if (!ld || (ld->op != OP_MOV) ||
((typeSizeof(ld->dType) != 4) && (typeSizeof(ld->dType) != 8)))
return false;
- return ld->src(0).getFile() == FILE_IMMEDIATE;
+
+ // A 0 can be replaced with a register, so it doesn't count as an immediate.
+ ImmediateValue val;
+ return ld->src(0).getImmediate(val) && !val.isInteger(0);
}
bool
void
LoadPropagation::checkSwapSrc01(Instruction *insn)
{
- if (!prog->getTarget()->getOpInfo(insn).commutative)
+ const Target *targ = prog->getTarget();
+ if (!targ->getOpInfo(insn).commutative)
if (insn->op != OP_SET && insn->op != OP_SLCT)
return;
if (insn->src(1).getFile() != FILE_GPR)
Instruction *i0 = insn->getSrc(0)->getInsn();
Instruction *i1 = insn->getSrc(1)->getInsn();
- if (isCSpaceLoad(i0)) {
- if (!isCSpaceLoad(i1))
- insn->swapSources(0, 1);
- else
- return;
- } else
- if (isImmdLoad(i0)) {
- if (!isCSpaceLoad(i1) && !isImmdLoad(i1))
+ // Swap sources to inline the less frequently used source. That way,
+ // optimistically, it will eventually be able to remove the instruction.
+ int i0refs = insn->getSrc(0)->refCount();
+ int i1refs = insn->getSrc(1)->refCount();
+
+ if ((isCSpaceLoad(i0) || isImmdLoad(i0)) && targ->insnCanLoad(insn, 1, i0)) {
+ if ((!isImmdLoad(i1) && !isCSpaceLoad(i1)) ||
+ !targ->insnCanLoad(insn, 1, i1) ||
+ i0refs < i1refs)
insn->swapSources(0, 1);
else
return;
void expr(Instruction *, ImmediateValue&, ImmediateValue&);
void expr(Instruction *, ImmediateValue&, ImmediateValue&, ImmediateValue&);
void opnd(Instruction *, ImmediateValue&, int s);
+ void opnd3(Instruction *, ImmediateValue&);
void unary(Instruction *, const ImmediateValue&);
else
if (i->srcExists(1) && i->src(1).getImmediate(src1))
opnd(i, src1, 1);
+ if (i->srcExists(2) && i->src(2).getImmediate(src2))
+ opnd3(i, src2);
}
return true;
}
switch (i->op) {
case OP_MAD:
case OP_FMA: {
- i->op = OP_ADD;
+ ImmediateValue src0, src1 = *i->getSrc(0)->asImm();
- /* Move the immediate to the second arg, otherwise the ADD operation
- * won't be emittable
- */
- i->setSrc(1, i->getSrc(0));
+ // Move the immediate into position 1, where we know it might be
+ // emittable. However it might not be anyways, as there may be other
+ // restrictions, so move it into a separate LValue.
+ bld.setPosition(i, false);
+ i->op = OP_ADD;
+ i->setSrc(1, bld.mkMov(bld.getSSA(type), i->getSrc(0), type)->getDef(0));
i->setSrc(0, i->getSrc(2));
i->src(0).mod = i->src(2).mod;
i->setSrc(2, NULL);
- ImmediateValue src0;
if (i->src(0).getImmediate(src0))
- expr(i, src0, *i->getSrc(1)->asImm());
- if (i->saturate && !prog->getTarget()->isSatSupported(i)) {
- bld.setPosition(i, false);
- i->setSrc(1, bld.loadImm(NULL, res.data.u32));
- }
+ expr(i, src0, src1);
+ else
+ opnd(i, src1, 1);
break;
}
case OP_PFETCH:
}
}
+void
+ConstantFolding::opnd3(Instruction *i, ImmediateValue &imm2)
+{
+ switch (i->op) {
+ case OP_MAD:
+ case OP_FMA:
+ if (imm2.isInteger(0)) {
+ i->op = OP_MUL;
+ i->setSrc(2, NULL);
+ foldCount++;
+ return;
+ }
+ break;
+ default:
+ return;
+ }
+}
+
void
ConstantFolding::opnd(Instruction *i, ImmediateValue &imm0, int s)
{
i->setSrc(1, bld.loadImm(NULL, imm0.reg.data.u32 + imm1.reg.data.u32));
}
break;
+ case OP_SHR:
+ if (si->src(1).getImmediate(imm1) && imm0.reg.data.u32 == imm1.reg.data.u32) {
+ bld.setPosition(i, false);
+ i->op = OP_AND;
+ i->setSrc(0, si->getSrc(0));
+ i->setSrc(1, bld.loadImm(NULL, ~((1 << imm0.reg.data.u32) - 1)));
+ }
+ break;
case OP_MUL:
int muls;
if (isFloatType(si->dType))
adds = 1;
else
return;
+ if (si->src(!adds).mod != Modifier(0))
+ return;
// SHL(ADD(x, y), z) = ADD(SHL(x, z), SHL(y, z))
// This is more operations, but if one of x, y is an immediate, then
void handleSLCT(Instruction *);
void handleLOGOP(Instruction *);
void handleCVT_NEG(Instruction *);
+ void handleCVT_CVT(Instruction *);
void handleCVT_EXTBF(Instruction *);
void handleSUCLAMP(Instruction *);
+ void handleNEG(Instruction *);
BuildUtil bld;
};
if (src->getUniqueInsn() && src->getUniqueInsn()->bb != add->bb)
return false;
- if (src->getInsn()->postFactor)
+ if (src->getInsn()->saturate || src->getInsn()->postFactor ||
+ src->getInsn()->dnz)
return false;
+
if (toOp == OP_SAD) {
ImmediateValue imm;
if (!src->getInsn()->src(2).getImmediate(imm))
delete_Instruction(prog, cvt);
}
+// F2I(TRUNC()) and so on can be expressed as a single CVT. If the earlier CVT
+// does a type conversion, this becomes trickier as there might be range
+// changes/etc. We could handle those in theory as long as the range was being
+// reduced or kept the same.
+void
+AlgebraicOpt::handleCVT_CVT(Instruction *cvt)
+{
+ Instruction *insn = cvt->getSrc(0)->getInsn();
+ RoundMode rnd = insn->rnd;
+
+ if (insn->saturate ||
+ insn->subOp ||
+ insn->dType != insn->sType ||
+ insn->dType != cvt->sType)
+ return;
+
+ switch (insn->op) {
+ case OP_CEIL:
+ rnd = ROUND_PI;
+ break;
+ case OP_FLOOR:
+ rnd = ROUND_MI;
+ break;
+ case OP_TRUNC:
+ rnd = ROUND_ZI;
+ break;
+ case OP_CVT:
+ break;
+ default:
+ return;
+ }
+
+ if (!isFloatType(cvt->dType) || !isFloatType(insn->sType))
+ rnd = (RoundMode)(rnd & 3);
+
+ cvt->rnd = rnd;
+ cvt->setSrc(0, insn->getSrc(0));
+ cvt->src(0).mod *= insn->src(0).mod;
+ cvt->sType = insn->sType;
+}
+
// Some shaders extract packed bytes out of words and convert them to
// e.g. float. The Fermi+ CVT instruction can extract those directly, as can
// nv50 for word sizes.
arg = shift->getSrc(0);
offset = imm.reg.data.u32;
}
+ // We just AND'd the high bits away, which means this is effectively an
+ // unsigned value.
+ cvt->sType = TYPE_U32;
} else if (insn->op == OP_SHR &&
insn->sType == cvt->sType &&
insn->src(1).getImmediate(imm)) {
insn->setSrc(0, add->getSrc(s));
}
+// NEG(AND(SET, 1)) -> SET
+void
+AlgebraicOpt::handleNEG(Instruction *i) {
+ Instruction *src = i->getSrc(0)->getInsn();
+ ImmediateValue imm;
+ int b;
+
+ if (isFloatType(i->sType) || !src || src->op != OP_AND)
+ return;
+
+ if (src->src(0).getImmediate(imm))
+ b = 1;
+ else if (src->src(1).getImmediate(imm))
+ b = 0;
+ else
+ return;
+
+ if (!imm.isInteger(1))
+ return;
+
+ Instruction *set = src->getSrc(b)->getInsn();
+ if ((set->op == OP_SET || set->op == OP_SET_AND ||
+ set->op == OP_SET_OR || set->op == OP_SET_XOR) &&
+ !isFloatType(set->dType)) {
+ i->def(0).replace(set->getDef(0), false);
+ }
+}
+
bool
AlgebraicOpt::visit(BasicBlock *bb)
{
break;
case OP_CVT:
handleCVT_NEG(i);
+ handleCVT_CVT(i);
if (prog->getTarget()->isOpSupported(OP_EXTBF, TYPE_U32))
handleCVT_EXTBF(i);
break;
case OP_SUCLAMP:
handleSUCLAMP(i);
break;
+ case OP_NEG:
+ handleNEG(i);
+ break;
default:
break;
}
if (((size == 0x8) && (MIN2(offLd, offRc) & 0x7)) ||
((size == 0xc) && (MIN2(offLd, offRc) & 0xf)))
return false;
+ // for compute indirect loads are not guaranteed to be aligned
+ if (prog->getType() == Program::TYPE_COMPUTE && rec->rel[0])
+ return false;
assert(sizeRc + sizeLd <= 16 && offRc != offLd);
if (!prog->getTarget()->
isAccessSupported(st->getSrc(0)->reg.file, typeOfSize(size)))
return false;
+ // no unaligned stores
if (size == 8 && MIN2(offRc, offSt) & 0x7)
return false;
+ // for compute indirect stores are not guaranteed to be aligned
+ if (prog->getType() == Program::TYPE_COMPUTE && rec->rel[0])
+ return false;
st->takeExtraSources(0, extra); // save predicate and indirect address
// get non-replaced sources after values covered by st
for (; offR < endR; offR += ri->getSrc(s)->reg.size, ++s)
vals[k++] = ri->getSrc(s);
- assert((unsigned int)k <= Elements(vals));
+ assert((unsigned int)k <= ARRAY_SIZE(vals));
for (s = 0; s < k; ++s)
st->setSrc(s + 1, vals[s]);
st->setSrc(0, ri->getSrc(0));
}
} else
if (ldst->op == OP_STORE || ldst->op == OP_EXPORT) {
+ if (typeSizeof(ldst->dType) == 4 &&
+ ldst->src(1).getFile() == FILE_GPR &&
+ ldst->getSrc(1)->getInsn()->op == OP_NOP) {
+ delete_Instruction(prog, ldst);
+ continue;
+ }
isLoad = false;
} else {
// TODO: maybe have all fixed ops act as barrier ?
!isSurfaceOp(insn->op) && // not confirmed
insn->op != OP_LINTERP && // probably just nve4
insn->op != OP_PINTERP && // probably just nve4
- ((insn->op != OP_LOAD && insn->op != OP_STORE) ||
+ ((insn->op != OP_LOAD && insn->op != OP_STORE && insn->op != OP_ATOM) ||
(typeSizeof(insn->dType) <= 4 && !insn->src(0).isIndirect(0))) &&
!insn->isNop()) {
insn->join = 1;
virtual bool visit(BasicBlock *);
};
+static bool
+post_ra_dead(Instruction *i)
+{
+ for (int d = 0; i->defExists(d); ++d)
+ if (i->getDef(d)->refCount())
+ return false;
+ return true;
+}
+
bool
NV50PostRaConstantFolding::visit(BasicBlock *bb)
{
i->src(0).getFile() != FILE_GPR ||
i->src(1).getFile() != FILE_GPR ||
i->src(2).getFile() != FILE_GPR ||
- i->getDef(0)->reg.data.id != i->getSrc(2)->reg.data.id ||
- !isFloatType(i->dType))
+ i->getDef(0)->reg.data.id != i->getSrc(2)->reg.data.id)
break;
if (i->getDef(0)->reg.data.id >= 64 ||
i->getSrc(0)->reg.data.id >= 64)
break;
+ if (i->flagsSrc >= 0 && i->getSrc(i->flagsSrc)->reg.data.id != 0)
+ break;
+
if (i->getPredicate())
break;
def = i->getSrc(1)->getInsn();
+ if (def && def->op == OP_SPLIT && typeSizeof(def->sType) == 4)
+ def = def->getSrc(0)->getInsn();
if (def && def->op == OP_MOV && def->src(0).getFile() == FILE_IMMEDIATE) {
vtmp = i->getSrc(1);
- i->setSrc(1, def->getSrc(0));
+ if (isFloatType(i->sType)) {
+ i->setSrc(1, def->getSrc(0));
+ } else {
+ ImmediateValue val;
+ bool ret = def->src(0).getImmediate(val);
+ assert(ret);
+ if (i->getSrc(1)->reg.data.id & 1)
+ val.reg.data.u32 >>= 16;
+ val.reg.data.u32 &= 0xffff;
+ i->setSrc(1, new_ImmediateValue(bb->getProgram(), val.reg.data.u32));
+ }
/* There's no post-RA dead code elimination, so do it here
* XXX: if we add more code-removing post-RA passes, we might
* want to create a post-RA dead-code elim pass */
- if (vtmp->refCount() == 0)
- delete_Instruction(bb->getProgram(), def);
+ if (post_ra_dead(vtmp->getInsn())) {
+ Value *src = vtmp->getInsn()->getSrc(0);
+ // Careful -- splits will have already been removed from the
+ // functions. Don't double-delete.
+ if (vtmp->getInsn()->bb)
+ delete_Instruction(prog, vtmp->getInsn());
+ if (src->getInsn() && post_ra_dead(src->getInsn()))
+ delete_Instruction(prog, src->getInsn());
+ }
break;
}
if (that->srcExists(s))
return false;
- if (op == OP_LOAD || op == OP_VFETCH) {
+ if (op == OP_LOAD || op == OP_VFETCH || op == OP_ATOM) {
switch (src(0).getFile()) {
case FILE_MEMORY_CONST:
case FILE_SHADER_INPUT:
ik = phi->getSrc(0)->getInsn();
if (!ik)
continue; // probably a function input
+ if (ik->defCount(0xff) > 1)
+ continue; // too painful to check if we can really push this forward
for (s = 1; phi->srcExists(s); ++s) {
if (phi->getSrc(s)->refCount() > 1)
break;
bool
DeadCodeElim::visit(BasicBlock *bb)
{
- Instruction *next;
+ Instruction *prev;
- for (Instruction *i = bb->getFirst(); i; i = next) {
- next = i->next;
+ for (Instruction *i = bb->getExit(); i; i = prev) {
+ prev = i->prev;
if (i->isDead()) {
++deadCount;
delete_Instruction(prog, i);
projects / mesa.git / blobdiff