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/IR/DerivedTypes.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/MC/MCAsmInfo.h"
30 #include "llvm/MC/MCInst.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
38 #define GET_INSTRINFO_CTOR
39 #include "X86GenInstrInfo.inc"
44 NoFusing("disable-spill-fusing",
45 cl::desc("Disable fusing of spill code into instructions"));
47 PrintFailedFusing("print-failed-fuse-candidates",
48 cl::desc("Print instructions that the allocator wants to"
49 " fuse, but the X86 backend currently can't"),
52 ReMatPICStubLoad("remat-pic-stub-load",
53 cl::desc("Re-materialize load from stub in PIC mode"),
54 cl::init(false), cl::Hidden);
57 // Select which memory operand is being unfolded.
58 // (stored in bits 0 - 3)
65 // Do not insert the reverse map (MemOp -> RegOp) into the table.
66 // This may be needed because there is a many -> one mapping.
67 TB_NO_REVERSE = 1 << 4,
69 // Do not insert the forward map (RegOp -> MemOp) into the table.
70 // This is needed for Native Client, which prohibits branch
71 // instructions from using a memory operand.
72 TB_NO_FORWARD = 1 << 5,
74 TB_FOLDED_LOAD = 1 << 6,
75 TB_FOLDED_STORE = 1 << 7,
77 // Minimum alignment required for load/store.
78 // Used for RegOp->MemOp conversion.
79 // (stored in bits 8 - 15)
81 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
82 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
83 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
87 struct X86OpTblEntry {
93 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
94 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
95 ? X86::ADJCALLSTACKDOWN64
96 : X86::ADJCALLSTACKDOWN32),
97 (tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKUP64
99 : X86::ADJCALLSTACKUP32)),
100 TM(tm), RI(tm, *this) {
102 static const X86OpTblEntry OpTbl2Addr[] = {
103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
270 unsigned RegOp = OpTbl2Addr[i].RegOp;
271 unsigned MemOp = OpTbl2Addr[i].MemOp;
272 unsigned Flags = OpTbl2Addr[i].Flags;
273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
275 // Index 0, folded load and store, no alignment requirement.
276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
279 static const X86OpTblEntry OpTbl0[] = {
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::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
381 unsigned RegOp = OpTbl0[i].RegOp;
382 unsigned MemOp = OpTbl0[i].MemOp;
383 unsigned Flags = OpTbl0[i].Flags;
384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
388 static const X86OpTblEntry OpTbl1[] = {
389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
417 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
418 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
419 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
420 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
421 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
422 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
423 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
424 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
425 { X86::MOV16rr, X86::MOV16rm, 0 },
426 { X86::MOV32rr, X86::MOV32rm, 0 },
427 { X86::MOV64rr, X86::MOV64rm, 0 },
428 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
429 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
430 { X86::MOV8rr, X86::MOV8rm, 0 },
431 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
432 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
433 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
434 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
435 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
436 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
437 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
438 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
439 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
440 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
441 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
442 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
443 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
444 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
445 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
446 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
447 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
448 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
449 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
450 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
451 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
452 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
453 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
454 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
455 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
456 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
457 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
458 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
459 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
460 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
461 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
462 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
463 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
464 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
465 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
466 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
467 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
468 { X86::SQRTSDr, X86::SQRTSDm, 0 },
469 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
470 { X86::SQRTSSr, X86::SQRTSSm, 0 },
471 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
472 { X86::TEST16rr, X86::TEST16rm, 0 },
473 { X86::TEST32rr, X86::TEST32rm, 0 },
474 { X86::TEST64rr, X86::TEST64rm, 0 },
475 { X86::TEST8rr, X86::TEST8rm, 0 },
476 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
477 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
478 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
479 // AVX 128-bit versions of foldable instructions
480 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
481 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
482 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
483 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
484 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
485 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
486 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
487 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
488 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
489 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
490 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
491 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
492 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
493 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
494 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
495 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
496 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
497 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
498 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
499 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
500 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
501 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
502 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
503 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
504 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
505 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
506 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
507 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
508 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
509 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
510 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
511 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
512 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
513 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
514 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
515 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
516 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
517 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
518 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
519 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
520 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
521 { X86::VRCPPSr, X86::VRCPPSm, 0 },
522 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
523 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
524 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
525 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
526 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
527 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
528 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
529 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
531 // AVX 256-bit foldable instructions
532 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
533 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
534 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
535 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
536 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
537 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
538 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
540 // AVX2 foldable instructions
541 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
542 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
543 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
544 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
545 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
546 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
547 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
548 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
549 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
550 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
551 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
552 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
553 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
555 // BMI/BMI2/LZCNT/POPCNT foldable instructions
556 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
557 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
558 { X86::BLSI32rr, X86::BLSI32rm, 0 },
559 { X86::BLSI64rr, X86::BLSI64rm, 0 },
560 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
561 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
562 { X86::BLSR32rr, X86::BLSR32rm, 0 },
563 { X86::BLSR64rr, X86::BLSR64rm, 0 },
564 { X86::BZHI32rr, X86::BZHI32rm, 0 },
565 { X86::BZHI64rr, X86::BZHI64rm, 0 },
566 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
567 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
568 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
569 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
570 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
571 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
572 { X86::RORX32ri, X86::RORX32mi, 0 },
573 { X86::RORX64ri, X86::RORX64mi, 0 },
574 { X86::SARX32rr, X86::SARX32rm, 0 },
575 { X86::SARX64rr, X86::SARX64rm, 0 },
576 { X86::SHRX32rr, X86::SHRX32rm, 0 },
577 { X86::SHRX64rr, X86::SHRX64rm, 0 },
578 { X86::SHLX32rr, X86::SHLX32rm, 0 },
579 { X86::SHLX64rr, X86::SHLX64rm, 0 },
580 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
581 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
582 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
585 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
586 unsigned RegOp = OpTbl1[i].RegOp;
587 unsigned MemOp = OpTbl1[i].MemOp;
588 unsigned Flags = OpTbl1[i].Flags;
589 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
591 // Index 1, folded load
592 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
595 static const X86OpTblEntry OpTbl2[] = {
596 { X86::ADC32rr, X86::ADC32rm, 0 },
597 { X86::ADC64rr, X86::ADC64rm, 0 },
598 { X86::ADD16rr, X86::ADD16rm, 0 },
599 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
600 { X86::ADD32rr, X86::ADD32rm, 0 },
601 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
602 { X86::ADD64rr, X86::ADD64rm, 0 },
603 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
604 { X86::ADD8rr, X86::ADD8rm, 0 },
605 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
606 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
607 { X86::ADDSDrr, X86::ADDSDrm, 0 },
608 { X86::ADDSSrr, X86::ADDSSrm, 0 },
609 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
610 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
611 { X86::AND16rr, X86::AND16rm, 0 },
612 { X86::AND32rr, X86::AND32rm, 0 },
613 { X86::AND64rr, X86::AND64rm, 0 },
614 { X86::AND8rr, X86::AND8rm, 0 },
615 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
616 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
617 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
618 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
619 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
620 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
621 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
622 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
623 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
624 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
625 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
626 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
627 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
628 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
629 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
630 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
631 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
632 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
633 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
634 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
635 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
636 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
637 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
638 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
639 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
640 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
641 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
642 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
643 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
644 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
645 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
646 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
647 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
648 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
649 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
650 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
651 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
652 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
653 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
654 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
655 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
656 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
657 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
658 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
659 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
660 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
661 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
662 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
663 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
664 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
665 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
666 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
667 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
668 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
669 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
670 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
671 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
672 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
673 { X86::CMPSDrr, X86::CMPSDrm, 0 },
674 { X86::CMPSSrr, X86::CMPSSrm, 0 },
675 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
676 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
677 { X86::DIVSDrr, X86::DIVSDrm, 0 },
678 { X86::DIVSSrr, X86::DIVSSrm, 0 },
679 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
680 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
681 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
682 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
683 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
684 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
685 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
686 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
687 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
688 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
689 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
690 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
691 { X86::IMUL16rr, X86::IMUL16rm, 0 },
692 { X86::IMUL32rr, X86::IMUL32rm, 0 },
693 { X86::IMUL64rr, X86::IMUL64rm, 0 },
694 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
695 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
696 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
697 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
698 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
699 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
700 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
701 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
702 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
703 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
704 { X86::MAXSDrr, X86::MAXSDrm, 0 },
705 { X86::MAXSSrr, X86::MAXSSrm, 0 },
706 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
707 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
708 { X86::MINSDrr, X86::MINSDrm, 0 },
709 { X86::MINSSrr, X86::MINSSrm, 0 },
710 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
711 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
712 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
713 { X86::MULSDrr, X86::MULSDrm, 0 },
714 { X86::MULSSrr, X86::MULSSrm, 0 },
715 { X86::OR16rr, X86::OR16rm, 0 },
716 { X86::OR32rr, X86::OR32rm, 0 },
717 { X86::OR64rr, X86::OR64rm, 0 },
718 { X86::OR8rr, X86::OR8rm, 0 },
719 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
720 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
721 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
722 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
723 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
724 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
725 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
726 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
727 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
728 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
729 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
730 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
731 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
732 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
733 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
734 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
735 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
736 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
737 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
738 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
739 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
740 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
741 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
742 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
743 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
744 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
745 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
746 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
747 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
748 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
749 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
750 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
751 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
752 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
753 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
754 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
755 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
756 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
757 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
758 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
759 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
760 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
761 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
762 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
763 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
764 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
765 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
766 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
767 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
768 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
769 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
770 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
771 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
772 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
773 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
774 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
775 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
776 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
777 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
778 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
779 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
780 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
781 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
782 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
783 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
784 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
785 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
786 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
787 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
788 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
789 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
790 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
791 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
792 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
793 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
794 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
795 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
796 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
797 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
798 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
799 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
800 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
801 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
802 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
803 { X86::SBB32rr, X86::SBB32rm, 0 },
804 { X86::SBB64rr, X86::SBB64rm, 0 },
805 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
806 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
807 { X86::SUB16rr, X86::SUB16rm, 0 },
808 { X86::SUB32rr, X86::SUB32rm, 0 },
809 { X86::SUB64rr, X86::SUB64rm, 0 },
810 { X86::SUB8rr, X86::SUB8rm, 0 },
811 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
812 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
813 { X86::SUBSDrr, X86::SUBSDrm, 0 },
814 { X86::SUBSSrr, X86::SUBSSrm, 0 },
815 // FIXME: TEST*rr -> swapped operand of TEST*mr.
816 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
817 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
818 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
819 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
820 { X86::XOR16rr, X86::XOR16rm, 0 },
821 { X86::XOR32rr, X86::XOR32rm, 0 },
822 { X86::XOR64rr, X86::XOR64rm, 0 },
823 { X86::XOR8rr, X86::XOR8rm, 0 },
824 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
825 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
826 // AVX 128-bit versions of foldable instructions
827 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
828 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
829 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
830 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
831 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
832 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
833 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
834 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
835 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
836 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
837 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
838 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
839 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
840 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
841 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
842 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
843 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
844 { X86::VADDPDrr, X86::VADDPDrm, 0 },
845 { X86::VADDPSrr, X86::VADDPSrm, 0 },
846 { X86::VADDSDrr, X86::VADDSDrm, 0 },
847 { X86::VADDSSrr, X86::VADDSSrm, 0 },
848 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
849 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
850 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
851 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
852 { X86::VANDPDrr, X86::VANDPDrm, 0 },
853 { X86::VANDPSrr, X86::VANDPSrm, 0 },
854 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
855 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
856 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
857 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
858 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
859 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
860 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
861 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
862 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
863 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
864 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
865 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
866 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
867 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
868 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
869 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
870 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
871 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
872 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
873 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
874 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
875 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
876 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
877 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
878 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
879 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
880 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
881 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
882 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
883 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
884 { X86::VMINPDrr, X86::VMINPDrm, 0 },
885 { X86::VMINPSrr, X86::VMINPSrm, 0 },
886 { X86::VMINSDrr, X86::VMINSDrm, 0 },
887 { X86::VMINSSrr, X86::VMINSSrm, 0 },
888 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
889 { X86::VMULPDrr, X86::VMULPDrm, 0 },
890 { X86::VMULPSrr, X86::VMULPSrm, 0 },
891 { X86::VMULSDrr, X86::VMULSDrm, 0 },
892 { X86::VMULSSrr, X86::VMULSSrm, 0 },
893 { X86::VORPDrr, X86::VORPDrm, 0 },
894 { X86::VORPSrr, X86::VORPSrm, 0 },
895 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
896 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
897 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
898 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
899 { X86::VPADDBrr, X86::VPADDBrm, 0 },
900 { X86::VPADDDrr, X86::VPADDDrm, 0 },
901 { X86::VPADDQrr, X86::VPADDQrm, 0 },
902 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
903 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
904 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
905 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
906 { X86::VPADDWrr, X86::VPADDWrm, 0 },
907 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
908 { X86::VPANDNrr, X86::VPANDNrm, 0 },
909 { X86::VPANDrr, X86::VPANDrm, 0 },
910 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
911 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
912 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
913 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
914 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
915 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
916 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
917 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
918 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
919 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
920 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
921 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
922 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
923 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
924 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
925 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
926 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
927 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
928 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
929 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
930 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
931 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
932 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
933 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
934 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
935 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
936 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
937 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
938 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
939 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
940 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
941 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
942 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
943 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
944 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
945 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
946 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
947 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
948 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
949 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
950 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
951 { X86::VPORrr, X86::VPORrm, 0 },
952 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
953 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
954 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
955 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
956 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
957 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
958 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
959 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
960 { X86::VPSRADrr, X86::VPSRADrm, 0 },
961 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
962 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
963 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
964 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
965 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
966 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
967 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
968 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
969 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
970 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
971 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
972 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
973 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
974 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
975 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
976 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
977 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
978 { X86::VPXORrr, X86::VPXORrm, 0 },
979 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
980 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
981 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
982 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
983 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
984 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
985 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
986 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
987 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
988 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
989 { X86::VXORPDrr, X86::VXORPDrm, 0 },
990 { X86::VXORPSrr, X86::VXORPSrm, 0 },
991 // AVX 256-bit foldable instructions
992 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
993 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
994 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
995 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
996 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
997 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
998 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
999 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1000 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1001 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1002 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1003 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1004 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1005 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1006 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1007 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1008 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1009 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1010 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1011 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1012 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1013 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
1014 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
1015 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
1016 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
1017 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1018 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1019 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1020 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1021 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1022 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1023 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1024 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1025 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1026 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1027 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1028 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1029 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1030 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1031 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1032 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1033 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
1034 // AVX2 foldable instructions
1035 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1036 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1037 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1038 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1039 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1040 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1041 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1042 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1043 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1044 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1045 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1046 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1047 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1048 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1049 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1050 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1051 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1052 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1053 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1054 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1055 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1056 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1057 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1058 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1059 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1060 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1061 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1062 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1063 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1064 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1065 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1066 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1067 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1068 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1069 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1070 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1071 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1072 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1073 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1074 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1075 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1076 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1077 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1078 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1079 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1080 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1081 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1082 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1083 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1084 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1085 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1086 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1087 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1088 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1089 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1090 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1091 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1092 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1093 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1094 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1095 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1096 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1097 { X86::VPORYrr, X86::VPORYrm, 0 },
1098 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1099 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1100 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1101 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1102 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1103 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1104 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1105 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1106 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1107 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1108 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1109 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1110 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1111 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1112 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1113 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1114 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1115 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1116 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1117 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1118 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1119 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1120 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1121 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1122 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1123 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1124 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1125 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1126 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1127 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1128 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1129 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1130 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1131 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1132 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1133 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1134 { X86::VPXORYrr, X86::VPXORYrm, 0 },
1135 // FIXME: add AVX 256-bit foldable instructions
1137 // FMA4 foldable patterns
1138 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1139 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
1140 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1141 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1142 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1143 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1144 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1145 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
1146 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1147 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1148 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1149 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1150 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1151 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
1152 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1153 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1154 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1155 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1156 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1157 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
1158 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1159 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1160 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1161 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1162 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1163 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1164 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1165 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1166 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1167 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1168 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1169 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
1171 // BMI/BMI2 foldable instructions
1172 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1173 { X86::ANDN64rr, X86::ANDN64rm, 0 },
1174 { X86::MULX32rr, X86::MULX32rm, 0 },
1175 { X86::MULX64rr, X86::MULX64rm, 0 },
1176 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1177 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1178 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1179 { X86::PEXT64rr, X86::PEXT64rm, 0 },
1182 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1183 unsigned RegOp = OpTbl2[i].RegOp;
1184 unsigned MemOp = OpTbl2[i].MemOp;
1185 unsigned Flags = OpTbl2[i].Flags;
1186 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1188 // Index 2, folded load
1189 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1192 static const X86OpTblEntry OpTbl3[] = {
1193 // FMA foldable instructions
1194 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1195 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1196 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1197 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1198 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1199 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
1200 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1201 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
1203 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1204 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1205 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1206 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1207 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1208 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1209 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1210 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1211 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1212 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1213 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1214 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
1216 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1217 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1218 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1219 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1220 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1221 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
1222 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1223 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
1225 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1226 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1227 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1228 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1229 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1230 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1231 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1232 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1233 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1234 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1235 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1236 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
1238 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1239 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1240 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1241 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1242 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1243 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
1244 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1245 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
1247 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1248 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1249 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1250 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1251 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1252 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1253 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1254 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1255 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1256 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1257 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1258 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
1260 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1261 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1262 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1263 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1264 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1265 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
1266 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1267 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
1269 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1270 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1271 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1272 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1273 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1274 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1275 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1276 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1277 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1278 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1279 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1280 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
1282 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1283 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1284 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1285 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1286 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1287 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1288 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1289 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1290 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1291 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1292 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1293 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
1295 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1296 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1297 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1298 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1299 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1300 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1301 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1302 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1303 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1304 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1305 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1306 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
1308 // FMA4 foldable patterns
1309 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1310 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
1311 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1312 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1313 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1314 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1315 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1316 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
1317 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1318 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1319 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1320 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1321 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1322 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
1323 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1324 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1325 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1326 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1327 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1328 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
1329 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1330 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1331 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1332 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1333 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1334 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1335 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1336 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1337 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1338 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1339 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1340 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
1343 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1344 unsigned RegOp = OpTbl3[i].RegOp;
1345 unsigned MemOp = OpTbl3[i].MemOp;
1346 unsigned Flags = OpTbl3[i].Flags;
1347 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1349 // Index 3, folded load
1350 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1356 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1357 MemOp2RegOpTableType &M2RTable,
1358 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1359 if ((Flags & TB_NO_FORWARD) == 0) {
1360 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1361 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1363 if ((Flags & TB_NO_REVERSE) == 0) {
1364 assert(!M2RTable.count(MemOp) &&
1365 "Duplicated entries in unfolding maps?");
1366 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1371 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1372 unsigned &SrcReg, unsigned &DstReg,
1373 unsigned &SubIdx) const {
1374 switch (MI.getOpcode()) {
1376 case X86::MOVSX16rr8:
1377 case X86::MOVZX16rr8:
1378 case X86::MOVSX32rr8:
1379 case X86::MOVZX32rr8:
1380 case X86::MOVSX64rr8:
1381 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1382 // It's not always legal to reference the low 8-bit of the larger
1383 // register in 32-bit mode.
1385 case X86::MOVSX32rr16:
1386 case X86::MOVZX32rr16:
1387 case X86::MOVSX64rr16:
1388 case X86::MOVSX64rr32: {
1389 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1392 SrcReg = MI.getOperand(1).getReg();
1393 DstReg = MI.getOperand(0).getReg();
1394 switch (MI.getOpcode()) {
1395 default: llvm_unreachable("Unreachable!");
1396 case X86::MOVSX16rr8:
1397 case X86::MOVZX16rr8:
1398 case X86::MOVSX32rr8:
1399 case X86::MOVZX32rr8:
1400 case X86::MOVSX64rr8:
1401 SubIdx = X86::sub_8bit;
1403 case X86::MOVSX32rr16:
1404 case X86::MOVZX32rr16:
1405 case X86::MOVSX64rr16:
1406 SubIdx = X86::sub_16bit;
1408 case X86::MOVSX64rr32:
1409 SubIdx = X86::sub_32bit;
1418 /// isFrameOperand - Return true and the FrameIndex if the specified
1419 /// operand and follow operands form a reference to the stack frame.
1420 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1421 int &FrameIndex) const {
1422 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1423 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1424 MI->getOperand(Op+1).getImm() == 1 &&
1425 MI->getOperand(Op+2).getReg() == 0 &&
1426 MI->getOperand(Op+3).getImm() == 0) {
1427 FrameIndex = MI->getOperand(Op).getIndex();
1433 static bool isFrameLoadOpcode(int Opcode) {
1449 case X86::VMOVAPSrm:
1450 case X86::VMOVAPDrm:
1451 case X86::VMOVDQArm:
1452 case X86::VMOVAPSYrm:
1453 case X86::VMOVAPDYrm:
1454 case X86::VMOVDQAYrm:
1455 case X86::MMX_MOVD64rm:
1456 case X86::MMX_MOVQ64rm:
1461 static bool isFrameStoreOpcode(int Opcode) {
1468 case X86::ST_FpP64m:
1476 case X86::VMOVAPSmr:
1477 case X86::VMOVAPDmr:
1478 case X86::VMOVDQAmr:
1479 case X86::VMOVAPSYmr:
1480 case X86::VMOVAPDYmr:
1481 case X86::VMOVDQAYmr:
1482 case X86::MMX_MOVD64mr:
1483 case X86::MMX_MOVQ64mr:
1484 case X86::MMX_MOVNTQmr:
1490 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1491 int &FrameIndex) const {
1492 if (isFrameLoadOpcode(MI->getOpcode()))
1493 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1494 return MI->getOperand(0).getReg();
1498 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1499 int &FrameIndex) const {
1500 if (isFrameLoadOpcode(MI->getOpcode())) {
1502 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1504 // Check for post-frame index elimination operations
1505 const MachineMemOperand *Dummy;
1506 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1511 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1512 int &FrameIndex) const {
1513 if (isFrameStoreOpcode(MI->getOpcode()))
1514 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1515 isFrameOperand(MI, 0, FrameIndex))
1516 return MI->getOperand(X86::AddrNumOperands).getReg();
1520 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1521 int &FrameIndex) const {
1522 if (isFrameStoreOpcode(MI->getOpcode())) {
1524 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1526 // Check for post-frame index elimination operations
1527 const MachineMemOperand *Dummy;
1528 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1533 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1535 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1536 // Don't waste compile time scanning use-def chains of physregs.
1537 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1539 bool isPICBase = false;
1540 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1541 E = MRI.def_end(); I != E; ++I) {
1542 MachineInstr *DefMI = I.getOperand().getParent();
1543 if (DefMI->getOpcode() != X86::MOVPC32r)
1545 assert(!isPICBase && "More than one PIC base?");
1552 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1553 AliasAnalysis *AA) const {
1554 switch (MI->getOpcode()) {
1570 case X86::VMOVAPSrm:
1571 case X86::VMOVUPSrm:
1572 case X86::VMOVAPDrm:
1573 case X86::VMOVDQArm:
1574 case X86::VMOVDQUrm:
1575 case X86::VMOVAPSYrm:
1576 case X86::VMOVUPSYrm:
1577 case X86::VMOVAPDYrm:
1578 case X86::VMOVDQAYrm:
1579 case X86::VMOVDQUYrm:
1580 case X86::MMX_MOVD64rm:
1581 case X86::MMX_MOVQ64rm:
1582 case X86::FsVMOVAPSrm:
1583 case X86::FsVMOVAPDrm:
1584 case X86::FsMOVAPSrm:
1585 case X86::FsMOVAPDrm: {
1586 // Loads from constant pools are trivially rematerializable.
1587 if (MI->getOperand(1).isReg() &&
1588 MI->getOperand(2).isImm() &&
1589 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1590 MI->isInvariantLoad(AA)) {
1591 unsigned BaseReg = MI->getOperand(1).getReg();
1592 if (BaseReg == 0 || BaseReg == X86::RIP)
1594 // Allow re-materialization of PIC load.
1595 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1597 const MachineFunction &MF = *MI->getParent()->getParent();
1598 const MachineRegisterInfo &MRI = MF.getRegInfo();
1599 return regIsPICBase(BaseReg, MRI);
1606 if (MI->getOperand(2).isImm() &&
1607 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1608 !MI->getOperand(4).isReg()) {
1609 // lea fi#, lea GV, etc. are all rematerializable.
1610 if (!MI->getOperand(1).isReg())
1612 unsigned BaseReg = MI->getOperand(1).getReg();
1615 // Allow re-materialization of lea PICBase + x.
1616 const MachineFunction &MF = *MI->getParent()->getParent();
1617 const MachineRegisterInfo &MRI = MF.getRegInfo();
1618 return regIsPICBase(BaseReg, MRI);
1624 // All other instructions marked M_REMATERIALIZABLE are always trivially
1625 // rematerializable.
1629 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1630 /// would clobber the EFLAGS condition register. Note the result may be
1631 /// conservative. If it cannot definitely determine the safety after visiting
1632 /// a few instructions in each direction it assumes it's not safe.
1633 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1634 MachineBasicBlock::iterator I) {
1635 MachineBasicBlock::iterator E = MBB.end();
1637 // For compile time consideration, if we are not able to determine the
1638 // safety after visiting 4 instructions in each direction, we will assume
1640 MachineBasicBlock::iterator Iter = I;
1641 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1642 bool SeenDef = false;
1643 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1644 MachineOperand &MO = Iter->getOperand(j);
1645 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1649 if (MO.getReg() == X86::EFLAGS) {
1657 // This instruction defines EFLAGS, no need to look any further.
1660 // Skip over DBG_VALUE.
1661 while (Iter != E && Iter->isDebugValue())
1665 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1668 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1669 SE = MBB.succ_end(); SI != SE; ++SI)
1670 if ((*SI)->isLiveIn(X86::EFLAGS))
1675 MachineBasicBlock::iterator B = MBB.begin();
1677 for (unsigned i = 0; i < 4; ++i) {
1678 // If we make it to the beginning of the block, it's safe to clobber
1679 // EFLAGS iff EFLAGS is not live-in.
1681 return !MBB.isLiveIn(X86::EFLAGS);
1684 // Skip over DBG_VALUE.
1685 while (Iter != B && Iter->isDebugValue())
1688 bool SawKill = false;
1689 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1690 MachineOperand &MO = Iter->getOperand(j);
1691 // A register mask may clobber EFLAGS, but we should still look for a
1693 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1695 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1696 if (MO.isDef()) return MO.isDead();
1697 if (MO.isKill()) SawKill = true;
1702 // This instruction kills EFLAGS and doesn't redefine it, so
1703 // there's no need to look further.
1707 // Conservative answer.
1711 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1712 MachineBasicBlock::iterator I,
1713 unsigned DestReg, unsigned SubIdx,
1714 const MachineInstr *Orig,
1715 const TargetRegisterInfo &TRI) const {
1716 // MOV32r0 is implemented with a xor which clobbers condition code.
1717 // Re-materialize it as movri instructions to avoid side effects.
1718 unsigned Opc = Orig->getOpcode();
1719 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1720 DebugLoc DL = Orig->getDebugLoc();
1721 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1724 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1728 MachineInstr *NewMI = prior(I);
1729 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1732 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1733 /// is not marked dead.
1734 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1735 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1736 MachineOperand &MO = MI->getOperand(i);
1737 if (MO.isReg() && MO.isDef() &&
1738 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1745 /// getTruncatedShiftCount - check whether the shift count for a machine operand
1747 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1748 unsigned ShiftAmtOperandIdx) {
1749 // The shift count is six bits with the REX.W prefix and five bits without.
1750 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1751 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1752 return Imm & ShiftCountMask;
1755 /// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1756 /// can be represented by a LEA instruction.
1757 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1758 // Left shift instructions can be transformed into load-effective-address
1759 // instructions if we can encode them appropriately.
1760 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1761 // The SIB.scale field is two bits wide which means that we can encode any
1762 // shift amount less than 4.
1763 return ShAmt < 4 && ShAmt > 0;
1766 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1767 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1768 /// to a 32-bit superregister and then truncating back down to a 16-bit
1771 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1772 MachineFunction::iterator &MFI,
1773 MachineBasicBlock::iterator &MBBI,
1774 LiveVariables *LV) const {
1775 MachineInstr *MI = MBBI;
1776 unsigned Dest = MI->getOperand(0).getReg();
1777 unsigned Src = MI->getOperand(1).getReg();
1778 bool isDead = MI->getOperand(0).isDead();
1779 bool isKill = MI->getOperand(1).isKill();
1781 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1782 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1783 unsigned Opc, leaInReg;
1784 if (TM.getSubtarget<X86Subtarget>().is64Bit()) {
1785 Opc = X86::LEA64_32r;
1786 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1789 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1792 // Build and insert into an implicit UNDEF value. This is OK because
1793 // well be shifting and then extracting the lower 16-bits.
1794 // This has the potential to cause partial register stall. e.g.
1795 // movw (%rbp,%rcx,2), %dx
1796 // leal -65(%rdx), %esi
1797 // But testing has shown this *does* help performance in 64-bit mode (at
1798 // least on modern x86 machines).
1799 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1800 MachineInstr *InsMI =
1801 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1802 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1803 .addReg(Src, getKillRegState(isKill));
1805 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1806 get(Opc), leaOutReg);
1808 default: llvm_unreachable("Unreachable!");
1809 case X86::SHL16ri: {
1810 unsigned ShAmt = MI->getOperand(2).getImm();
1811 MIB.addReg(0).addImm(1 << ShAmt)
1812 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1816 case X86::INC64_16r:
1817 addRegOffset(MIB, leaInReg, true, 1);
1820 case X86::DEC64_16r:
1821 addRegOffset(MIB, leaInReg, true, -1);
1825 case X86::ADD16ri_DB:
1826 case X86::ADD16ri8_DB:
1827 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
1830 case X86::ADD16rr_DB: {
1831 unsigned Src2 = MI->getOperand(2).getReg();
1832 bool isKill2 = MI->getOperand(2).isKill();
1833 unsigned leaInReg2 = 0;
1834 MachineInstr *InsMI2 = 0;
1836 // ADD16rr %reg1028<kill>, %reg1028
1837 // just a single insert_subreg.
1838 addRegReg(MIB, leaInReg, true, leaInReg, false);
1840 if (TM.getSubtarget<X86Subtarget>().is64Bit())
1841 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1843 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1844 // Build and insert into an implicit UNDEF value. This is OK because
1845 // well be shifting and then extracting the lower 16-bits.
1846 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
1848 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
1849 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1850 .addReg(Src2, getKillRegState(isKill2));
1851 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1853 if (LV && isKill2 && InsMI2)
1854 LV->replaceKillInstruction(Src2, MI, InsMI2);
1859 MachineInstr *NewMI = MIB;
1860 MachineInstr *ExtMI =
1861 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1862 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1863 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
1866 // Update live variables
1867 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1868 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1870 LV->replaceKillInstruction(Src, MI, InsMI);
1872 LV->replaceKillInstruction(Dest, MI, ExtMI);
1878 /// convertToThreeAddress - This method must be implemented by targets that
1879 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1880 /// may be able to convert a two-address instruction into a true
1881 /// three-address instruction on demand. This allows the X86 target (for
1882 /// example) to convert ADD and SHL instructions into LEA instructions if they
1883 /// would require register copies due to two-addressness.
1885 /// This method returns a null pointer if the transformation cannot be
1886 /// performed, otherwise it returns the new instruction.
1889 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1890 MachineBasicBlock::iterator &MBBI,
1891 LiveVariables *LV) const {
1892 MachineInstr *MI = MBBI;
1894 // The following opcodes also sets the condition code register(s). Only
1895 // convert them to equivalent lea if the condition code register def's
1897 if (hasLiveCondCodeDef(MI))
1900 MachineFunction &MF = *MI->getParent()->getParent();
1901 // All instructions input are two-addr instructions. Get the known operands.
1902 const MachineOperand &Dest = MI->getOperand(0);
1903 const MachineOperand &Src = MI->getOperand(1);
1905 MachineInstr *NewMI = NULL;
1906 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1907 // we have better subtarget support, enable the 16-bit LEA generation here.
1908 // 16-bit LEA is also slow on Core2.
1909 bool DisableLEA16 = true;
1910 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1912 unsigned MIOpc = MI->getOpcode();
1914 case X86::SHUFPSrri: {
1915 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1916 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1918 unsigned B = MI->getOperand(1).getReg();
1919 unsigned C = MI->getOperand(2).getReg();
1920 if (B != C) return 0;
1921 unsigned M = MI->getOperand(3).getImm();
1922 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1923 .addOperand(Dest).addOperand(Src).addImm(M);
1926 case X86::SHUFPDrri: {
1927 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1928 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1930 unsigned B = MI->getOperand(1).getReg();
1931 unsigned C = MI->getOperand(2).getReg();
1932 if (B != C) return 0;
1933 unsigned M = MI->getOperand(3).getImm();
1935 // Convert to PSHUFD mask.
1936 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1938 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1939 .addOperand(Dest).addOperand(Src).addImm(M);
1942 case X86::SHL64ri: {
1943 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1944 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1945 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
1947 // LEA can't handle RSP.
1948 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1949 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1950 &X86::GR64_NOSPRegClass))
1953 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1955 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
1958 case X86::SHL32ri: {
1959 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1960 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1961 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
1963 // LEA can't handle ESP.
1964 bool isKill = Src.isKill();
1965 unsigned SrcReg = Src.getReg();
1967 unsigned NewSrc = MF.getRegInfo().createVirtualRegister(&X86::GR64_NOSPRegClass);
1968 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1969 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1974 } else if (TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1975 !MF.getRegInfo().constrainRegClass(SrcReg,
1976 &X86::GR32_NOSPRegClass)) {
1980 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1981 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1983 .addReg(0).addImm(1 << ShAmt)
1984 .addReg(SrcReg, getKillRegState(isKill)).addImm(0).addReg(0);
1987 case X86::SHL16ri: {
1988 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1989 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
1990 if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
1993 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1994 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1996 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2005 case X86::INC64_32r: {
2006 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2007 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2008 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2009 const TargetRegisterClass *RC = is64Bit ?
2010 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
2011 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
2013 // LEA can't handle RSP.
2014 bool isKill = Src.isKill();
2015 unsigned SrcReg = Src.getReg();
2016 if (Opc == X86::LEA64_32r) {
2017 unsigned NewSrc = MF.getRegInfo().createVirtualRegister(&X86::GR64_NOSPRegClass);
2018 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
2019 get(TargetOpcode::COPY))
2020 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2025 } else if (TargetRegisterInfo::isVirtualRegister(SrcReg) &&
2026 !MF.getRegInfo().constrainRegClass(SrcReg, RC))
2029 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2030 .addOperand(Dest).addReg(SrcReg, getKillRegState(isKill)), 1);
2034 case X86::INC64_16r:
2036 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2037 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2038 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2039 .addOperand(Dest).addOperand(Src), 1);
2043 case X86::DEC64_32r: {
2044 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2045 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2046 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2047 const TargetRegisterClass *RC = is64Bit ?
2048 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
2049 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
2050 // LEA can't handle RSP.
2051 bool isKill = Src.isKill();
2052 unsigned SrcReg = Src.getReg();
2053 if (Opc == X86::LEA64_32r) {
2055 MF.getRegInfo().createVirtualRegister(&X86::GR64_NOSPRegClass);
2056 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
2057 get(TargetOpcode::COPY))
2058 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2063 } else if (TargetRegisterInfo::isVirtualRegister(SrcReg) &&
2064 !MF.getRegInfo().constrainRegClass(SrcReg, RC))
2067 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2068 .addOperand(Dest).addReg(SrcReg, getKillRegState(isKill)), -1);
2072 case X86::DEC64_16r:
2074 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2075 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2076 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2077 .addOperand(Dest).addOperand(Src), -1);
2080 case X86::ADD64rr_DB:
2082 case X86::ADD32rr_DB: {
2083 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2085 const TargetRegisterClass *RC;
2086 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2088 RC = &X86::GR64_NOSPRegClass;
2090 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2091 RC = is64Bit ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
2094 unsigned SrcReg = Src.getReg();
2095 unsigned isKill = Src.isKill();
2096 const MachineOperand &Src2 = MI->getOperand(2);
2097 unsigned Src2Reg = Src2.getReg();
2098 bool isKill2 = Src2.isKill();
2100 // LEA can't handle RSP.
2101 if (Opc == X86::LEA64_32r) {
2102 MachineBasicBlock &MBB = *MI->getParent();
2104 unsigned NewSrc = MF.getRegInfo().createVirtualRegister(RC);
2105 BuildMI(MBB, MI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2106 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2111 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
2112 BuildMI(MBB, MI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2113 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2117 } else if (TargetRegisterInfo::isVirtualRegister(Src2Reg) &&
2118 !MF.getRegInfo().constrainRegClass(Src2Reg, RC))
2121 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2123 SrcReg, isKill, Src2Reg, isKill2);
2125 // Preserve undefness of the operands.
2127 bool isUndef = MI->getOperand(1).isUndef();
2128 bool isUndef2 = MI->getOperand(2).isUndef();
2129 NewMI->getOperand(1).setIsUndef(isUndef);
2130 NewMI->getOperand(3).setIsUndef(isUndef2);
2133 if (LV && Src2.isKill())
2134 LV->replaceKillInstruction(Src2Reg, MI, NewMI);
2138 case X86::ADD16rr_DB: {
2140 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2141 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2142 unsigned Src2 = MI->getOperand(2).getReg();
2143 bool isKill2 = MI->getOperand(2).isKill();
2144 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2146 Src.getReg(), Src.isKill(), Src2, isKill2);
2148 // Preserve undefness of the operands.
2149 bool isUndef = MI->getOperand(1).isUndef();
2150 bool isUndef2 = MI->getOperand(2).isUndef();
2151 NewMI->getOperand(1).setIsUndef(isUndef);
2152 NewMI->getOperand(3).setIsUndef(isUndef2);
2155 LV->replaceKillInstruction(Src2, MI, NewMI);
2158 case X86::ADD64ri32:
2160 case X86::ADD64ri32_DB:
2161 case X86::ADD64ri8_DB:
2162 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2163 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2164 .addOperand(Dest).addOperand(Src),
2165 MI->getOperand(2).getImm());
2169 case X86::ADD32ri_DB:
2170 case X86::ADD32ri8_DB: {
2171 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2174 MF.getRegInfo().createVirtualRegister(&X86::GR64_NOSPRegClass);
2175 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
2176 get(TargetOpcode::COPY))
2177 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2180 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64_32r))
2182 .addReg(NewSrc, getKillRegState(true)),
2183 MI->getOperand(2).getImm());
2185 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA32r))
2186 .addOperand(Dest).addOperand(Src),
2187 MI->getOperand(2).getImm());
2194 case X86::ADD16ri_DB:
2195 case X86::ADD16ri8_DB:
2197 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2198 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2199 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2200 .addOperand(Dest).addOperand(Src),
2201 MI->getOperand(2).getImm());
2207 if (!NewMI) return 0;
2209 if (LV) { // Update live variables
2211 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2213 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2216 MFI->insert(MBBI, NewMI); // Insert the new inst
2220 /// commuteInstruction - We have a few instructions that must be hacked on to
2224 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2225 switch (MI->getOpcode()) {
2226 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2227 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2228 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2229 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2230 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2231 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2234 switch (MI->getOpcode()) {
2235 default: llvm_unreachable("Unreachable!");
2236 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2237 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2238 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2239 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2240 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2241 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2243 unsigned Amt = MI->getOperand(3).getImm();
2245 MachineFunction &MF = *MI->getParent()->getParent();
2246 MI = MF.CloneMachineInstr(MI);
2249 MI->setDesc(get(Opc));
2250 MI->getOperand(3).setImm(Size-Amt);
2251 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2253 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2254 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2255 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2256 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2257 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2258 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2259 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2260 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2261 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2262 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2263 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2264 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2265 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2266 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2267 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2268 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2270 switch (MI->getOpcode()) {
2271 default: llvm_unreachable("Unreachable!");
2272 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2273 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2274 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2275 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2276 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2277 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2278 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2279 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2280 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2281 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2282 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2283 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2284 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2285 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2286 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2287 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2288 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2289 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2290 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2291 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2292 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2293 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2294 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2295 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2296 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2297 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2298 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2299 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2300 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2301 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2302 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2303 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2304 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2305 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2306 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2307 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2308 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2309 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2310 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2311 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2312 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2313 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2314 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2315 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2316 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2317 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2318 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2319 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2322 MachineFunction &MF = *MI->getParent()->getParent();
2323 MI = MF.CloneMachineInstr(MI);
2326 MI->setDesc(get(Opc));
2327 // Fallthrough intended.
2330 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2334 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2336 default: return X86::COND_INVALID;
2337 case X86::JE_4: return X86::COND_E;
2338 case X86::JNE_4: return X86::COND_NE;
2339 case X86::JL_4: return X86::COND_L;
2340 case X86::JLE_4: return X86::COND_LE;
2341 case X86::JG_4: return X86::COND_G;
2342 case X86::JGE_4: return X86::COND_GE;
2343 case X86::JB_4: return X86::COND_B;
2344 case X86::JBE_4: return X86::COND_BE;
2345 case X86::JA_4: return X86::COND_A;
2346 case X86::JAE_4: return X86::COND_AE;
2347 case X86::JS_4: return X86::COND_S;
2348 case X86::JNS_4: return X86::COND_NS;
2349 case X86::JP_4: return X86::COND_P;
2350 case X86::JNP_4: return X86::COND_NP;
2351 case X86::JO_4: return X86::COND_O;
2352 case X86::JNO_4: return X86::COND_NO;
2356 /// getCondFromSETOpc - return condition code of a SET opcode.
2357 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2359 default: return X86::COND_INVALID;
2360 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2361 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2362 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2363 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2364 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2365 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2366 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2367 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2368 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2369 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2370 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2371 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2372 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2373 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2374 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2375 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2379 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2380 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2382 default: return X86::COND_INVALID;
2383 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2384 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2386 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2387 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2388 return X86::COND_AE;
2389 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2390 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2392 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2393 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2394 return X86::COND_BE;
2395 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2396 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2398 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2399 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2401 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2402 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2403 return X86::COND_GE;
2404 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2405 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2407 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2408 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2409 return X86::COND_LE;
2410 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2411 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2412 return X86::COND_NE;
2413 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2414 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2415 return X86::COND_NO;
2416 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2417 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2418 return X86::COND_NP;
2419 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2420 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2421 return X86::COND_NS;
2422 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2423 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2425 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2426 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2428 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2429 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2434 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2436 default: llvm_unreachable("Illegal condition code!");
2437 case X86::COND_E: return X86::JE_4;
2438 case X86::COND_NE: return X86::JNE_4;
2439 case X86::COND_L: return X86::JL_4;
2440 case X86::COND_LE: return X86::JLE_4;
2441 case X86::COND_G: return X86::JG_4;
2442 case X86::COND_GE: return X86::JGE_4;
2443 case X86::COND_B: return X86::JB_4;
2444 case X86::COND_BE: return X86::JBE_4;
2445 case X86::COND_A: return X86::JA_4;
2446 case X86::COND_AE: return X86::JAE_4;
2447 case X86::COND_S: return X86::JS_4;
2448 case X86::COND_NS: return X86::JNS_4;
2449 case X86::COND_P: return X86::JP_4;
2450 case X86::COND_NP: return X86::JNP_4;
2451 case X86::COND_O: return X86::JO_4;
2452 case X86::COND_NO: return X86::JNO_4;
2456 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2457 /// e.g. turning COND_E to COND_NE.
2458 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2460 default: llvm_unreachable("Illegal condition code!");
2461 case X86::COND_E: return X86::COND_NE;
2462 case X86::COND_NE: return X86::COND_E;
2463 case X86::COND_L: return X86::COND_GE;
2464 case X86::COND_LE: return X86::COND_G;
2465 case X86::COND_G: return X86::COND_LE;
2466 case X86::COND_GE: return X86::COND_L;
2467 case X86::COND_B: return X86::COND_AE;
2468 case X86::COND_BE: return X86::COND_A;
2469 case X86::COND_A: return X86::COND_BE;
2470 case X86::COND_AE: return X86::COND_B;
2471 case X86::COND_S: return X86::COND_NS;
2472 case X86::COND_NS: return X86::COND_S;
2473 case X86::COND_P: return X86::COND_NP;
2474 case X86::COND_NP: return X86::COND_P;
2475 case X86::COND_O: return X86::COND_NO;
2476 case X86::COND_NO: return X86::COND_O;
2480 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2481 /// the condition code if we modify the instructions such that flags are
2483 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2485 default: return X86::COND_INVALID;
2486 case X86::COND_E: return X86::COND_E;
2487 case X86::COND_NE: return X86::COND_NE;
2488 case X86::COND_L: return X86::COND_G;
2489 case X86::COND_LE: return X86::COND_GE;
2490 case X86::COND_G: return X86::COND_L;
2491 case X86::COND_GE: return X86::COND_LE;
2492 case X86::COND_B: return X86::COND_A;
2493 case X86::COND_BE: return X86::COND_AE;
2494 case X86::COND_A: return X86::COND_B;
2495 case X86::COND_AE: return X86::COND_BE;
2499 /// getSETFromCond - Return a set opcode for the given condition and
2500 /// whether it has memory operand.
2501 static unsigned getSETFromCond(X86::CondCode CC,
2502 bool HasMemoryOperand) {
2503 static const uint16_t Opc[16][2] = {
2504 { X86::SETAr, X86::SETAm },
2505 { X86::SETAEr, X86::SETAEm },
2506 { X86::SETBr, X86::SETBm },
2507 { X86::SETBEr, X86::SETBEm },
2508 { X86::SETEr, X86::SETEm },
2509 { X86::SETGr, X86::SETGm },
2510 { X86::SETGEr, X86::SETGEm },
2511 { X86::SETLr, X86::SETLm },
2512 { X86::SETLEr, X86::SETLEm },
2513 { X86::SETNEr, X86::SETNEm },
2514 { X86::SETNOr, X86::SETNOm },
2515 { X86::SETNPr, X86::SETNPm },
2516 { X86::SETNSr, X86::SETNSm },
2517 { X86::SETOr, X86::SETOm },
2518 { X86::SETPr, X86::SETPm },
2519 { X86::SETSr, X86::SETSm }
2522 assert(CC < 16 && "Can only handle standard cond codes");
2523 return Opc[CC][HasMemoryOperand ? 1 : 0];
2526 /// getCMovFromCond - Return a cmov opcode for the given condition,
2527 /// register size in bytes, and operand type.
2528 static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2529 bool HasMemoryOperand) {
2530 static const uint16_t Opc[32][3] = {
2531 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2532 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2533 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2534 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2535 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2536 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2537 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2538 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2539 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2540 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2541 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2542 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2543 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2544 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2545 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2546 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2547 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2548 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2549 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2550 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2551 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2552 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2553 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2554 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2555 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2556 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2557 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2558 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2559 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2560 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2561 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2562 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2565 assert(CC < 16 && "Can only handle standard cond codes");
2566 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2568 default: llvm_unreachable("Illegal register size!");
2569 case 2: return Opc[Idx][0];
2570 case 4: return Opc[Idx][1];
2571 case 8: return Opc[Idx][2];
2575 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2576 if (!MI->isTerminator()) return false;
2578 // Conditional branch is a special case.
2579 if (MI->isBranch() && !MI->isBarrier())
2581 if (!MI->isPredicable())
2583 return !isPredicated(MI);
2586 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2587 MachineBasicBlock *&TBB,
2588 MachineBasicBlock *&FBB,
2589 SmallVectorImpl<MachineOperand> &Cond,
2590 bool AllowModify) const {
2591 // Start from the bottom of the block and work up, examining the
2592 // terminator instructions.
2593 MachineBasicBlock::iterator I = MBB.end();
2594 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2595 while (I != MBB.begin()) {
2597 if (I->isDebugValue())
2600 // Working from the bottom, when we see a non-terminator instruction, we're
2602 if (!isUnpredicatedTerminator(I))
2605 // A terminator that isn't a branch can't easily be handled by this
2610 // Handle unconditional branches.
2611 if (I->getOpcode() == X86::JMP_4) {
2615 TBB = I->getOperand(0).getMBB();
2619 // If the block has any instructions after a JMP, delete them.
2620 while (llvm::next(I) != MBB.end())
2621 llvm::next(I)->eraseFromParent();
2626 // Delete the JMP if it's equivalent to a fall-through.
2627 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2629 I->eraseFromParent();
2631 UnCondBrIter = MBB.end();
2635 // TBB is used to indicate the unconditional destination.
2636 TBB = I->getOperand(0).getMBB();
2640 // Handle conditional branches.
2641 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
2642 if (BranchCode == X86::COND_INVALID)
2643 return true; // Can't handle indirect branch.
2645 // Working from the bottom, handle the first conditional branch.
2647 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2648 if (AllowModify && UnCondBrIter != MBB.end() &&
2649 MBB.isLayoutSuccessor(TargetBB)) {
2650 // If we can modify the code and it ends in something like:
2658 // Then we can change this to:
2665 // Which is a bit more efficient.
2666 // We conditionally jump to the fall-through block.
2667 BranchCode = GetOppositeBranchCondition(BranchCode);
2668 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2669 MachineBasicBlock::iterator OldInst = I;
2671 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2672 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2673 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2676 OldInst->eraseFromParent();
2677 UnCondBrIter->eraseFromParent();
2679 // Restart the analysis.
2680 UnCondBrIter = MBB.end();
2686 TBB = I->getOperand(0).getMBB();
2687 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2691 // Handle subsequent conditional branches. Only handle the case where all
2692 // conditional branches branch to the same destination and their condition
2693 // opcodes fit one of the special multi-branch idioms.
2694 assert(Cond.size() == 1);
2697 // Only handle the case where all conditional branches branch to the same
2699 if (TBB != I->getOperand(0).getMBB())
2702 // If the conditions are the same, we can leave them alone.
2703 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2704 if (OldBranchCode == BranchCode)
2707 // If they differ, see if they fit one of the known patterns. Theoretically,
2708 // we could handle more patterns here, but we shouldn't expect to see them
2709 // if instruction selection has done a reasonable job.
2710 if ((OldBranchCode == X86::COND_NP &&
2711 BranchCode == X86::COND_E) ||
2712 (OldBranchCode == X86::COND_E &&
2713 BranchCode == X86::COND_NP))
2714 BranchCode = X86::COND_NP_OR_E;
2715 else if ((OldBranchCode == X86::COND_P &&
2716 BranchCode == X86::COND_NE) ||
2717 (OldBranchCode == X86::COND_NE &&
2718 BranchCode == X86::COND_P))
2719 BranchCode = X86::COND_NE_OR_P;
2723 // Update the MachineOperand.
2724 Cond[0].setImm(BranchCode);
2730 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2731 MachineBasicBlock::iterator I = MBB.end();
2734 while (I != MBB.begin()) {
2736 if (I->isDebugValue())
2738 if (I->getOpcode() != X86::JMP_4 &&
2739 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2741 // Remove the branch.
2742 I->eraseFromParent();
2751 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2752 MachineBasicBlock *FBB,
2753 const SmallVectorImpl<MachineOperand> &Cond,
2754 DebugLoc DL) const {
2755 // Shouldn't be a fall through.
2756 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2757 assert((Cond.size() == 1 || Cond.size() == 0) &&
2758 "X86 branch conditions have one component!");
2761 // Unconditional branch?
2762 assert(!FBB && "Unconditional branch with multiple successors!");
2763 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2767 // Conditional branch.
2769 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2771 case X86::COND_NP_OR_E:
2772 // Synthesize NP_OR_E with two branches.
2773 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2775 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2778 case X86::COND_NE_OR_P:
2779 // Synthesize NE_OR_P with two branches.
2780 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2782 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2786 unsigned Opc = GetCondBranchFromCond(CC);
2787 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2792 // Two-way Conditional branch. Insert the second branch.
2793 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2800 canInsertSelect(const MachineBasicBlock &MBB,
2801 const SmallVectorImpl<MachineOperand> &Cond,
2802 unsigned TrueReg, unsigned FalseReg,
2803 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2804 // Not all subtargets have cmov instructions.
2805 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2807 if (Cond.size() != 1)
2809 // We cannot do the composite conditions, at least not in SSA form.
2810 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2813 // Check register classes.
2814 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2815 const TargetRegisterClass *RC =
2816 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2820 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2821 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2822 X86::GR32RegClass.hasSubClassEq(RC) ||
2823 X86::GR64RegClass.hasSubClassEq(RC)) {
2824 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2825 // Bridge. Probably Ivy Bridge as well.
2832 // Can't do vectors.
2836 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2837 MachineBasicBlock::iterator I, DebugLoc DL,
2839 const SmallVectorImpl<MachineOperand> &Cond,
2840 unsigned TrueReg, unsigned FalseReg) const {
2841 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2842 assert(Cond.size() == 1 && "Invalid Cond array");
2843 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2844 MRI.getRegClass(DstReg)->getSize(),
2845 false/*HasMemoryOperand*/);
2846 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2849 /// isHReg - Test if the given register is a physical h register.
2850 static bool isHReg(unsigned Reg) {
2851 return X86::GR8_ABCD_HRegClass.contains(Reg);
2854 // Try and copy between VR128/VR64 and GR64 registers.
2855 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2857 // SrcReg(VR128) -> DestReg(GR64)
2858 // SrcReg(VR64) -> DestReg(GR64)
2859 // SrcReg(GR64) -> DestReg(VR128)
2860 // SrcReg(GR64) -> DestReg(VR64)
2862 if (X86::GR64RegClass.contains(DestReg)) {
2863 if (X86::VR128RegClass.contains(SrcReg))
2864 // Copy from a VR128 register to a GR64 register.
2865 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
2866 if (X86::VR64RegClass.contains(SrcReg))
2867 // Copy from a VR64 register to a GR64 register.
2868 return X86::MOVSDto64rr;
2869 } else if (X86::GR64RegClass.contains(SrcReg)) {
2870 // Copy from a GR64 register to a VR128 register.
2871 if (X86::VR128RegClass.contains(DestReg))
2872 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
2873 // Copy from a GR64 register to a VR64 register.
2874 if (X86::VR64RegClass.contains(DestReg))
2875 return X86::MOV64toSDrr;
2878 // SrcReg(FR32) -> DestReg(GR32)
2879 // SrcReg(GR32) -> DestReg(FR32)
2881 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
2882 // Copy from a FR32 register to a GR32 register.
2883 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
2885 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
2886 // Copy from a GR32 register to a FR32 register.
2887 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
2892 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2893 MachineBasicBlock::iterator MI, DebugLoc DL,
2894 unsigned DestReg, unsigned SrcReg,
2895 bool KillSrc) const {
2896 // First deal with the normal symmetric copies.
2897 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2899 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2901 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2903 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2905 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2906 // Copying to or from a physical H register on x86-64 requires a NOREX
2907 // move. Otherwise use a normal move.
2908 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
2909 TM.getSubtarget<X86Subtarget>().is64Bit()) {
2910 Opc = X86::MOV8rr_NOREX;
2911 // Both operands must be encodable without an REX prefix.
2912 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2913 "8-bit H register can not be copied outside GR8_NOREX");
2916 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
2917 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
2918 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2919 Opc = X86::VMOVAPSYrr;
2920 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2921 Opc = X86::MMX_MOVQ64rr;
2923 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
2926 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2927 .addReg(SrcReg, getKillRegState(KillSrc));
2931 // Moving EFLAGS to / from another register requires a push and a pop.
2932 // Notice that we have to adjust the stack if we don't want to clobber the
2933 // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
2934 if (SrcReg == X86::EFLAGS) {
2935 if (X86::GR64RegClass.contains(DestReg)) {
2936 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2937 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2940 if (X86::GR32RegClass.contains(DestReg)) {
2941 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2942 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2946 if (DestReg == X86::EFLAGS) {
2947 if (X86::GR64RegClass.contains(SrcReg)) {
2948 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2949 .addReg(SrcReg, getKillRegState(KillSrc));
2950 BuildMI(MBB, MI, DL, get(X86::POPF64));
2953 if (X86::GR32RegClass.contains(SrcReg)) {
2954 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2955 .addReg(SrcReg, getKillRegState(KillSrc));
2956 BuildMI(MBB, MI, DL, get(X86::POPF32));
2961 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2962 << " to " << RI.getName(DestReg) << '\n');
2963 llvm_unreachable("Cannot emit physreg copy instruction");
2966 static unsigned getLoadStoreRegOpcode(unsigned Reg,
2967 const TargetRegisterClass *RC,
2968 bool isStackAligned,
2969 const TargetMachine &TM,
2971 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2972 switch (RC->getSize()) {
2974 llvm_unreachable("Unknown spill size");
2976 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
2977 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2978 // Copying to or from a physical H register on x86-64 requires a NOREX
2979 // move. Otherwise use a normal move.
2980 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2981 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2982 return load ? X86::MOV8rm : X86::MOV8mr;
2984 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2985 return load ? X86::MOV16rm : X86::MOV16mr;
2987 if (X86::GR32RegClass.hasSubClassEq(RC))
2988 return load ? X86::MOV32rm : X86::MOV32mr;
2989 if (X86::FR32RegClass.hasSubClassEq(RC))
2991 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2992 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
2993 if (X86::RFP32RegClass.hasSubClassEq(RC))
2994 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2995 llvm_unreachable("Unknown 4-byte regclass");
2997 if (X86::GR64RegClass.hasSubClassEq(RC))
2998 return load ? X86::MOV64rm : X86::MOV64mr;
2999 if (X86::FR64RegClass.hasSubClassEq(RC))
3001 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3002 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3003 if (X86::VR64RegClass.hasSubClassEq(RC))
3004 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3005 if (X86::RFP64RegClass.hasSubClassEq(RC))
3006 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3007 llvm_unreachable("Unknown 8-byte regclass");
3009 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3010 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3012 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
3013 // If stack is realigned we can use aligned stores.
3016 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3017 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
3020 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3021 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3024 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3025 // If stack is realigned we can use aligned stores.
3027 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3029 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
3033 static unsigned getStoreRegOpcode(unsigned SrcReg,
3034 const TargetRegisterClass *RC,
3035 bool isStackAligned,
3036 TargetMachine &TM) {
3037 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
3041 static unsigned getLoadRegOpcode(unsigned DestReg,
3042 const TargetRegisterClass *RC,
3043 bool isStackAligned,
3044 const TargetMachine &TM) {
3045 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
3048 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3049 MachineBasicBlock::iterator MI,
3050 unsigned SrcReg, bool isKill, int FrameIdx,
3051 const TargetRegisterClass *RC,
3052 const TargetRegisterInfo *TRI) const {
3053 const MachineFunction &MF = *MBB.getParent();
3054 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3055 "Stack slot too small for store");
3056 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3057 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3058 RI.canRealignStack(MF);
3059 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3060 DebugLoc DL = MBB.findDebugLoc(MI);
3061 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
3062 .addReg(SrcReg, getKillRegState(isKill));
3065 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3067 SmallVectorImpl<MachineOperand> &Addr,
3068 const TargetRegisterClass *RC,
3069 MachineInstr::mmo_iterator MMOBegin,
3070 MachineInstr::mmo_iterator MMOEnd,
3071 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3072 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3073 bool isAligned = MMOBegin != MMOEnd &&
3074 (*MMOBegin)->getAlignment() >= Alignment;
3075 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3077 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3078 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3079 MIB.addOperand(Addr[i]);
3080 MIB.addReg(SrcReg, getKillRegState(isKill));
3081 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3082 NewMIs.push_back(MIB);
3086 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3087 MachineBasicBlock::iterator MI,
3088 unsigned DestReg, int FrameIdx,
3089 const TargetRegisterClass *RC,
3090 const TargetRegisterInfo *TRI) const {
3091 const MachineFunction &MF = *MBB.getParent();
3092 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3093 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3094 RI.canRealignStack(MF);
3095 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3096 DebugLoc DL = MBB.findDebugLoc(MI);
3097 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
3100 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
3101 SmallVectorImpl<MachineOperand> &Addr,
3102 const TargetRegisterClass *RC,
3103 MachineInstr::mmo_iterator MMOBegin,
3104 MachineInstr::mmo_iterator MMOEnd,
3105 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3106 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3107 bool isAligned = MMOBegin != MMOEnd &&
3108 (*MMOBegin)->getAlignment() >= Alignment;
3109 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3111 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3112 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3113 MIB.addOperand(Addr[i]);
3114 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3115 NewMIs.push_back(MIB);
3119 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3120 int &CmpMask, int &CmpValue) const {
3121 switch (MI->getOpcode()) {
3123 case X86::CMP64ri32:
3130 SrcReg = MI->getOperand(0).getReg();
3133 CmpValue = MI->getOperand(1).getImm();
3135 // A SUB can be used to perform comparison.
3140 SrcReg = MI->getOperand(1).getReg();
3149 SrcReg = MI->getOperand(1).getReg();
3150 SrcReg2 = MI->getOperand(2).getReg();
3154 case X86::SUB64ri32:
3161 SrcReg = MI->getOperand(1).getReg();
3164 CmpValue = MI->getOperand(2).getImm();
3170 SrcReg = MI->getOperand(0).getReg();
3171 SrcReg2 = MI->getOperand(1).getReg();
3179 SrcReg = MI->getOperand(0).getReg();
3180 if (MI->getOperand(1).getReg() != SrcReg) return false;
3181 // Compare against zero.
3190 /// isRedundantFlagInstr - check whether the first instruction, whose only
3191 /// purpose is to update flags, can be made redundant.
3192 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3193 /// This function can be extended later on.
3194 /// SrcReg, SrcRegs: register operands for FlagI.
3195 /// ImmValue: immediate for FlagI if it takes an immediate.
3196 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3197 unsigned SrcReg2, int ImmValue,
3199 if (((FlagI->getOpcode() == X86::CMP64rr &&
3200 OI->getOpcode() == X86::SUB64rr) ||
3201 (FlagI->getOpcode() == X86::CMP32rr &&
3202 OI->getOpcode() == X86::SUB32rr)||
3203 (FlagI->getOpcode() == X86::CMP16rr &&
3204 OI->getOpcode() == X86::SUB16rr)||
3205 (FlagI->getOpcode() == X86::CMP8rr &&
3206 OI->getOpcode() == X86::SUB8rr)) &&
3207 ((OI->getOperand(1).getReg() == SrcReg &&
3208 OI->getOperand(2).getReg() == SrcReg2) ||
3209 (OI->getOperand(1).getReg() == SrcReg2 &&
3210 OI->getOperand(2).getReg() == SrcReg)))
3213 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3214 OI->getOpcode() == X86::SUB64ri32) ||
3215 (FlagI->getOpcode() == X86::CMP64ri8 &&
3216 OI->getOpcode() == X86::SUB64ri8) ||
3217 (FlagI->getOpcode() == X86::CMP32ri &&
3218 OI->getOpcode() == X86::SUB32ri) ||
3219 (FlagI->getOpcode() == X86::CMP32ri8 &&
3220 OI->getOpcode() == X86::SUB32ri8) ||
3221 (FlagI->getOpcode() == X86::CMP16ri &&
3222 OI->getOpcode() == X86::SUB16ri) ||
3223 (FlagI->getOpcode() == X86::CMP16ri8 &&
3224 OI->getOpcode() == X86::SUB16ri8) ||
3225 (FlagI->getOpcode() == X86::CMP8ri &&
3226 OI->getOpcode() == X86::SUB8ri)) &&
3227 OI->getOperand(1).getReg() == SrcReg &&
3228 OI->getOperand(2).getImm() == ImmValue)
3233 /// isDefConvertible - check whether the definition can be converted
3234 /// to remove a comparison against zero.
3235 inline static bool isDefConvertible(MachineInstr *MI) {
3236 switch (MI->getOpcode()) {
3237 default: return false;
3239 // The shift instructions only modify ZF if their shift count is non-zero.
3240 // N.B.: The processor truncates the shift count depending on the encoding.
3241 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3242 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3243 return getTruncatedShiftCount(MI, 2) != 0;
3245 // Some left shift instructions can be turned into LEA instructions but only
3246 // if their flags aren't used. Avoid transforming such instructions.
3247 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3248 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3249 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3253 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3254 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3255 return getTruncatedShiftCount(MI, 3) != 0;
3257 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3258 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3259 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3260 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3261 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3262 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3263 case X86::DEC64_32r: case X86::DEC64_16r:
3264 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3265 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3266 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3267 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3268 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3269 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3270 case X86::INC64_32r: case X86::INC64_16r:
3271 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3272 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3273 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3274 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3275 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3276 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3277 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3278 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3279 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3280 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3281 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3282 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3283 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3284 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3285 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3286 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3287 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3288 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3289 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3290 case X86::ADC32ri: case X86::ADC32ri8:
3291 case X86::ADC32rr: case X86::ADC64ri32:
3292 case X86::ADC64ri8: case X86::ADC64rr:
3293 case X86::SBB32ri: case X86::SBB32ri8:
3294 case X86::SBB32rr: case X86::SBB64ri32:
3295 case X86::SBB64ri8: case X86::SBB64rr:
3296 case X86::ANDN32rr: case X86::ANDN32rm:
3297 case X86::ANDN64rr: case X86::ANDN64rm:
3298 case X86::BEXTR32rr: case X86::BEXTR64rr:
3299 case X86::BEXTR32rm: case X86::BEXTR64rm:
3300 case X86::BLSI32rr: case X86::BLSI32rm:
3301 case X86::BLSI64rr: case X86::BLSI64rm:
3302 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3303 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3304 case X86::BLSR32rr: case X86::BLSR32rm:
3305 case X86::BLSR64rr: case X86::BLSR64rm:
3306 case X86::BZHI32rr: case X86::BZHI32rm:
3307 case X86::BZHI64rr: case X86::BZHI64rm:
3308 case X86::LZCNT16rr: case X86::LZCNT16rm:
3309 case X86::LZCNT32rr: case X86::LZCNT32rm:
3310 case X86::LZCNT64rr: case X86::LZCNT64rm:
3311 case X86::POPCNT16rr:case X86::POPCNT16rm:
3312 case X86::POPCNT32rr:case X86::POPCNT32rm:
3313 case X86::POPCNT64rr:case X86::POPCNT64rm:
3314 case X86::TZCNT16rr: case X86::TZCNT16rm:
3315 case X86::TZCNT32rr: case X86::TZCNT32rm:
3316 case X86::TZCNT64rr: case X86::TZCNT64rm:
3321 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3322 /// operates on the same source operands and sets flags in the same way as
3323 /// Compare; remove Compare if possible.
3325 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3326 int CmpMask, int CmpValue,
3327 const MachineRegisterInfo *MRI) const {
3328 // Check whether we can replace SUB with CMP.
3329 unsigned NewOpcode = 0;
3330 switch (CmpInstr->getOpcode()) {
3332 case X86::SUB64ri32:
3347 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3349 // There is no use of the destination register, we can replace SUB with CMP.
3350 switch (CmpInstr->getOpcode()) {
3351 default: llvm_unreachable("Unreachable!");
3352 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3353 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3354 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3355 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3356 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3357 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3358 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3359 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3360 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3361 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3362 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3363 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3364 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3365 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3366 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3368 CmpInstr->setDesc(get(NewOpcode));
3369 CmpInstr->RemoveOperand(0);
3370 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3371 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3372 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3377 // Get the unique definition of SrcReg.
3378 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3379 if (!MI) return false;
3381 // CmpInstr is the first instruction of the BB.
3382 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3384 // If we are comparing against zero, check whether we can use MI to update
3385 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3386 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3387 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3388 !isDefConvertible(MI)))
3391 // We are searching for an earlier instruction that can make CmpInstr
3392 // redundant and that instruction will be saved in Sub.
3393 MachineInstr *Sub = NULL;
3394 const TargetRegisterInfo *TRI = &getRegisterInfo();
3396 // We iterate backward, starting from the instruction before CmpInstr and
3397 // stop when reaching the definition of a source register or done with the BB.
3398 // RI points to the instruction before CmpInstr.
3399 // If the definition is in this basic block, RE points to the definition;
3400 // otherwise, RE is the rend of the basic block.
3401 MachineBasicBlock::reverse_iterator
3402 RI = MachineBasicBlock::reverse_iterator(I),
3403 RE = CmpInstr->getParent() == MI->getParent() ?
3404 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3405 CmpInstr->getParent()->rend();
3406 MachineInstr *Movr0Inst = 0;
3407 for (; RI != RE; ++RI) {
3408 MachineInstr *Instr = &*RI;
3409 // Check whether CmpInstr can be made redundant by the current instruction.
3411 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3416 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3417 Instr->readsRegister(X86::EFLAGS, TRI)) {
3418 // This instruction modifies or uses EFLAGS.
3420 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3421 // They are safe to move up, if the definition to EFLAGS is dead and
3422 // earlier instructions do not read or write EFLAGS.
3423 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
3424 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3429 // We can't remove CmpInstr.
3434 // Return false if no candidates exist.
3435 if (!IsCmpZero && !Sub)
3438 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3439 Sub->getOperand(2).getReg() == SrcReg);
3441 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3442 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3443 // If we are done with the basic block, we need to check whether EFLAGS is
3445 bool IsSafe = false;
3446 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3447 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3448 for (++I; I != E; ++I) {
3449 const MachineInstr &Instr = *I;
3450 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3451 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3452 // We should check the usage if this instruction uses and updates EFLAGS.
3453 if (!UseEFLAGS && ModifyEFLAGS) {
3454 // It is safe to remove CmpInstr if EFLAGS is updated again.
3458 if (!UseEFLAGS && !ModifyEFLAGS)
3461 // EFLAGS is used by this instruction.
3462 X86::CondCode OldCC;
3463 bool OpcIsSET = false;
3464 if (IsCmpZero || IsSwapped) {
3465 // We decode the condition code from opcode.
3466 if (Instr.isBranch())
3467 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3469 OldCC = getCondFromSETOpc(Instr.getOpcode());
3470 if (OldCC != X86::COND_INVALID)
3473 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3475 if (OldCC == X86::COND_INVALID) return false;
3480 case X86::COND_A: case X86::COND_AE:
3481 case X86::COND_B: case X86::COND_BE:
3482 case X86::COND_G: case X86::COND_GE:
3483 case X86::COND_L: case X86::COND_LE:
3484 case X86::COND_O: case X86::COND_NO:
3485 // CF and OF are used, we can't perform this optimization.
3488 } else if (IsSwapped) {
3489 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3490 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3491 // We swap the condition code and synthesize the new opcode.
3492 X86::CondCode NewCC = getSwappedCondition(OldCC);
3493 if (NewCC == X86::COND_INVALID) return false;
3495 // Synthesize the new opcode.
3496 bool HasMemoryOperand = Instr.hasOneMemOperand();
3498 if (Instr.isBranch())
3499 NewOpc = GetCondBranchFromCond(NewCC);
3501 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3503 unsigned DstReg = Instr.getOperand(0).getReg();
3504 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3508 // Push the MachineInstr to OpsToUpdate.
3509 // If it is safe to remove CmpInstr, the condition code of these
3510 // instructions will be modified.
3511 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3513 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3514 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3520 // If EFLAGS is not killed nor re-defined, we should check whether it is
3521 // live-out. If it is live-out, do not optimize.
3522 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3523 MachineBasicBlock *MBB = CmpInstr->getParent();
3524 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3525 SE = MBB->succ_end(); SI != SE; ++SI)
3526 if ((*SI)->isLiveIn(X86::EFLAGS))
3530 // The instruction to be updated is either Sub or MI.
3531 Sub = IsCmpZero ? MI : Sub;
3532 // Move Movr0Inst to the appropriate place before Sub.
3534 // Look backwards until we find a def that doesn't use the current EFLAGS.
3536 MachineBasicBlock::reverse_iterator
3537 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3538 InsertE = Sub->getParent()->rend();
3539 for (; InsertI != InsertE; ++InsertI) {
3540 MachineInstr *Instr = &*InsertI;
3541 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3542 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3543 Sub->getParent()->remove(Movr0Inst);
3544 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3549 if (InsertI == InsertE)
3553 // Make sure Sub instruction defines EFLAGS and mark the def live.
3554 unsigned i = 0, e = Sub->getNumOperands();
3555 for (; i != e; ++i) {
3556 MachineOperand &MO = Sub->getOperand(i);
3557 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3558 MO.setIsDead(false);
3562 assert(i != e && "Unable to locate a def EFLAGS operand");
3564 CmpInstr->eraseFromParent();
3566 // Modify the condition code of instructions in OpsToUpdate.
3567 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3568 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3572 /// optimizeLoadInstr - Try to remove the load by folding it to a register
3573 /// operand at the use. We fold the load instructions if load defines a virtual
3574 /// register, the virtual register is used once in the same BB, and the
3575 /// instructions in-between do not load or store, and have no side effects.
3576 MachineInstr* X86InstrInfo::
3577 optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3578 unsigned &FoldAsLoadDefReg,
3579 MachineInstr *&DefMI) const {
3580 if (FoldAsLoadDefReg == 0)
3582 // To be conservative, if there exists another load, clear the load candidate.
3583 if (MI->mayLoad()) {
3584 FoldAsLoadDefReg = 0;
3588 // Check whether we can move DefMI here.
3589 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3591 bool SawStore = false;
3592 if (!DefMI->isSafeToMove(this, 0, SawStore))
3595 // We try to commute MI if possible.
3596 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3597 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3598 // Collect information about virtual register operands of MI.
3599 unsigned SrcOperandId = 0;
3600 bool FoundSrcOperand = false;
3601 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3602 MachineOperand &MO = MI->getOperand(i);
3605 unsigned Reg = MO.getReg();
3606 if (Reg != FoldAsLoadDefReg)
3608 // Do not fold if we have a subreg use or a def or multiple uses.
3609 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3613 FoundSrcOperand = true;
3615 if (!FoundSrcOperand) return 0;
3617 // Check whether we can fold the def into SrcOperandId.
3618 SmallVector<unsigned, 8> Ops;
3619 Ops.push_back(SrcOperandId);
3620 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3622 FoldAsLoadDefReg = 0;
3627 // MI was changed but it didn't help, commute it back!
3628 commuteInstruction(MI, false);
3632 // Check whether we can commute MI and enable folding.
3633 if (MI->isCommutable()) {
3634 MachineInstr *NewMI = commuteInstruction(MI, false);
3635 // Unable to commute.
3636 if (!NewMI) return 0;
3638 // New instruction. It doesn't need to be kept.
3639 NewMI->eraseFromParent();
3647 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3648 /// instruction with two undef reads of the register being defined. This is
3649 /// used for mapping:
3652 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3654 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3655 const MCInstrDesc &Desc) {
3656 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3657 unsigned Reg = MIB->getOperand(0).getReg();
3660 // MachineInstr::addOperand() will insert explicit operands before any
3661 // implicit operands.
3662 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3663 // But we don't trust that.
3664 assert(MIB->getOperand(1).getReg() == Reg &&
3665 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3669 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3670 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3671 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3672 switch (MI->getOpcode()) {
3674 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
3675 case X86::SETB_C16r:
3676 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
3677 case X86::SETB_C32r:
3678 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
3679 case X86::SETB_C64r:
3680 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
3684 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
3686 assert(HasAVX && "AVX not supported");
3687 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
3688 case X86::V_SETALLONES:
3689 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3690 case X86::AVX2_SETALLONES:
3691 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
3692 case X86::TEST8ri_NOREX:
3693 MI->setDesc(get(X86::TEST8ri));
3700 X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
3701 int FrameIx, uint64_t Offset,
3702 const MDNode *MDPtr,
3703 DebugLoc DL) const {
3705 AM.BaseType = X86AddressMode::FrameIndexBase;
3706 AM.Base.FrameIndex = FrameIx;
3707 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3708 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3712 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
3713 const SmallVectorImpl<MachineOperand> &MOs,
3715 const TargetInstrInfo &TII) {
3716 // Create the base instruction with the memory operand as the first part.
3717 // Omit the implicit operands, something BuildMI can't do.
3718 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3719 MI->getDebugLoc(), true);
3720 MachineInstrBuilder MIB(MF, NewMI);
3721 unsigned NumAddrOps = MOs.size();
3722 for (unsigned i = 0; i != NumAddrOps; ++i)
3723 MIB.addOperand(MOs[i]);
3724 if (NumAddrOps < 4) // FrameIndex only
3727 // Loop over the rest of the ri operands, converting them over.
3728 unsigned NumOps = MI->getDesc().getNumOperands()-2;
3729 for (unsigned i = 0; i != NumOps; ++i) {
3730 MachineOperand &MO = MI->getOperand(i+2);
3733 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3734 MachineOperand &MO = MI->getOperand(i);
3740 static MachineInstr *FuseInst(MachineFunction &MF,
3741 unsigned Opcode, unsigned OpNo,
3742 const SmallVectorImpl<MachineOperand> &MOs,
3743 MachineInstr *MI, const TargetInstrInfo &TII) {
3744 // Omit the implicit operands, something BuildMI can't do.
3745 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3746 MI->getDebugLoc(), true);
3747 MachineInstrBuilder MIB(MF, NewMI);
3749 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3750 MachineOperand &MO = MI->getOperand(i);
3752 assert(MO.isReg() && "Expected to fold into reg operand!");
3753 unsigned NumAddrOps = MOs.size();
3754 for (unsigned i = 0; i != NumAddrOps; ++i)
3755 MIB.addOperand(MOs[i]);
3756 if (NumAddrOps < 4) // FrameIndex only
3765 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
3766 const SmallVectorImpl<MachineOperand> &MOs,
3768 MachineFunction &MF = *MI->getParent()->getParent();
3769 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
3771 unsigned NumAddrOps = MOs.size();
3772 for (unsigned i = 0; i != NumAddrOps; ++i)
3773 MIB.addOperand(MOs[i]);
3774 if (NumAddrOps < 4) // FrameIndex only
3776 return MIB.addImm(0);
3780 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3781 MachineInstr *MI, unsigned i,
3782 const SmallVectorImpl<MachineOperand> &MOs,
3783 unsigned Size, unsigned Align) const {
3784 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
3785 bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
3786 bool isTwoAddrFold = false;
3788 // Atom favors register form of call. So, we do not fold loads into calls
3789 // when X86Subtarget is Atom.
3790 if (isCallRegIndirect &&
3791 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
3795 unsigned NumOps = MI->getDesc().getNumOperands();
3796 bool isTwoAddr = NumOps > 1 &&
3797 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
3799 // FIXME: AsmPrinter doesn't know how to handle
3800 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3801 if (MI->getOpcode() == X86::ADD32ri &&
3802 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3805 MachineInstr *NewMI = NULL;
3806 // Folding a memory location into the two-address part of a two-address
3807 // instruction is different than folding it other places. It requires
3808 // replacing the *two* registers with the memory location.
3809 if (isTwoAddr && NumOps >= 2 && i < 2 &&
3810 MI->getOperand(0).isReg() &&
3811 MI->getOperand(1).isReg() &&
3812 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
3813 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3814 isTwoAddrFold = true;
3815 } else if (i == 0) { // If operand 0
3816 if (MI->getOpcode() == X86::MOV32r0) {
3817 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
3822 OpcodeTablePtr = &RegOp2MemOpTable0;
3823 } else if (i == 1) {
3824 OpcodeTablePtr = &RegOp2MemOpTable1;
3825 } else if (i == 2) {
3826 OpcodeTablePtr = &RegOp2MemOpTable2;
3827 } else if (i == 3) {
3828 OpcodeTablePtr = &RegOp2MemOpTable3;
3831 // If table selected...
3832 if (OpcodeTablePtr) {
3833 // Find the Opcode to fuse
3834 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3835 OpcodeTablePtr->find(MI->getOpcode());
3836 if (I != OpcodeTablePtr->end()) {
3837 unsigned Opcode = I->second.first;
3838 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
3839 if (Align < MinAlign)
3841 bool NarrowToMOV32rm = false;
3843 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
3844 if (Size < RCSize) {
3845 // Check if it's safe to fold the load. If the size of the object is
3846 // narrower than the load width, then it's not.
3847 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3849 // If this is a 64-bit load, but the spill slot is 32, then we can do
3850 // a 32-bit load which is implicitly zero-extended. This likely is due
3851 // to liveintervalanalysis remat'ing a load from stack slot.
3852 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3854 Opcode = X86::MOV32rm;
3855 NarrowToMOV32rm = true;
3860 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
3862 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
3864 if (NarrowToMOV32rm) {
3865 // If this is the special case where we use a MOV32rm to load a 32-bit
3866 // value and zero-extend the top bits. Change the destination register
3868 unsigned DstReg = NewMI->getOperand(0).getReg();
3869 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3870 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
3873 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
3880 if (PrintFailedFusing && !MI->isCopy())
3881 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
3885 /// hasPartialRegUpdate - Return true for all instructions that only update
3886 /// the first 32 or 64-bits of the destination register and leave the rest
3887 /// unmodified. This can be used to avoid folding loads if the instructions
3888 /// only update part of the destination register, and the non-updated part is
3889 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3890 /// instructions breaks the partial register dependency and it can improve
3891 /// performance. e.g.:
3893 /// movss (%rdi), %xmm0
3894 /// cvtss2sd %xmm0, %xmm0
3897 /// cvtss2sd (%rdi), %xmm0
3899 /// FIXME: This should be turned into a TSFlags.
3901 static bool hasPartialRegUpdate(unsigned Opcode) {
3903 case X86::CVTSI2SSrr:
3904 case X86::CVTSI2SS64rr:
3905 case X86::CVTSI2SDrr:
3906 case X86::CVTSI2SD64rr:
3907 case X86::CVTSD2SSrr:
3908 case X86::Int_CVTSD2SSrr:
3909 case X86::CVTSS2SDrr:
3910 case X86::Int_CVTSS2SDrr:
3912 case X86::RCPSSr_Int:
3914 case X86::ROUNDSDr_Int:
3916 case X86::ROUNDSSr_Int:
3918 case X86::RSQRTSSr_Int:
3920 case X86::SQRTSSr_Int:
3921 // AVX encoded versions
3922 case X86::VCVTSD2SSrr:
3923 case X86::Int_VCVTSD2SSrr:
3924 case X86::VCVTSS2SDrr:
3925 case X86::Int_VCVTSS2SDrr:
3927 case X86::VROUNDSDr:
3928 case X86::VROUNDSDr_Int:
3929 case X86::VROUNDSSr:
3930 case X86::VROUNDSSr_Int:
3931 case X86::VRSQRTSSr:
3939 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3940 /// instructions we would like before a partial register update.
3941 unsigned X86InstrInfo::
3942 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3943 const TargetRegisterInfo *TRI) const {
3944 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3947 // If MI is marked as reading Reg, the partial register update is wanted.
3948 const MachineOperand &MO = MI->getOperand(0);
3949 unsigned Reg = MO.getReg();
3950 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3951 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3954 if (MI->readsRegister(Reg, TRI))
3958 // If any of the preceding 16 instructions are reading Reg, insert a
3959 // dependency breaking instruction. The magic number is based on a few
3960 // Nehalem experiments.
3965 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3966 const TargetRegisterInfo *TRI) const {
3967 unsigned Reg = MI->getOperand(OpNum).getReg();
3968 if (X86::VR128RegClass.contains(Reg)) {
3969 // These instructions are all floating point domain, so xorps is the best
3971 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3972 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3973 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3974 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3975 } else if (X86::VR256RegClass.contains(Reg)) {
3976 // Use vxorps to clear the full ymm register.
3977 // It wants to read and write the xmm sub-register.
3978 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3979 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3980 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3981 .addReg(Reg, RegState::ImplicitDefine);
3984 MI->addRegisterKilled(Reg, TRI, true);
3987 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3989 const SmallVectorImpl<unsigned> &Ops,
3990 int FrameIndex) const {
3991 // Check switch flag
3992 if (NoFusing) return NULL;
3994 // Unless optimizing for size, don't fold to avoid partial
3995 // register update stalls
3996 if (!MF.getFunction()->getAttributes().
3997 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
3998 hasPartialRegUpdate(MI->getOpcode()))
4001 const MachineFrameInfo *MFI = MF.getFrameInfo();
4002 unsigned Size = MFI->getObjectSize(FrameIndex);
4003 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
4004 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4005 unsigned NewOpc = 0;
4006 unsigned RCSize = 0;
4007 switch (MI->getOpcode()) {
4008 default: return NULL;
4009 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
4010 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4011 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4012 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4014 // Check if it's safe to fold the load. If the size of the object is
4015 // narrower than the load width, then it's not.
4018 // Change to CMPXXri r, 0 first.
4019 MI->setDesc(get(NewOpc));
4020 MI->getOperand(1).ChangeToImmediate(0);
4021 } else if (Ops.size() != 1)
4024 SmallVector<MachineOperand,4> MOs;
4025 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
4026 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
4029 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4031 const SmallVectorImpl<unsigned> &Ops,
4032 MachineInstr *LoadMI) const {
4033 // Check switch flag
4034 if (NoFusing) return NULL;
4036 // Unless optimizing for size, don't fold to avoid partial
4037 // register update stalls
4038 if (!MF.getFunction()->getAttributes().
4039 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4040 hasPartialRegUpdate(MI->getOpcode()))
4043 // Determine the alignment of the load.
4044 unsigned Alignment = 0;
4045 if (LoadMI->hasOneMemOperand())
4046 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
4048 switch (LoadMI->getOpcode()) {
4049 case X86::AVX2_SETALLONES:
4054 case X86::V_SETALLONES:
4066 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4067 unsigned NewOpc = 0;
4068 switch (MI->getOpcode()) {
4069 default: return NULL;
4070 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
4071 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4072 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4073 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
4075 // Change to CMPXXri r, 0 first.
4076 MI->setDesc(get(NewOpc));
4077 MI->getOperand(1).ChangeToImmediate(0);
4078 } else if (Ops.size() != 1)
4081 // Make sure the subregisters match.
4082 // Otherwise we risk changing the size of the load.
4083 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4086 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
4087 switch (LoadMI->getOpcode()) {
4089 case X86::V_SETALLONES:
4090 case X86::AVX2_SETALLONES:
4093 case X86::FsFLD0SS: {
4094 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
4095 // Create a constant-pool entry and operands to load from it.
4097 // Medium and large mode can't fold loads this way.
4098 if (TM.getCodeModel() != CodeModel::Small &&
4099 TM.getCodeModel() != CodeModel::Kernel)
4102 // x86-32 PIC requires a PIC base register for constant pools.
4103 unsigned PICBase = 0;
4104 if (TM.getRelocationModel() == Reloc::PIC_) {
4105 if (TM.getSubtarget<X86Subtarget>().is64Bit())
4108 // FIXME: PICBase = getGlobalBaseReg(&MF);
4109 // This doesn't work for several reasons.
4110 // 1. GlobalBaseReg may have been spilled.
4111 // 2. It may not be live at MI.
4115 // Create a constant-pool entry.
4116 MachineConstantPool &MCP = *MF.getConstantPool();
4118 unsigned Opc = LoadMI->getOpcode();
4119 if (Opc == X86::FsFLD0SS)
4120 Ty = Type::getFloatTy(MF.getFunction()->getContext());
4121 else if (Opc == X86::FsFLD0SD)
4122 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
4123 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
4124 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
4126 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
4128 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
4129 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4130 Constant::getNullValue(Ty);
4131 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
4133 // Create operands to load from the constant pool entry.
4134 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4135 MOs.push_back(MachineOperand::CreateImm(1));
4136 MOs.push_back(MachineOperand::CreateReg(0, false));
4137 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
4138 MOs.push_back(MachineOperand::CreateReg(0, false));
4142 if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4143 LoadMI->getOpcode() == X86::VMOVSSrm) &&
4144 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4146 // These instructions only load 32 bits, we can't fold them if the
4147 // destination register is wider than 32 bits (4 bytes).
4149 if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4150 LoadMI->getOpcode() == X86::VMOVSDrm) &&
4151 MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4153 // These instructions only load 64 bits, we can't fold them if the
4154 // destination register is wider than 64 bits (8 bytes).
4157 // Folding a normal load. Just copy the load's address operands.
4158 unsigned NumOps = LoadMI->getDesc().getNumOperands();
4159 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
4160 MOs.push_back(LoadMI->getOperand(i));
4164 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
4168 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4169 const SmallVectorImpl<unsigned> &Ops) const {
4170 // Check switch flag
4171 if (NoFusing) return 0;
4173 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4174 switch (MI->getOpcode()) {
4175 default: return false;
4182 // FIXME: AsmPrinter doesn't know how to handle
4183 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4184 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4190 if (Ops.size() != 1)
4193 unsigned OpNum = Ops[0];
4194 unsigned Opc = MI->getOpcode();
4195 unsigned NumOps = MI->getDesc().getNumOperands();
4196 bool isTwoAddr = NumOps > 1 &&
4197 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4199 // Folding a memory location into the two-address part of a two-address
4200 // instruction is different than folding it other places. It requires
4201 // replacing the *two* registers with the memory location.
4202 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
4203 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4204 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4205 } else if (OpNum == 0) { // If operand 0
4206 if (Opc == X86::MOV32r0)
4209 OpcodeTablePtr = &RegOp2MemOpTable0;
4210 } else if (OpNum == 1) {
4211 OpcodeTablePtr = &RegOp2MemOpTable1;
4212 } else if (OpNum == 2) {
4213 OpcodeTablePtr = &RegOp2MemOpTable2;
4214 } else if (OpNum == 3) {
4215 OpcodeTablePtr = &RegOp2MemOpTable3;
4218 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4220 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
4223 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4224 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4225 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4226 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4227 MemOp2RegOpTable.find(MI->getOpcode());
4228 if (I == MemOp2RegOpTable.end())
4230 unsigned Opc = I->second.first;
4231 unsigned Index = I->second.second & TB_INDEX_MASK;
4232 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4233 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4234 if (UnfoldLoad && !FoldedLoad)
4236 UnfoldLoad &= FoldedLoad;
4237 if (UnfoldStore && !FoldedStore)
4239 UnfoldStore &= FoldedStore;
4241 const MCInstrDesc &MCID = get(Opc);
4242 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4243 if (!MI->hasOneMemOperand() &&
4244 RC == &X86::VR128RegClass &&
4245 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4246 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4247 // conservatively assume the address is unaligned. That's bad for
4250 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4251 SmallVector<MachineOperand,2> BeforeOps;
4252 SmallVector<MachineOperand,2> AfterOps;
4253 SmallVector<MachineOperand,4> ImpOps;
4254 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4255 MachineOperand &Op = MI->getOperand(i);
4256 if (i >= Index && i < Index + X86::AddrNumOperands)
4257 AddrOps.push_back(Op);
4258 else if (Op.isReg() && Op.isImplicit())
4259 ImpOps.push_back(Op);
4261 BeforeOps.push_back(Op);
4263 AfterOps.push_back(Op);
4266 // Emit the load instruction.
4268 std::pair<MachineInstr::mmo_iterator,
4269 MachineInstr::mmo_iterator> MMOs =
4270 MF.extractLoadMemRefs(MI->memoperands_begin(),
4271 MI->memoperands_end());
4272 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4274 // Address operands cannot be marked isKill.
4275 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4276 MachineOperand &MO = NewMIs[0]->getOperand(i);
4278 MO.setIsKill(false);
4283 // Emit the data processing instruction.
4284 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4285 MachineInstrBuilder MIB(MF, DataMI);
4288 MIB.addReg(Reg, RegState::Define);
4289 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4290 MIB.addOperand(BeforeOps[i]);
4293 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4294 MIB.addOperand(AfterOps[i]);
4295 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4296 MachineOperand &MO = ImpOps[i];
4297 MIB.addReg(MO.getReg(),
4298 getDefRegState(MO.isDef()) |
4299 RegState::Implicit |
4300 getKillRegState(MO.isKill()) |
4301 getDeadRegState(MO.isDead()) |
4302 getUndefRegState(MO.isUndef()));
4304 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4305 switch (DataMI->getOpcode()) {
4307 case X86::CMP64ri32:
4314 MachineOperand &MO0 = DataMI->getOperand(0);
4315 MachineOperand &MO1 = DataMI->getOperand(1);
4316 if (MO1.getImm() == 0) {
4318 switch (DataMI->getOpcode()) {
4319 default: llvm_unreachable("Unreachable!");
4321 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
4323 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
4325 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4326 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4328 DataMI->setDesc(get(NewOpc));
4329 MO1.ChangeToRegister(MO0.getReg(), false);
4333 NewMIs.push_back(DataMI);
4335 // Emit the store instruction.
4337 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
4338 std::pair<MachineInstr::mmo_iterator,
4339 MachineInstr::mmo_iterator> MMOs =
4340 MF.extractStoreMemRefs(MI->memoperands_begin(),
4341 MI->memoperands_end());
4342 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
4349 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
4350 SmallVectorImpl<SDNode*> &NewNodes) const {
4351 if (!N->isMachineOpcode())
4354 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4355 MemOp2RegOpTable.find(N->getMachineOpcode());
4356 if (I == MemOp2RegOpTable.end())
4358 unsigned Opc = I->second.first;
4359 unsigned Index = I->second.second & TB_INDEX_MASK;
4360 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4361 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4362 const MCInstrDesc &MCID = get(Opc);
4363 MachineFunction &MF = DAG.getMachineFunction();
4364 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4365 unsigned NumDefs = MCID.NumDefs;
4366 std::vector<SDValue> AddrOps;
4367 std::vector<SDValue> BeforeOps;
4368 std::vector<SDValue> AfterOps;
4370 unsigned NumOps = N->getNumOperands();
4371 for (unsigned i = 0; i != NumOps-1; ++i) {
4372 SDValue Op = N->getOperand(i);
4373 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
4374 AddrOps.push_back(Op);
4375 else if (i < Index-NumDefs)
4376 BeforeOps.push_back(Op);
4377 else if (i > Index-NumDefs)
4378 AfterOps.push_back(Op);
4380 SDValue Chain = N->getOperand(NumOps-1);
4381 AddrOps.push_back(Chain);
4383 // Emit the load instruction.
4386 EVT VT = *RC->vt_begin();
4387 std::pair<MachineInstr::mmo_iterator,
4388 MachineInstr::mmo_iterator> MMOs =
4389 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4390 cast<MachineSDNode>(N)->memoperands_end());
4391 if (!(*MMOs.first) &&
4392 RC == &X86::VR128RegClass &&
4393 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4394 // Do not introduce a slow unaligned load.
4396 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4397 bool isAligned = (*MMOs.first) &&
4398 (*MMOs.first)->getAlignment() >= Alignment;
4399 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4400 VT, MVT::Other, AddrOps);
4401 NewNodes.push_back(Load);
4403 // Preserve memory reference information.
4404 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4407 // Emit the data processing instruction.
4408 std::vector<EVT> VTs;
4409 const TargetRegisterClass *DstRC = 0;
4410 if (MCID.getNumDefs() > 0) {
4411 DstRC = getRegClass(MCID, 0, &RI, MF);
4412 VTs.push_back(*DstRC->vt_begin());
4414 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
4415 EVT VT = N->getValueType(i);
4416 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
4420 BeforeOps.push_back(SDValue(Load, 0));
4421 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
4422 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
4423 NewNodes.push_back(NewNode);
4425 // Emit the store instruction.
4428 AddrOps.push_back(SDValue(NewNode, 0));
4429 AddrOps.push_back(Chain);
4430 std::pair<MachineInstr::mmo_iterator,
4431 MachineInstr::mmo_iterator> MMOs =
4432 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4433 cast<MachineSDNode>(N)->memoperands_end());
4434 if (!(*MMOs.first) &&
4435 RC == &X86::VR128RegClass &&
4436 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4437 // Do not introduce a slow unaligned store.
4439 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4440 bool isAligned = (*MMOs.first) &&
4441 (*MMOs.first)->getAlignment() >= Alignment;
4442 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4444 dl, MVT::Other, AddrOps);
4445 NewNodes.push_back(Store);
4447 // Preserve memory reference information.
4448 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4454 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
4455 bool UnfoldLoad, bool UnfoldStore,
4456 unsigned *LoadRegIndex) const {
4457 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4458 MemOp2RegOpTable.find(Opc);
4459 if (I == MemOp2RegOpTable.end())
4461 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4462 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4463 if (UnfoldLoad && !FoldedLoad)
4465 if (UnfoldStore && !FoldedStore)
4468 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
4469 return I->second.first;
4473 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4474 int64_t &Offset1, int64_t &Offset2) const {
4475 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4477 unsigned Opc1 = Load1->getMachineOpcode();
4478 unsigned Opc2 = Load2->getMachineOpcode();
4480 default: return false;
4490 case X86::MMX_MOVD64rm:
4491 case X86::MMX_MOVQ64rm:
4492 case X86::FsMOVAPSrm:
4493 case X86::FsMOVAPDrm:
4499 // AVX load instructions
4502 case X86::FsVMOVAPSrm:
4503 case X86::FsVMOVAPDrm:
4504 case X86::VMOVAPSrm:
4505 case X86::VMOVUPSrm:
4506 case X86::VMOVAPDrm:
4507 case X86::VMOVDQArm:
4508 case X86::VMOVDQUrm:
4509 case X86::VMOVAPSYrm:
4510 case X86::VMOVUPSYrm:
4511 case X86::VMOVAPDYrm:
4512 case X86::VMOVDQAYrm:
4513 case X86::VMOVDQUYrm:
4517 default: return false;
4527 case X86::MMX_MOVD64rm:
4528 case X86::MMX_MOVQ64rm:
4529 case X86::FsMOVAPSrm:
4530 case X86::FsMOVAPDrm:
4536 // AVX load instructions
4539 case X86::FsVMOVAPSrm:
4540 case X86::FsVMOVAPDrm:
4541 case X86::VMOVAPSrm:
4542 case X86::VMOVUPSrm:
4543 case X86::VMOVAPDrm:
4544 case X86::VMOVDQArm:
4545 case X86::VMOVDQUrm:
4546 case X86::VMOVAPSYrm:
4547 case X86::VMOVUPSYrm:
4548 case X86::VMOVAPDYrm:
4549 case X86::VMOVDQAYrm:
4550 case X86::VMOVDQUYrm:
4554 // Check if chain operands and base addresses match.
4555 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4556 Load1->getOperand(5) != Load2->getOperand(5))
4558 // Segment operands should match as well.
4559 if (Load1->getOperand(4) != Load2->getOperand(4))
4561 // Scale should be 1, Index should be Reg0.
4562 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4563 Load1->getOperand(2) == Load2->getOperand(2)) {
4564 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4567 // Now let's examine the displacements.
4568 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4569 isa<ConstantSDNode>(Load2->getOperand(3))) {
4570 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4571 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4578 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4579 int64_t Offset1, int64_t Offset2,
4580 unsigned NumLoads) const {
4581 assert(Offset2 > Offset1);
4582 if ((Offset2 - Offset1) / 8 > 64)
4585 unsigned Opc1 = Load1->getMachineOpcode();
4586 unsigned Opc2 = Load2->getMachineOpcode();
4588 return false; // FIXME: overly conservative?
4595 case X86::MMX_MOVD64rm:
4596 case X86::MMX_MOVQ64rm:
4600 EVT VT = Load1->getValueType(0);
4601 switch (VT.getSimpleVT().SimpleTy) {
4603 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4604 // have 16 of them to play with.
4605 if (TM.getSubtargetImpl()->is64Bit()) {
4608 } else if (NumLoads) {
4628 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
4629 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
4630 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
4631 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4633 Cond[0].setImm(GetOppositeBranchCondition(CC));
4638 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4639 // FIXME: Return false for x87 stack register classes for now. We can't
4640 // allow any loads of these registers before FpGet_ST0_80.
4641 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4642 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
4645 /// getGlobalBaseReg - Return a virtual register initialized with the
4646 /// the global base register value. Output instructions required to
4647 /// initialize the register in the function entry block, if necessary.
4649 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4651 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4652 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4653 "X86-64 PIC uses RIP relative addressing");
4655 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4656 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4657 if (GlobalBaseReg != 0)
4658 return GlobalBaseReg;
4660 // Create the register. The code to initialize it is inserted
4661 // later, by the CGBR pass (below).
4662 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4663 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
4664 X86FI->setGlobalBaseReg(GlobalBaseReg);
4665 return GlobalBaseReg;
4668 // These are the replaceable SSE instructions. Some of these have Int variants
4669 // that we don't include here. We don't want to replace instructions selected
4671 static const uint16_t ReplaceableInstrs[][3] = {
4672 //PackedSingle PackedDouble PackedInt
4673 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4674 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4675 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4676 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4677 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4678 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4679 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4680 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4681 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4682 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4683 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4684 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4685 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4686 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
4687 // AVX 128-bit support
4688 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4689 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4690 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4691 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4692 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4693 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4694 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4695 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4696 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4697 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4698 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4699 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
4700 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4701 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
4702 // AVX 256-bit support
4703 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4704 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4705 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4706 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4707 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
4708 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4711 static const uint16_t ReplaceableInstrsAVX2[][3] = {
4712 //PackedSingle PackedDouble PackedInt
4713 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4714 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4715 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4716 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4717 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4718 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4719 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
4720 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4721 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4722 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4723 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4724 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4725 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4726 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
4729 // FIXME: Some shuffle and unpack instructions have equivalents in different
4730 // domains, but they require a bit more work than just switching opcodes.
4732 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
4733 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
4734 if (ReplaceableInstrs[i][domain-1] == opcode)
4735 return ReplaceableInstrs[i];
4739 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
4740 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4741 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4742 return ReplaceableInstrsAVX2[i];
4746 std::pair<uint16_t, uint16_t>
4747 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
4748 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4749 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
4750 uint16_t validDomains = 0;
4751 if (domain && lookup(MI->getOpcode(), domain))
4753 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4754 validDomains = hasAVX2 ? 0xe : 0x6;
4755 return std::make_pair(domain, validDomains);
4758 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
4759 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4760 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4761 assert(dom && "Not an SSE instruction");
4762 const uint16_t *table = lookup(MI->getOpcode(), dom);
4763 if (!table) { // try the other table
4764 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4765 "256-bit vector operations only available in AVX2");
4766 table = lookupAVX2(MI->getOpcode(), dom);
4768 assert(table && "Cannot change domain");
4769 MI->setDesc(get(table[Domain-1]));
4772 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4773 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4774 NopInst.setOpcode(X86::NOOP);
4777 bool X86InstrInfo::isHighLatencyDef(int opc) const {
4779 default: return false;
4781 case X86::DIVSDrm_Int:
4783 case X86::DIVSDrr_Int:
4785 case X86::DIVSSrm_Int:
4787 case X86::DIVSSrr_Int:
4793 case X86::SQRTSDm_Int:
4795 case X86::SQRTSDr_Int:
4797 case X86::SQRTSSm_Int:
4799 case X86::SQRTSSr_Int:
4800 // AVX instructions with high latency
4802 case X86::VDIVSDrm_Int:
4804 case X86::VDIVSDrr_Int:
4806 case X86::VDIVSSrm_Int:
4808 case X86::VDIVSSrr_Int:
4814 case X86::VSQRTSDm_Int:
4817 case X86::VSQRTSSm_Int:
4824 hasHighOperandLatency(const InstrItineraryData *ItinData,
4825 const MachineRegisterInfo *MRI,
4826 const MachineInstr *DefMI, unsigned DefIdx,
4827 const MachineInstr *UseMI, unsigned UseIdx) const {
4828 return isHighLatencyDef(DefMI->getOpcode());
4832 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4833 /// global base register for x86-32.
4834 struct CGBR : public MachineFunctionPass {
4836 CGBR() : MachineFunctionPass(ID) {}
4838 virtual bool runOnMachineFunction(MachineFunction &MF) {
4839 const X86TargetMachine *TM =
4840 static_cast<const X86TargetMachine *>(&MF.getTarget());
4842 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4843 "X86-64 PIC uses RIP relative addressing");
4845 // Only emit a global base reg in PIC mode.
4846 if (TM->getRelocationModel() != Reloc::PIC_)
4849 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4850 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4852 // If we didn't need a GlobalBaseReg, don't insert code.
4853 if (GlobalBaseReg == 0)
4856 // Insert the set of GlobalBaseReg into the first MBB of the function
4857 MachineBasicBlock &FirstMBB = MF.front();
4858 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4859 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4860 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4861 const X86InstrInfo *TII = TM->getInstrInfo();
4864 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
4865 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
4869 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4870 // only used in JIT code emission as displacement to pc.
4871 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
4873 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4874 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4875 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
4876 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4877 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4878 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4879 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4885 virtual const char *getPassName() const {
4886 return "X86 PIC Global Base Reg Initialization";
4889 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4890 AU.setPreservesCFG();
4891 MachineFunctionPass::getAnalysisUsage(AU);
4898 llvm::createGlobalBaseRegPass() { return new CGBR(); }
4901 struct LDTLSCleanup : public MachineFunctionPass {
4903 LDTLSCleanup() : MachineFunctionPass(ID) {}
4905 virtual bool runOnMachineFunction(MachineFunction &MF) {
4906 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4907 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4908 // No point folding accesses if there isn't at least two.
4912 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4913 return VisitNode(DT->getRootNode(), 0);
4916 // Visit the dominator subtree rooted at Node in pre-order.
4917 // If TLSBaseAddrReg is non-null, then use that to replace any
4918 // TLS_base_addr instructions. Otherwise, create the register
4919 // when the first such instruction is seen, and then use it
4920 // as we encounter more instructions.
4921 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4922 MachineBasicBlock *BB = Node->getBlock();
4923 bool Changed = false;
4925 // Traverse the current block.
4926 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4928 switch (I->getOpcode()) {
4929 case X86::TLS_base_addr32:
4930 case X86::TLS_base_addr64:
4932 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4934 I = SetRegister(I, &TLSBaseAddrReg);
4942 // Visit the children of this block in the dominator tree.
4943 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4945 Changed |= VisitNode(*I, TLSBaseAddrReg);
4951 // Replace the TLS_base_addr instruction I with a copy from
4952 // TLSBaseAddrReg, returning the new instruction.
4953 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4954 unsigned TLSBaseAddrReg) {
4955 MachineFunction *MF = I->getParent()->getParent();
4956 const X86TargetMachine *TM =
4957 static_cast<const X86TargetMachine *>(&MF->getTarget());
4958 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4959 const X86InstrInfo *TII = TM->getInstrInfo();
4961 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4962 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4963 TII->get(TargetOpcode::COPY),
4964 is64Bit ? X86::RAX : X86::EAX)
4965 .addReg(TLSBaseAddrReg);
4967 // Erase the TLS_base_addr instruction.
4968 I->eraseFromParent();
4973 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4974 // inserting a copy instruction after I. Returns the new instruction.
4975 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4976 MachineFunction *MF = I->getParent()->getParent();
4977 const X86TargetMachine *TM =
4978 static_cast<const X86TargetMachine *>(&MF->getTarget());
4979 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4980 const X86InstrInfo *TII = TM->getInstrInfo();
4982 // Create a virtual register for the TLS base address.
4983 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4984 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4985 ? &X86::GR64RegClass
4986 : &X86::GR32RegClass);
4988 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4989 MachineInstr *Next = I->getNextNode();
4990 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4991 TII->get(TargetOpcode::COPY),
4993 .addReg(is64Bit ? X86::RAX : X86::EAX);
4998 virtual const char *getPassName() const {
4999 return "Local Dynamic TLS Access Clean-up";
5002 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5003 AU.setPreservesCFG();
5004 AU.addRequired<MachineDominatorTree>();
5005 MachineFunctionPass::getAnalysisUsage(AU);
5010 char LDTLSCleanup::ID = 0;
5012 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
projects / oota-llvm.git / blob