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/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/MC/MCAsmInfo.h"
32 #include "llvm/MC/MCExpr.h"
33 #include "llvm/MC/MCInst.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetOptions.h"
43 #define DEBUG_TYPE "x86-instr-info"
45 #define GET_INSTRINFO_CTOR_DTOR
46 #include "X86GenInstrInfo.inc"
49 NoFusing("disable-spill-fusing",
50 cl::desc("Disable fusing of spill code into instructions"));
52 PrintFailedFusing("print-failed-fuse-candidates",
53 cl::desc("Print instructions that the allocator wants to"
54 " fuse, but the X86 backend currently can't"),
57 ReMatPICStubLoad("remat-pic-stub-load",
58 cl::desc("Re-materialize load from stub in PIC mode"),
59 cl::init(false), cl::Hidden);
62 // Select which memory operand is being unfolded.
63 // (stored in bits 0 - 3)
71 // Do not insert the reverse map (MemOp -> RegOp) into the table.
72 // This may be needed because there is a many -> one mapping.
73 TB_NO_REVERSE = 1 << 4,
75 // Do not insert the forward map (RegOp -> MemOp) into the table.
76 // This is needed for Native Client, which prohibits branch
77 // instructions from using a memory operand.
78 TB_NO_FORWARD = 1 << 5,
80 TB_FOLDED_LOAD = 1 << 6,
81 TB_FOLDED_STORE = 1 << 7,
83 // Minimum alignment required for load/store.
84 // Used for RegOp->MemOp conversion.
85 // (stored in bits 8 - 15)
87 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
88 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
89 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
90 TB_ALIGN_64 = 64 << TB_ALIGN_SHIFT,
91 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
94 struct X86OpTblEntry {
100 // Pin the vtable to this file.
101 void X86InstrInfo::anchor() {}
103 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
105 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64 : X86::ADJCALLSTACKDOWN32),
106 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64 : X86::ADJCALLSTACKUP32)),
107 Subtarget(STI), RI(STI) {
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::DEC64r, X86::DEC64m, 0 },
150 { X86::DEC8r, X86::DEC8m, 0 },
151 { X86::INC16r, X86::INC16m, 0 },
152 { X86::INC32r, X86::INC32m, 0 },
153 { X86::INC64r, X86::INC64m, 0 },
154 { X86::INC8r, X86::INC8m, 0 },
155 { X86::NEG16r, X86::NEG16m, 0 },
156 { X86::NEG32r, X86::NEG32m, 0 },
157 { X86::NEG64r, X86::NEG64m, 0 },
158 { X86::NEG8r, X86::NEG8m, 0 },
159 { X86::NOT16r, X86::NOT16m, 0 },
160 { X86::NOT32r, X86::NOT32m, 0 },
161 { X86::NOT64r, X86::NOT64m, 0 },
162 { X86::NOT8r, X86::NOT8m, 0 },
163 { X86::OR16ri, X86::OR16mi, 0 },
164 { X86::OR16ri8, X86::OR16mi8, 0 },
165 { X86::OR16rr, X86::OR16mr, 0 },
166 { X86::OR32ri, X86::OR32mi, 0 },
167 { X86::OR32ri8, X86::OR32mi8, 0 },
168 { X86::OR32rr, X86::OR32mr, 0 },
169 { X86::OR64ri32, X86::OR64mi32, 0 },
170 { X86::OR64ri8, X86::OR64mi8, 0 },
171 { X86::OR64rr, X86::OR64mr, 0 },
172 { X86::OR8ri, X86::OR8mi, 0 },
173 { X86::OR8rr, X86::OR8mr, 0 },
174 { X86::ROL16r1, X86::ROL16m1, 0 },
175 { X86::ROL16rCL, X86::ROL16mCL, 0 },
176 { X86::ROL16ri, X86::ROL16mi, 0 },
177 { X86::ROL32r1, X86::ROL32m1, 0 },
178 { X86::ROL32rCL, X86::ROL32mCL, 0 },
179 { X86::ROL32ri, X86::ROL32mi, 0 },
180 { X86::ROL64r1, X86::ROL64m1, 0 },
181 { X86::ROL64rCL, X86::ROL64mCL, 0 },
182 { X86::ROL64ri, X86::ROL64mi, 0 },
183 { X86::ROL8r1, X86::ROL8m1, 0 },
184 { X86::ROL8rCL, X86::ROL8mCL, 0 },
185 { X86::ROL8ri, X86::ROL8mi, 0 },
186 { X86::ROR16r1, X86::ROR16m1, 0 },
187 { X86::ROR16rCL, X86::ROR16mCL, 0 },
188 { X86::ROR16ri, X86::ROR16mi, 0 },
189 { X86::ROR32r1, X86::ROR32m1, 0 },
190 { X86::ROR32rCL, X86::ROR32mCL, 0 },
191 { X86::ROR32ri, X86::ROR32mi, 0 },
192 { X86::ROR64r1, X86::ROR64m1, 0 },
193 { X86::ROR64rCL, X86::ROR64mCL, 0 },
194 { X86::ROR64ri, X86::ROR64mi, 0 },
195 { X86::ROR8r1, X86::ROR8m1, 0 },
196 { X86::ROR8rCL, X86::ROR8mCL, 0 },
197 { X86::ROR8ri, X86::ROR8mi, 0 },
198 { X86::SAR16r1, X86::SAR16m1, 0 },
199 { X86::SAR16rCL, X86::SAR16mCL, 0 },
200 { X86::SAR16ri, X86::SAR16mi, 0 },
201 { X86::SAR32r1, X86::SAR32m1, 0 },
202 { X86::SAR32rCL, X86::SAR32mCL, 0 },
203 { X86::SAR32ri, X86::SAR32mi, 0 },
204 { X86::SAR64r1, X86::SAR64m1, 0 },
205 { X86::SAR64rCL, X86::SAR64mCL, 0 },
206 { X86::SAR64ri, X86::SAR64mi, 0 },
207 { X86::SAR8r1, X86::SAR8m1, 0 },
208 { X86::SAR8rCL, X86::SAR8mCL, 0 },
209 { X86::SAR8ri, X86::SAR8mi, 0 },
210 { X86::SBB32ri, X86::SBB32mi, 0 },
211 { X86::SBB32ri8, X86::SBB32mi8, 0 },
212 { X86::SBB32rr, X86::SBB32mr, 0 },
213 { X86::SBB64ri32, X86::SBB64mi32, 0 },
214 { X86::SBB64ri8, X86::SBB64mi8, 0 },
215 { X86::SBB64rr, X86::SBB64mr, 0 },
216 { X86::SHL16rCL, X86::SHL16mCL, 0 },
217 { X86::SHL16ri, X86::SHL16mi, 0 },
218 { X86::SHL32rCL, X86::SHL32mCL, 0 },
219 { X86::SHL32ri, X86::SHL32mi, 0 },
220 { X86::SHL64rCL, X86::SHL64mCL, 0 },
221 { X86::SHL64ri, X86::SHL64mi, 0 },
222 { X86::SHL8rCL, X86::SHL8mCL, 0 },
223 { X86::SHL8ri, X86::SHL8mi, 0 },
224 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
225 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
226 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
227 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
228 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
229 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
230 { X86::SHR16r1, X86::SHR16m1, 0 },
231 { X86::SHR16rCL, X86::SHR16mCL, 0 },
232 { X86::SHR16ri, X86::SHR16mi, 0 },
233 { X86::SHR32r1, X86::SHR32m1, 0 },
234 { X86::SHR32rCL, X86::SHR32mCL, 0 },
235 { X86::SHR32ri, X86::SHR32mi, 0 },
236 { X86::SHR64r1, X86::SHR64m1, 0 },
237 { X86::SHR64rCL, X86::SHR64mCL, 0 },
238 { X86::SHR64ri, X86::SHR64mi, 0 },
239 { X86::SHR8r1, X86::SHR8m1, 0 },
240 { X86::SHR8rCL, X86::SHR8mCL, 0 },
241 { X86::SHR8ri, X86::SHR8mi, 0 },
242 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
243 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
244 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
245 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
246 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
247 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
248 { X86::SUB16ri, X86::SUB16mi, 0 },
249 { X86::SUB16ri8, X86::SUB16mi8, 0 },
250 { X86::SUB16rr, X86::SUB16mr, 0 },
251 { X86::SUB32ri, X86::SUB32mi, 0 },
252 { X86::SUB32ri8, X86::SUB32mi8, 0 },
253 { X86::SUB32rr, X86::SUB32mr, 0 },
254 { X86::SUB64ri32, X86::SUB64mi32, 0 },
255 { X86::SUB64ri8, X86::SUB64mi8, 0 },
256 { X86::SUB64rr, X86::SUB64mr, 0 },
257 { X86::SUB8ri, X86::SUB8mi, 0 },
258 { X86::SUB8rr, X86::SUB8mr, 0 },
259 { X86::XOR16ri, X86::XOR16mi, 0 },
260 { X86::XOR16ri8, X86::XOR16mi8, 0 },
261 { X86::XOR16rr, X86::XOR16mr, 0 },
262 { X86::XOR32ri, X86::XOR32mi, 0 },
263 { X86::XOR32ri8, X86::XOR32mi8, 0 },
264 { X86::XOR32rr, X86::XOR32mr, 0 },
265 { X86::XOR64ri32, X86::XOR64mi32, 0 },
266 { X86::XOR64ri8, X86::XOR64mi8, 0 },
267 { X86::XOR64rr, X86::XOR64mr, 0 },
268 { X86::XOR8ri, X86::XOR8mi, 0 },
269 { X86::XOR8rr, X86::XOR8mr, 0 }
272 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
273 unsigned RegOp = OpTbl2Addr[i].RegOp;
274 unsigned MemOp = OpTbl2Addr[i].MemOp;
275 unsigned Flags = OpTbl2Addr[i].Flags;
276 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
278 // Index 0, folded load and store, no alignment requirement.
279 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
282 static const X86OpTblEntry OpTbl0[] = {
283 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
284 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
285 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
286 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
287 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
288 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
289 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
290 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
291 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
292 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
293 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
294 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
295 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
296 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
297 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
298 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
299 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
300 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
301 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
302 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
303 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
304 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
305 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
306 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
307 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
308 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
309 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
310 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
311 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
312 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
313 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
314 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
315 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
316 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
317 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
318 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
319 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
320 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
321 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
322 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
323 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
325 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
326 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
327 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
328 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
329 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
330 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
331 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
332 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
333 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
334 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
335 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
336 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
337 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
338 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
339 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
340 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
341 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
342 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
343 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
344 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
345 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
346 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
347 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
348 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
349 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
350 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
351 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
352 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
353 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
354 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
355 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
356 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
357 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
358 // AVX 128-bit versions of foldable instructions
359 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
360 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
361 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
362 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
365 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
366 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
367 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
368 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
369 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
370 // AVX 256-bit foldable instructions
371 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
372 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
373 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
376 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE },
377 // AVX-512 foldable instructions
378 { X86::VMOVPDI2DIZrr, X86::VMOVPDI2DIZmr, TB_FOLDED_STORE },
379 { X86::VMOVAPDZrr, X86::VMOVAPDZmr, TB_FOLDED_STORE | TB_ALIGN_64 },
380 { X86::VMOVAPSZrr, X86::VMOVAPSZmr, TB_FOLDED_STORE | TB_ALIGN_64 },
381 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zmr, TB_FOLDED_STORE | TB_ALIGN_64 },
382 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zmr, TB_FOLDED_STORE | TB_ALIGN_64 },
383 { X86::VMOVUPDZrr, X86::VMOVUPDZmr, TB_FOLDED_STORE },
384 { X86::VMOVUPSZrr, X86::VMOVUPSZmr, TB_FOLDED_STORE },
385 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zmr, TB_FOLDED_STORE },
386 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zmr, TB_FOLDED_STORE },
387 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zmr, TB_FOLDED_STORE },
388 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zmr, TB_FOLDED_STORE },
389 // AVX-512 foldable instructions (256-bit versions)
390 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
391 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
392 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
393 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
394 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256mr, TB_FOLDED_STORE },
395 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256mr, TB_FOLDED_STORE },
396 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256mr, TB_FOLDED_STORE },
397 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256mr, TB_FOLDED_STORE },
398 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256mr, TB_FOLDED_STORE },
399 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256mr, TB_FOLDED_STORE },
400 // AVX-512 foldable instructions (128-bit versions)
401 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
402 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
403 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
404 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
405 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128mr, TB_FOLDED_STORE },
406 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128mr, TB_FOLDED_STORE },
407 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128mr, TB_FOLDED_STORE },
408 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128mr, TB_FOLDED_STORE },
409 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128mr, TB_FOLDED_STORE },
410 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128mr, TB_FOLDED_STORE }
413 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
414 unsigned RegOp = OpTbl0[i].RegOp;
415 unsigned MemOp = OpTbl0[i].MemOp;
416 unsigned Flags = OpTbl0[i].Flags;
417 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
418 RegOp, MemOp, TB_INDEX_0 | Flags);
421 static const X86OpTblEntry OpTbl1[] = {
422 { X86::CMP16rr, X86::CMP16rm, 0 },
423 { X86::CMP32rr, X86::CMP32rm, 0 },
424 { X86::CMP64rr, X86::CMP64rm, 0 },
425 { X86::CMP8rr, X86::CMP8rm, 0 },
426 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
427 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
428 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
429 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
430 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
431 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
432 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
433 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
434 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
435 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
436 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
437 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
438 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
439 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
440 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
441 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
442 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
443 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
444 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
445 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
446 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
447 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
448 { X86::CVTDQ2PSrr, X86::CVTDQ2PSrm, TB_ALIGN_16 },
449 { X86::CVTPD2DQrr, X86::CVTPD2DQrm, TB_ALIGN_16 },
450 { X86::CVTPD2PSrr, X86::CVTPD2PSrm, TB_ALIGN_16 },
451 { X86::CVTPS2DQrr, X86::CVTPS2DQrm, TB_ALIGN_16 },
452 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
453 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
454 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
455 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
456 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
457 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
458 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
459 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
460 { X86::MOV16rr, X86::MOV16rm, 0 },
461 { X86::MOV32rr, X86::MOV32rm, 0 },
462 { X86::MOV64rr, X86::MOV64rm, 0 },
463 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
464 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
465 { X86::MOV8rr, X86::MOV8rm, 0 },
466 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
467 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
468 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
469 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
470 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
471 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
472 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
473 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
474 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
475 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
476 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
477 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
478 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
479 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
480 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
481 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
482 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
483 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
484 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
485 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
486 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
487 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
488 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
489 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
490 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
491 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
492 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
493 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
494 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
495 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
496 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
497 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
498 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
499 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
500 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
501 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
502 { X86::SQRTSDr, X86::SQRTSDm, 0 },
503 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
504 { X86::SQRTSSr, X86::SQRTSSm, 0 },
505 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
506 { X86::TEST16rr, X86::TEST16rm, 0 },
507 { X86::TEST32rr, X86::TEST32rm, 0 },
508 { X86::TEST64rr, X86::TEST64rm, 0 },
509 { X86::TEST8rr, X86::TEST8rm, 0 },
510 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
511 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
512 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
513 // AVX 128-bit versions of foldable instructions
514 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
515 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
516 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
517 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
518 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
519 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
520 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
521 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
522 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
523 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
524 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
525 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
526 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
527 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
528 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
529 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
530 { X86::VCVTDQ2PSrr, X86::VCVTDQ2PSrm, 0 },
531 { X86::VCVTPD2DQrr, X86::VCVTPD2DQXrm, 0 },
532 { X86::VCVTPD2PSrr, X86::VCVTPD2PSXrm, 0 },
533 { X86::VCVTPS2DQrr, X86::VCVTPS2DQrm, 0 },
534 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
535 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
536 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
537 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
538 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
539 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
540 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
541 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
542 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
543 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
544 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
545 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
546 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
547 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
548 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
549 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
550 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
551 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
552 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
553 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
554 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
555 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
556 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
557 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
558 { X86::VRCPPSr, X86::VRCPPSm, 0 },
559 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
560 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
561 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
562 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
563 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
564 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
565 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
566 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
568 // AVX 256-bit foldable instructions
569 { X86::VCVTDQ2PSYrr, X86::VCVTDQ2PSYrm, 0 },
570 { X86::VCVTPD2DQYrr, X86::VCVTPD2DQYrm, 0 },
571 { X86::VCVTPD2PSYrr, X86::VCVTPD2PSYrm, 0 },
572 { X86::VCVTPS2DQYrr, X86::VCVTPS2DQYrm, 0 },
573 { X86::VCVTTPD2DQYrr, X86::VCVTTPD2DQYrm, 0 },
574 { X86::VCVTTPS2DQYrr, X86::VCVTTPS2DQYrm, 0 },
575 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
576 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
577 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
578 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
579 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
580 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
581 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
582 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
583 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
584 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
585 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
586 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
587 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
588 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
590 // AVX2 foldable instructions
591 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
592 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
593 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
594 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
595 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
596 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
598 // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
599 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
600 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
601 { X86::BEXTRI32ri, X86::BEXTRI32mi, 0 },
602 { X86::BEXTRI64ri, X86::BEXTRI64mi, 0 },
603 { X86::BLCFILL32rr, X86::BLCFILL32rm, 0 },
604 { X86::BLCFILL64rr, X86::BLCFILL64rm, 0 },
605 { X86::BLCI32rr, X86::BLCI32rm, 0 },
606 { X86::BLCI64rr, X86::BLCI64rm, 0 },
607 { X86::BLCIC32rr, X86::BLCIC32rm, 0 },
608 { X86::BLCIC64rr, X86::BLCIC64rm, 0 },
609 { X86::BLCMSK32rr, X86::BLCMSK32rm, 0 },
610 { X86::BLCMSK64rr, X86::BLCMSK64rm, 0 },
611 { X86::BLCS32rr, X86::BLCS32rm, 0 },
612 { X86::BLCS64rr, X86::BLCS64rm, 0 },
613 { X86::BLSFILL32rr, X86::BLSFILL32rm, 0 },
614 { X86::BLSFILL64rr, X86::BLSFILL64rm, 0 },
615 { X86::BLSI32rr, X86::BLSI32rm, 0 },
616 { X86::BLSI64rr, X86::BLSI64rm, 0 },
617 { X86::BLSIC32rr, X86::BLSIC32rm, 0 },
618 { X86::BLSIC64rr, X86::BLSIC64rm, 0 },
619 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
620 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
621 { X86::BLSR32rr, X86::BLSR32rm, 0 },
622 { X86::BLSR64rr, X86::BLSR64rm, 0 },
623 { X86::BZHI32rr, X86::BZHI32rm, 0 },
624 { X86::BZHI64rr, X86::BZHI64rm, 0 },
625 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
626 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
627 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
628 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
629 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
630 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
631 { X86::RORX32ri, X86::RORX32mi, 0 },
632 { X86::RORX64ri, X86::RORX64mi, 0 },
633 { X86::SARX32rr, X86::SARX32rm, 0 },
634 { X86::SARX64rr, X86::SARX64rm, 0 },
635 { X86::SHRX32rr, X86::SHRX32rm, 0 },
636 { X86::SHRX64rr, X86::SHRX64rm, 0 },
637 { X86::SHLX32rr, X86::SHLX32rm, 0 },
638 { X86::SHLX64rr, X86::SHLX64rm, 0 },
639 { X86::T1MSKC32rr, X86::T1MSKC32rm, 0 },
640 { X86::T1MSKC64rr, X86::T1MSKC64rm, 0 },
641 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
642 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
643 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
644 { X86::TZMSK32rr, X86::TZMSK32rm, 0 },
645 { X86::TZMSK64rr, X86::TZMSK64rm, 0 },
647 // AVX-512 foldable instructions
648 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 },
649 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 },
650 { X86::VMOVAPDZrr, X86::VMOVAPDZrm, TB_ALIGN_64 },
651 { X86::VMOVAPSZrr, X86::VMOVAPSZrm, TB_ALIGN_64 },
652 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zrm, TB_ALIGN_64 },
653 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zrm, TB_ALIGN_64 },
654 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zrm, 0 },
655 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zrm, 0 },
656 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zrm, 0 },
657 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zrm, 0 },
658 { X86::VMOVUPDZrr, X86::VMOVUPDZrm, 0 },
659 { X86::VMOVUPSZrr, X86::VMOVUPSZrm, 0 },
660 { X86::VPABSDZrr, X86::VPABSDZrm, 0 },
661 { X86::VPABSQZrr, X86::VPABSQZrm, 0 },
662 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZm, TB_NO_REVERSE },
663 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZm, TB_NO_REVERSE },
664 // AVX-512 foldable instructions (256-bit versions)
665 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256rm, TB_ALIGN_32 },
666 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256rm, TB_ALIGN_32 },
667 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256rm, TB_ALIGN_32 },
668 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256rm, TB_ALIGN_32 },
669 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256rm, 0 },
670 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256rm, 0 },
671 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256rm, 0 },
672 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256rm, 0 },
673 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256rm, 0 },
674 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256rm, 0 },
675 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256m, TB_NO_REVERSE },
676 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256m, TB_NO_REVERSE },
677 // AVX-512 foldable instructions (256-bit versions)
678 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128rm, TB_ALIGN_16 },
679 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128rm, TB_ALIGN_16 },
680 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128rm, TB_ALIGN_16 },
681 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128rm, TB_ALIGN_16 },
682 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128rm, 0 },
683 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128rm, 0 },
684 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128rm, 0 },
685 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128rm, 0 },
686 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128rm, 0 },
687 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128rm, 0 },
688 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128m, TB_NO_REVERSE },
690 // AES foldable instructions
691 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 },
692 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 },
693 { X86::VAESIMCrr, X86::VAESIMCrm, TB_ALIGN_16 },
694 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, TB_ALIGN_16 }
697 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
698 unsigned RegOp = OpTbl1[i].RegOp;
699 unsigned MemOp = OpTbl1[i].MemOp;
700 unsigned Flags = OpTbl1[i].Flags;
701 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
703 // Index 1, folded load
704 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
707 static const X86OpTblEntry OpTbl2[] = {
708 { X86::ADC32rr, X86::ADC32rm, 0 },
709 { X86::ADC64rr, X86::ADC64rm, 0 },
710 { X86::ADD16rr, X86::ADD16rm, 0 },
711 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
712 { X86::ADD32rr, X86::ADD32rm, 0 },
713 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
714 { X86::ADD64rr, X86::ADD64rm, 0 },
715 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
716 { X86::ADD8rr, X86::ADD8rm, 0 },
717 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
718 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
719 { X86::ADDSDrr, X86::ADDSDrm, 0 },
720 { X86::ADDSSrr, X86::ADDSSrm, 0 },
721 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
722 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
723 { X86::AND16rr, X86::AND16rm, 0 },
724 { X86::AND32rr, X86::AND32rm, 0 },
725 { X86::AND64rr, X86::AND64rm, 0 },
726 { X86::AND8rr, X86::AND8rm, 0 },
727 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
728 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
729 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
730 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
731 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
732 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
733 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
734 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
735 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
736 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
737 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
738 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
739 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
740 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
741 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
742 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
743 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
744 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
745 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
746 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
747 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
748 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
749 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
750 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
751 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
752 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
753 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
754 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
755 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
756 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
757 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
758 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
759 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
760 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
761 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
762 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
763 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
764 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
765 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
766 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
767 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
768 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
769 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
770 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
771 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
772 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
773 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
774 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
775 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
776 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
777 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
778 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
779 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
780 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
781 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
782 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
783 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
784 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
785 { X86::CMPSDrr, X86::CMPSDrm, 0 },
786 { X86::CMPSSrr, X86::CMPSSrm, 0 },
787 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
788 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
789 { X86::DIVSDrr, X86::DIVSDrm, 0 },
790 { X86::DIVSSrr, X86::DIVSSrm, 0 },
791 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
792 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
793 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
794 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
795 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
796 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
797 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
798 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
799 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
800 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
801 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
802 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
803 { X86::IMUL16rr, X86::IMUL16rm, 0 },
804 { X86::IMUL32rr, X86::IMUL32rm, 0 },
805 { X86::IMUL64rr, X86::IMUL64rm, 0 },
806 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
807 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
808 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
809 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
810 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
811 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
812 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
813 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
814 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
815 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
816 { X86::MAXSDrr, X86::MAXSDrm, 0 },
817 { X86::MAXSSrr, X86::MAXSSrm, 0 },
818 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
819 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
820 { X86::MINSDrr, X86::MINSDrm, 0 },
821 { X86::MINSSrr, X86::MINSSrm, 0 },
822 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
823 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
824 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
825 { X86::MULSDrr, X86::MULSDrm, 0 },
826 { X86::MULSSrr, X86::MULSSrm, 0 },
827 { X86::OR16rr, X86::OR16rm, 0 },
828 { X86::OR32rr, X86::OR32rm, 0 },
829 { X86::OR64rr, X86::OR64rm, 0 },
830 { X86::OR8rr, X86::OR8rm, 0 },
831 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
832 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
833 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
834 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
835 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
836 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
837 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
838 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
839 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
840 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
841 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
842 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
843 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
844 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
845 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
846 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
847 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
848 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
849 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
850 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
851 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
852 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
853 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
854 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
855 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
856 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
857 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
858 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
859 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
860 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
861 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
862 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
863 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
864 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
865 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
866 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
867 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
868 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
869 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
870 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
871 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
872 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
873 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
874 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
875 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
876 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
877 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
878 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
879 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
880 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
881 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
882 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
883 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
884 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
885 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
886 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
887 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
888 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
889 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
890 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
891 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
892 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
893 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
894 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
895 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
896 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
897 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
898 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
899 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
900 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
901 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
902 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
903 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
904 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
905 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
906 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
907 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
908 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
909 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
910 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
911 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
912 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
913 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
914 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
915 { X86::SBB32rr, X86::SBB32rm, 0 },
916 { X86::SBB64rr, X86::SBB64rm, 0 },
917 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
918 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
919 { X86::SUB16rr, X86::SUB16rm, 0 },
920 { X86::SUB32rr, X86::SUB32rm, 0 },
921 { X86::SUB64rr, X86::SUB64rm, 0 },
922 { X86::SUB8rr, X86::SUB8rm, 0 },
923 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
924 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
925 { X86::SUBSDrr, X86::SUBSDrm, 0 },
926 { X86::SUBSSrr, X86::SUBSSrm, 0 },
927 // FIXME: TEST*rr -> swapped operand of TEST*mr.
928 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
929 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
930 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
931 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
932 { X86::XOR16rr, X86::XOR16rm, 0 },
933 { X86::XOR32rr, X86::XOR32rm, 0 },
934 { X86::XOR64rr, X86::XOR64rm, 0 },
935 { X86::XOR8rr, X86::XOR8rm, 0 },
936 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
937 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
938 // AVX 128-bit versions of foldable instructions
939 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
940 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
941 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
942 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
943 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
944 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
945 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
946 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
947 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
948 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
949 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
950 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
951 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
952 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
953 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
954 { X86::VADDPDrr, X86::VADDPDrm, 0 },
955 { X86::VADDPSrr, X86::VADDPSrm, 0 },
956 { X86::VADDSDrr, X86::VADDSDrm, 0 },
957 { X86::VADDSSrr, X86::VADDSSrm, 0 },
958 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
959 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
960 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
961 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
962 { X86::VANDPDrr, X86::VANDPDrm, 0 },
963 { X86::VANDPSrr, X86::VANDPSrm, 0 },
964 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
965 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
966 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
967 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
968 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
969 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
970 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
971 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
972 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
973 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
974 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
975 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
976 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
977 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
978 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
979 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
980 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
981 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
982 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
983 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
984 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
985 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
986 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
987 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
988 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
989 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
990 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
991 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
992 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
993 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
994 { X86::VMINPDrr, X86::VMINPDrm, 0 },
995 { X86::VMINPSrr, X86::VMINPSrm, 0 },
996 { X86::VMINSDrr, X86::VMINSDrm, 0 },
997 { X86::VMINSSrr, X86::VMINSSrm, 0 },
998 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
999 { X86::VMULPDrr, X86::VMULPDrm, 0 },
1000 { X86::VMULPSrr, X86::VMULPSrm, 0 },
1001 { X86::VMULSDrr, X86::VMULSDrm, 0 },
1002 { X86::VMULSSrr, X86::VMULSSrm, 0 },
1003 { X86::VORPDrr, X86::VORPDrm, 0 },
1004 { X86::VORPSrr, X86::VORPSrm, 0 },
1005 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
1006 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
1007 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
1008 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
1009 { X86::VPADDBrr, X86::VPADDBrm, 0 },
1010 { X86::VPADDDrr, X86::VPADDDrm, 0 },
1011 { X86::VPADDQrr, X86::VPADDQrm, 0 },
1012 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
1013 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
1014 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
1015 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
1016 { X86::VPADDWrr, X86::VPADDWrm, 0 },
1017 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
1018 { X86::VPANDNrr, X86::VPANDNrm, 0 },
1019 { X86::VPANDrr, X86::VPANDrm, 0 },
1020 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
1021 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
1022 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
1023 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
1024 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
1025 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
1026 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
1027 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
1028 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
1029 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
1030 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
1031 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
1032 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
1033 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
1034 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
1035 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
1036 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
1037 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
1038 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
1039 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
1040 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
1041 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
1042 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
1043 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
1044 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
1045 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
1046 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
1047 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
1048 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
1049 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
1050 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
1051 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
1052 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
1053 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
1054 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
1055 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
1056 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
1057 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
1058 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
1059 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
1060 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
1061 { X86::VPORrr, X86::VPORrm, 0 },
1062 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
1063 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
1064 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
1065 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
1066 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
1067 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
1068 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
1069 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
1070 { X86::VPSRADrr, X86::VPSRADrm, 0 },
1071 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
1072 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
1073 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
1074 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
1075 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
1076 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
1077 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
1078 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
1079 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
1080 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
1081 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
1082 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
1083 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
1084 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
1085 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
1086 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
1087 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
1088 { X86::VPXORrr, X86::VPXORrm, 0 },
1089 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
1090 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
1091 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
1092 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
1093 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
1094 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
1095 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
1096 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
1097 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
1098 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
1099 { X86::VXORPDrr, X86::VXORPDrm, 0 },
1100 { X86::VXORPSrr, X86::VXORPSrm, 0 },
1101 // AVX 256-bit foldable instructions
1102 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
1103 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
1104 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
1105 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
1106 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1107 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1108 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1109 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1110 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1111 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1112 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1113 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1114 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1115 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1116 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1117 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1118 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1119 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1120 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1121 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1122 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1123 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
1124 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
1125 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
1126 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
1127 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1128 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1129 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1130 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1131 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1132 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1133 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1134 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1135 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1136 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1137 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1138 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1139 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1140 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1141 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1142 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1143 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
1144 // AVX2 foldable instructions
1145 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1146 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1147 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1148 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1149 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1150 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1151 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1152 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1153 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1154 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1155 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1156 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1157 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1158 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1159 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1160 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1161 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1162 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1163 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1164 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1165 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1166 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1167 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1168 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1169 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1170 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1171 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1172 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1173 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1174 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1175 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1176 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1177 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1178 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1179 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1180 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1181 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1182 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1183 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1184 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1185 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1186 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1187 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1188 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1189 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1190 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1191 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1192 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1193 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1194 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1195 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1196 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1197 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1198 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1199 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1200 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1201 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1202 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1203 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1204 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1205 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1206 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1207 { X86::VPORYrr, X86::VPORYrm, 0 },
1208 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1209 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1210 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1211 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1212 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1213 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1214 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1215 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1216 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1217 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1218 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1219 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1220 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1221 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1222 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1223 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1224 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1225 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1226 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1227 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1228 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1229 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1230 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1231 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1232 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1233 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1234 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1235 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1236 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1237 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1238 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1239 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1240 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1241 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1242 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1243 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1244 { X86::VPXORYrr, X86::VPXORYrm, 0 },
1245 // FIXME: add AVX 256-bit foldable instructions
1247 // FMA4 foldable patterns
1248 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1249 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
1250 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1251 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1252 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1253 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1254 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1255 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
1256 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1257 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1258 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1259 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1260 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1261 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
1262 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1263 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1264 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1265 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1266 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1267 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
1268 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1269 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1270 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1271 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1272 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1273 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1274 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1275 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1276 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1277 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1278 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1279 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
1281 // BMI/BMI2 foldable instructions
1282 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1283 { X86::ANDN64rr, X86::ANDN64rm, 0 },
1284 { X86::MULX32rr, X86::MULX32rm, 0 },
1285 { X86::MULX64rr, X86::MULX64rm, 0 },
1286 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1287 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1288 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1289 { X86::PEXT64rr, X86::PEXT64rm, 0 },
1291 // AVX-512 foldable instructions
1292 { X86::VADDPSZrr, X86::VADDPSZrm, 0 },
1293 { X86::VADDPDZrr, X86::VADDPDZrm, 0 },
1294 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 },
1295 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 },
1296 { X86::VMULPSZrr, X86::VMULPSZrm, 0 },
1297 { X86::VMULPDZrr, X86::VMULPDZrm, 0 },
1298 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 },
1299 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 },
1300 { X86::VMINPSZrr, X86::VMINPSZrm, 0 },
1301 { X86::VMINPDZrr, X86::VMINPDZrm, 0 },
1302 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 },
1303 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 },
1304 { X86::VPADDDZrr, X86::VPADDDZrm, 0 },
1305 { X86::VPADDQZrr, X86::VPADDQZrm, 0 },
1306 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1307 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1308 { X86::VPMAXSDZrr, X86::VPMAXSDZrm, 0 },
1309 { X86::VPMAXSQZrr, X86::VPMAXSQZrm, 0 },
1310 { X86::VPMAXUDZrr, X86::VPMAXUDZrm, 0 },
1311 { X86::VPMAXUQZrr, X86::VPMAXUQZrm, 0 },
1312 { X86::VPMINSDZrr, X86::VPMINSDZrm, 0 },
1313 { X86::VPMINSQZrr, X86::VPMINSQZrm, 0 },
1314 { X86::VPMINUDZrr, X86::VPMINUDZrm, 0 },
1315 { X86::VPMINUQZrr, X86::VPMINUQZrm, 0 },
1316 { X86::VPMULDQZrr, X86::VPMULDQZrm, 0 },
1317 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 },
1318 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 },
1319 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 },
1320 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 },
1321 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 },
1322 { X86::VPSUBDZrr, X86::VPSUBDZrm, 0 },
1323 { X86::VPSUBQZrr, X86::VPSUBQZrm, 0 },
1324 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 },
1325 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 },
1326 { X86::VALIGNQrri, X86::VALIGNQrmi, 0 },
1327 { X86::VALIGNDrri, X86::VALIGNDrmi, 0 },
1328 { X86::VPMULUDQZrr, X86::VPMULUDQZrm, 0 },
1329 { X86::VBROADCASTSSZrkz, X86::VBROADCASTSSZmkz, TB_NO_REVERSE },
1330 { X86::VBROADCASTSDZrkz, X86::VBROADCASTSDZmkz, TB_NO_REVERSE },
1332 // AVX-512{F,VL} foldable instructions
1333 { X86::VBROADCASTSSZ256rkz, X86::VBROADCASTSSZ256mkz, TB_NO_REVERSE },
1334 { X86::VBROADCASTSDZ256rkz, X86::VBROADCASTSDZ256mkz, TB_NO_REVERSE },
1335 { X86::VBROADCASTSSZ128rkz, X86::VBROADCASTSSZ128mkz, TB_NO_REVERSE },
1337 // AVX-512{F,VL} foldable instructions
1338 { X86::VADDPDZ128rr, X86::VADDPDZ128rm, 0 },
1339 { X86::VADDPDZ256rr, X86::VADDPDZ256rm, 0 },
1340 { X86::VADDPSZ128rr, X86::VADDPSZ128rm, 0 },
1341 { X86::VADDPSZ256rr, X86::VADDPSZ256rm, 0 },
1343 // AES foldable instructions
1344 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 },
1345 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 },
1346 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 },
1347 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 },
1348 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, TB_ALIGN_16 },
1349 { X86::VAESDECrr, X86::VAESDECrm, TB_ALIGN_16 },
1350 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, TB_ALIGN_16 },
1351 { X86::VAESENCrr, X86::VAESENCrm, TB_ALIGN_16 },
1353 // SHA foldable instructions
1354 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 },
1355 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 },
1356 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 },
1357 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 },
1358 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 },
1359 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 },
1360 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 },
1363 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1364 unsigned RegOp = OpTbl2[i].RegOp;
1365 unsigned MemOp = OpTbl2[i].MemOp;
1366 unsigned Flags = OpTbl2[i].Flags;
1367 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1369 // Index 2, folded load
1370 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1373 static const X86OpTblEntry OpTbl3[] = {
1374 // FMA foldable instructions
1375 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, TB_ALIGN_NONE },
1376 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, TB_ALIGN_NONE },
1377 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, TB_ALIGN_NONE },
1378 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, TB_ALIGN_NONE },
1379 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, TB_ALIGN_NONE },
1380 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, TB_ALIGN_NONE },
1382 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_NONE },
1383 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_NONE },
1384 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_NONE },
1385 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_NONE },
1386 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_NONE },
1387 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_NONE },
1388 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_NONE },
1389 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_NONE },
1390 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_NONE },
1391 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_NONE },
1392 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_NONE },
1393 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_NONE },
1395 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, TB_ALIGN_NONE },
1396 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, TB_ALIGN_NONE },
1397 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, TB_ALIGN_NONE },
1398 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, TB_ALIGN_NONE },
1399 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, TB_ALIGN_NONE },
1400 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, TB_ALIGN_NONE },
1402 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_NONE },
1403 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_NONE },
1404 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_NONE },
1405 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_NONE },
1406 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_NONE },
1407 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_NONE },
1408 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_NONE },
1409 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_NONE },
1410 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_NONE },
1411 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_NONE },
1412 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_NONE },
1413 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_NONE },
1415 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, TB_ALIGN_NONE },
1416 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, TB_ALIGN_NONE },
1417 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, TB_ALIGN_NONE },
1418 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, TB_ALIGN_NONE },
1419 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, TB_ALIGN_NONE },
1420 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, TB_ALIGN_NONE },
1422 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_NONE },
1423 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_NONE },
1424 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_NONE },
1425 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_NONE },
1426 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_NONE },
1427 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_NONE },
1428 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_NONE },
1429 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_NONE },
1430 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_NONE },
1431 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_NONE },
1432 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_NONE },
1433 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_NONE },
1435 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, TB_ALIGN_NONE },
1436 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, TB_ALIGN_NONE },
1437 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, TB_ALIGN_NONE },
1438 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, TB_ALIGN_NONE },
1439 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, TB_ALIGN_NONE },
1440 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, TB_ALIGN_NONE },
1442 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_NONE },
1443 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_NONE },
1444 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_NONE },
1445 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_NONE },
1446 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_NONE },
1447 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_NONE },
1448 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_NONE },
1449 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_NONE },
1450 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_NONE },
1451 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_NONE },
1452 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_NONE },
1453 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_NONE },
1455 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_NONE },
1456 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_NONE },
1457 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_NONE },
1458 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_NONE },
1459 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_NONE },
1460 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_NONE },
1461 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_NONE },
1462 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_NONE },
1463 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_NONE },
1464 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_NONE },
1465 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_NONE },
1466 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_NONE },
1468 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_NONE },
1469 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_NONE },
1470 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_NONE },
1471 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_NONE },
1472 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_NONE },
1473 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_NONE },
1474 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_NONE },
1475 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_NONE },
1476 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_NONE },
1477 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_NONE },
1478 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_NONE },
1479 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_NONE },
1481 // FMA4 foldable patterns
1482 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1483 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
1484 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1485 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1486 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1487 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1488 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1489 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
1490 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1491 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1492 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1493 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1494 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1495 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
1496 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1497 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1498 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1499 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1500 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1501 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
1502 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1503 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1504 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1505 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1506 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1507 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1508 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1509 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1510 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1511 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1512 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1513 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
1514 // AVX-512 VPERMI instructions with 3 source operands.
1515 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1516 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1517 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1518 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
1519 { X86::VBLENDMPDZrr, X86::VBLENDMPDZrm, 0 },
1520 { X86::VBLENDMPSZrr, X86::VBLENDMPSZrm, 0 },
1521 { X86::VPBLENDMDZrr, X86::VPBLENDMDZrm, 0 },
1522 { X86::VPBLENDMQZrr, X86::VPBLENDMQZrm, 0 },
1523 { X86::VBROADCASTSSZrk, X86::VBROADCASTSSZmk, TB_NO_REVERSE },
1524 { X86::VBROADCASTSDZrk, X86::VBROADCASTSDZmk, TB_NO_REVERSE },
1525 { X86::VBROADCASTSSZ256rk, X86::VBROADCASTSSZ256mk, TB_NO_REVERSE },
1526 { X86::VBROADCASTSDZ256rk, X86::VBROADCASTSDZ256mk, TB_NO_REVERSE },
1527 { X86::VBROADCASTSSZ128rk, X86::VBROADCASTSSZ128mk, TB_NO_REVERSE },
1528 // AVX-512 arithmetic instructions
1529 { X86::VADDPSZrrkz, X86::VADDPSZrmkz, 0 },
1530 { X86::VADDPDZrrkz, X86::VADDPDZrmkz, 0 },
1531 { X86::VSUBPSZrrkz, X86::VSUBPSZrmkz, 0 },
1532 { X86::VSUBPDZrrkz, X86::VSUBPDZrmkz, 0 },
1533 { X86::VMULPSZrrkz, X86::VMULPSZrmkz, 0 },
1534 { X86::VMULPDZrrkz, X86::VMULPDZrmkz, 0 },
1535 { X86::VDIVPSZrrkz, X86::VDIVPSZrmkz, 0 },
1536 { X86::VDIVPDZrrkz, X86::VDIVPDZrmkz, 0 },
1537 { X86::VMINPSZrrkz, X86::VMINPSZrmkz, 0 },
1538 { X86::VMINPDZrrkz, X86::VMINPDZrmkz, 0 },
1539 { X86::VMAXPSZrrkz, X86::VMAXPSZrmkz, 0 },
1540 { X86::VMAXPDZrrkz, X86::VMAXPDZrmkz, 0 },
1541 // AVX-512{F,VL} arithmetic instructions 256-bit
1542 { X86::VADDPSZ256rrkz, X86::VADDPSZ256rmkz, 0 },
1543 { X86::VADDPDZ256rrkz, X86::VADDPDZ256rmkz, 0 },
1544 { X86::VSUBPSZ256rrkz, X86::VSUBPSZ256rmkz, 0 },
1545 { X86::VSUBPDZ256rrkz, X86::VSUBPDZ256rmkz, 0 },
1546 { X86::VMULPSZ256rrkz, X86::VMULPSZ256rmkz, 0 },
1547 { X86::VMULPDZ256rrkz, X86::VMULPDZ256rmkz, 0 },
1548 { X86::VDIVPSZ256rrkz, X86::VDIVPSZ256rmkz, 0 },
1549 { X86::VDIVPDZ256rrkz, X86::VDIVPDZ256rmkz, 0 },
1550 { X86::VMINPSZ256rrkz, X86::VMINPSZ256rmkz, 0 },
1551 { X86::VMINPDZ256rrkz, X86::VMINPDZ256rmkz, 0 },
1552 { X86::VMAXPSZ256rrkz, X86::VMAXPSZ256rmkz, 0 },
1553 { X86::VMAXPDZ256rrkz, X86::VMAXPDZ256rmkz, 0 },
1554 // AVX-512{F,VL} arithmetic instructions 128-bit
1555 { X86::VADDPSZ128rrkz, X86::VADDPSZ128rmkz, 0 },
1556 { X86::VADDPDZ128rrkz, X86::VADDPDZ128rmkz, 0 },
1557 { X86::VSUBPSZ128rrkz, X86::VSUBPSZ128rmkz, 0 },
1558 { X86::VSUBPDZ128rrkz, X86::VSUBPDZ128rmkz, 0 },
1559 { X86::VMULPSZ128rrkz, X86::VMULPSZ128rmkz, 0 },
1560 { X86::VMULPDZ128rrkz, X86::VMULPDZ128rmkz, 0 },
1561 { X86::VDIVPSZ128rrkz, X86::VDIVPSZ128rmkz, 0 },
1562 { X86::VDIVPDZ128rrkz, X86::VDIVPDZ128rmkz, 0 },
1563 { X86::VMINPSZ128rrkz, X86::VMINPSZ128rmkz, 0 },
1564 { X86::VMINPDZ128rrkz, X86::VMINPDZ128rmkz, 0 },
1565 { X86::VMAXPSZ128rrkz, X86::VMAXPSZ128rmkz, 0 },
1566 { X86::VMAXPDZ128rrkz, X86::VMAXPDZ128rmkz, 0 }
1569 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1570 unsigned RegOp = OpTbl3[i].RegOp;
1571 unsigned MemOp = OpTbl3[i].MemOp;
1572 unsigned Flags = OpTbl3[i].Flags;
1573 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1575 // Index 3, folded load
1576 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1579 static const X86OpTblEntry OpTbl4[] = {
1580 // AVX-512 foldable instructions
1581 { X86::VADDPSZrrk, X86::VADDPSZrmk, 0 },
1582 { X86::VADDPDZrrk, X86::VADDPDZrmk, 0 },
1583 { X86::VSUBPSZrrk, X86::VSUBPSZrmk, 0 },
1584 { X86::VSUBPDZrrk, X86::VSUBPDZrmk, 0 },
1585 { X86::VMULPSZrrk, X86::VMULPSZrmk, 0 },
1586 { X86::VMULPDZrrk, X86::VMULPDZrmk, 0 },
1587 { X86::VDIVPSZrrk, X86::VDIVPSZrmk, 0 },
1588 { X86::VDIVPDZrrk, X86::VDIVPDZrmk, 0 },
1589 { X86::VMINPSZrrk, X86::VMINPSZrmk, 0 },
1590 { X86::VMINPDZrrk, X86::VMINPDZrmk, 0 },
1591 { X86::VMAXPSZrrk, X86::VMAXPSZrmk, 0 },
1592 { X86::VMAXPDZrrk, X86::VMAXPDZrmk, 0 },
1593 // AVX-512{F,VL} foldable instructions 256-bit
1594 { X86::VADDPSZ256rrk, X86::VADDPSZ256rmk, 0 },
1595 { X86::VADDPDZ256rrk, X86::VADDPDZ256rmk, 0 },
1596 { X86::VSUBPSZ256rrk, X86::VSUBPSZ256rmk, 0 },
1597 { X86::VSUBPDZ256rrk, X86::VSUBPDZ256rmk, 0 },
1598 { X86::VMULPSZ256rrk, X86::VMULPSZ256rmk, 0 },
1599 { X86::VMULPDZ256rrk, X86::VMULPDZ256rmk, 0 },
1600 { X86::VDIVPSZ256rrk, X86::VDIVPSZ256rmk, 0 },
1601 { X86::VDIVPDZ256rrk, X86::VDIVPDZ256rmk, 0 },
1602 { X86::VMINPSZ256rrk, X86::VMINPSZ256rmk, 0 },
1603 { X86::VMINPDZ256rrk, X86::VMINPDZ256rmk, 0 },
1604 { X86::VMAXPSZ256rrk, X86::VMAXPSZ256rmk, 0 },
1605 { X86::VMAXPDZ256rrk, X86::VMAXPDZ256rmk, 0 },
1606 // AVX-512{F,VL} foldable instructions 128-bit
1607 { X86::VADDPSZ128rrk, X86::VADDPSZ128rmk, 0 },
1608 { X86::VADDPDZ128rrk, X86::VADDPDZ128rmk, 0 },
1609 { X86::VSUBPSZ128rrk, X86::VSUBPSZ128rmk, 0 },
1610 { X86::VSUBPDZ128rrk, X86::VSUBPDZ128rmk, 0 },
1611 { X86::VMULPSZ128rrk, X86::VMULPSZ128rmk, 0 },
1612 { X86::VMULPDZ128rrk, X86::VMULPDZ128rmk, 0 },
1613 { X86::VDIVPSZ128rrk, X86::VDIVPSZ128rmk, 0 },
1614 { X86::VDIVPDZ128rrk, X86::VDIVPDZ128rmk, 0 },
1615 { X86::VMINPSZ128rrk, X86::VMINPSZ128rmk, 0 },
1616 { X86::VMINPDZ128rrk, X86::VMINPDZ128rmk, 0 },
1617 { X86::VMAXPSZ128rrk, X86::VMAXPSZ128rmk, 0 },
1618 { X86::VMAXPDZ128rrk, X86::VMAXPDZ128rmk, 0 }
1621 for (unsigned i = 0, e = array_lengthof(OpTbl4); i != e; ++i) {
1622 unsigned RegOp = OpTbl4[i].RegOp;
1623 unsigned MemOp = OpTbl4[i].MemOp;
1624 unsigned Flags = OpTbl4[i].Flags;
1625 AddTableEntry(RegOp2MemOpTable4, MemOp2RegOpTable,
1627 // Index 4, folded load
1628 Flags | TB_INDEX_4 | TB_FOLDED_LOAD);
1633 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1634 MemOp2RegOpTableType &M2RTable,
1635 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1636 if ((Flags & TB_NO_FORWARD) == 0) {
1637 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1638 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1640 if ((Flags & TB_NO_REVERSE) == 0) {
1641 assert(!M2RTable.count(MemOp) &&
1642 "Duplicated entries in unfolding maps?");
1643 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1648 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1649 unsigned &SrcReg, unsigned &DstReg,
1650 unsigned &SubIdx) const {
1651 switch (MI.getOpcode()) {
1653 case X86::MOVSX16rr8:
1654 case X86::MOVZX16rr8:
1655 case X86::MOVSX32rr8:
1656 case X86::MOVZX32rr8:
1657 case X86::MOVSX64rr8:
1658 if (!Subtarget.is64Bit())
1659 // It's not always legal to reference the low 8-bit of the larger
1660 // register in 32-bit mode.
1662 case X86::MOVSX32rr16:
1663 case X86::MOVZX32rr16:
1664 case X86::MOVSX64rr16:
1665 case X86::MOVSX64rr32: {
1666 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1669 SrcReg = MI.getOperand(1).getReg();
1670 DstReg = MI.getOperand(0).getReg();
1671 switch (MI.getOpcode()) {
1672 default: llvm_unreachable("Unreachable!");
1673 case X86::MOVSX16rr8:
1674 case X86::MOVZX16rr8:
1675 case X86::MOVSX32rr8:
1676 case X86::MOVZX32rr8:
1677 case X86::MOVSX64rr8:
1678 SubIdx = X86::sub_8bit;
1680 case X86::MOVSX32rr16:
1681 case X86::MOVZX32rr16:
1682 case X86::MOVSX64rr16:
1683 SubIdx = X86::sub_16bit;
1685 case X86::MOVSX64rr32:
1686 SubIdx = X86::sub_32bit;
1695 /// isFrameOperand - Return true and the FrameIndex if the specified
1696 /// operand and follow operands form a reference to the stack frame.
1697 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1698 int &FrameIndex) const {
1699 if (MI->getOperand(Op+X86::AddrBaseReg).isFI() &&
1700 MI->getOperand(Op+X86::AddrScaleAmt).isImm() &&
1701 MI->getOperand(Op+X86::AddrIndexReg).isReg() &&
1702 MI->getOperand(Op+X86::AddrDisp).isImm() &&
1703 MI->getOperand(Op+X86::AddrScaleAmt).getImm() == 1 &&
1704 MI->getOperand(Op+X86::AddrIndexReg).getReg() == 0 &&
1705 MI->getOperand(Op+X86::AddrDisp).getImm() == 0) {
1706 FrameIndex = MI->getOperand(Op+X86::AddrBaseReg).getIndex();
1712 static bool isFrameLoadOpcode(int Opcode) {
1728 case X86::VMOVAPSrm:
1729 case X86::VMOVAPDrm:
1730 case X86::VMOVDQArm:
1731 case X86::VMOVUPSYrm:
1732 case X86::VMOVAPSYrm:
1733 case X86::VMOVUPDYrm:
1734 case X86::VMOVAPDYrm:
1735 case X86::VMOVDQUYrm:
1736 case X86::VMOVDQAYrm:
1737 case X86::MMX_MOVD64rm:
1738 case X86::MMX_MOVQ64rm:
1739 case X86::VMOVAPSZrm:
1740 case X86::VMOVUPSZrm:
1745 static bool isFrameStoreOpcode(int Opcode) {
1752 case X86::ST_FpP64m:
1760 case X86::VMOVAPSmr:
1761 case X86::VMOVAPDmr:
1762 case X86::VMOVDQAmr:
1763 case X86::VMOVUPSYmr:
1764 case X86::VMOVAPSYmr:
1765 case X86::VMOVUPDYmr:
1766 case X86::VMOVAPDYmr:
1767 case X86::VMOVDQUYmr:
1768 case X86::VMOVDQAYmr:
1769 case X86::VMOVUPSZmr:
1770 case X86::VMOVAPSZmr:
1771 case X86::MMX_MOVD64mr:
1772 case X86::MMX_MOVQ64mr:
1773 case X86::MMX_MOVNTQmr:
1779 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1780 int &FrameIndex) const {
1781 if (isFrameLoadOpcode(MI->getOpcode()))
1782 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1783 return MI->getOperand(0).getReg();
1787 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1788 int &FrameIndex) const {
1789 if (isFrameLoadOpcode(MI->getOpcode())) {
1791 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1793 // Check for post-frame index elimination operations
1794 const MachineMemOperand *Dummy;
1795 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1800 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1801 int &FrameIndex) const {
1802 if (isFrameStoreOpcode(MI->getOpcode()))
1803 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1804 isFrameOperand(MI, 0, FrameIndex))
1805 return MI->getOperand(X86::AddrNumOperands).getReg();
1809 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1810 int &FrameIndex) const {
1811 if (isFrameStoreOpcode(MI->getOpcode())) {
1813 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1815 // Check for post-frame index elimination operations
1816 const MachineMemOperand *Dummy;
1817 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1822 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1824 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1825 // Don't waste compile time scanning use-def chains of physregs.
1826 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1828 bool isPICBase = false;
1829 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
1830 E = MRI.def_instr_end(); I != E; ++I) {
1831 MachineInstr *DefMI = &*I;
1832 if (DefMI->getOpcode() != X86::MOVPC32r)
1834 assert(!isPICBase && "More than one PIC base?");
1841 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1842 AliasAnalysis *AA) const {
1843 switch (MI->getOpcode()) {
1859 case X86::VMOVAPSrm:
1860 case X86::VMOVUPSrm:
1861 case X86::VMOVAPDrm:
1862 case X86::VMOVDQArm:
1863 case X86::VMOVDQUrm:
1864 case X86::VMOVAPSYrm:
1865 case X86::VMOVUPSYrm:
1866 case X86::VMOVAPDYrm:
1867 case X86::VMOVDQAYrm:
1868 case X86::VMOVDQUYrm:
1869 case X86::MMX_MOVD64rm:
1870 case X86::MMX_MOVQ64rm:
1871 case X86::FsVMOVAPSrm:
1872 case X86::FsVMOVAPDrm:
1873 case X86::FsMOVAPSrm:
1874 case X86::FsMOVAPDrm: {
1875 // Loads from constant pools are trivially rematerializable.
1876 if (MI->getOperand(1+X86::AddrBaseReg).isReg() &&
1877 MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
1878 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
1879 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
1880 MI->isInvariantLoad(AA)) {
1881 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
1882 if (BaseReg == 0 || BaseReg == X86::RIP)
1884 // Allow re-materialization of PIC load.
1885 if (!ReMatPICStubLoad && MI->getOperand(1+X86::AddrDisp).isGlobal())
1887 const MachineFunction &MF = *MI->getParent()->getParent();
1888 const MachineRegisterInfo &MRI = MF.getRegInfo();
1889 return regIsPICBase(BaseReg, MRI);
1896 if (MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
1897 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
1898 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
1899 !MI->getOperand(1+X86::AddrDisp).isReg()) {
1900 // lea fi#, lea GV, etc. are all rematerializable.
1901 if (!MI->getOperand(1+X86::AddrBaseReg).isReg())
1903 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
1906 // Allow re-materialization of lea PICBase + x.
1907 const MachineFunction &MF = *MI->getParent()->getParent();
1908 const MachineRegisterInfo &MRI = MF.getRegInfo();
1909 return regIsPICBase(BaseReg, MRI);
1915 // All other instructions marked M_REMATERIALIZABLE are always trivially
1916 // rematerializable.
1920 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1921 MachineBasicBlock::iterator I) const {
1922 MachineBasicBlock::iterator E = MBB.end();
1924 // For compile time consideration, if we are not able to determine the
1925 // safety after visiting 4 instructions in each direction, we will assume
1927 MachineBasicBlock::iterator Iter = I;
1928 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1929 bool SeenDef = false;
1930 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1931 MachineOperand &MO = Iter->getOperand(j);
1932 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1936 if (MO.getReg() == X86::EFLAGS) {
1944 // This instruction defines EFLAGS, no need to look any further.
1947 // Skip over DBG_VALUE.
1948 while (Iter != E && Iter->isDebugValue())
1952 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1955 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1956 SE = MBB.succ_end(); SI != SE; ++SI)
1957 if ((*SI)->isLiveIn(X86::EFLAGS))
1962 MachineBasicBlock::iterator B = MBB.begin();
1964 for (unsigned i = 0; i < 4; ++i) {
1965 // If we make it to the beginning of the block, it's safe to clobber
1966 // EFLAGS iff EFLAGS is not live-in.
1968 return !MBB.isLiveIn(X86::EFLAGS);
1971 // Skip over DBG_VALUE.
1972 while (Iter != B && Iter->isDebugValue())
1975 bool SawKill = false;
1976 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1977 MachineOperand &MO = Iter->getOperand(j);
1978 // A register mask may clobber EFLAGS, but we should still look for a
1980 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1982 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1983 if (MO.isDef()) return MO.isDead();
1984 if (MO.isKill()) SawKill = true;
1989 // This instruction kills EFLAGS and doesn't redefine it, so
1990 // there's no need to look further.
1994 // Conservative answer.
1998 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1999 MachineBasicBlock::iterator I,
2000 unsigned DestReg, unsigned SubIdx,
2001 const MachineInstr *Orig,
2002 const TargetRegisterInfo &TRI) const {
2003 // MOV32r0 is implemented with a xor which clobbers condition code.
2004 // Re-materialize it as movri instructions to avoid side effects.
2005 unsigned Opc = Orig->getOpcode();
2006 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
2007 DebugLoc DL = Orig->getDebugLoc();
2008 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
2011 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
2015 MachineInstr *NewMI = std::prev(I);
2016 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
2019 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
2020 /// is not marked dead.
2021 static bool hasLiveCondCodeDef(MachineInstr *MI) {
2022 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2023 MachineOperand &MO = MI->getOperand(i);
2024 if (MO.isReg() && MO.isDef() &&
2025 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
2032 /// getTruncatedShiftCount - check whether the shift count for a machine operand
2034 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
2035 unsigned ShiftAmtOperandIdx) {
2036 // The shift count is six bits with the REX.W prefix and five bits without.
2037 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
2038 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
2039 return Imm & ShiftCountMask;
2042 /// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
2043 /// can be represented by a LEA instruction.
2044 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
2045 // Left shift instructions can be transformed into load-effective-address
2046 // instructions if we can encode them appropriately.
2047 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
2048 // The SIB.scale field is two bits wide which means that we can encode any
2049 // shift amount less than 4.
2050 return ShAmt < 4 && ShAmt > 0;
2053 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
2054 unsigned Opc, bool AllowSP,
2055 unsigned &NewSrc, bool &isKill, bool &isUndef,
2056 MachineOperand &ImplicitOp) const {
2057 MachineFunction &MF = *MI->getParent()->getParent();
2058 const TargetRegisterClass *RC;
2060 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
2062 RC = Opc != X86::LEA32r ?
2063 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
2065 unsigned SrcReg = Src.getReg();
2067 // For both LEA64 and LEA32 the register already has essentially the right
2068 // type (32-bit or 64-bit) we may just need to forbid SP.
2069 if (Opc != X86::LEA64_32r) {
2071 isKill = Src.isKill();
2072 isUndef = Src.isUndef();
2074 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
2075 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
2081 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
2082 // another we need to add 64-bit registers to the final MI.
2083 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
2085 ImplicitOp.setImplicit();
2087 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
2088 MachineBasicBlock::LivenessQueryResult LQR =
2089 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
2092 case MachineBasicBlock::LQR_Unknown:
2093 // We can't give sane liveness flags to the instruction, abandon LEA
2096 case MachineBasicBlock::LQR_Live:
2097 isKill = MI->killsRegister(SrcReg);
2101 // The physreg itself is dead, so we have to use it as an <undef>.
2107 // Virtual register of the wrong class, we have to create a temporary 64-bit
2108 // vreg to feed into the LEA.
2109 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
2110 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
2111 get(TargetOpcode::COPY))
2112 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2115 // Which is obviously going to be dead after we're done with it.
2120 // We've set all the parameters without issue.
2124 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
2125 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
2126 /// to a 32-bit superregister and then truncating back down to a 16-bit
2129 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
2130 MachineFunction::iterator &MFI,
2131 MachineBasicBlock::iterator &MBBI,
2132 LiveVariables *LV) const {
2133 MachineInstr *MI = MBBI;
2134 unsigned Dest = MI->getOperand(0).getReg();
2135 unsigned Src = MI->getOperand(1).getReg();
2136 bool isDead = MI->getOperand(0).isDead();
2137 bool isKill = MI->getOperand(1).isKill();
2139 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
2140 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
2141 unsigned Opc, leaInReg;
2142 if (Subtarget.is64Bit()) {
2143 Opc = X86::LEA64_32r;
2144 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2147 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2150 // Build and insert into an implicit UNDEF value. This is OK because
2151 // well be shifting and then extracting the lower 16-bits.
2152 // This has the potential to cause partial register stall. e.g.
2153 // movw (%rbp,%rcx,2), %dx
2154 // leal -65(%rdx), %esi
2155 // But testing has shown this *does* help performance in 64-bit mode (at
2156 // least on modern x86 machines).
2157 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
2158 MachineInstr *InsMI =
2159 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2160 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
2161 .addReg(Src, getKillRegState(isKill));
2163 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
2164 get(Opc), leaOutReg);
2166 default: llvm_unreachable("Unreachable!");
2167 case X86::SHL16ri: {
2168 unsigned ShAmt = MI->getOperand(2).getImm();
2169 MIB.addReg(0).addImm(1 << ShAmt)
2170 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
2174 addRegOffset(MIB, leaInReg, true, 1);
2177 addRegOffset(MIB, leaInReg, true, -1);
2181 case X86::ADD16ri_DB:
2182 case X86::ADD16ri8_DB:
2183 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
2186 case X86::ADD16rr_DB: {
2187 unsigned Src2 = MI->getOperand(2).getReg();
2188 bool isKill2 = MI->getOperand(2).isKill();
2189 unsigned leaInReg2 = 0;
2190 MachineInstr *InsMI2 = nullptr;
2192 // ADD16rr %reg1028<kill>, %reg1028
2193 // just a single insert_subreg.
2194 addRegReg(MIB, leaInReg, true, leaInReg, false);
2196 if (Subtarget.is64Bit())
2197 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2199 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2200 // Build and insert into an implicit UNDEF value. This is OK because
2201 // well be shifting and then extracting the lower 16-bits.
2202 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2204 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2205 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2206 .addReg(Src2, getKillRegState(isKill2));
2207 addRegReg(MIB, leaInReg, true, leaInReg2, true);
2209 if (LV && isKill2 && InsMI2)
2210 LV->replaceKillInstruction(Src2, MI, InsMI2);
2215 MachineInstr *NewMI = MIB;
2216 MachineInstr *ExtMI =
2217 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2218 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2219 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2222 // Update live variables
2223 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2224 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2226 LV->replaceKillInstruction(Src, MI, InsMI);
2228 LV->replaceKillInstruction(Dest, MI, ExtMI);
2234 /// convertToThreeAddress - This method must be implemented by targets that
2235 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
2236 /// may be able to convert a two-address instruction into a true
2237 /// three-address instruction on demand. This allows the X86 target (for
2238 /// example) to convert ADD and SHL instructions into LEA instructions if they
2239 /// would require register copies due to two-addressness.
2241 /// This method returns a null pointer if the transformation cannot be
2242 /// performed, otherwise it returns the new instruction.
2245 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2246 MachineBasicBlock::iterator &MBBI,
2247 LiveVariables *LV) const {
2248 MachineInstr *MI = MBBI;
2250 // The following opcodes also sets the condition code register(s). Only
2251 // convert them to equivalent lea if the condition code register def's
2253 if (hasLiveCondCodeDef(MI))
2256 MachineFunction &MF = *MI->getParent()->getParent();
2257 // All instructions input are two-addr instructions. Get the known operands.
2258 const MachineOperand &Dest = MI->getOperand(0);
2259 const MachineOperand &Src = MI->getOperand(1);
2261 MachineInstr *NewMI = nullptr;
2262 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
2263 // we have better subtarget support, enable the 16-bit LEA generation here.
2264 // 16-bit LEA is also slow on Core2.
2265 bool DisableLEA16 = true;
2266 bool is64Bit = Subtarget.is64Bit();
2268 unsigned MIOpc = MI->getOpcode();
2270 default: return nullptr;
2271 case X86::SHL64ri: {
2272 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2273 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2274 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2276 // LEA can't handle RSP.
2277 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2278 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2279 &X86::GR64_NOSPRegClass))
2282 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2284 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2287 case X86::SHL32ri: {
2288 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2289 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2290 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2292 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2294 // LEA can't handle ESP.
2295 bool isKill, isUndef;
2297 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2298 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2299 SrcReg, isKill, isUndef, ImplicitOp))
2302 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2304 .addReg(0).addImm(1 << ShAmt)
2305 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2306 .addImm(0).addReg(0);
2307 if (ImplicitOp.getReg() != 0)
2308 MIB.addOperand(ImplicitOp);
2313 case X86::SHL16ri: {
2314 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2315 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2316 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2319 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : nullptr;
2320 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2322 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2327 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2328 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2329 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2330 bool isKill, isUndef;
2332 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2333 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2334 SrcReg, isKill, isUndef, ImplicitOp))
2337 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2339 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2340 if (ImplicitOp.getReg() != 0)
2341 MIB.addOperand(ImplicitOp);
2343 NewMI = addOffset(MIB, 1);
2348 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2350 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2351 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2352 .addOperand(Dest).addOperand(Src), 1);
2356 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2357 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2358 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2360 bool isKill, isUndef;
2362 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2363 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2364 SrcReg, isKill, isUndef, ImplicitOp))
2367 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2369 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2370 if (ImplicitOp.getReg() != 0)
2371 MIB.addOperand(ImplicitOp);
2373 NewMI = addOffset(MIB, -1);
2379 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2381 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2382 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2383 .addOperand(Dest).addOperand(Src), -1);
2386 case X86::ADD64rr_DB:
2388 case X86::ADD32rr_DB: {
2389 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2391 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2394 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2396 bool isKill, isUndef;
2398 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2399 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2400 SrcReg, isKill, isUndef, ImplicitOp))
2403 const MachineOperand &Src2 = MI->getOperand(2);
2404 bool isKill2, isUndef2;
2406 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2407 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2408 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2411 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2413 if (ImplicitOp.getReg() != 0)
2414 MIB.addOperand(ImplicitOp);
2415 if (ImplicitOp2.getReg() != 0)
2416 MIB.addOperand(ImplicitOp2);
2418 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2420 // Preserve undefness of the operands.
2421 NewMI->getOperand(1).setIsUndef(isUndef);
2422 NewMI->getOperand(3).setIsUndef(isUndef2);
2424 if (LV && Src2.isKill())
2425 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2429 case X86::ADD16rr_DB: {
2431 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2433 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2434 unsigned Src2 = MI->getOperand(2).getReg();
2435 bool isKill2 = MI->getOperand(2).isKill();
2436 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2438 Src.getReg(), Src.isKill(), Src2, isKill2);
2440 // Preserve undefness of the operands.
2441 bool isUndef = MI->getOperand(1).isUndef();
2442 bool isUndef2 = MI->getOperand(2).isUndef();
2443 NewMI->getOperand(1).setIsUndef(isUndef);
2444 NewMI->getOperand(3).setIsUndef(isUndef2);
2447 LV->replaceKillInstruction(Src2, MI, NewMI);
2450 case X86::ADD64ri32:
2452 case X86::ADD64ri32_DB:
2453 case X86::ADD64ri8_DB:
2454 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2455 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2456 .addOperand(Dest).addOperand(Src),
2457 MI->getOperand(2).getImm());
2461 case X86::ADD32ri_DB:
2462 case X86::ADD32ri8_DB: {
2463 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2464 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2466 bool isKill, isUndef;
2468 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2469 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2470 SrcReg, isKill, isUndef, ImplicitOp))
2473 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2475 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2476 if (ImplicitOp.getReg() != 0)
2477 MIB.addOperand(ImplicitOp);
2479 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2484 case X86::ADD16ri_DB:
2485 case X86::ADD16ri8_DB:
2487 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2489 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2490 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2491 .addOperand(Dest).addOperand(Src),
2492 MI->getOperand(2).getImm());
2496 if (!NewMI) return nullptr;
2498 if (LV) { // Update live variables
2500 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2502 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2505 MFI->insert(MBBI, NewMI); // Insert the new inst
2509 /// commuteInstruction - We have a few instructions that must be hacked on to
2513 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2514 switch (MI->getOpcode()) {
2515 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2516 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2517 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2518 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2519 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2520 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2523 switch (MI->getOpcode()) {
2524 default: llvm_unreachable("Unreachable!");
2525 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2526 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2527 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2528 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2529 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2530 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2532 unsigned Amt = MI->getOperand(3).getImm();
2534 MachineFunction &MF = *MI->getParent()->getParent();
2535 MI = MF.CloneMachineInstr(MI);
2538 MI->setDesc(get(Opc));
2539 MI->getOperand(3).setImm(Size-Amt);
2540 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2542 case X86::BLENDPDrri:
2543 case X86::BLENDPSrri:
2544 case X86::PBLENDWrri:
2545 case X86::VBLENDPDrri:
2546 case X86::VBLENDPSrri:
2547 case X86::VBLENDPDYrri:
2548 case X86::VBLENDPSYrri:
2549 case X86::VPBLENDDrri:
2550 case X86::VPBLENDWrri:
2551 case X86::VPBLENDDYrri:
2552 case X86::VPBLENDWYrri:{
2554 switch (MI->getOpcode()) {
2555 default: llvm_unreachable("Unreachable!");
2556 case X86::BLENDPDrri: Mask = 0x03; break;
2557 case X86::BLENDPSrri: Mask = 0x0F; break;
2558 case X86::PBLENDWrri: Mask = 0xFF; break;
2559 case X86::VBLENDPDrri: Mask = 0x03; break;
2560 case X86::VBLENDPSrri: Mask = 0x0F; break;
2561 case X86::VBLENDPDYrri: Mask = 0x0F; break;
2562 case X86::VBLENDPSYrri: Mask = 0xFF; break;
2563 case X86::VPBLENDDrri: Mask = 0x0F; break;
2564 case X86::VPBLENDWrri: Mask = 0xFF; break;
2565 case X86::VPBLENDDYrri: Mask = 0xFF; break;
2566 case X86::VPBLENDWYrri: Mask = 0xFF; break;
2568 // Only the least significant bits of Imm are used.
2569 unsigned Imm = MI->getOperand(3).getImm() & Mask;
2571 MachineFunction &MF = *MI->getParent()->getParent();
2572 MI = MF.CloneMachineInstr(MI);
2575 MI->getOperand(3).setImm(Mask ^ Imm);
2576 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2578 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2579 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2580 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2581 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2582 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2583 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2584 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2585 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2586 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2587 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2588 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2589 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2590 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2591 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2592 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2593 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2595 switch (MI->getOpcode()) {
2596 default: llvm_unreachable("Unreachable!");
2597 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2598 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2599 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2600 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2601 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2602 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2603 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2604 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2605 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2606 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2607 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2608 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2609 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2610 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2611 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2612 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2613 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2614 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2615 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2616 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2617 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2618 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2619 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2620 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2621 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2622 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2623 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2624 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2625 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2626 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2627 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2628 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2629 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2630 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2631 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2632 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2633 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2634 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2635 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2636 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2637 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2638 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2639 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2640 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2641 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2642 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2643 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2644 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2647 MachineFunction &MF = *MI->getParent()->getParent();
2648 MI = MF.CloneMachineInstr(MI);
2651 MI->setDesc(get(Opc));
2652 // Fallthrough intended.
2655 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2659 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
2660 unsigned &SrcOpIdx2) const {
2661 switch (MI->getOpcode()) {
2662 case X86::BLENDPDrri:
2663 case X86::BLENDPSrri:
2664 case X86::PBLENDWrri:
2665 case X86::VBLENDPDrri:
2666 case X86::VBLENDPSrri:
2667 case X86::VBLENDPDYrri:
2668 case X86::VBLENDPSYrri:
2669 case X86::VPBLENDDrri:
2670 case X86::VPBLENDDYrri:
2671 case X86::VPBLENDWrri:
2672 case X86::VPBLENDWYrri:
2676 case X86::VFMADDPDr231r:
2677 case X86::VFMADDPSr231r:
2678 case X86::VFMADDSDr231r:
2679 case X86::VFMADDSSr231r:
2680 case X86::VFMSUBPDr231r:
2681 case X86::VFMSUBPSr231r:
2682 case X86::VFMSUBSDr231r:
2683 case X86::VFMSUBSSr231r:
2684 case X86::VFNMADDPDr231r:
2685 case X86::VFNMADDPSr231r:
2686 case X86::VFNMADDSDr231r:
2687 case X86::VFNMADDSSr231r:
2688 case X86::VFNMSUBPDr231r:
2689 case X86::VFNMSUBPSr231r:
2690 case X86::VFNMSUBSDr231r:
2691 case X86::VFNMSUBSSr231r:
2692 case X86::VFMADDPDr231rY:
2693 case X86::VFMADDPSr231rY:
2694 case X86::VFMSUBPDr231rY:
2695 case X86::VFMSUBPSr231rY:
2696 case X86::VFNMADDPDr231rY:
2697 case X86::VFNMADDPSr231rY:
2698 case X86::VFNMSUBPDr231rY:
2699 case X86::VFNMSUBPSr231rY:
2704 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2708 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2710 default: return X86::COND_INVALID;
2711 case X86::JE_1: return X86::COND_E;
2712 case X86::JNE_1: return X86::COND_NE;
2713 case X86::JL_1: return X86::COND_L;
2714 case X86::JLE_1: return X86::COND_LE;
2715 case X86::JG_1: return X86::COND_G;
2716 case X86::JGE_1: return X86::COND_GE;
2717 case X86::JB_1: return X86::COND_B;
2718 case X86::JBE_1: return X86::COND_BE;
2719 case X86::JA_1: return X86::COND_A;
2720 case X86::JAE_1: return X86::COND_AE;
2721 case X86::JS_1: return X86::COND_S;
2722 case X86::JNS_1: return X86::COND_NS;
2723 case X86::JP_1: return X86::COND_P;
2724 case X86::JNP_1: return X86::COND_NP;
2725 case X86::JO_1: return X86::COND_O;
2726 case X86::JNO_1: return X86::COND_NO;
2730 /// getCondFromSETOpc - return condition code of a SET opcode.
2731 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2733 default: return X86::COND_INVALID;
2734 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2735 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2736 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2737 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2738 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2739 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2740 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2741 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2742 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2743 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2744 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2745 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2746 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2747 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2748 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2749 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2753 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2754 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2756 default: return X86::COND_INVALID;
2757 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2758 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2760 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2761 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2762 return X86::COND_AE;
2763 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2764 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2766 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2767 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2768 return X86::COND_BE;
2769 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2770 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2772 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2773 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2775 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2776 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2777 return X86::COND_GE;
2778 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2779 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2781 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2782 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2783 return X86::COND_LE;
2784 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2785 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2786 return X86::COND_NE;
2787 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2788 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2789 return X86::COND_NO;
2790 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2791 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2792 return X86::COND_NP;
2793 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2794 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2795 return X86::COND_NS;
2796 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2797 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2799 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2800 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2802 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2803 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2808 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2810 default: llvm_unreachable("Illegal condition code!");
2811 case X86::COND_E: return X86::JE_1;
2812 case X86::COND_NE: return X86::JNE_1;
2813 case X86::COND_L: return X86::JL_1;
2814 case X86::COND_LE: return X86::JLE_1;
2815 case X86::COND_G: return X86::JG_1;
2816 case X86::COND_GE: return X86::JGE_1;
2817 case X86::COND_B: return X86::JB_1;
2818 case X86::COND_BE: return X86::JBE_1;
2819 case X86::COND_A: return X86::JA_1;
2820 case X86::COND_AE: return X86::JAE_1;
2821 case X86::COND_S: return X86::JS_1;
2822 case X86::COND_NS: return X86::JNS_1;
2823 case X86::COND_P: return X86::JP_1;
2824 case X86::COND_NP: return X86::JNP_1;
2825 case X86::COND_O: return X86::JO_1;
2826 case X86::COND_NO: return X86::JNO_1;
2830 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2831 /// e.g. turning COND_E to COND_NE.
2832 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2834 default: llvm_unreachable("Illegal condition code!");
2835 case X86::COND_E: return X86::COND_NE;
2836 case X86::COND_NE: return X86::COND_E;
2837 case X86::COND_L: return X86::COND_GE;
2838 case X86::COND_LE: return X86::COND_G;
2839 case X86::COND_G: return X86::COND_LE;
2840 case X86::COND_GE: return X86::COND_L;
2841 case X86::COND_B: return X86::COND_AE;
2842 case X86::COND_BE: return X86::COND_A;
2843 case X86::COND_A: return X86::COND_BE;
2844 case X86::COND_AE: return X86::COND_B;
2845 case X86::COND_S: return X86::COND_NS;
2846 case X86::COND_NS: return X86::COND_S;
2847 case X86::COND_P: return X86::COND_NP;
2848 case X86::COND_NP: return X86::COND_P;
2849 case X86::COND_O: return X86::COND_NO;
2850 case X86::COND_NO: return X86::COND_O;
2854 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2855 /// the condition code if we modify the instructions such that flags are
2857 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2859 default: return X86::COND_INVALID;
2860 case X86::COND_E: return X86::COND_E;
2861 case X86::COND_NE: return X86::COND_NE;
2862 case X86::COND_L: return X86::COND_G;
2863 case X86::COND_LE: return X86::COND_GE;
2864 case X86::COND_G: return X86::COND_L;
2865 case X86::COND_GE: return X86::COND_LE;
2866 case X86::COND_B: return X86::COND_A;
2867 case X86::COND_BE: return X86::COND_AE;
2868 case X86::COND_A: return X86::COND_B;
2869 case X86::COND_AE: return X86::COND_BE;
2873 /// getSETFromCond - Return a set opcode for the given condition and
2874 /// whether it has memory operand.
2875 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
2876 static const uint16_t Opc[16][2] = {
2877 { X86::SETAr, X86::SETAm },
2878 { X86::SETAEr, X86::SETAEm },
2879 { X86::SETBr, X86::SETBm },
2880 { X86::SETBEr, X86::SETBEm },
2881 { X86::SETEr, X86::SETEm },
2882 { X86::SETGr, X86::SETGm },
2883 { X86::SETGEr, X86::SETGEm },
2884 { X86::SETLr, X86::SETLm },
2885 { X86::SETLEr, X86::SETLEm },
2886 { X86::SETNEr, X86::SETNEm },
2887 { X86::SETNOr, X86::SETNOm },
2888 { X86::SETNPr, X86::SETNPm },
2889 { X86::SETNSr, X86::SETNSm },
2890 { X86::SETOr, X86::SETOm },
2891 { X86::SETPr, X86::SETPm },
2892 { X86::SETSr, X86::SETSm }
2895 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
2896 return Opc[CC][HasMemoryOperand ? 1 : 0];
2899 /// getCMovFromCond - Return a cmov opcode for the given condition,
2900 /// register size in bytes, and operand type.
2901 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
2902 bool HasMemoryOperand) {
2903 static const uint16_t Opc[32][3] = {
2904 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2905 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2906 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2907 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2908 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2909 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2910 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2911 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2912 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2913 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2914 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2915 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2916 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2917 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2918 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2919 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2920 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2921 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2922 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2923 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2924 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2925 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2926 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2927 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2928 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2929 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2930 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2931 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2932 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2933 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2934 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2935 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2938 assert(CC < 16 && "Can only handle standard cond codes");
2939 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2941 default: llvm_unreachable("Illegal register size!");
2942 case 2: return Opc[Idx][0];
2943 case 4: return Opc[Idx][1];
2944 case 8: return Opc[Idx][2];
2948 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2949 if (!MI->isTerminator()) return false;
2951 // Conditional branch is a special case.
2952 if (MI->isBranch() && !MI->isBarrier())
2954 if (!MI->isPredicable())
2956 return !isPredicated(MI);
2959 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2960 MachineBasicBlock *&TBB,
2961 MachineBasicBlock *&FBB,
2962 SmallVectorImpl<MachineOperand> &Cond,
2963 bool AllowModify) const {
2964 // Start from the bottom of the block and work up, examining the
2965 // terminator instructions.
2966 MachineBasicBlock::iterator I = MBB.end();
2967 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2968 while (I != MBB.begin()) {
2970 if (I->isDebugValue())
2973 // Working from the bottom, when we see a non-terminator instruction, we're
2975 if (!isUnpredicatedTerminator(I))
2978 // A terminator that isn't a branch can't easily be handled by this
2983 // Handle unconditional branches.
2984 if (I->getOpcode() == X86::JMP_1) {
2988 TBB = I->getOperand(0).getMBB();
2992 // If the block has any instructions after a JMP, delete them.
2993 while (std::next(I) != MBB.end())
2994 std::next(I)->eraseFromParent();
2999 // Delete the JMP if it's equivalent to a fall-through.
3000 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
3002 I->eraseFromParent();
3004 UnCondBrIter = MBB.end();
3008 // TBB is used to indicate the unconditional destination.
3009 TBB = I->getOperand(0).getMBB();
3013 // Handle conditional branches.
3014 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
3015 if (BranchCode == X86::COND_INVALID)
3016 return true; // Can't handle indirect branch.
3018 // Working from the bottom, handle the first conditional branch.
3020 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
3021 if (AllowModify && UnCondBrIter != MBB.end() &&
3022 MBB.isLayoutSuccessor(TargetBB)) {
3023 // If we can modify the code and it ends in something like:
3031 // Then we can change this to:
3038 // Which is a bit more efficient.
3039 // We conditionally jump to the fall-through block.
3040 BranchCode = GetOppositeBranchCondition(BranchCode);
3041 unsigned JNCC = GetCondBranchFromCond(BranchCode);
3042 MachineBasicBlock::iterator OldInst = I;
3044 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
3045 .addMBB(UnCondBrIter->getOperand(0).getMBB());
3046 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
3049 OldInst->eraseFromParent();
3050 UnCondBrIter->eraseFromParent();
3052 // Restart the analysis.
3053 UnCondBrIter = MBB.end();
3059 TBB = I->getOperand(0).getMBB();
3060 Cond.push_back(MachineOperand::CreateImm(BranchCode));
3064 // Handle subsequent conditional branches. Only handle the case where all
3065 // conditional branches branch to the same destination and their condition
3066 // opcodes fit one of the special multi-branch idioms.
3067 assert(Cond.size() == 1);
3070 // Only handle the case where all conditional branches branch to the same
3072 if (TBB != I->getOperand(0).getMBB())
3075 // If the conditions are the same, we can leave them alone.
3076 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
3077 if (OldBranchCode == BranchCode)
3080 // If they differ, see if they fit one of the known patterns. Theoretically,
3081 // we could handle more patterns here, but we shouldn't expect to see them
3082 // if instruction selection has done a reasonable job.
3083 if ((OldBranchCode == X86::COND_NP &&
3084 BranchCode == X86::COND_E) ||
3085 (OldBranchCode == X86::COND_E &&
3086 BranchCode == X86::COND_NP))
3087 BranchCode = X86::COND_NP_OR_E;
3088 else if ((OldBranchCode == X86::COND_P &&
3089 BranchCode == X86::COND_NE) ||
3090 (OldBranchCode == X86::COND_NE &&
3091 BranchCode == X86::COND_P))
3092 BranchCode = X86::COND_NE_OR_P;
3096 // Update the MachineOperand.
3097 Cond[0].setImm(BranchCode);
3103 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
3104 MachineBasicBlock::iterator I = MBB.end();
3107 while (I != MBB.begin()) {
3109 if (I->isDebugValue())
3111 if (I->getOpcode() != X86::JMP_1 &&
3112 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
3114 // Remove the branch.
3115 I->eraseFromParent();
3124 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
3125 MachineBasicBlock *FBB,
3126 const SmallVectorImpl<MachineOperand> &Cond,
3127 DebugLoc DL) const {
3128 // Shouldn't be a fall through.
3129 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
3130 assert((Cond.size() == 1 || Cond.size() == 0) &&
3131 "X86 branch conditions have one component!");
3134 // Unconditional branch?
3135 assert(!FBB && "Unconditional branch with multiple successors!");
3136 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
3140 // Conditional branch.
3142 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
3144 case X86::COND_NP_OR_E:
3145 // Synthesize NP_OR_E with two branches.
3146 BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
3148 BuildMI(&MBB, DL, get(X86::JE_1)).addMBB(TBB);
3151 case X86::COND_NE_OR_P:
3152 // Synthesize NE_OR_P with two branches.
3153 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
3155 BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
3159 unsigned Opc = GetCondBranchFromCond(CC);
3160 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
3165 // Two-way Conditional branch. Insert the second branch.
3166 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
3173 canInsertSelect(const MachineBasicBlock &MBB,
3174 const SmallVectorImpl<MachineOperand> &Cond,
3175 unsigned TrueReg, unsigned FalseReg,
3176 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
3177 // Not all subtargets have cmov instructions.
3178 if (!Subtarget.hasCMov())
3180 if (Cond.size() != 1)
3182 // We cannot do the composite conditions, at least not in SSA form.
3183 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
3186 // Check register classes.
3187 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3188 const TargetRegisterClass *RC =
3189 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
3193 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
3194 if (X86::GR16RegClass.hasSubClassEq(RC) ||
3195 X86::GR32RegClass.hasSubClassEq(RC) ||
3196 X86::GR64RegClass.hasSubClassEq(RC)) {
3197 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
3198 // Bridge. Probably Ivy Bridge as well.
3205 // Can't do vectors.
3209 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3210 MachineBasicBlock::iterator I, DebugLoc DL,
3212 const SmallVectorImpl<MachineOperand> &Cond,
3213 unsigned TrueReg, unsigned FalseReg) const {
3214 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3215 assert(Cond.size() == 1 && "Invalid Cond array");
3216 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
3217 MRI.getRegClass(DstReg)->getSize(),
3218 false/*HasMemoryOperand*/);
3219 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
3222 /// isHReg - Test if the given register is a physical h register.
3223 static bool isHReg(unsigned Reg) {
3224 return X86::GR8_ABCD_HRegClass.contains(Reg);
3227 // Try and copy between VR128/VR64 and GR64 registers.
3228 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3229 const X86Subtarget &Subtarget) {
3231 // SrcReg(VR128) -> DestReg(GR64)
3232 // SrcReg(VR64) -> DestReg(GR64)
3233 // SrcReg(GR64) -> DestReg(VR128)
3234 // SrcReg(GR64) -> DestReg(VR64)
3236 bool HasAVX = Subtarget.hasAVX();
3237 bool HasAVX512 = Subtarget.hasAVX512();
3238 if (X86::GR64RegClass.contains(DestReg)) {
3239 if (X86::VR128XRegClass.contains(SrcReg))
3240 // Copy from a VR128 register to a GR64 register.
3241 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
3243 if (X86::VR64RegClass.contains(SrcReg))
3244 // Copy from a VR64 register to a GR64 register.
3245 return X86::MOVSDto64rr;
3246 } else if (X86::GR64RegClass.contains(SrcReg)) {
3247 // Copy from a GR64 register to a VR128 register.
3248 if (X86::VR128XRegClass.contains(DestReg))
3249 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
3251 // Copy from a GR64 register to a VR64 register.
3252 if (X86::VR64RegClass.contains(DestReg))
3253 return X86::MOV64toSDrr;
3256 // SrcReg(FR32) -> DestReg(GR32)
3257 // SrcReg(GR32) -> DestReg(FR32)
3259 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3260 // Copy from a FR32 register to a GR32 register.
3261 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3263 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3264 // Copy from a GR32 register to a FR32 register.
3265 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3269 inline static bool MaskRegClassContains(unsigned Reg) {
3270 return X86::VK8RegClass.contains(Reg) ||
3271 X86::VK16RegClass.contains(Reg) ||
3272 X86::VK32RegClass.contains(Reg) ||
3273 X86::VK64RegClass.contains(Reg) ||
3274 X86::VK1RegClass.contains(Reg);
3277 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3278 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3279 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3280 X86::VR512RegClass.contains(DestReg, SrcReg)) {
3281 DestReg = get512BitSuperRegister(DestReg);
3282 SrcReg = get512BitSuperRegister(SrcReg);
3283 return X86::VMOVAPSZrr;
3285 if (MaskRegClassContains(DestReg) &&
3286 MaskRegClassContains(SrcReg))
3287 return X86::KMOVWkk;
3288 if (MaskRegClassContains(DestReg) &&
3289 (X86::GR32RegClass.contains(SrcReg) ||
3290 X86::GR16RegClass.contains(SrcReg) ||
3291 X86::GR8RegClass.contains(SrcReg))) {
3292 SrcReg = getX86SubSuperRegister(SrcReg, MVT::i32);
3293 return X86::KMOVWkr;
3295 if ((X86::GR32RegClass.contains(DestReg) ||
3296 X86::GR16RegClass.contains(DestReg) ||
3297 X86::GR8RegClass.contains(DestReg)) &&
3298 MaskRegClassContains(SrcReg)) {
3299 DestReg = getX86SubSuperRegister(DestReg, MVT::i32);
3300 return X86::KMOVWrk;
3305 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3306 MachineBasicBlock::iterator MI, DebugLoc DL,
3307 unsigned DestReg, unsigned SrcReg,
3308 bool KillSrc) const {
3309 // First deal with the normal symmetric copies.
3310 bool HasAVX = Subtarget.hasAVX();
3311 bool HasAVX512 = Subtarget.hasAVX512();
3313 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3315 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3317 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3319 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3320 // Copying to or from a physical H register on x86-64 requires a NOREX
3321 // move. Otherwise use a normal move.
3322 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3323 Subtarget.is64Bit()) {
3324 Opc = X86::MOV8rr_NOREX;
3325 // Both operands must be encodable without an REX prefix.
3326 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3327 "8-bit H register can not be copied outside GR8_NOREX");
3331 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3332 Opc = X86::MMX_MOVQ64rr;
3334 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3335 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3336 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3337 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3338 Opc = X86::VMOVAPSYrr;
3340 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3343 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3344 .addReg(SrcReg, getKillRegState(KillSrc));
3348 // Moving EFLAGS to / from another register requires a push and a pop.
3349 // Notice that we have to adjust the stack if we don't want to clobber the
3350 // first frame index. See X86FrameLowering.cpp - clobbersTheStack.
3351 if (SrcReg == X86::EFLAGS) {
3352 if (X86::GR64RegClass.contains(DestReg)) {
3353 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3354 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3357 if (X86::GR32RegClass.contains(DestReg)) {
3358 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3359 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3363 if (DestReg == X86::EFLAGS) {
3364 if (X86::GR64RegClass.contains(SrcReg)) {
3365 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3366 .addReg(SrcReg, getKillRegState(KillSrc));
3367 BuildMI(MBB, MI, DL, get(X86::POPF64));
3370 if (X86::GR32RegClass.contains(SrcReg)) {
3371 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3372 .addReg(SrcReg, getKillRegState(KillSrc));
3373 BuildMI(MBB, MI, DL, get(X86::POPF32));
3378 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3379 << " to " << RI.getName(DestReg) << '\n');
3380 llvm_unreachable("Cannot emit physreg copy instruction");
3383 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3384 const TargetRegisterClass *RC,
3385 bool isStackAligned,
3386 const X86Subtarget &STI,
3388 if (STI.hasAVX512()) {
3389 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3390 X86::VK16RegClass.hasSubClassEq(RC))
3391 return load ? X86::KMOVWkm : X86::KMOVWmk;
3392 if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3393 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3394 if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3395 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3396 if (X86::VR512RegClass.hasSubClassEq(RC))
3397 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3400 bool HasAVX = STI.hasAVX();
3401 switch (RC->getSize()) {
3403 llvm_unreachable("Unknown spill size");
3405 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3407 // Copying to or from a physical H register on x86-64 requires a NOREX
3408 // move. Otherwise use a normal move.
3409 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3410 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3411 return load ? X86::MOV8rm : X86::MOV8mr;
3413 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3414 return load ? X86::MOV16rm : X86::MOV16mr;
3416 if (X86::GR32RegClass.hasSubClassEq(RC))
3417 return load ? X86::MOV32rm : X86::MOV32mr;
3418 if (X86::FR32RegClass.hasSubClassEq(RC))
3420 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3421 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3422 if (X86::RFP32RegClass.hasSubClassEq(RC))
3423 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3424 llvm_unreachable("Unknown 4-byte regclass");
3426 if (X86::GR64RegClass.hasSubClassEq(RC))
3427 return load ? X86::MOV64rm : X86::MOV64mr;
3428 if (X86::FR64RegClass.hasSubClassEq(RC))
3430 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3431 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3432 if (X86::VR64RegClass.hasSubClassEq(RC))
3433 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3434 if (X86::RFP64RegClass.hasSubClassEq(RC))
3435 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3436 llvm_unreachable("Unknown 8-byte regclass");
3438 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3439 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3441 assert((X86::VR128RegClass.hasSubClassEq(RC) ||
3442 X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
3443 // If stack is realigned we can use aligned stores.
3446 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3447 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
3450 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3451 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3454 assert((X86::VR256RegClass.hasSubClassEq(RC) ||
3455 X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
3456 // If stack is realigned we can use aligned stores.
3458 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3460 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
3462 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3464 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3466 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3470 static unsigned getStoreRegOpcode(unsigned SrcReg,
3471 const TargetRegisterClass *RC,
3472 bool isStackAligned,
3473 const X86Subtarget &STI) {
3474 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
3478 static unsigned getLoadRegOpcode(unsigned DestReg,
3479 const TargetRegisterClass *RC,
3480 bool isStackAligned,
3481 const X86Subtarget &STI) {
3482 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
3485 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3486 MachineBasicBlock::iterator MI,
3487 unsigned SrcReg, bool isKill, int FrameIdx,
3488 const TargetRegisterClass *RC,
3489 const TargetRegisterInfo *TRI) const {
3490 const MachineFunction &MF = *MBB.getParent();
3491 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3492 "Stack slot too small for store");
3493 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3494 bool isAligned = (MF.getTarget()
3496 ->getFrameLowering()
3497 ->getStackAlignment() >= Alignment) ||
3498 RI.canRealignStack(MF);
3499 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3500 DebugLoc DL = MBB.findDebugLoc(MI);
3501 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
3502 .addReg(SrcReg, getKillRegState(isKill));
3505 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3507 SmallVectorImpl<MachineOperand> &Addr,
3508 const TargetRegisterClass *RC,
3509 MachineInstr::mmo_iterator MMOBegin,
3510 MachineInstr::mmo_iterator MMOEnd,
3511 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3512 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3513 bool isAligned = MMOBegin != MMOEnd &&
3514 (*MMOBegin)->getAlignment() >= Alignment;
3515 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3517 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3518 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3519 MIB.addOperand(Addr[i]);
3520 MIB.addReg(SrcReg, getKillRegState(isKill));
3521 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3522 NewMIs.push_back(MIB);
3526 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3527 MachineBasicBlock::iterator MI,
3528 unsigned DestReg, int FrameIdx,
3529 const TargetRegisterClass *RC,
3530 const TargetRegisterInfo *TRI) const {
3531 const MachineFunction &MF = *MBB.getParent();
3532 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3533 bool isAligned = (MF.getTarget()
3535 ->getFrameLowering()
3536 ->getStackAlignment() >= Alignment) ||
3537 RI.canRealignStack(MF);
3538 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3539 DebugLoc DL = MBB.findDebugLoc(MI);
3540 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
3543 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
3544 SmallVectorImpl<MachineOperand> &Addr,
3545 const TargetRegisterClass *RC,
3546 MachineInstr::mmo_iterator MMOBegin,
3547 MachineInstr::mmo_iterator MMOEnd,
3548 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3549 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3550 bool isAligned = MMOBegin != MMOEnd &&
3551 (*MMOBegin)->getAlignment() >= Alignment;
3552 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3554 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3555 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3556 MIB.addOperand(Addr[i]);
3557 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3558 NewMIs.push_back(MIB);
3562 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3563 int &CmpMask, int &CmpValue) const {
3564 switch (MI->getOpcode()) {
3566 case X86::CMP64ri32:
3573 SrcReg = MI->getOperand(0).getReg();
3576 CmpValue = MI->getOperand(1).getImm();
3578 // A SUB can be used to perform comparison.
3583 SrcReg = MI->getOperand(1).getReg();
3592 SrcReg = MI->getOperand(1).getReg();
3593 SrcReg2 = MI->getOperand(2).getReg();
3597 case X86::SUB64ri32:
3604 SrcReg = MI->getOperand(1).getReg();
3607 CmpValue = MI->getOperand(2).getImm();
3613 SrcReg = MI->getOperand(0).getReg();
3614 SrcReg2 = MI->getOperand(1).getReg();
3622 SrcReg = MI->getOperand(0).getReg();
3623 if (MI->getOperand(1).getReg() != SrcReg) return false;
3624 // Compare against zero.
3633 /// isRedundantFlagInstr - check whether the first instruction, whose only
3634 /// purpose is to update flags, can be made redundant.
3635 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3636 /// This function can be extended later on.
3637 /// SrcReg, SrcRegs: register operands for FlagI.
3638 /// ImmValue: immediate for FlagI if it takes an immediate.
3639 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3640 unsigned SrcReg2, int ImmValue,
3642 if (((FlagI->getOpcode() == X86::CMP64rr &&
3643 OI->getOpcode() == X86::SUB64rr) ||
3644 (FlagI->getOpcode() == X86::CMP32rr &&
3645 OI->getOpcode() == X86::SUB32rr)||
3646 (FlagI->getOpcode() == X86::CMP16rr &&
3647 OI->getOpcode() == X86::SUB16rr)||
3648 (FlagI->getOpcode() == X86::CMP8rr &&
3649 OI->getOpcode() == X86::SUB8rr)) &&
3650 ((OI->getOperand(1).getReg() == SrcReg &&
3651 OI->getOperand(2).getReg() == SrcReg2) ||
3652 (OI->getOperand(1).getReg() == SrcReg2 &&
3653 OI->getOperand(2).getReg() == SrcReg)))
3656 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3657 OI->getOpcode() == X86::SUB64ri32) ||
3658 (FlagI->getOpcode() == X86::CMP64ri8 &&
3659 OI->getOpcode() == X86::SUB64ri8) ||
3660 (FlagI->getOpcode() == X86::CMP32ri &&
3661 OI->getOpcode() == X86::SUB32ri) ||
3662 (FlagI->getOpcode() == X86::CMP32ri8 &&
3663 OI->getOpcode() == X86::SUB32ri8) ||
3664 (FlagI->getOpcode() == X86::CMP16ri &&
3665 OI->getOpcode() == X86::SUB16ri) ||
3666 (FlagI->getOpcode() == X86::CMP16ri8 &&
3667 OI->getOpcode() == X86::SUB16ri8) ||
3668 (FlagI->getOpcode() == X86::CMP8ri &&
3669 OI->getOpcode() == X86::SUB8ri)) &&
3670 OI->getOperand(1).getReg() == SrcReg &&
3671 OI->getOperand(2).getImm() == ImmValue)
3676 /// isDefConvertible - check whether the definition can be converted
3677 /// to remove a comparison against zero.
3678 inline static bool isDefConvertible(MachineInstr *MI) {
3679 switch (MI->getOpcode()) {
3680 default: return false;
3682 // The shift instructions only modify ZF if their shift count is non-zero.
3683 // N.B.: The processor truncates the shift count depending on the encoding.
3684 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3685 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3686 return getTruncatedShiftCount(MI, 2) != 0;
3688 // Some left shift instructions can be turned into LEA instructions but only
3689 // if their flags aren't used. Avoid transforming such instructions.
3690 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3691 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3692 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3696 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3697 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3698 return getTruncatedShiftCount(MI, 3) != 0;
3700 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3701 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3702 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3703 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3704 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3705 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3706 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3707 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3708 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3709 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3710 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3711 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3712 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3713 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3714 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3715 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3716 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3717 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3718 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3719 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3720 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3721 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3722 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3723 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3724 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3725 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3726 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3727 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3728 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3729 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3730 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3731 case X86::ADC32ri: case X86::ADC32ri8:
3732 case X86::ADC32rr: case X86::ADC64ri32:
3733 case X86::ADC64ri8: case X86::ADC64rr:
3734 case X86::SBB32ri: case X86::SBB32ri8:
3735 case X86::SBB32rr: case X86::SBB64ri32:
3736 case X86::SBB64ri8: case X86::SBB64rr:
3737 case X86::ANDN32rr: case X86::ANDN32rm:
3738 case X86::ANDN64rr: case X86::ANDN64rm:
3739 case X86::BEXTR32rr: case X86::BEXTR64rr:
3740 case X86::BEXTR32rm: case X86::BEXTR64rm:
3741 case X86::BLSI32rr: case X86::BLSI32rm:
3742 case X86::BLSI64rr: case X86::BLSI64rm:
3743 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3744 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3745 case X86::BLSR32rr: case X86::BLSR32rm:
3746 case X86::BLSR64rr: case X86::BLSR64rm:
3747 case X86::BZHI32rr: case X86::BZHI32rm:
3748 case X86::BZHI64rr: case X86::BZHI64rm:
3749 case X86::LZCNT16rr: case X86::LZCNT16rm:
3750 case X86::LZCNT32rr: case X86::LZCNT32rm:
3751 case X86::LZCNT64rr: case X86::LZCNT64rm:
3752 case X86::POPCNT16rr:case X86::POPCNT16rm:
3753 case X86::POPCNT32rr:case X86::POPCNT32rm:
3754 case X86::POPCNT64rr:case X86::POPCNT64rm:
3755 case X86::TZCNT16rr: case X86::TZCNT16rm:
3756 case X86::TZCNT32rr: case X86::TZCNT32rm:
3757 case X86::TZCNT64rr: case X86::TZCNT64rm:
3762 /// isUseDefConvertible - check whether the use can be converted
3763 /// to remove a comparison against zero.
3764 static X86::CondCode isUseDefConvertible(MachineInstr *MI) {
3765 switch (MI->getOpcode()) {
3766 default: return X86::COND_INVALID;
3767 case X86::LZCNT16rr: case X86::LZCNT16rm:
3768 case X86::LZCNT32rr: case X86::LZCNT32rm:
3769 case X86::LZCNT64rr: case X86::LZCNT64rm:
3771 case X86::POPCNT16rr:case X86::POPCNT16rm:
3772 case X86::POPCNT32rr:case X86::POPCNT32rm:
3773 case X86::POPCNT64rr:case X86::POPCNT64rm:
3775 case X86::TZCNT16rr: case X86::TZCNT16rm:
3776 case X86::TZCNT32rr: case X86::TZCNT32rm:
3777 case X86::TZCNT64rr: case X86::TZCNT64rm:
3782 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3783 /// operates on the same source operands and sets flags in the same way as
3784 /// Compare; remove Compare if possible.
3786 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3787 int CmpMask, int CmpValue,
3788 const MachineRegisterInfo *MRI) const {
3789 // Check whether we can replace SUB with CMP.
3790 unsigned NewOpcode = 0;
3791 switch (CmpInstr->getOpcode()) {
3793 case X86::SUB64ri32:
3808 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3810 // There is no use of the destination register, we can replace SUB with CMP.
3811 switch (CmpInstr->getOpcode()) {
3812 default: llvm_unreachable("Unreachable!");
3813 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3814 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3815 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3816 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3817 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3818 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3819 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3820 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3821 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3822 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3823 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3824 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3825 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3826 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3827 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3829 CmpInstr->setDesc(get(NewOpcode));
3830 CmpInstr->RemoveOperand(0);
3831 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3832 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3833 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3838 // Get the unique definition of SrcReg.
3839 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3840 if (!MI) return false;
3842 // CmpInstr is the first instruction of the BB.
3843 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3845 // If we are comparing against zero, check whether we can use MI to update
3846 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3847 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3848 if (IsCmpZero && MI->getParent() != CmpInstr->getParent())
3851 // If we have a use of the source register between the def and our compare
3852 // instruction we can eliminate the compare iff the use sets EFLAGS in the
3854 bool ShouldUpdateCC = false;
3855 X86::CondCode NewCC = X86::COND_INVALID;
3856 if (IsCmpZero && !isDefConvertible(MI)) {
3857 // Scan forward from the use until we hit the use we're looking for or the
3858 // compare instruction.
3859 for (MachineBasicBlock::iterator J = MI;; ++J) {
3860 // Do we have a convertible instruction?
3861 NewCC = isUseDefConvertible(J);
3862 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
3863 J->getOperand(1).getReg() == SrcReg) {
3864 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
3865 ShouldUpdateCC = true; // Update CC later on.
3866 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
3867 // with the new def.
3877 // We are searching for an earlier instruction that can make CmpInstr
3878 // redundant and that instruction will be saved in Sub.
3879 MachineInstr *Sub = nullptr;
3880 const TargetRegisterInfo *TRI = &getRegisterInfo();
3882 // We iterate backward, starting from the instruction before CmpInstr and
3883 // stop when reaching the definition of a source register or done with the BB.
3884 // RI points to the instruction before CmpInstr.
3885 // If the definition is in this basic block, RE points to the definition;
3886 // otherwise, RE is the rend of the basic block.
3887 MachineBasicBlock::reverse_iterator
3888 RI = MachineBasicBlock::reverse_iterator(I),
3889 RE = CmpInstr->getParent() == MI->getParent() ?
3890 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3891 CmpInstr->getParent()->rend();
3892 MachineInstr *Movr0Inst = nullptr;
3893 for (; RI != RE; ++RI) {
3894 MachineInstr *Instr = &*RI;
3895 // Check whether CmpInstr can be made redundant by the current instruction.
3897 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3902 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3903 Instr->readsRegister(X86::EFLAGS, TRI)) {
3904 // This instruction modifies or uses EFLAGS.
3906 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3907 // They are safe to move up, if the definition to EFLAGS is dead and
3908 // earlier instructions do not read or write EFLAGS.
3909 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
3910 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3915 // We can't remove CmpInstr.
3920 // Return false if no candidates exist.
3921 if (!IsCmpZero && !Sub)
3924 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3925 Sub->getOperand(2).getReg() == SrcReg);
3927 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3928 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3929 // If we are done with the basic block, we need to check whether EFLAGS is
3931 bool IsSafe = false;
3932 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3933 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3934 for (++I; I != E; ++I) {
3935 const MachineInstr &Instr = *I;
3936 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3937 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3938 // We should check the usage if this instruction uses and updates EFLAGS.
3939 if (!UseEFLAGS && ModifyEFLAGS) {
3940 // It is safe to remove CmpInstr if EFLAGS is updated again.
3944 if (!UseEFLAGS && !ModifyEFLAGS)
3947 // EFLAGS is used by this instruction.
3948 X86::CondCode OldCC = X86::COND_INVALID;
3949 bool OpcIsSET = false;
3950 if (IsCmpZero || IsSwapped) {
3951 // We decode the condition code from opcode.
3952 if (Instr.isBranch())
3953 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3955 OldCC = getCondFromSETOpc(Instr.getOpcode());
3956 if (OldCC != X86::COND_INVALID)
3959 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3961 if (OldCC == X86::COND_INVALID) return false;
3966 case X86::COND_A: case X86::COND_AE:
3967 case X86::COND_B: case X86::COND_BE:
3968 case X86::COND_G: case X86::COND_GE:
3969 case X86::COND_L: case X86::COND_LE:
3970 case X86::COND_O: case X86::COND_NO:
3971 // CF and OF are used, we can't perform this optimization.
3975 // If we're updating the condition code check if we have to reverse the
3984 NewCC = GetOppositeBranchCondition(NewCC);
3987 } else if (IsSwapped) {
3988 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3989 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3990 // We swap the condition code and synthesize the new opcode.
3991 NewCC = getSwappedCondition(OldCC);
3992 if (NewCC == X86::COND_INVALID) return false;
3995 if ((ShouldUpdateCC || IsSwapped) && NewCC != OldCC) {
3996 // Synthesize the new opcode.
3997 bool HasMemoryOperand = Instr.hasOneMemOperand();
3999 if (Instr.isBranch())
4000 NewOpc = GetCondBranchFromCond(NewCC);
4002 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
4004 unsigned DstReg = Instr.getOperand(0).getReg();
4005 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
4009 // Push the MachineInstr to OpsToUpdate.
4010 // If it is safe to remove CmpInstr, the condition code of these
4011 // instructions will be modified.
4012 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
4014 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
4015 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
4021 // If EFLAGS is not killed nor re-defined, we should check whether it is
4022 // live-out. If it is live-out, do not optimize.
4023 if ((IsCmpZero || IsSwapped) && !IsSafe) {
4024 MachineBasicBlock *MBB = CmpInstr->getParent();
4025 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
4026 SE = MBB->succ_end(); SI != SE; ++SI)
4027 if ((*SI)->isLiveIn(X86::EFLAGS))
4031 // The instruction to be updated is either Sub or MI.
4032 Sub = IsCmpZero ? MI : Sub;
4033 // Move Movr0Inst to the appropriate place before Sub.
4035 // Look backwards until we find a def that doesn't use the current EFLAGS.
4037 MachineBasicBlock::reverse_iterator
4038 InsertI = MachineBasicBlock::reverse_iterator(++Def),
4039 InsertE = Sub->getParent()->rend();
4040 for (; InsertI != InsertE; ++InsertI) {
4041 MachineInstr *Instr = &*InsertI;
4042 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
4043 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
4044 Sub->getParent()->remove(Movr0Inst);
4045 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
4050 if (InsertI == InsertE)
4054 // Make sure Sub instruction defines EFLAGS and mark the def live.
4055 unsigned i = 0, e = Sub->getNumOperands();
4056 for (; i != e; ++i) {
4057 MachineOperand &MO = Sub->getOperand(i);
4058 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
4059 MO.setIsDead(false);
4063 assert(i != e && "Unable to locate a def EFLAGS operand");
4065 CmpInstr->eraseFromParent();
4067 // Modify the condition code of instructions in OpsToUpdate.
4068 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
4069 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
4073 /// optimizeLoadInstr - Try to remove the load by folding it to a register
4074 /// operand at the use. We fold the load instructions if load defines a virtual
4075 /// register, the virtual register is used once in the same BB, and the
4076 /// instructions in-between do not load or store, and have no side effects.
4077 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr *MI,
4078 const MachineRegisterInfo *MRI,
4079 unsigned &FoldAsLoadDefReg,
4080 MachineInstr *&DefMI) const {
4081 if (FoldAsLoadDefReg == 0)
4083 // To be conservative, if there exists another load, clear the load candidate.
4084 if (MI->mayLoad()) {
4085 FoldAsLoadDefReg = 0;
4089 // Check whether we can move DefMI here.
4090 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
4092 bool SawStore = false;
4093 if (!DefMI->isSafeToMove(this, nullptr, SawStore))
4096 // Collect information about virtual register operands of MI.
4097 unsigned SrcOperandId = 0;
4098 bool FoundSrcOperand = false;
4099 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
4100 MachineOperand &MO = MI->getOperand(i);
4103 unsigned Reg = MO.getReg();
4104 if (Reg != FoldAsLoadDefReg)
4106 // Do not fold if we have a subreg use or a def or multiple uses.
4107 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
4111 FoundSrcOperand = true;
4113 if (!FoundSrcOperand)
4116 // Check whether we can fold the def into SrcOperandId.
4117 SmallVector<unsigned, 8> Ops;
4118 Ops.push_back(SrcOperandId);
4119 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
4121 FoldAsLoadDefReg = 0;
4128 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
4129 /// instruction with two undef reads of the register being defined. This is
4130 /// used for mapping:
4133 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
4135 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
4136 const MCInstrDesc &Desc) {
4137 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
4138 unsigned Reg = MIB->getOperand(0).getReg();
4141 // MachineInstr::addOperand() will insert explicit operands before any
4142 // implicit operands.
4143 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4144 // But we don't trust that.
4145 assert(MIB->getOperand(1).getReg() == Reg &&
4146 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
4150 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
4151 // code sequence is needed for other targets.
4152 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
4153 const TargetInstrInfo &TII) {
4154 MachineBasicBlock &MBB = *MIB->getParent();
4155 DebugLoc DL = MIB->getDebugLoc();
4156 unsigned Reg = MIB->getOperand(0).getReg();
4157 const GlobalValue *GV =
4158 cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
4159 unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
4160 MachineMemOperand *MMO = MBB.getParent()->
4161 getMachineMemOperand(MachinePointerInfo::getGOT(), Flag, 8, 8);
4162 MachineBasicBlock::iterator I = MIB.getInstr();
4164 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4165 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4166 .addMemOperand(MMO);
4167 MIB->setDebugLoc(DL);
4168 MIB->setDesc(TII.get(X86::MOV64rm));
4169 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4172 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
4173 bool HasAVX = Subtarget.hasAVX();
4174 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
4175 switch (MI->getOpcode()) {
4177 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4179 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4180 case X86::SETB_C16r:
4181 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4182 case X86::SETB_C32r:
4183 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4184 case X86::SETB_C64r:
4185 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4189 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4191 assert(HasAVX && "AVX not supported");
4192 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
4193 case X86::AVX512_512_SET0:
4194 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4195 case X86::V_SETALLONES:
4196 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4197 case X86::AVX2_SETALLONES:
4198 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4199 case X86::TEST8ri_NOREX:
4200 MI->setDesc(get(X86::TEST8ri));
4203 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
4205 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
4206 case TargetOpcode::LOAD_STACK_GUARD:
4207 expandLoadStackGuard(MIB, *this);
4213 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4214 const SmallVectorImpl<MachineOperand> &MOs,
4216 const TargetInstrInfo &TII) {
4217 // Create the base instruction with the memory operand as the first part.
4218 // Omit the implicit operands, something BuildMI can't do.
4219 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4220 MI->getDebugLoc(), true);
4221 MachineInstrBuilder MIB(MF, NewMI);
4222 unsigned NumAddrOps = MOs.size();
4223 for (unsigned i = 0; i != NumAddrOps; ++i)
4224 MIB.addOperand(MOs[i]);
4225 if (NumAddrOps < 4) // FrameIndex only
4228 // Loop over the rest of the ri operands, converting them over.
4229 unsigned NumOps = MI->getDesc().getNumOperands()-2;
4230 for (unsigned i = 0; i != NumOps; ++i) {
4231 MachineOperand &MO = MI->getOperand(i+2);
4234 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
4235 MachineOperand &MO = MI->getOperand(i);
4241 static MachineInstr *FuseInst(MachineFunction &MF,
4242 unsigned Opcode, unsigned OpNo,
4243 const SmallVectorImpl<MachineOperand> &MOs,
4244 MachineInstr *MI, const TargetInstrInfo &TII) {
4245 // Omit the implicit operands, something BuildMI can't do.
4246 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4247 MI->getDebugLoc(), true);
4248 MachineInstrBuilder MIB(MF, NewMI);
4250 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4251 MachineOperand &MO = MI->getOperand(i);
4253 assert(MO.isReg() && "Expected to fold into reg operand!");
4254 unsigned NumAddrOps = MOs.size();
4255 for (unsigned i = 0; i != NumAddrOps; ++i)
4256 MIB.addOperand(MOs[i]);
4257 if (NumAddrOps < 4) // FrameIndex only
4266 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4267 const SmallVectorImpl<MachineOperand> &MOs,
4269 MachineFunction &MF = *MI->getParent()->getParent();
4270 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
4272 unsigned NumAddrOps = MOs.size();
4273 for (unsigned i = 0; i != NumAddrOps; ++i)
4274 MIB.addOperand(MOs[i]);
4275 if (NumAddrOps < 4) // FrameIndex only
4277 return MIB.addImm(0);
4281 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4282 MachineInstr *MI, unsigned i,
4283 const SmallVectorImpl<MachineOperand> &MOs,
4284 unsigned Size, unsigned Align,
4285 bool AllowCommute) const {
4286 const DenseMap<unsigned,
4287 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4288 bool isCallRegIndirect = Subtarget.callRegIndirect();
4289 bool isTwoAddrFold = false;
4291 // Atom favors register form of call. So, we do not fold loads into calls
4292 // when X86Subtarget is Atom.
4293 if (isCallRegIndirect &&
4294 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
4298 unsigned NumOps = MI->getDesc().getNumOperands();
4299 bool isTwoAddr = NumOps > 1 &&
4300 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4302 // FIXME: AsmPrinter doesn't know how to handle
4303 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4304 if (MI->getOpcode() == X86::ADD32ri &&
4305 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4308 MachineInstr *NewMI = nullptr;
4309 // Folding a memory location into the two-address part of a two-address
4310 // instruction is different than folding it other places. It requires
4311 // replacing the *two* registers with the memory location.
4312 if (isTwoAddr && NumOps >= 2 && i < 2 &&
4313 MI->getOperand(0).isReg() &&
4314 MI->getOperand(1).isReg() &&
4315 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
4316 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4317 isTwoAddrFold = true;
4318 } else if (i == 0) { // If operand 0
4319 if (MI->getOpcode() == X86::MOV32r0) {
4320 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
4325 OpcodeTablePtr = &RegOp2MemOpTable0;
4326 } else if (i == 1) {
4327 OpcodeTablePtr = &RegOp2MemOpTable1;
4328 } else if (i == 2) {
4329 OpcodeTablePtr = &RegOp2MemOpTable2;
4330 } else if (i == 3) {
4331 OpcodeTablePtr = &RegOp2MemOpTable3;
4332 } else if (i == 4) {
4333 OpcodeTablePtr = &RegOp2MemOpTable4;
4336 // If table selected...
4337 if (OpcodeTablePtr) {
4338 // Find the Opcode to fuse
4339 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4340 OpcodeTablePtr->find(MI->getOpcode());
4341 if (I != OpcodeTablePtr->end()) {
4342 unsigned Opcode = I->second.first;
4343 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4344 if (Align < MinAlign)
4346 bool NarrowToMOV32rm = false;
4348 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
4349 if (Size < RCSize) {
4350 // Check if it's safe to fold the load. If the size of the object is
4351 // narrower than the load width, then it's not.
4352 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4354 // If this is a 64-bit load, but the spill slot is 32, then we can do
4355 // a 32-bit load which is implicitly zero-extended. This likely is
4356 // due to live interval analysis remat'ing a load from stack slot.
4357 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4359 Opcode = X86::MOV32rm;
4360 NarrowToMOV32rm = true;
4365 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
4367 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
4369 if (NarrowToMOV32rm) {
4370 // If this is the special case where we use a MOV32rm to load a 32-bit
4371 // value and zero-extend the top bits. Change the destination register
4373 unsigned DstReg = NewMI->getOperand(0).getReg();
4374 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4375 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4377 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4383 // If the instruction and target operand are commutable, commute the
4384 // instruction and try again.
4386 unsigned OriginalOpIdx = i, CommuteOpIdx1, CommuteOpIdx2;
4387 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4388 bool HasDef = MI->getDesc().getNumDefs();
4389 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
4390 unsigned Reg1 = MI->getOperand(CommuteOpIdx1).getReg();
4391 unsigned Reg2 = MI->getOperand(CommuteOpIdx2).getReg();
4393 0 == MI->getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
4395 0 == MI->getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
4397 // If either of the commutable operands are tied to the destination
4398 // then we can not commute + fold.
4399 if ((HasDef && Reg0 == Reg1 && Tied0) ||
4400 (HasDef && Reg0 == Reg2 && Tied1))
4403 if ((CommuteOpIdx1 == OriginalOpIdx) ||
4404 (CommuteOpIdx2 == OriginalOpIdx)) {
4405 MachineInstr *CommutedMI = commuteInstruction(MI, false);
4407 // Unable to commute.
4410 if (CommutedMI != MI) {
4411 // New instruction. We can't fold from this.
4412 CommutedMI->eraseFromParent();
4416 // Attempt to fold with the commuted version of the instruction.
4417 unsigned CommuteOp =
4418 (CommuteOpIdx1 == OriginalOpIdx ? CommuteOpIdx2 : CommuteOpIdx1);
4419 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOp, MOs, Size, Align,
4420 /*AllowCommute=*/false);
4424 // Folding failed again - undo the commute before returning.
4425 MachineInstr *UncommutedMI = commuteInstruction(MI, false);
4426 if (!UncommutedMI) {
4427 // Unable to commute.
4430 if (UncommutedMI != MI) {
4431 // New instruction. It doesn't need to be kept.
4432 UncommutedMI->eraseFromParent();
4436 // Return here to prevent duplicate fuse failure report.
4443 if (PrintFailedFusing && !MI->isCopy())
4444 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
4448 /// hasPartialRegUpdate - Return true for all instructions that only update
4449 /// the first 32 or 64-bits of the destination register and leave the rest
4450 /// unmodified. This can be used to avoid folding loads if the instructions
4451 /// only update part of the destination register, and the non-updated part is
4452 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4453 /// instructions breaks the partial register dependency and it can improve
4454 /// performance. e.g.:
4456 /// movss (%rdi), %xmm0
4457 /// cvtss2sd %xmm0, %xmm0
4460 /// cvtss2sd (%rdi), %xmm0
4462 /// FIXME: This should be turned into a TSFlags.
4464 static bool hasPartialRegUpdate(unsigned Opcode) {
4466 case X86::CVTSI2SSrr:
4467 case X86::CVTSI2SSrm:
4468 case X86::CVTSI2SS64rr:
4469 case X86::CVTSI2SS64rm:
4470 case X86::CVTSI2SDrr:
4471 case X86::CVTSI2SDrm:
4472 case X86::CVTSI2SD64rr:
4473 case X86::CVTSI2SD64rm:
4474 case X86::CVTSD2SSrr:
4475 case X86::CVTSD2SSrm:
4476 case X86::Int_CVTSD2SSrr:
4477 case X86::Int_CVTSD2SSrm:
4478 case X86::CVTSS2SDrr:
4479 case X86::CVTSS2SDrm:
4480 case X86::Int_CVTSS2SDrr:
4481 case X86::Int_CVTSS2SDrm:
4484 case X86::RCPSSr_Int:
4485 case X86::RCPSSm_Int:
4488 case X86::ROUNDSDr_Int:
4491 case X86::ROUNDSSr_Int:
4494 case X86::RSQRTSSr_Int:
4495 case X86::RSQRTSSm_Int:
4498 case X86::SQRTSSr_Int:
4499 case X86::SQRTSSm_Int:
4502 case X86::SQRTSDr_Int:
4503 case X86::SQRTSDm_Int:
4510 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4511 /// instructions we would like before a partial register update.
4512 unsigned X86InstrInfo::
4513 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4514 const TargetRegisterInfo *TRI) const {
4515 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4518 // If MI is marked as reading Reg, the partial register update is wanted.
4519 const MachineOperand &MO = MI->getOperand(0);
4520 unsigned Reg = MO.getReg();
4521 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4522 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4525 if (MI->readsRegister(Reg, TRI))
4529 // If any of the preceding 16 instructions are reading Reg, insert a
4530 // dependency breaking instruction. The magic number is based on a few
4531 // Nehalem experiments.
4535 // Return true for any instruction the copies the high bits of the first source
4536 // operand into the unused high bits of the destination operand.
4537 static bool hasUndefRegUpdate(unsigned Opcode) {
4539 case X86::VCVTSI2SSrr:
4540 case X86::VCVTSI2SSrm:
4541 case X86::Int_VCVTSI2SSrr:
4542 case X86::Int_VCVTSI2SSrm:
4543 case X86::VCVTSI2SS64rr:
4544 case X86::VCVTSI2SS64rm:
4545 case X86::Int_VCVTSI2SS64rr:
4546 case X86::Int_VCVTSI2SS64rm:
4547 case X86::VCVTSI2SDrr:
4548 case X86::VCVTSI2SDrm:
4549 case X86::Int_VCVTSI2SDrr:
4550 case X86::Int_VCVTSI2SDrm:
4551 case X86::VCVTSI2SD64rr:
4552 case X86::VCVTSI2SD64rm:
4553 case X86::Int_VCVTSI2SD64rr:
4554 case X86::Int_VCVTSI2SD64rm:
4555 case X86::VCVTSD2SSrr:
4556 case X86::VCVTSD2SSrm:
4557 case X86::Int_VCVTSD2SSrr:
4558 case X86::Int_VCVTSD2SSrm:
4559 case X86::VCVTSS2SDrr:
4560 case X86::VCVTSS2SDrm:
4561 case X86::Int_VCVTSS2SDrr:
4562 case X86::Int_VCVTSS2SDrm:
4565 case X86::VRCPSSm_Int:
4566 case X86::VROUNDSDr:
4567 case X86::VROUNDSDm:
4568 case X86::VROUNDSDr_Int:
4569 case X86::VROUNDSSr:
4570 case X86::VROUNDSSm:
4571 case X86::VROUNDSSr_Int:
4572 case X86::VRSQRTSSr:
4573 case X86::VRSQRTSSm:
4574 case X86::VRSQRTSSm_Int:
4577 case X86::VSQRTSSm_Int:
4580 case X86::VSQRTSDm_Int:
4582 case X86::VCVTSD2SSZrr:
4583 case X86::VCVTSD2SSZrm:
4584 case X86::VCVTSS2SDZrr:
4585 case X86::VCVTSS2SDZrm:
4592 /// Inform the ExeDepsFix pass how many idle instructions we would like before
4593 /// certain undef register reads.
4595 /// This catches the VCVTSI2SD family of instructions:
4597 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
4599 /// We should to be careful *not* to catch VXOR idioms which are presumably
4600 /// handled specially in the pipeline:
4602 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
4604 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4605 /// high bits that are passed-through are not live.
4606 unsigned X86InstrInfo::
4607 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
4608 const TargetRegisterInfo *TRI) const {
4609 if (!hasUndefRegUpdate(MI->getOpcode()))
4612 // Set the OpNum parameter to the first source operand.
4615 const MachineOperand &MO = MI->getOperand(OpNum);
4616 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4617 // Use the same magic number as getPartialRegUpdateClearance.
4624 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4625 const TargetRegisterInfo *TRI) const {
4626 unsigned Reg = MI->getOperand(OpNum).getReg();
4627 // If MI kills this register, the false dependence is already broken.
4628 if (MI->killsRegister(Reg, TRI))
4630 if (X86::VR128RegClass.contains(Reg)) {
4631 // These instructions are all floating point domain, so xorps is the best
4633 bool HasAVX = Subtarget.hasAVX();
4634 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4635 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4636 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4637 } else if (X86::VR256RegClass.contains(Reg)) {
4638 // Use vxorps to clear the full ymm register.
4639 // It wants to read and write the xmm sub-register.
4640 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4641 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4642 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4643 .addReg(Reg, RegState::ImplicitDefine);
4646 MI->addRegisterKilled(Reg, TRI, true);
4650 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
4651 const SmallVectorImpl<unsigned> &Ops,
4652 int FrameIndex) const {
4653 // Check switch flag
4654 if (NoFusing) return nullptr;
4656 // Unless optimizing for size, don't fold to avoid partial
4657 // register update stalls
4658 if (!MF.getFunction()->getAttributes().
4659 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4660 hasPartialRegUpdate(MI->getOpcode()))
4663 const MachineFrameInfo *MFI = MF.getFrameInfo();
4664 unsigned Size = MFI->getObjectSize(FrameIndex);
4665 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
4666 // If the function stack isn't realigned we don't want to fold instructions
4667 // that need increased alignment.
4668 if (!RI.needsStackRealignment(MF))
4669 Alignment = std::min(Alignment, MF.getTarget()
4671 ->getFrameLowering()
4672 ->getStackAlignment());
4673 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4674 unsigned NewOpc = 0;
4675 unsigned RCSize = 0;
4676 switch (MI->getOpcode()) {
4677 default: return nullptr;
4678 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
4679 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4680 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4681 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4683 // Check if it's safe to fold the load. If the size of the object is
4684 // narrower than the load width, then it's not.
4687 // Change to CMPXXri r, 0 first.
4688 MI->setDesc(get(NewOpc));
4689 MI->getOperand(1).ChangeToImmediate(0);
4690 } else if (Ops.size() != 1)
4693 SmallVector<MachineOperand,4> MOs;
4694 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
4695 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs,
4696 Size, Alignment, /*AllowCommute=*/true);
4699 static bool isPartialRegisterLoad(const MachineInstr &LoadMI,
4700 const MachineFunction &MF) {
4701 unsigned Opc = LoadMI.getOpcode();
4703 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg())->getSize();
4705 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm) && RegSize > 4)
4706 // These instructions only load 32 bits, we can't fold them if the
4707 // destination register is wider than 32 bits (4 bytes).
4710 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm) && RegSize > 8)
4711 // These instructions only load 64 bits, we can't fold them if the
4712 // destination register is wider than 64 bits (8 bytes).
4718 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4720 const SmallVectorImpl<unsigned> &Ops,
4721 MachineInstr *LoadMI) const {
4722 // If loading from a FrameIndex, fold directly from the FrameIndex.
4723 unsigned NumOps = LoadMI->getDesc().getNumOperands();
4725 if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
4726 if (isPartialRegisterLoad(*LoadMI, MF))
4728 return foldMemoryOperandImpl(MF, MI, Ops, FrameIndex);
4731 // Check switch flag
4732 if (NoFusing) return nullptr;
4734 // Unless optimizing for size, don't fold to avoid partial
4735 // register update stalls
4736 if (!MF.getFunction()->getAttributes().
4737 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4738 hasPartialRegUpdate(MI->getOpcode()))
4741 // Determine the alignment of the load.
4742 unsigned Alignment = 0;
4743 if (LoadMI->hasOneMemOperand())
4744 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
4746 switch (LoadMI->getOpcode()) {
4747 case X86::AVX2_SETALLONES:
4752 case X86::V_SETALLONES:
4764 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4765 unsigned NewOpc = 0;
4766 switch (MI->getOpcode()) {
4767 default: return nullptr;
4768 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
4769 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4770 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4771 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
4773 // Change to CMPXXri r, 0 first.
4774 MI->setDesc(get(NewOpc));
4775 MI->getOperand(1).ChangeToImmediate(0);
4776 } else if (Ops.size() != 1)
4779 // Make sure the subregisters match.
4780 // Otherwise we risk changing the size of the load.
4781 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4784 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
4785 switch (LoadMI->getOpcode()) {
4787 case X86::V_SETALLONES:
4788 case X86::AVX2_SETALLONES:
4791 case X86::FsFLD0SS: {
4792 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
4793 // Create a constant-pool entry and operands to load from it.
4795 // Medium and large mode can't fold loads this way.
4796 if (MF.getTarget().getCodeModel() != CodeModel::Small &&
4797 MF.getTarget().getCodeModel() != CodeModel::Kernel)
4800 // x86-32 PIC requires a PIC base register for constant pools.
4801 unsigned PICBase = 0;
4802 if (MF.getTarget().getRelocationModel() == Reloc::PIC_) {
4803 if (Subtarget.is64Bit())
4806 // FIXME: PICBase = getGlobalBaseReg(&MF);
4807 // This doesn't work for several reasons.
4808 // 1. GlobalBaseReg may have been spilled.
4809 // 2. It may not be live at MI.
4813 // Create a constant-pool entry.
4814 MachineConstantPool &MCP = *MF.getConstantPool();
4816 unsigned Opc = LoadMI->getOpcode();
4817 if (Opc == X86::FsFLD0SS)
4818 Ty = Type::getFloatTy(MF.getFunction()->getContext());
4819 else if (Opc == X86::FsFLD0SD)
4820 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
4821 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
4822 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
4824 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
4826 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
4827 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4828 Constant::getNullValue(Ty);
4829 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
4831 // Create operands to load from the constant pool entry.
4832 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4833 MOs.push_back(MachineOperand::CreateImm(1));
4834 MOs.push_back(MachineOperand::CreateReg(0, false));
4835 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
4836 MOs.push_back(MachineOperand::CreateReg(0, false));
4840 if (isPartialRegisterLoad(*LoadMI, MF))
4843 // Folding a normal load. Just copy the load's address operands.
4844 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
4845 MOs.push_back(LoadMI->getOperand(i));
4849 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs,
4850 /*Size=*/0, Alignment, /*AllowCommute=*/true);
4854 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4855 const SmallVectorImpl<unsigned> &Ops) const {
4856 // Check switch flag
4857 if (NoFusing) return 0;
4859 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4860 switch (MI->getOpcode()) {
4861 default: return false;
4868 // FIXME: AsmPrinter doesn't know how to handle
4869 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4870 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4876 if (Ops.size() != 1)
4879 unsigned OpNum = Ops[0];
4880 unsigned Opc = MI->getOpcode();
4881 unsigned NumOps = MI->getDesc().getNumOperands();
4882 bool isTwoAddr = NumOps > 1 &&
4883 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4885 // Folding a memory location into the two-address part of a two-address
4886 // instruction is different than folding it other places. It requires
4887 // replacing the *two* registers with the memory location.
4888 const DenseMap<unsigned,
4889 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4890 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4891 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4892 } else if (OpNum == 0) { // If operand 0
4893 if (Opc == X86::MOV32r0)
4896 OpcodeTablePtr = &RegOp2MemOpTable0;
4897 } else if (OpNum == 1) {
4898 OpcodeTablePtr = &RegOp2MemOpTable1;
4899 } else if (OpNum == 2) {
4900 OpcodeTablePtr = &RegOp2MemOpTable2;
4901 } else if (OpNum == 3) {
4902 OpcodeTablePtr = &RegOp2MemOpTable3;
4905 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4907 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
4910 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4911 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4912 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4913 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4914 MemOp2RegOpTable.find(MI->getOpcode());
4915 if (I == MemOp2RegOpTable.end())
4917 unsigned Opc = I->second.first;
4918 unsigned Index = I->second.second & TB_INDEX_MASK;
4919 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4920 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4921 if (UnfoldLoad && !FoldedLoad)
4923 UnfoldLoad &= FoldedLoad;
4924 if (UnfoldStore && !FoldedStore)
4926 UnfoldStore &= FoldedStore;
4928 const MCInstrDesc &MCID = get(Opc);
4929 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4930 if (!MI->hasOneMemOperand() &&
4931 RC == &X86::VR128RegClass &&
4932 !Subtarget.isUnalignedMemAccessFast())
4933 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4934 // conservatively assume the address is unaligned. That's bad for
4937 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4938 SmallVector<MachineOperand,2> BeforeOps;
4939 SmallVector<MachineOperand,2> AfterOps;
4940 SmallVector<MachineOperand,4> ImpOps;
4941 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4942 MachineOperand &Op = MI->getOperand(i);
4943 if (i >= Index && i < Index + X86::AddrNumOperands)
4944 AddrOps.push_back(Op);
4945 else if (Op.isReg() && Op.isImplicit())
4946 ImpOps.push_back(Op);
4948 BeforeOps.push_back(Op);
4950 AfterOps.push_back(Op);
4953 // Emit the load instruction.
4955 std::pair<MachineInstr::mmo_iterator,
4956 MachineInstr::mmo_iterator> MMOs =
4957 MF.extractLoadMemRefs(MI->memoperands_begin(),
4958 MI->memoperands_end());
4959 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4961 // Address operands cannot be marked isKill.
4962 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4963 MachineOperand &MO = NewMIs[0]->getOperand(i);
4965 MO.setIsKill(false);
4970 // Emit the data processing instruction.
4971 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4972 MachineInstrBuilder MIB(MF, DataMI);
4975 MIB.addReg(Reg, RegState::Define);
4976 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4977 MIB.addOperand(BeforeOps[i]);
4980 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4981 MIB.addOperand(AfterOps[i]);
4982 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4983 MachineOperand &MO = ImpOps[i];
4984 MIB.addReg(MO.getReg(),
4985 getDefRegState(MO.isDef()) |
4986 RegState::Implicit |
4987 getKillRegState(MO.isKill()) |
4988 getDeadRegState(MO.isDead()) |
4989 getUndefRegState(MO.isUndef()));
4991 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4992 switch (DataMI->getOpcode()) {
4994 case X86::CMP64ri32:
5001 MachineOperand &MO0 = DataMI->getOperand(0);
5002 MachineOperand &MO1 = DataMI->getOperand(1);
5003 if (MO1.getImm() == 0) {
5005 switch (DataMI->getOpcode()) {
5006 default: llvm_unreachable("Unreachable!");
5008 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5010 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
5012 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
5013 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
5015 DataMI->setDesc(get(NewOpc));
5016 MO1.ChangeToRegister(MO0.getReg(), false);
5020 NewMIs.push_back(DataMI);
5022 // Emit the store instruction.
5024 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5025 std::pair<MachineInstr::mmo_iterator,
5026 MachineInstr::mmo_iterator> MMOs =
5027 MF.extractStoreMemRefs(MI->memoperands_begin(),
5028 MI->memoperands_end());
5029 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
5036 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5037 SmallVectorImpl<SDNode*> &NewNodes) const {
5038 if (!N->isMachineOpcode())
5041 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5042 MemOp2RegOpTable.find(N->getMachineOpcode());
5043 if (I == MemOp2RegOpTable.end())
5045 unsigned Opc = I->second.first;
5046 unsigned Index = I->second.second & TB_INDEX_MASK;
5047 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5048 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5049 const MCInstrDesc &MCID = get(Opc);
5050 MachineFunction &MF = DAG.getMachineFunction();
5051 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5052 unsigned NumDefs = MCID.NumDefs;
5053 std::vector<SDValue> AddrOps;
5054 std::vector<SDValue> BeforeOps;
5055 std::vector<SDValue> AfterOps;
5057 unsigned NumOps = N->getNumOperands();
5058 for (unsigned i = 0; i != NumOps-1; ++i) {
5059 SDValue Op = N->getOperand(i);
5060 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5061 AddrOps.push_back(Op);
5062 else if (i < Index-NumDefs)
5063 BeforeOps.push_back(Op);
5064 else if (i > Index-NumDefs)
5065 AfterOps.push_back(Op);
5067 SDValue Chain = N->getOperand(NumOps-1);
5068 AddrOps.push_back(Chain);
5070 // Emit the load instruction.
5071 SDNode *Load = nullptr;
5073 EVT VT = *RC->vt_begin();
5074 std::pair<MachineInstr::mmo_iterator,
5075 MachineInstr::mmo_iterator> MMOs =
5076 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
5077 cast<MachineSDNode>(N)->memoperands_end());
5078 if (!(*MMOs.first) &&
5079 RC == &X86::VR128RegClass &&
5080 !Subtarget.isUnalignedMemAccessFast())
5081 // Do not introduce a slow unaligned load.
5083 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
5084 bool isAligned = (*MMOs.first) &&
5085 (*MMOs.first)->getAlignment() >= Alignment;
5086 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
5087 VT, MVT::Other, AddrOps);
5088 NewNodes.push_back(Load);
5090 // Preserve memory reference information.
5091 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
5094 // Emit the data processing instruction.
5095 std::vector<EVT> VTs;
5096 const TargetRegisterClass *DstRC = nullptr;
5097 if (MCID.getNumDefs() > 0) {
5098 DstRC = getRegClass(MCID, 0, &RI, MF);
5099 VTs.push_back(*DstRC->vt_begin());
5101 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5102 EVT VT = N->getValueType(i);
5103 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5107 BeforeOps.push_back(SDValue(Load, 0));
5108 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
5109 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5110 NewNodes.push_back(NewNode);
5112 // Emit the store instruction.
5115 AddrOps.push_back(SDValue(NewNode, 0));
5116 AddrOps.push_back(Chain);
5117 std::pair<MachineInstr::mmo_iterator,
5118 MachineInstr::mmo_iterator> MMOs =
5119 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
5120 cast<MachineSDNode>(N)->memoperands_end());
5121 if (!(*MMOs.first) &&
5122 RC == &X86::VR128RegClass &&
5123 !Subtarget.isUnalignedMemAccessFast())
5124 // Do not introduce a slow unaligned store.
5126 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
5127 bool isAligned = (*MMOs.first) &&
5128 (*MMOs.first)->getAlignment() >= Alignment;
5130 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5131 dl, MVT::Other, AddrOps);
5132 NewNodes.push_back(Store);
5134 // Preserve memory reference information.
5135 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
5141 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5142 bool UnfoldLoad, bool UnfoldStore,
5143 unsigned *LoadRegIndex) const {
5144 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5145 MemOp2RegOpTable.find(Opc);
5146 if (I == MemOp2RegOpTable.end())
5148 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5149 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5150 if (UnfoldLoad && !FoldedLoad)
5152 if (UnfoldStore && !FoldedStore)
5155 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
5156 return I->second.first;
5160 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5161 int64_t &Offset1, int64_t &Offset2) const {
5162 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5164 unsigned Opc1 = Load1->getMachineOpcode();
5165 unsigned Opc2 = Load2->getMachineOpcode();
5167 default: return false;
5177 case X86::MMX_MOVD64rm:
5178 case X86::MMX_MOVQ64rm:
5179 case X86::FsMOVAPSrm:
5180 case X86::FsMOVAPDrm:
5186 // AVX load instructions
5189 case X86::FsVMOVAPSrm:
5190 case X86::FsVMOVAPDrm:
5191 case X86::VMOVAPSrm:
5192 case X86::VMOVUPSrm:
5193 case X86::VMOVAPDrm:
5194 case X86::VMOVDQArm:
5195 case X86::VMOVDQUrm:
5196 case X86::VMOVAPSYrm:
5197 case X86::VMOVUPSYrm:
5198 case X86::VMOVAPDYrm:
5199 case X86::VMOVDQAYrm:
5200 case X86::VMOVDQUYrm:
5204 default: return false;
5214 case X86::MMX_MOVD64rm:
5215 case X86::MMX_MOVQ64rm:
5216 case X86::FsMOVAPSrm:
5217 case X86::FsMOVAPDrm:
5223 // AVX load instructions
5226 case X86::FsVMOVAPSrm:
5227 case X86::FsVMOVAPDrm:
5228 case X86::VMOVAPSrm:
5229 case X86::VMOVUPSrm:
5230 case X86::VMOVAPDrm:
5231 case X86::VMOVDQArm:
5232 case X86::VMOVDQUrm:
5233 case X86::VMOVAPSYrm:
5234 case X86::VMOVUPSYrm:
5235 case X86::VMOVAPDYrm:
5236 case X86::VMOVDQAYrm:
5237 case X86::VMOVDQUYrm:
5241 // Check if chain operands and base addresses match.
5242 if (Load1->getOperand(0) != Load2->getOperand(0) ||
5243 Load1->getOperand(5) != Load2->getOperand(5))
5245 // Segment operands should match as well.
5246 if (Load1->getOperand(4) != Load2->getOperand(4))
5248 // Scale should be 1, Index should be Reg0.
5249 if (Load1->getOperand(1) == Load2->getOperand(1) &&
5250 Load1->getOperand(2) == Load2->getOperand(2)) {
5251 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
5254 // Now let's examine the displacements.
5255 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
5256 isa<ConstantSDNode>(Load2->getOperand(3))) {
5257 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
5258 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
5265 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5266 int64_t Offset1, int64_t Offset2,
5267 unsigned NumLoads) const {
5268 assert(Offset2 > Offset1);
5269 if ((Offset2 - Offset1) / 8 > 64)
5272 unsigned Opc1 = Load1->getMachineOpcode();
5273 unsigned Opc2 = Load2->getMachineOpcode();
5275 return false; // FIXME: overly conservative?
5282 case X86::MMX_MOVD64rm:
5283 case X86::MMX_MOVQ64rm:
5287 EVT VT = Load1->getValueType(0);
5288 switch (VT.getSimpleVT().SimpleTy) {
5290 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5291 // have 16 of them to play with.
5292 if (Subtarget.is64Bit()) {
5295 } else if (NumLoads) {
5313 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
5314 MachineInstr *Second) const {
5315 // Check if this processor supports macro-fusion. Since this is a minor
5316 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
5317 // proxy for SandyBridge+.
5318 if (!Subtarget.hasAVX())
5327 switch(Second->getOpcode()) {
5350 FuseKind = FuseTest;
5353 switch (First->getOpcode()) {
5363 case X86::TEST32i32:
5364 case X86::TEST64i32:
5365 case X86::TEST64ri32:
5370 case X86::TEST8ri_NOREX:
5382 case X86::AND64ri32:
5402 case X86::CMP64ri32:
5413 case X86::ADD16ri8_DB:
5414 case X86::ADD16ri_DB:
5417 case X86::ADD16rr_DB:
5421 case X86::ADD32ri8_DB:
5422 case X86::ADD32ri_DB:
5425 case X86::ADD32rr_DB:
5427 case X86::ADD64ri32:
5428 case X86::ADD64ri32_DB:
5430 case X86::ADD64ri8_DB:
5433 case X86::ADD64rr_DB:
5451 case X86::SUB64ri32:
5459 return FuseKind == FuseCmp || FuseKind == FuseInc;
5468 return FuseKind == FuseInc;
5473 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5474 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5475 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5476 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
5478 Cond[0].setImm(GetOppositeBranchCondition(CC));
5483 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5484 // FIXME: Return false for x87 stack register classes for now. We can't
5485 // allow any loads of these registers before FpGet_ST0_80.
5486 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
5487 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
5490 /// getGlobalBaseReg - Return a virtual register initialized with the
5491 /// the global base register value. Output instructions required to
5492 /// initialize the register in the function entry block, if necessary.
5494 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5496 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5497 assert(!Subtarget.is64Bit() &&
5498 "X86-64 PIC uses RIP relative addressing");
5500 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5501 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5502 if (GlobalBaseReg != 0)
5503 return GlobalBaseReg;
5505 // Create the register. The code to initialize it is inserted
5506 // later, by the CGBR pass (below).
5507 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5508 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
5509 X86FI->setGlobalBaseReg(GlobalBaseReg);
5510 return GlobalBaseReg;
5513 // These are the replaceable SSE instructions. Some of these have Int variants
5514 // that we don't include here. We don't want to replace instructions selected
5516 static const uint16_t ReplaceableInstrs[][3] = {
5517 //PackedSingle PackedDouble PackedInt
5518 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
5519 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
5520 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
5521 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
5522 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
5523 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
5524 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
5525 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
5526 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
5527 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
5528 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
5529 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
5530 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
5531 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
5532 // AVX 128-bit support
5533 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
5534 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
5535 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
5536 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5537 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5538 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5539 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5540 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5541 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5542 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5543 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5544 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
5545 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
5546 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
5547 // AVX 256-bit support
5548 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
5549 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
5550 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
5551 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
5552 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
5553 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
5556 static const uint16_t ReplaceableInstrsAVX2[][3] = {
5557 //PackedSingle PackedDouble PackedInt
5558 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
5559 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
5560 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
5561 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
5562 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
5563 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
5564 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
5565 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
5566 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
5567 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
5568 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
5569 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
5570 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
5571 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
5572 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
5573 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
5574 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
5575 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
5576 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
5577 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
5580 // FIXME: Some shuffle and unpack instructions have equivalents in different
5581 // domains, but they require a bit more work than just switching opcodes.
5583 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
5584 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
5585 if (ReplaceableInstrs[i][domain-1] == opcode)
5586 return ReplaceableInstrs[i];
5590 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
5591 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
5592 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
5593 return ReplaceableInstrsAVX2[i];
5597 std::pair<uint16_t, uint16_t>
5598 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
5599 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5600 bool hasAVX2 = Subtarget.hasAVX2();
5601 uint16_t validDomains = 0;
5602 if (domain && lookup(MI->getOpcode(), domain))
5604 else if (domain && lookupAVX2(MI->getOpcode(), domain))
5605 validDomains = hasAVX2 ? 0xe : 0x6;
5606 return std::make_pair(domain, validDomains);
5609 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
5610 assert(Domain>0 && Domain<4 && "Invalid execution domain");
5611 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5612 assert(dom && "Not an SSE instruction");
5613 const uint16_t *table = lookup(MI->getOpcode(), dom);
5614 if (!table) { // try the other table
5615 assert((Subtarget.hasAVX2() || Domain < 3) &&
5616 "256-bit vector operations only available in AVX2");
5617 table = lookupAVX2(MI->getOpcode(), dom);
5619 assert(table && "Cannot change domain");
5620 MI->setDesc(get(table[Domain-1]));
5623 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5624 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5625 NopInst.setOpcode(X86::NOOP);
5628 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
5629 // In particular, getJumpInstrTableEntryBound must always return an upper bound
5630 // on the encoding lengths of the instructions generated by
5631 // getUnconditionalBranch and getTrap.
5632 void X86InstrInfo::getUnconditionalBranch(
5633 MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const {
5634 Branch.setOpcode(X86::JMP_1);
5635 Branch.addOperand(MCOperand::CreateExpr(BranchTarget));
5638 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
5639 // In particular, getJumpInstrTableEntryBound must always return an upper bound
5640 // on the encoding lengths of the instructions generated by
5641 // getUnconditionalBranch and getTrap.
5642 void X86InstrInfo::getTrap(MCInst &MI) const {
5643 MI.setOpcode(X86::TRAP);
5646 // See getTrap and getUnconditionalBranch for conditions on the value returned
5647 // by this function.
5648 unsigned X86InstrInfo::getJumpInstrTableEntryBound() const {
5649 // 5 bytes suffice: JMP_4 Symbol@PLT is uses 1 byte (E9) for the JMP_4 and 4
5650 // bytes for the symbol offset. And TRAP is ud2, which is two bytes (0F 0B).
5654 bool X86InstrInfo::isHighLatencyDef(int opc) const {
5656 default: return false;
5658 case X86::DIVSDrm_Int:
5660 case X86::DIVSDrr_Int:
5662 case X86::DIVSSrm_Int:
5664 case X86::DIVSSrr_Int:
5670 case X86::SQRTSDm_Int:
5672 case X86::SQRTSDr_Int:
5674 case X86::SQRTSSm_Int:
5676 case X86::SQRTSSr_Int:
5677 // AVX instructions with high latency
5679 case X86::VDIVSDrm_Int:
5681 case X86::VDIVSDrr_Int:
5683 case X86::VDIVSSrm_Int:
5685 case X86::VDIVSSrr_Int:
5691 case X86::VSQRTSDm_Int:
5694 case X86::VSQRTSSm_Int:
5696 case X86::VSQRTPDZm:
5697 case X86::VSQRTPDZr:
5698 case X86::VSQRTPSZm:
5699 case X86::VSQRTPSZr:
5700 case X86::VSQRTSDZm:
5701 case X86::VSQRTSDZm_Int:
5702 case X86::VSQRTSDZr:
5703 case X86::VSQRTSSZm_Int:
5704 case X86::VSQRTSSZr:
5705 case X86::VSQRTSSZm:
5706 case X86::VDIVSDZrm:
5707 case X86::VDIVSDZrr:
5708 case X86::VDIVSSZrm:
5709 case X86::VDIVSSZrr:
5711 case X86::VGATHERQPSZrm:
5712 case X86::VGATHERQPDZrm:
5713 case X86::VGATHERDPDZrm:
5714 case X86::VGATHERDPSZrm:
5715 case X86::VPGATHERQDZrm:
5716 case X86::VPGATHERQQZrm:
5717 case X86::VPGATHERDDZrm:
5718 case X86::VPGATHERDQZrm:
5719 case X86::VSCATTERQPDZmr:
5720 case X86::VSCATTERQPSZmr:
5721 case X86::VSCATTERDPDZmr:
5722 case X86::VSCATTERDPSZmr:
5723 case X86::VPSCATTERQDZmr:
5724 case X86::VPSCATTERQQZmr:
5725 case X86::VPSCATTERDDZmr:
5726 case X86::VPSCATTERDQZmr:
5732 hasHighOperandLatency(const InstrItineraryData *ItinData,
5733 const MachineRegisterInfo *MRI,
5734 const MachineInstr *DefMI, unsigned DefIdx,
5735 const MachineInstr *UseMI, unsigned UseIdx) const {
5736 return isHighLatencyDef(DefMI->getOpcode());
5740 /// CGBR - Create Global Base Reg pass. This initializes the PIC
5741 /// global base register for x86-32.
5742 struct CGBR : public MachineFunctionPass {
5744 CGBR() : MachineFunctionPass(ID) {}
5746 bool runOnMachineFunction(MachineFunction &MF) override {
5747 const X86TargetMachine *TM =
5748 static_cast<const X86TargetMachine *>(&MF.getTarget());
5750 // Don't do anything if this is 64-bit as 64-bit PIC
5751 // uses RIP relative addressing.
5752 if (TM->getSubtarget<X86Subtarget>().is64Bit())
5755 // Only emit a global base reg in PIC mode.
5756 if (TM->getRelocationModel() != Reloc::PIC_)
5759 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5760 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5762 // If we didn't need a GlobalBaseReg, don't insert code.
5763 if (GlobalBaseReg == 0)
5766 // Insert the set of GlobalBaseReg into the first MBB of the function
5767 MachineBasicBlock &FirstMBB = MF.front();
5768 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5769 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5770 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5771 const X86InstrInfo *TII = TM->getSubtargetImpl()->getInstrInfo();
5774 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
5775 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
5779 // Operand of MovePCtoStack is completely ignored by asm printer. It's
5780 // only used in JIT code emission as displacement to pc.
5781 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
5783 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5784 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5785 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
5786 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5787 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5788 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5789 X86II::MO_GOT_ABSOLUTE_ADDRESS);
5795 const char *getPassName() const override {
5796 return "X86 PIC Global Base Reg Initialization";
5799 void getAnalysisUsage(AnalysisUsage &AU) const override {
5800 AU.setPreservesCFG();
5801 MachineFunctionPass::getAnalysisUsage(AU);
5808 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
5811 struct LDTLSCleanup : public MachineFunctionPass {
5813 LDTLSCleanup() : MachineFunctionPass(ID) {}
5815 bool runOnMachineFunction(MachineFunction &MF) override {
5816 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5817 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5818 // No point folding accesses if there isn't at least two.
5822 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5823 return VisitNode(DT->getRootNode(), 0);
5826 // Visit the dominator subtree rooted at Node in pre-order.
5827 // If TLSBaseAddrReg is non-null, then use that to replace any
5828 // TLS_base_addr instructions. Otherwise, create the register
5829 // when the first such instruction is seen, and then use it
5830 // as we encounter more instructions.
5831 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5832 MachineBasicBlock *BB = Node->getBlock();
5833 bool Changed = false;
5835 // Traverse the current block.
5836 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5838 switch (I->getOpcode()) {
5839 case X86::TLS_base_addr32:
5840 case X86::TLS_base_addr64:
5842 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5844 I = SetRegister(I, &TLSBaseAddrReg);
5852 // Visit the children of this block in the dominator tree.
5853 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5855 Changed |= VisitNode(*I, TLSBaseAddrReg);
5861 // Replace the TLS_base_addr instruction I with a copy from
5862 // TLSBaseAddrReg, returning the new instruction.
5863 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5864 unsigned TLSBaseAddrReg) {
5865 MachineFunction *MF = I->getParent()->getParent();
5866 const X86TargetMachine *TM =
5867 static_cast<const X86TargetMachine *>(&MF->getTarget());
5868 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5869 const X86InstrInfo *TII = TM->getSubtargetImpl()->getInstrInfo();
5871 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5872 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5873 TII->get(TargetOpcode::COPY),
5874 is64Bit ? X86::RAX : X86::EAX)
5875 .addReg(TLSBaseAddrReg);
5877 // Erase the TLS_base_addr instruction.
5878 I->eraseFromParent();
5883 // Create a virtal register in *TLSBaseAddrReg, and populate it by
5884 // inserting a copy instruction after I. Returns the new instruction.
5885 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5886 MachineFunction *MF = I->getParent()->getParent();
5887 const X86TargetMachine *TM =
5888 static_cast<const X86TargetMachine *>(&MF->getTarget());
5889 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5890 const X86InstrInfo *TII = TM->getSubtargetImpl()->getInstrInfo();
5892 // Create a virtual register for the TLS base address.
5893 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5894 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5895 ? &X86::GR64RegClass
5896 : &X86::GR32RegClass);
5898 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5899 MachineInstr *Next = I->getNextNode();
5900 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5901 TII->get(TargetOpcode::COPY),
5903 .addReg(is64Bit ? X86::RAX : X86::EAX);
5908 const char *getPassName() const override {
5909 return "Local Dynamic TLS Access Clean-up";
5912 void getAnalysisUsage(AnalysisUsage &AU) const override {
5913 AU.setPreservesCFG();
5914 AU.addRequired<MachineDominatorTree>();
5915 MachineFunctionPass::getAnalysisUsage(AU);
5920 char LDTLSCleanup::ID = 0;
5922 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }