#include "codegen/nv50_ir.h"
#include "codegen/nv50_ir_target.h"
+#include <algorithm>
#include <stack>
#include <limits>
+#if __cplusplus >= 201103L
+#include <unordered_map>
+#else
#include <tr1/unordered_map>
+#endif
namespace nv50_ir {
+#if __cplusplus >= 201103L
+using std::hash;
+using std::unordered_map;
+#else
+using std::tr1::hash;
+using std::tr1::unordered_map;
+#endif
+
#define MAX_REGISTER_FILE_SIZE 256
class RegisterSet
return (size < 4) ? u : ((u << unit[f]) / 4);
}
- void print() const;
+ void print(DataFile f) const;
+
+ const bool restrictedGPR16Range;
private:
BitSet bits[LAST_REGISTER_FILE + 1];
int last[LAST_REGISTER_FILE + 1];
int fill[LAST_REGISTER_FILE + 1];
-
- const bool restrictedGPR16Range;
};
void
}
void
-RegisterSet::print() const
+RegisterSet::print(DataFile f) const
{
INFO("GPR:");
- bits[FILE_GPR].print();
+ bits[f].print();
INFO("\n");
}
private:
virtual bool visit(BasicBlock *);
+ void insertConstraintMove(Instruction *, int s);
bool insertConstraintMoves();
void condenseDefs(Instruction *);
struct PhiMapHash {
size_t operator()(const std::pair<Instruction *, BasicBlock *>& val) const {
- return std::tr1::hash<Instruction*>()(val.first) * 31 +
- std::tr1::hash<BasicBlock*>()(val.second);
+ return hash<Instruction*>()(val.first) * 31 +
+ hash<BasicBlock*>()(val.second);
}
};
-typedef std::tr1::unordered_map<
+typedef unordered_map<
std::pair<Instruction *, BasicBlock *>, Value *, PhiMapHash> PhiMap;
// Critical edges need to be split up so that work can be inserted along
bool coalesce(ArrayList&);
bool doCoalesce(ArrayList&, unsigned int mask);
void calculateSpillWeights();
- void simplify();
+ bool simplify();
bool selectRegisters();
void cleanup(const bool success);
}
}
+static bool
+isShortRegOp(Instruction *insn)
+{
+ // Immediates are always in src1. Every other situation can be resolved by
+ // using a long encoding.
+ return insn->srcExists(1) && insn->src(1).getFile() == FILE_IMMEDIATE;
+}
+
+// Check if this LValue is ever used in an instruction that can't be encoded
+// with long registers (i.e. > r63)
+static bool
+isShortRegVal(LValue *lval)
+{
+ if (lval->getInsn() == NULL)
+ return false;
+ for (Value::DefCIterator def = lval->defs.begin();
+ def != lval->defs.end(); ++def)
+ if (isShortRegOp((*def)->getInsn()))
+ return true;
+ for (Value::UseCIterator use = lval->uses.begin();
+ use != lval->uses.end(); ++use)
+ if (isShortRegOp((*use)->getInsn()))
+ return true;
+ return false;
+}
+
void
GCRA::RIG_Node::init(const RegisterSet& regs, LValue *lval)
{
weight = std::numeric_limits<float>::infinity();
degree = 0;
- degreeLimit = regs.getFileSize(f, lval->reg.size);
+ int size = regs.getFileSize(f, lval->reg.size);
+ // On nv50, we lose a bit of gpr encoding when there's an embedded
+ // immediate.
+ if (regs.restrictedGPR16Range && f == FILE_GPR && isShortRegVal(lval))
+ size /= 2;
+ degreeLimit = size;
degreeLimit -= relDegree[1][colors] - 1;
livei.insert(lval->livei);
case 0xf0:
case 0x100:
case 0x110:
+ case 0x120:
+ case 0x130:
ret = doCoalesce(insns, JOIN_MASK_UNION);
break;
default:
(node->degree < node->degreeLimit) ? "" : "(spill)");
}
-void
+bool
GCRA::simplify()
{
for (;;) {
}
if (isinf(bestScore)) {
ERROR("no viable spill candidates left\n");
- break;
+ return false;
}
simplifyNode(best);
} else {
- break;
+ return true;
}
}
}
continue;
LValue *lval = node->getValue();
if (prog->dbgFlags & NV50_IR_DEBUG_REG_ALLOC)
- regs.print();
+ regs.print(node->f);
bool ret = regs.assign(node->reg, node->f, node->colors);
if (ret) {
INFO_DBG(prog->dbgFlags, REG_ALLOC, "assigned reg %i\n", node->reg);
if (lval) {
nodes[i].init(regs, lval);
RIG.insert(&nodes[i]);
+
+ if (lval->inFile(FILE_GPR) && lval->getInsn() != NULL) {
+ Instruction *insn = lval->getInsn();
+ if (insn->op != OP_MAD && insn->op != OP_FMA && insn->op != OP_SAD)
+ continue;
+ // For both of the cases below, we only want to add the preference
+ // if all arguments are in registers.
+ if (insn->src(0).getFile() != FILE_GPR ||
+ insn->src(1).getFile() != FILE_GPR ||
+ insn->src(2).getFile() != FILE_GPR)
+ continue;
+ if (prog->getTarget()->getChipset() < 0xc0) {
+ // Outputting a flag is not supported with short encodings nor
+ // with immediate arguments.
+ // See handleMADforNV50.
+ if (insn->flagsDef >= 0)
+ continue;
+ } else {
+ // We can only fold immediate arguments if dst == src2. This
+ // only matters if one of the first two arguments is an
+ // immediate. This form is also only supported for floats.
+ // See handleMADforNVC0.
+ ImmediateValue imm;
+ if (insn->dType != TYPE_F32)
+ continue;
+ if (!insn->src(0).getImmediate(imm) &&
+ !insn->src(1).getImmediate(imm))
+ continue;
+ }
+
+ nodes[i].addRegPreference(getNode(insn->getSrc(2)->asLValue()));
+ }
}
}
buildRIG(insns);
calculateSpillWeights();
- simplify();
+ ret = simplify();
+ if (!ret)
+ goto out;
ret = selectRegisters();
if (!ret) {
delete[] nodes;
nodes = NULL;
+ hi.next = hi.prev = &hi;
+ lo[0].next = lo[0].prev = &lo[0];
+ lo[1].next = lo[1].prev = &lo[1];
}
Symbol *
Instruction *st;
if (slot->reg.file == FILE_MEMORY_LOCAL) {
- st = new_Instruction(func, OP_STORE, ty);
- st->setSrc(0, slot);
- st->setSrc(1, lval);
lval->noSpill = 1;
+ if (ty != TYPE_B96) {
+ st = new_Instruction(func, OP_STORE, ty);
+ st->setSrc(0, slot);
+ st->setSrc(1, lval);
+ } else {
+ st = new_Instruction(func, OP_SPLIT, ty);
+ st->setSrc(0, lval);
+ for (int d = 0; d < lval->reg.size / 4; ++d)
+ st->setDef(d, new_LValue(func, FILE_GPR));
+
+ for (int d = lval->reg.size / 4 - 1; d >= 0; --d) {
+ Value *tmp = cloneShallow(func, slot);
+ tmp->reg.size = 4;
+ tmp->reg.data.offset += 4 * d;
+
+ Instruction *s = new_Instruction(func, OP_STORE, TYPE_U32);
+ s->setSrc(0, tmp);
+ s->setSrc(1, st->getDef(d));
+ defi->bb->insertAfter(defi, s);
+ }
+ }
} else {
st = new_Instruction(func, OP_CVT, ty);
st->setDef(0, slot);
st->setSrc(0, lval);
+ if (lval->reg.file == FILE_FLAGS)
+ st->flagsSrc = 0;
}
defi->bb->insertAfter(defi, st);
}
Instruction *ld;
if (slot->reg.file == FILE_MEMORY_LOCAL) {
lval->noSpill = 1;
- ld = new_Instruction(func, OP_LOAD, ty);
+ if (ty != TYPE_B96) {
+ ld = new_Instruction(func, OP_LOAD, ty);
+ } else {
+ ld = new_Instruction(func, OP_MERGE, ty);
+ for (int d = 0; d < lval->reg.size / 4; ++d) {
+ Value *tmp = cloneShallow(func, slot);
+ LValue *val;
+ tmp->reg.size = 4;
+ tmp->reg.data.offset += 4 * d;
+
+ Instruction *l = new_Instruction(func, OP_LOAD, TYPE_U32);
+ l->setDef(0, (val = new_LValue(func, FILE_GPR)));
+ l->setSrc(0, tmp);
+ usei->bb->insertBefore(usei, l);
+ ld->setSrc(d, val);
+ val->noSpill = 1;
+ }
+ ld->setDef(0, lval);
+ usei->bb->insertBefore(usei, ld);
+ return lval;
+ }
} else {
ld = new_Instruction(func, OP_CVT, ty);
}
ld->setDef(0, lval);
ld->setSrc(0, slot);
+ if (lval->reg.file == FILE_FLAGS)
+ ld->flagsDef = 0;
usei->bb->insertBefore(usei, ld);
return lval;
}
+static bool
+value_cmp(ValueRef *a, ValueRef *b) {
+ Instruction *ai = a->getInsn(), *bi = b->getInsn();
+ if (ai->bb != bi->bb)
+ return ai->bb->getId() < bi->bb->getId();
+ return ai->serial < bi->serial;
+}
// For each value that is to be spilled, go through all its definitions.
// A value can have multiple definitions if it has been coalesced before.
LValue *dval = (*d)->get()->asLValue();
Instruction *defi = (*d)->getInsn();
+ // Sort all the uses by BB/instruction so that we don't unspill
+ // multiple times in a row, and also remove a source of
+ // non-determinism.
+ std::vector<ValueRef *> refs(dval->uses.begin(), dval->uses.end());
+ std::sort(refs.begin(), refs.end(), value_cmp);
+
// Unspill at each use *before* inserting spill instructions,
// we don't want to have the spill instructions in the use list here.
- while (!dval->uses.empty()) {
- ValueRef *u = *dval->uses.begin();
+ for (std::vector<ValueRef*>::const_iterator it = refs.begin();
+ it != refs.end(); ++it) {
+ ValueRef *u = *it;
Instruction *usei = u->getInsn();
assert(usei);
if (usei->isPseudo()) {
tmp = (slot->reg.file == FILE_MEMORY_LOCAL) ? NULL : slot;
last = NULL;
- } else
- if (!last || usei != last->next) { // TODO: sort uses
- tmp = unspill(usei, dval, slot);
+ } else {
+ if (!last || (usei != last->next && usei != last))
+ tmp = unspill(usei, dval, slot);
last = usei;
}
u->set(tmp);
// their registers should be identical.
if (v->getInsn()->op == OP_PHI || v->getInsn()->op == OP_UNION) {
Instruction *phi = v->getInsn();
- for (int phis = 0; phi->srcExists(phis); ++phis)
+ for (int phis = 0; phi->srcExists(phis); ++phis) {
phi->getSrc(phis)->join = v;
+ phi->getSrc(phis)->reg.data.id = v->reg.data.id;
+ }
}
reg += v->reg.size;
}
merge->setDef(0, lval);
for (int s = a, i = 0; s <= b; ++s, ++i) {
merge->setSrc(i, insn->getSrc(s));
- insn->setSrc(s, NULL);
}
+ insn->moveSources(b + 1, a - b);
insn->setSrc(a, lval);
-
- for (int k = a + 1, s = b + 1; insn->srcExists(s); ++s, ++k) {
- insn->setSrc(k, insn->getSrc(s));
- insn->setSrc(s, NULL);
- }
insn->bb->insertBefore(insn, merge);
insn->putExtraSources(0, save);
textureMask(tex);
condenseDefs(tex);
- if (tex->op == OP_SUSTB || tex->op == OP_SUSTP) {
- condenseSrcs(tex, 3, (3 + typeSizeof(tex->dType) / 4) - 1);
+ if (isSurfaceOp(tex->op)) {
+ int s = tex->tex.target.getDim() +
+ (tex->tex.target.isArray() || tex->tex.target.isCube());
+ int n = 0;
+
+ switch (tex->op) {
+ case OP_SUSTB:
+ case OP_SUSTP:
+ n = 4;
+ break;
+ case OP_SUREDB:
+ case OP_SUREDP:
+ if (tex->subOp == NV50_IR_SUBOP_ATOM_CAS)
+ n = 2;
+ break;
+ default:
+ break;
+ }
+
+ if (s > 1)
+ condenseSrcs(tex, 0, s - 1);
+ if (n > 1)
+ condenseSrcs(tex, 1, n); // do not condense the tex handle
} else
if (isTextureOp(tex->op)) {
if (tex->op != OP_TXQ) {
condenseDefs(tex);
if (tex->op == OP_SUSTB || tex->op == OP_SUSTP) {
- condenseSrcs(tex, 3, (3 + typeSizeof(tex->dType) / 4) - 1);
+ condenseSrcs(tex, 3, 6);
} else
if (isTextureOp(tex->op)) {
int n = tex->srcCount(0xff, true);
{
int n, s;
- textureMask(tex);
+ if (isTextureOp(tex->op))
+ textureMask(tex);
if (tex->op == OP_TXQ) {
s = tex->srcCount(0xff);
n = 0;
+ } else if (isSurfaceOp(tex->op)) {
+ s = tex->tex.target.getDim() + (tex->tex.target.isArray() || tex->tex.target.isCube());
+ if (tex->op == OP_SUSTB || tex->op == OP_SUSTP)
+ n = 4;
+ else
+ n = 0;
} else {
s = tex->tex.target.getArgCount() - tex->tex.target.isMS();
if (!tex->tex.target.isArray() &&
for (c = 0; tex->srcExists(c) || tex->defExists(c); ++c) {
if (!tex->srcExists(c))
tex->setSrc(c, new_LValue(func, tex->getSrc(0)->asLValue()));
+ else
+ insertConstraintMove(tex, c);
if (!tex->defExists(c))
tex->setDef(c, new_LValue(func, tex->getDef(0)->asLValue()));
}
texConstraintNVE0(tex);
break;
case 0x110:
+ case 0x120:
+ case 0x130:
texConstraintGM107(tex);
break;
default:
return true;
}
+void
+RegAlloc::InsertConstraintsPass::insertConstraintMove(Instruction *cst, int s)
+{
+ const uint8_t size = cst->src(s).getSize();
+
+ assert(cst->getSrc(s)->defs.size() == 1); // still SSA
+
+ Instruction *defi = cst->getSrc(s)->defs.front()->getInsn();
+ bool imm = defi->op == OP_MOV &&
+ defi->src(0).getFile() == FILE_IMMEDIATE;
+ bool load = defi->op == OP_LOAD &&
+ defi->src(0).getFile() == FILE_MEMORY_CONST &&
+ !defi->src(0).isIndirect(0);
+ // catch some cases where don't really need MOVs
+ if (cst->getSrc(s)->refCount() == 1 && !defi->constrainedDefs()) {
+ if (imm || load) {
+ // Move the defi right before the cst. No point in expanding
+ // the range.
+ defi->bb->remove(defi);
+ cst->bb->insertBefore(cst, defi);
+ }
+ return;
+ }
+
+ LValue *lval = new_LValue(func, cst->src(s).getFile());
+ lval->reg.size = size;
+
+ Instruction *mov = new_Instruction(func, OP_MOV, typeOfSize(size));
+ mov->setDef(0, lval);
+ mov->setSrc(0, cst->getSrc(s));
+
+ if (load) {
+ mov->op = OP_LOAD;
+ mov->setSrc(0, defi->getSrc(0));
+ } else if (imm) {
+ mov->setSrc(0, defi->getSrc(0));
+ }
+
+ if (defi->getPredicate())
+ mov->setPredicate(defi->cc, defi->getPredicate());
+
+ cst->setSrc(s, mov->getDef(0));
+ cst->bb->insertBefore(cst, mov);
+
+ cst->getDef(0)->asLValue()->noSpill = 1; // doesn't help
+}
+
// Insert extra moves so that, if multiple register constraints on a value are
// in conflict, these conflicts can be resolved.
bool
cst->bb->insertBefore(cst, mov);
continue;
}
- assert(cst->getSrc(s)->defs.size() == 1); // still SSA
-
- Instruction *defi = cst->getSrc(s)->defs.front()->getInsn();
- // catch some cases where don't really need MOVs
- if (cst->getSrc(s)->refCount() == 1 && !defi->constrainedDefs())
- continue;
-
- LValue *lval = new_LValue(func, cst->src(s).getFile());
- lval->reg.size = size;
-
- mov = new_Instruction(func, OP_MOV, typeOfSize(size));
- mov->setDef(0, lval);
- mov->setSrc(0, cst->getSrc(s));
- cst->setSrc(s, mov->getDef(0));
- cst->bb->insertBefore(cst, mov);
-
- cst->getDef(0)->asLValue()->noSpill = 1; // doesn't help
- if (cst->op == OP_UNION)
- mov->setPredicate(defi->cc, defi->getPredicate());
+ insertConstraintMove(cst, s);
}
}
}