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,
214 assert(false && "Duplicated entries?");
215 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
216 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
217 std::make_pair(RegOp,
219 AmbEntries.push_back(MemOp);
222 // If the third value is 1, then it's folding either a load or a store.
223 static const unsigned OpTbl0[][3] = {
224 { X86::CALL32r, X86::CALL32m, 1 },
225 { X86::CALL64r, X86::CALL64m, 1 },
226 { X86::CMP16ri, X86::CMP16mi, 1 },
227 { X86::CMP16ri8, X86::CMP16mi8, 1 },
228 { X86::CMP16rr, X86::CMP16mr, 1 },
229 { X86::CMP32ri, X86::CMP32mi, 1 },
230 { X86::CMP32ri8, X86::CMP32mi8, 1 },
231 { X86::CMP32rr, X86::CMP32mr, 1 },
232 { X86::CMP64ri32, X86::CMP64mi32, 1 },
233 { X86::CMP64ri8, X86::CMP64mi8, 1 },
234 { X86::CMP64rr, X86::CMP64mr, 1 },
235 { X86::CMP8ri, X86::CMP8mi, 1 },
236 { X86::CMP8rr, X86::CMP8mr, 1 },
237 { X86::DIV16r, X86::DIV16m, 1 },
238 { X86::DIV32r, X86::DIV32m, 1 },
239 { X86::DIV64r, X86::DIV64m, 1 },
240 { X86::DIV8r, X86::DIV8m, 1 },
241 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0 },
242 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
243 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
244 { X86::IDIV16r, X86::IDIV16m, 1 },
245 { X86::IDIV32r, X86::IDIV32m, 1 },
246 { X86::IDIV64r, X86::IDIV64m, 1 },
247 { X86::IDIV8r, X86::IDIV8m, 1 },
248 { X86::IMUL16r, X86::IMUL16m, 1 },
249 { X86::IMUL32r, X86::IMUL32m, 1 },
250 { X86::IMUL64r, X86::IMUL64m, 1 },
251 { X86::IMUL8r, X86::IMUL8m, 1 },
252 { X86::JMP32r, X86::JMP32m, 1 },
253 { X86::JMP64r, X86::JMP64m, 1 },
254 { X86::MOV16ri, X86::MOV16mi, 0 },
255 { X86::MOV16rr, X86::MOV16mr, 0 },
256 { X86::MOV16to16_, X86::MOV16_mr, 0 },
257 { X86::MOV32ri, X86::MOV32mi, 0 },
258 { X86::MOV32rr, X86::MOV32mr, 0 },
259 { X86::MOV32to32_, X86::MOV32_mr, 0 },
260 { X86::MOV64ri32, X86::MOV64mi32, 0 },
261 { X86::MOV64rr, X86::MOV64mr, 0 },
262 { X86::MOV8ri, X86::MOV8mi, 0 },
263 { X86::MOV8rr, X86::MOV8mr, 0 },
264 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
265 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
266 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
267 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
268 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
269 { X86::MOVSDrr, X86::MOVSDmr, 0 },
270 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
271 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
272 { X86::MOVSSrr, X86::MOVSSmr, 0 },
273 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
274 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
275 { X86::MUL16r, X86::MUL16m, 1 },
276 { X86::MUL32r, X86::MUL32m, 1 },
277 { X86::MUL64r, X86::MUL64m, 1 },
278 { X86::MUL8r, X86::MUL8m, 1 },
279 { X86::SETAEr, X86::SETAEm, 0 },
280 { X86::SETAr, X86::SETAm, 0 },
281 { X86::SETBEr, X86::SETBEm, 0 },
282 { X86::SETBr, X86::SETBm, 0 },
283 { X86::SETEr, X86::SETEm, 0 },
284 { X86::SETGEr, X86::SETGEm, 0 },
285 { X86::SETGr, X86::SETGm, 0 },
286 { X86::SETLEr, X86::SETLEm, 0 },
287 { X86::SETLr, X86::SETLm, 0 },
288 { X86::SETNEr, X86::SETNEm, 0 },
289 { X86::SETNPr, X86::SETNPm, 0 },
290 { X86::SETNSr, X86::SETNSm, 0 },
291 { X86::SETPr, X86::SETPm, 0 },
292 { X86::SETSr, X86::SETSm, 0 },
293 { X86::TAILJMPr, X86::TAILJMPm, 1 },
294 { X86::TEST16ri, X86::TEST16mi, 1 },
295 { X86::TEST32ri, X86::TEST32mi, 1 },
296 { X86::TEST64ri32, X86::TEST64mi32, 1 },
297 { X86::TEST8ri, X86::TEST8mi, 1 }
300 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
301 unsigned RegOp = OpTbl0[i][0];
302 unsigned MemOp = OpTbl0[i][1];
303 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp,
305 assert(false && "Duplicated entries?");
306 unsigned FoldedLoad = OpTbl0[i][2];
307 // Index 0, folded load or store.
308 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
309 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
310 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
311 std::make_pair(RegOp, AuxInfo))).second)
312 AmbEntries.push_back(MemOp);
315 static const unsigned OpTbl1[][2] = {
316 { X86::CMP16rr, X86::CMP16rm },
317 { X86::CMP32rr, X86::CMP32rm },
318 { X86::CMP64rr, X86::CMP64rm },
319 { X86::CMP8rr, X86::CMP8rm },
320 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
321 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
322 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
323 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
324 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
325 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
326 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
327 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
328 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
329 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
330 { X86::FsMOVAPDrr, X86::MOVSDrm },
331 { X86::FsMOVAPSrr, X86::MOVSSrm },
332 { X86::IMUL16rri, X86::IMUL16rmi },
333 { X86::IMUL16rri8, X86::IMUL16rmi8 },
334 { X86::IMUL32rri, X86::IMUL32rmi },
335 { X86::IMUL32rri8, X86::IMUL32rmi8 },
336 { X86::IMUL64rri32, X86::IMUL64rmi32 },
337 { X86::IMUL64rri8, X86::IMUL64rmi8 },
338 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
339 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
340 { X86::Int_COMISDrr, X86::Int_COMISDrm },
341 { X86::Int_COMISSrr, X86::Int_COMISSrm },
342 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
343 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
344 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
345 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
346 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
347 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
348 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
349 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
350 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
351 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
352 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
353 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
354 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
355 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
356 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
357 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
358 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
359 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
360 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
361 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
362 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
363 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
364 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
365 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
366 { X86::MOV16rr, X86::MOV16rm },
367 { X86::MOV16to16_, X86::MOV16_rm },
368 { X86::MOV32rr, X86::MOV32rm },
369 { X86::MOV32to32_, X86::MOV32_rm },
370 { X86::MOV64rr, X86::MOV64rm },
371 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
372 { X86::MOV64toSDrr, X86::MOV64toSDrm },
373 { X86::MOV8rr, X86::MOV8rm },
374 { X86::MOVAPDrr, X86::MOVAPDrm },
375 { X86::MOVAPSrr, X86::MOVAPSrm },
376 { X86::MOVDDUPrr, X86::MOVDDUPrm },
377 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
378 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
379 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
380 { X86::MOVSDrr, X86::MOVSDrm },
381 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
382 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
383 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
384 { X86::MOVSSrr, X86::MOVSSrm },
385 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
386 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
387 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
388 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
389 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
390 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
391 { X86::MOVUPDrr, X86::MOVUPDrm },
392 { X86::MOVUPSrr, X86::MOVUPSrm },
393 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
394 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
395 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
396 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
397 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
398 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
399 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
400 { X86::MOVZX64rr32, X86::MOVZX64rm32 },
401 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
402 { X86::PSHUFDri, X86::PSHUFDmi },
403 { X86::PSHUFHWri, X86::PSHUFHWmi },
404 { X86::PSHUFLWri, X86::PSHUFLWmi },
405 { X86::RCPPSr, X86::RCPPSm },
406 { X86::RCPPSr_Int, X86::RCPPSm_Int },
407 { X86::RSQRTPSr, X86::RSQRTPSm },
408 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
409 { X86::RSQRTSSr, X86::RSQRTSSm },
410 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
411 { X86::SQRTPDr, X86::SQRTPDm },
412 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
413 { X86::SQRTPSr, X86::SQRTPSm },
414 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
415 { X86::SQRTSDr, X86::SQRTSDm },
416 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
417 { X86::SQRTSSr, X86::SQRTSSm },
418 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
419 { X86::TEST16rr, X86::TEST16rm },
420 { X86::TEST32rr, X86::TEST32rm },
421 { X86::TEST64rr, X86::TEST64rm },
422 { X86::TEST8rr, X86::TEST8rm },
423 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
424 { X86::UCOMISDrr, X86::UCOMISDrm },
425 { X86::UCOMISSrr, X86::UCOMISSrm }
428 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
429 unsigned RegOp = OpTbl1[i][0];
430 unsigned MemOp = OpTbl1[i][1];
431 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp,
433 assert(false && "Duplicated entries?");
434 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
435 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
436 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
437 std::make_pair(RegOp, AuxInfo))).second)
438 AmbEntries.push_back(MemOp);
441 static const unsigned OpTbl2[][2] = {
442 { X86::ADC32rr, X86::ADC32rm },
443 { X86::ADC64rr, X86::ADC64rm },
444 { X86::ADD16rr, X86::ADD16rm },
445 { X86::ADD32rr, X86::ADD32rm },
446 { X86::ADD64rr, X86::ADD64rm },
447 { X86::ADD8rr, X86::ADD8rm },
448 { X86::ADDPDrr, X86::ADDPDrm },
449 { X86::ADDPSrr, X86::ADDPSrm },
450 { X86::ADDSDrr, X86::ADDSDrm },
451 { X86::ADDSSrr, X86::ADDSSrm },
452 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
453 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
454 { X86::AND16rr, X86::AND16rm },
455 { X86::AND32rr, X86::AND32rm },
456 { X86::AND64rr, X86::AND64rm },
457 { X86::AND8rr, X86::AND8rm },
458 { X86::ANDNPDrr, X86::ANDNPDrm },
459 { X86::ANDNPSrr, X86::ANDNPSrm },
460 { X86::ANDPDrr, X86::ANDPDrm },
461 { X86::ANDPSrr, X86::ANDPSrm },
462 { X86::CMOVA16rr, X86::CMOVA16rm },
463 { X86::CMOVA32rr, X86::CMOVA32rm },
464 { X86::CMOVA64rr, X86::CMOVA64rm },
465 { X86::CMOVAE16rr, X86::CMOVAE16rm },
466 { X86::CMOVAE32rr, X86::CMOVAE32rm },
467 { X86::CMOVAE64rr, X86::CMOVAE64rm },
468 { X86::CMOVB16rr, X86::CMOVB16rm },
469 { X86::CMOVB32rr, X86::CMOVB32rm },
470 { X86::CMOVB64rr, X86::CMOVB64rm },
471 { X86::CMOVBE16rr, X86::CMOVBE16rm },
472 { X86::CMOVBE32rr, X86::CMOVBE32rm },
473 { X86::CMOVBE64rr, X86::CMOVBE64rm },
474 { X86::CMOVE16rr, X86::CMOVE16rm },
475 { X86::CMOVE32rr, X86::CMOVE32rm },
476 { X86::CMOVE64rr, X86::CMOVE64rm },
477 { X86::CMOVG16rr, X86::CMOVG16rm },
478 { X86::CMOVG32rr, X86::CMOVG32rm },
479 { X86::CMOVG64rr, X86::CMOVG64rm },
480 { X86::CMOVGE16rr, X86::CMOVGE16rm },
481 { X86::CMOVGE32rr, X86::CMOVGE32rm },
482 { X86::CMOVGE64rr, X86::CMOVGE64rm },
483 { X86::CMOVL16rr, X86::CMOVL16rm },
484 { X86::CMOVL32rr, X86::CMOVL32rm },
485 { X86::CMOVL64rr, X86::CMOVL64rm },
486 { X86::CMOVLE16rr, X86::CMOVLE16rm },
487 { X86::CMOVLE32rr, X86::CMOVLE32rm },
488 { X86::CMOVLE64rr, X86::CMOVLE64rm },
489 { X86::CMOVNE16rr, X86::CMOVNE16rm },
490 { X86::CMOVNE32rr, X86::CMOVNE32rm },
491 { X86::CMOVNE64rr, X86::CMOVNE64rm },
492 { X86::CMOVNP16rr, X86::CMOVNP16rm },
493 { X86::CMOVNP32rr, X86::CMOVNP32rm },
494 { X86::CMOVNP64rr, X86::CMOVNP64rm },
495 { X86::CMOVNS16rr, X86::CMOVNS16rm },
496 { X86::CMOVNS32rr, X86::CMOVNS32rm },
497 { X86::CMOVNS64rr, X86::CMOVNS64rm },
498 { X86::CMOVP16rr, X86::CMOVP16rm },
499 { X86::CMOVP32rr, X86::CMOVP32rm },
500 { X86::CMOVP64rr, X86::CMOVP64rm },
501 { X86::CMOVS16rr, X86::CMOVS16rm },
502 { X86::CMOVS32rr, X86::CMOVS32rm },
503 { X86::CMOVS64rr, X86::CMOVS64rm },
504 { X86::CMPPDrri, X86::CMPPDrmi },
505 { X86::CMPPSrri, X86::CMPPSrmi },
506 { X86::CMPSDrr, X86::CMPSDrm },
507 { X86::CMPSSrr, X86::CMPSSrm },
508 { X86::DIVPDrr, X86::DIVPDrm },
509 { X86::DIVPSrr, X86::DIVPSrm },
510 { X86::DIVSDrr, X86::DIVSDrm },
511 { X86::DIVSSrr, X86::DIVSSrm },
512 { X86::FsANDNPDrr, X86::FsANDNPDrm },
513 { X86::FsANDNPSrr, X86::FsANDNPSrm },
514 { X86::FsANDPDrr, X86::FsANDPDrm },
515 { X86::FsANDPSrr, X86::FsANDPSrm },
516 { X86::FsORPDrr, X86::FsORPDrm },
517 { X86::FsORPSrr, X86::FsORPSrm },
518 { X86::FsXORPDrr, X86::FsXORPDrm },
519 { X86::FsXORPSrr, X86::FsXORPSrm },
520 { X86::HADDPDrr, X86::HADDPDrm },
521 { X86::HADDPSrr, X86::HADDPSrm },
522 { X86::HSUBPDrr, X86::HSUBPDrm },
523 { X86::HSUBPSrr, X86::HSUBPSrm },
524 { X86::IMUL16rr, X86::IMUL16rm },
525 { X86::IMUL32rr, X86::IMUL32rm },
526 { X86::IMUL64rr, X86::IMUL64rm },
527 { X86::MAXPDrr, X86::MAXPDrm },
528 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
529 { X86::MAXPSrr, X86::MAXPSrm },
530 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
531 { X86::MAXSDrr, X86::MAXSDrm },
532 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
533 { X86::MAXSSrr, X86::MAXSSrm },
534 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
535 { X86::MINPDrr, X86::MINPDrm },
536 { X86::MINPDrr_Int, X86::MINPDrm_Int },
537 { X86::MINPSrr, X86::MINPSrm },
538 { X86::MINPSrr_Int, X86::MINPSrm_Int },
539 { X86::MINSDrr, X86::MINSDrm },
540 { X86::MINSDrr_Int, X86::MINSDrm_Int },
541 { X86::MINSSrr, X86::MINSSrm },
542 { X86::MINSSrr_Int, X86::MINSSrm_Int },
543 { X86::MULPDrr, X86::MULPDrm },
544 { X86::MULPSrr, X86::MULPSrm },
545 { X86::MULSDrr, X86::MULSDrm },
546 { X86::MULSSrr, X86::MULSSrm },
547 { X86::OR16rr, X86::OR16rm },
548 { X86::OR32rr, X86::OR32rm },
549 { X86::OR64rr, X86::OR64rm },
550 { X86::OR8rr, X86::OR8rm },
551 { X86::ORPDrr, X86::ORPDrm },
552 { X86::ORPSrr, X86::ORPSrm },
553 { X86::PACKSSDWrr, X86::PACKSSDWrm },
554 { X86::PACKSSWBrr, X86::PACKSSWBrm },
555 { X86::PACKUSWBrr, X86::PACKUSWBrm },
556 { X86::PADDBrr, X86::PADDBrm },
557 { X86::PADDDrr, X86::PADDDrm },
558 { X86::PADDQrr, X86::PADDQrm },
559 { X86::PADDSBrr, X86::PADDSBrm },
560 { X86::PADDSWrr, X86::PADDSWrm },
561 { X86::PADDWrr, X86::PADDWrm },
562 { X86::PANDNrr, X86::PANDNrm },
563 { X86::PANDrr, X86::PANDrm },
564 { X86::PAVGBrr, X86::PAVGBrm },
565 { X86::PAVGWrr, X86::PAVGWrm },
566 { X86::PCMPEQBrr, X86::PCMPEQBrm },
567 { X86::PCMPEQDrr, X86::PCMPEQDrm },
568 { X86::PCMPEQWrr, X86::PCMPEQWrm },
569 { X86::PCMPGTBrr, X86::PCMPGTBrm },
570 { X86::PCMPGTDrr, X86::PCMPGTDrm },
571 { X86::PCMPGTWrr, X86::PCMPGTWrm },
572 { X86::PINSRWrri, X86::PINSRWrmi },
573 { X86::PMADDWDrr, X86::PMADDWDrm },
574 { X86::PMAXSWrr, X86::PMAXSWrm },
575 { X86::PMAXUBrr, X86::PMAXUBrm },
576 { X86::PMINSWrr, X86::PMINSWrm },
577 { X86::PMINUBrr, X86::PMINUBrm },
578 { X86::PMULDQrr, X86::PMULDQrm },
579 { X86::PMULDQrr_int, X86::PMULDQrm_int },
580 { X86::PMULHUWrr, X86::PMULHUWrm },
581 { X86::PMULHWrr, X86::PMULHWrm },
582 { X86::PMULLDrr, X86::PMULLDrm },
583 { X86::PMULLDrr_int, X86::PMULLDrm_int },
584 { X86::PMULLWrr, X86::PMULLWrm },
585 { X86::PMULUDQrr, X86::PMULUDQrm },
586 { X86::PORrr, X86::PORrm },
587 { X86::PSADBWrr, X86::PSADBWrm },
588 { X86::PSLLDrr, X86::PSLLDrm },
589 { X86::PSLLQrr, X86::PSLLQrm },
590 { X86::PSLLWrr, X86::PSLLWrm },
591 { X86::PSRADrr, X86::PSRADrm },
592 { X86::PSRAWrr, X86::PSRAWrm },
593 { X86::PSRLDrr, X86::PSRLDrm },
594 { X86::PSRLQrr, X86::PSRLQrm },
595 { X86::PSRLWrr, X86::PSRLWrm },
596 { X86::PSUBBrr, X86::PSUBBrm },
597 { X86::PSUBDrr, X86::PSUBDrm },
598 { X86::PSUBSBrr, X86::PSUBSBrm },
599 { X86::PSUBSWrr, X86::PSUBSWrm },
600 { X86::PSUBWrr, X86::PSUBWrm },
601 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
602 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
603 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
604 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
605 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
606 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
607 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
608 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
609 { X86::PXORrr, X86::PXORrm },
610 { X86::SBB32rr, X86::SBB32rm },
611 { X86::SBB64rr, X86::SBB64rm },
612 { X86::SHUFPDrri, X86::SHUFPDrmi },
613 { X86::SHUFPSrri, X86::SHUFPSrmi },
614 { X86::SUB16rr, X86::SUB16rm },
615 { X86::SUB32rr, X86::SUB32rm },
616 { X86::SUB64rr, X86::SUB64rm },
617 { X86::SUB8rr, X86::SUB8rm },
618 { X86::SUBPDrr, X86::SUBPDrm },
619 { X86::SUBPSrr, X86::SUBPSrm },
620 { X86::SUBSDrr, X86::SUBSDrm },
621 { X86::SUBSSrr, X86::SUBSSrm },
622 // FIXME: TEST*rr -> swapped operand of TEST*mr.
623 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
624 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
625 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
626 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
627 { X86::XOR16rr, X86::XOR16rm },
628 { X86::XOR32rr, X86::XOR32rm },
629 { X86::XOR64rr, X86::XOR64rm },
630 { X86::XOR8rr, X86::XOR8rm },
631 { X86::XORPDrr, X86::XORPDrm },
632 { X86::XORPSrr, X86::XORPSrm }
635 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
636 unsigned RegOp = OpTbl2[i][0];
637 unsigned MemOp = OpTbl2[i][1];
638 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp,
640 assert(false && "Duplicated entries?");
641 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
642 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
643 std::make_pair(RegOp, AuxInfo))).second)
644 AmbEntries.push_back(MemOp);
647 // Remove ambiguous entries.
648 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
651 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
653 unsigned& destReg) const {
654 switch (MI.getOpcode()) {
661 case X86::MOV16to16_:
662 case X86::MOV32to32_:
666 // FP Stack register class copies
667 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
668 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
669 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
671 case X86::FsMOVAPSrr:
672 case X86::FsMOVAPDrr:
675 case X86::MOVSS2PSrr:
676 case X86::MOVSD2PDrr:
677 case X86::MOVPS2SSrr:
678 case X86::MOVPD2SDrr:
679 case X86::MMX_MOVD64rr:
680 case X86::MMX_MOVQ64rr:
681 assert(MI.getNumOperands() >= 2 &&
682 MI.getOperand(0).isReg() &&
683 MI.getOperand(1).isReg() &&
684 "invalid register-register move instruction");
685 sourceReg = MI.getOperand(1).getReg();
686 destReg = MI.getOperand(0).getReg();
691 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
692 int &FrameIndex) const {
693 switch (MI->getOpcode()) {
706 case X86::MMX_MOVD64rm:
707 case X86::MMX_MOVQ64rm:
708 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
709 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
710 MI->getOperand(2).getImm() == 1 &&
711 MI->getOperand(3).getReg() == 0 &&
712 MI->getOperand(4).getImm() == 0) {
713 FrameIndex = MI->getOperand(1).getIndex();
714 return MI->getOperand(0).getReg();
721 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
722 int &FrameIndex) const {
723 switch (MI->getOpcode()) {
736 case X86::MMX_MOVD64mr:
737 case X86::MMX_MOVQ64mr:
738 case X86::MMX_MOVNTQmr:
739 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
740 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
741 MI->getOperand(1).getImm() == 1 &&
742 MI->getOperand(2).getReg() == 0 &&
743 MI->getOperand(3).getImm() == 0) {
744 FrameIndex = MI->getOperand(0).getIndex();
745 return MI->getOperand(4).getReg();
753 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
755 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
756 bool isPICBase = false;
757 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
758 E = MRI.def_end(); I != E; ++I) {
759 MachineInstr *DefMI = I.getOperand().getParent();
760 if (DefMI->getOpcode() != X86::MOVPC32r)
762 assert(!isPICBase && "More than one PIC base?");
768 /// isGVStub - Return true if the GV requires an extra load to get the
770 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
771 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
775 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI) const {
776 switch (MI->getOpcode()) {
789 case X86::MMX_MOVD64rm:
790 case X86::MMX_MOVQ64rm: {
791 // Loads from constant pools are trivially rematerializable.
792 if (MI->getOperand(1).isReg() &&
793 MI->getOperand(2).isImm() &&
794 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
795 (MI->getOperand(4).isCPI() ||
796 (MI->getOperand(4).isGlobal() &&
797 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
798 unsigned BaseReg = MI->getOperand(1).getReg();
801 // Allow re-materialization of PIC load.
802 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
804 const MachineFunction &MF = *MI->getParent()->getParent();
805 const MachineRegisterInfo &MRI = MF.getRegInfo();
806 bool isPICBase = false;
807 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
808 E = MRI.def_end(); I != E; ++I) {
809 MachineInstr *DefMI = I.getOperand().getParent();
810 if (DefMI->getOpcode() != X86::MOVPC32r)
812 assert(!isPICBase && "More than one PIC base?");
822 if (MI->getOperand(2).isImm() &&
823 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
824 !MI->getOperand(4).isReg()) {
825 // lea fi#, lea GV, etc. are all rematerializable.
826 if (!MI->getOperand(1).isReg())
828 unsigned BaseReg = MI->getOperand(1).getReg();
831 // Allow re-materialization of lea PICBase + x.
832 const MachineFunction &MF = *MI->getParent()->getParent();
833 const MachineRegisterInfo &MRI = MF.getRegInfo();
834 return regIsPICBase(BaseReg, MRI);
840 // All other instructions marked M_REMATERIALIZABLE are always trivially
845 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
846 /// would clobber the EFLAGS condition register. Note the result may be
847 /// conservative. If it cannot definitely determine the safety after visiting
848 /// two instructions it assumes it's not safe.
849 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
850 MachineBasicBlock::iterator I) {
851 // It's always safe to clobber EFLAGS at the end of a block.
855 // For compile time consideration, if we are not able to determine the
856 // safety after visiting 2 instructions, we will assume it's not safe.
857 for (unsigned i = 0; i < 2; ++i) {
858 bool SeenDef = false;
859 for (unsigned j = 0, e = I->getNumOperands(); j != e; ++j) {
860 MachineOperand &MO = I->getOperand(j);
863 if (MO.getReg() == X86::EFLAGS) {
871 // This instruction defines EFLAGS, no need to look any further.
875 // If we make it to the end of the block, it's safe to clobber EFLAGS.
880 // Conservative answer.
884 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
885 MachineBasicBlock::iterator I,
887 const MachineInstr *Orig) const {
888 unsigned SubIdx = Orig->getOperand(0).isReg()
889 ? Orig->getOperand(0).getSubReg() : 0;
890 bool ChangeSubIdx = SubIdx != 0;
891 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
892 DestReg = RI.getSubReg(DestReg, SubIdx);
896 // MOV32r0 etc. are implemented with xor which clobbers condition code.
897 // Re-materialize them as movri instructions to avoid side effects.
898 bool Emitted = false;
899 switch (Orig->getOpcode()) {
905 if (!isSafeToClobberEFLAGS(MBB, I)) {
907 switch (Orig->getOpcode()) {
909 case X86::MOV8r0: Opc = X86::MOV8ri; break;
910 case X86::MOV16r0: Opc = X86::MOV16ri; break;
911 case X86::MOV32r0: Opc = X86::MOV32ri; break;
912 case X86::MOV64r0: Opc = X86::MOV64ri32; break;
914 BuildMI(MBB, I, get(Opc), DestReg).addImm(0);
922 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
923 MI->getOperand(0).setReg(DestReg);
928 MachineInstr *NewMI = prior(I);
929 NewMI->getOperand(0).setSubReg(SubIdx);
933 /// isInvariantLoad - Return true if the specified instruction (which is marked
934 /// mayLoad) is loading from a location whose value is invariant across the
935 /// function. For example, loading a value from the constant pool or from
936 /// from the argument area of a function if it does not change. This should
937 /// only return true of *all* loads the instruction does are invariant (if it
938 /// does multiple loads).
939 bool X86InstrInfo::isInvariantLoad(const MachineInstr *MI) const {
940 // This code cares about loads from three cases: constant pool entries,
941 // invariant argument slots, and global stubs. In order to handle these cases
942 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
943 // operand and base our analysis on it. This is safe because the address of
944 // none of these three cases is ever used as anything other than a load base
945 // and X86 doesn't have any instructions that load from multiple places.
947 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
948 const MachineOperand &MO = MI->getOperand(i);
949 // Loads from constant pools are trivially invariant.
954 return isGVStub(MO.getGlobal(), TM);
956 // If this is a load from an invariant stack slot, the load is a constant.
958 const MachineFrameInfo &MFI =
959 *MI->getParent()->getParent()->getFrameInfo();
960 int Idx = MO.getIndex();
961 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
965 // All other instances of these instructions are presumed to have other
970 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
971 /// is not marked dead.
972 static bool hasLiveCondCodeDef(MachineInstr *MI) {
973 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
974 MachineOperand &MO = MI->getOperand(i);
975 if (MO.isReg() && MO.isDef() &&
976 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
983 /// convertToThreeAddress - This method must be implemented by targets that
984 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
985 /// may be able to convert a two-address instruction into a true
986 /// three-address instruction on demand. This allows the X86 target (for
987 /// example) to convert ADD and SHL instructions into LEA instructions if they
988 /// would require register copies due to two-addressness.
990 /// This method returns a null pointer if the transformation cannot be
991 /// performed, otherwise it returns the new instruction.
994 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
995 MachineBasicBlock::iterator &MBBI,
996 LiveVariables *LV) const {
997 MachineInstr *MI = MBBI;
998 MachineFunction &MF = *MI->getParent()->getParent();
999 // All instructions input are two-addr instructions. Get the known operands.
1000 unsigned Dest = MI->getOperand(0).getReg();
1001 unsigned Src = MI->getOperand(1).getReg();
1002 bool isDead = MI->getOperand(0).isDead();
1003 bool isKill = MI->getOperand(1).isKill();
1005 MachineInstr *NewMI = NULL;
1006 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1007 // we have better subtarget support, enable the 16-bit LEA generation here.
1008 bool DisableLEA16 = true;
1010 unsigned MIOpc = MI->getOpcode();
1012 case X86::SHUFPSrri: {
1013 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1014 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1016 unsigned B = MI->getOperand(1).getReg();
1017 unsigned C = MI->getOperand(2).getReg();
1018 if (B != C) return 0;
1019 unsigned A = MI->getOperand(0).getReg();
1020 unsigned M = MI->getOperand(3).getImm();
1021 NewMI = BuildMI(MF, get(X86::PSHUFDri)).addReg(A, true, false, false, isDead)
1022 .addReg(B, false, false, isKill).addImm(M);
1025 case X86::SHL64ri: {
1026 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1027 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1028 // the flags produced by a shift yet, so this is safe.
1029 unsigned ShAmt = MI->getOperand(2).getImm();
1030 if (ShAmt == 0 || ShAmt >= 4) return 0;
1032 NewMI = BuildMI(MF, get(X86::LEA64r)).addReg(Dest, true, false, false, isDead)
1033 .addReg(0).addImm(1 << ShAmt).addReg(Src, false, false, isKill).addImm(0);
1036 case X86::SHL32ri: {
1037 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1038 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1039 // the flags produced by a shift yet, so this is safe.
1040 unsigned ShAmt = MI->getOperand(2).getImm();
1041 if (ShAmt == 0 || ShAmt >= 4) return 0;
1043 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1044 X86::LEA64_32r : X86::LEA32r;
1045 NewMI = BuildMI(MF, get(Opc)).addReg(Dest, true, false, false, isDead)
1046 .addReg(0).addImm(1 << ShAmt)
1047 .addReg(Src, false, false, isKill).addImm(0);
1050 case X86::SHL16ri: {
1051 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1052 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1053 // the flags produced by a shift yet, so this is safe.
1054 unsigned ShAmt = MI->getOperand(2).getImm();
1055 if (ShAmt == 0 || ShAmt >= 4) return 0;
1058 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1059 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1060 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1061 ? X86::LEA64_32r : X86::LEA32r;
1062 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1063 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1065 // Build and insert into an implicit UNDEF value. This is OK because
1066 // well be shifting and then extracting the lower 16-bits.
1067 BuildMI(*MFI, MBBI, get(X86::IMPLICIT_DEF), leaInReg);
1068 MachineInstr *InsMI = BuildMI(*MFI, MBBI, get(X86::INSERT_SUBREG),leaInReg)
1069 .addReg(leaInReg).addReg(Src, false, false, isKill)
1070 .addImm(X86::SUBREG_16BIT);
1072 NewMI = BuildMI(*MFI, MBBI, get(Opc), leaOutReg).addReg(0).addImm(1 << ShAmt)
1073 .addReg(leaInReg, false, false, true).addImm(0);
1075 MachineInstr *ExtMI = BuildMI(*MFI, MBBI, get(X86::EXTRACT_SUBREG))
1076 .addReg(Dest, true, false, false, isDead)
1077 .addReg(leaOutReg, false, false, true).addImm(X86::SUBREG_16BIT);
1079 // Update live variables
1080 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1081 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1083 LV->replaceKillInstruction(Src, MI, InsMI);
1085 LV->replaceKillInstruction(Dest, MI, ExtMI);
1089 NewMI = BuildMI(MF, get(X86::LEA16r)).addReg(Dest, true, false, false, isDead)
1090 .addReg(0).addImm(1 << ShAmt)
1091 .addReg(Src, false, false, isKill).addImm(0);
1096 // The following opcodes also sets the condition code register(s). Only
1097 // convert them to equivalent lea if the condition code register def's
1099 if (hasLiveCondCodeDef(MI))
1102 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1107 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1108 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1109 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1110 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1111 .addReg(Dest, true, false, false, isDead),
1116 case X86::INC64_16r:
1117 if (DisableLEA16) return 0;
1118 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1119 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1120 .addReg(Dest, true, false, false, isDead),
1125 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1126 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1127 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1128 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1129 .addReg(Dest, true, false, false, isDead),
1134 case X86::DEC64_16r:
1135 if (DisableLEA16) return 0;
1136 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1137 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1138 .addReg(Dest, true, false, false, isDead),
1142 case X86::ADD32rr: {
1143 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1144 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1145 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1146 unsigned Src2 = MI->getOperand(2).getReg();
1147 bool isKill2 = MI->getOperand(2).isKill();
1148 NewMI = addRegReg(BuildMI(MF, get(Opc))
1149 .addReg(Dest, true, false, false, isDead),
1150 Src, isKill, Src2, isKill2);
1152 LV->replaceKillInstruction(Src2, MI, NewMI);
1155 case X86::ADD16rr: {
1156 if (DisableLEA16) return 0;
1157 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1158 unsigned Src2 = MI->getOperand(2).getReg();
1159 bool isKill2 = MI->getOperand(2).isKill();
1160 NewMI = addRegReg(BuildMI(MF, get(X86::LEA16r))
1161 .addReg(Dest, true, false, false, isDead),
1162 Src, isKill, Src2, isKill2);
1164 LV->replaceKillInstruction(Src2, MI, NewMI);
1167 case X86::ADD64ri32:
1169 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1170 if (MI->getOperand(2).isImm())
1171 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA64r))
1172 .addReg(Dest, true, false, false, isDead),
1173 Src, isKill, MI->getOperand(2).getImm());
1177 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1178 if (MI->getOperand(2).isImm()) {
1179 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1180 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1181 .addReg(Dest, true, false, false, isDead),
1182 Src, isKill, MI->getOperand(2).getImm());
1187 if (DisableLEA16) return 0;
1188 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1189 if (MI->getOperand(2).isImm())
1190 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1191 .addReg(Dest, true, false, false, isDead),
1192 Src, isKill, MI->getOperand(2).getImm());
1195 if (DisableLEA16) return 0;
1197 case X86::SHL64ri: {
1198 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImm() &&
1199 "Unknown shl instruction!");
1200 unsigned ShAmt = MI->getOperand(2).getImm();
1201 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1203 AM.Scale = 1 << ShAmt;
1205 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1206 : (MIOpc == X86::SHL32ri
1207 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1208 NewMI = addFullAddress(BuildMI(MF, get(Opc))
1209 .addReg(Dest, true, false, false, isDead), AM);
1211 NewMI->getOperand(3).setIsKill(true);
1219 if (!NewMI) return 0;
1221 if (LV) { // Update live variables
1223 LV->replaceKillInstruction(Src, MI, NewMI);
1225 LV->replaceKillInstruction(Dest, MI, NewMI);
1228 MFI->insert(MBBI, NewMI); // Insert the new inst
1232 /// commuteInstruction - We have a few instructions that must be hacked on to
1236 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1237 switch (MI->getOpcode()) {
1238 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1239 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1240 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1241 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1242 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1243 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1246 switch (MI->getOpcode()) {
1247 default: assert(0 && "Unreachable!");
1248 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1249 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1250 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1251 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1252 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1253 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1255 unsigned Amt = MI->getOperand(3).getImm();
1257 MachineFunction &MF = *MI->getParent()->getParent();
1258 MI = MF.CloneMachineInstr(MI);
1261 MI->setDesc(get(Opc));
1262 MI->getOperand(3).setImm(Size-Amt);
1263 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1265 case X86::CMOVB16rr:
1266 case X86::CMOVB32rr:
1267 case X86::CMOVB64rr:
1268 case X86::CMOVAE16rr:
1269 case X86::CMOVAE32rr:
1270 case X86::CMOVAE64rr:
1271 case X86::CMOVE16rr:
1272 case X86::CMOVE32rr:
1273 case X86::CMOVE64rr:
1274 case X86::CMOVNE16rr:
1275 case X86::CMOVNE32rr:
1276 case X86::CMOVNE64rr:
1277 case X86::CMOVBE16rr:
1278 case X86::CMOVBE32rr:
1279 case X86::CMOVBE64rr:
1280 case X86::CMOVA16rr:
1281 case X86::CMOVA32rr:
1282 case X86::CMOVA64rr:
1283 case X86::CMOVL16rr:
1284 case X86::CMOVL32rr:
1285 case X86::CMOVL64rr:
1286 case X86::CMOVGE16rr:
1287 case X86::CMOVGE32rr:
1288 case X86::CMOVGE64rr:
1289 case X86::CMOVLE16rr:
1290 case X86::CMOVLE32rr:
1291 case X86::CMOVLE64rr:
1292 case X86::CMOVG16rr:
1293 case X86::CMOVG32rr:
1294 case X86::CMOVG64rr:
1295 case X86::CMOVS16rr:
1296 case X86::CMOVS32rr:
1297 case X86::CMOVS64rr:
1298 case X86::CMOVNS16rr:
1299 case X86::CMOVNS32rr:
1300 case X86::CMOVNS64rr:
1301 case X86::CMOVP16rr:
1302 case X86::CMOVP32rr:
1303 case X86::CMOVP64rr:
1304 case X86::CMOVNP16rr:
1305 case X86::CMOVNP32rr:
1306 case X86::CMOVNP64rr: {
1308 switch (MI->getOpcode()) {
1310 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1311 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1312 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1313 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1314 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1315 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1316 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1317 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1318 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1319 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1320 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1321 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1322 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1323 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1324 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1325 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1326 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1327 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1328 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1329 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1330 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1331 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1332 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1333 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1334 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1335 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1336 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1337 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1338 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1339 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1340 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1341 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1342 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1343 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1344 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1345 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1346 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1347 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1348 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1349 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1350 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1351 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1354 MachineFunction &MF = *MI->getParent()->getParent();
1355 MI = MF.CloneMachineInstr(MI);
1358 MI->setDesc(get(Opc));
1359 // Fallthrough intended.
1362 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1366 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1368 default: return X86::COND_INVALID;
1369 case X86::JE: return X86::COND_E;
1370 case X86::JNE: return X86::COND_NE;
1371 case X86::JL: return X86::COND_L;
1372 case X86::JLE: return X86::COND_LE;
1373 case X86::JG: return X86::COND_G;
1374 case X86::JGE: return X86::COND_GE;
1375 case X86::JB: return X86::COND_B;
1376 case X86::JBE: return X86::COND_BE;
1377 case X86::JA: return X86::COND_A;
1378 case X86::JAE: return X86::COND_AE;
1379 case X86::JS: return X86::COND_S;
1380 case X86::JNS: return X86::COND_NS;
1381 case X86::JP: return X86::COND_P;
1382 case X86::JNP: return X86::COND_NP;
1383 case X86::JO: return X86::COND_O;
1384 case X86::JNO: return X86::COND_NO;
1385 case X86::JC: return X86::COND_C;
1386 case X86::JNC: return X86::COND_NC;
1390 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1392 default: assert(0 && "Illegal condition code!");
1393 case X86::COND_E: return X86::JE;
1394 case X86::COND_NE: return X86::JNE;
1395 case X86::COND_L: return X86::JL;
1396 case X86::COND_LE: return X86::JLE;
1397 case X86::COND_G: return X86::JG;
1398 case X86::COND_GE: return X86::JGE;
1399 case X86::COND_B: return X86::JB;
1400 case X86::COND_BE: return X86::JBE;
1401 case X86::COND_A: return X86::JA;
1402 case X86::COND_AE: return X86::JAE;
1403 case X86::COND_S: return X86::JS;
1404 case X86::COND_NS: return X86::JNS;
1405 case X86::COND_P: return X86::JP;
1406 case X86::COND_NP: return X86::JNP;
1407 case X86::COND_O: return X86::JO;
1408 case X86::COND_NO: return X86::JNO;
1409 case X86::COND_C: return X86::JC;
1410 case X86::COND_NC: return X86::JNC;
1414 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1415 /// e.g. turning COND_E to COND_NE.
1416 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1418 default: assert(0 && "Illegal condition code!");
1419 case X86::COND_E: return X86::COND_NE;
1420 case X86::COND_NE: return X86::COND_E;
1421 case X86::COND_L: return X86::COND_GE;
1422 case X86::COND_LE: return X86::COND_G;
1423 case X86::COND_G: return X86::COND_LE;
1424 case X86::COND_GE: return X86::COND_L;
1425 case X86::COND_B: return X86::COND_AE;
1426 case X86::COND_BE: return X86::COND_A;
1427 case X86::COND_A: return X86::COND_BE;
1428 case X86::COND_AE: return X86::COND_B;
1429 case X86::COND_S: return X86::COND_NS;
1430 case X86::COND_NS: return X86::COND_S;
1431 case X86::COND_P: return X86::COND_NP;
1432 case X86::COND_NP: return X86::COND_P;
1433 case X86::COND_O: return X86::COND_NO;
1434 case X86::COND_NO: return X86::COND_O;
1435 case X86::COND_C: return X86::COND_NC;
1436 case X86::COND_NC: return X86::COND_C;
1440 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1441 const TargetInstrDesc &TID = MI->getDesc();
1442 if (!TID.isTerminator()) return false;
1444 // Conditional branch is a special case.
1445 if (TID.isBranch() && !TID.isBarrier())
1447 if (!TID.isPredicable())
1449 return !isPredicated(MI);
1452 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1453 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1454 const X86InstrInfo &TII) {
1455 if (MI->getOpcode() == X86::FP_REG_KILL)
1457 return TII.isUnpredicatedTerminator(MI);
1460 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1461 MachineBasicBlock *&TBB,
1462 MachineBasicBlock *&FBB,
1463 SmallVectorImpl<MachineOperand> &Cond) const {
1464 // Start from the bottom of the block and work up, examining the
1465 // terminator instructions.
1466 MachineBasicBlock::iterator I = MBB.end();
1467 while (I != MBB.begin()) {
1469 // Working from the bottom, when we see a non-terminator
1470 // instruction, we're done.
1471 if (!isBrAnalysisUnpredicatedTerminator(I, *this))
1473 // A terminator that isn't a branch can't easily be handled
1474 // by this analysis.
1475 if (!I->getDesc().isBranch())
1477 // Handle unconditional branches.
1478 if (I->getOpcode() == X86::JMP) {
1479 // If the block has any instructions after a JMP, delete them.
1480 while (next(I) != MBB.end())
1481 next(I)->eraseFromParent();
1484 // Delete the JMP if it's equivalent to a fall-through.
1485 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1487 I->eraseFromParent();
1491 // TBB is used to indicate the unconditinal destination.
1492 TBB = I->getOperand(0).getMBB();
1495 // Handle conditional branches.
1496 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1497 if (BranchCode == X86::COND_INVALID)
1498 return true; // Can't handle indirect branch.
1499 // Working from the bottom, handle the first conditional branch.
1502 TBB = I->getOperand(0).getMBB();
1503 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1506 // Handle subsequent conditional branches. Only handle the case
1507 // where all conditional branches branch to the same destination
1508 // and their condition opcodes fit one of the special
1509 // multi-branch idioms.
1510 assert(Cond.size() == 1);
1512 // Only handle the case where all conditional branches branch to
1513 // the same destination.
1514 if (TBB != I->getOperand(0).getMBB())
1516 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1517 // If the conditions are the same, we can leave them alone.
1518 if (OldBranchCode == BranchCode)
1520 // If they differ, see if they fit one of the known patterns.
1521 // Theoretically we could handle more patterns here, but
1522 // we shouldn't expect to see them if instruction selection
1523 // has done a reasonable job.
1524 if ((OldBranchCode == X86::COND_NP &&
1525 BranchCode == X86::COND_E) ||
1526 (OldBranchCode == X86::COND_E &&
1527 BranchCode == X86::COND_NP))
1528 BranchCode = X86::COND_NP_OR_E;
1529 else if ((OldBranchCode == X86::COND_P &&
1530 BranchCode == X86::COND_NE) ||
1531 (OldBranchCode == X86::COND_NE &&
1532 BranchCode == X86::COND_P))
1533 BranchCode = X86::COND_NE_OR_P;
1536 // Update the MachineOperand.
1537 Cond[0].setImm(BranchCode);
1543 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1544 MachineBasicBlock::iterator I = MBB.end();
1547 while (I != MBB.begin()) {
1549 if (I->getOpcode() != X86::JMP &&
1550 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1552 // Remove the branch.
1553 I->eraseFromParent();
1561 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
1562 const MachineOperand &MO) {
1564 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
1565 MO.isKill(), MO.isDead(), MO.getSubReg());
1566 else if (MO.isImm())
1567 MIB = MIB.addImm(MO.getImm());
1569 MIB = MIB.addFrameIndex(MO.getIndex());
1570 else if (MO.isGlobal())
1571 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
1572 else if (MO.isCPI())
1573 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
1574 else if (MO.isJTI())
1575 MIB = MIB.addJumpTableIndex(MO.getIndex());
1576 else if (MO.isSymbol())
1577 MIB = MIB.addExternalSymbol(MO.getSymbolName());
1579 assert(0 && "Unknown operand for X86InstrAddOperand!");
1585 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1586 MachineBasicBlock *FBB,
1587 const SmallVectorImpl<MachineOperand> &Cond) const {
1588 // Shouldn't be a fall through.
1589 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1590 assert((Cond.size() == 1 || Cond.size() == 0) &&
1591 "X86 branch conditions have one component!");
1594 // Unconditional branch?
1595 assert(!FBB && "Unconditional branch with multiple successors!");
1596 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
1600 // Conditional branch.
1602 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1604 case X86::COND_NP_OR_E:
1605 // Synthesize NP_OR_E with two branches.
1606 BuildMI(&MBB, get(X86::JNP)).addMBB(TBB);
1608 BuildMI(&MBB, get(X86::JE)).addMBB(TBB);
1611 case X86::COND_NE_OR_P:
1612 // Synthesize NE_OR_P with two branches.
1613 BuildMI(&MBB, get(X86::JNE)).addMBB(TBB);
1615 BuildMI(&MBB, get(X86::JP)).addMBB(TBB);
1619 unsigned Opc = GetCondBranchFromCond(CC);
1620 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1625 // Two-way Conditional branch. Insert the second branch.
1626 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
1632 bool X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1633 MachineBasicBlock::iterator MI,
1634 unsigned DestReg, unsigned SrcReg,
1635 const TargetRegisterClass *DestRC,
1636 const TargetRegisterClass *SrcRC) const {
1637 if (DestRC == SrcRC) {
1639 if (DestRC == &X86::GR64RegClass) {
1641 } else if (DestRC == &X86::GR32RegClass) {
1643 } else if (DestRC == &X86::GR16RegClass) {
1645 } else if (DestRC == &X86::GR8RegClass) {
1647 } else if (DestRC == &X86::GR32_RegClass) {
1648 Opc = X86::MOV32_rr;
1649 } else if (DestRC == &X86::GR16_RegClass) {
1650 Opc = X86::MOV16_rr;
1651 } else if (DestRC == &X86::RFP32RegClass) {
1652 Opc = X86::MOV_Fp3232;
1653 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1654 Opc = X86::MOV_Fp6464;
1655 } else if (DestRC == &X86::RFP80RegClass) {
1656 Opc = X86::MOV_Fp8080;
1657 } else if (DestRC == &X86::FR32RegClass) {
1658 Opc = X86::FsMOVAPSrr;
1659 } else if (DestRC == &X86::FR64RegClass) {
1660 Opc = X86::FsMOVAPDrr;
1661 } else if (DestRC == &X86::VR128RegClass) {
1662 Opc = X86::MOVAPSrr;
1663 } else if (DestRC == &X86::VR64RegClass) {
1664 Opc = X86::MMX_MOVQ64rr;
1668 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
1672 // Moving EFLAGS to / from another register requires a push and a pop.
1673 if (SrcRC == &X86::CCRRegClass) {
1674 if (SrcReg != X86::EFLAGS)
1676 if (DestRC == &X86::GR64RegClass) {
1677 BuildMI(MBB, MI, get(X86::PUSHFQ));
1678 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
1680 } else if (DestRC == &X86::GR32RegClass) {
1681 BuildMI(MBB, MI, get(X86::PUSHFD));
1682 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
1685 } else if (DestRC == &X86::CCRRegClass) {
1686 if (DestReg != X86::EFLAGS)
1688 if (SrcRC == &X86::GR64RegClass) {
1689 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
1690 BuildMI(MBB, MI, get(X86::POPFQ));
1692 } else if (SrcRC == &X86::GR32RegClass) {
1693 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
1694 BuildMI(MBB, MI, get(X86::POPFD));
1699 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1700 if (SrcRC == &X86::RSTRegClass) {
1701 // Copying from ST(0)/ST(1).
1702 if (SrcReg != X86::ST0 && SrcReg != X86::ST1)
1703 // Can only copy from ST(0)/ST(1) right now
1705 bool isST0 = SrcReg == X86::ST0;
1707 if (DestRC == &X86::RFP32RegClass)
1708 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1709 else if (DestRC == &X86::RFP64RegClass)
1710 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1712 if (DestRC != &X86::RFP80RegClass)
1714 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1716 BuildMI(MBB, MI, get(Opc), DestReg);
1720 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1721 if (DestRC == &X86::RSTRegClass) {
1722 // Copying to ST(0). FIXME: handle ST(1) also
1723 if (DestReg != X86::ST0)
1724 // Can only copy to TOS right now
1727 if (SrcRC == &X86::RFP32RegClass)
1728 Opc = X86::FpSET_ST0_32;
1729 else if (SrcRC == &X86::RFP64RegClass)
1730 Opc = X86::FpSET_ST0_64;
1732 if (SrcRC != &X86::RFP80RegClass)
1734 Opc = X86::FpSET_ST0_80;
1736 BuildMI(MBB, MI, get(Opc)).addReg(SrcReg);
1740 // Not yet supported!
1744 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1745 bool isStackAligned) {
1747 if (RC == &X86::GR64RegClass) {
1749 } else if (RC == &X86::GR32RegClass) {
1751 } else if (RC == &X86::GR16RegClass) {
1753 } else if (RC == &X86::GR8RegClass) {
1755 } else if (RC == &X86::GR32_RegClass) {
1756 Opc = X86::MOV32_mr;
1757 } else if (RC == &X86::GR16_RegClass) {
1758 Opc = X86::MOV16_mr;
1759 } else if (RC == &X86::RFP80RegClass) {
1760 Opc = X86::ST_FpP80m; // pops
1761 } else if (RC == &X86::RFP64RegClass) {
1762 Opc = X86::ST_Fp64m;
1763 } else if (RC == &X86::RFP32RegClass) {
1764 Opc = X86::ST_Fp32m;
1765 } else if (RC == &X86::FR32RegClass) {
1767 } else if (RC == &X86::FR64RegClass) {
1769 } else if (RC == &X86::VR128RegClass) {
1770 // If stack is realigned we can use aligned stores.
1771 Opc = isStackAligned ? X86::MOVAPSmr : X86::MOVUPSmr;
1772 } else if (RC == &X86::VR64RegClass) {
1773 Opc = X86::MMX_MOVQ64mr;
1775 assert(0 && "Unknown regclass");
1782 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1783 MachineBasicBlock::iterator MI,
1784 unsigned SrcReg, bool isKill, int FrameIdx,
1785 const TargetRegisterClass *RC) const {
1786 const MachineFunction &MF = *MBB.getParent();
1787 bool isAligned = (RI.getStackAlignment() >= 16) ||
1788 RI.needsStackRealignment(MF);
1789 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1790 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
1791 .addReg(SrcReg, false, false, isKill);
1794 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1796 SmallVectorImpl<MachineOperand> &Addr,
1797 const TargetRegisterClass *RC,
1798 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1799 bool isAligned = (RI.getStackAlignment() >= 16) ||
1800 RI.needsStackRealignment(MF);
1801 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1802 MachineInstrBuilder MIB = BuildMI(MF, get(Opc));
1803 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1804 MIB = X86InstrAddOperand(MIB, Addr[i]);
1805 MIB.addReg(SrcReg, false, false, isKill);
1806 NewMIs.push_back(MIB);
1809 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1810 bool isStackAligned) {
1812 if (RC == &X86::GR64RegClass) {
1814 } else if (RC == &X86::GR32RegClass) {
1816 } else if (RC == &X86::GR16RegClass) {
1818 } else if (RC == &X86::GR8RegClass) {
1820 } else if (RC == &X86::GR32_RegClass) {
1821 Opc = X86::MOV32_rm;
1822 } else if (RC == &X86::GR16_RegClass) {
1823 Opc = X86::MOV16_rm;
1824 } else if (RC == &X86::RFP80RegClass) {
1825 Opc = X86::LD_Fp80m;
1826 } else if (RC == &X86::RFP64RegClass) {
1827 Opc = X86::LD_Fp64m;
1828 } else if (RC == &X86::RFP32RegClass) {
1829 Opc = X86::LD_Fp32m;
1830 } else if (RC == &X86::FR32RegClass) {
1832 } else if (RC == &X86::FR64RegClass) {
1834 } else if (RC == &X86::VR128RegClass) {
1835 // If stack is realigned we can use aligned loads.
1836 Opc = isStackAligned ? X86::MOVAPSrm : X86::MOVUPSrm;
1837 } else if (RC == &X86::VR64RegClass) {
1838 Opc = X86::MMX_MOVQ64rm;
1840 assert(0 && "Unknown regclass");
1847 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1848 MachineBasicBlock::iterator MI,
1849 unsigned DestReg, int FrameIdx,
1850 const TargetRegisterClass *RC) const{
1851 const MachineFunction &MF = *MBB.getParent();
1852 bool isAligned = (RI.getStackAlignment() >= 16) ||
1853 RI.needsStackRealignment(MF);
1854 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1855 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
1858 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1859 SmallVectorImpl<MachineOperand> &Addr,
1860 const TargetRegisterClass *RC,
1861 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1862 bool isAligned = (RI.getStackAlignment() >= 16) ||
1863 RI.needsStackRealignment(MF);
1864 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1865 MachineInstrBuilder MIB = BuildMI(MF, get(Opc), DestReg);
1866 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1867 MIB = X86InstrAddOperand(MIB, Addr[i]);
1868 NewMIs.push_back(MIB);
1871 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1872 MachineBasicBlock::iterator MI,
1873 const std::vector<CalleeSavedInfo> &CSI) const {
1877 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1878 unsigned SlotSize = is64Bit ? 8 : 4;
1880 MachineFunction &MF = *MBB.getParent();
1881 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1882 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1884 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1885 for (unsigned i = CSI.size(); i != 0; --i) {
1886 unsigned Reg = CSI[i-1].getReg();
1887 // Add the callee-saved register as live-in. It's killed at the spill.
1889 BuildMI(MBB, MI, get(Opc))
1890 .addReg(Reg, /*isDef=*/false, /*isImp=*/false, /*isKill=*/true);
1895 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1896 MachineBasicBlock::iterator MI,
1897 const std::vector<CalleeSavedInfo> &CSI) const {
1901 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1903 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1904 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1905 unsigned Reg = CSI[i].getReg();
1906 BuildMI(MBB, MI, get(Opc), Reg);
1911 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
1912 const SmallVector<MachineOperand,4> &MOs,
1913 MachineInstr *MI, const TargetInstrInfo &TII) {
1914 // Create the base instruction with the memory operand as the first part.
1915 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), true);
1916 MachineInstrBuilder MIB(NewMI);
1917 unsigned NumAddrOps = MOs.size();
1918 for (unsigned i = 0; i != NumAddrOps; ++i)
1919 MIB = X86InstrAddOperand(MIB, MOs[i]);
1920 if (NumAddrOps < 4) // FrameIndex only
1921 MIB.addImm(1).addReg(0).addImm(0);
1923 // Loop over the rest of the ri operands, converting them over.
1924 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1925 for (unsigned i = 0; i != NumOps; ++i) {
1926 MachineOperand &MO = MI->getOperand(i+2);
1927 MIB = X86InstrAddOperand(MIB, MO);
1929 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1930 MachineOperand &MO = MI->getOperand(i);
1931 MIB = X86InstrAddOperand(MIB, MO);
1936 static MachineInstr *FuseInst(MachineFunction &MF,
1937 unsigned Opcode, unsigned OpNo,
1938 const SmallVector<MachineOperand,4> &MOs,
1939 MachineInstr *MI, const TargetInstrInfo &TII) {
1940 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), true);
1941 MachineInstrBuilder MIB(NewMI);
1943 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1944 MachineOperand &MO = MI->getOperand(i);
1946 assert(MO.isReg() && "Expected to fold into reg operand!");
1947 unsigned NumAddrOps = MOs.size();
1948 for (unsigned i = 0; i != NumAddrOps; ++i)
1949 MIB = X86InstrAddOperand(MIB, MOs[i]);
1950 if (NumAddrOps < 4) // FrameIndex only
1951 MIB.addImm(1).addReg(0).addImm(0);
1953 MIB = X86InstrAddOperand(MIB, MO);
1959 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
1960 const SmallVector<MachineOperand,4> &MOs,
1962 MachineFunction &MF = *MI->getParent()->getParent();
1963 MachineInstrBuilder MIB = BuildMI(MF, TII.get(Opcode));
1965 unsigned NumAddrOps = MOs.size();
1966 for (unsigned i = 0; i != NumAddrOps; ++i)
1967 MIB = X86InstrAddOperand(MIB, MOs[i]);
1968 if (NumAddrOps < 4) // FrameIndex only
1969 MIB.addImm(1).addReg(0).addImm(0);
1970 return MIB.addImm(0);
1974 X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
1975 MachineInstr *MI, unsigned i,
1976 const SmallVector<MachineOperand,4> &MOs) const{
1977 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
1978 bool isTwoAddrFold = false;
1979 unsigned NumOps = MI->getDesc().getNumOperands();
1980 bool isTwoAddr = NumOps > 1 &&
1981 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
1983 MachineInstr *NewMI = NULL;
1984 // Folding a memory location into the two-address part of a two-address
1985 // instruction is different than folding it other places. It requires
1986 // replacing the *two* registers with the memory location.
1987 if (isTwoAddr && NumOps >= 2 && i < 2 &&
1988 MI->getOperand(0).isReg() &&
1989 MI->getOperand(1).isReg() &&
1990 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
1991 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
1992 isTwoAddrFold = true;
1993 } else if (i == 0) { // If operand 0
1994 if (MI->getOpcode() == X86::MOV16r0)
1995 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
1996 else if (MI->getOpcode() == X86::MOV32r0)
1997 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
1998 else if (MI->getOpcode() == X86::MOV64r0)
1999 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2000 else if (MI->getOpcode() == X86::MOV8r0)
2001 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2005 OpcodeTablePtr = &RegOp2MemOpTable0;
2006 } else if (i == 1) {
2007 OpcodeTablePtr = &RegOp2MemOpTable1;
2008 } else if (i == 2) {
2009 OpcodeTablePtr = &RegOp2MemOpTable2;
2012 // If table selected...
2013 if (OpcodeTablePtr) {
2014 // Find the Opcode to fuse
2015 DenseMap<unsigned*, unsigned>::iterator I =
2016 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
2017 if (I != OpcodeTablePtr->end()) {
2019 NewMI = FuseTwoAddrInst(MF, I->second, MOs, MI, *this);
2021 NewMI = FuseInst(MF, I->second, i, MOs, MI, *this);
2027 if (PrintFailedFusing)
2028 cerr << "We failed to fuse operand " << i << *MI;
2033 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
2035 const SmallVectorImpl<unsigned> &Ops,
2036 int FrameIndex) const {
2037 // Check switch flag
2038 if (NoFusing) return NULL;
2040 const MachineFrameInfo *MFI = MF.getFrameInfo();
2041 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2042 // FIXME: Move alignment requirement into tables?
2043 if (Alignment < 16) {
2044 switch (MI->getOpcode()) {
2046 // Not always safe to fold movsd into these instructions since their load
2047 // folding variants expects the address to be 16 byte aligned.
2048 case X86::FsANDNPDrr:
2049 case X86::FsANDNPSrr:
2050 case X86::FsANDPDrr:
2051 case X86::FsANDPSrr:
2054 case X86::FsXORPDrr:
2055 case X86::FsXORPSrr:
2060 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2061 unsigned NewOpc = 0;
2062 switch (MI->getOpcode()) {
2063 default: return NULL;
2064 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2065 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2066 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2067 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2069 // Change to CMPXXri r, 0 first.
2070 MI->setDesc(get(NewOpc));
2071 MI->getOperand(1).ChangeToImmediate(0);
2072 } else if (Ops.size() != 1)
2075 SmallVector<MachineOperand,4> MOs;
2076 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2077 return foldMemoryOperand(MF, MI, Ops[0], MOs);
2080 MachineInstr* X86InstrInfo::foldMemoryOperand(MachineFunction &MF,
2082 const SmallVectorImpl<unsigned> &Ops,
2083 MachineInstr *LoadMI) const {
2084 // Check switch flag
2085 if (NoFusing) return NULL;
2087 // Determine the alignment of the load.
2088 unsigned Alignment = 0;
2089 if (LoadMI->hasOneMemOperand())
2090 Alignment = LoadMI->memoperands_begin()->getAlignment();
2092 // FIXME: Move alignment requirement into tables?
2093 if (Alignment < 16) {
2094 switch (MI->getOpcode()) {
2096 // Not always safe to fold movsd into these instructions since their load
2097 // folding variants expects the address to be 16 byte aligned.
2098 case X86::FsANDNPDrr:
2099 case X86::FsANDNPSrr:
2100 case X86::FsANDPDrr:
2101 case X86::FsANDPSrr:
2104 case X86::FsXORPDrr:
2105 case X86::FsXORPSrr:
2110 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2111 unsigned NewOpc = 0;
2112 switch (MI->getOpcode()) {
2113 default: return NULL;
2114 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2115 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2116 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2117 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2119 // Change to CMPXXri r, 0 first.
2120 MI->setDesc(get(NewOpc));
2121 MI->getOperand(1).ChangeToImmediate(0);
2122 } else if (Ops.size() != 1)
2125 SmallVector<MachineOperand,4> MOs;
2126 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2127 for (unsigned i = NumOps - 4; i != NumOps; ++i)
2128 MOs.push_back(LoadMI->getOperand(i));
2129 return foldMemoryOperand(MF, MI, Ops[0], MOs);
2133 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2134 const SmallVectorImpl<unsigned> &Ops) const {
2135 // Check switch flag
2136 if (NoFusing) return 0;
2138 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2139 switch (MI->getOpcode()) {
2140 default: return false;
2149 if (Ops.size() != 1)
2152 unsigned OpNum = Ops[0];
2153 unsigned Opc = MI->getOpcode();
2154 unsigned NumOps = MI->getDesc().getNumOperands();
2155 bool isTwoAddr = NumOps > 1 &&
2156 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2158 // Folding a memory location into the two-address part of a two-address
2159 // instruction is different than folding it other places. It requires
2160 // replacing the *two* registers with the memory location.
2161 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2162 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2163 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2164 } else if (OpNum == 0) { // If operand 0
2173 OpcodeTablePtr = &RegOp2MemOpTable0;
2174 } else if (OpNum == 1) {
2175 OpcodeTablePtr = &RegOp2MemOpTable1;
2176 } else if (OpNum == 2) {
2177 OpcodeTablePtr = &RegOp2MemOpTable2;
2180 if (OpcodeTablePtr) {
2181 // Find the Opcode to fuse
2182 DenseMap<unsigned*, unsigned>::iterator I =
2183 OpcodeTablePtr->find((unsigned*)Opc);
2184 if (I != OpcodeTablePtr->end())
2190 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2191 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2192 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2193 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2194 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2195 if (I == MemOp2RegOpTable.end())
2197 unsigned Opc = I->second.first;
2198 unsigned Index = I->second.second & 0xf;
2199 bool FoldedLoad = I->second.second & (1 << 4);
2200 bool FoldedStore = I->second.second & (1 << 5);
2201 if (UnfoldLoad && !FoldedLoad)
2203 UnfoldLoad &= FoldedLoad;
2204 if (UnfoldStore && !FoldedStore)
2206 UnfoldStore &= FoldedStore;
2208 const TargetInstrDesc &TID = get(Opc);
2209 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2210 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2211 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2212 SmallVector<MachineOperand,4> AddrOps;
2213 SmallVector<MachineOperand,2> BeforeOps;
2214 SmallVector<MachineOperand,2> AfterOps;
2215 SmallVector<MachineOperand,4> ImpOps;
2216 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2217 MachineOperand &Op = MI->getOperand(i);
2218 if (i >= Index && i < Index+4)
2219 AddrOps.push_back(Op);
2220 else if (Op.isReg() && Op.isImplicit())
2221 ImpOps.push_back(Op);
2223 BeforeOps.push_back(Op);
2225 AfterOps.push_back(Op);
2228 // Emit the load instruction.
2230 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2232 // Address operands cannot be marked isKill.
2233 for (unsigned i = 1; i != 5; ++i) {
2234 MachineOperand &MO = NewMIs[0]->getOperand(i);
2236 MO.setIsKill(false);
2241 // Emit the data processing instruction.
2242 MachineInstr *DataMI = MF.CreateMachineInstr(TID, true);
2243 MachineInstrBuilder MIB(DataMI);
2246 MIB.addReg(Reg, true);
2247 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2248 MIB = X86InstrAddOperand(MIB, BeforeOps[i]);
2251 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2252 MIB = X86InstrAddOperand(MIB, AfterOps[i]);
2253 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2254 MachineOperand &MO = ImpOps[i];
2255 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2257 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2258 unsigned NewOpc = 0;
2259 switch (DataMI->getOpcode()) {
2261 case X86::CMP64ri32:
2265 MachineOperand &MO0 = DataMI->getOperand(0);
2266 MachineOperand &MO1 = DataMI->getOperand(1);
2267 if (MO1.getImm() == 0) {
2268 switch (DataMI->getOpcode()) {
2270 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2271 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2272 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2273 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2275 DataMI->setDesc(get(NewOpc));
2276 MO1.ChangeToRegister(MO0.getReg(), false);
2280 NewMIs.push_back(DataMI);
2282 // Emit the store instruction.
2284 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2285 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2286 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2287 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2294 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2295 SmallVectorImpl<SDNode*> &NewNodes) const {
2296 if (!N->isMachineOpcode())
2299 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2300 MemOp2RegOpTable.find((unsigned*)N->getMachineOpcode());
2301 if (I == MemOp2RegOpTable.end())
2303 unsigned Opc = I->second.first;
2304 unsigned Index = I->second.second & 0xf;
2305 bool FoldedLoad = I->second.second & (1 << 4);
2306 bool FoldedStore = I->second.second & (1 << 5);
2307 const TargetInstrDesc &TID = get(Opc);
2308 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2309 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2310 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2311 std::vector<SDValue> AddrOps;
2312 std::vector<SDValue> BeforeOps;
2313 std::vector<SDValue> AfterOps;
2314 unsigned NumOps = N->getNumOperands();
2315 for (unsigned i = 0; i != NumOps-1; ++i) {
2316 SDValue Op = N->getOperand(i);
2317 if (i >= Index && i < Index+4)
2318 AddrOps.push_back(Op);
2320 BeforeOps.push_back(Op);
2322 AfterOps.push_back(Op);
2324 SDValue Chain = N->getOperand(NumOps-1);
2325 AddrOps.push_back(Chain);
2327 // Emit the load instruction.
2329 const MachineFunction &MF = DAG.getMachineFunction();
2331 MVT VT = *RC->vt_begin();
2332 bool isAligned = (RI.getStackAlignment() >= 16) ||
2333 RI.needsStackRealignment(MF);
2334 Load = DAG.getTargetNode(getLoadRegOpcode(RC, isAligned),
2336 &AddrOps[0], AddrOps.size());
2337 NewNodes.push_back(Load);
2340 // Emit the data processing instruction.
2341 std::vector<MVT> VTs;
2342 const TargetRegisterClass *DstRC = 0;
2343 if (TID.getNumDefs() > 0) {
2344 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2345 DstRC = DstTOI.isLookupPtrRegClass()
2346 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2347 VTs.push_back(*DstRC->vt_begin());
2349 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2350 MVT VT = N->getValueType(i);
2351 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2355 BeforeOps.push_back(SDValue(Load, 0));
2356 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2357 SDNode *NewNode= DAG.getTargetNode(Opc, VTs, &BeforeOps[0], BeforeOps.size());
2358 NewNodes.push_back(NewNode);
2360 // Emit the store instruction.
2363 AddrOps.push_back(SDValue(NewNode, 0));
2364 AddrOps.push_back(Chain);
2365 bool isAligned = (RI.getStackAlignment() >= 16) ||
2366 RI.needsStackRealignment(MF);
2367 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, isAligned),
2368 MVT::Other, &AddrOps[0], AddrOps.size());
2369 NewNodes.push_back(Store);
2375 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2376 bool UnfoldLoad, bool UnfoldStore) const {
2377 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2378 MemOp2RegOpTable.find((unsigned*)Opc);
2379 if (I == MemOp2RegOpTable.end())
2381 bool FoldedLoad = I->second.second & (1 << 4);
2382 bool FoldedStore = I->second.second & (1 << 5);
2383 if (UnfoldLoad && !FoldedLoad)
2385 if (UnfoldStore && !FoldedStore)
2387 return I->second.first;
2390 bool X86InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
2391 if (MBB.empty()) return false;
2393 switch (MBB.back().getOpcode()) {
2394 case X86::TCRETURNri:
2395 case X86::TCRETURNdi:
2396 case X86::RET: // Return.
2401 case X86::JMP: // Uncond branch.
2402 case X86::JMP32r: // Indirect branch.
2403 case X86::JMP64r: // Indirect branch (64-bit).
2404 case X86::JMP32m: // Indirect branch through mem.
2405 case X86::JMP64m: // Indirect branch through mem (64-bit).
2407 default: return false;
2412 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
2413 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2414 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
2415 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
2417 Cond[0].setImm(GetOppositeBranchCondition(CC));
2422 IgnoreRegisterClassBarriers(const TargetRegisterClass *RC) const {
2423 // FIXME: Ignore bariers of x87 stack registers for now. We can't
2424 // allow any loads of these registers before FpGet_ST0_80.
2425 return RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
2426 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass;
2429 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
2430 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
2431 if (Subtarget->is64Bit())
2432 return &X86::GR64RegClass;
2434 return &X86::GR32RegClass;
2437 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2438 switch (Desc->TSFlags & X86II::ImmMask) {
2439 case X86II::Imm8: return 1;
2440 case X86II::Imm16: return 2;
2441 case X86II::Imm32: return 4;
2442 case X86II::Imm64: return 8;
2443 default: assert(0 && "Immediate size not set!");
2448 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2449 /// e.g. r8, xmm8, etc.
2450 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2451 if (!MO.isReg()) return false;
2452 switch (MO.getReg()) {
2454 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2455 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2456 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2457 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2458 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2459 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2460 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2461 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2462 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2463 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2470 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2471 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2472 /// size, and 3) use of X86-64 extended registers.
2473 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2475 const TargetInstrDesc &Desc = MI.getDesc();
2477 // Pseudo instructions do not need REX prefix byte.
2478 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2480 if (Desc.TSFlags & X86II::REX_W)
2483 unsigned NumOps = Desc.getNumOperands();
2485 bool isTwoAddr = NumOps > 1 &&
2486 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2488 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2489 unsigned i = isTwoAddr ? 1 : 0;
2490 for (unsigned e = NumOps; i != e; ++i) {
2491 const MachineOperand& MO = MI.getOperand(i);
2493 unsigned Reg = MO.getReg();
2494 if (isX86_64NonExtLowByteReg(Reg))
2499 switch (Desc.TSFlags & X86II::FormMask) {
2500 case X86II::MRMInitReg:
2501 if (isX86_64ExtendedReg(MI.getOperand(0)))
2502 REX |= (1 << 0) | (1 << 2);
2504 case X86II::MRMSrcReg: {
2505 if (isX86_64ExtendedReg(MI.getOperand(0)))
2507 i = isTwoAddr ? 2 : 1;
2508 for (unsigned e = NumOps; i != e; ++i) {
2509 const MachineOperand& MO = MI.getOperand(i);
2510 if (isX86_64ExtendedReg(MO))
2515 case X86II::MRMSrcMem: {
2516 if (isX86_64ExtendedReg(MI.getOperand(0)))
2519 i = isTwoAddr ? 2 : 1;
2520 for (; i != NumOps; ++i) {
2521 const MachineOperand& MO = MI.getOperand(i);
2523 if (isX86_64ExtendedReg(MO))
2530 case X86II::MRM0m: case X86II::MRM1m:
2531 case X86II::MRM2m: case X86II::MRM3m:
2532 case X86II::MRM4m: case X86II::MRM5m:
2533 case X86II::MRM6m: case X86II::MRM7m:
2534 case X86II::MRMDestMem: {
2535 unsigned e = isTwoAddr ? 5 : 4;
2536 i = isTwoAddr ? 1 : 0;
2537 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2540 for (; i != e; ++i) {
2541 const MachineOperand& MO = MI.getOperand(i);
2543 if (isX86_64ExtendedReg(MO))
2551 if (isX86_64ExtendedReg(MI.getOperand(0)))
2553 i = isTwoAddr ? 2 : 1;
2554 for (unsigned e = NumOps; i != e; ++i) {
2555 const MachineOperand& MO = MI.getOperand(i);
2556 if (isX86_64ExtendedReg(MO))
2566 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2567 /// relative block address instruction
2569 static unsigned sizePCRelativeBlockAddress() {
2573 /// sizeGlobalAddress - Give the size of the emission of this global address
2575 static unsigned sizeGlobalAddress(bool dword) {
2576 return dword ? 8 : 4;
2579 /// sizeConstPoolAddress - Give the size of the emission of this constant
2582 static unsigned sizeConstPoolAddress(bool dword) {
2583 return dword ? 8 : 4;
2586 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2589 static unsigned sizeExternalSymbolAddress(bool dword) {
2590 return dword ? 8 : 4;
2593 /// sizeJumpTableAddress - Give the size of the emission of this jump
2596 static unsigned sizeJumpTableAddress(bool dword) {
2597 return dword ? 8 : 4;
2600 static unsigned sizeConstant(unsigned Size) {
2604 static unsigned sizeRegModRMByte(){
2608 static unsigned sizeSIBByte(){
2612 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2613 unsigned FinalSize = 0;
2614 // If this is a simple integer displacement that doesn't require a relocation.
2616 FinalSize += sizeConstant(4);
2620 // Otherwise, this is something that requires a relocation.
2621 if (RelocOp->isGlobal()) {
2622 FinalSize += sizeGlobalAddress(false);
2623 } else if (RelocOp->isCPI()) {
2624 FinalSize += sizeConstPoolAddress(false);
2625 } else if (RelocOp->isJTI()) {
2626 FinalSize += sizeJumpTableAddress(false);
2628 assert(0 && "Unknown value to relocate!");
2633 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2634 bool IsPIC, bool Is64BitMode) {
2635 const MachineOperand &Op3 = MI.getOperand(Op+3);
2637 const MachineOperand *DispForReloc = 0;
2638 unsigned FinalSize = 0;
2640 // Figure out what sort of displacement we have to handle here.
2641 if (Op3.isGlobal()) {
2642 DispForReloc = &Op3;
2643 } else if (Op3.isCPI()) {
2644 if (Is64BitMode || IsPIC) {
2645 DispForReloc = &Op3;
2649 } else if (Op3.isJTI()) {
2650 if (Is64BitMode || IsPIC) {
2651 DispForReloc = &Op3;
2659 const MachineOperand &Base = MI.getOperand(Op);
2660 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2662 unsigned BaseReg = Base.getReg();
2664 // Is a SIB byte needed?
2665 if (IndexReg.getReg() == 0 &&
2666 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2667 if (BaseReg == 0) { // Just a displacement?
2668 // Emit special case [disp32] encoding
2670 FinalSize += getDisplacementFieldSize(DispForReloc);
2672 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2673 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2674 // Emit simple indirect register encoding... [EAX] f.e.
2676 // Be pessimistic and assume it's a disp32, not a disp8
2678 // Emit the most general non-SIB encoding: [REG+disp32]
2680 FinalSize += getDisplacementFieldSize(DispForReloc);
2684 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2685 assert(IndexReg.getReg() != X86::ESP &&
2686 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2688 bool ForceDisp32 = false;
2689 if (BaseReg == 0 || DispForReloc) {
2690 // Emit the normal disp32 encoding.
2697 FinalSize += sizeSIBByte();
2699 // Do we need to output a displacement?
2700 if (DispVal != 0 || ForceDisp32) {
2701 FinalSize += getDisplacementFieldSize(DispForReloc);
2708 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2709 const TargetInstrDesc *Desc,
2710 bool IsPIC, bool Is64BitMode) {
2712 unsigned Opcode = Desc->Opcode;
2713 unsigned FinalSize = 0;
2715 // Emit the lock opcode prefix as needed.
2716 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2718 // Emit segment overrid opcode prefix as needed.
2719 switch (Desc->TSFlags & X86II::SegOvrMask) {
2724 default: assert(0 && "Invalid segment!");
2725 case 0: break; // No segment override!
2728 // Emit the repeat opcode prefix as needed.
2729 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2731 // Emit the operand size opcode prefix as needed.
2732 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2734 // Emit the address size opcode prefix as needed.
2735 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2737 bool Need0FPrefix = false;
2738 switch (Desc->TSFlags & X86II::Op0Mask) {
2739 case X86II::TB: // Two-byte opcode prefix
2740 case X86II::T8: // 0F 38
2741 case X86II::TA: // 0F 3A
2742 Need0FPrefix = true;
2744 case X86II::REP: break; // already handled.
2745 case X86II::XS: // F3 0F
2747 Need0FPrefix = true;
2749 case X86II::XD: // F2 0F
2751 Need0FPrefix = true;
2753 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2754 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2756 break; // Two-byte opcode prefix
2757 default: assert(0 && "Invalid prefix!");
2758 case 0: break; // No prefix!
2763 unsigned REX = X86InstrInfo::determineREX(MI);
2768 // 0x0F escape code must be emitted just before the opcode.
2772 switch (Desc->TSFlags & X86II::Op0Mask) {
2773 case X86II::T8: // 0F 38
2776 case X86II::TA: // 0F 3A
2781 // If this is a two-address instruction, skip one of the register operands.
2782 unsigned NumOps = Desc->getNumOperands();
2784 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2787 switch (Desc->TSFlags & X86II::FormMask) {
2788 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
2790 // Remember the current PC offset, this is the PIC relocation
2795 case TargetInstrInfo::INLINEASM: {
2796 const MachineFunction *MF = MI.getParent()->getParent();
2797 const char *AsmStr = MI.getOperand(0).getSymbolName();
2798 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
2799 FinalSize += AI->getInlineAsmLength(AsmStr);
2802 case TargetInstrInfo::DBG_LABEL:
2803 case TargetInstrInfo::EH_LABEL:
2805 case TargetInstrInfo::IMPLICIT_DEF:
2806 case TargetInstrInfo::DECLARE:
2807 case X86::DWARF_LOC:
2808 case X86::FP_REG_KILL:
2810 case X86::MOVPC32r: {
2811 // This emits the "call" portion of this pseudo instruction.
2813 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2817 case X86::TLS_gs_ri:
2819 FinalSize += sizeGlobalAddress(false);
2827 if (CurOp != NumOps) {
2828 const MachineOperand &MO = MI.getOperand(CurOp++);
2830 FinalSize += sizePCRelativeBlockAddress();
2831 } else if (MO.isGlobal()) {
2832 FinalSize += sizeGlobalAddress(false);
2833 } else if (MO.isSymbol()) {
2834 FinalSize += sizeExternalSymbolAddress(false);
2835 } else if (MO.isImm()) {
2836 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2838 assert(0 && "Unknown RawFrm operand!");
2843 case X86II::AddRegFrm:
2847 if (CurOp != NumOps) {
2848 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2849 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2851 FinalSize += sizeConstant(Size);
2854 if (Opcode == X86::MOV64ri)
2856 if (MO1.isGlobal()) {
2857 FinalSize += sizeGlobalAddress(dword);
2858 } else if (MO1.isSymbol())
2859 FinalSize += sizeExternalSymbolAddress(dword);
2860 else if (MO1.isCPI())
2861 FinalSize += sizeConstPoolAddress(dword);
2862 else if (MO1.isJTI())
2863 FinalSize += sizeJumpTableAddress(dword);
2868 case X86II::MRMDestReg: {
2870 FinalSize += sizeRegModRMByte();
2872 if (CurOp != NumOps) {
2874 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2878 case X86II::MRMDestMem: {
2880 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2882 if (CurOp != NumOps) {
2884 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2889 case X86II::MRMSrcReg:
2891 FinalSize += sizeRegModRMByte();
2893 if (CurOp != NumOps) {
2895 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2899 case X86II::MRMSrcMem: {
2902 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
2904 if (CurOp != NumOps) {
2906 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2911 case X86II::MRM0r: case X86II::MRM1r:
2912 case X86II::MRM2r: case X86II::MRM3r:
2913 case X86II::MRM4r: case X86II::MRM5r:
2914 case X86II::MRM6r: case X86II::MRM7r:
2917 FinalSize += sizeRegModRMByte();
2919 if (CurOp != NumOps) {
2920 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2921 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2923 FinalSize += sizeConstant(Size);
2926 if (Opcode == X86::MOV64ri32)
2928 if (MO1.isGlobal()) {
2929 FinalSize += sizeGlobalAddress(dword);
2930 } else if (MO1.isSymbol())
2931 FinalSize += sizeExternalSymbolAddress(dword);
2932 else if (MO1.isCPI())
2933 FinalSize += sizeConstPoolAddress(dword);
2934 else if (MO1.isJTI())
2935 FinalSize += sizeJumpTableAddress(dword);
2940 case X86II::MRM0m: case X86II::MRM1m:
2941 case X86II::MRM2m: case X86II::MRM3m:
2942 case X86II::MRM4m: case X86II::MRM5m:
2943 case X86II::MRM6m: case X86II::MRM7m: {
2946 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2949 if (CurOp != NumOps) {
2950 const MachineOperand &MO = MI.getOperand(CurOp++);
2951 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2953 FinalSize += sizeConstant(Size);
2956 if (Opcode == X86::MOV64mi32)
2958 if (MO.isGlobal()) {
2959 FinalSize += sizeGlobalAddress(dword);
2960 } else if (MO.isSymbol())
2961 FinalSize += sizeExternalSymbolAddress(dword);
2962 else if (MO.isCPI())
2963 FinalSize += sizeConstPoolAddress(dword);
2964 else if (MO.isJTI())
2965 FinalSize += sizeJumpTableAddress(dword);
2971 case X86II::MRMInitReg:
2973 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
2974 FinalSize += sizeRegModRMByte();
2979 if (!Desc->isVariadic() && CurOp != NumOps) {
2980 cerr << "Cannot determine size: ";
2991 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
2992 const TargetInstrDesc &Desc = MI->getDesc();
2993 bool IsPIC = (TM.getRelocationModel() == Reloc::PIC_);
2994 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
2995 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
2996 if (Desc.getOpcode() == X86::MOVPC32r) {
2997 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);
3002 /// getGlobalBaseReg - Return a virtual register initialized with the
3003 /// the global base register value. Output instructions required to
3004 /// initialize the register in the function entry block, if necessary.
3006 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3007 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3008 "X86-64 PIC uses RIP relative addressing");
3010 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3011 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3012 if (GlobalBaseReg != 0)
3013 return GlobalBaseReg;
3015 // Insert the set of GlobalBaseReg into the first MBB of the function
3016 MachineBasicBlock &FirstMBB = MF->front();
3017 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3018 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3019 unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3021 const TargetInstrInfo *TII = TM.getInstrInfo();
3022 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3023 // only used in JIT code emission as displacement to pc.
3024 BuildMI(FirstMBB, MBBI, TII->get(X86::MOVPC32r), PC).addImm(0);
3026 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3027 // not to pc, but to _GLOBAL_ADDRESS_TABLE_ external
3028 if (TM.getRelocationModel() == Reloc::PIC_ &&
3029 TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3031 RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3032 BuildMI(FirstMBB, MBBI, TII->get(X86::ADD32ri), GlobalBaseReg)
3033 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_");
3038 X86FI->setGlobalBaseReg(GlobalBaseReg);
3039 return GlobalBaseReg;