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/DerivedTypes.h"
21 #include "llvm/LLVMContext.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/LiveVariables.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"
36 #include "llvm/ADT/Statistic.h"
39 #define GET_INSTRINFO_CTOR
40 #include "X86GenInstrInfo.inc"
44 STATISTIC(NumCmpsRemoved, "Number of Cmps removed due to an earlier Sub");
47 NoFusing("disable-spill-fusing",
48 cl::desc("Disable fusing of spill code into instructions"));
50 PrintFailedFusing("print-failed-fuse-candidates",
51 cl::desc("Print instructions that the allocator wants to"
52 " fuse, but the X86 backend currently can't"),
55 ReMatPICStubLoad("remat-pic-stub-load",
56 cl::desc("Re-materialize load from stub in PIC mode"),
57 cl::init(false), cl::Hidden);
60 // Select which memory operand is being unfolded.
61 // (stored in bits 0 - 7)
68 // Minimum alignment required for load/store.
69 // Used for RegOp->MemOp conversion.
70 // (stored in bits 8 - 15)
72 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
73 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
74 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
75 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT,
77 // Do not insert the reverse map (MemOp -> RegOp) into the table.
78 // This may be needed because there is a many -> one mapping.
79 TB_NO_REVERSE = 1 << 16,
81 // Do not insert the forward map (RegOp -> MemOp) into the table.
82 // This is needed for Native Client, which prohibits branch
83 // instructions from using a memory operand.
84 TB_NO_FORWARD = 1 << 17,
86 TB_FOLDED_LOAD = 1 << 18,
87 TB_FOLDED_STORE = 1 << 19
90 struct X86OpTblEntry {
96 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
97 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKDOWN64
99 : X86::ADJCALLSTACKDOWN32),
100 (tm.getSubtarget<X86Subtarget>().is64Bit()
101 ? X86::ADJCALLSTACKUP64
102 : X86::ADJCALLSTACKUP32)),
103 TM(tm), RI(tm, *this) {
105 static const X86OpTblEntry OpTbl2Addr[] = {
106 { X86::ADC32ri, X86::ADC32mi, 0 },
107 { X86::ADC32ri8, X86::ADC32mi8, 0 },
108 { X86::ADC32rr, X86::ADC32mr, 0 },
109 { X86::ADC64ri32, X86::ADC64mi32, 0 },
110 { X86::ADC64ri8, X86::ADC64mi8, 0 },
111 { X86::ADC64rr, X86::ADC64mr, 0 },
112 { X86::ADD16ri, X86::ADD16mi, 0 },
113 { X86::ADD16ri8, X86::ADD16mi8, 0 },
114 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
115 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
116 { X86::ADD16rr, X86::ADD16mr, 0 },
117 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
118 { X86::ADD32ri, X86::ADD32mi, 0 },
119 { X86::ADD32ri8, X86::ADD32mi8, 0 },
120 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
121 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
122 { X86::ADD32rr, X86::ADD32mr, 0 },
123 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
124 { X86::ADD64ri32, X86::ADD64mi32, 0 },
125 { X86::ADD64ri8, X86::ADD64mi8, 0 },
126 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
127 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
128 { X86::ADD64rr, X86::ADD64mr, 0 },
129 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
130 { X86::ADD8ri, X86::ADD8mi, 0 },
131 { X86::ADD8rr, X86::ADD8mr, 0 },
132 { X86::AND16ri, X86::AND16mi, 0 },
133 { X86::AND16ri8, X86::AND16mi8, 0 },
134 { X86::AND16rr, X86::AND16mr, 0 },
135 { X86::AND32ri, X86::AND32mi, 0 },
136 { X86::AND32ri8, X86::AND32mi8, 0 },
137 { X86::AND32rr, X86::AND32mr, 0 },
138 { X86::AND64ri32, X86::AND64mi32, 0 },
139 { X86::AND64ri8, X86::AND64mi8, 0 },
140 { X86::AND64rr, X86::AND64mr, 0 },
141 { X86::AND8ri, X86::AND8mi, 0 },
142 { X86::AND8rr, X86::AND8mr, 0 },
143 { X86::DEC16r, X86::DEC16m, 0 },
144 { X86::DEC32r, X86::DEC32m, 0 },
145 { X86::DEC64_16r, X86::DEC64_16m, 0 },
146 { X86::DEC64_32r, X86::DEC64_32m, 0 },
147 { X86::DEC64r, X86::DEC64m, 0 },
148 { X86::DEC8r, X86::DEC8m, 0 },
149 { X86::INC16r, X86::INC16m, 0 },
150 { X86::INC32r, X86::INC32m, 0 },
151 { X86::INC64_16r, X86::INC64_16m, 0 },
152 { X86::INC64_32r, X86::INC64_32m, 0 },
153 { X86::INC64r, X86::INC64m, 0 },
154 { X86::INC8r, X86::INC8m, 0 },
155 { X86::NEG16r, X86::NEG16m, 0 },
156 { X86::NEG32r, X86::NEG32m, 0 },
157 { X86::NEG64r, X86::NEG64m, 0 },
158 { X86::NEG8r, X86::NEG8m, 0 },
159 { X86::NOT16r, X86::NOT16m, 0 },
160 { X86::NOT32r, X86::NOT32m, 0 },
161 { X86::NOT64r, X86::NOT64m, 0 },
162 { X86::NOT8r, X86::NOT8m, 0 },
163 { X86::OR16ri, X86::OR16mi, 0 },
164 { X86::OR16ri8, X86::OR16mi8, 0 },
165 { X86::OR16rr, X86::OR16mr, 0 },
166 { X86::OR32ri, X86::OR32mi, 0 },
167 { X86::OR32ri8, X86::OR32mi8, 0 },
168 { X86::OR32rr, X86::OR32mr, 0 },
169 { X86::OR64ri32, X86::OR64mi32, 0 },
170 { X86::OR64ri8, X86::OR64mi8, 0 },
171 { X86::OR64rr, X86::OR64mr, 0 },
172 { X86::OR8ri, X86::OR8mi, 0 },
173 { X86::OR8rr, X86::OR8mr, 0 },
174 { X86::ROL16r1, X86::ROL16m1, 0 },
175 { X86::ROL16rCL, X86::ROL16mCL, 0 },
176 { X86::ROL16ri, X86::ROL16mi, 0 },
177 { X86::ROL32r1, X86::ROL32m1, 0 },
178 { X86::ROL32rCL, X86::ROL32mCL, 0 },
179 { X86::ROL32ri, X86::ROL32mi, 0 },
180 { X86::ROL64r1, X86::ROL64m1, 0 },
181 { X86::ROL64rCL, X86::ROL64mCL, 0 },
182 { X86::ROL64ri, X86::ROL64mi, 0 },
183 { X86::ROL8r1, X86::ROL8m1, 0 },
184 { X86::ROL8rCL, X86::ROL8mCL, 0 },
185 { X86::ROL8ri, X86::ROL8mi, 0 },
186 { X86::ROR16r1, X86::ROR16m1, 0 },
187 { X86::ROR16rCL, X86::ROR16mCL, 0 },
188 { X86::ROR16ri, X86::ROR16mi, 0 },
189 { X86::ROR32r1, X86::ROR32m1, 0 },
190 { X86::ROR32rCL, X86::ROR32mCL, 0 },
191 { X86::ROR32ri, X86::ROR32mi, 0 },
192 { X86::ROR64r1, X86::ROR64m1, 0 },
193 { X86::ROR64rCL, X86::ROR64mCL, 0 },
194 { X86::ROR64ri, X86::ROR64mi, 0 },
195 { X86::ROR8r1, X86::ROR8m1, 0 },
196 { X86::ROR8rCL, X86::ROR8mCL, 0 },
197 { X86::ROR8ri, X86::ROR8mi, 0 },
198 { X86::SAR16r1, X86::SAR16m1, 0 },
199 { X86::SAR16rCL, X86::SAR16mCL, 0 },
200 { X86::SAR16ri, X86::SAR16mi, 0 },
201 { X86::SAR32r1, X86::SAR32m1, 0 },
202 { X86::SAR32rCL, X86::SAR32mCL, 0 },
203 { X86::SAR32ri, X86::SAR32mi, 0 },
204 { X86::SAR64r1, X86::SAR64m1, 0 },
205 { X86::SAR64rCL, X86::SAR64mCL, 0 },
206 { X86::SAR64ri, X86::SAR64mi, 0 },
207 { X86::SAR8r1, X86::SAR8m1, 0 },
208 { X86::SAR8rCL, X86::SAR8mCL, 0 },
209 { X86::SAR8ri, X86::SAR8mi, 0 },
210 { X86::SBB32ri, X86::SBB32mi, 0 },
211 { X86::SBB32ri8, X86::SBB32mi8, 0 },
212 { X86::SBB32rr, X86::SBB32mr, 0 },
213 { X86::SBB64ri32, X86::SBB64mi32, 0 },
214 { X86::SBB64ri8, X86::SBB64mi8, 0 },
215 { X86::SBB64rr, X86::SBB64mr, 0 },
216 { X86::SHL16rCL, X86::SHL16mCL, 0 },
217 { X86::SHL16ri, X86::SHL16mi, 0 },
218 { X86::SHL32rCL, X86::SHL32mCL, 0 },
219 { X86::SHL32ri, X86::SHL32mi, 0 },
220 { X86::SHL64rCL, X86::SHL64mCL, 0 },
221 { X86::SHL64ri, X86::SHL64mi, 0 },
222 { X86::SHL8rCL, X86::SHL8mCL, 0 },
223 { X86::SHL8ri, X86::SHL8mi, 0 },
224 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
225 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
226 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
227 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
228 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
229 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
230 { X86::SHR16r1, X86::SHR16m1, 0 },
231 { X86::SHR16rCL, X86::SHR16mCL, 0 },
232 { X86::SHR16ri, X86::SHR16mi, 0 },
233 { X86::SHR32r1, X86::SHR32m1, 0 },
234 { X86::SHR32rCL, X86::SHR32mCL, 0 },
235 { X86::SHR32ri, X86::SHR32mi, 0 },
236 { X86::SHR64r1, X86::SHR64m1, 0 },
237 { X86::SHR64rCL, X86::SHR64mCL, 0 },
238 { X86::SHR64ri, X86::SHR64mi, 0 },
239 { X86::SHR8r1, X86::SHR8m1, 0 },
240 { X86::SHR8rCL, X86::SHR8mCL, 0 },
241 { X86::SHR8ri, X86::SHR8mi, 0 },
242 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
243 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
244 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
245 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
246 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
247 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
248 { X86::SUB16ri, X86::SUB16mi, 0 },
249 { X86::SUB16ri8, X86::SUB16mi8, 0 },
250 { X86::SUB16rr, X86::SUB16mr, 0 },
251 { X86::SUB32ri, X86::SUB32mi, 0 },
252 { X86::SUB32ri8, X86::SUB32mi8, 0 },
253 { X86::SUB32rr, X86::SUB32mr, 0 },
254 { X86::SUB64ri32, X86::SUB64mi32, 0 },
255 { X86::SUB64ri8, X86::SUB64mi8, 0 },
256 { X86::SUB64rr, X86::SUB64mr, 0 },
257 { X86::SUB8ri, X86::SUB8mi, 0 },
258 { X86::SUB8rr, X86::SUB8mr, 0 },
259 { X86::XOR16ri, X86::XOR16mi, 0 },
260 { X86::XOR16ri8, X86::XOR16mi8, 0 },
261 { X86::XOR16rr, X86::XOR16mr, 0 },
262 { X86::XOR32ri, X86::XOR32mi, 0 },
263 { X86::XOR32ri8, X86::XOR32mi8, 0 },
264 { X86::XOR32rr, X86::XOR32mr, 0 },
265 { X86::XOR64ri32, X86::XOR64mi32, 0 },
266 { X86::XOR64ri8, X86::XOR64mi8, 0 },
267 { X86::XOR64rr, X86::XOR64mr, 0 },
268 { X86::XOR8ri, X86::XOR8mi, 0 },
269 { X86::XOR8rr, X86::XOR8mr, 0 }
272 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
273 unsigned RegOp = OpTbl2Addr[i].RegOp;
274 unsigned MemOp = OpTbl2Addr[i].MemOp;
275 unsigned Flags = OpTbl2Addr[i].Flags;
276 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
278 // Index 0, folded load and store, no alignment requirement.
279 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
282 static const X86OpTblEntry OpTbl0[] = {
283 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
284 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
285 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
286 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
287 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
288 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
289 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
290 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
291 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
292 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
293 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
294 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
295 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
296 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
297 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
298 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
299 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
300 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
301 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
302 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
303 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
304 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
305 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
306 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
307 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
308 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
309 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
310 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
311 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
312 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
313 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
314 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
315 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
316 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
317 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
318 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
319 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
320 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
321 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
322 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
323 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
324 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
325 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
326 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
327 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
328 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
329 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
330 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
331 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
332 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
333 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
334 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
335 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
336 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
337 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
338 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
339 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
340 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
341 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
342 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
343 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
344 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
345 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
346 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
347 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
348 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
349 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
350 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
351 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
352 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
353 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
354 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
355 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
356 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
357 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
358 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
359 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
360 // AVX 128-bit versions of foldable instructions
361 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
362 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
363 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
364 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
366 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
367 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
368 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
369 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
370 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
371 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
372 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
373 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
374 // AVX 256-bit foldable instructions
375 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
376 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
377 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
378 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
379 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
380 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
383 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
384 unsigned RegOp = OpTbl0[i].RegOp;
385 unsigned MemOp = OpTbl0[i].MemOp;
386 unsigned Flags = OpTbl0[i].Flags;
387 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
388 RegOp, MemOp, TB_INDEX_0 | Flags);
391 static const X86OpTblEntry OpTbl1[] = {
392 { X86::CMP16rr, X86::CMP16rm, 0 },
393 { X86::CMP32rr, X86::CMP32rm, 0 },
394 { X86::CMP64rr, X86::CMP64rm, 0 },
395 { X86::CMP8rr, X86::CMP8rm, 0 },
396 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
397 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
398 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
399 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
400 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
401 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
402 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
403 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
404 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
405 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
406 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
407 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
408 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
409 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
410 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
411 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
412 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
413 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
414 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
415 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
416 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm, TB_ALIGN_16 },
417 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm, TB_ALIGN_16 },
418 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm, TB_ALIGN_16 },
419 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm, TB_ALIGN_16 },
420 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm, TB_ALIGN_16 },
421 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm, 0 },
422 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
423 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
424 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
425 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
426 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
427 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
428 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
429 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
430 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
431 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
432 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
433 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
434 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
435 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
436 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
437 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
438 { X86::MOV16rr, X86::MOV16rm, 0 },
439 { X86::MOV32rr, X86::MOV32rm, 0 },
440 { X86::MOV64rr, X86::MOV64rm, 0 },
441 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
442 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
443 { X86::MOV8rr, X86::MOV8rm, 0 },
444 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
445 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
446 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
447 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
448 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
449 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
450 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
451 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
452 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
453 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
454 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
455 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
456 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
457 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
458 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
459 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
460 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
461 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
462 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
463 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
464 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
465 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
466 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
467 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
468 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
469 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
470 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
471 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
472 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
473 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
474 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
475 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
476 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
477 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
478 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
479 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
480 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
481 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
482 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
483 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 },
484 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
485 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 },
486 { X86::SQRTSDr, X86::SQRTSDm, 0 },
487 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
488 { X86::SQRTSSr, X86::SQRTSSm, 0 },
489 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
490 { X86::TEST16rr, X86::TEST16rm, 0 },
491 { X86::TEST32rr, X86::TEST32rm, 0 },
492 { X86::TEST64rr, X86::TEST64rm, 0 },
493 { X86::TEST8rr, X86::TEST8rm, 0 },
494 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
495 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
496 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
497 // AVX 128-bit versions of foldable instructions
498 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
499 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
500 { X86::Int_VCVTDQ2PDrr, X86::Int_VCVTDQ2PDrm, TB_ALIGN_16 },
501 { X86::Int_VCVTDQ2PSrr, X86::Int_VCVTDQ2PSrm, TB_ALIGN_16 },
502 { X86::Int_VCVTPD2DQrr, X86::Int_VCVTPD2DQrm, TB_ALIGN_16 },
503 { X86::Int_VCVTPD2PSrr, X86::Int_VCVTPD2PSrm, TB_ALIGN_16 },
504 { X86::Int_VCVTPS2DQrr, X86::Int_VCVTPS2DQrm, TB_ALIGN_16 },
505 { X86::Int_VCVTPS2PDrr, X86::Int_VCVTPS2PDrm, 0 },
506 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
507 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
508 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
509 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
510 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
511 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
512 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
513 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
514 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
515 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
516 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
517 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
518 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
519 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
520 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 },
521 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
522 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
523 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
524 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
525 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 },
526 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 },
527 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 },
528 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
529 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
530 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 },
531 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 },
532 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 },
533 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 },
534 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 },
535 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 },
536 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 },
537 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 },
538 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 },
539 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 },
540 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 },
541 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
542 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
543 // AVX 256-bit foldable instructions
544 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
545 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
546 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
547 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
548 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
549 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
550 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
551 // AVX2 foldable instructions
552 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 },
553 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 },
554 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 },
555 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 },
556 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 },
557 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 },
558 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 },
559 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 },
560 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 },
561 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 },
562 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 },
563 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 },
564 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 },
565 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 },
568 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
569 unsigned RegOp = OpTbl1[i].RegOp;
570 unsigned MemOp = OpTbl1[i].MemOp;
571 unsigned Flags = OpTbl1[i].Flags;
572 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
574 // Index 1, folded load
575 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
578 static const X86OpTblEntry OpTbl2[] = {
579 { X86::ADC32rr, X86::ADC32rm, 0 },
580 { X86::ADC64rr, X86::ADC64rm, 0 },
581 { X86::ADD16rr, X86::ADD16rm, 0 },
582 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
583 { X86::ADD32rr, X86::ADD32rm, 0 },
584 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
585 { X86::ADD64rr, X86::ADD64rm, 0 },
586 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
587 { X86::ADD8rr, X86::ADD8rm, 0 },
588 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
589 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
590 { X86::ADDSDrr, X86::ADDSDrm, 0 },
591 { X86::ADDSSrr, X86::ADDSSrm, 0 },
592 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
593 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
594 { X86::AND16rr, X86::AND16rm, 0 },
595 { X86::AND32rr, X86::AND32rm, 0 },
596 { X86::AND64rr, X86::AND64rm, 0 },
597 { X86::AND8rr, X86::AND8rm, 0 },
598 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
599 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
600 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
601 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
602 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
603 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
604 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
605 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
606 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
607 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
608 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
609 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
610 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
611 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
612 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
613 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
614 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
615 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
616 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
617 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
618 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
619 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
620 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
621 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
622 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
623 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
624 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
625 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
626 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
627 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
628 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
629 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
630 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
631 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
632 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
633 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
634 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
635 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
636 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
637 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
638 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
639 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
640 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
641 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
642 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
643 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
644 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
645 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
646 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
647 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
648 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
649 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
650 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
651 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
652 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
653 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
654 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
655 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
656 { X86::CMPSDrr, X86::CMPSDrm, 0 },
657 { X86::CMPSSrr, X86::CMPSSrm, 0 },
658 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
659 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
660 { X86::DIVSDrr, X86::DIVSDrm, 0 },
661 { X86::DIVSSrr, X86::DIVSSrm, 0 },
662 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
663 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
664 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
665 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
666 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
667 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
668 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
669 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
670 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
671 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
672 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
673 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
674 { X86::IMUL16rr, X86::IMUL16rm, 0 },
675 { X86::IMUL32rr, X86::IMUL32rm, 0 },
676 { X86::IMUL64rr, X86::IMUL64rm, 0 },
677 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
678 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
679 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
680 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
681 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
682 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
683 { X86::MAXSDrr, X86::MAXSDrm, 0 },
684 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
685 { X86::MAXSSrr, X86::MAXSSrm, 0 },
686 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
687 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
688 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
689 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
690 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
691 { X86::MINSDrr, X86::MINSDrm, 0 },
692 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
693 { X86::MINSSrr, X86::MINSSrm, 0 },
694 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
695 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
696 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
697 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
698 { X86::MULSDrr, X86::MULSDrm, 0 },
699 { X86::MULSSrr, X86::MULSSrm, 0 },
700 { X86::OR16rr, X86::OR16rm, 0 },
701 { X86::OR32rr, X86::OR32rm, 0 },
702 { X86::OR64rr, X86::OR64rm, 0 },
703 { X86::OR8rr, X86::OR8rm, 0 },
704 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
705 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
706 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
707 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
708 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
709 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
710 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
711 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
712 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
713 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
714 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
715 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
716 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
717 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
718 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
719 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
720 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
721 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
722 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
723 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
724 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
725 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
726 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
727 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
728 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
729 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
730 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
731 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
732 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
733 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
734 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
735 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
736 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
737 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
738 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
739 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
740 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
741 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
742 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
743 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
744 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
745 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
746 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
747 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
748 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
749 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
750 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
751 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
752 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
753 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
754 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
755 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
756 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
757 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
758 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
759 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
760 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
761 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
762 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
763 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
764 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
765 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
766 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
767 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
768 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
769 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
770 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
771 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
772 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
773 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
774 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
775 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
776 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
777 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
778 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
779 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
780 { X86::SBB32rr, X86::SBB32rm, 0 },
781 { X86::SBB64rr, X86::SBB64rm, 0 },
782 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
783 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
784 { X86::SUB16rr, X86::SUB16rm, 0 },
785 { X86::SUB32rr, X86::SUB32rm, 0 },
786 { X86::SUB64rr, X86::SUB64rm, 0 },
787 { X86::SUB8rr, X86::SUB8rm, 0 },
788 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
789 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
790 { X86::SUBSDrr, X86::SUBSDrm, 0 },
791 { X86::SUBSSrr, X86::SUBSSrm, 0 },
792 // FIXME: TEST*rr -> swapped operand of TEST*mr.
793 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
794 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
795 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
796 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
797 { X86::XOR16rr, X86::XOR16rm, 0 },
798 { X86::XOR32rr, X86::XOR32rm, 0 },
799 { X86::XOR64rr, X86::XOR64rm, 0 },
800 { X86::XOR8rr, X86::XOR8rm, 0 },
801 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
802 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
803 // AVX 128-bit versions of foldable instructions
804 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
805 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
806 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
807 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
808 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
809 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
810 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
811 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
812 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
813 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
814 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
815 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
816 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
817 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm, 0 },
818 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
819 { X86::Int_VCVTTSD2SIrr, X86::Int_VCVTTSD2SIrm, 0 },
820 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
821 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm, 0 },
822 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
823 { X86::Int_VCVTTSS2SIrr, X86::Int_VCVTTSS2SIrm, 0 },
824 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
825 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
826 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQrm, TB_ALIGN_16 },
827 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
828 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
829 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
830 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
831 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
832 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
833 { X86::VADDSDrr, X86::VADDSDrm, 0 },
834 { X86::VADDSSrr, X86::VADDSSrm, 0 },
835 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
836 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
837 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
838 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
839 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
840 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
841 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
842 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
843 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
844 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
845 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
846 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
847 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
848 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
849 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
850 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
851 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
852 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
853 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
854 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
855 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
856 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
857 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
858 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
859 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
860 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
861 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
862 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
863 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
864 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
865 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
866 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
867 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
868 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
869 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
870 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
871 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
872 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
873 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
874 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
875 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
876 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
877 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
878 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
879 { X86::VMINSDrr, X86::VMINSDrm, 0 },
880 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
881 { X86::VMINSSrr, X86::VMINSSrm, 0 },
882 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
883 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
884 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
885 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
886 { X86::VMULSDrr, X86::VMULSDrm, 0 },
887 { X86::VMULSSrr, X86::VMULSSrm, 0 },
888 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
889 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
890 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
891 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
892 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
893 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
894 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
895 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
896 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
897 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
898 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
899 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
900 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
901 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
902 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
903 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
904 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
905 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
906 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
907 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
908 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
909 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
910 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
911 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
912 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
913 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
914 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
915 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
916 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
917 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
918 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
919 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
920 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
921 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
922 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
923 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
924 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
925 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
926 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
927 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
928 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
929 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
930 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
931 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
932 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
933 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
934 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
935 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
936 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
937 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
938 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
939 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
940 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
941 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
942 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
943 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
944 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
945 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
946 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
947 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
948 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
949 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
950 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
951 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
952 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
953 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
954 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
955 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
956 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
957 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
958 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
959 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
960 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
961 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
962 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
963 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
964 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
965 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
966 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
967 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
968 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
969 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
970 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
971 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
972 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
973 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
974 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
975 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
976 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
977 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
978 // AVX 256-bit foldable instructions
979 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
980 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
981 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
982 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
983 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
984 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
985 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
986 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
987 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
988 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
989 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
990 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
991 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
992 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
993 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
994 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
995 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
996 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
997 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
998 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
999 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
1000 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
1001 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
1002 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
1003 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
1004 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
1005 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
1006 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
1007 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
1008 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
1009 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
1010 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1011 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1012 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1013 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1014 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1015 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1016 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1017 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1018 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1019 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1020 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1021 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1022 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1023 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1024 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
1025 // AVX2 foldable instructions
1026 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1027 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1028 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1029 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1030 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1031 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1032 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1033 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1034 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1035 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1036 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1037 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1038 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1039 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1040 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1041 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1042 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1043 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1044 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1045 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1046 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1047 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1048 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1049 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1050 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1051 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1052 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1053 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1054 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1055 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
1056 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
1057 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
1058 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
1059 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
1060 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
1061 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
1062 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1063 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
1064 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
1065 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
1066 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1067 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1068 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1069 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1070 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1071 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1072 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1073 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1074 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1075 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1076 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1077 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1078 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1079 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1080 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1081 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
1082 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1083 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1084 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1085 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
1086 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1087 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1088 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1089 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
1090 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
1091 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
1092 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
1093 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1094 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1095 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
1096 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
1097 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1098 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1099 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1100 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
1101 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
1102 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
1103 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1104 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1105 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1106 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1107 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1108 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1109 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1110 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
1111 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
1112 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1113 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1114 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1115 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1116 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1117 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
1118 // FIXME: add AVX 256-bit foldable instructions
1121 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1122 unsigned RegOp = OpTbl2[i].RegOp;
1123 unsigned MemOp = OpTbl2[i].MemOp;
1124 unsigned Flags = OpTbl2[i].Flags;
1125 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1127 // Index 2, folded load
1128 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1131 static const X86OpTblEntry OpTbl3[] = {
1132 // FMA foldable instructions
1133 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1134 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1135 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1136 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1138 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1139 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1140 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1141 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1142 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1143 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1144 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1145 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1146 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1147 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1148 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1149 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
1151 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1152 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1153 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1154 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1156 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1157 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1158 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1159 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1160 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1161 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1162 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1163 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1164 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1165 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1166 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1167 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
1169 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1170 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1171 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1172 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1174 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1175 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1176 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1177 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1178 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1179 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1180 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1181 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1182 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1183 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1184 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1185 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
1187 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1188 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1189 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1190 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1192 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1193 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1194 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1195 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1196 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1197 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1198 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1199 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1200 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1201 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1202 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1203 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
1206 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1207 unsigned RegOp = OpTbl3[i].RegOp;
1208 unsigned MemOp = OpTbl3[i].MemOp;
1209 unsigned Flags = OpTbl3[i].Flags;
1210 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1212 // Index 3, folded load
1213 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1219 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1220 MemOp2RegOpTableType &M2RTable,
1221 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1222 if ((Flags & TB_NO_FORWARD) == 0) {
1223 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1224 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1226 if ((Flags & TB_NO_REVERSE) == 0) {
1227 assert(!M2RTable.count(MemOp) &&
1228 "Duplicated entries in unfolding maps?");
1229 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1234 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1235 unsigned &SrcReg, unsigned &DstReg,
1236 unsigned &SubIdx) const {
1237 switch (MI.getOpcode()) {
1239 case X86::MOVSX16rr8:
1240 case X86::MOVZX16rr8:
1241 case X86::MOVSX32rr8:
1242 case X86::MOVZX32rr8:
1243 case X86::MOVSX64rr8:
1244 case X86::MOVZX64rr8:
1245 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1246 // It's not always legal to reference the low 8-bit of the larger
1247 // register in 32-bit mode.
1249 case X86::MOVSX32rr16:
1250 case X86::MOVZX32rr16:
1251 case X86::MOVSX64rr16:
1252 case X86::MOVZX64rr16:
1253 case X86::MOVSX64rr32:
1254 case X86::MOVZX64rr32: {
1255 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1258 SrcReg = MI.getOperand(1).getReg();
1259 DstReg = MI.getOperand(0).getReg();
1260 switch (MI.getOpcode()) {
1262 llvm_unreachable(0);
1263 case X86::MOVSX16rr8:
1264 case X86::MOVZX16rr8:
1265 case X86::MOVSX32rr8:
1266 case X86::MOVZX32rr8:
1267 case X86::MOVSX64rr8:
1268 case X86::MOVZX64rr8:
1269 SubIdx = X86::sub_8bit;
1271 case X86::MOVSX32rr16:
1272 case X86::MOVZX32rr16:
1273 case X86::MOVSX64rr16:
1274 case X86::MOVZX64rr16:
1275 SubIdx = X86::sub_16bit;
1277 case X86::MOVSX64rr32:
1278 case X86::MOVZX64rr32:
1279 SubIdx = X86::sub_32bit;
1288 /// isFrameOperand - Return true and the FrameIndex if the specified
1289 /// operand and follow operands form a reference to the stack frame.
1290 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1291 int &FrameIndex) const {
1292 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1293 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1294 MI->getOperand(Op+1).getImm() == 1 &&
1295 MI->getOperand(Op+2).getReg() == 0 &&
1296 MI->getOperand(Op+3).getImm() == 0) {
1297 FrameIndex = MI->getOperand(Op).getIndex();
1303 static bool isFrameLoadOpcode(int Opcode) {
1319 case X86::VMOVAPSrm:
1320 case X86::VMOVAPDrm:
1321 case X86::VMOVDQArm:
1322 case X86::VMOVAPSYrm:
1323 case X86::VMOVAPDYrm:
1324 case X86::VMOVDQAYrm:
1325 case X86::MMX_MOVD64rm:
1326 case X86::MMX_MOVQ64rm:
1331 static bool isFrameStoreOpcode(int Opcode) {
1338 case X86::ST_FpP64m:
1346 case X86::VMOVAPSmr:
1347 case X86::VMOVAPDmr:
1348 case X86::VMOVDQAmr:
1349 case X86::VMOVAPSYmr:
1350 case X86::VMOVAPDYmr:
1351 case X86::VMOVDQAYmr:
1352 case X86::MMX_MOVD64mr:
1353 case X86::MMX_MOVQ64mr:
1354 case X86::MMX_MOVNTQmr:
1360 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1361 int &FrameIndex) const {
1362 if (isFrameLoadOpcode(MI->getOpcode()))
1363 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1364 return MI->getOperand(0).getReg();
1368 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1369 int &FrameIndex) const {
1370 if (isFrameLoadOpcode(MI->getOpcode())) {
1372 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1374 // Check for post-frame index elimination operations
1375 const MachineMemOperand *Dummy;
1376 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1381 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1382 int &FrameIndex) const {
1383 if (isFrameStoreOpcode(MI->getOpcode()))
1384 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1385 isFrameOperand(MI, 0, FrameIndex))
1386 return MI->getOperand(X86::AddrNumOperands).getReg();
1390 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1391 int &FrameIndex) const {
1392 if (isFrameStoreOpcode(MI->getOpcode())) {
1394 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1396 // Check for post-frame index elimination operations
1397 const MachineMemOperand *Dummy;
1398 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1403 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1405 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1406 bool isPICBase = false;
1407 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1408 E = MRI.def_end(); I != E; ++I) {
1409 MachineInstr *DefMI = I.getOperand().getParent();
1410 if (DefMI->getOpcode() != X86::MOVPC32r)
1412 assert(!isPICBase && "More than one PIC base?");
1419 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1420 AliasAnalysis *AA) const {
1421 switch (MI->getOpcode()) {
1436 case X86::VMOVAPSrm:
1437 case X86::VMOVUPSrm:
1438 case X86::VMOVAPDrm:
1439 case X86::VMOVDQArm:
1440 case X86::VMOVAPSYrm:
1441 case X86::VMOVUPSYrm:
1442 case X86::VMOVAPDYrm:
1443 case X86::VMOVDQAYrm:
1444 case X86::MMX_MOVD64rm:
1445 case X86::MMX_MOVQ64rm:
1446 case X86::FsVMOVAPSrm:
1447 case X86::FsVMOVAPDrm:
1448 case X86::FsMOVAPSrm:
1449 case X86::FsMOVAPDrm: {
1450 // Loads from constant pools are trivially rematerializable.
1451 if (MI->getOperand(1).isReg() &&
1452 MI->getOperand(2).isImm() &&
1453 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1454 MI->isInvariantLoad(AA)) {
1455 unsigned BaseReg = MI->getOperand(1).getReg();
1456 if (BaseReg == 0 || BaseReg == X86::RIP)
1458 // Allow re-materialization of PIC load.
1459 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1461 const MachineFunction &MF = *MI->getParent()->getParent();
1462 const MachineRegisterInfo &MRI = MF.getRegInfo();
1463 bool isPICBase = false;
1464 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1465 E = MRI.def_end(); I != E; ++I) {
1466 MachineInstr *DefMI = I.getOperand().getParent();
1467 if (DefMI->getOpcode() != X86::MOVPC32r)
1469 assert(!isPICBase && "More than one PIC base?");
1479 if (MI->getOperand(2).isImm() &&
1480 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1481 !MI->getOperand(4).isReg()) {
1482 // lea fi#, lea GV, etc. are all rematerializable.
1483 if (!MI->getOperand(1).isReg())
1485 unsigned BaseReg = MI->getOperand(1).getReg();
1488 // Allow re-materialization of lea PICBase + x.
1489 const MachineFunction &MF = *MI->getParent()->getParent();
1490 const MachineRegisterInfo &MRI = MF.getRegInfo();
1491 return regIsPICBase(BaseReg, MRI);
1497 // All other instructions marked M_REMATERIALIZABLE are always trivially
1498 // rematerializable.
1502 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1503 /// would clobber the EFLAGS condition register. Note the result may be
1504 /// conservative. If it cannot definitely determine the safety after visiting
1505 /// a few instructions in each direction it assumes it's not safe.
1506 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1507 MachineBasicBlock::iterator I) {
1508 MachineBasicBlock::iterator E = MBB.end();
1510 // For compile time consideration, if we are not able to determine the
1511 // safety after visiting 4 instructions in each direction, we will assume
1513 MachineBasicBlock::iterator Iter = I;
1514 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1515 bool SeenDef = false;
1516 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1517 MachineOperand &MO = Iter->getOperand(j);
1518 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1522 if (MO.getReg() == X86::EFLAGS) {
1530 // This instruction defines EFLAGS, no need to look any further.
1533 // Skip over DBG_VALUE.
1534 while (Iter != E && Iter->isDebugValue())
1538 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1541 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1542 SE = MBB.succ_end(); SI != SE; ++SI)
1543 if ((*SI)->isLiveIn(X86::EFLAGS))
1548 MachineBasicBlock::iterator B = MBB.begin();
1550 for (unsigned i = 0; i < 4; ++i) {
1551 // If we make it to the beginning of the block, it's safe to clobber
1552 // EFLAGS iff EFLAGS is not live-in.
1554 return !MBB.isLiveIn(X86::EFLAGS);
1557 // Skip over DBG_VALUE.
1558 while (Iter != B && Iter->isDebugValue())
1561 bool SawKill = false;
1562 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1563 MachineOperand &MO = Iter->getOperand(j);
1564 // A register mask may clobber EFLAGS, but we should still look for a
1566 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1568 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1569 if (MO.isDef()) return MO.isDead();
1570 if (MO.isKill()) SawKill = true;
1575 // This instruction kills EFLAGS and doesn't redefine it, so
1576 // there's no need to look further.
1580 // Conservative answer.
1584 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1585 MachineBasicBlock::iterator I,
1586 unsigned DestReg, unsigned SubIdx,
1587 const MachineInstr *Orig,
1588 const TargetRegisterInfo &TRI) const {
1589 DebugLoc DL = Orig->getDebugLoc();
1591 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1592 // Re-materialize them as movri instructions to avoid side effects.
1594 unsigned Opc = Orig->getOpcode();
1600 case X86::MOV64r0: {
1601 if (!isSafeToClobberEFLAGS(MBB, I)) {
1604 case X86::MOV8r0: Opc = X86::MOV8ri; break;
1605 case X86::MOV16r0: Opc = X86::MOV16ri; break;
1606 case X86::MOV32r0: Opc = X86::MOV32ri; break;
1607 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
1616 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1619 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
1622 MachineInstr *NewMI = prior(I);
1623 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1626 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1627 /// is not marked dead.
1628 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1629 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1630 MachineOperand &MO = MI->getOperand(i);
1631 if (MO.isReg() && MO.isDef() &&
1632 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1639 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1640 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1641 /// to a 32-bit superregister and then truncating back down to a 16-bit
1644 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1645 MachineFunction::iterator &MFI,
1646 MachineBasicBlock::iterator &MBBI,
1647 LiveVariables *LV) const {
1648 MachineInstr *MI = MBBI;
1649 unsigned Dest = MI->getOperand(0).getReg();
1650 unsigned Src = MI->getOperand(1).getReg();
1651 bool isDead = MI->getOperand(0).isDead();
1652 bool isKill = MI->getOperand(1).isKill();
1654 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1655 ? X86::LEA64_32r : X86::LEA32r;
1656 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1657 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1658 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1660 // Build and insert into an implicit UNDEF value. This is OK because
1661 // well be shifting and then extracting the lower 16-bits.
1662 // This has the potential to cause partial register stall. e.g.
1663 // movw (%rbp,%rcx,2), %dx
1664 // leal -65(%rdx), %esi
1665 // But testing has shown this *does* help performance in 64-bit mode (at
1666 // least on modern x86 machines).
1667 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1668 MachineInstr *InsMI =
1669 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1670 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1671 .addReg(Src, getKillRegState(isKill));
1673 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1674 get(Opc), leaOutReg);
1677 llvm_unreachable(0);
1678 case X86::SHL16ri: {
1679 unsigned ShAmt = MI->getOperand(2).getImm();
1680 MIB.addReg(0).addImm(1 << ShAmt)
1681 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1685 case X86::INC64_16r:
1686 addRegOffset(MIB, leaInReg, true, 1);
1689 case X86::DEC64_16r:
1690 addRegOffset(MIB, leaInReg, true, -1);
1694 case X86::ADD16ri_DB:
1695 case X86::ADD16ri8_DB:
1696 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
1699 case X86::ADD16rr_DB: {
1700 unsigned Src2 = MI->getOperand(2).getReg();
1701 bool isKill2 = MI->getOperand(2).isKill();
1702 unsigned leaInReg2 = 0;
1703 MachineInstr *InsMI2 = 0;
1705 // ADD16rr %reg1028<kill>, %reg1028
1706 // just a single insert_subreg.
1707 addRegReg(MIB, leaInReg, true, leaInReg, false);
1709 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1710 // Build and insert into an implicit UNDEF value. This is OK because
1711 // well be shifting and then extracting the lower 16-bits.
1712 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
1714 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
1715 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1716 .addReg(Src2, getKillRegState(isKill2));
1717 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1719 if (LV && isKill2 && InsMI2)
1720 LV->replaceKillInstruction(Src2, MI, InsMI2);
1725 MachineInstr *NewMI = MIB;
1726 MachineInstr *ExtMI =
1727 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1728 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1729 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
1732 // Update live variables
1733 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1734 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1736 LV->replaceKillInstruction(Src, MI, InsMI);
1738 LV->replaceKillInstruction(Dest, MI, ExtMI);
1744 /// convertToThreeAddress - This method must be implemented by targets that
1745 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1746 /// may be able to convert a two-address instruction into a true
1747 /// three-address instruction on demand. This allows the X86 target (for
1748 /// example) to convert ADD and SHL instructions into LEA instructions if they
1749 /// would require register copies due to two-addressness.
1751 /// This method returns a null pointer if the transformation cannot be
1752 /// performed, otherwise it returns the new instruction.
1755 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1756 MachineBasicBlock::iterator &MBBI,
1757 LiveVariables *LV) const {
1758 MachineInstr *MI = MBBI;
1759 MachineFunction &MF = *MI->getParent()->getParent();
1760 // All instructions input are two-addr instructions. Get the known operands.
1761 unsigned Dest = MI->getOperand(0).getReg();
1762 unsigned Src = MI->getOperand(1).getReg();
1763 bool isDead = MI->getOperand(0).isDead();
1764 bool isKill = MI->getOperand(1).isKill();
1766 MachineInstr *NewMI = NULL;
1767 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1768 // we have better subtarget support, enable the 16-bit LEA generation here.
1769 // 16-bit LEA is also slow on Core2.
1770 bool DisableLEA16 = true;
1771 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1773 unsigned MIOpc = MI->getOpcode();
1775 case X86::SHUFPSrri: {
1776 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1777 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1779 unsigned B = MI->getOperand(1).getReg();
1780 unsigned C = MI->getOperand(2).getReg();
1781 if (B != C) return 0;
1782 unsigned A = MI->getOperand(0).getReg();
1783 unsigned M = MI->getOperand(3).getImm();
1784 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1785 .addReg(A, RegState::Define | getDeadRegState(isDead))
1786 .addReg(B, getKillRegState(isKill)).addImm(M);
1789 case X86::SHUFPDrri: {
1790 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1791 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1793 unsigned B = MI->getOperand(1).getReg();
1794 unsigned C = MI->getOperand(2).getReg();
1795 if (B != C) return 0;
1796 unsigned A = MI->getOperand(0).getReg();
1797 unsigned M = MI->getOperand(3).getImm();
1799 // Convert to PSHUFD mask.
1800 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1802 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1803 .addReg(A, RegState::Define | getDeadRegState(isDead))
1804 .addReg(B, getKillRegState(isKill)).addImm(M);
1807 case X86::SHL64ri: {
1808 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1809 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1810 // the flags produced by a shift yet, so this is safe.
1811 unsigned ShAmt = MI->getOperand(2).getImm();
1812 if (ShAmt == 0 || ShAmt >= 4) return 0;
1814 // LEA can't handle RSP.
1815 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1816 !MF.getRegInfo().constrainRegClass(Src, &X86::GR64_NOSPRegClass))
1819 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1820 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1821 .addReg(0).addImm(1 << ShAmt)
1822 .addReg(Src, getKillRegState(isKill))
1823 .addImm(0).addReg(0);
1826 case X86::SHL32ri: {
1827 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1828 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1829 // the flags produced by a shift yet, so this is safe.
1830 unsigned ShAmt = MI->getOperand(2).getImm();
1831 if (ShAmt == 0 || ShAmt >= 4) return 0;
1833 // LEA can't handle ESP.
1834 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1835 !MF.getRegInfo().constrainRegClass(Src, &X86::GR32_NOSPRegClass))
1838 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1839 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1840 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1841 .addReg(0).addImm(1 << ShAmt)
1842 .addReg(Src, getKillRegState(isKill)).addImm(0).addReg(0);
1845 case X86::SHL16ri: {
1846 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1847 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1848 // the flags produced by a shift yet, so this is safe.
1849 unsigned ShAmt = MI->getOperand(2).getImm();
1850 if (ShAmt == 0 || ShAmt >= 4) return 0;
1853 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1854 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1855 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1856 .addReg(0).addImm(1 << ShAmt)
1857 .addReg(Src, getKillRegState(isKill))
1858 .addImm(0).addReg(0);
1862 // The following opcodes also sets the condition code register(s). Only
1863 // convert them to equivalent lea if the condition code register def's
1865 if (hasLiveCondCodeDef(MI))
1872 case X86::INC64_32r: {
1873 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1874 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1875 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1876 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1877 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1878 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
1880 // LEA can't handle RSP.
1881 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1882 !MF.getRegInfo().constrainRegClass(Src, RC))
1885 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1886 .addReg(Dest, RegState::Define |
1887 getDeadRegState(isDead)),
1892 case X86::INC64_16r:
1894 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1895 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1896 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1897 .addReg(Dest, RegState::Define |
1898 getDeadRegState(isDead)),
1903 case X86::DEC64_32r: {
1904 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1905 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1906 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1907 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1908 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1909 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
1910 // LEA can't handle RSP.
1911 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1912 !MF.getRegInfo().constrainRegClass(Src, RC))
1915 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1916 .addReg(Dest, RegState::Define |
1917 getDeadRegState(isDead)),
1922 case X86::DEC64_16r:
1924 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1925 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1926 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1927 .addReg(Dest, RegState::Define |
1928 getDeadRegState(isDead)),
1932 case X86::ADD64rr_DB:
1934 case X86::ADD32rr_DB: {
1935 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1937 const TargetRegisterClass *RC;
1938 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
1940 RC = &X86::GR64_NOSPRegClass;
1942 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1943 RC = &X86::GR32_NOSPRegClass;
1947 unsigned Src2 = MI->getOperand(2).getReg();
1948 bool isKill2 = MI->getOperand(2).isKill();
1950 // LEA can't handle RSP.
1951 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
1952 !MF.getRegInfo().constrainRegClass(Src2, RC))
1955 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1956 .addReg(Dest, RegState::Define |
1957 getDeadRegState(isDead)),
1958 Src, isKill, Src2, isKill2);
1960 LV->replaceKillInstruction(Src2, MI, NewMI);
1964 case X86::ADD16rr_DB: {
1966 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1967 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1968 unsigned Src2 = MI->getOperand(2).getReg();
1969 bool isKill2 = MI->getOperand(2).isKill();
1970 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1971 .addReg(Dest, RegState::Define |
1972 getDeadRegState(isDead)),
1973 Src, isKill, Src2, isKill2);
1975 LV->replaceKillInstruction(Src2, MI, NewMI);
1978 case X86::ADD64ri32:
1980 case X86::ADD64ri32_DB:
1981 case X86::ADD64ri8_DB:
1982 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1983 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1984 .addReg(Dest, RegState::Define |
1985 getDeadRegState(isDead)),
1986 Src, isKill, MI->getOperand(2).getImm());
1990 case X86::ADD32ri_DB:
1991 case X86::ADD32ri8_DB: {
1992 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
1993 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1994 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1995 .addReg(Dest, RegState::Define |
1996 getDeadRegState(isDead)),
1997 Src, isKill, MI->getOperand(2).getImm());
2002 case X86::ADD16ri_DB:
2003 case X86::ADD16ri8_DB:
2005 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2006 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2007 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2008 .addReg(Dest, RegState::Define |
2009 getDeadRegState(isDead)),
2010 Src, isKill, MI->getOperand(2).getImm());
2016 if (!NewMI) return 0;
2018 if (LV) { // Update live variables
2020 LV->replaceKillInstruction(Src, MI, NewMI);
2022 LV->replaceKillInstruction(Dest, MI, NewMI);
2025 MFI->insert(MBBI, NewMI); // Insert the new inst
2029 /// commuteInstruction - We have a few instructions that must be hacked on to
2033 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2034 switch (MI->getOpcode()) {
2035 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2036 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2037 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2038 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2039 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2040 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2043 switch (MI->getOpcode()) {
2044 default: llvm_unreachable("Unreachable!");
2045 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2046 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2047 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2048 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2049 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2050 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2052 unsigned Amt = MI->getOperand(3).getImm();
2054 MachineFunction &MF = *MI->getParent()->getParent();
2055 MI = MF.CloneMachineInstr(MI);
2058 MI->setDesc(get(Opc));
2059 MI->getOperand(3).setImm(Size-Amt);
2060 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
2062 case X86::CMOVB16rr:
2063 case X86::CMOVB32rr:
2064 case X86::CMOVB64rr:
2065 case X86::CMOVAE16rr:
2066 case X86::CMOVAE32rr:
2067 case X86::CMOVAE64rr:
2068 case X86::CMOVE16rr:
2069 case X86::CMOVE32rr:
2070 case X86::CMOVE64rr:
2071 case X86::CMOVNE16rr:
2072 case X86::CMOVNE32rr:
2073 case X86::CMOVNE64rr:
2074 case X86::CMOVBE16rr:
2075 case X86::CMOVBE32rr:
2076 case X86::CMOVBE64rr:
2077 case X86::CMOVA16rr:
2078 case X86::CMOVA32rr:
2079 case X86::CMOVA64rr:
2080 case X86::CMOVL16rr:
2081 case X86::CMOVL32rr:
2082 case X86::CMOVL64rr:
2083 case X86::CMOVGE16rr:
2084 case X86::CMOVGE32rr:
2085 case X86::CMOVGE64rr:
2086 case X86::CMOVLE16rr:
2087 case X86::CMOVLE32rr:
2088 case X86::CMOVLE64rr:
2089 case X86::CMOVG16rr:
2090 case X86::CMOVG32rr:
2091 case X86::CMOVG64rr:
2092 case X86::CMOVS16rr:
2093 case X86::CMOVS32rr:
2094 case X86::CMOVS64rr:
2095 case X86::CMOVNS16rr:
2096 case X86::CMOVNS32rr:
2097 case X86::CMOVNS64rr:
2098 case X86::CMOVP16rr:
2099 case X86::CMOVP32rr:
2100 case X86::CMOVP64rr:
2101 case X86::CMOVNP16rr:
2102 case X86::CMOVNP32rr:
2103 case X86::CMOVNP64rr:
2104 case X86::CMOVO16rr:
2105 case X86::CMOVO32rr:
2106 case X86::CMOVO64rr:
2107 case X86::CMOVNO16rr:
2108 case X86::CMOVNO32rr:
2109 case X86::CMOVNO64rr: {
2111 switch (MI->getOpcode()) {
2113 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2114 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2115 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2116 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2117 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2118 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2119 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2120 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2121 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2122 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2123 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2124 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2125 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2126 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2127 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2128 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2129 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2130 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2131 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2132 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2133 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2134 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2135 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2136 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2137 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2138 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2139 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2140 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2141 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2142 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2143 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2144 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2145 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2146 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2147 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2148 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2149 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2150 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2151 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2152 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2153 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2154 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2155 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2156 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2157 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2158 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2159 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2160 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2163 MachineFunction &MF = *MI->getParent()->getParent();
2164 MI = MF.CloneMachineInstr(MI);
2167 MI->setDesc(get(Opc));
2168 // Fallthrough intended.
2171 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
2175 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
2177 default: return X86::COND_INVALID;
2178 case X86::JE_4: return X86::COND_E;
2179 case X86::JNE_4: return X86::COND_NE;
2180 case X86::JL_4: return X86::COND_L;
2181 case X86::JLE_4: return X86::COND_LE;
2182 case X86::JG_4: return X86::COND_G;
2183 case X86::JGE_4: return X86::COND_GE;
2184 case X86::JB_4: return X86::COND_B;
2185 case X86::JBE_4: return X86::COND_BE;
2186 case X86::JA_4: return X86::COND_A;
2187 case X86::JAE_4: return X86::COND_AE;
2188 case X86::JS_4: return X86::COND_S;
2189 case X86::JNS_4: return X86::COND_NS;
2190 case X86::JP_4: return X86::COND_P;
2191 case X86::JNP_4: return X86::COND_NP;
2192 case X86::JO_4: return X86::COND_O;
2193 case X86::JNO_4: return X86::COND_NO;
2197 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2199 default: llvm_unreachable("Illegal condition code!");
2200 case X86::COND_E: return X86::JE_4;
2201 case X86::COND_NE: return X86::JNE_4;
2202 case X86::COND_L: return X86::JL_4;
2203 case X86::COND_LE: return X86::JLE_4;
2204 case X86::COND_G: return X86::JG_4;
2205 case X86::COND_GE: return X86::JGE_4;
2206 case X86::COND_B: return X86::JB_4;
2207 case X86::COND_BE: return X86::JBE_4;
2208 case X86::COND_A: return X86::JA_4;
2209 case X86::COND_AE: return X86::JAE_4;
2210 case X86::COND_S: return X86::JS_4;
2211 case X86::COND_NS: return X86::JNS_4;
2212 case X86::COND_P: return X86::JP_4;
2213 case X86::COND_NP: return X86::JNP_4;
2214 case X86::COND_O: return X86::JO_4;
2215 case X86::COND_NO: return X86::JNO_4;
2219 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2220 /// e.g. turning COND_E to COND_NE.
2221 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2223 default: llvm_unreachable("Illegal condition code!");
2224 case X86::COND_E: return X86::COND_NE;
2225 case X86::COND_NE: return X86::COND_E;
2226 case X86::COND_L: return X86::COND_GE;
2227 case X86::COND_LE: return X86::COND_G;
2228 case X86::COND_G: return X86::COND_LE;
2229 case X86::COND_GE: return X86::COND_L;
2230 case X86::COND_B: return X86::COND_AE;
2231 case X86::COND_BE: return X86::COND_A;
2232 case X86::COND_A: return X86::COND_BE;
2233 case X86::COND_AE: return X86::COND_B;
2234 case X86::COND_S: return X86::COND_NS;
2235 case X86::COND_NS: return X86::COND_S;
2236 case X86::COND_P: return X86::COND_NP;
2237 case X86::COND_NP: return X86::COND_P;
2238 case X86::COND_O: return X86::COND_NO;
2239 case X86::COND_NO: return X86::COND_O;
2243 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2244 if (!MI->isTerminator()) return false;
2246 // Conditional branch is a special case.
2247 if (MI->isBranch() && !MI->isBarrier())
2249 if (!MI->isPredicable())
2251 return !isPredicated(MI);
2254 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2255 MachineBasicBlock *&TBB,
2256 MachineBasicBlock *&FBB,
2257 SmallVectorImpl<MachineOperand> &Cond,
2258 bool AllowModify) const {
2259 // Start from the bottom of the block and work up, examining the
2260 // terminator instructions.
2261 MachineBasicBlock::iterator I = MBB.end();
2262 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2263 while (I != MBB.begin()) {
2265 if (I->isDebugValue())
2268 // Working from the bottom, when we see a non-terminator instruction, we're
2270 if (!isUnpredicatedTerminator(I))
2273 // A terminator that isn't a branch can't easily be handled by this
2278 // Handle unconditional branches.
2279 if (I->getOpcode() == X86::JMP_4) {
2283 TBB = I->getOperand(0).getMBB();
2287 // If the block has any instructions after a JMP, delete them.
2288 while (llvm::next(I) != MBB.end())
2289 llvm::next(I)->eraseFromParent();
2294 // Delete the JMP if it's equivalent to a fall-through.
2295 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2297 I->eraseFromParent();
2299 UnCondBrIter = MBB.end();
2303 // TBB is used to indicate the unconditional destination.
2304 TBB = I->getOperand(0).getMBB();
2308 // Handle conditional branches.
2309 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
2310 if (BranchCode == X86::COND_INVALID)
2311 return true; // Can't handle indirect branch.
2313 // Working from the bottom, handle the first conditional branch.
2315 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2316 if (AllowModify && UnCondBrIter != MBB.end() &&
2317 MBB.isLayoutSuccessor(TargetBB)) {
2318 // If we can modify the code and it ends in something like:
2326 // Then we can change this to:
2333 // Which is a bit more efficient.
2334 // We conditionally jump to the fall-through block.
2335 BranchCode = GetOppositeBranchCondition(BranchCode);
2336 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2337 MachineBasicBlock::iterator OldInst = I;
2339 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2340 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2341 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2344 OldInst->eraseFromParent();
2345 UnCondBrIter->eraseFromParent();
2347 // Restart the analysis.
2348 UnCondBrIter = MBB.end();
2354 TBB = I->getOperand(0).getMBB();
2355 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2359 // Handle subsequent conditional branches. Only handle the case where all
2360 // conditional branches branch to the same destination and their condition
2361 // opcodes fit one of the special multi-branch idioms.
2362 assert(Cond.size() == 1);
2365 // Only handle the case where all conditional branches branch to the same
2367 if (TBB != I->getOperand(0).getMBB())
2370 // If the conditions are the same, we can leave them alone.
2371 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2372 if (OldBranchCode == BranchCode)
2375 // If they differ, see if they fit one of the known patterns. Theoretically,
2376 // we could handle more patterns here, but we shouldn't expect to see them
2377 // if instruction selection has done a reasonable job.
2378 if ((OldBranchCode == X86::COND_NP &&
2379 BranchCode == X86::COND_E) ||
2380 (OldBranchCode == X86::COND_E &&
2381 BranchCode == X86::COND_NP))
2382 BranchCode = X86::COND_NP_OR_E;
2383 else if ((OldBranchCode == X86::COND_P &&
2384 BranchCode == X86::COND_NE) ||
2385 (OldBranchCode == X86::COND_NE &&
2386 BranchCode == X86::COND_P))
2387 BranchCode = X86::COND_NE_OR_P;
2391 // Update the MachineOperand.
2392 Cond[0].setImm(BranchCode);
2398 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2399 MachineBasicBlock::iterator I = MBB.end();
2402 while (I != MBB.begin()) {
2404 if (I->isDebugValue())
2406 if (I->getOpcode() != X86::JMP_4 &&
2407 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2409 // Remove the branch.
2410 I->eraseFromParent();
2419 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2420 MachineBasicBlock *FBB,
2421 const SmallVectorImpl<MachineOperand> &Cond,
2422 DebugLoc DL) const {
2423 // Shouldn't be a fall through.
2424 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2425 assert((Cond.size() == 1 || Cond.size() == 0) &&
2426 "X86 branch conditions have one component!");
2429 // Unconditional branch?
2430 assert(!FBB && "Unconditional branch with multiple successors!");
2431 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2435 // Conditional branch.
2437 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2439 case X86::COND_NP_OR_E:
2440 // Synthesize NP_OR_E with two branches.
2441 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2443 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2446 case X86::COND_NE_OR_P:
2447 // Synthesize NE_OR_P with two branches.
2448 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2450 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2454 unsigned Opc = GetCondBranchFromCond(CC);
2455 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2460 // Two-way Conditional branch. Insert the second branch.
2461 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2467 /// isHReg - Test if the given register is a physical h register.
2468 static bool isHReg(unsigned Reg) {
2469 return X86::GR8_ABCD_HRegClass.contains(Reg);
2472 // Try and copy between VR128/VR64 and GR64 registers.
2473 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2475 // SrcReg(VR128) -> DestReg(GR64)
2476 // SrcReg(VR64) -> DestReg(GR64)
2477 // SrcReg(GR64) -> DestReg(VR128)
2478 // SrcReg(GR64) -> DestReg(VR64)
2480 if (X86::GR64RegClass.contains(DestReg)) {
2481 if (X86::VR128RegClass.contains(SrcReg)) {
2482 // Copy from a VR128 register to a GR64 register.
2483 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
2484 } else if (X86::VR64RegClass.contains(SrcReg)) {
2485 // Copy from a VR64 register to a GR64 register.
2486 return X86::MOVSDto64rr;
2488 } else if (X86::GR64RegClass.contains(SrcReg)) {
2489 // Copy from a GR64 register to a VR128 register.
2490 if (X86::VR128RegClass.contains(DestReg))
2491 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
2492 // Copy from a GR64 register to a VR64 register.
2493 else if (X86::VR64RegClass.contains(DestReg))
2494 return X86::MOV64toSDrr;
2497 // SrcReg(FR32) -> DestReg(GR32)
2498 // SrcReg(GR32) -> DestReg(FR32)
2500 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
2501 // Copy from a FR32 register to a GR32 register.
2502 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
2504 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
2505 // Copy from a GR32 register to a FR32 register.
2506 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
2511 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2512 MachineBasicBlock::iterator MI, DebugLoc DL,
2513 unsigned DestReg, unsigned SrcReg,
2514 bool KillSrc) const {
2515 // First deal with the normal symmetric copies.
2516 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2518 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2520 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2522 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2524 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2525 // Copying to or from a physical H register on x86-64 requires a NOREX
2526 // move. Otherwise use a normal move.
2527 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
2528 TM.getSubtarget<X86Subtarget>().is64Bit()) {
2529 Opc = X86::MOV8rr_NOREX;
2530 // Both operands must be encodable without an REX prefix.
2531 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2532 "8-bit H register can not be copied outside GR8_NOREX");
2535 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
2536 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
2537 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2538 Opc = X86::VMOVAPSYrr;
2539 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2540 Opc = X86::MMX_MOVQ64rr;
2542 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
2545 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2546 .addReg(SrcReg, getKillRegState(KillSrc));
2550 // Moving EFLAGS to / from another register requires a push and a pop.
2551 if (SrcReg == X86::EFLAGS) {
2552 if (X86::GR64RegClass.contains(DestReg)) {
2553 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2554 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2556 } else if (X86::GR32RegClass.contains(DestReg)) {
2557 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2558 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2562 if (DestReg == X86::EFLAGS) {
2563 if (X86::GR64RegClass.contains(SrcReg)) {
2564 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2565 .addReg(SrcReg, getKillRegState(KillSrc));
2566 BuildMI(MBB, MI, DL, get(X86::POPF64));
2568 } else if (X86::GR32RegClass.contains(SrcReg)) {
2569 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2570 .addReg(SrcReg, getKillRegState(KillSrc));
2571 BuildMI(MBB, MI, DL, get(X86::POPF32));
2576 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2577 << " to " << RI.getName(DestReg) << '\n');
2578 llvm_unreachable("Cannot emit physreg copy instruction");
2581 static unsigned getLoadStoreRegOpcode(unsigned Reg,
2582 const TargetRegisterClass *RC,
2583 bool isStackAligned,
2584 const TargetMachine &TM,
2586 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2587 switch (RC->getSize()) {
2589 llvm_unreachable("Unknown spill size");
2591 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
2592 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2593 // Copying to or from a physical H register on x86-64 requires a NOREX
2594 // move. Otherwise use a normal move.
2595 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2596 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2597 return load ? X86::MOV8rm : X86::MOV8mr;
2599 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2600 return load ? X86::MOV16rm : X86::MOV16mr;
2602 if (X86::GR32RegClass.hasSubClassEq(RC))
2603 return load ? X86::MOV32rm : X86::MOV32mr;
2604 if (X86::FR32RegClass.hasSubClassEq(RC))
2606 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2607 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
2608 if (X86::RFP32RegClass.hasSubClassEq(RC))
2609 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2610 llvm_unreachable("Unknown 4-byte regclass");
2612 if (X86::GR64RegClass.hasSubClassEq(RC))
2613 return load ? X86::MOV64rm : X86::MOV64mr;
2614 if (X86::FR64RegClass.hasSubClassEq(RC))
2616 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2617 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
2618 if (X86::VR64RegClass.hasSubClassEq(RC))
2619 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2620 if (X86::RFP64RegClass.hasSubClassEq(RC))
2621 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2622 llvm_unreachable("Unknown 8-byte regclass");
2624 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
2625 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
2627 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
2628 // If stack is realigned we can use aligned stores.
2631 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2632 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
2635 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2636 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2639 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2640 // If stack is realigned we can use aligned stores.
2642 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2644 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
2648 static unsigned getStoreRegOpcode(unsigned SrcReg,
2649 const TargetRegisterClass *RC,
2650 bool isStackAligned,
2651 TargetMachine &TM) {
2652 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2656 static unsigned getLoadRegOpcode(unsigned DestReg,
2657 const TargetRegisterClass *RC,
2658 bool isStackAligned,
2659 const TargetMachine &TM) {
2660 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
2663 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2664 MachineBasicBlock::iterator MI,
2665 unsigned SrcReg, bool isKill, int FrameIdx,
2666 const TargetRegisterClass *RC,
2667 const TargetRegisterInfo *TRI) const {
2668 const MachineFunction &MF = *MBB.getParent();
2669 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2670 "Stack slot too small for store");
2671 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2672 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
2673 RI.canRealignStack(MF);
2674 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
2675 DebugLoc DL = MBB.findDebugLoc(MI);
2676 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
2677 .addReg(SrcReg, getKillRegState(isKill));
2680 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2682 SmallVectorImpl<MachineOperand> &Addr,
2683 const TargetRegisterClass *RC,
2684 MachineInstr::mmo_iterator MMOBegin,
2685 MachineInstr::mmo_iterator MMOEnd,
2686 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2687 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2688 bool isAligned = MMOBegin != MMOEnd &&
2689 (*MMOBegin)->getAlignment() >= Alignment;
2690 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
2692 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
2693 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2694 MIB.addOperand(Addr[i]);
2695 MIB.addReg(SrcReg, getKillRegState(isKill));
2696 (*MIB).setMemRefs(MMOBegin, MMOEnd);
2697 NewMIs.push_back(MIB);
2701 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
2702 MachineBasicBlock::iterator MI,
2703 unsigned DestReg, int FrameIdx,
2704 const TargetRegisterClass *RC,
2705 const TargetRegisterInfo *TRI) const {
2706 const MachineFunction &MF = *MBB.getParent();
2707 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2708 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
2709 RI.canRealignStack(MF);
2710 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2711 DebugLoc DL = MBB.findDebugLoc(MI);
2712 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
2715 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
2716 SmallVectorImpl<MachineOperand> &Addr,
2717 const TargetRegisterClass *RC,
2718 MachineInstr::mmo_iterator MMOBegin,
2719 MachineInstr::mmo_iterator MMOEnd,
2720 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2721 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2722 bool isAligned = MMOBegin != MMOEnd &&
2723 (*MMOBegin)->getAlignment() >= Alignment;
2724 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2726 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
2727 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2728 MIB.addOperand(Addr[i]);
2729 (*MIB).setMemRefs(MMOBegin, MMOEnd);
2730 NewMIs.push_back(MIB);
2734 AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpMask,
2735 int &CmpValue) const {
2736 switch (MI->getOpcode()) {
2738 case X86::CMP64ri32:
2745 SrcReg = MI->getOperand(0).getReg();
2747 CmpValue = MI->getOperand(1).getImm();
2753 SrcReg = MI->getOperand(0).getReg();
2762 // This function updates condition code for SET Opcodes.
2763 // The input condition code can be E,NE,L,LE,G,GE,B,BE,A,AE and
2764 // the updated condition code will be E,NE,G,GE,L,LE,A,AE,B,BE.
2765 // This is to convert from a > b to b < a, a >= b to b <= a etc.
2766 static unsigned UpdateSETCondToOptimizeCmp(unsigned SETOpc) {
2769 case X86::SETEr: return X86::SETEr;
2770 case X86::SETEm: return X86::SETEm;
2771 case X86::SETNEr: return X86::SETNEr;
2772 case X86::SETNEm: return X86::SETNEm;
2773 case X86::SETLr: return X86::SETGr;
2774 case X86::SETLm: return X86::SETGm;
2775 case X86::SETLEr: return X86::SETGEr;
2776 case X86::SETLEm: return X86::SETGEm;
2777 case X86::SETGr: return X86::SETLr;
2778 case X86::SETGm: return X86::SETLm;
2779 case X86::SETGEr: return X86::SETLEr;
2780 case X86::SETGEm: return X86::SETLEm;
2781 case X86::SETBr: return X86::SETAr;
2782 case X86::SETBm: return X86::SETAm;
2783 case X86::SETBEr: return X86::SETAEr;
2784 case X86::SETBEm: return X86::SETAEm;
2785 case X86::SETAr: return X86::SETBr;
2786 case X86::SETAm: return X86::SETBm;
2787 case X86::SETAEr: return X86::SETBEr;
2788 case X86::SETAEm: return X86::SETBEm;
2792 // This function updates condition code for Branch Opcodes.
2793 // The input condition code can be E,NE,L,LE,G,GE,B,BE,A,AE and
2794 // the updated condition code will be E,NE,G,GE,L,LE,A,AE,B,BE.
2795 // This is to convert from a > b to b < a, a >= b to b <= a etc.
2796 static unsigned UpdateBranchCondToOptimizeCmp(unsigned BranchOpc) {
2797 switch (BranchOpc) {
2799 case X86::JE_4: return X86::JE_4;
2800 case X86::JNE_4: return X86::JNE_4;
2801 case X86::JL_4: return X86::JG_4;
2802 case X86::JLE_4: return X86::JGE_4;
2803 case X86::JG_4: return X86::JL_4;
2804 case X86::JGE_4: return X86::JLE_4;
2805 case X86::JB_4: return X86::JA_4;
2806 case X86::JBE_4: return X86::JAE_4;
2807 case X86::JA_4: return X86::JB_4;
2808 case X86::JAE_4: return X86::JBE_4;
2812 // This function updates condition code for CMOV Opcodes.
2813 // The input condition code can be E,NE,L,LE,G,GE,B,BE,A,AE and
2814 // the updated condition code will be E,NE,G,GE,L,LE,A,AE,B,BE.
2815 // This is to convert from a > b to b < a, a >= b to b <= a etc.
2816 static unsigned UpdateCMovCondToOptimizeCmp(unsigned CMovOpc) {
2819 case X86::CMOVE16rm: return X86::CMOVE16rm;
2820 case X86::CMOVE16rr: return X86::CMOVE16rr;
2821 case X86::CMOVE32rm: return X86::CMOVE32rm;
2822 case X86::CMOVE32rr: return X86::CMOVE32rr;
2823 case X86::CMOVE64rm: return X86::CMOVE64rm;
2824 case X86::CMOVE64rr: return X86::CMOVE64rr;
2825 case X86::CMOVNE16rm: return X86::CMOVNE16rm;
2826 case X86::CMOVNE16rr: return X86::CMOVNE16rr;
2827 case X86::CMOVNE32rm: return X86::CMOVNE32rm;
2828 case X86::CMOVNE32rr: return X86::CMOVNE32rr;
2829 case X86::CMOVNE64rm: return X86::CMOVNE64rm;
2830 case X86::CMOVNE64rr: return X86::CMOVNE64rr;
2832 case X86::CMOVL16rm: return X86::CMOVG16rm;
2833 case X86::CMOVL16rr: return X86::CMOVG16rr;
2834 case X86::CMOVL32rm: return X86::CMOVG32rm;
2835 case X86::CMOVL32rr: return X86::CMOVG32rr;
2836 case X86::CMOVL64rm: return X86::CMOVG64rm;
2837 case X86::CMOVL64rr: return X86::CMOVG64rr;
2838 case X86::CMOVLE16rm: return X86::CMOVGE16rm;
2839 case X86::CMOVLE16rr: return X86::CMOVGE16rr;
2840 case X86::CMOVLE32rm: return X86::CMOVGE32rm;
2841 case X86::CMOVLE32rr: return X86::CMOVGE32rr;
2842 case X86::CMOVLE64rm: return X86::CMOVGE64rm;
2843 case X86::CMOVLE64rr: return X86::CMOVGE64rr;
2845 case X86::CMOVG16rm: return X86::CMOVL16rm;
2846 case X86::CMOVG16rr: return X86::CMOVL16rr;
2847 case X86::CMOVG32rm: return X86::CMOVL32rm;
2848 case X86::CMOVG32rr: return X86::CMOVL32rr;
2849 case X86::CMOVG64rm: return X86::CMOVL64rm;
2850 case X86::CMOVG64rr: return X86::CMOVL64rr;
2851 case X86::CMOVGE16rm: return X86::CMOVLE16rm;
2852 case X86::CMOVGE16rr: return X86::CMOVLE16rr;
2853 case X86::CMOVGE32rm: return X86::CMOVLE32rm;
2854 case X86::CMOVGE32rr: return X86::CMOVLE32rr;
2855 case X86::CMOVGE64rm: return X86::CMOVLE64rm;
2856 case X86::CMOVGE64rr: return X86::CMOVLE64rr;
2858 case X86::CMOVB16rm: return X86::CMOVA16rm;
2859 case X86::CMOVB16rr: return X86::CMOVA16rr;
2860 case X86::CMOVB32rm: return X86::CMOVA32rm;
2861 case X86::CMOVB32rr: return X86::CMOVA32rr;
2862 case X86::CMOVB64rm: return X86::CMOVA64rm;
2863 case X86::CMOVB64rr: return X86::CMOVA64rr;
2864 case X86::CMOVBE16rm: return X86::CMOVAE16rm;
2865 case X86::CMOVBE16rr: return X86::CMOVAE16rr;
2866 case X86::CMOVBE32rm: return X86::CMOVAE32rm;
2867 case X86::CMOVBE32rr: return X86::CMOVAE32rr;
2868 case X86::CMOVBE64rm: return X86::CMOVAE64rm;
2869 case X86::CMOVBE64rr: return X86::CMOVAE64rr;
2871 case X86::CMOVA16rm: return X86::CMOVB16rm;
2872 case X86::CMOVA16rr: return X86::CMOVB16rr;
2873 case X86::CMOVA32rm: return X86::CMOVB32rm;
2874 case X86::CMOVA32rr: return X86::CMOVB32rr;
2875 case X86::CMOVA64rm: return X86::CMOVB64rm;
2876 case X86::CMOVA64rr: return X86::CMOVB64rr;
2877 case X86::CMOVAE16rm: return X86::CMOVBE16rm;
2878 case X86::CMOVAE16rr: return X86::CMOVBE16rr;
2879 case X86::CMOVAE32rm: return X86::CMOVBE32rm;
2880 case X86::CMOVAE32rr: return X86::CMOVBE32rr;
2881 case X86::CMOVAE64rm: return X86::CMOVBE64rm;
2882 case X86::CMOVAE64rr: return X86::CMOVBE64rr;
2887 OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, int CmpMask,
2888 int CmpValue, const MachineRegisterInfo *MRI) const {
2889 MachineRegisterInfo::def_iterator DI = MRI->def_begin(SrcReg);
2890 if (llvm::next(DI) != MRI->def_end())
2891 // Only support one definition.
2894 MachineInstr *MI = &*DI;
2895 // Get ready to iterate backward from CmpInstr.
2896 MachineBasicBlock::iterator I = CmpInstr, E = MI,
2897 B = CmpInstr->getParent()->begin();
2899 // Early exit if CmpInstr is at the beginning of the BB.
2900 if (I == B) return false;
2902 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1).
2903 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue).
2904 MachineInstr *Sub = NULL;
2905 unsigned SrcReg2 = 0;
2906 if (CmpInstr->getOpcode() == X86::CMP64rr ||
2907 CmpInstr->getOpcode() == X86::CMP32rr ||
2908 CmpInstr->getOpcode() == X86::CMP16rr ||
2909 CmpInstr->getOpcode() == X86::CMP8rr) {
2910 SrcReg2 = CmpInstr->getOperand(1).getReg();
2913 // Search backwards for a SUB instruction.
2914 // If EFLAGS is updated in the middle, we can not remove the CMP instruction.
2916 for (; I != E; --I) {
2917 const MachineInstr &Instr = *I;
2919 // Check whether the current instruction is SUB(r1, r2) or SUB(r2, r1).
2920 if (((CmpInstr->getOpcode() == X86::CMP64rr &&
2921 Instr.getOpcode() == X86::SUB64rr) ||
2922 (CmpInstr->getOpcode() == X86::CMP32rr &&
2923 Instr.getOpcode() == X86::SUB32rr)||
2924 (CmpInstr->getOpcode() == X86::CMP16rr &&
2925 Instr.getOpcode() == X86::SUB16rr)||
2926 (CmpInstr->getOpcode() == X86::CMP8rr &&
2927 Instr.getOpcode() == X86::SUB8rr)) &&
2928 ((Instr.getOperand(1).getReg() == SrcReg &&
2929 Instr.getOperand(2).getReg() == SrcReg2) ||
2930 (Instr.getOperand(1).getReg() == SrcReg2 &&
2931 Instr.getOperand(2).getReg() == SrcReg))) {
2936 // Check whether the current instruction is SUBri(r1, CmpValue).
2937 if (((CmpInstr->getOpcode() == X86::CMP64ri32 &&
2938 Instr.getOpcode() == X86::SUB64ri32) ||
2939 (CmpInstr->getOpcode() == X86::CMP64ri8 &&
2940 Instr.getOpcode() == X86::SUB64ri8) ||
2941 (CmpInstr->getOpcode() == X86::CMP32ri &&
2942 Instr.getOpcode() == X86::SUB32ri) ||
2943 (CmpInstr->getOpcode() == X86::CMP32ri8 &&
2944 Instr.getOpcode() == X86::SUB32ri8) ||
2945 (CmpInstr->getOpcode() == X86::CMP16ri &&
2946 Instr.getOpcode() == X86::SUB16ri) ||
2947 (CmpInstr->getOpcode() == X86::CMP16ri8 &&
2948 Instr.getOpcode() == X86::SUB16ri8) ||
2949 (CmpInstr->getOpcode() == X86::CMP8ri &&
2950 Instr.getOpcode() == X86::SUB8ri)) &&
2952 Instr.getOperand(1).getReg() == SrcReg &&
2953 Instr.getOperand(2).getImm() == CmpValue) {
2958 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) {
2959 const MachineOperand &MO = Instr.getOperand(IO);
2960 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2962 if (!MO.isReg()) continue;
2964 // This instruction modifies or uses EFLAGS before we find a SUB.
2965 // We can't do this transformation.
2966 if (MO.getReg() == X86::EFLAGS)
2971 // Reaching beginning of the block.
2975 // Return false if no candidates exist.
2980 switch (MI->getOpcode()) {
2986 case X86::SUB64ri32:
2993 // Scan forward from CmpInstr for the use of EFLAGS.
2994 // Handle the condition codes GE, L, G, LE, B, L, BE, LE.
2995 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
2996 bool isSafe = false;
2998 E = CmpInstr->getParent()->end();
2999 while (!isSafe && ++I != E) {
3000 const MachineInstr &Instr = *I;
3001 for (unsigned IO = 0, EO = Instr.getNumOperands();
3002 !isSafe && IO != EO; ++IO) {
3003 const MachineOperand &MO = Instr.getOperand(IO);
3004 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) {
3008 if (!MO.isReg() || MO.getReg() != X86::EFLAGS)
3010 // EFLAGS is redefined by this instruction.
3015 // EFLAGS is used by this instruction.
3017 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3018 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3019 unsigned NewOpc = UpdateBranchCondToOptimizeCmp(Instr.getOpcode());
3020 if (!NewOpc) NewOpc = UpdateSETCondToOptimizeCmp(Instr.getOpcode());
3021 if (!NewOpc) NewOpc = UpdateCMovCondToOptimizeCmp(Instr.getOpcode());
3022 if (!NewOpc) return false;
3024 // Push the MachineInstr to OpsToUpdate.
3025 // If it is safe to remove CmpInstr, the condition code of these
3026 // instructions will be modified.
3027 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3028 Sub->getOperand(2).getReg() == SrcReg)
3029 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3033 // We may exit the loop at end of the basic block.
3034 // In that case, it is still safe to remove CmpInstr.
3036 // Make sure Sub instruction defines EFLAGS.
3037 assert(MI->getOperand(3).getReg() == X86::EFLAGS &&
3038 "Expected EFLAGS is the 4th operand of SUBrr or SUBri.");
3039 MI->getOperand(3).setIsDef(true);
3040 CmpInstr->eraseFromParent();
3042 // Modify the condition code of instructions in OpsToUpdate.
3043 for (unsigned i = 0; i < OpsToUpdate.size(); i++)
3044 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3054 OptimizeSubInstr(MachineInstr *SubInstr, const MachineRegisterInfo *MRI) const {
3055 // If destination is a memory operand, do not perform this optimization.
3056 if ((SubInstr->getOpcode() != X86::SUB64rr) &&
3057 (SubInstr->getOpcode() != X86::SUB32rr) &&
3058 (SubInstr->getOpcode() != X86::SUB16rr) &&
3059 (SubInstr->getOpcode() != X86::SUB8rr) &&
3060 (SubInstr->getOpcode() != X86::SUB64ri32) &&
3061 (SubInstr->getOpcode() != X86::SUB64ri8) &&
3062 (SubInstr->getOpcode() != X86::SUB32ri) &&
3063 (SubInstr->getOpcode() != X86::SUB32ri8) &&
3064 (SubInstr->getOpcode() != X86::SUB16ri) &&
3065 (SubInstr->getOpcode() != X86::SUB16ri8) &&
3066 (SubInstr->getOpcode() != X86::SUB8ri))
3068 unsigned DestReg = SubInstr->getOperand(0).getReg();
3069 if (MRI->use_begin(DestReg) != MRI->use_end())
3072 // There is no use of the destination register, we can replace SUB with CMP.
3073 switch (SubInstr->getOpcode()) {
3075 case X86::SUB64rr: SubInstr->setDesc(get(X86::CMP64rr)); break;
3076 case X86::SUB32rr: SubInstr->setDesc(get(X86::CMP32rr)); break;
3077 case X86::SUB16rr: SubInstr->setDesc(get(X86::CMP16rr)); break;
3078 case X86::SUB8rr: SubInstr->setDesc(get(X86::CMP8rr)); break;
3079 case X86::SUB64ri32: SubInstr->setDesc(get(X86::CMP64ri32)); break;
3080 case X86::SUB64ri8: SubInstr->setDesc(get(X86::CMP64ri8)); break;
3081 case X86::SUB32ri: SubInstr->setDesc(get(X86::CMP32ri)); break;
3082 case X86::SUB32ri8: SubInstr->setDesc(get(X86::CMP32ri8)); break;
3083 case X86::SUB16ri: SubInstr->setDesc(get(X86::CMP16ri)); break;
3084 case X86::SUB16ri8: SubInstr->setDesc(get(X86::CMP16ri8)); break;
3085 case X86::SUB8ri: SubInstr->setDesc(get(X86::CMP8ri)); break;
3087 SubInstr->RemoveOperand(0);
3091 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3092 /// instruction with two undef reads of the register being defined. This is
3093 /// used for mapping:
3096 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3098 static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
3099 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3100 unsigned Reg = MI->getOperand(0).getReg();
3103 // MachineInstr::addOperand() will insert explicit operands before any
3104 // implicit operands.
3105 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
3106 .addReg(Reg, RegState::Undef);
3107 // But we don't trust that.
3108 assert(MI->getOperand(1).getReg() == Reg &&
3109 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
3113 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3114 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3115 switch (MI->getOpcode()) {
3119 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
3120 case X86::TEST8ri_NOREX:
3121 MI->setDesc(get(X86::TEST8ri));
3128 X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
3129 int FrameIx, uint64_t Offset,
3130 const MDNode *MDPtr,
3131 DebugLoc DL) const {
3133 AM.BaseType = X86AddressMode::FrameIndexBase;
3134 AM.Base.FrameIndex = FrameIx;
3135 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3136 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3140 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
3141 const SmallVectorImpl<MachineOperand> &MOs,
3143 const TargetInstrInfo &TII) {
3144 // Create the base instruction with the memory operand as the first part.
3145 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3146 MI->getDebugLoc(), true);
3147 MachineInstrBuilder MIB(NewMI);
3148 unsigned NumAddrOps = MOs.size();
3149 for (unsigned i = 0; i != NumAddrOps; ++i)
3150 MIB.addOperand(MOs[i]);
3151 if (NumAddrOps < 4) // FrameIndex only
3154 // Loop over the rest of the ri operands, converting them over.
3155 unsigned NumOps = MI->getDesc().getNumOperands()-2;
3156 for (unsigned i = 0; i != NumOps; ++i) {
3157 MachineOperand &MO = MI->getOperand(i+2);
3160 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3161 MachineOperand &MO = MI->getOperand(i);
3167 static MachineInstr *FuseInst(MachineFunction &MF,
3168 unsigned Opcode, unsigned OpNo,
3169 const SmallVectorImpl<MachineOperand> &MOs,
3170 MachineInstr *MI, const TargetInstrInfo &TII) {
3171 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3172 MI->getDebugLoc(), true);
3173 MachineInstrBuilder MIB(NewMI);
3175 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3176 MachineOperand &MO = MI->getOperand(i);
3178 assert(MO.isReg() && "Expected to fold into reg operand!");
3179 unsigned NumAddrOps = MOs.size();
3180 for (unsigned i = 0; i != NumAddrOps; ++i)
3181 MIB.addOperand(MOs[i]);
3182 if (NumAddrOps < 4) // FrameIndex only
3191 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
3192 const SmallVectorImpl<MachineOperand> &MOs,
3194 MachineFunction &MF = *MI->getParent()->getParent();
3195 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
3197 unsigned NumAddrOps = MOs.size();
3198 for (unsigned i = 0; i != NumAddrOps; ++i)
3199 MIB.addOperand(MOs[i]);
3200 if (NumAddrOps < 4) // FrameIndex only
3202 return MIB.addImm(0);
3206 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3207 MachineInstr *MI, unsigned i,
3208 const SmallVectorImpl<MachineOperand> &MOs,
3209 unsigned Size, unsigned Align) const {
3210 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
3211 bool isTwoAddrFold = false;
3212 unsigned NumOps = MI->getDesc().getNumOperands();
3213 bool isTwoAddr = NumOps > 1 &&
3214 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
3216 // FIXME: AsmPrinter doesn't know how to handle
3217 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3218 if (MI->getOpcode() == X86::ADD32ri &&
3219 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3222 MachineInstr *NewMI = NULL;
3223 // Folding a memory location into the two-address part of a two-address
3224 // instruction is different than folding it other places. It requires
3225 // replacing the *two* registers with the memory location.
3226 if (isTwoAddr && NumOps >= 2 && i < 2 &&
3227 MI->getOperand(0).isReg() &&
3228 MI->getOperand(1).isReg() &&
3229 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
3230 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3231 isTwoAddrFold = true;
3232 } else if (i == 0) { // If operand 0
3233 if (MI->getOpcode() == X86::MOV64r0)
3234 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
3235 else if (MI->getOpcode() == X86::MOV32r0)
3236 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
3237 else if (MI->getOpcode() == X86::MOV16r0)
3238 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
3239 else if (MI->getOpcode() == X86::MOV8r0)
3240 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
3244 OpcodeTablePtr = &RegOp2MemOpTable0;
3245 } else if (i == 1) {
3246 OpcodeTablePtr = &RegOp2MemOpTable1;
3247 } else if (i == 2) {
3248 OpcodeTablePtr = &RegOp2MemOpTable2;
3251 // If table selected...
3252 if (OpcodeTablePtr) {
3253 // Find the Opcode to fuse
3254 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3255 OpcodeTablePtr->find(MI->getOpcode());
3256 if (I != OpcodeTablePtr->end()) {
3257 unsigned Opcode = I->second.first;
3258 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
3259 if (Align < MinAlign)
3261 bool NarrowToMOV32rm = false;
3263 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
3264 if (Size < RCSize) {
3265 // Check if it's safe to fold the load. If the size of the object is
3266 // narrower than the load width, then it's not.
3267 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3269 // If this is a 64-bit load, but the spill slot is 32, then we can do
3270 // a 32-bit load which is implicitly zero-extended. This likely is due
3271 // to liveintervalanalysis remat'ing a load from stack slot.
3272 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3274 Opcode = X86::MOV32rm;
3275 NarrowToMOV32rm = true;
3280 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
3282 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
3284 if (NarrowToMOV32rm) {
3285 // If this is the special case where we use a MOV32rm to load a 32-bit
3286 // value and zero-extend the top bits. Change the destination register
3288 unsigned DstReg = NewMI->getOperand(0).getReg();
3289 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3290 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
3293 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
3300 if (PrintFailedFusing && !MI->isCopy())
3301 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
3305 /// hasPartialRegUpdate - Return true for all instructions that only update
3306 /// the first 32 or 64-bits of the destination register and leave the rest
3307 /// unmodified. This can be used to avoid folding loads if the instructions
3308 /// only update part of the destination register, and the non-updated part is
3309 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3310 /// instructions breaks the partial register dependency and it can improve
3311 /// performance. e.g.:
3313 /// movss (%rdi), %xmm0
3314 /// cvtss2sd %xmm0, %xmm0
3317 /// cvtss2sd (%rdi), %xmm0
3319 /// FIXME: This should be turned into a TSFlags.
3321 static bool hasPartialRegUpdate(unsigned Opcode) {
3323 case X86::CVTSI2SSrr:
3324 case X86::CVTSI2SS64rr:
3325 case X86::CVTSI2SDrr:
3326 case X86::CVTSI2SD64rr:
3327 case X86::CVTSD2SSrr:
3328 case X86::Int_CVTSD2SSrr:
3329 case X86::CVTSS2SDrr:
3330 case X86::Int_CVTSS2SDrr:
3332 case X86::RCPSSr_Int:
3334 case X86::ROUNDSDr_Int:
3336 case X86::ROUNDSSr_Int:
3338 case X86::RSQRTSSr_Int:
3340 case X86::SQRTSSr_Int:
3341 // AVX encoded versions
3342 case X86::VCVTSD2SSrr:
3343 case X86::Int_VCVTSD2SSrr:
3344 case X86::VCVTSS2SDrr:
3345 case X86::Int_VCVTSS2SDrr:
3347 case X86::VROUNDSDr:
3348 case X86::VROUNDSDr_Int:
3349 case X86::VROUNDSSr:
3350 case X86::VROUNDSSr_Int:
3351 case X86::VRSQRTSSr:
3359 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3360 /// instructions we would like before a partial register update.
3361 unsigned X86InstrInfo::
3362 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3363 const TargetRegisterInfo *TRI) const {
3364 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3367 // If MI is marked as reading Reg, the partial register update is wanted.
3368 const MachineOperand &MO = MI->getOperand(0);
3369 unsigned Reg = MO.getReg();
3370 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3371 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3374 if (MI->readsRegister(Reg, TRI))
3378 // If any of the preceding 16 instructions are reading Reg, insert a
3379 // dependency breaking instruction. The magic number is based on a few
3380 // Nehalem experiments.
3385 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3386 const TargetRegisterInfo *TRI) const {
3387 unsigned Reg = MI->getOperand(OpNum).getReg();
3388 if (X86::VR128RegClass.contains(Reg)) {
3389 // These instructions are all floating point domain, so xorps is the best
3391 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3392 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3393 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3394 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3395 } else if (X86::VR256RegClass.contains(Reg)) {
3396 // Use vxorps to clear the full ymm register.
3397 // It wants to read and write the xmm sub-register.
3398 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3399 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3400 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3401 .addReg(Reg, RegState::ImplicitDefine);
3404 MI->addRegisterKilled(Reg, TRI, true);
3407 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3409 const SmallVectorImpl<unsigned> &Ops,
3410 int FrameIndex) const {
3411 // Check switch flag
3412 if (NoFusing) return NULL;
3414 // Unless optimizing for size, don't fold to avoid partial
3415 // register update stalls
3416 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3417 hasPartialRegUpdate(MI->getOpcode()))
3420 const MachineFrameInfo *MFI = MF.getFrameInfo();
3421 unsigned Size = MFI->getObjectSize(FrameIndex);
3422 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
3423 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3424 unsigned NewOpc = 0;
3425 unsigned RCSize = 0;
3426 switch (MI->getOpcode()) {
3427 default: return NULL;
3428 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
3429 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3430 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3431 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
3433 // Check if it's safe to fold the load. If the size of the object is
3434 // narrower than the load width, then it's not.
3437 // Change to CMPXXri r, 0 first.
3438 MI->setDesc(get(NewOpc));
3439 MI->getOperand(1).ChangeToImmediate(0);
3440 } else if (Ops.size() != 1)
3443 SmallVector<MachineOperand,4> MOs;
3444 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
3445 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
3448 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3450 const SmallVectorImpl<unsigned> &Ops,
3451 MachineInstr *LoadMI) const {
3452 // Check switch flag
3453 if (NoFusing) return NULL;
3455 // Unless optimizing for size, don't fold to avoid partial
3456 // register update stalls
3457 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3458 hasPartialRegUpdate(MI->getOpcode()))
3461 // Determine the alignment of the load.
3462 unsigned Alignment = 0;
3463 if (LoadMI->hasOneMemOperand())
3464 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
3466 switch (LoadMI->getOpcode()) {
3467 case X86::AVX_SET0PSY:
3468 case X86::AVX_SET0PDY:
3469 case X86::AVX2_SETALLONES:
3470 case X86::AVX2_SET0:
3474 case X86::V_SETALLONES:
3475 case X86::AVX_SETALLONES:
3487 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3488 unsigned NewOpc = 0;
3489 switch (MI->getOpcode()) {
3490 default: return NULL;
3491 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
3492 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3493 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3494 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
3496 // Change to CMPXXri r, 0 first.
3497 MI->setDesc(get(NewOpc));
3498 MI->getOperand(1).ChangeToImmediate(0);
3499 } else if (Ops.size() != 1)
3502 // Make sure the subregisters match.
3503 // Otherwise we risk changing the size of the load.
3504 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3507 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
3508 switch (LoadMI->getOpcode()) {
3510 case X86::V_SETALLONES:
3511 case X86::AVX_SET0PSY:
3512 case X86::AVX_SET0PDY:
3513 case X86::AVX_SETALLONES:
3514 case X86::AVX2_SETALLONES:
3515 case X86::AVX2_SET0:
3517 case X86::FsFLD0SS: {
3518 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
3519 // Create a constant-pool entry and operands to load from it.
3521 // Medium and large mode can't fold loads this way.
3522 if (TM.getCodeModel() != CodeModel::Small &&
3523 TM.getCodeModel() != CodeModel::Kernel)
3526 // x86-32 PIC requires a PIC base register for constant pools.
3527 unsigned PICBase = 0;
3528 if (TM.getRelocationModel() == Reloc::PIC_) {
3529 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3532 // FIXME: PICBase = getGlobalBaseReg(&MF);
3533 // This doesn't work for several reasons.
3534 // 1. GlobalBaseReg may have been spilled.
3535 // 2. It may not be live at MI.
3539 // Create a constant-pool entry.
3540 MachineConstantPool &MCP = *MF.getConstantPool();
3542 unsigned Opc = LoadMI->getOpcode();
3543 if (Opc == X86::FsFLD0SS)
3544 Ty = Type::getFloatTy(MF.getFunction()->getContext());
3545 else if (Opc == X86::FsFLD0SD)
3546 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
3547 else if (Opc == X86::AVX_SET0PSY || Opc == X86::AVX_SET0PDY)
3548 Ty = VectorType::get(Type::getFloatTy(MF.getFunction()->getContext()), 8);
3549 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX2_SET0)
3550 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
3552 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
3554 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX_SETALLONES ||
3555 Opc == X86::AVX2_SETALLONES);
3556 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3557 Constant::getNullValue(Ty);
3558 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
3560 // Create operands to load from the constant pool entry.
3561 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3562 MOs.push_back(MachineOperand::CreateImm(1));
3563 MOs.push_back(MachineOperand::CreateReg(0, false));
3564 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
3565 MOs.push_back(MachineOperand::CreateReg(0, false));
3569 // Folding a normal load. Just copy the load's address operands.
3570 unsigned NumOps = LoadMI->getDesc().getNumOperands();
3571 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
3572 MOs.push_back(LoadMI->getOperand(i));
3576 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
3580 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3581 const SmallVectorImpl<unsigned> &Ops) const {
3582 // Check switch flag
3583 if (NoFusing) return 0;
3585 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3586 switch (MI->getOpcode()) {
3587 default: return false;
3594 // FIXME: AsmPrinter doesn't know how to handle
3595 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3596 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3602 if (Ops.size() != 1)
3605 unsigned OpNum = Ops[0];
3606 unsigned Opc = MI->getOpcode();
3607 unsigned NumOps = MI->getDesc().getNumOperands();
3608 bool isTwoAddr = NumOps > 1 &&
3609 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
3611 // Folding a memory location into the two-address part of a two-address
3612 // instruction is different than folding it other places. It requires
3613 // replacing the *two* registers with the memory location.
3614 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
3615 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
3616 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3617 } else if (OpNum == 0) { // If operand 0
3622 case X86::MOV64r0: return true;
3625 OpcodeTablePtr = &RegOp2MemOpTable0;
3626 } else if (OpNum == 1) {
3627 OpcodeTablePtr = &RegOp2MemOpTable1;
3628 } else if (OpNum == 2) {
3629 OpcodeTablePtr = &RegOp2MemOpTable2;
3632 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
3634 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
3637 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
3638 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
3639 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3640 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3641 MemOp2RegOpTable.find(MI->getOpcode());
3642 if (I == MemOp2RegOpTable.end())
3644 unsigned Opc = I->second.first;
3645 unsigned Index = I->second.second & TB_INDEX_MASK;
3646 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3647 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
3648 if (UnfoldLoad && !FoldedLoad)
3650 UnfoldLoad &= FoldedLoad;
3651 if (UnfoldStore && !FoldedStore)
3653 UnfoldStore &= FoldedStore;
3655 const MCInstrDesc &MCID = get(Opc);
3656 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
3657 if (!MI->hasOneMemOperand() &&
3658 RC == &X86::VR128RegClass &&
3659 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3660 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
3661 // conservatively assume the address is unaligned. That's bad for
3664 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
3665 SmallVector<MachineOperand,2> BeforeOps;
3666 SmallVector<MachineOperand,2> AfterOps;
3667 SmallVector<MachineOperand,4> ImpOps;
3668 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3669 MachineOperand &Op = MI->getOperand(i);
3670 if (i >= Index && i < Index + X86::AddrNumOperands)
3671 AddrOps.push_back(Op);
3672 else if (Op.isReg() && Op.isImplicit())
3673 ImpOps.push_back(Op);
3675 BeforeOps.push_back(Op);
3677 AfterOps.push_back(Op);
3680 // Emit the load instruction.
3682 std::pair<MachineInstr::mmo_iterator,
3683 MachineInstr::mmo_iterator> MMOs =
3684 MF.extractLoadMemRefs(MI->memoperands_begin(),
3685 MI->memoperands_end());
3686 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
3688 // Address operands cannot be marked isKill.
3689 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
3690 MachineOperand &MO = NewMIs[0]->getOperand(i);
3692 MO.setIsKill(false);
3697 // Emit the data processing instruction.
3698 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
3699 MachineInstrBuilder MIB(DataMI);
3702 MIB.addReg(Reg, RegState::Define);
3703 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
3704 MIB.addOperand(BeforeOps[i]);
3707 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
3708 MIB.addOperand(AfterOps[i]);
3709 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
3710 MachineOperand &MO = ImpOps[i];
3711 MIB.addReg(MO.getReg(),
3712 getDefRegState(MO.isDef()) |
3713 RegState::Implicit |
3714 getKillRegState(MO.isKill()) |
3715 getDeadRegState(MO.isDead()) |
3716 getUndefRegState(MO.isUndef()));
3718 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
3719 unsigned NewOpc = 0;
3720 switch (DataMI->getOpcode()) {
3722 case X86::CMP64ri32:
3729 MachineOperand &MO0 = DataMI->getOperand(0);
3730 MachineOperand &MO1 = DataMI->getOperand(1);
3731 if (MO1.getImm() == 0) {
3732 switch (DataMI->getOpcode()) {
3735 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
3737 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
3739 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
3740 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
3742 DataMI->setDesc(get(NewOpc));
3743 MO1.ChangeToRegister(MO0.getReg(), false);
3747 NewMIs.push_back(DataMI);
3749 // Emit the store instruction.
3751 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
3752 std::pair<MachineInstr::mmo_iterator,
3753 MachineInstr::mmo_iterator> MMOs =
3754 MF.extractStoreMemRefs(MI->memoperands_begin(),
3755 MI->memoperands_end());
3756 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
3763 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
3764 SmallVectorImpl<SDNode*> &NewNodes) const {
3765 if (!N->isMachineOpcode())
3768 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3769 MemOp2RegOpTable.find(N->getMachineOpcode());
3770 if (I == MemOp2RegOpTable.end())
3772 unsigned Opc = I->second.first;
3773 unsigned Index = I->second.second & TB_INDEX_MASK;
3774 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3775 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
3776 const MCInstrDesc &MCID = get(Opc);
3777 MachineFunction &MF = DAG.getMachineFunction();
3778 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
3779 unsigned NumDefs = MCID.NumDefs;
3780 std::vector<SDValue> AddrOps;
3781 std::vector<SDValue> BeforeOps;
3782 std::vector<SDValue> AfterOps;
3783 DebugLoc dl = N->getDebugLoc();
3784 unsigned NumOps = N->getNumOperands();
3785 for (unsigned i = 0; i != NumOps-1; ++i) {
3786 SDValue Op = N->getOperand(i);
3787 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
3788 AddrOps.push_back(Op);
3789 else if (i < Index-NumDefs)
3790 BeforeOps.push_back(Op);
3791 else if (i > Index-NumDefs)
3792 AfterOps.push_back(Op);
3794 SDValue Chain = N->getOperand(NumOps-1);
3795 AddrOps.push_back(Chain);
3797 // Emit the load instruction.
3800 EVT VT = *RC->vt_begin();
3801 std::pair<MachineInstr::mmo_iterator,
3802 MachineInstr::mmo_iterator> MMOs =
3803 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
3804 cast<MachineSDNode>(N)->memoperands_end());
3805 if (!(*MMOs.first) &&
3806 RC == &X86::VR128RegClass &&
3807 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3808 // Do not introduce a slow unaligned load.
3810 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3811 bool isAligned = (*MMOs.first) &&
3812 (*MMOs.first)->getAlignment() >= Alignment;
3813 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
3814 VT, MVT::Other, &AddrOps[0], AddrOps.size());
3815 NewNodes.push_back(Load);
3817 // Preserve memory reference information.
3818 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
3821 // Emit the data processing instruction.
3822 std::vector<EVT> VTs;
3823 const TargetRegisterClass *DstRC = 0;
3824 if (MCID.getNumDefs() > 0) {
3825 DstRC = getRegClass(MCID, 0, &RI, MF);
3826 VTs.push_back(*DstRC->vt_begin());
3828 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
3829 EVT VT = N->getValueType(i);
3830 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
3834 BeforeOps.push_back(SDValue(Load, 0));
3835 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
3836 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
3838 NewNodes.push_back(NewNode);
3840 // Emit the store instruction.
3843 AddrOps.push_back(SDValue(NewNode, 0));
3844 AddrOps.push_back(Chain);
3845 std::pair<MachineInstr::mmo_iterator,
3846 MachineInstr::mmo_iterator> MMOs =
3847 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
3848 cast<MachineSDNode>(N)->memoperands_end());
3849 if (!(*MMOs.first) &&
3850 RC == &X86::VR128RegClass &&
3851 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3852 // Do not introduce a slow unaligned store.
3854 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3855 bool isAligned = (*MMOs.first) &&
3856 (*MMOs.first)->getAlignment() >= Alignment;
3857 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
3860 &AddrOps[0], AddrOps.size());
3861 NewNodes.push_back(Store);
3863 // Preserve memory reference information.
3864 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
3870 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
3871 bool UnfoldLoad, bool UnfoldStore,
3872 unsigned *LoadRegIndex) const {
3873 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3874 MemOp2RegOpTable.find(Opc);
3875 if (I == MemOp2RegOpTable.end())
3877 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3878 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
3879 if (UnfoldLoad && !FoldedLoad)
3881 if (UnfoldStore && !FoldedStore)
3884 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
3885 return I->second.first;
3889 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
3890 int64_t &Offset1, int64_t &Offset2) const {
3891 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
3893 unsigned Opc1 = Load1->getMachineOpcode();
3894 unsigned Opc2 = Load2->getMachineOpcode();
3896 default: return false;
3906 case X86::MMX_MOVD64rm:
3907 case X86::MMX_MOVQ64rm:
3908 case X86::FsMOVAPSrm:
3909 case X86::FsMOVAPDrm:
3915 // AVX load instructions
3918 case X86::FsVMOVAPSrm:
3919 case X86::FsVMOVAPDrm:
3920 case X86::VMOVAPSrm:
3921 case X86::VMOVUPSrm:
3922 case X86::VMOVAPDrm:
3923 case X86::VMOVDQArm:
3924 case X86::VMOVDQUrm:
3925 case X86::VMOVAPSYrm:
3926 case X86::VMOVUPSYrm:
3927 case X86::VMOVAPDYrm:
3928 case X86::VMOVDQAYrm:
3929 case X86::VMOVDQUYrm:
3933 default: return false;
3943 case X86::MMX_MOVD64rm:
3944 case X86::MMX_MOVQ64rm:
3945 case X86::FsMOVAPSrm:
3946 case X86::FsMOVAPDrm:
3952 // AVX load instructions
3955 case X86::FsVMOVAPSrm:
3956 case X86::FsVMOVAPDrm:
3957 case X86::VMOVAPSrm:
3958 case X86::VMOVUPSrm:
3959 case X86::VMOVAPDrm:
3960 case X86::VMOVDQArm:
3961 case X86::VMOVDQUrm:
3962 case X86::VMOVAPSYrm:
3963 case X86::VMOVUPSYrm:
3964 case X86::VMOVAPDYrm:
3965 case X86::VMOVDQAYrm:
3966 case X86::VMOVDQUYrm:
3970 // Check if chain operands and base addresses match.
3971 if (Load1->getOperand(0) != Load2->getOperand(0) ||
3972 Load1->getOperand(5) != Load2->getOperand(5))
3974 // Segment operands should match as well.
3975 if (Load1->getOperand(4) != Load2->getOperand(4))
3977 // Scale should be 1, Index should be Reg0.
3978 if (Load1->getOperand(1) == Load2->getOperand(1) &&
3979 Load1->getOperand(2) == Load2->getOperand(2)) {
3980 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
3983 // Now let's examine the displacements.
3984 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
3985 isa<ConstantSDNode>(Load2->getOperand(3))) {
3986 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
3987 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
3994 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
3995 int64_t Offset1, int64_t Offset2,
3996 unsigned NumLoads) const {
3997 assert(Offset2 > Offset1);
3998 if ((Offset2 - Offset1) / 8 > 64)
4001 unsigned Opc1 = Load1->getMachineOpcode();
4002 unsigned Opc2 = Load2->getMachineOpcode();
4004 return false; // FIXME: overly conservative?
4011 case X86::MMX_MOVD64rm:
4012 case X86::MMX_MOVQ64rm:
4016 EVT VT = Load1->getValueType(0);
4017 switch (VT.getSimpleVT().SimpleTy) {
4019 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4020 // have 16 of them to play with.
4021 if (TM.getSubtargetImpl()->is64Bit()) {
4024 } else if (NumLoads) {
4044 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
4045 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
4046 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
4047 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4049 Cond[0].setImm(GetOppositeBranchCondition(CC));
4054 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4055 // FIXME: Return false for x87 stack register classes for now. We can't
4056 // allow any loads of these registers before FpGet_ST0_80.
4057 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4058 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
4061 /// getGlobalBaseReg - Return a virtual register initialized with the
4062 /// the global base register value. Output instructions required to
4063 /// initialize the register in the function entry block, if necessary.
4065 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4067 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4068 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4069 "X86-64 PIC uses RIP relative addressing");
4071 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4072 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4073 if (GlobalBaseReg != 0)
4074 return GlobalBaseReg;
4076 // Create the register. The code to initialize it is inserted
4077 // later, by the CGBR pass (below).
4078 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4079 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
4080 X86FI->setGlobalBaseReg(GlobalBaseReg);
4081 return GlobalBaseReg;
4084 // These are the replaceable SSE instructions. Some of these have Int variants
4085 // that we don't include here. We don't want to replace instructions selected
4087 static const uint16_t ReplaceableInstrs[][3] = {
4088 //PackedSingle PackedDouble PackedInt
4089 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4090 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4091 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4092 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4093 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4094 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4095 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4096 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4097 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4098 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4099 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4100 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4101 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4102 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
4103 // AVX 128-bit support
4104 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4105 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4106 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4107 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4108 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4109 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4110 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4111 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4112 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4113 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4114 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4115 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
4116 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4117 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
4118 // AVX 256-bit support
4119 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4120 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4121 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4122 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4123 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
4124 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4127 static const uint16_t ReplaceableInstrsAVX2[][3] = {
4128 //PackedSingle PackedDouble PackedInt
4129 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4130 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4131 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4132 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4133 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4134 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4135 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
4136 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4137 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4138 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4139 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4140 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4141 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4142 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
4145 // FIXME: Some shuffle and unpack instructions have equivalents in different
4146 // domains, but they require a bit more work than just switching opcodes.
4148 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
4149 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
4150 if (ReplaceableInstrs[i][domain-1] == opcode)
4151 return ReplaceableInstrs[i];
4155 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
4156 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4157 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4158 return ReplaceableInstrsAVX2[i];
4162 std::pair<uint16_t, uint16_t>
4163 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
4164 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4165 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
4166 uint16_t validDomains = 0;
4167 if (domain && lookup(MI->getOpcode(), domain))
4169 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4170 validDomains = hasAVX2 ? 0xe : 0x6;
4171 return std::make_pair(domain, validDomains);
4174 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
4175 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4176 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4177 assert(dom && "Not an SSE instruction");
4178 const uint16_t *table = lookup(MI->getOpcode(), dom);
4179 if (!table) { // try the other table
4180 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4181 "256-bit vector operations only available in AVX2");
4182 table = lookupAVX2(MI->getOpcode(), dom);
4184 assert(table && "Cannot change domain");
4185 MI->setDesc(get(table[Domain-1]));
4188 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4189 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4190 NopInst.setOpcode(X86::NOOP);
4193 bool X86InstrInfo::isHighLatencyDef(int opc) const {
4195 default: return false;
4197 case X86::DIVSDrm_Int:
4199 case X86::DIVSDrr_Int:
4201 case X86::DIVSSrm_Int:
4203 case X86::DIVSSrr_Int:
4205 case X86::SQRTPDm_Int:
4207 case X86::SQRTPDr_Int:
4209 case X86::SQRTPSm_Int:
4211 case X86::SQRTPSr_Int:
4213 case X86::SQRTSDm_Int:
4215 case X86::SQRTSDr_Int:
4217 case X86::SQRTSSm_Int:
4219 case X86::SQRTSSr_Int:
4220 // AVX instructions with high latency
4222 case X86::VDIVSDrm_Int:
4224 case X86::VDIVSDrr_Int:
4226 case X86::VDIVSSrm_Int:
4228 case X86::VDIVSSrr_Int:
4230 case X86::VSQRTPDm_Int:
4232 case X86::VSQRTPDr_Int:
4234 case X86::VSQRTPSm_Int:
4236 case X86::VSQRTPSr_Int:
4238 case X86::VSQRTSDm_Int:
4241 case X86::VSQRTSSm_Int:
4248 hasHighOperandLatency(const InstrItineraryData *ItinData,
4249 const MachineRegisterInfo *MRI,
4250 const MachineInstr *DefMI, unsigned DefIdx,
4251 const MachineInstr *UseMI, unsigned UseIdx) const {
4252 return isHighLatencyDef(DefMI->getOpcode());
4256 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4257 /// global base register for x86-32.
4258 struct CGBR : public MachineFunctionPass {
4260 CGBR() : MachineFunctionPass(ID) {}
4262 virtual bool runOnMachineFunction(MachineFunction &MF) {
4263 const X86TargetMachine *TM =
4264 static_cast<const X86TargetMachine *>(&MF.getTarget());
4266 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4267 "X86-64 PIC uses RIP relative addressing");
4269 // Only emit a global base reg in PIC mode.
4270 if (TM->getRelocationModel() != Reloc::PIC_)
4273 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4274 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4276 // If we didn't need a GlobalBaseReg, don't insert code.
4277 if (GlobalBaseReg == 0)
4280 // Insert the set of GlobalBaseReg into the first MBB of the function
4281 MachineBasicBlock &FirstMBB = MF.front();
4282 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4283 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4284 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4285 const X86InstrInfo *TII = TM->getInstrInfo();
4288 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
4289 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
4293 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4294 // only used in JIT code emission as displacement to pc.
4295 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
4297 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4298 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4299 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
4300 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4301 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4302 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4303 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4309 virtual const char *getPassName() const {
4310 return "X86 PIC Global Base Reg Initialization";
4313 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4314 AU.setPreservesCFG();
4315 MachineFunctionPass::getAnalysisUsage(AU);
4325 llvm::createGlobalBaseRegPass() { return new CGBR(); }
4328 struct LDTLSCleanup : public MachineFunctionPass {
4330 LDTLSCleanup() : MachineFunctionPass(ID) {}
4332 virtual bool runOnMachineFunction(MachineFunction &MF) {
4333 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4334 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4335 // No point folding accesses if there isn't at least two.
4339 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4340 return VisitNode(DT->getRootNode(), 0);
4343 // Visit the dominator subtree rooted at Node in pre-order.
4344 // If TLSBaseAddrReg is non-null, then use that to replace any
4345 // TLS_base_addr instructions. Otherwise, create the register
4346 // when the first such instruction is seen, and then use it
4347 // as we encounter more instructions.
4348 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4349 MachineBasicBlock *BB = Node->getBlock();
4350 bool Changed = false;
4352 // Traverse the current block.
4353 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4355 switch (I->getOpcode()) {
4356 case X86::TLS_base_addr32:
4357 case X86::TLS_base_addr64:
4359 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4361 I = SetRegister(I, &TLSBaseAddrReg);
4369 // Visit the children of this block in the dominator tree.
4370 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4372 Changed |= VisitNode(*I, TLSBaseAddrReg);
4378 // Replace the TLS_base_addr instruction I with a copy from
4379 // TLSBaseAddrReg, returning the new instruction.
4380 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4381 unsigned TLSBaseAddrReg) {
4382 MachineFunction *MF = I->getParent()->getParent();
4383 const X86TargetMachine *TM =
4384 static_cast<const X86TargetMachine *>(&MF->getTarget());
4385 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4386 const X86InstrInfo *TII = TM->getInstrInfo();
4388 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4389 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4390 TII->get(TargetOpcode::COPY),
4391 is64Bit ? X86::RAX : X86::EAX)
4392 .addReg(TLSBaseAddrReg);
4394 // Erase the TLS_base_addr instruction.
4395 I->eraseFromParent();
4400 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4401 // inserting a copy instruction after I. Returns the new instruction.
4402 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4403 MachineFunction *MF = I->getParent()->getParent();
4404 const X86TargetMachine *TM =
4405 static_cast<const X86TargetMachine *>(&MF->getTarget());
4406 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4407 const X86InstrInfo *TII = TM->getInstrInfo();
4409 // Create a virtual register for the TLS base address.
4410 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4411 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4412 ? &X86::GR64RegClass
4413 : &X86::GR32RegClass);
4415 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4416 MachineInstr *Next = I->getNextNode();
4417 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4418 TII->get(TargetOpcode::COPY),
4420 .addReg(is64Bit ? X86::RAX : X86::EAX);
4425 virtual const char *getPassName() const {
4426 return "Local Dynamic TLS Access Clean-up";
4429 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4430 AU.setPreservesCFG();
4431 AU.addRequired<MachineDominatorTree>();
4432 MachineFunctionPass::getAnalysisUsage(AU);
4437 char LDTLSCleanup::ID = 0;
4439 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
projects / oota-llvm.git / blob