xref: /external/mesa3d/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp

/*
 * Copyright 2011 Christoph Bumiller
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 * OTHER DEALINGS IN THE SOFTWARE.
 */

#include "codegen/nv50_ir.h"
#include "codegen/nv50_ir_build_util.h"

#include "codegen/nv50_ir_target_nvc0.h"
#include "codegen/nv50_ir_lowering_nvc0.h"

#include <limits>

namespace nv50_ir {

#define QOP_ADD  0
#define QOP_SUBR 1
#define QOP_SUB  2
#define QOP_MOV2 3

//             UL UR LL LR
#define QUADOP(q, r, s, t)                      \
   ((QOP_##q << 6) | (QOP_##r << 4) |           \
    (QOP_##s << 2) | (QOP_##t << 0))

void
NVC0LegalizeSSA::handleDIV(Instruction *i)
{
   FlowInstruction *call;
   int builtin;
   Value *def[2];

   bld.setPosition(i, false);
   def[0] = bld.mkMovToReg(0, i->getSrc(0))->getDef(0);
   def[1] = bld.mkMovToReg(1, i->getSrc(1))->getDef(0);
   switch (i->dType) {
   case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break;
   case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break;
   default:
      return;
   }
   call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL);
   bld.mkMov(i->getDef(0), def[(i->op == OP_DIV) ? 0 : 1]);
   bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2);
   bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0);

   call->fixed = 1;
   call->absolute = call->builtin = 1;
   call->target.builtin = builtin;
   delete_Instruction(prog, i);
}

void
NVC0LegalizeSSA::handleRCPRSQ(Instruction *i)
{
   assert(i->dType == TYPE_F64);
   // There are instructions that will compute the high 32 bits of the 64-bit
   // float. We will just stick 0 in the bottom 32 bits.

   bld.setPosition(i, false);

   // 1. Take the source and it up.
   Value *src[2], *dst[2], *def = i->getDef(0);
   bld.mkSplit(src, 4, i->getSrc(0));

   // 2. We don't care about the low 32 bits of the destination. Stick a 0 in.
   dst[0] = bld.loadImm(NULL, 0);
   dst[1] = bld.getSSA();

   // 3. The new version of the instruction takes the high 32 bits of the
   // source and outputs the high 32 bits of the destination.
   i->setSrc(0, src[1]);
   i->setDef(0, dst[1]);
   i->setType(TYPE_F32);
   i->subOp = NV50_IR_SUBOP_RCPRSQ_64H;

   // 4. Recombine the two dst pieces back into the original destination.
   bld.setPosition(i, true);
   bld.mkOp2(OP_MERGE, TYPE_U64, def, dst[0], dst[1]);
}

void
NVC0LegalizeSSA::handleFTZ(Instruction *i)
{
   // Only want to flush float inputs
   assert(i->sType == TYPE_F32);

   // If we're already flushing denorms (and NaN's) to zero, no need for this.
   if (i->dnz)
      return;

   // Only certain classes of operations can flush
   OpClass cls = prog->getTarget()->getOpClass(i->op);
   if (cls != OPCLASS_ARITH && cls != OPCLASS_COMPARE &&
       cls != OPCLASS_CONVERT)
      return;

   i->ftz = true;
}

bool
NVC0LegalizeSSA::visit(Function *fn)
{
   bld.setProgram(fn->getProgram());
   return true;
}

bool
NVC0LegalizeSSA::visit(BasicBlock *bb)
{
   Instruction *next;
   for (Instruction *i = bb->getEntry(); i; i = next) {
      next = i->next;
      if (i->sType == TYPE_F32) {
         if (prog->getType() != Program::TYPE_COMPUTE)
            handleFTZ(i);
         continue;
      }
      switch (i->op) {
      case OP_DIV:
      case OP_MOD:
         handleDIV(i);
         break;
      case OP_RCP:
      case OP_RSQ:
         if (i->dType == TYPE_F64)
            handleRCPRSQ(i);
         break;
      default:
         break;
      }
   }
   return true;
}

NVC0LegalizePostRA::NVC0LegalizePostRA(const Program *prog)
   : rZero(NULL),
     carry(NULL),
     pOne(NULL),
     needTexBar(prog->getTarget()->getChipset() >= 0xe0)
{
}

bool
NVC0LegalizePostRA::insnDominatedBy(const Instruction *later,
                                    const Instruction *early) const
{
   if (early->bb == later->bb)
      return early->serial < later->serial;
   return later->bb->dominatedBy(early->bb);
}

void
NVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses,
                              Instruction *usei, const Instruction *texi)
{
   bool add = true;
   bool dominated = insnDominatedBy(usei, texi);
   // Uses before the tex have to all be included. Just because an earlier
   // instruction dominates another instruction doesn't mean that there's no
   // way to get from the tex to the later instruction. For example you could
   // have nested loops, with the tex in the inner loop, and uses before it in
   // both loops - even though the outer loop's instruction would dominate the
   // inner's, we still want a texbar before the inner loop's instruction.
   //
   // However we can still use the eliding logic between uses dominated by the
   // tex instruction, as that is unambiguously correct.
   if (dominated) {
      for (std::list<TexUse>::iterator it = uses.begin(); it != uses.end();) {
         if (it->after) {
            if (insnDominatedBy(usei, it->insn)) {
               add = false;
               break;
            }
            if (insnDominatedBy(it->insn, usei)) {
               it = uses.erase(it);
               continue;
            }
         }
         ++it;
      }
   }
   if (add)
      uses.push_back(TexUse(usei, texi, dominated));
}

// While it might be tempting to use the an algorithm that just looks at tex
// uses, not all texture results are guaranteed to be used on all paths. In
// the case where along some control flow path a texture result is never used,
// we might reuse that register for something else, creating a
// write-after-write hazard. So we have to manually look through all
// instructions looking for ones that reference the registers in question.
void
NVC0LegalizePostRA::findFirstUses(
   Instruction *texi, std::list<TexUse> &uses)
{
   int minGPR = texi->def(0).rep()->reg.data.id;
   int maxGPR = minGPR + texi->def(0).rep()->reg.size / 4 - 1;

   unordered_set<const BasicBlock *> visited;
   findFirstUsesBB(minGPR, maxGPR, texi->next, texi, uses, visited);
}

void
NVC0LegalizePostRA::findFirstUsesBB(
   int minGPR, int maxGPR, Instruction *start,
   const Instruction *texi, std::list<TexUse> &uses,
   unordered_set<const BasicBlock *> &visited)
{
   const BasicBlock *bb = start->bb;

   // We don't process the whole bb the first time around. This is correct,
   // however we might be in a loop and hit this BB again, and need to process
   // the full thing. So only mark a bb as visited if we processed it from the
   // beginning.
   if (start == bb->getEntry()) {
      if (visited.find(bb) != visited.end())
         return;
      visited.insert(bb);
   }

   for (Instruction *insn = start; insn != bb->getExit(); insn = insn->next) {
      if (insn->isNop())
         continue;

      for (int d = 0; insn->defExists(d); ++d) {
         const Value *def = insn->def(d).rep();
         if (insn->def(d).getFile() != FILE_GPR ||
             def->reg.data.id + def->reg.size / 4 - 1 < minGPR ||
             def->reg.data.id > maxGPR)
            continue;
         addTexUse(uses, insn, texi);
         return;
      }

      for (int s = 0; insn->srcExists(s); ++s) {
         const Value *src = insn->src(s).rep();
         if (insn->src(s).getFile() != FILE_GPR ||
             src->reg.data.id + src->reg.size / 4 - 1 < minGPR ||
             src->reg.data.id > maxGPR)
            continue;
         addTexUse(uses, insn, texi);
         return;
      }
   }

   for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) {
      findFirstUsesBB(minGPR, maxGPR, BasicBlock::get(ei.getNode())->getEntry(),
                      texi, uses, visited);
   }
}

// Texture barriers:
// This pass is a bit long and ugly and can probably be optimized.
//
// 1. obtain a list of TEXes and their outputs' first use(s)
// 2. calculate the barrier level of each first use (minimal number of TEXes,
//    over all paths, between the TEX and the use in question)
// 3. for each barrier, if all paths from the source TEX to that barrier
//    contain a barrier of lesser level, it can be culled
bool
NVC0LegalizePostRA::insertTextureBarriers(Function *fn)
{
   std::list<TexUse> *uses;
   std::vector<Instruction *> texes;
   std::vector<int> bbFirstTex;
   std::vector<int> bbFirstUse;
   std::vector<int> texCounts;
   std::vector<TexUse> useVec;
   ArrayList insns;

   fn->orderInstructions(insns);

   texCounts.resize(fn->allBBlocks.getSize(), 0);
   bbFirstTex.resize(fn->allBBlocks.getSize(), insns.getSize());
   bbFirstUse.resize(fn->allBBlocks.getSize(), insns.getSize());

   // tag BB CFG nodes by their id for later
   for (ArrayList::Iterator i = fn->allBBlocks.iterator(); !i.end(); i.next()) {
      BasicBlock *bb = reinterpret_cast<BasicBlock *>(i.get());
      if (bb)
         bb->cfg.tag = bb->getId();
   }

   // gather the first uses for each TEX
   for (int i = 0; i < insns.getSize(); ++i) {
      Instruction *tex = reinterpret_cast<Instruction *>(insns.get(i));
      if (isTextureOp(tex->op)) {
         texes.push_back(tex);
         if (!texCounts.at(tex->bb->getId()))
            bbFirstTex[tex->bb->getId()] = texes.size() - 1;
         texCounts[tex->bb->getId()]++;
      }
   }
   insns.clear();
   if (texes.empty())
      return false;
   uses = new std::list<TexUse>[texes.size()];
   if (!uses)
      return false;
   for (size_t i = 0; i < texes.size(); ++i) {
      findFirstUses(texes[i], uses[i]);
   }

   // determine the barrier level at each use
   for (size_t i = 0; i < texes.size(); ++i) {
      for (std::list<TexUse>::iterator u = uses[i].begin(); u != uses[i].end();
           ++u) {
         BasicBlock *tb = texes[i]->bb;
         BasicBlock *ub = u->insn->bb;
         if (tb == ub) {
            u->level = 0;
            for (size_t j = i + 1; j < texes.size() &&
                    texes[j]->bb == tb && texes[j]->serial < u->insn->serial;
                 ++j)
               u->level++;
         } else {
            u->level = fn->cfg.findLightestPathWeight(&tb->cfg,
                                                      &ub->cfg, texCounts);
            if (u->level < 0) {
               WARN("Failed to find path TEX -> TEXBAR\n");
               u->level = 0;
               continue;
            }
            // this counted all TEXes in the origin block, correct that
            u->level -= i - bbFirstTex.at(tb->getId()) + 1 /* this TEX */;
            // and did not count the TEXes in the destination block, add those
            for (size_t j = bbFirstTex.at(ub->getId()); j < texes.size() &&
                    texes[j]->bb == ub && texes[j]->serial < u->insn->serial;
                 ++j)
               u->level++;
         }
         assert(u->level >= 0);
         useVec.push_back(*u);
      }
   }
   delete[] uses;

   // insert the barriers
   for (size_t i = 0; i < useVec.size(); ++i) {
      Instruction *prev = useVec[i].insn->prev;
      if (useVec[i].level < 0)
         continue;
      if (prev && prev->op == OP_TEXBAR) {
         if (prev->subOp > useVec[i].level)
            prev->subOp = useVec[i].level;
         prev->setSrc(prev->srcCount(), useVec[i].tex->getDef(0));
      } else {
         Instruction *bar = new_Instruction(func, OP_TEXBAR, TYPE_NONE);
         bar->fixed = 1;
         bar->subOp = useVec[i].level;
         // make use explicit to ease latency calculation
         bar->setSrc(bar->srcCount(), useVec[i].tex->getDef(0));
         useVec[i].insn->bb->insertBefore(useVec[i].insn, bar);
      }
   }

   if (fn->getProgram()->optLevel < 3)
      return true;

   std::vector<Limits> limitT, limitB, limitS; // entry, exit, single

   limitT.resize(fn->allBBlocks.getSize(), Limits(0, 0));
   limitB.resize(fn->allBBlocks.getSize(), Limits(0, 0));
   limitS.resize(fn->allBBlocks.getSize());

   // cull unneeded barriers (should do that earlier, but for simplicity)
   IteratorRef bi = fn->cfg.iteratorCFG();
   // first calculate min/max outstanding TEXes for each BB
   for (bi->reset(); !bi->end(); bi->next()) {
      Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
      BasicBlock *bb = BasicBlock::get(n);
      int min = 0;
      int max = std::numeric_limits<int>::max();
      for (Instruction *i = bb->getFirst(); i; i = i->next) {
         if (isTextureOp(i->op)) {
            min++;
            if (max < std::numeric_limits<int>::max())
               max++;
         } else
         if (i->op == OP_TEXBAR) {
            min = MIN2(min, i->subOp);
            max = MIN2(max, i->subOp);
         }
      }
      // limits when looking at an isolated block
      limitS[bb->getId()].min = min;
      limitS[bb->getId()].max = max;
   }
   // propagate the min/max values
   for (unsigned int l = 0; l <= fn->loopNestingBound; ++l) {
      for (bi->reset(); !bi->end(); bi->next()) {
         Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
         BasicBlock *bb = BasicBlock::get(n);
         const int bbId = bb->getId();
         for (Graph::EdgeIterator ei = n->incident(); !ei.end(); ei.next()) {
            BasicBlock *in = BasicBlock::get(ei.getNode());
            const int inId = in->getId();
            limitT[bbId].min = MAX2(limitT[bbId].min, limitB[inId].min);
            limitT[bbId].max = MAX2(limitT[bbId].max, limitB[inId].max);
         }
         // I just hope this is correct ...
         if (limitS[bbId].max == std::numeric_limits<int>::max()) {
            // no barrier
            limitB[bbId].min = limitT[bbId].min + limitS[bbId].min;
            limitB[bbId].max = limitT[bbId].max + limitS[bbId].min;
         } else {
            // block contained a barrier
            limitB[bbId].min = MIN2(limitS[bbId].max,
                                    limitT[bbId].min + limitS[bbId].min);
            limitB[bbId].max = MIN2(limitS[bbId].max,
                                    limitT[bbId].max + limitS[bbId].min);
         }
      }
   }
   // finally delete unnecessary barriers
   for (bi->reset(); !bi->end(); bi->next()) {
      Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get());
      BasicBlock *bb = BasicBlock::get(n);
      Instruction *prev = NULL;
      Instruction *next;
      int max = limitT[bb->getId()].max;
      for (Instruction *i = bb->getFirst(); i; i = next) {
         next = i->next;
         if (i->op == OP_TEXBAR) {
            if (i->subOp >= max) {
               delete_Instruction(prog, i);
               i = NULL;
            } else {
               max = i->subOp;
               if (prev && prev->op == OP_TEXBAR && prev->subOp >= max) {
                  delete_Instruction(prog, prev);
                  prev = NULL;
               }
            }
         } else
         if (isTextureOp(i->op)) {
            max++;
         }
         if (i && !i->isNop())
            prev = i;
      }
   }
   return true;
}

bool
NVC0LegalizePostRA::visit(Function *fn)
{
   if (needTexBar)
      insertTextureBarriers(fn);

   rZero = new_LValue(fn, FILE_GPR);
   pOne = new_LValue(fn, FILE_PREDICATE);
   carry = new_LValue(fn, FILE_FLAGS);

   rZero->reg.data.id = (prog->getTarget()->getChipset() >= NVISA_GK20A_CHIPSET) ? 255 : 63;
   carry->reg.data.id = 0;
   pOne->reg.data.id = 7;

   return true;
}

void
NVC0LegalizePostRA::replaceZero(Instruction *i)
{
   for (int s = 0; i->srcExists(s); ++s) {
      if (s == 2 && i->op == OP_SUCLAMP)
         continue;
      ImmediateValue *imm = i->getSrc(s)->asImm();
      if (imm) {
         if (i->op == OP_SELP && s == 2) {
            i->setSrc(s, pOne);
            if (imm->reg.data.u64 == 0)
               i->src(s).mod = i->src(s).mod ^ Modifier(NV50_IR_MOD_NOT);
         } else if (imm->reg.data.u64 == 0) {
            i->setSrc(s, rZero);
         }
      }
   }
}

// replace CONT with BRA for single unconditional continue
bool
NVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock *bb)
{
   if (bb->cfg.incidentCount() != 2 || bb->getEntry()->op != OP_PRECONT)
      return false;
   Graph::EdgeIterator ei = bb->cfg.incident();
   if (ei.getType() != Graph::Edge::BACK)
      ei.next();
   if (ei.getType() != Graph::Edge::BACK)
      return false;
   BasicBlock *contBB = BasicBlock::get(ei.getNode());

   if (!contBB->getExit() || contBB->getExit()->op != OP_CONT ||
       contBB->getExit()->getPredicate())
      return false;
   contBB->getExit()->op = OP_BRA;
   bb->remove(bb->getEntry()); // delete PRECONT

   ei.next();
   assert(ei.end() || ei.getType() != Graph::Edge::BACK);
   return true;
}

// replace branches to join blocks with join ops
void
NVC0LegalizePostRA::propagateJoin(BasicBlock *bb)
{
   if (bb->getEntry()->op != OP_JOIN || bb->getEntry()->asFlow()->limit)
      return;
   for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
      BasicBlock *in = BasicBlock::get(ei.getNode());
      Instruction *exit = in->getExit();
      if (!exit) {
         in->insertTail(new FlowInstruction(func, OP_JOIN, bb));
         // there should always be a terminator instruction
         WARN("inserted missing terminator in BB:%i\n", in->getId());
      } else
      if (exit->op == OP_BRA) {
         exit->op = OP_JOIN;
         exit->asFlow()->limit = 1; // must-not-propagate marker
      }
   }
   bb->remove(bb->getEntry());
}

bool
NVC0LegalizePostRA::visit(BasicBlock *bb)
{
   Instruction *i, *next;

   // remove pseudo operations and non-fixed no-ops, split 64 bit operations
   for (i = bb->getFirst(); i; i = next) {
      next = i->next;
      if (i->op == OP_EMIT || i->op == OP_RESTART) {
         if (!i->getDef(0)->refCount())
            i->setDef(0, NULL);
         if (i->src(0).getFile() == FILE_IMMEDIATE)
            i->setSrc(0, rZero); // initial value must be 0
         replaceZero(i);
      } else
      if (i->isNop()) {
         bb->remove(i);
      } else
      if (i->op == OP_BAR && i->subOp == NV50_IR_SUBOP_BAR_SYNC &&
          prog->getType() != Program::TYPE_COMPUTE) {
         // It seems like barriers are never required for tessellation since
         // the warp size is 32, and there are always at most 32 tcs threads.
         bb->remove(i);
      } else
      if (i->op == OP_LOAD && i->subOp == NV50_IR_SUBOP_LDC_IS) {
         int offset = i->src(0).get()->reg.data.offset;
         if (abs(offset) > 0x10000)
            i->src(0).get()->reg.fileIndex += offset >> 16;
         i->src(0).get()->reg.data.offset = (int)(short)offset;
      } else {
         // TODO: Move this to before register allocation for operations that
         // need the $c register !
         if (typeSizeof(i->dType) == 8) {
            Instruction *hi;
            hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry);
            if (hi)
               next = hi;
         }

         if (i->op != OP_MOV && i->op != OP_PFETCH)
            replaceZero(i);
      }
   }
   if (!bb->getEntry())
      return true;

   if (!tryReplaceContWithBra(bb))
      propagateJoin(bb);

   return true;
}

NVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget())
{
   bld.setProgram(prog);
}

bool
NVC0LoweringPass::visit(Function *fn)
{
   if (prog->getType() == Program::TYPE_GEOMETRY) {
      assert(!strncmp(fn->getName(), "MAIN", 4));
      // TODO: when we generate actual functions pass this value along somehow
      bld.setPosition(BasicBlock::get(fn->cfg.getRoot()), false);
      gpEmitAddress = bld.loadImm(NULL, 0)->asLValue();
      if (fn->cfgExit) {
         bld.setPosition(BasicBlock::get(fn->cfgExit)->getExit(), false);
         bld.mkMovToReg(0, gpEmitAddress);
      }
   }
   return true;
}

bool
NVC0LoweringPass::visit(BasicBlock *bb)
{
   return true;
}

inline Value *
NVC0LoweringPass::loadTexHandle(Value *ptr, unsigned int slot)
{
   uint8_t b = prog->driver->io.auxCBSlot;
   uint32_t off = prog->driver->io.texBindBase + slot * 4;

   if (ptr)
      ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(2));

   return bld.
      mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
}

// move array source to first slot, convert to u16, add indirections
bool
NVC0LoweringPass::handleTEX(TexInstruction *i)
{
   const int dim = i->tex.target.getDim() + i->tex.target.isCube();
   const int arg = i->tex.target.getArgCount();
   const int lyr = arg - (i->tex.target.isMS() ? 2 : 1);
   const int chipset = prog->getTarget()->getChipset();

   /* Only normalize in the non-explicit derivatives case. For explicit
    * derivatives, this is handled in handleManualTXD.
    */
   if (i->tex.target.isCube() && i->dPdx[0].get() == NULL) {
      Value *src[3], *val;
      int c;
      for (c = 0; c < 3; ++c)
         src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), i->getSrc(c));
      val = bld.getScratch();
      bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
      bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
      bld.mkOp1(OP_RCP, TYPE_F32, val, val);
      for (c = 0; c < 3; ++c) {
         i->setSrc(c, bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(),
                                 i->getSrc(c), val));
      }
   }

   // Arguments to the TEX instruction are a little insane. Even though the
   // encoding is identical between SM20 and SM30, the arguments mean
   // different things between Fermi and Kepler+. A lot of arguments are
   // optional based on flags passed to the instruction. This summarizes the
   // order of things.
   //
   // Fermi:
   //  array/indirect
   //  coords
   //  sample
   //  lod bias
   //  depth compare
   //  offsets:
   //    - tg4: 8 bits each, either 2 (1 offset reg) or 8 (2 offset reg)
   //    - other: 4 bits each, single reg
   //
   // Kepler+:
   //  indirect handle
   //  array (+ offsets for txd in upper 16 bits)
   //  coords
   //  sample
   //  lod bias
   //  depth compare
   //  offsets (same as fermi, except txd which takes it with array)
   //
   // Maxwell (tex):
   //  array
   //  coords
   //  indirect handle
   //  sample
   //  lod bias
   //  depth compare
   //  offsets
   //
   // Maxwell (txd):
   //  indirect handle
   //  coords
   //  array + offsets
   //  derivatives

   if (chipset >= NVISA_GK104_CHIPSET) {
      if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
         // XXX this ignores tsc, and assumes a 1:1 mapping
         assert(i->tex.rIndirectSrc >= 0);
         Value *hnd = loadTexHandle(i->getIndirectR(), i->tex.r);
         i->tex.r = 0xff;
         i->tex.s = 0x1f;
         i->setIndirectR(hnd);
         i->setIndirectS(NULL);
      } else if (i->tex.r == i->tex.s || i->op == OP_TXF) {
         i->tex.r += prog->driver->io.texBindBase / 4;
         i->tex.s  = 0; // only a single cX[] value possible here
      } else {
         Value *hnd = bld.getScratch();
         Value *rHnd = loadTexHandle(NULL, i->tex.r);
         Value *sHnd = loadTexHandle(NULL, i->tex.s);

         bld.mkOp3(OP_INSBF, TYPE_U32, hnd, rHnd, bld.mkImm(0x1400), sHnd);

         i->tex.r = 0; // not used for indirect tex
         i->tex.s = 0;
         i->setIndirectR(hnd);
      }
      if (i->tex.target.isArray()) {
         LValue *layer = new_LValue(func, FILE_GPR);
         Value *src = i->getSrc(lyr);
         const int sat = (i->op == OP_TXF) ? 1 : 0;
         DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
         bld.mkCvt(OP_CVT, TYPE_U16, layer, sTy, src)->saturate = sat;
         if (i->op != OP_TXD || chipset < NVISA_GM107_CHIPSET) {
            for (int s = dim; s >= 1; --s)
               i->setSrc(s, i->getSrc(s - 1));
            i->setSrc(0, layer);
         } else {
            i->setSrc(dim, layer);
         }
      }
      // Move the indirect reference to the first place
      if (i->tex.rIndirectSrc >= 0 && (
                i->op == OP_TXD || chipset < NVISA_GM107_CHIPSET)) {
         Value *hnd = i->getIndirectR();

         i->setIndirectR(NULL);
         i->moveSources(0, 1);
         i->setSrc(0, hnd);
         i->tex.rIndirectSrc = 0;
         i->tex.sIndirectSrc = -1;
      }
      // Move the indirect reference to right after the coords
      else if (i->tex.rIndirectSrc >= 0 && chipset >= NVISA_GM107_CHIPSET) {
         Value *hnd = i->getIndirectR();

         i->setIndirectR(NULL);
         i->moveSources(arg, 1);
         i->setSrc(arg, hnd);
         i->tex.rIndirectSrc = 0;
         i->tex.sIndirectSrc = -1;
      }
   } else
   // (nvc0) generate and move the tsc/tic/array source to the front
   if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) {
      LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa

      Value *ticRel = i->getIndirectR();
      Value *tscRel = i->getIndirectS();

      if (ticRel) {
         i->setSrc(i->tex.rIndirectSrc, NULL);
         if (i->tex.r)
            ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
                                ticRel, bld.mkImm(i->tex.r));
      }
      if (tscRel) {
         i->setSrc(i->tex.sIndirectSrc, NULL);
         if (i->tex.s)
            tscRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
                                tscRel, bld.mkImm(i->tex.s));
      }

      Value *arrayIndex = i->tex.target.isArray() ? i->getSrc(lyr) : NULL;
      if (arrayIndex) {
         for (int s = dim; s >= 1; --s)
            i->setSrc(s, i->getSrc(s - 1));
         i->setSrc(0, arrayIndex);
      } else {
         i->moveSources(0, 1);
      }

      if (arrayIndex) {
         int sat = (i->op == OP_TXF) ? 1 : 0;
         DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32;
         bld.mkCvt(OP_CVT, TYPE_U16, src, sTy, arrayIndex)->saturate = sat;
      } else {
         bld.loadImm(src, 0);
      }

      if (ticRel)
         bld.mkOp3(OP_INSBF, TYPE_U32, src, ticRel, bld.mkImm(0x0917), src);
      if (tscRel)
         bld.mkOp3(OP_INSBF, TYPE_U32, src, tscRel, bld.mkImm(0x0710), src);

      i->setSrc(0, src);
   }

   // For nvc0, the sample id has to be in the second operand, as the offset
   // does. Right now we don't know how to pass both in, and this case can't
   // happen with OpenGL. On nve0, the sample id is part of the texture
   // coordinate argument.
   assert(chipset >= NVISA_GK104_CHIPSET ||
          !i->tex.useOffsets || !i->tex.target.isMS());

   // offset is between lod and dc
   if (i->tex.useOffsets) {
      int n, c;
      int s = i->srcCount(0xff, true);
      if (i->op != OP_TXD || chipset < NVISA_GK104_CHIPSET) {
         if (i->tex.target.isShadow())
            s--;
         if (i->srcExists(s)) // move potential predicate out of the way
            i->moveSources(s, 1);
         if (i->tex.useOffsets == 4 && i->srcExists(s + 1))
            i->moveSources(s + 1, 1);
      }
      if (i->op == OP_TXG) {
         // Either there is 1 offset, which goes into the 2 low bytes of the
         // first source, or there are 4 offsets, which go into 2 sources (8
         // values, 1 byte each).
         Value *offs[2] = {NULL, NULL};
         for (n = 0; n < i->tex.useOffsets; n++) {
            for (c = 0; c < 2; ++c) {
               if ((n % 2) == 0 && c == 0)
                  bld.mkMov(offs[n / 2] = bld.getScratch(), i->offset[n][c].get());
               else
                  bld.mkOp3(OP_INSBF, TYPE_U32,
                            offs[n / 2],
                            i->offset[n][c].get(),
                            bld.mkImm(0x800 | ((n * 16 + c * 8) % 32)),
                            offs[n / 2]);
            }
         }
         i->setSrc(s, offs[0]);
         if (offs[1])
            i->setSrc(s + 1, offs[1]);
      } else {
         unsigned imm = 0;
         assert(i->tex.useOffsets == 1);
         for (c = 0; c < 3; ++c) {
            ImmediateValue val;
            if (!i->offset[0][c].getImmediate(val))
               assert(!"non-immediate offset passed to non-TXG");
            imm |= (val.reg.data.u32 & 0xf) << (c * 4);
         }
         if (i->op == OP_TXD && chipset >= NVISA_GK104_CHIPSET) {
            // The offset goes into the upper 16 bits of the array index. So
            // create it if it's not already there, and INSBF it if it already
            // is.
            s = (i->tex.rIndirectSrc >= 0) ? 1 : 0;
            if (chipset >= NVISA_GM107_CHIPSET)
               s += dim;
            if (i->tex.target.isArray()) {
               bld.mkOp3(OP_INSBF, TYPE_U32, i->getSrc(s),
                         bld.loadImm(NULL, imm), bld.mkImm(0xc10),
                         i->getSrc(s));
            } else {
               i->moveSources(s, 1);
               i->setSrc(s, bld.loadImm(NULL, imm << 16));
            }
         } else {
            i->setSrc(s, bld.loadImm(NULL, imm));
         }
      }
   }

   if (chipset >= NVISA_GK104_CHIPSET) {
      //
      // If TEX requires more than 4 sources, the 2nd register tuple must be
      // aligned to 4, even if it consists of just a single 4-byte register.
      //
      // XXX HACK: We insert 0 sources to avoid the 5 or 6 regs case.
      //
      int s = i->srcCount(0xff, true);
      if (s > 4 && s < 7) {
         if (i->srcExists(s)) // move potential predicate out of the way
            i->moveSources(s, 7 - s);
         while (s < 7)
            i->setSrc(s++, bld.loadImm(NULL, 0));
      }
   }

   return true;
}

bool
NVC0LoweringPass::handleManualTXD(TexInstruction *i)
{
   static const uint8_t qOps[4][2] =
   {
      { QUADOP(MOV2, ADD,  MOV2, ADD),  QUADOP(MOV2, MOV2, ADD,  ADD) }, // l0
      { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD,  ADD) }, // l1
      { QUADOP(MOV2, ADD,  MOV2, ADD),  QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2
      { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3
   };
   Value *def[4][4];
   Value *crd[3];
   Instruction *tex;
   Value *zero = bld.loadImm(bld.getSSA(), 0);
   int l, c;
   const int dim = i->tex.target.getDim() + i->tex.target.isCube();

   // This function is invoked after handleTEX lowering, so we have to expect
   // the arguments in the order that the hw wants them. For Fermi, array and
   // indirect are both in the leading arg, while for Kepler, array and
   // indirect are separate (and both precede the coordinates). Maxwell is
   // handled in a separate function.
   unsigned array;
   if (targ->getChipset() < NVISA_GK104_CHIPSET)
      array = i->tex.target.isArray() || i->tex.rIndirectSrc >= 0;
   else
      array = i->tex.target.isArray() + (i->tex.rIndirectSrc >= 0);

   i->op = OP_TEX; // no need to clone dPdx/dPdy later

   for (c = 0; c < dim; ++c)
      crd[c] = bld.getScratch();

   bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
   for (l = 0; l < 4; ++l) {
      Value *src[3], *val;
      // mov coordinates from lane l to all lanes
      for (c = 0; c < dim; ++c)
         bld.mkQuadop(0x00, crd[c], l, i->getSrc(c + array), zero);
      // add dPdx from lane l to lanes dx
      for (c = 0; c < dim; ++c)
         bld.mkQuadop(qOps[l][0], crd[c], l, i->dPdx[c].get(), crd[c]);
      // add dPdy from lane l to lanes dy
      for (c = 0; c < dim; ++c)
         bld.mkQuadop(qOps[l][1], crd[c], l, i->dPdy[c].get(), crd[c]);
      // normalize cube coordinates
      if (i->tex.target.isCube()) {
         for (c = 0; c < 3; ++c)
            src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), crd[c]);
         val = bld.getScratch();
         bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]);
         bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val);
         bld.mkOp1(OP_RCP, TYPE_F32, val, val);
         for (c = 0; c < 3; ++c)
            src[c] = bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(), crd[c], val);
      } else {
         for (c = 0; c < dim; ++c)
            src[c] = crd[c];
      }
      // texture
      bld.insert(tex = cloneForward(func, i));
      for (c = 0; c < dim; ++c)
         tex->setSrc(c + array, src[c]);
      // save results
      for (c = 0; i->defExists(c); ++c) {
         Instruction *mov;
         def[c][l] = bld.getSSA();
         mov = bld.mkMov(def[c][l], tex->getDef(c));
         mov->fixed = 1;
         mov->lanes = 1 << l;
      }
   }
   bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);

   for (c = 0; i->defExists(c); ++c) {
      Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c));
      for (l = 0; l < 4; ++l)
         u->setSrc(l, def[c][l]);
   }

   i->bb->remove(i);
   return true;
}

bool
NVC0LoweringPass::handleTXD(TexInstruction *txd)
{
   int dim = txd->tex.target.getDim() + txd->tex.target.isCube();
   unsigned arg = txd->tex.target.getArgCount();
   unsigned expected_args = arg;
   const int chipset = prog->getTarget()->getChipset();

   if (chipset >= NVISA_GK104_CHIPSET) {
      if (!txd->tex.target.isArray() && txd->tex.useOffsets)
         expected_args++;
      if (txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0)
         expected_args++;
   } else {
      if (txd->tex.useOffsets)
         expected_args++;
      if (!txd->tex.target.isArray() && (
                txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0))
         expected_args++;
   }

   if (expected_args > 4 ||
       dim > 2 ||
       txd->tex.target.isShadow())
      txd->op = OP_TEX;

   handleTEX(txd);
   while (txd->srcExists(arg))
      ++arg;

   txd->tex.derivAll = true;
   if (txd->op == OP_TEX)
      return handleManualTXD(txd);

   assert(arg == expected_args);
   for (int c = 0; c < dim; ++c) {
      txd->setSrc(arg + c * 2 + 0, txd->dPdx[c]);
      txd->setSrc(arg + c * 2 + 1, txd->dPdy[c]);
      txd->dPdx[c].set(NULL);
      txd->dPdy[c].set(NULL);
   }

   // In this case we have fewer than 4 "real" arguments, which means that
   // handleTEX didn't apply any padding. However we have to make sure that
   // the second "group" of arguments still gets padded up to 4.
   if (chipset >= NVISA_GK104_CHIPSET) {
      int s = arg + 2 * dim;
      if (s >= 4 && s < 7) {
         if (txd->srcExists(s)) // move potential predicate out of the way
            txd->moveSources(s, 7 - s);
         while (s < 7)
            txd->setSrc(s++, bld.loadImm(NULL, 0));
      }
   }

   return true;
}

bool
NVC0LoweringPass::handleTXQ(TexInstruction *txq)
{
   const int chipset = prog->getTarget()->getChipset();
   if (chipset >= NVISA_GK104_CHIPSET && txq->tex.rIndirectSrc < 0)
      txq->tex.r += prog->driver->io.texBindBase / 4;

   if (txq->tex.rIndirectSrc < 0)
      return true;

   Value *ticRel = txq->getIndirectR();

   txq->setIndirectS(NULL);
   txq->tex.sIndirectSrc = -1;

   assert(ticRel);

   if (chipset < NVISA_GK104_CHIPSET) {
      LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa

      txq->setSrc(txq->tex.rIndirectSrc, NULL);
      if (txq->tex.r)
         ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(),
                             ticRel, bld.mkImm(txq->tex.r));

      bld.mkOp2(OP_SHL, TYPE_U32, src, ticRel, bld.mkImm(0x17));

      txq->moveSources(0, 1);
      txq->setSrc(0, src);
   } else {
      Value *hnd = loadTexHandle(txq->getIndirectR(), txq->tex.r);
      txq->tex.r = 0xff;
      txq->tex.s = 0x1f;

      txq->setIndirectR(NULL);
      txq->moveSources(0, 1);
      txq->setSrc(0, hnd);
      txq->tex.rIndirectSrc = 0;
   }

   return true;
}

bool
NVC0LoweringPass::handleTXLQ(TexInstruction *i)
{
   /* The outputs are inverted compared to what the TGSI instruction
    * expects. Take that into account in the mask.
    */
   assert((i->tex.mask & ~3) == 0);
   if (i->tex.mask == 1)
      i->tex.mask = 2;
   else if (i->tex.mask == 2)
      i->tex.mask = 1;
   handleTEX(i);
   bld.setPosition(i, true);

   /* The returned values are not quite what we want:
    * (a) convert from s16/u16 to f32
    * (b) multiply by 1/256
    */
   for (int def = 0; def < 2; ++def) {
      if (!i->defExists(def))
         continue;
      enum DataType type = TYPE_S16;
      if (i->tex.mask == 2 || def > 0)
         type = TYPE_U16;
      bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(def), type, i->getDef(def));
      bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(def),
                i->getDef(def), bld.loadImm(NULL, 1.0f / 256));
   }
   if (i->tex.mask == 3) {
      LValue *t = new_LValue(func, FILE_GPR);
      bld.mkMov(t, i->getDef(0));
      bld.mkMov(i->getDef(0), i->getDef(1));
      bld.mkMov(i->getDef(1), t);
   }
   return true;
}

bool
NVC0LoweringPass::handleBUFQ(Instruction *bufq)
{
   bufq->op = OP_MOV;
   bufq->setSrc(0, loadBufLength32(bufq->getIndirect(0, 1),
                                   bufq->getSrc(0)->reg.fileIndex * 16));
   bufq->setIndirect(0, 0, NULL);
   bufq->setIndirect(0, 1, NULL);
   return true;
}

void
NVC0LoweringPass::handleSharedATOMNVE4(Instruction *atom)
{
   assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);

   BasicBlock *currBB = atom->bb;
   BasicBlock *tryLockBB = atom->bb->splitBefore(atom, false);
   BasicBlock *joinBB = atom->bb->splitAfter(atom);
   BasicBlock *setAndUnlockBB = new BasicBlock(func);
   BasicBlock *failLockBB = new BasicBlock(func);

   bld.setPosition(currBB, true);
   assert(!currBB->joinAt);
   currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);

   CmpInstruction *pred =
      bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
                TYPE_U32, bld.mkImm(0), bld.mkImm(1));

   bld.mkFlow(OP_BRA, tryLockBB, CC_ALWAYS, NULL);
   currBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::TREE);

   bld.setPosition(tryLockBB, true);

   Instruction *ld =
      bld.mkLoad(TYPE_U32, atom->getDef(0), atom->getSrc(0)->asSym(),
                 atom->getIndirect(0, 0));
   ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
   ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;

   bld.mkFlow(OP_BRA, setAndUnlockBB, CC_P, ld->getDef(1));
   bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
   tryLockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::CROSS);
   tryLockBB->cfg.attach(&setAndUnlockBB->cfg, Graph::Edge::TREE);

   tryLockBB->cfg.detach(&joinBB->cfg);
   bld.remove(atom);

   bld.setPosition(setAndUnlockBB, true);
   Value *stVal;
   if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
      // Read the old value, and write the new one.
      stVal = atom->getSrc(1);
   } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
      CmpInstruction *set =
         bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(),
                   TYPE_U32, ld->getDef(0), atom->getSrc(1));

      bld.mkCmp(OP_SLCT, CC_NE, TYPE_U32, (stVal = bld.getSSA()),
                TYPE_U32, atom->getSrc(2), ld->getDef(0), set->getDef(0));
   } else {
      operation op;

      switch (atom->subOp) {
      case NV50_IR_SUBOP_ATOM_ADD:
         op = OP_ADD;
         break;
      case NV50_IR_SUBOP_ATOM_AND:
         op = OP_AND;
         break;
      case NV50_IR_SUBOP_ATOM_OR:
         op = OP_OR;
         break;
      case NV50_IR_SUBOP_ATOM_XOR:
         op = OP_XOR;
         break;
      case NV50_IR_SUBOP_ATOM_MIN:
         op = OP_MIN;
         break;
      case NV50_IR_SUBOP_ATOM_MAX:
         op = OP_MAX;
         break;
      default:
         assert(0);
         return;
      }

      stVal = bld.mkOp2v(op, atom->dType, bld.getSSA(), ld->getDef(0),
                         atom->getSrc(1));
   }

   Instruction *st =
      bld.mkStore(OP_STORE, TYPE_U32, atom->getSrc(0)->asSym(),
                  atom->getIndirect(0, 0), stVal);
   st->setDef(0, pred->getDef(0));
   st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;

   bld.mkFlow(OP_BRA, failLockBB, CC_ALWAYS, NULL);
   setAndUnlockBB->cfg.attach(&failLockBB->cfg, Graph::Edge::TREE);

   // Lock until the store has not been performed.
   bld.setPosition(failLockBB, true);
   bld.mkFlow(OP_BRA, tryLockBB, CC_NOT_P, pred->getDef(0));
   bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
   failLockBB->cfg.attach(&tryLockBB->cfg, Graph::Edge::BACK);
   failLockBB->cfg.attach(&joinBB->cfg, Graph::Edge::TREE);

   bld.setPosition(joinBB, false);
   bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
}

void
NVC0LoweringPass::handleSharedATOM(Instruction *atom)
{
   assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);

   BasicBlock *currBB = atom->bb;
   BasicBlock *tryLockAndSetBB = atom->bb->splitBefore(atom, false);
   BasicBlock *joinBB = atom->bb->splitAfter(atom);

   bld.setPosition(currBB, true);
   assert(!currBB->joinAt);
   currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);

   bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_ALWAYS, NULL);
   currBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::TREE);

   bld.setPosition(tryLockAndSetBB, true);

   Instruction *ld =
      bld.mkLoad(TYPE_U32, atom->getDef(0), atom->getSrc(0)->asSym(),
                 atom->getIndirect(0, 0));
   ld->setDef(1, bld.getSSA(1, FILE_PREDICATE));
   ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED;

   Value *stVal;
   if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) {
      // Read the old value, and write the new one.
      stVal = atom->getSrc(1);
   } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) {
      CmpInstruction *set =
         bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
                   TYPE_U32, ld->getDef(0), atom->getSrc(1));
      set->setPredicate(CC_P, ld->getDef(1));

      Instruction *selp =
         bld.mkOp3(OP_SELP, TYPE_U32, bld.getSSA(), ld->getDef(0),
                   atom->getSrc(2), set->getDef(0));
      selp->src(2).mod = Modifier(NV50_IR_MOD_NOT);
      selp->setPredicate(CC_P, ld->getDef(1));

      stVal = selp->getDef(0);
   } else {
      operation op;

      switch (atom->subOp) {
      case NV50_IR_SUBOP_ATOM_ADD:
         op = OP_ADD;
         break;
      case NV50_IR_SUBOP_ATOM_AND:
         op = OP_AND;
         break;
      case NV50_IR_SUBOP_ATOM_OR:
         op = OP_OR;
         break;
      case NV50_IR_SUBOP_ATOM_XOR:
         op = OP_XOR;
         break;
      case NV50_IR_SUBOP_ATOM_MIN:
         op = OP_MIN;
         break;
      case NV50_IR_SUBOP_ATOM_MAX:
         op = OP_MAX;
         break;
      default:
         assert(0);
         return;
      }

      Instruction *i =
         bld.mkOp2(op, atom->dType, bld.getSSA(), ld->getDef(0),
                   atom->getSrc(1));
      i->setPredicate(CC_P, ld->getDef(1));

      stVal = i->getDef(0);
   }

   Instruction *st =
      bld.mkStore(OP_STORE, TYPE_U32, atom->getSrc(0)->asSym(),
                  atom->getIndirect(0, 0), stVal);
   st->setPredicate(CC_P, ld->getDef(1));
   st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED;

   // Loop until the lock is acquired.
   bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_NOT_P, ld->getDef(1));
   tryLockAndSetBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::BACK);
   tryLockAndSetBB->cfg.attach(&joinBB->cfg, Graph::Edge::CROSS);
   bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);

   bld.remove(atom);

   bld.setPosition(joinBB, false);
   bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
}

bool
NVC0LoweringPass::handleATOM(Instruction *atom)
{
   SVSemantic sv;
   Value *ptr = atom->getIndirect(0, 0), *ind = atom->getIndirect(0, 1), *base;

   switch (atom->src(0).getFile()) {
   case FILE_MEMORY_LOCAL:
      sv = SV_LBASE;
      break;
   case FILE_MEMORY_SHARED:
      // For Fermi/Kepler, we have to use ld lock/st unlock to perform atomic
      // operations on shared memory. For Maxwell, ATOMS is enough.
      if (targ->getChipset() < NVISA_GK104_CHIPSET)
         handleSharedATOM(atom);
      else if (targ->getChipset() < NVISA_GM107_CHIPSET)
         handleSharedATOMNVE4(atom);
      return true;
   default:
      assert(atom->src(0).getFile() == FILE_MEMORY_BUFFER);
      base = loadBufInfo64(ind, atom->getSrc(0)->reg.fileIndex * 16);
      assert(base->reg.size == 8);
      if (ptr)
         base = bld.mkOp2v(OP_ADD, TYPE_U64, base, base, ptr);
      assert(base->reg.size == 8);
      atom->setIndirect(0, 0, base);
      atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;

      // Harden against out-of-bounds accesses
      Value *offset = bld.loadImm(NULL, atom->getSrc(0)->reg.data.offset + typeSizeof(atom->sType));
      Value *length = loadBufLength32(ind, atom->getSrc(0)->reg.fileIndex * 16);
      Value *pred = new_LValue(func, FILE_PREDICATE);
      if (ptr)
         bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, ptr);
      bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
      atom->setPredicate(CC_NOT_P, pred);
      if (atom->defExists(0)) {
         Value *zero, *dst = atom->getDef(0);
         atom->setDef(0, bld.getSSA());

         bld.setPosition(atom, true);
         bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
            ->setPredicate(CC_P, pred);
         bld.mkOp2(OP_UNION, TYPE_U32, dst, atom->getDef(0), zero);
      }

      return true;
   }
   base =
      bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(), bld.mkSysVal(sv, 0));

   atom->setSrc(0, cloneShallow(func, atom->getSrc(0)));
   atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
   if (ptr)
      base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr);
   atom->setIndirect(0, 1, NULL);
   atom->setIndirect(0, 0, base);

   return true;
}

bool
NVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl)
{
   if (targ->getChipset() < NVISA_GM107_CHIPSET) {
      if (cas->src(0).getFile() == FILE_MEMORY_SHARED) {
         // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM().
         return false;
      }
   }

   if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS &&
       cas->subOp != NV50_IR_SUBOP_ATOM_EXCH)
      return false;
   bld.setPosition(cas, true);

   if (needCctl) {
      Instruction *cctl = bld.mkOp1(OP_CCTL, TYPE_NONE, NULL, cas->getSrc(0));
      cctl->setIndirect(0, 0, cas->getIndirect(0, 0));
      cctl->fixed = 1;
      cctl->subOp = NV50_IR_SUBOP_CCTL_IV;
      if (cas->isPredicated())
         cctl->setPredicate(cas->cc, cas->getPredicate());
   }

   if (cas->subOp == NV50_IR_SUBOP_ATOM_CAS) {
      // CAS is crazy. It's 2nd source is a double reg, and the 3rd source
      // should be set to the high part of the double reg or bad things will
      // happen elsewhere in the universe.
      // Also, it sometimes returns the new value instead of the old one
      // under mysterious circumstances.
      Value *dreg = bld.getSSA(8);
      bld.setPosition(cas, false);
      bld.mkOp2(OP_MERGE, TYPE_U64, dreg, cas->getSrc(1), cas->getSrc(2));
      cas->setSrc(1, dreg);
      cas->setSrc(2, dreg);
   }

   return true;
}

inline Value *
NVC0LoweringPass::loadResInfo32(Value *ptr, uint32_t off, uint16_t base)
{
   uint8_t b = prog->driver->io.auxCBSlot;
   off += base;

   return bld.
      mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
}

inline Value *
NVC0LoweringPass::loadResInfo64(Value *ptr, uint32_t off, uint16_t base)
{
   uint8_t b = prog->driver->io.auxCBSlot;
   off += base;

   if (ptr)
      ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));

   return bld.
      mkLoadv(TYPE_U64, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off), ptr);
}

inline Value *
NVC0LoweringPass::loadResLength32(Value *ptr, uint32_t off, uint16_t base)
{
   uint8_t b = prog->driver->io.auxCBSlot;
   off += base;

   if (ptr)
      ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));

   return bld.
      mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr);
}

inline Value *
NVC0LoweringPass::loadBufInfo64(Value *ptr, uint32_t off)
{
   return loadResInfo64(ptr, off, prog->driver->io.bufInfoBase);
}

inline Value *
NVC0LoweringPass::loadBufLength32(Value *ptr, uint32_t off)
{
   return loadResLength32(ptr, off, prog->driver->io.bufInfoBase);
}

inline Value *
NVC0LoweringPass::loadUboInfo64(Value *ptr, uint32_t off)
{
   return loadResInfo64(ptr, off, prog->driver->io.uboInfoBase);
}

inline Value *
NVC0LoweringPass::loadUboLength32(Value *ptr, uint32_t off)
{
   return loadResLength32(ptr, off, prog->driver->io.uboInfoBase);
}

inline Value *
NVC0LoweringPass::loadMsInfo32(Value *ptr, uint32_t off)
{
   uint8_t b = prog->driver->io.msInfoCBSlot;
   off += prog->driver->io.msInfoBase;
   return bld.
      mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr);
}

/* On nvc0, surface info is obtained via the surface binding points passed
 * to the SULD/SUST instructions.
 * On nve4, surface info is stored in c[] and is used by various special
 * instructions, e.g. for clamping coordinates or generating an address.
 * They couldn't just have added an equivalent to TIC now, couldn't they ?
 */
#define NVC0_SU_INFO_ADDR   0x00
#define NVC0_SU_INFO_FMT    0x04
#define NVC0_SU_INFO_DIM_X  0x08
#define NVC0_SU_INFO_PITCH  0x0c
#define NVC0_SU_INFO_DIM_Y  0x10
#define NVC0_SU_INFO_ARRAY  0x14
#define NVC0_SU_INFO_DIM_Z  0x18
#define NVC0_SU_INFO_UNK1C  0x1c
#define NVC0_SU_INFO_WIDTH  0x20
#define NVC0_SU_INFO_HEIGHT 0x24
#define NVC0_SU_INFO_DEPTH  0x28
#define NVC0_SU_INFO_TARGET 0x2c
#define NVC0_SU_INFO_BSIZE  0x30
#define NVC0_SU_INFO_RAW_X  0x34
#define NVC0_SU_INFO_MS_X   0x38
#define NVC0_SU_INFO_MS_Y   0x3c

#define NVC0_SU_INFO__STRIDE 0x40

#define NVC0_SU_INFO_DIM(i)  (0x08 + (i) * 8)
#define NVC0_SU_INFO_SIZE(i) (0x20 + (i) * 4)
#define NVC0_SU_INFO_MS(i)   (0x38 + (i) * 4)

inline Value *
NVC0LoweringPass::loadSuInfo32(Value *ptr, int slot, uint32_t off)
{
   uint32_t base = slot * NVC0_SU_INFO__STRIDE;

   if (ptr) {
      ptr = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(slot));
      ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(7));
      ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(6));
      base = 0;
   }
   off += base;

   return loadResInfo32(ptr, off, prog->driver->io.suInfoBase);
}

static inline uint16_t getSuClampSubOp(const TexInstruction *su, int c)
{
   switch (su->tex.target.getEnum()) {
   case TEX_TARGET_BUFFER:      return NV50_IR_SUBOP_SUCLAMP_PL(0, 1);
   case TEX_TARGET_RECT:        return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_1D:          return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_1D_ARRAY:    return (c == 1) ?
                                   NV50_IR_SUBOP_SUCLAMP_PL(0, 2) :
                                   NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_2D:          return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
   case TEX_TARGET_2D_MS:       return NV50_IR_SUBOP_SUCLAMP_BL(0, 2);
   case TEX_TARGET_2D_ARRAY:    return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_2D_MS_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_3D:          return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_CUBE:        return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   case TEX_TARGET_CUBE_ARRAY:  return NV50_IR_SUBOP_SUCLAMP_SD(0, 2);
   default:
      assert(0);
      return 0;
   }
}

bool
NVC0LoweringPass::handleSUQ(TexInstruction *suq)
{
   int mask = suq->tex.mask;
   int dim = suq->tex.target.getDim();
   int arg = dim + (suq->tex.target.isArray() || suq->tex.target.isCube());
   Value *ind = suq->getIndirectR();
   int slot = suq->tex.r;
   int c, d;

   for (c = 0, d = 0; c < 3; ++c, mask >>= 1) {
      if (c >= arg || !(mask & 1))
         continue;

      int offset;

      if (c == 1 && suq->tex.target == TEX_TARGET_1D_ARRAY) {
         offset = NVC0_SU_INFO_SIZE(2);
      } else {
         offset = NVC0_SU_INFO_SIZE(c);
      }
      bld.mkMov(suq->getDef(d++), loadSuInfo32(ind, slot, offset));
      if (c == 2 && suq->tex.target.isCube())
         bld.mkOp2(OP_DIV, TYPE_U32, suq->getDef(d - 1), suq->getDef(d - 1),
                   bld.loadImm(NULL, 6));
   }

   if (mask & 1) {
      if (suq->tex.target.isMS()) {
         Value *ms_x = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(0));
         Value *ms_y = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(1));
         Value *ms = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(), ms_x, ms_y);
         bld.mkOp2(OP_SHL, TYPE_U32, suq->getDef(d++), bld.loadImm(NULL, 1), ms);
      } else {
         bld.mkMov(suq->getDef(d++), bld.loadImm(NULL, 1));
      }
   }

   bld.remove(suq);
   return true;
}

void
NVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex)
{
   const int arg = tex->tex.target.getArgCount();
   int slot = tex->tex.r;

   if (tex->tex.target == TEX_TARGET_2D_MS)
      tex->tex.target = TEX_TARGET_2D;
   else
   if (tex->tex.target == TEX_TARGET_2D_MS_ARRAY)
      tex->tex.target = TEX_TARGET_2D_ARRAY;
   else
      return;

   Value *x = tex->getSrc(0);
   Value *y = tex->getSrc(1);
   Value *s = tex->getSrc(arg - 1);

   Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA();
   Value *ind = tex->getIndirectR();

   Value *ms_x = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(0));
   Value *ms_y = loadSuInfo32(ind, slot, NVC0_SU_INFO_MS(1));

   bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x);
   bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y);

   s = bld.mkOp2v(OP_AND, TYPE_U32, ts, s, bld.loadImm(NULL, 0x7));
   s = bld.mkOp2v(OP_SHL, TYPE_U32, ts, ts, bld.mkImm(3));

   Value *dx = loadMsInfo32(ts, 0x0);
   Value *dy = loadMsInfo32(ts, 0x4);

   bld.mkOp2(OP_ADD, TYPE_U32, tx, tx, dx);
   bld.mkOp2(OP_ADD, TYPE_U32, ty, ty, dy);

   tex->setSrc(0, tx);
   tex->setSrc(1, ty);
   tex->moveSources(arg, -1);
}

// Sets 64-bit "generic address", predicate and format sources for SULD/SUST.
// They're computed from the coordinates using the surface info in c[] space.
void
NVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction *su)
{
   Instruction *insn;
   const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP;
   const bool raw =
      su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB;
   const int slot = su->tex.r;
   const int dim = su->tex.target.getDim();
   const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
   int c;
   Value *zero = bld.mkImm(0);
   Value *p1 = NULL;
   Value *v;
   Value *src[3];
   Value *bf, *eau, *off;
   Value *addr, *pred;
   Value *ind = su->getIndirectR();

   off = bld.getScratch(4);
   bf = bld.getScratch(4);
   addr = bld.getSSA(8);
   pred = bld.getScratch(1, FILE_PREDICATE);

   bld.setPosition(su, false);

   adjustCoordinatesMS(su);

   // calculate clamped coordinates
   for (c = 0; c < arg; ++c) {
      int dimc = c;

      if (c == 1 && su->tex.target == TEX_TARGET_1D_ARRAY) {
         // The array index is stored in the Z component for 1D arrays.
         dimc = 2;
      }

      src[c] = bld.getScratch();
      if (c == 0 && raw)
         v = loadSuInfo32(ind, slot, NVC0_SU_INFO_RAW_X);
      else
         v = loadSuInfo32(ind, slot, NVC0_SU_INFO_DIM(dimc));
      bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero)
         ->subOp = getSuClampSubOp(su, dimc);
   }
   for (; c < 3; ++c)
      src[c] = zero;

   // set predicate output
   if (su->tex.target == TEX_TARGET_BUFFER) {
      src[0]->getInsn()->setFlagsDef(1, pred);
   } else
   if (su->tex.target.isArray() || su->tex.target.isCube()) {
      p1 = bld.getSSA(1, FILE_PREDICATE);
      src[dim]->getInsn()->setFlagsDef(1, p1);
   }

   // calculate pixel offset
   if (dim == 1) {
      if (su->tex.target != TEX_TARGET_BUFFER)
         bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff));
   } else
   if (dim == 3) {
      v = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C);
      bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1])
         ->subOp = NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l

      v = loadSuInfo32(ind, slot, NVC0_SU_INFO_PITCH);
      bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0])
         ->subOp = NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l
   } else {
      assert(dim == 2);
      v = loadSuInfo32(ind, slot, NVC0_SU_INFO_PITCH);
      bld.mkOp3(OP_MADSP, TYPE_U32, off, src[1], v, src[0])
         ->subOp = (su->tex.target.isArray() || su->tex.target.isCube()) ?
         NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l
   }

   // calculate effective address part 1
   if (su->tex.target == TEX_TARGET_BUFFER) {
      if (raw) {
         bf = src[0];
      } else {
         v = loadSuInfo32(ind, slot, NVC0_SU_INFO_FMT);
         bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero)
            ->subOp = NV50_IR_SUBOP_V1(7,6,8|2);
      }
   } else {
      Value *y = src[1];
      Value *z = src[2];
      uint16_t subOp = 0;

      switch (dim) {
      case 1:
         y = zero;
         z = zero;
         break;
      case 2:
         z = off;
         if (!su->tex.target.isArray() && !su->tex.target.isCube()) {
            z = loadSuInfo32(ind, slot, NVC0_SU_INFO_UNK1C);
            subOp = NV50_IR_SUBOP_SUBFM_3D;
         }
         break;
      default:
         subOp = NV50_IR_SUBOP_SUBFM_3D;
         assert(dim == 3);
         break;
      }
      insn = bld.mkOp3(OP_SUBFM, TYPE_U32, bf, src[0], y, z);
      insn->subOp = subOp;
      insn->setFlagsDef(1, pred);
   }

   // part 2
   v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR);

   if (su->tex.target == TEX_TARGET_BUFFER) {
      eau = v;
   } else {
      eau = bld.mkOp3v(OP_SUEAU, TYPE_U32, bld.getScratch(4), off, bf, v);
   }
   // add array layer offset
   if (su->tex.target.isArray() || su->tex.target.isCube()) {
      v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ARRAY);
      if (dim == 1)
         bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau)
            ->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32
      else
         bld.mkOp3(OP_MADSP, TYPE_U32, eau, v, src[2], eau)
            ->subOp = NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32
      // combine predicates
      assert(p1);
      bld.mkOp2(OP_OR, TYPE_U8, pred, pred, p1);
   }

   if (atom) {
      Value *lo = bf;
      if (su->tex.target == TEX_TARGET_BUFFER) {
         lo = zero;
         bld.mkMov(off, bf);
      }
      //  bf == g[] address & 0xff
      // eau == g[] address >> 8
      bld.mkOp3(OP_PERMT, TYPE_U32,  bf,   lo, bld.loadImm(NULL, 0x6540), eau);
      bld.mkOp3(OP_PERMT, TYPE_U32, eau, zero, bld.loadImm(NULL, 0x0007), eau);
   } else
   if (su->op == OP_SULDP && su->tex.target == TEX_TARGET_BUFFER) {
      // Convert from u32 to u8 address format, which is what the library code
      // doing SULDP currently uses.
      // XXX: can SUEAU do this ?
      // XXX: does it matter that we don't mask high bytes in bf ?
      // Grrr.
      bld.mkOp2(OP_SHR, TYPE_U32, off, bf, bld.mkImm(8));
      bld.mkOp2(OP_ADD, TYPE_U32, eau, eau, off);
   }

   bld.mkOp2(OP_MERGE, TYPE_U64, addr, bf, eau);

   if (atom && su->tex.target == TEX_TARGET_BUFFER)
      bld.mkOp2(OP_ADD, TYPE_U64, addr, addr, off);

   // let's just set it 0 for raw access and hope it works
   v = raw ?
      bld.mkImm(0) : loadSuInfo32(ind, slot, NVC0_SU_INFO_FMT);

   // get rid of old coordinate sources, make space for fmt info and predicate
   su->moveSources(arg, 3 - arg);
   // set 64 bit address and 32-bit format sources
   su->setSrc(0, addr);
   su->setSrc(1, v);
   su->setSrc(2, pred);

   // prevent read fault when the image is not actually bound
   CmpInstruction *pred1 =
      bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
                TYPE_U32, bld.mkImm(0),
                loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR));

   if (su->op != OP_SUSTP && su->tex.format) {
      const TexInstruction::ImgFormatDesc *format = su->tex.format;
      int blockwidth = format->bits[0] + format->bits[1] +
                       format->bits[2] + format->bits[3];

      // make sure that the format doesn't mismatch
      assert(format->components != 0);
      bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred1->getDef(0),
                TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
                loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE),
                pred1->getDef(0));
   }
   su->setPredicate(CC_NOT_P, pred1->getDef(0));

   // TODO: initialize def values to 0 when the surface operation is not
   // performed (not needed for stores). Also, fix the "address bounds test"
   // subtests from arb_shader_image_load_store-invalid for buffers, because it
   // seems like that the predicate is not correctly set by suclamp.
}

static DataType
getSrcType(const TexInstruction::ImgFormatDesc *t, int c)
{
   switch (t->type) {
   case FLOAT: return t->bits[c] == 16 ? TYPE_F16 : TYPE_F32;
   case UNORM: return t->bits[c] == 8 ? TYPE_U8 : TYPE_U16;
   case SNORM: return t->bits[c] == 8 ? TYPE_S8 : TYPE_S16;
   case UINT:
      return (t->bits[c] == 8 ? TYPE_U8 :
              (t->bits[c] == 16 ? TYPE_U16 : TYPE_U32));
   case SINT:
      return (t->bits[c] == 8 ? TYPE_S8 :
              (t->bits[c] == 16 ? TYPE_S16 : TYPE_S32));
   }
   return TYPE_NONE;
}

static DataType
getDestType(const ImgType type) {
   switch (type) {
   case FLOAT:
   case UNORM:
   case SNORM:
      return TYPE_F32;
   case UINT:
      return TYPE_U32;
   case SINT:
      return TYPE_S32;
   default:
      assert(!"Impossible type");
      return TYPE_NONE;
   }
}

void
NVC0LoweringPass::convertSurfaceFormat(TexInstruction *su)
{
   const TexInstruction::ImgFormatDesc *format = su->tex.format;
   int width = format->bits[0] + format->bits[1] +
      format->bits[2] + format->bits[3];
   Value *untypedDst[4] = {};
   Value *typedDst[4] = {};

   // We must convert this to a generic load.
   su->op = OP_SULDB;

   su->dType = typeOfSize(width / 8);
   su->sType = TYPE_U8;

   for (int i = 0; i < width / 32; i++)
      untypedDst[i] = bld.getSSA();
   if (width < 32)
      untypedDst[0] = bld.getSSA();

   for (int i = 0; i < 4; i++) {
      typedDst[i] = su->getDef(i);
   }

   // Set the untyped dsts as the su's destinations
   for (int i = 0; i < 4; i++)
      su->setDef(i, untypedDst[i]);

   bld.setPosition(su, true);

   // Unpack each component into the typed dsts
   int bits = 0;
   for (int i = 0; i < 4; bits += format->bits[i], i++) {
      if (!typedDst[i])
         continue;
      if (i >= format->components) {
         if (format->type == FLOAT ||
             format->type == UNORM ||
             format->type == SNORM)
            bld.loadImm(typedDst[i], i == 3 ? 1.0f : 0.0f);
         else
            bld.loadImm(typedDst[i], i == 3 ? 1 : 0);
         continue;
      }

      // Get just that component's data into the relevant place
      if (format->bits[i] == 32)
         bld.mkMov(typedDst[i], untypedDst[i]);
      else if (format->bits[i] == 16)
         bld.mkCvt(OP_CVT, getDestType(format->type), typedDst[i],
                   getSrcType(format, i), untypedDst[i / 2])
         ->subOp = (i & 1) << (format->type == FLOAT ? 0 : 1);
      else if (format->bits[i] == 8)
         bld.mkCvt(OP_CVT, getDestType(format->type), typedDst[i],
                   getSrcType(format, i), untypedDst[0])->subOp = i;
      else {
         bld.mkOp2(OP_EXTBF, TYPE_U32, typedDst[i], untypedDst[bits / 32],
                   bld.mkImm((bits % 32) | (format->bits[i] << 8)));
         if (format->type == UNORM || format->type == SNORM)
            bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], getSrcType(format, i), typedDst[i]);
      }

      // Normalize / convert as necessary
      if (format->type == UNORM)
         bld.mkOp2(OP_MUL, TYPE_F32, typedDst[i], typedDst[i], bld.loadImm(NULL, 1.0f / ((1 << format->bits[i]) - 1)));
      else if (format->type == SNORM)
         bld.mkOp2(OP_MUL, TYPE_F32, typedDst[i], typedDst[i], bld.loadImm(NULL, 1.0f / ((1 << (format->bits[i] - 1)) - 1)));
      else if (format->type == FLOAT && format->bits[i] < 16) {
         bld.mkOp2(OP_SHL, TYPE_U32, typedDst[i], typedDst[i], bld.loadImm(NULL, 15 - format->bits[i]));
         bld.mkCvt(OP_CVT, TYPE_F32, typedDst[i], TYPE_F16, typedDst[i]);
      }
   }

   if (format->bgra) {
      std::swap(typedDst[0], typedDst[2]);
   }
}

void
NVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction *su)
{
   processSurfaceCoordsNVE4(su);

   if (su->op == OP_SULDP)
      convertSurfaceFormat(su);

   if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
      Value *pred = su->getSrc(2);
      CondCode cc = CC_NOT_P;
      if (su->getPredicate()) {
         pred = bld.getScratch(1, FILE_PREDICATE);
         cc = su->cc;
         if (cc == CC_NOT_P) {
            bld.mkOp2(OP_OR, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
         } else {
            bld.mkOp2(OP_AND, TYPE_U8, pred, su->getPredicate(), su->getSrc(2));
            pred->getInsn()->src(1).mod = Modifier(NV50_IR_MOD_NOT);
         }
      }
      Instruction *red = bld.mkOp(OP_ATOM, su->dType, bld.getSSA());
      red->subOp = su->subOp;
      red->setSrc(0, bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0));
      red->setSrc(1, su->getSrc(3));
      if (su->subOp == NV50_IR_SUBOP_ATOM_CAS)
         red->setSrc(2, su->getSrc(4));
      red->setIndirect(0, 0, su->getSrc(0));

      // make sure to initialize dst value when the atomic operation is not
      // performed
      Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));

      assert(cc == CC_NOT_P);
      red->setPredicate(cc, pred);
      mov->setPredicate(CC_P, pred);

      bld.mkOp2(OP_UNION, TYPE_U32, su->getDef(0),
                red->getDef(0), mov->getDef(0));

      delete_Instruction(bld.getProgram(), su);
      handleCasExch(red, true);
   }

   if (su->op == OP_SUSTB || su->op == OP_SUSTP)
      su->sType = (su->tex.target == TEX_TARGET_BUFFER) ? TYPE_U32 : TYPE_U8;
}

void
NVC0LoweringPass::processSurfaceCoordsNVC0(TexInstruction *su)
{
   const int slot = su->tex.r;
   const int dim = su->tex.target.getDim();
   const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
   int c;
   Value *zero = bld.mkImm(0);
   Value *src[3];
   Value *v;
   Value *ind = su->getIndirectR();

   bld.setPosition(su, false);

   adjustCoordinatesMS(su);

   if (ind) {
      Value *ptr;
      ptr = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(), ind, bld.mkImm(su->tex.r));
      ptr = bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ptr, bld.mkImm(7));
      su->setIndirectR(ptr);
   }

   // get surface coordinates
   for (c = 0; c < arg; ++c)
      src[c] = su->getSrc(c);
   for (; c < 3; ++c)
      src[c] = zero;

   // calculate pixel offset
   if (su->op == OP_SULDP || su->op == OP_SUREDP) {
      v = loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE);
      su->setSrc(0, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[0], v));
   }

   // add array layer offset
   if (su->tex.target.isArray() || su->tex.target.isCube()) {
      v = loadSuInfo32(ind, slot, NVC0_SU_INFO_ARRAY);
      assert(dim > 1);
      su->setSrc(2, bld.mkOp2v(OP_MUL, TYPE_U32, bld.getSSA(), src[2], v));
   }

   // prevent read fault when the image is not actually bound
   CmpInstruction *pred =
      bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
                TYPE_U32, bld.mkImm(0),
                loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR));
   if (su->op != OP_SUSTP && su->tex.format) {
      const TexInstruction::ImgFormatDesc *format = su->tex.format;
      int blockwidth = format->bits[0] + format->bits[1] +
                       format->bits[2] + format->bits[3];

      assert(format->components != 0);
      // make sure that the format doesn't mismatch when it's not FMT_NONE
      bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred->getDef(0),
                TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
                loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE),
                pred->getDef(0));
   }
   su->setPredicate(CC_NOT_P, pred->getDef(0));
}

void
NVC0LoweringPass::handleSurfaceOpNVC0(TexInstruction *su)
{
   if (su->tex.target == TEX_TARGET_1D_ARRAY) {
      /* As 1d arrays also need 3 coordinates, switching to TEX_TARGET_2D_ARRAY
       * will simplify the lowering pass and the texture constraints. */
      su->moveSources(1, 1);
      su->setSrc(1, bld.loadImm(NULL, 0));
      su->tex.target = TEX_TARGET_2D_ARRAY;
   }

   processSurfaceCoordsNVC0(su);

   if (su->op == OP_SULDP)
      convertSurfaceFormat(su);

   if (su->op == OP_SUREDB || su->op == OP_SUREDP) {
      const int dim = su->tex.target.getDim();
      const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
      LValue *addr = bld.getSSA(8);
      Value *def = su->getDef(0);

      su->op = OP_SULEA;

      // Set the destination to the address
      su->dType = TYPE_U64;
      su->setDef(0, addr);
      su->setDef(1, su->getPredicate());

      bld.setPosition(su, true);

      // Perform the atomic op
      Instruction *red = bld.mkOp(OP_ATOM, su->sType, bld.getSSA());
      red->subOp = su->subOp;
      red->setSrc(0, bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, su->sType, 0));
      red->setSrc(1, su->getSrc(arg));
      if (red->subOp == NV50_IR_SUBOP_ATOM_CAS)
         red->setSrc(2, su->getSrc(arg + 1));
      red->setIndirect(0, 0, addr);

      // make sure to initialize dst value when the atomic operation is not
      // performed
      Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));

      assert(su->cc == CC_NOT_P);
      red->setPredicate(su->cc, su->getPredicate());
      mov->setPredicate(CC_P, su->getPredicate());

      bld.mkOp2(OP_UNION, TYPE_U32, def, red->getDef(0), mov->getDef(0));

      handleCasExch(red, false);
   }
}

void
NVC0LoweringPass::processSurfaceCoordsGM107(TexInstruction *su)
{
   const int slot = su->tex.r;
   const int dim = su->tex.target.getDim();
   const int arg = dim + (su->tex.target.isArray() || su->tex.target.isCube());
   Value *ind = su->getIndirectR();
   int pos = 0;

   bld.setPosition(su, false);

   // add texture handle
   switch (su->op) {
   case OP_SUSTP:
      pos = 4;
      break;
   case OP_SUREDP:
      pos = (su->subOp == NV50_IR_SUBOP_ATOM_CAS) ? 2 : 1;
      break;
   default:
      assert(pos == 0);
      break;
   }
   su->setSrc(arg + pos, loadTexHandle(ind, slot + 32));

   // prevent read fault when the image is not actually bound
   CmpInstruction *pred =
      bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE),
                TYPE_U32, bld.mkImm(0),
                loadSuInfo32(ind, slot, NVC0_SU_INFO_ADDR));
   if (su->op != OP_SUSTP && su->tex.format) {
      const TexInstruction::ImgFormatDesc *format = su->tex.format;
      int blockwidth = format->bits[0] + format->bits[1] +
                       format->bits[2] + format->bits[3];

      assert(format->components != 0);
      // make sure that the format doesn't mismatch when it's not FMT_NONE
      bld.mkCmp(OP_SET_OR, CC_NE, TYPE_U32, pred->getDef(0),
                TYPE_U32, bld.loadImm(NULL, blockwidth / 8),
                loadSuInfo32(ind, slot, NVC0_SU_INFO_BSIZE),
                pred->getDef(0));
   }
   su->setPredicate(CC_NOT_P, pred->getDef(0));
}

void
NVC0LoweringPass::handleSurfaceOpGM107(TexInstruction *su)
{
   processSurfaceCoordsGM107(su);

   if (su->op == OP_SULDP)
      convertSurfaceFormat(su);

   if (su->op == OP_SUREDP) {
      Value *def = su->getDef(0);

      su->op = OP_SUREDB;
      su->setDef(0, bld.getSSA());

      bld.setPosition(su, true);

      // make sure to initialize dst value when the atomic operation is not
      // performed
      Instruction *mov = bld.mkMov(bld.getSSA(), bld.loadImm(NULL, 0));

      assert(su->cc == CC_NOT_P);
      mov->setPredicate(CC_P, su->getPredicate());

      bld.mkOp2(OP_UNION, TYPE_U32, def, su->getDef(0), mov->getDef(0));
   }
}

bool
NVC0LoweringPass::handleWRSV(Instruction *i)
{
   Instruction *st;
   Symbol *sym;
   uint32_t addr;

   // must replace, $sreg are not writeable
   addr = targ->getSVAddress(FILE_SHADER_OUTPUT, i->getSrc(0)->asSym());
   if (addr >= 0x400)
      return false;
   sym = bld.mkSymbol(FILE_SHADER_OUTPUT, 0, i->sType, addr);

   st = bld.mkStore(OP_EXPORT, i->dType, sym, i->getIndirect(0, 0),
                    i->getSrc(1));
   st->perPatch = i->perPatch;

   bld.getBB()->remove(i);
   return true;
}

void
NVC0LoweringPass::handleLDST(Instruction *i)
{
   if (i->src(0).getFile() == FILE_SHADER_INPUT) {
      if (prog->getType() == Program::TYPE_COMPUTE) {
         i->getSrc(0)->reg.file = FILE_MEMORY_CONST;
         i->getSrc(0)->reg.fileIndex = 0;
      } else
      if (prog->getType() == Program::TYPE_GEOMETRY &&
          i->src(0).isIndirect(0)) {
         // XXX: this assumes vec4 units
         Value *ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
                                 i->getIndirect(0, 0), bld.mkImm(4));
         i->setIndirect(0, 0, ptr);
         i->op = OP_VFETCH;
      } else {
         i->op = OP_VFETCH;
         assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP
      }
   } else if (i->src(0).getFile() == FILE_MEMORY_CONST) {
      if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
          prog->getType() == Program::TYPE_COMPUTE) {
         // The launch descriptor only allows to set up 8 CBs, but OpenGL
         // requires at least 12 UBOs. To bypass this limitation, we store the
         // addrs into the driver constbuf and we directly load from the global
         // memory.
         int8_t fileIndex = i->getSrc(0)->reg.fileIndex - 1;
         Value *ind = i->getIndirect(0, 1);

         if (ind) {
            // Clamp the UBO index when an indirect access is used to avoid
            // loading information from the wrong place in the driver cb.
            ind = bld.mkOp2v(OP_MIN, TYPE_U32, ind,
                             bld.mkOp2v(OP_ADD, TYPE_U32, bld.getSSA(),
                                        ind, bld.loadImm(NULL, fileIndex)),
                             bld.loadImm(NULL, 12));
         }

         if (i->src(0).isIndirect(1)) {
            Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
            Value *ptr = loadUboInfo64(ind, fileIndex * 16);
            Value *length = loadUboLength32(ind, fileIndex * 16);
            Value *pred = new_LValue(func, FILE_PREDICATE);
            if (i->src(0).isIndirect(0)) {
               bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
               bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
            }
            i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
            i->setIndirect(0, 1, NULL);
            i->setIndirect(0, 0, ptr);
            bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
            i->setPredicate(CC_NOT_P, pred);
            if (i->defExists(0)) {
               bld.mkMov(i->getDef(0), bld.mkImm(0));
            }
         } else if (fileIndex >= 0) {
            Value *ptr = loadUboInfo64(ind, fileIndex * 16);
            if (i->src(0).isIndirect(0)) {
               bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
            }
            i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
            i->setIndirect(0, 1, NULL);
            i->setIndirect(0, 0, ptr);
         }
      } else if (i->src(0).isIndirect(1)) {
         Value *ptr;
         if (i->src(0).isIndirect(0))
            ptr = bld.mkOp3v(OP_INSBF, TYPE_U32, bld.getSSA(),
                             i->getIndirect(0, 1), bld.mkImm(0x1010),
                             i->getIndirect(0, 0));
         else
            ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
                             i->getIndirect(0, 1), bld.mkImm(16));
         i->setIndirect(0, 1, NULL);
         i->setIndirect(0, 0, ptr);
         i->subOp = NV50_IR_SUBOP_LDC_IS;
      }
   } else if (i->src(0).getFile() == FILE_SHADER_OUTPUT) {
      assert(prog->getType() == Program::TYPE_TESSELLATION_CONTROL);
      i->op = OP_VFETCH;
   } else if (i->src(0).getFile() == FILE_MEMORY_BUFFER) {
      Value *ind = i->getIndirect(0, 1);
      Value *ptr = loadBufInfo64(ind, i->getSrc(0)->reg.fileIndex * 16);
      // XXX come up with a way not to do this for EVERY little access but
      // rather to batch these up somehow. Unfortunately we've lost the
      // information about the field width by the time we get here.
      Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
      Value *length = loadBufLength32(ind, i->getSrc(0)->reg.fileIndex * 16);
      Value *pred = new_LValue(func, FILE_PREDICATE);
      if (i->src(0).isIndirect(0)) {
         bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
         bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
      }
      i->setIndirect(0, 1, NULL);
      i->setIndirect(0, 0, ptr);
      i->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
      bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
      i->setPredicate(CC_NOT_P, pred);
      if (i->defExists(0)) {
         Value *zero, *dst = i->getDef(0);
         i->setDef(0, bld.getSSA());

         bld.setPosition(i, true);
         bld.mkMov((zero = bld.getSSA()), bld.mkImm(0))
            ->setPredicate(CC_P, pred);
         bld.mkOp2(OP_UNION, TYPE_U32, dst, i->getDef(0), zero);
      }
   }
}

void
NVC0LoweringPass::readTessCoord(LValue *dst, int c)
{
   Value *laneid = bld.getSSA();
   Value *x, *y;

   bld.mkOp1(OP_RDSV, TYPE_U32, laneid, bld.mkSysVal(SV_LANEID, 0));

   if (c == 0) {
      x = dst;
      y = NULL;
   } else
   if (c == 1) {
      x = NULL;
      y = dst;
   } else {
      assert(c == 2);
      if (prog->driver->prop.tp.domain != PIPE_PRIM_TRIANGLES) {
         bld.mkMov(dst, bld.loadImm(NULL, 0));
         return;
      }
      x = bld.getSSA();
      y = bld.getSSA();
   }
   if (x)
      bld.mkFetch(x, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f0, NULL, laneid);
   if (y)
      bld.mkFetch(y, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f4, NULL, laneid);

   if (c == 2) {
      bld.mkOp2(OP_ADD, TYPE_F32, dst, x, y);
      bld.mkOp2(OP_SUB, TYPE_F32, dst, bld.loadImm(NULL, 1.0f), dst);
   }
}

bool
NVC0LoweringPass::handleRDSV(Instruction *i)
{
   Symbol *sym = i->getSrc(0)->asSym();
   const SVSemantic sv = sym->reg.data.sv.sv;
   Value *vtx = NULL;
   Instruction *ld;
   uint32_t addr = targ->getSVAddress(FILE_SHADER_INPUT, sym);

   if (addr >= 0x400) {
      // mov $sreg
      if (sym->reg.data.sv.index == 3) {
         // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID
         i->op = OP_MOV;
         i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0));
      }
      if (sv == SV_VERTEX_COUNT) {
         bld.setPosition(i, true);
         bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0), bld.mkImm(0x808));
      }
      return true;
   }

   switch (sv) {
   case SV_POSITION:
      assert(prog->getType() == Program::TYPE_FRAGMENT);
      if (i->srcExists(1)) {
         // Pass offset through to the interpolation logic
         ld = bld.mkInterp(NV50_IR_INTERP_LINEAR | NV50_IR_INTERP_OFFSET,
                           i->getDef(0), addr, NULL);
         ld->setSrc(1, i->getSrc(1));
      } else {
         bld.mkInterp(NV50_IR_INTERP_LINEAR, i->getDef(0), addr, NULL);
      }
      break;
   case SV_FACE:
   {
      Value *face = i->getDef(0);
      bld.mkInterp(NV50_IR_INTERP_FLAT, face, addr, NULL);
      if (i->dType == TYPE_F32) {
         bld.mkOp2(OP_OR, TYPE_U32, face, face, bld.mkImm(0x00000001));
         bld.mkOp1(OP_NEG, TYPE_S32, face, face);
         bld.mkCvt(OP_CVT, TYPE_F32, face, TYPE_S32, face);
      }
   }
      break;
   case SV_TESS_COORD:
      assert(prog->getType() == Program::TYPE_TESSELLATION_EVAL);
      readTessCoord(i->getDef(0)->asLValue(), i->getSrc(0)->reg.data.sv.index);
      break;
   case SV_NTID:
   case SV_NCTAID:
   case SV_GRIDID:
      assert(targ->getChipset() >= NVISA_GK104_CHIPSET); // mov $sreg otherwise
      if (sym->reg.data.sv.index == 3) {
         i->op = OP_MOV;
         i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1));
         return true;
      }
      // Fallthrough
   case SV_WORK_DIM:
      addr += prog->driver->prop.cp.gridInfoBase;
      bld.mkLoad(TYPE_U32, i->getDef(0),
                 bld.mkSymbol(FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
                              TYPE_U32, addr), NULL);
      break;
   case SV_SAMPLE_INDEX:
      // TODO: Properly pass source as an address in the PIX address space
      // (which can be of the form [r0+offset]). But this is currently
      // unnecessary.
      ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
      ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
      break;
   case SV_SAMPLE_POS: {
      Value *off = new_LValue(func, FILE_GPR);
      ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
      ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
      bld.mkOp2(OP_SHL, TYPE_U32, off, i->getDef(0), bld.mkImm(3));
      bld.mkLoad(TYPE_F32,
                 i->getDef(0),
                 bld.mkSymbol(
                       FILE_MEMORY_CONST, prog->driver->io.auxCBSlot,
                       TYPE_U32, prog->driver->io.sampleInfoBase +
                       4 * sym->reg.data.sv.index),
                 off);
      break;
   }
   case SV_SAMPLE_MASK: {
      ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0));
      ld->subOp = NV50_IR_SUBOP_PIXLD_COVMASK;
      Instruction *sampleid =
         bld.mkOp1(OP_PIXLD, TYPE_U32, bld.getSSA(), bld.mkImm(0));
      sampleid->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID;
      Value *masked =
         bld.mkOp2v(OP_AND, TYPE_U32, bld.getSSA(), ld->getDef(0),
                    bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(),
                               bld.loadImm(NULL, 1), sampleid->getDef(0)));
      if (prog->driver->prop.fp.persampleInvocation) {
         bld.mkMov(i->getDef(0), masked);
      } else {
         bld.mkOp3(OP_SELP, TYPE_U32, i->getDef(0), ld->getDef(0), masked,
                   bld.mkImm(0))
            ->subOp = 1;
      }
      break;
   }
   case SV_BASEVERTEX:
   case SV_BASEINSTANCE:
   case SV_DRAWID:
      ld = bld.mkLoad(TYPE_U32, i->getDef(0),
                      bld.mkSymbol(FILE_MEMORY_CONST,
                                   prog->driver->io.auxCBSlot,
                                   TYPE_U32,
                                   prog->driver->io.drawInfoBase +
                                   4 * (sv - SV_BASEVERTEX)),
                      NULL);
      break;
   default:
      if (prog->getType() == Program::TYPE_TESSELLATION_EVAL && !i->perPatch)
         vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0));
      ld = bld.mkFetch(i->getDef(0), i->dType,
                       FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx);
      ld->perPatch = i->perPatch;
      break;
   }
   bld.getBB()->remove(i);
   return true;
}

bool
NVC0LoweringPass::handleDIV(Instruction *i)
{
   if (!isFloatType(i->dType))
      return true;
   bld.setPosition(i, false);
   Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(typeSizeof(i->dType)), i->getSrc(1));
   i->op = OP_MUL;
   i->setSrc(1, rcp->getDef(0));
   return true;
}

bool
NVC0LoweringPass::handleMOD(Instruction *i)
{
   if (!isFloatType(i->dType))
      return true;
   LValue *value = bld.getScratch(typeSizeof(i->dType));
   bld.mkOp1(OP_RCP, i->dType, value, i->getSrc(1));
   bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(0), value);
   bld.mkOp1(OP_TRUNC, i->dType, value, value);
   bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(1), value);
   i->op = OP_SUB;
   i->setSrc(1, value);
   return true;
}

bool
NVC0LoweringPass::handleSQRT(Instruction *i)
{
   if (i->dType == TYPE_F64) {
      Value *pred = bld.getSSA(1, FILE_PREDICATE);
      Value *zero = bld.loadImm(NULL, 0.0);
      Value *dst = bld.getSSA(8);
      bld.mkOp1(OP_RSQ, i->dType, dst, i->getSrc(0));
      bld.mkCmp(OP_SET, CC_LE, i->dType, pred, i->dType, i->getSrc(0), zero);
      bld.mkOp3(OP_SELP, TYPE_U64, dst, zero, dst, pred);
      i->op = OP_MUL;
      i->setSrc(1, dst);
      // TODO: Handle this properly with a library function
   } else {
      bld.setPosition(i, true);
      i->op = OP_RSQ;
      bld.mkOp1(OP_RCP, i->dType, i->getDef(0), i->getDef(0));
   }

   return true;
}

bool
NVC0LoweringPass::handlePOW(Instruction *i)
{
   LValue *val = bld.getScratch();

   bld.mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0));
   bld.mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1;
   bld.mkOp1(OP_PREEX2, TYPE_F32, val, val);

   i->op = OP_EX2;
   i->setSrc(0, val);
   i->setSrc(1, NULL);

   return true;
}

bool
NVC0LoweringPass::handleEXPORT(Instruction *i)
{
   if (prog->getType() == Program::TYPE_FRAGMENT) {
      int id = i->getSrc(0)->reg.data.offset / 4;

      if (i->src(0).isIndirect(0)) // TODO, ugly
         return false;
      i->op = OP_MOV;
      i->subOp = NV50_IR_SUBOP_MOV_FINAL;
      i->src(0).set(i->src(1));
      i->setSrc(1, NULL);
      i->setDef(0, new_LValue(func, FILE_GPR));
      i->getDef(0)->reg.data.id = id;

      prog->maxGPR = MAX2(prog->maxGPR, id);
   } else
   if (prog->getType() == Program::TYPE_GEOMETRY) {
      i->setIndirect(0, 1, gpEmitAddress);
   }
   return true;
}

bool
NVC0LoweringPass::handleOUT(Instruction *i)
{
   Instruction *prev = i->prev;
   ImmediateValue stream, prevStream;

   // Only merge if the stream ids match. Also, note that the previous
   // instruction would have already been lowered, so we take arg1 from it.
   if (i->op == OP_RESTART && prev && prev->op == OP_EMIT &&
       i->src(0).getImmediate(stream) &&
       prev->src(1).getImmediate(prevStream) &&
       stream.reg.data.u32 == prevStream.reg.data.u32) {
      i->prev->subOp = NV50_IR_SUBOP_EMIT_RESTART;
      delete_Instruction(prog, i);
   } else {
      assert(gpEmitAddress);
      i->setDef(0, gpEmitAddress);
      i->setSrc(1, i->getSrc(0));
      i->setSrc(0, gpEmitAddress);
   }
   return true;
}

// Generate a binary predicate if an instruction is predicated by
// e.g. an f32 value.
void
NVC0LoweringPass::checkPredicate(Instruction *insn)
{
   Value *pred = insn->getPredicate();
   Value *pdst;

   if (!pred || pred->reg.file == FILE_PREDICATE)
      return;
   pdst = new_LValue(func, FILE_PREDICATE);

   // CAUTION: don't use pdst->getInsn, the definition might not be unique,
   //  delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass

   bld.mkCmp(OP_SET, CC_NEU, insn->dType, pdst, insn->dType, bld.mkImm(0), pred);

   insn->setPredicate(insn->cc, pdst);
}

//
// - add quadop dance for texturing
// - put FP outputs in GPRs
// - convert instruction sequences
//
bool
NVC0LoweringPass::visit(Instruction *i)
{
   bool ret = true;
   bld.setPosition(i, false);

   if (i->cc != CC_ALWAYS)
      checkPredicate(i);

   switch (i->op) {
   case OP_TEX:
   case OP_TXB:
   case OP_TXL:
   case OP_TXF:
   case OP_TXG:
      return handleTEX(i->asTex());
   case OP_TXD:
      return handleTXD(i->asTex());
   case OP_TXLQ:
      return handleTXLQ(i->asTex());
   case OP_TXQ:
     return handleTXQ(i->asTex());
   case OP_EX2:
      bld.mkOp1(OP_PREEX2, TYPE_F32, i->getDef(0), i->getSrc(0));
      i->setSrc(0, i->getDef(0));
      break;
   case OP_POW:
      return handlePOW(i);
   case OP_DIV:
      return handleDIV(i);
   case OP_MOD:
      return handleMOD(i);
   case OP_SQRT:
      return handleSQRT(i);
   case OP_EXPORT:
      ret = handleEXPORT(i);
      break;
   case OP_EMIT:
   case OP_RESTART:
      return handleOUT(i);
   case OP_RDSV:
      return handleRDSV(i);
   case OP_WRSV:
      return handleWRSV(i);
   case OP_STORE:
   case OP_LOAD:
      handleLDST(i);
      break;
   case OP_ATOM:
   {
      const bool cctl = i->src(0).getFile() == FILE_MEMORY_BUFFER;
      handleATOM(i);
      handleCasExch(i, cctl);
   }
      break;
   case OP_SULDB:
   case OP_SULDP:
   case OP_SUSTB:
   case OP_SUSTP:
   case OP_SUREDB:
   case OP_SUREDP:
      if (targ->getChipset() >= NVISA_GM107_CHIPSET)
         handleSurfaceOpGM107(i->asTex());
      else if (targ->getChipset() >= NVISA_GK104_CHIPSET)
         handleSurfaceOpNVE4(i->asTex());
      else
         handleSurfaceOpNVC0(i->asTex());
      break;
   case OP_SUQ:
      handleSUQ(i->asTex());
      break;
   case OP_BUFQ:
      handleBUFQ(i);
      break;
   default:
      break;
   }

   /* Kepler+ has a special opcode to compute a new base address to be used
    * for indirect loads.
    *
    * Maxwell+ has an additional similar requirement for indirect
    * interpolation ops in frag shaders.
    */
   bool doAfetch = false;
   if (targ->getChipset() >= NVISA_GK104_CHIPSET &&
       !i->perPatch &&
       (i->op == OP_VFETCH || i->op == OP_EXPORT) &&
       i->src(0).isIndirect(0)) {
      doAfetch = true;
   }
   if (targ->getChipset() >= NVISA_GM107_CHIPSET &&
       (i->op == OP_LINTERP || i->op == OP_PINTERP) &&
       i->src(0).isIndirect(0)) {
      doAfetch = true;
   }

   if (doAfetch) {
      Value *addr = cloneShallow(func, i->getSrc(0));
      Instruction *afetch = bld.mkOp1(OP_AFETCH, TYPE_U32, bld.getSSA(),
                                      i->getSrc(0));
      afetch->setIndirect(0, 0, i->getIndirect(0, 0));
      addr->reg.data.offset = 0;
      i->setSrc(0, addr);
      i->setIndirect(0, 0, afetch->getDef(0));
   }

   return ret;
}

bool
TargetNVC0::runLegalizePass(Program *prog, CGStage stage) const
{
   if (stage == CG_STAGE_PRE_SSA) {
      NVC0LoweringPass pass(prog);
      return pass.run(prog, false, true);
   } else
   if (stage == CG_STAGE_POST_RA) {
      NVC0LegalizePostRA pass(prog);
      return pass.run(prog, false, true);
   } else
   if (stage == CG_STAGE_SSA) {
      NVC0LegalizeSSA pass;
      return pass.run(prog, false, true);
   }
   return false;
}

} // namespace nv50_ir