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/IR/DerivedTypes.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/MC/MCAsmInfo.h"
30 #include "llvm/MC/MCInst.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetOptions.h"
36 #include <limits>
37 
38 #define GET_INSTRINFO_CTOR
39 #include "X86GenInstrInfo.inc"
40 
41 using namespace llvm;
42 
43 static cl::opt<bool>
44 NoFusing("disable-spill-fusing",
45          cl::desc("Disable fusing of spill code into instructions"));
46 static cl::opt<bool>
47 PrintFailedFusing("print-failed-fuse-candidates",
48                   cl::desc("Print instructions that the allocator wants to"
49                            " fuse, but the X86 backend currently can't"),
50                   cl::Hidden);
51 static cl::opt<bool>
52 ReMatPICStubLoad("remat-pic-stub-load",
53                  cl::desc("Re-materialize load from stub in PIC mode"),
54                  cl::init(false), cl::Hidden);
55 
56 enum {
57   // Select which memory operand is being unfolded.
58   // (stored in bits 0 - 3)
59   TB_INDEX_0    = 0,
60   TB_INDEX_1    = 1,
61   TB_INDEX_2    = 2,
62   TB_INDEX_3    = 3,
63   TB_INDEX_MASK = 0xf,
64 
65   // Do not insert the reverse map (MemOp -> RegOp) into the table.
66   // This may be needed because there is a many -> one mapping.
67   TB_NO_REVERSE   = 1 << 4,
68 
69   // Do not insert the forward map (RegOp -> MemOp) into the table.
70   // This is needed for Native Client, which prohibits branch
71   // instructions from using a memory operand.
72   TB_NO_FORWARD   = 1 << 5,
73 
74   TB_FOLDED_LOAD  = 1 << 6,
75   TB_FOLDED_STORE = 1 << 7,
76 
77   // Minimum alignment required for load/store.
78   // Used for RegOp->MemOp conversion.
79   // (stored in bits 8 - 15)
80   TB_ALIGN_SHIFT = 8,
81   TB_ALIGN_NONE  =    0 << TB_ALIGN_SHIFT,
82   TB_ALIGN_16    =   16 << TB_ALIGN_SHIFT,
83   TB_ALIGN_32    =   32 << TB_ALIGN_SHIFT,
84   TB_ALIGN_64    =   64 << TB_ALIGN_SHIFT,
85   TB_ALIGN_MASK  = 0xff << TB_ALIGN_SHIFT
86 };
87 
88 struct X86OpTblEntry {
89   uint16_t RegOp;
90   uint16_t MemOp;
91   uint16_t Flags;
92 };
93 
94 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
95   : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
96                      ? X86::ADJCALLSTACKDOWN64
97                      : X86::ADJCALLSTACKDOWN32),
98                     (tm.getSubtarget<X86Subtarget>().is64Bit()
99                      ? X86::ADJCALLSTACKUP64
100                      : X86::ADJCALLSTACKUP32)),
101     TM(tm), RI(tm) {
102 
103   static const X86OpTblEntry OpTbl2Addr[] = {
104     { X86::ADC32ri,     X86::ADC32mi,    0 },
105     { X86::ADC32ri8,    X86::ADC32mi8,   0 },
106     { X86::ADC32rr,     X86::ADC32mr,    0 },
107     { X86::ADC64ri32,   X86::ADC64mi32,  0 },
108     { X86::ADC64ri8,    X86::ADC64mi8,   0 },
109     { X86::ADC64rr,     X86::ADC64mr,    0 },
110     { X86::ADD16ri,     X86::ADD16mi,    0 },
111     { X86::ADD16ri8,    X86::ADD16mi8,   0 },
112     { X86::ADD16ri_DB,  X86::ADD16mi,    TB_NO_REVERSE },
113     { X86::ADD16ri8_DB, X86::ADD16mi8,   TB_NO_REVERSE },
114     { X86::ADD16rr,     X86::ADD16mr,    0 },
115     { X86::ADD16rr_DB,  X86::ADD16mr,    TB_NO_REVERSE },
116     { X86::ADD32ri,     X86::ADD32mi,    0 },
117     { X86::ADD32ri8,    X86::ADD32mi8,   0 },
118     { X86::ADD32ri_DB,  X86::ADD32mi,    TB_NO_REVERSE },
119     { X86::ADD32ri8_DB, X86::ADD32mi8,   TB_NO_REVERSE },
120     { X86::ADD32rr,     X86::ADD32mr,    0 },
121     { X86::ADD32rr_DB,  X86::ADD32mr,    TB_NO_REVERSE },
122     { X86::ADD64ri32,   X86::ADD64mi32,  0 },
123     { X86::ADD64ri8,    X86::ADD64mi8,   0 },
124     { X86::ADD64ri32_DB,X86::ADD64mi32,  TB_NO_REVERSE },
125     { X86::ADD64ri8_DB, X86::ADD64mi8,   TB_NO_REVERSE },
126     { X86::ADD64rr,     X86::ADD64mr,    0 },
127     { X86::ADD64rr_DB,  X86::ADD64mr,    TB_NO_REVERSE },
128     { X86::ADD8ri,      X86::ADD8mi,     0 },
129     { X86::ADD8rr,      X86::ADD8mr,     0 },
130     { X86::AND16ri,     X86::AND16mi,    0 },
131     { X86::AND16ri8,    X86::AND16mi8,   0 },
132     { X86::AND16rr,     X86::AND16mr,    0 },
133     { X86::AND32ri,     X86::AND32mi,    0 },
134     { X86::AND32ri8,    X86::AND32mi8,   0 },
135     { X86::AND32rr,     X86::AND32mr,    0 },
136     { X86::AND64ri32,   X86::AND64mi32,  0 },
137     { X86::AND64ri8,    X86::AND64mi8,   0 },
138     { X86::AND64rr,     X86::AND64mr,    0 },
139     { X86::AND8ri,      X86::AND8mi,     0 },
140     { X86::AND8rr,      X86::AND8mr,     0 },
141     { X86::DEC16r,      X86::DEC16m,     0 },
142     { X86::DEC32r,      X86::DEC32m,     0 },
143     { X86::DEC64_16r,   X86::DEC64_16m,  0 },
144     { X86::DEC64_32r,   X86::DEC64_32m,  0 },
145     { X86::DEC64r,      X86::DEC64m,     0 },
146     { X86::DEC8r,       X86::DEC8m,      0 },
147     { X86::INC16r,      X86::INC16m,     0 },
148     { X86::INC32r,      X86::INC32m,     0 },
149     { X86::INC64_16r,   X86::INC64_16m,  0 },
150     { X86::INC64_32r,   X86::INC64_32m,  0 },
151     { X86::INC64r,      X86::INC64m,     0 },
152     { X86::INC8r,       X86::INC8m,      0 },
153     { X86::NEG16r,      X86::NEG16m,     0 },
154     { X86::NEG32r,      X86::NEG32m,     0 },
155     { X86::NEG64r,      X86::NEG64m,     0 },
156     { X86::NEG8r,       X86::NEG8m,      0 },
157     { X86::NOT16r,      X86::NOT16m,     0 },
158     { X86::NOT32r,      X86::NOT32m,     0 },
159     { X86::NOT64r,      X86::NOT64m,     0 },
160     { X86::NOT8r,       X86::NOT8m,      0 },
161     { X86::OR16ri,      X86::OR16mi,     0 },
162     { X86::OR16ri8,     X86::OR16mi8,    0 },
163     { X86::OR16rr,      X86::OR16mr,     0 },
164     { X86::OR32ri,      X86::OR32mi,     0 },
165     { X86::OR32ri8,     X86::OR32mi8,    0 },
166     { X86::OR32rr,      X86::OR32mr,     0 },
167     { X86::OR64ri32,    X86::OR64mi32,   0 },
168     { X86::OR64ri8,     X86::OR64mi8,    0 },
169     { X86::OR64rr,      X86::OR64mr,     0 },
170     { X86::OR8ri,       X86::OR8mi,      0 },
171     { X86::OR8rr,       X86::OR8mr,      0 },
172     { X86::ROL16r1,     X86::ROL16m1,    0 },
173     { X86::ROL16rCL,    X86::ROL16mCL,   0 },
174     { X86::ROL16ri,     X86::ROL16mi,    0 },
175     { X86::ROL32r1,     X86::ROL32m1,    0 },
176     { X86::ROL32rCL,    X86::ROL32mCL,   0 },
177     { X86::ROL32ri,     X86::ROL32mi,    0 },
178     { X86::ROL64r1,     X86::ROL64m1,    0 },
179     { X86::ROL64rCL,    X86::ROL64mCL,   0 },
180     { X86::ROL64ri,     X86::ROL64mi,    0 },
181     { X86::ROL8r1,      X86::ROL8m1,     0 },
182     { X86::ROL8rCL,     X86::ROL8mCL,    0 },
183     { X86::ROL8ri,      X86::ROL8mi,     0 },
184     { X86::ROR16r1,     X86::ROR16m1,    0 },
185     { X86::ROR16rCL,    X86::ROR16mCL,   0 },
186     { X86::ROR16ri,     X86::ROR16mi,    0 },
187     { X86::ROR32r1,     X86::ROR32m1,    0 },
188     { X86::ROR32rCL,    X86::ROR32mCL,   0 },
189     { X86::ROR32ri,     X86::ROR32mi,    0 },
190     { X86::ROR64r1,     X86::ROR64m1,    0 },
191     { X86::ROR64rCL,    X86::ROR64mCL,   0 },
192     { X86::ROR64ri,     X86::ROR64mi,    0 },
193     { X86::ROR8r1,      X86::ROR8m1,     0 },
194     { X86::ROR8rCL,     X86::ROR8mCL,    0 },
195     { X86::ROR8ri,      X86::ROR8mi,     0 },
196     { X86::SAR16r1,     X86::SAR16m1,    0 },
197     { X86::SAR16rCL,    X86::SAR16mCL,   0 },
198     { X86::SAR16ri,     X86::SAR16mi,    0 },
199     { X86::SAR32r1,     X86::SAR32m1,    0 },
200     { X86::SAR32rCL,    X86::SAR32mCL,   0 },
201     { X86::SAR32ri,     X86::SAR32mi,    0 },
202     { X86::SAR64r1,     X86::SAR64m1,    0 },
203     { X86::SAR64rCL,    X86::SAR64mCL,   0 },
204     { X86::SAR64ri,     X86::SAR64mi,    0 },
205     { X86::SAR8r1,      X86::SAR8m1,     0 },
206     { X86::SAR8rCL,     X86::SAR8mCL,    0 },
207     { X86::SAR8ri,      X86::SAR8mi,     0 },
208     { X86::SBB32ri,     X86::SBB32mi,    0 },
209     { X86::SBB32ri8,    X86::SBB32mi8,   0 },
210     { X86::SBB32rr,     X86::SBB32mr,    0 },
211     { X86::SBB64ri32,   X86::SBB64mi32,  0 },
212     { X86::SBB64ri8,    X86::SBB64mi8,   0 },
213     { X86::SBB64rr,     X86::SBB64mr,    0 },
214     { X86::SHL16rCL,    X86::SHL16mCL,   0 },
215     { X86::SHL16ri,     X86::SHL16mi,    0 },
216     { X86::SHL32rCL,    X86::SHL32mCL,   0 },
217     { X86::SHL32ri,     X86::SHL32mi,    0 },
218     { X86::SHL64rCL,    X86::SHL64mCL,   0 },
219     { X86::SHL64ri,     X86::SHL64mi,    0 },
220     { X86::SHL8rCL,     X86::SHL8mCL,    0 },
221     { X86::SHL8ri,      X86::SHL8mi,     0 },
222     { X86::SHLD16rrCL,  X86::SHLD16mrCL, 0 },
223     { X86::SHLD16rri8,  X86::SHLD16mri8, 0 },
224     { X86::SHLD32rrCL,  X86::SHLD32mrCL, 0 },
225     { X86::SHLD32rri8,  X86::SHLD32mri8, 0 },
226     { X86::SHLD64rrCL,  X86::SHLD64mrCL, 0 },
227     { X86::SHLD64rri8,  X86::SHLD64mri8, 0 },
228     { X86::SHR16r1,     X86::SHR16m1,    0 },
229     { X86::SHR16rCL,    X86::SHR16mCL,   0 },
230     { X86::SHR16ri,     X86::SHR16mi,    0 },
231     { X86::SHR32r1,     X86::SHR32m1,    0 },
232     { X86::SHR32rCL,    X86::SHR32mCL,   0 },
233     { X86::SHR32ri,     X86::SHR32mi,    0 },
234     { X86::SHR64r1,     X86::SHR64m1,    0 },
235     { X86::SHR64rCL,    X86::SHR64mCL,   0 },
236     { X86::SHR64ri,     X86::SHR64mi,    0 },
237     { X86::SHR8r1,      X86::SHR8m1,     0 },
238     { X86::SHR8rCL,     X86::SHR8mCL,    0 },
239     { X86::SHR8ri,      X86::SHR8mi,     0 },
240     { X86::SHRD16rrCL,  X86::SHRD16mrCL, 0 },
241     { X86::SHRD16rri8,  X86::SHRD16mri8, 0 },
242     { X86::SHRD32rrCL,  X86::SHRD32mrCL, 0 },
243     { X86::SHRD32rri8,  X86::SHRD32mri8, 0 },
244     { X86::SHRD64rrCL,  X86::SHRD64mrCL, 0 },
245     { X86::SHRD64rri8,  X86::SHRD64mri8, 0 },
246     { X86::SUB16ri,     X86::SUB16mi,    0 },
247     { X86::SUB16ri8,    X86::SUB16mi8,   0 },
248     { X86::SUB16rr,     X86::SUB16mr,    0 },
249     { X86::SUB32ri,     X86::SUB32mi,    0 },
250     { X86::SUB32ri8,    X86::SUB32mi8,   0 },
251     { X86::SUB32rr,     X86::SUB32mr,    0 },
252     { X86::SUB64ri32,   X86::SUB64mi32,  0 },
253     { X86::SUB64ri8,    X86::SUB64mi8,   0 },
254     { X86::SUB64rr,     X86::SUB64mr,    0 },
255     { X86::SUB8ri,      X86::SUB8mi,     0 },
256     { X86::SUB8rr,      X86::SUB8mr,     0 },
257     { X86::XOR16ri,     X86::XOR16mi,    0 },
258     { X86::XOR16ri8,    X86::XOR16mi8,   0 },
259     { X86::XOR16rr,     X86::XOR16mr,    0 },
260     { X86::XOR32ri,     X86::XOR32mi,    0 },
261     { X86::XOR32ri8,    X86::XOR32mi8,   0 },
262     { X86::XOR32rr,     X86::XOR32mr,    0 },
263     { X86::XOR64ri32,   X86::XOR64mi32,  0 },
264     { X86::XOR64ri8,    X86::XOR64mi8,   0 },
265     { X86::XOR64rr,     X86::XOR64mr,    0 },
266     { X86::XOR8ri,      X86::XOR8mi,     0 },
267     { X86::XOR8rr,      X86::XOR8mr,     0 }
268   };
269 
270   for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
271     unsigned RegOp = OpTbl2Addr[i].RegOp;
272     unsigned MemOp = OpTbl2Addr[i].MemOp;
273     unsigned Flags = OpTbl2Addr[i].Flags;
274     AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
275                   RegOp, MemOp,
276                   // Index 0, folded load and store, no alignment requirement.
277                   Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
278   }
279 
280   static const X86OpTblEntry OpTbl0[] = {
281     { X86::BT16ri8,     X86::BT16mi8,       TB_FOLDED_LOAD },
282     { X86::BT32ri8,     X86::BT32mi8,       TB_FOLDED_LOAD },
283     { X86::BT64ri8,     X86::BT64mi8,       TB_FOLDED_LOAD },
284     { X86::CALL32r,     X86::CALL32m,       TB_FOLDED_LOAD },
285     { X86::CALL64r,     X86::CALL64m,       TB_FOLDED_LOAD },
286     { X86::CMP16ri,     X86::CMP16mi,       TB_FOLDED_LOAD },
287     { X86::CMP16ri8,    X86::CMP16mi8,      TB_FOLDED_LOAD },
288     { X86::CMP16rr,     X86::CMP16mr,       TB_FOLDED_LOAD },
289     { X86::CMP32ri,     X86::CMP32mi,       TB_FOLDED_LOAD },
290     { X86::CMP32ri8,    X86::CMP32mi8,      TB_FOLDED_LOAD },
291     { X86::CMP32rr,     X86::CMP32mr,       TB_FOLDED_LOAD },
292     { X86::CMP64ri32,   X86::CMP64mi32,     TB_FOLDED_LOAD },
293     { X86::CMP64ri8,    X86::CMP64mi8,      TB_FOLDED_LOAD },
294     { X86::CMP64rr,     X86::CMP64mr,       TB_FOLDED_LOAD },
295     { X86::CMP8ri,      X86::CMP8mi,        TB_FOLDED_LOAD },
296     { X86::CMP8rr,      X86::CMP8mr,        TB_FOLDED_LOAD },
297     { X86::DIV16r,      X86::DIV16m,        TB_FOLDED_LOAD },
298     { X86::DIV32r,      X86::DIV32m,        TB_FOLDED_LOAD },
299     { X86::DIV64r,      X86::DIV64m,        TB_FOLDED_LOAD },
300     { X86::DIV8r,       X86::DIV8m,         TB_FOLDED_LOAD },
301     { X86::EXTRACTPSrr, X86::EXTRACTPSmr,   TB_FOLDED_STORE },
302     { X86::IDIV16r,     X86::IDIV16m,       TB_FOLDED_LOAD },
303     { X86::IDIV32r,     X86::IDIV32m,       TB_FOLDED_LOAD },
304     { X86::IDIV64r,     X86::IDIV64m,       TB_FOLDED_LOAD },
305     { X86::IDIV8r,      X86::IDIV8m,        TB_FOLDED_LOAD },
306     { X86::IMUL16r,     X86::IMUL16m,       TB_FOLDED_LOAD },
307     { X86::IMUL32r,     X86::IMUL32m,       TB_FOLDED_LOAD },
308     { X86::IMUL64r,     X86::IMUL64m,       TB_FOLDED_LOAD },
309     { X86::IMUL8r,      X86::IMUL8m,        TB_FOLDED_LOAD },
310     { X86::JMP32r,      X86::JMP32m,        TB_FOLDED_LOAD },
311     { X86::JMP64r,      X86::JMP64m,        TB_FOLDED_LOAD },
312     { X86::MOV16ri,     X86::MOV16mi,       TB_FOLDED_STORE },
313     { X86::MOV16rr,     X86::MOV16mr,       TB_FOLDED_STORE },
314     { X86::MOV32ri,     X86::MOV32mi,       TB_FOLDED_STORE },
315     { X86::MOV32rr,     X86::MOV32mr,       TB_FOLDED_STORE },
316     { X86::MOV64ri32,   X86::MOV64mi32,     TB_FOLDED_STORE },
317     { X86::MOV64rr,     X86::MOV64mr,       TB_FOLDED_STORE },
318     { X86::MOV8ri,      X86::MOV8mi,        TB_FOLDED_STORE },
319     { X86::MOV8rr,      X86::MOV8mr,        TB_FOLDED_STORE },
320     { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
321     { X86::MOVAPDrr,    X86::MOVAPDmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
322     { X86::MOVAPSrr,    X86::MOVAPSmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
323     { X86::MOVDQArr,    X86::MOVDQAmr,      TB_FOLDED_STORE | TB_ALIGN_16 },
324     { X86::MOVPDI2DIrr, X86::MOVPDI2DImr,   TB_FOLDED_STORE },
325     { X86::MOVPQIto64rr,X86::MOVPQI2QImr,   TB_FOLDED_STORE },
326     { X86::MOVSDto64rr, X86::MOVSDto64mr,   TB_FOLDED_STORE },
327     { X86::MOVSS2DIrr,  X86::MOVSS2DImr,    TB_FOLDED_STORE },
328     { X86::MOVUPDrr,    X86::MOVUPDmr,      TB_FOLDED_STORE },
329     { X86::MOVUPSrr,    X86::MOVUPSmr,      TB_FOLDED_STORE },
330     { X86::MUL16r,      X86::MUL16m,        TB_FOLDED_LOAD },
331     { X86::MUL32r,      X86::MUL32m,        TB_FOLDED_LOAD },
332     { X86::MUL64r,      X86::MUL64m,        TB_FOLDED_LOAD },
333     { X86::MUL8r,       X86::MUL8m,         TB_FOLDED_LOAD },
334     { X86::SETAEr,      X86::SETAEm,        TB_FOLDED_STORE },
335     { X86::SETAr,       X86::SETAm,         TB_FOLDED_STORE },
336     { X86::SETBEr,      X86::SETBEm,        TB_FOLDED_STORE },
337     { X86::SETBr,       X86::SETBm,         TB_FOLDED_STORE },
338     { X86::SETEr,       X86::SETEm,         TB_FOLDED_STORE },
339     { X86::SETGEr,      X86::SETGEm,        TB_FOLDED_STORE },
340     { X86::SETGr,       X86::SETGm,         TB_FOLDED_STORE },
341     { X86::SETLEr,      X86::SETLEm,        TB_FOLDED_STORE },
342     { X86::SETLr,       X86::SETLm,         TB_FOLDED_STORE },
343     { X86::SETNEr,      X86::SETNEm,        TB_FOLDED_STORE },
344     { X86::SETNOr,      X86::SETNOm,        TB_FOLDED_STORE },
345     { X86::SETNPr,      X86::SETNPm,        TB_FOLDED_STORE },
346     { X86::SETNSr,      X86::SETNSm,        TB_FOLDED_STORE },
347     { X86::SETOr,       X86::SETOm,         TB_FOLDED_STORE },
348     { X86::SETPr,       X86::SETPm,         TB_FOLDED_STORE },
349     { X86::SETSr,       X86::SETSm,         TB_FOLDED_STORE },
350     { X86::TAILJMPr,    X86::TAILJMPm,      TB_FOLDED_LOAD },
351     { X86::TAILJMPr64,  X86::TAILJMPm64,    TB_FOLDED_LOAD },
352     { X86::TEST16ri,    X86::TEST16mi,      TB_FOLDED_LOAD },
353     { X86::TEST32ri,    X86::TEST32mi,      TB_FOLDED_LOAD },
354     { X86::TEST64ri32,  X86::TEST64mi32,    TB_FOLDED_LOAD },
355     { X86::TEST8ri,     X86::TEST8mi,       TB_FOLDED_LOAD },
356     // AVX 128-bit versions of foldable instructions
357     { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr,  TB_FOLDED_STORE  },
358     { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
359     { X86::VMOVAPDrr,   X86::VMOVAPDmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
360     { X86::VMOVAPSrr,   X86::VMOVAPSmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
361     { X86::VMOVDQArr,   X86::VMOVDQAmr,     TB_FOLDED_STORE | TB_ALIGN_16 },
362     { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr,  TB_FOLDED_STORE },
363     { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
364     { X86::VMOVSDto64rr,X86::VMOVSDto64mr,  TB_FOLDED_STORE },
365     { X86::VMOVSS2DIrr, X86::VMOVSS2DImr,   TB_FOLDED_STORE },
366     { X86::VMOVUPDrr,   X86::VMOVUPDmr,     TB_FOLDED_STORE },
367     { X86::VMOVUPSrr,   X86::VMOVUPSmr,     TB_FOLDED_STORE },
368     // AVX 256-bit foldable instructions
369     { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
370     { X86::VMOVAPDYrr,  X86::VMOVAPDYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
371     { X86::VMOVAPSYrr,  X86::VMOVAPSYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
372     { X86::VMOVDQAYrr,  X86::VMOVDQAYmr,    TB_FOLDED_STORE | TB_ALIGN_32 },
373     { X86::VMOVUPDYrr,  X86::VMOVUPDYmr,    TB_FOLDED_STORE },
374     { X86::VMOVUPSYrr,  X86::VMOVUPSYmr,    TB_FOLDED_STORE },
375     // AVX-512 foldable instructions
376     { X86::VMOVPDI2DIZrr,X86::VMOVPDI2DIZmr,  TB_FOLDED_STORE }
377   };
378 
379   for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
380     unsigned RegOp      = OpTbl0[i].RegOp;
381     unsigned MemOp      = OpTbl0[i].MemOp;
382     unsigned Flags      = OpTbl0[i].Flags;
383     AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
384                   RegOp, MemOp, TB_INDEX_0 | Flags);
385   }
386 
387   static const X86OpTblEntry OpTbl1[] = {
388     { X86::CMP16rr,         X86::CMP16rm,             0 },
389     { X86::CMP32rr,         X86::CMP32rm,             0 },
390     { X86::CMP64rr,         X86::CMP64rm,             0 },
391     { X86::CMP8rr,          X86::CMP8rm,              0 },
392     { X86::CVTSD2SSrr,      X86::CVTSD2SSrm,          0 },
393     { X86::CVTSI2SD64rr,    X86::CVTSI2SD64rm,        0 },
394     { X86::CVTSI2SDrr,      X86::CVTSI2SDrm,          0 },
395     { X86::CVTSI2SS64rr,    X86::CVTSI2SS64rm,        0 },
396     { X86::CVTSI2SSrr,      X86::CVTSI2SSrm,          0 },
397     { X86::CVTSS2SDrr,      X86::CVTSS2SDrm,          0 },
398     { X86::CVTTSD2SI64rr,   X86::CVTTSD2SI64rm,       0 },
399     { X86::CVTTSD2SIrr,     X86::CVTTSD2SIrm,         0 },
400     { X86::CVTTSS2SI64rr,   X86::CVTTSS2SI64rm,       0 },
401     { X86::CVTTSS2SIrr,     X86::CVTTSS2SIrm,         0 },
402     { X86::IMUL16rri,       X86::IMUL16rmi,           0 },
403     { X86::IMUL16rri8,      X86::IMUL16rmi8,          0 },
404     { X86::IMUL32rri,       X86::IMUL32rmi,           0 },
405     { X86::IMUL32rri8,      X86::IMUL32rmi8,          0 },
406     { X86::IMUL64rri32,     X86::IMUL64rmi32,         0 },
407     { X86::IMUL64rri8,      X86::IMUL64rmi8,          0 },
408     { X86::Int_COMISDrr,    X86::Int_COMISDrm,        0 },
409     { X86::Int_COMISSrr,    X86::Int_COMISSrm,        0 },
410     { X86::CVTSD2SI64rr,    X86::CVTSD2SI64rm,        0 },
411     { X86::CVTSD2SIrr,      X86::CVTSD2SIrm,          0 },
412     { X86::CVTSS2SI64rr,    X86::CVTSS2SI64rm,        0 },
413     { X86::CVTSS2SIrr,      X86::CVTSS2SIrm,          0 },
414     { X86::CVTTPD2DQrr,     X86::CVTTPD2DQrm,         TB_ALIGN_16 },
415     { X86::CVTTPS2DQrr,     X86::CVTTPS2DQrm,         TB_ALIGN_16 },
416     { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm,  0 },
417     { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm,     0 },
418     { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm,  0 },
419     { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm,     0 },
420     { X86::Int_UCOMISDrr,   X86::Int_UCOMISDrm,       0 },
421     { X86::Int_UCOMISSrr,   X86::Int_UCOMISSrm,       0 },
422     { X86::MOV16rr,         X86::MOV16rm,             0 },
423     { X86::MOV32rr,         X86::MOV32rm,             0 },
424     { X86::MOV64rr,         X86::MOV64rm,             0 },
425     { X86::MOV64toPQIrr,    X86::MOVQI2PQIrm,         0 },
426     { X86::MOV64toSDrr,     X86::MOV64toSDrm,         0 },
427     { X86::MOV8rr,          X86::MOV8rm,              0 },
428     { X86::MOVAPDrr,        X86::MOVAPDrm,            TB_ALIGN_16 },
429     { X86::MOVAPSrr,        X86::MOVAPSrm,            TB_ALIGN_16 },
430     { X86::MOVDDUPrr,       X86::MOVDDUPrm,           0 },
431     { X86::MOVDI2PDIrr,     X86::MOVDI2PDIrm,         0 },
432     { X86::MOVDI2SSrr,      X86::MOVDI2SSrm,          0 },
433     { X86::MOVDQArr,        X86::MOVDQArm,            TB_ALIGN_16 },
434     { X86::MOVSHDUPrr,      X86::MOVSHDUPrm,          TB_ALIGN_16 },
435     { X86::MOVSLDUPrr,      X86::MOVSLDUPrm,          TB_ALIGN_16 },
436     { X86::MOVSX16rr8,      X86::MOVSX16rm8,          0 },
437     { X86::MOVSX32rr16,     X86::MOVSX32rm16,         0 },
438     { X86::MOVSX32rr8,      X86::MOVSX32rm8,          0 },
439     { X86::MOVSX64rr16,     X86::MOVSX64rm16,         0 },
440     { X86::MOVSX64rr32,     X86::MOVSX64rm32,         0 },
441     { X86::MOVSX64rr8,      X86::MOVSX64rm8,          0 },
442     { X86::MOVUPDrr,        X86::MOVUPDrm,            TB_ALIGN_16 },
443     { X86::MOVUPSrr,        X86::MOVUPSrm,            0 },
444     { X86::MOVZDI2PDIrr,    X86::MOVZDI2PDIrm,        0 },
445     { X86::MOVZQI2PQIrr,    X86::MOVZQI2PQIrm,        0 },
446     { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm,     TB_ALIGN_16 },
447     { X86::MOVZX16rr8,      X86::MOVZX16rm8,          0 },
448     { X86::MOVZX32rr16,     X86::MOVZX32rm16,         0 },
449     { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8,   0 },
450     { X86::MOVZX32rr8,      X86::MOVZX32rm8,          0 },
451     { X86::PABSBrr128,      X86::PABSBrm128,          TB_ALIGN_16 },
452     { X86::PABSDrr128,      X86::PABSDrm128,          TB_ALIGN_16 },
453     { X86::PABSWrr128,      X86::PABSWrm128,          TB_ALIGN_16 },
454     { X86::PSHUFDri,        X86::PSHUFDmi,            TB_ALIGN_16 },
455     { X86::PSHUFHWri,       X86::PSHUFHWmi,           TB_ALIGN_16 },
456     { X86::PSHUFLWri,       X86::PSHUFLWmi,           TB_ALIGN_16 },
457     { X86::RCPPSr,          X86::RCPPSm,              TB_ALIGN_16 },
458     { X86::RCPPSr_Int,      X86::RCPPSm_Int,          TB_ALIGN_16 },
459     { X86::RSQRTPSr,        X86::RSQRTPSm,            TB_ALIGN_16 },
460     { X86::RSQRTPSr_Int,    X86::RSQRTPSm_Int,        TB_ALIGN_16 },
461     { X86::RSQRTSSr,        X86::RSQRTSSm,            0 },
462     { X86::RSQRTSSr_Int,    X86::RSQRTSSm_Int,        0 },
463     { X86::SQRTPDr,         X86::SQRTPDm,             TB_ALIGN_16 },
464     { X86::SQRTPSr,         X86::SQRTPSm,             TB_ALIGN_16 },
465     { X86::SQRTSDr,         X86::SQRTSDm,             0 },
466     { X86::SQRTSDr_Int,     X86::SQRTSDm_Int,         0 },
467     { X86::SQRTSSr,         X86::SQRTSSm,             0 },
468     { X86::SQRTSSr_Int,     X86::SQRTSSm_Int,         0 },
469     { X86::TEST16rr,        X86::TEST16rm,            0 },
470     { X86::TEST32rr,        X86::TEST32rm,            0 },
471     { X86::TEST64rr,        X86::TEST64rm,            0 },
472     { X86::TEST8rr,         X86::TEST8rm,             0 },
473     // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
474     { X86::UCOMISDrr,       X86::UCOMISDrm,           0 },
475     { X86::UCOMISSrr,       X86::UCOMISSrm,           0 },
476     // AVX 128-bit versions of foldable instructions
477     { X86::Int_VCOMISDrr,   X86::Int_VCOMISDrm,       0 },
478     { X86::Int_VCOMISSrr,   X86::Int_VCOMISSrm,       0 },
479     { X86::Int_VUCOMISDrr,  X86::Int_VUCOMISDrm,      0 },
480     { X86::Int_VUCOMISSrr,  X86::Int_VUCOMISSrm,      0 },
481     { X86::VCVTTSD2SI64rr,  X86::VCVTTSD2SI64rm,      0 },
482     { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
483     { X86::VCVTTSD2SIrr,    X86::VCVTTSD2SIrm,        0 },
484     { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm,    0 },
485     { X86::VCVTTSS2SI64rr,  X86::VCVTTSS2SI64rm,      0 },
486     { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
487     { X86::VCVTTSS2SIrr,    X86::VCVTTSS2SIrm,        0 },
488     { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm,    0 },
489     { X86::VCVTSD2SI64rr,   X86::VCVTSD2SI64rm,       0 },
490     { X86::VCVTSD2SIrr,     X86::VCVTSD2SIrm,         0 },
491     { X86::VCVTSS2SI64rr,   X86::VCVTSS2SI64rm,       0 },
492     { X86::VCVTSS2SIrr,     X86::VCVTSS2SIrm,         0 },
493     { X86::VMOV64toPQIrr,   X86::VMOVQI2PQIrm,        0 },
494     { X86::VMOV64toSDrr,    X86::VMOV64toSDrm,        0 },
495     { X86::VMOVAPDrr,       X86::VMOVAPDrm,           TB_ALIGN_16 },
496     { X86::VMOVAPSrr,       X86::VMOVAPSrm,           TB_ALIGN_16 },
497     { X86::VMOVDDUPrr,      X86::VMOVDDUPrm,          0 },
498     { X86::VMOVDI2PDIrr,    X86::VMOVDI2PDIrm,        0 },
499     { X86::VMOVDI2SSrr,     X86::VMOVDI2SSrm,         0 },
500     { X86::VMOVDQArr,       X86::VMOVDQArm,           TB_ALIGN_16 },
501     { X86::VMOVSLDUPrr,     X86::VMOVSLDUPrm,         TB_ALIGN_16 },
502     { X86::VMOVSHDUPrr,     X86::VMOVSHDUPrm,         TB_ALIGN_16 },
503     { X86::VMOVUPDrr,       X86::VMOVUPDrm,           0 },
504     { X86::VMOVUPSrr,       X86::VMOVUPSrm,           0 },
505     { X86::VMOVZDI2PDIrr,   X86::VMOVZDI2PDIrm,       0 },
506     { X86::VMOVZQI2PQIrr,   X86::VMOVZQI2PQIrm,       0 },
507     { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm,    TB_ALIGN_16 },
508     { X86::VPABSBrr128,     X86::VPABSBrm128,         0 },
509     { X86::VPABSDrr128,     X86::VPABSDrm128,         0 },
510     { X86::VPABSWrr128,     X86::VPABSWrm128,         0 },
511     { X86::VPERMILPDri,     X86::VPERMILPDmi,         0 },
512     { X86::VPERMILPSri,     X86::VPERMILPSmi,         0 },
513     { X86::VPSHUFDri,       X86::VPSHUFDmi,           0 },
514     { X86::VPSHUFHWri,      X86::VPSHUFHWmi,          0 },
515     { X86::VPSHUFLWri,      X86::VPSHUFLWmi,          0 },
516     { X86::VRCPPSr,         X86::VRCPPSm,             0 },
517     { X86::VRCPPSr_Int,     X86::VRCPPSm_Int,         0 },
518     { X86::VRSQRTPSr,       X86::VRSQRTPSm,           0 },
519     { X86::VRSQRTPSr_Int,   X86::VRSQRTPSm_Int,       0 },
520     { X86::VSQRTPDr,        X86::VSQRTPDm,            0 },
521     { X86::VSQRTPSr,        X86::VSQRTPSm,            0 },
522     { X86::VUCOMISDrr,      X86::VUCOMISDrm,          0 },
523     { X86::VUCOMISSrr,      X86::VUCOMISSrm,          0 },
524     { X86::VBROADCASTSSrr,  X86::VBROADCASTSSrm,      TB_NO_REVERSE },
525 
526     // AVX 256-bit foldable instructions
527     { X86::VMOVAPDYrr,      X86::VMOVAPDYrm,          TB_ALIGN_32 },
528     { X86::VMOVAPSYrr,      X86::VMOVAPSYrm,          TB_ALIGN_32 },
529     { X86::VMOVDQAYrr,      X86::VMOVDQAYrm,          TB_ALIGN_32 },
530     { X86::VMOVUPDYrr,      X86::VMOVUPDYrm,          0 },
531     { X86::VMOVUPSYrr,      X86::VMOVUPSYrm,          0 },
532     { X86::VPERMILPDYri,    X86::VPERMILPDYmi,        0 },
533     { X86::VPERMILPSYri,    X86::VPERMILPSYmi,        0 },
534 
535     // AVX2 foldable instructions
536     { X86::VPABSBrr256,     X86::VPABSBrm256,         0 },
537     { X86::VPABSDrr256,     X86::VPABSDrm256,         0 },
538     { X86::VPABSWrr256,     X86::VPABSWrm256,         0 },
539     { X86::VPSHUFDYri,      X86::VPSHUFDYmi,          0 },
540     { X86::VPSHUFHWYri,     X86::VPSHUFHWYmi,         0 },
541     { X86::VPSHUFLWYri,     X86::VPSHUFLWYmi,         0 },
542     { X86::VRCPPSYr,        X86::VRCPPSYm,            0 },
543     { X86::VRCPPSYr_Int,    X86::VRCPPSYm_Int,        0 },
544     { X86::VRSQRTPSYr,      X86::VRSQRTPSYm,          0 },
545     { X86::VSQRTPDYr,       X86::VSQRTPDYm,           0 },
546     { X86::VSQRTPSYr,       X86::VSQRTPSYm,           0 },
547     { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm,     TB_NO_REVERSE },
548     { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm,     TB_NO_REVERSE },
549 
550     // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions
551     { X86::BEXTR32rr,       X86::BEXTR32rm,           0 },
552     { X86::BEXTR64rr,       X86::BEXTR64rm,           0 },
553     { X86::BEXTRI32ri,      X86::BEXTRI32mi,          0 },
554     { X86::BEXTRI64ri,      X86::BEXTRI64mi,          0 },
555     { X86::BLCFILL32rr,     X86::BLCFILL32rm,         0 },
556     { X86::BLCFILL64rr,     X86::BLCFILL64rm,         0 },
557     { X86::BLCI32rr,        X86::BLCI32rm,            0 },
558     { X86::BLCI64rr,        X86::BLCI64rm,            0 },
559     { X86::BLCIC32rr,       X86::BLCIC32rm,           0 },
560     { X86::BLCIC64rr,       X86::BLCIC64rm,           0 },
561     { X86::BLCMSK32rr,      X86::BLCMSK32rm,          0 },
562     { X86::BLCMSK64rr,      X86::BLCMSK64rm,          0 },
563     { X86::BLCS32rr,        X86::BLCS32rm,            0 },
564     { X86::BLCS64rr,        X86::BLCS64rm,            0 },
565     { X86::BLSFILL32rr,     X86::BLSFILL32rm,         0 },
566     { X86::BLSFILL64rr,     X86::BLSFILL64rm,         0 },
567     { X86::BLSI32rr,        X86::BLSI32rm,            0 },
568     { X86::BLSI64rr,        X86::BLSI64rm,            0 },
569     { X86::BLSIC32rr,       X86::BLSIC32rm,           0 },
570     { X86::BLSIC64rr,       X86::BLSIC64rm,           0 },
571     { X86::BLSMSK32rr,      X86::BLSMSK32rm,          0 },
572     { X86::BLSMSK64rr,      X86::BLSMSK64rm,          0 },
573     { X86::BLSR32rr,        X86::BLSR32rm,            0 },
574     { X86::BLSR64rr,        X86::BLSR64rm,            0 },
575     { X86::BZHI32rr,        X86::BZHI32rm,            0 },
576     { X86::BZHI64rr,        X86::BZHI64rm,            0 },
577     { X86::LZCNT16rr,       X86::LZCNT16rm,           0 },
578     { X86::LZCNT32rr,       X86::LZCNT32rm,           0 },
579     { X86::LZCNT64rr,       X86::LZCNT64rm,           0 },
580     { X86::POPCNT16rr,      X86::POPCNT16rm,          0 },
581     { X86::POPCNT32rr,      X86::POPCNT32rm,          0 },
582     { X86::POPCNT64rr,      X86::POPCNT64rm,          0 },
583     { X86::RORX32ri,        X86::RORX32mi,            0 },
584     { X86::RORX64ri,        X86::RORX64mi,            0 },
585     { X86::SARX32rr,        X86::SARX32rm,            0 },
586     { X86::SARX64rr,        X86::SARX64rm,            0 },
587     { X86::SHRX32rr,        X86::SHRX32rm,            0 },
588     { X86::SHRX64rr,        X86::SHRX64rm,            0 },
589     { X86::SHLX32rr,        X86::SHLX32rm,            0 },
590     { X86::SHLX64rr,        X86::SHLX64rm,            0 },
591     { X86::T1MSKC32rr,      X86::T1MSKC32rm,          0 },
592     { X86::T1MSKC64rr,      X86::T1MSKC64rm,          0 },
593     { X86::TZCNT16rr,       X86::TZCNT16rm,           0 },
594     { X86::TZCNT32rr,       X86::TZCNT32rm,           0 },
595     { X86::TZCNT64rr,       X86::TZCNT64rm,           0 },
596     { X86::TZMSK32rr,       X86::TZMSK32rm,           0 },
597     { X86::TZMSK64rr,       X86::TZMSK64rm,           0 },
598 
599     // AVX-512 foldable instructions
600     { X86::VMOV64toPQIZrr,  X86::VMOVQI2PQIZrm,       0 },
601     { X86::VMOVDI2SSZrr,    X86::VMOVDI2SSZrm,        0 },
602     { X86::VMOVDQA32rr,     X86::VMOVDQA32rm,         TB_ALIGN_64 },
603     { X86::VMOVDQA64rr,     X86::VMOVDQA64rm,         TB_ALIGN_64 },
604     { X86::VMOVDQU32rr,     X86::VMOVDQU32rm,         0 },
605     { X86::VMOVDQU64rr,     X86::VMOVDQU64rm,         0 },
606 
607     // AES foldable instructions
608     { X86::AESIMCrr,              X86::AESIMCrm,              TB_ALIGN_16 },
609     { X86::AESKEYGENASSIST128rr,  X86::AESKEYGENASSIST128rm,  TB_ALIGN_16 },
610     { X86::VAESIMCrr,             X86::VAESIMCrm,             TB_ALIGN_16 },
611     { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, TB_ALIGN_16 },
612   };
613 
614   for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
615     unsigned RegOp = OpTbl1[i].RegOp;
616     unsigned MemOp = OpTbl1[i].MemOp;
617     unsigned Flags = OpTbl1[i].Flags;
618     AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
619                   RegOp, MemOp,
620                   // Index 1, folded load
621                   Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
622   }
623 
624   static const X86OpTblEntry OpTbl2[] = {
625     { X86::ADC32rr,         X86::ADC32rm,       0 },
626     { X86::ADC64rr,         X86::ADC64rm,       0 },
627     { X86::ADD16rr,         X86::ADD16rm,       0 },
628     { X86::ADD16rr_DB,      X86::ADD16rm,       TB_NO_REVERSE },
629     { X86::ADD32rr,         X86::ADD32rm,       0 },
630     { X86::ADD32rr_DB,      X86::ADD32rm,       TB_NO_REVERSE },
631     { X86::ADD64rr,         X86::ADD64rm,       0 },
632     { X86::ADD64rr_DB,      X86::ADD64rm,       TB_NO_REVERSE },
633     { X86::ADD8rr,          X86::ADD8rm,        0 },
634     { X86::ADDPDrr,         X86::ADDPDrm,       TB_ALIGN_16 },
635     { X86::ADDPSrr,         X86::ADDPSrm,       TB_ALIGN_16 },
636     { X86::ADDSDrr,         X86::ADDSDrm,       0 },
637     { X86::ADDSSrr,         X86::ADDSSrm,       0 },
638     { X86::ADDSUBPDrr,      X86::ADDSUBPDrm,    TB_ALIGN_16 },
639     { X86::ADDSUBPSrr,      X86::ADDSUBPSrm,    TB_ALIGN_16 },
640     { X86::AND16rr,         X86::AND16rm,       0 },
641     { X86::AND32rr,         X86::AND32rm,       0 },
642     { X86::AND64rr,         X86::AND64rm,       0 },
643     { X86::AND8rr,          X86::AND8rm,        0 },
644     { X86::ANDNPDrr,        X86::ANDNPDrm,      TB_ALIGN_16 },
645     { X86::ANDNPSrr,        X86::ANDNPSrm,      TB_ALIGN_16 },
646     { X86::ANDPDrr,         X86::ANDPDrm,       TB_ALIGN_16 },
647     { X86::ANDPSrr,         X86::ANDPSrm,       TB_ALIGN_16 },
648     { X86::BLENDPDrri,      X86::BLENDPDrmi,    TB_ALIGN_16 },
649     { X86::BLENDPSrri,      X86::BLENDPSrmi,    TB_ALIGN_16 },
650     { X86::BLENDVPDrr0,     X86::BLENDVPDrm0,   TB_ALIGN_16 },
651     { X86::BLENDVPSrr0,     X86::BLENDVPSrm0,   TB_ALIGN_16 },
652     { X86::CMOVA16rr,       X86::CMOVA16rm,     0 },
653     { X86::CMOVA32rr,       X86::CMOVA32rm,     0 },
654     { X86::CMOVA64rr,       X86::CMOVA64rm,     0 },
655     { X86::CMOVAE16rr,      X86::CMOVAE16rm,    0 },
656     { X86::CMOVAE32rr,      X86::CMOVAE32rm,    0 },
657     { X86::CMOVAE64rr,      X86::CMOVAE64rm,    0 },
658     { X86::CMOVB16rr,       X86::CMOVB16rm,     0 },
659     { X86::CMOVB32rr,       X86::CMOVB32rm,     0 },
660     { X86::CMOVB64rr,       X86::CMOVB64rm,     0 },
661     { X86::CMOVBE16rr,      X86::CMOVBE16rm,    0 },
662     { X86::CMOVBE32rr,      X86::CMOVBE32rm,    0 },
663     { X86::CMOVBE64rr,      X86::CMOVBE64rm,    0 },
664     { X86::CMOVE16rr,       X86::CMOVE16rm,     0 },
665     { X86::CMOVE32rr,       X86::CMOVE32rm,     0 },
666     { X86::CMOVE64rr,       X86::CMOVE64rm,     0 },
667     { X86::CMOVG16rr,       X86::CMOVG16rm,     0 },
668     { X86::CMOVG32rr,       X86::CMOVG32rm,     0 },
669     { X86::CMOVG64rr,       X86::CMOVG64rm,     0 },
670     { X86::CMOVGE16rr,      X86::CMOVGE16rm,    0 },
671     { X86::CMOVGE32rr,      X86::CMOVGE32rm,    0 },
672     { X86::CMOVGE64rr,      X86::CMOVGE64rm,    0 },
673     { X86::CMOVL16rr,       X86::CMOVL16rm,     0 },
674     { X86::CMOVL32rr,       X86::CMOVL32rm,     0 },
675     { X86::CMOVL64rr,       X86::CMOVL64rm,     0 },
676     { X86::CMOVLE16rr,      X86::CMOVLE16rm,    0 },
677     { X86::CMOVLE32rr,      X86::CMOVLE32rm,    0 },
678     { X86::CMOVLE64rr,      X86::CMOVLE64rm,    0 },
679     { X86::CMOVNE16rr,      X86::CMOVNE16rm,    0 },
680     { X86::CMOVNE32rr,      X86::CMOVNE32rm,    0 },
681     { X86::CMOVNE64rr,      X86::CMOVNE64rm,    0 },
682     { X86::CMOVNO16rr,      X86::CMOVNO16rm,    0 },
683     { X86::CMOVNO32rr,      X86::CMOVNO32rm,    0 },
684     { X86::CMOVNO64rr,      X86::CMOVNO64rm,    0 },
685     { X86::CMOVNP16rr,      X86::CMOVNP16rm,    0 },
686     { X86::CMOVNP32rr,      X86::CMOVNP32rm,    0 },
687     { X86::CMOVNP64rr,      X86::CMOVNP64rm,    0 },
688     { X86::CMOVNS16rr,      X86::CMOVNS16rm,    0 },
689     { X86::CMOVNS32rr,      X86::CMOVNS32rm,    0 },
690     { X86::CMOVNS64rr,      X86::CMOVNS64rm,    0 },
691     { X86::CMOVO16rr,       X86::CMOVO16rm,     0 },
692     { X86::CMOVO32rr,       X86::CMOVO32rm,     0 },
693     { X86::CMOVO64rr,       X86::CMOVO64rm,     0 },
694     { X86::CMOVP16rr,       X86::CMOVP16rm,     0 },
695     { X86::CMOVP32rr,       X86::CMOVP32rm,     0 },
696     { X86::CMOVP64rr,       X86::CMOVP64rm,     0 },
697     { X86::CMOVS16rr,       X86::CMOVS16rm,     0 },
698     { X86::CMOVS32rr,       X86::CMOVS32rm,     0 },
699     { X86::CMOVS64rr,       X86::CMOVS64rm,     0 },
700     { X86::CMPPDrri,        X86::CMPPDrmi,      TB_ALIGN_16 },
701     { X86::CMPPSrri,        X86::CMPPSrmi,      TB_ALIGN_16 },
702     { X86::CMPSDrr,         X86::CMPSDrm,       0 },
703     { X86::CMPSSrr,         X86::CMPSSrm,       0 },
704     { X86::DIVPDrr,         X86::DIVPDrm,       TB_ALIGN_16 },
705     { X86::DIVPSrr,         X86::DIVPSrm,       TB_ALIGN_16 },
706     { X86::DIVSDrr,         X86::DIVSDrm,       0 },
707     { X86::DIVSSrr,         X86::DIVSSrm,       0 },
708     { X86::FsANDNPDrr,      X86::FsANDNPDrm,    TB_ALIGN_16 },
709     { X86::FsANDNPSrr,      X86::FsANDNPSrm,    TB_ALIGN_16 },
710     { X86::FsANDPDrr,       X86::FsANDPDrm,     TB_ALIGN_16 },
711     { X86::FsANDPSrr,       X86::FsANDPSrm,     TB_ALIGN_16 },
712     { X86::FsORPDrr,        X86::FsORPDrm,      TB_ALIGN_16 },
713     { X86::FsORPSrr,        X86::FsORPSrm,      TB_ALIGN_16 },
714     { X86::FsXORPDrr,       X86::FsXORPDrm,     TB_ALIGN_16 },
715     { X86::FsXORPSrr,       X86::FsXORPSrm,     TB_ALIGN_16 },
716     { X86::HADDPDrr,        X86::HADDPDrm,      TB_ALIGN_16 },
717     { X86::HADDPSrr,        X86::HADDPSrm,      TB_ALIGN_16 },
718     { X86::HSUBPDrr,        X86::HSUBPDrm,      TB_ALIGN_16 },
719     { X86::HSUBPSrr,        X86::HSUBPSrm,      TB_ALIGN_16 },
720     { X86::IMUL16rr,        X86::IMUL16rm,      0 },
721     { X86::IMUL32rr,        X86::IMUL32rm,      0 },
722     { X86::IMUL64rr,        X86::IMUL64rm,      0 },
723     { X86::Int_CMPSDrr,     X86::Int_CMPSDrm,   0 },
724     { X86::Int_CMPSSrr,     X86::Int_CMPSSrm,   0 },
725     { X86::Int_CVTSD2SSrr,  X86::Int_CVTSD2SSrm,      0 },
726     { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm,    0 },
727     { X86::Int_CVTSI2SDrr,  X86::Int_CVTSI2SDrm,      0 },
728     { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm,    0 },
729     { X86::Int_CVTSI2SSrr,  X86::Int_CVTSI2SSrm,      0 },
730     { X86::Int_CVTSS2SDrr,  X86::Int_CVTSS2SDrm,      0 },
731     { X86::MAXPDrr,         X86::MAXPDrm,       TB_ALIGN_16 },
732     { X86::MAXPSrr,         X86::MAXPSrm,       TB_ALIGN_16 },
733     { X86::MAXSDrr,         X86::MAXSDrm,       0 },
734     { X86::MAXSSrr,         X86::MAXSSrm,       0 },
735     { X86::MINPDrr,         X86::MINPDrm,       TB_ALIGN_16 },
736     { X86::MINPSrr,         X86::MINPSrm,       TB_ALIGN_16 },
737     { X86::MINSDrr,         X86::MINSDrm,       0 },
738     { X86::MINSSrr,         X86::MINSSrm,       0 },
739     { X86::MPSADBWrri,      X86::MPSADBWrmi,    TB_ALIGN_16 },
740     { X86::MULPDrr,         X86::MULPDrm,       TB_ALIGN_16 },
741     { X86::MULPSrr,         X86::MULPSrm,       TB_ALIGN_16 },
742     { X86::MULSDrr,         X86::MULSDrm,       0 },
743     { X86::MULSSrr,         X86::MULSSrm,       0 },
744     { X86::OR16rr,          X86::OR16rm,        0 },
745     { X86::OR32rr,          X86::OR32rm,        0 },
746     { X86::OR64rr,          X86::OR64rm,        0 },
747     { X86::OR8rr,           X86::OR8rm,         0 },
748     { X86::ORPDrr,          X86::ORPDrm,        TB_ALIGN_16 },
749     { X86::ORPSrr,          X86::ORPSrm,        TB_ALIGN_16 },
750     { X86::PACKSSDWrr,      X86::PACKSSDWrm,    TB_ALIGN_16 },
751     { X86::PACKSSWBrr,      X86::PACKSSWBrm,    TB_ALIGN_16 },
752     { X86::PACKUSDWrr,      X86::PACKUSDWrm,    TB_ALIGN_16 },
753     { X86::PACKUSWBrr,      X86::PACKUSWBrm,    TB_ALIGN_16 },
754     { X86::PADDBrr,         X86::PADDBrm,       TB_ALIGN_16 },
755     { X86::PADDDrr,         X86::PADDDrm,       TB_ALIGN_16 },
756     { X86::PADDQrr,         X86::PADDQrm,       TB_ALIGN_16 },
757     { X86::PADDSBrr,        X86::PADDSBrm,      TB_ALIGN_16 },
758     { X86::PADDSWrr,        X86::PADDSWrm,      TB_ALIGN_16 },
759     { X86::PADDUSBrr,       X86::PADDUSBrm,     TB_ALIGN_16 },
760     { X86::PADDUSWrr,       X86::PADDUSWrm,     TB_ALIGN_16 },
761     { X86::PADDWrr,         X86::PADDWrm,       TB_ALIGN_16 },
762     { X86::PALIGNR128rr,    X86::PALIGNR128rm,  TB_ALIGN_16 },
763     { X86::PANDNrr,         X86::PANDNrm,       TB_ALIGN_16 },
764     { X86::PANDrr,          X86::PANDrm,        TB_ALIGN_16 },
765     { X86::PAVGBrr,         X86::PAVGBrm,       TB_ALIGN_16 },
766     { X86::PAVGWrr,         X86::PAVGWrm,       TB_ALIGN_16 },
767     { X86::PBLENDWrri,      X86::PBLENDWrmi,    TB_ALIGN_16 },
768     { X86::PCMPEQBrr,       X86::PCMPEQBrm,     TB_ALIGN_16 },
769     { X86::PCMPEQDrr,       X86::PCMPEQDrm,     TB_ALIGN_16 },
770     { X86::PCMPEQQrr,       X86::PCMPEQQrm,     TB_ALIGN_16 },
771     { X86::PCMPEQWrr,       X86::PCMPEQWrm,     TB_ALIGN_16 },
772     { X86::PCMPGTBrr,       X86::PCMPGTBrm,     TB_ALIGN_16 },
773     { X86::PCMPGTDrr,       X86::PCMPGTDrm,     TB_ALIGN_16 },
774     { X86::PCMPGTQrr,       X86::PCMPGTQrm,     TB_ALIGN_16 },
775     { X86::PCMPGTWrr,       X86::PCMPGTWrm,     TB_ALIGN_16 },
776     { X86::PHADDDrr,        X86::PHADDDrm,      TB_ALIGN_16 },
777     { X86::PHADDWrr,        X86::PHADDWrm,      TB_ALIGN_16 },
778     { X86::PHADDSWrr128,    X86::PHADDSWrm128,  TB_ALIGN_16 },
779     { X86::PHSUBDrr,        X86::PHSUBDrm,      TB_ALIGN_16 },
780     { X86::PHSUBSWrr128,    X86::PHSUBSWrm128,  TB_ALIGN_16 },
781     { X86::PHSUBWrr,        X86::PHSUBWrm,      TB_ALIGN_16 },
782     { X86::PINSRWrri,       X86::PINSRWrmi,     TB_ALIGN_16 },
783     { X86::PMADDUBSWrr128,  X86::PMADDUBSWrm128, TB_ALIGN_16 },
784     { X86::PMADDWDrr,       X86::PMADDWDrm,     TB_ALIGN_16 },
785     { X86::PMAXSWrr,        X86::PMAXSWrm,      TB_ALIGN_16 },
786     { X86::PMAXUBrr,        X86::PMAXUBrm,      TB_ALIGN_16 },
787     { X86::PMINSWrr,        X86::PMINSWrm,      TB_ALIGN_16 },
788     { X86::PMINUBrr,        X86::PMINUBrm,      TB_ALIGN_16 },
789     { X86::PMINSBrr,        X86::PMINSBrm,      TB_ALIGN_16 },
790     { X86::PMINSDrr,        X86::PMINSDrm,      TB_ALIGN_16 },
791     { X86::PMINUDrr,        X86::PMINUDrm,      TB_ALIGN_16 },
792     { X86::PMINUWrr,        X86::PMINUWrm,      TB_ALIGN_16 },
793     { X86::PMAXSBrr,        X86::PMAXSBrm,      TB_ALIGN_16 },
794     { X86::PMAXSDrr,        X86::PMAXSDrm,      TB_ALIGN_16 },
795     { X86::PMAXUDrr,        X86::PMAXUDrm,      TB_ALIGN_16 },
796     { X86::PMAXUWrr,        X86::PMAXUWrm,      TB_ALIGN_16 },
797     { X86::PMULDQrr,        X86::PMULDQrm,      TB_ALIGN_16 },
798     { X86::PMULHRSWrr128,   X86::PMULHRSWrm128, TB_ALIGN_16 },
799     { X86::PMULHUWrr,       X86::PMULHUWrm,     TB_ALIGN_16 },
800     { X86::PMULHWrr,        X86::PMULHWrm,      TB_ALIGN_16 },
801     { X86::PMULLDrr,        X86::PMULLDrm,      TB_ALIGN_16 },
802     { X86::PMULLWrr,        X86::PMULLWrm,      TB_ALIGN_16 },
803     { X86::PMULUDQrr,       X86::PMULUDQrm,     TB_ALIGN_16 },
804     { X86::PORrr,           X86::PORrm,         TB_ALIGN_16 },
805     { X86::PSADBWrr,        X86::PSADBWrm,      TB_ALIGN_16 },
806     { X86::PSHUFBrr,        X86::PSHUFBrm,      TB_ALIGN_16 },
807     { X86::PSIGNBrr,        X86::PSIGNBrm,      TB_ALIGN_16 },
808     { X86::PSIGNWrr,        X86::PSIGNWrm,      TB_ALIGN_16 },
809     { X86::PSIGNDrr,        X86::PSIGNDrm,      TB_ALIGN_16 },
810     { X86::PSLLDrr,         X86::PSLLDrm,       TB_ALIGN_16 },
811     { X86::PSLLQrr,         X86::PSLLQrm,       TB_ALIGN_16 },
812     { X86::PSLLWrr,         X86::PSLLWrm,       TB_ALIGN_16 },
813     { X86::PSRADrr,         X86::PSRADrm,       TB_ALIGN_16 },
814     { X86::PSRAWrr,         X86::PSRAWrm,       TB_ALIGN_16 },
815     { X86::PSRLDrr,         X86::PSRLDrm,       TB_ALIGN_16 },
816     { X86::PSRLQrr,         X86::PSRLQrm,       TB_ALIGN_16 },
817     { X86::PSRLWrr,         X86::PSRLWrm,       TB_ALIGN_16 },
818     { X86::PSUBBrr,         X86::PSUBBrm,       TB_ALIGN_16 },
819     { X86::PSUBDrr,         X86::PSUBDrm,       TB_ALIGN_16 },
820     { X86::PSUBSBrr,        X86::PSUBSBrm,      TB_ALIGN_16 },
821     { X86::PSUBSWrr,        X86::PSUBSWrm,      TB_ALIGN_16 },
822     { X86::PSUBWrr,         X86::PSUBWrm,       TB_ALIGN_16 },
823     { X86::PUNPCKHBWrr,     X86::PUNPCKHBWrm,   TB_ALIGN_16 },
824     { X86::PUNPCKHDQrr,     X86::PUNPCKHDQrm,   TB_ALIGN_16 },
825     { X86::PUNPCKHQDQrr,    X86::PUNPCKHQDQrm,  TB_ALIGN_16 },
826     { X86::PUNPCKHWDrr,     X86::PUNPCKHWDrm,   TB_ALIGN_16 },
827     { X86::PUNPCKLBWrr,     X86::PUNPCKLBWrm,   TB_ALIGN_16 },
828     { X86::PUNPCKLDQrr,     X86::PUNPCKLDQrm,   TB_ALIGN_16 },
829     { X86::PUNPCKLQDQrr,    X86::PUNPCKLQDQrm,  TB_ALIGN_16 },
830     { X86::PUNPCKLWDrr,     X86::PUNPCKLWDrm,   TB_ALIGN_16 },
831     { X86::PXORrr,          X86::PXORrm,        TB_ALIGN_16 },
832     { X86::SBB32rr,         X86::SBB32rm,       0 },
833     { X86::SBB64rr,         X86::SBB64rm,       0 },
834     { X86::SHUFPDrri,       X86::SHUFPDrmi,     TB_ALIGN_16 },
835     { X86::SHUFPSrri,       X86::SHUFPSrmi,     TB_ALIGN_16 },
836     { X86::SUB16rr,         X86::SUB16rm,       0 },
837     { X86::SUB32rr,         X86::SUB32rm,       0 },
838     { X86::SUB64rr,         X86::SUB64rm,       0 },
839     { X86::SUB8rr,          X86::SUB8rm,        0 },
840     { X86::SUBPDrr,         X86::SUBPDrm,       TB_ALIGN_16 },
841     { X86::SUBPSrr,         X86::SUBPSrm,       TB_ALIGN_16 },
842     { X86::SUBSDrr,         X86::SUBSDrm,       0 },
843     { X86::SUBSSrr,         X86::SUBSSrm,       0 },
844     // FIXME: TEST*rr -> swapped operand of TEST*mr.
845     { X86::UNPCKHPDrr,      X86::UNPCKHPDrm,    TB_ALIGN_16 },
846     { X86::UNPCKHPSrr,      X86::UNPCKHPSrm,    TB_ALIGN_16 },
847     { X86::UNPCKLPDrr,      X86::UNPCKLPDrm,    TB_ALIGN_16 },
848     { X86::UNPCKLPSrr,      X86::UNPCKLPSrm,    TB_ALIGN_16 },
849     { X86::XOR16rr,         X86::XOR16rm,       0 },
850     { X86::XOR32rr,         X86::XOR32rm,       0 },
851     { X86::XOR64rr,         X86::XOR64rm,       0 },
852     { X86::XOR8rr,          X86::XOR8rm,        0 },
853     { X86::XORPDrr,         X86::XORPDrm,       TB_ALIGN_16 },
854     { X86::XORPSrr,         X86::XORPSrm,       TB_ALIGN_16 },
855     // AVX 128-bit versions of foldable instructions
856     { X86::VCVTSD2SSrr,       X86::VCVTSD2SSrm,        0 },
857     { X86::Int_VCVTSD2SSrr,   X86::Int_VCVTSD2SSrm,    0 },
858     { X86::VCVTSI2SD64rr,     X86::VCVTSI2SD64rm,      0 },
859     { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm,  0 },
860     { X86::VCVTSI2SDrr,       X86::VCVTSI2SDrm,        0 },
861     { X86::Int_VCVTSI2SDrr,   X86::Int_VCVTSI2SDrm,    0 },
862     { X86::VCVTSI2SS64rr,     X86::VCVTSI2SS64rm,      0 },
863     { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm,  0 },
864     { X86::VCVTSI2SSrr,       X86::VCVTSI2SSrm,        0 },
865     { X86::Int_VCVTSI2SSrr,   X86::Int_VCVTSI2SSrm,    0 },
866     { X86::VCVTSS2SDrr,       X86::VCVTSS2SDrm,        0 },
867     { X86::Int_VCVTSS2SDrr,   X86::Int_VCVTSS2SDrm,    0 },
868     { X86::VCVTTPD2DQrr,      X86::VCVTTPD2DQXrm,      0 },
869     { X86::VCVTTPS2DQrr,      X86::VCVTTPS2DQrm,       0 },
870     { X86::VRSQRTSSr,         X86::VRSQRTSSm,          0 },
871     { X86::VSQRTSDr,          X86::VSQRTSDm,           0 },
872     { X86::VSQRTSSr,          X86::VSQRTSSm,           0 },
873     { X86::VADDPDrr,          X86::VADDPDrm,           0 },
874     { X86::VADDPSrr,          X86::VADDPSrm,           0 },
875     { X86::VADDSDrr,          X86::VADDSDrm,           0 },
876     { X86::VADDSSrr,          X86::VADDSSrm,           0 },
877     { X86::VADDSUBPDrr,       X86::VADDSUBPDrm,        0 },
878     { X86::VADDSUBPSrr,       X86::VADDSUBPSrm,        0 },
879     { X86::VANDNPDrr,         X86::VANDNPDrm,          0 },
880     { X86::VANDNPSrr,         X86::VANDNPSrm,          0 },
881     { X86::VANDPDrr,          X86::VANDPDrm,           0 },
882     { X86::VANDPSrr,          X86::VANDPSrm,           0 },
883     { X86::VBLENDPDrri,       X86::VBLENDPDrmi,        0 },
884     { X86::VBLENDPSrri,       X86::VBLENDPSrmi,        0 },
885     { X86::VBLENDVPDrr,       X86::VBLENDVPDrm,        0 },
886     { X86::VBLENDVPSrr,       X86::VBLENDVPSrm,        0 },
887     { X86::VCMPPDrri,         X86::VCMPPDrmi,          0 },
888     { X86::VCMPPSrri,         X86::VCMPPSrmi,          0 },
889     { X86::VCMPSDrr,          X86::VCMPSDrm,           0 },
890     { X86::VCMPSSrr,          X86::VCMPSSrm,           0 },
891     { X86::VDIVPDrr,          X86::VDIVPDrm,           0 },
892     { X86::VDIVPSrr,          X86::VDIVPSrm,           0 },
893     { X86::VDIVSDrr,          X86::VDIVSDrm,           0 },
894     { X86::VDIVSSrr,          X86::VDIVSSrm,           0 },
895     { X86::VFsANDNPDrr,       X86::VFsANDNPDrm,        TB_ALIGN_16 },
896     { X86::VFsANDNPSrr,       X86::VFsANDNPSrm,        TB_ALIGN_16 },
897     { X86::VFsANDPDrr,        X86::VFsANDPDrm,         TB_ALIGN_16 },
898     { X86::VFsANDPSrr,        X86::VFsANDPSrm,         TB_ALIGN_16 },
899     { X86::VFsORPDrr,         X86::VFsORPDrm,          TB_ALIGN_16 },
900     { X86::VFsORPSrr,         X86::VFsORPSrm,          TB_ALIGN_16 },
901     { X86::VFsXORPDrr,        X86::VFsXORPDrm,         TB_ALIGN_16 },
902     { X86::VFsXORPSrr,        X86::VFsXORPSrm,         TB_ALIGN_16 },
903     { X86::VHADDPDrr,         X86::VHADDPDrm,          0 },
904     { X86::VHADDPSrr,         X86::VHADDPSrm,          0 },
905     { X86::VHSUBPDrr,         X86::VHSUBPDrm,          0 },
906     { X86::VHSUBPSrr,         X86::VHSUBPSrm,          0 },
907     { X86::Int_VCMPSDrr,      X86::Int_VCMPSDrm,       0 },
908     { X86::Int_VCMPSSrr,      X86::Int_VCMPSSrm,       0 },
909     { X86::VMAXPDrr,          X86::VMAXPDrm,           0 },
910     { X86::VMAXPSrr,          X86::VMAXPSrm,           0 },
911     { X86::VMAXSDrr,          X86::VMAXSDrm,           0 },
912     { X86::VMAXSSrr,          X86::VMAXSSrm,           0 },
913     { X86::VMINPDrr,          X86::VMINPDrm,           0 },
914     { X86::VMINPSrr,          X86::VMINPSrm,           0 },
915     { X86::VMINSDrr,          X86::VMINSDrm,           0 },
916     { X86::VMINSSrr,          X86::VMINSSrm,           0 },
917     { X86::VMPSADBWrri,       X86::VMPSADBWrmi,        0 },
918     { X86::VMULPDrr,          X86::VMULPDrm,           0 },
919     { X86::VMULPSrr,          X86::VMULPSrm,           0 },
920     { X86::VMULSDrr,          X86::VMULSDrm,           0 },
921     { X86::VMULSSrr,          X86::VMULSSrm,           0 },
922     { X86::VORPDrr,           X86::VORPDrm,            0 },
923     { X86::VORPSrr,           X86::VORPSrm,            0 },
924     { X86::VPACKSSDWrr,       X86::VPACKSSDWrm,        0 },
925     { X86::VPACKSSWBrr,       X86::VPACKSSWBrm,        0 },
926     { X86::VPACKUSDWrr,       X86::VPACKUSDWrm,        0 },
927     { X86::VPACKUSWBrr,       X86::VPACKUSWBrm,        0 },
928     { X86::VPADDBrr,          X86::VPADDBrm,           0 },
929     { X86::VPADDDrr,          X86::VPADDDrm,           0 },
930     { X86::VPADDQrr,          X86::VPADDQrm,           0 },
931     { X86::VPADDSBrr,         X86::VPADDSBrm,          0 },
932     { X86::VPADDSWrr,         X86::VPADDSWrm,          0 },
933     { X86::VPADDUSBrr,        X86::VPADDUSBrm,         0 },
934     { X86::VPADDUSWrr,        X86::VPADDUSWrm,         0 },
935     { X86::VPADDWrr,          X86::VPADDWrm,           0 },
936     { X86::VPALIGNR128rr,     X86::VPALIGNR128rm,      0 },
937     { X86::VPANDNrr,          X86::VPANDNrm,           0 },
938     { X86::VPANDrr,           X86::VPANDrm,            0 },
939     { X86::VPAVGBrr,          X86::VPAVGBrm,           0 },
940     { X86::VPAVGWrr,          X86::VPAVGWrm,           0 },
941     { X86::VPBLENDWrri,       X86::VPBLENDWrmi,        0 },
942     { X86::VPCMPEQBrr,        X86::VPCMPEQBrm,         0 },
943     { X86::VPCMPEQDrr,        X86::VPCMPEQDrm,         0 },
944     { X86::VPCMPEQQrr,        X86::VPCMPEQQrm,         0 },
945     { X86::VPCMPEQWrr,        X86::VPCMPEQWrm,         0 },
946     { X86::VPCMPGTBrr,        X86::VPCMPGTBrm,         0 },
947     { X86::VPCMPGTDrr,        X86::VPCMPGTDrm,         0 },
948     { X86::VPCMPGTQrr,        X86::VPCMPGTQrm,         0 },
949     { X86::VPCMPGTWrr,        X86::VPCMPGTWrm,         0 },
950     { X86::VPHADDDrr,         X86::VPHADDDrm,          0 },
951     { X86::VPHADDSWrr128,     X86::VPHADDSWrm128,      0 },
952     { X86::VPHADDWrr,         X86::VPHADDWrm,          0 },
953     { X86::VPHSUBDrr,         X86::VPHSUBDrm,          0 },
954     { X86::VPHSUBSWrr128,     X86::VPHSUBSWrm128,      0 },
955     { X86::VPHSUBWrr,         X86::VPHSUBWrm,          0 },
956     { X86::VPERMILPDrr,       X86::VPERMILPDrm,        0 },
957     { X86::VPERMILPSrr,       X86::VPERMILPSrm,        0 },
958     { X86::VPINSRWrri,        X86::VPINSRWrmi,         0 },
959     { X86::VPMADDUBSWrr128,   X86::VPMADDUBSWrm128,    0 },
960     { X86::VPMADDWDrr,        X86::VPMADDWDrm,         0 },
961     { X86::VPMAXSWrr,         X86::VPMAXSWrm,          0 },
962     { X86::VPMAXUBrr,         X86::VPMAXUBrm,          0 },
963     { X86::VPMINSWrr,         X86::VPMINSWrm,          0 },
964     { X86::VPMINUBrr,         X86::VPMINUBrm,          0 },
965     { X86::VPMINSBrr,         X86::VPMINSBrm,          0 },
966     { X86::VPMINSDrr,         X86::VPMINSDrm,          0 },
967     { X86::VPMINUDrr,         X86::VPMINUDrm,          0 },
968     { X86::VPMINUWrr,         X86::VPMINUWrm,          0 },
969     { X86::VPMAXSBrr,         X86::VPMAXSBrm,          0 },
970     { X86::VPMAXSDrr,         X86::VPMAXSDrm,          0 },
971     { X86::VPMAXUDrr,         X86::VPMAXUDrm,          0 },
972     { X86::VPMAXUWrr,         X86::VPMAXUWrm,          0 },
973     { X86::VPMULDQrr,         X86::VPMULDQrm,          0 },
974     { X86::VPMULHRSWrr128,    X86::VPMULHRSWrm128,     0 },
975     { X86::VPMULHUWrr,        X86::VPMULHUWrm,         0 },
976     { X86::VPMULHWrr,         X86::VPMULHWrm,          0 },
977     { X86::VPMULLDrr,         X86::VPMULLDrm,          0 },
978     { X86::VPMULLWrr,         X86::VPMULLWrm,          0 },
979     { X86::VPMULUDQrr,        X86::VPMULUDQrm,         0 },
980     { X86::VPORrr,            X86::VPORrm,             0 },
981     { X86::VPSADBWrr,         X86::VPSADBWrm,          0 },
982     { X86::VPSHUFBrr,         X86::VPSHUFBrm,          0 },
983     { X86::VPSIGNBrr,         X86::VPSIGNBrm,          0 },
984     { X86::VPSIGNWrr,         X86::VPSIGNWrm,          0 },
985     { X86::VPSIGNDrr,         X86::VPSIGNDrm,          0 },
986     { X86::VPSLLDrr,          X86::VPSLLDrm,           0 },
987     { X86::VPSLLQrr,          X86::VPSLLQrm,           0 },
988     { X86::VPSLLWrr,          X86::VPSLLWrm,           0 },
989     { X86::VPSRADrr,          X86::VPSRADrm,           0 },
990     { X86::VPSRAWrr,          X86::VPSRAWrm,           0 },
991     { X86::VPSRLDrr,          X86::VPSRLDrm,           0 },
992     { X86::VPSRLQrr,          X86::VPSRLQrm,           0 },
993     { X86::VPSRLWrr,          X86::VPSRLWrm,           0 },
994     { X86::VPSUBBrr,          X86::VPSUBBrm,           0 },
995     { X86::VPSUBDrr,          X86::VPSUBDrm,           0 },
996     { X86::VPSUBSBrr,         X86::VPSUBSBrm,          0 },
997     { X86::VPSUBSWrr,         X86::VPSUBSWrm,          0 },
998     { X86::VPSUBWrr,          X86::VPSUBWrm,           0 },
999     { X86::VPUNPCKHBWrr,      X86::VPUNPCKHBWrm,       0 },
1000     { X86::VPUNPCKHDQrr,      X86::VPUNPCKHDQrm,       0 },
1001     { X86::VPUNPCKHQDQrr,     X86::VPUNPCKHQDQrm,      0 },
1002     { X86::VPUNPCKHWDrr,      X86::VPUNPCKHWDrm,       0 },
1003     { X86::VPUNPCKLBWrr,      X86::VPUNPCKLBWrm,       0 },
1004     { X86::VPUNPCKLDQrr,      X86::VPUNPCKLDQrm,       0 },
1005     { X86::VPUNPCKLQDQrr,     X86::VPUNPCKLQDQrm,      0 },
1006     { X86::VPUNPCKLWDrr,      X86::VPUNPCKLWDrm,       0 },
1007     { X86::VPXORrr,           X86::VPXORrm,            0 },
1008     { X86::VSHUFPDrri,        X86::VSHUFPDrmi,         0 },
1009     { X86::VSHUFPSrri,        X86::VSHUFPSrmi,         0 },
1010     { X86::VSUBPDrr,          X86::VSUBPDrm,           0 },
1011     { X86::VSUBPSrr,          X86::VSUBPSrm,           0 },
1012     { X86::VSUBSDrr,          X86::VSUBSDrm,           0 },
1013     { X86::VSUBSSrr,          X86::VSUBSSrm,           0 },
1014     { X86::VUNPCKHPDrr,       X86::VUNPCKHPDrm,        0 },
1015     { X86::VUNPCKHPSrr,       X86::VUNPCKHPSrm,        0 },
1016     { X86::VUNPCKLPDrr,       X86::VUNPCKLPDrm,        0 },
1017     { X86::VUNPCKLPSrr,       X86::VUNPCKLPSrm,        0 },
1018     { X86::VXORPDrr,          X86::VXORPDrm,           0 },
1019     { X86::VXORPSrr,          X86::VXORPSrm,           0 },
1020     // AVX 256-bit foldable instructions
1021     { X86::VADDPDYrr,         X86::VADDPDYrm,          0 },
1022     { X86::VADDPSYrr,         X86::VADDPSYrm,          0 },
1023     { X86::VADDSUBPDYrr,      X86::VADDSUBPDYrm,       0 },
1024     { X86::VADDSUBPSYrr,      X86::VADDSUBPSYrm,       0 },
1025     { X86::VANDNPDYrr,        X86::VANDNPDYrm,         0 },
1026     { X86::VANDNPSYrr,        X86::VANDNPSYrm,         0 },
1027     { X86::VANDPDYrr,         X86::VANDPDYrm,          0 },
1028     { X86::VANDPSYrr,         X86::VANDPSYrm,          0 },
1029     { X86::VBLENDPDYrri,      X86::VBLENDPDYrmi,       0 },
1030     { X86::VBLENDPSYrri,      X86::VBLENDPSYrmi,       0 },
1031     { X86::VBLENDVPDYrr,      X86::VBLENDVPDYrm,       0 },
1032     { X86::VBLENDVPSYrr,      X86::VBLENDVPSYrm,       0 },
1033     { X86::VCMPPDYrri,        X86::VCMPPDYrmi,         0 },
1034     { X86::VCMPPSYrri,        X86::VCMPPSYrmi,         0 },
1035     { X86::VDIVPDYrr,         X86::VDIVPDYrm,          0 },
1036     { X86::VDIVPSYrr,         X86::VDIVPSYrm,          0 },
1037     { X86::VHADDPDYrr,        X86::VHADDPDYrm,         0 },
1038     { X86::VHADDPSYrr,        X86::VHADDPSYrm,         0 },
1039     { X86::VHSUBPDYrr,        X86::VHSUBPDYrm,         0 },
1040     { X86::VHSUBPSYrr,        X86::VHSUBPSYrm,         0 },
1041     { X86::VINSERTF128rr,     X86::VINSERTF128rm,      0 },
1042     { X86::VMAXPDYrr,         X86::VMAXPDYrm,          0 },
1043     { X86::VMAXPSYrr,         X86::VMAXPSYrm,          0 },
1044     { X86::VMINPDYrr,         X86::VMINPDYrm,          0 },
1045     { X86::VMINPSYrr,         X86::VMINPSYrm,          0 },
1046     { X86::VMULPDYrr,         X86::VMULPDYrm,          0 },
1047     { X86::VMULPSYrr,         X86::VMULPSYrm,          0 },
1048     { X86::VORPDYrr,          X86::VORPDYrm,           0 },
1049     { X86::VORPSYrr,          X86::VORPSYrm,           0 },
1050     { X86::VPERM2F128rr,      X86::VPERM2F128rm,       0 },
1051     { X86::VPERMILPDYrr,      X86::VPERMILPDYrm,       0 },
1052     { X86::VPERMILPSYrr,      X86::VPERMILPSYrm,       0 },
1053     { X86::VSHUFPDYrri,       X86::VSHUFPDYrmi,        0 },
1054     { X86::VSHUFPSYrri,       X86::VSHUFPSYrmi,        0 },
1055     { X86::VSUBPDYrr,         X86::VSUBPDYrm,          0 },
1056     { X86::VSUBPSYrr,         X86::VSUBPSYrm,          0 },
1057     { X86::VUNPCKHPDYrr,      X86::VUNPCKHPDYrm,       0 },
1058     { X86::VUNPCKHPSYrr,      X86::VUNPCKHPSYrm,       0 },
1059     { X86::VUNPCKLPDYrr,      X86::VUNPCKLPDYrm,       0 },
1060     { X86::VUNPCKLPSYrr,      X86::VUNPCKLPSYrm,       0 },
1061     { X86::VXORPDYrr,         X86::VXORPDYrm,          0 },
1062     { X86::VXORPSYrr,         X86::VXORPSYrm,          0 },
1063     // AVX2 foldable instructions
1064     { X86::VINSERTI128rr,     X86::VINSERTI128rm,      0 },
1065     { X86::VPACKSSDWYrr,      X86::VPACKSSDWYrm,       0 },
1066     { X86::VPACKSSWBYrr,      X86::VPACKSSWBYrm,       0 },
1067     { X86::VPACKUSDWYrr,      X86::VPACKUSDWYrm,       0 },
1068     { X86::VPACKUSWBYrr,      X86::VPACKUSWBYrm,       0 },
1069     { X86::VPADDBYrr,         X86::VPADDBYrm,          0 },
1070     { X86::VPADDDYrr,         X86::VPADDDYrm,          0 },
1071     { X86::VPADDQYrr,         X86::VPADDQYrm,          0 },
1072     { X86::VPADDSBYrr,        X86::VPADDSBYrm,         0 },
1073     { X86::VPADDSWYrr,        X86::VPADDSWYrm,         0 },
1074     { X86::VPADDUSBYrr,       X86::VPADDUSBYrm,        0 },
1075     { X86::VPADDUSWYrr,       X86::VPADDUSWYrm,        0 },
1076     { X86::VPADDWYrr,         X86::VPADDWYrm,          0 },
1077     { X86::VPALIGNR256rr,     X86::VPALIGNR256rm,      0 },
1078     { X86::VPANDNYrr,         X86::VPANDNYrm,          0 },
1079     { X86::VPANDYrr,          X86::VPANDYrm,           0 },
1080     { X86::VPAVGBYrr,         X86::VPAVGBYrm,          0 },
1081     { X86::VPAVGWYrr,         X86::VPAVGWYrm,          0 },
1082     { X86::VPBLENDDrri,       X86::VPBLENDDrmi,        0 },
1083     { X86::VPBLENDDYrri,      X86::VPBLENDDYrmi,       0 },
1084     { X86::VPBLENDWYrri,      X86::VPBLENDWYrmi,       0 },
1085     { X86::VPCMPEQBYrr,       X86::VPCMPEQBYrm,        0 },
1086     { X86::VPCMPEQDYrr,       X86::VPCMPEQDYrm,        0 },
1087     { X86::VPCMPEQQYrr,       X86::VPCMPEQQYrm,        0 },
1088     { X86::VPCMPEQWYrr,       X86::VPCMPEQWYrm,        0 },
1089     { X86::VPCMPGTBYrr,       X86::VPCMPGTBYrm,        0 },
1090     { X86::VPCMPGTDYrr,       X86::VPCMPGTDYrm,        0 },
1091     { X86::VPCMPGTQYrr,       X86::VPCMPGTQYrm,        0 },
1092     { X86::VPCMPGTWYrr,       X86::VPCMPGTWYrm,        0 },
1093     { X86::VPERM2I128rr,      X86::VPERM2I128rm,       0 },
1094     { X86::VPERMDYrr,         X86::VPERMDYrm,          0 },
1095     { X86::VPERMPDYri,        X86::VPERMPDYmi,         0 },
1096     { X86::VPERMPSYrr,        X86::VPERMPSYrm,         0 },
1097     { X86::VPERMQYri,         X86::VPERMQYmi,          0 },
1098     { X86::VPHADDDYrr,        X86::VPHADDDYrm,         0 },
1099     { X86::VPHADDSWrr256,     X86::VPHADDSWrm256,      0 },
1100     { X86::VPHADDWYrr,        X86::VPHADDWYrm,         0 },
1101     { X86::VPHSUBDYrr,        X86::VPHSUBDYrm,         0 },
1102     { X86::VPHSUBSWrr256,     X86::VPHSUBSWrm256,      0 },
1103     { X86::VPHSUBWYrr,        X86::VPHSUBWYrm,         0 },
1104     { X86::VPMADDUBSWrr256,   X86::VPMADDUBSWrm256,    0 },
1105     { X86::VPMADDWDYrr,       X86::VPMADDWDYrm,        0 },
1106     { X86::VPMAXSWYrr,        X86::VPMAXSWYrm,         0 },
1107     { X86::VPMAXUBYrr,        X86::VPMAXUBYrm,         0 },
1108     { X86::VPMINSWYrr,        X86::VPMINSWYrm,         0 },
1109     { X86::VPMINUBYrr,        X86::VPMINUBYrm,         0 },
1110     { X86::VPMINSBYrr,        X86::VPMINSBYrm,         0 },
1111     { X86::VPMINSDYrr,        X86::VPMINSDYrm,         0 },
1112     { X86::VPMINUDYrr,        X86::VPMINUDYrm,         0 },
1113     { X86::VPMINUWYrr,        X86::VPMINUWYrm,         0 },
1114     { X86::VPMAXSBYrr,        X86::VPMAXSBYrm,         0 },
1115     { X86::VPMAXSDYrr,        X86::VPMAXSDYrm,         0 },
1116     { X86::VPMAXUDYrr,        X86::VPMAXUDYrm,         0 },
1117     { X86::VPMAXUWYrr,        X86::VPMAXUWYrm,         0 },
1118     { X86::VMPSADBWYrri,      X86::VMPSADBWYrmi,       0 },
1119     { X86::VPMULDQYrr,        X86::VPMULDQYrm,         0 },
1120     { X86::VPMULHRSWrr256,    X86::VPMULHRSWrm256,     0 },
1121     { X86::VPMULHUWYrr,       X86::VPMULHUWYrm,        0 },
1122     { X86::VPMULHWYrr,        X86::VPMULHWYrm,         0 },
1123     { X86::VPMULLDYrr,        X86::VPMULLDYrm,         0 },
1124     { X86::VPMULLWYrr,        X86::VPMULLWYrm,         0 },
1125     { X86::VPMULUDQYrr,       X86::VPMULUDQYrm,        0 },
1126     { X86::VPORYrr,           X86::VPORYrm,            0 },
1127     { X86::VPSADBWYrr,        X86::VPSADBWYrm,         0 },
1128     { X86::VPSHUFBYrr,        X86::VPSHUFBYrm,         0 },
1129     { X86::VPSIGNBYrr,        X86::VPSIGNBYrm,         0 },
1130     { X86::VPSIGNWYrr,        X86::VPSIGNWYrm,         0 },
1131     { X86::VPSIGNDYrr,        X86::VPSIGNDYrm,         0 },
1132     { X86::VPSLLDYrr,         X86::VPSLLDYrm,          0 },
1133     { X86::VPSLLQYrr,         X86::VPSLLQYrm,          0 },
1134     { X86::VPSLLWYrr,         X86::VPSLLWYrm,          0 },
1135     { X86::VPSLLVDrr,         X86::VPSLLVDrm,          0 },
1136     { X86::VPSLLVDYrr,        X86::VPSLLVDYrm,         0 },
1137     { X86::VPSLLVQrr,         X86::VPSLLVQrm,          0 },
1138     { X86::VPSLLVQYrr,        X86::VPSLLVQYrm,         0 },
1139     { X86::VPSRADYrr,         X86::VPSRADYrm,          0 },
1140     { X86::VPSRAWYrr,         X86::VPSRAWYrm,          0 },
1141     { X86::VPSRAVDrr,         X86::VPSRAVDrm,          0 },
1142     { X86::VPSRAVDYrr,        X86::VPSRAVDYrm,         0 },
1143     { X86::VPSRLDYrr,         X86::VPSRLDYrm,          0 },
1144     { X86::VPSRLQYrr,         X86::VPSRLQYrm,          0 },
1145     { X86::VPSRLWYrr,         X86::VPSRLWYrm,          0 },
1146     { X86::VPSRLVDrr,         X86::VPSRLVDrm,          0 },
1147     { X86::VPSRLVDYrr,        X86::VPSRLVDYrm,         0 },
1148     { X86::VPSRLVQrr,         X86::VPSRLVQrm,          0 },
1149     { X86::VPSRLVQYrr,        X86::VPSRLVQYrm,         0 },
1150     { X86::VPSUBBYrr,         X86::VPSUBBYrm,          0 },
1151     { X86::VPSUBDYrr,         X86::VPSUBDYrm,          0 },
1152     { X86::VPSUBSBYrr,        X86::VPSUBSBYrm,         0 },
1153     { X86::VPSUBSWYrr,        X86::VPSUBSWYrm,         0 },
1154     { X86::VPSUBWYrr,         X86::VPSUBWYrm,          0 },
1155     { X86::VPUNPCKHBWYrr,     X86::VPUNPCKHBWYrm,      0 },
1156     { X86::VPUNPCKHDQYrr,     X86::VPUNPCKHDQYrm,      0 },
1157     { X86::VPUNPCKHQDQYrr,    X86::VPUNPCKHQDQYrm,     0 },
1158     { X86::VPUNPCKHWDYrr,     X86::VPUNPCKHWDYrm,      0 },
1159     { X86::VPUNPCKLBWYrr,     X86::VPUNPCKLBWYrm,      0 },
1160     { X86::VPUNPCKLDQYrr,     X86::VPUNPCKLDQYrm,      0 },
1161     { X86::VPUNPCKLQDQYrr,    X86::VPUNPCKLQDQYrm,     0 },
1162     { X86::VPUNPCKLWDYrr,     X86::VPUNPCKLWDYrm,      0 },
1163     { X86::VPXORYrr,          X86::VPXORYrm,           0 },
1164     // FIXME: add AVX 256-bit foldable instructions
1165 
1166     // FMA4 foldable patterns
1167     { X86::VFMADDSS4rr,       X86::VFMADDSS4mr,        0           },
1168     { X86::VFMADDSD4rr,       X86::VFMADDSD4mr,        0           },
1169     { X86::VFMADDPS4rr,       X86::VFMADDPS4mr,        TB_ALIGN_16 },
1170     { X86::VFMADDPD4rr,       X86::VFMADDPD4mr,        TB_ALIGN_16 },
1171     { X86::VFMADDPS4rrY,      X86::VFMADDPS4mrY,       TB_ALIGN_32 },
1172     { X86::VFMADDPD4rrY,      X86::VFMADDPD4mrY,       TB_ALIGN_32 },
1173     { X86::VFNMADDSS4rr,      X86::VFNMADDSS4mr,       0           },
1174     { X86::VFNMADDSD4rr,      X86::VFNMADDSD4mr,       0           },
1175     { X86::VFNMADDPS4rr,      X86::VFNMADDPS4mr,       TB_ALIGN_16 },
1176     { X86::VFNMADDPD4rr,      X86::VFNMADDPD4mr,       TB_ALIGN_16 },
1177     { X86::VFNMADDPS4rrY,     X86::VFNMADDPS4mrY,      TB_ALIGN_32 },
1178     { X86::VFNMADDPD4rrY,     X86::VFNMADDPD4mrY,      TB_ALIGN_32 },
1179     { X86::VFMSUBSS4rr,       X86::VFMSUBSS4mr,        0           },
1180     { X86::VFMSUBSD4rr,       X86::VFMSUBSD4mr,        0           },
1181     { X86::VFMSUBPS4rr,       X86::VFMSUBPS4mr,        TB_ALIGN_16 },
1182     { X86::VFMSUBPD4rr,       X86::VFMSUBPD4mr,        TB_ALIGN_16 },
1183     { X86::VFMSUBPS4rrY,      X86::VFMSUBPS4mrY,       TB_ALIGN_32 },
1184     { X86::VFMSUBPD4rrY,      X86::VFMSUBPD4mrY,       TB_ALIGN_32 },
1185     { X86::VFNMSUBSS4rr,      X86::VFNMSUBSS4mr,       0           },
1186     { X86::VFNMSUBSD4rr,      X86::VFNMSUBSD4mr,       0           },
1187     { X86::VFNMSUBPS4rr,      X86::VFNMSUBPS4mr,       TB_ALIGN_16 },
1188     { X86::VFNMSUBPD4rr,      X86::VFNMSUBPD4mr,       TB_ALIGN_16 },
1189     { X86::VFNMSUBPS4rrY,     X86::VFNMSUBPS4mrY,      TB_ALIGN_32 },
1190     { X86::VFNMSUBPD4rrY,     X86::VFNMSUBPD4mrY,      TB_ALIGN_32 },
1191     { X86::VFMADDSUBPS4rr,    X86::VFMADDSUBPS4mr,     TB_ALIGN_16 },
1192     { X86::VFMADDSUBPD4rr,    X86::VFMADDSUBPD4mr,     TB_ALIGN_16 },
1193     { X86::VFMADDSUBPS4rrY,   X86::VFMADDSUBPS4mrY,    TB_ALIGN_32 },
1194     { X86::VFMADDSUBPD4rrY,   X86::VFMADDSUBPD4mrY,    TB_ALIGN_32 },
1195     { X86::VFMSUBADDPS4rr,    X86::VFMSUBADDPS4mr,     TB_ALIGN_16 },
1196     { X86::VFMSUBADDPD4rr,    X86::VFMSUBADDPD4mr,     TB_ALIGN_16 },
1197     { X86::VFMSUBADDPS4rrY,   X86::VFMSUBADDPS4mrY,    TB_ALIGN_32 },
1198     { X86::VFMSUBADDPD4rrY,   X86::VFMSUBADDPD4mrY,    TB_ALIGN_32 },
1199 
1200     // BMI/BMI2 foldable instructions
1201     { X86::ANDN32rr,          X86::ANDN32rm,            0 },
1202     { X86::ANDN64rr,          X86::ANDN64rm,            0 },
1203     { X86::MULX32rr,          X86::MULX32rm,            0 },
1204     { X86::MULX64rr,          X86::MULX64rm,            0 },
1205     { X86::PDEP32rr,          X86::PDEP32rm,            0 },
1206     { X86::PDEP64rr,          X86::PDEP64rm,            0 },
1207     { X86::PEXT32rr,          X86::PEXT32rm,            0 },
1208     { X86::PEXT64rr,          X86::PEXT64rm,            0 },
1209 
1210     // AVX-512 foldable instructions
1211     { X86::VPADDDZrr,         X86::VPADDDZrm,           0 },
1212     { X86::VPADDQZrr,         X86::VPADDQZrm,           0 },
1213     { X86::VADDPSZrr,         X86::VADDPSZrm,           0 },
1214     { X86::VADDPDZrr,         X86::VADDPDZrm,           0 },
1215     { X86::VSUBPSZrr,         X86::VSUBPSZrm,           0 },
1216     { X86::VSUBPDZrr,         X86::VSUBPDZrm,           0 },
1217     { X86::VMULPSZrr,         X86::VMULPSZrm,           0 },
1218     { X86::VMULPDZrr,         X86::VMULPDZrm,           0 },
1219     { X86::VDIVPSZrr,         X86::VDIVPSZrm,           0 },
1220     { X86::VDIVPDZrr,         X86::VDIVPDZrm,           0 },
1221     { X86::VMINPSZrr,         X86::VMINPSZrm,           0 },
1222     { X86::VMINPDZrr,         X86::VMINPDZrm,           0 },
1223     { X86::VMAXPSZrr,         X86::VMAXPSZrm,           0 },
1224     { X86::VMAXPDZrr,         X86::VMAXPDZrm,           0 },
1225     { X86::VPERMPDZri,        X86::VPERMPDZmi,          0 },
1226     { X86::VPERMPSZrr,        X86::VPERMPSZrm,          0 },
1227     { X86::VPSLLVDZrr,        X86::VPSLLVDZrm,          0 },
1228     { X86::VPSLLVQZrr,        X86::VPSLLVQZrm,          0 },
1229     { X86::VPSRAVDZrr,        X86::VPSRAVDZrm,          0 },
1230     { X86::VPSRLVDZrr,        X86::VPSRLVDZrm,          0 },
1231     { X86::VPSRLVQZrr,        X86::VPSRLVQZrm,          0 },
1232     { X86::VSHUFPDZrri,       X86::VSHUFPDZrmi,         0 },
1233     { X86::VSHUFPSZrri,       X86::VSHUFPSZrmi,         0 },
1234     { X86::VALIGNQrri,        X86::VALIGNQrmi,          0 },
1235     { X86::VALIGNDrri,        X86::VALIGNDrmi,          0 },
1236 
1237     // AES foldable instructions
1238     { X86::AESDECLASTrr,      X86::AESDECLASTrm,        TB_ALIGN_16 },
1239     { X86::AESDECrr,          X86::AESDECrm,            TB_ALIGN_16 },
1240     { X86::AESENCLASTrr,      X86::AESENCLASTrm,        TB_ALIGN_16 },
1241     { X86::AESENCrr,          X86::AESENCrm,            TB_ALIGN_16 },
1242     { X86::VAESDECLASTrr,     X86::VAESDECLASTrm,       TB_ALIGN_16 },
1243     { X86::VAESDECrr,         X86::VAESDECrm,           TB_ALIGN_16 },
1244     { X86::VAESENCLASTrr,     X86::VAESENCLASTrm,       TB_ALIGN_16 },
1245     { X86::VAESENCrr,         X86::VAESENCrm,           TB_ALIGN_16 },
1246 
1247     // SHA foldable instructions
1248     { X86::SHA1MSG1rr,        X86::SHA1MSG1rm,          TB_ALIGN_16 },
1249     { X86::SHA1MSG2rr,        X86::SHA1MSG2rm,          TB_ALIGN_16 },
1250     { X86::SHA1NEXTErr,       X86::SHA1NEXTErm,         TB_ALIGN_16 },
1251     { X86::SHA1RNDS4rri,      X86::SHA1RNDS4rmi,        TB_ALIGN_16 },
1252     { X86::SHA256MSG1rr,      X86::SHA256MSG1rm,        TB_ALIGN_16 },
1253     { X86::SHA256MSG2rr,      X86::SHA256MSG2rm,        TB_ALIGN_16 },
1254     { X86::SHA256RNDS2rr,     X86::SHA256RNDS2rm,       TB_ALIGN_16 },
1255   };
1256 
1257   for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
1258     unsigned RegOp = OpTbl2[i].RegOp;
1259     unsigned MemOp = OpTbl2[i].MemOp;
1260     unsigned Flags = OpTbl2[i].Flags;
1261     AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1262                   RegOp, MemOp,
1263                   // Index 2, folded load
1264                   Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
1265   }
1266 
1267   static const X86OpTblEntry OpTbl3[] = {
1268     // FMA foldable instructions
1269     { X86::VFMADDSSr231r,         X86::VFMADDSSr231m,         0 },
1270     { X86::VFMADDSDr231r,         X86::VFMADDSDr231m,         0 },
1271     { X86::VFMADDSSr132r,         X86::VFMADDSSr132m,         0 },
1272     { X86::VFMADDSDr132r,         X86::VFMADDSDr132m,         0 },
1273     { X86::VFMADDSSr213r,         X86::VFMADDSSr213m,         0 },
1274     { X86::VFMADDSDr213r,         X86::VFMADDSDr213m,         0 },
1275     { X86::VFMADDSSr213r_Int,     X86::VFMADDSSr213m_Int,     0 },
1276     { X86::VFMADDSDr213r_Int,     X86::VFMADDSDr213m_Int,     0 },
1277 
1278     { X86::VFMADDPSr231r,         X86::VFMADDPSr231m,         TB_ALIGN_16 },
1279     { X86::VFMADDPDr231r,         X86::VFMADDPDr231m,         TB_ALIGN_16 },
1280     { X86::VFMADDPSr132r,         X86::VFMADDPSr132m,         TB_ALIGN_16 },
1281     { X86::VFMADDPDr132r,         X86::VFMADDPDr132m,         TB_ALIGN_16 },
1282     { X86::VFMADDPSr213r,         X86::VFMADDPSr213m,         TB_ALIGN_16 },
1283     { X86::VFMADDPDr213r,         X86::VFMADDPDr213m,         TB_ALIGN_16 },
1284     { X86::VFMADDPSr231rY,        X86::VFMADDPSr231mY,        TB_ALIGN_32 },
1285     { X86::VFMADDPDr231rY,        X86::VFMADDPDr231mY,        TB_ALIGN_32 },
1286     { X86::VFMADDPSr132rY,        X86::VFMADDPSr132mY,        TB_ALIGN_32 },
1287     { X86::VFMADDPDr132rY,        X86::VFMADDPDr132mY,        TB_ALIGN_32 },
1288     { X86::VFMADDPSr213rY,        X86::VFMADDPSr213mY,        TB_ALIGN_32 },
1289     { X86::VFMADDPDr213rY,        X86::VFMADDPDr213mY,        TB_ALIGN_32 },
1290 
1291     { X86::VFNMADDSSr231r,        X86::VFNMADDSSr231m,        0 },
1292     { X86::VFNMADDSDr231r,        X86::VFNMADDSDr231m,        0 },
1293     { X86::VFNMADDSSr132r,        X86::VFNMADDSSr132m,        0 },
1294     { X86::VFNMADDSDr132r,        X86::VFNMADDSDr132m,        0 },
1295     { X86::VFNMADDSSr213r,        X86::VFNMADDSSr213m,        0 },
1296     { X86::VFNMADDSDr213r,        X86::VFNMADDSDr213m,        0 },
1297     { X86::VFNMADDSSr213r_Int,    X86::VFNMADDSSr213m_Int,    0 },
1298     { X86::VFNMADDSDr213r_Int,    X86::VFNMADDSDr213m_Int,    0 },
1299 
1300     { X86::VFNMADDPSr231r,        X86::VFNMADDPSr231m,        TB_ALIGN_16 },
1301     { X86::VFNMADDPDr231r,        X86::VFNMADDPDr231m,        TB_ALIGN_16 },
1302     { X86::VFNMADDPSr132r,        X86::VFNMADDPSr132m,        TB_ALIGN_16 },
1303     { X86::VFNMADDPDr132r,        X86::VFNMADDPDr132m,        TB_ALIGN_16 },
1304     { X86::VFNMADDPSr213r,        X86::VFNMADDPSr213m,        TB_ALIGN_16 },
1305     { X86::VFNMADDPDr213r,        X86::VFNMADDPDr213m,        TB_ALIGN_16 },
1306     { X86::VFNMADDPSr231rY,       X86::VFNMADDPSr231mY,       TB_ALIGN_32 },
1307     { X86::VFNMADDPDr231rY,       X86::VFNMADDPDr231mY,       TB_ALIGN_32 },
1308     { X86::VFNMADDPSr132rY,       X86::VFNMADDPSr132mY,       TB_ALIGN_32 },
1309     { X86::VFNMADDPDr132rY,       X86::VFNMADDPDr132mY,       TB_ALIGN_32 },
1310     { X86::VFNMADDPSr213rY,       X86::VFNMADDPSr213mY,       TB_ALIGN_32 },
1311     { X86::VFNMADDPDr213rY,       X86::VFNMADDPDr213mY,       TB_ALIGN_32 },
1312 
1313     { X86::VFMSUBSSr231r,         X86::VFMSUBSSr231m,         0 },
1314     { X86::VFMSUBSDr231r,         X86::VFMSUBSDr231m,         0 },
1315     { X86::VFMSUBSSr132r,         X86::VFMSUBSSr132m,         0 },
1316     { X86::VFMSUBSDr132r,         X86::VFMSUBSDr132m,         0 },
1317     { X86::VFMSUBSSr213r,         X86::VFMSUBSSr213m,         0 },
1318     { X86::VFMSUBSDr213r,         X86::VFMSUBSDr213m,         0 },
1319     { X86::VFMSUBSSr213r_Int,     X86::VFMSUBSSr213m_Int,     0 },
1320     { X86::VFMSUBSDr213r_Int,     X86::VFMSUBSDr213m_Int,     0 },
1321 
1322     { X86::VFMSUBPSr231r,         X86::VFMSUBPSr231m,         TB_ALIGN_16 },
1323     { X86::VFMSUBPDr231r,         X86::VFMSUBPDr231m,         TB_ALIGN_16 },
1324     { X86::VFMSUBPSr132r,         X86::VFMSUBPSr132m,         TB_ALIGN_16 },
1325     { X86::VFMSUBPDr132r,         X86::VFMSUBPDr132m,         TB_ALIGN_16 },
1326     { X86::VFMSUBPSr213r,         X86::VFMSUBPSr213m,         TB_ALIGN_16 },
1327     { X86::VFMSUBPDr213r,         X86::VFMSUBPDr213m,         TB_ALIGN_16 },
1328     { X86::VFMSUBPSr231rY,        X86::VFMSUBPSr231mY,        TB_ALIGN_32 },
1329     { X86::VFMSUBPDr231rY,        X86::VFMSUBPDr231mY,        TB_ALIGN_32 },
1330     { X86::VFMSUBPSr132rY,        X86::VFMSUBPSr132mY,        TB_ALIGN_32 },
1331     { X86::VFMSUBPDr132rY,        X86::VFMSUBPDr132mY,        TB_ALIGN_32 },
1332     { X86::VFMSUBPSr213rY,        X86::VFMSUBPSr213mY,        TB_ALIGN_32 },
1333     { X86::VFMSUBPDr213rY,        X86::VFMSUBPDr213mY,        TB_ALIGN_32 },
1334 
1335     { X86::VFNMSUBSSr231r,        X86::VFNMSUBSSr231m,        0 },
1336     { X86::VFNMSUBSDr231r,        X86::VFNMSUBSDr231m,        0 },
1337     { X86::VFNMSUBSSr132r,        X86::VFNMSUBSSr132m,        0 },
1338     { X86::VFNMSUBSDr132r,        X86::VFNMSUBSDr132m,        0 },
1339     { X86::VFNMSUBSSr213r,        X86::VFNMSUBSSr213m,        0 },
1340     { X86::VFNMSUBSDr213r,        X86::VFNMSUBSDr213m,        0 },
1341     { X86::VFNMSUBSSr213r_Int,    X86::VFNMSUBSSr213m_Int,    0 },
1342     { X86::VFNMSUBSDr213r_Int,    X86::VFNMSUBSDr213m_Int,    0 },
1343 
1344     { X86::VFNMSUBPSr231r,        X86::VFNMSUBPSr231m,        TB_ALIGN_16 },
1345     { X86::VFNMSUBPDr231r,        X86::VFNMSUBPDr231m,        TB_ALIGN_16 },
1346     { X86::VFNMSUBPSr132r,        X86::VFNMSUBPSr132m,        TB_ALIGN_16 },
1347     { X86::VFNMSUBPDr132r,        X86::VFNMSUBPDr132m,        TB_ALIGN_16 },
1348     { X86::VFNMSUBPSr213r,        X86::VFNMSUBPSr213m,        TB_ALIGN_16 },
1349     { X86::VFNMSUBPDr213r,        X86::VFNMSUBPDr213m,        TB_ALIGN_16 },
1350     { X86::VFNMSUBPSr231rY,       X86::VFNMSUBPSr231mY,       TB_ALIGN_32 },
1351     { X86::VFNMSUBPDr231rY,       X86::VFNMSUBPDr231mY,       TB_ALIGN_32 },
1352     { X86::VFNMSUBPSr132rY,       X86::VFNMSUBPSr132mY,       TB_ALIGN_32 },
1353     { X86::VFNMSUBPDr132rY,       X86::VFNMSUBPDr132mY,       TB_ALIGN_32 },
1354     { X86::VFNMSUBPSr213rY,       X86::VFNMSUBPSr213mY,       TB_ALIGN_32 },
1355     { X86::VFNMSUBPDr213rY,       X86::VFNMSUBPDr213mY,       TB_ALIGN_32 },
1356 
1357     { X86::VFMADDSUBPSr231r,      X86::VFMADDSUBPSr231m,      TB_ALIGN_16 },
1358     { X86::VFMADDSUBPDr231r,      X86::VFMADDSUBPDr231m,      TB_ALIGN_16 },
1359     { X86::VFMADDSUBPSr132r,      X86::VFMADDSUBPSr132m,      TB_ALIGN_16 },
1360     { X86::VFMADDSUBPDr132r,      X86::VFMADDSUBPDr132m,      TB_ALIGN_16 },
1361     { X86::VFMADDSUBPSr213r,      X86::VFMADDSUBPSr213m,      TB_ALIGN_16 },
1362     { X86::VFMADDSUBPDr213r,      X86::VFMADDSUBPDr213m,      TB_ALIGN_16 },
1363     { X86::VFMADDSUBPSr231rY,     X86::VFMADDSUBPSr231mY,     TB_ALIGN_32 },
1364     { X86::VFMADDSUBPDr231rY,     X86::VFMADDSUBPDr231mY,     TB_ALIGN_32 },
1365     { X86::VFMADDSUBPSr132rY,     X86::VFMADDSUBPSr132mY,     TB_ALIGN_32 },
1366     { X86::VFMADDSUBPDr132rY,     X86::VFMADDSUBPDr132mY,     TB_ALIGN_32 },
1367     { X86::VFMADDSUBPSr213rY,     X86::VFMADDSUBPSr213mY,     TB_ALIGN_32 },
1368     { X86::VFMADDSUBPDr213rY,     X86::VFMADDSUBPDr213mY,     TB_ALIGN_32 },
1369 
1370     { X86::VFMSUBADDPSr231r,      X86::VFMSUBADDPSr231m,      TB_ALIGN_16 },
1371     { X86::VFMSUBADDPDr231r,      X86::VFMSUBADDPDr231m,      TB_ALIGN_16 },
1372     { X86::VFMSUBADDPSr132r,      X86::VFMSUBADDPSr132m,      TB_ALIGN_16 },
1373     { X86::VFMSUBADDPDr132r,      X86::VFMSUBADDPDr132m,      TB_ALIGN_16 },
1374     { X86::VFMSUBADDPSr213r,      X86::VFMSUBADDPSr213m,      TB_ALIGN_16 },
1375     { X86::VFMSUBADDPDr213r,      X86::VFMSUBADDPDr213m,      TB_ALIGN_16 },
1376     { X86::VFMSUBADDPSr231rY,     X86::VFMSUBADDPSr231mY,     TB_ALIGN_32 },
1377     { X86::VFMSUBADDPDr231rY,     X86::VFMSUBADDPDr231mY,     TB_ALIGN_32 },
1378     { X86::VFMSUBADDPSr132rY,     X86::VFMSUBADDPSr132mY,     TB_ALIGN_32 },
1379     { X86::VFMSUBADDPDr132rY,     X86::VFMSUBADDPDr132mY,     TB_ALIGN_32 },
1380     { X86::VFMSUBADDPSr213rY,     X86::VFMSUBADDPSr213mY,     TB_ALIGN_32 },
1381     { X86::VFMSUBADDPDr213rY,     X86::VFMSUBADDPDr213mY,     TB_ALIGN_32 },
1382 
1383     // FMA4 foldable patterns
1384     { X86::VFMADDSS4rr,           X86::VFMADDSS4rm,           0           },
1385     { X86::VFMADDSD4rr,           X86::VFMADDSD4rm,           0           },
1386     { X86::VFMADDPS4rr,           X86::VFMADDPS4rm,           TB_ALIGN_16 },
1387     { X86::VFMADDPD4rr,           X86::VFMADDPD4rm,           TB_ALIGN_16 },
1388     { X86::VFMADDPS4rrY,          X86::VFMADDPS4rmY,          TB_ALIGN_32 },
1389     { X86::VFMADDPD4rrY,          X86::VFMADDPD4rmY,          TB_ALIGN_32 },
1390     { X86::VFNMADDSS4rr,          X86::VFNMADDSS4rm,          0           },
1391     { X86::VFNMADDSD4rr,          X86::VFNMADDSD4rm,          0           },
1392     { X86::VFNMADDPS4rr,          X86::VFNMADDPS4rm,          TB_ALIGN_16 },
1393     { X86::VFNMADDPD4rr,          X86::VFNMADDPD4rm,          TB_ALIGN_16 },
1394     { X86::VFNMADDPS4rrY,         X86::VFNMADDPS4rmY,         TB_ALIGN_32 },
1395     { X86::VFNMADDPD4rrY,         X86::VFNMADDPD4rmY,         TB_ALIGN_32 },
1396     { X86::VFMSUBSS4rr,           X86::VFMSUBSS4rm,           0           },
1397     { X86::VFMSUBSD4rr,           X86::VFMSUBSD4rm,           0           },
1398     { X86::VFMSUBPS4rr,           X86::VFMSUBPS4rm,           TB_ALIGN_16 },
1399     { X86::VFMSUBPD4rr,           X86::VFMSUBPD4rm,           TB_ALIGN_16 },
1400     { X86::VFMSUBPS4rrY,          X86::VFMSUBPS4rmY,          TB_ALIGN_32 },
1401     { X86::VFMSUBPD4rrY,          X86::VFMSUBPD4rmY,          TB_ALIGN_32 },
1402     { X86::VFNMSUBSS4rr,          X86::VFNMSUBSS4rm,          0           },
1403     { X86::VFNMSUBSD4rr,          X86::VFNMSUBSD4rm,          0           },
1404     { X86::VFNMSUBPS4rr,          X86::VFNMSUBPS4rm,          TB_ALIGN_16 },
1405     { X86::VFNMSUBPD4rr,          X86::VFNMSUBPD4rm,          TB_ALIGN_16 },
1406     { X86::VFNMSUBPS4rrY,         X86::VFNMSUBPS4rmY,         TB_ALIGN_32 },
1407     { X86::VFNMSUBPD4rrY,         X86::VFNMSUBPD4rmY,         TB_ALIGN_32 },
1408     { X86::VFMADDSUBPS4rr,        X86::VFMADDSUBPS4rm,        TB_ALIGN_16 },
1409     { X86::VFMADDSUBPD4rr,        X86::VFMADDSUBPD4rm,        TB_ALIGN_16 },
1410     { X86::VFMADDSUBPS4rrY,       X86::VFMADDSUBPS4rmY,       TB_ALIGN_32 },
1411     { X86::VFMADDSUBPD4rrY,       X86::VFMADDSUBPD4rmY,       TB_ALIGN_32 },
1412     { X86::VFMSUBADDPS4rr,        X86::VFMSUBADDPS4rm,        TB_ALIGN_16 },
1413     { X86::VFMSUBADDPD4rr,        X86::VFMSUBADDPD4rm,        TB_ALIGN_16 },
1414     { X86::VFMSUBADDPS4rrY,       X86::VFMSUBADDPS4rmY,       TB_ALIGN_32 },
1415     { X86::VFMSUBADDPD4rrY,       X86::VFMSUBADDPD4rmY,       TB_ALIGN_32 },
1416     // AVX-512 VPERMI instructions with 3 source operands.
1417     { X86::VPERMI2Drr,            X86::VPERMI2Drm,            0 },
1418     { X86::VPERMI2Qrr,            X86::VPERMI2Qrm,            0 },
1419     { X86::VPERMI2PSrr,           X86::VPERMI2PSrm,           0 },
1420     { X86::VPERMI2PDrr,           X86::VPERMI2PDrm,           0 },
1421   };
1422 
1423   for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1424     unsigned RegOp = OpTbl3[i].RegOp;
1425     unsigned MemOp = OpTbl3[i].MemOp;
1426     unsigned Flags = OpTbl3[i].Flags;
1427     AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1428                   RegOp, MemOp,
1429                   // Index 3, folded load
1430                   Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1431   }
1432 
1433 }
1434 
1435 void
1436 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1437                             MemOp2RegOpTableType &M2RTable,
1438                             unsigned RegOp, unsigned MemOp, unsigned Flags) {
1439     if ((Flags & TB_NO_FORWARD) == 0) {
1440       assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1441       R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1442     }
1443     if ((Flags & TB_NO_REVERSE) == 0) {
1444       assert(!M2RTable.count(MemOp) &&
1445            "Duplicated entries in unfolding maps?");
1446       M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1447     }
1448 }
1449 
1450 bool
1451 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1452                                     unsigned &SrcReg, unsigned &DstReg,
1453                                     unsigned &SubIdx) const {
1454   switch (MI.getOpcode()) {
1455   default: break;
1456   case X86::MOVSX16rr8:
1457   case X86::MOVZX16rr8:
1458   case X86::MOVSX32rr8:
1459   case X86::MOVZX32rr8:
1460   case X86::MOVSX64rr8:
1461     if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1462       // It's not always legal to reference the low 8-bit of the larger
1463       // register in 32-bit mode.
1464       return false;
1465   case X86::MOVSX32rr16:
1466   case X86::MOVZX32rr16:
1467   case X86::MOVSX64rr16:
1468   case X86::MOVSX64rr32: {
1469     if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1470       // Be conservative.
1471       return false;
1472     SrcReg = MI.getOperand(1).getReg();
1473     DstReg = MI.getOperand(0).getReg();
1474     switch (MI.getOpcode()) {
1475     default: llvm_unreachable("Unreachable!");
1476     case X86::MOVSX16rr8:
1477     case X86::MOVZX16rr8:
1478     case X86::MOVSX32rr8:
1479     case X86::MOVZX32rr8:
1480     case X86::MOVSX64rr8:
1481       SubIdx = X86::sub_8bit;
1482       break;
1483     case X86::MOVSX32rr16:
1484     case X86::MOVZX32rr16:
1485     case X86::MOVSX64rr16:
1486       SubIdx = X86::sub_16bit;
1487       break;
1488     case X86::MOVSX64rr32:
1489       SubIdx = X86::sub_32bit;
1490       break;
1491     }
1492     return true;
1493   }
1494   }
1495   return false;
1496 }
1497 
1498 /// isFrameOperand - Return true and the FrameIndex if the specified
1499 /// operand and follow operands form a reference to the stack frame.
1500 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1501                                   int &FrameIndex) const {
1502   if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1503       MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1504       MI->getOperand(Op+1).getImm() == 1 &&
1505       MI->getOperand(Op+2).getReg() == 0 &&
1506       MI->getOperand(Op+3).getImm() == 0) {
1507     FrameIndex = MI->getOperand(Op).getIndex();
1508     return true;
1509   }
1510   return false;
1511 }
1512 
1513 static bool isFrameLoadOpcode(int Opcode) {
1514   switch (Opcode) {
1515   default:
1516     return false;
1517   case X86::MOV8rm:
1518   case X86::MOV16rm:
1519   case X86::MOV32rm:
1520   case X86::MOV64rm:
1521   case X86::LD_Fp64m:
1522   case X86::MOVSSrm:
1523   case X86::MOVSDrm:
1524   case X86::MOVAPSrm:
1525   case X86::MOVAPDrm:
1526   case X86::MOVDQArm:
1527   case X86::VMOVSSrm:
1528   case X86::VMOVSDrm:
1529   case X86::VMOVAPSrm:
1530   case X86::VMOVAPDrm:
1531   case X86::VMOVDQArm:
1532   case X86::VMOVAPSYrm:
1533   case X86::VMOVAPDYrm:
1534   case X86::VMOVDQAYrm:
1535   case X86::MMX_MOVD64rm:
1536   case X86::MMX_MOVQ64rm:
1537   case X86::VMOVDQA32rm:
1538   case X86::VMOVDQA64rm:
1539     return true;
1540   }
1541 }
1542 
1543 static bool isFrameStoreOpcode(int Opcode) {
1544   switch (Opcode) {
1545   default: break;
1546   case X86::MOV8mr:
1547   case X86::MOV16mr:
1548   case X86::MOV32mr:
1549   case X86::MOV64mr:
1550   case X86::ST_FpP64m:
1551   case X86::MOVSSmr:
1552   case X86::MOVSDmr:
1553   case X86::MOVAPSmr:
1554   case X86::MOVAPDmr:
1555   case X86::MOVDQAmr:
1556   case X86::VMOVSSmr:
1557   case X86::VMOVSDmr:
1558   case X86::VMOVAPSmr:
1559   case X86::VMOVAPDmr:
1560   case X86::VMOVDQAmr:
1561   case X86::VMOVAPSYmr:
1562   case X86::VMOVAPDYmr:
1563   case X86::VMOVDQAYmr:
1564   case X86::MMX_MOVD64mr:
1565   case X86::MMX_MOVQ64mr:
1566   case X86::MMX_MOVNTQmr:
1567     return true;
1568   }
1569   return false;
1570 }
1571 
1572 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
1573                                            int &FrameIndex) const {
1574   if (isFrameLoadOpcode(MI->getOpcode()))
1575     if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
1576       return MI->getOperand(0).getReg();
1577   return 0;
1578 }
1579 
1580 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
1581                                                  int &FrameIndex) const {
1582   if (isFrameLoadOpcode(MI->getOpcode())) {
1583     unsigned Reg;
1584     if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1585       return Reg;
1586     // Check for post-frame index elimination operations
1587     const MachineMemOperand *Dummy;
1588     return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
1589   }
1590   return 0;
1591 }
1592 
1593 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
1594                                           int &FrameIndex) const {
1595   if (isFrameStoreOpcode(MI->getOpcode()))
1596     if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1597         isFrameOperand(MI, 0, FrameIndex))
1598       return MI->getOperand(X86::AddrNumOperands).getReg();
1599   return 0;
1600 }
1601 
1602 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1603                                                 int &FrameIndex) const {
1604   if (isFrameStoreOpcode(MI->getOpcode())) {
1605     unsigned Reg;
1606     if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1607       return Reg;
1608     // Check for post-frame index elimination operations
1609     const MachineMemOperand *Dummy;
1610     return hasStoreToStackSlot(MI, Dummy, FrameIndex);
1611   }
1612   return 0;
1613 }
1614 
1615 /// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1616 /// X86::MOVPC32r.
1617 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
1618   // Don't waste compile time scanning use-def chains of physregs.
1619   if (!TargetRegisterInfo::isVirtualRegister(BaseReg))
1620     return false;
1621   bool isPICBase = false;
1622   for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1623          E = MRI.def_end(); I != E; ++I) {
1624     MachineInstr *DefMI = I.getOperand().getParent();
1625     if (DefMI->getOpcode() != X86::MOVPC32r)
1626       return false;
1627     assert(!isPICBase && "More than one PIC base?");
1628     isPICBase = true;
1629   }
1630   return isPICBase;
1631 }
1632 
1633 bool
1634 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1635                                                 AliasAnalysis *AA) const {
1636   switch (MI->getOpcode()) {
1637   default: break;
1638   case X86::MOV8rm:
1639   case X86::MOV16rm:
1640   case X86::MOV32rm:
1641   case X86::MOV64rm:
1642   case X86::LD_Fp64m:
1643   case X86::MOVSSrm:
1644   case X86::MOVSDrm:
1645   case X86::MOVAPSrm:
1646   case X86::MOVUPSrm:
1647   case X86::MOVAPDrm:
1648   case X86::MOVDQArm:
1649   case X86::MOVDQUrm:
1650   case X86::VMOVSSrm:
1651   case X86::VMOVSDrm:
1652   case X86::VMOVAPSrm:
1653   case X86::VMOVUPSrm:
1654   case X86::VMOVAPDrm:
1655   case X86::VMOVDQArm:
1656   case X86::VMOVDQUrm:
1657   case X86::VMOVAPSYrm:
1658   case X86::VMOVUPSYrm:
1659   case X86::VMOVAPDYrm:
1660   case X86::VMOVDQAYrm:
1661   case X86::VMOVDQUYrm:
1662   case X86::MMX_MOVD64rm:
1663   case X86::MMX_MOVQ64rm:
1664   case X86::FsVMOVAPSrm:
1665   case X86::FsVMOVAPDrm:
1666   case X86::FsMOVAPSrm:
1667   case X86::FsMOVAPDrm: {
1668     // Loads from constant pools are trivially rematerializable.
1669     if (MI->getOperand(1).isReg() &&
1670         MI->getOperand(2).isImm() &&
1671         MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1672         MI->isInvariantLoad(AA)) {
1673       unsigned BaseReg = MI->getOperand(1).getReg();
1674       if (BaseReg == 0 || BaseReg == X86::RIP)
1675         return true;
1676       // Allow re-materialization of PIC load.
1677       if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
1678         return false;
1679       const MachineFunction &MF = *MI->getParent()->getParent();
1680       const MachineRegisterInfo &MRI = MF.getRegInfo();
1681       return regIsPICBase(BaseReg, MRI);
1682     }
1683     return false;
1684   }
1685 
1686   case X86::LEA32r:
1687   case X86::LEA64r: {
1688     if (MI->getOperand(2).isImm() &&
1689         MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1690         !MI->getOperand(4).isReg()) {
1691       // lea fi#, lea GV, etc. are all rematerializable.
1692       if (!MI->getOperand(1).isReg())
1693         return true;
1694       unsigned BaseReg = MI->getOperand(1).getReg();
1695       if (BaseReg == 0)
1696         return true;
1697       // Allow re-materialization of lea PICBase + x.
1698       const MachineFunction &MF = *MI->getParent()->getParent();
1699       const MachineRegisterInfo &MRI = MF.getRegInfo();
1700       return regIsPICBase(BaseReg, MRI);
1701     }
1702     return false;
1703   }
1704   }
1705 
1706   // All other instructions marked M_REMATERIALIZABLE are always trivially
1707   // rematerializable.
1708   return true;
1709 }
1710 
1711 /// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1712 /// would clobber the EFLAGS condition register. Note the result may be
1713 /// conservative. If it cannot definitely determine the safety after visiting
1714 /// a few instructions in each direction it assumes it's not safe.
1715 static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1716                                   MachineBasicBlock::iterator I) {
1717   MachineBasicBlock::iterator E = MBB.end();
1718 
1719   // For compile time consideration, if we are not able to determine the
1720   // safety after visiting 4 instructions in each direction, we will assume
1721   // it's not safe.
1722   MachineBasicBlock::iterator Iter = I;
1723   for (unsigned i = 0; Iter != E && i < 4; ++i) {
1724     bool SeenDef = false;
1725     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1726       MachineOperand &MO = Iter->getOperand(j);
1727       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1728         SeenDef = true;
1729       if (!MO.isReg())
1730         continue;
1731       if (MO.getReg() == X86::EFLAGS) {
1732         if (MO.isUse())
1733           return false;
1734         SeenDef = true;
1735       }
1736     }
1737 
1738     if (SeenDef)
1739       // This instruction defines EFLAGS, no need to look any further.
1740       return true;
1741     ++Iter;
1742     // Skip over DBG_VALUE.
1743     while (Iter != E && Iter->isDebugValue())
1744       ++Iter;
1745   }
1746 
1747   // It is safe to clobber EFLAGS at the end of a block of no successor has it
1748   // live in.
1749   if (Iter == E) {
1750     for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1751            SE = MBB.succ_end(); SI != SE; ++SI)
1752       if ((*SI)->isLiveIn(X86::EFLAGS))
1753         return false;
1754     return true;
1755   }
1756 
1757   MachineBasicBlock::iterator B = MBB.begin();
1758   Iter = I;
1759   for (unsigned i = 0; i < 4; ++i) {
1760     // If we make it to the beginning of the block, it's safe to clobber
1761     // EFLAGS iff EFLAGS is not live-in.
1762     if (Iter == B)
1763       return !MBB.isLiveIn(X86::EFLAGS);
1764 
1765     --Iter;
1766     // Skip over DBG_VALUE.
1767     while (Iter != B && Iter->isDebugValue())
1768       --Iter;
1769 
1770     bool SawKill = false;
1771     for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1772       MachineOperand &MO = Iter->getOperand(j);
1773       // A register mask may clobber EFLAGS, but we should still look for a
1774       // live EFLAGS def.
1775       if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1776         SawKill = true;
1777       if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1778         if (MO.isDef()) return MO.isDead();
1779         if (MO.isKill()) SawKill = true;
1780       }
1781     }
1782 
1783     if (SawKill)
1784       // This instruction kills EFLAGS and doesn't redefine it, so
1785       // there's no need to look further.
1786       return true;
1787   }
1788 
1789   // Conservative answer.
1790   return false;
1791 }
1792 
1793 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1794                                  MachineBasicBlock::iterator I,
1795                                  unsigned DestReg, unsigned SubIdx,
1796                                  const MachineInstr *Orig,
1797                                  const TargetRegisterInfo &TRI) const {
1798   // MOV32r0 is implemented with a xor which clobbers condition code.
1799   // Re-materialize it as movri instructions to avoid side effects.
1800   unsigned Opc = Orig->getOpcode();
1801   if (Opc == X86::MOV32r0 && !isSafeToClobberEFLAGS(MBB, I)) {
1802     DebugLoc DL = Orig->getDebugLoc();
1803     BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0))
1804       .addImm(0);
1805   } else {
1806     MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
1807     MBB.insert(I, MI);
1808   }
1809 
1810   MachineInstr *NewMI = prior(I);
1811   NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
1812 }
1813 
1814 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1815 /// is not marked dead.
1816 static bool hasLiveCondCodeDef(MachineInstr *MI) {
1817   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1818     MachineOperand &MO = MI->getOperand(i);
1819     if (MO.isReg() && MO.isDef() &&
1820         MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1821       return true;
1822     }
1823   }
1824   return false;
1825 }
1826 
1827 /// getTruncatedShiftCount - check whether the shift count for a machine operand
1828 /// is non-zero.
1829 inline static unsigned getTruncatedShiftCount(MachineInstr *MI,
1830                                               unsigned ShiftAmtOperandIdx) {
1831   // The shift count is six bits with the REX.W prefix and five bits without.
1832   unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31;
1833   unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm();
1834   return Imm & ShiftCountMask;
1835 }
1836 
1837 /// isTruncatedShiftCountForLEA - check whether the given shift count is appropriate
1838 /// can be represented by a LEA instruction.
1839 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) {
1840   // Left shift instructions can be transformed into load-effective-address
1841   // instructions if we can encode them appropriately.
1842   // A LEA instruction utilizes a SIB byte to encode it's scale factor.
1843   // The SIB.scale field is two bits wide which means that we can encode any
1844   // shift amount less than 4.
1845   return ShAmt < 4 && ShAmt > 0;
1846 }
1847 
1848 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
1849                                   unsigned Opc, bool AllowSP,
1850                                   unsigned &NewSrc, bool &isKill, bool &isUndef,
1851                                   MachineOperand &ImplicitOp) const {
1852   MachineFunction &MF = *MI->getParent()->getParent();
1853   const TargetRegisterClass *RC;
1854   if (AllowSP) {
1855     RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass;
1856   } else {
1857     RC = Opc != X86::LEA32r ?
1858       &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass;
1859   }
1860   unsigned SrcReg = Src.getReg();
1861 
1862   // For both LEA64 and LEA32 the register already has essentially the right
1863   // type (32-bit or 64-bit) we may just need to forbid SP.
1864   if (Opc != X86::LEA64_32r) {
1865     NewSrc = SrcReg;
1866     isKill = Src.isKill();
1867     isUndef = Src.isUndef();
1868 
1869     if (TargetRegisterInfo::isVirtualRegister(NewSrc) &&
1870         !MF.getRegInfo().constrainRegClass(NewSrc, RC))
1871       return false;
1872 
1873     return true;
1874   }
1875 
1876   // This is for an LEA64_32r and incoming registers are 32-bit. One way or
1877   // another we need to add 64-bit registers to the final MI.
1878   if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
1879     ImplicitOp = Src;
1880     ImplicitOp.setImplicit();
1881 
1882     NewSrc = getX86SubSuperRegister(Src.getReg(), MVT::i64);
1883     MachineBasicBlock::LivenessQueryResult LQR =
1884       MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI);
1885 
1886     switch (LQR) {
1887     case MachineBasicBlock::LQR_Unknown:
1888       // We can't give sane liveness flags to the instruction, abandon LEA
1889       // formation.
1890       return false;
1891     case MachineBasicBlock::LQR_Live:
1892       isKill = MI->killsRegister(SrcReg);
1893       isUndef = false;
1894       break;
1895     default:
1896       // The physreg itself is dead, so we have to use it as an <undef>.
1897       isKill = false;
1898       isUndef = true;
1899       break;
1900     }
1901   } else {
1902     // Virtual register of the wrong class, we have to create a temporary 64-bit
1903     // vreg to feed into the LEA.
1904     NewSrc = MF.getRegInfo().createVirtualRegister(RC);
1905     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1906             get(TargetOpcode::COPY))
1907       .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit)
1908         .addOperand(Src);
1909 
1910     // Which is obviously going to be dead after we're done with it.
1911     isKill = true;
1912     isUndef = false;
1913   }
1914 
1915   // We've set all the parameters without issue.
1916   return true;
1917 }
1918 
1919 /// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
1920 /// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1921 /// to a 32-bit superregister and then truncating back down to a 16-bit
1922 /// subregister.
1923 MachineInstr *
1924 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1925                                            MachineFunction::iterator &MFI,
1926                                            MachineBasicBlock::iterator &MBBI,
1927                                            LiveVariables *LV) const {
1928   MachineInstr *MI = MBBI;
1929   unsigned Dest = MI->getOperand(0).getReg();
1930   unsigned Src = MI->getOperand(1).getReg();
1931   bool isDead = MI->getOperand(0).isDead();
1932   bool isKill = MI->getOperand(1).isKill();
1933 
1934   MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
1935   unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
1936   unsigned Opc, leaInReg;
1937   if (TM.getSubtarget<X86Subtarget>().is64Bit()) {
1938     Opc = X86::LEA64_32r;
1939     leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1940   } else {
1941     Opc = X86::LEA32r;
1942     leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1943   }
1944 
1945   // Build and insert into an implicit UNDEF value. This is OK because
1946   // well be shifting and then extracting the lower 16-bits.
1947   // This has the potential to cause partial register stall. e.g.
1948   //   movw    (%rbp,%rcx,2), %dx
1949   //   leal    -65(%rdx), %esi
1950   // But testing has shown this *does* help performance in 64-bit mode (at
1951   // least on modern x86 machines).
1952   BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1953   MachineInstr *InsMI =
1954     BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1955     .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1956     .addReg(Src, getKillRegState(isKill));
1957 
1958   MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1959                                     get(Opc), leaOutReg);
1960   switch (MIOpc) {
1961   default: llvm_unreachable("Unreachable!");
1962   case X86::SHL16ri: {
1963     unsigned ShAmt = MI->getOperand(2).getImm();
1964     MIB.addReg(0).addImm(1 << ShAmt)
1965        .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
1966     break;
1967   }
1968   case X86::INC16r:
1969   case X86::INC64_16r:
1970     addRegOffset(MIB, leaInReg, true, 1);
1971     break;
1972   case X86::DEC16r:
1973   case X86::DEC64_16r:
1974     addRegOffset(MIB, leaInReg, true, -1);
1975     break;
1976   case X86::ADD16ri:
1977   case X86::ADD16ri8:
1978   case X86::ADD16ri_DB:
1979   case X86::ADD16ri8_DB:
1980     addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
1981     break;
1982   case X86::ADD16rr:
1983   case X86::ADD16rr_DB: {
1984     unsigned Src2 = MI->getOperand(2).getReg();
1985     bool isKill2 = MI->getOperand(2).isKill();
1986     unsigned leaInReg2 = 0;
1987     MachineInstr *InsMI2 = 0;
1988     if (Src == Src2) {
1989       // ADD16rr %reg1028<kill>, %reg1028
1990       // just a single insert_subreg.
1991       addRegReg(MIB, leaInReg, true, leaInReg, false);
1992     } else {
1993       if (TM.getSubtarget<X86Subtarget>().is64Bit())
1994         leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass);
1995       else
1996         leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
1997       // Build and insert into an implicit UNDEF value. This is OK because
1998       // well be shifting and then extracting the lower 16-bits.
1999       BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
2000       InsMI2 =
2001         BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
2002         .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
2003         .addReg(Src2, getKillRegState(isKill2));
2004       addRegReg(MIB, leaInReg, true, leaInReg2, true);
2005     }
2006     if (LV && isKill2 && InsMI2)
2007       LV->replaceKillInstruction(Src2, MI, InsMI2);
2008     break;
2009   }
2010   }
2011 
2012   MachineInstr *NewMI = MIB;
2013   MachineInstr *ExtMI =
2014     BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
2015     .addReg(Dest, RegState::Define | getDeadRegState(isDead))
2016     .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
2017 
2018   if (LV) {
2019     // Update live variables
2020     LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
2021     LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
2022     if (isKill)
2023       LV->replaceKillInstruction(Src, MI, InsMI);
2024     if (isDead)
2025       LV->replaceKillInstruction(Dest, MI, ExtMI);
2026   }
2027 
2028   return ExtMI;
2029 }
2030 
2031 /// convertToThreeAddress - This method must be implemented by targets that
2032 /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
2033 /// may be able to convert a two-address instruction into a true
2034 /// three-address instruction on demand.  This allows the X86 target (for
2035 /// example) to convert ADD and SHL instructions into LEA instructions if they
2036 /// would require register copies due to two-addressness.
2037 ///
2038 /// This method returns a null pointer if the transformation cannot be
2039 /// performed, otherwise it returns the new instruction.
2040 ///
2041 MachineInstr *
2042 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
2043                                     MachineBasicBlock::iterator &MBBI,
2044                                     LiveVariables *LV) const {
2045   MachineInstr *MI = MBBI;
2046 
2047   // The following opcodes also sets the condition code register(s). Only
2048   // convert them to equivalent lea if the condition code register def's
2049   // are dead!
2050   if (hasLiveCondCodeDef(MI))
2051     return 0;
2052 
2053   MachineFunction &MF = *MI->getParent()->getParent();
2054   // All instructions input are two-addr instructions.  Get the known operands.
2055   const MachineOperand &Dest = MI->getOperand(0);
2056   const MachineOperand &Src = MI->getOperand(1);
2057 
2058   MachineInstr *NewMI = NULL;
2059   // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's.  When
2060   // we have better subtarget support, enable the 16-bit LEA generation here.
2061   // 16-bit LEA is also slow on Core2.
2062   bool DisableLEA16 = true;
2063   bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
2064 
2065   unsigned MIOpc = MI->getOpcode();
2066   switch (MIOpc) {
2067   case X86::SHUFPSrri: {
2068     assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
2069     if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
2070 
2071     unsigned B = MI->getOperand(1).getReg();
2072     unsigned C = MI->getOperand(2).getReg();
2073     if (B != C) return 0;
2074     unsigned M = MI->getOperand(3).getImm();
2075     NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
2076       .addOperand(Dest).addOperand(Src).addImm(M);
2077     break;
2078   }
2079   case X86::SHUFPDrri: {
2080     assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
2081     if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
2082 
2083     unsigned B = MI->getOperand(1).getReg();
2084     unsigned C = MI->getOperand(2).getReg();
2085     if (B != C) return 0;
2086     unsigned M = MI->getOperand(3).getImm();
2087 
2088     // Convert to PSHUFD mask.
2089     M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
2090 
2091     NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
2092       .addOperand(Dest).addOperand(Src).addImm(M);
2093     break;
2094   }
2095   case X86::SHL64ri: {
2096     assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2097     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2098     if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
2099 
2100     // LEA can't handle RSP.
2101     if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) &&
2102         !MF.getRegInfo().constrainRegClass(Src.getReg(),
2103                                            &X86::GR64_NOSPRegClass))
2104       return 0;
2105 
2106     NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2107       .addOperand(Dest)
2108       .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2109     break;
2110   }
2111   case X86::SHL32ri: {
2112     assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2113     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2114     if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
2115 
2116     unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2117 
2118     // LEA can't handle ESP.
2119     bool isKill, isUndef;
2120     unsigned SrcReg;
2121     MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2122     if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2123                         SrcReg, isKill, isUndef, ImplicitOp))
2124       return 0;
2125 
2126     MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2127       .addOperand(Dest)
2128       .addReg(0).addImm(1 << ShAmt)
2129       .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef))
2130       .addImm(0).addReg(0);
2131     if (ImplicitOp.getReg() != 0)
2132       MIB.addOperand(ImplicitOp);
2133     NewMI = MIB;
2134 
2135     break;
2136   }
2137   case X86::SHL16ri: {
2138     assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
2139     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
2140     if (!isTruncatedShiftCountForLEA(ShAmt)) return 0;
2141 
2142     if (DisableLEA16)
2143       return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2144     NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2145       .addOperand(Dest)
2146       .addReg(0).addImm(1 << ShAmt).addOperand(Src).addImm(0).addReg(0);
2147     break;
2148   }
2149   default: {
2150 
2151     switch (MIOpc) {
2152     default: return 0;
2153     case X86::INC64r:
2154     case X86::INC32r:
2155     case X86::INC64_32r: {
2156       assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2157       unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
2158         : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2159       bool isKill, isUndef;
2160       unsigned SrcReg;
2161       MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2162       if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2163                           SrcReg, isKill, isUndef, ImplicitOp))
2164         return 0;
2165 
2166       MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2167           .addOperand(Dest)
2168           .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef));
2169       if (ImplicitOp.getReg() != 0)
2170         MIB.addOperand(ImplicitOp);
2171 
2172       NewMI = addOffset(MIB, 1);
2173       break;
2174     }
2175     case X86::INC16r:
2176     case X86::INC64_16r:
2177       if (DisableLEA16)
2178         return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2179       assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
2180       NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2181                         .addOperand(Dest).addOperand(Src), 1);
2182       break;
2183     case X86::DEC64r:
2184     case X86::DEC32r:
2185     case X86::DEC64_32r: {
2186       assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2187       unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
2188         : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
2189 
2190       bool isKill, isUndef;
2191       unsigned SrcReg;
2192       MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2193       if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false,
2194                           SrcReg, isKill, isUndef, ImplicitOp))
2195         return 0;
2196 
2197       MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2198           .addOperand(Dest)
2199           .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2200       if (ImplicitOp.getReg() != 0)
2201         MIB.addOperand(ImplicitOp);
2202 
2203       NewMI = addOffset(MIB, -1);
2204 
2205       break;
2206     }
2207     case X86::DEC16r:
2208     case X86::DEC64_16r:
2209       if (DisableLEA16)
2210         return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2211       assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
2212       NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2213                         .addOperand(Dest).addOperand(Src), -1);
2214       break;
2215     case X86::ADD64rr:
2216     case X86::ADD64rr_DB:
2217     case X86::ADD32rr:
2218     case X86::ADD32rr_DB: {
2219       assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2220       unsigned Opc;
2221       if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB)
2222         Opc = X86::LEA64r;
2223       else
2224         Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2225 
2226       bool isKill, isUndef;
2227       unsigned SrcReg;
2228       MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2229       if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2230                           SrcReg, isKill, isUndef, ImplicitOp))
2231         return 0;
2232 
2233       const MachineOperand &Src2 = MI->getOperand(2);
2234       bool isKill2, isUndef2;
2235       unsigned SrcReg2;
2236       MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false);
2237       if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false,
2238                           SrcReg2, isKill2, isUndef2, ImplicitOp2))
2239         return 0;
2240 
2241       MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2242         .addOperand(Dest);
2243       if (ImplicitOp.getReg() != 0)
2244         MIB.addOperand(ImplicitOp);
2245       if (ImplicitOp2.getReg() != 0)
2246         MIB.addOperand(ImplicitOp2);
2247 
2248       NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2);
2249 
2250       // Preserve undefness of the operands.
2251       NewMI->getOperand(1).setIsUndef(isUndef);
2252       NewMI->getOperand(3).setIsUndef(isUndef2);
2253 
2254       if (LV && Src2.isKill())
2255         LV->replaceKillInstruction(SrcReg2, MI, NewMI);
2256       break;
2257     }
2258     case X86::ADD16rr:
2259     case X86::ADD16rr_DB: {
2260       if (DisableLEA16)
2261         return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2262       assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2263       unsigned Src2 = MI->getOperand(2).getReg();
2264       bool isKill2 = MI->getOperand(2).isKill();
2265       NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2266                         .addOperand(Dest),
2267                         Src.getReg(), Src.isKill(), Src2, isKill2);
2268 
2269       // Preserve undefness of the operands.
2270       bool isUndef = MI->getOperand(1).isUndef();
2271       bool isUndef2 = MI->getOperand(2).isUndef();
2272       NewMI->getOperand(1).setIsUndef(isUndef);
2273       NewMI->getOperand(3).setIsUndef(isUndef2);
2274 
2275       if (LV && isKill2)
2276         LV->replaceKillInstruction(Src2, MI, NewMI);
2277       break;
2278     }
2279     case X86::ADD64ri32:
2280     case X86::ADD64ri8:
2281     case X86::ADD64ri32_DB:
2282     case X86::ADD64ri8_DB:
2283       assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2284       NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
2285                         .addOperand(Dest).addOperand(Src),
2286                         MI->getOperand(2).getImm());
2287       break;
2288     case X86::ADD32ri:
2289     case X86::ADD32ri8:
2290     case X86::ADD32ri_DB:
2291     case X86::ADD32ri8_DB: {
2292       assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2293       unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
2294 
2295       bool isKill, isUndef;
2296       unsigned SrcReg;
2297       MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false);
2298       if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true,
2299                           SrcReg, isKill, isUndef, ImplicitOp))
2300         return 0;
2301 
2302       MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc))
2303           .addOperand(Dest)
2304           .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill));
2305       if (ImplicitOp.getReg() != 0)
2306         MIB.addOperand(ImplicitOp);
2307 
2308       NewMI = addOffset(MIB, MI->getOperand(2).getImm());
2309       break;
2310     }
2311     case X86::ADD16ri:
2312     case X86::ADD16ri8:
2313     case X86::ADD16ri_DB:
2314     case X86::ADD16ri8_DB:
2315       if (DisableLEA16)
2316         return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
2317       assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
2318       NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
2319                         .addOperand(Dest).addOperand(Src),
2320                         MI->getOperand(2).getImm());
2321       break;
2322     }
2323   }
2324   }
2325 
2326   if (!NewMI) return 0;
2327 
2328   if (LV) {  // Update live variables
2329     if (Src.isKill())
2330       LV->replaceKillInstruction(Src.getReg(), MI, NewMI);
2331     if (Dest.isDead())
2332       LV->replaceKillInstruction(Dest.getReg(), MI, NewMI);
2333   }
2334 
2335   MFI->insert(MBBI, NewMI);          // Insert the new inst
2336   return NewMI;
2337 }
2338 
2339 /// commuteInstruction - We have a few instructions that must be hacked on to
2340 /// commute them.
2341 ///
2342 MachineInstr *
2343 X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
2344   switch (MI->getOpcode()) {
2345   case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2346   case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
2347   case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
2348   case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2349   case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2350   case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
2351     unsigned Opc;
2352     unsigned Size;
2353     switch (MI->getOpcode()) {
2354     default: llvm_unreachable("Unreachable!");
2355     case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2356     case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2357     case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2358     case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
2359     case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2360     case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
2361     }
2362     unsigned Amt = MI->getOperand(3).getImm();
2363     if (NewMI) {
2364       MachineFunction &MF = *MI->getParent()->getParent();
2365       MI = MF.CloneMachineInstr(MI);
2366       NewMI = false;
2367     }
2368     MI->setDesc(get(Opc));
2369     MI->getOperand(3).setImm(Size-Amt);
2370     return TargetInstrInfo::commuteInstruction(MI, NewMI);
2371   }
2372   case X86::CMOVB16rr:  case X86::CMOVB32rr:  case X86::CMOVB64rr:
2373   case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
2374   case X86::CMOVE16rr:  case X86::CMOVE32rr:  case X86::CMOVE64rr:
2375   case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
2376   case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
2377   case X86::CMOVA16rr:  case X86::CMOVA32rr:  case X86::CMOVA64rr:
2378   case X86::CMOVL16rr:  case X86::CMOVL32rr:  case X86::CMOVL64rr:
2379   case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
2380   case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
2381   case X86::CMOVG16rr:  case X86::CMOVG32rr:  case X86::CMOVG64rr:
2382   case X86::CMOVS16rr:  case X86::CMOVS32rr:  case X86::CMOVS64rr:
2383   case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
2384   case X86::CMOVP16rr:  case X86::CMOVP32rr:  case X86::CMOVP64rr:
2385   case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
2386   case X86::CMOVO16rr:  case X86::CMOVO32rr:  case X86::CMOVO64rr:
2387   case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: {
2388     unsigned Opc;
2389     switch (MI->getOpcode()) {
2390     default: llvm_unreachable("Unreachable!");
2391     case X86::CMOVB16rr:  Opc = X86::CMOVAE16rr; break;
2392     case X86::CMOVB32rr:  Opc = X86::CMOVAE32rr; break;
2393     case X86::CMOVB64rr:  Opc = X86::CMOVAE64rr; break;
2394     case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2395     case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2396     case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2397     case X86::CMOVE16rr:  Opc = X86::CMOVNE16rr; break;
2398     case X86::CMOVE32rr:  Opc = X86::CMOVNE32rr; break;
2399     case X86::CMOVE64rr:  Opc = X86::CMOVNE64rr; break;
2400     case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2401     case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2402     case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
2403     case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2404     case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2405     case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2406     case X86::CMOVA16rr:  Opc = X86::CMOVBE16rr; break;
2407     case X86::CMOVA32rr:  Opc = X86::CMOVBE32rr; break;
2408     case X86::CMOVA64rr:  Opc = X86::CMOVBE64rr; break;
2409     case X86::CMOVL16rr:  Opc = X86::CMOVGE16rr; break;
2410     case X86::CMOVL32rr:  Opc = X86::CMOVGE32rr; break;
2411     case X86::CMOVL64rr:  Opc = X86::CMOVGE64rr; break;
2412     case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2413     case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2414     case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2415     case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2416     case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2417     case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2418     case X86::CMOVG16rr:  Opc = X86::CMOVLE16rr; break;
2419     case X86::CMOVG32rr:  Opc = X86::CMOVLE32rr; break;
2420     case X86::CMOVG64rr:  Opc = X86::CMOVLE64rr; break;
2421     case X86::CMOVS16rr:  Opc = X86::CMOVNS16rr; break;
2422     case X86::CMOVS32rr:  Opc = X86::CMOVNS32rr; break;
2423     case X86::CMOVS64rr:  Opc = X86::CMOVNS64rr; break;
2424     case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2425     case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2426     case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2427     case X86::CMOVP16rr:  Opc = X86::CMOVNP16rr; break;
2428     case X86::CMOVP32rr:  Opc = X86::CMOVNP32rr; break;
2429     case X86::CMOVP64rr:  Opc = X86::CMOVNP64rr; break;
2430     case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2431     case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2432     case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
2433     case X86::CMOVO16rr:  Opc = X86::CMOVNO16rr; break;
2434     case X86::CMOVO32rr:  Opc = X86::CMOVNO32rr; break;
2435     case X86::CMOVO64rr:  Opc = X86::CMOVNO64rr; break;
2436     case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2437     case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2438     case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
2439     }
2440     if (NewMI) {
2441       MachineFunction &MF = *MI->getParent()->getParent();
2442       MI = MF.CloneMachineInstr(MI);
2443       NewMI = false;
2444     }
2445     MI->setDesc(get(Opc));
2446     // Fallthrough intended.
2447   }
2448   default:
2449     return TargetInstrInfo::commuteInstruction(MI, NewMI);
2450   }
2451 }
2452 
2453 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) {
2454   switch (BrOpc) {
2455   default: return X86::COND_INVALID;
2456   case X86::JE_4:  return X86::COND_E;
2457   case X86::JNE_4: return X86::COND_NE;
2458   case X86::JL_4:  return X86::COND_L;
2459   case X86::JLE_4: return X86::COND_LE;
2460   case X86::JG_4:  return X86::COND_G;
2461   case X86::JGE_4: return X86::COND_GE;
2462   case X86::JB_4:  return X86::COND_B;
2463   case X86::JBE_4: return X86::COND_BE;
2464   case X86::JA_4:  return X86::COND_A;
2465   case X86::JAE_4: return X86::COND_AE;
2466   case X86::JS_4:  return X86::COND_S;
2467   case X86::JNS_4: return X86::COND_NS;
2468   case X86::JP_4:  return X86::COND_P;
2469   case X86::JNP_4: return X86::COND_NP;
2470   case X86::JO_4:  return X86::COND_O;
2471   case X86::JNO_4: return X86::COND_NO;
2472   }
2473 }
2474 
2475 /// getCondFromSETOpc - return condition code of a SET opcode.
2476 static X86::CondCode getCondFromSETOpc(unsigned Opc) {
2477   switch (Opc) {
2478   default: return X86::COND_INVALID;
2479   case X86::SETAr:  case X86::SETAm:  return X86::COND_A;
2480   case X86::SETAEr: case X86::SETAEm: return X86::COND_AE;
2481   case X86::SETBr:  case X86::SETBm:  return X86::COND_B;
2482   case X86::SETBEr: case X86::SETBEm: return X86::COND_BE;
2483   case X86::SETEr:  case X86::SETEm:  return X86::COND_E;
2484   case X86::SETGr:  case X86::SETGm:  return X86::COND_G;
2485   case X86::SETGEr: case X86::SETGEm: return X86::COND_GE;
2486   case X86::SETLr:  case X86::SETLm:  return X86::COND_L;
2487   case X86::SETLEr: case X86::SETLEm: return X86::COND_LE;
2488   case X86::SETNEr: case X86::SETNEm: return X86::COND_NE;
2489   case X86::SETNOr: case X86::SETNOm: return X86::COND_NO;
2490   case X86::SETNPr: case X86::SETNPm: return X86::COND_NP;
2491   case X86::SETNSr: case X86::SETNSm: return X86::COND_NS;
2492   case X86::SETOr:  case X86::SETOm:  return X86::COND_O;
2493   case X86::SETPr:  case X86::SETPm:  return X86::COND_P;
2494   case X86::SETSr:  case X86::SETSm:  return X86::COND_S;
2495   }
2496 }
2497 
2498 /// getCondFromCmovOpc - return condition code of a CMov opcode.
2499 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) {
2500   switch (Opc) {
2501   default: return X86::COND_INVALID;
2502   case X86::CMOVA16rm:  case X86::CMOVA16rr:  case X86::CMOVA32rm:
2503   case X86::CMOVA32rr:  case X86::CMOVA64rm:  case X86::CMOVA64rr:
2504     return X86::COND_A;
2505   case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm:
2506   case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr:
2507     return X86::COND_AE;
2508   case X86::CMOVB16rm:  case X86::CMOVB16rr:  case X86::CMOVB32rm:
2509   case X86::CMOVB32rr:  case X86::CMOVB64rm:  case X86::CMOVB64rr:
2510     return X86::COND_B;
2511   case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm:
2512   case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr:
2513     return X86::COND_BE;
2514   case X86::CMOVE16rm:  case X86::CMOVE16rr:  case X86::CMOVE32rm:
2515   case X86::CMOVE32rr:  case X86::CMOVE64rm:  case X86::CMOVE64rr:
2516     return X86::COND_E;
2517   case X86::CMOVG16rm:  case X86::CMOVG16rr:  case X86::CMOVG32rm:
2518   case X86::CMOVG32rr:  case X86::CMOVG64rm:  case X86::CMOVG64rr:
2519     return X86::COND_G;
2520   case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm:
2521   case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr:
2522     return X86::COND_GE;
2523   case X86::CMOVL16rm:  case X86::CMOVL16rr:  case X86::CMOVL32rm:
2524   case X86::CMOVL32rr:  case X86::CMOVL64rm:  case X86::CMOVL64rr:
2525     return X86::COND_L;
2526   case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm:
2527   case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr:
2528     return X86::COND_LE;
2529   case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm:
2530   case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr:
2531     return X86::COND_NE;
2532   case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm:
2533   case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr:
2534     return X86::COND_NO;
2535   case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm:
2536   case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr:
2537     return X86::COND_NP;
2538   case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm:
2539   case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr:
2540     return X86::COND_NS;
2541   case X86::CMOVO16rm:  case X86::CMOVO16rr:  case X86::CMOVO32rm:
2542   case X86::CMOVO32rr:  case X86::CMOVO64rm:  case X86::CMOVO64rr:
2543     return X86::COND_O;
2544   case X86::CMOVP16rm:  case X86::CMOVP16rr:  case X86::CMOVP32rm:
2545   case X86::CMOVP32rr:  case X86::CMOVP64rm:  case X86::CMOVP64rr:
2546     return X86::COND_P;
2547   case X86::CMOVS16rm:  case X86::CMOVS16rr:  case X86::CMOVS32rm:
2548   case X86::CMOVS32rr:  case X86::CMOVS64rm:  case X86::CMOVS64rr:
2549     return X86::COND_S;
2550   }
2551 }
2552 
2553 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2554   switch (CC) {
2555   default: llvm_unreachable("Illegal condition code!");
2556   case X86::COND_E:  return X86::JE_4;
2557   case X86::COND_NE: return X86::JNE_4;
2558   case X86::COND_L:  return X86::JL_4;
2559   case X86::COND_LE: return X86::JLE_4;
2560   case X86::COND_G:  return X86::JG_4;
2561   case X86::COND_GE: return X86::JGE_4;
2562   case X86::COND_B:  return X86::JB_4;
2563   case X86::COND_BE: return X86::JBE_4;
2564   case X86::COND_A:  return X86::JA_4;
2565   case X86::COND_AE: return X86::JAE_4;
2566   case X86::COND_S:  return X86::JS_4;
2567   case X86::COND_NS: return X86::JNS_4;
2568   case X86::COND_P:  return X86::JP_4;
2569   case X86::COND_NP: return X86::JNP_4;
2570   case X86::COND_O:  return X86::JO_4;
2571   case X86::COND_NO: return X86::JNO_4;
2572   }
2573 }
2574 
2575 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
2576 /// e.g. turning COND_E to COND_NE.
2577 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2578   switch (CC) {
2579   default: llvm_unreachable("Illegal condition code!");
2580   case X86::COND_E:  return X86::COND_NE;
2581   case X86::COND_NE: return X86::COND_E;
2582   case X86::COND_L:  return X86::COND_GE;
2583   case X86::COND_LE: return X86::COND_G;
2584   case X86::COND_G:  return X86::COND_LE;
2585   case X86::COND_GE: return X86::COND_L;
2586   case X86::COND_B:  return X86::COND_AE;
2587   case X86::COND_BE: return X86::COND_A;
2588   case X86::COND_A:  return X86::COND_BE;
2589   case X86::COND_AE: return X86::COND_B;
2590   case X86::COND_S:  return X86::COND_NS;
2591   case X86::COND_NS: return X86::COND_S;
2592   case X86::COND_P:  return X86::COND_NP;
2593   case X86::COND_NP: return X86::COND_P;
2594   case X86::COND_O:  return X86::COND_NO;
2595   case X86::COND_NO: return X86::COND_O;
2596   }
2597 }
2598 
2599 /// getSwappedCondition - assume the flags are set by MI(a,b), return
2600 /// the condition code if we modify the instructions such that flags are
2601 /// set by MI(b,a).
2602 static X86::CondCode getSwappedCondition(X86::CondCode CC) {
2603   switch (CC) {
2604   default: return X86::COND_INVALID;
2605   case X86::COND_E:  return X86::COND_E;
2606   case X86::COND_NE: return X86::COND_NE;
2607   case X86::COND_L:  return X86::COND_G;
2608   case X86::COND_LE: return X86::COND_GE;
2609   case X86::COND_G:  return X86::COND_L;
2610   case X86::COND_GE: return X86::COND_LE;
2611   case X86::COND_B:  return X86::COND_A;
2612   case X86::COND_BE: return X86::COND_AE;
2613   case X86::COND_A:  return X86::COND_B;
2614   case X86::COND_AE: return X86::COND_BE;
2615   }
2616 }
2617 
2618 /// getSETFromCond - Return a set opcode for the given condition and
2619 /// whether it has memory operand.
2620 static unsigned getSETFromCond(X86::CondCode CC,
2621                                bool HasMemoryOperand) {
2622   static const uint16_t Opc[16][2] = {
2623     { X86::SETAr,  X86::SETAm  },
2624     { X86::SETAEr, X86::SETAEm },
2625     { X86::SETBr,  X86::SETBm  },
2626     { X86::SETBEr, X86::SETBEm },
2627     { X86::SETEr,  X86::SETEm  },
2628     { X86::SETGr,  X86::SETGm  },
2629     { X86::SETGEr, X86::SETGEm },
2630     { X86::SETLr,  X86::SETLm  },
2631     { X86::SETLEr, X86::SETLEm },
2632     { X86::SETNEr, X86::SETNEm },
2633     { X86::SETNOr, X86::SETNOm },
2634     { X86::SETNPr, X86::SETNPm },
2635     { X86::SETNSr, X86::SETNSm },
2636     { X86::SETOr,  X86::SETOm  },
2637     { X86::SETPr,  X86::SETPm  },
2638     { X86::SETSr,  X86::SETSm  }
2639   };
2640 
2641   assert(CC < 16 && "Can only handle standard cond codes");
2642   return Opc[CC][HasMemoryOperand ? 1 : 0];
2643 }
2644 
2645 /// getCMovFromCond - Return a cmov opcode for the given condition,
2646 /// register size in bytes, and operand type.
2647 static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes,
2648                                 bool HasMemoryOperand) {
2649   static const uint16_t Opc[32][3] = {
2650     { X86::CMOVA16rr,  X86::CMOVA32rr,  X86::CMOVA64rr  },
2651     { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr },
2652     { X86::CMOVB16rr,  X86::CMOVB32rr,  X86::CMOVB64rr  },
2653     { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr },
2654     { X86::CMOVE16rr,  X86::CMOVE32rr,  X86::CMOVE64rr  },
2655     { X86::CMOVG16rr,  X86::CMOVG32rr,  X86::CMOVG64rr  },
2656     { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr },
2657     { X86::CMOVL16rr,  X86::CMOVL32rr,  X86::CMOVL64rr  },
2658     { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr },
2659     { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr },
2660     { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr },
2661     { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr },
2662     { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr },
2663     { X86::CMOVO16rr,  X86::CMOVO32rr,  X86::CMOVO64rr  },
2664     { X86::CMOVP16rr,  X86::CMOVP32rr,  X86::CMOVP64rr  },
2665     { X86::CMOVS16rr,  X86::CMOVS32rr,  X86::CMOVS64rr  },
2666     { X86::CMOVA16rm,  X86::CMOVA32rm,  X86::CMOVA64rm  },
2667     { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm },
2668     { X86::CMOVB16rm,  X86::CMOVB32rm,  X86::CMOVB64rm  },
2669     { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm },
2670     { X86::CMOVE16rm,  X86::CMOVE32rm,  X86::CMOVE64rm  },
2671     { X86::CMOVG16rm,  X86::CMOVG32rm,  X86::CMOVG64rm  },
2672     { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm },
2673     { X86::CMOVL16rm,  X86::CMOVL32rm,  X86::CMOVL64rm  },
2674     { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm },
2675     { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm },
2676     { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm },
2677     { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm },
2678     { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm },
2679     { X86::CMOVO16rm,  X86::CMOVO32rm,  X86::CMOVO64rm  },
2680     { X86::CMOVP16rm,  X86::CMOVP32rm,  X86::CMOVP64rm  },
2681     { X86::CMOVS16rm,  X86::CMOVS32rm,  X86::CMOVS64rm  }
2682   };
2683 
2684   assert(CC < 16 && "Can only handle standard cond codes");
2685   unsigned Idx = HasMemoryOperand ? 16+CC : CC;
2686   switch(RegBytes) {
2687   default: llvm_unreachable("Illegal register size!");
2688   case 2: return Opc[Idx][0];
2689   case 4: return Opc[Idx][1];
2690   case 8: return Opc[Idx][2];
2691   }
2692 }
2693 
2694 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
2695   if (!MI->isTerminator()) return false;
2696 
2697   // Conditional branch is a special case.
2698   if (MI->isBranch() && !MI->isBarrier())
2699     return true;
2700   if (!MI->isPredicable())
2701     return true;
2702   return !isPredicated(MI);
2703 }
2704 
2705 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
2706                                  MachineBasicBlock *&TBB,
2707                                  MachineBasicBlock *&FBB,
2708                                  SmallVectorImpl<MachineOperand> &Cond,
2709                                  bool AllowModify) const {
2710   // Start from the bottom of the block and work up, examining the
2711   // terminator instructions.
2712   MachineBasicBlock::iterator I = MBB.end();
2713   MachineBasicBlock::iterator UnCondBrIter = MBB.end();
2714   while (I != MBB.begin()) {
2715     --I;
2716     if (I->isDebugValue())
2717       continue;
2718 
2719     // Working from the bottom, when we see a non-terminator instruction, we're
2720     // done.
2721     if (!isUnpredicatedTerminator(I))
2722       break;
2723 
2724     // A terminator that isn't a branch can't easily be handled by this
2725     // analysis.
2726     if (!I->isBranch())
2727       return true;
2728 
2729     // Handle unconditional branches.
2730     if (I->getOpcode() == X86::JMP_4) {
2731       UnCondBrIter = I;
2732 
2733       if (!AllowModify) {
2734         TBB = I->getOperand(0).getMBB();
2735         continue;
2736       }
2737 
2738       // If the block has any instructions after a JMP, delete them.
2739       while (llvm::next(I) != MBB.end())
2740         llvm::next(I)->eraseFromParent();
2741 
2742       Cond.clear();
2743       FBB = 0;
2744 
2745       // Delete the JMP if it's equivalent to a fall-through.
2746       if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2747         TBB = 0;
2748         I->eraseFromParent();
2749         I = MBB.end();
2750         UnCondBrIter = MBB.end();
2751         continue;
2752       }
2753 
2754       // TBB is used to indicate the unconditional destination.
2755       TBB = I->getOperand(0).getMBB();
2756       continue;
2757     }
2758 
2759     // Handle conditional branches.
2760     X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode());
2761     if (BranchCode == X86::COND_INVALID)
2762       return true;  // Can't handle indirect branch.
2763 
2764     // Working from the bottom, handle the first conditional branch.
2765     if (Cond.empty()) {
2766       MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2767       if (AllowModify && UnCondBrIter != MBB.end() &&
2768           MBB.isLayoutSuccessor(TargetBB)) {
2769         // If we can modify the code and it ends in something like:
2770         //
2771         //     jCC L1
2772         //     jmp L2
2773         //   L1:
2774         //     ...
2775         //   L2:
2776         //
2777         // Then we can change this to:
2778         //
2779         //     jnCC L2
2780         //   L1:
2781         //     ...
2782         //   L2:
2783         //
2784         // Which is a bit more efficient.
2785         // We conditionally jump to the fall-through block.
2786         BranchCode = GetOppositeBranchCondition(BranchCode);
2787         unsigned JNCC = GetCondBranchFromCond(BranchCode);
2788         MachineBasicBlock::iterator OldInst = I;
2789 
2790         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2791           .addMBB(UnCondBrIter->getOperand(0).getMBB());
2792         BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2793           .addMBB(TargetBB);
2794 
2795         OldInst->eraseFromParent();
2796         UnCondBrIter->eraseFromParent();
2797 
2798         // Restart the analysis.
2799         UnCondBrIter = MBB.end();
2800         I = MBB.end();
2801         continue;
2802       }
2803 
2804       FBB = TBB;
2805       TBB = I->getOperand(0).getMBB();
2806       Cond.push_back(MachineOperand::CreateImm(BranchCode));
2807       continue;
2808     }
2809 
2810     // Handle subsequent conditional branches. Only handle the case where all
2811     // conditional branches branch to the same destination and their condition
2812     // opcodes fit one of the special multi-branch idioms.
2813     assert(Cond.size() == 1);
2814     assert(TBB);
2815 
2816     // Only handle the case where all conditional branches branch to the same
2817     // destination.
2818     if (TBB != I->getOperand(0).getMBB())
2819       return true;
2820 
2821     // If the conditions are the same, we can leave them alone.
2822     X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
2823     if (OldBranchCode == BranchCode)
2824       continue;
2825 
2826     // If they differ, see if they fit one of the known patterns. Theoretically,
2827     // we could handle more patterns here, but we shouldn't expect to see them
2828     // if instruction selection has done a reasonable job.
2829     if ((OldBranchCode == X86::COND_NP &&
2830          BranchCode == X86::COND_E) ||
2831         (OldBranchCode == X86::COND_E &&
2832          BranchCode == X86::COND_NP))
2833       BranchCode = X86::COND_NP_OR_E;
2834     else if ((OldBranchCode == X86::COND_P &&
2835               BranchCode == X86::COND_NE) ||
2836              (OldBranchCode == X86::COND_NE &&
2837               BranchCode == X86::COND_P))
2838       BranchCode = X86::COND_NE_OR_P;
2839     else
2840       return true;
2841 
2842     // Update the MachineOperand.
2843     Cond[0].setImm(BranchCode);
2844   }
2845 
2846   return false;
2847 }
2848 
2849 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
2850   MachineBasicBlock::iterator I = MBB.end();
2851   unsigned Count = 0;
2852 
2853   while (I != MBB.begin()) {
2854     --I;
2855     if (I->isDebugValue())
2856       continue;
2857     if (I->getOpcode() != X86::JMP_4 &&
2858         getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2859       break;
2860     // Remove the branch.
2861     I->eraseFromParent();
2862     I = MBB.end();
2863     ++Count;
2864   }
2865 
2866   return Count;
2867 }
2868 
2869 unsigned
2870 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2871                            MachineBasicBlock *FBB,
2872                            const SmallVectorImpl<MachineOperand> &Cond,
2873                            DebugLoc DL) const {
2874   // Shouldn't be a fall through.
2875   assert(TBB && "InsertBranch must not be told to insert a fallthrough");
2876   assert((Cond.size() == 1 || Cond.size() == 0) &&
2877          "X86 branch conditions have one component!");
2878 
2879   if (Cond.empty()) {
2880     // Unconditional branch?
2881     assert(!FBB && "Unconditional branch with multiple successors!");
2882     BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
2883     return 1;
2884   }
2885 
2886   // Conditional branch.
2887   unsigned Count = 0;
2888   X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2889   switch (CC) {
2890   case X86::COND_NP_OR_E:
2891     // Synthesize NP_OR_E with two branches.
2892     BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
2893     ++Count;
2894     BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
2895     ++Count;
2896     break;
2897   case X86::COND_NE_OR_P:
2898     // Synthesize NE_OR_P with two branches.
2899     BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
2900     ++Count;
2901     BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
2902     ++Count;
2903     break;
2904   default: {
2905     unsigned Opc = GetCondBranchFromCond(CC);
2906     BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
2907     ++Count;
2908   }
2909   }
2910   if (FBB) {
2911     // Two-way Conditional branch. Insert the second branch.
2912     BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
2913     ++Count;
2914   }
2915   return Count;
2916 }
2917 
2918 bool X86InstrInfo::
2919 canInsertSelect(const MachineBasicBlock &MBB,
2920                 const SmallVectorImpl<MachineOperand> &Cond,
2921                 unsigned TrueReg, unsigned FalseReg,
2922                 int &CondCycles, int &TrueCycles, int &FalseCycles) const {
2923   // Not all subtargets have cmov instructions.
2924   if (!TM.getSubtarget<X86Subtarget>().hasCMov())
2925     return false;
2926   if (Cond.size() != 1)
2927     return false;
2928   // We cannot do the composite conditions, at least not in SSA form.
2929   if ((X86::CondCode)Cond[0].getImm() > X86::COND_S)
2930     return false;
2931 
2932   // Check register classes.
2933   const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2934   const TargetRegisterClass *RC =
2935     RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg));
2936   if (!RC)
2937     return false;
2938 
2939   // We have cmov instructions for 16, 32, and 64 bit general purpose registers.
2940   if (X86::GR16RegClass.hasSubClassEq(RC) ||
2941       X86::GR32RegClass.hasSubClassEq(RC) ||
2942       X86::GR64RegClass.hasSubClassEq(RC)) {
2943     // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy
2944     // Bridge. Probably Ivy Bridge as well.
2945     CondCycles = 2;
2946     TrueCycles = 2;
2947     FalseCycles = 2;
2948     return true;
2949   }
2950 
2951   // Can't do vectors.
2952   return false;
2953 }
2954 
2955 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB,
2956                                 MachineBasicBlock::iterator I, DebugLoc DL,
2957                                 unsigned DstReg,
2958                                 const SmallVectorImpl<MachineOperand> &Cond,
2959                                 unsigned TrueReg, unsigned FalseReg) const {
2960    MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2961    assert(Cond.size() == 1 && "Invalid Cond array");
2962    unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(),
2963                                   MRI.getRegClass(DstReg)->getSize(),
2964                                   false/*HasMemoryOperand*/);
2965    BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg);
2966 }
2967 
2968 /// isHReg - Test if the given register is a physical h register.
2969 static bool isHReg(unsigned Reg) {
2970   return X86::GR8_ABCD_HRegClass.contains(Reg);
2971 }
2972 
2973 // Try and copy between VR128/VR64 and GR64 registers.
2974 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2975                                         const X86Subtarget& Subtarget) {
2976 
2977 
2978   // SrcReg(VR128) -> DestReg(GR64)
2979   // SrcReg(VR64)  -> DestReg(GR64)
2980   // SrcReg(GR64)  -> DestReg(VR128)
2981   // SrcReg(GR64)  -> DestReg(VR64)
2982 
2983   bool HasAVX = Subtarget.hasAVX();
2984   bool HasAVX512 = Subtarget.hasAVX512();
2985   if (X86::GR64RegClass.contains(DestReg)) {
2986     if (X86::VR128XRegClass.contains(SrcReg))
2987       // Copy from a VR128 register to a GR64 register.
2988       return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr :
2989                                                X86::MOVPQIto64rr);
2990     if (X86::VR64RegClass.contains(SrcReg))
2991       // Copy from a VR64 register to a GR64 register.
2992       return X86::MOVSDto64rr;
2993   } else if (X86::GR64RegClass.contains(SrcReg)) {
2994     // Copy from a GR64 register to a VR128 register.
2995     if (X86::VR128XRegClass.contains(DestReg))
2996       return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr :
2997                                                X86::MOV64toPQIrr);
2998     // Copy from a GR64 register to a VR64 register.
2999     if (X86::VR64RegClass.contains(DestReg))
3000       return X86::MOV64toSDrr;
3001   }
3002 
3003   // SrcReg(FR32) -> DestReg(GR32)
3004   // SrcReg(GR32) -> DestReg(FR32)
3005 
3006   if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg))
3007     // Copy from a FR32 register to a GR32 register.
3008     return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr);
3009 
3010   if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
3011     // Copy from a GR32 register to a FR32 register.
3012     return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr);
3013   return 0;
3014 }
3015 
3016 static
3017 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg) {
3018   if (X86::VR128XRegClass.contains(DestReg, SrcReg) ||
3019       X86::VR256XRegClass.contains(DestReg, SrcReg) ||
3020       X86::VR512RegClass.contains(DestReg, SrcReg)) {
3021      DestReg = get512BitSuperRegister(DestReg);
3022      SrcReg = get512BitSuperRegister(SrcReg);
3023      return X86::VMOVAPSZrr;
3024   }
3025   if ((X86::VK8RegClass.contains(DestReg) ||
3026        X86::VK16RegClass.contains(DestReg)) &&
3027       (X86::VK8RegClass.contains(SrcReg) ||
3028        X86::VK16RegClass.contains(SrcReg)))
3029     return X86::KMOVWkk;
3030   return 0;
3031 }
3032 
3033 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
3034                                MachineBasicBlock::iterator MI, DebugLoc DL,
3035                                unsigned DestReg, unsigned SrcReg,
3036                                bool KillSrc) const {
3037   // First deal with the normal symmetric copies.
3038   bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3039   bool HasAVX512 = TM.getSubtarget<X86Subtarget>().hasAVX512();
3040   unsigned Opc = 0;
3041   if (X86::GR64RegClass.contains(DestReg, SrcReg))
3042     Opc = X86::MOV64rr;
3043   else if (X86::GR32RegClass.contains(DestReg, SrcReg))
3044     Opc = X86::MOV32rr;
3045   else if (X86::GR16RegClass.contains(DestReg, SrcReg))
3046     Opc = X86::MOV16rr;
3047   else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
3048     // Copying to or from a physical H register on x86-64 requires a NOREX
3049     // move.  Otherwise use a normal move.
3050     if ((isHReg(DestReg) || isHReg(SrcReg)) &&
3051         TM.getSubtarget<X86Subtarget>().is64Bit()) {
3052       Opc = X86::MOV8rr_NOREX;
3053       // Both operands must be encodable without an REX prefix.
3054       assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
3055              "8-bit H register can not be copied outside GR8_NOREX");
3056     } else
3057       Opc = X86::MOV8rr;
3058   }
3059   else if (X86::VR64RegClass.contains(DestReg, SrcReg))
3060     Opc = X86::MMX_MOVQ64rr;
3061   else if (HasAVX512)
3062     Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg);
3063   else if (X86::VR128RegClass.contains(DestReg, SrcReg))
3064     Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
3065   else if (X86::VR256RegClass.contains(DestReg, SrcReg))
3066     Opc = X86::VMOVAPSYrr;
3067   if (!Opc)
3068     Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, TM.getSubtarget<X86Subtarget>());
3069 
3070   if (Opc) {
3071     BuildMI(MBB, MI, DL, get(Opc), DestReg)
3072       .addReg(SrcReg, getKillRegState(KillSrc));
3073     return;
3074   }
3075 
3076   // Moving EFLAGS to / from another register requires a push and a pop.
3077   // Notice that we have to adjust the stack if we don't want to clobber the
3078   // first frame index. See X86FrameLowering.cpp - colobbersTheStack.
3079   if (SrcReg == X86::EFLAGS) {
3080     if (X86::GR64RegClass.contains(DestReg)) {
3081       BuildMI(MBB, MI, DL, get(X86::PUSHF64));
3082       BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
3083       return;
3084     }
3085     if (X86::GR32RegClass.contains(DestReg)) {
3086       BuildMI(MBB, MI, DL, get(X86::PUSHF32));
3087       BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
3088       return;
3089     }
3090   }
3091   if (DestReg == X86::EFLAGS) {
3092     if (X86::GR64RegClass.contains(SrcReg)) {
3093       BuildMI(MBB, MI, DL, get(X86::PUSH64r))
3094         .addReg(SrcReg, getKillRegState(KillSrc));
3095       BuildMI(MBB, MI, DL, get(X86::POPF64));
3096       return;
3097     }
3098     if (X86::GR32RegClass.contains(SrcReg)) {
3099       BuildMI(MBB, MI, DL, get(X86::PUSH32r))
3100         .addReg(SrcReg, getKillRegState(KillSrc));
3101       BuildMI(MBB, MI, DL, get(X86::POPF32));
3102       return;
3103     }
3104   }
3105 
3106   DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
3107                << " to " << RI.getName(DestReg) << '\n');
3108   llvm_unreachable("Cannot emit physreg copy instruction");
3109 }
3110 
3111 static unsigned getLoadStoreRegOpcode(unsigned Reg,
3112                                       const TargetRegisterClass *RC,
3113                                       bool isStackAligned,
3114                                       const TargetMachine &TM,
3115                                       bool load) {
3116   if (TM.getSubtarget<X86Subtarget>().hasAVX512()) {
3117     if (X86::VK8RegClass.hasSubClassEq(RC)  ||
3118       X86::VK16RegClass.hasSubClassEq(RC))
3119       return load ? X86::KMOVWkm : X86::KMOVWmk;
3120     if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC))
3121       return load ? X86::VMOVSSZrm : X86::VMOVSSZmr;
3122     if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC))
3123       return load ? X86::VMOVSDZrm : X86::VMOVSDZmr;
3124     if (X86::VR512RegClass.hasSubClassEq(RC))
3125       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3126   }
3127 
3128   bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3129   switch (RC->getSize()) {
3130   default:
3131     llvm_unreachable("Unknown spill size");
3132   case 1:
3133     assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
3134     if (TM.getSubtarget<X86Subtarget>().is64Bit())
3135       // Copying to or from a physical H register on x86-64 requires a NOREX
3136       // move.  Otherwise use a normal move.
3137       if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
3138         return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
3139     return load ? X86::MOV8rm : X86::MOV8mr;
3140   case 2:
3141     assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
3142     return load ? X86::MOV16rm : X86::MOV16mr;
3143   case 4:
3144     if (X86::GR32RegClass.hasSubClassEq(RC))
3145       return load ? X86::MOV32rm : X86::MOV32mr;
3146     if (X86::FR32RegClass.hasSubClassEq(RC))
3147       return load ?
3148         (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
3149         (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
3150     if (X86::RFP32RegClass.hasSubClassEq(RC))
3151       return load ? X86::LD_Fp32m : X86::ST_Fp32m;
3152     llvm_unreachable("Unknown 4-byte regclass");
3153   case 8:
3154     if (X86::GR64RegClass.hasSubClassEq(RC))
3155       return load ? X86::MOV64rm : X86::MOV64mr;
3156     if (X86::FR64RegClass.hasSubClassEq(RC))
3157       return load ?
3158         (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
3159         (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
3160     if (X86::VR64RegClass.hasSubClassEq(RC))
3161       return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
3162     if (X86::RFP64RegClass.hasSubClassEq(RC))
3163       return load ? X86::LD_Fp64m : X86::ST_Fp64m;
3164     llvm_unreachable("Unknown 8-byte regclass");
3165   case 10:
3166     assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
3167     return load ? X86::LD_Fp80m : X86::ST_FpP80m;
3168   case 16: {
3169     assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
3170     // If stack is realigned we can use aligned stores.
3171     if (isStackAligned)
3172       return load ?
3173         (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
3174         (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
3175     else
3176       return load ?
3177         (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
3178         (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
3179   }
3180   case 32:
3181     assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
3182     // If stack is realigned we can use aligned stores.
3183     if (isStackAligned)
3184       return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
3185     else
3186       return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
3187   case 64:
3188     assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass");
3189     if (isStackAligned)
3190       return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
3191     else
3192       return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
3193   }
3194 }
3195 
3196 static unsigned getStoreRegOpcode(unsigned SrcReg,
3197                                   const TargetRegisterClass *RC,
3198                                   bool isStackAligned,
3199                                   TargetMachine &TM) {
3200   return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
3201 }
3202 
3203 
3204 static unsigned getLoadRegOpcode(unsigned DestReg,
3205                                  const TargetRegisterClass *RC,
3206                                  bool isStackAligned,
3207                                  const TargetMachine &TM) {
3208   return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
3209 }
3210 
3211 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
3212                                        MachineBasicBlock::iterator MI,
3213                                        unsigned SrcReg, bool isKill, int FrameIdx,
3214                                        const TargetRegisterClass *RC,
3215                                        const TargetRegisterInfo *TRI) const {
3216   const MachineFunction &MF = *MBB.getParent();
3217   assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
3218          "Stack slot too small for store");
3219   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3220   bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3221     RI.canRealignStack(MF);
3222   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3223   DebugLoc DL = MBB.findDebugLoc(MI);
3224   addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
3225     .addReg(SrcReg, getKillRegState(isKill));
3226 }
3227 
3228 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
3229                                   bool isKill,
3230                                   SmallVectorImpl<MachineOperand> &Addr,
3231                                   const TargetRegisterClass *RC,
3232                                   MachineInstr::mmo_iterator MMOBegin,
3233                                   MachineInstr::mmo_iterator MMOEnd,
3234                                   SmallVectorImpl<MachineInstr*> &NewMIs) const {
3235   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3236   bool isAligned = MMOBegin != MMOEnd &&
3237                    (*MMOBegin)->getAlignment() >= Alignment;
3238   unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
3239   DebugLoc DL;
3240   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
3241   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3242     MIB.addOperand(Addr[i]);
3243   MIB.addReg(SrcReg, getKillRegState(isKill));
3244   (*MIB).setMemRefs(MMOBegin, MMOEnd);
3245   NewMIs.push_back(MIB);
3246 }
3247 
3248 
3249 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
3250                                         MachineBasicBlock::iterator MI,
3251                                         unsigned DestReg, int FrameIdx,
3252                                         const TargetRegisterClass *RC,
3253                                         const TargetRegisterInfo *TRI) const {
3254   const MachineFunction &MF = *MBB.getParent();
3255   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3256   bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
3257     RI.canRealignStack(MF);
3258   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3259   DebugLoc DL = MBB.findDebugLoc(MI);
3260   addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
3261 }
3262 
3263 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
3264                                  SmallVectorImpl<MachineOperand> &Addr,
3265                                  const TargetRegisterClass *RC,
3266                                  MachineInstr::mmo_iterator MMOBegin,
3267                                  MachineInstr::mmo_iterator MMOEnd,
3268                                  SmallVectorImpl<MachineInstr*> &NewMIs) const {
3269   unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16);
3270   bool isAligned = MMOBegin != MMOEnd &&
3271                    (*MMOBegin)->getAlignment() >= Alignment;
3272   unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
3273   DebugLoc DL;
3274   MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
3275   for (unsigned i = 0, e = Addr.size(); i != e; ++i)
3276     MIB.addOperand(Addr[i]);
3277   (*MIB).setMemRefs(MMOBegin, MMOEnd);
3278   NewMIs.push_back(MIB);
3279 }
3280 
3281 bool X86InstrInfo::
3282 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2,
3283                int &CmpMask, int &CmpValue) const {
3284   switch (MI->getOpcode()) {
3285   default: break;
3286   case X86::CMP64ri32:
3287   case X86::CMP64ri8:
3288   case X86::CMP32ri:
3289   case X86::CMP32ri8:
3290   case X86::CMP16ri:
3291   case X86::CMP16ri8:
3292   case X86::CMP8ri:
3293     SrcReg = MI->getOperand(0).getReg();
3294     SrcReg2 = 0;
3295     CmpMask = ~0;
3296     CmpValue = MI->getOperand(1).getImm();
3297     return true;
3298   // A SUB can be used to perform comparison.
3299   case X86::SUB64rm:
3300   case X86::SUB32rm:
3301   case X86::SUB16rm:
3302   case X86::SUB8rm:
3303     SrcReg = MI->getOperand(1).getReg();
3304     SrcReg2 = 0;
3305     CmpMask = ~0;
3306     CmpValue = 0;
3307     return true;
3308   case X86::SUB64rr:
3309   case X86::SUB32rr:
3310   case X86::SUB16rr:
3311   case X86::SUB8rr:
3312     SrcReg = MI->getOperand(1).getReg();
3313     SrcReg2 = MI->getOperand(2).getReg();
3314     CmpMask = ~0;
3315     CmpValue = 0;
3316     return true;
3317   case X86::SUB64ri32:
3318   case X86::SUB64ri8:
3319   case X86::SUB32ri:
3320   case X86::SUB32ri8:
3321   case X86::SUB16ri:
3322   case X86::SUB16ri8:
3323   case X86::SUB8ri:
3324     SrcReg = MI->getOperand(1).getReg();
3325     SrcReg2 = 0;
3326     CmpMask = ~0;
3327     CmpValue = MI->getOperand(2).getImm();
3328     return true;
3329   case X86::CMP64rr:
3330   case X86::CMP32rr:
3331   case X86::CMP16rr:
3332   case X86::CMP8rr:
3333     SrcReg = MI->getOperand(0).getReg();
3334     SrcReg2 = MI->getOperand(1).getReg();
3335     CmpMask = ~0;
3336     CmpValue = 0;
3337     return true;
3338   case X86::TEST8rr:
3339   case X86::TEST16rr:
3340   case X86::TEST32rr:
3341   case X86::TEST64rr:
3342     SrcReg = MI->getOperand(0).getReg();
3343     if (MI->getOperand(1).getReg() != SrcReg) return false;
3344     // Compare against zero.
3345     SrcReg2 = 0;
3346     CmpMask = ~0;
3347     CmpValue = 0;
3348     return true;
3349   }
3350   return false;
3351 }
3352 
3353 /// isRedundantFlagInstr - check whether the first instruction, whose only
3354 /// purpose is to update flags, can be made redundant.
3355 /// CMPrr can be made redundant by SUBrr if the operands are the same.
3356 /// This function can be extended later on.
3357 /// SrcReg, SrcRegs: register operands for FlagI.
3358 /// ImmValue: immediate for FlagI if it takes an immediate.
3359 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg,
3360                                         unsigned SrcReg2, int ImmValue,
3361                                         MachineInstr *OI) {
3362   if (((FlagI->getOpcode() == X86::CMP64rr &&
3363         OI->getOpcode() == X86::SUB64rr) ||
3364        (FlagI->getOpcode() == X86::CMP32rr &&
3365         OI->getOpcode() == X86::SUB32rr)||
3366        (FlagI->getOpcode() == X86::CMP16rr &&
3367         OI->getOpcode() == X86::SUB16rr)||
3368        (FlagI->getOpcode() == X86::CMP8rr &&
3369         OI->getOpcode() == X86::SUB8rr)) &&
3370       ((OI->getOperand(1).getReg() == SrcReg &&
3371         OI->getOperand(2).getReg() == SrcReg2) ||
3372        (OI->getOperand(1).getReg() == SrcReg2 &&
3373         OI->getOperand(2).getReg() == SrcReg)))
3374     return true;
3375 
3376   if (((FlagI->getOpcode() == X86::CMP64ri32 &&
3377         OI->getOpcode() == X86::SUB64ri32) ||
3378        (FlagI->getOpcode() == X86::CMP64ri8 &&
3379         OI->getOpcode() == X86::SUB64ri8) ||
3380        (FlagI->getOpcode() == X86::CMP32ri &&
3381         OI->getOpcode() == X86::SUB32ri) ||
3382        (FlagI->getOpcode() == X86::CMP32ri8 &&
3383         OI->getOpcode() == X86::SUB32ri8) ||
3384        (FlagI->getOpcode() == X86::CMP16ri &&
3385         OI->getOpcode() == X86::SUB16ri) ||
3386        (FlagI->getOpcode() == X86::CMP16ri8 &&
3387         OI->getOpcode() == X86::SUB16ri8) ||
3388        (FlagI->getOpcode() == X86::CMP8ri &&
3389         OI->getOpcode() == X86::SUB8ri)) &&
3390       OI->getOperand(1).getReg() == SrcReg &&
3391       OI->getOperand(2).getImm() == ImmValue)
3392     return true;
3393   return false;
3394 }
3395 
3396 /// isDefConvertible - check whether the definition can be converted
3397 /// to remove a comparison against zero.
3398 inline static bool isDefConvertible(MachineInstr *MI) {
3399   switch (MI->getOpcode()) {
3400   default: return false;
3401 
3402   // The shift instructions only modify ZF if their shift count is non-zero.
3403   // N.B.: The processor truncates the shift count depending on the encoding.
3404   case X86::SAR8ri:    case X86::SAR16ri:  case X86::SAR32ri:case X86::SAR64ri:
3405   case X86::SHR8ri:    case X86::SHR16ri:  case X86::SHR32ri:case X86::SHR64ri:
3406      return getTruncatedShiftCount(MI, 2) != 0;
3407 
3408   // Some left shift instructions can be turned into LEA instructions but only
3409   // if their flags aren't used. Avoid transforming such instructions.
3410   case X86::SHL8ri:    case X86::SHL16ri:  case X86::SHL32ri:case X86::SHL64ri:{
3411     unsigned ShAmt = getTruncatedShiftCount(MI, 2);
3412     if (isTruncatedShiftCountForLEA(ShAmt)) return false;
3413     return ShAmt != 0;
3414   }
3415 
3416   case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8:
3417   case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8:
3418      return getTruncatedShiftCount(MI, 3) != 0;
3419 
3420   case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri:
3421   case X86::SUB32ri8:  case X86::SUB16ri:  case X86::SUB16ri8:
3422   case X86::SUB8ri:    case X86::SUB64rr:  case X86::SUB32rr:
3423   case X86::SUB16rr:   case X86::SUB8rr:   case X86::SUB64rm:
3424   case X86::SUB32rm:   case X86::SUB16rm:  case X86::SUB8rm:
3425   case X86::DEC64r:    case X86::DEC32r:   case X86::DEC16r: case X86::DEC8r:
3426   case X86::DEC64_32r: case X86::DEC64_16r:
3427   case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri:
3428   case X86::ADD32ri8:  case X86::ADD16ri:  case X86::ADD16ri8:
3429   case X86::ADD8ri:    case X86::ADD64rr:  case X86::ADD32rr:
3430   case X86::ADD16rr:   case X86::ADD8rr:   case X86::ADD64rm:
3431   case X86::ADD32rm:   case X86::ADD16rm:  case X86::ADD8rm:
3432   case X86::INC64r:    case X86::INC32r:   case X86::INC16r: case X86::INC8r:
3433   case X86::INC64_32r: case X86::INC64_16r:
3434   case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri:
3435   case X86::AND32ri8:  case X86::AND16ri:  case X86::AND16ri8:
3436   case X86::AND8ri:    case X86::AND64rr:  case X86::AND32rr:
3437   case X86::AND16rr:   case X86::AND8rr:   case X86::AND64rm:
3438   case X86::AND32rm:   case X86::AND16rm:  case X86::AND8rm:
3439   case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri:
3440   case X86::XOR32ri8:  case X86::XOR16ri:  case X86::XOR16ri8:
3441   case X86::XOR8ri:    case X86::XOR64rr:  case X86::XOR32rr:
3442   case X86::XOR16rr:   case X86::XOR8rr:   case X86::XOR64rm:
3443   case X86::XOR32rm:   case X86::XOR16rm:  case X86::XOR8rm:
3444   case X86::OR64ri32:  case X86::OR64ri8:  case X86::OR32ri:
3445   case X86::OR32ri8:   case X86::OR16ri:   case X86::OR16ri8:
3446   case X86::OR8ri:     case X86::OR64rr:   case X86::OR32rr:
3447   case X86::OR16rr:    case X86::OR8rr:    case X86::OR64rm:
3448   case X86::OR32rm:    case X86::OR16rm:   case X86::OR8rm:
3449   case X86::NEG8r:     case X86::NEG16r:   case X86::NEG32r: case X86::NEG64r:
3450   case X86::SAR8r1:    case X86::SAR16r1:  case X86::SAR32r1:case X86::SAR64r1:
3451   case X86::SHR8r1:    case X86::SHR16r1:  case X86::SHR32r1:case X86::SHR64r1:
3452   case X86::SHL8r1:    case X86::SHL16r1:  case X86::SHL32r1:case X86::SHL64r1:
3453   case X86::ADC32ri:   case X86::ADC32ri8:
3454   case X86::ADC32rr:   case X86::ADC64ri32:
3455   case X86::ADC64ri8:  case X86::ADC64rr:
3456   case X86::SBB32ri:   case X86::SBB32ri8:
3457   case X86::SBB32rr:   case X86::SBB64ri32:
3458   case X86::SBB64ri8:  case X86::SBB64rr:
3459   case X86::ANDN32rr:  case X86::ANDN32rm:
3460   case X86::ANDN64rr:  case X86::ANDN64rm:
3461   case X86::BEXTR32rr: case X86::BEXTR64rr:
3462   case X86::BEXTR32rm: case X86::BEXTR64rm:
3463   case X86::BLSI32rr:  case X86::BLSI32rm:
3464   case X86::BLSI64rr:  case X86::BLSI64rm:
3465   case X86::BLSMSK32rr:case X86::BLSMSK32rm:
3466   case X86::BLSMSK64rr:case X86::BLSMSK64rm:
3467   case X86::BLSR32rr:  case X86::BLSR32rm:
3468   case X86::BLSR64rr:  case X86::BLSR64rm:
3469   case X86::BZHI32rr:  case X86::BZHI32rm:
3470   case X86::BZHI64rr:  case X86::BZHI64rm:
3471   case X86::LZCNT16rr: case X86::LZCNT16rm:
3472   case X86::LZCNT32rr: case X86::LZCNT32rm:
3473   case X86::LZCNT64rr: case X86::LZCNT64rm:
3474   case X86::POPCNT16rr:case X86::POPCNT16rm:
3475   case X86::POPCNT32rr:case X86::POPCNT32rm:
3476   case X86::POPCNT64rr:case X86::POPCNT64rm:
3477   case X86::TZCNT16rr: case X86::TZCNT16rm:
3478   case X86::TZCNT32rr: case X86::TZCNT32rm:
3479   case X86::TZCNT64rr: case X86::TZCNT64rm:
3480     return true;
3481   }
3482 }
3483 
3484 /// optimizeCompareInstr - Check if there exists an earlier instruction that
3485 /// operates on the same source operands and sets flags in the same way as
3486 /// Compare; remove Compare if possible.
3487 bool X86InstrInfo::
3488 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
3489                      int CmpMask, int CmpValue,
3490                      const MachineRegisterInfo *MRI) const {
3491   // Check whether we can replace SUB with CMP.
3492   unsigned NewOpcode = 0;
3493   switch (CmpInstr->getOpcode()) {
3494   default: break;
3495   case X86::SUB64ri32:
3496   case X86::SUB64ri8:
3497   case X86::SUB32ri:
3498   case X86::SUB32ri8:
3499   case X86::SUB16ri:
3500   case X86::SUB16ri8:
3501   case X86::SUB8ri:
3502   case X86::SUB64rm:
3503   case X86::SUB32rm:
3504   case X86::SUB16rm:
3505   case X86::SUB8rm:
3506   case X86::SUB64rr:
3507   case X86::SUB32rr:
3508   case X86::SUB16rr:
3509   case X86::SUB8rr: {
3510     if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg()))
3511       return false;
3512     // There is no use of the destination register, we can replace SUB with CMP.
3513     switch (CmpInstr->getOpcode()) {
3514     default: llvm_unreachable("Unreachable!");
3515     case X86::SUB64rm:   NewOpcode = X86::CMP64rm;   break;
3516     case X86::SUB32rm:   NewOpcode = X86::CMP32rm;   break;
3517     case X86::SUB16rm:   NewOpcode = X86::CMP16rm;   break;
3518     case X86::SUB8rm:    NewOpcode = X86::CMP8rm;    break;
3519     case X86::SUB64rr:   NewOpcode = X86::CMP64rr;   break;
3520     case X86::SUB32rr:   NewOpcode = X86::CMP32rr;   break;
3521     case X86::SUB16rr:   NewOpcode = X86::CMP16rr;   break;
3522     case X86::SUB8rr:    NewOpcode = X86::CMP8rr;    break;
3523     case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break;
3524     case X86::SUB64ri8:  NewOpcode = X86::CMP64ri8;  break;
3525     case X86::SUB32ri:   NewOpcode = X86::CMP32ri;   break;
3526     case X86::SUB32ri8:  NewOpcode = X86::CMP32ri8;  break;
3527     case X86::SUB16ri:   NewOpcode = X86::CMP16ri;   break;
3528     case X86::SUB16ri8:  NewOpcode = X86::CMP16ri8;  break;
3529     case X86::SUB8ri:    NewOpcode = X86::CMP8ri;    break;
3530     }
3531     CmpInstr->setDesc(get(NewOpcode));
3532     CmpInstr->RemoveOperand(0);
3533     // Fall through to optimize Cmp if Cmp is CMPrr or CMPri.
3534     if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm ||
3535         NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm)
3536       return false;
3537   }
3538   }
3539 
3540   // Get the unique definition of SrcReg.
3541   MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg);
3542   if (!MI) return false;
3543 
3544   // CmpInstr is the first instruction of the BB.
3545   MachineBasicBlock::iterator I = CmpInstr, Def = MI;
3546 
3547   // If we are comparing against zero, check whether we can use MI to update
3548   // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize.
3549   bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0);
3550   if (IsCmpZero && (MI->getParent() != CmpInstr->getParent() ||
3551       !isDefConvertible(MI)))
3552     return false;
3553 
3554   // We are searching for an earlier instruction that can make CmpInstr
3555   // redundant and that instruction will be saved in Sub.
3556   MachineInstr *Sub = NULL;
3557   const TargetRegisterInfo *TRI = &getRegisterInfo();
3558 
3559   // We iterate backward, starting from the instruction before CmpInstr and
3560   // stop when reaching the definition of a source register or done with the BB.
3561   // RI points to the instruction before CmpInstr.
3562   // If the definition is in this basic block, RE points to the definition;
3563   // otherwise, RE is the rend of the basic block.
3564   MachineBasicBlock::reverse_iterator
3565       RI = MachineBasicBlock::reverse_iterator(I),
3566       RE = CmpInstr->getParent() == MI->getParent() ?
3567            MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ :
3568            CmpInstr->getParent()->rend();
3569   MachineInstr *Movr0Inst = 0;
3570   for (; RI != RE; ++RI) {
3571     MachineInstr *Instr = &*RI;
3572     // Check whether CmpInstr can be made redundant by the current instruction.
3573     if (!IsCmpZero &&
3574         isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) {
3575       Sub = Instr;
3576       break;
3577     }
3578 
3579     if (Instr->modifiesRegister(X86::EFLAGS, TRI) ||
3580         Instr->readsRegister(X86::EFLAGS, TRI)) {
3581       // This instruction modifies or uses EFLAGS.
3582 
3583       // MOV32r0 etc. are implemented with xor which clobbers condition code.
3584       // They are safe to move up, if the definition to EFLAGS is dead and
3585       // earlier instructions do not read or write EFLAGS.
3586       if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 &&
3587           Instr->registerDefIsDead(X86::EFLAGS, TRI)) {
3588         Movr0Inst = Instr;
3589         continue;
3590       }
3591 
3592       // We can't remove CmpInstr.
3593       return false;
3594     }
3595   }
3596 
3597   // Return false if no candidates exist.
3598   if (!IsCmpZero && !Sub)
3599     return false;
3600 
3601   bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 &&
3602                     Sub->getOperand(2).getReg() == SrcReg);
3603 
3604   // Scan forward from the instruction after CmpInstr for uses of EFLAGS.
3605   // It is safe to remove CmpInstr if EFLAGS is redefined or killed.
3606   // If we are done with the basic block, we need to check whether EFLAGS is
3607   // live-out.
3608   bool IsSafe = false;
3609   SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate;
3610   MachineBasicBlock::iterator E = CmpInstr->getParent()->end();
3611   for (++I; I != E; ++I) {
3612     const MachineInstr &Instr = *I;
3613     bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI);
3614     bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI);
3615     // We should check the usage if this instruction uses and updates EFLAGS.
3616     if (!UseEFLAGS && ModifyEFLAGS) {
3617       // It is safe to remove CmpInstr if EFLAGS is updated again.
3618       IsSafe = true;
3619       break;
3620     }
3621     if (!UseEFLAGS && !ModifyEFLAGS)
3622       continue;
3623 
3624     // EFLAGS is used by this instruction.
3625     X86::CondCode OldCC;
3626     bool OpcIsSET = false;
3627     if (IsCmpZero || IsSwapped) {
3628       // We decode the condition code from opcode.
3629       if (Instr.isBranch())
3630         OldCC = getCondFromBranchOpc(Instr.getOpcode());
3631       else {
3632         OldCC = getCondFromSETOpc(Instr.getOpcode());
3633         if (OldCC != X86::COND_INVALID)
3634           OpcIsSET = true;
3635         else
3636           OldCC = X86::getCondFromCMovOpc(Instr.getOpcode());
3637       }
3638       if (OldCC == X86::COND_INVALID) return false;
3639     }
3640     if (IsCmpZero) {
3641       switch (OldCC) {
3642       default: break;
3643       case X86::COND_A: case X86::COND_AE:
3644       case X86::COND_B: case X86::COND_BE:
3645       case X86::COND_G: case X86::COND_GE:
3646       case X86::COND_L: case X86::COND_LE:
3647       case X86::COND_O: case X86::COND_NO:
3648         // CF and OF are used, we can't perform this optimization.
3649         return false;
3650       }
3651     } else if (IsSwapped) {
3652       // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs
3653       // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc.
3654       // We swap the condition code and synthesize the new opcode.
3655       X86::CondCode NewCC = getSwappedCondition(OldCC);
3656       if (NewCC == X86::COND_INVALID) return false;
3657 
3658       // Synthesize the new opcode.
3659       bool HasMemoryOperand = Instr.hasOneMemOperand();
3660       unsigned NewOpc;
3661       if (Instr.isBranch())
3662         NewOpc = GetCondBranchFromCond(NewCC);
3663       else if(OpcIsSET)
3664         NewOpc = getSETFromCond(NewCC, HasMemoryOperand);
3665       else {
3666         unsigned DstReg = Instr.getOperand(0).getReg();
3667         NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(),
3668                                  HasMemoryOperand);
3669       }
3670 
3671       // Push the MachineInstr to OpsToUpdate.
3672       // If it is safe to remove CmpInstr, the condition code of these
3673       // instructions will be modified.
3674       OpsToUpdate.push_back(std::make_pair(&*I, NewOpc));
3675     }
3676     if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) {
3677       // It is safe to remove CmpInstr if EFLAGS is updated again or killed.
3678       IsSafe = true;
3679       break;
3680     }
3681   }
3682 
3683   // If EFLAGS is not killed nor re-defined, we should check whether it is
3684   // live-out. If it is live-out, do not optimize.
3685   if ((IsCmpZero || IsSwapped) && !IsSafe) {
3686     MachineBasicBlock *MBB = CmpInstr->getParent();
3687     for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
3688              SE = MBB->succ_end(); SI != SE; ++SI)
3689       if ((*SI)->isLiveIn(X86::EFLAGS))
3690         return false;
3691   }
3692 
3693   // The instruction to be updated is either Sub or MI.
3694   Sub = IsCmpZero ? MI : Sub;
3695   // Move Movr0Inst to the appropriate place before Sub.
3696   if (Movr0Inst) {
3697     // Look backwards until we find a def that doesn't use the current EFLAGS.
3698     Def = Sub;
3699     MachineBasicBlock::reverse_iterator
3700       InsertI = MachineBasicBlock::reverse_iterator(++Def),
3701                 InsertE = Sub->getParent()->rend();
3702     for (; InsertI != InsertE; ++InsertI) {
3703       MachineInstr *Instr = &*InsertI;
3704       if (!Instr->readsRegister(X86::EFLAGS, TRI) &&
3705           Instr->modifiesRegister(X86::EFLAGS, TRI)) {
3706         Sub->getParent()->remove(Movr0Inst);
3707         Instr->getParent()->insert(MachineBasicBlock::iterator(Instr),
3708                                    Movr0Inst);
3709         break;
3710       }
3711     }
3712     if (InsertI == InsertE)
3713       return false;
3714   }
3715 
3716   // Make sure Sub instruction defines EFLAGS and mark the def live.
3717   unsigned i = 0, e = Sub->getNumOperands();
3718   for (; i != e; ++i) {
3719     MachineOperand &MO = Sub->getOperand(i);
3720     if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) {
3721       MO.setIsDead(false);
3722       break;
3723     }
3724   }
3725   assert(i != e && "Unable to locate a def EFLAGS operand");
3726 
3727   CmpInstr->eraseFromParent();
3728 
3729   // Modify the condition code of instructions in OpsToUpdate.
3730   for (unsigned i = 0, e = OpsToUpdate.size(); i < e; i++)
3731     OpsToUpdate[i].first->setDesc(get(OpsToUpdate[i].second));
3732   return true;
3733 }
3734 
3735 /// optimizeLoadInstr - Try to remove the load by folding it to a register
3736 /// operand at the use. We fold the load instructions if load defines a virtual
3737 /// register, the virtual register is used once in the same BB, and the
3738 /// instructions in-between do not load or store, and have no side effects.
3739 MachineInstr* X86InstrInfo::
3740 optimizeLoadInstr(MachineInstr *MI, const MachineRegisterInfo *MRI,
3741                   unsigned &FoldAsLoadDefReg,
3742                   MachineInstr *&DefMI) const {
3743   if (FoldAsLoadDefReg == 0)
3744     return 0;
3745   // To be conservative, if there exists another load, clear the load candidate.
3746   if (MI->mayLoad()) {
3747     FoldAsLoadDefReg = 0;
3748     return 0;
3749   }
3750 
3751   // Check whether we can move DefMI here.
3752   DefMI = MRI->getVRegDef(FoldAsLoadDefReg);
3753   assert(DefMI);
3754   bool SawStore = false;
3755   if (!DefMI->isSafeToMove(this, 0, SawStore))
3756     return 0;
3757 
3758   // We try to commute MI if possible.
3759   unsigned IdxEnd = (MI->isCommutable()) ? 2 : 1;
3760   for (unsigned Idx = 0; Idx < IdxEnd; Idx++) {
3761     // Collect information about virtual register operands of MI.
3762     unsigned SrcOperandId = 0;
3763     bool FoundSrcOperand = false;
3764     for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
3765       MachineOperand &MO = MI->getOperand(i);
3766       if (!MO.isReg())
3767         continue;
3768       unsigned Reg = MO.getReg();
3769       if (Reg != FoldAsLoadDefReg)
3770         continue;
3771       // Do not fold if we have a subreg use or a def or multiple uses.
3772       if (MO.getSubReg() || MO.isDef() || FoundSrcOperand)
3773         return 0;
3774 
3775       SrcOperandId = i;
3776       FoundSrcOperand = true;
3777     }
3778     if (!FoundSrcOperand) return 0;
3779 
3780     // Check whether we can fold the def into SrcOperandId.
3781     SmallVector<unsigned, 8> Ops;
3782     Ops.push_back(SrcOperandId);
3783     MachineInstr *FoldMI = foldMemoryOperand(MI, Ops, DefMI);
3784     if (FoldMI) {
3785       FoldAsLoadDefReg = 0;
3786       return FoldMI;
3787     }
3788 
3789     if (Idx == 1) {
3790       // MI was changed but it didn't help, commute it back!
3791       commuteInstruction(MI, false);
3792       return 0;
3793     }
3794 
3795     // Check whether we can commute MI and enable folding.
3796     if (MI->isCommutable()) {
3797       MachineInstr *NewMI = commuteInstruction(MI, false);
3798       // Unable to commute.
3799       if (!NewMI) return 0;
3800       if (NewMI != MI) {
3801         // New instruction. It doesn't need to be kept.
3802         NewMI->eraseFromParent();
3803         return 0;
3804       }
3805     }
3806   }
3807   return 0;
3808 }
3809 
3810 /// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
3811 /// instruction with two undef reads of the register being defined.  This is
3812 /// used for mapping:
3813 ///   %xmm4 = V_SET0
3814 /// to:
3815 ///   %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
3816 ///
3817 static bool Expand2AddrUndef(MachineInstrBuilder &MIB,
3818                              const MCInstrDesc &Desc) {
3819   assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
3820   unsigned Reg = MIB->getOperand(0).getReg();
3821   MIB->setDesc(Desc);
3822 
3823   // MachineInstr::addOperand() will insert explicit operands before any
3824   // implicit operands.
3825   MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3826   // But we don't trust that.
3827   assert(MIB->getOperand(1).getReg() == Reg &&
3828          MIB->getOperand(2).getReg() == Reg && "Misplaced operand");
3829   return true;
3830 }
3831 
3832 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
3833   bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3834   MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
3835   switch (MI->getOpcode()) {
3836   case X86::SETB_C8r:
3837     return Expand2AddrUndef(MIB, get(X86::SBB8rr));
3838   case X86::SETB_C16r:
3839     return Expand2AddrUndef(MIB, get(X86::SBB16rr));
3840   case X86::SETB_C32r:
3841     return Expand2AddrUndef(MIB, get(X86::SBB32rr));
3842   case X86::SETB_C64r:
3843     return Expand2AddrUndef(MIB, get(X86::SBB64rr));
3844   case X86::V_SET0:
3845   case X86::FsFLD0SS:
3846   case X86::FsFLD0SD:
3847     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
3848   case X86::AVX_SET0:
3849     assert(HasAVX && "AVX not supported");
3850     return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
3851   case X86::AVX512_512_SET0:
3852     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
3853   case X86::V_SETALLONES:
3854     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
3855   case X86::AVX2_SETALLONES:
3856     return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
3857   case X86::TEST8ri_NOREX:
3858     MI->setDesc(get(X86::TEST8ri));
3859     return true;
3860   case X86::KSET0W: return Expand2AddrUndef(MIB, get(X86::KXORWrr));
3861   case X86::KSET1B:
3862   case X86::KSET1W: return Expand2AddrUndef(MIB, get(X86::KXNORWrr));
3863   }
3864   return false;
3865 }
3866 
3867 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
3868                                      const SmallVectorImpl<MachineOperand> &MOs,
3869                                      MachineInstr *MI,
3870                                      const TargetInstrInfo &TII) {
3871   // Create the base instruction with the memory operand as the first part.
3872   // Omit the implicit operands, something BuildMI can't do.
3873   MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3874                                               MI->getDebugLoc(), true);
3875   MachineInstrBuilder MIB(MF, NewMI);
3876   unsigned NumAddrOps = MOs.size();
3877   for (unsigned i = 0; i != NumAddrOps; ++i)
3878     MIB.addOperand(MOs[i]);
3879   if (NumAddrOps < 4)  // FrameIndex only
3880     addOffset(MIB, 0);
3881 
3882   // Loop over the rest of the ri operands, converting them over.
3883   unsigned NumOps = MI->getDesc().getNumOperands()-2;
3884   for (unsigned i = 0; i != NumOps; ++i) {
3885     MachineOperand &MO = MI->getOperand(i+2);
3886     MIB.addOperand(MO);
3887   }
3888   for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
3889     MachineOperand &MO = MI->getOperand(i);
3890     MIB.addOperand(MO);
3891   }
3892   return MIB;
3893 }
3894 
3895 static MachineInstr *FuseInst(MachineFunction &MF,
3896                               unsigned Opcode, unsigned OpNo,
3897                               const SmallVectorImpl<MachineOperand> &MOs,
3898                               MachineInstr *MI, const TargetInstrInfo &TII) {
3899   // Omit the implicit operands, something BuildMI can't do.
3900   MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
3901                                               MI->getDebugLoc(), true);
3902   MachineInstrBuilder MIB(MF, NewMI);
3903 
3904   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3905     MachineOperand &MO = MI->getOperand(i);
3906     if (i == OpNo) {
3907       assert(MO.isReg() && "Expected to fold into reg operand!");
3908       unsigned NumAddrOps = MOs.size();
3909       for (unsigned i = 0; i != NumAddrOps; ++i)
3910         MIB.addOperand(MOs[i]);
3911       if (NumAddrOps < 4)  // FrameIndex only
3912         addOffset(MIB, 0);
3913     } else {
3914       MIB.addOperand(MO);
3915     }
3916   }
3917   return MIB;
3918 }
3919 
3920 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
3921                                 const SmallVectorImpl<MachineOperand> &MOs,
3922                                 MachineInstr *MI) {
3923   MachineFunction &MF = *MI->getParent()->getParent();
3924   MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
3925 
3926   unsigned NumAddrOps = MOs.size();
3927   for (unsigned i = 0; i != NumAddrOps; ++i)
3928     MIB.addOperand(MOs[i]);
3929   if (NumAddrOps < 4)  // FrameIndex only
3930     addOffset(MIB, 0);
3931   return MIB.addImm(0);
3932 }
3933 
3934 MachineInstr*
3935 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3936                                     MachineInstr *MI, unsigned i,
3937                                     const SmallVectorImpl<MachineOperand> &MOs,
3938                                     unsigned Size, unsigned Align) const {
3939   const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
3940   bool isCallRegIndirect = TM.getSubtarget<X86Subtarget>().callRegIndirect();
3941   bool isTwoAddrFold = false;
3942 
3943   // Atom favors register form of call. So, we do not fold loads into calls
3944   // when X86Subtarget is Atom.
3945   if (isCallRegIndirect &&
3946     (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r)) {
3947     return NULL;
3948   }
3949 
3950   unsigned NumOps = MI->getDesc().getNumOperands();
3951   bool isTwoAddr = NumOps > 1 &&
3952     MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
3953 
3954   // FIXME: AsmPrinter doesn't know how to handle
3955   // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3956   if (MI->getOpcode() == X86::ADD32ri &&
3957       MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3958     return NULL;
3959 
3960   MachineInstr *NewMI = NULL;
3961   // Folding a memory location into the two-address part of a two-address
3962   // instruction is different than folding it other places.  It requires
3963   // replacing the *two* registers with the memory location.
3964   if (isTwoAddr && NumOps >= 2 && i < 2 &&
3965       MI->getOperand(0).isReg() &&
3966       MI->getOperand(1).isReg() &&
3967       MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
3968     OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3969     isTwoAddrFold = true;
3970   } else if (i == 0) { // If operand 0
3971     if (MI->getOpcode() == X86::MOV32r0) {
3972       NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
3973       if (NewMI)
3974         return NewMI;
3975     }
3976 
3977     OpcodeTablePtr = &RegOp2MemOpTable0;
3978   } else if (i == 1) {
3979     OpcodeTablePtr = &RegOp2MemOpTable1;
3980   } else if (i == 2) {
3981     OpcodeTablePtr = &RegOp2MemOpTable2;
3982   } else if (i == 3) {
3983     OpcodeTablePtr = &RegOp2MemOpTable3;
3984   }
3985 
3986   // If table selected...
3987   if (OpcodeTablePtr) {
3988     // Find the Opcode to fuse
3989     DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3990       OpcodeTablePtr->find(MI->getOpcode());
3991     if (I != OpcodeTablePtr->end()) {
3992       unsigned Opcode = I->second.first;
3993       unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
3994       if (Align < MinAlign)
3995         return NULL;
3996       bool NarrowToMOV32rm = false;
3997       if (Size) {
3998         unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
3999         if (Size < RCSize) {
4000           // Check if it's safe to fold the load. If the size of the object is
4001           // narrower than the load width, then it's not.
4002           if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
4003             return NULL;
4004           // If this is a 64-bit load, but the spill slot is 32, then we can do
4005           // a 32-bit load which is implicitly zero-extended. This likely is due
4006           // to liveintervalanalysis remat'ing a load from stack slot.
4007           if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
4008             return NULL;
4009           Opcode = X86::MOV32rm;
4010           NarrowToMOV32rm = true;
4011         }
4012       }
4013 
4014       if (isTwoAddrFold)
4015         NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
4016       else
4017         NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
4018 
4019       if (NarrowToMOV32rm) {
4020         // If this is the special case where we use a MOV32rm to load a 32-bit
4021         // value and zero-extend the top bits. Change the destination register
4022         // to a 32-bit one.
4023         unsigned DstReg = NewMI->getOperand(0).getReg();
4024         if (TargetRegisterInfo::isPhysicalRegister(DstReg))
4025           NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
4026                                                    X86::sub_32bit));
4027         else
4028           NewMI->getOperand(0).setSubReg(X86::sub_32bit);
4029       }
4030       return NewMI;
4031     }
4032   }
4033 
4034   // No fusion
4035   if (PrintFailedFusing && !MI->isCopy())
4036     dbgs() << "We failed to fuse operand " << i << " in " << *MI;
4037   return NULL;
4038 }
4039 
4040 /// hasPartialRegUpdate - Return true for all instructions that only update
4041 /// the first 32 or 64-bits of the destination register and leave the rest
4042 /// unmodified. This can be used to avoid folding loads if the instructions
4043 /// only update part of the destination register, and the non-updated part is
4044 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
4045 /// instructions breaks the partial register dependency and it can improve
4046 /// performance. e.g.:
4047 ///
4048 ///   movss (%rdi), %xmm0
4049 ///   cvtss2sd %xmm0, %xmm0
4050 ///
4051 /// Instead of
4052 ///   cvtss2sd (%rdi), %xmm0
4053 ///
4054 /// FIXME: This should be turned into a TSFlags.
4055 ///
4056 static bool hasPartialRegUpdate(unsigned Opcode) {
4057   switch (Opcode) {
4058   case X86::CVTSI2SSrr:
4059   case X86::CVTSI2SS64rr:
4060   case X86::CVTSI2SDrr:
4061   case X86::CVTSI2SD64rr:
4062   case X86::CVTSD2SSrr:
4063   case X86::Int_CVTSD2SSrr:
4064   case X86::CVTSS2SDrr:
4065   case X86::Int_CVTSS2SDrr:
4066   case X86::RCPSSr:
4067   case X86::RCPSSr_Int:
4068   case X86::ROUNDSDr:
4069   case X86::ROUNDSDr_Int:
4070   case X86::ROUNDSSr:
4071   case X86::ROUNDSSr_Int:
4072   case X86::RSQRTSSr:
4073   case X86::RSQRTSSr_Int:
4074   case X86::SQRTSSr:
4075   case X86::SQRTSSr_Int:
4076   // AVX encoded versions
4077   case X86::VCVTSD2SSrr:
4078   case X86::Int_VCVTSD2SSrr:
4079   case X86::VCVTSS2SDrr:
4080   case X86::Int_VCVTSS2SDrr:
4081   case X86::VCVTSD2SSZrr:
4082   case X86::VCVTSS2SDZrr:
4083   case X86::VRCPSSr:
4084   case X86::VROUNDSDr:
4085   case X86::VROUNDSDr_Int:
4086   case X86::VROUNDSSr:
4087   case X86::VROUNDSSr_Int:
4088   case X86::VRSQRTSSr:
4089   case X86::VSQRTSSr:
4090     return true;
4091   }
4092 
4093   return false;
4094 }
4095 
4096 /// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
4097 /// instructions we would like before a partial register update.
4098 unsigned X86InstrInfo::
4099 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
4100                              const TargetRegisterInfo *TRI) const {
4101   if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
4102     return 0;
4103 
4104   // If MI is marked as reading Reg, the partial register update is wanted.
4105   const MachineOperand &MO = MI->getOperand(0);
4106   unsigned Reg = MO.getReg();
4107   if (TargetRegisterInfo::isVirtualRegister(Reg)) {
4108     if (MO.readsReg() || MI->readsVirtualRegister(Reg))
4109       return 0;
4110   } else {
4111     if (MI->readsRegister(Reg, TRI))
4112       return 0;
4113   }
4114 
4115   // If any of the preceding 16 instructions are reading Reg, insert a
4116   // dependency breaking instruction.  The magic number is based on a few
4117   // Nehalem experiments.
4118   return 16;
4119 }
4120 
4121 void X86InstrInfo::
4122 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
4123                           const TargetRegisterInfo *TRI) const {
4124   unsigned Reg = MI->getOperand(OpNum).getReg();
4125   if (X86::VR128RegClass.contains(Reg)) {
4126     // These instructions are all floating point domain, so xorps is the best
4127     // choice.
4128     bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
4129     unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
4130     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
4131       .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
4132   } else if (X86::VR256RegClass.contains(Reg)) {
4133     // Use vxorps to clear the full ymm register.
4134     // It wants to read and write the xmm sub-register.
4135     unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
4136     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
4137       .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
4138       .addReg(Reg, RegState::ImplicitDefine);
4139   } else
4140     return;
4141   MI->addRegisterKilled(Reg, TRI, true);
4142 }
4143 
4144 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4145                                                   MachineInstr *MI,
4146                                            const SmallVectorImpl<unsigned> &Ops,
4147                                                   int FrameIndex) const {
4148   // Check switch flag
4149   if (NoFusing) return NULL;
4150 
4151   // Unless optimizing for size, don't fold to avoid partial
4152   // register update stalls
4153   if (!MF.getFunction()->getAttributes().
4154         hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4155       hasPartialRegUpdate(MI->getOpcode()))
4156     return 0;
4157 
4158   const MachineFrameInfo *MFI = MF.getFrameInfo();
4159   unsigned Size = MFI->getObjectSize(FrameIndex);
4160   unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
4161   // If the function stack isn't realigned we don't want to fold instructions
4162   // that need increased alignment.
4163   if (!RI.needsStackRealignment(MF))
4164     Alignment = std::min(Alignment, TM.getFrameLowering()->getStackAlignment());
4165   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4166     unsigned NewOpc = 0;
4167     unsigned RCSize = 0;
4168     switch (MI->getOpcode()) {
4169     default: return NULL;
4170     case X86::TEST8rr:  NewOpc = X86::CMP8ri; RCSize = 1; break;
4171     case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
4172     case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
4173     case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
4174     }
4175     // Check if it's safe to fold the load. If the size of the object is
4176     // narrower than the load width, then it's not.
4177     if (Size < RCSize)
4178       return NULL;
4179     // Change to CMPXXri r, 0 first.
4180     MI->setDesc(get(NewOpc));
4181     MI->getOperand(1).ChangeToImmediate(0);
4182   } else if (Ops.size() != 1)
4183     return NULL;
4184 
4185   SmallVector<MachineOperand,4> MOs;
4186   MOs.push_back(MachineOperand::CreateFI(FrameIndex));
4187   return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
4188 }
4189 
4190 MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
4191                                                   MachineInstr *MI,
4192                                            const SmallVectorImpl<unsigned> &Ops,
4193                                                   MachineInstr *LoadMI) const {
4194   // Check switch flag
4195   if (NoFusing) return NULL;
4196 
4197   // Unless optimizing for size, don't fold to avoid partial
4198   // register update stalls
4199   if (!MF.getFunction()->getAttributes().
4200         hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize) &&
4201       hasPartialRegUpdate(MI->getOpcode()))
4202     return 0;
4203 
4204   // Determine the alignment of the load.
4205   unsigned Alignment = 0;
4206   if (LoadMI->hasOneMemOperand())
4207     Alignment = (*LoadMI->memoperands_begin())->getAlignment();
4208   else
4209     switch (LoadMI->getOpcode()) {
4210     case X86::AVX2_SETALLONES:
4211     case X86::AVX_SET0:
4212       Alignment = 32;
4213       break;
4214     case X86::V_SET0:
4215     case X86::V_SETALLONES:
4216       Alignment = 16;
4217       break;
4218     case X86::FsFLD0SD:
4219       Alignment = 8;
4220       break;
4221     case X86::FsFLD0SS:
4222       Alignment = 4;
4223       break;
4224     default:
4225       return 0;
4226     }
4227   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4228     unsigned NewOpc = 0;
4229     switch (MI->getOpcode()) {
4230     default: return NULL;
4231     case X86::TEST8rr:  NewOpc = X86::CMP8ri; break;
4232     case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
4233     case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
4234     case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
4235     }
4236     // Change to CMPXXri r, 0 first.
4237     MI->setDesc(get(NewOpc));
4238     MI->getOperand(1).ChangeToImmediate(0);
4239   } else if (Ops.size() != 1)
4240     return NULL;
4241 
4242   // Make sure the subregisters match.
4243   // Otherwise we risk changing the size of the load.
4244   if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
4245     return NULL;
4246 
4247   SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
4248   switch (LoadMI->getOpcode()) {
4249   case X86::V_SET0:
4250   case X86::V_SETALLONES:
4251   case X86::AVX2_SETALLONES:
4252   case X86::AVX_SET0:
4253   case X86::FsFLD0SD:
4254   case X86::FsFLD0SS: {
4255     // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
4256     // Create a constant-pool entry and operands to load from it.
4257 
4258     // Medium and large mode can't fold loads this way.
4259     if (TM.getCodeModel() != CodeModel::Small &&
4260         TM.getCodeModel() != CodeModel::Kernel)
4261       return NULL;
4262 
4263     // x86-32 PIC requires a PIC base register for constant pools.
4264     unsigned PICBase = 0;
4265     if (TM.getRelocationModel() == Reloc::PIC_) {
4266       if (TM.getSubtarget<X86Subtarget>().is64Bit())
4267         PICBase = X86::RIP;
4268       else
4269         // FIXME: PICBase = getGlobalBaseReg(&MF);
4270         // This doesn't work for several reasons.
4271         // 1. GlobalBaseReg may have been spilled.
4272         // 2. It may not be live at MI.
4273         return NULL;
4274     }
4275 
4276     // Create a constant-pool entry.
4277     MachineConstantPool &MCP = *MF.getConstantPool();
4278     Type *Ty;
4279     unsigned Opc = LoadMI->getOpcode();
4280     if (Opc == X86::FsFLD0SS)
4281       Ty = Type::getFloatTy(MF.getFunction()->getContext());
4282     else if (Opc == X86::FsFLD0SD)
4283       Ty = Type::getDoubleTy(MF.getFunction()->getContext());
4284     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0)
4285       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
4286     else
4287       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
4288 
4289     bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES);
4290     const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
4291                                     Constant::getNullValue(Ty);
4292     unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
4293 
4294     // Create operands to load from the constant pool entry.
4295     MOs.push_back(MachineOperand::CreateReg(PICBase, false));
4296     MOs.push_back(MachineOperand::CreateImm(1));
4297     MOs.push_back(MachineOperand::CreateReg(0, false));
4298     MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
4299     MOs.push_back(MachineOperand::CreateReg(0, false));
4300     break;
4301   }
4302   default: {
4303     if ((LoadMI->getOpcode() == X86::MOVSSrm ||
4304          LoadMI->getOpcode() == X86::VMOVSSrm) &&
4305         MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4306           > 4)
4307       // These instructions only load 32 bits, we can't fold them if the
4308       // destination register is wider than 32 bits (4 bytes).
4309       return NULL;
4310     if ((LoadMI->getOpcode() == X86::MOVSDrm ||
4311          LoadMI->getOpcode() == X86::VMOVSDrm) &&
4312         MF.getRegInfo().getRegClass(LoadMI->getOperand(0).getReg())->getSize()
4313           > 8)
4314       // These instructions only load 64 bits, we can't fold them if the
4315       // destination register is wider than 64 bits (8 bytes).
4316       return NULL;
4317 
4318     // Folding a normal load. Just copy the load's address operands.
4319     unsigned NumOps = LoadMI->getDesc().getNumOperands();
4320     for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
4321       MOs.push_back(LoadMI->getOperand(i));
4322     break;
4323   }
4324   }
4325   return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
4326 }
4327 
4328 
4329 bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
4330                                   const SmallVectorImpl<unsigned> &Ops) const {
4331   // Check switch flag
4332   if (NoFusing) return 0;
4333 
4334   if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
4335     switch (MI->getOpcode()) {
4336     default: return false;
4337     case X86::TEST8rr:
4338     case X86::TEST16rr:
4339     case X86::TEST32rr:
4340     case X86::TEST64rr:
4341       return true;
4342     case X86::ADD32ri:
4343       // FIXME: AsmPrinter doesn't know how to handle
4344       // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
4345       if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
4346         return false;
4347       break;
4348     }
4349   }
4350 
4351   if (Ops.size() != 1)
4352     return false;
4353 
4354   unsigned OpNum = Ops[0];
4355   unsigned Opc = MI->getOpcode();
4356   unsigned NumOps = MI->getDesc().getNumOperands();
4357   bool isTwoAddr = NumOps > 1 &&
4358     MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
4359 
4360   // Folding a memory location into the two-address part of a two-address
4361   // instruction is different than folding it other places.  It requires
4362   // replacing the *two* registers with the memory location.
4363   const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
4364   if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
4365     OpcodeTablePtr = &RegOp2MemOpTable2Addr;
4366   } else if (OpNum == 0) { // If operand 0
4367     if (Opc == X86::MOV32r0)
4368       return true;
4369 
4370     OpcodeTablePtr = &RegOp2MemOpTable0;
4371   } else if (OpNum == 1) {
4372     OpcodeTablePtr = &RegOp2MemOpTable1;
4373   } else if (OpNum == 2) {
4374     OpcodeTablePtr = &RegOp2MemOpTable2;
4375   } else if (OpNum == 3) {
4376     OpcodeTablePtr = &RegOp2MemOpTable3;
4377   }
4378 
4379   if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
4380     return true;
4381   return TargetInstrInfo::canFoldMemoryOperand(MI, Ops);
4382 }
4383 
4384 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
4385                                 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
4386                                 SmallVectorImpl<MachineInstr*> &NewMIs) const {
4387   DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4388     MemOp2RegOpTable.find(MI->getOpcode());
4389   if (I == MemOp2RegOpTable.end())
4390     return false;
4391   unsigned Opc = I->second.first;
4392   unsigned Index = I->second.second & TB_INDEX_MASK;
4393   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4394   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4395   if (UnfoldLoad && !FoldedLoad)
4396     return false;
4397   UnfoldLoad &= FoldedLoad;
4398   if (UnfoldStore && !FoldedStore)
4399     return false;
4400   UnfoldStore &= FoldedStore;
4401 
4402   const MCInstrDesc &MCID = get(Opc);
4403   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4404   if (!MI->hasOneMemOperand() &&
4405       RC == &X86::VR128RegClass &&
4406       !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4407     // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
4408     // conservatively assume the address is unaligned. That's bad for
4409     // performance.
4410     return false;
4411   SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
4412   SmallVector<MachineOperand,2> BeforeOps;
4413   SmallVector<MachineOperand,2> AfterOps;
4414   SmallVector<MachineOperand,4> ImpOps;
4415   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
4416     MachineOperand &Op = MI->getOperand(i);
4417     if (i >= Index && i < Index + X86::AddrNumOperands)
4418       AddrOps.push_back(Op);
4419     else if (Op.isReg() && Op.isImplicit())
4420       ImpOps.push_back(Op);
4421     else if (i < Index)
4422       BeforeOps.push_back(Op);
4423     else if (i > Index)
4424       AfterOps.push_back(Op);
4425   }
4426 
4427   // Emit the load instruction.
4428   if (UnfoldLoad) {
4429     std::pair<MachineInstr::mmo_iterator,
4430               MachineInstr::mmo_iterator> MMOs =
4431       MF.extractLoadMemRefs(MI->memoperands_begin(),
4432                             MI->memoperands_end());
4433     loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
4434     if (UnfoldStore) {
4435       // Address operands cannot be marked isKill.
4436       for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
4437         MachineOperand &MO = NewMIs[0]->getOperand(i);
4438         if (MO.isReg())
4439           MO.setIsKill(false);
4440       }
4441     }
4442   }
4443 
4444   // Emit the data processing instruction.
4445   MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
4446   MachineInstrBuilder MIB(MF, DataMI);
4447 
4448   if (FoldedStore)
4449     MIB.addReg(Reg, RegState::Define);
4450   for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
4451     MIB.addOperand(BeforeOps[i]);
4452   if (FoldedLoad)
4453     MIB.addReg(Reg);
4454   for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
4455     MIB.addOperand(AfterOps[i]);
4456   for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
4457     MachineOperand &MO = ImpOps[i];
4458     MIB.addReg(MO.getReg(),
4459                getDefRegState(MO.isDef()) |
4460                RegState::Implicit |
4461                getKillRegState(MO.isKill()) |
4462                getDeadRegState(MO.isDead()) |
4463                getUndefRegState(MO.isUndef()));
4464   }
4465   // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
4466   switch (DataMI->getOpcode()) {
4467   default: break;
4468   case X86::CMP64ri32:
4469   case X86::CMP64ri8:
4470   case X86::CMP32ri:
4471   case X86::CMP32ri8:
4472   case X86::CMP16ri:
4473   case X86::CMP16ri8:
4474   case X86::CMP8ri: {
4475     MachineOperand &MO0 = DataMI->getOperand(0);
4476     MachineOperand &MO1 = DataMI->getOperand(1);
4477     if (MO1.getImm() == 0) {
4478       unsigned NewOpc;
4479       switch (DataMI->getOpcode()) {
4480       default: llvm_unreachable("Unreachable!");
4481       case X86::CMP64ri8:
4482       case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
4483       case X86::CMP32ri8:
4484       case X86::CMP32ri:   NewOpc = X86::TEST32rr; break;
4485       case X86::CMP16ri8:
4486       case X86::CMP16ri:   NewOpc = X86::TEST16rr; break;
4487       case X86::CMP8ri:    NewOpc = X86::TEST8rr; break;
4488       }
4489       DataMI->setDesc(get(NewOpc));
4490       MO1.ChangeToRegister(MO0.getReg(), false);
4491     }
4492   }
4493   }
4494   NewMIs.push_back(DataMI);
4495 
4496   // Emit the store instruction.
4497   if (UnfoldStore) {
4498     const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
4499     std::pair<MachineInstr::mmo_iterator,
4500               MachineInstr::mmo_iterator> MMOs =
4501       MF.extractStoreMemRefs(MI->memoperands_begin(),
4502                              MI->memoperands_end());
4503     storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
4504   }
4505 
4506   return true;
4507 }
4508 
4509 bool
4510 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
4511                                   SmallVectorImpl<SDNode*> &NewNodes) const {
4512   if (!N->isMachineOpcode())
4513     return false;
4514 
4515   DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4516     MemOp2RegOpTable.find(N->getMachineOpcode());
4517   if (I == MemOp2RegOpTable.end())
4518     return false;
4519   unsigned Opc = I->second.first;
4520   unsigned Index = I->second.second & TB_INDEX_MASK;
4521   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4522   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4523   const MCInstrDesc &MCID = get(Opc);
4524   MachineFunction &MF = DAG.getMachineFunction();
4525   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
4526   unsigned NumDefs = MCID.NumDefs;
4527   std::vector<SDValue> AddrOps;
4528   std::vector<SDValue> BeforeOps;
4529   std::vector<SDValue> AfterOps;
4530   SDLoc dl(N);
4531   unsigned NumOps = N->getNumOperands();
4532   for (unsigned i = 0; i != NumOps-1; ++i) {
4533     SDValue Op = N->getOperand(i);
4534     if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
4535       AddrOps.push_back(Op);
4536     else if (i < Index-NumDefs)
4537       BeforeOps.push_back(Op);
4538     else if (i > Index-NumDefs)
4539       AfterOps.push_back(Op);
4540   }
4541   SDValue Chain = N->getOperand(NumOps-1);
4542   AddrOps.push_back(Chain);
4543 
4544   // Emit the load instruction.
4545   SDNode *Load = 0;
4546   if (FoldedLoad) {
4547     EVT VT = *RC->vt_begin();
4548     std::pair<MachineInstr::mmo_iterator,
4549               MachineInstr::mmo_iterator> MMOs =
4550       MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4551                             cast<MachineSDNode>(N)->memoperands_end());
4552     if (!(*MMOs.first) &&
4553         RC == &X86::VR128RegClass &&
4554         !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4555       // Do not introduce a slow unaligned load.
4556       return false;
4557     unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4558     bool isAligned = (*MMOs.first) &&
4559                      (*MMOs.first)->getAlignment() >= Alignment;
4560     Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
4561                               VT, MVT::Other, AddrOps);
4562     NewNodes.push_back(Load);
4563 
4564     // Preserve memory reference information.
4565     cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4566   }
4567 
4568   // Emit the data processing instruction.
4569   std::vector<EVT> VTs;
4570   const TargetRegisterClass *DstRC = 0;
4571   if (MCID.getNumDefs() > 0) {
4572     DstRC = getRegClass(MCID, 0, &RI, MF);
4573     VTs.push_back(*DstRC->vt_begin());
4574   }
4575   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
4576     EVT VT = N->getValueType(i);
4577     if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
4578       VTs.push_back(VT);
4579   }
4580   if (Load)
4581     BeforeOps.push_back(SDValue(Load, 0));
4582   std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
4583   SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps);
4584   NewNodes.push_back(NewNode);
4585 
4586   // Emit the store instruction.
4587   if (FoldedStore) {
4588     AddrOps.pop_back();
4589     AddrOps.push_back(SDValue(NewNode, 0));
4590     AddrOps.push_back(Chain);
4591     std::pair<MachineInstr::mmo_iterator,
4592               MachineInstr::mmo_iterator> MMOs =
4593       MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
4594                              cast<MachineSDNode>(N)->memoperands_end());
4595     if (!(*MMOs.first) &&
4596         RC == &X86::VR128RegClass &&
4597         !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
4598       // Do not introduce a slow unaligned store.
4599       return false;
4600     unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
4601     bool isAligned = (*MMOs.first) &&
4602                      (*MMOs.first)->getAlignment() >= Alignment;
4603     SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
4604                                                          isAligned, TM),
4605                                        dl, MVT::Other, AddrOps);
4606     NewNodes.push_back(Store);
4607 
4608     // Preserve memory reference information.
4609     cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
4610   }
4611 
4612   return true;
4613 }
4614 
4615 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
4616                                       bool UnfoldLoad, bool UnfoldStore,
4617                                       unsigned *LoadRegIndex) const {
4618   DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
4619     MemOp2RegOpTable.find(Opc);
4620   if (I == MemOp2RegOpTable.end())
4621     return 0;
4622   bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
4623   bool FoldedStore = I->second.second & TB_FOLDED_STORE;
4624   if (UnfoldLoad && !FoldedLoad)
4625     return 0;
4626   if (UnfoldStore && !FoldedStore)
4627     return 0;
4628   if (LoadRegIndex)
4629     *LoadRegIndex = I->second.second & TB_INDEX_MASK;
4630   return I->second.first;
4631 }
4632 
4633 bool
4634 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
4635                                      int64_t &Offset1, int64_t &Offset2) const {
4636   if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
4637     return false;
4638   unsigned Opc1 = Load1->getMachineOpcode();
4639   unsigned Opc2 = Load2->getMachineOpcode();
4640   switch (Opc1) {
4641   default: return false;
4642   case X86::MOV8rm:
4643   case X86::MOV16rm:
4644   case X86::MOV32rm:
4645   case X86::MOV64rm:
4646   case X86::LD_Fp32m:
4647   case X86::LD_Fp64m:
4648   case X86::LD_Fp80m:
4649   case X86::MOVSSrm:
4650   case X86::MOVSDrm:
4651   case X86::MMX_MOVD64rm:
4652   case X86::MMX_MOVQ64rm:
4653   case X86::FsMOVAPSrm:
4654   case X86::FsMOVAPDrm:
4655   case X86::MOVAPSrm:
4656   case X86::MOVUPSrm:
4657   case X86::MOVAPDrm:
4658   case X86::MOVDQArm:
4659   case X86::MOVDQUrm:
4660   // AVX load instructions
4661   case X86::VMOVSSrm:
4662   case X86::VMOVSDrm:
4663   case X86::FsVMOVAPSrm:
4664   case X86::FsVMOVAPDrm:
4665   case X86::VMOVAPSrm:
4666   case X86::VMOVUPSrm:
4667   case X86::VMOVAPDrm:
4668   case X86::VMOVDQArm:
4669   case X86::VMOVDQUrm:
4670   case X86::VMOVAPSYrm:
4671   case X86::VMOVUPSYrm:
4672   case X86::VMOVAPDYrm:
4673   case X86::VMOVDQAYrm:
4674   case X86::VMOVDQUYrm:
4675     break;
4676   }
4677   switch (Opc2) {
4678   default: return false;
4679   case X86::MOV8rm:
4680   case X86::MOV16rm:
4681   case X86::MOV32rm:
4682   case X86::MOV64rm:
4683   case X86::LD_Fp32m:
4684   case X86::LD_Fp64m:
4685   case X86::LD_Fp80m:
4686   case X86::MOVSSrm:
4687   case X86::MOVSDrm:
4688   case X86::MMX_MOVD64rm:
4689   case X86::MMX_MOVQ64rm:
4690   case X86::FsMOVAPSrm:
4691   case X86::FsMOVAPDrm:
4692   case X86::MOVAPSrm:
4693   case X86::MOVUPSrm:
4694   case X86::MOVAPDrm:
4695   case X86::MOVDQArm:
4696   case X86::MOVDQUrm:
4697   // AVX load instructions
4698   case X86::VMOVSSrm:
4699   case X86::VMOVSDrm:
4700   case X86::FsVMOVAPSrm:
4701   case X86::FsVMOVAPDrm:
4702   case X86::VMOVAPSrm:
4703   case X86::VMOVUPSrm:
4704   case X86::VMOVAPDrm:
4705   case X86::VMOVDQArm:
4706   case X86::VMOVDQUrm:
4707   case X86::VMOVAPSYrm:
4708   case X86::VMOVUPSYrm:
4709   case X86::VMOVAPDYrm:
4710   case X86::VMOVDQAYrm:
4711   case X86::VMOVDQUYrm:
4712     break;
4713   }
4714 
4715   // Check if chain operands and base addresses match.
4716   if (Load1->getOperand(0) != Load2->getOperand(0) ||
4717       Load1->getOperand(5) != Load2->getOperand(5))
4718     return false;
4719   // Segment operands should match as well.
4720   if (Load1->getOperand(4) != Load2->getOperand(4))
4721     return false;
4722   // Scale should be 1, Index should be Reg0.
4723   if (Load1->getOperand(1) == Load2->getOperand(1) &&
4724       Load1->getOperand(2) == Load2->getOperand(2)) {
4725     if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
4726       return false;
4727 
4728     // Now let's examine the displacements.
4729     if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
4730         isa<ConstantSDNode>(Load2->getOperand(3))) {
4731       Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
4732       Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
4733       return true;
4734     }
4735   }
4736   return false;
4737 }
4738 
4739 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
4740                                            int64_t Offset1, int64_t Offset2,
4741                                            unsigned NumLoads) const {
4742   assert(Offset2 > Offset1);
4743   if ((Offset2 - Offset1) / 8 > 64)
4744     return false;
4745 
4746   unsigned Opc1 = Load1->getMachineOpcode();
4747   unsigned Opc2 = Load2->getMachineOpcode();
4748   if (Opc1 != Opc2)
4749     return false;  // FIXME: overly conservative?
4750 
4751   switch (Opc1) {
4752   default: break;
4753   case X86::LD_Fp32m:
4754   case X86::LD_Fp64m:
4755   case X86::LD_Fp80m:
4756   case X86::MMX_MOVD64rm:
4757   case X86::MMX_MOVQ64rm:
4758     return false;
4759   }
4760 
4761   EVT VT = Load1->getValueType(0);
4762   switch (VT.getSimpleVT().SimpleTy) {
4763   default:
4764     // XMM registers. In 64-bit mode we can be a bit more aggressive since we
4765     // have 16 of them to play with.
4766     if (TM.getSubtargetImpl()->is64Bit()) {
4767       if (NumLoads >= 3)
4768         return false;
4769     } else if (NumLoads) {
4770       return false;
4771     }
4772     break;
4773   case MVT::i8:
4774   case MVT::i16:
4775   case MVT::i32:
4776   case MVT::i64:
4777   case MVT::f32:
4778   case MVT::f64:
4779     if (NumLoads)
4780       return false;
4781     break;
4782   }
4783 
4784   return true;
4785 }
4786 
4787 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First,
4788                                           MachineInstr *Second) const {
4789   // Check if this processor supports macro-fusion. Since this is a minor
4790   // heuristic, we haven't specifically reserved a feature. hasAVX is a decent
4791   // proxy for SandyBridge+.
4792   if (!TM.getSubtarget<X86Subtarget>().hasAVX())
4793     return false;
4794 
4795   enum {
4796     FuseTest,
4797     FuseCmp,
4798     FuseInc
4799   } FuseKind;
4800 
4801   switch(Second->getOpcode()) {
4802   default:
4803     return false;
4804   case X86::JE_4:
4805   case X86::JNE_4:
4806   case X86::JL_4:
4807   case X86::JLE_4:
4808   case X86::JG_4:
4809   case X86::JGE_4:
4810     FuseKind = FuseInc;
4811     break;
4812   case X86::JB_4:
4813   case X86::JBE_4:
4814   case X86::JA_4:
4815   case X86::JAE_4:
4816     FuseKind = FuseCmp;
4817     break;
4818   case X86::JS_4:
4819   case X86::JNS_4:
4820   case X86::JP_4:
4821   case X86::JNP_4:
4822   case X86::JO_4:
4823   case X86::JNO_4:
4824     FuseKind = FuseTest;
4825     break;
4826   }
4827   switch (First->getOpcode()) {
4828   default:
4829     return false;
4830   case X86::TEST8rr:
4831   case X86::TEST16rr:
4832   case X86::TEST32rr:
4833   case X86::TEST64rr:
4834   case X86::TEST8ri:
4835   case X86::TEST16ri:
4836   case X86::TEST32ri:
4837   case X86::TEST32i32:
4838   case X86::TEST64i32:
4839   case X86::TEST64ri32:
4840   case X86::TEST8rm:
4841   case X86::TEST16rm:
4842   case X86::TEST32rm:
4843   case X86::TEST64rm:
4844   case X86::AND16i16:
4845   case X86::AND16ri:
4846   case X86::AND16ri8:
4847   case X86::AND16rm:
4848   case X86::AND16rr:
4849   case X86::AND32i32:
4850   case X86::AND32ri:
4851   case X86::AND32ri8:
4852   case X86::AND32rm:
4853   case X86::AND32rr:
4854   case X86::AND64i32:
4855   case X86::AND64ri32:
4856   case X86::AND64ri8:
4857   case X86::AND64rm:
4858   case X86::AND64rr:
4859   case X86::AND8i8:
4860   case X86::AND8ri:
4861   case X86::AND8rm:
4862   case X86::AND8rr:
4863     return true;
4864   case X86::CMP16i16:
4865   case X86::CMP16ri:
4866   case X86::CMP16ri8:
4867   case X86::CMP16rm:
4868   case X86::CMP16rr:
4869   case X86::CMP32i32:
4870   case X86::CMP32ri:
4871   case X86::CMP32ri8:
4872   case X86::CMP32rm:
4873   case X86::CMP32rr:
4874   case X86::CMP64i32:
4875   case X86::CMP64ri32:
4876   case X86::CMP64ri8:
4877   case X86::CMP64rm:
4878   case X86::CMP64rr:
4879   case X86::CMP8i8:
4880   case X86::CMP8ri:
4881   case X86::CMP8rm:
4882   case X86::CMP8rr:
4883   case X86::ADD16i16:
4884   case X86::ADD16ri:
4885   case X86::ADD16ri8:
4886   case X86::ADD16ri8_DB:
4887   case X86::ADD16ri_DB:
4888   case X86::ADD16rm:
4889   case X86::ADD16rr:
4890   case X86::ADD16rr_DB:
4891   case X86::ADD32i32:
4892   case X86::ADD32ri:
4893   case X86::ADD32ri8:
4894   case X86::ADD32ri8_DB:
4895   case X86::ADD32ri_DB:
4896   case X86::ADD32rm:
4897   case X86::ADD32rr:
4898   case X86::ADD32rr_DB:
4899   case X86::ADD64i32:
4900   case X86::ADD64ri32:
4901   case X86::ADD64ri32_DB:
4902   case X86::ADD64ri8:
4903   case X86::ADD64ri8_DB:
4904   case X86::ADD64rm:
4905   case X86::ADD64rr:
4906   case X86::ADD64rr_DB:
4907   case X86::ADD8i8:
4908   case X86::ADD8mi:
4909   case X86::ADD8mr:
4910   case X86::ADD8ri:
4911   case X86::ADD8rm:
4912   case X86::ADD8rr:
4913   case X86::SUB16i16:
4914   case X86::SUB16ri:
4915   case X86::SUB16ri8:
4916   case X86::SUB16rm:
4917   case X86::SUB16rr:
4918   case X86::SUB32i32:
4919   case X86::SUB32ri:
4920   case X86::SUB32ri8:
4921   case X86::SUB32rm:
4922   case X86::SUB32rr:
4923   case X86::SUB64i32:
4924   case X86::SUB64ri32:
4925   case X86::SUB64ri8:
4926   case X86::SUB64rm:
4927   case X86::SUB64rr:
4928   case X86::SUB8i8:
4929   case X86::SUB8ri:
4930   case X86::SUB8rm:
4931   case X86::SUB8rr:
4932     return FuseKind == FuseCmp || FuseKind == FuseInc;
4933   case X86::INC16r:
4934   case X86::INC32r:
4935   case X86::INC64_16r:
4936   case X86::INC64_32r:
4937   case X86::INC64r:
4938   case X86::INC8r:
4939   case X86::DEC16r:
4940   case X86::DEC32r:
4941   case X86::DEC64_16r:
4942   case X86::DEC64_32r:
4943   case X86::DEC64r:
4944   case X86::DEC8r:
4945     return FuseKind == FuseInc;
4946   }
4947 }
4948 
4949 bool X86InstrInfo::
4950 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
4951   assert(Cond.size() == 1 && "Invalid X86 branch condition!");
4952   X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
4953   if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
4954     return true;
4955   Cond[0].setImm(GetOppositeBranchCondition(CC));
4956   return false;
4957 }
4958 
4959 bool X86InstrInfo::
4960 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
4961   // FIXME: Return false for x87 stack register classes for now. We can't
4962   // allow any loads of these registers before FpGet_ST0_80.
4963   return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
4964            RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
4965 }
4966 
4967 /// getGlobalBaseReg - Return a virtual register initialized with the
4968 /// the global base register value. Output instructions required to
4969 /// initialize the register in the function entry block, if necessary.
4970 ///
4971 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
4972 ///
4973 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
4974   assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
4975          "X86-64 PIC uses RIP relative addressing");
4976 
4977   X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
4978   unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
4979   if (GlobalBaseReg != 0)
4980     return GlobalBaseReg;
4981 
4982   // Create the register. The code to initialize it is inserted
4983   // later, by the CGBR pass (below).
4984   MachineRegisterInfo &RegInfo = MF->getRegInfo();
4985   GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
4986   X86FI->setGlobalBaseReg(GlobalBaseReg);
4987   return GlobalBaseReg;
4988 }
4989 
4990 // These are the replaceable SSE instructions. Some of these have Int variants
4991 // that we don't include here. We don't want to replace instructions selected
4992 // by intrinsics.
4993 static const uint16_t ReplaceableInstrs[][3] = {
4994   //PackedSingle     PackedDouble    PackedInt
4995   { X86::MOVAPSmr,   X86::MOVAPDmr,  X86::MOVDQAmr  },
4996   { X86::MOVAPSrm,   X86::MOVAPDrm,  X86::MOVDQArm  },
4997   { X86::MOVAPSrr,   X86::MOVAPDrr,  X86::MOVDQArr  },
4998   { X86::MOVUPSmr,   X86::MOVUPDmr,  X86::MOVDQUmr  },
4999   { X86::MOVUPSrm,   X86::MOVUPDrm,  X86::MOVDQUrm  },
5000   { X86::MOVNTPSmr,  X86::MOVNTPDmr, X86::MOVNTDQmr },
5001   { X86::ANDNPSrm,   X86::ANDNPDrm,  X86::PANDNrm   },
5002   { X86::ANDNPSrr,   X86::ANDNPDrr,  X86::PANDNrr   },
5003   { X86::ANDPSrm,    X86::ANDPDrm,   X86::PANDrm    },
5004   { X86::ANDPSrr,    X86::ANDPDrr,   X86::PANDrr    },
5005   { X86::ORPSrm,     X86::ORPDrm,    X86::PORrm     },
5006   { X86::ORPSrr,     X86::ORPDrr,    X86::PORrr     },
5007   { X86::XORPSrm,    X86::XORPDrm,   X86::PXORrm    },
5008   { X86::XORPSrr,    X86::XORPDrr,   X86::PXORrr    },
5009   // AVX 128-bit support
5010   { X86::VMOVAPSmr,  X86::VMOVAPDmr,  X86::VMOVDQAmr  },
5011   { X86::VMOVAPSrm,  X86::VMOVAPDrm,  X86::VMOVDQArm  },
5012   { X86::VMOVAPSrr,  X86::VMOVAPDrr,  X86::VMOVDQArr  },
5013   { X86::VMOVUPSmr,  X86::VMOVUPDmr,  X86::VMOVDQUmr  },
5014   { X86::VMOVUPSrm,  X86::VMOVUPDrm,  X86::VMOVDQUrm  },
5015   { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
5016   { X86::VANDNPSrm,  X86::VANDNPDrm,  X86::VPANDNrm   },
5017   { X86::VANDNPSrr,  X86::VANDNPDrr,  X86::VPANDNrr   },
5018   { X86::VANDPSrm,   X86::VANDPDrm,   X86::VPANDrm    },
5019   { X86::VANDPSrr,   X86::VANDPDrr,   X86::VPANDrr    },
5020   { X86::VORPSrm,    X86::VORPDrm,    X86::VPORrm     },
5021   { X86::VORPSrr,    X86::VORPDrr,    X86::VPORrr     },
5022   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORrm    },
5023   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORrr    },
5024   // AVX 256-bit support
5025   { X86::VMOVAPSYmr,   X86::VMOVAPDYmr,   X86::VMOVDQAYmr  },
5026   { X86::VMOVAPSYrm,   X86::VMOVAPDYrm,   X86::VMOVDQAYrm  },
5027   { X86::VMOVAPSYrr,   X86::VMOVAPDYrr,   X86::VMOVDQAYrr  },
5028   { X86::VMOVUPSYmr,   X86::VMOVUPDYmr,   X86::VMOVDQUYmr  },
5029   { X86::VMOVUPSYrm,   X86::VMOVUPDYrm,   X86::VMOVDQUYrm  },
5030   { X86::VMOVNTPSYmr,  X86::VMOVNTPDYmr,  X86::VMOVNTDQYmr }
5031 };
5032 
5033 static const uint16_t ReplaceableInstrsAVX2[][3] = {
5034   //PackedSingle       PackedDouble       PackedInt
5035   { X86::VANDNPSYrm,   X86::VANDNPDYrm,   X86::VPANDNYrm   },
5036   { X86::VANDNPSYrr,   X86::VANDNPDYrr,   X86::VPANDNYrr   },
5037   { X86::VANDPSYrm,    X86::VANDPDYrm,    X86::VPANDYrm    },
5038   { X86::VANDPSYrr,    X86::VANDPDYrr,    X86::VPANDYrr    },
5039   { X86::VORPSYrm,     X86::VORPDYrm,     X86::VPORYrm     },
5040   { X86::VORPSYrr,     X86::VORPDYrr,     X86::VPORYrr     },
5041   { X86::VXORPSYrm,    X86::VXORPDYrm,    X86::VPXORYrm    },
5042   { X86::VXORPSYrr,    X86::VXORPDYrr,    X86::VPXORYrr    },
5043   { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
5044   { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
5045   { X86::VINSERTF128rm,  X86::VINSERTF128rm,  X86::VINSERTI128rm },
5046   { X86::VINSERTF128rr,  X86::VINSERTF128rr,  X86::VINSERTI128rr },
5047   { X86::VPERM2F128rm,   X86::VPERM2F128rm,   X86::VPERM2I128rm },
5048   { X86::VPERM2F128rr,   X86::VPERM2F128rr,   X86::VPERM2I128rr }
5049 };
5050 
5051 // FIXME: Some shuffle and unpack instructions have equivalents in different
5052 // domains, but they require a bit more work than just switching opcodes.
5053 
5054 static const uint16_t *lookup(unsigned opcode, unsigned domain) {
5055   for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
5056     if (ReplaceableInstrs[i][domain-1] == opcode)
5057       return ReplaceableInstrs[i];
5058   return 0;
5059 }
5060 
5061 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
5062   for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
5063     if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
5064       return ReplaceableInstrsAVX2[i];
5065   return 0;
5066 }
5067 
5068 std::pair<uint16_t, uint16_t>
5069 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
5070   uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5071   bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
5072   uint16_t validDomains = 0;
5073   if (domain && lookup(MI->getOpcode(), domain))
5074     validDomains = 0xe;
5075   else if (domain && lookupAVX2(MI->getOpcode(), domain))
5076     validDomains = hasAVX2 ? 0xe : 0x6;
5077   return std::make_pair(domain, validDomains);
5078 }
5079 
5080 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
5081   assert(Domain>0 && Domain<4 && "Invalid execution domain");
5082   uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
5083   assert(dom && "Not an SSE instruction");
5084   const uint16_t *table = lookup(MI->getOpcode(), dom);
5085   if (!table) { // try the other table
5086     assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
5087            "256-bit vector operations only available in AVX2");
5088     table = lookupAVX2(MI->getOpcode(), dom);
5089   }
5090   assert(table && "Cannot change domain");
5091   MI->setDesc(get(table[Domain-1]));
5092 }
5093 
5094 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
5095 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
5096   NopInst.setOpcode(X86::NOOP);
5097 }
5098 
5099 bool X86InstrInfo::isHighLatencyDef(int opc) const {
5100   switch (opc) {
5101   default: return false;
5102   case X86::DIVSDrm:
5103   case X86::DIVSDrm_Int:
5104   case X86::DIVSDrr:
5105   case X86::DIVSDrr_Int:
5106   case X86::DIVSSrm:
5107   case X86::DIVSSrm_Int:
5108   case X86::DIVSSrr:
5109   case X86::DIVSSrr_Int:
5110   case X86::SQRTPDm:
5111   case X86::SQRTPDr:
5112   case X86::SQRTPSm:
5113   case X86::SQRTPSr:
5114   case X86::SQRTSDm:
5115   case X86::SQRTSDm_Int:
5116   case X86::SQRTSDr:
5117   case X86::SQRTSDr_Int:
5118   case X86::SQRTSSm:
5119   case X86::SQRTSSm_Int:
5120   case X86::SQRTSSr:
5121   case X86::SQRTSSr_Int:
5122   // AVX instructions with high latency
5123   case X86::VDIVSDrm:
5124   case X86::VDIVSDrm_Int:
5125   case X86::VDIVSDrr:
5126   case X86::VDIVSDrr_Int:
5127   case X86::VDIVSSrm:
5128   case X86::VDIVSSrm_Int:
5129   case X86::VDIVSSrr:
5130   case X86::VDIVSSrr_Int:
5131   case X86::VSQRTPDm:
5132   case X86::VSQRTPDr:
5133   case X86::VSQRTPSm:
5134   case X86::VSQRTPSr:
5135   case X86::VSQRTSDm:
5136   case X86::VSQRTSDm_Int:
5137   case X86::VSQRTSDr:
5138   case X86::VSQRTSSm:
5139   case X86::VSQRTSSm_Int:
5140   case X86::VSQRTSSr:
5141   case X86::VSQRTPDZrm:
5142   case X86::VSQRTPDZrr:
5143   case X86::VSQRTPSZrm:
5144   case X86::VSQRTPSZrr:
5145   case X86::VSQRTSDZm:
5146   case X86::VSQRTSDZm_Int:
5147   case X86::VSQRTSDZr:
5148   case X86::VSQRTSSZm_Int:
5149   case X86::VSQRTSSZr:
5150   case X86::VSQRTSSZm:
5151   case X86::VDIVSDZrm:
5152   case X86::VDIVSDZrr:
5153   case X86::VDIVSSZrm:
5154   case X86::VDIVSSZrr:
5155 
5156   case X86::VGATHERQPSZrm:
5157   case X86::VGATHERQPDZrm:
5158   case X86::VGATHERDPDZrm:
5159   case X86::VGATHERDPSZrm:
5160   case X86::VPGATHERQDZrm:
5161   case X86::VPGATHERQQZrm:
5162   case X86::VPGATHERDDZrm:
5163   case X86::VPGATHERDQZrm:
5164   case X86::VSCATTERQPDZmr:
5165   case X86::VSCATTERQPSZmr:
5166   case X86::VSCATTERDPDZmr:
5167   case X86::VSCATTERDPSZmr:
5168   case X86::VPSCATTERQDZmr:
5169   case X86::VPSCATTERQQZmr:
5170   case X86::VPSCATTERDDZmr:
5171   case X86::VPSCATTERDQZmr:
5172     return true;
5173   }
5174 }
5175 
5176 bool X86InstrInfo::
5177 hasHighOperandLatency(const InstrItineraryData *ItinData,
5178                       const MachineRegisterInfo *MRI,
5179                       const MachineInstr *DefMI, unsigned DefIdx,
5180                       const MachineInstr *UseMI, unsigned UseIdx) const {
5181   return isHighLatencyDef(DefMI->getOpcode());
5182 }
5183 
5184 namespace {
5185   /// CGBR - Create Global Base Reg pass. This initializes the PIC
5186   /// global base register for x86-32.
5187   struct CGBR : public MachineFunctionPass {
5188     static char ID;
5189     CGBR() : MachineFunctionPass(ID) {}
5190 
5191     virtual bool runOnMachineFunction(MachineFunction &MF) {
5192       const X86TargetMachine *TM =
5193         static_cast<const X86TargetMachine *>(&MF.getTarget());
5194 
5195       assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
5196              "X86-64 PIC uses RIP relative addressing");
5197 
5198       // Only emit a global base reg in PIC mode.
5199       if (TM->getRelocationModel() != Reloc::PIC_)
5200         return false;
5201 
5202       X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
5203       unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
5204 
5205       // If we didn't need a GlobalBaseReg, don't insert code.
5206       if (GlobalBaseReg == 0)
5207         return false;
5208 
5209       // Insert the set of GlobalBaseReg into the first MBB of the function
5210       MachineBasicBlock &FirstMBB = MF.front();
5211       MachineBasicBlock::iterator MBBI = FirstMBB.begin();
5212       DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
5213       MachineRegisterInfo &RegInfo = MF.getRegInfo();
5214       const X86InstrInfo *TII = TM->getInstrInfo();
5215 
5216       unsigned PC;
5217       if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
5218         PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
5219       else
5220         PC = GlobalBaseReg;
5221 
5222       // Operand of MovePCtoStack is completely ignored by asm printer. It's
5223       // only used in JIT code emission as displacement to pc.
5224       BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
5225 
5226       // If we're using vanilla 'GOT' PIC style, we should use relative addressing
5227       // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
5228       if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
5229         // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
5230         BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
5231           .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
5232                                         X86II::MO_GOT_ABSOLUTE_ADDRESS);
5233       }
5234 
5235       return true;
5236     }
5237 
5238     virtual const char *getPassName() const {
5239       return "X86 PIC Global Base Reg Initialization";
5240     }
5241 
5242     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5243       AU.setPreservesCFG();
5244       MachineFunctionPass::getAnalysisUsage(AU);
5245     }
5246   };
5247 }
5248 
5249 char CGBR::ID = 0;
5250 FunctionPass*
5251 llvm::createGlobalBaseRegPass() { return new CGBR(); }
5252 
5253 namespace {
5254   struct LDTLSCleanup : public MachineFunctionPass {
5255     static char ID;
5256     LDTLSCleanup() : MachineFunctionPass(ID) {}
5257 
5258     virtual bool runOnMachineFunction(MachineFunction &MF) {
5259       X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
5260       if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
5261         // No point folding accesses if there isn't at least two.
5262         return false;
5263       }
5264 
5265       MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
5266       return VisitNode(DT->getRootNode(), 0);
5267     }
5268 
5269     // Visit the dominator subtree rooted at Node in pre-order.
5270     // If TLSBaseAddrReg is non-null, then use that to replace any
5271     // TLS_base_addr instructions. Otherwise, create the register
5272     // when the first such instruction is seen, and then use it
5273     // as we encounter more instructions.
5274     bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
5275       MachineBasicBlock *BB = Node->getBlock();
5276       bool Changed = false;
5277 
5278       // Traverse the current block.
5279       for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
5280            ++I) {
5281         switch (I->getOpcode()) {
5282           case X86::TLS_base_addr32:
5283           case X86::TLS_base_addr64:
5284             if (TLSBaseAddrReg)
5285               I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
5286             else
5287               I = SetRegister(I, &TLSBaseAddrReg);
5288             Changed = true;
5289             break;
5290           default:
5291             break;
5292         }
5293       }
5294 
5295       // Visit the children of this block in the dominator tree.
5296       for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
5297            I != E; ++I) {
5298         Changed |= VisitNode(*I, TLSBaseAddrReg);
5299       }
5300 
5301       return Changed;
5302     }
5303 
5304     // Replace the TLS_base_addr instruction I with a copy from
5305     // TLSBaseAddrReg, returning the new instruction.
5306     MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
5307                                          unsigned TLSBaseAddrReg) {
5308       MachineFunction *MF = I->getParent()->getParent();
5309       const X86TargetMachine *TM =
5310           static_cast<const X86TargetMachine *>(&MF->getTarget());
5311       const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5312       const X86InstrInfo *TII = TM->getInstrInfo();
5313 
5314       // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
5315       MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
5316                                    TII->get(TargetOpcode::COPY),
5317                                    is64Bit ? X86::RAX : X86::EAX)
5318                                    .addReg(TLSBaseAddrReg);
5319 
5320       // Erase the TLS_base_addr instruction.
5321       I->eraseFromParent();
5322 
5323       return Copy;
5324     }
5325 
5326     // Create a virtal register in *TLSBaseAddrReg, and populate it by
5327     // inserting a copy instruction after I. Returns the new instruction.
5328     MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
5329       MachineFunction *MF = I->getParent()->getParent();
5330       const X86TargetMachine *TM =
5331           static_cast<const X86TargetMachine *>(&MF->getTarget());
5332       const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
5333       const X86InstrInfo *TII = TM->getInstrInfo();
5334 
5335       // Create a virtual register for the TLS base address.
5336       MachineRegisterInfo &RegInfo = MF->getRegInfo();
5337       *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
5338                                                       ? &X86::GR64RegClass
5339                                                       : &X86::GR32RegClass);
5340 
5341       // Insert a copy from RAX/EAX to TLSBaseAddrReg.
5342       MachineInstr *Next = I->getNextNode();
5343       MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
5344                                    TII->get(TargetOpcode::COPY),
5345                                    *TLSBaseAddrReg)
5346                                    .addReg(is64Bit ? X86::RAX : X86::EAX);
5347 
5348       return Copy;
5349     }
5350 
5351     virtual const char *getPassName() const {
5352       return "Local Dynamic TLS Access Clean-up";
5353     }
5354 
5355     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
5356       AU.setPreservesCFG();
5357       AU.addRequired<MachineDominatorTree>();
5358       MachineFunctionPass::getAnalysisUsage(AU);
5359     }
5360   };
5361 }
5362 
5363 char LDTLSCleanup::ID = 0;
5364 FunctionPass*
5365 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }
5366