This optimization greatly enhances efficiency of creating new instructions by
masking and shifting operands directly into their place in the instruction,
instead of the old-fashioned way of ORing in each bit separately.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@7179 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/utils/TableGen/CodeEmitterGen.cpp b/utils/TableGen/CodeEmitterGen.cpp
index 9d0d939..8cb7979 100644
--- a/utils/TableGen/CodeEmitterGen.cpp
+++ b/utils/TableGen/CodeEmitterGen.cpp
@@ -59,6 +59,7 @@
//
unsigned op = 0;
std::map<const std::string,unsigned> OpOrder;
+ std::map<const std::string,bool> OpContinuous;
for (unsigned i = 0, e = Vals.size(); i != e; ++i) {
if (Vals[i].getName() != "Inst" &&
!Vals[i].getValue()->isComplete() &&
@@ -71,6 +72,90 @@
<<" = getMachineOpValue(MI, MI.getOperand("<<op<<"));\n";
//<< " MachineOperand &op" << op <<" = MI.getOperand("<<op<<");\n";
OpOrder[Vals[i].getName()] = op++;
+
+ // Is the operand continuous? If so, we can just mask and OR it in
+ // instead of doing it bit-by-bit, saving a lot in runtime cost.
+ const BitsInit *InstInit = BI;
+ int beginBitInVar = -1, endBitInVar = -1,
+ beginBitInInst = -1, endBitInInst = -1;
+ bool continuous = true;
+
+ for (int bit = InstInit->getNumBits()-1; bit >= 0; --bit) {
+ if (VarBitInit *VBI =
+ dynamic_cast<VarBitInit*>(InstInit->getBit(bit))) {
+ TypedInit *TI = VBI->getVariable();
+ if (VarInit *VI = dynamic_cast<VarInit*>(TI)) {
+ // only process the current variable
+ if (VI->getName() != Vals[i].getName())
+ continue;
+
+ if (beginBitInVar == -1)
+ beginBitInVar = VBI->getBitNum();
+
+ if (endBitInVar == -1)
+ endBitInVar = VBI->getBitNum();
+ else {
+ if (endBitInVar == (int)VBI->getBitNum() + 1)
+ endBitInVar = VBI->getBitNum();
+ else {
+ continuous = false;
+ break;
+ }
+ }
+
+ if (beginBitInInst == -1)
+ beginBitInInst = bit;
+ if (endBitInInst == -1)
+ endBitInInst = bit;
+ else {
+ if (endBitInInst == bit + 1)
+ endBitInInst = bit;
+ else {
+ continuous = false;
+ break;
+ }
+ }
+
+ // maintain same distance between bits in field and bits in
+ // instruction. if the relative distances stay the same
+ // throughout,
+ if ((beginBitInVar - (int)VBI->getBitNum()) !=
+ (beginBitInInst - bit))
+ {
+ continuous = false;
+ break;
+ }
+ }
+ }
+ }
+
+ if (continuous) {
+ o << " // continuous: op" << OpOrder[Vals[i].getName()] << "\n";
+
+ // Mask off the right bits
+ // Low mask (ie. shift, if necessary)
+ if (endBitInVar != 0) {
+ o << " op" << OpOrder[Vals[i].getName()]
+ << " >>= endBitInVar;\n";
+ beginBitInVar -= endBitInVar;
+ endBitInVar = 0;
+ }
+
+ // High mask
+ o << " op" << OpOrder[Vals[i].getName()]
+ << " &= (1<<" << beginBitInVar+1 << ") - 1;\n";
+
+ // Shift the value to the correct place (according to place in instr)
+ if (endBitInInst != 0)
+ o << " op" << OpOrder[Vals[i].getName()]
+ << " <<= " << endBitInInst << ";\n";
+
+ // Just OR in the result
+ o << " Value |= op" << OpOrder[Vals[i].getName()] << ";\n";
+ }
+
+ // otherwise, will be taken care of in the loop below using this value:
+ OpContinuous[Vals[i].getName()] = continuous;
}
}
@@ -81,18 +166,32 @@
// Scan through the field looking for bit initializers of the current
// variable...
for (int i = FieldInitializer->getNumBits()-1; i >= 0; --i) {
- if (BitInit *BI=dynamic_cast<BitInit*>(FieldInitializer->getBit(i))){
+ if (BitInit *BI=dynamic_cast<BitInit*>(FieldInitializer->getBit(i)))
+ {
o << " // bit init: f: " << f << ", i: " << i << "\n";
- } else if (UnsetInit *UI =
+ } else if (UnsetInit *UI =
dynamic_cast<UnsetInit*>(FieldInitializer->getBit(i))) {
o << " // unset init: f: " << f << ", i: " << i << "\n";
} else if (VarBitInit *VBI =
dynamic_cast<VarBitInit*>(FieldInitializer->getBit(i))) {
TypedInit *TI = VBI->getVariable();
if (VarInit *VI = dynamic_cast<VarInit*>(TI)) {
- o << " Value |= getValueBit(op" << OpOrder[VI->getName()]
- << ", " << VBI->getBitNum()
- << ")" << " << " << i << ";\n";
+ // If the bits of the field are laid out consecutively in the
+ // instruction, then instead of separately ORing in bits, just
+ // mask and shift the entire field for efficiency.
+ if (OpContinuous[VI->getName()]) {
+ // already taken care of in the loop above, thus there is no
+ // need to individually OR in the bits
+
+ // for debugging, output the regular version anyway, commented
+ o << " // Value |= getValueBit(op"
+ << OpOrder[VI->getName()] << ", " << VBI->getBitNum()
+ << ")" << " << " << i << ";\n";
+ } else {
+ o << " Value |= getValueBit(op" << OpOrder[VI->getName()]
+ << ", " << VBI->getBitNum()
+ << ")" << " << " << i << ";\n";
+ }
} else if (FieldInit *FI = dynamic_cast<FieldInit*>(TI)) {
// FIXME: implement this!
o << "FIELD INIT not implemented yet!\n";
@@ -100,7 +199,7 @@
o << "Error: UNIMPLEMENTED\n";
}
}
- }
+ }
} else {
// ignore annul and predict bits since no one sets them yet
if (Vals[f].getName() == "annul" || Vals[f].getName() == "predict") {
@@ -112,6 +211,7 @@
o << " break;\n"
<< " }\n";
}
+
o << " default:\n"
<< " DEBUG(std::cerr << \"Not supported instr: \" << MI << \"\\n\");\n"
<< " abort();\n"