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"
41 #define DEBUG_TYPE "x86-instr-info"
43 #define GET_INSTRINFO_CTOR_DTOR
44 #include "X86GenInstrInfo.inc"
47 NoFusing("disable-spill-fusing",
48 cl::desc("Disable fusing of spill code into instructions"));
50 PrintFailedFusing("print-failed-fuse-candidates",
51 cl::desc("Print instructions that the allocator wants to"
52 " fuse, but the X86 backend currently can't"),
55 ReMatPICStubLoad("remat-pic-stub-load",
56 cl::desc("Re-materialize load from stub in PIC mode"),
57 cl::init(false), cl::Hidden);
60 // Select which memory operand is being unfolded.
61 // (stored in bits 0 - 3)
68 // Do not insert the reverse map (MemOp -> RegOp) into the table.
69 // This may be needed because there is a many -> one mapping.
70 TB_NO_REVERSE = 1 << 4,
72 // Do not insert the forward map (RegOp -> MemOp) into the table.
73 // This is needed for Native Client, which prohibits branch
74 // instructions from using a memory operand.
75 TB_NO_FORWARD = 1 << 5,
77 TB_FOLDED_LOAD = 1 << 6,
78 TB_FOLDED_STORE = 1 << 7,
80 // Minimum alignment required for load/store.
81 // Used for RegOp->MemOp conversion.
82 // (stored in bits 8 - 15)
84 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
85 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
86 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
87 TB_ALIGN_64 = 64 << TB_ALIGN_SHIFT,
88 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
91 struct X86OpTblEntry {
97 // Pin the vtable to this file.
98 void X86InstrInfo::anchor() {}
100 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
101 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
102 ? X86::ADJCALLSTACKDOWN64
103 : X86::ADJCALLSTACKDOWN32),
104 (tm.getSubtarget<X86Subtarget>().is64Bit()
105 ? X86::ADJCALLSTACKUP64
106 : X86::ADJCALLSTACKUP32)),
109 static const X86OpTblEntry OpTbl2Addr[] = {
110 { X86::ADC32ri, X86::ADC32mi, 0 },
111 { X86::ADC32ri8, X86::ADC32mi8, 0 },
112 { X86::ADC32rr, X86::ADC32mr, 0 },
113 { X86::ADC64ri32, X86::ADC64mi32, 0 },
114 { X86::ADC64ri8, X86::ADC64mi8, 0 },
115 { X86::ADC64rr, X86::ADC64mr, 0 },
116 { X86::ADD16ri, X86::ADD16mi, 0 },
117 { X86::ADD16ri8, X86::ADD16mi8, 0 },
118 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
119 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
120 { X86::ADD16rr, X86::ADD16mr, 0 },
121 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
122 { X86::ADD32ri, X86::ADD32mi, 0 },
123 { X86::ADD32ri8, X86::ADD32mi8, 0 },
124 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
125 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
126 { X86::ADD32rr, X86::ADD32mr, 0 },
127 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
128 { X86::ADD64ri32, X86::ADD64mi32, 0 },
129 { X86::ADD64ri8, X86::ADD64mi8, 0 },
130 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
131 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
132 { X86::ADD64rr, X86::ADD64mr, 0 },
133 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
134 { X86::ADD8ri, X86::ADD8mi, 0 },
135 { X86::ADD8rr, X86::ADD8mr, 0 },
136 { X86::AND16ri, X86::AND16mi, 0 },
137 { X86::AND16ri8, X86::AND16mi8, 0 },
138 { X86::AND16rr, X86::AND16mr, 0 },
139 { X86::AND32ri, X86::AND32mi, 0 },
140 { X86::AND32ri8, X86::AND32mi8, 0 },
141 { X86::AND32rr, X86::AND32mr, 0 },
142 { X86::AND64ri32, X86::AND64mi32, 0 },
143 { X86::AND64ri8, X86::AND64mi8, 0 },
144 { X86::AND64rr, X86::AND64mr, 0 },
145 { X86::AND8ri, X86::AND8mi, 0 },
146 { X86::AND8rr, X86::AND8mr, 0 },
147 { X86::DEC16r, X86::DEC16m, 0 },
148 { X86::DEC32r, X86::DEC32m, 0 },
149 { X86::DEC64_16r, X86::DEC64_16m, 0 },
150 { X86::DEC64_32r, X86::DEC64_32m, 0 },
151 { X86::DEC64r, X86::DEC64m, 0 },
152 { X86::DEC8r, X86::DEC8m, 0 },
153 { X86::INC16r, X86::INC16m, 0 },
154 { X86::INC32r, X86::INC32m, 0 },
155 { X86::INC64_16r, X86::INC64_16m, 0 },
156 { X86::INC64_32r, X86::INC64_32m, 0 },
157 { X86::INC64r, X86::INC64m, 0 },
158 { X86::INC8r, X86::INC8m, 0 },
159 { X86::NEG16r, X86::NEG16m, 0 },
160 { X86::NEG32r, X86::NEG32m, 0 },
161 { X86::NEG64r, X86::NEG64m, 0 },
162 { X86::NEG8r, X86::NEG8m, 0 },
163 { X86::NOT16r, X86::NOT16m, 0 },
164 { X86::NOT32r, X86::NOT32m, 0 },
165 { X86::NOT64r, X86::NOT64m, 0 },
166 { X86::NOT8r, X86::NOT8m, 0 },
167 { X86::OR16ri, X86::OR16mi, 0 },
168 { X86::OR16ri8, X86::OR16mi8, 0 },
169 { X86::OR16rr, X86::OR16mr, 0 },
170 { X86::OR32ri, X86::OR32mi, 0 },
171 { X86::OR32ri8, X86::OR32mi8, 0 },
172 { X86::OR32rr, X86::OR32mr, 0 },
173 { X86::OR64ri32, X86::OR64mi32, 0 },
174 { X86::OR64ri8, X86::OR64mi8, 0 },
175 { X86::OR64rr, X86::OR64mr, 0 },
176 { X86::OR8ri, X86::OR8mi, 0 },
177 { X86::OR8rr, X86::OR8mr, 0 },
178 { X86::ROL16r1, X86::ROL16m1, 0 },
179 { X86::ROL16rCL, X86::ROL16mCL, 0 },
180 { X86::ROL16ri, X86::ROL16mi, 0 },
181 { X86::ROL32r1, X86::ROL32m1, 0 },
182 { X86::ROL32rCL, X86::ROL32mCL, 0 },
183 { X86::ROL32ri, X86::ROL32mi, 0 },
184 { X86::ROL64r1, X86::ROL64m1, 0 },
185 { X86::ROL64rCL, X86::ROL64mCL, 0 },
186 { X86::ROL64ri, X86::ROL64mi, 0 },
187 { X86::ROL8r1, X86::ROL8m1, 0 },
188 { X86::ROL8rCL, X86::ROL8mCL, 0 },
189 { X86::ROL8ri, X86::ROL8mi, 0 },
190 { X86::ROR16r1, X86::ROR16m1, 0 },
191 { X86::ROR16rCL, X86::ROR16mCL, 0 },
192 { X86::ROR16ri, X86::ROR16mi, 0 },
193 { X86::ROR32r1, X86::ROR32m1, 0 },
194 { X86::ROR32rCL, X86::ROR32mCL, 0 },
195 { X86::ROR32ri, X86::ROR32mi, 0 },
196 { X86::ROR64r1, X86::ROR64m1, 0 },
197 { X86::ROR64rCL, X86::ROR64mCL, 0 },
198 { X86::ROR64ri, X86::ROR64mi, 0 },
199 { X86::ROR8r1, X86::ROR8m1, 0 },
200 { X86::ROR8rCL, X86::ROR8mCL, 0 },
201 { X86::ROR8ri, X86::ROR8mi, 0 },
202 { X86::SAR16r1, X86::SAR16m1, 0 },
203 { X86::SAR16rCL, X86::SAR16mCL, 0 },
204 { X86::SAR16ri, X86::SAR16mi, 0 },
205 { X86::SAR32r1, X86::SAR32m1, 0 },
206 { X86::SAR32rCL, X86::SAR32mCL, 0 },
207 { X86::SAR32ri, X86::SAR32mi, 0 },
208 { X86::SAR64r1, X86::SAR64m1, 0 },
209 { X86::SAR64rCL, X86::SAR64mCL, 0 },
210 { X86::SAR64ri, X86::SAR64mi, 0 },
211 { X86::SAR8r1, X86::SAR8m1, 0 },
212 { X86::SAR8rCL, X86::SAR8mCL, 0 },
213 { X86::SAR8ri, X86::SAR8mi, 0 },
214 { X86::SBB32ri, X86::SBB32mi, 0 },
215 { X86::SBB32ri8, X86::SBB32mi8, 0 },
216 { X86::SBB32rr, X86::SBB32mr, 0 },
217 { X86::SBB64ri32, X86::SBB64mi32, 0 },
218 { X86::SBB64ri8, X86::SBB64mi8, 0 },
219 { X86::SBB64rr, X86::SBB64mr, 0 },
220 { X86::SHL16rCL, X86::SHL16mCL, 0 },
221 { X86::SHL16ri, X86::SHL16mi, 0 },
222 { X86::SHL32rCL, X86::SHL32mCL, 0 },
223 { X86::SHL32ri, X86::SHL32mi, 0 },
224 { X86::SHL64rCL, X86::SHL64mCL, 0 },
225 { X86::SHL64ri, X86::SHL64mi, 0 },
226 { X86::SHL8rCL, X86::SHL8mCL, 0 },
227 { X86::SHL8ri, X86::SHL8mi, 0 },
228 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
229 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
230 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
231 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
232 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
233 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
234 { X86::SHR16r1, X86::SHR16m1, 0 },
235 { X86::SHR16rCL, X86::SHR16mCL, 0 },
236 { X86::SHR16ri, X86::SHR16mi, 0 },
237 { X86::SHR32r1, X86::SHR32m1, 0 },
238 { X86::SHR32rCL, X86::SHR32mCL, 0 },
239 { X86::SHR32ri, X86::SHR32mi, 0 },
240 { X86::SHR64r1, X86::SHR64m1, 0 },
241 { X86::SHR64rCL, X86::SHR64mCL, 0 },
242 { X86::SHR64ri, X86::SHR64mi, 0 },
243 { X86::SHR8r1, X86::SHR8m1, 0 },
244 { X86::SHR8rCL, X86::SHR8mCL, 0 },
245 { X86::SHR8ri, X86::SHR8mi, 0 },
246 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
247 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
248 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
249 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
250 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
251 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
252 { X86::SUB16ri, X86::SUB16mi, 0 },
253 { X86::SUB16ri8, X86::SUB16mi8, 0 },
254 { X86::SUB16rr, X86::SUB16mr, 0 },
255 { X86::SUB32ri, X86::SUB32mi, 0 },
256 { X86::SUB32ri8, X86::SUB32mi8, 0 },
257 { X86::SUB32rr, X86::SUB32mr, 0 },
258 { X86::SUB64ri32, X86::SUB64mi32, 0 },
259 { X86::SUB64ri8, X86::SUB64mi8, 0 },
260 { X86::SUB64rr, X86::SUB64mr, 0 },
261 { X86::SUB8ri, X86::SUB8mi, 0 },
262 { X86::SUB8rr, X86::SUB8mr, 0 },
263 { X86::XOR16ri, X86::XOR16mi, 0 },
264 { X86::XOR16ri8, X86::XOR16mi8, 0 },
265 { X86::XOR16rr, X86::XOR16mr, 0 },
266 { X86::XOR32ri, X86::XOR32mi, 0 },
267 { X86::XOR32ri8, X86::XOR32mi8, 0 },
268 { X86::XOR32rr, X86::XOR32mr, 0 },
269 { X86::XOR64ri32, X86::XOR64mi32, 0 },
270 { X86::XOR64ri8, X86::XOR64mi8, 0 },
271 { X86::XOR64rr, X86::XOR64mr, 0 },
272 { X86::XOR8ri, X86::XOR8mi, 0 },
273 { X86::XOR8rr, X86::XOR8mr, 0 }
276 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
277 unsigned RegOp = OpTbl2Addr[i].RegOp;
278 unsigned MemOp = OpTbl2Addr[i].MemOp;
279 unsigned Flags = OpTbl2Addr[i].Flags;
280 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
282 // Index 0, folded load and store, no alignment requirement.
283 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
286 static const X86OpTblEntry OpTbl0[] = {
287 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
288 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
289 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
290 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
291 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
292 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
293 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
294 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
295 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
296 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
297 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
298 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
299 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
300 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
301 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
302 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
303 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
304 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
305 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
306 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
307 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
308 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
309 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
310 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
311 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
312 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
313 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
314 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
315 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
316 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
317 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
318 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
319 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
320 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
321 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
322 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
323 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
324 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
325 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
326 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
327 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
328 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
329 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
330 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
331 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
332 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
333 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
334 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
335 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
336 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
337 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
338 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
339 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
340 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
341 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
342 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
343 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
344 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
345 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
346 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
347 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
348 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
349 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
350 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
351 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
352 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
353 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
354 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
355 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
356 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
357 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
358 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
359 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
360 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
361 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
362 // AVX 128-bit versions of foldable instructions
363 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
364 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
366 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
367 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
368 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
369 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
370 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
371 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
372 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
373 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
374 // AVX 256-bit foldable instructions
375 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
376 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
377 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
378 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
379 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
380 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE },
381 // AVX-512 foldable instructions
382 { X86::VMOVPDI2DIZrr,X86::VMOVPDI2DIZmr, TB_FOLDED_STORE }
385 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
386 unsigned RegOp = OpTbl0[i].RegOp;
387 unsigned MemOp = OpTbl0[i].MemOp;
388 unsigned Flags = OpTbl0[i].Flags;
389 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
390 RegOp, MemOp, TB_INDEX_0 | Flags);
393 static const X86OpTblEntry OpTbl1[] = {
394 { X86::CMP16rr, X86::CMP16rm, 0 },
395 { X86::CMP32rr, X86::CMP32rm, 0 },
396 { X86::CMP64rr, X86::CMP64rm, 0 },
397 { X86::CMP8rr, X86::CMP8rm, 0 },
398 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
399 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
400 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
401 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
402 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
403 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
404 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
405 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
406 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
407 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
408 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
409 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
410 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
411 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
412 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
413 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
414 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
415 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
416 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
417 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
418 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
419 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
420 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
421 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
422 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
423 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
424 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
425 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
426 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
427 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
428 { X86::MOV16rr, X86::MOV16rm, 0 },
429 { X86::MOV32rr, X86::MOV32rm, 0 },
430 { X86::MOV64rr, X86::MOV64rm, 0 },
431 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
432 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
433 { X86::MOV8rr, X86::MOV8rm, 0 },
434 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
435 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
436 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
437 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
438 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
439 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
440 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
441 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
442 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
443 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
444 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
445 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
446 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
447 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
448 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
449 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
450 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
451 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
452 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
453 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
454 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
455 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
456 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
457 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
458 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
459 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
460 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
461 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
462 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
463 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
464 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
465 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
466 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
467 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
468 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
469 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
470 { X86::SQRTSDr, X86::SQRTSDm, 0 },
471 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
472 { X86::SQRTSSr, X86::SQRTSSm, 0 },
473 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
474 { X86::TEST16rr, X86::TEST16rm, 0 },
475 { X86::TEST32rr, X86::TEST32rm, 0 },
476 { X86::TEST64rr, X86::TEST64rm, 0 },
477 { X86::TEST8rr, X86::TEST8rm, 0 },
478 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
479 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
480 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
481 // AVX 128-bit versions of foldable instructions
482 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
483 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
484 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
485 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
486 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
487 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
488 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
489 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
490 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
491 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
492 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
493 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
494 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
495 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
496 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
497 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
498 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
499 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
500 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
501 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
502 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
503 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
504 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
505 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
506 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
507 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
508 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
509 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
510 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
511 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
512 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
513 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
514 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
515 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
516 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
517 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
518 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
519 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
520 { X86::VRCPPSr, X86::VRCPPSm, 0 },
521 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
522 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
523 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
524 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
525 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
526 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
527 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
528 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
530 // AVX 256-bit foldable instructions
531 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
532 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
533 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
534 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
535 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
536 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
537 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
539 // AVX2 foldable instructions
540 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
541 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
542 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
543 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
544 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
545 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
546 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
547 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
548 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
549 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
550 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
551 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
552 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
554 // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
555 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
556 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
557 { X86::BEXTRI32ri, X86::BEXTRI32mi, 0 },
558 { X86::BEXTRI64ri, X86::BEXTRI64mi, 0 },
559 { X86::BLCFILL32rr, X86::BLCFILL32rm, 0 },
560 { X86::BLCFILL64rr, X86::BLCFILL64rm, 0 },
561 { X86::BLCI32rr, X86::BLCI32rm, 0 },
562 { X86::BLCI64rr, X86::BLCI64rm, 0 },
563 { X86::BLCIC32rr, X86::BLCIC32rm, 0 },
564 { X86::BLCIC64rr, X86::BLCIC64rm, 0 },
565 { X86::BLCMSK32rr, X86::BLCMSK32rm, 0 },
566 { X86::BLCMSK64rr, X86::BLCMSK64rm, 0 },
567 { X86::BLCS32rr, X86::BLCS32rm, 0 },
568 { X86::BLCS64rr, X86::BLCS64rm, 0 },
569 { X86::BLSFILL32rr, X86::BLSFILL32rm, 0 },
570 { X86::BLSFILL64rr, X86::BLSFILL64rm, 0 },
571 { X86::BLSI32rr, X86::BLSI32rm, 0 },
572 { X86::BLSI64rr, X86::BLSI64rm, 0 },
573 { X86::BLSIC32rr, X86::BLSIC32rm, 0 },
574 { X86::BLSIC64rr, X86::BLSIC64rm, 0 },
575 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
576 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
577 { X86::BLSR32rr, X86::BLSR32rm, 0 },
578 { X86::BLSR64rr, X86::BLSR64rm, 0 },
579 { X86::BZHI32rr, X86::BZHI32rm, 0 },
580 { X86::BZHI64rr, X86::BZHI64rm, 0 },
581 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
582 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
583 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
584 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
585 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
586 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
587 { X86::RORX32ri, X86::RORX32mi, 0 },
588 { X86::RORX64ri, X86::RORX64mi, 0 },
589 { X86::SARX32rr, X86::SARX32rm, 0 },
590 { X86::SARX64rr, X86::SARX64rm, 0 },
591 { X86::SHRX32rr, X86::SHRX32rm, 0 },
592 { X86::SHRX64rr, X86::SHRX64rm, 0 },
593 { X86::SHLX32rr, X86::SHLX32rm, 0 },
594 { X86::SHLX64rr, X86::SHLX64rm, 0 },
595 { X86::T1MSKC32rr, X86::T1MSKC32rm, 0 },
596 { X86::T1MSKC64rr, X86::T1MSKC64rm, 0 },
597 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
598 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
599 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
600 { X86::TZMSK32rr, X86::TZMSK32rm, 0 },
601 { X86::TZMSK64rr, X86::TZMSK64rm, 0 },
603 // AVX-512 foldable instructions
604 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 },
605 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 },
606 { X86::VMOVDQA32rr, X86::VMOVDQA32rm, TB_ALIGN_64 },
607 { X86::VMOVDQA64rr, X86::VMOVDQA64rm, TB_ALIGN_64 },
608 { X86::VMOVDQU32rr, X86::VMOVDQU32rm, 0 },
609 { X86::VMOVDQU64rr, X86::VMOVDQU64rm, 0 },
610 { X86::VPABSDZrr, X86::VPABSDZrm, 0 },
611 { X86::VPABSQZrr, X86::VPABSQZrm, 0 },
613 // AES foldable instructions
614 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 },
615 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 },
616 { X86::VAESIMCrr, X86::VAESIMCrm, TB_ALIGN_16 },
617 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, TB_ALIGN_16 },
620 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
621 unsigned RegOp = OpTbl1[i].RegOp;
622 unsigned MemOp = OpTbl1[i].MemOp;
623 unsigned Flags = OpTbl1[i].Flags;
624 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
626 // Index 1, folded load
627 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
630 static const X86OpTblEntry OpTbl2[] = {
631 { X86::ADC32rr, X86::ADC32rm, 0 },
632 { X86::ADC64rr, X86::ADC64rm, 0 },
633 { X86::ADD16rr, X86::ADD16rm, 0 },
634 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
635 { X86::ADD32rr, X86::ADD32rm, 0 },
636 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
637 { X86::ADD64rr, X86::ADD64rm, 0 },
638 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
639 { X86::ADD8rr, X86::ADD8rm, 0 },
640 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
641 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
642 { X86::ADDSDrr, X86::ADDSDrm, 0 },
643 { X86::ADDSSrr, X86::ADDSSrm, 0 },
644 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
645 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
646 { X86::AND16rr, X86::AND16rm, 0 },
647 { X86::AND32rr, X86::AND32rm, 0 },
648 { X86::AND64rr, X86::AND64rm, 0 },
649 { X86::AND8rr, X86::AND8rm, 0 },
650 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
651 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
652 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
653 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
654 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
655 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
656 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
657 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
658 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
659 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
660 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
661 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
662 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
663 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
664 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
665 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
666 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
667 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
668 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
669 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
670 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
671 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
672 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
673 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
674 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
675 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
676 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
677 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
678 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
679 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
680 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
681 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
682 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
683 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
684 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
685 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
686 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
687 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
688 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
689 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
690 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
691 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
692 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
693 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
694 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
695 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
696 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
697 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
698 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
699 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
700 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
701 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
702 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
703 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
704 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
705 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
706 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
707 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
708 { X86::CMPSDrr, X86::CMPSDrm, 0 },
709 { X86::CMPSSrr, X86::CMPSSrm, 0 },
710 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
711 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
712 { X86::DIVSDrr, X86::DIVSDrm, 0 },
713 { X86::DIVSSrr, X86::DIVSSrm, 0 },
714 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
715 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
716 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
717 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
718 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
719 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
720 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
721 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
722 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
723 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
724 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
725 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
726 { X86::IMUL16rr, X86::IMUL16rm, 0 },
727 { X86::IMUL32rr, X86::IMUL32rm, 0 },
728 { X86::IMUL64rr, X86::IMUL64rm, 0 },
729 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
730 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
731 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
732 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
733 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
734 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
735 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
736 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
737 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
738 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
739 { X86::MAXSDrr, X86::MAXSDrm, 0 },
740 { X86::MAXSSrr, X86::MAXSSrm, 0 },
741 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
742 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
743 { X86::MINSDrr, X86::MINSDrm, 0 },
744 { X86::MINSSrr, X86::MINSSrm, 0 },
745 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
746 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
747 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
748 { X86::MULSDrr, X86::MULSDrm, 0 },
749 { X86::MULSSrr, X86::MULSSrm, 0 },
750 { X86::OR16rr, X86::OR16rm, 0 },
751 { X86::OR32rr, X86::OR32rm, 0 },
752 { X86::OR64rr, X86::OR64rm, 0 },
753 { X86::OR8rr, X86::OR8rm, 0 },
754 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
755 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
756 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
757 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
758 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
759 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
760 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
761 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
762 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
763 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
764 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
765 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
766 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
767 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
768 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
769 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
770 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
771 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
772 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
773 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
774 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
775 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
776 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
777 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
778 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
779 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
780 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
781 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
782 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
783 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
784 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
785 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
786 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
787 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
788 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
789 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
790 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
791 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
792 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
793 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
794 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
795 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
796 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
797 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
798 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
799 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
800 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
801 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
802 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
803 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
804 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
805 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
806 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
807 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
808 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
809 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
810 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
811 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
812 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
813 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
814 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
815 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
816 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
817 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
818 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
819 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
820 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
821 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
822 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
823 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
824 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
825 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
826 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
827 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
828 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
829 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
830 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
831 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
832 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
833 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
834 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
835 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
836 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
837 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
838 { X86::SBB32rr, X86::SBB32rm, 0 },
839 { X86::SBB64rr, X86::SBB64rm, 0 },
840 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
841 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
842 { X86::SUB16rr, X86::SUB16rm, 0 },
843 { X86::SUB32rr, X86::SUB32rm, 0 },
844 { X86::SUB64rr, X86::SUB64rm, 0 },
845 { X86::SUB8rr, X86::SUB8rm, 0 },
846 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
847 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
848 { X86::SUBSDrr, X86::SUBSDrm, 0 },
849 { X86::SUBSSrr, X86::SUBSSrm, 0 },
850 // FIXME: TEST*rr -> swapped operand of TEST*mr.
851 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
852 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
853 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
854 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
855 { X86::XOR16rr, X86::XOR16rm, 0 },
856 { X86::XOR32rr, X86::XOR32rm, 0 },
857 { X86::XOR64rr, X86::XOR64rm, 0 },
858 { X86::XOR8rr, X86::XOR8rm, 0 },
859 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
860 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
861 // AVX 128-bit versions of foldable instructions
862 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
863 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
864 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
865 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
866 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
867 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
868 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
869 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
870 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
871 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
872 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
873 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
874 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
875 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
876 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
877 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
878 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
879 { X86::VADDPDrr, X86::VADDPDrm, 0 },
880 { X86::VADDPSrr, X86::VADDPSrm, 0 },
881 { X86::VADDSDrr, X86::VADDSDrm, 0 },
882 { X86::VADDSSrr, X86::VADDSSrm, 0 },
883 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
884 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
885 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
886 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
887 { X86::VANDPDrr, X86::VANDPDrm, 0 },
888 { X86::VANDPSrr, X86::VANDPSrm, 0 },
889 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
890 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
891 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
892 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
893 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
894 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
895 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
896 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
897 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
898 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
899 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
900 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
901 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
902 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
903 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
904 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
905 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
906 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
907 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
908 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
909 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
910 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
911 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
912 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
913 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
914 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
915 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
916 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
917 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
918 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
919 { X86::VMINPDrr, X86::VMINPDrm, 0 },
920 { X86::VMINPSrr, X86::VMINPSrm, 0 },
921 { X86::VMINSDrr, X86::VMINSDrm, 0 },
922 { X86::VMINSSrr, X86::VMINSSrm, 0 },
923 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
924 { X86::VMULPDrr, X86::VMULPDrm, 0 },
925 { X86::VMULPSrr, X86::VMULPSrm, 0 },
926 { X86::VMULSDrr, X86::VMULSDrm, 0 },
927 { X86::VMULSSrr, X86::VMULSSrm, 0 },
928 { X86::VORPDrr, X86::VORPDrm, 0 },
929 { X86::VORPSrr, X86::VORPSrm, 0 },
930 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
931 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
932 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
933 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
934 { X86::VPADDBrr, X86::VPADDBrm, 0 },
935 { X86::VPADDDrr, X86::VPADDDrm, 0 },
936 { X86::VPADDQrr, X86::VPADDQrm, 0 },
937 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
938 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
939 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
940 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
941 { X86::VPADDWrr, X86::VPADDWrm, 0 },
942 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
943 { X86::VPANDNrr, X86::VPANDNrm, 0 },
944 { X86::VPANDrr, X86::VPANDrm, 0 },
945 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
946 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
947 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
948 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
949 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
950 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
951 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
952 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
953 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
954 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
955 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
956 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
957 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
958 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
959 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
960 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
961 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
962 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
963 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
964 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
965 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
966 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
967 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
968 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
969 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
970 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
971 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
972 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
973 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
974 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
975 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
976 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
977 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
978 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
979 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
980 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
981 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
982 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
983 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
984 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
985 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
986 { X86::VPORrr, X86::VPORrm, 0 },
987 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
988 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
989 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
990 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
991 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
992 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
993 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
994 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
995 { X86::VPSRADrr, X86::VPSRADrm, 0 },
996 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
997 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
998 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
999 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
1000 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
1001 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
1002 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
1003 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
1004 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
1005 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
1006 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
1007 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
1008 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
1009 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
1010 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
1011 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
1012 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
1013 { X86::VPXORrr, X86::VPXORrm, 0 },
1014 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
1015 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
1016 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
1017 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
1018 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
1019 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
1020 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
1021 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
1022 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
1023 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
1024 { X86::VXORPDrr, X86::VXORPDrm, 0 },
1025 { X86::VXORPSrr, X86::VXORPSrm, 0 },
1026 // AVX 256-bit foldable instructions
1027 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
1028 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
1029 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
1030 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
1031 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1032 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1033 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1034 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1035 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1036 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1037 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1038 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1039 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1040 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1041 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1042 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1043 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1044 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1045 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1046 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1047 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1048 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
1049 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
1050 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
1051 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
1052 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1053 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1054 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1055 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1056 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1057 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1058 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1059 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1060 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1061 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1062 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1063 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1064 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1065 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1066 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1067 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1068 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
1069 // AVX2 foldable instructions
1070 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1071 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1072 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1073 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1074 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1075 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1076 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1077 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1078 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1079 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1080 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1081 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1082 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1083 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1084 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1085 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1086 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1087 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1088 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1089 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1090 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1091 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1092 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1093 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1094 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1095 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1096 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1097 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1098 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1099 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1100 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1101 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1102 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1103 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1104 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1105 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1106 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1107 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1108 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1109 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1110 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1111 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1112 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1113 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1114 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1115 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1116 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1117 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1118 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1119 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1120 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1121 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1122 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1123 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1124 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1125 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1126 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1127 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1128 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1129 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1130 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1131 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1132 { X86::VPORYrr, X86::VPORYrm, 0 },
1133 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1134 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1135 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1136 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1137 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1138 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1139 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1140 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1141 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1142 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1143 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1144 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1145 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1146 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1147 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1148 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1149 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1150 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1151 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1152 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1153 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1154 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1155 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1156 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1157 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1158 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1159 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1160 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1161 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1162 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1163 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1164 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1165 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1166 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1167 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1168 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1169 { X86::VPXORYrr, X86::VPXORYrm, 0 },
1170 // FIXME: add AVX 256-bit foldable instructions
1172 // FMA4 foldable patterns
1173 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1174 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
1175 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1176 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1177 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1178 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1179 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1180 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
1181 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1182 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1183 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1184 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1185 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1186 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
1187 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1188 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1189 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1190 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1191 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1192 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
1193 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1194 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1195 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1196 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1197 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1198 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1199 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1200 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1201 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1202 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1203 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1204 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
1206 // BMI/BMI2 foldable instructions
1207 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1208 { X86::ANDN64rr, X86::ANDN64rm, 0 },
1209 { X86::MULX32rr, X86::MULX32rm, 0 },
1210 { X86::MULX64rr, X86::MULX64rm, 0 },
1211 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1212 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1213 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1214 { X86::PEXT64rr, X86::PEXT64rm, 0 },
1216 // AVX-512 foldable instructions
1217 { X86::VADDPSZrr, X86::VADDPSZrm, 0 },
1218 { X86::VADDPDZrr, X86::VADDPDZrm, 0 },
1219 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 },
1220 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 },
1221 { X86::VMULPSZrr, X86::VMULPSZrm, 0 },
1222 { X86::VMULPDZrr, X86::VMULPDZrm, 0 },
1223 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 },
1224 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 },
1225 { X86::VMINPSZrr, X86::VMINPSZrm, 0 },
1226 { X86::VMINPDZrr, X86::VMINPDZrm, 0 },
1227 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 },
1228 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 },
1229 { X86::VPADDDZrr, X86::VPADDDZrm, 0 },
1230 { X86::VPADDQZrr, X86::VPADDQZrm, 0 },
1231 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1232 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1233 { X86::VPMAXSDZrr, X86::VPMAXSDZrm, 0 },
1234 { X86::VPMAXSQZrr, X86::VPMAXSQZrm, 0 },
1235 { X86::VPMAXUDZrr, X86::VPMAXUDZrm, 0 },
1236 { X86::VPMAXUQZrr, X86::VPMAXUQZrm, 0 },
1237 { X86::VPMINSDZrr, X86::VPMINSDZrm, 0 },
1238 { X86::VPMINSQZrr, X86::VPMINSQZrm, 0 },
1239 { X86::VPMINUDZrr, X86::VPMINUDZrm, 0 },
1240 { X86::VPMINUQZrr, X86::VPMINUQZrm, 0 },
1241 { X86::VPMULDQZrr, X86::VPMULDQZrm, 0 },
1242 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 },
1243 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 },
1244 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 },
1245 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 },
1246 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 },
1247 { X86::VPSUBDZrr, X86::VPSUBDZrm, 0 },
1248 { X86::VPSUBQZrr, X86::VPSUBQZrm, 0 },
1249 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 },
1250 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 },
1251 { X86::VALIGNQrri, X86::VALIGNQrmi, 0 },
1252 { X86::VALIGNDrri, X86::VALIGNDrmi, 0 },
1253 { X86::VPMULUDQZrr, X86::VPMULUDQZrm, 0 },
1255 // AES foldable instructions
1256 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 },
1257 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 },
1258 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 },
1259 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 },
1260 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, TB_ALIGN_16 },
1261 { X86::VAESDECrr, X86::VAESDECrm, TB_ALIGN_16 },
1262 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, TB_ALIGN_16 },
1263 { X86::VAESENCrr, X86::VAESENCrm, TB_ALIGN_16 },
1265 // SHA foldable instructions
1266 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 },
1267 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 },
1268 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 },
1269 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 },
1270 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 },
1271 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 },
1272 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 },
1275 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1276 unsigned RegOp = OpTbl2[i].RegOp;
1277 unsigned MemOp = OpTbl2[i].MemOp;
1278 unsigned Flags = OpTbl2[i].Flags;
1279 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1281 // Index 2, folded load
1282 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1285 static const X86OpTblEntry OpTbl3[] = {
1286 // FMA foldable instructions
1287 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, TB_ALIGN_NONE },
1288 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, TB_ALIGN_NONE },
1289 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, TB_ALIGN_NONE },
1290 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, TB_ALIGN_NONE },
1291 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, TB_ALIGN_NONE },
1292 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, TB_ALIGN_NONE },
1294 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_NONE },
1295 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_NONE },
1296 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_NONE },
1297 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_NONE },
1298 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_NONE },
1299 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_NONE },
1300 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_NONE },
1301 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_NONE },
1302 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_NONE },
1303 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_NONE },
1304 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_NONE },
1305 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_NONE },
1307 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, TB_ALIGN_NONE },
1308 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, TB_ALIGN_NONE },
1309 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, TB_ALIGN_NONE },
1310 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, TB_ALIGN_NONE },
1311 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, TB_ALIGN_NONE },
1312 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, TB_ALIGN_NONE },
1314 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_NONE },
1315 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_NONE },
1316 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_NONE },
1317 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_NONE },
1318 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_NONE },
1319 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_NONE },
1320 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_NONE },
1321 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_NONE },
1322 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_NONE },
1323 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_NONE },
1324 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_NONE },
1325 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_NONE },
1327 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, TB_ALIGN_NONE },
1328 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, TB_ALIGN_NONE },
1329 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, TB_ALIGN_NONE },
1330 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, TB_ALIGN_NONE },
1331 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, TB_ALIGN_NONE },
1332 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, TB_ALIGN_NONE },
1334 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_NONE },
1335 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_NONE },
1336 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_NONE },
1337 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_NONE },
1338 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_NONE },
1339 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_NONE },
1340 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_NONE },
1341 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_NONE },
1342 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_NONE },
1343 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_NONE },
1344 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_NONE },
1345 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_NONE },
1347 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, TB_ALIGN_NONE },
1348 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, TB_ALIGN_NONE },
1349 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, TB_ALIGN_NONE },
1350 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, TB_ALIGN_NONE },
1351 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, TB_ALIGN_NONE },
1352 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, TB_ALIGN_NONE },
1354 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_NONE },
1355 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_NONE },
1356 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_NONE },
1357 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_NONE },
1358 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_NONE },
1359 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_NONE },
1360 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_NONE },
1361 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_NONE },
1362 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_NONE },
1363 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_NONE },
1364 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_NONE },
1365 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_NONE },
1367 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_NONE },
1368 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_NONE },
1369 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_NONE },
1370 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_NONE },
1371 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_NONE },
1372 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_NONE },
1373 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_NONE },
1374 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_NONE },
1375 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_NONE },
1376 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_NONE },
1377 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_NONE },
1378 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_NONE },
1380 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_NONE },
1381 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_NONE },
1382 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_NONE },
1383 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_NONE },
1384 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_NONE },
1385 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_NONE },
1386 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_NONE },
1387 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_NONE },
1388 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_NONE },
1389 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_NONE },
1390 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_NONE },
1391 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_NONE },
1393 // FMA4 foldable patterns
1394 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1395 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
1396 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1397 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1398 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1399 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1400 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1401 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
1402 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1403 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1404 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1405 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1406 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1407 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
1408 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1409 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1410 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1411 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1412 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1413 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
1414 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1415 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1416 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1417 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1418 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1419 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1420 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1421 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1422 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1423 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1424 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1425 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
1426 // AVX-512 VPERMI instructions with 3 source operands.
1427 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1428 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1429 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1430 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
1431 { X86::VBLENDMPDZrr, X86::VBLENDMPDZrm, 0 },
1432 { X86::VBLENDMPSZrr, X86::VBLENDMPSZrm, 0 },
1433 { X86::VPBLENDMDZrr, X86::VPBLENDMDZrm, 0 },
1434 { X86::VPBLENDMQZrr, X86::VPBLENDMQZrm, 0 }
1437 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1438 unsigned RegOp = OpTbl3[i].RegOp;
1439 unsigned MemOp = OpTbl3[i].MemOp;
1440 unsigned Flags = OpTbl3[i].Flags;
1441 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1443 // Index 3, folded load
1444 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1450 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1451 MemOp2RegOpTableType &M2RTable,
1452 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1453 if ((Flags & TB_NO_FORWARD) == 0) {
1454 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1455 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1457 if ((Flags & TB_NO_REVERSE) == 0) {
1458 assert(!M2RTable.count(MemOp) &&
1459 "Duplicated entries in unfolding maps?");
1460 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1465 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1466 unsigned &SrcReg, unsigned &DstReg,
1467 unsigned &SubIdx) const {
1468 switch (MI.getOpcode()) {
1470 case X86::MOVSX16rr8:
1471 case X86::MOVZX16rr8:
1472 case X86::MOVSX32rr8:
1473 case X86::MOVZX32rr8:
1474 case X86::MOVSX64rr8:
1475 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1476 // It's not always legal to reference the low 8-bit of the larger
1477 // register in 32-bit mode.
1479 case X86::MOVSX32rr16:
1480 case X86::MOVZX32rr16:
1481 case X86::MOVSX64rr16:
1482 case X86::MOVSX64rr32: {
1483 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1486 SrcReg = MI.getOperand(1).getReg();
1487 DstReg = MI.getOperand(0).getReg();
1488 switch (MI.getOpcode()) {
1489 default: llvm_unreachable("Unreachable!");
1490 case X86::MOVSX16rr8:
1491 case X86::MOVZX16rr8:
1492 case X86::MOVSX32rr8:
1493 case X86::MOVZX32rr8:
1494 case X86::MOVSX64rr8:
1495 SubIdx = X86::sub_8bit;
1497 case X86::MOVSX32rr16:
1498 case X86::MOVZX32rr16:
1499 case X86::MOVSX64rr16:
1500 SubIdx = X86::sub_16bit;
1502 case X86::MOVSX64rr32:
1503 SubIdx = X86::sub_32bit;
1512 /// isFrameOperand - Return true and the FrameIndex if the specified
1513 /// operand and follow operands form a reference to the stack frame.
1514 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1515 int &FrameIndex) const {
1516 if (MI->getOperand(Op+X86::AddrBaseReg).isFI() &&
1517 MI->getOperand(Op+X86::AddrScaleAmt).isImm() &&
1518 MI->getOperand(Op+X86::AddrIndexReg).isReg() &&
1519 MI->getOperand(Op+X86::AddrDisp).isImm() &&
1520 MI->getOperand(Op+X86::AddrScaleAmt).getImm() == 1 &&
1521 MI->getOperand(Op+X86::AddrIndexReg).getReg() == 0 &&
1522 MI->getOperand(Op+X86::AddrDisp).getImm() == 0) {
1523 FrameIndex = MI->getOperand(Op+X86::AddrBaseReg).getIndex();
1529 static bool isFrameLoadOpcode(int Opcode) {
1545 case X86::VMOVAPSrm:
1546 case X86::VMOVAPDrm:
1547 case X86::VMOVDQArm:
1548 case X86::VMOVAPSYrm:
1549 case X86::VMOVAPDYrm:
1550 case X86::VMOVDQAYrm:
1551 case X86::MMX_MOVD64rm:
1552 case X86::MMX_MOVQ64rm:
1553 case X86::VMOVAPSZrm:
1554 case X86::VMOVUPSZrm:
1559 static bool isFrameStoreOpcode(int Opcode) {
1566 case X86::ST_FpP64m:
1574 case X86::VMOVAPSmr:
1575 case X86::VMOVAPDmr:
1576 case X86::VMOVDQAmr:
1577 case X86::VMOVAPSYmr:
1578 case X86::VMOVAPDYmr:
1579 case X86::VMOVDQAYmr:
1580 case X86::VMOVUPSZmr:
1581 case X86::VMOVAPSZmr:
1582 case X86::MMX_MOVD64mr:
1583 case X86::MMX_MOVQ64mr:
1584 case X86::MMX_MOVNTQmr:
1590 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1591 int &FrameIndex) const {
1592 if (isFrameLoadOpcode(MI->getOpcode()))
1593 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1594 return MI->getOperand(0).getReg();
1598 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1599 int &FrameIndex) const {
1600 if (isFrameLoadOpcode(MI->getOpcode())) {
1602 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1604 // Check for post-frame index elimination operations
1605 const MachineMemOperand *Dummy;
1606 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1611 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1612 int &FrameIndex) const {
1613 if (isFrameStoreOpcode(MI->getOpcode()))
1614 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1615 isFrameOperand(MI, 0, FrameIndex))
1616 return MI->getOperand(X86::AddrNumOperands).getReg();
1620 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1621 int &FrameIndex) const {
1622 if (isFrameStoreOpcode(MI->getOpcode())) {
1624 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1626 // Check for post-frame index elimination operations
1627 const MachineMemOperand *Dummy;
1628 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1633 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1635 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1636 // Don't waste compile time scanning use-def chains of physregs.
1637 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1639 bool isPICBase = false;
1640 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
1641 E = MRI.def_instr_end(); I != E; ++I) {
1642 MachineInstr *DefMI = &*I;
1643 if (DefMI->getOpcode() != X86::MOVPC32r)
1645 assert(!isPICBase && "More than one PIC base?");
1652 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1653 AliasAnalysis *AA) const {
1654 switch (MI->getOpcode()) {
1670 case X86::VMOVAPSrm:
1671 case X86::VMOVUPSrm:
1672 case X86::VMOVAPDrm:
1673 case X86::VMOVDQArm:
1674 case X86::VMOVDQUrm:
1675 case X86::VMOVAPSYrm:
1676 case X86::VMOVUPSYrm:
1677 case X86::VMOVAPDYrm:
1678 case X86::VMOVDQAYrm:
1679 case X86::VMOVDQUYrm:
1680 case X86::MMX_MOVD64rm:
1681 case X86::MMX_MOVQ64rm:
1682 case X86::FsVMOVAPSrm:
1683 case X86::FsVMOVAPDrm:
1684 case X86::FsMOVAPSrm:
1685 case X86::FsMOVAPDrm: {
1686 // Loads from constant pools are trivially rematerializable.
1687 if (MI->getOperand(1+X86::AddrBaseReg).isReg() &&
1688 MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
1689 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
1690 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
1691 MI->isInvariantLoad(AA)) {
1692 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
1693 if (BaseReg == 0 || BaseReg == X86::RIP)
1695 // Allow re-materialization of PIC load.
1696 if (!ReMatPICStubLoad && MI->getOperand(1+X86::AddrDisp).isGlobal())
1698 const MachineFunction &MF = *MI->getParent()->getParent();
1699 const MachineRegisterInfo &MRI = MF.getRegInfo();
1700 return regIsPICBase(BaseReg, MRI);
1707 if (MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
1708 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
1709 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
1710 !MI->getOperand(1+X86::AddrDisp).isReg()) {
1711 // lea fi#, lea GV, etc. are all rematerializable.
1712 if (!MI->getOperand(1+X86::AddrBaseReg).isReg())
1714 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
1717 // Allow re-materialization of lea PICBase + x.
1718 const MachineFunction &MF = *MI->getParent()->getParent();
1719 const MachineRegisterInfo &MRI = MF.getRegInfo();
1720 return regIsPICBase(BaseReg, MRI);
1726 // All other instructions marked M_REMATERIALIZABLE are always trivially
1727 // rematerializable.
1731 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1732 /// would clobber the EFLAGS condition register. Note the result may be
1733 /// conservative. If it cannot definitely determine the safety after visiting
1734 /// a few instructions in each direction it assumes it's not safe.
1735 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1736 MachineBasicBlock::iterator I) {
1737 MachineBasicBlock::iterator E = MBB.end();
1739 // For compile time consideration, if we are not able to determine the
1740 // safety after visiting 4 instructions in each direction, we will assume
1742 MachineBasicBlock::iterator Iter = I;
1743 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1744 bool SeenDef = false;
1745 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1746 MachineOperand &MO = Iter->getOperand(j);
1747 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1751 if (MO.getReg() == X86::EFLAGS) {
1759 // This instruction defines EFLAGS, no need to look any further.
1762 // Skip over DBG_VALUE.
1763 while (Iter != E && Iter->isDebugValue())
1767 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1770 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1771 SE = MBB.succ_end(); SI != SE; ++SI)
1772 if ((*SI)->isLiveIn(X86::EFLAGS))
1777 MachineBasicBlock::iterator B = MBB.begin();
1779 for (unsigned i = 0; i < 4; ++i) {
1780 // If we make it to the beginning of the block, it's safe to clobber
1781 // EFLAGS iff EFLAGS is not live-in.
1783 return !MBB.isLiveIn(X86::EFLAGS);
1786 // Skip over DBG_VALUE.
1787 while (Iter != B && Iter->isDebugValue())
1790 bool SawKill = false;
1791 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1792 MachineOperand &MO = Iter->getOperand(j);
1793 // A register mask may clobber EFLAGS, but we should still look for a
1795 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1797 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1798 if (MO.isDef()) return MO.isDead();
1799 if (MO.isKill()) SawKill = true;
1804 // This instruction kills EFLAGS and doesn't redefine it, so
1805 // there's no need to look further.
1809 // Conservative answer.
1813 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1814 MachineBasicBlock::iterator I,
1815 unsigned DestReg, unsigned SubIdx,
1816 const MachineInstr *Orig,
1817 const TargetRegisterInfo &TRI) const {
1818 // MOV32r0 is implemented with a xor which clobbers condition code.
1819 // Re-materialize it as movri instructions to avoid side effects.
1820 unsigned Opc = Orig->getOpcode();
1821 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1822 DebugLoc DL = Orig->getDebugLoc();
1823 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1826 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1830 MachineInstr *NewMI = std::prev(I);
1831 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1834 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1835 /// is not marked dead.
1836 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1837 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1838 MachineOperand &MO = MI->getOperand(i);
1839 if (MO.isReg() && MO.isDef() &&
1840 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1847 /// getTruncatedShiftCount - check whether the shift count for a machine operand
1849 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1850 unsigned ShiftAmtOperandIdx) {
1851 // The shift count is six bits with the REX.W prefix and five bits without.
1852 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1853 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1854 return Imm & ShiftCountMask;
1857 /// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1858 /// can be represented by a LEA instruction.
1859 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1860 // Left shift instructions can be transformed into load-effective-address
1861 // instructions if we can encode them appropriately.
1862 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1863 // The SIB.scale field is two bits wide which means that we can encode any
1864 // shift amount less than 4.
1865 return ShAmt < 4 && ShAmt > 0;
1868 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
1869 unsigned Opc, bool AllowSP,
1870 unsigned &NewSrc, bool &isKill, bool &isUndef,
1871 MachineOperand &ImplicitOp) const {
1872 MachineFunction &MF = *MI->getParent()->getParent();
1873 const TargetRegisterClass *RC;
1875 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1877 RC = Opc != X86::LEA32r ?
1878 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1880 unsigned SrcReg = Src.getReg();
1882 // For both LEA64 and LEA32 the register already has essentially the right
1883 // type (32-bit or 64-bit) we may just need to forbid SP.
1884 if (Opc != X86::LEA64_32r) {
1886 isKill = Src.isKill();
1887 isUndef = Src.isUndef();
1889 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
1890 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1896 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1897 // another we need to add 64-bit registers to the final MI.
1898 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
1900 ImplicitOp.setImplicit();
1902 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
1903 MachineBasicBlock::LivenessQueryResult LQR =
1904 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
1907 case MachineBasicBlock::LQR_Unknown:
1908 // We can't give sane liveness flags to the instruction, abandon LEA
1911 case MachineBasicBlock::LQR_Live:
1912 isKill = MI->killsRegister(SrcReg);
1916 // The physreg itself is dead, so we have to use it as an <undef>.
1922 // Virtual register of the wrong class, we have to create a temporary 64-bit
1923 // vreg to feed into the LEA.
1924 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1925 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1926 get(TargetOpcode::COPY))
1927 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1930 // Which is obviously going to be dead after we're done with it.
1935 // We've set all the parameters without issue.
1939 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1940 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1941 /// to a 32-bit superregister and then truncating back down to a 16-bit
1944 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1945 MachineFunction::iterator &MFI,
1946 MachineBasicBlock::iterator &MBBI,
1947 LiveVariables *LV) const {
1948 MachineInstr *MI = MBBI;
1949 unsigned Dest = MI->getOperand(0).getReg();
1950 unsigned Src = MI->getOperand(1).getReg();
1951 bool isDead = MI->getOperand(0).isDead();
1952 bool isKill = MI->getOperand(1).isKill();
1954 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1955 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1956 unsigned Opc, leaInReg;
1957 if (TM.getSubtarget<X86Subtarget>().is64Bit()) {
1958 Opc = X86::LEA64_32r;
1959 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1962 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1965 // Build and insert into an implicit UNDEF value. This is OK because
1966 // well be shifting and then extracting the lower 16-bits.
1967 // This has the potential to cause partial register stall. e.g.
1968 // movw (%rbp,%rcx,2), %dx
1969 // leal -65(%rdx), %esi
1970 // But testing has shown this *does* help performance in 64-bit mode (at
1971 // least on modern x86 machines).
1972 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1973 MachineInstr *InsMI =
1974 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1975 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1976 .addReg(Src, getKillRegState(isKill));
1978 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1979 get(Opc), leaOutReg);
1981 default: llvm_unreachable("Unreachable!");
1982 case X86::SHL16ri: {
1983 unsigned ShAmt = MI->getOperand(2).getImm();
1984 MIB.addReg(0).addImm(1 << ShAmt)
1985 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1989 case X86::INC64_16r:
1990 addRegOffset(MIB, leaInReg, true, 1);
1993 case X86::DEC64_16r:
1994 addRegOffset(MIB, leaInReg, true, -1);
1998 case X86::ADD16ri_DB:
1999 case X86::ADD16ri8_DB:
2000 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
2003 case X86::ADD16rr_DB: {
2004 unsigned Src2 = MI->getOperand(2).getReg();
2005 bool isKill2 = MI->getOperand(2).isKill();
2006 unsigned leaInReg2 = 0;
2007 MachineInstr *InsMI2 = nullptr;
2009 // ADD16rr %reg1028<kill>, %reg1028
2010 // just a single insert_subreg.
2011 addRegReg(MIB, leaInReg, true, leaInReg, false);
2013 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2014 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2016 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2017 // Build and insert into an implicit UNDEF value. This is OK because
2018 // well be shifting and then extracting the lower 16-bits.
2019 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2021 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2022 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2023 .addReg(Src2, getKillRegState(isKill2));
2024 addRegReg(MIB, leaInReg, true, leaInReg2, true);
2026 if (LV && isKill2 && InsMI2)
2027 LV->replaceKillInstruction(Src2, MI, InsMI2);
2032 MachineInstr *NewMI = MIB;
2033 MachineInstr *ExtMI =
2034 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2035 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2036 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2039 // Update live variables
2040 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2041 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2043 LV->replaceKillInstruction(Src, MI, InsMI);
2045 LV->replaceKillInstruction(Dest, MI, ExtMI);
2051 /// convertToThreeAddress - This method must be implemented by targets that
2052 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
2053 /// may be able to convert a two-address instruction into a true
2054 /// three-address instruction on demand. This allows the X86 target (for
2055 /// example) to convert ADD and SHL instructions into LEA instructions if they
2056 /// would require register copies due to two-addressness.
2058 /// This method returns a null pointer if the transformation cannot be
2059 /// performed, otherwise it returns the new instruction.
2062 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2063 MachineBasicBlock::iterator &MBBI,
2064 LiveVariables *LV) const {
2065 MachineInstr *MI = MBBI;
2067 // The following opcodes also sets the condition code register(s). Only
2068 // convert them to equivalent lea if the condition code register def's
2070 if (hasLiveCondCodeDef(MI))
2073 MachineFunction &MF = *MI->getParent()->getParent();
2074 // All instructions input are two-addr instructions. Get the known operands.
2075 const MachineOperand &Dest = MI->getOperand(0);
2076 const MachineOperand &Src = MI->getOperand(1);
2078 MachineInstr *NewMI = nullptr;
2079 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
2080 // we have better subtarget support, enable the 16-bit LEA generation here.
2081 // 16-bit LEA is also slow on Core2.
2082 bool DisableLEA16 = true;
2083 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2085 unsigned MIOpc = MI->getOpcode();
2087 case X86::SHUFPSrri: {
2088 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
2089 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return nullptr;
2091 unsigned B = MI->getOperand(1).getReg();
2092 unsigned C = MI->getOperand(2).getReg();
2093 if (B != C) return nullptr;
2094 unsigned M = MI->getOperand(3).getImm();
2095 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
2096 .addOperand(Dest).addOperand(Src).addImm(M);
2099 case X86::SHUFPDrri: {
2100 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
2101 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return nullptr;
2103 unsigned B = MI->getOperand(1).getReg();
2104 unsigned C = MI->getOperand(2).getReg();
2105 if (B != C) return nullptr;
2106 unsigned M = MI->getOperand(3).getImm();
2108 // Convert to PSHUFD mask.
2109 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
2111 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
2112 .addOperand(Dest).addOperand(Src).addImm(M);
2115 case X86::SHL64ri: {
2116 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2117 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2118 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2120 // LEA can't handle RSP.
2121 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2122 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2123 &X86::GR64_NOSPRegClass))
2126 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2128 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2131 case X86::SHL32ri: {
2132 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2133 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2134 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2136 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2138 // LEA can't handle ESP.
2139 bool isKill, isUndef;
2141 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2142 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2143 SrcReg, isKill, isUndef, ImplicitOp))
2146 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2148 .addReg(0).addImm(1 << ShAmt)
2149 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2150 .addImm(0).addReg(0);
2151 if (ImplicitOp.getReg() != 0)
2152 MIB.addOperand(ImplicitOp);
2157 case X86::SHL16ri: {
2158 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2159 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2160 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2163 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : nullptr;
2164 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2166 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2172 default: return nullptr;
2175 case X86::INC64_32r: {
2176 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2177 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2178 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2179 bool isKill, isUndef;
2181 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2182 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2183 SrcReg, isKill, isUndef, ImplicitOp))
2186 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2188 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2189 if (ImplicitOp.getReg() != 0)
2190 MIB.addOperand(ImplicitOp);
2192 NewMI = addOffset(MIB, 1);
2196 case X86::INC64_16r:
2198 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2200 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2201 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2202 .addOperand(Dest).addOperand(Src), 1);
2206 case X86::DEC64_32r: {
2207 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2208 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2209 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2211 bool isKill, isUndef;
2213 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2214 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2215 SrcReg, isKill, isUndef, ImplicitOp))
2218 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2220 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2221 if (ImplicitOp.getReg() != 0)
2222 MIB.addOperand(ImplicitOp);
2224 NewMI = addOffset(MIB, -1);
2229 case X86::DEC64_16r:
2231 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2233 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2234 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2235 .addOperand(Dest).addOperand(Src), -1);
2238 case X86::ADD64rr_DB:
2240 case X86::ADD32rr_DB: {
2241 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2243 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2246 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2248 bool isKill, isUndef;
2250 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2251 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2252 SrcReg, isKill, isUndef, ImplicitOp))
2255 const MachineOperand &Src2 = MI->getOperand(2);
2256 bool isKill2, isUndef2;
2258 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2259 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2260 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2263 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2265 if (ImplicitOp.getReg() != 0)
2266 MIB.addOperand(ImplicitOp);
2267 if (ImplicitOp2.getReg() != 0)
2268 MIB.addOperand(ImplicitOp2);
2270 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2272 // Preserve undefness of the operands.
2273 NewMI->getOperand(1).setIsUndef(isUndef);
2274 NewMI->getOperand(3).setIsUndef(isUndef2);
2276 if (LV && Src2.isKill())
2277 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2281 case X86::ADD16rr_DB: {
2283 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2285 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2286 unsigned Src2 = MI->getOperand(2).getReg();
2287 bool isKill2 = MI->getOperand(2).isKill();
2288 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2290 Src.getReg(), Src.isKill(), Src2, isKill2);
2292 // Preserve undefness of the operands.
2293 bool isUndef = MI->getOperand(1).isUndef();
2294 bool isUndef2 = MI->getOperand(2).isUndef();
2295 NewMI->getOperand(1).setIsUndef(isUndef);
2296 NewMI->getOperand(3).setIsUndef(isUndef2);
2299 LV->replaceKillInstruction(Src2, MI, NewMI);
2302 case X86::ADD64ri32:
2304 case X86::ADD64ri32_DB:
2305 case X86::ADD64ri8_DB:
2306 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2307 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2308 .addOperand(Dest).addOperand(Src),
2309 MI->getOperand(2).getImm());
2313 case X86::ADD32ri_DB:
2314 case X86::ADD32ri8_DB: {
2315 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2316 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2318 bool isKill, isUndef;
2320 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2321 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2322 SrcReg, isKill, isUndef, ImplicitOp))
2325 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2327 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2328 if (ImplicitOp.getReg() != 0)
2329 MIB.addOperand(ImplicitOp);
2331 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2336 case X86::ADD16ri_DB:
2337 case X86::ADD16ri8_DB:
2339 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2341 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2342 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2343 .addOperand(Dest).addOperand(Src),
2344 MI->getOperand(2).getImm());
2350 if (!NewMI) return nullptr;
2352 if (LV) { // Update live variables
2354 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2356 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2359 MFI->insert(MBBI, NewMI); // Insert the new inst
2363 /// commuteInstruction - We have a few instructions that must be hacked on to
2367 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2368 switch (MI->getOpcode()) {
2369 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2370 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2371 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2372 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2373 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2374 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2377 switch (MI->getOpcode()) {
2378 default: llvm_unreachable("Unreachable!");
2379 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2380 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2381 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2382 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2383 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2384 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2386 unsigned Amt = MI->getOperand(3).getImm();
2388 MachineFunction &MF = *MI->getParent()->getParent();
2389 MI = MF.CloneMachineInstr(MI);
2392 MI->setDesc(get(Opc));
2393 MI->getOperand(3).setImm(Size-Amt);
2394 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2396 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2397 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2398 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2399 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2400 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2401 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2402 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2403 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2404 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2405 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2406 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2407 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2408 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2409 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2410 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2411 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2413 switch (MI->getOpcode()) {
2414 default: llvm_unreachable("Unreachable!");
2415 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2416 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2417 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2418 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2419 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2420 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2421 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2422 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2423 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2424 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2425 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2426 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2427 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2428 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2429 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2430 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2431 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2432 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2433 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2434 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2435 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2436 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2437 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2438 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2439 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2440 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2441 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2442 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2443 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2444 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2445 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2446 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2447 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2448 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2449 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2450 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2451 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2452 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2453 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2454 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2455 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2456 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2457 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2458 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2459 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2460 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2461 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2462 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2465 MachineFunction &MF = *MI->getParent()->getParent();
2466 MI = MF.CloneMachineInstr(MI);
2469 MI->setDesc(get(Opc));
2470 // Fallthrough intended.
2473 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2477 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
2478 unsigned &SrcOpIdx2) const {
2479 switch (MI->getOpcode()) {
2480 case X86::VFMADDPDr231r:
2481 case X86::VFMADDPSr231r:
2482 case X86::VFMADDSDr231r:
2483 case X86::VFMADDSSr231r:
2484 case X86::VFMSUBPDr231r:
2485 case X86::VFMSUBPSr231r:
2486 case X86::VFMSUBSDr231r:
2487 case X86::VFMSUBSSr231r:
2488 case X86::VFNMADDPDr231r:
2489 case X86::VFNMADDPSr231r:
2490 case X86::VFNMADDSDr231r:
2491 case X86::VFNMADDSSr231r:
2492 case X86::VFNMSUBPDr231r:
2493 case X86::VFNMSUBPSr231r:
2494 case X86::VFNMSUBSDr231r:
2495 case X86::VFNMSUBSSr231r:
2496 case X86::VFMADDPDr231rY:
2497 case X86::VFMADDPSr231rY:
2498 case X86::VFMSUBPDr231rY:
2499 case X86::VFMSUBPSr231rY:
2500 case X86::VFNMADDPDr231rY:
2501 case X86::VFNMADDPSr231rY:
2502 case X86::VFNMSUBPDr231rY:
2503 case X86::VFNMSUBPSr231rY:
2508 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2512 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2514 default: return X86::COND_INVALID;
2515 case X86::JE_4: return X86::COND_E;
2516 case X86::JNE_4: return X86::COND_NE;
2517 case X86::JL_4: return X86::COND_L;
2518 case X86::JLE_4: return X86::COND_LE;
2519 case X86::JG_4: return X86::COND_G;
2520 case X86::JGE_4: return X86::COND_GE;
2521 case X86::JB_4: return X86::COND_B;
2522 case X86::JBE_4: return X86::COND_BE;
2523 case X86::JA_4: return X86::COND_A;
2524 case X86::JAE_4: return X86::COND_AE;
2525 case X86::JS_4: return X86::COND_S;
2526 case X86::JNS_4: return X86::COND_NS;
2527 case X86::JP_4: return X86::COND_P;
2528 case X86::JNP_4: return X86::COND_NP;
2529 case X86::JO_4: return X86::COND_O;
2530 case X86::JNO_4: return X86::COND_NO;
2534 /// getCondFromSETOpc - return condition code of a SET opcode.
2535 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2537 default: return X86::COND_INVALID;
2538 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2539 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2540 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2541 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2542 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2543 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2544 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2545 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2546 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2547 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2548 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2549 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2550 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2551 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2552 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2553 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2557 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2558 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2560 default: return X86::COND_INVALID;
2561 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2562 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2564 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2565 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2566 return X86::COND_AE;
2567 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2568 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2570 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2571 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2572 return X86::COND_BE;
2573 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2574 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2576 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2577 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2579 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2580 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2581 return X86::COND_GE;
2582 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2583 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2585 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2586 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2587 return X86::COND_LE;
2588 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2589 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2590 return X86::COND_NE;
2591 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2592 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2593 return X86::COND_NO;
2594 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2595 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2596 return X86::COND_NP;
2597 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2598 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2599 return X86::COND_NS;
2600 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2601 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2603 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2604 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2606 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2607 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2612 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2614 default: llvm_unreachable("Illegal condition code!");
2615 case X86::COND_E: return X86::JE_4;
2616 case X86::COND_NE: return X86::JNE_4;
2617 case X86::COND_L: return X86::JL_4;
2618 case X86::COND_LE: return X86::JLE_4;
2619 case X86::COND_G: return X86::JG_4;
2620 case X86::COND_GE: return X86::JGE_4;
2621 case X86::COND_B: return X86::JB_4;
2622 case X86::COND_BE: return X86::JBE_4;
2623 case X86::COND_A: return X86::JA_4;
2624 case X86::COND_AE: return X86::JAE_4;
2625 case X86::COND_S: return X86::JS_4;
2626 case X86::COND_NS: return X86::JNS_4;
2627 case X86::COND_P: return X86::JP_4;
2628 case X86::COND_NP: return X86::JNP_4;
2629 case X86::COND_O: return X86::JO_4;
2630 case X86::COND_NO: return X86::JNO_4;
2634 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2635 /// e.g. turning COND_E to COND_NE.
2636 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2638 default: llvm_unreachable("Illegal condition code!");
2639 case X86::COND_E: return X86::COND_NE;
2640 case X86::COND_NE: return X86::COND_E;
2641 case X86::COND_L: return X86::COND_GE;
2642 case X86::COND_LE: return X86::COND_G;
2643 case X86::COND_G: return X86::COND_LE;
2644 case X86::COND_GE: return X86::COND_L;
2645 case X86::COND_B: return X86::COND_AE;
2646 case X86::COND_BE: return X86::COND_A;
2647 case X86::COND_A: return X86::COND_BE;
2648 case X86::COND_AE: return X86::COND_B;
2649 case X86::COND_S: return X86::COND_NS;
2650 case X86::COND_NS: return X86::COND_S;
2651 case X86::COND_P: return X86::COND_NP;
2652 case X86::COND_NP: return X86::COND_P;
2653 case X86::COND_O: return X86::COND_NO;
2654 case X86::COND_NO: return X86::COND_O;
2658 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2659 /// the condition code if we modify the instructions such that flags are
2661 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2663 default: return X86::COND_INVALID;
2664 case X86::COND_E: return X86::COND_E;
2665 case X86::COND_NE: return X86::COND_NE;
2666 case X86::COND_L: return X86::COND_G;
2667 case X86::COND_LE: return X86::COND_GE;
2668 case X86::COND_G: return X86::COND_L;
2669 case X86::COND_GE: return X86::COND_LE;
2670 case X86::COND_B: return X86::COND_A;
2671 case X86::COND_BE: return X86::COND_AE;
2672 case X86::COND_A: return X86::COND_B;
2673 case X86::COND_AE: return X86::COND_BE;
2677 /// getSETFromCond - Return a set opcode for the given condition and
2678 /// whether it has memory operand.
2679 static unsigned getSETFromCond(X86::CondCode CC,
2680 bool HasMemoryOperand) {
2681 static const uint16_t Opc[16][2] = {
2682 { X86::SETAr, X86::SETAm },
2683 { X86::SETAEr, X86::SETAEm },
2684 { X86::SETBr, X86::SETBm },
2685 { X86::SETBEr, X86::SETBEm },
2686 { X86::SETEr, X86::SETEm },
2687 { X86::SETGr, X86::SETGm },
2688 { X86::SETGEr, X86::SETGEm },
2689 { X86::SETLr, X86::SETLm },
2690 { X86::SETLEr, X86::SETLEm },
2691 { X86::SETNEr, X86::SETNEm },
2692 { X86::SETNOr, X86::SETNOm },
2693 { X86::SETNPr, X86::SETNPm },
2694 { X86::SETNSr, X86::SETNSm },
2695 { X86::SETOr, X86::SETOm },
2696 { X86::SETPr, X86::SETPm },
2697 { X86::SETSr, X86::SETSm }
2700 assert(CC < 16 && "Can only handle standard cond codes");
2701 return Opc[CC][HasMemoryOperand ? 1 : 0];
2704 /// getCMovFromCond - Return a cmov opcode for the given condition,
2705 /// register size in bytes, and operand type.
2706 static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2707 bool HasMemoryOperand) {
2708 static const uint16_t Opc[32][3] = {
2709 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2710 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2711 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2712 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2713 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2714 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2715 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2716 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2717 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2718 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2719 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2720 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2721 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2722 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2723 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2724 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2725 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2726 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2727 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2728 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2729 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2730 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2731 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2732 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2733 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2734 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2735 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2736 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2737 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2738 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2739 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2740 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2743 assert(CC < 16 && "Can only handle standard cond codes");
2744 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2746 default: llvm_unreachable("Illegal register size!");
2747 case 2: return Opc[Idx][0];
2748 case 4: return Opc[Idx][1];
2749 case 8: return Opc[Idx][2];
2753 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2754 if (!MI->isTerminator()) return false;
2756 // Conditional branch is a special case.
2757 if (MI->isBranch() && !MI->isBarrier())
2759 if (!MI->isPredicable())
2761 return !isPredicated(MI);
2764 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2765 MachineBasicBlock *&TBB,
2766 MachineBasicBlock *&FBB,
2767 SmallVectorImpl<MachineOperand> &Cond,
2768 bool AllowModify) const {
2769 // Start from the bottom of the block and work up, examining the
2770 // terminator instructions.
2771 MachineBasicBlock::iterator I = MBB.end();
2772 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2773 while (I != MBB.begin()) {
2775 if (I->isDebugValue())
2778 // Working from the bottom, when we see a non-terminator instruction, we're
2780 if (!isUnpredicatedTerminator(I))
2783 // A terminator that isn't a branch can't easily be handled by this
2788 // Handle unconditional branches.
2789 if (I->getOpcode() == X86::JMP_4) {
2793 TBB = I->getOperand(0).getMBB();
2797 // If the block has any instructions after a JMP, delete them.
2798 while (std::next(I) != MBB.end())
2799 std::next(I)->eraseFromParent();
2804 // Delete the JMP if it's equivalent to a fall-through.
2805 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2807 I->eraseFromParent();
2809 UnCondBrIter = MBB.end();
2813 // TBB is used to indicate the unconditional destination.
2814 TBB = I->getOperand(0).getMBB();
2818 // Handle conditional branches.
2819 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
2820 if (BranchCode == X86::COND_INVALID)
2821 return true; // Can't handle indirect branch.
2823 // Working from the bottom, handle the first conditional branch.
2825 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2826 if (AllowModify && UnCondBrIter != MBB.end() &&
2827 MBB.isLayoutSuccessor(TargetBB)) {
2828 // If we can modify the code and it ends in something like:
2836 // Then we can change this to:
2843 // Which is a bit more efficient.
2844 // We conditionally jump to the fall-through block.
2845 BranchCode = GetOppositeBranchCondition(BranchCode);
2846 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2847 MachineBasicBlock::iterator OldInst = I;
2849 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2850 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2851 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2854 OldInst->eraseFromParent();
2855 UnCondBrIter->eraseFromParent();
2857 // Restart the analysis.
2858 UnCondBrIter = MBB.end();
2864 TBB = I->getOperand(0).getMBB();
2865 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2869 // Handle subsequent conditional branches. Only handle the case where all
2870 // conditional branches branch to the same destination and their condition
2871 // opcodes fit one of the special multi-branch idioms.
2872 assert(Cond.size() == 1);
2875 // Only handle the case where all conditional branches branch to the same
2877 if (TBB != I->getOperand(0).getMBB())
2880 // If the conditions are the same, we can leave them alone.
2881 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2882 if (OldBranchCode == BranchCode)
2885 // If they differ, see if they fit one of the known patterns. Theoretically,
2886 // we could handle more patterns here, but we shouldn't expect to see them
2887 // if instruction selection has done a reasonable job.
2888 if ((OldBranchCode == X86::COND_NP &&
2889 BranchCode == X86::COND_E) ||
2890 (OldBranchCode == X86::COND_E &&
2891 BranchCode == X86::COND_NP))
2892 BranchCode = X86::COND_NP_OR_E;
2893 else if ((OldBranchCode == X86::COND_P &&
2894 BranchCode == X86::COND_NE) ||
2895 (OldBranchCode == X86::COND_NE &&
2896 BranchCode == X86::COND_P))
2897 BranchCode = X86::COND_NE_OR_P;
2901 // Update the MachineOperand.
2902 Cond[0].setImm(BranchCode);
2908 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2909 MachineBasicBlock::iterator I = MBB.end();
2912 while (I != MBB.begin()) {
2914 if (I->isDebugValue())
2916 if (I->getOpcode() != X86::JMP_4 &&
2917 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2919 // Remove the branch.
2920 I->eraseFromParent();
2929 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2930 MachineBasicBlock *FBB,
2931 const SmallVectorImpl<MachineOperand> &Cond,
2932 DebugLoc DL) const {
2933 // Shouldn't be a fall through.
2934 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2935 assert((Cond.size() == 1 || Cond.size() == 0) &&
2936 "X86 branch conditions have one component!");
2939 // Unconditional branch?
2940 assert(!FBB && "Unconditional branch with multiple successors!");
2941 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2945 // Conditional branch.
2947 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2949 case X86::COND_NP_OR_E:
2950 // Synthesize NP_OR_E with two branches.
2951 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2953 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2956 case X86::COND_NE_OR_P:
2957 // Synthesize NE_OR_P with two branches.
2958 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2960 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2964 unsigned Opc = GetCondBranchFromCond(CC);
2965 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2970 // Two-way Conditional branch. Insert the second branch.
2971 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2978 canInsertSelect(const MachineBasicBlock &MBB,
2979 const SmallVectorImpl<MachineOperand> &Cond,
2980 unsigned TrueReg, unsigned FalseReg,
2981 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2982 // Not all subtargets have cmov instructions.
2983 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2985 if (Cond.size() != 1)
2987 // We cannot do the composite conditions, at least not in SSA form.
2988 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2991 // Check register classes.
2992 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2993 const TargetRegisterClass *RC =
2994 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2998 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2999 if (X86::GR16RegClass.hasSubClassEq(RC) ||
3000 X86::GR32RegClass.hasSubClassEq(RC) ||
3001 X86::GR64RegClass.hasSubClassEq(RC)) {
3002 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
3003 // Bridge. Probably Ivy Bridge as well.
3010 // Can't do vectors.
3014 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3015 MachineBasicBlock::iterator I, DebugLoc DL,
3017 const SmallVectorImpl<MachineOperand> &Cond,
3018 unsigned TrueReg, unsigned FalseReg) const {
3019 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3020 assert(Cond.size() == 1 && "Invalid Cond array");
3021 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
3022 MRI.getRegClass(DstReg)->getSize(),
3023 false/*HasMemoryOperand*/);
3024 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
3027 /// isHReg - Test if the given register is a physical h register.
3028 static bool isHReg(unsigned Reg) {
3029 return X86::GR8_ABCD_HRegClass.contains(Reg);
3032 // Try and copy between VR128/VR64 and GR64 registers.
3033 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3034 const X86Subtarget& Subtarget) {
3037 // SrcReg(VR128) -> DestReg(GR64)
3038 // SrcReg(VR64) -> DestReg(GR64)
3039 // SrcReg(GR64) -> DestReg(VR128)
3040 // SrcReg(GR64) -> DestReg(VR64)
3042 bool HasAVX = Subtarget.hasAVX();
3043 bool HasAVX512 = Subtarget.hasAVX512();
3044 if (X86::GR64RegClass.contains(DestReg)) {
3045 if (X86::VR128XRegClass.contains(SrcReg))
3046 // Copy from a VR128 register to a GR64 register.
3047 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
3049 if (X86::VR64RegClass.contains(SrcReg))
3050 // Copy from a VR64 register to a GR64 register.
3051 return X86::MOVSDto64rr;
3052 } else if (X86::GR64RegClass.contains(SrcReg)) {
3053 // Copy from a GR64 register to a VR128 register.
3054 if (X86::VR128XRegClass.contains(DestReg))
3055 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
3057 // Copy from a GR64 register to a VR64 register.
3058 if (X86::VR64RegClass.contains(DestReg))
3059 return X86::MOV64toSDrr;
3062 // SrcReg(FR32) -> DestReg(GR32)
3063 // SrcReg(GR32) -> DestReg(FR32)
3065 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3066 // Copy from a FR32 register to a GR32 register.
3067 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3069 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3070 // Copy from a GR32 register to a FR32 register.
3071 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3075 inline static bool MaskRegClassContains(unsigned Reg) {
3076 return X86::VK8RegClass.contains(Reg) ||
3077 X86::VK16RegClass.contains(Reg) ||
3078 X86::VK1RegClass.contains(Reg);
3081 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3082 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3083 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3084 X86::VR512RegClass.contains(DestReg, SrcReg)) {
3085 DestReg = get512BitSuperRegister(DestReg);
3086 SrcReg = get512BitSuperRegister(SrcReg);
3087 return X86::VMOVAPSZrr;
3089 if (MaskRegClassContains(DestReg) &&
3090 MaskRegClassContains(SrcReg))
3091 return X86::KMOVWkk;
3092 if (MaskRegClassContains(DestReg) &&
3093 (X86::GR32RegClass.contains(SrcReg) ||
3094 X86::GR16RegClass.contains(SrcReg) ||
3095 X86::GR8RegClass.contains(SrcReg))) {
3096 SrcReg = getX86SubSuperRegister(SrcReg, MVT::i32);
3097 return X86::KMOVWkr;
3099 if ((X86::GR32RegClass.contains(DestReg) ||
3100 X86::GR16RegClass.contains(DestReg) ||
3101 X86::GR8RegClass.contains(DestReg)) &&
3102 MaskRegClassContains(SrcReg)) {
3103 DestReg = getX86SubSuperRegister(DestReg, MVT::i32);
3104 return X86::KMOVWrk;
3109 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3110 MachineBasicBlock::iterator MI, DebugLoc DL,
3111 unsigned DestReg, unsigned SrcReg,
3112 bool KillSrc) const {
3113 // First deal with the normal symmetric copies.
3114 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3115 bool HasAVX512 = TM.getSubtarget<X86Subtarget>().hasAVX512();
3117 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3119 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3121 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3123 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3124 // Copying to or from a physical H register on x86-64 requires a NOREX
3125 // move. Otherwise use a normal move.
3126 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3127 TM.getSubtarget<X86Subtarget>().is64Bit()) {
3128 Opc = X86::MOV8rr_NOREX;
3129 // Both operands must be encodable without an REX prefix.
3130 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3131 "8-bit H register can not be copied outside GR8_NOREX");
3135 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3136 Opc = X86::MMX_MOVQ64rr;
3138 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3139 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3140 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3141 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3142 Opc = X86::VMOVAPSYrr;
3144 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, TM.getSubtarget<X86Subtarget>());
3147 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3148 .addReg(SrcReg, getKillRegState(KillSrc));
3152 // Moving EFLAGS to / from another register requires a push and a pop.
3153 // Notice that we have to adjust the stack if we don't want to clobber the
3154 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
3155 if (SrcReg == X86::EFLAGS) {
3156 if (X86::GR64RegClass.contains(DestReg)) {
3157 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3158 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3161 if (X86::GR32RegClass.contains(DestReg)) {
3162 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3163 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3167 if (DestReg == X86::EFLAGS) {
3168 if (X86::GR64RegClass.contains(SrcReg)) {
3169 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3170 .addReg(SrcReg, getKillRegState(KillSrc));
3171 BuildMI(MBB, MI, DL, get(X86::POPF64));
3174 if (X86::GR32RegClass.contains(SrcReg)) {
3175 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3176 .addReg(SrcReg, getKillRegState(KillSrc));
3177 BuildMI(MBB, MI, DL, get(X86::POPF32));
3182 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3183 << " to " << RI.getName(DestReg) << '\n');
3184 llvm_unreachable("Cannot emit physreg copy instruction");
3187 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3188 const TargetRegisterClass *RC,
3189 bool isStackAligned,
3190 const TargetMachine &TM,
3192 if (TM.getSubtarget<X86Subtarget>().hasAVX512()) {
3193 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3194 X86::VK16RegClass.hasSubClassEq(RC))
3195 return load ? X86::KMOVWkm : X86::KMOVWmk;
3196 if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3197 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3198 if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3199 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3200 if (X86::VR512RegClass.hasSubClassEq(RC))
3201 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3204 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3205 switch (RC->getSize()) {
3207 llvm_unreachable("Unknown spill size");
3209 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3210 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3211 // Copying to or from a physical H register on x86-64 requires a NOREX
3212 // move. Otherwise use a normal move.
3213 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3214 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3215 return load ? X86::MOV8rm : X86::MOV8mr;
3217 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3218 return load ? X86::MOV16rm : X86::MOV16mr;
3220 if (X86::GR32RegClass.hasSubClassEq(RC))
3221 return load ? X86::MOV32rm : X86::MOV32mr;
3222 if (X86::FR32RegClass.hasSubClassEq(RC))
3224 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3225 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3226 if (X86::RFP32RegClass.hasSubClassEq(RC))
3227 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3228 llvm_unreachable("Unknown 4-byte regclass");
3230 if (X86::GR64RegClass.hasSubClassEq(RC))
3231 return load ? X86::MOV64rm : X86::MOV64mr;
3232 if (X86::FR64RegClass.hasSubClassEq(RC))
3234 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3235 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3236 if (X86::VR64RegClass.hasSubClassEq(RC))
3237 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3238 if (X86::RFP64RegClass.hasSubClassEq(RC))
3239 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3240 llvm_unreachable("Unknown 8-byte regclass");
3242 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3243 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3245 assert((X86::VR128RegClass.hasSubClassEq(RC) ||
3246 X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
3247 // If stack is realigned we can use aligned stores.
3250 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3251 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
3254 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3255 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3258 assert((X86::VR256RegClass.hasSubClassEq(RC) ||
3259 X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
3260 // If stack is realigned we can use aligned stores.
3262 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3264 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
3266 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3268 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3270 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3274 static unsigned getStoreRegOpcode(unsigned SrcReg,
3275 const TargetRegisterClass *RC,
3276 bool isStackAligned,
3277 TargetMachine &TM) {
3278 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
3282 static unsigned getLoadRegOpcode(unsigned DestReg,
3283 const TargetRegisterClass *RC,
3284 bool isStackAligned,
3285 const TargetMachine &TM) {
3286 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
3289 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3290 MachineBasicBlock::iterator MI,
3291 unsigned SrcReg, bool isKill, int FrameIdx,
3292 const TargetRegisterClass *RC,
3293 const TargetRegisterInfo *TRI) const {
3294 const MachineFunction &MF = *MBB.getParent();
3295 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3296 "Stack slot too small for store");
3297 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3298 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3299 RI.canRealignStack(MF);
3300 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3301 DebugLoc DL = MBB.findDebugLoc(MI);
3302 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
3303 .addReg(SrcReg, getKillRegState(isKill));
3306 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3308 SmallVectorImpl<MachineOperand> &Addr,
3309 const TargetRegisterClass *RC,
3310 MachineInstr::mmo_iterator MMOBegin,
3311 MachineInstr::mmo_iterator MMOEnd,
3312 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3313 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3314 bool isAligned = MMOBegin != MMOEnd &&
3315 (*MMOBegin)->getAlignment() >= Alignment;
3316 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3318 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3319 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3320 MIB.addOperand(Addr[i]);
3321 MIB.addReg(SrcReg, getKillRegState(isKill));
3322 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3323 NewMIs.push_back(MIB);
3327 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3328 MachineBasicBlock::iterator MI,
3329 unsigned DestReg, int FrameIdx,
3330 const TargetRegisterClass *RC,
3331 const TargetRegisterInfo *TRI) const {
3332 const MachineFunction &MF = *MBB.getParent();
3333 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3334 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3335 RI.canRealignStack(MF);
3336 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3337 DebugLoc DL = MBB.findDebugLoc(MI);
3338 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
3341 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
3342 SmallVectorImpl<MachineOperand> &Addr,
3343 const TargetRegisterClass *RC,
3344 MachineInstr::mmo_iterator MMOBegin,
3345 MachineInstr::mmo_iterator MMOEnd,
3346 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3347 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3348 bool isAligned = MMOBegin != MMOEnd &&
3349 (*MMOBegin)->getAlignment() >= Alignment;
3350 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3352 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3353 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3354 MIB.addOperand(Addr[i]);
3355 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3356 NewMIs.push_back(MIB);
3360 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3361 int &CmpMask, int &CmpValue) const {
3362 switch (MI->getOpcode()) {
3364 case X86::CMP64ri32:
3371 SrcReg = MI->getOperand(0).getReg();
3374 CmpValue = MI->getOperand(1).getImm();
3376 // A SUB can be used to perform comparison.
3381 SrcReg = MI->getOperand(1).getReg();
3390 SrcReg = MI->getOperand(1).getReg();
3391 SrcReg2 = MI->getOperand(2).getReg();
3395 case X86::SUB64ri32:
3402 SrcReg = MI->getOperand(1).getReg();
3405 CmpValue = MI->getOperand(2).getImm();
3411 SrcReg = MI->getOperand(0).getReg();
3412 SrcReg2 = MI->getOperand(1).getReg();
3420 SrcReg = MI->getOperand(0).getReg();
3421 if (MI->getOperand(1).getReg() != SrcReg) return false;
3422 // Compare against zero.
3431 /// isRedundantFlagInstr - check whether the first instruction, whose only
3432 /// purpose is to update flags, can be made redundant.
3433 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3434 /// This function can be extended later on.
3435 /// SrcReg, SrcRegs: register operands for FlagI.
3436 /// ImmValue: immediate for FlagI if it takes an immediate.
3437 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3438 unsigned SrcReg2, int ImmValue,
3440 if (((FlagI->getOpcode() == X86::CMP64rr &&
3441 OI->getOpcode() == X86::SUB64rr) ||
3442 (FlagI->getOpcode() == X86::CMP32rr &&
3443 OI->getOpcode() == X86::SUB32rr)||
3444 (FlagI->getOpcode() == X86::CMP16rr &&
3445 OI->getOpcode() == X86::SUB16rr)||
3446 (FlagI->getOpcode() == X86::CMP8rr &&
3447 OI->getOpcode() == X86::SUB8rr)) &&
3448 ((OI->getOperand(1).getReg() == SrcReg &&
3449 OI->getOperand(2).getReg() == SrcReg2) ||
3450 (OI->getOperand(1).getReg() == SrcReg2 &&
3451 OI->getOperand(2).getReg() == SrcReg)))
3454 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3455 OI->getOpcode() == X86::SUB64ri32) ||
3456 (FlagI->getOpcode() == X86::CMP64ri8 &&
3457 OI->getOpcode() == X86::SUB64ri8) ||
3458 (FlagI->getOpcode() == X86::CMP32ri &&
3459 OI->getOpcode() == X86::SUB32ri) ||
3460 (FlagI->getOpcode() == X86::CMP32ri8 &&
3461 OI->getOpcode() == X86::SUB32ri8) ||
3462 (FlagI->getOpcode() == X86::CMP16ri &&
3463 OI->getOpcode() == X86::SUB16ri) ||
3464 (FlagI->getOpcode() == X86::CMP16ri8 &&
3465 OI->getOpcode() == X86::SUB16ri8) ||
3466 (FlagI->getOpcode() == X86::CMP8ri &&
3467 OI->getOpcode() == X86::SUB8ri)) &&
3468 OI->getOperand(1).getReg() == SrcReg &&
3469 OI->getOperand(2).getImm() == ImmValue)
3474 /// isDefConvertible - check whether the definition can be converted
3475 /// to remove a comparison against zero.
3476 inline static bool isDefConvertible(MachineInstr *MI) {
3477 switch (MI->getOpcode()) {
3478 default: return false;
3480 // The shift instructions only modify ZF if their shift count is non-zero.
3481 // N.B.: The processor truncates the shift count depending on the encoding.
3482 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3483 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3484 return getTruncatedShiftCount(MI, 2) != 0;
3486 // Some left shift instructions can be turned into LEA instructions but only
3487 // if their flags aren't used. Avoid transforming such instructions.
3488 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3489 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3490 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3494 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3495 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3496 return getTruncatedShiftCount(MI, 3) != 0;
3498 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3499 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3500 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3501 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3502 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3503 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3504 case X86::DEC64_32r: case X86::DEC64_16r:
3505 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3506 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3507 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3508 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3509 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3510 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3511 case X86::INC64_32r: case X86::INC64_16r:
3512 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3513 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3514 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3515 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3516 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3517 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3518 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3519 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3520 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3521 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3522 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3523 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3524 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3525 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3526 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3527 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3528 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3529 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3530 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3531 case X86::ADC32ri: case X86::ADC32ri8:
3532 case X86::ADC32rr: case X86::ADC64ri32:
3533 case X86::ADC64ri8: case X86::ADC64rr:
3534 case X86::SBB32ri: case X86::SBB32ri8:
3535 case X86::SBB32rr: case X86::SBB64ri32:
3536 case X86::SBB64ri8: case X86::SBB64rr:
3537 case X86::ANDN32rr: case X86::ANDN32rm:
3538 case X86::ANDN64rr: case X86::ANDN64rm:
3539 case X86::BEXTR32rr: case X86::BEXTR64rr:
3540 case X86::BEXTR32rm: case X86::BEXTR64rm:
3541 case X86::BLSI32rr: case X86::BLSI32rm:
3542 case X86::BLSI64rr: case X86::BLSI64rm:
3543 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3544 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3545 case X86::BLSR32rr: case X86::BLSR32rm:
3546 case X86::BLSR64rr: case X86::BLSR64rm:
3547 case X86::BZHI32rr: case X86::BZHI32rm:
3548 case X86::BZHI64rr: case X86::BZHI64rm:
3549 case X86::LZCNT16rr: case X86::LZCNT16rm:
3550 case X86::LZCNT32rr: case X86::LZCNT32rm:
3551 case X86::LZCNT64rr: case X86::LZCNT64rm:
3552 case X86::POPCNT16rr:case X86::POPCNT16rm:
3553 case X86::POPCNT32rr:case X86::POPCNT32rm:
3554 case X86::POPCNT64rr:case X86::POPCNT64rm:
3555 case X86::TZCNT16rr: case X86::TZCNT16rm:
3556 case X86::TZCNT32rr: case X86::TZCNT32rm:
3557 case X86::TZCNT64rr: case X86::TZCNT64rm:
3562 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3563 /// operates on the same source operands and sets flags in the same way as
3564 /// Compare; remove Compare if possible.
3566 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3567 int CmpMask, int CmpValue,
3568 const MachineRegisterInfo *MRI) const {
3569 // Check whether we can replace SUB with CMP.
3570 unsigned NewOpcode = 0;
3571 switch (CmpInstr->getOpcode()) {
3573 case X86::SUB64ri32:
3588 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3590 // There is no use of the destination register, we can replace SUB with CMP.
3591 switch (CmpInstr->getOpcode()) {
3592 default: llvm_unreachable("Unreachable!");
3593 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3594 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3595 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3596 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3597 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3598 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3599 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3600 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3601 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3602 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3603 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3604 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3605 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3606 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3607 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3609 CmpInstr->setDesc(get(NewOpcode));
3610 CmpInstr->RemoveOperand(0);
3611 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3612 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3613 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3618 // Get the unique definition of SrcReg.
3619 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3620 if (!MI) return false;
3622 // CmpInstr is the first instruction of the BB.
3623 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3625 // If we are comparing against zero, check whether we can use MI to update
3626 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3627 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3628 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3629 !isDefConvertible(MI)))
3632 // We are searching for an earlier instruction that can make CmpInstr
3633 // redundant and that instruction will be saved in Sub.
3634 MachineInstr *Sub = nullptr;
3635 const TargetRegisterInfo *TRI = &getRegisterInfo();
3637 // We iterate backward, starting from the instruction before CmpInstr and
3638 // stop when reaching the definition of a source register or done with the BB.
3639 // RI points to the instruction before CmpInstr.
3640 // If the definition is in this basic block, RE points to the definition;
3641 // otherwise, RE is the rend of the basic block.
3642 MachineBasicBlock::reverse_iterator
3643 RI = MachineBasicBlock::reverse_iterator(I),
3644 RE = CmpInstr->getParent() == MI->getParent() ?
3645 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3646 CmpInstr->getParent()->rend();
3647 MachineInstr *Movr0Inst = nullptr;
3648 for (; RI != RE; ++RI) {
3649 MachineInstr *Instr = &*RI;
3650 // Check whether CmpInstr can be made redundant by the current instruction.
3652 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3657 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3658 Instr->readsRegister(X86::EFLAGS, TRI)) {
3659 // This instruction modifies or uses EFLAGS.
3661 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3662 // They are safe to move up, if the definition to EFLAGS is dead and
3663 // earlier instructions do not read or write EFLAGS.
3664 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
3665 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3670 // We can't remove CmpInstr.
3675 // Return false if no candidates exist.
3676 if (!IsCmpZero && !Sub)
3679 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3680 Sub->getOperand(2).getReg() == SrcReg);
3682 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3683 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3684 // If we are done with the basic block, we need to check whether EFLAGS is
3686 bool IsSafe = false;
3687 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3688 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3689 for (++I; I != E; ++I) {
3690 const MachineInstr &Instr = *I;
3691 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3692 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3693 // We should check the usage if this instruction uses and updates EFLAGS.
3694 if (!UseEFLAGS && ModifyEFLAGS) {
3695 // It is safe to remove CmpInstr if EFLAGS is updated again.
3699 if (!UseEFLAGS && !ModifyEFLAGS)
3702 // EFLAGS is used by this instruction.
3703 X86::CondCode OldCC;
3704 bool OpcIsSET = false;
3705 if (IsCmpZero || IsSwapped) {
3706 // We decode the condition code from opcode.
3707 if (Instr.isBranch())
3708 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3710 OldCC = getCondFromSETOpc(Instr.getOpcode());
3711 if (OldCC != X86::COND_INVALID)
3714 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3716 if (OldCC == X86::COND_INVALID) return false;
3721 case X86::COND_A: case X86::COND_AE:
3722 case X86::COND_B: case X86::COND_BE:
3723 case X86::COND_G: case X86::COND_GE:
3724 case X86::COND_L: case X86::COND_LE:
3725 case X86::COND_O: case X86::COND_NO:
3726 // CF and OF are used, we can't perform this optimization.
3729 } else if (IsSwapped) {
3730 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3731 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3732 // We swap the condition code and synthesize the new opcode.
3733 X86::CondCode NewCC = getSwappedCondition(OldCC);
3734 if (NewCC == X86::COND_INVALID) return false;
3736 // Synthesize the new opcode.
3737 bool HasMemoryOperand = Instr.hasOneMemOperand();
3739 if (Instr.isBranch())
3740 NewOpc = GetCondBranchFromCond(NewCC);
3742 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3744 unsigned DstReg = Instr.getOperand(0).getReg();
3745 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3749 // Push the MachineInstr to OpsToUpdate.
3750 // If it is safe to remove CmpInstr, the condition code of these
3751 // instructions will be modified.
3752 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3754 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3755 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3761 // If EFLAGS is not killed nor re-defined, we should check whether it is
3762 // live-out. If it is live-out, do not optimize.
3763 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3764 MachineBasicBlock *MBB = CmpInstr->getParent();
3765 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3766 SE = MBB->succ_end(); SI != SE; ++SI)
3767 if ((*SI)->isLiveIn(X86::EFLAGS))
3771 // The instruction to be updated is either Sub or MI.
3772 Sub = IsCmpZero ? MI : Sub;
3773 // Move Movr0Inst to the appropriate place before Sub.
3775 // Look backwards until we find a def that doesn't use the current EFLAGS.
3777 MachineBasicBlock::reverse_iterator
3778 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3779 InsertE = Sub->getParent()->rend();
3780 for (; InsertI != InsertE; ++InsertI) {
3781 MachineInstr *Instr = &*InsertI;
3782 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3783 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3784 Sub->getParent()->remove(Movr0Inst);
3785 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3790 if (InsertI == InsertE)
3794 // Make sure Sub instruction defines EFLAGS and mark the def live.
3795 unsigned i = 0, e = Sub->getNumOperands();
3796 for (; i != e; ++i) {
3797 MachineOperand &MO = Sub->getOperand(i);
3798 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3799 MO.setIsDead(false);
3803 assert(i != e && "Unable to locate a def EFLAGS operand");
3805 CmpInstr->eraseFromParent();
3807 // Modify the condition code of instructions in OpsToUpdate.
3808 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3809 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3813 /// optimizeLoadInstr - Try to remove the load by folding it to a register
3814 /// operand at the use. We fold the load instructions if load defines a virtual
3815 /// register, the virtual register is used once in the same BB, and the
3816 /// instructions in-between do not load or store, and have no side effects.
3817 MachineInstr* X86InstrInfo::
3818 optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3819 unsigned &FoldAsLoadDefReg,
3820 MachineInstr *&DefMI) const {
3821 if (FoldAsLoadDefReg == 0)
3823 // To be conservative, if there exists another load, clear the load candidate.
3824 if (MI->mayLoad()) {
3825 FoldAsLoadDefReg = 0;
3829 // Check whether we can move DefMI here.
3830 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3832 bool SawStore = false;
3833 if (!DefMI->isSafeToMove(this, nullptr, SawStore))
3836 // We try to commute MI if possible.
3837 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3838 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3839 // Collect information about virtual register operands of MI.
3840 unsigned SrcOperandId = 0;
3841 bool FoundSrcOperand = false;
3842 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3843 MachineOperand &MO = MI->getOperand(i);
3846 unsigned Reg = MO.getReg();
3847 if (Reg != FoldAsLoadDefReg)
3849 // Do not fold if we have a subreg use or a def or multiple uses.
3850 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3854 FoundSrcOperand = true;
3856 if (!FoundSrcOperand) return nullptr;
3858 // Check whether we can fold the def into SrcOperandId.
3859 SmallVector<unsigned, 8> Ops;
3860 Ops.push_back(SrcOperandId);
3861 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3863 FoldAsLoadDefReg = 0;
3868 // MI was changed but it didn't help, commute it back!
3869 commuteInstruction(MI, false);
3873 // Check whether we can commute MI and enable folding.
3874 if (MI->isCommutable()) {
3875 MachineInstr *NewMI = commuteInstruction(MI, false);
3876 // Unable to commute.
3877 if (!NewMI) return nullptr;
3879 // New instruction. It doesn't need to be kept.
3880 NewMI->eraseFromParent();
3888 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3889 /// instruction with two undef reads of the register being defined. This is
3890 /// used for mapping:
3893 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3895 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3896 const MCInstrDesc &Desc) {
3897 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3898 unsigned Reg = MIB->getOperand(0).getReg();
3901 // MachineInstr::addOperand() will insert explicit operands before any
3902 // implicit operands.
3903 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3904 // But we don't trust that.
3905 assert(MIB->getOperand(1).getReg() == Reg &&
3906 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3910 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3911 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3912 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3913 switch (MI->getOpcode()) {
3915 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
3917 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
3918 case X86::SETB_C16r:
3919 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
3920 case X86::SETB_C32r:
3921 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
3922 case X86::SETB_C64r:
3923 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
3927 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
3929 assert(HasAVX && "AVX not supported");
3930 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
3931 case X86::AVX512_512_SET0:
3932 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
3933 case X86::V_SETALLONES:
3934 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3935 case X86::AVX2_SETALLONES:
3936 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
3937 case X86::TEST8ri_NOREX:
3938 MI->setDesc(get(X86::TEST8ri));
3941 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
3943 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
3948 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
3949 const SmallVectorImpl<MachineOperand> &MOs,
3951 const TargetInstrInfo &TII) {
3952 // Create the base instruction with the memory operand as the first part.
3953 // Omit the implicit operands, something BuildMI can't do.
3954 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3955 MI->getDebugLoc(), true);
3956 MachineInstrBuilder MIB(MF, NewMI);
3957 unsigned NumAddrOps = MOs.size();
3958 for (unsigned i = 0; i != NumAddrOps; ++i)
3959 MIB.addOperand(MOs[i]);
3960 if (NumAddrOps < 4) // FrameIndex only
3963 // Loop over the rest of the ri operands, converting them over.
3964 unsigned NumOps = MI->getDesc().getNumOperands()-2;
3965 for (unsigned i = 0; i != NumOps; ++i) {
3966 MachineOperand &MO = MI->getOperand(i+2);
3969 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3970 MachineOperand &MO = MI->getOperand(i);
3976 static MachineInstr *FuseInst(MachineFunction &MF,
3977 unsigned Opcode, unsigned OpNo,
3978 const SmallVectorImpl<MachineOperand> &MOs,
3979 MachineInstr *MI, const TargetInstrInfo &TII) {
3980 // Omit the implicit operands, something BuildMI can't do.
3981 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3982 MI->getDebugLoc(), true);
3983 MachineInstrBuilder MIB(MF, NewMI);
3985 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3986 MachineOperand &MO = MI->getOperand(i);
3988 assert(MO.isReg() && "Expected to fold into reg operand!");
3989 unsigned NumAddrOps = MOs.size();
3990 for (unsigned i = 0; i != NumAddrOps; ++i)
3991 MIB.addOperand(MOs[i]);
3992 if (NumAddrOps < 4) // FrameIndex only
4001 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4002 const SmallVectorImpl<MachineOperand> &MOs,
4004 MachineFunction &MF = *MI->getParent()->getParent();
4005 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
4007 unsigned NumAddrOps = MOs.size();
4008 for (unsigned i = 0; i != NumAddrOps; ++i)
4009 MIB.addOperand(MOs[i]);
4010 if (NumAddrOps < 4) // FrameIndex only
4012 return MIB.addImm(0);
4016 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4017 MachineInstr *MI, unsigned i,
4018 const SmallVectorImpl<MachineOperand> &MOs,
4019 unsigned Size, unsigned Align) const {
4020 const DenseMap<unsigned,
4021 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4022 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
4023 bool isTwoAddrFold = false;
4025 // Atom favors register form of call. So, we do not fold loads into calls
4026 // when X86Subtarget is Atom.
4027 if (isCallRegIndirect &&
4028 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
4032 unsigned NumOps = MI->getDesc().getNumOperands();
4033 bool isTwoAddr = NumOps > 1 &&
4034 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4036 // FIXME: AsmPrinter doesn't know how to handle
4037 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4038 if (MI->getOpcode() == X86::ADD32ri &&
4039 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4042 MachineInstr *NewMI = nullptr;
4043 // Folding a memory location into the two-address part of a two-address
4044 // instruction is different than folding it other places. It requires
4045 // replacing the *two* registers with the memory location.
4046 if (isTwoAddr && NumOps >= 2 && i < 2 &&
4047 MI->getOperand(0).isReg() &&
4048 MI->getOperand(1).isReg() &&
4049 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
4050 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4051 isTwoAddrFold = true;
4052 } else if (i == 0) { // If operand 0
4053 if (MI->getOpcode() == X86::MOV32r0) {
4054 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
4059 OpcodeTablePtr = &RegOp2MemOpTable0;
4060 } else if (i == 1) {
4061 OpcodeTablePtr = &RegOp2MemOpTable1;
4062 } else if (i == 2) {
4063 OpcodeTablePtr = &RegOp2MemOpTable2;
4064 } else if (i == 3) {
4065 OpcodeTablePtr = &RegOp2MemOpTable3;
4068 // If table selected...
4069 if (OpcodeTablePtr) {
4070 // Find the Opcode to fuse
4071 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4072 OpcodeTablePtr->find(MI->getOpcode());
4073 if (I != OpcodeTablePtr->end()) {
4074 unsigned Opcode = I->second.first;
4075 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4076 if (Align < MinAlign)
4078 bool NarrowToMOV32rm = false;
4080 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
4081 if (Size < RCSize) {
4082 // Check if it's safe to fold the load. If the size of the object is
4083 // narrower than the load width, then it's not.
4084 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4086 // If this is a 64-bit load, but the spill slot is 32, then we can do
4087 // a 32-bit load which is implicitly zero-extended. This likely is due
4088 // to liveintervalanalysis remat'ing a load from stack slot.
4089 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4091 Opcode = X86::MOV32rm;
4092 NarrowToMOV32rm = true;
4097 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
4099 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
4101 if (NarrowToMOV32rm) {
4102 // If this is the special case where we use a MOV32rm to load a 32-bit
4103 // value and zero-extend the top bits. Change the destination register
4105 unsigned DstReg = NewMI->getOperand(0).getReg();
4106 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4107 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
4110 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4117 if (PrintFailedFusing && !MI->isCopy())
4118 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
4122 /// hasPartialRegUpdate - Return true for all instructions that only update
4123 /// the first 32 or 64-bits of the destination register and leave the rest
4124 /// unmodified. This can be used to avoid folding loads if the instructions
4125 /// only update part of the destination register, and the non-updated part is
4126 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4127 /// instructions breaks the partial register dependency and it can improve
4128 /// performance. e.g.:
4130 /// movss (%rdi), %xmm0
4131 /// cvtss2sd %xmm0, %xmm0
4134 /// cvtss2sd (%rdi), %xmm0
4136 /// FIXME: This should be turned into a TSFlags.
4138 static bool hasPartialRegUpdate(unsigned Opcode) {
4140 case X86::CVTSI2SSrr:
4141 case X86::CVTSI2SS64rr:
4142 case X86::CVTSI2SDrr:
4143 case X86::CVTSI2SD64rr:
4144 case X86::CVTSD2SSrr:
4145 case X86::Int_CVTSD2SSrr:
4146 case X86::CVTSS2SDrr:
4147 case X86::Int_CVTSS2SDrr:
4149 case X86::RCPSSr_Int:
4151 case X86::ROUNDSDr_Int:
4153 case X86::ROUNDSSr_Int:
4155 case X86::RSQRTSSr_Int:
4157 case X86::SQRTSSr_Int:
4164 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4165 /// instructions we would like before a partial register update.
4166 unsigned X86InstrInfo::
4167 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4168 const TargetRegisterInfo *TRI) const {
4169 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4172 // If MI is marked as reading Reg, the partial register update is wanted.
4173 const MachineOperand &MO = MI->getOperand(0);
4174 unsigned Reg = MO.getReg();
4175 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4176 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4179 if (MI->readsRegister(Reg, TRI))
4183 // If any of the preceding 16 instructions are reading Reg, insert a
4184 // dependency breaking instruction. The magic number is based on a few
4185 // Nehalem experiments.
4189 // Return true for any instruction the copies the high bits of the first source
4190 // operand into the unused high bits of the destination operand.
4191 static bool hasUndefRegUpdate(unsigned Opcode) {
4193 case X86::VCVTSI2SSrr:
4194 case X86::Int_VCVTSI2SSrr:
4195 case X86::VCVTSI2SS64rr:
4196 case X86::Int_VCVTSI2SS64rr:
4197 case X86::VCVTSI2SDrr:
4198 case X86::Int_VCVTSI2SDrr:
4199 case X86::VCVTSI2SD64rr:
4200 case X86::Int_VCVTSI2SD64rr:
4201 case X86::VCVTSD2SSrr:
4202 case X86::Int_VCVTSD2SSrr:
4203 case X86::VCVTSS2SDrr:
4204 case X86::Int_VCVTSS2SDrr:
4206 case X86::VROUNDSDr:
4207 case X86::VROUNDSDr_Int:
4208 case X86::VROUNDSSr:
4209 case X86::VROUNDSSr_Int:
4210 case X86::VRSQRTSSr:
4214 case X86::VCVTSD2SSZrr:
4215 case X86::VCVTSS2SDZrr:
4222 /// Inform the ExeDepsFix pass how many idle instructions we would like before
4223 /// certain undef register reads.
4225 /// This catches the VCVTSI2SD family of instructions:
4227 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
4229 /// We should to be careful *not* to catch VXOR idioms which are presumably
4230 /// handled specially in the pipeline:
4232 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
4234 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4235 /// high bits that are passed-through are not live.
4236 unsigned X86InstrInfo::
4237 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
4238 const TargetRegisterInfo *TRI) const {
4239 if (!hasUndefRegUpdate(MI->getOpcode()))
4242 // Set the OpNum parameter to the first source operand.
4245 const MachineOperand &MO = MI->getOperand(OpNum);
4246 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4247 // Use the same magic number as getPartialRegUpdateClearance.
4254 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4255 const TargetRegisterInfo *TRI) const {
4256 unsigned Reg = MI->getOperand(OpNum).getReg();
4257 // If MI kills this register, the false dependence is already broken.
4258 if (MI->killsRegister(Reg, TRI))
4260 if (X86::VR128RegClass.contains(Reg)) {
4261 // These instructions are all floating point domain, so xorps is the best
4263 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
4264 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4265 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4266 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4267 } else if (X86::VR256RegClass.contains(Reg)) {
4268 // Use vxorps to clear the full ymm register.
4269 // It wants to read and write the xmm sub-register.
4270 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4271 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4272 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4273 .addReg(Reg, RegState::ImplicitDefine);
4276 MI->addRegisterKilled(Reg, TRI, true);
4280 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
4281 const SmallVectorImpl<unsigned> &Ops,
4282 int FrameIndex) const {
4283 // Check switch flag
4284 if (NoFusing) return nullptr;
4286 // Unless optimizing for size, don't fold to avoid partial
4287 // register update stalls
4288 if (!MF.getFunction()->getAttributes().
4289 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4290 hasPartialRegUpdate(MI->getOpcode()))
4293 const MachineFrameInfo *MFI = MF.getFrameInfo();
4294 unsigned Size = MFI->getObjectSize(FrameIndex);
4295 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
4296 // If the function stack isn't realigned we don't want to fold instructions
4297 // that need increased alignment.
4298 if (!RI.needsStackRealignment(MF))
4299 Alignment = std::min(Alignment, TM.getFrameLowering()->getStackAlignment());
4300 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4301 unsigned NewOpc = 0;
4302 unsigned RCSize = 0;
4303 switch (MI->getOpcode()) {
4304 default: return nullptr;
4305 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
4306 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4307 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4308 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4310 // Check if it's safe to fold the load. If the size of the object is
4311 // narrower than the load width, then it's not.
4314 // Change to CMPXXri r, 0 first.
4315 MI->setDesc(get(NewOpc));
4316 MI->getOperand(1).ChangeToImmediate(0);
4317 } else if (Ops.size() != 1)
4320 SmallVector<MachineOperand,4> MOs;
4321 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
4322 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
4325 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4327 const SmallVectorImpl<unsigned> &Ops,
4328 MachineInstr *LoadMI) const {
4329 // If loading from a FrameIndex, fold directly from the FrameIndex.
4330 unsigned NumOps = LoadMI->getDesc().getNumOperands();
4332 if (isLoadFromStackSlot(LoadMI, FrameIndex))
4333 return foldMemoryOperandImpl(MF, MI, Ops, FrameIndex);
4335 // Check switch flag
4336 if (NoFusing) return nullptr;
4338 // Unless optimizing for size, don't fold to avoid partial
4339 // register update stalls
4340 if (!MF.getFunction()->getAttributes().
4341 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4342 hasPartialRegUpdate(MI->getOpcode()))
4345 // Determine the alignment of the load.
4346 unsigned Alignment = 0;
4347 if (LoadMI->hasOneMemOperand())
4348 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
4350 switch (LoadMI->getOpcode()) {
4351 case X86::AVX2_SETALLONES:
4356 case X86::V_SETALLONES:
4368 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4369 unsigned NewOpc = 0;
4370 switch (MI->getOpcode()) {
4371 default: return nullptr;
4372 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
4373 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4374 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4375 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
4377 // Change to CMPXXri r, 0 first.
4378 MI->setDesc(get(NewOpc));
4379 MI->getOperand(1).ChangeToImmediate(0);
4380 } else if (Ops.size() != 1)
4383 // Make sure the subregisters match.
4384 // Otherwise we risk changing the size of the load.
4385 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4388 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
4389 switch (LoadMI->getOpcode()) {
4391 case X86::V_SETALLONES:
4392 case X86::AVX2_SETALLONES:
4395 case X86::FsFLD0SS: {
4396 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
4397 // Create a constant-pool entry and operands to load from it.
4399 // Medium and large mode can't fold loads this way.
4400 if (TM.getCodeModel() != CodeModel::Small &&
4401 TM.getCodeModel() != CodeModel::Kernel)
4404 // x86-32 PIC requires a PIC base register for constant pools.
4405 unsigned PICBase = 0;
4406 if (TM.getRelocationModel() == Reloc::PIC_) {
4407 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4410 // FIXME: PICBase = getGlobalBaseReg(&MF);
4411 // This doesn't work for several reasons.
4412 // 1. GlobalBaseReg may have been spilled.
4413 // 2. It may not be live at MI.
4417 // Create a constant-pool entry.
4418 MachineConstantPool &MCP = *MF.getConstantPool();
4420 unsigned Opc = LoadMI->getOpcode();
4421 if (Opc == X86::FsFLD0SS)
4422 Ty = Type::getFloatTy(MF.getFunction()->getContext());
4423 else if (Opc == X86::FsFLD0SD)
4424 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
4425 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
4426 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
4428 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
4430 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
4431 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4432 Constant::getNullValue(Ty);
4433 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
4435 // Create operands to load from the constant pool entry.
4436 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4437 MOs.push_back(MachineOperand::CreateImm(1));
4438 MOs.push_back(MachineOperand::CreateReg(0, false));
4439 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
4440 MOs.push_back(MachineOperand::CreateReg(0, false));
4444 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4445 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4446 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4448 // These instructions only load 32 bits, we can't fold them if the
4449 // destination register is wider than 32 bits (4 bytes).
4451 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4452 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4453 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4455 // These instructions only load 64 bits, we can't fold them if the
4456 // destination register is wider than 64 bits (8 bytes).
4459 // Folding a normal load. Just copy the load's address operands.
4460 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
4461 MOs.push_back(LoadMI->getOperand(i));
4465 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
4469 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4470 const SmallVectorImpl<unsigned> &Ops) const {
4471 // Check switch flag
4472 if (NoFusing) return 0;
4474 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4475 switch (MI->getOpcode()) {
4476 default: return false;
4483 // FIXME: AsmPrinter doesn't know how to handle
4484 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4485 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4491 if (Ops.size() != 1)
4494 unsigned OpNum = Ops[0];
4495 unsigned Opc = MI->getOpcode();
4496 unsigned NumOps = MI->getDesc().getNumOperands();
4497 bool isTwoAddr = NumOps > 1 &&
4498 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4500 // Folding a memory location into the two-address part of a two-address
4501 // instruction is different than folding it other places. It requires
4502 // replacing the *two* registers with the memory location.
4503 const DenseMap<unsigned,
4504 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4505 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4506 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4507 } else if (OpNum == 0) { // If operand 0
4508 if (Opc == X86::MOV32r0)
4511 OpcodeTablePtr = &RegOp2MemOpTable0;
4512 } else if (OpNum == 1) {
4513 OpcodeTablePtr = &RegOp2MemOpTable1;
4514 } else if (OpNum == 2) {
4515 OpcodeTablePtr = &RegOp2MemOpTable2;
4516 } else if (OpNum == 3) {
4517 OpcodeTablePtr = &RegOp2MemOpTable3;
4520 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4522 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
4525 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4526 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4527 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4528 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4529 MemOp2RegOpTable.find(MI->getOpcode());
4530 if (I == MemOp2RegOpTable.end())
4532 unsigned Opc = I->second.first;
4533 unsigned Index = I->second.second & TB_INDEX_MASK;
4534 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4535 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4536 if (UnfoldLoad && !FoldedLoad)
4538 UnfoldLoad &= FoldedLoad;
4539 if (UnfoldStore && !FoldedStore)
4541 UnfoldStore &= FoldedStore;
4543 const MCInstrDesc &MCID = get(Opc);
4544 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4545 if (!MI->hasOneMemOperand() &&
4546 RC == &X86::VR128RegClass &&
4547 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4548 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4549 // conservatively assume the address is unaligned. That's bad for
4552 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4553 SmallVector<MachineOperand,2> BeforeOps;
4554 SmallVector<MachineOperand,2> AfterOps;
4555 SmallVector<MachineOperand,4> ImpOps;
4556 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4557 MachineOperand &Op = MI->getOperand(i);
4558 if (i >= Index && i < Index + X86::AddrNumOperands)
4559 AddrOps.push_back(Op);
4560 else if (Op.isReg() && Op.isImplicit())
4561 ImpOps.push_back(Op);
4563 BeforeOps.push_back(Op);
4565 AfterOps.push_back(Op);
4568 // Emit the load instruction.
4570 std::pair<MachineInstr::mmo_iterator,
4571 MachineInstr::mmo_iterator> MMOs =
4572 MF.extractLoadMemRefs(MI->memoperands_begin(),
4573 MI->memoperands_end());
4574 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4576 // Address operands cannot be marked isKill.
4577 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4578 MachineOperand &MO = NewMIs[0]->getOperand(i);
4580 MO.setIsKill(false);
4585 // Emit the data processing instruction.
4586 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4587 MachineInstrBuilder MIB(MF, DataMI);
4590 MIB.addReg(Reg, RegState::Define);
4591 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4592 MIB.addOperand(BeforeOps[i]);
4595 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4596 MIB.addOperand(AfterOps[i]);
4597 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4598 MachineOperand &MO = ImpOps[i];
4599 MIB.addReg(MO.getReg(),
4600 getDefRegState(MO.isDef()) |
4601 RegState::Implicit |
4602 getKillRegState(MO.isKill()) |
4603 getDeadRegState(MO.isDead()) |
4604 getUndefRegState(MO.isUndef()));
4606 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4607 switch (DataMI->getOpcode()) {
4609 case X86::CMP64ri32:
4616 MachineOperand &MO0 = DataMI->getOperand(0);
4617 MachineOperand &MO1 = DataMI->getOperand(1);
4618 if (MO1.getImm() == 0) {
4620 switch (DataMI->getOpcode()) {
4621 default: llvm_unreachable("Unreachable!");
4623 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
4625 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
4627 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4628 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4630 DataMI->setDesc(get(NewOpc));
4631 MO1.ChangeToRegister(MO0.getReg(), false);
4635 NewMIs.push_back(DataMI);
4637 // Emit the store instruction.
4639 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
4640 std::pair<MachineInstr::mmo_iterator,
4641 MachineInstr::mmo_iterator> MMOs =
4642 MF.extractStoreMemRefs(MI->memoperands_begin(),
4643 MI->memoperands_end());
4644 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
4651 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
4652 SmallVectorImpl<SDNode*> &NewNodes) const {
4653 if (!N->isMachineOpcode())
4656 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4657 MemOp2RegOpTable.find(N->getMachineOpcode());
4658 if (I == MemOp2RegOpTable.end())
4660 unsigned Opc = I->second.first;
4661 unsigned Index = I->second.second & TB_INDEX_MASK;
4662 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4663 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4664 const MCInstrDesc &MCID = get(Opc);
4665 MachineFunction &MF = DAG.getMachineFunction();
4666 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4667 unsigned NumDefs = MCID.NumDefs;
4668 std::vector<SDValue> AddrOps;
4669 std::vector<SDValue> BeforeOps;
4670 std::vector<SDValue> AfterOps;
4672 unsigned NumOps = N->getNumOperands();
4673 for (unsigned i = 0; i != NumOps-1; ++i) {
4674 SDValue Op = N->getOperand(i);
4675 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
4676 AddrOps.push_back(Op);
4677 else if (i < Index-NumDefs)
4678 BeforeOps.push_back(Op);
4679 else if (i > Index-NumDefs)
4680 AfterOps.push_back(Op);
4682 SDValue Chain = N->getOperand(NumOps-1);
4683 AddrOps.push_back(Chain);
4685 // Emit the load instruction.
4686 SDNode *Load = nullptr;
4688 EVT VT = *RC->vt_begin();
4689 std::pair<MachineInstr::mmo_iterator,
4690 MachineInstr::mmo_iterator> MMOs =
4691 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4692 cast<MachineSDNode>(N)->memoperands_end());
4693 if (!(*MMOs.first) &&
4694 RC == &X86::VR128RegClass &&
4695 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4696 // Do not introduce a slow unaligned load.
4698 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4699 bool isAligned = (*MMOs.first) &&
4700 (*MMOs.first)->getAlignment() >= Alignment;
4701 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4702 VT, MVT::Other, AddrOps);
4703 NewNodes.push_back(Load);
4705 // Preserve memory reference information.
4706 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4709 // Emit the data processing instruction.
4710 std::vector<EVT> VTs;
4711 const TargetRegisterClass *DstRC = nullptr;
4712 if (MCID.getNumDefs() > 0) {
4713 DstRC = getRegClass(MCID, 0, &RI, MF);
4714 VTs.push_back(*DstRC->vt_begin());
4716 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
4717 EVT VT = N->getValueType(i);
4718 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
4722 BeforeOps.push_back(SDValue(Load, 0));
4723 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
4724 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
4725 NewNodes.push_back(NewNode);
4727 // Emit the store instruction.
4730 AddrOps.push_back(SDValue(NewNode, 0));
4731 AddrOps.push_back(Chain);
4732 std::pair<MachineInstr::mmo_iterator,
4733 MachineInstr::mmo_iterator> MMOs =
4734 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4735 cast<MachineSDNode>(N)->memoperands_end());
4736 if (!(*MMOs.first) &&
4737 RC == &X86::VR128RegClass &&
4738 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4739 // Do not introduce a slow unaligned store.
4741 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4742 bool isAligned = (*MMOs.first) &&
4743 (*MMOs.first)->getAlignment() >= Alignment;
4744 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4746 dl, MVT::Other, AddrOps);
4747 NewNodes.push_back(Store);
4749 // Preserve memory reference information.
4750 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4756 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
4757 bool UnfoldLoad, bool UnfoldStore,
4758 unsigned *LoadRegIndex) const {
4759 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4760 MemOp2RegOpTable.find(Opc);
4761 if (I == MemOp2RegOpTable.end())
4763 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4764 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4765 if (UnfoldLoad && !FoldedLoad)
4767 if (UnfoldStore && !FoldedStore)
4770 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
4771 return I->second.first;
4775 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4776 int64_t &Offset1, int64_t &Offset2) const {
4777 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4779 unsigned Opc1 = Load1->getMachineOpcode();
4780 unsigned Opc2 = Load2->getMachineOpcode();
4782 default: return false;
4792 case X86::MMX_MOVD64rm:
4793 case X86::MMX_MOVQ64rm:
4794 case X86::FsMOVAPSrm:
4795 case X86::FsMOVAPDrm:
4801 // AVX load instructions
4804 case X86::FsVMOVAPSrm:
4805 case X86::FsVMOVAPDrm:
4806 case X86::VMOVAPSrm:
4807 case X86::VMOVUPSrm:
4808 case X86::VMOVAPDrm:
4809 case X86::VMOVDQArm:
4810 case X86::VMOVDQUrm:
4811 case X86::VMOVAPSYrm:
4812 case X86::VMOVUPSYrm:
4813 case X86::VMOVAPDYrm:
4814 case X86::VMOVDQAYrm:
4815 case X86::VMOVDQUYrm:
4819 default: return false;
4829 case X86::MMX_MOVD64rm:
4830 case X86::MMX_MOVQ64rm:
4831 case X86::FsMOVAPSrm:
4832 case X86::FsMOVAPDrm:
4838 // AVX load instructions
4841 case X86::FsVMOVAPSrm:
4842 case X86::FsVMOVAPDrm:
4843 case X86::VMOVAPSrm:
4844 case X86::VMOVUPSrm:
4845 case X86::VMOVAPDrm:
4846 case X86::VMOVDQArm:
4847 case X86::VMOVDQUrm:
4848 case X86::VMOVAPSYrm:
4849 case X86::VMOVUPSYrm:
4850 case X86::VMOVAPDYrm:
4851 case X86::VMOVDQAYrm:
4852 case X86::VMOVDQUYrm:
4856 // Check if chain operands and base addresses match.
4857 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4858 Load1->getOperand(5) != Load2->getOperand(5))
4860 // Segment operands should match as well.
4861 if (Load1->getOperand(4) != Load2->getOperand(4))
4863 // Scale should be 1, Index should be Reg0.
4864 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4865 Load1->getOperand(2) == Load2->getOperand(2)) {
4866 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4869 // Now let's examine the displacements.
4870 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4871 isa<ConstantSDNode>(Load2->getOperand(3))) {
4872 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4873 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4880 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4881 int64_t Offset1, int64_t Offset2,
4882 unsigned NumLoads) const {
4883 assert(Offset2 > Offset1);
4884 if ((Offset2 - Offset1) / 8 > 64)
4887 unsigned Opc1 = Load1->getMachineOpcode();
4888 unsigned Opc2 = Load2->getMachineOpcode();
4890 return false; // FIXME: overly conservative?
4897 case X86::MMX_MOVD64rm:
4898 case X86::MMX_MOVQ64rm:
4902 EVT VT = Load1->getValueType(0);
4903 switch (VT.getSimpleVT().SimpleTy) {
4905 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4906 // have 16 of them to play with.
4907 if (TM.getSubtargetImpl()->is64Bit()) {
4910 } else if (NumLoads) {
4928 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
4929 MachineInstr *Second) const {
4930 // Check if this processor supports macro-fusion. Since this is a minor
4931 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
4932 // proxy for SandyBridge+.
4933 if (!TM.getSubtarget<X86Subtarget>().hasAVX())
4942 switch(Second->getOpcode()) {
4965 FuseKind = FuseTest;
4968 switch (First->getOpcode()) {
4978 case X86::TEST32i32:
4979 case X86::TEST64i32:
4980 case X86::TEST64ri32:
4996 case X86::AND64ri32:
5016 case X86::CMP64ri32:
5027 case X86::ADD16ri8_DB:
5028 case X86::ADD16ri_DB:
5031 case X86::ADD16rr_DB:
5035 case X86::ADD32ri8_DB:
5036 case X86::ADD32ri_DB:
5039 case X86::ADD32rr_DB:
5041 case X86::ADD64ri32:
5042 case X86::ADD64ri32_DB:
5044 case X86::ADD64ri8_DB:
5047 case X86::ADD64rr_DB:
5065 case X86::SUB64ri32:
5073 return FuseKind == FuseCmp || FuseKind == FuseInc;
5076 case X86::INC64_16r:
5077 case X86::INC64_32r:
5082 case X86::DEC64_16r:
5083 case X86::DEC64_32r:
5086 return FuseKind == FuseInc;
5091 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5092 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5093 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5094 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
5096 Cond[0].setImm(GetOppositeBranchCondition(CC));
5101 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5102 // FIXME: Return false for x87 stack register classes for now. We can't
5103 // allow any loads of these registers before FpGet_ST0_80.
5104 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
5105 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
5108 /// getGlobalBaseReg - Return a virtual register initialized with the
5109 /// the global base register value. Output instructions required to
5110 /// initialize the register in the function entry block, if necessary.
5112 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5114 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5115 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
5116 "X86-64 PIC uses RIP relative addressing");
5118 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5119 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5120 if (GlobalBaseReg != 0)
5121 return GlobalBaseReg;
5123 // Create the register. The code to initialize it is inserted
5124 // later, by the CGBR pass (below).
5125 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5126 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
5127 X86FI->setGlobalBaseReg(GlobalBaseReg);
5128 return GlobalBaseReg;
5131 // These are the replaceable SSE instructions. Some of these have Int variants
5132 // that we don't include here. We don't want to replace instructions selected
5134 static const uint16_t ReplaceableInstrs[][3] = {
5135 //PackedSingle PackedDouble PackedInt
5136 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
5137 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
5138 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
5139 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
5140 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
5141 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
5142 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
5143 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
5144 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
5145 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
5146 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
5147 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
5148 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
5149 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
5150 // AVX 128-bit support
5151 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
5152 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
5153 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
5154 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5155 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5156 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5157 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5158 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5159 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5160 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5161 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5162 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
5163 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
5164 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
5165 // AVX 256-bit support
5166 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
5167 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
5168 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
5169 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
5170 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
5171 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
5174 static const uint16_t ReplaceableInstrsAVX2[][3] = {
5175 //PackedSingle PackedDouble PackedInt
5176 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
5177 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
5178 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
5179 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
5180 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
5181 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
5182 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
5183 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
5184 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
5185 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
5186 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
5187 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
5188 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
5189 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
5190 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
5191 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
5192 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
5193 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
5194 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
5195 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
5198 // FIXME: Some shuffle and unpack instructions have equivalents in different
5199 // domains, but they require a bit more work than just switching opcodes.
5201 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
5202 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
5203 if (ReplaceableInstrs[i][domain-1] == opcode)
5204 return ReplaceableInstrs[i];
5208 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
5209 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
5210 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
5211 return ReplaceableInstrsAVX2[i];
5215 std::pair<uint16_t, uint16_t>
5216 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
5217 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5218 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
5219 uint16_t validDomains = 0;
5220 if (domain && lookup(MI->getOpcode(), domain))
5222 else if (domain && lookupAVX2(MI->getOpcode(), domain))
5223 validDomains = hasAVX2 ? 0xe : 0x6;
5224 return std::make_pair(domain, validDomains);
5227 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
5228 assert(Domain>0 && Domain<4 && "Invalid execution domain");
5229 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5230 assert(dom && "Not an SSE instruction");
5231 const uint16_t *table = lookup(MI->getOpcode(), dom);
5232 if (!table) { // try the other table
5233 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
5234 "256-bit vector operations only available in AVX2");
5235 table = lookupAVX2(MI->getOpcode(), dom);
5237 assert(table && "Cannot change domain");
5238 MI->setDesc(get(table[Domain-1]));
5241 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5242 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5243 NopInst.setOpcode(X86::NOOP);
5246 bool X86InstrInfo::isHighLatencyDef(int opc) const {
5248 default: return false;
5250 case X86::DIVSDrm_Int:
5252 case X86::DIVSDrr_Int:
5254 case X86::DIVSSrm_Int:
5256 case X86::DIVSSrr_Int:
5262 case X86::SQRTSDm_Int:
5264 case X86::SQRTSDr_Int:
5266 case X86::SQRTSSm_Int:
5268 case X86::SQRTSSr_Int:
5269 // AVX instructions with high latency
5271 case X86::VDIVSDrm_Int:
5273 case X86::VDIVSDrr_Int:
5275 case X86::VDIVSSrm_Int:
5277 case X86::VDIVSSrr_Int:
5283 case X86::VSQRTSDm_Int:
5286 case X86::VSQRTSSm_Int:
5288 case X86::VSQRTPDZrm:
5289 case X86::VSQRTPDZrr:
5290 case X86::VSQRTPSZrm:
5291 case X86::VSQRTPSZrr:
5292 case X86::VSQRTSDZm:
5293 case X86::VSQRTSDZm_Int:
5294 case X86::VSQRTSDZr:
5295 case X86::VSQRTSSZm_Int:
5296 case X86::VSQRTSSZr:
5297 case X86::VSQRTSSZm:
5298 case X86::VDIVSDZrm:
5299 case X86::VDIVSDZrr:
5300 case X86::VDIVSSZrm:
5301 case X86::VDIVSSZrr:
5303 case X86::VGATHERQPSZrm:
5304 case X86::VGATHERQPDZrm:
5305 case X86::VGATHERDPDZrm:
5306 case X86::VGATHERDPSZrm:
5307 case X86::VPGATHERQDZrm:
5308 case X86::VPGATHERQQZrm:
5309 case X86::VPGATHERDDZrm:
5310 case X86::VPGATHERDQZrm:
5311 case X86::VSCATTERQPDZmr:
5312 case X86::VSCATTERQPSZmr:
5313 case X86::VSCATTERDPDZmr:
5314 case X86::VSCATTERDPSZmr:
5315 case X86::VPSCATTERQDZmr:
5316 case X86::VPSCATTERQQZmr:
5317 case X86::VPSCATTERDDZmr:
5318 case X86::VPSCATTERDQZmr:
5324 hasHighOperandLatency(const InstrItineraryData *ItinData,
5325 const MachineRegisterInfo *MRI,
5326 const MachineInstr *DefMI, unsigned DefIdx,
5327 const MachineInstr *UseMI, unsigned UseIdx) const {
5328 return isHighLatencyDef(DefMI->getOpcode());
5332 /// CGBR - Create Global Base Reg pass. This initializes the PIC
5333 /// global base register for x86-32.
5334 struct CGBR : public MachineFunctionPass {
5336 CGBR() : MachineFunctionPass(ID) {}
5338 bool runOnMachineFunction(MachineFunction &MF) override {
5339 const X86TargetMachine *TM =
5340 static_cast<const X86TargetMachine *>(&MF.getTarget());
5342 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
5343 "X86-64 PIC uses RIP relative addressing");
5345 // Only emit a global base reg in PIC mode.
5346 if (TM->getRelocationModel() != Reloc::PIC_)
5349 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5350 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5352 // If we didn't need a GlobalBaseReg, don't insert code.
5353 if (GlobalBaseReg == 0)
5356 // Insert the set of GlobalBaseReg into the first MBB of the function
5357 MachineBasicBlock &FirstMBB = MF.front();
5358 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5359 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5360 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5361 const X86InstrInfo *TII = TM->getInstrInfo();
5364 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
5365 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
5369 // Operand of MovePCtoStack is completely ignored by asm printer. It's
5370 // only used in JIT code emission as displacement to pc.
5371 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
5373 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5374 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5375 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
5376 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5377 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5378 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5379 X86II::MO_GOT_ABSOLUTE_ADDRESS);
5385 const char *getPassName() const override {
5386 return "X86 PIC Global Base Reg Initialization";
5389 void getAnalysisUsage(AnalysisUsage &AU) const override {
5390 AU.setPreservesCFG();
5391 MachineFunctionPass::getAnalysisUsage(AU);
5398 llvm::createGlobalBaseRegPass() { return new CGBR(); }
5401 struct LDTLSCleanup : public MachineFunctionPass {
5403 LDTLSCleanup() : MachineFunctionPass(ID) {}
5405 bool runOnMachineFunction(MachineFunction &MF) override {
5406 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5407 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5408 // No point folding accesses if there isn't at least two.
5412 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5413 return VisitNode(DT->getRootNode(), 0);
5416 // Visit the dominator subtree rooted at Node in pre-order.
5417 // If TLSBaseAddrReg is non-null, then use that to replace any
5418 // TLS_base_addr instructions. Otherwise, create the register
5419 // when the first such instruction is seen, and then use it
5420 // as we encounter more instructions.
5421 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5422 MachineBasicBlock *BB = Node->getBlock();
5423 bool Changed = false;
5425 // Traverse the current block.
5426 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5428 switch (I->getOpcode()) {
5429 case X86::TLS_base_addr32:
5430 case X86::TLS_base_addr64:
5432 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5434 I = SetRegister(I, &TLSBaseAddrReg);
5442 // Visit the children of this block in the dominator tree.
5443 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5445 Changed |= VisitNode(*I, TLSBaseAddrReg);
5451 // Replace the TLS_base_addr instruction I with a copy from
5452 // TLSBaseAddrReg, returning the new instruction.
5453 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5454 unsigned TLSBaseAddrReg) {
5455 MachineFunction *MF = I->getParent()->getParent();
5456 const X86TargetMachine *TM =
5457 static_cast<const X86TargetMachine *>(&MF->getTarget());
5458 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5459 const X86InstrInfo *TII = TM->getInstrInfo();
5461 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5462 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5463 TII->get(TargetOpcode::COPY),
5464 is64Bit ? X86::RAX : X86::EAX)
5465 .addReg(TLSBaseAddrReg);
5467 // Erase the TLS_base_addr instruction.
5468 I->eraseFromParent();
5473 // Create a virtal register in *TLSBaseAddrReg, and populate it by
5474 // inserting a copy instruction after I. Returns the new instruction.
5475 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5476 MachineFunction *MF = I->getParent()->getParent();
5477 const X86TargetMachine *TM =
5478 static_cast<const X86TargetMachine *>(&MF->getTarget());
5479 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5480 const X86InstrInfo *TII = TM->getInstrInfo();
5482 // Create a virtual register for the TLS base address.
5483 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5484 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5485 ? &X86::GR64RegClass
5486 : &X86::GR32RegClass);
5488 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5489 MachineInstr *Next = I->getNextNode();
5490 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5491 TII->get(TargetOpcode::COPY),
5493 .addReg(is64Bit ? X86::RAX : X86::EAX);
5498 const char *getPassName() const override {
5499 return "Local Dynamic TLS Access Clean-up";
5502 void getAnalysisUsage(AnalysisUsage &AU) const override {
5503 AU.setPreservesCFG();
5504 AU.addRequired<MachineDominatorTree>();
5505 MachineFunctionPass::getAnalysisUsage(AU);
5510 char LDTLSCleanup::ID = 0;
5512 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
projects / oota-llvm.git / blob