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/DerivedTypes.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/LiveVariables.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/Target/TargetAsmInfo.h"
36 NoFusing("disable-spill-fusing",
37 cl::desc("Disable fusing of spill code into instructions"));
39 PrintFailedFusing("print-failed-fuse-candidates",
40 cl::desc("Print instructions that the allocator wants to"
41 " fuse, but the X86 backend currently can't"),
44 ReMatPICStubLoad("remat-pic-stub-load",
45 cl::desc("Re-materialize load from stub in PIC mode"),
46 cl::init(false), cl::Hidden);
49 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
50 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
51 TM(tm), RI(tm, *this) {
52 SmallVector<unsigned,16> AmbEntries;
53 static const unsigned OpTbl2Addr[][2] = {
54 { X86::ADC32ri, X86::ADC32mi },
55 { X86::ADC32ri8, X86::ADC32mi8 },
56 { X86::ADC32rr, X86::ADC32mr },
57 { X86::ADC64ri32, X86::ADC64mi32 },
58 { X86::ADC64ri8, X86::ADC64mi8 },
59 { X86::ADC64rr, X86::ADC64mr },
60 { X86::ADD16ri, X86::ADD16mi },
61 { X86::ADD16ri8, X86::ADD16mi8 },
62 { X86::ADD16rr, X86::ADD16mr },
63 { X86::ADD32ri, X86::ADD32mi },
64 { X86::ADD32ri8, X86::ADD32mi8 },
65 { X86::ADD32rr, X86::ADD32mr },
66 { X86::ADD64ri32, X86::ADD64mi32 },
67 { X86::ADD64ri8, X86::ADD64mi8 },
68 { X86::ADD64rr, X86::ADD64mr },
69 { X86::ADD8ri, X86::ADD8mi },
70 { X86::ADD8rr, X86::ADD8mr },
71 { X86::AND16ri, X86::AND16mi },
72 { X86::AND16ri8, X86::AND16mi8 },
73 { X86::AND16rr, X86::AND16mr },
74 { X86::AND32ri, X86::AND32mi },
75 { X86::AND32ri8, X86::AND32mi8 },
76 { X86::AND32rr, X86::AND32mr },
77 { X86::AND64ri32, X86::AND64mi32 },
78 { X86::AND64ri8, X86::AND64mi8 },
79 { X86::AND64rr, X86::AND64mr },
80 { X86::AND8ri, X86::AND8mi },
81 { X86::AND8rr, X86::AND8mr },
82 { X86::DEC16r, X86::DEC16m },
83 { X86::DEC32r, X86::DEC32m },
84 { X86::DEC64_16r, X86::DEC64_16m },
85 { X86::DEC64_32r, X86::DEC64_32m },
86 { X86::DEC64r, X86::DEC64m },
87 { X86::DEC8r, X86::DEC8m },
88 { X86::INC16r, X86::INC16m },
89 { X86::INC32r, X86::INC32m },
90 { X86::INC64_16r, X86::INC64_16m },
91 { X86::INC64_32r, X86::INC64_32m },
92 { X86::INC64r, X86::INC64m },
93 { X86::INC8r, X86::INC8m },
94 { X86::NEG16r, X86::NEG16m },
95 { X86::NEG32r, X86::NEG32m },
96 { X86::NEG64r, X86::NEG64m },
97 { X86::NEG8r, X86::NEG8m },
98 { X86::NOT16r, X86::NOT16m },
99 { X86::NOT32r, X86::NOT32m },
100 { X86::NOT64r, X86::NOT64m },
101 { X86::NOT8r, X86::NOT8m },
102 { X86::OR16ri, X86::OR16mi },
103 { X86::OR16ri8, X86::OR16mi8 },
104 { X86::OR16rr, X86::OR16mr },
105 { X86::OR32ri, X86::OR32mi },
106 { X86::OR32ri8, X86::OR32mi8 },
107 { X86::OR32rr, X86::OR32mr },
108 { X86::OR64ri32, X86::OR64mi32 },
109 { X86::OR64ri8, X86::OR64mi8 },
110 { X86::OR64rr, X86::OR64mr },
111 { X86::OR8ri, X86::OR8mi },
112 { X86::OR8rr, X86::OR8mr },
113 { X86::ROL16r1, X86::ROL16m1 },
114 { X86::ROL16rCL, X86::ROL16mCL },
115 { X86::ROL16ri, X86::ROL16mi },
116 { X86::ROL32r1, X86::ROL32m1 },
117 { X86::ROL32rCL, X86::ROL32mCL },
118 { X86::ROL32ri, X86::ROL32mi },
119 { X86::ROL64r1, X86::ROL64m1 },
120 { X86::ROL64rCL, X86::ROL64mCL },
121 { X86::ROL64ri, X86::ROL64mi },
122 { X86::ROL8r1, X86::ROL8m1 },
123 { X86::ROL8rCL, X86::ROL8mCL },
124 { X86::ROL8ri, X86::ROL8mi },
125 { X86::ROR16r1, X86::ROR16m1 },
126 { X86::ROR16rCL, X86::ROR16mCL },
127 { X86::ROR16ri, X86::ROR16mi },
128 { X86::ROR32r1, X86::ROR32m1 },
129 { X86::ROR32rCL, X86::ROR32mCL },
130 { X86::ROR32ri, X86::ROR32mi },
131 { X86::ROR64r1, X86::ROR64m1 },
132 { X86::ROR64rCL, X86::ROR64mCL },
133 { X86::ROR64ri, X86::ROR64mi },
134 { X86::ROR8r1, X86::ROR8m1 },
135 { X86::ROR8rCL, X86::ROR8mCL },
136 { X86::ROR8ri, X86::ROR8mi },
137 { X86::SAR16r1, X86::SAR16m1 },
138 { X86::SAR16rCL, X86::SAR16mCL },
139 { X86::SAR16ri, X86::SAR16mi },
140 { X86::SAR32r1, X86::SAR32m1 },
141 { X86::SAR32rCL, X86::SAR32mCL },
142 { X86::SAR32ri, X86::SAR32mi },
143 { X86::SAR64r1, X86::SAR64m1 },
144 { X86::SAR64rCL, X86::SAR64mCL },
145 { X86::SAR64ri, X86::SAR64mi },
146 { X86::SAR8r1, X86::SAR8m1 },
147 { X86::SAR8rCL, X86::SAR8mCL },
148 { X86::SAR8ri, X86::SAR8mi },
149 { X86::SBB32ri, X86::SBB32mi },
150 { X86::SBB32ri8, X86::SBB32mi8 },
151 { X86::SBB32rr, X86::SBB32mr },
152 { X86::SBB64ri32, X86::SBB64mi32 },
153 { X86::SBB64ri8, X86::SBB64mi8 },
154 { X86::SBB64rr, X86::SBB64mr },
155 { X86::SHL16rCL, X86::SHL16mCL },
156 { X86::SHL16ri, X86::SHL16mi },
157 { X86::SHL32rCL, X86::SHL32mCL },
158 { X86::SHL32ri, X86::SHL32mi },
159 { X86::SHL64rCL, X86::SHL64mCL },
160 { X86::SHL64ri, X86::SHL64mi },
161 { X86::SHL8rCL, X86::SHL8mCL },
162 { X86::SHL8ri, X86::SHL8mi },
163 { X86::SHLD16rrCL, X86::SHLD16mrCL },
164 { X86::SHLD16rri8, X86::SHLD16mri8 },
165 { X86::SHLD32rrCL, X86::SHLD32mrCL },
166 { X86::SHLD32rri8, X86::SHLD32mri8 },
167 { X86::SHLD64rrCL, X86::SHLD64mrCL },
168 { X86::SHLD64rri8, X86::SHLD64mri8 },
169 { X86::SHR16r1, X86::SHR16m1 },
170 { X86::SHR16rCL, X86::SHR16mCL },
171 { X86::SHR16ri, X86::SHR16mi },
172 { X86::SHR32r1, X86::SHR32m1 },
173 { X86::SHR32rCL, X86::SHR32mCL },
174 { X86::SHR32ri, X86::SHR32mi },
175 { X86::SHR64r1, X86::SHR64m1 },
176 { X86::SHR64rCL, X86::SHR64mCL },
177 { X86::SHR64ri, X86::SHR64mi },
178 { X86::SHR8r1, X86::SHR8m1 },
179 { X86::SHR8rCL, X86::SHR8mCL },
180 { X86::SHR8ri, X86::SHR8mi },
181 { X86::SHRD16rrCL, X86::SHRD16mrCL },
182 { X86::SHRD16rri8, X86::SHRD16mri8 },
183 { X86::SHRD32rrCL, X86::SHRD32mrCL },
184 { X86::SHRD32rri8, X86::SHRD32mri8 },
185 { X86::SHRD64rrCL, X86::SHRD64mrCL },
186 { X86::SHRD64rri8, X86::SHRD64mri8 },
187 { X86::SUB16ri, X86::SUB16mi },
188 { X86::SUB16ri8, X86::SUB16mi8 },
189 { X86::SUB16rr, X86::SUB16mr },
190 { X86::SUB32ri, X86::SUB32mi },
191 { X86::SUB32ri8, X86::SUB32mi8 },
192 { X86::SUB32rr, X86::SUB32mr },
193 { X86::SUB64ri32, X86::SUB64mi32 },
194 { X86::SUB64ri8, X86::SUB64mi8 },
195 { X86::SUB64rr, X86::SUB64mr },
196 { X86::SUB8ri, X86::SUB8mi },
197 { X86::SUB8rr, X86::SUB8mr },
198 { X86::XOR16ri, X86::XOR16mi },
199 { X86::XOR16ri8, X86::XOR16mi8 },
200 { X86::XOR16rr, X86::XOR16mr },
201 { X86::XOR32ri, X86::XOR32mi },
202 { X86::XOR32ri8, X86::XOR32mi8 },
203 { X86::XOR32rr, X86::XOR32mr },
204 { X86::XOR64ri32, X86::XOR64mi32 },
205 { X86::XOR64ri8, X86::XOR64mi8 },
206 { X86::XOR64rr, X86::XOR64mr },
207 { X86::XOR8ri, X86::XOR8mi },
208 { X86::XOR8rr, X86::XOR8mr }
211 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
212 unsigned RegOp = OpTbl2Addr[i][0];
213 unsigned MemOp = OpTbl2Addr[i][1];
214 if (!RegOp2MemOpTable2Addr.insert(std::make_pair((unsigned*)RegOp,
216 assert(false && "Duplicated entries?");
217 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5); // Index 0,folded load and store
218 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
219 std::make_pair(RegOp,
221 AmbEntries.push_back(MemOp);
224 // If the third value is 1, then it's folding either a load or a store.
225 static const unsigned OpTbl0[][3] = {
226 { X86::BT16ri8, X86::BT16mi8, 1 },
227 { X86::BT32ri8, X86::BT32mi8, 1 },
228 { X86::BT64ri8, X86::BT64mi8, 1 },
229 { X86::CALL32r, X86::CALL32m, 1 },
230 { X86::CALL64r, X86::CALL64m, 1 },
231 { X86::CMP16ri, X86::CMP16mi, 1 },
232 { X86::CMP16ri8, X86::CMP16mi8, 1 },
233 { X86::CMP16rr, X86::CMP16mr, 1 },
234 { X86::CMP32ri, X86::CMP32mi, 1 },
235 { X86::CMP32ri8, X86::CMP32mi8, 1 },
236 { X86::CMP32rr, X86::CMP32mr, 1 },
237 { X86::CMP64ri32, X86::CMP64mi32, 1 },
238 { X86::CMP64ri8, X86::CMP64mi8, 1 },
239 { X86::CMP64rr, X86::CMP64mr, 1 },
240 { X86::CMP8ri, X86::CMP8mi, 1 },
241 { X86::CMP8rr, X86::CMP8mr, 1 },
242 { X86::DIV16r, X86::DIV16m, 1 },
243 { X86::DIV32r, X86::DIV32m, 1 },
244 { X86::DIV64r, X86::DIV64m, 1 },
245 { X86::DIV8r, X86::DIV8m, 1 },
246 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0 },
247 { X86::FsMOVAPDrr, X86::MOVSDmr, 0 },
248 { X86::FsMOVAPSrr, X86::MOVSSmr, 0 },
249 { X86::IDIV16r, X86::IDIV16m, 1 },
250 { X86::IDIV32r, X86::IDIV32m, 1 },
251 { X86::IDIV64r, X86::IDIV64m, 1 },
252 { X86::IDIV8r, X86::IDIV8m, 1 },
253 { X86::IMUL16r, X86::IMUL16m, 1 },
254 { X86::IMUL32r, X86::IMUL32m, 1 },
255 { X86::IMUL64r, X86::IMUL64m, 1 },
256 { X86::IMUL8r, X86::IMUL8m, 1 },
257 { X86::JMP32r, X86::JMP32m, 1 },
258 { X86::JMP64r, X86::JMP64m, 1 },
259 { X86::MOV16ri, X86::MOV16mi, 0 },
260 { X86::MOV16rr, X86::MOV16mr, 0 },
261 { X86::MOV16to16_, X86::MOV16_mr, 0 },
262 { X86::MOV32ri, X86::MOV32mi, 0 },
263 { X86::MOV32rr, X86::MOV32mr, 0 },
264 { X86::MOV32to32_, X86::MOV32_mr, 0 },
265 { X86::MOV64ri32, X86::MOV64mi32, 0 },
266 { X86::MOV64rr, X86::MOV64mr, 0 },
267 { X86::MOV8ri, X86::MOV8mi, 0 },
268 { X86::MOV8rr, X86::MOV8mr, 0 },
269 { X86::MOVAPDrr, X86::MOVAPDmr, 0 },
270 { X86::MOVAPSrr, X86::MOVAPSmr, 0 },
271 { X86::MOVDQArr, X86::MOVDQAmr, 0 },
272 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0 },
273 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0 },
274 { X86::MOVPS2SSrr, X86::MOVPS2SSmr, 0 },
275 { X86::MOVSDrr, X86::MOVSDmr, 0 },
276 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0 },
277 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0 },
278 { X86::MOVSSrr, X86::MOVSSmr, 0 },
279 { X86::MOVUPDrr, X86::MOVUPDmr, 0 },
280 { X86::MOVUPSrr, X86::MOVUPSmr, 0 },
281 { X86::MUL16r, X86::MUL16m, 1 },
282 { X86::MUL32r, X86::MUL32m, 1 },
283 { X86::MUL64r, X86::MUL64m, 1 },
284 { X86::MUL8r, X86::MUL8m, 1 },
285 { X86::SETAEr, X86::SETAEm, 0 },
286 { X86::SETAr, X86::SETAm, 0 },
287 { X86::SETBEr, X86::SETBEm, 0 },
288 { X86::SETBr, X86::SETBm, 0 },
289 { X86::SETEr, X86::SETEm, 0 },
290 { X86::SETGEr, X86::SETGEm, 0 },
291 { X86::SETGr, X86::SETGm, 0 },
292 { X86::SETLEr, X86::SETLEm, 0 },
293 { X86::SETLr, X86::SETLm, 0 },
294 { X86::SETNEr, X86::SETNEm, 0 },
295 { X86::SETNOr, X86::SETNOm, 0 },
296 { X86::SETNPr, X86::SETNPm, 0 },
297 { X86::SETNSr, X86::SETNSm, 0 },
298 { X86::SETOr, X86::SETOm, 0 },
299 { X86::SETPr, X86::SETPm, 0 },
300 { X86::SETSr, X86::SETSm, 0 },
301 { X86::TAILJMPr, X86::TAILJMPm, 1 },
302 { X86::TEST16ri, X86::TEST16mi, 1 },
303 { X86::TEST32ri, X86::TEST32mi, 1 },
304 { X86::TEST64ri32, X86::TEST64mi32, 1 },
305 { X86::TEST8ri, X86::TEST8mi, 1 }
308 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
309 unsigned RegOp = OpTbl0[i][0];
310 unsigned MemOp = OpTbl0[i][1];
311 if (!RegOp2MemOpTable0.insert(std::make_pair((unsigned*)RegOp,
313 assert(false && "Duplicated entries?");
314 unsigned FoldedLoad = OpTbl0[i][2];
315 // Index 0, folded load or store.
316 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
317 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
318 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
319 std::make_pair(RegOp, AuxInfo))).second)
320 AmbEntries.push_back(MemOp);
323 static const unsigned OpTbl1[][2] = {
324 { X86::CMP16rr, X86::CMP16rm },
325 { X86::CMP32rr, X86::CMP32rm },
326 { X86::CMP64rr, X86::CMP64rm },
327 { X86::CMP8rr, X86::CMP8rm },
328 { X86::CVTSD2SSrr, X86::CVTSD2SSrm },
329 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm },
330 { X86::CVTSI2SDrr, X86::CVTSI2SDrm },
331 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm },
332 { X86::CVTSI2SSrr, X86::CVTSI2SSrm },
333 { X86::CVTSS2SDrr, X86::CVTSS2SDrm },
334 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm },
335 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm },
336 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm },
337 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm },
338 { X86::FsMOVAPDrr, X86::MOVSDrm },
339 { X86::FsMOVAPSrr, X86::MOVSSrm },
340 { X86::IMUL16rri, X86::IMUL16rmi },
341 { X86::IMUL16rri8, X86::IMUL16rmi8 },
342 { X86::IMUL32rri, X86::IMUL32rmi },
343 { X86::IMUL32rri8, X86::IMUL32rmi8 },
344 { X86::IMUL64rri32, X86::IMUL64rmi32 },
345 { X86::IMUL64rri8, X86::IMUL64rmi8 },
346 { X86::Int_CMPSDrr, X86::Int_CMPSDrm },
347 { X86::Int_CMPSSrr, X86::Int_CMPSSrm },
348 { X86::Int_COMISDrr, X86::Int_COMISDrm },
349 { X86::Int_COMISSrr, X86::Int_COMISSrm },
350 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm },
351 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm },
352 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm },
353 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm },
354 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm },
355 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm },
356 { X86::Int_CVTSD2SI64rr,X86::Int_CVTSD2SI64rm },
357 { X86::Int_CVTSD2SIrr, X86::Int_CVTSD2SIrm },
358 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm },
359 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm },
360 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm },
361 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm },
362 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm },
363 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm },
364 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm },
365 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm },
366 { X86::Int_CVTTPD2DQrr, X86::Int_CVTTPD2DQrm },
367 { X86::Int_CVTTPS2DQrr, X86::Int_CVTTPS2DQrm },
368 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm },
369 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm },
370 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm },
371 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm },
372 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm },
373 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm },
374 { X86::MOV16rr, X86::MOV16rm },
375 { X86::MOV16to16_, X86::MOV16_rm },
376 { X86::MOV32rr, X86::MOV32rm },
377 { X86::MOV32to32_, X86::MOV32_rm },
378 { X86::MOV64rr, X86::MOV64rm },
379 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm },
380 { X86::MOV64toSDrr, X86::MOV64toSDrm },
381 { X86::MOV8rr, X86::MOV8rm },
382 { X86::MOVAPDrr, X86::MOVAPDrm },
383 { X86::MOVAPSrr, X86::MOVAPSrm },
384 { X86::MOVDDUPrr, X86::MOVDDUPrm },
385 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm },
386 { X86::MOVDI2SSrr, X86::MOVDI2SSrm },
387 { X86::MOVDQArr, X86::MOVDQArm },
388 { X86::MOVSD2PDrr, X86::MOVSD2PDrm },
389 { X86::MOVSDrr, X86::MOVSDrm },
390 { X86::MOVSHDUPrr, X86::MOVSHDUPrm },
391 { X86::MOVSLDUPrr, X86::MOVSLDUPrm },
392 { X86::MOVSS2PSrr, X86::MOVSS2PSrm },
393 { X86::MOVSSrr, X86::MOVSSrm },
394 { X86::MOVSX16rr8, X86::MOVSX16rm8 },
395 { X86::MOVSX32rr16, X86::MOVSX32rm16 },
396 { X86::MOVSX32rr8, X86::MOVSX32rm8 },
397 { X86::MOVSX64rr16, X86::MOVSX64rm16 },
398 { X86::MOVSX64rr32, X86::MOVSX64rm32 },
399 { X86::MOVSX64rr8, X86::MOVSX64rm8 },
400 { X86::MOVUPDrr, X86::MOVUPDrm },
401 { X86::MOVUPSrr, X86::MOVUPSrm },
402 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm },
403 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm },
404 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm },
405 { X86::MOVZX16rr8, X86::MOVZX16rm8 },
406 { X86::MOVZX32rr16, X86::MOVZX32rm16 },
407 { X86::MOVZX32rr8, X86::MOVZX32rm8 },
408 { X86::MOVZX64rr16, X86::MOVZX64rm16 },
409 { X86::MOVZX64rr32, X86::MOVZX64rm32 },
410 { X86::MOVZX64rr8, X86::MOVZX64rm8 },
411 { X86::PSHUFDri, X86::PSHUFDmi },
412 { X86::PSHUFHWri, X86::PSHUFHWmi },
413 { X86::PSHUFLWri, X86::PSHUFLWmi },
414 { X86::RCPPSr, X86::RCPPSm },
415 { X86::RCPPSr_Int, X86::RCPPSm_Int },
416 { X86::RSQRTPSr, X86::RSQRTPSm },
417 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int },
418 { X86::RSQRTSSr, X86::RSQRTSSm },
419 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int },
420 { X86::SQRTPDr, X86::SQRTPDm },
421 { X86::SQRTPDr_Int, X86::SQRTPDm_Int },
422 { X86::SQRTPSr, X86::SQRTPSm },
423 { X86::SQRTPSr_Int, X86::SQRTPSm_Int },
424 { X86::SQRTSDr, X86::SQRTSDm },
425 { X86::SQRTSDr_Int, X86::SQRTSDm_Int },
426 { X86::SQRTSSr, X86::SQRTSSm },
427 { X86::SQRTSSr_Int, X86::SQRTSSm_Int },
428 { X86::TEST16rr, X86::TEST16rm },
429 { X86::TEST32rr, X86::TEST32rm },
430 { X86::TEST64rr, X86::TEST64rm },
431 { X86::TEST8rr, X86::TEST8rm },
432 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
433 { X86::UCOMISDrr, X86::UCOMISDrm },
434 { X86::UCOMISSrr, X86::UCOMISSrm }
437 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
438 unsigned RegOp = OpTbl1[i][0];
439 unsigned MemOp = OpTbl1[i][1];
440 if (!RegOp2MemOpTable1.insert(std::make_pair((unsigned*)RegOp,
442 assert(false && "Duplicated entries?");
443 unsigned AuxInfo = 1 | (1 << 4); // Index 1, folded load
444 if (RegOp != X86::FsMOVAPDrr && RegOp != X86::FsMOVAPSrr)
445 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
446 std::make_pair(RegOp, AuxInfo))).second)
447 AmbEntries.push_back(MemOp);
450 static const unsigned OpTbl2[][2] = {
451 { X86::ADC32rr, X86::ADC32rm },
452 { X86::ADC64rr, X86::ADC64rm },
453 { X86::ADD16rr, X86::ADD16rm },
454 { X86::ADD32rr, X86::ADD32rm },
455 { X86::ADD64rr, X86::ADD64rm },
456 { X86::ADD8rr, X86::ADD8rm },
457 { X86::ADDPDrr, X86::ADDPDrm },
458 { X86::ADDPSrr, X86::ADDPSrm },
459 { X86::ADDSDrr, X86::ADDSDrm },
460 { X86::ADDSSrr, X86::ADDSSrm },
461 { X86::ADDSUBPDrr, X86::ADDSUBPDrm },
462 { X86::ADDSUBPSrr, X86::ADDSUBPSrm },
463 { X86::AND16rr, X86::AND16rm },
464 { X86::AND32rr, X86::AND32rm },
465 { X86::AND64rr, X86::AND64rm },
466 { X86::AND8rr, X86::AND8rm },
467 { X86::ANDNPDrr, X86::ANDNPDrm },
468 { X86::ANDNPSrr, X86::ANDNPSrm },
469 { X86::ANDPDrr, X86::ANDPDrm },
470 { X86::ANDPSrr, X86::ANDPSrm },
471 { X86::CMOVA16rr, X86::CMOVA16rm },
472 { X86::CMOVA32rr, X86::CMOVA32rm },
473 { X86::CMOVA64rr, X86::CMOVA64rm },
474 { X86::CMOVAE16rr, X86::CMOVAE16rm },
475 { X86::CMOVAE32rr, X86::CMOVAE32rm },
476 { X86::CMOVAE64rr, X86::CMOVAE64rm },
477 { X86::CMOVB16rr, X86::CMOVB16rm },
478 { X86::CMOVB32rr, X86::CMOVB32rm },
479 { X86::CMOVB64rr, X86::CMOVB64rm },
480 { X86::CMOVBE16rr, X86::CMOVBE16rm },
481 { X86::CMOVBE32rr, X86::CMOVBE32rm },
482 { X86::CMOVBE64rr, X86::CMOVBE64rm },
483 { X86::CMOVE16rr, X86::CMOVE16rm },
484 { X86::CMOVE32rr, X86::CMOVE32rm },
485 { X86::CMOVE64rr, X86::CMOVE64rm },
486 { X86::CMOVG16rr, X86::CMOVG16rm },
487 { X86::CMOVG32rr, X86::CMOVG32rm },
488 { X86::CMOVG64rr, X86::CMOVG64rm },
489 { X86::CMOVGE16rr, X86::CMOVGE16rm },
490 { X86::CMOVGE32rr, X86::CMOVGE32rm },
491 { X86::CMOVGE64rr, X86::CMOVGE64rm },
492 { X86::CMOVL16rr, X86::CMOVL16rm },
493 { X86::CMOVL32rr, X86::CMOVL32rm },
494 { X86::CMOVL64rr, X86::CMOVL64rm },
495 { X86::CMOVLE16rr, X86::CMOVLE16rm },
496 { X86::CMOVLE32rr, X86::CMOVLE32rm },
497 { X86::CMOVLE64rr, X86::CMOVLE64rm },
498 { X86::CMOVNE16rr, X86::CMOVNE16rm },
499 { X86::CMOVNE32rr, X86::CMOVNE32rm },
500 { X86::CMOVNE64rr, X86::CMOVNE64rm },
501 { X86::CMOVNO16rr, X86::CMOVNO16rm },
502 { X86::CMOVNO32rr, X86::CMOVNO32rm },
503 { X86::CMOVNO64rr, X86::CMOVNO64rm },
504 { X86::CMOVNP16rr, X86::CMOVNP16rm },
505 { X86::CMOVNP32rr, X86::CMOVNP32rm },
506 { X86::CMOVNP64rr, X86::CMOVNP64rm },
507 { X86::CMOVNS16rr, X86::CMOVNS16rm },
508 { X86::CMOVNS32rr, X86::CMOVNS32rm },
509 { X86::CMOVNS64rr, X86::CMOVNS64rm },
510 { X86::CMOVO16rr, X86::CMOVO16rm },
511 { X86::CMOVO32rr, X86::CMOVO32rm },
512 { X86::CMOVO64rr, X86::CMOVO64rm },
513 { X86::CMOVP16rr, X86::CMOVP16rm },
514 { X86::CMOVP32rr, X86::CMOVP32rm },
515 { X86::CMOVP64rr, X86::CMOVP64rm },
516 { X86::CMOVS16rr, X86::CMOVS16rm },
517 { X86::CMOVS32rr, X86::CMOVS32rm },
518 { X86::CMOVS64rr, X86::CMOVS64rm },
519 { X86::CMPPDrri, X86::CMPPDrmi },
520 { X86::CMPPSrri, X86::CMPPSrmi },
521 { X86::CMPSDrr, X86::CMPSDrm },
522 { X86::CMPSSrr, X86::CMPSSrm },
523 { X86::DIVPDrr, X86::DIVPDrm },
524 { X86::DIVPSrr, X86::DIVPSrm },
525 { X86::DIVSDrr, X86::DIVSDrm },
526 { X86::DIVSSrr, X86::DIVSSrm },
527 { X86::FsANDNPDrr, X86::FsANDNPDrm },
528 { X86::FsANDNPSrr, X86::FsANDNPSrm },
529 { X86::FsANDPDrr, X86::FsANDPDrm },
530 { X86::FsANDPSrr, X86::FsANDPSrm },
531 { X86::FsORPDrr, X86::FsORPDrm },
532 { X86::FsORPSrr, X86::FsORPSrm },
533 { X86::FsXORPDrr, X86::FsXORPDrm },
534 { X86::FsXORPSrr, X86::FsXORPSrm },
535 { X86::HADDPDrr, X86::HADDPDrm },
536 { X86::HADDPSrr, X86::HADDPSrm },
537 { X86::HSUBPDrr, X86::HSUBPDrm },
538 { X86::HSUBPSrr, X86::HSUBPSrm },
539 { X86::IMUL16rr, X86::IMUL16rm },
540 { X86::IMUL32rr, X86::IMUL32rm },
541 { X86::IMUL64rr, X86::IMUL64rm },
542 { X86::MAXPDrr, X86::MAXPDrm },
543 { X86::MAXPDrr_Int, X86::MAXPDrm_Int },
544 { X86::MAXPSrr, X86::MAXPSrm },
545 { X86::MAXPSrr_Int, X86::MAXPSrm_Int },
546 { X86::MAXSDrr, X86::MAXSDrm },
547 { X86::MAXSDrr_Int, X86::MAXSDrm_Int },
548 { X86::MAXSSrr, X86::MAXSSrm },
549 { X86::MAXSSrr_Int, X86::MAXSSrm_Int },
550 { X86::MINPDrr, X86::MINPDrm },
551 { X86::MINPDrr_Int, X86::MINPDrm_Int },
552 { X86::MINPSrr, X86::MINPSrm },
553 { X86::MINPSrr_Int, X86::MINPSrm_Int },
554 { X86::MINSDrr, X86::MINSDrm },
555 { X86::MINSDrr_Int, X86::MINSDrm_Int },
556 { X86::MINSSrr, X86::MINSSrm },
557 { X86::MINSSrr_Int, X86::MINSSrm_Int },
558 { X86::MULPDrr, X86::MULPDrm },
559 { X86::MULPSrr, X86::MULPSrm },
560 { X86::MULSDrr, X86::MULSDrm },
561 { X86::MULSSrr, X86::MULSSrm },
562 { X86::OR16rr, X86::OR16rm },
563 { X86::OR32rr, X86::OR32rm },
564 { X86::OR64rr, X86::OR64rm },
565 { X86::OR8rr, X86::OR8rm },
566 { X86::ORPDrr, X86::ORPDrm },
567 { X86::ORPSrr, X86::ORPSrm },
568 { X86::PACKSSDWrr, X86::PACKSSDWrm },
569 { X86::PACKSSWBrr, X86::PACKSSWBrm },
570 { X86::PACKUSWBrr, X86::PACKUSWBrm },
571 { X86::PADDBrr, X86::PADDBrm },
572 { X86::PADDDrr, X86::PADDDrm },
573 { X86::PADDQrr, X86::PADDQrm },
574 { X86::PADDSBrr, X86::PADDSBrm },
575 { X86::PADDSWrr, X86::PADDSWrm },
576 { X86::PADDWrr, X86::PADDWrm },
577 { X86::PANDNrr, X86::PANDNrm },
578 { X86::PANDrr, X86::PANDrm },
579 { X86::PAVGBrr, X86::PAVGBrm },
580 { X86::PAVGWrr, X86::PAVGWrm },
581 { X86::PCMPEQBrr, X86::PCMPEQBrm },
582 { X86::PCMPEQDrr, X86::PCMPEQDrm },
583 { X86::PCMPEQWrr, X86::PCMPEQWrm },
584 { X86::PCMPGTBrr, X86::PCMPGTBrm },
585 { X86::PCMPGTDrr, X86::PCMPGTDrm },
586 { X86::PCMPGTWrr, X86::PCMPGTWrm },
587 { X86::PINSRWrri, X86::PINSRWrmi },
588 { X86::PMADDWDrr, X86::PMADDWDrm },
589 { X86::PMAXSWrr, X86::PMAXSWrm },
590 { X86::PMAXUBrr, X86::PMAXUBrm },
591 { X86::PMINSWrr, X86::PMINSWrm },
592 { X86::PMINUBrr, X86::PMINUBrm },
593 { X86::PMULDQrr, X86::PMULDQrm },
594 { X86::PMULHUWrr, X86::PMULHUWrm },
595 { X86::PMULHWrr, X86::PMULHWrm },
596 { X86::PMULLDrr, X86::PMULLDrm },
597 { X86::PMULLDrr_int, X86::PMULLDrm_int },
598 { X86::PMULLWrr, X86::PMULLWrm },
599 { X86::PMULUDQrr, X86::PMULUDQrm },
600 { X86::PORrr, X86::PORrm },
601 { X86::PSADBWrr, X86::PSADBWrm },
602 { X86::PSLLDrr, X86::PSLLDrm },
603 { X86::PSLLQrr, X86::PSLLQrm },
604 { X86::PSLLWrr, X86::PSLLWrm },
605 { X86::PSRADrr, X86::PSRADrm },
606 { X86::PSRAWrr, X86::PSRAWrm },
607 { X86::PSRLDrr, X86::PSRLDrm },
608 { X86::PSRLQrr, X86::PSRLQrm },
609 { X86::PSRLWrr, X86::PSRLWrm },
610 { X86::PSUBBrr, X86::PSUBBrm },
611 { X86::PSUBDrr, X86::PSUBDrm },
612 { X86::PSUBSBrr, X86::PSUBSBrm },
613 { X86::PSUBSWrr, X86::PSUBSWrm },
614 { X86::PSUBWrr, X86::PSUBWrm },
615 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm },
616 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm },
617 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm },
618 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm },
619 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm },
620 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm },
621 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm },
622 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm },
623 { X86::PXORrr, X86::PXORrm },
624 { X86::SBB32rr, X86::SBB32rm },
625 { X86::SBB64rr, X86::SBB64rm },
626 { X86::SHUFPDrri, X86::SHUFPDrmi },
627 { X86::SHUFPSrri, X86::SHUFPSrmi },
628 { X86::SUB16rr, X86::SUB16rm },
629 { X86::SUB32rr, X86::SUB32rm },
630 { X86::SUB64rr, X86::SUB64rm },
631 { X86::SUB8rr, X86::SUB8rm },
632 { X86::SUBPDrr, X86::SUBPDrm },
633 { X86::SUBPSrr, X86::SUBPSrm },
634 { X86::SUBSDrr, X86::SUBSDrm },
635 { X86::SUBSSrr, X86::SUBSSrm },
636 // FIXME: TEST*rr -> swapped operand of TEST*mr.
637 { X86::UNPCKHPDrr, X86::UNPCKHPDrm },
638 { X86::UNPCKHPSrr, X86::UNPCKHPSrm },
639 { X86::UNPCKLPDrr, X86::UNPCKLPDrm },
640 { X86::UNPCKLPSrr, X86::UNPCKLPSrm },
641 { X86::XOR16rr, X86::XOR16rm },
642 { X86::XOR32rr, X86::XOR32rm },
643 { X86::XOR64rr, X86::XOR64rm },
644 { X86::XOR8rr, X86::XOR8rm },
645 { X86::XORPDrr, X86::XORPDrm },
646 { X86::XORPSrr, X86::XORPSrm }
649 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
650 unsigned RegOp = OpTbl2[i][0];
651 unsigned MemOp = OpTbl2[i][1];
652 if (!RegOp2MemOpTable2.insert(std::make_pair((unsigned*)RegOp,
654 assert(false && "Duplicated entries?");
655 unsigned AuxInfo = 2 | (1 << 4); // Index 1, folded load
656 if (!MemOp2RegOpTable.insert(std::make_pair((unsigned*)MemOp,
657 std::make_pair(RegOp, AuxInfo))).second)
658 AmbEntries.push_back(MemOp);
661 // Remove ambiguous entries.
662 assert(AmbEntries.empty() && "Duplicated entries in unfolding maps?");
665 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
666 unsigned &SrcReg, unsigned &DstReg,
667 unsigned &SrcSubIdx, unsigned &DstSubIdx) const {
668 switch (MI.getOpcode()) {
675 case X86::MOV16to16_:
676 case X86::MOV32to32_:
680 // FP Stack register class copies
681 case X86::MOV_Fp3232: case X86::MOV_Fp6464: case X86::MOV_Fp8080:
682 case X86::MOV_Fp3264: case X86::MOV_Fp3280:
683 case X86::MOV_Fp6432: case X86::MOV_Fp8032:
685 case X86::FsMOVAPSrr:
686 case X86::FsMOVAPDrr:
690 case X86::MOVSS2PSrr:
691 case X86::MOVSD2PDrr:
692 case X86::MOVPS2SSrr:
693 case X86::MOVPD2SDrr:
694 case X86::MMX_MOVD64rr:
695 case X86::MMX_MOVQ64rr:
696 assert(MI.getNumOperands() >= 2 &&
697 MI.getOperand(0).isReg() &&
698 MI.getOperand(1).isReg() &&
699 "invalid register-register move instruction");
700 SrcReg = MI.getOperand(1).getReg();
701 DstReg = MI.getOperand(0).getReg();
702 SrcSubIdx = MI.getOperand(1).getSubReg();
703 DstSubIdx = MI.getOperand(0).getSubReg();
708 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
709 int &FrameIndex) const {
710 switch (MI->getOpcode()) {
724 case X86::MMX_MOVD64rm:
725 case X86::MMX_MOVQ64rm:
726 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
727 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
728 MI->getOperand(2).getImm() == 1 &&
729 MI->getOperand(3).getReg() == 0 &&
730 MI->getOperand(4).getImm() == 0) {
731 FrameIndex = MI->getOperand(1).getIndex();
732 return MI->getOperand(0).getReg();
739 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
740 int &FrameIndex) const {
741 switch (MI->getOpcode()) {
755 case X86::MMX_MOVD64mr:
756 case X86::MMX_MOVQ64mr:
757 case X86::MMX_MOVNTQmr:
758 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
759 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
760 MI->getOperand(1).getImm() == 1 &&
761 MI->getOperand(2).getReg() == 0 &&
762 MI->getOperand(3).getImm() == 0) {
763 FrameIndex = MI->getOperand(0).getIndex();
764 return MI->getOperand(4).getReg();
772 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
774 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
775 bool isPICBase = false;
776 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
777 E = MRI.def_end(); I != E; ++I) {
778 MachineInstr *DefMI = I.getOperand().getParent();
779 if (DefMI->getOpcode() != X86::MOVPC32r)
781 assert(!isPICBase && "More than one PIC base?");
787 /// isGVStub - Return true if the GV requires an extra load to get the
789 static inline bool isGVStub(GlobalValue *GV, X86TargetMachine &TM) {
790 return TM.getSubtarget<X86Subtarget>().GVRequiresExtraLoad(GV, TM, false);
794 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI) const {
795 switch (MI->getOpcode()) {
809 case X86::MMX_MOVD64rm:
810 case X86::MMX_MOVQ64rm: {
811 // Loads from constant pools are trivially rematerializable.
812 if (MI->getOperand(1).isReg() &&
813 MI->getOperand(2).isImm() &&
814 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
815 (MI->getOperand(4).isCPI() ||
816 (MI->getOperand(4).isGlobal() &&
817 isGVStub(MI->getOperand(4).getGlobal(), TM)))) {
818 unsigned BaseReg = MI->getOperand(1).getReg();
821 // Allow re-materialization of PIC load.
822 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
824 const MachineFunction &MF = *MI->getParent()->getParent();
825 const MachineRegisterInfo &MRI = MF.getRegInfo();
826 bool isPICBase = false;
827 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
828 E = MRI.def_end(); I != E; ++I) {
829 MachineInstr *DefMI = I.getOperand().getParent();
830 if (DefMI->getOpcode() != X86::MOVPC32r)
832 assert(!isPICBase && "More than one PIC base?");
842 if (MI->getOperand(2).isImm() &&
843 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
844 !MI->getOperand(4).isReg()) {
845 // lea fi#, lea GV, etc. are all rematerializable.
846 if (!MI->getOperand(1).isReg())
848 unsigned BaseReg = MI->getOperand(1).getReg();
851 // Allow re-materialization of lea PICBase + x.
852 const MachineFunction &MF = *MI->getParent()->getParent();
853 const MachineRegisterInfo &MRI = MF.getRegInfo();
854 return regIsPICBase(BaseReg, MRI);
860 // All other instructions marked M_REMATERIALIZABLE are always trivially
865 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
866 /// would clobber the EFLAGS condition register. Note the result may be
867 /// conservative. If it cannot definitely determine the safety after visiting
868 /// two instructions it assumes it's not safe.
869 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
870 MachineBasicBlock::iterator I) {
871 // It's always safe to clobber EFLAGS at the end of a block.
875 // For compile time consideration, if we are not able to determine the
876 // safety after visiting 2 instructions, we will assume it's not safe.
877 for (unsigned i = 0; i < 2; ++i) {
878 bool SeenDef = false;
879 for (unsigned j = 0, e = I->getNumOperands(); j != e; ++j) {
880 MachineOperand &MO = I->getOperand(j);
883 if (MO.getReg() == X86::EFLAGS) {
891 // This instruction defines EFLAGS, no need to look any further.
895 // If we make it to the end of the block, it's safe to clobber EFLAGS.
900 // Conservative answer.
904 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
905 MachineBasicBlock::iterator I,
907 const MachineInstr *Orig) const {
908 DebugLoc DL = DebugLoc::getUnknownLoc();
909 if (I != MBB.end()) DL = I->getDebugLoc();
911 unsigned SubIdx = Orig->getOperand(0).isReg()
912 ? Orig->getOperand(0).getSubReg() : 0;
913 bool ChangeSubIdx = SubIdx != 0;
914 if (SubIdx && TargetRegisterInfo::isPhysicalRegister(DestReg)) {
915 DestReg = RI.getSubReg(DestReg, SubIdx);
919 // MOV32r0 etc. are implemented with xor which clobbers condition code.
920 // Re-materialize them as movri instructions to avoid side effects.
921 bool Emitted = false;
922 switch (Orig->getOpcode()) {
928 if (!isSafeToClobberEFLAGS(MBB, I)) {
930 switch (Orig->getOpcode()) {
932 case X86::MOV8r0: Opc = X86::MOV8ri; break;
933 case X86::MOV16r0: Opc = X86::MOV16ri; break;
934 case X86::MOV32r0: Opc = X86::MOV32ri; break;
935 case X86::MOV64r0: Opc = X86::MOV64ri32; break;
937 BuildMI(MBB, I, DL, get(Opc), DestReg).addImm(0);
945 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
946 MI->getOperand(0).setReg(DestReg);
951 MachineInstr *NewMI = prior(I);
952 NewMI->getOperand(0).setSubReg(SubIdx);
956 /// isInvariantLoad - Return true if the specified instruction (which is marked
957 /// mayLoad) is loading from a location whose value is invariant across the
958 /// function. For example, loading a value from the constant pool or from
959 /// from the argument area of a function if it does not change. This should
960 /// only return true of *all* loads the instruction does are invariant (if it
961 /// does multiple loads).
962 bool X86InstrInfo::isInvariantLoad(const MachineInstr *MI) const {
963 // This code cares about loads from three cases: constant pool entries,
964 // invariant argument slots, and global stubs. In order to handle these cases
965 // for all of the myriad of X86 instructions, we just scan for a CP/FI/GV
966 // operand and base our analysis on it. This is safe because the address of
967 // none of these three cases is ever used as anything other than a load base
968 // and X86 doesn't have any instructions that load from multiple places.
970 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
971 const MachineOperand &MO = MI->getOperand(i);
972 // Loads from constant pools are trivially invariant.
977 return isGVStub(MO.getGlobal(), TM);
979 // If this is a load from an invariant stack slot, the load is a constant.
981 const MachineFrameInfo &MFI =
982 *MI->getParent()->getParent()->getFrameInfo();
983 int Idx = MO.getIndex();
984 return MFI.isFixedObjectIndex(Idx) && MFI.isImmutableObjectIndex(Idx);
988 // All other instances of these instructions are presumed to have other
993 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
994 /// is not marked dead.
995 static bool hasLiveCondCodeDef(MachineInstr *MI) {
996 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
997 MachineOperand &MO = MI->getOperand(i);
998 if (MO.isReg() && MO.isDef() &&
999 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1006 /// convertToThreeAddress - This method must be implemented by targets that
1007 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1008 /// may be able to convert a two-address instruction into a true
1009 /// three-address instruction on demand. This allows the X86 target (for
1010 /// example) to convert ADD and SHL instructions into LEA instructions if they
1011 /// would require register copies due to two-addressness.
1013 /// This method returns a null pointer if the transformation cannot be
1014 /// performed, otherwise it returns the new instruction.
1017 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1018 MachineBasicBlock::iterator &MBBI,
1019 LiveVariables *LV) const {
1020 MachineInstr *MI = MBBI;
1021 MachineFunction &MF = *MI->getParent()->getParent();
1022 // All instructions input are two-addr instructions. Get the known operands.
1023 unsigned Dest = MI->getOperand(0).getReg();
1024 unsigned Src = MI->getOperand(1).getReg();
1025 bool isDead = MI->getOperand(0).isDead();
1026 bool isKill = MI->getOperand(1).isKill();
1028 MachineInstr *NewMI = NULL;
1029 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1030 // we have better subtarget support, enable the 16-bit LEA generation here.
1031 bool DisableLEA16 = true;
1033 unsigned MIOpc = MI->getOpcode();
1035 case X86::SHUFPSrri: {
1036 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1037 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1039 unsigned B = MI->getOperand(1).getReg();
1040 unsigned C = MI->getOperand(2).getReg();
1041 if (B != C) return 0;
1042 unsigned A = MI->getOperand(0).getReg();
1043 unsigned M = MI->getOperand(3).getImm();
1044 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1045 .addReg(A, true, false, false, isDead)
1046 .addReg(B, false, false, isKill).addImm(M);
1049 case X86::SHL64ri: {
1050 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1051 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1052 // the flags produced by a shift yet, so this is safe.
1053 unsigned ShAmt = MI->getOperand(2).getImm();
1054 if (ShAmt == 0 || ShAmt >= 4) return 0;
1056 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1057 .addReg(Dest, true, false, false, isDead)
1058 .addReg(0).addImm(1 << ShAmt).addReg(Src, false, false, isKill).addImm(0);
1061 case X86::SHL32ri: {
1062 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1063 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1064 // the flags produced by a shift yet, so this is safe.
1065 unsigned ShAmt = MI->getOperand(2).getImm();
1066 if (ShAmt == 0 || ShAmt >= 4) return 0;
1068 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
1069 X86::LEA64_32r : X86::LEA32r;
1070 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1071 .addReg(Dest, true, false, false, isDead)
1072 .addReg(0).addImm(1 << ShAmt)
1073 .addReg(Src, false, false, isKill).addImm(0);
1076 case X86::SHL16ri: {
1077 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1078 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1079 // the flags produced by a shift yet, so this is safe.
1080 unsigned ShAmt = MI->getOperand(2).getImm();
1081 if (ShAmt == 0 || ShAmt >= 4) return 0;
1084 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
1085 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1086 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1087 ? X86::LEA64_32r : X86::LEA32r;
1088 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1089 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1091 // Build and insert into an implicit UNDEF value. This is OK because
1092 // well be shifting and then extracting the lower 16-bits.
1093 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1094 MachineInstr *InsMI =
1095 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::INSERT_SUBREG),leaInReg)
1096 .addReg(leaInReg).addReg(Src, false, false, isKill)
1097 .addImm(X86::SUBREG_16BIT);
1099 NewMI = BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(Opc), leaOutReg)
1100 .addReg(0).addImm(1 << ShAmt)
1101 .addReg(leaInReg, false, false, true).addImm(0);
1103 MachineInstr *ExtMI =
1104 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::EXTRACT_SUBREG))
1105 .addReg(Dest, true, false, false, isDead)
1106 .addReg(leaOutReg, false, false, true).addImm(X86::SUBREG_16BIT);
1109 // Update live variables
1110 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1111 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1113 LV->replaceKillInstruction(Src, MI, InsMI);
1115 LV->replaceKillInstruction(Dest, MI, ExtMI);
1119 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1120 .addReg(Dest, true, false, false, isDead)
1121 .addReg(0).addImm(1 << ShAmt)
1122 .addReg(Src, false, false, isKill).addImm(0);
1127 // The following opcodes also sets the condition code register(s). Only
1128 // convert them to equivalent lea if the condition code register def's
1130 if (hasLiveCondCodeDef(MI))
1133 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1138 case X86::INC64_32r: {
1139 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1140 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1141 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1142 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1143 .addReg(Dest, true, false, false, isDead),
1148 case X86::INC64_16r:
1149 if (DisableLEA16) return 0;
1150 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1151 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1152 .addReg(Dest, true, false, false, isDead),
1157 case X86::DEC64_32r: {
1158 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1159 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1160 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1161 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1162 .addReg(Dest, true, false, false, isDead),
1167 case X86::DEC64_16r:
1168 if (DisableLEA16) return 0;
1169 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1170 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1171 .addReg(Dest, true, false, false, isDead),
1175 case X86::ADD32rr: {
1176 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1177 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
1178 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1179 unsigned Src2 = MI->getOperand(2).getReg();
1180 bool isKill2 = MI->getOperand(2).isKill();
1181 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1182 .addReg(Dest, true, false, false, isDead),
1183 Src, isKill, Src2, isKill2);
1185 LV->replaceKillInstruction(Src2, MI, NewMI);
1188 case X86::ADD16rr: {
1189 if (DisableLEA16) return 0;
1190 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1191 unsigned Src2 = MI->getOperand(2).getReg();
1192 bool isKill2 = MI->getOperand(2).isKill();
1193 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1194 .addReg(Dest, true, false, false, isDead),
1195 Src, isKill, Src2, isKill2);
1197 LV->replaceKillInstruction(Src2, MI, NewMI);
1200 case X86::ADD64ri32:
1202 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1203 if (MI->getOperand(2).isImm())
1204 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1205 .addReg(Dest, true, false, false, isDead),
1206 Src, isKill, MI->getOperand(2).getImm());
1210 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1211 if (MI->getOperand(2).isImm()) {
1212 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1213 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1214 .addReg(Dest, true, false, false, isDead),
1215 Src, isKill, MI->getOperand(2).getImm());
1220 if (DisableLEA16) return 0;
1221 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1222 if (MI->getOperand(2).isImm())
1223 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1224 .addReg(Dest, true, false, false, isDead),
1225 Src, isKill, MI->getOperand(2).getImm());
1228 if (DisableLEA16) return 0;
1230 case X86::SHL64ri: {
1231 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImm() &&
1232 "Unknown shl instruction!");
1233 unsigned ShAmt = MI->getOperand(2).getImm();
1234 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
1236 AM.Scale = 1 << ShAmt;
1238 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
1239 : (MIOpc == X86::SHL32ri
1240 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
1241 NewMI = addFullAddress(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1242 .addReg(Dest, true, false, false, isDead), AM);
1244 NewMI->getOperand(3).setIsKill(true);
1252 if (!NewMI) return 0;
1254 if (LV) { // Update live variables
1256 LV->replaceKillInstruction(Src, MI, NewMI);
1258 LV->replaceKillInstruction(Dest, MI, NewMI);
1261 MFI->insert(MBBI, NewMI); // Insert the new inst
1265 /// commuteInstruction - We have a few instructions that must be hacked on to
1269 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1270 switch (MI->getOpcode()) {
1271 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1272 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1273 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1274 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1275 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1276 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1279 switch (MI->getOpcode()) {
1280 default: assert(0 && "Unreachable!");
1281 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1282 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1283 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1284 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1285 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1286 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1288 unsigned Amt = MI->getOperand(3).getImm();
1290 MachineFunction &MF = *MI->getParent()->getParent();
1291 MI = MF.CloneMachineInstr(MI);
1294 MI->setDesc(get(Opc));
1295 MI->getOperand(3).setImm(Size-Amt);
1296 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1298 case X86::CMOVB16rr:
1299 case X86::CMOVB32rr:
1300 case X86::CMOVB64rr:
1301 case X86::CMOVAE16rr:
1302 case X86::CMOVAE32rr:
1303 case X86::CMOVAE64rr:
1304 case X86::CMOVE16rr:
1305 case X86::CMOVE32rr:
1306 case X86::CMOVE64rr:
1307 case X86::CMOVNE16rr:
1308 case X86::CMOVNE32rr:
1309 case X86::CMOVNE64rr:
1310 case X86::CMOVBE16rr:
1311 case X86::CMOVBE32rr:
1312 case X86::CMOVBE64rr:
1313 case X86::CMOVA16rr:
1314 case X86::CMOVA32rr:
1315 case X86::CMOVA64rr:
1316 case X86::CMOVL16rr:
1317 case X86::CMOVL32rr:
1318 case X86::CMOVL64rr:
1319 case X86::CMOVGE16rr:
1320 case X86::CMOVGE32rr:
1321 case X86::CMOVGE64rr:
1322 case X86::CMOVLE16rr:
1323 case X86::CMOVLE32rr:
1324 case X86::CMOVLE64rr:
1325 case X86::CMOVG16rr:
1326 case X86::CMOVG32rr:
1327 case X86::CMOVG64rr:
1328 case X86::CMOVS16rr:
1329 case X86::CMOVS32rr:
1330 case X86::CMOVS64rr:
1331 case X86::CMOVNS16rr:
1332 case X86::CMOVNS32rr:
1333 case X86::CMOVNS64rr:
1334 case X86::CMOVP16rr:
1335 case X86::CMOVP32rr:
1336 case X86::CMOVP64rr:
1337 case X86::CMOVNP16rr:
1338 case X86::CMOVNP32rr:
1339 case X86::CMOVNP64rr:
1340 case X86::CMOVO16rr:
1341 case X86::CMOVO32rr:
1342 case X86::CMOVO64rr:
1343 case X86::CMOVNO16rr:
1344 case X86::CMOVNO32rr:
1345 case X86::CMOVNO64rr: {
1347 switch (MI->getOpcode()) {
1349 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1350 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1351 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1352 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1353 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1354 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1355 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1356 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1357 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1358 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1359 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1360 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1361 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1362 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1363 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1364 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1365 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1366 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1367 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1368 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1369 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1370 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1371 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1372 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1373 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1374 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1375 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1376 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1377 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1378 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1379 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1380 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1381 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
1382 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1383 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1384 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1385 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1386 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1387 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
1388 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1389 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1390 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1391 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1392 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1393 case X86::CMOVO64rr: Opc = X86::CMOVNO32rr; break;
1394 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1395 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1396 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1399 MachineFunction &MF = *MI->getParent()->getParent();
1400 MI = MF.CloneMachineInstr(MI);
1403 MI->setDesc(get(Opc));
1404 // Fallthrough intended.
1407 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1411 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1413 default: return X86::COND_INVALID;
1414 case X86::JE: return X86::COND_E;
1415 case X86::JNE: return X86::COND_NE;
1416 case X86::JL: return X86::COND_L;
1417 case X86::JLE: return X86::COND_LE;
1418 case X86::JG: return X86::COND_G;
1419 case X86::JGE: return X86::COND_GE;
1420 case X86::JB: return X86::COND_B;
1421 case X86::JBE: return X86::COND_BE;
1422 case X86::JA: return X86::COND_A;
1423 case X86::JAE: return X86::COND_AE;
1424 case X86::JS: return X86::COND_S;
1425 case X86::JNS: return X86::COND_NS;
1426 case X86::JP: return X86::COND_P;
1427 case X86::JNP: return X86::COND_NP;
1428 case X86::JO: return X86::COND_O;
1429 case X86::JNO: return X86::COND_NO;
1433 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1435 default: assert(0 && "Illegal condition code!");
1436 case X86::COND_E: return X86::JE;
1437 case X86::COND_NE: return X86::JNE;
1438 case X86::COND_L: return X86::JL;
1439 case X86::COND_LE: return X86::JLE;
1440 case X86::COND_G: return X86::JG;
1441 case X86::COND_GE: return X86::JGE;
1442 case X86::COND_B: return X86::JB;
1443 case X86::COND_BE: return X86::JBE;
1444 case X86::COND_A: return X86::JA;
1445 case X86::COND_AE: return X86::JAE;
1446 case X86::COND_S: return X86::JS;
1447 case X86::COND_NS: return X86::JNS;
1448 case X86::COND_P: return X86::JP;
1449 case X86::COND_NP: return X86::JNP;
1450 case X86::COND_O: return X86::JO;
1451 case X86::COND_NO: return X86::JNO;
1455 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1456 /// e.g. turning COND_E to COND_NE.
1457 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1459 default: assert(0 && "Illegal condition code!");
1460 case X86::COND_E: return X86::COND_NE;
1461 case X86::COND_NE: return X86::COND_E;
1462 case X86::COND_L: return X86::COND_GE;
1463 case X86::COND_LE: return X86::COND_G;
1464 case X86::COND_G: return X86::COND_LE;
1465 case X86::COND_GE: return X86::COND_L;
1466 case X86::COND_B: return X86::COND_AE;
1467 case X86::COND_BE: return X86::COND_A;
1468 case X86::COND_A: return X86::COND_BE;
1469 case X86::COND_AE: return X86::COND_B;
1470 case X86::COND_S: return X86::COND_NS;
1471 case X86::COND_NS: return X86::COND_S;
1472 case X86::COND_P: return X86::COND_NP;
1473 case X86::COND_NP: return X86::COND_P;
1474 case X86::COND_O: return X86::COND_NO;
1475 case X86::COND_NO: return X86::COND_O;
1479 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1480 const TargetInstrDesc &TID = MI->getDesc();
1481 if (!TID.isTerminator()) return false;
1483 // Conditional branch is a special case.
1484 if (TID.isBranch() && !TID.isBarrier())
1486 if (!TID.isPredicable())
1488 return !isPredicated(MI);
1491 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
1492 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
1493 const X86InstrInfo &TII) {
1494 if (MI->getOpcode() == X86::FP_REG_KILL)
1496 return TII.isUnpredicatedTerminator(MI);
1499 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1500 MachineBasicBlock *&TBB,
1501 MachineBasicBlock *&FBB,
1502 SmallVectorImpl<MachineOperand> &Cond,
1503 bool AllowModify) const {
1504 // Start from the bottom of the block and work up, examining the
1505 // terminator instructions.
1506 MachineBasicBlock::iterator I = MBB.end();
1507 while (I != MBB.begin()) {
1509 // Working from the bottom, when we see a non-terminator
1510 // instruction, we're done.
1511 if (!isBrAnalysisUnpredicatedTerminator(I, *this))
1513 // A terminator that isn't a branch can't easily be handled
1514 // by this analysis.
1515 if (!I->getDesc().isBranch())
1517 // Handle unconditional branches.
1518 if (I->getOpcode() == X86::JMP) {
1520 TBB = I->getOperand(0).getMBB();
1524 // If the block has any instructions after a JMP, delete them.
1525 while (next(I) != MBB.end())
1526 next(I)->eraseFromParent();
1529 // Delete the JMP if it's equivalent to a fall-through.
1530 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1532 I->eraseFromParent();
1536 // TBB is used to indicate the unconditinal destination.
1537 TBB = I->getOperand(0).getMBB();
1540 // Handle conditional branches.
1541 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1542 if (BranchCode == X86::COND_INVALID)
1543 return true; // Can't handle indirect branch.
1544 // Working from the bottom, handle the first conditional branch.
1547 TBB = I->getOperand(0).getMBB();
1548 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1551 // Handle subsequent conditional branches. Only handle the case
1552 // where all conditional branches branch to the same destination
1553 // and their condition opcodes fit one of the special
1554 // multi-branch idioms.
1555 assert(Cond.size() == 1);
1557 // Only handle the case where all conditional branches branch to
1558 // the same destination.
1559 if (TBB != I->getOperand(0).getMBB())
1561 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1562 // If the conditions are the same, we can leave them alone.
1563 if (OldBranchCode == BranchCode)
1565 // If they differ, see if they fit one of the known patterns.
1566 // Theoretically we could handle more patterns here, but
1567 // we shouldn't expect to see them if instruction selection
1568 // has done a reasonable job.
1569 if ((OldBranchCode == X86::COND_NP &&
1570 BranchCode == X86::COND_E) ||
1571 (OldBranchCode == X86::COND_E &&
1572 BranchCode == X86::COND_NP))
1573 BranchCode = X86::COND_NP_OR_E;
1574 else if ((OldBranchCode == X86::COND_P &&
1575 BranchCode == X86::COND_NE) ||
1576 (OldBranchCode == X86::COND_NE &&
1577 BranchCode == X86::COND_P))
1578 BranchCode = X86::COND_NE_OR_P;
1581 // Update the MachineOperand.
1582 Cond[0].setImm(BranchCode);
1588 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1589 MachineBasicBlock::iterator I = MBB.end();
1592 while (I != MBB.begin()) {
1594 if (I->getOpcode() != X86::JMP &&
1595 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1597 // Remove the branch.
1598 I->eraseFromParent();
1607 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1608 MachineBasicBlock *FBB,
1609 const SmallVectorImpl<MachineOperand> &Cond) const {
1610 // FIXME this should probably have a DebugLoc operand
1611 DebugLoc dl = DebugLoc::getUnknownLoc();
1612 // Shouldn't be a fall through.
1613 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1614 assert((Cond.size() == 1 || Cond.size() == 0) &&
1615 "X86 branch conditions have one component!");
1618 // Unconditional branch?
1619 assert(!FBB && "Unconditional branch with multiple successors!");
1620 BuildMI(&MBB, dl, get(X86::JMP)).addMBB(TBB);
1624 // Conditional branch.
1626 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1628 case X86::COND_NP_OR_E:
1629 // Synthesize NP_OR_E with two branches.
1630 BuildMI(&MBB, dl, get(X86::JNP)).addMBB(TBB);
1632 BuildMI(&MBB, dl, get(X86::JE)).addMBB(TBB);
1635 case X86::COND_NE_OR_P:
1636 // Synthesize NE_OR_P with two branches.
1637 BuildMI(&MBB, dl, get(X86::JNE)).addMBB(TBB);
1639 BuildMI(&MBB, dl, get(X86::JP)).addMBB(TBB);
1643 unsigned Opc = GetCondBranchFromCond(CC);
1644 BuildMI(&MBB, dl, get(Opc)).addMBB(TBB);
1649 // Two-way Conditional branch. Insert the second branch.
1650 BuildMI(&MBB, dl, get(X86::JMP)).addMBB(FBB);
1656 bool X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
1657 MachineBasicBlock::iterator MI,
1658 unsigned DestReg, unsigned SrcReg,
1659 const TargetRegisterClass *DestRC,
1660 const TargetRegisterClass *SrcRC) const {
1661 DebugLoc DL = DebugLoc::getUnknownLoc();
1662 if (MI != MBB.end()) DL = MI->getDebugLoc();
1664 if (DestRC == SrcRC) {
1666 if (DestRC == &X86::GR64RegClass) {
1668 } else if (DestRC == &X86::GR32RegClass) {
1670 } else if (DestRC == &X86::GR16RegClass) {
1672 } else if (DestRC == &X86::GR8RegClass) {
1674 } else if (DestRC == &X86::GR32_RegClass) {
1675 Opc = X86::MOV32_rr;
1676 } else if (DestRC == &X86::GR16_RegClass) {
1677 Opc = X86::MOV16_rr;
1678 } else if (DestRC == &X86::RFP32RegClass) {
1679 Opc = X86::MOV_Fp3232;
1680 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
1681 Opc = X86::MOV_Fp6464;
1682 } else if (DestRC == &X86::RFP80RegClass) {
1683 Opc = X86::MOV_Fp8080;
1684 } else if (DestRC == &X86::FR32RegClass) {
1685 Opc = X86::FsMOVAPSrr;
1686 } else if (DestRC == &X86::FR64RegClass) {
1687 Opc = X86::FsMOVAPDrr;
1688 } else if (DestRC == &X86::VR128RegClass) {
1689 Opc = X86::MOVAPSrr;
1690 } else if (DestRC == &X86::VR64RegClass) {
1691 Opc = X86::MMX_MOVQ64rr;
1695 BuildMI(MBB, MI, DL, get(Opc), DestReg).addReg(SrcReg);
1699 // Moving EFLAGS to / from another register requires a push and a pop.
1700 if (SrcRC == &X86::CCRRegClass) {
1701 if (SrcReg != X86::EFLAGS)
1703 if (DestRC == &X86::GR64RegClass) {
1704 BuildMI(MBB, MI, DL, get(X86::PUSHFQ));
1705 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
1707 } else if (DestRC == &X86::GR32RegClass) {
1708 BuildMI(MBB, MI, DL, get(X86::PUSHFD));
1709 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
1712 } else if (DestRC == &X86::CCRRegClass) {
1713 if (DestReg != X86::EFLAGS)
1715 if (SrcRC == &X86::GR64RegClass) {
1716 BuildMI(MBB, MI, DL, get(X86::PUSH64r)).addReg(SrcReg);
1717 BuildMI(MBB, MI, DL, get(X86::POPFQ));
1719 } else if (SrcRC == &X86::GR32RegClass) {
1720 BuildMI(MBB, MI, DL, get(X86::PUSH32r)).addReg(SrcReg);
1721 BuildMI(MBB, MI, DL, get(X86::POPFD));
1726 // Moving from ST(0) turns into FpGET_ST0_32 etc.
1727 if (SrcRC == &X86::RSTRegClass) {
1728 // Copying from ST(0)/ST(1).
1729 if (SrcReg != X86::ST0 && SrcReg != X86::ST1)
1730 // Can only copy from ST(0)/ST(1) right now
1732 bool isST0 = SrcReg == X86::ST0;
1734 if (DestRC == &X86::RFP32RegClass)
1735 Opc = isST0 ? X86::FpGET_ST0_32 : X86::FpGET_ST1_32;
1736 else if (DestRC == &X86::RFP64RegClass)
1737 Opc = isST0 ? X86::FpGET_ST0_64 : X86::FpGET_ST1_64;
1739 if (DestRC != &X86::RFP80RegClass)
1741 Opc = isST0 ? X86::FpGET_ST0_80 : X86::FpGET_ST1_80;
1743 BuildMI(MBB, MI, DL, get(Opc), DestReg);
1747 // Moving to ST(0) turns into FpSET_ST0_32 etc.
1748 if (DestRC == &X86::RSTRegClass) {
1749 // Copying to ST(0) / ST(1).
1750 if (DestReg != X86::ST0 && DestReg != X86::ST1)
1751 // Can only copy to TOS right now
1753 bool isST0 = DestReg == X86::ST0;
1755 if (SrcRC == &X86::RFP32RegClass)
1756 Opc = isST0 ? X86::FpSET_ST0_32 : X86::FpSET_ST1_32;
1757 else if (SrcRC == &X86::RFP64RegClass)
1758 Opc = isST0 ? X86::FpSET_ST0_64 : X86::FpSET_ST1_64;
1760 if (SrcRC != &X86::RFP80RegClass)
1762 Opc = isST0 ? X86::FpSET_ST0_80 : X86::FpSET_ST1_80;
1764 BuildMI(MBB, MI, DL, get(Opc)).addReg(SrcReg);
1768 // Not yet supported!
1772 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
1773 bool isStackAligned) {
1775 if (RC == &X86::GR64RegClass) {
1777 } else if (RC == &X86::GR32RegClass) {
1779 } else if (RC == &X86::GR16RegClass) {
1781 } else if (RC == &X86::GR8RegClass) {
1783 } else if (RC == &X86::GR32_RegClass) {
1784 Opc = X86::MOV32_mr;
1785 } else if (RC == &X86::GR16_RegClass) {
1786 Opc = X86::MOV16_mr;
1787 } else if (RC == &X86::RFP80RegClass) {
1788 Opc = X86::ST_FpP80m; // pops
1789 } else if (RC == &X86::RFP64RegClass) {
1790 Opc = X86::ST_Fp64m;
1791 } else if (RC == &X86::RFP32RegClass) {
1792 Opc = X86::ST_Fp32m;
1793 } else if (RC == &X86::FR32RegClass) {
1795 } else if (RC == &X86::FR64RegClass) {
1797 } else if (RC == &X86::VR128RegClass) {
1798 // If stack is realigned we can use aligned stores.
1799 Opc = isStackAligned ? X86::MOVAPSmr : X86::MOVUPSmr;
1800 } else if (RC == &X86::VR64RegClass) {
1801 Opc = X86::MMX_MOVQ64mr;
1803 assert(0 && "Unknown regclass");
1810 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
1811 MachineBasicBlock::iterator MI,
1812 unsigned SrcReg, bool isKill, int FrameIdx,
1813 const TargetRegisterClass *RC) const {
1814 const MachineFunction &MF = *MBB.getParent();
1815 bool isAligned = (RI.getStackAlignment() >= 16) ||
1816 RI.needsStackRealignment(MF);
1817 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1818 DebugLoc DL = DebugLoc::getUnknownLoc();
1819 if (MI != MBB.end()) DL = MI->getDebugLoc();
1820 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
1821 .addReg(SrcReg, false, false, isKill);
1824 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
1826 SmallVectorImpl<MachineOperand> &Addr,
1827 const TargetRegisterClass *RC,
1828 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1829 bool isAligned = (RI.getStackAlignment() >= 16) ||
1830 RI.needsStackRealignment(MF);
1831 unsigned Opc = getStoreRegOpcode(RC, isAligned);
1832 DebugLoc DL = DebugLoc::getUnknownLoc();
1833 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
1834 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1835 MIB.addOperand(Addr[i]);
1836 MIB.addReg(SrcReg, false, false, isKill);
1837 NewMIs.push_back(MIB);
1840 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
1841 bool isStackAligned) {
1843 if (RC == &X86::GR64RegClass) {
1845 } else if (RC == &X86::GR32RegClass) {
1847 } else if (RC == &X86::GR16RegClass) {
1849 } else if (RC == &X86::GR8RegClass) {
1851 } else if (RC == &X86::GR32_RegClass) {
1852 Opc = X86::MOV32_rm;
1853 } else if (RC == &X86::GR16_RegClass) {
1854 Opc = X86::MOV16_rm;
1855 } else if (RC == &X86::RFP80RegClass) {
1856 Opc = X86::LD_Fp80m;
1857 } else if (RC == &X86::RFP64RegClass) {
1858 Opc = X86::LD_Fp64m;
1859 } else if (RC == &X86::RFP32RegClass) {
1860 Opc = X86::LD_Fp32m;
1861 } else if (RC == &X86::FR32RegClass) {
1863 } else if (RC == &X86::FR64RegClass) {
1865 } else if (RC == &X86::VR128RegClass) {
1866 // If stack is realigned we can use aligned loads.
1867 Opc = isStackAligned ? X86::MOVAPSrm : X86::MOVUPSrm;
1868 } else if (RC == &X86::VR64RegClass) {
1869 Opc = X86::MMX_MOVQ64rm;
1871 assert(0 && "Unknown regclass");
1878 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1879 MachineBasicBlock::iterator MI,
1880 unsigned DestReg, int FrameIdx,
1881 const TargetRegisterClass *RC) const{
1882 const MachineFunction &MF = *MBB.getParent();
1883 bool isAligned = (RI.getStackAlignment() >= 16) ||
1884 RI.needsStackRealignment(MF);
1885 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1886 DebugLoc DL = DebugLoc::getUnknownLoc();
1887 if (MI != MBB.end()) DL = MI->getDebugLoc();
1888 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
1891 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
1892 SmallVectorImpl<MachineOperand> &Addr,
1893 const TargetRegisterClass *RC,
1894 SmallVectorImpl<MachineInstr*> &NewMIs) const {
1895 bool isAligned = (RI.getStackAlignment() >= 16) ||
1896 RI.needsStackRealignment(MF);
1897 unsigned Opc = getLoadRegOpcode(RC, isAligned);
1898 DebugLoc DL = DebugLoc::getUnknownLoc();
1899 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
1900 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
1901 MIB.addOperand(Addr[i]);
1902 NewMIs.push_back(MIB);
1905 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
1906 MachineBasicBlock::iterator MI,
1907 const std::vector<CalleeSavedInfo> &CSI) const {
1911 DebugLoc DL = DebugLoc::getUnknownLoc();
1912 if (MI != MBB.end()) DL = MI->getDebugLoc();
1914 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1915 unsigned SlotSize = is64Bit ? 8 : 4;
1917 MachineFunction &MF = *MBB.getParent();
1918 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
1919 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
1921 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
1922 for (unsigned i = CSI.size(); i != 0; --i) {
1923 unsigned Reg = CSI[i-1].getReg();
1924 // Add the callee-saved register as live-in. It's killed at the spill.
1926 BuildMI(MBB, MI, DL, get(Opc))
1927 .addReg(Reg, /*isDef=*/false, /*isImp=*/false, /*isKill=*/true);
1932 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
1933 MachineBasicBlock::iterator MI,
1934 const std::vector<CalleeSavedInfo> &CSI) const {
1938 DebugLoc DL = DebugLoc::getUnknownLoc();
1939 if (MI != MBB.end()) DL = MI->getDebugLoc();
1941 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1943 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
1944 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
1945 unsigned Reg = CSI[i].getReg();
1946 BuildMI(MBB, MI, DL, get(Opc), Reg);
1951 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
1952 const SmallVectorImpl<MachineOperand> &MOs,
1954 const TargetInstrInfo &TII) {
1955 // Create the base instruction with the memory operand as the first part.
1956 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
1957 MI->getDebugLoc(), true);
1958 MachineInstrBuilder MIB(NewMI);
1959 unsigned NumAddrOps = MOs.size();
1960 for (unsigned i = 0; i != NumAddrOps; ++i)
1961 MIB.addOperand(MOs[i]);
1962 if (NumAddrOps < 4) // FrameIndex only
1963 MIB.addImm(1).addReg(0).addImm(0);
1965 // Loop over the rest of the ri operands, converting them over.
1966 unsigned NumOps = MI->getDesc().getNumOperands()-2;
1967 for (unsigned i = 0; i != NumOps; ++i) {
1968 MachineOperand &MO = MI->getOperand(i+2);
1971 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
1972 MachineOperand &MO = MI->getOperand(i);
1978 static MachineInstr *FuseInst(MachineFunction &MF,
1979 unsigned Opcode, unsigned OpNo,
1980 const SmallVectorImpl<MachineOperand> &MOs,
1981 MachineInstr *MI, const TargetInstrInfo &TII) {
1982 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
1983 MI->getDebugLoc(), true);
1984 MachineInstrBuilder MIB(NewMI);
1986 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1987 MachineOperand &MO = MI->getOperand(i);
1989 assert(MO.isReg() && "Expected to fold into reg operand!");
1990 unsigned NumAddrOps = MOs.size();
1991 for (unsigned i = 0; i != NumAddrOps; ++i)
1992 MIB.addOperand(MOs[i]);
1993 if (NumAddrOps < 4) // FrameIndex only
1994 MIB.addImm(1).addReg(0).addImm(0);
2002 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
2003 const SmallVectorImpl<MachineOperand> &MOs,
2005 MachineFunction &MF = *MI->getParent()->getParent();
2006 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
2008 unsigned NumAddrOps = MOs.size();
2009 for (unsigned i = 0; i != NumAddrOps; ++i)
2010 MIB.addOperand(MOs[i]);
2011 if (NumAddrOps < 4) // FrameIndex only
2012 MIB.addImm(1).addReg(0).addImm(0);
2013 return MIB.addImm(0);
2017 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2018 MachineInstr *MI, unsigned i,
2019 const SmallVectorImpl<MachineOperand> &MOs) const{
2020 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2021 bool isTwoAddrFold = false;
2022 unsigned NumOps = MI->getDesc().getNumOperands();
2023 bool isTwoAddr = NumOps > 1 &&
2024 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2026 MachineInstr *NewMI = NULL;
2027 // Folding a memory location into the two-address part of a two-address
2028 // instruction is different than folding it other places. It requires
2029 // replacing the *two* registers with the memory location.
2030 if (isTwoAddr && NumOps >= 2 && i < 2 &&
2031 MI->getOperand(0).isReg() &&
2032 MI->getOperand(1).isReg() &&
2033 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
2034 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2035 isTwoAddrFold = true;
2036 } else if (i == 0) { // If operand 0
2037 if (MI->getOpcode() == X86::MOV16r0)
2038 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
2039 else if (MI->getOpcode() == X86::MOV32r0)
2040 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
2041 else if (MI->getOpcode() == X86::MOV64r0)
2042 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2043 else if (MI->getOpcode() == X86::MOV8r0)
2044 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2048 OpcodeTablePtr = &RegOp2MemOpTable0;
2049 } else if (i == 1) {
2050 OpcodeTablePtr = &RegOp2MemOpTable1;
2051 } else if (i == 2) {
2052 OpcodeTablePtr = &RegOp2MemOpTable2;
2055 // If table selected...
2056 if (OpcodeTablePtr) {
2057 // Find the Opcode to fuse
2058 DenseMap<unsigned*, unsigned>::iterator I =
2059 OpcodeTablePtr->find((unsigned*)MI->getOpcode());
2060 if (I != OpcodeTablePtr->end()) {
2062 NewMI = FuseTwoAddrInst(MF, I->second, MOs, MI, *this);
2064 NewMI = FuseInst(MF, I->second, i, MOs, MI, *this);
2070 if (PrintFailedFusing)
2071 cerr << "We failed to fuse operand " << i << " in " << *MI;
2076 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2078 const SmallVectorImpl<unsigned> &Ops,
2079 int FrameIndex) const {
2080 // Check switch flag
2081 if (NoFusing) return NULL;
2083 const MachineFrameInfo *MFI = MF.getFrameInfo();
2084 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2085 // FIXME: Move alignment requirement into tables?
2086 if (Alignment < 16) {
2087 switch (MI->getOpcode()) {
2089 // Not always safe to fold movsd into these instructions since their load
2090 // folding variants expects the address to be 16 byte aligned.
2091 case X86::FsANDNPDrr:
2092 case X86::FsANDNPSrr:
2093 case X86::FsANDPDrr:
2094 case X86::FsANDPSrr:
2097 case X86::FsXORPDrr:
2098 case X86::FsXORPSrr:
2103 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2104 unsigned NewOpc = 0;
2105 switch (MI->getOpcode()) {
2106 default: return NULL;
2107 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2108 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2109 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2110 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2112 // Change to CMPXXri r, 0 first.
2113 MI->setDesc(get(NewOpc));
2114 MI->getOperand(1).ChangeToImmediate(0);
2115 } else if (Ops.size() != 1)
2118 SmallVector<MachineOperand,4> MOs;
2119 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2120 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2123 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2125 const SmallVectorImpl<unsigned> &Ops,
2126 MachineInstr *LoadMI) const {
2127 // Check switch flag
2128 if (NoFusing) return NULL;
2130 // Determine the alignment of the load.
2131 unsigned Alignment = 0;
2132 if (LoadMI->hasOneMemOperand())
2133 Alignment = LoadMI->memoperands_begin()->getAlignment();
2135 // FIXME: Move alignment requirement into tables?
2136 if (Alignment < 16) {
2137 switch (MI->getOpcode()) {
2139 // Not always safe to fold movsd into these instructions since their load
2140 // folding variants expects the address to be 16 byte aligned.
2141 case X86::FsANDNPDrr:
2142 case X86::FsANDNPSrr:
2143 case X86::FsANDPDrr:
2144 case X86::FsANDPSrr:
2147 case X86::FsXORPDrr:
2148 case X86::FsXORPSrr:
2153 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2154 unsigned NewOpc = 0;
2155 switch (MI->getOpcode()) {
2156 default: return NULL;
2157 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2158 case X86::TEST16rr: NewOpc = X86::CMP16ri; break;
2159 case X86::TEST32rr: NewOpc = X86::CMP32ri; break;
2160 case X86::TEST64rr: NewOpc = X86::CMP64ri32; break;
2162 // Change to CMPXXri r, 0 first.
2163 MI->setDesc(get(NewOpc));
2164 MI->getOperand(1).ChangeToImmediate(0);
2165 } else if (Ops.size() != 1)
2168 SmallVector<MachineOperand,4> MOs;
2169 if (LoadMI->getOpcode() == X86::V_SET0 ||
2170 LoadMI->getOpcode() == X86::V_SETALLONES) {
2171 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
2172 // Create a constant-pool entry and operands to load from it.
2174 // x86-32 PIC requires a PIC base register for constant pools.
2175 unsigned PICBase = 0;
2176 if (TM.getRelocationModel() == Reloc::PIC_ &&
2177 !TM.getSubtarget<X86Subtarget>().is64Bit())
2178 // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
2179 // This doesn't work for several reasons.
2180 // 1. GlobalBaseReg may have been spilled.
2181 // 2. It may not be live at MI.
2184 // Create a v4i32 constant-pool entry.
2185 MachineConstantPool &MCP = *MF.getConstantPool();
2186 const VectorType *Ty = VectorType::get(Type::Int32Ty, 4);
2187 Constant *C = LoadMI->getOpcode() == X86::V_SET0 ?
2188 ConstantVector::getNullValue(Ty) :
2189 ConstantVector::getAllOnesValue(Ty);
2190 unsigned CPI = MCP.getConstantPoolIndex(C, /*AlignmentLog2=*/4);
2192 // Create operands to load from the constant pool entry.
2193 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
2194 MOs.push_back(MachineOperand::CreateImm(1));
2195 MOs.push_back(MachineOperand::CreateReg(0, false));
2196 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
2198 // Folding a normal load. Just copy the load's address operands.
2199 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2200 for (unsigned i = NumOps - 4; i != NumOps; ++i)
2201 MOs.push_back(LoadMI->getOperand(i));
2203 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs);
2207 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2208 const SmallVectorImpl<unsigned> &Ops) const {
2209 // Check switch flag
2210 if (NoFusing) return 0;
2212 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2213 switch (MI->getOpcode()) {
2214 default: return false;
2223 if (Ops.size() != 1)
2226 unsigned OpNum = Ops[0];
2227 unsigned Opc = MI->getOpcode();
2228 unsigned NumOps = MI->getDesc().getNumOperands();
2229 bool isTwoAddr = NumOps > 1 &&
2230 MI->getDesc().getOperandConstraint(1, TOI::TIED_TO) != -1;
2232 // Folding a memory location into the two-address part of a two-address
2233 // instruction is different than folding it other places. It requires
2234 // replacing the *two* registers with the memory location.
2235 const DenseMap<unsigned*, unsigned> *OpcodeTablePtr = NULL;
2236 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2237 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2238 } else if (OpNum == 0) { // If operand 0
2247 OpcodeTablePtr = &RegOp2MemOpTable0;
2248 } else if (OpNum == 1) {
2249 OpcodeTablePtr = &RegOp2MemOpTable1;
2250 } else if (OpNum == 2) {
2251 OpcodeTablePtr = &RegOp2MemOpTable2;
2254 if (OpcodeTablePtr) {
2255 // Find the Opcode to fuse
2256 DenseMap<unsigned*, unsigned>::iterator I =
2257 OpcodeTablePtr->find((unsigned*)Opc);
2258 if (I != OpcodeTablePtr->end())
2264 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2265 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2266 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2267 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2268 MemOp2RegOpTable.find((unsigned*)MI->getOpcode());
2269 if (I == MemOp2RegOpTable.end())
2271 DebugLoc dl = MI->getDebugLoc();
2272 unsigned Opc = I->second.first;
2273 unsigned Index = I->second.second & 0xf;
2274 bool FoldedLoad = I->second.second & (1 << 4);
2275 bool FoldedStore = I->second.second & (1 << 5);
2276 if (UnfoldLoad && !FoldedLoad)
2278 UnfoldLoad &= FoldedLoad;
2279 if (UnfoldStore && !FoldedStore)
2281 UnfoldStore &= FoldedStore;
2283 const TargetInstrDesc &TID = get(Opc);
2284 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2285 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2286 ? RI.getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2287 SmallVector<MachineOperand,4> AddrOps;
2288 SmallVector<MachineOperand,2> BeforeOps;
2289 SmallVector<MachineOperand,2> AfterOps;
2290 SmallVector<MachineOperand,4> ImpOps;
2291 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2292 MachineOperand &Op = MI->getOperand(i);
2293 if (i >= Index && i < Index+4)
2294 AddrOps.push_back(Op);
2295 else if (Op.isReg() && Op.isImplicit())
2296 ImpOps.push_back(Op);
2298 BeforeOps.push_back(Op);
2300 AfterOps.push_back(Op);
2303 // Emit the load instruction.
2305 loadRegFromAddr(MF, Reg, AddrOps, RC, NewMIs);
2307 // Address operands cannot be marked isKill.
2308 for (unsigned i = 1; i != 5; ++i) {
2309 MachineOperand &MO = NewMIs[0]->getOperand(i);
2311 MO.setIsKill(false);
2316 // Emit the data processing instruction.
2317 MachineInstr *DataMI = MF.CreateMachineInstr(TID, MI->getDebugLoc(), true);
2318 MachineInstrBuilder MIB(DataMI);
2321 MIB.addReg(Reg, true);
2322 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2323 MIB.addOperand(BeforeOps[i]);
2326 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2327 MIB.addOperand(AfterOps[i]);
2328 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2329 MachineOperand &MO = ImpOps[i];
2330 MIB.addReg(MO.getReg(), MO.isDef(), true, MO.isKill(), MO.isDead());
2332 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2333 unsigned NewOpc = 0;
2334 switch (DataMI->getOpcode()) {
2336 case X86::CMP64ri32:
2340 MachineOperand &MO0 = DataMI->getOperand(0);
2341 MachineOperand &MO1 = DataMI->getOperand(1);
2342 if (MO1.getImm() == 0) {
2343 switch (DataMI->getOpcode()) {
2345 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2346 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2347 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2348 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2350 DataMI->setDesc(get(NewOpc));
2351 MO1.ChangeToRegister(MO0.getReg(), false);
2355 NewMIs.push_back(DataMI);
2357 // Emit the store instruction.
2359 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2360 const TargetRegisterClass *DstRC = DstTOI.isLookupPtrRegClass()
2361 ? RI.getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2362 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, NewMIs);
2369 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2370 SmallVectorImpl<SDNode*> &NewNodes) const {
2371 if (!N->isMachineOpcode())
2374 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2375 MemOp2RegOpTable.find((unsigned*)N->getMachineOpcode());
2376 if (I == MemOp2RegOpTable.end())
2378 unsigned Opc = I->second.first;
2379 unsigned Index = I->second.second & 0xf;
2380 bool FoldedLoad = I->second.second & (1 << 4);
2381 bool FoldedStore = I->second.second & (1 << 5);
2382 const TargetInstrDesc &TID = get(Opc);
2383 const TargetOperandInfo &TOI = TID.OpInfo[Index];
2384 const TargetRegisterClass *RC = TOI.isLookupPtrRegClass()
2385 ? RI.getPointerRegClass() : RI.getRegClass(TOI.RegClass);
2386 std::vector<SDValue> AddrOps;
2387 std::vector<SDValue> BeforeOps;
2388 std::vector<SDValue> AfterOps;
2389 DebugLoc dl = N->getDebugLoc();
2390 unsigned NumOps = N->getNumOperands();
2391 for (unsigned i = 0; i != NumOps-1; ++i) {
2392 SDValue Op = N->getOperand(i);
2393 if (i >= Index && i < Index+4)
2394 AddrOps.push_back(Op);
2396 BeforeOps.push_back(Op);
2398 AfterOps.push_back(Op);
2400 SDValue Chain = N->getOperand(NumOps-1);
2401 AddrOps.push_back(Chain);
2403 // Emit the load instruction.
2405 const MachineFunction &MF = DAG.getMachineFunction();
2407 MVT VT = *RC->vt_begin();
2408 bool isAligned = (RI.getStackAlignment() >= 16) ||
2409 RI.needsStackRealignment(MF);
2410 Load = DAG.getTargetNode(getLoadRegOpcode(RC, isAligned), dl,
2412 &AddrOps[0], AddrOps.size());
2413 NewNodes.push_back(Load);
2416 // Emit the data processing instruction.
2417 std::vector<MVT> VTs;
2418 const TargetRegisterClass *DstRC = 0;
2419 if (TID.getNumDefs() > 0) {
2420 const TargetOperandInfo &DstTOI = TID.OpInfo[0];
2421 DstRC = DstTOI.isLookupPtrRegClass()
2422 ? RI.getPointerRegClass() : RI.getRegClass(DstTOI.RegClass);
2423 VTs.push_back(*DstRC->vt_begin());
2425 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2426 MVT VT = N->getValueType(i);
2427 if (VT != MVT::Other && i >= (unsigned)TID.getNumDefs())
2431 BeforeOps.push_back(SDValue(Load, 0));
2432 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2433 SDNode *NewNode= DAG.getTargetNode(Opc, dl, VTs, &BeforeOps[0],
2435 NewNodes.push_back(NewNode);
2437 // Emit the store instruction.
2440 AddrOps.push_back(SDValue(NewNode, 0));
2441 AddrOps.push_back(Chain);
2442 bool isAligned = (RI.getStackAlignment() >= 16) ||
2443 RI.needsStackRealignment(MF);
2444 SDNode *Store = DAG.getTargetNode(getStoreRegOpcode(DstRC, isAligned), dl,
2445 MVT::Other, &AddrOps[0], AddrOps.size());
2446 NewNodes.push_back(Store);
2452 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2453 bool UnfoldLoad, bool UnfoldStore) const {
2454 DenseMap<unsigned*, std::pair<unsigned,unsigned> >::iterator I =
2455 MemOp2RegOpTable.find((unsigned*)Opc);
2456 if (I == MemOp2RegOpTable.end())
2458 bool FoldedLoad = I->second.second & (1 << 4);
2459 bool FoldedStore = I->second.second & (1 << 5);
2460 if (UnfoldLoad && !FoldedLoad)
2462 if (UnfoldStore && !FoldedStore)
2464 return I->second.first;
2467 bool X86InstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
2468 if (MBB.empty()) return false;
2470 switch (MBB.back().getOpcode()) {
2471 case X86::TCRETURNri:
2472 case X86::TCRETURNdi:
2473 case X86::RET: // Return.
2478 case X86::JMP: // Uncond branch.
2479 case X86::JMP32r: // Indirect branch.
2480 case X86::JMP64r: // Indirect branch (64-bit).
2481 case X86::JMP32m: // Indirect branch through mem.
2482 case X86::JMP64m: // Indirect branch through mem (64-bit).
2484 default: return false;
2489 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
2490 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
2491 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
2492 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
2494 Cond[0].setImm(GetOppositeBranchCondition(CC));
2499 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
2500 // FIXME: Return false for x87 stack register classes for now. We can't
2501 // allow any loads of these registers before FpGet_ST0_80.
2502 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
2503 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
2506 unsigned X86InstrInfo::sizeOfImm(const TargetInstrDesc *Desc) {
2507 switch (Desc->TSFlags & X86II::ImmMask) {
2508 case X86II::Imm8: return 1;
2509 case X86II::Imm16: return 2;
2510 case X86II::Imm32: return 4;
2511 case X86II::Imm64: return 8;
2512 default: assert(0 && "Immediate size not set!");
2517 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended register?
2518 /// e.g. r8, xmm8, etc.
2519 bool X86InstrInfo::isX86_64ExtendedReg(const MachineOperand &MO) {
2520 if (!MO.isReg()) return false;
2521 switch (MO.getReg()) {
2523 case X86::R8: case X86::R9: case X86::R10: case X86::R11:
2524 case X86::R12: case X86::R13: case X86::R14: case X86::R15:
2525 case X86::R8D: case X86::R9D: case X86::R10D: case X86::R11D:
2526 case X86::R12D: case X86::R13D: case X86::R14D: case X86::R15D:
2527 case X86::R8W: case X86::R9W: case X86::R10W: case X86::R11W:
2528 case X86::R12W: case X86::R13W: case X86::R14W: case X86::R15W:
2529 case X86::R8B: case X86::R9B: case X86::R10B: case X86::R11B:
2530 case X86::R12B: case X86::R13B: case X86::R14B: case X86::R15B:
2531 case X86::XMM8: case X86::XMM9: case X86::XMM10: case X86::XMM11:
2532 case X86::XMM12: case X86::XMM13: case X86::XMM14: case X86::XMM15:
2539 /// determineREX - Determine if the MachineInstr has to be encoded with a X86-64
2540 /// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand
2541 /// size, and 3) use of X86-64 extended registers.
2542 unsigned X86InstrInfo::determineREX(const MachineInstr &MI) {
2544 const TargetInstrDesc &Desc = MI.getDesc();
2546 // Pseudo instructions do not need REX prefix byte.
2547 if ((Desc.TSFlags & X86II::FormMask) == X86II::Pseudo)
2549 if (Desc.TSFlags & X86II::REX_W)
2552 unsigned NumOps = Desc.getNumOperands();
2554 bool isTwoAddr = NumOps > 1 &&
2555 Desc.getOperandConstraint(1, TOI::TIED_TO) != -1;
2557 // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix.
2558 unsigned i = isTwoAddr ? 1 : 0;
2559 for (unsigned e = NumOps; i != e; ++i) {
2560 const MachineOperand& MO = MI.getOperand(i);
2562 unsigned Reg = MO.getReg();
2563 if (isX86_64NonExtLowByteReg(Reg))
2568 switch (Desc.TSFlags & X86II::FormMask) {
2569 case X86II::MRMInitReg:
2570 if (isX86_64ExtendedReg(MI.getOperand(0)))
2571 REX |= (1 << 0) | (1 << 2);
2573 case X86II::MRMSrcReg: {
2574 if (isX86_64ExtendedReg(MI.getOperand(0)))
2576 i = isTwoAddr ? 2 : 1;
2577 for (unsigned e = NumOps; i != e; ++i) {
2578 const MachineOperand& MO = MI.getOperand(i);
2579 if (isX86_64ExtendedReg(MO))
2584 case X86II::MRMSrcMem: {
2585 if (isX86_64ExtendedReg(MI.getOperand(0)))
2588 i = isTwoAddr ? 2 : 1;
2589 for (; i != NumOps; ++i) {
2590 const MachineOperand& MO = MI.getOperand(i);
2592 if (isX86_64ExtendedReg(MO))
2599 case X86II::MRM0m: case X86II::MRM1m:
2600 case X86II::MRM2m: case X86II::MRM3m:
2601 case X86II::MRM4m: case X86II::MRM5m:
2602 case X86II::MRM6m: case X86II::MRM7m:
2603 case X86II::MRMDestMem: {
2604 unsigned e = isTwoAddr ? 5 : 4;
2605 i = isTwoAddr ? 1 : 0;
2606 if (NumOps > e && isX86_64ExtendedReg(MI.getOperand(e)))
2609 for (; i != e; ++i) {
2610 const MachineOperand& MO = MI.getOperand(i);
2612 if (isX86_64ExtendedReg(MO))
2620 if (isX86_64ExtendedReg(MI.getOperand(0)))
2622 i = isTwoAddr ? 2 : 1;
2623 for (unsigned e = NumOps; i != e; ++i) {
2624 const MachineOperand& MO = MI.getOperand(i);
2625 if (isX86_64ExtendedReg(MO))
2635 /// sizePCRelativeBlockAddress - This method returns the size of a PC
2636 /// relative block address instruction
2638 static unsigned sizePCRelativeBlockAddress() {
2642 /// sizeGlobalAddress - Give the size of the emission of this global address
2644 static unsigned sizeGlobalAddress(bool dword) {
2645 return dword ? 8 : 4;
2648 /// sizeConstPoolAddress - Give the size of the emission of this constant
2651 static unsigned sizeConstPoolAddress(bool dword) {
2652 return dword ? 8 : 4;
2655 /// sizeExternalSymbolAddress - Give the size of the emission of this external
2658 static unsigned sizeExternalSymbolAddress(bool dword) {
2659 return dword ? 8 : 4;
2662 /// sizeJumpTableAddress - Give the size of the emission of this jump
2665 static unsigned sizeJumpTableAddress(bool dword) {
2666 return dword ? 8 : 4;
2669 static unsigned sizeConstant(unsigned Size) {
2673 static unsigned sizeRegModRMByte(){
2677 static unsigned sizeSIBByte(){
2681 static unsigned getDisplacementFieldSize(const MachineOperand *RelocOp) {
2682 unsigned FinalSize = 0;
2683 // If this is a simple integer displacement that doesn't require a relocation.
2685 FinalSize += sizeConstant(4);
2689 // Otherwise, this is something that requires a relocation.
2690 if (RelocOp->isGlobal()) {
2691 FinalSize += sizeGlobalAddress(false);
2692 } else if (RelocOp->isCPI()) {
2693 FinalSize += sizeConstPoolAddress(false);
2694 } else if (RelocOp->isJTI()) {
2695 FinalSize += sizeJumpTableAddress(false);
2697 assert(0 && "Unknown value to relocate!");
2702 static unsigned getMemModRMByteSize(const MachineInstr &MI, unsigned Op,
2703 bool IsPIC, bool Is64BitMode) {
2704 const MachineOperand &Op3 = MI.getOperand(Op+3);
2706 const MachineOperand *DispForReloc = 0;
2707 unsigned FinalSize = 0;
2709 // Figure out what sort of displacement we have to handle here.
2710 if (Op3.isGlobal()) {
2711 DispForReloc = &Op3;
2712 } else if (Op3.isCPI()) {
2713 if (Is64BitMode || IsPIC) {
2714 DispForReloc = &Op3;
2718 } else if (Op3.isJTI()) {
2719 if (Is64BitMode || IsPIC) {
2720 DispForReloc = &Op3;
2728 const MachineOperand &Base = MI.getOperand(Op);
2729 const MachineOperand &IndexReg = MI.getOperand(Op+2);
2731 unsigned BaseReg = Base.getReg();
2733 // Is a SIB byte needed?
2734 if (IndexReg.getReg() == 0 &&
2735 (BaseReg == 0 || X86RegisterInfo::getX86RegNum(BaseReg) != N86::ESP)) {
2736 if (BaseReg == 0) { // Just a displacement?
2737 // Emit special case [disp32] encoding
2739 FinalSize += getDisplacementFieldSize(DispForReloc);
2741 unsigned BaseRegNo = X86RegisterInfo::getX86RegNum(BaseReg);
2742 if (!DispForReloc && DispVal == 0 && BaseRegNo != N86::EBP) {
2743 // Emit simple indirect register encoding... [EAX] f.e.
2745 // Be pessimistic and assume it's a disp32, not a disp8
2747 // Emit the most general non-SIB encoding: [REG+disp32]
2749 FinalSize += getDisplacementFieldSize(DispForReloc);
2753 } else { // We need a SIB byte, so start by outputting the ModR/M byte first
2754 assert(IndexReg.getReg() != X86::ESP &&
2755 IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!");
2757 bool ForceDisp32 = false;
2758 if (BaseReg == 0 || DispForReloc) {
2759 // Emit the normal disp32 encoding.
2766 FinalSize += sizeSIBByte();
2768 // Do we need to output a displacement?
2769 if (DispVal != 0 || ForceDisp32) {
2770 FinalSize += getDisplacementFieldSize(DispForReloc);
2777 static unsigned GetInstSizeWithDesc(const MachineInstr &MI,
2778 const TargetInstrDesc *Desc,
2779 bool IsPIC, bool Is64BitMode) {
2781 unsigned Opcode = Desc->Opcode;
2782 unsigned FinalSize = 0;
2784 // Emit the lock opcode prefix as needed.
2785 if (Desc->TSFlags & X86II::LOCK) ++FinalSize;
2787 // Emit segment overrid opcode prefix as needed.
2788 switch (Desc->TSFlags & X86II::SegOvrMask) {
2793 default: assert(0 && "Invalid segment!");
2794 case 0: break; // No segment override!
2797 // Emit the repeat opcode prefix as needed.
2798 if ((Desc->TSFlags & X86II::Op0Mask) == X86II::REP) ++FinalSize;
2800 // Emit the operand size opcode prefix as needed.
2801 if (Desc->TSFlags & X86II::OpSize) ++FinalSize;
2803 // Emit the address size opcode prefix as needed.
2804 if (Desc->TSFlags & X86II::AdSize) ++FinalSize;
2806 bool Need0FPrefix = false;
2807 switch (Desc->TSFlags & X86II::Op0Mask) {
2808 case X86II::TB: // Two-byte opcode prefix
2809 case X86II::T8: // 0F 38
2810 case X86II::TA: // 0F 3A
2811 Need0FPrefix = true;
2813 case X86II::REP: break; // already handled.
2814 case X86II::XS: // F3 0F
2816 Need0FPrefix = true;
2818 case X86II::XD: // F2 0F
2820 Need0FPrefix = true;
2822 case X86II::D8: case X86II::D9: case X86II::DA: case X86II::DB:
2823 case X86II::DC: case X86II::DD: case X86II::DE: case X86II::DF:
2825 break; // Two-byte opcode prefix
2826 default: assert(0 && "Invalid prefix!");
2827 case 0: break; // No prefix!
2832 unsigned REX = X86InstrInfo::determineREX(MI);
2837 // 0x0F escape code must be emitted just before the opcode.
2841 switch (Desc->TSFlags & X86II::Op0Mask) {
2842 case X86II::T8: // 0F 38
2845 case X86II::TA: // 0F 3A
2850 // If this is a two-address instruction, skip one of the register operands.
2851 unsigned NumOps = Desc->getNumOperands();
2853 if (NumOps > 1 && Desc->getOperandConstraint(1, TOI::TIED_TO) != -1)
2856 switch (Desc->TSFlags & X86II::FormMask) {
2857 default: assert(0 && "Unknown FormMask value in X86 MachineCodeEmitter!");
2859 // Remember the current PC offset, this is the PIC relocation
2864 case TargetInstrInfo::INLINEASM: {
2865 const MachineFunction *MF = MI.getParent()->getParent();
2866 const char *AsmStr = MI.getOperand(0).getSymbolName();
2867 const TargetAsmInfo* AI = MF->getTarget().getTargetAsmInfo();
2868 FinalSize += AI->getInlineAsmLength(AsmStr);
2871 case TargetInstrInfo::DBG_LABEL:
2872 case TargetInstrInfo::EH_LABEL:
2874 case TargetInstrInfo::IMPLICIT_DEF:
2875 case TargetInstrInfo::DECLARE:
2876 case X86::DWARF_LOC:
2877 case X86::FP_REG_KILL:
2879 case X86::MOVPC32r: {
2880 // This emits the "call" portion of this pseudo instruction.
2882 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2886 case X86::TLS_gs_ri:
2888 FinalSize += sizeGlobalAddress(false);
2896 if (CurOp != NumOps) {
2897 const MachineOperand &MO = MI.getOperand(CurOp++);
2899 FinalSize += sizePCRelativeBlockAddress();
2900 } else if (MO.isGlobal()) {
2901 FinalSize += sizeGlobalAddress(false);
2902 } else if (MO.isSymbol()) {
2903 FinalSize += sizeExternalSymbolAddress(false);
2904 } else if (MO.isImm()) {
2905 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2907 assert(0 && "Unknown RawFrm operand!");
2912 case X86II::AddRegFrm:
2916 if (CurOp != NumOps) {
2917 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2918 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2920 FinalSize += sizeConstant(Size);
2923 if (Opcode == X86::MOV64ri)
2925 if (MO1.isGlobal()) {
2926 FinalSize += sizeGlobalAddress(dword);
2927 } else if (MO1.isSymbol())
2928 FinalSize += sizeExternalSymbolAddress(dword);
2929 else if (MO1.isCPI())
2930 FinalSize += sizeConstPoolAddress(dword);
2931 else if (MO1.isJTI())
2932 FinalSize += sizeJumpTableAddress(dword);
2937 case X86II::MRMDestReg: {
2939 FinalSize += sizeRegModRMByte();
2941 if (CurOp != NumOps) {
2943 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2947 case X86II::MRMDestMem: {
2949 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
2951 if (CurOp != NumOps) {
2953 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2958 case X86II::MRMSrcReg:
2960 FinalSize += sizeRegModRMByte();
2962 if (CurOp != NumOps) {
2964 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2968 case X86II::MRMSrcMem: {
2971 FinalSize += getMemModRMByteSize(MI, CurOp+1, IsPIC, Is64BitMode);
2973 if (CurOp != NumOps) {
2975 FinalSize += sizeConstant(X86InstrInfo::sizeOfImm(Desc));
2980 case X86II::MRM0r: case X86II::MRM1r:
2981 case X86II::MRM2r: case X86II::MRM3r:
2982 case X86II::MRM4r: case X86II::MRM5r:
2983 case X86II::MRM6r: case X86II::MRM7r:
2986 FinalSize += sizeRegModRMByte();
2988 if (CurOp != NumOps) {
2989 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2990 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2992 FinalSize += sizeConstant(Size);
2995 if (Opcode == X86::MOV64ri32)
2997 if (MO1.isGlobal()) {
2998 FinalSize += sizeGlobalAddress(dword);
2999 } else if (MO1.isSymbol())
3000 FinalSize += sizeExternalSymbolAddress(dword);
3001 else if (MO1.isCPI())
3002 FinalSize += sizeConstPoolAddress(dword);
3003 else if (MO1.isJTI())
3004 FinalSize += sizeJumpTableAddress(dword);
3009 case X86II::MRM0m: case X86II::MRM1m:
3010 case X86II::MRM2m: case X86II::MRM3m:
3011 case X86II::MRM4m: case X86II::MRM5m:
3012 case X86II::MRM6m: case X86II::MRM7m: {
3015 FinalSize += getMemModRMByteSize(MI, CurOp, IsPIC, Is64BitMode);
3018 if (CurOp != NumOps) {
3019 const MachineOperand &MO = MI.getOperand(CurOp++);
3020 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
3022 FinalSize += sizeConstant(Size);
3025 if (Opcode == X86::MOV64mi32)
3027 if (MO.isGlobal()) {
3028 FinalSize += sizeGlobalAddress(dword);
3029 } else if (MO.isSymbol())
3030 FinalSize += sizeExternalSymbolAddress(dword);
3031 else if (MO.isCPI())
3032 FinalSize += sizeConstPoolAddress(dword);
3033 else if (MO.isJTI())
3034 FinalSize += sizeJumpTableAddress(dword);
3040 case X86II::MRMInitReg:
3042 // Duplicate register, used by things like MOV8r0 (aka xor reg,reg).
3043 FinalSize += sizeRegModRMByte();
3048 if (!Desc->isVariadic() && CurOp != NumOps) {
3049 cerr << "Cannot determine size: ";
3060 unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
3061 const TargetInstrDesc &Desc = MI->getDesc();
3062 bool IsPIC = (TM.getRelocationModel() == Reloc::PIC_);
3063 bool Is64BitMode = TM.getSubtargetImpl()->is64Bit();
3064 unsigned Size = GetInstSizeWithDesc(*MI, &Desc, IsPIC, Is64BitMode);
3065 if (Desc.getOpcode() == X86::MOVPC32r) {
3066 Size += GetInstSizeWithDesc(*MI, &get(X86::POP32r), IsPIC, Is64BitMode);
3071 /// getGlobalBaseReg - Return a virtual register initialized with the
3072 /// the global base register value. Output instructions required to
3073 /// initialize the register in the function entry block, if necessary.
3075 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3076 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3077 "X86-64 PIC uses RIP relative addressing");
3079 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3080 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3081 if (GlobalBaseReg != 0)
3082 return GlobalBaseReg;
3084 // Insert the set of GlobalBaseReg into the first MBB of the function
3085 MachineBasicBlock &FirstMBB = MF->front();
3086 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3087 DebugLoc DL = DebugLoc::getUnknownLoc();
3088 if (MBBI != FirstMBB.end()) DL = MBBI->getDebugLoc();
3089 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3090 unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3092 const TargetInstrInfo *TII = TM.getInstrInfo();
3093 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3094 // only used in JIT code emission as displacement to pc.
3095 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC)
3098 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3099 // not to pc, but to _GLOBAL_ADDRESS_TABLE_ external
3100 if (TM.getRelocationModel() == Reloc::PIC_ &&
3101 TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3103 RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3104 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
3105 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_");
3110 X86FI->setGlobalBaseReg(GlobalBaseReg);
3111 return GlobalBaseReg;