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/GlobalVariable.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/LLVMContext.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/LiveVariables.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/Target/TargetAsmInfo.h"
39 NoFusing("disable-spill-fusing",
40 cl::desc("Disable fusing of spill code into instructions"));
42 PrintFailedFusing("print-failed-fuse-candidates",
43 cl::desc("Print instructions that the allocator wants to"
44 " fuse, but the X86 backend currently can't"),
47 ReMatPICStubLoad("remat-pic-stub-load",
48 cl::desc("Re-materialize load from stub in PIC mode"),
49 cl::init(false), cl::Hidden);
52 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
53 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
54 TM(tm), RI(tm, *this) {
55 SmallVector<unsigned,16> AmbEntries;
56 static const unsigned OpTbl2Addr[][2] = {
57 { X86::ADC32ri, X86::ADC32mi },
58 { X86::ADC32ri8, X86::ADC32mi8 },
59 { X86::ADC32rr, X86::ADC32mr },
60 { X86::ADC64ri32, X86::ADC64mi32 },
61 { X86::ADC64ri8, X86::ADC64mi8 },
62 { X86::ADC64rr, X86::ADC64mr },
63 { X86::ADD16ri, X86::ADD16mi },
64 { X86::ADD16ri8, X86::ADD16mi8 },
65 { X86::ADD16rr, X86::ADD16mr },
66 { X86::ADD32ri, X86::ADD32mi },
67 { X86::ADD32ri8, X86::ADD32mi8 },
68 { X86::ADD32rr, X86::ADD32mr },
69 { X86::ADD64ri32, X86::ADD64mi32 },
70 { X86::ADD64ri8, X86::ADD64mi8 },
71 { X86::ADD64rr, X86::ADD64mr },
72 { X86::ADD8ri, X86::ADD8mi },
73 { X86::ADD8rr, X86::ADD8mr },
74 { X86::AND16ri, X86::AND16mi },
75 { X86::AND16ri8, X86::AND16mi8 },
76 { X86::AND16rr, X86::AND16mr },
77 { X86::AND32ri, X86::AND32mi },
78 { X86::AND32ri8, X86::AND32mi8 },
79 { X86::AND32rr, X86::AND32mr },
80 { X86::AND64ri32, X86::AND64mi32 },
81 { X86::AND64ri8, X86::AND64mi8 },
82 { X86::AND64rr, X86::AND64mr },
83 { X86::AND8ri, X86::AND8mi },
84 { X86::AND8rr, X86::AND8mr },
85 { X86::DEC16r, X86::DEC16m },
86 { X86::DEC32r, X86::DEC32m },
87 { X86::DEC64_16r, X86::DEC64_16m },
88 { X86::DEC64_32r, X86::DEC64_32m },
89 { X86::DEC64r, X86::DEC64m },
90 { X86::DEC8r, X86::DEC8m },
91 { X86::INC16r, X86::INC16m },
92 { X86::INC32r, X86::INC32m },
93 { X86::INC64_16r, X86::INC64_16m },
94 { X86::INC64_32r, X86::INC64_32m },
95 { X86::INC64r, X86::INC64m },
96 { X86::INC8r, X86::INC8m },
97 { X86::NEG16r, X86::NEG16m },
98 { X86::NEG32r, X86::NEG32m },
99 { X86::NEG64r, X86::NEG64m },
100 { X86::NEG8r, X86::NEG8m },
101 { X86::NOT16r, X86::NOT16m },
102 { X86::NOT32r, X86::NOT32m },
103 { X86::NOT64r, X86::NOT64m },
104 { X86::NOT8r, X86::NOT8m },
105 { X86::OR16ri, X86::OR16mi },
106 { X86::OR16ri8, X86::OR16mi8 },
107 { X86::OR16rr, X86::OR16mr },
108 { X86::OR32ri, X86::OR32mi },
109 { X86::OR32ri8, X86::OR32mi8 },
110 { X86::OR32rr, X86::OR32mr },
111 { X86::OR64ri32, X86::OR64mi32 },
112 { X86::OR64ri8, X86::OR64mi8 },
113 { X86::OR64rr, X86::OR64mr },
114 { X86::OR8ri, X86::OR8mi },
115 { X86::OR8rr, X86::OR8mr },
116 { X86::ROL16r1, X86::ROL16m1 },
117 { X86::ROL16rCL, X86::ROL16mCL },
118 { X86::ROL16ri, X86::ROL16mi },
119 { X86::ROL32r1, X86::ROL32m1 },
120 { X86::ROL32rCL, X86::ROL32mCL },
121 { X86::ROL32ri, X86::ROL32mi },
122 { X86::ROL64r1, X86::ROL64m1 },
123 { X86::ROL64rCL, X86::ROL64mCL },
124 { X86::ROL64ri, X86::ROL64mi },
125 { X86::ROL8r1, X86::ROL8m1 },
126 { X86::ROL8rCL, X86::ROL8mCL },
127 { X86::ROL8ri, X86::ROL8mi },
128 { X86::ROR16r1, X86::ROR16m1 },
129 { X86::ROR16rCL, X86::ROR16mCL },
130 { X86::ROR16ri, X86::ROR16mi },
131 { X86::ROR32r1, X86::ROR32m1 },
132 { X86::ROR32rCL, X86::ROR32mCL },
133 { X86::ROR32ri, X86::ROR32mi },
134 { X86::ROR64r1, X86::ROR64m1 },
135 { X86::ROR64rCL, X86::ROR64mCL },
136 { X86::ROR64ri, X86::ROR64mi },
137 { X86::ROR8r1, X86::ROR8m1 },
138 { X86::ROR8rCL, X86::ROR8mCL },
139 { X86::ROR8ri, X86::ROR8mi },
140 { X86::SAR16r1, X86::SAR16m1 },
141 { X86::SAR16rCL, X86::SAR16mCL },
142 { X86::SAR16ri, X86::SAR16mi },
143 { X86::SAR32r1, X86::SAR32m1 },
144 { X86::SAR32rCL, X86::SAR32mCL },
145 { X86::SAR32ri, X86::SAR32mi },
146 { X86::SAR64r1, X86::SAR64m1 },
147 { X86::SAR64rCL, X86::SAR64mCL },
148 { X86::SAR64ri, X86::SAR64mi },
149 { X86::SAR8r1, X86::SAR8m1 },
150 { X86::SAR8rCL, X86::SAR8mCL },
151 { X86::SAR8ri, X86::SAR8mi },
152 { X86::SBB32ri, X86::SBB32mi },
153 { X86::SBB32ri8, X86::SBB32mi8 },
154 { X86::SBB32rr, X86::SBB32mr },
155 { X86::SBB64ri32, X86::SBB64mi32 },
156 { X86::SBB64ri8, X86::SBB64mi8 },
157 { X86::SBB64rr, X86::SBB64mr },
158 { X86::SHL16rCL, X86::SHL16mCL },
159 { X86::SHL16ri, X86::SHL16mi },
160 { X86::SHL32rCL, X86::SHL32mCL },
161 { X86::SHL32ri, X86::SHL32mi },
162 { X86::SHL64rCL, X86::SHL64mCL },
163 { X86::SHL64ri, X86::SHL64mi },
164 { X86::SHL8rCL, X86::SHL8mCL },
165 { X86::SHL8ri, X86::SHL8mi },
166 { X86::SHLD16rrCL, X86::SHLD16mrCL },
167 { X86::SHLD16rri8, X86::SHLD16mri8 },
168 { X86::SHLD32rrCL, X86::SHLD32mrCL },
169 { X86::SHLD32rri8, X86::SHLD32mri8 },
170 { X86::SHLD64rrCL, X86::SHLD64mrCL },
171 { X86::SHLD64rri8, X86::SHLD64mri8 },
172 { X86::SHR16r1, X86::SHR16m1 },
173 { X86::SHR16rCL, X86::SHR16mCL },
174 { X86::SHR16ri, X86::SHR16mi },
175 { X86::SHR32r1, X86::SHR32m1 },
176 { X86::SHR32rCL, X86::SHR32mCL },
177 { X86::SHR32ri, X86::SHR32mi },
178 { X86::SHR64r1, X86::SHR64m1 },
179 { X86::SHR64rCL, X86::SHR64mCL },
180 { X86::SHR64ri, X86::SHR64mi },
181 { X86::SHR8r1, X86::SHR8m1 },
182 { X86::SHR8rCL, X86::SHR8mCL },
183 { X86::SHR8ri, X86::SHR8mi },
184 { X86::SHRD16rrCL, X86::SHRD16mrCL },
185 { X86::SHRD16rri8, X86::SHRD16mri8 },
186 { X86::SHRD32rrCL, X86::SHRD32mrCL },
187 { X86::SHRD32rri8, X86::SHRD32mri8 },
188 { X86::SHRD64rrCL, X86::SHRD64mrCL },
189 { X86::SHRD64rri8, X86::SHRD64mri8 },
190 { X86::SUB16ri, X86::SUB16mi },
191 { X86::SUB16ri8, X86::SUB16mi8 },
192 { X86::SUB16rr, X86::SUB16mr },
193 { X86::SUB32ri, X86::SUB32mi },
194 { X86::SUB32ri8, X86::SUB32mi8 },
195 { X86::SUB32rr, X86::SUB32mr },
196 { X86::SUB64ri32, X86::SUB64mi32 },
197 { X86::SUB64ri8, X86::SUB64mi8 },
198 { X86::SUB64rr, X86::SUB64mr },
199 { X86::SUB8ri, X86::SUB8mi },
200 { X86::SUB8rr, X86::SUB8mr },
201 { X86::XOR16ri, X86::XOR16mi },
202 { X86::XOR16ri8, X86::XOR16mi8 },
203 { X86::XOR16rr, X86::XOR16mr },
204 { X86::XOR32ri, X86::XOR32mi },
205 { X86::XOR32ri8, X86::XOR32mi8 },
206 { X86::XOR32rr, X86::XOR32mr },
207 { X86::XOR64ri32, X86::XOR64mi32 },
208 { X86::XOR64ri8, X86::XOR64mi8 },
209 { X86::XOR64rr, X86::XOR64mr },
210 { X86::XOR8ri, X86::XOR8mi },
211 { X86::XOR8rr, X86::XOR8mr }
214 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
215 unsigned RegOp = OpTbl2Addr[i][0];
216 unsigned MemOp = OpTbl2Addr[i][1];
217 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp,
219 assert(false && "Duplicated entries?");
220 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
221 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
222 std::make_pair(RegOp,
224 AmbEntries.push_back(MemOp);
227 // If the third value is 1, then it's folding either a load or a store.
228 static const unsigned OpTbl0[][3] = {
229 { X86::BT16ri8, X86::BT16mi8, 1 },
230 { X86::BT32ri8, X86::BT32mi8, 1 },
231 { X86::BT64ri8, X86::BT64mi8, 1 },
232 { X86::CALL32r, X86::CALL32m, 1 },
233 { X86::CALL64r, X86::CALL64m, 1 },
234 { X86::CMP16ri, X86::CMP16mi, 1 },
235 { X86::CMP16ri8, X86::CMP16mi8, 1 },
236 { X86::CMP16rr, X86::CMP16mr, 1 },
237 { X86::CMP32ri, X86::CMP32mi, 1 },
238 { X86::CMP32ri8, X86::CMP32mi8, 1 },
239 { X86::CMP32rr, X86::CMP32mr, 1 },
240 { X86::CMP64ri32, X86::CMP64mi32, 1 },
241 { X86::CMP64ri8, X86::CMP64mi8, 1 },
242 { X86::CMP64rr, X86::CMP64mr, 1 },
243 { X86::CMP8ri, X86::CMP8mi, 1 },
244 { X86::CMP8rr, X86::CMP8mr, 1 },
245 { X86::DIV16r, X86::DIV16m, 1 },
246 { X86::DIV32r, X86::DIV32m, 1 },
247 { X86::DIV64r, X86::DIV64m, 1 },
248 { X86::DIV8r, X86::DIV8m, 1 },
249 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0 },
250 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
251 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
252 { X86::IDIV16r, X86::IDIV16m, 1 },
253 { X86::IDIV32r, X86::IDIV32m, 1 },
254 { X86::IDIV64r, X86::IDIV64m, 1 },
255 { X86::IDIV8r, X86::IDIV8m, 1 },
256 { X86::IMUL16r, X86::IMUL16m, 1 },
257 { X86::IMUL32r, X86::IMUL32m, 1 },
258 { X86::IMUL64r, X86::IMUL64m, 1 },
259 { X86::IMUL8r, X86::IMUL8m, 1 },
260 { X86::JMP32r, X86::JMP32m, 1 },
261 { X86::JMP64r, X86::JMP64m, 1 },
262 { X86::MOV16ri, X86::MOV16mi, 0 },
263 { X86::MOV16rr, X86::MOV16mr, 0 },
264 { X86::MOV32ri, X86::MOV32mi, 0 },
265 { X86::MOV32rr, X86::MOV32mr, 0 },
266 { X86::MOV64ri32, X86::MOV64mi32, 0 },
267 { X86::MOV64rr, X86::MOV64mr, 0 },
268 { X86::MOV8ri, X86::MOV8mi, 0 },
269 { X86::MOV8rr, X86::MOV8mr, 0 },
270 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, 0 },
271 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
272 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
273 { X86::MOVDQArr, X86::MOVDQAmr, 0 },
274 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
275 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
276 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
277 { X86::MOVSDrr, X86::MOVSDmr, 0 },
278 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
279 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
280 { X86::MOVSSrr, X86::MOVSSmr, 0 },
281 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
282 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
283 { X86::MUL16r, X86::MUL16m, 1 },
284 { X86::MUL32r, X86::MUL32m, 1 },
285 { X86::MUL64r, X86::MUL64m, 1 },
286 { X86::MUL8r, X86::MUL8m, 1 },
287 { X86::SETAEr, X86::SETAEm, 0 },
288 { X86::SETAr, X86::SETAm, 0 },
289 { X86::SETBEr, X86::SETBEm, 0 },
290 { X86::SETBr, X86::SETBm, 0 },
291 { X86::SETEr, X86::SETEm, 0 },
292 { X86::SETGEr, X86::SETGEm, 0 },
293 { X86::SETGr, X86::SETGm, 0 },
294 { X86::SETLEr, X86::SETLEm, 0 },
295 { X86::SETLr, X86::SETLm, 0 },
296 { X86::SETNEr, X86::SETNEm, 0 },
297 { X86::SETNOr, X86::SETNOm, 0 },
298 { X86::SETNPr, X86::SETNPm, 0 },
299 { X86::SETNSr, X86::SETNSm, 0 },
300 { X86::SETOr, X86::SETOm, 0 },
301 { X86::SETPr, X86::SETPm, 0 },
302 { X86::SETSr, X86::SETSm, 0 },
303 { X86::TAILJMPr, X86::TAILJMPm, 1 },
304 { X86::TEST16ri, X86::TEST16mi, 1 },
305 { X86::TEST32ri, X86::TEST32mi, 1 },
306 { X86::TEST64ri32, X86::TEST64mi32, 1 },
307 { X86::TEST8ri, X86::TEST8mi, 1 }
310 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
311 unsigned RegOp = OpTbl0[i][0];
312 unsigned MemOp = OpTbl0[i][1];
313 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp,
315 assert(false && "Duplicated entries?");
316 unsigned FoldedLoad = OpTbl0[i][2];
317 // Index 0, folded load or store.
318 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
319 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
320 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
321 std::make_pair(RegOp, AuxInfo))).second)
322 AmbEntries.push_back(MemOp);
325 static const unsigned OpTbl1[][2] = {
326 { X86::CMP16rr, X86::CMP16rm },
327 { X86::CMP32rr, X86::CMP32rm },
328 { X86::CMP64rr, X86::CMP64rm },
329 { X86::CMP8rr, X86::CMP8rm },
330 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
331 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
332 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
333 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
334 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
335 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
336 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
337 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
338 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
339 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
340 { X86::FsMOVAPDrr, X86::MOVSDrm },
341 { X86::FsMOVAPSrr, X86::MOVSSrm },
342 { X86::IMUL16rri, X86::IMUL16rmi },
343 { X86::IMUL16rri8, X86::IMUL16rmi8 },
344 { X86::IMUL32rri, X86::IMUL32rmi },
345 { X86::IMUL32rri8, X86::IMUL32rmi8 },
346 { X86::IMUL64rri32, X86::IMUL64rmi32 },
347 { X86::IMUL64rri8, X86::IMUL64rmi8 },
348 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
349 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
350 { X86::Int_COMISDrr, X86::Int_COMISDrm },
351 { X86::Int_COMISSrr, X86::Int_COMISSrm },
352 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
353 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
354 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
355 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
356 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
357 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
358 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
359 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
360 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
361 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
362 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
363 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
364 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
365 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
366 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
367 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
368 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
369 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
370 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
371 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
372 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
373 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
374 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
375 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
376 { X86::MOV16rr, X86::MOV16rm },
377 { X86::MOV32rr, X86::MOV32rm },
378 { X86::MOV64rr, X86::MOV64rm },
379 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
380 { X86::MOV64toSDrr, X86::MOV64toSDrm },
381 { X86::MOV8rr, X86::MOV8rm },
382 { X86::MOVAPDrr, X86::MOVAPDrm },
383 { X86::MOVAPSrr, X86::MOVAPSrm },
384 { X86::MOVDDUPrr, X86::MOVDDUPrm },
385 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
386 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
387 { X86::MOVDQArr, X86::MOVDQArm },
388 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
389 { X86::MOVSDrr, X86::MOVSDrm },
390 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
391 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
392 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
393 { X86::MOVSSrr, X86::MOVSSrm },
394 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
395 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
396 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
397 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
398 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
399 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
400 { X86::MOVUPDrr, X86::MOVUPDrm },
401 { X86::MOVUPSrr, X86::MOVUPSrm },
402 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
403 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
404 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
405 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
406 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
407 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8 },
408 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
409 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
410 { X86::MOVZX64rr32, X86::MOVZX64rm32 },
411 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
412 { X86::PSHUFDri, X86::PSHUFDmi },
413 { X86::PSHUFHWri, X86::PSHUFHWmi },
414 { X86::PSHUFLWri, X86::PSHUFLWmi },
415 { X86::RCPPSr, X86::RCPPSm },
416 { X86::RCPPSr_Int, X86::RCPPSm_Int },
417 { X86::RSQRTPSr, X86::RSQRTPSm },
418 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
419 { X86::RSQRTSSr, X86::RSQRTSSm },
420 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
421 { X86::SQRTPDr, X86::SQRTPDm },
422 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
423 { X86::SQRTPSr, X86::SQRTPSm },
424 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
425 { X86::SQRTSDr, X86::SQRTSDm },
426 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
427 { X86::SQRTSSr, X86::SQRTSSm },
428 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
429 { X86::TEST16rr, X86::TEST16rm },
430 { X86::TEST32rr, X86::TEST32rm },
431 { X86::TEST64rr, X86::TEST64rm },
432 { X86::TEST8rr, X86::TEST8rm },
433 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
434 { X86::UCOMISDrr, X86::UCOMISDrm },
435 { X86::UCOMISSrr, X86::UCOMISSrm }
438 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
439 unsigned RegOp = OpTbl1[i][0];
440 unsigned MemOp = OpTbl1[i][1];
441 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp,
443 assert(false && "Duplicated entries?");
444 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
445 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
446 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
447 std::make_pair(RegOp, AuxInfo))).second)
448 AmbEntries.push_back(MemOp);
451 static const unsigned OpTbl2[][2] = {
452 { X86::ADC32rr, X86::ADC32rm },
453 { X86::ADC64rr, X86::ADC64rm },
454 { X86::ADD16rr, X86::ADD16rm },
455 { X86::ADD32rr, X86::ADD32rm },
456 { X86::ADD64rr, X86::ADD64rm },
457 { X86::ADD8rr, X86::ADD8rm },
458 { X86::ADDPDrr, X86::ADDPDrm },
459 { X86::ADDPSrr, X86::ADDPSrm },
460 { X86::ADDSDrr, X86::ADDSDrm },
461 { X86::ADDSSrr, X86::ADDSSrm },
462 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
463 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
464 { X86::AND16rr, X86::AND16rm },
465 { X86::AND32rr, X86::AND32rm },
466 { X86::AND64rr, X86::AND64rm },
467 { X86::AND8rr, X86::AND8rm },
468 { X86::ANDNPDrr, X86::ANDNPDrm },
469 { X86::ANDNPSrr, X86::ANDNPSrm },
470 { X86::ANDPDrr, X86::ANDPDrm },
471 { X86::ANDPSrr, X86::ANDPSrm },
472 { X86::CMOVA16rr, X86::CMOVA16rm },
473 { X86::CMOVA32rr, X86::CMOVA32rm },
474 { X86::CMOVA64rr, X86::CMOVA64rm },
475 { X86::CMOVAE16rr, X86::CMOVAE16rm },
476 { X86::CMOVAE32rr, X86::CMOVAE32rm },
477 { X86::CMOVAE64rr, X86::CMOVAE64rm },
478 { X86::CMOVB16rr, X86::CMOVB16rm },
479 { X86::CMOVB32rr, X86::CMOVB32rm },
480 { X86::CMOVB64rr, X86::CMOVB64rm },
481 { X86::CMOVBE16rr, X86::CMOVBE16rm },
482 { X86::CMOVBE32rr, X86::CMOVBE32rm },
483 { X86::CMOVBE64rr, X86::CMOVBE64rm },
484 { X86::CMOVE16rr, X86::CMOVE16rm },
485 { X86::CMOVE32rr, X86::CMOVE32rm },
486 { X86::CMOVE64rr, X86::CMOVE64rm },
487 { X86::CMOVG16rr, X86::CMOVG16rm },
488 { X86::CMOVG32rr, X86::CMOVG32rm },
489 { X86::CMOVG64rr, X86::CMOVG64rm },
490 { X86::CMOVGE16rr, X86::CMOVGE16rm },
491 { X86::CMOVGE32rr, X86::CMOVGE32rm },
492 { X86::CMOVGE64rr, X86::CMOVGE64rm },
493 { X86::CMOVL16rr, X86::CMOVL16rm },
494 { X86::CMOVL32rr, X86::CMOVL32rm },
495 { X86::CMOVL64rr, X86::CMOVL64rm },
496 { X86::CMOVLE16rr, X86::CMOVLE16rm },
497 { X86::CMOVLE32rr, X86::CMOVLE32rm },
498 { X86::CMOVLE64rr, X86::CMOVLE64rm },
499 { X86::CMOVNE16rr, X86::CMOVNE16rm },
500 { X86::CMOVNE32rr, X86::CMOVNE32rm },
501 { X86::CMOVNE64rr, X86::CMOVNE64rm },
502 { X86::CMOVNO16rr, X86::CMOVNO16rm },
503 { X86::CMOVNO32rr, X86::CMOVNO32rm },
504 { X86::CMOVNO64rr, X86::CMOVNO64rm },
505 { X86::CMOVNP16rr, X86::CMOVNP16rm },
506 { X86::CMOVNP32rr, X86::CMOVNP32rm },
507 { X86::CMOVNP64rr, X86::CMOVNP64rm },
508 { X86::CMOVNS16rr, X86::CMOVNS16rm },
509 { X86::CMOVNS32rr, X86::CMOVNS32rm },
510 { X86::CMOVNS64rr, X86::CMOVNS64rm },
511 { X86::CMOVO16rr, X86::CMOVO16rm },
512 { X86::CMOVO32rr, X86::CMOVO32rm },
513 { X86::CMOVO64rr, X86::CMOVO64rm },
514 { X86::CMOVP16rr, X86::CMOVP16rm },
515 { X86::CMOVP32rr, X86::CMOVP32rm },
516 { X86::CMOVP64rr, X86::CMOVP64rm },
517 { X86::CMOVS16rr, X86::CMOVS16rm },
518 { X86::CMOVS32rr, X86::CMOVS32rm },
519 { X86::CMOVS64rr, X86::CMOVS64rm },
520 { X86::CMPPDrri, X86::CMPPDrmi },
521 { X86::CMPPSrri, X86::CMPPSrmi },
522 { X86::CMPSDrr, X86::CMPSDrm },
523 { X86::CMPSSrr, X86::CMPSSrm },
524 { X86::DIVPDrr, X86::DIVPDrm },
525 { X86::DIVPSrr, X86::DIVPSrm },
526 { X86::DIVSDrr, X86::DIVSDrm },
527 { X86::DIVSSrr, X86::DIVSSrm },
528 { X86::FsANDNPDrr, X86::FsANDNPDrm },
529 { X86::FsANDNPSrr, X86::FsANDNPSrm },
530 { X86::FsANDPDrr, X86::FsANDPDrm },
531 { X86::FsANDPSrr, X86::FsANDPSrm },
532 { X86::FsORPDrr, X86::FsORPDrm },
533 { X86::FsORPSrr, X86::FsORPSrm },
534 { X86::FsXORPDrr, X86::FsXORPDrm },
535 { X86::FsXORPSrr, X86::FsXORPSrm },
536 { X86::HADDPDrr, X86::HADDPDrm },
537 { X86::HADDPSrr, X86::HADDPSrm },
538 { X86::HSUBPDrr, X86::HSUBPDrm },
539 { X86::HSUBPSrr, X86::HSUBPSrm },
540 { X86::IMUL16rr, X86::IMUL16rm },
541 { X86::IMUL32rr, X86::IMUL32rm },
542 { X86::IMUL64rr, X86::IMUL64rm },
543 { X86::MAXPDrr, X86::MAXPDrm },
544 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
545 { X86::MAXPSrr, X86::MAXPSrm },
546 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
547 { X86::MAXSDrr, X86::MAXSDrm },
548 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
549 { X86::MAXSSrr, X86::MAXSSrm },
550 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
551 { X86::MINPDrr, X86::MINPDrm },
552 { X86::MINPDrr_Int, X86::MINPDrm_Int },
553 { X86::MINPSrr, X86::MINPSrm },
554 { X86::MINPSrr_Int, X86::MINPSrm_Int },
555 { X86::MINSDrr, X86::MINSDrm },
556 { X86::MINSDrr_Int, X86::MINSDrm_Int },
557 { X86::MINSSrr, X86::MINSSrm },
558 { X86::MINSSrr_Int, X86::MINSSrm_Int },
559 { X86::MULPDrr, X86::MULPDrm },
560 { X86::MULPSrr, X86::MULPSrm },
561 { X86::MULSDrr, X86::MULSDrm },
562 { X86::MULSSrr, X86::MULSSrm },
563 { X86::OR16rr, X86::OR16rm },
564 { X86::OR32rr, X86::OR32rm },
565 { X86::OR64rr, X86::OR64rm },
566 { X86::OR8rr, X86::OR8rm },
567 { X86::ORPDrr, X86::ORPDrm },
568 { X86::ORPSrr, X86::ORPSrm },
569 { X86::PACKSSDWrr, X86::PACKSSDWrm },
570 { X86::PACKSSWBrr, X86::PACKSSWBrm },
571 { X86::PACKUSWBrr, X86::PACKUSWBrm },
572 { X86::PADDBrr, X86::PADDBrm },
573 { X86::PADDDrr, X86::PADDDrm },
574 { X86::PADDQrr, X86::PADDQrm },
575 { X86::PADDSBrr, X86::PADDSBrm },
576 { X86::PADDSWrr, X86::PADDSWrm },
577 { X86::PADDWrr, X86::PADDWrm },
578 { X86::PANDNrr, X86::PANDNrm },
579 { X86::PANDrr, X86::PANDrm },
580 { X86::PAVGBrr, X86::PAVGBrm },
581 { X86::PAVGWrr, X86::PAVGWrm },
582 { X86::PCMPEQBrr, X86::PCMPEQBrm },
583 { X86::PCMPEQDrr, X86::PCMPEQDrm },
584 { X86::PCMPEQWrr, X86::PCMPEQWrm },
585 { X86::PCMPGTBrr, X86::PCMPGTBrm },
586 { X86::PCMPGTDrr, X86::PCMPGTDrm },
587 { X86::PCMPGTWrr, X86::PCMPGTWrm },
588 { X86::PINSRWrri, X86::PINSRWrmi },
589 { X86::PMADDWDrr, X86::PMADDWDrm },
590 { X86::PMAXSWrr, X86::PMAXSWrm },
591 { X86::PMAXUBrr, X86::PMAXUBrm },
592 { X86::PMINSWrr, X86::PMINSWrm },
593 { X86::PMINUBrr, X86::PMINUBrm },
594 { X86::PMULDQrr, X86::PMULDQrm },
595 { X86::PMULHUWrr, X86::PMULHUWrm },
596 { X86::PMULHWrr, X86::PMULHWrm },
597 { X86::PMULLDrr, X86::PMULLDrm },
598 { X86::PMULLDrr_int, X86::PMULLDrm_int },
599 { X86::PMULLWrr, X86::PMULLWrm },
600 { X86::PMULUDQrr, X86::PMULUDQrm },
601 { X86::PORrr, X86::PORrm },
602 { X86::PSADBWrr, X86::PSADBWrm },
603 { X86::PSLLDrr, X86::PSLLDrm },
604 { X86::PSLLQrr, X86::PSLLQrm },
605 { X86::PSLLWrr, X86::PSLLWrm },
606 { X86::PSRADrr, X86::PSRADrm },
607 { X86::PSRAWrr, X86::PSRAWrm },
608 { X86::PSRLDrr, X86::PSRLDrm },
609 { X86::PSRLQrr, X86::PSRLQrm },
610 { X86::PSRLWrr, X86::PSRLWrm },
611 { X86::PSUBBrr, X86::PSUBBrm },
612 { X86::PSUBDrr, X86::PSUBDrm },
613 { X86::PSUBSBrr, X86::PSUBSBrm },
614 { X86::PSUBSWrr, X86::PSUBSWrm },
615 { X86::PSUBWrr, X86::PSUBWrm },
616 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
617 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
618 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
619 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
620 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
621 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
622 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
623 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
624 { X86::PXORrr, X86::PXORrm },
625 { X86::SBB32rr, X86::SBB32rm },
626 { X86::SBB64rr, X86::SBB64rm },
627 { X86::SHUFPDrri, X86::SHUFPDrmi },
628 { X86::SHUFPSrri, X86::SHUFPSrmi },
629 { X86::SUB16rr, X86::SUB16rm },
630 { X86::SUB32rr, X86::SUB32rm },
631 { X86::SUB64rr, X86::SUB64rm },
632 { X86::SUB8rr, X86::SUB8rm },
633 { X86::SUBPDrr, X86::SUBPDrm },
634 { X86::SUBPSrr, X86::SUBPSrm },
635 { X86::SUBSDrr, X86::SUBSDrm },
636 { X86::SUBSSrr, X86::SUBSSrm },
637 // FIXME: TEST*rr -> swapped operand of TEST*mr.
638 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
639 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
640 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
641 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
642 { X86::XOR16rr, X86::XOR16rm },
643 { X86::XOR32rr, X86::XOR32rm },
644 { X86::XOR64rr, X86::XOR64rm },
645 { X86::XOR8rr, X86::XOR8rm },
646 { X86::XORPDrr, X86::XORPDrm },
647 { X86::XORPSrr, X86::XORPSrm }
650 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
651 unsigned RegOp = OpTbl2[i][0];
652 unsigned MemOp = OpTbl2[i][1];
653 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp,
655 assert(false && "Duplicated entries?");
656 unsigned AuxInfo = 2 | (1 << 4); // Index 2, folded load
657 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
658 std::make_pair(RegOp, AuxInfo))).second)
659 AmbEntries.push_back(MemOp);
662 // Remove ambiguous entries.
663 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
666 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
667 unsigned &SrcReg, unsigned &DstReg,
668 unsigned &SrcSubIdx, unsigned &DstSubIdx) const {
669 switch (MI.getOpcode()) {
673 case X86::MOV8rr_NOREX:
680 // FP Stack register class copies
681 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
682 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
683 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
685 case X86::FsMOVAPSrr:
686 case X86::FsMOVAPDrr:
690 case X86::MOVSS2PSrr:
691 case X86::MOVSD2PDrr:
692 case X86::MOVPS2SSrr:
693 case X86::MOVPD2SDrr:
694 case X86::MMX_MOVQ64rr:
695 assert(MI.getNumOperands() >= 2 &&
696 MI.getOperand(0).isReg() &&
697 MI.getOperand(1).isReg() &&
698 "invalid register-register move instruction");
699 SrcReg = MI.getOperand(1).getReg();
700 DstReg = MI.getOperand(0).getReg();
701 SrcSubIdx = MI.getOperand(1).getSubReg();
702 DstSubIdx = MI.getOperand(0).getSubReg();
707 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
708 int &FrameIndex) const {
709 switch (MI->getOpcode()) {
721 case X86::MMX_MOVD64rm:
722 case X86::MMX_MOVQ64rm:
723 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
724 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
725 MI->getOperand(2).getImm() == 1 &&
726 MI->getOperand(3).getReg() == 0 &&
727 MI->getOperand(4).getImm() == 0) {
728 FrameIndex = MI->getOperand(1).getIndex();
729 return MI->getOperand(0).getReg();
736 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
737 int &FrameIndex) const {
738 switch (MI->getOpcode()) {
750 case X86::MMX_MOVD64mr:
751 case X86::MMX_MOVQ64mr:
752 case X86::MMX_MOVNTQmr:
753 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
754 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
755 MI->getOperand(1).getImm() == 1 &&
756 MI->getOperand(2).getReg() == 0 &&
757 MI->getOperand(3).getImm() == 0) {
758 FrameIndex = MI->getOperand(0).getIndex();
759 return MI->getOperand(X86AddrNumOperands).getReg();
766 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
768 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
769 bool isPICBase = false;
770 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
771 E = MRI.def_end(); I != E; ++I) {
772 MachineInstr *DefMI = I.getOperand().getParent();
773 if (DefMI->getOpcode() != X86::MOVPC32r)
775 assert(!isPICBase && "More than one PIC base?");
781 /// CanRematLoadWithDispOperand - Return true if a load with the specified
782 /// operand is a candidate for remat: for this to be true we need to know that
783 /// the load will always return the same value, even if moved.
784 static bool CanRematLoadWithDispOperand(const MachineOperand &MO,
785 X86TargetMachine &TM) {
786 // Loads from constant pool entries can be remat'd.
787 if (MO.isCPI()) return true;
789 // We can remat globals in some cases.
791 // If this is a load of a stub, not of the global, we can remat it. This
792 // access will always return the address of the global.
793 if (isGlobalStubReference(MO.getTargetFlags()))
796 // If the global itself is constant, we can remat the load.
797 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(MO.getGlobal()))
798 if (GV->isConstant())
805 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI) const {
806 switch (MI->getOpcode()) {
818 case X86::MMX_MOVD64rm:
819 case X86::MMX_MOVQ64rm: {
820 // Loads from constant pools are trivially rematerializable.
821 if (MI->getOperand(1).isReg() &&
822 MI->getOperand(2).isImm() &&
823 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
824 CanRematLoadWithDispOperand(MI->getOperand(4), TM)) {
825 unsigned BaseReg = MI->getOperand(1).getReg();
826 if (BaseReg == 0 || BaseReg == X86::RIP)
828 // Allow re-materialization of PIC load.
829 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
831 const MachineFunction &MF = *MI->getParent()->getParent();
832 const MachineRegisterInfo &MRI = MF.getRegInfo();
833 bool isPICBase = false;
834 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
835 E = MRI.def_end(); I != E; ++I) {
836 MachineInstr *DefMI = I.getOperand().getParent();
837 if (DefMI->getOpcode() != X86::MOVPC32r)
839 assert(!isPICBase && "More than one PIC base?");
849 if (MI->getOperand(2).isImm() &&
850 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
851 !MI->getOperand(4).isReg()) {
852 // lea fi#, lea GV, etc. are all rematerializable.
853 if (!MI->getOperand(1).isReg())
855 unsigned BaseReg = MI->getOperand(1).getReg();
858 // Allow re-materialization of lea PICBase + x.
859 const MachineFunction &MF = *MI->getParent()->getParent();
860 const MachineRegisterInfo &MRI = MF.getRegInfo();
861 return regIsPICBase(BaseReg, MRI);
867 // All other instructions marked M_REMATERIALIZABLE are always trivially
872 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
873 /// would clobber the EFLAGS condition register. Note the result may be
874 /// conservative. If it cannot definitely determine the safety after visiting
875 /// two instructions it assumes it's not safe.
876 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
877 MachineBasicBlock::iterator I) {
878 // It's always safe to clobber EFLAGS at the end of a block.
882 // For compile time consideration, if we are not able to determine the
883 // safety after visiting 2 instructions, we will assume it's not safe.
884 for (unsigned i = 0; i < 2; ++i) {
885 bool SeenDef = false;
886 for (unsigned j = 0, e = I->getNumOperands(); j != e; ++j) {
887 MachineOperand &MO = I->getOperand(j);
890 if (MO.getReg() == X86::EFLAGS) {
898 // This instruction defines EFLAGS, no need to look any further.
902 // If we make it to the end of the block, it's safe to clobber EFLAGS.
907 // Conservative answer.
911 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
912 MachineBasicBlock::iterator I,
914 const MachineInstr *Orig) const {
915 DebugLoc DL = DebugLoc::getUnknownLoc();
916 if (I != MBB.end()) DL = I->getDebugLoc();
918 unsigned SubIdx = Orig->getOperand(0).isReg()
919 ? Orig->getOperand(0).getSubReg() : 0;
920 bool ChangeSubIdx = SubIdx != 0;
921 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
922 DestReg = RI.getSubReg(DestReg, SubIdx);
926 // MOV32r0 etc. are implemented with xor which clobbers condition code.
927 // Re-materialize them as movri instructions to avoid side effects.
928 bool Emitted = false;
929 switch (Orig->getOpcode()) {
935 if (!isSafeToClobberEFLAGS(MBB, I)) {
937 switch (Orig->getOpcode()) {
939 case X86::MOV8r0: Opc = X86::MOV8ri; break;
940 case X86::MOV16r0: Opc = X86::MOV16ri; break;
941 case X86::MOV32r0: Opc = X86::MOV32ri; break;
942 case X86::MOV64r0: Opc = X86::MOV64ri32; break;
944 BuildMI(MBB, I, DL, get(Opc), DestReg).addImm(0);
952 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
953 MI->getOperand(0).setReg(DestReg);
958 MachineInstr *NewMI = prior(I);
959 NewMI->getOperand(0).setSubReg(SubIdx);
963 /// isInvariantLoad - Return true if the specified instruction (which is marked
964 /// mayLoad) is loading from a location whose value is invariant across the
965 /// function. For example, loading a value from the constant pool or from
966 /// from the argument area of a function if it does not change. This should
967 /// only return true of *all* loads the instruction does are invariant (if it
968 /// does multiple loads).
969 bool X86InstrInfo::isInvariantLoad(const MachineInstr *MI) const {
970 // This code cares about loads from three cases: constant pool entries,
971 // invariant argument slots, and global stubs. In order to handle these cases
972 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
973 // operand and base our analysis on it. This is safe because the address of
974 // none of these three cases is ever used as anything other than a load base
975 // and X86 doesn't have any instructions that load from multiple places.
977 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
978 const MachineOperand &MO = MI->getOperand(i);
979 // Loads from constant pools are trivially invariant.
984 return isGlobalStubReference(MO.getTargetFlags());
986 // If this is a load from an invariant stack slot, the load is a constant.
988 const MachineFrameInfo &MFI =
989 *MI->getParent()->getParent()->getFrameInfo();
990 int Idx = MO.getIndex();
991 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
995 // All other instances of these instructions are presumed to have other
1000 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1001 /// is not marked dead.
1002 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1003 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1004 MachineOperand &MO = MI->getOperand(i);
1005 if (MO.isReg() && MO.isDef() &&
1006 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1013 /// convertToThreeAddress - This method must be implemented by targets that
1014 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1015 /// may be able to convert a two-address instruction into a true
1016 /// three-address instruction on demand. This allows the X86 target (for
1017 /// example) to convert ADD and SHL instructions into LEA instructions if they
1018 /// would require register copies due to two-addressness.
1020 /// This method returns a null pointer if the transformation cannot be
1021 /// performed, otherwise it returns the new instruction.
1024 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1025 MachineBasicBlock::iterator &MBBI,
1026 LiveVariables *LV) const {
1027 MachineInstr *MI = MBBI;
1028 MachineFunction &MF = *MI->getParent()->getParent();
1029 // All instructions input are two-addr instructions. Get the known operands.
1030 unsigned Dest = MI->getOperand(0).getReg();
1031 unsigned Src = MI->getOperand(1).getReg();
1032 bool isDead = MI->getOperand(0).isDead();
1033 bool isKill = MI->getOperand(1).isKill();
1035 MachineInstr *NewMI = NULL;
1036 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1037 // we have better subtarget support, enable the 16-bit LEA generation here.
1038 bool DisableLEA16 = true;
1040 unsigned MIOpc = MI->getOpcode();
1042 case X86::SHUFPSrri: {
1043 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1044 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1046 unsigned B = MI->getOperand(1).getReg();
1047 unsigned C = MI->getOperand(2).getReg();
1048 if (B != C) return 0;
1049 unsigned A = MI->getOperand(0).getReg();
1050 unsigned M = MI->getOperand(3).getImm();
1051 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1052 .addReg(A, RegState::Define | getDeadRegState(isDead))
1053 .addReg(B, getKillRegState(isKill)).addImm(M);
1056 case X86::SHL64ri: {
1057 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1058 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1059 // the flags produced by a shift yet, so this is safe.
1060 unsigned ShAmt = MI->getOperand(2).getImm();
1061 if (ShAmt == 0 || ShAmt >= 4) return 0;
1063 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1064 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1065 .addReg(0).addImm(1 << ShAmt)
1066 .addReg(Src, getKillRegState(isKill))
1070 case X86::SHL32ri: {
1071 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1072 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1073 // the flags produced by a shift yet, so this is safe.
1074 unsigned ShAmt = MI->getOperand(2).getImm();
1075 if (ShAmt == 0 || ShAmt >= 4) return 0;
1077 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1078 X86::LEA64_32r : X86::LEA32r;
1079 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1080 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1081 .addReg(0).addImm(1 << ShAmt)
1082 .addReg(Src, getKillRegState(isKill)).addImm(0);
1085 case X86::SHL16ri: {
1086 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1087 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1088 // the flags produced by a shift yet, so this is safe.
1089 unsigned ShAmt = MI->getOperand(2).getImm();
1090 if (ShAmt == 0 || ShAmt >= 4) return 0;
1093 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1094 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1095 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1096 ? X86::LEA64_32r : X86::LEA32r;
1097 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1098 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1100 // Build and insert into an implicit UNDEF value. This is OK because
1101 // well be shifting and then extracting the lower 16-bits.
1102 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1103 MachineInstr *InsMI =
1104 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::INSERT_SUBREG),leaInReg)
1106 .addReg(Src, getKillRegState(isKill))
1107 .addImm(X86::SUBREG_16BIT);
1109 NewMI = BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(Opc), leaOutReg)
1110 .addReg(0).addImm(1 << ShAmt)
1111 .addReg(leaInReg, RegState::Kill)
1114 MachineInstr *ExtMI =
1115 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::EXTRACT_SUBREG))
1116 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1117 .addReg(leaOutReg, RegState::Kill)
1118 .addImm(X86::SUBREG_16BIT);
1121 // Update live variables
1122 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1123 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1125 LV->replaceKillInstruction(Src, MI, InsMI);
1127 LV->replaceKillInstruction(Dest, MI, ExtMI);
1131 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1132 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1133 .addReg(0).addImm(1 << ShAmt)
1134 .addReg(Src, getKillRegState(isKill))
1140 // The following opcodes also sets the condition code register(s). Only
1141 // convert them to equivalent lea if the condition code register def's
1143 if (hasLiveCondCodeDef(MI))
1146 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1151 case X86::INC64_32r: {
1152 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1153 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1154 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1155 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1156 .addReg(Dest, RegState::Define |
1157 getDeadRegState(isDead)),
1162 case X86::INC64_16r:
1163 if (DisableLEA16) return 0;
1164 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1165 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1166 .addReg(Dest, RegState::Define |
1167 getDeadRegState(isDead)),
1172 case X86::DEC64_32r: {
1173 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1174 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1175 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1176 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1177 .addReg(Dest, RegState::Define |
1178 getDeadRegState(isDead)),
1183 case X86::DEC64_16r:
1184 if (DisableLEA16) return 0;
1185 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1186 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1187 .addReg(Dest, RegState::Define |
1188 getDeadRegState(isDead)),
1192 case X86::ADD32rr: {
1193 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1194 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1195 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1196 unsigned Src2 = MI->getOperand(2).getReg();
1197 bool isKill2 = MI->getOperand(2).isKill();
1198 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1199 .addReg(Dest, RegState::Define |
1200 getDeadRegState(isDead)),
1201 Src, isKill, Src2, isKill2);
1203 LV->replaceKillInstruction(Src2, MI, NewMI);
1206 case X86::ADD16rr: {
1207 if (DisableLEA16) return 0;
1208 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1209 unsigned Src2 = MI->getOperand(2).getReg();
1210 bool isKill2 = MI->getOperand(2).isKill();
1211 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1212 .addReg(Dest, RegState::Define |
1213 getDeadRegState(isDead)),
1214 Src, isKill, Src2, isKill2);
1216 LV->replaceKillInstruction(Src2, MI, NewMI);
1219 case X86::ADD64ri32:
1221 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1222 if (MI->getOperand(2).isImm())
1223 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1224 .addReg(Dest, RegState::Define |
1225 getDeadRegState(isDead)),
1226 Src, isKill, MI->getOperand(2).getImm());
1230 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1231 if (MI->getOperand(2).isImm()) {
1232 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1233 NewMI = addLeaRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1234 .addReg(Dest, RegState::Define |
1235 getDeadRegState(isDead)),
1236 Src, isKill, MI->getOperand(2).getImm());
1241 if (DisableLEA16) return 0;
1242 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1243 if (MI->getOperand(2).isImm())
1244 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1245 .addReg(Dest, RegState::Define |
1246 getDeadRegState(isDead)),
1247 Src, isKill, MI->getOperand(2).getImm());
1250 if (DisableLEA16) return 0;
1252 case X86::SHL64ri: {
1253 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImm() &&
1254 "Unknown shl instruction!");
1255 unsigned ShAmt = MI->getOperand(2).getImm();
1256 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1258 AM.Scale = 1 << ShAmt;
1260 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1261 : (MIOpc == X86::SHL32ri
1262 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1263 NewMI = addFullAddress(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1264 .addReg(Dest, RegState::Define |
1265 getDeadRegState(isDead)), AM);
1267 NewMI->getOperand(3).setIsKill(true);
1275 if (!NewMI) return 0;
1277 if (LV) { // Update live variables
1279 LV->replaceKillInstruction(Src, MI, NewMI);
1281 LV->replaceKillInstruction(Dest, MI, NewMI);
1284 MFI->insert(MBBI, NewMI); // Insert the new inst
1288 /// commuteInstruction - We have a few instructions that must be hacked on to
1292 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1293 switch (MI->getOpcode()) {
1294 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1295 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1296 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1297 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1298 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1299 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1302 switch (MI->getOpcode()) {
1303 default: LLVM_UNREACHABLE("Unreachable!");
1304 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1305 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1306 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1307 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1308 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1309 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1311 unsigned Amt = MI->getOperand(3).getImm();
1313 MachineFunction &MF = *MI->getParent()->getParent();
1314 MI = MF.CloneMachineInstr(MI);
1317 MI->setDesc(get(Opc));
1318 MI->getOperand(3).setImm(Size-Amt);
1319 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1321 case X86::CMOVB16rr:
1322 case X86::CMOVB32rr:
1323 case X86::CMOVB64rr:
1324 case X86::CMOVAE16rr:
1325 case X86::CMOVAE32rr:
1326 case X86::CMOVAE64rr:
1327 case X86::CMOVE16rr:
1328 case X86::CMOVE32rr:
1329 case X86::CMOVE64rr:
1330 case X86::CMOVNE16rr:
1331 case X86::CMOVNE32rr:
1332 case X86::CMOVNE64rr:
1333 case X86::CMOVBE16rr:
1334 case X86::CMOVBE32rr:
1335 case X86::CMOVBE64rr:
1336 case X86::CMOVA16rr:
1337 case X86::CMOVA32rr:
1338 case X86::CMOVA64rr:
1339 case X86::CMOVL16rr:
1340 case X86::CMOVL32rr:
1341 case X86::CMOVL64rr:
1342 case X86::CMOVGE16rr:
1343 case X86::CMOVGE32rr:
1344 case X86::CMOVGE64rr:
1345 case X86::CMOVLE16rr:
1346 case X86::CMOVLE32rr:
1347 case X86::CMOVLE64rr:
1348 case X86::CMOVG16rr:
1349 case X86::CMOVG32rr:
1350 case X86::CMOVG64rr:
1351 case X86::CMOVS16rr:
1352 case X86::CMOVS32rr:
1353 case X86::CMOVS64rr:
1354 case X86::CMOVNS16rr:
1355 case X86::CMOVNS32rr:
1356 case X86::CMOVNS64rr:
1357 case X86::CMOVP16rr:
1358 case X86::CMOVP32rr:
1359 case X86::CMOVP64rr:
1360 case X86::CMOVNP16rr:
1361 case X86::CMOVNP32rr:
1362 case X86::CMOVNP64rr:
1363 case X86::CMOVO16rr:
1364 case X86::CMOVO32rr:
1365 case X86::CMOVO64rr:
1366 case X86::CMOVNO16rr:
1367 case X86::CMOVNO32rr:
1368 case X86::CMOVNO64rr: {
1370 switch (MI->getOpcode()) {
1372 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1373 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1374 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1375 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1376 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1377 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1378 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1379 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1380 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1381 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1382 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1383 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1384 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1385 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1386 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1387 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1388 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1389 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1390 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1391 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1392 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1393 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1394 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1395 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1396 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1397 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1398 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1399 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1400 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1401 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1402 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1403 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1404 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
1405 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1406 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1407 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1408 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1409 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1410 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
1411 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1412 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1413 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1414 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1415 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1416 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
1417 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1418 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1419 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1422 MachineFunction &MF = *MI->getParent()->getParent();
1423 MI = MF.CloneMachineInstr(MI);
1426 MI->setDesc(get(Opc));
1427 // Fallthrough intended.
1430 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1434 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1436 default: return X86::COND_INVALID;
1437 case X86::JE: return X86::COND_E;
1438 case X86::JNE: return X86::COND_NE;
1439 case X86::JL: return X86::COND_L;
1440 case X86::JLE: return X86::COND_LE;
1441 case X86::JG: return X86::COND_G;
1442 case X86::JGE: return X86::COND_GE;
1443 case X86::JB: return X86::COND_B;
1444 case X86::JBE: return X86::COND_BE;
1445 case X86::JA: return X86::COND_A;
1446 case X86::JAE: return X86::COND_AE;
1447 case X86::JS: return X86::COND_S;
1448 case X86::JNS: return X86::COND_NS;
1449 case X86::JP: return X86::COND_P;
1450 case X86::JNP: return X86::COND_NP;
1451 case X86::JO: return X86::COND_O;
1452 case X86::JNO: return X86::COND_NO;
1456 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1458 default: LLVM_UNREACHABLE("Illegal condition code!");
1459 case X86::COND_E: return X86::JE;
1460 case X86::COND_NE: return X86::JNE;
1461 case X86::COND_L: return X86::JL;
1462 case X86::COND_LE: return X86::JLE;
1463 case X86::COND_G: return X86::JG;
1464 case X86::COND_GE: return X86::JGE;
1465 case X86::COND_B: return X86::JB;
1466 case X86::COND_BE: return X86::JBE;
1467 case X86::COND_A: return X86::JA;
1468 case X86::COND_AE: return X86::JAE;
1469 case X86::COND_S: return X86::JS;
1470 case X86::COND_NS: return X86::JNS;
1471 case X86::COND_P: return X86::JP;
1472 case X86::COND_NP: return X86::JNP;
1473 case X86::COND_O: return X86::JO;
1474 case X86::COND_NO: return X86::JNO;
1478 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1479 /// e.g. turning COND_E to COND_NE.
1480 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1482 default: LLVM_UNREACHABLE("Illegal condition code!");
1483 case X86::COND_E: return X86::COND_NE;
1484 case X86::COND_NE: return X86::COND_E;
1485 case X86::COND_L: return X86::COND_GE;
1486 case X86::COND_LE: return X86::COND_G;
1487 case X86::COND_G: return X86::COND_LE;
1488 case X86::COND_GE: return X86::COND_L;
1489 case X86::COND_B: return X86::COND_AE;
1490 case X86::COND_BE: return X86::COND_A;
1491 case X86::COND_A: return X86::COND_BE;
1492 case X86::COND_AE: return X86::COND_B;
1493 case X86::COND_S: return X86::COND_NS;
1494 case X86::COND_NS: return X86::COND_S;
1495 case X86::COND_P: return X86::COND_NP;
1496 case X86::COND_NP: return X86::COND_P;
1497 case X86::COND_O: return X86::COND_NO;
1498 case X86::COND_NO: return X86::COND_O;
1502 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1503 const TargetInstrDesc &TID = MI->getDesc();
1504 if (!TID.isTerminator()) return false;
1506 // Conditional branch is a special case.
1507 if (TID.isBranch() && !TID.isBarrier())
1509 if (!TID.isPredicable())
1511 return !isPredicated(MI);
1514 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1515 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1516 const X86InstrInfo &TII) {
1517 if (MI->getOpcode() == X86::FP_REG_KILL)
1519 return TII.isUnpredicatedTerminator(MI);
1522 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1523 MachineBasicBlock *&TBB,
1524 MachineBasicBlock *&FBB,
1525 SmallVectorImpl<MachineOperand> &Cond,
1526 bool AllowModify) const {
1527 // Start from the bottom of the block and work up, examining the
1528 // terminator instructions.
1529 MachineBasicBlock::iterator I = MBB.end();
1530 while (I != MBB.begin()) {
1532 // Working from the bottom, when we see a non-terminator
1533 // instruction, we're done.
1534 if (!isBrAnalysisUnpredicatedTerminator(I, *this))
1536 // A terminator that isn't a branch can't easily be handled
1537 // by this analysis.
1538 if (!I->getDesc().isBranch())
1540 // Handle unconditional branches.
1541 if (I->getOpcode() == X86::JMP) {
1543 TBB = I->getOperand(0).getMBB();
1547 // If the block has any instructions after a JMP, delete them.
1548 while (next(I) != MBB.end())
1549 next(I)->eraseFromParent();
1552 // Delete the JMP if it's equivalent to a fall-through.
1553 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1555 I->eraseFromParent();
1559 // TBB is used to indicate the unconditinal destination.
1560 TBB = I->getOperand(0).getMBB();
1563 // Handle conditional branches.
1564 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1565 if (BranchCode == X86::COND_INVALID)
1566 return true; // Can't handle indirect branch.
1567 // Working from the bottom, handle the first conditional branch.
1570 TBB = I->getOperand(0).getMBB();
1571 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1574 // Handle subsequent conditional branches. Only handle the case
1575 // where all conditional branches branch to the same destination
1576 // and their condition opcodes fit one of the special
1577 // multi-branch idioms.
1578 assert(Cond.size() == 1);
1580 // Only handle the case where all conditional branches branch to
1581 // the same destination.
1582 if (TBB != I->getOperand(0).getMBB())
1584 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1585 // If the conditions are the same, we can leave them alone.
1586 if (OldBranchCode == BranchCode)
1588 // If they differ, see if they fit one of the known patterns.
1589 // Theoretically we could handle more patterns here, but
1590 // we shouldn't expect to see them if instruction selection
1591 // has done a reasonable job.
1592 if ((OldBranchCode == X86::COND_NP &&
1593 BranchCode == X86::COND_E) ||
1594 (OldBranchCode == X86::COND_E &&
1595 BranchCode == X86::COND_NP))
1596 BranchCode = X86::COND_NP_OR_E;
1597 else if ((OldBranchCode == X86::COND_P &&
1598 BranchCode == X86::COND_NE) ||
1599 (OldBranchCode == X86::COND_NE &&
1600 BranchCode == X86::COND_P))
1601 BranchCode = X86::COND_NE_OR_P;
1604 // Update the MachineOperand.
1605 Cond[0].setImm(BranchCode);
1611 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1612 MachineBasicBlock::iterator I = MBB.end();
1615 while (I != MBB.begin()) {
1617 if (I->getOpcode() != X86::JMP &&
1618 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1620 // Remove the branch.
1621 I->eraseFromParent();
1630 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1631 MachineBasicBlock *FBB,
1632 const SmallVectorImpl<MachineOperand> &Cond) const {
1633 // FIXME this should probably have a DebugLoc operand
1634 DebugLoc dl = DebugLoc::getUnknownLoc();
1635 // Shouldn't be a fall through.
1636 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1637 assert((Cond.size() == 1 || Cond.size() == 0) &&
1638 "X86 branch conditions have one component!");
1641 // Unconditional branch?
1642 assert(!FBB && "Unconditional branch with multiple successors!");
1643 BuildMI(&MBB, dl, get(X86::JMP)).addMBB(TBB);
1647 // Conditional branch.
1649 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1651 case X86::COND_NP_OR_E:
1652 // Synthesize NP_OR_E with two branches.
1653 BuildMI(&MBB, dl, get(X86::JNP)).addMBB(TBB);
1655 BuildMI(&MBB, dl, get(X86::JE)).addMBB(TBB);
1658 case X86::COND_NE_OR_P:
1659 // Synthesize NE_OR_P with two branches.
1660 BuildMI(&MBB, dl, get(X86::JNE)).addMBB(TBB);
1662 BuildMI(&MBB, dl, get(X86::JP)).addMBB(TBB);
1666 unsigned Opc = GetCondBranchFromCond(CC);
1667 BuildMI(&MBB, dl, get(Opc)).addMBB(TBB);
1672 // Two-way Conditional branch. Insert the second branch.
1673 BuildMI(&MBB, dl, get(X86::JMP)).addMBB(FBB);
1679 /// isHReg - Test if the given register is a physical h register.
1680 static bool isHReg(unsigned Reg) {
1681 return X86::GR8_ABCD_HRegClass.contains(Reg);
1684 bool X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1685 MachineBasicBlock::iterator MI,
1686 unsigned DestReg, unsigned SrcReg,
1687 const TargetRegisterClass *DestRC,
1688 const TargetRegisterClass *SrcRC) const {
1689 DebugLoc DL = DebugLoc::getUnknownLoc();
1690 if (MI != MBB.end()) DL = MI->getDebugLoc();
1692 // Determine if DstRC and SrcRC have a common superclass in common.
1693 const TargetRegisterClass *CommonRC = DestRC;
1694 if (DestRC == SrcRC)
1695 /* Source and destination have the same register class. */;
1696 else if (CommonRC->hasSuperClass(SrcRC))
1698 else if (!DestRC->hasSubClass(SrcRC))
1703 if (CommonRC == &X86::GR64RegClass) {
1705 } else if (CommonRC == &X86::GR32RegClass) {
1707 } else if (CommonRC == &X86::GR16RegClass) {
1709 } else if (CommonRC == &X86::GR8RegClass) {
1710 // Copying to or from a physical H register on x86-64 requires a NOREX
1711 // move. Otherwise use a normal move.
1712 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
1713 TM.getSubtarget<X86Subtarget>().is64Bit())
1714 Opc = X86::MOV8rr_NOREX;
1717 } else if (CommonRC == &X86::GR64_ABCDRegClass) {
1719 } else if (CommonRC == &X86::GR32_ABCDRegClass) {
1721 } else if (CommonRC == &X86::GR16_ABCDRegClass) {
1723 } else if (CommonRC == &X86::GR8_ABCD_LRegClass) {
1725 } else if (CommonRC == &X86::GR8_ABCD_HRegClass) {
1726 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1727 Opc = X86::MOV8rr_NOREX;
1730 } else if (CommonRC == &X86::GR64_NOREXRegClass) {
1732 } else if (CommonRC == &X86::GR32_NOREXRegClass) {
1734 } else if (CommonRC == &X86::GR16_NOREXRegClass) {
1736 } else if (CommonRC == &X86::GR8_NOREXRegClass) {
1738 } else if (CommonRC == &X86::RFP32RegClass) {
1739 Opc = X86::MOV_Fp3232;
1740 } else if (CommonRC == &X86::RFP64RegClass || CommonRC == &X86::RSTRegClass) {
1741 Opc = X86::MOV_Fp6464;
1742 } else if (CommonRC == &X86::RFP80RegClass) {
1743 Opc = X86::MOV_Fp8080;
1744 } else if (CommonRC == &X86::FR32RegClass) {
1745 Opc = X86::FsMOVAPSrr;
1746 } else if (CommonRC == &X86::FR64RegClass) {
1747 Opc = X86::FsMOVAPDrr;
1748 } else if (CommonRC == &X86::VR128RegClass) {
1749 Opc = X86::MOVAPSrr;
1750 } else if (CommonRC == &X86::VR64RegClass) {
1751 Opc = X86::MMX_MOVQ64rr;
1755 BuildMI(MBB, MI, DL, get(Opc), DestReg).addReg(SrcReg);
1759 // Moving EFLAGS to / from another register requires a push and a pop.
1760 if (SrcRC == &X86::CCRRegClass) {
1761 if (SrcReg != X86::EFLAGS)
1763 if (DestRC == &X86::GR64RegClass) {
1764 BuildMI(MBB, MI, DL, get(X86::PUSHFQ));
1765 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
1767 } else if (DestRC == &X86::GR32RegClass) {
1768 BuildMI(MBB, MI, DL, get(X86::PUSHFD));
1769 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
1772 } else if (DestRC == &X86::CCRRegClass) {
1773 if (DestReg != X86::EFLAGS)
1775 if (SrcRC == &X86::GR64RegClass) {
1776 BuildMI(MBB, MI, DL, get(X86::PUSH64r)).addReg(SrcReg);
1777 BuildMI(MBB, MI, DL, get(X86::POPFQ));
1779 } else if (SrcRC == &X86::GR32RegClass) {
1780 BuildMI(MBB, MI, DL, get(X86::PUSH32r)).addReg(SrcReg);
1781 BuildMI(MBB, MI, DL, get(X86::POPFD));
1786 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1787 if (SrcRC == &X86::RSTRegClass) {
1788 // Copying from ST(0)/ST(1).
1789 if (SrcReg != X86::ST0 && SrcReg != X86::ST1)
1790 // Can only copy from ST(0)/ST(1) right now
1792 bool isST0 = SrcReg == X86::ST0;
1794 if (DestRC == &X86::RFP32RegClass)
1795 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1796 else if (DestRC == &X86::RFP64RegClass)
1797 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1799 if (DestRC != &X86::RFP80RegClass)
1801 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1803 BuildMI(MBB, MI, DL, get(Opc), DestReg);
1807 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1808 if (DestRC == &X86::RSTRegClass) {
1809 // Copying to ST(0) / ST(1).
1810 if (DestReg != X86::ST0 && DestReg != X86::ST1)
1811 // Can only copy to TOS right now
1813 bool isST0 = DestReg == X86::ST0;
1815 if (SrcRC == &X86::RFP32RegClass)
1816 Opc = isST0 ? X86::FpSET_ST0_32 : X86::FpSET_ST1_32;
1817 else if (SrcRC == &X86::RFP64RegClass)
1818 Opc = isST0 ? X86::FpSET_ST0_64 : X86::FpSET_ST1_64;
1820 if (SrcRC != &X86::RFP80RegClass)
1822 Opc = isST0 ? X86::FpSET_ST0_80 : X86::FpSET_ST1_80;
1824 BuildMI(MBB, MI, DL, get(Opc)).addReg(SrcReg);
1828 // Not yet supported!
1832 static unsigned getStoreRegOpcode(unsigned SrcReg,
1833 const TargetRegisterClass *RC,
1834 bool isStackAligned,
1835 TargetMachine &TM) {
1837 if (RC == &X86::GR64RegClass) {
1839 } else if (RC == &X86::GR32RegClass) {
1841 } else if (RC == &X86::GR16RegClass) {
1843 } else if (RC == &X86::GR8RegClass) {
1844 // Copying to or from a physical H register on x86-64 requires a NOREX
1845 // move. Otherwise use a normal move.
1846 if (isHReg(SrcReg) &&
1847 TM.getSubtarget<X86Subtarget>().is64Bit())
1848 Opc = X86::MOV8mr_NOREX;
1851 } else if (RC == &X86::GR64_ABCDRegClass) {
1853 } else if (RC == &X86::GR32_ABCDRegClass) {
1855 } else if (RC == &X86::GR16_ABCDRegClass) {
1857 } else if (RC == &X86::GR8_ABCD_LRegClass) {
1859 } else if (RC == &X86::GR8_ABCD_HRegClass) {
1860 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1861 Opc = X86::MOV8mr_NOREX;
1864 } else if (RC == &X86::GR64_NOREXRegClass) {
1866 } else if (RC == &X86::GR32_NOREXRegClass) {
1868 } else if (RC == &X86::GR16_NOREXRegClass) {
1870 } else if (RC == &X86::GR8_NOREXRegClass) {
1872 } else if (RC == &X86::RFP80RegClass) {
1873 Opc = X86::ST_FpP80m; // pops
1874 } else if (RC == &X86::RFP64RegClass) {
1875 Opc = X86::ST_Fp64m;
1876 } else if (RC == &X86::RFP32RegClass) {
1877 Opc = X86::ST_Fp32m;
1878 } else if (RC == &X86::FR32RegClass) {
1880 } else if (RC == &X86::FR64RegClass) {
1882 } else if (RC == &X86::VR128RegClass) {
1883 // If stack is realigned we can use aligned stores.
1884 Opc = isStackAligned ? X86::MOVAPSmr : X86::MOVUPSmr;
1885 } else if (RC == &X86::VR64RegClass) {
1886 Opc = X86::MMX_MOVQ64mr;
1888 LLVM_UNREACHABLE("Unknown regclass");
1894 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1895 MachineBasicBlock::iterator MI,
1896 unsigned SrcReg, bool isKill, int FrameIdx,
1897 const TargetRegisterClass *RC) const {
1898 const MachineFunction &MF = *MBB.getParent();
1899 bool isAligned = (RI.getStackAlignment() >= 16) ||
1900 RI.needsStackRealignment(MF);
1901 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
1902 DebugLoc DL = DebugLoc::getUnknownLoc();
1903 if (MI != MBB.end()) DL = MI->getDebugLoc();
1904 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
1905 .addReg(SrcReg, getKillRegState(isKill));
1908 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1910 SmallVectorImpl<MachineOperand> &Addr,
1911 const TargetRegisterClass *RC,
1912 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1913 bool isAligned = (RI.getStackAlignment() >= 16) ||
1914 RI.needsStackRealignment(MF);
1915 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
1916 DebugLoc DL = DebugLoc::getUnknownLoc();
1917 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
1918 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1919 MIB.addOperand(Addr[i]);
1920 MIB.addReg(SrcReg, getKillRegState(isKill));
1921 NewMIs.push_back(MIB);
1924 static unsigned getLoadRegOpcode(unsigned DestReg,
1925 const TargetRegisterClass *RC,
1926 bool isStackAligned,
1927 const TargetMachine &TM) {
1929 if (RC == &X86::GR64RegClass) {
1931 } else if (RC == &X86::GR32RegClass) {
1933 } else if (RC == &X86::GR16RegClass) {
1935 } else if (RC == &X86::GR8RegClass) {
1936 // Copying to or from a physical H register on x86-64 requires a NOREX
1937 // move. Otherwise use a normal move.
1938 if (isHReg(DestReg) &&
1939 TM.getSubtarget<X86Subtarget>().is64Bit())
1940 Opc = X86::MOV8rm_NOREX;
1943 } else if (RC == &X86::GR64_ABCDRegClass) {
1945 } else if (RC == &X86::GR32_ABCDRegClass) {
1947 } else if (RC == &X86::GR16_ABCDRegClass) {
1949 } else if (RC == &X86::GR8_ABCD_LRegClass) {
1951 } else if (RC == &X86::GR8_ABCD_HRegClass) {
1952 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1953 Opc = X86::MOV8rm_NOREX;
1956 } else if (RC == &X86::GR64_NOREXRegClass) {
1958 } else if (RC == &X86::GR32_NOREXRegClass) {
1960 } else if (RC == &X86::GR16_NOREXRegClass) {
1962 } else if (RC == &X86::GR8_NOREXRegClass) {
1964 } else if (RC == &X86::RFP80RegClass) {
1965 Opc = X86::LD_Fp80m;
1966 } else if (RC == &X86::RFP64RegClass) {
1967 Opc = X86::LD_Fp64m;
1968 } else if (RC == &X86::RFP32RegClass) {
1969 Opc = X86::LD_Fp32m;
1970 } else if (RC == &X86::FR32RegClass) {
1972 } else if (RC == &X86::FR64RegClass) {
1974 } else if (RC == &X86::VR128RegClass) {
1975 // If stack is realigned we can use aligned loads.
1976 Opc = isStackAligned ? X86::MOVAPSrm : X86::MOVUPSrm;
1977 } else if (RC == &X86::VR64RegClass) {
1978 Opc = X86::MMX_MOVQ64rm;
1980 LLVM_UNREACHABLE("Unknown regclass");
1986 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1987 MachineBasicBlock::iterator MI,
1988 unsigned DestReg, int FrameIdx,
1989 const TargetRegisterClass *RC) const{
1990 const MachineFunction &MF = *MBB.getParent();
1991 bool isAligned = (RI.getStackAlignment() >= 16) ||
1992 RI.needsStackRealignment(MF);
1993 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
1994 DebugLoc DL = DebugLoc::getUnknownLoc();
1995 if (MI != MBB.end()) DL = MI->getDebugLoc();
1996 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
1999 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
2000 SmallVectorImpl<MachineOperand> &Addr,
2001 const TargetRegisterClass *RC,
2002 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2003 bool isAligned = (RI.getStackAlignment() >= 16) ||
2004 RI.needsStackRealignment(MF);
2005 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2006 DebugLoc DL = DebugLoc::getUnknownLoc();
2007 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
2008 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2009 MIB.addOperand(Addr[i]);
2010 NewMIs.push_back(MIB);
2013 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
2014 MachineBasicBlock::iterator MI,
2015 const std::vector<CalleeSavedInfo> &CSI) const {
2019 DebugLoc DL = DebugLoc::getUnknownLoc();
2020 if (MI != MBB.end()) DL = MI->getDebugLoc();
2022 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2023 unsigned SlotSize = is64Bit ? 8 : 4;
2025 MachineFunction &MF = *MBB.getParent();
2026 unsigned FPReg = RI.getFrameRegister(MF);
2027 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
2028 unsigned CalleeFrameSize = 0;
2030 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
2031 for (unsigned i = CSI.size(); i != 0; --i) {
2032 unsigned Reg = CSI[i-1].getReg();
2033 const TargetRegisterClass *RegClass = CSI[i-1].getRegClass();
2034 // Add the callee-saved register as live-in. It's killed at the spill.
2037 // X86RegisterInfo::emitPrologue will handle spilling of frame register.
2039 if (RegClass != &X86::VR128RegClass) {
2040 CalleeFrameSize += SlotSize;
2041 BuildMI(MBB, MI, DL, get(Opc)).addReg(Reg, RegState::Kill);
2043 storeRegToStackSlot(MBB, MI, Reg, true, CSI[i-1].getFrameIdx(), RegClass);
2047 X86FI->setCalleeSavedFrameSize(CalleeFrameSize);
2051 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
2052 MachineBasicBlock::iterator MI,
2053 const std::vector<CalleeSavedInfo> &CSI) const {
2057 DebugLoc DL = DebugLoc::getUnknownLoc();
2058 if (MI != MBB.end()) DL = MI->getDebugLoc();
2060 MachineFunction &MF = *MBB.getParent();
2061 unsigned FPReg = RI.getFrameRegister(MF);
2062 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2063 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
2064 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
2065 unsigned Reg = CSI[i].getReg();
2067 // X86RegisterInfo::emitEpilogue will handle restoring of frame register.
2069 const TargetRegisterClass *RegClass = CSI[i].getRegClass();
2070 if (RegClass != &X86::VR128RegClass) {
2071 BuildMI(MBB, MI, DL, get(Opc), Reg);
2073 loadRegFromStackSlot(MBB, MI, Reg, CSI[i].getFrameIdx(), RegClass);
2079 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
2080 const SmallVectorImpl<MachineOperand> &MOs,
2082 const TargetInstrInfo &TII) {
2083 // Create the base instruction with the memory operand as the first part.
2084 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2085 MI->getDebugLoc(), true);
2086 MachineInstrBuilder MIB(NewMI);
2087 unsigned NumAddrOps = MOs.size();
2088 for (unsigned i = 0; i != NumAddrOps; ++i)
2089 MIB.addOperand(MOs[i]);
2090 if (NumAddrOps < 4) // FrameIndex only
2093 // Loop over the rest of the ri operands, converting them over.
2094 unsigned NumOps = MI->getDesc().getNumOperands()-2;
2095 for (unsigned i = 0; i != NumOps; ++i) {
2096 MachineOperand &MO = MI->getOperand(i+2);
2099 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
2100 MachineOperand &MO = MI->getOperand(i);
2106 static MachineInstr *FuseInst(MachineFunction &MF,
2107 unsigned Opcode, unsigned OpNo,
2108 const SmallVectorImpl<MachineOperand> &MOs,
2109 MachineInstr *MI, const TargetInstrInfo &TII) {
2110 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2111 MI->getDebugLoc(), true);
2112 MachineInstrBuilder MIB(NewMI);
2114 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2115 MachineOperand &MO = MI->getOperand(i);
2117 assert(MO.isReg() && "Expected to fold into reg operand!");
2118 unsigned NumAddrOps = MOs.size();
2119 for (unsigned i = 0; i != NumAddrOps; ++i)
2120 MIB.addOperand(MOs[i]);
2121 if (NumAddrOps < 4) // FrameIndex only
2130 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
2131 const SmallVectorImpl<MachineOperand> &MOs,
2133 MachineFunction &MF = *MI->getParent()->getParent();
2134 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
2136 unsigned NumAddrOps = MOs.size();
2137 for (unsigned i = 0; i != NumAddrOps; ++i)
2138 MIB.addOperand(MOs[i]);
2139 if (NumAddrOps < 4) // FrameIndex only
2141 return MIB.addImm(0);
2145 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2146 MachineInstr *MI, unsigned i,
2147 const SmallVectorImpl<MachineOperand> &MOs) const{
2148 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2149 bool isTwoAddrFold = false;
2150 unsigned NumOps = MI->getDesc().getNumOperands();
2151 bool isTwoAddr = NumOps > 1 &&
2152 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2154 MachineInstr *NewMI = NULL;
2155 // Folding a memory location into the two-address part of a two-address
2156 // instruction is different than folding it other places. It requires
2157 // replacing the *two* registers with the memory location.
2158 if (isTwoAddr && NumOps >= 2 && i < 2 &&
2159 MI->getOperand(0).isReg() &&
2160 MI->getOperand(1).isReg() &&
2161 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
2162 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2163 isTwoAddrFold = true;
2164 } else if (i == 0) { // If operand 0
2165 if (MI->getOpcode() == X86::MOV16r0)
2166 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
2167 else if (MI->getOpcode() == X86::MOV32r0)
2168 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
2169 else if (MI->getOpcode() == X86::MOV64r0)
2170 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2171 else if (MI->getOpcode() == X86::MOV8r0)
2172 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2176 OpcodeTablePtr = &RegOp2MemOpTable0;
2177 } else if (i == 1) {
2178 OpcodeTablePtr = &RegOp2MemOpTable1;
2179 } else if (i == 2) {
2180 OpcodeTablePtr = &RegOp2MemOpTable2;
2183 // If table selected...
2184 if (OpcodeTablePtr) {
2185 // Find the Opcode to fuse
2186 DenseMap<unsigned*, unsigned>::iterator I =
2187 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
2188 if (I != OpcodeTablePtr->end()) {
2190 NewMI = FuseTwoAddrInst(MF, I->second, MOs, MI, *this);
2192 NewMI = FuseInst(MF, I->second, i, MOs, MI, *this);
2198 if (PrintFailedFusing)
2199 cerr << "We failed to fuse operand " << i << " in " << *MI;
2204 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2206 const SmallVectorImpl<unsigned> &Ops,
2207 int FrameIndex) const {
2208 // Check switch flag
2209 if (NoFusing) return NULL;
2211 const MachineFrameInfo *MFI = MF.getFrameInfo();
2212 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2213 // FIXME: Move alignment requirement into tables?
2214 if (Alignment < 16) {
2215 switch (MI->getOpcode()) {
2217 // Not always safe to fold movsd into these instructions since their load
2218 // folding variants expects the address to be 16 byte aligned.
2219 case X86::FsANDNPDrr:
2220 case X86::FsANDNPSrr:
2221 case X86::FsANDPDrr:
2222 case X86::FsANDPSrr:
2225 case X86::FsXORPDrr:
2226 case X86::FsXORPSrr:
2231 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2232 unsigned NewOpc = 0;
2233 switch (MI->getOpcode()) {
2234 default: return NULL;
2235 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2236 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2237 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2238 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2240 // Change to CMPXXri r, 0 first.
2241 MI->setDesc(get(NewOpc));
2242 MI->getOperand(1).ChangeToImmediate(0);
2243 } else if (Ops.size() != 1)
2246 SmallVector<MachineOperand,4> MOs;
2247 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2248 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2251 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2253 const SmallVectorImpl<unsigned> &Ops,
2254 MachineInstr *LoadMI) const {
2255 // Check switch flag
2256 if (NoFusing) return NULL;
2258 // Determine the alignment of the load.
2259 unsigned Alignment = 0;
2260 if (LoadMI->hasOneMemOperand())
2261 Alignment = LoadMI->memoperands_begin()->getAlignment();
2263 // FIXME: Move alignment requirement into tables?
2264 if (Alignment < 16) {
2265 switch (MI->getOpcode()) {
2267 // Not always safe to fold movsd into these instructions since their load
2268 // folding variants expects the address to be 16 byte aligned.
2269 case X86::FsANDNPDrr:
2270 case X86::FsANDNPSrr:
2271 case X86::FsANDPDrr:
2272 case X86::FsANDPSrr:
2275 case X86::FsXORPDrr:
2276 case X86::FsXORPSrr:
2281 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2282 unsigned NewOpc = 0;
2283 switch (MI->getOpcode()) {
2284 default: return NULL;
2285 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2286 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2287 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2288 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2290 // Change to CMPXXri r, 0 first.
2291 MI->setDesc(get(NewOpc));
2292 MI->getOperand(1).ChangeToImmediate(0);
2293 } else if (Ops.size() != 1)
2296 SmallVector<MachineOperand,X86AddrNumOperands> MOs;
2297 if (LoadMI->getOpcode() == X86::V_SET0 ||
2298 LoadMI->getOpcode() == X86::V_SETALLONES) {
2299 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
2300 // Create a constant-pool entry and operands to load from it.
2302 // x86-32 PIC requires a PIC base register for constant pools.
2303 unsigned PICBase = 0;
2304 if (TM.getRelocationModel() == Reloc::PIC_ &&
2305 !TM.getSubtarget<X86Subtarget>().is64Bit())
2306 // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
2307 // This doesn't work for several reasons.
2308 // 1. GlobalBaseReg may have been spilled.
2309 // 2. It may not be live at MI.
2312 // Create a v4i32 constant-pool entry.
2313 MachineConstantPool &MCP = *MF.getConstantPool();
2314 const VectorType *Ty = VectorType::get(Type::Int32Ty, 4);
2315 Constant *C = LoadMI->getOpcode() == X86::V_SET0 ?
2316 MF.getFunction()->getContext()->getNullValue(Ty) :
2317 ConstantVector::getAllOnesValue(Ty);
2318 unsigned CPI = MCP.getConstantPoolIndex(C, 16);
2320 // Create operands to load from the constant pool entry.
2321 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
2322 MOs.push_back(MachineOperand::CreateImm(1));
2323 MOs.push_back(MachineOperand::CreateReg(0, false));
2324 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
2325 MOs.push_back(MachineOperand::CreateReg(0, false));
2327 // Folding a normal load. Just copy the load's address operands.
2328 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2329 for (unsigned i = NumOps - X86AddrNumOperands; i != NumOps; ++i)
2330 MOs.push_back(LoadMI->getOperand(i));
2332 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2336 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2337 const SmallVectorImpl<unsigned> &Ops) const {
2338 // Check switch flag
2339 if (NoFusing) return 0;
2341 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2342 switch (MI->getOpcode()) {
2343 default: return false;
2352 if (Ops.size() != 1)
2355 unsigned OpNum = Ops[0];
2356 unsigned Opc = MI->getOpcode();
2357 unsigned NumOps = MI->getDesc().getNumOperands();
2358 bool isTwoAddr = NumOps > 1 &&
2359 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2361 // Folding a memory location into the two-address part of a two-address
2362 // instruction is different than folding it other places. It requires
2363 // replacing the *two* registers with the memory location.
2364 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2365 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2366 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2367 } else if (OpNum == 0) { // If operand 0
2376 OpcodeTablePtr = &RegOp2MemOpTable0;
2377 } else if (OpNum == 1) {
2378 OpcodeTablePtr = &RegOp2MemOpTable1;
2379 } else if (OpNum == 2) {
2380 OpcodeTablePtr = &RegOp2MemOpTable2;
2383 if (OpcodeTablePtr) {
2384 // Find the Opcode to fuse
2385 DenseMap<unsigned*, unsigned>::iterator I =
2386 OpcodeTablePtr->find((unsigned*)Opc);
2387 if (I != OpcodeTablePtr->end())
2393 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2394 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2395 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2396 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2397 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2398 if (I == MemOp2RegOpTable.end())
2400 DebugLoc dl = MI->getDebugLoc();
2401 unsigned Opc = I->second.first;
2402 unsigned Index = I->second.second & 0xf;
2403 bool FoldedLoad = I->second.second & (1 << 4);
2404 bool FoldedStore = I->second.second & (1 << 5);
2405 if (UnfoldLoad && !FoldedLoad)
2407 UnfoldLoad &= FoldedLoad;
2408 if (UnfoldStore && !FoldedStore)
2410 UnfoldStore &= FoldedStore;
2412 const TargetInstrDesc &TID = get(Opc);
2413 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2414 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2415 ? RI.getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2416 SmallVector<MachineOperand, X86AddrNumOperands> AddrOps;
2417 SmallVector<MachineOperand,2> BeforeOps;
2418 SmallVector<MachineOperand,2> AfterOps;
2419 SmallVector<MachineOperand,4> ImpOps;
2420 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2421 MachineOperand &Op = MI->getOperand(i);
2422 if (i >= Index && i < Index + X86AddrNumOperands)
2423 AddrOps.push_back(Op);
2424 else if (Op.isReg() && Op.isImplicit())
2425 ImpOps.push_back(Op);
2427 BeforeOps.push_back(Op);
2429 AfterOps.push_back(Op);
2432 // Emit the load instruction.
2434 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2436 // Address operands cannot be marked isKill.
2437 for (unsigned i = 1; i != 1 + X86AddrNumOperands; ++i) {
2438 MachineOperand &MO = NewMIs[0]->getOperand(i);
2440 MO.setIsKill(false);
2445 // Emit the data processing instruction.
2446 MachineInstr *DataMI = MF.CreateMachineInstr(TID, MI->getDebugLoc(), true);
2447 MachineInstrBuilder MIB(DataMI);
2450 MIB.addReg(Reg, RegState::Define);
2451 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2452 MIB.addOperand(BeforeOps[i]);
2455 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2456 MIB.addOperand(AfterOps[i]);
2457 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2458 MachineOperand &MO = ImpOps[i];
2459 MIB.addReg(MO.getReg(),
2460 getDefRegState(MO.isDef()) |
2461 RegState::Implicit |
2462 getKillRegState(MO.isKill()) |
2463 getDeadRegState(MO.isDead()) |
2464 getUndefRegState(MO.isUndef()));
2466 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2467 unsigned NewOpc = 0;
2468 switch (DataMI->getOpcode()) {
2470 case X86::CMP64ri32:
2474 MachineOperand &MO0 = DataMI->getOperand(0);
2475 MachineOperand &MO1 = DataMI->getOperand(1);
2476 if (MO1.getImm() == 0) {
2477 switch (DataMI->getOpcode()) {
2479 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2480 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2481 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2482 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2484 DataMI->setDesc(get(NewOpc));
2485 MO1.ChangeToRegister(MO0.getReg(), false);
2489 NewMIs.push_back(DataMI);
2491 // Emit the store instruction.
2493 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2494 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2495 ? RI.getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2496 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2503 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2504 SmallVectorImpl<SDNode*> &NewNodes) const {
2505 if (!N->isMachineOpcode())
2508 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2509 MemOp2RegOpTable.find((unsigned*)N->getMachineOpcode());
2510 if (I == MemOp2RegOpTable.end())
2512 unsigned Opc = I->second.first;
2513 unsigned Index = I->second.second & 0xf;
2514 bool FoldedLoad = I->second.second & (1 << 4);
2515 bool FoldedStore = I->second.second & (1 << 5);
2516 const TargetInstrDesc &TID = get(Opc);
2517 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2518 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2519 ? RI.getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2520 unsigned NumDefs = TID.NumDefs;
2521 std::vector<SDValue> AddrOps;
2522 std::vector<SDValue> BeforeOps;
2523 std::vector<SDValue> AfterOps;
2524 DebugLoc dl = N->getDebugLoc();
2525 unsigned NumOps = N->getNumOperands();
2526 for (unsigned i = 0; i != NumOps-1; ++i) {
2527 SDValue Op = N->getOperand(i);
2528 if (i >= Index-NumDefs && i < Index-NumDefs + X86AddrNumOperands)
2529 AddrOps.push_back(Op);
2530 else if (i < Index-NumDefs)
2531 BeforeOps.push_back(Op);
2532 else if (i > Index-NumDefs)
2533 AfterOps.push_back(Op);
2535 SDValue Chain = N->getOperand(NumOps-1);
2536 AddrOps.push_back(Chain);
2538 // Emit the load instruction.
2540 const MachineFunction &MF = DAG.getMachineFunction();
2542 MVT VT = *RC->vt_begin();
2543 bool isAligned = (RI.getStackAlignment() >= 16) ||
2544 RI.needsStackRealignment(MF);
2545 Load = DAG.getTargetNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
2546 VT, MVT::Other, &AddrOps[0], AddrOps.size());
2547 NewNodes.push_back(Load);
2550 // Emit the data processing instruction.
2551 std::vector<MVT> VTs;
2552 const TargetRegisterClass *DstRC = 0;
2553 if (TID.getNumDefs() > 0) {
2554 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2555 DstRC = DstTOI.isLookupPtrRegClass()
2556 ? RI.getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2557 VTs.push_back(*DstRC->vt_begin());
2559 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2560 MVT VT = N->getValueType(i);
2561 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2565 BeforeOps.push_back(SDValue(Load, 0));
2566 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2567 SDNode *NewNode= DAG.getTargetNode(Opc, dl, VTs, &BeforeOps[0],
2569 NewNodes.push_back(NewNode);
2571 // Emit the store instruction.
2574 AddrOps.push_back(SDValue(NewNode, 0));
2575 AddrOps.push_back(Chain);
2576 bool isAligned = (RI.getStackAlignment() >= 16) ||
2577 RI.needsStackRealignment(MF);
2578 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(0, DstRC,
2581 &AddrOps[0], AddrOps.size());
2582 NewNodes.push_back(Store);
2588 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2589 bool UnfoldLoad, bool UnfoldStore) const {
2590 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2591 MemOp2RegOpTable.find((unsigned*)Opc);
2592 if (I == MemOp2RegOpTable.end())
2594 bool FoldedLoad = I->second.second & (1 << 4);
2595 bool FoldedStore = I->second.second & (1 << 5);
2596 if (UnfoldLoad && !FoldedLoad)
2598 if (UnfoldStore && !FoldedStore)
2600 return I->second.first;
2603 bool X86InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
2604 if (MBB.empty()) return false;
2606 switch (MBB.back().getOpcode()) {
2607 case X86::TCRETURNri:
2608 case X86::TCRETURNdi:
2609 case X86::RET: // Return.
2614 case X86::JMP: // Uncond branch.
2615 case X86::JMP32r: // Indirect branch.
2616 case X86::JMP64r: // Indirect branch (64-bit).
2617 case X86::JMP32m: // Indirect branch through mem.
2618 case X86::JMP64m: // Indirect branch through mem (64-bit).
2620 default: return false;
2625 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
2626 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2627 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
2628 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
2630 Cond[0].setImm(GetOppositeBranchCondition(CC));
2635 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
2636 // FIXME: Return false for x87 stack register classes for now. We can't
2637 // allow any loads of these registers before FpGet_ST0_80.
2638 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
2639 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
2642 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2643 switch (Desc->TSFlags & X86II::ImmMask) {
2644 case X86II::Imm8: return 1;
2645 case X86II::Imm16: return 2;
2646 case X86II::Imm32: return 4;
2647 case X86II::Imm64: return 8;
2648 default: LLVM_UNREACHABLE("Immediate size not set!");
2653 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2654 /// e.g. r8, xmm8, etc.
2655 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2656 if (!MO.isReg()) return false;
2657 switch (MO.getReg()) {
2659 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2660 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2661 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2662 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2663 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2664 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2665 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2666 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2667 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2668 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2675 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2676 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2677 /// size, and 3) use of X86-64 extended registers.
2678 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2680 const TargetInstrDesc &Desc = MI.getDesc();
2682 // Pseudo instructions do not need REX prefix byte.
2683 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2685 if (Desc.TSFlags & X86II::REX_W)
2688 unsigned NumOps = Desc.getNumOperands();
2690 bool isTwoAddr = NumOps > 1 &&
2691 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2693 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2694 unsigned i = isTwoAddr ? 1 : 0;
2695 for (unsigned e = NumOps; i != e; ++i) {
2696 const MachineOperand& MO = MI.getOperand(i);
2698 unsigned Reg = MO.getReg();
2699 if (isX86_64NonExtLowByteReg(Reg))
2704 switch (Desc.TSFlags & X86II::FormMask) {
2705 case X86II::MRMInitReg:
2706 if (isX86_64ExtendedReg(MI.getOperand(0)))
2707 REX |= (1 << 0) | (1 << 2);
2709 case X86II::MRMSrcReg: {
2710 if (isX86_64ExtendedReg(MI.getOperand(0)))
2712 i = isTwoAddr ? 2 : 1;
2713 for (unsigned e = NumOps; i != e; ++i) {
2714 const MachineOperand& MO = MI.getOperand(i);
2715 if (isX86_64ExtendedReg(MO))
2720 case X86II::MRMSrcMem: {
2721 if (isX86_64ExtendedReg(MI.getOperand(0)))
2724 i = isTwoAddr ? 2 : 1;
2725 for (; i != NumOps; ++i) {
2726 const MachineOperand& MO = MI.getOperand(i);
2728 if (isX86_64ExtendedReg(MO))
2735 case X86II::MRM0m: case X86II::MRM1m:
2736 case X86II::MRM2m: case X86II::MRM3m:
2737 case X86II::MRM4m: case X86II::MRM5m:
2738 case X86II::MRM6m: case X86II::MRM7m:
2739 case X86II::MRMDestMem: {
2740 unsigned e = (isTwoAddr ? X86AddrNumOperands+1 : X86AddrNumOperands);
2741 i = isTwoAddr ? 1 : 0;
2742 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2745 for (; i != e; ++i) {
2746 const MachineOperand& MO = MI.getOperand(i);
2748 if (isX86_64ExtendedReg(MO))
2756 if (isX86_64ExtendedReg(MI.getOperand(0)))
2758 i = isTwoAddr ? 2 : 1;
2759 for (unsigned e = NumOps; i != e; ++i) {
2760 const MachineOperand& MO = MI.getOperand(i);
2761 if (isX86_64ExtendedReg(MO))
2771 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2772 /// relative block address instruction
2774 static unsigned sizePCRelativeBlockAddress() {
2778 /// sizeGlobalAddress - Give the size of the emission of this global address
2780 static unsigned sizeGlobalAddress(bool dword) {
2781 return dword ? 8 : 4;
2784 /// sizeConstPoolAddress - Give the size of the emission of this constant
2787 static unsigned sizeConstPoolAddress(bool dword) {
2788 return dword ? 8 : 4;
2791 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2794 static unsigned sizeExternalSymbolAddress(bool dword) {
2795 return dword ? 8 : 4;
2798 /// sizeJumpTableAddress - Give the size of the emission of this jump
2801 static unsigned sizeJumpTableAddress(bool dword) {
2802 return dword ? 8 : 4;
2805 static unsigned sizeConstant(unsigned Size) {
2809 static unsigned sizeRegModRMByte(){
2813 static unsigned sizeSIBByte(){
2817 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2818 unsigned FinalSize = 0;
2819 // If this is a simple integer displacement that doesn't require a relocation.
2821 FinalSize += sizeConstant(4);
2825 // Otherwise, this is something that requires a relocation.
2826 if (RelocOp->isGlobal()) {
2827 FinalSize += sizeGlobalAddress(false);
2828 } else if (RelocOp->isCPI()) {
2829 FinalSize += sizeConstPoolAddress(false);
2830 } else if (RelocOp->isJTI()) {
2831 FinalSize += sizeJumpTableAddress(false);
2833 LLVM_UNREACHABLE("Unknown value to relocate!");
2838 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2839 bool IsPIC, bool Is64BitMode) {
2840 const MachineOperand &Op3 = MI.getOperand(Op+3);
2842 const MachineOperand *DispForReloc = 0;
2843 unsigned FinalSize = 0;
2845 // Figure out what sort of displacement we have to handle here.
2846 if (Op3.isGlobal()) {
2847 DispForReloc = &Op3;
2848 } else if (Op3.isCPI()) {
2849 if (Is64BitMode || IsPIC) {
2850 DispForReloc = &Op3;
2854 } else if (Op3.isJTI()) {
2855 if (Is64BitMode || IsPIC) {
2856 DispForReloc = &Op3;
2864 const MachineOperand &Base = MI.getOperand(Op);
2865 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2867 unsigned BaseReg = Base.getReg();
2869 // Is a SIB byte needed?
2870 if ((!Is64BitMode || DispForReloc || BaseReg != 0) &&
2871 IndexReg.getReg() == 0 &&
2872 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2873 if (BaseReg == 0) { // Just a displacement?
2874 // Emit special case [disp32] encoding
2876 FinalSize += getDisplacementFieldSize(DispForReloc);
2878 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2879 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2880 // Emit simple indirect register encoding... [EAX] f.e.
2882 // Be pessimistic and assume it's a disp32, not a disp8
2884 // Emit the most general non-SIB encoding: [REG+disp32]
2886 FinalSize += getDisplacementFieldSize(DispForReloc);
2890 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2891 assert(IndexReg.getReg() != X86::ESP &&
2892 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2894 bool ForceDisp32 = false;
2895 if (BaseReg == 0 || DispForReloc) {
2896 // Emit the normal disp32 encoding.
2903 FinalSize += sizeSIBByte();
2905 // Do we need to output a displacement?
2906 if (DispVal != 0 || ForceDisp32) {
2907 FinalSize += getDisplacementFieldSize(DispForReloc);
2914 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2915 const TargetInstrDesc *Desc,
2916 bool IsPIC, bool Is64BitMode) {
2918 unsigned Opcode = Desc->Opcode;
2919 unsigned FinalSize = 0;
2921 // Emit the lock opcode prefix as needed.
2922 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2924 // Emit segment override opcode prefix as needed.
2925 switch (Desc->TSFlags & X86II::SegOvrMask) {
2930 default: LLVM_UNREACHABLE("Invalid segment!");
2931 case 0: break; // No segment override!
2934 // Emit the repeat opcode prefix as needed.
2935 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2937 // Emit the operand size opcode prefix as needed.
2938 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2940 // Emit the address size opcode prefix as needed.
2941 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2943 bool Need0FPrefix = false;
2944 switch (Desc->TSFlags & X86II::Op0Mask) {
2945 case X86II::TB: // Two-byte opcode prefix
2946 case X86II::T8: // 0F 38
2947 case X86II::TA: // 0F 3A
2948 Need0FPrefix = true;
2950 case X86II::REP: break; // already handled.
2951 case X86II::XS: // F3 0F
2953 Need0FPrefix = true;
2955 case X86II::XD: // F2 0F
2957 Need0FPrefix = true;
2959 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2960 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2962 break; // Two-byte opcode prefix
2963 default: LLVM_UNREACHABLE("Invalid prefix!");
2964 case 0: break; // No prefix!
2969 unsigned REX = X86InstrInfo::determineREX(MI);
2974 // 0x0F escape code must be emitted just before the opcode.
2978 switch (Desc->TSFlags & X86II::Op0Mask) {
2979 case X86II::T8: // 0F 38
2982 case X86II::TA: // 0F 3A
2987 // If this is a two-address instruction, skip one of the register operands.
2988 unsigned NumOps = Desc->getNumOperands();
2990 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2992 else if (NumOps > 2 && Desc->getOperandConstraint(NumOps-1, TOI::TIED_TO)== 0)
2993 // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32
2996 switch (Desc->TSFlags & X86II::FormMask) {
2997 default: LLVM_UNREACHABLE("Unknown FormMask value in X86 MachineCodeEmitter!");
2999 // Remember the current PC offset, this is the PIC relocation
3004 case TargetInstrInfo::INLINEASM: {
3005 const MachineFunction *MF = MI.getParent()->getParent();
3006 const char *AsmStr = MI.getOperand(0).getSymbolName();
3007 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
3008 FinalSize += AI->getInlineAsmLength(AsmStr);
3011 case TargetInstrInfo::DBG_LABEL:
3012 case TargetInstrInfo::EH_LABEL:
3014 case TargetInstrInfo::IMPLICIT_DEF:
3015 case TargetInstrInfo::DECLARE:
3016 case X86::DWARF_LOC:
3017 case X86::FP_REG_KILL:
3019 case X86::MOVPC32r: {
3020 // This emits the "call" portion of this pseudo instruction.
3022 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3031 if (CurOp != NumOps) {
3032 const MachineOperand &MO = MI.getOperand(CurOp++);
3034 FinalSize += sizePCRelativeBlockAddress();
3035 } else if (MO.isGlobal()) {
3036 FinalSize += sizeGlobalAddress(false);
3037 } else if (MO.isSymbol()) {
3038 FinalSize += sizeExternalSymbolAddress(false);
3039 } else if (MO.isImm()) {
3040 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3042 LLVM_UNREACHABLE("Unknown RawFrm operand!");
3047 case X86II::AddRegFrm:
3051 if (CurOp != NumOps) {
3052 const MachineOperand &MO1 = MI.getOperand(CurOp++);
3053 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3055 FinalSize += sizeConstant(Size);
3058 if (Opcode == X86::MOV64ri)
3060 if (MO1.isGlobal()) {
3061 FinalSize += sizeGlobalAddress(dword);
3062 } else if (MO1.isSymbol())
3063 FinalSize += sizeExternalSymbolAddress(dword);
3064 else if (MO1.isCPI())
3065 FinalSize += sizeConstPoolAddress(dword);
3066 else if (MO1.isJTI())
3067 FinalSize += sizeJumpTableAddress(dword);
3072 case X86II::MRMDestReg: {
3074 FinalSize += sizeRegModRMByte();
3076 if (CurOp != NumOps) {
3078 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3082 case X86II::MRMDestMem: {
3084 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
3085 CurOp += X86AddrNumOperands + 1;
3086 if (CurOp != NumOps) {
3088 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3093 case X86II::MRMSrcReg:
3095 FinalSize += sizeRegModRMByte();
3097 if (CurOp != NumOps) {
3099 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3103 case X86II::MRMSrcMem: {
3105 if (Opcode == X86::LEA64r || Opcode == X86::LEA64_32r ||
3106 Opcode == X86::LEA16r || Opcode == X86::LEA32r)
3107 AddrOperands = X86AddrNumOperands - 1; // No segment register
3109 AddrOperands = X86AddrNumOperands;
3112 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
3113 CurOp += AddrOperands + 1;
3114 if (CurOp != NumOps) {
3116 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
3121 case X86II::MRM0r: case X86II::MRM1r:
3122 case X86II::MRM2r: case X86II::MRM3r:
3123 case X86II::MRM4r: case X86II::MRM5r:
3124 case X86II::MRM6r: case X86II::MRM7r:
3126 if (Desc->getOpcode() == X86::LFENCE ||
3127 Desc->getOpcode() == X86::MFENCE) {
3128 // Special handling of lfence and mfence;
3129 FinalSize += sizeRegModRMByte();
3130 } else if (Desc->getOpcode() == X86::MONITOR ||
3131 Desc->getOpcode() == X86::MWAIT) {
3132 // Special handling of monitor and mwait.
3133 FinalSize += sizeRegModRMByte() + 1; // +1 for the opcode.
3136 FinalSize += sizeRegModRMByte();
3139 if (CurOp != NumOps) {
3140 const MachineOperand &MO1 = MI.getOperand(CurOp++);
3141 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3143 FinalSize += sizeConstant(Size);
3146 if (Opcode == X86::MOV64ri32)
3148 if (MO1.isGlobal()) {
3149 FinalSize += sizeGlobalAddress(dword);
3150 } else if (MO1.isSymbol())
3151 FinalSize += sizeExternalSymbolAddress(dword);
3152 else if (MO1.isCPI())
3153 FinalSize += sizeConstPoolAddress(dword);
3154 else if (MO1.isJTI())
3155 FinalSize += sizeJumpTableAddress(dword);
3160 case X86II::MRM0m: case X86II::MRM1m:
3161 case X86II::MRM2m: case X86II::MRM3m:
3162 case X86II::MRM4m: case X86II::MRM5m:
3163 case X86II::MRM6m: case X86II::MRM7m: {
3166 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
3167 CurOp += X86AddrNumOperands;
3169 if (CurOp != NumOps) {
3170 const MachineOperand &MO = MI.getOperand(CurOp++);
3171 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3173 FinalSize += sizeConstant(Size);
3176 if (Opcode == X86::MOV64mi32)
3178 if (MO.isGlobal()) {
3179 FinalSize += sizeGlobalAddress(dword);
3180 } else if (MO.isSymbol())
3181 FinalSize += sizeExternalSymbolAddress(dword);
3182 else if (MO.isCPI())
3183 FinalSize += sizeConstPoolAddress(dword);
3184 else if (MO.isJTI())
3185 FinalSize += sizeJumpTableAddress(dword);
3191 case X86II::MRMInitReg:
3193 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
3194 FinalSize += sizeRegModRMByte();
3199 if (!Desc->isVariadic() && CurOp != NumOps) {
3201 raw_string_ostream Msg(msg);
3202 Msg << "Cannot determine size: " << MI;
3203 llvm_report_error(Msg.str());
3211 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
3212 const TargetInstrDesc &Desc = MI->getDesc();
3213 bool IsPIC = TM.getRelocationModel() == Reloc::PIC_;
3214 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
3215 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
3216 if (Desc.getOpcode() == X86::MOVPC32r)
3217 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);
3221 /// getGlobalBaseReg - Return a virtual register initialized with the
3222 /// the global base register value. Output instructions required to
3223 /// initialize the register in the function entry block, if necessary.
3225 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3226 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3227 "X86-64 PIC uses RIP relative addressing");
3229 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3230 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3231 if (GlobalBaseReg != 0)
3232 return GlobalBaseReg;
3234 // Insert the set of GlobalBaseReg into the first MBB of the function
3235 MachineBasicBlock &FirstMBB = MF->front();
3236 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3237 DebugLoc DL = DebugLoc::getUnknownLoc();
3238 if (MBBI != FirstMBB.end()) DL = MBBI->getDebugLoc();
3239 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3240 unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3242 const TargetInstrInfo *TII = TM.getInstrInfo();
3243 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3244 // only used in JIT code emission as displacement to pc.
3245 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
3247 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3248 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
3249 if (TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3250 GlobalBaseReg = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3251 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
3252 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
3253 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_", 0,
3254 X86II::MO_GOT_ABSOLUTE_ADDRESS);
3259 X86FI->setGlobalBaseReg(GlobalBaseReg);
3260 return GlobalBaseReg;