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/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/LiveVariables.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Target/TargetOptions.h"
29 #include "llvm/Target/TargetAsmInfo.h"
35 NoFusing("disable-spill-fusing",
36 cl::desc("Disable fusing of spill code into instructions"));
38 PrintFailedFusing("print-failed-fuse-candidates",
39 cl::desc("Print instructions that the allocator wants to"
40 " fuse, but the X86 backend currently can't"),
43 ReMatPICStubLoad("remat-pic-stub-load",
44 cl::desc("Re-materialize load from stub in PIC mode"),
45 cl::init(false), cl::Hidden);
48 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
49 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
50 TM(tm), RI(tm, *this) {
51 SmallVector<unsigned,16> AmbEntries;
52 static const unsigned OpTbl2Addr[][2] = {
53 { X86::ADC32ri, X86::ADC32mi },
54 { X86::ADC32ri8, X86::ADC32mi8 },
55 { X86::ADC32rr, X86::ADC32mr },
56 { X86::ADC64ri32, X86::ADC64mi32 },
57 { X86::ADC64ri8, X86::ADC64mi8 },
58 { X86::ADC64rr, X86::ADC64mr },
59 { X86::ADD16ri, X86::ADD16mi },
60 { X86::ADD16ri8, X86::ADD16mi8 },
61 { X86::ADD16rr, X86::ADD16mr },
62 { X86::ADD32ri, X86::ADD32mi },
63 { X86::ADD32ri8, X86::ADD32mi8 },
64 { X86::ADD32rr, X86::ADD32mr },
65 { X86::ADD64ri32, X86::ADD64mi32 },
66 { X86::ADD64ri8, X86::ADD64mi8 },
67 { X86::ADD64rr, X86::ADD64mr },
68 { X86::ADD8ri, X86::ADD8mi },
69 { X86::ADD8rr, X86::ADD8mr },
70 { X86::AND16ri, X86::AND16mi },
71 { X86::AND16ri8, X86::AND16mi8 },
72 { X86::AND16rr, X86::AND16mr },
73 { X86::AND32ri, X86::AND32mi },
74 { X86::AND32ri8, X86::AND32mi8 },
75 { X86::AND32rr, X86::AND32mr },
76 { X86::AND64ri32, X86::AND64mi32 },
77 { X86::AND64ri8, X86::AND64mi8 },
78 { X86::AND64rr, X86::AND64mr },
79 { X86::AND8ri, X86::AND8mi },
80 { X86::AND8rr, X86::AND8mr },
81 { X86::DEC16r, X86::DEC16m },
82 { X86::DEC32r, X86::DEC32m },
83 { X86::DEC64_16r, X86::DEC64_16m },
84 { X86::DEC64_32r, X86::DEC64_32m },
85 { X86::DEC64r, X86::DEC64m },
86 { X86::DEC8r, X86::DEC8m },
87 { X86::INC16r, X86::INC16m },
88 { X86::INC32r, X86::INC32m },
89 { X86::INC64_16r, X86::INC64_16m },
90 { X86::INC64_32r, X86::INC64_32m },
91 { X86::INC64r, X86::INC64m },
92 { X86::INC8r, X86::INC8m },
93 { X86::NEG16r, X86::NEG16m },
94 { X86::NEG32r, X86::NEG32m },
95 { X86::NEG64r, X86::NEG64m },
96 { X86::NEG8r, X86::NEG8m },
97 { X86::NOT16r, X86::NOT16m },
98 { X86::NOT32r, X86::NOT32m },
99 { X86::NOT64r, X86::NOT64m },
100 { X86::NOT8r, X86::NOT8m },
101 { X86::OR16ri, X86::OR16mi },
102 { X86::OR16ri8, X86::OR16mi8 },
103 { X86::OR16rr, X86::OR16mr },
104 { X86::OR32ri, X86::OR32mi },
105 { X86::OR32ri8, X86::OR32mi8 },
106 { X86::OR32rr, X86::OR32mr },
107 { X86::OR64ri32, X86::OR64mi32 },
108 { X86::OR64ri8, X86::OR64mi8 },
109 { X86::OR64rr, X86::OR64mr },
110 { X86::OR8ri, X86::OR8mi },
111 { X86::OR8rr, X86::OR8mr },
112 { X86::ROL16r1, X86::ROL16m1 },
113 { X86::ROL16rCL, X86::ROL16mCL },
114 { X86::ROL16ri, X86::ROL16mi },
115 { X86::ROL32r1, X86::ROL32m1 },
116 { X86::ROL32rCL, X86::ROL32mCL },
117 { X86::ROL32ri, X86::ROL32mi },
118 { X86::ROL64r1, X86::ROL64m1 },
119 { X86::ROL64rCL, X86::ROL64mCL },
120 { X86::ROL64ri, X86::ROL64mi },
121 { X86::ROL8r1, X86::ROL8m1 },
122 { X86::ROL8rCL, X86::ROL8mCL },
123 { X86::ROL8ri, X86::ROL8mi },
124 { X86::ROR16r1, X86::ROR16m1 },
125 { X86::ROR16rCL, X86::ROR16mCL },
126 { X86::ROR16ri, X86::ROR16mi },
127 { X86::ROR32r1, X86::ROR32m1 },
128 { X86::ROR32rCL, X86::ROR32mCL },
129 { X86::ROR32ri, X86::ROR32mi },
130 { X86::ROR64r1, X86::ROR64m1 },
131 { X86::ROR64rCL, X86::ROR64mCL },
132 { X86::ROR64ri, X86::ROR64mi },
133 { X86::ROR8r1, X86::ROR8m1 },
134 { X86::ROR8rCL, X86::ROR8mCL },
135 { X86::ROR8ri, X86::ROR8mi },
136 { X86::SAR16r1, X86::SAR16m1 },
137 { X86::SAR16rCL, X86::SAR16mCL },
138 { X86::SAR16ri, X86::SAR16mi },
139 { X86::SAR32r1, X86::SAR32m1 },
140 { X86::SAR32rCL, X86::SAR32mCL },
141 { X86::SAR32ri, X86::SAR32mi },
142 { X86::SAR64r1, X86::SAR64m1 },
143 { X86::SAR64rCL, X86::SAR64mCL },
144 { X86::SAR64ri, X86::SAR64mi },
145 { X86::SAR8r1, X86::SAR8m1 },
146 { X86::SAR8rCL, X86::SAR8mCL },
147 { X86::SAR8ri, X86::SAR8mi },
148 { X86::SBB32ri, X86::SBB32mi },
149 { X86::SBB32ri8, X86::SBB32mi8 },
150 { X86::SBB32rr, X86::SBB32mr },
151 { X86::SBB64ri32, X86::SBB64mi32 },
152 { X86::SBB64ri8, X86::SBB64mi8 },
153 { X86::SBB64rr, X86::SBB64mr },
154 { X86::SHL16rCL, X86::SHL16mCL },
155 { X86::SHL16ri, X86::SHL16mi },
156 { X86::SHL32rCL, X86::SHL32mCL },
157 { X86::SHL32ri, X86::SHL32mi },
158 { X86::SHL64rCL, X86::SHL64mCL },
159 { X86::SHL64ri, X86::SHL64mi },
160 { X86::SHL8rCL, X86::SHL8mCL },
161 { X86::SHL8ri, X86::SHL8mi },
162 { X86::SHLD16rrCL, X86::SHLD16mrCL },
163 { X86::SHLD16rri8, X86::SHLD16mri8 },
164 { X86::SHLD32rrCL, X86::SHLD32mrCL },
165 { X86::SHLD32rri8, X86::SHLD32mri8 },
166 { X86::SHLD64rrCL, X86::SHLD64mrCL },
167 { X86::SHLD64rri8, X86::SHLD64mri8 },
168 { X86::SHR16r1, X86::SHR16m1 },
169 { X86::SHR16rCL, X86::SHR16mCL },
170 { X86::SHR16ri, X86::SHR16mi },
171 { X86::SHR32r1, X86::SHR32m1 },
172 { X86::SHR32rCL, X86::SHR32mCL },
173 { X86::SHR32ri, X86::SHR32mi },
174 { X86::SHR64r1, X86::SHR64m1 },
175 { X86::SHR64rCL, X86::SHR64mCL },
176 { X86::SHR64ri, X86::SHR64mi },
177 { X86::SHR8r1, X86::SHR8m1 },
178 { X86::SHR8rCL, X86::SHR8mCL },
179 { X86::SHR8ri, X86::SHR8mi },
180 { X86::SHRD16rrCL, X86::SHRD16mrCL },
181 { X86::SHRD16rri8, X86::SHRD16mri8 },
182 { X86::SHRD32rrCL, X86::SHRD32mrCL },
183 { X86::SHRD32rri8, X86::SHRD32mri8 },
184 { X86::SHRD64rrCL, X86::SHRD64mrCL },
185 { X86::SHRD64rri8, X86::SHRD64mri8 },
186 { X86::SUB16ri, X86::SUB16mi },
187 { X86::SUB16ri8, X86::SUB16mi8 },
188 { X86::SUB16rr, X86::SUB16mr },
189 { X86::SUB32ri, X86::SUB32mi },
190 { X86::SUB32ri8, X86::SUB32mi8 },
191 { X86::SUB32rr, X86::SUB32mr },
192 { X86::SUB64ri32, X86::SUB64mi32 },
193 { X86::SUB64ri8, X86::SUB64mi8 },
194 { X86::SUB64rr, X86::SUB64mr },
195 { X86::SUB8ri, X86::SUB8mi },
196 { X86::SUB8rr, X86::SUB8mr },
197 { X86::XOR16ri, X86::XOR16mi },
198 { X86::XOR16ri8, X86::XOR16mi8 },
199 { X86::XOR16rr, X86::XOR16mr },
200 { X86::XOR32ri, X86::XOR32mi },
201 { X86::XOR32ri8, X86::XOR32mi8 },
202 { X86::XOR32rr, X86::XOR32mr },
203 { X86::XOR64ri32, X86::XOR64mi32 },
204 { X86::XOR64ri8, X86::XOR64mi8 },
205 { X86::XOR64rr, X86::XOR64mr },
206 { X86::XOR8ri, X86::XOR8mi },
207 { X86::XOR8rr, X86::XOR8mr }
210 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
211 unsigned RegOp = OpTbl2Addr[i][0];
212 unsigned MemOp = OpTbl2Addr[i][1];
213 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp,
215 assert(false && "Duplicated entries?");
216 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
217 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
218 std::make_pair(RegOp,
220 AmbEntries.push_back(MemOp);
223 // If the third value is 1, then it's folding either a load or a store.
224 static const unsigned OpTbl0[][3] = {
225 { X86::CALL32r, X86::CALL32m, 1 },
226 { X86::CALL64r, X86::CALL64m, 1 },
227 { X86::CMP16ri, X86::CMP16mi, 1 },
228 { X86::CMP16ri8, X86::CMP16mi8, 1 },
229 { X86::CMP16rr, X86::CMP16mr, 1 },
230 { X86::CMP32ri, X86::CMP32mi, 1 },
231 { X86::CMP32ri8, X86::CMP32mi8, 1 },
232 { X86::CMP32rr, X86::CMP32mr, 1 },
233 { X86::CMP64ri32, X86::CMP64mi32, 1 },
234 { X86::CMP64ri8, X86::CMP64mi8, 1 },
235 { X86::CMP64rr, X86::CMP64mr, 1 },
236 { X86::CMP8ri, X86::CMP8mi, 1 },
237 { X86::CMP8rr, X86::CMP8mr, 1 },
238 { X86::DIV16r, X86::DIV16m, 1 },
239 { X86::DIV32r, X86::DIV32m, 1 },
240 { X86::DIV64r, X86::DIV64m, 1 },
241 { X86::DIV8r, X86::DIV8m, 1 },
242 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0 },
243 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
244 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
245 { X86::IDIV16r, X86::IDIV16m, 1 },
246 { X86::IDIV32r, X86::IDIV32m, 1 },
247 { X86::IDIV64r, X86::IDIV64m, 1 },
248 { X86::IDIV8r, X86::IDIV8m, 1 },
249 { X86::IMUL16r, X86::IMUL16m, 1 },
250 { X86::IMUL32r, X86::IMUL32m, 1 },
251 { X86::IMUL64r, X86::IMUL64m, 1 },
252 { X86::IMUL8r, X86::IMUL8m, 1 },
253 { X86::JMP32r, X86::JMP32m, 1 },
254 { X86::JMP64r, X86::JMP64m, 1 },
255 { X86::MOV16ri, X86::MOV16mi, 0 },
256 { X86::MOV16rr, X86::MOV16mr, 0 },
257 { X86::MOV16to16_, X86::MOV16_mr, 0 },
258 { X86::MOV32ri, X86::MOV32mi, 0 },
259 { X86::MOV32rr, X86::MOV32mr, 0 },
260 { X86::MOV32to32_, X86::MOV32_mr, 0 },
261 { X86::MOV64ri32, X86::MOV64mi32, 0 },
262 { X86::MOV64rr, X86::MOV64mr, 0 },
263 { X86::MOV8ri, X86::MOV8mi, 0 },
264 { X86::MOV8rr, X86::MOV8mr, 0 },
265 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
266 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
267 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
268 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
269 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
270 { X86::MOVSDrr, X86::MOVSDmr, 0 },
271 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
272 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
273 { X86::MOVSSrr, X86::MOVSSmr, 0 },
274 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
275 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
276 { X86::MUL16r, X86::MUL16m, 1 },
277 { X86::MUL32r, X86::MUL32m, 1 },
278 { X86::MUL64r, X86::MUL64m, 1 },
279 { X86::MUL8r, X86::MUL8m, 1 },
280 { X86::SETAEr, X86::SETAEm, 0 },
281 { X86::SETAr, X86::SETAm, 0 },
282 { X86::SETBEr, X86::SETBEm, 0 },
283 { X86::SETBr, X86::SETBm, 0 },
284 { X86::SETCr, X86::SETCm, 0 },
285 { X86::SETEr, X86::SETEm, 0 },
286 { X86::SETGEr, X86::SETGEm, 0 },
287 { X86::SETGr, X86::SETGm, 0 },
288 { X86::SETLEr, X86::SETLEm, 0 },
289 { X86::SETLr, X86::SETLm, 0 },
290 { X86::SETNCr, X86::SETNCm, 0 },
291 { X86::SETNEr, X86::SETNEm, 0 },
292 { X86::SETNOr, X86::SETNOm, 0 },
293 { X86::SETNPr, X86::SETNPm, 0 },
294 { X86::SETNSr, X86::SETNSm, 0 },
295 { X86::SETOr, X86::SETOm, 0 },
296 { X86::SETPr, X86::SETPm, 0 },
297 { X86::SETSr, X86::SETSm, 0 },
298 { X86::TAILJMPr, X86::TAILJMPm, 1 },
299 { X86::TEST16ri, X86::TEST16mi, 1 },
300 { X86::TEST32ri, X86::TEST32mi, 1 },
301 { X86::TEST64ri32, X86::TEST64mi32, 1 },
302 { X86::TEST8ri, X86::TEST8mi, 1 }
305 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
306 unsigned RegOp = OpTbl0[i][0];
307 unsigned MemOp = OpTbl0[i][1];
308 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp,
310 assert(false && "Duplicated entries?");
311 unsigned FoldedLoad = OpTbl0[i][2];
312 // Index 0, folded load or store.
313 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
314 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
315 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
316 std::make_pair(RegOp, AuxInfo))).second)
317 AmbEntries.push_back(MemOp);
320 static const unsigned OpTbl1[][2] = {
321 { X86::CMP16rr, X86::CMP16rm },
322 { X86::CMP32rr, X86::CMP32rm },
323 { X86::CMP64rr, X86::CMP64rm },
324 { X86::CMP8rr, X86::CMP8rm },
325 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
326 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
327 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
328 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
329 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
330 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
331 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
332 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
333 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
334 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
335 { X86::FsMOVAPDrr, X86::MOVSDrm },
336 { X86::FsMOVAPSrr, X86::MOVSSrm },
337 { X86::IMUL16rri, X86::IMUL16rmi },
338 { X86::IMUL16rri8, X86::IMUL16rmi8 },
339 { X86::IMUL32rri, X86::IMUL32rmi },
340 { X86::IMUL32rri8, X86::IMUL32rmi8 },
341 { X86::IMUL64rri32, X86::IMUL64rmi32 },
342 { X86::IMUL64rri8, X86::IMUL64rmi8 },
343 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
344 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
345 { X86::Int_COMISDrr, X86::Int_COMISDrm },
346 { X86::Int_COMISSrr, X86::Int_COMISSrm },
347 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
348 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
349 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
350 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
351 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
352 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
353 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
354 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
355 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
356 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
357 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
358 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
359 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
360 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
361 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
362 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
363 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
364 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
365 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
366 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
367 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
368 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
369 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
370 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
371 { X86::MOV16rr, X86::MOV16rm },
372 { X86::MOV16to16_, X86::MOV16_rm },
373 { X86::MOV32rr, X86::MOV32rm },
374 { X86::MOV32to32_, X86::MOV32_rm },
375 { X86::MOV64rr, X86::MOV64rm },
376 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
377 { X86::MOV64toSDrr, X86::MOV64toSDrm },
378 { X86::MOV8rr, X86::MOV8rm },
379 { X86::MOVAPDrr, X86::MOVAPDrm },
380 { X86::MOVAPSrr, X86::MOVAPSrm },
381 { X86::MOVDDUPrr, X86::MOVDDUPrm },
382 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
383 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
384 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
385 { X86::MOVSDrr, X86::MOVSDrm },
386 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
387 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
388 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
389 { X86::MOVSSrr, X86::MOVSSrm },
390 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
391 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
392 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
393 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
394 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
395 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
396 { X86::MOVUPDrr, X86::MOVUPDrm },
397 { X86::MOVUPSrr, X86::MOVUPSrm },
398 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
399 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
400 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
401 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
402 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
403 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
404 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
405 { X86::MOVZX64rr32, X86::MOVZX64rm32 },
406 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
407 { X86::PSHUFDri, X86::PSHUFDmi },
408 { X86::PSHUFHWri, X86::PSHUFHWmi },
409 { X86::PSHUFLWri, X86::PSHUFLWmi },
410 { X86::RCPPSr, X86::RCPPSm },
411 { X86::RCPPSr_Int, X86::RCPPSm_Int },
412 { X86::RSQRTPSr, X86::RSQRTPSm },
413 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
414 { X86::RSQRTSSr, X86::RSQRTSSm },
415 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
416 { X86::SQRTPDr, X86::SQRTPDm },
417 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
418 { X86::SQRTPSr, X86::SQRTPSm },
419 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
420 { X86::SQRTSDr, X86::SQRTSDm },
421 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
422 { X86::SQRTSSr, X86::SQRTSSm },
423 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
424 { X86::TEST16rr, X86::TEST16rm },
425 { X86::TEST32rr, X86::TEST32rm },
426 { X86::TEST64rr, X86::TEST64rm },
427 { X86::TEST8rr, X86::TEST8rm },
428 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
429 { X86::UCOMISDrr, X86::UCOMISDrm },
430 { X86::UCOMISSrr, X86::UCOMISSrm }
433 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
434 unsigned RegOp = OpTbl1[i][0];
435 unsigned MemOp = OpTbl1[i][1];
436 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp,
438 assert(false && "Duplicated entries?");
439 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
440 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
441 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
442 std::make_pair(RegOp, AuxInfo))).second)
443 AmbEntries.push_back(MemOp);
446 static const unsigned OpTbl2[][2] = {
447 { X86::ADC32rr, X86::ADC32rm },
448 { X86::ADC64rr, X86::ADC64rm },
449 { X86::ADD16rr, X86::ADD16rm },
450 { X86::ADD32rr, X86::ADD32rm },
451 { X86::ADD64rr, X86::ADD64rm },
452 { X86::ADD8rr, X86::ADD8rm },
453 { X86::ADDPDrr, X86::ADDPDrm },
454 { X86::ADDPSrr, X86::ADDPSrm },
455 { X86::ADDSDrr, X86::ADDSDrm },
456 { X86::ADDSSrr, X86::ADDSSrm },
457 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
458 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
459 { X86::AND16rr, X86::AND16rm },
460 { X86::AND32rr, X86::AND32rm },
461 { X86::AND64rr, X86::AND64rm },
462 { X86::AND8rr, X86::AND8rm },
463 { X86::ANDNPDrr, X86::ANDNPDrm },
464 { X86::ANDNPSrr, X86::ANDNPSrm },
465 { X86::ANDPDrr, X86::ANDPDrm },
466 { X86::ANDPSrr, X86::ANDPSrm },
467 { X86::CMOVA16rr, X86::CMOVA16rm },
468 { X86::CMOVA32rr, X86::CMOVA32rm },
469 { X86::CMOVA64rr, X86::CMOVA64rm },
470 { X86::CMOVAE16rr, X86::CMOVAE16rm },
471 { X86::CMOVAE32rr, X86::CMOVAE32rm },
472 { X86::CMOVAE64rr, X86::CMOVAE64rm },
473 { X86::CMOVB16rr, X86::CMOVB16rm },
474 { X86::CMOVB32rr, X86::CMOVB32rm },
475 { X86::CMOVB64rr, X86::CMOVB64rm },
476 { X86::CMOVBE16rr, X86::CMOVBE16rm },
477 { X86::CMOVBE32rr, X86::CMOVBE32rm },
478 { X86::CMOVBE64rr, X86::CMOVBE64rm },
479 { X86::CMOVE16rr, X86::CMOVE16rm },
480 { X86::CMOVE32rr, X86::CMOVE32rm },
481 { X86::CMOVE64rr, X86::CMOVE64rm },
482 { X86::CMOVG16rr, X86::CMOVG16rm },
483 { X86::CMOVG32rr, X86::CMOVG32rm },
484 { X86::CMOVG64rr, X86::CMOVG64rm },
485 { X86::CMOVGE16rr, X86::CMOVGE16rm },
486 { X86::CMOVGE32rr, X86::CMOVGE32rm },
487 { X86::CMOVGE64rr, X86::CMOVGE64rm },
488 { X86::CMOVL16rr, X86::CMOVL16rm },
489 { X86::CMOVL32rr, X86::CMOVL32rm },
490 { X86::CMOVL64rr, X86::CMOVL64rm },
491 { X86::CMOVLE16rr, X86::CMOVLE16rm },
492 { X86::CMOVLE32rr, X86::CMOVLE32rm },
493 { X86::CMOVLE64rr, X86::CMOVLE64rm },
494 { X86::CMOVNE16rr, X86::CMOVNE16rm },
495 { X86::CMOVNE32rr, X86::CMOVNE32rm },
496 { X86::CMOVNE64rr, X86::CMOVNE64rm },
497 { X86::CMOVNP16rr, X86::CMOVNP16rm },
498 { X86::CMOVNP32rr, X86::CMOVNP32rm },
499 { X86::CMOVNP64rr, X86::CMOVNP64rm },
500 { X86::CMOVNS16rr, X86::CMOVNS16rm },
501 { X86::CMOVNS32rr, X86::CMOVNS32rm },
502 { X86::CMOVNS64rr, X86::CMOVNS64rm },
503 { X86::CMOVP16rr, X86::CMOVP16rm },
504 { X86::CMOVP32rr, X86::CMOVP32rm },
505 { X86::CMOVP64rr, X86::CMOVP64rm },
506 { X86::CMOVS16rr, X86::CMOVS16rm },
507 { X86::CMOVS32rr, X86::CMOVS32rm },
508 { X86::CMOVS64rr, X86::CMOVS64rm },
509 { X86::CMPPDrri, X86::CMPPDrmi },
510 { X86::CMPPSrri, X86::CMPPSrmi },
511 { X86::CMPSDrr, X86::CMPSDrm },
512 { X86::CMPSSrr, X86::CMPSSrm },
513 { X86::DIVPDrr, X86::DIVPDrm },
514 { X86::DIVPSrr, X86::DIVPSrm },
515 { X86::DIVSDrr, X86::DIVSDrm },
516 { X86::DIVSSrr, X86::DIVSSrm },
517 { X86::FsANDNPDrr, X86::FsANDNPDrm },
518 { X86::FsANDNPSrr, X86::FsANDNPSrm },
519 { X86::FsANDPDrr, X86::FsANDPDrm },
520 { X86::FsANDPSrr, X86::FsANDPSrm },
521 { X86::FsORPDrr, X86::FsORPDrm },
522 { X86::FsORPSrr, X86::FsORPSrm },
523 { X86::FsXORPDrr, X86::FsXORPDrm },
524 { X86::FsXORPSrr, X86::FsXORPSrm },
525 { X86::HADDPDrr, X86::HADDPDrm },
526 { X86::HADDPSrr, X86::HADDPSrm },
527 { X86::HSUBPDrr, X86::HSUBPDrm },
528 { X86::HSUBPSrr, X86::HSUBPSrm },
529 { X86::IMUL16rr, X86::IMUL16rm },
530 { X86::IMUL32rr, X86::IMUL32rm },
531 { X86::IMUL64rr, X86::IMUL64rm },
532 { X86::MAXPDrr, X86::MAXPDrm },
533 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
534 { X86::MAXPSrr, X86::MAXPSrm },
535 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
536 { X86::MAXSDrr, X86::MAXSDrm },
537 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
538 { X86::MAXSSrr, X86::MAXSSrm },
539 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
540 { X86::MINPDrr, X86::MINPDrm },
541 { X86::MINPDrr_Int, X86::MINPDrm_Int },
542 { X86::MINPSrr, X86::MINPSrm },
543 { X86::MINPSrr_Int, X86::MINPSrm_Int },
544 { X86::MINSDrr, X86::MINSDrm },
545 { X86::MINSDrr_Int, X86::MINSDrm_Int },
546 { X86::MINSSrr, X86::MINSSrm },
547 { X86::MINSSrr_Int, X86::MINSSrm_Int },
548 { X86::MULPDrr, X86::MULPDrm },
549 { X86::MULPSrr, X86::MULPSrm },
550 { X86::MULSDrr, X86::MULSDrm },
551 { X86::MULSSrr, X86::MULSSrm },
552 { X86::OR16rr, X86::OR16rm },
553 { X86::OR32rr, X86::OR32rm },
554 { X86::OR64rr, X86::OR64rm },
555 { X86::OR8rr, X86::OR8rm },
556 { X86::ORPDrr, X86::ORPDrm },
557 { X86::ORPSrr, X86::ORPSrm },
558 { X86::PACKSSDWrr, X86::PACKSSDWrm },
559 { X86::PACKSSWBrr, X86::PACKSSWBrm },
560 { X86::PACKUSWBrr, X86::PACKUSWBrm },
561 { X86::PADDBrr, X86::PADDBrm },
562 { X86::PADDDrr, X86::PADDDrm },
563 { X86::PADDQrr, X86::PADDQrm },
564 { X86::PADDSBrr, X86::PADDSBrm },
565 { X86::PADDSWrr, X86::PADDSWrm },
566 { X86::PADDWrr, X86::PADDWrm },
567 { X86::PANDNrr, X86::PANDNrm },
568 { X86::PANDrr, X86::PANDrm },
569 { X86::PAVGBrr, X86::PAVGBrm },
570 { X86::PAVGWrr, X86::PAVGWrm },
571 { X86::PCMPEQBrr, X86::PCMPEQBrm },
572 { X86::PCMPEQDrr, X86::PCMPEQDrm },
573 { X86::PCMPEQWrr, X86::PCMPEQWrm },
574 { X86::PCMPGTBrr, X86::PCMPGTBrm },
575 { X86::PCMPGTDrr, X86::PCMPGTDrm },
576 { X86::PCMPGTWrr, X86::PCMPGTWrm },
577 { X86::PINSRWrri, X86::PINSRWrmi },
578 { X86::PMADDWDrr, X86::PMADDWDrm },
579 { X86::PMAXSWrr, X86::PMAXSWrm },
580 { X86::PMAXUBrr, X86::PMAXUBrm },
581 { X86::PMINSWrr, X86::PMINSWrm },
582 { X86::PMINUBrr, X86::PMINUBrm },
583 { X86::PMULDQrr, X86::PMULDQrm },
584 { X86::PMULHUWrr, X86::PMULHUWrm },
585 { X86::PMULHWrr, X86::PMULHWrm },
586 { X86::PMULLDrr, X86::PMULLDrm },
587 { X86::PMULLDrr_int, X86::PMULLDrm_int },
588 { X86::PMULLWrr, X86::PMULLWrm },
589 { X86::PMULUDQrr, X86::PMULUDQrm },
590 { X86::PORrr, X86::PORrm },
591 { X86::PSADBWrr, X86::PSADBWrm },
592 { X86::PSLLDrr, X86::PSLLDrm },
593 { X86::PSLLQrr, X86::PSLLQrm },
594 { X86::PSLLWrr, X86::PSLLWrm },
595 { X86::PSRADrr, X86::PSRADrm },
596 { X86::PSRAWrr, X86::PSRAWrm },
597 { X86::PSRLDrr, X86::PSRLDrm },
598 { X86::PSRLQrr, X86::PSRLQrm },
599 { X86::PSRLWrr, X86::PSRLWrm },
600 { X86::PSUBBrr, X86::PSUBBrm },
601 { X86::PSUBDrr, X86::PSUBDrm },
602 { X86::PSUBSBrr, X86::PSUBSBrm },
603 { X86::PSUBSWrr, X86::PSUBSWrm },
604 { X86::PSUBWrr, X86::PSUBWrm },
605 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
606 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
607 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
608 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
609 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
610 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
611 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
612 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
613 { X86::PXORrr, X86::PXORrm },
614 { X86::SBB32rr, X86::SBB32rm },
615 { X86::SBB64rr, X86::SBB64rm },
616 { X86::SHUFPDrri, X86::SHUFPDrmi },
617 { X86::SHUFPSrri, X86::SHUFPSrmi },
618 { X86::SUB16rr, X86::SUB16rm },
619 { X86::SUB32rr, X86::SUB32rm },
620 { X86::SUB64rr, X86::SUB64rm },
621 { X86::SUB8rr, X86::SUB8rm },
622 { X86::SUBPDrr, X86::SUBPDrm },
623 { X86::SUBPSrr, X86::SUBPSrm },
624 { X86::SUBSDrr, X86::SUBSDrm },
625 { X86::SUBSSrr, X86::SUBSSrm },
626 // FIXME: TEST*rr -> swapped operand of TEST*mr.
627 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
628 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
629 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
630 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
631 { X86::XOR16rr, X86::XOR16rm },
632 { X86::XOR32rr, X86::XOR32rm },
633 { X86::XOR64rr, X86::XOR64rm },
634 { X86::XOR8rr, X86::XOR8rm },
635 { X86::XORPDrr, X86::XORPDrm },
636 { X86::XORPSrr, X86::XORPSrm }
639 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
640 unsigned RegOp = OpTbl2[i][0];
641 unsigned MemOp = OpTbl2[i][1];
642 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp,
644 assert(false && "Duplicated entries?");
645 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
646 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
647 std::make_pair(RegOp, AuxInfo))).second)
648 AmbEntries.push_back(MemOp);
651 // Remove ambiguous entries.
652 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
655 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
657 unsigned& destReg) const {
658 switch (MI.getOpcode()) {
665 case X86::MOV16to16_:
666 case X86::MOV32to32_:
670 // FP Stack register class copies
671 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
672 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
673 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
675 case X86::FsMOVAPSrr:
676 case X86::FsMOVAPDrr:
679 case X86::MOVSS2PSrr:
680 case X86::MOVSD2PDrr:
681 case X86::MOVPS2SSrr:
682 case X86::MOVPD2SDrr:
683 case X86::MMX_MOVD64rr:
684 case X86::MMX_MOVQ64rr:
685 assert(MI.getNumOperands() >= 2 &&
686 MI.getOperand(0).isReg() &&
687 MI.getOperand(1).isReg() &&
688 "invalid register-register move instruction");
689 sourceReg = MI.getOperand(1).getReg();
690 destReg = MI.getOperand(0).getReg();
695 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
696 int &FrameIndex) const {
697 switch (MI->getOpcode()) {
710 case X86::MMX_MOVD64rm:
711 case X86::MMX_MOVQ64rm:
712 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
713 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
714 MI->getOperand(2).getImm() == 1 &&
715 MI->getOperand(3).getReg() == 0 &&
716 MI->getOperand(4).getImm() == 0) {
717 FrameIndex = MI->getOperand(1).getIndex();
718 return MI->getOperand(0).getReg();
725 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
726 int &FrameIndex) const {
727 switch (MI->getOpcode()) {
740 case X86::MMX_MOVD64mr:
741 case X86::MMX_MOVQ64mr:
742 case X86::MMX_MOVNTQmr:
743 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
744 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
745 MI->getOperand(1).getImm() == 1 &&
746 MI->getOperand(2).getReg() == 0 &&
747 MI->getOperand(3).getImm() == 0) {
748 FrameIndex = MI->getOperand(0).getIndex();
749 return MI->getOperand(4).getReg();
757 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
759 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
760 bool isPICBase = false;
761 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
762 E = MRI.def_end(); I != E; ++I) {
763 MachineInstr *DefMI = I.getOperand().getParent();
764 if (DefMI->getOpcode() != X86::MOVPC32r)
766 assert(!isPICBase && "More than one PIC base?");
772 /// isGVStub - Return true if the GV requires an extra load to get the
774 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
775 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
779 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI) const {
780 switch (MI->getOpcode()) {
793 case X86::MMX_MOVD64rm:
794 case X86::MMX_MOVQ64rm: {
795 // Loads from constant pools are trivially rematerializable.
796 if (MI->getOperand(1).isReg() &&
797 MI->getOperand(2).isImm() &&
798 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
799 (MI->getOperand(4).isCPI() ||
800 (MI->getOperand(4).isGlobal() &&
801 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
802 unsigned BaseReg = MI->getOperand(1).getReg();
805 // Allow re-materialization of PIC load.
806 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
808 const MachineFunction &MF = *MI->getParent()->getParent();
809 const MachineRegisterInfo &MRI = MF.getRegInfo();
810 bool isPICBase = false;
811 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
812 E = MRI.def_end(); I != E; ++I) {
813 MachineInstr *DefMI = I.getOperand().getParent();
814 if (DefMI->getOpcode() != X86::MOVPC32r)
816 assert(!isPICBase && "More than one PIC base?");
826 if (MI->getOperand(2).isImm() &&
827 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
828 !MI->getOperand(4).isReg()) {
829 // lea fi#, lea GV, etc. are all rematerializable.
830 if (!MI->getOperand(1).isReg())
832 unsigned BaseReg = MI->getOperand(1).getReg();
835 // Allow re-materialization of lea PICBase + x.
836 const MachineFunction &MF = *MI->getParent()->getParent();
837 const MachineRegisterInfo &MRI = MF.getRegInfo();
838 return regIsPICBase(BaseReg, MRI);
844 // All other instructions marked M_REMATERIALIZABLE are always trivially
849 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
850 /// would clobber the EFLAGS condition register. Note the result may be
851 /// conservative. If it cannot definitely determine the safety after visiting
852 /// two instructions it assumes it's not safe.
853 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
854 MachineBasicBlock::iterator I) {
855 // It's always safe to clobber EFLAGS at the end of a block.
859 // For compile time consideration, if we are not able to determine the
860 // safety after visiting 2 instructions, we will assume it's not safe.
861 for (unsigned i = 0; i < 2; ++i) {
862 bool SeenDef = false;
863 for (unsigned j = 0, e = I->getNumOperands(); j != e; ++j) {
864 MachineOperand &MO = I->getOperand(j);
867 if (MO.getReg() == X86::EFLAGS) {
875 // This instruction defines EFLAGS, no need to look any further.
879 // If we make it to the end of the block, it's safe to clobber EFLAGS.
884 // Conservative answer.
888 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
889 MachineBasicBlock::iterator I,
891 const MachineInstr *Orig) const {
892 unsigned SubIdx = Orig->getOperand(0).isReg()
893 ? Orig->getOperand(0).getSubReg() : 0;
894 bool ChangeSubIdx = SubIdx != 0;
895 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
896 DestReg = RI.getSubReg(DestReg, SubIdx);
900 // MOV32r0 etc. are implemented with xor which clobbers condition code.
901 // Re-materialize them as movri instructions to avoid side effects.
902 bool Emitted = false;
903 switch (Orig->getOpcode()) {
909 if (!isSafeToClobberEFLAGS(MBB, I)) {
911 switch (Orig->getOpcode()) {
913 case X86::MOV8r0: Opc = X86::MOV8ri; break;
914 case X86::MOV16r0: Opc = X86::MOV16ri; break;
915 case X86::MOV32r0: Opc = X86::MOV32ri; break;
916 case X86::MOV64r0: Opc = X86::MOV64ri32; break;
918 BuildMI(MBB, I, get(Opc), DestReg).addImm(0);
926 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
927 MI->getOperand(0).setReg(DestReg);
932 MachineInstr *NewMI = prior(I);
933 NewMI->getOperand(0).setSubReg(SubIdx);
937 /// isInvariantLoad - Return true if the specified instruction (which is marked
938 /// mayLoad) is loading from a location whose value is invariant across the
939 /// function. For example, loading a value from the constant pool or from
940 /// from the argument area of a function if it does not change. This should
941 /// only return true of *all* loads the instruction does are invariant (if it
942 /// does multiple loads).
943 bool X86InstrInfo::isInvariantLoad(const MachineInstr *MI) const {
944 // This code cares about loads from three cases: constant pool entries,
945 // invariant argument slots, and global stubs. In order to handle these cases
946 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
947 // operand and base our analysis on it. This is safe because the address of
948 // none of these three cases is ever used as anything other than a load base
949 // and X86 doesn't have any instructions that load from multiple places.
951 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
952 const MachineOperand &MO = MI->getOperand(i);
953 // Loads from constant pools are trivially invariant.
958 return isGVStub(MO.getGlobal(), TM);
960 // If this is a load from an invariant stack slot, the load is a constant.
962 const MachineFrameInfo &MFI =
963 *MI->getParent()->getParent()->getFrameInfo();
964 int Idx = MO.getIndex();
965 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
969 // All other instances of these instructions are presumed to have other
974 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
975 /// is not marked dead.
976 static bool hasLiveCondCodeDef(MachineInstr *MI) {
977 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
978 MachineOperand &MO = MI->getOperand(i);
979 if (MO.isReg() && MO.isDef() &&
980 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
987 /// convertToThreeAddress - This method must be implemented by targets that
988 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
989 /// may be able to convert a two-address instruction into a true
990 /// three-address instruction on demand. This allows the X86 target (for
991 /// example) to convert ADD and SHL instructions into LEA instructions if they
992 /// would require register copies due to two-addressness.
994 /// This method returns a null pointer if the transformation cannot be
995 /// performed, otherwise it returns the new instruction.
998 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
999 MachineBasicBlock::iterator &MBBI,
1000 LiveVariables *LV) const {
1001 MachineInstr *MI = MBBI;
1002 MachineFunction &MF = *MI->getParent()->getParent();
1003 // All instructions input are two-addr instructions. Get the known operands.
1004 unsigned Dest = MI->getOperand(0).getReg();
1005 unsigned Src = MI->getOperand(1).getReg();
1006 bool isDead = MI->getOperand(0).isDead();
1007 bool isKill = MI->getOperand(1).isKill();
1009 MachineInstr *NewMI = NULL;
1010 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1011 // we have better subtarget support, enable the 16-bit LEA generation here.
1012 bool DisableLEA16 = true;
1014 unsigned MIOpc = MI->getOpcode();
1016 case X86::SHUFPSrri: {
1017 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1018 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1020 unsigned B = MI->getOperand(1).getReg();
1021 unsigned C = MI->getOperand(2).getReg();
1022 if (B != C) return 0;
1023 unsigned A = MI->getOperand(0).getReg();
1024 unsigned M = MI->getOperand(3).getImm();
1025 NewMI = BuildMI(MF, get(X86::PSHUFDri)).addReg(A, true, false, false, isDead)
1026 .addReg(B, false, false, isKill).addImm(M);
1029 case X86::SHL64ri: {
1030 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1031 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1032 // the flags produced by a shift yet, so this is safe.
1033 unsigned ShAmt = MI->getOperand(2).getImm();
1034 if (ShAmt == 0 || ShAmt >= 4) return 0;
1036 NewMI = BuildMI(MF, get(X86::LEA64r)).addReg(Dest, true, false, false, isDead)
1037 .addReg(0).addImm(1 << ShAmt).addReg(Src, false, false, isKill).addImm(0);
1040 case X86::SHL32ri: {
1041 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1042 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1043 // the flags produced by a shift yet, so this is safe.
1044 unsigned ShAmt = MI->getOperand(2).getImm();
1045 if (ShAmt == 0 || ShAmt >= 4) return 0;
1047 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1048 X86::LEA64_32r : X86::LEA32r;
1049 NewMI = BuildMI(MF, get(Opc)).addReg(Dest, true, false, false, isDead)
1050 .addReg(0).addImm(1 << ShAmt)
1051 .addReg(Src, false, false, isKill).addImm(0);
1054 case X86::SHL16ri: {
1055 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1056 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1057 // the flags produced by a shift yet, so this is safe.
1058 unsigned ShAmt = MI->getOperand(2).getImm();
1059 if (ShAmt == 0 || ShAmt >= 4) return 0;
1062 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1063 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1064 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1065 ? X86::LEA64_32r : X86::LEA32r;
1066 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1067 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1069 // Build and insert into an implicit UNDEF value. This is OK because
1070 // well be shifting and then extracting the lower 16-bits.
1071 BuildMI(*MFI, MBBI, get(X86::IMPLICIT_DEF), leaInReg);
1072 MachineInstr *InsMI = BuildMI(*MFI, MBBI, get(X86::INSERT_SUBREG),leaInReg)
1073 .addReg(leaInReg).addReg(Src, false, false, isKill)
1074 .addImm(X86::SUBREG_16BIT);
1076 NewMI = BuildMI(*MFI, MBBI, get(Opc), leaOutReg).addReg(0).addImm(1 << ShAmt)
1077 .addReg(leaInReg, false, false, true).addImm(0);
1079 MachineInstr *ExtMI = BuildMI(*MFI, MBBI, get(X86::EXTRACT_SUBREG))
1080 .addReg(Dest, true, false, false, isDead)
1081 .addReg(leaOutReg, false, false, true).addImm(X86::SUBREG_16BIT);
1083 // Update live variables
1084 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1085 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1087 LV->replaceKillInstruction(Src, MI, InsMI);
1089 LV->replaceKillInstruction(Dest, MI, ExtMI);
1093 NewMI = BuildMI(MF, get(X86::LEA16r)).addReg(Dest, true, false, false, isDead)
1094 .addReg(0).addImm(1 << ShAmt)
1095 .addReg(Src, false, false, isKill).addImm(0);
1100 // The following opcodes also sets the condition code register(s). Only
1101 // convert them to equivalent lea if the condition code register def's
1103 if (hasLiveCondCodeDef(MI))
1106 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1111 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1112 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1113 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1114 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1115 .addReg(Dest, true, false, false, isDead),
1120 case X86::INC64_16r:
1121 if (DisableLEA16) return 0;
1122 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1123 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1124 .addReg(Dest, true, false, false, isDead),
1129 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1130 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1131 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1132 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1133 .addReg(Dest, true, false, false, isDead),
1138 case X86::DEC64_16r:
1139 if (DisableLEA16) return 0;
1140 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1141 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1142 .addReg(Dest, true, false, false, isDead),
1146 case X86::ADD32rr: {
1147 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1148 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1149 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1150 unsigned Src2 = MI->getOperand(2).getReg();
1151 bool isKill2 = MI->getOperand(2).isKill();
1152 NewMI = addRegReg(BuildMI(MF, get(Opc))
1153 .addReg(Dest, true, false, false, isDead),
1154 Src, isKill, Src2, isKill2);
1156 LV->replaceKillInstruction(Src2, MI, NewMI);
1159 case X86::ADD16rr: {
1160 if (DisableLEA16) return 0;
1161 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1162 unsigned Src2 = MI->getOperand(2).getReg();
1163 bool isKill2 = MI->getOperand(2).isKill();
1164 NewMI = addRegReg(BuildMI(MF, get(X86::LEA16r))
1165 .addReg(Dest, true, false, false, isDead),
1166 Src, isKill, Src2, isKill2);
1168 LV->replaceKillInstruction(Src2, MI, NewMI);
1171 case X86::ADD64ri32:
1173 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1174 if (MI->getOperand(2).isImm())
1175 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA64r))
1176 .addReg(Dest, true, false, false, isDead),
1177 Src, isKill, MI->getOperand(2).getImm());
1181 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1182 if (MI->getOperand(2).isImm()) {
1183 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1184 NewMI = addRegOffset(BuildMI(MF, get(Opc))
1185 .addReg(Dest, true, false, false, isDead),
1186 Src, isKill, MI->getOperand(2).getImm());
1191 if (DisableLEA16) return 0;
1192 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1193 if (MI->getOperand(2).isImm())
1194 NewMI = addRegOffset(BuildMI(MF, get(X86::LEA16r))
1195 .addReg(Dest, true, false, false, isDead),
1196 Src, isKill, MI->getOperand(2).getImm());
1199 if (DisableLEA16) return 0;
1201 case X86::SHL64ri: {
1202 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImm() &&
1203 "Unknown shl instruction!");
1204 unsigned ShAmt = MI->getOperand(2).getImm();
1205 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1207 AM.Scale = 1 << ShAmt;
1209 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1210 : (MIOpc == X86::SHL32ri
1211 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1212 NewMI = addFullAddress(BuildMI(MF, get(Opc))
1213 .addReg(Dest, true, false, false, isDead), AM);
1215 NewMI->getOperand(3).setIsKill(true);
1223 if (!NewMI) return 0;
1225 if (LV) { // Update live variables
1227 LV->replaceKillInstruction(Src, MI, NewMI);
1229 LV->replaceKillInstruction(Dest, MI, NewMI);
1232 MFI->insert(MBBI, NewMI); // Insert the new inst
1236 /// commuteInstruction - We have a few instructions that must be hacked on to
1240 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1241 switch (MI->getOpcode()) {
1242 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1243 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1244 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1245 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1246 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1247 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1250 switch (MI->getOpcode()) {
1251 default: assert(0 && "Unreachable!");
1252 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1253 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1254 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1255 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1256 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1257 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1259 unsigned Amt = MI->getOperand(3).getImm();
1261 MachineFunction &MF = *MI->getParent()->getParent();
1262 MI = MF.CloneMachineInstr(MI);
1265 MI->setDesc(get(Opc));
1266 MI->getOperand(3).setImm(Size-Amt);
1267 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1269 case X86::CMOVB16rr:
1270 case X86::CMOVB32rr:
1271 case X86::CMOVB64rr:
1272 case X86::CMOVAE16rr:
1273 case X86::CMOVAE32rr:
1274 case X86::CMOVAE64rr:
1275 case X86::CMOVE16rr:
1276 case X86::CMOVE32rr:
1277 case X86::CMOVE64rr:
1278 case X86::CMOVNE16rr:
1279 case X86::CMOVNE32rr:
1280 case X86::CMOVNE64rr:
1281 case X86::CMOVBE16rr:
1282 case X86::CMOVBE32rr:
1283 case X86::CMOVBE64rr:
1284 case X86::CMOVA16rr:
1285 case X86::CMOVA32rr:
1286 case X86::CMOVA64rr:
1287 case X86::CMOVL16rr:
1288 case X86::CMOVL32rr:
1289 case X86::CMOVL64rr:
1290 case X86::CMOVGE16rr:
1291 case X86::CMOVGE32rr:
1292 case X86::CMOVGE64rr:
1293 case X86::CMOVLE16rr:
1294 case X86::CMOVLE32rr:
1295 case X86::CMOVLE64rr:
1296 case X86::CMOVG16rr:
1297 case X86::CMOVG32rr:
1298 case X86::CMOVG64rr:
1299 case X86::CMOVS16rr:
1300 case X86::CMOVS32rr:
1301 case X86::CMOVS64rr:
1302 case X86::CMOVNS16rr:
1303 case X86::CMOVNS32rr:
1304 case X86::CMOVNS64rr:
1305 case X86::CMOVP16rr:
1306 case X86::CMOVP32rr:
1307 case X86::CMOVP64rr:
1308 case X86::CMOVNP16rr:
1309 case X86::CMOVNP32rr:
1310 case X86::CMOVNP64rr: {
1312 switch (MI->getOpcode()) {
1314 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1315 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1316 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1317 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1318 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1319 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1320 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1321 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1322 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1323 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1324 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1325 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1326 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1327 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1328 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1329 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1330 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1331 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1332 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1333 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1334 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1335 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1336 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1337 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1338 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1339 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1340 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1341 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1342 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1343 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1344 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1345 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1346 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1347 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1348 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1349 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1350 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1351 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1352 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1353 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1354 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1355 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1358 MachineFunction &MF = *MI->getParent()->getParent();
1359 MI = MF.CloneMachineInstr(MI);
1362 MI->setDesc(get(Opc));
1363 // Fallthrough intended.
1366 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1370 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1372 default: return X86::COND_INVALID;
1373 case X86::JE: return X86::COND_E;
1374 case X86::JNE: return X86::COND_NE;
1375 case X86::JL: return X86::COND_L;
1376 case X86::JLE: return X86::COND_LE;
1377 case X86::JG: return X86::COND_G;
1378 case X86::JGE: return X86::COND_GE;
1379 case X86::JB: return X86::COND_B;
1380 case X86::JBE: return X86::COND_BE;
1381 case X86::JA: return X86::COND_A;
1382 case X86::JAE: return X86::COND_AE;
1383 case X86::JS: return X86::COND_S;
1384 case X86::JNS: return X86::COND_NS;
1385 case X86::JP: return X86::COND_P;
1386 case X86::JNP: return X86::COND_NP;
1387 case X86::JO: return X86::COND_O;
1388 case X86::JNO: return X86::COND_NO;
1389 case X86::JC: return X86::COND_C;
1390 case X86::JNC: return X86::COND_NC;
1394 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1396 default: assert(0 && "Illegal condition code!");
1397 case X86::COND_E: return X86::JE;
1398 case X86::COND_NE: return X86::JNE;
1399 case X86::COND_L: return X86::JL;
1400 case X86::COND_LE: return X86::JLE;
1401 case X86::COND_G: return X86::JG;
1402 case X86::COND_GE: return X86::JGE;
1403 case X86::COND_B: return X86::JB;
1404 case X86::COND_BE: return X86::JBE;
1405 case X86::COND_A: return X86::JA;
1406 case X86::COND_AE: return X86::JAE;
1407 case X86::COND_S: return X86::JS;
1408 case X86::COND_NS: return X86::JNS;
1409 case X86::COND_P: return X86::JP;
1410 case X86::COND_NP: return X86::JNP;
1411 case X86::COND_O: return X86::JO;
1412 case X86::COND_NO: return X86::JNO;
1413 case X86::COND_C: return X86::JC;
1414 case X86::COND_NC: return X86::JNC;
1418 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1419 /// e.g. turning COND_E to COND_NE.
1420 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1422 default: assert(0 && "Illegal condition code!");
1423 case X86::COND_E: return X86::COND_NE;
1424 case X86::COND_NE: return X86::COND_E;
1425 case X86::COND_L: return X86::COND_GE;
1426 case X86::COND_LE: return X86::COND_G;
1427 case X86::COND_G: return X86::COND_LE;
1428 case X86::COND_GE: return X86::COND_L;
1429 case X86::COND_B: return X86::COND_AE;
1430 case X86::COND_BE: return X86::COND_A;
1431 case X86::COND_A: return X86::COND_BE;
1432 case X86::COND_AE: return X86::COND_B;
1433 case X86::COND_S: return X86::COND_NS;
1434 case X86::COND_NS: return X86::COND_S;
1435 case X86::COND_P: return X86::COND_NP;
1436 case X86::COND_NP: return X86::COND_P;
1437 case X86::COND_O: return X86::COND_NO;
1438 case X86::COND_NO: return X86::COND_O;
1439 case X86::COND_C: return X86::COND_NC;
1440 case X86::COND_NC: return X86::COND_C;
1444 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1445 const TargetInstrDesc &TID = MI->getDesc();
1446 if (!TID.isTerminator()) return false;
1448 // Conditional branch is a special case.
1449 if (TID.isBranch() && !TID.isBarrier())
1451 if (!TID.isPredicable())
1453 return !isPredicated(MI);
1456 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1457 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1458 const X86InstrInfo &TII) {
1459 if (MI->getOpcode() == X86::FP_REG_KILL)
1461 return TII.isUnpredicatedTerminator(MI);
1464 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1465 MachineBasicBlock *&TBB,
1466 MachineBasicBlock *&FBB,
1467 SmallVectorImpl<MachineOperand> &Cond) const {
1468 // Start from the bottom of the block and work up, examining the
1469 // terminator instructions.
1470 MachineBasicBlock::iterator I = MBB.end();
1471 while (I != MBB.begin()) {
1473 // Working from the bottom, when we see a non-terminator
1474 // instruction, we're done.
1475 if (!isBrAnalysisUnpredicatedTerminator(I, *this))
1477 // A terminator that isn't a branch can't easily be handled
1478 // by this analysis.
1479 if (!I->getDesc().isBranch())
1481 // Handle unconditional branches.
1482 if (I->getOpcode() == X86::JMP) {
1483 // If the block has any instructions after a JMP, delete them.
1484 while (next(I) != MBB.end())
1485 next(I)->eraseFromParent();
1488 // Delete the JMP if it's equivalent to a fall-through.
1489 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1491 I->eraseFromParent();
1495 // TBB is used to indicate the unconditinal destination.
1496 TBB = I->getOperand(0).getMBB();
1499 // Handle conditional branches.
1500 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1501 if (BranchCode == X86::COND_INVALID)
1502 return true; // Can't handle indirect branch.
1503 // Working from the bottom, handle the first conditional branch.
1506 TBB = I->getOperand(0).getMBB();
1507 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1510 // Handle subsequent conditional branches. Only handle the case
1511 // where all conditional branches branch to the same destination
1512 // and their condition opcodes fit one of the special
1513 // multi-branch idioms.
1514 assert(Cond.size() == 1);
1516 // Only handle the case where all conditional branches branch to
1517 // the same destination.
1518 if (TBB != I->getOperand(0).getMBB())
1520 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1521 // If the conditions are the same, we can leave them alone.
1522 if (OldBranchCode == BranchCode)
1524 // If they differ, see if they fit one of the known patterns.
1525 // Theoretically we could handle more patterns here, but
1526 // we shouldn't expect to see them if instruction selection
1527 // has done a reasonable job.
1528 if ((OldBranchCode == X86::COND_NP &&
1529 BranchCode == X86::COND_E) ||
1530 (OldBranchCode == X86::COND_E &&
1531 BranchCode == X86::COND_NP))
1532 BranchCode = X86::COND_NP_OR_E;
1533 else if ((OldBranchCode == X86::COND_P &&
1534 BranchCode == X86::COND_NE) ||
1535 (OldBranchCode == X86::COND_NE &&
1536 BranchCode == X86::COND_P))
1537 BranchCode = X86::COND_NE_OR_P;
1540 // Update the MachineOperand.
1541 Cond[0].setImm(BranchCode);
1547 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1548 MachineBasicBlock::iterator I = MBB.end();
1551 while (I != MBB.begin()) {
1553 if (I->getOpcode() != X86::JMP &&
1554 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1556 // Remove the branch.
1557 I->eraseFromParent();
1565 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
1566 const MachineOperand &MO) {
1568 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
1569 MO.isKill(), MO.isDead(), MO.getSubReg());
1570 else if (MO.isImm())
1571 MIB = MIB.addImm(MO.getImm());
1573 MIB = MIB.addFrameIndex(MO.getIndex());
1574 else if (MO.isGlobal())
1575 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
1576 else if (MO.isCPI())
1577 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
1578 else if (MO.isJTI())
1579 MIB = MIB.addJumpTableIndex(MO.getIndex());
1580 else if (MO.isSymbol())
1581 MIB = MIB.addExternalSymbol(MO.getSymbolName());
1583 assert(0 && "Unknown operand for X86InstrAddOperand!");
1589 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1590 MachineBasicBlock *FBB,
1591 const SmallVectorImpl<MachineOperand> &Cond) const {
1592 // Shouldn't be a fall through.
1593 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1594 assert((Cond.size() == 1 || Cond.size() == 0) &&
1595 "X86 branch conditions have one component!");
1598 // Unconditional branch?
1599 assert(!FBB && "Unconditional branch with multiple successors!");
1600 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
1604 // Conditional branch.
1606 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1608 case X86::COND_NP_OR_E:
1609 // Synthesize NP_OR_E with two branches.
1610 BuildMI(&MBB, get(X86::JNP)).addMBB(TBB);
1612 BuildMI(&MBB, get(X86::JE)).addMBB(TBB);
1615 case X86::COND_NE_OR_P:
1616 // Synthesize NE_OR_P with two branches.
1617 BuildMI(&MBB, get(X86::JNE)).addMBB(TBB);
1619 BuildMI(&MBB, get(X86::JP)).addMBB(TBB);
1623 unsigned Opc = GetCondBranchFromCond(CC);
1624 BuildMI(&MBB, get(Opc)).addMBB(TBB);
1629 // Two-way Conditional branch. Insert the second branch.
1630 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
1636 bool X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1637 MachineBasicBlock::iterator MI,
1638 unsigned DestReg, unsigned SrcReg,
1639 const TargetRegisterClass *DestRC,
1640 const TargetRegisterClass *SrcRC) const {
1641 if (DestRC == SrcRC) {
1643 if (DestRC == &X86::GR64RegClass) {
1645 } else if (DestRC == &X86::GR32RegClass) {
1647 } else if (DestRC == &X86::GR16RegClass) {
1649 } else if (DestRC == &X86::GR8RegClass) {
1651 } else if (DestRC == &X86::GR32_RegClass) {
1652 Opc = X86::MOV32_rr;
1653 } else if (DestRC == &X86::GR16_RegClass) {
1654 Opc = X86::MOV16_rr;
1655 } else if (DestRC == &X86::RFP32RegClass) {
1656 Opc = X86::MOV_Fp3232;
1657 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1658 Opc = X86::MOV_Fp6464;
1659 } else if (DestRC == &X86::RFP80RegClass) {
1660 Opc = X86::MOV_Fp8080;
1661 } else if (DestRC == &X86::FR32RegClass) {
1662 Opc = X86::FsMOVAPSrr;
1663 } else if (DestRC == &X86::FR64RegClass) {
1664 Opc = X86::FsMOVAPDrr;
1665 } else if (DestRC == &X86::VR128RegClass) {
1666 Opc = X86::MOVAPSrr;
1667 } else if (DestRC == &X86::VR64RegClass) {
1668 Opc = X86::MMX_MOVQ64rr;
1672 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
1676 // Moving EFLAGS to / from another register requires a push and a pop.
1677 if (SrcRC == &X86::CCRRegClass) {
1678 if (SrcReg != X86::EFLAGS)
1680 if (DestRC == &X86::GR64RegClass) {
1681 BuildMI(MBB, MI, get(X86::PUSHFQ));
1682 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
1684 } else if (DestRC == &X86::GR32RegClass) {
1685 BuildMI(MBB, MI, get(X86::PUSHFD));
1686 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
1689 } else if (DestRC == &X86::CCRRegClass) {
1690 if (DestReg != X86::EFLAGS)
1692 if (SrcRC == &X86::GR64RegClass) {
1693 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
1694 BuildMI(MBB, MI, get(X86::POPFQ));
1696 } else if (SrcRC == &X86::GR32RegClass) {
1697 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
1698 BuildMI(MBB, MI, get(X86::POPFD));
1703 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1704 if (SrcRC == &X86::RSTRegClass) {
1705 // Copying from ST(0)/ST(1).
1706 if (SrcReg != X86::ST0 && SrcReg != X86::ST1)
1707 // Can only copy from ST(0)/ST(1) right now
1709 bool isST0 = SrcReg == X86::ST0;
1711 if (DestRC == &X86::RFP32RegClass)
1712 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1713 else if (DestRC == &X86::RFP64RegClass)
1714 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1716 if (DestRC != &X86::RFP80RegClass)
1718 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1720 BuildMI(MBB, MI, get(Opc), DestReg);
1724 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1725 if (DestRC == &X86::RSTRegClass) {
1726 // Copying to ST(0). FIXME: handle ST(1) also
1727 if (DestReg != X86::ST0)
1728 // Can only copy to TOS right now
1731 if (SrcRC == &X86::RFP32RegClass)
1732 Opc = X86::FpSET_ST0_32;
1733 else if (SrcRC == &X86::RFP64RegClass)
1734 Opc = X86::FpSET_ST0_64;
1736 if (SrcRC != &X86::RFP80RegClass)
1738 Opc = X86::FpSET_ST0_80;
1740 BuildMI(MBB, MI, get(Opc)).addReg(SrcReg);
1744 // Not yet supported!
1748 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1749 bool isStackAligned) {
1751 if (RC == &X86::GR64RegClass) {
1753 } else if (RC == &X86::GR32RegClass) {
1755 } else if (RC == &X86::GR16RegClass) {
1757 } else if (RC == &X86::GR8RegClass) {
1759 } else if (RC == &X86::GR32_RegClass) {
1760 Opc = X86::MOV32_mr;
1761 } else if (RC == &X86::GR16_RegClass) {
1762 Opc = X86::MOV16_mr;
1763 } else if (RC == &X86::RFP80RegClass) {
1764 Opc = X86::ST_FpP80m; // pops
1765 } else if (RC == &X86::RFP64RegClass) {
1766 Opc = X86::ST_Fp64m;
1767 } else if (RC == &X86::RFP32RegClass) {
1768 Opc = X86::ST_Fp32m;
1769 } else if (RC == &X86::FR32RegClass) {
1771 } else if (RC == &X86::FR64RegClass) {
1773 } else if (RC == &X86::VR128RegClass) {
1774 // If stack is realigned we can use aligned stores.
1775 Opc = isStackAligned ? X86::MOVAPSmr : X86::MOVUPSmr;
1776 } else if (RC == &X86::VR64RegClass) {
1777 Opc = X86::MMX_MOVQ64mr;
1779 assert(0 && "Unknown regclass");
1786 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1787 MachineBasicBlock::iterator MI,
1788 unsigned SrcReg, bool isKill, int FrameIdx,
1789 const TargetRegisterClass *RC) const {
1790 const MachineFunction &MF = *MBB.getParent();
1791 bool isAligned = (RI.getStackAlignment() >= 16) ||
1792 RI.needsStackRealignment(MF);
1793 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1794 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
1795 .addReg(SrcReg, false, false, isKill);
1798 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1800 SmallVectorImpl<MachineOperand> &Addr,
1801 const TargetRegisterClass *RC,
1802 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1803 bool isAligned = (RI.getStackAlignment() >= 16) ||
1804 RI.needsStackRealignment(MF);
1805 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1806 MachineInstrBuilder MIB = BuildMI(MF, get(Opc));
1807 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1808 MIB = X86InstrAddOperand(MIB, Addr[i]);
1809 MIB.addReg(SrcReg, false, false, isKill);
1810 NewMIs.push_back(MIB);
1813 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1814 bool isStackAligned) {
1816 if (RC == &X86::GR64RegClass) {
1818 } else if (RC == &X86::GR32RegClass) {
1820 } else if (RC == &X86::GR16RegClass) {
1822 } else if (RC == &X86::GR8RegClass) {
1824 } else if (RC == &X86::GR32_RegClass) {
1825 Opc = X86::MOV32_rm;
1826 } else if (RC == &X86::GR16_RegClass) {
1827 Opc = X86::MOV16_rm;
1828 } else if (RC == &X86::RFP80RegClass) {
1829 Opc = X86::LD_Fp80m;
1830 } else if (RC == &X86::RFP64RegClass) {
1831 Opc = X86::LD_Fp64m;
1832 } else if (RC == &X86::RFP32RegClass) {
1833 Opc = X86::LD_Fp32m;
1834 } else if (RC == &X86::FR32RegClass) {
1836 } else if (RC == &X86::FR64RegClass) {
1838 } else if (RC == &X86::VR128RegClass) {
1839 // If stack is realigned we can use aligned loads.
1840 Opc = isStackAligned ? X86::MOVAPSrm : X86::MOVUPSrm;
1841 } else if (RC == &X86::VR64RegClass) {
1842 Opc = X86::MMX_MOVQ64rm;
1844 assert(0 && "Unknown regclass");
1851 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1852 MachineBasicBlock::iterator MI,
1853 unsigned DestReg, int FrameIdx,
1854 const TargetRegisterClass *RC) const{
1855 const MachineFunction &MF = *MBB.getParent();
1856 bool isAligned = (RI.getStackAlignment() >= 16) ||
1857 RI.needsStackRealignment(MF);
1858 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1859 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
1862 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1863 SmallVectorImpl<MachineOperand> &Addr,
1864 const TargetRegisterClass *RC,
1865 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1866 bool isAligned = (RI.getStackAlignment() >= 16) ||
1867 RI.needsStackRealignment(MF);
1868 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1869 MachineInstrBuilder MIB = BuildMI(MF, get(Opc), DestReg);
1870 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1871 MIB = X86InstrAddOperand(MIB, Addr[i]);
1872 NewMIs.push_back(MIB);
1875 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1876 MachineBasicBlock::iterator MI,
1877 const std::vector<CalleeSavedInfo> &CSI) const {
1881 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1882 unsigned SlotSize = is64Bit ? 8 : 4;
1884 MachineFunction &MF = *MBB.getParent();
1885 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1886 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1888 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1889 for (unsigned i = CSI.size(); i != 0; --i) {
1890 unsigned Reg = CSI[i-1].getReg();
1891 // Add the callee-saved register as live-in. It's killed at the spill.
1893 BuildMI(MBB, MI, get(Opc))
1894 .addReg(Reg, /*isDef=*/false, /*isImp=*/false, /*isKill=*/true);
1899 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1900 MachineBasicBlock::iterator MI,
1901 const std::vector<CalleeSavedInfo> &CSI) const {
1905 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1907 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1908 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1909 unsigned Reg = CSI[i].getReg();
1910 BuildMI(MBB, MI, get(Opc), Reg);
1915 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
1916 const SmallVector<MachineOperand,4> &MOs,
1917 MachineInstr *MI, const TargetInstrInfo &TII) {
1918 // Create the base instruction with the memory operand as the first part.
1919 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), true);
1920 MachineInstrBuilder MIB(NewMI);
1921 unsigned NumAddrOps = MOs.size();
1922 for (unsigned i = 0; i != NumAddrOps; ++i)
1923 MIB = X86InstrAddOperand(MIB, MOs[i]);
1924 if (NumAddrOps < 4) // FrameIndex only
1925 MIB.addImm(1).addReg(0).addImm(0);
1927 // Loop over the rest of the ri operands, converting them over.
1928 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1929 for (unsigned i = 0; i != NumOps; ++i) {
1930 MachineOperand &MO = MI->getOperand(i+2);
1931 MIB = X86InstrAddOperand(MIB, MO);
1933 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1934 MachineOperand &MO = MI->getOperand(i);
1935 MIB = X86InstrAddOperand(MIB, MO);
1940 static MachineInstr *FuseInst(MachineFunction &MF,
1941 unsigned Opcode, unsigned OpNo,
1942 const SmallVector<MachineOperand,4> &MOs,
1943 MachineInstr *MI, const TargetInstrInfo &TII) {
1944 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), true);
1945 MachineInstrBuilder MIB(NewMI);
1947 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1948 MachineOperand &MO = MI->getOperand(i);
1950 assert(MO.isReg() && "Expected to fold into reg operand!");
1951 unsigned NumAddrOps = MOs.size();
1952 for (unsigned i = 0; i != NumAddrOps; ++i)
1953 MIB = X86InstrAddOperand(MIB, MOs[i]);
1954 if (NumAddrOps < 4) // FrameIndex only
1955 MIB.addImm(1).addReg(0).addImm(0);
1957 MIB = X86InstrAddOperand(MIB, MO);
1963 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
1964 const SmallVector<MachineOperand,4> &MOs,
1966 MachineFunction &MF = *MI->getParent()->getParent();
1967 MachineInstrBuilder MIB = BuildMI(MF, TII.get(Opcode));
1969 unsigned NumAddrOps = MOs.size();
1970 for (unsigned i = 0; i != NumAddrOps; ++i)
1971 MIB = X86InstrAddOperand(MIB, MOs[i]);
1972 if (NumAddrOps < 4) // FrameIndex only
1973 MIB.addImm(1).addReg(0).addImm(0);
1974 return MIB.addImm(0);
1978 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
1979 MachineInstr *MI, unsigned i,
1980 const SmallVector<MachineOperand,4> &MOs) const{
1981 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
1982 bool isTwoAddrFold = false;
1983 unsigned NumOps = MI->getDesc().getNumOperands();
1984 bool isTwoAddr = NumOps > 1 &&
1985 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
1987 MachineInstr *NewMI = NULL;
1988 // Folding a memory location into the two-address part of a two-address
1989 // instruction is different than folding it other places. It requires
1990 // replacing the *two* registers with the memory location.
1991 if (isTwoAddr && NumOps >= 2 && i < 2 &&
1992 MI->getOperand(0).isReg() &&
1993 MI->getOperand(1).isReg() &&
1994 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
1995 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
1996 isTwoAddrFold = true;
1997 } else if (i == 0) { // If operand 0
1998 if (MI->getOpcode() == X86::MOV16r0)
1999 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
2000 else if (MI->getOpcode() == X86::MOV32r0)
2001 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
2002 else if (MI->getOpcode() == X86::MOV64r0)
2003 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2004 else if (MI->getOpcode() == X86::MOV8r0)
2005 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2009 OpcodeTablePtr = &RegOp2MemOpTable0;
2010 } else if (i == 1) {
2011 OpcodeTablePtr = &RegOp2MemOpTable1;
2012 } else if (i == 2) {
2013 OpcodeTablePtr = &RegOp2MemOpTable2;
2016 // If table selected...
2017 if (OpcodeTablePtr) {
2018 // Find the Opcode to fuse
2019 DenseMap<unsigned*, unsigned>::iterator I =
2020 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
2021 if (I != OpcodeTablePtr->end()) {
2023 NewMI = FuseTwoAddrInst(MF, I->second, MOs, MI, *this);
2025 NewMI = FuseInst(MF, I->second, i, MOs, MI, *this);
2031 if (PrintFailedFusing)
2032 cerr << "We failed to fuse operand " << i << *MI;
2037 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2039 const SmallVectorImpl<unsigned> &Ops,
2040 int FrameIndex) const {
2041 // Check switch flag
2042 if (NoFusing) return NULL;
2044 const MachineFrameInfo *MFI = MF.getFrameInfo();
2045 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2046 // FIXME: Move alignment requirement into tables?
2047 if (Alignment < 16) {
2048 switch (MI->getOpcode()) {
2050 // Not always safe to fold movsd into these instructions since their load
2051 // folding variants expects the address to be 16 byte aligned.
2052 case X86::FsANDNPDrr:
2053 case X86::FsANDNPSrr:
2054 case X86::FsANDPDrr:
2055 case X86::FsANDPSrr:
2058 case X86::FsXORPDrr:
2059 case X86::FsXORPSrr:
2064 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2065 unsigned NewOpc = 0;
2066 switch (MI->getOpcode()) {
2067 default: return NULL;
2068 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2069 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2070 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2071 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2073 // Change to CMPXXri r, 0 first.
2074 MI->setDesc(get(NewOpc));
2075 MI->getOperand(1).ChangeToImmediate(0);
2076 } else if (Ops.size() != 1)
2079 SmallVector<MachineOperand,4> MOs;
2080 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2081 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2084 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2086 const SmallVectorImpl<unsigned> &Ops,
2087 MachineInstr *LoadMI) const {
2088 // Check switch flag
2089 if (NoFusing) return NULL;
2091 // Determine the alignment of the load.
2092 unsigned Alignment = 0;
2093 if (LoadMI->hasOneMemOperand())
2094 Alignment = LoadMI->memoperands_begin()->getAlignment();
2096 // FIXME: Move alignment requirement into tables?
2097 if (Alignment < 16) {
2098 switch (MI->getOpcode()) {
2100 // Not always safe to fold movsd into these instructions since their load
2101 // folding variants expects the address to be 16 byte aligned.
2102 case X86::FsANDNPDrr:
2103 case X86::FsANDNPSrr:
2104 case X86::FsANDPDrr:
2105 case X86::FsANDPSrr:
2108 case X86::FsXORPDrr:
2109 case X86::FsXORPSrr:
2114 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2115 unsigned NewOpc = 0;
2116 switch (MI->getOpcode()) {
2117 default: return NULL;
2118 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2119 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2120 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2121 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2123 // Change to CMPXXri r, 0 first.
2124 MI->setDesc(get(NewOpc));
2125 MI->getOperand(1).ChangeToImmediate(0);
2126 } else if (Ops.size() != 1)
2129 SmallVector<MachineOperand,4> MOs;
2130 if (LoadMI->getOpcode() == X86::V_SET0 ||
2131 LoadMI->getOpcode() == X86::V_SETALLONES) {
2132 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
2133 // Create a constant-pool entry and operands to load from it.
2135 // x86-32 PIC requires a PIC base register for constant pools.
2136 unsigned PICBase = 0;
2137 if (TM.getRelocationModel() == Reloc::PIC_ &&
2138 !TM.getSubtarget<X86Subtarget>().is64Bit())
2139 // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
2140 // This doesn't work for several reasons.
2141 // 1. GlobalBaseReg may have been spilled.
2142 // 2. It may not be live at MI.
2145 // Create a v4i32 constant-pool entry.
2146 MachineConstantPool &MCP = *MF.getConstantPool();
2147 const VectorType *Ty = VectorType::get(Type::Int32Ty, 4);
2148 Constant *C = LoadMI->getOpcode() == X86::V_SET0 ?
2149 ConstantVector::getNullValue(Ty) :
2150 ConstantVector::getAllOnesValue(Ty);
2151 unsigned CPI = MCP.getConstantPoolIndex(C, /*AlignmentLog2=*/4);
2153 // Create operands to load from the constant pool entry.
2154 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
2155 MOs.push_back(MachineOperand::CreateImm(1));
2156 MOs.push_back(MachineOperand::CreateReg(0, false));
2157 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
2159 // Folding a normal load. Just copy the load's address operands.
2160 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2161 for (unsigned i = NumOps - 4; i != NumOps; ++i)
2162 MOs.push_back(LoadMI->getOperand(i));
2164 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2168 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2169 const SmallVectorImpl<unsigned> &Ops) const {
2170 // Check switch flag
2171 if (NoFusing) return 0;
2173 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2174 switch (MI->getOpcode()) {
2175 default: return false;
2184 if (Ops.size() != 1)
2187 unsigned OpNum = Ops[0];
2188 unsigned Opc = MI->getOpcode();
2189 unsigned NumOps = MI->getDesc().getNumOperands();
2190 bool isTwoAddr = NumOps > 1 &&
2191 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2193 // Folding a memory location into the two-address part of a two-address
2194 // instruction is different than folding it other places. It requires
2195 // replacing the *two* registers with the memory location.
2196 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2197 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2198 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2199 } else if (OpNum == 0) { // If operand 0
2208 OpcodeTablePtr = &RegOp2MemOpTable0;
2209 } else if (OpNum == 1) {
2210 OpcodeTablePtr = &RegOp2MemOpTable1;
2211 } else if (OpNum == 2) {
2212 OpcodeTablePtr = &RegOp2MemOpTable2;
2215 if (OpcodeTablePtr) {
2216 // Find the Opcode to fuse
2217 DenseMap<unsigned*, unsigned>::iterator I =
2218 OpcodeTablePtr->find((unsigned*)Opc);
2219 if (I != OpcodeTablePtr->end())
2225 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2226 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2227 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2228 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2229 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2230 if (I == MemOp2RegOpTable.end())
2232 unsigned Opc = I->second.first;
2233 unsigned Index = I->second.second & 0xf;
2234 bool FoldedLoad = I->second.second & (1 << 4);
2235 bool FoldedStore = I->second.second & (1 << 5);
2236 if (UnfoldLoad && !FoldedLoad)
2238 UnfoldLoad &= FoldedLoad;
2239 if (UnfoldStore && !FoldedStore)
2241 UnfoldStore &= FoldedStore;
2243 const TargetInstrDesc &TID = get(Opc);
2244 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2245 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2246 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2247 SmallVector<MachineOperand,4> AddrOps;
2248 SmallVector<MachineOperand,2> BeforeOps;
2249 SmallVector<MachineOperand,2> AfterOps;
2250 SmallVector<MachineOperand,4> ImpOps;
2251 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2252 MachineOperand &Op = MI->getOperand(i);
2253 if (i >= Index && i < Index+4)
2254 AddrOps.push_back(Op);
2255 else if (Op.isReg() && Op.isImplicit())
2256 ImpOps.push_back(Op);
2258 BeforeOps.push_back(Op);
2260 AfterOps.push_back(Op);
2263 // Emit the load instruction.
2265 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2267 // Address operands cannot be marked isKill.
2268 for (unsigned i = 1; i != 5; ++i) {
2269 MachineOperand &MO = NewMIs[0]->getOperand(i);
2271 MO.setIsKill(false);
2276 // Emit the data processing instruction.
2277 MachineInstr *DataMI = MF.CreateMachineInstr(TID, true);
2278 MachineInstrBuilder MIB(DataMI);
2281 MIB.addReg(Reg, true);
2282 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2283 MIB = X86InstrAddOperand(MIB, BeforeOps[i]);
2286 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2287 MIB = X86InstrAddOperand(MIB, AfterOps[i]);
2288 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2289 MachineOperand &MO = ImpOps[i];
2290 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2292 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2293 unsigned NewOpc = 0;
2294 switch (DataMI->getOpcode()) {
2296 case X86::CMP64ri32:
2300 MachineOperand &MO0 = DataMI->getOperand(0);
2301 MachineOperand &MO1 = DataMI->getOperand(1);
2302 if (MO1.getImm() == 0) {
2303 switch (DataMI->getOpcode()) {
2305 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2306 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2307 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2308 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2310 DataMI->setDesc(get(NewOpc));
2311 MO1.ChangeToRegister(MO0.getReg(), false);
2315 NewMIs.push_back(DataMI);
2317 // Emit the store instruction.
2319 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2320 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2321 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2322 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2329 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2330 SmallVectorImpl<SDNode*> &NewNodes) const {
2331 if (!N->isMachineOpcode())
2334 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2335 MemOp2RegOpTable.find((unsigned*)N->getMachineOpcode());
2336 if (I == MemOp2RegOpTable.end())
2338 unsigned Opc = I->second.first;
2339 unsigned Index = I->second.second & 0xf;
2340 bool FoldedLoad = I->second.second & (1 << 4);
2341 bool FoldedStore = I->second.second & (1 << 5);
2342 const TargetInstrDesc &TID = get(Opc);
2343 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2344 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2345 ? getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2346 std::vector<SDValue> AddrOps;
2347 std::vector<SDValue> BeforeOps;
2348 std::vector<SDValue> AfterOps;
2349 unsigned NumOps = N->getNumOperands();
2350 for (unsigned i = 0; i != NumOps-1; ++i) {
2351 SDValue Op = N->getOperand(i);
2352 if (i >= Index && i < Index+4)
2353 AddrOps.push_back(Op);
2355 BeforeOps.push_back(Op);
2357 AfterOps.push_back(Op);
2359 SDValue Chain = N->getOperand(NumOps-1);
2360 AddrOps.push_back(Chain);
2362 // Emit the load instruction.
2364 const MachineFunction &MF = DAG.getMachineFunction();
2366 MVT VT = *RC->vt_begin();
2367 bool isAligned = (RI.getStackAlignment() >= 16) ||
2368 RI.needsStackRealignment(MF);
2369 Load = DAG.getTargetNode(getLoadRegOpcode(RC, isAligned),
2371 &AddrOps[0], AddrOps.size());
2372 NewNodes.push_back(Load);
2375 // Emit the data processing instruction.
2376 std::vector<MVT> VTs;
2377 const TargetRegisterClass *DstRC = 0;
2378 if (TID.getNumDefs() > 0) {
2379 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2380 DstRC = DstTOI.isLookupPtrRegClass()
2381 ? getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2382 VTs.push_back(*DstRC->vt_begin());
2384 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2385 MVT VT = N->getValueType(i);
2386 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2390 BeforeOps.push_back(SDValue(Load, 0));
2391 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2392 SDNode *NewNode= DAG.getTargetNode(Opc, VTs, &BeforeOps[0], BeforeOps.size());
2393 NewNodes.push_back(NewNode);
2395 // Emit the store instruction.
2398 AddrOps.push_back(SDValue(NewNode, 0));
2399 AddrOps.push_back(Chain);
2400 bool isAligned = (RI.getStackAlignment() >= 16) ||
2401 RI.needsStackRealignment(MF);
2402 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, isAligned),
2403 MVT::Other, &AddrOps[0], AddrOps.size());
2404 NewNodes.push_back(Store);
2410 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2411 bool UnfoldLoad, bool UnfoldStore) const {
2412 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2413 MemOp2RegOpTable.find((unsigned*)Opc);
2414 if (I == MemOp2RegOpTable.end())
2416 bool FoldedLoad = I->second.second & (1 << 4);
2417 bool FoldedStore = I->second.second & (1 << 5);
2418 if (UnfoldLoad && !FoldedLoad)
2420 if (UnfoldStore && !FoldedStore)
2422 return I->second.first;
2425 bool X86InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
2426 if (MBB.empty()) return false;
2428 switch (MBB.back().getOpcode()) {
2429 case X86::TCRETURNri:
2430 case X86::TCRETURNdi:
2431 case X86::RET: // Return.
2436 case X86::JMP: // Uncond branch.
2437 case X86::JMP32r: // Indirect branch.
2438 case X86::JMP64r: // Indirect branch (64-bit).
2439 case X86::JMP32m: // Indirect branch through mem.
2440 case X86::JMP64m: // Indirect branch through mem (64-bit).
2442 default: return false;
2447 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
2448 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2449 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
2450 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
2452 Cond[0].setImm(GetOppositeBranchCondition(CC));
2457 IgnoreRegisterClassBarriers(const TargetRegisterClass *RC) const {
2458 // FIXME: Ignore bariers of x87 stack registers for now. We can't
2459 // allow any loads of these registers before FpGet_ST0_80.
2460 return RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
2461 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass;
2464 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
2465 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
2466 if (Subtarget->is64Bit())
2467 return &X86::GR64RegClass;
2469 return &X86::GR32RegClass;
2472 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2473 switch (Desc->TSFlags & X86II::ImmMask) {
2474 case X86II::Imm8: return 1;
2475 case X86II::Imm16: return 2;
2476 case X86II::Imm32: return 4;
2477 case X86II::Imm64: return 8;
2478 default: assert(0 && "Immediate size not set!");
2483 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2484 /// e.g. r8, xmm8, etc.
2485 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2486 if (!MO.isReg()) return false;
2487 switch (MO.getReg()) {
2489 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2490 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2491 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2492 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2493 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2494 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2495 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2496 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2497 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2498 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2505 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2506 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2507 /// size, and 3) use of X86-64 extended registers.
2508 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2510 const TargetInstrDesc &Desc = MI.getDesc();
2512 // Pseudo instructions do not need REX prefix byte.
2513 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2515 if (Desc.TSFlags & X86II::REX_W)
2518 unsigned NumOps = Desc.getNumOperands();
2520 bool isTwoAddr = NumOps > 1 &&
2521 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2523 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2524 unsigned i = isTwoAddr ? 1 : 0;
2525 for (unsigned e = NumOps; i != e; ++i) {
2526 const MachineOperand& MO = MI.getOperand(i);
2528 unsigned Reg = MO.getReg();
2529 if (isX86_64NonExtLowByteReg(Reg))
2534 switch (Desc.TSFlags & X86II::FormMask) {
2535 case X86II::MRMInitReg:
2536 if (isX86_64ExtendedReg(MI.getOperand(0)))
2537 REX |= (1 << 0) | (1 << 2);
2539 case X86II::MRMSrcReg: {
2540 if (isX86_64ExtendedReg(MI.getOperand(0)))
2542 i = isTwoAddr ? 2 : 1;
2543 for (unsigned e = NumOps; i != e; ++i) {
2544 const MachineOperand& MO = MI.getOperand(i);
2545 if (isX86_64ExtendedReg(MO))
2550 case X86II::MRMSrcMem: {
2551 if (isX86_64ExtendedReg(MI.getOperand(0)))
2554 i = isTwoAddr ? 2 : 1;
2555 for (; i != NumOps; ++i) {
2556 const MachineOperand& MO = MI.getOperand(i);
2558 if (isX86_64ExtendedReg(MO))
2565 case X86II::MRM0m: case X86II::MRM1m:
2566 case X86II::MRM2m: case X86II::MRM3m:
2567 case X86II::MRM4m: case X86II::MRM5m:
2568 case X86II::MRM6m: case X86II::MRM7m:
2569 case X86II::MRMDestMem: {
2570 unsigned e = isTwoAddr ? 5 : 4;
2571 i = isTwoAddr ? 1 : 0;
2572 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2575 for (; i != e; ++i) {
2576 const MachineOperand& MO = MI.getOperand(i);
2578 if (isX86_64ExtendedReg(MO))
2586 if (isX86_64ExtendedReg(MI.getOperand(0)))
2588 i = isTwoAddr ? 2 : 1;
2589 for (unsigned e = NumOps; i != e; ++i) {
2590 const MachineOperand& MO = MI.getOperand(i);
2591 if (isX86_64ExtendedReg(MO))
2601 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2602 /// relative block address instruction
2604 static unsigned sizePCRelativeBlockAddress() {
2608 /// sizeGlobalAddress - Give the size of the emission of this global address
2610 static unsigned sizeGlobalAddress(bool dword) {
2611 return dword ? 8 : 4;
2614 /// sizeConstPoolAddress - Give the size of the emission of this constant
2617 static unsigned sizeConstPoolAddress(bool dword) {
2618 return dword ? 8 : 4;
2621 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2624 static unsigned sizeExternalSymbolAddress(bool dword) {
2625 return dword ? 8 : 4;
2628 /// sizeJumpTableAddress - Give the size of the emission of this jump
2631 static unsigned sizeJumpTableAddress(bool dword) {
2632 return dword ? 8 : 4;
2635 static unsigned sizeConstant(unsigned Size) {
2639 static unsigned sizeRegModRMByte(){
2643 static unsigned sizeSIBByte(){
2647 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2648 unsigned FinalSize = 0;
2649 // If this is a simple integer displacement that doesn't require a relocation.
2651 FinalSize += sizeConstant(4);
2655 // Otherwise, this is something that requires a relocation.
2656 if (RelocOp->isGlobal()) {
2657 FinalSize += sizeGlobalAddress(false);
2658 } else if (RelocOp->isCPI()) {
2659 FinalSize += sizeConstPoolAddress(false);
2660 } else if (RelocOp->isJTI()) {
2661 FinalSize += sizeJumpTableAddress(false);
2663 assert(0 && "Unknown value to relocate!");
2668 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2669 bool IsPIC, bool Is64BitMode) {
2670 const MachineOperand &Op3 = MI.getOperand(Op+3);
2672 const MachineOperand *DispForReloc = 0;
2673 unsigned FinalSize = 0;
2675 // Figure out what sort of displacement we have to handle here.
2676 if (Op3.isGlobal()) {
2677 DispForReloc = &Op3;
2678 } else if (Op3.isCPI()) {
2679 if (Is64BitMode || IsPIC) {
2680 DispForReloc = &Op3;
2684 } else if (Op3.isJTI()) {
2685 if (Is64BitMode || IsPIC) {
2686 DispForReloc = &Op3;
2694 const MachineOperand &Base = MI.getOperand(Op);
2695 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2697 unsigned BaseReg = Base.getReg();
2699 // Is a SIB byte needed?
2700 if (IndexReg.getReg() == 0 &&
2701 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2702 if (BaseReg == 0) { // Just a displacement?
2703 // Emit special case [disp32] encoding
2705 FinalSize += getDisplacementFieldSize(DispForReloc);
2707 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2708 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2709 // Emit simple indirect register encoding... [EAX] f.e.
2711 // Be pessimistic and assume it's a disp32, not a disp8
2713 // Emit the most general non-SIB encoding: [REG+disp32]
2715 FinalSize += getDisplacementFieldSize(DispForReloc);
2719 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2720 assert(IndexReg.getReg() != X86::ESP &&
2721 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2723 bool ForceDisp32 = false;
2724 if (BaseReg == 0 || DispForReloc) {
2725 // Emit the normal disp32 encoding.
2732 FinalSize += sizeSIBByte();
2734 // Do we need to output a displacement?
2735 if (DispVal != 0 || ForceDisp32) {
2736 FinalSize += getDisplacementFieldSize(DispForReloc);
2743 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2744 const TargetInstrDesc *Desc,
2745 bool IsPIC, bool Is64BitMode) {
2747 unsigned Opcode = Desc->Opcode;
2748 unsigned FinalSize = 0;
2750 // Emit the lock opcode prefix as needed.
2751 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2753 // Emit segment overrid opcode prefix as needed.
2754 switch (Desc->TSFlags & X86II::SegOvrMask) {
2759 default: assert(0 && "Invalid segment!");
2760 case 0: break; // No segment override!
2763 // Emit the repeat opcode prefix as needed.
2764 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2766 // Emit the operand size opcode prefix as needed.
2767 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2769 // Emit the address size opcode prefix as needed.
2770 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2772 bool Need0FPrefix = false;
2773 switch (Desc->TSFlags & X86II::Op0Mask) {
2774 case X86II::TB: // Two-byte opcode prefix
2775 case X86II::T8: // 0F 38
2776 case X86II::TA: // 0F 3A
2777 Need0FPrefix = true;
2779 case X86II::REP: break; // already handled.
2780 case X86II::XS: // F3 0F
2782 Need0FPrefix = true;
2784 case X86II::XD: // F2 0F
2786 Need0FPrefix = true;
2788 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2789 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2791 break; // Two-byte opcode prefix
2792 default: assert(0 && "Invalid prefix!");
2793 case 0: break; // No prefix!
2798 unsigned REX = X86InstrInfo::determineREX(MI);
2803 // 0x0F escape code must be emitted just before the opcode.
2807 switch (Desc->TSFlags & X86II::Op0Mask) {
2808 case X86II::T8: // 0F 38
2811 case X86II::TA: // 0F 3A
2816 // If this is a two-address instruction, skip one of the register operands.
2817 unsigned NumOps = Desc->getNumOperands();
2819 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2822 switch (Desc->TSFlags & X86II::FormMask) {
2823 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
2825 // Remember the current PC offset, this is the PIC relocation
2830 case TargetInstrInfo::INLINEASM: {
2831 const MachineFunction *MF = MI.getParent()->getParent();
2832 const char *AsmStr = MI.getOperand(0).getSymbolName();
2833 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
2834 FinalSize += AI->getInlineAsmLength(AsmStr);
2837 case TargetInstrInfo::DBG_LABEL:
2838 case TargetInstrInfo::EH_LABEL:
2840 case TargetInstrInfo::IMPLICIT_DEF:
2841 case TargetInstrInfo::DECLARE:
2842 case X86::DWARF_LOC:
2843 case X86::FP_REG_KILL:
2845 case X86::MOVPC32r: {
2846 // This emits the "call" portion of this pseudo instruction.
2848 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2852 case X86::TLS_gs_ri:
2854 FinalSize += sizeGlobalAddress(false);
2862 if (CurOp != NumOps) {
2863 const MachineOperand &MO = MI.getOperand(CurOp++);
2865 FinalSize += sizePCRelativeBlockAddress();
2866 } else if (MO.isGlobal()) {
2867 FinalSize += sizeGlobalAddress(false);
2868 } else if (MO.isSymbol()) {
2869 FinalSize += sizeExternalSymbolAddress(false);
2870 } else if (MO.isImm()) {
2871 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2873 assert(0 && "Unknown RawFrm operand!");
2878 case X86II::AddRegFrm:
2882 if (CurOp != NumOps) {
2883 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2884 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2886 FinalSize += sizeConstant(Size);
2889 if (Opcode == X86::MOV64ri)
2891 if (MO1.isGlobal()) {
2892 FinalSize += sizeGlobalAddress(dword);
2893 } else if (MO1.isSymbol())
2894 FinalSize += sizeExternalSymbolAddress(dword);
2895 else if (MO1.isCPI())
2896 FinalSize += sizeConstPoolAddress(dword);
2897 else if (MO1.isJTI())
2898 FinalSize += sizeJumpTableAddress(dword);
2903 case X86II::MRMDestReg: {
2905 FinalSize += sizeRegModRMByte();
2907 if (CurOp != NumOps) {
2909 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2913 case X86II::MRMDestMem: {
2915 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2917 if (CurOp != NumOps) {
2919 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2924 case X86II::MRMSrcReg:
2926 FinalSize += sizeRegModRMByte();
2928 if (CurOp != NumOps) {
2930 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2934 case X86II::MRMSrcMem: {
2937 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
2939 if (CurOp != NumOps) {
2941 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2946 case X86II::MRM0r: case X86II::MRM1r:
2947 case X86II::MRM2r: case X86II::MRM3r:
2948 case X86II::MRM4r: case X86II::MRM5r:
2949 case X86II::MRM6r: case X86II::MRM7r:
2952 FinalSize += sizeRegModRMByte();
2954 if (CurOp != NumOps) {
2955 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2956 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2958 FinalSize += sizeConstant(Size);
2961 if (Opcode == X86::MOV64ri32)
2963 if (MO1.isGlobal()) {
2964 FinalSize += sizeGlobalAddress(dword);
2965 } else if (MO1.isSymbol())
2966 FinalSize += sizeExternalSymbolAddress(dword);
2967 else if (MO1.isCPI())
2968 FinalSize += sizeConstPoolAddress(dword);
2969 else if (MO1.isJTI())
2970 FinalSize += sizeJumpTableAddress(dword);
2975 case X86II::MRM0m: case X86II::MRM1m:
2976 case X86II::MRM2m: case X86II::MRM3m:
2977 case X86II::MRM4m: case X86II::MRM5m:
2978 case X86II::MRM6m: case X86II::MRM7m: {
2981 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2984 if (CurOp != NumOps) {
2985 const MachineOperand &MO = MI.getOperand(CurOp++);
2986 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2988 FinalSize += sizeConstant(Size);
2991 if (Opcode == X86::MOV64mi32)
2993 if (MO.isGlobal()) {
2994 FinalSize += sizeGlobalAddress(dword);
2995 } else if (MO.isSymbol())
2996 FinalSize += sizeExternalSymbolAddress(dword);
2997 else if (MO.isCPI())
2998 FinalSize += sizeConstPoolAddress(dword);
2999 else if (MO.isJTI())
3000 FinalSize += sizeJumpTableAddress(dword);
3006 case X86II::MRMInitReg:
3008 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
3009 FinalSize += sizeRegModRMByte();
3014 if (!Desc->isVariadic() && CurOp != NumOps) {
3015 cerr << "Cannot determine size: ";
3026 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
3027 const TargetInstrDesc &Desc = MI->getDesc();
3028 bool IsPIC = (TM.getRelocationModel() == Reloc::PIC_);
3029 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
3030 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
3031 if (Desc.getOpcode() == X86::MOVPC32r) {
3032 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);
3037 /// getGlobalBaseReg - Return a virtual register initialized with the
3038 /// the global base register value. Output instructions required to
3039 /// initialize the register in the function entry block, if necessary.
3041 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3042 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3043 "X86-64 PIC uses RIP relative addressing");
3045 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3046 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3047 if (GlobalBaseReg != 0)
3048 return GlobalBaseReg;
3050 // Insert the set of GlobalBaseReg into the first MBB of the function
3051 MachineBasicBlock &FirstMBB = MF->front();
3052 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3053 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3054 unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3056 const TargetInstrInfo *TII = TM.getInstrInfo();
3057 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3058 // only used in JIT code emission as displacement to pc.
3059 BuildMI(FirstMBB, MBBI, TII->get(X86::MOVPC32r), PC).addImm(0);
3061 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3062 // not to pc, but to _GLOBAL_ADDRESS_TABLE_ external
3063 if (TM.getRelocationModel() == Reloc::PIC_ &&
3064 TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3066 RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3067 BuildMI(FirstMBB, MBBI, TII->get(X86::ADD32ri), GlobalBaseReg)
3068 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_");
3073 X86FI->setGlobalBaseReg(GlobalBaseReg);
3074 return GlobalBaseReg;