File: | lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp |
Location: | line 6453, column 14 |
Description: | Called C++ object pointer is null |
1 | //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===// | |||
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 implements routines for translating from LLVM IR into SelectionDAG IR. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "SelectionDAGBuilder.h" | |||
15 | #include "SDNodeDbgValue.h" | |||
16 | #include "llvm/ADT/BitVector.h" | |||
17 | #include "llvm/ADT/Optional.h" | |||
18 | #include "llvm/ADT/SmallSet.h" | |||
19 | #include "llvm/ADT/Statistic.h" | |||
20 | #include "llvm/Analysis/AliasAnalysis.h" | |||
21 | #include "llvm/Analysis/BranchProbabilityInfo.h" | |||
22 | #include "llvm/Analysis/ConstantFolding.h" | |||
23 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
24 | #include "llvm/Analysis/ValueTracking.h" | |||
25 | #include "llvm/CodeGen/Analysis.h" | |||
26 | #include "llvm/CodeGen/FastISel.h" | |||
27 | #include "llvm/CodeGen/FunctionLoweringInfo.h" | |||
28 | #include "llvm/CodeGen/GCMetadata.h" | |||
29 | #include "llvm/CodeGen/GCStrategy.h" | |||
30 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
31 | #include "llvm/CodeGen/MachineFunction.h" | |||
32 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
33 | #include "llvm/CodeGen/MachineJumpTableInfo.h" | |||
34 | #include "llvm/CodeGen/MachineModuleInfo.h" | |||
35 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
36 | #include "llvm/CodeGen/SelectionDAG.h" | |||
37 | #include "llvm/CodeGen/StackMaps.h" | |||
38 | #include "llvm/IR/CallingConv.h" | |||
39 | #include "llvm/IR/Constants.h" | |||
40 | #include "llvm/IR/DataLayout.h" | |||
41 | #include "llvm/IR/DebugInfo.h" | |||
42 | #include "llvm/IR/DerivedTypes.h" | |||
43 | #include "llvm/IR/Function.h" | |||
44 | #include "llvm/IR/GlobalVariable.h" | |||
45 | #include "llvm/IR/InlineAsm.h" | |||
46 | #include "llvm/IR/Instructions.h" | |||
47 | #include "llvm/IR/IntrinsicInst.h" | |||
48 | #include "llvm/IR/Intrinsics.h" | |||
49 | #include "llvm/IR/LLVMContext.h" | |||
50 | #include "llvm/IR/Module.h" | |||
51 | #include "llvm/IR/Statepoint.h" | |||
52 | #include "llvm/MC/MCSymbol.h" | |||
53 | #include "llvm/Support/CommandLine.h" | |||
54 | #include "llvm/Support/Debug.h" | |||
55 | #include "llvm/Support/ErrorHandling.h" | |||
56 | #include "llvm/Support/MathExtras.h" | |||
57 | #include "llvm/Support/raw_ostream.h" | |||
58 | #include "llvm/Target/TargetFrameLowering.h" | |||
59 | #include "llvm/Target/TargetInstrInfo.h" | |||
60 | #include "llvm/Target/TargetIntrinsicInfo.h" | |||
61 | #include "llvm/Target/TargetLowering.h" | |||
62 | #include "llvm/Target/TargetOptions.h" | |||
63 | #include "llvm/Target/TargetSelectionDAGInfo.h" | |||
64 | #include "llvm/Target/TargetSubtargetInfo.h" | |||
65 | #include <algorithm> | |||
66 | using namespace llvm; | |||
67 | ||||
68 | #define DEBUG_TYPE"isel" "isel" | |||
69 | ||||
70 | /// LimitFloatPrecision - Generate low-precision inline sequences for | |||
71 | /// some float libcalls (6, 8 or 12 bits). | |||
72 | static unsigned LimitFloatPrecision; | |||
73 | ||||
74 | static cl::opt<unsigned, true> | |||
75 | LimitFPPrecision("limit-float-precision", | |||
76 | cl::desc("Generate low-precision inline sequences " | |||
77 | "for some float libcalls"), | |||
78 | cl::location(LimitFloatPrecision), | |||
79 | cl::init(0)); | |||
80 | ||||
81 | // Limit the width of DAG chains. This is important in general to prevent | |||
82 | // prevent DAG-based analysis from blowing up. For example, alias analysis and | |||
83 | // load clustering may not complete in reasonable time. It is difficult to | |||
84 | // recognize and avoid this situation within each individual analysis, and | |||
85 | // future analyses are likely to have the same behavior. Limiting DAG width is | |||
86 | // the safe approach, and will be especially important with global DAGs. | |||
87 | // | |||
88 | // MaxParallelChains default is arbitrarily high to avoid affecting | |||
89 | // optimization, but could be lowered to improve compile time. Any ld-ld-st-st | |||
90 | // sequence over this should have been converted to llvm.memcpy by the | |||
91 | // frontend. It easy to induce this behavior with .ll code such as: | |||
92 | // %buffer = alloca [4096 x i8] | |||
93 | // %data = load [4096 x i8]* %argPtr | |||
94 | // store [4096 x i8] %data, [4096 x i8]* %buffer | |||
95 | static const unsigned MaxParallelChains = 64; | |||
96 | ||||
97 | static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL, | |||
98 | const SDValue *Parts, unsigned NumParts, | |||
99 | MVT PartVT, EVT ValueVT, const Value *V); | |||
100 | ||||
101 | /// getCopyFromParts - Create a value that contains the specified legal parts | |||
102 | /// combined into the value they represent. If the parts combine to a type | |||
103 | /// larger then ValueVT then AssertOp can be used to specify whether the extra | |||
104 | /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT | |||
105 | /// (ISD::AssertSext). | |||
106 | static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL, | |||
107 | const SDValue *Parts, | |||
108 | unsigned NumParts, MVT PartVT, EVT ValueVT, | |||
109 | const Value *V, | |||
110 | ISD::NodeType AssertOp = ISD::DELETED_NODE) { | |||
111 | if (ValueVT.isVector()) | |||
112 | return getCopyFromPartsVector(DAG, DL, Parts, NumParts, | |||
113 | PartVT, ValueVT, V); | |||
114 | ||||
115 | assert(NumParts > 0 && "No parts to assemble!")((NumParts > 0 && "No parts to assemble!") ? static_cast <void> (0) : __assert_fail ("NumParts > 0 && \"No parts to assemble!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 115, __PRETTY_FUNCTION__)); | |||
116 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
117 | SDValue Val = Parts[0]; | |||
118 | ||||
119 | if (NumParts > 1) { | |||
120 | // Assemble the value from multiple parts. | |||
121 | if (ValueVT.isInteger()) { | |||
122 | unsigned PartBits = PartVT.getSizeInBits(); | |||
123 | unsigned ValueBits = ValueVT.getSizeInBits(); | |||
124 | ||||
125 | // Assemble the power of 2 part. | |||
126 | unsigned RoundParts = NumParts & (NumParts - 1) ? | |||
127 | 1 << Log2_32(NumParts) : NumParts; | |||
128 | unsigned RoundBits = PartBits * RoundParts; | |||
129 | EVT RoundVT = RoundBits == ValueBits ? | |||
130 | ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits); | |||
131 | SDValue Lo, Hi; | |||
132 | ||||
133 | EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); | |||
134 | ||||
135 | if (RoundParts > 2) { | |||
136 | Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, | |||
137 | PartVT, HalfVT, V); | |||
138 | Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, | |||
139 | RoundParts / 2, PartVT, HalfVT, V); | |||
140 | } else { | |||
141 | Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]); | |||
142 | Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]); | |||
143 | } | |||
144 | ||||
145 | if (TLI.isBigEndian()) | |||
146 | std::swap(Lo, Hi); | |||
147 | ||||
148 | Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi); | |||
149 | ||||
150 | if (RoundParts < NumParts) { | |||
151 | // Assemble the trailing non-power-of-2 part. | |||
152 | unsigned OddParts = NumParts - RoundParts; | |||
153 | EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); | |||
154 | Hi = getCopyFromParts(DAG, DL, | |||
155 | Parts + RoundParts, OddParts, PartVT, OddVT, V); | |||
156 | ||||
157 | // Combine the round and odd parts. | |||
158 | Lo = Val; | |||
159 | if (TLI.isBigEndian()) | |||
160 | std::swap(Lo, Hi); | |||
161 | EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); | |||
162 | Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi); | |||
163 | Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi, | |||
164 | DAG.getConstant(Lo.getValueType().getSizeInBits(), | |||
165 | TLI.getPointerTy())); | |||
166 | Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo); | |||
167 | Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi); | |||
168 | } | |||
169 | } else if (PartVT.isFloatingPoint()) { | |||
170 | // FP split into multiple FP parts (for ppcf128) | |||
171 | assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&((ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && "Unexpected split") ? static_cast<void> (0) : __assert_fail ("ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 172, __PRETTY_FUNCTION__)) | |||
172 | "Unexpected split")((ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && "Unexpected split") ? static_cast<void> (0) : __assert_fail ("ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 172, __PRETTY_FUNCTION__)); | |||
173 | SDValue Lo, Hi; | |||
174 | Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]); | |||
175 | Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]); | |||
176 | if (TLI.hasBigEndianPartOrdering(ValueVT)) | |||
177 | std::swap(Lo, Hi); | |||
178 | Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi); | |||
179 | } else { | |||
180 | // FP split into integer parts (soft fp) | |||
181 | assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&((ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split") ? static_cast <void> (0) : __assert_fail ("ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 182, __PRETTY_FUNCTION__)) | |||
182 | !PartVT.isVector() && "Unexpected split")((ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split") ? static_cast <void> (0) : __assert_fail ("ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && \"Unexpected split\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 182, __PRETTY_FUNCTION__)); | |||
183 | EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); | |||
184 | Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V); | |||
185 | } | |||
186 | } | |||
187 | ||||
188 | // There is now one part, held in Val. Correct it to match ValueVT. | |||
189 | EVT PartEVT = Val.getValueType(); | |||
190 | ||||
191 | if (PartEVT == ValueVT) | |||
192 | return Val; | |||
193 | ||||
194 | if (PartEVT.isInteger() && ValueVT.isInteger()) { | |||
195 | if (ValueVT.bitsLT(PartEVT)) { | |||
196 | // For a truncate, see if we have any information to | |||
197 | // indicate whether the truncated bits will always be | |||
198 | // zero or sign-extension. | |||
199 | if (AssertOp != ISD::DELETED_NODE) | |||
200 | Val = DAG.getNode(AssertOp, DL, PartEVT, Val, | |||
201 | DAG.getValueType(ValueVT)); | |||
202 | return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); | |||
203 | } | |||
204 | return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val); | |||
205 | } | |||
206 | ||||
207 | if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { | |||
208 | // FP_ROUND's are always exact here. | |||
209 | if (ValueVT.bitsLT(Val.getValueType())) | |||
210 | return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, | |||
211 | DAG.getTargetConstant(1, TLI.getPointerTy())); | |||
212 | ||||
213 | return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val); | |||
214 | } | |||
215 | ||||
216 | if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits()) | |||
217 | return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); | |||
218 | ||||
219 | llvm_unreachable("Unknown mismatch!")::llvm::llvm_unreachable_internal("Unknown mismatch!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 219); | |||
220 | } | |||
221 | ||||
222 | static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V, | |||
223 | const Twine &ErrMsg) { | |||
224 | const Instruction *I = dyn_cast_or_null<Instruction>(V); | |||
225 | if (!V) | |||
226 | return Ctx.emitError(ErrMsg); | |||
227 | ||||
228 | const char *AsmError = ", possible invalid constraint for vector type"; | |||
229 | if (const CallInst *CI = dyn_cast<CallInst>(I)) | |||
230 | if (isa<InlineAsm>(CI->getCalledValue())) | |||
231 | return Ctx.emitError(I, ErrMsg + AsmError); | |||
232 | ||||
233 | return Ctx.emitError(I, ErrMsg); | |||
234 | } | |||
235 | ||||
236 | /// getCopyFromPartsVector - Create a value that contains the specified legal | |||
237 | /// parts combined into the value they represent. If the parts combine to a | |||
238 | /// type larger then ValueVT then AssertOp can be used to specify whether the | |||
239 | /// extra bits are known to be zero (ISD::AssertZext) or sign extended from | |||
240 | /// ValueVT (ISD::AssertSext). | |||
241 | static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL, | |||
242 | const SDValue *Parts, unsigned NumParts, | |||
243 | MVT PartVT, EVT ValueVT, const Value *V) { | |||
244 | assert(ValueVT.isVector() && "Not a vector value")((ValueVT.isVector() && "Not a vector value") ? static_cast <void> (0) : __assert_fail ("ValueVT.isVector() && \"Not a vector value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 244, __PRETTY_FUNCTION__)); | |||
245 | assert(NumParts > 0 && "No parts to assemble!")((NumParts > 0 && "No parts to assemble!") ? static_cast <void> (0) : __assert_fail ("NumParts > 0 && \"No parts to assemble!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 245, __PRETTY_FUNCTION__)); | |||
246 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
247 | SDValue Val = Parts[0]; | |||
248 | ||||
249 | // Handle a multi-element vector. | |||
250 | if (NumParts > 1) { | |||
251 | EVT IntermediateVT; | |||
252 | MVT RegisterVT; | |||
253 | unsigned NumIntermediates; | |||
254 | unsigned NumRegs = | |||
255 | TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, | |||
256 | NumIntermediates, RegisterVT); | |||
257 | assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!")((NumRegs == NumParts && "Part count doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("NumRegs == NumParts && \"Part count doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 257, __PRETTY_FUNCTION__)); | |||
258 | NumParts = NumRegs; // Silence a compiler warning. | |||
259 | assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!")((RegisterVT == PartVT && "Part type doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == PartVT && \"Part type doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 259, __PRETTY_FUNCTION__)); | |||
260 | assert(RegisterVT == Parts[0].getSimpleValueType() &&((RegisterVT == Parts[0].getSimpleValueType() && "Part type doesn't match part!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == Parts[0].getSimpleValueType() && \"Part type doesn't match part!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 261, __PRETTY_FUNCTION__)) | |||
261 | "Part type doesn't match part!")((RegisterVT == Parts[0].getSimpleValueType() && "Part type doesn't match part!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == Parts[0].getSimpleValueType() && \"Part type doesn't match part!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 261, __PRETTY_FUNCTION__)); | |||
262 | ||||
263 | // Assemble the parts into intermediate operands. | |||
264 | SmallVector<SDValue, 8> Ops(NumIntermediates); | |||
265 | if (NumIntermediates == NumParts) { | |||
266 | // If the register was not expanded, truncate or copy the value, | |||
267 | // as appropriate. | |||
268 | for (unsigned i = 0; i != NumParts; ++i) | |||
269 | Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, | |||
270 | PartVT, IntermediateVT, V); | |||
271 | } else if (NumParts > 0) { | |||
272 | // If the intermediate type was expanded, build the intermediate | |||
273 | // operands from the parts. | |||
274 | assert(NumParts % NumIntermediates == 0 &&((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 275, __PRETTY_FUNCTION__)) | |||
275 | "Must expand into a divisible number of parts!")((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 275, __PRETTY_FUNCTION__)); | |||
276 | unsigned Factor = NumParts / NumIntermediates; | |||
277 | for (unsigned i = 0; i != NumIntermediates; ++i) | |||
278 | Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, | |||
279 | PartVT, IntermediateVT, V); | |||
280 | } | |||
281 | ||||
282 | // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the | |||
283 | // intermediate operands. | |||
284 | Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS | |||
285 | : ISD::BUILD_VECTOR, | |||
286 | DL, ValueVT, Ops); | |||
287 | } | |||
288 | ||||
289 | // There is now one part, held in Val. Correct it to match ValueVT. | |||
290 | EVT PartEVT = Val.getValueType(); | |||
291 | ||||
292 | if (PartEVT == ValueVT) | |||
293 | return Val; | |||
294 | ||||
295 | if (PartEVT.isVector()) { | |||
296 | // If the element type of the source/dest vectors are the same, but the | |||
297 | // parts vector has more elements than the value vector, then we have a | |||
298 | // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the | |||
299 | // elements we want. | |||
300 | if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) { | |||
301 | assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&((PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements () && "Cannot narrow, it would be a lossy transformation" ) ? static_cast<void> (0) : __assert_fail ("PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() && \"Cannot narrow, it would be a lossy transformation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 302, __PRETTY_FUNCTION__)) | |||
302 | "Cannot narrow, it would be a lossy transformation")((PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements () && "Cannot narrow, it would be a lossy transformation" ) ? static_cast<void> (0) : __assert_fail ("PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() && \"Cannot narrow, it would be a lossy transformation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 302, __PRETTY_FUNCTION__)); | |||
303 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val, | |||
304 | DAG.getConstant(0, TLI.getVectorIdxTy())); | |||
305 | } | |||
306 | ||||
307 | // Vector/Vector bitcast. | |||
308 | if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits()) | |||
309 | return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); | |||
310 | ||||
311 | assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&((PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements () && "Cannot handle this kind of promotion") ? static_cast <void> (0) : __assert_fail ("PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() && \"Cannot handle this kind of promotion\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 312, __PRETTY_FUNCTION__)) | |||
312 | "Cannot handle this kind of promotion")((PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements () && "Cannot handle this kind of promotion") ? static_cast <void> (0) : __assert_fail ("PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() && \"Cannot handle this kind of promotion\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 312, __PRETTY_FUNCTION__)); | |||
313 | // Promoted vector extract | |||
314 | bool Smaller = ValueVT.bitsLE(PartEVT); | |||
315 | return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), | |||
316 | DL, ValueVT, Val); | |||
317 | ||||
318 | } | |||
319 | ||||
320 | // Trivial bitcast if the types are the same size and the destination | |||
321 | // vector type is legal. | |||
322 | if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() && | |||
323 | TLI.isTypeLegal(ValueVT)) | |||
324 | return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); | |||
325 | ||||
326 | // Handle cases such as i8 -> <1 x i1> | |||
327 | if (ValueVT.getVectorNumElements() != 1) { | |||
328 | diagnosePossiblyInvalidConstraint(*DAG.getContext(), V, | |||
329 | "non-trivial scalar-to-vector conversion"); | |||
330 | return DAG.getUNDEF(ValueVT); | |||
331 | } | |||
332 | ||||
333 | if (ValueVT.getVectorNumElements() == 1 && | |||
334 | ValueVT.getVectorElementType() != PartEVT) { | |||
335 | bool Smaller = ValueVT.bitsLE(PartEVT); | |||
336 | Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), | |||
337 | DL, ValueVT.getScalarType(), Val); | |||
338 | } | |||
339 | ||||
340 | return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val); | |||
341 | } | |||
342 | ||||
343 | static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl, | |||
344 | SDValue Val, SDValue *Parts, unsigned NumParts, | |||
345 | MVT PartVT, const Value *V); | |||
346 | ||||
347 | /// getCopyToParts - Create a series of nodes that contain the specified value | |||
348 | /// split into legal parts. If the parts contain more bits than Val, then, for | |||
349 | /// integers, ExtendKind can be used to specify how to generate the extra bits. | |||
350 | static void getCopyToParts(SelectionDAG &DAG, SDLoc DL, | |||
351 | SDValue Val, SDValue *Parts, unsigned NumParts, | |||
352 | MVT PartVT, const Value *V, | |||
353 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { | |||
354 | EVT ValueVT = Val.getValueType(); | |||
355 | ||||
356 | // Handle the vector case separately. | |||
357 | if (ValueVT.isVector()) | |||
358 | return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V); | |||
359 | ||||
360 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
361 | unsigned PartBits = PartVT.getSizeInBits(); | |||
362 | unsigned OrigNumParts = NumParts; | |||
363 | assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!")((TLI.isTypeLegal(PartVT) && "Copying to an illegal type!" ) ? static_cast<void> (0) : __assert_fail ("TLI.isTypeLegal(PartVT) && \"Copying to an illegal type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 363, __PRETTY_FUNCTION__)); | |||
364 | ||||
365 | if (NumParts == 0) | |||
366 | return; | |||
367 | ||||
368 | assert(!ValueVT.isVector() && "Vector case handled elsewhere")((!ValueVT.isVector() && "Vector case handled elsewhere" ) ? static_cast<void> (0) : __assert_fail ("!ValueVT.isVector() && \"Vector case handled elsewhere\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 368, __PRETTY_FUNCTION__)); | |||
369 | EVT PartEVT = PartVT; | |||
370 | if (PartEVT == ValueVT) { | |||
371 | assert(NumParts == 1 && "No-op copy with multiple parts!")((NumParts == 1 && "No-op copy with multiple parts!") ? static_cast<void> (0) : __assert_fail ("NumParts == 1 && \"No-op copy with multiple parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 371, __PRETTY_FUNCTION__)); | |||
372 | Parts[0] = Val; | |||
373 | return; | |||
374 | } | |||
375 | ||||
376 | if (NumParts * PartBits > ValueVT.getSizeInBits()) { | |||
377 | // If the parts cover more bits than the value has, promote the value. | |||
378 | if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { | |||
379 | assert(NumParts == 1 && "Do not know what to promote to!")((NumParts == 1 && "Do not know what to promote to!") ? static_cast<void> (0) : __assert_fail ("NumParts == 1 && \"Do not know what to promote to!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 379, __PRETTY_FUNCTION__)); | |||
380 | Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val); | |||
381 | } else { | |||
382 | assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 384, __PRETTY_FUNCTION__)) | |||
383 | ValueVT.isInteger() &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 384, __PRETTY_FUNCTION__)) | |||
384 | "Unknown mismatch!")(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 384, __PRETTY_FUNCTION__)); | |||
385 | ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); | |||
386 | Val = DAG.getNode(ExtendKind, DL, ValueVT, Val); | |||
387 | if (PartVT == MVT::x86mmx) | |||
388 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); | |||
389 | } | |||
390 | } else if (PartBits == ValueVT.getSizeInBits()) { | |||
391 | // Different types of the same size. | |||
392 | assert(NumParts == 1 && PartEVT != ValueVT)((NumParts == 1 && PartEVT != ValueVT) ? static_cast< void> (0) : __assert_fail ("NumParts == 1 && PartEVT != ValueVT" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 392, __PRETTY_FUNCTION__)); | |||
393 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); | |||
394 | } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { | |||
395 | // If the parts cover less bits than value has, truncate the value. | |||
396 | assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 398, __PRETTY_FUNCTION__)) | |||
397 | ValueVT.isInteger() &&(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 398, __PRETTY_FUNCTION__)) | |||
398 | "Unknown mismatch!")(((PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT .isInteger() && "Unknown mismatch!") ? static_cast< void> (0) : __assert_fail ("(PartVT.isInteger() || PartVT == MVT::x86mmx) && ValueVT.isInteger() && \"Unknown mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 398, __PRETTY_FUNCTION__)); | |||
399 | ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); | |||
400 | Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); | |||
401 | if (PartVT == MVT::x86mmx) | |||
402 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); | |||
403 | } | |||
404 | ||||
405 | // The value may have changed - recompute ValueVT. | |||
406 | ValueVT = Val.getValueType(); | |||
407 | assert(NumParts * PartBits == ValueVT.getSizeInBits() &&((NumParts * PartBits == ValueVT.getSizeInBits() && "Failed to tile the value with PartVT!" ) ? static_cast<void> (0) : __assert_fail ("NumParts * PartBits == ValueVT.getSizeInBits() && \"Failed to tile the value with PartVT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 408, __PRETTY_FUNCTION__)) | |||
408 | "Failed to tile the value with PartVT!")((NumParts * PartBits == ValueVT.getSizeInBits() && "Failed to tile the value with PartVT!" ) ? static_cast<void> (0) : __assert_fail ("NumParts * PartBits == ValueVT.getSizeInBits() && \"Failed to tile the value with PartVT!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 408, __PRETTY_FUNCTION__)); | |||
409 | ||||
410 | if (NumParts == 1) { | |||
411 | if (PartEVT != ValueVT) | |||
412 | diagnosePossiblyInvalidConstraint(*DAG.getContext(), V, | |||
413 | "scalar-to-vector conversion failed"); | |||
414 | ||||
415 | Parts[0] = Val; | |||
416 | return; | |||
417 | } | |||
418 | ||||
419 | // Expand the value into multiple parts. | |||
420 | if (NumParts & (NumParts - 1)) { | |||
421 | // The number of parts is not a power of 2. Split off and copy the tail. | |||
422 | assert(PartVT.isInteger() && ValueVT.isInteger() &&((PartVT.isInteger() && ValueVT.isInteger() && "Do not know what to expand to!") ? static_cast<void> ( 0) : __assert_fail ("PartVT.isInteger() && ValueVT.isInteger() && \"Do not know what to expand to!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 423, __PRETTY_FUNCTION__)) | |||
423 | "Do not know what to expand to!")((PartVT.isInteger() && ValueVT.isInteger() && "Do not know what to expand to!") ? static_cast<void> ( 0) : __assert_fail ("PartVT.isInteger() && ValueVT.isInteger() && \"Do not know what to expand to!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 423, __PRETTY_FUNCTION__)); | |||
424 | unsigned RoundParts = 1 << Log2_32(NumParts); | |||
425 | unsigned RoundBits = RoundParts * PartBits; | |||
426 | unsigned OddParts = NumParts - RoundParts; | |||
427 | SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val, | |||
428 | DAG.getIntPtrConstant(RoundBits)); | |||
429 | getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V); | |||
430 | ||||
431 | if (TLI.isBigEndian()) | |||
432 | // The odd parts were reversed by getCopyToParts - unreverse them. | |||
433 | std::reverse(Parts + RoundParts, Parts + NumParts); | |||
434 | ||||
435 | NumParts = RoundParts; | |||
436 | ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); | |||
437 | Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); | |||
438 | } | |||
439 | ||||
440 | // The number of parts is a power of 2. Repeatedly bisect the value using | |||
441 | // EXTRACT_ELEMENT. | |||
442 | Parts[0] = DAG.getNode(ISD::BITCAST, DL, | |||
443 | EVT::getIntegerVT(*DAG.getContext(), | |||
444 | ValueVT.getSizeInBits()), | |||
445 | Val); | |||
446 | ||||
447 | for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { | |||
448 | for (unsigned i = 0; i < NumParts; i += StepSize) { | |||
449 | unsigned ThisBits = StepSize * PartBits / 2; | |||
450 | EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); | |||
451 | SDValue &Part0 = Parts[i]; | |||
452 | SDValue &Part1 = Parts[i+StepSize/2]; | |||
453 | ||||
454 | Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, | |||
455 | ThisVT, Part0, DAG.getIntPtrConstant(1)); | |||
456 | Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, | |||
457 | ThisVT, Part0, DAG.getIntPtrConstant(0)); | |||
458 | ||||
459 | if (ThisBits == PartBits && ThisVT != PartVT) { | |||
460 | Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0); | |||
461 | Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1); | |||
462 | } | |||
463 | } | |||
464 | } | |||
465 | ||||
466 | if (TLI.isBigEndian()) | |||
467 | std::reverse(Parts, Parts + OrigNumParts); | |||
468 | } | |||
469 | ||||
470 | ||||
471 | /// getCopyToPartsVector - Create a series of nodes that contain the specified | |||
472 | /// value split into legal parts. | |||
473 | static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL, | |||
474 | SDValue Val, SDValue *Parts, unsigned NumParts, | |||
475 | MVT PartVT, const Value *V) { | |||
476 | EVT ValueVT = Val.getValueType(); | |||
477 | assert(ValueVT.isVector() && "Not a vector")((ValueVT.isVector() && "Not a vector") ? static_cast <void> (0) : __assert_fail ("ValueVT.isVector() && \"Not a vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 477, __PRETTY_FUNCTION__)); | |||
478 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
479 | ||||
480 | if (NumParts == 1) { | |||
481 | EVT PartEVT = PartVT; | |||
482 | if (PartEVT == ValueVT) { | |||
483 | // Nothing to do. | |||
484 | } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { | |||
485 | // Bitconvert vector->vector case. | |||
486 | Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); | |||
487 | } else if (PartVT.isVector() && | |||
488 | PartEVT.getVectorElementType() == ValueVT.getVectorElementType() && | |||
489 | PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) { | |||
490 | EVT ElementVT = PartVT.getVectorElementType(); | |||
491 | // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in | |||
492 | // undef elements. | |||
493 | SmallVector<SDValue, 16> Ops; | |||
494 | for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i) | |||
495 | Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, | |||
496 | ElementVT, Val, DAG.getConstant(i, | |||
497 | TLI.getVectorIdxTy()))); | |||
498 | ||||
499 | for (unsigned i = ValueVT.getVectorNumElements(), | |||
500 | e = PartVT.getVectorNumElements(); i != e; ++i) | |||
501 | Ops.push_back(DAG.getUNDEF(ElementVT)); | |||
502 | ||||
503 | Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops); | |||
504 | ||||
505 | // FIXME: Use CONCAT for 2x -> 4x. | |||
506 | ||||
507 | //SDValue UndefElts = DAG.getUNDEF(VectorTy); | |||
508 | //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts); | |||
509 | } else if (PartVT.isVector() && | |||
510 | PartEVT.getVectorElementType().bitsGE( | |||
511 | ValueVT.getVectorElementType()) && | |||
512 | PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) { | |||
513 | ||||
514 | // Promoted vector extract | |||
515 | bool Smaller = PartEVT.bitsLE(ValueVT); | |||
516 | Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), | |||
517 | DL, PartVT, Val); | |||
518 | } else{ | |||
519 | // Vector -> scalar conversion. | |||
520 | assert(ValueVT.getVectorNumElements() == 1 &&((ValueVT.getVectorNumElements() == 1 && "Only trivial vector-to-scalar conversions should get here!" ) ? static_cast<void> (0) : __assert_fail ("ValueVT.getVectorNumElements() == 1 && \"Only trivial vector-to-scalar conversions should get here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 521, __PRETTY_FUNCTION__)) | |||
521 | "Only trivial vector-to-scalar conversions should get here!")((ValueVT.getVectorNumElements() == 1 && "Only trivial vector-to-scalar conversions should get here!" ) ? static_cast<void> (0) : __assert_fail ("ValueVT.getVectorNumElements() == 1 && \"Only trivial vector-to-scalar conversions should get here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 521, __PRETTY_FUNCTION__)); | |||
522 | Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, | |||
523 | PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy())); | |||
524 | ||||
525 | bool Smaller = ValueVT.bitsLE(PartVT); | |||
526 | Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), | |||
527 | DL, PartVT, Val); | |||
528 | } | |||
529 | ||||
530 | Parts[0] = Val; | |||
531 | return; | |||
532 | } | |||
533 | ||||
534 | // Handle a multi-element vector. | |||
535 | EVT IntermediateVT; | |||
536 | MVT RegisterVT; | |||
537 | unsigned NumIntermediates; | |||
538 | unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, | |||
539 | IntermediateVT, | |||
540 | NumIntermediates, RegisterVT); | |||
541 | unsigned NumElements = ValueVT.getVectorNumElements(); | |||
542 | ||||
543 | assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!")((NumRegs == NumParts && "Part count doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("NumRegs == NumParts && \"Part count doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 543, __PRETTY_FUNCTION__)); | |||
544 | NumParts = NumRegs; // Silence a compiler warning. | |||
545 | assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!")((RegisterVT == PartVT && "Part type doesn't match vector breakdown!" ) ? static_cast<void> (0) : __assert_fail ("RegisterVT == PartVT && \"Part type doesn't match vector breakdown!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 545, __PRETTY_FUNCTION__)); | |||
546 | ||||
547 | // Split the vector into intermediate operands. | |||
548 | SmallVector<SDValue, 8> Ops(NumIntermediates); | |||
549 | for (unsigned i = 0; i != NumIntermediates; ++i) { | |||
550 | if (IntermediateVT.isVector()) | |||
551 | Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, | |||
552 | IntermediateVT, Val, | |||
553 | DAG.getConstant(i * (NumElements / NumIntermediates), | |||
554 | TLI.getVectorIdxTy())); | |||
555 | else | |||
556 | Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, | |||
557 | IntermediateVT, Val, | |||
558 | DAG.getConstant(i, TLI.getVectorIdxTy())); | |||
559 | } | |||
560 | ||||
561 | // Split the intermediate operands into legal parts. | |||
562 | if (NumParts == NumIntermediates) { | |||
563 | // If the register was not expanded, promote or copy the value, | |||
564 | // as appropriate. | |||
565 | for (unsigned i = 0; i != NumParts; ++i) | |||
566 | getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V); | |||
567 | } else if (NumParts > 0) { | |||
568 | // If the intermediate type was expanded, split each the value into | |||
569 | // legal parts. | |||
570 | assert(NumIntermediates != 0 && "division by zero")((NumIntermediates != 0 && "division by zero") ? static_cast <void> (0) : __assert_fail ("NumIntermediates != 0 && \"division by zero\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 570, __PRETTY_FUNCTION__)); | |||
571 | assert(NumParts % NumIntermediates == 0 &&((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 572, __PRETTY_FUNCTION__)) | |||
572 | "Must expand into a divisible number of parts!")((NumParts % NumIntermediates == 0 && "Must expand into a divisible number of parts!" ) ? static_cast<void> (0) : __assert_fail ("NumParts % NumIntermediates == 0 && \"Must expand into a divisible number of parts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 572, __PRETTY_FUNCTION__)); | |||
573 | unsigned Factor = NumParts / NumIntermediates; | |||
574 | for (unsigned i = 0; i != NumIntermediates; ++i) | |||
575 | getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V); | |||
576 | } | |||
577 | } | |||
578 | ||||
579 | namespace { | |||
580 | /// RegsForValue - This struct represents the registers (physical or virtual) | |||
581 | /// that a particular set of values is assigned, and the type information | |||
582 | /// about the value. The most common situation is to represent one value at a | |||
583 | /// time, but struct or array values are handled element-wise as multiple | |||
584 | /// values. The splitting of aggregates is performed recursively, so that we | |||
585 | /// never have aggregate-typed registers. The values at this point do not | |||
586 | /// necessarily have legal types, so each value may require one or more | |||
587 | /// registers of some legal type. | |||
588 | /// | |||
589 | struct RegsForValue { | |||
590 | /// ValueVTs - The value types of the values, which may not be legal, and | |||
591 | /// may need be promoted or synthesized from one or more registers. | |||
592 | /// | |||
593 | SmallVector<EVT, 4> ValueVTs; | |||
594 | ||||
595 | /// RegVTs - The value types of the registers. This is the same size as | |||
596 | /// ValueVTs and it records, for each value, what the type of the assigned | |||
597 | /// register or registers are. (Individual values are never synthesized | |||
598 | /// from more than one type of register.) | |||
599 | /// | |||
600 | /// With virtual registers, the contents of RegVTs is redundant with TLI's | |||
601 | /// getRegisterType member function, however when with physical registers | |||
602 | /// it is necessary to have a separate record of the types. | |||
603 | /// | |||
604 | SmallVector<MVT, 4> RegVTs; | |||
605 | ||||
606 | /// Regs - This list holds the registers assigned to the values. | |||
607 | /// Each legal or promoted value requires one register, and each | |||
608 | /// expanded value requires multiple registers. | |||
609 | /// | |||
610 | SmallVector<unsigned, 4> Regs; | |||
611 | ||||
612 | RegsForValue() {} | |||
613 | ||||
614 | RegsForValue(const SmallVector<unsigned, 4> ®s, | |||
615 | MVT regvt, EVT valuevt) | |||
616 | : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} | |||
617 | ||||
618 | RegsForValue(LLVMContext &Context, const TargetLowering &tli, | |||
619 | unsigned Reg, Type *Ty) { | |||
620 | ComputeValueVTs(tli, Ty, ValueVTs); | |||
621 | ||||
622 | for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { | |||
623 | EVT ValueVT = ValueVTs[Value]; | |||
624 | unsigned NumRegs = tli.getNumRegisters(Context, ValueVT); | |||
625 | MVT RegisterVT = tli.getRegisterType(Context, ValueVT); | |||
626 | for (unsigned i = 0; i != NumRegs; ++i) | |||
627 | Regs.push_back(Reg + i); | |||
628 | RegVTs.push_back(RegisterVT); | |||
629 | Reg += NumRegs; | |||
630 | } | |||
631 | } | |||
632 | ||||
633 | /// append - Add the specified values to this one. | |||
634 | void append(const RegsForValue &RHS) { | |||
635 | ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end()); | |||
636 | RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end()); | |||
637 | Regs.append(RHS.Regs.begin(), RHS.Regs.end()); | |||
638 | } | |||
639 | ||||
640 | /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from | |||
641 | /// this value and returns the result as a ValueVTs value. This uses | |||
642 | /// Chain/Flag as the input and updates them for the output Chain/Flag. | |||
643 | /// If the Flag pointer is NULL, no flag is used. | |||
644 | SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, | |||
645 | SDLoc dl, | |||
646 | SDValue &Chain, SDValue *Flag, | |||
647 | const Value *V = nullptr) const; | |||
648 | ||||
649 | /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the | |||
650 | /// specified value into the registers specified by this object. This uses | |||
651 | /// Chain/Flag as the input and updates them for the output Chain/Flag. | |||
652 | /// If the Flag pointer is NULL, no flag is used. | |||
653 | void | |||
654 | getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl, SDValue &Chain, | |||
655 | SDValue *Flag, const Value *V, | |||
656 | ISD::NodeType PreferredExtendType = ISD::ANY_EXTEND) const; | |||
657 | ||||
658 | /// AddInlineAsmOperands - Add this value to the specified inlineasm node | |||
659 | /// operand list. This adds the code marker, matching input operand index | |||
660 | /// (if applicable), and includes the number of values added into it. | |||
661 | void AddInlineAsmOperands(unsigned Kind, | |||
662 | bool HasMatching, unsigned MatchingIdx, | |||
663 | SelectionDAG &DAG, | |||
664 | std::vector<SDValue> &Ops) const; | |||
665 | }; | |||
666 | } | |||
667 | ||||
668 | /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from | |||
669 | /// this value and returns the result as a ValueVT value. This uses | |||
670 | /// Chain/Flag as the input and updates them for the output Chain/Flag. | |||
671 | /// If the Flag pointer is NULL, no flag is used. | |||
672 | SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, | |||
673 | FunctionLoweringInfo &FuncInfo, | |||
674 | SDLoc dl, | |||
675 | SDValue &Chain, SDValue *Flag, | |||
676 | const Value *V) const { | |||
677 | // A Value with type {} or [0 x %t] needs no registers. | |||
678 | if (ValueVTs.empty()) | |||
679 | return SDValue(); | |||
680 | ||||
681 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
682 | ||||
683 | // Assemble the legal parts into the final values. | |||
684 | SmallVector<SDValue, 4> Values(ValueVTs.size()); | |||
685 | SmallVector<SDValue, 8> Parts; | |||
686 | for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { | |||
687 | // Copy the legal parts from the registers. | |||
688 | EVT ValueVT = ValueVTs[Value]; | |||
689 | unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT); | |||
690 | MVT RegisterVT = RegVTs[Value]; | |||
691 | ||||
692 | Parts.resize(NumRegs); | |||
693 | for (unsigned i = 0; i != NumRegs; ++i) { | |||
694 | SDValue P; | |||
695 | if (!Flag) { | |||
696 | P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT); | |||
697 | } else { | |||
698 | P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag); | |||
699 | *Flag = P.getValue(2); | |||
700 | } | |||
701 | ||||
702 | Chain = P.getValue(1); | |||
703 | Parts[i] = P; | |||
704 | ||||
705 | // If the source register was virtual and if we know something about it, | |||
706 | // add an assert node. | |||
707 | if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) || | |||
708 | !RegisterVT.isInteger() || RegisterVT.isVector()) | |||
709 | continue; | |||
710 | ||||
711 | const FunctionLoweringInfo::LiveOutInfo *LOI = | |||
712 | FuncInfo.GetLiveOutRegInfo(Regs[Part+i]); | |||
713 | if (!LOI) | |||
714 | continue; | |||
715 | ||||
716 | unsigned RegSize = RegisterVT.getSizeInBits(); | |||
717 | unsigned NumSignBits = LOI->NumSignBits; | |||
718 | unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes(); | |||
719 | ||||
720 | if (NumZeroBits == RegSize) { | |||
721 | // The current value is a zero. | |||
722 | // Explicitly express that as it would be easier for | |||
723 | // optimizations to kick in. | |||
724 | Parts[i] = DAG.getConstant(0, RegisterVT); | |||
725 | continue; | |||
726 | } | |||
727 | ||||
728 | // FIXME: We capture more information than the dag can represent. For | |||
729 | // now, just use the tightest assertzext/assertsext possible. | |||
730 | bool isSExt = true; | |||
731 | EVT FromVT(MVT::Other); | |||
732 | if (NumSignBits == RegSize) | |||
733 | isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1 | |||
734 | else if (NumZeroBits >= RegSize-1) | |||
735 | isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1 | |||
736 | else if (NumSignBits > RegSize-8) | |||
737 | isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8 | |||
738 | else if (NumZeroBits >= RegSize-8) | |||
739 | isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8 | |||
740 | else if (NumSignBits > RegSize-16) | |||
741 | isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16 | |||
742 | else if (NumZeroBits >= RegSize-16) | |||
743 | isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16 | |||
744 | else if (NumSignBits > RegSize-32) | |||
745 | isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32 | |||
746 | else if (NumZeroBits >= RegSize-32) | |||
747 | isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32 | |||
748 | else | |||
749 | continue; | |||
750 | ||||
751 | // Add an assertion node. | |||
752 | assert(FromVT != MVT::Other)((FromVT != MVT::Other) ? static_cast<void> (0) : __assert_fail ("FromVT != MVT::Other", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 752, __PRETTY_FUNCTION__)); | |||
753 | Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl, | |||
754 | RegisterVT, P, DAG.getValueType(FromVT)); | |||
755 | } | |||
756 | ||||
757 | Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), | |||
758 | NumRegs, RegisterVT, ValueVT, V); | |||
759 | Part += NumRegs; | |||
760 | Parts.clear(); | |||
761 | } | |||
762 | ||||
763 | return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values); | |||
764 | } | |||
765 | ||||
766 | /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the | |||
767 | /// specified value into the registers specified by this object. This uses | |||
768 | /// Chain/Flag as the input and updates them for the output Chain/Flag. | |||
769 | /// If the Flag pointer is NULL, no flag is used. | |||
770 | void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl, | |||
771 | SDValue &Chain, SDValue *Flag, const Value *V, | |||
772 | ISD::NodeType PreferredExtendType) const { | |||
773 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
774 | ISD::NodeType ExtendKind = PreferredExtendType; | |||
775 | ||||
776 | // Get the list of the values's legal parts. | |||
777 | unsigned NumRegs = Regs.size(); | |||
778 | SmallVector<SDValue, 8> Parts(NumRegs); | |||
779 | for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { | |||
780 | EVT ValueVT = ValueVTs[Value]; | |||
781 | unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); | |||
782 | MVT RegisterVT = RegVTs[Value]; | |||
783 | ||||
784 | if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT)) | |||
785 | ExtendKind = ISD::ZERO_EXTEND; | |||
786 | ||||
787 | getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), | |||
788 | &Parts[Part], NumParts, RegisterVT, V, ExtendKind); | |||
789 | Part += NumParts; | |||
790 | } | |||
791 | ||||
792 | // Copy the parts into the registers. | |||
793 | SmallVector<SDValue, 8> Chains(NumRegs); | |||
794 | for (unsigned i = 0; i != NumRegs; ++i) { | |||
795 | SDValue Part; | |||
796 | if (!Flag) { | |||
797 | Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]); | |||
798 | } else { | |||
799 | Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag); | |||
800 | *Flag = Part.getValue(1); | |||
801 | } | |||
802 | ||||
803 | Chains[i] = Part.getValue(0); | |||
804 | } | |||
805 | ||||
806 | if (NumRegs == 1 || Flag) | |||
807 | // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is | |||
808 | // flagged to it. That is the CopyToReg nodes and the user are considered | |||
809 | // a single scheduling unit. If we create a TokenFactor and return it as | |||
810 | // chain, then the TokenFactor is both a predecessor (operand) of the | |||
811 | // user as well as a successor (the TF operands are flagged to the user). | |||
812 | // c1, f1 = CopyToReg | |||
813 | // c2, f2 = CopyToReg | |||
814 | // c3 = TokenFactor c1, c2 | |||
815 | // ... | |||
816 | // = op c3, ..., f2 | |||
817 | Chain = Chains[NumRegs-1]; | |||
818 | else | |||
819 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); | |||
820 | } | |||
821 | ||||
822 | /// AddInlineAsmOperands - Add this value to the specified inlineasm node | |||
823 | /// operand list. This adds the code marker and includes the number of | |||
824 | /// values added into it. | |||
825 | void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching, | |||
826 | unsigned MatchingIdx, | |||
827 | SelectionDAG &DAG, | |||
828 | std::vector<SDValue> &Ops) const { | |||
829 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
830 | ||||
831 | unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size()); | |||
832 | if (HasMatching) | |||
833 | Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx); | |||
834 | else if (!Regs.empty() && | |||
835 | TargetRegisterInfo::isVirtualRegister(Regs.front())) { | |||
836 | // Put the register class of the virtual registers in the flag word. That | |||
837 | // way, later passes can recompute register class constraints for inline | |||
838 | // assembly as well as normal instructions. | |||
839 | // Don't do this for tied operands that can use the regclass information | |||
840 | // from the def. | |||
841 | const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo(); | |||
842 | const TargetRegisterClass *RC = MRI.getRegClass(Regs.front()); | |||
843 | Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID()); | |||
844 | } | |||
845 | ||||
846 | SDValue Res = DAG.getTargetConstant(Flag, MVT::i32); | |||
847 | Ops.push_back(Res); | |||
848 | ||||
849 | unsigned SP = TLI.getStackPointerRegisterToSaveRestore(); | |||
850 | for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) { | |||
851 | unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]); | |||
852 | MVT RegisterVT = RegVTs[Value]; | |||
853 | for (unsigned i = 0; i != NumRegs; ++i) { | |||
854 | assert(Reg < Regs.size() && "Mismatch in # registers expected")((Reg < Regs.size() && "Mismatch in # registers expected" ) ? static_cast<void> (0) : __assert_fail ("Reg < Regs.size() && \"Mismatch in # registers expected\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 854, __PRETTY_FUNCTION__)); | |||
855 | unsigned TheReg = Regs[Reg++]; | |||
856 | Ops.push_back(DAG.getRegister(TheReg, RegisterVT)); | |||
857 | ||||
858 | if (TheReg == SP && Code == InlineAsm::Kind_Clobber) { | |||
859 | // If we clobbered the stack pointer, MFI should know about it. | |||
860 | assert(DAG.getMachineFunction().getFrameInfo()->((DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust ()) ? static_cast<void> (0) : __assert_fail ("DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 861, __PRETTY_FUNCTION__)) | |||
861 | hasInlineAsmWithSPAdjust())((DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust ()) ? static_cast<void> (0) : __assert_fail ("DAG.getMachineFunction().getFrameInfo()-> hasInlineAsmWithSPAdjust()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 861, __PRETTY_FUNCTION__)); | |||
862 | } | |||
863 | } | |||
864 | } | |||
865 | } | |||
866 | ||||
867 | void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa, | |||
868 | const TargetLibraryInfo *li) { | |||
869 | AA = &aa; | |||
870 | GFI = gfi; | |||
871 | LibInfo = li; | |||
872 | DL = DAG.getTarget().getDataLayout(); | |||
873 | Context = DAG.getContext(); | |||
874 | LPadToCallSiteMap.clear(); | |||
875 | } | |||
876 | ||||
877 | /// clear - Clear out the current SelectionDAG and the associated | |||
878 | /// state and prepare this SelectionDAGBuilder object to be used | |||
879 | /// for a new block. This doesn't clear out information about | |||
880 | /// additional blocks that are needed to complete switch lowering | |||
881 | /// or PHI node updating; that information is cleared out as it is | |||
882 | /// consumed. | |||
883 | void SelectionDAGBuilder::clear() { | |||
884 | NodeMap.clear(); | |||
885 | UnusedArgNodeMap.clear(); | |||
886 | PendingLoads.clear(); | |||
887 | PendingExports.clear(); | |||
888 | CurInst = nullptr; | |||
889 | HasTailCall = false; | |||
890 | SDNodeOrder = LowestSDNodeOrder; | |||
891 | StatepointLowering.clear(); | |||
892 | } | |||
893 | ||||
894 | /// clearDanglingDebugInfo - Clear the dangling debug information | |||
895 | /// map. This function is separated from the clear so that debug | |||
896 | /// information that is dangling in a basic block can be properly | |||
897 | /// resolved in a different basic block. This allows the | |||
898 | /// SelectionDAG to resolve dangling debug information attached | |||
899 | /// to PHI nodes. | |||
900 | void SelectionDAGBuilder::clearDanglingDebugInfo() { | |||
901 | DanglingDebugInfoMap.clear(); | |||
902 | } | |||
903 | ||||
904 | /// getRoot - Return the current virtual root of the Selection DAG, | |||
905 | /// flushing any PendingLoad items. This must be done before emitting | |||
906 | /// a store or any other node that may need to be ordered after any | |||
907 | /// prior load instructions. | |||
908 | /// | |||
909 | SDValue SelectionDAGBuilder::getRoot() { | |||
910 | if (PendingLoads.empty()) | |||
911 | return DAG.getRoot(); | |||
912 | ||||
913 | if (PendingLoads.size() == 1) { | |||
914 | SDValue Root = PendingLoads[0]; | |||
915 | DAG.setRoot(Root); | |||
916 | PendingLoads.clear(); | |||
917 | return Root; | |||
918 | } | |||
919 | ||||
920 | // Otherwise, we have to make a token factor node. | |||
921 | SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, | |||
922 | PendingLoads); | |||
923 | PendingLoads.clear(); | |||
924 | DAG.setRoot(Root); | |||
925 | return Root; | |||
926 | } | |||
927 | ||||
928 | /// getControlRoot - Similar to getRoot, but instead of flushing all the | |||
929 | /// PendingLoad items, flush all the PendingExports items. It is necessary | |||
930 | /// to do this before emitting a terminator instruction. | |||
931 | /// | |||
932 | SDValue SelectionDAGBuilder::getControlRoot() { | |||
933 | SDValue Root = DAG.getRoot(); | |||
934 | ||||
935 | if (PendingExports.empty()) | |||
936 | return Root; | |||
937 | ||||
938 | // Turn all of the CopyToReg chains into one factored node. | |||
939 | if (Root.getOpcode() != ISD::EntryToken) { | |||
940 | unsigned i = 0, e = PendingExports.size(); | |||
941 | for (; i != e; ++i) { | |||
942 | assert(PendingExports[i].getNode()->getNumOperands() > 1)((PendingExports[i].getNode()->getNumOperands() > 1) ? static_cast <void> (0) : __assert_fail ("PendingExports[i].getNode()->getNumOperands() > 1" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 942, __PRETTY_FUNCTION__)); | |||
943 | if (PendingExports[i].getNode()->getOperand(0) == Root) | |||
944 | break; // Don't add the root if we already indirectly depend on it. | |||
945 | } | |||
946 | ||||
947 | if (i == e) | |||
948 | PendingExports.push_back(Root); | |||
949 | } | |||
950 | ||||
951 | Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, | |||
952 | PendingExports); | |||
953 | PendingExports.clear(); | |||
954 | DAG.setRoot(Root); | |||
955 | return Root; | |||
956 | } | |||
957 | ||||
958 | void SelectionDAGBuilder::visit(const Instruction &I) { | |||
959 | // Set up outgoing PHI node register values before emitting the terminator. | |||
960 | if (isa<TerminatorInst>(&I)) | |||
961 | HandlePHINodesInSuccessorBlocks(I.getParent()); | |||
962 | ||||
963 | ++SDNodeOrder; | |||
964 | ||||
965 | CurInst = &I; | |||
966 | ||||
967 | visit(I.getOpcode(), I); | |||
968 | ||||
969 | if (!isa<TerminatorInst>(&I) && !HasTailCall) | |||
970 | CopyToExportRegsIfNeeded(&I); | |||
971 | ||||
972 | CurInst = nullptr; | |||
973 | } | |||
974 | ||||
975 | void SelectionDAGBuilder::visitPHI(const PHINode &) { | |||
976 | llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!")::llvm::llvm_unreachable_internal("SelectionDAGBuilder shouldn't visit PHI nodes!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 976); | |||
977 | } | |||
978 | ||||
979 | void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { | |||
980 | // Note: this doesn't use InstVisitor, because it has to work with | |||
981 | // ConstantExpr's in addition to instructions. | |||
982 | switch (Opcode) { | |||
983 | default: llvm_unreachable("Unknown instruction type encountered!")::llvm::llvm_unreachable_internal("Unknown instruction type encountered!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 983); | |||
984 | // Build the switch statement using the Instruction.def file. | |||
985 | #define HANDLE_INST(NUM, OPCODE, CLASS) \ | |||
986 | case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break; | |||
987 | #include "llvm/IR/Instruction.def" | |||
988 | } | |||
989 | } | |||
990 | ||||
991 | // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, | |||
992 | // generate the debug data structures now that we've seen its definition. | |||
993 | void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V, | |||
994 | SDValue Val) { | |||
995 | DanglingDebugInfo &DDI = DanglingDebugInfoMap[V]; | |||
996 | if (DDI.getDI()) { | |||
997 | const DbgValueInst *DI = DDI.getDI(); | |||
998 | DebugLoc dl = DDI.getdl(); | |||
999 | unsigned DbgSDNodeOrder = DDI.getSDNodeOrder(); | |||
1000 | MDNode *Variable = DI->getVariable(); | |||
1001 | MDNode *Expr = DI->getExpression(); | |||
1002 | uint64_t Offset = DI->getOffset(); | |||
1003 | // A dbg.value for an alloca is always indirect. | |||
1004 | bool IsIndirect = isa<AllocaInst>(V) || Offset != 0; | |||
1005 | SDDbgValue *SDV; | |||
1006 | if (Val.getNode()) { | |||
1007 | if (!EmitFuncArgumentDbgValue(V, Variable, Expr, Offset, IsIndirect, | |||
1008 | Val)) { | |||
1009 | SDV = DAG.getDbgValue(Variable, Expr, Val.getNode(), Val.getResNo(), | |||
1010 | IsIndirect, Offset, dl, DbgSDNodeOrder); | |||
1011 | DAG.AddDbgValue(SDV, Val.getNode(), false); | |||
1012 | } | |||
1013 | } else | |||
1014 | DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << *DI << "\n"; } } while (0); | |||
1015 | DanglingDebugInfoMap[V] = DanglingDebugInfo(); | |||
1016 | } | |||
1017 | } | |||
1018 | ||||
1019 | /// getValue - Return an SDValue for the given Value. | |||
1020 | SDValue SelectionDAGBuilder::getValue(const Value *V) { | |||
1021 | // If we already have an SDValue for this value, use it. It's important | |||
1022 | // to do this first, so that we don't create a CopyFromReg if we already | |||
1023 | // have a regular SDValue. | |||
1024 | SDValue &N = NodeMap[V]; | |||
1025 | if (N.getNode()) return N; | |||
1026 | ||||
1027 | // If there's a virtual register allocated and initialized for this | |||
1028 | // value, use it. | |||
1029 | DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V); | |||
1030 | if (It != FuncInfo.ValueMap.end()) { | |||
1031 | unsigned InReg = It->second; | |||
1032 | RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(), InReg, | |||
1033 | V->getType()); | |||
1034 | SDValue Chain = DAG.getEntryNode(); | |||
1035 | N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V); | |||
1036 | resolveDanglingDebugInfo(V, N); | |||
1037 | return N; | |||
1038 | } | |||
1039 | ||||
1040 | // Otherwise create a new SDValue and remember it. | |||
1041 | SDValue Val = getValueImpl(V); | |||
1042 | NodeMap[V] = Val; | |||
1043 | resolveDanglingDebugInfo(V, Val); | |||
1044 | return Val; | |||
1045 | } | |||
1046 | ||||
1047 | /// getNonRegisterValue - Return an SDValue for the given Value, but | |||
1048 | /// don't look in FuncInfo.ValueMap for a virtual register. | |||
1049 | SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) { | |||
1050 | // If we already have an SDValue for this value, use it. | |||
1051 | SDValue &N = NodeMap[V]; | |||
1052 | if (N.getNode()) return N; | |||
1053 | ||||
1054 | // Otherwise create a new SDValue and remember it. | |||
1055 | SDValue Val = getValueImpl(V); | |||
1056 | NodeMap[V] = Val; | |||
1057 | resolveDanglingDebugInfo(V, Val); | |||
1058 | return Val; | |||
1059 | } | |||
1060 | ||||
1061 | /// getValueImpl - Helper function for getValue and getNonRegisterValue. | |||
1062 | /// Create an SDValue for the given value. | |||
1063 | SDValue SelectionDAGBuilder::getValueImpl(const Value *V) { | |||
1064 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1065 | ||||
1066 | if (const Constant *C = dyn_cast<Constant>(V)) { | |||
1067 | EVT VT = TLI.getValueType(V->getType(), true); | |||
1068 | ||||
1069 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(C)) | |||
1070 | return DAG.getConstant(*CI, VT); | |||
1071 | ||||
1072 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) | |||
1073 | return DAG.getGlobalAddress(GV, getCurSDLoc(), VT); | |||
1074 | ||||
1075 | if (isa<ConstantPointerNull>(C)) { | |||
1076 | unsigned AS = V->getType()->getPointerAddressSpace(); | |||
1077 | return DAG.getConstant(0, TLI.getPointerTy(AS)); | |||
1078 | } | |||
1079 | ||||
1080 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) | |||
1081 | return DAG.getConstantFP(*CFP, VT); | |||
1082 | ||||
1083 | if (isa<UndefValue>(C) && !V->getType()->isAggregateType()) | |||
1084 | return DAG.getUNDEF(VT); | |||
1085 | ||||
1086 | if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { | |||
1087 | visit(CE->getOpcode(), *CE); | |||
1088 | SDValue N1 = NodeMap[V]; | |||
1089 | assert(N1.getNode() && "visit didn't populate the NodeMap!")((N1.getNode() && "visit didn't populate the NodeMap!" ) ? static_cast<void> (0) : __assert_fail ("N1.getNode() && \"visit didn't populate the NodeMap!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1089, __PRETTY_FUNCTION__)); | |||
1090 | return N1; | |||
1091 | } | |||
1092 | ||||
1093 | if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) { | |||
1094 | SmallVector<SDValue, 4> Constants; | |||
1095 | for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end(); | |||
1096 | OI != OE; ++OI) { | |||
1097 | SDNode *Val = getValue(*OI).getNode(); | |||
1098 | // If the operand is an empty aggregate, there are no values. | |||
1099 | if (!Val) continue; | |||
1100 | // Add each leaf value from the operand to the Constants list | |||
1101 | // to form a flattened list of all the values. | |||
1102 | for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) | |||
1103 | Constants.push_back(SDValue(Val, i)); | |||
1104 | } | |||
1105 | ||||
1106 | return DAG.getMergeValues(Constants, getCurSDLoc()); | |||
1107 | } | |||
1108 | ||||
1109 | if (const ConstantDataSequential *CDS = | |||
1110 | dyn_cast<ConstantDataSequential>(C)) { | |||
1111 | SmallVector<SDValue, 4> Ops; | |||
1112 | for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { | |||
1113 | SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode(); | |||
1114 | // Add each leaf value from the operand to the Constants list | |||
1115 | // to form a flattened list of all the values. | |||
1116 | for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) | |||
1117 | Ops.push_back(SDValue(Val, i)); | |||
1118 | } | |||
1119 | ||||
1120 | if (isa<ArrayType>(CDS->getType())) | |||
1121 | return DAG.getMergeValues(Ops, getCurSDLoc()); | |||
1122 | return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), | |||
1123 | VT, Ops); | |||
1124 | } | |||
1125 | ||||
1126 | if (C->getType()->isStructTy() || C->getType()->isArrayTy()) { | |||
1127 | assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&(((isa<ConstantAggregateZero>(C) || isa<UndefValue> (C)) && "Unknown struct or array constant!") ? static_cast <void> (0) : __assert_fail ("(isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && \"Unknown struct or array constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1128, __PRETTY_FUNCTION__)) | |||
1128 | "Unknown struct or array constant!")(((isa<ConstantAggregateZero>(C) || isa<UndefValue> (C)) && "Unknown struct or array constant!") ? static_cast <void> (0) : __assert_fail ("(isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && \"Unknown struct or array constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1128, __PRETTY_FUNCTION__)); | |||
1129 | ||||
1130 | SmallVector<EVT, 4> ValueVTs; | |||
1131 | ComputeValueVTs(TLI, C->getType(), ValueVTs); | |||
1132 | unsigned NumElts = ValueVTs.size(); | |||
1133 | if (NumElts == 0) | |||
1134 | return SDValue(); // empty struct | |||
1135 | SmallVector<SDValue, 4> Constants(NumElts); | |||
1136 | for (unsigned i = 0; i != NumElts; ++i) { | |||
1137 | EVT EltVT = ValueVTs[i]; | |||
1138 | if (isa<UndefValue>(C)) | |||
1139 | Constants[i] = DAG.getUNDEF(EltVT); | |||
1140 | else if (EltVT.isFloatingPoint()) | |||
1141 | Constants[i] = DAG.getConstantFP(0, EltVT); | |||
1142 | else | |||
1143 | Constants[i] = DAG.getConstant(0, EltVT); | |||
1144 | } | |||
1145 | ||||
1146 | return DAG.getMergeValues(Constants, getCurSDLoc()); | |||
1147 | } | |||
1148 | ||||
1149 | if (const BlockAddress *BA = dyn_cast<BlockAddress>(C)) | |||
1150 | return DAG.getBlockAddress(BA, VT); | |||
1151 | ||||
1152 | VectorType *VecTy = cast<VectorType>(V->getType()); | |||
1153 | unsigned NumElements = VecTy->getNumElements(); | |||
1154 | ||||
1155 | // Now that we know the number and type of the elements, get that number of | |||
1156 | // elements into the Ops array based on what kind of constant it is. | |||
1157 | SmallVector<SDValue, 16> Ops; | |||
1158 | if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) { | |||
1159 | for (unsigned i = 0; i != NumElements; ++i) | |||
1160 | Ops.push_back(getValue(CV->getOperand(i))); | |||
1161 | } else { | |||
1162 | assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!")((isa<ConstantAggregateZero>(C) && "Unknown vector constant!" ) ? static_cast<void> (0) : __assert_fail ("isa<ConstantAggregateZero>(C) && \"Unknown vector constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1162, __PRETTY_FUNCTION__)); | |||
1163 | EVT EltVT = TLI.getValueType(VecTy->getElementType()); | |||
1164 | ||||
1165 | SDValue Op; | |||
1166 | if (EltVT.isFloatingPoint()) | |||
1167 | Op = DAG.getConstantFP(0, EltVT); | |||
1168 | else | |||
1169 | Op = DAG.getConstant(0, EltVT); | |||
1170 | Ops.assign(NumElements, Op); | |||
1171 | } | |||
1172 | ||||
1173 | // Create a BUILD_VECTOR node. | |||
1174 | return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops); | |||
1175 | } | |||
1176 | ||||
1177 | // If this is a static alloca, generate it as the frameindex instead of | |||
1178 | // computation. | |||
1179 | if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { | |||
1180 | DenseMap<const AllocaInst*, int>::iterator SI = | |||
1181 | FuncInfo.StaticAllocaMap.find(AI); | |||
1182 | if (SI != FuncInfo.StaticAllocaMap.end()) | |||
1183 | return DAG.getFrameIndex(SI->second, TLI.getPointerTy()); | |||
1184 | } | |||
1185 | ||||
1186 | // If this is an instruction which fast-isel has deferred, select it now. | |||
1187 | if (const Instruction *Inst = dyn_cast<Instruction>(V)) { | |||
1188 | unsigned InReg = FuncInfo.InitializeRegForValue(Inst); | |||
1189 | RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType()); | |||
1190 | SDValue Chain = DAG.getEntryNode(); | |||
1191 | return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V); | |||
1192 | } | |||
1193 | ||||
1194 | llvm_unreachable("Can't get register for value!")::llvm::llvm_unreachable_internal("Can't get register for value!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1194); | |||
1195 | } | |||
1196 | ||||
1197 | void SelectionDAGBuilder::visitRet(const ReturnInst &I) { | |||
1198 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1199 | SDValue Chain = getControlRoot(); | |||
1200 | SmallVector<ISD::OutputArg, 8> Outs; | |||
1201 | SmallVector<SDValue, 8> OutVals; | |||
1202 | ||||
1203 | if (!FuncInfo.CanLowerReturn) { | |||
1204 | unsigned DemoteReg = FuncInfo.DemoteRegister; | |||
1205 | const Function *F = I.getParent()->getParent(); | |||
1206 | ||||
1207 | // Emit a store of the return value through the virtual register. | |||
1208 | // Leave Outs empty so that LowerReturn won't try to load return | |||
1209 | // registers the usual way. | |||
1210 | SmallVector<EVT, 1> PtrValueVTs; | |||
1211 | ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()), | |||
1212 | PtrValueVTs); | |||
1213 | ||||
1214 | SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]); | |||
1215 | SDValue RetOp = getValue(I.getOperand(0)); | |||
1216 | ||||
1217 | SmallVector<EVT, 4> ValueVTs; | |||
1218 | SmallVector<uint64_t, 4> Offsets; | |||
1219 | ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets); | |||
1220 | unsigned NumValues = ValueVTs.size(); | |||
1221 | ||||
1222 | SmallVector<SDValue, 4> Chains(NumValues); | |||
1223 | for (unsigned i = 0; i != NumValues; ++i) { | |||
1224 | SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), | |||
1225 | RetPtr.getValueType(), RetPtr, | |||
1226 | DAG.getIntPtrConstant(Offsets[i])); | |||
1227 | Chains[i] = | |||
1228 | DAG.getStore(Chain, getCurSDLoc(), | |||
1229 | SDValue(RetOp.getNode(), RetOp.getResNo() + i), | |||
1230 | // FIXME: better loc info would be nice. | |||
1231 | Add, MachinePointerInfo(), false, false, 0); | |||
1232 | } | |||
1233 | ||||
1234 | Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), | |||
1235 | MVT::Other, Chains); | |||
1236 | } else if (I.getNumOperands() != 0) { | |||
1237 | SmallVector<EVT, 4> ValueVTs; | |||
1238 | ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs); | |||
1239 | unsigned NumValues = ValueVTs.size(); | |||
1240 | if (NumValues) { | |||
1241 | SDValue RetOp = getValue(I.getOperand(0)); | |||
1242 | ||||
1243 | const Function *F = I.getParent()->getParent(); | |||
1244 | ||||
1245 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND; | |||
1246 | if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, | |||
1247 | Attribute::SExt)) | |||
1248 | ExtendKind = ISD::SIGN_EXTEND; | |||
1249 | else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, | |||
1250 | Attribute::ZExt)) | |||
1251 | ExtendKind = ISD::ZERO_EXTEND; | |||
1252 | ||||
1253 | LLVMContext &Context = F->getContext(); | |||
1254 | bool RetInReg = F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, | |||
1255 | Attribute::InReg); | |||
1256 | ||||
1257 | for (unsigned j = 0; j != NumValues; ++j) { | |||
1258 | EVT VT = ValueVTs[j]; | |||
1259 | ||||
1260 | if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) | |||
1261 | VT = TLI.getTypeForExtArgOrReturn(Context, VT, ExtendKind); | |||
1262 | ||||
1263 | unsigned NumParts = TLI.getNumRegisters(Context, VT); | |||
1264 | MVT PartVT = TLI.getRegisterType(Context, VT); | |||
1265 | SmallVector<SDValue, 4> Parts(NumParts); | |||
1266 | getCopyToParts(DAG, getCurSDLoc(), | |||
1267 | SDValue(RetOp.getNode(), RetOp.getResNo() + j), | |||
1268 | &Parts[0], NumParts, PartVT, &I, ExtendKind); | |||
1269 | ||||
1270 | // 'inreg' on function refers to return value | |||
1271 | ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); | |||
1272 | if (RetInReg) | |||
1273 | Flags.setInReg(); | |||
1274 | ||||
1275 | // Propagate extension type if any | |||
1276 | if (ExtendKind == ISD::SIGN_EXTEND) | |||
1277 | Flags.setSExt(); | |||
1278 | else if (ExtendKind == ISD::ZERO_EXTEND) | |||
1279 | Flags.setZExt(); | |||
1280 | ||||
1281 | for (unsigned i = 0; i < NumParts; ++i) { | |||
1282 | Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(), | |||
1283 | VT, /*isfixed=*/true, 0, 0)); | |||
1284 | OutVals.push_back(Parts[i]); | |||
1285 | } | |||
1286 | } | |||
1287 | } | |||
1288 | } | |||
1289 | ||||
1290 | bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg(); | |||
1291 | CallingConv::ID CallConv = | |||
1292 | DAG.getMachineFunction().getFunction()->getCallingConv(); | |||
1293 | Chain = DAG.getTargetLoweringInfo().LowerReturn( | |||
1294 | Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG); | |||
1295 | ||||
1296 | // Verify that the target's LowerReturn behaved as expected. | |||
1297 | assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&((Chain.getNode() && Chain.getValueType() == MVT::Other && "LowerReturn didn't return a valid chain!") ? static_cast <void> (0) : __assert_fail ("Chain.getNode() && Chain.getValueType() == MVT::Other && \"LowerReturn didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1298, __PRETTY_FUNCTION__)) | |||
1298 | "LowerReturn didn't return a valid chain!")((Chain.getNode() && Chain.getValueType() == MVT::Other && "LowerReturn didn't return a valid chain!") ? static_cast <void> (0) : __assert_fail ("Chain.getNode() && Chain.getValueType() == MVT::Other && \"LowerReturn didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1298, __PRETTY_FUNCTION__)); | |||
1299 | ||||
1300 | // Update the DAG with the new chain value resulting from return lowering. | |||
1301 | DAG.setRoot(Chain); | |||
1302 | } | |||
1303 | ||||
1304 | /// CopyToExportRegsIfNeeded - If the given value has virtual registers | |||
1305 | /// created for it, emit nodes to copy the value into the virtual | |||
1306 | /// registers. | |||
1307 | void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) { | |||
1308 | // Skip empty types | |||
1309 | if (V->getType()->isEmptyTy()) | |||
1310 | return; | |||
1311 | ||||
1312 | DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); | |||
1313 | if (VMI != FuncInfo.ValueMap.end()) { | |||
1314 | assert(!V->use_empty() && "Unused value assigned virtual registers!")((!V->use_empty() && "Unused value assigned virtual registers!" ) ? static_cast<void> (0) : __assert_fail ("!V->use_empty() && \"Unused value assigned virtual registers!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1314, __PRETTY_FUNCTION__)); | |||
1315 | CopyValueToVirtualRegister(V, VMI->second); | |||
1316 | } | |||
1317 | } | |||
1318 | ||||
1319 | /// ExportFromCurrentBlock - If this condition isn't known to be exported from | |||
1320 | /// the current basic block, add it to ValueMap now so that we'll get a | |||
1321 | /// CopyTo/FromReg. | |||
1322 | void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) { | |||
1323 | // No need to export constants. | |||
1324 | if (!isa<Instruction>(V) && !isa<Argument>(V)) return; | |||
1325 | ||||
1326 | // Already exported? | |||
1327 | if (FuncInfo.isExportedInst(V)) return; | |||
1328 | ||||
1329 | unsigned Reg = FuncInfo.InitializeRegForValue(V); | |||
1330 | CopyValueToVirtualRegister(V, Reg); | |||
1331 | } | |||
1332 | ||||
1333 | bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V, | |||
1334 | const BasicBlock *FromBB) { | |||
1335 | // The operands of the setcc have to be in this block. We don't know | |||
1336 | // how to export them from some other block. | |||
1337 | if (const Instruction *VI = dyn_cast<Instruction>(V)) { | |||
1338 | // Can export from current BB. | |||
1339 | if (VI->getParent() == FromBB) | |||
1340 | return true; | |||
1341 | ||||
1342 | // Is already exported, noop. | |||
1343 | return FuncInfo.isExportedInst(V); | |||
1344 | } | |||
1345 | ||||
1346 | // If this is an argument, we can export it if the BB is the entry block or | |||
1347 | // if it is already exported. | |||
1348 | if (isa<Argument>(V)) { | |||
1349 | if (FromBB == &FromBB->getParent()->getEntryBlock()) | |||
1350 | return true; | |||
1351 | ||||
1352 | // Otherwise, can only export this if it is already exported. | |||
1353 | return FuncInfo.isExportedInst(V); | |||
1354 | } | |||
1355 | ||||
1356 | // Otherwise, constants can always be exported. | |||
1357 | return true; | |||
1358 | } | |||
1359 | ||||
1360 | /// Return branch probability calculated by BranchProbabilityInfo for IR blocks. | |||
1361 | uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src, | |||
1362 | const MachineBasicBlock *Dst) const { | |||
1363 | BranchProbabilityInfo *BPI = FuncInfo.BPI; | |||
1364 | if (!BPI) | |||
1365 | return 0; | |||
1366 | const BasicBlock *SrcBB = Src->getBasicBlock(); | |||
1367 | const BasicBlock *DstBB = Dst->getBasicBlock(); | |||
1368 | return BPI->getEdgeWeight(SrcBB, DstBB); | |||
1369 | } | |||
1370 | ||||
1371 | void SelectionDAGBuilder:: | |||
1372 | addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst, | |||
1373 | uint32_t Weight /* = 0 */) { | |||
1374 | if (!Weight) | |||
1375 | Weight = getEdgeWeight(Src, Dst); | |||
1376 | Src->addSuccessor(Dst, Weight); | |||
1377 | } | |||
1378 | ||||
1379 | ||||
1380 | static bool InBlock(const Value *V, const BasicBlock *BB) { | |||
1381 | if (const Instruction *I = dyn_cast<Instruction>(V)) | |||
1382 | return I->getParent() == BB; | |||
1383 | return true; | |||
1384 | } | |||
1385 | ||||
1386 | /// EmitBranchForMergedCondition - Helper method for FindMergedConditions. | |||
1387 | /// This function emits a branch and is used at the leaves of an OR or an | |||
1388 | /// AND operator tree. | |||
1389 | /// | |||
1390 | void | |||
1391 | SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond, | |||
1392 | MachineBasicBlock *TBB, | |||
1393 | MachineBasicBlock *FBB, | |||
1394 | MachineBasicBlock *CurBB, | |||
1395 | MachineBasicBlock *SwitchBB, | |||
1396 | uint32_t TWeight, | |||
1397 | uint32_t FWeight) { | |||
1398 | const BasicBlock *BB = CurBB->getBasicBlock(); | |||
1399 | ||||
1400 | // If the leaf of the tree is a comparison, merge the condition into | |||
1401 | // the caseblock. | |||
1402 | if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) { | |||
1403 | // The operands of the cmp have to be in this block. We don't know | |||
1404 | // how to export them from some other block. If this is the first block | |||
1405 | // of the sequence, no exporting is needed. | |||
1406 | if (CurBB == SwitchBB || | |||
1407 | (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && | |||
1408 | isExportableFromCurrentBlock(BOp->getOperand(1), BB))) { | |||
1409 | ISD::CondCode Condition; | |||
1410 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { | |||
1411 | Condition = getICmpCondCode(IC->getPredicate()); | |||
1412 | } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { | |||
1413 | Condition = getFCmpCondCode(FC->getPredicate()); | |||
1414 | if (TM.Options.NoNaNsFPMath) | |||
1415 | Condition = getFCmpCodeWithoutNaN(Condition); | |||
1416 | } else { | |||
1417 | (void)Condition; // silence warning. | |||
1418 | llvm_unreachable("Unknown compare instruction")::llvm::llvm_unreachable_internal("Unknown compare instruction" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1418); | |||
1419 | } | |||
1420 | ||||
1421 | CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr, | |||
1422 | TBB, FBB, CurBB, TWeight, FWeight); | |||
1423 | SwitchCases.push_back(CB); | |||
1424 | return; | |||
1425 | } | |||
1426 | } | |||
1427 | ||||
1428 | // Create a CaseBlock record representing this branch. | |||
1429 | CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()), | |||
1430 | nullptr, TBB, FBB, CurBB, TWeight, FWeight); | |||
1431 | SwitchCases.push_back(CB); | |||
1432 | } | |||
1433 | ||||
1434 | /// Scale down both weights to fit into uint32_t. | |||
1435 | static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) { | |||
1436 | uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; | |||
1437 | uint32_t Scale = (NewMax / UINT32_MAX(4294967295U)) + 1; | |||
1438 | NewTrue = NewTrue / Scale; | |||
1439 | NewFalse = NewFalse / Scale; | |||
1440 | } | |||
1441 | ||||
1442 | /// FindMergedConditions - If Cond is an expression like | |||
1443 | void SelectionDAGBuilder::FindMergedConditions(const Value *Cond, | |||
1444 | MachineBasicBlock *TBB, | |||
1445 | MachineBasicBlock *FBB, | |||
1446 | MachineBasicBlock *CurBB, | |||
1447 | MachineBasicBlock *SwitchBB, | |||
1448 | unsigned Opc, uint32_t TWeight, | |||
1449 | uint32_t FWeight) { | |||
1450 | // If this node is not part of the or/and tree, emit it as a branch. | |||
1451 | const Instruction *BOp = dyn_cast<Instruction>(Cond); | |||
1452 | if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || | |||
1453 | (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || | |||
1454 | BOp->getParent() != CurBB->getBasicBlock() || | |||
1455 | !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || | |||
1456 | !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { | |||
1457 | EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB, | |||
1458 | TWeight, FWeight); | |||
1459 | return; | |||
1460 | } | |||
1461 | ||||
1462 | // Create TmpBB after CurBB. | |||
1463 | MachineFunction::iterator BBI = CurBB; | |||
1464 | MachineFunction &MF = DAG.getMachineFunction(); | |||
1465 | MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock()); | |||
1466 | CurBB->getParent()->insert(++BBI, TmpBB); | |||
1467 | ||||
1468 | if (Opc == Instruction::Or) { | |||
1469 | // Codegen X | Y as: | |||
1470 | // BB1: | |||
1471 | // jmp_if_X TBB | |||
1472 | // jmp TmpBB | |||
1473 | // TmpBB: | |||
1474 | // jmp_if_Y TBB | |||
1475 | // jmp FBB | |||
1476 | // | |||
1477 | ||||
1478 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. | |||
1479 | // The requirement is that | |||
1480 | // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) | |||
1481 | // = TrueProb for orignal BB. | |||
1482 | // Assuming the orignal weights are A and B, one choice is to set BB1's | |||
1483 | // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice | |||
1484 | // assumes that | |||
1485 | // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. | |||
1486 | // Another choice is to assume TrueProb for BB1 equals to TrueProb for | |||
1487 | // TmpBB, but the math is more complicated. | |||
1488 | ||||
1489 | uint64_t NewTrueWeight = TWeight; | |||
1490 | uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight; | |||
1491 | ScaleWeights(NewTrueWeight, NewFalseWeight); | |||
1492 | // Emit the LHS condition. | |||
1493 | FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc, | |||
1494 | NewTrueWeight, NewFalseWeight); | |||
1495 | ||||
1496 | NewTrueWeight = TWeight; | |||
1497 | NewFalseWeight = 2 * (uint64_t)FWeight; | |||
1498 | ScaleWeights(NewTrueWeight, NewFalseWeight); | |||
1499 | // Emit the RHS condition into TmpBB. | |||
1500 | FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc, | |||
1501 | NewTrueWeight, NewFalseWeight); | |||
1502 | } else { | |||
1503 | assert(Opc == Instruction::And && "Unknown merge op!")((Opc == Instruction::And && "Unknown merge op!") ? static_cast <void> (0) : __assert_fail ("Opc == Instruction::And && \"Unknown merge op!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1503, __PRETTY_FUNCTION__)); | |||
1504 | // Codegen X & Y as: | |||
1505 | // BB1: | |||
1506 | // jmp_if_X TmpBB | |||
1507 | // jmp FBB | |||
1508 | // TmpBB: | |||
1509 | // jmp_if_Y TBB | |||
1510 | // jmp FBB | |||
1511 | // | |||
1512 | // This requires creation of TmpBB after CurBB. | |||
1513 | ||||
1514 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. | |||
1515 | // The requirement is that | |||
1516 | // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) | |||
1517 | // = FalseProb for orignal BB. | |||
1518 | // Assuming the orignal weights are A and B, one choice is to set BB1's | |||
1519 | // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice | |||
1520 | // assumes that | |||
1521 | // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB. | |||
1522 | ||||
1523 | uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight; | |||
1524 | uint64_t NewFalseWeight = FWeight; | |||
1525 | ScaleWeights(NewTrueWeight, NewFalseWeight); | |||
1526 | // Emit the LHS condition. | |||
1527 | FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc, | |||
1528 | NewTrueWeight, NewFalseWeight); | |||
1529 | ||||
1530 | NewTrueWeight = 2 * (uint64_t)TWeight; | |||
1531 | NewFalseWeight = FWeight; | |||
1532 | ScaleWeights(NewTrueWeight, NewFalseWeight); | |||
1533 | // Emit the RHS condition into TmpBB. | |||
1534 | FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc, | |||
1535 | NewTrueWeight, NewFalseWeight); | |||
1536 | } | |||
1537 | } | |||
1538 | ||||
1539 | /// If the set of cases should be emitted as a series of branches, return true. | |||
1540 | /// If we should emit this as a bunch of and/or'd together conditions, return | |||
1541 | /// false. | |||
1542 | bool | |||
1543 | SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) { | |||
1544 | if (Cases.size() != 2) return true; | |||
1545 | ||||
1546 | // If this is two comparisons of the same values or'd or and'd together, they | |||
1547 | // will get folded into a single comparison, so don't emit two blocks. | |||
1548 | if ((Cases[0].CmpLHS == Cases[1].CmpLHS && | |||
1549 | Cases[0].CmpRHS == Cases[1].CmpRHS) || | |||
1550 | (Cases[0].CmpRHS == Cases[1].CmpLHS && | |||
1551 | Cases[0].CmpLHS == Cases[1].CmpRHS)) { | |||
1552 | return false; | |||
1553 | } | |||
1554 | ||||
1555 | // Handle: (X != null) | (Y != null) --> (X|Y) != 0 | |||
1556 | // Handle: (X == null) & (Y == null) --> (X|Y) == 0 | |||
1557 | if (Cases[0].CmpRHS == Cases[1].CmpRHS && | |||
1558 | Cases[0].CC == Cases[1].CC && | |||
1559 | isa<Constant>(Cases[0].CmpRHS) && | |||
1560 | cast<Constant>(Cases[0].CmpRHS)->isNullValue()) { | |||
1561 | if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB) | |||
1562 | return false; | |||
1563 | if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB) | |||
1564 | return false; | |||
1565 | } | |||
1566 | ||||
1567 | return true; | |||
1568 | } | |||
1569 | ||||
1570 | void SelectionDAGBuilder::visitBr(const BranchInst &I) { | |||
1571 | MachineBasicBlock *BrMBB = FuncInfo.MBB; | |||
1572 | ||||
1573 | // Update machine-CFG edges. | |||
1574 | MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; | |||
1575 | ||||
1576 | // Figure out which block is immediately after the current one. | |||
1577 | MachineBasicBlock *NextBlock = nullptr; | |||
1578 | MachineFunction::iterator BBI = BrMBB; | |||
1579 | if (++BBI != FuncInfo.MF->end()) | |||
1580 | NextBlock = BBI; | |||
1581 | ||||
1582 | if (I.isUnconditional()) { | |||
1583 | // Update machine-CFG edges. | |||
1584 | BrMBB->addSuccessor(Succ0MBB); | |||
1585 | ||||
1586 | // If this is not a fall-through branch or optimizations are switched off, | |||
1587 | // emit the branch. | |||
1588 | if (Succ0MBB != NextBlock || TM.getOptLevel() == CodeGenOpt::None) | |||
1589 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), | |||
1590 | MVT::Other, getControlRoot(), | |||
1591 | DAG.getBasicBlock(Succ0MBB))); | |||
1592 | ||||
1593 | return; | |||
1594 | } | |||
1595 | ||||
1596 | // If this condition is one of the special cases we handle, do special stuff | |||
1597 | // now. | |||
1598 | const Value *CondVal = I.getCondition(); | |||
1599 | MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; | |||
1600 | ||||
1601 | // If this is a series of conditions that are or'd or and'd together, emit | |||
1602 | // this as a sequence of branches instead of setcc's with and/or operations. | |||
1603 | // As long as jumps are not expensive, this should improve performance. | |||
1604 | // For example, instead of something like: | |||
1605 | // cmp A, B | |||
1606 | // C = seteq | |||
1607 | // cmp D, E | |||
1608 | // F = setle | |||
1609 | // or C, F | |||
1610 | // jnz foo | |||
1611 | // Emit: | |||
1612 | // cmp A, B | |||
1613 | // je foo | |||
1614 | // cmp D, E | |||
1615 | // jle foo | |||
1616 | // | |||
1617 | if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { | |||
1618 | if (!DAG.getTargetLoweringInfo().isJumpExpensive() && | |||
1619 | BOp->hasOneUse() && (BOp->getOpcode() == Instruction::And || | |||
1620 | BOp->getOpcode() == Instruction::Or)) { | |||
1621 | FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB, | |||
1622 | BOp->getOpcode(), getEdgeWeight(BrMBB, Succ0MBB), | |||
1623 | getEdgeWeight(BrMBB, Succ1MBB)); | |||
1624 | // If the compares in later blocks need to use values not currently | |||
1625 | // exported from this block, export them now. This block should always | |||
1626 | // be the first entry. | |||
1627 | assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!")((SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!" ) ? static_cast<void> (0) : __assert_fail ("SwitchCases[0].ThisBB == BrMBB && \"Unexpected lowering!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1627, __PRETTY_FUNCTION__)); | |||
1628 | ||||
1629 | // Allow some cases to be rejected. | |||
1630 | if (ShouldEmitAsBranches(SwitchCases)) { | |||
1631 | for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { | |||
1632 | ExportFromCurrentBlock(SwitchCases[i].CmpLHS); | |||
1633 | ExportFromCurrentBlock(SwitchCases[i].CmpRHS); | |||
1634 | } | |||
1635 | ||||
1636 | // Emit the branch for this block. | |||
1637 | visitSwitchCase(SwitchCases[0], BrMBB); | |||
1638 | SwitchCases.erase(SwitchCases.begin()); | |||
1639 | return; | |||
1640 | } | |||
1641 | ||||
1642 | // Okay, we decided not to do this, remove any inserted MBB's and clear | |||
1643 | // SwitchCases. | |||
1644 | for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) | |||
1645 | FuncInfo.MF->erase(SwitchCases[i].ThisBB); | |||
1646 | ||||
1647 | SwitchCases.clear(); | |||
1648 | } | |||
1649 | } | |||
1650 | ||||
1651 | // Create a CaseBlock record representing this branch. | |||
1652 | CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()), | |||
1653 | nullptr, Succ0MBB, Succ1MBB, BrMBB); | |||
1654 | ||||
1655 | // Use visitSwitchCase to actually insert the fast branch sequence for this | |||
1656 | // cond branch. | |||
1657 | visitSwitchCase(CB, BrMBB); | |||
1658 | } | |||
1659 | ||||
1660 | /// visitSwitchCase - Emits the necessary code to represent a single node in | |||
1661 | /// the binary search tree resulting from lowering a switch instruction. | |||
1662 | void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB, | |||
1663 | MachineBasicBlock *SwitchBB) { | |||
1664 | SDValue Cond; | |||
1665 | SDValue CondLHS = getValue(CB.CmpLHS); | |||
1666 | SDLoc dl = getCurSDLoc(); | |||
1667 | ||||
1668 | // Build the setcc now. | |||
1669 | if (!CB.CmpMHS) { | |||
1670 | // Fold "(X == true)" to X and "(X == false)" to !X to | |||
1671 | // handle common cases produced by branch lowering. | |||
1672 | if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) && | |||
1673 | CB.CC == ISD::SETEQ) | |||
1674 | Cond = CondLHS; | |||
1675 | else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) && | |||
1676 | CB.CC == ISD::SETEQ) { | |||
1677 | SDValue True = DAG.getConstant(1, CondLHS.getValueType()); | |||
1678 | Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True); | |||
1679 | } else | |||
1680 | Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); | |||
1681 | } else { | |||
1682 | assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now")((CB.CC == ISD::SETLE && "Can handle only LE ranges now" ) ? static_cast<void> (0) : __assert_fail ("CB.CC == ISD::SETLE && \"Can handle only LE ranges now\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1682, __PRETTY_FUNCTION__)); | |||
1683 | ||||
1684 | const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue(); | |||
1685 | const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue(); | |||
1686 | ||||
1687 | SDValue CmpOp = getValue(CB.CmpMHS); | |||
1688 | EVT VT = CmpOp.getValueType(); | |||
1689 | ||||
1690 | if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) { | |||
1691 | Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT), | |||
1692 | ISD::SETLE); | |||
1693 | } else { | |||
1694 | SDValue SUB = DAG.getNode(ISD::SUB, dl, | |||
1695 | VT, CmpOp, DAG.getConstant(Low, VT)); | |||
1696 | Cond = DAG.getSetCC(dl, MVT::i1, SUB, | |||
1697 | DAG.getConstant(High-Low, VT), ISD::SETULE); | |||
1698 | } | |||
1699 | } | |||
1700 | ||||
1701 | // Update successor info | |||
1702 | addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight); | |||
1703 | // TrueBB and FalseBB are always different unless the incoming IR is | |||
1704 | // degenerate. This only happens when running llc on weird IR. | |||
1705 | if (CB.TrueBB != CB.FalseBB) | |||
1706 | addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight); | |||
1707 | ||||
1708 | // Set NextBlock to be the MBB immediately after the current one, if any. | |||
1709 | // This is used to avoid emitting unnecessary branches to the next block. | |||
1710 | MachineBasicBlock *NextBlock = nullptr; | |||
1711 | MachineFunction::iterator BBI = SwitchBB; | |||
1712 | if (++BBI != FuncInfo.MF->end()) | |||
1713 | NextBlock = BBI; | |||
1714 | ||||
1715 | // If the lhs block is the next block, invert the condition so that we can | |||
1716 | // fall through to the lhs instead of the rhs block. | |||
1717 | if (CB.TrueBB == NextBlock) { | |||
1718 | std::swap(CB.TrueBB, CB.FalseBB); | |||
1719 | SDValue True = DAG.getConstant(1, Cond.getValueType()); | |||
1720 | Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True); | |||
1721 | } | |||
1722 | ||||
1723 | SDValue BrCond = DAG.getNode(ISD::BRCOND, dl, | |||
1724 | MVT::Other, getControlRoot(), Cond, | |||
1725 | DAG.getBasicBlock(CB.TrueBB)); | |||
1726 | ||||
1727 | // Insert the false branch. Do this even if it's a fall through branch, | |||
1728 | // this makes it easier to do DAG optimizations which require inverting | |||
1729 | // the branch condition. | |||
1730 | BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond, | |||
1731 | DAG.getBasicBlock(CB.FalseBB)); | |||
1732 | ||||
1733 | DAG.setRoot(BrCond); | |||
1734 | } | |||
1735 | ||||
1736 | /// visitJumpTable - Emit JumpTable node in the current MBB | |||
1737 | void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) { | |||
1738 | // Emit the code for the jump table | |||
1739 | assert(JT.Reg != -1U && "Should lower JT Header first!")((JT.Reg != -1U && "Should lower JT Header first!") ? static_cast<void> (0) : __assert_fail ("JT.Reg != -1U && \"Should lower JT Header first!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 1739, __PRETTY_FUNCTION__)); | |||
1740 | EVT PTy = DAG.getTargetLoweringInfo().getPointerTy(); | |||
1741 | SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(), | |||
1742 | JT.Reg, PTy); | |||
1743 | SDValue Table = DAG.getJumpTable(JT.JTI, PTy); | |||
1744 | SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(), | |||
1745 | MVT::Other, Index.getValue(1), | |||
1746 | Table, Index); | |||
1747 | DAG.setRoot(BrJumpTable); | |||
1748 | } | |||
1749 | ||||
1750 | /// visitJumpTableHeader - This function emits necessary code to produce index | |||
1751 | /// in the JumpTable from switch case. | |||
1752 | void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT, | |||
1753 | JumpTableHeader &JTH, | |||
1754 | MachineBasicBlock *SwitchBB) { | |||
1755 | // Subtract the lowest switch case value from the value being switched on and | |||
1756 | // conditional branch to default mbb if the result is greater than the | |||
1757 | // difference between smallest and largest cases. | |||
1758 | SDValue SwitchOp = getValue(JTH.SValue); | |||
1759 | EVT VT = SwitchOp.getValueType(); | |||
1760 | SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp, | |||
1761 | DAG.getConstant(JTH.First, VT)); | |||
1762 | ||||
1763 | // The SDNode we just created, which holds the value being switched on minus | |||
1764 | // the smallest case value, needs to be copied to a virtual register so it | |||
1765 | // can be used as an index into the jump table in a subsequent basic block. | |||
1766 | // This value may be smaller or larger than the target's pointer type, and | |||
1767 | // therefore require extension or truncating. | |||
1768 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1769 | SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI.getPointerTy()); | |||
1770 | ||||
1771 | unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy()); | |||
1772 | SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(), | |||
1773 | JumpTableReg, SwitchOp); | |||
1774 | JT.Reg = JumpTableReg; | |||
1775 | ||||
1776 | // Emit the range check for the jump table, and branch to the default block | |||
1777 | // for the switch statement if the value being switched on exceeds the largest | |||
1778 | // case in the switch. | |||
1779 | SDValue CMP = | |||
1780 | DAG.getSetCC(getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), | |||
1781 | Sub.getValueType()), | |||
1782 | Sub, DAG.getConstant(JTH.Last - JTH.First, VT), ISD::SETUGT); | |||
1783 | ||||
1784 | // Set NextBlock to be the MBB immediately after the current one, if any. | |||
1785 | // This is used to avoid emitting unnecessary branches to the next block. | |||
1786 | MachineBasicBlock *NextBlock = nullptr; | |||
1787 | MachineFunction::iterator BBI = SwitchBB; | |||
1788 | ||||
1789 | if (++BBI != FuncInfo.MF->end()) | |||
1790 | NextBlock = BBI; | |||
1791 | ||||
1792 | SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(), | |||
1793 | MVT::Other, CopyTo, CMP, | |||
1794 | DAG.getBasicBlock(JT.Default)); | |||
1795 | ||||
1796 | if (JT.MBB != NextBlock) | |||
1797 | BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond, | |||
1798 | DAG.getBasicBlock(JT.MBB)); | |||
1799 | ||||
1800 | DAG.setRoot(BrCond); | |||
1801 | } | |||
1802 | ||||
1803 | /// Codegen a new tail for a stack protector check ParentMBB which has had its | |||
1804 | /// tail spliced into a stack protector check success bb. | |||
1805 | /// | |||
1806 | /// For a high level explanation of how this fits into the stack protector | |||
1807 | /// generation see the comment on the declaration of class | |||
1808 | /// StackProtectorDescriptor. | |||
1809 | void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD, | |||
1810 | MachineBasicBlock *ParentBB) { | |||
1811 | ||||
1812 | // First create the loads to the guard/stack slot for the comparison. | |||
1813 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1814 | EVT PtrTy = TLI.getPointerTy(); | |||
1815 | ||||
1816 | MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo(); | |||
1817 | int FI = MFI->getStackProtectorIndex(); | |||
1818 | ||||
1819 | const Value *IRGuard = SPD.getGuard(); | |||
1820 | SDValue GuardPtr = getValue(IRGuard); | |||
1821 | SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy); | |||
1822 | ||||
1823 | unsigned Align = | |||
1824 | TLI.getDataLayout()->getPrefTypeAlignment(IRGuard->getType()); | |||
1825 | ||||
1826 | SDValue Guard; | |||
1827 | ||||
1828 | // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the | |||
1829 | // guard value from the virtual register holding the value. Otherwise, emit a | |||
1830 | // volatile load to retrieve the stack guard value. | |||
1831 | unsigned GuardReg = SPD.getGuardReg(); | |||
1832 | ||||
1833 | if (GuardReg && TLI.useLoadStackGuardNode()) | |||
1834 | Guard = DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), GuardReg, | |||
1835 | PtrTy); | |||
1836 | else | |||
1837 | Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(), | |||
1838 | GuardPtr, MachinePointerInfo(IRGuard, 0), | |||
1839 | true, false, false, Align); | |||
1840 | ||||
1841 | SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(), | |||
1842 | StackSlotPtr, | |||
1843 | MachinePointerInfo::getFixedStack(FI), | |||
1844 | true, false, false, Align); | |||
1845 | ||||
1846 | // Perform the comparison via a subtract/getsetcc. | |||
1847 | EVT VT = Guard.getValueType(); | |||
1848 | SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot); | |||
1849 | ||||
1850 | SDValue Cmp = | |||
1851 | DAG.getSetCC(getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), | |||
1852 | Sub.getValueType()), | |||
1853 | Sub, DAG.getConstant(0, VT), ISD::SETNE); | |||
1854 | ||||
1855 | // If the sub is not 0, then we know the guard/stackslot do not equal, so | |||
1856 | // branch to failure MBB. | |||
1857 | SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(), | |||
1858 | MVT::Other, StackSlot.getOperand(0), | |||
1859 | Cmp, DAG.getBasicBlock(SPD.getFailureMBB())); | |||
1860 | // Otherwise branch to success MBB. | |||
1861 | SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(), | |||
1862 | MVT::Other, BrCond, | |||
1863 | DAG.getBasicBlock(SPD.getSuccessMBB())); | |||
1864 | ||||
1865 | DAG.setRoot(Br); | |||
1866 | } | |||
1867 | ||||
1868 | /// Codegen the failure basic block for a stack protector check. | |||
1869 | /// | |||
1870 | /// A failure stack protector machine basic block consists simply of a call to | |||
1871 | /// __stack_chk_fail(). | |||
1872 | /// | |||
1873 | /// For a high level explanation of how this fits into the stack protector | |||
1874 | /// generation see the comment on the declaration of class | |||
1875 | /// StackProtectorDescriptor. | |||
1876 | void | |||
1877 | SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) { | |||
1878 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1879 | SDValue Chain = | |||
1880 | TLI.makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, MVT::isVoid, | |||
1881 | nullptr, 0, false, getCurSDLoc(), false, false).second; | |||
1882 | DAG.setRoot(Chain); | |||
1883 | } | |||
1884 | ||||
1885 | /// visitBitTestHeader - This function emits necessary code to produce value | |||
1886 | /// suitable for "bit tests" | |||
1887 | void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B, | |||
1888 | MachineBasicBlock *SwitchBB) { | |||
1889 | // Subtract the minimum value | |||
1890 | SDValue SwitchOp = getValue(B.SValue); | |||
1891 | EVT VT = SwitchOp.getValueType(); | |||
1892 | SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp, | |||
1893 | DAG.getConstant(B.First, VT)); | |||
1894 | ||||
1895 | // Check range | |||
1896 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1897 | SDValue RangeCmp = | |||
1898 | DAG.getSetCC(getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), | |||
1899 | Sub.getValueType()), | |||
1900 | Sub, DAG.getConstant(B.Range, VT), ISD::SETUGT); | |||
1901 | ||||
1902 | // Determine the type of the test operands. | |||
1903 | bool UsePtrType = false; | |||
1904 | if (!TLI.isTypeLegal(VT)) | |||
1905 | UsePtrType = true; | |||
1906 | else { | |||
1907 | for (unsigned i = 0, e = B.Cases.size(); i != e; ++i) | |||
1908 | if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) { | |||
1909 | // Switch table case range are encoded into series of masks. | |||
1910 | // Just use pointer type, it's guaranteed to fit. | |||
1911 | UsePtrType = true; | |||
1912 | break; | |||
1913 | } | |||
1914 | } | |||
1915 | if (UsePtrType) { | |||
1916 | VT = TLI.getPointerTy(); | |||
1917 | Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT); | |||
1918 | } | |||
1919 | ||||
1920 | B.RegVT = VT.getSimpleVT(); | |||
1921 | B.Reg = FuncInfo.CreateReg(B.RegVT); | |||
1922 | SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(), | |||
1923 | B.Reg, Sub); | |||
1924 | ||||
1925 | // Set NextBlock to be the MBB immediately after the current one, if any. | |||
1926 | // This is used to avoid emitting unnecessary branches to the next block. | |||
1927 | MachineBasicBlock *NextBlock = nullptr; | |||
1928 | MachineFunction::iterator BBI = SwitchBB; | |||
1929 | if (++BBI != FuncInfo.MF->end()) | |||
1930 | NextBlock = BBI; | |||
1931 | ||||
1932 | MachineBasicBlock* MBB = B.Cases[0].ThisBB; | |||
1933 | ||||
1934 | addSuccessorWithWeight(SwitchBB, B.Default); | |||
1935 | addSuccessorWithWeight(SwitchBB, MBB); | |||
1936 | ||||
1937 | SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(), | |||
1938 | MVT::Other, CopyTo, RangeCmp, | |||
1939 | DAG.getBasicBlock(B.Default)); | |||
1940 | ||||
1941 | if (MBB != NextBlock) | |||
1942 | BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo, | |||
1943 | DAG.getBasicBlock(MBB)); | |||
1944 | ||||
1945 | DAG.setRoot(BrRange); | |||
1946 | } | |||
1947 | ||||
1948 | /// visitBitTestCase - this function produces one "bit test" | |||
1949 | void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB, | |||
1950 | MachineBasicBlock* NextMBB, | |||
1951 | uint32_t BranchWeightToNext, | |||
1952 | unsigned Reg, | |||
1953 | BitTestCase &B, | |||
1954 | MachineBasicBlock *SwitchBB) { | |||
1955 | MVT VT = BB.RegVT; | |||
1956 | SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(), | |||
1957 | Reg, VT); | |||
1958 | SDValue Cmp; | |||
1959 | unsigned PopCount = countPopulation(B.Mask); | |||
1960 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
1961 | if (PopCount == 1) { | |||
1962 | // Testing for a single bit; just compare the shift count with what it | |||
1963 | // would need to be to shift a 1 bit in that position. | |||
1964 | Cmp = DAG.getSetCC( | |||
1965 | getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), VT), ShiftOp, | |||
1966 | DAG.getConstant(countTrailingZeros(B.Mask), VT), ISD::SETEQ); | |||
1967 | } else if (PopCount == BB.Range) { | |||
1968 | // There is only one zero bit in the range, test for it directly. | |||
1969 | Cmp = DAG.getSetCC( | |||
1970 | getCurSDLoc(), TLI.getSetCCResultType(*DAG.getContext(), VT), ShiftOp, | |||
1971 | DAG.getConstant(countTrailingOnes(B.Mask), VT), ISD::SETNE); | |||
1972 | } else { | |||
1973 | // Make desired shift | |||
1974 | SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT, | |||
1975 | DAG.getConstant(1, VT), ShiftOp); | |||
1976 | ||||
1977 | // Emit bit tests and jumps | |||
1978 | SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(), | |||
1979 | VT, SwitchVal, DAG.getConstant(B.Mask, VT)); | |||
1980 | Cmp = DAG.getSetCC(getCurSDLoc(), | |||
1981 | TLI.getSetCCResultType(*DAG.getContext(), VT), AndOp, | |||
1982 | DAG.getConstant(0, VT), ISD::SETNE); | |||
1983 | } | |||
1984 | ||||
1985 | // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight. | |||
1986 | addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight); | |||
1987 | // The branch weight from SwitchBB to NextMBB is BranchWeightToNext. | |||
1988 | addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext); | |||
1989 | ||||
1990 | SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(), | |||
1991 | MVT::Other, getControlRoot(), | |||
1992 | Cmp, DAG.getBasicBlock(B.TargetBB)); | |||
1993 | ||||
1994 | // Set NextBlock to be the MBB immediately after the current one, if any. | |||
1995 | // This is used to avoid emitting unnecessary branches to the next block. | |||
1996 | MachineBasicBlock *NextBlock = nullptr; | |||
1997 | MachineFunction::iterator BBI = SwitchBB; | |||
1998 | if (++BBI != FuncInfo.MF->end()) | |||
1999 | NextBlock = BBI; | |||
2000 | ||||
2001 | if (NextMBB != NextBlock) | |||
2002 | BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd, | |||
2003 | DAG.getBasicBlock(NextMBB)); | |||
2004 | ||||
2005 | DAG.setRoot(BrAnd); | |||
2006 | } | |||
2007 | ||||
2008 | void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) { | |||
2009 | MachineBasicBlock *InvokeMBB = FuncInfo.MBB; | |||
2010 | ||||
2011 | // Retrieve successors. | |||
2012 | MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)]; | |||
2013 | MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)]; | |||
2014 | ||||
2015 | const Value *Callee(I.getCalledValue()); | |||
2016 | const Function *Fn = dyn_cast<Function>(Callee); | |||
2017 | if (isa<InlineAsm>(Callee)) | |||
2018 | visitInlineAsm(&I); | |||
2019 | else if (Fn && Fn->isIntrinsic()) { | |||
2020 | switch (Fn->getIntrinsicID()) { | |||
2021 | default: | |||
2022 | llvm_unreachable("Cannot invoke this intrinsic")::llvm::llvm_unreachable_internal("Cannot invoke this intrinsic" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2022); | |||
2023 | case Intrinsic::donothing: | |||
2024 | // Ignore invokes to @llvm.donothing: jump directly to the next BB. | |||
2025 | break; | |||
2026 | case Intrinsic::experimental_patchpoint_void: | |||
2027 | case Intrinsic::experimental_patchpoint_i64: | |||
2028 | visitPatchpoint(&I, LandingPad); | |||
2029 | break; | |||
2030 | } | |||
2031 | } else | |||
2032 | LowerCallTo(&I, getValue(Callee), false, LandingPad); | |||
2033 | ||||
2034 | // If the value of the invoke is used outside of its defining block, make it | |||
2035 | // available as a virtual register. | |||
2036 | CopyToExportRegsIfNeeded(&I); | |||
2037 | ||||
2038 | // Update successor info | |||
2039 | addSuccessorWithWeight(InvokeMBB, Return); | |||
2040 | addSuccessorWithWeight(InvokeMBB, LandingPad); | |||
2041 | ||||
2042 | // Drop into normal successor. | |||
2043 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), | |||
2044 | MVT::Other, getControlRoot(), | |||
2045 | DAG.getBasicBlock(Return))); | |||
2046 | } | |||
2047 | ||||
2048 | void SelectionDAGBuilder::visitResume(const ResumeInst &RI) { | |||
2049 | llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!")::llvm::llvm_unreachable_internal("SelectionDAGBuilder shouldn't visit resume instructions!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2049); | |||
2050 | } | |||
2051 | ||||
2052 | void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) { | |||
2053 | assert(FuncInfo.MBB->isLandingPad() &&((FuncInfo.MBB->isLandingPad() && "Call to landingpad not in landing pad!" ) ? static_cast<void> (0) : __assert_fail ("FuncInfo.MBB->isLandingPad() && \"Call to landingpad not in landing pad!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2054, __PRETTY_FUNCTION__)) | |||
2054 | "Call to landingpad not in landing pad!")((FuncInfo.MBB->isLandingPad() && "Call to landingpad not in landing pad!" ) ? static_cast<void> (0) : __assert_fail ("FuncInfo.MBB->isLandingPad() && \"Call to landingpad not in landing pad!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2054, __PRETTY_FUNCTION__)); | |||
2055 | ||||
2056 | MachineBasicBlock *MBB = FuncInfo.MBB; | |||
2057 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); | |||
2058 | AddLandingPadInfo(LP, MMI, MBB); | |||
2059 | ||||
2060 | // If there aren't registers to copy the values into (e.g., during SjLj | |||
2061 | // exceptions), then don't bother to create these DAG nodes. | |||
2062 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
2063 | if (TLI.getExceptionPointerRegister() == 0 && | |||
2064 | TLI.getExceptionSelectorRegister() == 0) | |||
2065 | return; | |||
2066 | ||||
2067 | SmallVector<EVT, 2> ValueVTs; | |||
2068 | ComputeValueVTs(TLI, LP.getType(), ValueVTs); | |||
2069 | assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported")((ValueVTs.size() == 2 && "Only two-valued landingpads are supported" ) ? static_cast<void> (0) : __assert_fail ("ValueVTs.size() == 2 && \"Only two-valued landingpads are supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2069, __PRETTY_FUNCTION__)); | |||
2070 | ||||
2071 | // Get the two live-in registers as SDValues. The physregs have already been | |||
2072 | // copied into virtual registers. | |||
2073 | SDValue Ops[2]; | |||
2074 | if (FuncInfo.ExceptionPointerVirtReg) { | |||
2075 | Ops[0] = DAG.getZExtOrTrunc( | |||
2076 | DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), | |||
2077 | FuncInfo.ExceptionPointerVirtReg, TLI.getPointerTy()), | |||
2078 | getCurSDLoc(), ValueVTs[0]); | |||
2079 | } else { | |||
2080 | Ops[0] = DAG.getConstant(0, TLI.getPointerTy()); | |||
2081 | } | |||
2082 | Ops[1] = DAG.getZExtOrTrunc( | |||
2083 | DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), | |||
2084 | FuncInfo.ExceptionSelectorVirtReg, TLI.getPointerTy()), | |||
2085 | getCurSDLoc(), ValueVTs[1]); | |||
2086 | ||||
2087 | // Merge into one. | |||
2088 | SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), | |||
2089 | DAG.getVTList(ValueVTs), Ops); | |||
2090 | setValue(&LP, Res); | |||
2091 | } | |||
2092 | ||||
2093 | unsigned | |||
2094 | SelectionDAGBuilder::visitLandingPadClauseBB(GlobalValue *ClauseGV, | |||
2095 | MachineBasicBlock *LPadBB) { | |||
2096 | SDValue Chain = getControlRoot(); | |||
2097 | ||||
2098 | // Get the typeid that we will dispatch on later. | |||
2099 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
2100 | const TargetRegisterClass *RC = TLI.getRegClassFor(TLI.getPointerTy()); | |||
2101 | unsigned VReg = FuncInfo.MF->getRegInfo().createVirtualRegister(RC); | |||
2102 | unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(ClauseGV); | |||
2103 | SDValue Sel = DAG.getConstant(TypeID, TLI.getPointerTy()); | |||
2104 | Chain = DAG.getCopyToReg(Chain, getCurSDLoc(), VReg, Sel); | |||
2105 | ||||
2106 | // Branch to the main landing pad block. | |||
2107 | MachineBasicBlock *ClauseMBB = FuncInfo.MBB; | |||
2108 | ClauseMBB->addSuccessor(LPadBB); | |||
2109 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, Chain, | |||
2110 | DAG.getBasicBlock(LPadBB))); | |||
2111 | return VReg; | |||
2112 | } | |||
2113 | ||||
2114 | /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for | |||
2115 | /// small case ranges). | |||
2116 | bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR, | |||
2117 | CaseRecVector& WorkList, | |||
2118 | const Value* SV, | |||
2119 | MachineBasicBlock *Default, | |||
2120 | MachineBasicBlock *SwitchBB) { | |||
2121 | // Size is the number of Cases represented by this range. | |||
2122 | size_t Size = CR.Range.second - CR.Range.first; | |||
2123 | if (Size > 3) | |||
2124 | return false; | |||
2125 | ||||
2126 | // Get the MachineFunction which holds the current MBB. This is used when | |||
2127 | // inserting any additional MBBs necessary to represent the switch. | |||
2128 | MachineFunction *CurMF = FuncInfo.MF; | |||
2129 | ||||
2130 | // Figure out which block is immediately after the current one. | |||
2131 | MachineBasicBlock *NextBlock = nullptr; | |||
2132 | MachineFunction::iterator BBI = CR.CaseBB; | |||
2133 | ||||
2134 | if (++BBI != FuncInfo.MF->end()) | |||
2135 | NextBlock = BBI; | |||
2136 | ||||
2137 | BranchProbabilityInfo *BPI = FuncInfo.BPI; | |||
2138 | // If any two of the cases has the same destination, and if one value | |||
2139 | // is the same as the other, but has one bit unset that the other has set, | |||
2140 | // use bit manipulation to do two compares at once. For example: | |||
2141 | // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" | |||
2142 | // TODO: This could be extended to merge any 2 cases in switches with 3 cases. | |||
2143 | // TODO: Handle cases where CR.CaseBB != SwitchBB. | |||
2144 | if (Size == 2 && CR.CaseBB == SwitchBB) { | |||
2145 | Case &Small = *CR.Range.first; | |||
2146 | Case &Big = *(CR.Range.second-1); | |||
2147 | ||||
2148 | if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) { | |||
2149 | const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue(); | |||
2150 | const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue(); | |||
2151 | ||||
2152 | // Check that there is only one bit different. | |||
2153 | if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 && | |||
2154 | (SmallValue | BigValue) == BigValue) { | |||
2155 | // Isolate the common bit. | |||
2156 | APInt CommonBit = BigValue & ~SmallValue; | |||
2157 | assert((SmallValue | CommonBit) == BigValue &&(((SmallValue | CommonBit) == BigValue && CommonBit.countPopulation () == 1 && "Not a common bit?") ? static_cast<void > (0) : __assert_fail ("(SmallValue | CommonBit) == BigValue && CommonBit.countPopulation() == 1 && \"Not a common bit?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2158, __PRETTY_FUNCTION__)) | |||
2158 | CommonBit.countPopulation() == 1 && "Not a common bit?")(((SmallValue | CommonBit) == BigValue && CommonBit.countPopulation () == 1 && "Not a common bit?") ? static_cast<void > (0) : __assert_fail ("(SmallValue | CommonBit) == BigValue && CommonBit.countPopulation() == 1 && \"Not a common bit?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2158, __PRETTY_FUNCTION__)); | |||
2159 | ||||
2160 | SDValue CondLHS = getValue(SV); | |||
2161 | EVT VT = CondLHS.getValueType(); | |||
2162 | SDLoc DL = getCurSDLoc(); | |||
2163 | ||||
2164 | SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS, | |||
2165 | DAG.getConstant(CommonBit, VT)); | |||
2166 | SDValue Cond = DAG.getSetCC(DL, MVT::i1, | |||
2167 | Or, DAG.getConstant(BigValue, VT), | |||
2168 | ISD::SETEQ); | |||
2169 | ||||
2170 | // Update successor info. | |||
2171 | // Both Small and Big will jump to Small.BB, so we sum up the weights. | |||
2172 | addSuccessorWithWeight(SwitchBB, Small.BB, | |||
2173 | Small.ExtraWeight + Big.ExtraWeight); | |||
2174 | addSuccessorWithWeight(SwitchBB, Default, | |||
2175 | // The default destination is the first successor in IR. | |||
2176 | BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0); | |||
2177 | ||||
2178 | // Insert the true branch. | |||
2179 | SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other, | |||
2180 | getControlRoot(), Cond, | |||
2181 | DAG.getBasicBlock(Small.BB)); | |||
2182 | ||||
2183 | // Insert the false branch. | |||
2184 | BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond, | |||
2185 | DAG.getBasicBlock(Default)); | |||
2186 | ||||
2187 | DAG.setRoot(BrCond); | |||
2188 | return true; | |||
2189 | } | |||
2190 | } | |||
2191 | } | |||
2192 | ||||
2193 | // Order cases by weight so the most likely case will be checked first. | |||
2194 | uint32_t UnhandledWeights = 0; | |||
2195 | if (BPI) { | |||
2196 | for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) { | |||
2197 | uint32_t IWeight = I->ExtraWeight; | |||
2198 | UnhandledWeights += IWeight; | |||
2199 | for (CaseItr J = CR.Range.first; J < I; ++J) { | |||
2200 | uint32_t JWeight = J->ExtraWeight; | |||
2201 | if (IWeight > JWeight) | |||
2202 | std::swap(*I, *J); | |||
2203 | } | |||
2204 | } | |||
2205 | } | |||
2206 | // Rearrange the case blocks so that the last one falls through if possible. | |||
2207 | Case &BackCase = *(CR.Range.second-1); | |||
2208 | if (Size > 1 && | |||
2209 | NextBlock && Default != NextBlock && BackCase.BB != NextBlock) { | |||
2210 | // The last case block won't fall through into 'NextBlock' if we emit the | |||
2211 | // branches in this order. See if rearranging a case value would help. | |||
2212 | // We start at the bottom as it's the case with the least weight. | |||
2213 | for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I) | |||
2214 | if (I->BB == NextBlock) { | |||
2215 | std::swap(*I, BackCase); | |||
2216 | break; | |||
2217 | } | |||
2218 | } | |||
2219 | ||||
2220 | // Create a CaseBlock record representing a conditional branch to | |||
2221 | // the Case's target mbb if the value being switched on SV is equal | |||
2222 | // to C. | |||
2223 | MachineBasicBlock *CurBlock = CR.CaseBB; | |||
2224 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { | |||
2225 | MachineBasicBlock *FallThrough; | |||
2226 | if (I != E-1) { | |||
2227 | FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock()); | |||
2228 | CurMF->insert(BBI, FallThrough); | |||
2229 | ||||
2230 | // Put SV in a virtual register to make it available from the new blocks. | |||
2231 | ExportFromCurrentBlock(SV); | |||
2232 | } else { | |||
2233 | // If the last case doesn't match, go to the default block. | |||
2234 | FallThrough = Default; | |||
2235 | } | |||
2236 | ||||
2237 | const Value *RHS, *LHS, *MHS; | |||
2238 | ISD::CondCode CC; | |||
2239 | if (I->High == I->Low) { | |||
2240 | // This is just small small case range :) containing exactly 1 case | |||
2241 | CC = ISD::SETEQ; | |||
2242 | LHS = SV; RHS = I->High; MHS = nullptr; | |||
2243 | } else { | |||
2244 | CC = ISD::SETLE; | |||
2245 | LHS = I->Low; MHS = SV; RHS = I->High; | |||
2246 | } | |||
2247 | ||||
2248 | // The false weight should be sum of all un-handled cases. | |||
2249 | UnhandledWeights -= I->ExtraWeight; | |||
2250 | CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough, | |||
2251 | /* me */ CurBlock, | |||
2252 | /* trueweight */ I->ExtraWeight, | |||
2253 | /* falseweight */ UnhandledWeights); | |||
2254 | ||||
2255 | // If emitting the first comparison, just call visitSwitchCase to emit the | |||
2256 | // code into the current block. Otherwise, push the CaseBlock onto the | |||
2257 | // vector to be later processed by SDISel, and insert the node's MBB | |||
2258 | // before the next MBB. | |||
2259 | if (CurBlock == SwitchBB) | |||
2260 | visitSwitchCase(CB, SwitchBB); | |||
2261 | else | |||
2262 | SwitchCases.push_back(CB); | |||
2263 | ||||
2264 | CurBlock = FallThrough; | |||
2265 | } | |||
2266 | ||||
2267 | return true; | |||
2268 | } | |||
2269 | ||||
2270 | static inline bool areJTsAllowed(const TargetLowering &TLI) { | |||
2271 | return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) || | |||
2272 | TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other); | |||
2273 | } | |||
2274 | ||||
2275 | static APInt ComputeRange(const APInt &First, const APInt &Last) { | |||
2276 | uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1; | |||
2277 | APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth); | |||
2278 | return (LastExt - FirstExt + 1ULL); | |||
2279 | } | |||
2280 | ||||
2281 | /// handleJTSwitchCase - Emit jumptable for current switch case range | |||
2282 | bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR, | |||
2283 | CaseRecVector &WorkList, | |||
2284 | const Value *SV, | |||
2285 | MachineBasicBlock *Default, | |||
2286 | MachineBasicBlock *SwitchBB) { | |||
2287 | Case& FrontCase = *CR.Range.first; | |||
2288 | Case& BackCase = *(CR.Range.second-1); | |||
2289 | ||||
2290 | const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); | |||
2291 | const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); | |||
2292 | ||||
2293 | APInt TSize(First.getBitWidth(), 0); | |||
2294 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) | |||
2295 | TSize += I->size(); | |||
2296 | ||||
2297 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
2298 | if (!areJTsAllowed(TLI) || TSize.ult(TLI.getMinimumJumpTableEntries())) | |||
2299 | return false; | |||
2300 | ||||
2301 | APInt Range = ComputeRange(First, Last); | |||
2302 | // The density is TSize / Range. Require at least 40%. | |||
2303 | // It should not be possible for IntTSize to saturate for sane code, but make | |||
2304 | // sure we handle Range saturation correctly. | |||
2305 | uint64_t IntRange = Range.getLimitedValue(UINT64_MAX(18446744073709551615UL)/10); | |||
2306 | uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX(18446744073709551615UL)/10); | |||
2307 | if (IntTSize * 10 < IntRange * 4) | |||
2308 | return false; | |||
2309 | ||||
2310 | DEBUG(dbgs() << "Lowering jump table\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Lowering jump table\n" << "First entry: " << First << ". Last entry: " << Last << '\n' << "Range: " << Range << ". Size: " << TSize << ".\n\n"; } } while (0) | |||
2311 | << "First entry: " << First << ". Last entry: " << Last << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Lowering jump table\n" << "First entry: " << First << ". Last entry: " << Last << '\n' << "Range: " << Range << ". Size: " << TSize << ".\n\n"; } } while (0) | |||
2312 | << "Range: " << Range << ". Size: " << TSize << ".\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Lowering jump table\n" << "First entry: " << First << ". Last entry: " << Last << '\n' << "Range: " << Range << ". Size: " << TSize << ".\n\n"; } } while (0); | |||
2313 | ||||
2314 | // Get the MachineFunction which holds the current MBB. This is used when | |||
2315 | // inserting any additional MBBs necessary to represent the switch. | |||
2316 | MachineFunction *CurMF = FuncInfo.MF; | |||
2317 | ||||
2318 | // Figure out which block is immediately after the current one. | |||
2319 | MachineFunction::iterator BBI = CR.CaseBB; | |||
2320 | ++BBI; | |||
2321 | ||||
2322 | const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); | |||
2323 | ||||
2324 | // Create a new basic block to hold the code for loading the address | |||
2325 | // of the jump table, and jumping to it. Update successor information; | |||
2326 | // we will either branch to the default case for the switch, or the jump | |||
2327 | // table. | |||
2328 | MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB); | |||
2329 | CurMF->insert(BBI, JumpTableBB); | |||
2330 | ||||
2331 | addSuccessorWithWeight(CR.CaseBB, Default); | |||
2332 | addSuccessorWithWeight(CR.CaseBB, JumpTableBB); | |||
2333 | ||||
2334 | // Build a vector of destination BBs, corresponding to each target | |||
2335 | // of the jump table. If the value of the jump table slot corresponds to | |||
2336 | // a case statement, push the case's BB onto the vector, otherwise, push | |||
2337 | // the default BB. | |||
2338 | std::vector<MachineBasicBlock*> DestBBs; | |||
2339 | APInt TEI = First; | |||
2340 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) { | |||
2341 | const APInt &Low = cast<ConstantInt>(I->Low)->getValue(); | |||
2342 | const APInt &High = cast<ConstantInt>(I->High)->getValue(); | |||
2343 | ||||
2344 | if (Low.sle(TEI) && TEI.sle(High)) { | |||
2345 | DestBBs.push_back(I->BB); | |||
2346 | if (TEI==High) | |||
2347 | ++I; | |||
2348 | } else { | |||
2349 | DestBBs.push_back(Default); | |||
2350 | } | |||
2351 | } | |||
2352 | ||||
2353 | // Calculate weight for each unique destination in CR. | |||
2354 | DenseMap<MachineBasicBlock*, uint32_t> DestWeights; | |||
2355 | if (FuncInfo.BPI) | |||
2356 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { | |||
2357 | DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr = | |||
2358 | DestWeights.find(I->BB); | |||
2359 | if (Itr != DestWeights.end()) | |||
2360 | Itr->second += I->ExtraWeight; | |||
2361 | else | |||
2362 | DestWeights[I->BB] = I->ExtraWeight; | |||
2363 | } | |||
2364 | ||||
2365 | // Update successor info. Add one edge to each unique successor. | |||
2366 | BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs()); | |||
2367 | for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), | |||
2368 | E = DestBBs.end(); I != E; ++I) { | |||
2369 | if (!SuccsHandled[(*I)->getNumber()]) { | |||
2370 | SuccsHandled[(*I)->getNumber()] = true; | |||
2371 | DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr = | |||
2372 | DestWeights.find(*I); | |||
2373 | addSuccessorWithWeight(JumpTableBB, *I, | |||
2374 | Itr != DestWeights.end() ? Itr->second : 0); | |||
2375 | } | |||
2376 | } | |||
2377 | ||||
2378 | // Create a jump table index for this jump table. | |||
2379 | unsigned JTEncoding = TLI.getJumpTableEncoding(); | |||
2380 | unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding) | |||
2381 | ->createJumpTableIndex(DestBBs); | |||
2382 | ||||
2383 | // Set the jump table information so that we can codegen it as a second | |||
2384 | // MachineBasicBlock | |||
2385 | JumpTable JT(-1U, JTI, JumpTableBB, Default); | |||
2386 | JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB)); | |||
2387 | if (CR.CaseBB == SwitchBB) | |||
2388 | visitJumpTableHeader(JT, JTH, SwitchBB); | |||
2389 | ||||
2390 | JTCases.push_back(JumpTableBlock(JTH, JT)); | |||
2391 | return true; | |||
2392 | } | |||
2393 | ||||
2394 | /// handleBTSplitSwitchCase - emit comparison and split binary search tree into | |||
2395 | /// 2 subtrees. | |||
2396 | bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR, | |||
2397 | CaseRecVector& WorkList, | |||
2398 | const Value* SV, | |||
2399 | MachineBasicBlock* SwitchBB) { | |||
2400 | Case& FrontCase = *CR.Range.first; | |||
2401 | Case& BackCase = *(CR.Range.second-1); | |||
2402 | ||||
2403 | // Size is the number of Cases represented by this range. | |||
2404 | unsigned Size = CR.Range.second - CR.Range.first; | |||
2405 | ||||
2406 | const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); | |||
2407 | const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); | |||
2408 | double FMetric = 0; | |||
2409 | CaseItr Pivot = CR.Range.first + Size/2; | |||
2410 | ||||
2411 | // Select optimal pivot, maximizing sum density of LHS and RHS. This will | |||
2412 | // (heuristically) allow us to emit JumpTable's later. | |||
2413 | APInt TSize(First.getBitWidth(), 0); | |||
2414 | for (CaseItr I = CR.Range.first, E = CR.Range.second; | |||
2415 | I!=E; ++I) | |||
2416 | TSize += I->size(); | |||
2417 | ||||
2418 | APInt LSize = FrontCase.size(); | |||
2419 | APInt RSize = TSize-LSize; | |||
2420 | DEBUG(dbgs() << "Selecting best pivot: \n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Selecting best pivot: \n" << "First: " << First << ", Last: " << Last << '\n' << "LSize: " << LSize << ", RSize: " << RSize << '\n'; } } while (0) | |||
2421 | << "First: " << First << ", Last: " << Last <<'\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Selecting best pivot: \n" << "First: " << First << ", Last: " << Last << '\n' << "LSize: " << LSize << ", RSize: " << RSize << '\n'; } } while (0) | |||
2422 | << "LSize: " << LSize << ", RSize: " << RSize << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Selecting best pivot: \n" << "First: " << First << ", Last: " << Last << '\n' << "LSize: " << LSize << ", RSize: " << RSize << '\n'; } } while (0); | |||
2423 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
2424 | for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second; | |||
2425 | J!=E; ++I, ++J) { | |||
2426 | const APInt &LEnd = cast<ConstantInt>(I->High)->getValue(); | |||
2427 | const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue(); | |||
2428 | APInt Range = ComputeRange(LEnd, RBegin); | |||
2429 | assert((Range - 2ULL).isNonNegative() &&(((Range - 2ULL).isNonNegative() && "Invalid case distance" ) ? static_cast<void> (0) : __assert_fail ("(Range - 2ULL).isNonNegative() && \"Invalid case distance\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2430, __PRETTY_FUNCTION__)) | |||
2430 | "Invalid case distance")(((Range - 2ULL).isNonNegative() && "Invalid case distance" ) ? static_cast<void> (0) : __assert_fail ("(Range - 2ULL).isNonNegative() && \"Invalid case distance\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2430, __PRETTY_FUNCTION__)); | |||
2431 | // Use volatile double here to avoid excess precision issues on some hosts, | |||
2432 | // e.g. that use 80-bit X87 registers. | |||
2433 | // Only consider the density of sub-ranges that actually have sufficient | |||
2434 | // entries to be lowered as a jump table. | |||
2435 | volatile double LDensity = | |||
2436 | LSize.ult(TLI.getMinimumJumpTableEntries()) | |||
2437 | ? 0.0 | |||
2438 | : LSize.roundToDouble() / (LEnd - First + 1ULL).roundToDouble(); | |||
2439 | volatile double RDensity = | |||
2440 | RSize.ult(TLI.getMinimumJumpTableEntries()) | |||
2441 | ? 0.0 | |||
2442 | : RSize.roundToDouble() / (Last - RBegin + 1ULL).roundToDouble(); | |||
2443 | volatile double Metric = Range.logBase2() * (LDensity + RDensity); | |||
2444 | // Should always split in some non-trivial place | |||
2445 | DEBUG(dbgs() <<"=>Step\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) | |||
2446 | << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) | |||
2447 | << "LDensity: " << LDensitydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) | |||
2448 | << ", RDensity: " << RDensity << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0) | |||
2449 | << "Metric: " << Metric << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() <<"=>Step\n" << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' << "LDensity: " << LDensity << ", RDensity: " << RDensity << '\n' << "Metric: " << Metric << '\n'; } } while (0); | |||
2450 | if (FMetric < Metric) { | |||
2451 | Pivot = J; | |||
2452 | FMetric = Metric; | |||
2453 | DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Current metric set to: " << FMetric << '\n'; } } while (0); | |||
2454 | } | |||
2455 | ||||
2456 | LSize += J->size(); | |||
2457 | RSize -= J->size(); | |||
2458 | } | |||
2459 | ||||
2460 | if (FMetric == 0 || !areJTsAllowed(TLI)) | |||
2461 | Pivot = CR.Range.first + Size/2; | |||
2462 | splitSwitchCase(CR, Pivot, WorkList, SV, SwitchBB); | |||
2463 | return true; | |||
2464 | } | |||
2465 | ||||
2466 | void SelectionDAGBuilder::splitSwitchCase(CaseRec &CR, CaseItr Pivot, | |||
2467 | CaseRecVector &WorkList, | |||
2468 | const Value *SV, | |||
2469 | MachineBasicBlock *SwitchBB) { | |||
2470 | // Get the MachineFunction which holds the current MBB. This is used when | |||
2471 | // inserting any additional MBBs necessary to represent the switch. | |||
2472 | MachineFunction *CurMF = FuncInfo.MF; | |||
2473 | ||||
2474 | // Figure out which block is immediately after the current one. | |||
2475 | MachineFunction::iterator BBI = CR.CaseBB; | |||
2476 | ++BBI; | |||
2477 | ||||
2478 | const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); | |||
2479 | ||||
2480 | CaseRange LHSR(CR.Range.first, Pivot); | |||
2481 | CaseRange RHSR(Pivot, CR.Range.second); | |||
2482 | const Constant *C = Pivot->Low; | |||
2483 | MachineBasicBlock *FalseBB = nullptr, *TrueBB = nullptr; | |||
2484 | ||||
2485 | // We know that we branch to the LHS if the Value being switched on is | |||
2486 | // less than the Pivot value, C. We use this to optimize our binary | |||
2487 | // tree a bit, by recognizing that if SV is greater than or equal to the | |||
2488 | // LHS's Case Value, and that Case Value is exactly one less than the | |||
2489 | // Pivot's Value, then we can branch directly to the LHS's Target, | |||
2490 | // rather than creating a leaf node for it. | |||
2491 | if ((LHSR.second - LHSR.first) == 1 && LHSR.first->High == CR.GE && | |||
2492 | cast<ConstantInt>(C)->getValue() == | |||
2493 | (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) { | |||
2494 | TrueBB = LHSR.first->BB; | |||
2495 | } else { | |||
2496 | TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB); | |||
2497 | CurMF->insert(BBI, TrueBB); | |||
2498 | WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); | |||
2499 | ||||
2500 | // Put SV in a virtual register to make it available from the new blocks. | |||
2501 | ExportFromCurrentBlock(SV); | |||
2502 | } | |||
2503 | ||||
2504 | // Similar to the optimization above, if the Value being switched on is | |||
2505 | // known to be less than the Constant CR.LT, and the current Case Value | |||
2506 | // is CR.LT - 1, then we can branch directly to the target block for | |||
2507 | // the current Case Value, rather than emitting a RHS leaf node for it. | |||
2508 | if ((RHSR.second - RHSR.first) == 1 && CR.LT && | |||
2509 | cast<ConstantInt>(RHSR.first->Low)->getValue() == | |||
2510 | (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) { | |||
2511 | FalseBB = RHSR.first->BB; | |||
2512 | } else { | |||
2513 | FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB); | |||
2514 | CurMF->insert(BBI, FalseBB); | |||
2515 | WorkList.push_back(CaseRec(FalseBB, CR.LT, C, RHSR)); | |||
2516 | ||||
2517 | // Put SV in a virtual register to make it available from the new blocks. | |||
2518 | ExportFromCurrentBlock(SV); | |||
2519 | } | |||
2520 | ||||
2521 | // Create a CaseBlock record representing a conditional branch to | |||
2522 | // the LHS node if the value being switched on SV is less than C. | |||
2523 | // Otherwise, branch to LHS. | |||
2524 | CaseBlock CB(ISD::SETLT, SV, C, nullptr, TrueBB, FalseBB, CR.CaseBB); | |||
2525 | ||||
2526 | if (CR.CaseBB == SwitchBB) | |||
2527 | visitSwitchCase(CB, SwitchBB); | |||
2528 | else | |||
2529 | SwitchCases.push_back(CB); | |||
2530 | } | |||
2531 | ||||
2532 | /// handleBitTestsSwitchCase - if current case range has few destination and | |||
2533 | /// range span less, than machine word bitwidth, encode case range into series | |||
2534 | /// of masks and emit bit tests with these masks. | |||
2535 | bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR, | |||
2536 | CaseRecVector& WorkList, | |||
2537 | const Value* SV, | |||
2538 | MachineBasicBlock* Default, | |||
2539 | MachineBasicBlock* SwitchBB) { | |||
2540 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
2541 | EVT PTy = TLI.getPointerTy(); | |||
2542 | unsigned IntPtrBits = PTy.getSizeInBits(); | |||
2543 | ||||
2544 | Case& FrontCase = *CR.Range.first; | |||
2545 | Case& BackCase = *(CR.Range.second-1); | |||
2546 | ||||
2547 | // Get the MachineFunction which holds the current MBB. This is used when | |||
2548 | // inserting any additional MBBs necessary to represent the switch. | |||
2549 | MachineFunction *CurMF = FuncInfo.MF; | |||
2550 | ||||
2551 | // If target does not have legal shift left, do not emit bit tests at all. | |||
2552 | if (!TLI.isOperationLegal(ISD::SHL, PTy)) | |||
2553 | return false; | |||
2554 | ||||
2555 | size_t numCmps = 0; | |||
2556 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { | |||
2557 | // Single case counts one, case range - two. | |||
2558 | numCmps += (I->Low == I->High ? 1 : 2); | |||
2559 | } | |||
2560 | ||||
2561 | // Count unique destinations | |||
2562 | SmallSet<MachineBasicBlock*, 4> Dests; | |||
2563 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { | |||
2564 | Dests.insert(I->BB); | |||
2565 | if (Dests.size() > 3) | |||
2566 | // Don't bother the code below, if there are too much unique destinations | |||
2567 | return false; | |||
2568 | } | |||
2569 | DEBUG(dbgs() << "Total number of unique destinations: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Total number of unique destinations: " << Dests.size() << '\n' << "Total number of comparisons: " << numCmps << '\n'; } } while (0) | |||
2570 | << Dests.size() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Total number of unique destinations: " << Dests.size() << '\n' << "Total number of comparisons: " << numCmps << '\n'; } } while (0) | |||
2571 | << "Total number of comparisons: " << numCmps << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Total number of unique destinations: " << Dests.size() << '\n' << "Total number of comparisons: " << numCmps << '\n'; } } while (0); | |||
2572 | ||||
2573 | // Compute span of values. | |||
2574 | const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue(); | |||
2575 | const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue(); | |||
2576 | APInt cmpRange = maxValue - minValue; | |||
2577 | ||||
2578 | DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Compare range: " << cmpRange << '\n' << "Low bound: " << minValue << '\n' << "High bound: " << maxValue << '\n' ; } } while (0) | |||
2579 | << "Low bound: " << minValue << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Compare range: " << cmpRange << '\n' << "Low bound: " << minValue << '\n' << "High bound: " << maxValue << '\n' ; } } while (0) | |||
2580 | << "High bound: " << maxValue << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Compare range: " << cmpRange << '\n' << "Low bound: " << minValue << '\n' << "High bound: " << maxValue << '\n' ; } } while (0); | |||
2581 | ||||
2582 | if (cmpRange.uge(IntPtrBits) || | |||
2583 | (!(Dests.size() == 1 && numCmps >= 3) && | |||
2584 | !(Dests.size() == 2 && numCmps >= 5) && | |||
2585 | !(Dests.size() >= 3 && numCmps >= 6))) | |||
2586 | return false; | |||
2587 | ||||
2588 | DEBUG(dbgs() << "Emitting bit tests\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Emitting bit tests\n"; } } while (0); | |||
2589 | APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth()); | |||
2590 | ||||
2591 | // Optimize the case where all the case values fit in a | |||
2592 | // word without having to subtract minValue. In this case, | |||
2593 | // we can optimize away the subtraction. | |||
2594 | if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) { | |||
2595 | cmpRange = maxValue; | |||
2596 | } else { | |||
2597 | lowBound = minValue; | |||
2598 | } | |||
2599 | ||||
2600 | CaseBitsVector CasesBits; | |||
2601 | unsigned i, count = 0; | |||
2602 | ||||
2603 | for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { | |||
2604 | MachineBasicBlock* Dest = I->BB; | |||
2605 | for (i = 0; i < count; ++i) | |||
2606 | if (Dest == CasesBits[i].BB) | |||
2607 | break; | |||
2608 | ||||
2609 | if (i == count) { | |||
2610 | assert((count < 3) && "Too much destinations to test!")(((count < 3) && "Too much destinations to test!") ? static_cast<void> (0) : __assert_fail ("(count < 3) && \"Too much destinations to test!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2610, __PRETTY_FUNCTION__)); | |||
2611 | CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/)); | |||
2612 | count++; | |||
2613 | } | |||
2614 | ||||
2615 | const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue(); | |||
2616 | const APInt& highValue = cast<ConstantInt>(I->High)->getValue(); | |||
2617 | ||||
2618 | uint64_t lo = (lowValue - lowBound).getZExtValue(); | |||
2619 | uint64_t hi = (highValue - lowBound).getZExtValue(); | |||
2620 | CasesBits[i].ExtraWeight += I->ExtraWeight; | |||
2621 | ||||
2622 | for (uint64_t j = lo; j <= hi; j++) { | |||
2623 | CasesBits[i].Mask |= 1ULL << j; | |||
2624 | CasesBits[i].Bits++; | |||
2625 | } | |||
2626 | ||||
2627 | } | |||
2628 | std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp()); | |||
2629 | ||||
2630 | BitTestInfo BTC; | |||
2631 | ||||
2632 | // Figure out which block is immediately after the current one. | |||
2633 | MachineFunction::iterator BBI = CR.CaseBB; | |||
2634 | ++BBI; | |||
2635 | ||||
2636 | const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); | |||
2637 | ||||
2638 | DEBUG(dbgs() << "Cases:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Cases:\n"; } } while (0); | |||
2639 | for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) { | |||
2640 | DEBUG(dbgs() << "Mask: " << CasesBits[i].Maskdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits << ", BB: " << CasesBits[i].BB << '\n'; } } while (0) | |||
2641 | << ", Bits: " << CasesBits[i].Bitsdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits << ", BB: " << CasesBits[i].BB << '\n'; } } while (0) | |||
2642 | << ", BB: " << CasesBits[i].BB << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits << ", BB: " << CasesBits[i].BB << '\n'; } } while (0); | |||
2643 | ||||
2644 | MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB); | |||
2645 | CurMF->insert(BBI, CaseBB); | |||
2646 | BTC.push_back(BitTestCase(CasesBits[i].Mask, | |||
2647 | CaseBB, | |||
2648 | CasesBits[i].BB, CasesBits[i].ExtraWeight)); | |||
2649 | ||||
2650 | // Put SV in a virtual register to make it available from the new blocks. | |||
2651 | ExportFromCurrentBlock(SV); | |||
2652 | } | |||
2653 | ||||
2654 | BitTestBlock BTB(lowBound, cmpRange, SV, | |||
2655 | -1U, MVT::Other, (CR.CaseBB == SwitchBB), | |||
2656 | CR.CaseBB, Default, std::move(BTC)); | |||
2657 | ||||
2658 | if (CR.CaseBB == SwitchBB) | |||
2659 | visitBitTestHeader(BTB, SwitchBB); | |||
2660 | ||||
2661 | BitTestCases.push_back(std::move(BTB)); | |||
2662 | ||||
2663 | return true; | |||
2664 | } | |||
2665 | ||||
2666 | /// Clusterify - Transform simple list of Cases into list of CaseRange's | |||
2667 | void SelectionDAGBuilder::Clusterify(CaseVector& Cases, | |||
2668 | const SwitchInst& SI) { | |||
2669 | BranchProbabilityInfo *BPI = FuncInfo.BPI; | |||
2670 | // Start with "simple" cases. | |||
2671 | for (SwitchInst::ConstCaseIt i : SI.cases()) { | |||
2672 | const BasicBlock *SuccBB = i.getCaseSuccessor(); | |||
2673 | MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB]; | |||
2674 | ||||
2675 | uint32_t ExtraWeight = | |||
2676 | BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0; | |||
2677 | ||||
2678 | Cases.push_back(Case(i.getCaseValue(), i.getCaseValue(), | |||
2679 | SMBB, ExtraWeight)); | |||
2680 | } | |||
2681 | std::sort(Cases.begin(), Cases.end(), CaseCmp()); | |||
2682 | ||||
2683 | // Merge case into clusters | |||
2684 | if (Cases.size() >= 2) | |||
2685 | // Must recompute end() each iteration because it may be | |||
2686 | // invalidated by erase if we hold on to it | |||
2687 | for (CaseItr I = Cases.begin(), J = std::next(Cases.begin()); | |||
2688 | J != Cases.end(); ) { | |||
2689 | const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue(); | |||
2690 | const APInt& currentValue = cast<ConstantInt>(I->High)->getValue(); | |||
2691 | MachineBasicBlock* nextBB = J->BB; | |||
2692 | MachineBasicBlock* currentBB = I->BB; | |||
2693 | ||||
2694 | // If the two neighboring cases go to the same destination, merge them | |||
2695 | // into a single case. | |||
2696 | if ((nextValue - currentValue == 1) && (currentBB == nextBB)) { | |||
2697 | I->High = J->High; | |||
2698 | I->ExtraWeight += J->ExtraWeight; | |||
2699 | J = Cases.erase(J); | |||
2700 | } else { | |||
2701 | I = J++; | |||
2702 | } | |||
2703 | } | |||
2704 | ||||
2705 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2706 | size_t numCmps = 0;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2707 | for (auto &I : Cases)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2708 | // A range counts double, since it requires two compares.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2709 | numCmps += I.Low != I.High ? 2 : 1;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2710 | ||||
2711 | dbgs() << "Clusterify finished. Total clusters: " << Cases.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2712 | << ". Total compares: " << numCmps << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0) | |||
2713 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { size_t numCmps = 0; for (auto &I : Cases) numCmps += I.Low != I.High ? 2 : 1; dbgs() << "Clusterify finished. Total clusters: " << Cases.size() << ". Total compares: " << numCmps << '\n'; }; } } while (0); | |||
2714 | } | |||
2715 | ||||
2716 | void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First, | |||
2717 | MachineBasicBlock *Last) { | |||
2718 | // Update JTCases. | |||
2719 | for (unsigned i = 0, e = JTCases.size(); i != e; ++i) | |||
2720 | if (JTCases[i].first.HeaderBB == First) | |||
2721 | JTCases[i].first.HeaderBB = Last; | |||
2722 | ||||
2723 | // Update BitTestCases. | |||
2724 | for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i) | |||
2725 | if (BitTestCases[i].Parent == First) | |||
2726 | BitTestCases[i].Parent = Last; | |||
2727 | } | |||
2728 | ||||
2729 | void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) { | |||
2730 | MachineBasicBlock *SwitchMBB = FuncInfo.MBB; | |||
2731 | ||||
2732 | // Figure out which block is immediately after the current one. | |||
2733 | MachineBasicBlock *NextBlock = nullptr; | |||
2734 | if (SwitchMBB + 1 != FuncInfo.MF->end()) | |||
2735 | NextBlock = SwitchMBB + 1; | |||
2736 | ||||
2737 | ||||
2738 | // Create a vector of Cases, sorted so that we can efficiently create a binary | |||
2739 | // search tree from them. | |||
2740 | CaseVector Cases; | |||
2741 | Clusterify(Cases, SI); | |||
2742 | ||||
2743 | // Get the default destination MBB. | |||
2744 | MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()]; | |||
2745 | ||||
2746 | if (isa<UnreachableInst>(SI.getDefaultDest()->getFirstNonPHIOrDbg()) && | |||
2747 | !Cases.empty()) { | |||
2748 | // Replace an unreachable default destination with the most popular case | |||
2749 | // destination. | |||
2750 | DenseMap<const BasicBlock *, unsigned> Popularity; | |||
2751 | unsigned MaxPop = 0; | |||
2752 | const BasicBlock *MaxBB = nullptr; | |||
2753 | for (auto I : SI.cases()) { | |||
2754 | const BasicBlock *BB = I.getCaseSuccessor(); | |||
2755 | if (++Popularity[BB] > MaxPop) { | |||
2756 | MaxPop = Popularity[BB]; | |||
2757 | MaxBB = BB; | |||
2758 | } | |||
2759 | } | |||
2760 | ||||
2761 | // Set new default. | |||
2762 | assert(MaxPop > 0)((MaxPop > 0) ? static_cast<void> (0) : __assert_fail ("MaxPop > 0", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2762, __PRETTY_FUNCTION__)); | |||
2763 | assert(MaxBB)((MaxBB) ? static_cast<void> (0) : __assert_fail ("MaxBB" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 2763, __PRETTY_FUNCTION__)); | |||
2764 | Default = FuncInfo.MBBMap[MaxBB]; | |||
2765 | ||||
2766 | // Remove cases that were pointing to the destination that is now the default. | |||
2767 | Cases.erase(std::remove_if(Cases.begin(), Cases.end(), | |||
2768 | [&](const Case &C) { return C.BB == Default; }), | |||
2769 | Cases.end()); | |||
2770 | } | |||
2771 | ||||
2772 | // If there is only the default destination, go there directly. | |||
2773 | if (Cases.empty()) { | |||
2774 | // Update machine-CFG edges. | |||
2775 | SwitchMBB->addSuccessor(Default); | |||
2776 | ||||
2777 | // If this is not a fall-through branch, emit the branch. | |||
2778 | if (Default != NextBlock) { | |||
2779 | DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, | |||
2780 | getControlRoot(), DAG.getBasicBlock(Default))); | |||
2781 | } | |||
2782 | return; | |||
2783 | } | |||
2784 | ||||
2785 | // Get the Value to be switched on. | |||
2786 | const Value *SV = SI.getCondition(); | |||
2787 | ||||
2788 | // Push the initial CaseRec onto the worklist | |||
2789 | CaseRecVector WorkList; | |||
2790 | WorkList.push_back(CaseRec(SwitchMBB,nullptr,nullptr, | |||
2791 | CaseRange(Cases.begin(),Cases.end()))); | |||
2792 | ||||
2793 | while (!WorkList.empty()) { | |||
2794 | // Grab a record representing a case range to process off the worklist | |||
2795 | CaseRec CR = WorkList.back(); | |||
2796 | WorkList.pop_back(); | |||
2797 | ||||
2798 | if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) | |||
2799 | continue; | |||
2800 | ||||
2801 | // If the range has few cases (two or less) emit a series of specific | |||
2802 | // tests. | |||
2803 | if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB)) | |||
2804 | continue; | |||
2805 | ||||
2806 | // If the switch has more than N blocks, and is at least 40% dense, and the | |||
2807 | // target supports indirect branches, then emit a jump table rather than | |||
2808 | // lowering the switch to a binary tree of conditional branches. | |||
2809 | // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries(). | |||
2810 | if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) | |||
2811 | continue; | |||
2812 | ||||
2813 | // Emit binary tree. We need to pick a pivot, and push left and right ranges | |||
2814 | // onto the worklist. Leafs are handled via handleSmallSwitchRange() call. | |||
2815 | handleBTSplitSwitchCase(CR, WorkList, SV, SwitchMBB); | |||
2816 | } | |||
2817 | } | |||
2818 | ||||
2819 | void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) { | |||
2820 | MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB; | |||
2821 | ||||
2822 | // Update machine-CFG edges with unique successors. | |||
2823 | SmallSet<BasicBlock*, 32> Done; | |||
2824 | for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) { | |||
2825 | BasicBlock *BB = I.getSuccessor(i); | |||
2826 | bool Inserted = Done.insert(BB).second; | |||
2827 | if (!Inserted) | |||
2828 | continue; | |||
2829 | ||||
2830 | MachineBasicBlock *Succ = FuncInfo.MBBMap[BB]; | |||
2831 | addSuccessorWithWeight(IndirectBrMBB, Succ); | |||
2832 | } | |||
2833 | ||||
2834 | DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(), | |||
2835 | MVT::Other, getControlRoot(), | |||
2836 | getValue(I.getAddress()))); | |||
2837 | } | |||
2838 | ||||
2839 | void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) { | |||
2840 | if (DAG.getTarget().Options.TrapUnreachable) | |||
2841 | DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot())); | |||
2842 | } | |||
2843 | ||||
2844 | void SelectionDAGBuilder::visitFSub(const User &I) { | |||
2845 | // -0.0 - X --> fneg | |||
2846 | Type *Ty = I.getType(); | |||
2847 | if (isa<Constant>(I.getOperand(0)) && | |||
2848 | I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) { | |||
2849 | SDValue Op2 = getValue(I.getOperand(1)); | |||
2850 | setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(), | |||
2851 | Op2.getValueType(), Op2)); | |||
2852 | return; | |||
2853 | } | |||
2854 | ||||
2855 | visitBinary(I, ISD::FSUB); | |||
2856 | } | |||
2857 | ||||
2858 | void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) { | |||
2859 | SDValue Op1 = getValue(I.getOperand(0)); | |||
2860 | SDValue Op2 = getValue(I.getOperand(1)); | |||
2861 | ||||
2862 | bool nuw = false; | |||
2863 | bool nsw = false; | |||
2864 | bool exact = false; | |||
2865 | if (const OverflowingBinaryOperator *OFBinOp = | |||
2866 | dyn_cast<const OverflowingBinaryOperator>(&I)) { | |||
2867 | nuw = OFBinOp->hasNoUnsignedWrap(); | |||
2868 | nsw = OFBinOp->hasNoSignedWrap(); | |||
2869 | } | |||
2870 | if (const PossiblyExactOperator *ExactOp = | |||
2871 | dyn_cast<const PossiblyExactOperator>(&I)) | |||
2872 | exact = ExactOp->isExact(); | |||
2873 | ||||
2874 | SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(), | |||
2875 | Op1, Op2, nuw, nsw, exact); | |||
2876 | setValue(&I, BinNodeValue); | |||
2877 | } | |||
2878 | ||||
2879 | void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) { | |||
2880 | SDValue Op1 = getValue(I.getOperand(0)); | |||
2881 | SDValue Op2 = getValue(I.getOperand(1)); | |||
2882 | ||||
2883 | EVT ShiftTy = | |||
2884 | DAG.getTargetLoweringInfo().getShiftAmountTy(Op2.getValueType()); | |||
2885 | ||||
2886 | // Coerce the shift amount to the right type if we can. | |||
2887 | if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) { | |||
2888 | unsigned ShiftSize = ShiftTy.getSizeInBits(); | |||
2889 | unsigned Op2Size = Op2.getValueType().getSizeInBits(); | |||
2890 | SDLoc DL = getCurSDLoc(); | |||
2891 | ||||
2892 | // If the operand is smaller than the shift count type, promote it. | |||
2893 | if (ShiftSize > Op2Size) | |||
2894 | Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2); | |||
2895 | ||||
2896 | // If the operand is larger than the shift count type but the shift | |||
2897 | // count type has enough bits to represent any shift value, truncate | |||
2898 | // it now. This is a common case and it exposes the truncate to | |||
2899 | // optimization early. | |||
2900 | else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits())) | |||
2901 | Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2); | |||
2902 | // Otherwise we'll need to temporarily settle for some other convenient | |||
2903 | // type. Type legalization will make adjustments once the shiftee is split. | |||
2904 | else | |||
2905 | Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32); | |||
2906 | } | |||
2907 | ||||
2908 | bool nuw = false; | |||
2909 | bool nsw = false; | |||
2910 | bool exact = false; | |||
2911 | ||||
2912 | if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) { | |||
2913 | ||||
2914 | if (const OverflowingBinaryOperator *OFBinOp = | |||
2915 | dyn_cast<const OverflowingBinaryOperator>(&I)) { | |||
2916 | nuw = OFBinOp->hasNoUnsignedWrap(); | |||
2917 | nsw = OFBinOp->hasNoSignedWrap(); | |||
2918 | } | |||
2919 | if (const PossiblyExactOperator *ExactOp = | |||
2920 | dyn_cast<const PossiblyExactOperator>(&I)) | |||
2921 | exact = ExactOp->isExact(); | |||
2922 | } | |||
2923 | ||||
2924 | SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2, | |||
2925 | nuw, nsw, exact); | |||
2926 | setValue(&I, Res); | |||
2927 | } | |||
2928 | ||||
2929 | void SelectionDAGBuilder::visitSDiv(const User &I) { | |||
2930 | SDValue Op1 = getValue(I.getOperand(0)); | |||
2931 | SDValue Op2 = getValue(I.getOperand(1)); | |||
2932 | ||||
2933 | // Turn exact SDivs into multiplications. | |||
2934 | // FIXME: This should be in DAGCombiner, but it doesn't have access to the | |||
2935 | // exact bit. | |||
2936 | if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() && | |||
2937 | !isa<ConstantSDNode>(Op1) && | |||
2938 | isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue()) | |||
2939 | setValue(&I, DAG.getTargetLoweringInfo() | |||
2940 | .BuildExactSDIV(Op1, Op2, getCurSDLoc(), DAG)); | |||
2941 | else | |||
2942 | setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), | |||
2943 | Op1, Op2)); | |||
2944 | } | |||
2945 | ||||
2946 | void SelectionDAGBuilder::visitICmp(const User &I) { | |||
2947 | ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; | |||
2948 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I)) | |||
2949 | predicate = IC->getPredicate(); | |||
2950 | else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) | |||
2951 | predicate = ICmpInst::Predicate(IC->getPredicate()); | |||
2952 | SDValue Op1 = getValue(I.getOperand(0)); | |||
2953 | SDValue Op2 = getValue(I.getOperand(1)); | |||
2954 | ISD::CondCode Opcode = getICmpCondCode(predicate); | |||
2955 | ||||
2956 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
2957 | setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode)); | |||
2958 | } | |||
2959 | ||||
2960 | void SelectionDAGBuilder::visitFCmp(const User &I) { | |||
2961 | FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; | |||
2962 | if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I)) | |||
2963 | predicate = FC->getPredicate(); | |||
2964 | else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) | |||
2965 | predicate = FCmpInst::Predicate(FC->getPredicate()); | |||
2966 | SDValue Op1 = getValue(I.getOperand(0)); | |||
2967 | SDValue Op2 = getValue(I.getOperand(1)); | |||
2968 | ISD::CondCode Condition = getFCmpCondCode(predicate); | |||
2969 | if (TM.Options.NoNaNsFPMath) | |||
2970 | Condition = getFCmpCodeWithoutNaN(Condition); | |||
2971 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
2972 | setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition)); | |||
2973 | } | |||
2974 | ||||
2975 | void SelectionDAGBuilder::visitSelect(const User &I) { | |||
2976 | SmallVector<EVT, 4> ValueVTs; | |||
2977 | ComputeValueVTs(DAG.getTargetLoweringInfo(), I.getType(), ValueVTs); | |||
2978 | unsigned NumValues = ValueVTs.size(); | |||
2979 | if (NumValues == 0) return; | |||
2980 | ||||
2981 | SmallVector<SDValue, 4> Values(NumValues); | |||
2982 | SDValue Cond = getValue(I.getOperand(0)); | |||
2983 | SDValue TrueVal = getValue(I.getOperand(1)); | |||
2984 | SDValue FalseVal = getValue(I.getOperand(2)); | |||
2985 | ISD::NodeType OpCode = Cond.getValueType().isVector() ? | |||
2986 | ISD::VSELECT : ISD::SELECT; | |||
2987 | ||||
2988 | for (unsigned i = 0; i != NumValues; ++i) | |||
2989 | Values[i] = DAG.getNode(OpCode, getCurSDLoc(), | |||
2990 | TrueVal.getNode()->getValueType(TrueVal.getResNo()+i), | |||
2991 | Cond, | |||
2992 | SDValue(TrueVal.getNode(), | |||
2993 | TrueVal.getResNo() + i), | |||
2994 | SDValue(FalseVal.getNode(), | |||
2995 | FalseVal.getResNo() + i)); | |||
2996 | ||||
2997 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), | |||
2998 | DAG.getVTList(ValueVTs), Values)); | |||
2999 | } | |||
3000 | ||||
3001 | void SelectionDAGBuilder::visitTrunc(const User &I) { | |||
3002 | // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). | |||
3003 | SDValue N = getValue(I.getOperand(0)); | |||
3004 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3005 | setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N)); | |||
3006 | } | |||
3007 | ||||
3008 | void SelectionDAGBuilder::visitZExt(const User &I) { | |||
3009 | // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). | |||
3010 | // ZExt also can't be a cast to bool for same reason. So, nothing much to do | |||
3011 | SDValue N = getValue(I.getOperand(0)); | |||
3012 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3013 | setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N)); | |||
3014 | } | |||
3015 | ||||
3016 | void SelectionDAGBuilder::visitSExt(const User &I) { | |||
3017 | // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). | |||
3018 | // SExt also can't be a cast to bool for same reason. So, nothing much to do | |||
3019 | SDValue N = getValue(I.getOperand(0)); | |||
3020 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3021 | setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N)); | |||
3022 | } | |||
3023 | ||||
3024 | void SelectionDAGBuilder::visitFPTrunc(const User &I) { | |||
3025 | // FPTrunc is never a no-op cast, no need to check | |||
3026 | SDValue N = getValue(I.getOperand(0)); | |||
3027 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3028 | EVT DestVT = TLI.getValueType(I.getType()); | |||
3029 | setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(), DestVT, N, | |||
3030 | DAG.getTargetConstant(0, TLI.getPointerTy()))); | |||
3031 | } | |||
3032 | ||||
3033 | void SelectionDAGBuilder::visitFPExt(const User &I) { | |||
3034 | // FPExt is never a no-op cast, no need to check | |||
3035 | SDValue N = getValue(I.getOperand(0)); | |||
3036 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3037 | setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N)); | |||
3038 | } | |||
3039 | ||||
3040 | void SelectionDAGBuilder::visitFPToUI(const User &I) { | |||
3041 | // FPToUI is never a no-op cast, no need to check | |||
3042 | SDValue N = getValue(I.getOperand(0)); | |||
3043 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3044 | setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N)); | |||
3045 | } | |||
3046 | ||||
3047 | void SelectionDAGBuilder::visitFPToSI(const User &I) { | |||
3048 | // FPToSI is never a no-op cast, no need to check | |||
3049 | SDValue N = getValue(I.getOperand(0)); | |||
3050 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3051 | setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N)); | |||
3052 | } | |||
3053 | ||||
3054 | void SelectionDAGBuilder::visitUIToFP(const User &I) { | |||
3055 | // UIToFP is never a no-op cast, no need to check | |||
3056 | SDValue N = getValue(I.getOperand(0)); | |||
3057 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3058 | setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N)); | |||
3059 | } | |||
3060 | ||||
3061 | void SelectionDAGBuilder::visitSIToFP(const User &I) { | |||
3062 | // SIToFP is never a no-op cast, no need to check | |||
3063 | SDValue N = getValue(I.getOperand(0)); | |||
3064 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3065 | setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N)); | |||
3066 | } | |||
3067 | ||||
3068 | void SelectionDAGBuilder::visitPtrToInt(const User &I) { | |||
3069 | // What to do depends on the size of the integer and the size of the pointer. | |||
3070 | // We can either truncate, zero extend, or no-op, accordingly. | |||
3071 | SDValue N = getValue(I.getOperand(0)); | |||
3072 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3073 | setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT)); | |||
3074 | } | |||
3075 | ||||
3076 | void SelectionDAGBuilder::visitIntToPtr(const User &I) { | |||
3077 | // What to do depends on the size of the integer and the size of the pointer. | |||
3078 | // We can either truncate, zero extend, or no-op, accordingly. | |||
3079 | SDValue N = getValue(I.getOperand(0)); | |||
3080 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3081 | setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT)); | |||
3082 | } | |||
3083 | ||||
3084 | void SelectionDAGBuilder::visitBitCast(const User &I) { | |||
3085 | SDValue N = getValue(I.getOperand(0)); | |||
3086 | EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType()); | |||
3087 | ||||
3088 | // BitCast assures us that source and destination are the same size so this is | |||
3089 | // either a BITCAST or a no-op. | |||
3090 | if (DestVT != N.getValueType()) | |||
3091 | setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(), | |||
3092 | DestVT, N)); // convert types. | |||
3093 | // Check if the original LLVM IR Operand was a ConstantInt, because getValue() | |||
3094 | // might fold any kind of constant expression to an integer constant and that | |||
3095 | // is not what we are looking for. Only regcognize a bitcast of a genuine | |||
3096 | // constant integer as an opaque constant. | |||
3097 | else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0))) | |||
3098 | setValue(&I, DAG.getConstant(C->getValue(), DestVT, /*isTarget=*/false, | |||
3099 | /*isOpaque*/true)); | |||
3100 | else | |||
3101 | setValue(&I, N); // noop cast. | |||
3102 | } | |||
3103 | ||||
3104 | void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) { | |||
3105 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3106 | const Value *SV = I.getOperand(0); | |||
3107 | SDValue N = getValue(SV); | |||
3108 | EVT DestVT = TLI.getValueType(I.getType()); | |||
3109 | ||||
3110 | unsigned SrcAS = SV->getType()->getPointerAddressSpace(); | |||
3111 | unsigned DestAS = I.getType()->getPointerAddressSpace(); | |||
3112 | ||||
3113 | if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS)) | |||
3114 | N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS); | |||
3115 | ||||
3116 | setValue(&I, N); | |||
3117 | } | |||
3118 | ||||
3119 | void SelectionDAGBuilder::visitInsertElement(const User &I) { | |||
3120 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3121 | SDValue InVec = getValue(I.getOperand(0)); | |||
3122 | SDValue InVal = getValue(I.getOperand(1)); | |||
3123 | SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)), | |||
3124 | getCurSDLoc(), TLI.getVectorIdxTy()); | |||
3125 | setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(), | |||
3126 | TLI.getValueType(I.getType()), InVec, InVal, InIdx)); | |||
3127 | } | |||
3128 | ||||
3129 | void SelectionDAGBuilder::visitExtractElement(const User &I) { | |||
3130 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3131 | SDValue InVec = getValue(I.getOperand(0)); | |||
3132 | SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), | |||
3133 | getCurSDLoc(), TLI.getVectorIdxTy()); | |||
3134 | setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(), | |||
3135 | TLI.getValueType(I.getType()), InVec, InIdx)); | |||
3136 | } | |||
3137 | ||||
3138 | // Utility for visitShuffleVector - Return true if every element in Mask, | |||
3139 | // beginning from position Pos and ending in Pos+Size, falls within the | |||
3140 | // specified sequential range [L, L+Pos). or is undef. | |||
3141 | static bool isSequentialInRange(const SmallVectorImpl<int> &Mask, | |||
3142 | unsigned Pos, unsigned Size, int Low) { | |||
3143 | for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low) | |||
3144 | if (Mask[i] >= 0 && Mask[i] != Low) | |||
3145 | return false; | |||
3146 | return true; | |||
3147 | } | |||
3148 | ||||
3149 | void SelectionDAGBuilder::visitShuffleVector(const User &I) { | |||
3150 | SDValue Src1 = getValue(I.getOperand(0)); | |||
3151 | SDValue Src2 = getValue(I.getOperand(1)); | |||
3152 | ||||
3153 | SmallVector<int, 8> Mask; | |||
3154 | ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask); | |||
3155 | unsigned MaskNumElts = Mask.size(); | |||
3156 | ||||
3157 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3158 | EVT VT = TLI.getValueType(I.getType()); | |||
3159 | EVT SrcVT = Src1.getValueType(); | |||
3160 | unsigned SrcNumElts = SrcVT.getVectorNumElements(); | |||
3161 | ||||
3162 | if (SrcNumElts == MaskNumElts) { | |||
3163 | setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, | |||
3164 | &Mask[0])); | |||
3165 | return; | |||
3166 | } | |||
3167 | ||||
3168 | // Normalize the shuffle vector since mask and vector length don't match. | |||
3169 | if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) { | |||
3170 | // Mask is longer than the source vectors and is a multiple of the source | |||
3171 | // vectors. We can use concatenate vector to make the mask and vectors | |||
3172 | // lengths match. | |||
3173 | if (SrcNumElts*2 == MaskNumElts) { | |||
3174 | // First check for Src1 in low and Src2 in high | |||
3175 | if (isSequentialInRange(Mask, 0, SrcNumElts, 0) && | |||
3176 | isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) { | |||
3177 | // The shuffle is concatenating two vectors together. | |||
3178 | setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(), | |||
3179 | VT, Src1, Src2)); | |||
3180 | return; | |||
3181 | } | |||
3182 | // Then check for Src2 in low and Src1 in high | |||
3183 | if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) && | |||
3184 | isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) { | |||
3185 | // The shuffle is concatenating two vectors together. | |||
3186 | setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(), | |||
3187 | VT, Src2, Src1)); | |||
3188 | return; | |||
3189 | } | |||
3190 | } | |||
3191 | ||||
3192 | // Pad both vectors with undefs to make them the same length as the mask. | |||
3193 | unsigned NumConcat = MaskNumElts / SrcNumElts; | |||
3194 | bool Src1U = Src1.getOpcode() == ISD::UNDEF; | |||
3195 | bool Src2U = Src2.getOpcode() == ISD::UNDEF; | |||
3196 | SDValue UndefVal = DAG.getUNDEF(SrcVT); | |||
3197 | ||||
3198 | SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal); | |||
3199 | SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal); | |||
3200 | MOps1[0] = Src1; | |||
3201 | MOps2[0] = Src2; | |||
3202 | ||||
3203 | Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, | |||
3204 | getCurSDLoc(), VT, MOps1); | |||
3205 | Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, | |||
3206 | getCurSDLoc(), VT, MOps2); | |||
3207 | ||||
3208 | // Readjust mask for new input vector length. | |||
3209 | SmallVector<int, 8> MappedOps; | |||
3210 | for (unsigned i = 0; i != MaskNumElts; ++i) { | |||
3211 | int Idx = Mask[i]; | |||
3212 | if (Idx >= (int)SrcNumElts) | |||
3213 | Idx -= SrcNumElts - MaskNumElts; | |||
3214 | MappedOps.push_back(Idx); | |||
3215 | } | |||
3216 | ||||
3217 | setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, | |||
3218 | &MappedOps[0])); | |||
3219 | return; | |||
3220 | } | |||
3221 | ||||
3222 | if (SrcNumElts > MaskNumElts) { | |||
3223 | // Analyze the access pattern of the vector to see if we can extract | |||
3224 | // two subvectors and do the shuffle. The analysis is done by calculating | |||
3225 | // the range of elements the mask access on both vectors. | |||
3226 | int MinRange[2] = { static_cast<int>(SrcNumElts), | |||
3227 | static_cast<int>(SrcNumElts)}; | |||
3228 | int MaxRange[2] = {-1, -1}; | |||
3229 | ||||
3230 | for (unsigned i = 0; i != MaskNumElts; ++i) { | |||
3231 | int Idx = Mask[i]; | |||
3232 | unsigned Input = 0; | |||
3233 | if (Idx < 0) | |||
3234 | continue; | |||
3235 | ||||
3236 | if (Idx >= (int)SrcNumElts) { | |||
3237 | Input = 1; | |||
3238 | Idx -= SrcNumElts; | |||
3239 | } | |||
3240 | if (Idx > MaxRange[Input]) | |||
3241 | MaxRange[Input] = Idx; | |||
3242 | if (Idx < MinRange[Input]) | |||
3243 | MinRange[Input] = Idx; | |||
3244 | } | |||
3245 | ||||
3246 | // Check if the access is smaller than the vector size and can we find | |||
3247 | // a reasonable extract index. | |||
3248 | int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not | |||
3249 | // Extract. | |||
3250 | int StartIdx[2]; // StartIdx to extract from | |||
3251 | for (unsigned Input = 0; Input < 2; ++Input) { | |||
3252 | if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) { | |||
3253 | RangeUse[Input] = 0; // Unused | |||
3254 | StartIdx[Input] = 0; | |||
3255 | continue; | |||
3256 | } | |||
3257 | ||||
3258 | // Find a good start index that is a multiple of the mask length. Then | |||
3259 | // see if the rest of the elements are in range. | |||
3260 | StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts; | |||
3261 | if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts && | |||
3262 | StartIdx[Input] + MaskNumElts <= SrcNumElts) | |||
3263 | RangeUse[Input] = 1; // Extract from a multiple of the mask length. | |||
3264 | } | |||
3265 | ||||
3266 | if (RangeUse[0] == 0 && RangeUse[1] == 0) { | |||
3267 | setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used. | |||
3268 | return; | |||
3269 | } | |||
3270 | if (RangeUse[0] >= 0 && RangeUse[1] >= 0) { | |||
3271 | // Extract appropriate subvector and generate a vector shuffle | |||
3272 | for (unsigned Input = 0; Input < 2; ++Input) { | |||
3273 | SDValue &Src = Input == 0 ? Src1 : Src2; | |||
3274 | if (RangeUse[Input] == 0) | |||
3275 | Src = DAG.getUNDEF(VT); | |||
3276 | else | |||
3277 | Src = DAG.getNode( | |||
3278 | ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT, Src, | |||
3279 | DAG.getConstant(StartIdx[Input], TLI.getVectorIdxTy())); | |||
3280 | } | |||
3281 | ||||
3282 | // Calculate new mask. | |||
3283 | SmallVector<int, 8> MappedOps; | |||
3284 | for (unsigned i = 0; i != MaskNumElts; ++i) { | |||
3285 | int Idx = Mask[i]; | |||
3286 | if (Idx >= 0) { | |||
3287 | if (Idx < (int)SrcNumElts) | |||
3288 | Idx -= StartIdx[0]; | |||
3289 | else | |||
3290 | Idx -= SrcNumElts + StartIdx[1] - MaskNumElts; | |||
3291 | } | |||
3292 | MappedOps.push_back(Idx); | |||
3293 | } | |||
3294 | ||||
3295 | setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, | |||
3296 | &MappedOps[0])); | |||
3297 | return; | |||
3298 | } | |||
3299 | } | |||
3300 | ||||
3301 | // We can't use either concat vectors or extract subvectors so fall back to | |||
3302 | // replacing the shuffle with extract and build vector. | |||
3303 | // to insert and build vector. | |||
3304 | EVT EltVT = VT.getVectorElementType(); | |||
3305 | EVT IdxVT = TLI.getVectorIdxTy(); | |||
3306 | SmallVector<SDValue,8> Ops; | |||
3307 | for (unsigned i = 0; i != MaskNumElts; ++i) { | |||
3308 | int Idx = Mask[i]; | |||
3309 | SDValue Res; | |||
3310 | ||||
3311 | if (Idx < 0) { | |||
3312 | Res = DAG.getUNDEF(EltVT); | |||
3313 | } else { | |||
3314 | SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2; | |||
3315 | if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts; | |||
3316 | ||||
3317 | Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(), | |||
3318 | EltVT, Src, DAG.getConstant(Idx, IdxVT)); | |||
3319 | } | |||
3320 | ||||
3321 | Ops.push_back(Res); | |||
3322 | } | |||
3323 | ||||
3324 | setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops)); | |||
3325 | } | |||
3326 | ||||
3327 | void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) { | |||
3328 | const Value *Op0 = I.getOperand(0); | |||
3329 | const Value *Op1 = I.getOperand(1); | |||
3330 | Type *AggTy = I.getType(); | |||
3331 | Type *ValTy = Op1->getType(); | |||
3332 | bool IntoUndef = isa<UndefValue>(Op0); | |||
3333 | bool FromUndef = isa<UndefValue>(Op1); | |||
3334 | ||||
3335 | unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices()); | |||
3336 | ||||
3337 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3338 | SmallVector<EVT, 4> AggValueVTs; | |||
3339 | ComputeValueVTs(TLI, AggTy, AggValueVTs); | |||
3340 | SmallVector<EVT, 4> ValValueVTs; | |||
3341 | ComputeValueVTs(TLI, ValTy, ValValueVTs); | |||
3342 | ||||
3343 | unsigned NumAggValues = AggValueVTs.size(); | |||
3344 | unsigned NumValValues = ValValueVTs.size(); | |||
3345 | SmallVector<SDValue, 4> Values(NumAggValues); | |||
3346 | ||||
3347 | // Ignore an insertvalue that produces an empty object | |||
3348 | if (!NumAggValues) { | |||
3349 | setValue(&I, DAG.getUNDEF(MVT(MVT::Other))); | |||
3350 | return; | |||
3351 | } | |||
3352 | ||||
3353 | SDValue Agg = getValue(Op0); | |||
3354 | unsigned i = 0; | |||
3355 | // Copy the beginning value(s) from the original aggregate. | |||
3356 | for (; i != LinearIndex; ++i) | |||
3357 | Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : | |||
3358 | SDValue(Agg.getNode(), Agg.getResNo() + i); | |||
3359 | // Copy values from the inserted value(s). | |||
3360 | if (NumValValues) { | |||
3361 | SDValue Val = getValue(Op1); | |||
3362 | for (; i != LinearIndex + NumValValues; ++i) | |||
3363 | Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) : | |||
3364 | SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex); | |||
3365 | } | |||
3366 | // Copy remaining value(s) from the original aggregate. | |||
3367 | for (; i != NumAggValues; ++i) | |||
3368 | Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : | |||
3369 | SDValue(Agg.getNode(), Agg.getResNo() + i); | |||
3370 | ||||
3371 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), | |||
3372 | DAG.getVTList(AggValueVTs), Values)); | |||
3373 | } | |||
3374 | ||||
3375 | void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) { | |||
3376 | const Value *Op0 = I.getOperand(0); | |||
3377 | Type *AggTy = Op0->getType(); | |||
3378 | Type *ValTy = I.getType(); | |||
3379 | bool OutOfUndef = isa<UndefValue>(Op0); | |||
3380 | ||||
3381 | unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices()); | |||
3382 | ||||
3383 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3384 | SmallVector<EVT, 4> ValValueVTs; | |||
3385 | ComputeValueVTs(TLI, ValTy, ValValueVTs); | |||
3386 | ||||
3387 | unsigned NumValValues = ValValueVTs.size(); | |||
3388 | ||||
3389 | // Ignore a extractvalue that produces an empty object | |||
3390 | if (!NumValValues) { | |||
3391 | setValue(&I, DAG.getUNDEF(MVT(MVT::Other))); | |||
3392 | return; | |||
3393 | } | |||
3394 | ||||
3395 | SmallVector<SDValue, 4> Values(NumValValues); | |||
3396 | ||||
3397 | SDValue Agg = getValue(Op0); | |||
3398 | // Copy out the selected value(s). | |||
3399 | for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i) | |||
3400 | Values[i - LinearIndex] = | |||
3401 | OutOfUndef ? | |||
3402 | DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) : | |||
3403 | SDValue(Agg.getNode(), Agg.getResNo() + i); | |||
3404 | ||||
3405 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), | |||
3406 | DAG.getVTList(ValValueVTs), Values)); | |||
3407 | } | |||
3408 | ||||
3409 | void SelectionDAGBuilder::visitGetElementPtr(const User &I) { | |||
3410 | Value *Op0 = I.getOperand(0); | |||
3411 | // Note that the pointer operand may be a vector of pointers. Take the scalar | |||
3412 | // element which holds a pointer. | |||
3413 | Type *Ty = Op0->getType()->getScalarType(); | |||
3414 | unsigned AS = Ty->getPointerAddressSpace(); | |||
3415 | SDValue N = getValue(Op0); | |||
3416 | ||||
3417 | for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end(); | |||
3418 | OI != E; ++OI) { | |||
3419 | const Value *Idx = *OI; | |||
3420 | if (StructType *StTy = dyn_cast<StructType>(Ty)) { | |||
3421 | unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue(); | |||
3422 | if (Field) { | |||
3423 | // N = N + Offset | |||
3424 | uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field); | |||
3425 | N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N, | |||
3426 | DAG.getConstant(Offset, N.getValueType())); | |||
3427 | } | |||
3428 | ||||
3429 | Ty = StTy->getElementType(Field); | |||
3430 | } else { | |||
3431 | Ty = cast<SequentialType>(Ty)->getElementType(); | |||
3432 | ||||
3433 | // If this is a constant subscript, handle it quickly. | |||
3434 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3435 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { | |||
3436 | if (CI->isZero()) continue; | |||
3437 | uint64_t Offs = | |||
3438 | DL->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); | |||
3439 | SDValue OffsVal; | |||
3440 | EVT PTy = TLI.getPointerTy(AS); | |||
3441 | unsigned PtrBits = PTy.getSizeInBits(); | |||
3442 | if (PtrBits < 64) | |||
3443 | OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), PTy, | |||
3444 | DAG.getConstant(Offs, MVT::i64)); | |||
3445 | else | |||
3446 | OffsVal = DAG.getConstant(Offs, PTy); | |||
3447 | ||||
3448 | N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N, | |||
3449 | OffsVal); | |||
3450 | continue; | |||
3451 | } | |||
3452 | ||||
3453 | // N = N + Idx * ElementSize; | |||
3454 | APInt ElementSize = | |||
3455 | APInt(TLI.getPointerSizeInBits(AS), DL->getTypeAllocSize(Ty)); | |||
3456 | SDValue IdxN = getValue(Idx); | |||
3457 | ||||
3458 | // If the index is smaller or larger than intptr_t, truncate or extend | |||
3459 | // it. | |||
3460 | IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType()); | |||
3461 | ||||
3462 | // If this is a multiply by a power of two, turn it into a shl | |||
3463 | // immediately. This is a very common case. | |||
3464 | if (ElementSize != 1) { | |||
3465 | if (ElementSize.isPowerOf2()) { | |||
3466 | unsigned Amt = ElementSize.logBase2(); | |||
3467 | IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(), | |||
3468 | N.getValueType(), IdxN, | |||
3469 | DAG.getConstant(Amt, IdxN.getValueType())); | |||
3470 | } else { | |||
3471 | SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType()); | |||
3472 | IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(), | |||
3473 | N.getValueType(), IdxN, Scale); | |||
3474 | } | |||
3475 | } | |||
3476 | ||||
3477 | N = DAG.getNode(ISD::ADD, getCurSDLoc(), | |||
3478 | N.getValueType(), N, IdxN); | |||
3479 | } | |||
3480 | } | |||
3481 | ||||
3482 | setValue(&I, N); | |||
3483 | } | |||
3484 | ||||
3485 | void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { | |||
3486 | // If this is a fixed sized alloca in the entry block of the function, | |||
3487 | // allocate it statically on the stack. | |||
3488 | if (FuncInfo.StaticAllocaMap.count(&I)) | |||
3489 | return; // getValue will auto-populate this. | |||
3490 | ||||
3491 | Type *Ty = I.getAllocatedType(); | |||
3492 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3493 | uint64_t TySize = TLI.getDataLayout()->getTypeAllocSize(Ty); | |||
3494 | unsigned Align = | |||
3495 | std::max((unsigned)TLI.getDataLayout()->getPrefTypeAlignment(Ty), | |||
3496 | I.getAlignment()); | |||
3497 | ||||
3498 | SDValue AllocSize = getValue(I.getArraySize()); | |||
3499 | ||||
3500 | EVT IntPtr = TLI.getPointerTy(); | |||
3501 | if (AllocSize.getValueType() != IntPtr) | |||
3502 | AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr); | |||
3503 | ||||
3504 | AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr, | |||
3505 | AllocSize, | |||
3506 | DAG.getConstant(TySize, IntPtr)); | |||
3507 | ||||
3508 | // Handle alignment. If the requested alignment is less than or equal to | |||
3509 | // the stack alignment, ignore it. If the size is greater than or equal to | |||
3510 | // the stack alignment, we note this in the DYNAMIC_STACKALLOC node. | |||
3511 | unsigned StackAlign = | |||
3512 | DAG.getSubtarget().getFrameLowering()->getStackAlignment(); | |||
3513 | if (Align <= StackAlign) | |||
3514 | Align = 0; | |||
3515 | ||||
3516 | // Round the size of the allocation up to the stack alignment size | |||
3517 | // by add SA-1 to the size. | |||
3518 | AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(), | |||
3519 | AllocSize.getValueType(), AllocSize, | |||
3520 | DAG.getIntPtrConstant(StackAlign-1)); | |||
3521 | ||||
3522 | // Mask out the low bits for alignment purposes. | |||
3523 | AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(), | |||
3524 | AllocSize.getValueType(), AllocSize, | |||
3525 | DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1))); | |||
3526 | ||||
3527 | SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) }; | |||
3528 | SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other); | |||
3529 | SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(), VTs, Ops); | |||
3530 | setValue(&I, DSA); | |||
3531 | DAG.setRoot(DSA.getValue(1)); | |||
3532 | ||||
3533 | assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects())((FuncInfo.MF->getFrameInfo()->hasVarSizedObjects()) ? static_cast <void> (0) : __assert_fail ("FuncInfo.MF->getFrameInfo()->hasVarSizedObjects()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 3533, __PRETTY_FUNCTION__)); | |||
3534 | } | |||
3535 | ||||
3536 | void SelectionDAGBuilder::visitLoad(const LoadInst &I) { | |||
3537 | if (I.isAtomic()) | |||
3538 | return visitAtomicLoad(I); | |||
3539 | ||||
3540 | const Value *SV = I.getOperand(0); | |||
3541 | SDValue Ptr = getValue(SV); | |||
3542 | ||||
3543 | Type *Ty = I.getType(); | |||
3544 | ||||
3545 | bool isVolatile = I.isVolatile(); | |||
3546 | bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr; | |||
3547 | bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr; | |||
3548 | unsigned Alignment = I.getAlignment(); | |||
3549 | ||||
3550 | AAMDNodes AAInfo; | |||
3551 | I.getAAMetadata(AAInfo); | |||
3552 | const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range); | |||
3553 | ||||
3554 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3555 | SmallVector<EVT, 4> ValueVTs; | |||
3556 | SmallVector<uint64_t, 4> Offsets; | |||
3557 | ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets); | |||
3558 | unsigned NumValues = ValueVTs.size(); | |||
3559 | if (NumValues == 0) | |||
3560 | return; | |||
3561 | ||||
3562 | SDValue Root; | |||
3563 | bool ConstantMemory = false; | |||
3564 | if (isVolatile || NumValues > MaxParallelChains) | |||
3565 | // Serialize volatile loads with other side effects. | |||
3566 | Root = getRoot(); | |||
3567 | else if (AA->pointsToConstantMemory( | |||
3568 | AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), AAInfo))) { | |||
3569 | // Do not serialize (non-volatile) loads of constant memory with anything. | |||
3570 | Root = DAG.getEntryNode(); | |||
3571 | ConstantMemory = true; | |||
3572 | } else { | |||
3573 | // Do not serialize non-volatile loads against each other. | |||
3574 | Root = DAG.getRoot(); | |||
3575 | } | |||
3576 | ||||
3577 | if (isVolatile) | |||
3578 | Root = TLI.prepareVolatileOrAtomicLoad(Root, getCurSDLoc(), DAG); | |||
3579 | ||||
3580 | SmallVector<SDValue, 4> Values(NumValues); | |||
3581 | SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains), | |||
3582 | NumValues)); | |||
3583 | EVT PtrVT = Ptr.getValueType(); | |||
3584 | unsigned ChainI = 0; | |||
3585 | for (unsigned i = 0; i != NumValues; ++i, ++ChainI) { | |||
3586 | // Serializing loads here may result in excessive register pressure, and | |||
3587 | // TokenFactor places arbitrary choke points on the scheduler. SD scheduling | |||
3588 | // could recover a bit by hoisting nodes upward in the chain by recognizing | |||
3589 | // they are side-effect free or do not alias. The optimizer should really | |||
3590 | // avoid this case by converting large object/array copies to llvm.memcpy | |||
3591 | // (MaxParallelChains should always remain as failsafe). | |||
3592 | if (ChainI == MaxParallelChains) { | |||
3593 | assert(PendingLoads.empty() && "PendingLoads must be serialized first")((PendingLoads.empty() && "PendingLoads must be serialized first" ) ? static_cast<void> (0) : __assert_fail ("PendingLoads.empty() && \"PendingLoads must be serialized first\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 3593, __PRETTY_FUNCTION__)); | |||
3594 | SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, | |||
3595 | makeArrayRef(Chains.data(), ChainI)); | |||
3596 | Root = Chain; | |||
3597 | ChainI = 0; | |||
3598 | } | |||
3599 | SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(), | |||
3600 | PtrVT, Ptr, | |||
3601 | DAG.getConstant(Offsets[i], PtrVT)); | |||
3602 | SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root, | |||
3603 | A, MachinePointerInfo(SV, Offsets[i]), isVolatile, | |||
3604 | isNonTemporal, isInvariant, Alignment, AAInfo, | |||
3605 | Ranges); | |||
3606 | ||||
3607 | Values[i] = L; | |||
3608 | Chains[ChainI] = L.getValue(1); | |||
3609 | } | |||
3610 | ||||
3611 | if (!ConstantMemory) { | |||
3612 | SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, | |||
3613 | makeArrayRef(Chains.data(), ChainI)); | |||
3614 | if (isVolatile) | |||
3615 | DAG.setRoot(Chain); | |||
3616 | else | |||
3617 | PendingLoads.push_back(Chain); | |||
3618 | } | |||
3619 | ||||
3620 | setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), | |||
3621 | DAG.getVTList(ValueVTs), Values)); | |||
3622 | } | |||
3623 | ||||
3624 | void SelectionDAGBuilder::visitStore(const StoreInst &I) { | |||
3625 | if (I.isAtomic()) | |||
3626 | return visitAtomicStore(I); | |||
3627 | ||||
3628 | const Value *SrcV = I.getOperand(0); | |||
3629 | const Value *PtrV = I.getOperand(1); | |||
3630 | ||||
3631 | SmallVector<EVT, 4> ValueVTs; | |||
3632 | SmallVector<uint64_t, 4> Offsets; | |||
3633 | ComputeValueVTs(DAG.getTargetLoweringInfo(), SrcV->getType(), | |||
3634 | ValueVTs, &Offsets); | |||
3635 | unsigned NumValues = ValueVTs.size(); | |||
3636 | if (NumValues == 0) | |||
3637 | return; | |||
3638 | ||||
3639 | // Get the lowered operands. Note that we do this after | |||
3640 | // checking if NumResults is zero, because with zero results | |||
3641 | // the operands won't have values in the map. | |||
3642 | SDValue Src = getValue(SrcV); | |||
3643 | SDValue Ptr = getValue(PtrV); | |||
3644 | ||||
3645 | SDValue Root = getRoot(); | |||
3646 | SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains), | |||
3647 | NumValues)); | |||
3648 | EVT PtrVT = Ptr.getValueType(); | |||
3649 | bool isVolatile = I.isVolatile(); | |||
3650 | bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr; | |||
3651 | unsigned Alignment = I.getAlignment(); | |||
3652 | ||||
3653 | AAMDNodes AAInfo; | |||
3654 | I.getAAMetadata(AAInfo); | |||
3655 | ||||
3656 | unsigned ChainI = 0; | |||
3657 | for (unsigned i = 0; i != NumValues; ++i, ++ChainI) { | |||
3658 | // See visitLoad comments. | |||
3659 | if (ChainI == MaxParallelChains) { | |||
3660 | SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, | |||
3661 | makeArrayRef(Chains.data(), ChainI)); | |||
3662 | Root = Chain; | |||
3663 | ChainI = 0; | |||
3664 | } | |||
3665 | SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr, | |||
3666 | DAG.getConstant(Offsets[i], PtrVT)); | |||
3667 | SDValue St = DAG.getStore(Root, getCurSDLoc(), | |||
3668 | SDValue(Src.getNode(), Src.getResNo() + i), | |||
3669 | Add, MachinePointerInfo(PtrV, Offsets[i]), | |||
3670 | isVolatile, isNonTemporal, Alignment, AAInfo); | |||
3671 | Chains[ChainI] = St; | |||
3672 | } | |||
3673 | ||||
3674 | SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, | |||
3675 | makeArrayRef(Chains.data(), ChainI)); | |||
3676 | DAG.setRoot(StoreNode); | |||
3677 | } | |||
3678 | ||||
3679 | void SelectionDAGBuilder::visitMaskedStore(const CallInst &I) { | |||
3680 | SDLoc sdl = getCurSDLoc(); | |||
3681 | ||||
3682 | // llvm.masked.store.*(Src0, Ptr, alignemt, Mask) | |||
3683 | Value *PtrOperand = I.getArgOperand(1); | |||
3684 | SDValue Ptr = getValue(PtrOperand); | |||
3685 | SDValue Src0 = getValue(I.getArgOperand(0)); | |||
3686 | SDValue Mask = getValue(I.getArgOperand(3)); | |||
3687 | EVT VT = Src0.getValueType(); | |||
3688 | unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue(); | |||
3689 | if (!Alignment) | |||
3690 | Alignment = DAG.getEVTAlignment(VT); | |||
3691 | ||||
3692 | AAMDNodes AAInfo; | |||
3693 | I.getAAMetadata(AAInfo); | |||
3694 | ||||
3695 | MachineMemOperand *MMO = | |||
3696 | DAG.getMachineFunction(). | |||
3697 | getMachineMemOperand(MachinePointerInfo(PtrOperand), | |||
3698 | MachineMemOperand::MOStore, VT.getStoreSize(), | |||
3699 | Alignment, AAInfo); | |||
3700 | SDValue StoreNode = DAG.getMaskedStore(getRoot(), sdl, Src0, Ptr, Mask, VT, | |||
3701 | MMO, false); | |||
3702 | DAG.setRoot(StoreNode); | |||
3703 | setValue(&I, StoreNode); | |||
3704 | } | |||
3705 | ||||
3706 | void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I) { | |||
3707 | SDLoc sdl = getCurSDLoc(); | |||
3708 | ||||
3709 | // @llvm.masked.load.*(Ptr, alignment, Mask, Src0) | |||
3710 | Value *PtrOperand = I.getArgOperand(0); | |||
3711 | SDValue Ptr = getValue(PtrOperand); | |||
3712 | SDValue Src0 = getValue(I.getArgOperand(3)); | |||
3713 | SDValue Mask = getValue(I.getArgOperand(2)); | |||
3714 | ||||
3715 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3716 | EVT VT = TLI.getValueType(I.getType()); | |||
3717 | unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue(); | |||
3718 | if (!Alignment) | |||
3719 | Alignment = DAG.getEVTAlignment(VT); | |||
3720 | ||||
3721 | AAMDNodes AAInfo; | |||
3722 | I.getAAMetadata(AAInfo); | |||
3723 | const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range); | |||
3724 | ||||
3725 | SDValue InChain = DAG.getRoot(); | |||
3726 | if (AA->pointsToConstantMemory( | |||
3727 | AliasAnalysis::Location(PtrOperand, | |||
3728 | AA->getTypeStoreSize(I.getType()), | |||
3729 | AAInfo))) { | |||
3730 | // Do not serialize (non-volatile) loads of constant memory with anything. | |||
3731 | InChain = DAG.getEntryNode(); | |||
3732 | } | |||
3733 | ||||
3734 | MachineMemOperand *MMO = | |||
3735 | DAG.getMachineFunction(). | |||
3736 | getMachineMemOperand(MachinePointerInfo(PtrOperand), | |||
3737 | MachineMemOperand::MOLoad, VT.getStoreSize(), | |||
3738 | Alignment, AAInfo, Ranges); | |||
3739 | ||||
3740 | SDValue Load = DAG.getMaskedLoad(VT, sdl, InChain, Ptr, Mask, Src0, VT, MMO, | |||
3741 | ISD::NON_EXTLOAD); | |||
3742 | SDValue OutChain = Load.getValue(1); | |||
3743 | DAG.setRoot(OutChain); | |||
3744 | setValue(&I, Load); | |||
3745 | } | |||
3746 | ||||
3747 | void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) { | |||
3748 | SDLoc dl = getCurSDLoc(); | |||
3749 | AtomicOrdering SuccessOrder = I.getSuccessOrdering(); | |||
3750 | AtomicOrdering FailureOrder = I.getFailureOrdering(); | |||
3751 | SynchronizationScope Scope = I.getSynchScope(); | |||
3752 | ||||
3753 | SDValue InChain = getRoot(); | |||
3754 | ||||
3755 | MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType(); | |||
3756 | SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other); | |||
3757 | SDValue L = DAG.getAtomicCmpSwap( | |||
3758 | ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain, | |||
3759 | getValue(I.getPointerOperand()), getValue(I.getCompareOperand()), | |||
3760 | getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()), | |||
3761 | /*Alignment=*/ 0, SuccessOrder, FailureOrder, Scope); | |||
3762 | ||||
3763 | SDValue OutChain = L.getValue(2); | |||
3764 | ||||
3765 | setValue(&I, L); | |||
3766 | DAG.setRoot(OutChain); | |||
3767 | } | |||
3768 | ||||
3769 | void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) { | |||
3770 | SDLoc dl = getCurSDLoc(); | |||
3771 | ISD::NodeType NT; | |||
3772 | switch (I.getOperation()) { | |||
3773 | default: llvm_unreachable("Unknown atomicrmw operation")::llvm::llvm_unreachable_internal("Unknown atomicrmw operation" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 3773); | |||
3774 | case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break; | |||
3775 | case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break; | |||
3776 | case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break; | |||
3777 | case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break; | |||
3778 | case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break; | |||
3779 | case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break; | |||
3780 | case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break; | |||
3781 | case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break; | |||
3782 | case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break; | |||
3783 | case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break; | |||
3784 | case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break; | |||
3785 | } | |||
3786 | AtomicOrdering Order = I.getOrdering(); | |||
3787 | SynchronizationScope Scope = I.getSynchScope(); | |||
3788 | ||||
3789 | SDValue InChain = getRoot(); | |||
3790 | ||||
3791 | SDValue L = | |||
3792 | DAG.getAtomic(NT, dl, | |||
3793 | getValue(I.getValOperand()).getSimpleValueType(), | |||
3794 | InChain, | |||
3795 | getValue(I.getPointerOperand()), | |||
3796 | getValue(I.getValOperand()), | |||
3797 | I.getPointerOperand(), | |||
3798 | /* Alignment=*/ 0, Order, Scope); | |||
3799 | ||||
3800 | SDValue OutChain = L.getValue(1); | |||
3801 | ||||
3802 | setValue(&I, L); | |||
3803 | DAG.setRoot(OutChain); | |||
3804 | } | |||
3805 | ||||
3806 | void SelectionDAGBuilder::visitFence(const FenceInst &I) { | |||
3807 | SDLoc dl = getCurSDLoc(); | |||
3808 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3809 | SDValue Ops[3]; | |||
3810 | Ops[0] = getRoot(); | |||
3811 | Ops[1] = DAG.getConstant(I.getOrdering(), TLI.getPointerTy()); | |||
3812 | Ops[2] = DAG.getConstant(I.getSynchScope(), TLI.getPointerTy()); | |||
3813 | DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops)); | |||
3814 | } | |||
3815 | ||||
3816 | void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) { | |||
3817 | SDLoc dl = getCurSDLoc(); | |||
3818 | AtomicOrdering Order = I.getOrdering(); | |||
3819 | SynchronizationScope Scope = I.getSynchScope(); | |||
3820 | ||||
3821 | SDValue InChain = getRoot(); | |||
3822 | ||||
3823 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3824 | EVT VT = TLI.getValueType(I.getType()); | |||
3825 | ||||
3826 | if (I.getAlignment() < VT.getSizeInBits() / 8) | |||
3827 | report_fatal_error("Cannot generate unaligned atomic load"); | |||
3828 | ||||
3829 | MachineMemOperand *MMO = | |||
3830 | DAG.getMachineFunction(). | |||
3831 | getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()), | |||
3832 | MachineMemOperand::MOVolatile | | |||
3833 | MachineMemOperand::MOLoad, | |||
3834 | VT.getStoreSize(), | |||
3835 | I.getAlignment() ? I.getAlignment() : | |||
3836 | DAG.getEVTAlignment(VT)); | |||
3837 | ||||
3838 | InChain = TLI.prepareVolatileOrAtomicLoad(InChain, dl, DAG); | |||
3839 | SDValue L = | |||
3840 | DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain, | |||
3841 | getValue(I.getPointerOperand()), MMO, | |||
3842 | Order, Scope); | |||
3843 | ||||
3844 | SDValue OutChain = L.getValue(1); | |||
3845 | ||||
3846 | setValue(&I, L); | |||
3847 | DAG.setRoot(OutChain); | |||
3848 | } | |||
3849 | ||||
3850 | void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) { | |||
3851 | SDLoc dl = getCurSDLoc(); | |||
3852 | ||||
3853 | AtomicOrdering Order = I.getOrdering(); | |||
3854 | SynchronizationScope Scope = I.getSynchScope(); | |||
3855 | ||||
3856 | SDValue InChain = getRoot(); | |||
3857 | ||||
3858 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3859 | EVT VT = TLI.getValueType(I.getValueOperand()->getType()); | |||
3860 | ||||
3861 | if (I.getAlignment() < VT.getSizeInBits() / 8) | |||
3862 | report_fatal_error("Cannot generate unaligned atomic store"); | |||
3863 | ||||
3864 | SDValue OutChain = | |||
3865 | DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT, | |||
3866 | InChain, | |||
3867 | getValue(I.getPointerOperand()), | |||
3868 | getValue(I.getValueOperand()), | |||
3869 | I.getPointerOperand(), I.getAlignment(), | |||
3870 | Order, Scope); | |||
3871 | ||||
3872 | DAG.setRoot(OutChain); | |||
3873 | } | |||
3874 | ||||
3875 | /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC | |||
3876 | /// node. | |||
3877 | void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I, | |||
3878 | unsigned Intrinsic) { | |||
3879 | bool HasChain = !I.doesNotAccessMemory(); | |||
3880 | bool OnlyLoad = HasChain && I.onlyReadsMemory(); | |||
3881 | ||||
3882 | // Build the operand list. | |||
3883 | SmallVector<SDValue, 8> Ops; | |||
3884 | if (HasChain) { // If this intrinsic has side-effects, chainify it. | |||
3885 | if (OnlyLoad) { | |||
3886 | // We don't need to serialize loads against other loads. | |||
3887 | Ops.push_back(DAG.getRoot()); | |||
3888 | } else { | |||
3889 | Ops.push_back(getRoot()); | |||
3890 | } | |||
3891 | } | |||
3892 | ||||
3893 | // Info is set by getTgtMemInstrinsic | |||
3894 | TargetLowering::IntrinsicInfo Info; | |||
3895 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
3896 | bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic); | |||
3897 | ||||
3898 | // Add the intrinsic ID as an integer operand if it's not a target intrinsic. | |||
3899 | if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID || | |||
3900 | Info.opc == ISD::INTRINSIC_W_CHAIN) | |||
3901 | Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI.getPointerTy())); | |||
3902 | ||||
3903 | // Add all operands of the call to the operand list. | |||
3904 | for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) { | |||
3905 | SDValue Op = getValue(I.getArgOperand(i)); | |||
3906 | Ops.push_back(Op); | |||
3907 | } | |||
3908 | ||||
3909 | SmallVector<EVT, 4> ValueVTs; | |||
3910 | ComputeValueVTs(TLI, I.getType(), ValueVTs); | |||
3911 | ||||
3912 | if (HasChain) | |||
3913 | ValueVTs.push_back(MVT::Other); | |||
3914 | ||||
3915 | SDVTList VTs = DAG.getVTList(ValueVTs); | |||
3916 | ||||
3917 | // Create the node. | |||
3918 | SDValue Result; | |||
3919 | if (IsTgtIntrinsic) { | |||
3920 | // This is target intrinsic that touches memory | |||
3921 | Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(), | |||
3922 | VTs, Ops, Info.memVT, | |||
3923 | MachinePointerInfo(Info.ptrVal, Info.offset), | |||
3924 | Info.align, Info.vol, | |||
3925 | Info.readMem, Info.writeMem, Info.size); | |||
3926 | } else if (!HasChain) { | |||
3927 | Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops); | |||
3928 | } else if (!I.getType()->isVoidTy()) { | |||
3929 | Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops); | |||
3930 | } else { | |||
3931 | Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops); | |||
3932 | } | |||
3933 | ||||
3934 | if (HasChain) { | |||
3935 | SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1); | |||
3936 | if (OnlyLoad) | |||
3937 | PendingLoads.push_back(Chain); | |||
3938 | else | |||
3939 | DAG.setRoot(Chain); | |||
3940 | } | |||
3941 | ||||
3942 | if (!I.getType()->isVoidTy()) { | |||
3943 | if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) { | |||
3944 | EVT VT = TLI.getValueType(PTy); | |||
3945 | Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result); | |||
3946 | } | |||
3947 | ||||
3948 | setValue(&I, Result); | |||
3949 | } | |||
3950 | } | |||
3951 | ||||
3952 | /// GetSignificand - Get the significand and build it into a floating-point | |||
3953 | /// number with exponent of 1: | |||
3954 | /// | |||
3955 | /// Op = (Op & 0x007fffff) | 0x3f800000; | |||
3956 | /// | |||
3957 | /// where Op is the hexadecimal representation of floating point value. | |||
3958 | static SDValue | |||
3959 | GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) { | |||
3960 | SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, | |||
3961 | DAG.getConstant(0x007fffff, MVT::i32)); | |||
3962 | SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1, | |||
3963 | DAG.getConstant(0x3f800000, MVT::i32)); | |||
3964 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2); | |||
3965 | } | |||
3966 | ||||
3967 | /// GetExponent - Get the exponent: | |||
3968 | /// | |||
3969 | /// (float)(int)(((Op & 0x7f800000) >> 23) - 127); | |||
3970 | /// | |||
3971 | /// where Op is the hexadecimal representation of floating point value. | |||
3972 | static SDValue | |||
3973 | GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI, | |||
3974 | SDLoc dl) { | |||
3975 | SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, | |||
3976 | DAG.getConstant(0x7f800000, MVT::i32)); | |||
3977 | SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0, | |||
3978 | DAG.getConstant(23, TLI.getPointerTy())); | |||
3979 | SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1, | |||
3980 | DAG.getConstant(127, MVT::i32)); | |||
3981 | return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2); | |||
3982 | } | |||
3983 | ||||
3984 | /// getF32Constant - Get 32-bit floating point constant. | |||
3985 | static SDValue | |||
3986 | getF32Constant(SelectionDAG &DAG, unsigned Flt) { | |||
3987 | return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), | |||
3988 | MVT::f32); | |||
3989 | } | |||
3990 | ||||
3991 | /// expandExp - Lower an exp intrinsic. Handles the special sequences for | |||
3992 | /// limited-precision mode. | |||
3993 | static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG, | |||
3994 | const TargetLowering &TLI) { | |||
3995 | if (Op.getValueType() == MVT::f32 && | |||
3996 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { | |||
3997 | ||||
3998 | // Put the exponent in the right bit position for later addition to the | |||
3999 | // final result: | |||
4000 | // | |||
4001 | // #define LOG2OFe 1.4426950f | |||
4002 | // IntegerPartOfX = ((int32_t)(X * LOG2OFe)); | |||
4003 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, | |||
4004 | getF32Constant(DAG, 0x3fb8aa3b)); | |||
4005 | SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); | |||
4006 | ||||
4007 | // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX; | |||
4008 | SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); | |||
4009 | SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); | |||
4010 | ||||
4011 | // IntegerPartOfX <<= 23; | |||
4012 | IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, | |||
4013 | DAG.getConstant(23, TLI.getPointerTy())); | |||
4014 | ||||
4015 | SDValue TwoToFracPartOfX; | |||
4016 | if (LimitFloatPrecision <= 6) { | |||
4017 | // For floating-point precision of 6: | |||
4018 | // | |||
4019 | // TwoToFractionalPartOfX = | |||
4020 | // 0.997535578f + | |||
4021 | // (0.735607626f + 0.252464424f * x) * x; | |||
4022 | // | |||
4023 | // error 0.0144103317, which is 6 bits | |||
4024 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4025 | getF32Constant(DAG, 0x3e814304)); | |||
4026 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4027 | getF32Constant(DAG, 0x3f3c50c8)); | |||
4028 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4029 | TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4030 | getF32Constant(DAG, 0x3f7f5e7e)); | |||
4031 | } else if (LimitFloatPrecision <= 12) { | |||
4032 | // For floating-point precision of 12: | |||
4033 | // | |||
4034 | // TwoToFractionalPartOfX = | |||
4035 | // 0.999892986f + | |||
4036 | // (0.696457318f + | |||
4037 | // (0.224338339f + 0.792043434e-1f * x) * x) * x; | |||
4038 | // | |||
4039 | // 0.000107046256 error, which is 13 to 14 bits | |||
4040 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4041 | getF32Constant(DAG, 0x3da235e3)); | |||
4042 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4043 | getF32Constant(DAG, 0x3e65b8f3)); | |||
4044 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4045 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4046 | getF32Constant(DAG, 0x3f324b07)); | |||
4047 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4048 | TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4049 | getF32Constant(DAG, 0x3f7ff8fd)); | |||
4050 | } else { // LimitFloatPrecision <= 18 | |||
4051 | // For floating-point precision of 18: | |||
4052 | // | |||
4053 | // TwoToFractionalPartOfX = | |||
4054 | // 0.999999982f + | |||
4055 | // (0.693148872f + | |||
4056 | // (0.240227044f + | |||
4057 | // (0.554906021e-1f + | |||
4058 | // (0.961591928e-2f + | |||
4059 | // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; | |||
4060 | // | |||
4061 | // error 2.47208000*10^(-7), which is better than 18 bits | |||
4062 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4063 | getF32Constant(DAG, 0x3924b03e)); | |||
4064 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4065 | getF32Constant(DAG, 0x3ab24b87)); | |||
4066 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4067 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4068 | getF32Constant(DAG, 0x3c1d8c17)); | |||
4069 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4070 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4071 | getF32Constant(DAG, 0x3d634a1d)); | |||
4072 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); | |||
4073 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, | |||
4074 | getF32Constant(DAG, 0x3e75fe14)); | |||
4075 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); | |||
4076 | SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, | |||
4077 | getF32Constant(DAG, 0x3f317234)); | |||
4078 | SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); | |||
4079 | TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, | |||
4080 | getF32Constant(DAG, 0x3f800000)); | |||
4081 | } | |||
4082 | ||||
4083 | // Add the exponent into the result in integer domain. | |||
4084 | SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX); | |||
4085 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, | |||
4086 | DAG.getNode(ISD::ADD, dl, MVT::i32, | |||
4087 | t13, IntegerPartOfX)); | |||
4088 | } | |||
4089 | ||||
4090 | // No special expansion. | |||
4091 | return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op); | |||
4092 | } | |||
4093 | ||||
4094 | /// expandLog - Lower a log intrinsic. Handles the special sequences for | |||
4095 | /// limited-precision mode. | |||
4096 | static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG, | |||
4097 | const TargetLowering &TLI) { | |||
4098 | if (Op.getValueType() == MVT::f32 && | |||
4099 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { | |||
4100 | SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); | |||
4101 | ||||
4102 | // Scale the exponent by log(2) [0.69314718f]. | |||
4103 | SDValue Exp = GetExponent(DAG, Op1, TLI, dl); | |||
4104 | SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, | |||
4105 | getF32Constant(DAG, 0x3f317218)); | |||
4106 | ||||
4107 | // Get the significand and build it into a floating-point number with | |||
4108 | // exponent of 1. | |||
4109 | SDValue X = GetSignificand(DAG, Op1, dl); | |||
4110 | ||||
4111 | SDValue LogOfMantissa; | |||
4112 | if (LimitFloatPrecision <= 6) { | |||
4113 | // For floating-point precision of 6: | |||
4114 | // | |||
4115 | // LogofMantissa = | |||
4116 | // -1.1609546f + | |||
4117 | // (1.4034025f - 0.23903021f * x) * x; | |||
4118 | // | |||
4119 | // error 0.0034276066, which is better than 8 bits | |||
4120 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4121 | getF32Constant(DAG, 0xbe74c456)); | |||
4122 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4123 | getF32Constant(DAG, 0x3fb3a2b1)); | |||
4124 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4125 | LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4126 | getF32Constant(DAG, 0x3f949a29)); | |||
4127 | } else if (LimitFloatPrecision <= 12) { | |||
4128 | // For floating-point precision of 12: | |||
4129 | // | |||
4130 | // LogOfMantissa = | |||
4131 | // -1.7417939f + | |||
4132 | // (2.8212026f + | |||
4133 | // (-1.4699568f + | |||
4134 | // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x; | |||
4135 | // | |||
4136 | // error 0.000061011436, which is 14 bits | |||
4137 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4138 | getF32Constant(DAG, 0xbd67b6d6)); | |||
4139 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4140 | getF32Constant(DAG, 0x3ee4f4b8)); | |||
4141 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4142 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4143 | getF32Constant(DAG, 0x3fbc278b)); | |||
4144 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4145 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4146 | getF32Constant(DAG, 0x40348e95)); | |||
4147 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4148 | LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, | |||
4149 | getF32Constant(DAG, 0x3fdef31a)); | |||
4150 | } else { // LimitFloatPrecision <= 18 | |||
4151 | // For floating-point precision of 18: | |||
4152 | // | |||
4153 | // LogOfMantissa = | |||
4154 | // -2.1072184f + | |||
4155 | // (4.2372794f + | |||
4156 | // (-3.7029485f + | |||
4157 | // (2.2781945f + | |||
4158 | // (-0.87823314f + | |||
4159 | // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x; | |||
4160 | // | |||
4161 | // error 0.0000023660568, which is better than 18 bits | |||
4162 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4163 | getF32Constant(DAG, 0xbc91e5ac)); | |||
4164 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4165 | getF32Constant(DAG, 0x3e4350aa)); | |||
4166 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4167 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4168 | getF32Constant(DAG, 0x3f60d3e3)); | |||
4169 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4170 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4171 | getF32Constant(DAG, 0x4011cdf0)); | |||
4172 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4173 | SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, | |||
4174 | getF32Constant(DAG, 0x406cfd1c)); | |||
4175 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); | |||
4176 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, | |||
4177 | getF32Constant(DAG, 0x408797cb)); | |||
4178 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); | |||
4179 | LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, | |||
4180 | getF32Constant(DAG, 0x4006dcab)); | |||
4181 | } | |||
4182 | ||||
4183 | return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa); | |||
4184 | } | |||
4185 | ||||
4186 | // No special expansion. | |||
4187 | return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op); | |||
4188 | } | |||
4189 | ||||
4190 | /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for | |||
4191 | /// limited-precision mode. | |||
4192 | static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG, | |||
4193 | const TargetLowering &TLI) { | |||
4194 | if (Op.getValueType() == MVT::f32 && | |||
4195 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { | |||
4196 | SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); | |||
4197 | ||||
4198 | // Get the exponent. | |||
4199 | SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl); | |||
4200 | ||||
4201 | // Get the significand and build it into a floating-point number with | |||
4202 | // exponent of 1. | |||
4203 | SDValue X = GetSignificand(DAG, Op1, dl); | |||
4204 | ||||
4205 | // Different possible minimax approximations of significand in | |||
4206 | // floating-point for various degrees of accuracy over [1,2]. | |||
4207 | SDValue Log2ofMantissa; | |||
4208 | if (LimitFloatPrecision <= 6) { | |||
4209 | // For floating-point precision of 6: | |||
4210 | // | |||
4211 | // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x; | |||
4212 | // | |||
4213 | // error 0.0049451742, which is more than 7 bits | |||
4214 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4215 | getF32Constant(DAG, 0xbeb08fe0)); | |||
4216 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4217 | getF32Constant(DAG, 0x40019463)); | |||
4218 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4219 | Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4220 | getF32Constant(DAG, 0x3fd6633d)); | |||
4221 | } else if (LimitFloatPrecision <= 12) { | |||
4222 | // For floating-point precision of 12: | |||
4223 | // | |||
4224 | // Log2ofMantissa = | |||
4225 | // -2.51285454f + | |||
4226 | // (4.07009056f + | |||
4227 | // (-2.12067489f + | |||
4228 | // (.645142248f - 0.816157886e-1f * x) * x) * x) * x; | |||
4229 | // | |||
4230 | // error 0.0000876136000, which is better than 13 bits | |||
4231 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4232 | getF32Constant(DAG, 0xbda7262e)); | |||
4233 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4234 | getF32Constant(DAG, 0x3f25280b)); | |||
4235 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4236 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4237 | getF32Constant(DAG, 0x4007b923)); | |||
4238 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4239 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4240 | getF32Constant(DAG, 0x40823e2f)); | |||
4241 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4242 | Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, | |||
4243 | getF32Constant(DAG, 0x4020d29c)); | |||
4244 | } else { // LimitFloatPrecision <= 18 | |||
4245 | // For floating-point precision of 18: | |||
4246 | // | |||
4247 | // Log2ofMantissa = | |||
4248 | // -3.0400495f + | |||
4249 | // (6.1129976f + | |||
4250 | // (-5.3420409f + | |||
4251 | // (3.2865683f + | |||
4252 | // (-1.2669343f + | |||
4253 | // (0.27515199f - | |||
4254 | // 0.25691327e-1f * x) * x) * x) * x) * x) * x; | |||
4255 | // | |||
4256 | // error 0.0000018516, which is better than 18 bits | |||
4257 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4258 | getF32Constant(DAG, 0xbcd2769e)); | |||
4259 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4260 | getF32Constant(DAG, 0x3e8ce0b9)); | |||
4261 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4262 | SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4263 | getF32Constant(DAG, 0x3fa22ae7)); | |||
4264 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4265 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4266 | getF32Constant(DAG, 0x40525723)); | |||
4267 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4268 | SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, | |||
4269 | getF32Constant(DAG, 0x40aaf200)); | |||
4270 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); | |||
4271 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, | |||
4272 | getF32Constant(DAG, 0x40c39dad)); | |||
4273 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); | |||
4274 | Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, | |||
4275 | getF32Constant(DAG, 0x4042902c)); | |||
4276 | } | |||
4277 | ||||
4278 | return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa); | |||
4279 | } | |||
4280 | ||||
4281 | // No special expansion. | |||
4282 | return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op); | |||
4283 | } | |||
4284 | ||||
4285 | /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for | |||
4286 | /// limited-precision mode. | |||
4287 | static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG, | |||
4288 | const TargetLowering &TLI) { | |||
4289 | if (Op.getValueType() == MVT::f32 && | |||
4290 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { | |||
4291 | SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); | |||
4292 | ||||
4293 | // Scale the exponent by log10(2) [0.30102999f]. | |||
4294 | SDValue Exp = GetExponent(DAG, Op1, TLI, dl); | |||
4295 | SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, | |||
4296 | getF32Constant(DAG, 0x3e9a209a)); | |||
4297 | ||||
4298 | // Get the significand and build it into a floating-point number with | |||
4299 | // exponent of 1. | |||
4300 | SDValue X = GetSignificand(DAG, Op1, dl); | |||
4301 | ||||
4302 | SDValue Log10ofMantissa; | |||
4303 | if (LimitFloatPrecision <= 6) { | |||
4304 | // For floating-point precision of 6: | |||
4305 | // | |||
4306 | // Log10ofMantissa = | |||
4307 | // -0.50419619f + | |||
4308 | // (0.60948995f - 0.10380950f * x) * x; | |||
4309 | // | |||
4310 | // error 0.0014886165, which is 6 bits | |||
4311 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4312 | getF32Constant(DAG, 0xbdd49a13)); | |||
4313 | SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, | |||
4314 | getF32Constant(DAG, 0x3f1c0789)); | |||
4315 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4316 | Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, | |||
4317 | getF32Constant(DAG, 0x3f011300)); | |||
4318 | } else if (LimitFloatPrecision <= 12) { | |||
4319 | // For floating-point precision of 12: | |||
4320 | // | |||
4321 | // Log10ofMantissa = | |||
4322 | // -0.64831180f + | |||
4323 | // (0.91751397f + | |||
4324 | // (-0.31664806f + 0.47637168e-1f * x) * x) * x; | |||
4325 | // | |||
4326 | // error 0.00019228036, which is better than 12 bits | |||
4327 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4328 | getF32Constant(DAG, 0x3d431f31)); | |||
4329 | SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, | |||
4330 | getF32Constant(DAG, 0x3ea21fb2)); | |||
4331 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4332 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4333 | getF32Constant(DAG, 0x3f6ae232)); | |||
4334 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4335 | Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, | |||
4336 | getF32Constant(DAG, 0x3f25f7c3)); | |||
4337 | } else { // LimitFloatPrecision <= 18 | |||
4338 | // For floating-point precision of 18: | |||
4339 | // | |||
4340 | // Log10ofMantissa = | |||
4341 | // -0.84299375f + | |||
4342 | // (1.5327582f + | |||
4343 | // (-1.0688956f + | |||
4344 | // (0.49102474f + | |||
4345 | // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x; | |||
4346 | // | |||
4347 | // error 0.0000037995730, which is better than 18 bits | |||
4348 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4349 | getF32Constant(DAG, 0x3c5d51ce)); | |||
4350 | SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, | |||
4351 | getF32Constant(DAG, 0x3e00685a)); | |||
4352 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); | |||
4353 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4354 | getF32Constant(DAG, 0x3efb6798)); | |||
4355 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4356 | SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, | |||
4357 | getF32Constant(DAG, 0x3f88d192)); | |||
4358 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4359 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4360 | getF32Constant(DAG, 0x3fc4316c)); | |||
4361 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); | |||
4362 | Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8, | |||
4363 | getF32Constant(DAG, 0x3f57ce70)); | |||
4364 | } | |||
4365 | ||||
4366 | return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa); | |||
4367 | } | |||
4368 | ||||
4369 | // No special expansion. | |||
4370 | return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op); | |||
4371 | } | |||
4372 | ||||
4373 | /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for | |||
4374 | /// limited-precision mode. | |||
4375 | static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG, | |||
4376 | const TargetLowering &TLI) { | |||
4377 | if (Op.getValueType() == MVT::f32 && | |||
4378 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { | |||
4379 | SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op); | |||
4380 | ||||
4381 | // FractionalPartOfX = x - (float)IntegerPartOfX; | |||
4382 | SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); | |||
4383 | SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1); | |||
4384 | ||||
4385 | // IntegerPartOfX <<= 23; | |||
4386 | IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, | |||
4387 | DAG.getConstant(23, TLI.getPointerTy())); | |||
4388 | ||||
4389 | SDValue TwoToFractionalPartOfX; | |||
4390 | if (LimitFloatPrecision <= 6) { | |||
4391 | // For floating-point precision of 6: | |||
4392 | // | |||
4393 | // TwoToFractionalPartOfX = | |||
4394 | // 0.997535578f + | |||
4395 | // (0.735607626f + 0.252464424f * x) * x; | |||
4396 | // | |||
4397 | // error 0.0144103317, which is 6 bits | |||
4398 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4399 | getF32Constant(DAG, 0x3e814304)); | |||
4400 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4401 | getF32Constant(DAG, 0x3f3c50c8)); | |||
4402 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4403 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4404 | getF32Constant(DAG, 0x3f7f5e7e)); | |||
4405 | } else if (LimitFloatPrecision <= 12) { | |||
4406 | // For floating-point precision of 12: | |||
4407 | // | |||
4408 | // TwoToFractionalPartOfX = | |||
4409 | // 0.999892986f + | |||
4410 | // (0.696457318f + | |||
4411 | // (0.224338339f + 0.792043434e-1f * x) * x) * x; | |||
4412 | // | |||
4413 | // error 0.000107046256, which is 13 to 14 bits | |||
4414 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4415 | getF32Constant(DAG, 0x3da235e3)); | |||
4416 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4417 | getF32Constant(DAG, 0x3e65b8f3)); | |||
4418 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4419 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4420 | getF32Constant(DAG, 0x3f324b07)); | |||
4421 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4422 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4423 | getF32Constant(DAG, 0x3f7ff8fd)); | |||
4424 | } else { // LimitFloatPrecision <= 18 | |||
4425 | // For floating-point precision of 18: | |||
4426 | // | |||
4427 | // TwoToFractionalPartOfX = | |||
4428 | // 0.999999982f + | |||
4429 | // (0.693148872f + | |||
4430 | // (0.240227044f + | |||
4431 | // (0.554906021e-1f + | |||
4432 | // (0.961591928e-2f + | |||
4433 | // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; | |||
4434 | // error 2.47208000*10^(-7), which is better than 18 bits | |||
4435 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4436 | getF32Constant(DAG, 0x3924b03e)); | |||
4437 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4438 | getF32Constant(DAG, 0x3ab24b87)); | |||
4439 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4440 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4441 | getF32Constant(DAG, 0x3c1d8c17)); | |||
4442 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4443 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4444 | getF32Constant(DAG, 0x3d634a1d)); | |||
4445 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); | |||
4446 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, | |||
4447 | getF32Constant(DAG, 0x3e75fe14)); | |||
4448 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); | |||
4449 | SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, | |||
4450 | getF32Constant(DAG, 0x3f317234)); | |||
4451 | SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); | |||
4452 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, | |||
4453 | getF32Constant(DAG, 0x3f800000)); | |||
4454 | } | |||
4455 | ||||
4456 | // Add the exponent into the result in integer domain. | |||
4457 | SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, | |||
4458 | TwoToFractionalPartOfX); | |||
4459 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, | |||
4460 | DAG.getNode(ISD::ADD, dl, MVT::i32, | |||
4461 | t13, IntegerPartOfX)); | |||
4462 | } | |||
4463 | ||||
4464 | // No special expansion. | |||
4465 | return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op); | |||
4466 | } | |||
4467 | ||||
4468 | /// visitPow - Lower a pow intrinsic. Handles the special sequences for | |||
4469 | /// limited-precision mode with x == 10.0f. | |||
4470 | static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS, | |||
4471 | SelectionDAG &DAG, const TargetLowering &TLI) { | |||
4472 | bool IsExp10 = false; | |||
4473 | if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 && | |||
4474 | LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { | |||
4475 | if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) { | |||
4476 | APFloat Ten(10.0f); | |||
4477 | IsExp10 = LHSC->isExactlyValue(Ten); | |||
4478 | } | |||
4479 | } | |||
4480 | ||||
4481 | if (IsExp10) { | |||
4482 | // Put the exponent in the right bit position for later addition to the | |||
4483 | // final result: | |||
4484 | // | |||
4485 | // #define LOG2OF10 3.3219281f | |||
4486 | // IntegerPartOfX = (int32_t)(x * LOG2OF10); | |||
4487 | SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS, | |||
4488 | getF32Constant(DAG, 0x40549a78)); | |||
4489 | SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); | |||
4490 | ||||
4491 | // FractionalPartOfX = x - (float)IntegerPartOfX; | |||
4492 | SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); | |||
4493 | SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); | |||
4494 | ||||
4495 | // IntegerPartOfX <<= 23; | |||
4496 | IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, | |||
4497 | DAG.getConstant(23, TLI.getPointerTy())); | |||
4498 | ||||
4499 | SDValue TwoToFractionalPartOfX; | |||
4500 | if (LimitFloatPrecision <= 6) { | |||
4501 | // For floating-point precision of 6: | |||
4502 | // | |||
4503 | // twoToFractionalPartOfX = | |||
4504 | // 0.997535578f + | |||
4505 | // (0.735607626f + 0.252464424f * x) * x; | |||
4506 | // | |||
4507 | // error 0.0144103317, which is 6 bits | |||
4508 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4509 | getF32Constant(DAG, 0x3e814304)); | |||
4510 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4511 | getF32Constant(DAG, 0x3f3c50c8)); | |||
4512 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4513 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4514 | getF32Constant(DAG, 0x3f7f5e7e)); | |||
4515 | } else if (LimitFloatPrecision <= 12) { | |||
4516 | // For floating-point precision of 12: | |||
4517 | // | |||
4518 | // TwoToFractionalPartOfX = | |||
4519 | // 0.999892986f + | |||
4520 | // (0.696457318f + | |||
4521 | // (0.224338339f + 0.792043434e-1f * x) * x) * x; | |||
4522 | // | |||
4523 | // error 0.000107046256, which is 13 to 14 bits | |||
4524 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4525 | getF32Constant(DAG, 0x3da235e3)); | |||
4526 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4527 | getF32Constant(DAG, 0x3e65b8f3)); | |||
4528 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4529 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4530 | getF32Constant(DAG, 0x3f324b07)); | |||
4531 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4532 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4533 | getF32Constant(DAG, 0x3f7ff8fd)); | |||
4534 | } else { // LimitFloatPrecision <= 18 | |||
4535 | // For floating-point precision of 18: | |||
4536 | // | |||
4537 | // TwoToFractionalPartOfX = | |||
4538 | // 0.999999982f + | |||
4539 | // (0.693148872f + | |||
4540 | // (0.240227044f + | |||
4541 | // (0.554906021e-1f + | |||
4542 | // (0.961591928e-2f + | |||
4543 | // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; | |||
4544 | // error 2.47208000*10^(-7), which is better than 18 bits | |||
4545 | SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, | |||
4546 | getF32Constant(DAG, 0x3924b03e)); | |||
4547 | SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, | |||
4548 | getF32Constant(DAG, 0x3ab24b87)); | |||
4549 | SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); | |||
4550 | SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, | |||
4551 | getF32Constant(DAG, 0x3c1d8c17)); | |||
4552 | SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); | |||
4553 | SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, | |||
4554 | getF32Constant(DAG, 0x3d634a1d)); | |||
4555 | SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); | |||
4556 | SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, | |||
4557 | getF32Constant(DAG, 0x3e75fe14)); | |||
4558 | SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); | |||
4559 | SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, | |||
4560 | getF32Constant(DAG, 0x3f317234)); | |||
4561 | SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); | |||
4562 | TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, | |||
4563 | getF32Constant(DAG, 0x3f800000)); | |||
4564 | } | |||
4565 | ||||
4566 | SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX); | |||
4567 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, | |||
4568 | DAG.getNode(ISD::ADD, dl, MVT::i32, | |||
4569 | t13, IntegerPartOfX)); | |||
4570 | } | |||
4571 | ||||
4572 | // No special expansion. | |||
4573 | return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS); | |||
4574 | } | |||
4575 | ||||
4576 | ||||
4577 | /// ExpandPowI - Expand a llvm.powi intrinsic. | |||
4578 | static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS, | |||
4579 | SelectionDAG &DAG) { | |||
4580 | // If RHS is a constant, we can expand this out to a multiplication tree, | |||
4581 | // otherwise we end up lowering to a call to __powidf2 (for example). When | |||
4582 | // optimizing for size, we only want to do this if the expansion would produce | |||
4583 | // a small number of multiplies, otherwise we do the full expansion. | |||
4584 | if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { | |||
4585 | // Get the exponent as a positive value. | |||
4586 | unsigned Val = RHSC->getSExtValue(); | |||
4587 | if ((int)Val < 0) Val = -Val; | |||
4588 | ||||
4589 | // powi(x, 0) -> 1.0 | |||
4590 | if (Val == 0) | |||
4591 | return DAG.getConstantFP(1.0, LHS.getValueType()); | |||
4592 | ||||
4593 | const Function *F = DAG.getMachineFunction().getFunction(); | |||
4594 | if (!F->hasFnAttribute(Attribute::OptimizeForSize) || | |||
4595 | // If optimizing for size, don't insert too many multiplies. This | |||
4596 | // inserts up to 5 multiplies. | |||
4597 | countPopulation(Val) + Log2_32(Val) < 7) { | |||
4598 | // We use the simple binary decomposition method to generate the multiply | |||
4599 | // sequence. There are more optimal ways to do this (for example, | |||
4600 | // powi(x,15) generates one more multiply than it should), but this has | |||
4601 | // the benefit of being both really simple and much better than a libcall. | |||
4602 | SDValue Res; // Logically starts equal to 1.0 | |||
4603 | SDValue CurSquare = LHS; | |||
4604 | while (Val) { | |||
4605 | if (Val & 1) { | |||
4606 | if (Res.getNode()) | |||
4607 | Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare); | |||
4608 | else | |||
4609 | Res = CurSquare; // 1.0*CurSquare. | |||
4610 | } | |||
4611 | ||||
4612 | CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(), | |||
4613 | CurSquare, CurSquare); | |||
4614 | Val >>= 1; | |||
4615 | } | |||
4616 | ||||
4617 | // If the original was negative, invert the result, producing 1/(x*x*x). | |||
4618 | if (RHSC->getSExtValue() < 0) | |||
4619 | Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(), | |||
4620 | DAG.getConstantFP(1.0, LHS.getValueType()), Res); | |||
4621 | return Res; | |||
4622 | } | |||
4623 | } | |||
4624 | ||||
4625 | // Otherwise, expand to a libcall. | |||
4626 | return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS); | |||
4627 | } | |||
4628 | ||||
4629 | // getTruncatedArgReg - Find underlying register used for an truncated | |||
4630 | // argument. | |||
4631 | static unsigned getTruncatedArgReg(const SDValue &N) { | |||
4632 | if (N.getOpcode() != ISD::TRUNCATE) | |||
4633 | return 0; | |||
4634 | ||||
4635 | const SDValue &Ext = N.getOperand(0); | |||
4636 | if (Ext.getOpcode() == ISD::AssertZext || | |||
4637 | Ext.getOpcode() == ISD::AssertSext) { | |||
4638 | const SDValue &CFR = Ext.getOperand(0); | |||
4639 | if (CFR.getOpcode() == ISD::CopyFromReg) | |||
4640 | return cast<RegisterSDNode>(CFR.getOperand(1))->getReg(); | |||
4641 | if (CFR.getOpcode() == ISD::TRUNCATE) | |||
4642 | return getTruncatedArgReg(CFR); | |||
4643 | } | |||
4644 | return 0; | |||
4645 | } | |||
4646 | ||||
4647 | /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function | |||
4648 | /// argument, create the corresponding DBG_VALUE machine instruction for it now. | |||
4649 | /// At the end of instruction selection, they will be inserted to the entry BB. | |||
4650 | bool SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, | |||
4651 | MDNode *Variable, | |||
4652 | MDNode *Expr, int64_t Offset, | |||
4653 | bool IsIndirect, | |||
4654 | const SDValue &N) { | |||
4655 | const Argument *Arg = dyn_cast<Argument>(V); | |||
4656 | if (!Arg) | |||
4657 | return false; | |||
4658 | ||||
4659 | MachineFunction &MF = DAG.getMachineFunction(); | |||
4660 | const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo(); | |||
4661 | ||||
4662 | // Ignore inlined function arguments here. | |||
4663 | DIVariable DV(Variable); | |||
4664 | if (DV.isInlinedFnArgument(MF.getFunction())) | |||
4665 | return false; | |||
4666 | ||||
4667 | Optional<MachineOperand> Op; | |||
4668 | // Some arguments' frame index is recorded during argument lowering. | |||
4669 | if (int FI = FuncInfo.getArgumentFrameIndex(Arg)) | |||
4670 | Op = MachineOperand::CreateFI(FI); | |||
4671 | ||||
4672 | if (!Op && N.getNode()) { | |||
4673 | unsigned Reg; | |||
4674 | if (N.getOpcode() == ISD::CopyFromReg) | |||
4675 | Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); | |||
4676 | else | |||
4677 | Reg = getTruncatedArgReg(N); | |||
4678 | if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { | |||
4679 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); | |||
4680 | unsigned PR = RegInfo.getLiveInPhysReg(Reg); | |||
4681 | if (PR) | |||
4682 | Reg = PR; | |||
4683 | } | |||
4684 | if (Reg) | |||
4685 | Op = MachineOperand::CreateReg(Reg, false); | |||
4686 | } | |||
4687 | ||||
4688 | if (!Op) { | |||
4689 | // Check if ValueMap has reg number. | |||
4690 | DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); | |||
4691 | if (VMI != FuncInfo.ValueMap.end()) | |||
4692 | Op = MachineOperand::CreateReg(VMI->second, false); | |||
4693 | } | |||
4694 | ||||
4695 | if (!Op && N.getNode()) | |||
4696 | // Check if frame index is available. | |||
4697 | if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode())) | |||
4698 | if (FrameIndexSDNode *FINode = | |||
4699 | dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) | |||
4700 | Op = MachineOperand::CreateFI(FINode->getIndex()); | |||
4701 | ||||
4702 | if (!Op) | |||
4703 | return false; | |||
4704 | ||||
4705 | if (Op->isReg()) | |||
4706 | FuncInfo.ArgDbgValues.push_back( | |||
4707 | BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE), | |||
4708 | IsIndirect, Op->getReg(), Offset, Variable, Expr)); | |||
4709 | else | |||
4710 | FuncInfo.ArgDbgValues.push_back( | |||
4711 | BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE)) | |||
4712 | .addOperand(*Op) | |||
4713 | .addImm(Offset) | |||
4714 | .addMetadata(Variable) | |||
4715 | .addMetadata(Expr)); | |||
4716 | ||||
4717 | return true; | |||
4718 | } | |||
4719 | ||||
4720 | // VisualStudio defines setjmp as _setjmp | |||
4721 | #if defined(_MSC_VER) && defined(setjmp) && \ | |||
4722 | !defined(setjmp_undefined_for_msvc) | |||
4723 | # pragma push_macro("setjmp") | |||
4724 | # undef setjmp | |||
4725 | # define setjmp_undefined_for_msvc | |||
4726 | #endif | |||
4727 | ||||
4728 | /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If | |||
4729 | /// we want to emit this as a call to a named external function, return the name | |||
4730 | /// otherwise lower it and return null. | |||
4731 | const char * | |||
4732 | SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { | |||
4733 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
4734 | SDLoc sdl = getCurSDLoc(); | |||
4735 | DebugLoc dl = getCurDebugLoc(); | |||
4736 | SDValue Res; | |||
4737 | ||||
4738 | switch (Intrinsic) { | |||
4739 | default: | |||
4740 | // By default, turn this into a target intrinsic node. | |||
4741 | visitTargetIntrinsic(I, Intrinsic); | |||
4742 | return nullptr; | |||
4743 | case Intrinsic::vastart: visitVAStart(I); return nullptr; | |||
4744 | case Intrinsic::vaend: visitVAEnd(I); return nullptr; | |||
4745 | case Intrinsic::vacopy: visitVACopy(I); return nullptr; | |||
4746 | case Intrinsic::returnaddress: | |||
4747 | setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI.getPointerTy(), | |||
4748 | getValue(I.getArgOperand(0)))); | |||
4749 | return nullptr; | |||
4750 | case Intrinsic::frameaddress: | |||
4751 | setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI.getPointerTy(), | |||
4752 | getValue(I.getArgOperand(0)))); | |||
4753 | return nullptr; | |||
4754 | case Intrinsic::read_register: { | |||
4755 | Value *Reg = I.getArgOperand(0); | |||
4756 | SDValue RegName = | |||
4757 | DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata())); | |||
4758 | EVT VT = TLI.getValueType(I.getType()); | |||
4759 | setValue(&I, DAG.getNode(ISD::READ_REGISTER, sdl, VT, RegName)); | |||
4760 | return nullptr; | |||
4761 | } | |||
4762 | case Intrinsic::write_register: { | |||
4763 | Value *Reg = I.getArgOperand(0); | |||
4764 | Value *RegValue = I.getArgOperand(1); | |||
4765 | SDValue Chain = getValue(RegValue).getOperand(0); | |||
4766 | SDValue RegName = | |||
4767 | DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata())); | |||
4768 | DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain, | |||
4769 | RegName, getValue(RegValue))); | |||
4770 | return nullptr; | |||
4771 | } | |||
4772 | case Intrinsic::setjmp: | |||
4773 | return &"_setjmp"[!TLI.usesUnderscoreSetJmp()]; | |||
4774 | case Intrinsic::longjmp: | |||
4775 | return &"_longjmp"[!TLI.usesUnderscoreLongJmp()]; | |||
4776 | case Intrinsic::memcpy: { | |||
4777 | // FIXME: this definition of "user defined address space" is x86-specific | |||
4778 | // Assert for address < 256 since we support only user defined address | |||
4779 | // spaces. | |||
4780 | assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4784, __PRETTY_FUNCTION__)) | |||
4781 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4784, __PRETTY_FUNCTION__)) | |||
4782 | cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4784, __PRETTY_FUNCTION__)) | |||
4783 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4784, __PRETTY_FUNCTION__)) | |||
4784 | "Unknown address space")((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4784, __PRETTY_FUNCTION__)); | |||
4785 | SDValue Op1 = getValue(I.getArgOperand(0)); | |||
4786 | SDValue Op2 = getValue(I.getArgOperand(1)); | |||
4787 | SDValue Op3 = getValue(I.getArgOperand(2)); | |||
4788 | unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); | |||
4789 | if (!Align) | |||
4790 | Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment. | |||
4791 | bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); | |||
4792 | DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false, | |||
4793 | MachinePointerInfo(I.getArgOperand(0)), | |||
4794 | MachinePointerInfo(I.getArgOperand(1)))); | |||
4795 | return nullptr; | |||
4796 | } | |||
4797 | case Intrinsic::memset: { | |||
4798 | // FIXME: this definition of "user defined address space" is x86-specific | |||
4799 | // Assert for address < 256 since we support only user defined address | |||
4800 | // spaces. | |||
4801 | assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && "Unknown address space" ) ? static_cast<void> (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4803, __PRETTY_FUNCTION__)) | |||
4802 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && "Unknown address space" ) ? static_cast<void> (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4803, __PRETTY_FUNCTION__)) | |||
4803 | "Unknown address space")((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && "Unknown address space" ) ? static_cast<void> (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4803, __PRETTY_FUNCTION__)); | |||
4804 | SDValue Op1 = getValue(I.getArgOperand(0)); | |||
4805 | SDValue Op2 = getValue(I.getArgOperand(1)); | |||
4806 | SDValue Op3 = getValue(I.getArgOperand(2)); | |||
4807 | unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); | |||
4808 | if (!Align) | |||
4809 | Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment. | |||
4810 | bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); | |||
4811 | DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, | |||
4812 | MachinePointerInfo(I.getArgOperand(0)))); | |||
4813 | return nullptr; | |||
4814 | } | |||
4815 | case Intrinsic::memmove: { | |||
4816 | // FIXME: this definition of "user defined address space" is x86-specific | |||
4817 | // Assert for address < 256 since we support only user defined address | |||
4818 | // spaces. | |||
4819 | assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4823, __PRETTY_FUNCTION__)) | |||
4820 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4823, __PRETTY_FUNCTION__)) | |||
4821 | cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4823, __PRETTY_FUNCTION__)) | |||
4822 | < 256 &&((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4823, __PRETTY_FUNCTION__)) | |||
4823 | "Unknown address space")((cast<PointerType>(I.getArgOperand(0)->getType())-> getAddressSpace() < 256 && cast<PointerType> (I.getArgOperand(1)->getType())->getAddressSpace() < 256 && "Unknown address space") ? static_cast<void > (0) : __assert_fail ("cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() < 256 && cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() < 256 && \"Unknown address space\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4823, __PRETTY_FUNCTION__)); | |||
4824 | SDValue Op1 = getValue(I.getArgOperand(0)); | |||
4825 | SDValue Op2 = getValue(I.getArgOperand(1)); | |||
4826 | SDValue Op3 = getValue(I.getArgOperand(2)); | |||
4827 | unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); | |||
4828 | if (!Align) | |||
4829 | Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment. | |||
4830 | bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); | |||
4831 | DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, | |||
4832 | MachinePointerInfo(I.getArgOperand(0)), | |||
4833 | MachinePointerInfo(I.getArgOperand(1)))); | |||
4834 | return nullptr; | |||
4835 | } | |||
4836 | case Intrinsic::dbg_declare: { | |||
4837 | const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); | |||
4838 | MDNode *Variable = DI.getVariable(); | |||
4839 | MDNode *Expression = DI.getExpression(); | |||
4840 | const Value *Address = DI.getAddress(); | |||
4841 | DIVariable DIVar(Variable); | |||
4842 | assert((!DIVar || DIVar.isVariable()) &&(((!DIVar || DIVar.isVariable()) && "Variable in DbgDeclareInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgDeclareInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4843, __PRETTY_FUNCTION__)) | |||
4843 | "Variable in DbgDeclareInst should be either null or a DIVariable.")(((!DIVar || DIVar.isVariable()) && "Variable in DbgDeclareInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgDeclareInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4843, __PRETTY_FUNCTION__)); | |||
4844 | if (!Address || !DIVar) { | |||
4845 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); | |||
4846 | return nullptr; | |||
4847 | } | |||
4848 | ||||
4849 | // Check if address has undef value. | |||
4850 | if (isa<UndefValue>(Address) || | |||
4851 | (Address->use_empty() && !isa<Argument>(Address))) { | |||
4852 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); | |||
4853 | return nullptr; | |||
4854 | } | |||
4855 | ||||
4856 | SDValue &N = NodeMap[Address]; | |||
4857 | if (!N.getNode() && isa<Argument>(Address)) | |||
4858 | // Check unused arguments map. | |||
4859 | N = UnusedArgNodeMap[Address]; | |||
4860 | SDDbgValue *SDV; | |||
4861 | if (N.getNode()) { | |||
4862 | if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) | |||
4863 | Address = BCI->getOperand(0); | |||
4864 | // Parameters are handled specially. | |||
4865 | bool isParameter = | |||
4866 | (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable || | |||
4867 | isa<Argument>(Address)); | |||
4868 | ||||
4869 | const AllocaInst *AI = dyn_cast<AllocaInst>(Address); | |||
4870 | ||||
4871 | if (isParameter && !AI) { | |||
4872 | FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); | |||
4873 | if (FINode) | |||
4874 | // Byval parameter. We have a frame index at this point. | |||
4875 | SDV = DAG.getFrameIndexDbgValue( | |||
4876 | Variable, Expression, FINode->getIndex(), 0, dl, SDNodeOrder); | |||
4877 | else { | |||
4878 | // Address is an argument, so try to emit its dbg value using | |||
4879 | // virtual register info from the FuncInfo.ValueMap. | |||
4880 | EmitFuncArgumentDbgValue(Address, Variable, Expression, 0, false, N); | |||
4881 | return nullptr; | |||
4882 | } | |||
4883 | } else if (AI) | |||
4884 | SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(), | |||
4885 | true, 0, dl, SDNodeOrder); | |||
4886 | else { | |||
4887 | // Can't do anything with other non-AI cases yet. | |||
4888 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); | |||
4889 | DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "non-AllocaInst issue for Address: \n\t" ; } } while (0); | |||
4890 | DEBUG(Address->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { Address->dump(); } } while (0); | |||
4891 | return nullptr; | |||
4892 | } | |||
4893 | DAG.AddDbgValue(SDV, N.getNode(), isParameter); | |||
4894 | } else { | |||
4895 | // If Address is an argument then try to emit its dbg value using | |||
4896 | // virtual register info from the FuncInfo.ValueMap. | |||
4897 | if (!EmitFuncArgumentDbgValue(Address, Variable, Expression, 0, false, | |||
4898 | N)) { | |||
4899 | // If variable is pinned by a alloca in dominating bb then | |||
4900 | // use StaticAllocaMap. | |||
4901 | if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) { | |||
4902 | if (AI->getParent() != DI.getParent()) { | |||
4903 | DenseMap<const AllocaInst*, int>::iterator SI = | |||
4904 | FuncInfo.StaticAllocaMap.find(AI); | |||
4905 | if (SI != FuncInfo.StaticAllocaMap.end()) { | |||
4906 | SDV = DAG.getFrameIndexDbgValue(Variable, Expression, SI->second, | |||
4907 | 0, dl, SDNodeOrder); | |||
4908 | DAG.AddDbgValue(SDV, nullptr, false); | |||
4909 | return nullptr; | |||
4910 | } | |||
4911 | } | |||
4912 | } | |||
4913 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); | |||
4914 | } | |||
4915 | } | |||
4916 | return nullptr; | |||
4917 | } | |||
4918 | case Intrinsic::dbg_value: { | |||
4919 | const DbgValueInst &DI = cast<DbgValueInst>(I); | |||
4920 | DIVariable DIVar(DI.getVariable()); | |||
4921 | assert((!DIVar || DIVar.isVariable()) &&(((!DIVar || DIVar.isVariable()) && "Variable in DbgValueInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgValueInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4922, __PRETTY_FUNCTION__)) | |||
4922 | "Variable in DbgValueInst should be either null or a DIVariable.")(((!DIVar || DIVar.isVariable()) && "Variable in DbgValueInst should be either null or a DIVariable." ) ? static_cast<void> (0) : __assert_fail ("(!DIVar || DIVar.isVariable()) && \"Variable in DbgValueInst should be either null or a DIVariable.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 4922, __PRETTY_FUNCTION__)); | |||
4923 | if (!DIVar) | |||
4924 | return nullptr; | |||
4925 | ||||
4926 | MDNode *Variable = DI.getVariable(); | |||
4927 | MDNode *Expression = DI.getExpression(); | |||
4928 | uint64_t Offset = DI.getOffset(); | |||
4929 | const Value *V = DI.getValue(); | |||
4930 | if (!V) | |||
4931 | return nullptr; | |||
4932 | ||||
4933 | SDDbgValue *SDV; | |||
4934 | if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) { | |||
4935 | SDV = DAG.getConstantDbgValue(Variable, Expression, V, Offset, dl, | |||
4936 | SDNodeOrder); | |||
4937 | DAG.AddDbgValue(SDV, nullptr, false); | |||
4938 | } else { | |||
4939 | // Do not use getValue() in here; we don't want to generate code at | |||
4940 | // this point if it hasn't been done yet. | |||
4941 | SDValue N = NodeMap[V]; | |||
4942 | if (!N.getNode() && isa<Argument>(V)) | |||
4943 | // Check unused arguments map. | |||
4944 | N = UnusedArgNodeMap[V]; | |||
4945 | if (N.getNode()) { | |||
4946 | // A dbg.value for an alloca is always indirect. | |||
4947 | bool IsIndirect = isa<AllocaInst>(V) || Offset != 0; | |||
4948 | if (!EmitFuncArgumentDbgValue(V, Variable, Expression, Offset, | |||
4949 | IsIndirect, N)) { | |||
4950 | SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(), | |||
4951 | IsIndirect, Offset, dl, SDNodeOrder); | |||
4952 | DAG.AddDbgValue(SDV, N.getNode(), false); | |||
4953 | } | |||
4954 | } else if (!V->use_empty() ) { | |||
4955 | // Do not call getValue(V) yet, as we don't want to generate code. | |||
4956 | // Remember it for later. | |||
4957 | DanglingDebugInfo DDI(&DI, dl, SDNodeOrder); | |||
4958 | DanglingDebugInfoMap[V] = DDI; | |||
4959 | } else { | |||
4960 | // We may expand this to cover more cases. One case where we have no | |||
4961 | // data available is an unreferenced parameter. | |||
4962 | DEBUG(dbgs() << "Dropping debug info for " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug info for " << DI << "\n"; } } while (0); | |||
4963 | } | |||
4964 | } | |||
4965 | ||||
4966 | // Build a debug info table entry. | |||
4967 | if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V)) | |||
4968 | V = BCI->getOperand(0); | |||
4969 | const AllocaInst *AI = dyn_cast<AllocaInst>(V); | |||
4970 | // Don't handle byval struct arguments or VLAs, for example. | |||
4971 | if (!AI) { | |||
4972 | DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << "Dropping debug location info for:\n " << DI << "\n"; } } while (0); | |||
4973 | DEBUG(dbgs() << " Last seen at:\n " << *V << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { dbgs() << " Last seen at:\n " << * V << "\n"; } } while (0); | |||
4974 | return nullptr; | |||
4975 | } | |||
4976 | DenseMap<const AllocaInst*, int>::iterator SI = | |||
4977 | FuncInfo.StaticAllocaMap.find(AI); | |||
4978 | if (SI == FuncInfo.StaticAllocaMap.end()) | |||
4979 | return nullptr; // VLAs. | |||
4980 | return nullptr; | |||
4981 | } | |||
4982 | ||||
4983 | case Intrinsic::eh_typeid_for: { | |||
4984 | // Find the type id for the given typeinfo. | |||
4985 | GlobalValue *GV = ExtractTypeInfo(I.getArgOperand(0)); | |||
4986 | unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV); | |||
4987 | Res = DAG.getConstant(TypeID, MVT::i32); | |||
4988 | setValue(&I, Res); | |||
4989 | return nullptr; | |||
4990 | } | |||
4991 | ||||
4992 | case Intrinsic::eh_return_i32: | |||
4993 | case Intrinsic::eh_return_i64: | |||
4994 | DAG.getMachineFunction().getMMI().setCallsEHReturn(true); | |||
4995 | DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl, | |||
4996 | MVT::Other, | |||
4997 | getControlRoot(), | |||
4998 | getValue(I.getArgOperand(0)), | |||
4999 | getValue(I.getArgOperand(1)))); | |||
5000 | return nullptr; | |||
5001 | case Intrinsic::eh_unwind_init: | |||
5002 | DAG.getMachineFunction().getMMI().setCallsUnwindInit(true); | |||
5003 | return nullptr; | |||
5004 | case Intrinsic::eh_dwarf_cfa: { | |||
5005 | SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl, | |||
5006 | TLI.getPointerTy()); | |||
5007 | SDValue Offset = DAG.getNode(ISD::ADD, sdl, | |||
5008 | CfaArg.getValueType(), | |||
5009 | DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl, | |||
5010 | CfaArg.getValueType()), | |||
5011 | CfaArg); | |||
5012 | SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl, TLI.getPointerTy(), | |||
5013 | DAG.getConstant(0, TLI.getPointerTy())); | |||
5014 | setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(), | |||
5015 | FA, Offset)); | |||
5016 | return nullptr; | |||
5017 | } | |||
5018 | case Intrinsic::eh_sjlj_callsite: { | |||
5019 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); | |||
5020 | ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0)); | |||
5021 | assert(CI && "Non-constant call site value in eh.sjlj.callsite!")((CI && "Non-constant call site value in eh.sjlj.callsite!" ) ? static_cast<void> (0) : __assert_fail ("CI && \"Non-constant call site value in eh.sjlj.callsite!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5021, __PRETTY_FUNCTION__)); | |||
5022 | assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!")((MMI.getCurrentCallSite() == 0 && "Overlapping call sites!" ) ? static_cast<void> (0) : __assert_fail ("MMI.getCurrentCallSite() == 0 && \"Overlapping call sites!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5022, __PRETTY_FUNCTION__)); | |||
5023 | ||||
5024 | MMI.setCurrentCallSite(CI->getZExtValue()); | |||
5025 | return nullptr; | |||
5026 | } | |||
5027 | case Intrinsic::eh_sjlj_functioncontext: { | |||
5028 | // Get and store the index of the function context. | |||
5029 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); | |||
5030 | AllocaInst *FnCtx = | |||
5031 | cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts()); | |||
5032 | int FI = FuncInfo.StaticAllocaMap[FnCtx]; | |||
5033 | MFI->setFunctionContextIndex(FI); | |||
5034 | return nullptr; | |||
5035 | } | |||
5036 | case Intrinsic::eh_sjlj_setjmp: { | |||
5037 | SDValue Ops[2]; | |||
5038 | Ops[0] = getRoot(); | |||
5039 | Ops[1] = getValue(I.getArgOperand(0)); | |||
5040 | SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl, | |||
5041 | DAG.getVTList(MVT::i32, MVT::Other), Ops); | |||
5042 | setValue(&I, Op.getValue(0)); | |||
5043 | DAG.setRoot(Op.getValue(1)); | |||
5044 | return nullptr; | |||
5045 | } | |||
5046 | case Intrinsic::eh_sjlj_longjmp: { | |||
5047 | DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other, | |||
5048 | getRoot(), getValue(I.getArgOperand(0)))); | |||
5049 | return nullptr; | |||
5050 | } | |||
5051 | ||||
5052 | case Intrinsic::masked_load: | |||
5053 | visitMaskedLoad(I); | |||
5054 | return nullptr; | |||
5055 | case Intrinsic::masked_store: | |||
5056 | visitMaskedStore(I); | |||
5057 | return nullptr; | |||
5058 | case Intrinsic::x86_mmx_pslli_w: | |||
5059 | case Intrinsic::x86_mmx_pslli_d: | |||
5060 | case Intrinsic::x86_mmx_pslli_q: | |||
5061 | case Intrinsic::x86_mmx_psrli_w: | |||
5062 | case Intrinsic::x86_mmx_psrli_d: | |||
5063 | case Intrinsic::x86_mmx_psrli_q: | |||
5064 | case Intrinsic::x86_mmx_psrai_w: | |||
5065 | case Intrinsic::x86_mmx_psrai_d: { | |||
5066 | SDValue ShAmt = getValue(I.getArgOperand(1)); | |||
5067 | if (isa<ConstantSDNode>(ShAmt)) { | |||
5068 | visitTargetIntrinsic(I, Intrinsic); | |||
5069 | return nullptr; | |||
5070 | } | |||
5071 | unsigned NewIntrinsic = 0; | |||
5072 | EVT ShAmtVT = MVT::v2i32; | |||
5073 | switch (Intrinsic) { | |||
5074 | case Intrinsic::x86_mmx_pslli_w: | |||
5075 | NewIntrinsic = Intrinsic::x86_mmx_psll_w; | |||
5076 | break; | |||
5077 | case Intrinsic::x86_mmx_pslli_d: | |||
5078 | NewIntrinsic = Intrinsic::x86_mmx_psll_d; | |||
5079 | break; | |||
5080 | case Intrinsic::x86_mmx_pslli_q: | |||
5081 | NewIntrinsic = Intrinsic::x86_mmx_psll_q; | |||
5082 | break; | |||
5083 | case Intrinsic::x86_mmx_psrli_w: | |||
5084 | NewIntrinsic = Intrinsic::x86_mmx_psrl_w; | |||
5085 | break; | |||
5086 | case Intrinsic::x86_mmx_psrli_d: | |||
5087 | NewIntrinsic = Intrinsic::x86_mmx_psrl_d; | |||
5088 | break; | |||
5089 | case Intrinsic::x86_mmx_psrli_q: | |||
5090 | NewIntrinsic = Intrinsic::x86_mmx_psrl_q; | |||
5091 | break; | |||
5092 | case Intrinsic::x86_mmx_psrai_w: | |||
5093 | NewIntrinsic = Intrinsic::x86_mmx_psra_w; | |||
5094 | break; | |||
5095 | case Intrinsic::x86_mmx_psrai_d: | |||
5096 | NewIntrinsic = Intrinsic::x86_mmx_psra_d; | |||
5097 | break; | |||
5098 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5098); // Can't reach here. | |||
5099 | } | |||
5100 | ||||
5101 | // The vector shift intrinsics with scalars uses 32b shift amounts but | |||
5102 | // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits | |||
5103 | // to be zero. | |||
5104 | // We must do this early because v2i32 is not a legal type. | |||
5105 | SDValue ShOps[2]; | |||
5106 | ShOps[0] = ShAmt; | |||
5107 | ShOps[1] = DAG.getConstant(0, MVT::i32); | |||
5108 | ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps); | |||
5109 | EVT DestVT = TLI.getValueType(I.getType()); | |||
5110 | ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt); | |||
5111 | Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT, | |||
5112 | DAG.getConstant(NewIntrinsic, MVT::i32), | |||
5113 | getValue(I.getArgOperand(0)), ShAmt); | |||
5114 | setValue(&I, Res); | |||
5115 | return nullptr; | |||
5116 | } | |||
5117 | case Intrinsic::x86_avx_vinsertf128_pd_256: | |||
5118 | case Intrinsic::x86_avx_vinsertf128_ps_256: | |||
5119 | case Intrinsic::x86_avx_vinsertf128_si_256: | |||
5120 | case Intrinsic::x86_avx2_vinserti128: { | |||
5121 | EVT DestVT = TLI.getValueType(I.getType()); | |||
5122 | EVT ElVT = TLI.getValueType(I.getArgOperand(1)->getType()); | |||
5123 | uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) * | |||
5124 | ElVT.getVectorNumElements(); | |||
5125 | Res = | |||
5126 | DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT, | |||
5127 | getValue(I.getArgOperand(0)), getValue(I.getArgOperand(1)), | |||
5128 | DAG.getConstant(Idx, TLI.getVectorIdxTy())); | |||
5129 | setValue(&I, Res); | |||
5130 | return nullptr; | |||
5131 | } | |||
5132 | case Intrinsic::x86_avx_vextractf128_pd_256: | |||
5133 | case Intrinsic::x86_avx_vextractf128_ps_256: | |||
5134 | case Intrinsic::x86_avx_vextractf128_si_256: | |||
5135 | case Intrinsic::x86_avx2_vextracti128: { | |||
5136 | EVT DestVT = TLI.getValueType(I.getType()); | |||
5137 | uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) * | |||
5138 | DestVT.getVectorNumElements(); | |||
5139 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT, | |||
5140 | getValue(I.getArgOperand(0)), | |||
5141 | DAG.getConstant(Idx, TLI.getVectorIdxTy())); | |||
5142 | setValue(&I, Res); | |||
5143 | return nullptr; | |||
5144 | } | |||
5145 | case Intrinsic::convertff: | |||
5146 | case Intrinsic::convertfsi: | |||
5147 | case Intrinsic::convertfui: | |||
5148 | case Intrinsic::convertsif: | |||
5149 | case Intrinsic::convertuif: | |||
5150 | case Intrinsic::convertss: | |||
5151 | case Intrinsic::convertsu: | |||
5152 | case Intrinsic::convertus: | |||
5153 | case Intrinsic::convertuu: { | |||
5154 | ISD::CvtCode Code = ISD::CVT_INVALID; | |||
5155 | switch (Intrinsic) { | |||
5156 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5156); // Can't reach here. | |||
5157 | case Intrinsic::convertff: Code = ISD::CVT_FF; break; | |||
5158 | case Intrinsic::convertfsi: Code = ISD::CVT_FS; break; | |||
5159 | case Intrinsic::convertfui: Code = ISD::CVT_FU; break; | |||
5160 | case Intrinsic::convertsif: Code = ISD::CVT_SF; break; | |||
5161 | case Intrinsic::convertuif: Code = ISD::CVT_UF; break; | |||
5162 | case Intrinsic::convertss: Code = ISD::CVT_SS; break; | |||
5163 | case Intrinsic::convertsu: Code = ISD::CVT_SU; break; | |||
5164 | case Intrinsic::convertus: Code = ISD::CVT_US; break; | |||
5165 | case Intrinsic::convertuu: Code = ISD::CVT_UU; break; | |||
5166 | } | |||
5167 | EVT DestVT = TLI.getValueType(I.getType()); | |||
5168 | const Value *Op1 = I.getArgOperand(0); | |||
5169 | Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1), | |||
5170 | DAG.getValueType(DestVT), | |||
5171 | DAG.getValueType(getValue(Op1).getValueType()), | |||
5172 | getValue(I.getArgOperand(1)), | |||
5173 | getValue(I.getArgOperand(2)), | |||
5174 | Code); | |||
5175 | setValue(&I, Res); | |||
5176 | return nullptr; | |||
5177 | } | |||
5178 | case Intrinsic::powi: | |||
5179 | setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)), | |||
5180 | getValue(I.getArgOperand(1)), DAG)); | |||
5181 | return nullptr; | |||
5182 | case Intrinsic::log: | |||
5183 | setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); | |||
5184 | return nullptr; | |||
5185 | case Intrinsic::log2: | |||
5186 | setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); | |||
5187 | return nullptr; | |||
5188 | case Intrinsic::log10: | |||
5189 | setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); | |||
5190 | return nullptr; | |||
5191 | case Intrinsic::exp: | |||
5192 | setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); | |||
5193 | return nullptr; | |||
5194 | case Intrinsic::exp2: | |||
5195 | setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, TLI)); | |||
5196 | return nullptr; | |||
5197 | case Intrinsic::pow: | |||
5198 | setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)), | |||
5199 | getValue(I.getArgOperand(1)), DAG, TLI)); | |||
5200 | return nullptr; | |||
5201 | case Intrinsic::sqrt: | |||
5202 | case Intrinsic::fabs: | |||
5203 | case Intrinsic::sin: | |||
5204 | case Intrinsic::cos: | |||
5205 | case Intrinsic::floor: | |||
5206 | case Intrinsic::ceil: | |||
5207 | case Intrinsic::trunc: | |||
5208 | case Intrinsic::rint: | |||
5209 | case Intrinsic::nearbyint: | |||
5210 | case Intrinsic::round: { | |||
5211 | unsigned Opcode; | |||
5212 | switch (Intrinsic) { | |||
5213 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5213); // Can't reach here. | |||
5214 | case Intrinsic::sqrt: Opcode = ISD::FSQRT; break; | |||
5215 | case Intrinsic::fabs: Opcode = ISD::FABS; break; | |||
5216 | case Intrinsic::sin: Opcode = ISD::FSIN; break; | |||
5217 | case Intrinsic::cos: Opcode = ISD::FCOS; break; | |||
5218 | case Intrinsic::floor: Opcode = ISD::FFLOOR; break; | |||
5219 | case Intrinsic::ceil: Opcode = ISD::FCEIL; break; | |||
5220 | case Intrinsic::trunc: Opcode = ISD::FTRUNC; break; | |||
5221 | case Intrinsic::rint: Opcode = ISD::FRINT; break; | |||
5222 | case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break; | |||
5223 | case Intrinsic::round: Opcode = ISD::FROUND; break; | |||
5224 | } | |||
5225 | ||||
5226 | setValue(&I, DAG.getNode(Opcode, sdl, | |||
5227 | getValue(I.getArgOperand(0)).getValueType(), | |||
5228 | getValue(I.getArgOperand(0)))); | |||
5229 | return nullptr; | |||
5230 | } | |||
5231 | case Intrinsic::minnum: | |||
5232 | setValue(&I, DAG.getNode(ISD::FMINNUM, sdl, | |||
5233 | getValue(I.getArgOperand(0)).getValueType(), | |||
5234 | getValue(I.getArgOperand(0)), | |||
5235 | getValue(I.getArgOperand(1)))); | |||
5236 | return nullptr; | |||
5237 | case Intrinsic::maxnum: | |||
5238 | setValue(&I, DAG.getNode(ISD::FMAXNUM, sdl, | |||
5239 | getValue(I.getArgOperand(0)).getValueType(), | |||
5240 | getValue(I.getArgOperand(0)), | |||
5241 | getValue(I.getArgOperand(1)))); | |||
5242 | return nullptr; | |||
5243 | case Intrinsic::copysign: | |||
5244 | setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl, | |||
5245 | getValue(I.getArgOperand(0)).getValueType(), | |||
5246 | getValue(I.getArgOperand(0)), | |||
5247 | getValue(I.getArgOperand(1)))); | |||
5248 | return nullptr; | |||
5249 | case Intrinsic::fma: | |||
5250 | setValue(&I, DAG.getNode(ISD::FMA, sdl, | |||
5251 | getValue(I.getArgOperand(0)).getValueType(), | |||
5252 | getValue(I.getArgOperand(0)), | |||
5253 | getValue(I.getArgOperand(1)), | |||
5254 | getValue(I.getArgOperand(2)))); | |||
5255 | return nullptr; | |||
5256 | case Intrinsic::fmuladd: { | |||
5257 | EVT VT = TLI.getValueType(I.getType()); | |||
5258 | if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict && | |||
5259 | TLI.isFMAFasterThanFMulAndFAdd(VT)) { | |||
5260 | setValue(&I, DAG.getNode(ISD::FMA, sdl, | |||
5261 | getValue(I.getArgOperand(0)).getValueType(), | |||
5262 | getValue(I.getArgOperand(0)), | |||
5263 | getValue(I.getArgOperand(1)), | |||
5264 | getValue(I.getArgOperand(2)))); | |||
5265 | } else { | |||
5266 | SDValue Mul = DAG.getNode(ISD::FMUL, sdl, | |||
5267 | getValue(I.getArgOperand(0)).getValueType(), | |||
5268 | getValue(I.getArgOperand(0)), | |||
5269 | getValue(I.getArgOperand(1))); | |||
5270 | SDValue Add = DAG.getNode(ISD::FADD, sdl, | |||
5271 | getValue(I.getArgOperand(0)).getValueType(), | |||
5272 | Mul, | |||
5273 | getValue(I.getArgOperand(2))); | |||
5274 | setValue(&I, Add); | |||
5275 | } | |||
5276 | return nullptr; | |||
5277 | } | |||
5278 | case Intrinsic::convert_to_fp16: | |||
5279 | setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16, | |||
5280 | DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16, | |||
5281 | getValue(I.getArgOperand(0)), | |||
5282 | DAG.getTargetConstant(0, MVT::i32)))); | |||
5283 | return nullptr; | |||
5284 | case Intrinsic::convert_from_fp16: | |||
5285 | setValue(&I, | |||
5286 | DAG.getNode(ISD::FP_EXTEND, sdl, TLI.getValueType(I.getType()), | |||
5287 | DAG.getNode(ISD::BITCAST, sdl, MVT::f16, | |||
5288 | getValue(I.getArgOperand(0))))); | |||
5289 | return nullptr; | |||
5290 | case Intrinsic::pcmarker: { | |||
5291 | SDValue Tmp = getValue(I.getArgOperand(0)); | |||
5292 | DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp)); | |||
5293 | return nullptr; | |||
5294 | } | |||
5295 | case Intrinsic::readcyclecounter: { | |||
5296 | SDValue Op = getRoot(); | |||
5297 | Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl, | |||
5298 | DAG.getVTList(MVT::i64, MVT::Other), Op); | |||
5299 | setValue(&I, Res); | |||
5300 | DAG.setRoot(Res.getValue(1)); | |||
5301 | return nullptr; | |||
5302 | } | |||
5303 | case Intrinsic::bswap: | |||
5304 | setValue(&I, DAG.getNode(ISD::BSWAP, sdl, | |||
5305 | getValue(I.getArgOperand(0)).getValueType(), | |||
5306 | getValue(I.getArgOperand(0)))); | |||
5307 | return nullptr; | |||
5308 | case Intrinsic::cttz: { | |||
5309 | SDValue Arg = getValue(I.getArgOperand(0)); | |||
5310 | ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1)); | |||
5311 | EVT Ty = Arg.getValueType(); | |||
5312 | setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF, | |||
5313 | sdl, Ty, Arg)); | |||
5314 | return nullptr; | |||
5315 | } | |||
5316 | case Intrinsic::ctlz: { | |||
5317 | SDValue Arg = getValue(I.getArgOperand(0)); | |||
5318 | ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1)); | |||
5319 | EVT Ty = Arg.getValueType(); | |||
5320 | setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF, | |||
5321 | sdl, Ty, Arg)); | |||
5322 | return nullptr; | |||
5323 | } | |||
5324 | case Intrinsic::ctpop: { | |||
5325 | SDValue Arg = getValue(I.getArgOperand(0)); | |||
5326 | EVT Ty = Arg.getValueType(); | |||
5327 | setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg)); | |||
5328 | return nullptr; | |||
5329 | } | |||
5330 | case Intrinsic::stacksave: { | |||
5331 | SDValue Op = getRoot(); | |||
5332 | Res = DAG.getNode(ISD::STACKSAVE, sdl, | |||
5333 | DAG.getVTList(TLI.getPointerTy(), MVT::Other), Op); | |||
5334 | setValue(&I, Res); | |||
5335 | DAG.setRoot(Res.getValue(1)); | |||
5336 | return nullptr; | |||
5337 | } | |||
5338 | case Intrinsic::stackrestore: { | |||
5339 | Res = getValue(I.getArgOperand(0)); | |||
5340 | DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res)); | |||
5341 | return nullptr; | |||
5342 | } | |||
5343 | case Intrinsic::stackprotector: { | |||
5344 | // Emit code into the DAG to store the stack guard onto the stack. | |||
5345 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5346 | MachineFrameInfo *MFI = MF.getFrameInfo(); | |||
5347 | EVT PtrTy = TLI.getPointerTy(); | |||
5348 | SDValue Src, Chain = getRoot(); | |||
5349 | const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand(); | |||
5350 | const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr); | |||
5351 | ||||
5352 | // See if Ptr is a bitcast. If it is, look through it and see if we can get | |||
5353 | // global variable __stack_chk_guard. | |||
5354 | if (!GV) | |||
5355 | if (const Operator *BC = dyn_cast<Operator>(Ptr)) | |||
5356 | if (BC->getOpcode() == Instruction::BitCast) | |||
5357 | GV = dyn_cast<GlobalVariable>(BC->getOperand(0)); | |||
5358 | ||||
5359 | if (GV && TLI.useLoadStackGuardNode()) { | |||
5360 | // Emit a LOAD_STACK_GUARD node. | |||
5361 | MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD, | |||
5362 | sdl, PtrTy, Chain); | |||
5363 | MachinePointerInfo MPInfo(GV); | |||
5364 | MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1); | |||
5365 | unsigned Flags = MachineMemOperand::MOLoad | | |||
5366 | MachineMemOperand::MOInvariant; | |||
5367 | *MemRefs = MF.getMachineMemOperand(MPInfo, Flags, | |||
5368 | PtrTy.getSizeInBits() / 8, | |||
5369 | DAG.getEVTAlignment(PtrTy)); | |||
5370 | Node->setMemRefs(MemRefs, MemRefs + 1); | |||
5371 | ||||
5372 | // Copy the guard value to a virtual register so that it can be | |||
5373 | // retrieved in the epilogue. | |||
5374 | Src = SDValue(Node, 0); | |||
5375 | const TargetRegisterClass *RC = | |||
5376 | TLI.getRegClassFor(Src.getSimpleValueType()); | |||
5377 | unsigned Reg = MF.getRegInfo().createVirtualRegister(RC); | |||
5378 | ||||
5379 | SPDescriptor.setGuardReg(Reg); | |||
5380 | Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src); | |||
5381 | } else { | |||
5382 | Src = getValue(I.getArgOperand(0)); // The guard's value. | |||
5383 | } | |||
5384 | ||||
5385 | AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1)); | |||
5386 | ||||
5387 | int FI = FuncInfo.StaticAllocaMap[Slot]; | |||
5388 | MFI->setStackProtectorIndex(FI); | |||
5389 | ||||
5390 | SDValue FIN = DAG.getFrameIndex(FI, PtrTy); | |||
5391 | ||||
5392 | // Store the stack protector onto the stack. | |||
5393 | Res = DAG.getStore(Chain, sdl, Src, FIN, | |||
5394 | MachinePointerInfo::getFixedStack(FI), | |||
5395 | true, false, 0); | |||
5396 | setValue(&I, Res); | |||
5397 | DAG.setRoot(Res); | |||
5398 | return nullptr; | |||
5399 | } | |||
5400 | case Intrinsic::objectsize: { | |||
5401 | // If we don't know by now, we're never going to know. | |||
5402 | ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1)); | |||
5403 | ||||
5404 | assert(CI && "Non-constant type in __builtin_object_size?")((CI && "Non-constant type in __builtin_object_size?" ) ? static_cast<void> (0) : __assert_fail ("CI && \"Non-constant type in __builtin_object_size?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5404, __PRETTY_FUNCTION__)); | |||
5405 | ||||
5406 | SDValue Arg = getValue(I.getCalledValue()); | |||
5407 | EVT Ty = Arg.getValueType(); | |||
5408 | ||||
5409 | if (CI->isZero()) | |||
5410 | Res = DAG.getConstant(-1ULL, Ty); | |||
5411 | else | |||
5412 | Res = DAG.getConstant(0, Ty); | |||
5413 | ||||
5414 | setValue(&I, Res); | |||
5415 | return nullptr; | |||
5416 | } | |||
5417 | case Intrinsic::annotation: | |||
5418 | case Intrinsic::ptr_annotation: | |||
5419 | // Drop the intrinsic, but forward the value | |||
5420 | setValue(&I, getValue(I.getOperand(0))); | |||
5421 | return nullptr; | |||
5422 | case Intrinsic::assume: | |||
5423 | case Intrinsic::var_annotation: | |||
5424 | // Discard annotate attributes and assumptions | |||
5425 | return nullptr; | |||
5426 | ||||
5427 | case Intrinsic::init_trampoline: { | |||
5428 | const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts()); | |||
5429 | ||||
5430 | SDValue Ops[6]; | |||
5431 | Ops[0] = getRoot(); | |||
5432 | Ops[1] = getValue(I.getArgOperand(0)); | |||
5433 | Ops[2] = getValue(I.getArgOperand(1)); | |||
5434 | Ops[3] = getValue(I.getArgOperand(2)); | |||
5435 | Ops[4] = DAG.getSrcValue(I.getArgOperand(0)); | |||
5436 | Ops[5] = DAG.getSrcValue(F); | |||
5437 | ||||
5438 | Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops); | |||
5439 | ||||
5440 | DAG.setRoot(Res); | |||
5441 | return nullptr; | |||
5442 | } | |||
5443 | case Intrinsic::adjust_trampoline: { | |||
5444 | setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl, | |||
5445 | TLI.getPointerTy(), | |||
5446 | getValue(I.getArgOperand(0)))); | |||
5447 | return nullptr; | |||
5448 | } | |||
5449 | case Intrinsic::gcroot: | |||
5450 | if (GFI) { | |||
5451 | const Value *Alloca = I.getArgOperand(0)->stripPointerCasts(); | |||
5452 | const Constant *TypeMap = cast<Constant>(I.getArgOperand(1)); | |||
5453 | ||||
5454 | FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode()); | |||
5455 | GFI->addStackRoot(FI->getIndex(), TypeMap); | |||
5456 | } | |||
5457 | return nullptr; | |||
5458 | case Intrinsic::gcread: | |||
5459 | case Intrinsic::gcwrite: | |||
5460 | llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!")::llvm::llvm_unreachable_internal("GC failed to lower gcread/gcwrite intrinsics!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5460); | |||
5461 | case Intrinsic::flt_rounds: | |||
5462 | setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32)); | |||
5463 | return nullptr; | |||
5464 | ||||
5465 | case Intrinsic::expect: { | |||
5466 | // Just replace __builtin_expect(exp, c) with EXP. | |||
5467 | setValue(&I, getValue(I.getArgOperand(0))); | |||
5468 | return nullptr; | |||
5469 | } | |||
5470 | ||||
5471 | case Intrinsic::debugtrap: | |||
5472 | case Intrinsic::trap: { | |||
5473 | StringRef TrapFuncName = TM.Options.getTrapFunctionName(); | |||
5474 | if (TrapFuncName.empty()) { | |||
5475 | ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ? | |||
5476 | ISD::TRAP : ISD::DEBUGTRAP; | |||
5477 | DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot())); | |||
5478 | return nullptr; | |||
5479 | } | |||
5480 | TargetLowering::ArgListTy Args; | |||
5481 | ||||
5482 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
5483 | CLI.setDebugLoc(sdl).setChain(getRoot()) | |||
5484 | .setCallee(CallingConv::C, I.getType(), | |||
5485 | DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()), | |||
5486 | std::move(Args), 0); | |||
5487 | ||||
5488 | std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI); | |||
5489 | DAG.setRoot(Result.second); | |||
5490 | return nullptr; | |||
5491 | } | |||
5492 | ||||
5493 | case Intrinsic::uadd_with_overflow: | |||
5494 | case Intrinsic::sadd_with_overflow: | |||
5495 | case Intrinsic::usub_with_overflow: | |||
5496 | case Intrinsic::ssub_with_overflow: | |||
5497 | case Intrinsic::umul_with_overflow: | |||
5498 | case Intrinsic::smul_with_overflow: { | |||
5499 | ISD::NodeType Op; | |||
5500 | switch (Intrinsic) { | |||
5501 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5501); // Can't reach here. | |||
5502 | case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break; | |||
5503 | case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break; | |||
5504 | case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break; | |||
5505 | case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break; | |||
5506 | case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break; | |||
5507 | case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break; | |||
5508 | } | |||
5509 | SDValue Op1 = getValue(I.getArgOperand(0)); | |||
5510 | SDValue Op2 = getValue(I.getArgOperand(1)); | |||
5511 | ||||
5512 | SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1); | |||
5513 | setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2)); | |||
5514 | return nullptr; | |||
5515 | } | |||
5516 | case Intrinsic::prefetch: { | |||
5517 | SDValue Ops[5]; | |||
5518 | unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue(); | |||
5519 | Ops[0] = getRoot(); | |||
5520 | Ops[1] = getValue(I.getArgOperand(0)); | |||
5521 | Ops[2] = getValue(I.getArgOperand(1)); | |||
5522 | Ops[3] = getValue(I.getArgOperand(2)); | |||
5523 | Ops[4] = getValue(I.getArgOperand(3)); | |||
5524 | DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl, | |||
5525 | DAG.getVTList(MVT::Other), Ops, | |||
5526 | EVT::getIntegerVT(*Context, 8), | |||
5527 | MachinePointerInfo(I.getArgOperand(0)), | |||
5528 | 0, /* align */ | |||
5529 | false, /* volatile */ | |||
5530 | rw==0, /* read */ | |||
5531 | rw==1)); /* write */ | |||
5532 | return nullptr; | |||
5533 | } | |||
5534 | case Intrinsic::lifetime_start: | |||
5535 | case Intrinsic::lifetime_end: { | |||
5536 | bool IsStart = (Intrinsic == Intrinsic::lifetime_start); | |||
5537 | // Stack coloring is not enabled in O0, discard region information. | |||
5538 | if (TM.getOptLevel() == CodeGenOpt::None) | |||
5539 | return nullptr; | |||
5540 | ||||
5541 | SmallVector<Value *, 4> Allocas; | |||
5542 | GetUnderlyingObjects(I.getArgOperand(1), Allocas, DL); | |||
5543 | ||||
5544 | for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(), | |||
5545 | E = Allocas.end(); Object != E; ++Object) { | |||
5546 | AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object); | |||
5547 | ||||
5548 | // Could not find an Alloca. | |||
5549 | if (!LifetimeObject) | |||
5550 | continue; | |||
5551 | ||||
5552 | // First check that the Alloca is static, otherwise it won't have a | |||
5553 | // valid frame index. | |||
5554 | auto SI = FuncInfo.StaticAllocaMap.find(LifetimeObject); | |||
5555 | if (SI == FuncInfo.StaticAllocaMap.end()) | |||
5556 | return nullptr; | |||
5557 | ||||
5558 | int FI = SI->second; | |||
5559 | ||||
5560 | SDValue Ops[2]; | |||
5561 | Ops[0] = getRoot(); | |||
5562 | Ops[1] = DAG.getFrameIndex(FI, TLI.getPointerTy(), true); | |||
5563 | unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END); | |||
5564 | ||||
5565 | Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops); | |||
5566 | DAG.setRoot(Res); | |||
5567 | } | |||
5568 | return nullptr; | |||
5569 | } | |||
5570 | case Intrinsic::invariant_start: | |||
5571 | // Discard region information. | |||
5572 | setValue(&I, DAG.getUNDEF(TLI.getPointerTy())); | |||
5573 | return nullptr; | |||
5574 | case Intrinsic::invariant_end: | |||
5575 | // Discard region information. | |||
5576 | return nullptr; | |||
5577 | case Intrinsic::stackprotectorcheck: { | |||
5578 | // Do not actually emit anything for this basic block. Instead we initialize | |||
5579 | // the stack protector descriptor and export the guard variable so we can | |||
5580 | // access it in FinishBasicBlock. | |||
5581 | const BasicBlock *BB = I.getParent(); | |||
5582 | SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I); | |||
5583 | ExportFromCurrentBlock(SPDescriptor.getGuard()); | |||
5584 | ||||
5585 | // Flush our exports since we are going to process a terminator. | |||
5586 | (void)getControlRoot(); | |||
5587 | return nullptr; | |||
5588 | } | |||
5589 | case Intrinsic::clear_cache: | |||
5590 | return TLI.getClearCacheBuiltinName(); | |||
5591 | case Intrinsic::donothing: | |||
5592 | // ignore | |||
5593 | return nullptr; | |||
5594 | case Intrinsic::experimental_stackmap: { | |||
5595 | visitStackmap(I); | |||
5596 | return nullptr; | |||
5597 | } | |||
5598 | case Intrinsic::experimental_patchpoint_void: | |||
5599 | case Intrinsic::experimental_patchpoint_i64: { | |||
5600 | visitPatchpoint(&I); | |||
5601 | return nullptr; | |||
5602 | } | |||
5603 | case Intrinsic::experimental_gc_statepoint: { | |||
5604 | visitStatepoint(I); | |||
5605 | return nullptr; | |||
5606 | } | |||
5607 | case Intrinsic::experimental_gc_result_int: | |||
5608 | case Intrinsic::experimental_gc_result_float: | |||
5609 | case Intrinsic::experimental_gc_result_ptr: | |||
5610 | case Intrinsic::experimental_gc_result: { | |||
5611 | visitGCResult(I); | |||
5612 | return nullptr; | |||
5613 | } | |||
5614 | case Intrinsic::experimental_gc_relocate: { | |||
5615 | visitGCRelocate(I); | |||
5616 | return nullptr; | |||
5617 | } | |||
5618 | case Intrinsic::instrprof_increment: | |||
5619 | llvm_unreachable("instrprof failed to lower an increment")::llvm::llvm_unreachable_internal("instrprof failed to lower an increment" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5619); | |||
5620 | ||||
5621 | case Intrinsic::frameallocate: { | |||
5622 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5623 | const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo(); | |||
5624 | ||||
5625 | // Do the allocation and map it as a normal value. | |||
5626 | // FIXME: Maybe we should add this to the alloca map so that we don't have | |||
5627 | // to register allocate it? | |||
5628 | uint64_t Size = cast<ConstantInt>(I.getArgOperand(0))->getZExtValue(); | |||
5629 | int Alloc = MF.getFrameInfo()->CreateFrameAllocation(Size); | |||
5630 | MVT PtrVT = TLI.getPointerTy(0); | |||
5631 | SDValue FIVal = DAG.getFrameIndex(Alloc, PtrVT); | |||
5632 | setValue(&I, FIVal); | |||
5633 | ||||
5634 | // Directly emit a FRAME_ALLOC machine instr. Label assignment emission is | |||
5635 | // the same on all targets. | |||
5636 | MCSymbol *FrameAllocSym = | |||
5637 | MF.getMMI().getContext().getOrCreateFrameAllocSymbol(MF.getName()); | |||
5638 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, dl, | |||
5639 | TII->get(TargetOpcode::FRAME_ALLOC)) | |||
5640 | .addSym(FrameAllocSym) | |||
5641 | .addFrameIndex(Alloc); | |||
5642 | ||||
5643 | return nullptr; | |||
5644 | } | |||
5645 | ||||
5646 | case Intrinsic::framerecover: { | |||
5647 | // i8* @llvm.framerecover(i8* %fn, i8* %fp) | |||
5648 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5649 | MVT PtrVT = TLI.getPointerTy(0); | |||
5650 | ||||
5651 | // Get the symbol that defines the frame offset. | |||
5652 | Function *Fn = cast<Function>(I.getArgOperand(0)->stripPointerCasts()); | |||
5653 | MCSymbol *FrameAllocSym = | |||
5654 | MF.getMMI().getContext().getOrCreateFrameAllocSymbol(Fn->getName()); | |||
5655 | ||||
5656 | // Create a TargetExternalSymbol for the label to avoid any target lowering | |||
5657 | // that would make this PC relative. | |||
5658 | StringRef Name = FrameAllocSym->getName(); | |||
5659 | assert(Name.size() == strlen(Name.data()) && "not null terminated")((Name.size() == strlen(Name.data()) && "not null terminated" ) ? static_cast<void> (0) : __assert_fail ("Name.size() == strlen(Name.data()) && \"not null terminated\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5659, __PRETTY_FUNCTION__)); | |||
5660 | SDValue OffsetSym = DAG.getTargetExternalSymbol(Name.data(), PtrVT); | |||
5661 | SDValue OffsetVal = | |||
5662 | DAG.getNode(ISD::FRAME_ALLOC_RECOVER, sdl, PtrVT, OffsetSym); | |||
5663 | ||||
5664 | // Add the offset to the FP. | |||
5665 | Value *FP = I.getArgOperand(1); | |||
5666 | SDValue FPVal = getValue(FP); | |||
5667 | SDValue Add = DAG.getNode(ISD::ADD, sdl, PtrVT, FPVal, OffsetVal); | |||
5668 | setValue(&I, Add); | |||
5669 | ||||
5670 | return nullptr; | |||
5671 | } | |||
5672 | case Intrinsic::eh_begincatch: | |||
5673 | case Intrinsic::eh_endcatch: | |||
5674 | llvm_unreachable("begin/end catch intrinsics not lowered in codegen")::llvm::llvm_unreachable_internal("begin/end catch intrinsics not lowered in codegen" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5674); | |||
5675 | } | |||
5676 | } | |||
5677 | ||||
5678 | std::pair<SDValue, SDValue> | |||
5679 | SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI, | |||
5680 | MachineBasicBlock *LandingPad) { | |||
5681 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); | |||
5682 | MCSymbol *BeginLabel = nullptr; | |||
5683 | ||||
5684 | if (LandingPad) { | |||
5685 | // Insert a label before the invoke call to mark the try range. This can be | |||
5686 | // used to detect deletion of the invoke via the MachineModuleInfo. | |||
5687 | BeginLabel = MMI.getContext().CreateTempSymbol(); | |||
5688 | ||||
5689 | // For SjLj, keep track of which landing pads go with which invokes | |||
5690 | // so as to maintain the ordering of pads in the LSDA. | |||
5691 | unsigned CallSiteIndex = MMI.getCurrentCallSite(); | |||
5692 | if (CallSiteIndex) { | |||
5693 | MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex); | |||
5694 | LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex); | |||
5695 | ||||
5696 | // Now that the call site is handled, stop tracking it. | |||
5697 | MMI.setCurrentCallSite(0); | |||
5698 | } | |||
5699 | ||||
5700 | // Both PendingLoads and PendingExports must be flushed here; | |||
5701 | // this call might not return. | |||
5702 | (void)getRoot(); | |||
5703 | DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel)); | |||
5704 | ||||
5705 | CLI.setChain(getRoot()); | |||
5706 | } | |||
5707 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
5708 | std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI); | |||
5709 | ||||
5710 | assert((CLI.IsTailCall || Result.second.getNode()) &&(((CLI.IsTailCall || Result.second.getNode()) && "Non-null chain expected with non-tail call!" ) ? static_cast<void> (0) : __assert_fail ("(CLI.IsTailCall || Result.second.getNode()) && \"Non-null chain expected with non-tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5711, __PRETTY_FUNCTION__)) | |||
5711 | "Non-null chain expected with non-tail call!")(((CLI.IsTailCall || Result.second.getNode()) && "Non-null chain expected with non-tail call!" ) ? static_cast<void> (0) : __assert_fail ("(CLI.IsTailCall || Result.second.getNode()) && \"Non-null chain expected with non-tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5711, __PRETTY_FUNCTION__)); | |||
5712 | assert((Result.second.getNode() || !Result.first.getNode()) &&(((Result.second.getNode() || !Result.first.getNode()) && "Null value expected with tail call!") ? static_cast<void > (0) : __assert_fail ("(Result.second.getNode() || !Result.first.getNode()) && \"Null value expected with tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5713, __PRETTY_FUNCTION__)) | |||
5713 | "Null value expected with tail call!")(((Result.second.getNode() || !Result.first.getNode()) && "Null value expected with tail call!") ? static_cast<void > (0) : __assert_fail ("(Result.second.getNode() || !Result.first.getNode()) && \"Null value expected with tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 5713, __PRETTY_FUNCTION__)); | |||
5714 | ||||
5715 | if (!Result.second.getNode()) { | |||
5716 | // As a special case, a null chain means that a tail call has been emitted | |||
5717 | // and the DAG root is already updated. | |||
5718 | HasTailCall = true; | |||
5719 | ||||
5720 | // Since there's no actual continuation from this block, nothing can be | |||
5721 | // relying on us setting vregs for them. | |||
5722 | PendingExports.clear(); | |||
5723 | } else { | |||
5724 | DAG.setRoot(Result.second); | |||
5725 | } | |||
5726 | ||||
5727 | if (LandingPad) { | |||
5728 | // Insert a label at the end of the invoke call to mark the try range. This | |||
5729 | // can be used to detect deletion of the invoke via the MachineModuleInfo. | |||
5730 | MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol(); | |||
5731 | DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel)); | |||
5732 | ||||
5733 | // Inform MachineModuleInfo of range. | |||
5734 | MMI.addInvoke(LandingPad, BeginLabel, EndLabel); | |||
5735 | } | |||
5736 | ||||
5737 | return Result; | |||
5738 | } | |||
5739 | ||||
5740 | void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, | |||
5741 | bool isTailCall, | |||
5742 | MachineBasicBlock *LandingPad) { | |||
5743 | PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); | |||
5744 | FunctionType *FTy = cast<FunctionType>(PT->getElementType()); | |||
5745 | Type *RetTy = FTy->getReturnType(); | |||
5746 | ||||
5747 | TargetLowering::ArgListTy Args; | |||
5748 | TargetLowering::ArgListEntry Entry; | |||
5749 | Args.reserve(CS.arg_size()); | |||
5750 | ||||
5751 | for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); | |||
5752 | i != e; ++i) { | |||
5753 | const Value *V = *i; | |||
5754 | ||||
5755 | // Skip empty types | |||
5756 | if (V->getType()->isEmptyTy()) | |||
5757 | continue; | |||
5758 | ||||
5759 | SDValue ArgNode = getValue(V); | |||
5760 | Entry.Node = ArgNode; Entry.Ty = V->getType(); | |||
5761 | ||||
5762 | // Skip the first return-type Attribute to get to params. | |||
5763 | Entry.setAttributes(&CS, i - CS.arg_begin() + 1); | |||
5764 | Args.push_back(Entry); | |||
5765 | } | |||
5766 | ||||
5767 | // Check if target-independent constraints permit a tail call here. | |||
5768 | // Target-dependent constraints are checked within TLI->LowerCallTo. | |||
5769 | if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget())) | |||
5770 | isTailCall = false; | |||
5771 | ||||
5772 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
5773 | CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot()) | |||
5774 | .setCallee(RetTy, FTy, Callee, std::move(Args), CS) | |||
5775 | .setTailCall(isTailCall); | |||
5776 | std::pair<SDValue,SDValue> Result = lowerInvokable(CLI, LandingPad); | |||
5777 | ||||
5778 | if (Result.first.getNode()) | |||
5779 | setValue(CS.getInstruction(), Result.first); | |||
5780 | } | |||
5781 | ||||
5782 | /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the | |||
5783 | /// value is equal or not-equal to zero. | |||
5784 | static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) { | |||
5785 | for (const User *U : V->users()) { | |||
5786 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(U)) | |||
5787 | if (IC->isEquality()) | |||
5788 | if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1))) | |||
5789 | if (C->isNullValue()) | |||
5790 | continue; | |||
5791 | // Unknown instruction. | |||
5792 | return false; | |||
5793 | } | |||
5794 | return true; | |||
5795 | } | |||
5796 | ||||
5797 | static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT, | |||
5798 | Type *LoadTy, | |||
5799 | SelectionDAGBuilder &Builder) { | |||
5800 | ||||
5801 | // Check to see if this load can be trivially constant folded, e.g. if the | |||
5802 | // input is from a string literal. | |||
5803 | if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) { | |||
5804 | // Cast pointer to the type we really want to load. | |||
5805 | LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput), | |||
5806 | PointerType::getUnqual(LoadTy)); | |||
5807 | ||||
5808 | if (const Constant *LoadCst = | |||
5809 | ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput), | |||
5810 | Builder.DL)) | |||
5811 | return Builder.getValue(LoadCst); | |||
5812 | } | |||
5813 | ||||
5814 | // Otherwise, we have to emit the load. If the pointer is to unfoldable but | |||
5815 | // still constant memory, the input chain can be the entry node. | |||
5816 | SDValue Root; | |||
5817 | bool ConstantMemory = false; | |||
5818 | ||||
5819 | // Do not serialize (non-volatile) loads of constant memory with anything. | |||
5820 | if (Builder.AA->pointsToConstantMemory(PtrVal)) { | |||
5821 | Root = Builder.DAG.getEntryNode(); | |||
5822 | ConstantMemory = true; | |||
5823 | } else { | |||
5824 | // Do not serialize non-volatile loads against each other. | |||
5825 | Root = Builder.DAG.getRoot(); | |||
5826 | } | |||
5827 | ||||
5828 | SDValue Ptr = Builder.getValue(PtrVal); | |||
5829 | SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root, | |||
5830 | Ptr, MachinePointerInfo(PtrVal), | |||
5831 | false /*volatile*/, | |||
5832 | false /*nontemporal*/, | |||
5833 | false /*isinvariant*/, 1 /* align=1 */); | |||
5834 | ||||
5835 | if (!ConstantMemory) | |||
5836 | Builder.PendingLoads.push_back(LoadVal.getValue(1)); | |||
5837 | return LoadVal; | |||
5838 | } | |||
5839 | ||||
5840 | /// processIntegerCallValue - Record the value for an instruction that | |||
5841 | /// produces an integer result, converting the type where necessary. | |||
5842 | void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I, | |||
5843 | SDValue Value, | |||
5844 | bool IsSigned) { | |||
5845 | EVT VT = DAG.getTargetLoweringInfo().getValueType(I.getType(), true); | |||
5846 | if (IsSigned) | |||
5847 | Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT); | |||
5848 | else | |||
5849 | Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT); | |||
5850 | setValue(&I, Value); | |||
5851 | } | |||
5852 | ||||
5853 | /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form. | |||
5854 | /// If so, return true and lower it, otherwise return false and it will be | |||
5855 | /// lowered like a normal call. | |||
5856 | bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) { | |||
5857 | // Verify that the prototype makes sense. int memcmp(void*,void*,size_t) | |||
5858 | if (I.getNumArgOperands() != 3) | |||
5859 | return false; | |||
5860 | ||||
5861 | const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1); | |||
5862 | if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() || | |||
5863 | !I.getArgOperand(2)->getType()->isIntegerTy() || | |||
5864 | !I.getType()->isIntegerTy()) | |||
5865 | return false; | |||
5866 | ||||
5867 | const Value *Size = I.getArgOperand(2); | |||
5868 | const ConstantInt *CSize = dyn_cast<ConstantInt>(Size); | |||
5869 | if (CSize && CSize->getZExtValue() == 0) { | |||
5870 | EVT CallVT = DAG.getTargetLoweringInfo().getValueType(I.getType(), true); | |||
5871 | setValue(&I, DAG.getConstant(0, CallVT)); | |||
5872 | return true; | |||
5873 | } | |||
5874 | ||||
5875 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); | |||
5876 | std::pair<SDValue, SDValue> Res = | |||
5877 | TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(), | |||
5878 | getValue(LHS), getValue(RHS), getValue(Size), | |||
5879 | MachinePointerInfo(LHS), | |||
5880 | MachinePointerInfo(RHS)); | |||
5881 | if (Res.first.getNode()) { | |||
5882 | processIntegerCallValue(I, Res.first, true); | |||
5883 | PendingLoads.push_back(Res.second); | |||
5884 | return true; | |||
5885 | } | |||
5886 | ||||
5887 | // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0 | |||
5888 | // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0 | |||
5889 | if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) { | |||
5890 | bool ActuallyDoIt = true; | |||
5891 | MVT LoadVT; | |||
5892 | Type *LoadTy; | |||
5893 | switch (CSize->getZExtValue()) { | |||
5894 | default: | |||
5895 | LoadVT = MVT::Other; | |||
5896 | LoadTy = nullptr; | |||
5897 | ActuallyDoIt = false; | |||
5898 | break; | |||
5899 | case 2: | |||
5900 | LoadVT = MVT::i16; | |||
5901 | LoadTy = Type::getInt16Ty(CSize->getContext()); | |||
5902 | break; | |||
5903 | case 4: | |||
5904 | LoadVT = MVT::i32; | |||
5905 | LoadTy = Type::getInt32Ty(CSize->getContext()); | |||
5906 | break; | |||
5907 | case 8: | |||
5908 | LoadVT = MVT::i64; | |||
5909 | LoadTy = Type::getInt64Ty(CSize->getContext()); | |||
5910 | break; | |||
5911 | /* | |||
5912 | case 16: | |||
5913 | LoadVT = MVT::v4i32; | |||
5914 | LoadTy = Type::getInt32Ty(CSize->getContext()); | |||
5915 | LoadTy = VectorType::get(LoadTy, 4); | |||
5916 | break; | |||
5917 | */ | |||
5918 | } | |||
5919 | ||||
5920 | // This turns into unaligned loads. We only do this if the target natively | |||
5921 | // supports the MVT we'll be loading or if it is small enough (<= 4) that | |||
5922 | // we'll only produce a small number of byte loads. | |||
5923 | ||||
5924 | // Require that we can find a legal MVT, and only do this if the target | |||
5925 | // supports unaligned loads of that type. Expanding into byte loads would | |||
5926 | // bloat the code. | |||
5927 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
5928 | if (ActuallyDoIt && CSize->getZExtValue() > 4) { | |||
5929 | unsigned DstAS = LHS->getType()->getPointerAddressSpace(); | |||
5930 | unsigned SrcAS = RHS->getType()->getPointerAddressSpace(); | |||
5931 | // TODO: Handle 5 byte compare as 4-byte + 1 byte. | |||
5932 | // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads. | |||
5933 | // TODO: Check alignment of src and dest ptrs. | |||
5934 | if (!TLI.isTypeLegal(LoadVT) || | |||
5935 | !TLI.allowsMisalignedMemoryAccesses(LoadVT, SrcAS) || | |||
5936 | !TLI.allowsMisalignedMemoryAccesses(LoadVT, DstAS)) | |||
5937 | ActuallyDoIt = false; | |||
5938 | } | |||
5939 | ||||
5940 | if (ActuallyDoIt) { | |||
5941 | SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this); | |||
5942 | SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this); | |||
5943 | ||||
5944 | SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal, | |||
5945 | ISD::SETNE); | |||
5946 | processIntegerCallValue(I, Res, false); | |||
5947 | return true; | |||
5948 | } | |||
5949 | } | |||
5950 | ||||
5951 | ||||
5952 | return false; | |||
5953 | } | |||
5954 | ||||
5955 | /// visitMemChrCall -- See if we can lower a memchr call into an optimized | |||
5956 | /// form. If so, return true and lower it, otherwise return false and it | |||
5957 | /// will be lowered like a normal call. | |||
5958 | bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) { | |||
5959 | // Verify that the prototype makes sense. void *memchr(void *, int, size_t) | |||
5960 | if (I.getNumArgOperands() != 3) | |||
5961 | return false; | |||
5962 | ||||
5963 | const Value *Src = I.getArgOperand(0); | |||
5964 | const Value *Char = I.getArgOperand(1); | |||
5965 | const Value *Length = I.getArgOperand(2); | |||
5966 | if (!Src->getType()->isPointerTy() || | |||
5967 | !Char->getType()->isIntegerTy() || | |||
5968 | !Length->getType()->isIntegerTy() || | |||
5969 | !I.getType()->isPointerTy()) | |||
5970 | return false; | |||
5971 | ||||
5972 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); | |||
5973 | std::pair<SDValue, SDValue> Res = | |||
5974 | TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(), | |||
5975 | getValue(Src), getValue(Char), getValue(Length), | |||
5976 | MachinePointerInfo(Src)); | |||
5977 | if (Res.first.getNode()) { | |||
5978 | setValue(&I, Res.first); | |||
5979 | PendingLoads.push_back(Res.second); | |||
5980 | return true; | |||
5981 | } | |||
5982 | ||||
5983 | return false; | |||
5984 | } | |||
5985 | ||||
5986 | /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an | |||
5987 | /// optimized form. If so, return true and lower it, otherwise return false | |||
5988 | /// and it will be lowered like a normal call. | |||
5989 | bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) { | |||
5990 | // Verify that the prototype makes sense. char *strcpy(char *, char *) | |||
5991 | if (I.getNumArgOperands() != 2) | |||
5992 | return false; | |||
5993 | ||||
5994 | const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); | |||
5995 | if (!Arg0->getType()->isPointerTy() || | |||
5996 | !Arg1->getType()->isPointerTy() || | |||
5997 | !I.getType()->isPointerTy()) | |||
5998 | return false; | |||
5999 | ||||
6000 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); | |||
6001 | std::pair<SDValue, SDValue> Res = | |||
6002 | TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(), | |||
6003 | getValue(Arg0), getValue(Arg1), | |||
6004 | MachinePointerInfo(Arg0), | |||
6005 | MachinePointerInfo(Arg1), isStpcpy); | |||
6006 | if (Res.first.getNode()) { | |||
6007 | setValue(&I, Res.first); | |||
6008 | DAG.setRoot(Res.second); | |||
6009 | return true; | |||
6010 | } | |||
6011 | ||||
6012 | return false; | |||
6013 | } | |||
6014 | ||||
6015 | /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form. | |||
6016 | /// If so, return true and lower it, otherwise return false and it will be | |||
6017 | /// lowered like a normal call. | |||
6018 | bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) { | |||
6019 | // Verify that the prototype makes sense. int strcmp(void*,void*) | |||
6020 | if (I.getNumArgOperands() != 2) | |||
6021 | return false; | |||
6022 | ||||
6023 | const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); | |||
6024 | if (!Arg0->getType()->isPointerTy() || | |||
6025 | !Arg1->getType()->isPointerTy() || | |||
6026 | !I.getType()->isIntegerTy()) | |||
6027 | return false; | |||
6028 | ||||
6029 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); | |||
6030 | std::pair<SDValue, SDValue> Res = | |||
6031 | TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(), | |||
6032 | getValue(Arg0), getValue(Arg1), | |||
6033 | MachinePointerInfo(Arg0), | |||
6034 | MachinePointerInfo(Arg1)); | |||
6035 | if (Res.first.getNode()) { | |||
6036 | processIntegerCallValue(I, Res.first, true); | |||
6037 | PendingLoads.push_back(Res.second); | |||
6038 | return true; | |||
6039 | } | |||
6040 | ||||
6041 | return false; | |||
6042 | } | |||
6043 | ||||
6044 | /// visitStrLenCall -- See if we can lower a strlen call into an optimized | |||
6045 | /// form. If so, return true and lower it, otherwise return false and it | |||
6046 | /// will be lowered like a normal call. | |||
6047 | bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) { | |||
6048 | // Verify that the prototype makes sense. size_t strlen(char *) | |||
6049 | if (I.getNumArgOperands() != 1) | |||
6050 | return false; | |||
6051 | ||||
6052 | const Value *Arg0 = I.getArgOperand(0); | |||
6053 | if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy()) | |||
6054 | return false; | |||
6055 | ||||
6056 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); | |||
6057 | std::pair<SDValue, SDValue> Res = | |||
6058 | TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(), | |||
6059 | getValue(Arg0), MachinePointerInfo(Arg0)); | |||
6060 | if (Res.first.getNode()) { | |||
6061 | processIntegerCallValue(I, Res.first, false); | |||
6062 | PendingLoads.push_back(Res.second); | |||
6063 | return true; | |||
6064 | } | |||
6065 | ||||
6066 | return false; | |||
6067 | } | |||
6068 | ||||
6069 | /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized | |||
6070 | /// form. If so, return true and lower it, otherwise return false and it | |||
6071 | /// will be lowered like a normal call. | |||
6072 | bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) { | |||
6073 | // Verify that the prototype makes sense. size_t strnlen(char *, size_t) | |||
6074 | if (I.getNumArgOperands() != 2) | |||
6075 | return false; | |||
6076 | ||||
6077 | const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); | |||
6078 | if (!Arg0->getType()->isPointerTy() || | |||
6079 | !Arg1->getType()->isIntegerTy() || | |||
6080 | !I.getType()->isIntegerTy()) | |||
6081 | return false; | |||
6082 | ||||
6083 | const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); | |||
6084 | std::pair<SDValue, SDValue> Res = | |||
6085 | TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(), | |||
6086 | getValue(Arg0), getValue(Arg1), | |||
6087 | MachinePointerInfo(Arg0)); | |||
6088 | if (Res.first.getNode()) { | |||
6089 | processIntegerCallValue(I, Res.first, false); | |||
6090 | PendingLoads.push_back(Res.second); | |||
6091 | return true; | |||
6092 | } | |||
6093 | ||||
6094 | return false; | |||
6095 | } | |||
6096 | ||||
6097 | /// visitUnaryFloatCall - If a call instruction is a unary floating-point | |||
6098 | /// operation (as expected), translate it to an SDNode with the specified opcode | |||
6099 | /// and return true. | |||
6100 | bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I, | |||
6101 | unsigned Opcode) { | |||
6102 | // Sanity check that it really is a unary floating-point call. | |||
6103 | if (I.getNumArgOperands() != 1 || | |||
6104 | !I.getArgOperand(0)->getType()->isFloatingPointTy() || | |||
6105 | I.getType() != I.getArgOperand(0)->getType() || | |||
6106 | !I.onlyReadsMemory()) | |||
6107 | return false; | |||
6108 | ||||
6109 | SDValue Tmp = getValue(I.getArgOperand(0)); | |||
6110 | setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp)); | |||
6111 | return true; | |||
6112 | } | |||
6113 | ||||
6114 | /// visitBinaryFloatCall - If a call instruction is a binary floating-point | |||
6115 | /// operation (as expected), translate it to an SDNode with the specified opcode | |||
6116 | /// and return true. | |||
6117 | bool SelectionDAGBuilder::visitBinaryFloatCall(const CallInst &I, | |||
6118 | unsigned Opcode) { | |||
6119 | // Sanity check that it really is a binary floating-point call. | |||
6120 | if (I.getNumArgOperands() != 2 || | |||
6121 | !I.getArgOperand(0)->getType()->isFloatingPointTy() || | |||
6122 | I.getType() != I.getArgOperand(0)->getType() || | |||
6123 | I.getType() != I.getArgOperand(1)->getType() || | |||
6124 | !I.onlyReadsMemory()) | |||
6125 | return false; | |||
6126 | ||||
6127 | SDValue Tmp0 = getValue(I.getArgOperand(0)); | |||
6128 | SDValue Tmp1 = getValue(I.getArgOperand(1)); | |||
6129 | EVT VT = Tmp0.getValueType(); | |||
6130 | setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), VT, Tmp0, Tmp1)); | |||
6131 | return true; | |||
6132 | } | |||
6133 | ||||
6134 | void SelectionDAGBuilder::visitCall(const CallInst &I) { | |||
6135 | // Handle inline assembly differently. | |||
6136 | if (isa<InlineAsm>(I.getCalledValue())) { | |||
6137 | visitInlineAsm(&I); | |||
6138 | return; | |||
6139 | } | |||
6140 | ||||
6141 | MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); | |||
6142 | ComputeUsesVAFloatArgument(I, &MMI); | |||
6143 | ||||
6144 | const char *RenameFn = nullptr; | |||
6145 | if (Function *F = I.getCalledFunction()) { | |||
6146 | if (F->isDeclaration()) { | |||
6147 | if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) { | |||
6148 | if (unsigned IID = II->getIntrinsicID(F)) { | |||
6149 | RenameFn = visitIntrinsicCall(I, IID); | |||
6150 | if (!RenameFn) | |||
6151 | return; | |||
6152 | } | |||
6153 | } | |||
6154 | if (unsigned IID = F->getIntrinsicID()) { | |||
6155 | RenameFn = visitIntrinsicCall(I, IID); | |||
6156 | if (!RenameFn) | |||
6157 | return; | |||
6158 | } | |||
6159 | } | |||
6160 | ||||
6161 | // Check for well-known libc/libm calls. If the function is internal, it | |||
6162 | // can't be a library call. | |||
6163 | LibFunc::Func Func; | |||
6164 | if (!F->hasLocalLinkage() && F->hasName() && | |||
6165 | LibInfo->getLibFunc(F->getName(), Func) && | |||
6166 | LibInfo->hasOptimizedCodeGen(Func)) { | |||
6167 | switch (Func) { | |||
6168 | default: break; | |||
6169 | case LibFunc::copysign: | |||
6170 | case LibFunc::copysignf: | |||
6171 | case LibFunc::copysignl: | |||
6172 | if (I.getNumArgOperands() == 2 && // Basic sanity checks. | |||
6173 | I.getArgOperand(0)->getType()->isFloatingPointTy() && | |||
6174 | I.getType() == I.getArgOperand(0)->getType() && | |||
6175 | I.getType() == I.getArgOperand(1)->getType() && | |||
6176 | I.onlyReadsMemory()) { | |||
6177 | SDValue LHS = getValue(I.getArgOperand(0)); | |||
6178 | SDValue RHS = getValue(I.getArgOperand(1)); | |||
6179 | setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(), | |||
6180 | LHS.getValueType(), LHS, RHS)); | |||
6181 | return; | |||
6182 | } | |||
6183 | break; | |||
6184 | case LibFunc::fabs: | |||
6185 | case LibFunc::fabsf: | |||
6186 | case LibFunc::fabsl: | |||
6187 | if (visitUnaryFloatCall(I, ISD::FABS)) | |||
6188 | return; | |||
6189 | break; | |||
6190 | case LibFunc::fmin: | |||
6191 | case LibFunc::fminf: | |||
6192 | case LibFunc::fminl: | |||
6193 | if (visitBinaryFloatCall(I, ISD::FMINNUM)) | |||
6194 | return; | |||
6195 | break; | |||
6196 | case LibFunc::fmax: | |||
6197 | case LibFunc::fmaxf: | |||
6198 | case LibFunc::fmaxl: | |||
6199 | if (visitBinaryFloatCall(I, ISD::FMAXNUM)) | |||
6200 | return; | |||
6201 | break; | |||
6202 | case LibFunc::sin: | |||
6203 | case LibFunc::sinf: | |||
6204 | case LibFunc::sinl: | |||
6205 | if (visitUnaryFloatCall(I, ISD::FSIN)) | |||
6206 | return; | |||
6207 | break; | |||
6208 | case LibFunc::cos: | |||
6209 | case LibFunc::cosf: | |||
6210 | case LibFunc::cosl: | |||
6211 | if (visitUnaryFloatCall(I, ISD::FCOS)) | |||
6212 | return; | |||
6213 | break; | |||
6214 | case LibFunc::sqrt: | |||
6215 | case LibFunc::sqrtf: | |||
6216 | case LibFunc::sqrtl: | |||
6217 | case LibFunc::sqrt_finite: | |||
6218 | case LibFunc::sqrtf_finite: | |||
6219 | case LibFunc::sqrtl_finite: | |||
6220 | if (visitUnaryFloatCall(I, ISD::FSQRT)) | |||
6221 | return; | |||
6222 | break; | |||
6223 | case LibFunc::floor: | |||
6224 | case LibFunc::floorf: | |||
6225 | case LibFunc::floorl: | |||
6226 | if (visitUnaryFloatCall(I, ISD::FFLOOR)) | |||
6227 | return; | |||
6228 | break; | |||
6229 | case LibFunc::nearbyint: | |||
6230 | case LibFunc::nearbyintf: | |||
6231 | case LibFunc::nearbyintl: | |||
6232 | if (visitUnaryFloatCall(I, ISD::FNEARBYINT)) | |||
6233 | return; | |||
6234 | break; | |||
6235 | case LibFunc::ceil: | |||
6236 | case LibFunc::ceilf: | |||
6237 | case LibFunc::ceill: | |||
6238 | if (visitUnaryFloatCall(I, ISD::FCEIL)) | |||
6239 | return; | |||
6240 | break; | |||
6241 | case LibFunc::rint: | |||
6242 | case LibFunc::rintf: | |||
6243 | case LibFunc::rintl: | |||
6244 | if (visitUnaryFloatCall(I, ISD::FRINT)) | |||
6245 | return; | |||
6246 | break; | |||
6247 | case LibFunc::round: | |||
6248 | case LibFunc::roundf: | |||
6249 | case LibFunc::roundl: | |||
6250 | if (visitUnaryFloatCall(I, ISD::FROUND)) | |||
6251 | return; | |||
6252 | break; | |||
6253 | case LibFunc::trunc: | |||
6254 | case LibFunc::truncf: | |||
6255 | case LibFunc::truncl: | |||
6256 | if (visitUnaryFloatCall(I, ISD::FTRUNC)) | |||
6257 | return; | |||
6258 | break; | |||
6259 | case LibFunc::log2: | |||
6260 | case LibFunc::log2f: | |||
6261 | case LibFunc::log2l: | |||
6262 | if (visitUnaryFloatCall(I, ISD::FLOG2)) | |||
6263 | return; | |||
6264 | break; | |||
6265 | case LibFunc::exp2: | |||
6266 | case LibFunc::exp2f: | |||
6267 | case LibFunc::exp2l: | |||
6268 | if (visitUnaryFloatCall(I, ISD::FEXP2)) | |||
6269 | return; | |||
6270 | break; | |||
6271 | case LibFunc::memcmp: | |||
6272 | if (visitMemCmpCall(I)) | |||
6273 | return; | |||
6274 | break; | |||
6275 | case LibFunc::memchr: | |||
6276 | if (visitMemChrCall(I)) | |||
6277 | return; | |||
6278 | break; | |||
6279 | case LibFunc::strcpy: | |||
6280 | if (visitStrCpyCall(I, false)) | |||
6281 | return; | |||
6282 | break; | |||
6283 | case LibFunc::stpcpy: | |||
6284 | if (visitStrCpyCall(I, true)) | |||
6285 | return; | |||
6286 | break; | |||
6287 | case LibFunc::strcmp: | |||
6288 | if (visitStrCmpCall(I)) | |||
6289 | return; | |||
6290 | break; | |||
6291 | case LibFunc::strlen: | |||
6292 | if (visitStrLenCall(I)) | |||
6293 | return; | |||
6294 | break; | |||
6295 | case LibFunc::strnlen: | |||
6296 | if (visitStrNLenCall(I)) | |||
6297 | return; | |||
6298 | break; | |||
6299 | } | |||
6300 | } | |||
6301 | } | |||
6302 | ||||
6303 | SDValue Callee; | |||
6304 | if (!RenameFn) | |||
6305 | Callee = getValue(I.getCalledValue()); | |||
6306 | else | |||
6307 | Callee = DAG.getExternalSymbol(RenameFn, | |||
6308 | DAG.getTargetLoweringInfo().getPointerTy()); | |||
6309 | ||||
6310 | // Check if we can potentially perform a tail call. More detailed checking is | |||
6311 | // be done within LowerCallTo, after more information about the call is known. | |||
6312 | LowerCallTo(&I, Callee, I.isTailCall()); | |||
6313 | } | |||
6314 | ||||
6315 | namespace { | |||
6316 | ||||
6317 | /// AsmOperandInfo - This contains information for each constraint that we are | |||
6318 | /// lowering. | |||
6319 | class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo { | |||
6320 | public: | |||
6321 | /// CallOperand - If this is the result output operand or a clobber | |||
6322 | /// this is null, otherwise it is the incoming operand to the CallInst. | |||
6323 | /// This gets modified as the asm is processed. | |||
6324 | SDValue CallOperand; | |||
6325 | ||||
6326 | /// AssignedRegs - If this is a register or register class operand, this | |||
6327 | /// contains the set of register corresponding to the operand. | |||
6328 | RegsForValue AssignedRegs; | |||
6329 | ||||
6330 | explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info) | |||
6331 | : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) { | |||
6332 | } | |||
6333 | ||||
6334 | /// getCallOperandValEVT - Return the EVT of the Value* that this operand | |||
6335 | /// corresponds to. If there is no Value* for this operand, it returns | |||
6336 | /// MVT::Other. | |||
6337 | EVT getCallOperandValEVT(LLVMContext &Context, | |||
6338 | const TargetLowering &TLI, | |||
6339 | const DataLayout *DL) const { | |||
6340 | if (!CallOperandVal) return MVT::Other; | |||
6341 | ||||
6342 | if (isa<BasicBlock>(CallOperandVal)) | |||
6343 | return TLI.getPointerTy(); | |||
6344 | ||||
6345 | llvm::Type *OpTy = CallOperandVal->getType(); | |||
6346 | ||||
6347 | // FIXME: code duplicated from TargetLowering::ParseConstraints(). | |||
6348 | // If this is an indirect operand, the operand is a pointer to the | |||
6349 | // accessed type. | |||
6350 | if (isIndirect) { | |||
6351 | llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); | |||
6352 | if (!PtrTy) | |||
6353 | report_fatal_error("Indirect operand for inline asm not a pointer!"); | |||
6354 | OpTy = PtrTy->getElementType(); | |||
6355 | } | |||
6356 | ||||
6357 | // Look for vector wrapped in a struct. e.g. { <16 x i8> }. | |||
6358 | if (StructType *STy = dyn_cast<StructType>(OpTy)) | |||
6359 | if (STy->getNumElements() == 1) | |||
6360 | OpTy = STy->getElementType(0); | |||
6361 | ||||
6362 | // If OpTy is not a single value, it may be a struct/union that we | |||
6363 | // can tile with integers. | |||
6364 | if (!OpTy->isSingleValueType() && OpTy->isSized()) { | |||
6365 | unsigned BitSize = DL->getTypeSizeInBits(OpTy); | |||
6366 | switch (BitSize) { | |||
6367 | default: break; | |||
6368 | case 1: | |||
6369 | case 8: | |||
6370 | case 16: | |||
6371 | case 32: | |||
6372 | case 64: | |||
6373 | case 128: | |||
6374 | OpTy = IntegerType::get(Context, BitSize); | |||
6375 | break; | |||
6376 | } | |||
6377 | } | |||
6378 | ||||
6379 | return TLI.getValueType(OpTy, true); | |||
6380 | } | |||
6381 | }; | |||
6382 | ||||
6383 | typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector; | |||
6384 | ||||
6385 | } // end anonymous namespace | |||
6386 | ||||
6387 | /// GetRegistersForValue - Assign registers (virtual or physical) for the | |||
6388 | /// specified operand. We prefer to assign virtual registers, to allow the | |||
6389 | /// register allocator to handle the assignment process. However, if the asm | |||
6390 | /// uses features that we can't model on machineinstrs, we have SDISel do the | |||
6391 | /// allocation. This produces generally horrible, but correct, code. | |||
6392 | /// | |||
6393 | /// OpInfo describes the operand. | |||
6394 | /// | |||
6395 | static void GetRegistersForValue(SelectionDAG &DAG, | |||
6396 | const TargetLowering &TLI, | |||
6397 | SDLoc DL, | |||
6398 | SDISelAsmOperandInfo &OpInfo) { | |||
6399 | LLVMContext &Context = *DAG.getContext(); | |||
6400 | ||||
6401 | MachineFunction &MF = DAG.getMachineFunction(); | |||
6402 | SmallVector<unsigned, 4> Regs; | |||
6403 | ||||
6404 | // If this is a constraint for a single physreg, or a constraint for a | |||
6405 | // register class, find it. | |||
6406 | std::pair<unsigned, const TargetRegisterClass*> PhysReg = | |||
6407 | TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, | |||
| ||||
6408 | OpInfo.ConstraintVT); | |||
6409 | ||||
6410 | unsigned NumRegs = 1; | |||
6411 | if (OpInfo.ConstraintVT != MVT::Other) { | |||
6412 | // If this is a FP input in an integer register (or visa versa) insert a bit | |||
6413 | // cast of the input value. More generally, handle any case where the input | |||
6414 | // value disagrees with the register class we plan to stick this in. | |||
6415 | if (OpInfo.Type == InlineAsm::isInput && | |||
6416 | PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) { | |||
6417 | // Try to convert to the first EVT that the reg class contains. If the | |||
6418 | // types are identical size, use a bitcast to convert (e.g. two differing | |||
6419 | // vector types). | |||
6420 | MVT RegVT = *PhysReg.second->vt_begin(); | |||
6421 | if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) { | |||
6422 | OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL, | |||
6423 | RegVT, OpInfo.CallOperand); | |||
6424 | OpInfo.ConstraintVT = RegVT; | |||
6425 | } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) { | |||
6426 | // If the input is a FP value and we want it in FP registers, do a | |||
6427 | // bitcast to the corresponding integer type. This turns an f64 value | |||
6428 | // into i64, which can be passed with two i32 values on a 32-bit | |||
6429 | // machine. | |||
6430 | RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits()); | |||
6431 | OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL, | |||
6432 | RegVT, OpInfo.CallOperand); | |||
6433 | OpInfo.ConstraintVT = RegVT; | |||
6434 | } | |||
6435 | } | |||
6436 | ||||
6437 | NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT); | |||
6438 | } | |||
6439 | ||||
6440 | MVT RegVT; | |||
6441 | EVT ValueVT = OpInfo.ConstraintVT; | |||
6442 | ||||
6443 | // If this is a constraint for a specific physical register, like {r17}, | |||
6444 | // assign it now. | |||
6445 | if (unsigned AssignedReg = PhysReg.first) { | |||
6446 | const TargetRegisterClass *RC = PhysReg.second; | |||
6447 | if (OpInfo.ConstraintVT == MVT::Other) | |||
6448 | ValueVT = *RC->vt_begin(); | |||
6449 | ||||
6450 | // Get the actual register value type. This is important, because the user | |||
6451 | // may have asked for (e.g.) the AX register in i32 type. We need to | |||
6452 | // remember that AX is actually i16 to get the right extension. | |||
6453 | RegVT = *RC->vt_begin(); | |||
| ||||
6454 | ||||
6455 | // This is a explicit reference to a physical register. | |||
6456 | Regs.push_back(AssignedReg); | |||
6457 | ||||
6458 | // If this is an expanded reference, add the rest of the regs to Regs. | |||
6459 | if (NumRegs != 1) { | |||
6460 | TargetRegisterClass::iterator I = RC->begin(); | |||
6461 | for (; *I != AssignedReg; ++I) | |||
6462 | assert(I != RC->end() && "Didn't find reg!")((I != RC->end() && "Didn't find reg!") ? static_cast <void> (0) : __assert_fail ("I != RC->end() && \"Didn't find reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6462, __PRETTY_FUNCTION__)); | |||
6463 | ||||
6464 | // Already added the first reg. | |||
6465 | --NumRegs; ++I; | |||
6466 | for (; NumRegs; --NumRegs, ++I) { | |||
6467 | assert(I != RC->end() && "Ran out of registers to allocate!")((I != RC->end() && "Ran out of registers to allocate!" ) ? static_cast<void> (0) : __assert_fail ("I != RC->end() && \"Ran out of registers to allocate!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6467, __PRETTY_FUNCTION__)); | |||
6468 | Regs.push_back(*I); | |||
6469 | } | |||
6470 | } | |||
6471 | ||||
6472 | OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); | |||
6473 | return; | |||
6474 | } | |||
6475 | ||||
6476 | // Otherwise, if this was a reference to an LLVM register class, create vregs | |||
6477 | // for this reference. | |||
6478 | if (const TargetRegisterClass *RC = PhysReg.second) { | |||
6479 | RegVT = *RC->vt_begin(); | |||
6480 | if (OpInfo.ConstraintVT == MVT::Other) | |||
6481 | ValueVT = RegVT; | |||
6482 | ||||
6483 | // Create the appropriate number of virtual registers. | |||
6484 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); | |||
6485 | for (; NumRegs; --NumRegs) | |||
6486 | Regs.push_back(RegInfo.createVirtualRegister(RC)); | |||
6487 | ||||
6488 | OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); | |||
6489 | return; | |||
6490 | } | |||
6491 | ||||
6492 | // Otherwise, we couldn't allocate enough registers for this. | |||
6493 | } | |||
6494 | ||||
6495 | /// visitInlineAsm - Handle a call to an InlineAsm object. | |||
6496 | /// | |||
6497 | void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) { | |||
6498 | const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); | |||
6499 | ||||
6500 | /// ConstraintOperands - Information about all of the constraints. | |||
6501 | SDISelAsmOperandInfoVector ConstraintOperands; | |||
6502 | ||||
6503 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
6504 | TargetLowering::AsmOperandInfoVector | |||
6505 | TargetConstraints = TLI.ParseConstraints(CS); | |||
6506 | ||||
6507 | bool hasMemory = false; | |||
6508 | ||||
6509 | unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. | |||
6510 | unsigned ResNo = 0; // ResNo - The result number of the next output. | |||
6511 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { | |||
6512 | ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i])); | |||
6513 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back(); | |||
6514 | ||||
6515 | MVT OpVT = MVT::Other; | |||
6516 | ||||
6517 | // Compute the value type for each operand. | |||
6518 | switch (OpInfo.Type) { | |||
6519 | case InlineAsm::isOutput: | |||
6520 | // Indirect outputs just consume an argument. | |||
6521 | if (OpInfo.isIndirect) { | |||
6522 | OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); | |||
6523 | break; | |||
6524 | } | |||
6525 | ||||
6526 | // The return value of the call is this value. As such, there is no | |||
6527 | // corresponding argument. | |||
6528 | assert(!CS.getType()->isVoidTy() && "Bad inline asm!")((!CS.getType()->isVoidTy() && "Bad inline asm!") ? static_cast<void> (0) : __assert_fail ("!CS.getType()->isVoidTy() && \"Bad inline asm!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6528, __PRETTY_FUNCTION__)); | |||
6529 | if (StructType *STy = dyn_cast<StructType>(CS.getType())) { | |||
6530 | OpVT = TLI.getSimpleValueType(STy->getElementType(ResNo)); | |||
6531 | } else { | |||
6532 | assert(ResNo == 0 && "Asm only has one result!")((ResNo == 0 && "Asm only has one result!") ? static_cast <void> (0) : __assert_fail ("ResNo == 0 && \"Asm only has one result!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6532, __PRETTY_FUNCTION__)); | |||
6533 | OpVT = TLI.getSimpleValueType(CS.getType()); | |||
6534 | } | |||
6535 | ++ResNo; | |||
6536 | break; | |||
6537 | case InlineAsm::isInput: | |||
6538 | OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); | |||
6539 | break; | |||
6540 | case InlineAsm::isClobber: | |||
6541 | // Nothing to do. | |||
6542 | break; | |||
6543 | } | |||
6544 | ||||
6545 | // If this is an input or an indirect output, process the call argument. | |||
6546 | // BasicBlocks are labels, currently appearing only in asm's. | |||
6547 | if (OpInfo.CallOperandVal) { | |||
6548 | if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) { | |||
6549 | OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]); | |||
6550 | } else { | |||
6551 | OpInfo.CallOperand = getValue(OpInfo.CallOperandVal); | |||
6552 | } | |||
6553 | ||||
6554 | OpVT = | |||
6555 | OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, DL).getSimpleVT(); | |||
6556 | } | |||
6557 | ||||
6558 | OpInfo.ConstraintVT = OpVT; | |||
6559 | ||||
6560 | // Indirect operand accesses access memory. | |||
6561 | if (OpInfo.isIndirect) | |||
6562 | hasMemory = true; | |||
6563 | else { | |||
6564 | for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) { | |||
6565 | TargetLowering::ConstraintType | |||
6566 | CType = TLI.getConstraintType(OpInfo.Codes[j]); | |||
6567 | if (CType == TargetLowering::C_Memory) { | |||
6568 | hasMemory = true; | |||
6569 | break; | |||
6570 | } | |||
6571 | } | |||
6572 | } | |||
6573 | } | |||
6574 | ||||
6575 | SDValue Chain, Flag; | |||
6576 | ||||
6577 | // We won't need to flush pending loads if this asm doesn't touch | |||
6578 | // memory and is nonvolatile. | |||
6579 | if (hasMemory || IA->hasSideEffects()) | |||
6580 | Chain = getRoot(); | |||
6581 | else | |||
6582 | Chain = DAG.getRoot(); | |||
6583 | ||||
6584 | // Second pass over the constraints: compute which constraint option to use | |||
6585 | // and assign registers to constraints that want a specific physreg. | |||
6586 | for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { | |||
6587 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; | |||
6588 | ||||
6589 | // If this is an output operand with a matching input operand, look up the | |||
6590 | // matching input. If their types mismatch, e.g. one is an integer, the | |||
6591 | // other is floating point, or their sizes are different, flag it as an | |||
6592 | // error. | |||
6593 | if (OpInfo.hasMatchingInput()) { | |||
6594 | SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; | |||
6595 | ||||
6596 | if (OpInfo.ConstraintVT != Input.ConstraintVT) { | |||
6597 | std::pair<unsigned, const TargetRegisterClass*> MatchRC = | |||
6598 | TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, | |||
6599 | OpInfo.ConstraintVT); | |||
6600 | std::pair<unsigned, const TargetRegisterClass*> InputRC = | |||
6601 | TLI.getRegForInlineAsmConstraint(Input.ConstraintCode, | |||
6602 | Input.ConstraintVT); | |||
6603 | if ((OpInfo.ConstraintVT.isInteger() != | |||
6604 | Input.ConstraintVT.isInteger()) || | |||
6605 | (MatchRC.second != InputRC.second)) { | |||
6606 | report_fatal_error("Unsupported asm: input constraint" | |||
6607 | " with a matching output constraint of" | |||
6608 | " incompatible type!"); | |||
6609 | } | |||
6610 | Input.ConstraintVT = OpInfo.ConstraintVT; | |||
6611 | } | |||
6612 | } | |||
6613 | ||||
6614 | // Compute the constraint code and ConstraintType to use. | |||
6615 | TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG); | |||
6616 | ||||
6617 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && | |||
6618 | OpInfo.Type == InlineAsm::isClobber) | |||
6619 | continue; | |||
6620 | ||||
6621 | // If this is a memory input, and if the operand is not indirect, do what we | |||
6622 | // need to to provide an address for the memory input. | |||
6623 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && | |||
6624 | !OpInfo.isIndirect) { | |||
6625 | assert((OpInfo.isMultipleAlternative ||(((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm:: isInput)) && "Can only indirectify direct input operands!" ) ? static_cast<void> (0) : __assert_fail ("(OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) && \"Can only indirectify direct input operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6627, __PRETTY_FUNCTION__)) | |||
6626 | (OpInfo.Type == InlineAsm::isInput)) &&(((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm:: isInput)) && "Can only indirectify direct input operands!" ) ? static_cast<void> (0) : __assert_fail ("(OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) && \"Can only indirectify direct input operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6627, __PRETTY_FUNCTION__)) | |||
6627 | "Can only indirectify direct input operands!")(((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm:: isInput)) && "Can only indirectify direct input operands!" ) ? static_cast<void> (0) : __assert_fail ("(OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) && \"Can only indirectify direct input operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6627, __PRETTY_FUNCTION__)); | |||
6628 | ||||
6629 | // Memory operands really want the address of the value. If we don't have | |||
6630 | // an indirect input, put it in the constpool if we can, otherwise spill | |||
6631 | // it to a stack slot. | |||
6632 | // TODO: This isn't quite right. We need to handle these according to | |||
6633 | // the addressing mode that the constraint wants. Also, this may take | |||
6634 | // an additional register for the computation and we don't want that | |||
6635 | // either. | |||
6636 | ||||
6637 | // If the operand is a float, integer, or vector constant, spill to a | |||
6638 | // constant pool entry to get its address. | |||
6639 | const Value *OpVal = OpInfo.CallOperandVal; | |||
6640 | if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) || | |||
6641 | isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) { | |||
6642 | OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal), | |||
6643 | TLI.getPointerTy()); | |||
6644 | } else { | |||
6645 | // Otherwise, create a stack slot and emit a store to it before the | |||
6646 | // asm. | |||
6647 | Type *Ty = OpVal->getType(); | |||
6648 | uint64_t TySize = TLI.getDataLayout()->getTypeAllocSize(Ty); | |||
6649 | unsigned Align = TLI.getDataLayout()->getPrefTypeAlignment(Ty); | |||
6650 | MachineFunction &MF = DAG.getMachineFunction(); | |||
6651 | int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); | |||
6652 | SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); | |||
6653 | Chain = DAG.getStore(Chain, getCurSDLoc(), | |||
6654 | OpInfo.CallOperand, StackSlot, | |||
6655 | MachinePointerInfo::getFixedStack(SSFI), | |||
6656 | false, false, 0); | |||
6657 | OpInfo.CallOperand = StackSlot; | |||
6658 | } | |||
6659 | ||||
6660 | // There is no longer a Value* corresponding to this operand. | |||
6661 | OpInfo.CallOperandVal = nullptr; | |||
6662 | ||||
6663 | // It is now an indirect operand. | |||
6664 | OpInfo.isIndirect = true; | |||
6665 | } | |||
6666 | ||||
6667 | // If this constraint is for a specific register, allocate it before | |||
6668 | // anything else. | |||
6669 | if (OpInfo.ConstraintType == TargetLowering::C_Register) | |||
6670 | GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo); | |||
6671 | } | |||
6672 | ||||
6673 | // Second pass - Loop over all of the operands, assigning virtual or physregs | |||
6674 | // to register class operands. | |||
6675 | for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { | |||
6676 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; | |||
6677 | ||||
6678 | // C_Register operands have already been allocated, Other/Memory don't need | |||
6679 | // to be. | |||
6680 | if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass) | |||
6681 | GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo); | |||
6682 | } | |||
6683 | ||||
6684 | // AsmNodeOperands - The operands for the ISD::INLINEASM node. | |||
6685 | std::vector<SDValue> AsmNodeOperands; | |||
6686 | AsmNodeOperands.push_back(SDValue()); // reserve space for input chain | |||
6687 | AsmNodeOperands.push_back( | |||
6688 | DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), | |||
6689 | TLI.getPointerTy())); | |||
6690 | ||||
6691 | // If we have a !srcloc metadata node associated with it, we want to attach | |||
6692 | // this to the ultimately generated inline asm machineinstr. To do this, we | |||
6693 | // pass in the third operand as this (potentially null) inline asm MDNode. | |||
6694 | const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc"); | |||
6695 | AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc)); | |||
6696 | ||||
6697 | // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore | |||
6698 | // bits as operand 3. | |||
6699 | unsigned ExtraInfo = 0; | |||
6700 | if (IA->hasSideEffects()) | |||
6701 | ExtraInfo |= InlineAsm::Extra_HasSideEffects; | |||
6702 | if (IA->isAlignStack()) | |||
6703 | ExtraInfo |= InlineAsm::Extra_IsAlignStack; | |||
6704 | // Set the asm dialect. | |||
6705 | ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect; | |||
6706 | ||||
6707 | // Determine if this InlineAsm MayLoad or MayStore based on the constraints. | |||
6708 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { | |||
6709 | TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; | |||
6710 | ||||
6711 | // Compute the constraint code and ConstraintType to use. | |||
6712 | TLI.ComputeConstraintToUse(OpInfo, SDValue()); | |||
6713 | ||||
6714 | // Ideally, we would only check against memory constraints. However, the | |||
6715 | // meaning of an other constraint can be target-specific and we can't easily | |||
6716 | // reason about it. Therefore, be conservative and set MayLoad/MayStore | |||
6717 | // for other constriants as well. | |||
6718 | if (OpInfo.ConstraintType == TargetLowering::C_Memory || | |||
6719 | OpInfo.ConstraintType == TargetLowering::C_Other) { | |||
6720 | if (OpInfo.Type == InlineAsm::isInput) | |||
6721 | ExtraInfo |= InlineAsm::Extra_MayLoad; | |||
6722 | else if (OpInfo.Type == InlineAsm::isOutput) | |||
6723 | ExtraInfo |= InlineAsm::Extra_MayStore; | |||
6724 | else if (OpInfo.Type == InlineAsm::isClobber) | |||
6725 | ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore); | |||
6726 | } | |||
6727 | } | |||
6728 | ||||
6729 | AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo, | |||
6730 | TLI.getPointerTy())); | |||
6731 | ||||
6732 | // Loop over all of the inputs, copying the operand values into the | |||
6733 | // appropriate registers and processing the output regs. | |||
6734 | RegsForValue RetValRegs; | |||
6735 | ||||
6736 | // IndirectStoresToEmit - The set of stores to emit after the inline asm node. | |||
6737 | std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; | |||
6738 | ||||
6739 | for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { | |||
6740 | SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; | |||
6741 | ||||
6742 | switch (OpInfo.Type) { | |||
6743 | case InlineAsm::isOutput: { | |||
6744 | if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass && | |||
6745 | OpInfo.ConstraintType != TargetLowering::C_Register) { | |||
6746 | // Memory output, or 'other' output (e.g. 'X' constraint). | |||
6747 | assert(OpInfo.isIndirect && "Memory output must be indirect operand")((OpInfo.isIndirect && "Memory output must be indirect operand" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.isIndirect && \"Memory output must be indirect operand\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6747, __PRETTY_FUNCTION__)); | |||
6748 | ||||
6749 | // Add information to the INLINEASM node to know about this output. | |||
6750 | unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); | |||
6751 | AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, | |||
6752 | TLI.getPointerTy())); | |||
6753 | AsmNodeOperands.push_back(OpInfo.CallOperand); | |||
6754 | break; | |||
6755 | } | |||
6756 | ||||
6757 | // Otherwise, this is a register or register class output. | |||
6758 | ||||
6759 | // Copy the output from the appropriate register. Find a register that | |||
6760 | // we can use. | |||
6761 | if (OpInfo.AssignedRegs.Regs.empty()) { | |||
6762 | LLVMContext &Ctx = *DAG.getContext(); | |||
6763 | Ctx.emitError(CS.getInstruction(), | |||
6764 | "couldn't allocate output register for constraint '" + | |||
6765 | Twine(OpInfo.ConstraintCode) + "'"); | |||
6766 | return; | |||
6767 | } | |||
6768 | ||||
6769 | // If this is an indirect operand, store through the pointer after the | |||
6770 | // asm. | |||
6771 | if (OpInfo.isIndirect) { | |||
6772 | IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs, | |||
6773 | OpInfo.CallOperandVal)); | |||
6774 | } else { | |||
6775 | // This is the result value of the call. | |||
6776 | assert(!CS.getType()->isVoidTy() && "Bad inline asm!")((!CS.getType()->isVoidTy() && "Bad inline asm!") ? static_cast<void> (0) : __assert_fail ("!CS.getType()->isVoidTy() && \"Bad inline asm!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6776, __PRETTY_FUNCTION__)); | |||
6777 | // Concatenate this output onto the outputs list. | |||
6778 | RetValRegs.append(OpInfo.AssignedRegs); | |||
6779 | } | |||
6780 | ||||
6781 | // Add information to the INLINEASM node to know that this register is | |||
6782 | // set. | |||
6783 | OpInfo.AssignedRegs | |||
6784 | .AddInlineAsmOperands(OpInfo.isEarlyClobber | |||
6785 | ? InlineAsm::Kind_RegDefEarlyClobber | |||
6786 | : InlineAsm::Kind_RegDef, | |||
6787 | false, 0, DAG, AsmNodeOperands); | |||
6788 | break; | |||
6789 | } | |||
6790 | case InlineAsm::isInput: { | |||
6791 | SDValue InOperandVal = OpInfo.CallOperand; | |||
6792 | ||||
6793 | if (OpInfo.isMatchingInputConstraint()) { // Matching constraint? | |||
6794 | // If this is required to match an output register we have already set, | |||
6795 | // just use its register. | |||
6796 | unsigned OperandNo = OpInfo.getMatchedOperand(); | |||
6797 | ||||
6798 | // Scan until we find the definition we already emitted of this operand. | |||
6799 | // When we find it, create a RegsForValue operand. | |||
6800 | unsigned CurOp = InlineAsm::Op_FirstOperand; | |||
6801 | for (; OperandNo; --OperandNo) { | |||
6802 | // Advance to the next operand. | |||
6803 | unsigned OpFlag = | |||
6804 | cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); | |||
6805 | assert((InlineAsm::isRegDefKind(OpFlag) ||(((InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind (OpFlag) || InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?" ) ? static_cast<void> (0) : __assert_fail ("(InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind(OpFlag) || InlineAsm::isMemKind(OpFlag)) && \"Skipped past definitions?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6807, __PRETTY_FUNCTION__)) | |||
6806 | InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||(((InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind (OpFlag) || InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?" ) ? static_cast<void> (0) : __assert_fail ("(InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind(OpFlag) || InlineAsm::isMemKind(OpFlag)) && \"Skipped past definitions?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6807, __PRETTY_FUNCTION__)) | |||
6807 | InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?")(((InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind (OpFlag) || InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?" ) ? static_cast<void> (0) : __assert_fail ("(InlineAsm::isRegDefKind(OpFlag) || InlineAsm::isRegDefEarlyClobberKind(OpFlag) || InlineAsm::isMemKind(OpFlag)) && \"Skipped past definitions?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6807, __PRETTY_FUNCTION__)); | |||
6808 | CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1; | |||
6809 | } | |||
6810 | ||||
6811 | unsigned OpFlag = | |||
6812 | cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); | |||
6813 | if (InlineAsm::isRegDefKind(OpFlag) || | |||
6814 | InlineAsm::isRegDefEarlyClobberKind(OpFlag)) { | |||
6815 | // Add (OpFlag&0xffff)>>3 registers to MatchedRegs. | |||
6816 | if (OpInfo.isIndirect) { | |||
6817 | // This happens on gcc/testsuite/gcc.dg/pr8788-1.c | |||
6818 | LLVMContext &Ctx = *DAG.getContext(); | |||
6819 | Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:" | |||
6820 | " don't know how to handle tied " | |||
6821 | "indirect register inputs"); | |||
6822 | return; | |||
6823 | } | |||
6824 | ||||
6825 | RegsForValue MatchedRegs; | |||
6826 | MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType()); | |||
6827 | MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType(); | |||
6828 | MatchedRegs.RegVTs.push_back(RegVT); | |||
6829 | MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo(); | |||
6830 | for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag); | |||
6831 | i != e; ++i) { | |||
6832 | if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT)) | |||
6833 | MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC)); | |||
6834 | else { | |||
6835 | LLVMContext &Ctx = *DAG.getContext(); | |||
6836 | Ctx.emitError(CS.getInstruction(), | |||
6837 | "inline asm error: This value" | |||
6838 | " type register class is not natively supported!"); | |||
6839 | return; | |||
6840 | } | |||
6841 | } | |||
6842 | // Use the produced MatchedRegs object to | |||
6843 | MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(), | |||
6844 | Chain, &Flag, CS.getInstruction()); | |||
6845 | MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, | |||
6846 | true, OpInfo.getMatchedOperand(), | |||
6847 | DAG, AsmNodeOperands); | |||
6848 | break; | |||
6849 | } | |||
6850 | ||||
6851 | assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!")((InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!" ) ? static_cast<void> (0) : __assert_fail ("InlineAsm::isMemKind(OpFlag) && \"Unknown matching constraint!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6851, __PRETTY_FUNCTION__)); | |||
6852 | assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&((InlineAsm::getNumOperandRegisters(OpFlag) == 1 && "Unexpected number of operands" ) ? static_cast<void> (0) : __assert_fail ("InlineAsm::getNumOperandRegisters(OpFlag) == 1 && \"Unexpected number of operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6853, __PRETTY_FUNCTION__)) | |||
6853 | "Unexpected number of operands")((InlineAsm::getNumOperandRegisters(OpFlag) == 1 && "Unexpected number of operands" ) ? static_cast<void> (0) : __assert_fail ("InlineAsm::getNumOperandRegisters(OpFlag) == 1 && \"Unexpected number of operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6853, __PRETTY_FUNCTION__)); | |||
6854 | // Add information to the INLINEASM node to know about this input. | |||
6855 | // See InlineAsm.h isUseOperandTiedToDef. | |||
6856 | OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag, | |||
6857 | OpInfo.getMatchedOperand()); | |||
6858 | AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag, | |||
6859 | TLI.getPointerTy())); | |||
6860 | AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]); | |||
6861 | break; | |||
6862 | } | |||
6863 | ||||
6864 | // Treat indirect 'X' constraint as memory. | |||
6865 | if (OpInfo.ConstraintType == TargetLowering::C_Other && | |||
6866 | OpInfo.isIndirect) | |||
6867 | OpInfo.ConstraintType = TargetLowering::C_Memory; | |||
6868 | ||||
6869 | if (OpInfo.ConstraintType == TargetLowering::C_Other) { | |||
6870 | std::vector<SDValue> Ops; | |||
6871 | TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode, | |||
6872 | Ops, DAG); | |||
6873 | if (Ops.empty()) { | |||
6874 | LLVMContext &Ctx = *DAG.getContext(); | |||
6875 | Ctx.emitError(CS.getInstruction(), | |||
6876 | "invalid operand for inline asm constraint '" + | |||
6877 | Twine(OpInfo.ConstraintCode) + "'"); | |||
6878 | return; | |||
6879 | } | |||
6880 | ||||
6881 | // Add information to the INLINEASM node to know about this input. | |||
6882 | unsigned ResOpType = | |||
6883 | InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size()); | |||
6884 | AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, | |||
6885 | TLI.getPointerTy())); | |||
6886 | AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end()); | |||
6887 | break; | |||
6888 | } | |||
6889 | ||||
6890 | if (OpInfo.ConstraintType == TargetLowering::C_Memory) { | |||
6891 | assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!")((OpInfo.isIndirect && "Operand must be indirect to be a mem!" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.isIndirect && \"Operand must be indirect to be a mem!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6891, __PRETTY_FUNCTION__)); | |||
6892 | assert(InOperandVal.getValueType() == TLI.getPointerTy() &&((InOperandVal.getValueType() == TLI.getPointerTy() && "Memory operands expect pointer values") ? static_cast<void > (0) : __assert_fail ("InOperandVal.getValueType() == TLI.getPointerTy() && \"Memory operands expect pointer values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6893, __PRETTY_FUNCTION__)) | |||
6893 | "Memory operands expect pointer values")((InOperandVal.getValueType() == TLI.getPointerTy() && "Memory operands expect pointer values") ? static_cast<void > (0) : __assert_fail ("InOperandVal.getValueType() == TLI.getPointerTy() && \"Memory operands expect pointer values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6893, __PRETTY_FUNCTION__)); | |||
6894 | ||||
6895 | // Add information to the INLINEASM node to know about this input. | |||
6896 | unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); | |||
6897 | AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, | |||
6898 | TLI.getPointerTy())); | |||
6899 | AsmNodeOperands.push_back(InOperandVal); | |||
6900 | break; | |||
6901 | } | |||
6902 | ||||
6903 | assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||(((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && "Unknown constraint type!") ? static_cast<void> (0) : __assert_fail ("(OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && \"Unknown constraint type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6905, __PRETTY_FUNCTION__)) | |||
6904 | OpInfo.ConstraintType == TargetLowering::C_Register) &&(((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && "Unknown constraint type!") ? static_cast<void> (0) : __assert_fail ("(OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && \"Unknown constraint type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6905, __PRETTY_FUNCTION__)) | |||
6905 | "Unknown constraint type!")(((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && "Unknown constraint type!") ? static_cast<void> (0) : __assert_fail ("(OpInfo.ConstraintType == TargetLowering::C_RegisterClass || OpInfo.ConstraintType == TargetLowering::C_Register) && \"Unknown constraint type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6905, __PRETTY_FUNCTION__)); | |||
6906 | ||||
6907 | // TODO: Support this. | |||
6908 | if (OpInfo.isIndirect) { | |||
6909 | LLVMContext &Ctx = *DAG.getContext(); | |||
6910 | Ctx.emitError(CS.getInstruction(), | |||
6911 | "Don't know how to handle indirect register inputs yet " | |||
6912 | "for constraint '" + | |||
6913 | Twine(OpInfo.ConstraintCode) + "'"); | |||
6914 | return; | |||
6915 | } | |||
6916 | ||||
6917 | // Copy the input into the appropriate registers. | |||
6918 | if (OpInfo.AssignedRegs.Regs.empty()) { | |||
6919 | LLVMContext &Ctx = *DAG.getContext(); | |||
6920 | Ctx.emitError(CS.getInstruction(), | |||
6921 | "couldn't allocate input reg for constraint '" + | |||
6922 | Twine(OpInfo.ConstraintCode) + "'"); | |||
6923 | return; | |||
6924 | } | |||
6925 | ||||
6926 | OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(), | |||
6927 | Chain, &Flag, CS.getInstruction()); | |||
6928 | ||||
6929 | OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0, | |||
6930 | DAG, AsmNodeOperands); | |||
6931 | break; | |||
6932 | } | |||
6933 | case InlineAsm::isClobber: { | |||
6934 | // Add the clobbered value to the operand list, so that the register | |||
6935 | // allocator is aware that the physreg got clobbered. | |||
6936 | if (!OpInfo.AssignedRegs.Regs.empty()) | |||
6937 | OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber, | |||
6938 | false, 0, DAG, | |||
6939 | AsmNodeOperands); | |||
6940 | break; | |||
6941 | } | |||
6942 | } | |||
6943 | } | |||
6944 | ||||
6945 | // Finish up input operands. Set the input chain and add the flag last. | |||
6946 | AsmNodeOperands[InlineAsm::Op_InputChain] = Chain; | |||
6947 | if (Flag.getNode()) AsmNodeOperands.push_back(Flag); | |||
6948 | ||||
6949 | Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(), | |||
6950 | DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands); | |||
6951 | Flag = Chain.getValue(1); | |||
6952 | ||||
6953 | // If this asm returns a register value, copy the result from that register | |||
6954 | // and set it as the value of the call. | |||
6955 | if (!RetValRegs.Regs.empty()) { | |||
6956 | SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), | |||
6957 | Chain, &Flag, CS.getInstruction()); | |||
6958 | ||||
6959 | // FIXME: Why don't we do this for inline asms with MRVs? | |||
6960 | if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) { | |||
6961 | EVT ResultType = TLI.getValueType(CS.getType()); | |||
6962 | ||||
6963 | // If any of the results of the inline asm is a vector, it may have the | |||
6964 | // wrong width/num elts. This can happen for register classes that can | |||
6965 | // contain multiple different value types. The preg or vreg allocated may | |||
6966 | // not have the same VT as was expected. Convert it to the right type | |||
6967 | // with bit_convert. | |||
6968 | if (ResultType != Val.getValueType() && Val.getValueType().isVector()) { | |||
6969 | Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(), | |||
6970 | ResultType, Val); | |||
6971 | ||||
6972 | } else if (ResultType != Val.getValueType() && | |||
6973 | ResultType.isInteger() && Val.getValueType().isInteger()) { | |||
6974 | // If a result value was tied to an input value, the computed result may | |||
6975 | // have a wider width than the expected result. Extract the relevant | |||
6976 | // portion. | |||
6977 | Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val); | |||
6978 | } | |||
6979 | ||||
6980 | assert(ResultType == Val.getValueType() && "Asm result value mismatch!")((ResultType == Val.getValueType() && "Asm result value mismatch!" ) ? static_cast<void> (0) : __assert_fail ("ResultType == Val.getValueType() && \"Asm result value mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 6980, __PRETTY_FUNCTION__)); | |||
6981 | } | |||
6982 | ||||
6983 | setValue(CS.getInstruction(), Val); | |||
6984 | // Don't need to use this as a chain in this case. | |||
6985 | if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty()) | |||
6986 | return; | |||
6987 | } | |||
6988 | ||||
6989 | std::vector<std::pair<SDValue, const Value *> > StoresToEmit; | |||
6990 | ||||
6991 | // Process indirect outputs, first output all of the flagged copies out of | |||
6992 | // physregs. | |||
6993 | for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { | |||
6994 | RegsForValue &OutRegs = IndirectStoresToEmit[i].first; | |||
6995 | const Value *Ptr = IndirectStoresToEmit[i].second; | |||
6996 | SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), | |||
6997 | Chain, &Flag, IA); | |||
6998 | StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); | |||
6999 | } | |||
7000 | ||||
7001 | // Emit the non-flagged stores from the physregs. | |||
7002 | SmallVector<SDValue, 8> OutChains; | |||
7003 | for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) { | |||
7004 | SDValue Val = DAG.getStore(Chain, getCurSDLoc(), | |||
7005 | StoresToEmit[i].first, | |||
7006 | getValue(StoresToEmit[i].second), | |||
7007 | MachinePointerInfo(StoresToEmit[i].second), | |||
7008 | false, false, 0); | |||
7009 | OutChains.push_back(Val); | |||
7010 | } | |||
7011 | ||||
7012 | if (!OutChains.empty()) | |||
7013 | Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains); | |||
7014 | ||||
7015 | DAG.setRoot(Chain); | |||
7016 | } | |||
7017 | ||||
7018 | void SelectionDAGBuilder::visitVAStart(const CallInst &I) { | |||
7019 | DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(), | |||
7020 | MVT::Other, getRoot(), | |||
7021 | getValue(I.getArgOperand(0)), | |||
7022 | DAG.getSrcValue(I.getArgOperand(0)))); | |||
7023 | } | |||
7024 | ||||
7025 | void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) { | |||
7026 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
7027 | const DataLayout &DL = *TLI.getDataLayout(); | |||
7028 | SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurSDLoc(), | |||
7029 | getRoot(), getValue(I.getOperand(0)), | |||
7030 | DAG.getSrcValue(I.getOperand(0)), | |||
7031 | DL.getABITypeAlignment(I.getType())); | |||
7032 | setValue(&I, V); | |||
7033 | DAG.setRoot(V.getValue(1)); | |||
7034 | } | |||
7035 | ||||
7036 | void SelectionDAGBuilder::visitVAEnd(const CallInst &I) { | |||
7037 | DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(), | |||
7038 | MVT::Other, getRoot(), | |||
7039 | getValue(I.getArgOperand(0)), | |||
7040 | DAG.getSrcValue(I.getArgOperand(0)))); | |||
7041 | } | |||
7042 | ||||
7043 | void SelectionDAGBuilder::visitVACopy(const CallInst &I) { | |||
7044 | DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(), | |||
7045 | MVT::Other, getRoot(), | |||
7046 | getValue(I.getArgOperand(0)), | |||
7047 | getValue(I.getArgOperand(1)), | |||
7048 | DAG.getSrcValue(I.getArgOperand(0)), | |||
7049 | DAG.getSrcValue(I.getArgOperand(1)))); | |||
7050 | } | |||
7051 | ||||
7052 | /// \brief Lower an argument list according to the target calling convention. | |||
7053 | /// | |||
7054 | /// \return A tuple of <return-value, token-chain> | |||
7055 | /// | |||
7056 | /// This is a helper for lowering intrinsics that follow a target calling | |||
7057 | /// convention or require stack pointer adjustment. Only a subset of the | |||
7058 | /// intrinsic's operands need to participate in the calling convention. | |||
7059 | std::pair<SDValue, SDValue> | |||
7060 | SelectionDAGBuilder::lowerCallOperands(ImmutableCallSite CS, unsigned ArgIdx, | |||
7061 | unsigned NumArgs, SDValue Callee, | |||
7062 | bool UseVoidTy, | |||
7063 | MachineBasicBlock *LandingPad, | |||
7064 | bool IsPatchPoint) { | |||
7065 | TargetLowering::ArgListTy Args; | |||
7066 | Args.reserve(NumArgs); | |||
7067 | ||||
7068 | // Populate the argument list. | |||
7069 | // Attributes for args start at offset 1, after the return attribute. | |||
7070 | for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1; | |||
7071 | ArgI != ArgE; ++ArgI) { | |||
7072 | const Value *V = CS->getOperand(ArgI); | |||
7073 | ||||
7074 | assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.")((!V->getType()->isEmptyTy() && "Empty type passed to intrinsic." ) ? static_cast<void> (0) : __assert_fail ("!V->getType()->isEmptyTy() && \"Empty type passed to intrinsic.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7074, __PRETTY_FUNCTION__)); | |||
7075 | ||||
7076 | TargetLowering::ArgListEntry Entry; | |||
7077 | Entry.Node = getValue(V); | |||
7078 | Entry.Ty = V->getType(); | |||
7079 | Entry.setAttributes(&CS, AttrI); | |||
7080 | Args.push_back(Entry); | |||
7081 | } | |||
7082 | ||||
7083 | Type *retTy = UseVoidTy ? Type::getVoidTy(*DAG.getContext()) : CS->getType(); | |||
7084 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
7085 | CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot()) | |||
7086 | .setCallee(CS.getCallingConv(), retTy, Callee, std::move(Args), NumArgs) | |||
7087 | .setDiscardResult(CS->use_empty()).setIsPatchPoint(IsPatchPoint); | |||
7088 | ||||
7089 | return lowerInvokable(CLI, LandingPad); | |||
7090 | } | |||
7091 | ||||
7092 | /// \brief Add a stack map intrinsic call's live variable operands to a stackmap | |||
7093 | /// or patchpoint target node's operand list. | |||
7094 | /// | |||
7095 | /// Constants are converted to TargetConstants purely as an optimization to | |||
7096 | /// avoid constant materialization and register allocation. | |||
7097 | /// | |||
7098 | /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not | |||
7099 | /// generate addess computation nodes, and so ExpandISelPseudo can convert the | |||
7100 | /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids | |||
7101 | /// address materialization and register allocation, but may also be required | |||
7102 | /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an | |||
7103 | /// alloca in the entry block, then the runtime may assume that the alloca's | |||
7104 | /// StackMap location can be read immediately after compilation and that the | |||
7105 | /// location is valid at any point during execution (this is similar to the | |||
7106 | /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were | |||
7107 | /// only available in a register, then the runtime would need to trap when | |||
7108 | /// execution reaches the StackMap in order to read the alloca's location. | |||
7109 | static void addStackMapLiveVars(ImmutableCallSite CS, unsigned StartIdx, | |||
7110 | SmallVectorImpl<SDValue> &Ops, | |||
7111 | SelectionDAGBuilder &Builder) { | |||
7112 | for (unsigned i = StartIdx, e = CS.arg_size(); i != e; ++i) { | |||
7113 | SDValue OpVal = Builder.getValue(CS.getArgument(i)); | |||
7114 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) { | |||
7115 | Ops.push_back( | |||
7116 | Builder.DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64)); | |||
7117 | Ops.push_back( | |||
7118 | Builder.DAG.getTargetConstant(C->getSExtValue(), MVT::i64)); | |||
7119 | } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) { | |||
7120 | const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo(); | |||
7121 | Ops.push_back( | |||
7122 | Builder.DAG.getTargetFrameIndex(FI->getIndex(), TLI.getPointerTy())); | |||
7123 | } else | |||
7124 | Ops.push_back(OpVal); | |||
7125 | } | |||
7126 | } | |||
7127 | ||||
7128 | /// \brief Lower llvm.experimental.stackmap directly to its target opcode. | |||
7129 | void SelectionDAGBuilder::visitStackmap(const CallInst &CI) { | |||
7130 | // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>, | |||
7131 | // [live variables...]) | |||
7132 | ||||
7133 | assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.")((CI.getType()->isVoidTy() && "Stackmap cannot return a value." ) ? static_cast<void> (0) : __assert_fail ("CI.getType()->isVoidTy() && \"Stackmap cannot return a value.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7133, __PRETTY_FUNCTION__)); | |||
7134 | ||||
7135 | SDValue Chain, InFlag, Callee, NullPtr; | |||
7136 | SmallVector<SDValue, 32> Ops; | |||
7137 | ||||
7138 | SDLoc DL = getCurSDLoc(); | |||
7139 | Callee = getValue(CI.getCalledValue()); | |||
7140 | NullPtr = DAG.getIntPtrConstant(0, true); | |||
7141 | ||||
7142 | // The stackmap intrinsic only records the live variables (the arguemnts | |||
7143 | // passed to it) and emits NOPS (if requested). Unlike the patchpoint | |||
7144 | // intrinsic, this won't be lowered to a function call. This means we don't | |||
7145 | // have to worry about calling conventions and target specific lowering code. | |||
7146 | // Instead we perform the call lowering right here. | |||
7147 | // | |||
7148 | // chain, flag = CALLSEQ_START(chain, 0) | |||
7149 | // chain, flag = STACKMAP(id, nbytes, ..., chain, flag) | |||
7150 | // chain, flag = CALLSEQ_END(chain, 0, 0, flag) | |||
7151 | // | |||
7152 | Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL); | |||
7153 | InFlag = Chain.getValue(1); | |||
7154 | ||||
7155 | // Add the <id> and <numBytes> constants. | |||
7156 | SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos)); | |||
7157 | Ops.push_back(DAG.getTargetConstant( | |||
7158 | cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64)); | |||
7159 | SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos)); | |||
7160 | Ops.push_back(DAG.getTargetConstant( | |||
7161 | cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32)); | |||
7162 | ||||
7163 | // Push live variables for the stack map. | |||
7164 | addStackMapLiveVars(&CI, 2, Ops, *this); | |||
7165 | ||||
7166 | // We are not pushing any register mask info here on the operands list, | |||
7167 | // because the stackmap doesn't clobber anything. | |||
7168 | ||||
7169 | // Push the chain and the glue flag. | |||
7170 | Ops.push_back(Chain); | |||
7171 | Ops.push_back(InFlag); | |||
7172 | ||||
7173 | // Create the STACKMAP node. | |||
7174 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
7175 | SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops); | |||
7176 | Chain = SDValue(SM, 0); | |||
7177 | InFlag = Chain.getValue(1); | |||
7178 | ||||
7179 | Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL); | |||
7180 | ||||
7181 | // Stackmaps don't generate values, so nothing goes into the NodeMap. | |||
7182 | ||||
7183 | // Set the root to the target-lowered call chain. | |||
7184 | DAG.setRoot(Chain); | |||
7185 | ||||
7186 | // Inform the Frame Information that we have a stackmap in this function. | |||
7187 | FuncInfo.MF->getFrameInfo()->setHasStackMap(); | |||
7188 | } | |||
7189 | ||||
7190 | /// \brief Lower llvm.experimental.patchpoint directly to its target opcode. | |||
7191 | void SelectionDAGBuilder::visitPatchpoint(ImmutableCallSite CS, | |||
7192 | MachineBasicBlock *LandingPad) { | |||
7193 | // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>, | |||
7194 | // i32 <numBytes>, | |||
7195 | // i8* <target>, | |||
7196 | // i32 <numArgs>, | |||
7197 | // [Args...], | |||
7198 | // [live variables...]) | |||
7199 | ||||
7200 | CallingConv::ID CC = CS.getCallingConv(); | |||
7201 | bool IsAnyRegCC = CC == CallingConv::AnyReg; | |||
7202 | bool HasDef = !CS->getType()->isVoidTy(); | |||
7203 | SDValue Callee = getValue(CS->getOperand(2)); // <target> | |||
7204 | ||||
7205 | // Get the real number of arguments participating in the call <numArgs> | |||
7206 | SDValue NArgVal = getValue(CS.getArgument(PatchPointOpers::NArgPos)); | |||
7207 | unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue(); | |||
7208 | ||||
7209 | // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs> | |||
7210 | // Intrinsics include all meta-operands up to but not including CC. | |||
7211 | unsigned NumMetaOpers = PatchPointOpers::CCPos; | |||
7212 | assert(CS.arg_size() >= NumMetaOpers + NumArgs &&((CS.arg_size() >= NumMetaOpers + NumArgs && "Not enough arguments provided to the patchpoint intrinsic" ) ? static_cast<void> (0) : __assert_fail ("CS.arg_size() >= NumMetaOpers + NumArgs && \"Not enough arguments provided to the patchpoint intrinsic\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7213, __PRETTY_FUNCTION__)) | |||
7213 | "Not enough arguments provided to the patchpoint intrinsic")((CS.arg_size() >= NumMetaOpers + NumArgs && "Not enough arguments provided to the patchpoint intrinsic" ) ? static_cast<void> (0) : __assert_fail ("CS.arg_size() >= NumMetaOpers + NumArgs && \"Not enough arguments provided to the patchpoint intrinsic\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7213, __PRETTY_FUNCTION__)); | |||
7214 | ||||
7215 | // For AnyRegCC the arguments are lowered later on manually. | |||
7216 | unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs; | |||
7217 | std::pair<SDValue, SDValue> Result = | |||
7218 | lowerCallOperands(CS, NumMetaOpers, NumCallArgs, Callee, IsAnyRegCC, | |||
7219 | LandingPad, true); | |||
7220 | ||||
7221 | SDNode *CallEnd = Result.second.getNode(); | |||
7222 | if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg)) | |||
7223 | CallEnd = CallEnd->getOperand(0).getNode(); | |||
7224 | ||||
7225 | /// Get a call instruction from the call sequence chain. | |||
7226 | /// Tail calls are not allowed. | |||
7227 | assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&((CallEnd->getOpcode() == ISD::CALLSEQ_END && "Expected a callseq node." ) ? static_cast<void> (0) : __assert_fail ("CallEnd->getOpcode() == ISD::CALLSEQ_END && \"Expected a callseq node.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7228, __PRETTY_FUNCTION__)) | |||
7228 | "Expected a callseq node.")((CallEnd->getOpcode() == ISD::CALLSEQ_END && "Expected a callseq node." ) ? static_cast<void> (0) : __assert_fail ("CallEnd->getOpcode() == ISD::CALLSEQ_END && \"Expected a callseq node.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7228, __PRETTY_FUNCTION__)); | |||
7229 | SDNode *Call = CallEnd->getOperand(0).getNode(); | |||
7230 | bool HasGlue = Call->getGluedNode(); | |||
7231 | ||||
7232 | // Replace the target specific call node with the patchable intrinsic. | |||
7233 | SmallVector<SDValue, 8> Ops; | |||
7234 | ||||
7235 | // Add the <id> and <numBytes> constants. | |||
7236 | SDValue IDVal = getValue(CS->getOperand(PatchPointOpers::IDPos)); | |||
7237 | Ops.push_back(DAG.getTargetConstant( | |||
7238 | cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64)); | |||
7239 | SDValue NBytesVal = getValue(CS->getOperand(PatchPointOpers::NBytesPos)); | |||
7240 | Ops.push_back(DAG.getTargetConstant( | |||
7241 | cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32)); | |||
7242 | ||||
7243 | // Assume that the Callee is a constant address. | |||
7244 | // FIXME: handle function symbols in the future. | |||
7245 | Ops.push_back( | |||
7246 | DAG.getIntPtrConstant(cast<ConstantSDNode>(Callee)->getZExtValue(), | |||
7247 | /*isTarget=*/true)); | |||
7248 | ||||
7249 | // Adjust <numArgs> to account for any arguments that have been passed on the | |||
7250 | // stack instead. | |||
7251 | // Call Node: Chain, Target, {Args}, RegMask, [Glue] | |||
7252 | unsigned NumCallRegArgs = Call->getNumOperands() - (HasGlue ? 4 : 3); | |||
7253 | NumCallRegArgs = IsAnyRegCC ? NumArgs : NumCallRegArgs; | |||
7254 | Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, MVT::i32)); | |||
7255 | ||||
7256 | // Add the calling convention | |||
7257 | Ops.push_back(DAG.getTargetConstant((unsigned)CC, MVT::i32)); | |||
7258 | ||||
7259 | // Add the arguments we omitted previously. The register allocator should | |||
7260 | // place these in any free register. | |||
7261 | if (IsAnyRegCC) | |||
7262 | for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i) | |||
7263 | Ops.push_back(getValue(CS.getArgument(i))); | |||
7264 | ||||
7265 | // Push the arguments from the call instruction up to the register mask. | |||
7266 | SDNode::op_iterator e = HasGlue ? Call->op_end()-2 : Call->op_end()-1; | |||
7267 | Ops.append(Call->op_begin() + 2, e); | |||
7268 | ||||
7269 | // Push live variables for the stack map. | |||
7270 | addStackMapLiveVars(CS, NumMetaOpers + NumArgs, Ops, *this); | |||
7271 | ||||
7272 | // Push the register mask info. | |||
7273 | if (HasGlue) | |||
7274 | Ops.push_back(*(Call->op_end()-2)); | |||
7275 | else | |||
7276 | Ops.push_back(*(Call->op_end()-1)); | |||
7277 | ||||
7278 | // Push the chain (this is originally the first operand of the call, but | |||
7279 | // becomes now the last or second to last operand). | |||
7280 | Ops.push_back(*(Call->op_begin())); | |||
7281 | ||||
7282 | // Push the glue flag (last operand). | |||
7283 | if (HasGlue) | |||
7284 | Ops.push_back(*(Call->op_end()-1)); | |||
7285 | ||||
7286 | SDVTList NodeTys; | |||
7287 | if (IsAnyRegCC && HasDef) { | |||
7288 | // Create the return types based on the intrinsic definition | |||
7289 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
7290 | SmallVector<EVT, 3> ValueVTs; | |||
7291 | ComputeValueVTs(TLI, CS->getType(), ValueVTs); | |||
7292 | assert(ValueVTs.size() == 1 && "Expected only one return value type.")((ValueVTs.size() == 1 && "Expected only one return value type." ) ? static_cast<void> (0) : __assert_fail ("ValueVTs.size() == 1 && \"Expected only one return value type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7292, __PRETTY_FUNCTION__)); | |||
7293 | ||||
7294 | // There is always a chain and a glue type at the end | |||
7295 | ValueVTs.push_back(MVT::Other); | |||
7296 | ValueVTs.push_back(MVT::Glue); | |||
7297 | NodeTys = DAG.getVTList(ValueVTs); | |||
7298 | } else | |||
7299 | NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
7300 | ||||
7301 | // Replace the target specific call node with a PATCHPOINT node. | |||
7302 | MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT, | |||
7303 | getCurSDLoc(), NodeTys, Ops); | |||
7304 | ||||
7305 | // Update the NodeMap. | |||
7306 | if (HasDef) { | |||
7307 | if (IsAnyRegCC) | |||
7308 | setValue(CS.getInstruction(), SDValue(MN, 0)); | |||
7309 | else | |||
7310 | setValue(CS.getInstruction(), Result.first); | |||
7311 | } | |||
7312 | ||||
7313 | // Fixup the consumers of the intrinsic. The chain and glue may be used in the | |||
7314 | // call sequence. Furthermore the location of the chain and glue can change | |||
7315 | // when the AnyReg calling convention is used and the intrinsic returns a | |||
7316 | // value. | |||
7317 | if (IsAnyRegCC && HasDef) { | |||
7318 | SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)}; | |||
7319 | SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)}; | |||
7320 | DAG.ReplaceAllUsesOfValuesWith(From, To, 2); | |||
7321 | } else | |||
7322 | DAG.ReplaceAllUsesWith(Call, MN); | |||
7323 | DAG.DeleteNode(Call); | |||
7324 | ||||
7325 | // Inform the Frame Information that we have a patchpoint in this function. | |||
7326 | FuncInfo.MF->getFrameInfo()->setHasPatchPoint(); | |||
7327 | } | |||
7328 | ||||
7329 | /// Returns an AttributeSet representing the attributes applied to the return | |||
7330 | /// value of the given call. | |||
7331 | static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) { | |||
7332 | SmallVector<Attribute::AttrKind, 2> Attrs; | |||
7333 | if (CLI.RetSExt) | |||
7334 | Attrs.push_back(Attribute::SExt); | |||
7335 | if (CLI.RetZExt) | |||
7336 | Attrs.push_back(Attribute::ZExt); | |||
7337 | if (CLI.IsInReg) | |||
7338 | Attrs.push_back(Attribute::InReg); | |||
7339 | ||||
7340 | return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex, | |||
7341 | Attrs); | |||
7342 | } | |||
7343 | ||||
7344 | /// TargetLowering::LowerCallTo - This is the default LowerCallTo | |||
7345 | /// implementation, which just calls LowerCall. | |||
7346 | /// FIXME: When all targets are | |||
7347 | /// migrated to using LowerCall, this hook should be integrated into SDISel. | |||
7348 | std::pair<SDValue, SDValue> | |||
7349 | TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const { | |||
7350 | // Handle the incoming return values from the call. | |||
7351 | CLI.Ins.clear(); | |||
7352 | Type *OrigRetTy = CLI.RetTy; | |||
7353 | SmallVector<EVT, 4> RetTys; | |||
7354 | SmallVector<uint64_t, 4> Offsets; | |||
7355 | ComputeValueVTs(*this, CLI.RetTy, RetTys, &Offsets); | |||
7356 | ||||
7357 | SmallVector<ISD::OutputArg, 4> Outs; | |||
7358 | GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this); | |||
7359 | ||||
7360 | bool CanLowerReturn = | |||
7361 | this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(), | |||
7362 | CLI.IsVarArg, Outs, CLI.RetTy->getContext()); | |||
7363 | ||||
7364 | SDValue DemoteStackSlot; | |||
7365 | int DemoteStackIdx = -100; | |||
7366 | if (!CanLowerReturn) { | |||
7367 | // FIXME: equivalent assert? | |||
7368 | // assert(!CS.hasInAllocaArgument() && | |||
7369 | // "sret demotion is incompatible with inalloca"); | |||
7370 | uint64_t TySize = getDataLayout()->getTypeAllocSize(CLI.RetTy); | |||
7371 | unsigned Align = getDataLayout()->getPrefTypeAlignment(CLI.RetTy); | |||
7372 | MachineFunction &MF = CLI.DAG.getMachineFunction(); | |||
7373 | DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); | |||
7374 | Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy); | |||
7375 | ||||
7376 | DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy()); | |||
7377 | ArgListEntry Entry; | |||
7378 | Entry.Node = DemoteStackSlot; | |||
7379 | Entry.Ty = StackSlotPtrType; | |||
7380 | Entry.isSExt = false; | |||
7381 | Entry.isZExt = false; | |||
7382 | Entry.isInReg = false; | |||
7383 | Entry.isSRet = true; | |||
7384 | Entry.isNest = false; | |||
7385 | Entry.isByVal = false; | |||
7386 | Entry.isReturned = false; | |||
7387 | Entry.Alignment = Align; | |||
7388 | CLI.getArgs().insert(CLI.getArgs().begin(), Entry); | |||
7389 | CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext()); | |||
7390 | } else { | |||
7391 | for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { | |||
7392 | EVT VT = RetTys[I]; | |||
7393 | MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT); | |||
7394 | unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT); | |||
7395 | for (unsigned i = 0; i != NumRegs; ++i) { | |||
7396 | ISD::InputArg MyFlags; | |||
7397 | MyFlags.VT = RegisterVT; | |||
7398 | MyFlags.ArgVT = VT; | |||
7399 | MyFlags.Used = CLI.IsReturnValueUsed; | |||
7400 | if (CLI.RetSExt) | |||
7401 | MyFlags.Flags.setSExt(); | |||
7402 | if (CLI.RetZExt) | |||
7403 | MyFlags.Flags.setZExt(); | |||
7404 | if (CLI.IsInReg) | |||
7405 | MyFlags.Flags.setInReg(); | |||
7406 | CLI.Ins.push_back(MyFlags); | |||
7407 | } | |||
7408 | } | |||
7409 | } | |||
7410 | ||||
7411 | // Handle all of the outgoing arguments. | |||
7412 | CLI.Outs.clear(); | |||
7413 | CLI.OutVals.clear(); | |||
7414 | ArgListTy &Args = CLI.getArgs(); | |||
7415 | for (unsigned i = 0, e = Args.size(); i != e; ++i) { | |||
7416 | SmallVector<EVT, 4> ValueVTs; | |||
7417 | ComputeValueVTs(*this, Args[i].Ty, ValueVTs); | |||
7418 | Type *FinalType = Args[i].Ty; | |||
7419 | if (Args[i].isByVal) | |||
7420 | FinalType = cast<PointerType>(Args[i].Ty)->getElementType(); | |||
7421 | bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters( | |||
7422 | FinalType, CLI.CallConv, CLI.IsVarArg); | |||
7423 | for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues; | |||
7424 | ++Value) { | |||
7425 | EVT VT = ValueVTs[Value]; | |||
7426 | Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext()); | |||
7427 | SDValue Op = SDValue(Args[i].Node.getNode(), | |||
7428 | Args[i].Node.getResNo() + Value); | |||
7429 | ISD::ArgFlagsTy Flags; | |||
7430 | unsigned OriginalAlignment = getDataLayout()->getABITypeAlignment(ArgTy); | |||
7431 | ||||
7432 | if (Args[i].isZExt) | |||
7433 | Flags.setZExt(); | |||
7434 | if (Args[i].isSExt) | |||
7435 | Flags.setSExt(); | |||
7436 | if (Args[i].isInReg) | |||
7437 | Flags.setInReg(); | |||
7438 | if (Args[i].isSRet) | |||
7439 | Flags.setSRet(); | |||
7440 | if (Args[i].isByVal) | |||
7441 | Flags.setByVal(); | |||
7442 | if (Args[i].isInAlloca) { | |||
7443 | Flags.setInAlloca(); | |||
7444 | // Set the byval flag for CCAssignFn callbacks that don't know about | |||
7445 | // inalloca. This way we can know how many bytes we should've allocated | |||
7446 | // and how many bytes a callee cleanup function will pop. If we port | |||
7447 | // inalloca to more targets, we'll have to add custom inalloca handling | |||
7448 | // in the various CC lowering callbacks. | |||
7449 | Flags.setByVal(); | |||
7450 | } | |||
7451 | if (Args[i].isByVal || Args[i].isInAlloca) { | |||
7452 | PointerType *Ty = cast<PointerType>(Args[i].Ty); | |||
7453 | Type *ElementTy = Ty->getElementType(); | |||
7454 | Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy)); | |||
7455 | // For ByVal, alignment should come from FE. BE will guess if this | |||
7456 | // info is not there but there are cases it cannot get right. | |||
7457 | unsigned FrameAlign; | |||
7458 | if (Args[i].Alignment) | |||
7459 | FrameAlign = Args[i].Alignment; | |||
7460 | else | |||
7461 | FrameAlign = getByValTypeAlignment(ElementTy); | |||
7462 | Flags.setByValAlign(FrameAlign); | |||
7463 | } | |||
7464 | if (Args[i].isNest) | |||
7465 | Flags.setNest(); | |||
7466 | if (NeedsRegBlock) | |||
7467 | Flags.setInConsecutiveRegs(); | |||
7468 | Flags.setOrigAlign(OriginalAlignment); | |||
7469 | ||||
7470 | MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT); | |||
7471 | unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT); | |||
7472 | SmallVector<SDValue, 4> Parts(NumParts); | |||
7473 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND; | |||
7474 | ||||
7475 | if (Args[i].isSExt) | |||
7476 | ExtendKind = ISD::SIGN_EXTEND; | |||
7477 | else if (Args[i].isZExt) | |||
7478 | ExtendKind = ISD::ZERO_EXTEND; | |||
7479 | ||||
7480 | // Conservatively only handle 'returned' on non-vectors for now | |||
7481 | if (Args[i].isReturned && !Op.getValueType().isVector()) { | |||
7482 | assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&((CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && "unexpected use of 'returned'") ? static_cast< void> (0) : __assert_fail ("CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && \"unexpected use of 'returned'\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7483, __PRETTY_FUNCTION__)) | |||
7483 | "unexpected use of 'returned'")((CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && "unexpected use of 'returned'") ? static_cast< void> (0) : __assert_fail ("CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && \"unexpected use of 'returned'\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7483, __PRETTY_FUNCTION__)); | |||
7484 | // Before passing 'returned' to the target lowering code, ensure that | |||
7485 | // either the register MVT and the actual EVT are the same size or that | |||
7486 | // the return value and argument are extended in the same way; in these | |||
7487 | // cases it's safe to pass the argument register value unchanged as the | |||
7488 | // return register value (although it's at the target's option whether | |||
7489 | // to do so) | |||
7490 | // TODO: allow code generation to take advantage of partially preserved | |||
7491 | // registers rather than clobbering the entire register when the | |||
7492 | // parameter extension method is not compatible with the return | |||
7493 | // extension method | |||
7494 | if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) || | |||
7495 | (ExtendKind != ISD::ANY_EXTEND && | |||
7496 | CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt)) | |||
7497 | Flags.setReturned(); | |||
7498 | } | |||
7499 | ||||
7500 | getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT, | |||
7501 | CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind); | |||
7502 | ||||
7503 | for (unsigned j = 0; j != NumParts; ++j) { | |||
7504 | // if it isn't first piece, alignment must be 1 | |||
7505 | ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT, | |||
7506 | i < CLI.NumFixedArgs, | |||
7507 | i, j*Parts[j].getValueType().getStoreSize()); | |||
7508 | if (NumParts > 1 && j == 0) | |||
7509 | MyFlags.Flags.setSplit(); | |||
7510 | else if (j != 0) | |||
7511 | MyFlags.Flags.setOrigAlign(1); | |||
7512 | ||||
7513 | CLI.Outs.push_back(MyFlags); | |||
7514 | CLI.OutVals.push_back(Parts[j]); | |||
7515 | } | |||
7516 | ||||
7517 | if (NeedsRegBlock && Value == NumValues - 1) | |||
7518 | CLI.Outs[CLI.Outs.size() - 1].Flags.setInConsecutiveRegsLast(); | |||
7519 | } | |||
7520 | } | |||
7521 | ||||
7522 | SmallVector<SDValue, 4> InVals; | |||
7523 | CLI.Chain = LowerCall(CLI, InVals); | |||
7524 | ||||
7525 | // Verify that the target's LowerCall behaved as expected. | |||
7526 | assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&((CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT ::Other && "LowerCall didn't return a valid chain!") ? static_cast<void> (0) : __assert_fail ("CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other && \"LowerCall didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7527, __PRETTY_FUNCTION__)) | |||
7527 | "LowerCall didn't return a valid chain!")((CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT ::Other && "LowerCall didn't return a valid chain!") ? static_cast<void> (0) : __assert_fail ("CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other && \"LowerCall didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7527, __PRETTY_FUNCTION__)); | |||
7528 | assert((!CLI.IsTailCall || InVals.empty()) &&(((!CLI.IsTailCall || InVals.empty()) && "LowerCall emitted a return value for a tail call!" ) ? static_cast<void> (0) : __assert_fail ("(!CLI.IsTailCall || InVals.empty()) && \"LowerCall emitted a return value for a tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7529, __PRETTY_FUNCTION__)) | |||
7529 | "LowerCall emitted a return value for a tail call!")(((!CLI.IsTailCall || InVals.empty()) && "LowerCall emitted a return value for a tail call!" ) ? static_cast<void> (0) : __assert_fail ("(!CLI.IsTailCall || InVals.empty()) && \"LowerCall emitted a return value for a tail call!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7529, __PRETTY_FUNCTION__)); | |||
7530 | assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&(((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && "LowerCall didn't emit the correct number of values!") ? static_cast <void> (0) : __assert_fail ("(CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && \"LowerCall didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7531, __PRETTY_FUNCTION__)) | |||
7531 | "LowerCall didn't emit the correct number of values!")(((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && "LowerCall didn't emit the correct number of values!") ? static_cast <void> (0) : __assert_fail ("(CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && \"LowerCall didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7531, __PRETTY_FUNCTION__)); | |||
7532 | ||||
7533 | // For a tail call, the return value is merely live-out and there aren't | |||
7534 | // any nodes in the DAG representing it. Return a special value to | |||
7535 | // indicate that a tail call has been emitted and no more Instructions | |||
7536 | // should be processed in the current block. | |||
7537 | if (CLI.IsTailCall) { | |||
7538 | CLI.DAG.setRoot(CLI.Chain); | |||
7539 | return std::make_pair(SDValue(), SDValue()); | |||
7540 | } | |||
7541 | ||||
7542 | DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7544, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7546, __PRETTY_FUNCTION__)); }; } } while (0) | |||
7543 | assert(InVals[i].getNode() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7544, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7546, __PRETTY_FUNCTION__)); }; } } while (0) | |||
7544 | "LowerCall emitted a null value!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7544, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7546, __PRETTY_FUNCTION__)); }; } } while (0) | |||
7545 | assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7544, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7546, __PRETTY_FUNCTION__)); }; } } while (0) | |||
7546 | "LowerCall emitted a value with the wrong type!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7544, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7546, __PRETTY_FUNCTION__)); }; } } while (0) | |||
7547 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerCall emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerCall emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7544, __PRETTY_FUNCTION__)); ((EVT(CLI.Ins[i].VT) == InVals [i].getValueType() && "LowerCall emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && \"LowerCall emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7546, __PRETTY_FUNCTION__)); }; } } while (0); | |||
7548 | ||||
7549 | SmallVector<SDValue, 4> ReturnValues; | |||
7550 | if (!CanLowerReturn) { | |||
7551 | // The instruction result is the result of loading from the | |||
7552 | // hidden sret parameter. | |||
7553 | SmallVector<EVT, 1> PVTs; | |||
7554 | Type *PtrRetTy = PointerType::getUnqual(OrigRetTy); | |||
7555 | ||||
7556 | ComputeValueVTs(*this, PtrRetTy, PVTs); | |||
7557 | assert(PVTs.size() == 1 && "Pointers should fit in one register")((PVTs.size() == 1 && "Pointers should fit in one register" ) ? static_cast<void> (0) : __assert_fail ("PVTs.size() == 1 && \"Pointers should fit in one register\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7557, __PRETTY_FUNCTION__)); | |||
7558 | EVT PtrVT = PVTs[0]; | |||
7559 | ||||
7560 | unsigned NumValues = RetTys.size(); | |||
7561 | ReturnValues.resize(NumValues); | |||
7562 | SmallVector<SDValue, 4> Chains(NumValues); | |||
7563 | ||||
7564 | for (unsigned i = 0; i < NumValues; ++i) { | |||
7565 | SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot, | |||
7566 | CLI.DAG.getConstant(Offsets[i], PtrVT)); | |||
7567 | SDValue L = CLI.DAG.getLoad( | |||
7568 | RetTys[i], CLI.DL, CLI.Chain, Add, | |||
7569 | MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), false, | |||
7570 | false, false, 1); | |||
7571 | ReturnValues[i] = L; | |||
7572 | Chains[i] = L.getValue(1); | |||
7573 | } | |||
7574 | ||||
7575 | CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains); | |||
7576 | } else { | |||
7577 | // Collect the legal value parts into potentially illegal values | |||
7578 | // that correspond to the original function's return values. | |||
7579 | ISD::NodeType AssertOp = ISD::DELETED_NODE; | |||
7580 | if (CLI.RetSExt) | |||
7581 | AssertOp = ISD::AssertSext; | |||
7582 | else if (CLI.RetZExt) | |||
7583 | AssertOp = ISD::AssertZext; | |||
7584 | unsigned CurReg = 0; | |||
7585 | for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { | |||
7586 | EVT VT = RetTys[I]; | |||
7587 | MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT); | |||
7588 | unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT); | |||
7589 | ||||
7590 | ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg], | |||
7591 | NumRegs, RegisterVT, VT, nullptr, | |||
7592 | AssertOp)); | |||
7593 | CurReg += NumRegs; | |||
7594 | } | |||
7595 | ||||
7596 | // For a function returning void, there is no return value. We can't create | |||
7597 | // such a node, so we just return a null return value in that case. In | |||
7598 | // that case, nothing will actually look at the value. | |||
7599 | if (ReturnValues.empty()) | |||
7600 | return std::make_pair(SDValue(), CLI.Chain); | |||
7601 | } | |||
7602 | ||||
7603 | SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL, | |||
7604 | CLI.DAG.getVTList(RetTys), ReturnValues); | |||
7605 | return std::make_pair(Res, CLI.Chain); | |||
7606 | } | |||
7607 | ||||
7608 | void TargetLowering::LowerOperationWrapper(SDNode *N, | |||
7609 | SmallVectorImpl<SDValue> &Results, | |||
7610 | SelectionDAG &DAG) const { | |||
7611 | SDValue Res = LowerOperation(SDValue(N, 0), DAG); | |||
7612 | if (Res.getNode()) | |||
7613 | Results.push_back(Res); | |||
7614 | } | |||
7615 | ||||
7616 | SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { | |||
7617 | llvm_unreachable("LowerOperation not implemented for this target!")::llvm::llvm_unreachable_internal("LowerOperation not implemented for this target!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7617); | |||
7618 | } | |||
7619 | ||||
7620 | void | |||
7621 | SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) { | |||
7622 | SDValue Op = getNonRegisterValue(V); | |||
7623 | assert((Op.getOpcode() != ISD::CopyFromReg ||(((Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode >(Op.getOperand(1))->getReg() != Reg) && "Copy from a reg to the same reg!" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && \"Copy from a reg to the same reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7625, __PRETTY_FUNCTION__)) | |||
7624 | cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&(((Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode >(Op.getOperand(1))->getReg() != Reg) && "Copy from a reg to the same reg!" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && \"Copy from a reg to the same reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7625, __PRETTY_FUNCTION__)) | |||
7625 | "Copy from a reg to the same reg!")(((Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode >(Op.getOperand(1))->getReg() != Reg) && "Copy from a reg to the same reg!" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::CopyFromReg || cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && \"Copy from a reg to the same reg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7625, __PRETTY_FUNCTION__)); | |||
7626 | assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg")((!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg" ) ? static_cast<void> (0) : __assert_fail ("!TargetRegisterInfo::isPhysicalRegister(Reg) && \"Is a physreg\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7626, __PRETTY_FUNCTION__)); | |||
7627 | ||||
7628 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
7629 | RegsForValue RFV(V->getContext(), TLI, Reg, V->getType()); | |||
7630 | SDValue Chain = DAG.getEntryNode(); | |||
7631 | ||||
7632 | ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) == | |||
7633 | FuncInfo.PreferredExtendType.end()) | |||
7634 | ? ISD::ANY_EXTEND | |||
7635 | : FuncInfo.PreferredExtendType[V]; | |||
7636 | RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType); | |||
7637 | PendingExports.push_back(Chain); | |||
7638 | } | |||
7639 | ||||
7640 | #include "llvm/CodeGen/SelectionDAGISel.h" | |||
7641 | ||||
7642 | /// isOnlyUsedInEntryBlock - If the specified argument is only used in the | |||
7643 | /// entry block, return true. This includes arguments used by switches, since | |||
7644 | /// the switch may expand into multiple basic blocks. | |||
7645 | static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) { | |||
7646 | // With FastISel active, we may be splitting blocks, so force creation | |||
7647 | // of virtual registers for all non-dead arguments. | |||
7648 | if (FastISel) | |||
7649 | return A->use_empty(); | |||
7650 | ||||
7651 | const BasicBlock *Entry = A->getParent()->begin(); | |||
7652 | for (const User *U : A->users()) | |||
7653 | if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U)) | |||
7654 | return false; // Use not in entry block. | |||
7655 | ||||
7656 | return true; | |||
7657 | } | |||
7658 | ||||
7659 | void SelectionDAGISel::LowerArguments(const Function &F) { | |||
7660 | SelectionDAG &DAG = SDB->DAG; | |||
7661 | SDLoc dl = SDB->getCurSDLoc(); | |||
7662 | const DataLayout *DL = TLI->getDataLayout(); | |||
7663 | SmallVector<ISD::InputArg, 16> Ins; | |||
7664 | ||||
7665 | if (!FuncInfo->CanLowerReturn) { | |||
7666 | // Put in an sret pointer parameter before all the other parameters. | |||
7667 | SmallVector<EVT, 1> ValueVTs; | |||
7668 | ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); | |||
7669 | ||||
7670 | // NOTE: Assuming that a pointer will never break down to more than one VT | |||
7671 | // or one register. | |||
7672 | ISD::ArgFlagsTy Flags; | |||
7673 | Flags.setSRet(); | |||
7674 | MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]); | |||
7675 | ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true, | |||
7676 | ISD::InputArg::NoArgIndex, 0); | |||
7677 | Ins.push_back(RetArg); | |||
7678 | } | |||
7679 | ||||
7680 | // Set up the incoming argument description vector. | |||
7681 | unsigned Idx = 1; | |||
7682 | for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); | |||
7683 | I != E; ++I, ++Idx) { | |||
7684 | SmallVector<EVT, 4> ValueVTs; | |||
7685 | ComputeValueVTs(*TLI, I->getType(), ValueVTs); | |||
7686 | bool isArgValueUsed = !I->use_empty(); | |||
7687 | unsigned PartBase = 0; | |||
7688 | Type *FinalType = I->getType(); | |||
7689 | if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) | |||
7690 | FinalType = cast<PointerType>(FinalType)->getElementType(); | |||
7691 | bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters( | |||
7692 | FinalType, F.getCallingConv(), F.isVarArg()); | |||
7693 | for (unsigned Value = 0, NumValues = ValueVTs.size(); | |||
7694 | Value != NumValues; ++Value) { | |||
7695 | EVT VT = ValueVTs[Value]; | |||
7696 | Type *ArgTy = VT.getTypeForEVT(*DAG.getContext()); | |||
7697 | ISD::ArgFlagsTy Flags; | |||
7698 | unsigned OriginalAlignment = DL->getABITypeAlignment(ArgTy); | |||
7699 | ||||
7700 | if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt)) | |||
7701 | Flags.setZExt(); | |||
7702 | if (F.getAttributes().hasAttribute(Idx, Attribute::SExt)) | |||
7703 | Flags.setSExt(); | |||
7704 | if (F.getAttributes().hasAttribute(Idx, Attribute::InReg)) | |||
7705 | Flags.setInReg(); | |||
7706 | if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet)) | |||
7707 | Flags.setSRet(); | |||
7708 | if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) | |||
7709 | Flags.setByVal(); | |||
7710 | if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) { | |||
7711 | Flags.setInAlloca(); | |||
7712 | // Set the byval flag for CCAssignFn callbacks that don't know about | |||
7713 | // inalloca. This way we can know how many bytes we should've allocated | |||
7714 | // and how many bytes a callee cleanup function will pop. If we port | |||
7715 | // inalloca to more targets, we'll have to add custom inalloca handling | |||
7716 | // in the various CC lowering callbacks. | |||
7717 | Flags.setByVal(); | |||
7718 | } | |||
7719 | if (Flags.isByVal() || Flags.isInAlloca()) { | |||
7720 | PointerType *Ty = cast<PointerType>(I->getType()); | |||
7721 | Type *ElementTy = Ty->getElementType(); | |||
7722 | Flags.setByValSize(DL->getTypeAllocSize(ElementTy)); | |||
7723 | // For ByVal, alignment should be passed from FE. BE will guess if | |||
7724 | // this info is not there but there are cases it cannot get right. | |||
7725 | unsigned FrameAlign; | |||
7726 | if (F.getParamAlignment(Idx)) | |||
7727 | FrameAlign = F.getParamAlignment(Idx); | |||
7728 | else | |||
7729 | FrameAlign = TLI->getByValTypeAlignment(ElementTy); | |||
7730 | Flags.setByValAlign(FrameAlign); | |||
7731 | } | |||
7732 | if (F.getAttributes().hasAttribute(Idx, Attribute::Nest)) | |||
7733 | Flags.setNest(); | |||
7734 | if (NeedsRegBlock) | |||
7735 | Flags.setInConsecutiveRegs(); | |||
7736 | Flags.setOrigAlign(OriginalAlignment); | |||
7737 | ||||
7738 | MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT); | |||
7739 | unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT); | |||
7740 | for (unsigned i = 0; i != NumRegs; ++i) { | |||
7741 | ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed, | |||
7742 | Idx-1, PartBase+i*RegisterVT.getStoreSize()); | |||
7743 | if (NumRegs > 1 && i == 0) | |||
7744 | MyFlags.Flags.setSplit(); | |||
7745 | // if it isn't first piece, alignment must be 1 | |||
7746 | else if (i > 0) | |||
7747 | MyFlags.Flags.setOrigAlign(1); | |||
7748 | Ins.push_back(MyFlags); | |||
7749 | } | |||
7750 | if (NeedsRegBlock && Value == NumValues - 1) | |||
7751 | Ins[Ins.size() - 1].Flags.setInConsecutiveRegsLast(); | |||
7752 | PartBase += VT.getStoreSize(); | |||
7753 | } | |||
7754 | } | |||
7755 | ||||
7756 | // Call the target to set up the argument values. | |||
7757 | SmallVector<SDValue, 8> InVals; | |||
7758 | SDValue NewRoot = TLI->LowerFormalArguments( | |||
7759 | DAG.getRoot(), F.getCallingConv(), F.isVarArg(), Ins, dl, DAG, InVals); | |||
7760 | ||||
7761 | // Verify that the target's LowerFormalArguments behaved as expected. | |||
7762 | assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&((NewRoot.getNode() && NewRoot.getValueType() == MVT:: Other && "LowerFormalArguments didn't return a valid chain!" ) ? static_cast<void> (0) : __assert_fail ("NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && \"LowerFormalArguments didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7763, __PRETTY_FUNCTION__)) | |||
7763 | "LowerFormalArguments didn't return a valid chain!")((NewRoot.getNode() && NewRoot.getValueType() == MVT:: Other && "LowerFormalArguments didn't return a valid chain!" ) ? static_cast<void> (0) : __assert_fail ("NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && \"LowerFormalArguments didn't return a valid chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7763, __PRETTY_FUNCTION__)); | |||
7764 | assert(InVals.size() == Ins.size() &&((InVals.size() == Ins.size() && "LowerFormalArguments didn't emit the correct number of values!" ) ? static_cast<void> (0) : __assert_fail ("InVals.size() == Ins.size() && \"LowerFormalArguments didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7765, __PRETTY_FUNCTION__)) | |||
7765 | "LowerFormalArguments didn't emit the correct number of values!")((InVals.size() == Ins.size() && "LowerFormalArguments didn't emit the correct number of values!" ) ? static_cast<void> (0) : __assert_fail ("InVals.size() == Ins.size() && \"LowerFormalArguments didn't emit the correct number of values!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7765, __PRETTY_FUNCTION__)); | |||
7766 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7767 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7768 | assert(InVals[i].getNode() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7769 | "LowerFormalArguments emitted a null value!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7770 | assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7771 | "LowerFormalArguments emitted a value with the wrong type!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7772 | }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0) | |||
7773 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("isel")) { { for (unsigned i = 0, e = Ins.size(); i != e; ++ i) { ((InVals[i].getNode() && "LowerFormalArguments emitted a null value!" ) ? static_cast<void> (0) : __assert_fail ("InVals[i].getNode() && \"LowerFormalArguments emitted a null value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7769, __PRETTY_FUNCTION__)); ((EVT(Ins[i].VT) == InVals[i]. getValueType() && "LowerFormalArguments emitted a value with the wrong type!" ) ? static_cast<void> (0) : __assert_fail ("EVT(Ins[i].VT) == InVals[i].getValueType() && \"LowerFormalArguments emitted a value with the wrong type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7771, __PRETTY_FUNCTION__)); } }; } } while (0); | |||
7774 | ||||
7775 | // Update the DAG with the new chain value resulting from argument lowering. | |||
7776 | DAG.setRoot(NewRoot); | |||
7777 | ||||
7778 | // Set up the argument values. | |||
7779 | unsigned i = 0; | |||
7780 | Idx = 1; | |||
7781 | if (!FuncInfo->CanLowerReturn) { | |||
7782 | // Create a virtual register for the sret pointer, and put in a copy | |||
7783 | // from the sret argument into it. | |||
7784 | SmallVector<EVT, 1> ValueVTs; | |||
7785 | ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); | |||
7786 | MVT VT = ValueVTs[0].getSimpleVT(); | |||
7787 | MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT); | |||
7788 | ISD::NodeType AssertOp = ISD::DELETED_NODE; | |||
7789 | SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, | |||
7790 | RegVT, VT, nullptr, AssertOp); | |||
7791 | ||||
7792 | MachineFunction& MF = SDB->DAG.getMachineFunction(); | |||
7793 | MachineRegisterInfo& RegInfo = MF.getRegInfo(); | |||
7794 | unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT)); | |||
7795 | FuncInfo->DemoteRegister = SRetReg; | |||
7796 | NewRoot = | |||
7797 | SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), SRetReg, ArgValue); | |||
7798 | DAG.setRoot(NewRoot); | |||
7799 | ||||
7800 | // i indexes lowered arguments. Bump it past the hidden sret argument. | |||
7801 | // Idx indexes LLVM arguments. Don't touch it. | |||
7802 | ++i; | |||
7803 | } | |||
7804 | ||||
7805 | for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; | |||
7806 | ++I, ++Idx) { | |||
7807 | SmallVector<SDValue, 4> ArgValues; | |||
7808 | SmallVector<EVT, 4> ValueVTs; | |||
7809 | ComputeValueVTs(*TLI, I->getType(), ValueVTs); | |||
7810 | unsigned NumValues = ValueVTs.size(); | |||
7811 | ||||
7812 | // If this argument is unused then remember its value. It is used to generate | |||
7813 | // debugging information. | |||
7814 | if (I->use_empty() && NumValues) { | |||
7815 | SDB->setUnusedArgValue(I, InVals[i]); | |||
7816 | ||||
7817 | // Also remember any frame index for use in FastISel. | |||
7818 | if (FrameIndexSDNode *FI = | |||
7819 | dyn_cast<FrameIndexSDNode>(InVals[i].getNode())) | |||
7820 | FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); | |||
7821 | } | |||
7822 | ||||
7823 | for (unsigned Val = 0; Val != NumValues; ++Val) { | |||
7824 | EVT VT = ValueVTs[Val]; | |||
7825 | MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT); | |||
7826 | unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT); | |||
7827 | ||||
7828 | if (!I->use_empty()) { | |||
7829 | ISD::NodeType AssertOp = ISD::DELETED_NODE; | |||
7830 | if (F.getAttributes().hasAttribute(Idx, Attribute::SExt)) | |||
7831 | AssertOp = ISD::AssertSext; | |||
7832 | else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt)) | |||
7833 | AssertOp = ISD::AssertZext; | |||
7834 | ||||
7835 | ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], | |||
7836 | NumParts, PartVT, VT, | |||
7837 | nullptr, AssertOp)); | |||
7838 | } | |||
7839 | ||||
7840 | i += NumParts; | |||
7841 | } | |||
7842 | ||||
7843 | // We don't need to do anything else for unused arguments. | |||
7844 | if (ArgValues.empty()) | |||
7845 | continue; | |||
7846 | ||||
7847 | // Note down frame index. | |||
7848 | if (FrameIndexSDNode *FI = | |||
7849 | dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode())) | |||
7850 | FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); | |||
7851 | ||||
7852 | SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues), | |||
7853 | SDB->getCurSDLoc()); | |||
7854 | ||||
7855 | SDB->setValue(I, Res); | |||
7856 | if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) { | |||
7857 | if (LoadSDNode *LNode = | |||
7858 | dyn_cast<LoadSDNode>(Res.getOperand(0).getNode())) | |||
7859 | if (FrameIndexSDNode *FI = | |||
7860 | dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) | |||
7861 | FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); | |||
7862 | } | |||
7863 | ||||
7864 | // If this argument is live outside of the entry block, insert a copy from | |||
7865 | // wherever we got it to the vreg that other BB's will reference it as. | |||
7866 | if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) { | |||
7867 | // If we can, though, try to skip creating an unnecessary vreg. | |||
7868 | // FIXME: This isn't very clean... it would be nice to make this more | |||
7869 | // general. It's also subtly incompatible with the hacks FastISel | |||
7870 | // uses with vregs. | |||
7871 | unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg(); | |||
7872 | if (TargetRegisterInfo::isVirtualRegister(Reg)) { | |||
7873 | FuncInfo->ValueMap[I] = Reg; | |||
7874 | continue; | |||
7875 | } | |||
7876 | } | |||
7877 | if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) { | |||
7878 | FuncInfo->InitializeRegForValue(I); | |||
7879 | SDB->CopyToExportRegsIfNeeded(I); | |||
7880 | } | |||
7881 | } | |||
7882 | ||||
7883 | assert(i == InVals.size() && "Argument register count mismatch!")((i == InVals.size() && "Argument register count mismatch!" ) ? static_cast<void> (0) : __assert_fail ("i == InVals.size() && \"Argument register count mismatch!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7883, __PRETTY_FUNCTION__)); | |||
7884 | ||||
7885 | // Finally, if the target has anything special to do, allow it to do so. | |||
7886 | EmitFunctionEntryCode(); | |||
7887 | } | |||
7888 | ||||
7889 | /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to | |||
7890 | /// ensure constants are generated when needed. Remember the virtual registers | |||
7891 | /// that need to be added to the Machine PHI nodes as input. We cannot just | |||
7892 | /// directly add them, because expansion might result in multiple MBB's for one | |||
7893 | /// BB. As such, the start of the BB might correspond to a different MBB than | |||
7894 | /// the end. | |||
7895 | /// | |||
7896 | void | |||
7897 | SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { | |||
7898 | const TerminatorInst *TI = LLVMBB->getTerminator(); | |||
7899 | ||||
7900 | SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; | |||
7901 | ||||
7902 | // Check successor nodes' PHI nodes that expect a constant to be available | |||
7903 | // from this block. | |||
7904 | for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { | |||
7905 | const BasicBlock *SuccBB = TI->getSuccessor(succ); | |||
7906 | if (!isa<PHINode>(SuccBB->begin())) continue; | |||
7907 | MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; | |||
7908 | ||||
7909 | // If this terminator has multiple identical successors (common for | |||
7910 | // switches), only handle each succ once. | |||
7911 | if (!SuccsHandled.insert(SuccMBB).second) | |||
7912 | continue; | |||
7913 | ||||
7914 | MachineBasicBlock::iterator MBBI = SuccMBB->begin(); | |||
7915 | ||||
7916 | // At this point we know that there is a 1-1 correspondence between LLVM PHI | |||
7917 | // nodes and Machine PHI nodes, but the incoming operands have not been | |||
7918 | // emitted yet. | |||
7919 | for (BasicBlock::const_iterator I = SuccBB->begin(); | |||
7920 | const PHINode *PN = dyn_cast<PHINode>(I); ++I) { | |||
7921 | // Ignore dead phi's. | |||
7922 | if (PN->use_empty()) continue; | |||
7923 | ||||
7924 | // Skip empty types | |||
7925 | if (PN->getType()->isEmptyTy()) | |||
7926 | continue; | |||
7927 | ||||
7928 | unsigned Reg; | |||
7929 | const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); | |||
7930 | ||||
7931 | if (const Constant *C = dyn_cast<Constant>(PHIOp)) { | |||
7932 | unsigned &RegOut = ConstantsOut[C]; | |||
7933 | if (RegOut == 0) { | |||
7934 | RegOut = FuncInfo.CreateRegs(C->getType()); | |||
7935 | CopyValueToVirtualRegister(C, RegOut); | |||
7936 | } | |||
7937 | Reg = RegOut; | |||
7938 | } else { | |||
7939 | DenseMap<const Value *, unsigned>::iterator I = | |||
7940 | FuncInfo.ValueMap.find(PHIOp); | |||
7941 | if (I != FuncInfo.ValueMap.end()) | |||
7942 | Reg = I->second; | |||
7943 | else { | |||
7944 | assert(isa<AllocaInst>(PHIOp) &&((isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap .count(cast<AllocaInst>(PHIOp)) && "Didn't codegen value into a register!??" ) ? static_cast<void> (0) : __assert_fail ("isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && \"Didn't codegen value into a register!??\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7946, __PRETTY_FUNCTION__)) | |||
7945 | FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&((isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap .count(cast<AllocaInst>(PHIOp)) && "Didn't codegen value into a register!??" ) ? static_cast<void> (0) : __assert_fail ("isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && \"Didn't codegen value into a register!??\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7946, __PRETTY_FUNCTION__)) | |||
7946 | "Didn't codegen value into a register!??")((isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap .count(cast<AllocaInst>(PHIOp)) && "Didn't codegen value into a register!??" ) ? static_cast<void> (0) : __assert_fail ("isa<AllocaInst>(PHIOp) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && \"Didn't codegen value into a register!??\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn230682/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp" , 7946, __PRETTY_FUNCTION__)); | |||
7947 | Reg = FuncInfo.CreateRegs(PHIOp->getType()); | |||
7948 | CopyValueToVirtualRegister(PHIOp, Reg); | |||
7949 | } | |||
7950 | } | |||
7951 | ||||
7952 | // Remember that this register needs to added to the machine PHI node as | |||
7953 | // the input for this MBB. | |||
7954 | SmallVector<EVT, 4> ValueVTs; | |||
7955 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
7956 | ComputeValueVTs(TLI, PN->getType(), ValueVTs); | |||
7957 | for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) { | |||
7958 | EVT VT = ValueVTs[vti]; | |||
7959 | unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT); | |||
7960 | for (unsigned i = 0, e = NumRegisters; i != e; ++i) | |||
7961 | FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); | |||
7962 | Reg += NumRegisters; | |||
7963 | } | |||
7964 | } | |||
7965 | } | |||
7966 | ||||
7967 | ConstantsOut.clear(); | |||
7968 | } | |||
7969 | ||||
7970 | /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB | |||
7971 | /// is 0. | |||
7972 | MachineBasicBlock * | |||
7973 | SelectionDAGBuilder::StackProtectorDescriptor:: | |||
7974 | AddSuccessorMBB(const BasicBlock *BB, | |||
7975 | MachineBasicBlock *ParentMBB, | |||
7976 | bool IsLikely, | |||
7977 | MachineBasicBlock *SuccMBB) { | |||
7978 | // If SuccBB has not been created yet, create it. | |||
7979 | if (!SuccMBB) { | |||
7980 | MachineFunction *MF = ParentMBB->getParent(); | |||
7981 | MachineFunction::iterator BBI = ParentMBB; | |||
7982 | SuccMBB = MF->CreateMachineBasicBlock(BB); | |||
7983 | MF->insert(++BBI, SuccMBB); | |||
7984 | } | |||
7985 | // Add it as a successor of ParentMBB. | |||
7986 | ParentMBB->addSuccessor( | |||
7987 | SuccMBB, BranchProbabilityInfo::getBranchWeightStackProtector(IsLikely)); | |||
7988 | return SuccMBB; | |||
7989 | } |