1 //===- X86InstrInfo.cpp - X86 Instruction Information -----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the X86 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "X86InstrInfo.h"
16 #include "X86GenInstrInfo.inc"
17 #include "X86InstrBuilder.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/LiveVariables.h"
29 #include "llvm/CodeGen/PseudoSourceValue.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/MC/MCAsmInfo.h"
41 NoFusing("disable-spill-fusing",
42 cl::desc("Disable fusing of spill code into instructions"));
44 PrintFailedFusing("print-failed-fuse-candidates",
45 cl::desc("Print instructions that the allocator wants to"
46 " fuse, but the X86 backend currently can't"),
49 ReMatPICStubLoad("remat-pic-stub-load",
50 cl::desc("Re-materialize load from stub in PIC mode"),
51 cl::init(false), cl::Hidden);
53 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
54 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
55 TM(tm), RI(tm, *this) {
56 SmallVector<unsigned,16> AmbEntries;
57 static const unsigned OpTbl2Addr[][2] = {
58 { X86::ADC32ri, X86::ADC32mi },
59 { X86::ADC32ri8, X86::ADC32mi8 },
60 { X86::ADC32rr, X86::ADC32mr },
61 { X86::ADC64ri32, X86::ADC64mi32 },
62 { X86::ADC64ri8, X86::ADC64mi8 },
63 { X86::ADC64rr, X86::ADC64mr },
64 { X86::ADD16ri, X86::ADD16mi },
65 { X86::ADD16ri8, X86::ADD16mi8 },
66 { X86::ADD16rr, X86::ADD16mr },
67 { X86::ADD32ri, X86::ADD32mi },
68 { X86::ADD32ri8, X86::ADD32mi8 },
69 { X86::ADD32rr, X86::ADD32mr },
70 { X86::ADD64ri32, X86::ADD64mi32 },
71 { X86::ADD64ri8, X86::ADD64mi8 },
72 { X86::ADD64rr, X86::ADD64mr },
73 { X86::ADD8ri, X86::ADD8mi },
74 { X86::ADD8rr, X86::ADD8mr },
75 { X86::AND16ri, X86::AND16mi },
76 { X86::AND16ri8, X86::AND16mi8 },
77 { X86::AND16rr, X86::AND16mr },
78 { X86::AND32ri, X86::AND32mi },
79 { X86::AND32ri8, X86::AND32mi8 },
80 { X86::AND32rr, X86::AND32mr },
81 { X86::AND64ri32, X86::AND64mi32 },
82 { X86::AND64ri8, X86::AND64mi8 },
83 { X86::AND64rr, X86::AND64mr },
84 { X86::AND8ri, X86::AND8mi },
85 { X86::AND8rr, X86::AND8mr },
86 { X86::DEC16r, X86::DEC16m },
87 { X86::DEC32r, X86::DEC32m },
88 { X86::DEC64_16r, X86::DEC64_16m },
89 { X86::DEC64_32r, X86::DEC64_32m },
90 { X86::DEC64r, X86::DEC64m },
91 { X86::DEC8r, X86::DEC8m },
92 { X86::INC16r, X86::INC16m },
93 { X86::INC32r, X86::INC32m },
94 { X86::INC64_16r, X86::INC64_16m },
95 { X86::INC64_32r, X86::INC64_32m },
96 { X86::INC64r, X86::INC64m },
97 { X86::INC8r, X86::INC8m },
98 { X86::NEG16r, X86::NEG16m },
99 { X86::NEG32r, X86::NEG32m },
100 { X86::NEG64r, X86::NEG64m },
101 { X86::NEG8r, X86::NEG8m },
102 { X86::NOT16r, X86::NOT16m },
103 { X86::NOT32r, X86::NOT32m },
104 { X86::NOT64r, X86::NOT64m },
105 { X86::NOT8r, X86::NOT8m },
106 { X86::OR16ri, X86::OR16mi },
107 { X86::OR16ri8, X86::OR16mi8 },
108 { X86::OR16rr, X86::OR16mr },
109 { X86::OR32ri, X86::OR32mi },
110 { X86::OR32ri8, X86::OR32mi8 },
111 { X86::OR32rr, X86::OR32mr },
112 { X86::OR64ri32, X86::OR64mi32 },
113 { X86::OR64ri8, X86::OR64mi8 },
114 { X86::OR64rr, X86::OR64mr },
115 { X86::OR8ri, X86::OR8mi },
116 { X86::OR8rr, X86::OR8mr },
117 { X86::ROL16r1, X86::ROL16m1 },
118 { X86::ROL16rCL, X86::ROL16mCL },
119 { X86::ROL16ri, X86::ROL16mi },
120 { X86::ROL32r1, X86::ROL32m1 },
121 { X86::ROL32rCL, X86::ROL32mCL },
122 { X86::ROL32ri, X86::ROL32mi },
123 { X86::ROL64r1, X86::ROL64m1 },
124 { X86::ROL64rCL, X86::ROL64mCL },
125 { X86::ROL64ri, X86::ROL64mi },
126 { X86::ROL8r1, X86::ROL8m1 },
127 { X86::ROL8rCL, X86::ROL8mCL },
128 { X86::ROL8ri, X86::ROL8mi },
129 { X86::ROR16r1, X86::ROR16m1 },
130 { X86::ROR16rCL, X86::ROR16mCL },
131 { X86::ROR16ri, X86::ROR16mi },
132 { X86::ROR32r1, X86::ROR32m1 },
133 { X86::ROR32rCL, X86::ROR32mCL },
134 { X86::ROR32ri, X86::ROR32mi },
135 { X86::ROR64r1, X86::ROR64m1 },
136 { X86::ROR64rCL, X86::ROR64mCL },
137 { X86::ROR64ri, X86::ROR64mi },
138 { X86::ROR8r1, X86::ROR8m1 },
139 { X86::ROR8rCL, X86::ROR8mCL },
140 { X86::ROR8ri, X86::ROR8mi },
141 { X86::SAR16r1, X86::SAR16m1 },
142 { X86::SAR16rCL, X86::SAR16mCL },
143 { X86::SAR16ri, X86::SAR16mi },
144 { X86::SAR32r1, X86::SAR32m1 },
145 { X86::SAR32rCL, X86::SAR32mCL },
146 { X86::SAR32ri, X86::SAR32mi },
147 { X86::SAR64r1, X86::SAR64m1 },
148 { X86::SAR64rCL, X86::SAR64mCL },
149 { X86::SAR64ri, X86::SAR64mi },
150 { X86::SAR8r1, X86::SAR8m1 },
151 { X86::SAR8rCL, X86::SAR8mCL },
152 { X86::SAR8ri, X86::SAR8mi },
153 { X86::SBB32ri, X86::SBB32mi },
154 { X86::SBB32ri8, X86::SBB32mi8 },
155 { X86::SBB32rr, X86::SBB32mr },
156 { X86::SBB64ri32, X86::SBB64mi32 },
157 { X86::SBB64ri8, X86::SBB64mi8 },
158 { X86::SBB64rr, X86::SBB64mr },
159 { X86::SHL16rCL, X86::SHL16mCL },
160 { X86::SHL16ri, X86::SHL16mi },
161 { X86::SHL32rCL, X86::SHL32mCL },
162 { X86::SHL32ri, X86::SHL32mi },
163 { X86::SHL64rCL, X86::SHL64mCL },
164 { X86::SHL64ri, X86::SHL64mi },
165 { X86::SHL8rCL, X86::SHL8mCL },
166 { X86::SHL8ri, X86::SHL8mi },
167 { X86::SHLD16rrCL, X86::SHLD16mrCL },
168 { X86::SHLD16rri8, X86::SHLD16mri8 },
169 { X86::SHLD32rrCL, X86::SHLD32mrCL },
170 { X86::SHLD32rri8, X86::SHLD32mri8 },
171 { X86::SHLD64rrCL, X86::SHLD64mrCL },
172 { X86::SHLD64rri8, X86::SHLD64mri8 },
173 { X86::SHR16r1, X86::SHR16m1 },
174 { X86::SHR16rCL, X86::SHR16mCL },
175 { X86::SHR16ri, X86::SHR16mi },
176 { X86::SHR32r1, X86::SHR32m1 },
177 { X86::SHR32rCL, X86::SHR32mCL },
178 { X86::SHR32ri, X86::SHR32mi },
179 { X86::SHR64r1, X86::SHR64m1 },
180 { X86::SHR64rCL, X86::SHR64mCL },
181 { X86::SHR64ri, X86::SHR64mi },
182 { X86::SHR8r1, X86::SHR8m1 },
183 { X86::SHR8rCL, X86::SHR8mCL },
184 { X86::SHR8ri, X86::SHR8mi },
185 { X86::SHRD16rrCL, X86::SHRD16mrCL },
186 { X86::SHRD16rri8, X86::SHRD16mri8 },
187 { X86::SHRD32rrCL, X86::SHRD32mrCL },
188 { X86::SHRD32rri8, X86::SHRD32mri8 },
189 { X86::SHRD64rrCL, X86::SHRD64mrCL },
190 { X86::SHRD64rri8, X86::SHRD64mri8 },
191 { X86::SUB16ri, X86::SUB16mi },
192 { X86::SUB16ri8, X86::SUB16mi8 },
193 { X86::SUB16rr, X86::SUB16mr },
194 { X86::SUB32ri, X86::SUB32mi },
195 { X86::SUB32ri8, X86::SUB32mi8 },
196 { X86::SUB32rr, X86::SUB32mr },
197 { X86::SUB64ri32, X86::SUB64mi32 },
198 { X86::SUB64ri8, X86::SUB64mi8 },
199 { X86::SUB64rr, X86::SUB64mr },
200 { X86::SUB8ri, X86::SUB8mi },
201 { X86::SUB8rr, X86::SUB8mr },
202 { X86::XOR16ri, X86::XOR16mi },
203 { X86::XOR16ri8, X86::XOR16mi8 },
204 { X86::XOR16rr, X86::XOR16mr },
205 { X86::XOR32ri, X86::XOR32mi },
206 { X86::XOR32ri8, X86::XOR32mi8 },
207 { X86::XOR32rr, X86::XOR32mr },
208 { X86::XOR64ri32, X86::XOR64mi32 },
209 { X86::XOR64ri8, X86::XOR64mi8 },
210 { X86::XOR64rr, X86::XOR64mr },
211 { X86::XOR8ri, X86::XOR8mi },
212 { X86::XOR8rr, X86::XOR8mr }
215 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
216 unsigned RegOp = OpTbl2Addr[i][0];
217 unsigned MemOp = OpTbl2Addr[i][1];
218 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp,
219 std::make_pair(MemOp,0))).second)
220 assert(false && "Duplicated entries?");
221 // Index 0, folded load and store, no alignment requirement.
222 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5);
223 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
224 std::make_pair(RegOp,
226 AmbEntries.push_back(MemOp);
229 // If the third value is 1, then it's folding either a load or a store.
230 static const unsigned OpTbl0[][4] = {
231 { X86::BT16ri8, X86::BT16mi8, 1, 0 },
232 { X86::BT32ri8, X86::BT32mi8, 1, 0 },
233 { X86::BT64ri8, X86::BT64mi8, 1, 0 },
234 { X86::CALL32r, X86::CALL32m, 1, 0 },
235 { X86::CALL64r, X86::CALL64m, 1, 0 },
236 { X86::CMP16ri, X86::CMP16mi, 1, 0 },
237 { X86::CMP16ri8, X86::CMP16mi8, 1, 0 },
238 { X86::CMP16rr, X86::CMP16mr, 1, 0 },
239 { X86::CMP32ri, X86::CMP32mi, 1, 0 },
240 { X86::CMP32ri8, X86::CMP32mi8, 1, 0 },
241 { X86::CMP32rr, X86::CMP32mr, 1, 0 },
242 { X86::CMP64ri32, X86::CMP64mi32, 1, 0 },
243 { X86::CMP64ri8, X86::CMP64mi8, 1, 0 },
244 { X86::CMP64rr, X86::CMP64mr, 1, 0 },
245 { X86::CMP8ri, X86::CMP8mi, 1, 0 },
246 { X86::CMP8rr, X86::CMP8mr, 1, 0 },
247 { X86::DIV16r, X86::DIV16m, 1, 0 },
248 { X86::DIV32r, X86::DIV32m, 1, 0 },
249 { X86::DIV64r, X86::DIV64m, 1, 0 },
250 { X86::DIV8r, X86::DIV8m, 1, 0 },
251 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0, 16 },
252 { X86::FsMOVAPDrr, X86::MOVSDmr, 0, 0 },
253 { X86::FsMOVAPSrr, X86::MOVSSmr, 0, 0 },
254 { X86::IDIV16r, X86::IDIV16m, 1, 0 },
255 { X86::IDIV32r, X86::IDIV32m, 1, 0 },
256 { X86::IDIV64r, X86::IDIV64m, 1, 0 },
257 { X86::IDIV8r, X86::IDIV8m, 1, 0 },
258 { X86::IMUL16r, X86::IMUL16m, 1, 0 },
259 { X86::IMUL32r, X86::IMUL32m, 1, 0 },
260 { X86::IMUL64r, X86::IMUL64m, 1, 0 },
261 { X86::IMUL8r, X86::IMUL8m, 1, 0 },
262 { X86::JMP32r, X86::JMP32m, 1, 0 },
263 { X86::JMP64r, X86::JMP64m, 1, 0 },
264 { X86::MOV16ri, X86::MOV16mi, 0, 0 },
265 { X86::MOV16rr, X86::MOV16mr, 0, 0 },
266 { X86::MOV32ri, X86::MOV32mi, 0, 0 },
267 { X86::MOV32rr, X86::MOV32mr, 0, 0 },
268 { X86::MOV64ri32, X86::MOV64mi32, 0, 0 },
269 { X86::MOV64rr, X86::MOV64mr, 0, 0 },
270 { X86::MOV8ri, X86::MOV8mi, 0, 0 },
271 { X86::MOV8rr, X86::MOV8mr, 0, 0 },
272 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, 0, 0 },
273 { X86::MOVAPDrr, X86::MOVAPDmr, 0, 16 },
274 { X86::MOVAPSrr, X86::MOVAPSmr, 0, 16 },
275 { X86::MOVDQArr, X86::MOVDQAmr, 0, 16 },
276 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0, 0 },
277 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0, 0 },
278 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0, 0 },
279 { X86::MOVSDrr, X86::MOVSDmr, 0, 0 },
280 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0, 0 },
281 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0, 0 },
282 { X86::MOVSSrr, X86::MOVSSmr, 0, 0 },
283 { X86::MOVUPDrr, X86::MOVUPDmr, 0, 0 },
284 { X86::MOVUPSrr, X86::MOVUPSmr, 0, 0 },
285 { X86::MUL16r, X86::MUL16m, 1, 0 },
286 { X86::MUL32r, X86::MUL32m, 1, 0 },
287 { X86::MUL64r, X86::MUL64m, 1, 0 },
288 { X86::MUL8r, X86::MUL8m, 1, 0 },
289 { X86::SETAEr, X86::SETAEm, 0, 0 },
290 { X86::SETAr, X86::SETAm, 0, 0 },
291 { X86::SETBEr, X86::SETBEm, 0, 0 },
292 { X86::SETBr, X86::SETBm, 0, 0 },
293 { X86::SETEr, X86::SETEm, 0, 0 },
294 { X86::SETGEr, X86::SETGEm, 0, 0 },
295 { X86::SETGr, X86::SETGm, 0, 0 },
296 { X86::SETLEr, X86::SETLEm, 0, 0 },
297 { X86::SETLr, X86::SETLm, 0, 0 },
298 { X86::SETNEr, X86::SETNEm, 0, 0 },
299 { X86::SETNOr, X86::SETNOm, 0, 0 },
300 { X86::SETNPr, X86::SETNPm, 0, 0 },
301 { X86::SETNSr, X86::SETNSm, 0, 0 },
302 { X86::SETOr, X86::SETOm, 0, 0 },
303 { X86::SETPr, X86::SETPm, 0, 0 },
304 { X86::SETSr, X86::SETSm, 0, 0 },
305 { X86::TAILJMPr, X86::TAILJMPm, 1, 0 },
306 { X86::TEST16ri, X86::TEST16mi, 1, 0 },
307 { X86::TEST32ri, X86::TEST32mi, 1, 0 },
308 { X86::TEST64ri32, X86::TEST64mi32, 1, 0 },
309 { X86::TEST8ri, X86::TEST8mi, 1, 0 }
312 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
313 unsigned RegOp = OpTbl0[i][0];
314 unsigned MemOp = OpTbl0[i][1];
315 unsigned Align = OpTbl0[i][3];
316 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp,
317 std::make_pair(MemOp,Align))).second)
318 assert(false && "Duplicated entries?");
319 unsigned FoldedLoad = OpTbl0[i][2];
320 // Index 0, folded load or store.
321 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
322 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
323 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
324 std::make_pair(RegOp, AuxInfo))).second)
325 AmbEntries.push_back(MemOp);
328 static const unsigned OpTbl1[][3] = {
329 { X86::CMP16rr, X86::CMP16rm, 0 },
330 { X86::CMP32rr, X86::CMP32rm, 0 },
331 { X86::CMP64rr, X86::CMP64rm, 0 },
332 { X86::CMP8rr, X86::CMP8rm, 0 },
333 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
334 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
335 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
336 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
337 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
338 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
339 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
340 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
341 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
342 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
343 { X86::FsMOVAPDrr, X86::MOVSDrm, 0 },
344 { X86::FsMOVAPSrr, X86::MOVSSrm, 0 },
345 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
346 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
347 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
348 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
349 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
350 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
351 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
352 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
353 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
354 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
355 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm, 16 },
356 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm, 16 },
357 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm, 16 },
358 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm, 16 },
359 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm, 16 },
360 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm, 0 },
361 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm, 0 },
362 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm, 0 },
363 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
364 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
365 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
366 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
367 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
368 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
369 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm, 0 },
370 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm, 0 },
371 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm, 16 },
372 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm, 16 },
373 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
374 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
375 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
376 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
377 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
378 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
379 { X86::MOV16rr, X86::MOV16rm, 0 },
380 { X86::MOV32rr, X86::MOV32rm, 0 },
381 { X86::MOV64rr, X86::MOV64rm, 0 },
382 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
383 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
384 { X86::MOV8rr, X86::MOV8rm, 0 },
385 { X86::MOVAPDrr, X86::MOVAPDrm, 16 },
386 { X86::MOVAPSrr, X86::MOVAPSrm, 16 },
387 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
388 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
389 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
390 { X86::MOVDQArr, X86::MOVDQArm, 16 },
391 { X86::MOVSD2PDrr, X86::MOVSD2PDrm, 0 },
392 { X86::MOVSDrr, X86::MOVSDrm, 0 },
393 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, 16 },
394 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, 16 },
395 { X86::MOVSS2PSrr, X86::MOVSS2PSrm, 0 },
396 { X86::MOVSSrr, X86::MOVSSrm, 0 },
397 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
398 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
399 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
400 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
401 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
402 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
403 { X86::MOVUPDrr, X86::MOVUPDrm, 16 },
404 { X86::MOVUPSrr, X86::MOVUPSrm, 16 },
405 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
406 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
407 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, 16 },
408 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
409 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
410 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
411 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
412 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
413 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
414 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
415 { X86::PSHUFDri, X86::PSHUFDmi, 16 },
416 { X86::PSHUFHWri, X86::PSHUFHWmi, 16 },
417 { X86::PSHUFLWri, X86::PSHUFLWmi, 16 },
418 { X86::RCPPSr, X86::RCPPSm, 16 },
419 { X86::RCPPSr_Int, X86::RCPPSm_Int, 16 },
420 { X86::RSQRTPSr, X86::RSQRTPSm, 16 },
421 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, 16 },
422 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
423 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
424 { X86::SQRTPDr, X86::SQRTPDm, 16 },
425 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, 16 },
426 { X86::SQRTPSr, X86::SQRTPSm, 16 },
427 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, 16 },
428 { X86::SQRTSDr, X86::SQRTSDm, 0 },
429 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
430 { X86::SQRTSSr, X86::SQRTSSm, 0 },
431 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
432 { X86::TEST16rr, X86::TEST16rm, 0 },
433 { X86::TEST32rr, X86::TEST32rm, 0 },
434 { X86::TEST64rr, X86::TEST64rm, 0 },
435 { X86::TEST8rr, X86::TEST8rm, 0 },
436 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
437 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
438 { X86::UCOMISSrr, X86::UCOMISSrm, 0 }
441 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
442 unsigned RegOp = OpTbl1[i][0];
443 unsigned MemOp = OpTbl1[i][1];
444 unsigned Align = OpTbl1[i][2];
445 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp,
446 std::make_pair(MemOp,Align))).second)
447 assert(false && "Duplicated entries?");
448 // Index 1, folded load
449 unsigned AuxInfo = 1 | (1 << 4);
450 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
451 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
452 std::make_pair(RegOp, AuxInfo))).second)
453 AmbEntries.push_back(MemOp);
456 static const unsigned OpTbl2[][3] = {
457 { X86::ADC32rr, X86::ADC32rm, 0 },
458 { X86::ADC64rr, X86::ADC64rm, 0 },
459 { X86::ADD16rr, X86::ADD16rm, 0 },
460 { X86::ADD32rr, X86::ADD32rm, 0 },
461 { X86::ADD64rr, X86::ADD64rm, 0 },
462 { X86::ADD8rr, X86::ADD8rm, 0 },
463 { X86::ADDPDrr, X86::ADDPDrm, 16 },
464 { X86::ADDPSrr, X86::ADDPSrm, 16 },
465 { X86::ADDSDrr, X86::ADDSDrm, 0 },
466 { X86::ADDSSrr, X86::ADDSSrm, 0 },
467 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, 16 },
468 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, 16 },
469 { X86::AND16rr, X86::AND16rm, 0 },
470 { X86::AND32rr, X86::AND32rm, 0 },
471 { X86::AND64rr, X86::AND64rm, 0 },
472 { X86::AND8rr, X86::AND8rm, 0 },
473 { X86::ANDNPDrr, X86::ANDNPDrm, 16 },
474 { X86::ANDNPSrr, X86::ANDNPSrm, 16 },
475 { X86::ANDPDrr, X86::ANDPDrm, 16 },
476 { X86::ANDPSrr, X86::ANDPSrm, 16 },
477 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
478 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
479 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
480 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
481 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
482 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
483 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
484 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
485 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
486 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
487 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
488 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
489 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
490 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
491 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
492 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
493 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
494 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
495 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
496 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
497 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
498 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
499 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
500 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
501 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
502 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
503 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
504 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
505 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
506 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
507 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
508 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
509 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
510 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
511 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
512 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
513 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
514 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
515 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
516 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
517 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
518 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
519 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
520 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
521 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
522 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
523 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
524 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
525 { X86::CMPPDrri, X86::CMPPDrmi, 16 },
526 { X86::CMPPSrri, X86::CMPPSrmi, 16 },
527 { X86::CMPSDrr, X86::CMPSDrm, 0 },
528 { X86::CMPSSrr, X86::CMPSSrm, 0 },
529 { X86::DIVPDrr, X86::DIVPDrm, 16 },
530 { X86::DIVPSrr, X86::DIVPSrm, 16 },
531 { X86::DIVSDrr, X86::DIVSDrm, 0 },
532 { X86::DIVSSrr, X86::DIVSSrm, 0 },
533 { X86::FsANDNPDrr, X86::FsANDNPDrm, 16 },
534 { X86::FsANDNPSrr, X86::FsANDNPSrm, 16 },
535 { X86::FsANDPDrr, X86::FsANDPDrm, 16 },
536 { X86::FsANDPSrr, X86::FsANDPSrm, 16 },
537 { X86::FsORPDrr, X86::FsORPDrm, 16 },
538 { X86::FsORPSrr, X86::FsORPSrm, 16 },
539 { X86::FsXORPDrr, X86::FsXORPDrm, 16 },
540 { X86::FsXORPSrr, X86::FsXORPSrm, 16 },
541 { X86::HADDPDrr, X86::HADDPDrm, 16 },
542 { X86::HADDPSrr, X86::HADDPSrm, 16 },
543 { X86::HSUBPDrr, X86::HSUBPDrm, 16 },
544 { X86::HSUBPSrr, X86::HSUBPSrm, 16 },
545 { X86::IMUL16rr, X86::IMUL16rm, 0 },
546 { X86::IMUL32rr, X86::IMUL32rm, 0 },
547 { X86::IMUL64rr, X86::IMUL64rm, 0 },
548 { X86::MAXPDrr, X86::MAXPDrm, 16 },
549 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, 16 },
550 { X86::MAXPSrr, X86::MAXPSrm, 16 },
551 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, 16 },
552 { X86::MAXSDrr, X86::MAXSDrm, 0 },
553 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
554 { X86::MAXSSrr, X86::MAXSSrm, 0 },
555 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
556 { X86::MINPDrr, X86::MINPDrm, 16 },
557 { X86::MINPDrr_Int, X86::MINPDrm_Int, 16 },
558 { X86::MINPSrr, X86::MINPSrm, 16 },
559 { X86::MINPSrr_Int, X86::MINPSrm_Int, 16 },
560 { X86::MINSDrr, X86::MINSDrm, 0 },
561 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
562 { X86::MINSSrr, X86::MINSSrm, 0 },
563 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
564 { X86::MULPDrr, X86::MULPDrm, 16 },
565 { X86::MULPSrr, X86::MULPSrm, 16 },
566 { X86::MULSDrr, X86::MULSDrm, 0 },
567 { X86::MULSSrr, X86::MULSSrm, 0 },
568 { X86::OR16rr, X86::OR16rm, 0 },
569 { X86::OR32rr, X86::OR32rm, 0 },
570 { X86::OR64rr, X86::OR64rm, 0 },
571 { X86::OR8rr, X86::OR8rm, 0 },
572 { X86::ORPDrr, X86::ORPDrm, 16 },
573 { X86::ORPSrr, X86::ORPSrm, 16 },
574 { X86::PACKSSDWrr, X86::PACKSSDWrm, 16 },
575 { X86::PACKSSWBrr, X86::PACKSSWBrm, 16 },
576 { X86::PACKUSWBrr, X86::PACKUSWBrm, 16 },
577 { X86::PADDBrr, X86::PADDBrm, 16 },
578 { X86::PADDDrr, X86::PADDDrm, 16 },
579 { X86::PADDQrr, X86::PADDQrm, 16 },
580 { X86::PADDSBrr, X86::PADDSBrm, 16 },
581 { X86::PADDSWrr, X86::PADDSWrm, 16 },
582 { X86::PADDWrr, X86::PADDWrm, 16 },
583 { X86::PANDNrr, X86::PANDNrm, 16 },
584 { X86::PANDrr, X86::PANDrm, 16 },
585 { X86::PAVGBrr, X86::PAVGBrm, 16 },
586 { X86::PAVGWrr, X86::PAVGWrm, 16 },
587 { X86::PCMPEQBrr, X86::PCMPEQBrm, 16 },
588 { X86::PCMPEQDrr, X86::PCMPEQDrm, 16 },
589 { X86::PCMPEQWrr, X86::PCMPEQWrm, 16 },
590 { X86::PCMPGTBrr, X86::PCMPGTBrm, 16 },
591 { X86::PCMPGTDrr, X86::PCMPGTDrm, 16 },
592 { X86::PCMPGTWrr, X86::PCMPGTWrm, 16 },
593 { X86::PINSRWrri, X86::PINSRWrmi, 16 },
594 { X86::PMADDWDrr, X86::PMADDWDrm, 16 },
595 { X86::PMAXSWrr, X86::PMAXSWrm, 16 },
596 { X86::PMAXUBrr, X86::PMAXUBrm, 16 },
597 { X86::PMINSWrr, X86::PMINSWrm, 16 },
598 { X86::PMINUBrr, X86::PMINUBrm, 16 },
599 { X86::PMULDQrr, X86::PMULDQrm, 16 },
600 { X86::PMULHUWrr, X86::PMULHUWrm, 16 },
601 { X86::PMULHWrr, X86::PMULHWrm, 16 },
602 { X86::PMULLDrr, X86::PMULLDrm, 16 },
603 { X86::PMULLDrr_int, X86::PMULLDrm_int, 16 },
604 { X86::PMULLWrr, X86::PMULLWrm, 16 },
605 { X86::PMULUDQrr, X86::PMULUDQrm, 16 },
606 { X86::PORrr, X86::PORrm, 16 },
607 { X86::PSADBWrr, X86::PSADBWrm, 16 },
608 { X86::PSLLDrr, X86::PSLLDrm, 16 },
609 { X86::PSLLQrr, X86::PSLLQrm, 16 },
610 { X86::PSLLWrr, X86::PSLLWrm, 16 },
611 { X86::PSRADrr, X86::PSRADrm, 16 },
612 { X86::PSRAWrr, X86::PSRAWrm, 16 },
613 { X86::PSRLDrr, X86::PSRLDrm, 16 },
614 { X86::PSRLQrr, X86::PSRLQrm, 16 },
615 { X86::PSRLWrr, X86::PSRLWrm, 16 },
616 { X86::PSUBBrr, X86::PSUBBrm, 16 },
617 { X86::PSUBDrr, X86::PSUBDrm, 16 },
618 { X86::PSUBSBrr, X86::PSUBSBrm, 16 },
619 { X86::PSUBSWrr, X86::PSUBSWrm, 16 },
620 { X86::PSUBWrr, X86::PSUBWrm, 16 },
621 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, 16 },
622 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, 16 },
623 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, 16 },
624 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, 16 },
625 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, 16 },
626 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, 16 },
627 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, 16 },
628 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, 16 },
629 { X86::PXORrr, X86::PXORrm, 16 },
630 { X86::SBB32rr, X86::SBB32rm, 0 },
631 { X86::SBB64rr, X86::SBB64rm, 0 },
632 { X86::SHUFPDrri, X86::SHUFPDrmi, 16 },
633 { X86::SHUFPSrri, X86::SHUFPSrmi, 16 },
634 { X86::SUB16rr, X86::SUB16rm, 0 },
635 { X86::SUB32rr, X86::SUB32rm, 0 },
636 { X86::SUB64rr, X86::SUB64rm, 0 },
637 { X86::SUB8rr, X86::SUB8rm, 0 },
638 { X86::SUBPDrr, X86::SUBPDrm, 16 },
639 { X86::SUBPSrr, X86::SUBPSrm, 16 },
640 { X86::SUBSDrr, X86::SUBSDrm, 0 },
641 { X86::SUBSSrr, X86::SUBSSrm, 0 },
642 // FIXME: TEST*rr -> swapped operand of TEST*mr.
643 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, 16 },
644 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, 16 },
645 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, 16 },
646 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, 16 },
647 { X86::XOR16rr, X86::XOR16rm, 0 },
648 { X86::XOR32rr, X86::XOR32rm, 0 },
649 { X86::XOR64rr, X86::XOR64rm, 0 },
650 { X86::XOR8rr, X86::XOR8rm, 0 },
651 { X86::XORPDrr, X86::XORPDrm, 16 },
652 { X86::XORPSrr, X86::XORPSrm, 16 }
655 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
656 unsigned RegOp = OpTbl2[i][0];
657 unsigned MemOp = OpTbl2[i][1];
658 unsigned Align = OpTbl2[i][2];
659 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp,
660 std::make_pair(MemOp,Align))).second)
661 assert(false && "Duplicated entries?");
662 // Index 2, folded load
663 unsigned AuxInfo = 2 | (1 << 4);
664 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
665 std::make_pair(RegOp, AuxInfo))).second)
666 AmbEntries.push_back(MemOp);
669 // Remove ambiguous entries.
670 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
673 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
674 unsigned &SrcReg, unsigned &DstReg,
675 unsigned &SrcSubIdx, unsigned &DstSubIdx) const {
676 switch (MI.getOpcode()) {
680 case X86::MOV8rr_NOREX:
687 // FP Stack register class copies
688 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
689 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
690 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
692 case X86::FsMOVAPSrr:
693 case X86::FsMOVAPDrr:
697 case X86::MOVSS2PSrr:
698 case X86::MOVSD2PDrr:
699 case X86::MOVPS2SSrr:
700 case X86::MOVPD2SDrr:
701 case X86::MMX_MOVQ64rr:
702 assert(MI.getNumOperands() >= 2 &&
703 MI.getOperand(0).isReg() &&
704 MI.getOperand(1).isReg() &&
705 "invalid register-register move instruction");
706 SrcReg = MI.getOperand(1).getReg();
707 DstReg = MI.getOperand(0).getReg();
708 SrcSubIdx = MI.getOperand(1).getSubReg();
709 DstSubIdx = MI.getOperand(0).getSubReg();
714 /// isFrameOperand - Return true and the FrameIndex if the specified
715 /// operand and follow operands form a reference to the stack frame.
716 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
717 int &FrameIndex) const {
718 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
719 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
720 MI->getOperand(Op+1).getImm() == 1 &&
721 MI->getOperand(Op+2).getReg() == 0 &&
722 MI->getOperand(Op+3).getImm() == 0) {
723 FrameIndex = MI->getOperand(Op).getIndex();
729 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
730 int &FrameIndex) const {
731 switch (MI->getOpcode()) {
743 case X86::MMX_MOVD64rm:
744 case X86::MMX_MOVQ64rm:
745 if (isFrameOperand(MI, 1, FrameIndex)) {
746 return MI->getOperand(0).getReg();
748 // Check for post-frame index elimination operations
749 return hasLoadFromStackSlot(MI, FrameIndex);
755 bool X86InstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
756 int &FrameIndex) const {
757 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
758 oe = MI->memoperands_end();
761 if ((*o)->isLoad() && (*o)->getValue())
762 if (const FixedStackPseudoSourceValue *Value =
763 dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) {
764 FrameIndex = Value->getFrameIndex();
771 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
772 int &FrameIndex) const {
773 switch (MI->getOpcode()) {
785 case X86::MMX_MOVD64mr:
786 case X86::MMX_MOVQ64mr:
787 case X86::MMX_MOVNTQmr:
788 if (isFrameOperand(MI, 0, FrameIndex)) {
789 return MI->getOperand(X86AddrNumOperands).getReg();
791 // Check for post-frame index elimination operations
792 return hasStoreToStackSlot(MI, FrameIndex);
798 bool X86InstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
799 int &FrameIndex) const {
800 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
801 oe = MI->memoperands_end();
804 if ((*o)->isStore() && (*o)->getValue())
805 if (const FixedStackPseudoSourceValue *Value =
806 dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) {
807 FrameIndex = Value->getFrameIndex();
814 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
816 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
817 bool isPICBase = false;
818 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
819 E = MRI.def_end(); I != E; ++I) {
820 MachineInstr *DefMI = I.getOperand().getParent();
821 if (DefMI->getOpcode() != X86::MOVPC32r)
823 assert(!isPICBase && "More than one PIC base?");
830 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
831 AliasAnalysis *AA) const {
832 switch (MI->getOpcode()) {
844 case X86::MMX_MOVD64rm:
845 case X86::MMX_MOVQ64rm: {
846 // Loads from constant pools are trivially rematerializable.
847 if (MI->getOperand(1).isReg() &&
848 MI->getOperand(2).isImm() &&
849 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
850 MI->isInvariantLoad(AA)) {
851 unsigned BaseReg = MI->getOperand(1).getReg();
852 if (BaseReg == 0 || BaseReg == X86::RIP)
854 // Allow re-materialization of PIC load.
855 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
857 const MachineFunction &MF = *MI->getParent()->getParent();
858 const MachineRegisterInfo &MRI = MF.getRegInfo();
859 bool isPICBase = false;
860 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
861 E = MRI.def_end(); I != E; ++I) {
862 MachineInstr *DefMI = I.getOperand().getParent();
863 if (DefMI->getOpcode() != X86::MOVPC32r)
865 assert(!isPICBase && "More than one PIC base?");
875 if (MI->getOperand(2).isImm() &&
876 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
877 !MI->getOperand(4).isReg()) {
878 // lea fi#, lea GV, etc. are all rematerializable.
879 if (!MI->getOperand(1).isReg())
881 unsigned BaseReg = MI->getOperand(1).getReg();
884 // Allow re-materialization of lea PICBase + x.
885 const MachineFunction &MF = *MI->getParent()->getParent();
886 const MachineRegisterInfo &MRI = MF.getRegInfo();
887 return regIsPICBase(BaseReg, MRI);
893 // All other instructions marked M_REMATERIALIZABLE are always trivially
898 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
899 /// would clobber the EFLAGS condition register. Note the result may be
900 /// conservative. If it cannot definitely determine the safety after visiting
901 /// a few instructions in each direction it assumes it's not safe.
902 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
903 MachineBasicBlock::iterator I) {
904 // It's always safe to clobber EFLAGS at the end of a block.
908 // For compile time consideration, if we are not able to determine the
909 // safety after visiting 4 instructions in each direction, we will assume
911 MachineBasicBlock::iterator Iter = I;
912 for (unsigned i = 0; i < 4; ++i) {
913 bool SeenDef = false;
914 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
915 MachineOperand &MO = Iter->getOperand(j);
918 if (MO.getReg() == X86::EFLAGS) {
926 // This instruction defines EFLAGS, no need to look any further.
930 // If we make it to the end of the block, it's safe to clobber EFLAGS.
931 if (Iter == MBB.end())
936 for (unsigned i = 0; i < 4; ++i) {
937 // If we make it to the beginning of the block, it's safe to clobber
938 // EFLAGS iff EFLAGS is not live-in.
939 if (Iter == MBB.begin())
940 return !MBB.isLiveIn(X86::EFLAGS);
943 bool SawKill = false;
944 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
945 MachineOperand &MO = Iter->getOperand(j);
946 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
947 if (MO.isDef()) return MO.isDead();
948 if (MO.isKill()) SawKill = true;
953 // This instruction kills EFLAGS and doesn't redefine it, so
954 // there's no need to look further.
958 // Conservative answer.
962 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
963 MachineBasicBlock::iterator I,
964 unsigned DestReg, unsigned SubIdx,
965 const MachineInstr *Orig) const {
966 DebugLoc DL = DebugLoc::getUnknownLoc();
967 if (I != MBB.end()) DL = I->getDebugLoc();
969 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
970 DestReg = RI.getSubReg(DestReg, SubIdx);
974 // MOV32r0 etc. are implemented with xor which clobbers condition code.
975 // Re-materialize them as movri instructions to avoid side effects.
977 unsigned Opc = Orig->getOpcode();
983 if (!isSafeToClobberEFLAGS(MBB, I)) {
986 case X86::MOV8r0: Opc = X86::MOV8ri; break;
987 case X86::MOV16r0: Opc = X86::MOV16ri; break;
988 case X86::MOV32r0: Opc = X86::MOV32ri; break;
997 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
998 MI->getOperand(0).setReg(DestReg);
1001 BuildMI(MBB, I, DL, get(Opc), DestReg).addImm(0);
1004 MachineInstr *NewMI = prior(I);
1005 NewMI->getOperand(0).setSubReg(SubIdx);
1008 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1009 /// is not marked dead.
1010 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1011 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1012 MachineOperand &MO = MI->getOperand(i);
1013 if (MO.isReg() && MO.isDef() &&
1014 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1021 /// convertToThreeAddress - This method must be implemented by targets that
1022 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1023 /// may be able to convert a two-address instruction into a true
1024 /// three-address instruction on demand. This allows the X86 target (for
1025 /// example) to convert ADD and SHL instructions into LEA instructions if they
1026 /// would require register copies due to two-addressness.
1028 /// This method returns a null pointer if the transformation cannot be
1029 /// performed, otherwise it returns the new instruction.
1032 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1033 MachineBasicBlock::iterator &MBBI,
1034 LiveVariables *LV) const {
1035 MachineInstr *MI = MBBI;
1036 MachineFunction &MF = *MI->getParent()->getParent();
1037 // All instructions input are two-addr instructions. Get the known operands.
1038 unsigned Dest = MI->getOperand(0).getReg();
1039 unsigned Src = MI->getOperand(1).getReg();
1040 bool isDead = MI->getOperand(0).isDead();
1041 bool isKill = MI->getOperand(1).isKill();
1043 MachineInstr *NewMI = NULL;
1044 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1045 // we have better subtarget support, enable the 16-bit LEA generation here.
1046 bool DisableLEA16 = true;
1048 unsigned MIOpc = MI->getOpcode();
1050 case X86::SHUFPSrri: {
1051 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1052 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1054 unsigned B = MI->getOperand(1).getReg();
1055 unsigned C = MI->getOperand(2).getReg();
1056 if (B != C) return 0;
1057 unsigned A = MI->getOperand(0).getReg();
1058 unsigned M = MI->getOperand(3).getImm();
1059 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1060 .addReg(A, RegState::Define | getDeadRegState(isDead))
1061 .addReg(B, getKillRegState(isKill)).addImm(M);
1064 case X86::SHL64ri: {
1065 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1066 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1067 // the flags produced by a shift yet, so this is safe.
1068 unsigned ShAmt = MI->getOperand(2).getImm();
1069 if (ShAmt == 0 || ShAmt >= 4) return 0;
1071 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1072 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1073 .addReg(0).addImm(1 << ShAmt)
1074 .addReg(Src, getKillRegState(isKill))
1078 case X86::SHL32ri: {
1079 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1080 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1081 // the flags produced by a shift yet, so this is safe.
1082 unsigned ShAmt = MI->getOperand(2).getImm();
1083 if (ShAmt == 0 || ShAmt >= 4) return 0;
1085 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1086 X86::LEA64_32r : X86::LEA32r;
1087 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1088 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1089 .addReg(0).addImm(1 << ShAmt)
1090 .addReg(Src, getKillRegState(isKill)).addImm(0);
1093 case X86::SHL16ri: {
1094 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1095 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1096 // the flags produced by a shift yet, so this is safe.
1097 unsigned ShAmt = MI->getOperand(2).getImm();
1098 if (ShAmt == 0 || ShAmt >= 4) return 0;
1101 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1102 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1103 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1104 ? X86::LEA64_32r : X86::LEA32r;
1105 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1106 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1108 // Build and insert into an implicit UNDEF value. This is OK because
1109 // well be shifting and then extracting the lower 16-bits.
1110 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1111 MachineInstr *InsMI =
1112 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::INSERT_SUBREG),leaInReg)
1114 .addReg(Src, getKillRegState(isKill))
1115 .addImm(X86::SUBREG_16BIT);
1117 NewMI = BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(Opc), leaOutReg)
1118 .addReg(0).addImm(1 << ShAmt)
1119 .addReg(leaInReg, RegState::Kill)
1122 MachineInstr *ExtMI =
1123 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::EXTRACT_SUBREG))
1124 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1125 .addReg(leaOutReg, RegState::Kill)
1126 .addImm(X86::SUBREG_16BIT);
1129 // Update live variables
1130 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1131 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1133 LV->replaceKillInstruction(Src, MI, InsMI);
1135 LV->replaceKillInstruction(Dest, MI, ExtMI);
1139 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1140 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1141 .addReg(0).addImm(1 << ShAmt)
1142 .addReg(Src, getKillRegState(isKill))
1148 // The following opcodes also sets the condition code register(s). Only
1149 // convert them to equivalent lea if the condition code register def's
1151 if (hasLiveCondCodeDef(MI))
1154 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1159 case X86::INC64_32r: {
1160 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1161 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1162 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1163 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1164 .addReg(Dest, RegState::Define |
1165 getDeadRegState(isDead)),
1170 case X86::INC64_16r:
1171 if (DisableLEA16) return 0;
1172 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1173 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1174 .addReg(Dest, RegState::Define |
1175 getDeadRegState(isDead)),
1180 case X86::DEC64_32r: {
1181 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1182 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1183 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1184 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1185 .addReg(Dest, RegState::Define |
1186 getDeadRegState(isDead)),
1191 case X86::DEC64_16r:
1192 if (DisableLEA16) return 0;
1193 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1194 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1195 .addReg(Dest, RegState::Define |
1196 getDeadRegState(isDead)),
1200 case X86::ADD32rr: {
1201 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1202 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1203 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1204 unsigned Src2 = MI->getOperand(2).getReg();
1205 bool isKill2 = MI->getOperand(2).isKill();
1206 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1207 .addReg(Dest, RegState::Define |
1208 getDeadRegState(isDead)),
1209 Src, isKill, Src2, isKill2);
1211 LV->replaceKillInstruction(Src2, MI, NewMI);
1214 case X86::ADD16rr: {
1215 if (DisableLEA16) return 0;
1216 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1217 unsigned Src2 = MI->getOperand(2).getReg();
1218 bool isKill2 = MI->getOperand(2).isKill();
1219 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1220 .addReg(Dest, RegState::Define |
1221 getDeadRegState(isDead)),
1222 Src, isKill, Src2, isKill2);
1224 LV->replaceKillInstruction(Src2, MI, NewMI);
1227 case X86::ADD64ri32:
1229 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1230 if (MI->getOperand(2).isImm())
1231 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1232 .addReg(Dest, RegState::Define |
1233 getDeadRegState(isDead)),
1234 Src, isKill, MI->getOperand(2).getImm());
1238 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1239 if (MI->getOperand(2).isImm()) {
1240 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1241 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1242 .addReg(Dest, RegState::Define |
1243 getDeadRegState(isDead)),
1244 Src, isKill, MI->getOperand(2).getImm());
1249 if (DisableLEA16) return 0;
1250 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1251 if (MI->getOperand(2).isImm())
1252 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1253 .addReg(Dest, RegState::Define |
1254 getDeadRegState(isDead)),
1255 Src, isKill, MI->getOperand(2).getImm());
1258 if (DisableLEA16) return 0;
1260 case X86::SHL64ri: {
1261 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImm() &&
1262 "Unknown shl instruction!");
1263 unsigned ShAmt = MI->getOperand(2).getImm();
1264 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1266 AM.Scale = 1 << ShAmt;
1268 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1269 : (MIOpc == X86::SHL32ri
1270 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1271 NewMI = addFullAddress(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1272 .addReg(Dest, RegState::Define |
1273 getDeadRegState(isDead)), AM);
1275 NewMI->getOperand(3).setIsKill(true);
1283 if (!NewMI) return 0;
1285 if (LV) { // Update live variables
1287 LV->replaceKillInstruction(Src, MI, NewMI);
1289 LV->replaceKillInstruction(Dest, MI, NewMI);
1292 MFI->insert(MBBI, NewMI); // Insert the new inst
1296 /// commuteInstruction - We have a few instructions that must be hacked on to
1300 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1301 switch (MI->getOpcode()) {
1302 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1303 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1304 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1305 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1306 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1307 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1310 switch (MI->getOpcode()) {
1311 default: llvm_unreachable("Unreachable!");
1312 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1313 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1314 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1315 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1316 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1317 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1319 unsigned Amt = MI->getOperand(3).getImm();
1321 MachineFunction &MF = *MI->getParent()->getParent();
1322 MI = MF.CloneMachineInstr(MI);
1325 MI->setDesc(get(Opc));
1326 MI->getOperand(3).setImm(Size-Amt);
1327 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1329 case X86::CMOVB16rr:
1330 case X86::CMOVB32rr:
1331 case X86::CMOVB64rr:
1332 case X86::CMOVAE16rr:
1333 case X86::CMOVAE32rr:
1334 case X86::CMOVAE64rr:
1335 case X86::CMOVE16rr:
1336 case X86::CMOVE32rr:
1337 case X86::CMOVE64rr:
1338 case X86::CMOVNE16rr:
1339 case X86::CMOVNE32rr:
1340 case X86::CMOVNE64rr:
1341 case X86::CMOVBE16rr:
1342 case X86::CMOVBE32rr:
1343 case X86::CMOVBE64rr:
1344 case X86::CMOVA16rr:
1345 case X86::CMOVA32rr:
1346 case X86::CMOVA64rr:
1347 case X86::CMOVL16rr:
1348 case X86::CMOVL32rr:
1349 case X86::CMOVL64rr:
1350 case X86::CMOVGE16rr:
1351 case X86::CMOVGE32rr:
1352 case X86::CMOVGE64rr:
1353 case X86::CMOVLE16rr:
1354 case X86::CMOVLE32rr:
1355 case X86::CMOVLE64rr:
1356 case X86::CMOVG16rr:
1357 case X86::CMOVG32rr:
1358 case X86::CMOVG64rr:
1359 case X86::CMOVS16rr:
1360 case X86::CMOVS32rr:
1361 case X86::CMOVS64rr:
1362 case X86::CMOVNS16rr:
1363 case X86::CMOVNS32rr:
1364 case X86::CMOVNS64rr:
1365 case X86::CMOVP16rr:
1366 case X86::CMOVP32rr:
1367 case X86::CMOVP64rr:
1368 case X86::CMOVNP16rr:
1369 case X86::CMOVNP32rr:
1370 case X86::CMOVNP64rr:
1371 case X86::CMOVO16rr:
1372 case X86::CMOVO32rr:
1373 case X86::CMOVO64rr:
1374 case X86::CMOVNO16rr:
1375 case X86::CMOVNO32rr:
1376 case X86::CMOVNO64rr: {
1378 switch (MI->getOpcode()) {
1380 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1381 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1382 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1383 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1384 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1385 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1386 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1387 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1388 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1389 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1390 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1391 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1392 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1393 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1394 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1395 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1396 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1397 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1398 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1399 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1400 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1401 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1402 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1403 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1404 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1405 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1406 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1407 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1408 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1409 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1410 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1411 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1412 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
1413 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1414 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1415 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1416 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1417 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1418 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
1419 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1420 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1421 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1422 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1423 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1424 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
1425 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1426 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1427 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1430 MachineFunction &MF = *MI->getParent()->getParent();
1431 MI = MF.CloneMachineInstr(MI);
1434 MI->setDesc(get(Opc));
1435 // Fallthrough intended.
1438 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1442 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1444 default: return X86::COND_INVALID;
1445 case X86::JE: return X86::COND_E;
1446 case X86::JNE: return X86::COND_NE;
1447 case X86::JL: return X86::COND_L;
1448 case X86::JLE: return X86::COND_LE;
1449 case X86::JG: return X86::COND_G;
1450 case X86::JGE: return X86::COND_GE;
1451 case X86::JB: return X86::COND_B;
1452 case X86::JBE: return X86::COND_BE;
1453 case X86::JA: return X86::COND_A;
1454 case X86::JAE: return X86::COND_AE;
1455 case X86::JS: return X86::COND_S;
1456 case X86::JNS: return X86::COND_NS;
1457 case X86::JP: return X86::COND_P;
1458 case X86::JNP: return X86::COND_NP;
1459 case X86::JO: return X86::COND_O;
1460 case X86::JNO: return X86::COND_NO;
1464 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1466 default: llvm_unreachable("Illegal condition code!");
1467 case X86::COND_E: return X86::JE;
1468 case X86::COND_NE: return X86::JNE;
1469 case X86::COND_L: return X86::JL;
1470 case X86::COND_LE: return X86::JLE;
1471 case X86::COND_G: return X86::JG;
1472 case X86::COND_GE: return X86::JGE;
1473 case X86::COND_B: return X86::JB;
1474 case X86::COND_BE: return X86::JBE;
1475 case X86::COND_A: return X86::JA;
1476 case X86::COND_AE: return X86::JAE;
1477 case X86::COND_S: return X86::JS;
1478 case X86::COND_NS: return X86::JNS;
1479 case X86::COND_P: return X86::JP;
1480 case X86::COND_NP: return X86::JNP;
1481 case X86::COND_O: return X86::JO;
1482 case X86::COND_NO: return X86::JNO;
1486 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1487 /// e.g. turning COND_E to COND_NE.
1488 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1490 default: llvm_unreachable("Illegal condition code!");
1491 case X86::COND_E: return X86::COND_NE;
1492 case X86::COND_NE: return X86::COND_E;
1493 case X86::COND_L: return X86::COND_GE;
1494 case X86::COND_LE: return X86::COND_G;
1495 case X86::COND_G: return X86::COND_LE;
1496 case X86::COND_GE: return X86::COND_L;
1497 case X86::COND_B: return X86::COND_AE;
1498 case X86::COND_BE: return X86::COND_A;
1499 case X86::COND_A: return X86::COND_BE;
1500 case X86::COND_AE: return X86::COND_B;
1501 case X86::COND_S: return X86::COND_NS;
1502 case X86::COND_NS: return X86::COND_S;
1503 case X86::COND_P: return X86::COND_NP;
1504 case X86::COND_NP: return X86::COND_P;
1505 case X86::COND_O: return X86::COND_NO;
1506 case X86::COND_NO: return X86::COND_O;
1510 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1511 const TargetInstrDesc &TID = MI->getDesc();
1512 if (!TID.isTerminator()) return false;
1514 // Conditional branch is a special case.
1515 if (TID.isBranch() && !TID.isBarrier())
1517 if (!TID.isPredicable())
1519 return !isPredicated(MI);
1522 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1523 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1524 const X86InstrInfo &TII) {
1525 if (MI->getOpcode() == X86::FP_REG_KILL)
1527 return TII.isUnpredicatedTerminator(MI);
1530 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1531 MachineBasicBlock *&TBB,
1532 MachineBasicBlock *&FBB,
1533 SmallVectorImpl<MachineOperand> &Cond,
1534 bool AllowModify) const {
1535 // Start from the bottom of the block and work up, examining the
1536 // terminator instructions.
1537 MachineBasicBlock::iterator I = MBB.end();
1538 while (I != MBB.begin()) {
1540 // Working from the bottom, when we see a non-terminator
1541 // instruction, we're done.
1542 if (!isBrAnalysisUnpredicatedTerminator(I, *this))
1544 // A terminator that isn't a branch can't easily be handled
1545 // by this analysis.
1546 if (!I->getDesc().isBranch())
1548 // Handle unconditional branches.
1549 if (I->getOpcode() == X86::JMP) {
1551 TBB = I->getOperand(0).getMBB();
1555 // If the block has any instructions after a JMP, delete them.
1556 while (next(I) != MBB.end())
1557 next(I)->eraseFromParent();
1560 // Delete the JMP if it's equivalent to a fall-through.
1561 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1563 I->eraseFromParent();
1567 // TBB is used to indicate the unconditinal destination.
1568 TBB = I->getOperand(0).getMBB();
1571 // Handle conditional branches.
1572 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1573 if (BranchCode == X86::COND_INVALID)
1574 return true; // Can't handle indirect branch.
1575 // Working from the bottom, handle the first conditional branch.
1578 TBB = I->getOperand(0).getMBB();
1579 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1582 // Handle subsequent conditional branches. Only handle the case
1583 // where all conditional branches branch to the same destination
1584 // and their condition opcodes fit one of the special
1585 // multi-branch idioms.
1586 assert(Cond.size() == 1);
1588 // Only handle the case where all conditional branches branch to
1589 // the same destination.
1590 if (TBB != I->getOperand(0).getMBB())
1592 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1593 // If the conditions are the same, we can leave them alone.
1594 if (OldBranchCode == BranchCode)
1596 // If they differ, see if they fit one of the known patterns.
1597 // Theoretically we could handle more patterns here, but
1598 // we shouldn't expect to see them if instruction selection
1599 // has done a reasonable job.
1600 if ((OldBranchCode == X86::COND_NP &&
1601 BranchCode == X86::COND_E) ||
1602 (OldBranchCode == X86::COND_E &&
1603 BranchCode == X86::COND_NP))
1604 BranchCode = X86::COND_NP_OR_E;
1605 else if ((OldBranchCode == X86::COND_P &&
1606 BranchCode == X86::COND_NE) ||
1607 (OldBranchCode == X86::COND_NE &&
1608 BranchCode == X86::COND_P))
1609 BranchCode = X86::COND_NE_OR_P;
1612 // Update the MachineOperand.
1613 Cond[0].setImm(BranchCode);
1619 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1620 MachineBasicBlock::iterator I = MBB.end();
1623 while (I != MBB.begin()) {
1625 if (I->getOpcode() != X86::JMP &&
1626 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1628 // Remove the branch.
1629 I->eraseFromParent();
1638 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1639 MachineBasicBlock *FBB,
1640 const SmallVectorImpl<MachineOperand> &Cond) const {
1641 // FIXME this should probably have a DebugLoc operand
1642 DebugLoc dl = DebugLoc::getUnknownLoc();
1643 // Shouldn't be a fall through.
1644 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1645 assert((Cond.size() == 1 || Cond.size() == 0) &&
1646 "X86 branch conditions have one component!");
1649 // Unconditional branch?
1650 assert(!FBB && "Unconditional branch with multiple successors!");
1651 BuildMI(&MBB, dl, get(X86::JMP)).addMBB(TBB);
1655 // Conditional branch.
1657 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1659 case X86::COND_NP_OR_E:
1660 // Synthesize NP_OR_E with two branches.
1661 BuildMI(&MBB, dl, get(X86::JNP)).addMBB(TBB);
1663 BuildMI(&MBB, dl, get(X86::JE)).addMBB(TBB);
1666 case X86::COND_NE_OR_P:
1667 // Synthesize NE_OR_P with two branches.
1668 BuildMI(&MBB, dl, get(X86::JNE)).addMBB(TBB);
1670 BuildMI(&MBB, dl, get(X86::JP)).addMBB(TBB);
1674 unsigned Opc = GetCondBranchFromCond(CC);
1675 BuildMI(&MBB, dl, get(Opc)).addMBB(TBB);
1680 // Two-way Conditional branch. Insert the second branch.
1681 BuildMI(&MBB, dl, get(X86::JMP)).addMBB(FBB);
1687 /// isHReg - Test if the given register is a physical h register.
1688 static bool isHReg(unsigned Reg) {
1689 return X86::GR8_ABCD_HRegClass.contains(Reg);
1692 bool X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1693 MachineBasicBlock::iterator MI,
1694 unsigned DestReg, unsigned SrcReg,
1695 const TargetRegisterClass *DestRC,
1696 const TargetRegisterClass *SrcRC) const {
1697 DebugLoc DL = DebugLoc::getUnknownLoc();
1698 if (MI != MBB.end()) DL = MI->getDebugLoc();
1700 // Determine if DstRC and SrcRC have a common superclass in common.
1701 const TargetRegisterClass *CommonRC = DestRC;
1702 if (DestRC == SrcRC)
1703 /* Source and destination have the same register class. */;
1704 else if (CommonRC->hasSuperClass(SrcRC))
1706 else if (!DestRC->hasSubClass(SrcRC)) {
1707 // Neither of GR64_NOREX or GR64_NOSP is a superclass of the other,
1708 // but we want to copy then as GR64. Similarly, for GR32_NOREX and
1709 // GR32_NOSP, copy as GR32.
1710 if (SrcRC->hasSuperClass(&X86::GR64RegClass) &&
1711 DestRC->hasSuperClass(&X86::GR64RegClass))
1712 CommonRC = &X86::GR64RegClass;
1713 else if (SrcRC->hasSuperClass(&X86::GR32RegClass) &&
1714 DestRC->hasSuperClass(&X86::GR32RegClass))
1715 CommonRC = &X86::GR32RegClass;
1722 if (CommonRC == &X86::GR64RegClass || CommonRC == &X86::GR64_NOSPRegClass) {
1724 } else if (CommonRC == &X86::GR32RegClass ||
1725 CommonRC == &X86::GR32_NOSPRegClass) {
1727 } else if (CommonRC == &X86::GR16RegClass) {
1729 } else if (CommonRC == &X86::GR8RegClass) {
1730 // Copying to or from a physical H register on x86-64 requires a NOREX
1731 // move. Otherwise use a normal move.
1732 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
1733 TM.getSubtarget<X86Subtarget>().is64Bit())
1734 Opc = X86::MOV8rr_NOREX;
1737 } else if (CommonRC == &X86::GR64_ABCDRegClass) {
1739 } else if (CommonRC == &X86::GR32_ABCDRegClass) {
1741 } else if (CommonRC == &X86::GR16_ABCDRegClass) {
1743 } else if (CommonRC == &X86::GR8_ABCD_LRegClass) {
1745 } else if (CommonRC == &X86::GR8_ABCD_HRegClass) {
1746 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1747 Opc = X86::MOV8rr_NOREX;
1750 } else if (CommonRC == &X86::GR64_NOREXRegClass ||
1751 CommonRC == &X86::GR64_NOREX_NOSPRegClass) {
1753 } else if (CommonRC == &X86::GR32_NOREXRegClass) {
1755 } else if (CommonRC == &X86::GR16_NOREXRegClass) {
1757 } else if (CommonRC == &X86::GR8_NOREXRegClass) {
1759 } else if (CommonRC == &X86::RFP32RegClass) {
1760 Opc = X86::MOV_Fp3232;
1761 } else if (CommonRC == &X86::RFP64RegClass || CommonRC == &X86::RSTRegClass) {
1762 Opc = X86::MOV_Fp6464;
1763 } else if (CommonRC == &X86::RFP80RegClass) {
1764 Opc = X86::MOV_Fp8080;
1765 } else if (CommonRC == &X86::FR32RegClass) {
1766 Opc = X86::FsMOVAPSrr;
1767 } else if (CommonRC == &X86::FR64RegClass) {
1768 Opc = X86::FsMOVAPDrr;
1769 } else if (CommonRC == &X86::VR128RegClass) {
1770 Opc = X86::MOVAPSrr;
1771 } else if (CommonRC == &X86::VR64RegClass) {
1772 Opc = X86::MMX_MOVQ64rr;
1776 BuildMI(MBB, MI, DL, get(Opc), DestReg).addReg(SrcReg);
1780 // Moving EFLAGS to / from another register requires a push and a pop.
1781 if (SrcRC == &X86::CCRRegClass) {
1782 if (SrcReg != X86::EFLAGS)
1784 if (DestRC == &X86::GR64RegClass || DestRC == &X86::GR64_NOSPRegClass) {
1785 BuildMI(MBB, MI, DL, get(X86::PUSHFQ));
1786 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
1788 } else if (DestRC == &X86::GR32RegClass ||
1789 DestRC == &X86::GR32_NOSPRegClass) {
1790 BuildMI(MBB, MI, DL, get(X86::PUSHFD));
1791 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
1794 } else if (DestRC == &X86::CCRRegClass) {
1795 if (DestReg != X86::EFLAGS)
1797 if (SrcRC == &X86::GR64RegClass || DestRC == &X86::GR64_NOSPRegClass) {
1798 BuildMI(MBB, MI, DL, get(X86::PUSH64r)).addReg(SrcReg);
1799 BuildMI(MBB, MI, DL, get(X86::POPFQ));
1801 } else if (SrcRC == &X86::GR32RegClass ||
1802 DestRC == &X86::GR32_NOSPRegClass) {
1803 BuildMI(MBB, MI, DL, get(X86::PUSH32r)).addReg(SrcReg);
1804 BuildMI(MBB, MI, DL, get(X86::POPFD));
1809 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1810 if (SrcRC == &X86::RSTRegClass) {
1811 // Copying from ST(0)/ST(1).
1812 if (SrcReg != X86::ST0 && SrcReg != X86::ST1)
1813 // Can only copy from ST(0)/ST(1) right now
1815 bool isST0 = SrcReg == X86::ST0;
1817 if (DestRC == &X86::RFP32RegClass)
1818 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1819 else if (DestRC == &X86::RFP64RegClass)
1820 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1822 if (DestRC != &X86::RFP80RegClass)
1824 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1826 BuildMI(MBB, MI, DL, get(Opc), DestReg);
1830 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1831 if (DestRC == &X86::RSTRegClass) {
1832 // Copying to ST(0) / ST(1).
1833 if (DestReg != X86::ST0 && DestReg != X86::ST1)
1834 // Can only copy to TOS right now
1836 bool isST0 = DestReg == X86::ST0;
1838 if (SrcRC == &X86::RFP32RegClass)
1839 Opc = isST0 ? X86::FpSET_ST0_32 : X86::FpSET_ST1_32;
1840 else if (SrcRC == &X86::RFP64RegClass)
1841 Opc = isST0 ? X86::FpSET_ST0_64 : X86::FpSET_ST1_64;
1843 if (SrcRC != &X86::RFP80RegClass)
1845 Opc = isST0 ? X86::FpSET_ST0_80 : X86::FpSET_ST1_80;
1847 BuildMI(MBB, MI, DL, get(Opc)).addReg(SrcReg);
1851 // Not yet supported!
1855 static unsigned getStoreRegOpcode(unsigned SrcReg,
1856 const TargetRegisterClass *RC,
1857 bool isStackAligned,
1858 TargetMachine &TM) {
1860 if (RC == &X86::GR64RegClass || RC == &X86::GR64_NOSPRegClass) {
1862 } else if (RC == &X86::GR32RegClass || RC == &X86::GR32_NOSPRegClass) {
1864 } else if (RC == &X86::GR16RegClass) {
1866 } else if (RC == &X86::GR8RegClass) {
1867 // Copying to or from a physical H register on x86-64 requires a NOREX
1868 // move. Otherwise use a normal move.
1869 if (isHReg(SrcReg) &&
1870 TM.getSubtarget<X86Subtarget>().is64Bit())
1871 Opc = X86::MOV8mr_NOREX;
1874 } else if (RC == &X86::GR64_ABCDRegClass) {
1876 } else if (RC == &X86::GR32_ABCDRegClass) {
1878 } else if (RC == &X86::GR16_ABCDRegClass) {
1880 } else if (RC == &X86::GR8_ABCD_LRegClass) {
1882 } else if (RC == &X86::GR8_ABCD_HRegClass) {
1883 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1884 Opc = X86::MOV8mr_NOREX;
1887 } else if (RC == &X86::GR64_NOREXRegClass ||
1888 RC == &X86::GR64_NOREX_NOSPRegClass) {
1890 } else if (RC == &X86::GR32_NOREXRegClass) {
1892 } else if (RC == &X86::GR16_NOREXRegClass) {
1894 } else if (RC == &X86::GR8_NOREXRegClass) {
1896 } else if (RC == &X86::RFP80RegClass) {
1897 Opc = X86::ST_FpP80m; // pops
1898 } else if (RC == &X86::RFP64RegClass) {
1899 Opc = X86::ST_Fp64m;
1900 } else if (RC == &X86::RFP32RegClass) {
1901 Opc = X86::ST_Fp32m;
1902 } else if (RC == &X86::FR32RegClass) {
1904 } else if (RC == &X86::FR64RegClass) {
1906 } else if (RC == &X86::VR128RegClass) {
1907 // If stack is realigned we can use aligned stores.
1908 Opc = isStackAligned ? X86::MOVAPSmr : X86::MOVUPSmr;
1909 } else if (RC == &X86::VR64RegClass) {
1910 Opc = X86::MMX_MOVQ64mr;
1912 llvm_unreachable("Unknown regclass");
1918 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1919 MachineBasicBlock::iterator MI,
1920 unsigned SrcReg, bool isKill, int FrameIdx,
1921 const TargetRegisterClass *RC) const {
1922 const MachineFunction &MF = *MBB.getParent();
1923 bool isAligned = (RI.getStackAlignment() >= 16) ||
1924 RI.needsStackRealignment(MF);
1925 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
1926 DebugLoc DL = DebugLoc::getUnknownLoc();
1927 if (MI != MBB.end()) DL = MI->getDebugLoc();
1928 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
1929 .addReg(SrcReg, getKillRegState(isKill));
1932 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1934 SmallVectorImpl<MachineOperand> &Addr,
1935 const TargetRegisterClass *RC,
1936 MachineInstr::mmo_iterator MMOBegin,
1937 MachineInstr::mmo_iterator MMOEnd,
1938 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1939 bool isAligned = (RI.getStackAlignment() >= 16) ||
1940 RI.needsStackRealignment(MF);
1941 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
1942 DebugLoc DL = DebugLoc::getUnknownLoc();
1943 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
1944 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1945 MIB.addOperand(Addr[i]);
1946 MIB.addReg(SrcReg, getKillRegState(isKill));
1947 (*MIB).setMemRefs(MMOBegin, MMOEnd);
1948 NewMIs.push_back(MIB);
1951 static unsigned getLoadRegOpcode(unsigned DestReg,
1952 const TargetRegisterClass *RC,
1953 bool isStackAligned,
1954 const TargetMachine &TM) {
1956 if (RC == &X86::GR64RegClass || RC == &X86::GR64_NOSPRegClass) {
1958 } else if (RC == &X86::GR32RegClass || RC == &X86::GR32_NOSPRegClass) {
1960 } else if (RC == &X86::GR16RegClass) {
1962 } else if (RC == &X86::GR8RegClass) {
1963 // Copying to or from a physical H register on x86-64 requires a NOREX
1964 // move. Otherwise use a normal move.
1965 if (isHReg(DestReg) &&
1966 TM.getSubtarget<X86Subtarget>().is64Bit())
1967 Opc = X86::MOV8rm_NOREX;
1970 } else if (RC == &X86::GR64_ABCDRegClass) {
1972 } else if (RC == &X86::GR32_ABCDRegClass) {
1974 } else if (RC == &X86::GR16_ABCDRegClass) {
1976 } else if (RC == &X86::GR8_ABCD_LRegClass) {
1978 } else if (RC == &X86::GR8_ABCD_HRegClass) {
1979 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1980 Opc = X86::MOV8rm_NOREX;
1983 } else if (RC == &X86::GR64_NOREXRegClass ||
1984 RC == &X86::GR64_NOREX_NOSPRegClass) {
1986 } else if (RC == &X86::GR32_NOREXRegClass) {
1988 } else if (RC == &X86::GR16_NOREXRegClass) {
1990 } else if (RC == &X86::GR8_NOREXRegClass) {
1992 } else if (RC == &X86::RFP80RegClass) {
1993 Opc = X86::LD_Fp80m;
1994 } else if (RC == &X86::RFP64RegClass) {
1995 Opc = X86::LD_Fp64m;
1996 } else if (RC == &X86::RFP32RegClass) {
1997 Opc = X86::LD_Fp32m;
1998 } else if (RC == &X86::FR32RegClass) {
2000 } else if (RC == &X86::FR64RegClass) {
2002 } else if (RC == &X86::VR128RegClass) {
2003 // If stack is realigned we can use aligned loads.
2004 Opc = isStackAligned ? X86::MOVAPSrm : X86::MOVUPSrm;
2005 } else if (RC == &X86::VR64RegClass) {
2006 Opc = X86::MMX_MOVQ64rm;
2008 llvm_unreachable("Unknown regclass");
2014 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
2015 MachineBasicBlock::iterator MI,
2016 unsigned DestReg, int FrameIdx,
2017 const TargetRegisterClass *RC) const{
2018 const MachineFunction &MF = *MBB.getParent();
2019 bool isAligned = (RI.getStackAlignment() >= 16) ||
2020 RI.needsStackRealignment(MF);
2021 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2022 DebugLoc DL = DebugLoc::getUnknownLoc();
2023 if (MI != MBB.end()) DL = MI->getDebugLoc();
2024 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
2027 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
2028 SmallVectorImpl<MachineOperand> &Addr,
2029 const TargetRegisterClass *RC,
2030 MachineInstr::mmo_iterator MMOBegin,
2031 MachineInstr::mmo_iterator MMOEnd,
2032 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2033 bool isAligned = (RI.getStackAlignment() >= 16) ||
2034 RI.needsStackRealignment(MF);
2035 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2036 DebugLoc DL = DebugLoc::getUnknownLoc();
2037 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
2038 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2039 MIB.addOperand(Addr[i]);
2040 (*MIB).setMemRefs(MMOBegin, MMOEnd);
2041 NewMIs.push_back(MIB);
2044 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
2045 MachineBasicBlock::iterator MI,
2046 const std::vector<CalleeSavedInfo> &CSI) const {
2050 DebugLoc DL = DebugLoc::getUnknownLoc();
2051 if (MI != MBB.end()) DL = MI->getDebugLoc();
2053 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2054 bool isWin64 = TM.getSubtarget<X86Subtarget>().isTargetWin64();
2055 unsigned SlotSize = is64Bit ? 8 : 4;
2057 MachineFunction &MF = *MBB.getParent();
2058 unsigned FPReg = RI.getFrameRegister(MF);
2059 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
2060 unsigned CalleeFrameSize = 0;
2062 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
2063 for (unsigned i = CSI.size(); i != 0; --i) {
2064 unsigned Reg = CSI[i-1].getReg();
2065 const TargetRegisterClass *RegClass = CSI[i-1].getRegClass();
2066 // Add the callee-saved register as live-in. It's killed at the spill.
2069 // X86RegisterInfo::emitPrologue will handle spilling of frame register.
2071 if (RegClass != &X86::VR128RegClass && !isWin64) {
2072 CalleeFrameSize += SlotSize;
2073 BuildMI(MBB, MI, DL, get(Opc)).addReg(Reg, RegState::Kill);
2075 storeRegToStackSlot(MBB, MI, Reg, true, CSI[i-1].getFrameIdx(), RegClass);
2079 X86FI->setCalleeSavedFrameSize(CalleeFrameSize);
2083 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
2084 MachineBasicBlock::iterator MI,
2085 const std::vector<CalleeSavedInfo> &CSI) const {
2089 DebugLoc DL = DebugLoc::getUnknownLoc();
2090 if (MI != MBB.end()) DL = MI->getDebugLoc();
2092 MachineFunction &MF = *MBB.getParent();
2093 unsigned FPReg = RI.getFrameRegister(MF);
2094 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2095 bool isWin64 = TM.getSubtarget<X86Subtarget>().isTargetWin64();
2096 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
2097 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
2098 unsigned Reg = CSI[i].getReg();
2100 // X86RegisterInfo::emitEpilogue will handle restoring of frame register.
2102 const TargetRegisterClass *RegClass = CSI[i].getRegClass();
2103 if (RegClass != &X86::VR128RegClass && !isWin64) {
2104 BuildMI(MBB, MI, DL, get(Opc), Reg);
2106 loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(), RegClass);
2112 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
2113 const SmallVectorImpl<MachineOperand> &MOs,
2115 const TargetInstrInfo &TII) {
2116 // Create the base instruction with the memory operand as the first part.
2117 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2118 MI->getDebugLoc(), true);
2119 MachineInstrBuilder MIB(NewMI);
2120 unsigned NumAddrOps = MOs.size();
2121 for (unsigned i = 0; i != NumAddrOps; ++i)
2122 MIB.addOperand(MOs[i]);
2123 if (NumAddrOps < 4) // FrameIndex only
2126 // Loop over the rest of the ri operands, converting them over.
2127 unsigned NumOps = MI->getDesc().getNumOperands()-2;
2128 for (unsigned i = 0; i != NumOps; ++i) {
2129 MachineOperand &MO = MI->getOperand(i+2);
2132 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
2133 MachineOperand &MO = MI->getOperand(i);
2139 static MachineInstr *FuseInst(MachineFunction &MF,
2140 unsigned Opcode, unsigned OpNo,
2141 const SmallVectorImpl<MachineOperand> &MOs,
2142 MachineInstr *MI, const TargetInstrInfo &TII) {
2143 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2144 MI->getDebugLoc(), true);
2145 MachineInstrBuilder MIB(NewMI);
2147 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2148 MachineOperand &MO = MI->getOperand(i);
2150 assert(MO.isReg() && "Expected to fold into reg operand!");
2151 unsigned NumAddrOps = MOs.size();
2152 for (unsigned i = 0; i != NumAddrOps; ++i)
2153 MIB.addOperand(MOs[i]);
2154 if (NumAddrOps < 4) // FrameIndex only
2163 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
2164 const SmallVectorImpl<MachineOperand> &MOs,
2166 MachineFunction &MF = *MI->getParent()->getParent();
2167 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
2169 unsigned NumAddrOps = MOs.size();
2170 for (unsigned i = 0; i != NumAddrOps; ++i)
2171 MIB.addOperand(MOs[i]);
2172 if (NumAddrOps < 4) // FrameIndex only
2174 return MIB.addImm(0);
2178 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2179 MachineInstr *MI, unsigned i,
2180 const SmallVectorImpl<MachineOperand> &MOs,
2181 unsigned Size, unsigned Align) const {
2182 const DenseMap<unsigned*, std::pair<unsigned,unsigned> > *OpcodeTablePtr=NULL;
2183 bool isTwoAddrFold = false;
2184 unsigned NumOps = MI->getDesc().getNumOperands();
2185 bool isTwoAddr = NumOps > 1 &&
2186 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2188 MachineInstr *NewMI = NULL;
2189 // Folding a memory location into the two-address part of a two-address
2190 // instruction is different than folding it other places. It requires
2191 // replacing the *two* registers with the memory location.
2192 if (isTwoAddr && NumOps >= 2 && i < 2 &&
2193 MI->getOperand(0).isReg() &&
2194 MI->getOperand(1).isReg() &&
2195 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
2196 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2197 isTwoAddrFold = true;
2198 } else if (i == 0) { // If operand 0
2199 if (MI->getOpcode() == X86::MOV16r0)
2200 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
2201 else if (MI->getOpcode() == X86::MOV32r0)
2202 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
2203 else if (MI->getOpcode() == X86::MOV8r0)
2204 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2208 OpcodeTablePtr = &RegOp2MemOpTable0;
2209 } else if (i == 1) {
2210 OpcodeTablePtr = &RegOp2MemOpTable1;
2211 } else if (i == 2) {
2212 OpcodeTablePtr = &RegOp2MemOpTable2;
2215 // If table selected...
2216 if (OpcodeTablePtr) {
2217 // Find the Opcode to fuse
2218 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::const_iterator I =
2219 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
2220 if (I != OpcodeTablePtr->end()) {
2221 unsigned Opcode = I->second.first;
2222 unsigned MinAlign = I->second.second;
2223 if (Align < MinAlign)
2225 bool NarrowToMOV32rm = false;
2227 unsigned RCSize = MI->getDesc().OpInfo[i].getRegClass(&RI)->getSize();
2228 if (Size < RCSize) {
2229 // Check if it's safe to fold the load. If the size of the object is
2230 // narrower than the load width, then it's not.
2231 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
2233 // If this is a 64-bit load, but the spill slot is 32, then we can do
2234 // a 32-bit load which is implicitly zero-extended. This likely is due
2235 // to liveintervalanalysis remat'ing a load from stack slot.
2236 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
2238 Opcode = X86::MOV32rm;
2239 NarrowToMOV32rm = true;
2244 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
2246 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
2248 if (NarrowToMOV32rm) {
2249 // If this is the special case where we use a MOV32rm to load a 32-bit
2250 // value and zero-extend the top bits. Change the destination register
2252 unsigned DstReg = NewMI->getOperand(0).getReg();
2253 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
2254 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
2255 4/*x86_subreg_32bit*/));
2257 NewMI->getOperand(0).setSubReg(4/*x86_subreg_32bit*/);
2264 if (PrintFailedFusing)
2265 errs() << "We failed to fuse operand " << i << " in " << *MI;
2270 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2272 const SmallVectorImpl<unsigned> &Ops,
2273 int FrameIndex) const {
2274 // Check switch flag
2275 if (NoFusing) return NULL;
2277 const MachineFrameInfo *MFI = MF.getFrameInfo();
2278 unsigned Size = MFI->getObjectSize(FrameIndex);
2279 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2280 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2281 unsigned NewOpc = 0;
2282 unsigned RCSize = 0;
2283 switch (MI->getOpcode()) {
2284 default: return NULL;
2285 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
2286 case X86::TEST16rr: NewOpc = X86::CMP16ri; RCSize = 2; break;
2287 case X86::TEST32rr: NewOpc = X86::CMP32ri; RCSize = 4; break;
2288 case X86::TEST64rr: NewOpc = X86::CMP64ri32; RCSize = 8; break;
2290 // Check if it's safe to fold the load. If the size of the object is
2291 // narrower than the load width, then it's not.
2294 // Change to CMPXXri r, 0 first.
2295 MI->setDesc(get(NewOpc));
2296 MI->getOperand(1).ChangeToImmediate(0);
2297 } else if (Ops.size() != 1)
2300 SmallVector<MachineOperand,4> MOs;
2301 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2302 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
2305 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2307 const SmallVectorImpl<unsigned> &Ops,
2308 MachineInstr *LoadMI) const {
2309 // Check switch flag
2310 if (NoFusing) return NULL;
2312 // Determine the alignment of the load.
2313 unsigned Alignment = 0;
2314 if (LoadMI->hasOneMemOperand())
2315 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
2317 switch (LoadMI->getOpcode()) {
2319 case X86::V_SETALLONES:
2329 llvm_unreachable("Don't know how to fold this instruction!");
2331 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2332 unsigned NewOpc = 0;
2333 switch (MI->getOpcode()) {
2334 default: return NULL;
2335 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2336 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2337 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2338 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2340 // Change to CMPXXri r, 0 first.
2341 MI->setDesc(get(NewOpc));
2342 MI->getOperand(1).ChangeToImmediate(0);
2343 } else if (Ops.size() != 1)
2346 SmallVector<MachineOperand,X86AddrNumOperands> MOs;
2347 switch (LoadMI->getOpcode()) {
2349 case X86::V_SETALLONES:
2351 case X86::FsFLD0SS: {
2352 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
2353 // Create a constant-pool entry and operands to load from it.
2355 // x86-32 PIC requires a PIC base register for constant pools.
2356 unsigned PICBase = 0;
2357 if (TM.getRelocationModel() == Reloc::PIC_) {
2358 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2361 // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
2362 // This doesn't work for several reasons.
2363 // 1. GlobalBaseReg may have been spilled.
2364 // 2. It may not be live at MI.
2368 // Create a constant-pool entry.
2369 MachineConstantPool &MCP = *MF.getConstantPool();
2371 if (LoadMI->getOpcode() == X86::FsFLD0SS)
2372 Ty = Type::getFloatTy(MF.getFunction()->getContext());
2373 else if (LoadMI->getOpcode() == X86::FsFLD0SD)
2374 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
2376 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
2377 Constant *C = LoadMI->getOpcode() == X86::V_SETALLONES ?
2378 Constant::getAllOnesValue(Ty) :
2379 Constant::getNullValue(Ty);
2380 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
2382 // Create operands to load from the constant pool entry.
2383 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
2384 MOs.push_back(MachineOperand::CreateImm(1));
2385 MOs.push_back(MachineOperand::CreateReg(0, false));
2386 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
2387 MOs.push_back(MachineOperand::CreateReg(0, false));
2391 // Folding a normal load. Just copy the load's address operands.
2392 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2393 for (unsigned i = NumOps - X86AddrNumOperands; i != NumOps; ++i)
2394 MOs.push_back(LoadMI->getOperand(i));
2398 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
2402 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2403 const SmallVectorImpl<unsigned> &Ops) const {
2404 // Check switch flag
2405 if (NoFusing) return 0;
2407 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2408 switch (MI->getOpcode()) {
2409 default: return false;
2418 if (Ops.size() != 1)
2421 unsigned OpNum = Ops[0];
2422 unsigned Opc = MI->getOpcode();
2423 unsigned NumOps = MI->getDesc().getNumOperands();
2424 bool isTwoAddr = NumOps > 1 &&
2425 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2427 // Folding a memory location into the two-address part of a two-address
2428 // instruction is different than folding it other places. It requires
2429 // replacing the *two* registers with the memory location.
2430 const DenseMap<unsigned*, std::pair<unsigned,unsigned> > *OpcodeTablePtr=NULL;
2431 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2432 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2433 } else if (OpNum == 0) { // If operand 0
2441 OpcodeTablePtr = &RegOp2MemOpTable0;
2442 } else if (OpNum == 1) {
2443 OpcodeTablePtr = &RegOp2MemOpTable1;
2444 } else if (OpNum == 2) {
2445 OpcodeTablePtr = &RegOp2MemOpTable2;
2448 if (OpcodeTablePtr) {
2449 // Find the Opcode to fuse
2450 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::const_iterator I =
2451 OpcodeTablePtr->find((unsigned*)Opc);
2452 if (I != OpcodeTablePtr->end())
2458 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2459 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2460 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2461 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::const_iterator I =
2462 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2463 if (I == MemOp2RegOpTable.end())
2465 DebugLoc dl = MI->getDebugLoc();
2466 unsigned Opc = I->second.first;
2467 unsigned Index = I->second.second & 0xf;
2468 bool FoldedLoad = I->second.second & (1 << 4);
2469 bool FoldedStore = I->second.second & (1 << 5);
2470 if (UnfoldLoad && !FoldedLoad)
2472 UnfoldLoad &= FoldedLoad;
2473 if (UnfoldStore && !FoldedStore)
2475 UnfoldStore &= FoldedStore;
2477 const TargetInstrDesc &TID = get(Opc);
2478 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2479 const TargetRegisterClass *RC = TOI.getRegClass(&RI);
2480 SmallVector<MachineOperand, X86AddrNumOperands> AddrOps;
2481 SmallVector<MachineOperand,2> BeforeOps;
2482 SmallVector<MachineOperand,2> AfterOps;
2483 SmallVector<MachineOperand,4> ImpOps;
2484 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2485 MachineOperand &Op = MI->getOperand(i);
2486 if (i >= Index && i < Index + X86AddrNumOperands)
2487 AddrOps.push_back(Op);
2488 else if (Op.isReg() && Op.isImplicit())
2489 ImpOps.push_back(Op);
2491 BeforeOps.push_back(Op);
2493 AfterOps.push_back(Op);
2496 // Emit the load instruction.
2498 std::pair<MachineInstr::mmo_iterator,
2499 MachineInstr::mmo_iterator> MMOs =
2500 MF.extractLoadMemRefs(MI->memoperands_begin(),
2501 MI->memoperands_end());
2502 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
2504 // Address operands cannot be marked isKill.
2505 for (unsigned i = 1; i != 1 + X86AddrNumOperands; ++i) {
2506 MachineOperand &MO = NewMIs[0]->getOperand(i);
2508 MO.setIsKill(false);
2513 // Emit the data processing instruction.
2514 MachineInstr *DataMI = MF.CreateMachineInstr(TID, MI->getDebugLoc(), true);
2515 MachineInstrBuilder MIB(DataMI);
2518 MIB.addReg(Reg, RegState::Define);
2519 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2520 MIB.addOperand(BeforeOps[i]);
2523 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2524 MIB.addOperand(AfterOps[i]);
2525 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2526 MachineOperand &MO = ImpOps[i];
2527 MIB.addReg(MO.getReg(),
2528 getDefRegState(MO.isDef()) |
2529 RegState::Implicit |
2530 getKillRegState(MO.isKill()) |
2531 getDeadRegState(MO.isDead()) |
2532 getUndefRegState(MO.isUndef()));
2534 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2535 unsigned NewOpc = 0;
2536 switch (DataMI->getOpcode()) {
2538 case X86::CMP64ri32:
2542 MachineOperand &MO0 = DataMI->getOperand(0);
2543 MachineOperand &MO1 = DataMI->getOperand(1);
2544 if (MO1.getImm() == 0) {
2545 switch (DataMI->getOpcode()) {
2547 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2548 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2549 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2550 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2552 DataMI->setDesc(get(NewOpc));
2553 MO1.ChangeToRegister(MO0.getReg(), false);
2557 NewMIs.push_back(DataMI);
2559 // Emit the store instruction.
2561 const TargetRegisterClass *DstRC = TID.OpInfo[0].getRegClass(&RI);
2562 std::pair<MachineInstr::mmo_iterator,
2563 MachineInstr::mmo_iterator> MMOs =
2564 MF.extractStoreMemRefs(MI->memoperands_begin(),
2565 MI->memoperands_end());
2566 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
2573 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2574 SmallVectorImpl<SDNode*> &NewNodes) const {
2575 if (!N->isMachineOpcode())
2578 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::const_iterator I =
2579 MemOp2RegOpTable.find((unsigned*)N->getMachineOpcode());
2580 if (I == MemOp2RegOpTable.end())
2582 unsigned Opc = I->second.first;
2583 unsigned Index = I->second.second & 0xf;
2584 bool FoldedLoad = I->second.second & (1 << 4);
2585 bool FoldedStore = I->second.second & (1 << 5);
2586 const TargetInstrDesc &TID = get(Opc);
2587 const TargetRegisterClass *RC = TID.OpInfo[Index].getRegClass(&RI);
2588 unsigned NumDefs = TID.NumDefs;
2589 std::vector<SDValue> AddrOps;
2590 std::vector<SDValue> BeforeOps;
2591 std::vector<SDValue> AfterOps;
2592 DebugLoc dl = N->getDebugLoc();
2593 unsigned NumOps = N->getNumOperands();
2594 for (unsigned i = 0; i != NumOps-1; ++i) {
2595 SDValue Op = N->getOperand(i);
2596 if (i >= Index-NumDefs && i < Index-NumDefs + X86AddrNumOperands)
2597 AddrOps.push_back(Op);
2598 else if (i < Index-NumDefs)
2599 BeforeOps.push_back(Op);
2600 else if (i > Index-NumDefs)
2601 AfterOps.push_back(Op);
2603 SDValue Chain = N->getOperand(NumOps-1);
2604 AddrOps.push_back(Chain);
2606 // Emit the load instruction.
2608 MachineFunction &MF = DAG.getMachineFunction();
2610 EVT VT = *RC->vt_begin();
2611 bool isAligned = (RI.getStackAlignment() >= 16) ||
2612 RI.needsStackRealignment(MF);
2613 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
2614 VT, MVT::Other, &AddrOps[0], AddrOps.size());
2615 NewNodes.push_back(Load);
2617 // Preserve memory reference information.
2618 std::pair<MachineInstr::mmo_iterator,
2619 MachineInstr::mmo_iterator> MMOs =
2620 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
2621 cast<MachineSDNode>(N)->memoperands_end());
2622 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
2625 // Emit the data processing instruction.
2626 std::vector<EVT> VTs;
2627 const TargetRegisterClass *DstRC = 0;
2628 if (TID.getNumDefs() > 0) {
2629 DstRC = TID.OpInfo[0].getRegClass(&RI);
2630 VTs.push_back(*DstRC->vt_begin());
2632 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2633 EVT VT = N->getValueType(i);
2634 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2638 BeforeOps.push_back(SDValue(Load, 0));
2639 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2640 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
2642 NewNodes.push_back(NewNode);
2644 // Emit the store instruction.
2647 AddrOps.push_back(SDValue(NewNode, 0));
2648 AddrOps.push_back(Chain);
2649 bool isAligned = (RI.getStackAlignment() >= 16) ||
2650 RI.needsStackRealignment(MF);
2651 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
2654 &AddrOps[0], AddrOps.size());
2655 NewNodes.push_back(Store);
2657 // Preserve memory reference information.
2658 std::pair<MachineInstr::mmo_iterator,
2659 MachineInstr::mmo_iterator> MMOs =
2660 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
2661 cast<MachineSDNode>(N)->memoperands_end());
2662 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
2668 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2669 bool UnfoldLoad, bool UnfoldStore,
2670 unsigned *LoadRegIndex) const {
2671 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::const_iterator I =
2672 MemOp2RegOpTable.find((unsigned*)Opc);
2673 if (I == MemOp2RegOpTable.end())
2675 bool FoldedLoad = I->second.second & (1 << 4);
2676 bool FoldedStore = I->second.second & (1 << 5);
2677 if (UnfoldLoad && !FoldedLoad)
2679 if (UnfoldStore && !FoldedStore)
2682 *LoadRegIndex = I->second.second & 0xf;
2683 return I->second.first;
2686 bool X86InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
2687 if (MBB.empty()) return false;
2689 switch (MBB.back().getOpcode()) {
2690 case X86::TCRETURNri:
2691 case X86::TCRETURNdi:
2692 case X86::RET: // Return.
2697 case X86::JMP: // Uncond branch.
2698 case X86::JMP32r: // Indirect branch.
2699 case X86::JMP64r: // Indirect branch (64-bit).
2700 case X86::JMP32m: // Indirect branch through mem.
2701 case X86::JMP64m: // Indirect branch through mem (64-bit).
2703 default: return false;
2708 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
2709 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2710 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
2711 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
2713 Cond[0].setImm(GetOppositeBranchCondition(CC));
2718 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
2719 // FIXME: Return false for x87 stack register classes for now. We can't
2720 // allow any loads of these registers before FpGet_ST0_80.
2721 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
2722 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
2725 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2726 switch (Desc->TSFlags & X86II::ImmMask) {
2727 case X86II::Imm8: return 1;
2728 case X86II::Imm16: return 2;
2729 case X86II::Imm32: return 4;
2730 case X86II::Imm64: return 8;
2731 default: llvm_unreachable("Immediate size not set!");
2736 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2737 /// e.g. r8, xmm8, etc.
2738 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2739 if (!MO.isReg()) return false;
2740 switch (MO.getReg()) {
2742 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2743 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2744 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2745 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2746 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2747 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2748 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2749 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2750 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2751 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2758 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2759 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2760 /// size, and 3) use of X86-64 extended registers.
2761 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2763 const TargetInstrDesc &Desc = MI.getDesc();
2765 // Pseudo instructions do not need REX prefix byte.
2766 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2768 if (Desc.TSFlags & X86II::REX_W)
2771 unsigned NumOps = Desc.getNumOperands();
2773 bool isTwoAddr = NumOps > 1 &&
2774 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2776 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2777 unsigned i = isTwoAddr ? 1 : 0;
2778 for (unsigned e = NumOps; i != e; ++i) {
2779 const MachineOperand& MO = MI.getOperand(i);
2781 unsigned Reg = MO.getReg();
2782 if (isX86_64NonExtLowByteReg(Reg))
2787 switch (Desc.TSFlags & X86II::FormMask) {
2788 case X86II::MRMInitReg:
2789 if (isX86_64ExtendedReg(MI.getOperand(0)))
2790 REX |= (1 << 0) | (1 << 2);
2792 case X86II::MRMSrcReg: {
2793 if (isX86_64ExtendedReg(MI.getOperand(0)))
2795 i = isTwoAddr ? 2 : 1;
2796 for (unsigned e = NumOps; i != e; ++i) {
2797 const MachineOperand& MO = MI.getOperand(i);
2798 if (isX86_64ExtendedReg(MO))
2803 case X86II::MRMSrcMem: {
2804 if (isX86_64ExtendedReg(MI.getOperand(0)))
2807 i = isTwoAddr ? 2 : 1;
2808 for (; i != NumOps; ++i) {
2809 const MachineOperand& MO = MI.getOperand(i);
2811 if (isX86_64ExtendedReg(MO))
2818 case X86II::MRM0m: case X86II::MRM1m:
2819 case X86II::MRM2m: case X86II::MRM3m:
2820 case X86II::MRM4m: case X86II::MRM5m:
2821 case X86II::MRM6m: case X86II::MRM7m:
2822 case X86II::MRMDestMem: {
2823 unsigned e = (isTwoAddr ? X86AddrNumOperands+1 : X86AddrNumOperands);
2824 i = isTwoAddr ? 1 : 0;
2825 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2828 for (; i != e; ++i) {
2829 const MachineOperand& MO = MI.getOperand(i);
2831 if (isX86_64ExtendedReg(MO))
2839 if (isX86_64ExtendedReg(MI.getOperand(0)))
2841 i = isTwoAddr ? 2 : 1;
2842 for (unsigned e = NumOps; i != e; ++i) {
2843 const MachineOperand& MO = MI.getOperand(i);
2844 if (isX86_64ExtendedReg(MO))
2854 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2855 /// relative block address instruction
2857 static unsigned sizePCRelativeBlockAddress() {
2861 /// sizeGlobalAddress - Give the size of the emission of this global address
2863 static unsigned sizeGlobalAddress(bool dword) {
2864 return dword ? 8 : 4;
2867 /// sizeConstPoolAddress - Give the size of the emission of this constant
2870 static unsigned sizeConstPoolAddress(bool dword) {
2871 return dword ? 8 : 4;
2874 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2877 static unsigned sizeExternalSymbolAddress(bool dword) {
2878 return dword ? 8 : 4;
2881 /// sizeJumpTableAddress - Give the size of the emission of this jump
2884 static unsigned sizeJumpTableAddress(bool dword) {
2885 return dword ? 8 : 4;
2888 static unsigned sizeConstant(unsigned Size) {
2892 static unsigned sizeRegModRMByte(){
2896 static unsigned sizeSIBByte(){
2900 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2901 unsigned FinalSize = 0;
2902 // If this is a simple integer displacement that doesn't require a relocation.
2904 FinalSize += sizeConstant(4);
2908 // Otherwise, this is something that requires a relocation.
2909 if (RelocOp->isGlobal()) {
2910 FinalSize += sizeGlobalAddress(false);
2911 } else if (RelocOp->isCPI()) {
2912 FinalSize += sizeConstPoolAddress(false);
2913 } else if (RelocOp->isJTI()) {
2914 FinalSize += sizeJumpTableAddress(false);
2916 llvm_unreachable("Unknown value to relocate!");
2921 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2922 bool IsPIC, bool Is64BitMode) {
2923 const MachineOperand &Op3 = MI.getOperand(Op+3);
2925 const MachineOperand *DispForReloc = 0;
2926 unsigned FinalSize = 0;
2928 // Figure out what sort of displacement we have to handle here.
2929 if (Op3.isGlobal()) {
2930 DispForReloc = &Op3;
2931 } else if (Op3.isCPI()) {
2932 if (Is64BitMode || IsPIC) {
2933 DispForReloc = &Op3;
2937 } else if (Op3.isJTI()) {
2938 if (Is64BitMode || IsPIC) {
2939 DispForReloc = &Op3;
2947 const MachineOperand &Base = MI.getOperand(Op);
2948 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2950 unsigned BaseReg = Base.getReg();
2952 // Is a SIB byte needed?
2953 if ((!Is64BitMode || DispForReloc || BaseReg != 0) &&
2954 IndexReg.getReg() == 0 &&
2955 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2956 if (BaseReg == 0) { // Just a displacement?
2957 // Emit special case [disp32] encoding
2959 FinalSize += getDisplacementFieldSize(DispForReloc);
2961 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2962 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2963 // Emit simple indirect register encoding... [EAX] f.e.
2965 // Be pessimistic and assume it's a disp32, not a disp8
2967 // Emit the most general non-SIB encoding: [REG+disp32]
2969 FinalSize += getDisplacementFieldSize(DispForReloc);
2973 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2974 assert(IndexReg.getReg() != X86::ESP &&
2975 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2977 bool ForceDisp32 = false;
2978 if (BaseReg == 0 || DispForReloc) {
2979 // Emit the normal disp32 encoding.
2986 FinalSize += sizeSIBByte();
2988 // Do we need to output a displacement?
2989 if (DispVal != 0 || ForceDisp32) {
2990 FinalSize += getDisplacementFieldSize(DispForReloc);
2997 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2998 const TargetInstrDesc *Desc,
2999 bool IsPIC, bool Is64BitMode) {
3001 unsigned Opcode = Desc->Opcode;
3002 unsigned FinalSize = 0;
3004 // Emit the lock opcode prefix as needed.
3005 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
3007 // Emit segment override opcode prefix as needed.
3008 switch (Desc->TSFlags & X86II::SegOvrMask) {
3013 default: llvm_unreachable("Invalid segment!");
3014 case 0: break; // No segment override!
3017 // Emit the repeat opcode prefix as needed.
3018 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
3020 // Emit the operand size opcode prefix as needed.
3021 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
3023 // Emit the address size opcode prefix as needed.
3024 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
3026 bool Need0FPrefix = false;
3027 switch (Desc->TSFlags & X86II::Op0Mask) {
3028 case X86II::TB: // Two-byte opcode prefix
3029 case X86II::T8: // 0F 38
3030 case X86II::TA: // 0F 3A
3031 Need0FPrefix = true;
3033 case X86II::TF: // F2 0F 38
3035 Need0FPrefix = true;
3037 case X86II::REP: break; // already handled.
3038 case X86II::XS: // F3 0F
3040 Need0FPrefix = true;
3042 case X86II::XD: // F2 0F
3044 Need0FPrefix = true;
3046 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
3047 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
3049 break; // Two-byte opcode prefix
3050 default: llvm_unreachable("Invalid prefix!");
3051 case 0: break; // No prefix!
3056 unsigned REX = X86InstrInfo::determineREX(MI);
3061 // 0x0F escape code must be emitted just before the opcode.
3065 switch (Desc->TSFlags & X86II::Op0Mask) {
3066 case X86II::T8: // 0F 38
3069 case X86II::TA: // 0F 3A
3072 case X86II::TF: // F2 0F 38
3077 // If this is a two-address instruction, skip one of the register operands.
3078 unsigned NumOps = Desc->getNumOperands();
3080 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
3082 else if (NumOps > 2 && Desc->getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0)
3083 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
3086 switch (Desc->TSFlags & X86II::FormMask) {
3087 default: llvm_unreachable("Unknown FormMask value in X86 MachineCodeEmitter!");
3089 // Remember the current PC offset, this is the PIC relocation
3094 case TargetInstrInfo::INLINEASM: {
3095 const MachineFunction *MF = MI.getParent()->getParent();
3096 const TargetInstrInfo &TII = *MF->getTarget().getInstrInfo();
3097 FinalSize += TII.getInlineAsmLength(MI.getOperand(0).getSymbolName(),
3098 *MF->getTarget().getMCAsmInfo());
3101 case TargetInstrInfo::DBG_LABEL:
3102 case TargetInstrInfo::EH_LABEL:
3104 case TargetInstrInfo::IMPLICIT_DEF:
3105 case TargetInstrInfo::KILL:
3106 case X86::DWARF_LOC:
3107 case X86::FP_REG_KILL:
3109 case X86::MOVPC32r: {
3110 // This emits the "call" portion of this pseudo instruction.
3112 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3121 if (CurOp != NumOps) {
3122 const MachineOperand &MO = MI.getOperand(CurOp++);
3124 FinalSize += sizePCRelativeBlockAddress();
3125 } else if (MO.isGlobal()) {
3126 FinalSize += sizeGlobalAddress(false);
3127 } else if (MO.isSymbol()) {
3128 FinalSize += sizeExternalSymbolAddress(false);
3129 } else if (MO.isImm()) {
3130 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3132 llvm_unreachable("Unknown RawFrm operand!");
3137 case X86II::AddRegFrm:
3141 if (CurOp != NumOps) {
3142 const MachineOperand &MO1 = MI.getOperand(CurOp++);
3143 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3145 FinalSize += sizeConstant(Size);
3148 if (Opcode == X86::MOV64ri)
3150 if (MO1.isGlobal()) {
3151 FinalSize += sizeGlobalAddress(dword);
3152 } else if (MO1.isSymbol())
3153 FinalSize += sizeExternalSymbolAddress(dword);
3154 else if (MO1.isCPI())
3155 FinalSize += sizeConstPoolAddress(dword);
3156 else if (MO1.isJTI())
3157 FinalSize += sizeJumpTableAddress(dword);
3162 case X86II::MRMDestReg: {
3164 FinalSize += sizeRegModRMByte();
3166 if (CurOp != NumOps) {
3168 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3172 case X86II::MRMDestMem: {
3174 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
3175 CurOp += X86AddrNumOperands + 1;
3176 if (CurOp != NumOps) {
3178 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3183 case X86II::MRMSrcReg:
3185 FinalSize += sizeRegModRMByte();
3187 if (CurOp != NumOps) {
3189 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3193 case X86II::MRMSrcMem: {
3195 if (Opcode == X86::LEA64r || Opcode == X86::LEA64_32r ||
3196 Opcode == X86::LEA16r || Opcode == X86::LEA32r)
3197 AddrOperands = X86AddrNumOperands - 1; // No segment register
3199 AddrOperands = X86AddrNumOperands;
3202 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
3203 CurOp += AddrOperands + 1;
3204 if (CurOp != NumOps) {
3206 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3211 case X86II::MRM0r: case X86II::MRM1r:
3212 case X86II::MRM2r: case X86II::MRM3r:
3213 case X86II::MRM4r: case X86II::MRM5r:
3214 case X86II::MRM6r: case X86II::MRM7r:
3216 if (Desc->getOpcode() == X86::LFENCE ||
3217 Desc->getOpcode() == X86::MFENCE) {
3218 // Special handling of lfence and mfence;
3219 FinalSize += sizeRegModRMByte();
3220 } else if (Desc->getOpcode() == X86::MONITOR ||
3221 Desc->getOpcode() == X86::MWAIT) {
3222 // Special handling of monitor and mwait.
3223 FinalSize += sizeRegModRMByte() + 1; // +1 for the opcode.
3226 FinalSize += sizeRegModRMByte();
3229 if (CurOp != NumOps) {
3230 const MachineOperand &MO1 = MI.getOperand(CurOp++);
3231 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3233 FinalSize += sizeConstant(Size);
3236 if (Opcode == X86::MOV64ri32)
3238 if (MO1.isGlobal()) {
3239 FinalSize += sizeGlobalAddress(dword);
3240 } else if (MO1.isSymbol())
3241 FinalSize += sizeExternalSymbolAddress(dword);
3242 else if (MO1.isCPI())
3243 FinalSize += sizeConstPoolAddress(dword);
3244 else if (MO1.isJTI())
3245 FinalSize += sizeJumpTableAddress(dword);
3250 case X86II::MRM0m: case X86II::MRM1m:
3251 case X86II::MRM2m: case X86II::MRM3m:
3252 case X86II::MRM4m: case X86II::MRM5m:
3253 case X86II::MRM6m: case X86II::MRM7m: {
3256 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
3257 CurOp += X86AddrNumOperands;
3259 if (CurOp != NumOps) {
3260 const MachineOperand &MO = MI.getOperand(CurOp++);
3261 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3263 FinalSize += sizeConstant(Size);
3266 if (Opcode == X86::MOV64mi32)
3268 if (MO.isGlobal()) {
3269 FinalSize += sizeGlobalAddress(dword);
3270 } else if (MO.isSymbol())
3271 FinalSize += sizeExternalSymbolAddress(dword);
3272 else if (MO.isCPI())
3273 FinalSize += sizeConstPoolAddress(dword);
3274 else if (MO.isJTI())
3275 FinalSize += sizeJumpTableAddress(dword);
3281 case X86II::MRMInitReg:
3283 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
3284 FinalSize += sizeRegModRMByte();
3289 if (!Desc->isVariadic() && CurOp != NumOps) {
3291 raw_string_ostream Msg(msg);
3292 Msg << "Cannot determine size: " << MI;
3293 llvm_report_error(Msg.str());
3301 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
3302 const TargetInstrDesc &Desc = MI->getDesc();
3303 bool IsPIC = TM.getRelocationModel() == Reloc::PIC_;
3304 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
3305 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
3306 if (Desc.getOpcode() == X86::MOVPC32r)
3307 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);
3311 /// getGlobalBaseReg - Return a virtual register initialized with the
3312 /// the global base register value. Output instructions required to
3313 /// initialize the register in the function entry block, if necessary.
3315 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3316 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3317 "X86-64 PIC uses RIP relative addressing");
3319 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3320 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3321 if (GlobalBaseReg != 0)
3322 return GlobalBaseReg;
3324 // Insert the set of GlobalBaseReg into the first MBB of the function
3325 MachineBasicBlock &FirstMBB = MF->front();
3326 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3327 DebugLoc DL = DebugLoc::getUnknownLoc();
3328 if (MBBI != FirstMBB.end()) DL = MBBI->getDebugLoc();
3329 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3330 unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3332 const TargetInstrInfo *TII = TM.getInstrInfo();
3333 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3334 // only used in JIT code emission as displacement to pc.
3335 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
3337 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3338 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
3339 if (TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3340 GlobalBaseReg = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3341 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
3342 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
3343 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
3344 X86II::MO_GOT_ABSOLUTE_ADDRESS);
3349 X86FI->setGlobalBaseReg(GlobalBaseReg);
3350 return GlobalBaseReg;