1 //===- X86InstrInfo.cpp - X86 Instruction Information -----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the X86 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "X86InstrInfo.h"
16 #include "X86GenInstrInfo.inc"
17 #include "X86InstrBuilder.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/LiveVariables.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/Target/TargetAsmInfo.h"
36 NoFusing("disable-spill-fusing",
37 cl::desc("Disable fusing of spill code into instructions"));
39 PrintFailedFusing("print-failed-fuse-candidates",
40 cl::desc("Print instructions that the allocator wants to"
41 " fuse, but the X86 backend currently can't"),
44 ReMatPICStubLoad("remat-pic-stub-load",
45 cl::desc("Re-materialize load from stub in PIC mode"),
46 cl::init(false), cl::Hidden);
49 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
50 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
51 TM(tm), RI(tm, *this) {
52 SmallVector<unsigned,16> AmbEntries;
53 static const unsigned OpTbl2Addr[][2] = {
54 { X86::ADC32ri, X86::ADC32mi },
55 { X86::ADC32ri8, X86::ADC32mi8 },
56 { X86::ADC32rr, X86::ADC32mr },
57 { X86::ADC64ri32, X86::ADC64mi32 },
58 { X86::ADC64ri8, X86::ADC64mi8 },
59 { X86::ADC64rr, X86::ADC64mr },
60 { X86::ADD16ri, X86::ADD16mi },
61 { X86::ADD16ri8, X86::ADD16mi8 },
62 { X86::ADD16rr, X86::ADD16mr },
63 { X86::ADD32ri, X86::ADD32mi },
64 { X86::ADD32ri8, X86::ADD32mi8 },
65 { X86::ADD32rr, X86::ADD32mr },
66 { X86::ADD64ri32, X86::ADD64mi32 },
67 { X86::ADD64ri8, X86::ADD64mi8 },
68 { X86::ADD64rr, X86::ADD64mr },
69 { X86::ADD8ri, X86::ADD8mi },
70 { X86::ADD8rr, X86::ADD8mr },
71 { X86::AND16ri, X86::AND16mi },
72 { X86::AND16ri8, X86::AND16mi8 },
73 { X86::AND16rr, X86::AND16mr },
74 { X86::AND32ri, X86::AND32mi },
75 { X86::AND32ri8, X86::AND32mi8 },
76 { X86::AND32rr, X86::AND32mr },
77 { X86::AND64ri32, X86::AND64mi32 },
78 { X86::AND64ri8, X86::AND64mi8 },
79 { X86::AND64rr, X86::AND64mr },
80 { X86::AND8ri, X86::AND8mi },
81 { X86::AND8rr, X86::AND8mr },
82 { X86::DEC16r, X86::DEC16m },
83 { X86::DEC32r, X86::DEC32m },
84 { X86::DEC64_16r, X86::DEC64_16m },
85 { X86::DEC64_32r, X86::DEC64_32m },
86 { X86::DEC64r, X86::DEC64m },
87 { X86::DEC8r, X86::DEC8m },
88 { X86::INC16r, X86::INC16m },
89 { X86::INC32r, X86::INC32m },
90 { X86::INC64_16r, X86::INC64_16m },
91 { X86::INC64_32r, X86::INC64_32m },
92 { X86::INC64r, X86::INC64m },
93 { X86::INC8r, X86::INC8m },
94 { X86::NEG16r, X86::NEG16m },
95 { X86::NEG32r, X86::NEG32m },
96 { X86::NEG64r, X86::NEG64m },
97 { X86::NEG8r, X86::NEG8m },
98 { X86::NOT16r, X86::NOT16m },
99 { X86::NOT32r, X86::NOT32m },
100 { X86::NOT64r, X86::NOT64m },
101 { X86::NOT8r, X86::NOT8m },
102 { X86::OR16ri, X86::OR16mi },
103 { X86::OR16ri8, X86::OR16mi8 },
104 { X86::OR16rr, X86::OR16mr },
105 { X86::OR32ri, X86::OR32mi },
106 { X86::OR32ri8, X86::OR32mi8 },
107 { X86::OR32rr, X86::OR32mr },
108 { X86::OR64ri32, X86::OR64mi32 },
109 { X86::OR64ri8, X86::OR64mi8 },
110 { X86::OR64rr, X86::OR64mr },
111 { X86::OR8ri, X86::OR8mi },
112 { X86::OR8rr, X86::OR8mr },
113 { X86::ROL16r1, X86::ROL16m1 },
114 { X86::ROL16rCL, X86::ROL16mCL },
115 { X86::ROL16ri, X86::ROL16mi },
116 { X86::ROL32r1, X86::ROL32m1 },
117 { X86::ROL32rCL, X86::ROL32mCL },
118 { X86::ROL32ri, X86::ROL32mi },
119 { X86::ROL64r1, X86::ROL64m1 },
120 { X86::ROL64rCL, X86::ROL64mCL },
121 { X86::ROL64ri, X86::ROL64mi },
122 { X86::ROL8r1, X86::ROL8m1 },
123 { X86::ROL8rCL, X86::ROL8mCL },
124 { X86::ROL8ri, X86::ROL8mi },
125 { X86::ROR16r1, X86::ROR16m1 },
126 { X86::ROR16rCL, X86::ROR16mCL },
127 { X86::ROR16ri, X86::ROR16mi },
128 { X86::ROR32r1, X86::ROR32m1 },
129 { X86::ROR32rCL, X86::ROR32mCL },
130 { X86::ROR32ri, X86::ROR32mi },
131 { X86::ROR64r1, X86::ROR64m1 },
132 { X86::ROR64rCL, X86::ROR64mCL },
133 { X86::ROR64ri, X86::ROR64mi },
134 { X86::ROR8r1, X86::ROR8m1 },
135 { X86::ROR8rCL, X86::ROR8mCL },
136 { X86::ROR8ri, X86::ROR8mi },
137 { X86::SAR16r1, X86::SAR16m1 },
138 { X86::SAR16rCL, X86::SAR16mCL },
139 { X86::SAR16ri, X86::SAR16mi },
140 { X86::SAR32r1, X86::SAR32m1 },
141 { X86::SAR32rCL, X86::SAR32mCL },
142 { X86::SAR32ri, X86::SAR32mi },
143 { X86::SAR64r1, X86::SAR64m1 },
144 { X86::SAR64rCL, X86::SAR64mCL },
145 { X86::SAR64ri, X86::SAR64mi },
146 { X86::SAR8r1, X86::SAR8m1 },
147 { X86::SAR8rCL, X86::SAR8mCL },
148 { X86::SAR8ri, X86::SAR8mi },
149 { X86::SBB32ri, X86::SBB32mi },
150 { X86::SBB32ri8, X86::SBB32mi8 },
151 { X86::SBB32rr, X86::SBB32mr },
152 { X86::SBB64ri32, X86::SBB64mi32 },
153 { X86::SBB64ri8, X86::SBB64mi8 },
154 { X86::SBB64rr, X86::SBB64mr },
155 { X86::SHL16rCL, X86::SHL16mCL },
156 { X86::SHL16ri, X86::SHL16mi },
157 { X86::SHL32rCL, X86::SHL32mCL },
158 { X86::SHL32ri, X86::SHL32mi },
159 { X86::SHL64rCL, X86::SHL64mCL },
160 { X86::SHL64ri, X86::SHL64mi },
161 { X86::SHL8rCL, X86::SHL8mCL },
162 { X86::SHL8ri, X86::SHL8mi },
163 { X86::SHLD16rrCL, X86::SHLD16mrCL },
164 { X86::SHLD16rri8, X86::SHLD16mri8 },
165 { X86::SHLD32rrCL, X86::SHLD32mrCL },
166 { X86::SHLD32rri8, X86::SHLD32mri8 },
167 { X86::SHLD64rrCL, X86::SHLD64mrCL },
168 { X86::SHLD64rri8, X86::SHLD64mri8 },
169 { X86::SHR16r1, X86::SHR16m1 },
170 { X86::SHR16rCL, X86::SHR16mCL },
171 { X86::SHR16ri, X86::SHR16mi },
172 { X86::SHR32r1, X86::SHR32m1 },
173 { X86::SHR32rCL, X86::SHR32mCL },
174 { X86::SHR32ri, X86::SHR32mi },
175 { X86::SHR64r1, X86::SHR64m1 },
176 { X86::SHR64rCL, X86::SHR64mCL },
177 { X86::SHR64ri, X86::SHR64mi },
178 { X86::SHR8r1, X86::SHR8m1 },
179 { X86::SHR8rCL, X86::SHR8mCL },
180 { X86::SHR8ri, X86::SHR8mi },
181 { X86::SHRD16rrCL, X86::SHRD16mrCL },
182 { X86::SHRD16rri8, X86::SHRD16mri8 },
183 { X86::SHRD32rrCL, X86::SHRD32mrCL },
184 { X86::SHRD32rri8, X86::SHRD32mri8 },
185 { X86::SHRD64rrCL, X86::SHRD64mrCL },
186 { X86::SHRD64rri8, X86::SHRD64mri8 },
187 { X86::SUB16ri, X86::SUB16mi },
188 { X86::SUB16ri8, X86::SUB16mi8 },
189 { X86::SUB16rr, X86::SUB16mr },
190 { X86::SUB32ri, X86::SUB32mi },
191 { X86::SUB32ri8, X86::SUB32mi8 },
192 { X86::SUB32rr, X86::SUB32mr },
193 { X86::SUB64ri32, X86::SUB64mi32 },
194 { X86::SUB64ri8, X86::SUB64mi8 },
195 { X86::SUB64rr, X86::SUB64mr },
196 { X86::SUB8ri, X86::SUB8mi },
197 { X86::SUB8rr, X86::SUB8mr },
198 { X86::XOR16ri, X86::XOR16mi },
199 { X86::XOR16ri8, X86::XOR16mi8 },
200 { X86::XOR16rr, X86::XOR16mr },
201 { X86::XOR32ri, X86::XOR32mi },
202 { X86::XOR32ri8, X86::XOR32mi8 },
203 { X86::XOR32rr, X86::XOR32mr },
204 { X86::XOR64ri32, X86::XOR64mi32 },
205 { X86::XOR64ri8, X86::XOR64mi8 },
206 { X86::XOR64rr, X86::XOR64mr },
207 { X86::XOR8ri, X86::XOR8mi },
208 { X86::XOR8rr, X86::XOR8mr }
211 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
212 unsigned RegOp = OpTbl2Addr[i][0];
213 unsigned MemOp = OpTbl2Addr[i][1];
214 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp,
216 assert(false && "Duplicated entries?");
217 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
218 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
219 std::make_pair(RegOp,
221 AmbEntries.push_back(MemOp);
224 // If the third value is 1, then it's folding either a load or a store.
225 static const unsigned OpTbl0[][3] = {
226 { X86::CALL32r, X86::CALL32m, 1 },
227 { X86::CALL64r, X86::CALL64m, 1 },
228 { X86::CMP16ri, X86::CMP16mi, 1 },
229 { X86::CMP16ri8, X86::CMP16mi8, 1 },
230 { X86::CMP16rr, X86::CMP16mr, 1 },
231 { X86::CMP32ri, X86::CMP32mi, 1 },
232 { X86::CMP32ri8, X86::CMP32mi8, 1 },
233 { X86::CMP32rr, X86::CMP32mr, 1 },
234 { X86::CMP64ri32, X86::CMP64mi32, 1 },
235 { X86::CMP64ri8, X86::CMP64mi8, 1 },
236 { X86::CMP64rr, X86::CMP64mr, 1 },
237 { X86::CMP8ri, X86::CMP8mi, 1 },
238 { X86::CMP8rr, X86::CMP8mr, 1 },
239 { X86::DIV16r, X86::DIV16m, 1 },
240 { X86::DIV32r, X86::DIV32m, 1 },
241 { X86::DIV64r, X86::DIV64m, 1 },
242 { X86::DIV8r, X86::DIV8m, 1 },
243 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0 },
244 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
245 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
246 { X86::IDIV16r, X86::IDIV16m, 1 },
247 { X86::IDIV32r, X86::IDIV32m, 1 },
248 { X86::IDIV64r, X86::IDIV64m, 1 },
249 { X86::IDIV8r, X86::IDIV8m, 1 },
250 { X86::IMUL16r, X86::IMUL16m, 1 },
251 { X86::IMUL32r, X86::IMUL32m, 1 },
252 { X86::IMUL64r, X86::IMUL64m, 1 },
253 { X86::IMUL8r, X86::IMUL8m, 1 },
254 { X86::JMP32r, X86::JMP32m, 1 },
255 { X86::JMP64r, X86::JMP64m, 1 },
256 { X86::MOV16ri, X86::MOV16mi, 0 },
257 { X86::MOV16rr, X86::MOV16mr, 0 },
258 { X86::MOV16to16_, X86::MOV16_mr, 0 },
259 { X86::MOV32ri, X86::MOV32mi, 0 },
260 { X86::MOV32rr, X86::MOV32mr, 0 },
261 { X86::MOV32to32_, X86::MOV32_mr, 0 },
262 { X86::MOV64ri32, X86::MOV64mi32, 0 },
263 { X86::MOV64rr, X86::MOV64mr, 0 },
264 { X86::MOV8ri, X86::MOV8mi, 0 },
265 { X86::MOV8rr, X86::MOV8mr, 0 },
266 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
267 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
268 { X86::MOVDQArr, X86::MOVDQAmr, 0 },
269 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
270 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
271 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
272 { X86::MOVSDrr, X86::MOVSDmr, 0 },
273 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
274 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
275 { X86::MOVSSrr, X86::MOVSSmr, 0 },
276 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
277 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
278 { X86::MUL16r, X86::MUL16m, 1 },
279 { X86::MUL32r, X86::MUL32m, 1 },
280 { X86::MUL64r, X86::MUL64m, 1 },
281 { X86::MUL8r, X86::MUL8m, 1 },
282 { X86::SETAEr, X86::SETAEm, 0 },
283 { X86::SETAr, X86::SETAm, 0 },
284 { X86::SETBEr, X86::SETBEm, 0 },
285 { X86::SETBr, X86::SETBm, 0 },
286 { X86::SETEr, X86::SETEm, 0 },
287 { X86::SETGEr, X86::SETGEm, 0 },
288 { X86::SETGr, X86::SETGm, 0 },
289 { X86::SETLEr, X86::SETLEm, 0 },
290 { X86::SETLr, X86::SETLm, 0 },
291 { X86::SETNEr, X86::SETNEm, 0 },
292 { X86::SETNOr, X86::SETNOm, 0 },
293 { X86::SETNPr, X86::SETNPm, 0 },
294 { X86::SETNSr, X86::SETNSm, 0 },
295 { X86::SETOr, X86::SETOm, 0 },
296 { X86::SETPr, X86::SETPm, 0 },
297 { X86::SETSr, X86::SETSm, 0 },
298 { X86::TAILJMPr, X86::TAILJMPm, 1 },
299 { X86::TEST16ri, X86::TEST16mi, 1 },
300 { X86::TEST32ri, X86::TEST32mi, 1 },
301 { X86::TEST64ri32, X86::TEST64mi32, 1 },
302 { X86::TEST8ri, X86::TEST8mi, 1 }
305 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
306 unsigned RegOp = OpTbl0[i][0];
307 unsigned MemOp = OpTbl0[i][1];
308 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp,
310 assert(false && "Duplicated entries?");
311 unsigned FoldedLoad = OpTbl0[i][2];
312 // Index 0, folded load or store.
313 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
314 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
315 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
316 std::make_pair(RegOp, AuxInfo))).second)
317 AmbEntries.push_back(MemOp);
320 static const unsigned OpTbl1[][2] = {
321 { X86::CMP16rr, X86::CMP16rm },
322 { X86::CMP32rr, X86::CMP32rm },
323 { X86::CMP64rr, X86::CMP64rm },
324 { X86::CMP8rr, X86::CMP8rm },
325 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
326 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
327 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
328 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
329 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
330 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
331 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
332 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
333 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
334 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
335 { X86::FsMOVAPDrr, X86::MOVSDrm },
336 { X86::FsMOVAPSrr, X86::MOVSSrm },
337 { X86::IMUL16rri, X86::IMUL16rmi },
338 { X86::IMUL16rri8, X86::IMUL16rmi8 },
339 { X86::IMUL32rri, X86::IMUL32rmi },
340 { X86::IMUL32rri8, X86::IMUL32rmi8 },
341 { X86::IMUL64rri32, X86::IMUL64rmi32 },
342 { X86::IMUL64rri8, X86::IMUL64rmi8 },
343 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
344 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
345 { X86::Int_COMISDrr, X86::Int_COMISDrm },
346 { X86::Int_COMISSrr, X86::Int_COMISSrm },
347 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
348 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
349 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
350 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
351 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
352 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
353 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
354 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
355 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
356 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
357 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
358 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
359 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
360 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
361 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
362 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
363 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
364 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
365 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
366 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
367 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
368 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
369 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
370 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
371 { X86::MOV16rr, X86::MOV16rm },
372 { X86::MOV16to16_, X86::MOV16_rm },
373 { X86::MOV32rr, X86::MOV32rm },
374 { X86::MOV32to32_, X86::MOV32_rm },
375 { X86::MOV64rr, X86::MOV64rm },
376 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
377 { X86::MOV64toSDrr, X86::MOV64toSDrm },
378 { X86::MOV8rr, X86::MOV8rm },
379 { X86::MOVAPDrr, X86::MOVAPDrm },
380 { X86::MOVAPSrr, X86::MOVAPSrm },
381 { X86::MOVDDUPrr, X86::MOVDDUPrm },
382 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
383 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
384 { X86::MOVDQArr, X86::MOVDQArm },
385 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
386 { X86::MOVSDrr, X86::MOVSDrm },
387 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
388 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
389 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
390 { X86::MOVSSrr, X86::MOVSSrm },
391 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
392 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
393 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
394 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
395 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
396 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
397 { X86::MOVUPDrr, X86::MOVUPDrm },
398 { X86::MOVUPSrr, X86::MOVUPSrm },
399 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
400 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
401 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
402 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
403 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
404 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
405 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
406 { X86::MOVZX64rr32, X86::MOVZX64rm32 },
407 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
408 { X86::PSHUFDri, X86::PSHUFDmi },
409 { X86::PSHUFHWri, X86::PSHUFHWmi },
410 { X86::PSHUFLWri, X86::PSHUFLWmi },
411 { X86::RCPPSr, X86::RCPPSm },
412 { X86::RCPPSr_Int, X86::RCPPSm_Int },
413 { X86::RSQRTPSr, X86::RSQRTPSm },
414 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
415 { X86::RSQRTSSr, X86::RSQRTSSm },
416 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
417 { X86::SQRTPDr, X86::SQRTPDm },
418 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
419 { X86::SQRTPSr, X86::SQRTPSm },
420 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
421 { X86::SQRTSDr, X86::SQRTSDm },
422 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
423 { X86::SQRTSSr, X86::SQRTSSm },
424 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
425 { X86::TEST16rr, X86::TEST16rm },
426 { X86::TEST32rr, X86::TEST32rm },
427 { X86::TEST64rr, X86::TEST64rm },
428 { X86::TEST8rr, X86::TEST8rm },
429 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
430 { X86::UCOMISDrr, X86::UCOMISDrm },
431 { X86::UCOMISSrr, X86::UCOMISSrm }
434 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
435 unsigned RegOp = OpTbl1[i][0];
436 unsigned MemOp = OpTbl1[i][1];
437 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp,
439 assert(false && "Duplicated entries?");
440 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
441 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
442 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
443 std::make_pair(RegOp, AuxInfo))).second)
444 AmbEntries.push_back(MemOp);
447 static const unsigned OpTbl2[][2] = {
448 { X86::ADC32rr, X86::ADC32rm },
449 { X86::ADC64rr, X86::ADC64rm },
450 { X86::ADD16rr, X86::ADD16rm },
451 { X86::ADD32rr, X86::ADD32rm },
452 { X86::ADD64rr, X86::ADD64rm },
453 { X86::ADD8rr, X86::ADD8rm },
454 { X86::ADDPDrr, X86::ADDPDrm },
455 { X86::ADDPSrr, X86::ADDPSrm },
456 { X86::ADDSDrr, X86::ADDSDrm },
457 { X86::ADDSSrr, X86::ADDSSrm },
458 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
459 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
460 { X86::AND16rr, X86::AND16rm },
461 { X86::AND32rr, X86::AND32rm },
462 { X86::AND64rr, X86::AND64rm },
463 { X86::AND8rr, X86::AND8rm },
464 { X86::ANDNPDrr, X86::ANDNPDrm },
465 { X86::ANDNPSrr, X86::ANDNPSrm },
466 { X86::ANDPDrr, X86::ANDPDrm },
467 { X86::ANDPSrr, X86::ANDPSrm },
468 { X86::CMOVA16rr, X86::CMOVA16rm },
469 { X86::CMOVA32rr, X86::CMOVA32rm },
470 { X86::CMOVA64rr, X86::CMOVA64rm },
471 { X86::CMOVAE16rr, X86::CMOVAE16rm },
472 { X86::CMOVAE32rr, X86::CMOVAE32rm },
473 { X86::CMOVAE64rr, X86::CMOVAE64rm },
474 { X86::CMOVB16rr, X86::CMOVB16rm },
475 { X86::CMOVB32rr, X86::CMOVB32rm },
476 { X86::CMOVB64rr, X86::CMOVB64rm },
477 { X86::CMOVBE16rr, X86::CMOVBE16rm },
478 { X86::CMOVBE32rr, X86::CMOVBE32rm },
479 { X86::CMOVBE64rr, X86::CMOVBE64rm },
480 { X86::CMOVE16rr, X86::CMOVE16rm },
481 { X86::CMOVE32rr, X86::CMOVE32rm },
482 { X86::CMOVE64rr, X86::CMOVE64rm },
483 { X86::CMOVG16rr, X86::CMOVG16rm },
484 { X86::CMOVG32rr, X86::CMOVG32rm },
485 { X86::CMOVG64rr, X86::CMOVG64rm },
486 { X86::CMOVGE16rr, X86::CMOVGE16rm },
487 { X86::CMOVGE32rr, X86::CMOVGE32rm },
488 { X86::CMOVGE64rr, X86::CMOVGE64rm },
489 { X86::CMOVL16rr, X86::CMOVL16rm },
490 { X86::CMOVL32rr, X86::CMOVL32rm },
491 { X86::CMOVL64rr, X86::CMOVL64rm },
492 { X86::CMOVLE16rr, X86::CMOVLE16rm },
493 { X86::CMOVLE32rr, X86::CMOVLE32rm },
494 { X86::CMOVLE64rr, X86::CMOVLE64rm },
495 { X86::CMOVNE16rr, X86::CMOVNE16rm },
496 { X86::CMOVNE32rr, X86::CMOVNE32rm },
497 { X86::CMOVNE64rr, X86::CMOVNE64rm },
498 { X86::CMOVNO16rr, X86::CMOVNO16rm },
499 { X86::CMOVNO32rr, X86::CMOVNO32rm },
500 { X86::CMOVNO64rr, X86::CMOVNO64rm },
501 { X86::CMOVNP16rr, X86::CMOVNP16rm },
502 { X86::CMOVNP32rr, X86::CMOVNP32rm },
503 { X86::CMOVNP64rr, X86::CMOVNP64rm },
504 { X86::CMOVNS16rr, X86::CMOVNS16rm },
505 { X86::CMOVNS32rr, X86::CMOVNS32rm },
506 { X86::CMOVNS64rr, X86::CMOVNS64rm },
507 { X86::CMOVO16rr, X86::CMOVO16rm },
508 { X86::CMOVO32rr, X86::CMOVO32rm },
509 { X86::CMOVO64rr, X86::CMOVO64rm },
510 { X86::CMOVP16rr, X86::CMOVP16rm },
511 { X86::CMOVP32rr, X86::CMOVP32rm },
512 { X86::CMOVP64rr, X86::CMOVP64rm },
513 { X86::CMOVS16rr, X86::CMOVS16rm },
514 { X86::CMOVS32rr, X86::CMOVS32rm },
515 { X86::CMOVS64rr, X86::CMOVS64rm },
516 { X86::CMPPDrri, X86::CMPPDrmi },
517 { X86::CMPPSrri, X86::CMPPSrmi },
518 { X86::CMPSDrr, X86::CMPSDrm },
519 { X86::CMPSSrr, X86::CMPSSrm },
520 { X86::DIVPDrr, X86::DIVPDrm },
521 { X86::DIVPSrr, X86::DIVPSrm },
522 { X86::DIVSDrr, X86::DIVSDrm },
523 { X86::DIVSSrr, X86::DIVSSrm },
524 { X86::FsANDNPDrr, X86::FsANDNPDrm },
525 { X86::FsANDNPSrr, X86::FsANDNPSrm },
526 { X86::FsANDPDrr, X86::FsANDPDrm },
527 { X86::FsANDPSrr, X86::FsANDPSrm },
528 { X86::FsORPDrr, X86::FsORPDrm },
529 { X86::FsORPSrr, X86::FsORPSrm },
530 { X86::FsXORPDrr, X86::FsXORPDrm },
531 { X86::FsXORPSrr, X86::FsXORPSrm },
532 { X86::HADDPDrr, X86::HADDPDrm },
533 { X86::HADDPSrr, X86::HADDPSrm },
534 { X86::HSUBPDrr, X86::HSUBPDrm },
535 { X86::HSUBPSrr, X86::HSUBPSrm },
536 { X86::IMUL16rr, X86::IMUL16rm },
537 { X86::IMUL32rr, X86::IMUL32rm },
538 { X86::IMUL64rr, X86::IMUL64rm },
539 { X86::MAXPDrr, X86::MAXPDrm },
540 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
541 { X86::MAXPSrr, X86::MAXPSrm },
542 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
543 { X86::MAXSDrr, X86::MAXSDrm },
544 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
545 { X86::MAXSSrr, X86::MAXSSrm },
546 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
547 { X86::MINPDrr, X86::MINPDrm },
548 { X86::MINPDrr_Int, X86::MINPDrm_Int },
549 { X86::MINPSrr, X86::MINPSrm },
550 { X86::MINPSrr_Int, X86::MINPSrm_Int },
551 { X86::MINSDrr, X86::MINSDrm },
552 { X86::MINSDrr_Int, X86::MINSDrm_Int },
553 { X86::MINSSrr, X86::MINSSrm },
554 { X86::MINSSrr_Int, X86::MINSSrm_Int },
555 { X86::MULPDrr, X86::MULPDrm },
556 { X86::MULPSrr, X86::MULPSrm },
557 { X86::MULSDrr, X86::MULSDrm },
558 { X86::MULSSrr, X86::MULSSrm },
559 { X86::OR16rr, X86::OR16rm },
560 { X86::OR32rr, X86::OR32rm },
561 { X86::OR64rr, X86::OR64rm },
562 { X86::OR8rr, X86::OR8rm },
563 { X86::ORPDrr, X86::ORPDrm },
564 { X86::ORPSrr, X86::ORPSrm },
565 { X86::PACKSSDWrr, X86::PACKSSDWrm },
566 { X86::PACKSSWBrr, X86::PACKSSWBrm },
567 { X86::PACKUSWBrr, X86::PACKUSWBrm },
568 { X86::PADDBrr, X86::PADDBrm },
569 { X86::PADDDrr, X86::PADDDrm },
570 { X86::PADDQrr, X86::PADDQrm },
571 { X86::PADDSBrr, X86::PADDSBrm },
572 { X86::PADDSWrr, X86::PADDSWrm },
573 { X86::PADDWrr, X86::PADDWrm },
574 { X86::PANDNrr, X86::PANDNrm },
575 { X86::PANDrr, X86::PANDrm },
576 { X86::PAVGBrr, X86::PAVGBrm },
577 { X86::PAVGWrr, X86::PAVGWrm },
578 { X86::PCMPEQBrr, X86::PCMPEQBrm },
579 { X86::PCMPEQDrr, X86::PCMPEQDrm },
580 { X86::PCMPEQWrr, X86::PCMPEQWrm },
581 { X86::PCMPGTBrr, X86::PCMPGTBrm },
582 { X86::PCMPGTDrr, X86::PCMPGTDrm },
583 { X86::PCMPGTWrr, X86::PCMPGTWrm },
584 { X86::PINSRWrri, X86::PINSRWrmi },
585 { X86::PMADDWDrr, X86::PMADDWDrm },
586 { X86::PMAXSWrr, X86::PMAXSWrm },
587 { X86::PMAXUBrr, X86::PMAXUBrm },
588 { X86::PMINSWrr, X86::PMINSWrm },
589 { X86::PMINUBrr, X86::PMINUBrm },
590 { X86::PMULDQrr, X86::PMULDQrm },
591 { X86::PMULHUWrr, X86::PMULHUWrm },
592 { X86::PMULHWrr, X86::PMULHWrm },
593 { X86::PMULLDrr, X86::PMULLDrm },
594 { X86::PMULLDrr_int, X86::PMULLDrm_int },
595 { X86::PMULLWrr, X86::PMULLWrm },
596 { X86::PMULUDQrr, X86::PMULUDQrm },
597 { X86::PORrr, X86::PORrm },
598 { X86::PSADBWrr, X86::PSADBWrm },
599 { X86::PSLLDrr, X86::PSLLDrm },
600 { X86::PSLLQrr, X86::PSLLQrm },
601 { X86::PSLLWrr, X86::PSLLWrm },
602 { X86::PSRADrr, X86::PSRADrm },
603 { X86::PSRAWrr, X86::PSRAWrm },
604 { X86::PSRLDrr, X86::PSRLDrm },
605 { X86::PSRLQrr, X86::PSRLQrm },
606 { X86::PSRLWrr, X86::PSRLWrm },
607 { X86::PSUBBrr, X86::PSUBBrm },
608 { X86::PSUBDrr, X86::PSUBDrm },
609 { X86::PSUBSBrr, X86::PSUBSBrm },
610 { X86::PSUBSWrr, X86::PSUBSWrm },
611 { X86::PSUBWrr, X86::PSUBWrm },
612 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
613 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
614 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
615 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
616 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
617 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
618 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
619 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
620 { X86::PXORrr, X86::PXORrm },
621 { X86::SBB32rr, X86::SBB32rm },
622 { X86::SBB64rr, X86::SBB64rm },
623 { X86::SHUFPDrri, X86::SHUFPDrmi },
624 { X86::SHUFPSrri, X86::SHUFPSrmi },
625 { X86::SUB16rr, X86::SUB16rm },
626 { X86::SUB32rr, X86::SUB32rm },
627 { X86::SUB64rr, X86::SUB64rm },
628 { X86::SUB8rr, X86::SUB8rm },
629 { X86::SUBPDrr, X86::SUBPDrm },
630 { X86::SUBPSrr, X86::SUBPSrm },
631 { X86::SUBSDrr, X86::SUBSDrm },
632 { X86::SUBSSrr, X86::SUBSSrm },
633 // FIXME: TEST*rr -> swapped operand of TEST*mr.
634 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
635 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
636 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
637 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
638 { X86::XOR16rr, X86::XOR16rm },
639 { X86::XOR32rr, X86::XOR32rm },
640 { X86::XOR64rr, X86::XOR64rm },
641 { X86::XOR8rr, X86::XOR8rm },
642 { X86::XORPDrr, X86::XORPDrm },
643 { X86::XORPSrr, X86::XORPSrm }
646 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
647 unsigned RegOp = OpTbl2[i][0];
648 unsigned MemOp = OpTbl2[i][1];
649 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp,
651 assert(false && "Duplicated entries?");
652 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
653 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
654 std::make_pair(RegOp, AuxInfo))).second)
655 AmbEntries.push_back(MemOp);
658 // Remove ambiguous entries.
659 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
662 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
664 unsigned& destReg) const {
665 switch (MI.getOpcode()) {
672 case X86::MOV16to16_:
673 case X86::MOV32to32_:
677 // FP Stack register class copies
678 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
679 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
680 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
682 case X86::FsMOVAPSrr:
683 case X86::FsMOVAPDrr:
687 case X86::MOVSS2PSrr:
688 case X86::MOVSD2PDrr:
689 case X86::MOVPS2SSrr:
690 case X86::MOVPD2SDrr:
691 case X86::MMX_MOVD64rr:
692 case X86::MMX_MOVQ64rr:
693 assert(MI.getNumOperands() >= 2 &&
694 MI.getOperand(0).isReg() &&
695 MI.getOperand(1).isReg() &&
696 "invalid register-register move instruction");
697 sourceReg = MI.getOperand(1).getReg();
698 destReg = MI.getOperand(0).getReg();
703 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
704 int &FrameIndex) const {
705 switch (MI->getOpcode()) {
719 case X86::MMX_MOVD64rm:
720 case X86::MMX_MOVQ64rm:
721 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
722 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
723 MI->getOperand(2).getImm() == 1 &&
724 MI->getOperand(3).getReg() == 0 &&
725 MI->getOperand(4).getImm() == 0) {
726 FrameIndex = MI->getOperand(1).getIndex();
727 return MI->getOperand(0).getReg();
734 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
735 int &FrameIndex) const {
736 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(4).getReg();
767 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
769 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
770 bool isPICBase = false;
771 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
772 E = MRI.def_end(); I != E; ++I) {
773 MachineInstr *DefMI = I.getOperand().getParent();
774 if (DefMI->getOpcode() != X86::MOVPC32r)
776 assert(!isPICBase && "More than one PIC base?");
782 /// isGVStub - Return true if the GV requires an extra load to get the
784 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
785 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
789 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI) const {
790 switch (MI->getOpcode()) {
804 case X86::MMX_MOVD64rm:
805 case X86::MMX_MOVQ64rm: {
806 // Loads from constant pools are trivially rematerializable.
807 if (MI->getOperand(1).isReg() &&
808 MI->getOperand(2).isImm() &&
809 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
810 (MI->getOperand(4).isCPI() ||
811 (MI->getOperand(4).isGlobal() &&
812 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
813 unsigned BaseReg = MI->getOperand(1).getReg();
816 // Allow re-materialization of PIC load.
817 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
819 const MachineFunction &MF = *MI->getParent()->getParent();
820 const MachineRegisterInfo &MRI = MF.getRegInfo();
821 bool isPICBase = false;
822 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
823 E = MRI.def_end(); I != E; ++I) {
824 MachineInstr *DefMI = I.getOperand().getParent();
825 if (DefMI->getOpcode() != X86::MOVPC32r)
827 assert(!isPICBase && "More than one PIC base?");
837 if (MI->getOperand(2).isImm() &&
838 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
839 !MI->getOperand(4).isReg()) {
840 // lea fi#, lea GV, etc. are all rematerializable.
841 if (!MI->getOperand(1).isReg())
843 unsigned BaseReg = MI->getOperand(1).getReg();
846 // Allow re-materialization of lea PICBase + x.
847 const MachineFunction &MF = *MI->getParent()->getParent();
848 const MachineRegisterInfo &MRI = MF.getRegInfo();
849 return regIsPICBase(BaseReg, MRI);
855 // All other instructions marked M_REMATERIALIZABLE are always trivially
860 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
861 /// would clobber the EFLAGS condition register. Note the result may be
862 /// conservative. If it cannot definitely determine the safety after visiting
863 /// two instructions it assumes it's not safe.
864 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
865 MachineBasicBlock::iterator I) {
866 // It's always safe to clobber EFLAGS at the end of a block.
870 // For compile time consideration, if we are not able to determine the
871 // safety after visiting 2 instructions, we will assume it's not safe.
872 for (unsigned i = 0; i < 2; ++i) {
873 bool SeenDef = false;
874 for (unsigned j = 0, e = I->getNumOperands(); j != e; ++j) {
875 MachineOperand &MO = I->getOperand(j);
878 if (MO.getReg() == X86::EFLAGS) {
886 // This instruction defines EFLAGS, no need to look any further.
890 // If we make it to the end of the block, it's safe to clobber EFLAGS.
895 // Conservative answer.
899 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
900 MachineBasicBlock::iterator I,
902 const MachineInstr *Orig) const {
903 unsigned SubIdx = Orig->getOperand(0).isReg()
904 ? Orig->getOperand(0).getSubReg() : 0;
905 bool ChangeSubIdx = SubIdx != 0;
906 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
907 DestReg = RI.getSubReg(DestReg, SubIdx);
911 // MOV32r0 etc. are implemented with xor which clobbers condition code.
912 // Re-materialize them as movri instructions to avoid side effects.
913 bool Emitted = false;
914 switch (Orig->getOpcode()) {
920 if (!isSafeToClobberEFLAGS(MBB, I)) {
922 switch (Orig->getOpcode()) {
924 case X86::MOV8r0: Opc = X86::MOV8ri; break;
925 case X86::MOV16r0: Opc = X86::MOV16ri; break;
926 case X86::MOV32r0: Opc = X86::MOV32ri; break;
927 case X86::MOV64r0: Opc = X86::MOV64ri32; break;
929 BuildMI(MBB, I, get(Opc), DestReg).addImm(0);
937 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
938 MI->getOperand(0).setReg(DestReg);
943 MachineInstr *NewMI = prior(I);
944 NewMI->getOperand(0).setSubReg(SubIdx);
948 /// isInvariantLoad - Return true if the specified instruction (which is marked
949 /// mayLoad) is loading from a location whose value is invariant across the
950 /// function. For example, loading a value from the constant pool or from
951 /// from the argument area of a function if it does not change. This should
952 /// only return true of *all* loads the instruction does are invariant (if it
953 /// does multiple loads).
954 bool X86InstrInfo::isInvariantLoad(const MachineInstr *MI) const {
955 // This code cares about loads from three cases: constant pool entries,
956 // invariant argument slots, and global stubs. In order to handle these cases
957 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
958 // operand and base our analysis on it. This is safe because the address of
959 // none of these three cases is ever used as anything other than a load base
960 // and X86 doesn't have any instructions that load from multiple places.
962 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
963 const MachineOperand &MO = MI->getOperand(i);
964 // Loads from constant pools are trivially invariant.
969 return isGVStub(MO.getGlobal(), TM);
971 // If this is a load from an invariant stack slot, the load is a constant.
973 const MachineFrameInfo &MFI =
974 *MI->getParent()->getParent()->getFrameInfo();
975 int Idx = MO.getIndex();
976 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
980 // All other instances of these instructions are presumed to have other
985 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
986 /// is not marked dead.
987 static bool hasLiveCondCodeDef(MachineInstr *MI) {
988 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
989 MachineOperand &MO = MI->getOperand(i);
990 if (MO.isReg() && MO.isDef() &&
991 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
998 /// convertToThreeAddress - This method must be implemented by targets that
999 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1000 /// may be able to convert a two-address instruction into a true
1001 /// three-address instruction on demand. This allows the X86 target (for
1002 /// example) to convert ADD and SHL instructions into LEA instructions if they
1003 /// would require register copies due to two-addressness.
1005 /// This method returns a null pointer if the transformation cannot be
1006 /// performed, otherwise it returns the new instruction.
1009 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1010 MachineBasicBlock::iterator &MBBI,
1011 LiveVariables *LV) const {
1012 MachineInstr *MI = MBBI;
1013 MachineFunction &MF = *MI->getParent()->getParent();
1014 // All instructions input are two-addr instructions. Get the known operands.
1015 unsigned Dest = MI->getOperand(0).getReg();
1016 unsigned Src = MI->getOperand(1).getReg();
1017 bool isDead = MI->getOperand(0).isDead();
1018 bool isKill = MI->getOperand(1).isKill();
1020 MachineInstr *NewMI = NULL;
1021 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1022 // we have better subtarget support, enable the 16-bit LEA generation here.
1023 bool DisableLEA16 = true;
1025 unsigned MIOpc = MI->getOpcode();
1027 case X86::SHUFPSrri: {
1028 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1029 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1031 unsigned B = MI->getOperand(1).getReg();
1032 unsigned C = MI->getOperand(2).getReg();
1033 if (B != C) return 0;
1034 unsigned A = MI->getOperand(0).getReg();
1035 unsigned M = MI->getOperand(3).getImm();
1036 NewMI = BuildMI(MF, get(X86::PSHUFDri)).addReg(A, true, false, false, isDead)
1037 .addReg(B, false, false, isKill).addImm(M);
1040 case X86::SHL64ri: {
1041 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1042 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1043 // the flags produced by a shift yet, so this is safe.
1044 unsigned ShAmt = MI->getOperand(2).getImm();
1045 if (ShAmt == 0 || ShAmt >= 4) return 0;
1047 NewMI = BuildMI(MF, get(X86::LEA64r)).addReg(Dest, true, false, false, isDead)
1048 .addReg(0).addImm(1 << ShAmt).addReg(Src, false, false, isKill).addImm(0);
1051 case X86::SHL32ri: {
1052 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1053 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1054 // the flags produced by a shift yet, so this is safe.
1055 unsigned ShAmt = MI->getOperand(2).getImm();
1056 if (ShAmt == 0 || ShAmt >= 4) return 0;
1058 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1059 X86::LEA64_32r : X86::LEA32r;
1060 NewMI = BuildMI(MF, get(Opc)).addReg(Dest, true, false, false, isDead)
1061 .addReg(0).addImm(1 << ShAmt)
1062 .addReg(Src, false, false, isKill).addImm(0);
1065 case X86::SHL16ri: {
1066 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1067 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1068 // the flags produced by a shift yet, so this is safe.
1069 unsigned ShAmt = MI->getOperand(2).getImm();
1070 if (ShAmt == 0 || ShAmt >= 4) return 0;
1073 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1074 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1075 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1076 ? X86::LEA64_32r : X86::LEA32r;
1077 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1078 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1080 // Build and insert into an implicit UNDEF value. This is OK because
1081 // well be shifting and then extracting the lower 16-bits.
1082 BuildMI(*MFI, MBBI, get(X86::IMPLICIT_DEF), leaInReg);
1083 MachineInstr *InsMI = BuildMI(*MFI, MBBI, get(X86::INSERT_SUBREG),leaInReg)
1084 .addReg(leaInReg).addReg(Src, false, false, isKill)
1085 .addImm(X86::SUBREG_16BIT);
1087 NewMI = BuildMI(*MFI, MBBI, get(Opc), leaOutReg).addReg(0).addImm(1 << ShAmt)
1088 .addReg(leaInReg, false, false, true).addImm(0);
1090 MachineInstr *ExtMI = BuildMI(*MFI, MBBI, get(X86::EXTRACT_SUBREG))
1091 .addReg(Dest, true, false, false, isDead)
1092 .addReg(leaOutReg, false, false, true).addImm(X86::SUBREG_16BIT);
1094 // Update live variables
1095 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1096 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1098 LV->replaceKillInstruction(Src, MI, InsMI);
1100 LV->replaceKillInstruction(Dest, MI, ExtMI);
1104 NewMI = BuildMI(MF, get(X86::LEA16r)).addReg(Dest, true, false, false, isDead)
1105 .addReg(0).addImm(1 << ShAmt)
1106 .addReg(Src, false, false, isKill).addImm(0);
1111 // The following opcodes also sets the condition code register(s). Only
1112 // convert them to equivalent lea if the condition code register def's
1114 if (hasLiveCondCodeDef(MI))
1117 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1122 case X86::INC64_32r: {
1123 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1124 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1125 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1126 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1127 .addReg(Dest, true, false, false, isDead),
1132 case X86::INC64_16r:
1133 if (DisableLEA16) return 0;
1134 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1135 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1136 .addReg(Dest, true, false, false, isDead),
1141 case X86::DEC64_32r: {
1142 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1143 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1144 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1145 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1146 .addReg(Dest, true, false, false, isDead),
1151 case X86::DEC64_16r:
1152 if (DisableLEA16) return 0;
1153 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1154 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1155 .addReg(Dest, true, false, false, isDead),
1159 case X86::ADD32rr: {
1160 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1161 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1162 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1163 unsigned Src2 = MI->getOperand(2).getReg();
1164 bool isKill2 = MI->getOperand(2).isKill();
1165 NewMI = addRegReg(BuildMI(MF, get(Opc))
1166 .addReg(Dest, true, false, false, isDead),
1167 Src, isKill, Src2, isKill2);
1169 LV->replaceKillInstruction(Src2, MI, NewMI);
1172 case X86::ADD16rr: {
1173 if (DisableLEA16) return 0;
1174 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1175 unsigned Src2 = MI->getOperand(2).getReg();
1176 bool isKill2 = MI->getOperand(2).isKill();
1177 NewMI = addRegReg(BuildMI(MF, get(X86::LEA16r))
1178 .addReg(Dest, true, false, false, isDead),
1179 Src, isKill, Src2, isKill2);
1181 LV->replaceKillInstruction(Src2, MI, NewMI);
1184 case X86::ADD64ri32:
1186 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1187 if (MI->getOperand(2).isImm())
1188 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA64r))
1189 .addReg(Dest, true, false, false, isDead),
1190 Src, isKill, MI->getOperand(2).getImm());
1194 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1195 if (MI->getOperand(2).isImm()) {
1196 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1197 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1198 .addReg(Dest, true, false, false, isDead),
1199 Src, isKill, MI->getOperand(2).getImm());
1204 if (DisableLEA16) return 0;
1205 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1206 if (MI->getOperand(2).isImm())
1207 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1208 .addReg(Dest, true, false, false, isDead),
1209 Src, isKill, MI->getOperand(2).getImm());
1212 if (DisableLEA16) return 0;
1214 case X86::SHL64ri: {
1215 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImm() &&
1216 "Unknown shl instruction!");
1217 unsigned ShAmt = MI->getOperand(2).getImm();
1218 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1220 AM.Scale = 1 << ShAmt;
1222 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1223 : (MIOpc == X86::SHL32ri
1224 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1225 NewMI = addFullAddress(BuildMI(MF, get(Opc))
1226 .addReg(Dest, true, false, false, isDead), AM);
1228 NewMI->getOperand(3).setIsKill(true);
1236 if (!NewMI) return 0;
1238 if (LV) { // Update live variables
1240 LV->replaceKillInstruction(Src, MI, NewMI);
1242 LV->replaceKillInstruction(Dest, MI, NewMI);
1245 MFI->insert(MBBI, NewMI); // Insert the new inst
1249 /// commuteInstruction - We have a few instructions that must be hacked on to
1253 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1254 switch (MI->getOpcode()) {
1255 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1256 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1257 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1258 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1259 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1260 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1263 switch (MI->getOpcode()) {
1264 default: assert(0 && "Unreachable!");
1265 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1266 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1267 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1268 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1269 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1270 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1272 unsigned Amt = MI->getOperand(3).getImm();
1274 MachineFunction &MF = *MI->getParent()->getParent();
1275 MI = MF.CloneMachineInstr(MI);
1278 MI->setDesc(get(Opc));
1279 MI->getOperand(3).setImm(Size-Amt);
1280 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1282 case X86::CMOVB16rr:
1283 case X86::CMOVB32rr:
1284 case X86::CMOVB64rr:
1285 case X86::CMOVAE16rr:
1286 case X86::CMOVAE32rr:
1287 case X86::CMOVAE64rr:
1288 case X86::CMOVE16rr:
1289 case X86::CMOVE32rr:
1290 case X86::CMOVE64rr:
1291 case X86::CMOVNE16rr:
1292 case X86::CMOVNE32rr:
1293 case X86::CMOVNE64rr:
1294 case X86::CMOVBE16rr:
1295 case X86::CMOVBE32rr:
1296 case X86::CMOVBE64rr:
1297 case X86::CMOVA16rr:
1298 case X86::CMOVA32rr:
1299 case X86::CMOVA64rr:
1300 case X86::CMOVL16rr:
1301 case X86::CMOVL32rr:
1302 case X86::CMOVL64rr:
1303 case X86::CMOVGE16rr:
1304 case X86::CMOVGE32rr:
1305 case X86::CMOVGE64rr:
1306 case X86::CMOVLE16rr:
1307 case X86::CMOVLE32rr:
1308 case X86::CMOVLE64rr:
1309 case X86::CMOVG16rr:
1310 case X86::CMOVG32rr:
1311 case X86::CMOVG64rr:
1312 case X86::CMOVS16rr:
1313 case X86::CMOVS32rr:
1314 case X86::CMOVS64rr:
1315 case X86::CMOVNS16rr:
1316 case X86::CMOVNS32rr:
1317 case X86::CMOVNS64rr:
1318 case X86::CMOVP16rr:
1319 case X86::CMOVP32rr:
1320 case X86::CMOVP64rr:
1321 case X86::CMOVNP16rr:
1322 case X86::CMOVNP32rr:
1323 case X86::CMOVNP64rr:
1324 case X86::CMOVO16rr:
1325 case X86::CMOVO32rr:
1326 case X86::CMOVO64rr:
1327 case X86::CMOVNO16rr:
1328 case X86::CMOVNO32rr:
1329 case X86::CMOVNO64rr: {
1331 switch (MI->getOpcode()) {
1333 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1334 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1335 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1336 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1337 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1338 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1339 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1340 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1341 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1342 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1343 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1344 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1345 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1346 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1347 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1348 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1349 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1350 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1351 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1352 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1353 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1354 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1355 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1356 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1357 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1358 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1359 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1360 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1361 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1362 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1363 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1364 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1365 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1366 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1367 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1368 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1369 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1370 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1371 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1372 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1373 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1374 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1375 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1376 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1377 case X86::CMOVO64rr: Opc = X86::CMOVNO32rr; break;
1378 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1379 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1380 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1383 MachineFunction &MF = *MI->getParent()->getParent();
1384 MI = MF.CloneMachineInstr(MI);
1387 MI->setDesc(get(Opc));
1388 // Fallthrough intended.
1391 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1395 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1397 default: return X86::COND_INVALID;
1398 case X86::JE: return X86::COND_E;
1399 case X86::JNE: return X86::COND_NE;
1400 case X86::JL: return X86::COND_L;
1401 case X86::JLE: return X86::COND_LE;
1402 case X86::JG: return X86::COND_G;
1403 case X86::JGE: return X86::COND_GE;
1404 case X86::JB: return X86::COND_B;
1405 case X86::JBE: return X86::COND_BE;
1406 case X86::JA: return X86::COND_A;
1407 case X86::JAE: return X86::COND_AE;
1408 case X86::JS: return X86::COND_S;
1409 case X86::JNS: return X86::COND_NS;
1410 case X86::JP: return X86::COND_P;
1411 case X86::JNP: return X86::COND_NP;
1412 case X86::JO: return X86::COND_O;
1413 case X86::JNO: return X86::COND_NO;
1417 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1419 default: assert(0 && "Illegal condition code!");
1420 case X86::COND_E: return X86::JE;
1421 case X86::COND_NE: return X86::JNE;
1422 case X86::COND_L: return X86::JL;
1423 case X86::COND_LE: return X86::JLE;
1424 case X86::COND_G: return X86::JG;
1425 case X86::COND_GE: return X86::JGE;
1426 case X86::COND_B: return X86::JB;
1427 case X86::COND_BE: return X86::JBE;
1428 case X86::COND_A: return X86::JA;
1429 case X86::COND_AE: return X86::JAE;
1430 case X86::COND_S: return X86::JS;
1431 case X86::COND_NS: return X86::JNS;
1432 case X86::COND_P: return X86::JP;
1433 case X86::COND_NP: return X86::JNP;
1434 case X86::COND_O: return X86::JO;
1435 case X86::COND_NO: return X86::JNO;
1439 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1440 /// e.g. turning COND_E to COND_NE.
1441 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1443 default: assert(0 && "Illegal condition code!");
1444 case X86::COND_E: return X86::COND_NE;
1445 case X86::COND_NE: return X86::COND_E;
1446 case X86::COND_L: return X86::COND_GE;
1447 case X86::COND_LE: return X86::COND_G;
1448 case X86::COND_G: return X86::COND_LE;
1449 case X86::COND_GE: return X86::COND_L;
1450 case X86::COND_B: return X86::COND_AE;
1451 case X86::COND_BE: return X86::COND_A;
1452 case X86::COND_A: return X86::COND_BE;
1453 case X86::COND_AE: return X86::COND_B;
1454 case X86::COND_S: return X86::COND_NS;
1455 case X86::COND_NS: return X86::COND_S;
1456 case X86::COND_P: return X86::COND_NP;
1457 case X86::COND_NP: return X86::COND_P;
1458 case X86::COND_O: return X86::COND_NO;
1459 case X86::COND_NO: return X86::COND_O;
1463 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1464 const TargetInstrDesc &TID = MI->getDesc();
1465 if (!TID.isTerminator()) return false;
1467 // Conditional branch is a special case.
1468 if (TID.isBranch() && !TID.isBarrier())
1470 if (!TID.isPredicable())
1472 return !isPredicated(MI);
1475 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1476 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1477 const X86InstrInfo &TII) {
1478 if (MI->getOpcode() == X86::FP_REG_KILL)
1480 return TII.isUnpredicatedTerminator(MI);
1483 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1484 MachineBasicBlock *&TBB,
1485 MachineBasicBlock *&FBB,
1486 SmallVectorImpl<MachineOperand> &Cond) const {
1487 // Start from the bottom of the block and work up, examining the
1488 // terminator instructions.
1489 MachineBasicBlock::iterator I = MBB.end();
1490 while (I != MBB.begin()) {
1492 // Working from the bottom, when we see a non-terminator
1493 // instruction, we're done.
1494 if (!isBrAnalysisUnpredicatedTerminator(I, *this))
1496 // A terminator that isn't a branch can't easily be handled
1497 // by this analysis.
1498 if (!I->getDesc().isBranch())
1500 // Handle unconditional branches.
1501 if (I->getOpcode() == X86::JMP) {
1502 // If the block has any instructions after a JMP, delete them.
1503 while (next(I) != MBB.end())
1504 next(I)->eraseFromParent();
1507 // Delete the JMP if it's equivalent to a fall-through.
1508 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1510 I->eraseFromParent();
1514 // TBB is used to indicate the unconditinal destination.
1515 TBB = I->getOperand(0).getMBB();
1518 // Handle conditional branches.
1519 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1520 if (BranchCode == X86::COND_INVALID)
1521 return true; // Can't handle indirect branch.
1522 // Working from the bottom, handle the first conditional branch.
1525 TBB = I->getOperand(0).getMBB();
1526 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1529 // Handle subsequent conditional branches. Only handle the case
1530 // where all conditional branches branch to the same destination
1531 // and their condition opcodes fit one of the special
1532 // multi-branch idioms.
1533 assert(Cond.size() == 1);
1535 // Only handle the case where all conditional branches branch to
1536 // the same destination.
1537 if (TBB != I->getOperand(0).getMBB())
1539 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1540 // If the conditions are the same, we can leave them alone.
1541 if (OldBranchCode == BranchCode)
1543 // If they differ, see if they fit one of the known patterns.
1544 // Theoretically we could handle more patterns here, but
1545 // we shouldn't expect to see them if instruction selection
1546 // has done a reasonable job.
1547 if ((OldBranchCode == X86::COND_NP &&
1548 BranchCode == X86::COND_E) ||
1549 (OldBranchCode == X86::COND_E &&
1550 BranchCode == X86::COND_NP))
1551 BranchCode = X86::COND_NP_OR_E;
1552 else if ((OldBranchCode == X86::COND_P &&
1553 BranchCode == X86::COND_NE) ||
1554 (OldBranchCode == X86::COND_NE &&
1555 BranchCode == X86::COND_P))
1556 BranchCode = X86::COND_NE_OR_P;
1559 // Update the MachineOperand.
1560 Cond[0].setImm(BranchCode);
1566 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1567 MachineBasicBlock::iterator I = MBB.end();
1570 while (I != MBB.begin()) {
1572 if (I->getOpcode() != X86::JMP &&
1573 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1575 // Remove the branch.
1576 I->eraseFromParent();
1584 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
1585 const MachineOperand &MO) {
1587 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
1588 MO.isKill(), MO.isDead(), MO.getSubReg());
1589 else if (MO.isImm())
1590 MIB = MIB.addImm(MO.getImm());
1592 MIB = MIB.addFrameIndex(MO.getIndex());
1593 else if (MO.isGlobal())
1594 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
1595 else if (MO.isCPI())
1596 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
1597 else if (MO.isJTI())
1598 MIB = MIB.addJumpTableIndex(MO.getIndex());
1599 else if (MO.isSymbol())
1600 MIB = MIB.addExternalSymbol(MO.getSymbolName());
1602 assert(0 && "Unknown operand for X86InstrAddOperand!");
1608 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1609 MachineBasicBlock *FBB,
1610 const SmallVectorImpl<MachineOperand> &Cond) const {
1611 // Shouldn't be a fall through.
1612 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1613 assert((Cond.size() == 1 || Cond.size() == 0) &&
1614 "X86 branch conditions have one component!");
1617 // Unconditional branch?
1618 assert(!FBB && "Unconditional branch with multiple successors!");
1619 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
1623 // Conditional branch.
1625 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1627 case X86::COND_NP_OR_E:
1628 // Synthesize NP_OR_E with two branches.
1629 BuildMI(&MBB, get(X86::JNP)).addMBB(TBB);
1631 BuildMI(&MBB, get(X86::JE)).addMBB(TBB);
1634 case X86::COND_NE_OR_P:
1635 // Synthesize NE_OR_P with two branches.
1636 BuildMI(&MBB, get(X86::JNE)).addMBB(TBB);
1638 BuildMI(&MBB, get(X86::JP)).addMBB(TBB);
1642 unsigned Opc = GetCondBranchFromCond(CC);
1643 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1648 // Two-way Conditional branch. Insert the second branch.
1649 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
1655 bool X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1656 MachineBasicBlock::iterator MI,
1657 unsigned DestReg, unsigned SrcReg,
1658 const TargetRegisterClass *DestRC,
1659 const TargetRegisterClass *SrcRC) const {
1660 if (DestRC == SrcRC) {
1662 if (DestRC == &X86::GR64RegClass) {
1664 } else if (DestRC == &X86::GR32RegClass) {
1666 } else if (DestRC == &X86::GR16RegClass) {
1668 } else if (DestRC == &X86::GR8RegClass) {
1670 } else if (DestRC == &X86::GR32_RegClass) {
1671 Opc = X86::MOV32_rr;
1672 } else if (DestRC == &X86::GR16_RegClass) {
1673 Opc = X86::MOV16_rr;
1674 } else if (DestRC == &X86::RFP32RegClass) {
1675 Opc = X86::MOV_Fp3232;
1676 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1677 Opc = X86::MOV_Fp6464;
1678 } else if (DestRC == &X86::RFP80RegClass) {
1679 Opc = X86::MOV_Fp8080;
1680 } else if (DestRC == &X86::FR32RegClass) {
1681 Opc = X86::FsMOVAPSrr;
1682 } else if (DestRC == &X86::FR64RegClass) {
1683 Opc = X86::FsMOVAPDrr;
1684 } else if (DestRC == &X86::VR128RegClass) {
1685 Opc = X86::MOVAPSrr;
1686 } else if (DestRC == &X86::VR64RegClass) {
1687 Opc = X86::MMX_MOVQ64rr;
1691 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
1695 // Moving EFLAGS to / from another register requires a push and a pop.
1696 if (SrcRC == &X86::CCRRegClass) {
1697 if (SrcReg != X86::EFLAGS)
1699 if (DestRC == &X86::GR64RegClass) {
1700 BuildMI(MBB, MI, get(X86::PUSHFQ));
1701 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
1703 } else if (DestRC == &X86::GR32RegClass) {
1704 BuildMI(MBB, MI, get(X86::PUSHFD));
1705 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
1708 } else if (DestRC == &X86::CCRRegClass) {
1709 if (DestReg != X86::EFLAGS)
1711 if (SrcRC == &X86::GR64RegClass) {
1712 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
1713 BuildMI(MBB, MI, get(X86::POPFQ));
1715 } else if (SrcRC == &X86::GR32RegClass) {
1716 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
1717 BuildMI(MBB, MI, get(X86::POPFD));
1722 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1723 if (SrcRC == &X86::RSTRegClass) {
1724 // Copying from ST(0)/ST(1).
1725 if (SrcReg != X86::ST0 && SrcReg != X86::ST1)
1726 // Can only copy from ST(0)/ST(1) right now
1728 bool isST0 = SrcReg == X86::ST0;
1730 if (DestRC == &X86::RFP32RegClass)
1731 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1732 else if (DestRC == &X86::RFP64RegClass)
1733 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1735 if (DestRC != &X86::RFP80RegClass)
1737 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1739 BuildMI(MBB, MI, get(Opc), DestReg);
1743 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1744 if (DestRC == &X86::RSTRegClass) {
1745 // Copying to ST(0). FIXME: handle ST(1) also
1746 if (DestReg != X86::ST0)
1747 // Can only copy to TOS right now
1750 if (SrcRC == &X86::RFP32RegClass)
1751 Opc = X86::FpSET_ST0_32;
1752 else if (SrcRC == &X86::RFP64RegClass)
1753 Opc = X86::FpSET_ST0_64;
1755 if (SrcRC != &X86::RFP80RegClass)
1757 Opc = X86::FpSET_ST0_80;
1759 BuildMI(MBB, MI, get(Opc)).addReg(SrcReg);
1763 // Not yet supported!
1767 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1768 bool isStackAligned) {
1770 if (RC == &X86::GR64RegClass) {
1772 } else if (RC == &X86::GR32RegClass) {
1774 } else if (RC == &X86::GR16RegClass) {
1776 } else if (RC == &X86::GR8RegClass) {
1778 } else if (RC == &X86::GR32_RegClass) {
1779 Opc = X86::MOV32_mr;
1780 } else if (RC == &X86::GR16_RegClass) {
1781 Opc = X86::MOV16_mr;
1782 } else if (RC == &X86::RFP80RegClass) {
1783 Opc = X86::ST_FpP80m; // pops
1784 } else if (RC == &X86::RFP64RegClass) {
1785 Opc = X86::ST_Fp64m;
1786 } else if (RC == &X86::RFP32RegClass) {
1787 Opc = X86::ST_Fp32m;
1788 } else if (RC == &X86::FR32RegClass) {
1790 } else if (RC == &X86::FR64RegClass) {
1792 } else if (RC == &X86::VR128RegClass) {
1793 // If stack is realigned we can use aligned stores.
1794 Opc = isStackAligned ? X86::MOVAPSmr : X86::MOVUPSmr;
1795 } else if (RC == &X86::VR64RegClass) {
1796 Opc = X86::MMX_MOVQ64mr;
1798 assert(0 && "Unknown regclass");
1805 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1806 MachineBasicBlock::iterator MI,
1807 unsigned SrcReg, bool isKill, int FrameIdx,
1808 const TargetRegisterClass *RC) const {
1809 const MachineFunction &MF = *MBB.getParent();
1810 bool isAligned = (RI.getStackAlignment() >= 16) ||
1811 RI.needsStackRealignment(MF);
1812 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1813 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
1814 .addReg(SrcReg, false, false, isKill);
1817 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1819 SmallVectorImpl<MachineOperand> &Addr,
1820 const TargetRegisterClass *RC,
1821 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1822 bool isAligned = (RI.getStackAlignment() >= 16) ||
1823 RI.needsStackRealignment(MF);
1824 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1825 MachineInstrBuilder MIB = BuildMI(MF, get(Opc));
1826 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1827 MIB = X86InstrAddOperand(MIB, Addr[i]);
1828 MIB.addReg(SrcReg, false, false, isKill);
1829 NewMIs.push_back(MIB);
1832 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1833 bool isStackAligned) {
1835 if (RC == &X86::GR64RegClass) {
1837 } else if (RC == &X86::GR32RegClass) {
1839 } else if (RC == &X86::GR16RegClass) {
1841 } else if (RC == &X86::GR8RegClass) {
1843 } else if (RC == &X86::GR32_RegClass) {
1844 Opc = X86::MOV32_rm;
1845 } else if (RC == &X86::GR16_RegClass) {
1846 Opc = X86::MOV16_rm;
1847 } else if (RC == &X86::RFP80RegClass) {
1848 Opc = X86::LD_Fp80m;
1849 } else if (RC == &X86::RFP64RegClass) {
1850 Opc = X86::LD_Fp64m;
1851 } else if (RC == &X86::RFP32RegClass) {
1852 Opc = X86::LD_Fp32m;
1853 } else if (RC == &X86::FR32RegClass) {
1855 } else if (RC == &X86::FR64RegClass) {
1857 } else if (RC == &X86::VR128RegClass) {
1858 // If stack is realigned we can use aligned loads.
1859 Opc = isStackAligned ? X86::MOVAPSrm : X86::MOVUPSrm;
1860 } else if (RC == &X86::VR64RegClass) {
1861 Opc = X86::MMX_MOVQ64rm;
1863 assert(0 && "Unknown regclass");
1870 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1871 MachineBasicBlock::iterator MI,
1872 unsigned DestReg, int FrameIdx,
1873 const TargetRegisterClass *RC) const{
1874 const MachineFunction &MF = *MBB.getParent();
1875 bool isAligned = (RI.getStackAlignment() >= 16) ||
1876 RI.needsStackRealignment(MF);
1877 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1878 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
1881 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1882 SmallVectorImpl<MachineOperand> &Addr,
1883 const TargetRegisterClass *RC,
1884 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1885 bool isAligned = (RI.getStackAlignment() >= 16) ||
1886 RI.needsStackRealignment(MF);
1887 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1888 MachineInstrBuilder MIB = BuildMI(MF, get(Opc), DestReg);
1889 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1890 MIB = X86InstrAddOperand(MIB, Addr[i]);
1891 NewMIs.push_back(MIB);
1894 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1895 MachineBasicBlock::iterator MI,
1896 const std::vector<CalleeSavedInfo> &CSI) const {
1900 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1901 unsigned SlotSize = is64Bit ? 8 : 4;
1903 MachineFunction &MF = *MBB.getParent();
1904 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1905 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1907 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1908 for (unsigned i = CSI.size(); i != 0; --i) {
1909 unsigned Reg = CSI[i-1].getReg();
1910 // Add the callee-saved register as live-in. It's killed at the spill.
1912 BuildMI(MBB, MI, get(Opc))
1913 .addReg(Reg, /*isDef=*/false, /*isImp=*/false, /*isKill=*/true);
1918 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1919 MachineBasicBlock::iterator MI,
1920 const std::vector<CalleeSavedInfo> &CSI) const {
1924 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1926 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1927 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1928 unsigned Reg = CSI[i].getReg();
1929 BuildMI(MBB, MI, get(Opc), Reg);
1934 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
1935 const SmallVectorImpl<MachineOperand> &MOs,
1936 MachineInstr *MI, const TargetInstrInfo &TII) {
1937 // Create the base instruction with the memory operand as the first part.
1938 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), true);
1939 MachineInstrBuilder MIB(NewMI);
1940 unsigned NumAddrOps = MOs.size();
1941 for (unsigned i = 0; i != NumAddrOps; ++i)
1942 MIB = X86InstrAddOperand(MIB, MOs[i]);
1943 if (NumAddrOps < 4) // FrameIndex only
1944 MIB.addImm(1).addReg(0).addImm(0);
1946 // Loop over the rest of the ri operands, converting them over.
1947 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1948 for (unsigned i = 0; i != NumOps; ++i) {
1949 MachineOperand &MO = MI->getOperand(i+2);
1950 MIB = X86InstrAddOperand(MIB, MO);
1952 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1953 MachineOperand &MO = MI->getOperand(i);
1954 MIB = X86InstrAddOperand(MIB, MO);
1959 static MachineInstr *FuseInst(MachineFunction &MF,
1960 unsigned Opcode, unsigned OpNo,
1961 const SmallVectorImpl<MachineOperand> &MOs,
1962 MachineInstr *MI, const TargetInstrInfo &TII) {
1963 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), true);
1964 MachineInstrBuilder MIB(NewMI);
1966 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1967 MachineOperand &MO = MI->getOperand(i);
1969 assert(MO.isReg() && "Expected to fold into reg operand!");
1970 unsigned NumAddrOps = MOs.size();
1971 for (unsigned i = 0; i != NumAddrOps; ++i)
1972 MIB = X86InstrAddOperand(MIB, MOs[i]);
1973 if (NumAddrOps < 4) // FrameIndex only
1974 MIB.addImm(1).addReg(0).addImm(0);
1976 MIB = X86InstrAddOperand(MIB, MO);
1982 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
1983 const SmallVectorImpl<MachineOperand> &MOs,
1985 MachineFunction &MF = *MI->getParent()->getParent();
1986 MachineInstrBuilder MIB = BuildMI(MF, TII.get(Opcode));
1988 unsigned NumAddrOps = MOs.size();
1989 for (unsigned i = 0; i != NumAddrOps; ++i)
1990 MIB = X86InstrAddOperand(MIB, MOs[i]);
1991 if (NumAddrOps < 4) // FrameIndex only
1992 MIB.addImm(1).addReg(0).addImm(0);
1993 return MIB.addImm(0);
1997 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
1998 MachineInstr *MI, unsigned i,
1999 const SmallVectorImpl<MachineOperand> &MOs) const{
2000 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2001 bool isTwoAddrFold = false;
2002 unsigned NumOps = MI->getDesc().getNumOperands();
2003 bool isTwoAddr = NumOps > 1 &&
2004 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2006 MachineInstr *NewMI = NULL;
2007 // Folding a memory location into the two-address part of a two-address
2008 // instruction is different than folding it other places. It requires
2009 // replacing the *two* registers with the memory location.
2010 if (isTwoAddr && NumOps >= 2 && i < 2 &&
2011 MI->getOperand(0).isReg() &&
2012 MI->getOperand(1).isReg() &&
2013 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
2014 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2015 isTwoAddrFold = true;
2016 } else if (i == 0) { // If operand 0
2017 if (MI->getOpcode() == X86::MOV16r0)
2018 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
2019 else if (MI->getOpcode() == X86::MOV32r0)
2020 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
2021 else if (MI->getOpcode() == X86::MOV64r0)
2022 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2023 else if (MI->getOpcode() == X86::MOV8r0)
2024 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2028 OpcodeTablePtr = &RegOp2MemOpTable0;
2029 } else if (i == 1) {
2030 OpcodeTablePtr = &RegOp2MemOpTable1;
2031 } else if (i == 2) {
2032 OpcodeTablePtr = &RegOp2MemOpTable2;
2035 // If table selected...
2036 if (OpcodeTablePtr) {
2037 // Find the Opcode to fuse
2038 DenseMap<unsigned*, unsigned>::iterator I =
2039 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
2040 if (I != OpcodeTablePtr->end()) {
2042 NewMI = FuseTwoAddrInst(MF, I->second, MOs, MI, *this);
2044 NewMI = FuseInst(MF, I->second, i, MOs, MI, *this);
2050 if (PrintFailedFusing)
2051 cerr << "We failed to fuse operand " << i << " in " << *MI;
2056 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2058 const SmallVectorImpl<unsigned> &Ops,
2059 int FrameIndex) const {
2060 // Check switch flag
2061 if (NoFusing) return NULL;
2063 const MachineFrameInfo *MFI = MF.getFrameInfo();
2064 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2065 // FIXME: Move alignment requirement into tables?
2066 if (Alignment < 16) {
2067 switch (MI->getOpcode()) {
2069 // Not always safe to fold movsd into these instructions since their load
2070 // folding variants expects the address to be 16 byte aligned.
2071 case X86::FsANDNPDrr:
2072 case X86::FsANDNPSrr:
2073 case X86::FsANDPDrr:
2074 case X86::FsANDPSrr:
2077 case X86::FsXORPDrr:
2078 case X86::FsXORPSrr:
2083 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2084 unsigned NewOpc = 0;
2085 switch (MI->getOpcode()) {
2086 default: return NULL;
2087 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2088 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2089 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2090 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2092 // Change to CMPXXri r, 0 first.
2093 MI->setDesc(get(NewOpc));
2094 MI->getOperand(1).ChangeToImmediate(0);
2095 } else if (Ops.size() != 1)
2098 SmallVector<MachineOperand,4> MOs;
2099 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2100 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2103 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2105 const SmallVectorImpl<unsigned> &Ops,
2106 MachineInstr *LoadMI) const {
2107 // Check switch flag
2108 if (NoFusing) return NULL;
2110 // Determine the alignment of the load.
2111 unsigned Alignment = 0;
2112 if (LoadMI->hasOneMemOperand())
2113 Alignment = LoadMI->memoperands_begin()->getAlignment();
2115 // FIXME: Move alignment requirement into tables?
2116 if (Alignment < 16) {
2117 switch (MI->getOpcode()) {
2119 // Not always safe to fold movsd into these instructions since their load
2120 // folding variants expects the address to be 16 byte aligned.
2121 case X86::FsANDNPDrr:
2122 case X86::FsANDNPSrr:
2123 case X86::FsANDPDrr:
2124 case X86::FsANDPSrr:
2127 case X86::FsXORPDrr:
2128 case X86::FsXORPSrr:
2133 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2134 unsigned NewOpc = 0;
2135 switch (MI->getOpcode()) {
2136 default: return NULL;
2137 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2138 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2139 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2140 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2142 // Change to CMPXXri r, 0 first.
2143 MI->setDesc(get(NewOpc));
2144 MI->getOperand(1).ChangeToImmediate(0);
2145 } else if (Ops.size() != 1)
2148 SmallVector<MachineOperand,4> MOs;
2149 if (LoadMI->getOpcode() == X86::V_SET0 ||
2150 LoadMI->getOpcode() == X86::V_SETALLONES) {
2151 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
2152 // Create a constant-pool entry and operands to load from it.
2154 // x86-32 PIC requires a PIC base register for constant pools.
2155 unsigned PICBase = 0;
2156 if (TM.getRelocationModel() == Reloc::PIC_ &&
2157 !TM.getSubtarget<X86Subtarget>().is64Bit())
2158 // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
2159 // This doesn't work for several reasons.
2160 // 1. GlobalBaseReg may have been spilled.
2161 // 2. It may not be live at MI.
2164 // Create a v4i32 constant-pool entry.
2165 MachineConstantPool &MCP = *MF.getConstantPool();
2166 const VectorType *Ty = VectorType::get(Type::Int32Ty, 4);
2167 Constant *C = LoadMI->getOpcode() == X86::V_SET0 ?
2168 ConstantVector::getNullValue(Ty) :
2169 ConstantVector::getAllOnesValue(Ty);
2170 unsigned CPI = MCP.getConstantPoolIndex(C, /*AlignmentLog2=*/4);
2172 // Create operands to load from the constant pool entry.
2173 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
2174 MOs.push_back(MachineOperand::CreateImm(1));
2175 MOs.push_back(MachineOperand::CreateReg(0, false));
2176 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
2178 // Folding a normal load. Just copy the load's address operands.
2179 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2180 for (unsigned i = NumOps - 4; i != NumOps; ++i)
2181 MOs.push_back(LoadMI->getOperand(i));
2183 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2187 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2188 const SmallVectorImpl<unsigned> &Ops) const {
2189 // Check switch flag
2190 if (NoFusing) return 0;
2192 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2193 switch (MI->getOpcode()) {
2194 default: return false;
2203 if (Ops.size() != 1)
2206 unsigned OpNum = Ops[0];
2207 unsigned Opc = MI->getOpcode();
2208 unsigned NumOps = MI->getDesc().getNumOperands();
2209 bool isTwoAddr = NumOps > 1 &&
2210 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2212 // Folding a memory location into the two-address part of a two-address
2213 // instruction is different than folding it other places. It requires
2214 // replacing the *two* registers with the memory location.
2215 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2216 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2217 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2218 } else if (OpNum == 0) { // If operand 0
2227 OpcodeTablePtr = &RegOp2MemOpTable0;
2228 } else if (OpNum == 1) {
2229 OpcodeTablePtr = &RegOp2MemOpTable1;
2230 } else if (OpNum == 2) {
2231 OpcodeTablePtr = &RegOp2MemOpTable2;
2234 if (OpcodeTablePtr) {
2235 // Find the Opcode to fuse
2236 DenseMap<unsigned*, unsigned>::iterator I =
2237 OpcodeTablePtr->find((unsigned*)Opc);
2238 if (I != OpcodeTablePtr->end())
2244 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2245 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2246 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2247 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2248 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2249 if (I == MemOp2RegOpTable.end())
2251 unsigned Opc = I->second.first;
2252 unsigned Index = I->second.second & 0xf;
2253 bool FoldedLoad = I->second.second & (1 << 4);
2254 bool FoldedStore = I->second.second & (1 << 5);
2255 if (UnfoldLoad && !FoldedLoad)
2257 UnfoldLoad &= FoldedLoad;
2258 if (UnfoldStore && !FoldedStore)
2260 UnfoldStore &= FoldedStore;
2262 const TargetInstrDesc &TID = get(Opc);
2263 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2264 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2265 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2266 SmallVector<MachineOperand,4> AddrOps;
2267 SmallVector<MachineOperand,2> BeforeOps;
2268 SmallVector<MachineOperand,2> AfterOps;
2269 SmallVector<MachineOperand,4> ImpOps;
2270 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2271 MachineOperand &Op = MI->getOperand(i);
2272 if (i >= Index && i < Index+4)
2273 AddrOps.push_back(Op);
2274 else if (Op.isReg() && Op.isImplicit())
2275 ImpOps.push_back(Op);
2277 BeforeOps.push_back(Op);
2279 AfterOps.push_back(Op);
2282 // Emit the load instruction.
2284 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2286 // Address operands cannot be marked isKill.
2287 for (unsigned i = 1; i != 5; ++i) {
2288 MachineOperand &MO = NewMIs[0]->getOperand(i);
2290 MO.setIsKill(false);
2295 // Emit the data processing instruction.
2296 MachineInstr *DataMI = MF.CreateMachineInstr(TID, true);
2297 MachineInstrBuilder MIB(DataMI);
2300 MIB.addReg(Reg, true);
2301 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2302 MIB = X86InstrAddOperand(MIB, BeforeOps[i]);
2305 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2306 MIB = X86InstrAddOperand(MIB, AfterOps[i]);
2307 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2308 MachineOperand &MO = ImpOps[i];
2309 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2311 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2312 unsigned NewOpc = 0;
2313 switch (DataMI->getOpcode()) {
2315 case X86::CMP64ri32:
2319 MachineOperand &MO0 = DataMI->getOperand(0);
2320 MachineOperand &MO1 = DataMI->getOperand(1);
2321 if (MO1.getImm() == 0) {
2322 switch (DataMI->getOpcode()) {
2324 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2325 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2326 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2327 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2329 DataMI->setDesc(get(NewOpc));
2330 MO1.ChangeToRegister(MO0.getReg(), false);
2334 NewMIs.push_back(DataMI);
2336 // Emit the store instruction.
2338 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2339 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2340 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2341 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2348 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2349 SmallVectorImpl<SDNode*> &NewNodes) const {
2350 if (!N->isMachineOpcode())
2353 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2354 MemOp2RegOpTable.find((unsigned*)N->getMachineOpcode());
2355 if (I == MemOp2RegOpTable.end())
2357 unsigned Opc = I->second.first;
2358 unsigned Index = I->second.second & 0xf;
2359 bool FoldedLoad = I->second.second & (1 << 4);
2360 bool FoldedStore = I->second.second & (1 << 5);
2361 const TargetInstrDesc &TID = get(Opc);
2362 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2363 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2364 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2365 std::vector<SDValue> AddrOps;
2366 std::vector<SDValue> BeforeOps;
2367 std::vector<SDValue> AfterOps;
2368 unsigned NumOps = N->getNumOperands();
2369 for (unsigned i = 0; i != NumOps-1; ++i) {
2370 SDValue Op = N->getOperand(i);
2371 if (i >= Index && i < Index+4)
2372 AddrOps.push_back(Op);
2374 BeforeOps.push_back(Op);
2376 AfterOps.push_back(Op);
2378 SDValue Chain = N->getOperand(NumOps-1);
2379 AddrOps.push_back(Chain);
2381 // Emit the load instruction.
2383 const MachineFunction &MF = DAG.getMachineFunction();
2385 MVT VT = *RC->vt_begin();
2386 bool isAligned = (RI.getStackAlignment() >= 16) ||
2387 RI.needsStackRealignment(MF);
2388 Load = DAG.getTargetNode(getLoadRegOpcode(RC, isAligned),
2390 &AddrOps[0], AddrOps.size());
2391 NewNodes.push_back(Load);
2394 // Emit the data processing instruction.
2395 std::vector<MVT> VTs;
2396 const TargetRegisterClass *DstRC = 0;
2397 if (TID.getNumDefs() > 0) {
2398 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2399 DstRC = DstTOI.isLookupPtrRegClass()
2400 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2401 VTs.push_back(*DstRC->vt_begin());
2403 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2404 MVT VT = N->getValueType(i);
2405 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2409 BeforeOps.push_back(SDValue(Load, 0));
2410 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2411 SDNode *NewNode= DAG.getTargetNode(Opc, VTs, &BeforeOps[0], BeforeOps.size());
2412 NewNodes.push_back(NewNode);
2414 // Emit the store instruction.
2417 AddrOps.push_back(SDValue(NewNode, 0));
2418 AddrOps.push_back(Chain);
2419 bool isAligned = (RI.getStackAlignment() >= 16) ||
2420 RI.needsStackRealignment(MF);
2421 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, isAligned),
2422 MVT::Other, &AddrOps[0], AddrOps.size());
2423 NewNodes.push_back(Store);
2429 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2430 bool UnfoldLoad, bool UnfoldStore) const {
2431 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2432 MemOp2RegOpTable.find((unsigned*)Opc);
2433 if (I == MemOp2RegOpTable.end())
2435 bool FoldedLoad = I->second.second & (1 << 4);
2436 bool FoldedStore = I->second.second & (1 << 5);
2437 if (UnfoldLoad && !FoldedLoad)
2439 if (UnfoldStore && !FoldedStore)
2441 return I->second.first;
2444 bool X86InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
2445 if (MBB.empty()) return false;
2447 switch (MBB.back().getOpcode()) {
2448 case X86::TCRETURNri:
2449 case X86::TCRETURNdi:
2450 case X86::RET: // Return.
2455 case X86::JMP: // Uncond branch.
2456 case X86::JMP32r: // Indirect branch.
2457 case X86::JMP64r: // Indirect branch (64-bit).
2458 case X86::JMP32m: // Indirect branch through mem.
2459 case X86::JMP64m: // Indirect branch through mem (64-bit).
2461 default: return false;
2466 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
2467 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2468 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
2469 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
2471 Cond[0].setImm(GetOppositeBranchCondition(CC));
2476 IgnoreRegisterClassBarriers(const TargetRegisterClass *RC) const {
2477 // FIXME: Ignore bariers of x87 stack registers for now. We can't
2478 // allow any loads of these registers before FpGet_ST0_80.
2479 return RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
2480 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass;
2483 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
2484 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
2485 if (Subtarget->is64Bit())
2486 return &X86::GR64RegClass;
2488 return &X86::GR32RegClass;
2491 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2492 switch (Desc->TSFlags & X86II::ImmMask) {
2493 case X86II::Imm8: return 1;
2494 case X86II::Imm16: return 2;
2495 case X86II::Imm32: return 4;
2496 case X86II::Imm64: return 8;
2497 default: assert(0 && "Immediate size not set!");
2502 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2503 /// e.g. r8, xmm8, etc.
2504 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2505 if (!MO.isReg()) return false;
2506 switch (MO.getReg()) {
2508 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2509 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2510 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2511 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2512 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2513 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2514 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2515 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2516 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2517 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2524 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2525 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2526 /// size, and 3) use of X86-64 extended registers.
2527 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2529 const TargetInstrDesc &Desc = MI.getDesc();
2531 // Pseudo instructions do not need REX prefix byte.
2532 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2534 if (Desc.TSFlags & X86II::REX_W)
2537 unsigned NumOps = Desc.getNumOperands();
2539 bool isTwoAddr = NumOps > 1 &&
2540 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2542 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2543 unsigned i = isTwoAddr ? 1 : 0;
2544 for (unsigned e = NumOps; i != e; ++i) {
2545 const MachineOperand& MO = MI.getOperand(i);
2547 unsigned Reg = MO.getReg();
2548 if (isX86_64NonExtLowByteReg(Reg))
2553 switch (Desc.TSFlags & X86II::FormMask) {
2554 case X86II::MRMInitReg:
2555 if (isX86_64ExtendedReg(MI.getOperand(0)))
2556 REX |= (1 << 0) | (1 << 2);
2558 case X86II::MRMSrcReg: {
2559 if (isX86_64ExtendedReg(MI.getOperand(0)))
2561 i = isTwoAddr ? 2 : 1;
2562 for (unsigned e = NumOps; i != e; ++i) {
2563 const MachineOperand& MO = MI.getOperand(i);
2564 if (isX86_64ExtendedReg(MO))
2569 case X86II::MRMSrcMem: {
2570 if (isX86_64ExtendedReg(MI.getOperand(0)))
2573 i = isTwoAddr ? 2 : 1;
2574 for (; i != NumOps; ++i) {
2575 const MachineOperand& MO = MI.getOperand(i);
2577 if (isX86_64ExtendedReg(MO))
2584 case X86II::MRM0m: case X86II::MRM1m:
2585 case X86II::MRM2m: case X86II::MRM3m:
2586 case X86II::MRM4m: case X86II::MRM5m:
2587 case X86II::MRM6m: case X86II::MRM7m:
2588 case X86II::MRMDestMem: {
2589 unsigned e = isTwoAddr ? 5 : 4;
2590 i = isTwoAddr ? 1 : 0;
2591 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2594 for (; i != e; ++i) {
2595 const MachineOperand& MO = MI.getOperand(i);
2597 if (isX86_64ExtendedReg(MO))
2605 if (isX86_64ExtendedReg(MI.getOperand(0)))
2607 i = isTwoAddr ? 2 : 1;
2608 for (unsigned e = NumOps; i != e; ++i) {
2609 const MachineOperand& MO = MI.getOperand(i);
2610 if (isX86_64ExtendedReg(MO))
2620 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2621 /// relative block address instruction
2623 static unsigned sizePCRelativeBlockAddress() {
2627 /// sizeGlobalAddress - Give the size of the emission of this global address
2629 static unsigned sizeGlobalAddress(bool dword) {
2630 return dword ? 8 : 4;
2633 /// sizeConstPoolAddress - Give the size of the emission of this constant
2636 static unsigned sizeConstPoolAddress(bool dword) {
2637 return dword ? 8 : 4;
2640 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2643 static unsigned sizeExternalSymbolAddress(bool dword) {
2644 return dword ? 8 : 4;
2647 /// sizeJumpTableAddress - Give the size of the emission of this jump
2650 static unsigned sizeJumpTableAddress(bool dword) {
2651 return dword ? 8 : 4;
2654 static unsigned sizeConstant(unsigned Size) {
2658 static unsigned sizeRegModRMByte(){
2662 static unsigned sizeSIBByte(){
2666 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2667 unsigned FinalSize = 0;
2668 // If this is a simple integer displacement that doesn't require a relocation.
2670 FinalSize += sizeConstant(4);
2674 // Otherwise, this is something that requires a relocation.
2675 if (RelocOp->isGlobal()) {
2676 FinalSize += sizeGlobalAddress(false);
2677 } else if (RelocOp->isCPI()) {
2678 FinalSize += sizeConstPoolAddress(false);
2679 } else if (RelocOp->isJTI()) {
2680 FinalSize += sizeJumpTableAddress(false);
2682 assert(0 && "Unknown value to relocate!");
2687 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2688 bool IsPIC, bool Is64BitMode) {
2689 const MachineOperand &Op3 = MI.getOperand(Op+3);
2691 const MachineOperand *DispForReloc = 0;
2692 unsigned FinalSize = 0;
2694 // Figure out what sort of displacement we have to handle here.
2695 if (Op3.isGlobal()) {
2696 DispForReloc = &Op3;
2697 } else if (Op3.isCPI()) {
2698 if (Is64BitMode || IsPIC) {
2699 DispForReloc = &Op3;
2703 } else if (Op3.isJTI()) {
2704 if (Is64BitMode || IsPIC) {
2705 DispForReloc = &Op3;
2713 const MachineOperand &Base = MI.getOperand(Op);
2714 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2716 unsigned BaseReg = Base.getReg();
2718 // Is a SIB byte needed?
2719 if (IndexReg.getReg() == 0 &&
2720 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2721 if (BaseReg == 0) { // Just a displacement?
2722 // Emit special case [disp32] encoding
2724 FinalSize += getDisplacementFieldSize(DispForReloc);
2726 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2727 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2728 // Emit simple indirect register encoding... [EAX] f.e.
2730 // Be pessimistic and assume it's a disp32, not a disp8
2732 // Emit the most general non-SIB encoding: [REG+disp32]
2734 FinalSize += getDisplacementFieldSize(DispForReloc);
2738 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2739 assert(IndexReg.getReg() != X86::ESP &&
2740 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2742 bool ForceDisp32 = false;
2743 if (BaseReg == 0 || DispForReloc) {
2744 // Emit the normal disp32 encoding.
2751 FinalSize += sizeSIBByte();
2753 // Do we need to output a displacement?
2754 if (DispVal != 0 || ForceDisp32) {
2755 FinalSize += getDisplacementFieldSize(DispForReloc);
2762 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2763 const TargetInstrDesc *Desc,
2764 bool IsPIC, bool Is64BitMode) {
2766 unsigned Opcode = Desc->Opcode;
2767 unsigned FinalSize = 0;
2769 // Emit the lock opcode prefix as needed.
2770 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2772 // Emit segment overrid opcode prefix as needed.
2773 switch (Desc->TSFlags & X86II::SegOvrMask) {
2778 default: assert(0 && "Invalid segment!");
2779 case 0: break; // No segment override!
2782 // Emit the repeat opcode prefix as needed.
2783 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2785 // Emit the operand size opcode prefix as needed.
2786 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2788 // Emit the address size opcode prefix as needed.
2789 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2791 bool Need0FPrefix = false;
2792 switch (Desc->TSFlags & X86II::Op0Mask) {
2793 case X86II::TB: // Two-byte opcode prefix
2794 case X86II::T8: // 0F 38
2795 case X86II::TA: // 0F 3A
2796 Need0FPrefix = true;
2798 case X86II::REP: break; // already handled.
2799 case X86II::XS: // F3 0F
2801 Need0FPrefix = true;
2803 case X86II::XD: // F2 0F
2805 Need0FPrefix = true;
2807 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2808 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2810 break; // Two-byte opcode prefix
2811 default: assert(0 && "Invalid prefix!");
2812 case 0: break; // No prefix!
2817 unsigned REX = X86InstrInfo::determineREX(MI);
2822 // 0x0F escape code must be emitted just before the opcode.
2826 switch (Desc->TSFlags & X86II::Op0Mask) {
2827 case X86II::T8: // 0F 38
2830 case X86II::TA: // 0F 3A
2835 // If this is a two-address instruction, skip one of the register operands.
2836 unsigned NumOps = Desc->getNumOperands();
2838 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2841 switch (Desc->TSFlags & X86II::FormMask) {
2842 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
2844 // Remember the current PC offset, this is the PIC relocation
2849 case TargetInstrInfo::INLINEASM: {
2850 const MachineFunction *MF = MI.getParent()->getParent();
2851 const char *AsmStr = MI.getOperand(0).getSymbolName();
2852 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
2853 FinalSize += AI->getInlineAsmLength(AsmStr);
2856 case TargetInstrInfo::DBG_LABEL:
2857 case TargetInstrInfo::EH_LABEL:
2859 case TargetInstrInfo::IMPLICIT_DEF:
2860 case TargetInstrInfo::DECLARE:
2861 case X86::DWARF_LOC:
2862 case X86::FP_REG_KILL:
2864 case X86::MOVPC32r: {
2865 // This emits the "call" portion of this pseudo instruction.
2867 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2871 case X86::TLS_gs_ri:
2873 FinalSize += sizeGlobalAddress(false);
2881 if (CurOp != NumOps) {
2882 const MachineOperand &MO = MI.getOperand(CurOp++);
2884 FinalSize += sizePCRelativeBlockAddress();
2885 } else if (MO.isGlobal()) {
2886 FinalSize += sizeGlobalAddress(false);
2887 } else if (MO.isSymbol()) {
2888 FinalSize += sizeExternalSymbolAddress(false);
2889 } else if (MO.isImm()) {
2890 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2892 assert(0 && "Unknown RawFrm operand!");
2897 case X86II::AddRegFrm:
2901 if (CurOp != NumOps) {
2902 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2903 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2905 FinalSize += sizeConstant(Size);
2908 if (Opcode == X86::MOV64ri)
2910 if (MO1.isGlobal()) {
2911 FinalSize += sizeGlobalAddress(dword);
2912 } else if (MO1.isSymbol())
2913 FinalSize += sizeExternalSymbolAddress(dword);
2914 else if (MO1.isCPI())
2915 FinalSize += sizeConstPoolAddress(dword);
2916 else if (MO1.isJTI())
2917 FinalSize += sizeJumpTableAddress(dword);
2922 case X86II::MRMDestReg: {
2924 FinalSize += sizeRegModRMByte();
2926 if (CurOp != NumOps) {
2928 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2932 case X86II::MRMDestMem: {
2934 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2936 if (CurOp != NumOps) {
2938 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2943 case X86II::MRMSrcReg:
2945 FinalSize += sizeRegModRMByte();
2947 if (CurOp != NumOps) {
2949 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2953 case X86II::MRMSrcMem: {
2956 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
2958 if (CurOp != NumOps) {
2960 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2965 case X86II::MRM0r: case X86II::MRM1r:
2966 case X86II::MRM2r: case X86II::MRM3r:
2967 case X86II::MRM4r: case X86II::MRM5r:
2968 case X86II::MRM6r: case X86II::MRM7r:
2971 FinalSize += sizeRegModRMByte();
2973 if (CurOp != NumOps) {
2974 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2975 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2977 FinalSize += sizeConstant(Size);
2980 if (Opcode == X86::MOV64ri32)
2982 if (MO1.isGlobal()) {
2983 FinalSize += sizeGlobalAddress(dword);
2984 } else if (MO1.isSymbol())
2985 FinalSize += sizeExternalSymbolAddress(dword);
2986 else if (MO1.isCPI())
2987 FinalSize += sizeConstPoolAddress(dword);
2988 else if (MO1.isJTI())
2989 FinalSize += sizeJumpTableAddress(dword);
2994 case X86II::MRM0m: case X86II::MRM1m:
2995 case X86II::MRM2m: case X86II::MRM3m:
2996 case X86II::MRM4m: case X86II::MRM5m:
2997 case X86II::MRM6m: case X86II::MRM7m: {
3000 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
3003 if (CurOp != NumOps) {
3004 const MachineOperand &MO = MI.getOperand(CurOp++);
3005 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3007 FinalSize += sizeConstant(Size);
3010 if (Opcode == X86::MOV64mi32)
3012 if (MO.isGlobal()) {
3013 FinalSize += sizeGlobalAddress(dword);
3014 } else if (MO.isSymbol())
3015 FinalSize += sizeExternalSymbolAddress(dword);
3016 else if (MO.isCPI())
3017 FinalSize += sizeConstPoolAddress(dword);
3018 else if (MO.isJTI())
3019 FinalSize += sizeJumpTableAddress(dword);
3025 case X86II::MRMInitReg:
3027 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
3028 FinalSize += sizeRegModRMByte();
3033 if (!Desc->isVariadic() && CurOp != NumOps) {
3034 cerr << "Cannot determine size: ";
3045 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
3046 const TargetInstrDesc &Desc = MI->getDesc();
3047 bool IsPIC = (TM.getRelocationModel() == Reloc::PIC_);
3048 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
3049 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
3050 if (Desc.getOpcode() == X86::MOVPC32r) {
3051 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);
3056 /// getGlobalBaseReg - Return a virtual register initialized with the
3057 /// the global base register value. Output instructions required to
3058 /// initialize the register in the function entry block, if necessary.
3060 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3061 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3062 "X86-64 PIC uses RIP relative addressing");
3064 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3065 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3066 if (GlobalBaseReg != 0)
3067 return GlobalBaseReg;
3069 // Insert the set of GlobalBaseReg into the first MBB of the function
3070 MachineBasicBlock &FirstMBB = MF->front();
3071 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3072 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3073 unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3075 const TargetInstrInfo *TII = TM.getInstrInfo();
3076 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3077 // only used in JIT code emission as displacement to pc.
3078 BuildMI(FirstMBB, MBBI, TII->get(X86::MOVPC32r), PC).addImm(0);
3080 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3081 // not to pc, but to _GLOBAL_ADDRESS_TABLE_ external
3082 if (TM.getRelocationModel() == Reloc::PIC_ &&
3083 TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3085 RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3086 BuildMI(FirstMBB, MBBI, TII->get(X86::ADD32ri), GlobalBaseReg)
3087 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_");
3092 X86FI->setGlobalBaseReg(GlobalBaseReg);
3093 return GlobalBaseReg;