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/ADT/STLExtras.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/LiveVariables.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Target/TargetOptions.h"
28 #include "llvm/Target/TargetAsmInfo.h"
34 NoFusing("disable-spill-fusing",
35 cl::desc("Disable fusing of spill code into instructions"));
37 PrintFailedFusing("print-failed-fuse-candidates",
38 cl::desc("Print instructions that the allocator wants to"
39 " fuse, but the X86 backend currently can't"),
42 ReMatPICStubLoad("remat-pic-stub-load",
43 cl::desc("Re-materialize load from stub in PIC mode"),
44 cl::init(false), cl::Hidden);
47 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
48 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
49 TM(tm), RI(tm, *this) {
50 SmallVector<unsigned,16> AmbEntries;
51 static const unsigned OpTbl2Addr[][2] = {
52 { X86::ADC32ri, X86::ADC32mi },
53 { X86::ADC32ri8, X86::ADC32mi8 },
54 { X86::ADC32rr, X86::ADC32mr },
55 { X86::ADC64ri32, X86::ADC64mi32 },
56 { X86::ADC64ri8, X86::ADC64mi8 },
57 { X86::ADC64rr, X86::ADC64mr },
58 { X86::ADD16ri, X86::ADD16mi },
59 { X86::ADD16ri8, X86::ADD16mi8 },
60 { X86::ADD16rr, X86::ADD16mr },
61 { X86::ADD32ri, X86::ADD32mi },
62 { X86::ADD32ri8, X86::ADD32mi8 },
63 { X86::ADD32rr, X86::ADD32mr },
64 { X86::ADD64ri32, X86::ADD64mi32 },
65 { X86::ADD64ri8, X86::ADD64mi8 },
66 { X86::ADD64rr, X86::ADD64mr },
67 { X86::ADD8ri, X86::ADD8mi },
68 { X86::ADD8rr, X86::ADD8mr },
69 { X86::AND16ri, X86::AND16mi },
70 { X86::AND16ri8, X86::AND16mi8 },
71 { X86::AND16rr, X86::AND16mr },
72 { X86::AND32ri, X86::AND32mi },
73 { X86::AND32ri8, X86::AND32mi8 },
74 { X86::AND32rr, X86::AND32mr },
75 { X86::AND64ri32, X86::AND64mi32 },
76 { X86::AND64ri8, X86::AND64mi8 },
77 { X86::AND64rr, X86::AND64mr },
78 { X86::AND8ri, X86::AND8mi },
79 { X86::AND8rr, X86::AND8mr },
80 { X86::DEC16r, X86::DEC16m },
81 { X86::DEC32r, X86::DEC32m },
82 { X86::DEC64_16r, X86::DEC64_16m },
83 { X86::DEC64_32r, X86::DEC64_32m },
84 { X86::DEC64r, X86::DEC64m },
85 { X86::DEC8r, X86::DEC8m },
86 { X86::INC16r, X86::INC16m },
87 { X86::INC32r, X86::INC32m },
88 { X86::INC64_16r, X86::INC64_16m },
89 { X86::INC64_32r, X86::INC64_32m },
90 { X86::INC64r, X86::INC64m },
91 { X86::INC8r, X86::INC8m },
92 { X86::NEG16r, X86::NEG16m },
93 { X86::NEG32r, X86::NEG32m },
94 { X86::NEG64r, X86::NEG64m },
95 { X86::NEG8r, X86::NEG8m },
96 { X86::NOT16r, X86::NOT16m },
97 { X86::NOT32r, X86::NOT32m },
98 { X86::NOT64r, X86::NOT64m },
99 { X86::NOT8r, X86::NOT8m },
100 { X86::OR16ri, X86::OR16mi },
101 { X86::OR16ri8, X86::OR16mi8 },
102 { X86::OR16rr, X86::OR16mr },
103 { X86::OR32ri, X86::OR32mi },
104 { X86::OR32ri8, X86::OR32mi8 },
105 { X86::OR32rr, X86::OR32mr },
106 { X86::OR64ri32, X86::OR64mi32 },
107 { X86::OR64ri8, X86::OR64mi8 },
108 { X86::OR64rr, X86::OR64mr },
109 { X86::OR8ri, X86::OR8mi },
110 { X86::OR8rr, X86::OR8mr },
111 { X86::ROL16r1, X86::ROL16m1 },
112 { X86::ROL16rCL, X86::ROL16mCL },
113 { X86::ROL16ri, X86::ROL16mi },
114 { X86::ROL32r1, X86::ROL32m1 },
115 { X86::ROL32rCL, X86::ROL32mCL },
116 { X86::ROL32ri, X86::ROL32mi },
117 { X86::ROL64r1, X86::ROL64m1 },
118 { X86::ROL64rCL, X86::ROL64mCL },
119 { X86::ROL64ri, X86::ROL64mi },
120 { X86::ROL8r1, X86::ROL8m1 },
121 { X86::ROL8rCL, X86::ROL8mCL },
122 { X86::ROL8ri, X86::ROL8mi },
123 { X86::ROR16r1, X86::ROR16m1 },
124 { X86::ROR16rCL, X86::ROR16mCL },
125 { X86::ROR16ri, X86::ROR16mi },
126 { X86::ROR32r1, X86::ROR32m1 },
127 { X86::ROR32rCL, X86::ROR32mCL },
128 { X86::ROR32ri, X86::ROR32mi },
129 { X86::ROR64r1, X86::ROR64m1 },
130 { X86::ROR64rCL, X86::ROR64mCL },
131 { X86::ROR64ri, X86::ROR64mi },
132 { X86::ROR8r1, X86::ROR8m1 },
133 { X86::ROR8rCL, X86::ROR8mCL },
134 { X86::ROR8ri, X86::ROR8mi },
135 { X86::SAR16r1, X86::SAR16m1 },
136 { X86::SAR16rCL, X86::SAR16mCL },
137 { X86::SAR16ri, X86::SAR16mi },
138 { X86::SAR32r1, X86::SAR32m1 },
139 { X86::SAR32rCL, X86::SAR32mCL },
140 { X86::SAR32ri, X86::SAR32mi },
141 { X86::SAR64r1, X86::SAR64m1 },
142 { X86::SAR64rCL, X86::SAR64mCL },
143 { X86::SAR64ri, X86::SAR64mi },
144 { X86::SAR8r1, X86::SAR8m1 },
145 { X86::SAR8rCL, X86::SAR8mCL },
146 { X86::SAR8ri, X86::SAR8mi },
147 { X86::SBB32ri, X86::SBB32mi },
148 { X86::SBB32ri8, X86::SBB32mi8 },
149 { X86::SBB32rr, X86::SBB32mr },
150 { X86::SBB64ri32, X86::SBB64mi32 },
151 { X86::SBB64ri8, X86::SBB64mi8 },
152 { X86::SBB64rr, X86::SBB64mr },
153 { X86::SHL16rCL, X86::SHL16mCL },
154 { X86::SHL16ri, X86::SHL16mi },
155 { X86::SHL32rCL, X86::SHL32mCL },
156 { X86::SHL32ri, X86::SHL32mi },
157 { X86::SHL64rCL, X86::SHL64mCL },
158 { X86::SHL64ri, X86::SHL64mi },
159 { X86::SHL8rCL, X86::SHL8mCL },
160 { X86::SHL8ri, X86::SHL8mi },
161 { X86::SHLD16rrCL, X86::SHLD16mrCL },
162 { X86::SHLD16rri8, X86::SHLD16mri8 },
163 { X86::SHLD32rrCL, X86::SHLD32mrCL },
164 { X86::SHLD32rri8, X86::SHLD32mri8 },
165 { X86::SHLD64rrCL, X86::SHLD64mrCL },
166 { X86::SHLD64rri8, X86::SHLD64mri8 },
167 { X86::SHR16r1, X86::SHR16m1 },
168 { X86::SHR16rCL, X86::SHR16mCL },
169 { X86::SHR16ri, X86::SHR16mi },
170 { X86::SHR32r1, X86::SHR32m1 },
171 { X86::SHR32rCL, X86::SHR32mCL },
172 { X86::SHR32ri, X86::SHR32mi },
173 { X86::SHR64r1, X86::SHR64m1 },
174 { X86::SHR64rCL, X86::SHR64mCL },
175 { X86::SHR64ri, X86::SHR64mi },
176 { X86::SHR8r1, X86::SHR8m1 },
177 { X86::SHR8rCL, X86::SHR8mCL },
178 { X86::SHR8ri, X86::SHR8mi },
179 { X86::SHRD16rrCL, X86::SHRD16mrCL },
180 { X86::SHRD16rri8, X86::SHRD16mri8 },
181 { X86::SHRD32rrCL, X86::SHRD32mrCL },
182 { X86::SHRD32rri8, X86::SHRD32mri8 },
183 { X86::SHRD64rrCL, X86::SHRD64mrCL },
184 { X86::SHRD64rri8, X86::SHRD64mri8 },
185 { X86::SUB16ri, X86::SUB16mi },
186 { X86::SUB16ri8, X86::SUB16mi8 },
187 { X86::SUB16rr, X86::SUB16mr },
188 { X86::SUB32ri, X86::SUB32mi },
189 { X86::SUB32ri8, X86::SUB32mi8 },
190 { X86::SUB32rr, X86::SUB32mr },
191 { X86::SUB64ri32, X86::SUB64mi32 },
192 { X86::SUB64ri8, X86::SUB64mi8 },
193 { X86::SUB64rr, X86::SUB64mr },
194 { X86::SUB8ri, X86::SUB8mi },
195 { X86::SUB8rr, X86::SUB8mr },
196 { X86::XOR16ri, X86::XOR16mi },
197 { X86::XOR16ri8, X86::XOR16mi8 },
198 { X86::XOR16rr, X86::XOR16mr },
199 { X86::XOR32ri, X86::XOR32mi },
200 { X86::XOR32ri8, X86::XOR32mi8 },
201 { X86::XOR32rr, X86::XOR32mr },
202 { X86::XOR64ri32, X86::XOR64mi32 },
203 { X86::XOR64ri8, X86::XOR64mi8 },
204 { X86::XOR64rr, X86::XOR64mr },
205 { X86::XOR8ri, X86::XOR8mi },
206 { X86::XOR8rr, X86::XOR8mr }
209 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
210 unsigned RegOp = OpTbl2Addr[i][0];
211 unsigned MemOp = OpTbl2Addr[i][1];
212 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp, MemOp)))
213 assert(false && "Duplicated entries?");
214 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
215 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
216 std::make_pair(RegOp, AuxInfo))))
217 AmbEntries.push_back(MemOp);
220 // If the third value is 1, then it's folding either a load or a store.
221 static const unsigned OpTbl0[][3] = {
222 { X86::CALL32r, X86::CALL32m, 1 },
223 { X86::CALL64r, X86::CALL64m, 1 },
224 { X86::CMP16ri, X86::CMP16mi, 1 },
225 { X86::CMP16ri8, X86::CMP16mi8, 1 },
226 { X86::CMP16rr, X86::CMP16mr, 1 },
227 { X86::CMP32ri, X86::CMP32mi, 1 },
228 { X86::CMP32ri8, X86::CMP32mi8, 1 },
229 { X86::CMP32rr, X86::CMP32mr, 1 },
230 { X86::CMP64ri32, X86::CMP64mi32, 1 },
231 { X86::CMP64ri8, X86::CMP64mi8, 1 },
232 { X86::CMP64rr, X86::CMP64mr, 1 },
233 { X86::CMP8ri, X86::CMP8mi, 1 },
234 { X86::CMP8rr, X86::CMP8mr, 1 },
235 { X86::DIV16r, X86::DIV16m, 1 },
236 { X86::DIV32r, X86::DIV32m, 1 },
237 { X86::DIV64r, X86::DIV64m, 1 },
238 { X86::DIV8r, X86::DIV8m, 1 },
239 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
240 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
241 { X86::IDIV16r, X86::IDIV16m, 1 },
242 { X86::IDIV32r, X86::IDIV32m, 1 },
243 { X86::IDIV64r, X86::IDIV64m, 1 },
244 { X86::IDIV8r, X86::IDIV8m, 1 },
245 { X86::IMUL16r, X86::IMUL16m, 1 },
246 { X86::IMUL32r, X86::IMUL32m, 1 },
247 { X86::IMUL64r, X86::IMUL64m, 1 },
248 { X86::IMUL8r, X86::IMUL8m, 1 },
249 { X86::JMP32r, X86::JMP32m, 1 },
250 { X86::JMP64r, X86::JMP64m, 1 },
251 { X86::MOV16ri, X86::MOV16mi, 0 },
252 { X86::MOV16rr, X86::MOV16mr, 0 },
253 { X86::MOV16to16_, X86::MOV16_mr, 0 },
254 { X86::MOV32ri, X86::MOV32mi, 0 },
255 { X86::MOV32rr, X86::MOV32mr, 0 },
256 { X86::MOV32to32_, X86::MOV32_mr, 0 },
257 { X86::MOV64ri32, X86::MOV64mi32, 0 },
258 { X86::MOV64rr, X86::MOV64mr, 0 },
259 { X86::MOV8ri, X86::MOV8mi, 0 },
260 { X86::MOV8rr, X86::MOV8mr, 0 },
261 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
262 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
263 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
264 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
265 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
266 { X86::MOVSDrr, X86::MOVSDmr, 0 },
267 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
268 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
269 { X86::MOVSSrr, X86::MOVSSmr, 0 },
270 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
271 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
272 { X86::MUL16r, X86::MUL16m, 1 },
273 { X86::MUL32r, X86::MUL32m, 1 },
274 { X86::MUL64r, X86::MUL64m, 1 },
275 { X86::MUL8r, X86::MUL8m, 1 },
276 { X86::SETAEr, X86::SETAEm, 0 },
277 { X86::SETAr, X86::SETAm, 0 },
278 { X86::SETBEr, X86::SETBEm, 0 },
279 { X86::SETBr, X86::SETBm, 0 },
280 { X86::SETEr, X86::SETEm, 0 },
281 { X86::SETGEr, X86::SETGEm, 0 },
282 { X86::SETGr, X86::SETGm, 0 },
283 { X86::SETLEr, X86::SETLEm, 0 },
284 { X86::SETLr, X86::SETLm, 0 },
285 { X86::SETNEr, X86::SETNEm, 0 },
286 { X86::SETNPr, X86::SETNPm, 0 },
287 { X86::SETNSr, X86::SETNSm, 0 },
288 { X86::SETPr, X86::SETPm, 0 },
289 { X86::SETSr, X86::SETSm, 0 },
290 { X86::TAILJMPr, X86::TAILJMPm, 1 },
291 { X86::TEST16ri, X86::TEST16mi, 1 },
292 { X86::TEST32ri, X86::TEST32mi, 1 },
293 { X86::TEST64ri32, X86::TEST64mi32, 1 },
294 { X86::TEST8ri, X86::TEST8mi, 1 }
297 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
298 unsigned RegOp = OpTbl0[i][0];
299 unsigned MemOp = OpTbl0[i][1];
300 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp, MemOp)))
301 assert(false && "Duplicated entries?");
302 unsigned FoldedLoad = OpTbl0[i][2];
303 // Index 0, folded load or store.
304 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
305 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
306 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
307 std::make_pair(RegOp, AuxInfo))))
308 AmbEntries.push_back(MemOp);
311 static const unsigned OpTbl1[][2] = {
312 { X86::CMP16rr, X86::CMP16rm },
313 { X86::CMP32rr, X86::CMP32rm },
314 { X86::CMP64rr, X86::CMP64rm },
315 { X86::CMP8rr, X86::CMP8rm },
316 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
317 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
318 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
319 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
320 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
321 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
322 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
323 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
324 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
325 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
326 { X86::FsMOVAPDrr, X86::MOVSDrm },
327 { X86::FsMOVAPSrr, X86::MOVSSrm },
328 { X86::IMUL16rri, X86::IMUL16rmi },
329 { X86::IMUL16rri8, X86::IMUL16rmi8 },
330 { X86::IMUL32rri, X86::IMUL32rmi },
331 { X86::IMUL32rri8, X86::IMUL32rmi8 },
332 { X86::IMUL64rri32, X86::IMUL64rmi32 },
333 { X86::IMUL64rri8, X86::IMUL64rmi8 },
334 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
335 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
336 { X86::Int_COMISDrr, X86::Int_COMISDrm },
337 { X86::Int_COMISSrr, X86::Int_COMISSrm },
338 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
339 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
340 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
341 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
342 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
343 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
344 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
345 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
346 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
347 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
348 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
349 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
350 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
351 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
352 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
353 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
354 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
355 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
356 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
357 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
358 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
359 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
360 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
361 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
362 { X86::MOV16rr, X86::MOV16rm },
363 { X86::MOV16to16_, X86::MOV16_rm },
364 { X86::MOV32rr, X86::MOV32rm },
365 { X86::MOV32to32_, X86::MOV32_rm },
366 { X86::MOV64rr, X86::MOV64rm },
367 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
368 { X86::MOV64toSDrr, X86::MOV64toSDrm },
369 { X86::MOV8rr, X86::MOV8rm },
370 { X86::MOVAPDrr, X86::MOVAPDrm },
371 { X86::MOVAPSrr, X86::MOVAPSrm },
372 { X86::MOVDDUPrr, X86::MOVDDUPrm },
373 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
374 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
375 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
376 { X86::MOVSDrr, X86::MOVSDrm },
377 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
378 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
379 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
380 { X86::MOVSSrr, X86::MOVSSrm },
381 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
382 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
383 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
384 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
385 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
386 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
387 { X86::MOVUPDrr, X86::MOVUPDrm },
388 { X86::MOVUPSrr, X86::MOVUPSrm },
389 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
390 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
391 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
392 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
393 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
394 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
395 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
396 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
397 { X86::PSHUFDri, X86::PSHUFDmi },
398 { X86::PSHUFHWri, X86::PSHUFHWmi },
399 { X86::PSHUFLWri, X86::PSHUFLWmi },
400 { X86::RCPPSr, X86::RCPPSm },
401 { X86::RCPPSr_Int, X86::RCPPSm_Int },
402 { X86::RSQRTPSr, X86::RSQRTPSm },
403 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
404 { X86::RSQRTSSr, X86::RSQRTSSm },
405 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
406 { X86::SQRTPDr, X86::SQRTPDm },
407 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
408 { X86::SQRTPSr, X86::SQRTPSm },
409 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
410 { X86::SQRTSDr, X86::SQRTSDm },
411 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
412 { X86::SQRTSSr, X86::SQRTSSm },
413 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
414 { X86::TEST16rr, X86::TEST16rm },
415 { X86::TEST32rr, X86::TEST32rm },
416 { X86::TEST64rr, X86::TEST64rm },
417 { X86::TEST8rr, X86::TEST8rm },
418 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
419 { X86::UCOMISDrr, X86::UCOMISDrm },
420 { X86::UCOMISSrr, X86::UCOMISSrm }
423 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
424 unsigned RegOp = OpTbl1[i][0];
425 unsigned MemOp = OpTbl1[i][1];
426 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp, MemOp)))
427 assert(false && "Duplicated entries?");
428 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
429 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
430 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
431 std::make_pair(RegOp, AuxInfo))))
432 AmbEntries.push_back(MemOp);
435 static const unsigned OpTbl2[][2] = {
436 { X86::ADC32rr, X86::ADC32rm },
437 { X86::ADC64rr, X86::ADC64rm },
438 { X86::ADD16rr, X86::ADD16rm },
439 { X86::ADD32rr, X86::ADD32rm },
440 { X86::ADD64rr, X86::ADD64rm },
441 { X86::ADD8rr, X86::ADD8rm },
442 { X86::ADDPDrr, X86::ADDPDrm },
443 { X86::ADDPSrr, X86::ADDPSrm },
444 { X86::ADDSDrr, X86::ADDSDrm },
445 { X86::ADDSSrr, X86::ADDSSrm },
446 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
447 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
448 { X86::AND16rr, X86::AND16rm },
449 { X86::AND32rr, X86::AND32rm },
450 { X86::AND64rr, X86::AND64rm },
451 { X86::AND8rr, X86::AND8rm },
452 { X86::ANDNPDrr, X86::ANDNPDrm },
453 { X86::ANDNPSrr, X86::ANDNPSrm },
454 { X86::ANDPDrr, X86::ANDPDrm },
455 { X86::ANDPSrr, X86::ANDPSrm },
456 { X86::CMOVA16rr, X86::CMOVA16rm },
457 { X86::CMOVA32rr, X86::CMOVA32rm },
458 { X86::CMOVA64rr, X86::CMOVA64rm },
459 { X86::CMOVAE16rr, X86::CMOVAE16rm },
460 { X86::CMOVAE32rr, X86::CMOVAE32rm },
461 { X86::CMOVAE64rr, X86::CMOVAE64rm },
462 { X86::CMOVB16rr, X86::CMOVB16rm },
463 { X86::CMOVB32rr, X86::CMOVB32rm },
464 { X86::CMOVB64rr, X86::CMOVB64rm },
465 { X86::CMOVBE16rr, X86::CMOVBE16rm },
466 { X86::CMOVBE32rr, X86::CMOVBE32rm },
467 { X86::CMOVBE64rr, X86::CMOVBE64rm },
468 { X86::CMOVE16rr, X86::CMOVE16rm },
469 { X86::CMOVE32rr, X86::CMOVE32rm },
470 { X86::CMOVE64rr, X86::CMOVE64rm },
471 { X86::CMOVG16rr, X86::CMOVG16rm },
472 { X86::CMOVG32rr, X86::CMOVG32rm },
473 { X86::CMOVG64rr, X86::CMOVG64rm },
474 { X86::CMOVGE16rr, X86::CMOVGE16rm },
475 { X86::CMOVGE32rr, X86::CMOVGE32rm },
476 { X86::CMOVGE64rr, X86::CMOVGE64rm },
477 { X86::CMOVL16rr, X86::CMOVL16rm },
478 { X86::CMOVL32rr, X86::CMOVL32rm },
479 { X86::CMOVL64rr, X86::CMOVL64rm },
480 { X86::CMOVLE16rr, X86::CMOVLE16rm },
481 { X86::CMOVLE32rr, X86::CMOVLE32rm },
482 { X86::CMOVLE64rr, X86::CMOVLE64rm },
483 { X86::CMOVNE16rr, X86::CMOVNE16rm },
484 { X86::CMOVNE32rr, X86::CMOVNE32rm },
485 { X86::CMOVNE64rr, X86::CMOVNE64rm },
486 { X86::CMOVNP16rr, X86::CMOVNP16rm },
487 { X86::CMOVNP32rr, X86::CMOVNP32rm },
488 { X86::CMOVNP64rr, X86::CMOVNP64rm },
489 { X86::CMOVNS16rr, X86::CMOVNS16rm },
490 { X86::CMOVNS32rr, X86::CMOVNS32rm },
491 { X86::CMOVNS64rr, X86::CMOVNS64rm },
492 { X86::CMOVP16rr, X86::CMOVP16rm },
493 { X86::CMOVP32rr, X86::CMOVP32rm },
494 { X86::CMOVP64rr, X86::CMOVP64rm },
495 { X86::CMOVS16rr, X86::CMOVS16rm },
496 { X86::CMOVS32rr, X86::CMOVS32rm },
497 { X86::CMOVS64rr, X86::CMOVS64rm },
498 { X86::CMPPDrri, X86::CMPPDrmi },
499 { X86::CMPPSrri, X86::CMPPSrmi },
500 { X86::CMPSDrr, X86::CMPSDrm },
501 { X86::CMPSSrr, X86::CMPSSrm },
502 { X86::DIVPDrr, X86::DIVPDrm },
503 { X86::DIVPSrr, X86::DIVPSrm },
504 { X86::DIVSDrr, X86::DIVSDrm },
505 { X86::DIVSSrr, X86::DIVSSrm },
506 { X86::FsANDNPDrr, X86::FsANDNPDrm },
507 { X86::FsANDNPSrr, X86::FsANDNPSrm },
508 { X86::FsANDPDrr, X86::FsANDPDrm },
509 { X86::FsANDPSrr, X86::FsANDPSrm },
510 { X86::FsORPDrr, X86::FsORPDrm },
511 { X86::FsORPSrr, X86::FsORPSrm },
512 { X86::FsXORPDrr, X86::FsXORPDrm },
513 { X86::FsXORPSrr, X86::FsXORPSrm },
514 { X86::HADDPDrr, X86::HADDPDrm },
515 { X86::HADDPSrr, X86::HADDPSrm },
516 { X86::HSUBPDrr, X86::HSUBPDrm },
517 { X86::HSUBPSrr, X86::HSUBPSrm },
518 { X86::IMUL16rr, X86::IMUL16rm },
519 { X86::IMUL32rr, X86::IMUL32rm },
520 { X86::IMUL64rr, X86::IMUL64rm },
521 { X86::MAXPDrr, X86::MAXPDrm },
522 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
523 { X86::MAXPSrr, X86::MAXPSrm },
524 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
525 { X86::MAXSDrr, X86::MAXSDrm },
526 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
527 { X86::MAXSSrr, X86::MAXSSrm },
528 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
529 { X86::MINPDrr, X86::MINPDrm },
530 { X86::MINPDrr_Int, X86::MINPDrm_Int },
531 { X86::MINPSrr, X86::MINPSrm },
532 { X86::MINPSrr_Int, X86::MINPSrm_Int },
533 { X86::MINSDrr, X86::MINSDrm },
534 { X86::MINSDrr_Int, X86::MINSDrm_Int },
535 { X86::MINSSrr, X86::MINSSrm },
536 { X86::MINSSrr_Int, X86::MINSSrm_Int },
537 { X86::MULPDrr, X86::MULPDrm },
538 { X86::MULPSrr, X86::MULPSrm },
539 { X86::MULSDrr, X86::MULSDrm },
540 { X86::MULSSrr, X86::MULSSrm },
541 { X86::OR16rr, X86::OR16rm },
542 { X86::OR32rr, X86::OR32rm },
543 { X86::OR64rr, X86::OR64rm },
544 { X86::OR8rr, X86::OR8rm },
545 { X86::ORPDrr, X86::ORPDrm },
546 { X86::ORPSrr, X86::ORPSrm },
547 { X86::PACKSSDWrr, X86::PACKSSDWrm },
548 { X86::PACKSSWBrr, X86::PACKSSWBrm },
549 { X86::PACKUSWBrr, X86::PACKUSWBrm },
550 { X86::PADDBrr, X86::PADDBrm },
551 { X86::PADDDrr, X86::PADDDrm },
552 { X86::PADDQrr, X86::PADDQrm },
553 { X86::PADDSBrr, X86::PADDSBrm },
554 { X86::PADDSWrr, X86::PADDSWrm },
555 { X86::PADDWrr, X86::PADDWrm },
556 { X86::PANDNrr, X86::PANDNrm },
557 { X86::PANDrr, X86::PANDrm },
558 { X86::PAVGBrr, X86::PAVGBrm },
559 { X86::PAVGWrr, X86::PAVGWrm },
560 { X86::PCMPEQBrr, X86::PCMPEQBrm },
561 { X86::PCMPEQDrr, X86::PCMPEQDrm },
562 { X86::PCMPEQWrr, X86::PCMPEQWrm },
563 { X86::PCMPGTBrr, X86::PCMPGTBrm },
564 { X86::PCMPGTDrr, X86::PCMPGTDrm },
565 { X86::PCMPGTWrr, X86::PCMPGTWrm },
566 { X86::PINSRWrri, X86::PINSRWrmi },
567 { X86::PMADDWDrr, X86::PMADDWDrm },
568 { X86::PMAXSWrr, X86::PMAXSWrm },
569 { X86::PMAXUBrr, X86::PMAXUBrm },
570 { X86::PMINSWrr, X86::PMINSWrm },
571 { X86::PMINUBrr, X86::PMINUBrm },
572 { X86::PMULHUWrr, X86::PMULHUWrm },
573 { X86::PMULHWrr, X86::PMULHWrm },
574 { X86::PMULLWrr, X86::PMULLWrm },
575 { X86::PMULUDQrr, X86::PMULUDQrm },
576 { X86::PORrr, X86::PORrm },
577 { X86::PSADBWrr, X86::PSADBWrm },
578 { X86::PSLLDrr, X86::PSLLDrm },
579 { X86::PSLLQrr, X86::PSLLQrm },
580 { X86::PSLLWrr, X86::PSLLWrm },
581 { X86::PSRADrr, X86::PSRADrm },
582 { X86::PSRAWrr, X86::PSRAWrm },
583 { X86::PSRLDrr, X86::PSRLDrm },
584 { X86::PSRLQrr, X86::PSRLQrm },
585 { X86::PSRLWrr, X86::PSRLWrm },
586 { X86::PSUBBrr, X86::PSUBBrm },
587 { X86::PSUBDrr, X86::PSUBDrm },
588 { X86::PSUBSBrr, X86::PSUBSBrm },
589 { X86::PSUBSWrr, X86::PSUBSWrm },
590 { X86::PSUBWrr, X86::PSUBWrm },
591 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
592 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
593 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
594 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
595 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
596 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
597 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
598 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
599 { X86::PXORrr, X86::PXORrm },
600 { X86::SBB32rr, X86::SBB32rm },
601 { X86::SBB64rr, X86::SBB64rm },
602 { X86::SHUFPDrri, X86::SHUFPDrmi },
603 { X86::SHUFPSrri, X86::SHUFPSrmi },
604 { X86::SUB16rr, X86::SUB16rm },
605 { X86::SUB32rr, X86::SUB32rm },
606 { X86::SUB64rr, X86::SUB64rm },
607 { X86::SUB8rr, X86::SUB8rm },
608 { X86::SUBPDrr, X86::SUBPDrm },
609 { X86::SUBPSrr, X86::SUBPSrm },
610 { X86::SUBSDrr, X86::SUBSDrm },
611 { X86::SUBSSrr, X86::SUBSSrm },
612 // FIXME: TEST*rr -> swapped operand of TEST*mr.
613 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
614 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
615 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
616 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
617 { X86::XOR16rr, X86::XOR16rm },
618 { X86::XOR32rr, X86::XOR32rm },
619 { X86::XOR64rr, X86::XOR64rm },
620 { X86::XOR8rr, X86::XOR8rm },
621 { X86::XORPDrr, X86::XORPDrm },
622 { X86::XORPSrr, X86::XORPSrm }
625 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
626 unsigned RegOp = OpTbl2[i][0];
627 unsigned MemOp = OpTbl2[i][1];
628 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp, MemOp)))
629 assert(false && "Duplicated entries?");
630 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
631 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
632 std::make_pair(RegOp, AuxInfo))))
633 AmbEntries.push_back(MemOp);
636 // Remove ambiguous entries.
637 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
640 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
642 unsigned& destReg) const {
643 switch (MI.getOpcode()) {
650 case X86::MOV16to16_:
651 case X86::MOV32to32_:
655 // FP Stack register class copies
656 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
657 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
658 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
660 case X86::FsMOVAPSrr:
661 case X86::FsMOVAPDrr:
664 case X86::MOVSS2PSrr:
665 case X86::MOVSD2PDrr:
666 case X86::MOVPS2SSrr:
667 case X86::MOVPD2SDrr:
668 case X86::MMX_MOVD64rr:
669 case X86::MMX_MOVQ64rr:
670 assert(MI.getNumOperands() >= 2 &&
671 MI.getOperand(0).isRegister() &&
672 MI.getOperand(1).isRegister() &&
673 "invalid register-register move instruction");
674 sourceReg = MI.getOperand(1).getReg();
675 destReg = MI.getOperand(0).getReg();
680 unsigned X86InstrInfo::isLoadFromStackSlot(MachineInstr *MI,
681 int &FrameIndex) const {
682 switch (MI->getOpcode()) {
695 case X86::MMX_MOVD64rm:
696 case X86::MMX_MOVQ64rm:
697 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
698 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
699 MI->getOperand(2).getImm() == 1 &&
700 MI->getOperand(3).getReg() == 0 &&
701 MI->getOperand(4).getImm() == 0) {
702 FrameIndex = MI->getOperand(1).getIndex();
703 return MI->getOperand(0).getReg();
710 unsigned X86InstrInfo::isStoreToStackSlot(MachineInstr *MI,
711 int &FrameIndex) const {
712 switch (MI->getOpcode()) {
725 case X86::MMX_MOVD64mr:
726 case X86::MMX_MOVQ64mr:
727 case X86::MMX_MOVNTQmr:
728 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
729 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
730 MI->getOperand(1).getImm() == 1 &&
731 MI->getOperand(2).getReg() == 0 &&
732 MI->getOperand(3).getImm() == 0) {
733 FrameIndex = MI->getOperand(0).getIndex();
734 return MI->getOperand(4).getReg();
742 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
744 static bool regIsPICBase(unsigned BaseReg, MachineRegisterInfo &MRI) {
745 bool isPICBase = false;
746 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
747 E = MRI.def_end(); I != E; ++I) {
748 MachineInstr *DefMI = I.getOperand().getParent();
749 if (DefMI->getOpcode() != X86::MOVPC32r)
751 assert(!isPICBase && "More than one PIC base?");
757 /// isGVStub - Return true if the GV requires an extra load to get the
759 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
760 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
763 bool X86InstrInfo::isReallyTriviallyReMaterializable(MachineInstr *MI) const {
764 switch (MI->getOpcode()) {
777 case X86::MMX_MOVD64rm:
778 case X86::MMX_MOVQ64rm: {
779 // Loads from constant pools are trivially rematerializable.
780 if (MI->getOperand(1).isReg() &&
781 MI->getOperand(2).isImm() &&
782 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
783 (MI->getOperand(4).isCPI() ||
784 (MI->getOperand(4).isGlobal() &&
785 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
786 unsigned BaseReg = MI->getOperand(1).getReg();
789 // Allow re-materialization of PIC load.
790 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
792 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
793 bool isPICBase = false;
794 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
795 E = MRI.def_end(); I != E; ++I) {
796 MachineInstr *DefMI = I.getOperand().getParent();
797 if (DefMI->getOpcode() != X86::MOVPC32r)
799 assert(!isPICBase && "More than one PIC base?");
809 if (MI->getOperand(1).isReg() &&
810 MI->getOperand(2).isImm() &&
811 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
812 !MI->getOperand(4).isReg()) {
813 // lea fi#, lea GV, etc. are all rematerializable.
814 unsigned BaseReg = MI->getOperand(1).getReg();
817 // Allow re-materialization of lea PICBase + x.
818 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
819 return regIsPICBase(BaseReg, MRI);
825 // All other instructions marked M_REMATERIALIZABLE are always trivially
830 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
831 MachineBasicBlock::iterator I,
833 const MachineInstr *Orig) const {
834 unsigned SubIdx = Orig->getOperand(0).isReg()
835 ? Orig->getOperand(0).getSubReg() : 0;
836 bool ChangeSubIdx = SubIdx != 0;
837 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
838 DestReg = RI.getSubReg(DestReg, SubIdx);
842 // MOV32r0 etc. are implemented with xor which clobbers condition code.
843 // Re-materialize them as movri instructions to avoid side effects.
844 switch (Orig->getOpcode()) {
846 BuildMI(MBB, I, get(X86::MOV8ri), DestReg).addImm(0);
849 BuildMI(MBB, I, get(X86::MOV16ri), DestReg).addImm(0);
852 BuildMI(MBB, I, get(X86::MOV32ri), DestReg).addImm(0);
855 BuildMI(MBB, I, get(X86::MOV64ri32), DestReg).addImm(0);
858 MachineInstr *MI = Orig->clone();
859 MI->getOperand(0).setReg(DestReg);
866 MachineInstr *NewMI = prior(I);
867 NewMI->getOperand(0).setSubReg(SubIdx);
871 /// isInvariantLoad - Return true if the specified instruction (which is marked
872 /// mayLoad) is loading from a location whose value is invariant across the
873 /// function. For example, loading a value from the constant pool or from
874 /// from the argument area of a function if it does not change. This should
875 /// only return true of *all* loads the instruction does are invariant (if it
876 /// does multiple loads).
877 bool X86InstrInfo::isInvariantLoad(MachineInstr *MI) const {
878 // This code cares about loads from three cases: constant pool entries,
879 // invariant argument slots, and global stubs. In order to handle these cases
880 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
881 // operand and base our analysis on it. This is safe because the address of
882 // none of these three cases is ever used as anything other than a load base
883 // and X86 doesn't have any instructions that load from multiple places.
885 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
886 const MachineOperand &MO = MI->getOperand(i);
887 // Loads from constant pools are trivially invariant.
892 return isGVStub(MO.getGlobal(), TM);
894 // If this is a load from an invariant stack slot, the load is a constant.
896 const MachineFrameInfo &MFI =
897 *MI->getParent()->getParent()->getFrameInfo();
898 int Idx = MO.getIndex();
899 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
903 // All other instances of these instructions are presumed to have other
908 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
909 /// is not marked dead.
910 static bool hasLiveCondCodeDef(MachineInstr *MI) {
911 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
912 MachineOperand &MO = MI->getOperand(i);
913 if (MO.isRegister() && MO.isDef() &&
914 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
921 /// convertToThreeAddress - This method must be implemented by targets that
922 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
923 /// may be able to convert a two-address instruction into a true
924 /// three-address instruction on demand. This allows the X86 target (for
925 /// example) to convert ADD and SHL instructions into LEA instructions if they
926 /// would require register copies due to two-addressness.
928 /// This method returns a null pointer if the transformation cannot be
929 /// performed, otherwise it returns the new instruction.
932 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
933 MachineBasicBlock::iterator &MBBI,
934 LiveVariables &LV) const {
935 MachineInstr *MI = MBBI;
936 // All instructions input are two-addr instructions. Get the known operands.
937 unsigned Dest = MI->getOperand(0).getReg();
938 unsigned Src = MI->getOperand(1).getReg();
940 MachineInstr *NewMI = NULL;
941 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
942 // we have better subtarget support, enable the 16-bit LEA generation here.
943 bool DisableLEA16 = true;
945 unsigned MIOpc = MI->getOpcode();
947 case X86::SHUFPSrri: {
948 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
949 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
951 unsigned A = MI->getOperand(0).getReg();
952 unsigned B = MI->getOperand(1).getReg();
953 unsigned C = MI->getOperand(2).getReg();
954 unsigned M = MI->getOperand(3).getImm();
955 if (B != C) return 0;
956 NewMI = BuildMI(get(X86::PSHUFDri), A).addReg(B).addImm(M);
960 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
961 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
962 // the flags produced by a shift yet, so this is safe.
963 unsigned Dest = MI->getOperand(0).getReg();
964 unsigned Src = MI->getOperand(1).getReg();
965 unsigned ShAmt = MI->getOperand(2).getImm();
966 if (ShAmt == 0 || ShAmt >= 4) return 0;
968 NewMI = BuildMI(get(X86::LEA64r), Dest)
969 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
973 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
974 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
975 // the flags produced by a shift yet, so this is safe.
976 unsigned Dest = MI->getOperand(0).getReg();
977 unsigned Src = MI->getOperand(1).getReg();
978 unsigned ShAmt = MI->getOperand(2).getImm();
979 if (ShAmt == 0 || ShAmt >= 4) return 0;
981 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
982 X86::LEA64_32r : X86::LEA32r;
983 NewMI = BuildMI(get(Opc), Dest)
984 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
988 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
989 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
990 // the flags produced by a shift yet, so this is safe.
991 unsigned Dest = MI->getOperand(0).getReg();
992 unsigned Src = MI->getOperand(1).getReg();
993 unsigned ShAmt = MI->getOperand(2).getImm();
994 if (ShAmt == 0 || ShAmt >= 4) return 0;
997 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
998 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
999 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1000 ? X86::LEA64_32r : X86::LEA32r;
1001 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1002 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1004 // Build and insert into an implicit UNDEF value. This is OK because
1005 // well be shifting and then extracting the lower 16-bits.
1006 MachineInstr *Undef = BuildMI(get(X86::IMPLICIT_DEF), leaInReg);
1009 BuildMI(get(X86::INSERT_SUBREG),leaInReg)
1010 .addReg(leaInReg).addReg(Src).addImm(X86::SUBREG_16BIT);
1012 NewMI = BuildMI(get(Opc), leaOutReg)
1013 .addReg(0).addImm(1 << ShAmt).addReg(leaInReg).addImm(0);
1016 BuildMI(get(X86::EXTRACT_SUBREG), Dest)
1017 .addReg(leaOutReg).addImm(X86::SUBREG_16BIT);
1018 Ext->copyKillDeadInfo(MI);
1020 MFI->insert(MBBI, Undef);
1021 MFI->insert(MBBI, Ins); // Insert the insert_subreg
1022 LV.instructionChanged(MI, NewMI); // Update live variables
1023 LV.addVirtualRegisterKilled(leaInReg, NewMI);
1024 MFI->insert(MBBI, NewMI); // Insert the new inst
1025 LV.addVirtualRegisterKilled(leaOutReg, Ext);
1026 MFI->insert(MBBI, Ext); // Insert the extract_subreg
1029 NewMI = BuildMI(get(X86::LEA16r), Dest)
1030 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
1035 // The following opcodes also sets the condition code register(s). Only
1036 // convert them to equivalent lea if the condition code register def's
1038 if (hasLiveCondCodeDef(MI))
1041 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1046 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1047 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1048 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1049 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, 1);
1053 case X86::INC64_16r:
1054 if (DisableLEA16) return 0;
1055 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1056 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, 1);
1060 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1061 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1062 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1063 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, -1);
1067 case X86::DEC64_16r:
1068 if (DisableLEA16) return 0;
1069 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1070 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, -1);
1073 case X86::ADD32rr: {
1074 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1075 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1076 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1077 NewMI = addRegReg(BuildMI(get(Opc), Dest), Src,
1078 MI->getOperand(2).getReg());
1082 if (DisableLEA16) return 0;
1083 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1084 NewMI = addRegReg(BuildMI(get(X86::LEA16r), Dest), Src,
1085 MI->getOperand(2).getReg());
1087 case X86::ADD64ri32:
1089 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1090 if (MI->getOperand(2).isImmediate())
1091 NewMI = addRegOffset(BuildMI(get(X86::LEA64r), Dest), Src,
1092 MI->getOperand(2).getImm());
1096 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1097 if (MI->getOperand(2).isImmediate()) {
1098 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1099 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src,
1100 MI->getOperand(2).getImm());
1105 if (DisableLEA16) return 0;
1106 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1107 if (MI->getOperand(2).isImmediate())
1108 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src,
1109 MI->getOperand(2).getImm());
1112 if (DisableLEA16) return 0;
1114 case X86::SHL64ri: {
1115 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImmediate() &&
1116 "Unknown shl instruction!");
1117 unsigned ShAmt = MI->getOperand(2).getImm();
1118 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1120 AM.Scale = 1 << ShAmt;
1122 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1123 : (MIOpc == X86::SHL32ri
1124 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1125 NewMI = addFullAddress(BuildMI(get(Opc), Dest), AM);
1133 if (!NewMI) return 0;
1135 NewMI->copyKillDeadInfo(MI);
1136 LV.instructionChanged(MI, NewMI); // Update live variables
1137 MFI->insert(MBBI, NewMI); // Insert the new inst
1141 /// commuteInstruction - We have a few instructions that must be hacked on to
1144 MachineInstr *X86InstrInfo::commuteInstruction(MachineInstr *MI) const {
1145 switch (MI->getOpcode()) {
1146 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1147 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1148 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1149 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1150 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1151 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1154 switch (MI->getOpcode()) {
1155 default: assert(0 && "Unreachable!");
1156 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1157 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1158 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1159 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1160 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1161 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1163 unsigned Amt = MI->getOperand(3).getImm();
1164 unsigned A = MI->getOperand(0).getReg();
1165 unsigned B = MI->getOperand(1).getReg();
1166 unsigned C = MI->getOperand(2).getReg();
1167 bool BisKill = MI->getOperand(1).isKill();
1168 bool CisKill = MI->getOperand(2).isKill();
1169 // If machine instrs are no longer in two-address forms, update
1170 // destination register as well.
1172 // Must be two address instruction!
1173 assert(MI->getDesc().getOperandConstraint(0, TOI::TIED_TO) &&
1174 "Expecting a two-address instruction!");
1178 return BuildMI(get(Opc), A).addReg(C, false, false, CisKill)
1179 .addReg(B, false, false, BisKill).addImm(Size-Amt);
1181 case X86::CMOVB16rr:
1182 case X86::CMOVB32rr:
1183 case X86::CMOVB64rr:
1184 case X86::CMOVAE16rr:
1185 case X86::CMOVAE32rr:
1186 case X86::CMOVAE64rr:
1187 case X86::CMOVE16rr:
1188 case X86::CMOVE32rr:
1189 case X86::CMOVE64rr:
1190 case X86::CMOVNE16rr:
1191 case X86::CMOVNE32rr:
1192 case X86::CMOVNE64rr:
1193 case X86::CMOVBE16rr:
1194 case X86::CMOVBE32rr:
1195 case X86::CMOVBE64rr:
1196 case X86::CMOVA16rr:
1197 case X86::CMOVA32rr:
1198 case X86::CMOVA64rr:
1199 case X86::CMOVL16rr:
1200 case X86::CMOVL32rr:
1201 case X86::CMOVL64rr:
1202 case X86::CMOVGE16rr:
1203 case X86::CMOVGE32rr:
1204 case X86::CMOVGE64rr:
1205 case X86::CMOVLE16rr:
1206 case X86::CMOVLE32rr:
1207 case X86::CMOVLE64rr:
1208 case X86::CMOVG16rr:
1209 case X86::CMOVG32rr:
1210 case X86::CMOVG64rr:
1211 case X86::CMOVS16rr:
1212 case X86::CMOVS32rr:
1213 case X86::CMOVS64rr:
1214 case X86::CMOVNS16rr:
1215 case X86::CMOVNS32rr:
1216 case X86::CMOVNS64rr:
1217 case X86::CMOVP16rr:
1218 case X86::CMOVP32rr:
1219 case X86::CMOVP64rr:
1220 case X86::CMOVNP16rr:
1221 case X86::CMOVNP32rr:
1222 case X86::CMOVNP64rr: {
1224 switch (MI->getOpcode()) {
1226 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1227 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1228 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1229 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1230 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1231 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1232 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1233 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1234 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1235 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1236 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1237 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1238 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1239 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1240 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1241 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1242 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1243 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1244 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1245 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1246 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1247 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1248 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1249 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1250 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1251 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1252 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1253 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1254 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1255 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1256 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1257 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1258 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1259 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1260 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1261 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1262 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1263 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1264 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1265 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1266 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1267 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1270 MI->setDesc(get(Opc));
1271 // Fallthrough intended.
1274 return TargetInstrInfoImpl::commuteInstruction(MI);
1278 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1280 default: return X86::COND_INVALID;
1281 case X86::JE: return X86::COND_E;
1282 case X86::JNE: return X86::COND_NE;
1283 case X86::JL: return X86::COND_L;
1284 case X86::JLE: return X86::COND_LE;
1285 case X86::JG: return X86::COND_G;
1286 case X86::JGE: return X86::COND_GE;
1287 case X86::JB: return X86::COND_B;
1288 case X86::JBE: return X86::COND_BE;
1289 case X86::JA: return X86::COND_A;
1290 case X86::JAE: return X86::COND_AE;
1291 case X86::JS: return X86::COND_S;
1292 case X86::JNS: return X86::COND_NS;
1293 case X86::JP: return X86::COND_P;
1294 case X86::JNP: return X86::COND_NP;
1295 case X86::JO: return X86::COND_O;
1296 case X86::JNO: return X86::COND_NO;
1300 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1302 default: assert(0 && "Illegal condition code!");
1303 case X86::COND_E: return X86::JE;
1304 case X86::COND_NE: return X86::JNE;
1305 case X86::COND_L: return X86::JL;
1306 case X86::COND_LE: return X86::JLE;
1307 case X86::COND_G: return X86::JG;
1308 case X86::COND_GE: return X86::JGE;
1309 case X86::COND_B: return X86::JB;
1310 case X86::COND_BE: return X86::JBE;
1311 case X86::COND_A: return X86::JA;
1312 case X86::COND_AE: return X86::JAE;
1313 case X86::COND_S: return X86::JS;
1314 case X86::COND_NS: return X86::JNS;
1315 case X86::COND_P: return X86::JP;
1316 case X86::COND_NP: return X86::JNP;
1317 case X86::COND_O: return X86::JO;
1318 case X86::COND_NO: return X86::JNO;
1322 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1323 /// e.g. turning COND_E to COND_NE.
1324 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1326 default: assert(0 && "Illegal condition code!");
1327 case X86::COND_E: return X86::COND_NE;
1328 case X86::COND_NE: return X86::COND_E;
1329 case X86::COND_L: return X86::COND_GE;
1330 case X86::COND_LE: return X86::COND_G;
1331 case X86::COND_G: return X86::COND_LE;
1332 case X86::COND_GE: return X86::COND_L;
1333 case X86::COND_B: return X86::COND_AE;
1334 case X86::COND_BE: return X86::COND_A;
1335 case X86::COND_A: return X86::COND_BE;
1336 case X86::COND_AE: return X86::COND_B;
1337 case X86::COND_S: return X86::COND_NS;
1338 case X86::COND_NS: return X86::COND_S;
1339 case X86::COND_P: return X86::COND_NP;
1340 case X86::COND_NP: return X86::COND_P;
1341 case X86::COND_O: return X86::COND_NO;
1342 case X86::COND_NO: return X86::COND_O;
1346 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1347 const TargetInstrDesc &TID = MI->getDesc();
1348 if (!TID.isTerminator()) return false;
1350 // Conditional branch is a special case.
1351 if (TID.isBranch() && !TID.isBarrier())
1353 if (!TID.isPredicable())
1355 return !isPredicated(MI);
1358 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1359 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1360 const X86InstrInfo &TII) {
1361 if (MI->getOpcode() == X86::FP_REG_KILL)
1363 return TII.isUnpredicatedTerminator(MI);
1366 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1367 MachineBasicBlock *&TBB,
1368 MachineBasicBlock *&FBB,
1369 std::vector<MachineOperand> &Cond) const {
1370 // If the block has no terminators, it just falls into the block after it.
1371 MachineBasicBlock::iterator I = MBB.end();
1372 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this))
1375 // Get the last instruction in the block.
1376 MachineInstr *LastInst = I;
1378 // If there is only one terminator instruction, process it.
1379 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this)) {
1380 if (!LastInst->getDesc().isBranch())
1383 // If the block ends with a branch there are 3 possibilities:
1384 // it's an unconditional, conditional, or indirect branch.
1386 if (LastInst->getOpcode() == X86::JMP) {
1387 TBB = LastInst->getOperand(0).getMBB();
1390 X86::CondCode BranchCode = GetCondFromBranchOpc(LastInst->getOpcode());
1391 if (BranchCode == X86::COND_INVALID)
1392 return true; // Can't handle indirect branch.
1394 // Otherwise, block ends with fall-through condbranch.
1395 TBB = LastInst->getOperand(0).getMBB();
1396 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1400 // Get the instruction before it if it's a terminator.
1401 MachineInstr *SecondLastInst = I;
1403 // If there are three terminators, we don't know what sort of block this is.
1404 if (SecondLastInst && I != MBB.begin() &&
1405 isBrAnalysisUnpredicatedTerminator(--I, *this))
1408 // If the block ends with X86::JMP and a conditional branch, handle it.
1409 X86::CondCode BranchCode = GetCondFromBranchOpc(SecondLastInst->getOpcode());
1410 if (BranchCode != X86::COND_INVALID && LastInst->getOpcode() == X86::JMP) {
1411 TBB = SecondLastInst->getOperand(0).getMBB();
1412 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1413 FBB = LastInst->getOperand(0).getMBB();
1417 // If the block ends with two X86::JMPs, handle it. The second one is not
1418 // executed, so remove it.
1419 if (SecondLastInst->getOpcode() == X86::JMP &&
1420 LastInst->getOpcode() == X86::JMP) {
1421 TBB = SecondLastInst->getOperand(0).getMBB();
1423 I->eraseFromParent();
1427 // Otherwise, can't handle this.
1431 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1432 MachineBasicBlock::iterator I = MBB.end();
1433 if (I == MBB.begin()) return 0;
1435 if (I->getOpcode() != X86::JMP &&
1436 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1439 // Remove the branch.
1440 I->eraseFromParent();
1444 if (I == MBB.begin()) return 1;
1446 if (GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1449 // Remove the branch.
1450 I->eraseFromParent();
1454 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
1455 MachineOperand &MO) {
1456 if (MO.isRegister())
1457 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
1458 false, false, MO.getSubReg());
1459 else if (MO.isImmediate())
1460 MIB = MIB.addImm(MO.getImm());
1461 else if (MO.isFrameIndex())
1462 MIB = MIB.addFrameIndex(MO.getIndex());
1463 else if (MO.isGlobalAddress())
1464 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
1465 else if (MO.isConstantPoolIndex())
1466 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
1467 else if (MO.isJumpTableIndex())
1468 MIB = MIB.addJumpTableIndex(MO.getIndex());
1469 else if (MO.isExternalSymbol())
1470 MIB = MIB.addExternalSymbol(MO.getSymbolName());
1472 assert(0 && "Unknown operand for X86InstrAddOperand!");
1478 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1479 MachineBasicBlock *FBB,
1480 const std::vector<MachineOperand> &Cond) const {
1481 // Shouldn't be a fall through.
1482 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1483 assert((Cond.size() == 1 || Cond.size() == 0) &&
1484 "X86 branch conditions have one component!");
1486 if (FBB == 0) { // One way branch.
1488 // Unconditional branch?
1489 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
1491 // Conditional branch.
1492 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
1493 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1498 // Two-way Conditional branch.
1499 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
1500 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1501 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
1505 void X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1506 MachineBasicBlock::iterator MI,
1507 unsigned DestReg, unsigned SrcReg,
1508 const TargetRegisterClass *DestRC,
1509 const TargetRegisterClass *SrcRC) const {
1510 if (DestRC == SrcRC) {
1512 if (DestRC == &X86::GR64RegClass) {
1514 } else if (DestRC == &X86::GR32RegClass) {
1516 } else if (DestRC == &X86::GR16RegClass) {
1518 } else if (DestRC == &X86::GR8RegClass) {
1520 } else if (DestRC == &X86::GR32_RegClass) {
1521 Opc = X86::MOV32_rr;
1522 } else if (DestRC == &X86::GR16_RegClass) {
1523 Opc = X86::MOV16_rr;
1524 } else if (DestRC == &X86::RFP32RegClass) {
1525 Opc = X86::MOV_Fp3232;
1526 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1527 Opc = X86::MOV_Fp6464;
1528 } else if (DestRC == &X86::RFP80RegClass) {
1529 Opc = X86::MOV_Fp8080;
1530 } else if (DestRC == &X86::FR32RegClass) {
1531 Opc = X86::FsMOVAPSrr;
1532 } else if (DestRC == &X86::FR64RegClass) {
1533 Opc = X86::FsMOVAPDrr;
1534 } else if (DestRC == &X86::VR128RegClass) {
1535 Opc = X86::MOVAPSrr;
1536 } else if (DestRC == &X86::VR64RegClass) {
1537 Opc = X86::MMX_MOVQ64rr;
1539 assert(0 && "Unknown regclass");
1542 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
1546 // Moving EFLAGS to / from another register requires a push and a pop.
1547 if (SrcRC == &X86::CCRRegClass) {
1548 assert(SrcReg == X86::EFLAGS);
1549 if (DestRC == &X86::GR64RegClass) {
1550 BuildMI(MBB, MI, get(X86::PUSHFQ));
1551 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
1553 } else if (DestRC == &X86::GR32RegClass) {
1554 BuildMI(MBB, MI, get(X86::PUSHFD));
1555 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
1558 } else if (DestRC == &X86::CCRRegClass) {
1559 assert(DestReg == X86::EFLAGS);
1560 if (SrcRC == &X86::GR64RegClass) {
1561 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
1562 BuildMI(MBB, MI, get(X86::POPFQ));
1564 } else if (SrcRC == &X86::GR32RegClass) {
1565 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
1566 BuildMI(MBB, MI, get(X86::POPFD));
1571 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1572 if (SrcRC == &X86::RSTRegClass) {
1573 // Copying from ST(0)/ST(1).
1574 assert((SrcReg == X86::ST0 || SrcReg == X86::ST1) &&
1575 "Can only copy from ST(0)/ST(1) right now");
1576 bool isST0 = SrcReg == X86::ST0;
1578 if (DestRC == &X86::RFP32RegClass)
1579 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1580 else if (DestRC == &X86::RFP64RegClass)
1581 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1583 assert(DestRC == &X86::RFP80RegClass);
1584 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1586 BuildMI(MBB, MI, get(Opc), DestReg);
1590 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1591 if (DestRC == &X86::RSTRegClass) {
1592 // Copying to ST(0). FIXME: handle ST(1) also
1593 assert(DestReg == X86::ST0 && "Can only copy to TOS right now");
1595 if (SrcRC == &X86::RFP32RegClass)
1596 Opc = X86::FpSET_ST0_32;
1597 else if (SrcRC == &X86::RFP64RegClass)
1598 Opc = X86::FpSET_ST0_64;
1600 assert(SrcRC == &X86::RFP80RegClass);
1601 Opc = X86::FpSET_ST0_80;
1603 BuildMI(MBB, MI, get(Opc)).addReg(SrcReg);
1607 assert(0 && "Not yet supported!");
1611 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1612 unsigned StackAlign) {
1614 if (RC == &X86::GR64RegClass) {
1616 } else if (RC == &X86::GR32RegClass) {
1618 } else if (RC == &X86::GR16RegClass) {
1620 } else if (RC == &X86::GR8RegClass) {
1622 } else if (RC == &X86::GR32_RegClass) {
1623 Opc = X86::MOV32_mr;
1624 } else if (RC == &X86::GR16_RegClass) {
1625 Opc = X86::MOV16_mr;
1626 } else if (RC == &X86::RFP80RegClass) {
1627 Opc = X86::ST_FpP80m; // pops
1628 } else if (RC == &X86::RFP64RegClass) {
1629 Opc = X86::ST_Fp64m;
1630 } else if (RC == &X86::RFP32RegClass) {
1631 Opc = X86::ST_Fp32m;
1632 } else if (RC == &X86::FR32RegClass) {
1634 } else if (RC == &X86::FR64RegClass) {
1636 } else if (RC == &X86::VR128RegClass) {
1637 // FIXME: Use movaps once we are capable of selectively
1638 // aligning functions that spill SSE registers on 16-byte boundaries.
1639 Opc = StackAlign >= 16 ? X86::MOVAPSmr : X86::MOVUPSmr;
1640 } else if (RC == &X86::VR64RegClass) {
1641 Opc = X86::MMX_MOVQ64mr;
1643 assert(0 && "Unknown regclass");
1650 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1651 MachineBasicBlock::iterator MI,
1652 unsigned SrcReg, bool isKill, int FrameIdx,
1653 const TargetRegisterClass *RC) const {
1654 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
1655 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
1656 .addReg(SrcReg, false, false, isKill);
1659 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1661 SmallVectorImpl<MachineOperand> &Addr,
1662 const TargetRegisterClass *RC,
1663 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1664 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
1665 MachineInstrBuilder MIB = BuildMI(get(Opc));
1666 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1667 MIB = X86InstrAddOperand(MIB, Addr[i]);
1668 MIB.addReg(SrcReg, false, false, isKill);
1669 NewMIs.push_back(MIB);
1672 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1673 unsigned StackAlign) {
1675 if (RC == &X86::GR64RegClass) {
1677 } else if (RC == &X86::GR32RegClass) {
1679 } else if (RC == &X86::GR16RegClass) {
1681 } else if (RC == &X86::GR8RegClass) {
1683 } else if (RC == &X86::GR32_RegClass) {
1684 Opc = X86::MOV32_rm;
1685 } else if (RC == &X86::GR16_RegClass) {
1686 Opc = X86::MOV16_rm;
1687 } else if (RC == &X86::RFP80RegClass) {
1688 Opc = X86::LD_Fp80m;
1689 } else if (RC == &X86::RFP64RegClass) {
1690 Opc = X86::LD_Fp64m;
1691 } else if (RC == &X86::RFP32RegClass) {
1692 Opc = X86::LD_Fp32m;
1693 } else if (RC == &X86::FR32RegClass) {
1695 } else if (RC == &X86::FR64RegClass) {
1697 } else if (RC == &X86::VR128RegClass) {
1698 // FIXME: Use movaps once we are capable of selectively
1699 // aligning functions that spill SSE registers on 16-byte boundaries.
1700 Opc = StackAlign >= 16 ? X86::MOVAPSrm : X86::MOVUPSrm;
1701 } else if (RC == &X86::VR64RegClass) {
1702 Opc = X86::MMX_MOVQ64rm;
1704 assert(0 && "Unknown regclass");
1711 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1712 MachineBasicBlock::iterator MI,
1713 unsigned DestReg, int FrameIdx,
1714 const TargetRegisterClass *RC) const{
1715 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
1716 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
1719 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1720 SmallVectorImpl<MachineOperand> &Addr,
1721 const TargetRegisterClass *RC,
1722 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1723 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
1724 MachineInstrBuilder MIB = BuildMI(get(Opc), DestReg);
1725 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1726 MIB = X86InstrAddOperand(MIB, Addr[i]);
1727 NewMIs.push_back(MIB);
1730 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1731 MachineBasicBlock::iterator MI,
1732 const std::vector<CalleeSavedInfo> &CSI) const {
1736 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1737 unsigned SlotSize = is64Bit ? 8 : 4;
1739 MachineFunction &MF = *MBB.getParent();
1740 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1741 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1743 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1744 for (unsigned i = CSI.size(); i != 0; --i) {
1745 unsigned Reg = CSI[i-1].getReg();
1746 // Add the callee-saved register as live-in. It's killed at the spill.
1748 BuildMI(MBB, MI, get(Opc)).addReg(Reg);
1753 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1754 MachineBasicBlock::iterator MI,
1755 const std::vector<CalleeSavedInfo> &CSI) const {
1759 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1761 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1762 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1763 unsigned Reg = CSI[i].getReg();
1764 BuildMI(MBB, MI, get(Opc), Reg);
1769 static MachineInstr *FuseTwoAddrInst(unsigned Opcode,
1770 SmallVector<MachineOperand,4> &MOs,
1771 MachineInstr *MI, const TargetInstrInfo &TII) {
1772 // Create the base instruction with the memory operand as the first part.
1773 MachineInstr *NewMI = new MachineInstr(TII.get(Opcode), true);
1774 MachineInstrBuilder MIB(NewMI);
1775 unsigned NumAddrOps = MOs.size();
1776 for (unsigned i = 0; i != NumAddrOps; ++i)
1777 MIB = X86InstrAddOperand(MIB, MOs[i]);
1778 if (NumAddrOps < 4) // FrameIndex only
1779 MIB.addImm(1).addReg(0).addImm(0);
1781 // Loop over the rest of the ri operands, converting them over.
1782 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1783 for (unsigned i = 0; i != NumOps; ++i) {
1784 MachineOperand &MO = MI->getOperand(i+2);
1785 MIB = X86InstrAddOperand(MIB, MO);
1787 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1788 MachineOperand &MO = MI->getOperand(i);
1789 MIB = X86InstrAddOperand(MIB, MO);
1794 static MachineInstr *FuseInst(unsigned Opcode, unsigned OpNo,
1795 SmallVector<MachineOperand,4> &MOs,
1796 MachineInstr *MI, const TargetInstrInfo &TII) {
1797 MachineInstr *NewMI = new MachineInstr(TII.get(Opcode), true);
1798 MachineInstrBuilder MIB(NewMI);
1800 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1801 MachineOperand &MO = MI->getOperand(i);
1803 assert(MO.isRegister() && "Expected to fold into reg operand!");
1804 unsigned NumAddrOps = MOs.size();
1805 for (unsigned i = 0; i != NumAddrOps; ++i)
1806 MIB = X86InstrAddOperand(MIB, MOs[i]);
1807 if (NumAddrOps < 4) // FrameIndex only
1808 MIB.addImm(1).addReg(0).addImm(0);
1810 MIB = X86InstrAddOperand(MIB, MO);
1816 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
1817 SmallVector<MachineOperand,4> &MOs,
1819 MachineInstrBuilder MIB = BuildMI(TII.get(Opcode));
1821 unsigned NumAddrOps = MOs.size();
1822 for (unsigned i = 0; i != NumAddrOps; ++i)
1823 MIB = X86InstrAddOperand(MIB, MOs[i]);
1824 if (NumAddrOps < 4) // FrameIndex only
1825 MIB.addImm(1).addReg(0).addImm(0);
1826 return MIB.addImm(0);
1830 X86InstrInfo::foldMemoryOperand(MachineInstr *MI, unsigned i,
1831 SmallVector<MachineOperand,4> &MOs) const {
1832 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
1833 bool isTwoAddrFold = false;
1834 unsigned NumOps = MI->getDesc().getNumOperands();
1835 bool isTwoAddr = NumOps > 1 &&
1836 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
1838 MachineInstr *NewMI = NULL;
1839 // Folding a memory location into the two-address part of a two-address
1840 // instruction is different than folding it other places. It requires
1841 // replacing the *two* registers with the memory location.
1842 if (isTwoAddr && NumOps >= 2 && i < 2 &&
1843 MI->getOperand(0).isRegister() &&
1844 MI->getOperand(1).isRegister() &&
1845 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
1846 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
1847 isTwoAddrFold = true;
1848 } else if (i == 0) { // If operand 0
1849 if (MI->getOpcode() == X86::MOV16r0)
1850 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
1851 else if (MI->getOpcode() == X86::MOV32r0)
1852 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
1853 else if (MI->getOpcode() == X86::MOV64r0)
1854 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
1855 else if (MI->getOpcode() == X86::MOV8r0)
1856 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
1858 NewMI->copyKillDeadInfo(MI);
1862 OpcodeTablePtr = &RegOp2MemOpTable0;
1863 } else if (i == 1) {
1864 OpcodeTablePtr = &RegOp2MemOpTable1;
1865 } else if (i == 2) {
1866 OpcodeTablePtr = &RegOp2MemOpTable2;
1869 // If table selected...
1870 if (OpcodeTablePtr) {
1871 // Find the Opcode to fuse
1872 DenseMap<unsigned*, unsigned>::iterator I =
1873 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
1874 if (I != OpcodeTablePtr->end()) {
1876 NewMI = FuseTwoAddrInst(I->second, MOs, MI, *this);
1878 NewMI = FuseInst(I->second, i, MOs, MI, *this);
1879 NewMI->copyKillDeadInfo(MI);
1885 if (PrintFailedFusing)
1886 cerr << "We failed to fuse operand " << i << *MI;
1891 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1893 SmallVectorImpl<unsigned> &Ops,
1894 int FrameIndex) const {
1895 // Check switch flag
1896 if (NoFusing) return NULL;
1898 const MachineFrameInfo *MFI = MF.getFrameInfo();
1899 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
1900 // FIXME: Move alignment requirement into tables?
1901 if (Alignment < 16) {
1902 switch (MI->getOpcode()) {
1904 // Not always safe to fold movsd into these instructions since their load
1905 // folding variants expects the address to be 16 byte aligned.
1906 case X86::FsANDNPDrr:
1907 case X86::FsANDNPSrr:
1908 case X86::FsANDPDrr:
1909 case X86::FsANDPSrr:
1912 case X86::FsXORPDrr:
1913 case X86::FsXORPSrr:
1918 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1919 unsigned NewOpc = 0;
1920 switch (MI->getOpcode()) {
1921 default: return NULL;
1922 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
1923 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
1924 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
1925 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
1927 // Change to CMPXXri r, 0 first.
1928 MI->setDesc(get(NewOpc));
1929 MI->getOperand(1).ChangeToImmediate(0);
1930 } else if (Ops.size() != 1)
1933 SmallVector<MachineOperand,4> MOs;
1934 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
1935 return foldMemoryOperand(MI, Ops[0], MOs);
1938 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1940 SmallVectorImpl<unsigned> &Ops,
1941 MachineInstr *LoadMI) const {
1942 // Check switch flag
1943 if (NoFusing) return NULL;
1945 unsigned Alignment = 0;
1946 for (unsigned i = 0, e = LoadMI->getNumMemOperands(); i != e; ++i) {
1947 const MachineMemOperand &MRO = LoadMI->getMemOperand(i);
1948 unsigned Align = MRO.getAlignment();
1949 if (Align > Alignment)
1953 // FIXME: Move alignment requirement into tables?
1954 if (Alignment < 16) {
1955 switch (MI->getOpcode()) {
1957 // Not always safe to fold movsd into these instructions since their load
1958 // folding variants expects the address to be 16 byte aligned.
1959 case X86::FsANDNPDrr:
1960 case X86::FsANDNPSrr:
1961 case X86::FsANDPDrr:
1962 case X86::FsANDPSrr:
1965 case X86::FsXORPDrr:
1966 case X86::FsXORPSrr:
1971 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1972 unsigned NewOpc = 0;
1973 switch (MI->getOpcode()) {
1974 default: return NULL;
1975 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
1976 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
1977 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
1978 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
1980 // Change to CMPXXri r, 0 first.
1981 MI->setDesc(get(NewOpc));
1982 MI->getOperand(1).ChangeToImmediate(0);
1983 } else if (Ops.size() != 1)
1986 SmallVector<MachineOperand,4> MOs;
1987 unsigned NumOps = LoadMI->getDesc().getNumOperands();
1988 for (unsigned i = NumOps - 4; i != NumOps; ++i)
1989 MOs.push_back(LoadMI->getOperand(i));
1990 return foldMemoryOperand(MI, Ops[0], MOs);
1994 bool X86InstrInfo::canFoldMemoryOperand(MachineInstr *MI,
1995 SmallVectorImpl<unsigned> &Ops) const {
1996 // Check switch flag
1997 if (NoFusing) return 0;
1999 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2000 switch (MI->getOpcode()) {
2001 default: return false;
2010 if (Ops.size() != 1)
2013 unsigned OpNum = Ops[0];
2014 unsigned Opc = MI->getOpcode();
2015 unsigned NumOps = MI->getDesc().getNumOperands();
2016 bool isTwoAddr = NumOps > 1 &&
2017 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2019 // Folding a memory location into the two-address part of a two-address
2020 // instruction is different than folding it other places. It requires
2021 // replacing the *two* registers with the memory location.
2022 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2023 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2024 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2025 } else if (OpNum == 0) { // If operand 0
2034 OpcodeTablePtr = &RegOp2MemOpTable0;
2035 } else if (OpNum == 1) {
2036 OpcodeTablePtr = &RegOp2MemOpTable1;
2037 } else if (OpNum == 2) {
2038 OpcodeTablePtr = &RegOp2MemOpTable2;
2041 if (OpcodeTablePtr) {
2042 // Find the Opcode to fuse
2043 DenseMap<unsigned*, unsigned>::iterator I =
2044 OpcodeTablePtr->find((unsigned*)Opc);
2045 if (I != OpcodeTablePtr->end())
2051 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2052 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2053 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2054 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2055 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2056 if (I == MemOp2RegOpTable.end())
2058 unsigned Opc = I->second.first;
2059 unsigned Index = I->second.second & 0xf;
2060 bool FoldedLoad = I->second.second & (1 << 4);
2061 bool FoldedStore = I->second.second & (1 << 5);
2062 if (UnfoldLoad && !FoldedLoad)
2064 UnfoldLoad &= FoldedLoad;
2065 if (UnfoldStore && !FoldedStore)
2067 UnfoldStore &= FoldedStore;
2069 const TargetInstrDesc &TID = get(Opc);
2070 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2071 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2072 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2073 SmallVector<MachineOperand,4> AddrOps;
2074 SmallVector<MachineOperand,2> BeforeOps;
2075 SmallVector<MachineOperand,2> AfterOps;
2076 SmallVector<MachineOperand,4> ImpOps;
2077 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2078 MachineOperand &Op = MI->getOperand(i);
2079 if (i >= Index && i < Index+4)
2080 AddrOps.push_back(Op);
2081 else if (Op.isRegister() && Op.isImplicit())
2082 ImpOps.push_back(Op);
2084 BeforeOps.push_back(Op);
2086 AfterOps.push_back(Op);
2089 // Emit the load instruction.
2091 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2093 // Address operands cannot be marked isKill.
2094 for (unsigned i = 1; i != 5; ++i) {
2095 MachineOperand &MO = NewMIs[0]->getOperand(i);
2096 if (MO.isRegister())
2097 MO.setIsKill(false);
2102 // Emit the data processing instruction.
2103 MachineInstr *DataMI = new MachineInstr(TID, true);
2104 MachineInstrBuilder MIB(DataMI);
2107 MIB.addReg(Reg, true);
2108 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2109 MIB = X86InstrAddOperand(MIB, BeforeOps[i]);
2112 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2113 MIB = X86InstrAddOperand(MIB, AfterOps[i]);
2114 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2115 MachineOperand &MO = ImpOps[i];
2116 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2118 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2119 unsigned NewOpc = 0;
2120 switch (DataMI->getOpcode()) {
2122 case X86::CMP64ri32:
2126 MachineOperand &MO0 = DataMI->getOperand(0);
2127 MachineOperand &MO1 = DataMI->getOperand(1);
2128 if (MO1.getImm() == 0) {
2129 switch (DataMI->getOpcode()) {
2131 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2132 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2133 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2134 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2136 DataMI->setDesc(get(NewOpc));
2137 MO1.ChangeToRegister(MO0.getReg(), false);
2141 NewMIs.push_back(DataMI);
2143 // Emit the store instruction.
2145 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2146 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2147 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2148 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2155 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2156 SmallVectorImpl<SDNode*> &NewNodes) const {
2157 if (!N->isTargetOpcode())
2160 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2161 MemOp2RegOpTable.find((unsigned*)N->getTargetOpcode());
2162 if (I == MemOp2RegOpTable.end())
2164 unsigned Opc = I->second.first;
2165 unsigned Index = I->second.second & 0xf;
2166 bool FoldedLoad = I->second.second & (1 << 4);
2167 bool FoldedStore = I->second.second & (1 << 5);
2168 const TargetInstrDesc &TID = get(Opc);
2169 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2170 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2171 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2172 std::vector<SDOperand> AddrOps;
2173 std::vector<SDOperand> BeforeOps;
2174 std::vector<SDOperand> AfterOps;
2175 unsigned NumOps = N->getNumOperands();
2176 for (unsigned i = 0; i != NumOps-1; ++i) {
2177 SDOperand Op = N->getOperand(i);
2178 if (i >= Index && i < Index+4)
2179 AddrOps.push_back(Op);
2181 BeforeOps.push_back(Op);
2183 AfterOps.push_back(Op);
2185 SDOperand Chain = N->getOperand(NumOps-1);
2186 AddrOps.push_back(Chain);
2188 // Emit the load instruction.
2191 MVT::ValueType VT = *RC->vt_begin();
2192 Load = DAG.getTargetNode(getLoadRegOpcode(RC, RI.getStackAlignment()), VT,
2193 MVT::Other, &AddrOps[0], AddrOps.size());
2194 NewNodes.push_back(Load);
2197 // Emit the data processing instruction.
2198 std::vector<MVT::ValueType> VTs;
2199 const TargetRegisterClass *DstRC = 0;
2200 if (TID.getNumDefs() > 0) {
2201 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2202 DstRC = DstTOI.isLookupPtrRegClass()
2203 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2204 VTs.push_back(*DstRC->vt_begin());
2206 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2207 MVT::ValueType VT = N->getValueType(i);
2208 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2212 BeforeOps.push_back(SDOperand(Load, 0));
2213 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2214 SDNode *NewNode= DAG.getTargetNode(Opc, VTs, &BeforeOps[0], BeforeOps.size());
2215 NewNodes.push_back(NewNode);
2217 // Emit the store instruction.
2220 AddrOps.push_back(SDOperand(NewNode, 0));
2221 AddrOps.push_back(Chain);
2222 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, RI.getStackAlignment()),
2223 MVT::Other, &AddrOps[0], AddrOps.size());
2224 NewNodes.push_back(Store);
2230 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2231 bool UnfoldLoad, bool UnfoldStore) const {
2232 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2233 MemOp2RegOpTable.find((unsigned*)Opc);
2234 if (I == MemOp2RegOpTable.end())
2236 bool FoldedLoad = I->second.second & (1 << 4);
2237 bool FoldedStore = I->second.second & (1 << 5);
2238 if (UnfoldLoad && !FoldedLoad)
2240 if (UnfoldStore && !FoldedStore)
2242 return I->second.first;
2245 bool X86InstrInfo::BlockHasNoFallThrough(MachineBasicBlock &MBB) const {
2246 if (MBB.empty()) return false;
2248 switch (MBB.back().getOpcode()) {
2249 case X86::TCRETURNri:
2250 case X86::TCRETURNdi:
2251 case X86::RET: // Return.
2256 case X86::JMP: // Uncond branch.
2257 case X86::JMP32r: // Indirect branch.
2258 case X86::JMP64r: // Indirect branch (64-bit).
2259 case X86::JMP32m: // Indirect branch through mem.
2260 case X86::JMP64m: // Indirect branch through mem (64-bit).
2262 default: return false;
2267 ReverseBranchCondition(std::vector<MachineOperand> &Cond) const {
2268 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2269 Cond[0].setImm(GetOppositeBranchCondition((X86::CondCode)Cond[0].getImm()));
2273 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
2274 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
2275 if (Subtarget->is64Bit())
2276 return &X86::GR64RegClass;
2278 return &X86::GR32RegClass;
2281 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2282 switch (Desc->TSFlags & X86II::ImmMask) {
2283 case X86II::Imm8: return 1;
2284 case X86II::Imm16: return 2;
2285 case X86II::Imm32: return 4;
2286 case X86II::Imm64: return 8;
2287 default: assert(0 && "Immediate size not set!");
2292 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2293 /// e.g. r8, xmm8, etc.
2294 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2295 if (!MO.isRegister()) return false;
2296 switch (MO.getReg()) {
2298 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2299 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2300 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2301 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2302 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2303 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2304 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2305 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2306 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2307 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2314 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2315 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2316 /// size, and 3) use of X86-64 extended registers.
2317 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2319 const TargetInstrDesc &Desc = MI.getDesc();
2321 // Pseudo instructions do not need REX prefix byte.
2322 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2324 if (Desc.TSFlags & X86II::REX_W)
2327 unsigned NumOps = Desc.getNumOperands();
2329 bool isTwoAddr = NumOps > 1 &&
2330 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2332 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2333 unsigned i = isTwoAddr ? 1 : 0;
2334 for (unsigned e = NumOps; i != e; ++i) {
2335 const MachineOperand& MO = MI.getOperand(i);
2336 if (MO.isRegister()) {
2337 unsigned Reg = MO.getReg();
2338 if (isX86_64NonExtLowByteReg(Reg))
2343 switch (Desc.TSFlags & X86II::FormMask) {
2344 case X86II::MRMInitReg:
2345 if (isX86_64ExtendedReg(MI.getOperand(0)))
2346 REX |= (1 << 0) | (1 << 2);
2348 case X86II::MRMSrcReg: {
2349 if (isX86_64ExtendedReg(MI.getOperand(0)))
2351 i = isTwoAddr ? 2 : 1;
2352 for (unsigned e = NumOps; i != e; ++i) {
2353 const MachineOperand& MO = MI.getOperand(i);
2354 if (isX86_64ExtendedReg(MO))
2359 case X86II::MRMSrcMem: {
2360 if (isX86_64ExtendedReg(MI.getOperand(0)))
2363 i = isTwoAddr ? 2 : 1;
2364 for (; i != NumOps; ++i) {
2365 const MachineOperand& MO = MI.getOperand(i);
2366 if (MO.isRegister()) {
2367 if (isX86_64ExtendedReg(MO))
2374 case X86II::MRM0m: case X86II::MRM1m:
2375 case X86II::MRM2m: case X86II::MRM3m:
2376 case X86II::MRM4m: case X86II::MRM5m:
2377 case X86II::MRM6m: case X86II::MRM7m:
2378 case X86II::MRMDestMem: {
2379 unsigned e = isTwoAddr ? 5 : 4;
2380 i = isTwoAddr ? 1 : 0;
2381 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2384 for (; i != e; ++i) {
2385 const MachineOperand& MO = MI.getOperand(i);
2386 if (MO.isRegister()) {
2387 if (isX86_64ExtendedReg(MO))
2395 if (isX86_64ExtendedReg(MI.getOperand(0)))
2397 i = isTwoAddr ? 2 : 1;
2398 for (unsigned e = NumOps; i != e; ++i) {
2399 const MachineOperand& MO = MI.getOperand(i);
2400 if (isX86_64ExtendedReg(MO))
2410 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2411 /// relative block address instruction
2413 static unsigned sizePCRelativeBlockAddress() {
2417 /// sizeGlobalAddress - Give the size of the emission of this global address
2419 static unsigned sizeGlobalAddress(bool dword) {
2420 return dword ? 8 : 4;
2423 /// sizeConstPoolAddress - Give the size of the emission of this constant
2426 static unsigned sizeConstPoolAddress(bool dword) {
2427 return dword ? 8 : 4;
2430 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2433 static unsigned sizeExternalSymbolAddress(bool dword) {
2434 return dword ? 8 : 4;
2437 /// sizeJumpTableAddress - Give the size of the emission of this jump
2440 static unsigned sizeJumpTableAddress(bool dword) {
2441 return dword ? 8 : 4;
2444 static unsigned sizeConstant(unsigned Size) {
2448 static unsigned sizeRegModRMByte(){
2452 static unsigned sizeSIBByte(){
2456 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2457 unsigned FinalSize = 0;
2458 // If this is a simple integer displacement that doesn't require a relocation.
2460 FinalSize += sizeConstant(4);
2464 // Otherwise, this is something that requires a relocation.
2465 if (RelocOp->isGlobalAddress()) {
2466 FinalSize += sizeGlobalAddress(false);
2467 } else if (RelocOp->isConstantPoolIndex()) {
2468 FinalSize += sizeConstPoolAddress(false);
2469 } else if (RelocOp->isJumpTableIndex()) {
2470 FinalSize += sizeJumpTableAddress(false);
2472 assert(0 && "Unknown value to relocate!");
2477 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2478 bool IsPIC, bool Is64BitMode) {
2479 const MachineOperand &Op3 = MI.getOperand(Op+3);
2481 const MachineOperand *DispForReloc = 0;
2482 unsigned FinalSize = 0;
2484 // Figure out what sort of displacement we have to handle here.
2485 if (Op3.isGlobalAddress()) {
2486 DispForReloc = &Op3;
2487 } else if (Op3.isConstantPoolIndex()) {
2488 if (Is64BitMode || IsPIC) {
2489 DispForReloc = &Op3;
2493 } else if (Op3.isJumpTableIndex()) {
2494 if (Is64BitMode || IsPIC) {
2495 DispForReloc = &Op3;
2503 const MachineOperand &Base = MI.getOperand(Op);
2504 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2506 unsigned BaseReg = Base.getReg();
2508 // Is a SIB byte needed?
2509 if (IndexReg.getReg() == 0 &&
2510 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2511 if (BaseReg == 0) { // Just a displacement?
2512 // Emit special case [disp32] encoding
2514 FinalSize += getDisplacementFieldSize(DispForReloc);
2516 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2517 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2518 // Emit simple indirect register encoding... [EAX] f.e.
2520 // Be pessimistic and assume it's a disp32, not a disp8
2522 // Emit the most general non-SIB encoding: [REG+disp32]
2524 FinalSize += getDisplacementFieldSize(DispForReloc);
2528 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2529 assert(IndexReg.getReg() != X86::ESP &&
2530 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2532 bool ForceDisp32 = false;
2533 if (BaseReg == 0 || DispForReloc) {
2534 // Emit the normal disp32 encoding.
2541 FinalSize += sizeSIBByte();
2543 // Do we need to output a displacement?
2544 if (DispVal != 0 || ForceDisp32) {
2545 FinalSize += getDisplacementFieldSize(DispForReloc);
2552 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2553 const TargetInstrDesc *Desc,
2554 bool IsPIC, bool Is64BitMode) {
2556 unsigned Opcode = Desc->Opcode;
2557 unsigned FinalSize = 0;
2559 // Emit the lock opcode prefix as needed.
2560 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2562 // Emit the repeat opcode prefix as needed.
2563 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2565 // Emit the operand size opcode prefix as needed.
2566 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2568 // Emit the address size opcode prefix as needed.
2569 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2571 bool Need0FPrefix = false;
2572 switch (Desc->TSFlags & X86II::Op0Mask) {
2573 case X86II::TB: // Two-byte opcode prefix
2574 case X86II::T8: // 0F 38
2575 case X86II::TA: // 0F 3A
2576 Need0FPrefix = true;
2578 case X86II::REP: break; // already handled.
2579 case X86II::XS: // F3 0F
2581 Need0FPrefix = true;
2583 case X86II::XD: // F2 0F
2585 Need0FPrefix = true;
2587 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2588 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2590 break; // Two-byte opcode prefix
2591 default: assert(0 && "Invalid prefix!");
2592 case 0: break; // No prefix!
2597 unsigned REX = X86InstrInfo::determineREX(MI);
2602 // 0x0F escape code must be emitted just before the opcode.
2606 switch (Desc->TSFlags & X86II::Op0Mask) {
2607 case X86II::T8: // 0F 38
2610 case X86II::TA: // 0F 3A
2615 // If this is a two-address instruction, skip one of the register operands.
2616 unsigned NumOps = Desc->getNumOperands();
2618 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2621 switch (Desc->TSFlags & X86II::FormMask) {
2622 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
2624 // Remember the current PC offset, this is the PIC relocation
2629 case TargetInstrInfo::INLINEASM: {
2630 const MachineFunction *MF = MI.getParent()->getParent();
2631 const char *AsmStr = MI.getOperand(0).getSymbolName();
2632 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
2633 FinalSize += AI->getInlineAsmLength(AsmStr);
2636 case TargetInstrInfo::LABEL:
2638 case TargetInstrInfo::IMPLICIT_DEF:
2639 case TargetInstrInfo::DECLARE:
2640 case X86::DWARF_LOC:
2641 case X86::FP_REG_KILL:
2643 case X86::MOVPC32r: {
2644 // This emits the "call" portion of this pseudo instruction.
2646 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2655 if (CurOp != NumOps) {
2656 const MachineOperand &MO = MI.getOperand(CurOp++);
2657 if (MO.isMachineBasicBlock()) {
2658 FinalSize += sizePCRelativeBlockAddress();
2659 } else if (MO.isGlobalAddress()) {
2660 FinalSize += sizeGlobalAddress(false);
2661 } else if (MO.isExternalSymbol()) {
2662 FinalSize += sizeExternalSymbolAddress(false);
2663 } else if (MO.isImmediate()) {
2664 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2666 assert(0 && "Unknown RawFrm operand!");
2671 case X86II::AddRegFrm:
2675 if (CurOp != NumOps) {
2676 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2677 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2678 if (MO1.isImmediate())
2679 FinalSize += sizeConstant(Size);
2682 if (Opcode == X86::MOV64ri)
2684 if (MO1.isGlobalAddress()) {
2685 FinalSize += sizeGlobalAddress(dword);
2686 } else if (MO1.isExternalSymbol())
2687 FinalSize += sizeExternalSymbolAddress(dword);
2688 else if (MO1.isConstantPoolIndex())
2689 FinalSize += sizeConstPoolAddress(dword);
2690 else if (MO1.isJumpTableIndex())
2691 FinalSize += sizeJumpTableAddress(dword);
2696 case X86II::MRMDestReg: {
2698 FinalSize += sizeRegModRMByte();
2700 if (CurOp != NumOps) {
2702 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2706 case X86II::MRMDestMem: {
2708 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2710 if (CurOp != NumOps) {
2712 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2717 case X86II::MRMSrcReg:
2719 FinalSize += sizeRegModRMByte();
2721 if (CurOp != NumOps) {
2723 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2727 case X86II::MRMSrcMem: {
2730 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
2732 if (CurOp != NumOps) {
2734 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2739 case X86II::MRM0r: case X86II::MRM1r:
2740 case X86II::MRM2r: case X86II::MRM3r:
2741 case X86II::MRM4r: case X86II::MRM5r:
2742 case X86II::MRM6r: case X86II::MRM7r:
2745 FinalSize += sizeRegModRMByte();
2747 if (CurOp != NumOps) {
2748 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2749 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2750 if (MO1.isImmediate())
2751 FinalSize += sizeConstant(Size);
2754 if (Opcode == X86::MOV64ri32)
2756 if (MO1.isGlobalAddress()) {
2757 FinalSize += sizeGlobalAddress(dword);
2758 } else if (MO1.isExternalSymbol())
2759 FinalSize += sizeExternalSymbolAddress(dword);
2760 else if (MO1.isConstantPoolIndex())
2761 FinalSize += sizeConstPoolAddress(dword);
2762 else if (MO1.isJumpTableIndex())
2763 FinalSize += sizeJumpTableAddress(dword);
2768 case X86II::MRM0m: case X86II::MRM1m:
2769 case X86II::MRM2m: case X86II::MRM3m:
2770 case X86II::MRM4m: case X86II::MRM5m:
2771 case X86II::MRM6m: case X86II::MRM7m: {
2774 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2777 if (CurOp != NumOps) {
2778 const MachineOperand &MO = MI.getOperand(CurOp++);
2779 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2780 if (MO.isImmediate())
2781 FinalSize += sizeConstant(Size);
2784 if (Opcode == X86::MOV64mi32)
2786 if (MO.isGlobalAddress()) {
2787 FinalSize += sizeGlobalAddress(dword);
2788 } else if (MO.isExternalSymbol())
2789 FinalSize += sizeExternalSymbolAddress(dword);
2790 else if (MO.isConstantPoolIndex())
2791 FinalSize += sizeConstPoolAddress(dword);
2792 else if (MO.isJumpTableIndex())
2793 FinalSize += sizeJumpTableAddress(dword);
2799 case X86II::MRMInitReg:
2801 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
2802 FinalSize += sizeRegModRMByte();
2807 if (!Desc->isVariadic() && CurOp != NumOps) {
2808 cerr << "Cannot determine size: ";
2819 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
2820 const TargetInstrDesc &Desc = MI->getDesc();
2821 bool IsPIC = (TM.getRelocationModel() == Reloc::PIC_);
2822 bool Is64BitMode = ((X86Subtarget*)TM.getSubtargetImpl())->is64Bit();
2823 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
2824 if (Desc.getOpcode() == X86::MOVPC32r) {
2825 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);