1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
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/ADT/STLExtras.h"
21 #include "llvm/CodeGen/LiveVariables.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineDominators.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/StackMaps.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/MC/MCAsmInfo.h"
31 #include "llvm/MC/MCInst.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 #define GET_INSTRINFO_CTOR_DTOR
40 #include "X86GenInstrInfo.inc"
45 NoFusing("disable-spill-fusing",
46 cl::desc("Disable fusing of spill code into instructions"));
48 PrintFailedFusing("print-failed-fuse-candidates",
49 cl::desc("Print instructions that the allocator wants to"
50 " fuse, but the X86 backend currently can't"),
53 ReMatPICStubLoad("remat-pic-stub-load",
54 cl::desc("Re-materialize load from stub in PIC mode"),
55 cl::init(false), cl::Hidden);
58 // Select which memory operand is being unfolded.
59 // (stored in bits 0 - 3)
66 // Do not insert the reverse map (MemOp -> RegOp) into the table.
67 // This may be needed because there is a many -> one mapping.
68 TB_NO_REVERSE = 1 << 4,
70 // Do not insert the forward map (RegOp -> MemOp) into the table.
71 // This is needed for Native Client, which prohibits branch
72 // instructions from using a memory operand.
73 TB_NO_FORWARD = 1 << 5,
75 TB_FOLDED_LOAD = 1 << 6,
76 TB_FOLDED_STORE = 1 << 7,
78 // Minimum alignment required for load/store.
79 // Used for RegOp->MemOp conversion.
80 // (stored in bits 8 - 15)
82 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
83 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
84 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
85 TB_ALIGN_64 = 64 << TB_ALIGN_SHIFT,
86 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
89 struct X86OpTblEntry {
95 // Pin the vtable to this file.
96 void X86InstrInfo::anchor() {}
98 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
99 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
100 ? X86::ADJCALLSTACKDOWN64
101 : X86::ADJCALLSTACKDOWN32),
102 (tm.getSubtarget<X86Subtarget>().is64Bit()
103 ? X86::ADJCALLSTACKUP64
104 : X86::ADJCALLSTACKUP32)),
107 static const X86OpTblEntry OpTbl2Addr[] = {
108 { X86::ADC32ri, X86::ADC32mi, 0 },
109 { X86::ADC32ri8, X86::ADC32mi8, 0 },
110 { X86::ADC32rr, X86::ADC32mr, 0 },
111 { X86::ADC64ri32, X86::ADC64mi32, 0 },
112 { X86::ADC64ri8, X86::ADC64mi8, 0 },
113 { X86::ADC64rr, X86::ADC64mr, 0 },
114 { X86::ADD16ri, X86::ADD16mi, 0 },
115 { X86::ADD16ri8, X86::ADD16mi8, 0 },
116 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
117 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
118 { X86::ADD16rr, X86::ADD16mr, 0 },
119 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
120 { X86::ADD32ri, X86::ADD32mi, 0 },
121 { X86::ADD32ri8, X86::ADD32mi8, 0 },
122 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
123 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
124 { X86::ADD32rr, X86::ADD32mr, 0 },
125 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
126 { X86::ADD64ri32, X86::ADD64mi32, 0 },
127 { X86::ADD64ri8, X86::ADD64mi8, 0 },
128 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
129 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
130 { X86::ADD64rr, X86::ADD64mr, 0 },
131 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
132 { X86::ADD8ri, X86::ADD8mi, 0 },
133 { X86::ADD8rr, X86::ADD8mr, 0 },
134 { X86::AND16ri, X86::AND16mi, 0 },
135 { X86::AND16ri8, X86::AND16mi8, 0 },
136 { X86::AND16rr, X86::AND16mr, 0 },
137 { X86::AND32ri, X86::AND32mi, 0 },
138 { X86::AND32ri8, X86::AND32mi8, 0 },
139 { X86::AND32rr, X86::AND32mr, 0 },
140 { X86::AND64ri32, X86::AND64mi32, 0 },
141 { X86::AND64ri8, X86::AND64mi8, 0 },
142 { X86::AND64rr, X86::AND64mr, 0 },
143 { X86::AND8ri, X86::AND8mi, 0 },
144 { X86::AND8rr, X86::AND8mr, 0 },
145 { X86::DEC16r, X86::DEC16m, 0 },
146 { X86::DEC32r, X86::DEC32m, 0 },
147 { X86::DEC64_16r, X86::DEC64_16m, 0 },
148 { X86::DEC64_32r, X86::DEC64_32m, 0 },
149 { X86::DEC64r, X86::DEC64m, 0 },
150 { X86::DEC8r, X86::DEC8m, 0 },
151 { X86::INC16r, X86::INC16m, 0 },
152 { X86::INC32r, X86::INC32m, 0 },
153 { X86::INC64_16r, X86::INC64_16m, 0 },
154 { X86::INC64_32r, X86::INC64_32m, 0 },
155 { X86::INC64r, X86::INC64m, 0 },
156 { X86::INC8r, X86::INC8m, 0 },
157 { X86::NEG16r, X86::NEG16m, 0 },
158 { X86::NEG32r, X86::NEG32m, 0 },
159 { X86::NEG64r, X86::NEG64m, 0 },
160 { X86::NEG8r, X86::NEG8m, 0 },
161 { X86::NOT16r, X86::NOT16m, 0 },
162 { X86::NOT32r, X86::NOT32m, 0 },
163 { X86::NOT64r, X86::NOT64m, 0 },
164 { X86::NOT8r, X86::NOT8m, 0 },
165 { X86::OR16ri, X86::OR16mi, 0 },
166 { X86::OR16ri8, X86::OR16mi8, 0 },
167 { X86::OR16rr, X86::OR16mr, 0 },
168 { X86::OR32ri, X86::OR32mi, 0 },
169 { X86::OR32ri8, X86::OR32mi8, 0 },
170 { X86::OR32rr, X86::OR32mr, 0 },
171 { X86::OR64ri32, X86::OR64mi32, 0 },
172 { X86::OR64ri8, X86::OR64mi8, 0 },
173 { X86::OR64rr, X86::OR64mr, 0 },
174 { X86::OR8ri, X86::OR8mi, 0 },
175 { X86::OR8rr, X86::OR8mr, 0 },
176 { X86::ROL16r1, X86::ROL16m1, 0 },
177 { X86::ROL16rCL, X86::ROL16mCL, 0 },
178 { X86::ROL16ri, X86::ROL16mi, 0 },
179 { X86::ROL32r1, X86::ROL32m1, 0 },
180 { X86::ROL32rCL, X86::ROL32mCL, 0 },
181 { X86::ROL32ri, X86::ROL32mi, 0 },
182 { X86::ROL64r1, X86::ROL64m1, 0 },
183 { X86::ROL64rCL, X86::ROL64mCL, 0 },
184 { X86::ROL64ri, X86::ROL64mi, 0 },
185 { X86::ROL8r1, X86::ROL8m1, 0 },
186 { X86::ROL8rCL, X86::ROL8mCL, 0 },
187 { X86::ROL8ri, X86::ROL8mi, 0 },
188 { X86::ROR16r1, X86::ROR16m1, 0 },
189 { X86::ROR16rCL, X86::ROR16mCL, 0 },
190 { X86::ROR16ri, X86::ROR16mi, 0 },
191 { X86::ROR32r1, X86::ROR32m1, 0 },
192 { X86::ROR32rCL, X86::ROR32mCL, 0 },
193 { X86::ROR32ri, X86::ROR32mi, 0 },
194 { X86::ROR64r1, X86::ROR64m1, 0 },
195 { X86::ROR64rCL, X86::ROR64mCL, 0 },
196 { X86::ROR64ri, X86::ROR64mi, 0 },
197 { X86::ROR8r1, X86::ROR8m1, 0 },
198 { X86::ROR8rCL, X86::ROR8mCL, 0 },
199 { X86::ROR8ri, X86::ROR8mi, 0 },
200 { X86::SAR16r1, X86::SAR16m1, 0 },
201 { X86::SAR16rCL, X86::SAR16mCL, 0 },
202 { X86::SAR16ri, X86::SAR16mi, 0 },
203 { X86::SAR32r1, X86::SAR32m1, 0 },
204 { X86::SAR32rCL, X86::SAR32mCL, 0 },
205 { X86::SAR32ri, X86::SAR32mi, 0 },
206 { X86::SAR64r1, X86::SAR64m1, 0 },
207 { X86::SAR64rCL, X86::SAR64mCL, 0 },
208 { X86::SAR64ri, X86::SAR64mi, 0 },
209 { X86::SAR8r1, X86::SAR8m1, 0 },
210 { X86::SAR8rCL, X86::SAR8mCL, 0 },
211 { X86::SAR8ri, X86::SAR8mi, 0 },
212 { X86::SBB32ri, X86::SBB32mi, 0 },
213 { X86::SBB32ri8, X86::SBB32mi8, 0 },
214 { X86::SBB32rr, X86::SBB32mr, 0 },
215 { X86::SBB64ri32, X86::SBB64mi32, 0 },
216 { X86::SBB64ri8, X86::SBB64mi8, 0 },
217 { X86::SBB64rr, X86::SBB64mr, 0 },
218 { X86::SHL16rCL, X86::SHL16mCL, 0 },
219 { X86::SHL16ri, X86::SHL16mi, 0 },
220 { X86::SHL32rCL, X86::SHL32mCL, 0 },
221 { X86::SHL32ri, X86::SHL32mi, 0 },
222 { X86::SHL64rCL, X86::SHL64mCL, 0 },
223 { X86::SHL64ri, X86::SHL64mi, 0 },
224 { X86::SHL8rCL, X86::SHL8mCL, 0 },
225 { X86::SHL8ri, X86::SHL8mi, 0 },
226 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
227 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
228 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
229 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
230 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
231 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
232 { X86::SHR16r1, X86::SHR16m1, 0 },
233 { X86::SHR16rCL, X86::SHR16mCL, 0 },
234 { X86::SHR16ri, X86::SHR16mi, 0 },
235 { X86::SHR32r1, X86::SHR32m1, 0 },
236 { X86::SHR32rCL, X86::SHR32mCL, 0 },
237 { X86::SHR32ri, X86::SHR32mi, 0 },
238 { X86::SHR64r1, X86::SHR64m1, 0 },
239 { X86::SHR64rCL, X86::SHR64mCL, 0 },
240 { X86::SHR64ri, X86::SHR64mi, 0 },
241 { X86::SHR8r1, X86::SHR8m1, 0 },
242 { X86::SHR8rCL, X86::SHR8mCL, 0 },
243 { X86::SHR8ri, X86::SHR8mi, 0 },
244 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
245 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
246 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
247 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
248 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
249 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
250 { X86::SUB16ri, X86::SUB16mi, 0 },
251 { X86::SUB16ri8, X86::SUB16mi8, 0 },
252 { X86::SUB16rr, X86::SUB16mr, 0 },
253 { X86::SUB32ri, X86::SUB32mi, 0 },
254 { X86::SUB32ri8, X86::SUB32mi8, 0 },
255 { X86::SUB32rr, X86::SUB32mr, 0 },
256 { X86::SUB64ri32, X86::SUB64mi32, 0 },
257 { X86::SUB64ri8, X86::SUB64mi8, 0 },
258 { X86::SUB64rr, X86::SUB64mr, 0 },
259 { X86::SUB8ri, X86::SUB8mi, 0 },
260 { X86::SUB8rr, X86::SUB8mr, 0 },
261 { X86::XOR16ri, X86::XOR16mi, 0 },
262 { X86::XOR16ri8, X86::XOR16mi8, 0 },
263 { X86::XOR16rr, X86::XOR16mr, 0 },
264 { X86::XOR32ri, X86::XOR32mi, 0 },
265 { X86::XOR32ri8, X86::XOR32mi8, 0 },
266 { X86::XOR32rr, X86::XOR32mr, 0 },
267 { X86::XOR64ri32, X86::XOR64mi32, 0 },
268 { X86::XOR64ri8, X86::XOR64mi8, 0 },
269 { X86::XOR64rr, X86::XOR64mr, 0 },
270 { X86::XOR8ri, X86::XOR8mi, 0 },
271 { X86::XOR8rr, X86::XOR8mr, 0 }
274 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
275 unsigned RegOp = OpTbl2Addr[i].RegOp;
276 unsigned MemOp = OpTbl2Addr[i].MemOp;
277 unsigned Flags = OpTbl2Addr[i].Flags;
278 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
280 // Index 0, folded load and store, no alignment requirement.
281 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
284 static const X86OpTblEntry OpTbl0[] = {
285 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
286 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
287 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
288 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
289 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
290 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
291 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
292 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
293 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
294 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
295 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
296 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
297 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
298 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
299 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
300 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
301 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
302 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
303 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
304 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
305 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
306 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
307 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
308 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
309 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
310 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
311 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
312 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
313 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
314 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
315 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
316 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
317 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
318 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
319 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
320 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
321 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
322 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
323 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
324 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
325 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
326 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
327 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
328 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
329 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
330 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
331 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
332 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
333 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
334 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
335 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
336 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
337 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
338 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
339 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
340 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
341 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
342 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
343 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
344 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
345 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
346 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
347 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
348 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
349 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
350 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
351 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
352 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
353 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
354 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
355 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
356 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
357 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
358 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
359 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
360 // AVX 128-bit versions of foldable instructions
361 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
362 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
366 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
367 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
368 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
369 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
370 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
371 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
372 // AVX 256-bit foldable instructions
373 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
374 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
377 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
378 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE },
379 // AVX-512 foldable instructions
380 { X86::VMOVPDI2DIZrr,X86::VMOVPDI2DIZmr, TB_FOLDED_STORE }
383 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
384 unsigned RegOp = OpTbl0[i].RegOp;
385 unsigned MemOp = OpTbl0[i].MemOp;
386 unsigned Flags = OpTbl0[i].Flags;
387 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
388 RegOp, MemOp, TB_INDEX_0 | Flags);
391 static const X86OpTblEntry OpTbl1[] = {
392 { X86::CMP16rr, X86::CMP16rm, 0 },
393 { X86::CMP32rr, X86::CMP32rm, 0 },
394 { X86::CMP64rr, X86::CMP64rm, 0 },
395 { X86::CMP8rr, X86::CMP8rm, 0 },
396 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
397 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
398 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
399 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
400 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
401 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
402 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
403 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
404 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
405 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
406 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
407 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
408 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
409 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
410 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
411 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
412 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
413 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
414 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
415 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
416 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
417 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
418 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
419 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
420 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
421 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
422 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
423 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
424 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
425 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
426 { X86::MOV16rr, X86::MOV16rm, 0 },
427 { X86::MOV32rr, X86::MOV32rm, 0 },
428 { X86::MOV64rr, X86::MOV64rm, 0 },
429 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
430 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
431 { X86::MOV8rr, X86::MOV8rm, 0 },
432 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
433 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
434 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
435 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
436 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
437 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
438 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
439 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
440 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
441 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
442 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
443 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
444 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
445 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
446 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
447 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
448 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
449 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
450 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
451 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
452 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
453 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
454 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
455 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
456 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
457 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
458 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
459 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
460 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
461 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
462 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
463 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
464 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
465 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
466 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
467 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
468 { X86::SQRTSDr, X86::SQRTSDm, 0 },
469 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
470 { X86::SQRTSSr, X86::SQRTSSm, 0 },
471 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
472 { X86::TEST16rr, X86::TEST16rm, 0 },
473 { X86::TEST32rr, X86::TEST32rm, 0 },
474 { X86::TEST64rr, X86::TEST64rm, 0 },
475 { X86::TEST8rr, X86::TEST8rm, 0 },
476 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
477 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
478 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
479 // AVX 128-bit versions of foldable instructions
480 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
481 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
482 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
483 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
484 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
485 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
486 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
487 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
488 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
489 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
490 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
491 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
492 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
493 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
494 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
495 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
496 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
497 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
498 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
499 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
500 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
501 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
502 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
503 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
504 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
505 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
506 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
507 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
508 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
509 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
510 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
511 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
512 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
513 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
514 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
515 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
516 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
517 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
518 { X86::VRCPPSr, X86::VRCPPSm, 0 },
519 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
520 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
521 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
522 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
523 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
524 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
525 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
526 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
528 // AVX 256-bit foldable instructions
529 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
530 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
531 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
532 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
533 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
534 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
535 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
537 // AVX2 foldable instructions
538 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
539 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
540 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
541 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
542 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
543 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
544 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
545 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
546 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
547 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
548 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
549 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
550 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
552 // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
553 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
554 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
555 { X86::BEXTRI32ri, X86::BEXTRI32mi, 0 },
556 { X86::BEXTRI64ri, X86::BEXTRI64mi, 0 },
557 { X86::BLCFILL32rr, X86::BLCFILL32rm, 0 },
558 { X86::BLCFILL64rr, X86::BLCFILL64rm, 0 },
559 { X86::BLCI32rr, X86::BLCI32rm, 0 },
560 { X86::BLCI64rr, X86::BLCI64rm, 0 },
561 { X86::BLCIC32rr, X86::BLCIC32rm, 0 },
562 { X86::BLCIC64rr, X86::BLCIC64rm, 0 },
563 { X86::BLCMSK32rr, X86::BLCMSK32rm, 0 },
564 { X86::BLCMSK64rr, X86::BLCMSK64rm, 0 },
565 { X86::BLCS32rr, X86::BLCS32rm, 0 },
566 { X86::BLCS64rr, X86::BLCS64rm, 0 },
567 { X86::BLSFILL32rr, X86::BLSFILL32rm, 0 },
568 { X86::BLSFILL64rr, X86::BLSFILL64rm, 0 },
569 { X86::BLSI32rr, X86::BLSI32rm, 0 },
570 { X86::BLSI64rr, X86::BLSI64rm, 0 },
571 { X86::BLSIC32rr, X86::BLSIC32rm, 0 },
572 { X86::BLSIC64rr, X86::BLSIC64rm, 0 },
573 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
574 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
575 { X86::BLSR32rr, X86::BLSR32rm, 0 },
576 { X86::BLSR64rr, X86::BLSR64rm, 0 },
577 { X86::BZHI32rr, X86::BZHI32rm, 0 },
578 { X86::BZHI64rr, X86::BZHI64rm, 0 },
579 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
580 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
581 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
582 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
583 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
584 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
585 { X86::RORX32ri, X86::RORX32mi, 0 },
586 { X86::RORX64ri, X86::RORX64mi, 0 },
587 { X86::SARX32rr, X86::SARX32rm, 0 },
588 { X86::SARX64rr, X86::SARX64rm, 0 },
589 { X86::SHRX32rr, X86::SHRX32rm, 0 },
590 { X86::SHRX64rr, X86::SHRX64rm, 0 },
591 { X86::SHLX32rr, X86::SHLX32rm, 0 },
592 { X86::SHLX64rr, X86::SHLX64rm, 0 },
593 { X86::T1MSKC32rr, X86::T1MSKC32rm, 0 },
594 { X86::T1MSKC64rr, X86::T1MSKC64rm, 0 },
595 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
596 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
597 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
598 { X86::TZMSK32rr, X86::TZMSK32rm, 0 },
599 { X86::TZMSK64rr, X86::TZMSK64rm, 0 },
601 // AVX-512 foldable instructions
602 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 },
603 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 },
604 { X86::VMOVDQA32rr, X86::VMOVDQA32rm, TB_ALIGN_64 },
605 { X86::VMOVDQA64rr, X86::VMOVDQA64rm, TB_ALIGN_64 },
606 { X86::VMOVDQU32rr, X86::VMOVDQU32rm, 0 },
607 { X86::VMOVDQU64rr, X86::VMOVDQU64rm, 0 },
608 { X86::VPABSDZrr, X86::VPABSDZrm, 0 },
609 { X86::VPABSQZrr, X86::VPABSQZrm, 0 },
611 // AES foldable instructions
612 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 },
613 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 },
614 { X86::VAESIMCrr, X86::VAESIMCrm, TB_ALIGN_16 },
615 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, TB_ALIGN_16 },
618 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
619 unsigned RegOp = OpTbl1[i].RegOp;
620 unsigned MemOp = OpTbl1[i].MemOp;
621 unsigned Flags = OpTbl1[i].Flags;
622 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
624 // Index 1, folded load
625 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
628 static const X86OpTblEntry OpTbl2[] = {
629 { X86::ADC32rr, X86::ADC32rm, 0 },
630 { X86::ADC64rr, X86::ADC64rm, 0 },
631 { X86::ADD16rr, X86::ADD16rm, 0 },
632 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
633 { X86::ADD32rr, X86::ADD32rm, 0 },
634 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
635 { X86::ADD64rr, X86::ADD64rm, 0 },
636 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
637 { X86::ADD8rr, X86::ADD8rm, 0 },
638 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
639 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
640 { X86::ADDSDrr, X86::ADDSDrm, 0 },
641 { X86::ADDSSrr, X86::ADDSSrm, 0 },
642 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
643 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
644 { X86::AND16rr, X86::AND16rm, 0 },
645 { X86::AND32rr, X86::AND32rm, 0 },
646 { X86::AND64rr, X86::AND64rm, 0 },
647 { X86::AND8rr, X86::AND8rm, 0 },
648 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
649 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
650 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
651 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
652 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
653 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
654 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
655 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
656 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
657 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
658 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
659 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
660 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
661 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
662 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
663 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
664 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
665 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
666 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
667 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
668 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
669 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
670 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
671 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
672 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
673 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
674 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
675 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
676 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
677 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
678 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
679 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
680 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
681 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
682 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
683 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
684 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
685 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
686 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
687 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
688 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
689 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
690 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
691 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
692 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
693 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
694 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
695 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
696 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
697 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
698 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
699 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
700 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
701 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
702 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
703 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
704 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
705 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
706 { X86::CMPSDrr, X86::CMPSDrm, 0 },
707 { X86::CMPSSrr, X86::CMPSSrm, 0 },
708 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
709 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
710 { X86::DIVSDrr, X86::DIVSDrm, 0 },
711 { X86::DIVSSrr, X86::DIVSSrm, 0 },
712 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
713 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
714 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
715 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
716 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
717 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
718 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
719 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
720 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
721 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
722 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
723 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
724 { X86::IMUL16rr, X86::IMUL16rm, 0 },
725 { X86::IMUL32rr, X86::IMUL32rm, 0 },
726 { X86::IMUL64rr, X86::IMUL64rm, 0 },
727 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
728 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
729 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
730 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
731 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
732 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
733 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
734 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
735 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
736 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
737 { X86::MAXSDrr, X86::MAXSDrm, 0 },
738 { X86::MAXSSrr, X86::MAXSSrm, 0 },
739 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
740 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
741 { X86::MINSDrr, X86::MINSDrm, 0 },
742 { X86::MINSSrr, X86::MINSSrm, 0 },
743 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
744 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
745 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
746 { X86::MULSDrr, X86::MULSDrm, 0 },
747 { X86::MULSSrr, X86::MULSSrm, 0 },
748 { X86::OR16rr, X86::OR16rm, 0 },
749 { X86::OR32rr, X86::OR32rm, 0 },
750 { X86::OR64rr, X86::OR64rm, 0 },
751 { X86::OR8rr, X86::OR8rm, 0 },
752 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
753 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
754 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
755 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
756 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
757 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
758 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
759 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
760 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
761 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
762 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
763 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
764 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
765 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
766 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
767 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
768 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
769 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
770 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
771 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
772 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
773 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
774 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
775 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
776 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
777 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
778 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
779 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
780 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
781 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
782 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
783 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
784 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
785 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
786 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
787 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
788 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
789 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
790 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
791 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
792 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
793 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
794 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
795 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
796 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
797 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
798 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
799 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
800 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
801 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
802 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
803 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
804 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
805 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
806 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
807 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
808 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
809 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
810 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
811 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
812 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
813 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
814 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
815 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
816 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
817 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
818 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
819 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
820 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
821 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
822 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
823 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
824 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
825 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
826 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
827 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
828 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
829 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
830 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
831 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
832 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
833 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
834 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
835 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
836 { X86::SBB32rr, X86::SBB32rm, 0 },
837 { X86::SBB64rr, X86::SBB64rm, 0 },
838 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
839 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
840 { X86::SUB16rr, X86::SUB16rm, 0 },
841 { X86::SUB32rr, X86::SUB32rm, 0 },
842 { X86::SUB64rr, X86::SUB64rm, 0 },
843 { X86::SUB8rr, X86::SUB8rm, 0 },
844 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
845 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
846 { X86::SUBSDrr, X86::SUBSDrm, 0 },
847 { X86::SUBSSrr, X86::SUBSSrm, 0 },
848 // FIXME: TEST*rr -> swapped operand of TEST*mr.
849 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
850 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
851 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
852 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
853 { X86::XOR16rr, X86::XOR16rm, 0 },
854 { X86::XOR32rr, X86::XOR32rm, 0 },
855 { X86::XOR64rr, X86::XOR64rm, 0 },
856 { X86::XOR8rr, X86::XOR8rm, 0 },
857 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
858 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
859 // AVX 128-bit versions of foldable instructions
860 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
861 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
862 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
863 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
864 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
865 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
866 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
867 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
868 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
869 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
870 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
871 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
872 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
873 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
874 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
875 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
876 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
877 { X86::VADDPDrr, X86::VADDPDrm, 0 },
878 { X86::VADDPSrr, X86::VADDPSrm, 0 },
879 { X86::VADDSDrr, X86::VADDSDrm, 0 },
880 { X86::VADDSSrr, X86::VADDSSrm, 0 },
881 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
882 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
883 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
884 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
885 { X86::VANDPDrr, X86::VANDPDrm, 0 },
886 { X86::VANDPSrr, X86::VANDPSrm, 0 },
887 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
888 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
889 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
890 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
891 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
892 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
893 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
894 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
895 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
896 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
897 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
898 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
899 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
900 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
901 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
902 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
903 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
904 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
905 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
906 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
907 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
908 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
909 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
910 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
911 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
912 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
913 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
914 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
915 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
916 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
917 { X86::VMINPDrr, X86::VMINPDrm, 0 },
918 { X86::VMINPSrr, X86::VMINPSrm, 0 },
919 { X86::VMINSDrr, X86::VMINSDrm, 0 },
920 { X86::VMINSSrr, X86::VMINSSrm, 0 },
921 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
922 { X86::VMULPDrr, X86::VMULPDrm, 0 },
923 { X86::VMULPSrr, X86::VMULPSrm, 0 },
924 { X86::VMULSDrr, X86::VMULSDrm, 0 },
925 { X86::VMULSSrr, X86::VMULSSrm, 0 },
926 { X86::VORPDrr, X86::VORPDrm, 0 },
927 { X86::VORPSrr, X86::VORPSrm, 0 },
928 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
929 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
930 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
931 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
932 { X86::VPADDBrr, X86::VPADDBrm, 0 },
933 { X86::VPADDDrr, X86::VPADDDrm, 0 },
934 { X86::VPADDQrr, X86::VPADDQrm, 0 },
935 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
936 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
937 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
938 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
939 { X86::VPADDWrr, X86::VPADDWrm, 0 },
940 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
941 { X86::VPANDNrr, X86::VPANDNrm, 0 },
942 { X86::VPANDrr, X86::VPANDrm, 0 },
943 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
944 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
945 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
946 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
947 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
948 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
949 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
950 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
951 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
952 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
953 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
954 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
955 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
956 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
957 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
958 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
959 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
960 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
961 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
962 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
963 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
964 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
965 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
966 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
967 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
968 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
969 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
970 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
971 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
972 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
973 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
974 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
975 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
976 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
977 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
978 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
979 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
980 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
981 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
982 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
983 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
984 { X86::VPORrr, X86::VPORrm, 0 },
985 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
986 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
987 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
988 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
989 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
990 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
991 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
992 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
993 { X86::VPSRADrr, X86::VPSRADrm, 0 },
994 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
995 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
996 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
997 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
998 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
999 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
1000 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
1001 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
1002 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
1003 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
1004 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
1005 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
1006 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
1007 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
1008 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
1009 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
1010 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
1011 { X86::VPXORrr, X86::VPXORrm, 0 },
1012 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
1013 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
1014 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
1015 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
1016 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
1017 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
1018 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
1019 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
1020 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
1021 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
1022 { X86::VXORPDrr, X86::VXORPDrm, 0 },
1023 { X86::VXORPSrr, X86::VXORPSrm, 0 },
1024 // AVX 256-bit foldable instructions
1025 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
1026 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
1027 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
1028 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
1029 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1030 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1031 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1032 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1033 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1034 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1035 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1036 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1037 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1038 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1039 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1040 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1041 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1042 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1043 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1044 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1045 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1046 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
1047 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
1048 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
1049 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
1050 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1051 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1052 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1053 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1054 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1055 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1056 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1057 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1058 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1059 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1060 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1061 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1062 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1063 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1064 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1065 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1066 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
1067 // AVX2 foldable instructions
1068 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1069 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1070 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1071 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1072 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1073 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1074 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1075 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1076 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1077 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1078 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1079 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1080 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1081 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1082 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1083 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1084 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1085 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1086 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1087 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1088 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1089 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1090 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1091 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1092 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1093 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1094 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1095 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1096 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1097 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1098 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1099 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1100 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1101 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1102 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1103 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1104 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1105 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1106 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1107 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1108 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1109 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1110 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1111 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1112 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1113 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1114 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1115 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1116 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1117 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1118 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1119 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1120 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1121 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1122 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1123 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1124 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1125 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1126 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1127 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1128 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1129 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1130 { X86::VPORYrr, X86::VPORYrm, 0 },
1131 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1132 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1133 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1134 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1135 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1136 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1137 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1138 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1139 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1140 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1141 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1142 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1143 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1144 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1145 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1146 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1147 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1148 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1149 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1150 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1151 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1152 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1153 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1154 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1155 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1156 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1157 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1158 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1159 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1160 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1161 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1162 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1163 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1164 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1165 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1166 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1167 { X86::VPXORYrr, X86::VPXORYrm, 0 },
1168 // FIXME: add AVX 256-bit foldable instructions
1170 // FMA4 foldable patterns
1171 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1172 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
1173 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1174 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1175 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1176 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1177 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1178 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
1179 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1180 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1181 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1182 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1183 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1184 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
1185 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1186 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1187 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1188 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1189 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1190 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
1191 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1192 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1193 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1194 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1195 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1196 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1197 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1198 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1199 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1200 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1201 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1202 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
1204 // BMI/BMI2 foldable instructions
1205 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1206 { X86::ANDN64rr, X86::ANDN64rm, 0 },
1207 { X86::MULX32rr, X86::MULX32rm, 0 },
1208 { X86::MULX64rr, X86::MULX64rm, 0 },
1209 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1210 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1211 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1212 { X86::PEXT64rr, X86::PEXT64rm, 0 },
1214 // AVX-512 foldable instructions
1215 { X86::VADDPSZrr, X86::VADDPSZrm, 0 },
1216 { X86::VADDPDZrr, X86::VADDPDZrm, 0 },
1217 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 },
1218 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 },
1219 { X86::VMULPSZrr, X86::VMULPSZrm, 0 },
1220 { X86::VMULPDZrr, X86::VMULPDZrm, 0 },
1221 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 },
1222 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 },
1223 { X86::VMINPSZrr, X86::VMINPSZrm, 0 },
1224 { X86::VMINPDZrr, X86::VMINPDZrm, 0 },
1225 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 },
1226 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 },
1227 { X86::VPADDDZrr, X86::VPADDDZrm, 0 },
1228 { X86::VPADDQZrr, X86::VPADDQZrm, 0 },
1229 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1230 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1231 { X86::VPMAXSDZrr, X86::VPMAXSDZrm, 0 },
1232 { X86::VPMAXSQZrr, X86::VPMAXSQZrm, 0 },
1233 { X86::VPMAXUDZrr, X86::VPMAXUDZrm, 0 },
1234 { X86::VPMAXUQZrr, X86::VPMAXUQZrm, 0 },
1235 { X86::VPMINSDZrr, X86::VPMINSDZrm, 0 },
1236 { X86::VPMINSQZrr, X86::VPMINSQZrm, 0 },
1237 { X86::VPMINUDZrr, X86::VPMINUDZrm, 0 },
1238 { X86::VPMINUQZrr, X86::VPMINUQZrm, 0 },
1239 { X86::VPMULDQZrr, X86::VPMULDQZrm, 0 },
1240 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 },
1241 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 },
1242 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 },
1243 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 },
1244 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 },
1245 { X86::VPSUBDZrr, X86::VPSUBDZrm, 0 },
1246 { X86::VPSUBQZrr, X86::VPSUBQZrm, 0 },
1247 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 },
1248 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 },
1249 { X86::VALIGNQrri, X86::VALIGNQrmi, 0 },
1250 { X86::VALIGNDrri, X86::VALIGNDrmi, 0 },
1251 { X86::VPMULUDQZrr, X86::VPMULUDQZrm, 0 },
1253 // AES foldable instructions
1254 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 },
1255 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 },
1256 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 },
1257 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 },
1258 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, TB_ALIGN_16 },
1259 { X86::VAESDECrr, X86::VAESDECrm, TB_ALIGN_16 },
1260 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, TB_ALIGN_16 },
1261 { X86::VAESENCrr, X86::VAESENCrm, TB_ALIGN_16 },
1263 // SHA foldable instructions
1264 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 },
1265 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 },
1266 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 },
1267 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 },
1268 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 },
1269 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 },
1270 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 },
1273 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1274 unsigned RegOp = OpTbl2[i].RegOp;
1275 unsigned MemOp = OpTbl2[i].MemOp;
1276 unsigned Flags = OpTbl2[i].Flags;
1277 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1279 // Index 2, folded load
1280 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1283 static const X86OpTblEntry OpTbl3[] = {
1284 // FMA foldable instructions
1285 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, TB_ALIGN_NONE },
1286 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, TB_ALIGN_NONE },
1287 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, TB_ALIGN_NONE },
1288 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, TB_ALIGN_NONE },
1289 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, TB_ALIGN_NONE },
1290 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, TB_ALIGN_NONE },
1292 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_NONE },
1293 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_NONE },
1294 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_NONE },
1295 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_NONE },
1296 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_NONE },
1297 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_NONE },
1298 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_NONE },
1299 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_NONE },
1300 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_NONE },
1301 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_NONE },
1302 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_NONE },
1303 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_NONE },
1305 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, TB_ALIGN_NONE },
1306 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, TB_ALIGN_NONE },
1307 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, TB_ALIGN_NONE },
1308 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, TB_ALIGN_NONE },
1309 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, TB_ALIGN_NONE },
1310 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, TB_ALIGN_NONE },
1312 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_NONE },
1313 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_NONE },
1314 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_NONE },
1315 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_NONE },
1316 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_NONE },
1317 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_NONE },
1318 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_NONE },
1319 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_NONE },
1320 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_NONE },
1321 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_NONE },
1322 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_NONE },
1323 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_NONE },
1325 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, TB_ALIGN_NONE },
1326 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, TB_ALIGN_NONE },
1327 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, TB_ALIGN_NONE },
1328 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, TB_ALIGN_NONE },
1329 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, TB_ALIGN_NONE },
1330 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, TB_ALIGN_NONE },
1332 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_NONE },
1333 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_NONE },
1334 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_NONE },
1335 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_NONE },
1336 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_NONE },
1337 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_NONE },
1338 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_NONE },
1339 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_NONE },
1340 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_NONE },
1341 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_NONE },
1342 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_NONE },
1343 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_NONE },
1345 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, TB_ALIGN_NONE },
1346 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, TB_ALIGN_NONE },
1347 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, TB_ALIGN_NONE },
1348 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, TB_ALIGN_NONE },
1349 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, TB_ALIGN_NONE },
1350 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, TB_ALIGN_NONE },
1352 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_NONE },
1353 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_NONE },
1354 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_NONE },
1355 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_NONE },
1356 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_NONE },
1357 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_NONE },
1358 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_NONE },
1359 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_NONE },
1360 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_NONE },
1361 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_NONE },
1362 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_NONE },
1363 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_NONE },
1365 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_NONE },
1366 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_NONE },
1367 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_NONE },
1368 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_NONE },
1369 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_NONE },
1370 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_NONE },
1371 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_NONE },
1372 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_NONE },
1373 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_NONE },
1374 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_NONE },
1375 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_NONE },
1376 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_NONE },
1378 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_NONE },
1379 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_NONE },
1380 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_NONE },
1381 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_NONE },
1382 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_NONE },
1383 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_NONE },
1384 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_NONE },
1385 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_NONE },
1386 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_NONE },
1387 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_NONE },
1388 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_NONE },
1389 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_NONE },
1391 // FMA4 foldable patterns
1392 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1393 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
1394 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1395 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1396 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1397 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1398 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1399 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
1400 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1401 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1402 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1403 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1404 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1405 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
1406 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1407 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1408 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1409 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1410 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1411 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
1412 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1413 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1414 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1415 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1416 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1417 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1418 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1419 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1420 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1421 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1422 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1423 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
1424 // AVX-512 VPERMI instructions with 3 source operands.
1425 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1426 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1427 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1428 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
1429 { X86::VBLENDMPDZrr, X86::VBLENDMPDZrm, 0 },
1430 { X86::VBLENDMPSZrr, X86::VBLENDMPSZrm, 0 },
1431 { X86::VPBLENDMDZrr, X86::VPBLENDMDZrm, 0 },
1432 { X86::VPBLENDMQZrr, X86::VPBLENDMQZrm, 0 }
1435 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1436 unsigned RegOp = OpTbl3[i].RegOp;
1437 unsigned MemOp = OpTbl3[i].MemOp;
1438 unsigned Flags = OpTbl3[i].Flags;
1439 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1441 // Index 3, folded load
1442 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1448 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1449 MemOp2RegOpTableType &M2RTable,
1450 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1451 if ((Flags & TB_NO_FORWARD) == 0) {
1452 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1453 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1455 if ((Flags & TB_NO_REVERSE) == 0) {
1456 assert(!M2RTable.count(MemOp) &&
1457 "Duplicated entries in unfolding maps?");
1458 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1463 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1464 unsigned &SrcReg, unsigned &DstReg,
1465 unsigned &SubIdx) const {
1466 switch (MI.getOpcode()) {
1468 case X86::MOVSX16rr8:
1469 case X86::MOVZX16rr8:
1470 case X86::MOVSX32rr8:
1471 case X86::MOVZX32rr8:
1472 case X86::MOVSX64rr8:
1473 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1474 // It's not always legal to reference the low 8-bit of the larger
1475 // register in 32-bit mode.
1477 case X86::MOVSX32rr16:
1478 case X86::MOVZX32rr16:
1479 case X86::MOVSX64rr16:
1480 case X86::MOVSX64rr32: {
1481 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1484 SrcReg = MI.getOperand(1).getReg();
1485 DstReg = MI.getOperand(0).getReg();
1486 switch (MI.getOpcode()) {
1487 default: llvm_unreachable("Unreachable!");
1488 case X86::MOVSX16rr8:
1489 case X86::MOVZX16rr8:
1490 case X86::MOVSX32rr8:
1491 case X86::MOVZX32rr8:
1492 case X86::MOVSX64rr8:
1493 SubIdx = X86::sub_8bit;
1495 case X86::MOVSX32rr16:
1496 case X86::MOVZX32rr16:
1497 case X86::MOVSX64rr16:
1498 SubIdx = X86::sub_16bit;
1500 case X86::MOVSX64rr32:
1501 SubIdx = X86::sub_32bit;
1510 /// isFrameOperand - Return true and the FrameIndex if the specified
1511 /// operand and follow operands form a reference to the stack frame.
1512 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1513 int &FrameIndex) const {
1514 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1515 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1516 MI->getOperand(Op+1).getImm() == 1 &&
1517 MI->getOperand(Op+2).getReg() == 0 &&
1518 MI->getOperand(Op+3).getImm() == 0) {
1519 FrameIndex = MI->getOperand(Op).getIndex();
1525 static bool isFrameLoadOpcode(int Opcode) {
1541 case X86::VMOVAPSrm:
1542 case X86::VMOVAPDrm:
1543 case X86::VMOVDQArm:
1544 case X86::VMOVAPSYrm:
1545 case X86::VMOVAPDYrm:
1546 case X86::VMOVDQAYrm:
1547 case X86::MMX_MOVD64rm:
1548 case X86::MMX_MOVQ64rm:
1549 case X86::VMOVAPSZrm:
1550 case X86::VMOVUPSZrm:
1555 static bool isFrameStoreOpcode(int Opcode) {
1562 case X86::ST_FpP64m:
1570 case X86::VMOVAPSmr:
1571 case X86::VMOVAPDmr:
1572 case X86::VMOVDQAmr:
1573 case X86::VMOVAPSYmr:
1574 case X86::VMOVAPDYmr:
1575 case X86::VMOVDQAYmr:
1576 case X86::VMOVUPSZmr:
1577 case X86::VMOVAPSZmr:
1578 case X86::MMX_MOVD64mr:
1579 case X86::MMX_MOVQ64mr:
1580 case X86::MMX_MOVNTQmr:
1586 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1587 int &FrameIndex) const {
1588 if (isFrameLoadOpcode(MI->getOpcode()))
1589 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1590 return MI->getOperand(0).getReg();
1594 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1595 int &FrameIndex) const {
1596 if (isFrameLoadOpcode(MI->getOpcode())) {
1598 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1600 // Check for post-frame index elimination operations
1601 const MachineMemOperand *Dummy;
1602 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1607 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1608 int &FrameIndex) const {
1609 if (isFrameStoreOpcode(MI->getOpcode()))
1610 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1611 isFrameOperand(MI, 0, FrameIndex))
1612 return MI->getOperand(X86::AddrNumOperands).getReg();
1616 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1617 int &FrameIndex) const {
1618 if (isFrameStoreOpcode(MI->getOpcode())) {
1620 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1622 // Check for post-frame index elimination operations
1623 const MachineMemOperand *Dummy;
1624 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1629 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1631 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1632 // Don't waste compile time scanning use-def chains of physregs.
1633 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1635 bool isPICBase = false;
1636 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
1637 E = MRI.def_instr_end(); I != E; ++I) {
1638 MachineInstr *DefMI = &*I;
1639 if (DefMI->getOpcode() != X86::MOVPC32r)
1641 assert(!isPICBase && "More than one PIC base?");
1648 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1649 AliasAnalysis *AA) const {
1650 switch (MI->getOpcode()) {
1666 case X86::VMOVAPSrm:
1667 case X86::VMOVUPSrm:
1668 case X86::VMOVAPDrm:
1669 case X86::VMOVDQArm:
1670 case X86::VMOVDQUrm:
1671 case X86::VMOVAPSYrm:
1672 case X86::VMOVUPSYrm:
1673 case X86::VMOVAPDYrm:
1674 case X86::VMOVDQAYrm:
1675 case X86::VMOVDQUYrm:
1676 case X86::MMX_MOVD64rm:
1677 case X86::MMX_MOVQ64rm:
1678 case X86::FsVMOVAPSrm:
1679 case X86::FsVMOVAPDrm:
1680 case X86::FsMOVAPSrm:
1681 case X86::FsMOVAPDrm: {
1682 // Loads from constant pools are trivially rematerializable.
1683 if (MI->getOperand(1).isReg() &&
1684 MI->getOperand(2).isImm() &&
1685 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1686 MI->isInvariantLoad(AA)) {
1687 unsigned BaseReg = MI->getOperand(1).getReg();
1688 if (BaseReg == 0 || BaseReg == X86::RIP)
1690 // Allow re-materialization of PIC load.
1691 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1693 const MachineFunction &MF = *MI->getParent()->getParent();
1694 const MachineRegisterInfo &MRI = MF.getRegInfo();
1695 return regIsPICBase(BaseReg, MRI);
1702 if (MI->getOperand(2).isImm() &&
1703 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1704 !MI->getOperand(4).isReg()) {
1705 // lea fi#, lea GV, etc. are all rematerializable.
1706 if (!MI->getOperand(1).isReg())
1708 unsigned BaseReg = MI->getOperand(1).getReg();
1711 // Allow re-materialization of lea PICBase + x.
1712 const MachineFunction &MF = *MI->getParent()->getParent();
1713 const MachineRegisterInfo &MRI = MF.getRegInfo();
1714 return regIsPICBase(BaseReg, MRI);
1720 // All other instructions marked M_REMATERIALIZABLE are always trivially
1721 // rematerializable.
1725 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1726 /// would clobber the EFLAGS condition register. Note the result may be
1727 /// conservative. If it cannot definitely determine the safety after visiting
1728 /// a few instructions in each direction it assumes it's not safe.
1729 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1730 MachineBasicBlock::iterator I) {
1731 MachineBasicBlock::iterator E = MBB.end();
1733 // For compile time consideration, if we are not able to determine the
1734 // safety after visiting 4 instructions in each direction, we will assume
1736 MachineBasicBlock::iterator Iter = I;
1737 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1738 bool SeenDef = false;
1739 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1740 MachineOperand &MO = Iter->getOperand(j);
1741 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1745 if (MO.getReg() == X86::EFLAGS) {
1753 // This instruction defines EFLAGS, no need to look any further.
1756 // Skip over DBG_VALUE.
1757 while (Iter != E && Iter->isDebugValue())
1761 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1764 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1765 SE = MBB.succ_end(); SI != SE; ++SI)
1766 if ((*SI)->isLiveIn(X86::EFLAGS))
1771 MachineBasicBlock::iterator B = MBB.begin();
1773 for (unsigned i = 0; i < 4; ++i) {
1774 // If we make it to the beginning of the block, it's safe to clobber
1775 // EFLAGS iff EFLAGS is not live-in.
1777 return !MBB.isLiveIn(X86::EFLAGS);
1780 // Skip over DBG_VALUE.
1781 while (Iter != B && Iter->isDebugValue())
1784 bool SawKill = false;
1785 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1786 MachineOperand &MO = Iter->getOperand(j);
1787 // A register mask may clobber EFLAGS, but we should still look for a
1789 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1791 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1792 if (MO.isDef()) return MO.isDead();
1793 if (MO.isKill()) SawKill = true;
1798 // This instruction kills EFLAGS and doesn't redefine it, so
1799 // there's no need to look further.
1803 // Conservative answer.
1807 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1808 MachineBasicBlock::iterator I,
1809 unsigned DestReg, unsigned SubIdx,
1810 const MachineInstr *Orig,
1811 const TargetRegisterInfo &TRI) const {
1812 // MOV32r0 is implemented with a xor which clobbers condition code.
1813 // Re-materialize it as movri instructions to avoid side effects.
1814 unsigned Opc = Orig->getOpcode();
1815 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1816 DebugLoc DL = Orig->getDebugLoc();
1817 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1820 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1824 MachineInstr *NewMI = std::prev(I);
1825 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1828 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1829 /// is not marked dead.
1830 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1831 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1832 MachineOperand &MO = MI->getOperand(i);
1833 if (MO.isReg() && MO.isDef() &&
1834 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1841 /// getTruncatedShiftCount - check whether the shift count for a machine operand
1843 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1844 unsigned ShiftAmtOperandIdx) {
1845 // The shift count is six bits with the REX.W prefix and five bits without.
1846 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1847 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1848 return Imm & ShiftCountMask;
1851 /// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1852 /// can be represented by a LEA instruction.
1853 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1854 // Left shift instructions can be transformed into load-effective-address
1855 // instructions if we can encode them appropriately.
1856 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1857 // The SIB.scale field is two bits wide which means that we can encode any
1858 // shift amount less than 4.
1859 return ShAmt < 4 && ShAmt > 0;
1862 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
1863 unsigned Opc, bool AllowSP,
1864 unsigned &NewSrc, bool &isKill, bool &isUndef,
1865 MachineOperand &ImplicitOp) const {
1866 MachineFunction &MF = *MI->getParent()->getParent();
1867 const TargetRegisterClass *RC;
1869 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1871 RC = Opc != X86::LEA32r ?
1872 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1874 unsigned SrcReg = Src.getReg();
1876 // For both LEA64 and LEA32 the register already has essentially the right
1877 // type (32-bit or 64-bit) we may just need to forbid SP.
1878 if (Opc != X86::LEA64_32r) {
1880 isKill = Src.isKill();
1881 isUndef = Src.isUndef();
1883 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
1884 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1890 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1891 // another we need to add 64-bit registers to the final MI.
1892 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
1894 ImplicitOp.setImplicit();
1896 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
1897 MachineBasicBlock::LivenessQueryResult LQR =
1898 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
1901 case MachineBasicBlock::LQR_Unknown:
1902 // We can't give sane liveness flags to the instruction, abandon LEA
1905 case MachineBasicBlock::LQR_Live:
1906 isKill = MI->killsRegister(SrcReg);
1910 // The physreg itself is dead, so we have to use it as an <undef>.
1916 // Virtual register of the wrong class, we have to create a temporary 64-bit
1917 // vreg to feed into the LEA.
1918 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1919 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1920 get(TargetOpcode::COPY))
1921 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1924 // Which is obviously going to be dead after we're done with it.
1929 // We've set all the parameters without issue.
1933 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1934 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1935 /// to a 32-bit superregister and then truncating back down to a 16-bit
1938 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1939 MachineFunction::iterator &MFI,
1940 MachineBasicBlock::iterator &MBBI,
1941 LiveVariables *LV) const {
1942 MachineInstr *MI = MBBI;
1943 unsigned Dest = MI->getOperand(0).getReg();
1944 unsigned Src = MI->getOperand(1).getReg();
1945 bool isDead = MI->getOperand(0).isDead();
1946 bool isKill = MI->getOperand(1).isKill();
1948 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1949 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1950 unsigned Opc, leaInReg;
1951 if (TM.getSubtarget<X86Subtarget>().is64Bit()) {
1952 Opc = X86::LEA64_32r;
1953 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1956 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1959 // Build and insert into an implicit UNDEF value. This is OK because
1960 // well be shifting and then extracting the lower 16-bits.
1961 // This has the potential to cause partial register stall. e.g.
1962 // movw (%rbp,%rcx,2), %dx
1963 // leal -65(%rdx), %esi
1964 // But testing has shown this *does* help performance in 64-bit mode (at
1965 // least on modern x86 machines).
1966 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1967 MachineInstr *InsMI =
1968 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1969 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1970 .addReg(Src, getKillRegState(isKill));
1972 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1973 get(Opc), leaOutReg);
1975 default: llvm_unreachable("Unreachable!");
1976 case X86::SHL16ri: {
1977 unsigned ShAmt = MI->getOperand(2).getImm();
1978 MIB.addReg(0).addImm(1 << ShAmt)
1979 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1983 case X86::INC64_16r:
1984 addRegOffset(MIB, leaInReg, true, 1);
1987 case X86::DEC64_16r:
1988 addRegOffset(MIB, leaInReg, true, -1);
1992 case X86::ADD16ri_DB:
1993 case X86::ADD16ri8_DB:
1994 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
1997 case X86::ADD16rr_DB: {
1998 unsigned Src2 = MI->getOperand(2).getReg();
1999 bool isKill2 = MI->getOperand(2).isKill();
2000 unsigned leaInReg2 = 0;
2001 MachineInstr *InsMI2 = 0;
2003 // ADD16rr %reg1028<kill>, %reg1028
2004 // just a single insert_subreg.
2005 addRegReg(MIB, leaInReg, true, leaInReg, false);
2007 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2008 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2010 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2011 // Build and insert into an implicit UNDEF value. This is OK because
2012 // well be shifting and then extracting the lower 16-bits.
2013 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2015 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2016 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2017 .addReg(Src2, getKillRegState(isKill2));
2018 addRegReg(MIB, leaInReg, true, leaInReg2, true);
2020 if (LV && isKill2 && InsMI2)
2021 LV->replaceKillInstruction(Src2, MI, InsMI2);
2026 MachineInstr *NewMI = MIB;
2027 MachineInstr *ExtMI =
2028 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2029 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2030 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2033 // Update live variables
2034 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2035 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2037 LV->replaceKillInstruction(Src, MI, InsMI);
2039 LV->replaceKillInstruction(Dest, MI, ExtMI);
2045 /// convertToThreeAddress - This method must be implemented by targets that
2046 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
2047 /// may be able to convert a two-address instruction into a true
2048 /// three-address instruction on demand. This allows the X86 target (for
2049 /// example) to convert ADD and SHL instructions into LEA instructions if they
2050 /// would require register copies due to two-addressness.
2052 /// This method returns a null pointer if the transformation cannot be
2053 /// performed, otherwise it returns the new instruction.
2056 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2057 MachineBasicBlock::iterator &MBBI,
2058 LiveVariables *LV) const {
2059 MachineInstr *MI = MBBI;
2061 // The following opcodes also sets the condition code register(s). Only
2062 // convert them to equivalent lea if the condition code register def's
2064 if (hasLiveCondCodeDef(MI))
2067 MachineFunction &MF = *MI->getParent()->getParent();
2068 // All instructions input are two-addr instructions. Get the known operands.
2069 const MachineOperand &Dest = MI->getOperand(0);
2070 const MachineOperand &Src = MI->getOperand(1);
2072 MachineInstr *NewMI = NULL;
2073 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
2074 // we have better subtarget support, enable the 16-bit LEA generation here.
2075 // 16-bit LEA is also slow on Core2.
2076 bool DisableLEA16 = true;
2077 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2079 unsigned MIOpc = MI->getOpcode();
2081 case X86::SHUFPSrri: {
2082 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
2083 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
2085 unsigned B = MI->getOperand(1).getReg();
2086 unsigned C = MI->getOperand(2).getReg();
2087 if (B != C) return 0;
2088 unsigned M = MI->getOperand(3).getImm();
2089 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
2090 .addOperand(Dest).addOperand(Src).addImm(M);
2093 case X86::SHUFPDrri: {
2094 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
2095 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
2097 unsigned B = MI->getOperand(1).getReg();
2098 unsigned C = MI->getOperand(2).getReg();
2099 if (B != C) return 0;
2100 unsigned M = MI->getOperand(3).getImm();
2102 // Convert to PSHUFD mask.
2103 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
2105 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
2106 .addOperand(Dest).addOperand(Src).addImm(M);
2109 case X86::SHL64ri: {
2110 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2111 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2112 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
2114 // LEA can't handle RSP.
2115 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2116 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2117 &X86::GR64_NOSPRegClass))
2120 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2122 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2125 case X86::SHL32ri: {
2126 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2127 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2128 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
2130 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2132 // LEA can't handle ESP.
2133 bool isKill, isUndef;
2135 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2136 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2137 SrcReg, isKill, isUndef, ImplicitOp))
2140 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2142 .addReg(0).addImm(1 << ShAmt)
2143 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2144 .addImm(0).addReg(0);
2145 if (ImplicitOp.getReg() != 0)
2146 MIB.addOperand(ImplicitOp);
2151 case X86::SHL16ri: {
2152 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2153 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2154 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
2157 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2158 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2160 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2169 case X86::INC64_32r: {
2170 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2171 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2172 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2173 bool isKill, isUndef;
2175 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2176 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2177 SrcReg, isKill, isUndef, ImplicitOp))
2180 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2182 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2183 if (ImplicitOp.getReg() != 0)
2184 MIB.addOperand(ImplicitOp);
2186 NewMI = addOffset(MIB, 1);
2190 case X86::INC64_16r:
2192 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2193 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2194 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2195 .addOperand(Dest).addOperand(Src), 1);
2199 case X86::DEC64_32r: {
2200 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2201 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2202 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2204 bool isKill, isUndef;
2206 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2207 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2208 SrcReg, isKill, isUndef, ImplicitOp))
2211 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2213 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2214 if (ImplicitOp.getReg() != 0)
2215 MIB.addOperand(ImplicitOp);
2217 NewMI = addOffset(MIB, -1);
2222 case X86::DEC64_16r:
2224 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2225 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2226 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2227 .addOperand(Dest).addOperand(Src), -1);
2230 case X86::ADD64rr_DB:
2232 case X86::ADD32rr_DB: {
2233 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2235 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2238 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2240 bool isKill, isUndef;
2242 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2243 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2244 SrcReg, isKill, isUndef, ImplicitOp))
2247 const MachineOperand &Src2 = MI->getOperand(2);
2248 bool isKill2, isUndef2;
2250 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2251 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2252 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2255 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2257 if (ImplicitOp.getReg() != 0)
2258 MIB.addOperand(ImplicitOp);
2259 if (ImplicitOp2.getReg() != 0)
2260 MIB.addOperand(ImplicitOp2);
2262 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2264 // Preserve undefness of the operands.
2265 NewMI->getOperand(1).setIsUndef(isUndef);
2266 NewMI->getOperand(3).setIsUndef(isUndef2);
2268 if (LV && Src2.isKill())
2269 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2273 case X86::ADD16rr_DB: {
2275 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2276 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2277 unsigned Src2 = MI->getOperand(2).getReg();
2278 bool isKill2 = MI->getOperand(2).isKill();
2279 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2281 Src.getReg(), Src.isKill(), Src2, isKill2);
2283 // Preserve undefness of the operands.
2284 bool isUndef = MI->getOperand(1).isUndef();
2285 bool isUndef2 = MI->getOperand(2).isUndef();
2286 NewMI->getOperand(1).setIsUndef(isUndef);
2287 NewMI->getOperand(3).setIsUndef(isUndef2);
2290 LV->replaceKillInstruction(Src2, MI, NewMI);
2293 case X86::ADD64ri32:
2295 case X86::ADD64ri32_DB:
2296 case X86::ADD64ri8_DB:
2297 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2298 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2299 .addOperand(Dest).addOperand(Src),
2300 MI->getOperand(2).getImm());
2304 case X86::ADD32ri_DB:
2305 case X86::ADD32ri8_DB: {
2306 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2307 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2309 bool isKill, isUndef;
2311 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2312 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2313 SrcReg, isKill, isUndef, ImplicitOp))
2316 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2318 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2319 if (ImplicitOp.getReg() != 0)
2320 MIB.addOperand(ImplicitOp);
2322 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2327 case X86::ADD16ri_DB:
2328 case X86::ADD16ri8_DB:
2330 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2331 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2332 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2333 .addOperand(Dest).addOperand(Src),
2334 MI->getOperand(2).getImm());
2340 if (!NewMI) return 0;
2342 if (LV) { // Update live variables
2344 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2346 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2349 MFI->insert(MBBI, NewMI); // Insert the new inst
2353 /// commuteInstruction - We have a few instructions that must be hacked on to
2357 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2358 switch (MI->getOpcode()) {
2359 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2360 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2361 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2362 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2363 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2364 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2367 switch (MI->getOpcode()) {
2368 default: llvm_unreachable("Unreachable!");
2369 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2370 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2371 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2372 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2373 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2374 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2376 unsigned Amt = MI->getOperand(3).getImm();
2378 MachineFunction &MF = *MI->getParent()->getParent();
2379 MI = MF.CloneMachineInstr(MI);
2382 MI->setDesc(get(Opc));
2383 MI->getOperand(3).setImm(Size-Amt);
2384 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2386 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2387 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2388 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2389 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2390 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2391 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2392 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2393 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2394 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2395 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2396 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2397 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2398 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2399 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2400 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2401 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2403 switch (MI->getOpcode()) {
2404 default: llvm_unreachable("Unreachable!");
2405 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2406 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2407 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2408 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2409 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2410 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2411 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2412 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2413 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2414 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2415 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2416 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2417 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2418 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2419 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2420 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2421 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2422 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2423 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2424 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2425 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2426 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2427 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2428 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2429 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2430 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2431 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2432 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2433 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2434 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2435 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2436 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2437 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2438 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2439 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2440 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2441 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2442 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2443 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2444 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2445 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2446 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2447 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2448 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2449 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2450 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2451 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2452 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2455 MachineFunction &MF = *MI->getParent()->getParent();
2456 MI = MF.CloneMachineInstr(MI);
2459 MI->setDesc(get(Opc));
2460 // Fallthrough intended.
2463 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2467 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
2468 unsigned &SrcOpIdx2) const {
2469 switch (MI->getOpcode()) {
2470 case X86::VFMADDPDr231r:
2471 case X86::VFMADDPSr231r:
2472 case X86::VFMADDSDr231r:
2473 case X86::VFMADDSSr231r:
2474 case X86::VFMSUBPDr231r:
2475 case X86::VFMSUBPSr231r:
2476 case X86::VFMSUBSDr231r:
2477 case X86::VFMSUBSSr231r:
2478 case X86::VFNMADDPDr231r:
2479 case X86::VFNMADDPSr231r:
2480 case X86::VFNMADDSDr231r:
2481 case X86::VFNMADDSSr231r:
2482 case X86::VFNMSUBPDr231r:
2483 case X86::VFNMSUBPSr231r:
2484 case X86::VFNMSUBSDr231r:
2485 case X86::VFNMSUBSSr231r:
2486 case X86::VFMADDPDr231rY:
2487 case X86::VFMADDPSr231rY:
2488 case X86::VFMSUBPDr231rY:
2489 case X86::VFMSUBPSr231rY:
2490 case X86::VFNMADDPDr231rY:
2491 case X86::VFNMADDPSr231rY:
2492 case X86::VFNMSUBPDr231rY:
2493 case X86::VFNMSUBPSr231rY:
2498 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2502 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2504 default: return X86::COND_INVALID;
2505 case X86::JE_4: return X86::COND_E;
2506 case X86::JNE_4: return X86::COND_NE;
2507 case X86::JL_4: return X86::COND_L;
2508 case X86::JLE_4: return X86::COND_LE;
2509 case X86::JG_4: return X86::COND_G;
2510 case X86::JGE_4: return X86::COND_GE;
2511 case X86::JB_4: return X86::COND_B;
2512 case X86::JBE_4: return X86::COND_BE;
2513 case X86::JA_4: return X86::COND_A;
2514 case X86::JAE_4: return X86::COND_AE;
2515 case X86::JS_4: return X86::COND_S;
2516 case X86::JNS_4: return X86::COND_NS;
2517 case X86::JP_4: return X86::COND_P;
2518 case X86::JNP_4: return X86::COND_NP;
2519 case X86::JO_4: return X86::COND_O;
2520 case X86::JNO_4: return X86::COND_NO;
2524 /// getCondFromSETOpc - return condition code of a SET opcode.
2525 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2527 default: return X86::COND_INVALID;
2528 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2529 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2530 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2531 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2532 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2533 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2534 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2535 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2536 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2537 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2538 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2539 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2540 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2541 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2542 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2543 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2547 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2548 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2550 default: return X86::COND_INVALID;
2551 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2552 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2554 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2555 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2556 return X86::COND_AE;
2557 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2558 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2560 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2561 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2562 return X86::COND_BE;
2563 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2564 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2566 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2567 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2569 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2570 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2571 return X86::COND_GE;
2572 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2573 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2575 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2576 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2577 return X86::COND_LE;
2578 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2579 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2580 return X86::COND_NE;
2581 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2582 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2583 return X86::COND_NO;
2584 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2585 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2586 return X86::COND_NP;
2587 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2588 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2589 return X86::COND_NS;
2590 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2591 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2593 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2594 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2596 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2597 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2602 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2604 default: llvm_unreachable("Illegal condition code!");
2605 case X86::COND_E: return X86::JE_4;
2606 case X86::COND_NE: return X86::JNE_4;
2607 case X86::COND_L: return X86::JL_4;
2608 case X86::COND_LE: return X86::JLE_4;
2609 case X86::COND_G: return X86::JG_4;
2610 case X86::COND_GE: return X86::JGE_4;
2611 case X86::COND_B: return X86::JB_4;
2612 case X86::COND_BE: return X86::JBE_4;
2613 case X86::COND_A: return X86::JA_4;
2614 case X86::COND_AE: return X86::JAE_4;
2615 case X86::COND_S: return X86::JS_4;
2616 case X86::COND_NS: return X86::JNS_4;
2617 case X86::COND_P: return X86::JP_4;
2618 case X86::COND_NP: return X86::JNP_4;
2619 case X86::COND_O: return X86::JO_4;
2620 case X86::COND_NO: return X86::JNO_4;
2624 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2625 /// e.g. turning COND_E to COND_NE.
2626 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2628 default: llvm_unreachable("Illegal condition code!");
2629 case X86::COND_E: return X86::COND_NE;
2630 case X86::COND_NE: return X86::COND_E;
2631 case X86::COND_L: return X86::COND_GE;
2632 case X86::COND_LE: return X86::COND_G;
2633 case X86::COND_G: return X86::COND_LE;
2634 case X86::COND_GE: return X86::COND_L;
2635 case X86::COND_B: return X86::COND_AE;
2636 case X86::COND_BE: return X86::COND_A;
2637 case X86::COND_A: return X86::COND_BE;
2638 case X86::COND_AE: return X86::COND_B;
2639 case X86::COND_S: return X86::COND_NS;
2640 case X86::COND_NS: return X86::COND_S;
2641 case X86::COND_P: return X86::COND_NP;
2642 case X86::COND_NP: return X86::COND_P;
2643 case X86::COND_O: return X86::COND_NO;
2644 case X86::COND_NO: return X86::COND_O;
2648 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2649 /// the condition code if we modify the instructions such that flags are
2651 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2653 default: return X86::COND_INVALID;
2654 case X86::COND_E: return X86::COND_E;
2655 case X86::COND_NE: return X86::COND_NE;
2656 case X86::COND_L: return X86::COND_G;
2657 case X86::COND_LE: return X86::COND_GE;
2658 case X86::COND_G: return X86::COND_L;
2659 case X86::COND_GE: return X86::COND_LE;
2660 case X86::COND_B: return X86::COND_A;
2661 case X86::COND_BE: return X86::COND_AE;
2662 case X86::COND_A: return X86::COND_B;
2663 case X86::COND_AE: return X86::COND_BE;
2667 /// getSETFromCond - Return a set opcode for the given condition and
2668 /// whether it has memory operand.
2669 static unsigned getSETFromCond(X86::CondCode CC,
2670 bool HasMemoryOperand) {
2671 static const uint16_t Opc[16][2] = {
2672 { X86::SETAr, X86::SETAm },
2673 { X86::SETAEr, X86::SETAEm },
2674 { X86::SETBr, X86::SETBm },
2675 { X86::SETBEr, X86::SETBEm },
2676 { X86::SETEr, X86::SETEm },
2677 { X86::SETGr, X86::SETGm },
2678 { X86::SETGEr, X86::SETGEm },
2679 { X86::SETLr, X86::SETLm },
2680 { X86::SETLEr, X86::SETLEm },
2681 { X86::SETNEr, X86::SETNEm },
2682 { X86::SETNOr, X86::SETNOm },
2683 { X86::SETNPr, X86::SETNPm },
2684 { X86::SETNSr, X86::SETNSm },
2685 { X86::SETOr, X86::SETOm },
2686 { X86::SETPr, X86::SETPm },
2687 { X86::SETSr, X86::SETSm }
2690 assert(CC < 16 && "Can only handle standard cond codes");
2691 return Opc[CC][HasMemoryOperand ? 1 : 0];
2694 /// getCMovFromCond - Return a cmov opcode for the given condition,
2695 /// register size in bytes, and operand type.
2696 static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2697 bool HasMemoryOperand) {
2698 static const uint16_t Opc[32][3] = {
2699 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2700 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2701 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2702 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2703 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2704 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2705 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2706 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2707 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2708 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2709 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2710 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2711 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2712 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2713 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2714 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2715 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2716 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2717 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2718 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2719 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2720 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2721 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2722 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2723 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2724 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2725 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2726 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2727 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2728 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2729 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2730 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2733 assert(CC < 16 && "Can only handle standard cond codes");
2734 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2736 default: llvm_unreachable("Illegal register size!");
2737 case 2: return Opc[Idx][0];
2738 case 4: return Opc[Idx][1];
2739 case 8: return Opc[Idx][2];
2743 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2744 if (!MI->isTerminator()) return false;
2746 // Conditional branch is a special case.
2747 if (MI->isBranch() && !MI->isBarrier())
2749 if (!MI->isPredicable())
2751 return !isPredicated(MI);
2754 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2755 MachineBasicBlock *&TBB,
2756 MachineBasicBlock *&FBB,
2757 SmallVectorImpl<MachineOperand> &Cond,
2758 bool AllowModify) const {
2759 // Start from the bottom of the block and work up, examining the
2760 // terminator instructions.
2761 MachineBasicBlock::iterator I = MBB.end();
2762 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2763 while (I != MBB.begin()) {
2765 if (I->isDebugValue())
2768 // Working from the bottom, when we see a non-terminator instruction, we're
2770 if (!isUnpredicatedTerminator(I))
2773 // A terminator that isn't a branch can't easily be handled by this
2778 // Handle unconditional branches.
2779 if (I->getOpcode() == X86::JMP_4) {
2783 TBB = I->getOperand(0).getMBB();
2787 // If the block has any instructions after a JMP, delete them.
2788 while (std::next(I) != MBB.end())
2789 std::next(I)->eraseFromParent();
2794 // Delete the JMP if it's equivalent to a fall-through.
2795 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2797 I->eraseFromParent();
2799 UnCondBrIter = MBB.end();
2803 // TBB is used to indicate the unconditional destination.
2804 TBB = I->getOperand(0).getMBB();
2808 // Handle conditional branches.
2809 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
2810 if (BranchCode == X86::COND_INVALID)
2811 return true; // Can't handle indirect branch.
2813 // Working from the bottom, handle the first conditional branch.
2815 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2816 if (AllowModify && UnCondBrIter != MBB.end() &&
2817 MBB.isLayoutSuccessor(TargetBB)) {
2818 // If we can modify the code and it ends in something like:
2826 // Then we can change this to:
2833 // Which is a bit more efficient.
2834 // We conditionally jump to the fall-through block.
2835 BranchCode = GetOppositeBranchCondition(BranchCode);
2836 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2837 MachineBasicBlock::iterator OldInst = I;
2839 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2840 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2841 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2844 OldInst->eraseFromParent();
2845 UnCondBrIter->eraseFromParent();
2847 // Restart the analysis.
2848 UnCondBrIter = MBB.end();
2854 TBB = I->getOperand(0).getMBB();
2855 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2859 // Handle subsequent conditional branches. Only handle the case where all
2860 // conditional branches branch to the same destination and their condition
2861 // opcodes fit one of the special multi-branch idioms.
2862 assert(Cond.size() == 1);
2865 // Only handle the case where all conditional branches branch to the same
2867 if (TBB != I->getOperand(0).getMBB())
2870 // If the conditions are the same, we can leave them alone.
2871 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2872 if (OldBranchCode == BranchCode)
2875 // If they differ, see if they fit one of the known patterns. Theoretically,
2876 // we could handle more patterns here, but we shouldn't expect to see them
2877 // if instruction selection has done a reasonable job.
2878 if ((OldBranchCode == X86::COND_NP &&
2879 BranchCode == X86::COND_E) ||
2880 (OldBranchCode == X86::COND_E &&
2881 BranchCode == X86::COND_NP))
2882 BranchCode = X86::COND_NP_OR_E;
2883 else if ((OldBranchCode == X86::COND_P &&
2884 BranchCode == X86::COND_NE) ||
2885 (OldBranchCode == X86::COND_NE &&
2886 BranchCode == X86::COND_P))
2887 BranchCode = X86::COND_NE_OR_P;
2891 // Update the MachineOperand.
2892 Cond[0].setImm(BranchCode);
2898 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2899 MachineBasicBlock::iterator I = MBB.end();
2902 while (I != MBB.begin()) {
2904 if (I->isDebugValue())
2906 if (I->getOpcode() != X86::JMP_4 &&
2907 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2909 // Remove the branch.
2910 I->eraseFromParent();
2919 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2920 MachineBasicBlock *FBB,
2921 const SmallVectorImpl<MachineOperand> &Cond,
2922 DebugLoc DL) const {
2923 // Shouldn't be a fall through.
2924 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2925 assert((Cond.size() == 1 || Cond.size() == 0) &&
2926 "X86 branch conditions have one component!");
2929 // Unconditional branch?
2930 assert(!FBB && "Unconditional branch with multiple successors!");
2931 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2935 // Conditional branch.
2937 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2939 case X86::COND_NP_OR_E:
2940 // Synthesize NP_OR_E with two branches.
2941 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2943 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2946 case X86::COND_NE_OR_P:
2947 // Synthesize NE_OR_P with two branches.
2948 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2950 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2954 unsigned Opc = GetCondBranchFromCond(CC);
2955 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2960 // Two-way Conditional branch. Insert the second branch.
2961 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2968 canInsertSelect(const MachineBasicBlock &MBB,
2969 const SmallVectorImpl<MachineOperand> &Cond,
2970 unsigned TrueReg, unsigned FalseReg,
2971 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2972 // Not all subtargets have cmov instructions.
2973 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2975 if (Cond.size() != 1)
2977 // We cannot do the composite conditions, at least not in SSA form.
2978 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2981 // Check register classes.
2982 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2983 const TargetRegisterClass *RC =
2984 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2988 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2989 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2990 X86::GR32RegClass.hasSubClassEq(RC) ||
2991 X86::GR64RegClass.hasSubClassEq(RC)) {
2992 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2993 // Bridge. Probably Ivy Bridge as well.
3000 // Can't do vectors.
3004 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3005 MachineBasicBlock::iterator I, DebugLoc DL,
3007 const SmallVectorImpl<MachineOperand> &Cond,
3008 unsigned TrueReg, unsigned FalseReg) const {
3009 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3010 assert(Cond.size() == 1 && "Invalid Cond array");
3011 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
3012 MRI.getRegClass(DstReg)->getSize(),
3013 false/*HasMemoryOperand*/);
3014 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
3017 /// isHReg - Test if the given register is a physical h register.
3018 static bool isHReg(unsigned Reg) {
3019 return X86::GR8_ABCD_HRegClass.contains(Reg);
3022 // Try and copy between VR128/VR64 and GR64 registers.
3023 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3024 const X86Subtarget& Subtarget) {
3027 // SrcReg(VR128) -> DestReg(GR64)
3028 // SrcReg(VR64) -> DestReg(GR64)
3029 // SrcReg(GR64) -> DestReg(VR128)
3030 // SrcReg(GR64) -> DestReg(VR64)
3032 bool HasAVX = Subtarget.hasAVX();
3033 bool HasAVX512 = Subtarget.hasAVX512();
3034 if (X86::GR64RegClass.contains(DestReg)) {
3035 if (X86::VR128XRegClass.contains(SrcReg))
3036 // Copy from a VR128 register to a GR64 register.
3037 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
3039 if (X86::VR64RegClass.contains(SrcReg))
3040 // Copy from a VR64 register to a GR64 register.
3041 return X86::MOVSDto64rr;
3042 } else if (X86::GR64RegClass.contains(SrcReg)) {
3043 // Copy from a GR64 register to a VR128 register.
3044 if (X86::VR128XRegClass.contains(DestReg))
3045 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
3047 // Copy from a GR64 register to a VR64 register.
3048 if (X86::VR64RegClass.contains(DestReg))
3049 return X86::MOV64toSDrr;
3052 // SrcReg(FR32) -> DestReg(GR32)
3053 // SrcReg(GR32) -> DestReg(FR32)
3055 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3056 // Copy from a FR32 register to a GR32 register.
3057 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3059 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3060 // Copy from a GR32 register to a FR32 register.
3061 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3065 inline static bool MaskRegClassContains(unsigned Reg) {
3066 return X86::VK8RegClass.contains(Reg) ||
3067 X86::VK16RegClass.contains(Reg) ||
3068 X86::VK1RegClass.contains(Reg);
3071 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3072 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3073 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3074 X86::VR512RegClass.contains(DestReg, SrcReg)) {
3075 DestReg = get512BitSuperRegister(DestReg);
3076 SrcReg = get512BitSuperRegister(SrcReg);
3077 return X86::VMOVAPSZrr;
3079 if (MaskRegClassContains(DestReg) &&
3080 MaskRegClassContains(SrcReg))
3081 return X86::KMOVWkk;
3082 if (MaskRegClassContains(DestReg) &&
3083 (X86::GR32RegClass.contains(SrcReg) ||
3084 X86::GR16RegClass.contains(SrcReg) ||
3085 X86::GR8RegClass.contains(SrcReg))) {
3086 SrcReg = getX86SubSuperRegister(SrcReg, MVT::i32);
3087 return X86::KMOVWkr;
3089 if ((X86::GR32RegClass.contains(DestReg) ||
3090 X86::GR16RegClass.contains(DestReg) ||
3091 X86::GR8RegClass.contains(DestReg)) &&
3092 MaskRegClassContains(SrcReg)) {
3093 DestReg = getX86SubSuperRegister(DestReg, MVT::i32);
3094 return X86::KMOVWrk;
3099 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3100 MachineBasicBlock::iterator MI, DebugLoc DL,
3101 unsigned DestReg, unsigned SrcReg,
3102 bool KillSrc) const {
3103 // First deal with the normal symmetric copies.
3104 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3105 bool HasAVX512 = TM.getSubtarget<X86Subtarget>().hasAVX512();
3107 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3109 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3111 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3113 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3114 // Copying to or from a physical H register on x86-64 requires a NOREX
3115 // move. Otherwise use a normal move.
3116 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3117 TM.getSubtarget<X86Subtarget>().is64Bit()) {
3118 Opc = X86::MOV8rr_NOREX;
3119 // Both operands must be encodable without an REX prefix.
3120 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3121 "8-bit H register can not be copied outside GR8_NOREX");
3125 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3126 Opc = X86::MMX_MOVQ64rr;
3128 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3129 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3130 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3131 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3132 Opc = X86::VMOVAPSYrr;
3134 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, TM.getSubtarget<X86Subtarget>());
3137 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3138 .addReg(SrcReg, getKillRegState(KillSrc));
3142 // Moving EFLAGS to / from another register requires a push and a pop.
3143 // Notice that we have to adjust the stack if we don't want to clobber the
3144 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
3145 if (SrcReg == X86::EFLAGS) {
3146 if (X86::GR64RegClass.contains(DestReg)) {
3147 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3148 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3151 if (X86::GR32RegClass.contains(DestReg)) {
3152 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3153 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3157 if (DestReg == X86::EFLAGS) {
3158 if (X86::GR64RegClass.contains(SrcReg)) {
3159 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3160 .addReg(SrcReg, getKillRegState(KillSrc));
3161 BuildMI(MBB, MI, DL, get(X86::POPF64));
3164 if (X86::GR32RegClass.contains(SrcReg)) {
3165 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3166 .addReg(SrcReg, getKillRegState(KillSrc));
3167 BuildMI(MBB, MI, DL, get(X86::POPF32));
3172 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3173 << " to " << RI.getName(DestReg) << '\n');
3174 llvm_unreachable("Cannot emit physreg copy instruction");
3177 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3178 const TargetRegisterClass *RC,
3179 bool isStackAligned,
3180 const TargetMachine &TM,
3182 if (TM.getSubtarget<X86Subtarget>().hasAVX512()) {
3183 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3184 X86::VK16RegClass.hasSubClassEq(RC))
3185 return load ? X86::KMOVWkm : X86::KMOVWmk;
3186 if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3187 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3188 if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3189 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3190 if (X86::VR512RegClass.hasSubClassEq(RC))
3191 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3194 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3195 switch (RC->getSize()) {
3197 llvm_unreachable("Unknown spill size");
3199 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3200 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3201 // Copying to or from a physical H register on x86-64 requires a NOREX
3202 // move. Otherwise use a normal move.
3203 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3204 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3205 return load ? X86::MOV8rm : X86::MOV8mr;
3207 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3208 return load ? X86::MOV16rm : X86::MOV16mr;
3210 if (X86::GR32RegClass.hasSubClassEq(RC))
3211 return load ? X86::MOV32rm : X86::MOV32mr;
3212 if (X86::FR32RegClass.hasSubClassEq(RC))
3214 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3215 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3216 if (X86::RFP32RegClass.hasSubClassEq(RC))
3217 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3218 llvm_unreachable("Unknown 4-byte regclass");
3220 if (X86::GR64RegClass.hasSubClassEq(RC))
3221 return load ? X86::MOV64rm : X86::MOV64mr;
3222 if (X86::FR64RegClass.hasSubClassEq(RC))
3224 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3225 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3226 if (X86::VR64RegClass.hasSubClassEq(RC))
3227 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3228 if (X86::RFP64RegClass.hasSubClassEq(RC))
3229 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3230 llvm_unreachable("Unknown 8-byte regclass");
3232 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3233 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3235 assert((X86::VR128RegClass.hasSubClassEq(RC) ||
3236 X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
3237 // If stack is realigned we can use aligned stores.
3240 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3241 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
3244 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3245 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3248 assert((X86::VR256RegClass.hasSubClassEq(RC) ||
3249 X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
3250 // If stack is realigned we can use aligned stores.
3252 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3254 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
3256 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3258 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3260 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3264 static unsigned getStoreRegOpcode(unsigned SrcReg,
3265 const TargetRegisterClass *RC,
3266 bool isStackAligned,
3267 TargetMachine &TM) {
3268 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
3272 static unsigned getLoadRegOpcode(unsigned DestReg,
3273 const TargetRegisterClass *RC,
3274 bool isStackAligned,
3275 const TargetMachine &TM) {
3276 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
3279 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3280 MachineBasicBlock::iterator MI,
3281 unsigned SrcReg, bool isKill, int FrameIdx,
3282 const TargetRegisterClass *RC,
3283 const TargetRegisterInfo *TRI) const {
3284 const MachineFunction &MF = *MBB.getParent();
3285 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3286 "Stack slot too small for store");
3287 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3288 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3289 RI.canRealignStack(MF);
3290 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3291 DebugLoc DL = MBB.findDebugLoc(MI);
3292 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
3293 .addReg(SrcReg, getKillRegState(isKill));
3296 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3298 SmallVectorImpl<MachineOperand> &Addr,
3299 const TargetRegisterClass *RC,
3300 MachineInstr::mmo_iterator MMOBegin,
3301 MachineInstr::mmo_iterator MMOEnd,
3302 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3303 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3304 bool isAligned = MMOBegin != MMOEnd &&
3305 (*MMOBegin)->getAlignment() >= Alignment;
3306 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3308 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3309 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3310 MIB.addOperand(Addr[i]);
3311 MIB.addReg(SrcReg, getKillRegState(isKill));
3312 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3313 NewMIs.push_back(MIB);
3317 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3318 MachineBasicBlock::iterator MI,
3319 unsigned DestReg, int FrameIdx,
3320 const TargetRegisterClass *RC,
3321 const TargetRegisterInfo *TRI) const {
3322 const MachineFunction &MF = *MBB.getParent();
3323 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3324 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3325 RI.canRealignStack(MF);
3326 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3327 DebugLoc DL = MBB.findDebugLoc(MI);
3328 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
3331 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
3332 SmallVectorImpl<MachineOperand> &Addr,
3333 const TargetRegisterClass *RC,
3334 MachineInstr::mmo_iterator MMOBegin,
3335 MachineInstr::mmo_iterator MMOEnd,
3336 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3337 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3338 bool isAligned = MMOBegin != MMOEnd &&
3339 (*MMOBegin)->getAlignment() >= Alignment;
3340 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3342 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3343 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3344 MIB.addOperand(Addr[i]);
3345 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3346 NewMIs.push_back(MIB);
3350 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3351 int &CmpMask, int &CmpValue) const {
3352 switch (MI->getOpcode()) {
3354 case X86::CMP64ri32:
3361 SrcReg = MI->getOperand(0).getReg();
3364 CmpValue = MI->getOperand(1).getImm();
3366 // A SUB can be used to perform comparison.
3371 SrcReg = MI->getOperand(1).getReg();
3380 SrcReg = MI->getOperand(1).getReg();
3381 SrcReg2 = MI->getOperand(2).getReg();
3385 case X86::SUB64ri32:
3392 SrcReg = MI->getOperand(1).getReg();
3395 CmpValue = MI->getOperand(2).getImm();
3401 SrcReg = MI->getOperand(0).getReg();
3402 SrcReg2 = MI->getOperand(1).getReg();
3410 SrcReg = MI->getOperand(0).getReg();
3411 if (MI->getOperand(1).getReg() != SrcReg) return false;
3412 // Compare against zero.
3421 /// isRedundantFlagInstr - check whether the first instruction, whose only
3422 /// purpose is to update flags, can be made redundant.
3423 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3424 /// This function can be extended later on.
3425 /// SrcReg, SrcRegs: register operands for FlagI.
3426 /// ImmValue: immediate for FlagI if it takes an immediate.
3427 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3428 unsigned SrcReg2, int ImmValue,
3430 if (((FlagI->getOpcode() == X86::CMP64rr &&
3431 OI->getOpcode() == X86::SUB64rr) ||
3432 (FlagI->getOpcode() == X86::CMP32rr &&
3433 OI->getOpcode() == X86::SUB32rr)||
3434 (FlagI->getOpcode() == X86::CMP16rr &&
3435 OI->getOpcode() == X86::SUB16rr)||
3436 (FlagI->getOpcode() == X86::CMP8rr &&
3437 OI->getOpcode() == X86::SUB8rr)) &&
3438 ((OI->getOperand(1).getReg() == SrcReg &&
3439 OI->getOperand(2).getReg() == SrcReg2) ||
3440 (OI->getOperand(1).getReg() == SrcReg2 &&
3441 OI->getOperand(2).getReg() == SrcReg)))
3444 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3445 OI->getOpcode() == X86::SUB64ri32) ||
3446 (FlagI->getOpcode() == X86::CMP64ri8 &&
3447 OI->getOpcode() == X86::SUB64ri8) ||
3448 (FlagI->getOpcode() == X86::CMP32ri &&
3449 OI->getOpcode() == X86::SUB32ri) ||
3450 (FlagI->getOpcode() == X86::CMP32ri8 &&
3451 OI->getOpcode() == X86::SUB32ri8) ||
3452 (FlagI->getOpcode() == X86::CMP16ri &&
3453 OI->getOpcode() == X86::SUB16ri) ||
3454 (FlagI->getOpcode() == X86::CMP16ri8 &&
3455 OI->getOpcode() == X86::SUB16ri8) ||
3456 (FlagI->getOpcode() == X86::CMP8ri &&
3457 OI->getOpcode() == X86::SUB8ri)) &&
3458 OI->getOperand(1).getReg() == SrcReg &&
3459 OI->getOperand(2).getImm() == ImmValue)
3464 /// isDefConvertible - check whether the definition can be converted
3465 /// to remove a comparison against zero.
3466 inline static bool isDefConvertible(MachineInstr *MI) {
3467 switch (MI->getOpcode()) {
3468 default: return false;
3470 // The shift instructions only modify ZF if their shift count is non-zero.
3471 // N.B.: The processor truncates the shift count depending on the encoding.
3472 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3473 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3474 return getTruncatedShiftCount(MI, 2) != 0;
3476 // Some left shift instructions can be turned into LEA instructions but only
3477 // if their flags aren't used. Avoid transforming such instructions.
3478 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3479 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3480 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3484 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3485 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3486 return getTruncatedShiftCount(MI, 3) != 0;
3488 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3489 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3490 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3491 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3492 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3493 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3494 case X86::DEC64_32r: case X86::DEC64_16r:
3495 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3496 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3497 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3498 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3499 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3500 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3501 case X86::INC64_32r: case X86::INC64_16r:
3502 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3503 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3504 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3505 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3506 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3507 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3508 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3509 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3510 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3511 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3512 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3513 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3514 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3515 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3516 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3517 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3518 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3519 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3520 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3521 case X86::ADC32ri: case X86::ADC32ri8:
3522 case X86::ADC32rr: case X86::ADC64ri32:
3523 case X86::ADC64ri8: case X86::ADC64rr:
3524 case X86::SBB32ri: case X86::SBB32ri8:
3525 case X86::SBB32rr: case X86::SBB64ri32:
3526 case X86::SBB64ri8: case X86::SBB64rr:
3527 case X86::ANDN32rr: case X86::ANDN32rm:
3528 case X86::ANDN64rr: case X86::ANDN64rm:
3529 case X86::BEXTR32rr: case X86::BEXTR64rr:
3530 case X86::BEXTR32rm: case X86::BEXTR64rm:
3531 case X86::BLSI32rr: case X86::BLSI32rm:
3532 case X86::BLSI64rr: case X86::BLSI64rm:
3533 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3534 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3535 case X86::BLSR32rr: case X86::BLSR32rm:
3536 case X86::BLSR64rr: case X86::BLSR64rm:
3537 case X86::BZHI32rr: case X86::BZHI32rm:
3538 case X86::BZHI64rr: case X86::BZHI64rm:
3539 case X86::LZCNT16rr: case X86::LZCNT16rm:
3540 case X86::LZCNT32rr: case X86::LZCNT32rm:
3541 case X86::LZCNT64rr: case X86::LZCNT64rm:
3542 case X86::POPCNT16rr:case X86::POPCNT16rm:
3543 case X86::POPCNT32rr:case X86::POPCNT32rm:
3544 case X86::POPCNT64rr:case X86::POPCNT64rm:
3545 case X86::TZCNT16rr: case X86::TZCNT16rm:
3546 case X86::TZCNT32rr: case X86::TZCNT32rm:
3547 case X86::TZCNT64rr: case X86::TZCNT64rm:
3552 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3553 /// operates on the same source operands and sets flags in the same way as
3554 /// Compare; remove Compare if possible.
3556 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3557 int CmpMask, int CmpValue,
3558 const MachineRegisterInfo *MRI) const {
3559 // Check whether we can replace SUB with CMP.
3560 unsigned NewOpcode = 0;
3561 switch (CmpInstr->getOpcode()) {
3563 case X86::SUB64ri32:
3578 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3580 // There is no use of the destination register, we can replace SUB with CMP.
3581 switch (CmpInstr->getOpcode()) {
3582 default: llvm_unreachable("Unreachable!");
3583 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3584 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3585 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3586 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3587 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3588 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3589 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3590 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3591 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3592 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3593 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3594 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3595 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3596 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3597 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3599 CmpInstr->setDesc(get(NewOpcode));
3600 CmpInstr->RemoveOperand(0);
3601 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3602 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3603 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3608 // Get the unique definition of SrcReg.
3609 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3610 if (!MI) return false;
3612 // CmpInstr is the first instruction of the BB.
3613 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3615 // If we are comparing against zero, check whether we can use MI to update
3616 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3617 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3618 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3619 !isDefConvertible(MI)))
3622 // We are searching for an earlier instruction that can make CmpInstr
3623 // redundant and that instruction will be saved in Sub.
3624 MachineInstr *Sub = NULL;
3625 const TargetRegisterInfo *TRI = &getRegisterInfo();
3627 // We iterate backward, starting from the instruction before CmpInstr and
3628 // stop when reaching the definition of a source register or done with the BB.
3629 // RI points to the instruction before CmpInstr.
3630 // If the definition is in this basic block, RE points to the definition;
3631 // otherwise, RE is the rend of the basic block.
3632 MachineBasicBlock::reverse_iterator
3633 RI = MachineBasicBlock::reverse_iterator(I),
3634 RE = CmpInstr->getParent() == MI->getParent() ?
3635 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3636 CmpInstr->getParent()->rend();
3637 MachineInstr *Movr0Inst = 0;
3638 for (; RI != RE; ++RI) {
3639 MachineInstr *Instr = &*RI;
3640 // Check whether CmpInstr can be made redundant by the current instruction.
3642 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3647 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3648 Instr->readsRegister(X86::EFLAGS, TRI)) {
3649 // This instruction modifies or uses EFLAGS.
3651 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3652 // They are safe to move up, if the definition to EFLAGS is dead and
3653 // earlier instructions do not read or write EFLAGS.
3654 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
3655 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3660 // We can't remove CmpInstr.
3665 // Return false if no candidates exist.
3666 if (!IsCmpZero && !Sub)
3669 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3670 Sub->getOperand(2).getReg() == SrcReg);
3672 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3673 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3674 // If we are done with the basic block, we need to check whether EFLAGS is
3676 bool IsSafe = false;
3677 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3678 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3679 for (++I; I != E; ++I) {
3680 const MachineInstr &Instr = *I;
3681 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3682 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3683 // We should check the usage if this instruction uses and updates EFLAGS.
3684 if (!UseEFLAGS && ModifyEFLAGS) {
3685 // It is safe to remove CmpInstr if EFLAGS is updated again.
3689 if (!UseEFLAGS && !ModifyEFLAGS)
3692 // EFLAGS is used by this instruction.
3693 X86::CondCode OldCC;
3694 bool OpcIsSET = false;
3695 if (IsCmpZero || IsSwapped) {
3696 // We decode the condition code from opcode.
3697 if (Instr.isBranch())
3698 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3700 OldCC = getCondFromSETOpc(Instr.getOpcode());
3701 if (OldCC != X86::COND_INVALID)
3704 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3706 if (OldCC == X86::COND_INVALID) return false;
3711 case X86::COND_A: case X86::COND_AE:
3712 case X86::COND_B: case X86::COND_BE:
3713 case X86::COND_G: case X86::COND_GE:
3714 case X86::COND_L: case X86::COND_LE:
3715 case X86::COND_O: case X86::COND_NO:
3716 // CF and OF are used, we can't perform this optimization.
3719 } else if (IsSwapped) {
3720 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3721 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3722 // We swap the condition code and synthesize the new opcode.
3723 X86::CondCode NewCC = getSwappedCondition(OldCC);
3724 if (NewCC == X86::COND_INVALID) return false;
3726 // Synthesize the new opcode.
3727 bool HasMemoryOperand = Instr.hasOneMemOperand();
3729 if (Instr.isBranch())
3730 NewOpc = GetCondBranchFromCond(NewCC);
3732 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3734 unsigned DstReg = Instr.getOperand(0).getReg();
3735 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3739 // Push the MachineInstr to OpsToUpdate.
3740 // If it is safe to remove CmpInstr, the condition code of these
3741 // instructions will be modified.
3742 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3744 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3745 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3751 // If EFLAGS is not killed nor re-defined, we should check whether it is
3752 // live-out. If it is live-out, do not optimize.
3753 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3754 MachineBasicBlock *MBB = CmpInstr->getParent();
3755 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3756 SE = MBB->succ_end(); SI != SE; ++SI)
3757 if ((*SI)->isLiveIn(X86::EFLAGS))
3761 // The instruction to be updated is either Sub or MI.
3762 Sub = IsCmpZero ? MI : Sub;
3763 // Move Movr0Inst to the appropriate place before Sub.
3765 // Look backwards until we find a def that doesn't use the current EFLAGS.
3767 MachineBasicBlock::reverse_iterator
3768 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3769 InsertE = Sub->getParent()->rend();
3770 for (; InsertI != InsertE; ++InsertI) {
3771 MachineInstr *Instr = &*InsertI;
3772 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3773 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3774 Sub->getParent()->remove(Movr0Inst);
3775 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3780 if (InsertI == InsertE)
3784 // Make sure Sub instruction defines EFLAGS and mark the def live.
3785 unsigned i = 0, e = Sub->getNumOperands();
3786 for (; i != e; ++i) {
3787 MachineOperand &MO = Sub->getOperand(i);
3788 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3789 MO.setIsDead(false);
3793 assert(i != e && "Unable to locate a def EFLAGS operand");
3795 CmpInstr->eraseFromParent();
3797 // Modify the condition code of instructions in OpsToUpdate.
3798 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3799 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3803 /// optimizeLoadInstr - Try to remove the load by folding it to a register
3804 /// operand at the use. We fold the load instructions if load defines a virtual
3805 /// register, the virtual register is used once in the same BB, and the
3806 /// instructions in-between do not load or store, and have no side effects.
3807 MachineInstr* X86InstrInfo::
3808 optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3809 unsigned &FoldAsLoadDefReg,
3810 MachineInstr *&DefMI) const {
3811 if (FoldAsLoadDefReg == 0)
3813 // To be conservative, if there exists another load, clear the load candidate.
3814 if (MI->mayLoad()) {
3815 FoldAsLoadDefReg = 0;
3819 // Check whether we can move DefMI here.
3820 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3822 bool SawStore = false;
3823 if (!DefMI->isSafeToMove(this, 0, SawStore))
3826 // We try to commute MI if possible.
3827 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3828 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3829 // Collect information about virtual register operands of MI.
3830 unsigned SrcOperandId = 0;
3831 bool FoundSrcOperand = false;
3832 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3833 MachineOperand &MO = MI->getOperand(i);
3836 unsigned Reg = MO.getReg();
3837 if (Reg != FoldAsLoadDefReg)
3839 // Do not fold if we have a subreg use or a def or multiple uses.
3840 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3844 FoundSrcOperand = true;
3846 if (!FoundSrcOperand) return 0;
3848 // Check whether we can fold the def into SrcOperandId.
3849 SmallVector<unsigned, 8> Ops;
3850 Ops.push_back(SrcOperandId);
3851 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3853 FoldAsLoadDefReg = 0;
3858 // MI was changed but it didn't help, commute it back!
3859 commuteInstruction(MI, false);
3863 // Check whether we can commute MI and enable folding.
3864 if (MI->isCommutable()) {
3865 MachineInstr *NewMI = commuteInstruction(MI, false);
3866 // Unable to commute.
3867 if (!NewMI) return 0;
3869 // New instruction. It doesn't need to be kept.
3870 NewMI->eraseFromParent();
3878 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3879 /// instruction with two undef reads of the register being defined. This is
3880 /// used for mapping:
3883 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3885 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3886 const MCInstrDesc &Desc) {
3887 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3888 unsigned Reg = MIB->getOperand(0).getReg();
3891 // MachineInstr::addOperand() will insert explicit operands before any
3892 // implicit operands.
3893 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3894 // But we don't trust that.
3895 assert(MIB->getOperand(1).getReg() == Reg &&
3896 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3900 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3901 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3902 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3903 switch (MI->getOpcode()) {
3905 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
3907 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
3908 case X86::SETB_C16r:
3909 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
3910 case X86::SETB_C32r:
3911 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
3912 case X86::SETB_C64r:
3913 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
3917 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
3919 assert(HasAVX && "AVX not supported");
3920 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
3921 case X86::AVX512_512_SET0:
3922 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
3923 case X86::V_SETALLONES:
3924 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3925 case X86::AVX2_SETALLONES:
3926 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
3927 case X86::TEST8ri_NOREX:
3928 MI->setDesc(get(X86::TEST8ri));
3931 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
3933 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
3938 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
3939 const SmallVectorImpl<MachineOperand> &MOs,
3941 const TargetInstrInfo &TII) {
3942 // Create the base instruction with the memory operand as the first part.
3943 // Omit the implicit operands, something BuildMI can't do.
3944 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3945 MI->getDebugLoc(), true);
3946 MachineInstrBuilder MIB(MF, NewMI);
3947 unsigned NumAddrOps = MOs.size();
3948 for (unsigned i = 0; i != NumAddrOps; ++i)
3949 MIB.addOperand(MOs[i]);
3950 if (NumAddrOps < 4) // FrameIndex only
3953 // Loop over the rest of the ri operands, converting them over.
3954 unsigned NumOps = MI->getDesc().getNumOperands()-2;
3955 for (unsigned i = 0; i != NumOps; ++i) {
3956 MachineOperand &MO = MI->getOperand(i+2);
3959 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3960 MachineOperand &MO = MI->getOperand(i);
3966 static MachineInstr *FuseInst(MachineFunction &MF,
3967 unsigned Opcode, unsigned OpNo,
3968 const SmallVectorImpl<MachineOperand> &MOs,
3969 MachineInstr *MI, const TargetInstrInfo &TII) {
3970 // Omit the implicit operands, something BuildMI can't do.
3971 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3972 MI->getDebugLoc(), true);
3973 MachineInstrBuilder MIB(MF, NewMI);
3975 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3976 MachineOperand &MO = MI->getOperand(i);
3978 assert(MO.isReg() && "Expected to fold into reg operand!");
3979 unsigned NumAddrOps = MOs.size();
3980 for (unsigned i = 0; i != NumAddrOps; ++i)
3981 MIB.addOperand(MOs[i]);
3982 if (NumAddrOps < 4) // FrameIndex only
3991 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
3992 const SmallVectorImpl<MachineOperand> &MOs,
3994 MachineFunction &MF = *MI->getParent()->getParent();
3995 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
3997 unsigned NumAddrOps = MOs.size();
3998 for (unsigned i = 0; i != NumAddrOps; ++i)
3999 MIB.addOperand(MOs[i]);
4000 if (NumAddrOps < 4) // FrameIndex only
4002 return MIB.addImm(0);
4006 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4007 MachineInstr *MI, unsigned i,
4008 const SmallVectorImpl<MachineOperand> &MOs,
4009 unsigned Size, unsigned Align) const {
4010 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
4011 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
4012 bool isTwoAddrFold = false;
4014 // Atom favors register form of call. So, we do not fold loads into calls
4015 // when X86Subtarget is Atom.
4016 if (isCallRegIndirect &&
4017 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
4021 unsigned NumOps = MI->getDesc().getNumOperands();
4022 bool isTwoAddr = NumOps > 1 &&
4023 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4025 // FIXME: AsmPrinter doesn't know how to handle
4026 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4027 if (MI->getOpcode() == X86::ADD32ri &&
4028 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4031 MachineInstr *NewMI = NULL;
4032 // Folding a memory location into the two-address part of a two-address
4033 // instruction is different than folding it other places. It requires
4034 // replacing the *two* registers with the memory location.
4035 if (isTwoAddr && NumOps >= 2 && i < 2 &&
4036 MI->getOperand(0).isReg() &&
4037 MI->getOperand(1).isReg() &&
4038 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
4039 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4040 isTwoAddrFold = true;
4041 } else if (i == 0) { // If operand 0
4042 if (MI->getOpcode() == X86::MOV32r0) {
4043 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
4048 OpcodeTablePtr = &RegOp2MemOpTable0;
4049 } else if (i == 1) {
4050 OpcodeTablePtr = &RegOp2MemOpTable1;
4051 } else if (i == 2) {
4052 OpcodeTablePtr = &RegOp2MemOpTable2;
4053 } else if (i == 3) {
4054 OpcodeTablePtr = &RegOp2MemOpTable3;
4057 // If table selected...
4058 if (OpcodeTablePtr) {
4059 // Find the Opcode to fuse
4060 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4061 OpcodeTablePtr->find(MI->getOpcode());
4062 if (I != OpcodeTablePtr->end()) {
4063 unsigned Opcode = I->second.first;
4064 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4065 if (Align < MinAlign)
4067 bool NarrowToMOV32rm = false;
4069 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
4070 if (Size < RCSize) {
4071 // Check if it's safe to fold the load. If the size of the object is
4072 // narrower than the load width, then it's not.
4073 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4075 // If this is a 64-bit load, but the spill slot is 32, then we can do
4076 // a 32-bit load which is implicitly zero-extended. This likely is due
4077 // to liveintervalanalysis remat'ing a load from stack slot.
4078 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4080 Opcode = X86::MOV32rm;
4081 NarrowToMOV32rm = true;
4086 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
4088 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
4090 if (NarrowToMOV32rm) {
4091 // If this is the special case where we use a MOV32rm to load a 32-bit
4092 // value and zero-extend the top bits. Change the destination register
4094 unsigned DstReg = NewMI->getOperand(0).getReg();
4095 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4096 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
4099 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4106 if (PrintFailedFusing && !MI->isCopy())
4107 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
4111 /// hasPartialRegUpdate - Return true for all instructions that only update
4112 /// the first 32 or 64-bits of the destination register and leave the rest
4113 /// unmodified. This can be used to avoid folding loads if the instructions
4114 /// only update part of the destination register, and the non-updated part is
4115 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4116 /// instructions breaks the partial register dependency and it can improve
4117 /// performance. e.g.:
4119 /// movss (%rdi), %xmm0
4120 /// cvtss2sd %xmm0, %xmm0
4123 /// cvtss2sd (%rdi), %xmm0
4125 /// FIXME: This should be turned into a TSFlags.
4127 static bool hasPartialRegUpdate(unsigned Opcode) {
4129 case X86::CVTSI2SSrr:
4130 case X86::CVTSI2SS64rr:
4131 case X86::CVTSI2SDrr:
4132 case X86::CVTSI2SD64rr:
4133 case X86::CVTSD2SSrr:
4134 case X86::Int_CVTSD2SSrr:
4135 case X86::CVTSS2SDrr:
4136 case X86::Int_CVTSS2SDrr:
4138 case X86::RCPSSr_Int:
4140 case X86::ROUNDSDr_Int:
4142 case X86::ROUNDSSr_Int:
4144 case X86::RSQRTSSr_Int:
4146 case X86::SQRTSSr_Int:
4153 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4154 /// instructions we would like before a partial register update.
4155 unsigned X86InstrInfo::
4156 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4157 const TargetRegisterInfo *TRI) const {
4158 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4161 // If MI is marked as reading Reg, the partial register update is wanted.
4162 const MachineOperand &MO = MI->getOperand(0);
4163 unsigned Reg = MO.getReg();
4164 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4165 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4168 if (MI->readsRegister(Reg, TRI))
4172 // If any of the preceding 16 instructions are reading Reg, insert a
4173 // dependency breaking instruction. The magic number is based on a few
4174 // Nehalem experiments.
4178 // Return true for any instruction the copies the high bits of the first source
4179 // operand into the unused high bits of the destination operand.
4180 static bool hasUndefRegUpdate(unsigned Opcode) {
4182 case X86::VCVTSI2SSrr:
4183 case X86::Int_VCVTSI2SSrr:
4184 case X86::VCVTSI2SS64rr:
4185 case X86::Int_VCVTSI2SS64rr:
4186 case X86::VCVTSI2SDrr:
4187 case X86::Int_VCVTSI2SDrr:
4188 case X86::VCVTSI2SD64rr:
4189 case X86::Int_VCVTSI2SD64rr:
4190 case X86::VCVTSD2SSrr:
4191 case X86::Int_VCVTSD2SSrr:
4192 case X86::VCVTSS2SDrr:
4193 case X86::Int_VCVTSS2SDrr:
4195 case X86::VROUNDSDr:
4196 case X86::VROUNDSDr_Int:
4197 case X86::VROUNDSSr:
4198 case X86::VROUNDSSr_Int:
4199 case X86::VRSQRTSSr:
4203 case X86::VCVTSD2SSZrr:
4204 case X86::VCVTSS2SDZrr:
4211 /// Inform the ExeDepsFix pass how many idle instructions we would like before
4212 /// certain undef register reads.
4214 /// This catches the VCVTSI2SD family of instructions:
4216 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
4218 /// We should to be careful *not* to catch VXOR idioms which are presumably
4219 /// handled specially in the pipeline:
4221 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
4223 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4224 /// high bits that are passed-through are not live.
4225 unsigned X86InstrInfo::
4226 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
4227 const TargetRegisterInfo *TRI) const {
4228 if (!hasUndefRegUpdate(MI->getOpcode()))
4231 // Set the OpNum parameter to the first source operand.
4234 const MachineOperand &MO = MI->getOperand(OpNum);
4235 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4236 // Use the same magic number as getPartialRegUpdateClearance.
4243 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4244 const TargetRegisterInfo *TRI) const {
4245 unsigned Reg = MI->getOperand(OpNum).getReg();
4246 // If MI kills this register, the false dependence is already broken.
4247 if (MI->killsRegister(Reg, TRI))
4249 if (X86::VR128RegClass.contains(Reg)) {
4250 // These instructions are all floating point domain, so xorps is the best
4252 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
4253 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4254 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4255 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4256 } else if (X86::VR256RegClass.contains(Reg)) {
4257 // Use vxorps to clear the full ymm register.
4258 // It wants to read and write the xmm sub-register.
4259 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4260 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4261 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4262 .addReg(Reg, RegState::ImplicitDefine);
4265 MI->addRegisterKilled(Reg, TRI, true);
4269 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
4270 const SmallVectorImpl<unsigned> &Ops,
4271 int FrameIndex) const {
4272 // Check switch flag
4273 if (NoFusing) return NULL;
4275 // Unless optimizing for size, don't fold to avoid partial
4276 // register update stalls
4277 if (!MF.getFunction()->getAttributes().
4278 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4279 hasPartialRegUpdate(MI->getOpcode()))
4282 const MachineFrameInfo *MFI = MF.getFrameInfo();
4283 unsigned Size = MFI->getObjectSize(FrameIndex);
4284 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
4285 // If the function stack isn't realigned we don't want to fold instructions
4286 // that need increased alignment.
4287 if (!RI.needsStackRealignment(MF))
4288 Alignment = std::min(Alignment, TM.getFrameLowering()->getStackAlignment());
4289 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4290 unsigned NewOpc = 0;
4291 unsigned RCSize = 0;
4292 switch (MI->getOpcode()) {
4293 default: return NULL;
4294 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
4295 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4296 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4297 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4299 // Check if it's safe to fold the load. If the size of the object is
4300 // narrower than the load width, then it's not.
4303 // Change to CMPXXri r, 0 first.
4304 MI->setDesc(get(NewOpc));
4305 MI->getOperand(1).ChangeToImmediate(0);
4306 } else if (Ops.size() != 1)
4309 SmallVector<MachineOperand,4> MOs;
4310 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
4311 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
4314 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4316 const SmallVectorImpl<unsigned> &Ops,
4317 MachineInstr *LoadMI) const {
4318 // If loading from a FrameIndex, fold directly from the FrameIndex.
4319 unsigned NumOps = LoadMI->getDesc().getNumOperands();
4321 if (isLoadFromStackSlot(LoadMI, FrameIndex))
4322 return foldMemoryOperandImpl(MF, MI, Ops, FrameIndex);
4324 // Check switch flag
4325 if (NoFusing) return NULL;
4327 // Unless optimizing for size, don't fold to avoid partial
4328 // register update stalls
4329 if (!MF.getFunction()->getAttributes().
4330 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4331 hasPartialRegUpdate(MI->getOpcode()))
4334 // Determine the alignment of the load.
4335 unsigned Alignment = 0;
4336 if (LoadMI->hasOneMemOperand())
4337 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
4339 switch (LoadMI->getOpcode()) {
4340 case X86::AVX2_SETALLONES:
4345 case X86::V_SETALLONES:
4357 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4358 unsigned NewOpc = 0;
4359 switch (MI->getOpcode()) {
4360 default: return NULL;
4361 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
4362 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4363 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4364 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
4366 // Change to CMPXXri r, 0 first.
4367 MI->setDesc(get(NewOpc));
4368 MI->getOperand(1).ChangeToImmediate(0);
4369 } else if (Ops.size() != 1)
4372 // Make sure the subregisters match.
4373 // Otherwise we risk changing the size of the load.
4374 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4377 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
4378 switch (LoadMI->getOpcode()) {
4380 case X86::V_SETALLONES:
4381 case X86::AVX2_SETALLONES:
4384 case X86::FsFLD0SS: {
4385 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
4386 // Create a constant-pool entry and operands to load from it.
4388 // Medium and large mode can't fold loads this way.
4389 if (TM.getCodeModel() != CodeModel::Small &&
4390 TM.getCodeModel() != CodeModel::Kernel)
4393 // x86-32 PIC requires a PIC base register for constant pools.
4394 unsigned PICBase = 0;
4395 if (TM.getRelocationModel() == Reloc::PIC_) {
4396 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4399 // FIXME: PICBase = getGlobalBaseReg(&MF);
4400 // This doesn't work for several reasons.
4401 // 1. GlobalBaseReg may have been spilled.
4402 // 2. It may not be live at MI.
4406 // Create a constant-pool entry.
4407 MachineConstantPool &MCP = *MF.getConstantPool();
4409 unsigned Opc = LoadMI->getOpcode();
4410 if (Opc == X86::FsFLD0SS)
4411 Ty = Type::getFloatTy(MF.getFunction()->getContext());
4412 else if (Opc == X86::FsFLD0SD)
4413 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
4414 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
4415 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
4417 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
4419 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
4420 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4421 Constant::getNullValue(Ty);
4422 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
4424 // Create operands to load from the constant pool entry.
4425 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4426 MOs.push_back(MachineOperand::CreateImm(1));
4427 MOs.push_back(MachineOperand::CreateReg(0, false));
4428 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
4429 MOs.push_back(MachineOperand::CreateReg(0, false));
4433 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4434 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4435 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4437 // These instructions only load 32 bits, we can't fold them if the
4438 // destination register is wider than 32 bits (4 bytes).
4440 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4441 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4442 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4444 // These instructions only load 64 bits, we can't fold them if the
4445 // destination register is wider than 64 bits (8 bytes).
4448 // Folding a normal load. Just copy the load's address operands.
4449 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
4450 MOs.push_back(LoadMI->getOperand(i));
4454 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
4458 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4459 const SmallVectorImpl<unsigned> &Ops) const {
4460 // Check switch flag
4461 if (NoFusing) return 0;
4463 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4464 switch (MI->getOpcode()) {
4465 default: return false;
4472 // FIXME: AsmPrinter doesn't know how to handle
4473 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4474 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4480 if (Ops.size() != 1)
4483 unsigned OpNum = Ops[0];
4484 unsigned Opc = MI->getOpcode();
4485 unsigned NumOps = MI->getDesc().getNumOperands();
4486 bool isTwoAddr = NumOps > 1 &&
4487 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4489 // Folding a memory location into the two-address part of a two-address
4490 // instruction is different than folding it other places. It requires
4491 // replacing the *two* registers with the memory location.
4492 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
4493 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4494 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4495 } else if (OpNum == 0) { // If operand 0
4496 if (Opc == X86::MOV32r0)
4499 OpcodeTablePtr = &RegOp2MemOpTable0;
4500 } else if (OpNum == 1) {
4501 OpcodeTablePtr = &RegOp2MemOpTable1;
4502 } else if (OpNum == 2) {
4503 OpcodeTablePtr = &RegOp2MemOpTable2;
4504 } else if (OpNum == 3) {
4505 OpcodeTablePtr = &RegOp2MemOpTable3;
4508 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4510 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
4513 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4514 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4515 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4516 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4517 MemOp2RegOpTable.find(MI->getOpcode());
4518 if (I == MemOp2RegOpTable.end())
4520 unsigned Opc = I->second.first;
4521 unsigned Index = I->second.second & TB_INDEX_MASK;
4522 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4523 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4524 if (UnfoldLoad && !FoldedLoad)
4526 UnfoldLoad &= FoldedLoad;
4527 if (UnfoldStore && !FoldedStore)
4529 UnfoldStore &= FoldedStore;
4531 const MCInstrDesc &MCID = get(Opc);
4532 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4533 if (!MI->hasOneMemOperand() &&
4534 RC == &X86::VR128RegClass &&
4535 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4536 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4537 // conservatively assume the address is unaligned. That's bad for
4540 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4541 SmallVector<MachineOperand,2> BeforeOps;
4542 SmallVector<MachineOperand,2> AfterOps;
4543 SmallVector<MachineOperand,4> ImpOps;
4544 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4545 MachineOperand &Op = MI->getOperand(i);
4546 if (i >= Index && i < Index + X86::AddrNumOperands)
4547 AddrOps.push_back(Op);
4548 else if (Op.isReg() && Op.isImplicit())
4549 ImpOps.push_back(Op);
4551 BeforeOps.push_back(Op);
4553 AfterOps.push_back(Op);
4556 // Emit the load instruction.
4558 std::pair<MachineInstr::mmo_iterator,
4559 MachineInstr::mmo_iterator> MMOs =
4560 MF.extractLoadMemRefs(MI->memoperands_begin(),
4561 MI->memoperands_end());
4562 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4564 // Address operands cannot be marked isKill.
4565 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4566 MachineOperand &MO = NewMIs[0]->getOperand(i);
4568 MO.setIsKill(false);
4573 // Emit the data processing instruction.
4574 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4575 MachineInstrBuilder MIB(MF, DataMI);
4578 MIB.addReg(Reg, RegState::Define);
4579 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4580 MIB.addOperand(BeforeOps[i]);
4583 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4584 MIB.addOperand(AfterOps[i]);
4585 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4586 MachineOperand &MO = ImpOps[i];
4587 MIB.addReg(MO.getReg(),
4588 getDefRegState(MO.isDef()) |
4589 RegState::Implicit |
4590 getKillRegState(MO.isKill()) |
4591 getDeadRegState(MO.isDead()) |
4592 getUndefRegState(MO.isUndef()));
4594 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4595 switch (DataMI->getOpcode()) {
4597 case X86::CMP64ri32:
4604 MachineOperand &MO0 = DataMI->getOperand(0);
4605 MachineOperand &MO1 = DataMI->getOperand(1);
4606 if (MO1.getImm() == 0) {
4608 switch (DataMI->getOpcode()) {
4609 default: llvm_unreachable("Unreachable!");
4611 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
4613 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
4615 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4616 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4618 DataMI->setDesc(get(NewOpc));
4619 MO1.ChangeToRegister(MO0.getReg(), false);
4623 NewMIs.push_back(DataMI);
4625 // Emit the store instruction.
4627 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
4628 std::pair<MachineInstr::mmo_iterator,
4629 MachineInstr::mmo_iterator> MMOs =
4630 MF.extractStoreMemRefs(MI->memoperands_begin(),
4631 MI->memoperands_end());
4632 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
4639 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
4640 SmallVectorImpl<SDNode*> &NewNodes) const {
4641 if (!N->isMachineOpcode())
4644 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4645 MemOp2RegOpTable.find(N->getMachineOpcode());
4646 if (I == MemOp2RegOpTable.end())
4648 unsigned Opc = I->second.first;
4649 unsigned Index = I->second.second & TB_INDEX_MASK;
4650 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4651 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4652 const MCInstrDesc &MCID = get(Opc);
4653 MachineFunction &MF = DAG.getMachineFunction();
4654 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4655 unsigned NumDefs = MCID.NumDefs;
4656 std::vector<SDValue> AddrOps;
4657 std::vector<SDValue> BeforeOps;
4658 std::vector<SDValue> AfterOps;
4660 unsigned NumOps = N->getNumOperands();
4661 for (unsigned i = 0; i != NumOps-1; ++i) {
4662 SDValue Op = N->getOperand(i);
4663 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
4664 AddrOps.push_back(Op);
4665 else if (i < Index-NumDefs)
4666 BeforeOps.push_back(Op);
4667 else if (i > Index-NumDefs)
4668 AfterOps.push_back(Op);
4670 SDValue Chain = N->getOperand(NumOps-1);
4671 AddrOps.push_back(Chain);
4673 // Emit the load instruction.
4676 EVT VT = *RC->vt_begin();
4677 std::pair<MachineInstr::mmo_iterator,
4678 MachineInstr::mmo_iterator> MMOs =
4679 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4680 cast<MachineSDNode>(N)->memoperands_end());
4681 if (!(*MMOs.first) &&
4682 RC == &X86::VR128RegClass &&
4683 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4684 // Do not introduce a slow unaligned load.
4686 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4687 bool isAligned = (*MMOs.first) &&
4688 (*MMOs.first)->getAlignment() >= Alignment;
4689 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4690 VT, MVT::Other, AddrOps);
4691 NewNodes.push_back(Load);
4693 // Preserve memory reference information.
4694 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4697 // Emit the data processing instruction.
4698 std::vector<EVT> VTs;
4699 const TargetRegisterClass *DstRC = 0;
4700 if (MCID.getNumDefs() > 0) {
4701 DstRC = getRegClass(MCID, 0, &RI, MF);
4702 VTs.push_back(*DstRC->vt_begin());
4704 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
4705 EVT VT = N->getValueType(i);
4706 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
4710 BeforeOps.push_back(SDValue(Load, 0));
4711 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
4712 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
4713 NewNodes.push_back(NewNode);
4715 // Emit the store instruction.
4718 AddrOps.push_back(SDValue(NewNode, 0));
4719 AddrOps.push_back(Chain);
4720 std::pair<MachineInstr::mmo_iterator,
4721 MachineInstr::mmo_iterator> MMOs =
4722 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4723 cast<MachineSDNode>(N)->memoperands_end());
4724 if (!(*MMOs.first) &&
4725 RC == &X86::VR128RegClass &&
4726 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4727 // Do not introduce a slow unaligned store.
4729 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4730 bool isAligned = (*MMOs.first) &&
4731 (*MMOs.first)->getAlignment() >= Alignment;
4732 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4734 dl, MVT::Other, AddrOps);
4735 NewNodes.push_back(Store);
4737 // Preserve memory reference information.
4738 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4744 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
4745 bool UnfoldLoad, bool UnfoldStore,
4746 unsigned *LoadRegIndex) const {
4747 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4748 MemOp2RegOpTable.find(Opc);
4749 if (I == MemOp2RegOpTable.end())
4751 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4752 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4753 if (UnfoldLoad && !FoldedLoad)
4755 if (UnfoldStore && !FoldedStore)
4758 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
4759 return I->second.first;
4763 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4764 int64_t &Offset1, int64_t &Offset2) const {
4765 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4767 unsigned Opc1 = Load1->getMachineOpcode();
4768 unsigned Opc2 = Load2->getMachineOpcode();
4770 default: return false;
4780 case X86::MMX_MOVD64rm:
4781 case X86::MMX_MOVQ64rm:
4782 case X86::FsMOVAPSrm:
4783 case X86::FsMOVAPDrm:
4789 // AVX load instructions
4792 case X86::FsVMOVAPSrm:
4793 case X86::FsVMOVAPDrm:
4794 case X86::VMOVAPSrm:
4795 case X86::VMOVUPSrm:
4796 case X86::VMOVAPDrm:
4797 case X86::VMOVDQArm:
4798 case X86::VMOVDQUrm:
4799 case X86::VMOVAPSYrm:
4800 case X86::VMOVUPSYrm:
4801 case X86::VMOVAPDYrm:
4802 case X86::VMOVDQAYrm:
4803 case X86::VMOVDQUYrm:
4807 default: return false;
4817 case X86::MMX_MOVD64rm:
4818 case X86::MMX_MOVQ64rm:
4819 case X86::FsMOVAPSrm:
4820 case X86::FsMOVAPDrm:
4826 // AVX load instructions
4829 case X86::FsVMOVAPSrm:
4830 case X86::FsVMOVAPDrm:
4831 case X86::VMOVAPSrm:
4832 case X86::VMOVUPSrm:
4833 case X86::VMOVAPDrm:
4834 case X86::VMOVDQArm:
4835 case X86::VMOVDQUrm:
4836 case X86::VMOVAPSYrm:
4837 case X86::VMOVUPSYrm:
4838 case X86::VMOVAPDYrm:
4839 case X86::VMOVDQAYrm:
4840 case X86::VMOVDQUYrm:
4844 // Check if chain operands and base addresses match.
4845 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4846 Load1->getOperand(5) != Load2->getOperand(5))
4848 // Segment operands should match as well.
4849 if (Load1->getOperand(4) != Load2->getOperand(4))
4851 // Scale should be 1, Index should be Reg0.
4852 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4853 Load1->getOperand(2) == Load2->getOperand(2)) {
4854 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4857 // Now let's examine the displacements.
4858 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4859 isa<ConstantSDNode>(Load2->getOperand(3))) {
4860 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4861 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4868 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4869 int64_t Offset1, int64_t Offset2,
4870 unsigned NumLoads) const {
4871 assert(Offset2 > Offset1);
4872 if ((Offset2 - Offset1) / 8 > 64)
4875 unsigned Opc1 = Load1->getMachineOpcode();
4876 unsigned Opc2 = Load2->getMachineOpcode();
4878 return false; // FIXME: overly conservative?
4885 case X86::MMX_MOVD64rm:
4886 case X86::MMX_MOVQ64rm:
4890 EVT VT = Load1->getValueType(0);
4891 switch (VT.getSimpleVT().SimpleTy) {
4893 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4894 // have 16 of them to play with.
4895 if (TM.getSubtargetImpl()->is64Bit()) {
4898 } else if (NumLoads) {
4916 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
4917 MachineInstr *Second) const {
4918 // Check if this processor supports macro-fusion. Since this is a minor
4919 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
4920 // proxy for SandyBridge+.
4921 if (!TM.getSubtarget<X86Subtarget>().hasAVX())
4930 switch(Second->getOpcode()) {
4953 FuseKind = FuseTest;
4956 switch (First->getOpcode()) {
4966 case X86::TEST32i32:
4967 case X86::TEST64i32:
4968 case X86::TEST64ri32:
4984 case X86::AND64ri32:
5004 case X86::CMP64ri32:
5015 case X86::ADD16ri8_DB:
5016 case X86::ADD16ri_DB:
5019 case X86::ADD16rr_DB:
5023 case X86::ADD32ri8_DB:
5024 case X86::ADD32ri_DB:
5027 case X86::ADD32rr_DB:
5029 case X86::ADD64ri32:
5030 case X86::ADD64ri32_DB:
5032 case X86::ADD64ri8_DB:
5035 case X86::ADD64rr_DB:
5053 case X86::SUB64ri32:
5061 return FuseKind == FuseCmp || FuseKind == FuseInc;
5064 case X86::INC64_16r:
5065 case X86::INC64_32r:
5070 case X86::DEC64_16r:
5071 case X86::DEC64_32r:
5074 return FuseKind == FuseInc;
5079 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5080 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5081 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5082 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
5084 Cond[0].setImm(GetOppositeBranchCondition(CC));
5089 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5090 // FIXME: Return false for x87 stack register classes for now. We can't
5091 // allow any loads of these registers before FpGet_ST0_80.
5092 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
5093 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
5096 /// getGlobalBaseReg - Return a virtual register initialized with the
5097 /// the global base register value. Output instructions required to
5098 /// initialize the register in the function entry block, if necessary.
5100 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5102 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5103 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
5104 "X86-64 PIC uses RIP relative addressing");
5106 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5107 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5108 if (GlobalBaseReg != 0)
5109 return GlobalBaseReg;
5111 // Create the register. The code to initialize it is inserted
5112 // later, by the CGBR pass (below).
5113 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5114 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
5115 X86FI->setGlobalBaseReg(GlobalBaseReg);
5116 return GlobalBaseReg;
5119 // These are the replaceable SSE instructions. Some of these have Int variants
5120 // that we don't include here. We don't want to replace instructions selected
5122 static const uint16_t ReplaceableInstrs[][3] = {
5123 //PackedSingle PackedDouble PackedInt
5124 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
5125 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
5126 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
5127 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
5128 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
5129 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
5130 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
5131 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
5132 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
5133 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
5134 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
5135 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
5136 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
5137 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
5138 // AVX 128-bit support
5139 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
5140 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
5141 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
5142 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5143 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5144 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5145 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5146 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5147 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5148 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5149 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5150 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
5151 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
5152 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
5153 // AVX 256-bit support
5154 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
5155 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
5156 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
5157 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
5158 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
5159 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
5162 static const uint16_t ReplaceableInstrsAVX2[][3] = {
5163 //PackedSingle PackedDouble PackedInt
5164 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
5165 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
5166 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
5167 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
5168 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
5169 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
5170 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
5171 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
5172 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
5173 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
5174 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
5175 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
5176 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
5177 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
5178 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
5179 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
5180 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
5181 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
5182 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
5183 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
5186 // FIXME: Some shuffle and unpack instructions have equivalents in different
5187 // domains, but they require a bit more work than just switching opcodes.
5189 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
5190 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
5191 if (ReplaceableInstrs[i][domain-1] == opcode)
5192 return ReplaceableInstrs[i];
5196 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
5197 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
5198 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
5199 return ReplaceableInstrsAVX2[i];
5203 std::pair<uint16_t, uint16_t>
5204 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
5205 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5206 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
5207 uint16_t validDomains = 0;
5208 if (domain && lookup(MI->getOpcode(), domain))
5210 else if (domain && lookupAVX2(MI->getOpcode(), domain))
5211 validDomains = hasAVX2 ? 0xe : 0x6;
5212 return std::make_pair(domain, validDomains);
5215 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
5216 assert(Domain>0 && Domain<4 && "Invalid execution domain");
5217 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5218 assert(dom && "Not an SSE instruction");
5219 const uint16_t *table = lookup(MI->getOpcode(), dom);
5220 if (!table) { // try the other table
5221 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
5222 "256-bit vector operations only available in AVX2");
5223 table = lookupAVX2(MI->getOpcode(), dom);
5225 assert(table && "Cannot change domain");
5226 MI->setDesc(get(table[Domain-1]));
5229 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5230 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5231 NopInst.setOpcode(X86::NOOP);
5234 bool X86InstrInfo::isHighLatencyDef(int opc) const {
5236 default: return false;
5238 case X86::DIVSDrm_Int:
5240 case X86::DIVSDrr_Int:
5242 case X86::DIVSSrm_Int:
5244 case X86::DIVSSrr_Int:
5250 case X86::SQRTSDm_Int:
5252 case X86::SQRTSDr_Int:
5254 case X86::SQRTSSm_Int:
5256 case X86::SQRTSSr_Int:
5257 // AVX instructions with high latency
5259 case X86::VDIVSDrm_Int:
5261 case X86::VDIVSDrr_Int:
5263 case X86::VDIVSSrm_Int:
5265 case X86::VDIVSSrr_Int:
5271 case X86::VSQRTSDm_Int:
5274 case X86::VSQRTSSm_Int:
5276 case X86::VSQRTPDZrm:
5277 case X86::VSQRTPDZrr:
5278 case X86::VSQRTPSZrm:
5279 case X86::VSQRTPSZrr:
5280 case X86::VSQRTSDZm:
5281 case X86::VSQRTSDZm_Int:
5282 case X86::VSQRTSDZr:
5283 case X86::VSQRTSSZm_Int:
5284 case X86::VSQRTSSZr:
5285 case X86::VSQRTSSZm:
5286 case X86::VDIVSDZrm:
5287 case X86::VDIVSDZrr:
5288 case X86::VDIVSSZrm:
5289 case X86::VDIVSSZrr:
5291 case X86::VGATHERQPSZrm:
5292 case X86::VGATHERQPDZrm:
5293 case X86::VGATHERDPDZrm:
5294 case X86::VGATHERDPSZrm:
5295 case X86::VPGATHERQDZrm:
5296 case X86::VPGATHERQQZrm:
5297 case X86::VPGATHERDDZrm:
5298 case X86::VPGATHERDQZrm:
5299 case X86::VSCATTERQPDZmr:
5300 case X86::VSCATTERQPSZmr:
5301 case X86::VSCATTERDPDZmr:
5302 case X86::VSCATTERDPSZmr:
5303 case X86::VPSCATTERQDZmr:
5304 case X86::VPSCATTERQQZmr:
5305 case X86::VPSCATTERDDZmr:
5306 case X86::VPSCATTERDQZmr:
5312 hasHighOperandLatency(const InstrItineraryData *ItinData,
5313 const MachineRegisterInfo *MRI,
5314 const MachineInstr *DefMI, unsigned DefIdx,
5315 const MachineInstr *UseMI, unsigned UseIdx) const {
5316 return isHighLatencyDef(DefMI->getOpcode());
5320 /// CGBR - Create Global Base Reg pass. This initializes the PIC
5321 /// global base register for x86-32.
5322 struct CGBR : public MachineFunctionPass {
5324 CGBR() : MachineFunctionPass(ID) {}
5326 bool runOnMachineFunction(MachineFunction &MF) override {
5327 const X86TargetMachine *TM =
5328 static_cast<const X86TargetMachine *>(&MF.getTarget());
5330 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
5331 "X86-64 PIC uses RIP relative addressing");
5333 // Only emit a global base reg in PIC mode.
5334 if (TM->getRelocationModel() != Reloc::PIC_)
5337 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5338 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5340 // If we didn't need a GlobalBaseReg, don't insert code.
5341 if (GlobalBaseReg == 0)
5344 // Insert the set of GlobalBaseReg into the first MBB of the function
5345 MachineBasicBlock &FirstMBB = MF.front();
5346 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5347 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5348 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5349 const X86InstrInfo *TII = TM->getInstrInfo();
5352 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
5353 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
5357 // Operand of MovePCtoStack is completely ignored by asm printer. It's
5358 // only used in JIT code emission as displacement to pc.
5359 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
5361 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5362 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5363 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
5364 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5365 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5366 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5367 X86II::MO_GOT_ABSOLUTE_ADDRESS);
5373 const char *getPassName() const override {
5374 return "X86 PIC Global Base Reg Initialization";
5377 void getAnalysisUsage(AnalysisUsage &AU) const override {
5378 AU.setPreservesCFG();
5379 MachineFunctionPass::getAnalysisUsage(AU);
5386 llvm::createGlobalBaseRegPass() { return new CGBR(); }
5389 struct LDTLSCleanup : public MachineFunctionPass {
5391 LDTLSCleanup() : MachineFunctionPass(ID) {}
5393 bool runOnMachineFunction(MachineFunction &MF) override {
5394 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5395 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5396 // No point folding accesses if there isn't at least two.
5400 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5401 return VisitNode(DT->getRootNode(), 0);
5404 // Visit the dominator subtree rooted at Node in pre-order.
5405 // If TLSBaseAddrReg is non-null, then use that to replace any
5406 // TLS_base_addr instructions. Otherwise, create the register
5407 // when the first such instruction is seen, and then use it
5408 // as we encounter more instructions.
5409 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5410 MachineBasicBlock *BB = Node->getBlock();
5411 bool Changed = false;
5413 // Traverse the current block.
5414 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5416 switch (I->getOpcode()) {
5417 case X86::TLS_base_addr32:
5418 case X86::TLS_base_addr64:
5420 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5422 I = SetRegister(I, &TLSBaseAddrReg);
5430 // Visit the children of this block in the dominator tree.
5431 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5433 Changed |= VisitNode(*I, TLSBaseAddrReg);
5439 // Replace the TLS_base_addr instruction I with a copy from
5440 // TLSBaseAddrReg, returning the new instruction.
5441 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5442 unsigned TLSBaseAddrReg) {
5443 MachineFunction *MF = I->getParent()->getParent();
5444 const X86TargetMachine *TM =
5445 static_cast<const X86TargetMachine *>(&MF->getTarget());
5446 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5447 const X86InstrInfo *TII = TM->getInstrInfo();
5449 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5450 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5451 TII->get(TargetOpcode::COPY),
5452 is64Bit ? X86::RAX : X86::EAX)
5453 .addReg(TLSBaseAddrReg);
5455 // Erase the TLS_base_addr instruction.
5456 I->eraseFromParent();
5461 // Create a virtal register in *TLSBaseAddrReg, and populate it by
5462 // inserting a copy instruction after I. Returns the new instruction.
5463 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5464 MachineFunction *MF = I->getParent()->getParent();
5465 const X86TargetMachine *TM =
5466 static_cast<const X86TargetMachine *>(&MF->getTarget());
5467 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5468 const X86InstrInfo *TII = TM->getInstrInfo();
5470 // Create a virtual register for the TLS base address.
5471 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5472 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5473 ? &X86::GR64RegClass
5474 : &X86::GR32RegClass);
5476 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5477 MachineInstr *Next = I->getNextNode();
5478 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5479 TII->get(TargetOpcode::COPY),
5481 .addReg(is64Bit ? X86::RAX : X86::EAX);
5486 const char *getPassName() const override {
5487 return "Local Dynamic TLS Access Clean-up";
5490 void getAnalysisUsage(AnalysisUsage &AU) const override {
5491 AU.setPreservesCFG();
5492 AU.addRequired<MachineDominatorTree>();
5493 MachineFunctionPass::getAnalysisUsage(AU);
5498 char LDTLSCleanup::ID = 0;
5500 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }