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::PMULDQrr, X86::PMULDQrm },
573 { X86::PMULDQrr_int, X86::PMULDQrm_int },
574 { X86::PMULHUWrr, X86::PMULHUWrm },
575 { X86::PMULHWrr, X86::PMULHWrm },
576 { X86::PMULLDrr, X86::PMULLDrm },
577 { X86::PMULLDrr_int, X86::PMULLDrm_int },
578 { X86::PMULLWrr, X86::PMULLWrm },
579 { X86::PMULUDQrr, X86::PMULUDQrm },
580 { X86::PORrr, X86::PORrm },
581 { X86::PSADBWrr, X86::PSADBWrm },
582 { X86::PSLLDrr, X86::PSLLDrm },
583 { X86::PSLLQrr, X86::PSLLQrm },
584 { X86::PSLLWrr, X86::PSLLWrm },
585 { X86::PSRADrr, X86::PSRADrm },
586 { X86::PSRAWrr, X86::PSRAWrm },
587 { X86::PSRLDrr, X86::PSRLDrm },
588 { X86::PSRLQrr, X86::PSRLQrm },
589 { X86::PSRLWrr, X86::PSRLWrm },
590 { X86::PSUBBrr, X86::PSUBBrm },
591 { X86::PSUBDrr, X86::PSUBDrm },
592 { X86::PSUBSBrr, X86::PSUBSBrm },
593 { X86::PSUBSWrr, X86::PSUBSWrm },
594 { X86::PSUBWrr, X86::PSUBWrm },
595 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
596 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
597 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
598 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
599 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
600 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
601 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
602 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
603 { X86::PXORrr, X86::PXORrm },
604 { X86::SBB32rr, X86::SBB32rm },
605 { X86::SBB64rr, X86::SBB64rm },
606 { X86::SHUFPDrri, X86::SHUFPDrmi },
607 { X86::SHUFPSrri, X86::SHUFPSrmi },
608 { X86::SUB16rr, X86::SUB16rm },
609 { X86::SUB32rr, X86::SUB32rm },
610 { X86::SUB64rr, X86::SUB64rm },
611 { X86::SUB8rr, X86::SUB8rm },
612 { X86::SUBPDrr, X86::SUBPDrm },
613 { X86::SUBPSrr, X86::SUBPSrm },
614 { X86::SUBSDrr, X86::SUBSDrm },
615 { X86::SUBSSrr, X86::SUBSSrm },
616 // FIXME: TEST*rr -> swapped operand of TEST*mr.
617 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
618 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
619 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
620 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
621 { X86::XOR16rr, X86::XOR16rm },
622 { X86::XOR32rr, X86::XOR32rm },
623 { X86::XOR64rr, X86::XOR64rm },
624 { X86::XOR8rr, X86::XOR8rm },
625 { X86::XORPDrr, X86::XORPDrm },
626 { X86::XORPSrr, X86::XORPSrm }
629 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
630 unsigned RegOp = OpTbl2[i][0];
631 unsigned MemOp = OpTbl2[i][1];
632 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp, MemOp)))
633 assert(false && "Duplicated entries?");
634 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
635 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
636 std::make_pair(RegOp, AuxInfo))))
637 AmbEntries.push_back(MemOp);
640 // Remove ambiguous entries.
641 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
644 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
646 unsigned& destReg) const {
647 switch (MI.getOpcode()) {
654 case X86::MOV16to16_:
655 case X86::MOV32to32_:
659 // FP Stack register class copies
660 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
661 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
662 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
664 case X86::FsMOVAPSrr:
665 case X86::FsMOVAPDrr:
668 case X86::MOVSS2PSrr:
669 case X86::MOVSD2PDrr:
670 case X86::MOVPS2SSrr:
671 case X86::MOVPD2SDrr:
672 case X86::MMX_MOVD64rr:
673 case X86::MMX_MOVQ64rr:
674 assert(MI.getNumOperands() >= 2 &&
675 MI.getOperand(0).isRegister() &&
676 MI.getOperand(1).isRegister() &&
677 "invalid register-register move instruction");
678 sourceReg = MI.getOperand(1).getReg();
679 destReg = MI.getOperand(0).getReg();
684 unsigned X86InstrInfo::isLoadFromStackSlot(MachineInstr *MI,
685 int &FrameIndex) const {
686 switch (MI->getOpcode()) {
699 case X86::MMX_MOVD64rm:
700 case X86::MMX_MOVQ64rm:
701 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
702 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
703 MI->getOperand(2).getImm() == 1 &&
704 MI->getOperand(3).getReg() == 0 &&
705 MI->getOperand(4).getImm() == 0) {
706 FrameIndex = MI->getOperand(1).getIndex();
707 return MI->getOperand(0).getReg();
714 unsigned X86InstrInfo::isStoreToStackSlot(MachineInstr *MI,
715 int &FrameIndex) const {
716 switch (MI->getOpcode()) {
729 case X86::MMX_MOVD64mr:
730 case X86::MMX_MOVQ64mr:
731 case X86::MMX_MOVNTQmr:
732 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
733 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
734 MI->getOperand(1).getImm() == 1 &&
735 MI->getOperand(2).getReg() == 0 &&
736 MI->getOperand(3).getImm() == 0) {
737 FrameIndex = MI->getOperand(0).getIndex();
738 return MI->getOperand(4).getReg();
746 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
748 static bool regIsPICBase(unsigned BaseReg, MachineRegisterInfo &MRI) {
749 bool isPICBase = false;
750 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
751 E = MRI.def_end(); I != E; ++I) {
752 MachineInstr *DefMI = I.getOperand().getParent();
753 if (DefMI->getOpcode() != X86::MOVPC32r)
755 assert(!isPICBase && "More than one PIC base?");
761 /// isGVStub - Return true if the GV requires an extra load to get the
763 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
764 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
768 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI) const {
769 switch (MI->getOpcode()) {
782 case X86::MMX_MOVD64rm:
783 case X86::MMX_MOVQ64rm: {
784 // Loads from constant pools are trivially rematerializable.
785 if (MI->getOperand(1).isReg() &&
786 MI->getOperand(2).isImm() &&
787 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
788 (MI->getOperand(4).isCPI() ||
789 (MI->getOperand(4).isGlobal() &&
790 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
791 unsigned BaseReg = MI->getOperand(1).getReg();
794 // Allow re-materialization of PIC load.
795 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
797 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
798 bool isPICBase = false;
799 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
800 E = MRI.def_end(); I != E; ++I) {
801 MachineInstr *DefMI = I.getOperand().getParent();
802 if (DefMI->getOpcode() != X86::MOVPC32r)
804 assert(!isPICBase && "More than one PIC base?");
814 if (MI->getOperand(1).isReg() &&
815 MI->getOperand(2).isImm() &&
816 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
817 !MI->getOperand(4).isReg()) {
818 // lea fi#, lea GV, etc. are all rematerializable.
819 unsigned BaseReg = MI->getOperand(1).getReg();
822 // Allow re-materialization of lea PICBase + x.
823 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
824 return regIsPICBase(BaseReg, MRI);
830 // All other instructions marked M_REMATERIALIZABLE are always trivially
835 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
836 /// would clobber the EFLAGS condition register. Note the result may be
837 /// conservative. If it cannot definitely determine the safety after visiting
838 /// two instructions it assumes it's not safe.
839 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
840 MachineBasicBlock::iterator I) {
841 // For compile time consideration, if we are not able to determine the
842 // safety after visiting 2 instructions, we will assume it's not safe.
843 for (unsigned i = 0; i < 2; ++i) {
845 // Reached end of block, it's safe.
847 bool SeenDef = false;
848 for (unsigned j = 0, e = I->getNumOperands(); j != e; ++j) {
849 MachineOperand &MO = I->getOperand(j);
850 if (!MO.isRegister())
852 if (MO.getReg() == X86::EFLAGS) {
860 // This instruction defines EFLAGS, no need to look any further.
865 // Conservative answer.
869 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
870 MachineBasicBlock::iterator I,
872 const MachineInstr *Orig) const {
873 unsigned SubIdx = Orig->getOperand(0).isReg()
874 ? Orig->getOperand(0).getSubReg() : 0;
875 bool ChangeSubIdx = SubIdx != 0;
876 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
877 DestReg = RI.getSubReg(DestReg, SubIdx);
881 // MOV32r0 etc. are implemented with xor which clobbers condition code.
882 // Re-materialize them as movri instructions to avoid side effects.
883 bool Emitted = false;
884 switch (Orig->getOpcode()) {
890 if (!isSafeToClobberEFLAGS(MBB, I)) {
892 switch (Orig->getOpcode()) {
894 case X86::MOV8r0: Opc = X86::MOV8ri; break;
895 case X86::MOV16r0: Opc = X86::MOV16ri; break;
896 case X86::MOV32r0: Opc = X86::MOV32ri; break;
897 case X86::MOV64r0: Opc = X86::MOV64ri32; break;
899 BuildMI(MBB, I, get(Opc), DestReg).addImm(0);
907 MachineInstr *MI = Orig->clone();
908 MI->getOperand(0).setReg(DestReg);
913 MachineInstr *NewMI = prior(I);
914 NewMI->getOperand(0).setSubReg(SubIdx);
918 /// isInvariantLoad - Return true if the specified instruction (which is marked
919 /// mayLoad) is loading from a location whose value is invariant across the
920 /// function. For example, loading a value from the constant pool or from
921 /// from the argument area of a function if it does not change. This should
922 /// only return true of *all* loads the instruction does are invariant (if it
923 /// does multiple loads).
924 bool X86InstrInfo::isInvariantLoad(MachineInstr *MI) const {
925 // This code cares about loads from three cases: constant pool entries,
926 // invariant argument slots, and global stubs. In order to handle these cases
927 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
928 // operand and base our analysis on it. This is safe because the address of
929 // none of these three cases is ever used as anything other than a load base
930 // and X86 doesn't have any instructions that load from multiple places.
932 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
933 const MachineOperand &MO = MI->getOperand(i);
934 // Loads from constant pools are trivially invariant.
939 return isGVStub(MO.getGlobal(), TM);
941 // If this is a load from an invariant stack slot, the load is a constant.
943 const MachineFrameInfo &MFI =
944 *MI->getParent()->getParent()->getFrameInfo();
945 int Idx = MO.getIndex();
946 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
950 // All other instances of these instructions are presumed to have other
955 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
956 /// is not marked dead.
957 static bool hasLiveCondCodeDef(MachineInstr *MI) {
958 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
959 MachineOperand &MO = MI->getOperand(i);
960 if (MO.isRegister() && MO.isDef() &&
961 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
968 /// convertToThreeAddress - This method must be implemented by targets that
969 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
970 /// may be able to convert a two-address instruction into a true
971 /// three-address instruction on demand. This allows the X86 target (for
972 /// example) to convert ADD and SHL instructions into LEA instructions if they
973 /// would require register copies due to two-addressness.
975 /// This method returns a null pointer if the transformation cannot be
976 /// performed, otherwise it returns the new instruction.
979 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
980 MachineBasicBlock::iterator &MBBI,
981 LiveVariables *LV) const {
982 MachineInstr *MI = MBBI;
983 // All instructions input are two-addr instructions. Get the known operands.
984 unsigned Dest = MI->getOperand(0).getReg();
985 unsigned Src = MI->getOperand(1).getReg();
987 MachineInstr *NewMI = NULL;
988 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
989 // we have better subtarget support, enable the 16-bit LEA generation here.
990 bool DisableLEA16 = true;
992 unsigned MIOpc = MI->getOpcode();
994 case X86::SHUFPSrri: {
995 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
996 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
998 unsigned A = MI->getOperand(0).getReg();
999 unsigned B = MI->getOperand(1).getReg();
1000 unsigned C = MI->getOperand(2).getReg();
1001 unsigned M = MI->getOperand(3).getImm();
1002 if (B != C) return 0;
1003 NewMI = BuildMI(get(X86::PSHUFDri), A).addReg(B).addImm(M);
1006 case X86::SHL64ri: {
1007 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1008 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1009 // the flags produced by a shift yet, so this is safe.
1010 unsigned Dest = MI->getOperand(0).getReg();
1011 unsigned Src = MI->getOperand(1).getReg();
1012 unsigned ShAmt = MI->getOperand(2).getImm();
1013 if (ShAmt == 0 || ShAmt >= 4) return 0;
1015 NewMI = BuildMI(get(X86::LEA64r), Dest)
1016 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
1019 case X86::SHL32ri: {
1020 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1021 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1022 // the flags produced by a shift yet, so this is safe.
1023 unsigned Dest = MI->getOperand(0).getReg();
1024 unsigned Src = MI->getOperand(1).getReg();
1025 unsigned ShAmt = MI->getOperand(2).getImm();
1026 if (ShAmt == 0 || ShAmt >= 4) return 0;
1028 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1029 X86::LEA64_32r : X86::LEA32r;
1030 NewMI = BuildMI(get(Opc), Dest)
1031 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
1034 case X86::SHL16ri: {
1035 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1036 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1037 // the flags produced by a shift yet, so this is safe.
1038 unsigned Dest = MI->getOperand(0).getReg();
1039 unsigned Src = MI->getOperand(1).getReg();
1040 unsigned ShAmt = MI->getOperand(2).getImm();
1041 if (ShAmt == 0 || ShAmt >= 4) return 0;
1044 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1045 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1046 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1047 ? X86::LEA64_32r : X86::LEA32r;
1048 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1049 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1051 // Build and insert into an implicit UNDEF value. This is OK because
1052 // well be shifting and then extracting the lower 16-bits.
1053 MachineInstr *Undef = BuildMI(get(X86::IMPLICIT_DEF), leaInReg);
1056 BuildMI(get(X86::INSERT_SUBREG),leaInReg)
1057 .addReg(leaInReg).addReg(Src).addImm(X86::SUBREG_16BIT);
1059 NewMI = BuildMI(get(Opc), leaOutReg)
1060 .addReg(0).addImm(1 << ShAmt).addReg(leaInReg).addImm(0);
1063 BuildMI(get(X86::EXTRACT_SUBREG), Dest)
1064 .addReg(leaOutReg).addImm(X86::SUBREG_16BIT);
1065 Ext->copyKillDeadInfo(MI);
1067 MFI->insert(MBBI, Undef);
1068 MFI->insert(MBBI, Ins); // Insert the insert_subreg
1070 LV->instructionChanged(MI, NewMI); // Update live variables
1071 LV->addVirtualRegisterKilled(leaInReg, NewMI);
1073 MFI->insert(MBBI, NewMI); // Insert the new inst
1074 if (LV) LV->addVirtualRegisterKilled(leaOutReg, Ext);
1075 MFI->insert(MBBI, Ext); // Insert the extract_subreg
1078 NewMI = BuildMI(get(X86::LEA16r), Dest)
1079 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
1084 // The following opcodes also sets the condition code register(s). Only
1085 // convert them to equivalent lea if the condition code register def's
1087 if (hasLiveCondCodeDef(MI))
1090 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1095 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1096 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1097 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1098 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, 1);
1102 case X86::INC64_16r:
1103 if (DisableLEA16) return 0;
1104 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1105 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, 1);
1109 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1110 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1111 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1112 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, -1);
1116 case X86::DEC64_16r:
1117 if (DisableLEA16) return 0;
1118 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1119 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, -1);
1122 case X86::ADD32rr: {
1123 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1124 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1125 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1126 NewMI = addRegReg(BuildMI(get(Opc), Dest), Src,
1127 MI->getOperand(2).getReg());
1131 if (DisableLEA16) return 0;
1132 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1133 NewMI = addRegReg(BuildMI(get(X86::LEA16r), Dest), Src,
1134 MI->getOperand(2).getReg());
1136 case X86::ADD64ri32:
1138 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1139 if (MI->getOperand(2).isImmediate())
1140 NewMI = addRegOffset(BuildMI(get(X86::LEA64r), Dest), Src,
1141 MI->getOperand(2).getImm());
1145 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1146 if (MI->getOperand(2).isImmediate()) {
1147 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1148 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src,
1149 MI->getOperand(2).getImm());
1154 if (DisableLEA16) return 0;
1155 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1156 if (MI->getOperand(2).isImmediate())
1157 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src,
1158 MI->getOperand(2).getImm());
1161 if (DisableLEA16) return 0;
1163 case X86::SHL64ri: {
1164 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImmediate() &&
1165 "Unknown shl instruction!");
1166 unsigned ShAmt = MI->getOperand(2).getImm();
1167 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1169 AM.Scale = 1 << ShAmt;
1171 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1172 : (MIOpc == X86::SHL32ri
1173 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1174 NewMI = addFullAddress(BuildMI(get(Opc), Dest), AM);
1182 if (!NewMI) return 0;
1184 NewMI->copyKillDeadInfo(MI);
1185 if (LV) LV->instructionChanged(MI, NewMI); // Update live variables
1186 MFI->insert(MBBI, NewMI); // Insert the new inst
1190 /// commuteInstruction - We have a few instructions that must be hacked on to
1194 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1195 switch (MI->getOpcode()) {
1196 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1197 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1198 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1199 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1200 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1201 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1204 switch (MI->getOpcode()) {
1205 default: assert(0 && "Unreachable!");
1206 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1207 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1208 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1209 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1210 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1211 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1213 unsigned Amt = MI->getOperand(3).getImm();
1214 unsigned A = MI->getOperand(0).getReg();
1215 unsigned B = MI->getOperand(1).getReg();
1216 unsigned C = MI->getOperand(2).getReg();
1217 bool AisDead = MI->getOperand(0).isDead();
1218 bool BisKill = MI->getOperand(1).isKill();
1219 bool CisKill = MI->getOperand(2).isKill();
1220 // If machine instrs are no longer in two-address forms, update
1221 // destination register as well.
1223 // Must be two address instruction!
1224 assert(MI->getDesc().getOperandConstraint(0, TOI::TIED_TO) &&
1225 "Expecting a two-address instruction!");
1229 return BuildMI(get(Opc)).addReg(A, true, false, false, AisDead)
1230 .addReg(C, false, false, CisKill)
1231 .addReg(B, false, false, BisKill).addImm(Size-Amt);
1233 case X86::CMOVB16rr:
1234 case X86::CMOVB32rr:
1235 case X86::CMOVB64rr:
1236 case X86::CMOVAE16rr:
1237 case X86::CMOVAE32rr:
1238 case X86::CMOVAE64rr:
1239 case X86::CMOVE16rr:
1240 case X86::CMOVE32rr:
1241 case X86::CMOVE64rr:
1242 case X86::CMOVNE16rr:
1243 case X86::CMOVNE32rr:
1244 case X86::CMOVNE64rr:
1245 case X86::CMOVBE16rr:
1246 case X86::CMOVBE32rr:
1247 case X86::CMOVBE64rr:
1248 case X86::CMOVA16rr:
1249 case X86::CMOVA32rr:
1250 case X86::CMOVA64rr:
1251 case X86::CMOVL16rr:
1252 case X86::CMOVL32rr:
1253 case X86::CMOVL64rr:
1254 case X86::CMOVGE16rr:
1255 case X86::CMOVGE32rr:
1256 case X86::CMOVGE64rr:
1257 case X86::CMOVLE16rr:
1258 case X86::CMOVLE32rr:
1259 case X86::CMOVLE64rr:
1260 case X86::CMOVG16rr:
1261 case X86::CMOVG32rr:
1262 case X86::CMOVG64rr:
1263 case X86::CMOVS16rr:
1264 case X86::CMOVS32rr:
1265 case X86::CMOVS64rr:
1266 case X86::CMOVNS16rr:
1267 case X86::CMOVNS32rr:
1268 case X86::CMOVNS64rr:
1269 case X86::CMOVP16rr:
1270 case X86::CMOVP32rr:
1271 case X86::CMOVP64rr:
1272 case X86::CMOVNP16rr:
1273 case X86::CMOVNP32rr:
1274 case X86::CMOVNP64rr: {
1276 switch (MI->getOpcode()) {
1278 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1279 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1280 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1281 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1282 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1283 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1284 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1285 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1286 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1287 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1288 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1289 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1290 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1291 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1292 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1293 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1294 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1295 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1296 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1297 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1298 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1299 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1300 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1301 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1302 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1303 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1304 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1305 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1306 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1307 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1308 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1309 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1310 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1311 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1312 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1313 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1314 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1315 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1316 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1317 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1318 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1319 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1322 MI->setDesc(get(Opc));
1323 // Fallthrough intended.
1326 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1330 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1332 default: return X86::COND_INVALID;
1333 case X86::JE: return X86::COND_E;
1334 case X86::JNE: return X86::COND_NE;
1335 case X86::JL: return X86::COND_L;
1336 case X86::JLE: return X86::COND_LE;
1337 case X86::JG: return X86::COND_G;
1338 case X86::JGE: return X86::COND_GE;
1339 case X86::JB: return X86::COND_B;
1340 case X86::JBE: return X86::COND_BE;
1341 case X86::JA: return X86::COND_A;
1342 case X86::JAE: return X86::COND_AE;
1343 case X86::JS: return X86::COND_S;
1344 case X86::JNS: return X86::COND_NS;
1345 case X86::JP: return X86::COND_P;
1346 case X86::JNP: return X86::COND_NP;
1347 case X86::JO: return X86::COND_O;
1348 case X86::JNO: return X86::COND_NO;
1352 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1354 default: assert(0 && "Illegal condition code!");
1355 case X86::COND_E: return X86::JE;
1356 case X86::COND_NE: return X86::JNE;
1357 case X86::COND_L: return X86::JL;
1358 case X86::COND_LE: return X86::JLE;
1359 case X86::COND_G: return X86::JG;
1360 case X86::COND_GE: return X86::JGE;
1361 case X86::COND_B: return X86::JB;
1362 case X86::COND_BE: return X86::JBE;
1363 case X86::COND_A: return X86::JA;
1364 case X86::COND_AE: return X86::JAE;
1365 case X86::COND_S: return X86::JS;
1366 case X86::COND_NS: return X86::JNS;
1367 case X86::COND_P: return X86::JP;
1368 case X86::COND_NP: return X86::JNP;
1369 case X86::COND_O: return X86::JO;
1370 case X86::COND_NO: return X86::JNO;
1374 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1375 /// e.g. turning COND_E to COND_NE.
1376 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1378 default: assert(0 && "Illegal condition code!");
1379 case X86::COND_E: return X86::COND_NE;
1380 case X86::COND_NE: return X86::COND_E;
1381 case X86::COND_L: return X86::COND_GE;
1382 case X86::COND_LE: return X86::COND_G;
1383 case X86::COND_G: return X86::COND_LE;
1384 case X86::COND_GE: return X86::COND_L;
1385 case X86::COND_B: return X86::COND_AE;
1386 case X86::COND_BE: return X86::COND_A;
1387 case X86::COND_A: return X86::COND_BE;
1388 case X86::COND_AE: return X86::COND_B;
1389 case X86::COND_S: return X86::COND_NS;
1390 case X86::COND_NS: return X86::COND_S;
1391 case X86::COND_P: return X86::COND_NP;
1392 case X86::COND_NP: return X86::COND_P;
1393 case X86::COND_O: return X86::COND_NO;
1394 case X86::COND_NO: return X86::COND_O;
1398 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1399 const TargetInstrDesc &TID = MI->getDesc();
1400 if (!TID.isTerminator()) return false;
1402 // Conditional branch is a special case.
1403 if (TID.isBranch() && !TID.isBarrier())
1405 if (!TID.isPredicable())
1407 return !isPredicated(MI);
1410 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1411 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1412 const X86InstrInfo &TII) {
1413 if (MI->getOpcode() == X86::FP_REG_KILL)
1415 return TII.isUnpredicatedTerminator(MI);
1418 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1419 MachineBasicBlock *&TBB,
1420 MachineBasicBlock *&FBB,
1421 std::vector<MachineOperand> &Cond) const {
1422 // If the block has no terminators, it just falls into the block after it.
1423 MachineBasicBlock::iterator I = MBB.end();
1424 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this))
1427 // Get the last instruction in the block.
1428 MachineInstr *LastInst = I;
1430 // If there is only one terminator instruction, process it.
1431 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this)) {
1432 if (!LastInst->getDesc().isBranch())
1435 // If the block ends with a branch there are 3 possibilities:
1436 // it's an unconditional, conditional, or indirect branch.
1438 if (LastInst->getOpcode() == X86::JMP) {
1439 TBB = LastInst->getOperand(0).getMBB();
1442 X86::CondCode BranchCode = GetCondFromBranchOpc(LastInst->getOpcode());
1443 if (BranchCode == X86::COND_INVALID)
1444 return true; // Can't handle indirect branch.
1446 // Otherwise, block ends with fall-through condbranch.
1447 TBB = LastInst->getOperand(0).getMBB();
1448 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1452 // Get the instruction before it if it's a terminator.
1453 MachineInstr *SecondLastInst = I;
1455 // If there are three terminators, we don't know what sort of block this is.
1456 if (SecondLastInst && I != MBB.begin() &&
1457 isBrAnalysisUnpredicatedTerminator(--I, *this))
1460 // If the block ends with X86::JMP and a conditional branch, handle it.
1461 X86::CondCode BranchCode = GetCondFromBranchOpc(SecondLastInst->getOpcode());
1462 if (BranchCode != X86::COND_INVALID && LastInst->getOpcode() == X86::JMP) {
1463 TBB = SecondLastInst->getOperand(0).getMBB();
1464 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1465 FBB = LastInst->getOperand(0).getMBB();
1469 // If the block ends with two X86::JMPs, handle it. The second one is not
1470 // executed, so remove it.
1471 if (SecondLastInst->getOpcode() == X86::JMP &&
1472 LastInst->getOpcode() == X86::JMP) {
1473 TBB = SecondLastInst->getOperand(0).getMBB();
1475 I->eraseFromParent();
1479 // Otherwise, can't handle this.
1483 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1484 MachineBasicBlock::iterator I = MBB.end();
1485 if (I == MBB.begin()) return 0;
1487 if (I->getOpcode() != X86::JMP &&
1488 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1491 // Remove the branch.
1492 I->eraseFromParent();
1496 if (I == MBB.begin()) return 1;
1498 if (GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1501 // Remove the branch.
1502 I->eraseFromParent();
1506 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
1507 MachineOperand &MO) {
1508 if (MO.isRegister())
1509 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
1510 false, false, MO.getSubReg());
1511 else if (MO.isImmediate())
1512 MIB = MIB.addImm(MO.getImm());
1513 else if (MO.isFrameIndex())
1514 MIB = MIB.addFrameIndex(MO.getIndex());
1515 else if (MO.isGlobalAddress())
1516 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
1517 else if (MO.isConstantPoolIndex())
1518 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
1519 else if (MO.isJumpTableIndex())
1520 MIB = MIB.addJumpTableIndex(MO.getIndex());
1521 else if (MO.isExternalSymbol())
1522 MIB = MIB.addExternalSymbol(MO.getSymbolName());
1524 assert(0 && "Unknown operand for X86InstrAddOperand!");
1530 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1531 MachineBasicBlock *FBB,
1532 const std::vector<MachineOperand> &Cond) const {
1533 // Shouldn't be a fall through.
1534 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1535 assert((Cond.size() == 1 || Cond.size() == 0) &&
1536 "X86 branch conditions have one component!");
1538 if (FBB == 0) { // One way branch.
1540 // Unconditional branch?
1541 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
1543 // Conditional branch.
1544 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
1545 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1550 // Two-way Conditional branch.
1551 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
1552 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1553 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
1557 void X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1558 MachineBasicBlock::iterator MI,
1559 unsigned DestReg, unsigned SrcReg,
1560 const TargetRegisterClass *DestRC,
1561 const TargetRegisterClass *SrcRC) const {
1562 if (DestRC == SrcRC) {
1564 if (DestRC == &X86::GR64RegClass) {
1566 } else if (DestRC == &X86::GR32RegClass) {
1568 } else if (DestRC == &X86::GR16RegClass) {
1570 } else if (DestRC == &X86::GR8RegClass) {
1572 } else if (DestRC == &X86::GR32_RegClass) {
1573 Opc = X86::MOV32_rr;
1574 } else if (DestRC == &X86::GR16_RegClass) {
1575 Opc = X86::MOV16_rr;
1576 } else if (DestRC == &X86::RFP32RegClass) {
1577 Opc = X86::MOV_Fp3232;
1578 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1579 Opc = X86::MOV_Fp6464;
1580 } else if (DestRC == &X86::RFP80RegClass) {
1581 Opc = X86::MOV_Fp8080;
1582 } else if (DestRC == &X86::FR32RegClass) {
1583 Opc = X86::FsMOVAPSrr;
1584 } else if (DestRC == &X86::FR64RegClass) {
1585 Opc = X86::FsMOVAPDrr;
1586 } else if (DestRC == &X86::VR128RegClass) {
1587 Opc = X86::MOVAPSrr;
1588 } else if (DestRC == &X86::VR64RegClass) {
1589 Opc = X86::MMX_MOVQ64rr;
1591 assert(0 && "Unknown regclass");
1594 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
1598 // Moving EFLAGS to / from another register requires a push and a pop.
1599 if (SrcRC == &X86::CCRRegClass) {
1600 assert(SrcReg == X86::EFLAGS);
1601 if (DestRC == &X86::GR64RegClass) {
1602 BuildMI(MBB, MI, get(X86::PUSHFQ));
1603 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
1605 } else if (DestRC == &X86::GR32RegClass) {
1606 BuildMI(MBB, MI, get(X86::PUSHFD));
1607 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
1610 } else if (DestRC == &X86::CCRRegClass) {
1611 assert(DestReg == X86::EFLAGS);
1612 if (SrcRC == &X86::GR64RegClass) {
1613 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
1614 BuildMI(MBB, MI, get(X86::POPFQ));
1616 } else if (SrcRC == &X86::GR32RegClass) {
1617 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
1618 BuildMI(MBB, MI, get(X86::POPFD));
1623 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1624 if (SrcRC == &X86::RSTRegClass) {
1625 // Copying from ST(0)/ST(1).
1626 assert((SrcReg == X86::ST0 || SrcReg == X86::ST1) &&
1627 "Can only copy from ST(0)/ST(1) right now");
1628 bool isST0 = SrcReg == X86::ST0;
1630 if (DestRC == &X86::RFP32RegClass)
1631 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1632 else if (DestRC == &X86::RFP64RegClass)
1633 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1635 assert(DestRC == &X86::RFP80RegClass);
1636 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1638 BuildMI(MBB, MI, get(Opc), DestReg);
1642 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1643 if (DestRC == &X86::RSTRegClass) {
1644 // Copying to ST(0). FIXME: handle ST(1) also
1645 assert(DestReg == X86::ST0 && "Can only copy to TOS right now");
1647 if (SrcRC == &X86::RFP32RegClass)
1648 Opc = X86::FpSET_ST0_32;
1649 else if (SrcRC == &X86::RFP64RegClass)
1650 Opc = X86::FpSET_ST0_64;
1652 assert(SrcRC == &X86::RFP80RegClass);
1653 Opc = X86::FpSET_ST0_80;
1655 BuildMI(MBB, MI, get(Opc)).addReg(SrcReg);
1659 assert(0 && "Not yet supported!");
1663 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1664 unsigned StackAlign) {
1666 if (RC == &X86::GR64RegClass) {
1668 } else if (RC == &X86::GR32RegClass) {
1670 } else if (RC == &X86::GR16RegClass) {
1672 } else if (RC == &X86::GR8RegClass) {
1674 } else if (RC == &X86::GR32_RegClass) {
1675 Opc = X86::MOV32_mr;
1676 } else if (RC == &X86::GR16_RegClass) {
1677 Opc = X86::MOV16_mr;
1678 } else if (RC == &X86::RFP80RegClass) {
1679 Opc = X86::ST_FpP80m; // pops
1680 } else if (RC == &X86::RFP64RegClass) {
1681 Opc = X86::ST_Fp64m;
1682 } else if (RC == &X86::RFP32RegClass) {
1683 Opc = X86::ST_Fp32m;
1684 } else if (RC == &X86::FR32RegClass) {
1686 } else if (RC == &X86::FR64RegClass) {
1688 } else if (RC == &X86::VR128RegClass) {
1689 // FIXME: Use movaps once we are capable of selectively
1690 // aligning functions that spill SSE registers on 16-byte boundaries.
1691 Opc = StackAlign >= 16 ? X86::MOVAPSmr : X86::MOVUPSmr;
1692 } else if (RC == &X86::VR64RegClass) {
1693 Opc = X86::MMX_MOVQ64mr;
1695 assert(0 && "Unknown regclass");
1702 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1703 MachineBasicBlock::iterator MI,
1704 unsigned SrcReg, bool isKill, int FrameIdx,
1705 const TargetRegisterClass *RC) const {
1706 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
1707 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
1708 .addReg(SrcReg, false, false, isKill);
1711 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1713 SmallVectorImpl<MachineOperand> &Addr,
1714 const TargetRegisterClass *RC,
1715 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1716 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
1717 MachineInstrBuilder MIB = BuildMI(get(Opc));
1718 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1719 MIB = X86InstrAddOperand(MIB, Addr[i]);
1720 MIB.addReg(SrcReg, false, false, isKill);
1721 NewMIs.push_back(MIB);
1724 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1725 unsigned StackAlign) {
1727 if (RC == &X86::GR64RegClass) {
1729 } else if (RC == &X86::GR32RegClass) {
1731 } else if (RC == &X86::GR16RegClass) {
1733 } else if (RC == &X86::GR8RegClass) {
1735 } else if (RC == &X86::GR32_RegClass) {
1736 Opc = X86::MOV32_rm;
1737 } else if (RC == &X86::GR16_RegClass) {
1738 Opc = X86::MOV16_rm;
1739 } else if (RC == &X86::RFP80RegClass) {
1740 Opc = X86::LD_Fp80m;
1741 } else if (RC == &X86::RFP64RegClass) {
1742 Opc = X86::LD_Fp64m;
1743 } else if (RC == &X86::RFP32RegClass) {
1744 Opc = X86::LD_Fp32m;
1745 } else if (RC == &X86::FR32RegClass) {
1747 } else if (RC == &X86::FR64RegClass) {
1749 } else if (RC == &X86::VR128RegClass) {
1750 // FIXME: Use movaps once we are capable of selectively
1751 // aligning functions that spill SSE registers on 16-byte boundaries.
1752 Opc = StackAlign >= 16 ? X86::MOVAPSrm : X86::MOVUPSrm;
1753 } else if (RC == &X86::VR64RegClass) {
1754 Opc = X86::MMX_MOVQ64rm;
1756 assert(0 && "Unknown regclass");
1763 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1764 MachineBasicBlock::iterator MI,
1765 unsigned DestReg, int FrameIdx,
1766 const TargetRegisterClass *RC) const{
1767 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
1768 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
1771 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1772 SmallVectorImpl<MachineOperand> &Addr,
1773 const TargetRegisterClass *RC,
1774 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1775 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
1776 MachineInstrBuilder MIB = BuildMI(get(Opc), DestReg);
1777 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1778 MIB = X86InstrAddOperand(MIB, Addr[i]);
1779 NewMIs.push_back(MIB);
1782 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1783 MachineBasicBlock::iterator MI,
1784 const std::vector<CalleeSavedInfo> &CSI) const {
1788 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1789 unsigned SlotSize = is64Bit ? 8 : 4;
1791 MachineFunction &MF = *MBB.getParent();
1792 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1793 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1795 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1796 for (unsigned i = CSI.size(); i != 0; --i) {
1797 unsigned Reg = CSI[i-1].getReg();
1798 // Add the callee-saved register as live-in. It's killed at the spill.
1800 BuildMI(MBB, MI, get(Opc)).addReg(Reg);
1805 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1806 MachineBasicBlock::iterator MI,
1807 const std::vector<CalleeSavedInfo> &CSI) const {
1811 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1813 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1814 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1815 unsigned Reg = CSI[i].getReg();
1816 BuildMI(MBB, MI, get(Opc), Reg);
1821 static MachineInstr *FuseTwoAddrInst(unsigned Opcode,
1822 SmallVector<MachineOperand,4> &MOs,
1823 MachineInstr *MI, const TargetInstrInfo &TII) {
1824 // Create the base instruction with the memory operand as the first part.
1825 MachineInstr *NewMI = new MachineInstr(TII.get(Opcode), true);
1826 MachineInstrBuilder MIB(NewMI);
1827 unsigned NumAddrOps = MOs.size();
1828 for (unsigned i = 0; i != NumAddrOps; ++i)
1829 MIB = X86InstrAddOperand(MIB, MOs[i]);
1830 if (NumAddrOps < 4) // FrameIndex only
1831 MIB.addImm(1).addReg(0).addImm(0);
1833 // Loop over the rest of the ri operands, converting them over.
1834 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1835 for (unsigned i = 0; i != NumOps; ++i) {
1836 MachineOperand &MO = MI->getOperand(i+2);
1837 MIB = X86InstrAddOperand(MIB, MO);
1839 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1840 MachineOperand &MO = MI->getOperand(i);
1841 MIB = X86InstrAddOperand(MIB, MO);
1846 static MachineInstr *FuseInst(unsigned Opcode, unsigned OpNo,
1847 SmallVector<MachineOperand,4> &MOs,
1848 MachineInstr *MI, const TargetInstrInfo &TII) {
1849 MachineInstr *NewMI = new MachineInstr(TII.get(Opcode), true);
1850 MachineInstrBuilder MIB(NewMI);
1852 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1853 MachineOperand &MO = MI->getOperand(i);
1855 assert(MO.isRegister() && "Expected to fold into reg operand!");
1856 unsigned NumAddrOps = MOs.size();
1857 for (unsigned i = 0; i != NumAddrOps; ++i)
1858 MIB = X86InstrAddOperand(MIB, MOs[i]);
1859 if (NumAddrOps < 4) // FrameIndex only
1860 MIB.addImm(1).addReg(0).addImm(0);
1862 MIB = X86InstrAddOperand(MIB, MO);
1868 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
1869 SmallVector<MachineOperand,4> &MOs,
1871 MachineInstrBuilder MIB = BuildMI(TII.get(Opcode));
1873 unsigned NumAddrOps = MOs.size();
1874 for (unsigned i = 0; i != NumAddrOps; ++i)
1875 MIB = X86InstrAddOperand(MIB, MOs[i]);
1876 if (NumAddrOps < 4) // FrameIndex only
1877 MIB.addImm(1).addReg(0).addImm(0);
1878 return MIB.addImm(0);
1882 X86InstrInfo::foldMemoryOperand(MachineInstr *MI, unsigned i,
1883 SmallVector<MachineOperand,4> &MOs) const {
1884 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
1885 bool isTwoAddrFold = false;
1886 unsigned NumOps = MI->getDesc().getNumOperands();
1887 bool isTwoAddr = NumOps > 1 &&
1888 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
1890 MachineInstr *NewMI = NULL;
1891 // Folding a memory location into the two-address part of a two-address
1892 // instruction is different than folding it other places. It requires
1893 // replacing the *two* registers with the memory location.
1894 if (isTwoAddr && NumOps >= 2 && i < 2 &&
1895 MI->getOperand(0).isRegister() &&
1896 MI->getOperand(1).isRegister() &&
1897 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
1898 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
1899 isTwoAddrFold = true;
1900 } else if (i == 0) { // If operand 0
1901 if (MI->getOpcode() == X86::MOV16r0)
1902 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
1903 else if (MI->getOpcode() == X86::MOV32r0)
1904 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
1905 else if (MI->getOpcode() == X86::MOV64r0)
1906 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
1907 else if (MI->getOpcode() == X86::MOV8r0)
1908 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
1910 NewMI->copyKillDeadInfo(MI);
1914 OpcodeTablePtr = &RegOp2MemOpTable0;
1915 } else if (i == 1) {
1916 OpcodeTablePtr = &RegOp2MemOpTable1;
1917 } else if (i == 2) {
1918 OpcodeTablePtr = &RegOp2MemOpTable2;
1921 // If table selected...
1922 if (OpcodeTablePtr) {
1923 // Find the Opcode to fuse
1924 DenseMap<unsigned*, unsigned>::iterator I =
1925 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
1926 if (I != OpcodeTablePtr->end()) {
1928 NewMI = FuseTwoAddrInst(I->second, MOs, MI, *this);
1930 NewMI = FuseInst(I->second, i, MOs, MI, *this);
1931 NewMI->copyKillDeadInfo(MI);
1937 if (PrintFailedFusing)
1938 cerr << "We failed to fuse operand " << i << *MI;
1943 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1945 SmallVectorImpl<unsigned> &Ops,
1946 int FrameIndex) const {
1947 // Check switch flag
1948 if (NoFusing) return NULL;
1950 const MachineFrameInfo *MFI = MF.getFrameInfo();
1951 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
1952 // FIXME: Move alignment requirement into tables?
1953 if (Alignment < 16) {
1954 switch (MI->getOpcode()) {
1956 // Not always safe to fold movsd into these instructions since their load
1957 // folding variants expects the address to be 16 byte aligned.
1958 case X86::FsANDNPDrr:
1959 case X86::FsANDNPSrr:
1960 case X86::FsANDPDrr:
1961 case X86::FsANDPSrr:
1964 case X86::FsXORPDrr:
1965 case X86::FsXORPSrr:
1970 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1971 unsigned NewOpc = 0;
1972 switch (MI->getOpcode()) {
1973 default: return NULL;
1974 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
1975 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
1976 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
1977 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
1979 // Change to CMPXXri r, 0 first.
1980 MI->setDesc(get(NewOpc));
1981 MI->getOperand(1).ChangeToImmediate(0);
1982 } else if (Ops.size() != 1)
1985 SmallVector<MachineOperand,4> MOs;
1986 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
1987 return foldMemoryOperand(MI, Ops[0], MOs);
1990 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1992 SmallVectorImpl<unsigned> &Ops,
1993 MachineInstr *LoadMI) const {
1994 // Check switch flag
1995 if (NoFusing) return NULL;
1997 unsigned Alignment = 0;
1998 for (unsigned i = 0, e = LoadMI->getNumMemOperands(); i != e; ++i) {
1999 const MachineMemOperand &MRO = LoadMI->getMemOperand(i);
2000 unsigned Align = MRO.getAlignment();
2001 if (Align > Alignment)
2005 // FIXME: Move alignment requirement into tables?
2006 if (Alignment < 16) {
2007 switch (MI->getOpcode()) {
2009 // Not always safe to fold movsd into these instructions since their load
2010 // folding variants expects the address to be 16 byte aligned.
2011 case X86::FsANDNPDrr:
2012 case X86::FsANDNPSrr:
2013 case X86::FsANDPDrr:
2014 case X86::FsANDPSrr:
2017 case X86::FsXORPDrr:
2018 case X86::FsXORPSrr:
2023 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2024 unsigned NewOpc = 0;
2025 switch (MI->getOpcode()) {
2026 default: return NULL;
2027 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2028 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2029 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2030 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2032 // Change to CMPXXri r, 0 first.
2033 MI->setDesc(get(NewOpc));
2034 MI->getOperand(1).ChangeToImmediate(0);
2035 } else if (Ops.size() != 1)
2038 SmallVector<MachineOperand,4> MOs;
2039 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2040 for (unsigned i = NumOps - 4; i != NumOps; ++i)
2041 MOs.push_back(LoadMI->getOperand(i));
2042 return foldMemoryOperand(MI, Ops[0], MOs);
2046 bool X86InstrInfo::canFoldMemoryOperand(MachineInstr *MI,
2047 SmallVectorImpl<unsigned> &Ops) const {
2048 // Check switch flag
2049 if (NoFusing) return 0;
2051 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2052 switch (MI->getOpcode()) {
2053 default: return false;
2062 if (Ops.size() != 1)
2065 unsigned OpNum = Ops[0];
2066 unsigned Opc = MI->getOpcode();
2067 unsigned NumOps = MI->getDesc().getNumOperands();
2068 bool isTwoAddr = NumOps > 1 &&
2069 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2071 // Folding a memory location into the two-address part of a two-address
2072 // instruction is different than folding it other places. It requires
2073 // replacing the *two* registers with the memory location.
2074 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2075 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2076 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2077 } else if (OpNum == 0) { // If operand 0
2086 OpcodeTablePtr = &RegOp2MemOpTable0;
2087 } else if (OpNum == 1) {
2088 OpcodeTablePtr = &RegOp2MemOpTable1;
2089 } else if (OpNum == 2) {
2090 OpcodeTablePtr = &RegOp2MemOpTable2;
2093 if (OpcodeTablePtr) {
2094 // Find the Opcode to fuse
2095 DenseMap<unsigned*, unsigned>::iterator I =
2096 OpcodeTablePtr->find((unsigned*)Opc);
2097 if (I != OpcodeTablePtr->end())
2103 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2104 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2105 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2106 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2107 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2108 if (I == MemOp2RegOpTable.end())
2110 unsigned Opc = I->second.first;
2111 unsigned Index = I->second.second & 0xf;
2112 bool FoldedLoad = I->second.second & (1 << 4);
2113 bool FoldedStore = I->second.second & (1 << 5);
2114 if (UnfoldLoad && !FoldedLoad)
2116 UnfoldLoad &= FoldedLoad;
2117 if (UnfoldStore && !FoldedStore)
2119 UnfoldStore &= FoldedStore;
2121 const TargetInstrDesc &TID = get(Opc);
2122 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2123 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2124 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2125 SmallVector<MachineOperand,4> AddrOps;
2126 SmallVector<MachineOperand,2> BeforeOps;
2127 SmallVector<MachineOperand,2> AfterOps;
2128 SmallVector<MachineOperand,4> ImpOps;
2129 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2130 MachineOperand &Op = MI->getOperand(i);
2131 if (i >= Index && i < Index+4)
2132 AddrOps.push_back(Op);
2133 else if (Op.isRegister() && Op.isImplicit())
2134 ImpOps.push_back(Op);
2136 BeforeOps.push_back(Op);
2138 AfterOps.push_back(Op);
2141 // Emit the load instruction.
2143 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2145 // Address operands cannot be marked isKill.
2146 for (unsigned i = 1; i != 5; ++i) {
2147 MachineOperand &MO = NewMIs[0]->getOperand(i);
2148 if (MO.isRegister())
2149 MO.setIsKill(false);
2154 // Emit the data processing instruction.
2155 MachineInstr *DataMI = new MachineInstr(TID, true);
2156 MachineInstrBuilder MIB(DataMI);
2159 MIB.addReg(Reg, true);
2160 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2161 MIB = X86InstrAddOperand(MIB, BeforeOps[i]);
2164 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2165 MIB = X86InstrAddOperand(MIB, AfterOps[i]);
2166 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2167 MachineOperand &MO = ImpOps[i];
2168 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2170 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2171 unsigned NewOpc = 0;
2172 switch (DataMI->getOpcode()) {
2174 case X86::CMP64ri32:
2178 MachineOperand &MO0 = DataMI->getOperand(0);
2179 MachineOperand &MO1 = DataMI->getOperand(1);
2180 if (MO1.getImm() == 0) {
2181 switch (DataMI->getOpcode()) {
2183 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2184 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2185 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2186 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2188 DataMI->setDesc(get(NewOpc));
2189 MO1.ChangeToRegister(MO0.getReg(), false);
2193 NewMIs.push_back(DataMI);
2195 // Emit the store instruction.
2197 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2198 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2199 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2200 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2207 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2208 SmallVectorImpl<SDNode*> &NewNodes) const {
2209 if (!N->isTargetOpcode())
2212 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2213 MemOp2RegOpTable.find((unsigned*)N->getTargetOpcode());
2214 if (I == MemOp2RegOpTable.end())
2216 unsigned Opc = I->second.first;
2217 unsigned Index = I->second.second & 0xf;
2218 bool FoldedLoad = I->second.second & (1 << 4);
2219 bool FoldedStore = I->second.second & (1 << 5);
2220 const TargetInstrDesc &TID = get(Opc);
2221 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2222 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2223 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2224 std::vector<SDOperand> AddrOps;
2225 std::vector<SDOperand> BeforeOps;
2226 std::vector<SDOperand> AfterOps;
2227 unsigned NumOps = N->getNumOperands();
2228 for (unsigned i = 0; i != NumOps-1; ++i) {
2229 SDOperand Op = N->getOperand(i);
2230 if (i >= Index && i < Index+4)
2231 AddrOps.push_back(Op);
2233 BeforeOps.push_back(Op);
2235 AfterOps.push_back(Op);
2237 SDOperand Chain = N->getOperand(NumOps-1);
2238 AddrOps.push_back(Chain);
2240 // Emit the load instruction.
2243 MVT VT = *RC->vt_begin();
2244 Load = DAG.getTargetNode(getLoadRegOpcode(RC, RI.getStackAlignment()), VT,
2245 MVT::Other, &AddrOps[0], AddrOps.size());
2246 NewNodes.push_back(Load);
2249 // Emit the data processing instruction.
2250 std::vector<MVT> VTs;
2251 const TargetRegisterClass *DstRC = 0;
2252 if (TID.getNumDefs() > 0) {
2253 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2254 DstRC = DstTOI.isLookupPtrRegClass()
2255 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2256 VTs.push_back(*DstRC->vt_begin());
2258 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2259 MVT VT = N->getValueType(i);
2260 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2264 BeforeOps.push_back(SDOperand(Load, 0));
2265 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2266 SDNode *NewNode= DAG.getTargetNode(Opc, VTs, &BeforeOps[0], BeforeOps.size());
2267 NewNodes.push_back(NewNode);
2269 // Emit the store instruction.
2272 AddrOps.push_back(SDOperand(NewNode, 0));
2273 AddrOps.push_back(Chain);
2274 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, RI.getStackAlignment()),
2275 MVT::Other, &AddrOps[0], AddrOps.size());
2276 NewNodes.push_back(Store);
2282 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2283 bool UnfoldLoad, bool UnfoldStore) const {
2284 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2285 MemOp2RegOpTable.find((unsigned*)Opc);
2286 if (I == MemOp2RegOpTable.end())
2288 bool FoldedLoad = I->second.second & (1 << 4);
2289 bool FoldedStore = I->second.second & (1 << 5);
2290 if (UnfoldLoad && !FoldedLoad)
2292 if (UnfoldStore && !FoldedStore)
2294 return I->second.first;
2297 bool X86InstrInfo::BlockHasNoFallThrough(MachineBasicBlock &MBB) const {
2298 if (MBB.empty()) return false;
2300 switch (MBB.back().getOpcode()) {
2301 case X86::TCRETURNri:
2302 case X86::TCRETURNdi:
2303 case X86::RET: // Return.
2308 case X86::JMP: // Uncond branch.
2309 case X86::JMP32r: // Indirect branch.
2310 case X86::JMP64r: // Indirect branch (64-bit).
2311 case X86::JMP32m: // Indirect branch through mem.
2312 case X86::JMP64m: // Indirect branch through mem (64-bit).
2314 default: return false;
2319 ReverseBranchCondition(std::vector<MachineOperand> &Cond) const {
2320 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2321 Cond[0].setImm(GetOppositeBranchCondition((X86::CondCode)Cond[0].getImm()));
2325 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
2326 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
2327 if (Subtarget->is64Bit())
2328 return &X86::GR64RegClass;
2330 return &X86::GR32RegClass;
2333 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2334 switch (Desc->TSFlags & X86II::ImmMask) {
2335 case X86II::Imm8: return 1;
2336 case X86II::Imm16: return 2;
2337 case X86II::Imm32: return 4;
2338 case X86II::Imm64: return 8;
2339 default: assert(0 && "Immediate size not set!");
2344 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2345 /// e.g. r8, xmm8, etc.
2346 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2347 if (!MO.isRegister()) return false;
2348 switch (MO.getReg()) {
2350 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2351 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2352 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2353 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2354 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2355 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2356 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2357 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2358 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2359 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2366 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2367 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2368 /// size, and 3) use of X86-64 extended registers.
2369 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2371 const TargetInstrDesc &Desc = MI.getDesc();
2373 // Pseudo instructions do not need REX prefix byte.
2374 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2376 if (Desc.TSFlags & X86II::REX_W)
2379 unsigned NumOps = Desc.getNumOperands();
2381 bool isTwoAddr = NumOps > 1 &&
2382 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2384 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2385 unsigned i = isTwoAddr ? 1 : 0;
2386 for (unsigned e = NumOps; i != e; ++i) {
2387 const MachineOperand& MO = MI.getOperand(i);
2388 if (MO.isRegister()) {
2389 unsigned Reg = MO.getReg();
2390 if (isX86_64NonExtLowByteReg(Reg))
2395 switch (Desc.TSFlags & X86II::FormMask) {
2396 case X86II::MRMInitReg:
2397 if (isX86_64ExtendedReg(MI.getOperand(0)))
2398 REX |= (1 << 0) | (1 << 2);
2400 case X86II::MRMSrcReg: {
2401 if (isX86_64ExtendedReg(MI.getOperand(0)))
2403 i = isTwoAddr ? 2 : 1;
2404 for (unsigned e = NumOps; i != e; ++i) {
2405 const MachineOperand& MO = MI.getOperand(i);
2406 if (isX86_64ExtendedReg(MO))
2411 case X86II::MRMSrcMem: {
2412 if (isX86_64ExtendedReg(MI.getOperand(0)))
2415 i = isTwoAddr ? 2 : 1;
2416 for (; i != NumOps; ++i) {
2417 const MachineOperand& MO = MI.getOperand(i);
2418 if (MO.isRegister()) {
2419 if (isX86_64ExtendedReg(MO))
2426 case X86II::MRM0m: case X86II::MRM1m:
2427 case X86II::MRM2m: case X86II::MRM3m:
2428 case X86II::MRM4m: case X86II::MRM5m:
2429 case X86II::MRM6m: case X86II::MRM7m:
2430 case X86II::MRMDestMem: {
2431 unsigned e = isTwoAddr ? 5 : 4;
2432 i = isTwoAddr ? 1 : 0;
2433 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2436 for (; i != e; ++i) {
2437 const MachineOperand& MO = MI.getOperand(i);
2438 if (MO.isRegister()) {
2439 if (isX86_64ExtendedReg(MO))
2447 if (isX86_64ExtendedReg(MI.getOperand(0)))
2449 i = isTwoAddr ? 2 : 1;
2450 for (unsigned e = NumOps; i != e; ++i) {
2451 const MachineOperand& MO = MI.getOperand(i);
2452 if (isX86_64ExtendedReg(MO))
2462 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2463 /// relative block address instruction
2465 static unsigned sizePCRelativeBlockAddress() {
2469 /// sizeGlobalAddress - Give the size of the emission of this global address
2471 static unsigned sizeGlobalAddress(bool dword) {
2472 return dword ? 8 : 4;
2475 /// sizeConstPoolAddress - Give the size of the emission of this constant
2478 static unsigned sizeConstPoolAddress(bool dword) {
2479 return dword ? 8 : 4;
2482 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2485 static unsigned sizeExternalSymbolAddress(bool dword) {
2486 return dword ? 8 : 4;
2489 /// sizeJumpTableAddress - Give the size of the emission of this jump
2492 static unsigned sizeJumpTableAddress(bool dword) {
2493 return dword ? 8 : 4;
2496 static unsigned sizeConstant(unsigned Size) {
2500 static unsigned sizeRegModRMByte(){
2504 static unsigned sizeSIBByte(){
2508 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2509 unsigned FinalSize = 0;
2510 // If this is a simple integer displacement that doesn't require a relocation.
2512 FinalSize += sizeConstant(4);
2516 // Otherwise, this is something that requires a relocation.
2517 if (RelocOp->isGlobalAddress()) {
2518 FinalSize += sizeGlobalAddress(false);
2519 } else if (RelocOp->isConstantPoolIndex()) {
2520 FinalSize += sizeConstPoolAddress(false);
2521 } else if (RelocOp->isJumpTableIndex()) {
2522 FinalSize += sizeJumpTableAddress(false);
2524 assert(0 && "Unknown value to relocate!");
2529 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2530 bool IsPIC, bool Is64BitMode) {
2531 const MachineOperand &Op3 = MI.getOperand(Op+3);
2533 const MachineOperand *DispForReloc = 0;
2534 unsigned FinalSize = 0;
2536 // Figure out what sort of displacement we have to handle here.
2537 if (Op3.isGlobalAddress()) {
2538 DispForReloc = &Op3;
2539 } else if (Op3.isConstantPoolIndex()) {
2540 if (Is64BitMode || IsPIC) {
2541 DispForReloc = &Op3;
2545 } else if (Op3.isJumpTableIndex()) {
2546 if (Is64BitMode || IsPIC) {
2547 DispForReloc = &Op3;
2555 const MachineOperand &Base = MI.getOperand(Op);
2556 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2558 unsigned BaseReg = Base.getReg();
2560 // Is a SIB byte needed?
2561 if (IndexReg.getReg() == 0 &&
2562 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2563 if (BaseReg == 0) { // Just a displacement?
2564 // Emit special case [disp32] encoding
2566 FinalSize += getDisplacementFieldSize(DispForReloc);
2568 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2569 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2570 // Emit simple indirect register encoding... [EAX] f.e.
2572 // Be pessimistic and assume it's a disp32, not a disp8
2574 // Emit the most general non-SIB encoding: [REG+disp32]
2576 FinalSize += getDisplacementFieldSize(DispForReloc);
2580 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2581 assert(IndexReg.getReg() != X86::ESP &&
2582 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2584 bool ForceDisp32 = false;
2585 if (BaseReg == 0 || DispForReloc) {
2586 // Emit the normal disp32 encoding.
2593 FinalSize += sizeSIBByte();
2595 // Do we need to output a displacement?
2596 if (DispVal != 0 || ForceDisp32) {
2597 FinalSize += getDisplacementFieldSize(DispForReloc);
2604 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2605 const TargetInstrDesc *Desc,
2606 bool IsPIC, bool Is64BitMode) {
2608 unsigned Opcode = Desc->Opcode;
2609 unsigned FinalSize = 0;
2611 // Emit the lock opcode prefix as needed.
2612 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2614 // Emit the repeat opcode prefix as needed.
2615 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2617 // Emit the operand size opcode prefix as needed.
2618 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2620 // Emit the address size opcode prefix as needed.
2621 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2623 bool Need0FPrefix = false;
2624 switch (Desc->TSFlags & X86II::Op0Mask) {
2625 case X86II::TB: // Two-byte opcode prefix
2626 case X86II::T8: // 0F 38
2627 case X86II::TA: // 0F 3A
2628 Need0FPrefix = true;
2630 case X86II::REP: break; // already handled.
2631 case X86II::XS: // F3 0F
2633 Need0FPrefix = true;
2635 case X86II::XD: // F2 0F
2637 Need0FPrefix = true;
2639 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2640 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2642 break; // Two-byte opcode prefix
2643 default: assert(0 && "Invalid prefix!");
2644 case 0: break; // No prefix!
2649 unsigned REX = X86InstrInfo::determineREX(MI);
2654 // 0x0F escape code must be emitted just before the opcode.
2658 switch (Desc->TSFlags & X86II::Op0Mask) {
2659 case X86II::T8: // 0F 38
2662 case X86II::TA: // 0F 3A
2667 // If this is a two-address instruction, skip one of the register operands.
2668 unsigned NumOps = Desc->getNumOperands();
2670 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2673 switch (Desc->TSFlags & X86II::FormMask) {
2674 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
2676 // Remember the current PC offset, this is the PIC relocation
2681 case TargetInstrInfo::INLINEASM: {
2682 const MachineFunction *MF = MI.getParent()->getParent();
2683 const char *AsmStr = MI.getOperand(0).getSymbolName();
2684 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
2685 FinalSize += AI->getInlineAsmLength(AsmStr);
2688 case TargetInstrInfo::DBG_LABEL:
2689 case TargetInstrInfo::EH_LABEL:
2691 case TargetInstrInfo::IMPLICIT_DEF:
2692 case TargetInstrInfo::DECLARE:
2693 case X86::DWARF_LOC:
2694 case X86::FP_REG_KILL:
2696 case X86::MOVPC32r: {
2697 // This emits the "call" portion of this pseudo instruction.
2699 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2708 if (CurOp != NumOps) {
2709 const MachineOperand &MO = MI.getOperand(CurOp++);
2710 if (MO.isMachineBasicBlock()) {
2711 FinalSize += sizePCRelativeBlockAddress();
2712 } else if (MO.isGlobalAddress()) {
2713 FinalSize += sizeGlobalAddress(false);
2714 } else if (MO.isExternalSymbol()) {
2715 FinalSize += sizeExternalSymbolAddress(false);
2716 } else if (MO.isImmediate()) {
2717 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2719 assert(0 && "Unknown RawFrm operand!");
2724 case X86II::AddRegFrm:
2728 if (CurOp != NumOps) {
2729 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2730 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2731 if (MO1.isImmediate())
2732 FinalSize += sizeConstant(Size);
2735 if (Opcode == X86::MOV64ri)
2737 if (MO1.isGlobalAddress()) {
2738 FinalSize += sizeGlobalAddress(dword);
2739 } else if (MO1.isExternalSymbol())
2740 FinalSize += sizeExternalSymbolAddress(dword);
2741 else if (MO1.isConstantPoolIndex())
2742 FinalSize += sizeConstPoolAddress(dword);
2743 else if (MO1.isJumpTableIndex())
2744 FinalSize += sizeJumpTableAddress(dword);
2749 case X86II::MRMDestReg: {
2751 FinalSize += sizeRegModRMByte();
2753 if (CurOp != NumOps) {
2755 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2759 case X86II::MRMDestMem: {
2761 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2763 if (CurOp != NumOps) {
2765 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2770 case X86II::MRMSrcReg:
2772 FinalSize += sizeRegModRMByte();
2774 if (CurOp != NumOps) {
2776 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2780 case X86II::MRMSrcMem: {
2783 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
2785 if (CurOp != NumOps) {
2787 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2792 case X86II::MRM0r: case X86II::MRM1r:
2793 case X86II::MRM2r: case X86II::MRM3r:
2794 case X86II::MRM4r: case X86II::MRM5r:
2795 case X86II::MRM6r: case X86II::MRM7r:
2798 FinalSize += sizeRegModRMByte();
2800 if (CurOp != NumOps) {
2801 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2802 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2803 if (MO1.isImmediate())
2804 FinalSize += sizeConstant(Size);
2807 if (Opcode == X86::MOV64ri32)
2809 if (MO1.isGlobalAddress()) {
2810 FinalSize += sizeGlobalAddress(dword);
2811 } else if (MO1.isExternalSymbol())
2812 FinalSize += sizeExternalSymbolAddress(dword);
2813 else if (MO1.isConstantPoolIndex())
2814 FinalSize += sizeConstPoolAddress(dword);
2815 else if (MO1.isJumpTableIndex())
2816 FinalSize += sizeJumpTableAddress(dword);
2821 case X86II::MRM0m: case X86II::MRM1m:
2822 case X86II::MRM2m: case X86II::MRM3m:
2823 case X86II::MRM4m: case X86II::MRM5m:
2824 case X86II::MRM6m: case X86II::MRM7m: {
2827 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2830 if (CurOp != NumOps) {
2831 const MachineOperand &MO = MI.getOperand(CurOp++);
2832 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2833 if (MO.isImmediate())
2834 FinalSize += sizeConstant(Size);
2837 if (Opcode == X86::MOV64mi32)
2839 if (MO.isGlobalAddress()) {
2840 FinalSize += sizeGlobalAddress(dword);
2841 } else if (MO.isExternalSymbol())
2842 FinalSize += sizeExternalSymbolAddress(dword);
2843 else if (MO.isConstantPoolIndex())
2844 FinalSize += sizeConstPoolAddress(dword);
2845 else if (MO.isJumpTableIndex())
2846 FinalSize += sizeJumpTableAddress(dword);
2852 case X86II::MRMInitReg:
2854 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
2855 FinalSize += sizeRegModRMByte();
2860 if (!Desc->isVariadic() && CurOp != NumOps) {
2861 cerr << "Cannot determine size: ";
2872 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
2873 const TargetInstrDesc &Desc = MI->getDesc();
2874 bool IsPIC = (TM.getRelocationModel() == Reloc::PIC_);
2875 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
2876 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
2877 if (Desc.getOpcode() == X86::MOVPC32r) {
2878 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);