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"
33 NoFusing("disable-spill-fusing",
34 cl::desc("Disable fusing of spill code into instructions"));
36 PrintFailedFusing("print-failed-fuse-candidates",
37 cl::desc("Print instructions that the allocator wants to"
38 " fuse, but the X86 backend currently can't"),
41 ReMatPICStubLoad("remat-pic-stub-load",
42 cl::desc("Re-materialize load from stub in PIC mode"),
43 cl::init(false), cl::Hidden);
46 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
47 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
48 TM(tm), RI(tm, *this) {
49 SmallVector<unsigned,16> AmbEntries;
50 static const unsigned OpTbl2Addr[][2] = {
51 { X86::ADC32ri, X86::ADC32mi },
52 { X86::ADC32ri8, X86::ADC32mi8 },
53 { X86::ADC32rr, X86::ADC32mr },
54 { X86::ADC64ri32, X86::ADC64mi32 },
55 { X86::ADC64ri8, X86::ADC64mi8 },
56 { X86::ADC64rr, X86::ADC64mr },
57 { X86::ADD16ri, X86::ADD16mi },
58 { X86::ADD16ri8, X86::ADD16mi8 },
59 { X86::ADD16rr, X86::ADD16mr },
60 { X86::ADD32ri, X86::ADD32mi },
61 { X86::ADD32ri8, X86::ADD32mi8 },
62 { X86::ADD32rr, X86::ADD32mr },
63 { X86::ADD64ri32, X86::ADD64mi32 },
64 { X86::ADD64ri8, X86::ADD64mi8 },
65 { X86::ADD64rr, X86::ADD64mr },
66 { X86::ADD8ri, X86::ADD8mi },
67 { X86::ADD8rr, X86::ADD8mr },
68 { X86::AND16ri, X86::AND16mi },
69 { X86::AND16ri8, X86::AND16mi8 },
70 { X86::AND16rr, X86::AND16mr },
71 { X86::AND32ri, X86::AND32mi },
72 { X86::AND32ri8, X86::AND32mi8 },
73 { X86::AND32rr, X86::AND32mr },
74 { X86::AND64ri32, X86::AND64mi32 },
75 { X86::AND64ri8, X86::AND64mi8 },
76 { X86::AND64rr, X86::AND64mr },
77 { X86::AND8ri, X86::AND8mi },
78 { X86::AND8rr, X86::AND8mr },
79 { X86::DEC16r, X86::DEC16m },
80 { X86::DEC32r, X86::DEC32m },
81 { X86::DEC64_16r, X86::DEC64_16m },
82 { X86::DEC64_32r, X86::DEC64_32m },
83 { X86::DEC64r, X86::DEC64m },
84 { X86::DEC8r, X86::DEC8m },
85 { X86::INC16r, X86::INC16m },
86 { X86::INC32r, X86::INC32m },
87 { X86::INC64_16r, X86::INC64_16m },
88 { X86::INC64_32r, X86::INC64_32m },
89 { X86::INC64r, X86::INC64m },
90 { X86::INC8r, X86::INC8m },
91 { X86::NEG16r, X86::NEG16m },
92 { X86::NEG32r, X86::NEG32m },
93 { X86::NEG64r, X86::NEG64m },
94 { X86::NEG8r, X86::NEG8m },
95 { X86::NOT16r, X86::NOT16m },
96 { X86::NOT32r, X86::NOT32m },
97 { X86::NOT64r, X86::NOT64m },
98 { X86::NOT8r, X86::NOT8m },
99 { X86::OR16ri, X86::OR16mi },
100 { X86::OR16ri8, X86::OR16mi8 },
101 { X86::OR16rr, X86::OR16mr },
102 { X86::OR32ri, X86::OR32mi },
103 { X86::OR32ri8, X86::OR32mi8 },
104 { X86::OR32rr, X86::OR32mr },
105 { X86::OR64ri32, X86::OR64mi32 },
106 { X86::OR64ri8, X86::OR64mi8 },
107 { X86::OR64rr, X86::OR64mr },
108 { X86::OR8ri, X86::OR8mi },
109 { X86::OR8rr, X86::OR8mr },
110 { X86::ROL16r1, X86::ROL16m1 },
111 { X86::ROL16rCL, X86::ROL16mCL },
112 { X86::ROL16ri, X86::ROL16mi },
113 { X86::ROL32r1, X86::ROL32m1 },
114 { X86::ROL32rCL, X86::ROL32mCL },
115 { X86::ROL32ri, X86::ROL32mi },
116 { X86::ROL64r1, X86::ROL64m1 },
117 { X86::ROL64rCL, X86::ROL64mCL },
118 { X86::ROL64ri, X86::ROL64mi },
119 { X86::ROL8r1, X86::ROL8m1 },
120 { X86::ROL8rCL, X86::ROL8mCL },
121 { X86::ROL8ri, X86::ROL8mi },
122 { X86::ROR16r1, X86::ROR16m1 },
123 { X86::ROR16rCL, X86::ROR16mCL },
124 { X86::ROR16ri, X86::ROR16mi },
125 { X86::ROR32r1, X86::ROR32m1 },
126 { X86::ROR32rCL, X86::ROR32mCL },
127 { X86::ROR32ri, X86::ROR32mi },
128 { X86::ROR64r1, X86::ROR64m1 },
129 { X86::ROR64rCL, X86::ROR64mCL },
130 { X86::ROR64ri, X86::ROR64mi },
131 { X86::ROR8r1, X86::ROR8m1 },
132 { X86::ROR8rCL, X86::ROR8mCL },
133 { X86::ROR8ri, X86::ROR8mi },
134 { X86::SAR16r1, X86::SAR16m1 },
135 { X86::SAR16rCL, X86::SAR16mCL },
136 { X86::SAR16ri, X86::SAR16mi },
137 { X86::SAR32r1, X86::SAR32m1 },
138 { X86::SAR32rCL, X86::SAR32mCL },
139 { X86::SAR32ri, X86::SAR32mi },
140 { X86::SAR64r1, X86::SAR64m1 },
141 { X86::SAR64rCL, X86::SAR64mCL },
142 { X86::SAR64ri, X86::SAR64mi },
143 { X86::SAR8r1, X86::SAR8m1 },
144 { X86::SAR8rCL, X86::SAR8mCL },
145 { X86::SAR8ri, X86::SAR8mi },
146 { X86::SBB32ri, X86::SBB32mi },
147 { X86::SBB32ri8, X86::SBB32mi8 },
148 { X86::SBB32rr, X86::SBB32mr },
149 { X86::SBB64ri32, X86::SBB64mi32 },
150 { X86::SBB64ri8, X86::SBB64mi8 },
151 { X86::SBB64rr, X86::SBB64mr },
152 { X86::SHL16rCL, X86::SHL16mCL },
153 { X86::SHL16ri, X86::SHL16mi },
154 { X86::SHL32rCL, X86::SHL32mCL },
155 { X86::SHL32ri, X86::SHL32mi },
156 { X86::SHL64rCL, X86::SHL64mCL },
157 { X86::SHL64ri, X86::SHL64mi },
158 { X86::SHL8rCL, X86::SHL8mCL },
159 { X86::SHL8ri, X86::SHL8mi },
160 { X86::SHLD16rrCL, X86::SHLD16mrCL },
161 { X86::SHLD16rri8, X86::SHLD16mri8 },
162 { X86::SHLD32rrCL, X86::SHLD32mrCL },
163 { X86::SHLD32rri8, X86::SHLD32mri8 },
164 { X86::SHLD64rrCL, X86::SHLD64mrCL },
165 { X86::SHLD64rri8, X86::SHLD64mri8 },
166 { X86::SHR16r1, X86::SHR16m1 },
167 { X86::SHR16rCL, X86::SHR16mCL },
168 { X86::SHR16ri, X86::SHR16mi },
169 { X86::SHR32r1, X86::SHR32m1 },
170 { X86::SHR32rCL, X86::SHR32mCL },
171 { X86::SHR32ri, X86::SHR32mi },
172 { X86::SHR64r1, X86::SHR64m1 },
173 { X86::SHR64rCL, X86::SHR64mCL },
174 { X86::SHR64ri, X86::SHR64mi },
175 { X86::SHR8r1, X86::SHR8m1 },
176 { X86::SHR8rCL, X86::SHR8mCL },
177 { X86::SHR8ri, X86::SHR8mi },
178 { X86::SHRD16rrCL, X86::SHRD16mrCL },
179 { X86::SHRD16rri8, X86::SHRD16mri8 },
180 { X86::SHRD32rrCL, X86::SHRD32mrCL },
181 { X86::SHRD32rri8, X86::SHRD32mri8 },
182 { X86::SHRD64rrCL, X86::SHRD64mrCL },
183 { X86::SHRD64rri8, X86::SHRD64mri8 },
184 { X86::SUB16ri, X86::SUB16mi },
185 { X86::SUB16ri8, X86::SUB16mi8 },
186 { X86::SUB16rr, X86::SUB16mr },
187 { X86::SUB32ri, X86::SUB32mi },
188 { X86::SUB32ri8, X86::SUB32mi8 },
189 { X86::SUB32rr, X86::SUB32mr },
190 { X86::SUB64ri32, X86::SUB64mi32 },
191 { X86::SUB64ri8, X86::SUB64mi8 },
192 { X86::SUB64rr, X86::SUB64mr },
193 { X86::SUB8ri, X86::SUB8mi },
194 { X86::SUB8rr, X86::SUB8mr },
195 { X86::XOR16ri, X86::XOR16mi },
196 { X86::XOR16ri8, X86::XOR16mi8 },
197 { X86::XOR16rr, X86::XOR16mr },
198 { X86::XOR32ri, X86::XOR32mi },
199 { X86::XOR32ri8, X86::XOR32mi8 },
200 { X86::XOR32rr, X86::XOR32mr },
201 { X86::XOR64ri32, X86::XOR64mi32 },
202 { X86::XOR64ri8, X86::XOR64mi8 },
203 { X86::XOR64rr, X86::XOR64mr },
204 { X86::XOR8ri, X86::XOR8mi },
205 { X86::XOR8rr, X86::XOR8mr }
208 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
209 unsigned RegOp = OpTbl2Addr[i][0];
210 unsigned MemOp = OpTbl2Addr[i][1];
211 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp, MemOp)))
212 assert(false && "Duplicated entries?");
213 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
214 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
215 std::make_pair(RegOp, AuxInfo))))
216 AmbEntries.push_back(MemOp);
219 // If the third value is 1, then it's folding either a load or a store.
220 static const unsigned OpTbl0[][3] = {
221 { X86::CALL32r, X86::CALL32m, 1 },
222 { X86::CALL64r, X86::CALL64m, 1 },
223 { X86::CMP16ri, X86::CMP16mi, 1 },
224 { X86::CMP16ri8, X86::CMP16mi8, 1 },
225 { X86::CMP16rr, X86::CMP16mr, 1 },
226 { X86::CMP32ri, X86::CMP32mi, 1 },
227 { X86::CMP32ri8, X86::CMP32mi8, 1 },
228 { X86::CMP32rr, X86::CMP32mr, 1 },
229 { X86::CMP64ri32, X86::CMP64mi32, 1 },
230 { X86::CMP64ri8, X86::CMP64mi8, 1 },
231 { X86::CMP64rr, X86::CMP64mr, 1 },
232 { X86::CMP8ri, X86::CMP8mi, 1 },
233 { X86::CMP8rr, X86::CMP8mr, 1 },
234 { X86::DIV16r, X86::DIV16m, 1 },
235 { X86::DIV32r, X86::DIV32m, 1 },
236 { X86::DIV64r, X86::DIV64m, 1 },
237 { X86::DIV8r, X86::DIV8m, 1 },
238 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
239 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
240 { X86::IDIV16r, X86::IDIV16m, 1 },
241 { X86::IDIV32r, X86::IDIV32m, 1 },
242 { X86::IDIV64r, X86::IDIV64m, 1 },
243 { X86::IDIV8r, X86::IDIV8m, 1 },
244 { X86::IMUL16r, X86::IMUL16m, 1 },
245 { X86::IMUL32r, X86::IMUL32m, 1 },
246 { X86::IMUL64r, X86::IMUL64m, 1 },
247 { X86::IMUL8r, X86::IMUL8m, 1 },
248 { X86::JMP32r, X86::JMP32m, 1 },
249 { X86::JMP64r, X86::JMP64m, 1 },
250 { X86::MOV16ri, X86::MOV16mi, 0 },
251 { X86::MOV16rr, X86::MOV16mr, 0 },
252 { X86::MOV16to16_, X86::MOV16_mr, 0 },
253 { X86::MOV32ri, X86::MOV32mi, 0 },
254 { X86::MOV32rr, X86::MOV32mr, 0 },
255 { X86::MOV32to32_, X86::MOV32_mr, 0 },
256 { X86::MOV64ri32, X86::MOV64mi32, 0 },
257 { X86::MOV64rr, X86::MOV64mr, 0 },
258 { X86::MOV8ri, X86::MOV8mi, 0 },
259 { X86::MOV8rr, X86::MOV8mr, 0 },
260 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
261 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
262 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
263 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
264 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
265 { X86::MOVSDrr, X86::MOVSDmr, 0 },
266 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
267 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
268 { X86::MOVSSrr, X86::MOVSSmr, 0 },
269 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
270 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
271 { X86::MUL16r, X86::MUL16m, 1 },
272 { X86::MUL32r, X86::MUL32m, 1 },
273 { X86::MUL64r, X86::MUL64m, 1 },
274 { X86::MUL8r, X86::MUL8m, 1 },
275 { X86::SETAEr, X86::SETAEm, 0 },
276 { X86::SETAr, X86::SETAm, 0 },
277 { X86::SETBEr, X86::SETBEm, 0 },
278 { X86::SETBr, X86::SETBm, 0 },
279 { X86::SETEr, X86::SETEm, 0 },
280 { X86::SETGEr, X86::SETGEm, 0 },
281 { X86::SETGr, X86::SETGm, 0 },
282 { X86::SETLEr, X86::SETLEm, 0 },
283 { X86::SETLr, X86::SETLm, 0 },
284 { X86::SETNEr, X86::SETNEm, 0 },
285 { X86::SETNPr, X86::SETNPm, 0 },
286 { X86::SETNSr, X86::SETNSm, 0 },
287 { X86::SETPr, X86::SETPm, 0 },
288 { X86::SETSr, X86::SETSm, 0 },
289 { X86::TAILJMPr, X86::TAILJMPm, 1 },
290 { X86::TEST16ri, X86::TEST16mi, 1 },
291 { X86::TEST32ri, X86::TEST32mi, 1 },
292 { X86::TEST64ri32, X86::TEST64mi32, 1 },
293 { X86::TEST8ri, X86::TEST8mi, 1 }
296 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
297 unsigned RegOp = OpTbl0[i][0];
298 unsigned MemOp = OpTbl0[i][1];
299 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp, MemOp)))
300 assert(false && "Duplicated entries?");
301 unsigned FoldedLoad = OpTbl0[i][2];
302 // Index 0, folded load or store.
303 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
304 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
305 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
306 std::make_pair(RegOp, AuxInfo))))
307 AmbEntries.push_back(MemOp);
310 static const unsigned OpTbl1[][2] = {
311 { X86::CMP16rr, X86::CMP16rm },
312 { X86::CMP32rr, X86::CMP32rm },
313 { X86::CMP64rr, X86::CMP64rm },
314 { X86::CMP8rr, X86::CMP8rm },
315 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
316 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
317 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
318 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
319 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
320 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
321 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
322 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
323 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
324 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
325 { X86::FsMOVAPDrr, X86::MOVSDrm },
326 { X86::FsMOVAPSrr, X86::MOVSSrm },
327 { X86::IMUL16rri, X86::IMUL16rmi },
328 { X86::IMUL16rri8, X86::IMUL16rmi8 },
329 { X86::IMUL32rri, X86::IMUL32rmi },
330 { X86::IMUL32rri8, X86::IMUL32rmi8 },
331 { X86::IMUL64rri32, X86::IMUL64rmi32 },
332 { X86::IMUL64rri8, X86::IMUL64rmi8 },
333 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
334 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
335 { X86::Int_COMISDrr, X86::Int_COMISDrm },
336 { X86::Int_COMISSrr, X86::Int_COMISSrm },
337 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
338 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
339 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
340 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
341 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
342 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
343 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
344 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
345 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
346 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
347 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
348 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
349 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
350 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
351 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
352 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
353 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
354 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
355 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
356 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
357 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
358 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
359 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
360 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
361 { X86::MOV16rr, X86::MOV16rm },
362 { X86::MOV16to16_, X86::MOV16_rm },
363 { X86::MOV32rr, X86::MOV32rm },
364 { X86::MOV32to32_, X86::MOV32_rm },
365 { X86::MOV64rr, X86::MOV64rm },
366 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
367 { X86::MOV64toSDrr, X86::MOV64toSDrm },
368 { X86::MOV8rr, X86::MOV8rm },
369 { X86::MOVAPDrr, X86::MOVAPDrm },
370 { X86::MOVAPSrr, X86::MOVAPSrm },
371 { X86::MOVDDUPrr, X86::MOVDDUPrm },
372 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
373 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
374 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
375 { X86::MOVSDrr, X86::MOVSDrm },
376 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
377 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
378 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
379 { X86::MOVSSrr, X86::MOVSSrm },
380 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
381 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
382 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
383 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
384 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
385 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
386 { X86::MOVUPDrr, X86::MOVUPDrm },
387 { X86::MOVUPSrr, X86::MOVUPSrm },
388 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
389 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
390 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
391 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
392 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
393 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
394 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
395 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
396 { X86::PSHUFDri, X86::PSHUFDmi },
397 { X86::PSHUFHWri, X86::PSHUFHWmi },
398 { X86::PSHUFLWri, X86::PSHUFLWmi },
399 { X86::RCPPSr, X86::RCPPSm },
400 { X86::RCPPSr_Int, X86::RCPPSm_Int },
401 { X86::RSQRTPSr, X86::RSQRTPSm },
402 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
403 { X86::RSQRTSSr, X86::RSQRTSSm },
404 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
405 { X86::SQRTPDr, X86::SQRTPDm },
406 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
407 { X86::SQRTPSr, X86::SQRTPSm },
408 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
409 { X86::SQRTSDr, X86::SQRTSDm },
410 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
411 { X86::SQRTSSr, X86::SQRTSSm },
412 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
413 { X86::TEST16rr, X86::TEST16rm },
414 { X86::TEST32rr, X86::TEST32rm },
415 { X86::TEST64rr, X86::TEST64rm },
416 { X86::TEST8rr, X86::TEST8rm },
417 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
418 { X86::UCOMISDrr, X86::UCOMISDrm },
419 { X86::UCOMISSrr, X86::UCOMISSrm }
422 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
423 unsigned RegOp = OpTbl1[i][0];
424 unsigned MemOp = OpTbl1[i][1];
425 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp, MemOp)))
426 assert(false && "Duplicated entries?");
427 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
428 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
429 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
430 std::make_pair(RegOp, AuxInfo))))
431 AmbEntries.push_back(MemOp);
434 static const unsigned OpTbl2[][2] = {
435 { X86::ADC32rr, X86::ADC32rm },
436 { X86::ADC64rr, X86::ADC64rm },
437 { X86::ADD16rr, X86::ADD16rm },
438 { X86::ADD32rr, X86::ADD32rm },
439 { X86::ADD64rr, X86::ADD64rm },
440 { X86::ADD8rr, X86::ADD8rm },
441 { X86::ADDPDrr, X86::ADDPDrm },
442 { X86::ADDPSrr, X86::ADDPSrm },
443 { X86::ADDSDrr, X86::ADDSDrm },
444 { X86::ADDSSrr, X86::ADDSSrm },
445 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
446 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
447 { X86::AND16rr, X86::AND16rm },
448 { X86::AND32rr, X86::AND32rm },
449 { X86::AND64rr, X86::AND64rm },
450 { X86::AND8rr, X86::AND8rm },
451 { X86::ANDNPDrr, X86::ANDNPDrm },
452 { X86::ANDNPSrr, X86::ANDNPSrm },
453 { X86::ANDPDrr, X86::ANDPDrm },
454 { X86::ANDPSrr, X86::ANDPSrm },
455 { X86::CMOVA16rr, X86::CMOVA16rm },
456 { X86::CMOVA32rr, X86::CMOVA32rm },
457 { X86::CMOVA64rr, X86::CMOVA64rm },
458 { X86::CMOVAE16rr, X86::CMOVAE16rm },
459 { X86::CMOVAE32rr, X86::CMOVAE32rm },
460 { X86::CMOVAE64rr, X86::CMOVAE64rm },
461 { X86::CMOVB16rr, X86::CMOVB16rm },
462 { X86::CMOVB32rr, X86::CMOVB32rm },
463 { X86::CMOVB64rr, X86::CMOVB64rm },
464 { X86::CMOVBE16rr, X86::CMOVBE16rm },
465 { X86::CMOVBE32rr, X86::CMOVBE32rm },
466 { X86::CMOVBE64rr, X86::CMOVBE64rm },
467 { X86::CMOVE16rr, X86::CMOVE16rm },
468 { X86::CMOVE32rr, X86::CMOVE32rm },
469 { X86::CMOVE64rr, X86::CMOVE64rm },
470 { X86::CMOVG16rr, X86::CMOVG16rm },
471 { X86::CMOVG32rr, X86::CMOVG32rm },
472 { X86::CMOVG64rr, X86::CMOVG64rm },
473 { X86::CMOVGE16rr, X86::CMOVGE16rm },
474 { X86::CMOVGE32rr, X86::CMOVGE32rm },
475 { X86::CMOVGE64rr, X86::CMOVGE64rm },
476 { X86::CMOVL16rr, X86::CMOVL16rm },
477 { X86::CMOVL32rr, X86::CMOVL32rm },
478 { X86::CMOVL64rr, X86::CMOVL64rm },
479 { X86::CMOVLE16rr, X86::CMOVLE16rm },
480 { X86::CMOVLE32rr, X86::CMOVLE32rm },
481 { X86::CMOVLE64rr, X86::CMOVLE64rm },
482 { X86::CMOVNE16rr, X86::CMOVNE16rm },
483 { X86::CMOVNE32rr, X86::CMOVNE32rm },
484 { X86::CMOVNE64rr, X86::CMOVNE64rm },
485 { X86::CMOVNP16rr, X86::CMOVNP16rm },
486 { X86::CMOVNP32rr, X86::CMOVNP32rm },
487 { X86::CMOVNP64rr, X86::CMOVNP64rm },
488 { X86::CMOVNS16rr, X86::CMOVNS16rm },
489 { X86::CMOVNS32rr, X86::CMOVNS32rm },
490 { X86::CMOVNS64rr, X86::CMOVNS64rm },
491 { X86::CMOVP16rr, X86::CMOVP16rm },
492 { X86::CMOVP32rr, X86::CMOVP32rm },
493 { X86::CMOVP64rr, X86::CMOVP64rm },
494 { X86::CMOVS16rr, X86::CMOVS16rm },
495 { X86::CMOVS32rr, X86::CMOVS32rm },
496 { X86::CMOVS64rr, X86::CMOVS64rm },
497 { X86::CMPPDrri, X86::CMPPDrmi },
498 { X86::CMPPSrri, X86::CMPPSrmi },
499 { X86::CMPSDrr, X86::CMPSDrm },
500 { X86::CMPSSrr, X86::CMPSSrm },
501 { X86::DIVPDrr, X86::DIVPDrm },
502 { X86::DIVPSrr, X86::DIVPSrm },
503 { X86::DIVSDrr, X86::DIVSDrm },
504 { X86::DIVSSrr, X86::DIVSSrm },
505 { X86::FsANDNPDrr, X86::FsANDNPDrm },
506 { X86::FsANDNPSrr, X86::FsANDNPSrm },
507 { X86::FsANDPDrr, X86::FsANDPDrm },
508 { X86::FsANDPSrr, X86::FsANDPSrm },
509 { X86::FsORPDrr, X86::FsORPDrm },
510 { X86::FsORPSrr, X86::FsORPSrm },
511 { X86::FsXORPDrr, X86::FsXORPDrm },
512 { X86::FsXORPSrr, X86::FsXORPSrm },
513 { X86::HADDPDrr, X86::HADDPDrm },
514 { X86::HADDPSrr, X86::HADDPSrm },
515 { X86::HSUBPDrr, X86::HSUBPDrm },
516 { X86::HSUBPSrr, X86::HSUBPSrm },
517 { X86::IMUL16rr, X86::IMUL16rm },
518 { X86::IMUL32rr, X86::IMUL32rm },
519 { X86::IMUL64rr, X86::IMUL64rm },
520 { X86::MAXPDrr, X86::MAXPDrm },
521 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
522 { X86::MAXPSrr, X86::MAXPSrm },
523 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
524 { X86::MAXSDrr, X86::MAXSDrm },
525 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
526 { X86::MAXSSrr, X86::MAXSSrm },
527 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
528 { X86::MINPDrr, X86::MINPDrm },
529 { X86::MINPDrr_Int, X86::MINPDrm_Int },
530 { X86::MINPSrr, X86::MINPSrm },
531 { X86::MINPSrr_Int, X86::MINPSrm_Int },
532 { X86::MINSDrr, X86::MINSDrm },
533 { X86::MINSDrr_Int, X86::MINSDrm_Int },
534 { X86::MINSSrr, X86::MINSSrm },
535 { X86::MINSSrr_Int, X86::MINSSrm_Int },
536 { X86::MULPDrr, X86::MULPDrm },
537 { X86::MULPSrr, X86::MULPSrm },
538 { X86::MULSDrr, X86::MULSDrm },
539 { X86::MULSSrr, X86::MULSSrm },
540 { X86::OR16rr, X86::OR16rm },
541 { X86::OR32rr, X86::OR32rm },
542 { X86::OR64rr, X86::OR64rm },
543 { X86::OR8rr, X86::OR8rm },
544 { X86::ORPDrr, X86::ORPDrm },
545 { X86::ORPSrr, X86::ORPSrm },
546 { X86::PACKSSDWrr, X86::PACKSSDWrm },
547 { X86::PACKSSWBrr, X86::PACKSSWBrm },
548 { X86::PACKUSWBrr, X86::PACKUSWBrm },
549 { X86::PADDBrr, X86::PADDBrm },
550 { X86::PADDDrr, X86::PADDDrm },
551 { X86::PADDQrr, X86::PADDQrm },
552 { X86::PADDSBrr, X86::PADDSBrm },
553 { X86::PADDSWrr, X86::PADDSWrm },
554 { X86::PADDWrr, X86::PADDWrm },
555 { X86::PANDNrr, X86::PANDNrm },
556 { X86::PANDrr, X86::PANDrm },
557 { X86::PAVGBrr, X86::PAVGBrm },
558 { X86::PAVGWrr, X86::PAVGWrm },
559 { X86::PCMPEQBrr, X86::PCMPEQBrm },
560 { X86::PCMPEQDrr, X86::PCMPEQDrm },
561 { X86::PCMPEQWrr, X86::PCMPEQWrm },
562 { X86::PCMPGTBrr, X86::PCMPGTBrm },
563 { X86::PCMPGTDrr, X86::PCMPGTDrm },
564 { X86::PCMPGTWrr, X86::PCMPGTWrm },
565 { X86::PINSRWrri, X86::PINSRWrmi },
566 { X86::PMADDWDrr, X86::PMADDWDrm },
567 { X86::PMAXSWrr, X86::PMAXSWrm },
568 { X86::PMAXUBrr, X86::PMAXUBrm },
569 { X86::PMINSWrr, X86::PMINSWrm },
570 { X86::PMINUBrr, X86::PMINUBrm },
571 { X86::PMULHUWrr, X86::PMULHUWrm },
572 { X86::PMULHWrr, X86::PMULHWrm },
573 { X86::PMULLWrr, X86::PMULLWrm },
574 { X86::PMULUDQrr, X86::PMULUDQrm },
575 { X86::PORrr, X86::PORrm },
576 { X86::PSADBWrr, X86::PSADBWrm },
577 { X86::PSLLDrr, X86::PSLLDrm },
578 { X86::PSLLQrr, X86::PSLLQrm },
579 { X86::PSLLWrr, X86::PSLLWrm },
580 { X86::PSRADrr, X86::PSRADrm },
581 { X86::PSRAWrr, X86::PSRAWrm },
582 { X86::PSRLDrr, X86::PSRLDrm },
583 { X86::PSRLQrr, X86::PSRLQrm },
584 { X86::PSRLWrr, X86::PSRLWrm },
585 { X86::PSUBBrr, X86::PSUBBrm },
586 { X86::PSUBDrr, X86::PSUBDrm },
587 { X86::PSUBSBrr, X86::PSUBSBrm },
588 { X86::PSUBSWrr, X86::PSUBSWrm },
589 { X86::PSUBWrr, X86::PSUBWrm },
590 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
591 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
592 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
593 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
594 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
595 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
596 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
597 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
598 { X86::PXORrr, X86::PXORrm },
599 { X86::SBB32rr, X86::SBB32rm },
600 { X86::SBB64rr, X86::SBB64rm },
601 { X86::SHUFPDrri, X86::SHUFPDrmi },
602 { X86::SHUFPSrri, X86::SHUFPSrmi },
603 { X86::SUB16rr, X86::SUB16rm },
604 { X86::SUB32rr, X86::SUB32rm },
605 { X86::SUB64rr, X86::SUB64rm },
606 { X86::SUB8rr, X86::SUB8rm },
607 { X86::SUBPDrr, X86::SUBPDrm },
608 { X86::SUBPSrr, X86::SUBPSrm },
609 { X86::SUBSDrr, X86::SUBSDrm },
610 { X86::SUBSSrr, X86::SUBSSrm },
611 // FIXME: TEST*rr -> swapped operand of TEST*mr.
612 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
613 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
614 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
615 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
616 { X86::XOR16rr, X86::XOR16rm },
617 { X86::XOR32rr, X86::XOR32rm },
618 { X86::XOR64rr, X86::XOR64rm },
619 { X86::XOR8rr, X86::XOR8rm },
620 { X86::XORPDrr, X86::XORPDrm },
621 { X86::XORPSrr, X86::XORPSrm }
624 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
625 unsigned RegOp = OpTbl2[i][0];
626 unsigned MemOp = OpTbl2[i][1];
627 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp, MemOp)))
628 assert(false && "Duplicated entries?");
629 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
630 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
631 std::make_pair(RegOp, AuxInfo))))
632 AmbEntries.push_back(MemOp);
635 // Remove ambiguous entries.
636 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
639 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
641 unsigned& destReg) const {
642 switch (MI.getOpcode()) {
649 case X86::MOV16to16_:
650 case X86::MOV32to32_:
654 // FP Stack register class copies
655 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
656 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
657 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
659 case X86::FsMOVAPSrr:
660 case X86::FsMOVAPDrr:
663 case X86::MOVSS2PSrr:
664 case X86::MOVSD2PDrr:
665 case X86::MOVPS2SSrr:
666 case X86::MOVPD2SDrr:
667 case X86::MMX_MOVD64rr:
668 case X86::MMX_MOVQ64rr:
669 assert(MI.getNumOperands() >= 2 &&
670 MI.getOperand(0).isRegister() &&
671 MI.getOperand(1).isRegister() &&
672 "invalid register-register move instruction");
673 sourceReg = MI.getOperand(1).getReg();
674 destReg = MI.getOperand(0).getReg();
679 unsigned X86InstrInfo::isLoadFromStackSlot(MachineInstr *MI,
680 int &FrameIndex) const {
681 switch (MI->getOpcode()) {
694 case X86::MMX_MOVD64rm:
695 case X86::MMX_MOVQ64rm:
696 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
697 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
698 MI->getOperand(2).getImm() == 1 &&
699 MI->getOperand(3).getReg() == 0 &&
700 MI->getOperand(4).getImm() == 0) {
701 FrameIndex = MI->getOperand(1).getIndex();
702 return MI->getOperand(0).getReg();
709 unsigned X86InstrInfo::isStoreToStackSlot(MachineInstr *MI,
710 int &FrameIndex) const {
711 switch (MI->getOpcode()) {
724 case X86::MMX_MOVD64mr:
725 case X86::MMX_MOVQ64mr:
726 case X86::MMX_MOVNTQmr:
727 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
728 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
729 MI->getOperand(1).getImm() == 1 &&
730 MI->getOperand(2).getReg() == 0 &&
731 MI->getOperand(3).getImm() == 0) {
732 FrameIndex = MI->getOperand(0).getIndex();
733 return MI->getOperand(4).getReg();
741 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
743 static bool regIsPICBase(unsigned BaseReg, MachineRegisterInfo &MRI) {
744 bool isPICBase = false;
745 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
746 E = MRI.def_end(); I != E; ++I) {
747 MachineInstr *DefMI = I.getOperand().getParent();
748 if (DefMI->getOpcode() != X86::MOVPC32r)
750 assert(!isPICBase && "More than one PIC base?");
756 /// isGVStub - Return true if the GV requires an extra load to get the
758 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
759 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
762 bool X86InstrInfo::isReallyTriviallyReMaterializable(MachineInstr *MI) const {
763 switch (MI->getOpcode()) {
776 case X86::MMX_MOVD64rm:
777 case X86::MMX_MOVQ64rm: {
778 // Loads from constant pools are trivially rematerializable.
779 if (MI->getOperand(1).isReg() &&
780 MI->getOperand(2).isImm() &&
781 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
782 (MI->getOperand(4).isCPI() ||
783 (MI->getOperand(4).isGlobal() &&
784 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
785 unsigned BaseReg = MI->getOperand(1).getReg();
788 // Allow re-materialization of PIC load.
789 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
791 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
792 bool isPICBase = false;
793 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
794 E = MRI.def_end(); I != E; ++I) {
795 MachineInstr *DefMI = I.getOperand().getParent();
796 if (DefMI->getOpcode() != X86::MOVPC32r)
798 assert(!isPICBase && "More than one PIC base?");
808 if (MI->getOperand(1).isReg() &&
809 MI->getOperand(2).isImm() &&
810 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
811 !MI->getOperand(4).isReg()) {
812 // lea fi#, lea GV, etc. are all rematerializable.
813 unsigned BaseReg = MI->getOperand(1).getReg();
816 // Allow re-materialization of lea PICBase + x.
817 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
818 return regIsPICBase(BaseReg, MRI);
824 // All other instructions marked M_REMATERIALIZABLE are always trivially
829 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
830 MachineBasicBlock::iterator I,
832 const MachineInstr *Orig) const {
833 // MOV32r0 etc. are implemented with xor which clobbers condition code.
834 // Re-materialize them as movri instructions to avoid side effects.
835 switch (Orig->getOpcode()) {
837 BuildMI(MBB, I, get(X86::MOV8ri), DestReg).addImm(0);
840 BuildMI(MBB, I, get(X86::MOV16ri), DestReg).addImm(0);
843 BuildMI(MBB, I, get(X86::MOV32ri), DestReg).addImm(0);
846 BuildMI(MBB, I, get(X86::MOV64ri32), DestReg).addImm(0);
849 MachineInstr *MI = Orig->clone();
850 MI->getOperand(0).setReg(DestReg);
857 /// isInvariantLoad - Return true if the specified instruction (which is marked
858 /// mayLoad) is loading from a location whose value is invariant across the
859 /// function. For example, loading a value from the constant pool or from
860 /// from the argument area of a function if it does not change. This should
861 /// only return true of *all* loads the instruction does are invariant (if it
862 /// does multiple loads).
863 bool X86InstrInfo::isInvariantLoad(MachineInstr *MI) const {
864 // This code cares about loads from three cases: constant pool entries,
865 // invariant argument slots, and global stubs. In order to handle these cases
866 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
867 // operand and base our analysis on it. This is safe because the address of
868 // none of these three cases is ever used as anything other than a load base
869 // and X86 doesn't have any instructions that load from multiple places.
871 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
872 const MachineOperand &MO = MI->getOperand(i);
873 // Loads from constant pools are trivially invariant.
878 return isGVStub(MO.getGlobal(), TM);
880 // If this is a load from an invariant stack slot, the load is a constant.
882 const MachineFrameInfo &MFI =
883 *MI->getParent()->getParent()->getFrameInfo();
884 int Idx = MO.getIndex();
885 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
889 // All other instances of these instructions are presumed to have other
894 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
895 /// is not marked dead.
896 static bool hasLiveCondCodeDef(MachineInstr *MI) {
897 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
898 MachineOperand &MO = MI->getOperand(i);
899 if (MO.isRegister() && MO.isDef() &&
900 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
907 /// convertToThreeAddress - This method must be implemented by targets that
908 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
909 /// may be able to convert a two-address instruction into a true
910 /// three-address instruction on demand. This allows the X86 target (for
911 /// example) to convert ADD and SHL instructions into LEA instructions if they
912 /// would require register copies due to two-addressness.
914 /// This method returns a null pointer if the transformation cannot be
915 /// performed, otherwise it returns the new instruction.
918 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
919 MachineBasicBlock::iterator &MBBI,
920 LiveVariables &LV) const {
921 MachineInstr *MI = MBBI;
922 // All instructions input are two-addr instructions. Get the known operands.
923 unsigned Dest = MI->getOperand(0).getReg();
924 unsigned Src = MI->getOperand(1).getReg();
926 MachineInstr *NewMI = NULL;
927 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
928 // we have better subtarget support, enable the 16-bit LEA generation here.
929 bool DisableLEA16 = true;
931 unsigned MIOpc = MI->getOpcode();
933 case X86::SHUFPSrri: {
934 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
935 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
937 unsigned A = MI->getOperand(0).getReg();
938 unsigned B = MI->getOperand(1).getReg();
939 unsigned C = MI->getOperand(2).getReg();
940 unsigned M = MI->getOperand(3).getImm();
941 if (B != C) return 0;
942 NewMI = BuildMI(get(X86::PSHUFDri), A).addReg(B).addImm(M);
946 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
947 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
948 // the flags produced by a shift yet, so this is safe.
949 unsigned Dest = MI->getOperand(0).getReg();
950 unsigned Src = MI->getOperand(1).getReg();
951 unsigned ShAmt = MI->getOperand(2).getImm();
952 if (ShAmt == 0 || ShAmt >= 4) return 0;
954 NewMI = BuildMI(get(X86::LEA64r), Dest)
955 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
959 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
960 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
961 // the flags produced by a shift yet, so this is safe.
962 unsigned Dest = MI->getOperand(0).getReg();
963 unsigned Src = MI->getOperand(1).getReg();
964 unsigned ShAmt = MI->getOperand(2).getImm();
965 if (ShAmt == 0 || ShAmt >= 4) return 0;
967 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
968 X86::LEA64_32r : X86::LEA32r;
969 NewMI = BuildMI(get(Opc), Dest)
970 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
974 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
975 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
976 // the flags produced by a shift yet, so this is safe.
977 unsigned Dest = MI->getOperand(0).getReg();
978 unsigned Src = MI->getOperand(1).getReg();
979 unsigned ShAmt = MI->getOperand(2).getImm();
980 if (ShAmt == 0 || ShAmt >= 4) return 0;
983 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
984 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
985 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
986 ? X86::LEA64_32r : X86::LEA32r;
987 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
988 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
990 // Build and insert into an implicit UNDEF value. This is OK because
991 // well be shifting and then extracting the lower 16-bits.
992 MachineInstr *Undef = BuildMI(get(X86::IMPLICIT_DEF), leaInReg);
995 BuildMI(get(X86::INSERT_SUBREG),leaInReg)
996 .addReg(leaInReg).addReg(Src).addImm(X86::SUBREG_16BIT);
998 NewMI = BuildMI(get(Opc), leaOutReg)
999 .addReg(0).addImm(1 << ShAmt).addReg(leaInReg).addImm(0);
1002 BuildMI(get(X86::EXTRACT_SUBREG), Dest)
1003 .addReg(leaOutReg).addImm(X86::SUBREG_16BIT);
1004 Ext->copyKillDeadInfo(MI);
1006 MFI->insert(MBBI, Undef);
1007 MFI->insert(MBBI, Ins); // Insert the insert_subreg
1008 LV.instructionChanged(MI, NewMI); // Update live variables
1009 LV.addVirtualRegisterKilled(leaInReg, NewMI);
1010 MFI->insert(MBBI, NewMI); // Insert the new inst
1011 LV.addVirtualRegisterKilled(leaOutReg, Ext);
1012 MFI->insert(MBBI, Ext); // Insert the extract_subreg
1015 NewMI = BuildMI(get(X86::LEA16r), Dest)
1016 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
1021 // The following opcodes also sets the condition code register(s). Only
1022 // convert them to equivalent lea if the condition code register def's
1024 if (hasLiveCondCodeDef(MI))
1027 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1032 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1033 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1034 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1035 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, 1);
1039 case X86::INC64_16r:
1040 if (DisableLEA16) return 0;
1041 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1042 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, 1);
1046 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1047 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1048 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1049 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, -1);
1053 case X86::DEC64_16r:
1054 if (DisableLEA16) return 0;
1055 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1056 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, -1);
1059 case X86::ADD32rr: {
1060 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1061 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1062 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1063 NewMI = addRegReg(BuildMI(get(Opc), Dest), Src,
1064 MI->getOperand(2).getReg());
1068 if (DisableLEA16) return 0;
1069 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1070 NewMI = addRegReg(BuildMI(get(X86::LEA16r), Dest), Src,
1071 MI->getOperand(2).getReg());
1073 case X86::ADD64ri32:
1075 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1076 if (MI->getOperand(2).isImmediate())
1077 NewMI = addRegOffset(BuildMI(get(X86::LEA64r), Dest), Src,
1078 MI->getOperand(2).getImm());
1082 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1083 if (MI->getOperand(2).isImmediate()) {
1084 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1085 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src,
1086 MI->getOperand(2).getImm());
1091 if (DisableLEA16) return 0;
1092 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1093 if (MI->getOperand(2).isImmediate())
1094 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src,
1095 MI->getOperand(2).getImm());
1098 if (DisableLEA16) return 0;
1100 case X86::SHL64ri: {
1101 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImmediate() &&
1102 "Unknown shl instruction!");
1103 unsigned ShAmt = MI->getOperand(2).getImm();
1104 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1106 AM.Scale = 1 << ShAmt;
1108 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1109 : (MIOpc == X86::SHL32ri
1110 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1111 NewMI = addFullAddress(BuildMI(get(Opc), Dest), AM);
1119 if (!NewMI) return 0;
1121 NewMI->copyKillDeadInfo(MI);
1122 LV.instructionChanged(MI, NewMI); // Update live variables
1123 MFI->insert(MBBI, NewMI); // Insert the new inst
1127 /// commuteInstruction - We have a few instructions that must be hacked on to
1130 MachineInstr *X86InstrInfo::commuteInstruction(MachineInstr *MI) const {
1131 switch (MI->getOpcode()) {
1132 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1133 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1134 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1135 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1136 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1137 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1140 switch (MI->getOpcode()) {
1141 default: assert(0 && "Unreachable!");
1142 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1143 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1144 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1145 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1146 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1147 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1149 unsigned Amt = MI->getOperand(3).getImm();
1150 unsigned A = MI->getOperand(0).getReg();
1151 unsigned B = MI->getOperand(1).getReg();
1152 unsigned C = MI->getOperand(2).getReg();
1153 bool BisKill = MI->getOperand(1).isKill();
1154 bool CisKill = MI->getOperand(2).isKill();
1155 // If machine instrs are no longer in two-address forms, update
1156 // destination register as well.
1158 // Must be two address instruction!
1159 assert(MI->getDesc().getOperandConstraint(0, TOI::TIED_TO) &&
1160 "Expecting a two-address instruction!");
1164 return BuildMI(get(Opc), A).addReg(C, false, false, CisKill)
1165 .addReg(B, false, false, BisKill).addImm(Size-Amt);
1167 case X86::CMOVB16rr:
1168 case X86::CMOVB32rr:
1169 case X86::CMOVB64rr:
1170 case X86::CMOVAE16rr:
1171 case X86::CMOVAE32rr:
1172 case X86::CMOVAE64rr:
1173 case X86::CMOVE16rr:
1174 case X86::CMOVE32rr:
1175 case X86::CMOVE64rr:
1176 case X86::CMOVNE16rr:
1177 case X86::CMOVNE32rr:
1178 case X86::CMOVNE64rr:
1179 case X86::CMOVBE16rr:
1180 case X86::CMOVBE32rr:
1181 case X86::CMOVBE64rr:
1182 case X86::CMOVA16rr:
1183 case X86::CMOVA32rr:
1184 case X86::CMOVA64rr:
1185 case X86::CMOVL16rr:
1186 case X86::CMOVL32rr:
1187 case X86::CMOVL64rr:
1188 case X86::CMOVGE16rr:
1189 case X86::CMOVGE32rr:
1190 case X86::CMOVGE64rr:
1191 case X86::CMOVLE16rr:
1192 case X86::CMOVLE32rr:
1193 case X86::CMOVLE64rr:
1194 case X86::CMOVG16rr:
1195 case X86::CMOVG32rr:
1196 case X86::CMOVG64rr:
1197 case X86::CMOVS16rr:
1198 case X86::CMOVS32rr:
1199 case X86::CMOVS64rr:
1200 case X86::CMOVNS16rr:
1201 case X86::CMOVNS32rr:
1202 case X86::CMOVNS64rr:
1203 case X86::CMOVP16rr:
1204 case X86::CMOVP32rr:
1205 case X86::CMOVP64rr:
1206 case X86::CMOVNP16rr:
1207 case X86::CMOVNP32rr:
1208 case X86::CMOVNP64rr: {
1210 switch (MI->getOpcode()) {
1212 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1213 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1214 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1215 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1216 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1217 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1218 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1219 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1220 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1221 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1222 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1223 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1224 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1225 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1226 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1227 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1228 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1229 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1230 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1231 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1232 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1233 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1234 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1235 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1236 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1237 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1238 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1239 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1240 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1241 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1242 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1243 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1244 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1245 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1246 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1247 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1248 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1249 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1250 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1251 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1252 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1253 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1256 MI->setDesc(get(Opc));
1257 // Fallthrough intended.
1260 return TargetInstrInfoImpl::commuteInstruction(MI);
1264 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1266 default: return X86::COND_INVALID;
1267 case X86::JE: return X86::COND_E;
1268 case X86::JNE: return X86::COND_NE;
1269 case X86::JL: return X86::COND_L;
1270 case X86::JLE: return X86::COND_LE;
1271 case X86::JG: return X86::COND_G;
1272 case X86::JGE: return X86::COND_GE;
1273 case X86::JB: return X86::COND_B;
1274 case X86::JBE: return X86::COND_BE;
1275 case X86::JA: return X86::COND_A;
1276 case X86::JAE: return X86::COND_AE;
1277 case X86::JS: return X86::COND_S;
1278 case X86::JNS: return X86::COND_NS;
1279 case X86::JP: return X86::COND_P;
1280 case X86::JNP: return X86::COND_NP;
1281 case X86::JO: return X86::COND_O;
1282 case X86::JNO: return X86::COND_NO;
1286 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1288 default: assert(0 && "Illegal condition code!");
1289 case X86::COND_E: return X86::JE;
1290 case X86::COND_NE: return X86::JNE;
1291 case X86::COND_L: return X86::JL;
1292 case X86::COND_LE: return X86::JLE;
1293 case X86::COND_G: return X86::JG;
1294 case X86::COND_GE: return X86::JGE;
1295 case X86::COND_B: return X86::JB;
1296 case X86::COND_BE: return X86::JBE;
1297 case X86::COND_A: return X86::JA;
1298 case X86::COND_AE: return X86::JAE;
1299 case X86::COND_S: return X86::JS;
1300 case X86::COND_NS: return X86::JNS;
1301 case X86::COND_P: return X86::JP;
1302 case X86::COND_NP: return X86::JNP;
1303 case X86::COND_O: return X86::JO;
1304 case X86::COND_NO: return X86::JNO;
1308 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1309 /// e.g. turning COND_E to COND_NE.
1310 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1312 default: assert(0 && "Illegal condition code!");
1313 case X86::COND_E: return X86::COND_NE;
1314 case X86::COND_NE: return X86::COND_E;
1315 case X86::COND_L: return X86::COND_GE;
1316 case X86::COND_LE: return X86::COND_G;
1317 case X86::COND_G: return X86::COND_LE;
1318 case X86::COND_GE: return X86::COND_L;
1319 case X86::COND_B: return X86::COND_AE;
1320 case X86::COND_BE: return X86::COND_A;
1321 case X86::COND_A: return X86::COND_BE;
1322 case X86::COND_AE: return X86::COND_B;
1323 case X86::COND_S: return X86::COND_NS;
1324 case X86::COND_NS: return X86::COND_S;
1325 case X86::COND_P: return X86::COND_NP;
1326 case X86::COND_NP: return X86::COND_P;
1327 case X86::COND_O: return X86::COND_NO;
1328 case X86::COND_NO: return X86::COND_O;
1332 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1333 const TargetInstrDesc &TID = MI->getDesc();
1334 if (!TID.isTerminator()) return false;
1336 // Conditional branch is a special case.
1337 if (TID.isBranch() && !TID.isBarrier())
1339 if (!TID.isPredicable())
1341 return !isPredicated(MI);
1344 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1345 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1346 const X86InstrInfo &TII) {
1347 if (MI->getOpcode() == X86::FP_REG_KILL)
1349 return TII.isUnpredicatedTerminator(MI);
1352 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1353 MachineBasicBlock *&TBB,
1354 MachineBasicBlock *&FBB,
1355 std::vector<MachineOperand> &Cond) const {
1356 // If the block has no terminators, it just falls into the block after it.
1357 MachineBasicBlock::iterator I = MBB.end();
1358 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this))
1361 // Get the last instruction in the block.
1362 MachineInstr *LastInst = I;
1364 // If there is only one terminator instruction, process it.
1365 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this)) {
1366 if (!LastInst->getDesc().isBranch())
1369 // If the block ends with a branch there are 3 possibilities:
1370 // it's an unconditional, conditional, or indirect branch.
1372 if (LastInst->getOpcode() == X86::JMP) {
1373 TBB = LastInst->getOperand(0).getMBB();
1376 X86::CondCode BranchCode = GetCondFromBranchOpc(LastInst->getOpcode());
1377 if (BranchCode == X86::COND_INVALID)
1378 return true; // Can't handle indirect branch.
1380 // Otherwise, block ends with fall-through condbranch.
1381 TBB = LastInst->getOperand(0).getMBB();
1382 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1386 // Get the instruction before it if it's a terminator.
1387 MachineInstr *SecondLastInst = I;
1389 // If there are three terminators, we don't know what sort of block this is.
1390 if (SecondLastInst && I != MBB.begin() &&
1391 isBrAnalysisUnpredicatedTerminator(--I, *this))
1394 // If the block ends with X86::JMP and a conditional branch, handle it.
1395 X86::CondCode BranchCode = GetCondFromBranchOpc(SecondLastInst->getOpcode());
1396 if (BranchCode != X86::COND_INVALID && LastInst->getOpcode() == X86::JMP) {
1397 TBB = SecondLastInst->getOperand(0).getMBB();
1398 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1399 FBB = LastInst->getOperand(0).getMBB();
1403 // If the block ends with two X86::JMPs, handle it. The second one is not
1404 // executed, so remove it.
1405 if (SecondLastInst->getOpcode() == X86::JMP &&
1406 LastInst->getOpcode() == X86::JMP) {
1407 TBB = SecondLastInst->getOperand(0).getMBB();
1409 I->eraseFromParent();
1413 // Otherwise, can't handle this.
1417 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1418 MachineBasicBlock::iterator I = MBB.end();
1419 if (I == MBB.begin()) return 0;
1421 if (I->getOpcode() != X86::JMP &&
1422 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1425 // Remove the branch.
1426 I->eraseFromParent();
1430 if (I == MBB.begin()) return 1;
1432 if (GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1435 // Remove the branch.
1436 I->eraseFromParent();
1440 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
1441 MachineOperand &MO) {
1442 if (MO.isRegister())
1443 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
1444 false, false, MO.getSubReg());
1445 else if (MO.isImmediate())
1446 MIB = MIB.addImm(MO.getImm());
1447 else if (MO.isFrameIndex())
1448 MIB = MIB.addFrameIndex(MO.getIndex());
1449 else if (MO.isGlobalAddress())
1450 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
1451 else if (MO.isConstantPoolIndex())
1452 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
1453 else if (MO.isJumpTableIndex())
1454 MIB = MIB.addJumpTableIndex(MO.getIndex());
1455 else if (MO.isExternalSymbol())
1456 MIB = MIB.addExternalSymbol(MO.getSymbolName());
1458 assert(0 && "Unknown operand for X86InstrAddOperand!");
1464 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1465 MachineBasicBlock *FBB,
1466 const std::vector<MachineOperand> &Cond) const {
1467 // Shouldn't be a fall through.
1468 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1469 assert((Cond.size() == 1 || Cond.size() == 0) &&
1470 "X86 branch conditions have one component!");
1472 if (FBB == 0) { // One way branch.
1474 // Unconditional branch?
1475 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
1477 // Conditional branch.
1478 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
1479 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1484 // Two-way Conditional branch.
1485 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
1486 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1487 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
1491 void X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1492 MachineBasicBlock::iterator MI,
1493 unsigned DestReg, unsigned SrcReg,
1494 const TargetRegisterClass *DestRC,
1495 const TargetRegisterClass *SrcRC) const {
1496 if (DestRC == SrcRC) {
1498 if (DestRC == &X86::GR64RegClass) {
1500 } else if (DestRC == &X86::GR32RegClass) {
1502 } else if (DestRC == &X86::GR16RegClass) {
1504 } else if (DestRC == &X86::GR8RegClass) {
1506 } else if (DestRC == &X86::GR32_RegClass) {
1507 Opc = X86::MOV32_rr;
1508 } else if (DestRC == &X86::GR16_RegClass) {
1509 Opc = X86::MOV16_rr;
1510 } else if (DestRC == &X86::RFP32RegClass) {
1511 Opc = X86::MOV_Fp3232;
1512 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1513 Opc = X86::MOV_Fp6464;
1514 } else if (DestRC == &X86::RFP80RegClass) {
1515 Opc = X86::MOV_Fp8080;
1516 } else if (DestRC == &X86::FR32RegClass) {
1517 Opc = X86::FsMOVAPSrr;
1518 } else if (DestRC == &X86::FR64RegClass) {
1519 Opc = X86::FsMOVAPDrr;
1520 } else if (DestRC == &X86::VR128RegClass) {
1521 Opc = X86::MOVAPSrr;
1522 } else if (DestRC == &X86::VR64RegClass) {
1523 Opc = X86::MMX_MOVQ64rr;
1525 assert(0 && "Unknown regclass");
1528 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
1532 // Moving EFLAGS to / from another register requires a push and a pop.
1533 if (SrcRC == &X86::CCRRegClass) {
1534 assert(SrcReg == X86::EFLAGS);
1535 if (DestRC == &X86::GR64RegClass) {
1536 BuildMI(MBB, MI, get(X86::PUSHFQ));
1537 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
1539 } else if (DestRC == &X86::GR32RegClass) {
1540 BuildMI(MBB, MI, get(X86::PUSHFD));
1541 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
1544 } else if (DestRC == &X86::CCRRegClass) {
1545 assert(DestReg == X86::EFLAGS);
1546 if (SrcRC == &X86::GR64RegClass) {
1547 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
1548 BuildMI(MBB, MI, get(X86::POPFQ));
1550 } else if (SrcRC == &X86::GR32RegClass) {
1551 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
1552 BuildMI(MBB, MI, get(X86::POPFD));
1557 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1558 if (SrcRC == &X86::RSTRegClass) {
1559 // Copying from ST(0)/ST(1).
1560 assert((SrcReg == X86::ST0 || SrcReg == X86::ST1) &&
1561 "Can only copy from ST(0)/ST(1) right now");
1562 bool isST0 = SrcReg == X86::ST0;
1564 if (DestRC == &X86::RFP32RegClass)
1565 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1566 else if (DestRC == &X86::RFP64RegClass)
1567 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1569 assert(DestRC == &X86::RFP80RegClass);
1570 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1572 BuildMI(MBB, MI, get(Opc), DestReg);
1576 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1577 if (DestRC == &X86::RSTRegClass) {
1578 // Copying to ST(0). FIXME: handle ST(1) also
1579 assert(DestReg == X86::ST0 && "Can only copy to TOS right now");
1581 if (SrcRC == &X86::RFP32RegClass)
1582 Opc = X86::FpSET_ST0_32;
1583 else if (SrcRC == &X86::RFP64RegClass)
1584 Opc = X86::FpSET_ST0_64;
1586 assert(SrcRC == &X86::RFP80RegClass);
1587 Opc = X86::FpSET_ST0_80;
1589 BuildMI(MBB, MI, get(Opc)).addReg(SrcReg);
1593 assert(0 && "Not yet supported!");
1597 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1598 unsigned StackAlign) {
1600 if (RC == &X86::GR64RegClass) {
1602 } else if (RC == &X86::GR32RegClass) {
1604 } else if (RC == &X86::GR16RegClass) {
1606 } else if (RC == &X86::GR8RegClass) {
1608 } else if (RC == &X86::GR32_RegClass) {
1609 Opc = X86::MOV32_mr;
1610 } else if (RC == &X86::GR16_RegClass) {
1611 Opc = X86::MOV16_mr;
1612 } else if (RC == &X86::RFP80RegClass) {
1613 Opc = X86::ST_FpP80m; // pops
1614 } else if (RC == &X86::RFP64RegClass) {
1615 Opc = X86::ST_Fp64m;
1616 } else if (RC == &X86::RFP32RegClass) {
1617 Opc = X86::ST_Fp32m;
1618 } else if (RC == &X86::FR32RegClass) {
1620 } else if (RC == &X86::FR64RegClass) {
1622 } else if (RC == &X86::VR128RegClass) {
1623 // FIXME: Use movaps once we are capable of selectively
1624 // aligning functions that spill SSE registers on 16-byte boundaries.
1625 Opc = StackAlign >= 16 ? X86::MOVAPSmr : X86::MOVUPSmr;
1626 } else if (RC == &X86::VR64RegClass) {
1627 Opc = X86::MMX_MOVQ64mr;
1629 assert(0 && "Unknown regclass");
1636 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1637 MachineBasicBlock::iterator MI,
1638 unsigned SrcReg, bool isKill, int FrameIdx,
1639 const TargetRegisterClass *RC) const {
1640 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
1641 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
1642 .addReg(SrcReg, false, false, isKill);
1645 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1647 SmallVectorImpl<MachineOperand> &Addr,
1648 const TargetRegisterClass *RC,
1649 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1650 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
1651 MachineInstrBuilder MIB = BuildMI(get(Opc));
1652 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1653 MIB = X86InstrAddOperand(MIB, Addr[i]);
1654 MIB.addReg(SrcReg, false, false, isKill);
1655 NewMIs.push_back(MIB);
1658 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1659 unsigned StackAlign) {
1661 if (RC == &X86::GR64RegClass) {
1663 } else if (RC == &X86::GR32RegClass) {
1665 } else if (RC == &X86::GR16RegClass) {
1667 } else if (RC == &X86::GR8RegClass) {
1669 } else if (RC == &X86::GR32_RegClass) {
1670 Opc = X86::MOV32_rm;
1671 } else if (RC == &X86::GR16_RegClass) {
1672 Opc = X86::MOV16_rm;
1673 } else if (RC == &X86::RFP80RegClass) {
1674 Opc = X86::LD_Fp80m;
1675 } else if (RC == &X86::RFP64RegClass) {
1676 Opc = X86::LD_Fp64m;
1677 } else if (RC == &X86::RFP32RegClass) {
1678 Opc = X86::LD_Fp32m;
1679 } else if (RC == &X86::FR32RegClass) {
1681 } else if (RC == &X86::FR64RegClass) {
1683 } else if (RC == &X86::VR128RegClass) {
1684 // FIXME: Use movaps once we are capable of selectively
1685 // aligning functions that spill SSE registers on 16-byte boundaries.
1686 Opc = StackAlign >= 16 ? X86::MOVAPSrm : X86::MOVUPSrm;
1687 } else if (RC == &X86::VR64RegClass) {
1688 Opc = X86::MMX_MOVQ64rm;
1690 assert(0 && "Unknown regclass");
1697 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1698 MachineBasicBlock::iterator MI,
1699 unsigned DestReg, int FrameIdx,
1700 const TargetRegisterClass *RC) const{
1701 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
1702 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
1705 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1706 SmallVectorImpl<MachineOperand> &Addr,
1707 const TargetRegisterClass *RC,
1708 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1709 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
1710 MachineInstrBuilder MIB = BuildMI(get(Opc), DestReg);
1711 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1712 MIB = X86InstrAddOperand(MIB, Addr[i]);
1713 NewMIs.push_back(MIB);
1716 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1717 MachineBasicBlock::iterator MI,
1718 const std::vector<CalleeSavedInfo> &CSI) const {
1722 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1723 unsigned SlotSize = is64Bit ? 8 : 4;
1725 MachineFunction &MF = *MBB.getParent();
1726 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1727 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1729 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1730 for (unsigned i = CSI.size(); i != 0; --i) {
1731 unsigned Reg = CSI[i-1].getReg();
1732 // Add the callee-saved register as live-in. It's killed at the spill.
1734 BuildMI(MBB, MI, get(Opc)).addReg(Reg);
1739 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1740 MachineBasicBlock::iterator MI,
1741 const std::vector<CalleeSavedInfo> &CSI) const {
1745 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1747 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1748 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1749 unsigned Reg = CSI[i].getReg();
1750 BuildMI(MBB, MI, get(Opc), Reg);
1755 static MachineInstr *FuseTwoAddrInst(unsigned Opcode,
1756 SmallVector<MachineOperand,4> &MOs,
1757 MachineInstr *MI, const TargetInstrInfo &TII) {
1758 // Create the base instruction with the memory operand as the first part.
1759 MachineInstr *NewMI = new MachineInstr(TII.get(Opcode), true);
1760 MachineInstrBuilder MIB(NewMI);
1761 unsigned NumAddrOps = MOs.size();
1762 for (unsigned i = 0; i != NumAddrOps; ++i)
1763 MIB = X86InstrAddOperand(MIB, MOs[i]);
1764 if (NumAddrOps < 4) // FrameIndex only
1765 MIB.addImm(1).addReg(0).addImm(0);
1767 // Loop over the rest of the ri operands, converting them over.
1768 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1769 for (unsigned i = 0; i != NumOps; ++i) {
1770 MachineOperand &MO = MI->getOperand(i+2);
1771 MIB = X86InstrAddOperand(MIB, MO);
1773 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1774 MachineOperand &MO = MI->getOperand(i);
1775 MIB = X86InstrAddOperand(MIB, MO);
1780 static MachineInstr *FuseInst(unsigned Opcode, unsigned OpNo,
1781 SmallVector<MachineOperand,4> &MOs,
1782 MachineInstr *MI, const TargetInstrInfo &TII) {
1783 MachineInstr *NewMI = new MachineInstr(TII.get(Opcode), true);
1784 MachineInstrBuilder MIB(NewMI);
1786 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1787 MachineOperand &MO = MI->getOperand(i);
1789 assert(MO.isRegister() && "Expected to fold into reg operand!");
1790 unsigned NumAddrOps = MOs.size();
1791 for (unsigned i = 0; i != NumAddrOps; ++i)
1792 MIB = X86InstrAddOperand(MIB, MOs[i]);
1793 if (NumAddrOps < 4) // FrameIndex only
1794 MIB.addImm(1).addReg(0).addImm(0);
1796 MIB = X86InstrAddOperand(MIB, MO);
1802 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
1803 SmallVector<MachineOperand,4> &MOs,
1805 MachineInstrBuilder MIB = BuildMI(TII.get(Opcode));
1807 unsigned NumAddrOps = MOs.size();
1808 for (unsigned i = 0; i != NumAddrOps; ++i)
1809 MIB = X86InstrAddOperand(MIB, MOs[i]);
1810 if (NumAddrOps < 4) // FrameIndex only
1811 MIB.addImm(1).addReg(0).addImm(0);
1812 return MIB.addImm(0);
1816 X86InstrInfo::foldMemoryOperand(MachineInstr *MI, unsigned i,
1817 SmallVector<MachineOperand,4> &MOs) const {
1818 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
1819 bool isTwoAddrFold = false;
1820 unsigned NumOps = MI->getDesc().getNumOperands();
1821 bool isTwoAddr = NumOps > 1 &&
1822 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
1824 MachineInstr *NewMI = NULL;
1825 // Folding a memory location into the two-address part of a two-address
1826 // instruction is different than folding it other places. It requires
1827 // replacing the *two* registers with the memory location.
1828 if (isTwoAddr && NumOps >= 2 && i < 2 &&
1829 MI->getOperand(0).isRegister() &&
1830 MI->getOperand(1).isRegister() &&
1831 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
1832 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
1833 isTwoAddrFold = true;
1834 } else if (i == 0) { // If operand 0
1835 if (MI->getOpcode() == X86::MOV16r0)
1836 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
1837 else if (MI->getOpcode() == X86::MOV32r0)
1838 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
1839 else if (MI->getOpcode() == X86::MOV64r0)
1840 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
1841 else if (MI->getOpcode() == X86::MOV8r0)
1842 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
1844 NewMI->copyKillDeadInfo(MI);
1848 OpcodeTablePtr = &RegOp2MemOpTable0;
1849 } else if (i == 1) {
1850 OpcodeTablePtr = &RegOp2MemOpTable1;
1851 } else if (i == 2) {
1852 OpcodeTablePtr = &RegOp2MemOpTable2;
1855 // If table selected...
1856 if (OpcodeTablePtr) {
1857 // Find the Opcode to fuse
1858 DenseMap<unsigned*, unsigned>::iterator I =
1859 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
1860 if (I != OpcodeTablePtr->end()) {
1862 NewMI = FuseTwoAddrInst(I->second, MOs, MI, *this);
1864 NewMI = FuseInst(I->second, i, MOs, MI, *this);
1865 NewMI->copyKillDeadInfo(MI);
1871 if (PrintFailedFusing)
1872 cerr << "We failed to fuse operand " << i << *MI;
1877 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1879 SmallVectorImpl<unsigned> &Ops,
1880 int FrameIndex) const {
1881 // Check switch flag
1882 if (NoFusing) return NULL;
1884 const MachineFrameInfo *MFI = MF.getFrameInfo();
1885 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
1886 // FIXME: Move alignment requirement into tables?
1887 if (Alignment < 16) {
1888 switch (MI->getOpcode()) {
1890 // Not always safe to fold movsd into these instructions since their load
1891 // folding variants expects the address to be 16 byte aligned.
1892 case X86::FsANDNPDrr:
1893 case X86::FsANDNPSrr:
1894 case X86::FsANDPDrr:
1895 case X86::FsANDPSrr:
1898 case X86::FsXORPDrr:
1899 case X86::FsXORPSrr:
1904 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1905 unsigned NewOpc = 0;
1906 switch (MI->getOpcode()) {
1907 default: return NULL;
1908 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
1909 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
1910 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
1911 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
1913 // Change to CMPXXri r, 0 first.
1914 MI->setDesc(get(NewOpc));
1915 MI->getOperand(1).ChangeToImmediate(0);
1916 } else if (Ops.size() != 1)
1919 SmallVector<MachineOperand,4> MOs;
1920 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
1921 return foldMemoryOperand(MI, Ops[0], MOs);
1924 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1926 SmallVectorImpl<unsigned> &Ops,
1927 MachineInstr *LoadMI) const {
1928 // Check switch flag
1929 if (NoFusing) return NULL;
1931 unsigned Alignment = 0;
1932 for (unsigned i = 0, e = LoadMI->getNumMemOperands(); i != e; ++i) {
1933 const MachineMemOperand &MRO = LoadMI->getMemOperand(i);
1934 unsigned Align = MRO.getAlignment();
1935 if (Align > Alignment)
1939 // FIXME: Move alignment requirement into tables?
1940 if (Alignment < 16) {
1941 switch (MI->getOpcode()) {
1943 // Not always safe to fold movsd into these instructions since their load
1944 // folding variants expects the address to be 16 byte aligned.
1945 case X86::FsANDNPDrr:
1946 case X86::FsANDNPSrr:
1947 case X86::FsANDPDrr:
1948 case X86::FsANDPSrr:
1951 case X86::FsXORPDrr:
1952 case X86::FsXORPSrr:
1957 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1958 unsigned NewOpc = 0;
1959 switch (MI->getOpcode()) {
1960 default: return NULL;
1961 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
1962 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
1963 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
1964 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
1966 // Change to CMPXXri r, 0 first.
1967 MI->setDesc(get(NewOpc));
1968 MI->getOperand(1).ChangeToImmediate(0);
1969 } else if (Ops.size() != 1)
1972 SmallVector<MachineOperand,4> MOs;
1973 unsigned NumOps = LoadMI->getDesc().getNumOperands();
1974 for (unsigned i = NumOps - 4; i != NumOps; ++i)
1975 MOs.push_back(LoadMI->getOperand(i));
1976 return foldMemoryOperand(MI, Ops[0], MOs);
1980 bool X86InstrInfo::canFoldMemoryOperand(MachineInstr *MI,
1981 SmallVectorImpl<unsigned> &Ops) const {
1982 // Check switch flag
1983 if (NoFusing) return 0;
1985 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
1986 switch (MI->getOpcode()) {
1987 default: return false;
1996 if (Ops.size() != 1)
1999 unsigned OpNum = Ops[0];
2000 unsigned Opc = MI->getOpcode();
2001 unsigned NumOps = MI->getDesc().getNumOperands();
2002 bool isTwoAddr = NumOps > 1 &&
2003 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2005 // Folding a memory location into the two-address part of a two-address
2006 // instruction is different than folding it other places. It requires
2007 // replacing the *two* registers with the memory location.
2008 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2009 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2010 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2011 } else if (OpNum == 0) { // If operand 0
2020 OpcodeTablePtr = &RegOp2MemOpTable0;
2021 } else if (OpNum == 1) {
2022 OpcodeTablePtr = &RegOp2MemOpTable1;
2023 } else if (OpNum == 2) {
2024 OpcodeTablePtr = &RegOp2MemOpTable2;
2027 if (OpcodeTablePtr) {
2028 // Find the Opcode to fuse
2029 DenseMap<unsigned*, unsigned>::iterator I =
2030 OpcodeTablePtr->find((unsigned*)Opc);
2031 if (I != OpcodeTablePtr->end())
2037 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2038 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2039 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2040 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2041 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2042 if (I == MemOp2RegOpTable.end())
2044 unsigned Opc = I->second.first;
2045 unsigned Index = I->second.second & 0xf;
2046 bool FoldedLoad = I->second.second & (1 << 4);
2047 bool FoldedStore = I->second.second & (1 << 5);
2048 if (UnfoldLoad && !FoldedLoad)
2050 UnfoldLoad &= FoldedLoad;
2051 if (UnfoldStore && !FoldedStore)
2053 UnfoldStore &= FoldedStore;
2055 const TargetInstrDesc &TID = get(Opc);
2056 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2057 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2058 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2059 SmallVector<MachineOperand,4> AddrOps;
2060 SmallVector<MachineOperand,2> BeforeOps;
2061 SmallVector<MachineOperand,2> AfterOps;
2062 SmallVector<MachineOperand,4> ImpOps;
2063 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2064 MachineOperand &Op = MI->getOperand(i);
2065 if (i >= Index && i < Index+4)
2066 AddrOps.push_back(Op);
2067 else if (Op.isRegister() && Op.isImplicit())
2068 ImpOps.push_back(Op);
2070 BeforeOps.push_back(Op);
2072 AfterOps.push_back(Op);
2075 // Emit the load instruction.
2077 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2079 // Address operands cannot be marked isKill.
2080 for (unsigned i = 1; i != 5; ++i) {
2081 MachineOperand &MO = NewMIs[0]->getOperand(i);
2082 if (MO.isRegister())
2083 MO.setIsKill(false);
2088 // Emit the data processing instruction.
2089 MachineInstr *DataMI = new MachineInstr(TID, true);
2090 MachineInstrBuilder MIB(DataMI);
2093 MIB.addReg(Reg, true);
2094 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2095 MIB = X86InstrAddOperand(MIB, BeforeOps[i]);
2098 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2099 MIB = X86InstrAddOperand(MIB, AfterOps[i]);
2100 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2101 MachineOperand &MO = ImpOps[i];
2102 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2104 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2105 unsigned NewOpc = 0;
2106 switch (DataMI->getOpcode()) {
2108 case X86::CMP64ri32:
2112 MachineOperand &MO0 = DataMI->getOperand(0);
2113 MachineOperand &MO1 = DataMI->getOperand(1);
2114 if (MO1.getImm() == 0) {
2115 switch (DataMI->getOpcode()) {
2117 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2118 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2119 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2120 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2122 DataMI->setDesc(get(NewOpc));
2123 MO1.ChangeToRegister(MO0.getReg(), false);
2127 NewMIs.push_back(DataMI);
2129 // Emit the store instruction.
2131 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2132 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2133 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2134 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2141 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2142 SmallVectorImpl<SDNode*> &NewNodes) const {
2143 if (!N->isTargetOpcode())
2146 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2147 MemOp2RegOpTable.find((unsigned*)N->getTargetOpcode());
2148 if (I == MemOp2RegOpTable.end())
2150 unsigned Opc = I->second.first;
2151 unsigned Index = I->second.second & 0xf;
2152 bool FoldedLoad = I->second.second & (1 << 4);
2153 bool FoldedStore = I->second.second & (1 << 5);
2154 const TargetInstrDesc &TID = get(Opc);
2155 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2156 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2157 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2158 std::vector<SDOperand> AddrOps;
2159 std::vector<SDOperand> BeforeOps;
2160 std::vector<SDOperand> AfterOps;
2161 unsigned NumOps = N->getNumOperands();
2162 for (unsigned i = 0; i != NumOps-1; ++i) {
2163 SDOperand Op = N->getOperand(i);
2164 if (i >= Index && i < Index+4)
2165 AddrOps.push_back(Op);
2167 BeforeOps.push_back(Op);
2169 AfterOps.push_back(Op);
2171 SDOperand Chain = N->getOperand(NumOps-1);
2172 AddrOps.push_back(Chain);
2174 // Emit the load instruction.
2177 MVT::ValueType VT = *RC->vt_begin();
2178 Load = DAG.getTargetNode(getLoadRegOpcode(RC, RI.getStackAlignment()), VT,
2179 MVT::Other, &AddrOps[0], AddrOps.size());
2180 NewNodes.push_back(Load);
2183 // Emit the data processing instruction.
2184 std::vector<MVT::ValueType> VTs;
2185 const TargetRegisterClass *DstRC = 0;
2186 if (TID.getNumDefs() > 0) {
2187 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2188 DstRC = DstTOI.isLookupPtrRegClass()
2189 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2190 VTs.push_back(*DstRC->vt_begin());
2192 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2193 MVT::ValueType VT = N->getValueType(i);
2194 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2198 BeforeOps.push_back(SDOperand(Load, 0));
2199 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2200 SDNode *NewNode= DAG.getTargetNode(Opc, VTs, &BeforeOps[0], BeforeOps.size());
2201 NewNodes.push_back(NewNode);
2203 // Emit the store instruction.
2206 AddrOps.push_back(SDOperand(NewNode, 0));
2207 AddrOps.push_back(Chain);
2208 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, RI.getStackAlignment()),
2209 MVT::Other, &AddrOps[0], AddrOps.size());
2210 NewNodes.push_back(Store);
2216 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2217 bool UnfoldLoad, bool UnfoldStore) const {
2218 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2219 MemOp2RegOpTable.find((unsigned*)Opc);
2220 if (I == MemOp2RegOpTable.end())
2222 bool FoldedLoad = I->second.second & (1 << 4);
2223 bool FoldedStore = I->second.second & (1 << 5);
2224 if (UnfoldLoad && !FoldedLoad)
2226 if (UnfoldStore && !FoldedStore)
2228 return I->second.first;
2231 bool X86InstrInfo::BlockHasNoFallThrough(MachineBasicBlock &MBB) const {
2232 if (MBB.empty()) return false;
2234 switch (MBB.back().getOpcode()) {
2235 case X86::TCRETURNri:
2236 case X86::TCRETURNdi:
2237 case X86::RET: // Return.
2242 case X86::JMP: // Uncond branch.
2243 case X86::JMP32r: // Indirect branch.
2244 case X86::JMP64r: // Indirect branch (64-bit).
2245 case X86::JMP32m: // Indirect branch through mem.
2246 case X86::JMP64m: // Indirect branch through mem (64-bit).
2248 default: return false;
2253 ReverseBranchCondition(std::vector<MachineOperand> &Cond) const {
2254 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2255 Cond[0].setImm(GetOppositeBranchCondition((X86::CondCode)Cond[0].getImm()));
2259 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
2260 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
2261 if (Subtarget->is64Bit())
2262 return &X86::GR64RegClass;
2264 return &X86::GR32RegClass;