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 X86MemoryFoldTableEntry {
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.getTargetTriple()) {
109 static const X86MemoryFoldTableEntry MemoryFoldTable2Addr[] = {
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 (X86MemoryFoldTableEntry Entry : MemoryFoldTable2Addr) {
273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
274 Entry.RegOp, Entry.MemOp,
275 // Index 0, folded load and store, no alignment requirement.
276 Entry.Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
279 static const X86MemoryFoldTableEntry MemoryFoldTable0[] = {
280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
301 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
302 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
303 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
304 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
305 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
306 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
307 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
308 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
309 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
310 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
311 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
312 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
313 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
314 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
315 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
316 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
317 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
318 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
319 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
320 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
321 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
322 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
324 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
325 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
326 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
327 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
328 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
329 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
330 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
331 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
332 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
333 { X86::PEXTRDrr, X86::PEXTRDmr, TB_FOLDED_STORE },
334 { X86::PEXTRQrr, X86::PEXTRQmr, TB_FOLDED_STORE },
335 { X86::PUSH16r, X86::PUSH16rmm, TB_FOLDED_LOAD },
336 { X86::PUSH32r, X86::PUSH32rmm, TB_FOLDED_LOAD },
337 { X86::PUSH64r, X86::PUSH64rmm, TB_FOLDED_LOAD },
338 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
339 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
340 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
341 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
342 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
343 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
344 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
345 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
346 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
347 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
348 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
349 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
350 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
351 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
352 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
353 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
354 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
355 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
356 { X86::TAILJMPr64_REX, X86::TAILJMPm64_REX, TB_FOLDED_LOAD },
357 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
358 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
359 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
360 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
362 // AVX 128-bit versions of foldable instructions
363 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
364 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
366 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
367 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
368 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
369 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
370 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
371 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
372 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
373 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
374 { X86::VPEXTRDrr, X86::VPEXTRDmr, TB_FOLDED_STORE },
375 { X86::VPEXTRQrr, X86::VPEXTRQmr, TB_FOLDED_STORE },
377 // AVX 256-bit foldable instructions
378 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
379 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
380 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
381 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
382 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
383 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE },
385 // AVX-512 foldable instructions
386 { X86::VMOVPDI2DIZrr, X86::VMOVPDI2DIZmr, TB_FOLDED_STORE },
387 { X86::VMOVAPDZrr, X86::VMOVAPDZmr, TB_FOLDED_STORE | TB_ALIGN_64 },
388 { X86::VMOVAPSZrr, X86::VMOVAPSZmr, TB_FOLDED_STORE | TB_ALIGN_64 },
389 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zmr, TB_FOLDED_STORE | TB_ALIGN_64 },
390 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zmr, TB_FOLDED_STORE | TB_ALIGN_64 },
391 { X86::VMOVUPDZrr, X86::VMOVUPDZmr, TB_FOLDED_STORE },
392 { X86::VMOVUPSZrr, X86::VMOVUPSZmr, TB_FOLDED_STORE },
393 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zmr, TB_FOLDED_STORE },
394 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zmr, TB_FOLDED_STORE },
395 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zmr, TB_FOLDED_STORE },
396 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zmr, TB_FOLDED_STORE },
398 // AVX-512 foldable instructions (256-bit versions)
399 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
400 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
401 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
402 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
403 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256mr, TB_FOLDED_STORE },
404 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256mr, TB_FOLDED_STORE },
405 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256mr, TB_FOLDED_STORE },
406 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256mr, TB_FOLDED_STORE },
407 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256mr, TB_FOLDED_STORE },
408 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256mr, TB_FOLDED_STORE },
410 // AVX-512 foldable instructions (128-bit versions)
411 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
412 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
413 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
414 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
415 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128mr, TB_FOLDED_STORE },
416 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128mr, TB_FOLDED_STORE },
417 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128mr, TB_FOLDED_STORE },
418 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128mr, TB_FOLDED_STORE },
419 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128mr, TB_FOLDED_STORE },
420 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128mr, TB_FOLDED_STORE },
422 // F16C foldable instructions
423 { X86::VCVTPS2PHrr, X86::VCVTPS2PHmr, TB_FOLDED_STORE },
424 { X86::VCVTPS2PHYrr, X86::VCVTPS2PHYmr, TB_FOLDED_STORE }
427 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable0) {
428 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
429 Entry.RegOp, Entry.MemOp, TB_INDEX_0 | Entry.Flags);
432 static const X86MemoryFoldTableEntry MemoryFoldTable1[] = {
433 { X86::BSF16rr, X86::BSF16rm, 0 },
434 { X86::BSF32rr, X86::BSF32rm, 0 },
435 { X86::BSF64rr, X86::BSF64rm, 0 },
436 { X86::BSR16rr, X86::BSR16rm, 0 },
437 { X86::BSR32rr, X86::BSR32rm, 0 },
438 { X86::BSR64rr, X86::BSR64rm, 0 },
439 { X86::CMP16rr, X86::CMP16rm, 0 },
440 { X86::CMP32rr, X86::CMP32rm, 0 },
441 { X86::CMP64rr, X86::CMP64rm, 0 },
442 { X86::CMP8rr, X86::CMP8rm, 0 },
443 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
444 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
445 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
446 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
447 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
448 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
449 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
450 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
451 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
452 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
453 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
454 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
455 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
456 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
457 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
458 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
459 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
460 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
461 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
462 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
463 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
464 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
465 { X86::CVTDQ2PDrr, X86::CVTDQ2PDrm, TB_ALIGN_16 },
466 { X86::CVTDQ2PSrr, X86::CVTDQ2PSrm, TB_ALIGN_16 },
467 { X86::CVTPD2DQrr, X86::CVTPD2DQrm, TB_ALIGN_16 },
468 { X86::CVTPD2PSrr, X86::CVTPD2PSrm, TB_ALIGN_16 },
469 { X86::CVTPS2DQrr, X86::CVTPS2DQrm, TB_ALIGN_16 },
470 { X86::CVTPS2PDrr, X86::CVTPS2PDrm, TB_ALIGN_16 },
471 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
472 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
473 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
474 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
475 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
476 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
477 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
478 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
479 { X86::MOV16rr, X86::MOV16rm, 0 },
480 { X86::MOV32rr, X86::MOV32rm, 0 },
481 { X86::MOV64rr, X86::MOV64rm, 0 },
482 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
483 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
484 { X86::MOV8rr, X86::MOV8rm, 0 },
485 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
486 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
487 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
488 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
489 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
490 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
491 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
492 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
493 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
494 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
495 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
496 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
497 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
498 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
499 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
500 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
501 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
502 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
503 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
504 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
505 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
506 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
507 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
508 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
509 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
510 { X86::PCMPESTRIrr, X86::PCMPESTRIrm, TB_ALIGN_16 },
511 { X86::PCMPESTRM128rr, X86::PCMPESTRM128rm, TB_ALIGN_16 },
512 { X86::PCMPISTRIrr, X86::PCMPISTRIrm, TB_ALIGN_16 },
513 { X86::PCMPISTRM128rr, X86::PCMPISTRM128rm, TB_ALIGN_16 },
514 { X86::PHMINPOSUWrr128, X86::PHMINPOSUWrm128, TB_ALIGN_16 },
515 { X86::PMOVSXBDrr, X86::PMOVSXBDrm, TB_ALIGN_16 },
516 { X86::PMOVSXBQrr, X86::PMOVSXBQrm, TB_ALIGN_16 },
517 { X86::PMOVSXBWrr, X86::PMOVSXBWrm, TB_ALIGN_16 },
518 { X86::PMOVSXDQrr, X86::PMOVSXDQrm, TB_ALIGN_16 },
519 { X86::PMOVSXWDrr, X86::PMOVSXWDrm, TB_ALIGN_16 },
520 { X86::PMOVSXWQrr, X86::PMOVSXWQrm, TB_ALIGN_16 },
521 { X86::PMOVZXBDrr, X86::PMOVZXBDrm, TB_ALIGN_16 },
522 { X86::PMOVZXBQrr, X86::PMOVZXBQrm, TB_ALIGN_16 },
523 { X86::PMOVZXBWrr, X86::PMOVZXBWrm, TB_ALIGN_16 },
524 { X86::PMOVZXDQrr, X86::PMOVZXDQrm, TB_ALIGN_16 },
525 { X86::PMOVZXWDrr, X86::PMOVZXWDrm, TB_ALIGN_16 },
526 { X86::PMOVZXWQrr, X86::PMOVZXWQrm, TB_ALIGN_16 },
527 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
528 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
529 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
530 { X86::PTESTrr, X86::PTESTrm, TB_ALIGN_16 },
531 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
532 { X86::RCPSSr, X86::RCPSSm, 0 },
533 { X86::RCPSSr_Int, X86::RCPSSm_Int, 0 },
534 { X86::ROUNDPDr, X86::ROUNDPDm, TB_ALIGN_16 },
535 { X86::ROUNDPSr, X86::ROUNDPSm, TB_ALIGN_16 },
536 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
537 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
538 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
539 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
540 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
541 { X86::SQRTSDr, X86::SQRTSDm, 0 },
542 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
543 { X86::SQRTSSr, X86::SQRTSSm, 0 },
544 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
545 { X86::TEST16rr, X86::TEST16rm, 0 },
546 { X86::TEST32rr, X86::TEST32rm, 0 },
547 { X86::TEST64rr, X86::TEST64rm, 0 },
548 { X86::TEST8rr, X86::TEST8rm, 0 },
549 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
550 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
551 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
553 // MMX version of foldable instructions
554 { X86::MMX_CVTPD2PIirr, X86::MMX_CVTPD2PIirm, 0 },
555 { X86::MMX_CVTPI2PDirr, X86::MMX_CVTPI2PDirm, 0 },
556 { X86::MMX_CVTPS2PIirr, X86::MMX_CVTPS2PIirm, 0 },
557 { X86::MMX_CVTTPD2PIirr, X86::MMX_CVTTPD2PIirm, 0 },
558 { X86::MMX_CVTTPS2PIirr, X86::MMX_CVTTPS2PIirm, 0 },
559 { X86::MMX_MOVD64to64rr, X86::MMX_MOVQ64rm, 0 },
560 { X86::MMX_PABSBrr64, X86::MMX_PABSBrm64, 0 },
561 { X86::MMX_PABSDrr64, X86::MMX_PABSDrm64, 0 },
562 { X86::MMX_PABSWrr64, X86::MMX_PABSWrm64, 0 },
563 { X86::MMX_PSHUFWri, X86::MMX_PSHUFWmi, 0 },
565 // 3DNow! version of foldable instructions
566 { X86::PF2IDrr, X86::PF2IDrm, 0 },
567 { X86::PF2IWrr, X86::PF2IWrm, 0 },
568 { X86::PFRCPrr, X86::PFRCPrm, 0 },
569 { X86::PFRSQRTrr, X86::PFRSQRTrm, 0 },
570 { X86::PI2FDrr, X86::PI2FDrm, 0 },
571 { X86::PI2FWrr, X86::PI2FWrm, 0 },
572 { X86::PSWAPDrr, X86::PSWAPDrm, 0 },
574 // AVX 128-bit versions of foldable instructions
575 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
576 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
577 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
578 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
579 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
580 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
581 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
582 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
583 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
584 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
585 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
586 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
587 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
588 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
589 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
590 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
591 { X86::VCVTDQ2PDrr, X86::VCVTDQ2PDrm, 0 },
592 { X86::VCVTDQ2PSrr, X86::VCVTDQ2PSrm, 0 },
593 { X86::VCVTPD2DQrr, X86::VCVTPD2DQXrm, 0 },
594 { X86::VCVTPD2PSrr, X86::VCVTPD2PSXrm, 0 },
595 { X86::VCVTPS2DQrr, X86::VCVTPS2DQrm, 0 },
596 { X86::VCVTPS2PDrr, X86::VCVTPS2PDrm, 0 },
597 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
598 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
599 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
600 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
601 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
602 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
603 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
604 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
605 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
606 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
607 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, 0 },
608 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, 0 },
609 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
610 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
611 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
612 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
613 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
614 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
615 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
616 { X86::VPCMPESTRIrr, X86::VPCMPESTRIrm, 0 },
617 { X86::VPCMPESTRM128rr, X86::VPCMPESTRM128rm, 0 },
618 { X86::VPCMPISTRIrr, X86::VPCMPISTRIrm, 0 },
619 { X86::VPCMPISTRM128rr, X86::VPCMPISTRM128rm, 0 },
620 { X86::VPHMINPOSUWrr128, X86::VPHMINPOSUWrm128, 0 },
621 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
622 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
623 { X86::VPMOVSXBDrr, X86::VPMOVSXBDrm, 0 },
624 { X86::VPMOVSXBQrr, X86::VPMOVSXBQrm, 0 },
625 { X86::VPMOVSXBWrr, X86::VPMOVSXBWrm, 0 },
626 { X86::VPMOVSXDQrr, X86::VPMOVSXDQrm, 0 },
627 { X86::VPMOVSXWDrr, X86::VPMOVSXWDrm, 0 },
628 { X86::VPMOVSXWQrr, X86::VPMOVSXWQrm, 0 },
629 { X86::VPMOVZXBDrr, X86::VPMOVZXBDrm, 0 },
630 { X86::VPMOVZXBQrr, X86::VPMOVZXBQrm, 0 },
631 { X86::VPMOVZXBWrr, X86::VPMOVZXBWrm, 0 },
632 { X86::VPMOVZXDQrr, X86::VPMOVZXDQrm, 0 },
633 { X86::VPMOVZXWDrr, X86::VPMOVZXWDrm, 0 },
634 { X86::VPMOVZXWQrr, X86::VPMOVZXWQrm, 0 },
635 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
636 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
637 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
638 { X86::VPTESTrr, X86::VPTESTrm, 0 },
639 { X86::VRCPPSr, X86::VRCPPSm, 0 },
640 { X86::VROUNDPDr, X86::VROUNDPDm, 0 },
641 { X86::VROUNDPSr, X86::VROUNDPSm, 0 },
642 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
643 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
644 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
645 { X86::VTESTPDrr, X86::VTESTPDrm, 0 },
646 { X86::VTESTPSrr, X86::VTESTPSrm, 0 },
647 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
648 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
650 // AVX 256-bit foldable instructions
651 { X86::VCVTDQ2PDYrr, X86::VCVTDQ2PDYrm, 0 },
652 { X86::VCVTDQ2PSYrr, X86::VCVTDQ2PSYrm, 0 },
653 { X86::VCVTPD2DQYrr, X86::VCVTPD2DQYrm, 0 },
654 { X86::VCVTPD2PSYrr, X86::VCVTPD2PSYrm, 0 },
655 { X86::VCVTPS2DQYrr, X86::VCVTPS2DQYrm, 0 },
656 { X86::VCVTPS2PDYrr, X86::VCVTPS2PDYrm, 0 },
657 { X86::VCVTTPD2DQYrr, X86::VCVTTPD2DQYrm, 0 },
658 { X86::VCVTTPS2DQYrr, X86::VCVTTPS2DQYrm, 0 },
659 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
660 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
661 { X86::VMOVDDUPYrr, X86::VMOVDDUPYrm, 0 },
662 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
663 { X86::VMOVSLDUPYrr, X86::VMOVSLDUPYrm, 0 },
664 { X86::VMOVSHDUPYrr, X86::VMOVSHDUPYrm, 0 },
665 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
666 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
667 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
668 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
669 { X86::VPTESTYrr, X86::VPTESTYrm, 0 },
670 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
671 { X86::VROUNDYPDr, X86::VROUNDYPDm, 0 },
672 { X86::VROUNDYPSr, X86::VROUNDYPSm, 0 },
673 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
674 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
675 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
676 { X86::VTESTPDYrr, X86::VTESTPDYrm, 0 },
677 { X86::VTESTPSYrr, X86::VTESTPSYrm, 0 },
679 // AVX2 foldable instructions
681 // VBROADCASTS{SD}rr register instructions were an AVX2 addition while the
682 // VBROADCASTS{SD}rm memory instructions were available from AVX1.
683 // TB_NO_REVERSE prevents unfolding from introducing an illegal instruction
684 // on AVX1 targets. The VPBROADCAST instructions are all AVX2 instructions
685 // so they don't need an equivalent limitation.
686 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
687 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
688 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
689 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
690 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
691 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
692 { X86::VPBROADCASTBrr, X86::VPBROADCASTBrm, 0 },
693 { X86::VPBROADCASTBYrr, X86::VPBROADCASTBYrm, 0 },
694 { X86::VPBROADCASTDrr, X86::VPBROADCASTDrm, 0 },
695 { X86::VPBROADCASTDYrr, X86::VPBROADCASTDYrm, 0 },
696 { X86::VPBROADCASTQrr, X86::VPBROADCASTQrm, 0 },
697 { X86::VPBROADCASTQYrr, X86::VPBROADCASTQYrm, 0 },
698 { X86::VPBROADCASTWrr, X86::VPBROADCASTWrm, 0 },
699 { X86::VPBROADCASTWYrr, X86::VPBROADCASTWYrm, 0 },
700 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
701 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
702 { X86::VPMOVSXBDYrr, X86::VPMOVSXBDYrm, 0 },
703 { X86::VPMOVSXBQYrr, X86::VPMOVSXBQYrm, 0 },
704 { X86::VPMOVSXBWYrr, X86::VPMOVSXBWYrm, 0 },
705 { X86::VPMOVSXDQYrr, X86::VPMOVSXDQYrm, 0 },
706 { X86::VPMOVSXWDYrr, X86::VPMOVSXWDYrm, 0 },
707 { X86::VPMOVSXWQYrr, X86::VPMOVSXWQYrm, 0 },
708 { X86::VPMOVZXBDYrr, X86::VPMOVZXBDYrm, 0 },
709 { X86::VPMOVZXBQYrr, X86::VPMOVZXBQYrm, 0 },
710 { X86::VPMOVZXBWYrr, X86::VPMOVZXBWYrm, 0 },
711 { X86::VPMOVZXDQYrr, X86::VPMOVZXDQYrm, 0 },
712 { X86::VPMOVZXWDYrr, X86::VPMOVZXWDYrm, 0 },
713 { X86::VPMOVZXWQYrr, X86::VPMOVZXWQYrm, 0 },
714 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
715 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
716 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
718 // XOP foldable instructions
719 { X86::VFRCZPDrr, X86::VFRCZPDrm, 0 },
720 { X86::VFRCZPDrrY, X86::VFRCZPDrmY, 0 },
721 { X86::VFRCZPSrr, X86::VFRCZPSrm, 0 },
722 { X86::VFRCZPSrrY, X86::VFRCZPSrmY, 0 },
723 { X86::VFRCZSDrr, X86::VFRCZSDrm, 0 },
724 { X86::VFRCZSSrr, X86::VFRCZSSrm, 0 },
725 { X86::VPHADDBDrr, X86::VPHADDBDrm, 0 },
726 { X86::VPHADDBQrr, X86::VPHADDBQrm, 0 },
727 { X86::VPHADDBWrr, X86::VPHADDBWrm, 0 },
728 { X86::VPHADDDQrr, X86::VPHADDDQrm, 0 },
729 { X86::VPHADDWDrr, X86::VPHADDWDrm, 0 },
730 { X86::VPHADDWQrr, X86::VPHADDWQrm, 0 },
731 { X86::VPHADDUBDrr, X86::VPHADDUBDrm, 0 },
732 { X86::VPHADDUBQrr, X86::VPHADDUBQrm, 0 },
733 { X86::VPHADDUBWrr, X86::VPHADDUBWrm, 0 },
734 { X86::VPHADDUDQrr, X86::VPHADDUDQrm, 0 },
735 { X86::VPHADDUWDrr, X86::VPHADDUWDrm, 0 },
736 { X86::VPHADDUWQrr, X86::VPHADDUWQrm, 0 },
737 { X86::VPHSUBBWrr, X86::VPHSUBBWrm, 0 },
738 { X86::VPHSUBDQrr, X86::VPHSUBDQrm, 0 },
739 { X86::VPHSUBWDrr, X86::VPHSUBWDrm, 0 },
740 { X86::VPROTBri, X86::VPROTBmi, 0 },
741 { X86::VPROTBrr, X86::VPROTBmr, 0 },
742 { X86::VPROTDri, X86::VPROTDmi, 0 },
743 { X86::VPROTDrr, X86::VPROTDmr, 0 },
744 { X86::VPROTQri, X86::VPROTQmi, 0 },
745 { X86::VPROTQrr, X86::VPROTQmr, 0 },
746 { X86::VPROTWri, X86::VPROTWmi, 0 },
747 { X86::VPROTWrr, X86::VPROTWmr, 0 },
748 { X86::VPSHABrr, X86::VPSHABmr, 0 },
749 { X86::VPSHADrr, X86::VPSHADmr, 0 },
750 { X86::VPSHAQrr, X86::VPSHAQmr, 0 },
751 { X86::VPSHAWrr, X86::VPSHAWmr, 0 },
752 { X86::VPSHLBrr, X86::VPSHLBmr, 0 },
753 { X86::VPSHLDrr, X86::VPSHLDmr, 0 },
754 { X86::VPSHLQrr, X86::VPSHLQmr, 0 },
755 { X86::VPSHLWrr, X86::VPSHLWmr, 0 },
757 // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
758 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
759 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
760 { X86::BEXTRI32ri, X86::BEXTRI32mi, 0 },
761 { X86::BEXTRI64ri, X86::BEXTRI64mi, 0 },
762 { X86::BLCFILL32rr, X86::BLCFILL32rm, 0 },
763 { X86::BLCFILL64rr, X86::BLCFILL64rm, 0 },
764 { X86::BLCI32rr, X86::BLCI32rm, 0 },
765 { X86::BLCI64rr, X86::BLCI64rm, 0 },
766 { X86::BLCIC32rr, X86::BLCIC32rm, 0 },
767 { X86::BLCIC64rr, X86::BLCIC64rm, 0 },
768 { X86::BLCMSK32rr, X86::BLCMSK32rm, 0 },
769 { X86::BLCMSK64rr, X86::BLCMSK64rm, 0 },
770 { X86::BLCS32rr, X86::BLCS32rm, 0 },
771 { X86::BLCS64rr, X86::BLCS64rm, 0 },
772 { X86::BLSFILL32rr, X86::BLSFILL32rm, 0 },
773 { X86::BLSFILL64rr, X86::BLSFILL64rm, 0 },
774 { X86::BLSI32rr, X86::BLSI32rm, 0 },
775 { X86::BLSI64rr, X86::BLSI64rm, 0 },
776 { X86::BLSIC32rr, X86::BLSIC32rm, 0 },
777 { X86::BLSIC64rr, X86::BLSIC64rm, 0 },
778 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
779 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
780 { X86::BLSR32rr, X86::BLSR32rm, 0 },
781 { X86::BLSR64rr, X86::BLSR64rm, 0 },
782 { X86::BZHI32rr, X86::BZHI32rm, 0 },
783 { X86::BZHI64rr, X86::BZHI64rm, 0 },
784 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
785 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
786 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
787 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
788 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
789 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
790 { X86::RORX32ri, X86::RORX32mi, 0 },
791 { X86::RORX64ri, X86::RORX64mi, 0 },
792 { X86::SARX32rr, X86::SARX32rm, 0 },
793 { X86::SARX64rr, X86::SARX64rm, 0 },
794 { X86::SHRX32rr, X86::SHRX32rm, 0 },
795 { X86::SHRX64rr, X86::SHRX64rm, 0 },
796 { X86::SHLX32rr, X86::SHLX32rm, 0 },
797 { X86::SHLX64rr, X86::SHLX64rm, 0 },
798 { X86::T1MSKC32rr, X86::T1MSKC32rm, 0 },
799 { X86::T1MSKC64rr, X86::T1MSKC64rm, 0 },
800 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
801 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
802 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
803 { X86::TZMSK32rr, X86::TZMSK32rm, 0 },
804 { X86::TZMSK64rr, X86::TZMSK64rm, 0 },
806 // AVX-512 foldable instructions
807 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 },
808 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 },
809 { X86::VMOVAPDZrr, X86::VMOVAPDZrm, TB_ALIGN_64 },
810 { X86::VMOVAPSZrr, X86::VMOVAPSZrm, TB_ALIGN_64 },
811 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zrm, TB_ALIGN_64 },
812 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zrm, TB_ALIGN_64 },
813 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zrm, 0 },
814 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zrm, 0 },
815 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zrm, 0 },
816 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zrm, 0 },
817 { X86::VMOVUPDZrr, X86::VMOVUPDZrm, 0 },
818 { X86::VMOVUPSZrr, X86::VMOVUPSZrm, 0 },
819 { X86::VPABSDZrr, X86::VPABSDZrm, 0 },
820 { X86::VPABSQZrr, X86::VPABSQZrm, 0 },
821 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZm, TB_NO_REVERSE },
822 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZm, TB_NO_REVERSE },
824 // AVX-512 foldable instructions (256-bit versions)
825 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256rm, TB_ALIGN_32 },
826 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256rm, TB_ALIGN_32 },
827 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256rm, TB_ALIGN_32 },
828 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256rm, TB_ALIGN_32 },
829 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256rm, 0 },
830 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256rm, 0 },
831 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256rm, 0 },
832 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256rm, 0 },
833 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256rm, 0 },
834 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256rm, 0 },
835 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256m, TB_NO_REVERSE },
836 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256m, TB_NO_REVERSE },
838 // AVX-512 foldable instructions (256-bit versions)
839 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128rm, TB_ALIGN_16 },
840 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128rm, TB_ALIGN_16 },
841 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128rm, TB_ALIGN_16 },
842 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128rm, TB_ALIGN_16 },
843 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128rm, 0 },
844 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128rm, 0 },
845 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128rm, 0 },
846 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128rm, 0 },
847 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128rm, 0 },
848 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128rm, 0 },
849 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128m, TB_NO_REVERSE },
851 // F16C foldable instructions
852 { X86::VCVTPH2PSrr, X86::VCVTPH2PSrm, 0 },
853 { X86::VCVTPH2PSYrr, X86::VCVTPH2PSYrm, 0 },
855 // AES foldable instructions
856 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 },
857 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 },
858 { X86::VAESIMCrr, X86::VAESIMCrm, 0 },
859 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, 0 }
862 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable1) {
863 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
864 Entry.RegOp, Entry.MemOp,
865 // Index 1, folded load
866 Entry.Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
869 static const X86MemoryFoldTableEntry MemoryFoldTable2[] = {
870 { X86::ADC32rr, X86::ADC32rm, 0 },
871 { X86::ADC64rr, X86::ADC64rm, 0 },
872 { X86::ADD16rr, X86::ADD16rm, 0 },
873 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
874 { X86::ADD32rr, X86::ADD32rm, 0 },
875 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
876 { X86::ADD64rr, X86::ADD64rm, 0 },
877 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
878 { X86::ADD8rr, X86::ADD8rm, 0 },
879 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
880 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
881 { X86::ADDSDrr, X86::ADDSDrm, 0 },
882 { X86::ADDSDrr_Int, X86::ADDSDrm_Int, 0 },
883 { X86::ADDSSrr, X86::ADDSSrm, 0 },
884 { X86::ADDSSrr_Int, X86::ADDSSrm_Int, 0 },
885 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
886 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
887 { X86::AND16rr, X86::AND16rm, 0 },
888 { X86::AND32rr, X86::AND32rm, 0 },
889 { X86::AND64rr, X86::AND64rm, 0 },
890 { X86::AND8rr, X86::AND8rm, 0 },
891 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
892 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
893 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
894 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
895 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
896 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
897 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
898 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
899 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
900 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
901 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
902 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
903 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
904 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
905 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
906 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
907 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
908 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
909 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
910 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
911 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
912 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
913 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
914 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
915 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
916 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
917 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
918 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
919 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
920 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
921 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
922 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
923 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
924 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
925 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
926 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
927 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
928 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
929 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
930 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
931 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
932 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
933 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
934 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
935 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
936 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
937 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
938 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
939 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
940 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
941 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
942 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
943 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
944 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
945 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
946 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
947 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
948 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
949 { X86::CMPSDrr, X86::CMPSDrm, 0 },
950 { X86::CMPSSrr, X86::CMPSSrm, 0 },
951 { X86::CRC32r32r32, X86::CRC32r32m32, 0 },
952 { X86::CRC32r64r64, X86::CRC32r64m64, 0 },
953 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
954 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
955 { X86::DIVSDrr, X86::DIVSDrm, 0 },
956 { X86::DIVSDrr_Int, X86::DIVSDrm_Int, 0 },
957 { X86::DIVSSrr, X86::DIVSSrm, 0 },
958 { X86::DIVSSrr_Int, X86::DIVSSrm_Int, 0 },
959 { X86::DPPDrri, X86::DPPDrmi, TB_ALIGN_16 },
960 { X86::DPPSrri, X86::DPPSrmi, TB_ALIGN_16 },
962 // Do not fold Fs* scalar logical op loads because there are no scalar
963 // load variants for these instructions. When folded, the load is required
964 // to be 128-bits, so the load size would not match.
966 { X86::FvANDNPDrr, X86::FvANDNPDrm, TB_ALIGN_16 },
967 { X86::FvANDNPSrr, X86::FvANDNPSrm, TB_ALIGN_16 },
968 { X86::FvANDPDrr, X86::FvANDPDrm, TB_ALIGN_16 },
969 { X86::FvANDPSrr, X86::FvANDPSrm, TB_ALIGN_16 },
970 { X86::FvORPDrr, X86::FvORPDrm, TB_ALIGN_16 },
971 { X86::FvORPSrr, X86::FvORPSrm, TB_ALIGN_16 },
972 { X86::FvXORPDrr, X86::FvXORPDrm, TB_ALIGN_16 },
973 { X86::FvXORPSrr, X86::FvXORPSrm, TB_ALIGN_16 },
974 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
975 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
976 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
977 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
978 { X86::IMUL16rr, X86::IMUL16rm, 0 },
979 { X86::IMUL32rr, X86::IMUL32rm, 0 },
980 { X86::IMUL64rr, X86::IMUL64rm, 0 },
981 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
982 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
983 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
984 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
985 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
986 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
987 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
988 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
989 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
990 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
991 { X86::MAXSDrr, X86::MAXSDrm, 0 },
992 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
993 { X86::MAXSSrr, X86::MAXSSrm, 0 },
994 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
995 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
996 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
997 { X86::MINSDrr, X86::MINSDrm, 0 },
998 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
999 { X86::MINSSrr, X86::MINSSrm, 0 },
1000 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
1001 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
1002 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
1003 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
1004 { X86::MULSDrr, X86::MULSDrm, 0 },
1005 { X86::MULSDrr_Int, X86::MULSDrm_Int, 0 },
1006 { X86::MULSSrr, X86::MULSSrm, 0 },
1007 { X86::MULSSrr_Int, X86::MULSSrm_Int, 0 },
1008 { X86::OR16rr, X86::OR16rm, 0 },
1009 { X86::OR32rr, X86::OR32rm, 0 },
1010 { X86::OR64rr, X86::OR64rm, 0 },
1011 { X86::OR8rr, X86::OR8rm, 0 },
1012 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
1013 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
1014 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
1015 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
1016 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
1017 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
1018 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
1019 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
1020 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
1021 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
1022 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
1023 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
1024 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
1025 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
1026 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
1027 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
1028 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
1029 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
1030 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
1031 { X86::PBLENDVBrr0, X86::PBLENDVBrm0, TB_ALIGN_16 },
1032 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
1033 { X86::PCLMULQDQrr, X86::PCLMULQDQrm, TB_ALIGN_16 },
1034 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
1035 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
1036 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
1037 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
1038 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
1039 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
1040 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
1041 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
1042 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
1043 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
1044 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
1045 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
1046 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
1047 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
1048 { X86::PINSRBrr, X86::PINSRBrm, 0 },
1049 { X86::PINSRDrr, X86::PINSRDrm, 0 },
1050 { X86::PINSRQrr, X86::PINSRQrm, 0 },
1051 { X86::PINSRWrri, X86::PINSRWrmi, 0 },
1052 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
1053 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
1054 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
1055 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
1056 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
1057 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
1058 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
1059 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
1060 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
1061 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
1062 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
1063 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
1064 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
1065 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
1066 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
1067 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
1068 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
1069 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
1070 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
1071 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
1072 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
1073 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
1074 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
1075 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
1076 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
1077 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
1078 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
1079 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
1080 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
1081 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
1082 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
1083 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
1084 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
1085 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
1086 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
1087 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
1088 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
1089 { X86::PSUBQrr, X86::PSUBQrm, TB_ALIGN_16 },
1090 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
1091 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
1092 { X86::PSUBUSBrr, X86::PSUBUSBrm, TB_ALIGN_16 },
1093 { X86::PSUBUSWrr, X86::PSUBUSWrm, TB_ALIGN_16 },
1094 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
1095 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
1096 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
1097 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
1098 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
1099 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
1100 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
1101 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
1102 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
1103 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
1104 { X86::ROUNDSDr, X86::ROUNDSDm, 0 },
1105 { X86::ROUNDSSr, X86::ROUNDSSm, 0 },
1106 { X86::SBB32rr, X86::SBB32rm, 0 },
1107 { X86::SBB64rr, X86::SBB64rm, 0 },
1108 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
1109 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
1110 { X86::SUB16rr, X86::SUB16rm, 0 },
1111 { X86::SUB32rr, X86::SUB32rm, 0 },
1112 { X86::SUB64rr, X86::SUB64rm, 0 },
1113 { X86::SUB8rr, X86::SUB8rm, 0 },
1114 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
1115 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
1116 { X86::SUBSDrr, X86::SUBSDrm, 0 },
1117 { X86::SUBSDrr_Int, X86::SUBSDrm_Int, 0 },
1118 { X86::SUBSSrr, X86::SUBSSrm, 0 },
1119 { X86::SUBSSrr_Int, X86::SUBSSrm_Int, 0 },
1120 // FIXME: TEST*rr -> swapped operand of TEST*mr.
1121 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
1122 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
1123 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
1124 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
1125 { X86::XOR16rr, X86::XOR16rm, 0 },
1126 { X86::XOR32rr, X86::XOR32rm, 0 },
1127 { X86::XOR64rr, X86::XOR64rm, 0 },
1128 { X86::XOR8rr, X86::XOR8rm, 0 },
1129 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
1130 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
1132 // MMX version of foldable instructions
1133 { X86::MMX_CVTPI2PSirr, X86::MMX_CVTPI2PSirm, 0 },
1134 { X86::MMX_PACKSSDWirr, X86::MMX_PACKSSDWirm, 0 },
1135 { X86::MMX_PACKSSWBirr, X86::MMX_PACKSSWBirm, 0 },
1136 { X86::MMX_PACKUSWBirr, X86::MMX_PACKUSWBirm, 0 },
1137 { X86::MMX_PADDBirr, X86::MMX_PADDBirm, 0 },
1138 { X86::MMX_PADDDirr, X86::MMX_PADDDirm, 0 },
1139 { X86::MMX_PADDQirr, X86::MMX_PADDQirm, 0 },
1140 { X86::MMX_PADDSBirr, X86::MMX_PADDSBirm, 0 },
1141 { X86::MMX_PADDSWirr, X86::MMX_PADDSWirm, 0 },
1142 { X86::MMX_PADDUSBirr, X86::MMX_PADDUSBirm, 0 },
1143 { X86::MMX_PADDUSWirr, X86::MMX_PADDUSWirm, 0 },
1144 { X86::MMX_PADDWirr, X86::MMX_PADDWirm, 0 },
1145 { X86::MMX_PALIGNR64irr, X86::MMX_PALIGNR64irm, 0 },
1146 { X86::MMX_PANDNirr, X86::MMX_PANDNirm, 0 },
1147 { X86::MMX_PANDirr, X86::MMX_PANDirm, 0 },
1148 { X86::MMX_PAVGBirr, X86::MMX_PAVGBirm, 0 },
1149 { X86::MMX_PAVGWirr, X86::MMX_PAVGWirm, 0 },
1150 { X86::MMX_PCMPEQBirr, X86::MMX_PCMPEQBirm, 0 },
1151 { X86::MMX_PCMPEQDirr, X86::MMX_PCMPEQDirm, 0 },
1152 { X86::MMX_PCMPEQWirr, X86::MMX_PCMPEQWirm, 0 },
1153 { X86::MMX_PCMPGTBirr, X86::MMX_PCMPGTBirm, 0 },
1154 { X86::MMX_PCMPGTDirr, X86::MMX_PCMPGTDirm, 0 },
1155 { X86::MMX_PCMPGTWirr, X86::MMX_PCMPGTWirm, 0 },
1156 { X86::MMX_PHADDSWrr64, X86::MMX_PHADDSWrm64, 0 },
1157 { X86::MMX_PHADDWrr64, X86::MMX_PHADDWrm64, 0 },
1158 { X86::MMX_PHADDrr64, X86::MMX_PHADDrm64, 0 },
1159 { X86::MMX_PHSUBDrr64, X86::MMX_PHSUBDrm64, 0 },
1160 { X86::MMX_PHSUBSWrr64, X86::MMX_PHSUBSWrm64, 0 },
1161 { X86::MMX_PHSUBWrr64, X86::MMX_PHSUBWrm64, 0 },
1162 { X86::MMX_PINSRWirri, X86::MMX_PINSRWirmi, 0 },
1163 { X86::MMX_PMADDUBSWrr64, X86::MMX_PMADDUBSWrm64, 0 },
1164 { X86::MMX_PMADDWDirr, X86::MMX_PMADDWDirm, 0 },
1165 { X86::MMX_PMAXSWirr, X86::MMX_PMAXSWirm, 0 },
1166 { X86::MMX_PMAXUBirr, X86::MMX_PMAXUBirm, 0 },
1167 { X86::MMX_PMINSWirr, X86::MMX_PMINSWirm, 0 },
1168 { X86::MMX_PMINUBirr, X86::MMX_PMINUBirm, 0 },
1169 { X86::MMX_PMULHRSWrr64, X86::MMX_PMULHRSWrm64, 0 },
1170 { X86::MMX_PMULHUWirr, X86::MMX_PMULHUWirm, 0 },
1171 { X86::MMX_PMULHWirr, X86::MMX_PMULHWirm, 0 },
1172 { X86::MMX_PMULLWirr, X86::MMX_PMULLWirm, 0 },
1173 { X86::MMX_PMULUDQirr, X86::MMX_PMULUDQirm, 0 },
1174 { X86::MMX_PORirr, X86::MMX_PORirm, 0 },
1175 { X86::MMX_PSADBWirr, X86::MMX_PSADBWirm, 0 },
1176 { X86::MMX_PSHUFBrr64, X86::MMX_PSHUFBrm64, 0 },
1177 { X86::MMX_PSIGNBrr64, X86::MMX_PSIGNBrm64, 0 },
1178 { X86::MMX_PSIGNDrr64, X86::MMX_PSIGNDrm64, 0 },
1179 { X86::MMX_PSIGNWrr64, X86::MMX_PSIGNWrm64, 0 },
1180 { X86::MMX_PSLLDrr, X86::MMX_PSLLDrm, 0 },
1181 { X86::MMX_PSLLQrr, X86::MMX_PSLLQrm, 0 },
1182 { X86::MMX_PSLLWrr, X86::MMX_PSLLWrm, 0 },
1183 { X86::MMX_PSRADrr, X86::MMX_PSRADrm, 0 },
1184 { X86::MMX_PSRAWrr, X86::MMX_PSRAWrm, 0 },
1185 { X86::MMX_PSRLDrr, X86::MMX_PSRLDrm, 0 },
1186 { X86::MMX_PSRLQrr, X86::MMX_PSRLQrm, 0 },
1187 { X86::MMX_PSRLWrr, X86::MMX_PSRLWrm, 0 },
1188 { X86::MMX_PSUBBirr, X86::MMX_PSUBBirm, 0 },
1189 { X86::MMX_PSUBDirr, X86::MMX_PSUBDirm, 0 },
1190 { X86::MMX_PSUBQirr, X86::MMX_PSUBQirm, 0 },
1191 { X86::MMX_PSUBSBirr, X86::MMX_PSUBSBirm, 0 },
1192 { X86::MMX_PSUBSWirr, X86::MMX_PSUBSWirm, 0 },
1193 { X86::MMX_PSUBUSBirr, X86::MMX_PSUBUSBirm, 0 },
1194 { X86::MMX_PSUBUSWirr, X86::MMX_PSUBUSWirm, 0 },
1195 { X86::MMX_PSUBWirr, X86::MMX_PSUBWirm, 0 },
1196 { X86::MMX_PUNPCKHBWirr, X86::MMX_PUNPCKHBWirm, 0 },
1197 { X86::MMX_PUNPCKHDQirr, X86::MMX_PUNPCKHDQirm, 0 },
1198 { X86::MMX_PUNPCKHWDirr, X86::MMX_PUNPCKHWDirm, 0 },
1199 { X86::MMX_PUNPCKLBWirr, X86::MMX_PUNPCKLBWirm, 0 },
1200 { X86::MMX_PUNPCKLDQirr, X86::MMX_PUNPCKLDQirm, 0 },
1201 { X86::MMX_PUNPCKLWDirr, X86::MMX_PUNPCKLWDirm, 0 },
1202 { X86::MMX_PXORirr, X86::MMX_PXORirm, 0 },
1204 // 3DNow! version of foldable instructions
1205 { X86::PAVGUSBrr, X86::PAVGUSBrm, 0 },
1206 { X86::PFACCrr, X86::PFACCrm, 0 },
1207 { X86::PFADDrr, X86::PFADDrm, 0 },
1208 { X86::PFCMPEQrr, X86::PFCMPEQrm, 0 },
1209 { X86::PFCMPGErr, X86::PFCMPGErm, 0 },
1210 { X86::PFCMPGTrr, X86::PFCMPGTrm, 0 },
1211 { X86::PFMAXrr, X86::PFMAXrm, 0 },
1212 { X86::PFMINrr, X86::PFMINrm, 0 },
1213 { X86::PFMULrr, X86::PFMULrm, 0 },
1214 { X86::PFNACCrr, X86::PFNACCrm, 0 },
1215 { X86::PFPNACCrr, X86::PFPNACCrm, 0 },
1216 { X86::PFRCPIT1rr, X86::PFRCPIT1rm, 0 },
1217 { X86::PFRCPIT2rr, X86::PFRCPIT2rm, 0 },
1218 { X86::PFRSQIT1rr, X86::PFRSQIT1rm, 0 },
1219 { X86::PFSUBrr, X86::PFSUBrm, 0 },
1220 { X86::PFSUBRrr, X86::PFSUBRrm, 0 },
1221 { X86::PMULHRWrr, X86::PMULHRWrm, 0 },
1223 // AVX 128-bit versions of foldable instructions
1224 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
1225 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
1226 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
1227 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
1228 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
1229 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
1230 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
1231 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
1232 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
1233 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
1234 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
1235 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
1236 { X86::VRCPSSr, X86::VRCPSSm, 0 },
1237 { X86::VRCPSSr_Int, X86::VRCPSSm_Int, 0 },
1238 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
1239 { X86::VRSQRTSSr_Int, X86::VRSQRTSSm_Int, 0 },
1240 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
1241 { X86::VSQRTSDr_Int, X86::VSQRTSDm_Int, 0 },
1242 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
1243 { X86::VSQRTSSr_Int, X86::VSQRTSSm_Int, 0 },
1244 { X86::VADDPDrr, X86::VADDPDrm, 0 },
1245 { X86::VADDPSrr, X86::VADDPSrm, 0 },
1246 { X86::VADDSDrr, X86::VADDSDrm, 0 },
1247 { X86::VADDSDrr_Int, X86::VADDSDrm_Int, 0 },
1248 { X86::VADDSSrr, X86::VADDSSrm, 0 },
1249 { X86::VADDSSrr_Int, X86::VADDSSrm_Int, 0 },
1250 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
1251 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
1252 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
1253 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
1254 { X86::VANDPDrr, X86::VANDPDrm, 0 },
1255 { X86::VANDPSrr, X86::VANDPSrm, 0 },
1256 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
1257 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
1258 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
1259 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
1260 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
1261 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
1262 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
1263 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
1264 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
1265 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
1266 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
1267 { X86::VDIVSDrr_Int, X86::VDIVSDrm_Int, 0 },
1268 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
1269 { X86::VDIVSSrr_Int, X86::VDIVSSrm_Int, 0 },
1270 { X86::VDPPDrri, X86::VDPPDrmi, 0 },
1271 { X86::VDPPSrri, X86::VDPPSrmi, 0 },
1272 // Do not fold VFs* loads because there are no scalar load variants for
1273 // these instructions. When folded, the load is required to be 128-bits, so
1274 // the load size would not match.
1275 { X86::VFvANDNPDrr, X86::VFvANDNPDrm, 0 },
1276 { X86::VFvANDNPSrr, X86::VFvANDNPSrm, 0 },
1277 { X86::VFvANDPDrr, X86::VFvANDPDrm, 0 },
1278 { X86::VFvANDPSrr, X86::VFvANDPSrm, 0 },
1279 { X86::VFvORPDrr, X86::VFvORPDrm, 0 },
1280 { X86::VFvORPSrr, X86::VFvORPSrm, 0 },
1281 { X86::VFvXORPDrr, X86::VFvXORPDrm, 0 },
1282 { X86::VFvXORPSrr, X86::VFvXORPSrm, 0 },
1283 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
1284 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
1285 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
1286 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
1287 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
1288 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
1289 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
1290 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
1291 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
1292 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
1293 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
1294 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
1295 { X86::VMINPDrr, X86::VMINPDrm, 0 },
1296 { X86::VMINPSrr, X86::VMINPSrm, 0 },
1297 { X86::VMINSDrr, X86::VMINSDrm, 0 },
1298 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
1299 { X86::VMINSSrr, X86::VMINSSrm, 0 },
1300 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
1301 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
1302 { X86::VMULPDrr, X86::VMULPDrm, 0 },
1303 { X86::VMULPSrr, X86::VMULPSrm, 0 },
1304 { X86::VMULSDrr, X86::VMULSDrm, 0 },
1305 { X86::VMULSDrr_Int, X86::VMULSDrm_Int, 0 },
1306 { X86::VMULSSrr, X86::VMULSSrm, 0 },
1307 { X86::VMULSSrr_Int, X86::VMULSSrm_Int, 0 },
1308 { X86::VORPDrr, X86::VORPDrm, 0 },
1309 { X86::VORPSrr, X86::VORPSrm, 0 },
1310 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
1311 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
1312 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
1313 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
1314 { X86::VPADDBrr, X86::VPADDBrm, 0 },
1315 { X86::VPADDDrr, X86::VPADDDrm, 0 },
1316 { X86::VPADDQrr, X86::VPADDQrm, 0 },
1317 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
1318 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
1319 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
1320 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
1321 { X86::VPADDWrr, X86::VPADDWrm, 0 },
1322 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
1323 { X86::VPANDNrr, X86::VPANDNrm, 0 },
1324 { X86::VPANDrr, X86::VPANDrm, 0 },
1325 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
1326 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
1327 { X86::VPBLENDVBrr, X86::VPBLENDVBrm, 0 },
1328 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
1329 { X86::VPCLMULQDQrr, X86::VPCLMULQDQrm, 0 },
1330 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
1331 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
1332 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
1333 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
1334 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
1335 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
1336 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
1337 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
1338 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
1339 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
1340 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
1341 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
1342 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
1343 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
1344 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
1345 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
1346 { X86::VPINSRBrr, X86::VPINSRBrm, 0 },
1347 { X86::VPINSRDrr, X86::VPINSRDrm, 0 },
1348 { X86::VPINSRQrr, X86::VPINSRQrm, 0 },
1349 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
1350 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
1351 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
1352 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
1353 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
1354 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
1355 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
1356 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
1357 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
1358 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
1359 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
1360 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
1361 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
1362 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
1363 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
1364 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
1365 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
1366 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
1367 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
1368 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
1369 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
1370 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
1371 { X86::VPORrr, X86::VPORrm, 0 },
1372 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
1373 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
1374 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
1375 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
1376 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
1377 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
1378 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
1379 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
1380 { X86::VPSRADrr, X86::VPSRADrm, 0 },
1381 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
1382 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
1383 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
1384 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
1385 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
1386 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
1387 { X86::VPSUBQrr, X86::VPSUBQrm, 0 },
1388 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
1389 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
1390 { X86::VPSUBUSBrr, X86::VPSUBUSBrm, 0 },
1391 { X86::VPSUBUSWrr, X86::VPSUBUSWrm, 0 },
1392 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
1393 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
1394 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
1395 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
1396 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
1397 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
1398 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
1399 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
1400 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
1401 { X86::VPXORrr, X86::VPXORrm, 0 },
1402 { X86::VROUNDSDr, X86::VROUNDSDm, 0 },
1403 { X86::VROUNDSSr, X86::VROUNDSSm, 0 },
1404 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
1405 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
1406 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
1407 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
1408 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
1409 { X86::VSUBSDrr_Int, X86::VSUBSDrm_Int, 0 },
1410 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
1411 { X86::VSUBSSrr_Int, X86::VSUBSSrm_Int, 0 },
1412 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
1413 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
1414 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
1415 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
1416 { X86::VXORPDrr, X86::VXORPDrm, 0 },
1417 { X86::VXORPSrr, X86::VXORPSrm, 0 },
1419 // AVX 256-bit foldable instructions
1420 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
1421 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
1422 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
1423 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
1424 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1425 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1426 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1427 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1428 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1429 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1430 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1431 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1432 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1433 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1434 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1435 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1436 { X86::VDPPSYrri, X86::VDPPSYrmi, 0 },
1437 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1438 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1439 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1440 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1441 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1442 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
1443 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
1444 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
1445 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
1446 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1447 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1448 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1449 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1450 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1451 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1452 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1453 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1454 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1455 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1456 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1457 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1458 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1459 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1460 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1461 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1462 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
1464 // AVX2 foldable instructions
1465 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1466 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1467 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1468 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1469 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1470 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1471 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1472 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1473 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1474 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1475 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1476 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1477 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1478 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1479 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1480 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1481 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1482 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1483 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1484 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1485 { X86::VPBLENDVBYrr, X86::VPBLENDVBYrm, 0 },
1486 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1487 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1488 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1489 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1490 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1491 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1492 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1493 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1494 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1495 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1496 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1497 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1498 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1499 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1500 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1501 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1502 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1503 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1504 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1505 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1506 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1507 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1508 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1509 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1510 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1511 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1512 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1513 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1514 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1515 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1516 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1517 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1518 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1519 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1520 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1521 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1522 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1523 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1524 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1525 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1526 { X86::VPORYrr, X86::VPORYrm, 0 },
1527 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1528 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1529 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1530 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1531 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1532 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1533 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1534 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1535 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1536 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1537 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1538 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1539 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1540 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1541 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1542 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1543 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1544 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1545 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1546 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1547 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1548 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1549 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1550 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1551 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1552 { X86::VPSUBQYrr, X86::VPSUBQYrm, 0 },
1553 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1554 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1555 { X86::VPSUBUSBYrr, X86::VPSUBUSBYrm, 0 },
1556 { X86::VPSUBUSWYrr, X86::VPSUBUSWYrm, 0 },
1557 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1558 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1559 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1560 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1561 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1562 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1563 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1564 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1565 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1566 { X86::VPXORYrr, X86::VPXORYrm, 0 },
1568 // FMA4 foldable patterns
1569 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, TB_ALIGN_NONE },
1570 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, TB_ALIGN_NONE },
1571 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_NONE },
1572 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_NONE },
1573 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_NONE },
1574 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_NONE },
1575 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, TB_ALIGN_NONE },
1576 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, TB_ALIGN_NONE },
1577 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_NONE },
1578 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_NONE },
1579 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_NONE },
1580 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_NONE },
1581 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, TB_ALIGN_NONE },
1582 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, TB_ALIGN_NONE },
1583 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_NONE },
1584 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_NONE },
1585 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_NONE },
1586 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_NONE },
1587 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, TB_ALIGN_NONE },
1588 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, TB_ALIGN_NONE },
1589 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_NONE },
1590 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_NONE },
1591 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_NONE },
1592 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_NONE },
1593 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_NONE },
1594 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_NONE },
1595 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_NONE },
1596 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_NONE },
1597 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_NONE },
1598 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_NONE },
1599 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_NONE },
1600 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_NONE },
1602 // XOP foldable instructions
1603 { X86::VPCMOVrr, X86::VPCMOVmr, 0 },
1604 { X86::VPCMOVrrY, X86::VPCMOVmrY, 0 },
1605 { X86::VPCOMBri, X86::VPCOMBmi, 0 },
1606 { X86::VPCOMDri, X86::VPCOMDmi, 0 },
1607 { X86::VPCOMQri, X86::VPCOMQmi, 0 },
1608 { X86::VPCOMWri, X86::VPCOMWmi, 0 },
1609 { X86::VPCOMUBri, X86::VPCOMUBmi, 0 },
1610 { X86::VPCOMUDri, X86::VPCOMUDmi, 0 },
1611 { X86::VPCOMUQri, X86::VPCOMUQmi, 0 },
1612 { X86::VPCOMUWri, X86::VPCOMUWmi, 0 },
1613 { X86::VPERMIL2PDrr, X86::VPERMIL2PDmr, 0 },
1614 { X86::VPERMIL2PDrrY, X86::VPERMIL2PDmrY, 0 },
1615 { X86::VPERMIL2PSrr, X86::VPERMIL2PSmr, 0 },
1616 { X86::VPERMIL2PSrrY, X86::VPERMIL2PSmrY, 0 },
1617 { X86::VPMACSDDrr, X86::VPMACSDDrm, 0 },
1618 { X86::VPMACSDQHrr, X86::VPMACSDQHrm, 0 },
1619 { X86::VPMACSDQLrr, X86::VPMACSDQLrm, 0 },
1620 { X86::VPMACSSDDrr, X86::VPMACSSDDrm, 0 },
1621 { X86::VPMACSSDQHrr, X86::VPMACSSDQHrm, 0 },
1622 { X86::VPMACSSDQLrr, X86::VPMACSSDQLrm, 0 },
1623 { X86::VPMACSSWDrr, X86::VPMACSSWDrm, 0 },
1624 { X86::VPMACSSWWrr, X86::VPMACSSWWrm, 0 },
1625 { X86::VPMACSWDrr, X86::VPMACSWDrm, 0 },
1626 { X86::VPMACSWWrr, X86::VPMACSWWrm, 0 },
1627 { X86::VPMADCSSWDrr, X86::VPMADCSSWDrm, 0 },
1628 { X86::VPMADCSWDrr, X86::VPMADCSWDrm, 0 },
1629 { X86::VPPERMrr, X86::VPPERMmr, 0 },
1630 { X86::VPROTBrr, X86::VPROTBrm, 0 },
1631 { X86::VPROTDrr, X86::VPROTDrm, 0 },
1632 { X86::VPROTQrr, X86::VPROTQrm, 0 },
1633 { X86::VPROTWrr, X86::VPROTWrm, 0 },
1634 { X86::VPSHABrr, X86::VPSHABrm, 0 },
1635 { X86::VPSHADrr, X86::VPSHADrm, 0 },
1636 { X86::VPSHAQrr, X86::VPSHAQrm, 0 },
1637 { X86::VPSHAWrr, X86::VPSHAWrm, 0 },
1638 { X86::VPSHLBrr, X86::VPSHLBrm, 0 },
1639 { X86::VPSHLDrr, X86::VPSHLDrm, 0 },
1640 { X86::VPSHLQrr, X86::VPSHLQrm, 0 },
1641 { X86::VPSHLWrr, X86::VPSHLWrm, 0 },
1643 // BMI/BMI2 foldable instructions
1644 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1645 { X86::ANDN64rr, X86::ANDN64rm, 0 },
1646 { X86::MULX32rr, X86::MULX32rm, 0 },
1647 { X86::MULX64rr, X86::MULX64rm, 0 },
1648 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1649 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1650 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1651 { X86::PEXT64rr, X86::PEXT64rm, 0 },
1653 // AVX-512 foldable instructions
1654 { X86::VADDPSZrr, X86::VADDPSZrm, 0 },
1655 { X86::VADDPDZrr, X86::VADDPDZrm, 0 },
1656 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 },
1657 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 },
1658 { X86::VMULPSZrr, X86::VMULPSZrm, 0 },
1659 { X86::VMULPDZrr, X86::VMULPDZrm, 0 },
1660 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 },
1661 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 },
1662 { X86::VMINPSZrr, X86::VMINPSZrm, 0 },
1663 { X86::VMINPDZrr, X86::VMINPDZrm, 0 },
1664 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 },
1665 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 },
1666 { X86::VPADDDZrr, X86::VPADDDZrm, 0 },
1667 { X86::VPADDQZrr, X86::VPADDQZrm, 0 },
1668 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1669 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1670 { X86::VPMAXSDZrr, X86::VPMAXSDZrm, 0 },
1671 { X86::VPMAXSQZrr, X86::VPMAXSQZrm, 0 },
1672 { X86::VPMAXUDZrr, X86::VPMAXUDZrm, 0 },
1673 { X86::VPMAXUQZrr, X86::VPMAXUQZrm, 0 },
1674 { X86::VPMINSDZrr, X86::VPMINSDZrm, 0 },
1675 { X86::VPMINSQZrr, X86::VPMINSQZrm, 0 },
1676 { X86::VPMINUDZrr, X86::VPMINUDZrm, 0 },
1677 { X86::VPMINUQZrr, X86::VPMINUQZrm, 0 },
1678 { X86::VPMULDQZrr, X86::VPMULDQZrm, 0 },
1679 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 },
1680 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 },
1681 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 },
1682 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 },
1683 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 },
1684 { X86::VPSUBDZrr, X86::VPSUBDZrm, 0 },
1685 { X86::VPSUBQZrr, X86::VPSUBQZrm, 0 },
1686 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 },
1687 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 },
1688 { X86::VALIGNQZrri, X86::VALIGNQZrmi, 0 },
1689 { X86::VALIGNDZrri, X86::VALIGNDZrmi, 0 },
1690 { X86::VPMULUDQZrr, X86::VPMULUDQZrm, 0 },
1691 { X86::VBROADCASTSSZrkz, X86::VBROADCASTSSZmkz, TB_NO_REVERSE },
1692 { X86::VBROADCASTSDZrkz, X86::VBROADCASTSDZmkz, TB_NO_REVERSE },
1694 // AVX-512{F,VL} foldable instructions
1695 { X86::VBROADCASTSSZ256rkz, X86::VBROADCASTSSZ256mkz, TB_NO_REVERSE },
1696 { X86::VBROADCASTSDZ256rkz, X86::VBROADCASTSDZ256mkz, TB_NO_REVERSE },
1697 { X86::VBROADCASTSSZ128rkz, X86::VBROADCASTSSZ128mkz, TB_NO_REVERSE },
1699 // AVX-512{F,VL} foldable instructions
1700 { X86::VADDPDZ128rr, X86::VADDPDZ128rm, 0 },
1701 { X86::VADDPDZ256rr, X86::VADDPDZ256rm, 0 },
1702 { X86::VADDPSZ128rr, X86::VADDPSZ128rm, 0 },
1703 { X86::VADDPSZ256rr, X86::VADDPSZ256rm, 0 },
1705 // AES foldable instructions
1706 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 },
1707 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 },
1708 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 },
1709 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 },
1710 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, 0 },
1711 { X86::VAESDECrr, X86::VAESDECrm, 0 },
1712 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, 0 },
1713 { X86::VAESENCrr, X86::VAESENCrm, 0 },
1715 // SHA foldable instructions
1716 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 },
1717 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 },
1718 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 },
1719 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 },
1720 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 },
1721 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 },
1722 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 }
1725 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2) {
1726 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1727 Entry.RegOp, Entry.MemOp,
1728 // Index 2, folded load
1729 Entry.Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1732 static const X86MemoryFoldTableEntry MemoryFoldTable3[] = {
1733 // FMA foldable instructions
1734 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, TB_ALIGN_NONE },
1735 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, TB_ALIGN_NONE },
1736 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, TB_ALIGN_NONE },
1737 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, TB_ALIGN_NONE },
1738 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, TB_ALIGN_NONE },
1739 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, TB_ALIGN_NONE },
1741 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_NONE },
1742 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_NONE },
1743 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_NONE },
1744 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_NONE },
1745 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_NONE },
1746 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_NONE },
1747 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_NONE },
1748 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_NONE },
1749 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_NONE },
1750 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_NONE },
1751 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_NONE },
1752 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_NONE },
1754 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, TB_ALIGN_NONE },
1755 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, TB_ALIGN_NONE },
1756 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, TB_ALIGN_NONE },
1757 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, TB_ALIGN_NONE },
1758 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, TB_ALIGN_NONE },
1759 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, TB_ALIGN_NONE },
1761 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_NONE },
1762 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_NONE },
1763 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_NONE },
1764 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_NONE },
1765 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_NONE },
1766 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_NONE },
1767 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_NONE },
1768 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_NONE },
1769 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_NONE },
1770 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_NONE },
1771 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_NONE },
1772 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_NONE },
1774 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, TB_ALIGN_NONE },
1775 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, TB_ALIGN_NONE },
1776 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, TB_ALIGN_NONE },
1777 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, TB_ALIGN_NONE },
1778 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, TB_ALIGN_NONE },
1779 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, TB_ALIGN_NONE },
1781 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_NONE },
1782 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_NONE },
1783 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_NONE },
1784 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_NONE },
1785 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_NONE },
1786 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_NONE },
1787 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_NONE },
1788 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_NONE },
1789 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_NONE },
1790 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_NONE },
1791 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_NONE },
1792 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_NONE },
1794 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, TB_ALIGN_NONE },
1795 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, TB_ALIGN_NONE },
1796 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, TB_ALIGN_NONE },
1797 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, TB_ALIGN_NONE },
1798 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, TB_ALIGN_NONE },
1799 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, TB_ALIGN_NONE },
1801 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_NONE },
1802 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_NONE },
1803 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_NONE },
1804 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_NONE },
1805 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_NONE },
1806 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_NONE },
1807 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_NONE },
1808 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_NONE },
1809 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_NONE },
1810 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_NONE },
1811 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_NONE },
1812 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_NONE },
1814 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_NONE },
1815 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_NONE },
1816 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_NONE },
1817 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_NONE },
1818 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_NONE },
1819 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_NONE },
1820 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_NONE },
1821 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_NONE },
1822 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_NONE },
1823 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_NONE },
1824 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_NONE },
1825 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_NONE },
1827 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_NONE },
1828 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_NONE },
1829 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_NONE },
1830 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_NONE },
1831 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_NONE },
1832 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_NONE },
1833 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_NONE },
1834 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_NONE },
1835 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_NONE },
1836 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_NONE },
1837 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_NONE },
1838 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_NONE },
1840 // FMA4 foldable patterns
1841 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, TB_ALIGN_NONE },
1842 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, TB_ALIGN_NONE },
1843 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_NONE },
1844 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_NONE },
1845 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_NONE },
1846 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_NONE },
1847 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, TB_ALIGN_NONE },
1848 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, TB_ALIGN_NONE },
1849 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_NONE },
1850 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_NONE },
1851 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_NONE },
1852 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_NONE },
1853 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, TB_ALIGN_NONE },
1854 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, TB_ALIGN_NONE },
1855 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_NONE },
1856 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_NONE },
1857 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_NONE },
1858 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_NONE },
1859 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, TB_ALIGN_NONE },
1860 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, TB_ALIGN_NONE },
1861 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_NONE },
1862 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_NONE },
1863 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_NONE },
1864 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_NONE },
1865 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_NONE },
1866 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_NONE },
1867 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_NONE },
1868 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_NONE },
1869 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_NONE },
1870 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_NONE },
1871 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_NONE },
1872 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_NONE },
1874 // XOP foldable instructions
1875 { X86::VPCMOVrr, X86::VPCMOVrm, 0 },
1876 { X86::VPCMOVrrY, X86::VPCMOVrmY, 0 },
1877 { X86::VPERMIL2PDrr, X86::VPERMIL2PDrm, 0 },
1878 { X86::VPERMIL2PDrrY, X86::VPERMIL2PDrmY, 0 },
1879 { X86::VPERMIL2PSrr, X86::VPERMIL2PSrm, 0 },
1880 { X86::VPERMIL2PSrrY, X86::VPERMIL2PSrmY, 0 },
1881 { X86::VPPERMrr, X86::VPPERMrm, 0 },
1883 // AVX-512 VPERMI instructions with 3 source operands.
1884 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1885 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1886 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1887 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
1888 { X86::VBLENDMPDZrr, X86::VBLENDMPDZrm, 0 },
1889 { X86::VBLENDMPSZrr, X86::VBLENDMPSZrm, 0 },
1890 { X86::VPBLENDMDZrr, X86::VPBLENDMDZrm, 0 },
1891 { X86::VPBLENDMQZrr, X86::VPBLENDMQZrm, 0 },
1892 { X86::VBROADCASTSSZrk, X86::VBROADCASTSSZmk, TB_NO_REVERSE },
1893 { X86::VBROADCASTSDZrk, X86::VBROADCASTSDZmk, TB_NO_REVERSE },
1894 { X86::VBROADCASTSSZ256rk, X86::VBROADCASTSSZ256mk, TB_NO_REVERSE },
1895 { X86::VBROADCASTSDZ256rk, X86::VBROADCASTSDZ256mk, TB_NO_REVERSE },
1896 { X86::VBROADCASTSSZ128rk, X86::VBROADCASTSSZ128mk, TB_NO_REVERSE },
1897 // AVX-512 arithmetic instructions
1898 { X86::VADDPSZrrkz, X86::VADDPSZrmkz, 0 },
1899 { X86::VADDPDZrrkz, X86::VADDPDZrmkz, 0 },
1900 { X86::VSUBPSZrrkz, X86::VSUBPSZrmkz, 0 },
1901 { X86::VSUBPDZrrkz, X86::VSUBPDZrmkz, 0 },
1902 { X86::VMULPSZrrkz, X86::VMULPSZrmkz, 0 },
1903 { X86::VMULPDZrrkz, X86::VMULPDZrmkz, 0 },
1904 { X86::VDIVPSZrrkz, X86::VDIVPSZrmkz, 0 },
1905 { X86::VDIVPDZrrkz, X86::VDIVPDZrmkz, 0 },
1906 { X86::VMINPSZrrkz, X86::VMINPSZrmkz, 0 },
1907 { X86::VMINPDZrrkz, X86::VMINPDZrmkz, 0 },
1908 { X86::VMAXPSZrrkz, X86::VMAXPSZrmkz, 0 },
1909 { X86::VMAXPDZrrkz, X86::VMAXPDZrmkz, 0 },
1910 // AVX-512{F,VL} arithmetic instructions 256-bit
1911 { X86::VADDPSZ256rrkz, X86::VADDPSZ256rmkz, 0 },
1912 { X86::VADDPDZ256rrkz, X86::VADDPDZ256rmkz, 0 },
1913 { X86::VSUBPSZ256rrkz, X86::VSUBPSZ256rmkz, 0 },
1914 { X86::VSUBPDZ256rrkz, X86::VSUBPDZ256rmkz, 0 },
1915 { X86::VMULPSZ256rrkz, X86::VMULPSZ256rmkz, 0 },
1916 { X86::VMULPDZ256rrkz, X86::VMULPDZ256rmkz, 0 },
1917 { X86::VDIVPSZ256rrkz, X86::VDIVPSZ256rmkz, 0 },
1918 { X86::VDIVPDZ256rrkz, X86::VDIVPDZ256rmkz, 0 },
1919 { X86::VMINPSZ256rrkz, X86::VMINPSZ256rmkz, 0 },
1920 { X86::VMINPDZ256rrkz, X86::VMINPDZ256rmkz, 0 },
1921 { X86::VMAXPSZ256rrkz, X86::VMAXPSZ256rmkz, 0 },
1922 { X86::VMAXPDZ256rrkz, X86::VMAXPDZ256rmkz, 0 },
1923 // AVX-512{F,VL} arithmetic instructions 128-bit
1924 { X86::VADDPSZ128rrkz, X86::VADDPSZ128rmkz, 0 },
1925 { X86::VADDPDZ128rrkz, X86::VADDPDZ128rmkz, 0 },
1926 { X86::VSUBPSZ128rrkz, X86::VSUBPSZ128rmkz, 0 },
1927 { X86::VSUBPDZ128rrkz, X86::VSUBPDZ128rmkz, 0 },
1928 { X86::VMULPSZ128rrkz, X86::VMULPSZ128rmkz, 0 },
1929 { X86::VMULPDZ128rrkz, X86::VMULPDZ128rmkz, 0 },
1930 { X86::VDIVPSZ128rrkz, X86::VDIVPSZ128rmkz, 0 },
1931 { X86::VDIVPDZ128rrkz, X86::VDIVPDZ128rmkz, 0 },
1932 { X86::VMINPSZ128rrkz, X86::VMINPSZ128rmkz, 0 },
1933 { X86::VMINPDZ128rrkz, X86::VMINPDZ128rmkz, 0 },
1934 { X86::VMAXPSZ128rrkz, X86::VMAXPSZ128rmkz, 0 },
1935 { X86::VMAXPDZ128rrkz, X86::VMAXPDZ128rmkz, 0 }
1938 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable3) {
1939 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1940 Entry.RegOp, Entry.MemOp,
1941 // Index 3, folded load
1942 Entry.Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1945 static const X86MemoryFoldTableEntry MemoryFoldTable4[] = {
1946 // AVX-512 foldable instructions
1947 { X86::VADDPSZrrk, X86::VADDPSZrmk, 0 },
1948 { X86::VADDPDZrrk, X86::VADDPDZrmk, 0 },
1949 { X86::VSUBPSZrrk, X86::VSUBPSZrmk, 0 },
1950 { X86::VSUBPDZrrk, X86::VSUBPDZrmk, 0 },
1951 { X86::VMULPSZrrk, X86::VMULPSZrmk, 0 },
1952 { X86::VMULPDZrrk, X86::VMULPDZrmk, 0 },
1953 { X86::VDIVPSZrrk, X86::VDIVPSZrmk, 0 },
1954 { X86::VDIVPDZrrk, X86::VDIVPDZrmk, 0 },
1955 { X86::VMINPSZrrk, X86::VMINPSZrmk, 0 },
1956 { X86::VMINPDZrrk, X86::VMINPDZrmk, 0 },
1957 { X86::VMAXPSZrrk, X86::VMAXPSZrmk, 0 },
1958 { X86::VMAXPDZrrk, X86::VMAXPDZrmk, 0 },
1959 // AVX-512{F,VL} foldable instructions 256-bit
1960 { X86::VADDPSZ256rrk, X86::VADDPSZ256rmk, 0 },
1961 { X86::VADDPDZ256rrk, X86::VADDPDZ256rmk, 0 },
1962 { X86::VSUBPSZ256rrk, X86::VSUBPSZ256rmk, 0 },
1963 { X86::VSUBPDZ256rrk, X86::VSUBPDZ256rmk, 0 },
1964 { X86::VMULPSZ256rrk, X86::VMULPSZ256rmk, 0 },
1965 { X86::VMULPDZ256rrk, X86::VMULPDZ256rmk, 0 },
1966 { X86::VDIVPSZ256rrk, X86::VDIVPSZ256rmk, 0 },
1967 { X86::VDIVPDZ256rrk, X86::VDIVPDZ256rmk, 0 },
1968 { X86::VMINPSZ256rrk, X86::VMINPSZ256rmk, 0 },
1969 { X86::VMINPDZ256rrk, X86::VMINPDZ256rmk, 0 },
1970 { X86::VMAXPSZ256rrk, X86::VMAXPSZ256rmk, 0 },
1971 { X86::VMAXPDZ256rrk, X86::VMAXPDZ256rmk, 0 },
1972 // AVX-512{F,VL} foldable instructions 128-bit
1973 { X86::VADDPSZ128rrk, X86::VADDPSZ128rmk, 0 },
1974 { X86::VADDPDZ128rrk, X86::VADDPDZ128rmk, 0 },
1975 { X86::VSUBPSZ128rrk, X86::VSUBPSZ128rmk, 0 },
1976 { X86::VSUBPDZ128rrk, X86::VSUBPDZ128rmk, 0 },
1977 { X86::VMULPSZ128rrk, X86::VMULPSZ128rmk, 0 },
1978 { X86::VMULPDZ128rrk, X86::VMULPDZ128rmk, 0 },
1979 { X86::VDIVPSZ128rrk, X86::VDIVPSZ128rmk, 0 },
1980 { X86::VDIVPDZ128rrk, X86::VDIVPDZ128rmk, 0 },
1981 { X86::VMINPSZ128rrk, X86::VMINPSZ128rmk, 0 },
1982 { X86::VMINPDZ128rrk, X86::VMINPDZ128rmk, 0 },
1983 { X86::VMAXPSZ128rrk, X86::VMAXPSZ128rmk, 0 },
1984 { X86::VMAXPDZ128rrk, X86::VMAXPDZ128rmk, 0 }
1987 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable4) {
1988 AddTableEntry(RegOp2MemOpTable4, MemOp2RegOpTable,
1989 Entry.RegOp, Entry.MemOp,
1990 // Index 4, folded load
1991 Entry.Flags | TB_INDEX_4 | TB_FOLDED_LOAD);
1996 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1997 MemOp2RegOpTableType &M2RTable,
1998 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1999 if ((Flags & TB_NO_FORWARD) == 0) {
2000 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
2001 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
2003 if ((Flags & TB_NO_REVERSE) == 0) {
2004 assert(!M2RTable.count(MemOp) &&
2005 "Duplicated entries in unfolding maps?");
2006 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
2011 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
2012 unsigned &SrcReg, unsigned &DstReg,
2013 unsigned &SubIdx) const {
2014 switch (MI.getOpcode()) {
2016 case X86::MOVSX16rr8:
2017 case X86::MOVZX16rr8:
2018 case X86::MOVSX32rr8:
2019 case X86::MOVZX32rr8:
2020 case X86::MOVSX64rr8:
2021 if (!Subtarget.is64Bit())
2022 // It's not always legal to reference the low 8-bit of the larger
2023 // register in 32-bit mode.
2025 case X86::MOVSX32rr16:
2026 case X86::MOVZX32rr16:
2027 case X86::MOVSX64rr16:
2028 case X86::MOVSX64rr32: {
2029 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
2032 SrcReg = MI.getOperand(1).getReg();
2033 DstReg = MI.getOperand(0).getReg();
2034 switch (MI.getOpcode()) {
2035 default: llvm_unreachable("Unreachable!");
2036 case X86::MOVSX16rr8:
2037 case X86::MOVZX16rr8:
2038 case X86::MOVSX32rr8:
2039 case X86::MOVZX32rr8:
2040 case X86::MOVSX64rr8:
2041 SubIdx = X86::sub_8bit;
2043 case X86::MOVSX32rr16:
2044 case X86::MOVZX32rr16:
2045 case X86::MOVSX64rr16:
2046 SubIdx = X86::sub_16bit;
2048 case X86::MOVSX64rr32:
2049 SubIdx = X86::sub_32bit;
2058 int X86InstrInfo::getSPAdjust(const MachineInstr *MI) const {
2059 const MachineFunction *MF = MI->getParent()->getParent();
2060 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2062 if (MI->getOpcode() == getCallFrameSetupOpcode() ||
2063 MI->getOpcode() == getCallFrameDestroyOpcode()) {
2064 unsigned StackAlign = TFI->getStackAlignment();
2065 int SPAdj = (MI->getOperand(0).getImm() + StackAlign - 1) / StackAlign *
2068 SPAdj -= MI->getOperand(1).getImm();
2070 if (MI->getOpcode() == getCallFrameSetupOpcode())
2076 // To know whether a call adjusts the stack, we need information
2077 // that is bound to the following ADJCALLSTACKUP pseudo.
2078 // Look for the next ADJCALLSTACKUP that follows the call.
2080 const MachineBasicBlock* MBB = MI->getParent();
2081 auto I = ++MachineBasicBlock::const_iterator(MI);
2082 for (auto E = MBB->end(); I != E; ++I) {
2083 if (I->getOpcode() == getCallFrameDestroyOpcode() ||
2088 // If we could not find a frame destroy opcode, then it has already
2089 // been simplified, so we don't care.
2090 if (I->getOpcode() != getCallFrameDestroyOpcode())
2093 return -(I->getOperand(1).getImm());
2096 // Currently handle only PUSHes we can reasonably expect to see
2097 // in call sequences
2098 switch (MI->getOpcode()) {
2103 case X86::PUSH32rmm:
2104 case X86::PUSH32rmr:
2110 /// Return true and the FrameIndex if the specified
2111 /// operand and follow operands form a reference to the stack frame.
2112 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
2113 int &FrameIndex) const {
2114 if (MI->getOperand(Op+X86::AddrBaseReg).isFI() &&
2115 MI->getOperand(Op+X86::AddrScaleAmt).isImm() &&
2116 MI->getOperand(Op+X86::AddrIndexReg).isReg() &&
2117 MI->getOperand(Op+X86::AddrDisp).isImm() &&
2118 MI->getOperand(Op+X86::AddrScaleAmt).getImm() == 1 &&
2119 MI->getOperand(Op+X86::AddrIndexReg).getReg() == 0 &&
2120 MI->getOperand(Op+X86::AddrDisp).getImm() == 0) {
2121 FrameIndex = MI->getOperand(Op+X86::AddrBaseReg).getIndex();
2127 static bool isFrameLoadOpcode(int Opcode) {
2143 case X86::VMOVAPSrm:
2144 case X86::VMOVAPDrm:
2145 case X86::VMOVDQArm:
2146 case X86::VMOVUPSYrm:
2147 case X86::VMOVAPSYrm:
2148 case X86::VMOVUPDYrm:
2149 case X86::VMOVAPDYrm:
2150 case X86::VMOVDQUYrm:
2151 case X86::VMOVDQAYrm:
2152 case X86::MMX_MOVD64rm:
2153 case X86::MMX_MOVQ64rm:
2154 case X86::VMOVAPSZrm:
2155 case X86::VMOVUPSZrm:
2160 static bool isFrameStoreOpcode(int Opcode) {
2167 case X86::ST_FpP64m:
2175 case X86::VMOVAPSmr:
2176 case X86::VMOVAPDmr:
2177 case X86::VMOVDQAmr:
2178 case X86::VMOVUPSYmr:
2179 case X86::VMOVAPSYmr:
2180 case X86::VMOVUPDYmr:
2181 case X86::VMOVAPDYmr:
2182 case X86::VMOVDQUYmr:
2183 case X86::VMOVDQAYmr:
2184 case X86::VMOVUPSZmr:
2185 case X86::VMOVAPSZmr:
2186 case X86::MMX_MOVD64mr:
2187 case X86::MMX_MOVQ64mr:
2188 case X86::MMX_MOVNTQmr:
2194 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
2195 int &FrameIndex) const {
2196 if (isFrameLoadOpcode(MI->getOpcode()))
2197 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
2198 return MI->getOperand(0).getReg();
2202 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
2203 int &FrameIndex) const {
2204 if (isFrameLoadOpcode(MI->getOpcode())) {
2206 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
2208 // Check for post-frame index elimination operations
2209 const MachineMemOperand *Dummy;
2210 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
2215 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
2216 int &FrameIndex) const {
2217 if (isFrameStoreOpcode(MI->getOpcode()))
2218 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
2219 isFrameOperand(MI, 0, FrameIndex))
2220 return MI->getOperand(X86::AddrNumOperands).getReg();
2224 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
2225 int &FrameIndex) const {
2226 if (isFrameStoreOpcode(MI->getOpcode())) {
2228 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
2230 // Check for post-frame index elimination operations
2231 const MachineMemOperand *Dummy;
2232 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
2237 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
2238 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
2239 // Don't waste compile time scanning use-def chains of physregs.
2240 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
2242 bool isPICBase = false;
2243 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
2244 E = MRI.def_instr_end(); I != E; ++I) {
2245 MachineInstr *DefMI = &*I;
2246 if (DefMI->getOpcode() != X86::MOVPC32r)
2248 assert(!isPICBase && "More than one PIC base?");
2255 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
2256 AliasAnalysis *AA) const {
2257 switch (MI->getOpcode()) {
2273 case X86::VMOVAPSrm:
2274 case X86::VMOVUPSrm:
2275 case X86::VMOVAPDrm:
2276 case X86::VMOVDQArm:
2277 case X86::VMOVDQUrm:
2278 case X86::VMOVAPSYrm:
2279 case X86::VMOVUPSYrm:
2280 case X86::VMOVAPDYrm:
2281 case X86::VMOVDQAYrm:
2282 case X86::VMOVDQUYrm:
2283 case X86::MMX_MOVD64rm:
2284 case X86::MMX_MOVQ64rm:
2285 case X86::FsVMOVAPSrm:
2286 case X86::FsVMOVAPDrm:
2287 case X86::FsMOVAPSrm:
2288 case X86::FsMOVAPDrm: {
2289 // Loads from constant pools are trivially rematerializable.
2290 if (MI->getOperand(1+X86::AddrBaseReg).isReg() &&
2291 MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
2292 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
2293 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
2294 MI->isInvariantLoad(AA)) {
2295 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
2296 if (BaseReg == 0 || BaseReg == X86::RIP)
2298 // Allow re-materialization of PIC load.
2299 if (!ReMatPICStubLoad && MI->getOperand(1+X86::AddrDisp).isGlobal())
2301 const MachineFunction &MF = *MI->getParent()->getParent();
2302 const MachineRegisterInfo &MRI = MF.getRegInfo();
2303 return regIsPICBase(BaseReg, MRI);
2310 if (MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
2311 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
2312 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
2313 !MI->getOperand(1+X86::AddrDisp).isReg()) {
2314 // lea fi#, lea GV, etc. are all rematerializable.
2315 if (!MI->getOperand(1+X86::AddrBaseReg).isReg())
2317 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
2320 // Allow re-materialization of lea PICBase + x.
2321 const MachineFunction &MF = *MI->getParent()->getParent();
2322 const MachineRegisterInfo &MRI = MF.getRegInfo();
2323 return regIsPICBase(BaseReg, MRI);
2329 // All other instructions marked M_REMATERIALIZABLE are always trivially
2330 // rematerializable.
2334 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
2335 MachineBasicBlock::iterator I) const {
2336 MachineBasicBlock::iterator E = MBB.end();
2338 // For compile time consideration, if we are not able to determine the
2339 // safety after visiting 4 instructions in each direction, we will assume
2341 MachineBasicBlock::iterator Iter = I;
2342 for (unsigned i = 0; Iter != E && i < 4; ++i) {
2343 bool SeenDef = false;
2344 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
2345 MachineOperand &MO = Iter->getOperand(j);
2346 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2350 if (MO.getReg() == X86::EFLAGS) {
2358 // This instruction defines EFLAGS, no need to look any further.
2361 // Skip over DBG_VALUE.
2362 while (Iter != E && Iter->isDebugValue())
2366 // It is safe to clobber EFLAGS at the end of a block of no successor has it
2369 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
2370 SE = MBB.succ_end(); SI != SE; ++SI)
2371 if ((*SI)->isLiveIn(X86::EFLAGS))
2376 MachineBasicBlock::iterator B = MBB.begin();
2378 for (unsigned i = 0; i < 4; ++i) {
2379 // If we make it to the beginning of the block, it's safe to clobber
2380 // EFLAGS iff EFLAGS is not live-in.
2382 return !MBB.isLiveIn(X86::EFLAGS);
2385 // Skip over DBG_VALUE.
2386 while (Iter != B && Iter->isDebugValue())
2389 bool SawKill = false;
2390 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
2391 MachineOperand &MO = Iter->getOperand(j);
2392 // A register mask may clobber EFLAGS, but we should still look for a
2394 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2396 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
2397 if (MO.isDef()) return MO.isDead();
2398 if (MO.isKill()) SawKill = true;
2403 // This instruction kills EFLAGS and doesn't redefine it, so
2404 // there's no need to look further.
2408 // Conservative answer.
2412 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
2413 MachineBasicBlock::iterator I,
2414 unsigned DestReg, unsigned SubIdx,
2415 const MachineInstr *Orig,
2416 const TargetRegisterInfo &TRI) const {
2417 // MOV32r0 is implemented with a xor which clobbers condition code.
2418 // Re-materialize it as movri instructions to avoid side effects.
2419 unsigned Opc = Orig->getOpcode();
2420 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
2421 DebugLoc DL = Orig->getDebugLoc();
2422 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
2425 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
2429 MachineInstr *NewMI = std::prev(I);
2430 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
2433 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
2434 static bool hasLiveCondCodeDef(MachineInstr *MI) {
2435 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2436 MachineOperand &MO = MI->getOperand(i);
2437 if (MO.isReg() && MO.isDef() &&
2438 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
2445 /// Check whether the shift count for a machine operand is non-zero.
2446 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
2447 unsigned ShiftAmtOperandIdx) {
2448 // The shift count is six bits with the REX.W prefix and five bits without.
2449 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
2450 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
2451 return Imm & ShiftCountMask;
2454 /// Check whether the given shift count is appropriate
2455 /// can be represented by a LEA instruction.
2456 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
2457 // Left shift instructions can be transformed into load-effective-address
2458 // instructions if we can encode them appropriately.
2459 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
2460 // The SIB.scale field is two bits wide which means that we can encode any
2461 // shift amount less than 4.
2462 return ShAmt < 4 && ShAmt > 0;
2465 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
2466 unsigned Opc, bool AllowSP,
2467 unsigned &NewSrc, bool &isKill, bool &isUndef,
2468 MachineOperand &ImplicitOp) const {
2469 MachineFunction &MF = *MI->getParent()->getParent();
2470 const TargetRegisterClass *RC;
2472 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
2474 RC = Opc != X86::LEA32r ?
2475 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
2477 unsigned SrcReg = Src.getReg();
2479 // For both LEA64 and LEA32 the register already has essentially the right
2480 // type (32-bit or 64-bit) we may just need to forbid SP.
2481 if (Opc != X86::LEA64_32r) {
2483 isKill = Src.isKill();
2484 isUndef = Src.isUndef();
2486 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
2487 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
2493 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
2494 // another we need to add 64-bit registers to the final MI.
2495 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
2497 ImplicitOp.setImplicit();
2499 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
2500 MachineBasicBlock::LivenessQueryResult LQR =
2501 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
2504 case MachineBasicBlock::LQR_Unknown:
2505 // We can't give sane liveness flags to the instruction, abandon LEA
2508 case MachineBasicBlock::LQR_Live:
2509 isKill = MI->killsRegister(SrcReg);
2513 // The physreg itself is dead, so we have to use it as an <undef>.
2519 // Virtual register of the wrong class, we have to create a temporary 64-bit
2520 // vreg to feed into the LEA.
2521 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
2522 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
2523 get(TargetOpcode::COPY))
2524 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2527 // Which is obviously going to be dead after we're done with it.
2532 // We've set all the parameters without issue.
2536 /// Helper for convertToThreeAddress when 16-bit LEA is disabled, use 32-bit
2537 /// LEA to form 3-address code by promoting to a 32-bit superregister and then
2538 /// truncating back down to a 16-bit subregister.
2540 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
2541 MachineFunction::iterator &MFI,
2542 MachineBasicBlock::iterator &MBBI,
2543 LiveVariables *LV) const {
2544 MachineInstr *MI = MBBI;
2545 unsigned Dest = MI->getOperand(0).getReg();
2546 unsigned Src = MI->getOperand(1).getReg();
2547 bool isDead = MI->getOperand(0).isDead();
2548 bool isKill = MI->getOperand(1).isKill();
2550 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
2551 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
2552 unsigned Opc, leaInReg;
2553 if (Subtarget.is64Bit()) {
2554 Opc = X86::LEA64_32r;
2555 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2558 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2561 // Build and insert into an implicit UNDEF value. This is OK because
2562 // well be shifting and then extracting the lower 16-bits.
2563 // This has the potential to cause partial register stall. e.g.
2564 // movw (%rbp,%rcx,2), %dx
2565 // leal -65(%rdx), %esi
2566 // But testing has shown this *does* help performance in 64-bit mode (at
2567 // least on modern x86 machines).
2568 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
2569 MachineInstr *InsMI =
2570 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2571 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
2572 .addReg(Src, getKillRegState(isKill));
2574 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
2575 get(Opc), leaOutReg);
2577 default: llvm_unreachable("Unreachable!");
2578 case X86::SHL16ri: {
2579 unsigned ShAmt = MI->getOperand(2).getImm();
2580 MIB.addReg(0).addImm(1 << ShAmt)
2581 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
2585 addRegOffset(MIB, leaInReg, true, 1);
2588 addRegOffset(MIB, leaInReg, true, -1);
2592 case X86::ADD16ri_DB:
2593 case X86::ADD16ri8_DB:
2594 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
2597 case X86::ADD16rr_DB: {
2598 unsigned Src2 = MI->getOperand(2).getReg();
2599 bool isKill2 = MI->getOperand(2).isKill();
2600 unsigned leaInReg2 = 0;
2601 MachineInstr *InsMI2 = nullptr;
2603 // ADD16rr %reg1028<kill>, %reg1028
2604 // just a single insert_subreg.
2605 addRegReg(MIB, leaInReg, true, leaInReg, false);
2607 if (Subtarget.is64Bit())
2608 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2610 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2611 // Build and insert into an implicit UNDEF value. This is OK because
2612 // well be shifting and then extracting the lower 16-bits.
2613 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2615 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2616 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2617 .addReg(Src2, getKillRegState(isKill2));
2618 addRegReg(MIB, leaInReg, true, leaInReg2, true);
2620 if (LV && isKill2 && InsMI2)
2621 LV->replaceKillInstruction(Src2, MI, InsMI2);
2626 MachineInstr *NewMI = MIB;
2627 MachineInstr *ExtMI =
2628 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2629 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2630 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2633 // Update live variables
2634 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2635 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2637 LV->replaceKillInstruction(Src, MI, InsMI);
2639 LV->replaceKillInstruction(Dest, MI, ExtMI);
2645 /// This method must be implemented by targets that
2646 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
2647 /// may be able to convert a two-address instruction into a true
2648 /// three-address instruction on demand. This allows the X86 target (for
2649 /// example) to convert ADD and SHL instructions into LEA instructions if they
2650 /// would require register copies due to two-addressness.
2652 /// This method returns a null pointer if the transformation cannot be
2653 /// performed, otherwise it returns the new instruction.
2656 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2657 MachineBasicBlock::iterator &MBBI,
2658 LiveVariables *LV) const {
2659 MachineInstr *MI = MBBI;
2661 // The following opcodes also sets the condition code register(s). Only
2662 // convert them to equivalent lea if the condition code register def's
2664 if (hasLiveCondCodeDef(MI))
2667 MachineFunction &MF = *MI->getParent()->getParent();
2668 // All instructions input are two-addr instructions. Get the known operands.
2669 const MachineOperand &Dest = MI->getOperand(0);
2670 const MachineOperand &Src = MI->getOperand(1);
2672 MachineInstr *NewMI = nullptr;
2673 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
2674 // we have better subtarget support, enable the 16-bit LEA generation here.
2675 // 16-bit LEA is also slow on Core2.
2676 bool DisableLEA16 = true;
2677 bool is64Bit = Subtarget.is64Bit();
2679 unsigned MIOpc = MI->getOpcode();
2681 default: return nullptr;
2682 case X86::SHL64ri: {
2683 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2684 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2685 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2687 // LEA can't handle RSP.
2688 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2689 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2690 &X86::GR64_NOSPRegClass))
2693 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2695 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2698 case X86::SHL32ri: {
2699 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2700 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2701 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2703 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2705 // LEA can't handle ESP.
2706 bool isKill, isUndef;
2708 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2709 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2710 SrcReg, isKill, isUndef, ImplicitOp))
2713 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2715 .addReg(0).addImm(1 << ShAmt)
2716 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2717 .addImm(0).addReg(0);
2718 if (ImplicitOp.getReg() != 0)
2719 MIB.addOperand(ImplicitOp);
2724 case X86::SHL16ri: {
2725 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2726 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2727 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2730 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : nullptr;
2731 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2733 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2738 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2739 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2740 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2741 bool isKill, isUndef;
2743 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2744 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2745 SrcReg, isKill, isUndef, ImplicitOp))
2748 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2750 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2751 if (ImplicitOp.getReg() != 0)
2752 MIB.addOperand(ImplicitOp);
2754 NewMI = addOffset(MIB, 1);
2759 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2761 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2762 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2763 .addOperand(Dest).addOperand(Src), 1);
2767 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2768 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2769 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2771 bool isKill, isUndef;
2773 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2774 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2775 SrcReg, isKill, isUndef, ImplicitOp))
2778 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2780 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2781 if (ImplicitOp.getReg() != 0)
2782 MIB.addOperand(ImplicitOp);
2784 NewMI = addOffset(MIB, -1);
2790 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2792 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2793 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2794 .addOperand(Dest).addOperand(Src), -1);
2797 case X86::ADD64rr_DB:
2799 case X86::ADD32rr_DB: {
2800 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2802 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2805 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2807 bool isKill, isUndef;
2809 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2810 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2811 SrcReg, isKill, isUndef, ImplicitOp))
2814 const MachineOperand &Src2 = MI->getOperand(2);
2815 bool isKill2, isUndef2;
2817 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2818 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2819 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2822 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2824 if (ImplicitOp.getReg() != 0)
2825 MIB.addOperand(ImplicitOp);
2826 if (ImplicitOp2.getReg() != 0)
2827 MIB.addOperand(ImplicitOp2);
2829 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2831 // Preserve undefness of the operands.
2832 NewMI->getOperand(1).setIsUndef(isUndef);
2833 NewMI->getOperand(3).setIsUndef(isUndef2);
2835 if (LV && Src2.isKill())
2836 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2840 case X86::ADD16rr_DB: {
2842 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2844 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2845 unsigned Src2 = MI->getOperand(2).getReg();
2846 bool isKill2 = MI->getOperand(2).isKill();
2847 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2849 Src.getReg(), Src.isKill(), Src2, isKill2);
2851 // Preserve undefness of the operands.
2852 bool isUndef = MI->getOperand(1).isUndef();
2853 bool isUndef2 = MI->getOperand(2).isUndef();
2854 NewMI->getOperand(1).setIsUndef(isUndef);
2855 NewMI->getOperand(3).setIsUndef(isUndef2);
2858 LV->replaceKillInstruction(Src2, MI, NewMI);
2861 case X86::ADD64ri32:
2863 case X86::ADD64ri32_DB:
2864 case X86::ADD64ri8_DB:
2865 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2866 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2867 .addOperand(Dest).addOperand(Src),
2868 MI->getOperand(2).getImm());
2872 case X86::ADD32ri_DB:
2873 case X86::ADD32ri8_DB: {
2874 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2875 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2877 bool isKill, isUndef;
2879 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2880 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2881 SrcReg, isKill, isUndef, ImplicitOp))
2884 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2886 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2887 if (ImplicitOp.getReg() != 0)
2888 MIB.addOperand(ImplicitOp);
2890 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2895 case X86::ADD16ri_DB:
2896 case X86::ADD16ri8_DB:
2898 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2900 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2901 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2902 .addOperand(Dest).addOperand(Src),
2903 MI->getOperand(2).getImm());
2907 if (!NewMI) return nullptr;
2909 if (LV) { // Update live variables
2911 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2913 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2916 MFI->insert(MBBI, NewMI); // Insert the new inst
2920 /// We have a few instructions that must be hacked on to commute them.
2923 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2924 switch (MI->getOpcode()) {
2925 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2926 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2927 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2928 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2929 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2930 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2933 switch (MI->getOpcode()) {
2934 default: llvm_unreachable("Unreachable!");
2935 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2936 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2937 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2938 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2939 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2940 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2942 unsigned Amt = MI->getOperand(3).getImm();
2944 MachineFunction &MF = *MI->getParent()->getParent();
2945 MI = MF.CloneMachineInstr(MI);
2948 MI->setDesc(get(Opc));
2949 MI->getOperand(3).setImm(Size-Amt);
2950 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2952 case X86::BLENDPDrri:
2953 case X86::BLENDPSrri:
2954 case X86::PBLENDWrri:
2955 case X86::VBLENDPDrri:
2956 case X86::VBLENDPSrri:
2957 case X86::VBLENDPDYrri:
2958 case X86::VBLENDPSYrri:
2959 case X86::VPBLENDDrri:
2960 case X86::VPBLENDWrri:
2961 case X86::VPBLENDDYrri:
2962 case X86::VPBLENDWYrri:{
2964 switch (MI->getOpcode()) {
2965 default: llvm_unreachable("Unreachable!");
2966 case X86::BLENDPDrri: Mask = 0x03; break;
2967 case X86::BLENDPSrri: Mask = 0x0F; break;
2968 case X86::PBLENDWrri: Mask = 0xFF; break;
2969 case X86::VBLENDPDrri: Mask = 0x03; break;
2970 case X86::VBLENDPSrri: Mask = 0x0F; break;
2971 case X86::VBLENDPDYrri: Mask = 0x0F; break;
2972 case X86::VBLENDPSYrri: Mask = 0xFF; break;
2973 case X86::VPBLENDDrri: Mask = 0x0F; break;
2974 case X86::VPBLENDWrri: Mask = 0xFF; break;
2975 case X86::VPBLENDDYrri: Mask = 0xFF; break;
2976 case X86::VPBLENDWYrri: Mask = 0xFF; break;
2978 // Only the least significant bits of Imm are used.
2979 unsigned Imm = MI->getOperand(3).getImm() & Mask;
2981 MachineFunction &MF = *MI->getParent()->getParent();
2982 MI = MF.CloneMachineInstr(MI);
2985 MI->getOperand(3).setImm(Mask ^ Imm);
2986 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2988 case X86::PCLMULQDQrr:
2989 case X86::VPCLMULQDQrr:{
2990 // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
2991 // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
2992 unsigned Imm = MI->getOperand(3).getImm();
2993 unsigned Src1Hi = Imm & 0x01;
2994 unsigned Src2Hi = Imm & 0x10;
2996 MachineFunction &MF = *MI->getParent()->getParent();
2997 MI = MF.CloneMachineInstr(MI);
3000 MI->getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
3001 return TargetInstrInfo::commuteInstruction(MI, NewMI);
3005 case X86::VCMPPDrri:
3006 case X86::VCMPPSrri:
3007 case X86::VCMPPDYrri:
3008 case X86::VCMPPSYrri: {
3009 // Float comparison can be safely commuted for
3010 // Ordered/Unordered/Equal/NotEqual tests
3011 unsigned Imm = MI->getOperand(3).getImm() & 0x7;
3014 case 0x03: // UNORDERED
3015 case 0x04: // NOT EQUAL
3016 case 0x07: // ORDERED
3018 MachineFunction &MF = *MI->getParent()->getParent();
3019 MI = MF.CloneMachineInstr(MI);
3022 return TargetInstrInfo::commuteInstruction(MI, NewMI);
3027 case X86::VPCOMBri: case X86::VPCOMUBri:
3028 case X86::VPCOMDri: case X86::VPCOMUDri:
3029 case X86::VPCOMQri: case X86::VPCOMUQri:
3030 case X86::VPCOMWri: case X86::VPCOMUWri: {
3031 // Flip comparison mode immediate (if necessary).
3032 unsigned Imm = MI->getOperand(3).getImm() & 0x7;
3034 case 0x00: Imm = 0x02; break; // LT -> GT
3035 case 0x01: Imm = 0x03; break; // LE -> GE
3036 case 0x02: Imm = 0x00; break; // GT -> LT
3037 case 0x03: Imm = 0x01; break; // GE -> LE
3046 MachineFunction &MF = *MI->getParent()->getParent();
3047 MI = MF.CloneMachineInstr(MI);
3050 MI->getOperand(3).setImm(Imm);
3051 return TargetInstrInfo::commuteInstruction(MI, NewMI);
3053 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
3054 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
3055 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
3056 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
3057 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
3058 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
3059 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
3060 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
3061 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
3062 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
3063 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
3064 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
3065 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
3066 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
3067 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
3068 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
3070 switch (MI->getOpcode()) {
3071 default: llvm_unreachable("Unreachable!");
3072 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
3073 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
3074 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
3075 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
3076 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
3077 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
3078 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
3079 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
3080 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
3081 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
3082 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
3083 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
3084 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
3085 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
3086 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
3087 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
3088 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
3089 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
3090 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
3091 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
3092 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
3093 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
3094 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
3095 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
3096 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
3097 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
3098 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
3099 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
3100 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
3101 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
3102 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
3103 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
3104 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
3105 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
3106 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
3107 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
3108 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
3109 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
3110 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
3111 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
3112 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
3113 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
3114 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
3115 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
3116 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
3117 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
3118 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
3119 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
3122 MachineFunction &MF = *MI->getParent()->getParent();
3123 MI = MF.CloneMachineInstr(MI);
3126 MI->setDesc(get(Opc));
3127 // Fallthrough intended.
3130 return TargetInstrInfo::commuteInstruction(MI, NewMI);
3134 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
3135 unsigned &SrcOpIdx2) const {
3136 switch (MI->getOpcode()) {
3139 case X86::VCMPPDrri:
3140 case X86::VCMPPSrri:
3141 case X86::VCMPPDYrri:
3142 case X86::VCMPPSYrri: {
3143 // Float comparison can be safely commuted for
3144 // Ordered/Unordered/Equal/NotEqual tests
3145 unsigned Imm = MI->getOperand(3).getImm() & 0x7;
3148 case 0x03: // UNORDERED
3149 case 0x04: // NOT EQUAL
3150 case 0x07: // ORDERED
3157 case X86::VFMADDPDr231r:
3158 case X86::VFMADDPSr231r:
3159 case X86::VFMADDSDr231r:
3160 case X86::VFMADDSSr231r:
3161 case X86::VFMSUBPDr231r:
3162 case X86::VFMSUBPSr231r:
3163 case X86::VFMSUBSDr231r:
3164 case X86::VFMSUBSSr231r:
3165 case X86::VFNMADDPDr231r:
3166 case X86::VFNMADDPSr231r:
3167 case X86::VFNMADDSDr231r:
3168 case X86::VFNMADDSSr231r:
3169 case X86::VFNMSUBPDr231r:
3170 case X86::VFNMSUBPSr231r:
3171 case X86::VFNMSUBSDr231r:
3172 case X86::VFNMSUBSSr231r:
3173 case X86::VFMADDPDr231rY:
3174 case X86::VFMADDPSr231rY:
3175 case X86::VFMSUBPDr231rY:
3176 case X86::VFMSUBPSr231rY:
3177 case X86::VFNMADDPDr231rY:
3178 case X86::VFNMADDPSr231rY:
3179 case X86::VFNMSUBPDr231rY:
3180 case X86::VFNMSUBPSr231rY:
3185 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
3189 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
3191 default: return X86::COND_INVALID;
3192 case X86::JE_1: return X86::COND_E;
3193 case X86::JNE_1: return X86::COND_NE;
3194 case X86::JL_1: return X86::COND_L;
3195 case X86::JLE_1: return X86::COND_LE;
3196 case X86::JG_1: return X86::COND_G;
3197 case X86::JGE_1: return X86::COND_GE;
3198 case X86::JB_1: return X86::COND_B;
3199 case X86::JBE_1: return X86::COND_BE;
3200 case X86::JA_1: return X86::COND_A;
3201 case X86::JAE_1: return X86::COND_AE;
3202 case X86::JS_1: return X86::COND_S;
3203 case X86::JNS_1: return X86::COND_NS;
3204 case X86::JP_1: return X86::COND_P;
3205 case X86::JNP_1: return X86::COND_NP;
3206 case X86::JO_1: return X86::COND_O;
3207 case X86::JNO_1: return X86::COND_NO;
3211 /// Return condition code of a SET opcode.
3212 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
3214 default: return X86::COND_INVALID;
3215 case X86::SETAr: case X86::SETAm: return X86::COND_A;
3216 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
3217 case X86::SETBr: case X86::SETBm: return X86::COND_B;
3218 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
3219 case X86::SETEr: case X86::SETEm: return X86::COND_E;
3220 case X86::SETGr: case X86::SETGm: return X86::COND_G;
3221 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
3222 case X86::SETLr: case X86::SETLm: return X86::COND_L;
3223 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
3224 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
3225 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
3226 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
3227 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
3228 case X86::SETOr: case X86::SETOm: return X86::COND_O;
3229 case X86::SETPr: case X86::SETPm: return X86::COND_P;
3230 case X86::SETSr: case X86::SETSm: return X86::COND_S;
3234 /// Return condition code of a CMov opcode.
3235 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
3237 default: return X86::COND_INVALID;
3238 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
3239 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
3241 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
3242 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
3243 return X86::COND_AE;
3244 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
3245 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
3247 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
3248 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
3249 return X86::COND_BE;
3250 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
3251 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
3253 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
3254 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
3256 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
3257 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
3258 return X86::COND_GE;
3259 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
3260 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
3262 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
3263 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
3264 return X86::COND_LE;
3265 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
3266 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
3267 return X86::COND_NE;
3268 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
3269 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
3270 return X86::COND_NO;
3271 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
3272 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
3273 return X86::COND_NP;
3274 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
3275 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
3276 return X86::COND_NS;
3277 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
3278 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
3280 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
3281 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
3283 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
3284 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
3289 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
3291 default: llvm_unreachable("Illegal condition code!");
3292 case X86::COND_E: return X86::JE_1;
3293 case X86::COND_NE: return X86::JNE_1;
3294 case X86::COND_L: return X86::JL_1;
3295 case X86::COND_LE: return X86::JLE_1;
3296 case X86::COND_G: return X86::JG_1;
3297 case X86::COND_GE: return X86::JGE_1;
3298 case X86::COND_B: return X86::JB_1;
3299 case X86::COND_BE: return X86::JBE_1;
3300 case X86::COND_A: return X86::JA_1;
3301 case X86::COND_AE: return X86::JAE_1;
3302 case X86::COND_S: return X86::JS_1;
3303 case X86::COND_NS: return X86::JNS_1;
3304 case X86::COND_P: return X86::JP_1;
3305 case X86::COND_NP: return X86::JNP_1;
3306 case X86::COND_O: return X86::JO_1;
3307 case X86::COND_NO: return X86::JNO_1;
3311 /// Return the inverse of the specified condition,
3312 /// e.g. turning COND_E to COND_NE.
3313 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
3315 default: llvm_unreachable("Illegal condition code!");
3316 case X86::COND_E: return X86::COND_NE;
3317 case X86::COND_NE: return X86::COND_E;
3318 case X86::COND_L: return X86::COND_GE;
3319 case X86::COND_LE: return X86::COND_G;
3320 case X86::COND_G: return X86::COND_LE;
3321 case X86::COND_GE: return X86::COND_L;
3322 case X86::COND_B: return X86::COND_AE;
3323 case X86::COND_BE: return X86::COND_A;
3324 case X86::COND_A: return X86::COND_BE;
3325 case X86::COND_AE: return X86::COND_B;
3326 case X86::COND_S: return X86::COND_NS;
3327 case X86::COND_NS: return X86::COND_S;
3328 case X86::COND_P: return X86::COND_NP;
3329 case X86::COND_NP: return X86::COND_P;
3330 case X86::COND_O: return X86::COND_NO;
3331 case X86::COND_NO: return X86::COND_O;
3335 /// Assuming the flags are set by MI(a,b), return the condition code if we
3336 /// modify the instructions such that flags are set by MI(b,a).
3337 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
3339 default: return X86::COND_INVALID;
3340 case X86::COND_E: return X86::COND_E;
3341 case X86::COND_NE: return X86::COND_NE;
3342 case X86::COND_L: return X86::COND_G;
3343 case X86::COND_LE: return X86::COND_GE;
3344 case X86::COND_G: return X86::COND_L;
3345 case X86::COND_GE: return X86::COND_LE;
3346 case X86::COND_B: return X86::COND_A;
3347 case X86::COND_BE: return X86::COND_AE;
3348 case X86::COND_A: return X86::COND_B;
3349 case X86::COND_AE: return X86::COND_BE;
3353 /// Return a set opcode for the given condition and
3354 /// whether it has memory operand.
3355 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
3356 static const uint16_t Opc[16][2] = {
3357 { X86::SETAr, X86::SETAm },
3358 { X86::SETAEr, X86::SETAEm },
3359 { X86::SETBr, X86::SETBm },
3360 { X86::SETBEr, X86::SETBEm },
3361 { X86::SETEr, X86::SETEm },
3362 { X86::SETGr, X86::SETGm },
3363 { X86::SETGEr, X86::SETGEm },
3364 { X86::SETLr, X86::SETLm },
3365 { X86::SETLEr, X86::SETLEm },
3366 { X86::SETNEr, X86::SETNEm },
3367 { X86::SETNOr, X86::SETNOm },
3368 { X86::SETNPr, X86::SETNPm },
3369 { X86::SETNSr, X86::SETNSm },
3370 { X86::SETOr, X86::SETOm },
3371 { X86::SETPr, X86::SETPm },
3372 { X86::SETSr, X86::SETSm }
3375 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
3376 return Opc[CC][HasMemoryOperand ? 1 : 0];
3379 /// Return a cmov opcode for the given condition,
3380 /// register size in bytes, and operand type.
3381 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
3382 bool HasMemoryOperand) {
3383 static const uint16_t Opc[32][3] = {
3384 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
3385 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
3386 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
3387 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
3388 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
3389 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
3390 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
3391 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
3392 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
3393 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
3394 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
3395 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
3396 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
3397 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
3398 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
3399 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
3400 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
3401 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
3402 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
3403 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
3404 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
3405 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
3406 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
3407 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
3408 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
3409 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
3410 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
3411 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
3412 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
3413 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
3414 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
3415 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
3418 assert(CC < 16 && "Can only handle standard cond codes");
3419 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
3421 default: llvm_unreachable("Illegal register size!");
3422 case 2: return Opc[Idx][0];
3423 case 4: return Opc[Idx][1];
3424 case 8: return Opc[Idx][2];
3428 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
3429 if (!MI->isTerminator()) return false;
3431 // Conditional branch is a special case.
3432 if (MI->isBranch() && !MI->isBarrier())
3434 if (!MI->isPredicable())
3436 return !isPredicated(MI);
3439 bool X86InstrInfo::AnalyzeBranchImpl(
3440 MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
3441 SmallVectorImpl<MachineOperand> &Cond,
3442 SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
3444 // Start from the bottom of the block and work up, examining the
3445 // terminator instructions.
3446 MachineBasicBlock::iterator I = MBB.end();
3447 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
3448 while (I != MBB.begin()) {
3450 if (I->isDebugValue())
3453 // Working from the bottom, when we see a non-terminator instruction, we're
3455 if (!isUnpredicatedTerminator(I))
3458 // A terminator that isn't a branch can't easily be handled by this
3463 // Handle unconditional branches.
3464 if (I->getOpcode() == X86::JMP_1) {
3468 TBB = I->getOperand(0).getMBB();
3472 // If the block has any instructions after a JMP, delete them.
3473 while (std::next(I) != MBB.end())
3474 std::next(I)->eraseFromParent();
3479 // Delete the JMP if it's equivalent to a fall-through.
3480 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
3482 I->eraseFromParent();
3484 UnCondBrIter = MBB.end();
3488 // TBB is used to indicate the unconditional destination.
3489 TBB = I->getOperand(0).getMBB();
3493 // Handle conditional branches.
3494 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
3495 if (BranchCode == X86::COND_INVALID)
3496 return true; // Can't handle indirect branch.
3498 // Working from the bottom, handle the first conditional branch.
3500 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
3501 if (AllowModify && UnCondBrIter != MBB.end() &&
3502 MBB.isLayoutSuccessor(TargetBB)) {
3503 // If we can modify the code and it ends in something like:
3511 // Then we can change this to:
3518 // Which is a bit more efficient.
3519 // We conditionally jump to the fall-through block.
3520 BranchCode = GetOppositeBranchCondition(BranchCode);
3521 unsigned JNCC = GetCondBranchFromCond(BranchCode);
3522 MachineBasicBlock::iterator OldInst = I;
3524 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
3525 .addMBB(UnCondBrIter->getOperand(0).getMBB());
3526 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
3529 OldInst->eraseFromParent();
3530 UnCondBrIter->eraseFromParent();
3532 // Restart the analysis.
3533 UnCondBrIter = MBB.end();
3539 TBB = I->getOperand(0).getMBB();
3540 Cond.push_back(MachineOperand::CreateImm(BranchCode));
3541 CondBranches.push_back(I);
3545 // Handle subsequent conditional branches. Only handle the case where all
3546 // conditional branches branch to the same destination and their condition
3547 // opcodes fit one of the special multi-branch idioms.
3548 assert(Cond.size() == 1);
3551 // Only handle the case where all conditional branches branch to the same
3553 if (TBB != I->getOperand(0).getMBB())
3556 // If the conditions are the same, we can leave them alone.
3557 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
3558 if (OldBranchCode == BranchCode)
3561 // If they differ, see if they fit one of the known patterns. Theoretically,
3562 // we could handle more patterns here, but we shouldn't expect to see them
3563 // if instruction selection has done a reasonable job.
3564 if ((OldBranchCode == X86::COND_NP &&
3565 BranchCode == X86::COND_E) ||
3566 (OldBranchCode == X86::COND_E &&
3567 BranchCode == X86::COND_NP))
3568 BranchCode = X86::COND_NP_OR_E;
3569 else if ((OldBranchCode == X86::COND_P &&
3570 BranchCode == X86::COND_NE) ||
3571 (OldBranchCode == X86::COND_NE &&
3572 BranchCode == X86::COND_P))
3573 BranchCode = X86::COND_NE_OR_P;
3577 // Update the MachineOperand.
3578 Cond[0].setImm(BranchCode);
3579 CondBranches.push_back(I);
3585 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
3586 MachineBasicBlock *&TBB,
3587 MachineBasicBlock *&FBB,
3588 SmallVectorImpl<MachineOperand> &Cond,
3589 bool AllowModify) const {
3590 SmallVector<MachineInstr *, 4> CondBranches;
3591 return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
3594 bool X86InstrInfo::AnalyzeBranchPredicate(MachineBasicBlock &MBB,
3595 MachineBranchPredicate &MBP,
3596 bool AllowModify) const {
3597 using namespace std::placeholders;
3599 SmallVector<MachineOperand, 4> Cond;
3600 SmallVector<MachineInstr *, 4> CondBranches;
3601 if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
3605 if (Cond.size() != 1)
3608 assert(MBP.TrueDest && "expected!");
3611 MBP.FalseDest = MBB.getNextNode();
3613 const TargetRegisterInfo *TRI = &getRegisterInfo();
3615 MachineInstr *ConditionDef = nullptr;
3616 bool SingleUseCondition = true;
3618 for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
3619 if (I->modifiesRegister(X86::EFLAGS, TRI)) {
3624 if (I->readsRegister(X86::EFLAGS, TRI))
3625 SingleUseCondition = false;
3631 if (SingleUseCondition) {
3632 for (auto *Succ : MBB.successors())
3633 if (Succ->isLiveIn(X86::EFLAGS))
3634 SingleUseCondition = false;
3637 MBP.ConditionDef = ConditionDef;
3638 MBP.SingleUseCondition = SingleUseCondition;
3640 // Currently we only recognize the simple pattern:
3645 const unsigned TestOpcode =
3646 Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
3648 if (ConditionDef->getOpcode() == TestOpcode &&
3649 ConditionDef->getNumOperands() == 3 &&
3650 ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
3651 (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
3652 MBP.LHS = ConditionDef->getOperand(0);
3653 MBP.RHS = MachineOperand::CreateImm(0);
3654 MBP.Predicate = Cond[0].getImm() == X86::COND_NE
3655 ? MachineBranchPredicate::PRED_NE
3656 : MachineBranchPredicate::PRED_EQ;
3663 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
3664 MachineBasicBlock::iterator I = MBB.end();
3667 while (I != MBB.begin()) {
3669 if (I->isDebugValue())
3671 if (I->getOpcode() != X86::JMP_1 &&
3672 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
3674 // Remove the branch.
3675 I->eraseFromParent();
3684 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
3685 MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
3686 DebugLoc DL) const {
3687 // Shouldn't be a fall through.
3688 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
3689 assert((Cond.size() == 1 || Cond.size() == 0) &&
3690 "X86 branch conditions have one component!");
3693 // Unconditional branch?
3694 assert(!FBB && "Unconditional branch with multiple successors!");
3695 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
3699 // Conditional branch.
3701 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
3703 case X86::COND_NP_OR_E:
3704 // Synthesize NP_OR_E with two branches.
3705 BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
3707 BuildMI(&MBB, DL, get(X86::JE_1)).addMBB(TBB);
3710 case X86::COND_NE_OR_P:
3711 // Synthesize NE_OR_P with two branches.
3712 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
3714 BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
3718 unsigned Opc = GetCondBranchFromCond(CC);
3719 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
3724 // Two-way Conditional branch. Insert the second branch.
3725 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
3732 canInsertSelect(const MachineBasicBlock &MBB,
3733 ArrayRef<MachineOperand> Cond,
3734 unsigned TrueReg, unsigned FalseReg,
3735 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
3736 // Not all subtargets have cmov instructions.
3737 if (!Subtarget.hasCMov())
3739 if (Cond.size() != 1)
3741 // We cannot do the composite conditions, at least not in SSA form.
3742 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
3745 // Check register classes.
3746 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3747 const TargetRegisterClass *RC =
3748 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
3752 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
3753 if (X86::GR16RegClass.hasSubClassEq(RC) ||
3754 X86::GR32RegClass.hasSubClassEq(RC) ||
3755 X86::GR64RegClass.hasSubClassEq(RC)) {
3756 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
3757 // Bridge. Probably Ivy Bridge as well.
3764 // Can't do vectors.
3768 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3769 MachineBasicBlock::iterator I, DebugLoc DL,
3770 unsigned DstReg, ArrayRef<MachineOperand> Cond,
3771 unsigned TrueReg, unsigned FalseReg) const {
3772 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3773 assert(Cond.size() == 1 && "Invalid Cond array");
3774 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
3775 MRI.getRegClass(DstReg)->getSize(),
3776 false/*HasMemoryOperand*/);
3777 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
3780 /// Test if the given register is a physical h register.
3781 static bool isHReg(unsigned Reg) {
3782 return X86::GR8_ABCD_HRegClass.contains(Reg);
3785 // Try and copy between VR128/VR64 and GR64 registers.
3786 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3787 const X86Subtarget &Subtarget) {
3789 // SrcReg(VR128) -> DestReg(GR64)
3790 // SrcReg(VR64) -> DestReg(GR64)
3791 // SrcReg(GR64) -> DestReg(VR128)
3792 // SrcReg(GR64) -> DestReg(VR64)
3794 bool HasAVX = Subtarget.hasAVX();
3795 bool HasAVX512 = Subtarget.hasAVX512();
3796 if (X86::GR64RegClass.contains(DestReg)) {
3797 if (X86::VR128XRegClass.contains(SrcReg))
3798 // Copy from a VR128 register to a GR64 register.
3799 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
3801 if (X86::VR64RegClass.contains(SrcReg))
3802 // Copy from a VR64 register to a GR64 register.
3803 return X86::MMX_MOVD64from64rr;
3804 } else if (X86::GR64RegClass.contains(SrcReg)) {
3805 // Copy from a GR64 register to a VR128 register.
3806 if (X86::VR128XRegClass.contains(DestReg))
3807 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
3809 // Copy from a GR64 register to a VR64 register.
3810 if (X86::VR64RegClass.contains(DestReg))
3811 return X86::MMX_MOVD64to64rr;
3814 // SrcReg(FR32) -> DestReg(GR32)
3815 // SrcReg(GR32) -> DestReg(FR32)
3817 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3818 // Copy from a FR32 register to a GR32 register.
3819 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3821 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3822 // Copy from a GR32 register to a FR32 register.
3823 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3827 inline static bool MaskRegClassContains(unsigned Reg) {
3828 return X86::VK8RegClass.contains(Reg) ||
3829 X86::VK16RegClass.contains(Reg) ||
3830 X86::VK32RegClass.contains(Reg) ||
3831 X86::VK64RegClass.contains(Reg) ||
3832 X86::VK1RegClass.contains(Reg);
3835 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3836 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3837 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3838 X86::VR512RegClass.contains(DestReg, SrcReg)) {
3839 DestReg = get512BitSuperRegister(DestReg);
3840 SrcReg = get512BitSuperRegister(SrcReg);
3841 return X86::VMOVAPSZrr;
3843 if (MaskRegClassContains(DestReg) &&
3844 MaskRegClassContains(SrcReg))
3845 return X86::KMOVWkk;
3846 if (MaskRegClassContains(DestReg) &&
3847 (X86::GR32RegClass.contains(SrcReg) ||
3848 X86::GR16RegClass.contains(SrcReg) ||
3849 X86::GR8RegClass.contains(SrcReg))) {
3850 SrcReg = getX86SubSuperRegister(SrcReg, MVT::i32);
3851 return X86::KMOVWkr;
3853 if ((X86::GR32RegClass.contains(DestReg) ||
3854 X86::GR16RegClass.contains(DestReg) ||
3855 X86::GR8RegClass.contains(DestReg)) &&
3856 MaskRegClassContains(SrcReg)) {
3857 DestReg = getX86SubSuperRegister(DestReg, MVT::i32);
3858 return X86::KMOVWrk;
3863 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3864 MachineBasicBlock::iterator MI, DebugLoc DL,
3865 unsigned DestReg, unsigned SrcReg,
3866 bool KillSrc) const {
3867 // First deal with the normal symmetric copies.
3868 bool HasAVX = Subtarget.hasAVX();
3869 bool HasAVX512 = Subtarget.hasAVX512();
3871 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3873 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3875 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3877 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3878 // Copying to or from a physical H register on x86-64 requires a NOREX
3879 // move. Otherwise use a normal move.
3880 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3881 Subtarget.is64Bit()) {
3882 Opc = X86::MOV8rr_NOREX;
3883 // Both operands must be encodable without an REX prefix.
3884 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3885 "8-bit H register can not be copied outside GR8_NOREX");
3889 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3890 Opc = X86::MMX_MOVQ64rr;
3892 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3893 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3894 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3895 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3896 Opc = X86::VMOVAPSYrr;
3898 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3901 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3902 .addReg(SrcReg, getKillRegState(KillSrc));
3906 bool FromEFLAGS = SrcReg == X86::EFLAGS;
3907 bool ToEFLAGS = DestReg == X86::EFLAGS;
3908 int Reg = FromEFLAGS ? DestReg : SrcReg;
3909 bool is32 = X86::GR32RegClass.contains(Reg);
3910 bool is64 = X86::GR64RegClass.contains(Reg);
3911 if ((FromEFLAGS || ToEFLAGS) && (is32 || is64)) {
3912 // The flags need to be saved, but saving EFLAGS with PUSHF/POPF is
3913 // inefficient. Instead:
3914 // - Save the overflow flag OF into AL using SETO, and restore it using a
3915 // signed 8-bit addition of AL and INT8_MAX.
3916 // - Save/restore the bottom 8 EFLAGS bits (CF, PF, AF, ZF, SF) to/from AH
3918 // - When RAX/EAX is live and isn't the destination register, make sure it
3919 // isn't clobbered by PUSH/POP'ing it before and after saving/restoring
3921 // This approach is ~2.25x faster than using PUSHF/POPF.
3923 // This is still somewhat inefficient because we don't know which flags are
3924 // actually live inside EFLAGS. Were we able to do a single SETcc instead of
3925 // SETO+LAHF / ADDB+SAHF the code could be 1.02x faster.
3927 // PUSHF/POPF is also potentially incorrect because it affects other flags
3928 // such as TF/IF/DF, which LLVM doesn't model.
3930 // Notice that we have to adjust the stack if we don't want to clobber the
3931 // first frame index. See X86FrameLowering.cpp - clobbersTheStack.
3933 int Mov = is64 ? X86::MOV64rr : X86::MOV32rr;
3934 int Push = is64 ? X86::PUSH64r : X86::PUSH32r;
3935 int Pop = is64 ? X86::POP64r : X86::POP32r;
3936 int AX = is64 ? X86::RAX : X86::EAX;
3938 bool AXDead = (Reg == AX) ||
3939 (MachineBasicBlock::LQR_Dead ==
3940 MBB.computeRegisterLiveness(&getRegisterInfo(), AX, MI));
3943 BuildMI(MBB, MI, DL, get(Push)).addReg(AX, getKillRegState(true));
3945 BuildMI(MBB, MI, DL, get(X86::SETOr), X86::AL);
3946 BuildMI(MBB, MI, DL, get(X86::LAHF));
3947 BuildMI(MBB, MI, DL, get(Mov), Reg).addReg(AX);
3950 BuildMI(MBB, MI, DL, get(Mov), AX).addReg(Reg, getKillRegState(KillSrc));
3951 BuildMI(MBB, MI, DL, get(X86::ADD8ri), X86::AL)
3954 BuildMI(MBB, MI, DL, get(X86::SAHF));
3957 BuildMI(MBB, MI, DL, get(Pop), AX);
3961 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3962 << " to " << RI.getName(DestReg) << '\n');
3963 llvm_unreachable("Cannot emit physreg copy instruction");
3966 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3967 const TargetRegisterClass *RC,
3968 bool isStackAligned,
3969 const X86Subtarget &STI,
3971 if (STI.hasAVX512()) {
3972 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3973 X86::VK16RegClass.hasSubClassEq(RC))
3974 return load ? X86::KMOVWkm : X86::KMOVWmk;
3975 if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3976 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3977 if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3978 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3979 if (X86::VR512RegClass.hasSubClassEq(RC))
3980 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3983 bool HasAVX = STI.hasAVX();
3984 switch (RC->getSize()) {
3986 llvm_unreachable("Unknown spill size");
3988 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3990 // Copying to or from a physical H register on x86-64 requires a NOREX
3991 // move. Otherwise use a normal move.
3992 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3993 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3994 return load ? X86::MOV8rm : X86::MOV8mr;
3996 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3997 return load ? X86::MOV16rm : X86::MOV16mr;
3999 if (X86::GR32RegClass.hasSubClassEq(RC))
4000 return load ? X86::MOV32rm : X86::MOV32mr;
4001 if (X86::FR32RegClass.hasSubClassEq(RC))
4003 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
4004 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
4005 if (X86::RFP32RegClass.hasSubClassEq(RC))
4006 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
4007 llvm_unreachable("Unknown 4-byte regclass");
4009 if (X86::GR64RegClass.hasSubClassEq(RC))
4010 return load ? X86::MOV64rm : X86::MOV64mr;
4011 if (X86::FR64RegClass.hasSubClassEq(RC))
4013 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
4014 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
4015 if (X86::VR64RegClass.hasSubClassEq(RC))
4016 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
4017 if (X86::RFP64RegClass.hasSubClassEq(RC))
4018 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
4019 llvm_unreachable("Unknown 8-byte regclass");
4021 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
4022 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
4024 assert((X86::VR128RegClass.hasSubClassEq(RC) ||
4025 X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
4026 // If stack is realigned we can use aligned stores.
4029 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
4030 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
4033 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
4034 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
4037 assert((X86::VR256RegClass.hasSubClassEq(RC) ||
4038 X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
4039 // If stack is realigned we can use aligned stores.
4041 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
4043 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
4045 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
4047 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
4049 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
4053 bool X86InstrInfo::getMemOpBaseRegImmOfs(MachineInstr *MemOp, unsigned &BaseReg,
4055 const TargetRegisterInfo *TRI) const {
4056 const MCInstrDesc &Desc = MemOp->getDesc();
4057 int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags, MemOp->getOpcode());
4058 if (MemRefBegin < 0)
4061 MemRefBegin += X86II::getOperandBias(Desc);
4063 BaseReg = MemOp->getOperand(MemRefBegin + X86::AddrBaseReg).getReg();
4064 if (MemOp->getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
4067 if (MemOp->getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
4071 const MachineOperand &DispMO = MemOp->getOperand(MemRefBegin + X86::AddrDisp);
4073 // Displacement can be symbolic
4074 if (!DispMO.isImm())
4077 Offset = DispMO.getImm();
4079 return (MemOp->getOperand(MemRefBegin + X86::AddrIndexReg).getReg() ==
4083 static unsigned getStoreRegOpcode(unsigned SrcReg,
4084 const TargetRegisterClass *RC,
4085 bool isStackAligned,
4086 const X86Subtarget &STI) {
4087 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
4091 static unsigned getLoadRegOpcode(unsigned DestReg,
4092 const TargetRegisterClass *RC,
4093 bool isStackAligned,
4094 const X86Subtarget &STI) {
4095 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
4098 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
4099 MachineBasicBlock::iterator MI,
4100 unsigned SrcReg, bool isKill, int FrameIdx,
4101 const TargetRegisterClass *RC,
4102 const TargetRegisterInfo *TRI) const {
4103 const MachineFunction &MF = *MBB.getParent();
4104 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
4105 "Stack slot too small for store");
4106 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4108 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
4109 RI.canRealignStack(MF);
4110 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
4111 DebugLoc DL = MBB.findDebugLoc(MI);
4112 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
4113 .addReg(SrcReg, getKillRegState(isKill));
4116 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
4118 SmallVectorImpl<MachineOperand> &Addr,
4119 const TargetRegisterClass *RC,
4120 MachineInstr::mmo_iterator MMOBegin,
4121 MachineInstr::mmo_iterator MMOEnd,
4122 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4123 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4124 bool isAligned = MMOBegin != MMOEnd &&
4125 (*MMOBegin)->getAlignment() >= Alignment;
4126 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
4128 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
4129 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
4130 MIB.addOperand(Addr[i]);
4131 MIB.addReg(SrcReg, getKillRegState(isKill));
4132 (*MIB).setMemRefs(MMOBegin, MMOEnd);
4133 NewMIs.push_back(MIB);
4137 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
4138 MachineBasicBlock::iterator MI,
4139 unsigned DestReg, int FrameIdx,
4140 const TargetRegisterClass *RC,
4141 const TargetRegisterInfo *TRI) const {
4142 const MachineFunction &MF = *MBB.getParent();
4143 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4145 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
4146 RI.canRealignStack(MF);
4147 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
4148 DebugLoc DL = MBB.findDebugLoc(MI);
4149 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
4152 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
4153 SmallVectorImpl<MachineOperand> &Addr,
4154 const TargetRegisterClass *RC,
4155 MachineInstr::mmo_iterator MMOBegin,
4156 MachineInstr::mmo_iterator MMOEnd,
4157 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4158 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4159 bool isAligned = MMOBegin != MMOEnd &&
4160 (*MMOBegin)->getAlignment() >= Alignment;
4161 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
4163 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
4164 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
4165 MIB.addOperand(Addr[i]);
4166 (*MIB).setMemRefs(MMOBegin, MMOEnd);
4167 NewMIs.push_back(MIB);
4171 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
4172 int &CmpMask, int &CmpValue) const {
4173 switch (MI->getOpcode()) {
4175 case X86::CMP64ri32:
4182 SrcReg = MI->getOperand(0).getReg();
4185 CmpValue = MI->getOperand(1).getImm();
4187 // A SUB can be used to perform comparison.
4192 SrcReg = MI->getOperand(1).getReg();
4201 SrcReg = MI->getOperand(1).getReg();
4202 SrcReg2 = MI->getOperand(2).getReg();
4206 case X86::SUB64ri32:
4213 SrcReg = MI->getOperand(1).getReg();
4216 CmpValue = MI->getOperand(2).getImm();
4222 SrcReg = MI->getOperand(0).getReg();
4223 SrcReg2 = MI->getOperand(1).getReg();
4231 SrcReg = MI->getOperand(0).getReg();
4232 if (MI->getOperand(1).getReg() != SrcReg) return false;
4233 // Compare against zero.
4242 /// Check whether the first instruction, whose only
4243 /// purpose is to update flags, can be made redundant.
4244 /// CMPrr can be made redundant by SUBrr if the operands are the same.
4245 /// This function can be extended later on.
4246 /// SrcReg, SrcRegs: register operands for FlagI.
4247 /// ImmValue: immediate for FlagI if it takes an immediate.
4248 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
4249 unsigned SrcReg2, int ImmValue,
4251 if (((FlagI->getOpcode() == X86::CMP64rr &&
4252 OI->getOpcode() == X86::SUB64rr) ||
4253 (FlagI->getOpcode() == X86::CMP32rr &&
4254 OI->getOpcode() == X86::SUB32rr)||
4255 (FlagI->getOpcode() == X86::CMP16rr &&
4256 OI->getOpcode() == X86::SUB16rr)||
4257 (FlagI->getOpcode() == X86::CMP8rr &&
4258 OI->getOpcode() == X86::SUB8rr)) &&
4259 ((OI->getOperand(1).getReg() == SrcReg &&
4260 OI->getOperand(2).getReg() == SrcReg2) ||
4261 (OI->getOperand(1).getReg() == SrcReg2 &&
4262 OI->getOperand(2).getReg() == SrcReg)))
4265 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
4266 OI->getOpcode() == X86::SUB64ri32) ||
4267 (FlagI->getOpcode() == X86::CMP64ri8 &&
4268 OI->getOpcode() == X86::SUB64ri8) ||
4269 (FlagI->getOpcode() == X86::CMP32ri &&
4270 OI->getOpcode() == X86::SUB32ri) ||
4271 (FlagI->getOpcode() == X86::CMP32ri8 &&
4272 OI->getOpcode() == X86::SUB32ri8) ||
4273 (FlagI->getOpcode() == X86::CMP16ri &&
4274 OI->getOpcode() == X86::SUB16ri) ||
4275 (FlagI->getOpcode() == X86::CMP16ri8 &&
4276 OI->getOpcode() == X86::SUB16ri8) ||
4277 (FlagI->getOpcode() == X86::CMP8ri &&
4278 OI->getOpcode() == X86::SUB8ri)) &&
4279 OI->getOperand(1).getReg() == SrcReg &&
4280 OI->getOperand(2).getImm() == ImmValue)
4285 /// Check whether the definition can be converted
4286 /// to remove a comparison against zero.
4287 inline static bool isDefConvertible(MachineInstr *MI) {
4288 switch (MI->getOpcode()) {
4289 default: return false;
4291 // The shift instructions only modify ZF if their shift count is non-zero.
4292 // N.B.: The processor truncates the shift count depending on the encoding.
4293 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
4294 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
4295 return getTruncatedShiftCount(MI, 2) != 0;
4297 // Some left shift instructions can be turned into LEA instructions but only
4298 // if their flags aren't used. Avoid transforming such instructions.
4299 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
4300 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
4301 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
4305 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
4306 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
4307 return getTruncatedShiftCount(MI, 3) != 0;
4309 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
4310 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
4311 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
4312 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
4313 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
4314 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
4315 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
4316 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
4317 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
4318 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
4319 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
4320 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
4321 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
4322 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
4323 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
4324 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
4325 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
4326 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
4327 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
4328 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
4329 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
4330 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
4331 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
4332 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
4333 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
4334 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
4335 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
4336 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
4337 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
4338 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
4339 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
4340 case X86::ADC32ri: case X86::ADC32ri8:
4341 case X86::ADC32rr: case X86::ADC64ri32:
4342 case X86::ADC64ri8: case X86::ADC64rr:
4343 case X86::SBB32ri: case X86::SBB32ri8:
4344 case X86::SBB32rr: case X86::SBB64ri32:
4345 case X86::SBB64ri8: case X86::SBB64rr:
4346 case X86::ANDN32rr: case X86::ANDN32rm:
4347 case X86::ANDN64rr: case X86::ANDN64rm:
4348 case X86::BEXTR32rr: case X86::BEXTR64rr:
4349 case X86::BEXTR32rm: case X86::BEXTR64rm:
4350 case X86::BLSI32rr: case X86::BLSI32rm:
4351 case X86::BLSI64rr: case X86::BLSI64rm:
4352 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
4353 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
4354 case X86::BLSR32rr: case X86::BLSR32rm:
4355 case X86::BLSR64rr: case X86::BLSR64rm:
4356 case X86::BZHI32rr: case X86::BZHI32rm:
4357 case X86::BZHI64rr: case X86::BZHI64rm:
4358 case X86::LZCNT16rr: case X86::LZCNT16rm:
4359 case X86::LZCNT32rr: case X86::LZCNT32rm:
4360 case X86::LZCNT64rr: case X86::LZCNT64rm:
4361 case X86::POPCNT16rr:case X86::POPCNT16rm:
4362 case X86::POPCNT32rr:case X86::POPCNT32rm:
4363 case X86::POPCNT64rr:case X86::POPCNT64rm:
4364 case X86::TZCNT16rr: case X86::TZCNT16rm:
4365 case X86::TZCNT32rr: case X86::TZCNT32rm:
4366 case X86::TZCNT64rr: case X86::TZCNT64rm:
4371 /// Check whether the use can be converted to remove a comparison against zero.
4372 static X86::CondCode isUseDefConvertible(MachineInstr *MI) {
4373 switch (MI->getOpcode()) {
4374 default: return X86::COND_INVALID;
4375 case X86::LZCNT16rr: case X86::LZCNT16rm:
4376 case X86::LZCNT32rr: case X86::LZCNT32rm:
4377 case X86::LZCNT64rr: case X86::LZCNT64rm:
4379 case X86::POPCNT16rr:case X86::POPCNT16rm:
4380 case X86::POPCNT32rr:case X86::POPCNT32rm:
4381 case X86::POPCNT64rr:case X86::POPCNT64rm:
4383 case X86::TZCNT16rr: case X86::TZCNT16rm:
4384 case X86::TZCNT32rr: case X86::TZCNT32rm:
4385 case X86::TZCNT64rr: case X86::TZCNT64rm:
4390 /// Check if there exists an earlier instruction that
4391 /// operates on the same source operands and sets flags in the same way as
4392 /// Compare; remove Compare if possible.
4394 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
4395 int CmpMask, int CmpValue,
4396 const MachineRegisterInfo *MRI) const {
4397 // Check whether we can replace SUB with CMP.
4398 unsigned NewOpcode = 0;
4399 switch (CmpInstr->getOpcode()) {
4401 case X86::SUB64ri32:
4416 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
4418 // There is no use of the destination register, we can replace SUB with CMP.
4419 switch (CmpInstr->getOpcode()) {
4420 default: llvm_unreachable("Unreachable!");
4421 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
4422 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
4423 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
4424 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
4425 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
4426 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
4427 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
4428 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
4429 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
4430 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
4431 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
4432 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
4433 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
4434 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
4435 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
4437 CmpInstr->setDesc(get(NewOpcode));
4438 CmpInstr->RemoveOperand(0);
4439 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
4440 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
4441 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
4446 // Get the unique definition of SrcReg.
4447 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
4448 if (!MI) return false;
4450 // CmpInstr is the first instruction of the BB.
4451 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
4453 // If we are comparing against zero, check whether we can use MI to update
4454 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
4455 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
4456 if (IsCmpZero && MI->getParent() != CmpInstr->getParent())
4459 // If we have a use of the source register between the def and our compare
4460 // instruction we can eliminate the compare iff the use sets EFLAGS in the
4462 bool ShouldUpdateCC = false;
4463 X86::CondCode NewCC = X86::COND_INVALID;
4464 if (IsCmpZero && !isDefConvertible(MI)) {
4465 // Scan forward from the use until we hit the use we're looking for or the
4466 // compare instruction.
4467 for (MachineBasicBlock::iterator J = MI;; ++J) {
4468 // Do we have a convertible instruction?
4469 NewCC = isUseDefConvertible(J);
4470 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
4471 J->getOperand(1).getReg() == SrcReg) {
4472 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
4473 ShouldUpdateCC = true; // Update CC later on.
4474 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
4475 // with the new def.
4485 // We are searching for an earlier instruction that can make CmpInstr
4486 // redundant and that instruction will be saved in Sub.
4487 MachineInstr *Sub = nullptr;
4488 const TargetRegisterInfo *TRI = &getRegisterInfo();
4490 // We iterate backward, starting from the instruction before CmpInstr and
4491 // stop when reaching the definition of a source register or done with the BB.
4492 // RI points to the instruction before CmpInstr.
4493 // If the definition is in this basic block, RE points to the definition;
4494 // otherwise, RE is the rend of the basic block.
4495 MachineBasicBlock::reverse_iterator
4496 RI = MachineBasicBlock::reverse_iterator(I),
4497 RE = CmpInstr->getParent() == MI->getParent() ?
4498 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
4499 CmpInstr->getParent()->rend();
4500 MachineInstr *Movr0Inst = nullptr;
4501 for (; RI != RE; ++RI) {
4502 MachineInstr *Instr = &*RI;
4503 // Check whether CmpInstr can be made redundant by the current instruction.
4505 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
4510 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
4511 Instr->readsRegister(X86::EFLAGS, TRI)) {
4512 // This instruction modifies or uses EFLAGS.
4514 // MOV32r0 etc. are implemented with xor which clobbers condition code.
4515 // They are safe to move up, if the definition to EFLAGS is dead and
4516 // earlier instructions do not read or write EFLAGS.
4517 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
4518 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
4523 // We can't remove CmpInstr.
4528 // Return false if no candidates exist.
4529 if (!IsCmpZero && !Sub)
4532 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
4533 Sub->getOperand(2).getReg() == SrcReg);
4535 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
4536 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
4537 // If we are done with the basic block, we need to check whether EFLAGS is
4539 bool IsSafe = false;
4540 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
4541 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
4542 for (++I; I != E; ++I) {
4543 const MachineInstr &Instr = *I;
4544 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
4545 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
4546 // We should check the usage if this instruction uses and updates EFLAGS.
4547 if (!UseEFLAGS && ModifyEFLAGS) {
4548 // It is safe to remove CmpInstr if EFLAGS is updated again.
4552 if (!UseEFLAGS && !ModifyEFLAGS)
4555 // EFLAGS is used by this instruction.
4556 X86::CondCode OldCC = X86::COND_INVALID;
4557 bool OpcIsSET = false;
4558 if (IsCmpZero || IsSwapped) {
4559 // We decode the condition code from opcode.
4560 if (Instr.isBranch())
4561 OldCC = getCondFromBranchOpc(Instr.getOpcode());
4563 OldCC = getCondFromSETOpc(Instr.getOpcode());
4564 if (OldCC != X86::COND_INVALID)
4567 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
4569 if (OldCC == X86::COND_INVALID) return false;
4574 case X86::COND_A: case X86::COND_AE:
4575 case X86::COND_B: case X86::COND_BE:
4576 case X86::COND_G: case X86::COND_GE:
4577 case X86::COND_L: case X86::COND_LE:
4578 case X86::COND_O: case X86::COND_NO:
4579 // CF and OF are used, we can't perform this optimization.
4583 // If we're updating the condition code check if we have to reverse the
4592 NewCC = GetOppositeBranchCondition(NewCC);
4595 } else if (IsSwapped) {
4596 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
4597 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
4598 // We swap the condition code and synthesize the new opcode.
4599 NewCC = getSwappedCondition(OldCC);
4600 if (NewCC == X86::COND_INVALID) return false;
4603 if ((ShouldUpdateCC || IsSwapped) && NewCC != OldCC) {
4604 // Synthesize the new opcode.
4605 bool HasMemoryOperand = Instr.hasOneMemOperand();
4607 if (Instr.isBranch())
4608 NewOpc = GetCondBranchFromCond(NewCC);
4610 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
4612 unsigned DstReg = Instr.getOperand(0).getReg();
4613 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
4617 // Push the MachineInstr to OpsToUpdate.
4618 // If it is safe to remove CmpInstr, the condition code of these
4619 // instructions will be modified.
4620 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
4622 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
4623 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
4629 // If EFLAGS is not killed nor re-defined, we should check whether it is
4630 // live-out. If it is live-out, do not optimize.
4631 if ((IsCmpZero || IsSwapped) && !IsSafe) {
4632 MachineBasicBlock *MBB = CmpInstr->getParent();
4633 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
4634 SE = MBB->succ_end(); SI != SE; ++SI)
4635 if ((*SI)->isLiveIn(X86::EFLAGS))
4639 // The instruction to be updated is either Sub or MI.
4640 Sub = IsCmpZero ? MI : Sub;
4641 // Move Movr0Inst to the appropriate place before Sub.
4643 // Look backwards until we find a def that doesn't use the current EFLAGS.
4645 MachineBasicBlock::reverse_iterator
4646 InsertI = MachineBasicBlock::reverse_iterator(++Def),
4647 InsertE = Sub->getParent()->rend();
4648 for (; InsertI != InsertE; ++InsertI) {
4649 MachineInstr *Instr = &*InsertI;
4650 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
4651 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
4652 Sub->getParent()->remove(Movr0Inst);
4653 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
4658 if (InsertI == InsertE)
4662 // Make sure Sub instruction defines EFLAGS and mark the def live.
4663 unsigned i = 0, e = Sub->getNumOperands();
4664 for (; i != e; ++i) {
4665 MachineOperand &MO = Sub->getOperand(i);
4666 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
4667 MO.setIsDead(false);
4671 assert(i != e && "Unable to locate a def EFLAGS operand");
4673 CmpInstr->eraseFromParent();
4675 // Modify the condition code of instructions in OpsToUpdate.
4676 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
4677 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
4681 /// Try to remove the load by folding it to a register
4682 /// operand at the use. We fold the load instructions if load defines a virtual
4683 /// register, the virtual register is used once in the same BB, and the
4684 /// instructions in-between do not load or store, and have no side effects.
4685 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr *MI,
4686 const MachineRegisterInfo *MRI,
4687 unsigned &FoldAsLoadDefReg,
4688 MachineInstr *&DefMI) const {
4689 if (FoldAsLoadDefReg == 0)
4691 // To be conservative, if there exists another load, clear the load candidate.
4692 if (MI->mayLoad()) {
4693 FoldAsLoadDefReg = 0;
4697 // Check whether we can move DefMI here.
4698 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
4700 bool SawStore = false;
4701 if (!DefMI->isSafeToMove(nullptr, SawStore))
4704 // Collect information about virtual register operands of MI.
4705 unsigned SrcOperandId = 0;
4706 bool FoundSrcOperand = false;
4707 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
4708 MachineOperand &MO = MI->getOperand(i);
4711 unsigned Reg = MO.getReg();
4712 if (Reg != FoldAsLoadDefReg)
4714 // Do not fold if we have a subreg use or a def or multiple uses.
4715 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
4719 FoundSrcOperand = true;
4721 if (!FoundSrcOperand)
4724 // Check whether we can fold the def into SrcOperandId.
4725 MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandId, DefMI);
4727 FoldAsLoadDefReg = 0;
4734 /// Expand a single-def pseudo instruction to a two-addr
4735 /// instruction with two undef reads of the register being defined.
4736 /// This is used for mapping:
4739 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
4741 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
4742 const MCInstrDesc &Desc) {
4743 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
4744 unsigned Reg = MIB->getOperand(0).getReg();
4747 // MachineInstr::addOperand() will insert explicit operands before any
4748 // implicit operands.
4749 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4750 // But we don't trust that.
4751 assert(MIB->getOperand(1).getReg() == Reg &&
4752 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
4756 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
4757 // code sequence is needed for other targets.
4758 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
4759 const TargetInstrInfo &TII) {
4760 MachineBasicBlock &MBB = *MIB->getParent();
4761 DebugLoc DL = MIB->getDebugLoc();
4762 unsigned Reg = MIB->getOperand(0).getReg();
4763 const GlobalValue *GV =
4764 cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
4765 unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
4766 MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand(
4767 MachinePointerInfo::getGOT(*MBB.getParent()), Flag, 8, 8);
4768 MachineBasicBlock::iterator I = MIB.getInstr();
4770 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4771 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4772 .addMemOperand(MMO);
4773 MIB->setDebugLoc(DL);
4774 MIB->setDesc(TII.get(X86::MOV64rm));
4775 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4778 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
4779 bool HasAVX = Subtarget.hasAVX();
4780 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
4781 switch (MI->getOpcode()) {
4783 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4785 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4786 case X86::SETB_C16r:
4787 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4788 case X86::SETB_C32r:
4789 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4790 case X86::SETB_C64r:
4791 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4795 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4797 assert(HasAVX && "AVX not supported");
4798 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
4799 case X86::AVX512_512_SET0:
4800 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4801 case X86::V_SETALLONES:
4802 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4803 case X86::AVX2_SETALLONES:
4804 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4805 case X86::TEST8ri_NOREX:
4806 MI->setDesc(get(X86::TEST8ri));
4809 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
4811 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
4812 case TargetOpcode::LOAD_STACK_GUARD:
4813 expandLoadStackGuard(MIB, *this);
4819 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs) {
4820 unsigned NumAddrOps = MOs.size();
4821 for (unsigned i = 0; i != NumAddrOps; ++i)
4822 MIB.addOperand(MOs[i]);
4823 if (NumAddrOps < 4) // FrameIndex only
4827 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4828 ArrayRef<MachineOperand> MOs,
4829 MachineBasicBlock::iterator InsertPt,
4831 const TargetInstrInfo &TII) {
4832 // Create the base instruction with the memory operand as the first part.
4833 // Omit the implicit operands, something BuildMI can't do.
4834 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4835 MI->getDebugLoc(), true);
4836 MachineInstrBuilder MIB(MF, NewMI);
4837 addOperands(MIB, MOs);
4839 // Loop over the rest of the ri operands, converting them over.
4840 unsigned NumOps = MI->getDesc().getNumOperands()-2;
4841 for (unsigned i = 0; i != NumOps; ++i) {
4842 MachineOperand &MO = MI->getOperand(i+2);
4845 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
4846 MachineOperand &MO = MI->getOperand(i);
4850 MachineBasicBlock *MBB = InsertPt->getParent();
4851 MBB->insert(InsertPt, NewMI);
4856 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
4857 unsigned OpNo, ArrayRef<MachineOperand> MOs,
4858 MachineBasicBlock::iterator InsertPt,
4859 MachineInstr *MI, const TargetInstrInfo &TII) {
4860 // Omit the implicit operands, something BuildMI can't do.
4861 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4862 MI->getDebugLoc(), true);
4863 MachineInstrBuilder MIB(MF, NewMI);
4865 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4866 MachineOperand &MO = MI->getOperand(i);
4868 assert(MO.isReg() && "Expected to fold into reg operand!");
4869 addOperands(MIB, MOs);
4875 MachineBasicBlock *MBB = InsertPt->getParent();
4876 MBB->insert(InsertPt, NewMI);
4881 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4882 ArrayRef<MachineOperand> MOs,
4883 MachineBasicBlock::iterator InsertPt,
4885 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
4886 MI->getDebugLoc(), TII.get(Opcode));
4887 addOperands(MIB, MOs);
4888 return MIB.addImm(0);
4891 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
4892 MachineFunction &MF, MachineInstr *MI, unsigned OpNum,
4893 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4894 unsigned Size, unsigned Align, bool AllowCommute) const {
4895 const DenseMap<unsigned,
4896 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4897 bool isCallRegIndirect = Subtarget.callRegIndirect();
4898 bool isTwoAddrFold = false;
4900 // For CPUs that favor the register form of a call or push,
4901 // do not fold loads into calls or pushes, unless optimizing for size
4903 if (isCallRegIndirect && !MF.getFunction()->optForMinSize() &&
4904 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r ||
4905 MI->getOpcode() == X86::PUSH16r || MI->getOpcode() == X86::PUSH32r ||
4906 MI->getOpcode() == X86::PUSH64r))
4909 unsigned NumOps = MI->getDesc().getNumOperands();
4910 bool isTwoAddr = NumOps > 1 &&
4911 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4913 // FIXME: AsmPrinter doesn't know how to handle
4914 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4915 if (MI->getOpcode() == X86::ADD32ri &&
4916 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4919 MachineInstr *NewMI = nullptr;
4920 // Folding a memory location into the two-address part of a two-address
4921 // instruction is different than folding it other places. It requires
4922 // replacing the *two* registers with the memory location.
4923 if (isTwoAddr && NumOps >= 2 && OpNum < 2 &&
4924 MI->getOperand(0).isReg() &&
4925 MI->getOperand(1).isReg() &&
4926 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
4927 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4928 isTwoAddrFold = true;
4929 } else if (OpNum == 0) {
4930 if (MI->getOpcode() == X86::MOV32r0) {
4931 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
4936 OpcodeTablePtr = &RegOp2MemOpTable0;
4937 } else if (OpNum == 1) {
4938 OpcodeTablePtr = &RegOp2MemOpTable1;
4939 } else if (OpNum == 2) {
4940 OpcodeTablePtr = &RegOp2MemOpTable2;
4941 } else if (OpNum == 3) {
4942 OpcodeTablePtr = &RegOp2MemOpTable3;
4943 } else if (OpNum == 4) {
4944 OpcodeTablePtr = &RegOp2MemOpTable4;
4947 // If table selected...
4948 if (OpcodeTablePtr) {
4949 // Find the Opcode to fuse
4950 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4951 OpcodeTablePtr->find(MI->getOpcode());
4952 if (I != OpcodeTablePtr->end()) {
4953 unsigned Opcode = I->second.first;
4954 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4955 if (Align < MinAlign)
4957 bool NarrowToMOV32rm = false;
4959 unsigned RCSize = getRegClass(MI->getDesc(), OpNum, &RI, MF)->getSize();
4960 if (Size < RCSize) {
4961 // Check if it's safe to fold the load. If the size of the object is
4962 // narrower than the load width, then it's not.
4963 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4965 // If this is a 64-bit load, but the spill slot is 32, then we can do
4966 // a 32-bit load which is implicitly zero-extended. This likely is
4967 // due to live interval analysis remat'ing a load from stack slot.
4968 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4970 Opcode = X86::MOV32rm;
4971 NarrowToMOV32rm = true;
4976 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
4978 NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
4980 if (NarrowToMOV32rm) {
4981 // If this is the special case where we use a MOV32rm to load a 32-bit
4982 // value and zero-extend the top bits. Change the destination register
4984 unsigned DstReg = NewMI->getOperand(0).getReg();
4985 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4986 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4988 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4994 // If the instruction and target operand are commutable, commute the
4995 // instruction and try again.
4997 unsigned OriginalOpIdx = OpNum, CommuteOpIdx1, CommuteOpIdx2;
4998 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4999 bool HasDef = MI->getDesc().getNumDefs();
5000 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
5001 unsigned Reg1 = MI->getOperand(CommuteOpIdx1).getReg();
5002 unsigned Reg2 = MI->getOperand(CommuteOpIdx2).getReg();
5004 0 == MI->getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
5006 0 == MI->getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
5008 // If either of the commutable operands are tied to the destination
5009 // then we can not commute + fold.
5010 if ((HasDef && Reg0 == Reg1 && Tied0) ||
5011 (HasDef && Reg0 == Reg2 && Tied1))
5014 if ((CommuteOpIdx1 == OriginalOpIdx) ||
5015 (CommuteOpIdx2 == OriginalOpIdx)) {
5016 MachineInstr *CommutedMI = commuteInstruction(MI, false);
5018 // Unable to commute.
5021 if (CommutedMI != MI) {
5022 // New instruction. We can't fold from this.
5023 CommutedMI->eraseFromParent();
5027 // Attempt to fold with the commuted version of the instruction.
5028 unsigned CommuteOp =
5029 (CommuteOpIdx1 == OriginalOpIdx ? CommuteOpIdx2 : CommuteOpIdx1);
5031 foldMemoryOperandImpl(MF, MI, CommuteOp, MOs, InsertPt, Size, Align,
5032 /*AllowCommute=*/false);
5036 // Folding failed again - undo the commute before returning.
5037 MachineInstr *UncommutedMI = commuteInstruction(MI, false);
5038 if (!UncommutedMI) {
5039 // Unable to commute.
5042 if (UncommutedMI != MI) {
5043 // New instruction. It doesn't need to be kept.
5044 UncommutedMI->eraseFromParent();
5048 // Return here to prevent duplicate fuse failure report.
5055 if (PrintFailedFusing && !MI->isCopy())
5056 dbgs() << "We failed to fuse operand " << OpNum << " in " << *MI;
5060 /// Return true for all instructions that only update
5061 /// the first 32 or 64-bits of the destination register and leave the rest
5062 /// unmodified. This can be used to avoid folding loads if the instructions
5063 /// only update part of the destination register, and the non-updated part is
5064 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
5065 /// instructions breaks the partial register dependency and it can improve
5066 /// performance. e.g.:
5068 /// movss (%rdi), %xmm0
5069 /// cvtss2sd %xmm0, %xmm0
5072 /// cvtss2sd (%rdi), %xmm0
5074 /// FIXME: This should be turned into a TSFlags.
5076 static bool hasPartialRegUpdate(unsigned Opcode) {
5078 case X86::CVTSI2SSrr:
5079 case X86::CVTSI2SSrm:
5080 case X86::CVTSI2SS64rr:
5081 case X86::CVTSI2SS64rm:
5082 case X86::CVTSI2SDrr:
5083 case X86::CVTSI2SDrm:
5084 case X86::CVTSI2SD64rr:
5085 case X86::CVTSI2SD64rm:
5086 case X86::CVTSD2SSrr:
5087 case X86::CVTSD2SSrm:
5088 case X86::Int_CVTSD2SSrr:
5089 case X86::Int_CVTSD2SSrm:
5090 case X86::CVTSS2SDrr:
5091 case X86::CVTSS2SDrm:
5092 case X86::Int_CVTSS2SDrr:
5093 case X86::Int_CVTSS2SDrm:
5096 case X86::RCPSSr_Int:
5097 case X86::RCPSSm_Int:
5100 case X86::ROUNDSDr_Int:
5103 case X86::ROUNDSSr_Int:
5106 case X86::RSQRTSSr_Int:
5107 case X86::RSQRTSSm_Int:
5110 case X86::SQRTSSr_Int:
5111 case X86::SQRTSSm_Int:
5114 case X86::SQRTSDr_Int:
5115 case X86::SQRTSDm_Int:
5122 /// Inform the ExeDepsFix pass how many idle
5123 /// instructions we would like before a partial register update.
5124 unsigned X86InstrInfo::
5125 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
5126 const TargetRegisterInfo *TRI) const {
5127 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
5130 // If MI is marked as reading Reg, the partial register update is wanted.
5131 const MachineOperand &MO = MI->getOperand(0);
5132 unsigned Reg = MO.getReg();
5133 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
5134 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
5137 if (MI->readsRegister(Reg, TRI))
5141 // If any of the preceding 16 instructions are reading Reg, insert a
5142 // dependency breaking instruction. The magic number is based on a few
5143 // Nehalem experiments.
5147 // Return true for any instruction the copies the high bits of the first source
5148 // operand into the unused high bits of the destination operand.
5149 static bool hasUndefRegUpdate(unsigned Opcode) {
5151 case X86::VCVTSI2SSrr:
5152 case X86::VCVTSI2SSrm:
5153 case X86::Int_VCVTSI2SSrr:
5154 case X86::Int_VCVTSI2SSrm:
5155 case X86::VCVTSI2SS64rr:
5156 case X86::VCVTSI2SS64rm:
5157 case X86::Int_VCVTSI2SS64rr:
5158 case X86::Int_VCVTSI2SS64rm:
5159 case X86::VCVTSI2SDrr:
5160 case X86::VCVTSI2SDrm:
5161 case X86::Int_VCVTSI2SDrr:
5162 case X86::Int_VCVTSI2SDrm:
5163 case X86::VCVTSI2SD64rr:
5164 case X86::VCVTSI2SD64rm:
5165 case X86::Int_VCVTSI2SD64rr:
5166 case X86::Int_VCVTSI2SD64rm:
5167 case X86::VCVTSD2SSrr:
5168 case X86::VCVTSD2SSrm:
5169 case X86::Int_VCVTSD2SSrr:
5170 case X86::Int_VCVTSD2SSrm:
5171 case X86::VCVTSS2SDrr:
5172 case X86::VCVTSS2SDrm:
5173 case X86::Int_VCVTSS2SDrr:
5174 case X86::Int_VCVTSS2SDrm:
5177 case X86::VRCPSSm_Int:
5178 case X86::VROUNDSDr:
5179 case X86::VROUNDSDm:
5180 case X86::VROUNDSDr_Int:
5181 case X86::VROUNDSSr:
5182 case X86::VROUNDSSm:
5183 case X86::VROUNDSSr_Int:
5184 case X86::VRSQRTSSr:
5185 case X86::VRSQRTSSm:
5186 case X86::VRSQRTSSm_Int:
5189 case X86::VSQRTSSm_Int:
5192 case X86::VSQRTSDm_Int:
5194 case X86::VCVTSD2SSZrr:
5195 case X86::VCVTSD2SSZrm:
5196 case X86::VCVTSS2SDZrr:
5197 case X86::VCVTSS2SDZrm:
5204 /// Inform the ExeDepsFix pass how many idle instructions we would like before
5205 /// certain undef register reads.
5207 /// This catches the VCVTSI2SD family of instructions:
5209 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
5211 /// We should to be careful *not* to catch VXOR idioms which are presumably
5212 /// handled specially in the pipeline:
5214 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
5216 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
5217 /// high bits that are passed-through are not live.
5218 unsigned X86InstrInfo::
5219 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
5220 const TargetRegisterInfo *TRI) const {
5221 if (!hasUndefRegUpdate(MI->getOpcode()))
5224 // Set the OpNum parameter to the first source operand.
5227 const MachineOperand &MO = MI->getOperand(OpNum);
5228 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
5229 // Use the same magic number as getPartialRegUpdateClearance.
5236 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
5237 const TargetRegisterInfo *TRI) const {
5238 unsigned Reg = MI->getOperand(OpNum).getReg();
5239 // If MI kills this register, the false dependence is already broken.
5240 if (MI->killsRegister(Reg, TRI))
5242 if (X86::VR128RegClass.contains(Reg)) {
5243 // These instructions are all floating point domain, so xorps is the best
5245 bool HasAVX = Subtarget.hasAVX();
5246 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
5247 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
5248 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
5249 } else if (X86::VR256RegClass.contains(Reg)) {
5250 // Use vxorps to clear the full ymm register.
5251 // It wants to read and write the xmm sub-register.
5252 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
5253 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
5254 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
5255 .addReg(Reg, RegState::ImplicitDefine);
5258 MI->addRegisterKilled(Reg, TRI, true);
5261 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5262 MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops,
5263 MachineBasicBlock::iterator InsertPt, int FrameIndex) const {
5264 // Check switch flag
5265 if (NoFusing) return nullptr;
5267 // Unless optimizing for size, don't fold to avoid partial
5268 // register update stalls
5269 if (!MF.getFunction()->optForSize() && hasPartialRegUpdate(MI->getOpcode()))
5272 const MachineFrameInfo *MFI = MF.getFrameInfo();
5273 unsigned Size = MFI->getObjectSize(FrameIndex);
5274 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
5275 // If the function stack isn't realigned we don't want to fold instructions
5276 // that need increased alignment.
5277 if (!RI.needsStackRealignment(MF))
5279 std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
5280 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5281 unsigned NewOpc = 0;
5282 unsigned RCSize = 0;
5283 switch (MI->getOpcode()) {
5284 default: return nullptr;
5285 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
5286 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
5287 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
5288 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
5290 // Check if it's safe to fold the load. If the size of the object is
5291 // narrower than the load width, then it's not.
5294 // Change to CMPXXri r, 0 first.
5295 MI->setDesc(get(NewOpc));
5296 MI->getOperand(1).ChangeToImmediate(0);
5297 } else if (Ops.size() != 1)
5300 return foldMemoryOperandImpl(MF, MI, Ops[0],
5301 MachineOperand::CreateFI(FrameIndex), InsertPt,
5302 Size, Alignment, /*AllowCommute=*/true);
5305 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
5306 /// because the latter uses contents that wouldn't be defined in the folded
5307 /// version. For instance, this transformation isn't legal:
5308 /// movss (%rdi), %xmm0
5309 /// addps %xmm0, %xmm0
5311 /// addps (%rdi), %xmm0
5313 /// But this one is:
5314 /// movss (%rdi), %xmm0
5315 /// addss %xmm0, %xmm0
5317 /// addss (%rdi), %xmm0
5319 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
5320 const MachineInstr &UserMI,
5321 const MachineFunction &MF) {
5322 unsigned Opc = LoadMI.getOpcode();
5323 unsigned UserOpc = UserMI.getOpcode();
5325 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg())->getSize();
5327 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm) && RegSize > 4) {
5328 // These instructions only load 32 bits, we can't fold them if the
5329 // destination register is wider than 32 bits (4 bytes), and its user
5330 // instruction isn't scalar (SS).
5332 case X86::ADDSSrr_Int: case X86::VADDSSrr_Int:
5333 case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int:
5334 case X86::MULSSrr_Int: case X86::VMULSSrr_Int:
5335 case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int:
5342 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm) && RegSize > 8) {
5343 // These instructions only load 64 bits, we can't fold them if the
5344 // destination register is wider than 64 bits (8 bytes), and its user
5345 // instruction isn't scalar (SD).
5347 case X86::ADDSDrr_Int: case X86::VADDSDrr_Int:
5348 case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int:
5349 case X86::MULSDrr_Int: case X86::VMULSDrr_Int:
5350 case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int:
5360 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5361 MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops,
5362 MachineBasicBlock::iterator InsertPt, MachineInstr *LoadMI) const {
5363 // If loading from a FrameIndex, fold directly from the FrameIndex.
5364 unsigned NumOps = LoadMI->getDesc().getNumOperands();
5366 if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
5367 if (isNonFoldablePartialRegisterLoad(*LoadMI, *MI, MF))
5369 return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex);
5372 // Check switch flag
5373 if (NoFusing) return nullptr;
5375 // Avoid partial register update stalls unless optimizing for size.
5376 if (!MF.getFunction()->optForSize() && hasPartialRegUpdate(MI->getOpcode()))
5379 // Determine the alignment of the load.
5380 unsigned Alignment = 0;
5381 if (LoadMI->hasOneMemOperand())
5382 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
5384 switch (LoadMI->getOpcode()) {
5385 case X86::AVX2_SETALLONES:
5390 case X86::V_SETALLONES:
5402 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5403 unsigned NewOpc = 0;
5404 switch (MI->getOpcode()) {
5405 default: return nullptr;
5406 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
5407 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
5408 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
5409 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
5411 // Change to CMPXXri r, 0 first.
5412 MI->setDesc(get(NewOpc));
5413 MI->getOperand(1).ChangeToImmediate(0);
5414 } else if (Ops.size() != 1)
5417 // Make sure the subregisters match.
5418 // Otherwise we risk changing the size of the load.
5419 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
5422 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
5423 switch (LoadMI->getOpcode()) {
5425 case X86::V_SETALLONES:
5426 case X86::AVX2_SETALLONES:
5429 case X86::FsFLD0SS: {
5430 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
5431 // Create a constant-pool entry and operands to load from it.
5433 // Medium and large mode can't fold loads this way.
5434 if (MF.getTarget().getCodeModel() != CodeModel::Small &&
5435 MF.getTarget().getCodeModel() != CodeModel::Kernel)
5438 // x86-32 PIC requires a PIC base register for constant pools.
5439 unsigned PICBase = 0;
5440 if (MF.getTarget().getRelocationModel() == Reloc::PIC_) {
5441 if (Subtarget.is64Bit())
5444 // FIXME: PICBase = getGlobalBaseReg(&MF);
5445 // This doesn't work for several reasons.
5446 // 1. GlobalBaseReg may have been spilled.
5447 // 2. It may not be live at MI.
5451 // Create a constant-pool entry.
5452 MachineConstantPool &MCP = *MF.getConstantPool();
5454 unsigned Opc = LoadMI->getOpcode();
5455 if (Opc == X86::FsFLD0SS)
5456 Ty = Type::getFloatTy(MF.getFunction()->getContext());
5457 else if (Opc == X86::FsFLD0SD)
5458 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
5459 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
5460 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
5462 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
5464 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
5465 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
5466 Constant::getNullValue(Ty);
5467 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
5469 // Create operands to load from the constant pool entry.
5470 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
5471 MOs.push_back(MachineOperand::CreateImm(1));
5472 MOs.push_back(MachineOperand::CreateReg(0, false));
5473 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
5474 MOs.push_back(MachineOperand::CreateReg(0, false));
5478 if (isNonFoldablePartialRegisterLoad(*LoadMI, *MI, MF))
5481 // Folding a normal load. Just copy the load's address operands.
5482 MOs.append(LoadMI->operands_begin() + NumOps - X86::AddrNumOperands,
5483 LoadMI->operands_begin() + NumOps);
5487 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
5488 /*Size=*/0, Alignment, /*AllowCommute=*/true);
5491 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
5492 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
5493 SmallVectorImpl<MachineInstr*> &NewMIs) const {
5494 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5495 MemOp2RegOpTable.find(MI->getOpcode());
5496 if (I == MemOp2RegOpTable.end())
5498 unsigned Opc = I->second.first;
5499 unsigned Index = I->second.second & TB_INDEX_MASK;
5500 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5501 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5502 if (UnfoldLoad && !FoldedLoad)
5504 UnfoldLoad &= FoldedLoad;
5505 if (UnfoldStore && !FoldedStore)
5507 UnfoldStore &= FoldedStore;
5509 const MCInstrDesc &MCID = get(Opc);
5510 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5511 if (!MI->hasOneMemOperand() &&
5512 RC == &X86::VR128RegClass &&
5513 !Subtarget.isUnalignedMemAccessFast())
5514 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
5515 // conservatively assume the address is unaligned. That's bad for
5518 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
5519 SmallVector<MachineOperand,2> BeforeOps;
5520 SmallVector<MachineOperand,2> AfterOps;
5521 SmallVector<MachineOperand,4> ImpOps;
5522 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
5523 MachineOperand &Op = MI->getOperand(i);
5524 if (i >= Index && i < Index + X86::AddrNumOperands)
5525 AddrOps.push_back(Op);
5526 else if (Op.isReg() && Op.isImplicit())
5527 ImpOps.push_back(Op);
5529 BeforeOps.push_back(Op);
5531 AfterOps.push_back(Op);
5534 // Emit the load instruction.
5536 std::pair<MachineInstr::mmo_iterator,
5537 MachineInstr::mmo_iterator> MMOs =
5538 MF.extractLoadMemRefs(MI->memoperands_begin(),
5539 MI->memoperands_end());
5540 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
5542 // Address operands cannot be marked isKill.
5543 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
5544 MachineOperand &MO = NewMIs[0]->getOperand(i);
5546 MO.setIsKill(false);
5551 // Emit the data processing instruction.
5552 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
5553 MachineInstrBuilder MIB(MF, DataMI);
5556 MIB.addReg(Reg, RegState::Define);
5557 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
5558 MIB.addOperand(BeforeOps[i]);
5561 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
5562 MIB.addOperand(AfterOps[i]);
5563 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
5564 MachineOperand &MO = ImpOps[i];
5565 MIB.addReg(MO.getReg(),
5566 getDefRegState(MO.isDef()) |
5567 RegState::Implicit |
5568 getKillRegState(MO.isKill()) |
5569 getDeadRegState(MO.isDead()) |
5570 getUndefRegState(MO.isUndef()));
5572 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5573 switch (DataMI->getOpcode()) {
5575 case X86::CMP64ri32:
5582 MachineOperand &MO0 = DataMI->getOperand(0);
5583 MachineOperand &MO1 = DataMI->getOperand(1);
5584 if (MO1.getImm() == 0) {
5586 switch (DataMI->getOpcode()) {
5587 default: llvm_unreachable("Unreachable!");
5589 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5591 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
5593 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
5594 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
5596 DataMI->setDesc(get(NewOpc));
5597 MO1.ChangeToRegister(MO0.getReg(), false);
5601 NewMIs.push_back(DataMI);
5603 // Emit the store instruction.
5605 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5606 std::pair<MachineInstr::mmo_iterator,
5607 MachineInstr::mmo_iterator> MMOs =
5608 MF.extractStoreMemRefs(MI->memoperands_begin(),
5609 MI->memoperands_end());
5610 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
5617 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5618 SmallVectorImpl<SDNode*> &NewNodes) const {
5619 if (!N->isMachineOpcode())
5622 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5623 MemOp2RegOpTable.find(N->getMachineOpcode());
5624 if (I == MemOp2RegOpTable.end())
5626 unsigned Opc = I->second.first;
5627 unsigned Index = I->second.second & TB_INDEX_MASK;
5628 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5629 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5630 const MCInstrDesc &MCID = get(Opc);
5631 MachineFunction &MF = DAG.getMachineFunction();
5632 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5633 unsigned NumDefs = MCID.NumDefs;
5634 std::vector<SDValue> AddrOps;
5635 std::vector<SDValue> BeforeOps;
5636 std::vector<SDValue> AfterOps;
5638 unsigned NumOps = N->getNumOperands();
5639 for (unsigned i = 0; i != NumOps-1; ++i) {
5640 SDValue Op = N->getOperand(i);
5641 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5642 AddrOps.push_back(Op);
5643 else if (i < Index-NumDefs)
5644 BeforeOps.push_back(Op);
5645 else if (i > Index-NumDefs)
5646 AfterOps.push_back(Op);
5648 SDValue Chain = N->getOperand(NumOps-1);
5649 AddrOps.push_back(Chain);
5651 // Emit the load instruction.
5652 SDNode *Load = nullptr;
5654 EVT VT = *RC->vt_begin();
5655 std::pair<MachineInstr::mmo_iterator,
5656 MachineInstr::mmo_iterator> MMOs =
5657 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
5658 cast<MachineSDNode>(N)->memoperands_end());
5659 if (!(*MMOs.first) &&
5660 RC == &X86::VR128RegClass &&
5661 !Subtarget.isUnalignedMemAccessFast())
5662 // Do not introduce a slow unaligned load.
5664 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
5665 bool isAligned = (*MMOs.first) &&
5666 (*MMOs.first)->getAlignment() >= Alignment;
5667 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
5668 VT, MVT::Other, AddrOps);
5669 NewNodes.push_back(Load);
5671 // Preserve memory reference information.
5672 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
5675 // Emit the data processing instruction.
5676 std::vector<EVT> VTs;
5677 const TargetRegisterClass *DstRC = nullptr;
5678 if (MCID.getNumDefs() > 0) {
5679 DstRC = getRegClass(MCID, 0, &RI, MF);
5680 VTs.push_back(*DstRC->vt_begin());
5682 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5683 EVT VT = N->getValueType(i);
5684 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5688 BeforeOps.push_back(SDValue(Load, 0));
5689 BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
5690 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5691 NewNodes.push_back(NewNode);
5693 // Emit the store instruction.
5696 AddrOps.push_back(SDValue(NewNode, 0));
5697 AddrOps.push_back(Chain);
5698 std::pair<MachineInstr::mmo_iterator,
5699 MachineInstr::mmo_iterator> MMOs =
5700 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
5701 cast<MachineSDNode>(N)->memoperands_end());
5702 if (!(*MMOs.first) &&
5703 RC == &X86::VR128RegClass &&
5704 !Subtarget.isUnalignedMemAccessFast())
5705 // Do not introduce a slow unaligned store.
5707 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
5708 bool isAligned = (*MMOs.first) &&
5709 (*MMOs.first)->getAlignment() >= Alignment;
5711 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5712 dl, MVT::Other, AddrOps);
5713 NewNodes.push_back(Store);
5715 // Preserve memory reference information.
5716 cast<MachineSDNode>(Store)->setMemRefs(MMOs.first, MMOs.second);
5722 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5723 bool UnfoldLoad, bool UnfoldStore,
5724 unsigned *LoadRegIndex) const {
5725 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5726 MemOp2RegOpTable.find(Opc);
5727 if (I == MemOp2RegOpTable.end())
5729 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5730 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5731 if (UnfoldLoad && !FoldedLoad)
5733 if (UnfoldStore && !FoldedStore)
5736 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
5737 return I->second.first;
5741 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5742 int64_t &Offset1, int64_t &Offset2) const {
5743 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5745 unsigned Opc1 = Load1->getMachineOpcode();
5746 unsigned Opc2 = Load2->getMachineOpcode();
5748 default: return false;
5758 case X86::MMX_MOVD64rm:
5759 case X86::MMX_MOVQ64rm:
5760 case X86::FsMOVAPSrm:
5761 case X86::FsMOVAPDrm:
5767 // AVX load instructions
5770 case X86::FsVMOVAPSrm:
5771 case X86::FsVMOVAPDrm:
5772 case X86::VMOVAPSrm:
5773 case X86::VMOVUPSrm:
5774 case X86::VMOVAPDrm:
5775 case X86::VMOVDQArm:
5776 case X86::VMOVDQUrm:
5777 case X86::VMOVAPSYrm:
5778 case X86::VMOVUPSYrm:
5779 case X86::VMOVAPDYrm:
5780 case X86::VMOVDQAYrm:
5781 case X86::VMOVDQUYrm:
5785 default: return false;
5795 case X86::MMX_MOVD64rm:
5796 case X86::MMX_MOVQ64rm:
5797 case X86::FsMOVAPSrm:
5798 case X86::FsMOVAPDrm:
5804 // AVX load instructions
5807 case X86::FsVMOVAPSrm:
5808 case X86::FsVMOVAPDrm:
5809 case X86::VMOVAPSrm:
5810 case X86::VMOVUPSrm:
5811 case X86::VMOVAPDrm:
5812 case X86::VMOVDQArm:
5813 case X86::VMOVDQUrm:
5814 case X86::VMOVAPSYrm:
5815 case X86::VMOVUPSYrm:
5816 case X86::VMOVAPDYrm:
5817 case X86::VMOVDQAYrm:
5818 case X86::VMOVDQUYrm:
5822 // Check if chain operands and base addresses match.
5823 if (Load1->getOperand(0) != Load2->getOperand(0) ||
5824 Load1->getOperand(5) != Load2->getOperand(5))
5826 // Segment operands should match as well.
5827 if (Load1->getOperand(4) != Load2->getOperand(4))
5829 // Scale should be 1, Index should be Reg0.
5830 if (Load1->getOperand(1) == Load2->getOperand(1) &&
5831 Load1->getOperand(2) == Load2->getOperand(2)) {
5832 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
5835 // Now let's examine the displacements.
5836 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
5837 isa<ConstantSDNode>(Load2->getOperand(3))) {
5838 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
5839 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
5846 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5847 int64_t Offset1, int64_t Offset2,
5848 unsigned NumLoads) const {
5849 assert(Offset2 > Offset1);
5850 if ((Offset2 - Offset1) / 8 > 64)
5853 unsigned Opc1 = Load1->getMachineOpcode();
5854 unsigned Opc2 = Load2->getMachineOpcode();
5856 return false; // FIXME: overly conservative?
5863 case X86::MMX_MOVD64rm:
5864 case X86::MMX_MOVQ64rm:
5868 EVT VT = Load1->getValueType(0);
5869 switch (VT.getSimpleVT().SimpleTy) {
5871 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5872 // have 16 of them to play with.
5873 if (Subtarget.is64Bit()) {
5876 } else if (NumLoads) {
5894 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
5895 MachineInstr *Second) const {
5896 // Check if this processor supports macro-fusion. Since this is a minor
5897 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
5898 // proxy for SandyBridge+.
5899 if (!Subtarget.hasAVX())
5908 switch(Second->getOpcode()) {
5931 FuseKind = FuseTest;
5934 switch (First->getOpcode()) {
5944 case X86::TEST32i32:
5945 case X86::TEST64i32:
5946 case X86::TEST64ri32:
5951 case X86::TEST8ri_NOREX:
5963 case X86::AND64ri32:
5983 case X86::CMP64ri32:
5994 case X86::ADD16ri8_DB:
5995 case X86::ADD16ri_DB:
5998 case X86::ADD16rr_DB:
6002 case X86::ADD32ri8_DB:
6003 case X86::ADD32ri_DB:
6006 case X86::ADD32rr_DB:
6008 case X86::ADD64ri32:
6009 case X86::ADD64ri32_DB:
6011 case X86::ADD64ri8_DB:
6014 case X86::ADD64rr_DB:
6032 case X86::SUB64ri32:
6040 return FuseKind == FuseCmp || FuseKind == FuseInc;
6049 return FuseKind == FuseInc;
6054 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
6055 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
6056 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
6057 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
6059 Cond[0].setImm(GetOppositeBranchCondition(CC));
6064 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
6065 // FIXME: Return false for x87 stack register classes for now. We can't
6066 // allow any loads of these registers before FpGet_ST0_80.
6067 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
6068 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
6071 /// Return a virtual register initialized with the
6072 /// the global base register value. Output instructions required to
6073 /// initialize the register in the function entry block, if necessary.
6075 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
6077 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
6078 assert(!Subtarget.is64Bit() &&
6079 "X86-64 PIC uses RIP relative addressing");
6081 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
6082 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
6083 if (GlobalBaseReg != 0)
6084 return GlobalBaseReg;
6086 // Create the register. The code to initialize it is inserted
6087 // later, by the CGBR pass (below).
6088 MachineRegisterInfo &RegInfo = MF->getRegInfo();
6089 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
6090 X86FI->setGlobalBaseReg(GlobalBaseReg);
6091 return GlobalBaseReg;
6094 // These are the replaceable SSE instructions. Some of these have Int variants
6095 // that we don't include here. We don't want to replace instructions selected
6097 static const uint16_t ReplaceableInstrs[][3] = {
6098 //PackedSingle PackedDouble PackedInt
6099 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
6100 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
6101 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
6102 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
6103 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
6104 { X86::MOVLPSmr, X86::MOVLPDmr, X86::MOVPQI2QImr },
6105 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
6106 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
6107 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
6108 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
6109 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
6110 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
6111 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
6112 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
6113 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
6114 // AVX 128-bit support
6115 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
6116 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
6117 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
6118 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
6119 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
6120 { X86::VMOVLPSmr, X86::VMOVLPDmr, X86::VMOVPQI2QImr },
6121 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
6122 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
6123 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
6124 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
6125 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
6126 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
6127 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
6128 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
6129 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
6130 // AVX 256-bit support
6131 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
6132 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
6133 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
6134 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
6135 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
6136 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
6139 static const uint16_t ReplaceableInstrsAVX2[][3] = {
6140 //PackedSingle PackedDouble PackedInt
6141 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
6142 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
6143 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
6144 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
6145 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
6146 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
6147 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
6148 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
6149 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
6150 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
6151 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
6152 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
6153 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
6154 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
6155 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
6156 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
6157 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
6158 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
6159 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
6160 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
6163 // FIXME: Some shuffle and unpack instructions have equivalents in different
6164 // domains, but they require a bit more work than just switching opcodes.
6166 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
6167 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
6168 if (ReplaceableInstrs[i][domain-1] == opcode)
6169 return ReplaceableInstrs[i];
6173 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
6174 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
6175 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
6176 return ReplaceableInstrsAVX2[i];
6180 std::pair<uint16_t, uint16_t>
6181 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
6182 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6183 bool hasAVX2 = Subtarget.hasAVX2();
6184 uint16_t validDomains = 0;
6185 if (domain && lookup(MI->getOpcode(), domain))
6187 else if (domain && lookupAVX2(MI->getOpcode(), domain))
6188 validDomains = hasAVX2 ? 0xe : 0x6;
6189 return std::make_pair(domain, validDomains);
6192 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
6193 assert(Domain>0 && Domain<4 && "Invalid execution domain");
6194 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6195 assert(dom && "Not an SSE instruction");
6196 const uint16_t *table = lookup(MI->getOpcode(), dom);
6197 if (!table) { // try the other table
6198 assert((Subtarget.hasAVX2() || Domain < 3) &&
6199 "256-bit vector operations only available in AVX2");
6200 table = lookupAVX2(MI->getOpcode(), dom);
6202 assert(table && "Cannot change domain");
6203 MI->setDesc(get(table[Domain-1]));
6206 /// Return the noop instruction to use for a noop.
6207 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
6208 NopInst.setOpcode(X86::NOOP);
6211 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
6212 // In particular, getJumpInstrTableEntryBound must always return an upper bound
6213 // on the encoding lengths of the instructions generated by
6214 // getUnconditionalBranch and getTrap.
6215 void X86InstrInfo::getUnconditionalBranch(
6216 MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const {
6217 Branch.setOpcode(X86::JMP_1);
6218 Branch.addOperand(MCOperand::createExpr(BranchTarget));
6221 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
6222 // In particular, getJumpInstrTableEntryBound must always return an upper bound
6223 // on the encoding lengths of the instructions generated by
6224 // getUnconditionalBranch and getTrap.
6225 void X86InstrInfo::getTrap(MCInst &MI) const {
6226 MI.setOpcode(X86::TRAP);
6229 // See getTrap and getUnconditionalBranch for conditions on the value returned
6230 // by this function.
6231 unsigned X86InstrInfo::getJumpInstrTableEntryBound() const {
6232 // 5 bytes suffice: JMP_4 Symbol@PLT is uses 1 byte (E9) for the JMP_4 and 4
6233 // bytes for the symbol offset. And TRAP is ud2, which is two bytes (0F 0B).
6237 bool X86InstrInfo::isHighLatencyDef(int opc) const {
6239 default: return false;
6241 case X86::DIVSDrm_Int:
6243 case X86::DIVSDrr_Int:
6245 case X86::DIVSSrm_Int:
6247 case X86::DIVSSrr_Int:
6253 case X86::SQRTSDm_Int:
6255 case X86::SQRTSDr_Int:
6257 case X86::SQRTSSm_Int:
6259 case X86::SQRTSSr_Int:
6260 // AVX instructions with high latency
6262 case X86::VDIVSDrm_Int:
6264 case X86::VDIVSDrr_Int:
6266 case X86::VDIVSSrm_Int:
6268 case X86::VDIVSSrr_Int:
6274 case X86::VSQRTSDm_Int:
6277 case X86::VSQRTSSm_Int:
6279 case X86::VSQRTPDZm:
6280 case X86::VSQRTPDZr:
6281 case X86::VSQRTPSZm:
6282 case X86::VSQRTPSZr:
6283 case X86::VSQRTSDZm:
6284 case X86::VSQRTSDZm_Int:
6285 case X86::VSQRTSDZr:
6286 case X86::VSQRTSSZm_Int:
6287 case X86::VSQRTSSZr:
6288 case X86::VSQRTSSZm:
6289 case X86::VDIVSDZrm:
6290 case X86::VDIVSDZrr:
6291 case X86::VDIVSSZrm:
6292 case X86::VDIVSSZrr:
6294 case X86::VGATHERQPSZrm:
6295 case X86::VGATHERQPDZrm:
6296 case X86::VGATHERDPDZrm:
6297 case X86::VGATHERDPSZrm:
6298 case X86::VPGATHERQDZrm:
6299 case X86::VPGATHERQQZrm:
6300 case X86::VPGATHERDDZrm:
6301 case X86::VPGATHERDQZrm:
6302 case X86::VSCATTERQPDZmr:
6303 case X86::VSCATTERQPSZmr:
6304 case X86::VSCATTERDPDZmr:
6305 case X86::VSCATTERDPSZmr:
6306 case X86::VPSCATTERQDZmr:
6307 case X86::VPSCATTERQQZmr:
6308 case X86::VPSCATTERDDZmr:
6309 case X86::VPSCATTERDQZmr:
6315 hasHighOperandLatency(const TargetSchedModel &SchedModel,
6316 const MachineRegisterInfo *MRI,
6317 const MachineInstr *DefMI, unsigned DefIdx,
6318 const MachineInstr *UseMI, unsigned UseIdx) const {
6319 return isHighLatencyDef(DefMI->getOpcode());
6322 static bool hasReassociableOperands(const MachineInstr &Inst,
6323 const MachineBasicBlock *MBB) {
6324 assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) &&
6325 "Reassociation needs binary operators");
6326 const MachineOperand &Op1 = Inst.getOperand(1);
6327 const MachineOperand &Op2 = Inst.getOperand(2);
6328 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
6330 // Integer binary math/logic instructions have a third source operand:
6331 // the EFLAGS register. That operand must be both defined here and never
6332 // used; ie, it must be dead. If the EFLAGS operand is live, then we can
6333 // not change anything because rearranging the operands could affect other
6334 // instructions that depend on the exact status flags (zero, sign, etc.)
6335 // that are set by using these particular operands with this operation.
6336 if (Inst.getNumOperands() == 4) {
6337 assert(Inst.getOperand(3).isReg() &&
6338 Inst.getOperand(3).getReg() == X86::EFLAGS &&
6339 "Unexpected operand in reassociable instruction");
6340 if (!Inst.getOperand(3).isDead())
6344 // We need virtual register definitions for the operands that we will
6346 MachineInstr *MI1 = nullptr;
6347 MachineInstr *MI2 = nullptr;
6348 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
6349 MI1 = MRI.getUniqueVRegDef(Op1.getReg());
6350 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
6351 MI2 = MRI.getUniqueVRegDef(Op2.getReg());
6353 // And they need to be in the trace (otherwise, they won't have a depth).
6354 if (MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB)
6360 static bool hasReassociableSibling(const MachineInstr &Inst, bool &Commuted) {
6361 const MachineBasicBlock *MBB = Inst.getParent();
6362 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
6363 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
6364 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
6365 unsigned AssocOpcode = Inst.getOpcode();
6367 // If only one operand has the same opcode and it's the second source operand,
6368 // the operands must be commuted.
6369 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
6371 std::swap(MI1, MI2);
6373 // 1. The previous instruction must be the same type as Inst.
6374 // 2. The previous instruction must have virtual register definitions for its
6375 // operands in the same basic block as Inst.
6376 // 3. The previous instruction's result must only be used by Inst.
6377 if (MI1->getOpcode() == AssocOpcode &&
6378 hasReassociableOperands(*MI1, MBB) &&
6379 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()))
6385 // TODO: There are many more machine instruction opcodes to match:
6386 // 1. Other data types (integer, vectors)
6387 // 2. Other math / logic operations (and, or)
6388 static bool isAssociativeAndCommutative(const MachineInstr &Inst) {
6389 switch (Inst.getOpcode()) {
6404 case X86::VADDPDYrr:
6405 case X86::VADDPSYrr:
6410 case X86::VMULPDYrr:
6411 case X86::VMULPSYrr:
6414 return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
6420 /// Return true if the input instruction is part of a chain of dependent ops
6421 /// that are suitable for reassociation, otherwise return false.
6422 /// If the instruction's operands must be commuted to have a previous
6423 /// instruction of the same type define the first source operand, Commuted will
6425 static bool isReassociationCandidate(const MachineInstr &Inst, bool &Commuted) {
6426 // 1. The operation must be associative and commutative.
6427 // 2. The instruction must have virtual register definitions for its
6428 // operands in the same basic block.
6429 // 3. The instruction must have a reassociable sibling.
6430 if (isAssociativeAndCommutative(Inst) &&
6431 hasReassociableOperands(Inst, Inst.getParent()) &&
6432 hasReassociableSibling(Inst, Commuted))
6438 // FIXME: This has the potential to be expensive (compile time) while not
6439 // improving the code at all. Some ways to limit the overhead:
6440 // 1. Track successful transforms; bail out if hit rate gets too low.
6441 // 2. Only enable at -O3 or some other non-default optimization level.
6442 // 3. Pre-screen pattern candidates here: if an operand of the previous
6443 // instruction is known to not increase the critical path, then don't match
6445 bool X86InstrInfo::getMachineCombinerPatterns(MachineInstr &Root,
6446 SmallVectorImpl<MachineCombinerPattern::MC_PATTERN> &Patterns) const {
6447 // TODO: There is nothing x86-specific here except the instruction type.
6448 // This logic could be hoisted into the machine combiner pass itself.
6450 // Look for this reassociation pattern:
6451 // B = A op X (Prev)
6452 // C = B op Y (Root)
6455 if (isReassociationCandidate(Root, Commute)) {
6456 // We found a sequence of instructions that may be suitable for a
6457 // reassociation of operands to increase ILP. Specify each commutation
6458 // possibility for the Prev instruction in the sequence and let the
6459 // machine combiner decide if changing the operands is worthwhile.
6461 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_YB);
6462 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_YB);
6464 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_BY);
6465 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_BY);
6473 /// This is an architecture-specific helper function of reassociateOps.
6474 /// Set special operand attributes for new instructions after reassociation.
6475 static void setSpecialOperandAttr(MachineInstr &OldMI1, MachineInstr &OldMI2,
6476 MachineInstr &NewMI1, MachineInstr &NewMI2) {
6477 // Integer instructions define an implicit EFLAGS source register operand as
6478 // the third source (fourth total) operand.
6479 if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4)
6482 assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 &&
6483 "Unexpected instruction type for reassociation");
6485 MachineOperand &OldOp1 = OldMI1.getOperand(3);
6486 MachineOperand &OldOp2 = OldMI2.getOperand(3);
6487 MachineOperand &NewOp1 = NewMI1.getOperand(3);
6488 MachineOperand &NewOp2 = NewMI2.getOperand(3);
6490 assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() &&
6491 "Must have dead EFLAGS operand in reassociable instruction");
6492 assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() &&
6493 "Must have dead EFLAGS operand in reassociable instruction");
6498 assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS &&
6499 "Unexpected operand in reassociable instruction");
6500 assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS &&
6501 "Unexpected operand in reassociable instruction");
6503 // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations
6504 // of this pass or other passes. The EFLAGS operands must be dead in these new
6505 // instructions because the EFLAGS operands in the original instructions must
6506 // be dead in order for reassociation to occur.
6511 /// Attempt the following reassociation to reduce critical path length:
6512 /// B = A op X (Prev)
6513 /// C = B op Y (Root)
6517 static void reassociateOps(MachineInstr &Root, MachineInstr &Prev,
6518 MachineCombinerPattern::MC_PATTERN Pattern,
6519 SmallVectorImpl<MachineInstr *> &InsInstrs,
6520 SmallVectorImpl<MachineInstr *> &DelInstrs,
6521 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) {
6522 MachineFunction *MF = Root.getParent()->getParent();
6523 MachineRegisterInfo &MRI = MF->getRegInfo();
6524 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
6525 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
6526 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
6528 // This array encodes the operand index for each parameter because the
6529 // operands may be commuted. Each row corresponds to a pattern value,
6530 // and each column specifies the index of A, B, X, Y.
6531 unsigned OpIdx[4][4] = {
6538 MachineOperand &OpA = Prev.getOperand(OpIdx[Pattern][0]);
6539 MachineOperand &OpB = Root.getOperand(OpIdx[Pattern][1]);
6540 MachineOperand &OpX = Prev.getOperand(OpIdx[Pattern][2]);
6541 MachineOperand &OpY = Root.getOperand(OpIdx[Pattern][3]);
6542 MachineOperand &OpC = Root.getOperand(0);
6544 unsigned RegA = OpA.getReg();
6545 unsigned RegB = OpB.getReg();
6546 unsigned RegX = OpX.getReg();
6547 unsigned RegY = OpY.getReg();
6548 unsigned RegC = OpC.getReg();
6550 if (TargetRegisterInfo::isVirtualRegister(RegA))
6551 MRI.constrainRegClass(RegA, RC);
6552 if (TargetRegisterInfo::isVirtualRegister(RegB))
6553 MRI.constrainRegClass(RegB, RC);
6554 if (TargetRegisterInfo::isVirtualRegister(RegX))
6555 MRI.constrainRegClass(RegX, RC);
6556 if (TargetRegisterInfo::isVirtualRegister(RegY))
6557 MRI.constrainRegClass(RegY, RC);
6558 if (TargetRegisterInfo::isVirtualRegister(RegC))
6559 MRI.constrainRegClass(RegC, RC);
6561 // Create a new virtual register for the result of (X op Y) instead of
6562 // recycling RegB because the MachineCombiner's computation of the critical
6563 // path requires a new register definition rather than an existing one.
6564 unsigned NewVR = MRI.createVirtualRegister(RC);
6565 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
6567 unsigned Opcode = Root.getOpcode();
6568 bool KillA = OpA.isKill();
6569 bool KillX = OpX.isKill();
6570 bool KillY = OpY.isKill();
6572 // Create new instructions for insertion.
6573 MachineInstrBuilder MIB1 =
6574 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
6575 .addReg(RegX, getKillRegState(KillX))
6576 .addReg(RegY, getKillRegState(KillY));
6577 MachineInstrBuilder MIB2 =
6578 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
6579 .addReg(RegA, getKillRegState(KillA))
6580 .addReg(NewVR, getKillRegState(true));
6582 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
6584 // Record new instructions for insertion and old instructions for deletion.
6585 InsInstrs.push_back(MIB1);
6586 InsInstrs.push_back(MIB2);
6587 DelInstrs.push_back(&Prev);
6588 DelInstrs.push_back(&Root);
6591 void X86InstrInfo::genAlternativeCodeSequence(
6593 MachineCombinerPattern::MC_PATTERN Pattern,
6594 SmallVectorImpl<MachineInstr *> &InsInstrs,
6595 SmallVectorImpl<MachineInstr *> &DelInstrs,
6596 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
6597 MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo();
6599 // Select the previous instruction in the sequence based on the input pattern.
6600 MachineInstr *Prev = nullptr;
6602 case MachineCombinerPattern::MC_REASSOC_AX_BY:
6603 case MachineCombinerPattern::MC_REASSOC_XA_BY:
6604 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
6606 case MachineCombinerPattern::MC_REASSOC_AX_YB:
6607 case MachineCombinerPattern::MC_REASSOC_XA_YB:
6608 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
6610 assert(Prev && "Unknown pattern for machine combiner");
6612 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
6616 std::pair<unsigned, unsigned>
6617 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
6618 return std::make_pair(TF, 0u);
6621 ArrayRef<std::pair<unsigned, const char *>>
6622 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
6623 using namespace X86II;
6624 static std::pair<unsigned, const char *> TargetFlags[] = {
6625 {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"},
6626 {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"},
6627 {MO_GOT, "x86-got"},
6628 {MO_GOTOFF, "x86-gotoff"},
6629 {MO_GOTPCREL, "x86-gotpcrel"},
6630 {MO_PLT, "x86-plt"},
6631 {MO_TLSGD, "x86-tlsgd"},
6632 {MO_TLSLD, "x86-tlsld"},
6633 {MO_TLSLDM, "x86-tlsldm"},
6634 {MO_GOTTPOFF, "x86-gottpoff"},
6635 {MO_INDNTPOFF, "x86-indntpoff"},
6636 {MO_TPOFF, "x86-tpoff"},
6637 {MO_DTPOFF, "x86-dtpoff"},
6638 {MO_NTPOFF, "x86-ntpoff"},
6639 {MO_GOTNTPOFF, "x86-gotntpoff"},
6640 {MO_DLLIMPORT, "x86-dllimport"},
6641 {MO_DARWIN_STUB, "x86-darwin-stub"},
6642 {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"},
6643 {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"},
6644 {MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE, "x86-darwin-hidden-nonlazy-pic-base"},
6645 {MO_TLVP, "x86-tlvp"},
6646 {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"},
6647 {MO_SECREL, "x86-secrel"}};
6648 return makeArrayRef(TargetFlags);
6652 /// Create Global Base Reg pass. This initializes the PIC
6653 /// global base register for x86-32.
6654 struct CGBR : public MachineFunctionPass {
6656 CGBR() : MachineFunctionPass(ID) {}
6658 bool runOnMachineFunction(MachineFunction &MF) override {
6659 const X86TargetMachine *TM =
6660 static_cast<const X86TargetMachine *>(&MF.getTarget());
6661 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
6663 // Don't do anything if this is 64-bit as 64-bit PIC
6664 // uses RIP relative addressing.
6668 // Only emit a global base reg in PIC mode.
6669 if (TM->getRelocationModel() != Reloc::PIC_)
6672 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
6673 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
6675 // If we didn't need a GlobalBaseReg, don't insert code.
6676 if (GlobalBaseReg == 0)
6679 // Insert the set of GlobalBaseReg into the first MBB of the function
6680 MachineBasicBlock &FirstMBB = MF.front();
6681 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
6682 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
6683 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6684 const X86InstrInfo *TII = STI.getInstrInfo();
6687 if (STI.isPICStyleGOT())
6688 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
6692 // Operand of MovePCtoStack is completely ignored by asm printer. It's
6693 // only used in JIT code emission as displacement to pc.
6694 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
6696 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
6697 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
6698 if (STI.isPICStyleGOT()) {
6699 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
6700 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
6701 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
6702 X86II::MO_GOT_ABSOLUTE_ADDRESS);
6708 const char *getPassName() const override {
6709 return "X86 PIC Global Base Reg Initialization";
6712 void getAnalysisUsage(AnalysisUsage &AU) const override {
6713 AU.setPreservesCFG();
6714 MachineFunctionPass::getAnalysisUsage(AU);
6721 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
6724 struct LDTLSCleanup : public MachineFunctionPass {
6726 LDTLSCleanup() : MachineFunctionPass(ID) {}
6728 bool runOnMachineFunction(MachineFunction &MF) override {
6729 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
6730 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
6731 // No point folding accesses if there isn't at least two.
6735 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
6736 return VisitNode(DT->getRootNode(), 0);
6739 // Visit the dominator subtree rooted at Node in pre-order.
6740 // If TLSBaseAddrReg is non-null, then use that to replace any
6741 // TLS_base_addr instructions. Otherwise, create the register
6742 // when the first such instruction is seen, and then use it
6743 // as we encounter more instructions.
6744 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
6745 MachineBasicBlock *BB = Node->getBlock();
6746 bool Changed = false;
6748 // Traverse the current block.
6749 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
6751 switch (I->getOpcode()) {
6752 case X86::TLS_base_addr32:
6753 case X86::TLS_base_addr64:
6755 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
6757 I = SetRegister(I, &TLSBaseAddrReg);
6765 // Visit the children of this block in the dominator tree.
6766 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
6768 Changed |= VisitNode(*I, TLSBaseAddrReg);
6774 // Replace the TLS_base_addr instruction I with a copy from
6775 // TLSBaseAddrReg, returning the new instruction.
6776 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
6777 unsigned TLSBaseAddrReg) {
6778 MachineFunction *MF = I->getParent()->getParent();
6779 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
6780 const bool is64Bit = STI.is64Bit();
6781 const X86InstrInfo *TII = STI.getInstrInfo();
6783 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
6784 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
6785 TII->get(TargetOpcode::COPY),
6786 is64Bit ? X86::RAX : X86::EAX)
6787 .addReg(TLSBaseAddrReg);
6789 // Erase the TLS_base_addr instruction.
6790 I->eraseFromParent();
6795 // Create a virtal register in *TLSBaseAddrReg, and populate it by
6796 // inserting a copy instruction after I. Returns the new instruction.
6797 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
6798 MachineFunction *MF = I->getParent()->getParent();
6799 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
6800 const bool is64Bit = STI.is64Bit();
6801 const X86InstrInfo *TII = STI.getInstrInfo();
6803 // Create a virtual register for the TLS base address.
6804 MachineRegisterInfo &RegInfo = MF->getRegInfo();
6805 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
6806 ? &X86::GR64RegClass
6807 : &X86::GR32RegClass);
6809 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
6810 MachineInstr *Next = I->getNextNode();
6811 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
6812 TII->get(TargetOpcode::COPY),
6814 .addReg(is64Bit ? X86::RAX : X86::EAX);
6819 const char *getPassName() const override {
6820 return "Local Dynamic TLS Access Clean-up";
6823 void getAnalysisUsage(AnalysisUsage &AU) const override {
6824 AU.setPreservesCFG();
6825 AU.addRequired<MachineDominatorTree>();
6826 MachineFunctionPass::getAnalysisUsage(AU);
6831 char LDTLSCleanup::ID = 0;
6833 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }