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 "X86InstrBuilder.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/LLVMContext.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/CodeGen/PseudoSourceValue.h"
29 #include "llvm/MC/MCInst.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/MC/MCAsmInfo.h"
38 #define GET_INSTRINFO_CTOR
39 #include "X86GenInstrInfo.inc"
44 NoFusing("disable-spill-fusing",
45 cl::desc("Disable fusing of spill code into instructions"));
47 PrintFailedFusing("print-failed-fuse-candidates",
48 cl::desc("Print instructions that the allocator wants to"
49 " fuse, but the X86 backend currently can't"),
52 ReMatPICStubLoad("remat-pic-stub-load",
53 cl::desc("Re-materialize load from stub in PIC mode"),
54 cl::init(false), cl::Hidden);
56 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
57 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
58 ? X86::ADJCALLSTACKDOWN64
59 : X86::ADJCALLSTACKDOWN32),
60 (tm.getSubtarget<X86Subtarget>().is64Bit()
61 ? X86::ADJCALLSTACKUP64
62 : X86::ADJCALLSTACKUP32)),
63 TM(tm), RI(tm, *this) {
65 TB_NOT_REVERSABLE = 1U << 31,
66 TB_FLAGS = TB_NOT_REVERSABLE
69 static const unsigned OpTbl2Addr[][2] = {
70 { X86::ADC32ri, X86::ADC32mi },
71 { X86::ADC32ri8, X86::ADC32mi8 },
72 { X86::ADC32rr, X86::ADC32mr },
73 { X86::ADC64ri32, X86::ADC64mi32 },
74 { X86::ADC64ri8, X86::ADC64mi8 },
75 { X86::ADC64rr, X86::ADC64mr },
76 { X86::ADD16ri, X86::ADD16mi },
77 { X86::ADD16ri8, X86::ADD16mi8 },
78 { X86::ADD16ri_DB, X86::ADD16mi | TB_NOT_REVERSABLE },
79 { X86::ADD16ri8_DB, X86::ADD16mi8 | TB_NOT_REVERSABLE },
80 { X86::ADD16rr, X86::ADD16mr },
81 { X86::ADD16rr_DB, X86::ADD16mr | TB_NOT_REVERSABLE },
82 { X86::ADD32ri, X86::ADD32mi },
83 { X86::ADD32ri8, X86::ADD32mi8 },
84 { X86::ADD32ri_DB, X86::ADD32mi | TB_NOT_REVERSABLE },
85 { X86::ADD32ri8_DB, X86::ADD32mi8 | TB_NOT_REVERSABLE },
86 { X86::ADD32rr, X86::ADD32mr },
87 { X86::ADD32rr_DB, X86::ADD32mr | TB_NOT_REVERSABLE },
88 { X86::ADD64ri32, X86::ADD64mi32 },
89 { X86::ADD64ri8, X86::ADD64mi8 },
90 { X86::ADD64ri32_DB,X86::ADD64mi32 | TB_NOT_REVERSABLE },
91 { X86::ADD64ri8_DB, X86::ADD64mi8 | TB_NOT_REVERSABLE },
92 { X86::ADD64rr, X86::ADD64mr },
93 { X86::ADD64rr_DB, X86::ADD64mr | TB_NOT_REVERSABLE },
94 { X86::ADD8ri, X86::ADD8mi },
95 { X86::ADD8rr, X86::ADD8mr },
96 { X86::AND16ri, X86::AND16mi },
97 { X86::AND16ri8, X86::AND16mi8 },
98 { X86::AND16rr, X86::AND16mr },
99 { X86::AND32ri, X86::AND32mi },
100 { X86::AND32ri8, X86::AND32mi8 },
101 { X86::AND32rr, X86::AND32mr },
102 { X86::AND64ri32, X86::AND64mi32 },
103 { X86::AND64ri8, X86::AND64mi8 },
104 { X86::AND64rr, X86::AND64mr },
105 { X86::AND8ri, X86::AND8mi },
106 { X86::AND8rr, X86::AND8mr },
107 { X86::DEC16r, X86::DEC16m },
108 { X86::DEC32r, X86::DEC32m },
109 { X86::DEC64_16r, X86::DEC64_16m },
110 { X86::DEC64_32r, X86::DEC64_32m },
111 { X86::DEC64r, X86::DEC64m },
112 { X86::DEC8r, X86::DEC8m },
113 { X86::INC16r, X86::INC16m },
114 { X86::INC32r, X86::INC32m },
115 { X86::INC64_16r, X86::INC64_16m },
116 { X86::INC64_32r, X86::INC64_32m },
117 { X86::INC64r, X86::INC64m },
118 { X86::INC8r, X86::INC8m },
119 { X86::NEG16r, X86::NEG16m },
120 { X86::NEG32r, X86::NEG32m },
121 { X86::NEG64r, X86::NEG64m },
122 { X86::NEG8r, X86::NEG8m },
123 { X86::NOT16r, X86::NOT16m },
124 { X86::NOT32r, X86::NOT32m },
125 { X86::NOT64r, X86::NOT64m },
126 { X86::NOT8r, X86::NOT8m },
127 { X86::OR16ri, X86::OR16mi },
128 { X86::OR16ri8, X86::OR16mi8 },
129 { X86::OR16rr, X86::OR16mr },
130 { X86::OR32ri, X86::OR32mi },
131 { X86::OR32ri8, X86::OR32mi8 },
132 { X86::OR32rr, X86::OR32mr },
133 { X86::OR64ri32, X86::OR64mi32 },
134 { X86::OR64ri8, X86::OR64mi8 },
135 { X86::OR64rr, X86::OR64mr },
136 { X86::OR8ri, X86::OR8mi },
137 { X86::OR8rr, X86::OR8mr },
138 { X86::ROL16r1, X86::ROL16m1 },
139 { X86::ROL16rCL, X86::ROL16mCL },
140 { X86::ROL16ri, X86::ROL16mi },
141 { X86::ROL32r1, X86::ROL32m1 },
142 { X86::ROL32rCL, X86::ROL32mCL },
143 { X86::ROL32ri, X86::ROL32mi },
144 { X86::ROL64r1, X86::ROL64m1 },
145 { X86::ROL64rCL, X86::ROL64mCL },
146 { X86::ROL64ri, X86::ROL64mi },
147 { X86::ROL8r1, X86::ROL8m1 },
148 { X86::ROL8rCL, X86::ROL8mCL },
149 { X86::ROL8ri, X86::ROL8mi },
150 { X86::ROR16r1, X86::ROR16m1 },
151 { X86::ROR16rCL, X86::ROR16mCL },
152 { X86::ROR16ri, X86::ROR16mi },
153 { X86::ROR32r1, X86::ROR32m1 },
154 { X86::ROR32rCL, X86::ROR32mCL },
155 { X86::ROR32ri, X86::ROR32mi },
156 { X86::ROR64r1, X86::ROR64m1 },
157 { X86::ROR64rCL, X86::ROR64mCL },
158 { X86::ROR64ri, X86::ROR64mi },
159 { X86::ROR8r1, X86::ROR8m1 },
160 { X86::ROR8rCL, X86::ROR8mCL },
161 { X86::ROR8ri, X86::ROR8mi },
162 { X86::SAR16r1, X86::SAR16m1 },
163 { X86::SAR16rCL, X86::SAR16mCL },
164 { X86::SAR16ri, X86::SAR16mi },
165 { X86::SAR32r1, X86::SAR32m1 },
166 { X86::SAR32rCL, X86::SAR32mCL },
167 { X86::SAR32ri, X86::SAR32mi },
168 { X86::SAR64r1, X86::SAR64m1 },
169 { X86::SAR64rCL, X86::SAR64mCL },
170 { X86::SAR64ri, X86::SAR64mi },
171 { X86::SAR8r1, X86::SAR8m1 },
172 { X86::SAR8rCL, X86::SAR8mCL },
173 { X86::SAR8ri, X86::SAR8mi },
174 { X86::SBB32ri, X86::SBB32mi },
175 { X86::SBB32ri8, X86::SBB32mi8 },
176 { X86::SBB32rr, X86::SBB32mr },
177 { X86::SBB64ri32, X86::SBB64mi32 },
178 { X86::SBB64ri8, X86::SBB64mi8 },
179 { X86::SBB64rr, X86::SBB64mr },
180 { X86::SHL16rCL, X86::SHL16mCL },
181 { X86::SHL16ri, X86::SHL16mi },
182 { X86::SHL32rCL, X86::SHL32mCL },
183 { X86::SHL32ri, X86::SHL32mi },
184 { X86::SHL64rCL, X86::SHL64mCL },
185 { X86::SHL64ri, X86::SHL64mi },
186 { X86::SHL8rCL, X86::SHL8mCL },
187 { X86::SHL8ri, X86::SHL8mi },
188 { X86::SHLD16rrCL, X86::SHLD16mrCL },
189 { X86::SHLD16rri8, X86::SHLD16mri8 },
190 { X86::SHLD32rrCL, X86::SHLD32mrCL },
191 { X86::SHLD32rri8, X86::SHLD32mri8 },
192 { X86::SHLD64rrCL, X86::SHLD64mrCL },
193 { X86::SHLD64rri8, X86::SHLD64mri8 },
194 { X86::SHR16r1, X86::SHR16m1 },
195 { X86::SHR16rCL, X86::SHR16mCL },
196 { X86::SHR16ri, X86::SHR16mi },
197 { X86::SHR32r1, X86::SHR32m1 },
198 { X86::SHR32rCL, X86::SHR32mCL },
199 { X86::SHR32ri, X86::SHR32mi },
200 { X86::SHR64r1, X86::SHR64m1 },
201 { X86::SHR64rCL, X86::SHR64mCL },
202 { X86::SHR64ri, X86::SHR64mi },
203 { X86::SHR8r1, X86::SHR8m1 },
204 { X86::SHR8rCL, X86::SHR8mCL },
205 { X86::SHR8ri, X86::SHR8mi },
206 { X86::SHRD16rrCL, X86::SHRD16mrCL },
207 { X86::SHRD16rri8, X86::SHRD16mri8 },
208 { X86::SHRD32rrCL, X86::SHRD32mrCL },
209 { X86::SHRD32rri8, X86::SHRD32mri8 },
210 { X86::SHRD64rrCL, X86::SHRD64mrCL },
211 { X86::SHRD64rri8, X86::SHRD64mri8 },
212 { X86::SUB16ri, X86::SUB16mi },
213 { X86::SUB16ri8, X86::SUB16mi8 },
214 { X86::SUB16rr, X86::SUB16mr },
215 { X86::SUB32ri, X86::SUB32mi },
216 { X86::SUB32ri8, X86::SUB32mi8 },
217 { X86::SUB32rr, X86::SUB32mr },
218 { X86::SUB64ri32, X86::SUB64mi32 },
219 { X86::SUB64ri8, X86::SUB64mi8 },
220 { X86::SUB64rr, X86::SUB64mr },
221 { X86::SUB8ri, X86::SUB8mi },
222 { X86::SUB8rr, X86::SUB8mr },
223 { X86::XOR16ri, X86::XOR16mi },
224 { X86::XOR16ri8, X86::XOR16mi8 },
225 { X86::XOR16rr, X86::XOR16mr },
226 { X86::XOR32ri, X86::XOR32mi },
227 { X86::XOR32ri8, X86::XOR32mi8 },
228 { X86::XOR32rr, X86::XOR32mr },
229 { X86::XOR64ri32, X86::XOR64mi32 },
230 { X86::XOR64ri8, X86::XOR64mi8 },
231 { X86::XOR64rr, X86::XOR64mr },
232 { X86::XOR8ri, X86::XOR8mi },
233 { X86::XOR8rr, X86::XOR8mr }
236 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
237 unsigned RegOp = OpTbl2Addr[i][0];
238 unsigned MemOp = OpTbl2Addr[i][1] & ~TB_FLAGS;
239 assert(!RegOp2MemOpTable2Addr.count(RegOp) && "Duplicated entries?");
240 RegOp2MemOpTable2Addr[RegOp] = std::make_pair(MemOp, 0U);
242 // If this is not a reversible operation (because there is a many->one)
243 // mapping, don't insert the reverse of the operation into MemOp2RegOpTable.
244 if (OpTbl2Addr[i][1] & TB_NOT_REVERSABLE)
247 // Index 0, folded load and store, no alignment requirement.
248 unsigned AuxInfo = 0 | (1 << 4) | (1 << 5);
250 assert(!MemOp2RegOpTable.count(MemOp) &&
251 "Duplicated entries in unfolding maps?");
252 MemOp2RegOpTable[MemOp] = std::make_pair(RegOp, AuxInfo);
255 // If the third value is 1, then it's folding either a load or a store.
256 static const unsigned OpTbl0[][4] = {
257 { X86::BT16ri8, X86::BT16mi8, 1, 0 },
258 { X86::BT32ri8, X86::BT32mi8, 1, 0 },
259 { X86::BT64ri8, X86::BT64mi8, 1, 0 },
260 { X86::CALL32r, X86::CALL32m, 1, 0 },
261 { X86::CALL64r, X86::CALL64m, 1, 0 },
262 { X86::WINCALL64r, X86::WINCALL64m, 1, 0 },
263 { X86::CMP16ri, X86::CMP16mi, 1, 0 },
264 { X86::CMP16ri8, X86::CMP16mi8, 1, 0 },
265 { X86::CMP16rr, X86::CMP16mr, 1, 0 },
266 { X86::CMP32ri, X86::CMP32mi, 1, 0 },
267 { X86::CMP32ri8, X86::CMP32mi8, 1, 0 },
268 { X86::CMP32rr, X86::CMP32mr, 1, 0 },
269 { X86::CMP64ri32, X86::CMP64mi32, 1, 0 },
270 { X86::CMP64ri8, X86::CMP64mi8, 1, 0 },
271 { X86::CMP64rr, X86::CMP64mr, 1, 0 },
272 { X86::CMP8ri, X86::CMP8mi, 1, 0 },
273 { X86::CMP8rr, X86::CMP8mr, 1, 0 },
274 { X86::DIV16r, X86::DIV16m, 1, 0 },
275 { X86::DIV32r, X86::DIV32m, 1, 0 },
276 { X86::DIV64r, X86::DIV64m, 1, 0 },
277 { X86::DIV8r, X86::DIV8m, 1, 0 },
278 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, 0, 16 },
279 { X86::FsMOVAPDrr, X86::MOVSDmr | TB_NOT_REVERSABLE , 0, 0 },
280 { X86::FsMOVAPSrr, X86::MOVSSmr | TB_NOT_REVERSABLE , 0, 0 },
281 { X86::IDIV16r, X86::IDIV16m, 1, 0 },
282 { X86::IDIV32r, X86::IDIV32m, 1, 0 },
283 { X86::IDIV64r, X86::IDIV64m, 1, 0 },
284 { X86::IDIV8r, X86::IDIV8m, 1, 0 },
285 { X86::IMUL16r, X86::IMUL16m, 1, 0 },
286 { X86::IMUL32r, X86::IMUL32m, 1, 0 },
287 { X86::IMUL64r, X86::IMUL64m, 1, 0 },
288 { X86::IMUL8r, X86::IMUL8m, 1, 0 },
289 { X86::JMP32r, X86::JMP32m, 1, 0 },
290 { X86::JMP64r, X86::JMP64m, 1, 0 },
291 { X86::MOV16ri, X86::MOV16mi, 0, 0 },
292 { X86::MOV16rr, X86::MOV16mr, 0, 0 },
293 { X86::MOV32ri, X86::MOV32mi, 0, 0 },
294 { X86::MOV32rr, X86::MOV32mr, 0, 0 },
295 { X86::MOV64ri32, X86::MOV64mi32, 0, 0 },
296 { X86::MOV64rr, X86::MOV64mr, 0, 0 },
297 { X86::MOV8ri, X86::MOV8mi, 0, 0 },
298 { X86::MOV8rr, X86::MOV8mr, 0, 0 },
299 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, 0, 0 },
300 { X86::MOVAPDrr, X86::MOVAPDmr, 0, 16 },
301 { X86::MOVAPSrr, X86::MOVAPSmr, 0, 16 },
302 { X86::MOVDQArr, X86::MOVDQAmr, 0, 16 },
303 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, 0, 32 },
304 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, 0, 32 },
305 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, 0, 32 },
306 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0, 0 },
307 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0, 0 },
308 { X86::MOVSDto64rr, X86::MOVSDto64mr, 0, 0 },
309 { X86::MOVSS2DIrr, X86::MOVSS2DImr, 0, 0 },
310 { X86::MOVUPDrr, X86::MOVUPDmr, 0, 0 },
311 { X86::MOVUPSrr, X86::MOVUPSmr, 0, 0 },
312 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, 0, 0 },
313 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, 0, 0 },
314 { X86::MUL16r, X86::MUL16m, 1, 0 },
315 { X86::MUL32r, X86::MUL32m, 1, 0 },
316 { X86::MUL64r, X86::MUL64m, 1, 0 },
317 { X86::MUL8r, X86::MUL8m, 1, 0 },
318 { X86::SETAEr, X86::SETAEm, 0, 0 },
319 { X86::SETAr, X86::SETAm, 0, 0 },
320 { X86::SETBEr, X86::SETBEm, 0, 0 },
321 { X86::SETBr, X86::SETBm, 0, 0 },
322 { X86::SETEr, X86::SETEm, 0, 0 },
323 { X86::SETGEr, X86::SETGEm, 0, 0 },
324 { X86::SETGr, X86::SETGm, 0, 0 },
325 { X86::SETLEr, X86::SETLEm, 0, 0 },
326 { X86::SETLr, X86::SETLm, 0, 0 },
327 { X86::SETNEr, X86::SETNEm, 0, 0 },
328 { X86::SETNOr, X86::SETNOm, 0, 0 },
329 { X86::SETNPr, X86::SETNPm, 0, 0 },
330 { X86::SETNSr, X86::SETNSm, 0, 0 },
331 { X86::SETOr, X86::SETOm, 0, 0 },
332 { X86::SETPr, X86::SETPm, 0, 0 },
333 { X86::SETSr, X86::SETSm, 0, 0 },
334 { X86::TAILJMPr, X86::TAILJMPm, 1, 0 },
335 { X86::TAILJMPr64, X86::TAILJMPm64, 1, 0 },
336 { X86::TEST16ri, X86::TEST16mi, 1, 0 },
337 { X86::TEST32ri, X86::TEST32mi, 1, 0 },
338 { X86::TEST64ri32, X86::TEST64mi32, 1, 0 },
339 { X86::TEST8ri, X86::TEST8mi, 1, 0 }
342 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
343 unsigned RegOp = OpTbl0[i][0];
344 unsigned MemOp = OpTbl0[i][1] & ~TB_FLAGS;
345 unsigned FoldedLoad = OpTbl0[i][2];
346 unsigned Align = OpTbl0[i][3];
347 assert(!RegOp2MemOpTable0.count(RegOp) && "Duplicated entries?");
348 RegOp2MemOpTable0[RegOp] = std::make_pair(MemOp, Align);
350 // If this is not a reversible operation (because there is a many->one)
351 // mapping, don't insert the reverse of the operation into MemOp2RegOpTable.
352 if (OpTbl0[i][1] & TB_NOT_REVERSABLE)
355 // Index 0, folded load or store.
356 unsigned AuxInfo = 0 | (FoldedLoad << 4) | ((FoldedLoad^1) << 5);
357 assert(!MemOp2RegOpTable.count(MemOp) && "Duplicated entries?");
358 MemOp2RegOpTable[MemOp] = std::make_pair(RegOp, AuxInfo);
361 static const unsigned OpTbl1[][3] = {
362 { X86::CMP16rr, X86::CMP16rm, 0 },
363 { X86::CMP32rr, X86::CMP32rm, 0 },
364 { X86::CMP64rr, X86::CMP64rm, 0 },
365 { X86::CMP8rr, X86::CMP8rm, 0 },
366 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
367 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
368 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
369 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
370 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
371 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
372 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
373 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
374 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
375 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
376 { X86::FsMOVAPDrr, X86::MOVSDrm | TB_NOT_REVERSABLE , 0 },
377 { X86::FsMOVAPSrr, X86::MOVSSrm | TB_NOT_REVERSABLE , 0 },
378 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
379 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
380 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
381 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
382 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
383 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
384 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
385 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
386 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm, 16 },
387 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm, 16 },
388 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm, 16 },
389 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm, 16 },
390 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm, 16 },
391 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm, 0 },
392 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
393 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
394 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
395 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
396 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
397 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
398 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
399 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
400 { X86::Int_CVTSS2SI64rr,X86::Int_CVTSS2SI64rm, 0 },
401 { X86::Int_CVTSS2SIrr, X86::Int_CVTSS2SIrm, 0 },
402 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, 16 },
403 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, 16 },
404 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
405 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
406 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
407 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
408 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
409 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
410 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
411 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
412 { X86::MOV16rr, X86::MOV16rm, 0 },
413 { X86::MOV32rr, X86::MOV32rm, 0 },
414 { X86::MOV64rr, X86::MOV64rm, 0 },
415 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
416 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
417 { X86::MOV8rr, X86::MOV8rm, 0 },
418 { X86::MOVAPDrr, X86::MOVAPDrm, 16 },
419 { X86::MOVAPSrr, X86::MOVAPSrm, 16 },
420 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, 32 },
421 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, 32 },
422 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
423 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
424 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
425 { X86::MOVDQArr, X86::MOVDQArm, 16 },
426 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, 16 },
427 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, 16 },
428 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, 16 },
429 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
430 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
431 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
432 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
433 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
434 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
435 { X86::MOVUPDrr, X86::MOVUPDrm, 16 },
436 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
437 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
438 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
439 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
440 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
441 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, 16 },
442 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
443 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
444 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
445 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
446 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
447 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
448 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
449 { X86::PSHUFDri, X86::PSHUFDmi, 16 },
450 { X86::PSHUFHWri, X86::PSHUFHWmi, 16 },
451 { X86::PSHUFLWri, X86::PSHUFLWmi, 16 },
452 { X86::RCPPSr, X86::RCPPSm, 16 },
453 { X86::RCPPSr_Int, X86::RCPPSm_Int, 16 },
454 { X86::RSQRTPSr, X86::RSQRTPSm, 16 },
455 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, 16 },
456 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
457 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
458 { X86::SQRTPDr, X86::SQRTPDm, 16 },
459 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, 16 },
460 { X86::SQRTPSr, X86::SQRTPSm, 16 },
461 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, 16 },
462 { X86::SQRTSDr, X86::SQRTSDm, 0 },
463 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
464 { X86::SQRTSSr, X86::SQRTSSm, 0 },
465 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
466 { X86::TEST16rr, X86::TEST16rm, 0 },
467 { X86::TEST32rr, X86::TEST32rm, 0 },
468 { X86::TEST64rr, X86::TEST64rm, 0 },
469 { X86::TEST8rr, X86::TEST8rm, 0 },
470 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
471 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
472 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
473 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
474 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 }
477 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
478 unsigned RegOp = OpTbl1[i][0];
479 unsigned MemOp = OpTbl1[i][1] & ~TB_FLAGS;
480 unsigned Align = OpTbl1[i][2];
481 assert(!RegOp2MemOpTable1.count(RegOp) && "Duplicate entries");
482 RegOp2MemOpTable1[RegOp] = std::make_pair(MemOp, Align);
484 // If this is not a reversible operation (because there is a many->one)
485 // mapping, don't insert the reverse of the operation into MemOp2RegOpTable.
486 if (OpTbl1[i][1] & TB_NOT_REVERSABLE)
489 // Index 1, folded load
490 unsigned AuxInfo = 1 | (1 << 4);
491 assert(!MemOp2RegOpTable.count(MemOp) && "Duplicate entries");
492 MemOp2RegOpTable[MemOp] = std::make_pair(RegOp, AuxInfo);
495 static const unsigned OpTbl2[][3] = {
496 { X86::ADC32rr, X86::ADC32rm, 0 },
497 { X86::ADC64rr, X86::ADC64rm, 0 },
498 { X86::ADD16rr, X86::ADD16rm, 0 },
499 { X86::ADD16rr_DB, X86::ADD16rm | TB_NOT_REVERSABLE, 0 },
500 { X86::ADD32rr, X86::ADD32rm, 0 },
501 { X86::ADD32rr_DB, X86::ADD32rm | TB_NOT_REVERSABLE, 0 },
502 { X86::ADD64rr, X86::ADD64rm, 0 },
503 { X86::ADD64rr_DB, X86::ADD64rm | TB_NOT_REVERSABLE, 0 },
504 { X86::ADD8rr, X86::ADD8rm, 0 },
505 { X86::ADDPDrr, X86::ADDPDrm, 16 },
506 { X86::ADDPSrr, X86::ADDPSrm, 16 },
507 { X86::ADDSDrr, X86::ADDSDrm, 0 },
508 { X86::ADDSSrr, X86::ADDSSrm, 0 },
509 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, 16 },
510 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, 16 },
511 { X86::AND16rr, X86::AND16rm, 0 },
512 { X86::AND32rr, X86::AND32rm, 0 },
513 { X86::AND64rr, X86::AND64rm, 0 },
514 { X86::AND8rr, X86::AND8rm, 0 },
515 { X86::ANDNPDrr, X86::ANDNPDrm, 16 },
516 { X86::ANDNPSrr, X86::ANDNPSrm, 16 },
517 { X86::ANDPDrr, X86::ANDPDrm, 16 },
518 { X86::ANDPSrr, X86::ANDPSrm, 16 },
519 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
520 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
521 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
522 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
523 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
524 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
525 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
526 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
527 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
528 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
529 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
530 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
531 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
532 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
533 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
534 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
535 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
536 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
537 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
538 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
539 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
540 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
541 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
542 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
543 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
544 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
545 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
546 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
547 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
548 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
549 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
550 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
551 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
552 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
553 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
554 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
555 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
556 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
557 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
558 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
559 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
560 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
561 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
562 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
563 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
564 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
565 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
566 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
567 { X86::CMPPDrri, X86::CMPPDrmi, 16 },
568 { X86::CMPPSrri, X86::CMPPSrmi, 16 },
569 { X86::CMPSDrr, X86::CMPSDrm, 0 },
570 { X86::CMPSSrr, X86::CMPSSrm, 0 },
571 { X86::DIVPDrr, X86::DIVPDrm, 16 },
572 { X86::DIVPSrr, X86::DIVPSrm, 16 },
573 { X86::DIVSDrr, X86::DIVSDrm, 0 },
574 { X86::DIVSSrr, X86::DIVSSrm, 0 },
575 { X86::FsANDNPDrr, X86::FsANDNPDrm, 16 },
576 { X86::FsANDNPSrr, X86::FsANDNPSrm, 16 },
577 { X86::FsANDPDrr, X86::FsANDPDrm, 16 },
578 { X86::FsANDPSrr, X86::FsANDPSrm, 16 },
579 { X86::FsORPDrr, X86::FsORPDrm, 16 },
580 { X86::FsORPSrr, X86::FsORPSrm, 16 },
581 { X86::FsXORPDrr, X86::FsXORPDrm, 16 },
582 { X86::FsXORPSrr, X86::FsXORPSrm, 16 },
583 { X86::HADDPDrr, X86::HADDPDrm, 16 },
584 { X86::HADDPSrr, X86::HADDPSrm, 16 },
585 { X86::HSUBPDrr, X86::HSUBPDrm, 16 },
586 { X86::HSUBPSrr, X86::HSUBPSrm, 16 },
587 { X86::IMUL16rr, X86::IMUL16rm, 0 },
588 { X86::IMUL32rr, X86::IMUL32rm, 0 },
589 { X86::IMUL64rr, X86::IMUL64rm, 0 },
590 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
591 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
592 { X86::MAXPDrr, X86::MAXPDrm, 16 },
593 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, 16 },
594 { X86::MAXPSrr, X86::MAXPSrm, 16 },
595 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, 16 },
596 { X86::MAXSDrr, X86::MAXSDrm, 0 },
597 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
598 { X86::MAXSSrr, X86::MAXSSrm, 0 },
599 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
600 { X86::MINPDrr, X86::MINPDrm, 16 },
601 { X86::MINPDrr_Int, X86::MINPDrm_Int, 16 },
602 { X86::MINPSrr, X86::MINPSrm, 16 },
603 { X86::MINPSrr_Int, X86::MINPSrm_Int, 16 },
604 { X86::MINSDrr, X86::MINSDrm, 0 },
605 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
606 { X86::MINSSrr, X86::MINSSrm, 0 },
607 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
608 { X86::MULPDrr, X86::MULPDrm, 16 },
609 { X86::MULPSrr, X86::MULPSrm, 16 },
610 { X86::MULSDrr, X86::MULSDrm, 0 },
611 { X86::MULSSrr, X86::MULSSrm, 0 },
612 { X86::OR16rr, X86::OR16rm, 0 },
613 { X86::OR32rr, X86::OR32rm, 0 },
614 { X86::OR64rr, X86::OR64rm, 0 },
615 { X86::OR8rr, X86::OR8rm, 0 },
616 { X86::ORPDrr, X86::ORPDrm, 16 },
617 { X86::ORPSrr, X86::ORPSrm, 16 },
618 { X86::PACKSSDWrr, X86::PACKSSDWrm, 16 },
619 { X86::PACKSSWBrr, X86::PACKSSWBrm, 16 },
620 { X86::PACKUSWBrr, X86::PACKUSWBrm, 16 },
621 { X86::PADDBrr, X86::PADDBrm, 16 },
622 { X86::PADDDrr, X86::PADDDrm, 16 },
623 { X86::PADDQrr, X86::PADDQrm, 16 },
624 { X86::PADDSBrr, X86::PADDSBrm, 16 },
625 { X86::PADDSWrr, X86::PADDSWrm, 16 },
626 { X86::PADDWrr, X86::PADDWrm, 16 },
627 { X86::PANDNrr, X86::PANDNrm, 16 },
628 { X86::PANDrr, X86::PANDrm, 16 },
629 { X86::PAVGBrr, X86::PAVGBrm, 16 },
630 { X86::PAVGWrr, X86::PAVGWrm, 16 },
631 { X86::PCMPEQBrr, X86::PCMPEQBrm, 16 },
632 { X86::PCMPEQDrr, X86::PCMPEQDrm, 16 },
633 { X86::PCMPEQWrr, X86::PCMPEQWrm, 16 },
634 { X86::PCMPGTBrr, X86::PCMPGTBrm, 16 },
635 { X86::PCMPGTDrr, X86::PCMPGTDrm, 16 },
636 { X86::PCMPGTWrr, X86::PCMPGTWrm, 16 },
637 { X86::PINSRWrri, X86::PINSRWrmi, 16 },
638 { X86::PMADDWDrr, X86::PMADDWDrm, 16 },
639 { X86::PMAXSWrr, X86::PMAXSWrm, 16 },
640 { X86::PMAXUBrr, X86::PMAXUBrm, 16 },
641 { X86::PMINSWrr, X86::PMINSWrm, 16 },
642 { X86::PMINUBrr, X86::PMINUBrm, 16 },
643 { X86::PMULDQrr, X86::PMULDQrm, 16 },
644 { X86::PMULHUWrr, X86::PMULHUWrm, 16 },
645 { X86::PMULHWrr, X86::PMULHWrm, 16 },
646 { X86::PMULLDrr, X86::PMULLDrm, 16 },
647 { X86::PMULLWrr, X86::PMULLWrm, 16 },
648 { X86::PMULUDQrr, X86::PMULUDQrm, 16 },
649 { X86::PORrr, X86::PORrm, 16 },
650 { X86::PSADBWrr, X86::PSADBWrm, 16 },
651 { X86::PSLLDrr, X86::PSLLDrm, 16 },
652 { X86::PSLLQrr, X86::PSLLQrm, 16 },
653 { X86::PSLLWrr, X86::PSLLWrm, 16 },
654 { X86::PSRADrr, X86::PSRADrm, 16 },
655 { X86::PSRAWrr, X86::PSRAWrm, 16 },
656 { X86::PSRLDrr, X86::PSRLDrm, 16 },
657 { X86::PSRLQrr, X86::PSRLQrm, 16 },
658 { X86::PSRLWrr, X86::PSRLWrm, 16 },
659 { X86::PSUBBrr, X86::PSUBBrm, 16 },
660 { X86::PSUBDrr, X86::PSUBDrm, 16 },
661 { X86::PSUBSBrr, X86::PSUBSBrm, 16 },
662 { X86::PSUBSWrr, X86::PSUBSWrm, 16 },
663 { X86::PSUBWrr, X86::PSUBWrm, 16 },
664 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, 16 },
665 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, 16 },
666 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, 16 },
667 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, 16 },
668 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, 16 },
669 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, 16 },
670 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, 16 },
671 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, 16 },
672 { X86::PXORrr, X86::PXORrm, 16 },
673 { X86::SBB32rr, X86::SBB32rm, 0 },
674 { X86::SBB64rr, X86::SBB64rm, 0 },
675 { X86::SHUFPDrri, X86::SHUFPDrmi, 16 },
676 { X86::SHUFPSrri, X86::SHUFPSrmi, 16 },
677 { X86::SUB16rr, X86::SUB16rm, 0 },
678 { X86::SUB32rr, X86::SUB32rm, 0 },
679 { X86::SUB64rr, X86::SUB64rm, 0 },
680 { X86::SUB8rr, X86::SUB8rm, 0 },
681 { X86::SUBPDrr, X86::SUBPDrm, 16 },
682 { X86::SUBPSrr, X86::SUBPSrm, 16 },
683 { X86::SUBSDrr, X86::SUBSDrm, 0 },
684 { X86::SUBSSrr, X86::SUBSSrm, 0 },
685 // FIXME: TEST*rr -> swapped operand of TEST*mr.
686 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, 16 },
687 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, 16 },
688 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, 16 },
689 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, 16 },
690 { X86::XOR16rr, X86::XOR16rm, 0 },
691 { X86::XOR32rr, X86::XOR32rm, 0 },
692 { X86::XOR64rr, X86::XOR64rm, 0 },
693 { X86::XOR8rr, X86::XOR8rm, 0 },
694 { X86::XORPDrr, X86::XORPDrm, 16 },
695 { X86::XORPSrr, X86::XORPSrm, 16 }
698 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
699 unsigned RegOp = OpTbl2[i][0];
700 unsigned MemOp = OpTbl2[i][1] & ~TB_FLAGS;
701 unsigned Align = OpTbl2[i][2];
703 assert(!RegOp2MemOpTable2.count(RegOp) && "Duplicate entry!");
704 RegOp2MemOpTable2[RegOp] = std::make_pair(MemOp, Align);
706 // If this is not a reversible operation (because there is a many->one)
707 // mapping, don't insert the reverse of the operation into MemOp2RegOpTable.
708 if (OpTbl2[i][1] & TB_NOT_REVERSABLE)
711 // Index 2, folded load
712 unsigned AuxInfo = 2 | (1 << 4);
713 assert(!MemOp2RegOpTable.count(MemOp) &&
714 "Duplicated entries in unfolding maps?");
715 MemOp2RegOpTable[MemOp] = std::make_pair(RegOp, AuxInfo);
720 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
721 unsigned &SrcReg, unsigned &DstReg,
722 unsigned &SubIdx) const {
723 switch (MI.getOpcode()) {
725 case X86::MOVSX16rr8:
726 case X86::MOVZX16rr8:
727 case X86::MOVSX32rr8:
728 case X86::MOVZX32rr8:
729 case X86::MOVSX64rr8:
730 case X86::MOVZX64rr8:
731 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
732 // It's not always legal to reference the low 8-bit of the larger
733 // register in 32-bit mode.
735 case X86::MOVSX32rr16:
736 case X86::MOVZX32rr16:
737 case X86::MOVSX64rr16:
738 case X86::MOVZX64rr16:
739 case X86::MOVSX64rr32:
740 case X86::MOVZX64rr32: {
741 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
744 SrcReg = MI.getOperand(1).getReg();
745 DstReg = MI.getOperand(0).getReg();
746 switch (MI.getOpcode()) {
750 case X86::MOVSX16rr8:
751 case X86::MOVZX16rr8:
752 case X86::MOVSX32rr8:
753 case X86::MOVZX32rr8:
754 case X86::MOVSX64rr8:
755 case X86::MOVZX64rr8:
756 SubIdx = X86::sub_8bit;
758 case X86::MOVSX32rr16:
759 case X86::MOVZX32rr16:
760 case X86::MOVSX64rr16:
761 case X86::MOVZX64rr16:
762 SubIdx = X86::sub_16bit;
764 case X86::MOVSX64rr32:
765 case X86::MOVZX64rr32:
766 SubIdx = X86::sub_32bit;
775 /// isFrameOperand - Return true and the FrameIndex if the specified
776 /// operand and follow operands form a reference to the stack frame.
777 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
778 int &FrameIndex) const {
779 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
780 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
781 MI->getOperand(Op+1).getImm() == 1 &&
782 MI->getOperand(Op+2).getReg() == 0 &&
783 MI->getOperand(Op+3).getImm() == 0) {
784 FrameIndex = MI->getOperand(Op).getIndex();
790 static bool isFrameLoadOpcode(int Opcode) {
803 case X86::VMOVAPSYrm:
804 case X86::VMOVAPDYrm:
805 case X86::VMOVDQAYrm:
806 case X86::MMX_MOVD64rm:
807 case X86::MMX_MOVQ64rm:
814 static bool isFrameStoreOpcode(int Opcode) {
827 case X86::VMOVAPSYmr:
828 case X86::VMOVAPDYmr:
829 case X86::VMOVDQAYmr:
830 case X86::MMX_MOVD64mr:
831 case X86::MMX_MOVQ64mr:
832 case X86::MMX_MOVNTQmr:
838 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
839 int &FrameIndex) const {
840 if (isFrameLoadOpcode(MI->getOpcode()))
841 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
842 return MI->getOperand(0).getReg();
846 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
847 int &FrameIndex) const {
848 if (isFrameLoadOpcode(MI->getOpcode())) {
850 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
852 // Check for post-frame index elimination operations
853 const MachineMemOperand *Dummy;
854 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
859 bool X86InstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
860 const MachineMemOperand *&MMO,
861 int &FrameIndex) const {
862 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
863 oe = MI->memoperands_end();
866 if ((*o)->isLoad() && (*o)->getValue())
867 if (const FixedStackPseudoSourceValue *Value =
868 dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) {
869 FrameIndex = Value->getFrameIndex();
877 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
878 int &FrameIndex) const {
879 if (isFrameStoreOpcode(MI->getOpcode()))
880 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
881 isFrameOperand(MI, 0, FrameIndex))
882 return MI->getOperand(X86::AddrNumOperands).getReg();
886 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
887 int &FrameIndex) const {
888 if (isFrameStoreOpcode(MI->getOpcode())) {
890 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
892 // Check for post-frame index elimination operations
893 const MachineMemOperand *Dummy;
894 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
899 bool X86InstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
900 const MachineMemOperand *&MMO,
901 int &FrameIndex) const {
902 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
903 oe = MI->memoperands_end();
906 if ((*o)->isStore() && (*o)->getValue())
907 if (const FixedStackPseudoSourceValue *Value =
908 dyn_cast<const FixedStackPseudoSourceValue>((*o)->getValue())) {
909 FrameIndex = Value->getFrameIndex();
917 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
919 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
920 bool isPICBase = false;
921 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
922 E = MRI.def_end(); I != E; ++I) {
923 MachineInstr *DefMI = I.getOperand().getParent();
924 if (DefMI->getOpcode() != X86::MOVPC32r)
926 assert(!isPICBase && "More than one PIC base?");
933 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
934 AliasAnalysis *AA) const {
935 switch (MI->getOpcode()) {
948 case X86::VMOVAPSYrm:
949 case X86::VMOVUPSYrm:
950 case X86::VMOVAPDYrm:
951 case X86::VMOVDQAYrm:
952 case X86::MMX_MOVD64rm:
953 case X86::MMX_MOVQ64rm:
954 case X86::FsMOVAPSrm:
955 case X86::FsMOVAPDrm: {
956 // Loads from constant pools are trivially rematerializable.
957 if (MI->getOperand(1).isReg() &&
958 MI->getOperand(2).isImm() &&
959 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
960 MI->isInvariantLoad(AA)) {
961 unsigned BaseReg = MI->getOperand(1).getReg();
962 if (BaseReg == 0 || BaseReg == X86::RIP)
964 // Allow re-materialization of PIC load.
965 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
967 const MachineFunction &MF = *MI->getParent()->getParent();
968 const MachineRegisterInfo &MRI = MF.getRegInfo();
969 bool isPICBase = false;
970 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
971 E = MRI.def_end(); I != E; ++I) {
972 MachineInstr *DefMI = I.getOperand().getParent();
973 if (DefMI->getOpcode() != X86::MOVPC32r)
975 assert(!isPICBase && "More than one PIC base?");
985 if (MI->getOperand(2).isImm() &&
986 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
987 !MI->getOperand(4).isReg()) {
988 // lea fi#, lea GV, etc. are all rematerializable.
989 if (!MI->getOperand(1).isReg())
991 unsigned BaseReg = MI->getOperand(1).getReg();
994 // Allow re-materialization of lea PICBase + x.
995 const MachineFunction &MF = *MI->getParent()->getParent();
996 const MachineRegisterInfo &MRI = MF.getRegInfo();
997 return regIsPICBase(BaseReg, MRI);
1003 // All other instructions marked M_REMATERIALIZABLE are always trivially
1004 // rematerializable.
1008 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1009 /// would clobber the EFLAGS condition register. Note the result may be
1010 /// conservative. If it cannot definitely determine the safety after visiting
1011 /// a few instructions in each direction it assumes it's not safe.
1012 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1013 MachineBasicBlock::iterator I) {
1014 MachineBasicBlock::iterator E = MBB.end();
1016 // It's always safe to clobber EFLAGS at the end of a block.
1020 // For compile time consideration, if we are not able to determine the
1021 // safety after visiting 4 instructions in each direction, we will assume
1023 MachineBasicBlock::iterator Iter = I;
1024 for (unsigned i = 0; i < 4; ++i) {
1025 bool SeenDef = false;
1026 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1027 MachineOperand &MO = Iter->getOperand(j);
1030 if (MO.getReg() == X86::EFLAGS) {
1038 // This instruction defines EFLAGS, no need to look any further.
1041 // Skip over DBG_VALUE.
1042 while (Iter != E && Iter->isDebugValue())
1045 // If we make it to the end of the block, it's safe to clobber EFLAGS.
1050 MachineBasicBlock::iterator B = MBB.begin();
1052 for (unsigned i = 0; i < 4; ++i) {
1053 // If we make it to the beginning of the block, it's safe to clobber
1054 // EFLAGS iff EFLAGS is not live-in.
1056 return !MBB.isLiveIn(X86::EFLAGS);
1059 // Skip over DBG_VALUE.
1060 while (Iter != B && Iter->isDebugValue())
1063 bool SawKill = false;
1064 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1065 MachineOperand &MO = Iter->getOperand(j);
1066 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1067 if (MO.isDef()) return MO.isDead();
1068 if (MO.isKill()) SawKill = true;
1073 // This instruction kills EFLAGS and doesn't redefine it, so
1074 // there's no need to look further.
1078 // Conservative answer.
1082 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1083 MachineBasicBlock::iterator I,
1084 unsigned DestReg, unsigned SubIdx,
1085 const MachineInstr *Orig,
1086 const TargetRegisterInfo &TRI) const {
1087 DebugLoc DL = Orig->getDebugLoc();
1089 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1090 // Re-materialize them as movri instructions to avoid side effects.
1092 unsigned Opc = Orig->getOpcode();
1098 case X86::MOV64r0: {
1099 if (!isSafeToClobberEFLAGS(MBB, I)) {
1102 case X86::MOV8r0: Opc = X86::MOV8ri; break;
1103 case X86::MOV16r0: Opc = X86::MOV16ri; break;
1104 case X86::MOV32r0: Opc = X86::MOV32ri; break;
1105 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
1114 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1117 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
1120 MachineInstr *NewMI = prior(I);
1121 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1124 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1125 /// is not marked dead.
1126 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1127 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1128 MachineOperand &MO = MI->getOperand(i);
1129 if (MO.isReg() && MO.isDef() &&
1130 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1137 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1138 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1139 /// to a 32-bit superregister and then truncating back down to a 16-bit
1142 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1143 MachineFunction::iterator &MFI,
1144 MachineBasicBlock::iterator &MBBI,
1145 LiveVariables *LV) const {
1146 MachineInstr *MI = MBBI;
1147 unsigned Dest = MI->getOperand(0).getReg();
1148 unsigned Src = MI->getOperand(1).getReg();
1149 bool isDead = MI->getOperand(0).isDead();
1150 bool isKill = MI->getOperand(1).isKill();
1152 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1153 ? X86::LEA64_32r : X86::LEA32r;
1154 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1155 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1156 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1158 // Build and insert into an implicit UNDEF value. This is OK because
1159 // well be shifting and then extracting the lower 16-bits.
1160 // This has the potential to cause partial register stall. e.g.
1161 // movw (%rbp,%rcx,2), %dx
1162 // leal -65(%rdx), %esi
1163 // But testing has shown this *does* help performance in 64-bit mode (at
1164 // least on modern x86 machines).
1165 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1166 MachineInstr *InsMI =
1167 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1168 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1169 .addReg(Src, getKillRegState(isKill));
1171 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1172 get(Opc), leaOutReg);
1175 llvm_unreachable(0);
1177 case X86::SHL16ri: {
1178 unsigned ShAmt = MI->getOperand(2).getImm();
1179 MIB.addReg(0).addImm(1 << ShAmt)
1180 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1184 case X86::INC64_16r:
1185 addRegOffset(MIB, leaInReg, true, 1);
1188 case X86::DEC64_16r:
1189 addRegOffset(MIB, leaInReg, true, -1);
1193 case X86::ADD16ri_DB:
1194 case X86::ADD16ri8_DB:
1195 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
1198 case X86::ADD16rr_DB: {
1199 unsigned Src2 = MI->getOperand(2).getReg();
1200 bool isKill2 = MI->getOperand(2).isKill();
1201 unsigned leaInReg2 = 0;
1202 MachineInstr *InsMI2 = 0;
1204 // ADD16rr %reg1028<kill>, %reg1028
1205 // just a single insert_subreg.
1206 addRegReg(MIB, leaInReg, true, leaInReg, false);
1208 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1209 // Build and insert into an implicit UNDEF value. This is OK because
1210 // well be shifting and then extracting the lower 16-bits.
1211 BuildMI(*MFI, MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg2);
1213 BuildMI(*MFI, MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
1214 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1215 .addReg(Src2, getKillRegState(isKill2));
1216 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1218 if (LV && isKill2 && InsMI2)
1219 LV->replaceKillInstruction(Src2, MI, InsMI2);
1224 MachineInstr *NewMI = MIB;
1225 MachineInstr *ExtMI =
1226 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1227 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1228 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
1231 // Update live variables
1232 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1233 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1235 LV->replaceKillInstruction(Src, MI, InsMI);
1237 LV->replaceKillInstruction(Dest, MI, ExtMI);
1243 /// convertToThreeAddress - This method must be implemented by targets that
1244 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1245 /// may be able to convert a two-address instruction into a true
1246 /// three-address instruction on demand. This allows the X86 target (for
1247 /// example) to convert ADD and SHL instructions into LEA instructions if they
1248 /// would require register copies due to two-addressness.
1250 /// This method returns a null pointer if the transformation cannot be
1251 /// performed, otherwise it returns the new instruction.
1254 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1255 MachineBasicBlock::iterator &MBBI,
1256 LiveVariables *LV) const {
1257 MachineInstr *MI = MBBI;
1258 MachineFunction &MF = *MI->getParent()->getParent();
1259 // All instructions input are two-addr instructions. Get the known operands.
1260 unsigned Dest = MI->getOperand(0).getReg();
1261 unsigned Src = MI->getOperand(1).getReg();
1262 bool isDead = MI->getOperand(0).isDead();
1263 bool isKill = MI->getOperand(1).isKill();
1265 MachineInstr *NewMI = NULL;
1266 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1267 // we have better subtarget support, enable the 16-bit LEA generation here.
1268 // 16-bit LEA is also slow on Core2.
1269 bool DisableLEA16 = true;
1270 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1272 unsigned MIOpc = MI->getOpcode();
1274 case X86::SHUFPSrri: {
1275 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1276 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1278 unsigned B = MI->getOperand(1).getReg();
1279 unsigned C = MI->getOperand(2).getReg();
1280 if (B != C) return 0;
1281 unsigned A = MI->getOperand(0).getReg();
1282 unsigned M = MI->getOperand(3).getImm();
1283 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1284 .addReg(A, RegState::Define | getDeadRegState(isDead))
1285 .addReg(B, getKillRegState(isKill)).addImm(M);
1288 case X86::SHL64ri: {
1289 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1290 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1291 // the flags produced by a shift yet, so this is safe.
1292 unsigned ShAmt = MI->getOperand(2).getImm();
1293 if (ShAmt == 0 || ShAmt >= 4) return 0;
1295 // LEA can't handle RSP.
1296 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1297 !MF.getRegInfo().constrainRegClass(Src, &X86::GR64_NOSPRegClass))
1300 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1301 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1302 .addReg(0).addImm(1 << ShAmt)
1303 .addReg(Src, getKillRegState(isKill))
1304 .addImm(0).addReg(0);
1307 case X86::SHL32ri: {
1308 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1309 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1310 // the flags produced by a shift yet, so this is safe.
1311 unsigned ShAmt = MI->getOperand(2).getImm();
1312 if (ShAmt == 0 || ShAmt >= 4) return 0;
1314 // LEA can't handle ESP.
1315 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1316 !MF.getRegInfo().constrainRegClass(Src, &X86::GR32_NOSPRegClass))
1319 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1320 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1321 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1322 .addReg(0).addImm(1 << ShAmt)
1323 .addReg(Src, getKillRegState(isKill)).addImm(0).addReg(0);
1326 case X86::SHL16ri: {
1327 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1328 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1329 // the flags produced by a shift yet, so this is safe.
1330 unsigned ShAmt = MI->getOperand(2).getImm();
1331 if (ShAmt == 0 || ShAmt >= 4) return 0;
1334 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1335 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1336 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1337 .addReg(0).addImm(1 << ShAmt)
1338 .addReg(Src, getKillRegState(isKill))
1339 .addImm(0).addReg(0);
1343 // The following opcodes also sets the condition code register(s). Only
1344 // convert them to equivalent lea if the condition code register def's
1346 if (hasLiveCondCodeDef(MI))
1353 case X86::INC64_32r: {
1354 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1355 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1356 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1358 // LEA can't handle RSP.
1359 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1360 !MF.getRegInfo().constrainRegClass(Src,
1361 MIOpc == X86::INC64r ? X86::GR64_NOSPRegisterClass :
1362 X86::GR32_NOSPRegisterClass))
1365 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1366 .addReg(Dest, RegState::Define |
1367 getDeadRegState(isDead)),
1372 case X86::INC64_16r:
1374 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1375 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1376 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1377 .addReg(Dest, RegState::Define |
1378 getDeadRegState(isDead)),
1383 case X86::DEC64_32r: {
1384 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1385 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1386 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1387 // LEA can't handle RSP.
1388 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1389 !MF.getRegInfo().constrainRegClass(Src,
1390 MIOpc == X86::DEC64r ? X86::GR64_NOSPRegisterClass :
1391 X86::GR32_NOSPRegisterClass))
1394 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1395 .addReg(Dest, RegState::Define |
1396 getDeadRegState(isDead)),
1401 case X86::DEC64_16r:
1403 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1404 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1405 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1406 .addReg(Dest, RegState::Define |
1407 getDeadRegState(isDead)),
1411 case X86::ADD64rr_DB:
1413 case X86::ADD32rr_DB: {
1414 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1416 TargetRegisterClass *RC;
1417 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
1419 RC = X86::GR64_NOSPRegisterClass;
1421 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1422 RC = X86::GR32_NOSPRegisterClass;
1426 unsigned Src2 = MI->getOperand(2).getReg();
1427 bool isKill2 = MI->getOperand(2).isKill();
1429 // LEA can't handle RSP.
1430 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
1431 !MF.getRegInfo().constrainRegClass(Src2, RC))
1434 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1435 .addReg(Dest, RegState::Define |
1436 getDeadRegState(isDead)),
1437 Src, isKill, Src2, isKill2);
1439 LV->replaceKillInstruction(Src2, MI, NewMI);
1443 case X86::ADD16rr_DB: {
1445 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1446 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1447 unsigned Src2 = MI->getOperand(2).getReg();
1448 bool isKill2 = MI->getOperand(2).isKill();
1449 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1450 .addReg(Dest, RegState::Define |
1451 getDeadRegState(isDead)),
1452 Src, isKill, Src2, isKill2);
1454 LV->replaceKillInstruction(Src2, MI, NewMI);
1457 case X86::ADD64ri32:
1459 case X86::ADD64ri32_DB:
1460 case X86::ADD64ri8_DB:
1461 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1462 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1463 .addReg(Dest, RegState::Define |
1464 getDeadRegState(isDead)),
1465 Src, isKill, MI->getOperand(2).getImm());
1469 case X86::ADD32ri_DB:
1470 case X86::ADD32ri8_DB: {
1471 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1472 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1473 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1474 .addReg(Dest, RegState::Define |
1475 getDeadRegState(isDead)),
1476 Src, isKill, MI->getOperand(2).getImm());
1481 case X86::ADD16ri_DB:
1482 case X86::ADD16ri8_DB:
1484 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1485 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1486 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1487 .addReg(Dest, RegState::Define |
1488 getDeadRegState(isDead)),
1489 Src, isKill, MI->getOperand(2).getImm());
1495 if (!NewMI) return 0;
1497 if (LV) { // Update live variables
1499 LV->replaceKillInstruction(Src, MI, NewMI);
1501 LV->replaceKillInstruction(Dest, MI, NewMI);
1504 MFI->insert(MBBI, NewMI); // Insert the new inst
1508 /// commuteInstruction - We have a few instructions that must be hacked on to
1512 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
1513 switch (MI->getOpcode()) {
1514 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
1515 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
1516 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
1517 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
1518 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
1519 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
1522 switch (MI->getOpcode()) {
1523 default: llvm_unreachable("Unreachable!");
1524 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
1525 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
1526 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
1527 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
1528 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
1529 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
1531 unsigned Amt = MI->getOperand(3).getImm();
1533 MachineFunction &MF = *MI->getParent()->getParent();
1534 MI = MF.CloneMachineInstr(MI);
1537 MI->setDesc(get(Opc));
1538 MI->getOperand(3).setImm(Size-Amt);
1539 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1541 case X86::CMOVB16rr:
1542 case X86::CMOVB32rr:
1543 case X86::CMOVB64rr:
1544 case X86::CMOVAE16rr:
1545 case X86::CMOVAE32rr:
1546 case X86::CMOVAE64rr:
1547 case X86::CMOVE16rr:
1548 case X86::CMOVE32rr:
1549 case X86::CMOVE64rr:
1550 case X86::CMOVNE16rr:
1551 case X86::CMOVNE32rr:
1552 case X86::CMOVNE64rr:
1553 case X86::CMOVBE16rr:
1554 case X86::CMOVBE32rr:
1555 case X86::CMOVBE64rr:
1556 case X86::CMOVA16rr:
1557 case X86::CMOVA32rr:
1558 case X86::CMOVA64rr:
1559 case X86::CMOVL16rr:
1560 case X86::CMOVL32rr:
1561 case X86::CMOVL64rr:
1562 case X86::CMOVGE16rr:
1563 case X86::CMOVGE32rr:
1564 case X86::CMOVGE64rr:
1565 case X86::CMOVLE16rr:
1566 case X86::CMOVLE32rr:
1567 case X86::CMOVLE64rr:
1568 case X86::CMOVG16rr:
1569 case X86::CMOVG32rr:
1570 case X86::CMOVG64rr:
1571 case X86::CMOVS16rr:
1572 case X86::CMOVS32rr:
1573 case X86::CMOVS64rr:
1574 case X86::CMOVNS16rr:
1575 case X86::CMOVNS32rr:
1576 case X86::CMOVNS64rr:
1577 case X86::CMOVP16rr:
1578 case X86::CMOVP32rr:
1579 case X86::CMOVP64rr:
1580 case X86::CMOVNP16rr:
1581 case X86::CMOVNP32rr:
1582 case X86::CMOVNP64rr:
1583 case X86::CMOVO16rr:
1584 case X86::CMOVO32rr:
1585 case X86::CMOVO64rr:
1586 case X86::CMOVNO16rr:
1587 case X86::CMOVNO32rr:
1588 case X86::CMOVNO64rr: {
1590 switch (MI->getOpcode()) {
1592 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
1593 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
1594 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
1595 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
1596 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
1597 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
1598 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
1599 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
1600 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
1601 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
1602 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
1603 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
1604 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
1605 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
1606 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
1607 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
1608 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
1609 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
1610 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
1611 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
1612 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
1613 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
1614 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
1615 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
1616 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
1617 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
1618 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
1619 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
1620 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
1621 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
1622 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
1623 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
1624 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
1625 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
1626 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
1627 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
1628 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
1629 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
1630 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
1631 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
1632 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
1633 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
1634 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
1635 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
1636 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
1637 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
1638 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
1639 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
1642 MachineFunction &MF = *MI->getParent()->getParent();
1643 MI = MF.CloneMachineInstr(MI);
1646 MI->setDesc(get(Opc));
1647 // Fallthrough intended.
1650 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
1654 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
1656 default: return X86::COND_INVALID;
1657 case X86::JE_4: return X86::COND_E;
1658 case X86::JNE_4: return X86::COND_NE;
1659 case X86::JL_4: return X86::COND_L;
1660 case X86::JLE_4: return X86::COND_LE;
1661 case X86::JG_4: return X86::COND_G;
1662 case X86::JGE_4: return X86::COND_GE;
1663 case X86::JB_4: return X86::COND_B;
1664 case X86::JBE_4: return X86::COND_BE;
1665 case X86::JA_4: return X86::COND_A;
1666 case X86::JAE_4: return X86::COND_AE;
1667 case X86::JS_4: return X86::COND_S;
1668 case X86::JNS_4: return X86::COND_NS;
1669 case X86::JP_4: return X86::COND_P;
1670 case X86::JNP_4: return X86::COND_NP;
1671 case X86::JO_4: return X86::COND_O;
1672 case X86::JNO_4: return X86::COND_NO;
1676 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
1678 default: llvm_unreachable("Illegal condition code!");
1679 case X86::COND_E: return X86::JE_4;
1680 case X86::COND_NE: return X86::JNE_4;
1681 case X86::COND_L: return X86::JL_4;
1682 case X86::COND_LE: return X86::JLE_4;
1683 case X86::COND_G: return X86::JG_4;
1684 case X86::COND_GE: return X86::JGE_4;
1685 case X86::COND_B: return X86::JB_4;
1686 case X86::COND_BE: return X86::JBE_4;
1687 case X86::COND_A: return X86::JA_4;
1688 case X86::COND_AE: return X86::JAE_4;
1689 case X86::COND_S: return X86::JS_4;
1690 case X86::COND_NS: return X86::JNS_4;
1691 case X86::COND_P: return X86::JP_4;
1692 case X86::COND_NP: return X86::JNP_4;
1693 case X86::COND_O: return X86::JO_4;
1694 case X86::COND_NO: return X86::JNO_4;
1698 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
1699 /// e.g. turning COND_E to COND_NE.
1700 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
1702 default: llvm_unreachable("Illegal condition code!");
1703 case X86::COND_E: return X86::COND_NE;
1704 case X86::COND_NE: return X86::COND_E;
1705 case X86::COND_L: return X86::COND_GE;
1706 case X86::COND_LE: return X86::COND_G;
1707 case X86::COND_G: return X86::COND_LE;
1708 case X86::COND_GE: return X86::COND_L;
1709 case X86::COND_B: return X86::COND_AE;
1710 case X86::COND_BE: return X86::COND_A;
1711 case X86::COND_A: return X86::COND_BE;
1712 case X86::COND_AE: return X86::COND_B;
1713 case X86::COND_S: return X86::COND_NS;
1714 case X86::COND_NS: return X86::COND_S;
1715 case X86::COND_P: return X86::COND_NP;
1716 case X86::COND_NP: return X86::COND_P;
1717 case X86::COND_O: return X86::COND_NO;
1718 case X86::COND_NO: return X86::COND_O;
1722 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
1723 const MCInstrDesc &MCID = MI->getDesc();
1724 if (!MCID.isTerminator()) return false;
1726 // Conditional branch is a special case.
1727 if (MCID.isBranch() && !MCID.isBarrier())
1729 if (!MCID.isPredicable())
1731 return !isPredicated(MI);
1734 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1735 MachineBasicBlock *&TBB,
1736 MachineBasicBlock *&FBB,
1737 SmallVectorImpl<MachineOperand> &Cond,
1738 bool AllowModify) const {
1739 // Start from the bottom of the block and work up, examining the
1740 // terminator instructions.
1741 MachineBasicBlock::iterator I = MBB.end();
1742 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
1743 while (I != MBB.begin()) {
1745 if (I->isDebugValue())
1748 // Working from the bottom, when we see a non-terminator instruction, we're
1750 if (!isUnpredicatedTerminator(I))
1753 // A terminator that isn't a branch can't easily be handled by this
1755 if (!I->getDesc().isBranch())
1758 // Handle unconditional branches.
1759 if (I->getOpcode() == X86::JMP_4) {
1763 TBB = I->getOperand(0).getMBB();
1767 // If the block has any instructions after a JMP, delete them.
1768 while (llvm::next(I) != MBB.end())
1769 llvm::next(I)->eraseFromParent();
1774 // Delete the JMP if it's equivalent to a fall-through.
1775 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
1777 I->eraseFromParent();
1779 UnCondBrIter = MBB.end();
1783 // TBB is used to indicate the unconditional destination.
1784 TBB = I->getOperand(0).getMBB();
1788 // Handle conditional branches.
1789 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
1790 if (BranchCode == X86::COND_INVALID)
1791 return true; // Can't handle indirect branch.
1793 // Working from the bottom, handle the first conditional branch.
1795 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
1796 if (AllowModify && UnCondBrIter != MBB.end() &&
1797 MBB.isLayoutSuccessor(TargetBB)) {
1798 // If we can modify the code and it ends in something like:
1806 // Then we can change this to:
1813 // Which is a bit more efficient.
1814 // We conditionally jump to the fall-through block.
1815 BranchCode = GetOppositeBranchCondition(BranchCode);
1816 unsigned JNCC = GetCondBranchFromCond(BranchCode);
1817 MachineBasicBlock::iterator OldInst = I;
1819 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
1820 .addMBB(UnCondBrIter->getOperand(0).getMBB());
1821 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
1824 OldInst->eraseFromParent();
1825 UnCondBrIter->eraseFromParent();
1827 // Restart the analysis.
1828 UnCondBrIter = MBB.end();
1834 TBB = I->getOperand(0).getMBB();
1835 Cond.push_back(MachineOperand::CreateImm(BranchCode));
1839 // Handle subsequent conditional branches. Only handle the case where all
1840 // conditional branches branch to the same destination and their condition
1841 // opcodes fit one of the special multi-branch idioms.
1842 assert(Cond.size() == 1);
1845 // Only handle the case where all conditional branches branch to the same
1847 if (TBB != I->getOperand(0).getMBB())
1850 // If the conditions are the same, we can leave them alone.
1851 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
1852 if (OldBranchCode == BranchCode)
1855 // If they differ, see if they fit one of the known patterns. Theoretically,
1856 // we could handle more patterns here, but we shouldn't expect to see them
1857 // if instruction selection has done a reasonable job.
1858 if ((OldBranchCode == X86::COND_NP &&
1859 BranchCode == X86::COND_E) ||
1860 (OldBranchCode == X86::COND_E &&
1861 BranchCode == X86::COND_NP))
1862 BranchCode = X86::COND_NP_OR_E;
1863 else if ((OldBranchCode == X86::COND_P &&
1864 BranchCode == X86::COND_NE) ||
1865 (OldBranchCode == X86::COND_NE &&
1866 BranchCode == X86::COND_P))
1867 BranchCode = X86::COND_NE_OR_P;
1871 // Update the MachineOperand.
1872 Cond[0].setImm(BranchCode);
1878 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
1879 MachineBasicBlock::iterator I = MBB.end();
1882 while (I != MBB.begin()) {
1884 if (I->isDebugValue())
1886 if (I->getOpcode() != X86::JMP_4 &&
1887 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
1889 // Remove the branch.
1890 I->eraseFromParent();
1899 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
1900 MachineBasicBlock *FBB,
1901 const SmallVectorImpl<MachineOperand> &Cond,
1902 DebugLoc DL) const {
1903 // Shouldn't be a fall through.
1904 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
1905 assert((Cond.size() == 1 || Cond.size() == 0) &&
1906 "X86 branch conditions have one component!");
1909 // Unconditional branch?
1910 assert(!FBB && "Unconditional branch with multiple successors!");
1911 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
1915 // Conditional branch.
1917 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
1919 case X86::COND_NP_OR_E:
1920 // Synthesize NP_OR_E with two branches.
1921 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
1923 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
1926 case X86::COND_NE_OR_P:
1927 // Synthesize NE_OR_P with two branches.
1928 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
1930 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
1934 unsigned Opc = GetCondBranchFromCond(CC);
1935 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
1940 // Two-way Conditional branch. Insert the second branch.
1941 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
1947 /// isHReg - Test if the given register is a physical h register.
1948 static bool isHReg(unsigned Reg) {
1949 return X86::GR8_ABCD_HRegClass.contains(Reg);
1952 // Try and copy between VR128/VR64 and GR64 registers.
1953 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg) {
1954 // SrcReg(VR128) -> DestReg(GR64)
1955 // SrcReg(VR64) -> DestReg(GR64)
1956 // SrcReg(GR64) -> DestReg(VR128)
1957 // SrcReg(GR64) -> DestReg(VR64)
1959 if (X86::GR64RegClass.contains(DestReg)) {
1960 if (X86::VR128RegClass.contains(SrcReg)) {
1961 // Copy from a VR128 register to a GR64 register.
1962 return X86::MOVPQIto64rr;
1963 } else if (X86::VR64RegClass.contains(SrcReg)) {
1964 // Copy from a VR64 register to a GR64 register.
1965 return X86::MOVSDto64rr;
1967 } else if (X86::GR64RegClass.contains(SrcReg)) {
1968 // Copy from a GR64 register to a VR128 register.
1969 if (X86::VR128RegClass.contains(DestReg))
1970 return X86::MOV64toPQIrr;
1971 // Copy from a GR64 register to a VR64 register.
1972 else if (X86::VR64RegClass.contains(DestReg))
1973 return X86::MOV64toSDrr;
1979 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
1980 MachineBasicBlock::iterator MI, DebugLoc DL,
1981 unsigned DestReg, unsigned SrcReg,
1982 bool KillSrc) const {
1983 // First deal with the normal symmetric copies.
1985 if (X86::GR64RegClass.contains(DestReg, SrcReg))
1987 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
1989 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
1991 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
1992 // Copying to or from a physical H register on x86-64 requires a NOREX
1993 // move. Otherwise use a normal move.
1994 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
1995 TM.getSubtarget<X86Subtarget>().is64Bit())
1996 Opc = X86::MOV8rr_NOREX;
1999 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
2000 Opc = X86::MOVAPSrr;
2001 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2002 Opc = X86::VMOVAPSYrr;
2003 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2004 Opc = X86::MMX_MOVQ64rr;
2006 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg);
2009 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2010 .addReg(SrcReg, getKillRegState(KillSrc));
2014 // Moving EFLAGS to / from another register requires a push and a pop.
2015 if (SrcReg == X86::EFLAGS) {
2016 if (X86::GR64RegClass.contains(DestReg)) {
2017 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2018 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2020 } else if (X86::GR32RegClass.contains(DestReg)) {
2021 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2022 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2026 if (DestReg == X86::EFLAGS) {
2027 if (X86::GR64RegClass.contains(SrcReg)) {
2028 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2029 .addReg(SrcReg, getKillRegState(KillSrc));
2030 BuildMI(MBB, MI, DL, get(X86::POPF64));
2032 } else if (X86::GR32RegClass.contains(SrcReg)) {
2033 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2034 .addReg(SrcReg, getKillRegState(KillSrc));
2035 BuildMI(MBB, MI, DL, get(X86::POPF32));
2040 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2041 << " to " << RI.getName(DestReg) << '\n');
2042 llvm_unreachable("Cannot emit physreg copy instruction");
2045 static unsigned getLoadStoreRegOpcode(unsigned Reg,
2046 const TargetRegisterClass *RC,
2047 bool isStackAligned,
2048 const TargetMachine &TM,
2050 switch (RC->getSize()) {
2052 llvm_unreachable("Unknown spill size");
2054 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
2055 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2056 // Copying to or from a physical H register on x86-64 requires a NOREX
2057 // move. Otherwise use a normal move.
2058 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2059 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2060 return load ? X86::MOV8rm : X86::MOV8mr;
2062 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2063 return load ? X86::MOV16rm : X86::MOV16mr;
2065 if (X86::GR32RegClass.hasSubClassEq(RC))
2066 return load ? X86::MOV32rm : X86::MOV32mr;
2067 if (X86::FR32RegClass.hasSubClassEq(RC))
2068 return load ? X86::MOVSSrm : X86::MOVSSmr;
2069 if (X86::RFP32RegClass.hasSubClassEq(RC))
2070 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2071 llvm_unreachable("Unknown 4-byte regclass");
2073 if (X86::GR64RegClass.hasSubClassEq(RC))
2074 return load ? X86::MOV64rm : X86::MOV64mr;
2075 if (X86::FR64RegClass.hasSubClassEq(RC))
2076 return load ? X86::MOVSDrm : X86::MOVSDmr;
2077 if (X86::VR64RegClass.hasSubClassEq(RC))
2078 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2079 if (X86::RFP64RegClass.hasSubClassEq(RC))
2080 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2081 llvm_unreachable("Unknown 8-byte regclass");
2083 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
2084 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
2086 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
2087 // If stack is realigned we can use aligned stores.
2089 return load ? X86::MOVAPSrm : X86::MOVAPSmr;
2091 return load ? X86::MOVUPSrm : X86::MOVUPSmr;
2093 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2094 // If stack is realigned we can use aligned stores.
2096 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2098 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
2102 static unsigned getStoreRegOpcode(unsigned SrcReg,
2103 const TargetRegisterClass *RC,
2104 bool isStackAligned,
2105 TargetMachine &TM) {
2106 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2110 static unsigned getLoadRegOpcode(unsigned DestReg,
2111 const TargetRegisterClass *RC,
2112 bool isStackAligned,
2113 const TargetMachine &TM) {
2114 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
2117 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2118 MachineBasicBlock::iterator MI,
2119 unsigned SrcReg, bool isKill, int FrameIdx,
2120 const TargetRegisterClass *RC,
2121 const TargetRegisterInfo *TRI) const {
2122 const MachineFunction &MF = *MBB.getParent();
2123 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2124 "Stack slot too small for store");
2125 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= 16) ||
2126 RI.canRealignStack(MF);
2127 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
2128 DebugLoc DL = MBB.findDebugLoc(MI);
2129 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
2130 .addReg(SrcReg, getKillRegState(isKill));
2133 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2135 SmallVectorImpl<MachineOperand> &Addr,
2136 const TargetRegisterClass *RC,
2137 MachineInstr::mmo_iterator MMOBegin,
2138 MachineInstr::mmo_iterator MMOEnd,
2139 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2140 bool isAligned = MMOBegin != MMOEnd && (*MMOBegin)->getAlignment() >= 16;
2141 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
2143 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
2144 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2145 MIB.addOperand(Addr[i]);
2146 MIB.addReg(SrcReg, getKillRegState(isKill));
2147 (*MIB).setMemRefs(MMOBegin, MMOEnd);
2148 NewMIs.push_back(MIB);
2152 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
2153 MachineBasicBlock::iterator MI,
2154 unsigned DestReg, int FrameIdx,
2155 const TargetRegisterClass *RC,
2156 const TargetRegisterInfo *TRI) const {
2157 const MachineFunction &MF = *MBB.getParent();
2158 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= 16) ||
2159 RI.canRealignStack(MF);
2160 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2161 DebugLoc DL = MBB.findDebugLoc(MI);
2162 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
2165 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
2166 SmallVectorImpl<MachineOperand> &Addr,
2167 const TargetRegisterClass *RC,
2168 MachineInstr::mmo_iterator MMOBegin,
2169 MachineInstr::mmo_iterator MMOEnd,
2170 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2171 bool isAligned = MMOBegin != MMOEnd && (*MMOBegin)->getAlignment() >= 16;
2172 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2174 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
2175 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2176 MIB.addOperand(Addr[i]);
2177 (*MIB).setMemRefs(MMOBegin, MMOEnd);
2178 NewMIs.push_back(MIB);
2182 X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
2183 int FrameIx, uint64_t Offset,
2184 const MDNode *MDPtr,
2185 DebugLoc DL) const {
2187 AM.BaseType = X86AddressMode::FrameIndexBase;
2188 AM.Base.FrameIndex = FrameIx;
2189 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
2190 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
2194 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
2195 const SmallVectorImpl<MachineOperand> &MOs,
2197 const TargetInstrInfo &TII) {
2198 // Create the base instruction with the memory operand as the first part.
2199 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2200 MI->getDebugLoc(), true);
2201 MachineInstrBuilder MIB(NewMI);
2202 unsigned NumAddrOps = MOs.size();
2203 for (unsigned i = 0; i != NumAddrOps; ++i)
2204 MIB.addOperand(MOs[i]);
2205 if (NumAddrOps < 4) // FrameIndex only
2208 // Loop over the rest of the ri operands, converting them over.
2209 unsigned NumOps = MI->getDesc().getNumOperands()-2;
2210 for (unsigned i = 0; i != NumOps; ++i) {
2211 MachineOperand &MO = MI->getOperand(i+2);
2214 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
2215 MachineOperand &MO = MI->getOperand(i);
2221 static MachineInstr *FuseInst(MachineFunction &MF,
2222 unsigned Opcode, unsigned OpNo,
2223 const SmallVectorImpl<MachineOperand> &MOs,
2224 MachineInstr *MI, const TargetInstrInfo &TII) {
2225 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2226 MI->getDebugLoc(), true);
2227 MachineInstrBuilder MIB(NewMI);
2229 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2230 MachineOperand &MO = MI->getOperand(i);
2232 assert(MO.isReg() && "Expected to fold into reg operand!");
2233 unsigned NumAddrOps = MOs.size();
2234 for (unsigned i = 0; i != NumAddrOps; ++i)
2235 MIB.addOperand(MOs[i]);
2236 if (NumAddrOps < 4) // FrameIndex only
2245 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
2246 const SmallVectorImpl<MachineOperand> &MOs,
2248 MachineFunction &MF = *MI->getParent()->getParent();
2249 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
2251 unsigned NumAddrOps = MOs.size();
2252 for (unsigned i = 0; i != NumAddrOps; ++i)
2253 MIB.addOperand(MOs[i]);
2254 if (NumAddrOps < 4) // FrameIndex only
2256 return MIB.addImm(0);
2260 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2261 MachineInstr *MI, unsigned i,
2262 const SmallVectorImpl<MachineOperand> &MOs,
2263 unsigned Size, unsigned Align) const {
2264 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
2265 bool isTwoAddrFold = false;
2266 unsigned NumOps = MI->getDesc().getNumOperands();
2267 bool isTwoAddr = NumOps > 1 &&
2268 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
2270 // FIXME: AsmPrinter doesn't know how to handle
2271 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
2272 if (MI->getOpcode() == X86::ADD32ri &&
2273 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
2276 MachineInstr *NewMI = NULL;
2277 // Folding a memory location into the two-address part of a two-address
2278 // instruction is different than folding it other places. It requires
2279 // replacing the *two* registers with the memory location.
2280 if (isTwoAddr && NumOps >= 2 && i < 2 &&
2281 MI->getOperand(0).isReg() &&
2282 MI->getOperand(1).isReg() &&
2283 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
2284 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2285 isTwoAddrFold = true;
2286 } else if (i == 0) { // If operand 0
2287 if (MI->getOpcode() == X86::MOV64r0)
2288 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2289 else if (MI->getOpcode() == X86::MOV32r0)
2290 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
2291 else if (MI->getOpcode() == X86::MOV16r0)
2292 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
2293 else if (MI->getOpcode() == X86::MOV8r0)
2294 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
2298 OpcodeTablePtr = &RegOp2MemOpTable0;
2299 } else if (i == 1) {
2300 OpcodeTablePtr = &RegOp2MemOpTable1;
2301 } else if (i == 2) {
2302 OpcodeTablePtr = &RegOp2MemOpTable2;
2305 // If table selected...
2306 if (OpcodeTablePtr) {
2307 // Find the Opcode to fuse
2308 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
2309 OpcodeTablePtr->find(MI->getOpcode());
2310 if (I != OpcodeTablePtr->end()) {
2311 unsigned Opcode = I->second.first;
2312 unsigned MinAlign = I->second.second;
2313 if (Align < MinAlign)
2315 bool NarrowToMOV32rm = false;
2317 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI)->getSize();
2318 if (Size < RCSize) {
2319 // Check if it's safe to fold the load. If the size of the object is
2320 // narrower than the load width, then it's not.
2321 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
2323 // If this is a 64-bit load, but the spill slot is 32, then we can do
2324 // a 32-bit load which is implicitly zero-extended. This likely is due
2325 // to liveintervalanalysis remat'ing a load from stack slot.
2326 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
2328 Opcode = X86::MOV32rm;
2329 NarrowToMOV32rm = true;
2334 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
2336 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
2338 if (NarrowToMOV32rm) {
2339 // If this is the special case where we use a MOV32rm to load a 32-bit
2340 // value and zero-extend the top bits. Change the destination register
2342 unsigned DstReg = NewMI->getOperand(0).getReg();
2343 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
2344 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
2347 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
2354 if (PrintFailedFusing && !MI->isCopy())
2355 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
2360 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2362 const SmallVectorImpl<unsigned> &Ops,
2363 int FrameIndex) const {
2364 // Check switch flag
2365 if (NoFusing) return NULL;
2367 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
2368 switch (MI->getOpcode()) {
2369 case X86::CVTSD2SSrr:
2370 case X86::Int_CVTSD2SSrr:
2371 case X86::CVTSS2SDrr:
2372 case X86::Int_CVTSS2SDrr:
2374 case X86::RCPSSr_Int:
2378 case X86::RSQRTSSr_Int:
2380 case X86::SQRTSSr_Int:
2384 const MachineFrameInfo *MFI = MF.getFrameInfo();
2385 unsigned Size = MFI->getObjectSize(FrameIndex);
2386 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
2387 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2388 unsigned NewOpc = 0;
2389 unsigned RCSize = 0;
2390 switch (MI->getOpcode()) {
2391 default: return NULL;
2392 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
2393 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
2394 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
2395 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
2397 // Check if it's safe to fold the load. If the size of the object is
2398 // narrower than the load width, then it's not.
2401 // Change to CMPXXri r, 0 first.
2402 MI->setDesc(get(NewOpc));
2403 MI->getOperand(1).ChangeToImmediate(0);
2404 } else if (Ops.size() != 1)
2407 SmallVector<MachineOperand,4> MOs;
2408 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
2409 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
2412 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2414 const SmallVectorImpl<unsigned> &Ops,
2415 MachineInstr *LoadMI) const {
2416 // Check switch flag
2417 if (NoFusing) return NULL;
2419 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
2420 switch (MI->getOpcode()) {
2421 case X86::CVTSD2SSrr:
2422 case X86::Int_CVTSD2SSrr:
2423 case X86::CVTSS2SDrr:
2424 case X86::Int_CVTSS2SDrr:
2426 case X86::RCPSSr_Int:
2430 case X86::RSQRTSSr_Int:
2432 case X86::SQRTSSr_Int:
2436 // Determine the alignment of the load.
2437 unsigned Alignment = 0;
2438 if (LoadMI->hasOneMemOperand())
2439 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
2441 switch (LoadMI->getOpcode()) {
2442 case X86::AVX_SET0PSY:
2443 case X86::AVX_SET0PDY:
2449 case X86::V_SETALLONES:
2450 case X86::AVX_SET0PS:
2451 case X86::AVX_SET0PD:
2452 case X86::AVX_SET0PI:
2456 case X86::VFsFLD0SD:
2460 case X86::VFsFLD0SS:
2466 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2467 unsigned NewOpc = 0;
2468 switch (MI->getOpcode()) {
2469 default: return NULL;
2470 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
2471 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
2472 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
2473 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
2475 // Change to CMPXXri r, 0 first.
2476 MI->setDesc(get(NewOpc));
2477 MI->getOperand(1).ChangeToImmediate(0);
2478 } else if (Ops.size() != 1)
2481 // Make sure the subregisters match.
2482 // Otherwise we risk changing the size of the load.
2483 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
2486 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
2487 switch (LoadMI->getOpcode()) {
2491 case X86::V_SETALLONES:
2492 case X86::AVX_SET0PS:
2493 case X86::AVX_SET0PD:
2494 case X86::AVX_SET0PI:
2495 case X86::AVX_SET0PSY:
2496 case X86::AVX_SET0PDY:
2499 case X86::VFsFLD0SD:
2500 case X86::VFsFLD0SS: {
2501 // Folding a V_SET0P? or V_SETALLONES as a load, to ease register pressure.
2502 // Create a constant-pool entry and operands to load from it.
2504 // Medium and large mode can't fold loads this way.
2505 if (TM.getCodeModel() != CodeModel::Small &&
2506 TM.getCodeModel() != CodeModel::Kernel)
2509 // x86-32 PIC requires a PIC base register for constant pools.
2510 unsigned PICBase = 0;
2511 if (TM.getRelocationModel() == Reloc::PIC_) {
2512 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2515 // FIXME: PICBase = getGlobalBaseReg(&MF);
2516 // This doesn't work for several reasons.
2517 // 1. GlobalBaseReg may have been spilled.
2518 // 2. It may not be live at MI.
2522 // Create a constant-pool entry.
2523 MachineConstantPool &MCP = *MF.getConstantPool();
2525 unsigned Opc = LoadMI->getOpcode();
2526 if (Opc == X86::FsFLD0SS || Opc == X86::VFsFLD0SS)
2527 Ty = Type::getFloatTy(MF.getFunction()->getContext());
2528 else if (Opc == X86::FsFLD0SD || Opc == X86::VFsFLD0SD)
2529 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
2530 else if (Opc == X86::AVX_SET0PSY || Opc == X86::AVX_SET0PDY)
2531 Ty = VectorType::get(Type::getFloatTy(MF.getFunction()->getContext()), 8);
2533 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
2534 const Constant *C = LoadMI->getOpcode() == X86::V_SETALLONES ?
2535 Constant::getAllOnesValue(Ty) :
2536 Constant::getNullValue(Ty);
2537 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
2539 // Create operands to load from the constant pool entry.
2540 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
2541 MOs.push_back(MachineOperand::CreateImm(1));
2542 MOs.push_back(MachineOperand::CreateReg(0, false));
2543 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
2544 MOs.push_back(MachineOperand::CreateReg(0, false));
2548 // Folding a normal load. Just copy the load's address operands.
2549 unsigned NumOps = LoadMI->getDesc().getNumOperands();
2550 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
2551 MOs.push_back(LoadMI->getOperand(i));
2555 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
2559 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
2560 const SmallVectorImpl<unsigned> &Ops) const {
2561 // Check switch flag
2562 if (NoFusing) return 0;
2564 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
2565 switch (MI->getOpcode()) {
2566 default: return false;
2573 // FIXME: AsmPrinter doesn't know how to handle
2574 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
2575 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
2581 if (Ops.size() != 1)
2584 unsigned OpNum = Ops[0];
2585 unsigned Opc = MI->getOpcode();
2586 unsigned NumOps = MI->getDesc().getNumOperands();
2587 bool isTwoAddr = NumOps > 1 &&
2588 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
2590 // Folding a memory location into the two-address part of a two-address
2591 // instruction is different than folding it other places. It requires
2592 // replacing the *two* registers with the memory location.
2593 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
2594 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
2595 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2596 } else if (OpNum == 0) { // If operand 0
2601 case X86::MOV64r0: return true;
2604 OpcodeTablePtr = &RegOp2MemOpTable0;
2605 } else if (OpNum == 1) {
2606 OpcodeTablePtr = &RegOp2MemOpTable1;
2607 } else if (OpNum == 2) {
2608 OpcodeTablePtr = &RegOp2MemOpTable2;
2611 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
2613 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
2616 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
2617 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
2618 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2619 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
2620 MemOp2RegOpTable.find(MI->getOpcode());
2621 if (I == MemOp2RegOpTable.end())
2623 unsigned Opc = I->second.first;
2624 unsigned Index = I->second.second & 0xf;
2625 bool FoldedLoad = I->second.second & (1 << 4);
2626 bool FoldedStore = I->second.second & (1 << 5);
2627 if (UnfoldLoad && !FoldedLoad)
2629 UnfoldLoad &= FoldedLoad;
2630 if (UnfoldStore && !FoldedStore)
2632 UnfoldStore &= FoldedStore;
2634 const MCInstrDesc &MCID = get(Opc);
2635 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI);
2636 if (!MI->hasOneMemOperand() &&
2637 RC == &X86::VR128RegClass &&
2638 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
2639 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
2640 // conservatively assume the address is unaligned. That's bad for
2643 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
2644 SmallVector<MachineOperand,2> BeforeOps;
2645 SmallVector<MachineOperand,2> AfterOps;
2646 SmallVector<MachineOperand,4> ImpOps;
2647 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2648 MachineOperand &Op = MI->getOperand(i);
2649 if (i >= Index && i < Index + X86::AddrNumOperands)
2650 AddrOps.push_back(Op);
2651 else if (Op.isReg() && Op.isImplicit())
2652 ImpOps.push_back(Op);
2654 BeforeOps.push_back(Op);
2656 AfterOps.push_back(Op);
2659 // Emit the load instruction.
2661 std::pair<MachineInstr::mmo_iterator,
2662 MachineInstr::mmo_iterator> MMOs =
2663 MF.extractLoadMemRefs(MI->memoperands_begin(),
2664 MI->memoperands_end());
2665 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
2667 // Address operands cannot be marked isKill.
2668 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
2669 MachineOperand &MO = NewMIs[0]->getOperand(i);
2671 MO.setIsKill(false);
2676 // Emit the data processing instruction.
2677 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
2678 MachineInstrBuilder MIB(DataMI);
2681 MIB.addReg(Reg, RegState::Define);
2682 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
2683 MIB.addOperand(BeforeOps[i]);
2686 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
2687 MIB.addOperand(AfterOps[i]);
2688 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
2689 MachineOperand &MO = ImpOps[i];
2690 MIB.addReg(MO.getReg(),
2691 getDefRegState(MO.isDef()) |
2692 RegState::Implicit |
2693 getKillRegState(MO.isKill()) |
2694 getDeadRegState(MO.isDead()) |
2695 getUndefRegState(MO.isUndef()));
2697 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
2698 unsigned NewOpc = 0;
2699 switch (DataMI->getOpcode()) {
2701 case X86::CMP64ri32:
2708 MachineOperand &MO0 = DataMI->getOperand(0);
2709 MachineOperand &MO1 = DataMI->getOperand(1);
2710 if (MO1.getImm() == 0) {
2711 switch (DataMI->getOpcode()) {
2714 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
2716 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
2718 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
2719 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
2721 DataMI->setDesc(get(NewOpc));
2722 MO1.ChangeToRegister(MO0.getReg(), false);
2726 NewMIs.push_back(DataMI);
2728 // Emit the store instruction.
2730 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI);
2731 std::pair<MachineInstr::mmo_iterator,
2732 MachineInstr::mmo_iterator> MMOs =
2733 MF.extractStoreMemRefs(MI->memoperands_begin(),
2734 MI->memoperands_end());
2735 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
2742 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
2743 SmallVectorImpl<SDNode*> &NewNodes) const {
2744 if (!N->isMachineOpcode())
2747 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
2748 MemOp2RegOpTable.find(N->getMachineOpcode());
2749 if (I == MemOp2RegOpTable.end())
2751 unsigned Opc = I->second.first;
2752 unsigned Index = I->second.second & 0xf;
2753 bool FoldedLoad = I->second.second & (1 << 4);
2754 bool FoldedStore = I->second.second & (1 << 5);
2755 const MCInstrDesc &MCID = get(Opc);
2756 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI);
2757 unsigned NumDefs = MCID.NumDefs;
2758 std::vector<SDValue> AddrOps;
2759 std::vector<SDValue> BeforeOps;
2760 std::vector<SDValue> AfterOps;
2761 DebugLoc dl = N->getDebugLoc();
2762 unsigned NumOps = N->getNumOperands();
2763 for (unsigned i = 0; i != NumOps-1; ++i) {
2764 SDValue Op = N->getOperand(i);
2765 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
2766 AddrOps.push_back(Op);
2767 else if (i < Index-NumDefs)
2768 BeforeOps.push_back(Op);
2769 else if (i > Index-NumDefs)
2770 AfterOps.push_back(Op);
2772 SDValue Chain = N->getOperand(NumOps-1);
2773 AddrOps.push_back(Chain);
2775 // Emit the load instruction.
2777 MachineFunction &MF = DAG.getMachineFunction();
2779 EVT VT = *RC->vt_begin();
2780 std::pair<MachineInstr::mmo_iterator,
2781 MachineInstr::mmo_iterator> MMOs =
2782 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
2783 cast<MachineSDNode>(N)->memoperands_end());
2784 if (!(*MMOs.first) &&
2785 RC == &X86::VR128RegClass &&
2786 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
2787 // Do not introduce a slow unaligned load.
2789 bool isAligned = (*MMOs.first) && (*MMOs.first)->getAlignment() >= 16;
2790 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
2791 VT, MVT::Other, &AddrOps[0], AddrOps.size());
2792 NewNodes.push_back(Load);
2794 // Preserve memory reference information.
2795 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
2798 // Emit the data processing instruction.
2799 std::vector<EVT> VTs;
2800 const TargetRegisterClass *DstRC = 0;
2801 if (MCID.getNumDefs() > 0) {
2802 DstRC = getRegClass(MCID, 0, &RI);
2803 VTs.push_back(*DstRC->vt_begin());
2805 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
2806 EVT VT = N->getValueType(i);
2807 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
2811 BeforeOps.push_back(SDValue(Load, 0));
2812 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
2813 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
2815 NewNodes.push_back(NewNode);
2817 // Emit the store instruction.
2820 AddrOps.push_back(SDValue(NewNode, 0));
2821 AddrOps.push_back(Chain);
2822 std::pair<MachineInstr::mmo_iterator,
2823 MachineInstr::mmo_iterator> MMOs =
2824 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
2825 cast<MachineSDNode>(N)->memoperands_end());
2826 if (!(*MMOs.first) &&
2827 RC == &X86::VR128RegClass &&
2828 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
2829 // Do not introduce a slow unaligned store.
2831 bool isAligned = (*MMOs.first) && (*MMOs.first)->getAlignment() >= 16;
2832 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
2835 &AddrOps[0], AddrOps.size());
2836 NewNodes.push_back(Store);
2838 // Preserve memory reference information.
2839 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
2845 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
2846 bool UnfoldLoad, bool UnfoldStore,
2847 unsigned *LoadRegIndex) const {
2848 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
2849 MemOp2RegOpTable.find(Opc);
2850 if (I == MemOp2RegOpTable.end())
2852 bool FoldedLoad = I->second.second & (1 << 4);
2853 bool FoldedStore = I->second.second & (1 << 5);
2854 if (UnfoldLoad && !FoldedLoad)
2856 if (UnfoldStore && !FoldedStore)
2859 *LoadRegIndex = I->second.second & 0xf;
2860 return I->second.first;
2864 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
2865 int64_t &Offset1, int64_t &Offset2) const {
2866 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
2868 unsigned Opc1 = Load1->getMachineOpcode();
2869 unsigned Opc2 = Load2->getMachineOpcode();
2871 default: return false;
2881 case X86::MMX_MOVD64rm:
2882 case X86::MMX_MOVQ64rm:
2883 case X86::FsMOVAPSrm:
2884 case X86::FsMOVAPDrm:
2890 case X86::VMOVAPSYrm:
2891 case X86::VMOVUPSYrm:
2892 case X86::VMOVAPDYrm:
2893 case X86::VMOVDQAYrm:
2894 case X86::VMOVDQUYrm:
2898 default: return false;
2908 case X86::MMX_MOVD64rm:
2909 case X86::MMX_MOVQ64rm:
2910 case X86::FsMOVAPSrm:
2911 case X86::FsMOVAPDrm:
2917 case X86::VMOVAPSYrm:
2918 case X86::VMOVUPSYrm:
2919 case X86::VMOVAPDYrm:
2920 case X86::VMOVDQAYrm:
2921 case X86::VMOVDQUYrm:
2925 // Check if chain operands and base addresses match.
2926 if (Load1->getOperand(0) != Load2->getOperand(0) ||
2927 Load1->getOperand(5) != Load2->getOperand(5))
2929 // Segment operands should match as well.
2930 if (Load1->getOperand(4) != Load2->getOperand(4))
2932 // Scale should be 1, Index should be Reg0.
2933 if (Load1->getOperand(1) == Load2->getOperand(1) &&
2934 Load1->getOperand(2) == Load2->getOperand(2)) {
2935 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
2938 // Now let's examine the displacements.
2939 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
2940 isa<ConstantSDNode>(Load2->getOperand(3))) {
2941 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
2942 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
2949 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
2950 int64_t Offset1, int64_t Offset2,
2951 unsigned NumLoads) const {
2952 assert(Offset2 > Offset1);
2953 if ((Offset2 - Offset1) / 8 > 64)
2956 unsigned Opc1 = Load1->getMachineOpcode();
2957 unsigned Opc2 = Load2->getMachineOpcode();
2959 return false; // FIXME: overly conservative?
2966 case X86::MMX_MOVD64rm:
2967 case X86::MMX_MOVQ64rm:
2971 EVT VT = Load1->getValueType(0);
2972 switch (VT.getSimpleVT().SimpleTy) {
2974 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
2975 // have 16 of them to play with.
2976 if (TM.getSubtargetImpl()->is64Bit()) {
2979 } else if (NumLoads) {
2999 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
3000 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
3001 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
3002 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
3004 Cond[0].setImm(GetOppositeBranchCondition(CC));
3009 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
3010 // FIXME: Return false for x87 stack register classes for now. We can't
3011 // allow any loads of these registers before FpGet_ST0_80.
3012 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
3013 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
3016 /// getGlobalBaseReg - Return a virtual register initialized with the
3017 /// the global base register value. Output instructions required to
3018 /// initialize the register in the function entry block, if necessary.
3020 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
3022 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3023 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3024 "X86-64 PIC uses RIP relative addressing");
3026 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3027 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3028 if (GlobalBaseReg != 0)
3029 return GlobalBaseReg;
3031 // Create the register. The code to initialize it is inserted
3032 // later, by the CGBR pass (below).
3033 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3034 GlobalBaseReg = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3035 X86FI->setGlobalBaseReg(GlobalBaseReg);
3036 return GlobalBaseReg;
3039 // These are the replaceable SSE instructions. Some of these have Int variants
3040 // that we don't include here. We don't want to replace instructions selected
3042 static const unsigned ReplaceableInstrs[][3] = {
3043 //PackedSingle PackedDouble PackedInt
3044 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
3045 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
3046 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
3047 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
3048 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
3049 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
3050 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
3051 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
3052 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
3053 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
3054 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
3055 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
3056 { X86::V_SET0PS, X86::V_SET0PD, X86::V_SET0PI },
3057 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
3058 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
3059 // AVX 128-bit support
3060 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
3061 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
3062 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
3063 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
3064 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
3065 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
3066 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
3067 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
3068 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
3069 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
3070 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
3071 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
3072 { X86::AVX_SET0PS, X86::AVX_SET0PD, X86::AVX_SET0PI },
3073 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
3074 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
3075 // AVX 256-bit support
3076 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
3077 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
3078 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
3079 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
3080 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
3081 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr },
3084 // FIXME: Some shuffle and unpack instructions have equivalents in different
3085 // domains, but they require a bit more work than just switching opcodes.
3087 static const unsigned *lookup(unsigned opcode, unsigned domain) {
3088 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
3089 if (ReplaceableInstrs[i][domain-1] == opcode)
3090 return ReplaceableInstrs[i];
3094 std::pair<uint16_t, uint16_t>
3095 X86InstrInfo::GetSSEDomain(const MachineInstr *MI) const {
3096 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
3097 return std::make_pair(domain,
3098 domain && lookup(MI->getOpcode(), domain) ? 0xe : 0);
3101 void X86InstrInfo::SetSSEDomain(MachineInstr *MI, unsigned Domain) const {
3102 assert(Domain>0 && Domain<4 && "Invalid execution domain");
3103 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
3104 assert(dom && "Not an SSE instruction");
3105 const unsigned *table = lookup(MI->getOpcode(), dom);
3106 assert(table && "Cannot change domain");
3107 MI->setDesc(get(table[Domain-1]));
3110 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
3111 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
3112 NopInst.setOpcode(X86::NOOP);
3115 bool X86InstrInfo::isHighLatencyDef(int opc) const {
3117 default: return false;
3119 case X86::DIVSDrm_Int:
3121 case X86::DIVSDrr_Int:
3123 case X86::DIVSSrm_Int:
3125 case X86::DIVSSrr_Int:
3127 case X86::SQRTPDm_Int:
3129 case X86::SQRTPDr_Int:
3131 case X86::SQRTPSm_Int:
3133 case X86::SQRTPSr_Int:
3135 case X86::SQRTSDm_Int:
3137 case X86::SQRTSDr_Int:
3139 case X86::SQRTSSm_Int:
3141 case X86::SQRTSSr_Int:
3147 hasHighOperandLatency(const InstrItineraryData *ItinData,
3148 const MachineRegisterInfo *MRI,
3149 const MachineInstr *DefMI, unsigned DefIdx,
3150 const MachineInstr *UseMI, unsigned UseIdx) const {
3151 return isHighLatencyDef(DefMI->getOpcode());
3155 /// CGBR - Create Global Base Reg pass. This initializes the PIC
3156 /// global base register for x86-32.
3157 struct CGBR : public MachineFunctionPass {
3159 CGBR() : MachineFunctionPass(ID) {}
3161 virtual bool runOnMachineFunction(MachineFunction &MF) {
3162 const X86TargetMachine *TM =
3163 static_cast<const X86TargetMachine *>(&MF.getTarget());
3165 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
3166 "X86-64 PIC uses RIP relative addressing");
3168 // Only emit a global base reg in PIC mode.
3169 if (TM->getRelocationModel() != Reloc::PIC_)
3172 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
3173 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3175 // If we didn't need a GlobalBaseReg, don't insert code.
3176 if (GlobalBaseReg == 0)
3179 // Insert the set of GlobalBaseReg into the first MBB of the function
3180 MachineBasicBlock &FirstMBB = MF.front();
3181 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3182 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
3183 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3184 const X86InstrInfo *TII = TM->getInstrInfo();
3187 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
3188 PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
3192 // Operand of MovePCtoStack is completely ignored by asm printer. It's
3193 // only used in JIT code emission as displacement to pc.
3194 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
3196 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
3197 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
3198 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
3199 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
3200 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
3201 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
3202 X86II::MO_GOT_ABSOLUTE_ADDRESS);
3208 virtual const char *getPassName() const {
3209 return "X86 PIC Global Base Reg Initialization";
3212 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
3213 AU.setPreservesCFG();
3214 MachineFunctionPass::getAnalysisUsage(AU);
3221 llvm::createGlobalBaseRegPass() { return new CGBR(); }