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"
38 #define GET_INSTRINFO_CTOR
39 #include "X86GenInstrInfo.inc"
44 NoFusing("disable-spill-fusing",
45 cl::desc("Disable fusing of spill code into instructions"));
47 PrintFailedFusing("print-failed-fuse-candidates",
48 cl::desc("Print instructions that the allocator wants to"
49 " fuse, but the X86 backend currently can't"),
52 ReMatPICStubLoad("remat-pic-stub-load",
53 cl::desc("Re-materialize load from stub in PIC mode"),
54 cl::init(false), cl::Hidden);
57 // Select which memory operand is being unfolded.
58 // (stored in bits 0 - 3)
65 // Do not insert the reverse map (MemOp -> RegOp) into the table.
66 // This may be needed because there is a many -> one mapping.
67 TB_NO_REVERSE = 1 << 4,
69 // Do not insert the forward map (RegOp -> MemOp) into the table.
70 // This is needed for Native Client, which prohibits branch
71 // instructions from using a memory operand.
72 TB_NO_FORWARD = 1 << 5,
74 TB_FOLDED_LOAD = 1 << 6,
75 TB_FOLDED_STORE = 1 << 7,
77 // Minimum alignment required for load/store.
78 // Used for RegOp->MemOp conversion.
79 // (stored in bits 8 - 15)
81 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
82 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
83 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
87 struct X86OpTblEntry {
93 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
94 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
95 ? X86::ADJCALLSTACKDOWN64
96 : X86::ADJCALLSTACKDOWN32),
97 (tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKUP64
99 : X86::ADJCALLSTACKUP32)),
100 TM(tm), RI(tm, *this) {
102 static const X86OpTblEntry OpTbl2Addr[] = {
103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
270 unsigned RegOp = OpTbl2Addr[i].RegOp;
271 unsigned MemOp = OpTbl2Addr[i].MemOp;
272 unsigned Flags = OpTbl2Addr[i].Flags;
273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
275 // Index 0, folded load and store, no alignment requirement.
276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
279 static const X86OpTblEntry OpTbl0[] = {
280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
381 unsigned RegOp = OpTbl0[i].RegOp;
382 unsigned MemOp = OpTbl0[i].MemOp;
383 unsigned Flags = OpTbl0[i].Flags;
384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
388 static const X86OpTblEntry OpTbl1[] = {
389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
417 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
418 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
419 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
420 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
421 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
422 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
423 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
424 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
425 { X86::MOV16rr, X86::MOV16rm, 0 },
426 { X86::MOV32rr, X86::MOV32rm, 0 },
427 { X86::MOV64rr, X86::MOV64rm, 0 },
428 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
429 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
430 { X86::MOV8rr, X86::MOV8rm, 0 },
431 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
432 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
433 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
434 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
435 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
436 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
437 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
438 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
439 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
440 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
441 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
442 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
443 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
444 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
445 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
446 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
447 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
448 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
449 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
450 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
451 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
452 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
453 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
454 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
455 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
456 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
457 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
458 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
459 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
460 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
461 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
462 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
463 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
464 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
465 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
466 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
467 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
468 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
469 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
470 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 },
471 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
472 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 },
473 { X86::SQRTSDr, X86::SQRTSDm, 0 },
474 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
475 { X86::SQRTSSr, X86::SQRTSSm, 0 },
476 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
477 { X86::TEST16rr, X86::TEST16rm, 0 },
478 { X86::TEST32rr, X86::TEST32rm, 0 },
479 { X86::TEST64rr, X86::TEST64rm, 0 },
480 { X86::TEST8rr, X86::TEST8rm, 0 },
481 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
482 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
483 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
484 // AVX 128-bit versions of foldable instructions
485 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
486 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
487 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
488 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
489 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
490 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
491 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
492 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
493 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
494 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
495 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
496 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
497 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
498 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
499 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
500 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
501 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
502 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
503 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
504 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
505 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
506 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
507 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
508 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
509 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
510 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
511 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
512 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
513 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 },
514 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
515 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
516 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
517 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
518 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 },
519 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 },
520 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 },
521 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
522 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
523 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 },
524 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 },
525 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 },
526 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 },
527 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 },
528 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 },
529 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 },
530 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 },
531 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 },
532 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 },
533 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 },
534 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
535 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
536 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
538 // AVX 256-bit foldable instructions
539 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
540 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
541 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
542 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
543 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
544 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
545 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
547 // AVX2 foldable instructions
548 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 },
549 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 },
550 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 },
551 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 },
552 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 },
553 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 },
554 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 },
555 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 },
556 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 },
557 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 },
558 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 },
559 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 },
560 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 },
561 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 },
562 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
563 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
566 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
567 unsigned RegOp = OpTbl1[i].RegOp;
568 unsigned MemOp = OpTbl1[i].MemOp;
569 unsigned Flags = OpTbl1[i].Flags;
570 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
572 // Index 1, folded load
573 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
576 static const X86OpTblEntry OpTbl2[] = {
577 { X86::ADC32rr, X86::ADC32rm, 0 },
578 { X86::ADC64rr, X86::ADC64rm, 0 },
579 { X86::ADD16rr, X86::ADD16rm, 0 },
580 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
581 { X86::ADD32rr, X86::ADD32rm, 0 },
582 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
583 { X86::ADD64rr, X86::ADD64rm, 0 },
584 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
585 { X86::ADD8rr, X86::ADD8rm, 0 },
586 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
587 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
588 { X86::ADDSDrr, X86::ADDSDrm, 0 },
589 { X86::ADDSSrr, X86::ADDSSrm, 0 },
590 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
591 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
592 { X86::AND16rr, X86::AND16rm, 0 },
593 { X86::AND32rr, X86::AND32rm, 0 },
594 { X86::AND64rr, X86::AND64rm, 0 },
595 { X86::AND8rr, X86::AND8rm, 0 },
596 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
597 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
598 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
599 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
600 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
601 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
602 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
603 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
604 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
605 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
606 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
607 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
608 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
609 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
610 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
611 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
612 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
613 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
614 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
615 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
616 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
617 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
618 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
619 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
620 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
621 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
622 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
623 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
624 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
625 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
626 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
627 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
628 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
629 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
630 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
631 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
632 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
633 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
634 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
635 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
636 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
637 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
638 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
639 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
640 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
641 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
642 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
643 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
644 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
645 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
646 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
647 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
648 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
649 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
650 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
651 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
652 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
653 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
654 { X86::CMPSDrr, X86::CMPSDrm, 0 },
655 { X86::CMPSSrr, X86::CMPSSrm, 0 },
656 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
657 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
658 { X86::DIVSDrr, X86::DIVSDrm, 0 },
659 { X86::DIVSSrr, X86::DIVSSrm, 0 },
660 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
661 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
662 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
663 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
664 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
665 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
666 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
667 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
668 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
669 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
670 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
671 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
672 { X86::IMUL16rr, X86::IMUL16rm, 0 },
673 { X86::IMUL32rr, X86::IMUL32rm, 0 },
674 { X86::IMUL64rr, X86::IMUL64rm, 0 },
675 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
676 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
677 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
678 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
679 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
680 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
681 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
682 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
683 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
684 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
685 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
686 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
687 { X86::MAXSDrr, X86::MAXSDrm, 0 },
688 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
689 { X86::MAXSSrr, X86::MAXSSrm, 0 },
690 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
691 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
692 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
693 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
694 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
695 { X86::MINSDrr, X86::MINSDrm, 0 },
696 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
697 { X86::MINSSrr, X86::MINSSrm, 0 },
698 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
699 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
700 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
701 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
702 { X86::MULSDrr, X86::MULSDrm, 0 },
703 { X86::MULSSrr, X86::MULSSrm, 0 },
704 { X86::OR16rr, X86::OR16rm, 0 },
705 { X86::OR32rr, X86::OR32rm, 0 },
706 { X86::OR64rr, X86::OR64rm, 0 },
707 { X86::OR8rr, X86::OR8rm, 0 },
708 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
709 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
710 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
711 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
712 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
713 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
714 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
715 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
716 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
717 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
718 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
719 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
720 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
721 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
722 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
723 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
724 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
725 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
726 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
727 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
728 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
729 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
730 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
731 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
732 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
733 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
734 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
735 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
736 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
737 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
738 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
739 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
740 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
741 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
742 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
743 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
744 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
745 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
746 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
747 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
748 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
749 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
750 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
751 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
752 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
753 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
754 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
755 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
756 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
757 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
758 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
759 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
760 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
761 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
762 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
763 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
764 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
765 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
766 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
767 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
768 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
769 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
770 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
771 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
772 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
773 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
774 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
775 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
776 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
777 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
778 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
779 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
780 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
781 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
782 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
783 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
784 { X86::SBB32rr, X86::SBB32rm, 0 },
785 { X86::SBB64rr, X86::SBB64rm, 0 },
786 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
787 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
788 { X86::SUB16rr, X86::SUB16rm, 0 },
789 { X86::SUB32rr, X86::SUB32rm, 0 },
790 { X86::SUB64rr, X86::SUB64rm, 0 },
791 { X86::SUB8rr, X86::SUB8rm, 0 },
792 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
793 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
794 { X86::SUBSDrr, X86::SUBSDrm, 0 },
795 { X86::SUBSSrr, X86::SUBSSrm, 0 },
796 // FIXME: TEST*rr -> swapped operand of TEST*mr.
797 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
798 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
799 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
800 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
801 { X86::XOR16rr, X86::XOR16rm, 0 },
802 { X86::XOR32rr, X86::XOR32rm, 0 },
803 { X86::XOR64rr, X86::XOR64rm, 0 },
804 { X86::XOR8rr, X86::XOR8rm, 0 },
805 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
806 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
807 // AVX 128-bit versions of foldable instructions
808 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
809 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
810 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
811 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
812 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
813 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
814 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
815 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
816 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
817 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
818 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
819 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
820 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, TB_ALIGN_16 },
821 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
822 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
823 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
824 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
825 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
826 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
827 { X86::VADDSDrr, X86::VADDSDrm, 0 },
828 { X86::VADDSSrr, X86::VADDSSrm, 0 },
829 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
830 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
831 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
832 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
833 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
834 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
835 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
836 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
837 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
838 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
839 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
840 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
841 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
842 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
843 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
844 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
845 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
846 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
847 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
848 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
849 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
850 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
851 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
852 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
853 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
854 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
855 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
856 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
857 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
858 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
859 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
860 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
861 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
862 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
863 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
864 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
865 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
866 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
867 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
868 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
869 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
870 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
871 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
872 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
873 { X86::VMINSDrr, X86::VMINSDrm, 0 },
874 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
875 { X86::VMINSSrr, X86::VMINSSrm, 0 },
876 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
877 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
878 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
879 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
880 { X86::VMULSDrr, X86::VMULSDrm, 0 },
881 { X86::VMULSSrr, X86::VMULSSrm, 0 },
882 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
883 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
884 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
885 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
886 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
887 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
888 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
889 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
890 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
891 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
892 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
893 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
894 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
895 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
896 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
897 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
898 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
899 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
900 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
901 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
902 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
903 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
904 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
905 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
906 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
907 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
908 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
909 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
910 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
911 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
912 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
913 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
914 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
915 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
916 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
917 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
918 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
919 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
920 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
921 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
922 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
923 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
924 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
925 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
926 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
927 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
928 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
929 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
930 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
931 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
932 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
933 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
934 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
935 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
936 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
937 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
938 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
939 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
940 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
941 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
942 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
943 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
944 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
945 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
946 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
947 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
948 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
949 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
950 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
951 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
952 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
953 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
954 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
955 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
956 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
957 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
958 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
959 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
960 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
961 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
962 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
963 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
964 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
965 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
966 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
967 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
968 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
969 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
970 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
971 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
972 // AVX 256-bit foldable instructions
973 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
974 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
975 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
976 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
977 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
978 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
979 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
980 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
981 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
982 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
983 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
984 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
985 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
986 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
987 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
988 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
989 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
990 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
991 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
992 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
993 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
994 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
995 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
996 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
997 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
998 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
999 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
1000 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
1001 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
1002 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
1003 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
1004 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1005 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1006 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1007 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1008 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1009 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1010 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1011 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1012 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1013 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1014 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1015 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1016 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1017 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1018 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
1019 // AVX2 foldable instructions
1020 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1021 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1022 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1023 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1024 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1025 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1026 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1027 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1028 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1029 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1030 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1031 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1032 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1033 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1034 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1035 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1036 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1037 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1038 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1039 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1040 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1041 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1042 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1043 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1044 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1045 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1046 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1047 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1048 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1049 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
1050 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
1051 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
1052 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
1053 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
1054 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
1055 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
1056 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1057 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
1058 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
1059 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
1060 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1061 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1062 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1063 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1064 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1065 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1066 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1067 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1068 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1069 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1070 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1071 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1072 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1073 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1074 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1075 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
1076 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1077 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1078 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1079 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
1080 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1081 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1082 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1083 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
1084 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
1085 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
1086 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
1087 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1088 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1089 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
1090 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
1091 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1092 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1093 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1094 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
1095 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
1096 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
1097 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1098 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1099 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1100 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1101 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1102 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1103 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1104 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
1105 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
1106 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1107 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1108 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1109 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1110 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1111 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
1112 // FIXME: add AVX 256-bit foldable instructions
1114 // FMA4 foldable patterns
1115 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, TB_ALIGN_16 },
1116 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, TB_ALIGN_16 },
1117 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1118 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1119 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1120 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1121 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1122 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1123 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1124 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1125 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, TB_ALIGN_16 },
1126 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, TB_ALIGN_16 },
1127 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1128 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1129 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1130 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1131 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1132 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1133 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1134 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1135 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1136 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1137 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1138 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1139 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1140 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1141 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1142 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
1145 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1146 unsigned RegOp = OpTbl2[i].RegOp;
1147 unsigned MemOp = OpTbl2[i].MemOp;
1148 unsigned Flags = OpTbl2[i].Flags;
1149 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1151 // Index 2, folded load
1152 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1155 static const X86OpTblEntry OpTbl3[] = {
1156 // FMA foldable instructions
1157 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1158 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1159 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1160 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1161 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1162 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
1163 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, 0 },
1164 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, 0 },
1166 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1167 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1168 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1169 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1170 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1171 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1172 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1173 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1174 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1175 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1176 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1177 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
1179 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1180 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1181 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1182 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1183 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1184 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
1185 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, 0 },
1186 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, 0 },
1188 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1189 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1190 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1191 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1192 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1193 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1194 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1195 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1196 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1197 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1198 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1199 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
1201 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1202 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1203 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1204 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1205 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1206 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
1207 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, 0 },
1208 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, 0 },
1210 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1211 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1212 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1213 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1214 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1215 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1216 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1217 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1218 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1219 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1220 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1221 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
1223 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1224 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1225 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1226 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1227 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1228 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
1229 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, 0 },
1230 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, 0 },
1232 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1233 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1234 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1235 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1236 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1237 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1238 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1239 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1240 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1241 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1242 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1243 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
1245 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1246 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1247 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1248 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1249 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1250 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1251 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1252 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1253 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1254 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1255 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1256 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
1258 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1259 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1260 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1261 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1262 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1263 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1264 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1265 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1266 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1267 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1268 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1269 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
1271 // FMA4 foldable patterns
1272 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, TB_ALIGN_16 },
1273 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, TB_ALIGN_16 },
1274 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1275 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1276 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1277 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1278 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1279 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1280 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1281 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1282 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, TB_ALIGN_16 },
1283 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, TB_ALIGN_16 },
1284 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1285 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1286 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1287 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1288 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1289 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1290 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1291 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1292 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1293 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1294 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1295 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1296 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1297 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1298 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1299 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
1302 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1303 unsigned RegOp = OpTbl3[i].RegOp;
1304 unsigned MemOp = OpTbl3[i].MemOp;
1305 unsigned Flags = OpTbl3[i].Flags;
1306 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1308 // Index 3, folded load
1309 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1315 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1316 MemOp2RegOpTableType &M2RTable,
1317 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1318 if ((Flags & TB_NO_FORWARD) == 0) {
1319 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1320 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1322 if ((Flags & TB_NO_REVERSE) == 0) {
1323 assert(!M2RTable.count(MemOp) &&
1324 "Duplicated entries in unfolding maps?");
1325 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1330 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1331 unsigned &SrcReg, unsigned &DstReg,
1332 unsigned &SubIdx) const {
1333 switch (MI.getOpcode()) {
1335 case X86::MOVSX16rr8:
1336 case X86::MOVZX16rr8:
1337 case X86::MOVSX32rr8:
1338 case X86::MOVZX32rr8:
1339 case X86::MOVSX64rr8:
1340 case X86::MOVZX64rr8:
1341 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1342 // It's not always legal to reference the low 8-bit of the larger
1343 // register in 32-bit mode.
1345 case X86::MOVSX32rr16:
1346 case X86::MOVZX32rr16:
1347 case X86::MOVSX64rr16:
1348 case X86::MOVZX64rr16:
1349 case X86::MOVSX64rr32:
1350 case X86::MOVZX64rr32: {
1351 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1354 SrcReg = MI.getOperand(1).getReg();
1355 DstReg = MI.getOperand(0).getReg();
1356 switch (MI.getOpcode()) {
1357 default: llvm_unreachable("Unreachable!");
1358 case X86::MOVSX16rr8:
1359 case X86::MOVZX16rr8:
1360 case X86::MOVSX32rr8:
1361 case X86::MOVZX32rr8:
1362 case X86::MOVSX64rr8:
1363 case X86::MOVZX64rr8:
1364 SubIdx = X86::sub_8bit;
1366 case X86::MOVSX32rr16:
1367 case X86::MOVZX32rr16:
1368 case X86::MOVSX64rr16:
1369 case X86::MOVZX64rr16:
1370 SubIdx = X86::sub_16bit;
1372 case X86::MOVSX64rr32:
1373 case X86::MOVZX64rr32:
1374 SubIdx = X86::sub_32bit;
1383 /// isFrameOperand - Return true and the FrameIndex if the specified
1384 /// operand and follow operands form a reference to the stack frame.
1385 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1386 int &FrameIndex) const {
1387 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1388 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1389 MI->getOperand(Op+1).getImm() == 1 &&
1390 MI->getOperand(Op+2).getReg() == 0 &&
1391 MI->getOperand(Op+3).getImm() == 0) {
1392 FrameIndex = MI->getOperand(Op).getIndex();
1398 static bool isFrameLoadOpcode(int Opcode) {
1414 case X86::VMOVAPSrm:
1415 case X86::VMOVAPDrm:
1416 case X86::VMOVDQArm:
1417 case X86::VMOVAPSYrm:
1418 case X86::VMOVAPDYrm:
1419 case X86::VMOVDQAYrm:
1420 case X86::MMX_MOVD64rm:
1421 case X86::MMX_MOVQ64rm:
1426 static bool isFrameStoreOpcode(int Opcode) {
1433 case X86::ST_FpP64m:
1441 case X86::VMOVAPSmr:
1442 case X86::VMOVAPDmr:
1443 case X86::VMOVDQAmr:
1444 case X86::VMOVAPSYmr:
1445 case X86::VMOVAPDYmr:
1446 case X86::VMOVDQAYmr:
1447 case X86::MMX_MOVD64mr:
1448 case X86::MMX_MOVQ64mr:
1449 case X86::MMX_MOVNTQmr:
1455 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1456 int &FrameIndex) const {
1457 if (isFrameLoadOpcode(MI->getOpcode()))
1458 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1459 return MI->getOperand(0).getReg();
1463 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1464 int &FrameIndex) const {
1465 if (isFrameLoadOpcode(MI->getOpcode())) {
1467 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1469 // Check for post-frame index elimination operations
1470 const MachineMemOperand *Dummy;
1471 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1476 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1477 int &FrameIndex) const {
1478 if (isFrameStoreOpcode(MI->getOpcode()))
1479 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1480 isFrameOperand(MI, 0, FrameIndex))
1481 return MI->getOperand(X86::AddrNumOperands).getReg();
1485 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1486 int &FrameIndex) const {
1487 if (isFrameStoreOpcode(MI->getOpcode())) {
1489 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1491 // Check for post-frame index elimination operations
1492 const MachineMemOperand *Dummy;
1493 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1498 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1500 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1501 // Don't waste compile time scanning use-def chains of physregs.
1502 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1504 bool isPICBase = false;
1505 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1506 E = MRI.def_end(); I != E; ++I) {
1507 MachineInstr *DefMI = I.getOperand().getParent();
1508 if (DefMI->getOpcode() != X86::MOVPC32r)
1510 assert(!isPICBase && "More than one PIC base?");
1517 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1518 AliasAnalysis *AA) const {
1519 switch (MI->getOpcode()) {
1534 case X86::VMOVAPSrm:
1535 case X86::VMOVUPSrm:
1536 case X86::VMOVAPDrm:
1537 case X86::VMOVDQArm:
1538 case X86::VMOVAPSYrm:
1539 case X86::VMOVUPSYrm:
1540 case X86::VMOVAPDYrm:
1541 case X86::VMOVDQAYrm:
1542 case X86::MMX_MOVD64rm:
1543 case X86::MMX_MOVQ64rm:
1544 case X86::FsVMOVAPSrm:
1545 case X86::FsVMOVAPDrm:
1546 case X86::FsMOVAPSrm:
1547 case X86::FsMOVAPDrm: {
1548 // Loads from constant pools are trivially rematerializable.
1549 if (MI->getOperand(1).isReg() &&
1550 MI->getOperand(2).isImm() &&
1551 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1552 MI->isInvariantLoad(AA)) {
1553 unsigned BaseReg = MI->getOperand(1).getReg();
1554 if (BaseReg == 0 || BaseReg == X86::RIP)
1556 // Allow re-materialization of PIC load.
1557 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1559 const MachineFunction &MF = *MI->getParent()->getParent();
1560 const MachineRegisterInfo &MRI = MF.getRegInfo();
1561 return regIsPICBase(BaseReg, MRI);
1568 if (MI->getOperand(2).isImm() &&
1569 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1570 !MI->getOperand(4).isReg()) {
1571 // lea fi#, lea GV, etc. are all rematerializable.
1572 if (!MI->getOperand(1).isReg())
1574 unsigned BaseReg = MI->getOperand(1).getReg();
1577 // Allow re-materialization of lea PICBase + x.
1578 const MachineFunction &MF = *MI->getParent()->getParent();
1579 const MachineRegisterInfo &MRI = MF.getRegInfo();
1580 return regIsPICBase(BaseReg, MRI);
1586 // All other instructions marked M_REMATERIALIZABLE are always trivially
1587 // rematerializable.
1591 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1592 /// would clobber the EFLAGS condition register. Note the result may be
1593 /// conservative. If it cannot definitely determine the safety after visiting
1594 /// a few instructions in each direction it assumes it's not safe.
1595 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1596 MachineBasicBlock::iterator I) {
1597 MachineBasicBlock::iterator E = MBB.end();
1599 // For compile time consideration, if we are not able to determine the
1600 // safety after visiting 4 instructions in each direction, we will assume
1602 MachineBasicBlock::iterator Iter = I;
1603 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1604 bool SeenDef = false;
1605 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1606 MachineOperand &MO = Iter->getOperand(j);
1607 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1611 if (MO.getReg() == X86::EFLAGS) {
1619 // This instruction defines EFLAGS, no need to look any further.
1622 // Skip over DBG_VALUE.
1623 while (Iter != E && Iter->isDebugValue())
1627 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1630 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1631 SE = MBB.succ_end(); SI != SE; ++SI)
1632 if ((*SI)->isLiveIn(X86::EFLAGS))
1637 MachineBasicBlock::iterator B = MBB.begin();
1639 for (unsigned i = 0; i < 4; ++i) {
1640 // If we make it to the beginning of the block, it's safe to clobber
1641 // EFLAGS iff EFLAGS is not live-in.
1643 return !MBB.isLiveIn(X86::EFLAGS);
1646 // Skip over DBG_VALUE.
1647 while (Iter != B && Iter->isDebugValue())
1650 bool SawKill = false;
1651 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1652 MachineOperand &MO = Iter->getOperand(j);
1653 // A register mask may clobber EFLAGS, but we should still look for a
1655 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1657 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1658 if (MO.isDef()) return MO.isDead();
1659 if (MO.isKill()) SawKill = true;
1664 // This instruction kills EFLAGS and doesn't redefine it, so
1665 // there's no need to look further.
1669 // Conservative answer.
1673 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1674 MachineBasicBlock::iterator I,
1675 unsigned DestReg, unsigned SubIdx,
1676 const MachineInstr *Orig,
1677 const TargetRegisterInfo &TRI) const {
1678 DebugLoc DL = Orig->getDebugLoc();
1680 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1681 // Re-materialize them as movri instructions to avoid side effects.
1683 unsigned Opc = Orig->getOpcode();
1689 case X86::MOV64r0: {
1690 if (!isSafeToClobberEFLAGS(MBB, I)) {
1692 default: llvm_unreachable("Unreachable!");
1693 case X86::MOV8r0: Opc = X86::MOV8ri; break;
1694 case X86::MOV16r0: Opc = X86::MOV16ri; break;
1695 case X86::MOV32r0: Opc = X86::MOV32ri; break;
1696 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
1705 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1708 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
1711 MachineInstr *NewMI = prior(I);
1712 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1715 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1716 /// is not marked dead.
1717 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1718 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1719 MachineOperand &MO = MI->getOperand(i);
1720 if (MO.isReg() && MO.isDef() &&
1721 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1728 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1729 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1730 /// to a 32-bit superregister and then truncating back down to a 16-bit
1733 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1734 MachineFunction::iterator &MFI,
1735 MachineBasicBlock::iterator &MBBI,
1736 LiveVariables *LV) const {
1737 MachineInstr *MI = MBBI;
1738 unsigned Dest = MI->getOperand(0).getReg();
1739 unsigned Src = MI->getOperand(1).getReg();
1740 bool isDead = MI->getOperand(0).isDead();
1741 bool isKill = MI->getOperand(1).isKill();
1743 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1744 ? X86::LEA64_32r : X86::LEA32r;
1745 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1746 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1747 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1749 // Build and insert into an implicit UNDEF value. This is OK because
1750 // well be shifting and then extracting the lower 16-bits.
1751 // This has the potential to cause partial register stall. e.g.
1752 // movw (%rbp,%rcx,2), %dx
1753 // leal -65(%rdx), %esi
1754 // But testing has shown this *does* help performance in 64-bit mode (at
1755 // least on modern x86 machines).
1756 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1757 MachineInstr *InsMI =
1758 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1759 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1760 .addReg(Src, getKillRegState(isKill));
1762 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1763 get(Opc), leaOutReg);
1765 default: llvm_unreachable("Unreachable!");
1766 case X86::SHL16ri: {
1767 unsigned ShAmt = MI->getOperand(2).getImm();
1768 MIB.addReg(0).addImm(1 << ShAmt)
1769 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1773 case X86::INC64_16r:
1774 addRegOffset(MIB, leaInReg, true, 1);
1777 case X86::DEC64_16r:
1778 addRegOffset(MIB, leaInReg, true, -1);
1782 case X86::ADD16ri_DB:
1783 case X86::ADD16ri8_DB:
1784 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
1787 case X86::ADD16rr_DB: {
1788 unsigned Src2 = MI->getOperand(2).getReg();
1789 bool isKill2 = MI->getOperand(2).isKill();
1790 unsigned leaInReg2 = 0;
1791 MachineInstr *InsMI2 = 0;
1793 // ADD16rr %reg1028<kill>, %reg1028
1794 // just a single insert_subreg.
1795 addRegReg(MIB, leaInReg, true, leaInReg, false);
1797 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1798 // Build and insert into an implicit UNDEF value. This is OK because
1799 // well be shifting and then extracting the lower 16-bits.
1800 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
1802 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
1803 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1804 .addReg(Src2, getKillRegState(isKill2));
1805 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1807 if (LV && isKill2 && InsMI2)
1808 LV->replaceKillInstruction(Src2, MI, InsMI2);
1813 MachineInstr *NewMI = MIB;
1814 MachineInstr *ExtMI =
1815 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1816 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1817 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
1820 // Update live variables
1821 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1822 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1824 LV->replaceKillInstruction(Src, MI, InsMI);
1826 LV->replaceKillInstruction(Dest, MI, ExtMI);
1832 /// convertToThreeAddress - This method must be implemented by targets that
1833 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1834 /// may be able to convert a two-address instruction into a true
1835 /// three-address instruction on demand. This allows the X86 target (for
1836 /// example) to convert ADD and SHL instructions into LEA instructions if they
1837 /// would require register copies due to two-addressness.
1839 /// This method returns a null pointer if the transformation cannot be
1840 /// performed, otherwise it returns the new instruction.
1843 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1844 MachineBasicBlock::iterator &MBBI,
1845 LiveVariables *LV) const {
1846 MachineInstr *MI = MBBI;
1847 MachineFunction &MF = *MI->getParent()->getParent();
1848 // All instructions input are two-addr instructions. Get the known operands.
1849 const MachineOperand &Dest = MI->getOperand(0);
1850 const MachineOperand &Src = MI->getOperand(1);
1852 MachineInstr *NewMI = NULL;
1853 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
1854 // we have better subtarget support, enable the 16-bit LEA generation here.
1855 // 16-bit LEA is also slow on Core2.
1856 bool DisableLEA16 = true;
1857 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
1859 unsigned MIOpc = MI->getOpcode();
1861 case X86::SHUFPSrri: {
1862 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
1863 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1865 unsigned B = MI->getOperand(1).getReg();
1866 unsigned C = MI->getOperand(2).getReg();
1867 if (B != C) return 0;
1868 unsigned M = MI->getOperand(3).getImm();
1869 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1870 .addOperand(Dest).addOperand(Src).addImm(M);
1873 case X86::SHUFPDrri: {
1874 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1875 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1877 unsigned B = MI->getOperand(1).getReg();
1878 unsigned C = MI->getOperand(2).getReg();
1879 if (B != C) return 0;
1880 unsigned M = MI->getOperand(3).getImm();
1882 // Convert to PSHUFD mask.
1883 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1885 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1886 .addOperand(Dest).addOperand(Src).addImm(M);
1889 case X86::SHL64ri: {
1890 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1891 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1892 // the flags produced by a shift yet, so this is safe.
1893 unsigned ShAmt = MI->getOperand(2).getImm();
1894 if (ShAmt == 0 || ShAmt >= 4) return 0;
1896 // LEA can't handle RSP.
1897 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1898 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1899 &X86::GR64_NOSPRegClass))
1902 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
1904 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
1907 case X86::SHL32ri: {
1908 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1909 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1910 // the flags produced by a shift yet, so this is safe.
1911 unsigned ShAmt = MI->getOperand(2).getImm();
1912 if (ShAmt == 0 || ShAmt >= 4) return 0;
1914 // LEA can't handle ESP.
1915 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1916 !MF.getRegInfo().constrainRegClass(Src.getReg(),
1917 &X86::GR32_NOSPRegClass))
1920 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
1921 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
1923 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
1926 case X86::SHL16ri: {
1927 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
1928 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1929 // the flags produced by a shift yet, so this is safe.
1930 unsigned ShAmt = MI->getOperand(2).getImm();
1931 if (ShAmt == 0 || ShAmt >= 4) return 0;
1934 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1935 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1937 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
1941 // The following opcodes also sets the condition code register(s). Only
1942 // convert them to equivalent lea if the condition code register def's
1944 if (hasLiveCondCodeDef(MI))
1951 case X86::INC64_32r: {
1952 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1953 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1954 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1955 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1956 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1957 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
1959 // LEA can't handle RSP.
1960 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1961 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
1964 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1965 .addOperand(Dest).addOperand(Src), 1);
1969 case X86::INC64_16r:
1971 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1972 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
1973 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1974 .addOperand(Dest).addOperand(Src), 1);
1978 case X86::DEC64_32r: {
1979 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1980 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1981 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
1982 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1983 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1984 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
1985 // LEA can't handle RSP.
1986 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
1987 !MF.getRegInfo().constrainRegClass(Src.getReg(), RC))
1990 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
1991 .addOperand(Dest).addOperand(Src), -1);
1995 case X86::DEC64_16r:
1997 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
1998 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
1999 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2000 .addOperand(Dest).addOperand(Src), -1);
2003 case X86::ADD64rr_DB:
2005 case X86::ADD32rr_DB: {
2006 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2008 const TargetRegisterClass *RC;
2009 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2011 RC = &X86::GR64_NOSPRegClass;
2013 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2014 RC = &X86::GR32_NOSPRegClass;
2018 unsigned Src2 = MI->getOperand(2).getReg();
2019 bool isKill2 = MI->getOperand(2).isKill();
2021 // LEA can't handle RSP.
2022 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
2023 !MF.getRegInfo().constrainRegClass(Src2, RC))
2026 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2028 Src.getReg(), Src.isKill(), Src2, isKill2);
2030 // Preserve undefness of the operands.
2031 bool isUndef = MI->getOperand(1).isUndef();
2032 bool isUndef2 = MI->getOperand(2).isUndef();
2033 NewMI->getOperand(1).setIsUndef(isUndef);
2034 NewMI->getOperand(3).setIsUndef(isUndef2);
2037 LV->replaceKillInstruction(Src2, MI, NewMI);
2041 case X86::ADD16rr_DB: {
2043 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2044 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2045 unsigned Src2 = MI->getOperand(2).getReg();
2046 bool isKill2 = MI->getOperand(2).isKill();
2047 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2049 Src.getReg(), Src.isKill(), Src2, isKill2);
2051 // Preserve undefness of the operands.
2052 bool isUndef = MI->getOperand(1).isUndef();
2053 bool isUndef2 = MI->getOperand(2).isUndef();
2054 NewMI->getOperand(1).setIsUndef(isUndef);
2055 NewMI->getOperand(3).setIsUndef(isUndef2);
2058 LV->replaceKillInstruction(Src2, MI, NewMI);
2061 case X86::ADD64ri32:
2063 case X86::ADD64ri32_DB:
2064 case X86::ADD64ri8_DB:
2065 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2066 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2067 .addOperand(Dest).addOperand(Src),
2068 MI->getOperand(2).getImm());
2072 case X86::ADD32ri_DB:
2073 case X86::ADD32ri8_DB: {
2074 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2075 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2076 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
2077 .addOperand(Dest).addOperand(Src),
2078 MI->getOperand(2).getImm());
2083 case X86::ADD16ri_DB:
2084 case X86::ADD16ri8_DB:
2086 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2087 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2088 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2089 .addOperand(Dest).addOperand(Src),
2090 MI->getOperand(2).getImm());
2096 if (!NewMI) return 0;
2098 if (LV) { // Update live variables
2100 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2102 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2105 MFI->insert(MBBI, NewMI); // Insert the new inst
2109 /// commuteInstruction - We have a few instructions that must be hacked on to
2113 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2114 switch (MI->getOpcode()) {
2115 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2116 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2117 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2118 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2119 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2120 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2123 switch (MI->getOpcode()) {
2124 default: llvm_unreachable("Unreachable!");
2125 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2126 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2127 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2128 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2129 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2130 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2132 unsigned Amt = MI->getOperand(3).getImm();
2134 MachineFunction &MF = *MI->getParent()->getParent();
2135 MI = MF.CloneMachineInstr(MI);
2138 MI->setDesc(get(Opc));
2139 MI->getOperand(3).setImm(Size-Amt);
2140 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
2142 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2143 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2144 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2145 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2146 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2147 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2148 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2149 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2150 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2151 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2152 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2153 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2154 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2155 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2156 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2157 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2159 switch (MI->getOpcode()) {
2160 default: llvm_unreachable("Unreachable!");
2161 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2162 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2163 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2164 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2165 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2166 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2167 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2168 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2169 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2170 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2171 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2172 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2173 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2174 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2175 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2176 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2177 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2178 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2179 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2180 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2181 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2182 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2183 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2184 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2185 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2186 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2187 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2188 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2189 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2190 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2191 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2192 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2193 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2194 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2195 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2196 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2197 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2198 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2199 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2200 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2201 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2202 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2203 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2204 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2205 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2206 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2207 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2208 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2211 MachineFunction &MF = *MI->getParent()->getParent();
2212 MI = MF.CloneMachineInstr(MI);
2215 MI->setDesc(get(Opc));
2216 // Fallthrough intended.
2219 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
2223 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2225 default: return X86::COND_INVALID;
2226 case X86::JE_4: return X86::COND_E;
2227 case X86::JNE_4: return X86::COND_NE;
2228 case X86::JL_4: return X86::COND_L;
2229 case X86::JLE_4: return X86::COND_LE;
2230 case X86::JG_4: return X86::COND_G;
2231 case X86::JGE_4: return X86::COND_GE;
2232 case X86::JB_4: return X86::COND_B;
2233 case X86::JBE_4: return X86::COND_BE;
2234 case X86::JA_4: return X86::COND_A;
2235 case X86::JAE_4: return X86::COND_AE;
2236 case X86::JS_4: return X86::COND_S;
2237 case X86::JNS_4: return X86::COND_NS;
2238 case X86::JP_4: return X86::COND_P;
2239 case X86::JNP_4: return X86::COND_NP;
2240 case X86::JO_4: return X86::COND_O;
2241 case X86::JNO_4: return X86::COND_NO;
2245 /// getCondFromSETOpc - return condition code of a SET opcode.
2246 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2248 default: return X86::COND_INVALID;
2249 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2250 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2251 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2252 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2253 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2254 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2255 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2256 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2257 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2258 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2259 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2260 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2261 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2262 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2263 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2264 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2268 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2269 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2271 default: return X86::COND_INVALID;
2272 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2273 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2275 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2276 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2277 return X86::COND_AE;
2278 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2279 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2281 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2282 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2283 return X86::COND_BE;
2284 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2285 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2287 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2288 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2290 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2291 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2292 return X86::COND_GE;
2293 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2294 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2296 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2297 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2298 return X86::COND_LE;
2299 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2300 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2301 return X86::COND_NE;
2302 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2303 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2304 return X86::COND_NO;
2305 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2306 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2307 return X86::COND_NP;
2308 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2309 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2310 return X86::COND_NS;
2311 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2312 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2314 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2315 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2317 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2318 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2323 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2325 default: llvm_unreachable("Illegal condition code!");
2326 case X86::COND_E: return X86::JE_4;
2327 case X86::COND_NE: return X86::JNE_4;
2328 case X86::COND_L: return X86::JL_4;
2329 case X86::COND_LE: return X86::JLE_4;
2330 case X86::COND_G: return X86::JG_4;
2331 case X86::COND_GE: return X86::JGE_4;
2332 case X86::COND_B: return X86::JB_4;
2333 case X86::COND_BE: return X86::JBE_4;
2334 case X86::COND_A: return X86::JA_4;
2335 case X86::COND_AE: return X86::JAE_4;
2336 case X86::COND_S: return X86::JS_4;
2337 case X86::COND_NS: return X86::JNS_4;
2338 case X86::COND_P: return X86::JP_4;
2339 case X86::COND_NP: return X86::JNP_4;
2340 case X86::COND_O: return X86::JO_4;
2341 case X86::COND_NO: return X86::JNO_4;
2345 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2346 /// e.g. turning COND_E to COND_NE.
2347 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2349 default: llvm_unreachable("Illegal condition code!");
2350 case X86::COND_E: return X86::COND_NE;
2351 case X86::COND_NE: return X86::COND_E;
2352 case X86::COND_L: return X86::COND_GE;
2353 case X86::COND_LE: return X86::COND_G;
2354 case X86::COND_G: return X86::COND_LE;
2355 case X86::COND_GE: return X86::COND_L;
2356 case X86::COND_B: return X86::COND_AE;
2357 case X86::COND_BE: return X86::COND_A;
2358 case X86::COND_A: return X86::COND_BE;
2359 case X86::COND_AE: return X86::COND_B;
2360 case X86::COND_S: return X86::COND_NS;
2361 case X86::COND_NS: return X86::COND_S;
2362 case X86::COND_P: return X86::COND_NP;
2363 case X86::COND_NP: return X86::COND_P;
2364 case X86::COND_O: return X86::COND_NO;
2365 case X86::COND_NO: return X86::COND_O;
2369 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2370 /// the condition code if we modify the instructions such that flags are
2372 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2374 default: return X86::COND_INVALID;
2375 case X86::COND_E: return X86::COND_E;
2376 case X86::COND_NE: return X86::COND_NE;
2377 case X86::COND_L: return X86::COND_G;
2378 case X86::COND_LE: return X86::COND_GE;
2379 case X86::COND_G: return X86::COND_L;
2380 case X86::COND_GE: return X86::COND_LE;
2381 case X86::COND_B: return X86::COND_A;
2382 case X86::COND_BE: return X86::COND_AE;
2383 case X86::COND_A: return X86::COND_B;
2384 case X86::COND_AE: return X86::COND_BE;
2388 /// getSETFromCond - Return a set opcode for the given condition and
2389 /// whether it has memory operand.
2390 static unsigned getSETFromCond(X86::CondCode CC,
2391 bool HasMemoryOperand) {
2392 static const uint16_t Opc[16][2] = {
2393 { X86::SETAr, X86::SETAm },
2394 { X86::SETAEr, X86::SETAEm },
2395 { X86::SETBr, X86::SETBm },
2396 { X86::SETBEr, X86::SETBEm },
2397 { X86::SETEr, X86::SETEm },
2398 { X86::SETGr, X86::SETGm },
2399 { X86::SETGEr, X86::SETGEm },
2400 { X86::SETLr, X86::SETLm },
2401 { X86::SETLEr, X86::SETLEm },
2402 { X86::SETNEr, X86::SETNEm },
2403 { X86::SETNOr, X86::SETNOm },
2404 { X86::SETNPr, X86::SETNPm },
2405 { X86::SETNSr, X86::SETNSm },
2406 { X86::SETOr, X86::SETOm },
2407 { X86::SETPr, X86::SETPm },
2408 { X86::SETSr, X86::SETSm }
2411 assert(CC < 16 && "Can only handle standard cond codes");
2412 return Opc[CC][HasMemoryOperand ? 1 : 0];
2415 /// getCMovFromCond - Return a cmov opcode for the given condition,
2416 /// register size in bytes, and operand type.
2417 static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2418 bool HasMemoryOperand) {
2419 static const uint16_t Opc[32][3] = {
2420 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2421 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2422 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2423 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2424 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2425 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2426 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2427 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2428 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2429 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2430 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2431 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2432 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2433 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2434 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2435 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2436 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2437 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2438 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2439 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2440 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2441 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2442 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2443 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2444 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2445 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2446 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2447 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2448 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2449 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2450 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2451 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2454 assert(CC < 16 && "Can only handle standard cond codes");
2455 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2457 default: llvm_unreachable("Illegal register size!");
2458 case 2: return Opc[Idx][0];
2459 case 4: return Opc[Idx][1];
2460 case 8: return Opc[Idx][2];
2464 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2465 if (!MI->isTerminator()) return false;
2467 // Conditional branch is a special case.
2468 if (MI->isBranch() && !MI->isBarrier())
2470 if (!MI->isPredicable())
2472 return !isPredicated(MI);
2475 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2476 MachineBasicBlock *&TBB,
2477 MachineBasicBlock *&FBB,
2478 SmallVectorImpl<MachineOperand> &Cond,
2479 bool AllowModify) const {
2480 // Start from the bottom of the block and work up, examining the
2481 // terminator instructions.
2482 MachineBasicBlock::iterator I = MBB.end();
2483 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2484 while (I != MBB.begin()) {
2486 if (I->isDebugValue())
2489 // Working from the bottom, when we see a non-terminator instruction, we're
2491 if (!isUnpredicatedTerminator(I))
2494 // A terminator that isn't a branch can't easily be handled by this
2499 // Handle unconditional branches.
2500 if (I->getOpcode() == X86::JMP_4) {
2504 TBB = I->getOperand(0).getMBB();
2508 // If the block has any instructions after a JMP, delete them.
2509 while (llvm::next(I) != MBB.end())
2510 llvm::next(I)->eraseFromParent();
2515 // Delete the JMP if it's equivalent to a fall-through.
2516 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2518 I->eraseFromParent();
2520 UnCondBrIter = MBB.end();
2524 // TBB is used to indicate the unconditional destination.
2525 TBB = I->getOperand(0).getMBB();
2529 // Handle conditional branches.
2530 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
2531 if (BranchCode == X86::COND_INVALID)
2532 return true; // Can't handle indirect branch.
2534 // Working from the bottom, handle the first conditional branch.
2536 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2537 if (AllowModify && UnCondBrIter != MBB.end() &&
2538 MBB.isLayoutSuccessor(TargetBB)) {
2539 // If we can modify the code and it ends in something like:
2547 // Then we can change this to:
2554 // Which is a bit more efficient.
2555 // We conditionally jump to the fall-through block.
2556 BranchCode = GetOppositeBranchCondition(BranchCode);
2557 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2558 MachineBasicBlock::iterator OldInst = I;
2560 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2561 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2562 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2565 OldInst->eraseFromParent();
2566 UnCondBrIter->eraseFromParent();
2568 // Restart the analysis.
2569 UnCondBrIter = MBB.end();
2575 TBB = I->getOperand(0).getMBB();
2576 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2580 // Handle subsequent conditional branches. Only handle the case where all
2581 // conditional branches branch to the same destination and their condition
2582 // opcodes fit one of the special multi-branch idioms.
2583 assert(Cond.size() == 1);
2586 // Only handle the case where all conditional branches branch to the same
2588 if (TBB != I->getOperand(0).getMBB())
2591 // If the conditions are the same, we can leave them alone.
2592 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2593 if (OldBranchCode == BranchCode)
2596 // If they differ, see if they fit one of the known patterns. Theoretically,
2597 // we could handle more patterns here, but we shouldn't expect to see them
2598 // if instruction selection has done a reasonable job.
2599 if ((OldBranchCode == X86::COND_NP &&
2600 BranchCode == X86::COND_E) ||
2601 (OldBranchCode == X86::COND_E &&
2602 BranchCode == X86::COND_NP))
2603 BranchCode = X86::COND_NP_OR_E;
2604 else if ((OldBranchCode == X86::COND_P &&
2605 BranchCode == X86::COND_NE) ||
2606 (OldBranchCode == X86::COND_NE &&
2607 BranchCode == X86::COND_P))
2608 BranchCode = X86::COND_NE_OR_P;
2612 // Update the MachineOperand.
2613 Cond[0].setImm(BranchCode);
2619 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2620 MachineBasicBlock::iterator I = MBB.end();
2623 while (I != MBB.begin()) {
2625 if (I->isDebugValue())
2627 if (I->getOpcode() != X86::JMP_4 &&
2628 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2630 // Remove the branch.
2631 I->eraseFromParent();
2640 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2641 MachineBasicBlock *FBB,
2642 const SmallVectorImpl<MachineOperand> &Cond,
2643 DebugLoc DL) const {
2644 // Shouldn't be a fall through.
2645 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2646 assert((Cond.size() == 1 || Cond.size() == 0) &&
2647 "X86 branch conditions have one component!");
2650 // Unconditional branch?
2651 assert(!FBB && "Unconditional branch with multiple successors!");
2652 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2656 // Conditional branch.
2658 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2660 case X86::COND_NP_OR_E:
2661 // Synthesize NP_OR_E with two branches.
2662 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2664 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2667 case X86::COND_NE_OR_P:
2668 // Synthesize NE_OR_P with two branches.
2669 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2671 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2675 unsigned Opc = GetCondBranchFromCond(CC);
2676 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2681 // Two-way Conditional branch. Insert the second branch.
2682 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2689 canInsertSelect(const MachineBasicBlock &MBB,
2690 const SmallVectorImpl<MachineOperand> &Cond,
2691 unsigned TrueReg, unsigned FalseReg,
2692 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2693 // Not all subtargets have cmov instructions.
2694 if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2696 if (Cond.size() != 1)
2698 // We cannot do the composite conditions, at least not in SSA form.
2699 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2702 // Check register classes.
2703 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2704 const TargetRegisterClass *RC =
2705 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2709 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2710 if (X86::GR16RegClass.hasSubClassEq(RC) ||
2711 X86::GR32RegClass.hasSubClassEq(RC) ||
2712 X86::GR64RegClass.hasSubClassEq(RC)) {
2713 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2714 // Bridge. Probably Ivy Bridge as well.
2721 // Can't do vectors.
2725 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2726 MachineBasicBlock::iterator I, DebugLoc DL,
2728 const SmallVectorImpl<MachineOperand> &Cond,
2729 unsigned TrueReg, unsigned FalseReg) const {
2730 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2731 assert(Cond.size() == 1 && "Invalid Cond array");
2732 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2733 MRI.getRegClass(DstReg)->getSize(),
2734 false/*HasMemoryOperand*/);
2735 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2738 /// isHReg - Test if the given register is a physical h register.
2739 static bool isHReg(unsigned Reg) {
2740 return X86::GR8_ABCD_HRegClass.contains(Reg);
2743 // Try and copy between VR128/VR64 and GR64 registers.
2744 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2746 // SrcReg(VR128) -> DestReg(GR64)
2747 // SrcReg(VR64) -> DestReg(GR64)
2748 // SrcReg(GR64) -> DestReg(VR128)
2749 // SrcReg(GR64) -> DestReg(VR64)
2751 if (X86::GR64RegClass.contains(DestReg)) {
2752 if (X86::VR128RegClass.contains(SrcReg))
2753 // Copy from a VR128 register to a GR64 register.
2754 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
2755 if (X86::VR64RegClass.contains(SrcReg))
2756 // Copy from a VR64 register to a GR64 register.
2757 return X86::MOVSDto64rr;
2758 } else if (X86::GR64RegClass.contains(SrcReg)) {
2759 // Copy from a GR64 register to a VR128 register.
2760 if (X86::VR128RegClass.contains(DestReg))
2761 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
2762 // Copy from a GR64 register to a VR64 register.
2763 if (X86::VR64RegClass.contains(DestReg))
2764 return X86::MOV64toSDrr;
2767 // SrcReg(FR32) -> DestReg(GR32)
2768 // SrcReg(GR32) -> DestReg(FR32)
2770 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
2771 // Copy from a FR32 register to a GR32 register.
2772 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
2774 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
2775 // Copy from a GR32 register to a FR32 register.
2776 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
2781 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2782 MachineBasicBlock::iterator MI, DebugLoc DL,
2783 unsigned DestReg, unsigned SrcReg,
2784 bool KillSrc) const {
2785 // First deal with the normal symmetric copies.
2786 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2788 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2790 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2792 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2794 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2795 // Copying to or from a physical H register on x86-64 requires a NOREX
2796 // move. Otherwise use a normal move.
2797 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
2798 TM.getSubtarget<X86Subtarget>().is64Bit()) {
2799 Opc = X86::MOV8rr_NOREX;
2800 // Both operands must be encodable without an REX prefix.
2801 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2802 "8-bit H register can not be copied outside GR8_NOREX");
2805 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
2806 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
2807 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2808 Opc = X86::VMOVAPSYrr;
2809 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2810 Opc = X86::MMX_MOVQ64rr;
2812 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
2815 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2816 .addReg(SrcReg, getKillRegState(KillSrc));
2820 // Moving EFLAGS to / from another register requires a push and a pop.
2821 if (SrcReg == X86::EFLAGS) {
2822 if (X86::GR64RegClass.contains(DestReg)) {
2823 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2824 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2827 if (X86::GR32RegClass.contains(DestReg)) {
2828 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2829 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2833 if (DestReg == X86::EFLAGS) {
2834 if (X86::GR64RegClass.contains(SrcReg)) {
2835 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2836 .addReg(SrcReg, getKillRegState(KillSrc));
2837 BuildMI(MBB, MI, DL, get(X86::POPF64));
2840 if (X86::GR32RegClass.contains(SrcReg)) {
2841 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2842 .addReg(SrcReg, getKillRegState(KillSrc));
2843 BuildMI(MBB, MI, DL, get(X86::POPF32));
2848 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2849 << " to " << RI.getName(DestReg) << '\n');
2850 llvm_unreachable("Cannot emit physreg copy instruction");
2853 static unsigned getLoadStoreRegOpcode(unsigned Reg,
2854 const TargetRegisterClass *RC,
2855 bool isStackAligned,
2856 const TargetMachine &TM,
2858 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2859 switch (RC->getSize()) {
2861 llvm_unreachable("Unknown spill size");
2863 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
2864 if (TM.getSubtarget<X86Subtarget>().is64Bit())
2865 // Copying to or from a physical H register on x86-64 requires a NOREX
2866 // move. Otherwise use a normal move.
2867 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2868 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2869 return load ? X86::MOV8rm : X86::MOV8mr;
2871 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2872 return load ? X86::MOV16rm : X86::MOV16mr;
2874 if (X86::GR32RegClass.hasSubClassEq(RC))
2875 return load ? X86::MOV32rm : X86::MOV32mr;
2876 if (X86::FR32RegClass.hasSubClassEq(RC))
2878 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2879 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
2880 if (X86::RFP32RegClass.hasSubClassEq(RC))
2881 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2882 llvm_unreachable("Unknown 4-byte regclass");
2884 if (X86::GR64RegClass.hasSubClassEq(RC))
2885 return load ? X86::MOV64rm : X86::MOV64mr;
2886 if (X86::FR64RegClass.hasSubClassEq(RC))
2888 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2889 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
2890 if (X86::VR64RegClass.hasSubClassEq(RC))
2891 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2892 if (X86::RFP64RegClass.hasSubClassEq(RC))
2893 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2894 llvm_unreachable("Unknown 8-byte regclass");
2896 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
2897 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
2899 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
2900 // If stack is realigned we can use aligned stores.
2903 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2904 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
2907 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2908 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2911 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2912 // If stack is realigned we can use aligned stores.
2914 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2916 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
2920 static unsigned getStoreRegOpcode(unsigned SrcReg,
2921 const TargetRegisterClass *RC,
2922 bool isStackAligned,
2923 TargetMachine &TM) {
2924 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2928 static unsigned getLoadRegOpcode(unsigned DestReg,
2929 const TargetRegisterClass *RC,
2930 bool isStackAligned,
2931 const TargetMachine &TM) {
2932 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
2935 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2936 MachineBasicBlock::iterator MI,
2937 unsigned SrcReg, bool isKill, int FrameIdx,
2938 const TargetRegisterClass *RC,
2939 const TargetRegisterInfo *TRI) const {
2940 const MachineFunction &MF = *MBB.getParent();
2941 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2942 "Stack slot too small for store");
2943 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2944 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
2945 RI.canRealignStack(MF);
2946 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
2947 DebugLoc DL = MBB.findDebugLoc(MI);
2948 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
2949 .addReg(SrcReg, getKillRegState(isKill));
2952 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2954 SmallVectorImpl<MachineOperand> &Addr,
2955 const TargetRegisterClass *RC,
2956 MachineInstr::mmo_iterator MMOBegin,
2957 MachineInstr::mmo_iterator MMOEnd,
2958 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2959 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2960 bool isAligned = MMOBegin != MMOEnd &&
2961 (*MMOBegin)->getAlignment() >= Alignment;
2962 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
2964 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
2965 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
2966 MIB.addOperand(Addr[i]);
2967 MIB.addReg(SrcReg, getKillRegState(isKill));
2968 (*MIB).setMemRefs(MMOBegin, MMOEnd);
2969 NewMIs.push_back(MIB);
2973 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
2974 MachineBasicBlock::iterator MI,
2975 unsigned DestReg, int FrameIdx,
2976 const TargetRegisterClass *RC,
2977 const TargetRegisterInfo *TRI) const {
2978 const MachineFunction &MF = *MBB.getParent();
2979 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2980 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
2981 RI.canRealignStack(MF);
2982 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2983 DebugLoc DL = MBB.findDebugLoc(MI);
2984 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
2987 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
2988 SmallVectorImpl<MachineOperand> &Addr,
2989 const TargetRegisterClass *RC,
2990 MachineInstr::mmo_iterator MMOBegin,
2991 MachineInstr::mmo_iterator MMOEnd,
2992 SmallVectorImpl<MachineInstr*> &NewMIs) const {
2993 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2994 bool isAligned = MMOBegin != MMOEnd &&
2995 (*MMOBegin)->getAlignment() >= Alignment;
2996 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
2998 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
2999 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3000 MIB.addOperand(Addr[i]);
3001 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3002 NewMIs.push_back(MIB);
3006 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3007 int &CmpMask, int &CmpValue) const {
3008 switch (MI->getOpcode()) {
3010 case X86::CMP64ri32:
3017 SrcReg = MI->getOperand(0).getReg();
3020 CmpValue = MI->getOperand(1).getImm();
3022 // A SUB can be used to perform comparison.
3027 SrcReg = MI->getOperand(1).getReg();
3036 SrcReg = MI->getOperand(1).getReg();
3037 SrcReg2 = MI->getOperand(2).getReg();
3041 case X86::SUB64ri32:
3048 SrcReg = MI->getOperand(1).getReg();
3051 CmpValue = MI->getOperand(2).getImm();
3057 SrcReg = MI->getOperand(0).getReg();
3058 SrcReg2 = MI->getOperand(1).getReg();
3066 SrcReg = MI->getOperand(0).getReg();
3067 if (MI->getOperand(1).getReg() != SrcReg) return false;
3068 // Compare against zero.
3077 /// isRedundantFlagInstr - check whether the first instruction, whose only
3078 /// purpose is to update flags, can be made redundant.
3079 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3080 /// This function can be extended later on.
3081 /// SrcReg, SrcRegs: register operands for FlagI.
3082 /// ImmValue: immediate for FlagI if it takes an immediate.
3083 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3084 unsigned SrcReg2, int ImmValue,
3086 if (((FlagI->getOpcode() == X86::CMP64rr &&
3087 OI->getOpcode() == X86::SUB64rr) ||
3088 (FlagI->getOpcode() == X86::CMP32rr &&
3089 OI->getOpcode() == X86::SUB32rr)||
3090 (FlagI->getOpcode() == X86::CMP16rr &&
3091 OI->getOpcode() == X86::SUB16rr)||
3092 (FlagI->getOpcode() == X86::CMP8rr &&
3093 OI->getOpcode() == X86::SUB8rr)) &&
3094 ((OI->getOperand(1).getReg() == SrcReg &&
3095 OI->getOperand(2).getReg() == SrcReg2) ||
3096 (OI->getOperand(1).getReg() == SrcReg2 &&
3097 OI->getOperand(2).getReg() == SrcReg)))
3100 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3101 OI->getOpcode() == X86::SUB64ri32) ||
3102 (FlagI->getOpcode() == X86::CMP64ri8 &&
3103 OI->getOpcode() == X86::SUB64ri8) ||
3104 (FlagI->getOpcode() == X86::CMP32ri &&
3105 OI->getOpcode() == X86::SUB32ri) ||
3106 (FlagI->getOpcode() == X86::CMP32ri8 &&
3107 OI->getOpcode() == X86::SUB32ri8) ||
3108 (FlagI->getOpcode() == X86::CMP16ri &&
3109 OI->getOpcode() == X86::SUB16ri) ||
3110 (FlagI->getOpcode() == X86::CMP16ri8 &&
3111 OI->getOpcode() == X86::SUB16ri8) ||
3112 (FlagI->getOpcode() == X86::CMP8ri &&
3113 OI->getOpcode() == X86::SUB8ri)) &&
3114 OI->getOperand(1).getReg() == SrcReg &&
3115 OI->getOperand(2).getImm() == ImmValue)
3120 /// isDefConvertible - check whether the definition can be converted
3121 /// to remove a comparison against zero.
3122 inline static bool isDefConvertible(MachineInstr *MI) {
3123 switch (MI->getOpcode()) {
3124 default: return false;
3125 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3126 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3127 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3128 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3129 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3130 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3131 case X86::DEC64m: case X86::DEC32m: case X86::DEC16m: case X86::DEC8m:
3132 case X86::DEC64_32r: case X86::DEC64_16r:
3133 case X86::DEC64_32m: case X86::DEC64_16m:
3134 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3135 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3136 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3137 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3138 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3139 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3140 case X86::INC64m: case X86::INC32m: case X86::INC16m: case X86::INC8m:
3141 case X86::INC64_32r: case X86::INC64_16r:
3142 case X86::INC64_32m: case X86::INC64_16m:
3143 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3144 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3145 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3146 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3147 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3148 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3149 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3150 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3151 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3152 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3153 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3154 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3155 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3156 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3157 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3162 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3163 /// operates on the same source operands and sets flags in the same way as
3164 /// Compare; remove Compare if possible.
3166 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3167 int CmpMask, int CmpValue,
3168 const MachineRegisterInfo *MRI) const {
3169 // Check whether we can replace SUB with CMP.
3170 unsigned NewOpcode = 0;
3171 switch (CmpInstr->getOpcode()) {
3173 case X86::SUB64ri32:
3188 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3190 // There is no use of the destination register, we can replace SUB with CMP.
3191 switch (CmpInstr->getOpcode()) {
3192 default: llvm_unreachable("Unreachable!");
3193 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3194 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3195 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3196 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3197 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3198 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3199 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3200 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3201 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3202 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3203 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3204 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3205 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3206 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3207 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3209 CmpInstr->setDesc(get(NewOpcode));
3210 CmpInstr->RemoveOperand(0);
3211 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3212 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3213 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3218 // Get the unique definition of SrcReg.
3219 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3220 if (!MI) return false;
3222 // CmpInstr is the first instruction of the BB.
3223 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3225 // If we are comparing against zero, check whether we can use MI to update
3226 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3227 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3228 if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3229 !isDefConvertible(MI)))
3232 // We are searching for an earlier instruction that can make CmpInstr
3233 // redundant and that instruction will be saved in Sub.
3234 MachineInstr *Sub = NULL;
3235 const TargetRegisterInfo *TRI = &getRegisterInfo();
3237 // We iterate backward, starting from the instruction before CmpInstr and
3238 // stop when reaching the definition of a source register or done with the BB.
3239 // RI points to the instruction before CmpInstr.
3240 // If the definition is in this basic block, RE points to the definition;
3241 // otherwise, RE is the rend of the basic block.
3242 MachineBasicBlock::reverse_iterator
3243 RI = MachineBasicBlock::reverse_iterator(I),
3244 RE = CmpInstr->getParent() == MI->getParent() ?
3245 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3246 CmpInstr->getParent()->rend();
3247 MachineInstr *Movr0Inst = 0;
3248 for (; RI != RE; ++RI) {
3249 MachineInstr *Instr = &*RI;
3250 // Check whether CmpInstr can be made redundant by the current instruction.
3252 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3257 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3258 Instr->readsRegister(X86::EFLAGS, TRI)) {
3259 // This instruction modifies or uses EFLAGS.
3261 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3262 // They are safe to move up, if the definition to EFLAGS is dead and
3263 // earlier instructions do not read or write EFLAGS.
3264 if (!Movr0Inst && (Instr->getOpcode() == X86::MOV8r0 ||
3265 Instr->getOpcode() == X86::MOV16r0 ||
3266 Instr->getOpcode() == X86::MOV32r0 ||
3267 Instr->getOpcode() == X86::MOV64r0) &&
3268 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3273 // We can't remove CmpInstr.
3278 // Return false if no candidates exist.
3279 if (!IsCmpZero && !Sub)
3282 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3283 Sub->getOperand(2).getReg() == SrcReg);
3285 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3286 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3287 // If we are done with the basic block, we need to check whether EFLAGS is
3289 bool IsSafe = false;
3290 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3291 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3292 for (++I; I != E; ++I) {
3293 const MachineInstr &Instr = *I;
3294 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3295 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3296 // We should check the usage if this instruction uses and updates EFLAGS.
3297 if (!UseEFLAGS && ModifyEFLAGS) {
3298 // It is safe to remove CmpInstr if EFLAGS is updated again.
3302 if (!UseEFLAGS && !ModifyEFLAGS)
3305 // EFLAGS is used by this instruction.
3306 X86::CondCode OldCC;
3307 bool OpcIsSET = false;
3308 if (IsCmpZero || IsSwapped) {
3309 // We decode the condition code from opcode.
3310 if (Instr.isBranch())
3311 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3313 OldCC = getCondFromSETOpc(Instr.getOpcode());
3314 if (OldCC != X86::COND_INVALID)
3317 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3319 if (OldCC == X86::COND_INVALID) return false;
3324 case X86::COND_A: case X86::COND_AE:
3325 case X86::COND_B: case X86::COND_BE:
3326 case X86::COND_G: case X86::COND_GE:
3327 case X86::COND_L: case X86::COND_LE:
3328 case X86::COND_O: case X86::COND_NO:
3329 // CF and OF are used, we can't perform this optimization.
3332 } else if (IsSwapped) {
3333 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3334 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3335 // We swap the condition code and synthesize the new opcode.
3336 X86::CondCode NewCC = getSwappedCondition(OldCC);
3337 if (NewCC == X86::COND_INVALID) return false;
3339 // Synthesize the new opcode.
3340 bool HasMemoryOperand = Instr.hasOneMemOperand();
3342 if (Instr.isBranch())
3343 NewOpc = GetCondBranchFromCond(NewCC);
3345 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3347 unsigned DstReg = Instr.getOperand(0).getReg();
3348 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3352 // Push the MachineInstr to OpsToUpdate.
3353 // If it is safe to remove CmpInstr, the condition code of these
3354 // instructions will be modified.
3355 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3357 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3358 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3364 // If EFLAGS is not killed nor re-defined, we should check whether it is
3365 // live-out. If it is live-out, do not optimize.
3366 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3367 MachineBasicBlock *MBB = CmpInstr->getParent();
3368 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3369 SE = MBB->succ_end(); SI != SE; ++SI)
3370 if ((*SI)->isLiveIn(X86::EFLAGS))
3374 // The instruction to be updated is either Sub or MI.
3375 Sub = IsCmpZero ? MI : Sub;
3376 // Move Movr0Inst to the place right before Sub.
3378 Sub->getParent()->remove(Movr0Inst);
3379 Sub->getParent()->insert(MachineBasicBlock::iterator(Sub), Movr0Inst);
3382 // Make sure Sub instruction defines EFLAGS and mark the def live.
3383 unsigned LastOperand = Sub->getNumOperands() - 1;
3384 assert(Sub->getNumOperands() >= 2 &&
3385 Sub->getOperand(LastOperand).isReg() &&
3386 Sub->getOperand(LastOperand).getReg() == X86::EFLAGS &&
3387 "EFLAGS should be the last operand of SUB, ADD, OR, XOR, AND");
3388 Sub->getOperand(LastOperand).setIsDef(true);
3389 Sub->getOperand(LastOperand).setIsDead(false);
3390 CmpInstr->eraseFromParent();
3392 // Modify the condition code of instructions in OpsToUpdate.
3393 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3394 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3398 /// optimizeLoadInstr - Try to remove the load by folding it to a register
3399 /// operand at the use. We fold the load instructions if load defines a virtual
3400 /// register, the virtual register is used once in the same BB, and the
3401 /// instructions in-between do not load or store, and have no side effects.
3402 MachineInstr* X86InstrInfo::
3403 optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3404 unsigned &FoldAsLoadDefReg,
3405 MachineInstr *&DefMI) const {
3406 if (FoldAsLoadDefReg == 0)
3408 // To be conservative, if there exists another load, clear the load candidate.
3409 if (MI->mayLoad()) {
3410 FoldAsLoadDefReg = 0;
3414 // Check whether we can move DefMI here.
3415 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3417 bool SawStore = false;
3418 if (!DefMI->isSafeToMove(this, 0, SawStore))
3421 // We try to commute MI if possible.
3422 unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3423 for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3424 // Collect information about virtual register operands of MI.
3425 unsigned SrcOperandId = 0;
3426 bool FoundSrcOperand = false;
3427 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3428 MachineOperand &MO = MI->getOperand(i);
3431 unsigned Reg = MO.getReg();
3432 if (Reg != FoldAsLoadDefReg)
3434 // Do not fold if we have a subreg use or a def or multiple uses.
3435 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3439 FoundSrcOperand = true;
3441 if (!FoundSrcOperand) return 0;
3443 // Check whether we can fold the def into SrcOperandId.
3444 SmallVector<unsigned, 8> Ops;
3445 Ops.push_back(SrcOperandId);
3446 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3448 FoldAsLoadDefReg = 0;
3453 // MI was changed but it didn't help, commute it back!
3454 commuteInstruction(MI, false);
3458 // Check whether we can commute MI and enable folding.
3459 if (MI->isCommutable()) {
3460 MachineInstr *NewMI = commuteInstruction(MI, false);
3461 // Unable to commute.
3462 if (!NewMI) return 0;
3464 // New instruction. It doesn't need to be kept.
3465 NewMI->eraseFromParent();
3473 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3474 /// instruction with two undef reads of the register being defined. This is
3475 /// used for mapping:
3478 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3480 static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
3481 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3482 unsigned Reg = MI->getOperand(0).getReg();
3485 // MachineInstr::addOperand() will insert explicit operands before any
3486 // implicit operands.
3487 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
3488 .addReg(Reg, RegState::Undef);
3489 // But we don't trust that.
3490 assert(MI->getOperand(1).getReg() == Reg &&
3491 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
3495 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3496 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3497 switch (MI->getOpcode()) {
3501 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
3503 assert(HasAVX && "AVX not supported");
3504 return Expand2AddrUndef(MI, get(X86::VXORPSYrr));
3505 case X86::V_SETALLONES:
3506 return Expand2AddrUndef(MI, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3507 case X86::AVX2_SETALLONES:
3508 return Expand2AddrUndef(MI, get(X86::VPCMPEQDYrr));
3509 case X86::TEST8ri_NOREX:
3510 MI->setDesc(get(X86::TEST8ri));
3517 X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
3518 int FrameIx, uint64_t Offset,
3519 const MDNode *MDPtr,
3520 DebugLoc DL) const {
3522 AM.BaseType = X86AddressMode::FrameIndexBase;
3523 AM.Base.FrameIndex = FrameIx;
3524 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
3525 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
3529 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
3530 const SmallVectorImpl<MachineOperand> &MOs,
3532 const TargetInstrInfo &TII) {
3533 // Create the base instruction with the memory operand as the first part.
3534 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3535 MI->getDebugLoc(), true);
3536 MachineInstrBuilder MIB(NewMI);
3537 unsigned NumAddrOps = MOs.size();
3538 for (unsigned i = 0; i != NumAddrOps; ++i)
3539 MIB.addOperand(MOs[i]);
3540 if (NumAddrOps < 4) // FrameIndex only
3543 // Loop over the rest of the ri operands, converting them over.
3544 unsigned NumOps = MI->getDesc().getNumOperands()-2;
3545 for (unsigned i = 0; i != NumOps; ++i) {
3546 MachineOperand &MO = MI->getOperand(i+2);
3549 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3550 MachineOperand &MO = MI->getOperand(i);
3556 static MachineInstr *FuseInst(MachineFunction &MF,
3557 unsigned Opcode, unsigned OpNo,
3558 const SmallVectorImpl<MachineOperand> &MOs,
3559 MachineInstr *MI, const TargetInstrInfo &TII) {
3560 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3561 MI->getDebugLoc(), true);
3562 MachineInstrBuilder MIB(NewMI);
3564 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3565 MachineOperand &MO = MI->getOperand(i);
3567 assert(MO.isReg() && "Expected to fold into reg operand!");
3568 unsigned NumAddrOps = MOs.size();
3569 for (unsigned i = 0; i != NumAddrOps; ++i)
3570 MIB.addOperand(MOs[i]);
3571 if (NumAddrOps < 4) // FrameIndex only
3580 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
3581 const SmallVectorImpl<MachineOperand> &MOs,
3583 MachineFunction &MF = *MI->getParent()->getParent();
3584 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
3586 unsigned NumAddrOps = MOs.size();
3587 for (unsigned i = 0; i != NumAddrOps; ++i)
3588 MIB.addOperand(MOs[i]);
3589 if (NumAddrOps < 4) // FrameIndex only
3591 return MIB.addImm(0);
3595 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3596 MachineInstr *MI, unsigned i,
3597 const SmallVectorImpl<MachineOperand> &MOs,
3598 unsigned Size, unsigned Align) const {
3599 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
3600 bool isTwoAddrFold = false;
3601 unsigned NumOps = MI->getDesc().getNumOperands();
3602 bool isTwoAddr = NumOps > 1 &&
3603 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
3605 // FIXME: AsmPrinter doesn't know how to handle
3606 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3607 if (MI->getOpcode() == X86::ADD32ri &&
3608 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3611 MachineInstr *NewMI = NULL;
3612 // Folding a memory location into the two-address part of a two-address
3613 // instruction is different than folding it other places. It requires
3614 // replacing the *two* registers with the memory location.
3615 if (isTwoAddr && NumOps >= 2 && i < 2 &&
3616 MI->getOperand(0).isReg() &&
3617 MI->getOperand(1).isReg() &&
3618 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
3619 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3620 isTwoAddrFold = true;
3621 } else if (i == 0) { // If operand 0
3623 switch (MI->getOpcode()) {
3625 case X86::MOV64r0: Opc = X86::MOV64mi32; break;
3626 case X86::MOV32r0: Opc = X86::MOV32mi; break;
3627 case X86::MOV16r0: Opc = X86::MOV16mi; break;
3628 case X86::MOV8r0: Opc = X86::MOV8mi; break;
3631 NewMI = MakeM0Inst(*this, Opc, MOs, MI);
3635 OpcodeTablePtr = &RegOp2MemOpTable0;
3636 } else if (i == 1) {
3637 OpcodeTablePtr = &RegOp2MemOpTable1;
3638 } else if (i == 2) {
3639 OpcodeTablePtr = &RegOp2MemOpTable2;
3640 } else if (i == 3) {
3641 OpcodeTablePtr = &RegOp2MemOpTable3;
3644 // If table selected...
3645 if (OpcodeTablePtr) {
3646 // Find the Opcode to fuse
3647 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3648 OpcodeTablePtr->find(MI->getOpcode());
3649 if (I != OpcodeTablePtr->end()) {
3650 unsigned Opcode = I->second.first;
3651 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
3652 if (Align < MinAlign)
3654 bool NarrowToMOV32rm = false;
3656 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
3657 if (Size < RCSize) {
3658 // Check if it's safe to fold the load. If the size of the object is
3659 // narrower than the load width, then it's not.
3660 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
3662 // If this is a 64-bit load, but the spill slot is 32, then we can do
3663 // a 32-bit load which is implicitly zero-extended. This likely is due
3664 // to liveintervalanalysis remat'ing a load from stack slot.
3665 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
3667 Opcode = X86::MOV32rm;
3668 NarrowToMOV32rm = true;
3673 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
3675 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
3677 if (NarrowToMOV32rm) {
3678 // If this is the special case where we use a MOV32rm to load a 32-bit
3679 // value and zero-extend the top bits. Change the destination register
3681 unsigned DstReg = NewMI->getOperand(0).getReg();
3682 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
3683 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
3686 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
3693 if (PrintFailedFusing && !MI->isCopy())
3694 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
3698 /// hasPartialRegUpdate - Return true for all instructions that only update
3699 /// the first 32 or 64-bits of the destination register and leave the rest
3700 /// unmodified. This can be used to avoid folding loads if the instructions
3701 /// only update part of the destination register, and the non-updated part is
3702 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3703 /// instructions breaks the partial register dependency and it can improve
3704 /// performance. e.g.:
3706 /// movss (%rdi), %xmm0
3707 /// cvtss2sd %xmm0, %xmm0
3710 /// cvtss2sd (%rdi), %xmm0
3712 /// FIXME: This should be turned into a TSFlags.
3714 static bool hasPartialRegUpdate(unsigned Opcode) {
3716 case X86::CVTSI2SSrr:
3717 case X86::CVTSI2SS64rr:
3718 case X86::CVTSI2SDrr:
3719 case X86::CVTSI2SD64rr:
3720 case X86::CVTSD2SSrr:
3721 case X86::Int_CVTSD2SSrr:
3722 case X86::CVTSS2SDrr:
3723 case X86::Int_CVTSS2SDrr:
3725 case X86::RCPSSr_Int:
3727 case X86::ROUNDSDr_Int:
3729 case X86::ROUNDSSr_Int:
3731 case X86::RSQRTSSr_Int:
3733 case X86::SQRTSSr_Int:
3734 // AVX encoded versions
3735 case X86::VCVTSD2SSrr:
3736 case X86::Int_VCVTSD2SSrr:
3737 case X86::VCVTSS2SDrr:
3738 case X86::Int_VCVTSS2SDrr:
3740 case X86::VROUNDSDr:
3741 case X86::VROUNDSDr_Int:
3742 case X86::VROUNDSSr:
3743 case X86::VROUNDSSr_Int:
3744 case X86::VRSQRTSSr:
3752 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3753 /// instructions we would like before a partial register update.
3754 unsigned X86InstrInfo::
3755 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3756 const TargetRegisterInfo *TRI) const {
3757 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3760 // If MI is marked as reading Reg, the partial register update is wanted.
3761 const MachineOperand &MO = MI->getOperand(0);
3762 unsigned Reg = MO.getReg();
3763 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3764 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3767 if (MI->readsRegister(Reg, TRI))
3771 // If any of the preceding 16 instructions are reading Reg, insert a
3772 // dependency breaking instruction. The magic number is based on a few
3773 // Nehalem experiments.
3778 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3779 const TargetRegisterInfo *TRI) const {
3780 unsigned Reg = MI->getOperand(OpNum).getReg();
3781 if (X86::VR128RegClass.contains(Reg)) {
3782 // These instructions are all floating point domain, so xorps is the best
3784 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3785 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3786 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3787 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3788 } else if (X86::VR256RegClass.contains(Reg)) {
3789 // Use vxorps to clear the full ymm register.
3790 // It wants to read and write the xmm sub-register.
3791 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3792 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3793 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3794 .addReg(Reg, RegState::ImplicitDefine);
3797 MI->addRegisterKilled(Reg, TRI, true);
3800 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3802 const SmallVectorImpl<unsigned> &Ops,
3803 int FrameIndex) const {
3804 // Check switch flag
3805 if (NoFusing) return NULL;
3807 // Unless optimizing for size, don't fold to avoid partial
3808 // register update stalls
3809 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3810 hasPartialRegUpdate(MI->getOpcode()))
3813 const MachineFrameInfo *MFI = MF.getFrameInfo();
3814 unsigned Size = MFI->getObjectSize(FrameIndex);
3815 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
3816 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3817 unsigned NewOpc = 0;
3818 unsigned RCSize = 0;
3819 switch (MI->getOpcode()) {
3820 default: return NULL;
3821 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
3822 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3823 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3824 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
3826 // Check if it's safe to fold the load. If the size of the object is
3827 // narrower than the load width, then it's not.
3830 // Change to CMPXXri r, 0 first.
3831 MI->setDesc(get(NewOpc));
3832 MI->getOperand(1).ChangeToImmediate(0);
3833 } else if (Ops.size() != 1)
3836 SmallVector<MachineOperand,4> MOs;
3837 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
3838 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
3841 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3843 const SmallVectorImpl<unsigned> &Ops,
3844 MachineInstr *LoadMI) const {
3845 // Check switch flag
3846 if (NoFusing) return NULL;
3848 // Unless optimizing for size, don't fold to avoid partial
3849 // register update stalls
3850 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3851 hasPartialRegUpdate(MI->getOpcode()))
3854 // Determine the alignment of the load.
3855 unsigned Alignment = 0;
3856 if (LoadMI->hasOneMemOperand())
3857 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
3859 switch (LoadMI->getOpcode()) {
3860 case X86::AVX2_SETALLONES:
3865 case X86::V_SETALLONES:
3877 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3878 unsigned NewOpc = 0;
3879 switch (MI->getOpcode()) {
3880 default: return NULL;
3881 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
3882 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3883 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3884 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
3886 // Change to CMPXXri r, 0 first.
3887 MI->setDesc(get(NewOpc));
3888 MI->getOperand(1).ChangeToImmediate(0);
3889 } else if (Ops.size() != 1)
3892 // Make sure the subregisters match.
3893 // Otherwise we risk changing the size of the load.
3894 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3897 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
3898 switch (LoadMI->getOpcode()) {
3900 case X86::V_SETALLONES:
3901 case X86::AVX2_SETALLONES:
3904 case X86::FsFLD0SS: {
3905 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
3906 // Create a constant-pool entry and operands to load from it.
3908 // Medium and large mode can't fold loads this way.
3909 if (TM.getCodeModel() != CodeModel::Small &&
3910 TM.getCodeModel() != CodeModel::Kernel)
3913 // x86-32 PIC requires a PIC base register for constant pools.
3914 unsigned PICBase = 0;
3915 if (TM.getRelocationModel() == Reloc::PIC_) {
3916 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3919 // FIXME: PICBase = getGlobalBaseReg(&MF);
3920 // This doesn't work for several reasons.
3921 // 1. GlobalBaseReg may have been spilled.
3922 // 2. It may not be live at MI.
3926 // Create a constant-pool entry.
3927 MachineConstantPool &MCP = *MF.getConstantPool();
3929 unsigned Opc = LoadMI->getOpcode();
3930 if (Opc == X86::FsFLD0SS)
3931 Ty = Type::getFloatTy(MF.getFunction()->getContext());
3932 else if (Opc == X86::FsFLD0SD)
3933 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
3934 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
3935 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
3937 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
3939 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
3940 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3941 Constant::getNullValue(Ty);
3942 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
3944 // Create operands to load from the constant pool entry.
3945 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3946 MOs.push_back(MachineOperand::CreateImm(1));
3947 MOs.push_back(MachineOperand::CreateReg(0, false));
3948 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
3949 MOs.push_back(MachineOperand::CreateReg(0, false));
3953 // Folding a normal load. Just copy the load's address operands.
3954 unsigned NumOps = LoadMI->getDesc().getNumOperands();
3955 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
3956 MOs.push_back(LoadMI->getOperand(i));
3960 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
3964 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3965 const SmallVectorImpl<unsigned> &Ops) const {
3966 // Check switch flag
3967 if (NoFusing) return 0;
3969 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3970 switch (MI->getOpcode()) {
3971 default: return false;
3978 // FIXME: AsmPrinter doesn't know how to handle
3979 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3980 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3986 if (Ops.size() != 1)
3989 unsigned OpNum = Ops[0];
3990 unsigned Opc = MI->getOpcode();
3991 unsigned NumOps = MI->getDesc().getNumOperands();
3992 bool isTwoAddr = NumOps > 1 &&
3993 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
3995 // Folding a memory location into the two-address part of a two-address
3996 // instruction is different than folding it other places. It requires
3997 // replacing the *two* registers with the memory location.
3998 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
3999 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4000 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4001 } else if (OpNum == 0) { // If operand 0
4006 case X86::MOV64r0: return true;
4009 OpcodeTablePtr = &RegOp2MemOpTable0;
4010 } else if (OpNum == 1) {
4011 OpcodeTablePtr = &RegOp2MemOpTable1;
4012 } else if (OpNum == 2) {
4013 OpcodeTablePtr = &RegOp2MemOpTable2;
4014 } else if (OpNum == 3) {
4015 OpcodeTablePtr = &RegOp2MemOpTable3;
4018 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4020 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
4023 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4024 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4025 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4026 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4027 MemOp2RegOpTable.find(MI->getOpcode());
4028 if (I == MemOp2RegOpTable.end())
4030 unsigned Opc = I->second.first;
4031 unsigned Index = I->second.second & TB_INDEX_MASK;
4032 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4033 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4034 if (UnfoldLoad && !FoldedLoad)
4036 UnfoldLoad &= FoldedLoad;
4037 if (UnfoldStore && !FoldedStore)
4039 UnfoldStore &= FoldedStore;
4041 const MCInstrDesc &MCID = get(Opc);
4042 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4043 if (!MI->hasOneMemOperand() &&
4044 RC == &X86::VR128RegClass &&
4045 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4046 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4047 // conservatively assume the address is unaligned. That's bad for
4050 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4051 SmallVector<MachineOperand,2> BeforeOps;
4052 SmallVector<MachineOperand,2> AfterOps;
4053 SmallVector<MachineOperand,4> ImpOps;
4054 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4055 MachineOperand &Op = MI->getOperand(i);
4056 if (i >= Index && i < Index + X86::AddrNumOperands)
4057 AddrOps.push_back(Op);
4058 else if (Op.isReg() && Op.isImplicit())
4059 ImpOps.push_back(Op);
4061 BeforeOps.push_back(Op);
4063 AfterOps.push_back(Op);
4066 // Emit the load instruction.
4068 std::pair<MachineInstr::mmo_iterator,
4069 MachineInstr::mmo_iterator> MMOs =
4070 MF.extractLoadMemRefs(MI->memoperands_begin(),
4071 MI->memoperands_end());
4072 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4074 // Address operands cannot be marked isKill.
4075 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4076 MachineOperand &MO = NewMIs[0]->getOperand(i);
4078 MO.setIsKill(false);
4083 // Emit the data processing instruction.
4084 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4085 MachineInstrBuilder MIB(DataMI);
4088 MIB.addReg(Reg, RegState::Define);
4089 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4090 MIB.addOperand(BeforeOps[i]);
4093 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4094 MIB.addOperand(AfterOps[i]);
4095 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4096 MachineOperand &MO = ImpOps[i];
4097 MIB.addReg(MO.getReg(),
4098 getDefRegState(MO.isDef()) |
4099 RegState::Implicit |
4100 getKillRegState(MO.isKill()) |
4101 getDeadRegState(MO.isDead()) |
4102 getUndefRegState(MO.isUndef()));
4104 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4105 switch (DataMI->getOpcode()) {
4107 case X86::CMP64ri32:
4114 MachineOperand &MO0 = DataMI->getOperand(0);
4115 MachineOperand &MO1 = DataMI->getOperand(1);
4116 if (MO1.getImm() == 0) {
4118 switch (DataMI->getOpcode()) {
4119 default: llvm_unreachable("Unreachable!");
4121 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
4123 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
4125 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4126 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4128 DataMI->setDesc(get(NewOpc));
4129 MO1.ChangeToRegister(MO0.getReg(), false);
4133 NewMIs.push_back(DataMI);
4135 // Emit the store instruction.
4137 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
4138 std::pair<MachineInstr::mmo_iterator,
4139 MachineInstr::mmo_iterator> MMOs =
4140 MF.extractStoreMemRefs(MI->memoperands_begin(),
4141 MI->memoperands_end());
4142 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
4149 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
4150 SmallVectorImpl<SDNode*> &NewNodes) const {
4151 if (!N->isMachineOpcode())
4154 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4155 MemOp2RegOpTable.find(N->getMachineOpcode());
4156 if (I == MemOp2RegOpTable.end())
4158 unsigned Opc = I->second.first;
4159 unsigned Index = I->second.second & TB_INDEX_MASK;
4160 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4161 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4162 const MCInstrDesc &MCID = get(Opc);
4163 MachineFunction &MF = DAG.getMachineFunction();
4164 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4165 unsigned NumDefs = MCID.NumDefs;
4166 std::vector<SDValue> AddrOps;
4167 std::vector<SDValue> BeforeOps;
4168 std::vector<SDValue> AfterOps;
4169 DebugLoc dl = N->getDebugLoc();
4170 unsigned NumOps = N->getNumOperands();
4171 for (unsigned i = 0; i != NumOps-1; ++i) {
4172 SDValue Op = N->getOperand(i);
4173 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
4174 AddrOps.push_back(Op);
4175 else if (i < Index-NumDefs)
4176 BeforeOps.push_back(Op);
4177 else if (i > Index-NumDefs)
4178 AfterOps.push_back(Op);
4180 SDValue Chain = N->getOperand(NumOps-1);
4181 AddrOps.push_back(Chain);
4183 // Emit the load instruction.
4186 EVT VT = *RC->vt_begin();
4187 std::pair<MachineInstr::mmo_iterator,
4188 MachineInstr::mmo_iterator> MMOs =
4189 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4190 cast<MachineSDNode>(N)->memoperands_end());
4191 if (!(*MMOs.first) &&
4192 RC == &X86::VR128RegClass &&
4193 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4194 // Do not introduce a slow unaligned load.
4196 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4197 bool isAligned = (*MMOs.first) &&
4198 (*MMOs.first)->getAlignment() >= Alignment;
4199 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4200 VT, MVT::Other, &AddrOps[0], AddrOps.size());
4201 NewNodes.push_back(Load);
4203 // Preserve memory reference information.
4204 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4207 // Emit the data processing instruction.
4208 std::vector<EVT> VTs;
4209 const TargetRegisterClass *DstRC = 0;
4210 if (MCID.getNumDefs() > 0) {
4211 DstRC = getRegClass(MCID, 0, &RI, MF);
4212 VTs.push_back(*DstRC->vt_begin());
4214 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
4215 EVT VT = N->getValueType(i);
4216 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
4220 BeforeOps.push_back(SDValue(Load, 0));
4221 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
4222 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
4224 NewNodes.push_back(NewNode);
4226 // Emit the store instruction.
4229 AddrOps.push_back(SDValue(NewNode, 0));
4230 AddrOps.push_back(Chain);
4231 std::pair<MachineInstr::mmo_iterator,
4232 MachineInstr::mmo_iterator> MMOs =
4233 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4234 cast<MachineSDNode>(N)->memoperands_end());
4235 if (!(*MMOs.first) &&
4236 RC == &X86::VR128RegClass &&
4237 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4238 // Do not introduce a slow unaligned store.
4240 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4241 bool isAligned = (*MMOs.first) &&
4242 (*MMOs.first)->getAlignment() >= Alignment;
4243 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4246 &AddrOps[0], AddrOps.size());
4247 NewNodes.push_back(Store);
4249 // Preserve memory reference information.
4250 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4256 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
4257 bool UnfoldLoad, bool UnfoldStore,
4258 unsigned *LoadRegIndex) const {
4259 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4260 MemOp2RegOpTable.find(Opc);
4261 if (I == MemOp2RegOpTable.end())
4263 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4264 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4265 if (UnfoldLoad && !FoldedLoad)
4267 if (UnfoldStore && !FoldedStore)
4270 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
4271 return I->second.first;
4275 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4276 int64_t &Offset1, int64_t &Offset2) const {
4277 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4279 unsigned Opc1 = Load1->getMachineOpcode();
4280 unsigned Opc2 = Load2->getMachineOpcode();
4282 default: return false;
4292 case X86::MMX_MOVD64rm:
4293 case X86::MMX_MOVQ64rm:
4294 case X86::FsMOVAPSrm:
4295 case X86::FsMOVAPDrm:
4301 // AVX load instructions
4304 case X86::FsVMOVAPSrm:
4305 case X86::FsVMOVAPDrm:
4306 case X86::VMOVAPSrm:
4307 case X86::VMOVUPSrm:
4308 case X86::VMOVAPDrm:
4309 case X86::VMOVDQArm:
4310 case X86::VMOVDQUrm:
4311 case X86::VMOVAPSYrm:
4312 case X86::VMOVUPSYrm:
4313 case X86::VMOVAPDYrm:
4314 case X86::VMOVDQAYrm:
4315 case X86::VMOVDQUYrm:
4319 default: return false;
4329 case X86::MMX_MOVD64rm:
4330 case X86::MMX_MOVQ64rm:
4331 case X86::FsMOVAPSrm:
4332 case X86::FsMOVAPDrm:
4338 // AVX load instructions
4341 case X86::FsVMOVAPSrm:
4342 case X86::FsVMOVAPDrm:
4343 case X86::VMOVAPSrm:
4344 case X86::VMOVUPSrm:
4345 case X86::VMOVAPDrm:
4346 case X86::VMOVDQArm:
4347 case X86::VMOVDQUrm:
4348 case X86::VMOVAPSYrm:
4349 case X86::VMOVUPSYrm:
4350 case X86::VMOVAPDYrm:
4351 case X86::VMOVDQAYrm:
4352 case X86::VMOVDQUYrm:
4356 // Check if chain operands and base addresses match.
4357 if (Load1->getOperand(0) != Load2->getOperand(0) ||
4358 Load1->getOperand(5) != Load2->getOperand(5))
4360 // Segment operands should match as well.
4361 if (Load1->getOperand(4) != Load2->getOperand(4))
4363 // Scale should be 1, Index should be Reg0.
4364 if (Load1->getOperand(1) == Load2->getOperand(1) &&
4365 Load1->getOperand(2) == Load2->getOperand(2)) {
4366 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4369 // Now let's examine the displacements.
4370 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4371 isa<ConstantSDNode>(Load2->getOperand(3))) {
4372 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4373 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4380 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4381 int64_t Offset1, int64_t Offset2,
4382 unsigned NumLoads) const {
4383 assert(Offset2 > Offset1);
4384 if ((Offset2 - Offset1) / 8 > 64)
4387 unsigned Opc1 = Load1->getMachineOpcode();
4388 unsigned Opc2 = Load2->getMachineOpcode();
4390 return false; // FIXME: overly conservative?
4397 case X86::MMX_MOVD64rm:
4398 case X86::MMX_MOVQ64rm:
4402 EVT VT = Load1->getValueType(0);
4403 switch (VT.getSimpleVT().SimpleTy) {
4405 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4406 // have 16 of them to play with.
4407 if (TM.getSubtargetImpl()->is64Bit()) {
4410 } else if (NumLoads) {
4430 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
4431 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
4432 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
4433 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4435 Cond[0].setImm(GetOppositeBranchCondition(CC));
4440 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4441 // FIXME: Return false for x87 stack register classes for now. We can't
4442 // allow any loads of these registers before FpGet_ST0_80.
4443 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4444 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
4447 /// getGlobalBaseReg - Return a virtual register initialized with the
4448 /// the global base register value. Output instructions required to
4449 /// initialize the register in the function entry block, if necessary.
4451 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4453 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4454 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4455 "X86-64 PIC uses RIP relative addressing");
4457 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4458 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4459 if (GlobalBaseReg != 0)
4460 return GlobalBaseReg;
4462 // Create the register. The code to initialize it is inserted
4463 // later, by the CGBR pass (below).
4464 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4465 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
4466 X86FI->setGlobalBaseReg(GlobalBaseReg);
4467 return GlobalBaseReg;
4470 // These are the replaceable SSE instructions. Some of these have Int variants
4471 // that we don't include here. We don't want to replace instructions selected
4473 static const uint16_t ReplaceableInstrs[][3] = {
4474 //PackedSingle PackedDouble PackedInt
4475 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
4476 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
4477 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
4478 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
4479 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
4480 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
4481 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
4482 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
4483 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
4484 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
4485 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
4486 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
4487 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
4488 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
4489 // AVX 128-bit support
4490 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
4491 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
4492 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
4493 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
4494 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
4495 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
4496 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
4497 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
4498 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
4499 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
4500 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
4501 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
4502 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
4503 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
4504 // AVX 256-bit support
4505 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
4506 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
4507 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
4508 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
4509 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
4510 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
4513 static const uint16_t ReplaceableInstrsAVX2[][3] = {
4514 //PackedSingle PackedDouble PackedInt
4515 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
4516 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
4517 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
4518 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
4519 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
4520 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
4521 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
4522 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
4523 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
4524 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
4525 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
4526 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
4527 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
4528 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
4531 // FIXME: Some shuffle and unpack instructions have equivalents in different
4532 // domains, but they require a bit more work than just switching opcodes.
4534 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
4535 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
4536 if (ReplaceableInstrs[i][domain-1] == opcode)
4537 return ReplaceableInstrs[i];
4541 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
4542 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
4543 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
4544 return ReplaceableInstrsAVX2[i];
4548 std::pair<uint16_t, uint16_t>
4549 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
4550 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4551 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
4552 uint16_t validDomains = 0;
4553 if (domain && lookup(MI->getOpcode(), domain))
4555 else if (domain && lookupAVX2(MI->getOpcode(), domain))
4556 validDomains = hasAVX2 ? 0xe : 0x6;
4557 return std::make_pair(domain, validDomains);
4560 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
4561 assert(Domain>0 && Domain<4 && "Invalid execution domain");
4562 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
4563 assert(dom && "Not an SSE instruction");
4564 const uint16_t *table = lookup(MI->getOpcode(), dom);
4565 if (!table) { // try the other table
4566 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
4567 "256-bit vector operations only available in AVX2");
4568 table = lookupAVX2(MI->getOpcode(), dom);
4570 assert(table && "Cannot change domain");
4571 MI->setDesc(get(table[Domain-1]));
4574 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
4575 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
4576 NopInst.setOpcode(X86::NOOP);
4579 bool X86InstrInfo::isHighLatencyDef(int opc) const {
4581 default: return false;
4583 case X86::DIVSDrm_Int:
4585 case X86::DIVSDrr_Int:
4587 case X86::DIVSSrm_Int:
4589 case X86::DIVSSrr_Int:
4591 case X86::SQRTPDm_Int:
4593 case X86::SQRTPDr_Int:
4595 case X86::SQRTPSm_Int:
4597 case X86::SQRTPSr_Int:
4599 case X86::SQRTSDm_Int:
4601 case X86::SQRTSDr_Int:
4603 case X86::SQRTSSm_Int:
4605 case X86::SQRTSSr_Int:
4606 // AVX instructions with high latency
4608 case X86::VDIVSDrm_Int:
4610 case X86::VDIVSDrr_Int:
4612 case X86::VDIVSSrm_Int:
4614 case X86::VDIVSSrr_Int:
4616 case X86::VSQRTPDm_Int:
4618 case X86::VSQRTPDr_Int:
4620 case X86::VSQRTPSm_Int:
4622 case X86::VSQRTPSr_Int:
4624 case X86::VSQRTSDm_Int:
4627 case X86::VSQRTSSm_Int:
4634 hasHighOperandLatency(const InstrItineraryData *ItinData,
4635 const MachineRegisterInfo *MRI,
4636 const MachineInstr *DefMI, unsigned DefIdx,
4637 const MachineInstr *UseMI, unsigned UseIdx) const {
4638 return isHighLatencyDef(DefMI->getOpcode());
4642 /// CGBR - Create Global Base Reg pass. This initializes the PIC
4643 /// global base register for x86-32.
4644 struct CGBR : public MachineFunctionPass {
4646 CGBR() : MachineFunctionPass(ID) {}
4648 virtual bool runOnMachineFunction(MachineFunction &MF) {
4649 const X86TargetMachine *TM =
4650 static_cast<const X86TargetMachine *>(&MF.getTarget());
4652 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
4653 "X86-64 PIC uses RIP relative addressing");
4655 // Only emit a global base reg in PIC mode.
4656 if (TM->getRelocationModel() != Reloc::PIC_)
4659 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
4660 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4662 // If we didn't need a GlobalBaseReg, don't insert code.
4663 if (GlobalBaseReg == 0)
4666 // Insert the set of GlobalBaseReg into the first MBB of the function
4667 MachineBasicBlock &FirstMBB = MF.front();
4668 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
4669 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
4670 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4671 const X86InstrInfo *TII = TM->getInstrInfo();
4674 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
4675 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
4679 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4680 // only used in JIT code emission as displacement to pc.
4681 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
4683 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4684 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4685 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
4686 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4687 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4688 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4689 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4695 virtual const char *getPassName() const {
4696 return "X86 PIC Global Base Reg Initialization";
4699 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4700 AU.setPreservesCFG();
4701 MachineFunctionPass::getAnalysisUsage(AU);
4708 llvm::createGlobalBaseRegPass() { return new CGBR(); }
4711 struct LDTLSCleanup : public MachineFunctionPass {
4713 LDTLSCleanup() : MachineFunctionPass(ID) {}
4715 virtual bool runOnMachineFunction(MachineFunction &MF) {
4716 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4717 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4718 // No point folding accesses if there isn't at least two.
4722 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4723 return VisitNode(DT->getRootNode(), 0);
4726 // Visit the dominator subtree rooted at Node in pre-order.
4727 // If TLSBaseAddrReg is non-null, then use that to replace any
4728 // TLS_base_addr instructions. Otherwise, create the register
4729 // when the first such instruction is seen, and then use it
4730 // as we encounter more instructions.
4731 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4732 MachineBasicBlock *BB = Node->getBlock();
4733 bool Changed = false;
4735 // Traverse the current block.
4736 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4738 switch (I->getOpcode()) {
4739 case X86::TLS_base_addr32:
4740 case X86::TLS_base_addr64:
4742 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4744 I = SetRegister(I, &TLSBaseAddrReg);
4752 // Visit the children of this block in the dominator tree.
4753 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4755 Changed |= VisitNode(*I, TLSBaseAddrReg);
4761 // Replace the TLS_base_addr instruction I with a copy from
4762 // TLSBaseAddrReg, returning the new instruction.
4763 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4764 unsigned TLSBaseAddrReg) {
4765 MachineFunction *MF = I->getParent()->getParent();
4766 const X86TargetMachine *TM =
4767 static_cast<const X86TargetMachine *>(&MF->getTarget());
4768 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4769 const X86InstrInfo *TII = TM->getInstrInfo();
4771 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4772 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4773 TII->get(TargetOpcode::COPY),
4774 is64Bit ? X86::RAX : X86::EAX)
4775 .addReg(TLSBaseAddrReg);
4777 // Erase the TLS_base_addr instruction.
4778 I->eraseFromParent();
4783 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4784 // inserting a copy instruction after I. Returns the new instruction.
4785 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4786 MachineFunction *MF = I->getParent()->getParent();
4787 const X86TargetMachine *TM =
4788 static_cast<const X86TargetMachine *>(&MF->getTarget());
4789 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4790 const X86InstrInfo *TII = TM->getInstrInfo();
4792 // Create a virtual register for the TLS base address.
4793 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4794 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4795 ? &X86::GR64RegClass
4796 : &X86::GR32RegClass);
4798 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4799 MachineInstr *Next = I->getNextNode();
4800 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4801 TII->get(TargetOpcode::COPY),
4803 .addReg(is64Bit ? X86::RAX : X86::EAX);
4808 virtual const char *getPassName() const {
4809 return "Local Dynamic TLS Access Clean-up";
4812 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4813 AU.setPreservesCFG();
4814 AU.addRequired<MachineDominatorTree>();
4815 MachineFunctionPass::getAnalysisUsage(AU);
4820 char LDTLSCleanup::ID = 0;
4822 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }