // entry, exit, single
363 limitT.resize(fn->allBBlocks.getSize(), Limits(0, 0));
364 limitB.resize(fn->allBBlocks.getSize(), Limits(0, 0));
365 limitS.resize(fn->allBBlocks.getSize());
367 // cull unneeded barriers (should do that earlier, but for simplicity)
368 IteratorRef bi = fn->cfg.iteratorCFG();
369 // first calculate min/max outstanding TEXes for each BB
370 for (bi->reset(); !bi->end(); bi->next()) {
371 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
372 BasicBlock *bb = BasicBlock::get(n);
374 int max = std::numeric_limits<int>::max();
375 for (Instruction *i = bb->getFirst(); i; i = i->next) {
376 if (isTextureOp(i->op)) {
378 if (max < std::numeric_limits<int>::max())
381 if (i->op == OP_TEXBAR) {
382 min = MIN2(min, i->subOp);
383 max = MIN2(max, i->subOp);
386 // limits when looking at an isolated block
387 limitS[bb->getId()].min = min;
388 limitS[bb->getId()].max = max;
390 // propagate the min/max values
391 for (unsigned int l = 0; l <= fn->loopNestingBound; ++l) {
392 for (bi->reset(); !bi->end(); bi->next()) {
393 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
394 BasicBlock *bb = BasicBlock::get(n);
395 const int bbId = bb->getId();
396 for (Graph::EdgeIterator ei = n->incident(); !ei.end(); ei.next()) {
397 BasicBlock *in = BasicBlock::get(ei.getNode());
398 const int inId = in->getId();
399 limitT[bbId].min = MAX2(limitT[bbId].min, limitB[inId].min);
400 limitT[bbId].max = MAX2(limitT[bbId].max, limitB[inId].max);
402 // I just hope this is correct ...
403 if (limitS[bbId].max == std::numeric_limits<int>::max()) {
405 limitB[bbId].min = limitT[bbId].min + limitS[bbId].min;
406 limitB[bbId].max = limitT[bbId].max + limitS[bbId].min;
408 // block contained a barrier
409 limitB[bbId].min = MIN2(limitS[bbId].max,
410 limitT[bbId].min + limitS[bbId].min);
411 limitB[bbId].max = MIN2(limitS[bbId].max,
412 limitT[bbId].max + limitS[bbId].min);
416 // finally delete unnecessary barriers
417 for (bi->reset(); !bi->end(); bi->next()) {
418 Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
419 BasicBlock *bb = BasicBlock::get(n);
420 Instruction *prev = NULL;
422 int max = limitT[bb->getId()].max;
423 for (Instruction *i = bb->getFirst(); i; i = next) {
425 if (i->op == OP_TEXBAR) {
426 if (i->subOp >= max) {
427 delete_Instruction(prog, i);
431 if (prev && prev->op == OP_TEXBAR && prev->subOp >= max) {
432 delete_Instruction(prog, prev);
437 if (isTextureOp(i->op)) {
440 if (i && !i->isNop())
448 NVC0LegalizePostRA::visit(Function *fn)
451 insertTextureBarriers(fn);
453 rZero = new_LValue(fn, FILE_GPR);
454 carry = new_LValue(fn, FILE_FLAGS);
456 rZero->reg.data.id = prog->getTarget()->getFileSize(FILE_GPR);
457 carry->reg.data.id = 0;
463 NVC0LegalizePostRA::replaceZero(Instruction *i)
465 for (int s = 0; i->srcExists(s); ++s) {
466 if (s == 2 && i->op == OP_SUCLAMP)
468 ImmediateValue *imm = i->getSrc(s)->asImm();
469 if (imm && imm->reg.data.u64 == 0)
474 // replace CONT with BRA for single unconditional continue
476 NVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock *bb)
478 if (bb->cfg.incidentCount() != 2 || bb->getEntry()->op != OP_PRECONT)
480 Graph::EdgeIterator ei = bb->cfg.incident();
481 if (ei.getType() != Graph::Edge::BACK)
483 if (ei.getType() != Graph::Edge::BACK)
485 BasicBlock *contBB = BasicBlock::get(ei.getNode());
487 if (!contBB->getExit() || contBB->getExit()->op != OP_CONT ||
488 contBB->getExit()->getPredicate())
490 contBB->getExit()->op = OP_BRA;
491 bb->remove(bb->getEntry()); // delete PRECONT
494 assert(ei.end() || ei.getType() != Graph::Edge::BACK);
498 // replace branches to join blocks with join ops
500 NVC0LegalizePostRA::propagateJoin(BasicBlock *bb)
502 if (bb->getEntry()->op != OP_JOIN || bb->getEntry()->asFlow()->limit)
504 for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
505 BasicBlock *in = BasicBlock::get(ei.getNode());
506 Instruction *exit = in->getExit();
508 in->insertTail(new FlowInstruction(func, OP_JOIN, bb));
509 // there should always be a terminator instruction
510 WARN("inserted missing terminator in BB:%i\n", in->getId());
512 if (exit->op == OP_BRA) {
514 exit->asFlow()->limit = 1; // must-not-propagate marker
517 bb->remove(bb->getEntry());
521 NVC0LegalizePostRA::visit(BasicBlock *bb)
523 Instruction *i, *next;
525 // remove pseudo operations and non-fixed no-ops, split 64 bit operations
526 for (i = bb->getFirst(); i; i = next) {
528 if (i->op == OP_EMIT || i->op == OP_RESTART) {
529 if (!i->getDef(0)->refCount())
531 if (i->src(0).getFile() == FILE_IMMEDIATE)
532 i->setSrc(0, rZero); // initial value must be 0
538 if (i->op == OP_BAR && i->subOp == NV50_IR_SUBOP_BAR_SYNC &&
539 prog->getType() != Program::TYPE_COMPUTE) {
540 // It seems like barriers are never required for tessellation since
541 // the warp size is 32, and there are always at most 32 tcs threads.
544 if (i->op == OP_LOAD && i->subOp == NV50_IR_SUBOP_LDC_IS) {
545 int offset = i->src(0).get()->reg.data.offset;
546 if (abs(offset) > 0x10000)
547 i->src(0).get()->reg.fileIndex += offset >> 16;
548 i->src(0).get()->reg.data.offset = (int)(short)offset;
550 // TODO: Move this to before register allocation for operations that
551 // need the $c register !
552 if (typeSizeof(i->dType) == 8) {
554 hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry);
559 if (i->op != OP_MOV && i->op != OP_PFETCH)
566 if (!tryReplaceContWithBra(bb))
572 NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
574 bld.setProgram(prog);
579 NVC0LoweringPass::visit(Function *fn)
581 if (prog->getType() == Program::TYPE_GEOMETRY) {
582 assert(!strncmp(fn->getName(), "MAIN", 4));
583 // TODO: when we generate actual functions pass this value along somehow
584 bld.setPosition(BasicBlock::get(fn->cfg.getRoot()), false);
585 gpEmitAddress = bld.loadImm(NULL, 0)->asLValue();
587 bld.setPosition(BasicBlock::get(fn->cfgExit)->getExit(), false);
588 bld.mkMovToReg(0, gpEmitAddress);
595 NVC0LoweringPass::visit(BasicBlock *bb)
601 NVC0LoweringPass::loadTexHandle(Value *ptr, unsigned int slot)
603 uint8_t b = prog->driver->io.auxCBSlot;
604 uint32_t off = prog->driver->io.texBindBase + slot * 4;
606 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
609 // move array source to first slot, convert to u16, add indirections
611 NVC0LoweringPass::handleTEX(TexInstruction *i)
613 const int dim = i->tex.target.getDim() + i->tex.target.isCube();
614 const int arg = i->tex.target.getArgCount();
615 const int lyr = arg - (i->tex.target.isMS() ? 2 : 1);
616 const int chipset = prog->getTarget()->getChipset();
618 /* Only normalize in the non-explicit derivatives case. For explicit
619 * derivatives, this is handled in handleManualTXD.
621 if (i->tex.target.isCube() && i->dPdx[0].get() == NULL) {
624 for (c = 0; c < 3; ++c)
625 src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), i->getSrc(c));
626 val = bld.getScratch();
627 bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
628 bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
629 bld.mkOp1(OP_RCP, TYPE_F32, val, val);
630 for (c = 0; c < 3; ++c) {
631 i->setSrc(c, bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(),
636 // Arguments to the TEX instruction are a little insane. Even though the
637 // encoding is identical between SM20 and SM30, the arguments mean
638 // different things between Fermi and Kepler+. A lot of arguments are
639 // optional based on flags passed to the instruction. This summarizes the
649 // - tg4: 8 bits each, either 2 (1 offset reg) or 8 (2 offset reg)
650 // - other: 4 bits each, single reg
654 // array (+ offsets for txd in upper 16 bits)
659 // offsets (same as fermi, except txd which takes it with array)
676 if (chipset >= NVISA_GK104_CHIPSET) {
677 if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
678 // XXX this ignores tsc, and assumes a 1:1 mapping
679 assert(i->tex.rIndirectSrc >= 0);
680 Value *hnd = loadTexHandle(
681 bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
682 i->getIndirectR(), bld.mkImm(2)),
686 i->setIndirectR(hnd);
687 i->setIndirectS(NULL);
688 } else if (i->tex.r == i->tex.s || i->op == OP_TXF) {
689 i->tex.r += prog->driver->io.texBindBase / 4;
690 i->tex.s = 0; // only a single cX[] value possible here
692 Value *hnd = bld.getScratch();
693 Value *rHnd = loadTexHandle(NULL, i->tex.r);
694 Value *sHnd = loadTexHandle(NULL, i->tex.s);
696 bld.mkOp3(OP_INSBF, TYPE_U32, hnd, rHnd, bld.mkImm(0x1400), sHnd);
698 i->tex.r = 0; // not used for indirect tex
700 i->setIndirectR(hnd);
702 if (i->tex.target.isArray()) {
703 LValue *layer = new_LValue(func, FILE_GPR);
704 Value *src = i->getSrc(lyr);
705 const int sat = (i->op == OP_TXF) ? 1 : 0;
706 DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
707 bld.mkCvt(OP_CVT, TYPE_U16, layer, sTy, src)->saturate = sat;
708 if (i->op != OP_TXD || chipset < NVISA_GM107_CHIPSET) {
709 for (int s = dim; s >= 1; --s)
710 i->setSrc(s, i->getSrc(s - 1));
713 i->setSrc(dim, layer);
716 // Move the indirect reference to the first place
717 if (i->tex.rIndirectSrc >= 0 && (
718 i->op == OP_TXD || chipset < NVISA_GM107_CHIPSET)) {
719 Value *hnd = i->getIndirectR();
721 i->setIndirectR(NULL);
722 i->moveSources(0, 1);
724 i->tex.rIndirectSrc = 0;
725 i->tex.sIndirectSrc = -1;
728 // (nvc0) generate and move the tsc/tic/array source to the front
729 if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
730 LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
732 Value *ticRel = i->getIndirectR();
733 Value *tscRel = i->getIndirectS();
736 i->setSrc(i->tex.rIndirectSrc, NULL);
738 ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
739 ticRel, bld.mkImm(i->tex.r));
742 i->setSrc(i->tex.sIndirectSrc, NULL);
744 tscRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
745 tscRel, bld.mkImm(i->tex.s));
748 Value *arrayIndex = i->tex.target.isArray() ? i->getSrc(lyr) : NULL;
750 for (int s = dim; s >= 1; --s)
751 i->setSrc(s, i->getSrc(s - 1));
752 i->setSrc(0, arrayIndex);
754 i->moveSources(0, 1);
758 int sat = (i->op == OP_TXF) ? 1 : 0;
759 DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
760 bld.mkCvt(OP_CVT, TYPE_U16, src, sTy, arrayIndex)->saturate = sat;
766 bld.mkOp3(OP_INSBF, TYPE_U32, src, ticRel, bld.mkImm(0x0917), src);
768 bld.mkOp3(OP_INSBF, TYPE_U32, src, tscRel, bld.mkImm(0x0710), src);
773 // For nvc0, the sample id has to be in the second operand, as the offset
774 // does. Right now we don't know how to pass both in, and this case can't
775 // happen with OpenGL. On nve0, the sample id is part of the texture
776 // coordinate argument.
777 assert(chipset >= NVISA_GK104_CHIPSET ||
778 !i->tex.useOffsets || !i->tex.target.isMS());
780 // offset is between lod and dc
781 if (i->tex.useOffsets) {
783 int s = i->srcCount(0xff, true);
784 if (i->op != OP_TXD || chipset < NVISA_GK104_CHIPSET) {
785 if (i->tex.target.isShadow())
787 if (i->srcExists(s)) // move potential predicate out of the way
788 i->moveSources(s, 1);
789 if (i->tex.useOffsets == 4 && i->srcExists(s + 1))
790 i->moveSources(s + 1, 1);
792 if (i->op == OP_TXG) {
793 // Either there is 1 offset, which goes into the 2 low bytes of the
794 // first source, or there are 4 offsets, which go into 2 sources (8
795 // values, 1 byte each).
796 Value *offs[2] = {NULL, NULL};
797 for (n = 0; n < i->tex.useOffsets; n++) {
798 for (c = 0; c < 2; ++c) {
799 if ((n % 2) == 0 && c == 0)
800 offs[n / 2] = i->offset[n][c].get();
802 bld.mkOp3(OP_INSBF, TYPE_U32,
804 i->offset[n][c].get(),
805 bld.mkImm(0x800 | ((n * 16 + c * 8) % 32)),
809 i->setSrc(s, offs[0]);
811 i->setSrc(s + 1, offs[1]);
814 assert(i->tex.useOffsets == 1);
815 for (c = 0; c < 3; ++c) {
817 if (!i->offset[0][c].getImmediate(val))
818 assert(!"non-immediate offset passed to non-TXG");
819 imm |= (val.reg.data.u32 & 0xf) << (c * 4);
821 if (i->op == OP_TXD && chipset >= NVISA_GK104_CHIPSET) {
822 // The offset goes into the upper 16 bits of the array index. So
823 // create it if it's not already there, and INSBF it if it already
825 s = (i->tex.rIndirectSrc >= 0) ? 1 : 0;
826 if (chipset >= NVISA_GM107_CHIPSET)
828 if (i->tex.target.isArray()) {
829 bld.mkOp3(OP_INSBF, TYPE_U32, i->getSrc(s),
830 bld.loadImm(NULL, imm), bld.mkImm(0xc10),
833 i->moveSources(s, 1);
834 i->setSrc(s, bld.loadImm(NULL, imm << 16));
837 i->setSrc(s, bld.loadImm(NULL, imm));
842 if (chipset >= NVISA_GK104_CHIPSET) {
844 // If TEX requires more than 4 sources, the 2nd register tuple must be
845 // aligned to 4, even if it consists of just a single 4-byte register.
847 // XXX HACK: We insert 0 sources to avoid the 5 or 6 regs case.
849 int s = i->srcCount(0xff, true);
850 if (s > 4 && s < 7) {
851 if (i->srcExists(s)) // move potential predicate out of the way
852 i->moveSources(s, 7 - s);
854 i->setSrc(s++, bld.loadImm(NULL, 0));
862 NVC0LoweringPass::handleManualTXD(TexInstruction *i)
864 static const uint8_t qOps[4][2] =
866 { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(MOV2, MOV2, ADD, ADD) }, // l0
867 { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD, ADD) }, // l1
868 { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2
869 { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3
874 Value *zero = bld.loadImm(bld.getSSA(), 0);
876 const int dim = i->tex.target.getDim() + i->tex.target.isCube();
878 // This function is invoked after handleTEX lowering, so we have to expect
879 // the arguments in the order that the hw wants them. For Fermi, array and
880 // indirect are both in the leading arg, while for Kepler, array and
881 // indirect are separate (and both precede the coordinates). Maxwell is
882 // handled in a separate function.
884 if (targ->getChipset() < NVISA_GK104_CHIPSET)
885 array = i->tex.target.isArray() || i->tex.rIndirectSrc >= 0;
887 array = i->tex.target.isArray() + (i->tex.rIndirectSrc >= 0);
889 i->op = OP_TEX; // no need to clone dPdx/dPdy later
891 for (c = 0; c < dim; ++c)
892 crd[c] = bld.getScratch();
894 bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
895 for (l = 0; l < 4; ++l) {
897 // mov coordinates from lane l to all lanes
898 for (c = 0; c < dim; ++c)
899 bld.mkQuadop(0x00, crd[c], l, i->getSrc(c + array), zero);
900 // add dPdx from lane l to lanes dx
901 for (c = 0; c < dim; ++c)
902 bld.mkQuadop(qOps[l][0], crd[c], l, i->dPdx[c].get(), crd[c]);
903 // add dPdy from lane l to lanes dy
904 for (c = 0; c < dim; ++c)
905 bld.mkQuadop(qOps[l][1], crd[c], l, i->dPdy[c].get(), crd[c]);
906 // normalize cube coordinates
907 if (i->tex.target.isCube()) {
908 for (c = 0; c < 3; ++c)
909 src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), crd[c]);
910 val = bld.getScratch();
911 bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
912 bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
913 bld.mkOp1(OP_RCP, TYPE_F32, val, val);
914 for (c = 0; c < 3; ++c)
915 src[c] = bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(), crd[c], val);
917 for (c = 0; c < dim; ++c)
921 bld.insert(tex = cloneForward(func, i));
922 for (c = 0; c < dim; ++c)
923 tex->setSrc(c + array, src[c]);
925 for (c = 0; i->defExists(c); ++c) {
927 def[c][l] = bld.getSSA();
928 mov = bld.mkMov(def[c][l], tex->getDef(c));
933 bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
935 for (c = 0; i->defExists(c); ++c) {
936 Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c));
937 for (l = 0; l < 4; ++l)
938 u->setSrc(l, def[c][l]);
946 NVC0LoweringPass::handleTXD(TexInstruction *txd)
948 int dim = txd->tex.target.getDim() + txd->tex.target.isCube();
949 unsigned arg = txd->tex.target.getArgCount();
950 unsigned expected_args = arg;
951 const int chipset = prog->getTarget()->getChipset();
953 if (chipset >= NVISA_GK104_CHIPSET) {
954 if (!txd->tex.target.isArray() && txd->tex.useOffsets)
956 if (txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0)
959 if (txd->tex.useOffsets)
961 if (!txd->tex.target.isArray() && (
962 txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0))
966 if (expected_args > 4 ||
968 txd->tex.target.isShadow())
972 while (txd->srcExists(arg))
975 txd->tex.derivAll = true;
976 if (txd->op == OP_TEX)
977 return handleManualTXD(txd);
979 assert(arg == expected_args);
980 for (int c = 0; c < dim; ++c) {
981 txd->setSrc(arg + c * 2 + 0, txd->dPdx[c]);
982 txd->setSrc(arg + c * 2 + 1, txd->dPdy[c]);
983 txd->dPdx[c].set(NULL);
984 txd->dPdy[c].set(NULL);
990 NVC0LoweringPass::handleTXQ(TexInstruction *txq)
992 const int chipset = prog->getTarget()->getChipset();
993 if (chipset >= NVISA_GK104_CHIPSET && txq->tex.rIndirectSrc < 0)
994 txq->tex.r += prog->driver->io.texBindBase / 4;
996 if (txq->tex.rIndirectSrc < 0)
999 Value *ticRel = txq->getIndirectR();
1001 txq->setIndirectS(NULL);
1002 txq->tex.sIndirectSrc = -1;
1006 if (chipset < NVISA_GK104_CHIPSET) {
1007 LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa
1009 txq->setSrc(txq->tex.rIndirectSrc, NULL);
1011 ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
1012 ticRel, bld.mkImm(txq->tex.r));
1014 bld.mkOp2(OP_SHL, TYPE_U32, src, ticRel, bld.mkImm(0x17));
1016 txq->moveSources(0, 1);
1017 txq->setSrc(0, src);
1019 Value *hnd = loadTexHandle(
1020 bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
1021 txq->getIndirectR(), bld.mkImm(2)),
1026 txq->setIndirectR(NULL);
1027 txq->moveSources(0, 1);
1028 txq->setSrc(0, hnd);
1029 txq->tex.rIndirectSrc = 0;
1036 NVC0LoweringPass::handleTXLQ(TexInstruction *i)
1038 /* The outputs are inverted compared to what the TGSI instruction
1039 * expects. Take that into account in the mask.
1041 assert((i->tex.mask & ~3) == 0);
1042 if (i->tex.mask == 1)
1044 else if (i->tex.mask == 2)
1047 bld.setPosition(i, true);
1049 /* The returned values are not quite what we want:
1050 * (a) convert from s16/u16 to f32
1051 * (b) multiply by 1/256
1053 for (int def = 0; def < 2; ++def) {
1054 if (!i->defExists(def))
1056 enum DataType type = TYPE_S16;
1057 if (i->tex.mask == 2 || def > 0)
1059 bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(def), type, i->getDef(def));
1060 bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(def),
1061 i->getDef(def), bld.loadImm(NULL, 1.0f / 256));
1063 if (i->tex.mask == 3) {
1064 LValue *t = new_LValue(func, FILE_GPR);
1065 bld.mkMov(t, i->getDef(0));
1066 bld.mkMov(i->getDef(0), i->getDef(1));
1067 bld.mkMov(i->getDef(1), t);
1073 NVC0LoweringPass::handleSUQ(Instruction *suq)
1076 suq->setSrc(0, loadBufLength32(suq->getIndirect(0, 1),
1077 suq->getSrc(0)->reg.fileIndex * 16));
1078 suq->setIndirect(0, 0, NULL);
1079 suq->setIndirect(0, 1, NULL);
1084 NVC0LoweringPass::handleSharedATOMNVE4(Instruction *atom)
1086 assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
1088 BasicBlock *currBB = atom->bb;
1089 BasicBlock *tryLockBB = atom->bb->splitBefore(atom, false);
1090 BasicBlock *joinBB = atom->bb->splitAfter(atom);
1091 BasicBlock *setAndUnlockBB = new BasicBlock(func);
1092 BasicBlock *failLockBB = new BasicBlock(func);
1094 bld.setPosition(currBB, true);
1095 assert(!currBB->joinAt);
1096 currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
1098 CmpInstruction *pred =
1099 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1100 TYPE_U32, bld.mkImm(0), bld.mkImm(1));
1102 bld.mkFlow(OP_BRA, tryLockBB, CC_ALWAYS, NULL);
1103 currBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::TREE);
1105 bld.setPosition(tryLockBB, true);
1108 bld.mkLoad(TYPE_U32, atom->getDef(0),
1109 bld.mkSymbol(FILE_MEMORY_SHARED, 0, TYPE_U32, 0), NULL);
1110 ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
1111 ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;
1113 bld.mkFlow(OP_BRA, setAndUnlockBB, CC_P, ld->getDef(1));
1114 bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
1115 tryLockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::CROSS);
1116 tryLockBB->cfg.attach(&setAndUnlockBB->cfg, Graph::Edge::TREE);
1118 tryLockBB->cfg.detach(&joinBB->cfg);
1121 bld.setPosition(setAndUnlockBB, true);
1123 if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
1124 // Read the old value, and write the new one.
1125 stVal = atom->getSrc(1);
1126 } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1127 CmpInstruction *set =
1128 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(),
1129 TYPE_U32, ld->getDef(0), atom->getSrc(1));
1131 bld.mkCmp(OP_SLCT, CC_NE, TYPE_U32, (stVal = bld.getSSA()),
1132 TYPE_U32, atom->getSrc(2), ld->getDef(0), set->getDef(0));
1136 switch (atom->subOp) {
1137 case NV50_IR_SUBOP_ATOM_ADD:
1140 case NV50_IR_SUBOP_ATOM_AND:
1143 case NV50_IR_SUBOP_ATOM_OR:
1146 case NV50_IR_SUBOP_ATOM_XOR:
1149 case NV50_IR_SUBOP_ATOM_MIN:
1152 case NV50_IR_SUBOP_ATOM_MAX:
1160 stVal = bld.mkOp2v(op, atom->dType, bld.getSSA(), ld->getDef(0),
1165 bld.mkStore(OP_STORE, TYPE_U32,
1166 bld.mkSymbol(FILE_MEMORY_SHARED, 0, TYPE_U32, 0),
1168 st->setDef(0, pred->getDef(0));
1169 st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;
1171 bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
1172 setAndUnlockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::TREE);
1174 // Lock until the store has not been performed.
1175 bld.setPosition(failLockBB, true);
1176 bld.mkFlow(OP_BRA, tryLockBB, CC_NOT_P, pred->getDef(0));
1177 bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
1178 failLockBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::BACK);
1179 failLockBB->cfg.attach(&joinBB->cfg, Graph::Edge::TREE);
1181 bld.setPosition(joinBB, false);
1182 bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1186 NVC0LoweringPass::handleSharedATOM(Instruction *atom)
1188 assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
1190 BasicBlock *currBB = atom->bb;
1191 BasicBlock *tryLockAndSetBB = atom->bb->splitBefore(atom, false);
1192 BasicBlock *joinBB = atom->bb->splitAfter(atom);
1194 bld.setPosition(currBB, true);
1195 assert(!currBB->joinAt);
1196 currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
1198 bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_ALWAYS, NULL);
1199 currBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::TREE);
1201 bld.setPosition(tryLockAndSetBB, true);
1204 bld.mkLoad(TYPE_U32, atom->getDef(0),
1205 bld.mkSymbol(FILE_MEMORY_SHARED, 0, TYPE_U32, 0), NULL);
1206 ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
1207 ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;
1210 if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
1211 // Read the old value, and write the new one.
1212 stVal = atom->getSrc(1);
1213 } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1214 CmpInstruction *set =
1215 bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
1216 TYPE_U32, ld->getDef(0), atom->getSrc(1));
1217 set->setPredicate(CC_P, ld->getDef(1));
1220 bld.mkOp3(OP_SELP, TYPE_U32, bld.getSSA(), ld->getDef(0),
1221 atom->getSrc(2), set->getDef(0));
1222 selp->src(2).mod = Modifier(NV50_IR_MOD_NOT);
1223 selp->setPredicate(CC_P, ld->getDef(1));
1225 stVal = selp->getDef(0);
1229 switch (atom->subOp) {
1230 case NV50_IR_SUBOP_ATOM_ADD:
1233 case NV50_IR_SUBOP_ATOM_AND:
1236 case NV50_IR_SUBOP_ATOM_OR:
1239 case NV50_IR_SUBOP_ATOM_XOR:
1242 case NV50_IR_SUBOP_ATOM_MIN:
1245 case NV50_IR_SUBOP_ATOM_MAX:
1254 bld.mkOp2(op, atom->dType, bld.getSSA(), ld->getDef(0),
1256 i->setPredicate(CC_P, ld->getDef(1));
1258 stVal = i->getDef(0);
1262 bld.mkStore(OP_STORE, TYPE_U32,
1263 bld.mkSymbol(FILE_MEMORY_SHARED, 0, TYPE_U32, 0),
1265 st->setPredicate(CC_P, ld->getDef(1));
1266 st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;
1268 // Loop until the lock is acquired.
1269 bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_NOT_P, ld->getDef(1));
1270 tryLockAndSetBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::BACK);
1271 tryLockAndSetBB->cfg.attach(&joinBB->cfg, Graph::Edge::CROSS);
1272 bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
1276 bld.setPosition(joinBB, false);
1277 bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
1281 NVC0LoweringPass::handleATOM(Instruction *atom)
1284 Value *ptr = atom->getIndirect(0, 0), *ind = atom->getIndirect(0, 1), *base;
1286 switch (atom->src(0).getFile()) {
1287 case FILE_MEMORY_LOCAL:
1290 case FILE_MEMORY_SHARED:
1291 // For Fermi/Kepler, we have to use ld lock/st unlock to perform atomic
1292 // operations on shared memory. For Maxwell, ATOMS is enough.
1293 if (targ->getChipset() < NVISA_GK104_CHIPSET)
1294 handleSharedATOM(atom);
1295 else if (targ->getChipset() < NVISA_GM107_CHIPSET)
1296 handleSharedATOMNVE4(atom);
1299 assert(atom->src(0).getFile() == FILE_MEMORY_GLOBAL);
1300 base = loadBufInfo64(ind, atom->getSrc(0)->reg.fileIndex * 16);
1301 assert(base->reg.size == 8);
1303 base = bld.mkOp2v(OP_ADD, TYPE_U64, base, base, ptr);
1304 assert(base->reg.size == 8);
1305 atom->setIndirect(0, 0, base);
1309 bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(), bld.mkSysVal(sv, 0));
1311 atom->setSrc(0, cloneShallow(func, atom->getSrc(0)));
1312 atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1314 base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr);
1315 atom->setIndirect(0, 1, NULL);
1316 atom->setIndirect(0, 0, base);
1322 NVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl)
1324 if (targ->getChipset() < NVISA_GM107_CHIPSET) {
1325 if (cas->src(0).getFile() == FILE_MEMORY_SHARED) {
1326 // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM().
1331 if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS &&
1332 cas->subOp != NV50_IR_SUBOP_ATOM_EXCH)
1334 bld.setPosition(cas, true);
1337 Instruction *cctl = bld.mkOp1(OP_CCTL, TYPE_NONE, NULL, cas->getSrc(0));
1338 cctl->setIndirect(0, 0, cas->getIndirect(0, 0));
1340 cctl->subOp = NV50_IR_SUBOP_CCTL_IV;
1341 if (cas->isPredicated())
1342 cctl->setPredicate(cas->cc, cas->getPredicate());
1345 if (cas->subOp == NV50_IR_SUBOP_ATOM_CAS) {
1346 // CAS is crazy. It's 2nd source is a double reg, and the 3rd source
1347 // should be set to the high part of the double reg or bad things will
1348 // happen elsewhere in the universe.
1349 // Also, it sometimes returns the new value instead of the old one
1350 // under mysterious circumstances.
1351 Value *dreg = bld.getSSA(8);
1352 bld.setPosition(cas, false);
1353 bld.mkOp2(OP_MERGE, TYPE_U64, dreg, cas->getSrc(1), cas->getSrc(2));
1354 cas->setSrc(1, dreg);
1355 cas->setSrc(2, dreg);
1362 NVC0LoweringPass::loadResInfo32(Value *ptr, uint32_t off, uint16_t base)
1364 uint8_t b = prog->driver->io.auxCBSlot;
1368 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1372 NVC0LoweringPass::loadResInfo64(Value *ptr, uint32_t off, uint16_t base)
1374 uint8_t b = prog->driver->io.auxCBSlot;
1378 ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
1381 mkLoadv(TYPE_U64, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off), ptr);
1385 NVC0LoweringPass::loadResLength32(Value *ptr, uint32_t off, uint16_t base)
1387 uint8_t b = prog->driver->io.auxCBSlot;
1391 ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
1394 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr);
1398 NVC0LoweringPass::loadSuInfo32(Value *ptr, uint32_t off)
1400 return loadResInfo32(ptr, off, prog->driver->io.suInfoBase);
1404 NVC0LoweringPass::loadSuInfo64(Value *ptr, uint32_t off)
1406 return loadResInfo64(ptr, off, prog->driver->io.suInfoBase);
1410 NVC0LoweringPass::loadSuLength32(Value *ptr, uint32_t off)
1412 return loadResLength32(ptr, off, prog->driver->io.suInfoBase);
1416 NVC0LoweringPass::loadBufInfo32(Value *ptr, uint32_t off)
1418 return loadResInfo32(ptr, off, prog->driver->io.bufInfoBase);
1422 NVC0LoweringPass::loadBufInfo64(Value *ptr, uint32_t off)
1424 return loadResInfo64(ptr, off, prog->driver->io.bufInfoBase);
1428 NVC0LoweringPass::loadBufLength32(Value *ptr, uint32_t off)
1430 return loadResLength32(ptr, off, prog->driver->io.bufInfoBase);
1434 NVC0LoweringPass::loadUboInfo32(Value *ptr, uint32_t off)
1436 return loadResInfo32(ptr, off, prog->driver->io.uboInfoBase);
1440 NVC0LoweringPass::loadUboInfo64(Value *ptr, uint32_t off)
1442 return loadResInfo64(ptr, off, prog->driver->io.uboInfoBase);
1446 NVC0LoweringPass::loadUboLength32(Value *ptr, uint32_t off)
1448 return loadResLength32(ptr, off, prog->driver->io.uboInfoBase);
1452 NVC0LoweringPass::loadMsInfo32(Value *ptr, uint32_t off)
1454 uint8_t b = prog->driver->io.msInfoCBSlot;
1455 off += prog->driver->io.msInfoBase;
1457 mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
1460 /* On nvc0, surface info is obtained via the surface binding points passed
1461 * to the SULD/SUST instructions.
1462 * On nve4, surface info is stored in c[] and is used by various special
1463 * instructions, e.g. for clamping coordiantes or generating an address.
1464 * They couldn't just have added an equivalent to TIC now, couldn't they ?
1466 #define NVE4_SU_INFO_ADDR 0x00
1467 #define NVE4_SU_INFO_FMT 0x04
1468 #define NVE4_SU_INFO_DIM_X 0x08
1469 #define NVE4_SU_INFO_PITCH 0x0c
1470 #define NVE4_SU_INFO_DIM_Y 0x10
1471 #define NVE4_SU_INFO_ARRAY 0x14
1472 #define NVE4_SU_INFO_DIM_Z 0x18
1473 #define NVE4_SU_INFO_UNK1C 0x1c
1474 #define NVE4_SU_INFO_WIDTH 0x20
1475 #define NVE4_SU_INFO_HEIGHT 0x24
1476 #define NVE4_SU_INFO_DEPTH 0x28
1477 #define NVE4_SU_INFO_TARGET 0x2c
1478 #define NVE4_SU_INFO_CALL 0x30
1479 #define NVE4_SU_INFO_RAW_X 0x34
1480 #define NVE4_SU_INFO_MS_X 0x38
1481 #define NVE4_SU_INFO_MS_Y 0x3c
1483 #define NVE4_SU_INFO__STRIDE 0x40
1485 #define NVE4_SU_INFO_DIM(i) (0x08 + (i) * 8)
1486 #define NVE4_SU_INFO_SIZE(i) (0x20 + (i) * 4)
1487 #define NVE4_SU_INFO_MS(i) (0x38 + (i) * 4)
1489 static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
1491 switch (su->tex.target.getEnum()) {
1492 case TEX_TARGET_BUFFER: return NV50_IR_SUBOP_SUCLAMP_PL(0, 1);
1493 case TEX_TARGET_RECT: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1494 case TEX_TARGET_1D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1495 case TEX_TARGET_1D_ARRAY: return (c == 1) ?
1496 NV50_IR_SUBOP_SUCLAMP_PL(0, 2) :
1497 NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1498 case TEX_TARGET_2D: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1499 case TEX_TARGET_2D_MS: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
1500 case TEX_TARGET_2D_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1501 case TEX_TARGET_2D_MS_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1502 case TEX_TARGET_3D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1503 case TEX_TARGET_CUBE: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1504 case TEX_TARGET_CUBE_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
1512 NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
1514 const uint16_t base = tex->tex.r * NVE4_SU_INFO__STRIDE;
1515 const int arg = tex->tex.target.getArgCount();
1517 if (tex->tex.target == TEX_TARGET_2D_MS)
1518 tex->tex.target = TEX_TARGET_2D;
1520 if (tex->tex.target == TEX_TARGET_2D_MS_ARRAY)
1521 tex->tex.target = TEX_TARGET_2D_ARRAY;
1525 Value *x = tex->getSrc(0);
1526 Value *y = tex->getSrc(1);
1527 Value *s = tex->getSrc(arg - 1);
1529 Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
1531 Value *ms_x = loadSuInfo32(NULL, base + NVE4_SU_INFO_MS(0));
1532 Value *ms_y = loadSuInfo32(NULL, base + NVE4_SU_INFO_MS(1));
1534 bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
1535 bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);
1537 s = bld.mkOp2v(OP_AND, TYPE_U32, ts, s, bld.loadImm(NULL, 0x7));
1538 s = bld.mkOp2v(OP_SHL, TYPE_U32, ts, ts, bld.mkImm(3));
1540 Value *dx = loadMsInfo32(ts, 0x0);
1541 Value *dy = loadMsInfo32(ts, 0x4);
1543 bld.mkOp2(OP_ADD, TYPE_U32, tx, tx, dx);
1544 bld.mkOp2(OP_ADD, TYPE_U32, ty, ty, dy);
1548 tex->moveSources(arg, -1);
1551 // Sets 64-bit "generic address", predicate and format sources for SULD/SUST.
1552 // They're computed from the coordinates using the surface info in c[] space.
1554 NVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction *su)
1557 const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP;
1559 su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB;
1560 const int idx = su->tex.r;
1561 const int dim = su->tex.target.getDim();
1562 const int arg = dim + (su->tex.target.isArray() ? 1 : 0);
1563 const uint16_t base = idx * NVE4_SU_INFO__STRIDE;
1565 Value *zero = bld.mkImm(0);
1569 Value *bf, *eau, *off;
1572 off = bld.getScratch(4);
1573 bf = bld.getScratch(4);
1574 addr = bld.getSSA(8);
1575 pred = bld.getScratch(1, FILE_PREDICATE);
1577 bld.setPosition(su, false);
1579 adjustCoordinatesMS(su);
1581 // calculate clamped coordinates
1582 for (c = 0; c < arg; ++c) {
1583 src[c] = bld.getScratch();
1585 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_RAW_X);
1587 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_DIM(c));
1588 bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero)
1589 ->subOp = getSuClampSubOp(su, c);
1594 // set predicate output
1595 if (su->tex.target == TEX_TARGET_BUFFER) {
1596 src[0]->getInsn()->setFlagsDef(1, pred);
1598 if (su->tex.target.isArray()) {
1599 p1 = bld.getSSA(1, FILE_PREDICATE);
1600 src[dim]->getInsn()->setFlagsDef(1, p1);
1603 // calculate pixel offset
1605 if (su->tex.target != TEX_TARGET_BUFFER)
1606 bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff));
1609 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_UNK1C);
1610 bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1])
1611 ->subOp = NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1613 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_PITCH);
1614 bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0])
1615 ->subOp = NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
1618 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_PITCH);
1619 bld.mkOp3(OP_MADSP, TYPE_U32, off, src[1], v, src[0])
1620 ->subOp = su->tex.target.isArray() ?
1621 NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
1624 // calculate effective address part 1
1625 if (su->tex.target == TEX_TARGET_BUFFER) {
1629 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_FMT);
1630 bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero)
1631 ->subOp = NV50_IR_SUBOP_V1(7,6,8|2);
1645 if (!su->tex.target.isArray()) {
1646 z = loadSuInfo32(NULL, base + NVE4_SU_INFO_UNK1C);
1647 subOp = NV50_IR_SUBOP_SUBFM_3D;
1651 subOp = NV50_IR_SUBOP_SUBFM_3D;
1655 insn = bld.mkOp3(OP_SUBFM, TYPE_U32, bf, src[0], y, z);
1656 insn->subOp = subOp;
1657 insn->setFlagsDef(1, pred);
1661 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_ADDR);
1663 if (su->tex.target == TEX_TARGET_BUFFER) {
1666 eau = bld.mkOp3v(OP_SUEAU, TYPE_U32, bld.getScratch(4), off, bf, v);
1668 // add array layer offset
1669 if (su->tex.target.isArray()) {
1670 v = loadSuInfo32(NULL, base + NVE4_SU_INFO_ARRAY);
1672 bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau)
1673 ->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
1675 bld.mkOp3(OP_MADSP, TYPE_U32, eau, v, src[2], eau)
1676 ->subOp = NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32
1677 // combine predicates
1679 bld.mkOp2(OP_OR, TYPE_U8, pred, pred, p1);
1684 if (su->tex.target == TEX_TARGET_BUFFER) {
1688 // bf == g[] address & 0xff
1689 // eau == g[] address >> 8
1690 bld.mkOp3(OP_PERMT, TYPE_U32, bf, lo, bld.loadImm(NULL, 0x6540), eau);
1691 bld.mkOp3(OP_PERMT, TYPE_U32, eau, zero, bld.loadImm(NULL, 0x0007), eau);
1693 if (su->op == OP_SULDP && su->tex.target == TEX_TARGET_BUFFER) {
1694 // Convert from u32 to u8 address format, which is what the library code
1695 // doing SULDP currently uses.
1696 // XXX: can SUEAU do this ?
1697 // XXX: does it matter that we don't mask high bytes in bf ?
1699 bld.mkOp2(OP_SHR, TYPE_U32, off, bf, bld.mkImm(8));
1700 bld.mkOp2(OP_ADD, TYPE_U32, eau, eau, off);
1703 bld.mkOp2(OP_MERGE, TYPE_U64, addr, bf, eau);
1705 if (atom && su->tex.target == TEX_TARGET_BUFFER)
1706 bld.mkOp2(OP_ADD, TYPE_U64, addr, addr, off);
1708 // let's just set it 0 for raw access and hope it works
1710 bld.mkImm(0) : loadSuInfo32(NULL, base + NVE4_SU_INFO_FMT);
1712 // get rid of old coordinate sources, make space for fmt info and predicate
1713 su->moveSources(arg, 3 - arg);
1714 // set 64 bit address and 32-bit format sources
1715 su->setSrc(0, addr);
1717 su->setSrc(2, pred);
1721 NVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction *su)
1723 processSurfaceCoordsNVE4(su);
1725 // Who do we hate more ? The person who decided that nvc0's SULD doesn't
1726 // have to support conversion or the person who decided that, in OpenCL,
1727 // you don't have to specify the format here like you do in OpenGL ?
1729 if (su->op == OP_SULDP) {
1730 // We don't patch shaders. Ever.
1731 // You get an indirect call to our library blob here.
1732 // But at least it's uniform.
1733 FlowInstruction *call;
1736 uint16_t base = su->tex.r * NVE4_SU_INFO__STRIDE + NVE4_SU_INFO_CALL;
1738 for (int i = 0; i < 4; ++i)
1739 (r[i] = bld.getScratch(4, FILE_GPR))->reg.data.id = i;
1740 for (int i = 0; i < 3; ++i)
1741 (p[i] = bld.getScratch(1, FILE_PREDICATE))->reg.data.id = i;
1742 (r[4] = bld.getScratch(8, FILE_GPR))->reg.data.id = 4;
1744 bld.mkMov(p[1], bld.mkImm((su->cache == CACHE_CA) ? 1 : 0), TYPE_U8);
1745 bld.mkMov(p[2], bld.mkImm((su->cache == CACHE_CG) ? 1 : 0), TYPE_U8);
1746 bld.mkMov(p[0], su->getSrc(2), TYPE_U8);
1747 bld.mkMov(r[4], su->getSrc(0), TYPE_U64);
1748 bld.mkMov(r[2], su->getSrc(1), TYPE_U32);
1750 call = bld.mkFlow(OP_CALL, NULL, su->cc, su->getPredicate());
1754 call->setSrc(0, bld.mkSymbol(FILE_MEMORY_CONST,
1755 prog->driver->io.auxCBSlot, TYPE_U32,
1756 prog->driver->io.suInfoBase + base));
1757 call->setSrc(1, r[2]);
1758 call->setSrc(2, r[4]);
1759 for (int i = 0; i < 3; ++i)
1760 call->setSrc(3 + i, p[i]);
1761 for (int i = 0; i < 4; ++i) {
1762 call->setDef(i, r[i]);
1763 bld.mkMov(su->getDef(i), r[i]);
1765 call->setDef(4, p[1]);
1766 delete_Instruction(bld.getProgram(), su);
1769 if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
1770 // FIXME: for out of bounds access, destination value will be undefined !
1771 Value *pred = su->getSrc(2);
1772 CondCode cc = CC_NOT_P;
1773 if (su->getPredicate()) {
1774 pred = bld.getScratch(1, FILE_PREDICATE);
1776 if (cc == CC_NOT_P) {
1777 bld.mkOp2(OP_OR, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
1779 bld.mkOp2(OP_AND, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
1780 pred->getInsn()->src(1).mod = Modifier(NV50_IR_MOD_NOT);
1783 Instruction *red = bld.mkOp(OP_ATOM, su->dType, su->getDef(0));
1784 red->subOp = su->subOp;
1786 gMemBase = bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0);
1787 red->setSrc(0, gMemBase);
1788 red->setSrc(1, su->getSrc(3));
1789 if (su->subOp == NV50_IR_SUBOP_ATOM_CAS)
1790 red->setSrc(2, su->getSrc(4));
1791 red->setIndirect(0, 0, su->getSrc(0));
1792 red->setPredicate(cc, pred);
1793 delete_Instruction(bld.getProgram(), su);
1794 handleCasExch(red, true);
1796 su->sType = (su->tex.target == TEX_TARGET_BUFFER) ? TYPE_U32 : TYPE_U8;
1801 NVC0LoweringPass::handleWRSV(Instruction *i)
1807 // must replace, $sreg are not writeable
1808 addr = targ->getSVAddress(FILE_SHADER_OUTPUT, i->getSrc(0)->asSym());
1811 sym = bld.mkSymbol(FILE_SHADER_OUTPUT, 0, i->sType, addr);
1813 st = bld.mkStore(OP_EXPORT, i->dType, sym, i->getIndirect(0, 0),
1815 st->perPatch = i->perPatch;
1817 bld.getBB()->remove(i);
1822 NVC0LoweringPass::handleLDST(Instruction *i)
1824 if (i->src(0).getFile() == FILE_SHADER_INPUT) {
1825 if (prog->getType() == Program::TYPE_COMPUTE) {
1826 i->getSrc(0)->reg.file = FILE_MEMORY_CONST;
1827 i->getSrc(0)->reg.fileIndex = 0;
1829 if (prog->getType() == Program::TYPE_GEOMETRY &&
1830 i->src(0).isIndirect(0)) {
1831 // XXX: this assumes vec4 units
1832 Value *ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
1833 i->getIndirect(0, 0), bld.mkImm(4));
1834 i->setIndirect(0, 0, ptr);
1838 assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP
1840 } else if (i->src(0).getFile() == FILE_MEMORY_CONST) {
1841 if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
1842 prog->getType() == Program::TYPE_COMPUTE) {
1843 // The launch descriptor only allows to set up 8 CBs, but OpenGL
1844 // requires at least 12 UBOs. To bypass this limitation, we store the
1845 // addrs into the driver constbuf and we directly load from the global
1847 int8_t fileIndex = i->getSrc(0)->reg.fileIndex - 1;
1848 Value *ind = i->getIndirect(0, 1);
1849 Value *ptr = loadUboInfo64(ind, fileIndex * 16);
1851 // TODO: clamp the offset to the maximum number of const buf.
1852 if (i->src(0).isIndirect(1)) {
1853 Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
1854 Value *length = loadUboLength32(ind, fileIndex * 16);
1855 Value *pred = new_LValue(func, FILE_PREDICATE);
1856 if (i->src(0).isIndirect(0)) {
1857 bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
1858 bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
1860 i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1861 i->setIndirect(0, 1, NULL);
1862 i->setIndirect(0, 0, ptr);
1863 bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
1864 i->setPredicate(CC_NOT_P, pred);
1865 if (i->defExists(0)) {
1866 bld.mkMov(i->getDef(0), bld.mkImm(0));
1868 } else if (fileIndex >= 0) {
1869 if (i->src(0).isIndirect(0)) {
1870 bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
1872 i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
1873 i->setIndirect(0, 1, NULL);
1874 i->setIndirect(0, 0, ptr);
1876 } else if (i->src(0).isIndirect(1)) {
1878 if (i->src(0).isIndirect(0))
1879 ptr = bld.mkOp3v(OP_INSBF, TYPE_U32, bld.getSSA(),
1880 i->getIndirect(0, 1), bld.mkImm(0x1010),
1881 i->getIndirect(0, 0));
1883 ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
1884 i->getIndirect(0, 1), bld.mkImm(16));
1885 i->setIndirect(0, 1, NULL);
1886 i->setIndirect(0, 0, ptr);
1887 i->subOp = NV50_IR_SUBOP_LDC_IS;
1889 } else if (i->src(0).getFile() == FILE_SHADER_OUTPUT) {
1890 assert(prog->getType() == Program::TYPE_TESSELLATION_CONTROL);
1892 } else if (i->src(0).getFile() == FILE_MEMORY_GLOBAL) {
1893 Value *ind = i->getIndirect(0, 1);
1894 Value *ptr = loadBufInfo64(ind, i->getSrc(0)->reg.fileIndex * 16);
1895 // XXX come up with a way not to do this for EVERY little access but
1896 // rather to batch these up somehow. Unfortunately we've lost the
1897 // information about the field width by the time we get here.
1898 Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
1899 Value *length = loadBufLength32(ind, i->getSrc(0)->reg.fileIndex * 16);
1900 Value *pred = new_LValue(func, FILE_PREDICATE);
1901 if (i->src(0).isIndirect(0)) {
1902 bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
1903 bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
1905 i->setIndirect(0, 1, NULL);
1906 i->setIndirect(0, 0, ptr);
1907 bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
1908 i->setPredicate(CC_NOT_P, pred);
1909 if (i->defExists(0)) {
1910 bld.mkMov(i->getDef(0), bld.mkImm(0));
1916 NVC0LoweringPass::readTessCoord(LValue *dst, int c)
1918 Value *laneid = bld.getSSA();
1921 bld.mkOp1(OP_RDSV, TYPE_U32, laneid, bld.mkSysVal(SV_LANEID, 0));
1936 bld.mkFetch(x, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f0, NULL, laneid);
1938 bld.mkFetch(y, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f4, NULL, laneid);
1941 bld.mkOp2(OP_ADD, TYPE_F32, dst, x, y);
1942 bld.mkOp2(OP_SUB, TYPE_F32, dst, bld.loadImm(NULL, 1.0f), dst);
1947 NVC0LoweringPass::handleRDSV(Instruction *i)
1949 Symbol *sym = i->getSrc(0)->asSym();
1950 const SVSemantic sv = sym->reg.data.sv.sv;
1953 uint32_t addr = targ->getSVAddress(FILE_SHADER_INPUT, sym);
1955 if (addr >= 0x400) {
1957 if (sym->reg.data.sv.index == 3) {
1958 // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
1960 i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0));
1962 if (sv == SV_VERTEX_COUNT) {
1963 bld.setPosition(i, true);
1964 bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0), bld.mkImm(0x808));
1971 assert(prog->getType() == Program::TYPE_FRAGMENT);
1972 if (i->srcExists(1)) {
1973 // Pass offset through to the interpolation logic
1974 ld = bld.mkInterp(NV50_IR_INTERP_LINEAR | NV50_IR_INTERP_OFFSET,
1975 i->getDef(0), addr, NULL);
1976 ld->setSrc(1, i->getSrc(1));
1978 bld.mkInterp(NV50_IR_INTERP_LINEAR, i->getDef(0), addr, NULL);
1983 Value *face = i->getDef(0);
1984 bld.mkInterp(NV50_IR_INTERP_FLAT, face, addr, NULL);
1985 if (i->dType == TYPE_F32) {
1986 bld.mkOp2(OP_OR, TYPE_U32, face, face, bld.mkImm(0x00000001));
1987 bld.mkOp1(OP_NEG, TYPE_S32, face, face);
1988 bld.mkCvt(OP_CVT, TYPE_F32, face, TYPE_S32, face);
1993 assert(prog->getType() == Program::TYPE_TESSELLATION_EVAL);
1994 readTessCoord(i->getDef(0)->asLValue(), i->getSrc(0)->reg.data.sv.index);
1999 assert(targ->getChipset() >= NVISA_GK104_CHIPSET); // mov $sreg otherwise
2000 if (sym->reg.data.sv.index == 3) {
2002 i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1));
2005 addr += prog->driver->prop.cp.gridInfoBase;
2006 bld.mkLoad(TYPE_U32, i->getDef(0),
2007 bld.mkSymbol(FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2008 TYPE_U32, addr), NULL);
2010 case SV_SAMPLE_INDEX:
2011 // TODO: Properly pass source as an address in the PIX address space
2012 // (which can be of the form [r0+offset]). But this is currently
2014 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2015 ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2017 case SV_SAMPLE_POS: {
2018 Value *off = new_LValue(func, FILE_GPR);
2019 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2020 ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
2021 bld.mkOp2(OP_SHL, TYPE_U32, off, i->getDef(0), bld.mkImm(3));
2022 bld.mkLoad(TYPE_F32,
2025 FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
2026 TYPE_U32, prog->driver->io.sampleInfoBase +
2027 4 * sym->reg.data.sv.index),
2031 case SV_SAMPLE_MASK:
2032 ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
2033 ld->subOp = NV50_IR_SUBOP_PIXLD_COVMASK;
2036 case SV_BASEINSTANCE:
2038 ld = bld.mkLoad(TYPE_U32, i->getDef(0),
2039 bld.mkSymbol(FILE_MEMORY_CONST,
2040 prog->driver->io.auxCBSlot,
2042 prog->driver->io.drawInfoBase +
2043 4 * (sv - SV_BASEVERTEX)),
2047 if (prog->getType() == Program::TYPE_TESSELLATION_EVAL && !i->perPatch)
2048 vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0));
2049 ld = bld.mkFetch(i->getDef(0), i->dType,
2050 FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
2051 ld->perPatch = i->perPatch;
2054 bld.getBB()->remove(i);
2059 NVC0LoweringPass::handleDIV(Instruction *i)
2061 if (!isFloatType(i->dType))
2063 bld.setPosition(i, false);
2064 Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(typeSizeof(i->dType)), i->getSrc(1));
2066 i->setSrc(1, rcp->getDef(0));
2071 NVC0LoweringPass::handleMOD(Instruction *i)
2073 if (!isFloatType(i->dType))
2075 LValue *value = bld.getScratch(typeSizeof(i->dType));
2076 bld.mkOp1(OP_RCP, i->dType, value, i->getSrc(1));
2077 bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(0), value);
2078 bld.mkOp1(OP_TRUNC, i->dType, value, value);
2079 bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(1), value);
2081 i->setSrc(1, value);
2086 NVC0LoweringPass::handleSQRT(Instruction *i)
2088 if (i->dType == TYPE_F64) {
2089 Value *pred = bld.getSSA(1, FILE_PREDICATE);
2090 Value *zero = bld.loadImm(NULL, 0.0);
2091 Value *dst = bld.getSSA(8);
2092 bld.mkOp1(OP_RSQ, i->dType, dst, i->getSrc(0));
2093 bld.mkCmp(OP_SET, CC_LE, i->dType, pred, i->dType, i->getSrc(0), zero);
2094 bld.mkOp3(OP_SELP, TYPE_U64, dst, zero, dst, pred);
2097 // TODO: Handle this properly with a library function
2099 bld.setPosition(i, true);
2101 bld.mkOp1(OP_RCP, i->dType, i->getDef(0), i->getDef(0));
2108 NVC0LoweringPass::handlePOW(Instruction *i)
2110 LValue *val = bld.getScratch();
2112 bld.mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0));
2113 bld.mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1;
2114 bld.mkOp1(OP_PREEX2, TYPE_F32, val, val);
2124 NVC0LoweringPass::handleEXPORT(Instruction *i)
2126 if (prog->getType() == Program::TYPE_FRAGMENT) {
2127 int id = i->getSrc(0)->reg.data.offset / 4;
2129 if (i->src(0).isIndirect(0)) // TODO, ugly
2132 i->subOp = NV50_IR_SUBOP_MOV_FINAL;
2133 i->src(0).set(i->src(1));
2135 i->setDef(0, new_LValue(func, FILE_GPR));
2136 i->getDef(0)->reg.data.id = id;
2138 prog->maxGPR = MAX2(prog->maxGPR, id);
2140 if (prog->getType() == Program::TYPE_GEOMETRY) {
2141 i->setIndirect(0, 1, gpEmitAddress);
2147 NVC0LoweringPass::handleOUT(Instruction *i)
2149 Instruction *prev = i->prev;
2150 ImmediateValue stream, prevStream;
2152 // Only merge if the stream ids match. Also, note that the previous
2153 // instruction would have already been lowered, so we take arg1 from it.
2154 if (i->op == OP_RESTART && prev && prev->op == OP_EMIT &&
2155 i->src(0).getImmediate(stream) &&
2156 prev->src(1).getImmediate(prevStream) &&
2157 stream.reg.data.u32 == prevStream.reg.data.u32) {
2158 i->prev->subOp = NV50_IR_SUBOP_EMIT_RESTART;
2159 delete_Instruction(prog, i);
2161 assert(gpEmitAddress);
2162 i->setDef(0, gpEmitAddress);
2163 i->setSrc(1, i->getSrc(0));
2164 i->setSrc(0, gpEmitAddress);
2169 // Generate a binary predicate if an instruction is predicated by
2170 // e.g. an f32 value.
2172 NVC0LoweringPass::checkPredicate(Instruction *insn)
2174 Value *pred = insn->getPredicate();
2177 if (!pred || pred->reg.file == FILE_PREDICATE)
2179 pdst = new_LValue(func, FILE_PREDICATE);
2181 // CAUTION: don't use pdst->getInsn, the definition might not be unique,
2182 // delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass
2184 bld.mkCmp(OP_SET, CC_NEU, insn->dType, pdst, insn->dType, bld.mkImm(0), pred);
2186 insn->setPredicate(insn->cc, pdst);
2190 // - add quadop dance for texturing
2191 // - put FP outputs in GPRs
2192 // - convert instruction sequences
2195 NVC0LoweringPass::visit(Instruction *i)
2198 bld.setPosition(i, false);
2200 if (i->cc != CC_ALWAYS)
2209 return handleTEX(i->asTex());
2211 return handleTXD(i->asTex());
2213 return handleTXLQ(i->asTex());
2215 return handleTXQ(i->asTex());
2217 bld.mkOp1(OP_PREEX2, TYPE_F32, i->getDef(0), i->getSrc(0));
2218 i->setSrc(0, i->getDef(0));
2221 return handlePOW(i);
2223 return handleDIV(i);
2225 return handleMOD(i);
2227 return handleSQRT(i);
2229 ret = handleEXPORT(i);
2233 return handleOUT(i);
2235 return handleRDSV(i);
2237 return handleWRSV(i);
2244 const bool cctl = i->src(0).getFile() == FILE_MEMORY_GLOBAL;
2246 handleCasExch(i, cctl);
2255 if (targ->getChipset() >= NVISA_GK104_CHIPSET)
2256 handleSurfaceOpNVE4(i->asTex());
2265 /* Kepler+ has a special opcode to compute a new base address to be used
2266 * for indirect loads.
2268 if (targ->getChipset() >= NVISA_GK104_CHIPSET && !i->perPatch &&
2269 (i->op == OP_VFETCH || i->op == OP_EXPORT) && i->src(0).isIndirect(0)) {
2270 Instruction *afetch = bld.mkOp1(OP_AFETCH, TYPE_U32, bld.getSSA(),
2271 cloneShallow(func, i->getSrc(0)));
2272 afetch->setIndirect(0, 0, i->getIndirect(0, 0));
2273 i->src(0).get()->reg.data.offset = 0;
2274 i->setIndirect(0, 0, afetch->getDef(0));
2281 TargetNVC0::runLegalizePass(Program *prog, CGStage stage) const
2283 if (stage == CG_STAGE_PRE_SSA) {
2284 NVC0LoweringPass pass(prog);
2285 return pass.run(prog, false, true);
2287 if (stage == CG_STAGE_POST_RA) {
2288 NVC0LegalizePostRA pass(prog);
2289 return pass.run(prog, false, true);
2291 if (stage == CG_STAGE_SSA) {
2292 NVC0LegalizeSSA pass;
2293 return pass.run(prog, false, true);
2298 } // namespace nv50_ir