1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the X86 implementation of the TargetInstrInfo class.
11 //
12 //===----------------------------------------------------------------------===//
14 #include "X86InstrInfo.h"
15 #include "X86.h"
16 #include "X86InstrBuilder.h"
17 #include "X86MachineFunctionInfo.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/CodeGen/LiveVariables.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineDominators.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/StackMaps.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/MC/MCAsmInfo.h"
32 #include "llvm/MC/MCExpr.h"
33 #include "llvm/MC/MCInst.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetOptions.h"
39 #include <limits>
41 using namespace llvm;
43 #define DEBUG_TYPE "x86-instr-info"
45 #define GET_INSTRINFO_CTOR_DTOR
46 #include "X86GenInstrInfo.inc"
48 static cl::opt<bool>
49 NoFusing("disable-spill-fusing",
50 cl::desc("Disable fusing of spill code into instructions"));
51 static cl::opt<bool>
52 PrintFailedFusing("print-failed-fuse-candidates",
53 cl::desc("Print instructions that the allocator wants to"
54 " fuse, but the X86 backend currently can't"),
55 cl::Hidden);
56 static cl::opt<bool>
57 ReMatPICStubLoad("remat-pic-stub-load",
58 cl::desc("Re-materialize load from stub in PIC mode"),
59 cl::init(false), cl::Hidden);
61 enum {
62 // Select which memory operand is being unfolded.
63 // (stored in bits 0 - 3)
64 TB_INDEX_0 = 0,
65 TB_INDEX_1 = 1,
66 TB_INDEX_2 = 2,
67 TB_INDEX_3 = 3,
68 TB_INDEX_MASK = 0xf,
70 // Do not insert the reverse map (MemOp -> RegOp) into the table.
71 // This may be needed because there is a many -> one mapping.
72 TB_NO_REVERSE = 1 << 4,
74 // Do not insert the forward map (RegOp -> MemOp) into the table.
75 // This is needed for Native Client, which prohibits branch
76 // instructions from using a memory operand.
77 TB_NO_FORWARD = 1 << 5,
79 TB_FOLDED_LOAD = 1 << 6,
80 TB_FOLDED_STORE = 1 << 7,
82 // Minimum alignment required for load/store.
83 // Used for RegOp->MemOp conversion.
84 // (stored in bits 8 - 15)
85 TB_ALIGN_SHIFT = 8,
86 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
87 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
88 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
89 TB_ALIGN_64 = 64 << TB_ALIGN_SHIFT,
90 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT
91 };
93 struct X86OpTblEntry {
94 uint16_t RegOp;
95 uint16_t MemOp;
96 uint16_t Flags;
97 };
99 // Pin the vtable to this file.
100 void X86InstrInfo::anchor() {}
102 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
103 : X86GenInstrInfo(
104 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64 : X86::ADJCALLSTACKDOWN32),
105 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64 : X86::ADJCALLSTACKUP32)),
106 Subtarget(STI), RI(STI) {
108 static const X86OpTblEntry OpTbl2Addr[] = {
109 { X86::ADC32ri, X86::ADC32mi, 0 },
110 { X86::ADC32ri8, X86::ADC32mi8, 0 },
111 { X86::ADC32rr, X86::ADC32mr, 0 },
112 { X86::ADC64ri32, X86::ADC64mi32, 0 },
113 { X86::ADC64ri8, X86::ADC64mi8, 0 },
114 { X86::ADC64rr, X86::ADC64mr, 0 },
115 { X86::ADD16ri, X86::ADD16mi, 0 },
116 { X86::ADD16ri8, X86::ADD16mi8, 0 },
117 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
118 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
119 { X86::ADD16rr, X86::ADD16mr, 0 },
120 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
121 { X86::ADD32ri, X86::ADD32mi, 0 },
122 { X86::ADD32ri8, X86::ADD32mi8, 0 },
123 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
124 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
125 { X86::ADD32rr, X86::ADD32mr, 0 },
126 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
127 { X86::ADD64ri32, X86::ADD64mi32, 0 },
128 { X86::ADD64ri8, X86::ADD64mi8, 0 },
129 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
130 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
131 { X86::ADD64rr, X86::ADD64mr, 0 },
132 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
133 { X86::ADD8ri, X86::ADD8mi, 0 },
134 { X86::ADD8rr, X86::ADD8mr, 0 },
135 { X86::AND16ri, X86::AND16mi, 0 },
136 { X86::AND16ri8, X86::AND16mi8, 0 },
137 { X86::AND16rr, X86::AND16mr, 0 },
138 { X86::AND32ri, X86::AND32mi, 0 },
139 { X86::AND32ri8, X86::AND32mi8, 0 },
140 { X86::AND32rr, X86::AND32mr, 0 },
141 { X86::AND64ri32, X86::AND64mi32, 0 },
142 { X86::AND64ri8, X86::AND64mi8, 0 },
143 { X86::AND64rr, X86::AND64mr, 0 },
144 { X86::AND8ri, X86::AND8mi, 0 },
145 { X86::AND8rr, X86::AND8mr, 0 },
146 { X86::DEC16r, X86::DEC16m, 0 },
147 { X86::DEC32r, X86::DEC32m, 0 },
148 { X86::DEC64_16r, X86::DEC64_16m, 0 },
149 { X86::DEC64_32r, X86::DEC64_32m, 0 },
150 { X86::DEC64r, X86::DEC64m, 0 },
151 { X86::DEC8r, X86::DEC8m, 0 },
152 { X86::INC16r, X86::INC16m, 0 },
153 { X86::INC32r, X86::INC32m, 0 },
154 { X86::INC64_16r, X86::INC64_16m, 0 },
155 { X86::INC64_32r, X86::INC64_32m, 0 },
156 { X86::INC64r, X86::INC64m, 0 },
157 { X86::INC8r, X86::INC8m, 0 },
158 { X86::NEG16r, X86::NEG16m, 0 },
159 { X86::NEG32r, X86::NEG32m, 0 },
160 { X86::NEG64r, X86::NEG64m, 0 },
161 { X86::NEG8r, X86::NEG8m, 0 },
162 { X86::NOT16r, X86::NOT16m, 0 },
163 { X86::NOT32r, X86::NOT32m, 0 },
164 { X86::NOT64r, X86::NOT64m, 0 },
165 { X86::NOT8r, X86::NOT8m, 0 },
166 { X86::OR16ri, X86::OR16mi, 0 },
167 { X86::OR16ri8, X86::OR16mi8, 0 },
168 { X86::OR16rr, X86::OR16mr, 0 },
169 { X86::OR32ri, X86::OR32mi, 0 },
170 { X86::OR32ri8, X86::OR32mi8, 0 },
171 { X86::OR32rr, X86::OR32mr, 0 },
172 { X86::OR64ri32, X86::OR64mi32, 0 },
173 { X86::OR64ri8, X86::OR64mi8, 0 },
174 { X86::OR64rr, X86::OR64mr, 0 },
175 { X86::OR8ri, X86::OR8mi, 0 },
176 { X86::OR8rr, X86::OR8mr, 0 },
177 { X86::ROL16r1, X86::ROL16m1, 0 },
178 { X86::ROL16rCL, X86::ROL16mCL, 0 },
179 { X86::ROL16ri, X86::ROL16mi, 0 },
180 { X86::ROL32r1, X86::ROL32m1, 0 },
181 { X86::ROL32rCL, X86::ROL32mCL, 0 },
182 { X86::ROL32ri, X86::ROL32mi, 0 },
183 { X86::ROL64r1, X86::ROL64m1, 0 },
184 { X86::ROL64rCL, X86::ROL64mCL, 0 },
185 { X86::ROL64ri, X86::ROL64mi, 0 },
186 { X86::ROL8r1, X86::ROL8m1, 0 },
187 { X86::ROL8rCL, X86::ROL8mCL, 0 },
188 { X86::ROL8ri, X86::ROL8mi, 0 },
189 { X86::ROR16r1, X86::ROR16m1, 0 },
190 { X86::ROR16rCL, X86::ROR16mCL, 0 },
191 { X86::ROR16ri, X86::ROR16mi, 0 },
192 { X86::ROR32r1, X86::ROR32m1, 0 },
193 { X86::ROR32rCL, X86::ROR32mCL, 0 },
194 { X86::ROR32ri, X86::ROR32mi, 0 },
195 { X86::ROR64r1, X86::ROR64m1, 0 },
196 { X86::ROR64rCL, X86::ROR64mCL, 0 },
197 { X86::ROR64ri, X86::ROR64mi, 0 },
198 { X86::ROR8r1, X86::ROR8m1, 0 },
199 { X86::ROR8rCL, X86::ROR8mCL, 0 },
200 { X86::ROR8ri, X86::ROR8mi, 0 },
201 { X86::SAR16r1, X86::SAR16m1, 0 },
202 { X86::SAR16rCL, X86::SAR16mCL, 0 },
203 { X86::SAR16ri, X86::SAR16mi, 0 },
204 { X86::SAR32r1, X86::SAR32m1, 0 },
205 { X86::SAR32rCL, X86::SAR32mCL, 0 },
206 { X86::SAR32ri, X86::SAR32mi, 0 },
207 { X86::SAR64r1, X86::SAR64m1, 0 },
208 { X86::SAR64rCL, X86::SAR64mCL, 0 },
209 { X86::SAR64ri, X86::SAR64mi, 0 },
210 { X86::SAR8r1, X86::SAR8m1, 0 },
211 { X86::SAR8rCL, X86::SAR8mCL, 0 },
212 { X86::SAR8ri, X86::SAR8mi, 0 },
213 { X86::SBB32ri, X86::SBB32mi, 0 },
214 { X86::SBB32ri8, X86::SBB32mi8, 0 },
215 { X86::SBB32rr, X86::SBB32mr, 0 },
216 { X86::SBB64ri32, X86::SBB64mi32, 0 },
217 { X86::SBB64ri8, X86::SBB64mi8, 0 },
218 { X86::SBB64rr, X86::SBB64mr, 0 },
219 { X86::SHL16rCL, X86::SHL16mCL, 0 },
220 { X86::SHL16ri, X86::SHL16mi, 0 },
221 { X86::SHL32rCL, X86::SHL32mCL, 0 },
222 { X86::SHL32ri, X86::SHL32mi, 0 },
223 { X86::SHL64rCL, X86::SHL64mCL, 0 },
224 { X86::SHL64ri, X86::SHL64mi, 0 },
225 { X86::SHL8rCL, X86::SHL8mCL, 0 },
226 { X86::SHL8ri, X86::SHL8mi, 0 },
227 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
228 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
229 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
230 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
231 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
232 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
233 { X86::SHR16r1, X86::SHR16m1, 0 },
234 { X86::SHR16rCL, X86::SHR16mCL, 0 },
235 { X86::SHR16ri, X86::SHR16mi, 0 },
236 { X86::SHR32r1, X86::SHR32m1, 0 },
237 { X86::SHR32rCL, X86::SHR32mCL, 0 },
238 { X86::SHR32ri, X86::SHR32mi, 0 },
239 { X86::SHR64r1, X86::SHR64m1, 0 },
240 { X86::SHR64rCL, X86::SHR64mCL, 0 },
241 { X86::SHR64ri, X86::SHR64mi, 0 },
242 { X86::SHR8r1, X86::SHR8m1, 0 },
243 { X86::SHR8rCL, X86::SHR8mCL, 0 },
244 { X86::SHR8ri, X86::SHR8mi, 0 },
245 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
246 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
247 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
248 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
249 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
250 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
251 { X86::SUB16ri, X86::SUB16mi, 0 },
252 { X86::SUB16ri8, X86::SUB16mi8, 0 },
253 { X86::SUB16rr, X86::SUB16mr, 0 },
254 { X86::SUB32ri, X86::SUB32mi, 0 },
255 { X86::SUB32ri8, X86::SUB32mi8, 0 },
256 { X86::SUB32rr, X86::SUB32mr, 0 },
257 { X86::SUB64ri32, X86::SUB64mi32, 0 },
258 { X86::SUB64ri8, X86::SUB64mi8, 0 },
259 { X86::SUB64rr, X86::SUB64mr, 0 },
260 { X86::SUB8ri, X86::SUB8mi, 0 },
261 { X86::SUB8rr, X86::SUB8mr, 0 },
262 { X86::XOR16ri, X86::XOR16mi, 0 },
263 { X86::XOR16ri8, X86::XOR16mi8, 0 },
264 { X86::XOR16rr, X86::XOR16mr, 0 },
265 { X86::XOR32ri, X86::XOR32mi, 0 },
266 { X86::XOR32ri8, X86::XOR32mi8, 0 },
267 { X86::XOR32rr, X86::XOR32mr, 0 },
268 { X86::XOR64ri32, X86::XOR64mi32, 0 },
269 { X86::XOR64ri8, X86::XOR64mi8, 0 },
270 { X86::XOR64rr, X86::XOR64mr, 0 },
271 { X86::XOR8ri, X86::XOR8mi, 0 },
272 { X86::XOR8rr, X86::XOR8mr, 0 }
273 };
275 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
276 unsigned RegOp = OpTbl2Addr[i].RegOp;
277 unsigned MemOp = OpTbl2Addr[i].MemOp;
278 unsigned Flags = OpTbl2Addr[i].Flags;
279 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
280 RegOp, MemOp,
281 // Index 0, folded load and store, no alignment requirement.
282 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
283 }
285 static const X86OpTblEntry OpTbl0[] = {
286 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
287 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
288 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
289 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
290 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
291 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
292 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
293 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
294 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
295 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
296 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
297 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
298 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
299 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
300 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
301 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
302 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
303 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
304 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
305 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
306 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE },
307 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
308 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
309 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
310 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
311 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
312 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
313 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
314 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
315 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
316 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
317 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
318 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
319 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
320 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
321 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
322 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
323 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
324 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
325 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
326 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
327 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
328 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
329 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
330 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
331 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
332 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
333 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
334 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
335 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
336 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
337 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
338 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
339 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
340 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
341 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
342 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
343 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
344 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
345 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
346 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
347 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
348 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
349 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
350 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
351 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
352 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
353 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
354 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
355 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
356 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
357 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
358 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
359 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
360 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
361 // AVX 128-bit versions of foldable instructions
362 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE },
363 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
366 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
367 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
368 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
369 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
370 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
371 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
372 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
373 // AVX 256-bit foldable instructions
374 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
375 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
377 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
378 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
379 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE },
380 // AVX-512 foldable instructions
381 { X86::VMOVPDI2DIZrr, X86::VMOVPDI2DIZmr, TB_FOLDED_STORE },
382 { X86::VMOVAPDZrr, X86::VMOVAPDZmr, TB_FOLDED_STORE | TB_ALIGN_64 },
383 { X86::VMOVAPSZrr, X86::VMOVAPSZmr, TB_FOLDED_STORE | TB_ALIGN_64 },
384 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zmr, TB_FOLDED_STORE | TB_ALIGN_64 },
385 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zmr, TB_FOLDED_STORE | TB_ALIGN_64 },
386 { X86::VMOVUPDZrr, X86::VMOVUPDZmr, TB_FOLDED_STORE },
387 { X86::VMOVUPSZrr, X86::VMOVUPSZmr, TB_FOLDED_STORE },
388 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zmr, TB_FOLDED_STORE },
389 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zmr, TB_FOLDED_STORE },
390 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zmr, TB_FOLDED_STORE },
391 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zmr, TB_FOLDED_STORE },
392 // AVX-512 foldable instructions (256-bit versions)
393 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
394 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
395 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
396 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 },
397 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256mr, TB_FOLDED_STORE },
398 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256mr, TB_FOLDED_STORE },
399 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256mr, TB_FOLDED_STORE },
400 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256mr, TB_FOLDED_STORE },
401 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256mr, TB_FOLDED_STORE },
402 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256mr, TB_FOLDED_STORE },
403 // AVX-512 foldable instructions (128-bit versions)
404 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
405 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
406 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
407 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
408 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128mr, TB_FOLDED_STORE },
409 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128mr, TB_FOLDED_STORE },
410 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128mr, TB_FOLDED_STORE },
411 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128mr, TB_FOLDED_STORE },
412 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128mr, TB_FOLDED_STORE },
413 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128mr, TB_FOLDED_STORE }
414 };
416 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
417 unsigned RegOp = OpTbl0[i].RegOp;
418 unsigned MemOp = OpTbl0[i].MemOp;
419 unsigned Flags = OpTbl0[i].Flags;
420 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
421 RegOp, MemOp, TB_INDEX_0 | Flags);
422 }
424 static const X86OpTblEntry OpTbl1[] = {
425 { X86::CMP16rr, X86::CMP16rm, 0 },
426 { X86::CMP32rr, X86::CMP32rm, 0 },
427 { X86::CMP64rr, X86::CMP64rm, 0 },
428 { X86::CMP8rr, X86::CMP8rm, 0 },
429 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
430 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
431 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
432 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
433 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
434 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
435 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
436 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
437 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
438 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
439 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
440 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
441 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
442 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
443 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
444 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
445 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
446 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
447 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
448 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
449 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
450 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
451 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
452 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
453 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
454 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
455 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
456 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
457 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
458 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
459 { X86::MOV16rr, X86::MOV16rm, 0 },
460 { X86::MOV32rr, X86::MOV32rm, 0 },
461 { X86::MOV64rr, X86::MOV64rm, 0 },
462 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
463 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
464 { X86::MOV8rr, X86::MOV8rm, 0 },
465 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
466 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
467 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
468 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
469 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
470 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
471 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
472 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
473 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
474 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
475 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
476 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
477 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
478 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
479 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
480 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
481 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
482 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
483 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
484 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
485 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
486 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
487 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
488 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
489 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
490 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
491 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
492 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
493 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
494 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
495 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
496 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
497 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
498 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
499 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
500 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
501 { X86::SQRTSDr, X86::SQRTSDm, 0 },
502 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
503 { X86::SQRTSSr, X86::SQRTSSm, 0 },
504 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
505 { X86::TEST16rr, X86::TEST16rm, 0 },
506 { X86::TEST32rr, X86::TEST32rm, 0 },
507 { X86::TEST64rr, X86::TEST64rm, 0 },
508 { X86::TEST8rr, X86::TEST8rm, 0 },
509 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
510 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
511 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
512 // AVX 128-bit versions of foldable instructions
513 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
514 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
515 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
516 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
517 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
518 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
519 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
520 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
521 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
522 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
523 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
524 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
525 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
526 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
527 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
528 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
529 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
530 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
531 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
532 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
533 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
534 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
535 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
536 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
537 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
538 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
539 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 },
540 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
541 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
542 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
543 { X86::VPABSBrr128, X86::VPABSBrm128, 0 },
544 { X86::VPABSDrr128, X86::VPABSDrm128, 0 },
545 { X86::VPABSWrr128, X86::VPABSWrm128, 0 },
546 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 },
547 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 },
548 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 },
549 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 },
550 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 },
551 { X86::VRCPPSr, X86::VRCPPSm, 0 },
552 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, 0 },
553 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 },
554 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, 0 },
555 { X86::VSQRTPDr, X86::VSQRTPDm, 0 },
556 { X86::VSQRTPSr, X86::VSQRTPSm, 0 },
557 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
558 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
559 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE },
561 // AVX 256-bit foldable instructions
562 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
563 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
564 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
565 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
566 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
567 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 },
568 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 },
570 // AVX2 foldable instructions
571 { X86::VPABSBrr256, X86::VPABSBrm256, 0 },
572 { X86::VPABSDrr256, X86::VPABSDrm256, 0 },
573 { X86::VPABSWrr256, X86::VPABSWrm256, 0 },
574 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 },
575 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 },
576 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 },
577 { X86::VRCPPSYr, X86::VRCPPSYm, 0 },
578 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, 0 },
579 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 },
580 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 },
581 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 },
582 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE },
583 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE },
585 // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
586 { X86::BEXTR32rr, X86::BEXTR32rm, 0 },
587 { X86::BEXTR64rr, X86::BEXTR64rm, 0 },
588 { X86::BEXTRI32ri, X86::BEXTRI32mi, 0 },
589 { X86::BEXTRI64ri, X86::BEXTRI64mi, 0 },
590 { X86::BLCFILL32rr, X86::BLCFILL32rm, 0 },
591 { X86::BLCFILL64rr, X86::BLCFILL64rm, 0 },
592 { X86::BLCI32rr, X86::BLCI32rm, 0 },
593 { X86::BLCI64rr, X86::BLCI64rm, 0 },
594 { X86::BLCIC32rr, X86::BLCIC32rm, 0 },
595 { X86::BLCIC64rr, X86::BLCIC64rm, 0 },
596 { X86::BLCMSK32rr, X86::BLCMSK32rm, 0 },
597 { X86::BLCMSK64rr, X86::BLCMSK64rm, 0 },
598 { X86::BLCS32rr, X86::BLCS32rm, 0 },
599 { X86::BLCS64rr, X86::BLCS64rm, 0 },
600 { X86::BLSFILL32rr, X86::BLSFILL32rm, 0 },
601 { X86::BLSFILL64rr, X86::BLSFILL64rm, 0 },
602 { X86::BLSI32rr, X86::BLSI32rm, 0 },
603 { X86::BLSI64rr, X86::BLSI64rm, 0 },
604 { X86::BLSIC32rr, X86::BLSIC32rm, 0 },
605 { X86::BLSIC64rr, X86::BLSIC64rm, 0 },
606 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 },
607 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 },
608 { X86::BLSR32rr, X86::BLSR32rm, 0 },
609 { X86::BLSR64rr, X86::BLSR64rm, 0 },
610 { X86::BZHI32rr, X86::BZHI32rm, 0 },
611 { X86::BZHI64rr, X86::BZHI64rm, 0 },
612 { X86::LZCNT16rr, X86::LZCNT16rm, 0 },
613 { X86::LZCNT32rr, X86::LZCNT32rm, 0 },
614 { X86::LZCNT64rr, X86::LZCNT64rm, 0 },
615 { X86::POPCNT16rr, X86::POPCNT16rm, 0 },
616 { X86::POPCNT32rr, X86::POPCNT32rm, 0 },
617 { X86::POPCNT64rr, X86::POPCNT64rm, 0 },
618 { X86::RORX32ri, X86::RORX32mi, 0 },
619 { X86::RORX64ri, X86::RORX64mi, 0 },
620 { X86::SARX32rr, X86::SARX32rm, 0 },
621 { X86::SARX64rr, X86::SARX64rm, 0 },
622 { X86::SHRX32rr, X86::SHRX32rm, 0 },
623 { X86::SHRX64rr, X86::SHRX64rm, 0 },
624 { X86::SHLX32rr, X86::SHLX32rm, 0 },
625 { X86::SHLX64rr, X86::SHLX64rm, 0 },
626 { X86::T1MSKC32rr, X86::T1MSKC32rm, 0 },
627 { X86::T1MSKC64rr, X86::T1MSKC64rm, 0 },
628 { X86::TZCNT16rr, X86::TZCNT16rm, 0 },
629 { X86::TZCNT32rr, X86::TZCNT32rm, 0 },
630 { X86::TZCNT64rr, X86::TZCNT64rm, 0 },
631 { X86::TZMSK32rr, X86::TZMSK32rm, 0 },
632 { X86::TZMSK64rr, X86::TZMSK64rm, 0 },
634 // AVX-512 foldable instructions
635 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 },
636 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 },
637 { X86::VMOVAPDZrr, X86::VMOVAPDZrm, TB_ALIGN_64 },
638 { X86::VMOVAPSZrr, X86::VMOVAPSZrm, TB_ALIGN_64 },
639 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zrm, TB_ALIGN_64 },
640 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zrm, TB_ALIGN_64 },
641 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zrm, 0 },
642 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zrm, 0 },
643 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zrm, 0 },
644 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zrm, 0 },
645 { X86::VMOVUPDZrr, X86::VMOVUPDZrm, 0 },
646 { X86::VMOVUPSZrr, X86::VMOVUPSZrm, 0 },
647 { X86::VPABSDZrr, X86::VPABSDZrm, 0 },
648 { X86::VPABSQZrr, X86::VPABSQZrm, 0 },
649 // AVX-512 foldable instructions (256-bit versions)
650 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256rm, TB_ALIGN_32 },
651 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256rm, TB_ALIGN_32 },
652 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256rm, TB_ALIGN_32 },
653 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256rm, TB_ALIGN_32 },
654 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256rm, 0 },
655 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256rm, 0 },
656 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256rm, 0 },
657 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256rm, 0 },
658 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256rm, 0 },
659 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256rm, 0 },
660 // AVX-512 foldable instructions (256-bit versions)
661 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128rm, TB_ALIGN_16 },
662 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128rm, TB_ALIGN_16 },
663 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128rm, TB_ALIGN_16 },
664 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128rm, TB_ALIGN_16 },
665 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128rm, 0 },
666 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128rm, 0 },
667 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128rm, 0 },
668 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128rm, 0 },
669 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128rm, 0 },
670 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128rm, 0 },
672 // AES foldable instructions
673 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 },
674 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 },
675 { X86::VAESIMCrr, X86::VAESIMCrm, TB_ALIGN_16 },
676 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, TB_ALIGN_16 }
677 };
679 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
680 unsigned RegOp = OpTbl1[i].RegOp;
681 unsigned MemOp = OpTbl1[i].MemOp;
682 unsigned Flags = OpTbl1[i].Flags;
683 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
684 RegOp, MemOp,
685 // Index 1, folded load
686 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
687 }
689 static const X86OpTblEntry OpTbl2[] = {
690 { X86::ADC32rr, X86::ADC32rm, 0 },
691 { X86::ADC64rr, X86::ADC64rm, 0 },
692 { X86::ADD16rr, X86::ADD16rm, 0 },
693 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
694 { X86::ADD32rr, X86::ADD32rm, 0 },
695 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
696 { X86::ADD64rr, X86::ADD64rm, 0 },
697 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
698 { X86::ADD8rr, X86::ADD8rm, 0 },
699 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
700 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
701 { X86::ADDSDrr, X86::ADDSDrm, 0 },
702 { X86::ADDSSrr, X86::ADDSSrm, 0 },
703 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
704 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
705 { X86::AND16rr, X86::AND16rm, 0 },
706 { X86::AND32rr, X86::AND32rm, 0 },
707 { X86::AND64rr, X86::AND64rm, 0 },
708 { X86::AND8rr, X86::AND8rm, 0 },
709 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
710 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
711 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
712 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
713 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
714 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
715 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
716 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
717 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
718 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
719 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
720 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
721 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
722 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
723 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
724 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
725 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
726 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
727 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
728 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
729 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
730 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
731 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
732 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
733 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
734 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
735 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
736 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
737 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
738 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
739 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
740 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
741 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
742 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
743 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
744 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
745 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
746 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
747 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
748 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
749 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
750 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
751 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
752 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
753 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
754 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
755 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
756 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
757 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
758 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
759 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
760 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
761 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
762 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
763 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
764 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
765 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
766 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
767 { X86::CMPSDrr, X86::CMPSDrm, 0 },
768 { X86::CMPSSrr, X86::CMPSSrm, 0 },
769 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
770 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
771 { X86::DIVSDrr, X86::DIVSDrm, 0 },
772 { X86::DIVSSrr, X86::DIVSSrm, 0 },
773 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
774 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
775 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
776 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
777 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
778 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
779 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
780 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
781 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
782 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
783 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
784 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
785 { X86::IMUL16rr, X86::IMUL16rm, 0 },
786 { X86::IMUL32rr, X86::IMUL32rm, 0 },
787 { X86::IMUL64rr, X86::IMUL64rm, 0 },
788 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
789 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
790 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
791 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
792 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
793 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
794 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
795 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
796 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
797 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
798 { X86::MAXSDrr, X86::MAXSDrm, 0 },
799 { X86::MAXSSrr, X86::MAXSSrm, 0 },
800 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
801 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
802 { X86::MINSDrr, X86::MINSDrm, 0 },
803 { X86::MINSSrr, X86::MINSSrm, 0 },
804 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
805 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
806 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
807 { X86::MULSDrr, X86::MULSDrm, 0 },
808 { X86::MULSSrr, X86::MULSSrm, 0 },
809 { X86::OR16rr, X86::OR16rm, 0 },
810 { X86::OR32rr, X86::OR32rm, 0 },
811 { X86::OR64rr, X86::OR64rm, 0 },
812 { X86::OR8rr, X86::OR8rm, 0 },
813 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
814 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
815 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
816 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
817 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
818 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
819 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
820 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
821 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
822 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
823 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
824 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
825 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
826 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
827 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
828 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
829 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
830 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
831 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
832 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
833 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
834 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
835 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
836 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
837 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
838 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
839 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
840 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
841 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
842 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
843 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
844 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
845 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
846 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
847 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
848 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
849 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
850 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
851 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
852 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
853 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
854 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 },
855 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 },
856 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 },
857 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 },
858 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 },
859 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 },
860 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 },
861 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 },
862 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
863 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
864 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
865 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
866 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
867 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
868 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
869 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
870 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
871 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
872 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
873 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
874 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
875 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
876 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
877 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
878 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
879 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
880 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
881 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
882 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
883 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
884 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
885 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
886 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
887 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
888 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
889 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
890 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
891 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
892 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
893 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
894 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
895 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
896 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
897 { X86::SBB32rr, X86::SBB32rm, 0 },
898 { X86::SBB64rr, X86::SBB64rm, 0 },
899 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
900 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
901 { X86::SUB16rr, X86::SUB16rm, 0 },
902 { X86::SUB32rr, X86::SUB32rm, 0 },
903 { X86::SUB64rr, X86::SUB64rm, 0 },
904 { X86::SUB8rr, X86::SUB8rm, 0 },
905 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
906 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
907 { X86::SUBSDrr, X86::SUBSDrm, 0 },
908 { X86::SUBSSrr, X86::SUBSSrm, 0 },
909 // FIXME: TEST*rr -> swapped operand of TEST*mr.
910 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
911 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
912 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
913 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
914 { X86::XOR16rr, X86::XOR16rm, 0 },
915 { X86::XOR32rr, X86::XOR32rm, 0 },
916 { X86::XOR64rr, X86::XOR64rm, 0 },
917 { X86::XOR8rr, X86::XOR8rm, 0 },
918 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
919 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
920 // AVX 128-bit versions of foldable instructions
921 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
922 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
923 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
924 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
925 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
926 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
927 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
928 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
929 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
930 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
931 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
932 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
933 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 },
934 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 },
935 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
936 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
937 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
938 { X86::VADDPDrr, X86::VADDPDrm, 0 },
939 { X86::VADDPSrr, X86::VADDPSrm, 0 },
940 { X86::VADDSDrr, X86::VADDSDrm, 0 },
941 { X86::VADDSSrr, X86::VADDSSrm, 0 },
942 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 },
943 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 },
944 { X86::VANDNPDrr, X86::VANDNPDrm, 0 },
945 { X86::VANDNPSrr, X86::VANDNPSrm, 0 },
946 { X86::VANDPDrr, X86::VANDPDrm, 0 },
947 { X86::VANDPSrr, X86::VANDPSrm, 0 },
948 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 },
949 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 },
950 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 },
951 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 },
952 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 },
953 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 },
954 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
955 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
956 { X86::VDIVPDrr, X86::VDIVPDrm, 0 },
957 { X86::VDIVPSrr, X86::VDIVPSrm, 0 },
958 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
959 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
960 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
961 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
962 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
963 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
964 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
965 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
966 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
967 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
968 { X86::VHADDPDrr, X86::VHADDPDrm, 0 },
969 { X86::VHADDPSrr, X86::VHADDPSrm, 0 },
970 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 },
971 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 },
972 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
973 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
974 { X86::VMAXPDrr, X86::VMAXPDrm, 0 },
975 { X86::VMAXPSrr, X86::VMAXPSrm, 0 },
976 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
977 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
978 { X86::VMINPDrr, X86::VMINPDrm, 0 },
979 { X86::VMINPSrr, X86::VMINPSrm, 0 },
980 { X86::VMINSDrr, X86::VMINSDrm, 0 },
981 { X86::VMINSSrr, X86::VMINSSrm, 0 },
982 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 },
983 { X86::VMULPDrr, X86::VMULPDrm, 0 },
984 { X86::VMULPSrr, X86::VMULPSrm, 0 },
985 { X86::VMULSDrr, X86::VMULSDrm, 0 },
986 { X86::VMULSSrr, X86::VMULSSrm, 0 },
987 { X86::VORPDrr, X86::VORPDrm, 0 },
988 { X86::VORPSrr, X86::VORPSrm, 0 },
989 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 },
990 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 },
991 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 },
992 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 },
993 { X86::VPADDBrr, X86::VPADDBrm, 0 },
994 { X86::VPADDDrr, X86::VPADDDrm, 0 },
995 { X86::VPADDQrr, X86::VPADDQrm, 0 },
996 { X86::VPADDSBrr, X86::VPADDSBrm, 0 },
997 { X86::VPADDSWrr, X86::VPADDSWrm, 0 },
998 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 },
999 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 },
1000 { X86::VPADDWrr, X86::VPADDWrm, 0 },
1001 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, 0 },
1002 { X86::VPANDNrr, X86::VPANDNrm, 0 },
1003 { X86::VPANDrr, X86::VPANDrm, 0 },
1004 { X86::VPAVGBrr, X86::VPAVGBrm, 0 },
1005 { X86::VPAVGWrr, X86::VPAVGWrm, 0 },
1006 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 },
1007 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 },
1008 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 },
1009 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 },
1010 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 },
1011 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 },
1012 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 },
1013 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 },
1014 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 },
1015 { X86::VPHADDDrr, X86::VPHADDDrm, 0 },
1016 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 },
1017 { X86::VPHADDWrr, X86::VPHADDWrm, 0 },
1018 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 },
1019 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 },
1020 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 },
1021 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 },
1022 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 },
1023 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 },
1024 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 },
1025 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 },
1026 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 },
1027 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 },
1028 { X86::VPMINSWrr, X86::VPMINSWrm, 0 },
1029 { X86::VPMINUBrr, X86::VPMINUBrm, 0 },
1030 { X86::VPMINSBrr, X86::VPMINSBrm, 0 },
1031 { X86::VPMINSDrr, X86::VPMINSDrm, 0 },
1032 { X86::VPMINUDrr, X86::VPMINUDrm, 0 },
1033 { X86::VPMINUWrr, X86::VPMINUWrm, 0 },
1034 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 },
1035 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 },
1036 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 },
1037 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 },
1038 { X86::VPMULDQrr, X86::VPMULDQrm, 0 },
1039 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 },
1040 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 },
1041 { X86::VPMULHWrr, X86::VPMULHWrm, 0 },
1042 { X86::VPMULLDrr, X86::VPMULLDrm, 0 },
1043 { X86::VPMULLWrr, X86::VPMULLWrm, 0 },
1044 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 },
1045 { X86::VPORrr, X86::VPORrm, 0 },
1046 { X86::VPSADBWrr, X86::VPSADBWrm, 0 },
1047 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 },
1048 { X86::VPSIGNBrr, X86::VPSIGNBrm, 0 },
1049 { X86::VPSIGNWrr, X86::VPSIGNWrm, 0 },
1050 { X86::VPSIGNDrr, X86::VPSIGNDrm, 0 },
1051 { X86::VPSLLDrr, X86::VPSLLDrm, 0 },
1052 { X86::VPSLLQrr, X86::VPSLLQrm, 0 },
1053 { X86::VPSLLWrr, X86::VPSLLWrm, 0 },
1054 { X86::VPSRADrr, X86::VPSRADrm, 0 },
1055 { X86::VPSRAWrr, X86::VPSRAWrm, 0 },
1056 { X86::VPSRLDrr, X86::VPSRLDrm, 0 },
1057 { X86::VPSRLQrr, X86::VPSRLQrm, 0 },
1058 { X86::VPSRLWrr, X86::VPSRLWrm, 0 },
1059 { X86::VPSUBBrr, X86::VPSUBBrm, 0 },
1060 { X86::VPSUBDrr, X86::VPSUBDrm, 0 },
1061 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 },
1062 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 },
1063 { X86::VPSUBWrr, X86::VPSUBWrm, 0 },
1064 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 },
1065 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 },
1066 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 },
1067 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 },
1068 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 },
1069 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 },
1070 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 },
1071 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 },
1072 { X86::VPXORrr, X86::VPXORrm, 0 },
1073 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 },
1074 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 },
1075 { X86::VSUBPDrr, X86::VSUBPDrm, 0 },
1076 { X86::VSUBPSrr, X86::VSUBPSrm, 0 },
1077 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
1078 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
1079 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 },
1080 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 },
1081 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 },
1082 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 },
1083 { X86::VXORPDrr, X86::VXORPDrm, 0 },
1084 { X86::VXORPSrr, X86::VXORPSrm, 0 },
1085 // AVX 256-bit foldable instructions
1086 { X86::VADDPDYrr, X86::VADDPDYrm, 0 },
1087 { X86::VADDPSYrr, X86::VADDPSYrm, 0 },
1088 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 },
1089 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 },
1090 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 },
1091 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 },
1092 { X86::VANDPDYrr, X86::VANDPDYrm, 0 },
1093 { X86::VANDPSYrr, X86::VANDPSYrm, 0 },
1094 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 },
1095 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 },
1096 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 },
1097 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 },
1098 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 },
1099 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 },
1100 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 },
1101 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 },
1102 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 },
1103 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 },
1104 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 },
1105 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 },
1106 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 },
1107 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 },
1108 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 },
1109 { X86::VMINPDYrr, X86::VMINPDYrm, 0 },
1110 { X86::VMINPSYrr, X86::VMINPSYrm, 0 },
1111 { X86::VMULPDYrr, X86::VMULPDYrm, 0 },
1112 { X86::VMULPSYrr, X86::VMULPSYrm, 0 },
1113 { X86::VORPDYrr, X86::VORPDYrm, 0 },
1114 { X86::VORPSYrr, X86::VORPSYrm, 0 },
1115 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 },
1116 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 },
1117 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 },
1118 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 },
1119 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 },
1120 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 },
1121 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 },
1122 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 },
1123 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 },
1124 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 },
1125 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 },
1126 { X86::VXORPDYrr, X86::VXORPDYrm, 0 },
1127 { X86::VXORPSYrr, X86::VXORPSYrm, 0 },
1128 // AVX2 foldable instructions
1129 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 },
1130 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 },
1131 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 },
1132 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 },
1133 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 },
1134 { X86::VPADDBYrr, X86::VPADDBYrm, 0 },
1135 { X86::VPADDDYrr, X86::VPADDDYrm, 0 },
1136 { X86::VPADDQYrr, X86::VPADDQYrm, 0 },
1137 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 },
1138 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 },
1139 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 },
1140 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 },
1141 { X86::VPADDWYrr, X86::VPADDWYrm, 0 },
1142 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, 0 },
1143 { X86::VPANDNYrr, X86::VPANDNYrm, 0 },
1144 { X86::VPANDYrr, X86::VPANDYrm, 0 },
1145 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 },
1146 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 },
1147 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 },
1148 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 },
1149 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 },
1150 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 },
1151 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 },
1152 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 },
1153 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 },
1154 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 },
1155 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 },
1156 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 },
1157 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 },
1158 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 },
1159 { X86::VPERMDYrr, X86::VPERMDYrm, 0 },
1160 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 },
1161 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 },
1162 { X86::VPERMQYri, X86::VPERMQYmi, 0 },
1163 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 },
1164 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 },
1165 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 },
1166 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 },
1167 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 },
1168 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 },
1169 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 },
1170 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 },
1171 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 },
1172 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 },
1173 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 },
1174 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 },
1175 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 },
1176 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 },
1177 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 },
1178 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 },
1179 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 },
1180 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 },
1181 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 },
1182 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 },
1183 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 },
1184 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 },
1185 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 },
1186 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 },
1187 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 },
1188 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 },
1189 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 },
1190 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 },
1191 { X86::VPORYrr, X86::VPORYrm, 0 },
1192 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 },
1193 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 },
1194 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, 0 },
1195 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, 0 },
1196 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, 0 },
1197 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 },
1198 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 },
1199 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 },
1200 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 },
1201 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 },
1202 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 },
1203 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 },
1204 { X86::VPSRADYrr, X86::VPSRADYrm, 0 },
1205 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 },
1206 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 },
1207 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 },
1208 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 },
1209 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 },
1210 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 },
1211 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 },
1212 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 },
1213 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 },
1214 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 },
1215 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 },
1216 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 },
1217 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 },
1218 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 },
1219 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 },
1220 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 },
1221 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 },
1222 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 },
1223 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 },
1224 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 },
1225 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 },
1226 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 },
1227 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 },
1228 { X86::VPXORYrr, X86::VPXORYrm, 0 },
1229 // FIXME: add AVX 256-bit foldable instructions
1231 // FMA4 foldable patterns
1232 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, 0 },
1233 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, 0 },
1234 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_16 },
1235 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_16 },
1236 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_32 },
1237 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_32 },
1238 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, 0 },
1239 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, 0 },
1240 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_16 },
1241 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_16 },
1242 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_32 },
1243 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_32 },
1244 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, 0 },
1245 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, 0 },
1246 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_16 },
1247 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_16 },
1248 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_32 },
1249 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_32 },
1250 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, 0 },
1251 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, 0 },
1252 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_16 },
1253 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_16 },
1254 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_32 },
1255 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_32 },
1256 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_16 },
1257 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_16 },
1258 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_32 },
1259 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_32 },
1260 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_16 },
1261 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_16 },
1262 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_32 },
1263 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_32 },
1265 // BMI/BMI2 foldable instructions
1266 { X86::ANDN32rr, X86::ANDN32rm, 0 },
1267 { X86::ANDN64rr, X86::ANDN64rm, 0 },
1268 { X86::MULX32rr, X86::MULX32rm, 0 },
1269 { X86::MULX64rr, X86::MULX64rm, 0 },
1270 { X86::PDEP32rr, X86::PDEP32rm, 0 },
1271 { X86::PDEP64rr, X86::PDEP64rm, 0 },
1272 { X86::PEXT32rr, X86::PEXT32rm, 0 },
1273 { X86::PEXT64rr, X86::PEXT64rm, 0 },
1275 // AVX-512 foldable instructions
1276 { X86::VADDPSZrr, X86::VADDPSZrm, 0 },
1277 { X86::VADDPDZrr, X86::VADDPDZrm, 0 },
1278 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 },
1279 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 },
1280 { X86::VMULPSZrr, X86::VMULPSZrm, 0 },
1281 { X86::VMULPDZrr, X86::VMULPDZrm, 0 },
1282 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 },
1283 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 },
1284 { X86::VMINPSZrr, X86::VMINPSZrm, 0 },
1285 { X86::VMINPDZrr, X86::VMINPDZrm, 0 },
1286 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 },
1287 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 },
1288 { X86::VPADDDZrr, X86::VPADDDZrm, 0 },
1289 { X86::VPADDQZrr, X86::VPADDQZrm, 0 },
1290 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 },
1291 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 },
1292 { X86::VPMAXSDZrr, X86::VPMAXSDZrm, 0 },
1293 { X86::VPMAXSQZrr, X86::VPMAXSQZrm, 0 },
1294 { X86::VPMAXUDZrr, X86::VPMAXUDZrm, 0 },
1295 { X86::VPMAXUQZrr, X86::VPMAXUQZrm, 0 },
1296 { X86::VPMINSDZrr, X86::VPMINSDZrm, 0 },
1297 { X86::VPMINSQZrr, X86::VPMINSQZrm, 0 },
1298 { X86::VPMINUDZrr, X86::VPMINUDZrm, 0 },
1299 { X86::VPMINUQZrr, X86::VPMINUQZrm, 0 },
1300 { X86::VPMULDQZrr, X86::VPMULDQZrm, 0 },
1301 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 },
1302 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 },
1303 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 },
1304 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 },
1305 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 },
1306 { X86::VPSUBDZrr, X86::VPSUBDZrm, 0 },
1307 { X86::VPSUBQZrr, X86::VPSUBQZrm, 0 },
1308 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 },
1309 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 },
1310 { X86::VALIGNQrri, X86::VALIGNQrmi, 0 },
1311 { X86::VALIGNDrri, X86::VALIGNDrmi, 0 },
1312 { X86::VPMULUDQZrr, X86::VPMULUDQZrm, 0 },
1314 // AES foldable instructions
1315 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 },
1316 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 },
1317 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 },
1318 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 },
1319 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, TB_ALIGN_16 },
1320 { X86::VAESDECrr, X86::VAESDECrm, TB_ALIGN_16 },
1321 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, TB_ALIGN_16 },
1322 { X86::VAESENCrr, X86::VAESENCrm, TB_ALIGN_16 },
1324 // SHA foldable instructions
1325 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 },
1326 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 },
1327 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 },
1328 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 },
1329 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 },
1330 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 },
1331 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 },
1332 };
1334 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1335 unsigned RegOp = OpTbl2[i].RegOp;
1336 unsigned MemOp = OpTbl2[i].MemOp;
1337 unsigned Flags = OpTbl2[i].Flags;
1338 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1339 RegOp, MemOp,
1340 // Index 2, folded load
1341 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1342 }
1344 static const X86OpTblEntry OpTbl3[] = {
1345 // FMA foldable instructions
1346 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, TB_ALIGN_NONE },
1347 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, TB_ALIGN_NONE },
1348 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, TB_ALIGN_NONE },
1349 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, TB_ALIGN_NONE },
1350 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, TB_ALIGN_NONE },
1351 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, TB_ALIGN_NONE },
1353 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_NONE },
1354 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_NONE },
1355 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_NONE },
1356 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_NONE },
1357 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_NONE },
1358 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_NONE },
1359 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_NONE },
1360 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_NONE },
1361 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_NONE },
1362 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_NONE },
1363 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_NONE },
1364 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_NONE },
1366 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, TB_ALIGN_NONE },
1367 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, TB_ALIGN_NONE },
1368 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, TB_ALIGN_NONE },
1369 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, TB_ALIGN_NONE },
1370 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, TB_ALIGN_NONE },
1371 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, TB_ALIGN_NONE },
1373 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_NONE },
1374 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_NONE },
1375 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_NONE },
1376 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_NONE },
1377 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_NONE },
1378 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_NONE },
1379 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_NONE },
1380 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_NONE },
1381 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_NONE },
1382 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_NONE },
1383 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_NONE },
1384 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_NONE },
1386 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, TB_ALIGN_NONE },
1387 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, TB_ALIGN_NONE },
1388 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, TB_ALIGN_NONE },
1389 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, TB_ALIGN_NONE },
1390 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, TB_ALIGN_NONE },
1391 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, TB_ALIGN_NONE },
1393 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_NONE },
1394 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_NONE },
1395 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_NONE },
1396 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_NONE },
1397 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_NONE },
1398 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_NONE },
1399 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_NONE },
1400 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_NONE },
1401 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_NONE },
1402 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_NONE },
1403 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_NONE },
1404 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_NONE },
1406 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, TB_ALIGN_NONE },
1407 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, TB_ALIGN_NONE },
1408 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, TB_ALIGN_NONE },
1409 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, TB_ALIGN_NONE },
1410 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, TB_ALIGN_NONE },
1411 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, TB_ALIGN_NONE },
1413 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_NONE },
1414 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_NONE },
1415 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_NONE },
1416 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_NONE },
1417 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_NONE },
1418 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_NONE },
1419 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_NONE },
1420 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_NONE },
1421 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_NONE },
1422 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_NONE },
1423 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_NONE },
1424 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_NONE },
1426 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_NONE },
1427 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_NONE },
1428 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_NONE },
1429 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_NONE },
1430 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_NONE },
1431 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_NONE },
1432 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_NONE },
1433 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_NONE },
1434 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_NONE },
1435 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_NONE },
1436 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_NONE },
1437 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_NONE },
1439 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_NONE },
1440 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_NONE },
1441 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_NONE },
1442 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_NONE },
1443 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_NONE },
1444 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_NONE },
1445 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_NONE },
1446 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_NONE },
1447 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_NONE },
1448 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_NONE },
1449 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_NONE },
1450 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_NONE },
1452 // FMA4 foldable patterns
1453 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, 0 },
1454 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, 0 },
1455 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_16 },
1456 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_16 },
1457 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_32 },
1458 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_32 },
1459 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, 0 },
1460 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, 0 },
1461 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_16 },
1462 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_16 },
1463 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_32 },
1464 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_32 },
1465 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, 0 },
1466 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, 0 },
1467 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_16 },
1468 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_16 },
1469 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_32 },
1470 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_32 },
1471 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, 0 },
1472 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, 0 },
1473 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_16 },
1474 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_16 },
1475 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_32 },
1476 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_32 },
1477 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_16 },
1478 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_16 },
1479 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_32 },
1480 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_32 },
1481 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_16 },
1482 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_16 },
1483 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_32 },
1484 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_32 },
1485 // AVX-512 VPERMI instructions with 3 source operands.
1486 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 },
1487 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 },
1488 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 },
1489 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 },
1490 { X86::VBLENDMPDZrr, X86::VBLENDMPDZrm, 0 },
1491 { X86::VBLENDMPSZrr, X86::VBLENDMPSZrm, 0 },
1492 { X86::VPBLENDMDZrr, X86::VPBLENDMDZrm, 0 },
1493 { X86::VPBLENDMQZrr, X86::VPBLENDMQZrm, 0 }
1494 };
1496 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1497 unsigned RegOp = OpTbl3[i].RegOp;
1498 unsigned MemOp = OpTbl3[i].MemOp;
1499 unsigned Flags = OpTbl3[i].Flags;
1500 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1501 RegOp, MemOp,
1502 // Index 3, folded load
1503 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1504 }
1506 }
1508 void
1509 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1510 MemOp2RegOpTableType &M2RTable,
1511 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1512 if ((Flags & TB_NO_FORWARD) == 0) {
1513 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1514 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1515 }
1516 if ((Flags & TB_NO_REVERSE) == 0) {
1517 assert(!M2RTable.count(MemOp) &&
1518 "Duplicated entries in unfolding maps?");
1519 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1520 }
1521 }
1523 bool
1524 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1525 unsigned &SrcReg, unsigned &DstReg,
1526 unsigned &SubIdx) const {
1527 switch (MI.getOpcode()) {
1528 default: break;
1529 case X86::MOVSX16rr8:
1530 case X86::MOVZX16rr8:
1531 case X86::MOVSX32rr8:
1532 case X86::MOVZX32rr8:
1533 case X86::MOVSX64rr8:
1534 if (!Subtarget.is64Bit())
1535 // It's not always legal to reference the low 8-bit of the larger
1536 // register in 32-bit mode.
1537 return false;
1538 case X86::MOVSX32rr16:
1539 case X86::MOVZX32rr16:
1540 case X86::MOVSX64rr16:
1541 case X86::MOVSX64rr32: {
1542 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1543 // Be conservative.
1544 return false;
1545 SrcReg = MI.getOperand(1).getReg();
1546 DstReg = MI.getOperand(0).getReg();
1547 switch (MI.getOpcode()) {
1548 default: llvm_unreachable("Unreachable!");
1549 case X86::MOVSX16rr8:
1550 case X86::MOVZX16rr8:
1551 case X86::MOVSX32rr8:
1552 case X86::MOVZX32rr8:
1553 case X86::MOVSX64rr8:
1554 SubIdx = X86::sub_8bit;
1555 break;
1556 case X86::MOVSX32rr16:
1557 case X86::MOVZX32rr16:
1558 case X86::MOVSX64rr16:
1559 SubIdx = X86::sub_16bit;
1560 break;
1561 case X86::MOVSX64rr32:
1562 SubIdx = X86::sub_32bit;
1563 break;
1564 }
1565 return true;
1566 }
1567 }
1568 return false;
1569 }
1571 /// isFrameOperand - Return true and the FrameIndex if the specified
1572 /// operand and follow operands form a reference to the stack frame.
1573 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1574 int &FrameIndex) const {
1575 if (MI->getOperand(Op+X86::AddrBaseReg).isFI() &&
1576 MI->getOperand(Op+X86::AddrScaleAmt).isImm() &&
1577 MI->getOperand(Op+X86::AddrIndexReg).isReg() &&
1578 MI->getOperand(Op+X86::AddrDisp).isImm() &&
1579 MI->getOperand(Op+X86::AddrScaleAmt).getImm() == 1 &&
1580 MI->getOperand(Op+X86::AddrIndexReg).getReg() == 0 &&
1581 MI->getOperand(Op+X86::AddrDisp).getImm() == 0) {
1582 FrameIndex = MI->getOperand(Op+X86::AddrBaseReg).getIndex();
1583 return true;
1584 }
1585 return false;
1586 }
1588 static bool isFrameLoadOpcode(int Opcode) {
1589 switch (Opcode) {
1590 default:
1591 return false;
1592 case X86::MOV8rm:
1593 case X86::MOV16rm:
1594 case X86::MOV32rm:
1595 case X86::MOV64rm:
1596 case X86::LD_Fp64m:
1597 case X86::MOVSSrm:
1598 case X86::MOVSDrm:
1599 case X86::MOVAPSrm:
1600 case X86::MOVAPDrm:
1601 case X86::MOVDQArm:
1602 case X86::VMOVSSrm:
1603 case X86::VMOVSDrm:
1604 case X86::VMOVAPSrm:
1605 case X86::VMOVAPDrm:
1606 case X86::VMOVDQArm:
1607 case X86::VMOVAPSYrm:
1608 case X86::VMOVAPDYrm:
1609 case X86::VMOVDQAYrm:
1610 case X86::MMX_MOVD64rm:
1611 case X86::MMX_MOVQ64rm:
1612 case X86::VMOVAPSZrm:
1613 case X86::VMOVUPSZrm:
1614 return true;
1615 }
1616 }
1618 static bool isFrameStoreOpcode(int Opcode) {
1619 switch (Opcode) {
1620 default: break;
1621 case X86::MOV8mr:
1622 case X86::MOV16mr:
1623 case X86::MOV32mr:
1624 case X86::MOV64mr:
1625 case X86::ST_FpP64m:
1626 case X86::MOVSSmr:
1627 case X86::MOVSDmr:
1628 case X86::MOVAPSmr:
1629 case X86::MOVAPDmr:
1630 case X86::MOVDQAmr:
1631 case X86::VMOVSSmr:
1632 case X86::VMOVSDmr:
1633 case X86::VMOVAPSmr:
1634 case X86::VMOVAPDmr:
1635 case X86::VMOVDQAmr:
1636 case X86::VMOVAPSYmr:
1637 case X86::VMOVAPDYmr:
1638 case X86::VMOVDQAYmr:
1639 case X86::VMOVUPSZmr:
1640 case X86::VMOVAPSZmr:
1641 case X86::MMX_MOVD64mr:
1642 case X86::MMX_MOVQ64mr:
1643 case X86::MMX_MOVNTQmr:
1644 return true;
1645 }
1646 return false;
1647 }
1649 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1650 int &FrameIndex) const {
1651 if (isFrameLoadOpcode(MI->getOpcode()))
1652 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1653 return MI->getOperand(0).getReg();
1654 return 0;
1655 }
1657 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1658 int &FrameIndex) const {
1659 if (isFrameLoadOpcode(MI->getOpcode())) {
1660 unsigned Reg;
1661 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1662 return Reg;
1663 // Check for post-frame index elimination operations
1664 const MachineMemOperand *Dummy;
1665 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1666 }
1667 return 0;
1668 }
1670 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1671 int &FrameIndex) const {
1672 if (isFrameStoreOpcode(MI->getOpcode()))
1673 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1674 isFrameOperand(MI, 0, FrameIndex))
1675 return MI->getOperand(X86::AddrNumOperands).getReg();
1676 return 0;
1677 }
1679 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1680 int &FrameIndex) const {
1681 if (isFrameStoreOpcode(MI->getOpcode())) {
1682 unsigned Reg;
1683 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1684 return Reg;
1685 // Check for post-frame index elimination operations
1686 const MachineMemOperand *Dummy;
1687 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1688 }
1689 return 0;
1690 }
1692 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1693 /// X86::MOVPC32r.
1694 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1695 // Don't waste compile time scanning use-def chains of physregs.
1696 if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1697 return false;
1698 bool isPICBase = false;
1699 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
1700 E = MRI.def_instr_end(); I != E; ++I) {
1701 MachineInstr *DefMI = &*I;
1702 if (DefMI->getOpcode() != X86::MOVPC32r)
1703 return false;
1704 assert(!isPICBase && "More than one PIC base?");
1705 isPICBase = true;
1706 }
1707 return isPICBase;
1708 }
1710 bool
1711 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1712 AliasAnalysis *AA) const {
1713 switch (MI->getOpcode()) {
1714 default: break;
1715 case X86::MOV8rm:
1716 case X86::MOV16rm:
1717 case X86::MOV32rm:
1718 case X86::MOV64rm:
1719 case X86::LD_Fp64m:
1720 case X86::MOVSSrm:
1721 case X86::MOVSDrm:
1722 case X86::MOVAPSrm:
1723 case X86::MOVUPSrm:
1724 case X86::MOVAPDrm:
1725 case X86::MOVDQArm:
1726 case X86::MOVDQUrm:
1727 case X86::VMOVSSrm:
1728 case X86::VMOVSDrm:
1729 case X86::VMOVAPSrm:
1730 case X86::VMOVUPSrm:
1731 case X86::VMOVAPDrm:
1732 case X86::VMOVDQArm:
1733 case X86::VMOVDQUrm:
1734 case X86::VMOVAPSYrm:
1735 case X86::VMOVUPSYrm:
1736 case X86::VMOVAPDYrm:
1737 case X86::VMOVDQAYrm:
1738 case X86::VMOVDQUYrm:
1739 case X86::MMX_MOVD64rm:
1740 case X86::MMX_MOVQ64rm:
1741 case X86::FsVMOVAPSrm:
1742 case X86::FsVMOVAPDrm:
1743 case X86::FsMOVAPSrm:
1744 case X86::FsMOVAPDrm: {
1745 // Loads from constant pools are trivially rematerializable.
1746 if (MI->getOperand(1+X86::AddrBaseReg).isReg() &&
1747 MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
1748 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
1749 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
1750 MI->isInvariantLoad(AA)) {
1751 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
1752 if (BaseReg == 0 || BaseReg == X86::RIP)
1753 return true;
1754 // Allow re-materialization of PIC load.
1755 if (!ReMatPICStubLoad && MI->getOperand(1+X86::AddrDisp).isGlobal())
1756 return false;
1757 const MachineFunction &MF = *MI->getParent()->getParent();
1758 const MachineRegisterInfo &MRI = MF.getRegInfo();
1759 return regIsPICBase(BaseReg, MRI);
1760 }
1761 return false;
1762 }
1764 case X86::LEA32r:
1765 case X86::LEA64r: {
1766 if (MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
1767 MI->getOperand(1+X86::AddrIndexReg).isReg() &&
1768 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
1769 !MI->getOperand(1+X86::AddrDisp).isReg()) {
1770 // lea fi#, lea GV, etc. are all rematerializable.
1771 if (!MI->getOperand(1+X86::AddrBaseReg).isReg())
1772 return true;
1773 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
1774 if (BaseReg == 0)
1775 return true;
1776 // Allow re-materialization of lea PICBase + x.
1777 const MachineFunction &MF = *MI->getParent()->getParent();
1778 const MachineRegisterInfo &MRI = MF.getRegInfo();
1779 return regIsPICBase(BaseReg, MRI);
1780 }
1781 return false;
1782 }
1783 }
1785 // All other instructions marked M_REMATERIALIZABLE are always trivially
1786 // rematerializable.
1787 return true;
1788 }
1790 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1791 MachineBasicBlock::iterator I) const {
1792 MachineBasicBlock::iterator E = MBB.end();
1794 // For compile time consideration, if we are not able to determine the
1795 // safety after visiting 4 instructions in each direction, we will assume
1796 // it's not safe.
1797 MachineBasicBlock::iterator Iter = I;
1798 for (unsigned i = 0; Iter != E && i < 4; ++i) {
1799 bool SeenDef = false;
1800 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1801 MachineOperand &MO = Iter->getOperand(j);
1802 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1803 SeenDef = true;
1804 if (!MO.isReg())
1805 continue;
1806 if (MO.getReg() == X86::EFLAGS) {
1807 if (MO.isUse())
1808 return false;
1809 SeenDef = true;
1810 }
1811 }
1813 if (SeenDef)
1814 // This instruction defines EFLAGS, no need to look any further.
1815 return true;
1816 ++Iter;
1817 // Skip over DBG_VALUE.
1818 while (Iter != E && Iter->isDebugValue())
1819 ++Iter;
1820 }
1822 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1823 // live in.
1824 if (Iter == E) {
1825 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1826 SE = MBB.succ_end(); SI != SE; ++SI)
1827 if ((*SI)->isLiveIn(X86::EFLAGS))
1828 return false;
1829 return true;
1830 }
1832 MachineBasicBlock::iterator B = MBB.begin();
1833 Iter = I;
1834 for (unsigned i = 0; i < 4; ++i) {
1835 // If we make it to the beginning of the block, it's safe to clobber
1836 // EFLAGS iff EFLAGS is not live-in.
1837 if (Iter == B)
1838 return !MBB.isLiveIn(X86::EFLAGS);
1840 --Iter;
1841 // Skip over DBG_VALUE.
1842 while (Iter != B && Iter->isDebugValue())
1843 --Iter;
1845 bool SawKill = false;
1846 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1847 MachineOperand &MO = Iter->getOperand(j);
1848 // A register mask may clobber EFLAGS, but we should still look for a
1849 // live EFLAGS def.
1850 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1851 SawKill = true;
1852 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1853 if (MO.isDef()) return MO.isDead();
1854 if (MO.isKill()) SawKill = true;
1855 }
1856 }
1858 if (SawKill)
1859 // This instruction kills EFLAGS and doesn't redefine it, so
1860 // there's no need to look further.
1861 return true;
1862 }
1864 // Conservative answer.
1865 return false;
1866 }
1868 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1869 MachineBasicBlock::iterator I,
1870 unsigned DestReg, unsigned SubIdx,
1871 const MachineInstr *Orig,
1872 const TargetRegisterInfo &TRI) const {
1873 // MOV32r0 is implemented with a xor which clobbers condition code.
1874 // Re-materialize it as movri instructions to avoid side effects.
1875 unsigned Opc = Orig->getOpcode();
1876 if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1877 DebugLoc DL = Orig->getDebugLoc();
1878 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1879 .addImm(0);
1880 } else {
1881 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1882 MBB.insert(I, MI);
1883 }
1885 MachineInstr *NewMI = std::prev(I);
1886 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1887 }
1889 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1890 /// is not marked dead.
1891 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1892 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1893 MachineOperand &MO = MI->getOperand(i);
1894 if (MO.isReg() && MO.isDef() &&
1895 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1896 return true;
1897 }
1898 }
1899 return false;
1900 }
1902 /// getTruncatedShiftCount - check whether the shift count for a machine operand
1903 /// is non-zero.
1904 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1905 unsigned ShiftAmtOperandIdx) {
1906 // The shift count is six bits with the REX.W prefix and five bits without.
1907 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1908 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1909 return Imm & ShiftCountMask;
1910 }
1912 /// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1913 /// can be represented by a LEA instruction.
1914 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1915 // Left shift instructions can be transformed into load-effective-address
1916 // instructions if we can encode them appropriately.
1917 // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1918 // The SIB.scale field is two bits wide which means that we can encode any
1919 // shift amount less than 4.
1920 return ShAmt < 4 && ShAmt > 0;
1921 }
1923 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
1924 unsigned Opc, bool AllowSP,
1925 unsigned &NewSrc, bool &isKill, bool &isUndef,
1926 MachineOperand &ImplicitOp) const {
1927 MachineFunction &MF = *MI->getParent()->getParent();
1928 const TargetRegisterClass *RC;
1929 if (AllowSP) {
1930 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1931 } else {
1932 RC = Opc != X86::LEA32r ?
1933 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1934 }
1935 unsigned SrcReg = Src.getReg();
1937 // For both LEA64 and LEA32 the register already has essentially the right
1938 // type (32-bit or 64-bit) we may just need to forbid SP.
1939 if (Opc != X86::LEA64_32r) {
1940 NewSrc = SrcReg;
1941 isKill = Src.isKill();
1942 isUndef = Src.isUndef();
1944 if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
1945 !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1946 return false;
1948 return true;
1949 }
1951 // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1952 // another we need to add 64-bit registers to the final MI.
1953 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
1954 ImplicitOp = Src;
1955 ImplicitOp.setImplicit();
1957 NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
1958 MachineBasicBlock::LivenessQueryResult LQR =
1959 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
1961 switch (LQR) {
1962 case MachineBasicBlock::LQR_Unknown:
1963 // We can't give sane liveness flags to the instruction, abandon LEA
1964 // formation.
1965 return false;
1966 case MachineBasicBlock::LQR_Live:
1967 isKill = MI->killsRegister(SrcReg);
1968 isUndef = false;
1969 break;
1970 default:
1971 // The physreg itself is dead, so we have to use it as an <undef>.
1972 isKill = false;
1973 isUndef = true;
1974 break;
1975 }
1976 } else {
1977 // Virtual register of the wrong class, we have to create a temporary 64-bit
1978 // vreg to feed into the LEA.
1979 NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1980 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1981 get(TargetOpcode::COPY))
1982 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1983 .addOperand(Src);
1985 // Which is obviously going to be dead after we're done with it.
1986 isKill = true;
1987 isUndef = false;
1988 }
1990 // We've set all the parameters without issue.
1991 return true;
1992 }
1994 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1995 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1996 /// to a 32-bit superregister and then truncating back down to a 16-bit
1997 /// subregister.
1998 MachineInstr *
1999 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
2000 MachineFunction::iterator &MFI,
2001 MachineBasicBlock::iterator &MBBI,
2002 LiveVariables *LV) const {
2003 MachineInstr *MI = MBBI;
2004 unsigned Dest = MI->getOperand(0).getReg();
2005 unsigned Src = MI->getOperand(1).getReg();
2006 bool isDead = MI->getOperand(0).isDead();
2007 bool isKill = MI->getOperand(1).isKill();
2009 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
2010 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
2011 unsigned Opc, leaInReg;
2012 if (Subtarget.is64Bit()) {
2013 Opc = X86::LEA64_32r;
2014 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2015 } else {
2016 Opc = X86::LEA32r;
2017 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2018 }
2020 // Build and insert into an implicit UNDEF value. This is OK because
2021 // well be shifting and then extracting the lower 16-bits.
2022 // This has the potential to cause partial register stall. e.g.
2023 // movw (%rbp,%rcx,2), %dx
2024 // leal -65(%rdx), %esi
2025 // But testing has shown this *does* help performance in 64-bit mode (at
2026 // least on modern x86 machines).
2027 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
2028 MachineInstr *InsMI =
2029 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2030 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
2031 .addReg(Src, getKillRegState(isKill));
2033 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
2034 get(Opc), leaOutReg);
2035 switch (MIOpc) {
2036 default: llvm_unreachable("Unreachable!");
2037 case X86::SHL16ri: {
2038 unsigned ShAmt = MI->getOperand(2).getImm();
2039 MIB.addReg(0).addImm(1 << ShAmt)
2040 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
2041 break;
2042 }
2043 case X86::INC16r:
2044 case X86::INC64_16r:
2045 addRegOffset(MIB, leaInReg, true, 1);
2046 break;
2047 case X86::DEC16r:
2048 case X86::DEC64_16r:
2049 addRegOffset(MIB, leaInReg, true, -1);
2050 break;
2051 case X86::ADD16ri:
2052 case X86::ADD16ri8:
2053 case X86::ADD16ri_DB:
2054 case X86::ADD16ri8_DB:
2055 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
2056 break;
2057 case X86::ADD16rr:
2058 case X86::ADD16rr_DB: {
2059 unsigned Src2 = MI->getOperand(2).getReg();
2060 bool isKill2 = MI->getOperand(2).isKill();
2061 unsigned leaInReg2 = 0;
2062 MachineInstr *InsMI2 = nullptr;
2063 if (Src == Src2) {
2064 // ADD16rr %reg1028<kill>, %reg1028
2065 // just a single insert_subreg.
2066 addRegReg(MIB, leaInReg, true, leaInReg, false);
2067 } else {
2068 if (Subtarget.is64Bit())
2069 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2070 else
2071 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2072 // Build and insert into an implicit UNDEF value. This is OK because
2073 // well be shifting and then extracting the lower 16-bits.
2074 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2075 InsMI2 =
2076 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2077 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2078 .addReg(Src2, getKillRegState(isKill2));
2079 addRegReg(MIB, leaInReg, true, leaInReg2, true);
2080 }
2081 if (LV && isKill2 && InsMI2)
2082 LV->replaceKillInstruction(Src2, MI, InsMI2);
2083 break;
2084 }
2085 }
2087 MachineInstr *NewMI = MIB;
2088 MachineInstr *ExtMI =
2089 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2090 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2091 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2093 if (LV) {
2094 // Update live variables
2095 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2096 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2097 if (isKill)
2098 LV->replaceKillInstruction(Src, MI, InsMI);
2099 if (isDead)
2100 LV->replaceKillInstruction(Dest, MI, ExtMI);
2101 }
2103 return ExtMI;
2104 }
2106 /// convertToThreeAddress - This method must be implemented by targets that
2107 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
2108 /// may be able to convert a two-address instruction into a true
2109 /// three-address instruction on demand. This allows the X86 target (for
2110 /// example) to convert ADD and SHL instructions into LEA instructions if they
2111 /// would require register copies due to two-addressness.
2112 ///
2113 /// This method returns a null pointer if the transformation cannot be
2114 /// performed, otherwise it returns the new instruction.
2115 ///
2116 MachineInstr *
2117 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2118 MachineBasicBlock::iterator &MBBI,
2119 LiveVariables *LV) const {
2120 MachineInstr *MI = MBBI;
2122 // The following opcodes also sets the condition code register(s). Only
2123 // convert them to equivalent lea if the condition code register def's
2124 // are dead!
2125 if (hasLiveCondCodeDef(MI))
2126 return nullptr;
2128 MachineFunction &MF = *MI->getParent()->getParent();
2129 // All instructions input are two-addr instructions. Get the known operands.
2130 const MachineOperand &Dest = MI->getOperand(0);
2131 const MachineOperand &Src = MI->getOperand(1);
2133 MachineInstr *NewMI = nullptr;
2134 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
2135 // we have better subtarget support, enable the 16-bit LEA generation here.
2136 // 16-bit LEA is also slow on Core2.
2137 bool DisableLEA16 = true;
2138 bool is64Bit = Subtarget.is64Bit();
2140 unsigned MIOpc = MI->getOpcode();
2141 switch (MIOpc) {
2142 case X86::SHL64ri: {
2143 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2144 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2145 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2147 // LEA can't handle RSP.
2148 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2149 !MF.getRegInfo().constrainRegClass(Src.getReg(),
2150 &X86::GR64_NOSPRegClass))
2151 return nullptr;
2153 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2154 .addOperand(Dest)
2155 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2156 break;
2157 }
2158 case X86::SHL32ri: {
2159 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2160 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2161 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2163 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2165 // LEA can't handle ESP.
2166 bool isKill, isUndef;
2167 unsigned SrcReg;
2168 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2169 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2170 SrcReg, isKill, isUndef, ImplicitOp))
2171 return nullptr;
2173 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2174 .addOperand(Dest)
2175 .addReg(0).addImm(1 << ShAmt)
2176 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2177 .addImm(0).addReg(0);
2178 if (ImplicitOp.getReg() != 0)
2179 MIB.addOperand(ImplicitOp);
2180 NewMI = MIB;
2182 break;
2183 }
2184 case X86::SHL16ri: {
2185 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2186 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2187 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2189 if (DisableLEA16)
2190 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : nullptr;
2191 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2192 .addOperand(Dest)
2193 .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2194 break;
2195 }
2196 default: {
2198 switch (MIOpc) {
2199 default: return nullptr;
2200 case X86::INC64r:
2201 case X86::INC32r:
2202 case X86::INC64_32r: {
2203 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2204 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2205 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2206 bool isKill, isUndef;
2207 unsigned SrcReg;
2208 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2209 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2210 SrcReg, isKill, isUndef, ImplicitOp))
2211 return nullptr;
2213 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2214 .addOperand(Dest)
2215 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2216 if (ImplicitOp.getReg() != 0)
2217 MIB.addOperand(ImplicitOp);
2219 NewMI = addOffset(MIB, 1);
2220 break;
2221 }
2222 case X86::INC16r:
2223 case X86::INC64_16r:
2224 if (DisableLEA16)
2225 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2226 : nullptr;
2227 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2228 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2229 .addOperand(Dest).addOperand(Src), 1);
2230 break;
2231 case X86::DEC64r:
2232 case X86::DEC32r:
2233 case X86::DEC64_32r: {
2234 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2235 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2236 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2238 bool isKill, isUndef;
2239 unsigned SrcReg;
2240 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2241 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2242 SrcReg, isKill, isUndef, ImplicitOp))
2243 return nullptr;
2245 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2246 .addOperand(Dest)
2247 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2248 if (ImplicitOp.getReg() != 0)
2249 MIB.addOperand(ImplicitOp);
2251 NewMI = addOffset(MIB, -1);
2253 break;
2254 }
2255 case X86::DEC16r:
2256 case X86::DEC64_16r:
2257 if (DisableLEA16)
2258 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2259 : nullptr;
2260 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2261 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2262 .addOperand(Dest).addOperand(Src), -1);
2263 break;
2264 case X86::ADD64rr:
2265 case X86::ADD64rr_DB:
2266 case X86::ADD32rr:
2267 case X86::ADD32rr_DB: {
2268 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2269 unsigned Opc;
2270 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2271 Opc = X86::LEA64r;
2272 else
2273 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2275 bool isKill, isUndef;
2276 unsigned SrcReg;
2277 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2278 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2279 SrcReg, isKill, isUndef, ImplicitOp))
2280 return nullptr;
2282 const MachineOperand &Src2 = MI->getOperand(2);
2283 bool isKill2, isUndef2;
2284 unsigned SrcReg2;
2285 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2286 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2287 SrcReg2, isKill2, isUndef2, ImplicitOp2))
2288 return nullptr;
2290 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2291 .addOperand(Dest);
2292 if (ImplicitOp.getReg() != 0)
2293 MIB.addOperand(ImplicitOp);
2294 if (ImplicitOp2.getReg() != 0)
2295 MIB.addOperand(ImplicitOp2);
2297 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2299 // Preserve undefness of the operands.
2300 NewMI->getOperand(1).setIsUndef(isUndef);
2301 NewMI->getOperand(3).setIsUndef(isUndef2);
2303 if (LV && Src2.isKill())
2304 LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2305 break;
2306 }
2307 case X86::ADD16rr:
2308 case X86::ADD16rr_DB: {
2309 if (DisableLEA16)
2310 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2311 : nullptr;
2312 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2313 unsigned Src2 = MI->getOperand(2).getReg();
2314 bool isKill2 = MI->getOperand(2).isKill();
2315 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2316 .addOperand(Dest),
2317 Src.getReg(), Src.isKill(), Src2, isKill2);
2319 // Preserve undefness of the operands.
2320 bool isUndef = MI->getOperand(1).isUndef();
2321 bool isUndef2 = MI->getOperand(2).isUndef();
2322 NewMI->getOperand(1).setIsUndef(isUndef);
2323 NewMI->getOperand(3).setIsUndef(isUndef2);
2325 if (LV && isKill2)
2326 LV->replaceKillInstruction(Src2, MI, NewMI);
2327 break;
2328 }
2329 case X86::ADD64ri32:
2330 case X86::ADD64ri8:
2331 case X86::ADD64ri32_DB:
2332 case X86::ADD64ri8_DB:
2333 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2334 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2335 .addOperand(Dest).addOperand(Src),
2336 MI->getOperand(2).getImm());
2337 break;
2338 case X86::ADD32ri:
2339 case X86::ADD32ri8:
2340 case X86::ADD32ri_DB:
2341 case X86::ADD32ri8_DB: {
2342 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2343 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2345 bool isKill, isUndef;
2346 unsigned SrcReg;
2347 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2348 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2349 SrcReg, isKill, isUndef, ImplicitOp))
2350 return nullptr;
2352 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2353 .addOperand(Dest)
2354 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2355 if (ImplicitOp.getReg() != 0)
2356 MIB.addOperand(ImplicitOp);
2358 NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2359 break;
2360 }
2361 case X86::ADD16ri:
2362 case X86::ADD16ri8:
2363 case X86::ADD16ri_DB:
2364 case X86::ADD16ri8_DB:
2365 if (DisableLEA16)
2366 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2367 : nullptr;
2368 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2369 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2370 .addOperand(Dest).addOperand(Src),
2371 MI->getOperand(2).getImm());
2372 break;
2373 }
2374 }
2375 }
2377 if (!NewMI) return nullptr;
2379 if (LV) { // Update live variables
2380 if (Src.isKill())
2381 LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2382 if (Dest.isDead())
2383 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2384 }
2386 MFI->insert(MBBI, NewMI); // Insert the new inst
2387 return NewMI;
2388 }
2390 /// commuteInstruction - We have a few instructions that must be hacked on to
2391 /// commute them.
2392 ///
2393 MachineInstr *
2394 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2395 switch (MI->getOpcode()) {
2396 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2397 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2398 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2399 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2400 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2401 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2402 unsigned Opc;
2403 unsigned Size;
2404 switch (MI->getOpcode()) {
2405 default: llvm_unreachable("Unreachable!");
2406 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2407 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2408 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2409 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2410 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2411 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2412 }
2413 unsigned Amt = MI->getOperand(3).getImm();
2414 if (NewMI) {
2415 MachineFunction &MF = *MI->getParent()->getParent();
2416 MI = MF.CloneMachineInstr(MI);
2417 NewMI = false;
2418 }
2419 MI->setDesc(get(Opc));
2420 MI->getOperand(3).setImm(Size-Amt);
2421 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2422 }
2423 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr:
2424 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2425 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
2426 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2427 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2428 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr:
2429 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
2430 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2431 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2432 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
2433 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
2434 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2435 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
2436 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2437 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
2438 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2439 unsigned Opc;
2440 switch (MI->getOpcode()) {
2441 default: llvm_unreachable("Unreachable!");
2442 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2443 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2444 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2445 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2446 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2447 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2448 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2449 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2450 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2451 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2452 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2453 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2454 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2455 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2456 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2457 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2458 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2459 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
2460 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2461 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2462 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2463 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2464 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2465 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2466 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2467 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2468 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2469 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2470 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2471 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2472 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2473 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
2474 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
2475 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2476 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2477 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2478 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2479 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
2480 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
2481 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2482 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2483 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2484 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2485 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
2486 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
2487 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2488 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2489 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2490 }
2491 if (NewMI) {
2492 MachineFunction &MF = *MI->getParent()->getParent();
2493 MI = MF.CloneMachineInstr(MI);
2494 NewMI = false;
2495 }
2496 MI->setDesc(get(Opc));
2497 // Fallthrough intended.
2498 }
2499 default:
2500 return TargetInstrInfo::commuteInstruction(MI, NewMI);
2501 }
2502 }
2504 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
2505 unsigned &SrcOpIdx2) const {
2506 switch (MI->getOpcode()) {
2507 case X86::VFMADDPDr231r:
2508 case X86::VFMADDPSr231r:
2509 case X86::VFMADDSDr231r:
2510 case X86::VFMADDSSr231r:
2511 case X86::VFMSUBPDr231r:
2512 case X86::VFMSUBPSr231r:
2513 case X86::VFMSUBSDr231r:
2514 case X86::VFMSUBSSr231r:
2515 case X86::VFNMADDPDr231r:
2516 case X86::VFNMADDPSr231r:
2517 case X86::VFNMADDSDr231r:
2518 case X86::VFNMADDSSr231r:
2519 case X86::VFNMSUBPDr231r:
2520 case X86::VFNMSUBPSr231r:
2521 case X86::VFNMSUBSDr231r:
2522 case X86::VFNMSUBSSr231r:
2523 case X86::VFMADDPDr231rY:
2524 case X86::VFMADDPSr231rY:
2525 case X86::VFMSUBPDr231rY:
2526 case X86::VFMSUBPSr231rY:
2527 case X86::VFNMADDPDr231rY:
2528 case X86::VFNMADDPSr231rY:
2529 case X86::VFNMSUBPDr231rY:
2530 case X86::VFNMSUBPSr231rY:
2531 SrcOpIdx1 = 2;
2532 SrcOpIdx2 = 3;
2533 return true;
2534 default:
2535 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
2536 }
2537 }
2539 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2540 switch (BrOpc) {
2541 default: return X86::COND_INVALID;
2542 case X86::JE_4: return X86::COND_E;
2543 case X86::JNE_4: return X86::COND_NE;
2544 case X86::JL_4: return X86::COND_L;
2545 case X86::JLE_4: return X86::COND_LE;
2546 case X86::JG_4: return X86::COND_G;
2547 case X86::JGE_4: return X86::COND_GE;
2548 case X86::JB_4: return X86::COND_B;
2549 case X86::JBE_4: return X86::COND_BE;
2550 case X86::JA_4: return X86::COND_A;
2551 case X86::JAE_4: return X86::COND_AE;
2552 case X86::JS_4: return X86::COND_S;
2553 case X86::JNS_4: return X86::COND_NS;
2554 case X86::JP_4: return X86::COND_P;
2555 case X86::JNP_4: return X86::COND_NP;
2556 case X86::JO_4: return X86::COND_O;
2557 case X86::JNO_4: return X86::COND_NO;
2558 }
2559 }
2561 /// getCondFromSETOpc - return condition code of a SET opcode.
2562 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2563 switch (Opc) {
2564 default: return X86::COND_INVALID;
2565 case X86::SETAr: case X86::SETAm: return X86::COND_A;
2566 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2567 case X86::SETBr: case X86::SETBm: return X86::COND_B;
2568 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2569 case X86::SETEr: case X86::SETEm: return X86::COND_E;
2570 case X86::SETGr: case X86::SETGm: return X86::COND_G;
2571 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2572 case X86::SETLr: case X86::SETLm: return X86::COND_L;
2573 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2574 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2575 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2576 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2577 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2578 case X86::SETOr: case X86::SETOm: return X86::COND_O;
2579 case X86::SETPr: case X86::SETPm: return X86::COND_P;
2580 case X86::SETSr: case X86::SETSm: return X86::COND_S;
2581 }
2582 }
2584 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2585 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2586 switch (Opc) {
2587 default: return X86::COND_INVALID;
2588 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm:
2589 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr:
2590 return X86::COND_A;
2591 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2592 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2593 return X86::COND_AE;
2594 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm:
2595 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr:
2596 return X86::COND_B;
2597 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2598 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2599 return X86::COND_BE;
2600 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm:
2601 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr:
2602 return X86::COND_E;
2603 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm:
2604 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr:
2605 return X86::COND_G;
2606 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2607 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2608 return X86::COND_GE;
2609 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm:
2610 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr:
2611 return X86::COND_L;
2612 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2613 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2614 return X86::COND_LE;
2615 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2616 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2617 return X86::COND_NE;
2618 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2619 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2620 return X86::COND_NO;
2621 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2622 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2623 return X86::COND_NP;
2624 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2625 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2626 return X86::COND_NS;
2627 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm:
2628 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr:
2629 return X86::COND_O;
2630 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm:
2631 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr:
2632 return X86::COND_P;
2633 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm:
2634 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr:
2635 return X86::COND_S;
2636 }
2637 }
2639 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2640 switch (CC) {
2641 default: llvm_unreachable("Illegal condition code!");
2642 case X86::COND_E: return X86::JE_4;
2643 case X86::COND_NE: return X86::JNE_4;
2644 case X86::COND_L: return X86::JL_4;
2645 case X86::COND_LE: return X86::JLE_4;
2646 case X86::COND_G: return X86::JG_4;
2647 case X86::COND_GE: return X86::JGE_4;
2648 case X86::COND_B: return X86::JB_4;
2649 case X86::COND_BE: return X86::JBE_4;
2650 case X86::COND_A: return X86::JA_4;
2651 case X86::COND_AE: return X86::JAE_4;
2652 case X86::COND_S: return X86::JS_4;
2653 case X86::COND_NS: return X86::JNS_4;
2654 case X86::COND_P: return X86::JP_4;
2655 case X86::COND_NP: return X86::JNP_4;
2656 case X86::COND_O: return X86::JO_4;
2657 case X86::COND_NO: return X86::JNO_4;
2658 }
2659 }
2661 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2662 /// e.g. turning COND_E to COND_NE.
2663 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2664 switch (CC) {
2665 default: llvm_unreachable("Illegal condition code!");
2666 case X86::COND_E: return X86::COND_NE;
2667 case X86::COND_NE: return X86::COND_E;
2668 case X86::COND_L: return X86::COND_GE;
2669 case X86::COND_LE: return X86::COND_G;
2670 case X86::COND_G: return X86::COND_LE;
2671 case X86::COND_GE: return X86::COND_L;
2672 case X86::COND_B: return X86::COND_AE;
2673 case X86::COND_BE: return X86::COND_A;
2674 case X86::COND_A: return X86::COND_BE;
2675 case X86::COND_AE: return X86::COND_B;
2676 case X86::COND_S: return X86::COND_NS;
2677 case X86::COND_NS: return X86::COND_S;
2678 case X86::COND_P: return X86::COND_NP;
2679 case X86::COND_NP: return X86::COND_P;
2680 case X86::COND_O: return X86::COND_NO;
2681 case X86::COND_NO: return X86::COND_O;
2682 }
2683 }
2685 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2686 /// the condition code if we modify the instructions such that flags are
2687 /// set by MI(b,a).
2688 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2689 switch (CC) {
2690 default: return X86::COND_INVALID;
2691 case X86::COND_E: return X86::COND_E;
2692 case X86::COND_NE: return X86::COND_NE;
2693 case X86::COND_L: return X86::COND_G;
2694 case X86::COND_LE: return X86::COND_GE;
2695 case X86::COND_G: return X86::COND_L;
2696 case X86::COND_GE: return X86::COND_LE;
2697 case X86::COND_B: return X86::COND_A;
2698 case X86::COND_BE: return X86::COND_AE;
2699 case X86::COND_A: return X86::COND_B;
2700 case X86::COND_AE: return X86::COND_BE;
2701 }
2702 }
2704 /// getSETFromCond - Return a set opcode for the given condition and
2705 /// whether it has memory operand.
2706 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
2707 static const uint16_t Opc[16][2] = {
2708 { X86::SETAr, X86::SETAm },
2709 { X86::SETAEr, X86::SETAEm },
2710 { X86::SETBr, X86::SETBm },
2711 { X86::SETBEr, X86::SETBEm },
2712 { X86::SETEr, X86::SETEm },
2713 { X86::SETGr, X86::SETGm },
2714 { X86::SETGEr, X86::SETGEm },
2715 { X86::SETLr, X86::SETLm },
2716 { X86::SETLEr, X86::SETLEm },
2717 { X86::SETNEr, X86::SETNEm },
2718 { X86::SETNOr, X86::SETNOm },
2719 { X86::SETNPr, X86::SETNPm },
2720 { X86::SETNSr, X86::SETNSm },
2721 { X86::SETOr, X86::SETOm },
2722 { X86::SETPr, X86::SETPm },
2723 { X86::SETSr, X86::SETSm }
2724 };
2726 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
2727 return Opc[CC][HasMemoryOperand ? 1 : 0];
2728 }
2730 /// getCMovFromCond - Return a cmov opcode for the given condition,
2731 /// register size in bytes, and operand type.
2732 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
2733 bool HasMemoryOperand) {
2734 static const uint16_t Opc[32][3] = {
2735 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr },
2736 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2737 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr },
2738 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2739 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr },
2740 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr },
2741 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2742 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr },
2743 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2744 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2745 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2746 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2747 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2748 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr },
2749 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr },
2750 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr },
2751 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm },
2752 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2753 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm },
2754 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2755 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm },
2756 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm },
2757 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2758 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm },
2759 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2760 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2761 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2762 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2763 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2764 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm },
2765 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm },
2766 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm }
2767 };
2769 assert(CC < 16 && "Can only handle standard cond codes");
2770 unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2771 switch(RegBytes) {
2772 default: llvm_unreachable("Illegal register size!");
2773 case 2: return Opc[Idx][0];
2774 case 4: return Opc[Idx][1];
2775 case 8: return Opc[Idx][2];
2776 }
2777 }
2779 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2780 if (!MI->isTerminator()) return false;
2782 // Conditional branch is a special case.
2783 if (MI->isBranch() && !MI->isBarrier())
2784 return true;
2785 if (!MI->isPredicable())
2786 return true;
2787 return !isPredicated(MI);
2788 }
2790 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2791 MachineBasicBlock *&TBB,
2792 MachineBasicBlock *&FBB,
2793 SmallVectorImpl<MachineOperand> &Cond,
2794 bool AllowModify) const {
2795 // Start from the bottom of the block and work up, examining the
2796 // terminator instructions.
2797 MachineBasicBlock::iterator I = MBB.end();
2798 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2799 while (I != MBB.begin()) {
2800 --I;
2801 if (I->isDebugValue())
2802 continue;
2804 // Working from the bottom, when we see a non-terminator instruction, we're
2805 // done.
2806 if (!isUnpredicatedTerminator(I))
2807 break;
2809 // A terminator that isn't a branch can't easily be handled by this
2810 // analysis.
2811 if (!I->isBranch())
2812 return true;
2814 // Handle unconditional branches.
2815 if (I->getOpcode() == X86::JMP_4) {
2816 UnCondBrIter = I;
2818 if (!AllowModify) {
2819 TBB = I->getOperand(0).getMBB();
2820 continue;
2821 }
2823 // If the block has any instructions after a JMP, delete them.
2824 while (std::next(I) != MBB.end())
2825 std::next(I)->eraseFromParent();
2827 Cond.clear();
2828 FBB = nullptr;
2830 // Delete the JMP if it's equivalent to a fall-through.
2831 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2832 TBB = nullptr;
2833 I->eraseFromParent();
2834 I = MBB.end();
2835 UnCondBrIter = MBB.end();
2836 continue;
2837 }
2839 // TBB is used to indicate the unconditional destination.
2840 TBB = I->getOperand(0).getMBB();
2841 continue;
2842 }
2844 // Handle conditional branches.
2845 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
2846 if (BranchCode == X86::COND_INVALID)
2847 return true; // Can't handle indirect branch.
2849 // Working from the bottom, handle the first conditional branch.
2850 if (Cond.empty()) {
2851 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2852 if (AllowModify && UnCondBrIter != MBB.end() &&
2853 MBB.isLayoutSuccessor(TargetBB)) {
2854 // If we can modify the code and it ends in something like:
2855 //
2856 // jCC L1
2857 // jmp L2
2858 // L1:
2859 // ...
2860 // L2:
2861 //
2862 // Then we can change this to:
2863 //
2864 // jnCC L2
2865 // L1:
2866 // ...
2867 // L2:
2868 //
2869 // Which is a bit more efficient.
2870 // We conditionally jump to the fall-through block.
2871 BranchCode = GetOppositeBranchCondition(BranchCode);
2872 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2873 MachineBasicBlock::iterator OldInst = I;
2875 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2876 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2877 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2878 .addMBB(TargetBB);
2880 OldInst->eraseFromParent();
2881 UnCondBrIter->eraseFromParent();
2883 // Restart the analysis.
2884 UnCondBrIter = MBB.end();
2885 I = MBB.end();
2886 continue;
2887 }
2889 FBB = TBB;
2890 TBB = I->getOperand(0).getMBB();
2891 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2892 continue;
2893 }
2895 // Handle subsequent conditional branches. Only handle the case where all
2896 // conditional branches branch to the same destination and their condition
2897 // opcodes fit one of the special multi-branch idioms.
2898 assert(Cond.size() == 1);
2899 assert(TBB);
2901 // Only handle the case where all conditional branches branch to the same
2902 // destination.
2903 if (TBB != I->getOperand(0).getMBB())
2904 return true;
2906 // If the conditions are the same, we can leave them alone.
2907 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2908 if (OldBranchCode == BranchCode)
2909 continue;
2911 // If they differ, see if they fit one of the known patterns. Theoretically,
2912 // we could handle more patterns here, but we shouldn't expect to see them
2913 // if instruction selection has done a reasonable job.
2914 if ((OldBranchCode == X86::COND_NP &&
2915 BranchCode == X86::COND_E) ||
2916 (OldBranchCode == X86::COND_E &&
2917 BranchCode == X86::COND_NP))
2918 BranchCode = X86::COND_NP_OR_E;
2919 else if ((OldBranchCode == X86::COND_P &&
2920 BranchCode == X86::COND_NE) ||
2921 (OldBranchCode == X86::COND_NE &&
2922 BranchCode == X86::COND_P))
2923 BranchCode = X86::COND_NE_OR_P;
2924 else
2925 return true;
2927 // Update the MachineOperand.
2928 Cond[0].setImm(BranchCode);
2929 }
2931 return false;
2932 }
2934 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2935 MachineBasicBlock::iterator I = MBB.end();
2936 unsigned Count = 0;
2938 while (I != MBB.begin()) {
2939 --I;
2940 if (I->isDebugValue())
2941 continue;
2942 if (I->getOpcode() != X86::JMP_4 &&
2943 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2944 break;
2945 // Remove the branch.
2946 I->eraseFromParent();
2947 I = MBB.end();
2948 ++Count;
2949 }
2951 return Count;
2952 }
2954 unsigned
2955 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2956 MachineBasicBlock *FBB,
2957 const SmallVectorImpl<MachineOperand> &Cond,
2958 DebugLoc DL) const {
2959 // Shouldn't be a fall through.
2960 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2961 assert((Cond.size() == 1 || Cond.size() == 0) &&
2962 "X86 branch conditions have one component!");
2964 if (Cond.empty()) {
2965 // Unconditional branch?
2966 assert(!FBB && "Unconditional branch with multiple successors!");
2967 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2968 return 1;
2969 }
2971 // Conditional branch.
2972 unsigned Count = 0;
2973 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2974 switch (CC) {
2975 case X86::COND_NP_OR_E:
2976 // Synthesize NP_OR_E with two branches.
2977 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2978 ++Count;
2979 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2980 ++Count;
2981 break;
2982 case X86::COND_NE_OR_P:
2983 // Synthesize NE_OR_P with two branches.
2984 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2985 ++Count;
2986 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2987 ++Count;
2988 break;
2989 default: {
2990 unsigned Opc = GetCondBranchFromCond(CC);
2991 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2992 ++Count;
2993 }
2994 }
2995 if (FBB) {
2996 // Two-way Conditional branch. Insert the second branch.
2997 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2998 ++Count;
2999 }
3000 return Count;
3001 }
3003 bool X86InstrInfo::
3004 canInsertSelect(const MachineBasicBlock &MBB,
3005 const SmallVectorImpl<MachineOperand> &Cond,
3006 unsigned TrueReg, unsigned FalseReg,
3007 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
3008 // Not all subtargets have cmov instructions.
3009 if (!Subtarget.hasCMov())
3010 return false;
3011 if (Cond.size() != 1)
3012 return false;
3013 // We cannot do the composite conditions, at least not in SSA form.
3014 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
3015 return false;
3017 // Check register classes.
3018 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3019 const TargetRegisterClass *RC =
3020 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
3021 if (!RC)
3022 return false;
3024 // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
3025 if (X86::GR16RegClass.hasSubClassEq(RC) ||
3026 X86::GR32RegClass.hasSubClassEq(RC) ||
3027 X86::GR64RegClass.hasSubClassEq(RC)) {
3028 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
3029 // Bridge. Probably Ivy Bridge as well.
3030 CondCycles = 2;
3031 TrueCycles = 2;
3032 FalseCycles = 2;
3033 return true;
3034 }
3036 // Can't do vectors.
3037 return false;
3038 }
3040 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3041 MachineBasicBlock::iterator I, DebugLoc DL,
3042 unsigned DstReg,
3043 const SmallVectorImpl<MachineOperand> &Cond,
3044 unsigned TrueReg, unsigned FalseReg) const {
3045 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3046 assert(Cond.size() == 1 && "Invalid Cond array");
3047 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
3048 MRI.getRegClass(DstReg)->getSize(),
3049 false/*HasMemoryOperand*/);
3050 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
3051 }
3053 /// isHReg - Test if the given register is a physical h register.
3054 static bool isHReg(unsigned Reg) {
3055 return X86::GR8_ABCD_HRegClass.contains(Reg);
3056 }
3058 // Try and copy between VR128/VR64 and GR64 registers.
3059 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3060 const X86Subtarget &Subtarget) {
3062 // SrcReg(VR128) -> DestReg(GR64)
3063 // SrcReg(VR64) -> DestReg(GR64)
3064 // SrcReg(GR64) -> DestReg(VR128)
3065 // SrcReg(GR64) -> DestReg(VR64)
3067 bool HasAVX = Subtarget.hasAVX();
3068 bool HasAVX512 = Subtarget.hasAVX512();
3069 if (X86::GR64RegClass.contains(DestReg)) {
3070 if (X86::VR128XRegClass.contains(SrcReg))
3071 // Copy from a VR128 register to a GR64 register.
3072 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
3073 X86::MOVPQIto64rr);
3074 if (X86::VR64RegClass.contains(SrcReg))
3075 // Copy from a VR64 register to a GR64 register.
3076 return X86::MOVSDto64rr;
3077 } else if (X86::GR64RegClass.contains(SrcReg)) {
3078 // Copy from a GR64 register to a VR128 register.
3079 if (X86::VR128XRegClass.contains(DestReg))
3080 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
3081 X86::MOV64toPQIrr);
3082 // Copy from a GR64 register to a VR64 register.
3083 if (X86::VR64RegClass.contains(DestReg))
3084 return X86::MOV64toSDrr;
3085 }
3087 // SrcReg(FR32) -> DestReg(GR32)
3088 // SrcReg(GR32) -> DestReg(FR32)
3090 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3091 // Copy from a FR32 register to a GR32 register.
3092 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3094 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3095 // Copy from a GR32 register to a FR32 register.
3096 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3097 return 0;
3098 }
3100 inline static bool MaskRegClassContains(unsigned Reg) {
3101 return X86::VK8RegClass.contains(Reg) ||
3102 X86::VK16RegClass.contains(Reg) ||
3103 X86::VK32RegClass.contains(Reg) ||
3104 X86::VK64RegClass.contains(Reg) ||
3105 X86::VK1RegClass.contains(Reg);
3106 }
3107 static
3108 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3109 if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3110 X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3111 X86::VR512RegClass.contains(DestReg, SrcReg)) {
3112 DestReg = get512BitSuperRegister(DestReg);
3113 SrcReg = get512BitSuperRegister(SrcReg);
3114 return X86::VMOVAPSZrr;
3115 }
3116 if (MaskRegClassContains(DestReg) &&
3117 MaskRegClassContains(SrcReg))
3118 return X86::KMOVWkk;
3119 if (MaskRegClassContains(DestReg) &&
3120 (X86::GR32RegClass.contains(SrcReg) ||
3121 X86::GR16RegClass.contains(SrcReg) ||
3122 X86::GR8RegClass.contains(SrcReg))) {
3123 SrcReg = getX86SubSuperRegister(SrcReg, MVT::i32);
3124 return X86::KMOVWkr;
3125 }
3126 if ((X86::GR32RegClass.contains(DestReg) ||
3127 X86::GR16RegClass.contains(DestReg) ||
3128 X86::GR8RegClass.contains(DestReg)) &&
3129 MaskRegClassContains(SrcReg)) {
3130 DestReg = getX86SubSuperRegister(DestReg, MVT::i32);
3131 return X86::KMOVWrk;
3132 }
3133 return 0;
3134 }
3136 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3137 MachineBasicBlock::iterator MI, DebugLoc DL,
3138 unsigned DestReg, unsigned SrcReg,
3139 bool KillSrc) const {
3140 // First deal with the normal symmetric copies.
3141 bool HasAVX = Subtarget.hasAVX();
3142 bool HasAVX512 = Subtarget.hasAVX512();
3143 unsigned Opc = 0;
3144 if (X86::GR64RegClass.contains(DestReg, SrcReg))
3145 Opc = X86::MOV64rr;
3146 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3147 Opc = X86::MOV32rr;
3148 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3149 Opc = X86::MOV16rr;
3150 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3151 // Copying to or from a physical H register on x86-64 requires a NOREX
3152 // move. Otherwise use a normal move.
3153 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3154 Subtarget.is64Bit()) {
3155 Opc = X86::MOV8rr_NOREX;
3156 // Both operands must be encodable without an REX prefix.
3157 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3158 "8-bit H register can not be copied outside GR8_NOREX");
3159 } else
3160 Opc = X86::MOV8rr;
3161 }
3162 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3163 Opc = X86::MMX_MOVQ64rr;
3164 else if (HasAVX512)
3165 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3166 else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3167 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3168 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3169 Opc = X86::VMOVAPSYrr;
3170 if (!Opc)
3171 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3173 if (Opc) {
3174 BuildMI(MBB, MI, DL, get(Opc), DestReg)
3175 .addReg(SrcReg, getKillRegState(KillSrc));
3176 return;
3177 }
3179 // Moving EFLAGS to / from another register requires a push and a pop.
3180 // Notice that we have to adjust the stack if we don't want to clobber the
3181 // first frame index. See X86FrameLowering.cpp - clobbersTheStack.
3182 if (SrcReg == X86::EFLAGS) {
3183 if (X86::GR64RegClass.contains(DestReg)) {
3184 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3185 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3186 return;
3187 }
3188 if (X86::GR32RegClass.contains(DestReg)) {
3189 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3190 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3191 return;
3192 }
3193 }
3194 if (DestReg == X86::EFLAGS) {
3195 if (X86::GR64RegClass.contains(SrcReg)) {
3196 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3197 .addReg(SrcReg, getKillRegState(KillSrc));
3198 BuildMI(MBB, MI, DL, get(X86::POPF64));
3199 return;
3200 }
3201 if (X86::GR32RegClass.contains(SrcReg)) {
3202 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3203 .addReg(SrcReg, getKillRegState(KillSrc));
3204 BuildMI(MBB, MI, DL, get(X86::POPF32));
3205 return;
3206 }
3207 }
3209 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3210 << " to " << RI.getName(DestReg) << '\n');
3211 llvm_unreachable("Cannot emit physreg copy instruction");
3212 }
3214 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3215 const TargetRegisterClass *RC,
3216 bool isStackAligned,
3217 const X86Subtarget &STI,
3218 bool load) {
3219 if (STI.hasAVX512()) {
3220 if (X86::VK8RegClass.hasSubClassEq(RC) ||
3221 X86::VK16RegClass.hasSubClassEq(RC))
3222 return load ? X86::KMOVWkm : X86::KMOVWmk;
3223 if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3224 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3225 if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3226 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3227 if (X86::VR512RegClass.hasSubClassEq(RC))
3228 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3229 }
3231 bool HasAVX = STI.hasAVX();
3232 switch (RC->getSize()) {
3233 default:
3234 llvm_unreachable("Unknown spill size");
3235 case 1:
3236 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3237 if (STI.is64Bit())
3238 // Copying to or from a physical H register on x86-64 requires a NOREX
3239 // move. Otherwise use a normal move.
3240 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3241 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3242 return load ? X86::MOV8rm : X86::MOV8mr;
3243 case 2:
3244 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3245 return load ? X86::MOV16rm : X86::MOV16mr;
3246 case 4:
3247 if (X86::GR32RegClass.hasSubClassEq(RC))
3248 return load ? X86::MOV32rm : X86::MOV32mr;
3249 if (X86::FR32RegClass.hasSubClassEq(RC))
3250 return load ?
3251 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3252 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3253 if (X86::RFP32RegClass.hasSubClassEq(RC))
3254 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3255 llvm_unreachable("Unknown 4-byte regclass");
3256 case 8:
3257 if (X86::GR64RegClass.hasSubClassEq(RC))
3258 return load ? X86::MOV64rm : X86::MOV64mr;
3259 if (X86::FR64RegClass.hasSubClassEq(RC))
3260 return load ?
3261 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3262 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3263 if (X86::VR64RegClass.hasSubClassEq(RC))
3264 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3265 if (X86::RFP64RegClass.hasSubClassEq(RC))
3266 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3267 llvm_unreachable("Unknown 8-byte regclass");
3268 case 10:
3269 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3270 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3271 case 16: {
3272 assert((X86::VR128RegClass.hasSubClassEq(RC) ||
3273 X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
3274 // If stack is realigned we can use aligned stores.
3275 if (isStackAligned)
3276 return load ?
3277 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3278 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
3279 else
3280 return load ?
3281 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3282 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3283 }
3284 case 32:
3285 assert((X86::VR256RegClass.hasSubClassEq(RC) ||
3286 X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
3287 // If stack is realigned we can use aligned stores.
3288 if (isStackAligned)
3289 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3290 else
3291 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
3292 case 64:
3293 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3294 if (isStackAligned)
3295 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3296 else
3297 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3298 }
3299 }
3301 static unsigned getStoreRegOpcode(unsigned SrcReg,
3302 const TargetRegisterClass *RC,
3303 bool isStackAligned,
3304 const X86Subtarget &STI) {
3305 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
3306 }
3309 static unsigned getLoadRegOpcode(unsigned DestReg,
3310 const TargetRegisterClass *RC,
3311 bool isStackAligned,
3312 const X86Subtarget &STI) {
3313 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
3314 }
3316 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3317 MachineBasicBlock::iterator MI,
3318 unsigned SrcReg, bool isKill, int FrameIdx,
3319 const TargetRegisterClass *RC,
3320 const TargetRegisterInfo *TRI) const {
3321 const MachineFunction &MF = *MBB.getParent();
3322 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3323 "Stack slot too small for store");
3324 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3325 bool isAligned = (MF.getTarget()
3326 .getSubtargetImpl()
3327 ->getFrameLowering()
3328 ->getStackAlignment() >= Alignment) ||
3329 RI.canRealignStack(MF);
3330 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3331 DebugLoc DL = MBB.findDebugLoc(MI);
3332 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
3333 .addReg(SrcReg, getKillRegState(isKill));
3334 }
3336 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3337 bool isKill,
3338 SmallVectorImpl<MachineOperand> &Addr,
3339 const TargetRegisterClass *RC,
3340 MachineInstr::mmo_iterator MMOBegin,
3341 MachineInstr::mmo_iterator MMOEnd,
3342 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3343 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3344 bool isAligned = MMOBegin != MMOEnd &&
3345 (*MMOBegin)->getAlignment() >= Alignment;
3346 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
3347 DebugLoc DL;
3348 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3349 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3350 MIB.addOperand(Addr[i]);
3351 MIB.addReg(SrcReg, getKillRegState(isKill));
3352 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3353 NewMIs.push_back(MIB);
3354 }
3357 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3358 MachineBasicBlock::iterator MI,
3359 unsigned DestReg, int FrameIdx,
3360 const TargetRegisterClass *RC,
3361 const TargetRegisterInfo *TRI) const {
3362 const MachineFunction &MF = *MBB.getParent();
3363 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3364 bool isAligned = (MF.getTarget()
3365 .getSubtargetImpl()
3366 ->getFrameLowering()
3367 ->getStackAlignment() >= Alignment) ||
3368 RI.canRealignStack(MF);
3369 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3370 DebugLoc DL = MBB.findDebugLoc(MI);
3371 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
3372 }
3374 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
3375 SmallVectorImpl<MachineOperand> &Addr,
3376 const TargetRegisterClass *RC,
3377 MachineInstr::mmo_iterator MMOBegin,
3378 MachineInstr::mmo_iterator MMOEnd,
3379 SmallVectorImpl<MachineInstr*> &NewMIs) const {
3380 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3381 bool isAligned = MMOBegin != MMOEnd &&
3382 (*MMOBegin)->getAlignment() >= Alignment;
3383 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
3384 DebugLoc DL;
3385 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3386 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3387 MIB.addOperand(Addr[i]);
3388 (*MIB).setMemRefs(MMOBegin, MMOEnd);
3389 NewMIs.push_back(MIB);
3390 }
3392 bool X86InstrInfo::
3393 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3394 int &CmpMask, int &CmpValue) const {
3395 switch (MI->getOpcode()) {
3396 default: break;
3397 case X86::CMP64ri32:
3398 case X86::CMP64ri8:
3399 case X86::CMP32ri:
3400 case X86::CMP32ri8:
3401 case X86::CMP16ri:
3402 case X86::CMP16ri8:
3403 case X86::CMP8ri:
3404 SrcReg = MI->getOperand(0).getReg();
3405 SrcReg2 = 0;
3406 CmpMask = ~0;
3407 CmpValue = MI->getOperand(1).getImm();
3408 return true;
3409 // A SUB can be used to perform comparison.
3410 case X86::SUB64rm:
3411 case X86::SUB32rm:
3412 case X86::SUB16rm:
3413 case X86::SUB8rm:
3414 SrcReg = MI->getOperand(1).getReg();
3415 SrcReg2 = 0;
3416 CmpMask = ~0;
3417 CmpValue = 0;
3418 return true;
3419 case X86::SUB64rr:
3420 case X86::SUB32rr:
3421 case X86::SUB16rr:
3422 case X86::SUB8rr:
3423 SrcReg = MI->getOperand(1).getReg();
3424 SrcReg2 = MI->getOperand(2).getReg();
3425 CmpMask = ~0;
3426 CmpValue = 0;
3427 return true;
3428 case X86::SUB64ri32:
3429 case X86::SUB64ri8:
3430 case X86::SUB32ri:
3431 case X86::SUB32ri8:
3432 case X86::SUB16ri:
3433 case X86::SUB16ri8:
3434 case X86::SUB8ri:
3435 SrcReg = MI->getOperand(1).getReg();
3436 SrcReg2 = 0;
3437 CmpMask = ~0;
3438 CmpValue = MI->getOperand(2).getImm();
3439 return true;
3440 case X86::CMP64rr:
3441 case X86::CMP32rr:
3442 case X86::CMP16rr:
3443 case X86::CMP8rr:
3444 SrcReg = MI->getOperand(0).getReg();
3445 SrcReg2 = MI->getOperand(1).getReg();
3446 CmpMask = ~0;
3447 CmpValue = 0;
3448 return true;
3449 case X86::TEST8rr:
3450 case X86::TEST16rr:
3451 case X86::TEST32rr:
3452 case X86::TEST64rr:
3453 SrcReg = MI->getOperand(0).getReg();
3454 if (MI->getOperand(1).getReg() != SrcReg) return false;
3455 // Compare against zero.
3456 SrcReg2 = 0;
3457 CmpMask = ~0;
3458 CmpValue = 0;
3459 return true;
3460 }
3461 return false;
3462 }
3464 /// isRedundantFlagInstr - check whether the first instruction, whose only
3465 /// purpose is to update flags, can be made redundant.
3466 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3467 /// This function can be extended later on.
3468 /// SrcReg, SrcRegs: register operands for FlagI.
3469 /// ImmValue: immediate for FlagI if it takes an immediate.
3470 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3471 unsigned SrcReg2, int ImmValue,
3472 MachineInstr *OI) {
3473 if (((FlagI->getOpcode() == X86::CMP64rr &&
3474 OI->getOpcode() == X86::SUB64rr) ||
3475 (FlagI->getOpcode() == X86::CMP32rr &&
3476 OI->getOpcode() == X86::SUB32rr)||
3477 (FlagI->getOpcode() == X86::CMP16rr &&
3478 OI->getOpcode() == X86::SUB16rr)||
3479 (FlagI->getOpcode() == X86::CMP8rr &&
3480 OI->getOpcode() == X86::SUB8rr)) &&
3481 ((OI->getOperand(1).getReg() == SrcReg &&
3482 OI->getOperand(2).getReg() == SrcReg2) ||
3483 (OI->getOperand(1).getReg() == SrcReg2 &&
3484 OI->getOperand(2).getReg() == SrcReg)))
3485 return true;
3487 if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3488 OI->getOpcode() == X86::SUB64ri32) ||
3489 (FlagI->getOpcode() == X86::CMP64ri8 &&
3490 OI->getOpcode() == X86::SUB64ri8) ||
3491 (FlagI->getOpcode() == X86::CMP32ri &&
3492 OI->getOpcode() == X86::SUB32ri) ||
3493 (FlagI->getOpcode() == X86::CMP32ri8 &&
3494 OI->getOpcode() == X86::SUB32ri8) ||
3495 (FlagI->getOpcode() == X86::CMP16ri &&
3496 OI->getOpcode() == X86::SUB16ri) ||
3497 (FlagI->getOpcode() == X86::CMP16ri8 &&
3498 OI->getOpcode() == X86::SUB16ri8) ||
3499 (FlagI->getOpcode() == X86::CMP8ri &&
3500 OI->getOpcode() == X86::SUB8ri)) &&
3501 OI->getOperand(1).getReg() == SrcReg &&
3502 OI->getOperand(2).getImm() == ImmValue)
3503 return true;
3504 return false;
3505 }
3507 /// isDefConvertible - check whether the definition can be converted
3508 /// to remove a comparison against zero.
3509 inline static bool isDefConvertible(MachineInstr *MI) {
3510 switch (MI->getOpcode()) {
3511 default: return false;
3513 // The shift instructions only modify ZF if their shift count is non-zero.
3514 // N.B.: The processor truncates the shift count depending on the encoding.
3515 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri:
3516 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri:
3517 return getTruncatedShiftCount(MI, 2) != 0;
3519 // Some left shift instructions can be turned into LEA instructions but only
3520 // if their flags aren't used. Avoid transforming such instructions.
3521 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{
3522 unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3523 if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3524 return ShAmt != 0;
3525 }
3527 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3528 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3529 return getTruncatedShiftCount(MI, 3) != 0;
3531 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3532 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8:
3533 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr:
3534 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm:
3535 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm:
3536 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r:
3537 case X86::DEC64_32r: case X86::DEC64_16r:
3538 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3539 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8:
3540 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr:
3541 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm:
3542 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm:
3543 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r:
3544 case X86::INC64_32r: case X86::INC64_16r:
3545 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3546 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8:
3547 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr:
3548 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm:
3549 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm:
3550 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3551 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8:
3552 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr:
3553 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm:
3554 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm:
3555 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri:
3556 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8:
3557 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr:
3558 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm:
3559 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm:
3560 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r:
3561 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1:
3562 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1:
3563 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1:
3564 case X86::ADC32ri: case X86::ADC32ri8:
3565 case X86::ADC32rr: case X86::ADC64ri32:
3566 case X86::ADC64ri8: case X86::ADC64rr:
3567 case X86::SBB32ri: case X86::SBB32ri8:
3568 case X86::SBB32rr: case X86::SBB64ri32:
3569 case X86::SBB64ri8: case X86::SBB64rr:
3570 case X86::ANDN32rr: case X86::ANDN32rm:
3571 case X86::ANDN64rr: case X86::ANDN64rm:
3572 case X86::BEXTR32rr: case X86::BEXTR64rr:
3573 case X86::BEXTR32rm: case X86::BEXTR64rm:
3574 case X86::BLSI32rr: case X86::BLSI32rm:
3575 case X86::BLSI64rr: case X86::BLSI64rm:
3576 case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3577 case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3578 case X86::BLSR32rr: case X86::BLSR32rm:
3579 case X86::BLSR64rr: case X86::BLSR64rm:
3580 case X86::BZHI32rr: case X86::BZHI32rm:
3581 case X86::BZHI64rr: case X86::BZHI64rm:
3582 case X86::LZCNT16rr: case X86::LZCNT16rm:
3583 case X86::LZCNT32rr: case X86::LZCNT32rm:
3584 case X86::LZCNT64rr: case X86::LZCNT64rm:
3585 case X86::POPCNT16rr:case X86::POPCNT16rm:
3586 case X86::POPCNT32rr:case X86::POPCNT32rm:
3587 case X86::POPCNT64rr:case X86::POPCNT64rm:
3588 case X86::TZCNT16rr: case X86::TZCNT16rm:
3589 case X86::TZCNT32rr: case X86::TZCNT32rm:
3590 case X86::TZCNT64rr: case X86::TZCNT64rm:
3591 return true;
3592 }
3593 }
3595 /// isUseDefConvertible - check whether the use can be converted
3596 /// to remove a comparison against zero.
3597 static X86::CondCode isUseDefConvertible(MachineInstr *MI) {
3598 switch (MI->getOpcode()) {
3599 default: return X86::COND_INVALID;
3600 case X86::LZCNT16rr: case X86::LZCNT16rm:
3601 case X86::LZCNT32rr: case X86::LZCNT32rm:
3602 case X86::LZCNT64rr: case X86::LZCNT64rm:
3603 return X86::COND_B;
3604 case X86::POPCNT16rr:case X86::POPCNT16rm:
3605 case X86::POPCNT32rr:case X86::POPCNT32rm:
3606 case X86::POPCNT64rr:case X86::POPCNT64rm:
3607 return X86::COND_E;
3608 case X86::TZCNT16rr: case X86::TZCNT16rm:
3609 case X86::TZCNT32rr: case X86::TZCNT32rm:
3610 case X86::TZCNT64rr: case X86::TZCNT64rm:
3611 return X86::COND_B;
3612 }
3613 }
3615 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3616 /// operates on the same source operands and sets flags in the same way as
3617 /// Compare; remove Compare if possible.
3618 bool X86InstrInfo::
3619 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3620 int CmpMask, int CmpValue,
3621 const MachineRegisterInfo *MRI) const {
3622 // Check whether we can replace SUB with CMP.
3623 unsigned NewOpcode = 0;
3624 switch (CmpInstr->getOpcode()) {
3625 default: break;
3626 case X86::SUB64ri32:
3627 case X86::SUB64ri8:
3628 case X86::SUB32ri:
3629 case X86::SUB32ri8:
3630 case X86::SUB16ri:
3631 case X86::SUB16ri8:
3632 case X86::SUB8ri:
3633 case X86::SUB64rm:
3634 case X86::SUB32rm:
3635 case X86::SUB16rm:
3636 case X86::SUB8rm:
3637 case X86::SUB64rr:
3638 case X86::SUB32rr:
3639 case X86::SUB16rr:
3640 case X86::SUB8rr: {
3641 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3642 return false;
3643 // There is no use of the destination register, we can replace SUB with CMP.
3644 switch (CmpInstr->getOpcode()) {
3645 default: llvm_unreachable("Unreachable!");
3646 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break;
3647 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break;
3648 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break;
3649 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break;
3650 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break;
3651 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break;
3652 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break;
3653 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break;
3654 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3655 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break;
3656 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break;
3657 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break;
3658 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break;
3659 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break;
3660 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break;
3661 }
3662 CmpInstr->setDesc(get(NewOpcode));
3663 CmpInstr->RemoveOperand(0);
3664 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3665 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3666 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3667 return false;
3668 }
3669 }
3671 // Get the unique definition of SrcReg.
3672 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3673 if (!MI) return false;
3675 // CmpInstr is the first instruction of the BB.
3676 MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3678 // If we are comparing against zero, check whether we can use MI to update
3679 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3680 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3681 if (IsCmpZero && MI->getParent() != CmpInstr->getParent())
3682 return false;
3684 // If we have a use of the source register between the def and our compare
3685 // instruction we can eliminate the compare iff the use sets EFLAGS in the
3686 // right way.
3687 bool ShouldUpdateCC = false;
3688 X86::CondCode NewCC = X86::COND_INVALID;
3689 if (IsCmpZero && !isDefConvertible(MI)) {
3690 // Scan forward from the use until we hit the use we're looking for or the
3691 // compare instruction.
3692 for (MachineBasicBlock::iterator J = MI;; ++J) {
3693 // Do we have a convertible instruction?
3694 NewCC = isUseDefConvertible(J);
3695 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
3696 J->getOperand(1).getReg() == SrcReg) {
3697 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
3698 ShouldUpdateCC = true; // Update CC later on.
3699 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
3700 // with the new def.
3701 MI = Def = J;
3702 break;
3703 }
3705 if (J == I)
3706 return false;
3707 }
3708 }
3710 // We are searching for an earlier instruction that can make CmpInstr
3711 // redundant and that instruction will be saved in Sub.
3712 MachineInstr *Sub = nullptr;
3713 const TargetRegisterInfo *TRI = &getRegisterInfo();
3715 // We iterate backward, starting from the instruction before CmpInstr and
3716 // stop when reaching the definition of a source register or done with the BB.
3717 // RI points to the instruction before CmpInstr.
3718 // If the definition is in this basic block, RE points to the definition;
3719 // otherwise, RE is the rend of the basic block.
3720 MachineBasicBlock::reverse_iterator
3721 RI = MachineBasicBlock::reverse_iterator(I),
3722 RE = CmpInstr->getParent() == MI->getParent() ?
3723 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3724 CmpInstr->getParent()->rend();
3725 MachineInstr *Movr0Inst = nullptr;
3726 for (; RI != RE; ++RI) {
3727 MachineInstr *Instr = &*RI;
3728 // Check whether CmpInstr can be made redundant by the current instruction.
3729 if (!IsCmpZero &&
3730 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3731 Sub = Instr;
3732 break;
3733 }
3735 if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3736 Instr->readsRegister(X86::EFLAGS, TRI)) {
3737 // This instruction modifies or uses EFLAGS.
3739 // MOV32r0 etc. are implemented with xor which clobbers condition code.
3740 // They are safe to move up, if the definition to EFLAGS is dead and
3741 // earlier instructions do not read or write EFLAGS.
3742 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
3743 Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3744 Movr0Inst = Instr;
3745 continue;
3746 }
3748 // We can't remove CmpInstr.
3749 return false;
3750 }
3751 }
3753 // Return false if no candidates exist.
3754 if (!IsCmpZero && !Sub)
3755 return false;
3757 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3758 Sub->getOperand(2).getReg() == SrcReg);
3760 // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3761 // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3762 // If we are done with the basic block, we need to check whether EFLAGS is
3763 // live-out.
3764 bool IsSafe = false;
3765 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3766 MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3767 for (++I; I != E; ++I) {
3768 const MachineInstr &Instr = *I;
3769 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3770 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3771 // We should check the usage if this instruction uses and updates EFLAGS.
3772 if (!UseEFLAGS && ModifyEFLAGS) {
3773 // It is safe to remove CmpInstr if EFLAGS is updated again.
3774 IsSafe = true;
3775 break;
3776 }
3777 if (!UseEFLAGS && !ModifyEFLAGS)
3778 continue;
3780 // EFLAGS is used by this instruction.
3781 X86::CondCode OldCC = X86::COND_INVALID;
3782 bool OpcIsSET = false;
3783 if (IsCmpZero || IsSwapped) {
3784 // We decode the condition code from opcode.
3785 if (Instr.isBranch())
3786 OldCC = getCondFromBranchOpc(Instr.getOpcode());
3787 else {
3788 OldCC = getCondFromSETOpc(Instr.getOpcode());
3789 if (OldCC != X86::COND_INVALID)
3790 OpcIsSET = true;
3791 else
3792 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3793 }
3794 if (OldCC == X86::COND_INVALID) return false;
3795 }
3796 if (IsCmpZero) {
3797 switch (OldCC) {
3798 default: break;
3799 case X86::COND_A: case X86::COND_AE:
3800 case X86::COND_B: case X86::COND_BE:
3801 case X86::COND_G: case X86::COND_GE:
3802 case X86::COND_L: case X86::COND_LE:
3803 case X86::COND_O: case X86::COND_NO:
3804 // CF and OF are used, we can't perform this optimization.
3805 return false;
3806 }
3808 // If we're updating the condition code check if we have to reverse the
3809 // condition.
3810 if (ShouldUpdateCC)
3811 switch (OldCC) {
3812 default:
3813 return false;
3814 case X86::COND_E:
3815 break;
3816 case X86::COND_NE:
3817 NewCC = GetOppositeBranchCondition(NewCC);
3818 break;
3819 }
3820 } else if (IsSwapped) {
3821 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3822 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3823 // We swap the condition code and synthesize the new opcode.
3824 NewCC = getSwappedCondition(OldCC);
3825 if (NewCC == X86::COND_INVALID) return false;
3826 }
3828 if ((ShouldUpdateCC || IsSwapped) && NewCC != OldCC) {
3829 // Synthesize the new opcode.
3830 bool HasMemoryOperand = Instr.hasOneMemOperand();
3831 unsigned NewOpc;
3832 if (Instr.isBranch())
3833 NewOpc = GetCondBranchFromCond(NewCC);
3834 else if(OpcIsSET)
3835 NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3836 else {
3837 unsigned DstReg = Instr.getOperand(0).getReg();
3838 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3839 HasMemoryOperand);
3840 }
3842 // Push the MachineInstr to OpsToUpdate.
3843 // If it is safe to remove CmpInstr, the condition code of these
3844 // instructions will be modified.
3845 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3846 }
3847 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3848 // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3849 IsSafe = true;
3850 break;
3851 }
3852 }
3854 // If EFLAGS is not killed nor re-defined, we should check whether it is
3855 // live-out. If it is live-out, do not optimize.
3856 if ((IsCmpZero || IsSwapped) && !IsSafe) {
3857 MachineBasicBlock *MBB = CmpInstr->getParent();
3858 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3859 SE = MBB->succ_end(); SI != SE; ++SI)
3860 if ((*SI)->isLiveIn(X86::EFLAGS))
3861 return false;
3862 }
3864 // The instruction to be updated is either Sub or MI.
3865 Sub = IsCmpZero ? MI : Sub;
3866 // Move Movr0Inst to the appropriate place before Sub.
3867 if (Movr0Inst) {
3868 // Look backwards until we find a def that doesn't use the current EFLAGS.
3869 Def = Sub;
3870 MachineBasicBlock::reverse_iterator
3871 InsertI = MachineBasicBlock::reverse_iterator(++Def),
3872 InsertE = Sub->getParent()->rend();
3873 for (; InsertI != InsertE; ++InsertI) {
3874 MachineInstr *Instr = &*InsertI;
3875 if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3876 Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3877 Sub->getParent()->remove(Movr0Inst);
3878 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3879 Movr0Inst);
3880 break;
3881 }
3882 }
3883 if (InsertI == InsertE)
3884 return false;
3885 }
3887 // Make sure Sub instruction defines EFLAGS and mark the def live.
3888 unsigned i = 0, e = Sub->getNumOperands();
3889 for (; i != e; ++i) {
3890 MachineOperand &MO = Sub->getOperand(i);
3891 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3892 MO.setIsDead(false);
3893 break;
3894 }
3895 }
3896 assert(i != e && "Unable to locate a def EFLAGS operand");
3898 CmpInstr->eraseFromParent();
3900 // Modify the condition code of instructions in OpsToUpdate.
3901 for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3902 OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3903 return true;
3904 }
3906 /// optimizeLoadInstr - Try to remove the load by folding it to a register
3907 /// operand at the use. We fold the load instructions if load defines a virtual
3908 /// register, the virtual register is used once in the same BB, and the
3909 /// instructions in-between do not load or store, and have no side effects.
3910 MachineInstr* X86InstrInfo::
3911 optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3912 unsigned &FoldAsLoadDefReg,
3913 MachineInstr *&DefMI) const {
3914 if (FoldAsLoadDefReg == 0)
3915 return nullptr;
3916 // To be conservative, if there exists another load, clear the load candidate.
3917 if (MI->mayLoad()) {
3918 FoldAsLoadDefReg = 0;
3919 return nullptr;
3920 }
3922 // Check whether we can move DefMI here.
3923 DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3924 assert(DefMI);
3925 bool SawStore = false;
3926 if (!DefMI->isSafeToMove(this, nullptr, SawStore))
3927 return nullptr;
3929 // Collect information about virtual register operands of MI.
3930 unsigned SrcOperandId = 0;
3931 bool FoundSrcOperand = false;
3932 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3933 MachineOperand &MO = MI->getOperand(i);
3934 if (!MO.isReg())
3935 continue;
3936 unsigned Reg = MO.getReg();
3937 if (Reg != FoldAsLoadDefReg)
3938 continue;
3939 // Do not fold if we have a subreg use or a def or multiple uses.
3940 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3941 return nullptr;
3943 SrcOperandId = i;
3944 FoundSrcOperand = true;
3945 }
3946 if (!FoundSrcOperand) return nullptr;
3948 // Check whether we can fold the def into SrcOperandId.
3949 SmallVector<unsigned, 8> Ops;
3950 Ops.push_back(SrcOperandId);
3951 MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3952 if (FoldMI) {
3953 FoldAsLoadDefReg = 0;
3954 return FoldMI;
3955 }
3957 return nullptr;
3958 }
3960 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3961 /// instruction with two undef reads of the register being defined. This is
3962 /// used for mapping:
3963 /// %xmm4 = V_SET0
3964 /// to:
3965 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3966 ///
3967 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3968 const MCInstrDesc &Desc) {
3969 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3970 unsigned Reg = MIB->getOperand(0).getReg();
3971 MIB->setDesc(Desc);
3973 // MachineInstr::addOperand() will insert explicit operands before any
3974 // implicit operands.
3975 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3976 // But we don't trust that.
3977 assert(MIB->getOperand(1).getReg() == Reg &&
3978 MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3979 return true;
3980 }
3982 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
3983 // code sequence is needed for other targets.
3984 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
3985 const TargetInstrInfo &TII) {
3986 MachineBasicBlock &MBB = *MIB->getParent();
3987 DebugLoc DL = MIB->getDebugLoc();
3988 unsigned Reg = MIB->getOperand(0).getReg();
3989 const GlobalValue *GV =
3990 cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
3991 unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
3992 MachineMemOperand *MMO = MBB.getParent()->
3993 getMachineMemOperand(MachinePointerInfo::getGOT(), Flag, 8, 8);
3994 MachineBasicBlock::iterator I = MIB;
3996 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
3997 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
3998 .addMemOperand(MMO);
3999 MIB->setDebugLoc(DL);
4000 MIB->setDesc(TII.get(X86::MOV64rm));
4001 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4002 }
4004 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
4005 bool HasAVX = Subtarget.hasAVX();
4006 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
4007 switch (MI->getOpcode()) {
4008 case X86::MOV32r0:
4009 return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4010 case X86::SETB_C8r:
4011 return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4012 case X86::SETB_C16r:
4013 return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4014 case X86::SETB_C32r:
4015 return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4016 case X86::SETB_C64r:
4017 return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4018 case X86::V_SET0:
4019 case X86::FsFLD0SS:
4020 case X86::FsFLD0SD:
4021 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4022 case X86::AVX_SET0:
4023 assert(HasAVX && "AVX not supported");
4024 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
4025 case X86::AVX512_512_SET0:
4026 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4027 case X86::V_SETALLONES:
4028 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4029 case X86::AVX2_SETALLONES:
4030 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4031 case X86::TEST8ri_NOREX:
4032 MI->setDesc(get(X86::TEST8ri));
4033 return true;
4034 case X86::KSET0B:
4035 case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
4036 case X86::KSET1B:
4037 case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
4038 case TargetOpcode::LOAD_STACK_GUARD:
4039 expandLoadStackGuard(MIB, *this);
4040 return true;
4041 }
4042 return false;
4043 }
4045 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4046 const SmallVectorImpl<MachineOperand> &MOs,
4047 MachineInstr *MI,
4048 const TargetInstrInfo &TII) {
4049 // Create the base instruction with the memory operand as the first part.
4050 // Omit the implicit operands, something BuildMI can't do.
4051 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4052 MI->getDebugLoc(), true);
4053 MachineInstrBuilder MIB(MF, NewMI);
4054 unsigned NumAddrOps = MOs.size();
4055 for (unsigned i = 0; i != NumAddrOps; ++i)
4056 MIB.addOperand(MOs[i]);
4057 if (NumAddrOps < 4) // FrameIndex only
4058 addOffset(MIB, 0);
4060 // Loop over the rest of the ri operands, converting them over.
4061 unsigned NumOps = MI->getDesc().getNumOperands()-2;
4062 for (unsigned i = 0; i != NumOps; ++i) {
4063 MachineOperand &MO = MI->getOperand(i+2);
4064 MIB.addOperand(MO);
4065 }
4066 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
4067 MachineOperand &MO = MI->getOperand(i);
4068 MIB.addOperand(MO);
4069 }
4070 return MIB;
4071 }
4073 static MachineInstr *FuseInst(MachineFunction &MF,
4074 unsigned Opcode, unsigned OpNo,
4075 const SmallVectorImpl<MachineOperand> &MOs,
4076 MachineInstr *MI, const TargetInstrInfo &TII) {
4077 // Omit the implicit operands, something BuildMI can't do.
4078 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4079 MI->getDebugLoc(), true);
4080 MachineInstrBuilder MIB(MF, NewMI);
4082 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4083 MachineOperand &MO = MI->getOperand(i);
4084 if (i == OpNo) {
4085 assert(MO.isReg() && "Expected to fold into reg operand!");
4086 unsigned NumAddrOps = MOs.size();
4087 for (unsigned i = 0; i != NumAddrOps; ++i)
4088 MIB.addOperand(MOs[i]);
4089 if (NumAddrOps < 4) // FrameIndex only
4090 addOffset(MIB, 0);
4091 } else {
4092 MIB.addOperand(MO);
4093 }
4094 }
4095 return MIB;
4096 }
4098 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4099 const SmallVectorImpl<MachineOperand> &MOs,
4100 MachineInstr *MI) {
4101 MachineFunction &MF = *MI->getParent()->getParent();
4102 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
4104 unsigned NumAddrOps = MOs.size();
4105 for (unsigned i = 0; i != NumAddrOps; ++i)
4106 MIB.addOperand(MOs[i]);
4107 if (NumAddrOps < 4) // FrameIndex only
4108 addOffset(MIB, 0);
4109 return MIB.addImm(0);
4110 }
4112 MachineInstr*
4113 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4114 MachineInstr *MI, unsigned i,
4115 const SmallVectorImpl<MachineOperand> &MOs,
4116 unsigned Size, unsigned Align, bool AllowCommute) const {
4117 const DenseMap<unsigned,
4118 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4119 bool isCallRegIndirect = Subtarget.callRegIndirect();
4120 bool isTwoAddrFold = false;
4122 // Atom favors register form of call. So, we do not fold loads into calls
4123 // when X86Subtarget is Atom.
4124 if (isCallRegIndirect &&
4125 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
4126 return nullptr;
4127 }
4129 unsigned NumOps = MI->getDesc().getNumOperands();
4130 bool isTwoAddr = NumOps > 1 &&
4131 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4133 // FIXME: AsmPrinter doesn't know how to handle
4134 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4135 if (MI->getOpcode() == X86::ADD32ri &&
4136 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4137 return nullptr;
4139 MachineInstr *NewMI = nullptr;
4140 // Folding a memory location into the two-address part of a two-address
4141 // instruction is different than folding it other places. It requires
4142 // replacing the *two* registers with the memory location.
4143 if (isTwoAddr && NumOps >= 2 && i < 2 &&
4144 MI->getOperand(0).isReg() &&
4145 MI->getOperand(1).isReg() &&
4146 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
4147 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4148 isTwoAddrFold = true;
4149 } else if (i == 0) { // If operand 0
4150 if (MI->getOpcode() == X86::MOV32r0) {
4151 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
4152 if (NewMI)
4153 return NewMI;
4154 }
4156 OpcodeTablePtr = &RegOp2MemOpTable0;
4157 } else if (i == 1) {
4158 OpcodeTablePtr = &RegOp2MemOpTable1;
4159 } else if (i == 2) {
4160 OpcodeTablePtr = &RegOp2MemOpTable2;
4161 } else if (i == 3) {
4162 OpcodeTablePtr = &RegOp2MemOpTable3;
4163 }
4165 // If table selected...
4166 if (OpcodeTablePtr) {
4167 // Find the Opcode to fuse
4168 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4169 OpcodeTablePtr->find(MI->getOpcode());
4170 if (I != OpcodeTablePtr->end()) {
4171 unsigned Opcode = I->second.first;
4172 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4173 if (Align < MinAlign)
4174 return nullptr;
4175 bool NarrowToMOV32rm = false;
4176 if (Size) {
4177 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
4178 if (Size < RCSize) {
4179 // Check if it's safe to fold the load. If the size of the object is
4180 // narrower than the load width, then it's not.
4181 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4182 return nullptr;
4183 // If this is a 64-bit load, but the spill slot is 32, then we can do
4184 // a 32-bit load which is implicitly zero-extended. This likely is due
4185 // to liveintervalanalysis remat'ing a load from stack slot.
4186 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4187 return nullptr;
4188 Opcode = X86::MOV32rm;
4189 NarrowToMOV32rm = true;
4190 }
4191 }
4193 if (isTwoAddrFold)
4194 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
4195 else
4196 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
4198 if (NarrowToMOV32rm) {
4199 // If this is the special case where we use a MOV32rm to load a 32-bit
4200 // value and zero-extend the top bits. Change the destination register
4201 // to a 32-bit one.
4202 unsigned DstReg = NewMI->getOperand(0).getReg();
4203 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4204 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
4205 X86::sub_32bit));
4206 else
4207 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4208 }
4209 return NewMI;
4210 }
4211 }
4213 // If the instruction and target operand are commutable, commute the instruction and try again.
4214 if (AllowCommute) {
4215 unsigned OriginalOpIdx = i, CommuteOpIdx1, CommuteOpIdx2;
4216 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4217 if ((CommuteOpIdx1 == OriginalOpIdx) || (CommuteOpIdx2 == OriginalOpIdx)) {
4218 MachineInstr* CommutedMI = commuteInstruction(MI, false);
4219 if (!CommutedMI) {
4220 // Unable to commute.
4221 return nullptr;
4222 }
4223 if (CommutedMI != MI) {
4224 // New instruction. We can't fold from this.
4225 CommutedMI->eraseFromParent();
4226 return nullptr;
4227 }
4229 // Attempt to fold with the commuted version of the instruction.
4230 unsigned CommuteOpIdx = (CommuteOpIdx1 == OriginalOpIdx ? CommuteOpIdx2 : CommuteOpIdx1);
4231 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx, MOs, Size, Align, /*AllowCommute=*/ false);
4232 if (NewMI)
4233 return NewMI;
4235 // Folding failed again - undo the commute before returning.
4236 MachineInstr* UncommutedMI = commuteInstruction(MI, false);
4237 if (!UncommutedMI) {
4238 // Unable to commute.
4239 return nullptr;
4240 }
4241 if (UncommutedMI != MI) {
4242 // New instruction. It doesn't need to be kept.
4243 UncommutedMI->eraseFromParent();
4244 return nullptr;
4245 }
4247 // Return here to prevent duplicate fuse failure report.
4248 return nullptr;
4249 }
4250 }
4251 }
4253 // No fusion
4254 if (PrintFailedFusing && !MI->isCopy())
4255 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
4256 return nullptr;
4257 }
4259 /// hasPartialRegUpdate - Return true for all instructions that only update
4260 /// the first 32 or 64-bits of the destination register and leave the rest
4261 /// unmodified. This can be used to avoid folding loads if the instructions
4262 /// only update part of the destination register, and the non-updated part is
4263 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4264 /// instructions breaks the partial register dependency and it can improve
4265 /// performance. e.g.:
4266 ///
4267 /// movss (%rdi), %xmm0
4268 /// cvtss2sd %xmm0, %xmm0
4269 ///
4270 /// Instead of
4271 /// cvtss2sd (%rdi), %xmm0
4272 ///
4273 /// FIXME: This should be turned into a TSFlags.
4274 ///
4275 static bool hasPartialRegUpdate(unsigned Opcode) {
4276 switch (Opcode) {
4277 case X86::CVTSI2SSrr:
4278 case X86::CVTSI2SS64rr:
4279 case X86::CVTSI2SDrr:
4280 case X86::CVTSI2SD64rr:
4281 case X86::CVTSD2SSrr:
4282 case X86::Int_CVTSD2SSrr:
4283 case X86::CVTSS2SDrr:
4284 case X86::Int_CVTSS2SDrr:
4285 case X86::RCPSSr:
4286 case X86::RCPSSr_Int:
4287 case X86::ROUNDSDr:
4288 case X86::ROUNDSDr_Int:
4289 case X86::ROUNDSSr:
4290 case X86::ROUNDSSr_Int:
4291 case X86::RSQRTSSr:
4292 case X86::RSQRTSSr_Int:
4293 case X86::SQRTSSr:
4294 case X86::SQRTSSr_Int:
4295 return true;
4296 }
4298 return false;
4299 }
4301 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4302 /// instructions we would like before a partial register update.
4303 unsigned X86InstrInfo::
4304 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4305 const TargetRegisterInfo *TRI) const {
4306 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4307 return 0;
4309 // If MI is marked as reading Reg, the partial register update is wanted.
4310 const MachineOperand &MO = MI->getOperand(0);
4311 unsigned Reg = MO.getReg();
4312 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4313 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4314 return 0;
4315 } else {
4316 if (MI->readsRegister(Reg, TRI))
4317 return 0;
4318 }
4320 // If any of the preceding 16 instructions are reading Reg, insert a
4321 // dependency breaking instruction. The magic number is based on a few
4322 // Nehalem experiments.
4323 return 16;
4324 }
4326 // Return true for any instruction the copies the high bits of the first source
4327 // operand into the unused high bits of the destination operand.
4328 static bool hasUndefRegUpdate(unsigned Opcode) {
4329 switch (Opcode) {
4330 case X86::VCVTSI2SSrr:
4331 case X86::Int_VCVTSI2SSrr:
4332 case X86::VCVTSI2SS64rr:
4333 case X86::Int_VCVTSI2SS64rr:
4334 case X86::VCVTSI2SDrr:
4335 case X86::Int_VCVTSI2SDrr:
4336 case X86::VCVTSI2SD64rr:
4337 case X86::Int_VCVTSI2SD64rr:
4338 case X86::VCVTSD2SSrr:
4339 case X86::Int_VCVTSD2SSrr:
4340 case X86::VCVTSS2SDrr:
4341 case X86::Int_VCVTSS2SDrr:
4342 case X86::VRCPSSr:
4343 case X86::VROUNDSDr:
4344 case X86::VROUNDSDr_Int:
4345 case X86::VROUNDSSr:
4346 case X86::VROUNDSSr_Int:
4347 case X86::VRSQRTSSr:
4348 case X86::VSQRTSSr:
4350 // AVX-512
4351 case X86::VCVTSD2SSZrr:
4352 case X86::VCVTSS2SDZrr:
4353 return true;
4354 }
4356 return false;
4357 }
4359 /// Inform the ExeDepsFix pass how many idle instructions we would like before
4360 /// certain undef register reads.
4361 ///
4362 /// This catches the VCVTSI2SD family of instructions:
4363 ///
4364 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
4365 ///
4366 /// We should to be careful *not* to catch VXOR idioms which are presumably
4367 /// handled specially in the pipeline:
4368 ///
4369 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
4370 ///
4371 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
4372 /// high bits that are passed-through are not live.
4373 unsigned X86InstrInfo::
4374 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
4375 const TargetRegisterInfo *TRI) const {
4376 if (!hasUndefRegUpdate(MI->getOpcode()))
4377 return 0;
4379 // Set the OpNum parameter to the first source operand.
4380 OpNum = 1;
4382 const MachineOperand &MO = MI->getOperand(OpNum);
4383 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
4384 // Use the same magic number as getPartialRegUpdateClearance.
4385 return 16;
4386 }
4387 return 0;
4388 }
4390 void X86InstrInfo::
4391 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4392 const TargetRegisterInfo *TRI) const {
4393 unsigned Reg = MI->getOperand(OpNum).getReg();
4394 // If MI kills this register, the false dependence is already broken.
4395 if (MI->killsRegister(Reg, TRI))
4396 return;
4397 if (X86::VR128RegClass.contains(Reg)) {
4398 // These instructions are all floating point domain, so xorps is the best
4399 // choice.
4400 bool HasAVX = Subtarget.hasAVX();
4401 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4402 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4403 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4404 } else if (X86::VR256RegClass.contains(Reg)) {
4405 // Use vxorps to clear the full ymm register.
4406 // It wants to read and write the xmm sub-register.
4407 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4408 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4409 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4410 .addReg(Reg, RegState::ImplicitDefine);
4411 } else
4412 return;
4413 MI->addRegisterKilled(Reg, TRI, true);
4414 }
4416 MachineInstr*
4417 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
4418 const SmallVectorImpl<unsigned> &Ops,
4419 int FrameIndex) const {
4420 // Check switch flag
4421 if (NoFusing) return nullptr;
4423 // Unless optimizing for size, don't fold to avoid partial
4424 // register update stalls
4425 if (!MF.getFunction()->getAttributes().
4426 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4427 hasPartialRegUpdate(MI->getOpcode()))
4428 return nullptr;
4430 const MachineFrameInfo *MFI = MF.getFrameInfo();
4431 unsigned Size = MFI->getObjectSize(FrameIndex);
4432 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
4433 // If the function stack isn't realigned we don't want to fold instructions
4434 // that need increased alignment.
4435 if (!RI.needsStackRealignment(MF))
4436 Alignment = std::min(Alignment, MF.getTarget()
4437 .getSubtargetImpl()
4438 ->getFrameLowering()
4439 ->getStackAlignment());
4440 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4441 unsigned NewOpc = 0;
4442 unsigned RCSize = 0;
4443 switch (MI->getOpcode()) {
4444 default: return nullptr;
4445 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
4446 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4447 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4448 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4449 }
4450 // Check if it's safe to fold the load. If the size of the object is
4451 // narrower than the load width, then it's not.
4452 if (Size < RCSize)
4453 return nullptr;
4454 // Change to CMPXXri r, 0 first.
4455 MI->setDesc(get(NewOpc));
4456 MI->getOperand(1).ChangeToImmediate(0);
4457 } else if (Ops.size() != 1)
4458 return nullptr;
4460 SmallVector<MachineOperand,4> MOs;
4461 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
4462 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment, /*AllowCommute=*/ true);
4463 }
4465 static bool isPartialRegisterLoad(const MachineInstr &LoadMI,
4466 const MachineFunction &MF) {
4467 unsigned Opc = LoadMI.getOpcode();
4468 unsigned RegSize =
4469 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg())->getSize();
4471 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm) && RegSize > 4)
4472 // These instructions only load 32 bits, we can't fold them if the
4473 // destination register is wider than 32 bits (4 bytes).
4474 return true;
4476 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm) && RegSize > 8)
4477 // These instructions only load 64 bits, we can't fold them if the
4478 // destination register is wider than 64 bits (8 bytes).
4479 return true;
4481 return false;
4482 }
4484 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4485 MachineInstr *MI,
4486 const SmallVectorImpl<unsigned> &Ops,
4487 MachineInstr *LoadMI) const {
4488 // If loading from a FrameIndex, fold directly from the FrameIndex.
4489 unsigned NumOps = LoadMI->getDesc().getNumOperands();
4490 int FrameIndex;
4491 if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
4492 if (isPartialRegisterLoad(*LoadMI, MF))
4493 return nullptr;
4494 return foldMemoryOperandImpl(MF, MI, Ops, FrameIndex);
4495 }
4497 // Check switch flag
4498 if (NoFusing) return nullptr;
4500 // Unless optimizing for size, don't fold to avoid partial
4501 // register update stalls
4502 if (!MF.getFunction()->getAttributes().
4503 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4504 hasPartialRegUpdate(MI->getOpcode()))
4505 return nullptr;
4507 // Determine the alignment of the load.
4508 unsigned Alignment = 0;
4509 if (LoadMI->hasOneMemOperand())
4510 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
4511 else
4512 switch (LoadMI->getOpcode()) {
4513 case X86::AVX2_SETALLONES:
4514 case X86::AVX_SET0:
4515 Alignment = 32;
4516 break;
4517 case X86::V_SET0:
4518 case X86::V_SETALLONES:
4519 Alignment = 16;
4520 break;
4521 case X86::FsFLD0SD:
4522 Alignment = 8;
4523 break;
4524 case X86::FsFLD0SS:
4525 Alignment = 4;
4526 break;
4527 default:
4528 return nullptr;
4529 }
4530 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4531 unsigned NewOpc = 0;
4532 switch (MI->getOpcode()) {
4533 default: return nullptr;
4534 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
4535 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4536 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4537 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
4538 }
4539 // Change to CMPXXri r, 0 first.
4540 MI->setDesc(get(NewOpc));
4541 MI->getOperand(1).ChangeToImmediate(0);
4542 } else if (Ops.size() != 1)
4543 return nullptr;
4545 // Make sure the subregisters match.
4546 // Otherwise we risk changing the size of the load.
4547 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4548 return nullptr;
4550 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
4551 switch (LoadMI->getOpcode()) {
4552 case X86::V_SET0:
4553 case X86::V_SETALLONES:
4554 case X86::AVX2_SETALLONES:
4555 case X86::AVX_SET0:
4556 case X86::FsFLD0SD:
4557 case X86::FsFLD0SS: {
4558 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
4559 // Create a constant-pool entry and operands to load from it.
4561 // Medium and large mode can't fold loads this way.
4562 if (MF.getTarget().getCodeModel() != CodeModel::Small &&
4563 MF.getTarget().getCodeModel() != CodeModel::Kernel)
4564 return nullptr;
4566 // x86-32 PIC requires a PIC base register for constant pools.
4567 unsigned PICBase = 0;
4568 if (MF.getTarget().getRelocationModel() == Reloc::PIC_) {
4569 if (Subtarget.is64Bit())
4570 PICBase = X86::RIP;
4571 else
4572 // FIXME: PICBase = getGlobalBaseReg(&MF);
4573 // This doesn't work for several reasons.
4574 // 1. GlobalBaseReg may have been spilled.
4575 // 2. It may not be live at MI.
4576 return nullptr;
4577 }
4579 // Create a constant-pool entry.
4580 MachineConstantPool &MCP = *MF.getConstantPool();
4581 Type *Ty;
4582 unsigned Opc = LoadMI->getOpcode();
4583 if (Opc == X86::FsFLD0SS)
4584 Ty = Type::getFloatTy(MF.getFunction()->getContext());
4585 else if (Opc == X86::FsFLD0SD)
4586 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
4587 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
4588 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
4589 else
4590 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
4592 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
4593 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4594 Constant::getNullValue(Ty);
4595 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
4597 // Create operands to load from the constant pool entry.
4598 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4599 MOs.push_back(MachineOperand::CreateImm(1));
4600 MOs.push_back(MachineOperand::CreateReg(0, false));
4601 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
4602 MOs.push_back(MachineOperand::CreateReg(0, false));
4603 break;
4604 }
4605 default: {
4606 if (isPartialRegisterLoad(*LoadMI, MF))
4607 return nullptr;
4609 // Folding a normal load. Just copy the load's address operands.
4610 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
4611 MOs.push_back(LoadMI->getOperand(i));
4612 break;
4613 }
4614 }
4615 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment, /*AllowCommute=*/ true);
4616 }
4619 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4620 const SmallVectorImpl<unsigned> &Ops) const {
4621 // Check switch flag
4622 if (NoFusing) return 0;
4624 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4625 switch (MI->getOpcode()) {
4626 default: return false;
4627 case X86::TEST8rr:
4628 case X86::TEST16rr:
4629 case X86::TEST32rr:
4630 case X86::TEST64rr:
4631 return true;
4632 case X86::ADD32ri:
4633 // FIXME: AsmPrinter doesn't know how to handle
4634 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4635 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4636 return false;
4637 break;
4638 }
4639 }
4641 if (Ops.size() != 1)
4642 return false;
4644 unsigned OpNum = Ops[0];
4645 unsigned Opc = MI->getOpcode();
4646 unsigned NumOps = MI->getDesc().getNumOperands();
4647 bool isTwoAddr = NumOps > 1 &&
4648 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4650 // Folding a memory location into the two-address part of a two-address
4651 // instruction is different than folding it other places. It requires
4652 // replacing the *two* registers with the memory location.
4653 const DenseMap<unsigned,
4654 std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4655 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4656 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4657 } else if (OpNum == 0) { // If operand 0
4658 if (Opc == X86::MOV32r0)
4659 return true;
4661 OpcodeTablePtr = &RegOp2MemOpTable0;
4662 } else if (OpNum == 1) {
4663 OpcodeTablePtr = &RegOp2MemOpTable1;
4664 } else if (OpNum == 2) {
4665 OpcodeTablePtr = &RegOp2MemOpTable2;
4666 } else if (OpNum == 3) {
4667 OpcodeTablePtr = &RegOp2MemOpTable3;
4668 }
4670 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4671 return true;
4672 return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
4673 }
4675 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4676 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4677 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4678 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4679 MemOp2RegOpTable.find(MI->getOpcode());
4680 if (I == MemOp2RegOpTable.end())
4681 return false;
4682 unsigned Opc = I->second.first;
4683 unsigned Index = I->second.second & TB_INDEX_MASK;
4684 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4685 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4686 if (UnfoldLoad && !FoldedLoad)
4687 return false;
4688 UnfoldLoad &= FoldedLoad;
4689 if (UnfoldStore && !FoldedStore)
4690 return false;
4691 UnfoldStore &= FoldedStore;
4693 const MCInstrDesc &MCID = get(Opc);
4694 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4695 if (!MI->hasOneMemOperand() &&
4696 RC == &X86::VR128RegClass &&
4697 !Subtarget.isUnalignedMemAccessFast())
4698 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4699 // conservatively assume the address is unaligned. That's bad for
4700 // performance.
4701 return false;
4702 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4703 SmallVector<MachineOperand,2> BeforeOps;
4704 SmallVector<MachineOperand,2> AfterOps;
4705 SmallVector<MachineOperand,4> ImpOps;
4706 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4707 MachineOperand &Op = MI->getOperand(i);
4708 if (i >= Index && i < Index + X86::AddrNumOperands)
4709 AddrOps.push_back(Op);
4710 else if (Op.isReg() && Op.isImplicit())
4711 ImpOps.push_back(Op);
4712 else if (i < Index)
4713 BeforeOps.push_back(Op);
4714 else if (i > Index)
4715 AfterOps.push_back(Op);
4716 }
4718 // Emit the load instruction.
4719 if (UnfoldLoad) {
4720 std::pair<MachineInstr::mmo_iterator,
4721 MachineInstr::mmo_iterator> MMOs =
4722 MF.extractLoadMemRefs(MI->memoperands_begin(),
4723 MI->memoperands_end());
4724 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4725 if (UnfoldStore) {
4726 // Address operands cannot be marked isKill.
4727 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4728 MachineOperand &MO = NewMIs[0]->getOperand(i);
4729 if (MO.isReg())
4730 MO.setIsKill(false);
4731 }
4732 }
4733 }
4735 // Emit the data processing instruction.
4736 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4737 MachineInstrBuilder MIB(MF, DataMI);
4739 if (FoldedStore)
4740 MIB.addReg(Reg, RegState::Define);
4741 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4742 MIB.addOperand(BeforeOps[i]);
4743 if (FoldedLoad)
4744 MIB.addReg(Reg);
4745 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4746 MIB.addOperand(AfterOps[i]);
4747 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4748 MachineOperand &MO = ImpOps[i];
4749 MIB.addReg(MO.getReg(),
4750 getDefRegState(MO.isDef()) |
4751 RegState::Implicit |
4752 getKillRegState(MO.isKill()) |
4753 getDeadRegState(MO.isDead()) |
4754 getUndefRegState(MO.isUndef()));
4755 }
4756 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4757 switch (DataMI->getOpcode()) {
4758 default: break;
4759 case X86::CMP64ri32:
4760 case X86::CMP64ri8:
4761 case X86::CMP32ri:
4762 case X86::CMP32ri8:
4763 case X86::CMP16ri:
4764 case X86::CMP16ri8:
4765 case X86::CMP8ri: {
4766 MachineOperand &MO0 = DataMI->getOperand(0);
4767 MachineOperand &MO1 = DataMI->getOperand(1);
4768 if (MO1.getImm() == 0) {
4769 unsigned NewOpc;
4770 switch (DataMI->getOpcode()) {
4771 default: llvm_unreachable("Unreachable!");
4772 case X86::CMP64ri8:
4773 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
4774 case X86::CMP32ri8:
4775 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
4776 case X86::CMP16ri8:
4777 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
4778 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
4779 }
4780 DataMI->setDesc(get(NewOpc));
4781 MO1.ChangeToRegister(MO0.getReg(), false);
4782 }
4783 }
4784 }
4785 NewMIs.push_back(DataMI);
4787 // Emit the store instruction.
4788 if (UnfoldStore) {
4789 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
4790 std::pair<MachineInstr::mmo_iterator,
4791 MachineInstr::mmo_iterator> MMOs =
4792 MF.extractStoreMemRefs(MI->memoperands_begin(),
4793 MI->memoperands_end());
4794 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
4795 }
4797 return true;
4798 }
4800 bool
4801 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
4802 SmallVectorImpl<SDNode*> &NewNodes) const {
4803 if (!N->isMachineOpcode())
4804 return false;
4806 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4807 MemOp2RegOpTable.find(N->getMachineOpcode());
4808 if (I == MemOp2RegOpTable.end())
4809 return false;
4810 unsigned Opc = I->second.first;
4811 unsigned Index = I->second.second & TB_INDEX_MASK;
4812 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4813 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4814 const MCInstrDesc &MCID = get(Opc);
4815 MachineFunction &MF = DAG.getMachineFunction();
4816 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4817 unsigned NumDefs = MCID.NumDefs;
4818 std::vector<SDValue> AddrOps;
4819 std::vector<SDValue> BeforeOps;
4820 std::vector<SDValue> AfterOps;
4821 SDLoc dl(N);
4822 unsigned NumOps = N->getNumOperands();
4823 for (unsigned i = 0; i != NumOps-1; ++i) {
4824 SDValue Op = N->getOperand(i);
4825 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
4826 AddrOps.push_back(Op);
4827 else if (i < Index-NumDefs)
4828 BeforeOps.push_back(Op);
4829 else if (i > Index-NumDefs)
4830 AfterOps.push_back(Op);
4831 }
4832 SDValue Chain = N->getOperand(NumOps-1);
4833 AddrOps.push_back(Chain);
4835 // Emit the load instruction.
4836 SDNode *Load = nullptr;
4837 if (FoldedLoad) {
4838 EVT VT = *RC->vt_begin();
4839 std::pair<MachineInstr::mmo_iterator,
4840 MachineInstr::mmo_iterator> MMOs =
4841 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4842 cast<MachineSDNode>(N)->memoperands_end());
4843 if (!(*MMOs.first) &&
4844 RC == &X86::VR128RegClass &&
4845 !Subtarget.isUnalignedMemAccessFast())
4846 // Do not introduce a slow unaligned load.
4847 return false;
4848 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4849 bool isAligned = (*MMOs.first) &&
4850 (*MMOs.first)->getAlignment() >= Alignment;
4851 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
4852 VT, MVT::Other, AddrOps);
4853 NewNodes.push_back(Load);
4855 // Preserve memory reference information.
4856 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4857 }
4859 // Emit the data processing instruction.
4860 std::vector<EVT> VTs;
4861 const TargetRegisterClass *DstRC = nullptr;
4862 if (MCID.getNumDefs() > 0) {
4863 DstRC = getRegClass(MCID, 0, &RI, MF);
4864 VTs.push_back(*DstRC->vt_begin());
4865 }
4866 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
4867 EVT VT = N->getValueType(i);
4868 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
4869 VTs.push_back(VT);
4870 }
4871 if (Load)
4872 BeforeOps.push_back(SDValue(Load, 0));
4873 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
4874 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
4875 NewNodes.push_back(NewNode);
4877 // Emit the store instruction.
4878 if (FoldedStore) {
4879 AddrOps.pop_back();
4880 AddrOps.push_back(SDValue(NewNode, 0));
4881 AddrOps.push_back(Chain);
4882 std::pair<MachineInstr::mmo_iterator,
4883 MachineInstr::mmo_iterator> MMOs =
4884 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4885 cast<MachineSDNode>(N)->memoperands_end());
4886 if (!(*MMOs.first) &&
4887 RC == &X86::VR128RegClass &&
4888 !Subtarget.isUnalignedMemAccessFast())
4889 // Do not introduce a slow unaligned store.
4890 return false;
4891 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4892 bool isAligned = (*MMOs.first) &&
4893 (*MMOs.first)->getAlignment() >= Alignment;
4894 SDNode *Store =
4895 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
4896 dl, MVT::Other, AddrOps);
4897 NewNodes.push_back(Store);
4899 // Preserve memory reference information.
4900 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4901 }
4903 return true;
4904 }
4906 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
4907 bool UnfoldLoad, bool UnfoldStore,
4908 unsigned *LoadRegIndex) const {
4909 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4910 MemOp2RegOpTable.find(Opc);
4911 if (I == MemOp2RegOpTable.end())
4912 return 0;
4913 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4914 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4915 if (UnfoldLoad && !FoldedLoad)
4916 return 0;
4917 if (UnfoldStore && !FoldedStore)
4918 return 0;
4919 if (LoadRegIndex)
4920 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
4921 return I->second.first;
4922 }
4924 bool
4925 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4926 int64_t &Offset1, int64_t &Offset2) const {
4927 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4928 return false;
4929 unsigned Opc1 = Load1->getMachineOpcode();
4930 unsigned Opc2 = Load2->getMachineOpcode();
4931 switch (Opc1) {
4932 default: return false;
4933 case X86::MOV8rm:
4934 case X86::MOV16rm:
4935 case X86::MOV32rm:
4936 case X86::MOV64rm:
4937 case X86::LD_Fp32m:
4938 case X86::LD_Fp64m:
4939 case X86::LD_Fp80m:
4940 case X86::MOVSSrm:
4941 case X86::MOVSDrm:
4942 case X86::MMX_MOVD64rm:
4943 case X86::MMX_MOVQ64rm:
4944 case X86::FsMOVAPSrm:
4945 case X86::FsMOVAPDrm:
4946 case X86::MOVAPSrm:
4947 case X86::MOVUPSrm:
4948 case X86::MOVAPDrm:
4949 case X86::MOVDQArm:
4950 case X86::MOVDQUrm:
4951 // AVX load instructions
4952 case X86::VMOVSSrm:
4953 case X86::VMOVSDrm:
4954 case X86::FsVMOVAPSrm:
4955 case X86::FsVMOVAPDrm:
4956 case X86::VMOVAPSrm:
4957 case X86::VMOVUPSrm:
4958 case X86::VMOVAPDrm:
4959 case X86::VMOVDQArm:
4960 case X86::VMOVDQUrm:
4961 case X86::VMOVAPSYrm:
4962 case X86::VMOVUPSYrm:
4963 case X86::VMOVAPDYrm:
4964 case X86::VMOVDQAYrm:
4965 case X86::VMOVDQUYrm:
4966 break;
4967 }
4968 switch (Opc2) {
4969 default: return false;
4970 case X86::MOV8rm:
4971 case X86::MOV16rm:
4972 case X86::MOV32rm:
4973 case X86::MOV64rm:
4974 case X86::LD_Fp32m:
4975 case X86::LD_Fp64m:
4976 case X86::LD_Fp80m:
4977 case X86::MOVSSrm:
4978 case X86::MOVSDrm:
4979 case X86::MMX_MOVD64rm:
4980 case X86::MMX_MOVQ64rm:
4981 case X86::FsMOVAPSrm:
4982 case X86::FsMOVAPDrm:
4983 case X86::MOVAPSrm:
4984 case X86::MOVUPSrm:
4985 case X86::MOVAPDrm:
4986 case X86::MOVDQArm:
4987 case X86::MOVDQUrm:
4988 // AVX load instructions
4989 case X86::VMOVSSrm:
4990 case X86::VMOVSDrm:
4991 case X86::FsVMOVAPSrm:
4992 case X86::FsVMOVAPDrm:
4993 case X86::VMOVAPSrm:
4994 case X86::VMOVUPSrm:
4995 case X86::VMOVAPDrm:
4996 case X86::VMOVDQArm:
4997 case X86::VMOVDQUrm:
4998 case X86::VMOVAPSYrm:
4999 case X86::VMOVUPSYrm:
5000 case X86::VMOVAPDYrm:
5001 case X86::VMOVDQAYrm:
5002 case X86::VMOVDQUYrm:
5003 break;
5004 }
5006 // Check if chain operands and base addresses match.
5007 if (Load1->getOperand(0) != Load2->getOperand(0) ||
5008 Load1->getOperand(5) != Load2->getOperand(5))
5009 return false;
5010 // Segment operands should match as well.
5011 if (Load1->getOperand(4) != Load2->getOperand(4))
5012 return false;
5013 // Scale should be 1, Index should be Reg0.
5014 if (Load1->getOperand(1) == Load2->getOperand(1) &&
5015 Load1->getOperand(2) == Load2->getOperand(2)) {
5016 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
5017 return false;
5019 // Now let's examine the displacements.
5020 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
5021 isa<ConstantSDNode>(Load2->getOperand(3))) {
5022 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
5023 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
5024 return true;
5025 }
5026 }
5027 return false;
5028 }
5030 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5031 int64_t Offset1, int64_t Offset2,
5032 unsigned NumLoads) const {
5033 assert(Offset2 > Offset1);
5034 if ((Offset2 - Offset1) / 8 > 64)
5035 return false;
5037 unsigned Opc1 = Load1->getMachineOpcode();
5038 unsigned Opc2 = Load2->getMachineOpcode();
5039 if (Opc1 != Opc2)
5040 return false; // FIXME: overly conservative?
5042 switch (Opc1) {
5043 default: break;
5044 case X86::LD_Fp32m:
5045 case X86::LD_Fp64m:
5046 case X86::LD_Fp80m:
5047 case X86::MMX_MOVD64rm:
5048 case X86::MMX_MOVQ64rm:
5049 return false;
5050 }
5052 EVT VT = Load1->getValueType(0);
5053 switch (VT.getSimpleVT().SimpleTy) {
5054 default:
5055 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5056 // have 16 of them to play with.
5057 if (Subtarget.is64Bit()) {
5058 if (NumLoads >= 3)
5059 return false;
5060 } else if (NumLoads) {
5061 return false;
5062 }
5063 break;
5064 case MVT::i8:
5065 case MVT::i16:
5066 case MVT::i32:
5067 case MVT::i64:
5068 case MVT::f32:
5069 case MVT::f64:
5070 if (NumLoads)
5071 return false;
5072 break;
5073 }
5075 return true;
5076 }
5078 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
5079 MachineInstr *Second) const {
5080 // Check if this processor supports macro-fusion. Since this is a minor
5081 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
5082 // proxy for SandyBridge+.
5083 if (!Subtarget.hasAVX())
5084 return false;
5086 enum {
5087 FuseTest,
5088 FuseCmp,
5089 FuseInc
5090 } FuseKind;
5092 switch(Second->getOpcode()) {
5093 default:
5094 return false;
5095 case X86::JE_4:
5096 case X86::JNE_4:
5097 case X86::JL_4:
5098 case X86::JLE_4:
5099 case X86::JG_4:
5100 case X86::JGE_4:
5101 FuseKind = FuseInc;
5102 break;
5103 case X86::JB_4:
5104 case X86::JBE_4:
5105 case X86::JA_4:
5106 case X86::JAE_4:
5107 FuseKind = FuseCmp;
5108 break;
5109 case X86::JS_4:
5110 case X86::JNS_4:
5111 case X86::JP_4:
5112 case X86::JNP_4:
5113 case X86::JO_4:
5114 case X86::JNO_4:
5115 FuseKind = FuseTest;
5116 break;
5117 }
5118 switch (First->getOpcode()) {
5119 default:
5120 return false;
5121 case X86::TEST8rr:
5122 case X86::TEST16rr:
5123 case X86::TEST32rr:
5124 case X86::TEST64rr:
5125 case X86::TEST8ri:
5126 case X86::TEST16ri:
5127 case X86::TEST32ri:
5128 case X86::TEST32i32:
5129 case X86::TEST64i32:
5130 case X86::TEST64ri32:
5131 case X86::TEST8rm:
5132 case X86::TEST16rm:
5133 case X86::TEST32rm:
5134 case X86::TEST64rm:
5135 case X86::TEST8ri_NOREX:
5136 case X86::AND16i16:
5137 case X86::AND16ri:
5138 case X86::AND16ri8:
5139 case X86::AND16rm:
5140 case X86::AND16rr:
5141 case X86::AND32i32:
5142 case X86::AND32ri:
5143 case X86::AND32ri8:
5144 case X86::AND32rm:
5145 case X86::AND32rr:
5146 case X86::AND64i32:
5147 case X86::AND64ri32:
5148 case X86::AND64ri8:
5149 case X86::AND64rm:
5150 case X86::AND64rr:
5151 case X86::AND8i8:
5152 case X86::AND8ri:
5153 case X86::AND8rm:
5154 case X86::AND8rr:
5155 return true;
5156 case X86::CMP16i16:
5157 case X86::CMP16ri:
5158 case X86::CMP16ri8:
5159 case X86::CMP16rm:
5160 case X86::CMP16rr:
5161 case X86::CMP32i32:
5162 case X86::CMP32ri:
5163 case X86::CMP32ri8:
5164 case X86::CMP32rm:
5165 case X86::CMP32rr:
5166 case X86::CMP64i32:
5167 case X86::CMP64ri32:
5168 case X86::CMP64ri8:
5169 case X86::CMP64rm:
5170 case X86::CMP64rr:
5171 case X86::CMP8i8:
5172 case X86::CMP8ri:
5173 case X86::CMP8rm:
5174 case X86::CMP8rr:
5175 case X86::ADD16i16:
5176 case X86::ADD16ri:
5177 case X86::ADD16ri8:
5178 case X86::ADD16ri8_DB:
5179 case X86::ADD16ri_DB:
5180 case X86::ADD16rm:
5181 case X86::ADD16rr:
5182 case X86::ADD16rr_DB:
5183 case X86::ADD32i32:
5184 case X86::ADD32ri:
5185 case X86::ADD32ri8:
5186 case X86::ADD32ri8_DB:
5187 case X86::ADD32ri_DB:
5188 case X86::ADD32rm:
5189 case X86::ADD32rr:
5190 case X86::ADD32rr_DB:
5191 case X86::ADD64i32:
5192 case X86::ADD64ri32:
5193 case X86::ADD64ri32_DB:
5194 case X86::ADD64ri8:
5195 case X86::ADD64ri8_DB:
5196 case X86::ADD64rm:
5197 case X86::ADD64rr:
5198 case X86::ADD64rr_DB:
5199 case X86::ADD8i8:
5200 case X86::ADD8mi:
5201 case X86::ADD8mr:
5202 case X86::ADD8ri:
5203 case X86::ADD8rm:
5204 case X86::ADD8rr:
5205 case X86::SUB16i16:
5206 case X86::SUB16ri:
5207 case X86::SUB16ri8:
5208 case X86::SUB16rm:
5209 case X86::SUB16rr:
5210 case X86::SUB32i32:
5211 case X86::SUB32ri:
5212 case X86::SUB32ri8:
5213 case X86::SUB32rm:
5214 case X86::SUB32rr:
5215 case X86::SUB64i32:
5216 case X86::SUB64ri32:
5217 case X86::SUB64ri8:
5218 case X86::SUB64rm:
5219 case X86::SUB64rr:
5220 case X86::SUB8i8:
5221 case X86::SUB8ri:
5222 case X86::SUB8rm:
5223 case X86::SUB8rr:
5224 return FuseKind == FuseCmp || FuseKind == FuseInc;
5225 case X86::INC16r:
5226 case X86::INC32r:
5227 case X86::INC64_16r:
5228 case X86::INC64_32r:
5229 case X86::INC64r:
5230 case X86::INC8r:
5231 case X86::DEC16r:
5232 case X86::DEC32r:
5233 case X86::DEC64_16r:
5234 case X86::DEC64_32r:
5235 case X86::DEC64r:
5236 case X86::DEC8r:
5237 return FuseKind == FuseInc;
5238 }
5239 }
5241 bool X86InstrInfo::
5242 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
5243 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
5244 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
5245 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
5246 return true;
5247 Cond[0].setImm(GetOppositeBranchCondition(CC));
5248 return false;
5249 }
5251 bool X86InstrInfo::
5252 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
5253 // FIXME: Return false for x87 stack register classes for now. We can't
5254 // allow any loads of these registers before FpGet_ST0_80.
5255 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
5256 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
5257 }
5259 /// getGlobalBaseReg - Return a virtual register initialized with the
5260 /// the global base register value. Output instructions required to
5261 /// initialize the register in the function entry block, if necessary.
5262 ///
5263 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
5264 ///
5265 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
5266 assert(!Subtarget.is64Bit() &&
5267 "X86-64 PIC uses RIP relative addressing");
5269 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
5270 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5271 if (GlobalBaseReg != 0)
5272 return GlobalBaseReg;
5274 // Create the register. The code to initialize it is inserted
5275 // later, by the CGBR pass (below).
5276 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5277 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
5278 X86FI->setGlobalBaseReg(GlobalBaseReg);
5279 return GlobalBaseReg;
5280 }
5282 // These are the replaceable SSE instructions. Some of these have Int variants
5283 // that we don't include here. We don't want to replace instructions selected
5284 // by intrinsics.
5285 static const uint16_t ReplaceableInstrs[][3] = {
5286 //PackedSingle PackedDouble PackedInt
5287 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
5288 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
5289 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
5290 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
5291 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
5292 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
5293 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
5294 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
5295 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
5296 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
5297 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
5298 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
5299 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
5300 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
5301 // AVX 128-bit support
5302 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
5303 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
5304 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
5305 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
5306 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
5307 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5308 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
5309 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
5310 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
5311 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
5312 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
5313 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
5314 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
5315 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
5316 // AVX 256-bit support
5317 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
5318 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
5319 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
5320 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
5321 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
5322 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
5323 };
5325 static const uint16_t ReplaceableInstrsAVX2[][3] = {
5326 //PackedSingle PackedDouble PackedInt
5327 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
5328 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
5329 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
5330 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
5331 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
5332 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
5333 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
5334 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
5335 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
5336 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
5337 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
5338 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
5339 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
5340 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
5341 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
5342 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
5343 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
5344 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
5345 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
5346 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
5347 };
5349 // FIXME: Some shuffle and unpack instructions have equivalents in different
5350 // domains, but they require a bit more work than just switching opcodes.
5352 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
5353 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
5354 if (ReplaceableInstrs[i][domain-1] == opcode)
5355 return ReplaceableInstrs[i];
5356 return nullptr;
5357 }
5359 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
5360 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
5361 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
5362 return ReplaceableInstrsAVX2[i];
5363 return nullptr;
5364 }
5366 std::pair<uint16_t, uint16_t>
5367 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
5368 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5369 bool hasAVX2 = Subtarget.hasAVX2();
5370 uint16_t validDomains = 0;
5371 if (domain && lookup(MI->getOpcode(), domain))
5372 validDomains = 0xe;
5373 else if (domain && lookupAVX2(MI->getOpcode(), domain))
5374 validDomains = hasAVX2 ? 0xe : 0x6;
5375 return std::make_pair(domain, validDomains);
5376 }
5378 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
5379 assert(Domain>0 && Domain<4 && "Invalid execution domain");
5380 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5381 assert(dom && "Not an SSE instruction");
5382 const uint16_t *table = lookup(MI->getOpcode(), dom);
5383 if (!table) { // try the other table
5384 assert((Subtarget.hasAVX2() || Domain < 3) &&
5385 "256-bit vector operations only available in AVX2");
5386 table = lookupAVX2(MI->getOpcode(), dom);
5387 }
5388 assert(table && "Cannot change domain");
5389 MI->setDesc(get(table[Domain-1]));
5390 }
5392 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5393 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5394 NopInst.setOpcode(X86::NOOP);
5395 }
5397 void X86InstrInfo::getUnconditionalBranch(
5398 MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const {
5399 Branch.setOpcode(X86::JMP_4);
5400 Branch.addOperand(MCOperand::CreateExpr(BranchTarget));
5401 }
5403 void X86InstrInfo::getTrap(MCInst &MI) const {
5404 MI.setOpcode(X86::TRAP);
5405 }
5407 bool X86InstrInfo::isHighLatencyDef(int opc) const {
5408 switch (opc) {
5409 default: return false;
5410 case X86::DIVSDrm:
5411 case X86::DIVSDrm_Int:
5412 case X86::DIVSDrr:
5413 case X86::DIVSDrr_Int:
5414 case X86::DIVSSrm:
5415 case X86::DIVSSrm_Int:
5416 case X86::DIVSSrr:
5417 case X86::DIVSSrr_Int:
5418 case X86::SQRTPDm:
5419 case X86::SQRTPDr:
5420 case X86::SQRTPSm:
5421 case X86::SQRTPSr:
5422 case X86::SQRTSDm:
5423 case X86::SQRTSDm_Int:
5424 case X86::SQRTSDr:
5425 case X86::SQRTSDr_Int:
5426 case X86::SQRTSSm:
5427 case X86::SQRTSSm_Int:
5428 case X86::SQRTSSr:
5429 case X86::SQRTSSr_Int:
5430 // AVX instructions with high latency
5431 case X86::VDIVSDrm:
5432 case X86::VDIVSDrm_Int:
5433 case X86::VDIVSDrr:
5434 case X86::VDIVSDrr_Int:
5435 case X86::VDIVSSrm:
5436 case X86::VDIVSSrm_Int:
5437 case X86::VDIVSSrr:
5438 case X86::VDIVSSrr_Int:
5439 case X86::VSQRTPDm:
5440 case X86::VSQRTPDr:
5441 case X86::VSQRTPSm:
5442 case X86::VSQRTPSr:
5443 case X86::VSQRTSDm:
5444 case X86::VSQRTSDm_Int:
5445 case X86::VSQRTSDr:
5446 case X86::VSQRTSSm:
5447 case X86::VSQRTSSm_Int:
5448 case X86::VSQRTSSr:
5449 case X86::VSQRTPDZrm:
5450 case X86::VSQRTPDZrr:
5451 case X86::VSQRTPSZrm:
5452 case X86::VSQRTPSZrr:
5453 case X86::VSQRTSDZm:
5454 case X86::VSQRTSDZm_Int:
5455 case X86::VSQRTSDZr:
5456 case X86::VSQRTSSZm_Int:
5457 case X86::VSQRTSSZr:
5458 case X86::VSQRTSSZm:
5459 case X86::VDIVSDZrm:
5460 case X86::VDIVSDZrr:
5461 case X86::VDIVSSZrm:
5462 case X86::VDIVSSZrr:
5464 case X86::VGATHERQPSZrm:
5465 case X86::VGATHERQPDZrm:
5466 case X86::VGATHERDPDZrm:
5467 case X86::VGATHERDPSZrm:
5468 case X86::VPGATHERQDZrm:
5469 case X86::VPGATHERQQZrm:
5470 case X86::VPGATHERDDZrm:
5471 case X86::VPGATHERDQZrm:
5472 case X86::VSCATTERQPDZmr:
5473 case X86::VSCATTERQPSZmr:
5474 case X86::VSCATTERDPDZmr:
5475 case X86::VSCATTERDPSZmr:
5476 case X86::VPSCATTERQDZmr:
5477 case X86::VPSCATTERQQZmr:
5478 case X86::VPSCATTERDDZmr:
5479 case X86::VPSCATTERDQZmr:
5480 return true;
5481 }
5482 }
5484 bool X86InstrInfo::
5485 hasHighOperandLatency(const InstrItineraryData *ItinData,
5486 const MachineRegisterInfo *MRI,
5487 const MachineInstr *DefMI, unsigned DefIdx,
5488 const MachineInstr *UseMI, unsigned UseIdx) const {
5489 return isHighLatencyDef(DefMI->getOpcode());
5490 }
5492 namespace {
5493 /// CGBR - Create Global Base Reg pass. This initializes the PIC
5494 /// global base register for x86-32.
5495 struct CGBR : public MachineFunctionPass {
5496 static char ID;
5497 CGBR() : MachineFunctionPass(ID) {}
5499 bool runOnMachineFunction(MachineFunction &MF) override {
5500 const X86TargetMachine *TM =
5501 static_cast<const X86TargetMachine *>(&MF.getTarget());
5503 // Don't do anything if this is 64-bit as 64-bit PIC
5504 // uses RIP relative addressing.
5505 if (TM->getSubtarget<X86Subtarget>().is64Bit())
5506 return false;
5508 // Only emit a global base reg in PIC mode.
5509 if (TM->getRelocationModel() != Reloc::PIC_)
5510 return false;
5512 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5513 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5515 // If we didn't need a GlobalBaseReg, don't insert code.
5516 if (GlobalBaseReg == 0)
5517 return false;
5519 // Insert the set of GlobalBaseReg into the first MBB of the function
5520 MachineBasicBlock &FirstMBB = MF.front();
5521 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5522 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5523 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5524 const X86InstrInfo *TII = TM->getSubtargetImpl()->getInstrInfo();
5526 unsigned PC;
5527 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
5528 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
5529 else
5530 PC = GlobalBaseReg;
5532 // Operand of MovePCtoStack is completely ignored by asm printer. It's
5533 // only used in JIT code emission as displacement to pc.
5534 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
5536 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5537 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5538 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
5539 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5540 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5541 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5542 X86II::MO_GOT_ABSOLUTE_ADDRESS);
5543 }
5545 return true;
5546 }
5548 const char *getPassName() const override {
5549 return "X86 PIC Global Base Reg Initialization";
5550 }
5552 void getAnalysisUsage(AnalysisUsage &AU) const override {
5553 AU.setPreservesCFG();
5554 MachineFunctionPass::getAnalysisUsage(AU);
5555 }
5556 };
5557 }
5559 char CGBR::ID = 0;
5560 FunctionPass*
5561 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
5563 namespace {
5564 struct LDTLSCleanup : public MachineFunctionPass {
5565 static char ID;
5566 LDTLSCleanup() : MachineFunctionPass(ID) {}
5568 bool runOnMachineFunction(MachineFunction &MF) override {
5569 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5570 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5571 // No point folding accesses if there isn't at least two.
5572 return false;
5573 }
5575 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5576 return VisitNode(DT->getRootNode(), 0);
5577 }
5579 // Visit the dominator subtree rooted at Node in pre-order.
5580 // If TLSBaseAddrReg is non-null, then use that to replace any
5581 // TLS_base_addr instructions. Otherwise, create the register
5582 // when the first such instruction is seen, and then use it
5583 // as we encounter more instructions.
5584 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5585 MachineBasicBlock *BB = Node->getBlock();
5586 bool Changed = false;
5588 // Traverse the current block.
5589 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5590 ++I) {
5591 switch (I->getOpcode()) {
5592 case X86::TLS_base_addr32:
5593 case X86::TLS_base_addr64:
5594 if (TLSBaseAddrReg)
5595 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5596 else
5597 I = SetRegister(I, &TLSBaseAddrReg);
5598 Changed = true;
5599 break;
5600 default:
5601 break;
5602 }
5603 }
5605 // Visit the children of this block in the dominator tree.
5606 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5607 I != E; ++I) {
5608 Changed |= VisitNode(*I, TLSBaseAddrReg);
5609 }
5611 return Changed;
5612 }
5614 // Replace the TLS_base_addr instruction I with a copy from
5615 // TLSBaseAddrReg, returning the new instruction.
5616 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5617 unsigned TLSBaseAddrReg) {
5618 MachineFunction *MF = I->getParent()->getParent();
5619 const X86TargetMachine *TM =
5620 static_cast<const X86TargetMachine *>(&MF->getTarget());
5621 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5622 const X86InstrInfo *TII = TM->getSubtargetImpl()->getInstrInfo();
5624 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5625 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5626 TII->get(TargetOpcode::COPY),
5627 is64Bit ? X86::RAX : X86::EAX)
5628 .addReg(TLSBaseAddrReg);
5630 // Erase the TLS_base_addr instruction.
5631 I->eraseFromParent();
5633 return Copy;
5634 }
5636 // Create a virtal register in *TLSBaseAddrReg, and populate it by
5637 // inserting a copy instruction after I. Returns the new instruction.
5638 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5639 MachineFunction *MF = I->getParent()->getParent();
5640 const X86TargetMachine *TM =
5641 static_cast<const X86TargetMachine *>(&MF->getTarget());
5642 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5643 const X86InstrInfo *TII = TM->getSubtargetImpl()->getInstrInfo();
5645 // Create a virtual register for the TLS base address.
5646 MachineRegisterInfo &RegInfo = MF->getRegInfo();
5647 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5648 ? &X86::GR64RegClass
5649 : &X86::GR32RegClass);
5651 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5652 MachineInstr *Next = I->getNextNode();
5653 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5654 TII->get(TargetOpcode::COPY),
5655 *TLSBaseAddrReg)
5656 .addReg(is64Bit ? X86::RAX : X86::EAX);
5658 return Copy;
5659 }
5661 const char *getPassName() const override {
5662 return "Local Dynamic TLS Access Clean-up";
5663 }
5665 void getAnalysisUsage(AnalysisUsage &AU) const override {
5666 AU.setPreservesCFG();
5667 AU.addRequired<MachineDominatorTree>();
5668 MachineFunctionPass::getAnalysisUsage(AU);
5669 }
5670 };
5671 }
5673 char LDTLSCleanup::ID = 0;
5674 FunctionPass*
5675 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }