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 //===----------------------------------------------------------------------===//
13 
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>
40 
41 using namespace llvm;
42 
43 #define DEBUG_TYPE "x86-instr-info"
44 
45 #define GET_INSTRINFO_CTOR_DTOR
46 #include "X86GenInstrInfo.inc"
47 
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);
60 
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_4    = 4,
69   TB_INDEX_MASK = 0xf,
70 
71   // Do not insert the reverse map (MemOp -> RegOp) into the table.
72   // This may be needed because there is a many -> one mapping.
73   TB_NO_REVERSE   = 1 << 4,
74 
75   // Do not insert the forward map (RegOp -> MemOp) into the table.
76   // This is needed for Native Client, which prohibits branch
77   // instructions from using a memory operand.
78   TB_NO_FORWARD   = 1 << 5,
79 
80   TB_FOLDED_LOAD  = 1 << 6,
81   TB_FOLDED_STORE = 1 << 7,
82 
83   // Minimum alignment required for load/store.
84   // Used for RegOp->MemOp conversion.
85   // (stored in bits 8 - 15)
86   TB_ALIGN_SHIFT = 8,
87   TB_ALIGN_NONE  =    0 << TB_ALIGN_SHIFT,
88   TB_ALIGN_16    =   16 << TB_ALIGN_SHIFT,
89   TB_ALIGN_32    =   32 << TB_ALIGN_SHIFT,
90   TB_ALIGN_64    =   64 << TB_ALIGN_SHIFT,
91   TB_ALIGN_MASK  = 0xff << TB_ALIGN_SHIFT
92 };
93 
94 struct X86MemoryFoldTableEntry {
95   uint16_t RegOp;
96   uint16_t MemOp;
97   uint16_t Flags;
98 };
99 
100 // Pin the vtable to this file.
101 void X86InstrInfo::anchor() {}
102 
103 X86InstrInfo::X86InstrInfo(X86Subtarget &STI)
104     : X86GenInstrInfo(
105           (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64 : X86::ADJCALLSTACKDOWN32),
106           (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64 : X86::ADJCALLSTACKUP32)),
107       Subtarget(STI), RI(STI.getTargetTriple()) {
108 
109   static const X86MemoryFoldTableEntry MemoryFoldTable2Addr[] = {
110     { X86::ADC32ri,     X86::ADC32mi,    0 },
111     { X86::ADC32ri8,    X86::ADC32mi8,   0 },
112     { X86::ADC32rr,     X86::ADC32mr,    0 },
113     { X86::ADC64ri32,   X86::ADC64mi32,  0 },
114     { X86::ADC64ri8,    X86::ADC64mi8,   0 },
115     { X86::ADC64rr,     X86::ADC64mr,    0 },
116     { X86::ADD16ri,     X86::ADD16mi,    0 },
117     { X86::ADD16ri8,    X86::ADD16mi8,   0 },
118     { X86::ADD16ri_DB,  X86::ADD16mi,    TB_NO_REVERSE },
119     { X86::ADD16ri8_DB, X86::ADD16mi8,   TB_NO_REVERSE },
120     { X86::ADD16rr,     X86::ADD16mr,    0 },
121     { X86::ADD16rr_DB,  X86::ADD16mr,    TB_NO_REVERSE },
122     { X86::ADD32ri,     X86::ADD32mi,    0 },
123     { X86::ADD32ri8,    X86::ADD32mi8,   0 },
124     { X86::ADD32ri_DB,  X86::ADD32mi,    TB_NO_REVERSE },
125     { X86::ADD32ri8_DB, X86::ADD32mi8,   TB_NO_REVERSE },
126     { X86::ADD32rr,     X86::ADD32mr,    0 },
127     { X86::ADD32rr_DB,  X86::ADD32mr,    TB_NO_REVERSE },
128     { X86::ADD64ri32,   X86::ADD64mi32,  0 },
129     { X86::ADD64ri8,    X86::ADD64mi8,   0 },
130     { X86::ADD64ri32_DB,X86::ADD64mi32,  TB_NO_REVERSE },
131     { X86::ADD64ri8_DB, X86::ADD64mi8,   TB_NO_REVERSE },
132     { X86::ADD64rr,     X86::ADD64mr,    0 },
133     { X86::ADD64rr_DB,  X86::ADD64mr,    TB_NO_REVERSE },
134     { X86::ADD8ri,      X86::ADD8mi,     0 },
135     { X86::ADD8rr,      X86::ADD8mr,     0 },
136     { X86::AND16ri,     X86::AND16mi,    0 },
137     { X86::AND16ri8,    X86::AND16mi8,   0 },
138     { X86::AND16rr,     X86::AND16mr,    0 },
139     { X86::AND32ri,     X86::AND32mi,    0 },
140     { X86::AND32ri8,    X86::AND32mi8,   0 },
141     { X86::AND32rr,     X86::AND32mr,    0 },
142     { X86::AND64ri32,   X86::AND64mi32,  0 },
143     { X86::AND64ri8,    X86::AND64mi8,   0 },
144     { X86::AND64rr,     X86::AND64mr,    0 },
145     { X86::AND8ri,      X86::AND8mi,     0 },
146     { X86::AND8rr,      X86::AND8mr,     0 },
147     { X86::DEC16r,      X86::DEC16m,     0 },
148     { X86::DEC32r,      X86::DEC32m,     0 },
149     { X86::DEC64r,      X86::DEC64m,     0 },
150     { X86::DEC8r,       X86::DEC8m,      0 },
151     { X86::INC16r,      X86::INC16m,     0 },
152     { X86::INC32r,      X86::INC32m,     0 },
153     { X86::INC64r,      X86::INC64m,     0 },
154     { X86::INC8r,       X86::INC8m,      0 },
155     { X86::NEG16r,      X86::NEG16m,     0 },
156     { X86::NEG32r,      X86::NEG32m,     0 },
157     { X86::NEG64r,      X86::NEG64m,     0 },
158     { X86::NEG8r,       X86::NEG8m,      0 },
159     { X86::NOT16r,      X86::NOT16m,     0 },
160     { X86::NOT32r,      X86::NOT32m,     0 },
161     { X86::NOT64r,      X86::NOT64m,     0 },
162     { X86::NOT8r,       X86::NOT8m,      0 },
163     { X86::OR16ri,      X86::OR16mi,     0 },
164     { X86::OR16ri8,     X86::OR16mi8,    0 },
165     { X86::OR16rr,      X86::OR16mr,     0 },
166     { X86::OR32ri,      X86::OR32mi,     0 },
167     { X86::OR32ri8,     X86::OR32mi8,    0 },
168     { X86::OR32rr,      X86::OR32mr,     0 },
169     { X86::OR64ri32,    X86::OR64mi32,   0 },
170     { X86::OR64ri8,     X86::OR64mi8,    0 },
171     { X86::OR64rr,      X86::OR64mr,     0 },
172     { X86::OR8ri,       X86::OR8mi,      0 },
173     { X86::OR8rr,       X86::OR8mr,      0 },
174     { X86::ROL16r1,     X86::ROL16m1,    0 },
175     { X86::ROL16rCL,    X86::ROL16mCL,   0 },
176     { X86::ROL16ri,     X86::ROL16mi,    0 },
177     { X86::ROL32r1,     X86::ROL32m1,    0 },
178     { X86::ROL32rCL,    X86::ROL32mCL,   0 },
179     { X86::ROL32ri,     X86::ROL32mi,    0 },
180     { X86::ROL64r1,     X86::ROL64m1,    0 },
181     { X86::ROL64rCL,    X86::ROL64mCL,   0 },
182     { X86::ROL64ri,     X86::ROL64mi,    0 },
183     { X86::ROL8r1,      X86::ROL8m1,     0 },
184     { X86::ROL8rCL,     X86::ROL8mCL,    0 },
185     { X86::ROL8ri,      X86::ROL8mi,     0 },
186     { X86::ROR16r1,     X86::ROR16m1,    0 },
187     { X86::ROR16rCL,    X86::ROR16mCL,   0 },
188     { X86::ROR16ri,     X86::ROR16mi,    0 },
189     { X86::ROR32r1,     X86::ROR32m1,    0 },
190     { X86::ROR32rCL,    X86::ROR32mCL,   0 },
191     { X86::ROR32ri,     X86::ROR32mi,    0 },
192     { X86::ROR64r1,     X86::ROR64m1,    0 },
193     { X86::ROR64rCL,    X86::ROR64mCL,   0 },
194     { X86::ROR64ri,     X86::ROR64mi,    0 },
195     { X86::ROR8r1,      X86::ROR8m1,     0 },
196     { X86::ROR8rCL,     X86::ROR8mCL,    0 },
197     { X86::ROR8ri,      X86::ROR8mi,     0 },
198     { X86::SAR16r1,     X86::SAR16m1,    0 },
199     { X86::SAR16rCL,    X86::SAR16mCL,   0 },
200     { X86::SAR16ri,     X86::SAR16mi,    0 },
201     { X86::SAR32r1,     X86::SAR32m1,    0 },
202     { X86::SAR32rCL,    X86::SAR32mCL,   0 },
203     { X86::SAR32ri,     X86::SAR32mi,    0 },
204     { X86::SAR64r1,     X86::SAR64m1,    0 },
205     { X86::SAR64rCL,    X86::SAR64mCL,   0 },
206     { X86::SAR64ri,     X86::SAR64mi,    0 },
207     { X86::SAR8r1,      X86::SAR8m1,     0 },
208     { X86::SAR8rCL,     X86::SAR8mCL,    0 },
209     { X86::SAR8ri,      X86::SAR8mi,     0 },
210     { X86::SBB32ri,     X86::SBB32mi,    0 },
211     { X86::SBB32ri8,    X86::SBB32mi8,   0 },
212     { X86::SBB32rr,     X86::SBB32mr,    0 },
213     { X86::SBB64ri32,   X86::SBB64mi32,  0 },
214     { X86::SBB64ri8,    X86::SBB64mi8,   0 },
215     { X86::SBB64rr,     X86::SBB64mr,    0 },
216     { X86::SHL16rCL,    X86::SHL16mCL,   0 },
217     { X86::SHL16ri,     X86::SHL16mi,    0 },
218     { X86::SHL32rCL,    X86::SHL32mCL,   0 },
219     { X86::SHL32ri,     X86::SHL32mi,    0 },
220     { X86::SHL64rCL,    X86::SHL64mCL,   0 },
221     { X86::SHL64ri,     X86::SHL64mi,    0 },
222     { X86::SHL8rCL,     X86::SHL8mCL,    0 },
223     { X86::SHL8ri,      X86::SHL8mi,     0 },
224     { X86::SHLD16rrCL,  X86::SHLD16mrCL, 0 },
225     { X86::SHLD16rri8,  X86::SHLD16mri8, 0 },
226     { X86::SHLD32rrCL,  X86::SHLD32mrCL, 0 },
227     { X86::SHLD32rri8,  X86::SHLD32mri8, 0 },
228     { X86::SHLD64rrCL,  X86::SHLD64mrCL, 0 },
229     { X86::SHLD64rri8,  X86::SHLD64mri8, 0 },
230     { X86::SHR16r1,     X86::SHR16m1,    0 },
231     { X86::SHR16rCL,    X86::SHR16mCL,   0 },
232     { X86::SHR16ri,     X86::SHR16mi,    0 },
233     { X86::SHR32r1,     X86::SHR32m1,    0 },
234     { X86::SHR32rCL,    X86::SHR32mCL,   0 },
235     { X86::SHR32ri,     X86::SHR32mi,    0 },
236     { X86::SHR64r1,     X86::SHR64m1,    0 },
237     { X86::SHR64rCL,    X86::SHR64mCL,   0 },
238     { X86::SHR64ri,     X86::SHR64mi,    0 },
239     { X86::SHR8r1,      X86::SHR8m1,     0 },
240     { X86::SHR8rCL,     X86::SHR8mCL,    0 },
241     { X86::SHR8ri,      X86::SHR8mi,     0 },
242     { X86::SHRD16rrCL,  X86::SHRD16mrCL, 0 },
243     { X86::SHRD16rri8,  X86::SHRD16mri8, 0 },
244     { X86::SHRD32rrCL,  X86::SHRD32mrCL, 0 },
245     { X86::SHRD32rri8,  X86::SHRD32mri8, 0 },
246     { X86::SHRD64rrCL,  X86::SHRD64mrCL, 0 },
247     { X86::SHRD64rri8,  X86::SHRD64mri8, 0 },
248     { X86::SUB16ri,     X86::SUB16mi,    0 },
249     { X86::SUB16ri8,    X86::SUB16mi8,   0 },
250     { X86::SUB16rr,     X86::SUB16mr,    0 },
251     { X86::SUB32ri,     X86::SUB32mi,    0 },
252     { X86::SUB32ri8,    X86::SUB32mi8,   0 },
253     { X86::SUB32rr,     X86::SUB32mr,    0 },
254     { X86::SUB64ri32,   X86::SUB64mi32,  0 },
255     { X86::SUB64ri8,    X86::SUB64mi8,   0 },
256     { X86::SUB64rr,     X86::SUB64mr,    0 },
257     { X86::SUB8ri,      X86::SUB8mi,     0 },
258     { X86::SUB8rr,      X86::SUB8mr,     0 },
259     { X86::XOR16ri,     X86::XOR16mi,    0 },
260     { X86::XOR16ri8,    X86::XOR16mi8,   0 },
261     { X86::XOR16rr,     X86::XOR16mr,    0 },
262     { X86::XOR32ri,     X86::XOR32mi,    0 },
263     { X86::XOR32ri8,    X86::XOR32mi8,   0 },
264     { X86::XOR32rr,     X86::XOR32mr,    0 },
265     { X86::XOR64ri32,   X86::XOR64mi32,  0 },
266     { X86::XOR64ri8,    X86::XOR64mi8,   0 },
267     { X86::XOR64rr,     X86::XOR64mr,    0 },
268     { X86::XOR8ri,      X86::XOR8mi,     0 },
269     { X86::XOR8rr,      X86::XOR8mr,     0 }
270   };
271 
272   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2Addr) {
273     AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
274                   Entry.RegOp, Entry.MemOp,
275                   // Index 0, folded load and store, no alignment requirement.
276                   Entry.Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
277   }
278 
279   static const X86MemoryFoldTableEntry MemoryFoldTable0[] = {
280     { X86::BT16ri8,     X86::BT16mi8,       TB_FOLDED_LOAD },
281     { X86::BT32ri8,     X86::BT32mi8,       TB_FOLDED_LOAD },
282     { X86::BT64ri8,     X86::BT64mi8,       TB_FOLDED_LOAD },
283     { X86::CALL32r,     X86::CALL32m,       TB_FOLDED_LOAD },
284     { X86::CALL64r,     X86::CALL64m,       TB_FOLDED_LOAD },
285     { X86::CMP16ri,     X86::CMP16mi,       TB_FOLDED_LOAD },
286     { X86::CMP16ri8,    X86::CMP16mi8,      TB_FOLDED_LOAD },
287     { X86::CMP16rr,     X86::CMP16mr,       TB_FOLDED_LOAD },
288     { X86::CMP32ri,     X86::CMP32mi,       TB_FOLDED_LOAD },
289     { X86::CMP32ri8,    X86::CMP32mi8,      TB_FOLDED_LOAD },
290     { X86::CMP32rr,     X86::CMP32mr,       TB_FOLDED_LOAD },
291     { X86::CMP64ri32,   X86::CMP64mi32,     TB_FOLDED_LOAD },
292     { X86::CMP64ri8,    X86::CMP64mi8,      TB_FOLDED_LOAD },
293     { X86::CMP64rr,     X86::CMP64mr,       TB_FOLDED_LOAD },
294     { X86::CMP8ri,      X86::CMP8mi,        TB_FOLDED_LOAD },
295     { X86::CMP8rr,      X86::CMP8mr,        TB_FOLDED_LOAD },
296     { X86::DIV16r,      X86::DIV16m,        TB_FOLDED_LOAD },
297     { X86::DIV32r,      X86::DIV32m,        TB_FOLDED_LOAD },
298     { X86::DIV64r,      X86::DIV64m,        TB_FOLDED_LOAD },
299     { X86::DIV8r,       X86::DIV8m,         TB_FOLDED_LOAD },
300     { X86::EXTRACTPSrr, X86::EXTRACTPSmr,   TB_FOLDED_STORE },
301     { X86::IDIV16r,     X86::IDIV16m,       TB_FOLDED_LOAD },
302     { X86::IDIV32r,     X86::IDIV32m,       TB_FOLDED_LOAD },
303     { X86::IDIV64r,     X86::IDIV64m,       TB_FOLDED_LOAD },
304     { X86::IDIV8r,      X86::IDIV8m,        TB_FOLDED_LOAD },
305     { X86::IMUL16r,     X86::IMUL16m,       TB_FOLDED_LOAD },
306     { X86::IMUL32r,     X86::IMUL32m,       TB_FOLDED_LOAD },
307     { X86::IMUL64r,     X86::IMUL64m,       TB_FOLDED_LOAD },
308     { X86::IMUL8r,      X86::IMUL8m,        TB_FOLDED_LOAD },
309     { X86::JMP32r,      X86::JMP32m,        TB_FOLDED_LOAD },
310     { X86::JMP64r,      X86::JMP64m,        TB_FOLDED_LOAD },
311     { X86::MOV16ri,     X86::MOV16mi,       TB_FOLDED_STORE },
312     { X86::MOV16rr,     X86::MOV16mr,       TB_FOLDED_STORE },
313     { X86::MOV32ri,     X86::MOV32mi,       TB_FOLDED_STORE },
314     { X86::MOV32rr,     X86::MOV32mr,       TB_FOLDED_STORE },
315     { X86::MOV64ri32,   X86::MOV64mi32,     TB_FOLDED_STORE },
316     { X86::MOV64rr,     X86::MOV64mr,       TB_FOLDED_STORE },
317     { X86::MOV8ri,      X86::MOV8mi,        TB_FOLDED_STORE },
318     { X86::MOV8rr,      X86::MOV8mr,        TB_FOLDED_STORE },
319     { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
320     { X86::MOVAPDrr,    X86::MOVAPDmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
321     { X86::MOVAPSrr,    X86::MOVAPSmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
322     { X86::MOVDQArr,    X86::MOVDQAmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
323     { X86::MOVPDI2DIrr, X86::MOVPDI2DImr,   TB_FOLDED_STORE },
324     { X86::MOVPQIto64rr,X86::MOVPQI2QImr,   TB_FOLDED_STORE },
325     { X86::MOVSDto64rr, X86::MOVSDto64mr,   TB_FOLDED_STORE },
326     { X86::MOVSS2DIrr,  X86::MOVSS2DImr,    TB_FOLDED_STORE },
327     { X86::MOVUPDrr,    X86::MOVUPDmr,      TB_FOLDED_STORE },
328     { X86::MOVUPSrr,    X86::MOVUPSmr,      TB_FOLDED_STORE },
329     { X86::MUL16r,      X86::MUL16m,        TB_FOLDED_LOAD },
330     { X86::MUL32r,      X86::MUL32m,        TB_FOLDED_LOAD },
331     { X86::MUL64r,      X86::MUL64m,        TB_FOLDED_LOAD },
332     { X86::MUL8r,       X86::MUL8m,         TB_FOLDED_LOAD },
333     { X86::PEXTRDrr,    X86::PEXTRDmr,      TB_FOLDED_STORE },
334     { X86::PEXTRQrr,    X86::PEXTRQmr,      TB_FOLDED_STORE },
335     { X86::SETAEr,      X86::SETAEm,        TB_FOLDED_STORE },
336     { X86::SETAr,       X86::SETAm,         TB_FOLDED_STORE },
337     { X86::SETBEr,      X86::SETBEm,        TB_FOLDED_STORE },
338     { X86::SETBr,       X86::SETBm,         TB_FOLDED_STORE },
339     { X86::SETEr,       X86::SETEm,         TB_FOLDED_STORE },
340     { X86::SETGEr,      X86::SETGEm,        TB_FOLDED_STORE },
341     { X86::SETGr,       X86::SETGm,         TB_FOLDED_STORE },
342     { X86::SETLEr,      X86::SETLEm,        TB_FOLDED_STORE },
343     { X86::SETLr,       X86::SETLm,         TB_FOLDED_STORE },
344     { X86::SETNEr,      X86::SETNEm,        TB_FOLDED_STORE },
345     { X86::SETNOr,      X86::SETNOm,        TB_FOLDED_STORE },
346     { X86::SETNPr,      X86::SETNPm,        TB_FOLDED_STORE },
347     { X86::SETNSr,      X86::SETNSm,        TB_FOLDED_STORE },
348     { X86::SETOr,       X86::SETOm,         TB_FOLDED_STORE },
349     { X86::SETPr,       X86::SETPm,         TB_FOLDED_STORE },
350     { X86::SETSr,       X86::SETSm,         TB_FOLDED_STORE },
351     { X86::TAILJMPr,    X86::TAILJMPm,      TB_FOLDED_LOAD },
352     { X86::TAILJMPr64,  X86::TAILJMPm64,    TB_FOLDED_LOAD },
353     { X86::TAILJMPr64_REX, X86::TAILJMPm64_REX, TB_FOLDED_LOAD },
354     { X86::TEST16ri,    X86::TEST16mi,      TB_FOLDED_LOAD },
355     { X86::TEST32ri,    X86::TEST32mi,      TB_FOLDED_LOAD },
356     { X86::TEST64ri32,  X86::TEST64mi32,    TB_FOLDED_LOAD },
357     { X86::TEST8ri,     X86::TEST8mi,       TB_FOLDED_LOAD },
358 
359     // AVX 128-bit versions of foldable instructions
360     { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr,  TB_FOLDED_STORE  },
361     { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
362     { X86::VMOVAPDrr,   X86::VMOVAPDmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
363     { X86::VMOVAPSrr,   X86::VMOVAPSmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
364     { X86::VMOVDQArr,   X86::VMOVDQAmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
365     { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr,  TB_FOLDED_STORE },
366     { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367     { X86::VMOVSDto64rr,X86::VMOVSDto64mr,  TB_FOLDED_STORE },
368     { X86::VMOVSS2DIrr, X86::VMOVSS2DImr,   TB_FOLDED_STORE },
369     { X86::VMOVUPDrr,   X86::VMOVUPDmr,     TB_FOLDED_STORE },
370     { X86::VMOVUPSrr,   X86::VMOVUPSmr,     TB_FOLDED_STORE },
371     { X86::VPEXTRDrr,   X86::VPEXTRDmr,     TB_FOLDED_STORE },
372     { X86::VPEXTRQrr,   X86::VPEXTRQmr,     TB_FOLDED_STORE },
373 
374     // AVX 256-bit foldable instructions
375     { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
376     { X86::VMOVAPDYrr,  X86::VMOVAPDYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
377     { X86::VMOVAPSYrr,  X86::VMOVAPSYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
378     { X86::VMOVDQAYrr,  X86::VMOVDQAYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
379     { X86::VMOVUPDYrr,  X86::VMOVUPDYmr,    TB_FOLDED_STORE },
380     { X86::VMOVUPSYrr,  X86::VMOVUPSYmr,    TB_FOLDED_STORE },
381 
382     // AVX-512 foldable instructions
383     { X86::VMOVPDI2DIZrr,   X86::VMOVPDI2DIZmr, TB_FOLDED_STORE },
384     { X86::VMOVAPDZrr,      X86::VMOVAPDZmr,    TB_FOLDED_STORE | TB_ALIGN_64 },
385     { X86::VMOVAPSZrr,      X86::VMOVAPSZmr,    TB_FOLDED_STORE | TB_ALIGN_64 },
386     { X86::VMOVDQA32Zrr,    X86::VMOVDQA32Zmr,  TB_FOLDED_STORE | TB_ALIGN_64 },
387     { X86::VMOVDQA64Zrr,    X86::VMOVDQA64Zmr,  TB_FOLDED_STORE | TB_ALIGN_64 },
388     { X86::VMOVUPDZrr,      X86::VMOVUPDZmr,    TB_FOLDED_STORE },
389     { X86::VMOVUPSZrr,      X86::VMOVUPSZmr,    TB_FOLDED_STORE },
390     { X86::VMOVDQU8Zrr,     X86::VMOVDQU8Zmr,   TB_FOLDED_STORE },
391     { X86::VMOVDQU16Zrr,    X86::VMOVDQU16Zmr,  TB_FOLDED_STORE },
392     { X86::VMOVDQU32Zrr,    X86::VMOVDQU32Zmr,  TB_FOLDED_STORE },
393     { X86::VMOVDQU64Zrr,    X86::VMOVDQU64Zmr,  TB_FOLDED_STORE },
394 
395     // AVX-512 foldable instructions (256-bit versions)
396     { X86::VMOVAPDZ256rr,      X86::VMOVAPDZ256mr,    TB_FOLDED_STORE | TB_ALIGN_32 },
397     { X86::VMOVAPSZ256rr,      X86::VMOVAPSZ256mr,    TB_FOLDED_STORE | TB_ALIGN_32 },
398     { X86::VMOVDQA32Z256rr,    X86::VMOVDQA32Z256mr,  TB_FOLDED_STORE | TB_ALIGN_32 },
399     { X86::VMOVDQA64Z256rr,    X86::VMOVDQA64Z256mr,  TB_FOLDED_STORE | TB_ALIGN_32 },
400     { X86::VMOVUPDZ256rr,      X86::VMOVUPDZ256mr,    TB_FOLDED_STORE },
401     { X86::VMOVUPSZ256rr,      X86::VMOVUPSZ256mr,    TB_FOLDED_STORE },
402     { X86::VMOVDQU8Z256rr,     X86::VMOVDQU8Z256mr,   TB_FOLDED_STORE },
403     { X86::VMOVDQU16Z256rr,    X86::VMOVDQU16Z256mr,  TB_FOLDED_STORE },
404     { X86::VMOVDQU32Z256rr,    X86::VMOVDQU32Z256mr,  TB_FOLDED_STORE },
405     { X86::VMOVDQU64Z256rr,    X86::VMOVDQU64Z256mr,  TB_FOLDED_STORE },
406 
407     // AVX-512 foldable instructions (128-bit versions)
408     { X86::VMOVAPDZ128rr,      X86::VMOVAPDZ128mr,    TB_FOLDED_STORE | TB_ALIGN_16 },
409     { X86::VMOVAPSZ128rr,      X86::VMOVAPSZ128mr,    TB_FOLDED_STORE | TB_ALIGN_16 },
410     { X86::VMOVDQA32Z128rr,    X86::VMOVDQA32Z128mr,  TB_FOLDED_STORE | TB_ALIGN_16 },
411     { X86::VMOVDQA64Z128rr,    X86::VMOVDQA64Z128mr,  TB_FOLDED_STORE | TB_ALIGN_16 },
412     { X86::VMOVUPDZ128rr,      X86::VMOVUPDZ128mr,    TB_FOLDED_STORE },
413     { X86::VMOVUPSZ128rr,      X86::VMOVUPSZ128mr,    TB_FOLDED_STORE },
414     { X86::VMOVDQU8Z128rr,     X86::VMOVDQU8Z128mr,   TB_FOLDED_STORE },
415     { X86::VMOVDQU16Z128rr,    X86::VMOVDQU16Z128mr,  TB_FOLDED_STORE },
416     { X86::VMOVDQU32Z128rr,    X86::VMOVDQU32Z128mr,  TB_FOLDED_STORE },
417     { X86::VMOVDQU64Z128rr,    X86::VMOVDQU64Z128mr,  TB_FOLDED_STORE },
418 
419     // F16C foldable instructions
420     { X86::VCVTPS2PHrr,        X86::VCVTPS2PHmr,      TB_FOLDED_STORE },
421     { X86::VCVTPS2PHYrr,       X86::VCVTPS2PHYmr,     TB_FOLDED_STORE }
422   };
423 
424   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable0) {
425     AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
426                   Entry.RegOp, Entry.MemOp, TB_INDEX_0 | Entry.Flags);
427   }
428 
429   static const X86MemoryFoldTableEntry MemoryFoldTable1[] = {
430     { X86::BSF16rr,         X86::BSF16rm,             0 },
431     { X86::BSF32rr,         X86::BSF32rm,             0 },
432     { X86::BSF64rr,         X86::BSF64rm,             0 },
433     { X86::BSR16rr,         X86::BSR16rm,             0 },
434     { X86::BSR32rr,         X86::BSR32rm,             0 },
435     { X86::BSR64rr,         X86::BSR64rm,             0 },
436     { X86::CMP16rr,         X86::CMP16rm,             0 },
437     { X86::CMP32rr,         X86::CMP32rm,             0 },
438     { X86::CMP64rr,         X86::CMP64rm,             0 },
439     { X86::CMP8rr,          X86::CMP8rm,              0 },
440     { X86::CVTSD2SSrr,      X86::CVTSD2SSrm,          0 },
441     { X86::CVTSI2SD64rr,    X86::CVTSI2SD64rm,        0 },
442     { X86::CVTSI2SDrr,      X86::CVTSI2SDrm,          0 },
443     { X86::CVTSI2SS64rr,    X86::CVTSI2SS64rm,        0 },
444     { X86::CVTSI2SSrr,      X86::CVTSI2SSrm,          0 },
445     { X86::CVTSS2SDrr,      X86::CVTSS2SDrm,          0 },
446     { X86::CVTTSD2SI64rr,   X86::CVTTSD2SI64rm,       0 },
447     { X86::CVTTSD2SIrr,     X86::CVTTSD2SIrm,         0 },
448     { X86::CVTTSS2SI64rr,   X86::CVTTSS2SI64rm,       0 },
449     { X86::CVTTSS2SIrr,     X86::CVTTSS2SIrm,         0 },
450     { X86::IMUL16rri,       X86::IMUL16rmi,           0 },
451     { X86::IMUL16rri8,      X86::IMUL16rmi8,          0 },
452     { X86::IMUL32rri,       X86::IMUL32rmi,           0 },
453     { X86::IMUL32rri8,      X86::IMUL32rmi8,          0 },
454     { X86::IMUL64rri32,     X86::IMUL64rmi32,         0 },
455     { X86::IMUL64rri8,      X86::IMUL64rmi8,          0 },
456     { X86::Int_COMISDrr,    X86::Int_COMISDrm,        0 },
457     { X86::Int_COMISSrr,    X86::Int_COMISSrm,        0 },
458     { X86::CVTSD2SI64rr,    X86::CVTSD2SI64rm,        0 },
459     { X86::CVTSD2SIrr,      X86::CVTSD2SIrm,          0 },
460     { X86::CVTSS2SI64rr,    X86::CVTSS2SI64rm,        0 },
461     { X86::CVTSS2SIrr,      X86::CVTSS2SIrm,          0 },
462     { X86::CVTDQ2PDrr,      X86::CVTDQ2PDrm,          TB_ALIGN_16 },
463     { X86::CVTDQ2PSrr,      X86::CVTDQ2PSrm,          TB_ALIGN_16 },
464     { X86::CVTPD2DQrr,      X86::CVTPD2DQrm,          TB_ALIGN_16 },
465     { X86::CVTPD2PSrr,      X86::CVTPD2PSrm,          TB_ALIGN_16 },
466     { X86::CVTPS2DQrr,      X86::CVTPS2DQrm,          TB_ALIGN_16 },
467     { X86::CVTPS2PDrr,      X86::CVTPS2PDrm,          TB_ALIGN_16 },
468     { X86::CVTTPD2DQrr,     X86::CVTTPD2DQrm,         TB_ALIGN_16 },
469     { X86::CVTTPS2DQrr,     X86::CVTTPS2DQrm,         TB_ALIGN_16 },
470     { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm,  0 },
471     { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm,     0 },
472     { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm,  0 },
473     { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm,     0 },
474     { X86::Int_UCOMISDrr,   X86::Int_UCOMISDrm,       0 },
475     { X86::Int_UCOMISSrr,   X86::Int_UCOMISSrm,       0 },
476     { X86::MOV16rr,         X86::MOV16rm,             0 },
477     { X86::MOV32rr,         X86::MOV32rm,             0 },
478     { X86::MOV64rr,         X86::MOV64rm,             0 },
479     { X86::MOV64toPQIrr,    X86::MOVQI2PQIrm,         0 },
480     { X86::MOV64toSDrr,     X86::MOV64toSDrm,         0 },
481     { X86::MOV8rr,          X86::MOV8rm,              0 },
482     { X86::MOVAPDrr,        X86::MOVAPDrm,            TB_ALIGN_16 },
483     { X86::MOVAPSrr,        X86::MOVAPSrm,            TB_ALIGN_16 },
484     { X86::MOVDDUPrr,       X86::MOVDDUPrm,           0 },
485     { X86::MOVDI2PDIrr,     X86::MOVDI2PDIrm,         0 },
486     { X86::MOVDI2SSrr,      X86::MOVDI2SSrm,          0 },
487     { X86::MOVDQArr,        X86::MOVDQArm,            TB_ALIGN_16 },
488     { X86::MOVSHDUPrr,      X86::MOVSHDUPrm,          TB_ALIGN_16 },
489     { X86::MOVSLDUPrr,      X86::MOVSLDUPrm,          TB_ALIGN_16 },
490     { X86::MOVSX16rr8,      X86::MOVSX16rm8,          0 },
491     { X86::MOVSX32rr16,     X86::MOVSX32rm16,         0 },
492     { X86::MOVSX32rr8,      X86::MOVSX32rm8,          0 },
493     { X86::MOVSX64rr16,     X86::MOVSX64rm16,         0 },
494     { X86::MOVSX64rr32,     X86::MOVSX64rm32,         0 },
495     { X86::MOVSX64rr8,      X86::MOVSX64rm8,          0 },
496     { X86::MOVUPDrr,        X86::MOVUPDrm,            TB_ALIGN_16 },
497     { X86::MOVUPSrr,        X86::MOVUPSrm,            0 },
498     { X86::MOVZQI2PQIrr,    X86::MOVZQI2PQIrm,        0 },
499     { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm,     TB_ALIGN_16 },
500     { X86::MOVZX16rr8,      X86::MOVZX16rm8,          0 },
501     { X86::MOVZX32rr16,     X86::MOVZX32rm16,         0 },
502     { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8,   0 },
503     { X86::MOVZX32rr8,      X86::MOVZX32rm8,          0 },
504     { X86::PABSBrr128,      X86::PABSBrm128,          TB_ALIGN_16 },
505     { X86::PABSDrr128,      X86::PABSDrm128,          TB_ALIGN_16 },
506     { X86::PABSWrr128,      X86::PABSWrm128,          TB_ALIGN_16 },
507     { X86::PCMPESTRIrr,     X86::PCMPESTRIrm,         TB_ALIGN_16 },
508     { X86::PCMPESTRM128rr,  X86::PCMPESTRM128rm,      TB_ALIGN_16 },
509     { X86::PCMPISTRIrr,     X86::PCMPISTRIrm,         TB_ALIGN_16 },
510     { X86::PCMPISTRM128rr,  X86::PCMPISTRM128rm,      TB_ALIGN_16 },
511     { X86::PHMINPOSUWrr128, X86::PHMINPOSUWrm128,     TB_ALIGN_16 },
512     { X86::PMOVSXBDrr,      X86::PMOVSXBDrm,          TB_ALIGN_16 },
513     { X86::PMOVSXBQrr,      X86::PMOVSXBQrm,          TB_ALIGN_16 },
514     { X86::PMOVSXBWrr,      X86::PMOVSXBWrm,          TB_ALIGN_16 },
515     { X86::PMOVSXDQrr,      X86::PMOVSXDQrm,          TB_ALIGN_16 },
516     { X86::PMOVSXWDrr,      X86::PMOVSXWDrm,          TB_ALIGN_16 },
517     { X86::PMOVSXWQrr,      X86::PMOVSXWQrm,          TB_ALIGN_16 },
518     { X86::PMOVZXBDrr,      X86::PMOVZXBDrm,          TB_ALIGN_16 },
519     { X86::PMOVZXBQrr,      X86::PMOVZXBQrm,          TB_ALIGN_16 },
520     { X86::PMOVZXBWrr,      X86::PMOVZXBWrm,          TB_ALIGN_16 },
521     { X86::PMOVZXDQrr,      X86::PMOVZXDQrm,          TB_ALIGN_16 },
522     { X86::PMOVZXWDrr,      X86::PMOVZXWDrm,          TB_ALIGN_16 },
523     { X86::PMOVZXWQrr,      X86::PMOVZXWQrm,          TB_ALIGN_16 },
524     { X86::PSHUFDri,        X86::PSHUFDmi,            TB_ALIGN_16 },
525     { X86::PSHUFHWri,       X86::PSHUFHWmi,           TB_ALIGN_16 },
526     { X86::PSHUFLWri,       X86::PSHUFLWmi,           TB_ALIGN_16 },
527     { X86::PTESTrr,         X86::PTESTrm,             TB_ALIGN_16 },
528     { X86::RCPPSr,          X86::RCPPSm,              TB_ALIGN_16 },
529     { X86::RCPSSr,          X86::RCPSSm,              0 },
530     { X86::RCPSSr_Int,      X86::RCPSSm_Int,          0 },
531     { X86::ROUNDPDr,        X86::ROUNDPDm,            TB_ALIGN_16 },
532     { X86::ROUNDPSr,        X86::ROUNDPSm,            TB_ALIGN_16 },
533     { X86::RSQRTPSr,        X86::RSQRTPSm,            TB_ALIGN_16 },
534     { X86::RSQRTSSr,        X86::RSQRTSSm,            0 },
535     { X86::RSQRTSSr_Int,    X86::RSQRTSSm_Int,        0 },
536     { X86::SQRTPDr,         X86::SQRTPDm,             TB_ALIGN_16 },
537     { X86::SQRTPSr,         X86::SQRTPSm,             TB_ALIGN_16 },
538     { X86::SQRTSDr,         X86::SQRTSDm,             0 },
539     { X86::SQRTSDr_Int,     X86::SQRTSDm_Int,         0 },
540     { X86::SQRTSSr,         X86::SQRTSSm,             0 },
541     { X86::SQRTSSr_Int,     X86::SQRTSSm_Int,         0 },
542     { X86::TEST16rr,        X86::TEST16rm,            0 },
543     { X86::TEST32rr,        X86::TEST32rm,            0 },
544     { X86::TEST64rr,        X86::TEST64rm,            0 },
545     { X86::TEST8rr,         X86::TEST8rm,             0 },
546     // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
547     { X86::UCOMISDrr,       X86::UCOMISDrm,           0 },
548     { X86::UCOMISSrr,       X86::UCOMISSrm,           0 },
549 
550     // MMX version of foldable instructions
551     { X86::MMX_CVTPD2PIirr,   X86::MMX_CVTPD2PIirm,   0 },
552     { X86::MMX_CVTPI2PDirr,   X86::MMX_CVTPI2PDirm,   0 },
553     { X86::MMX_CVTPS2PIirr,   X86::MMX_CVTPS2PIirm,   0 },
554     { X86::MMX_CVTTPD2PIirr,  X86::MMX_CVTTPD2PIirm,  0 },
555     { X86::MMX_CVTTPS2PIirr,  X86::MMX_CVTTPS2PIirm,  0 },
556     { X86::MMX_MOVD64to64rr,  X86::MMX_MOVQ64rm,      0 },
557     { X86::MMX_PABSBrr64,     X86::MMX_PABSBrm64,     0 },
558     { X86::MMX_PABSDrr64,     X86::MMX_PABSDrm64,     0 },
559     { X86::MMX_PABSWrr64,     X86::MMX_PABSWrm64,     0 },
560     { X86::MMX_PSHUFWri,      X86::MMX_PSHUFWmi,      0 },
561 
562     // 3DNow! version of foldable instructions
563     { X86::PF2IDrr,         X86::PF2IDrm,             0 },
564     { X86::PF2IWrr,         X86::PF2IWrm,             0 },
565     { X86::PFRCPrr,         X86::PFRCPrm,             0 },
566     { X86::PFRSQRTrr,       X86::PFRSQRTrm,           0 },
567     { X86::PI2FDrr,         X86::PI2FDrm,             0 },
568     { X86::PI2FWrr,         X86::PI2FWrm,             0 },
569     { X86::PSWAPDrr,        X86::PSWAPDrm,            0 },
570 
571     // AVX 128-bit versions of foldable instructions
572     { X86::Int_VCOMISDrr,   X86::Int_VCOMISDrm,       0 },
573     { X86::Int_VCOMISSrr,   X86::Int_VCOMISSrm,       0 },
574     { X86::Int_VUCOMISDrr,  X86::Int_VUCOMISDrm,      0 },
575     { X86::Int_VUCOMISSrr,  X86::Int_VUCOMISSrm,      0 },
576     { X86::VCVTTSD2SI64rr,  X86::VCVTTSD2SI64rm,      0 },
577     { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
578     { X86::VCVTTSD2SIrr,    X86::VCVTTSD2SIrm,        0 },
579     { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm,    0 },
580     { X86::VCVTTSS2SI64rr,  X86::VCVTTSS2SI64rm,      0 },
581     { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
582     { X86::VCVTTSS2SIrr,    X86::VCVTTSS2SIrm,        0 },
583     { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm,    0 },
584     { X86::VCVTSD2SI64rr,   X86::VCVTSD2SI64rm,       0 },
585     { X86::VCVTSD2SIrr,     X86::VCVTSD2SIrm,         0 },
586     { X86::VCVTSS2SI64rr,   X86::VCVTSS2SI64rm,       0 },
587     { X86::VCVTSS2SIrr,     X86::VCVTSS2SIrm,         0 },
588     { X86::VCVTDQ2PDrr,     X86::VCVTDQ2PDrm,         0 },
589     { X86::VCVTDQ2PSrr,     X86::VCVTDQ2PSrm,         0 },
590     { X86::VCVTPD2DQrr,     X86::VCVTPD2DQXrm,        0 },
591     { X86::VCVTPD2PSrr,     X86::VCVTPD2PSXrm,        0 },
592     { X86::VCVTPS2DQrr,     X86::VCVTPS2DQrm,         0 },
593     { X86::VCVTPS2PDrr,     X86::VCVTPS2PDrm,         0 },
594     { X86::VCVTTPD2DQrr,    X86::VCVTTPD2DQXrm,       0 },
595     { X86::VCVTTPS2DQrr,    X86::VCVTTPS2DQrm,        0 },
596     { X86::VMOV64toPQIrr,   X86::VMOVQI2PQIrm,        0 },
597     { X86::VMOV64toSDrr,    X86::VMOV64toSDrm,        0 },
598     { X86::VMOVAPDrr,       X86::VMOVAPDrm,           TB_ALIGN_16 },
599     { X86::VMOVAPSrr,       X86::VMOVAPSrm,           TB_ALIGN_16 },
600     { X86::VMOVDDUPrr,      X86::VMOVDDUPrm,          0 },
601     { X86::VMOVDI2PDIrr,    X86::VMOVDI2PDIrm,        0 },
602     { X86::VMOVDI2SSrr,     X86::VMOVDI2SSrm,         0 },
603     { X86::VMOVDQArr,       X86::VMOVDQArm,           TB_ALIGN_16 },
604     { X86::VMOVSLDUPrr,     X86::VMOVSLDUPrm,         0 },
605     { X86::VMOVSHDUPrr,     X86::VMOVSHDUPrm,         0 },
606     { X86::VMOVUPDrr,       X86::VMOVUPDrm,           0 },
607     { X86::VMOVUPSrr,       X86::VMOVUPSrm,           0 },
608     { X86::VMOVZQI2PQIrr,   X86::VMOVZQI2PQIrm,       0 },
609     { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm,    TB_ALIGN_16 },
610     { X86::VPABSBrr128,     X86::VPABSBrm128,         0 },
611     { X86::VPABSDrr128,     X86::VPABSDrm128,         0 },
612     { X86::VPABSWrr128,     X86::VPABSWrm128,         0 },
613     { X86::VPCMPESTRIrr,    X86::VPCMPESTRIrm,        0 },
614     { X86::VPCMPESTRM128rr, X86::VPCMPESTRM128rm,     0 },
615     { X86::VPCMPISTRIrr,    X86::VPCMPISTRIrm,        0 },
616     { X86::VPCMPISTRM128rr, X86::VPCMPISTRM128rm,     0 },
617     { X86::VPHMINPOSUWrr128, X86::VPHMINPOSUWrm128,   0 },
618     { X86::VPERMILPDri,     X86::VPERMILPDmi,         0 },
619     { X86::VPERMILPSri,     X86::VPERMILPSmi,         0 },
620     { X86::VPMOVSXBDrr,     X86::VPMOVSXBDrm,         0 },
621     { X86::VPMOVSXBQrr,     X86::VPMOVSXBQrm,         0 },
622     { X86::VPMOVSXBWrr,     X86::VPMOVSXBWrm,         0 },
623     { X86::VPMOVSXDQrr,     X86::VPMOVSXDQrm,         0 },
624     { X86::VPMOVSXWDrr,     X86::VPMOVSXWDrm,         0 },
625     { X86::VPMOVSXWQrr,     X86::VPMOVSXWQrm,         0 },
626     { X86::VPMOVZXBDrr,     X86::VPMOVZXBDrm,         0 },
627     { X86::VPMOVZXBQrr,     X86::VPMOVZXBQrm,         0 },
628     { X86::VPMOVZXBWrr,     X86::VPMOVZXBWrm,         0 },
629     { X86::VPMOVZXDQrr,     X86::VPMOVZXDQrm,         0 },
630     { X86::VPMOVZXWDrr,     X86::VPMOVZXWDrm,         0 },
631     { X86::VPMOVZXWQrr,     X86::VPMOVZXWQrm,         0 },
632     { X86::VPSHUFDri,       X86::VPSHUFDmi,           0 },
633     { X86::VPSHUFHWri,      X86::VPSHUFHWmi,          0 },
634     { X86::VPSHUFLWri,      X86::VPSHUFLWmi,          0 },
635     { X86::VPTESTrr,        X86::VPTESTrm,            0 },
636     { X86::VRCPPSr,         X86::VRCPPSm,             0 },
637     { X86::VROUNDPDr,       X86::VROUNDPDm,           0 },
638     { X86::VROUNDPSr,       X86::VROUNDPSm,           0 },
639     { X86::VRSQRTPSr,       X86::VRSQRTPSm,           0 },
640     { X86::VSQRTPDr,        X86::VSQRTPDm,            0 },
641     { X86::VSQRTPSr,        X86::VSQRTPSm,            0 },
642     { X86::VTESTPDrr,       X86::VTESTPDrm,           0 },
643     { X86::VTESTPSrr,       X86::VTESTPSrm,           0 },
644     { X86::VUCOMISDrr,      X86::VUCOMISDrm,          0 },
645     { X86::VUCOMISSrr,      X86::VUCOMISSrm,          0 },
646 
647     // AVX 256-bit foldable instructions
648     { X86::VCVTDQ2PDYrr,    X86::VCVTDQ2PDYrm,        0 },
649     { X86::VCVTDQ2PSYrr,    X86::VCVTDQ2PSYrm,        0 },
650     { X86::VCVTPD2DQYrr,    X86::VCVTPD2DQYrm,        0 },
651     { X86::VCVTPD2PSYrr,    X86::VCVTPD2PSYrm,        0 },
652     { X86::VCVTPS2DQYrr,    X86::VCVTPS2DQYrm,        0 },
653     { X86::VCVTPS2PDYrr,    X86::VCVTPS2PDYrm,        0 },
654     { X86::VCVTTPD2DQYrr,   X86::VCVTTPD2DQYrm,       0 },
655     { X86::VCVTTPS2DQYrr,   X86::VCVTTPS2DQYrm,       0 },
656     { X86::VMOVAPDYrr,      X86::VMOVAPDYrm,          TB_ALIGN_32 },
657     { X86::VMOVAPSYrr,      X86::VMOVAPSYrm,          TB_ALIGN_32 },
658     { X86::VMOVDDUPYrr,     X86::VMOVDDUPYrm,         0 },
659     { X86::VMOVDQAYrr,      X86::VMOVDQAYrm,          TB_ALIGN_32 },
660     { X86::VMOVSLDUPYrr,    X86::VMOVSLDUPYrm,        0 },
661     { X86::VMOVSHDUPYrr,    X86::VMOVSHDUPYrm,        0 },
662     { X86::VMOVUPDYrr,      X86::VMOVUPDYrm,          0 },
663     { X86::VMOVUPSYrr,      X86::VMOVUPSYrm,          0 },
664     { X86::VPERMILPDYri,    X86::VPERMILPDYmi,        0 },
665     { X86::VPERMILPSYri,    X86::VPERMILPSYmi,        0 },
666     { X86::VPTESTYrr,       X86::VPTESTYrm,           0 },
667     { X86::VRCPPSYr,        X86::VRCPPSYm,            0 },
668     { X86::VROUNDYPDr,      X86::VROUNDYPDm,          0 },
669     { X86::VROUNDYPSr,      X86::VROUNDYPSm,          0 },
670     { X86::VRSQRTPSYr,      X86::VRSQRTPSYm,          0 },
671     { X86::VSQRTPDYr,       X86::VSQRTPDYm,           0 },
672     { X86::VSQRTPSYr,       X86::VSQRTPSYm,           0 },
673     { X86::VTESTPDYrr,      X86::VTESTPDYrm,          0 },
674     { X86::VTESTPSYrr,      X86::VTESTPSYrm,          0 },
675 
676     // AVX2 foldable instructions
677 
678     // VBROADCASTS{SD}rr register instructions were an AVX2 addition while the
679     // VBROADCASTS{SD}rm memory instructions were available from AVX1.
680     // TB_NO_REVERSE prevents unfolding from introducing an illegal instruction
681     // on AVX1 targets. The VPBROADCAST instructions are all AVX2 instructions
682     // so they don't need an equivalent limitation.
683     { X86::VBROADCASTSSrr,  X86::VBROADCASTSSrm,      TB_NO_REVERSE },
684     { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm,     TB_NO_REVERSE },
685     { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm,     TB_NO_REVERSE },
686     { X86::VPABSBrr256,     X86::VPABSBrm256,         0 },
687     { X86::VPABSDrr256,     X86::VPABSDrm256,         0 },
688     { X86::VPABSWrr256,     X86::VPABSWrm256,         0 },
689     { X86::VPBROADCASTBrr,  X86::VPBROADCASTBrm,      0 },
690     { X86::VPBROADCASTBYrr, X86::VPBROADCASTBYrm,     0 },
691     { X86::VPBROADCASTDrr,  X86::VPBROADCASTDrm,      0 },
692     { X86::VPBROADCASTDYrr, X86::VPBROADCASTDYrm,     0 },
693     { X86::VPBROADCASTQrr,  X86::VPBROADCASTQrm,      0 },
694     { X86::VPBROADCASTQYrr, X86::VPBROADCASTQYrm,     0 },
695     { X86::VPBROADCASTWrr,  X86::VPBROADCASTWrm,      0 },
696     { X86::VPBROADCASTWYrr, X86::VPBROADCASTWYrm,     0 },
697     { X86::VPERMPDYri,      X86::VPERMPDYmi,          0 },
698     { X86::VPERMQYri,       X86::VPERMQYmi,           0 },
699     { X86::VPMOVSXBDYrr,    X86::VPMOVSXBDYrm,        0 },
700     { X86::VPMOVSXBQYrr,    X86::VPMOVSXBQYrm,        0 },
701     { X86::VPMOVSXBWYrr,    X86::VPMOVSXBWYrm,        0 },
702     { X86::VPMOVSXDQYrr,    X86::VPMOVSXDQYrm,        0 },
703     { X86::VPMOVSXWDYrr,    X86::VPMOVSXWDYrm,        0 },
704     { X86::VPMOVSXWQYrr,    X86::VPMOVSXWQYrm,        0 },
705     { X86::VPMOVZXBDYrr,    X86::VPMOVZXBDYrm,        0 },
706     { X86::VPMOVZXBQYrr,    X86::VPMOVZXBQYrm,        0 },
707     { X86::VPMOVZXBWYrr,    X86::VPMOVZXBWYrm,        0 },
708     { X86::VPMOVZXDQYrr,    X86::VPMOVZXDQYrm,        0 },
709     { X86::VPMOVZXWDYrr,    X86::VPMOVZXWDYrm,        0 },
710     { X86::VPMOVZXWQYrr,    X86::VPMOVZXWQYrm,        0 },
711     { X86::VPSHUFDYri,      X86::VPSHUFDYmi,          0 },
712     { X86::VPSHUFHWYri,     X86::VPSHUFHWYmi,         0 },
713     { X86::VPSHUFLWYri,     X86::VPSHUFLWYmi,         0 },
714 
715     // XOP foldable instructions
716     { X86::VFRCZPDrr,          X86::VFRCZPDrm,        0 },
717     { X86::VFRCZPDrrY,         X86::VFRCZPDrmY,       0 },
718     { X86::VFRCZPSrr,          X86::VFRCZPSrm,        0 },
719     { X86::VFRCZPSrrY,         X86::VFRCZPSrmY,       0 },
720     { X86::VFRCZSDrr,          X86::VFRCZSDrm,        0 },
721     { X86::VFRCZSSrr,          X86::VFRCZSSrm,        0 },
722     { X86::VPHADDBDrr,         X86::VPHADDBDrm,       0 },
723     { X86::VPHADDBQrr,         X86::VPHADDBQrm,       0 },
724     { X86::VPHADDBWrr,         X86::VPHADDBWrm,       0 },
725     { X86::VPHADDDQrr,         X86::VPHADDDQrm,       0 },
726     { X86::VPHADDWDrr,         X86::VPHADDWDrm,       0 },
727     { X86::VPHADDWQrr,         X86::VPHADDWQrm,       0 },
728     { X86::VPHADDUBDrr,        X86::VPHADDUBDrm,      0 },
729     { X86::VPHADDUBQrr,        X86::VPHADDUBQrm,      0 },
730     { X86::VPHADDUBWrr,        X86::VPHADDUBWrm,      0 },
731     { X86::VPHADDUDQrr,        X86::VPHADDUDQrm,      0 },
732     { X86::VPHADDUWDrr,        X86::VPHADDUWDrm,      0 },
733     { X86::VPHADDUWQrr,        X86::VPHADDUWQrm,      0 },
734     { X86::VPHSUBBWrr,         X86::VPHSUBBWrm,       0 },
735     { X86::VPHSUBDQrr,         X86::VPHSUBDQrm,       0 },
736     { X86::VPHSUBWDrr,         X86::VPHSUBWDrm,       0 },
737     { X86::VPROTBri,           X86::VPROTBmi,         0 },
738     { X86::VPROTBrr,           X86::VPROTBmr,         0 },
739     { X86::VPROTDri,           X86::VPROTDmi,         0 },
740     { X86::VPROTDrr,           X86::VPROTDmr,         0 },
741     { X86::VPROTQri,           X86::VPROTQmi,         0 },
742     { X86::VPROTQrr,           X86::VPROTQmr,         0 },
743     { X86::VPROTWri,           X86::VPROTWmi,         0 },
744     { X86::VPROTWrr,           X86::VPROTWmr,         0 },
745     { X86::VPSHABrr,           X86::VPSHABmr,         0 },
746     { X86::VPSHADrr,           X86::VPSHADmr,         0 },
747     { X86::VPSHAQrr,           X86::VPSHAQmr,         0 },
748     { X86::VPSHAWrr,           X86::VPSHAWmr,         0 },
749     { X86::VPSHLBrr,           X86::VPSHLBmr,         0 },
750     { X86::VPSHLDrr,           X86::VPSHLDmr,         0 },
751     { X86::VPSHLQrr,           X86::VPSHLQmr,         0 },
752     { X86::VPSHLWrr,           X86::VPSHLWmr,         0 },
753 
754     // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
755     { X86::BEXTR32rr,       X86::BEXTR32rm,           0 },
756     { X86::BEXTR64rr,       X86::BEXTR64rm,           0 },
757     { X86::BEXTRI32ri,      X86::BEXTRI32mi,          0 },
758     { X86::BEXTRI64ri,      X86::BEXTRI64mi,          0 },
759     { X86::BLCFILL32rr,     X86::BLCFILL32rm,         0 },
760     { X86::BLCFILL64rr,     X86::BLCFILL64rm,         0 },
761     { X86::BLCI32rr,        X86::BLCI32rm,            0 },
762     { X86::BLCI64rr,        X86::BLCI64rm,            0 },
763     { X86::BLCIC32rr,       X86::BLCIC32rm,           0 },
764     { X86::BLCIC64rr,       X86::BLCIC64rm,           0 },
765     { X86::BLCMSK32rr,      X86::BLCMSK32rm,          0 },
766     { X86::BLCMSK64rr,      X86::BLCMSK64rm,          0 },
767     { X86::BLCS32rr,        X86::BLCS32rm,            0 },
768     { X86::BLCS64rr,        X86::BLCS64rm,            0 },
769     { X86::BLSFILL32rr,     X86::BLSFILL32rm,         0 },
770     { X86::BLSFILL64rr,     X86::BLSFILL64rm,         0 },
771     { X86::BLSI32rr,        X86::BLSI32rm,            0 },
772     { X86::BLSI64rr,        X86::BLSI64rm,            0 },
773     { X86::BLSIC32rr,       X86::BLSIC32rm,           0 },
774     { X86::BLSIC64rr,       X86::BLSIC64rm,           0 },
775     { X86::BLSMSK32rr,      X86::BLSMSK32rm,          0 },
776     { X86::BLSMSK64rr,      X86::BLSMSK64rm,          0 },
777     { X86::BLSR32rr,        X86::BLSR32rm,            0 },
778     { X86::BLSR64rr,        X86::BLSR64rm,            0 },
779     { X86::BZHI32rr,        X86::BZHI32rm,            0 },
780     { X86::BZHI64rr,        X86::BZHI64rm,            0 },
781     { X86::LZCNT16rr,       X86::LZCNT16rm,           0 },
782     { X86::LZCNT32rr,       X86::LZCNT32rm,           0 },
783     { X86::LZCNT64rr,       X86::LZCNT64rm,           0 },
784     { X86::POPCNT16rr,      X86::POPCNT16rm,          0 },
785     { X86::POPCNT32rr,      X86::POPCNT32rm,          0 },
786     { X86::POPCNT64rr,      X86::POPCNT64rm,          0 },
787     { X86::RORX32ri,        X86::RORX32mi,            0 },
788     { X86::RORX64ri,        X86::RORX64mi,            0 },
789     { X86::SARX32rr,        X86::SARX32rm,            0 },
790     { X86::SARX64rr,        X86::SARX64rm,            0 },
791     { X86::SHRX32rr,        X86::SHRX32rm,            0 },
792     { X86::SHRX64rr,        X86::SHRX64rm,            0 },
793     { X86::SHLX32rr,        X86::SHLX32rm,            0 },
794     { X86::SHLX64rr,        X86::SHLX64rm,            0 },
795     { X86::T1MSKC32rr,      X86::T1MSKC32rm,          0 },
796     { X86::T1MSKC64rr,      X86::T1MSKC64rm,          0 },
797     { X86::TZCNT16rr,       X86::TZCNT16rm,           0 },
798     { X86::TZCNT32rr,       X86::TZCNT32rm,           0 },
799     { X86::TZCNT64rr,       X86::TZCNT64rm,           0 },
800     { X86::TZMSK32rr,       X86::TZMSK32rm,           0 },
801     { X86::TZMSK64rr,       X86::TZMSK64rm,           0 },
802 
803     // AVX-512 foldable instructions
804     { X86::VMOV64toPQIZrr,  X86::VMOVQI2PQIZrm,       0 },
805     { X86::VMOVDI2SSZrr,    X86::VMOVDI2SSZrm,        0 },
806     { X86::VMOVAPDZrr,      X86::VMOVAPDZrm,          TB_ALIGN_64 },
807     { X86::VMOVAPSZrr,      X86::VMOVAPSZrm,          TB_ALIGN_64 },
808     { X86::VMOVDQA32Zrr,    X86::VMOVDQA32Zrm,        TB_ALIGN_64 },
809     { X86::VMOVDQA64Zrr,    X86::VMOVDQA64Zrm,        TB_ALIGN_64 },
810     { X86::VMOVDQU8Zrr,     X86::VMOVDQU8Zrm,         0 },
811     { X86::VMOVDQU16Zrr,    X86::VMOVDQU16Zrm,        0 },
812     { X86::VMOVDQU32Zrr,    X86::VMOVDQU32Zrm,        0 },
813     { X86::VMOVDQU64Zrr,    X86::VMOVDQU64Zrm,        0 },
814     { X86::VMOVUPDZrr,      X86::VMOVUPDZrm,          0 },
815     { X86::VMOVUPSZrr,      X86::VMOVUPSZrm,          0 },
816     { X86::VPABSDZrr,       X86::VPABSDZrm,           0 },
817     { X86::VPABSQZrr,       X86::VPABSQZrm,           0 },
818     { X86::VBROADCASTSSZr,  X86::VBROADCASTSSZm,      TB_NO_REVERSE },
819     { X86::VBROADCASTSDZr,  X86::VBROADCASTSDZm,      TB_NO_REVERSE },
820 
821     // AVX-512 foldable instructions (256-bit versions)
822     { X86::VMOVAPDZ256rr,      X86::VMOVAPDZ256rm,          TB_ALIGN_32 },
823     { X86::VMOVAPSZ256rr,      X86::VMOVAPSZ256rm,          TB_ALIGN_32 },
824     { X86::VMOVDQA32Z256rr,    X86::VMOVDQA32Z256rm,        TB_ALIGN_32 },
825     { X86::VMOVDQA64Z256rr,    X86::VMOVDQA64Z256rm,        TB_ALIGN_32 },
826     { X86::VMOVDQU8Z256rr,     X86::VMOVDQU8Z256rm,         0 },
827     { X86::VMOVDQU16Z256rr,    X86::VMOVDQU16Z256rm,        0 },
828     { X86::VMOVDQU32Z256rr,    X86::VMOVDQU32Z256rm,        0 },
829     { X86::VMOVDQU64Z256rr,    X86::VMOVDQU64Z256rm,        0 },
830     { X86::VMOVUPDZ256rr,      X86::VMOVUPDZ256rm,          0 },
831     { X86::VMOVUPSZ256rr,      X86::VMOVUPSZ256rm,          0 },
832     { X86::VBROADCASTSSZ256r,  X86::VBROADCASTSSZ256m,      TB_NO_REVERSE },
833     { X86::VBROADCASTSDZ256r,  X86::VBROADCASTSDZ256m,      TB_NO_REVERSE },
834 
835     // AVX-512 foldable instructions (256-bit versions)
836     { X86::VMOVAPDZ128rr,      X86::VMOVAPDZ128rm,          TB_ALIGN_16 },
837     { X86::VMOVAPSZ128rr,      X86::VMOVAPSZ128rm,          TB_ALIGN_16 },
838     { X86::VMOVDQA32Z128rr,    X86::VMOVDQA32Z128rm,        TB_ALIGN_16 },
839     { X86::VMOVDQA64Z128rr,    X86::VMOVDQA64Z128rm,        TB_ALIGN_16 },
840     { X86::VMOVDQU8Z128rr,     X86::VMOVDQU8Z128rm,         0 },
841     { X86::VMOVDQU16Z128rr,    X86::VMOVDQU16Z128rm,        0 },
842     { X86::VMOVDQU32Z128rr,    X86::VMOVDQU32Z128rm,        0 },
843     { X86::VMOVDQU64Z128rr,    X86::VMOVDQU64Z128rm,        0 },
844     { X86::VMOVUPDZ128rr,      X86::VMOVUPDZ128rm,          0 },
845     { X86::VMOVUPSZ128rr,      X86::VMOVUPSZ128rm,          0 },
846     { X86::VBROADCASTSSZ128r,  X86::VBROADCASTSSZ128m,      TB_NO_REVERSE },
847 
848     // F16C foldable instructions
849     { X86::VCVTPH2PSrr,        X86::VCVTPH2PSrm,            0 },
850     { X86::VCVTPH2PSYrr,       X86::VCVTPH2PSYrm,           0 },
851 
852     // AES foldable instructions
853     { X86::AESIMCrr,              X86::AESIMCrm,              TB_ALIGN_16 },
854     { X86::AESKEYGENASSIST128rr,  X86::AESKEYGENASSIST128rm,  TB_ALIGN_16 },
855     { X86::VAESIMCrr,             X86::VAESIMCrm,             0 },
856     { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, 0 }
857   };
858 
859   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable1) {
860     AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
861                   Entry.RegOp, Entry.MemOp,
862                   // Index 1, folded load
863                   Entry.Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
864   }
865 
866   static const X86MemoryFoldTableEntry MemoryFoldTable2[] = {
867     { X86::ADC32rr,         X86::ADC32rm,       0 },
868     { X86::ADC64rr,         X86::ADC64rm,       0 },
869     { X86::ADD16rr,         X86::ADD16rm,       0 },
870     { X86::ADD16rr_DB,      X86::ADD16rm,       TB_NO_REVERSE },
871     { X86::ADD32rr,         X86::ADD32rm,       0 },
872     { X86::ADD32rr_DB,      X86::ADD32rm,       TB_NO_REVERSE },
873     { X86::ADD64rr,         X86::ADD64rm,       0 },
874     { X86::ADD64rr_DB,      X86::ADD64rm,       TB_NO_REVERSE },
875     { X86::ADD8rr,          X86::ADD8rm,        0 },
876     { X86::ADDPDrr,         X86::ADDPDrm,       TB_ALIGN_16 },
877     { X86::ADDPSrr,         X86::ADDPSrm,       TB_ALIGN_16 },
878     { X86::ADDSDrr,         X86::ADDSDrm,       0 },
879     { X86::ADDSDrr_Int,     X86::ADDSDrm_Int,   0 },
880     { X86::ADDSSrr,         X86::ADDSSrm,       0 },
881     { X86::ADDSSrr_Int,     X86::ADDSSrm_Int,   0 },
882     { X86::ADDSUBPDrr,      X86::ADDSUBPDrm,    TB_ALIGN_16 },
883     { X86::ADDSUBPSrr,      X86::ADDSUBPSrm,    TB_ALIGN_16 },
884     { X86::AND16rr,         X86::AND16rm,       0 },
885     { X86::AND32rr,         X86::AND32rm,       0 },
886     { X86::AND64rr,         X86::AND64rm,       0 },
887     { X86::AND8rr,          X86::AND8rm,        0 },
888     { X86::ANDNPDrr,        X86::ANDNPDrm,      TB_ALIGN_16 },
889     { X86::ANDNPSrr,        X86::ANDNPSrm,      TB_ALIGN_16 },
890     { X86::ANDPDrr,         X86::ANDPDrm,       TB_ALIGN_16 },
891     { X86::ANDPSrr,         X86::ANDPSrm,       TB_ALIGN_16 },
892     { X86::BLENDPDrri,      X86::BLENDPDrmi,    TB_ALIGN_16 },
893     { X86::BLENDPSrri,      X86::BLENDPSrmi,    TB_ALIGN_16 },
894     { X86::BLENDVPDrr0,     X86::BLENDVPDrm0,   TB_ALIGN_16 },
895     { X86::BLENDVPSrr0,     X86::BLENDVPSrm0,   TB_ALIGN_16 },
896     { X86::CMOVA16rr,       X86::CMOVA16rm,     0 },
897     { X86::CMOVA32rr,       X86::CMOVA32rm,     0 },
898     { X86::CMOVA64rr,       X86::CMOVA64rm,     0 },
899     { X86::CMOVAE16rr,      X86::CMOVAE16rm,    0 },
900     { X86::CMOVAE32rr,      X86::CMOVAE32rm,    0 },
901     { X86::CMOVAE64rr,      X86::CMOVAE64rm,    0 },
902     { X86::CMOVB16rr,       X86::CMOVB16rm,     0 },
903     { X86::CMOVB32rr,       X86::CMOVB32rm,     0 },
904     { X86::CMOVB64rr,       X86::CMOVB64rm,     0 },
905     { X86::CMOVBE16rr,      X86::CMOVBE16rm,    0 },
906     { X86::CMOVBE32rr,      X86::CMOVBE32rm,    0 },
907     { X86::CMOVBE64rr,      X86::CMOVBE64rm,    0 },
908     { X86::CMOVE16rr,       X86::CMOVE16rm,     0 },
909     { X86::CMOVE32rr,       X86::CMOVE32rm,     0 },
910     { X86::CMOVE64rr,       X86::CMOVE64rm,     0 },
911     { X86::CMOVG16rr,       X86::CMOVG16rm,     0 },
912     { X86::CMOVG32rr,       X86::CMOVG32rm,     0 },
913     { X86::CMOVG64rr,       X86::CMOVG64rm,     0 },
914     { X86::CMOVGE16rr,      X86::CMOVGE16rm,    0 },
915     { X86::CMOVGE32rr,      X86::CMOVGE32rm,    0 },
916     { X86::CMOVGE64rr,      X86::CMOVGE64rm,    0 },
917     { X86::CMOVL16rr,       X86::CMOVL16rm,     0 },
918     { X86::CMOVL32rr,       X86::CMOVL32rm,     0 },
919     { X86::CMOVL64rr,       X86::CMOVL64rm,     0 },
920     { X86::CMOVLE16rr,      X86::CMOVLE16rm,    0 },
921     { X86::CMOVLE32rr,      X86::CMOVLE32rm,    0 },
922     { X86::CMOVLE64rr,      X86::CMOVLE64rm,    0 },
923     { X86::CMOVNE16rr,      X86::CMOVNE16rm,    0 },
924     { X86::CMOVNE32rr,      X86::CMOVNE32rm,    0 },
925     { X86::CMOVNE64rr,      X86::CMOVNE64rm,    0 },
926     { X86::CMOVNO16rr,      X86::CMOVNO16rm,    0 },
927     { X86::CMOVNO32rr,      X86::CMOVNO32rm,    0 },
928     { X86::CMOVNO64rr,      X86::CMOVNO64rm,    0 },
929     { X86::CMOVNP16rr,      X86::CMOVNP16rm,    0 },
930     { X86::CMOVNP32rr,      X86::CMOVNP32rm,    0 },
931     { X86::CMOVNP64rr,      X86::CMOVNP64rm,    0 },
932     { X86::CMOVNS16rr,      X86::CMOVNS16rm,    0 },
933     { X86::CMOVNS32rr,      X86::CMOVNS32rm,    0 },
934     { X86::CMOVNS64rr,      X86::CMOVNS64rm,    0 },
935     { X86::CMOVO16rr,       X86::CMOVO16rm,     0 },
936     { X86::CMOVO32rr,       X86::CMOVO32rm,     0 },
937     { X86::CMOVO64rr,       X86::CMOVO64rm,     0 },
938     { X86::CMOVP16rr,       X86::CMOVP16rm,     0 },
939     { X86::CMOVP32rr,       X86::CMOVP32rm,     0 },
940     { X86::CMOVP64rr,       X86::CMOVP64rm,     0 },
941     { X86::CMOVS16rr,       X86::CMOVS16rm,     0 },
942     { X86::CMOVS32rr,       X86::CMOVS32rm,     0 },
943     { X86::CMOVS64rr,       X86::CMOVS64rm,     0 },
944     { X86::CMPPDrri,        X86::CMPPDrmi,      TB_ALIGN_16 },
945     { X86::CMPPSrri,        X86::CMPPSrmi,      TB_ALIGN_16 },
946     { X86::CMPSDrr,         X86::CMPSDrm,       0 },
947     { X86::CMPSSrr,         X86::CMPSSrm,       0 },
948     { X86::CRC32r32r32,     X86::CRC32r32m32,   0 },
949     { X86::CRC32r64r64,     X86::CRC32r64m64,   0 },
950     { X86::DIVPDrr,         X86::DIVPDrm,       TB_ALIGN_16 },
951     { X86::DIVPSrr,         X86::DIVPSrm,       TB_ALIGN_16 },
952     { X86::DIVSDrr,         X86::DIVSDrm,       0 },
953     { X86::DIVSDrr_Int,     X86::DIVSDrm_Int,   0 },
954     { X86::DIVSSrr,         X86::DIVSSrm,       0 },
955     { X86::DIVSSrr_Int,     X86::DIVSSrm_Int,   0 },
956     { X86::DPPDrri,         X86::DPPDrmi,       TB_ALIGN_16 },
957     { X86::DPPSrri,         X86::DPPSrmi,       TB_ALIGN_16 },
958 
959     // FIXME: We should not be folding Fs* scalar loads into vector
960     // instructions because the vector instructions require vector-sized
961     // loads. Lowering should create vector-sized instructions (the Fv*
962     // variants below) to allow load folding.
963     { X86::FsANDNPDrr,      X86::FsANDNPDrm,    TB_ALIGN_16 },
964     { X86::FsANDNPSrr,      X86::FsANDNPSrm,    TB_ALIGN_16 },
965     { X86::FsANDPDrr,       X86::FsANDPDrm,     TB_ALIGN_16 },
966     { X86::FsANDPSrr,       X86::FsANDPSrm,     TB_ALIGN_16 },
967     { X86::FsORPDrr,        X86::FsORPDrm,      TB_ALIGN_16 },
968     { X86::FsORPSrr,        X86::FsORPSrm,      TB_ALIGN_16 },
969     { X86::FsXORPDrr,       X86::FsXORPDrm,     TB_ALIGN_16 },
970     { X86::FsXORPSrr,       X86::FsXORPSrm,     TB_ALIGN_16 },
971 
972     { X86::FvANDNPDrr,      X86::FvANDNPDrm,    TB_ALIGN_16 },
973     { X86::FvANDNPSrr,      X86::FvANDNPSrm,    TB_ALIGN_16 },
974     { X86::FvANDPDrr,       X86::FvANDPDrm,     TB_ALIGN_16 },
975     { X86::FvANDPSrr,       X86::FvANDPSrm,     TB_ALIGN_16 },
976     { X86::FvORPDrr,        X86::FvORPDrm,      TB_ALIGN_16 },
977     { X86::FvORPSrr,        X86::FvORPSrm,      TB_ALIGN_16 },
978     { X86::FvXORPDrr,       X86::FvXORPDrm,     TB_ALIGN_16 },
979     { X86::FvXORPSrr,       X86::FvXORPSrm,     TB_ALIGN_16 },
980     { X86::HADDPDrr,        X86::HADDPDrm,      TB_ALIGN_16 },
981     { X86::HADDPSrr,        X86::HADDPSrm,      TB_ALIGN_16 },
982     { X86::HSUBPDrr,        X86::HSUBPDrm,      TB_ALIGN_16 },
983     { X86::HSUBPSrr,        X86::HSUBPSrm,      TB_ALIGN_16 },
984     { X86::IMUL16rr,        X86::IMUL16rm,      0 },
985     { X86::IMUL32rr,        X86::IMUL32rm,      0 },
986     { X86::IMUL64rr,        X86::IMUL64rm,      0 },
987     { X86::Int_CMPSDrr,     X86::Int_CMPSDrm,   0 },
988     { X86::Int_CMPSSrr,     X86::Int_CMPSSrm,   0 },
989     { X86::Int_CVTSD2SSrr,  X86::Int_CVTSD2SSrm,      0 },
990     { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm,    0 },
991     { X86::Int_CVTSI2SDrr,  X86::Int_CVTSI2SDrm,      0 },
992     { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm,    0 },
993     { X86::Int_CVTSI2SSrr,  X86::Int_CVTSI2SSrm,      0 },
994     { X86::Int_CVTSS2SDrr,  X86::Int_CVTSS2SDrm,      0 },
995     { X86::MAXPDrr,         X86::MAXPDrm,       TB_ALIGN_16 },
996     { X86::MAXPSrr,         X86::MAXPSrm,       TB_ALIGN_16 },
997     { X86::MAXSDrr,         X86::MAXSDrm,       0 },
998     { X86::MAXSDrr_Int,     X86::MAXSDrm_Int,   0 },
999     { X86::MAXSSrr,         X86::MAXSSrm,       0 },
1000     { X86::MAXSSrr_Int,     X86::MAXSSrm_Int,   0 },
1001     { X86::MINPDrr,         X86::MINPDrm,       TB_ALIGN_16 },
1002     { X86::MINPSrr,         X86::MINPSrm,       TB_ALIGN_16 },
1003     { X86::MINSDrr,         X86::MINSDrm,       0 },
1004     { X86::MINSDrr_Int,     X86::MINSDrm_Int,   0 },
1005     { X86::MINSSrr,         X86::MINSSrm,       0 },
1006     { X86::MINSSrr_Int,     X86::MINSSrm_Int,   0 },
1007     { X86::MPSADBWrri,      X86::MPSADBWrmi,    TB_ALIGN_16 },
1008     { X86::MULPDrr,         X86::MULPDrm,       TB_ALIGN_16 },
1009     { X86::MULPSrr,         X86::MULPSrm,       TB_ALIGN_16 },
1010     { X86::MULSDrr,         X86::MULSDrm,       0 },
1011     { X86::MULSDrr_Int,     X86::MULSDrm_Int,   0 },
1012     { X86::MULSSrr,         X86::MULSSrm,       0 },
1013     { X86::MULSSrr_Int,     X86::MULSSrm_Int,   0 },
1014     { X86::OR16rr,          X86::OR16rm,        0 },
1015     { X86::OR32rr,          X86::OR32rm,        0 },
1016     { X86::OR64rr,          X86::OR64rm,        0 },
1017     { X86::OR8rr,           X86::OR8rm,         0 },
1018     { X86::ORPDrr,          X86::ORPDrm,        TB_ALIGN_16 },
1019     { X86::ORPSrr,          X86::ORPSrm,        TB_ALIGN_16 },
1020     { X86::PACKSSDWrr,      X86::PACKSSDWrm,    TB_ALIGN_16 },
1021     { X86::PACKSSWBrr,      X86::PACKSSWBrm,    TB_ALIGN_16 },
1022     { X86::PACKUSDWrr,      X86::PACKUSDWrm,    TB_ALIGN_16 },
1023     { X86::PACKUSWBrr,      X86::PACKUSWBrm,    TB_ALIGN_16 },
1024     { X86::PADDBrr,         X86::PADDBrm,       TB_ALIGN_16 },
1025     { X86::PADDDrr,         X86::PADDDrm,       TB_ALIGN_16 },
1026     { X86::PADDQrr,         X86::PADDQrm,       TB_ALIGN_16 },
1027     { X86::PADDSBrr,        X86::PADDSBrm,      TB_ALIGN_16 },
1028     { X86::PADDSWrr,        X86::PADDSWrm,      TB_ALIGN_16 },
1029     { X86::PADDUSBrr,       X86::PADDUSBrm,     TB_ALIGN_16 },
1030     { X86::PADDUSWrr,       X86::PADDUSWrm,     TB_ALIGN_16 },
1031     { X86::PADDWrr,         X86::PADDWrm,       TB_ALIGN_16 },
1032     { X86::PALIGNR128rr,    X86::PALIGNR128rm,  TB_ALIGN_16 },
1033     { X86::PANDNrr,         X86::PANDNrm,       TB_ALIGN_16 },
1034     { X86::PANDrr,          X86::PANDrm,        TB_ALIGN_16 },
1035     { X86::PAVGBrr,         X86::PAVGBrm,       TB_ALIGN_16 },
1036     { X86::PAVGWrr,         X86::PAVGWrm,       TB_ALIGN_16 },
1037     { X86::PBLENDVBrr0,     X86::PBLENDVBrm0,   TB_ALIGN_16 },
1038     { X86::PBLENDWrri,      X86::PBLENDWrmi,    TB_ALIGN_16 },
1039     { X86::PCLMULQDQrr,     X86::PCLMULQDQrm,   TB_ALIGN_16 },
1040     { X86::PCMPEQBrr,       X86::PCMPEQBrm,     TB_ALIGN_16 },
1041     { X86::PCMPEQDrr,       X86::PCMPEQDrm,     TB_ALIGN_16 },
1042     { X86::PCMPEQQrr,       X86::PCMPEQQrm,     TB_ALIGN_16 },
1043     { X86::PCMPEQWrr,       X86::PCMPEQWrm,     TB_ALIGN_16 },
1044     { X86::PCMPGTBrr,       X86::PCMPGTBrm,     TB_ALIGN_16 },
1045     { X86::PCMPGTDrr,       X86::PCMPGTDrm,     TB_ALIGN_16 },
1046     { X86::PCMPGTQrr,       X86::PCMPGTQrm,     TB_ALIGN_16 },
1047     { X86::PCMPGTWrr,       X86::PCMPGTWrm,     TB_ALIGN_16 },
1048     { X86::PHADDDrr,        X86::PHADDDrm,      TB_ALIGN_16 },
1049     { X86::PHADDWrr,        X86::PHADDWrm,      TB_ALIGN_16 },
1050     { X86::PHADDSWrr128,    X86::PHADDSWrm128,  TB_ALIGN_16 },
1051     { X86::PHSUBDrr,        X86::PHSUBDrm,      TB_ALIGN_16 },
1052     { X86::PHSUBSWrr128,    X86::PHSUBSWrm128,  TB_ALIGN_16 },
1053     { X86::PHSUBWrr,        X86::PHSUBWrm,      TB_ALIGN_16 },
1054     { X86::PINSRBrr,        X86::PINSRBrm,      0 },
1055     { X86::PINSRDrr,        X86::PINSRDrm,      0 },
1056     { X86::PINSRQrr,        X86::PINSRQrm,      0 },
1057     { X86::PINSRWrri,       X86::PINSRWrmi,     0 },
1058     { X86::PMADDUBSWrr128,  X86::PMADDUBSWrm128, TB_ALIGN_16 },
1059     { X86::PMADDWDrr,       X86::PMADDWDrm,     TB_ALIGN_16 },
1060     { X86::PMAXSWrr,        X86::PMAXSWrm,      TB_ALIGN_16 },
1061     { X86::PMAXUBrr,        X86::PMAXUBrm,      TB_ALIGN_16 },
1062     { X86::PMINSWrr,        X86::PMINSWrm,      TB_ALIGN_16 },
1063     { X86::PMINUBrr,        X86::PMINUBrm,      TB_ALIGN_16 },
1064     { X86::PMINSBrr,        X86::PMINSBrm,      TB_ALIGN_16 },
1065     { X86::PMINSDrr,        X86::PMINSDrm,      TB_ALIGN_16 },
1066     { X86::PMINUDrr,        X86::PMINUDrm,      TB_ALIGN_16 },
1067     { X86::PMINUWrr,        X86::PMINUWrm,      TB_ALIGN_16 },
1068     { X86::PMAXSBrr,        X86::PMAXSBrm,      TB_ALIGN_16 },
1069     { X86::PMAXSDrr,        X86::PMAXSDrm,      TB_ALIGN_16 },
1070     { X86::PMAXUDrr,        X86::PMAXUDrm,      TB_ALIGN_16 },
1071     { X86::PMAXUWrr,        X86::PMAXUWrm,      TB_ALIGN_16 },
1072     { X86::PMULDQrr,        X86::PMULDQrm,      TB_ALIGN_16 },
1073     { X86::PMULHRSWrr128,   X86::PMULHRSWrm128, TB_ALIGN_16 },
1074     { X86::PMULHUWrr,       X86::PMULHUWrm,     TB_ALIGN_16 },
1075     { X86::PMULHWrr,        X86::PMULHWrm,      TB_ALIGN_16 },
1076     { X86::PMULLDrr,        X86::PMULLDrm,      TB_ALIGN_16 },
1077     { X86::PMULLWrr,        X86::PMULLWrm,      TB_ALIGN_16 },
1078     { X86::PMULUDQrr,       X86::PMULUDQrm,     TB_ALIGN_16 },
1079     { X86::PORrr,           X86::PORrm,         TB_ALIGN_16 },
1080     { X86::PSADBWrr,        X86::PSADBWrm,      TB_ALIGN_16 },
1081     { X86::PSHUFBrr,        X86::PSHUFBrm,      TB_ALIGN_16 },
1082     { X86::PSIGNBrr,        X86::PSIGNBrm,      TB_ALIGN_16 },
1083     { X86::PSIGNWrr,        X86::PSIGNWrm,      TB_ALIGN_16 },
1084     { X86::PSIGNDrr,        X86::PSIGNDrm,      TB_ALIGN_16 },
1085     { X86::PSLLDrr,         X86::PSLLDrm,       TB_ALIGN_16 },
1086     { X86::PSLLQrr,         X86::PSLLQrm,       TB_ALIGN_16 },
1087     { X86::PSLLWrr,         X86::PSLLWrm,       TB_ALIGN_16 },
1088     { X86::PSRADrr,         X86::PSRADrm,       TB_ALIGN_16 },
1089     { X86::PSRAWrr,         X86::PSRAWrm,       TB_ALIGN_16 },
1090     { X86::PSRLDrr,         X86::PSRLDrm,       TB_ALIGN_16 },
1091     { X86::PSRLQrr,         X86::PSRLQrm,       TB_ALIGN_16 },
1092     { X86::PSRLWrr,         X86::PSRLWrm,       TB_ALIGN_16 },
1093     { X86::PSUBBrr,         X86::PSUBBrm,       TB_ALIGN_16 },
1094     { X86::PSUBDrr,         X86::PSUBDrm,       TB_ALIGN_16 },
1095     { X86::PSUBQrr,         X86::PSUBQrm,       TB_ALIGN_16 },
1096     { X86::PSUBSBrr,        X86::PSUBSBrm,      TB_ALIGN_16 },
1097     { X86::PSUBSWrr,        X86::PSUBSWrm,      TB_ALIGN_16 },
1098     { X86::PSUBUSBrr,       X86::PSUBUSBrm,     TB_ALIGN_16 },
1099     { X86::PSUBUSWrr,       X86::PSUBUSWrm,     TB_ALIGN_16 },
1100     { X86::PSUBWrr,         X86::PSUBWrm,       TB_ALIGN_16 },
1101     { X86::PUNPCKHBWrr,     X86::PUNPCKHBWrm,   TB_ALIGN_16 },
1102     { X86::PUNPCKHDQrr,     X86::PUNPCKHDQrm,   TB_ALIGN_16 },
1103     { X86::PUNPCKHQDQrr,    X86::PUNPCKHQDQrm,  TB_ALIGN_16 },
1104     { X86::PUNPCKHWDrr,     X86::PUNPCKHWDrm,   TB_ALIGN_16 },
1105     { X86::PUNPCKLBWrr,     X86::PUNPCKLBWrm,   TB_ALIGN_16 },
1106     { X86::PUNPCKLDQrr,     X86::PUNPCKLDQrm,   TB_ALIGN_16 },
1107     { X86::PUNPCKLQDQrr,    X86::PUNPCKLQDQrm,  TB_ALIGN_16 },
1108     { X86::PUNPCKLWDrr,     X86::PUNPCKLWDrm,   TB_ALIGN_16 },
1109     { X86::PXORrr,          X86::PXORrm,        TB_ALIGN_16 },
1110     { X86::ROUNDSDr,        X86::ROUNDSDm,      0 },
1111     { X86::ROUNDSSr,        X86::ROUNDSSm,      0 },
1112     { X86::SBB32rr,         X86::SBB32rm,       0 },
1113     { X86::SBB64rr,         X86::SBB64rm,       0 },
1114     { X86::SHUFPDrri,       X86::SHUFPDrmi,     TB_ALIGN_16 },
1115     { X86::SHUFPSrri,       X86::SHUFPSrmi,     TB_ALIGN_16 },
1116     { X86::SUB16rr,         X86::SUB16rm,       0 },
1117     { X86::SUB32rr,         X86::SUB32rm,       0 },
1118     { X86::SUB64rr,         X86::SUB64rm,       0 },
1119     { X86::SUB8rr,          X86::SUB8rm,        0 },
1120     { X86::SUBPDrr,         X86::SUBPDrm,       TB_ALIGN_16 },
1121     { X86::SUBPSrr,         X86::SUBPSrm,       TB_ALIGN_16 },
1122     { X86::SUBSDrr,         X86::SUBSDrm,       0 },
1123     { X86::SUBSDrr_Int,     X86::SUBSDrm_Int,   0 },
1124     { X86::SUBSSrr,         X86::SUBSSrm,       0 },
1125     { X86::SUBSSrr_Int,     X86::SUBSSrm_Int,   0 },
1126     // FIXME: TEST*rr -> swapped operand of TEST*mr.
1127     { X86::UNPCKHPDrr,      X86::UNPCKHPDrm,    TB_ALIGN_16 },
1128     { X86::UNPCKHPSrr,      X86::UNPCKHPSrm,    TB_ALIGN_16 },
1129     { X86::UNPCKLPDrr,      X86::UNPCKLPDrm,    TB_ALIGN_16 },
1130     { X86::UNPCKLPSrr,      X86::UNPCKLPSrm,    TB_ALIGN_16 },
1131     { X86::XOR16rr,         X86::XOR16rm,       0 },
1132     { X86::XOR32rr,         X86::XOR32rm,       0 },
1133     { X86::XOR64rr,         X86::XOR64rm,       0 },
1134     { X86::XOR8rr,          X86::XOR8rm,        0 },
1135     { X86::XORPDrr,         X86::XORPDrm,       TB_ALIGN_16 },
1136     { X86::XORPSrr,         X86::XORPSrm,       TB_ALIGN_16 },
1137 
1138     // MMX version of foldable instructions
1139     { X86::MMX_CVTPI2PSirr,   X86::MMX_CVTPI2PSirm,   0 },
1140     { X86::MMX_PACKSSDWirr,   X86::MMX_PACKSSDWirm,   0 },
1141     { X86::MMX_PACKSSWBirr,   X86::MMX_PACKSSWBirm,   0 },
1142     { X86::MMX_PACKUSWBirr,   X86::MMX_PACKUSWBirm,   0 },
1143     { X86::MMX_PADDBirr,      X86::MMX_PADDBirm,      0 },
1144     { X86::MMX_PADDDirr,      X86::MMX_PADDDirm,      0 },
1145     { X86::MMX_PADDQirr,      X86::MMX_PADDQirm,      0 },
1146     { X86::MMX_PADDSBirr,     X86::MMX_PADDSBirm,     0 },
1147     { X86::MMX_PADDSWirr,     X86::MMX_PADDSWirm,     0 },
1148     { X86::MMX_PADDUSBirr,    X86::MMX_PADDUSBirm,    0 },
1149     { X86::MMX_PADDUSWirr,    X86::MMX_PADDUSWirm,    0 },
1150     { X86::MMX_PADDWirr,      X86::MMX_PADDWirm,      0 },
1151     { X86::MMX_PALIGNR64irr,  X86::MMX_PALIGNR64irm,  0 },
1152     { X86::MMX_PANDNirr,      X86::MMX_PANDNirm,      0 },
1153     { X86::MMX_PANDirr,       X86::MMX_PANDirm,       0 },
1154     { X86::MMX_PAVGBirr,      X86::MMX_PAVGBirm,      0 },
1155     { X86::MMX_PAVGWirr,      X86::MMX_PAVGWirm,      0 },
1156     { X86::MMX_PCMPEQBirr,    X86::MMX_PCMPEQBirm,    0 },
1157     { X86::MMX_PCMPEQDirr,    X86::MMX_PCMPEQDirm,    0 },
1158     { X86::MMX_PCMPEQWirr,    X86::MMX_PCMPEQWirm,    0 },
1159     { X86::MMX_PCMPGTBirr,    X86::MMX_PCMPGTBirm,    0 },
1160     { X86::MMX_PCMPGTDirr,    X86::MMX_PCMPGTDirm,    0 },
1161     { X86::MMX_PCMPGTWirr,    X86::MMX_PCMPGTWirm,    0 },
1162     { X86::MMX_PHADDSWrr64,   X86::MMX_PHADDSWrm64,   0 },
1163     { X86::MMX_PHADDWrr64,    X86::MMX_PHADDWrm64,    0 },
1164     { X86::MMX_PHADDrr64,     X86::MMX_PHADDrm64,     0 },
1165     { X86::MMX_PHSUBDrr64,    X86::MMX_PHSUBDrm64,    0 },
1166     { X86::MMX_PHSUBSWrr64,   X86::MMX_PHSUBSWrm64,   0 },
1167     { X86::MMX_PHSUBWrr64,    X86::MMX_PHSUBWrm64,    0 },
1168     { X86::MMX_PINSRWirri,    X86::MMX_PINSRWirmi,    0 },
1169     { X86::MMX_PMADDUBSWrr64, X86::MMX_PMADDUBSWrm64, 0 },
1170     { X86::MMX_PMADDWDirr,    X86::MMX_PMADDWDirm,    0 },
1171     { X86::MMX_PMAXSWirr,     X86::MMX_PMAXSWirm,     0 },
1172     { X86::MMX_PMAXUBirr,     X86::MMX_PMAXUBirm,     0 },
1173     { X86::MMX_PMINSWirr,     X86::MMX_PMINSWirm,     0 },
1174     { X86::MMX_PMINUBirr,     X86::MMX_PMINUBirm,     0 },
1175     { X86::MMX_PMULHRSWrr64,  X86::MMX_PMULHRSWrm64,  0 },
1176     { X86::MMX_PMULHUWirr,    X86::MMX_PMULHUWirm,    0 },
1177     { X86::MMX_PMULHWirr,     X86::MMX_PMULHWirm,     0 },
1178     { X86::MMX_PMULLWirr,     X86::MMX_PMULLWirm,     0 },
1179     { X86::MMX_PMULUDQirr,    X86::MMX_PMULUDQirm,    0 },
1180     { X86::MMX_PORirr,        X86::MMX_PORirm,        0 },
1181     { X86::MMX_PSADBWirr,     X86::MMX_PSADBWirm,     0 },
1182     { X86::MMX_PSHUFBrr64,    X86::MMX_PSHUFBrm64,    0 },
1183     { X86::MMX_PSIGNBrr64,    X86::MMX_PSIGNBrm64,    0 },
1184     { X86::MMX_PSIGNDrr64,    X86::MMX_PSIGNDrm64,    0 },
1185     { X86::MMX_PSIGNWrr64,    X86::MMX_PSIGNWrm64,    0 },
1186     { X86::MMX_PSLLDrr,       X86::MMX_PSLLDrm,       0 },
1187     { X86::MMX_PSLLQrr,       X86::MMX_PSLLQrm,       0 },
1188     { X86::MMX_PSLLWrr,       X86::MMX_PSLLWrm,       0 },
1189     { X86::MMX_PSRADrr,       X86::MMX_PSRADrm,       0 },
1190     { X86::MMX_PSRAWrr,       X86::MMX_PSRAWrm,       0 },
1191     { X86::MMX_PSRLDrr,       X86::MMX_PSRLDrm,       0 },
1192     { X86::MMX_PSRLQrr,       X86::MMX_PSRLQrm,       0 },
1193     { X86::MMX_PSRLWrr,       X86::MMX_PSRLWrm,       0 },
1194     { X86::MMX_PSUBBirr,      X86::MMX_PSUBBirm,      0 },
1195     { X86::MMX_PSUBDirr,      X86::MMX_PSUBDirm,      0 },
1196     { X86::MMX_PSUBQirr,      X86::MMX_PSUBQirm,      0 },
1197     { X86::MMX_PSUBSBirr,     X86::MMX_PSUBSBirm,     0 },
1198     { X86::MMX_PSUBSWirr,     X86::MMX_PSUBSWirm,     0 },
1199     { X86::MMX_PSUBUSBirr,    X86::MMX_PSUBUSBirm,    0 },
1200     { X86::MMX_PSUBUSWirr,    X86::MMX_PSUBUSWirm,    0 },
1201     { X86::MMX_PSUBWirr,      X86::MMX_PSUBWirm,      0 },
1202     { X86::MMX_PUNPCKHBWirr,  X86::MMX_PUNPCKHBWirm,  0 },
1203     { X86::MMX_PUNPCKHDQirr,  X86::MMX_PUNPCKHDQirm,  0 },
1204     { X86::MMX_PUNPCKHWDirr,  X86::MMX_PUNPCKHWDirm,  0 },
1205     { X86::MMX_PUNPCKLBWirr,  X86::MMX_PUNPCKLBWirm,  0 },
1206     { X86::MMX_PUNPCKLDQirr,  X86::MMX_PUNPCKLDQirm,  0 },
1207     { X86::MMX_PUNPCKLWDirr,  X86::MMX_PUNPCKLWDirm,  0 },
1208     { X86::MMX_PXORirr,       X86::MMX_PXORirm,       0 },
1209 
1210     // 3DNow! version of foldable instructions
1211     { X86::PAVGUSBrr,         X86::PAVGUSBrm,         0 },
1212     { X86::PFACCrr,           X86::PFACCrm,           0 },
1213     { X86::PFADDrr,           X86::PFADDrm,           0 },
1214     { X86::PFCMPEQrr,         X86::PFCMPEQrm,         0 },
1215     { X86::PFCMPGErr,         X86::PFCMPGErm,         0 },
1216     { X86::PFCMPGTrr,         X86::PFCMPGTrm,         0 },
1217     { X86::PFMAXrr,           X86::PFMAXrm,           0 },
1218     { X86::PFMINrr,           X86::PFMINrm,           0 },
1219     { X86::PFMULrr,           X86::PFMULrm,           0 },
1220     { X86::PFNACCrr,          X86::PFNACCrm,          0 },
1221     { X86::PFPNACCrr,         X86::PFPNACCrm,         0 },
1222     { X86::PFRCPIT1rr,        X86::PFRCPIT1rm,        0 },
1223     { X86::PFRCPIT2rr,        X86::PFRCPIT2rm,        0 },
1224     { X86::PFRSQIT1rr,        X86::PFRSQIT1rm,        0 },
1225     { X86::PFSUBrr,           X86::PFSUBrm,           0 },
1226     { X86::PFSUBRrr,          X86::PFSUBRrm,          0 },
1227     { X86::PMULHRWrr,         X86::PMULHRWrm,         0 },
1228 
1229     // AVX 128-bit versions of foldable instructions
1230     { X86::VCVTSD2SSrr,       X86::VCVTSD2SSrm,        0 },
1231     { X86::Int_VCVTSD2SSrr,   X86::Int_VCVTSD2SSrm,    0 },
1232     { X86::VCVTSI2SD64rr,     X86::VCVTSI2SD64rm,      0 },
1233     { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm,  0 },
1234     { X86::VCVTSI2SDrr,       X86::VCVTSI2SDrm,        0 },
1235     { X86::Int_VCVTSI2SDrr,   X86::Int_VCVTSI2SDrm,    0 },
1236     { X86::VCVTSI2SS64rr,     X86::VCVTSI2SS64rm,      0 },
1237     { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm,  0 },
1238     { X86::VCVTSI2SSrr,       X86::VCVTSI2SSrm,        0 },
1239     { X86::Int_VCVTSI2SSrr,   X86::Int_VCVTSI2SSrm,    0 },
1240     { X86::VCVTSS2SDrr,       X86::VCVTSS2SDrm,        0 },
1241     { X86::Int_VCVTSS2SDrr,   X86::Int_VCVTSS2SDrm,    0 },
1242     { X86::VRCPSSr,           X86::VRCPSSm,            0 },
1243     { X86::VRCPSSr_Int,       X86::VRCPSSm_Int,        0 },
1244     { X86::VRSQRTSSr,         X86::VRSQRTSSm,          0 },
1245     { X86::VRSQRTSSr_Int,     X86::VRSQRTSSm_Int,      0 },
1246     { X86::VSQRTSDr,          X86::VSQRTSDm,           0 },
1247     { X86::VSQRTSDr_Int,      X86::VSQRTSDm_Int,       0 },
1248     { X86::VSQRTSSr,          X86::VSQRTSSm,           0 },
1249     { X86::VSQRTSSr_Int,      X86::VSQRTSSm_Int,       0 },
1250     { X86::VADDPDrr,          X86::VADDPDrm,           0 },
1251     { X86::VADDPSrr,          X86::VADDPSrm,           0 },
1252     { X86::VADDSDrr,          X86::VADDSDrm,           0 },
1253     { X86::VADDSDrr_Int,      X86::VADDSDrm_Int,       0 },
1254     { X86::VADDSSrr,          X86::VADDSSrm,           0 },
1255     { X86::VADDSSrr_Int,      X86::VADDSSrm_Int,       0 },
1256     { X86::VADDSUBPDrr,       X86::VADDSUBPDrm,        0 },
1257     { X86::VADDSUBPSrr,       X86::VADDSUBPSrm,        0 },
1258     { X86::VANDNPDrr,         X86::VANDNPDrm,          0 },
1259     { X86::VANDNPSrr,         X86::VANDNPSrm,          0 },
1260     { X86::VANDPDrr,          X86::VANDPDrm,           0 },
1261     { X86::VANDPSrr,          X86::VANDPSrm,           0 },
1262     { X86::VBLENDPDrri,       X86::VBLENDPDrmi,        0 },
1263     { X86::VBLENDPSrri,       X86::VBLENDPSrmi,        0 },
1264     { X86::VBLENDVPDrr,       X86::VBLENDVPDrm,        0 },
1265     { X86::VBLENDVPSrr,       X86::VBLENDVPSrm,        0 },
1266     { X86::VCMPPDrri,         X86::VCMPPDrmi,          0 },
1267     { X86::VCMPPSrri,         X86::VCMPPSrmi,          0 },
1268     { X86::VCMPSDrr,          X86::VCMPSDrm,           0 },
1269     { X86::VCMPSSrr,          X86::VCMPSSrm,           0 },
1270     { X86::VDIVPDrr,          X86::VDIVPDrm,           0 },
1271     { X86::VDIVPSrr,          X86::VDIVPSrm,           0 },
1272     { X86::VDIVSDrr,          X86::VDIVSDrm,           0 },
1273     { X86::VDIVSDrr_Int,      X86::VDIVSDrm_Int,       0 },
1274     { X86::VDIVSSrr,          X86::VDIVSSrm,           0 },
1275     { X86::VDIVSSrr_Int,      X86::VDIVSSrm_Int,       0 },
1276     { X86::VDPPDrri,          X86::VDPPDrmi,           0 },
1277     { X86::VDPPSrri,          X86::VDPPSrmi,           0 },
1278     // Do not fold VFs* loads because there are no scalar load variants for
1279     // these instructions. When folded, the load is required to be 128-bits, so
1280     // the load size would not match.
1281     { X86::VFvANDNPDrr,       X86::VFvANDNPDrm,        0 },
1282     { X86::VFvANDNPSrr,       X86::VFvANDNPSrm,        0 },
1283     { X86::VFvANDPDrr,        X86::VFvANDPDrm,         0 },
1284     { X86::VFvANDPSrr,        X86::VFvANDPSrm,         0 },
1285     { X86::VFvORPDrr,         X86::VFvORPDrm,          0 },
1286     { X86::VFvORPSrr,         X86::VFvORPSrm,          0 },
1287     { X86::VFvXORPDrr,        X86::VFvXORPDrm,         0 },
1288     { X86::VFvXORPSrr,        X86::VFvXORPSrm,         0 },
1289     { X86::VHADDPDrr,         X86::VHADDPDrm,          0 },
1290     { X86::VHADDPSrr,         X86::VHADDPSrm,          0 },
1291     { X86::VHSUBPDrr,         X86::VHSUBPDrm,          0 },
1292     { X86::VHSUBPSrr,         X86::VHSUBPSrm,          0 },
1293     { X86::Int_VCMPSDrr,      X86::Int_VCMPSDrm,       0 },
1294     { X86::Int_VCMPSSrr,      X86::Int_VCMPSSrm,       0 },
1295     { X86::VMAXPDrr,          X86::VMAXPDrm,           0 },
1296     { X86::VMAXPSrr,          X86::VMAXPSrm,           0 },
1297     { X86::VMAXSDrr,          X86::VMAXSDrm,           0 },
1298     { X86::VMAXSDrr_Int,      X86::VMAXSDrm_Int,       0 },
1299     { X86::VMAXSSrr,          X86::VMAXSSrm,           0 },
1300     { X86::VMAXSSrr_Int,      X86::VMAXSSrm_Int,       0 },
1301     { X86::VMINPDrr,          X86::VMINPDrm,           0 },
1302     { X86::VMINPSrr,          X86::VMINPSrm,           0 },
1303     { X86::VMINSDrr,          X86::VMINSDrm,           0 },
1304     { X86::VMINSDrr_Int,      X86::VMINSDrm_Int,       0 },
1305     { X86::VMINSSrr,          X86::VMINSSrm,           0 },
1306     { X86::VMINSSrr_Int,      X86::VMINSSrm_Int,       0 },
1307     { X86::VMPSADBWrri,       X86::VMPSADBWrmi,        0 },
1308     { X86::VMULPDrr,          X86::VMULPDrm,           0 },
1309     { X86::VMULPSrr,          X86::VMULPSrm,           0 },
1310     { X86::VMULSDrr,          X86::VMULSDrm,           0 },
1311     { X86::VMULSDrr_Int,      X86::VMULSDrm_Int,       0 },
1312     { X86::VMULSSrr,          X86::VMULSSrm,           0 },
1313     { X86::VMULSSrr_Int,      X86::VMULSSrm_Int,       0 },
1314     { X86::VORPDrr,           X86::VORPDrm,            0 },
1315     { X86::VORPSrr,           X86::VORPSrm,            0 },
1316     { X86::VPACKSSDWrr,       X86::VPACKSSDWrm,        0 },
1317     { X86::VPACKSSWBrr,       X86::VPACKSSWBrm,        0 },
1318     { X86::VPACKUSDWrr,       X86::VPACKUSDWrm,        0 },
1319     { X86::VPACKUSWBrr,       X86::VPACKUSWBrm,        0 },
1320     { X86::VPADDBrr,          X86::VPADDBrm,           0 },
1321     { X86::VPADDDrr,          X86::VPADDDrm,           0 },
1322     { X86::VPADDQrr,          X86::VPADDQrm,           0 },
1323     { X86::VPADDSBrr,         X86::VPADDSBrm,          0 },
1324     { X86::VPADDSWrr,         X86::VPADDSWrm,          0 },
1325     { X86::VPADDUSBrr,        X86::VPADDUSBrm,         0 },
1326     { X86::VPADDUSWrr,        X86::VPADDUSWrm,         0 },
1327     { X86::VPADDWrr,          X86::VPADDWrm,           0 },
1328     { X86::VPALIGNR128rr,     X86::VPALIGNR128rm,      0 },
1329     { X86::VPANDNrr,          X86::VPANDNrm,           0 },
1330     { X86::VPANDrr,           X86::VPANDrm,            0 },
1331     { X86::VPAVGBrr,          X86::VPAVGBrm,           0 },
1332     { X86::VPAVGWrr,          X86::VPAVGWrm,           0 },
1333     { X86::VPBLENDVBrr,       X86::VPBLENDVBrm,        0 },
1334     { X86::VPBLENDWrri,       X86::VPBLENDWrmi,        0 },
1335     { X86::VPCLMULQDQrr,      X86::VPCLMULQDQrm,       0 },
1336     { X86::VPCMPEQBrr,        X86::VPCMPEQBrm,         0 },
1337     { X86::VPCMPEQDrr,        X86::VPCMPEQDrm,         0 },
1338     { X86::VPCMPEQQrr,        X86::VPCMPEQQrm,         0 },
1339     { X86::VPCMPEQWrr,        X86::VPCMPEQWrm,         0 },
1340     { X86::VPCMPGTBrr,        X86::VPCMPGTBrm,         0 },
1341     { X86::VPCMPGTDrr,        X86::VPCMPGTDrm,         0 },
1342     { X86::VPCMPGTQrr,        X86::VPCMPGTQrm,         0 },
1343     { X86::VPCMPGTWrr,        X86::VPCMPGTWrm,         0 },
1344     { X86::VPHADDDrr,         X86::VPHADDDrm,          0 },
1345     { X86::VPHADDSWrr128,     X86::VPHADDSWrm128,      0 },
1346     { X86::VPHADDWrr,         X86::VPHADDWrm,          0 },
1347     { X86::VPHSUBDrr,         X86::VPHSUBDrm,          0 },
1348     { X86::VPHSUBSWrr128,     X86::VPHSUBSWrm128,      0 },
1349     { X86::VPHSUBWrr,         X86::VPHSUBWrm,          0 },
1350     { X86::VPERMILPDrr,       X86::VPERMILPDrm,        0 },
1351     { X86::VPERMILPSrr,       X86::VPERMILPSrm,        0 },
1352     { X86::VPINSRBrr,         X86::VPINSRBrm,          0 },
1353     { X86::VPINSRDrr,         X86::VPINSRDrm,          0 },
1354     { X86::VPINSRQrr,         X86::VPINSRQrm,          0 },
1355     { X86::VPINSRWrri,        X86::VPINSRWrmi,         0 },
1356     { X86::VPMADDUBSWrr128,   X86::VPMADDUBSWrm128,    0 },
1357     { X86::VPMADDWDrr,        X86::VPMADDWDrm,         0 },
1358     { X86::VPMAXSWrr,         X86::VPMAXSWrm,          0 },
1359     { X86::VPMAXUBrr,         X86::VPMAXUBrm,          0 },
1360     { X86::VPMINSWrr,         X86::VPMINSWrm,          0 },
1361     { X86::VPMINUBrr,         X86::VPMINUBrm,          0 },
1362     { X86::VPMINSBrr,         X86::VPMINSBrm,          0 },
1363     { X86::VPMINSDrr,         X86::VPMINSDrm,          0 },
1364     { X86::VPMINUDrr,         X86::VPMINUDrm,          0 },
1365     { X86::VPMINUWrr,         X86::VPMINUWrm,          0 },
1366     { X86::VPMAXSBrr,         X86::VPMAXSBrm,          0 },
1367     { X86::VPMAXSDrr,         X86::VPMAXSDrm,          0 },
1368     { X86::VPMAXUDrr,         X86::VPMAXUDrm,          0 },
1369     { X86::VPMAXUWrr,         X86::VPMAXUWrm,          0 },
1370     { X86::VPMULDQrr,         X86::VPMULDQrm,          0 },
1371     { X86::VPMULHRSWrr128,    X86::VPMULHRSWrm128,     0 },
1372     { X86::VPMULHUWrr,        X86::VPMULHUWrm,         0 },
1373     { X86::VPMULHWrr,         X86::VPMULHWrm,          0 },
1374     { X86::VPMULLDrr,         X86::VPMULLDrm,          0 },
1375     { X86::VPMULLWrr,         X86::VPMULLWrm,          0 },
1376     { X86::VPMULUDQrr,        X86::VPMULUDQrm,         0 },
1377     { X86::VPORrr,            X86::VPORrm,             0 },
1378     { X86::VPSADBWrr,         X86::VPSADBWrm,          0 },
1379     { X86::VPSHUFBrr,         X86::VPSHUFBrm,          0 },
1380     { X86::VPSIGNBrr,         X86::VPSIGNBrm,          0 },
1381     { X86::VPSIGNWrr,         X86::VPSIGNWrm,          0 },
1382     { X86::VPSIGNDrr,         X86::VPSIGNDrm,          0 },
1383     { X86::VPSLLDrr,          X86::VPSLLDrm,           0 },
1384     { X86::VPSLLQrr,          X86::VPSLLQrm,           0 },
1385     { X86::VPSLLWrr,          X86::VPSLLWrm,           0 },
1386     { X86::VPSRADrr,          X86::VPSRADrm,           0 },
1387     { X86::VPSRAWrr,          X86::VPSRAWrm,           0 },
1388     { X86::VPSRLDrr,          X86::VPSRLDrm,           0 },
1389     { X86::VPSRLQrr,          X86::VPSRLQrm,           0 },
1390     { X86::VPSRLWrr,          X86::VPSRLWrm,           0 },
1391     { X86::VPSUBBrr,          X86::VPSUBBrm,           0 },
1392     { X86::VPSUBDrr,          X86::VPSUBDrm,           0 },
1393     { X86::VPSUBQrr,          X86::VPSUBQrm,           0 },
1394     { X86::VPSUBSBrr,         X86::VPSUBSBrm,          0 },
1395     { X86::VPSUBSWrr,         X86::VPSUBSWrm,          0 },
1396     { X86::VPSUBUSBrr,        X86::VPSUBUSBrm,         0 },
1397     { X86::VPSUBUSWrr,        X86::VPSUBUSWrm,         0 },
1398     { X86::VPSUBWrr,          X86::VPSUBWrm,           0 },
1399     { X86::VPUNPCKHBWrr,      X86::VPUNPCKHBWrm,       0 },
1400     { X86::VPUNPCKHDQrr,      X86::VPUNPCKHDQrm,       0 },
1401     { X86::VPUNPCKHQDQrr,     X86::VPUNPCKHQDQrm,      0 },
1402     { X86::VPUNPCKHWDrr,      X86::VPUNPCKHWDrm,       0 },
1403     { X86::VPUNPCKLBWrr,      X86::VPUNPCKLBWrm,       0 },
1404     { X86::VPUNPCKLDQrr,      X86::VPUNPCKLDQrm,       0 },
1405     { X86::VPUNPCKLQDQrr,     X86::VPUNPCKLQDQrm,      0 },
1406     { X86::VPUNPCKLWDrr,      X86::VPUNPCKLWDrm,       0 },
1407     { X86::VPXORrr,           X86::VPXORrm,            0 },
1408     { X86::VROUNDSDr,         X86::VROUNDSDm,          0 },
1409     { X86::VROUNDSSr,         X86::VROUNDSSm,          0 },
1410     { X86::VSHUFPDrri,        X86::VSHUFPDrmi,         0 },
1411     { X86::VSHUFPSrri,        X86::VSHUFPSrmi,         0 },
1412     { X86::VSUBPDrr,          X86::VSUBPDrm,           0 },
1413     { X86::VSUBPSrr,          X86::VSUBPSrm,           0 },
1414     { X86::VSUBSDrr,          X86::VSUBSDrm,           0 },
1415     { X86::VSUBSDrr_Int,      X86::VSUBSDrm_Int,       0 },
1416     { X86::VSUBSSrr,          X86::VSUBSSrm,           0 },
1417     { X86::VSUBSSrr_Int,      X86::VSUBSSrm_Int,       0 },
1418     { X86::VUNPCKHPDrr,       X86::VUNPCKHPDrm,        0 },
1419     { X86::VUNPCKHPSrr,       X86::VUNPCKHPSrm,        0 },
1420     { X86::VUNPCKLPDrr,       X86::VUNPCKLPDrm,        0 },
1421     { X86::VUNPCKLPSrr,       X86::VUNPCKLPSrm,        0 },
1422     { X86::VXORPDrr,          X86::VXORPDrm,           0 },
1423     { X86::VXORPSrr,          X86::VXORPSrm,           0 },
1424 
1425     // AVX 256-bit foldable instructions
1426     { X86::VADDPDYrr,         X86::VADDPDYrm,          0 },
1427     { X86::VADDPSYrr,         X86::VADDPSYrm,          0 },
1428     { X86::VADDSUBPDYrr,      X86::VADDSUBPDYrm,       0 },
1429     { X86::VADDSUBPSYrr,      X86::VADDSUBPSYrm,       0 },
1430     { X86::VANDNPDYrr,        X86::VANDNPDYrm,         0 },
1431     { X86::VANDNPSYrr,        X86::VANDNPSYrm,         0 },
1432     { X86::VANDPDYrr,         X86::VANDPDYrm,          0 },
1433     { X86::VANDPSYrr,         X86::VANDPSYrm,          0 },
1434     { X86::VBLENDPDYrri,      X86::VBLENDPDYrmi,       0 },
1435     { X86::VBLENDPSYrri,      X86::VBLENDPSYrmi,       0 },
1436     { X86::VBLENDVPDYrr,      X86::VBLENDVPDYrm,       0 },
1437     { X86::VBLENDVPSYrr,      X86::VBLENDVPSYrm,       0 },
1438     { X86::VCMPPDYrri,        X86::VCMPPDYrmi,         0 },
1439     { X86::VCMPPSYrri,        X86::VCMPPSYrmi,         0 },
1440     { X86::VDIVPDYrr,         X86::VDIVPDYrm,          0 },
1441     { X86::VDIVPSYrr,         X86::VDIVPSYrm,          0 },
1442     { X86::VDPPSYrri,         X86::VDPPSYrmi,          0 },
1443     { X86::VHADDPDYrr,        X86::VHADDPDYrm,         0 },
1444     { X86::VHADDPSYrr,        X86::VHADDPSYrm,         0 },
1445     { X86::VHSUBPDYrr,        X86::VHSUBPDYrm,         0 },
1446     { X86::VHSUBPSYrr,        X86::VHSUBPSYrm,         0 },
1447     { X86::VINSERTF128rr,     X86::VINSERTF128rm,      0 },
1448     { X86::VMAXPDYrr,         X86::VMAXPDYrm,          0 },
1449     { X86::VMAXPSYrr,         X86::VMAXPSYrm,          0 },
1450     { X86::VMINPDYrr,         X86::VMINPDYrm,          0 },
1451     { X86::VMINPSYrr,         X86::VMINPSYrm,          0 },
1452     { X86::VMULPDYrr,         X86::VMULPDYrm,          0 },
1453     { X86::VMULPSYrr,         X86::VMULPSYrm,          0 },
1454     { X86::VORPDYrr,          X86::VORPDYrm,           0 },
1455     { X86::VORPSYrr,          X86::VORPSYrm,           0 },
1456     { X86::VPERM2F128rr,      X86::VPERM2F128rm,       0 },
1457     { X86::VPERMILPDYrr,      X86::VPERMILPDYrm,       0 },
1458     { X86::VPERMILPSYrr,      X86::VPERMILPSYrm,       0 },
1459     { X86::VSHUFPDYrri,       X86::VSHUFPDYrmi,        0 },
1460     { X86::VSHUFPSYrri,       X86::VSHUFPSYrmi,        0 },
1461     { X86::VSUBPDYrr,         X86::VSUBPDYrm,          0 },
1462     { X86::VSUBPSYrr,         X86::VSUBPSYrm,          0 },
1463     { X86::VUNPCKHPDYrr,      X86::VUNPCKHPDYrm,       0 },
1464     { X86::VUNPCKHPSYrr,      X86::VUNPCKHPSYrm,       0 },
1465     { X86::VUNPCKLPDYrr,      X86::VUNPCKLPDYrm,       0 },
1466     { X86::VUNPCKLPSYrr,      X86::VUNPCKLPSYrm,       0 },
1467     { X86::VXORPDYrr,         X86::VXORPDYrm,          0 },
1468     { X86::VXORPSYrr,         X86::VXORPSYrm,          0 },
1469 
1470     // AVX2 foldable instructions
1471     { X86::VINSERTI128rr,     X86::VINSERTI128rm,      0 },
1472     { X86::VPACKSSDWYrr,      X86::VPACKSSDWYrm,       0 },
1473     { X86::VPACKSSWBYrr,      X86::VPACKSSWBYrm,       0 },
1474     { X86::VPACKUSDWYrr,      X86::VPACKUSDWYrm,       0 },
1475     { X86::VPACKUSWBYrr,      X86::VPACKUSWBYrm,       0 },
1476     { X86::VPADDBYrr,         X86::VPADDBYrm,          0 },
1477     { X86::VPADDDYrr,         X86::VPADDDYrm,          0 },
1478     { X86::VPADDQYrr,         X86::VPADDQYrm,          0 },
1479     { X86::VPADDSBYrr,        X86::VPADDSBYrm,         0 },
1480     { X86::VPADDSWYrr,        X86::VPADDSWYrm,         0 },
1481     { X86::VPADDUSBYrr,       X86::VPADDUSBYrm,        0 },
1482     { X86::VPADDUSWYrr,       X86::VPADDUSWYrm,        0 },
1483     { X86::VPADDWYrr,         X86::VPADDWYrm,          0 },
1484     { X86::VPALIGNR256rr,     X86::VPALIGNR256rm,      0 },
1485     { X86::VPANDNYrr,         X86::VPANDNYrm,          0 },
1486     { X86::VPANDYrr,          X86::VPANDYrm,           0 },
1487     { X86::VPAVGBYrr,         X86::VPAVGBYrm,          0 },
1488     { X86::VPAVGWYrr,         X86::VPAVGWYrm,          0 },
1489     { X86::VPBLENDDrri,       X86::VPBLENDDrmi,        0 },
1490     { X86::VPBLENDDYrri,      X86::VPBLENDDYrmi,       0 },
1491     { X86::VPBLENDVBYrr,      X86::VPBLENDVBYrm,       0 },
1492     { X86::VPBLENDWYrri,      X86::VPBLENDWYrmi,       0 },
1493     { X86::VPCMPEQBYrr,       X86::VPCMPEQBYrm,        0 },
1494     { X86::VPCMPEQDYrr,       X86::VPCMPEQDYrm,        0 },
1495     { X86::VPCMPEQQYrr,       X86::VPCMPEQQYrm,        0 },
1496     { X86::VPCMPEQWYrr,       X86::VPCMPEQWYrm,        0 },
1497     { X86::VPCMPGTBYrr,       X86::VPCMPGTBYrm,        0 },
1498     { X86::VPCMPGTDYrr,       X86::VPCMPGTDYrm,        0 },
1499     { X86::VPCMPGTQYrr,       X86::VPCMPGTQYrm,        0 },
1500     { X86::VPCMPGTWYrr,       X86::VPCMPGTWYrm,        0 },
1501     { X86::VPERM2I128rr,      X86::VPERM2I128rm,       0 },
1502     { X86::VPERMDYrr,         X86::VPERMDYrm,          0 },
1503     { X86::VPERMPSYrr,        X86::VPERMPSYrm,         0 },
1504     { X86::VPHADDDYrr,        X86::VPHADDDYrm,         0 },
1505     { X86::VPHADDSWrr256,     X86::VPHADDSWrm256,      0 },
1506     { X86::VPHADDWYrr,        X86::VPHADDWYrm,         0 },
1507     { X86::VPHSUBDYrr,        X86::VPHSUBDYrm,         0 },
1508     { X86::VPHSUBSWrr256,     X86::VPHSUBSWrm256,      0 },
1509     { X86::VPHSUBWYrr,        X86::VPHSUBWYrm,         0 },
1510     { X86::VPMADDUBSWrr256,   X86::VPMADDUBSWrm256,    0 },
1511     { X86::VPMADDWDYrr,       X86::VPMADDWDYrm,        0 },
1512     { X86::VPMAXSWYrr,        X86::VPMAXSWYrm,         0 },
1513     { X86::VPMAXUBYrr,        X86::VPMAXUBYrm,         0 },
1514     { X86::VPMINSWYrr,        X86::VPMINSWYrm,         0 },
1515     { X86::VPMINUBYrr,        X86::VPMINUBYrm,         0 },
1516     { X86::VPMINSBYrr,        X86::VPMINSBYrm,         0 },
1517     { X86::VPMINSDYrr,        X86::VPMINSDYrm,         0 },
1518     { X86::VPMINUDYrr,        X86::VPMINUDYrm,         0 },
1519     { X86::VPMINUWYrr,        X86::VPMINUWYrm,         0 },
1520     { X86::VPMAXSBYrr,        X86::VPMAXSBYrm,         0 },
1521     { X86::VPMAXSDYrr,        X86::VPMAXSDYrm,         0 },
1522     { X86::VPMAXUDYrr,        X86::VPMAXUDYrm,         0 },
1523     { X86::VPMAXUWYrr,        X86::VPMAXUWYrm,         0 },
1524     { X86::VMPSADBWYrri,      X86::VMPSADBWYrmi,       0 },
1525     { X86::VPMULDQYrr,        X86::VPMULDQYrm,         0 },
1526     { X86::VPMULHRSWrr256,    X86::VPMULHRSWrm256,     0 },
1527     { X86::VPMULHUWYrr,       X86::VPMULHUWYrm,        0 },
1528     { X86::VPMULHWYrr,        X86::VPMULHWYrm,         0 },
1529     { X86::VPMULLDYrr,        X86::VPMULLDYrm,         0 },
1530     { X86::VPMULLWYrr,        X86::VPMULLWYrm,         0 },
1531     { X86::VPMULUDQYrr,       X86::VPMULUDQYrm,        0 },
1532     { X86::VPORYrr,           X86::VPORYrm,            0 },
1533     { X86::VPSADBWYrr,        X86::VPSADBWYrm,         0 },
1534     { X86::VPSHUFBYrr,        X86::VPSHUFBYrm,         0 },
1535     { X86::VPSIGNBYrr,        X86::VPSIGNBYrm,         0 },
1536     { X86::VPSIGNWYrr,        X86::VPSIGNWYrm,         0 },
1537     { X86::VPSIGNDYrr,        X86::VPSIGNDYrm,         0 },
1538     { X86::VPSLLDYrr,         X86::VPSLLDYrm,          0 },
1539     { X86::VPSLLQYrr,         X86::VPSLLQYrm,          0 },
1540     { X86::VPSLLWYrr,         X86::VPSLLWYrm,          0 },
1541     { X86::VPSLLVDrr,         X86::VPSLLVDrm,          0 },
1542     { X86::VPSLLVDYrr,        X86::VPSLLVDYrm,         0 },
1543     { X86::VPSLLVQrr,         X86::VPSLLVQrm,          0 },
1544     { X86::VPSLLVQYrr,        X86::VPSLLVQYrm,         0 },
1545     { X86::VPSRADYrr,         X86::VPSRADYrm,          0 },
1546     { X86::VPSRAWYrr,         X86::VPSRAWYrm,          0 },
1547     { X86::VPSRAVDrr,         X86::VPSRAVDrm,          0 },
1548     { X86::VPSRAVDYrr,        X86::VPSRAVDYrm,         0 },
1549     { X86::VPSRLDYrr,         X86::VPSRLDYrm,          0 },
1550     { X86::VPSRLQYrr,         X86::VPSRLQYrm,          0 },
1551     { X86::VPSRLWYrr,         X86::VPSRLWYrm,          0 },
1552     { X86::VPSRLVDrr,         X86::VPSRLVDrm,          0 },
1553     { X86::VPSRLVDYrr,        X86::VPSRLVDYrm,         0 },
1554     { X86::VPSRLVQrr,         X86::VPSRLVQrm,          0 },
1555     { X86::VPSRLVQYrr,        X86::VPSRLVQYrm,         0 },
1556     { X86::VPSUBBYrr,         X86::VPSUBBYrm,          0 },
1557     { X86::VPSUBDYrr,         X86::VPSUBDYrm,          0 },
1558     { X86::VPSUBQYrr,         X86::VPSUBQYrm,          0 },
1559     { X86::VPSUBSBYrr,        X86::VPSUBSBYrm,         0 },
1560     { X86::VPSUBSWYrr,        X86::VPSUBSWYrm,         0 },
1561     { X86::VPSUBUSBYrr,       X86::VPSUBUSBYrm,        0 },
1562     { X86::VPSUBUSWYrr,       X86::VPSUBUSWYrm,        0 },
1563     { X86::VPSUBWYrr,         X86::VPSUBWYrm,          0 },
1564     { X86::VPUNPCKHBWYrr,     X86::VPUNPCKHBWYrm,      0 },
1565     { X86::VPUNPCKHDQYrr,     X86::VPUNPCKHDQYrm,      0 },
1566     { X86::VPUNPCKHQDQYrr,    X86::VPUNPCKHQDQYrm,     0 },
1567     { X86::VPUNPCKHWDYrr,     X86::VPUNPCKHWDYrm,      0 },
1568     { X86::VPUNPCKLBWYrr,     X86::VPUNPCKLBWYrm,      0 },
1569     { X86::VPUNPCKLDQYrr,     X86::VPUNPCKLDQYrm,      0 },
1570     { X86::VPUNPCKLQDQYrr,    X86::VPUNPCKLQDQYrm,     0 },
1571     { X86::VPUNPCKLWDYrr,     X86::VPUNPCKLWDYrm,      0 },
1572     { X86::VPXORYrr,          X86::VPXORYrm,           0 },
1573 
1574     // FMA4 foldable patterns
1575     { X86::VFMADDSS4rr,       X86::VFMADDSS4mr,        TB_ALIGN_NONE },
1576     { X86::VFMADDSD4rr,       X86::VFMADDSD4mr,        TB_ALIGN_NONE },
1577     { X86::VFMADDPS4rr,       X86::VFMADDPS4mr,        TB_ALIGN_NONE },
1578     { X86::VFMADDPD4rr,       X86::VFMADDPD4mr,        TB_ALIGN_NONE },
1579     { X86::VFMADDPS4rrY,      X86::VFMADDPS4mrY,       TB_ALIGN_NONE },
1580     { X86::VFMADDPD4rrY,      X86::VFMADDPD4mrY,       TB_ALIGN_NONE },
1581     { X86::VFNMADDSS4rr,      X86::VFNMADDSS4mr,       TB_ALIGN_NONE },
1582     { X86::VFNMADDSD4rr,      X86::VFNMADDSD4mr,       TB_ALIGN_NONE },
1583     { X86::VFNMADDPS4rr,      X86::VFNMADDPS4mr,       TB_ALIGN_NONE },
1584     { X86::VFNMADDPD4rr,      X86::VFNMADDPD4mr,       TB_ALIGN_NONE },
1585     { X86::VFNMADDPS4rrY,     X86::VFNMADDPS4mrY,      TB_ALIGN_NONE },
1586     { X86::VFNMADDPD4rrY,     X86::VFNMADDPD4mrY,      TB_ALIGN_NONE },
1587     { X86::VFMSUBSS4rr,       X86::VFMSUBSS4mr,        TB_ALIGN_NONE },
1588     { X86::VFMSUBSD4rr,       X86::VFMSUBSD4mr,        TB_ALIGN_NONE },
1589     { X86::VFMSUBPS4rr,       X86::VFMSUBPS4mr,        TB_ALIGN_NONE },
1590     { X86::VFMSUBPD4rr,       X86::VFMSUBPD4mr,        TB_ALIGN_NONE },
1591     { X86::VFMSUBPS4rrY,      X86::VFMSUBPS4mrY,       TB_ALIGN_NONE },
1592     { X86::VFMSUBPD4rrY,      X86::VFMSUBPD4mrY,       TB_ALIGN_NONE },
1593     { X86::VFNMSUBSS4rr,      X86::VFNMSUBSS4mr,       TB_ALIGN_NONE },
1594     { X86::VFNMSUBSD4rr,      X86::VFNMSUBSD4mr,       TB_ALIGN_NONE },
1595     { X86::VFNMSUBPS4rr,      X86::VFNMSUBPS4mr,       TB_ALIGN_NONE },
1596     { X86::VFNMSUBPD4rr,      X86::VFNMSUBPD4mr,       TB_ALIGN_NONE },
1597     { X86::VFNMSUBPS4rrY,     X86::VFNMSUBPS4mrY,      TB_ALIGN_NONE },
1598     { X86::VFNMSUBPD4rrY,     X86::VFNMSUBPD4mrY,      TB_ALIGN_NONE },
1599     { X86::VFMADDSUBPS4rr,    X86::VFMADDSUBPS4mr,     TB_ALIGN_NONE },
1600     { X86::VFMADDSUBPD4rr,    X86::VFMADDSUBPD4mr,     TB_ALIGN_NONE },
1601     { X86::VFMADDSUBPS4rrY,   X86::VFMADDSUBPS4mrY,    TB_ALIGN_NONE },
1602     { X86::VFMADDSUBPD4rrY,   X86::VFMADDSUBPD4mrY,    TB_ALIGN_NONE },
1603     { X86::VFMSUBADDPS4rr,    X86::VFMSUBADDPS4mr,     TB_ALIGN_NONE },
1604     { X86::VFMSUBADDPD4rr,    X86::VFMSUBADDPD4mr,     TB_ALIGN_NONE },
1605     { X86::VFMSUBADDPS4rrY,   X86::VFMSUBADDPS4mrY,    TB_ALIGN_NONE },
1606     { X86::VFMSUBADDPD4rrY,   X86::VFMSUBADDPD4mrY,    TB_ALIGN_NONE },
1607 
1608     // XOP foldable instructions
1609     { X86::VPCMOVrr,          X86::VPCMOVmr,            0 },
1610     { X86::VPCMOVrrY,         X86::VPCMOVmrY,           0 },
1611     { X86::VPCOMBri,          X86::VPCOMBmi,            0 },
1612     { X86::VPCOMDri,          X86::VPCOMDmi,            0 },
1613     { X86::VPCOMQri,          X86::VPCOMQmi,            0 },
1614     { X86::VPCOMWri,          X86::VPCOMWmi,            0 },
1615     { X86::VPCOMUBri,         X86::VPCOMUBmi,           0 },
1616     { X86::VPCOMUDri,         X86::VPCOMUDmi,           0 },
1617     { X86::VPCOMUQri,         X86::VPCOMUQmi,           0 },
1618     { X86::VPCOMUWri,         X86::VPCOMUWmi,           0 },
1619     { X86::VPERMIL2PDrr,      X86::VPERMIL2PDmr,        0 },
1620     { X86::VPERMIL2PDrrY,     X86::VPERMIL2PDmrY,       0 },
1621     { X86::VPERMIL2PSrr,      X86::VPERMIL2PSmr,        0 },
1622     { X86::VPERMIL2PSrrY,     X86::VPERMIL2PSmrY,       0 },
1623     { X86::VPMACSDDrr,        X86::VPMACSDDrm,          0 },
1624     { X86::VPMACSDQHrr,       X86::VPMACSDQHrm,         0 },
1625     { X86::VPMACSDQLrr,       X86::VPMACSDQLrm,         0 },
1626     { X86::VPMACSSDDrr,       X86::VPMACSSDDrm,         0 },
1627     { X86::VPMACSSDQHrr,      X86::VPMACSSDQHrm,        0 },
1628     { X86::VPMACSSDQLrr,      X86::VPMACSSDQLrm,        0 },
1629     { X86::VPMACSSWDrr,       X86::VPMACSSWDrm,         0 },
1630     { X86::VPMACSSWWrr,       X86::VPMACSSWWrm,         0 },
1631     { X86::VPMACSWDrr,        X86::VPMACSWDrm,          0 },
1632     { X86::VPMACSWWrr,        X86::VPMACSWWrm,          0 },
1633     { X86::VPMADCSSWDrr,      X86::VPMADCSSWDrm,        0 },
1634     { X86::VPMADCSWDrr,       X86::VPMADCSWDrm,         0 },
1635     { X86::VPPERMrr,          X86::VPPERMmr,            0 },
1636     { X86::VPROTBrr,          X86::VPROTBrm,            0 },
1637     { X86::VPROTDrr,          X86::VPROTDrm,            0 },
1638     { X86::VPROTQrr,          X86::VPROTQrm,            0 },
1639     { X86::VPROTWrr,          X86::VPROTWrm,            0 },
1640     { X86::VPSHABrr,          X86::VPSHABrm,            0 },
1641     { X86::VPSHADrr,          X86::VPSHADrm,            0 },
1642     { X86::VPSHAQrr,          X86::VPSHAQrm,            0 },
1643     { X86::VPSHAWrr,          X86::VPSHAWrm,            0 },
1644     { X86::VPSHLBrr,          X86::VPSHLBrm,            0 },
1645     { X86::VPSHLDrr,          X86::VPSHLDrm,            0 },
1646     { X86::VPSHLQrr,          X86::VPSHLQrm,            0 },
1647     { X86::VPSHLWrr,          X86::VPSHLWrm,            0 },
1648 
1649     // BMI/BMI2 foldable instructions
1650     { X86::ANDN32rr,          X86::ANDN32rm,            0 },
1651     { X86::ANDN64rr,          X86::ANDN64rm,            0 },
1652     { X86::MULX32rr,          X86::MULX32rm,            0 },
1653     { X86::MULX64rr,          X86::MULX64rm,            0 },
1654     { X86::PDEP32rr,          X86::PDEP32rm,            0 },
1655     { X86::PDEP64rr,          X86::PDEP64rm,            0 },
1656     { X86::PEXT32rr,          X86::PEXT32rm,            0 },
1657     { X86::PEXT64rr,          X86::PEXT64rm,            0 },
1658 
1659     // AVX-512 foldable instructions
1660     { X86::VADDPSZrr,         X86::VADDPSZrm,           0 },
1661     { X86::VADDPDZrr,         X86::VADDPDZrm,           0 },
1662     { X86::VSUBPSZrr,         X86::VSUBPSZrm,           0 },
1663     { X86::VSUBPDZrr,         X86::VSUBPDZrm,           0 },
1664     { X86::VMULPSZrr,         X86::VMULPSZrm,           0 },
1665     { X86::VMULPDZrr,         X86::VMULPDZrm,           0 },
1666     { X86::VDIVPSZrr,         X86::VDIVPSZrm,           0 },
1667     { X86::VDIVPDZrr,         X86::VDIVPDZrm,           0 },
1668     { X86::VMINPSZrr,         X86::VMINPSZrm,           0 },
1669     { X86::VMINPDZrr,         X86::VMINPDZrm,           0 },
1670     { X86::VMAXPSZrr,         X86::VMAXPSZrm,           0 },
1671     { X86::VMAXPDZrr,         X86::VMAXPDZrm,           0 },
1672     { X86::VPADDDZrr,         X86::VPADDDZrm,           0 },
1673     { X86::VPADDQZrr,         X86::VPADDQZrm,           0 },
1674     { X86::VPERMPDZri,        X86::VPERMPDZmi,          0 },
1675     { X86::VPERMPSZrr,        X86::VPERMPSZrm,          0 },
1676     { X86::VPMAXSDZrr,        X86::VPMAXSDZrm,          0 },
1677     { X86::VPMAXSQZrr,        X86::VPMAXSQZrm,          0 },
1678     { X86::VPMAXUDZrr,        X86::VPMAXUDZrm,          0 },
1679     { X86::VPMAXUQZrr,        X86::VPMAXUQZrm,          0 },
1680     { X86::VPMINSDZrr,        X86::VPMINSDZrm,          0 },
1681     { X86::VPMINSQZrr,        X86::VPMINSQZrm,          0 },
1682     { X86::VPMINUDZrr,        X86::VPMINUDZrm,          0 },
1683     { X86::VPMINUQZrr,        X86::VPMINUQZrm,          0 },
1684     { X86::VPMULDQZrr,        X86::VPMULDQZrm,          0 },
1685     { X86::VPSLLVDZrr,        X86::VPSLLVDZrm,          0 },
1686     { X86::VPSLLVQZrr,        X86::VPSLLVQZrm,          0 },
1687     { X86::VPSRAVDZrr,        X86::VPSRAVDZrm,          0 },
1688     { X86::VPSRLVDZrr,        X86::VPSRLVDZrm,          0 },
1689     { X86::VPSRLVQZrr,        X86::VPSRLVQZrm,          0 },
1690     { X86::VPSUBDZrr,         X86::VPSUBDZrm,           0 },
1691     { X86::VPSUBQZrr,         X86::VPSUBQZrm,           0 },
1692     { X86::VSHUFPDZrri,       X86::VSHUFPDZrmi,         0 },
1693     { X86::VSHUFPSZrri,       X86::VSHUFPSZrmi,         0 },
1694     { X86::VALIGNQZrri,       X86::VALIGNQZrmi,         0 },
1695     { X86::VALIGNDZrri,       X86::VALIGNDZrmi,         0 },
1696     { X86::VPMULUDQZrr,       X86::VPMULUDQZrm,         0 },
1697     { X86::VBROADCASTSSZrkz,  X86::VBROADCASTSSZmkz,    TB_NO_REVERSE },
1698     { X86::VBROADCASTSDZrkz,  X86::VBROADCASTSDZmkz,    TB_NO_REVERSE },
1699 
1700     // AVX-512{F,VL} foldable instructions
1701     { X86::VBROADCASTSSZ256rkz,  X86::VBROADCASTSSZ256mkz,      TB_NO_REVERSE },
1702     { X86::VBROADCASTSDZ256rkz,  X86::VBROADCASTSDZ256mkz,      TB_NO_REVERSE },
1703     { X86::VBROADCASTSSZ128rkz,  X86::VBROADCASTSSZ128mkz,      TB_NO_REVERSE },
1704 
1705     // AVX-512{F,VL} foldable instructions
1706     { X86::VADDPDZ128rr,      X86::VADDPDZ128rm,        0 },
1707     { X86::VADDPDZ256rr,      X86::VADDPDZ256rm,        0 },
1708     { X86::VADDPSZ128rr,      X86::VADDPSZ128rm,        0 },
1709     { X86::VADDPSZ256rr,      X86::VADDPSZ256rm,        0 },
1710 
1711     // AES foldable instructions
1712     { X86::AESDECLASTrr,      X86::AESDECLASTrm,        TB_ALIGN_16 },
1713     { X86::AESDECrr,          X86::AESDECrm,            TB_ALIGN_16 },
1714     { X86::AESENCLASTrr,      X86::AESENCLASTrm,        TB_ALIGN_16 },
1715     { X86::AESENCrr,          X86::AESENCrm,            TB_ALIGN_16 },
1716     { X86::VAESDECLASTrr,     X86::VAESDECLASTrm,       0 },
1717     { X86::VAESDECrr,         X86::VAESDECrm,           0 },
1718     { X86::VAESENCLASTrr,     X86::VAESENCLASTrm,       0 },
1719     { X86::VAESENCrr,         X86::VAESENCrm,           0 },
1720 
1721     // SHA foldable instructions
1722     { X86::SHA1MSG1rr,        X86::SHA1MSG1rm,          TB_ALIGN_16 },
1723     { X86::SHA1MSG2rr,        X86::SHA1MSG2rm,          TB_ALIGN_16 },
1724     { X86::SHA1NEXTErr,       X86::SHA1NEXTErm,         TB_ALIGN_16 },
1725     { X86::SHA1RNDS4rri,      X86::SHA1RNDS4rmi,        TB_ALIGN_16 },
1726     { X86::SHA256MSG1rr,      X86::SHA256MSG1rm,        TB_ALIGN_16 },
1727     { X86::SHA256MSG2rr,      X86::SHA256MSG2rm,        TB_ALIGN_16 },
1728     { X86::SHA256RNDS2rr,     X86::SHA256RNDS2rm,       TB_ALIGN_16 }
1729   };
1730 
1731   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2) {
1732     AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1733                   Entry.RegOp, Entry.MemOp,
1734                   // Index 2, folded load
1735                   Entry.Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1736   }
1737 
1738   static const X86MemoryFoldTableEntry MemoryFoldTable3[] = {
1739     // FMA foldable instructions
1740     { X86::VFMADDSSr231r,         X86::VFMADDSSr231m,         TB_ALIGN_NONE },
1741     { X86::VFMADDSDr231r,         X86::VFMADDSDr231m,         TB_ALIGN_NONE },
1742     { X86::VFMADDSSr132r,         X86::VFMADDSSr132m,         TB_ALIGN_NONE },
1743     { X86::VFMADDSDr132r,         X86::VFMADDSDr132m,         TB_ALIGN_NONE },
1744     { X86::VFMADDSSr213r,         X86::VFMADDSSr213m,         TB_ALIGN_NONE },
1745     { X86::VFMADDSDr213r,         X86::VFMADDSDr213m,         TB_ALIGN_NONE },
1746 
1747     { X86::VFMADDPSr231r,         X86::VFMADDPSr231m,         TB_ALIGN_NONE },
1748     { X86::VFMADDPDr231r,         X86::VFMADDPDr231m,         TB_ALIGN_NONE },
1749     { X86::VFMADDPSr132r,         X86::VFMADDPSr132m,         TB_ALIGN_NONE },
1750     { X86::VFMADDPDr132r,         X86::VFMADDPDr132m,         TB_ALIGN_NONE },
1751     { X86::VFMADDPSr213r,         X86::VFMADDPSr213m,         TB_ALIGN_NONE },
1752     { X86::VFMADDPDr213r,         X86::VFMADDPDr213m,         TB_ALIGN_NONE },
1753     { X86::VFMADDPSr231rY,        X86::VFMADDPSr231mY,        TB_ALIGN_NONE },
1754     { X86::VFMADDPDr231rY,        X86::VFMADDPDr231mY,        TB_ALIGN_NONE },
1755     { X86::VFMADDPSr132rY,        X86::VFMADDPSr132mY,        TB_ALIGN_NONE },
1756     { X86::VFMADDPDr132rY,        X86::VFMADDPDr132mY,        TB_ALIGN_NONE },
1757     { X86::VFMADDPSr213rY,        X86::VFMADDPSr213mY,        TB_ALIGN_NONE },
1758     { X86::VFMADDPDr213rY,        X86::VFMADDPDr213mY,        TB_ALIGN_NONE },
1759 
1760     { X86::VFNMADDSSr231r,        X86::VFNMADDSSr231m,        TB_ALIGN_NONE },
1761     { X86::VFNMADDSDr231r,        X86::VFNMADDSDr231m,        TB_ALIGN_NONE },
1762     { X86::VFNMADDSSr132r,        X86::VFNMADDSSr132m,        TB_ALIGN_NONE },
1763     { X86::VFNMADDSDr132r,        X86::VFNMADDSDr132m,        TB_ALIGN_NONE },
1764     { X86::VFNMADDSSr213r,        X86::VFNMADDSSr213m,        TB_ALIGN_NONE },
1765     { X86::VFNMADDSDr213r,        X86::VFNMADDSDr213m,        TB_ALIGN_NONE },
1766 
1767     { X86::VFNMADDPSr231r,        X86::VFNMADDPSr231m,        TB_ALIGN_NONE },
1768     { X86::VFNMADDPDr231r,        X86::VFNMADDPDr231m,        TB_ALIGN_NONE },
1769     { X86::VFNMADDPSr132r,        X86::VFNMADDPSr132m,        TB_ALIGN_NONE },
1770     { X86::VFNMADDPDr132r,        X86::VFNMADDPDr132m,        TB_ALIGN_NONE },
1771     { X86::VFNMADDPSr213r,        X86::VFNMADDPSr213m,        TB_ALIGN_NONE },
1772     { X86::VFNMADDPDr213r,        X86::VFNMADDPDr213m,        TB_ALIGN_NONE },
1773     { X86::VFNMADDPSr231rY,       X86::VFNMADDPSr231mY,       TB_ALIGN_NONE },
1774     { X86::VFNMADDPDr231rY,       X86::VFNMADDPDr231mY,       TB_ALIGN_NONE },
1775     { X86::VFNMADDPSr132rY,       X86::VFNMADDPSr132mY,       TB_ALIGN_NONE },
1776     { X86::VFNMADDPDr132rY,       X86::VFNMADDPDr132mY,       TB_ALIGN_NONE },
1777     { X86::VFNMADDPSr213rY,       X86::VFNMADDPSr213mY,       TB_ALIGN_NONE },
1778     { X86::VFNMADDPDr213rY,       X86::VFNMADDPDr213mY,       TB_ALIGN_NONE },
1779 
1780     { X86::VFMSUBSSr231r,         X86::VFMSUBSSr231m,         TB_ALIGN_NONE },
1781     { X86::VFMSUBSDr231r,         X86::VFMSUBSDr231m,         TB_ALIGN_NONE },
1782     { X86::VFMSUBSSr132r,         X86::VFMSUBSSr132m,         TB_ALIGN_NONE },
1783     { X86::VFMSUBSDr132r,         X86::VFMSUBSDr132m,         TB_ALIGN_NONE },
1784     { X86::VFMSUBSSr213r,         X86::VFMSUBSSr213m,         TB_ALIGN_NONE },
1785     { X86::VFMSUBSDr213r,         X86::VFMSUBSDr213m,         TB_ALIGN_NONE },
1786 
1787     { X86::VFMSUBPSr231r,         X86::VFMSUBPSr231m,         TB_ALIGN_NONE },
1788     { X86::VFMSUBPDr231r,         X86::VFMSUBPDr231m,         TB_ALIGN_NONE },
1789     { X86::VFMSUBPSr132r,         X86::VFMSUBPSr132m,         TB_ALIGN_NONE },
1790     { X86::VFMSUBPDr132r,         X86::VFMSUBPDr132m,         TB_ALIGN_NONE },
1791     { X86::VFMSUBPSr213r,         X86::VFMSUBPSr213m,         TB_ALIGN_NONE },
1792     { X86::VFMSUBPDr213r,         X86::VFMSUBPDr213m,         TB_ALIGN_NONE },
1793     { X86::VFMSUBPSr231rY,        X86::VFMSUBPSr231mY,        TB_ALIGN_NONE },
1794     { X86::VFMSUBPDr231rY,        X86::VFMSUBPDr231mY,        TB_ALIGN_NONE },
1795     { X86::VFMSUBPSr132rY,        X86::VFMSUBPSr132mY,        TB_ALIGN_NONE },
1796     { X86::VFMSUBPDr132rY,        X86::VFMSUBPDr132mY,        TB_ALIGN_NONE },
1797     { X86::VFMSUBPSr213rY,        X86::VFMSUBPSr213mY,        TB_ALIGN_NONE },
1798     { X86::VFMSUBPDr213rY,        X86::VFMSUBPDr213mY,        TB_ALIGN_NONE },
1799 
1800     { X86::VFNMSUBSSr231r,        X86::VFNMSUBSSr231m,        TB_ALIGN_NONE },
1801     { X86::VFNMSUBSDr231r,        X86::VFNMSUBSDr231m,        TB_ALIGN_NONE },
1802     { X86::VFNMSUBSSr132r,        X86::VFNMSUBSSr132m,        TB_ALIGN_NONE },
1803     { X86::VFNMSUBSDr132r,        X86::VFNMSUBSDr132m,        TB_ALIGN_NONE },
1804     { X86::VFNMSUBSSr213r,        X86::VFNMSUBSSr213m,        TB_ALIGN_NONE },
1805     { X86::VFNMSUBSDr213r,        X86::VFNMSUBSDr213m,        TB_ALIGN_NONE },
1806 
1807     { X86::VFNMSUBPSr231r,        X86::VFNMSUBPSr231m,        TB_ALIGN_NONE },
1808     { X86::VFNMSUBPDr231r,        X86::VFNMSUBPDr231m,        TB_ALIGN_NONE },
1809     { X86::VFNMSUBPSr132r,        X86::VFNMSUBPSr132m,        TB_ALIGN_NONE },
1810     { X86::VFNMSUBPDr132r,        X86::VFNMSUBPDr132m,        TB_ALIGN_NONE },
1811     { X86::VFNMSUBPSr213r,        X86::VFNMSUBPSr213m,        TB_ALIGN_NONE },
1812     { X86::VFNMSUBPDr213r,        X86::VFNMSUBPDr213m,        TB_ALIGN_NONE },
1813     { X86::VFNMSUBPSr231rY,       X86::VFNMSUBPSr231mY,       TB_ALIGN_NONE },
1814     { X86::VFNMSUBPDr231rY,       X86::VFNMSUBPDr231mY,       TB_ALIGN_NONE },
1815     { X86::VFNMSUBPSr132rY,       X86::VFNMSUBPSr132mY,       TB_ALIGN_NONE },
1816     { X86::VFNMSUBPDr132rY,       X86::VFNMSUBPDr132mY,       TB_ALIGN_NONE },
1817     { X86::VFNMSUBPSr213rY,       X86::VFNMSUBPSr213mY,       TB_ALIGN_NONE },
1818     { X86::VFNMSUBPDr213rY,       X86::VFNMSUBPDr213mY,       TB_ALIGN_NONE },
1819 
1820     { X86::VFMADDSUBPSr231r,      X86::VFMADDSUBPSr231m,      TB_ALIGN_NONE },
1821     { X86::VFMADDSUBPDr231r,      X86::VFMADDSUBPDr231m,      TB_ALIGN_NONE },
1822     { X86::VFMADDSUBPSr132r,      X86::VFMADDSUBPSr132m,      TB_ALIGN_NONE },
1823     { X86::VFMADDSUBPDr132r,      X86::VFMADDSUBPDr132m,      TB_ALIGN_NONE },
1824     { X86::VFMADDSUBPSr213r,      X86::VFMADDSUBPSr213m,      TB_ALIGN_NONE },
1825     { X86::VFMADDSUBPDr213r,      X86::VFMADDSUBPDr213m,      TB_ALIGN_NONE },
1826     { X86::VFMADDSUBPSr231rY,     X86::VFMADDSUBPSr231mY,     TB_ALIGN_NONE },
1827     { X86::VFMADDSUBPDr231rY,     X86::VFMADDSUBPDr231mY,     TB_ALIGN_NONE },
1828     { X86::VFMADDSUBPSr132rY,     X86::VFMADDSUBPSr132mY,     TB_ALIGN_NONE },
1829     { X86::VFMADDSUBPDr132rY,     X86::VFMADDSUBPDr132mY,     TB_ALIGN_NONE },
1830     { X86::VFMADDSUBPSr213rY,     X86::VFMADDSUBPSr213mY,     TB_ALIGN_NONE },
1831     { X86::VFMADDSUBPDr213rY,     X86::VFMADDSUBPDr213mY,     TB_ALIGN_NONE },
1832 
1833     { X86::VFMSUBADDPSr231r,      X86::VFMSUBADDPSr231m,      TB_ALIGN_NONE },
1834     { X86::VFMSUBADDPDr231r,      X86::VFMSUBADDPDr231m,      TB_ALIGN_NONE },
1835     { X86::VFMSUBADDPSr132r,      X86::VFMSUBADDPSr132m,      TB_ALIGN_NONE },
1836     { X86::VFMSUBADDPDr132r,      X86::VFMSUBADDPDr132m,      TB_ALIGN_NONE },
1837     { X86::VFMSUBADDPSr213r,      X86::VFMSUBADDPSr213m,      TB_ALIGN_NONE },
1838     { X86::VFMSUBADDPDr213r,      X86::VFMSUBADDPDr213m,      TB_ALIGN_NONE },
1839     { X86::VFMSUBADDPSr231rY,     X86::VFMSUBADDPSr231mY,     TB_ALIGN_NONE },
1840     { X86::VFMSUBADDPDr231rY,     X86::VFMSUBADDPDr231mY,     TB_ALIGN_NONE },
1841     { X86::VFMSUBADDPSr132rY,     X86::VFMSUBADDPSr132mY,     TB_ALIGN_NONE },
1842     { X86::VFMSUBADDPDr132rY,     X86::VFMSUBADDPDr132mY,     TB_ALIGN_NONE },
1843     { X86::VFMSUBADDPSr213rY,     X86::VFMSUBADDPSr213mY,     TB_ALIGN_NONE },
1844     { X86::VFMSUBADDPDr213rY,     X86::VFMSUBADDPDr213mY,     TB_ALIGN_NONE },
1845 
1846     // FMA4 foldable patterns
1847     { X86::VFMADDSS4rr,           X86::VFMADDSS4rm,           TB_ALIGN_NONE },
1848     { X86::VFMADDSD4rr,           X86::VFMADDSD4rm,           TB_ALIGN_NONE },
1849     { X86::VFMADDPS4rr,           X86::VFMADDPS4rm,           TB_ALIGN_NONE },
1850     { X86::VFMADDPD4rr,           X86::VFMADDPD4rm,           TB_ALIGN_NONE },
1851     { X86::VFMADDPS4rrY,          X86::VFMADDPS4rmY,          TB_ALIGN_NONE },
1852     { X86::VFMADDPD4rrY,          X86::VFMADDPD4rmY,          TB_ALIGN_NONE },
1853     { X86::VFNMADDSS4rr,          X86::VFNMADDSS4rm,          TB_ALIGN_NONE },
1854     { X86::VFNMADDSD4rr,          X86::VFNMADDSD4rm,          TB_ALIGN_NONE },
1855     { X86::VFNMADDPS4rr,          X86::VFNMADDPS4rm,          TB_ALIGN_NONE },
1856     { X86::VFNMADDPD4rr,          X86::VFNMADDPD4rm,          TB_ALIGN_NONE },
1857     { X86::VFNMADDPS4rrY,         X86::VFNMADDPS4rmY,         TB_ALIGN_NONE },
1858     { X86::VFNMADDPD4rrY,         X86::VFNMADDPD4rmY,         TB_ALIGN_NONE },
1859     { X86::VFMSUBSS4rr,           X86::VFMSUBSS4rm,           TB_ALIGN_NONE },
1860     { X86::VFMSUBSD4rr,           X86::VFMSUBSD4rm,           TB_ALIGN_NONE },
1861     { X86::VFMSUBPS4rr,           X86::VFMSUBPS4rm,           TB_ALIGN_NONE },
1862     { X86::VFMSUBPD4rr,           X86::VFMSUBPD4rm,           TB_ALIGN_NONE },
1863     { X86::VFMSUBPS4rrY,          X86::VFMSUBPS4rmY,          TB_ALIGN_NONE },
1864     { X86::VFMSUBPD4rrY,          X86::VFMSUBPD4rmY,          TB_ALIGN_NONE },
1865     { X86::VFNMSUBSS4rr,          X86::VFNMSUBSS4rm,          TB_ALIGN_NONE },
1866     { X86::VFNMSUBSD4rr,          X86::VFNMSUBSD4rm,          TB_ALIGN_NONE },
1867     { X86::VFNMSUBPS4rr,          X86::VFNMSUBPS4rm,          TB_ALIGN_NONE },
1868     { X86::VFNMSUBPD4rr,          X86::VFNMSUBPD4rm,          TB_ALIGN_NONE },
1869     { X86::VFNMSUBPS4rrY,         X86::VFNMSUBPS4rmY,         TB_ALIGN_NONE },
1870     { X86::VFNMSUBPD4rrY,         X86::VFNMSUBPD4rmY,         TB_ALIGN_NONE },
1871     { X86::VFMADDSUBPS4rr,        X86::VFMADDSUBPS4rm,        TB_ALIGN_NONE },
1872     { X86::VFMADDSUBPD4rr,        X86::VFMADDSUBPD4rm,        TB_ALIGN_NONE },
1873     { X86::VFMADDSUBPS4rrY,       X86::VFMADDSUBPS4rmY,       TB_ALIGN_NONE },
1874     { X86::VFMADDSUBPD4rrY,       X86::VFMADDSUBPD4rmY,       TB_ALIGN_NONE },
1875     { X86::VFMSUBADDPS4rr,        X86::VFMSUBADDPS4rm,        TB_ALIGN_NONE },
1876     { X86::VFMSUBADDPD4rr,        X86::VFMSUBADDPD4rm,        TB_ALIGN_NONE },
1877     { X86::VFMSUBADDPS4rrY,       X86::VFMSUBADDPS4rmY,       TB_ALIGN_NONE },
1878     { X86::VFMSUBADDPD4rrY,       X86::VFMSUBADDPD4rmY,       TB_ALIGN_NONE },
1879 
1880     // XOP foldable instructions
1881     { X86::VPCMOVrr,              X86::VPCMOVrm,              0 },
1882     { X86::VPCMOVrrY,             X86::VPCMOVrmY,             0 },
1883     { X86::VPERMIL2PDrr,          X86::VPERMIL2PDrm,          0 },
1884     { X86::VPERMIL2PDrrY,         X86::VPERMIL2PDrmY,         0 },
1885     { X86::VPERMIL2PSrr,          X86::VPERMIL2PSrm,          0 },
1886     { X86::VPERMIL2PSrrY,         X86::VPERMIL2PSrmY,         0 },
1887     { X86::VPPERMrr,              X86::VPPERMrm,              0 },
1888 
1889     // AVX-512 VPERMI instructions with 3 source operands.
1890     { X86::VPERMI2Drr,            X86::VPERMI2Drm,            0 },
1891     { X86::VPERMI2Qrr,            X86::VPERMI2Qrm,            0 },
1892     { X86::VPERMI2PSrr,           X86::VPERMI2PSrm,           0 },
1893     { X86::VPERMI2PDrr,           X86::VPERMI2PDrm,           0 },
1894     { X86::VBLENDMPDZrr,          X86::VBLENDMPDZrm,          0 },
1895     { X86::VBLENDMPSZrr,          X86::VBLENDMPSZrm,          0 },
1896     { X86::VPBLENDMDZrr,          X86::VPBLENDMDZrm,          0 },
1897     { X86::VPBLENDMQZrr,          X86::VPBLENDMQZrm,          0 },
1898     { X86::VBROADCASTSSZrk,       X86::VBROADCASTSSZmk,       TB_NO_REVERSE },
1899     { X86::VBROADCASTSDZrk,       X86::VBROADCASTSDZmk,       TB_NO_REVERSE },
1900     { X86::VBROADCASTSSZ256rk,    X86::VBROADCASTSSZ256mk,    TB_NO_REVERSE },
1901     { X86::VBROADCASTSDZ256rk,    X86::VBROADCASTSDZ256mk,    TB_NO_REVERSE },
1902     { X86::VBROADCASTSSZ128rk,    X86::VBROADCASTSSZ128mk,    TB_NO_REVERSE },
1903      // AVX-512 arithmetic instructions
1904     { X86::VADDPSZrrkz,           X86::VADDPSZrmkz,           0 },
1905     { X86::VADDPDZrrkz,           X86::VADDPDZrmkz,           0 },
1906     { X86::VSUBPSZrrkz,           X86::VSUBPSZrmkz,           0 },
1907     { X86::VSUBPDZrrkz,           X86::VSUBPDZrmkz,           0 },
1908     { X86::VMULPSZrrkz,           X86::VMULPSZrmkz,           0 },
1909     { X86::VMULPDZrrkz,           X86::VMULPDZrmkz,           0 },
1910     { X86::VDIVPSZrrkz,           X86::VDIVPSZrmkz,           0 },
1911     { X86::VDIVPDZrrkz,           X86::VDIVPDZrmkz,           0 },
1912     { X86::VMINPSZrrkz,           X86::VMINPSZrmkz,           0 },
1913     { X86::VMINPDZrrkz,           X86::VMINPDZrmkz,           0 },
1914     { X86::VMAXPSZrrkz,           X86::VMAXPSZrmkz,           0 },
1915     { X86::VMAXPDZrrkz,           X86::VMAXPDZrmkz,           0 },
1916     // AVX-512{F,VL} arithmetic instructions 256-bit
1917     { X86::VADDPSZ256rrkz,        X86::VADDPSZ256rmkz,        0 },
1918     { X86::VADDPDZ256rrkz,        X86::VADDPDZ256rmkz,        0 },
1919     { X86::VSUBPSZ256rrkz,        X86::VSUBPSZ256rmkz,        0 },
1920     { X86::VSUBPDZ256rrkz,        X86::VSUBPDZ256rmkz,        0 },
1921     { X86::VMULPSZ256rrkz,        X86::VMULPSZ256rmkz,        0 },
1922     { X86::VMULPDZ256rrkz,        X86::VMULPDZ256rmkz,        0 },
1923     { X86::VDIVPSZ256rrkz,        X86::VDIVPSZ256rmkz,        0 },
1924     { X86::VDIVPDZ256rrkz,        X86::VDIVPDZ256rmkz,        0 },
1925     { X86::VMINPSZ256rrkz,        X86::VMINPSZ256rmkz,        0 },
1926     { X86::VMINPDZ256rrkz,        X86::VMINPDZ256rmkz,        0 },
1927     { X86::VMAXPSZ256rrkz,        X86::VMAXPSZ256rmkz,        0 },
1928     { X86::VMAXPDZ256rrkz,        X86::VMAXPDZ256rmkz,        0 },
1929     // AVX-512{F,VL} arithmetic instructions 128-bit
1930     { X86::VADDPSZ128rrkz,        X86::VADDPSZ128rmkz,        0 },
1931     { X86::VADDPDZ128rrkz,        X86::VADDPDZ128rmkz,        0 },
1932     { X86::VSUBPSZ128rrkz,        X86::VSUBPSZ128rmkz,        0 },
1933     { X86::VSUBPDZ128rrkz,        X86::VSUBPDZ128rmkz,        0 },
1934     { X86::VMULPSZ128rrkz,        X86::VMULPSZ128rmkz,        0 },
1935     { X86::VMULPDZ128rrkz,        X86::VMULPDZ128rmkz,        0 },
1936     { X86::VDIVPSZ128rrkz,        X86::VDIVPSZ128rmkz,        0 },
1937     { X86::VDIVPDZ128rrkz,        X86::VDIVPDZ128rmkz,        0 },
1938     { X86::VMINPSZ128rrkz,        X86::VMINPSZ128rmkz,        0 },
1939     { X86::VMINPDZ128rrkz,        X86::VMINPDZ128rmkz,        0 },
1940     { X86::VMAXPSZ128rrkz,        X86::VMAXPSZ128rmkz,        0 },
1941     { X86::VMAXPDZ128rrkz,        X86::VMAXPDZ128rmkz,        0 }
1942   };
1943 
1944   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable3) {
1945     AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1946                   Entry.RegOp, Entry.MemOp,
1947                   // Index 3, folded load
1948                   Entry.Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1949   }
1950 
1951   static const X86MemoryFoldTableEntry MemoryFoldTable4[] = {
1952      // AVX-512 foldable instructions
1953     { X86::VADDPSZrrk,         X86::VADDPSZrmk,           0 },
1954     { X86::VADDPDZrrk,         X86::VADDPDZrmk,           0 },
1955     { X86::VSUBPSZrrk,         X86::VSUBPSZrmk,           0 },
1956     { X86::VSUBPDZrrk,         X86::VSUBPDZrmk,           0 },
1957     { X86::VMULPSZrrk,         X86::VMULPSZrmk,           0 },
1958     { X86::VMULPDZrrk,         X86::VMULPDZrmk,           0 },
1959     { X86::VDIVPSZrrk,         X86::VDIVPSZrmk,           0 },
1960     { X86::VDIVPDZrrk,         X86::VDIVPDZrmk,           0 },
1961     { X86::VMINPSZrrk,         X86::VMINPSZrmk,           0 },
1962     { X86::VMINPDZrrk,         X86::VMINPDZrmk,           0 },
1963     { X86::VMAXPSZrrk,         X86::VMAXPSZrmk,           0 },
1964     { X86::VMAXPDZrrk,         X86::VMAXPDZrmk,           0 },
1965     // AVX-512{F,VL} foldable instructions 256-bit
1966     { X86::VADDPSZ256rrk,      X86::VADDPSZ256rmk,        0 },
1967     { X86::VADDPDZ256rrk,      X86::VADDPDZ256rmk,        0 },
1968     { X86::VSUBPSZ256rrk,      X86::VSUBPSZ256rmk,        0 },
1969     { X86::VSUBPDZ256rrk,      X86::VSUBPDZ256rmk,        0 },
1970     { X86::VMULPSZ256rrk,      X86::VMULPSZ256rmk,        0 },
1971     { X86::VMULPDZ256rrk,      X86::VMULPDZ256rmk,        0 },
1972     { X86::VDIVPSZ256rrk,      X86::VDIVPSZ256rmk,        0 },
1973     { X86::VDIVPDZ256rrk,      X86::VDIVPDZ256rmk,        0 },
1974     { X86::VMINPSZ256rrk,      X86::VMINPSZ256rmk,        0 },
1975     { X86::VMINPDZ256rrk,      X86::VMINPDZ256rmk,        0 },
1976     { X86::VMAXPSZ256rrk,      X86::VMAXPSZ256rmk,        0 },
1977     { X86::VMAXPDZ256rrk,      X86::VMAXPDZ256rmk,        0 },
1978     // AVX-512{F,VL} foldable instructions 128-bit
1979     { X86::VADDPSZ128rrk,      X86::VADDPSZ128rmk,        0 },
1980     { X86::VADDPDZ128rrk,      X86::VADDPDZ128rmk,        0 },
1981     { X86::VSUBPSZ128rrk,      X86::VSUBPSZ128rmk,        0 },
1982     { X86::VSUBPDZ128rrk,      X86::VSUBPDZ128rmk,        0 },
1983     { X86::VMULPSZ128rrk,      X86::VMULPSZ128rmk,        0 },
1984     { X86::VMULPDZ128rrk,      X86::VMULPDZ128rmk,        0 },
1985     { X86::VDIVPSZ128rrk,      X86::VDIVPSZ128rmk,        0 },
1986     { X86::VDIVPDZ128rrk,      X86::VDIVPDZ128rmk,        0 },
1987     { X86::VMINPSZ128rrk,      X86::VMINPSZ128rmk,        0 },
1988     { X86::VMINPDZ128rrk,      X86::VMINPDZ128rmk,        0 },
1989     { X86::VMAXPSZ128rrk,      X86::VMAXPSZ128rmk,        0 },
1990     { X86::VMAXPDZ128rrk,      X86::VMAXPDZ128rmk,        0 }
1991   };
1992 
1993   for (X86MemoryFoldTableEntry Entry : MemoryFoldTable4) {
1994     AddTableEntry(RegOp2MemOpTable4, MemOp2RegOpTable,
1995                   Entry.RegOp, Entry.MemOp,
1996                   // Index 4, folded load
1997                   Entry.Flags | TB_INDEX_4 | TB_FOLDED_LOAD);
1998   }
1999 }
2000 
2001 void
2002 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
2003                             MemOp2RegOpTableType &M2RTable,
2004                             unsigned RegOp, unsigned MemOp, unsigned Flags) {
2005     if ((Flags & TB_NO_FORWARD) == 0) {
2006       assert(!R2MTable.count(RegOp) && "Duplicate entry!");
2007       R2MTable[RegOp] = std::make_pair(MemOp, Flags);
2008     }
2009     if ((Flags & TB_NO_REVERSE) == 0) {
2010       assert(!M2RTable.count(MemOp) &&
2011            "Duplicated entries in unfolding maps?");
2012       M2RTable[MemOp] = std::make_pair(RegOp, Flags);
2013     }
2014 }
2015 
2016 bool
2017 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
2018                                     unsigned &SrcReg, unsigned &DstReg,
2019                                     unsigned &SubIdx) const {
2020   switch (MI.getOpcode()) {
2021   default: break;
2022   case X86::MOVSX16rr8:
2023   case X86::MOVZX16rr8:
2024   case X86::MOVSX32rr8:
2025   case X86::MOVZX32rr8:
2026   case X86::MOVSX64rr8:
2027     if (!Subtarget.is64Bit())
2028       // It's not always legal to reference the low 8-bit of the larger
2029       // register in 32-bit mode.
2030       return false;
2031   case X86::MOVSX32rr16:
2032   case X86::MOVZX32rr16:
2033   case X86::MOVSX64rr16:
2034   case X86::MOVSX64rr32: {
2035     if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
2036       // Be conservative.
2037       return false;
2038     SrcReg = MI.getOperand(1).getReg();
2039     DstReg = MI.getOperand(0).getReg();
2040     switch (MI.getOpcode()) {
2041     default: llvm_unreachable("Unreachable!");
2042     case X86::MOVSX16rr8:
2043     case X86::MOVZX16rr8:
2044     case X86::MOVSX32rr8:
2045     case X86::MOVZX32rr8:
2046     case X86::MOVSX64rr8:
2047       SubIdx = X86::sub_8bit;
2048       break;
2049     case X86::MOVSX32rr16:
2050     case X86::MOVZX32rr16:
2051     case X86::MOVSX64rr16:
2052       SubIdx = X86::sub_16bit;
2053       break;
2054     case X86::MOVSX64rr32:
2055       SubIdx = X86::sub_32bit;
2056       break;
2057     }
2058     return true;
2059   }
2060   }
2061   return false;
2062 }
2063 
2064 int X86InstrInfo::getSPAdjust(const MachineInstr *MI) const {
2065   const MachineFunction *MF = MI->getParent()->getParent();
2066   const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
2067 
2068   if (MI->getOpcode() == getCallFrameSetupOpcode() ||
2069       MI->getOpcode() == getCallFrameDestroyOpcode()) {
2070     unsigned StackAlign = TFI->getStackAlignment();
2071     int SPAdj = (MI->getOperand(0).getImm() + StackAlign - 1) / StackAlign *
2072                  StackAlign;
2073 
2074     SPAdj -= MI->getOperand(1).getImm();
2075 
2076     if (MI->getOpcode() == getCallFrameSetupOpcode())
2077       return SPAdj;
2078     else
2079       return -SPAdj;
2080   }
2081 
2082   // To know whether a call adjusts the stack, we need information
2083   // that is bound to the following ADJCALLSTACKUP pseudo.
2084   // Look for the next ADJCALLSTACKUP that follows the call.
2085   if (MI->isCall()) {
2086     const MachineBasicBlock* MBB = MI->getParent();
2087     auto I = ++MachineBasicBlock::const_iterator(MI);
2088     for (auto E = MBB->end(); I != E; ++I) {
2089       if (I->getOpcode() == getCallFrameDestroyOpcode() ||
2090           I->isCall())
2091         break;
2092     }
2093 
2094     // If we could not find a frame destroy opcode, then it has already
2095     // been simplified, so we don't care.
2096     if (I->getOpcode() != getCallFrameDestroyOpcode())
2097       return 0;
2098 
2099     return -(I->getOperand(1).getImm());
2100   }
2101 
2102   // Currently handle only PUSHes we can reasonably expect to see
2103   // in call sequences
2104   switch (MI->getOpcode()) {
2105   default:
2106     return 0;
2107   case X86::PUSH32i8:
2108   case X86::PUSH32r:
2109   case X86::PUSH32rmm:
2110   case X86::PUSH32rmr:
2111   case X86::PUSHi32:
2112     return 4;
2113   }
2114 }
2115 
2116 /// Return true and the FrameIndex if the specified
2117 /// operand and follow operands form a reference to the stack frame.
2118 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
2119                                   int &FrameIndex) const {
2120   if (MI->getOperand(Op+X86::AddrBaseReg).isFI() &&
2121       MI->getOperand(Op+X86::AddrScaleAmt).isImm() &&
2122       MI->getOperand(Op+X86::AddrIndexReg).isReg() &&
2123       MI->getOperand(Op+X86::AddrDisp).isImm() &&
2124       MI->getOperand(Op+X86::AddrScaleAmt).getImm() == 1 &&
2125       MI->getOperand(Op+X86::AddrIndexReg).getReg() == 0 &&
2126       MI->getOperand(Op+X86::AddrDisp).getImm() == 0) {
2127     FrameIndex = MI->getOperand(Op+X86::AddrBaseReg).getIndex();
2128     return true;
2129   }
2130   return false;
2131 }
2132 
2133 static bool isFrameLoadOpcode(int Opcode) {
2134   switch (Opcode) {
2135   default:
2136     return false;
2137   case X86::MOV8rm:
2138   case X86::MOV16rm:
2139   case X86::MOV32rm:
2140   case X86::MOV64rm:
2141   case X86::LD_Fp64m:
2142   case X86::MOVSSrm:
2143   case X86::MOVSDrm:
2144   case X86::MOVAPSrm:
2145   case X86::MOVAPDrm:
2146   case X86::MOVDQArm:
2147   case X86::VMOVSSrm:
2148   case X86::VMOVSDrm:
2149   case X86::VMOVAPSrm:
2150   case X86::VMOVAPDrm:
2151   case X86::VMOVDQArm:
2152   case X86::VMOVUPSYrm:
2153   case X86::VMOVAPSYrm:
2154   case X86::VMOVUPDYrm:
2155   case X86::VMOVAPDYrm:
2156   case X86::VMOVDQUYrm:
2157   case X86::VMOVDQAYrm:
2158   case X86::MMX_MOVD64rm:
2159   case X86::MMX_MOVQ64rm:
2160   case X86::VMOVAPSZrm:
2161   case X86::VMOVUPSZrm:
2162     return true;
2163   }
2164 }
2165 
2166 static bool isFrameStoreOpcode(int Opcode) {
2167   switch (Opcode) {
2168   default: break;
2169   case X86::MOV8mr:
2170   case X86::MOV16mr:
2171   case X86::MOV32mr:
2172   case X86::MOV64mr:
2173   case X86::ST_FpP64m:
2174   case X86::MOVSSmr:
2175   case X86::MOVSDmr:
2176   case X86::MOVAPSmr:
2177   case X86::MOVAPDmr:
2178   case X86::MOVDQAmr:
2179   case X86::VMOVSSmr:
2180   case X86::VMOVSDmr:
2181   case X86::VMOVAPSmr:
2182   case X86::VMOVAPDmr:
2183   case X86::VMOVDQAmr:
2184   case X86::VMOVUPSYmr:
2185   case X86::VMOVAPSYmr:
2186   case X86::VMOVUPDYmr:
2187   case X86::VMOVAPDYmr:
2188   case X86::VMOVDQUYmr:
2189   case X86::VMOVDQAYmr:
2190   case X86::VMOVUPSZmr:
2191   case X86::VMOVAPSZmr:
2192   case X86::MMX_MOVD64mr:
2193   case X86::MMX_MOVQ64mr:
2194   case X86::MMX_MOVNTQmr:
2195     return true;
2196   }
2197   return false;
2198 }
2199 
2200 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
2201                                            int &FrameIndex) const {
2202   if (isFrameLoadOpcode(MI->getOpcode()))
2203     if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
2204       return MI->getOperand(0).getReg();
2205   return 0;
2206 }
2207 
2208 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
2209                                                  int &FrameIndex) const {
2210   if (isFrameLoadOpcode(MI->getOpcode())) {
2211     unsigned Reg;
2212     if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
2213       return Reg;
2214     // Check for post-frame index elimination operations
2215     const MachineMemOperand *Dummy;
2216     return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
2217   }
2218   return 0;
2219 }
2220 
2221 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
2222                                           int &FrameIndex) const {
2223   if (isFrameStoreOpcode(MI->getOpcode()))
2224     if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
2225         isFrameOperand(MI, 0, FrameIndex))
2226       return MI->getOperand(X86::AddrNumOperands).getReg();
2227   return 0;
2228 }
2229 
2230 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
2231                                                 int &FrameIndex) const {
2232   if (isFrameStoreOpcode(MI->getOpcode())) {
2233     unsigned Reg;
2234     if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
2235       return Reg;
2236     // Check for post-frame index elimination operations
2237     const MachineMemOperand *Dummy;
2238     return hasStoreToStackSlot(MI, Dummy, FrameIndex);
2239   }
2240   return 0;
2241 }
2242 
2243 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r.
2244 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
2245   // Don't waste compile time scanning use-def chains of physregs.
2246   if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
2247     return false;
2248   bool isPICBase = false;
2249   for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg),
2250          E = MRI.def_instr_end(); I != E; ++I) {
2251     MachineInstr *DefMI = &*I;
2252     if (DefMI->getOpcode() != X86::MOVPC32r)
2253       return false;
2254     assert(!isPICBase && "More than one PIC base?");
2255     isPICBase = true;
2256   }
2257   return isPICBase;
2258 }
2259 
2260 bool
2261 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
2262                                                 AliasAnalysis *AA) const {
2263   switch (MI->getOpcode()) {
2264   default: break;
2265   case X86::MOV8rm:
2266   case X86::MOV16rm:
2267   case X86::MOV32rm:
2268   case X86::MOV64rm:
2269   case X86::LD_Fp64m:
2270   case X86::MOVSSrm:
2271   case X86::MOVSDrm:
2272   case X86::MOVAPSrm:
2273   case X86::MOVUPSrm:
2274   case X86::MOVAPDrm:
2275   case X86::MOVDQArm:
2276   case X86::MOVDQUrm:
2277   case X86::VMOVSSrm:
2278   case X86::VMOVSDrm:
2279   case X86::VMOVAPSrm:
2280   case X86::VMOVUPSrm:
2281   case X86::VMOVAPDrm:
2282   case X86::VMOVDQArm:
2283   case X86::VMOVDQUrm:
2284   case X86::VMOVAPSYrm:
2285   case X86::VMOVUPSYrm:
2286   case X86::VMOVAPDYrm:
2287   case X86::VMOVDQAYrm:
2288   case X86::VMOVDQUYrm:
2289   case X86::MMX_MOVD64rm:
2290   case X86::MMX_MOVQ64rm:
2291   case X86::FsVMOVAPSrm:
2292   case X86::FsVMOVAPDrm:
2293   case X86::FsMOVAPSrm:
2294   case X86::FsMOVAPDrm: {
2295     // Loads from constant pools are trivially rematerializable.
2296     if (MI->getOperand(1+X86::AddrBaseReg).isReg() &&
2297         MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
2298         MI->getOperand(1+X86::AddrIndexReg).isReg() &&
2299         MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
2300         MI->isInvariantLoad(AA)) {
2301       unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
2302       if (BaseReg == 0 || BaseReg == X86::RIP)
2303         return true;
2304       // Allow re-materialization of PIC load.
2305       if (!ReMatPICStubLoad && MI->getOperand(1+X86::AddrDisp).isGlobal())
2306         return false;
2307       const MachineFunction &MF = *MI->getParent()->getParent();
2308       const MachineRegisterInfo &MRI = MF.getRegInfo();
2309       return regIsPICBase(BaseReg, MRI);
2310     }
2311     return false;
2312   }
2313 
2314   case X86::LEA32r:
2315   case X86::LEA64r: {
2316     if (MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
2317         MI->getOperand(1+X86::AddrIndexReg).isReg() &&
2318         MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
2319         !MI->getOperand(1+X86::AddrDisp).isReg()) {
2320       // lea fi#, lea GV, etc. are all rematerializable.
2321       if (!MI->getOperand(1+X86::AddrBaseReg).isReg())
2322         return true;
2323       unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
2324       if (BaseReg == 0)
2325         return true;
2326       // Allow re-materialization of lea PICBase + x.
2327       const MachineFunction &MF = *MI->getParent()->getParent();
2328       const MachineRegisterInfo &MRI = MF.getRegInfo();
2329       return regIsPICBase(BaseReg, MRI);
2330     }
2331     return false;
2332   }
2333   }
2334 
2335   // All other instructions marked M_REMATERIALIZABLE are always trivially
2336   // rematerializable.
2337   return true;
2338 }
2339 
2340 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
2341                                          MachineBasicBlock::iterator I) const {
2342   MachineBasicBlock::iterator E = MBB.end();
2343 
2344   // For compile time consideration, if we are not able to determine the
2345   // safety after visiting 4 instructions in each direction, we will assume
2346   // it's not safe.
2347   MachineBasicBlock::iterator Iter = I;
2348   for (unsigned i = 0; Iter != E && i < 4; ++i) {
2349     bool SeenDef = false;
2350     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
2351       MachineOperand &MO = Iter->getOperand(j);
2352       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2353         SeenDef = true;
2354       if (!MO.isReg())
2355         continue;
2356       if (MO.getReg() == X86::EFLAGS) {
2357         if (MO.isUse())
2358           return false;
2359         SeenDef = true;
2360       }
2361     }
2362 
2363     if (SeenDef)
2364       // This instruction defines EFLAGS, no need to look any further.
2365       return true;
2366     ++Iter;
2367     // Skip over DBG_VALUE.
2368     while (Iter != E && Iter->isDebugValue())
2369       ++Iter;
2370   }
2371 
2372   // It is safe to clobber EFLAGS at the end of a block of no successor has it
2373   // live in.
2374   if (Iter == E) {
2375     for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
2376            SE = MBB.succ_end(); SI != SE; ++SI)
2377       if ((*SI)->isLiveIn(X86::EFLAGS))
2378         return false;
2379     return true;
2380   }
2381 
2382   MachineBasicBlock::iterator B = MBB.begin();
2383   Iter = I;
2384   for (unsigned i = 0; i < 4; ++i) {
2385     // If we make it to the beginning of the block, it's safe to clobber
2386     // EFLAGS iff EFLAGS is not live-in.
2387     if (Iter == B)
2388       return !MBB.isLiveIn(X86::EFLAGS);
2389 
2390     --Iter;
2391     // Skip over DBG_VALUE.
2392     while (Iter != B && Iter->isDebugValue())
2393       --Iter;
2394 
2395     bool SawKill = false;
2396     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
2397       MachineOperand &MO = Iter->getOperand(j);
2398       // A register mask may clobber EFLAGS, but we should still look for a
2399       // live EFLAGS def.
2400       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
2401         SawKill = true;
2402       if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
2403         if (MO.isDef()) return MO.isDead();
2404         if (MO.isKill()) SawKill = true;
2405       }
2406     }
2407 
2408     if (SawKill)
2409       // This instruction kills EFLAGS and doesn't redefine it, so
2410       // there's no need to look further.
2411       return true;
2412   }
2413 
2414   // Conservative answer.
2415   return false;
2416 }
2417 
2418 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
2419                                  MachineBasicBlock::iterator I,
2420                                  unsigned DestReg, unsigned SubIdx,
2421                                  const MachineInstr *Orig,
2422                                  const TargetRegisterInfo &TRI) const {
2423   // MOV32r0 is implemented with a xor which clobbers condition code.
2424   // Re-materialize it as movri instructions to avoid side effects.
2425   unsigned Opc = Orig->getOpcode();
2426   if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
2427     DebugLoc DL = Orig->getDebugLoc();
2428     BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
2429       .addImm(0);
2430   } else {
2431     MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
2432     MBB.insert(I, MI);
2433   }
2434 
2435   MachineInstr *NewMI = std::prev(I);
2436   NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
2437 }
2438 
2439 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead.
2440 static bool hasLiveCondCodeDef(MachineInstr *MI) {
2441   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2442     MachineOperand &MO = MI->getOperand(i);
2443     if (MO.isReg() && MO.isDef() &&
2444         MO.getReg() == X86::EFLAGS && !MO.isDead()) {
2445       return true;
2446     }
2447   }
2448   return false;
2449 }
2450 
2451 /// Check whether the shift count for a machine operand is non-zero.
2452 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
2453                                               unsigned ShiftAmtOperandIdx) {
2454   // The shift count is six bits with the REX.W prefix and five bits without.
2455   unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
2456   unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
2457   return Imm & ShiftCountMask;
2458 }
2459 
2460 /// Check whether the given shift count is appropriate
2461 /// can be represented by a LEA instruction.
2462 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
2463   // Left shift instructions can be transformed into load-effective-address
2464   // instructions if we can encode them appropriately.
2465   // A LEA instruction utilizes a SIB byte to encode it's scale factor.
2466   // The SIB.scale field is two bits wide which means that we can encode any
2467   // shift amount less than 4.
2468   return ShAmt < 4 && ShAmt > 0;
2469 }
2470 
2471 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
2472                                   unsigned Opc, bool AllowSP,
2473                                   unsigned &NewSrc, bool &isKill, bool &isUndef,
2474                                   MachineOperand &ImplicitOp) const {
2475   MachineFunction &MF = *MI->getParent()->getParent();
2476   const TargetRegisterClass *RC;
2477   if (AllowSP) {
2478     RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
2479   } else {
2480     RC = Opc != X86::LEA32r ?
2481       &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
2482   }
2483   unsigned SrcReg = Src.getReg();
2484 
2485   // For both LEA64 and LEA32 the register already has essentially the right
2486   // type (32-bit or 64-bit) we may just need to forbid SP.
2487   if (Opc != X86::LEA64_32r) {
2488     NewSrc = SrcReg;
2489     isKill = Src.isKill();
2490     isUndef = Src.isUndef();
2491 
2492     if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
2493         !MF.getRegInfo().constrainRegClass(NewSrc, RC))
2494       return false;
2495 
2496     return true;
2497   }
2498 
2499   // This is for an LEA64_32r and incoming registers are 32-bit. One way or
2500   // another we need to add 64-bit registers to the final MI.
2501   if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
2502     ImplicitOp = Src;
2503     ImplicitOp.setImplicit();
2504 
2505     NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
2506     MachineBasicBlock::LivenessQueryResult LQR =
2507       MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
2508 
2509     switch (LQR) {
2510     case MachineBasicBlock::LQR_Unknown:
2511       // We can't give sane liveness flags to the instruction, abandon LEA
2512       // formation.
2513       return false;
2514     case MachineBasicBlock::LQR_Live:
2515       isKill = MI->killsRegister(SrcReg);
2516       isUndef = false;
2517       break;
2518     default:
2519       // The physreg itself is dead, so we have to use it as an <undef>.
2520       isKill = false;
2521       isUndef = true;
2522       break;
2523     }
2524   } else {
2525     // Virtual register of the wrong class, we have to create a temporary 64-bit
2526     // vreg to feed into the LEA.
2527     NewSrc = MF.getRegInfo().createVirtualRegister(RC);
2528     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
2529             get(TargetOpcode::COPY))
2530       .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
2531         .addOperand(Src);
2532 
2533     // Which is obviously going to be dead after we're done with it.
2534     isKill = true;
2535     isUndef = false;
2536   }
2537 
2538   // We've set all the parameters without issue.
2539   return true;
2540 }
2541 
2542 /// Helper for convertToThreeAddress when 16-bit LEA is disabled, use 32-bit
2543 /// LEA to form 3-address code by promoting to a 32-bit superregister and then
2544 /// truncating back down to a 16-bit subregister.
2545 MachineInstr *
2546 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
2547                                            MachineFunction::iterator &MFI,
2548                                            MachineBasicBlock::iterator &MBBI,
2549                                            LiveVariables *LV) const {
2550   MachineInstr *MI = MBBI;
2551   unsigned Dest = MI->getOperand(0).getReg();
2552   unsigned Src = MI->getOperand(1).getReg();
2553   bool isDead = MI->getOperand(0).isDead();
2554   bool isKill = MI->getOperand(1).isKill();
2555 
2556   MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
2557   unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
2558   unsigned Opc, leaInReg;
2559   if (Subtarget.is64Bit()) {
2560     Opc = X86::LEA64_32r;
2561     leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2562   } else {
2563     Opc = X86::LEA32r;
2564     leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2565   }
2566 
2567   // Build and insert into an implicit UNDEF value. This is OK because
2568   // well be shifting and then extracting the lower 16-bits.
2569   // This has the potential to cause partial register stall. e.g.
2570   //   movw    (%rbp,%rcx,2), %dx
2571   //   leal    -65(%rdx), %esi
2572   // But testing has shown this *does* help performance in 64-bit mode (at
2573   // least on modern x86 machines).
2574   BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
2575   MachineInstr *InsMI =
2576     BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2577     .addReg(leaInReg, RegState::Define, X86::sub_16bit)
2578     .addReg(Src, getKillRegState(isKill));
2579 
2580   MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
2581                                     get(Opc), leaOutReg);
2582   switch (MIOpc) {
2583   default: llvm_unreachable("Unreachable!");
2584   case X86::SHL16ri: {
2585     unsigned ShAmt = MI->getOperand(2).getImm();
2586     MIB.addReg(0).addImm(1 << ShAmt)
2587        .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
2588     break;
2589   }
2590   case X86::INC16r:
2591     addRegOffset(MIB, leaInReg, true, 1);
2592     break;
2593   case X86::DEC16r:
2594     addRegOffset(MIB, leaInReg, true, -1);
2595     break;
2596   case X86::ADD16ri:
2597   case X86::ADD16ri8:
2598   case X86::ADD16ri_DB:
2599   case X86::ADD16ri8_DB:
2600     addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
2601     break;
2602   case X86::ADD16rr:
2603   case X86::ADD16rr_DB: {
2604     unsigned Src2 = MI->getOperand(2).getReg();
2605     bool isKill2 = MI->getOperand(2).isKill();
2606     unsigned leaInReg2 = 0;
2607     MachineInstr *InsMI2 = nullptr;
2608     if (Src == Src2) {
2609       // ADD16rr %reg1028<kill>, %reg1028
2610       // just a single insert_subreg.
2611       addRegReg(MIB, leaInReg, true, leaInReg, false);
2612     } else {
2613       if (Subtarget.is64Bit())
2614         leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
2615       else
2616         leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
2617       // Build and insert into an implicit UNDEF value. This is OK because
2618       // well be shifting and then extracting the lower 16-bits.
2619       BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2620       InsMI2 =
2621         BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2622         .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2623         .addReg(Src2, getKillRegState(isKill2));
2624       addRegReg(MIB, leaInReg, true, leaInReg2, true);
2625     }
2626     if (LV && isKill2 && InsMI2)
2627       LV->replaceKillInstruction(Src2, MI, InsMI2);
2628     break;
2629   }
2630   }
2631 
2632   MachineInstr *NewMI = MIB;
2633   MachineInstr *ExtMI =
2634     BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2635     .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2636     .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2637 
2638   if (LV) {
2639     // Update live variables
2640     LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2641     LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2642     if (isKill)
2643       LV->replaceKillInstruction(Src, MI, InsMI);
2644     if (isDead)
2645       LV->replaceKillInstruction(Dest, MI, ExtMI);
2646   }
2647 
2648   return ExtMI;
2649 }
2650 
2651 /// This method must be implemented by targets that
2652 /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
2653 /// may be able to convert a two-address instruction into a true
2654 /// three-address instruction on demand.  This allows the X86 target (for
2655 /// example) to convert ADD and SHL instructions into LEA instructions if they
2656 /// would require register copies due to two-addressness.
2657 ///
2658 /// This method returns a null pointer if the transformation cannot be
2659 /// performed, otherwise it returns the new instruction.
2660 ///
2661 MachineInstr *
2662 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2663                                     MachineBasicBlock::iterator &MBBI,
2664                                     LiveVariables *LV) const {
2665   MachineInstr *MI = MBBI;
2666 
2667   // The following opcodes also sets the condition code register(s). Only
2668   // convert them to equivalent lea if the condition code register def's
2669   // are dead!
2670   if (hasLiveCondCodeDef(MI))
2671     return nullptr;
2672 
2673   MachineFunction &MF = *MI->getParent()->getParent();
2674   // All instructions input are two-addr instructions.  Get the known operands.
2675   const MachineOperand &Dest = MI->getOperand(0);
2676   const MachineOperand &Src = MI->getOperand(1);
2677 
2678   MachineInstr *NewMI = nullptr;
2679   // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's.  When
2680   // we have better subtarget support, enable the 16-bit LEA generation here.
2681   // 16-bit LEA is also slow on Core2.
2682   bool DisableLEA16 = true;
2683   bool is64Bit = Subtarget.is64Bit();
2684 
2685   unsigned MIOpc = MI->getOpcode();
2686   switch (MIOpc) {
2687   default: return nullptr;
2688   case X86::SHL64ri: {
2689     assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2690     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2691     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2692 
2693     // LEA can't handle RSP.
2694     if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2695         !MF.getRegInfo().constrainRegClass(Src.getReg(),
2696                                            &X86::GR64_NOSPRegClass))
2697       return nullptr;
2698 
2699     NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2700       .addOperand(Dest)
2701       .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2702     break;
2703   }
2704   case X86::SHL32ri: {
2705     assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2706     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2707     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2708 
2709     unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2710 
2711     // LEA can't handle ESP.
2712     bool isKill, isUndef;
2713     unsigned SrcReg;
2714     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2715     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2716                         SrcReg, isKill, isUndef, ImplicitOp))
2717       return nullptr;
2718 
2719     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2720       .addOperand(Dest)
2721       .addReg(0).addImm(1 << ShAmt)
2722       .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2723       .addImm(0).addReg(0);
2724     if (ImplicitOp.getReg() != 0)
2725       MIB.addOperand(ImplicitOp);
2726     NewMI = MIB;
2727 
2728     break;
2729   }
2730   case X86::SHL16ri: {
2731     assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2732     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2733     if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr;
2734 
2735     if (DisableLEA16)
2736       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : nullptr;
2737     NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2738       .addOperand(Dest)
2739       .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2740     break;
2741   }
2742   case X86::INC64r:
2743   case X86::INC32r: {
2744     assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2745     unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2746       : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2747     bool isKill, isUndef;
2748     unsigned SrcReg;
2749     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2750     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2751                         SrcReg, isKill, isUndef, ImplicitOp))
2752       return nullptr;
2753 
2754     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2755         .addOperand(Dest)
2756         .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2757     if (ImplicitOp.getReg() != 0)
2758       MIB.addOperand(ImplicitOp);
2759 
2760     NewMI = addOffset(MIB, 1);
2761     break;
2762   }
2763   case X86::INC16r:
2764     if (DisableLEA16)
2765       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2766                      : nullptr;
2767     assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2768     NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2769                       .addOperand(Dest).addOperand(Src), 1);
2770     break;
2771   case X86::DEC64r:
2772   case X86::DEC32r: {
2773     assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2774     unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2775       : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2776 
2777     bool isKill, isUndef;
2778     unsigned SrcReg;
2779     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2780     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2781                         SrcReg, isKill, isUndef, ImplicitOp))
2782       return nullptr;
2783 
2784     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2785         .addOperand(Dest)
2786         .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2787     if (ImplicitOp.getReg() != 0)
2788       MIB.addOperand(ImplicitOp);
2789 
2790     NewMI = addOffset(MIB, -1);
2791 
2792     break;
2793   }
2794   case X86::DEC16r:
2795     if (DisableLEA16)
2796       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2797                      : nullptr;
2798     assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2799     NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2800                       .addOperand(Dest).addOperand(Src), -1);
2801     break;
2802   case X86::ADD64rr:
2803   case X86::ADD64rr_DB:
2804   case X86::ADD32rr:
2805   case X86::ADD32rr_DB: {
2806     assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2807     unsigned Opc;
2808     if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2809       Opc = X86::LEA64r;
2810     else
2811       Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2812 
2813     bool isKill, isUndef;
2814     unsigned SrcReg;
2815     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2816     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2817                         SrcReg, isKill, isUndef, ImplicitOp))
2818       return nullptr;
2819 
2820     const MachineOperand &Src2 = MI->getOperand(2);
2821     bool isKill2, isUndef2;
2822     unsigned SrcReg2;
2823     MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2824     if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2825                         SrcReg2, isKill2, isUndef2, ImplicitOp2))
2826       return nullptr;
2827 
2828     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2829       .addOperand(Dest);
2830     if (ImplicitOp.getReg() != 0)
2831       MIB.addOperand(ImplicitOp);
2832     if (ImplicitOp2.getReg() != 0)
2833       MIB.addOperand(ImplicitOp2);
2834 
2835     NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2836 
2837     // Preserve undefness of the operands.
2838     NewMI->getOperand(1).setIsUndef(isUndef);
2839     NewMI->getOperand(3).setIsUndef(isUndef2);
2840 
2841     if (LV && Src2.isKill())
2842       LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2843     break;
2844   }
2845   case X86::ADD16rr:
2846   case X86::ADD16rr_DB: {
2847     if (DisableLEA16)
2848       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2849                      : nullptr;
2850     assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2851     unsigned Src2 = MI->getOperand(2).getReg();
2852     bool isKill2 = MI->getOperand(2).isKill();
2853     NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2854                       .addOperand(Dest),
2855                       Src.getReg(), Src.isKill(), Src2, isKill2);
2856 
2857     // Preserve undefness of the operands.
2858     bool isUndef = MI->getOperand(1).isUndef();
2859     bool isUndef2 = MI->getOperand(2).isUndef();
2860     NewMI->getOperand(1).setIsUndef(isUndef);
2861     NewMI->getOperand(3).setIsUndef(isUndef2);
2862 
2863     if (LV && isKill2)
2864       LV->replaceKillInstruction(Src2, MI, NewMI);
2865     break;
2866   }
2867   case X86::ADD64ri32:
2868   case X86::ADD64ri8:
2869   case X86::ADD64ri32_DB:
2870   case X86::ADD64ri8_DB:
2871     assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2872     NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2873                       .addOperand(Dest).addOperand(Src),
2874                       MI->getOperand(2).getImm());
2875     break;
2876   case X86::ADD32ri:
2877   case X86::ADD32ri8:
2878   case X86::ADD32ri_DB:
2879   case X86::ADD32ri8_DB: {
2880     assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2881     unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2882 
2883     bool isKill, isUndef;
2884     unsigned SrcReg;
2885     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2886     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2887                         SrcReg, isKill, isUndef, ImplicitOp))
2888       return nullptr;
2889 
2890     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2891         .addOperand(Dest)
2892         .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2893     if (ImplicitOp.getReg() != 0)
2894       MIB.addOperand(ImplicitOp);
2895 
2896     NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2897     break;
2898   }
2899   case X86::ADD16ri:
2900   case X86::ADD16ri8:
2901   case X86::ADD16ri_DB:
2902   case X86::ADD16ri8_DB:
2903     if (DisableLEA16)
2904       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV)
2905                      : nullptr;
2906     assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2907     NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2908                       .addOperand(Dest).addOperand(Src),
2909                       MI->getOperand(2).getImm());
2910     break;
2911   }
2912 
2913   if (!NewMI) return nullptr;
2914 
2915   if (LV) {  // Update live variables
2916     if (Src.isKill())
2917       LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2918     if (Dest.isDead())
2919       LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2920   }
2921 
2922   MFI->insert(MBBI, NewMI);          // Insert the new inst
2923   return NewMI;
2924 }
2925 
2926 /// We have a few instructions that must be hacked on to commute them.
2927 ///
2928 MachineInstr *
2929 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2930   switch (MI->getOpcode()) {
2931   case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2932   case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2933   case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2934   case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2935   case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2936   case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2937     unsigned Opc;
2938     unsigned Size;
2939     switch (MI->getOpcode()) {
2940     default: llvm_unreachable("Unreachable!");
2941     case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2942     case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2943     case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2944     case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2945     case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2946     case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2947     }
2948     unsigned Amt = MI->getOperand(3).getImm();
2949     if (NewMI) {
2950       MachineFunction &MF = *MI->getParent()->getParent();
2951       MI = MF.CloneMachineInstr(MI);
2952       NewMI = false;
2953     }
2954     MI->setDesc(get(Opc));
2955     MI->getOperand(3).setImm(Size-Amt);
2956     return TargetInstrInfo::commuteInstruction(MI, NewMI);
2957   }
2958   case X86::BLENDPDrri:
2959   case X86::BLENDPSrri:
2960   case X86::PBLENDWrri:
2961   case X86::VBLENDPDrri:
2962   case X86::VBLENDPSrri:
2963   case X86::VBLENDPDYrri:
2964   case X86::VBLENDPSYrri:
2965   case X86::VPBLENDDrri:
2966   case X86::VPBLENDWrri:
2967   case X86::VPBLENDDYrri:
2968   case X86::VPBLENDWYrri:{
2969     unsigned Mask;
2970     switch (MI->getOpcode()) {
2971     default: llvm_unreachable("Unreachable!");
2972     case X86::BLENDPDrri:    Mask = 0x03; break;
2973     case X86::BLENDPSrri:    Mask = 0x0F; break;
2974     case X86::PBLENDWrri:    Mask = 0xFF; break;
2975     case X86::VBLENDPDrri:   Mask = 0x03; break;
2976     case X86::VBLENDPSrri:   Mask = 0x0F; break;
2977     case X86::VBLENDPDYrri:  Mask = 0x0F; break;
2978     case X86::VBLENDPSYrri:  Mask = 0xFF; break;
2979     case X86::VPBLENDDrri:   Mask = 0x0F; break;
2980     case X86::VPBLENDWrri:   Mask = 0xFF; break;
2981     case X86::VPBLENDDYrri:  Mask = 0xFF; break;
2982     case X86::VPBLENDWYrri:  Mask = 0xFF; break;
2983     }
2984     // Only the least significant bits of Imm are used.
2985     unsigned Imm = MI->getOperand(3).getImm() & Mask;
2986     if (NewMI) {
2987       MachineFunction &MF = *MI->getParent()->getParent();
2988       MI = MF.CloneMachineInstr(MI);
2989       NewMI = false;
2990     }
2991     MI->getOperand(3).setImm(Mask ^ Imm);
2992     return TargetInstrInfo::commuteInstruction(MI, NewMI);
2993   }
2994   case X86::PCLMULQDQrr:
2995   case X86::VPCLMULQDQrr:{
2996     // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0]
2997     // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0]
2998     unsigned Imm = MI->getOperand(3).getImm();
2999     unsigned Src1Hi = Imm & 0x01;
3000     unsigned Src2Hi = Imm & 0x10;
3001     if (NewMI) {
3002       MachineFunction &MF = *MI->getParent()->getParent();
3003       MI = MF.CloneMachineInstr(MI);
3004       NewMI = false;
3005     }
3006     MI->getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
3007     return TargetInstrInfo::commuteInstruction(MI, NewMI);
3008   }
3009   case X86::CMPPDrri:
3010   case X86::CMPPSrri:
3011   case X86::VCMPPDrri:
3012   case X86::VCMPPSrri:
3013   case X86::VCMPPDYrri:
3014   case X86::VCMPPSYrri: {
3015     // Float comparison can be safely commuted for
3016     // Ordered/Unordered/Equal/NotEqual tests
3017     unsigned Imm = MI->getOperand(3).getImm() & 0x7;
3018     switch (Imm) {
3019     case 0x00: // EQUAL
3020     case 0x03: // UNORDERED
3021     case 0x04: // NOT EQUAL
3022     case 0x07: // ORDERED
3023       if (NewMI) {
3024         MachineFunction &MF = *MI->getParent()->getParent();
3025         MI = MF.CloneMachineInstr(MI);
3026         NewMI = false;
3027       }
3028       return TargetInstrInfo::commuteInstruction(MI, NewMI);
3029     default:
3030       return nullptr;
3031     }
3032   }
3033   case X86::VPCOMBri: case X86::VPCOMUBri:
3034   case X86::VPCOMDri: case X86::VPCOMUDri:
3035   case X86::VPCOMQri: case X86::VPCOMUQri:
3036   case X86::VPCOMWri: case X86::VPCOMUWri: {
3037     // Flip comparison mode immediate (if necessary).
3038     unsigned Imm = MI->getOperand(3).getImm() & 0x7;
3039     switch (Imm) {
3040     case 0x00: Imm = 0x02; break; // LT -> GT
3041     case 0x01: Imm = 0x03; break; // LE -> GE
3042     case 0x02: Imm = 0x00; break; // GT -> LT
3043     case 0x03: Imm = 0x01; break; // GE -> LE
3044     case 0x04: // EQ
3045     case 0x05: // NE
3046     case 0x06: // FALSE
3047     case 0x07: // TRUE
3048     default:
3049       break;
3050     }
3051     if (NewMI) {
3052       MachineFunction &MF = *MI->getParent()->getParent();
3053       MI = MF.CloneMachineInstr(MI);
3054       NewMI = false;
3055     }
3056     MI->getOperand(3).setImm(Imm);
3057     return TargetInstrInfo::commuteInstruction(MI, NewMI);
3058   }
3059   case X86::CMOVB16rr:  case X86::CMOVB32rr:  case X86::CMOVB64rr:
3060   case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
3061   case X86::CMOVE16rr:  case X86::CMOVE32rr:  case X86::CMOVE64rr:
3062   case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
3063   case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
3064   case X86::CMOVA16rr:  case X86::CMOVA32rr:  case X86::CMOVA64rr:
3065   case X86::CMOVL16rr:  case X86::CMOVL32rr:  case X86::CMOVL64rr:
3066   case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
3067   case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
3068   case X86::CMOVG16rr:  case X86::CMOVG32rr:  case X86::CMOVG64rr:
3069   case X86::CMOVS16rr:  case X86::CMOVS32rr:  case X86::CMOVS64rr:
3070   case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
3071   case X86::CMOVP16rr:  case X86::CMOVP32rr:  case X86::CMOVP64rr:
3072   case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
3073   case X86::CMOVO16rr:  case X86::CMOVO32rr:  case X86::CMOVO64rr:
3074   case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
3075     unsigned Opc;
3076     switch (MI->getOpcode()) {
3077     default: llvm_unreachable("Unreachable!");
3078     case X86::CMOVB16rr:  Opc = X86::CMOVAE16rr; break;
3079     case X86::CMOVB32rr:  Opc = X86::CMOVAE32rr; break;
3080     case X86::CMOVB64rr:  Opc = X86::CMOVAE64rr; break;
3081     case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
3082     case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
3083     case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
3084     case X86::CMOVE16rr:  Opc = X86::CMOVNE16rr; break;
3085     case X86::CMOVE32rr:  Opc = X86::CMOVNE32rr; break;
3086     case X86::CMOVE64rr:  Opc = X86::CMOVNE64rr; break;
3087     case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
3088     case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
3089     case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
3090     case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
3091     case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
3092     case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
3093     case X86::CMOVA16rr:  Opc = X86::CMOVBE16rr; break;
3094     case X86::CMOVA32rr:  Opc = X86::CMOVBE32rr; break;
3095     case X86::CMOVA64rr:  Opc = X86::CMOVBE64rr; break;
3096     case X86::CMOVL16rr:  Opc = X86::CMOVGE16rr; break;
3097     case X86::CMOVL32rr:  Opc = X86::CMOVGE32rr; break;
3098     case X86::CMOVL64rr:  Opc = X86::CMOVGE64rr; break;
3099     case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
3100     case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
3101     case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
3102     case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
3103     case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
3104     case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
3105     case X86::CMOVG16rr:  Opc = X86::CMOVLE16rr; break;
3106     case X86::CMOVG32rr:  Opc = X86::CMOVLE32rr; break;
3107     case X86::CMOVG64rr:  Opc = X86::CMOVLE64rr; break;
3108     case X86::CMOVS16rr:  Opc = X86::CMOVNS16rr; break;
3109     case X86::CMOVS32rr:  Opc = X86::CMOVNS32rr; break;
3110     case X86::CMOVS64rr:  Opc = X86::CMOVNS64rr; break;
3111     case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
3112     case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
3113     case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
3114     case X86::CMOVP16rr:  Opc = X86::CMOVNP16rr; break;
3115     case X86::CMOVP32rr:  Opc = X86::CMOVNP32rr; break;
3116     case X86::CMOVP64rr:  Opc = X86::CMOVNP64rr; break;
3117     case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
3118     case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
3119     case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
3120     case X86::CMOVO16rr:  Opc = X86::CMOVNO16rr; break;
3121     case X86::CMOVO32rr:  Opc = X86::CMOVNO32rr; break;
3122     case X86::CMOVO64rr:  Opc = X86::CMOVNO64rr; break;
3123     case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
3124     case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
3125     case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
3126     }
3127     if (NewMI) {
3128       MachineFunction &MF = *MI->getParent()->getParent();
3129       MI = MF.CloneMachineInstr(MI);
3130       NewMI = false;
3131     }
3132     MI->setDesc(get(Opc));
3133     // Fallthrough intended.
3134   }
3135   default:
3136     return TargetInstrInfo::commuteInstruction(MI, NewMI);
3137   }
3138 }
3139 
3140 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
3141                                          unsigned &SrcOpIdx2) const {
3142   switch (MI->getOpcode()) {
3143     case X86::CMPPDrri:
3144     case X86::CMPPSrri:
3145     case X86::VCMPPDrri:
3146     case X86::VCMPPSrri:
3147     case X86::VCMPPDYrri:
3148     case X86::VCMPPSYrri: {
3149       // Float comparison can be safely commuted for
3150       // Ordered/Unordered/Equal/NotEqual tests
3151       unsigned Imm = MI->getOperand(3).getImm() & 0x7;
3152       switch (Imm) {
3153       case 0x00: // EQUAL
3154       case 0x03: // UNORDERED
3155       case 0x04: // NOT EQUAL
3156       case 0x07: // ORDERED
3157         SrcOpIdx1 = 1;
3158         SrcOpIdx2 = 2;
3159         return true;
3160       }
3161       return false;
3162     }
3163     case X86::VFMADDPDr231r:
3164     case X86::VFMADDPSr231r:
3165     case X86::VFMADDSDr231r:
3166     case X86::VFMADDSSr231r:
3167     case X86::VFMSUBPDr231r:
3168     case X86::VFMSUBPSr231r:
3169     case X86::VFMSUBSDr231r:
3170     case X86::VFMSUBSSr231r:
3171     case X86::VFNMADDPDr231r:
3172     case X86::VFNMADDPSr231r:
3173     case X86::VFNMADDSDr231r:
3174     case X86::VFNMADDSSr231r:
3175     case X86::VFNMSUBPDr231r:
3176     case X86::VFNMSUBPSr231r:
3177     case X86::VFNMSUBSDr231r:
3178     case X86::VFNMSUBSSr231r:
3179     case X86::VFMADDPDr231rY:
3180     case X86::VFMADDPSr231rY:
3181     case X86::VFMSUBPDr231rY:
3182     case X86::VFMSUBPSr231rY:
3183     case X86::VFNMADDPDr231rY:
3184     case X86::VFNMADDPSr231rY:
3185     case X86::VFNMSUBPDr231rY:
3186     case X86::VFNMSUBPSr231rY:
3187       SrcOpIdx1 = 2;
3188       SrcOpIdx2 = 3;
3189       return true;
3190     default:
3191       return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
3192   }
3193 }
3194 
3195 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
3196   switch (BrOpc) {
3197   default: return X86::COND_INVALID;
3198   case X86::JE_1:  return X86::COND_E;
3199   case X86::JNE_1: return X86::COND_NE;
3200   case X86::JL_1:  return X86::COND_L;
3201   case X86::JLE_1: return X86::COND_LE;
3202   case X86::JG_1:  return X86::COND_G;
3203   case X86::JGE_1: return X86::COND_GE;
3204   case X86::JB_1:  return X86::COND_B;
3205   case X86::JBE_1: return X86::COND_BE;
3206   case X86::JA_1:  return X86::COND_A;
3207   case X86::JAE_1: return X86::COND_AE;
3208   case X86::JS_1:  return X86::COND_S;
3209   case X86::JNS_1: return X86::COND_NS;
3210   case X86::JP_1:  return X86::COND_P;
3211   case X86::JNP_1: return X86::COND_NP;
3212   case X86::JO_1:  return X86::COND_O;
3213   case X86::JNO_1: return X86::COND_NO;
3214   }
3215 }
3216 
3217 /// Return condition code of a SET opcode.
3218 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
3219   switch (Opc) {
3220   default: return X86::COND_INVALID;
3221   case X86::SETAr:  case X86::SETAm:  return X86::COND_A;
3222   case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
3223   case X86::SETBr:  case X86::SETBm:  return X86::COND_B;
3224   case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
3225   case X86::SETEr:  case X86::SETEm:  return X86::COND_E;
3226   case X86::SETGr:  case X86::SETGm:  return X86::COND_G;
3227   case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
3228   case X86::SETLr:  case X86::SETLm:  return X86::COND_L;
3229   case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
3230   case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
3231   case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
3232   case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
3233   case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
3234   case X86::SETOr:  case X86::SETOm:  return X86::COND_O;
3235   case X86::SETPr:  case X86::SETPm:  return X86::COND_P;
3236   case X86::SETSr:  case X86::SETSm:  return X86::COND_S;
3237   }
3238 }
3239 
3240 /// Return condition code of a CMov opcode.
3241 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
3242   switch (Opc) {
3243   default: return X86::COND_INVALID;
3244   case X86::CMOVA16rm:  case X86::CMOVA16rr:  case X86::CMOVA32rm:
3245   case X86::CMOVA32rr:  case X86::CMOVA64rm:  case X86::CMOVA64rr:
3246     return X86::COND_A;
3247   case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
3248   case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
3249     return X86::COND_AE;
3250   case X86::CMOVB16rm:  case X86::CMOVB16rr:  case X86::CMOVB32rm:
3251   case X86::CMOVB32rr:  case X86::CMOVB64rm:  case X86::CMOVB64rr:
3252     return X86::COND_B;
3253   case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
3254   case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
3255     return X86::COND_BE;
3256   case X86::CMOVE16rm:  case X86::CMOVE16rr:  case X86::CMOVE32rm:
3257   case X86::CMOVE32rr:  case X86::CMOVE64rm:  case X86::CMOVE64rr:
3258     return X86::COND_E;
3259   case X86::CMOVG16rm:  case X86::CMOVG16rr:  case X86::CMOVG32rm:
3260   case X86::CMOVG32rr:  case X86::CMOVG64rm:  case X86::CMOVG64rr:
3261     return X86::COND_G;
3262   case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
3263   case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
3264     return X86::COND_GE;
3265   case X86::CMOVL16rm:  case X86::CMOVL16rr:  case X86::CMOVL32rm:
3266   case X86::CMOVL32rr:  case X86::CMOVL64rm:  case X86::CMOVL64rr:
3267     return X86::COND_L;
3268   case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
3269   case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
3270     return X86::COND_LE;
3271   case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
3272   case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
3273     return X86::COND_NE;
3274   case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
3275   case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
3276     return X86::COND_NO;
3277   case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
3278   case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
3279     return X86::COND_NP;
3280   case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
3281   case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
3282     return X86::COND_NS;
3283   case X86::CMOVO16rm:  case X86::CMOVO16rr:  case X86::CMOVO32rm:
3284   case X86::CMOVO32rr:  case X86::CMOVO64rm:  case X86::CMOVO64rr:
3285     return X86::COND_O;
3286   case X86::CMOVP16rm:  case X86::CMOVP16rr:  case X86::CMOVP32rm:
3287   case X86::CMOVP32rr:  case X86::CMOVP64rm:  case X86::CMOVP64rr:
3288     return X86::COND_P;
3289   case X86::CMOVS16rm:  case X86::CMOVS16rr:  case X86::CMOVS32rm:
3290   case X86::CMOVS32rr:  case X86::CMOVS64rm:  case X86::CMOVS64rr:
3291     return X86::COND_S;
3292   }
3293 }
3294 
3295 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
3296   switch (CC) {
3297   default: llvm_unreachable("Illegal condition code!");
3298   case X86::COND_E:  return X86::JE_1;
3299   case X86::COND_NE: return X86::JNE_1;
3300   case X86::COND_L:  return X86::JL_1;
3301   case X86::COND_LE: return X86::JLE_1;
3302   case X86::COND_G:  return X86::JG_1;
3303   case X86::COND_GE: return X86::JGE_1;
3304   case X86::COND_B:  return X86::JB_1;
3305   case X86::COND_BE: return X86::JBE_1;
3306   case X86::COND_A:  return X86::JA_1;
3307   case X86::COND_AE: return X86::JAE_1;
3308   case X86::COND_S:  return X86::JS_1;
3309   case X86::COND_NS: return X86::JNS_1;
3310   case X86::COND_P:  return X86::JP_1;
3311   case X86::COND_NP: return X86::JNP_1;
3312   case X86::COND_O:  return X86::JO_1;
3313   case X86::COND_NO: return X86::JNO_1;
3314   }
3315 }
3316 
3317 /// Return the inverse of the specified condition,
3318 /// e.g. turning COND_E to COND_NE.
3319 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
3320   switch (CC) {
3321   default: llvm_unreachable("Illegal condition code!");
3322   case X86::COND_E:  return X86::COND_NE;
3323   case X86::COND_NE: return X86::COND_E;
3324   case X86::COND_L:  return X86::COND_GE;
3325   case X86::COND_LE: return X86::COND_G;
3326   case X86::COND_G:  return X86::COND_LE;
3327   case X86::COND_GE: return X86::COND_L;
3328   case X86::COND_B:  return X86::COND_AE;
3329   case X86::COND_BE: return X86::COND_A;
3330   case X86::COND_A:  return X86::COND_BE;
3331   case X86::COND_AE: return X86::COND_B;
3332   case X86::COND_S:  return X86::COND_NS;
3333   case X86::COND_NS: return X86::COND_S;
3334   case X86::COND_P:  return X86::COND_NP;
3335   case X86::COND_NP: return X86::COND_P;
3336   case X86::COND_O:  return X86::COND_NO;
3337   case X86::COND_NO: return X86::COND_O;
3338   }
3339 }
3340 
3341 /// Assuming the flags are set by MI(a,b), return the condition code if we
3342 /// modify the instructions such that flags are set by MI(b,a).
3343 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
3344   switch (CC) {
3345   default: return X86::COND_INVALID;
3346   case X86::COND_E:  return X86::COND_E;
3347   case X86::COND_NE: return X86::COND_NE;
3348   case X86::COND_L:  return X86::COND_G;
3349   case X86::COND_LE: return X86::COND_GE;
3350   case X86::COND_G:  return X86::COND_L;
3351   case X86::COND_GE: return X86::COND_LE;
3352   case X86::COND_B:  return X86::COND_A;
3353   case X86::COND_BE: return X86::COND_AE;
3354   case X86::COND_A:  return X86::COND_B;
3355   case X86::COND_AE: return X86::COND_BE;
3356   }
3357 }
3358 
3359 /// Return a set opcode for the given condition and
3360 /// whether it has memory operand.
3361 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) {
3362   static const uint16_t Opc[16][2] = {
3363     { X86::SETAr,  X86::SETAm  },
3364     { X86::SETAEr, X86::SETAEm },
3365     { X86::SETBr,  X86::SETBm  },
3366     { X86::SETBEr, X86::SETBEm },
3367     { X86::SETEr,  X86::SETEm  },
3368     { X86::SETGr,  X86::SETGm  },
3369     { X86::SETGEr, X86::SETGEm },
3370     { X86::SETLr,  X86::SETLm  },
3371     { X86::SETLEr, X86::SETLEm },
3372     { X86::SETNEr, X86::SETNEm },
3373     { X86::SETNOr, X86::SETNOm },
3374     { X86::SETNPr, X86::SETNPm },
3375     { X86::SETNSr, X86::SETNSm },
3376     { X86::SETOr,  X86::SETOm  },
3377     { X86::SETPr,  X86::SETPm  },
3378     { X86::SETSr,  X86::SETSm  }
3379   };
3380 
3381   assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes");
3382   return Opc[CC][HasMemoryOperand ? 1 : 0];
3383 }
3384 
3385 /// Return a cmov opcode for the given condition,
3386 /// register size in bytes, and operand type.
3387 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes,
3388                               bool HasMemoryOperand) {
3389   static const uint16_t Opc[32][3] = {
3390     { X86::CMOVA16rr,  X86::CMOVA32rr,  X86::CMOVA64rr  },
3391     { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
3392     { X86::CMOVB16rr,  X86::CMOVB32rr,  X86::CMOVB64rr  },
3393     { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
3394     { X86::CMOVE16rr,  X86::CMOVE32rr,  X86::CMOVE64rr  },
3395     { X86::CMOVG16rr,  X86::CMOVG32rr,  X86::CMOVG64rr  },
3396     { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
3397     { X86::CMOVL16rr,  X86::CMOVL32rr,  X86::CMOVL64rr  },
3398     { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
3399     { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
3400     { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
3401     { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
3402     { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
3403     { X86::CMOVO16rr,  X86::CMOVO32rr,  X86::CMOVO64rr  },
3404     { X86::CMOVP16rr,  X86::CMOVP32rr,  X86::CMOVP64rr  },
3405     { X86::CMOVS16rr,  X86::CMOVS32rr,  X86::CMOVS64rr  },
3406     { X86::CMOVA16rm,  X86::CMOVA32rm,  X86::CMOVA64rm  },
3407     { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
3408     { X86::CMOVB16rm,  X86::CMOVB32rm,  X86::CMOVB64rm  },
3409     { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
3410     { X86::CMOVE16rm,  X86::CMOVE32rm,  X86::CMOVE64rm  },
3411     { X86::CMOVG16rm,  X86::CMOVG32rm,  X86::CMOVG64rm  },
3412     { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
3413     { X86::CMOVL16rm,  X86::CMOVL32rm,  X86::CMOVL64rm  },
3414     { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
3415     { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
3416     { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
3417     { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
3418     { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
3419     { X86::CMOVO16rm,  X86::CMOVO32rm,  X86::CMOVO64rm  },
3420     { X86::CMOVP16rm,  X86::CMOVP32rm,  X86::CMOVP64rm  },
3421     { X86::CMOVS16rm,  X86::CMOVS32rm,  X86::CMOVS64rm  }
3422   };
3423 
3424   assert(CC < 16 && "Can only handle standard cond codes");
3425   unsigned Idx = HasMemoryOperand ? 16+CC : CC;
3426   switch(RegBytes) {
3427   default: llvm_unreachable("Illegal register size!");
3428   case 2: return Opc[Idx][0];
3429   case 4: return Opc[Idx][1];
3430   case 8: return Opc[Idx][2];
3431   }
3432 }
3433 
3434 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
3435   if (!MI->isTerminator()) return false;
3436 
3437   // Conditional branch is a special case.
3438   if (MI->isBranch() && !MI->isBarrier())
3439     return true;
3440   if (!MI->isPredicable())
3441     return true;
3442   return !isPredicated(MI);
3443 }
3444 
3445 bool X86InstrInfo::AnalyzeBranchImpl(
3446     MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
3447     SmallVectorImpl<MachineOperand> &Cond,
3448     SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const {
3449 
3450   // Start from the bottom of the block and work up, examining the
3451   // terminator instructions.
3452   MachineBasicBlock::iterator I = MBB.end();
3453   MachineBasicBlock::iterator UnCondBrIter = MBB.end();
3454   while (I != MBB.begin()) {
3455     --I;
3456     if (I->isDebugValue())
3457       continue;
3458 
3459     // Working from the bottom, when we see a non-terminator instruction, we're
3460     // done.
3461     if (!isUnpredicatedTerminator(I))
3462       break;
3463 
3464     // A terminator that isn't a branch can't easily be handled by this
3465     // analysis.
3466     if (!I->isBranch())
3467       return true;
3468 
3469     // Handle unconditional branches.
3470     if (I->getOpcode() == X86::JMP_1) {
3471       UnCondBrIter = I;
3472 
3473       if (!AllowModify) {
3474         TBB = I->getOperand(0).getMBB();
3475         continue;
3476       }
3477 
3478       // If the block has any instructions after a JMP, delete them.
3479       while (std::next(I) != MBB.end())
3480         std::next(I)->eraseFromParent();
3481 
3482       Cond.clear();
3483       FBB = nullptr;
3484 
3485       // Delete the JMP if it's equivalent to a fall-through.
3486       if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
3487         TBB = nullptr;
3488         I->eraseFromParent();
3489         I = MBB.end();
3490         UnCondBrIter = MBB.end();
3491         continue;
3492       }
3493 
3494       // TBB is used to indicate the unconditional destination.
3495       TBB = I->getOperand(0).getMBB();
3496       continue;
3497     }
3498 
3499     // Handle conditional branches.
3500     X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
3501     if (BranchCode == X86::COND_INVALID)
3502       return true;  // Can't handle indirect branch.
3503 
3504     // Working from the bottom, handle the first conditional branch.
3505     if (Cond.empty()) {
3506       MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
3507       if (AllowModify && UnCondBrIter != MBB.end() &&
3508           MBB.isLayoutSuccessor(TargetBB)) {
3509         // If we can modify the code and it ends in something like:
3510         //
3511         //     jCC L1
3512         //     jmp L2
3513         //   L1:
3514         //     ...
3515         //   L2:
3516         //
3517         // Then we can change this to:
3518         //
3519         //     jnCC L2
3520         //   L1:
3521         //     ...
3522         //   L2:
3523         //
3524         // Which is a bit more efficient.
3525         // We conditionally jump to the fall-through block.
3526         BranchCode = GetOppositeBranchCondition(BranchCode);
3527         unsigned JNCC = GetCondBranchFromCond(BranchCode);
3528         MachineBasicBlock::iterator OldInst = I;
3529 
3530         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
3531           .addMBB(UnCondBrIter->getOperand(0).getMBB());
3532         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1))
3533           .addMBB(TargetBB);
3534 
3535         OldInst->eraseFromParent();
3536         UnCondBrIter->eraseFromParent();
3537 
3538         // Restart the analysis.
3539         UnCondBrIter = MBB.end();
3540         I = MBB.end();
3541         continue;
3542       }
3543 
3544       FBB = TBB;
3545       TBB = I->getOperand(0).getMBB();
3546       Cond.push_back(MachineOperand::CreateImm(BranchCode));
3547       CondBranches.push_back(I);
3548       continue;
3549     }
3550 
3551     // Handle subsequent conditional branches. Only handle the case where all
3552     // conditional branches branch to the same destination and their condition
3553     // opcodes fit one of the special multi-branch idioms.
3554     assert(Cond.size() == 1);
3555     assert(TBB);
3556 
3557     // Only handle the case where all conditional branches branch to the same
3558     // destination.
3559     if (TBB != I->getOperand(0).getMBB())
3560       return true;
3561 
3562     // If the conditions are the same, we can leave them alone.
3563     X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
3564     if (OldBranchCode == BranchCode)
3565       continue;
3566 
3567     // If they differ, see if they fit one of the known patterns. Theoretically,
3568     // we could handle more patterns here, but we shouldn't expect to see them
3569     // if instruction selection has done a reasonable job.
3570     if ((OldBranchCode == X86::COND_NP &&
3571          BranchCode == X86::COND_E) ||
3572         (OldBranchCode == X86::COND_E &&
3573          BranchCode == X86::COND_NP))
3574       BranchCode = X86::COND_NP_OR_E;
3575     else if ((OldBranchCode == X86::COND_P &&
3576               BranchCode == X86::COND_NE) ||
3577              (OldBranchCode == X86::COND_NE &&
3578               BranchCode == X86::COND_P))
3579       BranchCode = X86::COND_NE_OR_P;
3580     else
3581       return true;
3582 
3583     // Update the MachineOperand.
3584     Cond[0].setImm(BranchCode);
3585     CondBranches.push_back(I);
3586   }
3587 
3588   return false;
3589 }
3590 
3591 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
3592                                  MachineBasicBlock *&TBB,
3593                                  MachineBasicBlock *&FBB,
3594                                  SmallVectorImpl<MachineOperand> &Cond,
3595                                  bool AllowModify) const {
3596   SmallVector<MachineInstr *, 4> CondBranches;
3597   return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify);
3598 }
3599 
3600 bool X86InstrInfo::AnalyzeBranchPredicate(MachineBasicBlock &MBB,
3601                                           MachineBranchPredicate &MBP,
3602                                           bool AllowModify) const {
3603   using namespace std::placeholders;
3604 
3605   SmallVector<MachineOperand, 4> Cond;
3606   SmallVector<MachineInstr *, 4> CondBranches;
3607   if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches,
3608                         AllowModify))
3609     return true;
3610 
3611   if (Cond.size() != 1)
3612     return true;
3613 
3614   assert(MBP.TrueDest && "expected!");
3615 
3616   if (!MBP.FalseDest)
3617     MBP.FalseDest = MBB.getNextNode();
3618 
3619   const TargetRegisterInfo *TRI = &getRegisterInfo();
3620 
3621   MachineInstr *ConditionDef = nullptr;
3622   bool SingleUseCondition = true;
3623 
3624   for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) {
3625     if (I->modifiesRegister(X86::EFLAGS, TRI)) {
3626       ConditionDef = &*I;
3627       break;
3628     }
3629 
3630     if (I->readsRegister(X86::EFLAGS, TRI))
3631       SingleUseCondition = false;
3632   }
3633 
3634   if (!ConditionDef)
3635     return true;
3636 
3637   if (SingleUseCondition) {
3638     for (auto *Succ : MBB.successors())
3639       if (Succ->isLiveIn(X86::EFLAGS))
3640         SingleUseCondition = false;
3641   }
3642 
3643   MBP.ConditionDef = ConditionDef;
3644   MBP.SingleUseCondition = SingleUseCondition;
3645 
3646   // Currently we only recognize the simple pattern:
3647   //
3648   //   test %reg, %reg
3649   //   je %label
3650   //
3651   const unsigned TestOpcode =
3652       Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr;
3653 
3654   if (ConditionDef->getOpcode() == TestOpcode &&
3655       ConditionDef->getNumOperands() == 3 &&
3656       ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) &&
3657       (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) {
3658     MBP.LHS = ConditionDef->getOperand(0);
3659     MBP.RHS = MachineOperand::CreateImm(0);
3660     MBP.Predicate = Cond[0].getImm() == X86::COND_NE
3661                         ? MachineBranchPredicate::PRED_NE
3662                         : MachineBranchPredicate::PRED_EQ;
3663     return false;
3664   }
3665 
3666   return true;
3667 }
3668 
3669 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
3670   MachineBasicBlock::iterator I = MBB.end();
3671   unsigned Count = 0;
3672 
3673   while (I != MBB.begin()) {
3674     --I;
3675     if (I->isDebugValue())
3676       continue;
3677     if (I->getOpcode() != X86::JMP_1 &&
3678         getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
3679       break;
3680     // Remove the branch.
3681     I->eraseFromParent();
3682     I = MBB.end();
3683     ++Count;
3684   }
3685 
3686   return Count;
3687 }
3688 
3689 unsigned
3690 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
3691                            MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
3692                            DebugLoc DL) const {
3693   // Shouldn't be a fall through.
3694   assert(TBB && "InsertBranch must not be told to insert a fallthrough");
3695   assert((Cond.size() == 1 || Cond.size() == 0) &&
3696          "X86 branch conditions have one component!");
3697 
3698   if (Cond.empty()) {
3699     // Unconditional branch?
3700     assert(!FBB && "Unconditional branch with multiple successors!");
3701     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB);
3702     return 1;
3703   }
3704 
3705   // Conditional branch.
3706   unsigned Count = 0;
3707   X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
3708   switch (CC) {
3709   case X86::COND_NP_OR_E:
3710     // Synthesize NP_OR_E with two branches.
3711     BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB);
3712     ++Count;
3713     BuildMI(&MBB, DL, get(X86::JE_1)).addMBB(TBB);
3714     ++Count;
3715     break;
3716   case X86::COND_NE_OR_P:
3717     // Synthesize NE_OR_P with two branches.
3718     BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB);
3719     ++Count;
3720     BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB);
3721     ++Count;
3722     break;
3723   default: {
3724     unsigned Opc = GetCondBranchFromCond(CC);
3725     BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
3726     ++Count;
3727   }
3728   }
3729   if (FBB) {
3730     // Two-way Conditional branch. Insert the second branch.
3731     BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB);
3732     ++Count;
3733   }
3734   return Count;
3735 }
3736 
3737 bool X86InstrInfo::
3738 canInsertSelect(const MachineBasicBlock &MBB,
3739                 ArrayRef<MachineOperand> Cond,
3740                 unsigned TrueReg, unsigned FalseReg,
3741                 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
3742   // Not all subtargets have cmov instructions.
3743   if (!Subtarget.hasCMov())
3744     return false;
3745   if (Cond.size() != 1)
3746     return false;
3747   // We cannot do the composite conditions, at least not in SSA form.
3748   if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
3749     return false;
3750 
3751   // Check register classes.
3752   const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3753   const TargetRegisterClass *RC =
3754     RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
3755   if (!RC)
3756     return false;
3757 
3758   // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
3759   if (X86::GR16RegClass.hasSubClassEq(RC) ||
3760       X86::GR32RegClass.hasSubClassEq(RC) ||
3761       X86::GR64RegClass.hasSubClassEq(RC)) {
3762     // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
3763     // Bridge. Probably Ivy Bridge as well.
3764     CondCycles = 2;
3765     TrueCycles = 2;
3766     FalseCycles = 2;
3767     return true;
3768   }
3769 
3770   // Can't do vectors.
3771   return false;
3772 }
3773 
3774 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
3775                                 MachineBasicBlock::iterator I, DebugLoc DL,
3776                                 unsigned DstReg, ArrayRef<MachineOperand> Cond,
3777                                 unsigned TrueReg, unsigned FalseReg) const {
3778    MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3779    assert(Cond.size() == 1 && "Invalid Cond array");
3780    unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
3781                                   MRI.getRegClass(DstReg)->getSize(),
3782                                   false/*HasMemoryOperand*/);
3783    BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
3784 }
3785 
3786 /// Test if the given register is a physical h register.
3787 static bool isHReg(unsigned Reg) {
3788   return X86::GR8_ABCD_HRegClass.contains(Reg);
3789 }
3790 
3791 // Try and copy between VR128/VR64 and GR64 registers.
3792 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
3793                                         const X86Subtarget &Subtarget) {
3794 
3795   // SrcReg(VR128) -> DestReg(GR64)
3796   // SrcReg(VR64)  -> DestReg(GR64)
3797   // SrcReg(GR64)  -> DestReg(VR128)
3798   // SrcReg(GR64)  -> DestReg(VR64)
3799 
3800   bool HasAVX = Subtarget.hasAVX();
3801   bool HasAVX512 = Subtarget.hasAVX512();
3802   if (X86::GR64RegClass.contains(DestReg)) {
3803     if (X86::VR128XRegClass.contains(SrcReg))
3804       // Copy from a VR128 register to a GR64 register.
3805       return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
3806                                                X86::MOVPQIto64rr);
3807     if (X86::VR64RegClass.contains(SrcReg))
3808       // Copy from a VR64 register to a GR64 register.
3809       return X86::MMX_MOVD64from64rr;
3810   } else if (X86::GR64RegClass.contains(SrcReg)) {
3811     // Copy from a GR64 register to a VR128 register.
3812     if (X86::VR128XRegClass.contains(DestReg))
3813       return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
3814                                                X86::MOV64toPQIrr);
3815     // Copy from a GR64 register to a VR64 register.
3816     if (X86::VR64RegClass.contains(DestReg))
3817       return X86::MMX_MOVD64to64rr;
3818   }
3819 
3820   // SrcReg(FR32) -> DestReg(GR32)
3821   // SrcReg(GR32) -> DestReg(FR32)
3822 
3823   if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3824     // Copy from a FR32 register to a GR32 register.
3825     return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3826 
3827   if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3828     // Copy from a GR32 register to a FR32 register.
3829     return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3830   return 0;
3831 }
3832 
3833 inline static bool MaskRegClassContains(unsigned Reg) {
3834   return X86::VK8RegClass.contains(Reg) ||
3835          X86::VK16RegClass.contains(Reg) ||
3836          X86::VK32RegClass.contains(Reg) ||
3837          X86::VK64RegClass.contains(Reg) ||
3838          X86::VK1RegClass.contains(Reg);
3839 }
3840 static
3841 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3842   if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3843       X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3844       X86::VR512RegClass.contains(DestReg, SrcReg)) {
3845      DestReg = get512BitSuperRegister(DestReg);
3846      SrcReg = get512BitSuperRegister(SrcReg);
3847      return X86::VMOVAPSZrr;
3848   }
3849   if (MaskRegClassContains(DestReg) &&
3850       MaskRegClassContains(SrcReg))
3851     return X86::KMOVWkk;
3852   if (MaskRegClassContains(DestReg) &&
3853       (X86::GR32RegClass.contains(SrcReg) ||
3854        X86::GR16RegClass.contains(SrcReg) ||
3855        X86::GR8RegClass.contains(SrcReg))) {
3856     SrcReg = getX86SubSuperRegister(SrcReg, MVT::i32);
3857     return X86::KMOVWkr;
3858   }
3859   if ((X86::GR32RegClass.contains(DestReg) ||
3860        X86::GR16RegClass.contains(DestReg) ||
3861        X86::GR8RegClass.contains(DestReg)) &&
3862        MaskRegClassContains(SrcReg)) {
3863     DestReg = getX86SubSuperRegister(DestReg, MVT::i32);
3864     return X86::KMOVWrk;
3865   }
3866   return 0;
3867 }
3868 
3869 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3870                                MachineBasicBlock::iterator MI, DebugLoc DL,
3871                                unsigned DestReg, unsigned SrcReg,
3872                                bool KillSrc) const {
3873   // First deal with the normal symmetric copies.
3874   bool HasAVX = Subtarget.hasAVX();
3875   bool HasAVX512 = Subtarget.hasAVX512();
3876   unsigned Opc = 0;
3877   if (X86::GR64RegClass.contains(DestReg, SrcReg))
3878     Opc = X86::MOV64rr;
3879   else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3880     Opc = X86::MOV32rr;
3881   else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3882     Opc = X86::MOV16rr;
3883   else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3884     // Copying to or from a physical H register on x86-64 requires a NOREX
3885     // move.  Otherwise use a normal move.
3886     if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3887         Subtarget.is64Bit()) {
3888       Opc = X86::MOV8rr_NOREX;
3889       // Both operands must be encodable without an REX prefix.
3890       assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3891              "8-bit H register can not be copied outside GR8_NOREX");
3892     } else
3893       Opc = X86::MOV8rr;
3894   }
3895   else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3896     Opc = X86::MMX_MOVQ64rr;
3897   else if (HasAVX512)
3898     Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3899   else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3900     Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3901   else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3902     Opc = X86::VMOVAPSYrr;
3903   if (!Opc)
3904     Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget);
3905 
3906   if (Opc) {
3907     BuildMI(MBB, MI, DL, get(Opc), DestReg)
3908       .addReg(SrcReg, getKillRegState(KillSrc));
3909     return;
3910   }
3911 
3912   // Moving EFLAGS to / from another register requires a push and a pop.
3913   // Notice that we have to adjust the stack if we don't want to clobber the
3914   // first frame index. See X86FrameLowering.cpp - clobbersTheStack.
3915   if (SrcReg == X86::EFLAGS) {
3916     if (X86::GR64RegClass.contains(DestReg)) {
3917       BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3918       BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3919       return;
3920     }
3921     if (X86::GR32RegClass.contains(DestReg)) {
3922       BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3923       BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3924       return;
3925     }
3926   }
3927   if (DestReg == X86::EFLAGS) {
3928     if (X86::GR64RegClass.contains(SrcReg)) {
3929       BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3930         .addReg(SrcReg, getKillRegState(KillSrc));
3931       BuildMI(MBB, MI, DL, get(X86::POPF64));
3932       return;
3933     }
3934     if (X86::GR32RegClass.contains(SrcReg)) {
3935       BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3936         .addReg(SrcReg, getKillRegState(KillSrc));
3937       BuildMI(MBB, MI, DL, get(X86::POPF32));
3938       return;
3939     }
3940   }
3941 
3942   DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3943                << " to " << RI.getName(DestReg) << '\n');
3944   llvm_unreachable("Cannot emit physreg copy instruction");
3945 }
3946 
3947 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3948                                       const TargetRegisterClass *RC,
3949                                       bool isStackAligned,
3950                                       const X86Subtarget &STI,
3951                                       bool load) {
3952   if (STI.hasAVX512()) {
3953     if (X86::VK8RegClass.hasSubClassEq(RC)  ||
3954       X86::VK16RegClass.hasSubClassEq(RC))
3955       return load ? X86::KMOVWkm : X86::KMOVWmk;
3956     if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3957       return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3958     if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3959       return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3960     if (X86::VR512RegClass.hasSubClassEq(RC))
3961       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3962   }
3963 
3964   bool HasAVX = STI.hasAVX();
3965   switch (RC->getSize()) {
3966   default:
3967     llvm_unreachable("Unknown spill size");
3968   case 1:
3969     assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3970     if (STI.is64Bit())
3971       // Copying to or from a physical H register on x86-64 requires a NOREX
3972       // move.  Otherwise use a normal move.
3973       if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3974         return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3975     return load ? X86::MOV8rm : X86::MOV8mr;
3976   case 2:
3977     assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3978     return load ? X86::MOV16rm : X86::MOV16mr;
3979   case 4:
3980     if (X86::GR32RegClass.hasSubClassEq(RC))
3981       return load ? X86::MOV32rm : X86::MOV32mr;
3982     if (X86::FR32RegClass.hasSubClassEq(RC))
3983       return load ?
3984         (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3985         (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3986     if (X86::RFP32RegClass.hasSubClassEq(RC))
3987       return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3988     llvm_unreachable("Unknown 4-byte regclass");
3989   case 8:
3990     if (X86::GR64RegClass.hasSubClassEq(RC))
3991       return load ? X86::MOV64rm : X86::MOV64mr;
3992     if (X86::FR64RegClass.hasSubClassEq(RC))
3993       return load ?
3994         (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3995         (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3996     if (X86::VR64RegClass.hasSubClassEq(RC))
3997       return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3998     if (X86::RFP64RegClass.hasSubClassEq(RC))
3999       return load ? X86::LD_Fp64m : X86::ST_Fp64m;
4000     llvm_unreachable("Unknown 8-byte regclass");
4001   case 10:
4002     assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
4003     return load ? X86::LD_Fp80m : X86::ST_FpP80m;
4004   case 16: {
4005     assert((X86::VR128RegClass.hasSubClassEq(RC) ||
4006             X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass");
4007     // If stack is realigned we can use aligned stores.
4008     if (isStackAligned)
4009       return load ?
4010         (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
4011         (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
4012     else
4013       return load ?
4014         (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
4015         (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
4016   }
4017   case 32:
4018     assert((X86::VR256RegClass.hasSubClassEq(RC) ||
4019             X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass");
4020     // If stack is realigned we can use aligned stores.
4021     if (isStackAligned)
4022       return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
4023     else
4024       return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
4025   case 64:
4026     assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
4027     if (isStackAligned)
4028       return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
4029     else
4030       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
4031   }
4032 }
4033 
4034 bool X86InstrInfo::getMemOpBaseRegImmOfs(MachineInstr *MemOp, unsigned &BaseReg,
4035                                          unsigned &Offset,
4036                                          const TargetRegisterInfo *TRI) const {
4037   const MCInstrDesc &Desc = MemOp->getDesc();
4038   int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags, MemOp->getOpcode());
4039   if (MemRefBegin < 0)
4040     return false;
4041 
4042   MemRefBegin += X86II::getOperandBias(Desc);
4043 
4044   BaseReg = MemOp->getOperand(MemRefBegin + X86::AddrBaseReg).getReg();
4045   if (MemOp->getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1)
4046     return false;
4047 
4048   if (MemOp->getOperand(MemRefBegin + X86::AddrIndexReg).getReg() !=
4049       X86::NoRegister)
4050     return false;
4051 
4052   const MachineOperand &DispMO = MemOp->getOperand(MemRefBegin + X86::AddrDisp);
4053 
4054   // Displacement can be symbolic
4055   if (!DispMO.isImm())
4056     return false;
4057 
4058   Offset = DispMO.getImm();
4059 
4060   return (MemOp->getOperand(MemRefBegin + X86::AddrIndexReg).getReg() ==
4061           X86::NoRegister);
4062 }
4063 
4064 static unsigned getStoreRegOpcode(unsigned SrcReg,
4065                                   const TargetRegisterClass *RC,
4066                                   bool isStackAligned,
4067                                   const X86Subtarget &STI) {
4068   return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false);
4069 }
4070 
4071 
4072 static unsigned getLoadRegOpcode(unsigned DestReg,
4073                                  const TargetRegisterClass *RC,
4074                                  bool isStackAligned,
4075                                  const X86Subtarget &STI) {
4076   return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true);
4077 }
4078 
4079 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
4080                                        MachineBasicBlock::iterator MI,
4081                                        unsigned SrcReg, bool isKill, int FrameIdx,
4082                                        const TargetRegisterClass *RC,
4083                                        const TargetRegisterInfo *TRI) const {
4084   const MachineFunction &MF = *MBB.getParent();
4085   assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
4086          "Stack slot too small for store");
4087   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4088   bool isAligned =
4089       (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
4090       RI.canRealignStack(MF);
4091   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
4092   DebugLoc DL = MBB.findDebugLoc(MI);
4093   addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
4094     .addReg(SrcReg, getKillRegState(isKill));
4095 }
4096 
4097 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
4098                                   bool isKill,
4099                                   SmallVectorImpl<MachineOperand> &Addr,
4100                                   const TargetRegisterClass *RC,
4101                                   MachineInstr::mmo_iterator MMOBegin,
4102                                   MachineInstr::mmo_iterator MMOEnd,
4103                                   SmallVectorImpl<MachineInstr*> &NewMIs) const {
4104   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4105   bool isAligned = MMOBegin != MMOEnd &&
4106                    (*MMOBegin)->getAlignment() >= Alignment;
4107   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget);
4108   DebugLoc DL;
4109   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
4110   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
4111     MIB.addOperand(Addr[i]);
4112   MIB.addReg(SrcReg, getKillRegState(isKill));
4113   (*MIB).setMemRefs(MMOBegin, MMOEnd);
4114   NewMIs.push_back(MIB);
4115 }
4116 
4117 
4118 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
4119                                         MachineBasicBlock::iterator MI,
4120                                         unsigned DestReg, int FrameIdx,
4121                                         const TargetRegisterClass *RC,
4122                                         const TargetRegisterInfo *TRI) const {
4123   const MachineFunction &MF = *MBB.getParent();
4124   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4125   bool isAligned =
4126       (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) ||
4127       RI.canRealignStack(MF);
4128   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
4129   DebugLoc DL = MBB.findDebugLoc(MI);
4130   addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
4131 }
4132 
4133 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
4134                                  SmallVectorImpl<MachineOperand> &Addr,
4135                                  const TargetRegisterClass *RC,
4136                                  MachineInstr::mmo_iterator MMOBegin,
4137                                  MachineInstr::mmo_iterator MMOEnd,
4138                                  SmallVectorImpl<MachineInstr*> &NewMIs) const {
4139   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
4140   bool isAligned = MMOBegin != MMOEnd &&
4141                    (*MMOBegin)->getAlignment() >= Alignment;
4142   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget);
4143   DebugLoc DL;
4144   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
4145   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
4146     MIB.addOperand(Addr[i]);
4147   (*MIB).setMemRefs(MMOBegin, MMOEnd);
4148   NewMIs.push_back(MIB);
4149 }
4150 
4151 bool X86InstrInfo::
4152 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
4153                int &CmpMask, int &CmpValue) const {
4154   switch (MI->getOpcode()) {
4155   default: break;
4156   case X86::CMP64ri32:
4157   case X86::CMP64ri8:
4158   case X86::CMP32ri:
4159   case X86::CMP32ri8:
4160   case X86::CMP16ri:
4161   case X86::CMP16ri8:
4162   case X86::CMP8ri:
4163     SrcReg = MI->getOperand(0).getReg();
4164     SrcReg2 = 0;
4165     CmpMask = ~0;
4166     CmpValue = MI->getOperand(1).getImm();
4167     return true;
4168   // A SUB can be used to perform comparison.
4169   case X86::SUB64rm:
4170   case X86::SUB32rm:
4171   case X86::SUB16rm:
4172   case X86::SUB8rm:
4173     SrcReg = MI->getOperand(1).getReg();
4174     SrcReg2 = 0;
4175     CmpMask = ~0;
4176     CmpValue = 0;
4177     return true;
4178   case X86::SUB64rr:
4179   case X86::SUB32rr:
4180   case X86::SUB16rr:
4181   case X86::SUB8rr:
4182     SrcReg = MI->getOperand(1).getReg();
4183     SrcReg2 = MI->getOperand(2).getReg();
4184     CmpMask = ~0;
4185     CmpValue = 0;
4186     return true;
4187   case X86::SUB64ri32:
4188   case X86::SUB64ri8:
4189   case X86::SUB32ri:
4190   case X86::SUB32ri8:
4191   case X86::SUB16ri:
4192   case X86::SUB16ri8:
4193   case X86::SUB8ri:
4194     SrcReg = MI->getOperand(1).getReg();
4195     SrcReg2 = 0;
4196     CmpMask = ~0;
4197     CmpValue = MI->getOperand(2).getImm();
4198     return true;
4199   case X86::CMP64rr:
4200   case X86::CMP32rr:
4201   case X86::CMP16rr:
4202   case X86::CMP8rr:
4203     SrcReg = MI->getOperand(0).getReg();
4204     SrcReg2 = MI->getOperand(1).getReg();
4205     CmpMask = ~0;
4206     CmpValue = 0;
4207     return true;
4208   case X86::TEST8rr:
4209   case X86::TEST16rr:
4210   case X86::TEST32rr:
4211   case X86::TEST64rr:
4212     SrcReg = MI->getOperand(0).getReg();
4213     if (MI->getOperand(1).getReg() != SrcReg) return false;
4214     // Compare against zero.
4215     SrcReg2 = 0;
4216     CmpMask = ~0;
4217     CmpValue = 0;
4218     return true;
4219   }
4220   return false;
4221 }
4222 
4223 /// Check whether the first instruction, whose only
4224 /// purpose is to update flags, can be made redundant.
4225 /// CMPrr can be made redundant by SUBrr if the operands are the same.
4226 /// This function can be extended later on.
4227 /// SrcReg, SrcRegs: register operands for FlagI.
4228 /// ImmValue: immediate for FlagI if it takes an immediate.
4229 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
4230                                         unsigned SrcReg2, int ImmValue,
4231                                         MachineInstr *OI) {
4232   if (((FlagI->getOpcode() == X86::CMP64rr &&
4233         OI->getOpcode() == X86::SUB64rr) ||
4234        (FlagI->getOpcode() == X86::CMP32rr &&
4235         OI->getOpcode() == X86::SUB32rr)||
4236        (FlagI->getOpcode() == X86::CMP16rr &&
4237         OI->getOpcode() == X86::SUB16rr)||
4238        (FlagI->getOpcode() == X86::CMP8rr &&
4239         OI->getOpcode() == X86::SUB8rr)) &&
4240       ((OI->getOperand(1).getReg() == SrcReg &&
4241         OI->getOperand(2).getReg() == SrcReg2) ||
4242        (OI->getOperand(1).getReg() == SrcReg2 &&
4243         OI->getOperand(2).getReg() == SrcReg)))
4244     return true;
4245 
4246   if (((FlagI->getOpcode() == X86::CMP64ri32 &&
4247         OI->getOpcode() == X86::SUB64ri32) ||
4248        (FlagI->getOpcode() == X86::CMP64ri8 &&
4249         OI->getOpcode() == X86::SUB64ri8) ||
4250        (FlagI->getOpcode() == X86::CMP32ri &&
4251         OI->getOpcode() == X86::SUB32ri) ||
4252        (FlagI->getOpcode() == X86::CMP32ri8 &&
4253         OI->getOpcode() == X86::SUB32ri8) ||
4254        (FlagI->getOpcode() == X86::CMP16ri &&
4255         OI->getOpcode() == X86::SUB16ri) ||
4256        (FlagI->getOpcode() == X86::CMP16ri8 &&
4257         OI->getOpcode() == X86::SUB16ri8) ||
4258        (FlagI->getOpcode() == X86::CMP8ri &&
4259         OI->getOpcode() == X86::SUB8ri)) &&
4260       OI->getOperand(1).getReg() == SrcReg &&
4261       OI->getOperand(2).getImm() == ImmValue)
4262     return true;
4263   return false;
4264 }
4265 
4266 /// Check whether the definition can be converted
4267 /// to remove a comparison against zero.
4268 inline static bool isDefConvertible(MachineInstr *MI) {
4269   switch (MI->getOpcode()) {
4270   default: return false;
4271 
4272   // The shift instructions only modify ZF if their shift count is non-zero.
4273   // N.B.: The processor truncates the shift count depending on the encoding.
4274   case X86::SAR8ri:    case X86::SAR16ri:  case X86::SAR32ri:case X86::SAR64ri:
4275   case X86::SHR8ri:    case X86::SHR16ri:  case X86::SHR32ri:case X86::SHR64ri:
4276      return getTruncatedShiftCount(MI, 2) != 0;
4277 
4278   // Some left shift instructions can be turned into LEA instructions but only
4279   // if their flags aren't used. Avoid transforming such instructions.
4280   case X86::SHL8ri:    case X86::SHL16ri:  case X86::SHL32ri:case X86::SHL64ri:{
4281     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
4282     if (isTruncatedShiftCountForLEA(ShAmt)) return false;
4283     return ShAmt != 0;
4284   }
4285 
4286   case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
4287   case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
4288      return getTruncatedShiftCount(MI, 3) != 0;
4289 
4290   case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
4291   case X86::SUB32ri8:  case X86::SUB16ri:  case X86::SUB16ri8:
4292   case X86::SUB8ri:    case X86::SUB64rr:  case X86::SUB32rr:
4293   case X86::SUB16rr:   case X86::SUB8rr:   case X86::SUB64rm:
4294   case X86::SUB32rm:   case X86::SUB16rm:  case X86::SUB8rm:
4295   case X86::DEC64r:    case X86::DEC32r:   case X86::DEC16r: case X86::DEC8r:
4296   case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
4297   case X86::ADD32ri8:  case X86::ADD16ri:  case X86::ADD16ri8:
4298   case X86::ADD8ri:    case X86::ADD64rr:  case X86::ADD32rr:
4299   case X86::ADD16rr:   case X86::ADD8rr:   case X86::ADD64rm:
4300   case X86::ADD32rm:   case X86::ADD16rm:  case X86::ADD8rm:
4301   case X86::INC64r:    case X86::INC32r:   case X86::INC16r: case X86::INC8r:
4302   case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
4303   case X86::AND32ri8:  case X86::AND16ri:  case X86::AND16ri8:
4304   case X86::AND8ri:    case X86::AND64rr:  case X86::AND32rr:
4305   case X86::AND16rr:   case X86::AND8rr:   case X86::AND64rm:
4306   case X86::AND32rm:   case X86::AND16rm:  case X86::AND8rm:
4307   case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
4308   case X86::XOR32ri8:  case X86::XOR16ri:  case X86::XOR16ri8:
4309   case X86::XOR8ri:    case X86::XOR64rr:  case X86::XOR32rr:
4310   case X86::XOR16rr:   case X86::XOR8rr:   case X86::XOR64rm:
4311   case X86::XOR32rm:   case X86::XOR16rm:  case X86::XOR8rm:
4312   case X86::OR64ri32:  case X86::OR64ri8:  case X86::OR32ri:
4313   case X86::OR32ri8:   case X86::OR16ri:   case X86::OR16ri8:
4314   case X86::OR8ri:     case X86::OR64rr:   case X86::OR32rr:
4315   case X86::OR16rr:    case X86::OR8rr:    case X86::OR64rm:
4316   case X86::OR32rm:    case X86::OR16rm:   case X86::OR8rm:
4317   case X86::NEG8r:     case X86::NEG16r:   case X86::NEG32r: case X86::NEG64r:
4318   case X86::SAR8r1:    case X86::SAR16r1:  case X86::SAR32r1:case X86::SAR64r1:
4319   case X86::SHR8r1:    case X86::SHR16r1:  case X86::SHR32r1:case X86::SHR64r1:
4320   case X86::SHL8r1:    case X86::SHL16r1:  case X86::SHL32r1:case X86::SHL64r1:
4321   case X86::ADC32ri:   case X86::ADC32ri8:
4322   case X86::ADC32rr:   case X86::ADC64ri32:
4323   case X86::ADC64ri8:  case X86::ADC64rr:
4324   case X86::SBB32ri:   case X86::SBB32ri8:
4325   case X86::SBB32rr:   case X86::SBB64ri32:
4326   case X86::SBB64ri8:  case X86::SBB64rr:
4327   case X86::ANDN32rr:  case X86::ANDN32rm:
4328   case X86::ANDN64rr:  case X86::ANDN64rm:
4329   case X86::BEXTR32rr: case X86::BEXTR64rr:
4330   case X86::BEXTR32rm: case X86::BEXTR64rm:
4331   case X86::BLSI32rr:  case X86::BLSI32rm:
4332   case X86::BLSI64rr:  case X86::BLSI64rm:
4333   case X86::BLSMSK32rr:case X86::BLSMSK32rm:
4334   case X86::BLSMSK64rr:case X86::BLSMSK64rm:
4335   case X86::BLSR32rr:  case X86::BLSR32rm:
4336   case X86::BLSR64rr:  case X86::BLSR64rm:
4337   case X86::BZHI32rr:  case X86::BZHI32rm:
4338   case X86::BZHI64rr:  case X86::BZHI64rm:
4339   case X86::LZCNT16rr: case X86::LZCNT16rm:
4340   case X86::LZCNT32rr: case X86::LZCNT32rm:
4341   case X86::LZCNT64rr: case X86::LZCNT64rm:
4342   case X86::POPCNT16rr:case X86::POPCNT16rm:
4343   case X86::POPCNT32rr:case X86::POPCNT32rm:
4344   case X86::POPCNT64rr:case X86::POPCNT64rm:
4345   case X86::TZCNT16rr: case X86::TZCNT16rm:
4346   case X86::TZCNT32rr: case X86::TZCNT32rm:
4347   case X86::TZCNT64rr: case X86::TZCNT64rm:
4348     return true;
4349   }
4350 }
4351 
4352 /// Check whether the use can be converted to remove a comparison against zero.
4353 static X86::CondCode isUseDefConvertible(MachineInstr *MI) {
4354   switch (MI->getOpcode()) {
4355   default: return X86::COND_INVALID;
4356   case X86::LZCNT16rr: case X86::LZCNT16rm:
4357   case X86::LZCNT32rr: case X86::LZCNT32rm:
4358   case X86::LZCNT64rr: case X86::LZCNT64rm:
4359     return X86::COND_B;
4360   case X86::POPCNT16rr:case X86::POPCNT16rm:
4361   case X86::POPCNT32rr:case X86::POPCNT32rm:
4362   case X86::POPCNT64rr:case X86::POPCNT64rm:
4363     return X86::COND_E;
4364   case X86::TZCNT16rr: case X86::TZCNT16rm:
4365   case X86::TZCNT32rr: case X86::TZCNT32rm:
4366   case X86::TZCNT64rr: case X86::TZCNT64rm:
4367     return X86::COND_B;
4368   }
4369 }
4370 
4371 /// Check if there exists an earlier instruction that
4372 /// operates on the same source operands and sets flags in the same way as
4373 /// Compare; remove Compare if possible.
4374 bool X86InstrInfo::
4375 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
4376                      int CmpMask, int CmpValue,
4377                      const MachineRegisterInfo *MRI) const {
4378   // Check whether we can replace SUB with CMP.
4379   unsigned NewOpcode = 0;
4380   switch (CmpInstr->getOpcode()) {
4381   default: break;
4382   case X86::SUB64ri32:
4383   case X86::SUB64ri8:
4384   case X86::SUB32ri:
4385   case X86::SUB32ri8:
4386   case X86::SUB16ri:
4387   case X86::SUB16ri8:
4388   case X86::SUB8ri:
4389   case X86::SUB64rm:
4390   case X86::SUB32rm:
4391   case X86::SUB16rm:
4392   case X86::SUB8rm:
4393   case X86::SUB64rr:
4394   case X86::SUB32rr:
4395   case X86::SUB16rr:
4396   case X86::SUB8rr: {
4397     if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
4398       return false;
4399     // There is no use of the destination register, we can replace SUB with CMP.
4400     switch (CmpInstr->getOpcode()) {
4401     default: llvm_unreachable("Unreachable!");
4402     case X86::SUB64rm:   NewOpcode = X86::CMP64rm;   break;
4403     case X86::SUB32rm:   NewOpcode = X86::CMP32rm;   break;
4404     case X86::SUB16rm:   NewOpcode = X86::CMP16rm;   break;
4405     case X86::SUB8rm:    NewOpcode = X86::CMP8rm;    break;
4406     case X86::SUB64rr:   NewOpcode = X86::CMP64rr;   break;
4407     case X86::SUB32rr:   NewOpcode = X86::CMP32rr;   break;
4408     case X86::SUB16rr:   NewOpcode = X86::CMP16rr;   break;
4409     case X86::SUB8rr:    NewOpcode = X86::CMP8rr;    break;
4410     case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
4411     case X86::SUB64ri8:  NewOpcode = X86::CMP64ri8;  break;
4412     case X86::SUB32ri:   NewOpcode = X86::CMP32ri;   break;
4413     case X86::SUB32ri8:  NewOpcode = X86::CMP32ri8;  break;
4414     case X86::SUB16ri:   NewOpcode = X86::CMP16ri;   break;
4415     case X86::SUB16ri8:  NewOpcode = X86::CMP16ri8;  break;
4416     case X86::SUB8ri:    NewOpcode = X86::CMP8ri;    break;
4417     }
4418     CmpInstr->setDesc(get(NewOpcode));
4419     CmpInstr->RemoveOperand(0);
4420     // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
4421     if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
4422         NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
4423       return false;
4424   }
4425   }
4426 
4427   // Get the unique definition of SrcReg.
4428   MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
4429   if (!MI) return false;
4430 
4431   // CmpInstr is the first instruction of the BB.
4432   MachineBasicBlock::iterator I = CmpInstr, Def = MI;
4433 
4434   // If we are comparing against zero, check whether we can use MI to update
4435   // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
4436   bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
4437   if (IsCmpZero && MI->getParent() != CmpInstr->getParent())
4438     return false;
4439 
4440   // If we have a use of the source register between the def and our compare
4441   // instruction we can eliminate the compare iff the use sets EFLAGS in the
4442   // right way.
4443   bool ShouldUpdateCC = false;
4444   X86::CondCode NewCC = X86::COND_INVALID;
4445   if (IsCmpZero && !isDefConvertible(MI)) {
4446     // Scan forward from the use until we hit the use we're looking for or the
4447     // compare instruction.
4448     for (MachineBasicBlock::iterator J = MI;; ++J) {
4449       // Do we have a convertible instruction?
4450       NewCC = isUseDefConvertible(J);
4451       if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() &&
4452           J->getOperand(1).getReg() == SrcReg) {
4453         assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!");
4454         ShouldUpdateCC = true; // Update CC later on.
4455         // This is not a def of SrcReg, but still a def of EFLAGS. Keep going
4456         // with the new def.
4457         MI = Def = J;
4458         break;
4459       }
4460 
4461       if (J == I)
4462         return false;
4463     }
4464   }
4465 
4466   // We are searching for an earlier instruction that can make CmpInstr
4467   // redundant and that instruction will be saved in Sub.
4468   MachineInstr *Sub = nullptr;
4469   const TargetRegisterInfo *TRI = &getRegisterInfo();
4470 
4471   // We iterate backward, starting from the instruction before CmpInstr and
4472   // stop when reaching the definition of a source register or done with the BB.
4473   // RI points to the instruction before CmpInstr.
4474   // If the definition is in this basic block, RE points to the definition;
4475   // otherwise, RE is the rend of the basic block.
4476   MachineBasicBlock::reverse_iterator
4477       RI = MachineBasicBlock::reverse_iterator(I),
4478       RE = CmpInstr->getParent() == MI->getParent() ?
4479            MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
4480            CmpInstr->getParent()->rend();
4481   MachineInstr *Movr0Inst = nullptr;
4482   for (; RI != RE; ++RI) {
4483     MachineInstr *Instr = &*RI;
4484     // Check whether CmpInstr can be made redundant by the current instruction.
4485     if (!IsCmpZero &&
4486         isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
4487       Sub = Instr;
4488       break;
4489     }
4490 
4491     if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
4492         Instr->readsRegister(X86::EFLAGS, TRI)) {
4493       // This instruction modifies or uses EFLAGS.
4494 
4495       // MOV32r0 etc. are implemented with xor which clobbers condition code.
4496       // They are safe to move up, if the definition to EFLAGS is dead and
4497       // earlier instructions do not read or write EFLAGS.
4498       if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
4499           Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
4500         Movr0Inst = Instr;
4501         continue;
4502       }
4503 
4504       // We can't remove CmpInstr.
4505       return false;
4506     }
4507   }
4508 
4509   // Return false if no candidates exist.
4510   if (!IsCmpZero && !Sub)
4511     return false;
4512 
4513   bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
4514                     Sub->getOperand(2).getReg() == SrcReg);
4515 
4516   // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
4517   // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
4518   // If we are done with the basic block, we need to check whether EFLAGS is
4519   // live-out.
4520   bool IsSafe = false;
4521   SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
4522   MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
4523   for (++I; I != E; ++I) {
4524     const MachineInstr &Instr = *I;
4525     bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
4526     bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
4527     // We should check the usage if this instruction uses and updates EFLAGS.
4528     if (!UseEFLAGS && ModifyEFLAGS) {
4529       // It is safe to remove CmpInstr if EFLAGS is updated again.
4530       IsSafe = true;
4531       break;
4532     }
4533     if (!UseEFLAGS && !ModifyEFLAGS)
4534       continue;
4535 
4536     // EFLAGS is used by this instruction.
4537     X86::CondCode OldCC = X86::COND_INVALID;
4538     bool OpcIsSET = false;
4539     if (IsCmpZero || IsSwapped) {
4540       // We decode the condition code from opcode.
4541       if (Instr.isBranch())
4542         OldCC = getCondFromBranchOpc(Instr.getOpcode());
4543       else {
4544         OldCC = getCondFromSETOpc(Instr.getOpcode());
4545         if (OldCC != X86::COND_INVALID)
4546           OpcIsSET = true;
4547         else
4548           OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
4549       }
4550       if (OldCC == X86::COND_INVALID) return false;
4551     }
4552     if (IsCmpZero) {
4553       switch (OldCC) {
4554       default: break;
4555       case X86::COND_A: case X86::COND_AE:
4556       case X86::COND_B: case X86::COND_BE:
4557       case X86::COND_G: case X86::COND_GE:
4558       case X86::COND_L: case X86::COND_LE:
4559       case X86::COND_O: case X86::COND_NO:
4560         // CF and OF are used, we can't perform this optimization.
4561         return false;
4562       }
4563 
4564       // If we're updating the condition code check if we have to reverse the
4565       // condition.
4566       if (ShouldUpdateCC)
4567         switch (OldCC) {
4568         default:
4569           return false;
4570         case X86::COND_E:
4571           break;
4572         case X86::COND_NE:
4573           NewCC = GetOppositeBranchCondition(NewCC);
4574           break;
4575         }
4576     } else if (IsSwapped) {
4577       // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
4578       // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
4579       // We swap the condition code and synthesize the new opcode.
4580       NewCC = getSwappedCondition(OldCC);
4581       if (NewCC == X86::COND_INVALID) return false;
4582     }
4583 
4584     if ((ShouldUpdateCC || IsSwapped) && NewCC != OldCC) {
4585       // Synthesize the new opcode.
4586       bool HasMemoryOperand = Instr.hasOneMemOperand();
4587       unsigned NewOpc;
4588       if (Instr.isBranch())
4589         NewOpc = GetCondBranchFromCond(NewCC);
4590       else if(OpcIsSET)
4591         NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
4592       else {
4593         unsigned DstReg = Instr.getOperand(0).getReg();
4594         NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
4595                                  HasMemoryOperand);
4596       }
4597 
4598       // Push the MachineInstr to OpsToUpdate.
4599       // If it is safe to remove CmpInstr, the condition code of these
4600       // instructions will be modified.
4601       OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
4602     }
4603     if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
4604       // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
4605       IsSafe = true;
4606       break;
4607     }
4608   }
4609 
4610   // If EFLAGS is not killed nor re-defined, we should check whether it is
4611   // live-out. If it is live-out, do not optimize.
4612   if ((IsCmpZero || IsSwapped) && !IsSafe) {
4613     MachineBasicBlock *MBB = CmpInstr->getParent();
4614     for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
4615              SE = MBB->succ_end(); SI != SE; ++SI)
4616       if ((*SI)->isLiveIn(X86::EFLAGS))
4617         return false;
4618   }
4619 
4620   // The instruction to be updated is either Sub or MI.
4621   Sub = IsCmpZero ? MI : Sub;
4622   // Move Movr0Inst to the appropriate place before Sub.
4623   if (Movr0Inst) {
4624     // Look backwards until we find a def that doesn't use the current EFLAGS.
4625     Def = Sub;
4626     MachineBasicBlock::reverse_iterator
4627       InsertI = MachineBasicBlock::reverse_iterator(++Def),
4628                 InsertE = Sub->getParent()->rend();
4629     for (; InsertI != InsertE; ++InsertI) {
4630       MachineInstr *Instr = &*InsertI;
4631       if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
4632           Instr->modifiesRegister(X86::EFLAGS, TRI)) {
4633         Sub->getParent()->remove(Movr0Inst);
4634         Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
4635                                    Movr0Inst);
4636         break;
4637       }
4638     }
4639     if (InsertI == InsertE)
4640       return false;
4641   }
4642 
4643   // Make sure Sub instruction defines EFLAGS and mark the def live.
4644   unsigned i = 0, e = Sub->getNumOperands();
4645   for (; i != e; ++i) {
4646     MachineOperand &MO = Sub->getOperand(i);
4647     if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
4648       MO.setIsDead(false);
4649       break;
4650     }
4651   }
4652   assert(i != e && "Unable to locate a def EFLAGS operand");
4653 
4654   CmpInstr->eraseFromParent();
4655 
4656   // Modify the condition code of instructions in OpsToUpdate.
4657   for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
4658     OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
4659   return true;
4660 }
4661 
4662 /// Try to remove the load by folding it to a register
4663 /// operand at the use. We fold the load instructions if load defines a virtual
4664 /// register, the virtual register is used once in the same BB, and the
4665 /// instructions in-between do not load or store, and have no side effects.
4666 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr *MI,
4667                                               const MachineRegisterInfo *MRI,
4668                                               unsigned &FoldAsLoadDefReg,
4669                                               MachineInstr *&DefMI) const {
4670   if (FoldAsLoadDefReg == 0)
4671     return nullptr;
4672   // To be conservative, if there exists another load, clear the load candidate.
4673   if (MI->mayLoad()) {
4674     FoldAsLoadDefReg = 0;
4675     return nullptr;
4676   }
4677 
4678   // Check whether we can move DefMI here.
4679   DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
4680   assert(DefMI);
4681   bool SawStore = false;
4682   if (!DefMI->isSafeToMove(nullptr, SawStore))
4683     return nullptr;
4684 
4685   // Collect information about virtual register operands of MI.
4686   unsigned SrcOperandId = 0;
4687   bool FoundSrcOperand = false;
4688   for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
4689     MachineOperand &MO = MI->getOperand(i);
4690     if (!MO.isReg())
4691       continue;
4692     unsigned Reg = MO.getReg();
4693     if (Reg != FoldAsLoadDefReg)
4694       continue;
4695     // Do not fold if we have a subreg use or a def or multiple uses.
4696     if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
4697       return nullptr;
4698 
4699     SrcOperandId = i;
4700     FoundSrcOperand = true;
4701   }
4702   if (!FoundSrcOperand)
4703     return nullptr;
4704 
4705   // Check whether we can fold the def into SrcOperandId.
4706   MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandId, DefMI);
4707   if (FoldMI) {
4708     FoldAsLoadDefReg = 0;
4709     return FoldMI;
4710   }
4711 
4712   return nullptr;
4713 }
4714 
4715 /// Expand a single-def pseudo instruction to a two-addr
4716 /// instruction with two undef reads of the register being defined.
4717 /// This is used for mapping:
4718 ///   %xmm4 = V_SET0
4719 /// to:
4720 ///   %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
4721 ///
4722 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
4723                              const MCInstrDesc &Desc) {
4724   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
4725   unsigned Reg = MIB->getOperand(0).getReg();
4726   MIB->setDesc(Desc);
4727 
4728   // MachineInstr::addOperand() will insert explicit operands before any
4729   // implicit operands.
4730   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4731   // But we don't trust that.
4732   assert(MIB->getOperand(1).getReg() == Reg &&
4733          MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
4734   return true;
4735 }
4736 
4737 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
4738 // code sequence is needed for other targets.
4739 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
4740                                  const TargetInstrInfo &TII) {
4741   MachineBasicBlock &MBB = *MIB->getParent();
4742   DebugLoc DL = MIB->getDebugLoc();
4743   unsigned Reg = MIB->getOperand(0).getReg();
4744   const GlobalValue *GV =
4745       cast<GlobalValue>((*MIB->memoperands_begin())->getValue());
4746   unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant;
4747   MachineMemOperand *MMO = MBB.getParent()->
4748       getMachineMemOperand(MachinePointerInfo::getGOT(), Flag, 8, 8);
4749   MachineBasicBlock::iterator I = MIB.getInstr();
4750 
4751   BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1)
4752       .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0)
4753       .addMemOperand(MMO);
4754   MIB->setDebugLoc(DL);
4755   MIB->setDesc(TII.get(X86::MOV64rm));
4756   MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0);
4757 }
4758 
4759 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
4760   bool HasAVX = Subtarget.hasAVX();
4761   MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
4762   switch (MI->getOpcode()) {
4763   case X86::MOV32r0:
4764     return Expand2AddrUndef(MIB, get(X86::XOR32rr));
4765   case X86::SETB_C8r:
4766     return Expand2AddrUndef(MIB, get(X86::SBB8rr));
4767   case X86::SETB_C16r:
4768     return Expand2AddrUndef(MIB, get(X86::SBB16rr));
4769   case X86::SETB_C32r:
4770     return Expand2AddrUndef(MIB, get(X86::SBB32rr));
4771   case X86::SETB_C64r:
4772     return Expand2AddrUndef(MIB, get(X86::SBB64rr));
4773   case X86::V_SET0:
4774   case X86::FsFLD0SS:
4775   case X86::FsFLD0SD:
4776     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
4777   case X86::AVX_SET0:
4778     assert(HasAVX && "AVX not supported");
4779     return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
4780   case X86::AVX512_512_SET0:
4781     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
4782   case X86::V_SETALLONES:
4783     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
4784   case X86::AVX2_SETALLONES:
4785     return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
4786   case X86::TEST8ri_NOREX:
4787     MI->setDesc(get(X86::TEST8ri));
4788     return true;
4789   case X86::KSET0B:
4790   case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
4791   case X86::KSET1B:
4792   case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
4793   case TargetOpcode::LOAD_STACK_GUARD:
4794     expandLoadStackGuard(MIB, *this);
4795     return true;
4796   }
4797   return false;
4798 }
4799 
4800 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs) {
4801   unsigned NumAddrOps = MOs.size();
4802   for (unsigned i = 0; i != NumAddrOps; ++i)
4803     MIB.addOperand(MOs[i]);
4804   if (NumAddrOps < 4) // FrameIndex only
4805     addOffset(MIB, 0);
4806 }
4807 
4808 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
4809                                      ArrayRef<MachineOperand> MOs,
4810                                      MachineBasicBlock::iterator InsertPt,
4811                                      MachineInstr *MI,
4812                                      const TargetInstrInfo &TII) {
4813   // Create the base instruction with the memory operand as the first part.
4814   // Omit the implicit operands, something BuildMI can't do.
4815   MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4816                                               MI->getDebugLoc(), true);
4817   MachineInstrBuilder MIB(MF, NewMI);
4818   addOperands(MIB, MOs);
4819 
4820   // Loop over the rest of the ri operands, converting them over.
4821   unsigned NumOps = MI->getDesc().getNumOperands()-2;
4822   for (unsigned i = 0; i != NumOps; ++i) {
4823     MachineOperand &MO = MI->getOperand(i+2);
4824     MIB.addOperand(MO);
4825   }
4826   for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
4827     MachineOperand &MO = MI->getOperand(i);
4828     MIB.addOperand(MO);
4829   }
4830 
4831   MachineBasicBlock *MBB = InsertPt->getParent();
4832   MBB->insert(InsertPt, NewMI);
4833 
4834   return MIB;
4835 }
4836 
4837 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode,
4838                               unsigned OpNo, ArrayRef<MachineOperand> MOs,
4839                               MachineBasicBlock::iterator InsertPt,
4840                               MachineInstr *MI, const TargetInstrInfo &TII) {
4841   // Omit the implicit operands, something BuildMI can't do.
4842   MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
4843                                               MI->getDebugLoc(), true);
4844   MachineInstrBuilder MIB(MF, NewMI);
4845 
4846   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4847     MachineOperand &MO = MI->getOperand(i);
4848     if (i == OpNo) {
4849       assert(MO.isReg() && "Expected to fold into reg operand!");
4850       addOperands(MIB, MOs);
4851     } else {
4852       MIB.addOperand(MO);
4853     }
4854   }
4855 
4856   MachineBasicBlock *MBB = InsertPt->getParent();
4857   MBB->insert(InsertPt, NewMI);
4858 
4859   return MIB;
4860 }
4861 
4862 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
4863                                 ArrayRef<MachineOperand> MOs,
4864                                 MachineBasicBlock::iterator InsertPt,
4865                                 MachineInstr *MI) {
4866   MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt,
4867                                     MI->getDebugLoc(), TII.get(Opcode));
4868   addOperands(MIB, MOs);
4869   return MIB.addImm(0);
4870 }
4871 
4872 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
4873     MachineFunction &MF, MachineInstr *MI, unsigned OpNum,
4874     ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt,
4875     unsigned Size, unsigned Align, bool AllowCommute) const {
4876   const DenseMap<unsigned,
4877                  std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
4878   bool isCallRegIndirect = Subtarget.callRegIndirect();
4879   bool isTwoAddrFold = false;
4880 
4881   // For CPUs that favor the register form of a call,
4882   // do not fold loads into calls.
4883   if (isCallRegIndirect &&
4884     (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r))
4885     return nullptr;
4886 
4887   unsigned NumOps = MI->getDesc().getNumOperands();
4888   bool isTwoAddr = NumOps > 1 &&
4889     MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4890 
4891   // FIXME: AsmPrinter doesn't know how to handle
4892   // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4893   if (MI->getOpcode() == X86::ADD32ri &&
4894       MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4895     return nullptr;
4896 
4897   MachineInstr *NewMI = nullptr;
4898   // Folding a memory location into the two-address part of a two-address
4899   // instruction is different than folding it other places.  It requires
4900   // replacing the *two* registers with the memory location.
4901   if (isTwoAddr && NumOps >= 2 && OpNum < 2 &&
4902       MI->getOperand(0).isReg() &&
4903       MI->getOperand(1).isReg() &&
4904       MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
4905     OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4906     isTwoAddrFold = true;
4907   } else if (OpNum == 0) {
4908     if (MI->getOpcode() == X86::MOV32r0) {
4909       NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI);
4910       if (NewMI)
4911         return NewMI;
4912     }
4913 
4914     OpcodeTablePtr = &RegOp2MemOpTable0;
4915   } else if (OpNum == 1) {
4916     OpcodeTablePtr = &RegOp2MemOpTable1;
4917   } else if (OpNum == 2) {
4918     OpcodeTablePtr = &RegOp2MemOpTable2;
4919   } else if (OpNum == 3) {
4920     OpcodeTablePtr = &RegOp2MemOpTable3;
4921   } else if (OpNum == 4) {
4922     OpcodeTablePtr = &RegOp2MemOpTable4;
4923   }
4924 
4925   // If table selected...
4926   if (OpcodeTablePtr) {
4927     // Find the Opcode to fuse
4928     DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4929       OpcodeTablePtr->find(MI->getOpcode());
4930     if (I != OpcodeTablePtr->end()) {
4931       unsigned Opcode = I->second.first;
4932       unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
4933       if (Align < MinAlign)
4934         return nullptr;
4935       bool NarrowToMOV32rm = false;
4936       if (Size) {
4937         unsigned RCSize = getRegClass(MI->getDesc(), OpNum, &RI, MF)->getSize();
4938         if (Size < RCSize) {
4939           // Check if it's safe to fold the load. If the size of the object is
4940           // narrower than the load width, then it's not.
4941           if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4942             return nullptr;
4943           // If this is a 64-bit load, but the spill slot is 32, then we can do
4944           // a 32-bit load which is implicitly zero-extended. This likely is
4945           // due to live interval analysis remat'ing a load from stack slot.
4946           if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4947             return nullptr;
4948           Opcode = X86::MOV32rm;
4949           NarrowToMOV32rm = true;
4950         }
4951       }
4952 
4953       if (isTwoAddrFold)
4954         NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this);
4955       else
4956         NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this);
4957 
4958       if (NarrowToMOV32rm) {
4959         // If this is the special case where we use a MOV32rm to load a 32-bit
4960         // value and zero-extend the top bits. Change the destination register
4961         // to a 32-bit one.
4962         unsigned DstReg = NewMI->getOperand(0).getReg();
4963         if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4964           NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit));
4965         else
4966           NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4967       }
4968       return NewMI;
4969     }
4970   }
4971 
4972   // If the instruction and target operand are commutable, commute the
4973   // instruction and try again.
4974   if (AllowCommute) {
4975     unsigned OriginalOpIdx = OpNum, CommuteOpIdx1, CommuteOpIdx2;
4976     if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) {
4977       bool HasDef = MI->getDesc().getNumDefs();
4978       unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
4979       unsigned Reg1 = MI->getOperand(CommuteOpIdx1).getReg();
4980       unsigned Reg2 = MI->getOperand(CommuteOpIdx2).getReg();
4981       bool Tied0 =
4982           0 == MI->getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO);
4983       bool Tied1 =
4984           0 == MI->getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO);
4985 
4986       // If either of the commutable operands are tied to the destination
4987       // then we can not commute + fold.
4988       if ((HasDef && Reg0 == Reg1 && Tied0) ||
4989           (HasDef && Reg0 == Reg2 && Tied1))
4990         return nullptr;
4991 
4992       if ((CommuteOpIdx1 == OriginalOpIdx) ||
4993           (CommuteOpIdx2 == OriginalOpIdx)) {
4994         MachineInstr *CommutedMI = commuteInstruction(MI, false);
4995         if (!CommutedMI) {
4996           // Unable to commute.
4997           return nullptr;
4998         }
4999         if (CommutedMI != MI) {
5000           // New instruction. We can't fold from this.
5001           CommutedMI->eraseFromParent();
5002           return nullptr;
5003         }
5004 
5005         // Attempt to fold with the commuted version of the instruction.
5006         unsigned CommuteOp =
5007             (CommuteOpIdx1 == OriginalOpIdx ? CommuteOpIdx2 : CommuteOpIdx1);
5008         NewMI =
5009             foldMemoryOperandImpl(MF, MI, CommuteOp, MOs, InsertPt, Size, Align,
5010                                   /*AllowCommute=*/false);
5011         if (NewMI)
5012           return NewMI;
5013 
5014         // Folding failed again - undo the commute before returning.
5015         MachineInstr *UncommutedMI = commuteInstruction(MI, false);
5016         if (!UncommutedMI) {
5017           // Unable to commute.
5018           return nullptr;
5019         }
5020         if (UncommutedMI != MI) {
5021           // New instruction. It doesn't need to be kept.
5022           UncommutedMI->eraseFromParent();
5023           return nullptr;
5024         }
5025 
5026         // Return here to prevent duplicate fuse failure report.
5027         return nullptr;
5028       }
5029     }
5030   }
5031 
5032   // No fusion
5033   if (PrintFailedFusing && !MI->isCopy())
5034     dbgs() << "We failed to fuse operand " << OpNum << " in " << *MI;
5035   return nullptr;
5036 }
5037 
5038 /// Return true for all instructions that only update
5039 /// the first 32 or 64-bits of the destination register and leave the rest
5040 /// unmodified. This can be used to avoid folding loads if the instructions
5041 /// only update part of the destination register, and the non-updated part is
5042 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
5043 /// instructions breaks the partial register dependency and it can improve
5044 /// performance. e.g.:
5045 ///
5046 ///   movss (%rdi), %xmm0
5047 ///   cvtss2sd %xmm0, %xmm0
5048 ///
5049 /// Instead of
5050 ///   cvtss2sd (%rdi), %xmm0
5051 ///
5052 /// FIXME: This should be turned into a TSFlags.
5053 ///
5054 static bool hasPartialRegUpdate(unsigned Opcode) {
5055   switch (Opcode) {
5056   case X86::CVTSI2SSrr:
5057   case X86::CVTSI2SSrm:
5058   case X86::CVTSI2SS64rr:
5059   case X86::CVTSI2SS64rm:
5060   case X86::CVTSI2SDrr:
5061   case X86::CVTSI2SDrm:
5062   case X86::CVTSI2SD64rr:
5063   case X86::CVTSI2SD64rm:
5064   case X86::CVTSD2SSrr:
5065   case X86::CVTSD2SSrm:
5066   case X86::Int_CVTSD2SSrr:
5067   case X86::Int_CVTSD2SSrm:
5068   case X86::CVTSS2SDrr:
5069   case X86::CVTSS2SDrm:
5070   case X86::Int_CVTSS2SDrr:
5071   case X86::Int_CVTSS2SDrm:
5072   case X86::RCPSSr:
5073   case X86::RCPSSm:
5074   case X86::RCPSSr_Int:
5075   case X86::RCPSSm_Int:
5076   case X86::ROUNDSDr:
5077   case X86::ROUNDSDm:
5078   case X86::ROUNDSDr_Int:
5079   case X86::ROUNDSSr:
5080   case X86::ROUNDSSm:
5081   case X86::ROUNDSSr_Int:
5082   case X86::RSQRTSSr:
5083   case X86::RSQRTSSm:
5084   case X86::RSQRTSSr_Int:
5085   case X86::RSQRTSSm_Int:
5086   case X86::SQRTSSr:
5087   case X86::SQRTSSm:
5088   case X86::SQRTSSr_Int:
5089   case X86::SQRTSSm_Int:
5090   case X86::SQRTSDr:
5091   case X86::SQRTSDm:
5092   case X86::SQRTSDr_Int:
5093   case X86::SQRTSDm_Int:
5094     return true;
5095   }
5096 
5097   return false;
5098 }
5099 
5100 /// Inform the ExeDepsFix pass how many idle
5101 /// instructions we would like before a partial register update.
5102 unsigned X86InstrInfo::
5103 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
5104                              const TargetRegisterInfo *TRI) const {
5105   if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
5106     return 0;
5107 
5108   // If MI is marked as reading Reg, the partial register update is wanted.
5109   const MachineOperand &MO = MI->getOperand(0);
5110   unsigned Reg = MO.getReg();
5111   if (TargetRegisterInfo::isVirtualRegister(Reg)) {
5112     if (MO.readsReg() || MI->readsVirtualRegister(Reg))
5113       return 0;
5114   } else {
5115     if (MI->readsRegister(Reg, TRI))
5116       return 0;
5117   }
5118 
5119   // If any of the preceding 16 instructions are reading Reg, insert a
5120   // dependency breaking instruction.  The magic number is based on a few
5121   // Nehalem experiments.
5122   return 16;
5123 }
5124 
5125 // Return true for any instruction the copies the high bits of the first source
5126 // operand into the unused high bits of the destination operand.
5127 static bool hasUndefRegUpdate(unsigned Opcode) {
5128   switch (Opcode) {
5129   case X86::VCVTSI2SSrr:
5130   case X86::VCVTSI2SSrm:
5131   case X86::Int_VCVTSI2SSrr:
5132   case X86::Int_VCVTSI2SSrm:
5133   case X86::VCVTSI2SS64rr:
5134   case X86::VCVTSI2SS64rm:
5135   case X86::Int_VCVTSI2SS64rr:
5136   case X86::Int_VCVTSI2SS64rm:
5137   case X86::VCVTSI2SDrr:
5138   case X86::VCVTSI2SDrm:
5139   case X86::Int_VCVTSI2SDrr:
5140   case X86::Int_VCVTSI2SDrm:
5141   case X86::VCVTSI2SD64rr:
5142   case X86::VCVTSI2SD64rm:
5143   case X86::Int_VCVTSI2SD64rr:
5144   case X86::Int_VCVTSI2SD64rm:
5145   case X86::VCVTSD2SSrr:
5146   case X86::VCVTSD2SSrm:
5147   case X86::Int_VCVTSD2SSrr:
5148   case X86::Int_VCVTSD2SSrm:
5149   case X86::VCVTSS2SDrr:
5150   case X86::VCVTSS2SDrm:
5151   case X86::Int_VCVTSS2SDrr:
5152   case X86::Int_VCVTSS2SDrm:
5153   case X86::VRCPSSr:
5154   case X86::VRCPSSm:
5155   case X86::VRCPSSm_Int:
5156   case X86::VROUNDSDr:
5157   case X86::VROUNDSDm:
5158   case X86::VROUNDSDr_Int:
5159   case X86::VROUNDSSr:
5160   case X86::VROUNDSSm:
5161   case X86::VROUNDSSr_Int:
5162   case X86::VRSQRTSSr:
5163   case X86::VRSQRTSSm:
5164   case X86::VRSQRTSSm_Int:
5165   case X86::VSQRTSSr:
5166   case X86::VSQRTSSm:
5167   case X86::VSQRTSSm_Int:
5168   case X86::VSQRTSDr:
5169   case X86::VSQRTSDm:
5170   case X86::VSQRTSDm_Int:
5171     // AVX-512
5172   case X86::VCVTSD2SSZrr:
5173   case X86::VCVTSD2SSZrm:
5174   case X86::VCVTSS2SDZrr:
5175   case X86::VCVTSS2SDZrm:
5176     return true;
5177   }
5178 
5179   return false;
5180 }
5181 
5182 /// Inform the ExeDepsFix pass how many idle instructions we would like before
5183 /// certain undef register reads.
5184 ///
5185 /// This catches the VCVTSI2SD family of instructions:
5186 ///
5187 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
5188 ///
5189 /// We should to be careful *not* to catch VXOR idioms which are presumably
5190 /// handled specially in the pipeline:
5191 ///
5192 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1
5193 ///
5194 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the
5195 /// high bits that are passed-through are not live.
5196 unsigned X86InstrInfo::
5197 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
5198                      const TargetRegisterInfo *TRI) const {
5199   if (!hasUndefRegUpdate(MI->getOpcode()))
5200     return 0;
5201 
5202   // Set the OpNum parameter to the first source operand.
5203   OpNum = 1;
5204 
5205   const MachineOperand &MO = MI->getOperand(OpNum);
5206   if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
5207     // Use the same magic number as getPartialRegUpdateClearance.
5208     return 16;
5209   }
5210   return 0;
5211 }
5212 
5213 void X86InstrInfo::
5214 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
5215                           const TargetRegisterInfo *TRI) const {
5216   unsigned Reg = MI->getOperand(OpNum).getReg();
5217   // If MI kills this register, the false dependence is already broken.
5218   if (MI->killsRegister(Reg, TRI))
5219     return;
5220   if (X86::VR128RegClass.contains(Reg)) {
5221     // These instructions are all floating point domain, so xorps is the best
5222     // choice.
5223     bool HasAVX = Subtarget.hasAVX();
5224     unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
5225     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
5226       .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
5227   } else if (X86::VR256RegClass.contains(Reg)) {
5228     // Use vxorps to clear the full ymm register.
5229     // It wants to read and write the xmm sub-register.
5230     unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
5231     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
5232       .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
5233       .addReg(Reg, RegState::ImplicitDefine);
5234   } else
5235     return;
5236   MI->addRegisterKilled(Reg, TRI, true);
5237 }
5238 
5239 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5240     MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops,
5241     MachineBasicBlock::iterator InsertPt, int FrameIndex) const {
5242   // Check switch flag
5243   if (NoFusing) return nullptr;
5244 
5245   // Unless optimizing for size, don't fold to avoid partial
5246   // register update stalls
5247   if (!MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize) &&
5248       hasPartialRegUpdate(MI->getOpcode()))
5249     return nullptr;
5250 
5251   const MachineFrameInfo *MFI = MF.getFrameInfo();
5252   unsigned Size = MFI->getObjectSize(FrameIndex);
5253   unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
5254   // If the function stack isn't realigned we don't want to fold instructions
5255   // that need increased alignment.
5256   if (!RI.needsStackRealignment(MF))
5257     Alignment =
5258         std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment());
5259   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5260     unsigned NewOpc = 0;
5261     unsigned RCSize = 0;
5262     switch (MI->getOpcode()) {
5263     default: return nullptr;
5264     case X86::TEST8rr:  NewOpc = X86::CMP8ri; RCSize = 1; break;
5265     case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
5266     case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
5267     case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
5268     }
5269     // Check if it's safe to fold the load. If the size of the object is
5270     // narrower than the load width, then it's not.
5271     if (Size < RCSize)
5272       return nullptr;
5273     // Change to CMPXXri r, 0 first.
5274     MI->setDesc(get(NewOpc));
5275     MI->getOperand(1).ChangeToImmediate(0);
5276   } else if (Ops.size() != 1)
5277     return nullptr;
5278 
5279   return foldMemoryOperandImpl(MF, MI, Ops[0],
5280                                MachineOperand::CreateFI(FrameIndex), InsertPt,
5281                                Size, Alignment, /*AllowCommute=*/true);
5282 }
5283 
5284 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI
5285 /// because the latter uses contents that wouldn't be defined in the folded
5286 /// version.  For instance, this transformation isn't legal:
5287 ///   movss (%rdi), %xmm0
5288 ///   addps %xmm0, %xmm0
5289 /// ->
5290 ///   addps (%rdi), %xmm0
5291 ///
5292 /// But this one is:
5293 ///   movss (%rdi), %xmm0
5294 ///   addss %xmm0, %xmm0
5295 /// ->
5296 ///   addss (%rdi), %xmm0
5297 ///
5298 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI,
5299                                              const MachineInstr &UserMI,
5300                                              const MachineFunction &MF) {
5301   unsigned Opc = LoadMI.getOpcode();
5302   unsigned UserOpc = UserMI.getOpcode();
5303   unsigned RegSize =
5304       MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg())->getSize();
5305 
5306   if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm) && RegSize > 4) {
5307     // These instructions only load 32 bits, we can't fold them if the
5308     // destination register is wider than 32 bits (4 bytes), and its user
5309     // instruction isn't scalar (SS).
5310     switch (UserOpc) {
5311     case X86::ADDSSrr_Int: case X86::VADDSSrr_Int:
5312     case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int:
5313     case X86::MULSSrr_Int: case X86::VMULSSrr_Int:
5314     case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int:
5315       return false;
5316     default:
5317       return true;
5318     }
5319   }
5320 
5321   if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm) && RegSize > 8) {
5322     // These instructions only load 64 bits, we can't fold them if the
5323     // destination register is wider than 64 bits (8 bytes), and its user
5324     // instruction isn't scalar (SD).
5325     switch (UserOpc) {
5326     case X86::ADDSDrr_Int: case X86::VADDSDrr_Int:
5327     case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int:
5328     case X86::MULSDrr_Int: case X86::VMULSDrr_Int:
5329     case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int:
5330       return false;
5331     default:
5332       return true;
5333     }
5334   }
5335 
5336   return false;
5337 }
5338 
5339 MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
5340     MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops,
5341     MachineBasicBlock::iterator InsertPt, MachineInstr *LoadMI) const {
5342   // If loading from a FrameIndex, fold directly from the FrameIndex.
5343   unsigned NumOps = LoadMI->getDesc().getNumOperands();
5344   int FrameIndex;
5345   if (isLoadFromStackSlot(LoadMI, FrameIndex)) {
5346     if (isNonFoldablePartialRegisterLoad(*LoadMI, *MI, MF))
5347       return nullptr;
5348     return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex);
5349   }
5350 
5351   // Check switch flag
5352   if (NoFusing) return nullptr;
5353 
5354   // Unless optimizing for size, don't fold to avoid partial
5355   // register update stalls
5356   if (!MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize) &&
5357       hasPartialRegUpdate(MI->getOpcode()))
5358     return nullptr;
5359 
5360   // Determine the alignment of the load.
5361   unsigned Alignment = 0;
5362   if (LoadMI->hasOneMemOperand())
5363     Alignment = (*LoadMI->memoperands_begin())->getAlignment();
5364   else
5365     switch (LoadMI->getOpcode()) {
5366     case X86::AVX2_SETALLONES:
5367     case X86::AVX_SET0:
5368       Alignment = 32;
5369       break;
5370     case X86::V_SET0:
5371     case X86::V_SETALLONES:
5372       Alignment = 16;
5373       break;
5374     case X86::FsFLD0SD:
5375       Alignment = 8;
5376       break;
5377     case X86::FsFLD0SS:
5378       Alignment = 4;
5379       break;
5380     default:
5381       return nullptr;
5382     }
5383   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5384     unsigned NewOpc = 0;
5385     switch (MI->getOpcode()) {
5386     default: return nullptr;
5387     case X86::TEST8rr:  NewOpc = X86::CMP8ri; break;
5388     case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
5389     case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
5390     case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
5391     }
5392     // Change to CMPXXri r, 0 first.
5393     MI->setDesc(get(NewOpc));
5394     MI->getOperand(1).ChangeToImmediate(0);
5395   } else if (Ops.size() != 1)
5396     return nullptr;
5397 
5398   // Make sure the subregisters match.
5399   // Otherwise we risk changing the size of the load.
5400   if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
5401     return nullptr;
5402 
5403   SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
5404   switch (LoadMI->getOpcode()) {
5405   case X86::V_SET0:
5406   case X86::V_SETALLONES:
5407   case X86::AVX2_SETALLONES:
5408   case X86::AVX_SET0:
5409   case X86::FsFLD0SD:
5410   case X86::FsFLD0SS: {
5411     // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
5412     // Create a constant-pool entry and operands to load from it.
5413 
5414     // Medium and large mode can't fold loads this way.
5415     if (MF.getTarget().getCodeModel() != CodeModel::Small &&
5416         MF.getTarget().getCodeModel() != CodeModel::Kernel)
5417       return nullptr;
5418 
5419     // x86-32 PIC requires a PIC base register for constant pools.
5420     unsigned PICBase = 0;
5421     if (MF.getTarget().getRelocationModel() == Reloc::PIC_) {
5422       if (Subtarget.is64Bit())
5423         PICBase = X86::RIP;
5424       else
5425         // FIXME: PICBase = getGlobalBaseReg(&MF);
5426         // This doesn't work for several reasons.
5427         // 1. GlobalBaseReg may have been spilled.
5428         // 2. It may not be live at MI.
5429         return nullptr;
5430     }
5431 
5432     // Create a constant-pool entry.
5433     MachineConstantPool &MCP = *MF.getConstantPool();
5434     Type *Ty;
5435     unsigned Opc = LoadMI->getOpcode();
5436     if (Opc == X86::FsFLD0SS)
5437       Ty = Type::getFloatTy(MF.getFunction()->getContext());
5438     else if (Opc == X86::FsFLD0SD)
5439       Ty = Type::getDoubleTy(MF.getFunction()->getContext());
5440     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
5441       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
5442     else
5443       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
5444 
5445     bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
5446     const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
5447                                     Constant::getNullValue(Ty);
5448     unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
5449 
5450     // Create operands to load from the constant pool entry.
5451     MOs.push_back(MachineOperand::CreateReg(PICBase, false));
5452     MOs.push_back(MachineOperand::CreateImm(1));
5453     MOs.push_back(MachineOperand::CreateReg(0, false));
5454     MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
5455     MOs.push_back(MachineOperand::CreateReg(0, false));
5456     break;
5457   }
5458   default: {
5459     if (isNonFoldablePartialRegisterLoad(*LoadMI, *MI, MF))
5460       return nullptr;
5461 
5462     // Folding a normal load. Just copy the load's address operands.
5463     MOs.append(LoadMI->operands_begin() + NumOps - X86::AddrNumOperands,
5464                LoadMI->operands_begin() + NumOps);
5465     break;
5466   }
5467   }
5468   return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt,
5469                                /*Size=*/0, Alignment, /*AllowCommute=*/true);
5470 }
5471 
5472 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
5473                                         ArrayRef<unsigned> Ops) const {
5474   // Check switch flag
5475   if (NoFusing) return 0;
5476 
5477   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
5478     switch (MI->getOpcode()) {
5479     default: return false;
5480     case X86::TEST8rr:
5481     case X86::TEST16rr:
5482     case X86::TEST32rr:
5483     case X86::TEST64rr:
5484       return true;
5485     case X86::ADD32ri:
5486       // FIXME: AsmPrinter doesn't know how to handle
5487       // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
5488       if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
5489         return false;
5490       break;
5491     }
5492   }
5493 
5494   if (Ops.size() != 1)
5495     return false;
5496 
5497   unsigned OpNum = Ops[0];
5498   unsigned Opc = MI->getOpcode();
5499   unsigned NumOps = MI->getDesc().getNumOperands();
5500   bool isTwoAddr = NumOps > 1 &&
5501     MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
5502 
5503   // Folding a memory location into the two-address part of a two-address
5504   // instruction is different than folding it other places.  It requires
5505   // replacing the *two* registers with the memory location.
5506   const DenseMap<unsigned,
5507                  std::pair<unsigned,unsigned> > *OpcodeTablePtr = nullptr;
5508   if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
5509     OpcodeTablePtr = &RegOp2MemOpTable2Addr;
5510   } else if (OpNum == 0) {
5511     if (Opc == X86::MOV32r0)
5512       return true;
5513 
5514     OpcodeTablePtr = &RegOp2MemOpTable0;
5515   } else if (OpNum == 1) {
5516     OpcodeTablePtr = &RegOp2MemOpTable1;
5517   } else if (OpNum == 2) {
5518     OpcodeTablePtr = &RegOp2MemOpTable2;
5519   } else if (OpNum == 3) {
5520     OpcodeTablePtr = &RegOp2MemOpTable3;
5521   }
5522 
5523   if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
5524     return true;
5525   return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
5526 }
5527 
5528 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
5529                                 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
5530                                 SmallVectorImpl<MachineInstr*> &NewMIs) const {
5531   DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5532     MemOp2RegOpTable.find(MI->getOpcode());
5533   if (I == MemOp2RegOpTable.end())
5534     return false;
5535   unsigned Opc = I->second.first;
5536   unsigned Index = I->second.second & TB_INDEX_MASK;
5537   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5538   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5539   if (UnfoldLoad && !FoldedLoad)
5540     return false;
5541   UnfoldLoad &= FoldedLoad;
5542   if (UnfoldStore && !FoldedStore)
5543     return false;
5544   UnfoldStore &= FoldedStore;
5545 
5546   const MCInstrDesc &MCID = get(Opc);
5547   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5548   if (!MI->hasOneMemOperand() &&
5549       RC == &X86::VR128RegClass &&
5550       !Subtarget.isUnalignedMemAccessFast())
5551     // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
5552     // conservatively assume the address is unaligned. That's bad for
5553     // performance.
5554     return false;
5555   SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
5556   SmallVector<MachineOperand,2> BeforeOps;
5557   SmallVector<MachineOperand,2> AfterOps;
5558   SmallVector<MachineOperand,4> ImpOps;
5559   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
5560     MachineOperand &Op = MI->getOperand(i);
5561     if (i >= Index && i < Index + X86::AddrNumOperands)
5562       AddrOps.push_back(Op);
5563     else if (Op.isReg() && Op.isImplicit())
5564       ImpOps.push_back(Op);
5565     else if (i < Index)
5566       BeforeOps.push_back(Op);
5567     else if (i > Index)
5568       AfterOps.push_back(Op);
5569   }
5570 
5571   // Emit the load instruction.
5572   if (UnfoldLoad) {
5573     std::pair<MachineInstr::mmo_iterator,
5574               MachineInstr::mmo_iterator> MMOs =
5575       MF.extractLoadMemRefs(MI->memoperands_begin(),
5576                             MI->memoperands_end());
5577     loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
5578     if (UnfoldStore) {
5579       // Address operands cannot be marked isKill.
5580       for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
5581         MachineOperand &MO = NewMIs[0]->getOperand(i);
5582         if (MO.isReg())
5583           MO.setIsKill(false);
5584       }
5585     }
5586   }
5587 
5588   // Emit the data processing instruction.
5589   MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
5590   MachineInstrBuilder MIB(MF, DataMI);
5591 
5592   if (FoldedStore)
5593     MIB.addReg(Reg, RegState::Define);
5594   for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
5595     MIB.addOperand(BeforeOps[i]);
5596   if (FoldedLoad)
5597     MIB.addReg(Reg);
5598   for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
5599     MIB.addOperand(AfterOps[i]);
5600   for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
5601     MachineOperand &MO = ImpOps[i];
5602     MIB.addReg(MO.getReg(),
5603                getDefRegState(MO.isDef()) |
5604                RegState::Implicit |
5605                getKillRegState(MO.isKill()) |
5606                getDeadRegState(MO.isDead()) |
5607                getUndefRegState(MO.isUndef()));
5608   }
5609   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
5610   switch (DataMI->getOpcode()) {
5611   default: break;
5612   case X86::CMP64ri32:
5613   case X86::CMP64ri8:
5614   case X86::CMP32ri:
5615   case X86::CMP32ri8:
5616   case X86::CMP16ri:
5617   case X86::CMP16ri8:
5618   case X86::CMP8ri: {
5619     MachineOperand &MO0 = DataMI->getOperand(0);
5620     MachineOperand &MO1 = DataMI->getOperand(1);
5621     if (MO1.getImm() == 0) {
5622       unsigned NewOpc;
5623       switch (DataMI->getOpcode()) {
5624       default: llvm_unreachable("Unreachable!");
5625       case X86::CMP64ri8:
5626       case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
5627       case X86::CMP32ri8:
5628       case X86::CMP32ri:   NewOpc = X86::TEST32rr; break;
5629       case X86::CMP16ri8:
5630       case X86::CMP16ri:   NewOpc = X86::TEST16rr; break;
5631       case X86::CMP8ri:    NewOpc = X86::TEST8rr; break;
5632       }
5633       DataMI->setDesc(get(NewOpc));
5634       MO1.ChangeToRegister(MO0.getReg(), false);
5635     }
5636   }
5637   }
5638   NewMIs.push_back(DataMI);
5639 
5640   // Emit the store instruction.
5641   if (UnfoldStore) {
5642     const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
5643     std::pair<MachineInstr::mmo_iterator,
5644               MachineInstr::mmo_iterator> MMOs =
5645       MF.extractStoreMemRefs(MI->memoperands_begin(),
5646                              MI->memoperands_end());
5647     storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
5648   }
5649 
5650   return true;
5651 }
5652 
5653 bool
5654 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
5655                                   SmallVectorImpl<SDNode*> &NewNodes) const {
5656   if (!N->isMachineOpcode())
5657     return false;
5658 
5659   DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5660     MemOp2RegOpTable.find(N->getMachineOpcode());
5661   if (I == MemOp2RegOpTable.end())
5662     return false;
5663   unsigned Opc = I->second.first;
5664   unsigned Index = I->second.second & TB_INDEX_MASK;
5665   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5666   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5667   const MCInstrDesc &MCID = get(Opc);
5668   MachineFunction &MF = DAG.getMachineFunction();
5669   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
5670   unsigned NumDefs = MCID.NumDefs;
5671   std::vector<SDValue> AddrOps;
5672   std::vector<SDValue> BeforeOps;
5673   std::vector<SDValue> AfterOps;
5674   SDLoc dl(N);
5675   unsigned NumOps = N->getNumOperands();
5676   for (unsigned i = 0; i != NumOps-1; ++i) {
5677     SDValue Op = N->getOperand(i);
5678     if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
5679       AddrOps.push_back(Op);
5680     else if (i < Index-NumDefs)
5681       BeforeOps.push_back(Op);
5682     else if (i > Index-NumDefs)
5683       AfterOps.push_back(Op);
5684   }
5685   SDValue Chain = N->getOperand(NumOps-1);
5686   AddrOps.push_back(Chain);
5687 
5688   // Emit the load instruction.
5689   SDNode *Load = nullptr;
5690   if (FoldedLoad) {
5691     EVT VT = *RC->vt_begin();
5692     std::pair<MachineInstr::mmo_iterator,
5693               MachineInstr::mmo_iterator> MMOs =
5694       MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
5695                             cast<MachineSDNode>(N)->memoperands_end());
5696     if (!(*MMOs.first) &&
5697         RC == &X86::VR128RegClass &&
5698         !Subtarget.isUnalignedMemAccessFast())
5699       // Do not introduce a slow unaligned load.
5700       return false;
5701     unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
5702     bool isAligned = (*MMOs.first) &&
5703                      (*MMOs.first)->getAlignment() >= Alignment;
5704     Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
5705                               VT, MVT::Other, AddrOps);
5706     NewNodes.push_back(Load);
5707 
5708     // Preserve memory reference information.
5709     cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
5710   }
5711 
5712   // Emit the data processing instruction.
5713   std::vector<EVT> VTs;
5714   const TargetRegisterClass *DstRC = nullptr;
5715   if (MCID.getNumDefs() > 0) {
5716     DstRC = getRegClass(MCID, 0, &RI, MF);
5717     VTs.push_back(*DstRC->vt_begin());
5718   }
5719   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
5720     EVT VT = N->getValueType(i);
5721     if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
5722       VTs.push_back(VT);
5723   }
5724   if (Load)
5725     BeforeOps.push_back(SDValue(Load, 0));
5726   BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end());
5727   SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
5728   NewNodes.push_back(NewNode);
5729 
5730   // Emit the store instruction.
5731   if (FoldedStore) {
5732     AddrOps.pop_back();
5733     AddrOps.push_back(SDValue(NewNode, 0));
5734     AddrOps.push_back(Chain);
5735     std::pair<MachineInstr::mmo_iterator,
5736               MachineInstr::mmo_iterator> MMOs =
5737       MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
5738                              cast<MachineSDNode>(N)->memoperands_end());
5739     if (!(*MMOs.first) &&
5740         RC == &X86::VR128RegClass &&
5741         !Subtarget.isUnalignedMemAccessFast())
5742       // Do not introduce a slow unaligned store.
5743       return false;
5744     unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
5745     bool isAligned = (*MMOs.first) &&
5746                      (*MMOs.first)->getAlignment() >= Alignment;
5747     SDNode *Store =
5748         DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget),
5749                            dl, MVT::Other, AddrOps);
5750     NewNodes.push_back(Store);
5751 
5752     // Preserve memory reference information.
5753     cast<MachineSDNode>(Store)->setMemRefs(MMOs.first, MMOs.second);
5754   }
5755 
5756   return true;
5757 }
5758 
5759 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
5760                                       bool UnfoldLoad, bool UnfoldStore,
5761                                       unsigned *LoadRegIndex) const {
5762   DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
5763     MemOp2RegOpTable.find(Opc);
5764   if (I == MemOp2RegOpTable.end())
5765     return 0;
5766   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
5767   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
5768   if (UnfoldLoad && !FoldedLoad)
5769     return 0;
5770   if (UnfoldStore && !FoldedStore)
5771     return 0;
5772   if (LoadRegIndex)
5773     *LoadRegIndex = I->second.second & TB_INDEX_MASK;
5774   return I->second.first;
5775 }
5776 
5777 bool
5778 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
5779                                      int64_t &Offset1, int64_t &Offset2) const {
5780   if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
5781     return false;
5782   unsigned Opc1 = Load1->getMachineOpcode();
5783   unsigned Opc2 = Load2->getMachineOpcode();
5784   switch (Opc1) {
5785   default: return false;
5786   case X86::MOV8rm:
5787   case X86::MOV16rm:
5788   case X86::MOV32rm:
5789   case X86::MOV64rm:
5790   case X86::LD_Fp32m:
5791   case X86::LD_Fp64m:
5792   case X86::LD_Fp80m:
5793   case X86::MOVSSrm:
5794   case X86::MOVSDrm:
5795   case X86::MMX_MOVD64rm:
5796   case X86::MMX_MOVQ64rm:
5797   case X86::FsMOVAPSrm:
5798   case X86::FsMOVAPDrm:
5799   case X86::MOVAPSrm:
5800   case X86::MOVUPSrm:
5801   case X86::MOVAPDrm:
5802   case X86::MOVDQArm:
5803   case X86::MOVDQUrm:
5804   // AVX load instructions
5805   case X86::VMOVSSrm:
5806   case X86::VMOVSDrm:
5807   case X86::FsVMOVAPSrm:
5808   case X86::FsVMOVAPDrm:
5809   case X86::VMOVAPSrm:
5810   case X86::VMOVUPSrm:
5811   case X86::VMOVAPDrm:
5812   case X86::VMOVDQArm:
5813   case X86::VMOVDQUrm:
5814   case X86::VMOVAPSYrm:
5815   case X86::VMOVUPSYrm:
5816   case X86::VMOVAPDYrm:
5817   case X86::VMOVDQAYrm:
5818   case X86::VMOVDQUYrm:
5819     break;
5820   }
5821   switch (Opc2) {
5822   default: return false;
5823   case X86::MOV8rm:
5824   case X86::MOV16rm:
5825   case X86::MOV32rm:
5826   case X86::MOV64rm:
5827   case X86::LD_Fp32m:
5828   case X86::LD_Fp64m:
5829   case X86::LD_Fp80m:
5830   case X86::MOVSSrm:
5831   case X86::MOVSDrm:
5832   case X86::MMX_MOVD64rm:
5833   case X86::MMX_MOVQ64rm:
5834   case X86::FsMOVAPSrm:
5835   case X86::FsMOVAPDrm:
5836   case X86::MOVAPSrm:
5837   case X86::MOVUPSrm:
5838   case X86::MOVAPDrm:
5839   case X86::MOVDQArm:
5840   case X86::MOVDQUrm:
5841   // AVX load instructions
5842   case X86::VMOVSSrm:
5843   case X86::VMOVSDrm:
5844   case X86::FsVMOVAPSrm:
5845   case X86::FsVMOVAPDrm:
5846   case X86::VMOVAPSrm:
5847   case X86::VMOVUPSrm:
5848   case X86::VMOVAPDrm:
5849   case X86::VMOVDQArm:
5850   case X86::VMOVDQUrm:
5851   case X86::VMOVAPSYrm:
5852   case X86::VMOVUPSYrm:
5853   case X86::VMOVAPDYrm:
5854   case X86::VMOVDQAYrm:
5855   case X86::VMOVDQUYrm:
5856     break;
5857   }
5858 
5859   // Check if chain operands and base addresses match.
5860   if (Load1->getOperand(0) != Load2->getOperand(0) ||
5861       Load1->getOperand(5) != Load2->getOperand(5))
5862     return false;
5863   // Segment operands should match as well.
5864   if (Load1->getOperand(4) != Load2->getOperand(4))
5865     return false;
5866   // Scale should be 1, Index should be Reg0.
5867   if (Load1->getOperand(1) == Load2->getOperand(1) &&
5868       Load1->getOperand(2) == Load2->getOperand(2)) {
5869     if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
5870       return false;
5871 
5872     // Now let's examine the displacements.
5873     if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
5874         isa<ConstantSDNode>(Load2->getOperand(3))) {
5875       Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
5876       Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
5877       return true;
5878     }
5879   }
5880   return false;
5881 }
5882 
5883 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
5884                                            int64_t Offset1, int64_t Offset2,
5885                                            unsigned NumLoads) const {
5886   assert(Offset2 > Offset1);
5887   if ((Offset2 - Offset1) / 8 > 64)
5888     return false;
5889 
5890   unsigned Opc1 = Load1->getMachineOpcode();
5891   unsigned Opc2 = Load2->getMachineOpcode();
5892   if (Opc1 != Opc2)
5893     return false;  // FIXME: overly conservative?
5894 
5895   switch (Opc1) {
5896   default: break;
5897   case X86::LD_Fp32m:
5898   case X86::LD_Fp64m:
5899   case X86::LD_Fp80m:
5900   case X86::MMX_MOVD64rm:
5901   case X86::MMX_MOVQ64rm:
5902     return false;
5903   }
5904 
5905   EVT VT = Load1->getValueType(0);
5906   switch (VT.getSimpleVT().SimpleTy) {
5907   default:
5908     // XMM registers. In 64-bit mode we can be a bit more aggressive since we
5909     // have 16 of them to play with.
5910     if (Subtarget.is64Bit()) {
5911       if (NumLoads >= 3)
5912         return false;
5913     } else if (NumLoads) {
5914       return false;
5915     }
5916     break;
5917   case MVT::i8:
5918   case MVT::i16:
5919   case MVT::i32:
5920   case MVT::i64:
5921   case MVT::f32:
5922   case MVT::f64:
5923     if (NumLoads)
5924       return false;
5925     break;
5926   }
5927 
5928   return true;
5929 }
5930 
5931 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
5932                                           MachineInstr *Second) const {
5933   // Check if this processor supports macro-fusion. Since this is a minor
5934   // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
5935   // proxy for SandyBridge+.
5936   if (!Subtarget.hasAVX())
5937     return false;
5938 
5939   enum {
5940     FuseTest,
5941     FuseCmp,
5942     FuseInc
5943   } FuseKind;
5944 
5945   switch(Second->getOpcode()) {
5946   default:
5947     return false;
5948   case X86::JE_1:
5949   case X86::JNE_1:
5950   case X86::JL_1:
5951   case X86::JLE_1:
5952   case X86::JG_1:
5953   case X86::JGE_1:
5954     FuseKind = FuseInc;
5955     break;
5956   case X86::JB_1:
5957   case X86::JBE_1:
5958   case X86::JA_1:
5959   case X86::JAE_1:
5960     FuseKind = FuseCmp;
5961     break;
5962   case X86::JS_1:
5963   case X86::JNS_1:
5964   case X86::JP_1:
5965   case X86::JNP_1:
5966   case X86::JO_1:
5967   case X86::JNO_1:
5968     FuseKind = FuseTest;
5969     break;
5970   }
5971   switch (First->getOpcode()) {
5972   default:
5973     return false;
5974   case X86::TEST8rr:
5975   case X86::TEST16rr:
5976   case X86::TEST32rr:
5977   case X86::TEST64rr:
5978   case X86::TEST8ri:
5979   case X86::TEST16ri:
5980   case X86::TEST32ri:
5981   case X86::TEST32i32:
5982   case X86::TEST64i32:
5983   case X86::TEST64ri32:
5984   case X86::TEST8rm:
5985   case X86::TEST16rm:
5986   case X86::TEST32rm:
5987   case X86::TEST64rm:
5988   case X86::TEST8ri_NOREX:
5989   case X86::AND16i16:
5990   case X86::AND16ri:
5991   case X86::AND16ri8:
5992   case X86::AND16rm:
5993   case X86::AND16rr:
5994   case X86::AND32i32:
5995   case X86::AND32ri:
5996   case X86::AND32ri8:
5997   case X86::AND32rm:
5998   case X86::AND32rr:
5999   case X86::AND64i32:
6000   case X86::AND64ri32:
6001   case X86::AND64ri8:
6002   case X86::AND64rm:
6003   case X86::AND64rr:
6004   case X86::AND8i8:
6005   case X86::AND8ri:
6006   case X86::AND8rm:
6007   case X86::AND8rr:
6008     return true;
6009   case X86::CMP16i16:
6010   case X86::CMP16ri:
6011   case X86::CMP16ri8:
6012   case X86::CMP16rm:
6013   case X86::CMP16rr:
6014   case X86::CMP32i32:
6015   case X86::CMP32ri:
6016   case X86::CMP32ri8:
6017   case X86::CMP32rm:
6018   case X86::CMP32rr:
6019   case X86::CMP64i32:
6020   case X86::CMP64ri32:
6021   case X86::CMP64ri8:
6022   case X86::CMP64rm:
6023   case X86::CMP64rr:
6024   case X86::CMP8i8:
6025   case X86::CMP8ri:
6026   case X86::CMP8rm:
6027   case X86::CMP8rr:
6028   case X86::ADD16i16:
6029   case X86::ADD16ri:
6030   case X86::ADD16ri8:
6031   case X86::ADD16ri8_DB:
6032   case X86::ADD16ri_DB:
6033   case X86::ADD16rm:
6034   case X86::ADD16rr:
6035   case X86::ADD16rr_DB:
6036   case X86::ADD32i32:
6037   case X86::ADD32ri:
6038   case X86::ADD32ri8:
6039   case X86::ADD32ri8_DB:
6040   case X86::ADD32ri_DB:
6041   case X86::ADD32rm:
6042   case X86::ADD32rr:
6043   case X86::ADD32rr_DB:
6044   case X86::ADD64i32:
6045   case X86::ADD64ri32:
6046   case X86::ADD64ri32_DB:
6047   case X86::ADD64ri8:
6048   case X86::ADD64ri8_DB:
6049   case X86::ADD64rm:
6050   case X86::ADD64rr:
6051   case X86::ADD64rr_DB:
6052   case X86::ADD8i8:
6053   case X86::ADD8mi:
6054   case X86::ADD8mr:
6055   case X86::ADD8ri:
6056   case X86::ADD8rm:
6057   case X86::ADD8rr:
6058   case X86::SUB16i16:
6059   case X86::SUB16ri:
6060   case X86::SUB16ri8:
6061   case X86::SUB16rm:
6062   case X86::SUB16rr:
6063   case X86::SUB32i32:
6064   case X86::SUB32ri:
6065   case X86::SUB32ri8:
6066   case X86::SUB32rm:
6067   case X86::SUB32rr:
6068   case X86::SUB64i32:
6069   case X86::SUB64ri32:
6070   case X86::SUB64ri8:
6071   case X86::SUB64rm:
6072   case X86::SUB64rr:
6073   case X86::SUB8i8:
6074   case X86::SUB8ri:
6075   case X86::SUB8rm:
6076   case X86::SUB8rr:
6077     return FuseKind == FuseCmp || FuseKind == FuseInc;
6078   case X86::INC16r:
6079   case X86::INC32r:
6080   case X86::INC64r:
6081   case X86::INC8r:
6082   case X86::DEC16r:
6083   case X86::DEC32r:
6084   case X86::DEC64r:
6085   case X86::DEC8r:
6086     return FuseKind == FuseInc;
6087   }
6088 }
6089 
6090 bool X86InstrInfo::
6091 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
6092   assert(Cond.size() == 1 && "Invalid X86 branch condition!");
6093   X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
6094   if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
6095     return true;
6096   Cond[0].setImm(GetOppositeBranchCondition(CC));
6097   return false;
6098 }
6099 
6100 bool X86InstrInfo::
6101 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
6102   // FIXME: Return false for x87 stack register classes for now. We can't
6103   // allow any loads of these registers before FpGet_ST0_80.
6104   return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
6105            RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
6106 }
6107 
6108 /// Return a virtual register initialized with the
6109 /// the global base register value. Output instructions required to
6110 /// initialize the register in the function entry block, if necessary.
6111 ///
6112 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
6113 ///
6114 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
6115   assert(!Subtarget.is64Bit() &&
6116          "X86-64 PIC uses RIP relative addressing");
6117 
6118   X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
6119   unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
6120   if (GlobalBaseReg != 0)
6121     return GlobalBaseReg;
6122 
6123   // Create the register. The code to initialize it is inserted
6124   // later, by the CGBR pass (below).
6125   MachineRegisterInfo &RegInfo = MF->getRegInfo();
6126   GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
6127   X86FI->setGlobalBaseReg(GlobalBaseReg);
6128   return GlobalBaseReg;
6129 }
6130 
6131 // These are the replaceable SSE instructions. Some of these have Int variants
6132 // that we don't include here. We don't want to replace instructions selected
6133 // by intrinsics.
6134 static const uint16_t ReplaceableInstrs[][3] = {
6135   //PackedSingle     PackedDouble    PackedInt
6136   { X86::MOVAPSmr,   X86::MOVAPDmr,  X86::MOVDQAmr  },
6137   { X86::MOVAPSrm,   X86::MOVAPDrm,  X86::MOVDQArm  },
6138   { X86::MOVAPSrr,   X86::MOVAPDrr,  X86::MOVDQArr  },
6139   { X86::MOVUPSmr,   X86::MOVUPDmr,  X86::MOVDQUmr  },
6140   { X86::MOVUPSrm,   X86::MOVUPDrm,  X86::MOVDQUrm  },
6141   { X86::MOVLPSmr,   X86::MOVLPDmr,  X86::MOVPQI2QImr  },
6142   { X86::MOVNTPSmr,  X86::MOVNTPDmr, X86::MOVNTDQmr },
6143   { X86::ANDNPSrm,   X86::ANDNPDrm,  X86::PANDNrm   },
6144   { X86::ANDNPSrr,   X86::ANDNPDrr,  X86::PANDNrr   },
6145   { X86::ANDPSrm,    X86::ANDPDrm,   X86::PANDrm    },
6146   { X86::ANDPSrr,    X86::ANDPDrr,   X86::PANDrr    },
6147   { X86::ORPSrm,     X86::ORPDrm,    X86::PORrm     },
6148   { X86::ORPSrr,     X86::ORPDrr,    X86::PORrr     },
6149   { X86::XORPSrm,    X86::XORPDrm,   X86::PXORrm    },
6150   { X86::XORPSrr,    X86::XORPDrr,   X86::PXORrr    },
6151   // AVX 128-bit support
6152   { X86::VMOVAPSmr,  X86::VMOVAPDmr,  X86::VMOVDQAmr  },
6153   { X86::VMOVAPSrm,  X86::VMOVAPDrm,  X86::VMOVDQArm  },
6154   { X86::VMOVAPSrr,  X86::VMOVAPDrr,  X86::VMOVDQArr  },
6155   { X86::VMOVUPSmr,  X86::VMOVUPDmr,  X86::VMOVDQUmr  },
6156   { X86::VMOVUPSrm,  X86::VMOVUPDrm,  X86::VMOVDQUrm  },
6157   { X86::VMOVLPSmr,  X86::VMOVLPDmr,  X86::VMOVPQI2QImr  },
6158   { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
6159   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNrm   },
6160   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNrr   },
6161   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDrm    },
6162   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDrr    },
6163   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORrm     },
6164   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORrr     },
6165   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORrm    },
6166   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORrr    },
6167   // AVX 256-bit support
6168   { X86::VMOVAPSYmr,   X86::VMOVAPDYmr,   X86::VMOVDQAYmr  },
6169   { X86::VMOVAPSYrm,   X86::VMOVAPDYrm,   X86::VMOVDQAYrm  },
6170   { X86::VMOVAPSYrr,   X86::VMOVAPDYrr,   X86::VMOVDQAYrr  },
6171   { X86::VMOVUPSYmr,   X86::VMOVUPDYmr,   X86::VMOVDQUYmr  },
6172   { X86::VMOVUPSYrm,   X86::VMOVUPDYrm,   X86::VMOVDQUYrm  },
6173   { X86::VMOVNTPSYmr,  X86::VMOVNTPDYmr,  X86::VMOVNTDQYmr }
6174 };
6175 
6176 static const uint16_t ReplaceableInstrsAVX2[][3] = {
6177   //PackedSingle       PackedDouble       PackedInt
6178   { X86::VANDNPSYrm,   X86::VANDNPDYrm,   X86::VPANDNYrm   },
6179   { X86::VANDNPSYrr,   X86::VANDNPDYrr,   X86::VPANDNYrr   },
6180   { X86::VANDPSYrm,    X86::VANDPDYrm,    X86::VPANDYrm    },
6181   { X86::VANDPSYrr,    X86::VANDPDYrr,    X86::VPANDYrr    },
6182   { X86::VORPSYrm,     X86::VORPDYrm,     X86::VPORYrm     },
6183   { X86::VORPSYrr,     X86::VORPDYrr,     X86::VPORYrr     },
6184   { X86::VXORPSYrm,    X86::VXORPDYrm,    X86::VPXORYrm    },
6185   { X86::VXORPSYrr,    X86::VXORPDYrr,    X86::VPXORYrr    },
6186   { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
6187   { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
6188   { X86::VINSERTF128rm,  X86::VINSERTF128rm,  X86::VINSERTI128rm },
6189   { X86::VINSERTF128rr,  X86::VINSERTF128rr,  X86::VINSERTI128rr },
6190   { X86::VPERM2F128rm,   X86::VPERM2F128rm,   X86::VPERM2I128rm },
6191   { X86::VPERM2F128rr,   X86::VPERM2F128rr,   X86::VPERM2I128rr },
6192   { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
6193   { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
6194   { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
6195   { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
6196   { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
6197   { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
6198 };
6199 
6200 // FIXME: Some shuffle and unpack instructions have equivalents in different
6201 // domains, but they require a bit more work than just switching opcodes.
6202 
6203 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
6204   for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
6205     if (ReplaceableInstrs[i][domain-1] == opcode)
6206       return ReplaceableInstrs[i];
6207   return nullptr;
6208 }
6209 
6210 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
6211   for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
6212     if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
6213       return ReplaceableInstrsAVX2[i];
6214   return nullptr;
6215 }
6216 
6217 std::pair<uint16_t, uint16_t>
6218 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
6219   uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6220   bool hasAVX2 = Subtarget.hasAVX2();
6221   uint16_t validDomains = 0;
6222   if (domain && lookup(MI->getOpcode(), domain))
6223     validDomains = 0xe;
6224   else if (domain && lookupAVX2(MI->getOpcode(), domain))
6225     validDomains = hasAVX2 ? 0xe : 0x6;
6226   return std::make_pair(domain, validDomains);
6227 }
6228 
6229 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
6230   assert(Domain>0 && Domain<4 && "Invalid execution domain");
6231   uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
6232   assert(dom && "Not an SSE instruction");
6233   const uint16_t *table = lookup(MI->getOpcode(), dom);
6234   if (!table) { // try the other table
6235     assert((Subtarget.hasAVX2() || Domain < 3) &&
6236            "256-bit vector operations only available in AVX2");
6237     table = lookupAVX2(MI->getOpcode(), dom);
6238   }
6239   assert(table && "Cannot change domain");
6240   MI->setDesc(get(table[Domain-1]));
6241 }
6242 
6243 /// Return the noop instruction to use for a noop.
6244 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
6245   NopInst.setOpcode(X86::NOOP);
6246 }
6247 
6248 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
6249 // In particular, getJumpInstrTableEntryBound must always return an upper bound
6250 // on the encoding lengths of the instructions generated by
6251 // getUnconditionalBranch and getTrap.
6252 void X86InstrInfo::getUnconditionalBranch(
6253     MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const {
6254   Branch.setOpcode(X86::JMP_1);
6255   Branch.addOperand(MCOperand::createExpr(BranchTarget));
6256 }
6257 
6258 // This code must remain in sync with getJumpInstrTableEntryBound in this class!
6259 // In particular, getJumpInstrTableEntryBound must always return an upper bound
6260 // on the encoding lengths of the instructions generated by
6261 // getUnconditionalBranch and getTrap.
6262 void X86InstrInfo::getTrap(MCInst &MI) const {
6263   MI.setOpcode(X86::TRAP);
6264 }
6265 
6266 // See getTrap and getUnconditionalBranch for conditions on the value returned
6267 // by this function.
6268 unsigned X86InstrInfo::getJumpInstrTableEntryBound() const {
6269   // 5 bytes suffice: JMP_4 Symbol@PLT is uses 1 byte (E9) for the JMP_4 and 4
6270   // bytes for the symbol offset. And TRAP is ud2, which is two bytes (0F 0B).
6271   return 5;
6272 }
6273 
6274 bool X86InstrInfo::isHighLatencyDef(int opc) const {
6275   switch (opc) {
6276   default: return false;
6277   case X86::DIVSDrm:
6278   case X86::DIVSDrm_Int:
6279   case X86::DIVSDrr:
6280   case X86::DIVSDrr_Int:
6281   case X86::DIVSSrm:
6282   case X86::DIVSSrm_Int:
6283   case X86::DIVSSrr:
6284   case X86::DIVSSrr_Int:
6285   case X86::SQRTPDm:
6286   case X86::SQRTPDr:
6287   case X86::SQRTPSm:
6288   case X86::SQRTPSr:
6289   case X86::SQRTSDm:
6290   case X86::SQRTSDm_Int:
6291   case X86::SQRTSDr:
6292   case X86::SQRTSDr_Int:
6293   case X86::SQRTSSm:
6294   case X86::SQRTSSm_Int:
6295   case X86::SQRTSSr:
6296   case X86::SQRTSSr_Int:
6297   // AVX instructions with high latency
6298   case X86::VDIVSDrm:
6299   case X86::VDIVSDrm_Int:
6300   case X86::VDIVSDrr:
6301   case X86::VDIVSDrr_Int:
6302   case X86::VDIVSSrm:
6303   case X86::VDIVSSrm_Int:
6304   case X86::VDIVSSrr:
6305   case X86::VDIVSSrr_Int:
6306   case X86::VSQRTPDm:
6307   case X86::VSQRTPDr:
6308   case X86::VSQRTPSm:
6309   case X86::VSQRTPSr:
6310   case X86::VSQRTSDm:
6311   case X86::VSQRTSDm_Int:
6312   case X86::VSQRTSDr:
6313   case X86::VSQRTSSm:
6314   case X86::VSQRTSSm_Int:
6315   case X86::VSQRTSSr:
6316   case X86::VSQRTPDZm:
6317   case X86::VSQRTPDZr:
6318   case X86::VSQRTPSZm:
6319   case X86::VSQRTPSZr:
6320   case X86::VSQRTSDZm:
6321   case X86::VSQRTSDZm_Int:
6322   case X86::VSQRTSDZr:
6323   case X86::VSQRTSSZm_Int:
6324   case X86::VSQRTSSZr:
6325   case X86::VSQRTSSZm:
6326   case X86::VDIVSDZrm:
6327   case X86::VDIVSDZrr:
6328   case X86::VDIVSSZrm:
6329   case X86::VDIVSSZrr:
6330 
6331   case X86::VGATHERQPSZrm:
6332   case X86::VGATHERQPDZrm:
6333   case X86::VGATHERDPDZrm:
6334   case X86::VGATHERDPSZrm:
6335   case X86::VPGATHERQDZrm:
6336   case X86::VPGATHERQQZrm:
6337   case X86::VPGATHERDDZrm:
6338   case X86::VPGATHERDQZrm:
6339   case X86::VSCATTERQPDZmr:
6340   case X86::VSCATTERQPSZmr:
6341   case X86::VSCATTERDPDZmr:
6342   case X86::VSCATTERDPSZmr:
6343   case X86::VPSCATTERQDZmr:
6344   case X86::VPSCATTERQQZmr:
6345   case X86::VPSCATTERDDZmr:
6346   case X86::VPSCATTERDQZmr:
6347     return true;
6348   }
6349 }
6350 
6351 bool X86InstrInfo::
6352 hasHighOperandLatency(const TargetSchedModel &SchedModel,
6353                       const MachineRegisterInfo *MRI,
6354                       const MachineInstr *DefMI, unsigned DefIdx,
6355                       const MachineInstr *UseMI, unsigned UseIdx) const {
6356   return isHighLatencyDef(DefMI->getOpcode());
6357 }
6358 
6359 static bool hasVirtualRegDefsInBasicBlock(const MachineInstr &Inst,
6360                                           const MachineBasicBlock *MBB) {
6361   assert(Inst.getNumOperands() == 3 && "Reassociation needs binary operators");
6362   const MachineOperand &Op1 = Inst.getOperand(1);
6363   const MachineOperand &Op2 = Inst.getOperand(2);
6364   const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
6365 
6366   // We need virtual register definitions.
6367   MachineInstr *MI1 = nullptr;
6368   MachineInstr *MI2 = nullptr;
6369   if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
6370     MI1 = MRI.getUniqueVRegDef(Op1.getReg());
6371   if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
6372     MI2 = MRI.getUniqueVRegDef(Op2.getReg());
6373 
6374   // And they need to be in the trace (otherwise, they won't have a depth).
6375   if (MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB)
6376     return true;
6377 
6378   return false;
6379 }
6380 
6381 static bool hasReassocSibling(const MachineInstr &Inst, bool &Commuted) {
6382   const MachineBasicBlock *MBB = Inst.getParent();
6383   const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
6384   MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
6385   MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
6386   unsigned AssocOpcode = Inst.getOpcode();
6387 
6388   // If only one operand has the same opcode and it's the second source operand,
6389   // the operands must be commuted.
6390   Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
6391   if (Commuted)
6392     std::swap(MI1, MI2);
6393 
6394   // 1. The previous instruction must be the same type as Inst.
6395   // 2. The previous instruction must have virtual register definitions for its
6396   //    operands in the same basic block as Inst.
6397   // 3. The previous instruction's result must only be used by Inst.
6398   if (MI1->getOpcode() == AssocOpcode &&
6399       hasVirtualRegDefsInBasicBlock(*MI1, MBB) &&
6400       MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()))
6401     return true;
6402 
6403   return false;
6404 }
6405 
6406 // TODO: There are many more machine instruction opcodes to match:
6407 //       1. Other data types (integer, vectors)
6408 //       2. Other math / logic operations (and, or)
6409 static bool isAssociativeAndCommutative(unsigned Opcode) {
6410   switch (Opcode) {
6411   case X86::ADDSDrr:
6412   case X86::ADDSSrr:
6413   case X86::VADDSDrr:
6414   case X86::VADDSSrr:
6415   case X86::MULSDrr:
6416   case X86::MULSSrr:
6417   case X86::VMULSDrr:
6418   case X86::VMULSSrr:
6419     return true;
6420   default:
6421     return false;
6422   }
6423 }
6424 
6425 /// Return true if the input instruction is part of a chain of dependent ops
6426 /// that are suitable for reassociation, otherwise return false.
6427 /// If the instruction's operands must be commuted to have a previous
6428 /// instruction of the same type define the first source operand, Commuted will
6429 /// be set to true.
6430 static bool isReassocCandidate(const MachineInstr &Inst, bool &Commuted) {
6431   // 1. The operation must be associative and commutative.
6432   // 2. The instruction must have virtual register definitions for its
6433   //    operands in the same basic block.
6434   // 3. The instruction must have a reassociable sibling.
6435   if (isAssociativeAndCommutative(Inst.getOpcode()) &&
6436       hasVirtualRegDefsInBasicBlock(Inst, Inst.getParent()) &&
6437       hasReassocSibling(Inst, Commuted))
6438     return true;
6439 
6440   return false;
6441 }
6442 
6443 // FIXME: This has the potential to be expensive (compile time) while not
6444 // improving the code at all. Some ways to limit the overhead:
6445 // 1. Track successful transforms; bail out if hit rate gets too low.
6446 // 2. Only enable at -O3 or some other non-default optimization level.
6447 // 3. Pre-screen pattern candidates here: if an operand of the previous
6448 //    instruction is known to not increase the critical path, then don't match
6449 //    that pattern.
6450 bool X86InstrInfo::getMachineCombinerPatterns(MachineInstr &Root,
6451         SmallVectorImpl<MachineCombinerPattern::MC_PATTERN> &Patterns) const {
6452   if (!Root.getParent()->getParent()->getTarget().Options.UnsafeFPMath)
6453     return false;
6454 
6455   // TODO: There is nothing x86-specific here except the instruction type.
6456   // This logic could be hoisted into the machine combiner pass itself.
6457 
6458   // Look for this reassociation pattern:
6459   //   B = A op X (Prev)
6460   //   C = B op Y (Root)
6461 
6462   bool Commute;
6463   if (isReassocCandidate(Root, Commute)) {
6464     // We found a sequence of instructions that may be suitable for a
6465     // reassociation of operands to increase ILP. Specify each commutation
6466     // possibility for the Prev instruction in the sequence and let the
6467     // machine combiner decide if changing the operands is worthwhile.
6468     if (Commute) {
6469       Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_YB);
6470       Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_YB);
6471     } else {
6472       Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_BY);
6473       Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_BY);
6474     }
6475     return true;
6476   }
6477 
6478   return false;
6479 }
6480 
6481 /// Attempt the following reassociation to reduce critical path length:
6482 ///   B = A op X (Prev)
6483 ///   C = B op Y (Root)
6484 ///   ===>
6485 ///   B = X op Y
6486 ///   C = A op B
6487 static void reassociateOps(MachineInstr &Root, MachineInstr &Prev,
6488                            MachineCombinerPattern::MC_PATTERN Pattern,
6489                            SmallVectorImpl<MachineInstr *> &InsInstrs,
6490                            SmallVectorImpl<MachineInstr *> &DelInstrs,
6491                            DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) {
6492   MachineFunction *MF = Root.getParent()->getParent();
6493   MachineRegisterInfo &MRI = MF->getRegInfo();
6494   const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
6495   const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
6496   const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
6497 
6498   // This array encodes the operand index for each parameter because the
6499   // operands may be commuted. Each row corresponds to a pattern value,
6500   // and each column specifies the index of A, B, X, Y.
6501   unsigned OpIdx[4][4] = {
6502     { 1, 1, 2, 2 },
6503     { 1, 2, 2, 1 },
6504     { 2, 1, 1, 2 },
6505     { 2, 2, 1, 1 }
6506   };
6507 
6508   MachineOperand &OpA = Prev.getOperand(OpIdx[Pattern][0]);
6509   MachineOperand &OpB = Root.getOperand(OpIdx[Pattern][1]);
6510   MachineOperand &OpX = Prev.getOperand(OpIdx[Pattern][2]);
6511   MachineOperand &OpY = Root.getOperand(OpIdx[Pattern][3]);
6512   MachineOperand &OpC = Root.getOperand(0);
6513 
6514   unsigned RegA = OpA.getReg();
6515   unsigned RegB = OpB.getReg();
6516   unsigned RegX = OpX.getReg();
6517   unsigned RegY = OpY.getReg();
6518   unsigned RegC = OpC.getReg();
6519 
6520   if (TargetRegisterInfo::isVirtualRegister(RegA))
6521     MRI.constrainRegClass(RegA, RC);
6522   if (TargetRegisterInfo::isVirtualRegister(RegB))
6523     MRI.constrainRegClass(RegB, RC);
6524   if (TargetRegisterInfo::isVirtualRegister(RegX))
6525     MRI.constrainRegClass(RegX, RC);
6526   if (TargetRegisterInfo::isVirtualRegister(RegY))
6527     MRI.constrainRegClass(RegY, RC);
6528   if (TargetRegisterInfo::isVirtualRegister(RegC))
6529     MRI.constrainRegClass(RegC, RC);
6530 
6531   // Create a new virtual register for the result of (X op Y) instead of
6532   // recycling RegB because the MachineCombiner's computation of the critical
6533   // path requires a new register definition rather than an existing one.
6534   unsigned NewVR = MRI.createVirtualRegister(RC);
6535   InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
6536 
6537   unsigned Opcode = Root.getOpcode();
6538   bool KillA = OpA.isKill();
6539   bool KillX = OpX.isKill();
6540   bool KillY = OpY.isKill();
6541 
6542   // Create new instructions for insertion.
6543   MachineInstrBuilder MIB1 =
6544     BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
6545       .addReg(RegX, getKillRegState(KillX))
6546       .addReg(RegY, getKillRegState(KillY));
6547   InsInstrs.push_back(MIB1);
6548 
6549   MachineInstrBuilder MIB2 =
6550     BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
6551       .addReg(RegA, getKillRegState(KillA))
6552       .addReg(NewVR, getKillRegState(true));
6553   InsInstrs.push_back(MIB2);
6554 
6555   // Record old instructions for deletion.
6556   DelInstrs.push_back(&Prev);
6557   DelInstrs.push_back(&Root);
6558 }
6559 
6560 void X86InstrInfo::genAlternativeCodeSequence(
6561     MachineInstr &Root,
6562     MachineCombinerPattern::MC_PATTERN Pattern,
6563     SmallVectorImpl<MachineInstr *> &InsInstrs,
6564     SmallVectorImpl<MachineInstr *> &DelInstrs,
6565     DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
6566   MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo();
6567 
6568   // Select the previous instruction in the sequence based on the input pattern.
6569   MachineInstr *Prev = nullptr;
6570   switch (Pattern) {
6571     case MachineCombinerPattern::MC_REASSOC_AX_BY:
6572     case MachineCombinerPattern::MC_REASSOC_XA_BY:
6573       Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
6574       break;
6575     case MachineCombinerPattern::MC_REASSOC_AX_YB:
6576     case MachineCombinerPattern::MC_REASSOC_XA_YB:
6577       Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
6578   }
6579   assert(Prev && "Unknown pattern for machine combiner");
6580 
6581   reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
6582   return;
6583 }
6584 
6585 namespace {
6586   /// Create Global Base Reg pass. This initializes the PIC
6587   /// global base register for x86-32.
6588   struct CGBR : public MachineFunctionPass {
6589     static char ID;
6590     CGBR() : MachineFunctionPass(ID) {}
6591 
6592     bool runOnMachineFunction(MachineFunction &MF) override {
6593       const X86TargetMachine *TM =
6594         static_cast<const X86TargetMachine *>(&MF.getTarget());
6595       const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>();
6596 
6597       // Don't do anything if this is 64-bit as 64-bit PIC
6598       // uses RIP relative addressing.
6599       if (STI.is64Bit())
6600         return false;
6601 
6602       // Only emit a global base reg in PIC mode.
6603       if (TM->getRelocationModel() != Reloc::PIC_)
6604         return false;
6605 
6606       X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
6607       unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
6608 
6609       // If we didn't need a GlobalBaseReg, don't insert code.
6610       if (GlobalBaseReg == 0)
6611         return false;
6612 
6613       // Insert the set of GlobalBaseReg into the first MBB of the function
6614       MachineBasicBlock &FirstMBB = MF.front();
6615       MachineBasicBlock::iterator MBBI = FirstMBB.begin();
6616       DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
6617       MachineRegisterInfo &RegInfo = MF.getRegInfo();
6618       const X86InstrInfo *TII = STI.getInstrInfo();
6619 
6620       unsigned PC;
6621       if (STI.isPICStyleGOT())
6622         PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
6623       else
6624         PC = GlobalBaseReg;
6625 
6626       // Operand of MovePCtoStack is completely ignored by asm printer. It's
6627       // only used in JIT code emission as displacement to pc.
6628       BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
6629 
6630       // If we're using vanilla 'GOT' PIC style, we should use relative addressing
6631       // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
6632       if (STI.isPICStyleGOT()) {
6633         // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
6634         BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
6635           .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
6636                                         X86II::MO_GOT_ABSOLUTE_ADDRESS);
6637       }
6638 
6639       return true;
6640     }
6641 
6642     const char *getPassName() const override {
6643       return "X86 PIC Global Base Reg Initialization";
6644     }
6645 
6646     void getAnalysisUsage(AnalysisUsage &AU) const override {
6647       AU.setPreservesCFG();
6648       MachineFunctionPass::getAnalysisUsage(AU);
6649     }
6650   };
6651 }
6652 
6653 char CGBR::ID = 0;
6654 FunctionPass*
6655 llvm::createX86GlobalBaseRegPass() { return new CGBR(); }
6656 
6657 namespace {
6658   struct LDTLSCleanup : public MachineFunctionPass {
6659     static char ID;
6660     LDTLSCleanup() : MachineFunctionPass(ID) {}
6661 
6662     bool runOnMachineFunction(MachineFunction &MF) override {
6663       X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
6664       if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
6665         // No point folding accesses if there isn't at least two.
6666         return false;
6667       }
6668 
6669       MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
6670       return VisitNode(DT->getRootNode(), 0);
6671     }
6672 
6673     // Visit the dominator subtree rooted at Node in pre-order.
6674     // If TLSBaseAddrReg is non-null, then use that to replace any
6675     // TLS_base_addr instructions. Otherwise, create the register
6676     // when the first such instruction is seen, and then use it
6677     // as we encounter more instructions.
6678     bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
6679       MachineBasicBlock *BB = Node->getBlock();
6680       bool Changed = false;
6681 
6682       // Traverse the current block.
6683       for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
6684            ++I) {
6685         switch (I->getOpcode()) {
6686           case X86::TLS_base_addr32:
6687           case X86::TLS_base_addr64:
6688             if (TLSBaseAddrReg)
6689               I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
6690             else
6691               I = SetRegister(I, &TLSBaseAddrReg);
6692             Changed = true;
6693             break;
6694           default:
6695             break;
6696         }
6697       }
6698 
6699       // Visit the children of this block in the dominator tree.
6700       for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
6701            I != E; ++I) {
6702         Changed |= VisitNode(*I, TLSBaseAddrReg);
6703       }
6704 
6705       return Changed;
6706     }
6707 
6708     // Replace the TLS_base_addr instruction I with a copy from
6709     // TLSBaseAddrReg, returning the new instruction.
6710     MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
6711                                          unsigned TLSBaseAddrReg) {
6712       MachineFunction *MF = I->getParent()->getParent();
6713       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
6714       const bool is64Bit = STI.is64Bit();
6715       const X86InstrInfo *TII = STI.getInstrInfo();
6716 
6717       // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
6718       MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
6719                                    TII->get(TargetOpcode::COPY),
6720                                    is64Bit ? X86::RAX : X86::EAX)
6721                                    .addReg(TLSBaseAddrReg);
6722 
6723       // Erase the TLS_base_addr instruction.
6724       I->eraseFromParent();
6725 
6726       return Copy;
6727     }
6728 
6729     // Create a virtal register in *TLSBaseAddrReg, and populate it by
6730     // inserting a copy instruction after I. Returns the new instruction.
6731     MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
6732       MachineFunction *MF = I->getParent()->getParent();
6733       const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>();
6734       const bool is64Bit = STI.is64Bit();
6735       const X86InstrInfo *TII = STI.getInstrInfo();
6736 
6737       // Create a virtual register for the TLS base address.
6738       MachineRegisterInfo &RegInfo = MF->getRegInfo();
6739       *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
6740                                                       ? &X86::GR64RegClass
6741                                                       : &X86::GR32RegClass);
6742 
6743       // Insert a copy from RAX/EAX to TLSBaseAddrReg.
6744       MachineInstr *Next = I->getNextNode();
6745       MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
6746                                    TII->get(TargetOpcode::COPY),
6747                                    *TLSBaseAddrReg)
6748                                    .addReg(is64Bit ? X86::RAX : X86::EAX);
6749 
6750       return Copy;
6751     }
6752 
6753     const char *getPassName() const override {
6754       return "Local Dynamic TLS Access Clean-up";
6755     }
6756 
6757     void getAnalysisUsage(AnalysisUsage &AU) const override {
6758       AU.setPreservesCFG();
6759       AU.addRequired<MachineDominatorTree>();
6760       MachineFunctionPass::getAnalysisUsage(AU);
6761     }
6762   };
6763 }
6764 
6765 char LDTLSCleanup::ID = 0;
6766 FunctionPass*
6767 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
6768