It's not legal to fold a load from a narrower stack slot into a wider instruction...
[llvm/avr.git] / lib / CodeGen / SelectionDAG / LegalizeVectorOps.cpp
blobca194305d9898a7ac246f3e1d38dfc2369198932
1 //===-- LegalizeVectorOps.cpp - Implement SelectionDAG::LegalizeVectors ---===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the SelectionDAG::LegalizeVectors method.
12 // The vector legalizer looks for vector operations which might need to be
13 // scalarized and legalizes them. This is a separate step from Legalize because
14 // scalarizing can introduce illegal types. For example, suppose we have an
15 // ISD::SDIV of type v2i64 on x86-32. The type is legal (for example, addition
16 // on a v2i64 is legal), but ISD::SDIV isn't legal, so we have to unroll the
17 // operation, which introduces nodes with the illegal type i64 which must be
18 // expanded. Similarly, suppose we have an ISD::SRA of type v16i8 on PowerPC;
19 // the operation must be unrolled, which introduces nodes with the illegal
20 // type i8 which must be promoted.
22 // This does not legalize vector manipulations like ISD::BUILD_VECTOR,
23 // or operations that happen to take a vector which are custom-lowered;
24 // the legalization for such operations never produces nodes
25 // with illegal types, so it's okay to put off legalizing them until
26 // SelectionDAG::Legalize runs.
28 //===----------------------------------------------------------------------===//
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/Target/TargetLowering.h"
32 using namespace llvm;
34 namespace {
35 class VectorLegalizer {
36 SelectionDAG& DAG;
37 TargetLowering& TLI;
38 bool Changed; // Keep track of whether anything changed
40 /// LegalizedNodes - For nodes that are of legal width, and that have more
41 /// than one use, this map indicates what regularized operand to use. This
42 /// allows us to avoid legalizing the same thing more than once.
43 DenseMap<SDValue, SDValue> LegalizedNodes;
45 // Adds a node to the translation cache
46 void AddLegalizedOperand(SDValue From, SDValue To) {
47 LegalizedNodes.insert(std::make_pair(From, To));
48 // If someone requests legalization of the new node, return itself.
49 if (From != To)
50 LegalizedNodes.insert(std::make_pair(To, To));
53 // Legalizes the given node
54 SDValue LegalizeOp(SDValue Op);
55 // Assuming the node is legal, "legalize" the results
56 SDValue TranslateLegalizeResults(SDValue Op, SDValue Result);
57 // Implements unrolling a generic vector operation, i.e. turning it into
58 // scalar operations.
59 SDValue UnrollVectorOp(SDValue Op);
60 // Implements unrolling a VSETCC.
61 SDValue UnrollVSETCC(SDValue Op);
62 // Implements expansion for FNEG; falls back to UnrollVectorOp if FSUB
63 // isn't legal.
64 SDValue ExpandFNEG(SDValue Op);
65 // Implements vector promotion; this is essentially just bitcasting the
66 // operands to a different type and bitcasting the result back to the
67 // original type.
68 SDValue PromoteVectorOp(SDValue Op);
70 public:
71 bool Run();
72 VectorLegalizer(SelectionDAG& dag) :
73 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
76 bool VectorLegalizer::Run() {
77 // The legalize process is inherently a bottom-up recursive process (users
78 // legalize their uses before themselves). Given infinite stack space, we
79 // could just start legalizing on the root and traverse the whole graph. In
80 // practice however, this causes us to run out of stack space on large basic
81 // blocks. To avoid this problem, compute an ordering of the nodes where each
82 // node is only legalized after all of its operands are legalized.
83 DAG.AssignTopologicalOrder();
84 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
85 E = prior(DAG.allnodes_end()); I != next(E); ++I)
86 LegalizeOp(SDValue(I, 0));
88 // Finally, it's possible the root changed. Get the new root.
89 SDValue OldRoot = DAG.getRoot();
90 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
91 DAG.setRoot(LegalizedNodes[OldRoot]);
93 LegalizedNodes.clear();
95 // Remove dead nodes now.
96 DAG.RemoveDeadNodes();
98 return Changed;
101 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
102 // Generic legalization: just pass the operand through.
103 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
104 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
105 return Result.getValue(Op.getResNo());
108 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
109 // Note that LegalizeOp may be reentered even from single-use nodes, which
110 // means that we always must cache transformed nodes.
111 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
112 if (I != LegalizedNodes.end()) return I->second;
114 SDNode* Node = Op.getNode();
116 // Legalize the operands
117 SmallVector<SDValue, 8> Ops;
118 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
119 Ops.push_back(LegalizeOp(Node->getOperand(i)));
121 SDValue Result =
122 DAG.UpdateNodeOperands(Op.getValue(0), Ops.data(), Ops.size());
124 bool HasVectorValue = false;
125 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
126 J != E;
127 ++J)
128 HasVectorValue |= J->isVector();
129 if (!HasVectorValue)
130 return TranslateLegalizeResults(Op, Result);
132 EVT QueryType;
133 switch (Op.getOpcode()) {
134 default:
135 return TranslateLegalizeResults(Op, Result);
136 case ISD::ADD:
137 case ISD::SUB:
138 case ISD::MUL:
139 case ISD::SDIV:
140 case ISD::UDIV:
141 case ISD::SREM:
142 case ISD::UREM:
143 case ISD::FADD:
144 case ISD::FSUB:
145 case ISD::FMUL:
146 case ISD::FDIV:
147 case ISD::FREM:
148 case ISD::AND:
149 case ISD::OR:
150 case ISD::XOR:
151 case ISD::SHL:
152 case ISD::SRA:
153 case ISD::SRL:
154 case ISD::ROTL:
155 case ISD::ROTR:
156 case ISD::CTTZ:
157 case ISD::CTLZ:
158 case ISD::CTPOP:
159 case ISD::SELECT:
160 case ISD::SELECT_CC:
161 case ISD::VSETCC:
162 case ISD::ZERO_EXTEND:
163 case ISD::ANY_EXTEND:
164 case ISD::TRUNCATE:
165 case ISD::SIGN_EXTEND:
166 case ISD::FP_TO_SINT:
167 case ISD::FP_TO_UINT:
168 case ISD::FNEG:
169 case ISD::FABS:
170 case ISD::FSQRT:
171 case ISD::FSIN:
172 case ISD::FCOS:
173 case ISD::FPOWI:
174 case ISD::FPOW:
175 case ISD::FLOG:
176 case ISD::FLOG2:
177 case ISD::FLOG10:
178 case ISD::FEXP:
179 case ISD::FEXP2:
180 case ISD::FCEIL:
181 case ISD::FTRUNC:
182 case ISD::FRINT:
183 case ISD::FNEARBYINT:
184 case ISD::FFLOOR:
185 QueryType = Node->getValueType(0);
186 break;
187 case ISD::SINT_TO_FP:
188 case ISD::UINT_TO_FP:
189 QueryType = Node->getOperand(0).getValueType();
190 break;
193 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
194 case TargetLowering::Promote:
195 // "Promote" the operation by bitcasting
196 Result = PromoteVectorOp(Op);
197 Changed = true;
198 break;
199 case TargetLowering::Legal: break;
200 case TargetLowering::Custom: {
201 SDValue Tmp1 = TLI.LowerOperation(Op, DAG);
202 if (Tmp1.getNode()) {
203 Result = Tmp1;
204 break;
206 // FALL THROUGH
208 case TargetLowering::Expand:
209 if (Node->getOpcode() == ISD::FNEG)
210 Result = ExpandFNEG(Op);
211 else if (Node->getOpcode() == ISD::VSETCC)
212 Result = UnrollVSETCC(Op);
213 else
214 Result = UnrollVectorOp(Op);
215 break;
218 // Make sure that the generated code is itself legal.
219 if (Result != Op) {
220 Result = LegalizeOp(Result);
221 Changed = true;
224 // Note that LegalizeOp may be reentered even from single-use nodes, which
225 // means that we always must cache transformed nodes.
226 AddLegalizedOperand(Op, Result);
227 return Result;
230 SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) {
231 // Vector "promotion" is basically just bitcasting and doing the operation
232 // in a different type. For example, x86 promotes ISD::AND on v2i32 to
233 // v1i64.
234 EVT VT = Op.getValueType();
235 assert(Op.getNode()->getNumValues() == 1 &&
236 "Can't promote a vector with multiple results!");
237 EVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
238 DebugLoc dl = Op.getDebugLoc();
239 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
241 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
242 if (Op.getOperand(j).getValueType().isVector())
243 Operands[j] = DAG.getNode(ISD::BIT_CONVERT, dl, NVT, Op.getOperand(j));
244 else
245 Operands[j] = Op.getOperand(j);
248 Op = DAG.getNode(Op.getOpcode(), dl, NVT, &Operands[0], Operands.size());
250 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Op);
253 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
254 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
255 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
256 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
257 Zero, Op.getOperand(0));
259 return UnrollVectorOp(Op);
262 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
263 EVT VT = Op.getValueType();
264 unsigned NumElems = VT.getVectorNumElements();
265 EVT EltVT = VT.getVectorElementType();
266 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
267 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
268 DebugLoc dl = Op.getDebugLoc();
269 SmallVector<SDValue, 8> Ops(NumElems);
270 for (unsigned i = 0; i < NumElems; ++i) {
271 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
272 DAG.getIntPtrConstant(i));
273 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
274 DAG.getIntPtrConstant(i));
275 Ops[i] = DAG.getNode(ISD::SETCC, dl, TLI.getSetCCResultType(TmpEltVT),
276 LHSElem, RHSElem, CC);
277 Ops[i] = DAG.getNode(ISD::SELECT, dl, EltVT, Ops[i],
278 DAG.getConstant(APInt::getAllOnesValue
279 (EltVT.getSizeInBits()), EltVT),
280 DAG.getConstant(0, EltVT));
282 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], NumElems);
285 /// UnrollVectorOp - We know that the given vector has a legal type, however
286 /// the operation it performs is not legal, and the target has requested that
287 /// the operation be expanded. "Unroll" the vector, splitting out the scalars
288 /// and operating on each element individually.
289 SDValue VectorLegalizer::UnrollVectorOp(SDValue Op) {
290 EVT VT = Op.getValueType();
291 assert(Op.getNode()->getNumValues() == 1 &&
292 "Can't unroll a vector with multiple results!");
293 unsigned NE = VT.getVectorNumElements();
294 EVT EltVT = VT.getVectorElementType();
295 DebugLoc dl = Op.getDebugLoc();
297 SmallVector<SDValue, 8> Scalars;
298 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
299 for (unsigned i = 0; i != NE; ++i) {
300 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
301 SDValue Operand = Op.getOperand(j);
302 EVT OperandVT = Operand.getValueType();
303 if (OperandVT.isVector()) {
304 // A vector operand; extract a single element.
305 EVT OperandEltVT = OperandVT.getVectorElementType();
306 Operands[j] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
307 OperandEltVT,
308 Operand,
309 DAG.getConstant(i, MVT::i32));
310 } else {
311 // A scalar operand; just use it as is.
312 Operands[j] = Operand;
316 switch (Op.getOpcode()) {
317 default:
318 Scalars.push_back(DAG.getNode(Op.getOpcode(), dl, EltVT,
319 &Operands[0], Operands.size()));
320 break;
321 case ISD::SHL:
322 case ISD::SRA:
323 case ISD::SRL:
324 case ISD::ROTL:
325 case ISD::ROTR:
326 Scalars.push_back(DAG.getNode(Op.getOpcode(), dl, EltVT, Operands[0],
327 DAG.getShiftAmountOperand(Operands[1])));
328 break;
332 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Scalars[0], Scalars.size());
337 bool SelectionDAG::LegalizeVectors() {
338 return VectorLegalizer(*this).Run();