blob: 306be80dc104d47ddc85058be38678dc4a178b71 [file] [log] [blame]
Jia Liu31d157a2012-02-18 12:03:15 +00001//===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===//
Misha Brukman0e0a7a452005-04-21 23:38:14 +00002//
John Criswellb576c942003-10-20 19:43:21 +00003// The LLVM Compiler Infrastructure
4//
Chris Lattner4ee451d2007-12-29 20:36:04 +00005// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
Misha Brukman0e0a7a452005-04-21 23:38:14 +00007//
John Criswellb576c942003-10-20 19:43:21 +00008//===----------------------------------------------------------------------===//
Chris Lattner72614082002-10-25 22:55:53 +00009//
Chris Lattner3501fea2003-01-14 22:00:31 +000010// This file contains the X86 implementation of the TargetInstrInfo class.
Chris Lattner72614082002-10-25 22:55:53 +000011//
12//===----------------------------------------------------------------------===//
13
Chris Lattner055c9652002-10-29 21:05:24 +000014#include "X86InstrInfo.h"
Chris Lattner4ce42a72002-12-03 05:42:53 +000015#include "X86.h"
Evan Chengaa3c1412006-05-30 21:45:53 +000016#include "X86InstrBuilder.h"
Owen Andersond94b6a12008-01-04 23:57:37 +000017#include "X86MachineFunctionInfo.h"
Evan Chengaa3c1412006-05-30 21:45:53 +000018#include "X86Subtarget.h"
19#include "X86TargetMachine.h"
Dan Gohmand68a0762009-01-05 17:59:02 +000020#include "llvm/DerivedTypes.h"
Owen Anderson0a5372e2009-07-13 04:09:18 +000021#include "llvm/LLVMContext.h"
Owen Anderson718cb662007-09-07 04:06:50 +000022#include "llvm/ADT/STLExtras.h"
Dan Gohman62c939d2008-12-03 05:21:24 +000023#include "llvm/CodeGen/MachineConstantPool.h"
Hans Wennborgf0234fc2012-06-01 16:27:21 +000024#include "llvm/CodeGen/MachineDominators.h"
Owen Andersond94b6a12008-01-04 23:57:37 +000025#include "llvm/CodeGen/MachineFrameInfo.h"
Evan Chengaa3c1412006-05-30 21:45:53 +000026#include "llvm/CodeGen/MachineInstrBuilder.h"
Chris Lattner84bc5422007-12-31 04:13:23 +000027#include "llvm/CodeGen/MachineRegisterInfo.h"
Evan Cheng258ff672006-12-01 21:52:41 +000028#include "llvm/CodeGen/LiveVariables.h"
Craig Topper79aa3412012-03-17 18:46:09 +000029#include "llvm/MC/MCAsmInfo.h"
Chris Lattneree9eb412010-04-26 23:37:21 +000030#include "llvm/MC/MCInst.h"
Owen Anderson43dbe052008-01-07 01:35:02 +000031#include "llvm/Support/CommandLine.h"
David Greene5b901322010-01-05 01:29:29 +000032#include "llvm/Support/Debug.h"
Torok Edwinab7c09b2009-07-08 18:01:40 +000033#include "llvm/Support/ErrorHandling.h"
34#include "llvm/Support/raw_ostream.h"
Evan Cheng0488db92007-09-25 01:57:46 +000035#include "llvm/Target/TargetOptions.h"
David Greeneb87bc952009-11-12 20:55:29 +000036#include <limits>
37
Evan Cheng4db3cff2011-07-01 17:57:27 +000038#define GET_INSTRINFO_CTOR
Evan Cheng22fee2d2011-06-28 20:07:07 +000039#include "X86GenInstrInfo.inc"
40
Brian Gaeked0fde302003-11-11 22:41:34 +000041using namespace llvm;
42
Chris Lattner705e07f2009-08-23 03:41:05 +000043static cl::opt<bool>
44NoFusing("disable-spill-fusing",
45 cl::desc("Disable fusing of spill code into instructions"));
46static cl::opt<bool>
47PrintFailedFusing("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);
51static cl::opt<bool>
52ReMatPICStubLoad("remat-pic-stub-load",
53 cl::desc("Re-materialize load from stub in PIC mode"),
54 cl::init(false), cl::Hidden);
Owen Anderson43dbe052008-01-07 01:35:02 +000055
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +000056enum {
57 // Select which memory operand is being unfolded.
58 // (stored in bits 0 - 7)
59 TB_INDEX_0 = 0,
60 TB_INDEX_1 = 1,
61 TB_INDEX_2 = 2,
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +000062 TB_INDEX_3 = 3,
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +000063 TB_INDEX_MASK = 0xff,
64
65 // Minimum alignment required for load/store.
66 // Used for RegOp->MemOp conversion.
67 // (stored in bits 8 - 15)
68 TB_ALIGN_SHIFT = 8,
69 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT,
70 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT,
71 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT,
72 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT,
73
74 // Do not insert the reverse map (MemOp -> RegOp) into the table.
75 // This may be needed because there is a many -> one mapping.
76 TB_NO_REVERSE = 1 << 16,
77
78 // Do not insert the forward map (RegOp -> MemOp) into the table.
79 // This is needed for Native Client, which prohibits branch
80 // instructions from using a memory operand.
81 TB_NO_FORWARD = 1 << 17,
82
83 TB_FOLDED_LOAD = 1 << 18,
84 TB_FOLDED_STORE = 1 << 19
85};
86
Craig Topper72051bf2012-03-09 07:45:21 +000087struct X86OpTblEntry {
88 uint16_t RegOp;
89 uint16_t MemOp;
90 uint32_t Flags;
91};
92
Evan Chengaa3c1412006-05-30 21:45:53 +000093X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
Evan Cheng4db3cff2011-07-01 17:57:27 +000094 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit()
95 ? X86::ADJCALLSTACKDOWN64
96 : X86::ADJCALLSTACKDOWN32),
97 (tm.getSubtarget<X86Subtarget>().is64Bit()
98 ? X86::ADJCALLSTACKUP64
99 : X86::ADJCALLSTACKUP32)),
Evan Cheng25ab6902006-09-08 06:48:29 +0000100 TM(tm), RI(tm, *this) {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +0000101
Craig Topper72051bf2012-03-09 07:45:21 +0000102 static const X86OpTblEntry OpTbl2Addr[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000103 { X86::ADC32ri, X86::ADC32mi, 0 },
104 { X86::ADC32ri8, X86::ADC32mi8, 0 },
105 { X86::ADC32rr, X86::ADC32mr, 0 },
106 { X86::ADC64ri32, X86::ADC64mi32, 0 },
107 { X86::ADC64ri8, X86::ADC64mi8, 0 },
108 { X86::ADC64rr, X86::ADC64mr, 0 },
109 { X86::ADD16ri, X86::ADD16mi, 0 },
110 { X86::ADD16ri8, X86::ADD16mi8, 0 },
111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE },
112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE },
113 { X86::ADD16rr, X86::ADD16mr, 0 },
114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE },
115 { X86::ADD32ri, X86::ADD32mi, 0 },
116 { X86::ADD32ri8, X86::ADD32mi8, 0 },
117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE },
118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE },
119 { X86::ADD32rr, X86::ADD32mr, 0 },
120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE },
121 { X86::ADD64ri32, X86::ADD64mi32, 0 },
122 { X86::ADD64ri8, X86::ADD64mi8, 0 },
123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE },
124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE },
125 { X86::ADD64rr, X86::ADD64mr, 0 },
126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE },
127 { X86::ADD8ri, X86::ADD8mi, 0 },
128 { X86::ADD8rr, X86::ADD8mr, 0 },
129 { X86::AND16ri, X86::AND16mi, 0 },
130 { X86::AND16ri8, X86::AND16mi8, 0 },
131 { X86::AND16rr, X86::AND16mr, 0 },
132 { X86::AND32ri, X86::AND32mi, 0 },
133 { X86::AND32ri8, X86::AND32mi8, 0 },
134 { X86::AND32rr, X86::AND32mr, 0 },
135 { X86::AND64ri32, X86::AND64mi32, 0 },
136 { X86::AND64ri8, X86::AND64mi8, 0 },
137 { X86::AND64rr, X86::AND64mr, 0 },
138 { X86::AND8ri, X86::AND8mi, 0 },
139 { X86::AND8rr, X86::AND8mr, 0 },
140 { X86::DEC16r, X86::DEC16m, 0 },
141 { X86::DEC32r, X86::DEC32m, 0 },
142 { X86::DEC64_16r, X86::DEC64_16m, 0 },
143 { X86::DEC64_32r, X86::DEC64_32m, 0 },
144 { X86::DEC64r, X86::DEC64m, 0 },
145 { X86::DEC8r, X86::DEC8m, 0 },
146 { X86::INC16r, X86::INC16m, 0 },
147 { X86::INC32r, X86::INC32m, 0 },
148 { X86::INC64_16r, X86::INC64_16m, 0 },
149 { X86::INC64_32r, X86::INC64_32m, 0 },
150 { X86::INC64r, X86::INC64m, 0 },
151 { X86::INC8r, X86::INC8m, 0 },
152 { X86::NEG16r, X86::NEG16m, 0 },
153 { X86::NEG32r, X86::NEG32m, 0 },
154 { X86::NEG64r, X86::NEG64m, 0 },
155 { X86::NEG8r, X86::NEG8m, 0 },
156 { X86::NOT16r, X86::NOT16m, 0 },
157 { X86::NOT32r, X86::NOT32m, 0 },
158 { X86::NOT64r, X86::NOT64m, 0 },
159 { X86::NOT8r, X86::NOT8m, 0 },
160 { X86::OR16ri, X86::OR16mi, 0 },
161 { X86::OR16ri8, X86::OR16mi8, 0 },
162 { X86::OR16rr, X86::OR16mr, 0 },
163 { X86::OR32ri, X86::OR32mi, 0 },
164 { X86::OR32ri8, X86::OR32mi8, 0 },
165 { X86::OR32rr, X86::OR32mr, 0 },
166 { X86::OR64ri32, X86::OR64mi32, 0 },
167 { X86::OR64ri8, X86::OR64mi8, 0 },
168 { X86::OR64rr, X86::OR64mr, 0 },
169 { X86::OR8ri, X86::OR8mi, 0 },
170 { X86::OR8rr, X86::OR8mr, 0 },
171 { X86::ROL16r1, X86::ROL16m1, 0 },
172 { X86::ROL16rCL, X86::ROL16mCL, 0 },
173 { X86::ROL16ri, X86::ROL16mi, 0 },
174 { X86::ROL32r1, X86::ROL32m1, 0 },
175 { X86::ROL32rCL, X86::ROL32mCL, 0 },
176 { X86::ROL32ri, X86::ROL32mi, 0 },
177 { X86::ROL64r1, X86::ROL64m1, 0 },
178 { X86::ROL64rCL, X86::ROL64mCL, 0 },
179 { X86::ROL64ri, X86::ROL64mi, 0 },
180 { X86::ROL8r1, X86::ROL8m1, 0 },
181 { X86::ROL8rCL, X86::ROL8mCL, 0 },
182 { X86::ROL8ri, X86::ROL8mi, 0 },
183 { X86::ROR16r1, X86::ROR16m1, 0 },
184 { X86::ROR16rCL, X86::ROR16mCL, 0 },
185 { X86::ROR16ri, X86::ROR16mi, 0 },
186 { X86::ROR32r1, X86::ROR32m1, 0 },
187 { X86::ROR32rCL, X86::ROR32mCL, 0 },
188 { X86::ROR32ri, X86::ROR32mi, 0 },
189 { X86::ROR64r1, X86::ROR64m1, 0 },
190 { X86::ROR64rCL, X86::ROR64mCL, 0 },
191 { X86::ROR64ri, X86::ROR64mi, 0 },
192 { X86::ROR8r1, X86::ROR8m1, 0 },
193 { X86::ROR8rCL, X86::ROR8mCL, 0 },
194 { X86::ROR8ri, X86::ROR8mi, 0 },
195 { X86::SAR16r1, X86::SAR16m1, 0 },
196 { X86::SAR16rCL, X86::SAR16mCL, 0 },
197 { X86::SAR16ri, X86::SAR16mi, 0 },
198 { X86::SAR32r1, X86::SAR32m1, 0 },
199 { X86::SAR32rCL, X86::SAR32mCL, 0 },
200 { X86::SAR32ri, X86::SAR32mi, 0 },
201 { X86::SAR64r1, X86::SAR64m1, 0 },
202 { X86::SAR64rCL, X86::SAR64mCL, 0 },
203 { X86::SAR64ri, X86::SAR64mi, 0 },
204 { X86::SAR8r1, X86::SAR8m1, 0 },
205 { X86::SAR8rCL, X86::SAR8mCL, 0 },
206 { X86::SAR8ri, X86::SAR8mi, 0 },
207 { X86::SBB32ri, X86::SBB32mi, 0 },
208 { X86::SBB32ri8, X86::SBB32mi8, 0 },
209 { X86::SBB32rr, X86::SBB32mr, 0 },
210 { X86::SBB64ri32, X86::SBB64mi32, 0 },
211 { X86::SBB64ri8, X86::SBB64mi8, 0 },
212 { X86::SBB64rr, X86::SBB64mr, 0 },
213 { X86::SHL16rCL, X86::SHL16mCL, 0 },
214 { X86::SHL16ri, X86::SHL16mi, 0 },
215 { X86::SHL32rCL, X86::SHL32mCL, 0 },
216 { X86::SHL32ri, X86::SHL32mi, 0 },
217 { X86::SHL64rCL, X86::SHL64mCL, 0 },
218 { X86::SHL64ri, X86::SHL64mi, 0 },
219 { X86::SHL8rCL, X86::SHL8mCL, 0 },
220 { X86::SHL8ri, X86::SHL8mi, 0 },
221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 },
222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 },
223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 },
224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 },
225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 },
226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 },
227 { X86::SHR16r1, X86::SHR16m1, 0 },
228 { X86::SHR16rCL, X86::SHR16mCL, 0 },
229 { X86::SHR16ri, X86::SHR16mi, 0 },
230 { X86::SHR32r1, X86::SHR32m1, 0 },
231 { X86::SHR32rCL, X86::SHR32mCL, 0 },
232 { X86::SHR32ri, X86::SHR32mi, 0 },
233 { X86::SHR64r1, X86::SHR64m1, 0 },
234 { X86::SHR64rCL, X86::SHR64mCL, 0 },
235 { X86::SHR64ri, X86::SHR64mi, 0 },
236 { X86::SHR8r1, X86::SHR8m1, 0 },
237 { X86::SHR8rCL, X86::SHR8mCL, 0 },
238 { X86::SHR8ri, X86::SHR8mi, 0 },
239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 },
240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 },
241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 },
242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 },
243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 },
244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 },
245 { X86::SUB16ri, X86::SUB16mi, 0 },
246 { X86::SUB16ri8, X86::SUB16mi8, 0 },
247 { X86::SUB16rr, X86::SUB16mr, 0 },
248 { X86::SUB32ri, X86::SUB32mi, 0 },
249 { X86::SUB32ri8, X86::SUB32mi8, 0 },
250 { X86::SUB32rr, X86::SUB32mr, 0 },
251 { X86::SUB64ri32, X86::SUB64mi32, 0 },
252 { X86::SUB64ri8, X86::SUB64mi8, 0 },
253 { X86::SUB64rr, X86::SUB64mr, 0 },
254 { X86::SUB8ri, X86::SUB8mi, 0 },
255 { X86::SUB8rr, X86::SUB8mr, 0 },
256 { X86::XOR16ri, X86::XOR16mi, 0 },
257 { X86::XOR16ri8, X86::XOR16mi8, 0 },
258 { X86::XOR16rr, X86::XOR16mr, 0 },
259 { X86::XOR32ri, X86::XOR32mi, 0 },
260 { X86::XOR32ri8, X86::XOR32mi8, 0 },
261 { X86::XOR32rr, X86::XOR32mr, 0 },
262 { X86::XOR64ri32, X86::XOR64mi32, 0 },
263 { X86::XOR64ri8, X86::XOR64mi8, 0 },
264 { X86::XOR64rr, X86::XOR64mr, 0 },
265 { X86::XOR8ri, X86::XOR8mi, 0 },
266 { X86::XOR8rr, X86::XOR8mr, 0 }
Owen Anderson43dbe052008-01-07 01:35:02 +0000267 };
268
269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000270 unsigned RegOp = OpTbl2Addr[i].RegOp;
271 unsigned MemOp = OpTbl2Addr[i].MemOp;
272 unsigned Flags = OpTbl2Addr[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable,
274 RegOp, MemOp,
275 // Index 0, folded load and store, no alignment requirement.
276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE);
Owen Anderson43dbe052008-01-07 01:35:02 +0000277 }
278
Craig Topper72051bf2012-03-09 07:45:21 +0000279 static const X86OpTblEntry OpTbl0[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD },
281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD },
282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD },
283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD },
284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD },
286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD },
287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD },
288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD },
289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD },
290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD },
291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD },
292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD },
293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD },
294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD },
295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD },
296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD },
297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD },
298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD },
299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD },
300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD },
304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD },
305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD },
306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD },
307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD },
308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD },
309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD },
310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD },
311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD },
312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD },
313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE },
314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE },
315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE },
316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE },
317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE },
318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE },
319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE },
320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE },
321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE },
322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE },
326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE },
327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE },
328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE },
329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE },
330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD },
332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD },
333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD },
334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD },
335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE },
336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE },
337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE },
338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE },
339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE },
340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE },
341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE },
342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE },
343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE },
344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE },
345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE },
346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE },
347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE },
348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE },
349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE },
350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE },
351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD },
352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD },
353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD },
354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD },
355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD },
357 // AVX 128-bit versions of foldable instructions
358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE },
360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE },
Craig Topper446626d2012-01-14 18:14:53 +0000361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 },
363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 },
364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 },
365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE },
366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE },
367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE },
368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE },
369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE },
370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE },
371 // AVX 256-bit foldable instructions
Craig Topper446626d2012-01-14 18:14:53 +0000372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 },
376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE },
377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }
Owen Anderson43dbe052008-01-07 01:35:02 +0000378 };
379
380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000381 unsigned RegOp = OpTbl0[i].RegOp;
382 unsigned MemOp = OpTbl0[i].MemOp;
383 unsigned Flags = OpTbl0[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable,
385 RegOp, MemOp, TB_INDEX_0 | Flags);
Owen Anderson43dbe052008-01-07 01:35:02 +0000386 }
387
Craig Topper72051bf2012-03-09 07:45:21 +0000388 static const X86OpTblEntry OpTbl1[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000389 { X86::CMP16rr, X86::CMP16rm, 0 },
390 { X86::CMP32rr, X86::CMP32rm, 0 },
391 { X86::CMP64rr, X86::CMP64rm, 0 },
392 { X86::CMP8rr, X86::CMP8rm, 0 },
393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 },
394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 },
395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 },
396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 },
397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 },
398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 },
399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 },
400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 },
401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 },
402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 },
403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE },
404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000405 { X86::IMUL16rri, X86::IMUL16rmi, 0 },
406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 },
407 { X86::IMUL32rri, X86::IMUL32rmi, 0 },
408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 },
409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 },
410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 },
411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 },
412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 },
413 { X86::Int_CVTDQ2PDrr, X86::Int_CVTDQ2PDrm, TB_ALIGN_16 },
414 { X86::Int_CVTDQ2PSrr, X86::Int_CVTDQ2PSrm, TB_ALIGN_16 },
415 { X86::Int_CVTPD2DQrr, X86::Int_CVTPD2DQrm, TB_ALIGN_16 },
416 { X86::Int_CVTPD2PSrr, X86::Int_CVTPD2PSrm, TB_ALIGN_16 },
417 { X86::Int_CVTPS2DQrr, X86::Int_CVTPS2DQrm, TB_ALIGN_16 },
418 { X86::Int_CVTPS2PDrr, X86::Int_CVTPS2PDrm, 0 },
419 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 },
420 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000421 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 },
422 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000423 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 },
424 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 },
425 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 },
426 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 },
427 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 },
428 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 },
429 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 },
430 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 },
431 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 },
432 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 },
433 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 },
434 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 },
435 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 },
436 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000437 { X86::MOV16rr, X86::MOV16rm, 0 },
438 { X86::MOV32rr, X86::MOV32rm, 0 },
439 { X86::MOV64rr, X86::MOV64rm, 0 },
440 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 },
441 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 },
442 { X86::MOV8rr, X86::MOV8rm, 0 },
443 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 },
444 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000445 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 },
446 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 },
447 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 },
448 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000449 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 },
450 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 },
451 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 },
452 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 },
453 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 },
454 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 },
455 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 },
456 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 },
457 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 },
458 { X86::MOVUPSrr, X86::MOVUPSrm, 0 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000459 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 },
460 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 },
461 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 },
462 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 },
463 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 },
464 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 },
465 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 },
466 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 },
467 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 },
468 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000469 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 },
470 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 },
471 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000472 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 },
473 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 },
474 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 },
475 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 },
476 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 },
477 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 },
478 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 },
479 { X86::RSQRTSSr, X86::RSQRTSSm, 0 },
480 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 },
481 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 },
482 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 },
483 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 },
484 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 },
485 { X86::SQRTSDr, X86::SQRTSDm, 0 },
486 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 },
487 { X86::SQRTSSr, X86::SQRTSSm, 0 },
488 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 },
489 { X86::TEST16rr, X86::TEST16rm, 0 },
490 { X86::TEST32rr, X86::TEST32rm, 0 },
491 { X86::TEST64rr, X86::TEST64rm, 0 },
492 { X86::TEST8rr, X86::TEST8rm, 0 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000493 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000494 { X86::UCOMISDrr, X86::UCOMISDrm, 0 },
495 { X86::UCOMISSrr, X86::UCOMISSrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000496 // AVX 128-bit versions of foldable instructions
497 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 },
498 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 },
499 { X86::Int_VCVTDQ2PDrr, X86::Int_VCVTDQ2PDrm, TB_ALIGN_16 },
500 { X86::Int_VCVTDQ2PSrr, X86::Int_VCVTDQ2PSrm, TB_ALIGN_16 },
501 { X86::Int_VCVTPD2DQrr, X86::Int_VCVTPD2DQrm, TB_ALIGN_16 },
502 { X86::Int_VCVTPD2PSrr, X86::Int_VCVTPD2PSrm, TB_ALIGN_16 },
503 { X86::Int_VCVTPS2DQrr, X86::Int_VCVTPS2DQrm, TB_ALIGN_16 },
504 { X86::Int_VCVTPS2PDrr, X86::Int_VCVTPS2PDrm, 0 },
505 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 },
506 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000507 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 },
508 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 },
Pete Cooper312091e2012-06-14 22:12:58 +0000509 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 },
Craig Topper8e58b3e2012-06-15 07:02:58 +0000510 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 },
511 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 },
512 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 },
513 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 },
514 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 },
515 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 },
516 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 },
517 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 },
518 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000519 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE },
520 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE },
521 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 },
522 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 },
523 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 },
524 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 },
525 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 },
526 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 },
527 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 },
528 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 },
529 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 },
530 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 },
531 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 },
532 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 },
533 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 },
534 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 },
535 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000536 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 },
537 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 },
538 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000539 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000540 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000541 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 },
542 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 },
543 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 },
544 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 },
545 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 },
546 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 },
547 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 },
548 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 },
549 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 },
550 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 },
551 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000552 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000553 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 },
554 // AVX 256-bit foldable instructions
555 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 },
556 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +0000557 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000558 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000559 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 },
Craig Topper40385c82012-01-19 08:50:38 +0000560 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 },
561 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +0000562 // AVX2 foldable instructions
Craig Topper40385c82012-01-19 08:50:38 +0000563 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 },
564 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 },
565 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 },
566 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 },
567 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 },
568 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 },
569 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 },
570 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 },
571 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 },
572 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 },
573 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 },
574 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 },
575 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 },
576 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000577 };
578
579 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +0000580 unsigned RegOp = OpTbl1[i].RegOp;
581 unsigned MemOp = OpTbl1[i].MemOp;
582 unsigned Flags = OpTbl1[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000583 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable,
584 RegOp, MemOp,
585 // Index 1, folded load
586 Flags | TB_INDEX_1 | TB_FOLDED_LOAD);
Owen Anderson43dbe052008-01-07 01:35:02 +0000587 }
588
Craig Topper72051bf2012-03-09 07:45:21 +0000589 static const X86OpTblEntry OpTbl2[] = {
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000590 { X86::ADC32rr, X86::ADC32rm, 0 },
591 { X86::ADC64rr, X86::ADC64rm, 0 },
592 { X86::ADD16rr, X86::ADD16rm, 0 },
593 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE },
594 { X86::ADD32rr, X86::ADD32rm, 0 },
595 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE },
596 { X86::ADD64rr, X86::ADD64rm, 0 },
597 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE },
598 { X86::ADD8rr, X86::ADD8rm, 0 },
599 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 },
600 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 },
601 { X86::ADDSDrr, X86::ADDSDrm, 0 },
602 { X86::ADDSSrr, X86::ADDSSrm, 0 },
603 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 },
604 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 },
605 { X86::AND16rr, X86::AND16rm, 0 },
606 { X86::AND32rr, X86::AND32rm, 0 },
607 { X86::AND64rr, X86::AND64rm, 0 },
608 { X86::AND8rr, X86::AND8rm, 0 },
609 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 },
610 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 },
611 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 },
612 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000613 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 },
614 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 },
615 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 },
616 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000617 { X86::CMOVA16rr, X86::CMOVA16rm, 0 },
618 { X86::CMOVA32rr, X86::CMOVA32rm, 0 },
619 { X86::CMOVA64rr, X86::CMOVA64rm, 0 },
620 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 },
621 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 },
622 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 },
623 { X86::CMOVB16rr, X86::CMOVB16rm, 0 },
624 { X86::CMOVB32rr, X86::CMOVB32rm, 0 },
625 { X86::CMOVB64rr, X86::CMOVB64rm, 0 },
626 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 },
627 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 },
628 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 },
629 { X86::CMOVE16rr, X86::CMOVE16rm, 0 },
630 { X86::CMOVE32rr, X86::CMOVE32rm, 0 },
631 { X86::CMOVE64rr, X86::CMOVE64rm, 0 },
632 { X86::CMOVG16rr, X86::CMOVG16rm, 0 },
633 { X86::CMOVG32rr, X86::CMOVG32rm, 0 },
634 { X86::CMOVG64rr, X86::CMOVG64rm, 0 },
635 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 },
636 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 },
637 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 },
638 { X86::CMOVL16rr, X86::CMOVL16rm, 0 },
639 { X86::CMOVL32rr, X86::CMOVL32rm, 0 },
640 { X86::CMOVL64rr, X86::CMOVL64rm, 0 },
641 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 },
642 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 },
643 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 },
644 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 },
645 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 },
646 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 },
647 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 },
648 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 },
649 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 },
650 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 },
651 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 },
652 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 },
653 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 },
654 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 },
655 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 },
656 { X86::CMOVO16rr, X86::CMOVO16rm, 0 },
657 { X86::CMOVO32rr, X86::CMOVO32rm, 0 },
658 { X86::CMOVO64rr, X86::CMOVO64rm, 0 },
659 { X86::CMOVP16rr, X86::CMOVP16rm, 0 },
660 { X86::CMOVP32rr, X86::CMOVP32rm, 0 },
661 { X86::CMOVP64rr, X86::CMOVP64rm, 0 },
662 { X86::CMOVS16rr, X86::CMOVS16rm, 0 },
663 { X86::CMOVS32rr, X86::CMOVS32rm, 0 },
664 { X86::CMOVS64rr, X86::CMOVS64rm, 0 },
665 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 },
666 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 },
667 { X86::CMPSDrr, X86::CMPSDrm, 0 },
668 { X86::CMPSSrr, X86::CMPSSrm, 0 },
669 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 },
670 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 },
671 { X86::DIVSDrr, X86::DIVSDrm, 0 },
672 { X86::DIVSSrr, X86::DIVSSrm, 0 },
673 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 },
674 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 },
675 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 },
676 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 },
677 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 },
678 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 },
679 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 },
680 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 },
681 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 },
682 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 },
683 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 },
684 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 },
685 { X86::IMUL16rr, X86::IMUL16rm, 0 },
686 { X86::IMUL32rr, X86::IMUL32rm, 0 },
687 { X86::IMUL64rr, X86::IMUL64rm, 0 },
688 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 },
689 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 },
690 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 },
691 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 },
692 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 },
693 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 },
694 { X86::MAXSDrr, X86::MAXSDrm, 0 },
695 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 },
696 { X86::MAXSSrr, X86::MAXSSrm, 0 },
697 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 },
698 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 },
699 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 },
700 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 },
701 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 },
702 { X86::MINSDrr, X86::MINSDrm, 0 },
703 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 },
704 { X86::MINSSrr, X86::MINSSrm, 0 },
705 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000706 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000707 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 },
708 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 },
709 { X86::MULSDrr, X86::MULSDrm, 0 },
710 { X86::MULSSrr, X86::MULSSrm, 0 },
711 { X86::OR16rr, X86::OR16rm, 0 },
712 { X86::OR32rr, X86::OR32rm, 0 },
713 { X86::OR64rr, X86::OR64rm, 0 },
714 { X86::OR8rr, X86::OR8rm, 0 },
715 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 },
716 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 },
717 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 },
718 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000719 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000720 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 },
721 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 },
722 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 },
723 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 },
724 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 },
725 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000726 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 },
727 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000728 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000729 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000730 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 },
731 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 },
732 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 },
733 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000734 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000735 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 },
736 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000737 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000738 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 },
739 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 },
740 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000741 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000742 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000743 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 },
744 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000745 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000746 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000747 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000748 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000749 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000750 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000751 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 },
752 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 },
753 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 },
754 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 },
755 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 },
756 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000757 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000758 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 },
759 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 },
760 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 },
761 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 },
762 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 },
763 { X86::PORrr, X86::PORrm, TB_ALIGN_16 },
764 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 },
Craig Topper969ba282012-01-25 06:43:11 +0000765 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 },
766 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 },
767 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 },
768 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000769 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 },
770 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 },
771 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 },
772 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 },
773 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 },
774 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 },
775 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 },
776 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 },
777 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 },
778 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 },
779 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 },
780 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 },
781 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 },
782 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 },
783 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 },
784 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 },
785 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 },
786 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 },
787 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 },
788 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 },
789 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 },
790 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 },
791 { X86::SBB32rr, X86::SBB32rm, 0 },
792 { X86::SBB64rr, X86::SBB64rm, 0 },
793 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 },
794 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 },
795 { X86::SUB16rr, X86::SUB16rm, 0 },
796 { X86::SUB32rr, X86::SUB32rm, 0 },
797 { X86::SUB64rr, X86::SUB64rm, 0 },
798 { X86::SUB8rr, X86::SUB8rm, 0 },
799 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 },
800 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 },
801 { X86::SUBSDrr, X86::SUBSDrm, 0 },
802 { X86::SUBSSrr, X86::SUBSSrm, 0 },
Owen Anderson43dbe052008-01-07 01:35:02 +0000803 // FIXME: TEST*rr -> swapped operand of TEST*mr.
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +0000804 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 },
805 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 },
806 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 },
807 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 },
808 { X86::XOR16rr, X86::XOR16rm, 0 },
809 { X86::XOR32rr, X86::XOR32rm, 0 },
810 { X86::XOR64rr, X86::XOR64rm, 0 },
811 { X86::XOR8rr, X86::XOR8rm, 0 },
812 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000813 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 },
814 // AVX 128-bit versions of foldable instructions
815 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 },
816 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 },
817 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 },
818 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 },
819 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 },
820 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 },
821 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 },
822 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 },
823 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 },
824 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 },
825 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 },
826 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000827 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQrm, TB_ALIGN_16 },
828 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 },
829 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 },
830 { X86::VSQRTSDr, X86::VSQRTSDm, 0 },
831 { X86::VSQRTSSr, X86::VSQRTSSm, 0 },
832 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 },
833 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 },
834 { X86::VADDSDrr, X86::VADDSDrm, 0 },
835 { X86::VADDSSrr, X86::VADDSSrm, 0 },
836 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 },
837 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 },
838 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 },
839 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 },
840 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 },
841 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000842 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 },
843 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 },
844 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 },
845 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000846 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 },
847 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 },
848 { X86::VCMPSDrr, X86::VCMPSDrm, 0 },
849 { X86::VCMPSSrr, X86::VCMPSSrm, 0 },
850 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 },
851 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 },
852 { X86::VDIVSDrr, X86::VDIVSDrm, 0 },
853 { X86::VDIVSSrr, X86::VDIVSSrm, 0 },
854 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 },
855 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 },
856 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 },
857 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 },
858 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 },
859 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 },
860 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 },
861 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 },
862 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 },
863 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 },
864 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 },
865 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 },
866 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 },
867 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 },
868 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 },
869 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 },
870 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 },
871 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 },
872 { X86::VMAXSDrr, X86::VMAXSDrm, 0 },
873 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 },
874 { X86::VMAXSSrr, X86::VMAXSSrm, 0 },
875 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 },
876 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 },
877 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 },
878 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 },
879 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 },
880 { X86::VMINSDrr, X86::VMINSDrm, 0 },
881 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 },
882 { X86::VMINSSrr, X86::VMINSSrm, 0 },
883 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 },
Craig Topperdcce2442011-11-14 08:07:55 +0000884 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000885 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 },
886 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 },
887 { X86::VMULSDrr, X86::VMULSDrm, 0 },
888 { X86::VMULSSrr, X86::VMULSSrm, 0 },
889 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 },
890 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 },
891 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 },
892 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000893 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000894 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 },
895 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 },
896 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 },
897 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 },
898 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 },
899 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000900 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 },
901 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000902 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000903 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000904 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 },
905 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000906 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 },
907 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000908 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000909 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 },
910 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000911 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000912 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 },
913 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 },
914 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000915 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000916 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000917 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000918 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000919 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 },
920 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000921 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 },
Craig Topper4bb3f342012-01-25 05:37:32 +0000922 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000923 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 },
924 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000925 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000926 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000927 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 },
928 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 },
929 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 },
930 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 },
931 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 },
932 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000933 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000934 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 },
935 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 },
936 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 },
937 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 },
938 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 },
939 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 },
940 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 },
Craig Topper969ba282012-01-25 06:43:11 +0000941 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 },
942 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 },
943 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 },
944 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +0000945 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 },
946 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 },
947 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 },
948 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 },
949 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 },
950 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 },
951 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 },
952 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 },
953 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 },
954 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 },
955 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 },
956 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 },
957 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 },
958 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 },
959 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 },
960 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 },
961 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 },
962 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 },
963 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 },
964 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 },
965 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 },
966 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 },
967 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 },
968 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 },
969 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 },
970 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 },
971 { X86::VSUBSDrr, X86::VSUBSDrm, 0 },
972 { X86::VSUBSSrr, X86::VSUBSSrm, 0 },
973 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 },
974 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 },
975 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 },
976 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 },
977 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 },
Craig Topperdcce2442011-11-14 08:07:55 +0000978 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +0000979 // AVX 256-bit foldable instructions
980 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 },
981 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 },
982 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 },
983 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 },
984 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 },
985 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 },
986 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 },
987 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 },
988 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 },
989 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 },
990 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 },
991 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 },
992 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 },
993 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 },
994 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 },
995 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 },
996 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 },
997 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 },
998 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 },
999 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 },
1000 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 },
1001 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 },
1002 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 },
1003 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 },
1004 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 },
1005 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 },
1006 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 },
1007 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 },
1008 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 },
1009 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 },
1010 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 },
1011 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 },
1012 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 },
1013 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 },
1014 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 },
1015 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 },
1016 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 },
1017 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 },
1018 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 },
1019 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 },
1020 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 },
1021 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 },
1022 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 },
1023 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 },
1024 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 },
1025 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001026 // AVX2 foldable instructions
Craig Topper446626d2012-01-14 18:14:53 +00001027 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 },
1028 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 },
1029 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 },
1030 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 },
1031 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 },
1032 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 },
1033 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 },
1034 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 },
1035 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 },
1036 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 },
1037 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 },
1038 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 },
1039 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 },
1040 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 },
1041 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 },
1042 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 },
1043 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 },
1044 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 },
1045 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 },
1046 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 },
1047 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 },
1048 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 },
1049 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 },
1050 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 },
1051 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 },
1052 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 },
1053 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 },
1054 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 },
1055 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 },
1056 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +00001057 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 },
Elena Demikhovsky73c504a2012-04-15 11:18:59 +00001058 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 },
Craig Topper40385c82012-01-19 08:50:38 +00001059 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 },
Elena Demikhovsky73c504a2012-04-15 11:18:59 +00001060 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +00001061 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +00001062 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +00001063 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 },
1064 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +00001065 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 },
Craig Topper4bb3f342012-01-25 05:37:32 +00001066 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 },
Craig Topper446626d2012-01-14 18:14:53 +00001067 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 },
1068 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 },
1069 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 },
1070 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 },
1071 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 },
1072 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 },
1073 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 },
1074 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 },
1075 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 },
1076 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 },
1077 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 },
1078 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 },
1079 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 },
1080 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 },
1081 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 },
1082 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 },
Craig Topper969ba282012-01-25 06:43:11 +00001083 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 },
1084 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 },
1085 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 },
1086 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001087 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 },
1088 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 },
1089 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 },
1090 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +00001091 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001092 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +00001093 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001094 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 },
1095 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 },
1096 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +00001097 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001098 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 },
1099 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 },
1100 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 },
1101 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +00001102 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001103 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +00001104 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 },
1105 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 },
1106 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 },
1107 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 },
1108 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 },
1109 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 },
1110 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 },
1111 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 },
Craig Topperdcce2442011-11-14 08:07:55 +00001112 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 },
Craig Topper446626d2012-01-14 18:14:53 +00001113 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 },
1114 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 },
1115 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 },
1116 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 },
1117 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 },
1118 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 },
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00001119 // FIXME: add AVX 256-bit foldable instructions
Owen Anderson43dbe052008-01-07 01:35:02 +00001120 };
1121
1122 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) {
Craig Topper72051bf2012-03-09 07:45:21 +00001123 unsigned RegOp = OpTbl2[i].RegOp;
1124 unsigned MemOp = OpTbl2[i].MemOp;
1125 unsigned Flags = OpTbl2[i].Flags;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00001126 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable,
1127 RegOp, MemOp,
1128 // Index 2, folded load
1129 Flags | TB_INDEX_2 | TB_FOLDED_LOAD);
Owen Anderson43dbe052008-01-07 01:35:02 +00001130 }
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001131
1132 static const X86OpTblEntry OpTbl3[] = {
1133 // FMA foldable instructions
Craig Toppercaea5e22012-06-04 07:46:16 +00001134 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 },
1135 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 },
1136 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 },
1137 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 },
1138 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 },
1139 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 },
1140 { X86::VFMADDSSr132r_Int, X86::VFMADDSSr132m_Int, 0 },
1141 { X86::VFMADDSDr132r_Int, X86::VFMADDSDr132m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001142
Craig Toppercaea5e22012-06-04 07:46:16 +00001143 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 },
1144 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 },
1145 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 },
1146 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 },
1147 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 },
1148 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 },
1149 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 },
1150 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 },
1151 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 },
1152 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 },
1153 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 },
1154 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 },
1155 { X86::VFMADDPSr132r_Int, X86::VFMADDPSr132m_Int, TB_ALIGN_16 },
1156 { X86::VFMADDPDr132r_Int, X86::VFMADDPDr132m_Int, TB_ALIGN_16 },
1157 { X86::VFMADDPSr132rY_Int, X86::VFMADDPSr132mY_Int, TB_ALIGN_32 },
1158 { X86::VFMADDPDr132rY_Int, X86::VFMADDPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001159
Craig Toppercaea5e22012-06-04 07:46:16 +00001160 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 },
1161 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 },
1162 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 },
1163 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 },
1164 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 },
1165 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 },
1166 { X86::VFNMADDSSr132r_Int, X86::VFNMADDSSr132m_Int, 0 },
1167 { X86::VFNMADDSDr132r_Int, X86::VFNMADDSDr132m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001168
Craig Toppercaea5e22012-06-04 07:46:16 +00001169 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 },
1170 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 },
1171 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 },
1172 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 },
1173 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 },
1174 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 },
1175 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 },
1176 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 },
1177 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 },
1178 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 },
1179 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 },
1180 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 },
1181 { X86::VFNMADDPSr132r_Int, X86::VFNMADDPSr132m_Int, TB_ALIGN_16 },
1182 { X86::VFNMADDPDr132r_Int, X86::VFNMADDPDr132m_Int, TB_ALIGN_16 },
1183 { X86::VFNMADDPSr132rY_Int, X86::VFNMADDPSr132mY_Int, TB_ALIGN_32 },
1184 { X86::VFNMADDPDr132rY_Int, X86::VFNMADDPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001185
Craig Toppercaea5e22012-06-04 07:46:16 +00001186 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 },
1187 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 },
1188 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 },
1189 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 },
1190 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 },
1191 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 },
1192 { X86::VFMSUBSSr132r_Int, X86::VFMSUBSSr132m_Int, 0 },
1193 { X86::VFMSUBSDr132r_Int, X86::VFMSUBSDr132m_Int, 0 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001194
Craig Toppercaea5e22012-06-04 07:46:16 +00001195 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 },
1196 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 },
1197 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 },
1198 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 },
1199 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 },
1200 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 },
1201 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 },
1202 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 },
1203 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 },
1204 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 },
1205 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 },
1206 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 },
1207 { X86::VFMSUBPSr132r_Int, X86::VFMSUBPSr132m_Int, TB_ALIGN_16 },
1208 { X86::VFMSUBPDr132r_Int, X86::VFMSUBPDr132m_Int, TB_ALIGN_16 },
1209 { X86::VFMSUBPSr132rY_Int, X86::VFMSUBPSr132mY_Int, TB_ALIGN_32 },
1210 { X86::VFMSUBPDr132rY_Int, X86::VFMSUBPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001211
Craig Toppercaea5e22012-06-04 07:46:16 +00001212 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 },
1213 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 },
1214 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 },
1215 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 },
1216 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 },
1217 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 },
1218 { X86::VFNMSUBSSr132r_Int, X86::VFNMSUBSSr132m_Int, 0 },
1219 { X86::VFNMSUBSDr132r_Int, X86::VFNMSUBSDr132m_Int, 0 },
Craig Topper78fc72d2012-06-01 05:48:39 +00001220
Craig Toppercaea5e22012-06-04 07:46:16 +00001221 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 },
1222 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 },
1223 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 },
1224 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 },
1225 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 },
1226 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 },
1227 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 },
1228 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 },
1229 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 },
1230 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 },
1231 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 },
1232 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 },
1233 { X86::VFNMSUBPSr132r_Int, X86::VFNMSUBPSr132m_Int, TB_ALIGN_16 },
1234 { X86::VFNMSUBPDr132r_Int, X86::VFNMSUBPDr132m_Int, TB_ALIGN_16 },
1235 { X86::VFNMSUBPSr132rY_Int, X86::VFNMSUBPSr132mY_Int, TB_ALIGN_32 },
1236 { X86::VFNMSUBPDr132rY_Int, X86::VFNMSUBPDr132mY_Int, TB_ALIGN_32 },
Craig Topperfc5ab242012-06-04 07:08:21 +00001237
Craig Toppercaea5e22012-06-04 07:46:16 +00001238 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 },
1239 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 },
1240 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 },
1241 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 },
1242 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 },
1243 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 },
1244 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 },
1245 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 },
1246 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 },
1247 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 },
1248 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 },
1249 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 },
1250 { X86::VFMADDSUBPSr132r_Int, X86::VFMADDSUBPSr132m_Int, TB_ALIGN_16 },
1251 { X86::VFMADDSUBPDr132r_Int, X86::VFMADDSUBPDr132m_Int, TB_ALIGN_16 },
1252 { X86::VFMADDSUBPSr132rY_Int, X86::VFMADDSUBPSr132mY_Int, TB_ALIGN_32 },
1253 { X86::VFMADDSUBPDr132rY_Int, X86::VFMADDSUBPDr132mY_Int, TB_ALIGN_32 },
Craig Topperfc5ab242012-06-04 07:08:21 +00001254
Craig Toppercaea5e22012-06-04 07:46:16 +00001255 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 },
1256 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 },
1257 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 },
1258 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 },
1259 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 },
1260 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 },
1261 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 },
1262 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 },
1263 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 },
1264 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 },
1265 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 },
1266 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 },
1267 { X86::VFMSUBADDPSr132r_Int, X86::VFMSUBADDPSr132m_Int, TB_ALIGN_16 },
1268 { X86::VFMSUBADDPDr132r_Int, X86::VFMSUBADDPDr132m_Int, TB_ALIGN_16 },
1269 { X86::VFMSUBADDPSr132rY_Int, X86::VFMSUBADDPSr132mY_Int, TB_ALIGN_32 },
1270 { X86::VFMSUBADDPDr132rY_Int, X86::VFMSUBADDPDr132mY_Int, TB_ALIGN_32 },
Elena Demikhovsky177cf1e2012-05-31 09:20:20 +00001271 };
1272
1273 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) {
1274 unsigned RegOp = OpTbl3[i].RegOp;
1275 unsigned MemOp = OpTbl3[i].MemOp;
1276 unsigned Flags = OpTbl3[i].Flags;
1277 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable,
1278 RegOp, MemOp,
1279 // Index 3, folded load
1280 Flags | TB_INDEX_3 | TB_FOLDED_LOAD);
1281 }
1282
Chris Lattner72614082002-10-25 22:55:53 +00001283}
1284
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00001285void
1286X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable,
1287 MemOp2RegOpTableType &M2RTable,
1288 unsigned RegOp, unsigned MemOp, unsigned Flags) {
1289 if ((Flags & TB_NO_FORWARD) == 0) {
1290 assert(!R2MTable.count(RegOp) && "Duplicate entry!");
1291 R2MTable[RegOp] = std::make_pair(MemOp, Flags);
1292 }
1293 if ((Flags & TB_NO_REVERSE) == 0) {
1294 assert(!M2RTable.count(MemOp) &&
1295 "Duplicated entries in unfolding maps?");
1296 M2RTable[MemOp] = std::make_pair(RegOp, Flags);
1297 }
1298}
1299
Evan Chenga5a81d72010-01-12 00:09:37 +00001300bool
Evan Cheng7da9ecf2010-01-13 00:30:23 +00001301X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
1302 unsigned &SrcReg, unsigned &DstReg,
1303 unsigned &SubIdx) const {
Evan Chenga5a81d72010-01-12 00:09:37 +00001304 switch (MI.getOpcode()) {
1305 default: break;
1306 case X86::MOVSX16rr8:
1307 case X86::MOVZX16rr8:
1308 case X86::MOVSX32rr8:
1309 case X86::MOVZX32rr8:
1310 case X86::MOVSX64rr8:
1311 case X86::MOVZX64rr8:
Evan Cheng57d1d932010-01-13 08:01:32 +00001312 if (!TM.getSubtarget<X86Subtarget>().is64Bit())
1313 // It's not always legal to reference the low 8-bit of the larger
1314 // register in 32-bit mode.
1315 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +00001316 case X86::MOVSX32rr16:
1317 case X86::MOVZX32rr16:
1318 case X86::MOVSX64rr16:
1319 case X86::MOVZX64rr16:
1320 case X86::MOVSX64rr32:
1321 case X86::MOVZX64rr32: {
1322 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg())
1323 // Be conservative.
1324 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +00001325 SrcReg = MI.getOperand(1).getReg();
1326 DstReg = MI.getOperand(0).getReg();
Evan Chenga5a81d72010-01-12 00:09:37 +00001327 switch (MI.getOpcode()) {
1328 default:
1329 llvm_unreachable(0);
Evan Chenga5a81d72010-01-12 00:09:37 +00001330 case X86::MOVSX16rr8:
1331 case X86::MOVZX16rr8:
1332 case X86::MOVSX32rr8:
1333 case X86::MOVZX32rr8:
1334 case X86::MOVSX64rr8:
1335 case X86::MOVZX64rr8:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +00001336 SubIdx = X86::sub_8bit;
Evan Chenga5a81d72010-01-12 00:09:37 +00001337 break;
1338 case X86::MOVSX32rr16:
1339 case X86::MOVZX32rr16:
1340 case X86::MOVSX64rr16:
1341 case X86::MOVZX64rr16:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +00001342 SubIdx = X86::sub_16bit;
Evan Chenga5a81d72010-01-12 00:09:37 +00001343 break;
1344 case X86::MOVSX64rr32:
1345 case X86::MOVZX64rr32:
Jakob Stoklund Olesen22c0e972010-05-25 17:04:16 +00001346 SubIdx = X86::sub_32bit;
Evan Chenga5a81d72010-01-12 00:09:37 +00001347 break;
1348 }
Evan Cheng7da9ecf2010-01-13 00:30:23 +00001349 return true;
Evan Chenga5a81d72010-01-12 00:09:37 +00001350 }
1351 }
Evan Cheng7da9ecf2010-01-13 00:30:23 +00001352 return false;
Evan Chenga5a81d72010-01-12 00:09:37 +00001353}
1354
David Greeneb87bc952009-11-12 20:55:29 +00001355/// isFrameOperand - Return true and the FrameIndex if the specified
1356/// operand and follow operands form a reference to the stack frame.
1357bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
1358 int &FrameIndex) const {
1359 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() &&
1360 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() &&
1361 MI->getOperand(Op+1).getImm() == 1 &&
1362 MI->getOperand(Op+2).getReg() == 0 &&
1363 MI->getOperand(Op+3).getImm() == 0) {
1364 FrameIndex = MI->getOperand(Op).getIndex();
1365 return true;
1366 }
1367 return false;
1368}
1369
David Greenedda39782009-11-13 00:29:53 +00001370static bool isFrameLoadOpcode(int Opcode) {
1371 switch (Opcode) {
David Blaikie4d6ccb52012-01-20 21:51:11 +00001372 default:
1373 return false;
Chris Lattner40839602006-02-02 20:12:32 +00001374 case X86::MOV8rm:
1375 case X86::MOV16rm:
1376 case X86::MOV32rm:
Evan Cheng25ab6902006-09-08 06:48:29 +00001377 case X86::MOV64rm:
Dale Johannesene377d4d2007-07-04 21:07:47 +00001378 case X86::LD_Fp64m:
Chris Lattner40839602006-02-02 20:12:32 +00001379 case X86::MOVSSrm:
1380 case X86::MOVSDrm:
Chris Lattner993c8972006-04-18 16:44:51 +00001381 case X86::MOVAPSrm:
1382 case X86::MOVAPDrm:
Dan Gohman54462742009-01-09 02:40:34 +00001383 case X86::MOVDQArm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00001384 case X86::VMOVSSrm:
1385 case X86::VMOVSDrm:
1386 case X86::VMOVAPSrm:
1387 case X86::VMOVAPDrm:
1388 case X86::VMOVDQArm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00001389 case X86::VMOVAPSYrm:
1390 case X86::VMOVAPDYrm:
1391 case X86::VMOVDQAYrm:
Bill Wendling823efee2007-04-03 06:00:37 +00001392 case X86::MMX_MOVD64rm:
1393 case X86::MMX_MOVQ64rm:
David Greenedda39782009-11-13 00:29:53 +00001394 return true;
David Greenedda39782009-11-13 00:29:53 +00001395 }
David Greenedda39782009-11-13 00:29:53 +00001396}
1397
1398static bool isFrameStoreOpcode(int Opcode) {
1399 switch (Opcode) {
1400 default: break;
1401 case X86::MOV8mr:
1402 case X86::MOV16mr:
1403 case X86::MOV32mr:
1404 case X86::MOV64mr:
1405 case X86::ST_FpP64m:
1406 case X86::MOVSSmr:
1407 case X86::MOVSDmr:
1408 case X86::MOVAPSmr:
1409 case X86::MOVAPDmr:
1410 case X86::MOVDQAmr:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00001411 case X86::VMOVSSmr:
1412 case X86::VMOVSDmr:
1413 case X86::VMOVAPSmr:
1414 case X86::VMOVAPDmr:
1415 case X86::VMOVDQAmr:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00001416 case X86::VMOVAPSYmr:
1417 case X86::VMOVAPDYmr:
1418 case X86::VMOVDQAYmr:
David Greenedda39782009-11-13 00:29:53 +00001419 case X86::MMX_MOVD64mr:
1420 case X86::MMX_MOVQ64mr:
1421 case X86::MMX_MOVNTQmr:
1422 return true;
1423 }
1424 return false;
1425}
1426
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001427unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
David Greenedda39782009-11-13 00:29:53 +00001428 int &FrameIndex) const {
1429 if (isFrameLoadOpcode(MI->getOpcode()))
Jakob Stoklund Olesen81c7b192010-07-27 04:17:01 +00001430 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex))
Chris Lattner40839602006-02-02 20:12:32 +00001431 return MI->getOperand(0).getReg();
David Greenedda39782009-11-13 00:29:53 +00001432 return 0;
1433}
1434
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001435unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
David Greenedda39782009-11-13 00:29:53 +00001436 int &FrameIndex) const {
1437 if (isFrameLoadOpcode(MI->getOpcode())) {
1438 unsigned Reg;
1439 if ((Reg = isLoadFromStackSlot(MI, FrameIndex)))
1440 return Reg;
David Greeneb87bc952009-11-12 20:55:29 +00001441 // Check for post-frame index elimination operations
David Greene29dbf502009-12-04 22:38:46 +00001442 const MachineMemOperand *Dummy;
1443 return hasLoadFromStackSlot(MI, Dummy, FrameIndex);
Chris Lattner40839602006-02-02 20:12:32 +00001444 }
1445 return 0;
1446}
1447
Dan Gohmancbad42c2008-11-18 19:49:32 +00001448unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI,
Chris Lattner40839602006-02-02 20:12:32 +00001449 int &FrameIndex) const {
David Greenedda39782009-11-13 00:29:53 +00001450 if (isFrameStoreOpcode(MI->getOpcode()))
Jakob Stoklund Olesen81c7b192010-07-27 04:17:01 +00001451 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 &&
1452 isFrameOperand(MI, 0, FrameIndex))
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00001453 return MI->getOperand(X86::AddrNumOperands).getReg();
David Greenedda39782009-11-13 00:29:53 +00001454 return 0;
1455}
1456
1457unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI,
1458 int &FrameIndex) const {
1459 if (isFrameStoreOpcode(MI->getOpcode())) {
1460 unsigned Reg;
1461 if ((Reg = isStoreToStackSlot(MI, FrameIndex)))
1462 return Reg;
David Greeneb87bc952009-11-12 20:55:29 +00001463 // Check for post-frame index elimination operations
David Greene29dbf502009-12-04 22:38:46 +00001464 const MachineMemOperand *Dummy;
1465 return hasStoreToStackSlot(MI, Dummy, FrameIndex);
Chris Lattner40839602006-02-02 20:12:32 +00001466 }
1467 return 0;
1468}
1469
Evan Chenge3d8dbf2008-03-27 01:45:11 +00001470/// regIsPICBase - Return true if register is PIC base (i.e.g defined by
1471/// X86::MOVPC32r.
Dan Gohman8e5f2c62008-07-07 23:14:23 +00001472static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) {
Evan Chenge3d8dbf2008-03-27 01:45:11 +00001473 bool isPICBase = false;
1474 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1475 E = MRI.def_end(); I != E; ++I) {
1476 MachineInstr *DefMI = I.getOperand().getParent();
1477 if (DefMI->getOpcode() != X86::MOVPC32r)
1478 return false;
1479 assert(!isPICBase && "More than one PIC base?");
1480 isPICBase = true;
1481 }
1482 return isPICBase;
1483}
Evan Cheng9d15abe2008-03-31 07:54:19 +00001484
Bill Wendling9f8fea32008-05-12 20:54:26 +00001485bool
Dan Gohman3731bc02009-10-10 00:34:18 +00001486X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
1487 AliasAnalysis *AA) const {
Dan Gohmanc101e952007-06-14 20:50:44 +00001488 switch (MI->getOpcode()) {
1489 default: break;
Evan Chenge771ebd2008-03-27 01:41:09 +00001490 case X86::MOV8rm:
1491 case X86::MOV16rm:
Evan Chenge771ebd2008-03-27 01:41:09 +00001492 case X86::MOV32rm:
Evan Chenge771ebd2008-03-27 01:41:09 +00001493 case X86::MOV64rm:
1494 case X86::LD_Fp64m:
1495 case X86::MOVSSrm:
1496 case X86::MOVSDrm:
1497 case X86::MOVAPSrm:
Evan Cheng600c0432009-11-16 21:56:03 +00001498 case X86::MOVUPSrm:
Evan Chenge771ebd2008-03-27 01:41:09 +00001499 case X86::MOVAPDrm:
Dan Gohman54462742009-01-09 02:40:34 +00001500 case X86::MOVDQArm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00001501 case X86::VMOVSSrm:
1502 case X86::VMOVSDrm:
1503 case X86::VMOVAPSrm:
1504 case X86::VMOVUPSrm:
1505 case X86::VMOVAPDrm:
1506 case X86::VMOVDQArm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00001507 case X86::VMOVAPSYrm:
1508 case X86::VMOVUPSYrm:
1509 case X86::VMOVAPDYrm:
1510 case X86::VMOVDQAYrm:
Evan Chenge771ebd2008-03-27 01:41:09 +00001511 case X86::MMX_MOVD64rm:
Evan Chengd15ac2f2009-11-17 09:51:18 +00001512 case X86::MMX_MOVQ64rm:
Bruno Cardoso Lopes0e59a042011-09-03 00:46:45 +00001513 case X86::FsVMOVAPSrm:
1514 case X86::FsVMOVAPDrm:
Evan Chengd15ac2f2009-11-17 09:51:18 +00001515 case X86::FsMOVAPSrm:
1516 case X86::FsMOVAPDrm: {
Evan Chenge771ebd2008-03-27 01:41:09 +00001517 // Loads from constant pools are trivially rematerializable.
Dan Gohmand735b802008-10-03 15:45:36 +00001518 if (MI->getOperand(1).isReg() &&
1519 MI->getOperand(2).isImm() &&
1520 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
Dan Gohman3731bc02009-10-10 00:34:18 +00001521 MI->isInvariantLoad(AA)) {
Evan Chenge771ebd2008-03-27 01:41:09 +00001522 unsigned BaseReg = MI->getOperand(1).getReg();
Chris Lattner18c59872009-06-27 04:16:01 +00001523 if (BaseReg == 0 || BaseReg == X86::RIP)
Evan Chenge771ebd2008-03-27 01:41:09 +00001524 return true;
1525 // Allow re-materialization of PIC load.
Dan Gohmand735b802008-10-03 15:45:36 +00001526 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal())
Evan Chengffe2eb02008-04-01 23:26:12 +00001527 return false;
Dan Gohman8e5f2c62008-07-07 23:14:23 +00001528 const MachineFunction &MF = *MI->getParent()->getParent();
1529 const MachineRegisterInfo &MRI = MF.getRegInfo();
Evan Chenge771ebd2008-03-27 01:41:09 +00001530 bool isPICBase = false;
1531 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg),
1532 E = MRI.def_end(); I != E; ++I) {
1533 MachineInstr *DefMI = I.getOperand().getParent();
1534 if (DefMI->getOpcode() != X86::MOVPC32r)
1535 return false;
1536 assert(!isPICBase && "More than one PIC base?");
1537 isPICBase = true;
1538 }
1539 return isPICBase;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001540 }
Evan Chenge771ebd2008-03-27 01:41:09 +00001541 return false;
Evan Chengd8850a52008-02-22 09:25:47 +00001542 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001543
Evan Chenge771ebd2008-03-27 01:41:09 +00001544 case X86::LEA32r:
1545 case X86::LEA64r: {
Dan Gohmand735b802008-10-03 15:45:36 +00001546 if (MI->getOperand(2).isImm() &&
1547 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 &&
1548 !MI->getOperand(4).isReg()) {
Evan Chenge771ebd2008-03-27 01:41:09 +00001549 // lea fi#, lea GV, etc. are all rematerializable.
Dan Gohmand735b802008-10-03 15:45:36 +00001550 if (!MI->getOperand(1).isReg())
Dan Gohman83ccd142008-09-26 21:30:20 +00001551 return true;
Evan Chenge771ebd2008-03-27 01:41:09 +00001552 unsigned BaseReg = MI->getOperand(1).getReg();
1553 if (BaseReg == 0)
1554 return true;
1555 // Allow re-materialization of lea PICBase + x.
Dan Gohman8e5f2c62008-07-07 23:14:23 +00001556 const MachineFunction &MF = *MI->getParent()->getParent();
1557 const MachineRegisterInfo &MRI = MF.getRegInfo();
Evan Chenge3d8dbf2008-03-27 01:45:11 +00001558 return regIsPICBase(BaseReg, MRI);
Evan Chenge771ebd2008-03-27 01:41:09 +00001559 }
1560 return false;
1561 }
Dan Gohmanc101e952007-06-14 20:50:44 +00001562 }
Evan Chenge771ebd2008-03-27 01:41:09 +00001563
Dan Gohmand45eddd2007-06-26 00:48:07 +00001564 // All other instructions marked M_REMATERIALIZABLE are always trivially
1565 // rematerializable.
1566 return true;
Dan Gohmanc101e952007-06-14 20:50:44 +00001567}
1568
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001569/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that
1570/// would clobber the EFLAGS condition register. Note the result may be
1571/// conservative. If it cannot definitely determine the safety after visiting
Dan Gohman1b1764b2009-10-14 00:08:59 +00001572/// a few instructions in each direction it assumes it's not safe.
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001573static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB,
1574 MachineBasicBlock::iterator I) {
Evan Cheng8d1f0dd2010-03-23 20:35:45 +00001575 MachineBasicBlock::iterator E = MBB.end();
1576
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001577 // For compile time consideration, if we are not able to determine the
Dan Gohman1b1764b2009-10-14 00:08:59 +00001578 // safety after visiting 4 instructions in each direction, we will assume
1579 // it's not safe.
1580 MachineBasicBlock::iterator Iter = I;
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +00001581 for (unsigned i = 0; Iter != E && i < 4; ++i) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001582 bool SeenDef = false;
Dan Gohman1b1764b2009-10-14 00:08:59 +00001583 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1584 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen450b3852012-02-09 00:17:22 +00001585 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1586 SeenDef = true;
Dan Gohmand735b802008-10-03 15:45:36 +00001587 if (!MO.isReg())
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001588 continue;
1589 if (MO.getReg() == X86::EFLAGS) {
1590 if (MO.isUse())
1591 return false;
1592 SeenDef = true;
1593 }
1594 }
1595
1596 if (SeenDef)
1597 // This instruction defines EFLAGS, no need to look any further.
1598 return true;
Dan Gohman1b1764b2009-10-14 00:08:59 +00001599 ++Iter;
Evan Cheng8d1f0dd2010-03-23 20:35:45 +00001600 // Skip over DBG_VALUE.
1601 while (Iter != E && Iter->isDebugValue())
1602 ++Iter;
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +00001603 }
Dan Gohman3afda6e2008-10-21 03:24:31 +00001604
Jakob Stoklund Olesenb8e052e2011-09-02 23:52:52 +00001605 // It is safe to clobber EFLAGS at the end of a block of no successor has it
1606 // live in.
1607 if (Iter == E) {
1608 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
1609 SE = MBB.succ_end(); SI != SE; ++SI)
1610 if ((*SI)->isLiveIn(X86::EFLAGS))
1611 return false;
1612 return true;
Dan Gohman1b1764b2009-10-14 00:08:59 +00001613 }
1614
Evan Cheng8d1f0dd2010-03-23 20:35:45 +00001615 MachineBasicBlock::iterator B = MBB.begin();
Dan Gohman1b1764b2009-10-14 00:08:59 +00001616 Iter = I;
1617 for (unsigned i = 0; i < 4; ++i) {
1618 // If we make it to the beginning of the block, it's safe to clobber
1619 // EFLAGS iff EFLAGS is not live-in.
Evan Cheng8d1f0dd2010-03-23 20:35:45 +00001620 if (Iter == B)
Dan Gohman1b1764b2009-10-14 00:08:59 +00001621 return !MBB.isLiveIn(X86::EFLAGS);
1622
1623 --Iter;
Evan Cheng8d1f0dd2010-03-23 20:35:45 +00001624 // Skip over DBG_VALUE.
1625 while (Iter != B && Iter->isDebugValue())
1626 --Iter;
1627
Dan Gohman1b1764b2009-10-14 00:08:59 +00001628 bool SawKill = false;
1629 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) {
1630 MachineOperand &MO = Iter->getOperand(j);
Jakob Stoklund Olesen450b3852012-02-09 00:17:22 +00001631 // A register mask may clobber EFLAGS, but we should still look for a
1632 // live EFLAGS def.
1633 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS))
1634 SawKill = true;
Dan Gohman1b1764b2009-10-14 00:08:59 +00001635 if (MO.isReg() && MO.getReg() == X86::EFLAGS) {
1636 if (MO.isDef()) return MO.isDead();
1637 if (MO.isKill()) SawKill = true;
1638 }
1639 }
1640
1641 if (SawKill)
1642 // This instruction kills EFLAGS and doesn't redefine it, so
1643 // there's no need to look further.
Dan Gohman3afda6e2008-10-21 03:24:31 +00001644 return true;
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001645 }
1646
1647 // Conservative answer.
1648 return false;
1649}
1650
Evan Chengca1267c2008-03-31 20:40:39 +00001651void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB,
1652 MachineBasicBlock::iterator I,
Evan Cheng37844532009-07-16 09:20:10 +00001653 unsigned DestReg, unsigned SubIdx,
Evan Chengd57cdd52009-11-14 02:55:43 +00001654 const MachineInstr *Orig,
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +00001655 const TargetRegisterInfo &TRI) const {
Dan Gohman0d881042010-05-07 01:28:10 +00001656 DebugLoc DL = Orig->getDebugLoc();
Bill Wendlingfbef3102009-02-11 21:51:19 +00001657
Evan Chengca1267c2008-03-31 20:40:39 +00001658 // MOV32r0 etc. are implemented with xor which clobbers condition code.
1659 // Re-materialize them as movri instructions to avoid side effects.
Evan Cheng37844532009-07-16 09:20:10 +00001660 bool Clone = true;
1661 unsigned Opc = Orig->getOpcode();
1662 switch (Opc) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001663 default: break;
Evan Chengca1267c2008-03-31 20:40:39 +00001664 case X86::MOV8r0:
Dan Gohmanf1b4d262010-01-12 04:42:54 +00001665 case X86::MOV16r0:
1666 case X86::MOV32r0:
1667 case X86::MOV64r0: {
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001668 if (!isSafeToClobberEFLAGS(MBB, I)) {
Evan Cheng37844532009-07-16 09:20:10 +00001669 switch (Opc) {
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001670 default: break;
1671 case X86::MOV8r0: Opc = X86::MOV8ri; break;
Dan Gohmanf1b4d262010-01-12 04:42:54 +00001672 case X86::MOV16r0: Opc = X86::MOV16ri; break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001673 case X86::MOV32r0: Opc = X86::MOV32ri; break;
Dan Gohman6fe0df22010-02-26 16:49:27 +00001674 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001675 }
Evan Cheng37844532009-07-16 09:20:10 +00001676 Clone = false;
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001677 }
Evan Chengca1267c2008-03-31 20:40:39 +00001678 break;
Evan Cheng9ef4ca22008-06-24 07:10:51 +00001679 }
1680 }
1681
Evan Cheng37844532009-07-16 09:20:10 +00001682 if (Clone) {
Dan Gohman8e5f2c62008-07-07 23:14:23 +00001683 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
Evan Chengca1267c2008-03-31 20:40:39 +00001684 MBB.insert(I, MI);
Evan Cheng37844532009-07-16 09:20:10 +00001685 } else {
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +00001686 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0);
Evan Chengca1267c2008-03-31 20:40:39 +00001687 }
Evan Cheng03eb3882008-04-16 23:44:44 +00001688
Evan Cheng37844532009-07-16 09:20:10 +00001689 MachineInstr *NewMI = prior(I);
Jakob Stoklund Olesen9edf7de2010-06-02 22:47:25 +00001690 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI);
Evan Chengca1267c2008-03-31 20:40:39 +00001691}
1692
Evan Cheng3f411c72007-10-05 08:04:01 +00001693/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
1694/// is not marked dead.
1695static bool hasLiveCondCodeDef(MachineInstr *MI) {
Evan Cheng3f411c72007-10-05 08:04:01 +00001696 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1697 MachineOperand &MO = MI->getOperand(i);
Dan Gohmand735b802008-10-03 15:45:36 +00001698 if (MO.isReg() && MO.isDef() &&
Evan Cheng3f411c72007-10-05 08:04:01 +00001699 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
1700 return true;
1701 }
1702 }
1703 return false;
1704}
1705
Evan Chengdd99f3a2009-12-12 20:03:14 +00001706/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when
Evan Cheng656e5142009-12-11 06:01:48 +00001707/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting
1708/// to a 32-bit superregister and then truncating back down to a 16-bit
1709/// subregister.
1710MachineInstr *
1711X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc,
1712 MachineFunction::iterator &MFI,
1713 MachineBasicBlock::iterator &MBBI,
1714 LiveVariables *LV) const {
1715 MachineInstr *MI = MBBI;
1716 unsigned Dest = MI->getOperand(0).getReg();
1717 unsigned Src = MI->getOperand(1).getReg();
1718 bool isDead = MI->getOperand(0).isDead();
1719 bool isKill = MI->getOperand(1).isKill();
1720
1721 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
1722 ? X86::LEA64_32r : X86::LEA32r;
1723 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001724 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng656e5142009-12-11 06:01:48 +00001725 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001726
Evan Cheng656e5142009-12-11 06:01:48 +00001727 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001728 // well be shifting and then extracting the lower 16-bits.
Evan Chengdd99f3a2009-12-12 20:03:14 +00001729 // This has the potential to cause partial register stall. e.g.
Evan Cheng04ab19c2009-12-12 18:55:26 +00001730 // movw (%rbp,%rcx,2), %dx
1731 // leal -65(%rdx), %esi
Evan Chengdd99f3a2009-12-12 20:03:14 +00001732 // But testing has shown this *does* help performance in 64-bit mode (at
1733 // least on modern x86 machines).
Evan Cheng656e5142009-12-11 06:01:48 +00001734 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg);
1735 MachineInstr *InsMI =
Jakob Stoklund Olesen5c00e072010-07-08 16:40:15 +00001736 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
1737 .addReg(leaInReg, RegState::Define, X86::sub_16bit)
1738 .addReg(Src, getKillRegState(isKill));
Evan Cheng656e5142009-12-11 06:01:48 +00001739
1740 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(),
1741 get(Opc), leaOutReg);
1742 switch (MIOpc) {
1743 default:
1744 llvm_unreachable(0);
Evan Cheng656e5142009-12-11 06:01:48 +00001745 case X86::SHL16ri: {
1746 unsigned ShAmt = MI->getOperand(2).getImm();
1747 MIB.addReg(0).addImm(1 << ShAmt)
Chris Lattner599b5312010-07-08 23:46:44 +00001748 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0);
Evan Cheng656e5142009-12-11 06:01:48 +00001749 break;
1750 }
1751 case X86::INC16r:
1752 case X86::INC64_16r:
Chris Lattner599b5312010-07-08 23:46:44 +00001753 addRegOffset(MIB, leaInReg, true, 1);
Evan Cheng656e5142009-12-11 06:01:48 +00001754 break;
1755 case X86::DEC16r:
1756 case X86::DEC64_16r:
Chris Lattner599b5312010-07-08 23:46:44 +00001757 addRegOffset(MIB, leaInReg, true, -1);
Evan Cheng656e5142009-12-11 06:01:48 +00001758 break;
1759 case X86::ADD16ri:
1760 case X86::ADD16ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +00001761 case X86::ADD16ri_DB:
1762 case X86::ADD16ri8_DB:
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001763 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm());
Evan Cheng656e5142009-12-11 06:01:48 +00001764 break;
Chris Lattner99ae6652010-10-08 03:54:52 +00001765 case X86::ADD16rr:
1766 case X86::ADD16rr_DB: {
Evan Cheng656e5142009-12-11 06:01:48 +00001767 unsigned Src2 = MI->getOperand(2).getReg();
1768 bool isKill2 = MI->getOperand(2).isKill();
1769 unsigned leaInReg2 = 0;
1770 MachineInstr *InsMI2 = 0;
1771 if (Src == Src2) {
1772 // ADD16rr %reg1028<kill>, %reg1028
1773 // just a single insert_subreg.
1774 addRegReg(MIB, leaInReg, true, leaInReg, false);
1775 } else {
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001776 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Evan Cheng656e5142009-12-11 06:01:48 +00001777 // Build and insert into an implicit UNDEF value. This is OK because
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001778 // well be shifting and then extracting the lower 16-bits.
Evan Chengddfd1372011-12-14 02:11:42 +00001779 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2);
Evan Cheng656e5142009-12-11 06:01:48 +00001780 InsMI2 =
Evan Chengddfd1372011-12-14 02:11:42 +00001781 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY))
Jakob Stoklund Olesen5c00e072010-07-08 16:40:15 +00001782 .addReg(leaInReg2, RegState::Define, X86::sub_16bit)
1783 .addReg(Src2, getKillRegState(isKill2));
Evan Cheng656e5142009-12-11 06:01:48 +00001784 addRegReg(MIB, leaInReg, true, leaInReg2, true);
1785 }
1786 if (LV && isKill2 && InsMI2)
1787 LV->replaceKillInstruction(Src2, MI, InsMI2);
1788 break;
1789 }
1790 }
1791
1792 MachineInstr *NewMI = MIB;
1793 MachineInstr *ExtMI =
Jakob Stoklund Olesen0bc25f42010-07-08 16:40:22 +00001794 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY))
Evan Cheng656e5142009-12-11 06:01:48 +00001795 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Jakob Stoklund Olesen0bc25f42010-07-08 16:40:22 +00001796 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit);
Evan Cheng656e5142009-12-11 06:01:48 +00001797
1798 if (LV) {
1799 // Update live variables
1800 LV->getVarInfo(leaInReg).Kills.push_back(NewMI);
1801 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI);
1802 if (isKill)
1803 LV->replaceKillInstruction(Src, MI, InsMI);
1804 if (isDead)
1805 LV->replaceKillInstruction(Dest, MI, ExtMI);
1806 }
1807
1808 return ExtMI;
1809}
1810
Chris Lattnerbcea4d62005-01-02 02:37:07 +00001811/// convertToThreeAddress - This method must be implemented by targets that
1812/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
1813/// may be able to convert a two-address instruction into a true
1814/// three-address instruction on demand. This allows the X86 target (for
1815/// example) to convert ADD and SHL instructions into LEA instructions if they
1816/// would require register copies due to two-addressness.
1817///
1818/// This method returns a null pointer if the transformation cannot be
1819/// performed, otherwise it returns the new instruction.
1820///
Evan Cheng258ff672006-12-01 21:52:41 +00001821MachineInstr *
1822X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
1823 MachineBasicBlock::iterator &MBBI,
Owen Andersonf660c172008-07-02 23:41:07 +00001824 LiveVariables *LV) const {
Evan Cheng258ff672006-12-01 21:52:41 +00001825 MachineInstr *MI = MBBI;
Dan Gohman8e5f2c62008-07-07 23:14:23 +00001826 MachineFunction &MF = *MI->getParent()->getParent();
Chris Lattnerbcea4d62005-01-02 02:37:07 +00001827 // All instructions input are two-addr instructions. Get the known operands.
1828 unsigned Dest = MI->getOperand(0).getReg();
1829 unsigned Src = MI->getOperand(1).getReg();
Evan Cheng9f1c8312008-07-03 09:09:37 +00001830 bool isDead = MI->getOperand(0).isDead();
1831 bool isKill = MI->getOperand(1).isKill();
Chris Lattnerbcea4d62005-01-02 02:37:07 +00001832
Evan Cheng6ce7dc22006-11-15 20:58:11 +00001833 MachineInstr *NewMI = NULL;
Evan Cheng258ff672006-12-01 21:52:41 +00001834 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001835 // we have better subtarget support, enable the 16-bit LEA generation here.
Evan Chengdd99f3a2009-12-12 20:03:14 +00001836 // 16-bit LEA is also slow on Core2.
Evan Cheng258ff672006-12-01 21:52:41 +00001837 bool DisableLEA16 = true;
Evan Chengdd99f3a2009-12-12 20:03:14 +00001838 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
Evan Cheng258ff672006-12-01 21:52:41 +00001839
Evan Cheng559dc462007-10-05 20:34:26 +00001840 unsigned MIOpc = MI->getOpcode();
1841 switch (MIOpc) {
Evan Chengccba76b2006-05-30 20:26:50 +00001842 case X86::SHUFPSrri: {
1843 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001844 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00001845
Evan Chengaa3c1412006-05-30 21:45:53 +00001846 unsigned B = MI->getOperand(1).getReg();
1847 unsigned C = MI->getOperand(2).getReg();
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001848 if (B != C) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +00001849 unsigned A = MI->getOperand(0).getReg();
1850 unsigned M = MI->getOperand(3).getImm();
Bill Wendlingfbef3102009-02-11 21:51:19 +00001851 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
Bill Wendling587daed2009-05-13 21:33:08 +00001852 .addReg(A, RegState::Define | getDeadRegState(isDead))
1853 .addReg(B, getKillRegState(isKill)).addImm(M);
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001854 break;
1855 }
Craig Topper05189702012-01-13 09:21:41 +00001856 case X86::SHUFPDrri: {
1857 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!");
1858 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
1859
1860 unsigned B = MI->getOperand(1).getReg();
1861 unsigned C = MI->getOperand(2).getReg();
1862 if (B != C) return 0;
1863 unsigned A = MI->getOperand(0).getReg();
1864 unsigned M = MI->getOperand(3).getImm();
1865
1866 // Convert to PSHUFD mask.
1867 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44;
1868
1869 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri))
1870 .addReg(A, RegState::Define | getDeadRegState(isDead))
1871 .addReg(B, getKillRegState(isKill)).addImm(M);
1872 break;
1873 }
Chris Lattner995f5502007-03-28 18:12:31 +00001874 case X86::SHL64ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +00001875 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattner995f5502007-03-28 18:12:31 +00001876 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1877 // the flags produced by a shift yet, so this is safe.
Chris Lattner995f5502007-03-28 18:12:31 +00001878 unsigned ShAmt = MI->getOperand(2).getImm();
1879 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +00001880
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001881 // LEA can't handle RSP.
1882 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1883 !MF.getRegInfo().constrainRegClass(Src, &X86::GR64_NOSPRegClass))
1884 return 0;
1885
Bill Wendlingfbef3102009-02-11 21:51:19 +00001886 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Bill Wendling587daed2009-05-13 21:33:08 +00001887 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1888 .addReg(0).addImm(1 << ShAmt)
1889 .addReg(Src, getKillRegState(isKill))
Chris Lattner599b5312010-07-08 23:46:44 +00001890 .addImm(0).addReg(0);
Chris Lattner995f5502007-03-28 18:12:31 +00001891 break;
1892 }
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001893 case X86::SHL32ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +00001894 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001895 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1896 // the flags produced by a shift yet, so this is safe.
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001897 unsigned ShAmt = MI->getOperand(2).getImm();
1898 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +00001899
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001900 // LEA can't handle ESP.
1901 if (TargetRegisterInfo::isVirtualRegister(Src) &&
1902 !MF.getRegInfo().constrainRegClass(Src, &X86::GR32_NOSPRegClass))
1903 return 0;
1904
Evan Chengdd99f3a2009-12-12 20:03:14 +00001905 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Bill Wendlingfbef3102009-02-11 21:51:19 +00001906 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +00001907 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
Evan Cheng9f1c8312008-07-03 09:09:37 +00001908 .addReg(0).addImm(1 << ShAmt)
Chris Lattner599b5312010-07-08 23:46:44 +00001909 .addReg(Src, getKillRegState(isKill)).addImm(0).addReg(0);
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001910 break;
1911 }
1912 case X86::SHL16ri: {
Evan Cheng24f2ea32007-09-14 21:48:26 +00001913 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
Evan Cheng61d9c862007-09-06 00:14:41 +00001914 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
1915 // the flags produced by a shift yet, so this is safe.
Evan Cheng61d9c862007-09-06 00:14:41 +00001916 unsigned ShAmt = MI->getOperand(2).getImm();
1917 if (ShAmt == 0 || ShAmt >= 4) return 0;
Evan Cheng9f1c8312008-07-03 09:09:37 +00001918
Evan Cheng656e5142009-12-11 06:01:48 +00001919 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +00001920 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng656e5142009-12-11 06:01:48 +00001921 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
1922 .addReg(Dest, RegState::Define | getDeadRegState(isDead))
1923 .addReg(0).addImm(1 << ShAmt)
1924 .addReg(Src, getKillRegState(isKill))
Chris Lattner599b5312010-07-08 23:46:44 +00001925 .addImm(0).addReg(0);
Chris Lattnera16b7cb2007-03-20 06:08:29 +00001926 break;
Evan Chengccba76b2006-05-30 20:26:50 +00001927 }
Evan Cheng559dc462007-10-05 20:34:26 +00001928 default: {
1929 // The following opcodes also sets the condition code register(s). Only
1930 // convert them to equivalent lea if the condition code register def's
1931 // are dead!
1932 if (hasLiveCondCodeDef(MI))
1933 return 0;
Evan Chengccba76b2006-05-30 20:26:50 +00001934
Evan Cheng559dc462007-10-05 20:34:26 +00001935 switch (MIOpc) {
1936 default: return 0;
1937 case X86::INC64r:
Dan Gohmancca29832009-01-06 23:34:46 +00001938 case X86::INC32r:
1939 case X86::INC64_32r: {
Evan Cheng559dc462007-10-05 20:34:26 +00001940 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Evan Chengb76143c2007-10-09 07:14:53 +00001941 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
1942 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperc9099502012-04-20 06:31:50 +00001943 const TargetRegisterClass *RC = MIOpc == X86::INC64r ?
1944 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1945 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001946
1947 // LEA can't handle RSP.
1948 if (TargetRegisterInfo::isVirtualRegister(Src) &&
Craig Topperc9099502012-04-20 06:31:50 +00001949 !MF.getRegInfo().constrainRegClass(Src, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001950 return 0;
1951
Chris Lattner599b5312010-07-08 23:46:44 +00001952 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +00001953 .addReg(Dest, RegState::Define |
1954 getDeadRegState(isDead)),
Rafael Espindola094fad32009-04-08 21:14:34 +00001955 Src, isKill, 1);
Evan Cheng559dc462007-10-05 20:34:26 +00001956 break;
Chris Lattnerbcea4d62005-01-02 02:37:07 +00001957 }
Evan Cheng559dc462007-10-05 20:34:26 +00001958 case X86::INC16r:
1959 case X86::INC64_16r:
Evan Cheng656e5142009-12-11 06:01:48 +00001960 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +00001961 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +00001962 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
Bill Wendlingfbef3102009-02-11 21:51:19 +00001963 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendling587daed2009-05-13 21:33:08 +00001964 .addReg(Dest, RegState::Define |
1965 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +00001966 Src, isKill, 1);
Evan Cheng559dc462007-10-05 20:34:26 +00001967 break;
1968 case X86::DEC64r:
Dan Gohmancca29832009-01-06 23:34:46 +00001969 case X86::DEC32r:
1970 case X86::DEC64_32r: {
Evan Cheng559dc462007-10-05 20:34:26 +00001971 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Evan Chengb76143c2007-10-09 07:14:53 +00001972 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
1973 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
Craig Topperc9099502012-04-20 06:31:50 +00001974 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ?
1975 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass :
1976 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass;
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001977 // LEA can't handle RSP.
1978 if (TargetRegisterInfo::isVirtualRegister(Src) &&
Craig Topperc9099502012-04-20 06:31:50 +00001979 !MF.getRegInfo().constrainRegClass(Src, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00001980 return 0;
1981
Chris Lattner599b5312010-07-08 23:46:44 +00001982 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +00001983 .addReg(Dest, RegState::Define |
1984 getDeadRegState(isDead)),
Rafael Espindola094fad32009-04-08 21:14:34 +00001985 Src, isKill, -1);
Evan Cheng559dc462007-10-05 20:34:26 +00001986 break;
1987 }
1988 case X86::DEC16r:
1989 case X86::DEC64_16r:
Evan Cheng656e5142009-12-11 06:01:48 +00001990 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +00001991 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +00001992 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
Bill Wendlingfbef3102009-02-11 21:51:19 +00001993 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendling587daed2009-05-13 21:33:08 +00001994 .addReg(Dest, RegState::Define |
1995 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +00001996 Src, isKill, -1);
Evan Cheng559dc462007-10-05 20:34:26 +00001997 break;
1998 case X86::ADD64rr:
Chris Lattner99ae6652010-10-08 03:54:52 +00001999 case X86::ADD64rr_DB:
2000 case X86::ADD32rr:
2001 case X86::ADD32rr_DB: {
Evan Cheng559dc462007-10-05 20:34:26 +00002002 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner99ae6652010-10-08 03:54:52 +00002003 unsigned Opc;
Craig Topper44d23822012-02-22 05:59:10 +00002004 const TargetRegisterClass *RC;
Chris Lattner99ae6652010-10-08 03:54:52 +00002005 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) {
2006 Opc = X86::LEA64r;
Craig Topperc9099502012-04-20 06:31:50 +00002007 RC = &X86::GR64_NOSPRegClass;
Chris Lattner99ae6652010-10-08 03:54:52 +00002008 } else {
2009 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Craig Topperc9099502012-04-20 06:31:50 +00002010 RC = &X86::GR32_NOSPRegClass;
Chris Lattner99ae6652010-10-08 03:54:52 +00002011 }
2012
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002013
Evan Cheng9f1c8312008-07-03 09:09:37 +00002014 unsigned Src2 = MI->getOperand(2).getReg();
2015 bool isKill2 = MI->getOperand(2).isKill();
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00002016
2017 // LEA can't handle RSP.
2018 if (TargetRegisterInfo::isVirtualRegister(Src2) &&
Chris Lattner99ae6652010-10-08 03:54:52 +00002019 !MF.getRegInfo().constrainRegClass(Src2, RC))
Jakob Stoklund Olesen635127a2010-10-07 00:07:26 +00002020 return 0;
2021
Bill Wendlingfbef3102009-02-11 21:51:19 +00002022 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Bill Wendling587daed2009-05-13 21:33:08 +00002023 .addReg(Dest, RegState::Define |
2024 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +00002025 Src, isKill, Src2, isKill2);
2026 if (LV && isKill2)
2027 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Cheng559dc462007-10-05 20:34:26 +00002028 break;
2029 }
Chris Lattner99ae6652010-10-08 03:54:52 +00002030 case X86::ADD16rr:
2031 case X86::ADD16rr_DB: {
Evan Cheng656e5142009-12-11 06:01:48 +00002032 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +00002033 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng559dc462007-10-05 20:34:26 +00002034 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng9f1c8312008-07-03 09:09:37 +00002035 unsigned Src2 = MI->getOperand(2).getReg();
2036 bool isKill2 = MI->getOperand(2).isKill();
Bill Wendlingfbef3102009-02-11 21:51:19 +00002037 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Bill Wendling587daed2009-05-13 21:33:08 +00002038 .addReg(Dest, RegState::Define |
2039 getDeadRegState(isDead)),
Evan Cheng9f1c8312008-07-03 09:09:37 +00002040 Src, isKill, Src2, isKill2);
2041 if (LV && isKill2)
2042 LV->replaceKillInstruction(Src2, MI, NewMI);
Evan Cheng559dc462007-10-05 20:34:26 +00002043 break;
Evan Cheng9f1c8312008-07-03 09:09:37 +00002044 }
Evan Cheng559dc462007-10-05 20:34:26 +00002045 case X86::ADD64ri32:
2046 case X86::ADD64ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +00002047 case X86::ADD64ri32_DB:
2048 case X86::ADD64ri8_DB:
Evan Cheng559dc462007-10-05 20:34:26 +00002049 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner599b5312010-07-08 23:46:44 +00002050 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r))
Evan Cheng656e5142009-12-11 06:01:48 +00002051 .addReg(Dest, RegState::Define |
2052 getDeadRegState(isDead)),
2053 Src, isKill, MI->getOperand(2).getImm());
Evan Cheng559dc462007-10-05 20:34:26 +00002054 break;
2055 case X86::ADD32ri:
Chris Lattner15df55d2010-10-08 03:57:25 +00002056 case X86::ADD32ri8:
2057 case X86::ADD32ri_DB:
2058 case X86::ADD32ri8_DB: {
Evan Cheng559dc462007-10-05 20:34:26 +00002059 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Evan Cheng656e5142009-12-11 06:01:48 +00002060 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
Chris Lattner599b5312010-07-08 23:46:44 +00002061 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc))
Evan Cheng656e5142009-12-11 06:01:48 +00002062 .addReg(Dest, RegState::Define |
2063 getDeadRegState(isDead)),
Rafael Espindola094fad32009-04-08 21:14:34 +00002064 Src, isKill, MI->getOperand(2).getImm());
Evan Cheng559dc462007-10-05 20:34:26 +00002065 break;
2066 }
Evan Cheng656e5142009-12-11 06:01:48 +00002067 case X86::ADD16ri:
2068 case X86::ADD16ri8:
Chris Lattner15df55d2010-10-08 03:57:25 +00002069 case X86::ADD16ri_DB:
2070 case X86::ADD16ri8_DB:
Evan Cheng656e5142009-12-11 06:01:48 +00002071 if (DisableLEA16)
Evan Chengdd99f3a2009-12-12 20:03:14 +00002072 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0;
Evan Cheng656e5142009-12-11 06:01:48 +00002073 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
Chris Lattner599b5312010-07-08 23:46:44 +00002074 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r))
Evan Cheng656e5142009-12-11 06:01:48 +00002075 .addReg(Dest, RegState::Define |
2076 getDeadRegState(isDead)),
2077 Src, isKill, MI->getOperand(2).getImm());
2078 break;
Evan Cheng559dc462007-10-05 20:34:26 +00002079 }
2080 }
Chris Lattnerbcea4d62005-01-02 02:37:07 +00002081 }
2082
Evan Cheng15246732008-02-07 08:29:53 +00002083 if (!NewMI) return 0;
2084
Evan Cheng9f1c8312008-07-03 09:09:37 +00002085 if (LV) { // Update live variables
2086 if (isKill)
2087 LV->replaceKillInstruction(Src, MI, NewMI);
2088 if (isDead)
2089 LV->replaceKillInstruction(Dest, MI, NewMI);
2090 }
2091
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002092 MFI->insert(MBBI, NewMI); // Insert the new inst
Evan Cheng6ce7dc22006-11-15 20:58:11 +00002093 return NewMI;
Chris Lattnerbcea4d62005-01-02 02:37:07 +00002094}
2095
Chris Lattner41e431b2005-01-19 07:11:01 +00002096/// commuteInstruction - We have a few instructions that must be hacked on to
2097/// commute them.
2098///
Evan Cheng58dcb0e2008-06-16 07:33:11 +00002099MachineInstr *
2100X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
Chris Lattner41e431b2005-01-19 07:11:01 +00002101 switch (MI->getOpcode()) {
Chris Lattner0df53d22005-01-19 07:31:24 +00002102 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
2103 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
Chris Lattner41e431b2005-01-19 07:11:01 +00002104 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
Dan Gohmane47f1f92007-09-14 23:17:45 +00002105 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
2106 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
2107 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
Chris Lattner0df53d22005-01-19 07:31:24 +00002108 unsigned Opc;
2109 unsigned Size;
2110 switch (MI->getOpcode()) {
Torok Edwinc23197a2009-07-14 16:55:14 +00002111 default: llvm_unreachable("Unreachable!");
Chris Lattner0df53d22005-01-19 07:31:24 +00002112 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
2113 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
2114 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
2115 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
Dan Gohmane47f1f92007-09-14 23:17:45 +00002116 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
2117 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
Chris Lattner0df53d22005-01-19 07:31:24 +00002118 }
Chris Lattner9a1ceae2007-12-30 20:49:49 +00002119 unsigned Amt = MI->getOperand(3).getImm();
Dan Gohman74feef22008-10-17 01:23:35 +00002120 if (NewMI) {
2121 MachineFunction &MF = *MI->getParent()->getParent();
2122 MI = MF.CloneMachineInstr(MI);
2123 NewMI = false;
Evan Chenga4d16a12008-02-13 02:46:49 +00002124 }
Dan Gohman74feef22008-10-17 01:23:35 +00002125 MI->setDesc(get(Opc));
2126 MI->getOperand(3).setImm(Size-Amt);
2127 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner41e431b2005-01-19 07:11:01 +00002128 }
Evan Cheng7ad42d92007-10-05 23:13:21 +00002129 case X86::CMOVB16rr:
2130 case X86::CMOVB32rr:
2131 case X86::CMOVB64rr:
2132 case X86::CMOVAE16rr:
2133 case X86::CMOVAE32rr:
2134 case X86::CMOVAE64rr:
2135 case X86::CMOVE16rr:
2136 case X86::CMOVE32rr:
2137 case X86::CMOVE64rr:
2138 case X86::CMOVNE16rr:
2139 case X86::CMOVNE32rr:
2140 case X86::CMOVNE64rr:
Chris Lattner25cbf502010-10-05 23:00:14 +00002141 case X86::CMOVBE16rr:
2142 case X86::CMOVBE32rr:
2143 case X86::CMOVBE64rr:
Evan Cheng7ad42d92007-10-05 23:13:21 +00002144 case X86::CMOVA16rr:
2145 case X86::CMOVA32rr:
2146 case X86::CMOVA64rr:
2147 case X86::CMOVL16rr:
2148 case X86::CMOVL32rr:
2149 case X86::CMOVL64rr:
2150 case X86::CMOVGE16rr:
2151 case X86::CMOVGE32rr:
2152 case X86::CMOVGE64rr:
2153 case X86::CMOVLE16rr:
2154 case X86::CMOVLE32rr:
2155 case X86::CMOVLE64rr:
2156 case X86::CMOVG16rr:
2157 case X86::CMOVG32rr:
2158 case X86::CMOVG64rr:
2159 case X86::CMOVS16rr:
2160 case X86::CMOVS32rr:
2161 case X86::CMOVS64rr:
2162 case X86::CMOVNS16rr:
2163 case X86::CMOVNS32rr:
2164 case X86::CMOVNS64rr:
2165 case X86::CMOVP16rr:
2166 case X86::CMOVP32rr:
2167 case X86::CMOVP64rr:
2168 case X86::CMOVNP16rr:
2169 case X86::CMOVNP32rr:
Dan Gohman305fceb2009-01-07 00:35:10 +00002170 case X86::CMOVNP64rr:
2171 case X86::CMOVO16rr:
2172 case X86::CMOVO32rr:
2173 case X86::CMOVO64rr:
2174 case X86::CMOVNO16rr:
2175 case X86::CMOVNO32rr:
2176 case X86::CMOVNO64rr: {
Evan Cheng7ad42d92007-10-05 23:13:21 +00002177 unsigned Opc = 0;
2178 switch (MI->getOpcode()) {
2179 default: break;
2180 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
2181 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
2182 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
2183 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
2184 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
2185 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
2186 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
2187 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
2188 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
2189 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
2190 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
2191 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
Chris Lattner25cbf502010-10-05 23:00:14 +00002192 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
2193 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
2194 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
2195 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
2196 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
2197 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +00002198 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
2199 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
2200 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
2201 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
2202 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
2203 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
2204 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
2205 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
2206 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
2207 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
2208 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
2209 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
2210 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
2211 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +00002212 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +00002213 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
2214 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
2215 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
2216 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
2217 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +00002218 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +00002219 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
2220 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
2221 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
Dan Gohman305fceb2009-01-07 00:35:10 +00002222 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break;
2223 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break;
Mon P Wang0bd07fc2009-04-18 05:16:01 +00002224 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break;
Dan Gohman305fceb2009-01-07 00:35:10 +00002225 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break;
2226 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break;
2227 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break;
Evan Cheng7ad42d92007-10-05 23:13:21 +00002228 }
Dan Gohman74feef22008-10-17 01:23:35 +00002229 if (NewMI) {
2230 MachineFunction &MF = *MI->getParent()->getParent();
2231 MI = MF.CloneMachineInstr(MI);
2232 NewMI = false;
2233 }
Chris Lattner5080f4d2008-01-11 18:10:50 +00002234 MI->setDesc(get(Opc));
Evan Cheng7ad42d92007-10-05 23:13:21 +00002235 // Fallthrough intended.
2236 }
Chris Lattner41e431b2005-01-19 07:11:01 +00002237 default:
Evan Cheng58dcb0e2008-06-16 07:33:11 +00002238 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
Chris Lattner41e431b2005-01-19 07:11:01 +00002239 }
2240}
2241
Chris Lattner7fbe9722006-10-20 17:42:20 +00002242static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
2243 switch (BrOpc) {
2244 default: return X86::COND_INVALID;
Chris Lattnerbd13fb62010-02-11 19:25:55 +00002245 case X86::JE_4: return X86::COND_E;
2246 case X86::JNE_4: return X86::COND_NE;
2247 case X86::JL_4: return X86::COND_L;
2248 case X86::JLE_4: return X86::COND_LE;
2249 case X86::JG_4: return X86::COND_G;
2250 case X86::JGE_4: return X86::COND_GE;
2251 case X86::JB_4: return X86::COND_B;
2252 case X86::JBE_4: return X86::COND_BE;
2253 case X86::JA_4: return X86::COND_A;
2254 case X86::JAE_4: return X86::COND_AE;
2255 case X86::JS_4: return X86::COND_S;
2256 case X86::JNS_4: return X86::COND_NS;
2257 case X86::JP_4: return X86::COND_P;
2258 case X86::JNP_4: return X86::COND_NP;
2259 case X86::JO_4: return X86::COND_O;
2260 case X86::JNO_4: return X86::COND_NO;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002261 }
2262}
2263
2264unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
2265 switch (CC) {
Torok Edwinc23197a2009-07-14 16:55:14 +00002266 default: llvm_unreachable("Illegal condition code!");
Chris Lattnerbd13fb62010-02-11 19:25:55 +00002267 case X86::COND_E: return X86::JE_4;
2268 case X86::COND_NE: return X86::JNE_4;
2269 case X86::COND_L: return X86::JL_4;
2270 case X86::COND_LE: return X86::JLE_4;
2271 case X86::COND_G: return X86::JG_4;
2272 case X86::COND_GE: return X86::JGE_4;
2273 case X86::COND_B: return X86::JB_4;
2274 case X86::COND_BE: return X86::JBE_4;
2275 case X86::COND_A: return X86::JA_4;
2276 case X86::COND_AE: return X86::JAE_4;
2277 case X86::COND_S: return X86::JS_4;
2278 case X86::COND_NS: return X86::JNS_4;
2279 case X86::COND_P: return X86::JP_4;
2280 case X86::COND_NP: return X86::JNP_4;
2281 case X86::COND_O: return X86::JO_4;
2282 case X86::COND_NO: return X86::JNO_4;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002283 }
2284}
2285
Chris Lattner9cd68752006-10-21 05:52:40 +00002286/// GetOppositeBranchCondition - Return the inverse of the specified condition,
2287/// e.g. turning COND_E to COND_NE.
2288X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
2289 switch (CC) {
Torok Edwinc23197a2009-07-14 16:55:14 +00002290 default: llvm_unreachable("Illegal condition code!");
Chris Lattner9cd68752006-10-21 05:52:40 +00002291 case X86::COND_E: return X86::COND_NE;
2292 case X86::COND_NE: return X86::COND_E;
2293 case X86::COND_L: return X86::COND_GE;
2294 case X86::COND_LE: return X86::COND_G;
2295 case X86::COND_G: return X86::COND_LE;
2296 case X86::COND_GE: return X86::COND_L;
2297 case X86::COND_B: return X86::COND_AE;
2298 case X86::COND_BE: return X86::COND_A;
2299 case X86::COND_A: return X86::COND_BE;
2300 case X86::COND_AE: return X86::COND_B;
2301 case X86::COND_S: return X86::COND_NS;
2302 case X86::COND_NS: return X86::COND_S;
2303 case X86::COND_P: return X86::COND_NP;
2304 case X86::COND_NP: return X86::COND_P;
2305 case X86::COND_O: return X86::COND_NO;
2306 case X86::COND_NO: return X86::COND_O;
2307 }
2308}
2309
Dale Johannesen318093b2007-06-14 22:03:45 +00002310bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
Evan Cheng5a96b3d2011-12-07 07:15:52 +00002311 if (!MI->isTerminator()) return false;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002312
Chris Lattner69244302008-01-07 01:56:04 +00002313 // Conditional branch is a special case.
Evan Cheng5a96b3d2011-12-07 07:15:52 +00002314 if (MI->isBranch() && !MI->isBarrier())
Chris Lattner69244302008-01-07 01:56:04 +00002315 return true;
Evan Cheng5a96b3d2011-12-07 07:15:52 +00002316 if (!MI->isPredicable())
Chris Lattner69244302008-01-07 01:56:04 +00002317 return true;
2318 return !isPredicated(MI);
Dale Johannesen318093b2007-06-14 22:03:45 +00002319}
Chris Lattner9cd68752006-10-21 05:52:40 +00002320
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002321bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
Chris Lattner7fbe9722006-10-20 17:42:20 +00002322 MachineBasicBlock *&TBB,
2323 MachineBasicBlock *&FBB,
Evan Chengdc54d312009-02-09 07:14:22 +00002324 SmallVectorImpl<MachineOperand> &Cond,
2325 bool AllowModify) const {
Dan Gohman279c22e2008-10-21 03:29:32 +00002326 // Start from the bottom of the block and work up, examining the
2327 // terminator instructions.
Chris Lattner7fbe9722006-10-20 17:42:20 +00002328 MachineBasicBlock::iterator I = MBB.end();
Evan Chengfc5a03e2010-04-13 18:50:27 +00002329 MachineBasicBlock::iterator UnCondBrIter = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +00002330 while (I != MBB.begin()) {
2331 --I;
Dale Johannesen93d6a7e2010-04-02 01:38:09 +00002332 if (I->isDebugValue())
2333 continue;
Bill Wendling85de1e52009-12-14 06:51:19 +00002334
2335 // Working from the bottom, when we see a non-terminator instruction, we're
2336 // done.
Jakob Stoklund Olesen468a2a42010-07-16 17:41:44 +00002337 if (!isUnpredicatedTerminator(I))
Dan Gohman279c22e2008-10-21 03:29:32 +00002338 break;
Bill Wendling85de1e52009-12-14 06:51:19 +00002339
2340 // A terminator that isn't a branch can't easily be handled by this
2341 // analysis.
Evan Cheng5a96b3d2011-12-07 07:15:52 +00002342 if (!I->isBranch())
Chris Lattner7fbe9722006-10-20 17:42:20 +00002343 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +00002344
Dan Gohman279c22e2008-10-21 03:29:32 +00002345 // Handle unconditional branches.
Chris Lattnerbd13fb62010-02-11 19:25:55 +00002346 if (I->getOpcode() == X86::JMP_4) {
Evan Chengfc5a03e2010-04-13 18:50:27 +00002347 UnCondBrIter = I;
2348
Evan Chengdc54d312009-02-09 07:14:22 +00002349 if (!AllowModify) {
2350 TBB = I->getOperand(0).getMBB();
Evan Cheng45e00102009-05-08 06:34:09 +00002351 continue;
Evan Chengdc54d312009-02-09 07:14:22 +00002352 }
2353
Dan Gohman279c22e2008-10-21 03:29:32 +00002354 // If the block has any instructions after a JMP, delete them.
Chris Lattner7896c9f2009-12-03 00:50:42 +00002355 while (llvm::next(I) != MBB.end())
2356 llvm::next(I)->eraseFromParent();
Bill Wendling85de1e52009-12-14 06:51:19 +00002357
Dan Gohman279c22e2008-10-21 03:29:32 +00002358 Cond.clear();
2359 FBB = 0;
Bill Wendling85de1e52009-12-14 06:51:19 +00002360
Dan Gohman279c22e2008-10-21 03:29:32 +00002361 // Delete the JMP if it's equivalent to a fall-through.
2362 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
2363 TBB = 0;
2364 I->eraseFromParent();
2365 I = MBB.end();
Evan Chengfc5a03e2010-04-13 18:50:27 +00002366 UnCondBrIter = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +00002367 continue;
2368 }
Bill Wendling85de1e52009-12-14 06:51:19 +00002369
Evan Chengfc5a03e2010-04-13 18:50:27 +00002370 // TBB is used to indicate the unconditional destination.
Dan Gohman279c22e2008-10-21 03:29:32 +00002371 TBB = I->getOperand(0).getMBB();
2372 continue;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002373 }
Bill Wendling85de1e52009-12-14 06:51:19 +00002374
Dan Gohman279c22e2008-10-21 03:29:32 +00002375 // Handle conditional branches.
2376 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode());
Chris Lattner7fbe9722006-10-20 17:42:20 +00002377 if (BranchCode == X86::COND_INVALID)
2378 return true; // Can't handle indirect branch.
Bill Wendling85de1e52009-12-14 06:51:19 +00002379
Dan Gohman279c22e2008-10-21 03:29:32 +00002380 // Working from the bottom, handle the first conditional branch.
2381 if (Cond.empty()) {
Evan Chengfc5a03e2010-04-13 18:50:27 +00002382 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB();
2383 if (AllowModify && UnCondBrIter != MBB.end() &&
2384 MBB.isLayoutSuccessor(TargetBB)) {
2385 // If we can modify the code and it ends in something like:
2386 //
2387 // jCC L1
2388 // jmp L2
2389 // L1:
2390 // ...
2391 // L2:
2392 //
2393 // Then we can change this to:
2394 //
2395 // jnCC L2
2396 // L1:
2397 // ...
2398 // L2:
2399 //
2400 // Which is a bit more efficient.
2401 // We conditionally jump to the fall-through block.
2402 BranchCode = GetOppositeBranchCondition(BranchCode);
2403 unsigned JNCC = GetCondBranchFromCond(BranchCode);
2404 MachineBasicBlock::iterator OldInst = I;
2405
2406 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC))
2407 .addMBB(UnCondBrIter->getOperand(0).getMBB());
2408 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4))
2409 .addMBB(TargetBB);
Evan Chengfc5a03e2010-04-13 18:50:27 +00002410
2411 OldInst->eraseFromParent();
2412 UnCondBrIter->eraseFromParent();
2413
2414 // Restart the analysis.
2415 UnCondBrIter = MBB.end();
2416 I = MBB.end();
2417 continue;
2418 }
2419
Dan Gohman279c22e2008-10-21 03:29:32 +00002420 FBB = TBB;
2421 TBB = I->getOperand(0).getMBB();
2422 Cond.push_back(MachineOperand::CreateImm(BranchCode));
2423 continue;
2424 }
Bill Wendling85de1e52009-12-14 06:51:19 +00002425
2426 // Handle subsequent conditional branches. Only handle the case where all
2427 // conditional branches branch to the same destination and their condition
2428 // opcodes fit one of the special multi-branch idioms.
Dan Gohman279c22e2008-10-21 03:29:32 +00002429 assert(Cond.size() == 1);
2430 assert(TBB);
Bill Wendling85de1e52009-12-14 06:51:19 +00002431
2432 // Only handle the case where all conditional branches branch to the same
2433 // destination.
Dan Gohman279c22e2008-10-21 03:29:32 +00002434 if (TBB != I->getOperand(0).getMBB())
2435 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +00002436
Dan Gohman279c22e2008-10-21 03:29:32 +00002437 // If the conditions are the same, we can leave them alone.
Bill Wendling85de1e52009-12-14 06:51:19 +00002438 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm();
Dan Gohman279c22e2008-10-21 03:29:32 +00002439 if (OldBranchCode == BranchCode)
2440 continue;
Bill Wendling85de1e52009-12-14 06:51:19 +00002441
2442 // If they differ, see if they fit one of the known patterns. Theoretically,
2443 // we could handle more patterns here, but we shouldn't expect to see them
2444 // if instruction selection has done a reasonable job.
Dan Gohman279c22e2008-10-21 03:29:32 +00002445 if ((OldBranchCode == X86::COND_NP &&
2446 BranchCode == X86::COND_E) ||
2447 (OldBranchCode == X86::COND_E &&
2448 BranchCode == X86::COND_NP))
2449 BranchCode = X86::COND_NP_OR_E;
2450 else if ((OldBranchCode == X86::COND_P &&
2451 BranchCode == X86::COND_NE) ||
2452 (OldBranchCode == X86::COND_NE &&
2453 BranchCode == X86::COND_P))
2454 BranchCode = X86::COND_NE_OR_P;
2455 else
2456 return true;
Bill Wendling85de1e52009-12-14 06:51:19 +00002457
Dan Gohman279c22e2008-10-21 03:29:32 +00002458 // Update the MachineOperand.
2459 Cond[0].setImm(BranchCode);
Chris Lattner6ce64432006-10-30 22:27:23 +00002460 }
Chris Lattner7fbe9722006-10-20 17:42:20 +00002461
Dan Gohman279c22e2008-10-21 03:29:32 +00002462 return false;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002463}
2464
Evan Cheng6ae36262007-05-18 00:18:17 +00002465unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
Chris Lattner7fbe9722006-10-20 17:42:20 +00002466 MachineBasicBlock::iterator I = MBB.end();
Dan Gohman279c22e2008-10-21 03:29:32 +00002467 unsigned Count = 0;
2468
2469 while (I != MBB.begin()) {
2470 --I;
Dale Johannesen93d6a7e2010-04-02 01:38:09 +00002471 if (I->isDebugValue())
2472 continue;
Chris Lattnerbd13fb62010-02-11 19:25:55 +00002473 if (I->getOpcode() != X86::JMP_4 &&
Dan Gohman279c22e2008-10-21 03:29:32 +00002474 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
2475 break;
2476 // Remove the branch.
2477 I->eraseFromParent();
2478 I = MBB.end();
2479 ++Count;
2480 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002481
Dan Gohman279c22e2008-10-21 03:29:32 +00002482 return Count;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002483}
2484
Evan Cheng6ae36262007-05-18 00:18:17 +00002485unsigned
2486X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
2487 MachineBasicBlock *FBB,
Stuart Hastings3bf91252010-06-17 22:43:56 +00002488 const SmallVectorImpl<MachineOperand> &Cond,
2489 DebugLoc DL) const {
Chris Lattner7fbe9722006-10-20 17:42:20 +00002490 // Shouldn't be a fall through.
2491 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
Chris Lattner34a84ac2006-10-21 05:34:23 +00002492 assert((Cond.size() == 1 || Cond.size() == 0) &&
2493 "X86 branch conditions have one component!");
2494
Dan Gohman279c22e2008-10-21 03:29:32 +00002495 if (Cond.empty()) {
2496 // Unconditional branch?
2497 assert(!FBB && "Unconditional branch with multiple successors!");
Stuart Hastings3bf91252010-06-17 22:43:56 +00002498 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB);
Evan Cheng6ae36262007-05-18 00:18:17 +00002499 return 1;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002500 }
Dan Gohman279c22e2008-10-21 03:29:32 +00002501
2502 // Conditional branch.
2503 unsigned Count = 0;
2504 X86::CondCode CC = (X86::CondCode)Cond[0].getImm();
2505 switch (CC) {
2506 case X86::COND_NP_OR_E:
2507 // Synthesize NP_OR_E with two branches.
Stuart Hastings3bf91252010-06-17 22:43:56 +00002508 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +00002509 ++Count;
Stuart Hastings3bf91252010-06-17 22:43:56 +00002510 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +00002511 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +00002512 break;
2513 case X86::COND_NE_OR_P:
2514 // Synthesize NE_OR_P with two branches.
Stuart Hastings3bf91252010-06-17 22:43:56 +00002515 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +00002516 ++Count;
Stuart Hastings3bf91252010-06-17 22:43:56 +00002517 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +00002518 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +00002519 break;
Bill Wendling18ce64e2010-03-05 00:33:59 +00002520 default: {
2521 unsigned Opc = GetCondBranchFromCond(CC);
Stuart Hastings3bf91252010-06-17 22:43:56 +00002522 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB);
Bill Wendling18ce64e2010-03-05 00:33:59 +00002523 ++Count;
Dan Gohman279c22e2008-10-21 03:29:32 +00002524 }
Bill Wendling18ce64e2010-03-05 00:33:59 +00002525 }
Dan Gohman279c22e2008-10-21 03:29:32 +00002526 if (FBB) {
2527 // Two-way Conditional branch. Insert the second branch.
Stuart Hastings3bf91252010-06-17 22:43:56 +00002528 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB);
Dan Gohman279c22e2008-10-21 03:29:32 +00002529 ++Count;
2530 }
2531 return Count;
Chris Lattner7fbe9722006-10-20 17:42:20 +00002532}
2533
Dan Gohman6d9305c2009-04-15 00:04:23 +00002534/// isHReg - Test if the given register is a physical h register.
2535static bool isHReg(unsigned Reg) {
Dan Gohman4af325d2009-04-27 16:41:36 +00002536 return X86::GR8_ABCD_HRegClass.contains(Reg);
Dan Gohman6d9305c2009-04-15 00:04:23 +00002537}
2538
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +00002539// Try and copy between VR128/VR64 and GR64 registers.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002540static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg,
2541 bool HasAVX) {
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +00002542 // SrcReg(VR128) -> DestReg(GR64)
2543 // SrcReg(VR64) -> DestReg(GR64)
2544 // SrcReg(GR64) -> DestReg(VR128)
2545 // SrcReg(GR64) -> DestReg(VR64)
2546
2547 if (X86::GR64RegClass.contains(DestReg)) {
2548 if (X86::VR128RegClass.contains(SrcReg)) {
2549 // Copy from a VR128 register to a GR64 register.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002550 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +00002551 } else if (X86::VR64RegClass.contains(SrcReg)) {
2552 // Copy from a VR64 register to a GR64 register.
2553 return X86::MOVSDto64rr;
2554 }
2555 } else if (X86::GR64RegClass.contains(SrcReg)) {
2556 // Copy from a GR64 register to a VR128 register.
2557 if (X86::VR128RegClass.contains(DestReg))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002558 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +00002559 // Copy from a GR64 register to a VR64 register.
2560 else if (X86::VR64RegClass.contains(DestReg))
2561 return X86::MOV64toSDrr;
2562 }
2563
Jakob Stoklund Olesen4bd89872011-09-22 22:45:24 +00002564 // SrcReg(FR32) -> DestReg(GR32)
2565 // SrcReg(GR32) -> DestReg(FR32)
2566
2567 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg))
2568 // Copy from a FR32 register to a GR32 register.
2569 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr;
2570
2571 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg))
2572 // Copy from a GR32 register to a FR32 register.
2573 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr;
2574
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +00002575 return 0;
2576}
2577
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +00002578void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
2579 MachineBasicBlock::iterator MI, DebugLoc DL,
2580 unsigned DestReg, unsigned SrcReg,
2581 bool KillSrc) const {
2582 // First deal with the normal symmetric copies.
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002583 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +00002584 unsigned Opc = 0;
2585 if (X86::GR64RegClass.contains(DestReg, SrcReg))
2586 Opc = X86::MOV64rr;
2587 else if (X86::GR32RegClass.contains(DestReg, SrcReg))
2588 Opc = X86::MOV32rr;
2589 else if (X86::GR16RegClass.contains(DestReg, SrcReg))
2590 Opc = X86::MOV16rr;
2591 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) {
2592 // Copying to or from a physical H register on x86-64 requires a NOREX
2593 // move. Otherwise use a normal move.
2594 if ((isHReg(DestReg) || isHReg(SrcReg)) &&
Jakob Stoklund Olesenb66f1842011-10-07 20:15:54 +00002595 TM.getSubtarget<X86Subtarget>().is64Bit()) {
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +00002596 Opc = X86::MOV8rr_NOREX;
Jakob Stoklund Olesenb66f1842011-10-07 20:15:54 +00002597 // Both operands must be encodable without an REX prefix.
2598 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) &&
2599 "8-bit H register can not be copied outside GR8_NOREX");
2600 } else
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +00002601 Opc = X86::MOV8rr;
2602 } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002603 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr;
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00002604 else if (X86::VR256RegClass.contains(DestReg, SrcReg))
2605 Opc = X86::VMOVAPSYrr;
Jakob Stoklund Olesen61c8ecc2010-07-08 22:30:35 +00002606 else if (X86::VR64RegClass.contains(DestReg, SrcReg))
2607 Opc = X86::MMX_MOVQ64rr;
Anton Korobeynikovc52bedb2010-08-27 14:43:06 +00002608 else
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002609 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX);
Jakob Stoklund Olesen320bdcb2010-07-08 19:46:25 +00002610
2611 if (Opc) {
2612 BuildMI(MBB, MI, DL, get(Opc), DestReg)
2613 .addReg(SrcReg, getKillRegState(KillSrc));
2614 return;
2615 }
2616
2617 // Moving EFLAGS to / from another register requires a push and a pop.
2618 if (SrcReg == X86::EFLAGS) {
2619 if (X86::GR64RegClass.contains(DestReg)) {
2620 BuildMI(MBB, MI, DL, get(X86::PUSHF64));
2621 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg);
2622 return;
2623 } else if (X86::GR32RegClass.contains(DestReg)) {
2624 BuildMI(MBB, MI, DL, get(X86::PUSHF32));
2625 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg);
2626 return;
2627 }
2628 }
2629 if (DestReg == X86::EFLAGS) {
2630 if (X86::GR64RegClass.contains(SrcReg)) {
2631 BuildMI(MBB, MI, DL, get(X86::PUSH64r))
2632 .addReg(SrcReg, getKillRegState(KillSrc));
2633 BuildMI(MBB, MI, DL, get(X86::POPF64));
2634 return;
2635 } else if (X86::GR32RegClass.contains(SrcReg)) {
2636 BuildMI(MBB, MI, DL, get(X86::PUSH32r))
2637 .addReg(SrcReg, getKillRegState(KillSrc));
2638 BuildMI(MBB, MI, DL, get(X86::POPF32));
2639 return;
2640 }
2641 }
2642
2643 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg)
2644 << " to " << RI.getName(DestReg) << '\n');
2645 llvm_unreachable("Cannot emit physreg copy instruction");
2646}
2647
Rafael Espindola21d238f2010-06-12 20:13:29 +00002648static unsigned getLoadStoreRegOpcode(unsigned Reg,
2649 const TargetRegisterClass *RC,
2650 bool isStackAligned,
2651 const TargetMachine &TM,
2652 bool load) {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002653 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +00002654 switch (RC->getSize()) {
Rafael Espindola5a717a32010-07-12 03:43:04 +00002655 default:
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +00002656 llvm_unreachable("Unknown spill size");
2657 case 1:
2658 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +00002659 if (TM.getSubtarget<X86Subtarget>().is64Bit())
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +00002660 // Copying to or from a physical H register on x86-64 requires a NOREX
2661 // move. Otherwise use a normal move.
2662 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC))
2663 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX;
2664 return load ? X86::MOV8rm : X86::MOV8mr;
2665 case 2:
2666 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass");
2667 return load ? X86::MOV16rm : X86::MOV16mr;
2668 case 4:
2669 if (X86::GR32RegClass.hasSubClassEq(RC))
2670 return load ? X86::MOV32rm : X86::MOV32mr;
2671 if (X86::FR32RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002672 return load ?
2673 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) :
2674 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr);
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +00002675 if (X86::RFP32RegClass.hasSubClassEq(RC))
2676 return load ? X86::LD_Fp32m : X86::ST_Fp32m;
2677 llvm_unreachable("Unknown 4-byte regclass");
2678 case 8:
2679 if (X86::GR64RegClass.hasSubClassEq(RC))
2680 return load ? X86::MOV64rm : X86::MOV64mr;
2681 if (X86::FR64RegClass.hasSubClassEq(RC))
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002682 return load ?
2683 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) :
2684 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr);
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +00002685 if (X86::VR64RegClass.hasSubClassEq(RC))
2686 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr;
2687 if (X86::RFP64RegClass.hasSubClassEq(RC))
2688 return load ? X86::LD_Fp64m : X86::ST_Fp64m;
2689 llvm_unreachable("Unknown 8-byte regclass");
2690 case 10:
2691 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +00002692 return load ? X86::LD_Fp80m : X86::ST_FpP80m;
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +00002693 case 16: {
Jakob Stoklund Olesen1f9a09c2011-06-01 15:32:10 +00002694 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass");
Rafael Espindola21d238f2010-06-12 20:13:29 +00002695 // If stack is realigned we can use aligned stores.
2696 if (isStackAligned)
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +00002697 return load ?
2698 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) :
2699 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
Rafael Espindola21d238f2010-06-12 20:13:29 +00002700 else
Bruno Cardoso Lopes5affa512011-08-31 03:04:09 +00002701 return load ?
2702 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) :
2703 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
2704 }
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00002705 case 32:
2706 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
2707 // If stack is realigned we can use aligned stores.
2708 if (isStackAligned)
2709 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
2710 else
2711 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
Rafael Espindola21d238f2010-06-12 20:13:29 +00002712 }
2713}
2714
Dan Gohman4af325d2009-04-27 16:41:36 +00002715static unsigned getStoreRegOpcode(unsigned SrcReg,
2716 const TargetRegisterClass *RC,
2717 bool isStackAligned,
2718 TargetMachine &TM) {
Rafael Espindola21d238f2010-06-12 20:13:29 +00002719 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false);
2720}
Owen Andersonf6372aa2008-01-01 21:11:32 +00002721
Rafael Espindola21d238f2010-06-12 20:13:29 +00002722
2723static unsigned getLoadRegOpcode(unsigned DestReg,
2724 const TargetRegisterClass *RC,
2725 bool isStackAligned,
2726 const TargetMachine &TM) {
2727 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true);
Owen Andersonf6372aa2008-01-01 21:11:32 +00002728}
2729
2730void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
2731 MachineBasicBlock::iterator MI,
2732 unsigned SrcReg, bool isKill, int FrameIdx,
Evan Cheng746ad692010-05-06 19:06:44 +00002733 const TargetRegisterClass *RC,
2734 const TargetRegisterInfo *TRI) const {
Anton Korobeynikov88bbf692008-07-19 06:30:51 +00002735 const MachineFunction &MF = *MBB.getParent();
Jakob Stoklund Olesen516cd452010-07-27 04:16:58 +00002736 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() &&
2737 "Stack slot too small for store");
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002738 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2739 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Cheng2fa82bc2011-06-23 01:53:43 +00002740 RI.canRealignStack(MF);
Dan Gohman4af325d2009-04-27 16:41:36 +00002741 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Dale Johannesen6ec25f52010-01-26 00:03:12 +00002742 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendlingfbef3102009-02-11 21:51:19 +00002743 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx)
Bill Wendling587daed2009-05-13 21:33:08 +00002744 .addReg(SrcReg, getKillRegState(isKill));
Owen Andersonf6372aa2008-01-01 21:11:32 +00002745}
2746
2747void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
2748 bool isKill,
2749 SmallVectorImpl<MachineOperand> &Addr,
2750 const TargetRegisterClass *RC,
Dan Gohman91e69c32009-10-09 18:10:05 +00002751 MachineInstr::mmo_iterator MMOBegin,
2752 MachineInstr::mmo_iterator MMOEnd,
Owen Andersonf6372aa2008-01-01 21:11:32 +00002753 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002754 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2755 bool isAligned = MMOBegin != MMOEnd &&
2756 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman4af325d2009-04-27 16:41:36 +00002757 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM);
Chris Lattnerc7f3ace2010-04-02 20:16:16 +00002758 DebugLoc DL;
Dale Johannesen21b55412009-02-12 23:08:38 +00002759 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc));
Owen Andersonf6372aa2008-01-01 21:11:32 +00002760 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00002761 MIB.addOperand(Addr[i]);
Bill Wendling587daed2009-05-13 21:33:08 +00002762 MIB.addReg(SrcReg, getKillRegState(isKill));
Dan Gohman91e69c32009-10-09 18:10:05 +00002763 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersonf6372aa2008-01-01 21:11:32 +00002764 NewMIs.push_back(MIB);
2765}
2766
Owen Andersonf6372aa2008-01-01 21:11:32 +00002767
2768void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
Anton Korobeynikov88bbf692008-07-19 06:30:51 +00002769 MachineBasicBlock::iterator MI,
2770 unsigned DestReg, int FrameIdx,
Evan Cheng746ad692010-05-06 19:06:44 +00002771 const TargetRegisterClass *RC,
2772 const TargetRegisterInfo *TRI) const {
Anton Korobeynikov88bbf692008-07-19 06:30:51 +00002773 const MachineFunction &MF = *MBB.getParent();
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002774 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2775 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) ||
Evan Cheng2fa82bc2011-06-23 01:53:43 +00002776 RI.canRealignStack(MF);
Dan Gohman4af325d2009-04-27 16:41:36 +00002777 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Dale Johannesen6ec25f52010-01-26 00:03:12 +00002778 DebugLoc DL = MBB.findDebugLoc(MI);
Bill Wendlingfbef3102009-02-11 21:51:19 +00002779 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx);
Owen Andersonf6372aa2008-01-01 21:11:32 +00002780}
2781
2782void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
Evan Cheng9f1c8312008-07-03 09:09:37 +00002783 SmallVectorImpl<MachineOperand> &Addr,
2784 const TargetRegisterClass *RC,
Dan Gohman91e69c32009-10-09 18:10:05 +00002785 MachineInstr::mmo_iterator MMOBegin,
2786 MachineInstr::mmo_iterator MMOEnd,
Owen Andersonf6372aa2008-01-01 21:11:32 +00002787 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00002788 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
2789 bool isAligned = MMOBegin != MMOEnd &&
2790 (*MMOBegin)->getAlignment() >= Alignment;
Dan Gohman4af325d2009-04-27 16:41:36 +00002791 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM);
Chris Lattnerc7f3ace2010-04-02 20:16:16 +00002792 DebugLoc DL;
Dale Johannesen21b55412009-02-12 23:08:38 +00002793 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg);
Owen Andersonf6372aa2008-01-01 21:11:32 +00002794 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00002795 MIB.addOperand(Addr[i]);
Dan Gohman91e69c32009-10-09 18:10:05 +00002796 (*MIB).setMemRefs(MMOBegin, MMOEnd);
Owen Andersonf6372aa2008-01-01 21:11:32 +00002797 NewMIs.push_back(MIB);
2798}
2799
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +00002800/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr
2801/// instruction with two undef reads of the register being defined. This is
2802/// used for mapping:
2803/// %xmm4 = V_SET0
2804/// to:
2805/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef>
2806///
2807static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) {
2808 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction.");
2809 unsigned Reg = MI->getOperand(0).getReg();
2810 MI->setDesc(Desc);
2811
2812 // MachineInstr::addOperand() will insert explicit operands before any
2813 // implicit operands.
2814 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef)
2815 .addReg(Reg, RegState::Undef);
2816 // But we don't trust that.
2817 assert(MI->getOperand(1).getReg() == Reg &&
2818 MI->getOperand(2).getReg() == Reg && "Misplaced operand");
2819 return true;
2820}
2821
2822bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
2823 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
2824 switch (MI->getOpcode()) {
2825 case X86::V_SET0:
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +00002826 case X86::FsFLD0SS:
2827 case X86::FsFLD0SD:
Jakob Stoklund Olesen3e5d5c52011-11-07 19:15:58 +00002828 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
Jakob Stoklund Olesened744822011-10-08 18:28:28 +00002829 case X86::TEST8ri_NOREX:
2830 MI->setDesc(get(X86::TEST8ri));
2831 return true;
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +00002832 }
2833 return false;
2834}
2835
Evan Cheng962021b2010-04-26 07:38:55 +00002836MachineInstr*
2837X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF,
Evan Cheng8601a3d2010-04-29 01:13:30 +00002838 int FrameIx, uint64_t Offset,
Evan Cheng962021b2010-04-26 07:38:55 +00002839 const MDNode *MDPtr,
2840 DebugLoc DL) const {
Evan Cheng962021b2010-04-26 07:38:55 +00002841 X86AddressMode AM;
2842 AM.BaseType = X86AddressMode::FrameIndexBase;
2843 AM.Base.FrameIndex = FrameIx;
2844 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE));
2845 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr);
2846 return &*MIB;
2847}
2848
Dan Gohman8e5f2c62008-07-07 23:14:23 +00002849static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode,
Dan Gohmand68a0762009-01-05 17:59:02 +00002850 const SmallVectorImpl<MachineOperand> &MOs,
Bill Wendling9bc96a52009-02-03 00:55:04 +00002851 MachineInstr *MI,
2852 const TargetInstrInfo &TII) {
Owen Anderson43dbe052008-01-07 01:35:02 +00002853 // Create the base instruction with the memory operand as the first part.
Bill Wendling9bc96a52009-02-03 00:55:04 +00002854 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2855 MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +00002856 MachineInstrBuilder MIB(NewMI);
2857 unsigned NumAddrOps = MOs.size();
2858 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00002859 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +00002860 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +00002861 addOffset(MIB, 0);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002862
Owen Anderson43dbe052008-01-07 01:35:02 +00002863 // Loop over the rest of the ri operands, converting them over.
Chris Lattner749c6f62008-01-07 07:27:27 +00002864 unsigned NumOps = MI->getDesc().getNumOperands()-2;
Owen Anderson43dbe052008-01-07 01:35:02 +00002865 for (unsigned i = 0; i != NumOps; ++i) {
2866 MachineOperand &MO = MI->getOperand(i+2);
Dan Gohman97357612009-02-18 05:45:50 +00002867 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +00002868 }
2869 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) {
2870 MachineOperand &MO = MI->getOperand(i);
Dan Gohman97357612009-02-18 05:45:50 +00002871 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +00002872 }
2873 return MIB;
2874}
2875
Dan Gohman8e5f2c62008-07-07 23:14:23 +00002876static MachineInstr *FuseInst(MachineFunction &MF,
2877 unsigned Opcode, unsigned OpNo,
Dan Gohmand68a0762009-01-05 17:59:02 +00002878 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson43dbe052008-01-07 01:35:02 +00002879 MachineInstr *MI, const TargetInstrInfo &TII) {
Bill Wendling9bc96a52009-02-03 00:55:04 +00002880 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode),
2881 MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +00002882 MachineInstrBuilder MIB(NewMI);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002883
Owen Anderson43dbe052008-01-07 01:35:02 +00002884 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2885 MachineOperand &MO = MI->getOperand(i);
2886 if (i == OpNo) {
Dan Gohmand735b802008-10-03 15:45:36 +00002887 assert(MO.isReg() && "Expected to fold into reg operand!");
Owen Anderson43dbe052008-01-07 01:35:02 +00002888 unsigned NumAddrOps = MOs.size();
2889 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00002890 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +00002891 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +00002892 addOffset(MIB, 0);
Owen Anderson43dbe052008-01-07 01:35:02 +00002893 } else {
Dan Gohman97357612009-02-18 05:45:50 +00002894 MIB.addOperand(MO);
Owen Anderson43dbe052008-01-07 01:35:02 +00002895 }
2896 }
2897 return MIB;
2898}
2899
2900static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode,
Dan Gohmand68a0762009-01-05 17:59:02 +00002901 const SmallVectorImpl<MachineOperand> &MOs,
Owen Anderson43dbe052008-01-07 01:35:02 +00002902 MachineInstr *MI) {
Dan Gohman8e5f2c62008-07-07 23:14:23 +00002903 MachineFunction &MF = *MI->getParent()->getParent();
Bill Wendlingfbef3102009-02-11 21:51:19 +00002904 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode));
Owen Anderson43dbe052008-01-07 01:35:02 +00002905
2906 unsigned NumAddrOps = MOs.size();
2907 for (unsigned i = 0; i != NumAddrOps; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00002908 MIB.addOperand(MOs[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +00002909 if (NumAddrOps < 4) // FrameIndex only
Rafael Espindola094fad32009-04-08 21:14:34 +00002910 addOffset(MIB, 0);
Owen Anderson43dbe052008-01-07 01:35:02 +00002911 return MIB.addImm(0);
2912}
2913
2914MachineInstr*
Dan Gohmanc54baa22008-12-03 18:43:12 +00002915X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
2916 MachineInstr *MI, unsigned i,
Evan Chengf9b36f02009-07-15 06:10:07 +00002917 const SmallVectorImpl<MachineOperand> &MOs,
Evan Cheng9cef48e2009-09-11 00:39:26 +00002918 unsigned Size, unsigned Align) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +00002919 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
Owen Anderson43dbe052008-01-07 01:35:02 +00002920 bool isTwoAddrFold = false;
Chris Lattner749c6f62008-01-07 07:27:27 +00002921 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson43dbe052008-01-07 01:35:02 +00002922 bool isTwoAddr = NumOps > 1 &&
Evan Chenge837dea2011-06-28 19:10:37 +00002923 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson43dbe052008-01-07 01:35:02 +00002924
Jakob Stoklund Olesen60045c22011-04-30 23:00:05 +00002925 // FIXME: AsmPrinter doesn't know how to handle
2926 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
2927 if (MI->getOpcode() == X86::ADD32ri &&
2928 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
2929 return NULL;
2930
Owen Anderson43dbe052008-01-07 01:35:02 +00002931 MachineInstr *NewMI = NULL;
2932 // Folding a memory location into the two-address part of a two-address
2933 // instruction is different than folding it other places. It requires
2934 // replacing the *two* registers with the memory location.
2935 if (isTwoAddr && NumOps >= 2 && i < 2 &&
Dan Gohmand735b802008-10-03 15:45:36 +00002936 MI->getOperand(0).isReg() &&
2937 MI->getOperand(1).isReg() &&
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002938 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) {
Owen Anderson43dbe052008-01-07 01:35:02 +00002939 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
2940 isTwoAddrFold = true;
2941 } else if (i == 0) { // If operand 0
Dan Gohmanf1b4d262010-01-12 04:42:54 +00002942 if (MI->getOpcode() == X86::MOV64r0)
2943 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI);
2944 else if (MI->getOpcode() == X86::MOV32r0)
Owen Anderson43dbe052008-01-07 01:35:02 +00002945 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI);
Dan Gohmanf1b4d262010-01-12 04:42:54 +00002946 else if (MI->getOpcode() == X86::MOV16r0)
2947 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI);
Owen Anderson43dbe052008-01-07 01:35:02 +00002948 else if (MI->getOpcode() == X86::MOV8r0)
2949 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI);
Evan Cheng9f1c8312008-07-03 09:09:37 +00002950 if (NewMI)
Owen Anderson43dbe052008-01-07 01:35:02 +00002951 return NewMI;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002952
Owen Anderson43dbe052008-01-07 01:35:02 +00002953 OpcodeTablePtr = &RegOp2MemOpTable0;
2954 } else if (i == 1) {
2955 OpcodeTablePtr = &RegOp2MemOpTable1;
2956 } else if (i == 2) {
2957 OpcodeTablePtr = &RegOp2MemOpTable2;
2958 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00002959
Owen Anderson43dbe052008-01-07 01:35:02 +00002960 // If table selected...
2961 if (OpcodeTablePtr) {
2962 // Find the Opcode to fuse
Chris Lattner45a1cb22010-10-07 23:08:41 +00002963 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
2964 OpcodeTablePtr->find(MI->getOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +00002965 if (I != OpcodeTablePtr->end()) {
Evan Cheng9cef48e2009-09-11 00:39:26 +00002966 unsigned Opcode = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00002967 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
Evan Chengf9b36f02009-07-15 06:10:07 +00002968 if (Align < MinAlign)
2969 return NULL;
Evan Cheng879caea2009-09-11 01:01:31 +00002970 bool NarrowToMOV32rm = false;
Evan Cheng9cef48e2009-09-11 00:39:26 +00002971 if (Size) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +00002972 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize();
Evan Cheng9cef48e2009-09-11 00:39:26 +00002973 if (Size < RCSize) {
2974 // Check if it's safe to fold the load. If the size of the object is
2975 // narrower than the load width, then it's not.
2976 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4)
2977 return NULL;
2978 // If this is a 64-bit load, but the spill slot is 32, then we can do
2979 // a 32-bit load which is implicitly zero-extended. This likely is due
2980 // to liveintervalanalysis remat'ing a load from stack slot.
Evan Cheng879caea2009-09-11 01:01:31 +00002981 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg())
2982 return NULL;
Evan Cheng9cef48e2009-09-11 00:39:26 +00002983 Opcode = X86::MOV32rm;
Evan Cheng879caea2009-09-11 01:01:31 +00002984 NarrowToMOV32rm = true;
Evan Cheng9cef48e2009-09-11 00:39:26 +00002985 }
2986 }
2987
Owen Anderson43dbe052008-01-07 01:35:02 +00002988 if (isTwoAddrFold)
Evan Cheng9cef48e2009-09-11 00:39:26 +00002989 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this);
Owen Anderson43dbe052008-01-07 01:35:02 +00002990 else
Evan Cheng9cef48e2009-09-11 00:39:26 +00002991 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this);
Evan Cheng879caea2009-09-11 01:01:31 +00002992
2993 if (NarrowToMOV32rm) {
2994 // If this is the special case where we use a MOV32rm to load a 32-bit
2995 // value and zero-extend the top bits. Change the destination register
2996 // to a 32-bit one.
2997 unsigned DstReg = NewMI->getOperand(0).getReg();
2998 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
2999 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg,
Jakob Stoklund Olesen3458e9e2010-05-24 14:48:17 +00003000 X86::sub_32bit));
Evan Cheng879caea2009-09-11 01:01:31 +00003001 else
Jakob Stoklund Olesen3458e9e2010-05-24 14:48:17 +00003002 NewMI->getOperand(0).setSubReg(X86::sub_32bit);
Evan Cheng879caea2009-09-11 01:01:31 +00003003 }
Owen Anderson43dbe052008-01-07 01:35:02 +00003004 return NewMI;
3005 }
3006 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003007
3008 // No fusion
Jakob Stoklund Olesen9c50e8b2010-07-09 20:43:09 +00003009 if (PrintFailedFusing && !MI->isCopy())
David Greene5b901322010-01-05 01:29:29 +00003010 dbgs() << "We failed to fuse operand " << i << " in " << *MI;
Owen Anderson43dbe052008-01-07 01:35:02 +00003011 return NULL;
3012}
3013
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003014/// hasPartialRegUpdate - Return true for all instructions that only update
3015/// the first 32 or 64-bits of the destination register and leave the rest
3016/// unmodified. This can be used to avoid folding loads if the instructions
3017/// only update part of the destination register, and the non-updated part is
3018/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these
3019/// instructions breaks the partial register dependency and it can improve
3020/// performance. e.g.:
3021///
3022/// movss (%rdi), %xmm0
3023/// cvtss2sd %xmm0, %xmm0
3024///
3025/// Instead of
3026/// cvtss2sd (%rdi), %xmm0
3027///
Bruno Cardoso Lopes6b5b79c2011-09-15 23:04:24 +00003028/// FIXME: This should be turned into a TSFlags.
3029///
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003030static bool hasPartialRegUpdate(unsigned Opcode) {
3031 switch (Opcode) {
Jakob Stoklund Olesenc2ecf3e2011-11-15 01:15:30 +00003032 case X86::CVTSI2SSrr:
3033 case X86::CVTSI2SS64rr:
3034 case X86::CVTSI2SDrr:
3035 case X86::CVTSI2SD64rr:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003036 case X86::CVTSD2SSrr:
3037 case X86::Int_CVTSD2SSrr:
3038 case X86::CVTSS2SDrr:
3039 case X86::Int_CVTSS2SDrr:
3040 case X86::RCPSSr:
3041 case X86::RCPSSr_Int:
3042 case X86::ROUNDSDr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +00003043 case X86::ROUNDSDr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003044 case X86::ROUNDSSr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +00003045 case X86::ROUNDSSr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003046 case X86::RSQRTSSr:
3047 case X86::RSQRTSSr_Int:
3048 case X86::SQRTSSr:
3049 case X86::SQRTSSr_Int:
3050 // AVX encoded versions
3051 case X86::VCVTSD2SSrr:
3052 case X86::Int_VCVTSD2SSrr:
3053 case X86::VCVTSS2SDrr:
3054 case X86::Int_VCVTSS2SDrr:
3055 case X86::VRCPSSr:
3056 case X86::VROUNDSDr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +00003057 case X86::VROUNDSDr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003058 case X86::VROUNDSSr:
Benjamin Kramera73fb9a2011-12-09 15:43:55 +00003059 case X86::VROUNDSSr_Int:
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003060 case X86::VRSQRTSSr:
3061 case X86::VSQRTSSr:
3062 return true;
3063 }
3064
3065 return false;
3066}
Owen Anderson43dbe052008-01-07 01:35:02 +00003067
Jakob Stoklund Olesenc2ecf3e2011-11-15 01:15:30 +00003068/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle
3069/// instructions we would like before a partial register update.
3070unsigned X86InstrInfo::
3071getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
3072 const TargetRegisterInfo *TRI) const {
3073 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode()))
3074 return 0;
3075
3076 // If MI is marked as reading Reg, the partial register update is wanted.
3077 const MachineOperand &MO = MI->getOperand(0);
3078 unsigned Reg = MO.getReg();
3079 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
3080 if (MO.readsReg() || MI->readsVirtualRegister(Reg))
3081 return 0;
3082 } else {
3083 if (MI->readsRegister(Reg, TRI))
3084 return 0;
3085 }
3086
3087 // If any of the preceding 16 instructions are reading Reg, insert a
3088 // dependency breaking instruction. The magic number is based on a few
3089 // Nehalem experiments.
3090 return 16;
3091}
3092
3093void X86InstrInfo::
3094breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
3095 const TargetRegisterInfo *TRI) const {
3096 unsigned Reg = MI->getOperand(OpNum).getReg();
3097 if (X86::VR128RegClass.contains(Reg)) {
3098 // These instructions are all floating point domain, so xorps is the best
3099 // choice.
3100 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX();
3101 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr;
3102 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg)
3103 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef);
3104 } else if (X86::VR256RegClass.contains(Reg)) {
3105 // Use vxorps to clear the full ymm register.
3106 // It wants to read and write the xmm sub-register.
3107 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm);
3108 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg)
3109 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef)
3110 .addReg(Reg, RegState::ImplicitDefine);
3111 } else
3112 return;
3113 MI->addRegisterKilled(Reg, TRI, true);
3114}
3115
Dan Gohmanc54baa22008-12-03 18:43:12 +00003116MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3117 MachineInstr *MI,
Evan Chengf9b36f02009-07-15 06:10:07 +00003118 const SmallVectorImpl<unsigned> &Ops,
Dan Gohmanc54baa22008-12-03 18:43:12 +00003119 int FrameIndex) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003120 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +00003121 if (NoFusing) return NULL;
3122
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003123 // Unless optimizing for size, don't fold to avoid partial
3124 // register update stalls
3125 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3126 hasPartialRegUpdate(MI->getOpcode()))
3127 return 0;
Evan Cheng400073d2009-12-18 07:40:29 +00003128
Evan Cheng5fd79d02008-02-08 21:20:40 +00003129 const MachineFrameInfo *MFI = MF.getFrameInfo();
Evan Cheng9cef48e2009-09-11 00:39:26 +00003130 unsigned Size = MFI->getObjectSize(FrameIndex);
Evan Cheng5fd79d02008-02-08 21:20:40 +00003131 unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
Owen Anderson43dbe052008-01-07 01:35:02 +00003132 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3133 unsigned NewOpc = 0;
Evan Cheng9cef48e2009-09-11 00:39:26 +00003134 unsigned RCSize = 0;
Owen Anderson43dbe052008-01-07 01:35:02 +00003135 switch (MI->getOpcode()) {
3136 default: return NULL;
Evan Cheng9cef48e2009-09-11 00:39:26 +00003137 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break;
Dan Gohmane5efbaf2010-05-18 21:42:03 +00003138 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break;
3139 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break;
3140 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break;
Owen Anderson43dbe052008-01-07 01:35:02 +00003141 }
Evan Cheng9cef48e2009-09-11 00:39:26 +00003142 // Check if it's safe to fold the load. If the size of the object is
3143 // narrower than the load width, then it's not.
3144 if (Size < RCSize)
3145 return NULL;
Owen Anderson43dbe052008-01-07 01:35:02 +00003146 // Change to CMPXXri r, 0 first.
Chris Lattner5080f4d2008-01-11 18:10:50 +00003147 MI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +00003148 MI->getOperand(1).ChangeToImmediate(0);
3149 } else if (Ops.size() != 1)
3150 return NULL;
3151
3152 SmallVector<MachineOperand,4> MOs;
3153 MOs.push_back(MachineOperand::CreateFI(FrameIndex));
Evan Cheng9cef48e2009-09-11 00:39:26 +00003154 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment);
Owen Anderson43dbe052008-01-07 01:35:02 +00003155}
3156
Dan Gohmanc54baa22008-12-03 18:43:12 +00003157MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
3158 MachineInstr *MI,
Evan Chengf9b36f02009-07-15 06:10:07 +00003159 const SmallVectorImpl<unsigned> &Ops,
Dan Gohmanc54baa22008-12-03 18:43:12 +00003160 MachineInstr *LoadMI) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003161 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +00003162 if (NoFusing) return NULL;
3163
Bruno Cardoso Lopescd2857e2011-09-15 21:42:23 +00003164 // Unless optimizing for size, don't fold to avoid partial
3165 // register update stalls
3166 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
3167 hasPartialRegUpdate(MI->getOpcode()))
3168 return 0;
Evan Cheng400073d2009-12-18 07:40:29 +00003169
Dan Gohmancddc11e2008-07-12 00:10:52 +00003170 // Determine the alignment of the load.
Evan Cheng5fd79d02008-02-08 21:20:40 +00003171 unsigned Alignment = 0;
Dan Gohmancddc11e2008-07-12 00:10:52 +00003172 if (LoadMI->hasOneMemOperand())
Dan Gohmanc76909a2009-09-25 20:36:54 +00003173 Alignment = (*LoadMI->memoperands_begin())->getAlignment();
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003174 else
3175 switch (LoadMI->getOpcode()) {
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003176 case X86::AVX_SET0PSY:
3177 case X86::AVX_SET0PDY:
Craig Topper745a86b2011-11-19 22:34:59 +00003178 case X86::AVX2_SETALLONES:
Craig Topper12216172012-01-13 08:12:35 +00003179 case X86::AVX2_SET0:
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003180 Alignment = 32;
3181 break;
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +00003182 case X86::V_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003183 case X86::V_SETALLONES:
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +00003184 case X86::AVX_SETALLONES:
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003185 Alignment = 16;
3186 break;
3187 case X86::FsFLD0SD:
3188 Alignment = 8;
3189 break;
3190 case X86::FsFLD0SS:
3191 Alignment = 4;
3192 break;
3193 default:
Eli Friedmanbe5cbaa2011-06-10 01:13:01 +00003194 return 0;
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003195 }
Owen Anderson43dbe052008-01-07 01:35:02 +00003196 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3197 unsigned NewOpc = 0;
3198 switch (MI->getOpcode()) {
3199 default: return NULL;
3200 case X86::TEST8rr: NewOpc = X86::CMP8ri; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003201 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break;
3202 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break;
3203 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break;
Owen Anderson43dbe052008-01-07 01:35:02 +00003204 }
3205 // Change to CMPXXri r, 0 first.
Chris Lattner5080f4d2008-01-11 18:10:50 +00003206 MI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +00003207 MI->getOperand(1).ChangeToImmediate(0);
3208 } else if (Ops.size() != 1)
3209 return NULL;
3210
Jakob Stoklund Olesend29583b2010-08-11 23:08:22 +00003211 // Make sure the subregisters match.
3212 // Otherwise we risk changing the size of the load.
3213 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg())
3214 return NULL;
3215
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00003216 SmallVector<MachineOperand,X86::AddrNumOperands> MOs;
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003217 switch (LoadMI->getOpcode()) {
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +00003218 case X86::V_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003219 case X86::V_SETALLONES:
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003220 case X86::AVX_SET0PSY:
3221 case X86::AVX_SET0PDY:
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +00003222 case X86::AVX_SETALLONES:
Craig Topper745a86b2011-11-19 22:34:59 +00003223 case X86::AVX2_SETALLONES:
Craig Topper12216172012-01-13 08:12:35 +00003224 case X86::AVX2_SET0:
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003225 case X86::FsFLD0SD:
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +00003226 case X86::FsFLD0SS: {
Jakob Stoklund Olesen92fb79b2011-09-29 05:10:54 +00003227 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
Dan Gohman62c939d2008-12-03 05:21:24 +00003228 // Create a constant-pool entry and operands to load from it.
3229
Dan Gohman81d0c362010-03-09 03:01:40 +00003230 // Medium and large mode can't fold loads this way.
3231 if (TM.getCodeModel() != CodeModel::Small &&
3232 TM.getCodeModel() != CodeModel::Kernel)
3233 return NULL;
3234
Dan Gohman62c939d2008-12-03 05:21:24 +00003235 // x86-32 PIC requires a PIC base register for constant pools.
3236 unsigned PICBase = 0;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +00003237 if (TM.getRelocationModel() == Reloc::PIC_) {
Evan Cheng2b48ab92009-07-16 18:44:05 +00003238 if (TM.getSubtarget<X86Subtarget>().is64Bit())
3239 PICBase = X86::RIP;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +00003240 else
Dan Gohman84023e02010-07-10 09:00:22 +00003241 // FIXME: PICBase = getGlobalBaseReg(&MF);
Evan Cheng2b48ab92009-07-16 18:44:05 +00003242 // This doesn't work for several reasons.
3243 // 1. GlobalBaseReg may have been spilled.
3244 // 2. It may not be live at MI.
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003245 return NULL;
Jakob Stoklund Olesen93e55de2009-07-16 21:24:13 +00003246 }
Dan Gohman62c939d2008-12-03 05:21:24 +00003247
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003248 // Create a constant-pool entry.
Dan Gohman62c939d2008-12-03 05:21:24 +00003249 MachineConstantPool &MCP = *MF.getConstantPool();
Chris Lattnerdb125cf2011-07-18 04:54:35 +00003250 Type *Ty;
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003251 unsigned Opc = LoadMI->getOpcode();
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +00003252 if (Opc == X86::FsFLD0SS)
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003253 Ty = Type::getFloatTy(MF.getFunction()->getContext());
Jakob Stoklund Olesen0edd83b2011-11-29 22:27:25 +00003254 else if (Opc == X86::FsFLD0SD)
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003255 Ty = Type::getDoubleTy(MF.getFunction()->getContext());
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003256 else if (Opc == X86::AVX_SET0PSY || Opc == X86::AVX_SET0PDY)
3257 Ty = VectorType::get(Type::getFloatTy(MF.getFunction()->getContext()), 8);
Craig Topper12216172012-01-13 08:12:35 +00003258 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX2_SET0)
Craig Topperb9c7f652012-01-13 06:12:41 +00003259 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003260 else
3261 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +00003262
Craig Topper745a86b2011-11-19 22:34:59 +00003263 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX_SETALLONES ||
3264 Opc == X86::AVX2_SETALLONES);
Bruno Cardoso Lopes863bd9d2011-07-25 23:05:32 +00003265 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
3266 Constant::getNullValue(Ty);
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003267 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
Dan Gohman62c939d2008-12-03 05:21:24 +00003268
3269 // Create operands to load from the constant pool entry.
3270 MOs.push_back(MachineOperand::CreateReg(PICBase, false));
3271 MOs.push_back(MachineOperand::CreateImm(1));
3272 MOs.push_back(MachineOperand::CreateReg(0, false));
3273 MOs.push_back(MachineOperand::CreateCPI(CPI, 0));
Rafael Espindola094fad32009-04-08 21:14:34 +00003274 MOs.push_back(MachineOperand::CreateReg(0, false));
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003275 break;
3276 }
3277 default: {
Dan Gohman62c939d2008-12-03 05:21:24 +00003278 // Folding a normal load. Just copy the load's address operands.
3279 unsigned NumOps = LoadMI->getDesc().getNumOperands();
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00003280 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
Dan Gohman62c939d2008-12-03 05:21:24 +00003281 MOs.push_back(LoadMI->getOperand(i));
Dan Gohman4a0b3e12009-09-21 18:30:38 +00003282 break;
3283 }
Dan Gohman62c939d2008-12-03 05:21:24 +00003284 }
Evan Cheng9cef48e2009-09-11 00:39:26 +00003285 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment);
Owen Anderson43dbe052008-01-07 01:35:02 +00003286}
3287
3288
Dan Gohman8e8b8a22008-10-16 01:49:15 +00003289bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
3290 const SmallVectorImpl<unsigned> &Ops) const {
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003291 // Check switch flag
Owen Anderson43dbe052008-01-07 01:35:02 +00003292 if (NoFusing) return 0;
3293
3294 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) {
3295 switch (MI->getOpcode()) {
3296 default: return false;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003297 case X86::TEST8rr:
Owen Anderson43dbe052008-01-07 01:35:02 +00003298 case X86::TEST16rr:
3299 case X86::TEST32rr:
3300 case X86::TEST64rr:
3301 return true;
Jakob Stoklund Olesen60045c22011-04-30 23:00:05 +00003302 case X86::ADD32ri:
3303 // FIXME: AsmPrinter doesn't know how to handle
3304 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding.
3305 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS)
3306 return false;
3307 break;
Owen Anderson43dbe052008-01-07 01:35:02 +00003308 }
3309 }
3310
3311 if (Ops.size() != 1)
3312 return false;
3313
3314 unsigned OpNum = Ops[0];
3315 unsigned Opc = MI->getOpcode();
Chris Lattner749c6f62008-01-07 07:27:27 +00003316 unsigned NumOps = MI->getDesc().getNumOperands();
Owen Anderson43dbe052008-01-07 01:35:02 +00003317 bool isTwoAddr = NumOps > 1 &&
Evan Chenge837dea2011-06-28 19:10:37 +00003318 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1;
Owen Anderson43dbe052008-01-07 01:35:02 +00003319
3320 // Folding a memory location into the two-address part of a two-address
3321 // instruction is different than folding it other places. It requires
3322 // replacing the *two* registers with the memory location.
Chris Lattner45a1cb22010-10-07 23:08:41 +00003323 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003324 if (isTwoAddr && NumOps >= 2 && OpNum < 2) {
Owen Anderson43dbe052008-01-07 01:35:02 +00003325 OpcodeTablePtr = &RegOp2MemOpTable2Addr;
3326 } else if (OpNum == 0) { // If operand 0
3327 switch (Opc) {
Chris Lattner9ac75422009-07-14 20:19:57 +00003328 case X86::MOV8r0:
Dan Gohmanf1b4d262010-01-12 04:42:54 +00003329 case X86::MOV16r0:
Owen Anderson43dbe052008-01-07 01:35:02 +00003330 case X86::MOV32r0:
Chris Lattner45a1cb22010-10-07 23:08:41 +00003331 case X86::MOV64r0: return true;
Owen Anderson43dbe052008-01-07 01:35:02 +00003332 default: break;
3333 }
3334 OpcodeTablePtr = &RegOp2MemOpTable0;
3335 } else if (OpNum == 1) {
3336 OpcodeTablePtr = &RegOp2MemOpTable1;
3337 } else if (OpNum == 2) {
3338 OpcodeTablePtr = &RegOp2MemOpTable2;
3339 }
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003340
Chris Lattner99ae6652010-10-08 03:54:52 +00003341 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc))
3342 return true;
Jakob Stoklund Olesen1f323402010-07-09 20:43:13 +00003343 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops);
Owen Anderson43dbe052008-01-07 01:35:02 +00003344}
3345
3346bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
3347 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
Bill Wendlingfbef3102009-02-11 21:51:19 +00003348 SmallVectorImpl<MachineInstr*> &NewMIs) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +00003349 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3350 MemOp2RegOpTable.find(MI->getOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +00003351 if (I == MemOp2RegOpTable.end())
3352 return false;
3353 unsigned Opc = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00003354 unsigned Index = I->second.second & TB_INDEX_MASK;
3355 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3356 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson43dbe052008-01-07 01:35:02 +00003357 if (UnfoldLoad && !FoldedLoad)
3358 return false;
3359 UnfoldLoad &= FoldedLoad;
3360 if (UnfoldStore && !FoldedStore)
3361 return false;
3362 UnfoldStore &= FoldedStore;
3363
Evan Chenge837dea2011-06-28 19:10:37 +00003364 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +00003365 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Cheng98ec91e2010-07-02 20:36:18 +00003366 if (!MI->hasOneMemOperand() &&
3367 RC == &X86::VR128RegClass &&
3368 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3369 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
3370 // conservatively assume the address is unaligned. That's bad for
3371 // performance.
3372 return false;
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00003373 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps;
Owen Anderson43dbe052008-01-07 01:35:02 +00003374 SmallVector<MachineOperand,2> BeforeOps;
3375 SmallVector<MachineOperand,2> AfterOps;
3376 SmallVector<MachineOperand,4> ImpOps;
3377 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
3378 MachineOperand &Op = MI->getOperand(i);
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00003379 if (i >= Index && i < Index + X86::AddrNumOperands)
Owen Anderson43dbe052008-01-07 01:35:02 +00003380 AddrOps.push_back(Op);
Dan Gohmand735b802008-10-03 15:45:36 +00003381 else if (Op.isReg() && Op.isImplicit())
Owen Anderson43dbe052008-01-07 01:35:02 +00003382 ImpOps.push_back(Op);
3383 else if (i < Index)
3384 BeforeOps.push_back(Op);
3385 else if (i > Index)
3386 AfterOps.push_back(Op);
3387 }
3388
3389 // Emit the load instruction.
3390 if (UnfoldLoad) {
Dan Gohman91e69c32009-10-09 18:10:05 +00003391 std::pair<MachineInstr::mmo_iterator,
3392 MachineInstr::mmo_iterator> MMOs =
3393 MF.extractLoadMemRefs(MI->memoperands_begin(),
3394 MI->memoperands_end());
3395 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson43dbe052008-01-07 01:35:02 +00003396 if (UnfoldStore) {
3397 // Address operands cannot be marked isKill.
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00003398 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) {
Owen Anderson43dbe052008-01-07 01:35:02 +00003399 MachineOperand &MO = NewMIs[0]->getOperand(i);
Dan Gohmand735b802008-10-03 15:45:36 +00003400 if (MO.isReg())
Owen Anderson43dbe052008-01-07 01:35:02 +00003401 MO.setIsKill(false);
3402 }
3403 }
3404 }
3405
3406 // Emit the data processing instruction.
Evan Chenge837dea2011-06-28 19:10:37 +00003407 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true);
Owen Anderson43dbe052008-01-07 01:35:02 +00003408 MachineInstrBuilder MIB(DataMI);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00003409
Owen Anderson43dbe052008-01-07 01:35:02 +00003410 if (FoldedStore)
Bill Wendling587daed2009-05-13 21:33:08 +00003411 MIB.addReg(Reg, RegState::Define);
Owen Anderson43dbe052008-01-07 01:35:02 +00003412 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00003413 MIB.addOperand(BeforeOps[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +00003414 if (FoldedLoad)
3415 MIB.addReg(Reg);
3416 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i)
Dan Gohman97357612009-02-18 05:45:50 +00003417 MIB.addOperand(AfterOps[i]);
Owen Anderson43dbe052008-01-07 01:35:02 +00003418 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) {
3419 MachineOperand &MO = ImpOps[i];
Bill Wendling587daed2009-05-13 21:33:08 +00003420 MIB.addReg(MO.getReg(),
3421 getDefRegState(MO.isDef()) |
3422 RegState::Implicit |
3423 getKillRegState(MO.isKill()) |
Evan Cheng4784f1f2009-06-30 08:49:04 +00003424 getDeadRegState(MO.isDead()) |
3425 getUndefRegState(MO.isUndef()));
Owen Anderson43dbe052008-01-07 01:35:02 +00003426 }
3427 // Change CMP32ri r, 0 back to TEST32rr r, r, etc.
3428 unsigned NewOpc = 0;
3429 switch (DataMI->getOpcode()) {
3430 default: break;
3431 case X86::CMP64ri32:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003432 case X86::CMP64ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +00003433 case X86::CMP32ri:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003434 case X86::CMP32ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +00003435 case X86::CMP16ri:
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003436 case X86::CMP16ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +00003437 case X86::CMP8ri: {
3438 MachineOperand &MO0 = DataMI->getOperand(0);
3439 MachineOperand &MO1 = DataMI->getOperand(1);
3440 if (MO1.getImm() == 0) {
3441 switch (DataMI->getOpcode()) {
3442 default: break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003443 case X86::CMP64ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +00003444 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003445 case X86::CMP32ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +00003446 case X86::CMP32ri: NewOpc = X86::TEST32rr; break;
Dan Gohmanf8c1ef02010-05-18 21:54:15 +00003447 case X86::CMP16ri8:
Owen Anderson43dbe052008-01-07 01:35:02 +00003448 case X86::CMP16ri: NewOpc = X86::TEST16rr; break;
3449 case X86::CMP8ri: NewOpc = X86::TEST8rr; break;
3450 }
Chris Lattner5080f4d2008-01-11 18:10:50 +00003451 DataMI->setDesc(get(NewOpc));
Owen Anderson43dbe052008-01-07 01:35:02 +00003452 MO1.ChangeToRegister(MO0.getReg(), false);
3453 }
3454 }
3455 }
3456 NewMIs.push_back(DataMI);
3457
3458 // Emit the store instruction.
3459 if (UnfoldStore) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +00003460 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF);
Dan Gohman91e69c32009-10-09 18:10:05 +00003461 std::pair<MachineInstr::mmo_iterator,
3462 MachineInstr::mmo_iterator> MMOs =
3463 MF.extractStoreMemRefs(MI->memoperands_begin(),
3464 MI->memoperands_end());
3465 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs);
Owen Anderson43dbe052008-01-07 01:35:02 +00003466 }
3467
3468 return true;
3469}
3470
3471bool
3472X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
Bill Wendlingfbef3102009-02-11 21:51:19 +00003473 SmallVectorImpl<SDNode*> &NewNodes) const {
Dan Gohmane8be6c62008-07-17 19:10:17 +00003474 if (!N->isMachineOpcode())
Owen Anderson43dbe052008-01-07 01:35:02 +00003475 return false;
3476
Chris Lattner45a1cb22010-10-07 23:08:41 +00003477 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3478 MemOp2RegOpTable.find(N->getMachineOpcode());
Owen Anderson43dbe052008-01-07 01:35:02 +00003479 if (I == MemOp2RegOpTable.end())
3480 return false;
3481 unsigned Opc = I->second.first;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00003482 unsigned Index = I->second.second & TB_INDEX_MASK;
3483 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3484 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Evan Chenge837dea2011-06-28 19:10:37 +00003485 const MCInstrDesc &MCID = get(Opc);
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +00003486 MachineFunction &MF = DAG.getMachineFunction();
3487 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
Evan Chenge837dea2011-06-28 19:10:37 +00003488 unsigned NumDefs = MCID.NumDefs;
Dan Gohman475871a2008-07-27 21:46:04 +00003489 std::vector<SDValue> AddrOps;
3490 std::vector<SDValue> BeforeOps;
3491 std::vector<SDValue> AfterOps;
Dale Johannesened2eee62009-02-06 01:31:28 +00003492 DebugLoc dl = N->getDebugLoc();
Owen Anderson43dbe052008-01-07 01:35:02 +00003493 unsigned NumOps = N->getNumOperands();
Dan Gohmanc76909a2009-09-25 20:36:54 +00003494 for (unsigned i = 0; i != NumOps-1; ++i) {
Dan Gohman475871a2008-07-27 21:46:04 +00003495 SDValue Op = N->getOperand(i);
Chris Lattnerac0ed5d2010-07-08 22:41:28 +00003496 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands)
Owen Anderson43dbe052008-01-07 01:35:02 +00003497 AddrOps.push_back(Op);
Dan Gohmanb37a8202009-03-04 19:23:38 +00003498 else if (i < Index-NumDefs)
Owen Anderson43dbe052008-01-07 01:35:02 +00003499 BeforeOps.push_back(Op);
Dan Gohmanb37a8202009-03-04 19:23:38 +00003500 else if (i > Index-NumDefs)
Owen Anderson43dbe052008-01-07 01:35:02 +00003501 AfterOps.push_back(Op);
3502 }
Dan Gohman475871a2008-07-27 21:46:04 +00003503 SDValue Chain = N->getOperand(NumOps-1);
Owen Anderson43dbe052008-01-07 01:35:02 +00003504 AddrOps.push_back(Chain);
3505
3506 // Emit the load instruction.
3507 SDNode *Load = 0;
3508 if (FoldedLoad) {
Owen Andersone50ed302009-08-10 22:56:29 +00003509 EVT VT = *RC->vt_begin();
Evan Cheng600c0432009-11-16 21:56:03 +00003510 std::pair<MachineInstr::mmo_iterator,
3511 MachineInstr::mmo_iterator> MMOs =
3512 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
3513 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng98ec91e2010-07-02 20:36:18 +00003514 if (!(*MMOs.first) &&
3515 RC == &X86::VR128RegClass &&
3516 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3517 // Do not introduce a slow unaligned load.
3518 return false;
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00003519 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3520 bool isAligned = (*MMOs.first) &&
3521 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman602b0c82009-09-25 18:54:59 +00003522 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl,
3523 VT, MVT::Other, &AddrOps[0], AddrOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +00003524 NewNodes.push_back(Load);
Dan Gohman91e69c32009-10-09 18:10:05 +00003525
3526 // Preserve memory reference information.
Dan Gohman91e69c32009-10-09 18:10:05 +00003527 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson43dbe052008-01-07 01:35:02 +00003528 }
3529
3530 // Emit the data processing instruction.
Owen Andersone50ed302009-08-10 22:56:29 +00003531 std::vector<EVT> VTs;
Owen Anderson43dbe052008-01-07 01:35:02 +00003532 const TargetRegisterClass *DstRC = 0;
Evan Chenge837dea2011-06-28 19:10:37 +00003533 if (MCID.getNumDefs() > 0) {
Jakob Stoklund Olesen397fc482012-05-07 22:10:26 +00003534 DstRC = getRegClass(MCID, 0, &RI, MF);
Owen Anderson43dbe052008-01-07 01:35:02 +00003535 VTs.push_back(*DstRC->vt_begin());
3536 }
3537 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
Owen Andersone50ed302009-08-10 22:56:29 +00003538 EVT VT = N->getValueType(i);
Evan Chenge837dea2011-06-28 19:10:37 +00003539 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs())
Owen Anderson43dbe052008-01-07 01:35:02 +00003540 VTs.push_back(VT);
3541 }
3542 if (Load)
Dan Gohman475871a2008-07-27 21:46:04 +00003543 BeforeOps.push_back(SDValue(Load, 0));
Owen Anderson43dbe052008-01-07 01:35:02 +00003544 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps));
Dan Gohman602b0c82009-09-25 18:54:59 +00003545 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0],
3546 BeforeOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +00003547 NewNodes.push_back(NewNode);
3548
3549 // Emit the store instruction.
3550 if (FoldedStore) {
3551 AddrOps.pop_back();
Dan Gohman475871a2008-07-27 21:46:04 +00003552 AddrOps.push_back(SDValue(NewNode, 0));
Owen Anderson43dbe052008-01-07 01:35:02 +00003553 AddrOps.push_back(Chain);
Evan Cheng600c0432009-11-16 21:56:03 +00003554 std::pair<MachineInstr::mmo_iterator,
3555 MachineInstr::mmo_iterator> MMOs =
3556 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(),
3557 cast<MachineSDNode>(N)->memoperands_end());
Evan Cheng98ec91e2010-07-02 20:36:18 +00003558 if (!(*MMOs.first) &&
3559 RC == &X86::VR128RegClass &&
3560 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast())
3561 // Do not introduce a slow unaligned store.
3562 return false;
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00003563 unsigned Alignment = RC->getSize() == 32 ? 32 : 16;
3564 bool isAligned = (*MMOs.first) &&
3565 (*MMOs.first)->getAlignment() >= Alignment;
Dan Gohman602b0c82009-09-25 18:54:59 +00003566 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC,
3567 isAligned, TM),
3568 dl, MVT::Other,
3569 &AddrOps[0], AddrOps.size());
Owen Anderson43dbe052008-01-07 01:35:02 +00003570 NewNodes.push_back(Store);
Dan Gohman91e69c32009-10-09 18:10:05 +00003571
3572 // Preserve memory reference information.
Dan Gohman91e69c32009-10-09 18:10:05 +00003573 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second);
Owen Anderson43dbe052008-01-07 01:35:02 +00003574 }
3575
3576 return true;
3577}
3578
3579unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
Dan Gohman0115e162009-10-30 22:18:41 +00003580 bool UnfoldLoad, bool UnfoldStore,
3581 unsigned *LoadRegIndex) const {
Chris Lattner45a1cb22010-10-07 23:08:41 +00003582 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I =
3583 MemOp2RegOpTable.find(Opc);
Owen Anderson43dbe052008-01-07 01:35:02 +00003584 if (I == MemOp2RegOpTable.end())
3585 return 0;
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00003586 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD;
3587 bool FoldedStore = I->second.second & TB_FOLDED_STORE;
Owen Anderson43dbe052008-01-07 01:35:02 +00003588 if (UnfoldLoad && !FoldedLoad)
3589 return 0;
3590 if (UnfoldStore && !FoldedStore)
3591 return 0;
Dan Gohman0115e162009-10-30 22:18:41 +00003592 if (LoadRegIndex)
Bruno Cardoso Lopescbf479d2011-09-08 18:35:57 +00003593 *LoadRegIndex = I->second.second & TB_INDEX_MASK;
Owen Anderson43dbe052008-01-07 01:35:02 +00003594 return I->second.first;
3595}
3596
Evan Cheng96dc1152010-01-22 03:34:51 +00003597bool
3598X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
3599 int64_t &Offset1, int64_t &Offset2) const {
3600 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode())
3601 return false;
3602 unsigned Opc1 = Load1->getMachineOpcode();
3603 unsigned Opc2 = Load2->getMachineOpcode();
3604 switch (Opc1) {
3605 default: return false;
3606 case X86::MOV8rm:
3607 case X86::MOV16rm:
3608 case X86::MOV32rm:
3609 case X86::MOV64rm:
3610 case X86::LD_Fp32m:
3611 case X86::LD_Fp64m:
3612 case X86::LD_Fp80m:
3613 case X86::MOVSSrm:
3614 case X86::MOVSDrm:
3615 case X86::MMX_MOVD64rm:
3616 case X86::MMX_MOVQ64rm:
3617 case X86::FsMOVAPSrm:
3618 case X86::FsMOVAPDrm:
3619 case X86::MOVAPSrm:
3620 case X86::MOVUPSrm:
Evan Cheng96dc1152010-01-22 03:34:51 +00003621 case X86::MOVAPDrm:
3622 case X86::MOVDQArm:
3623 case X86::MOVDQUrm:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +00003624 // AVX load instructions
3625 case X86::VMOVSSrm:
3626 case X86::VMOVSDrm:
3627 case X86::FsVMOVAPSrm:
3628 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00003629 case X86::VMOVAPSrm:
3630 case X86::VMOVUPSrm:
3631 case X86::VMOVAPDrm:
3632 case X86::VMOVDQArm:
3633 case X86::VMOVDQUrm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00003634 case X86::VMOVAPSYrm:
3635 case X86::VMOVUPSYrm:
3636 case X86::VMOVAPDYrm:
3637 case X86::VMOVDQAYrm:
3638 case X86::VMOVDQUYrm:
Evan Cheng96dc1152010-01-22 03:34:51 +00003639 break;
3640 }
3641 switch (Opc2) {
3642 default: return false;
3643 case X86::MOV8rm:
3644 case X86::MOV16rm:
3645 case X86::MOV32rm:
3646 case X86::MOV64rm:
3647 case X86::LD_Fp32m:
3648 case X86::LD_Fp64m:
3649 case X86::LD_Fp80m:
3650 case X86::MOVSSrm:
3651 case X86::MOVSDrm:
3652 case X86::MMX_MOVD64rm:
3653 case X86::MMX_MOVQ64rm:
3654 case X86::FsMOVAPSrm:
3655 case X86::FsMOVAPDrm:
3656 case X86::MOVAPSrm:
3657 case X86::MOVUPSrm:
Evan Cheng96dc1152010-01-22 03:34:51 +00003658 case X86::MOVAPDrm:
3659 case X86::MOVDQArm:
3660 case X86::MOVDQUrm:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +00003661 // AVX load instructions
3662 case X86::VMOVSSrm:
3663 case X86::VMOVSDrm:
3664 case X86::FsVMOVAPSrm:
3665 case X86::FsVMOVAPDrm:
Bruno Cardoso Lopes484ddf52011-09-14 02:36:58 +00003666 case X86::VMOVAPSrm:
3667 case X86::VMOVUPSrm:
3668 case X86::VMOVAPDrm:
3669 case X86::VMOVDQArm:
3670 case X86::VMOVDQUrm:
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00003671 case X86::VMOVAPSYrm:
3672 case X86::VMOVUPSYrm:
3673 case X86::VMOVAPDYrm:
3674 case X86::VMOVDQAYrm:
3675 case X86::VMOVDQUYrm:
Evan Cheng96dc1152010-01-22 03:34:51 +00003676 break;
3677 }
3678
3679 // Check if chain operands and base addresses match.
3680 if (Load1->getOperand(0) != Load2->getOperand(0) ||
3681 Load1->getOperand(5) != Load2->getOperand(5))
3682 return false;
3683 // Segment operands should match as well.
3684 if (Load1->getOperand(4) != Load2->getOperand(4))
3685 return false;
3686 // Scale should be 1, Index should be Reg0.
3687 if (Load1->getOperand(1) == Load2->getOperand(1) &&
3688 Load1->getOperand(2) == Load2->getOperand(2)) {
3689 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1)
3690 return false;
Evan Cheng96dc1152010-01-22 03:34:51 +00003691
3692 // Now let's examine the displacements.
3693 if (isa<ConstantSDNode>(Load1->getOperand(3)) &&
3694 isa<ConstantSDNode>(Load2->getOperand(3))) {
3695 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue();
3696 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue();
3697 return true;
3698 }
3699 }
3700 return false;
3701}
3702
3703bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
3704 int64_t Offset1, int64_t Offset2,
3705 unsigned NumLoads) const {
3706 assert(Offset2 > Offset1);
3707 if ((Offset2 - Offset1) / 8 > 64)
3708 return false;
3709
3710 unsigned Opc1 = Load1->getMachineOpcode();
3711 unsigned Opc2 = Load2->getMachineOpcode();
3712 if (Opc1 != Opc2)
3713 return false; // FIXME: overly conservative?
3714
3715 switch (Opc1) {
3716 default: break;
3717 case X86::LD_Fp32m:
3718 case X86::LD_Fp64m:
3719 case X86::LD_Fp80m:
3720 case X86::MMX_MOVD64rm:
3721 case X86::MMX_MOVQ64rm:
3722 return false;
3723 }
3724
3725 EVT VT = Load1->getValueType(0);
3726 switch (VT.getSimpleVT().SimpleTy) {
Bill Wendling19d85972010-06-22 22:16:17 +00003727 default:
Evan Cheng96dc1152010-01-22 03:34:51 +00003728 // XMM registers. In 64-bit mode we can be a bit more aggressive since we
3729 // have 16 of them to play with.
3730 if (TM.getSubtargetImpl()->is64Bit()) {
3731 if (NumLoads >= 3)
3732 return false;
Bill Wendling19d85972010-06-22 22:16:17 +00003733 } else if (NumLoads) {
Evan Cheng96dc1152010-01-22 03:34:51 +00003734 return false;
Bill Wendling19d85972010-06-22 22:16:17 +00003735 }
Evan Cheng96dc1152010-01-22 03:34:51 +00003736 break;
Evan Cheng96dc1152010-01-22 03:34:51 +00003737 case MVT::i8:
3738 case MVT::i16:
3739 case MVT::i32:
3740 case MVT::i64:
Evan Chengafc36732010-01-22 23:49:11 +00003741 case MVT::f32:
3742 case MVT::f64:
Evan Cheng96dc1152010-01-22 03:34:51 +00003743 if (NumLoads)
3744 return false;
Bill Wendling19d85972010-06-22 22:16:17 +00003745 break;
Evan Cheng96dc1152010-01-22 03:34:51 +00003746 }
3747
3748 return true;
3749}
3750
3751
Chris Lattner7fbe9722006-10-20 17:42:20 +00003752bool X86InstrInfo::
Owen Anderson44eb65c2008-08-14 22:49:33 +00003753ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
Chris Lattner9cd68752006-10-21 05:52:40 +00003754 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
Evan Cheng97af60b2008-08-29 23:21:31 +00003755 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm());
Dan Gohman279c22e2008-10-21 03:29:32 +00003756 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E)
3757 return true;
Evan Cheng97af60b2008-08-29 23:21:31 +00003758 Cond[0].setImm(GetOppositeBranchCondition(CC));
Chris Lattner9cd68752006-10-21 05:52:40 +00003759 return false;
Chris Lattner7fbe9722006-10-20 17:42:20 +00003760}
3761
Evan Cheng23066282008-10-27 07:14:50 +00003762bool X86InstrInfo::
Evan Cheng4350eb82009-02-06 17:17:30 +00003763isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
3764 // FIXME: Return false for x87 stack register classes for now. We can't
Evan Cheng23066282008-10-27 07:14:50 +00003765 // allow any loads of these registers before FpGet_ST0_80.
Evan Cheng4350eb82009-02-06 17:17:30 +00003766 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass ||
3767 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass);
Evan Cheng23066282008-10-27 07:14:50 +00003768}
3769
Dan Gohman57c3dac2008-09-30 00:58:23 +00003770/// getGlobalBaseReg - Return a virtual register initialized with the
3771/// the global base register value. Output instructions required to
3772/// initialize the register in the function entry block, if necessary.
Dan Gohman8b746962008-09-23 18:22:58 +00003773///
Dan Gohman84023e02010-07-10 09:00:22 +00003774/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
3775///
Dan Gohman57c3dac2008-09-30 00:58:23 +00003776unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
3777 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
3778 "X86-64 PIC uses RIP relative addressing");
3779
3780 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
3781 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3782 if (GlobalBaseReg != 0)
3783 return GlobalBaseReg;
3784
Dan Gohman84023e02010-07-10 09:00:22 +00003785 // Create the register. The code to initialize it is inserted
3786 // later, by the CGBR pass (below).
Dan Gohman8b746962008-09-23 18:22:58 +00003787 MachineRegisterInfo &RegInfo = MF->getRegInfo();
Jakob Stoklund Olesen53df9252012-05-20 18:43:00 +00003788 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass);
Dan Gohman57c3dac2008-09-30 00:58:23 +00003789 X86FI->setGlobalBaseReg(GlobalBaseReg);
3790 return GlobalBaseReg;
Dan Gohman8b746962008-09-23 18:22:58 +00003791}
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +00003792
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003793// These are the replaceable SSE instructions. Some of these have Int variants
3794// that we don't include here. We don't want to replace instructions selected
3795// by intrinsics.
Craig Topper72051bf2012-03-09 07:45:21 +00003796static const uint16_t ReplaceableInstrs[][3] = {
Bruno Cardoso Lopes4d043622010-08-12 02:08:52 +00003797 //PackedSingle PackedDouble PackedInt
Jakob Stoklund Olesen357be7f2010-03-30 22:46:53 +00003798 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr },
3799 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm },
3800 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr },
3801 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr },
3802 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm },
3803 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr },
3804 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm },
3805 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr },
3806 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm },
3807 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr },
3808 { X86::ORPSrm, X86::ORPDrm, X86::PORrm },
3809 { X86::ORPSrr, X86::ORPDrr, X86::PORrr },
3810 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm },
3811 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr },
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003812 // AVX 128-bit support
3813 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr },
3814 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm },
3815 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr },
3816 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr },
3817 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm },
3818 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr },
3819 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm },
3820 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr },
3821 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm },
3822 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr },
3823 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm },
3824 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr },
Bruno Cardoso Lopes642eb022010-08-12 20:20:53 +00003825 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm },
3826 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr },
Bruno Cardoso Lopes62f67f82011-07-14 18:50:58 +00003827 // AVX 256-bit support
3828 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr },
3829 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm },
3830 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr },
3831 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr },
3832 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm },
Craig Topper4c077a12011-11-15 05:55:35 +00003833 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }
3834};
3835
Craig Topper72051bf2012-03-09 07:45:21 +00003836static const uint16_t ReplaceableInstrsAVX2[][3] = {
Craig Topper4c077a12011-11-15 05:55:35 +00003837 //PackedSingle PackedDouble PackedInt
Craig Topperb80ada92011-11-09 09:37:21 +00003838 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm },
3839 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr },
3840 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm },
3841 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr },
3842 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm },
3843 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr },
3844 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm },
Craig Topperfe2a6c52011-11-29 05:37:58 +00003845 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr },
3846 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr },
3847 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr },
3848 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
3849 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
3850 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
3851 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003852};
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +00003853
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003854// FIXME: Some shuffle and unpack instructions have equivalents in different
3855// domains, but they require a bit more work than just switching opcodes.
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +00003856
Craig Topper72051bf2012-03-09 07:45:21 +00003857static const uint16_t *lookup(unsigned opcode, unsigned domain) {
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +00003858 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i)
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003859 if (ReplaceableInstrs[i][domain-1] == opcode)
3860 return ReplaceableInstrs[i];
Craig Topper44ec9fd2011-11-15 06:39:01 +00003861 return 0;
3862}
3863
Craig Topper72051bf2012-03-09 07:45:21 +00003864static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) {
Craig Topper44ec9fd2011-11-15 06:39:01 +00003865 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i)
3866 if (ReplaceableInstrsAVX2[i][domain-1] == opcode)
3867 return ReplaceableInstrsAVX2[i];
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003868 return 0;
3869}
3870
3871std::pair<uint16_t, uint16_t>
Jakob Stoklund Olesen98e933f2011-09-27 22:57:18 +00003872X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const {
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003873 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
Craig Topper4c077a12011-11-15 05:55:35 +00003874 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2();
Craig Topper44ec9fd2011-11-15 06:39:01 +00003875 uint16_t validDomains = 0;
3876 if (domain && lookup(MI->getOpcode(), domain))
3877 validDomains = 0xe;
3878 else if (domain && lookupAVX2(MI->getOpcode(), domain))
3879 validDomains = hasAVX2 ? 0xe : 0x6;
3880 return std::make_pair(domain, validDomains);
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003881}
3882
Jakob Stoklund Olesen98e933f2011-09-27 22:57:18 +00003883void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003884 assert(Domain>0 && Domain<4 && "Invalid execution domain");
3885 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3;
3886 assert(dom && "Not an SSE instruction");
Craig Topper72051bf2012-03-09 07:45:21 +00003887 const uint16_t *table = lookup(MI->getOpcode(), dom);
Jakob Stoklund Olesen7f5e43f2011-11-23 04:03:08 +00003888 if (!table) { // try the other table
3889 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) &&
3890 "256-bit vector operations only available in AVX2");
Craig Topper44ec9fd2011-11-15 06:39:01 +00003891 table = lookupAVX2(MI->getOpcode(), dom);
Jakob Stoklund Olesen7f5e43f2011-11-23 04:03:08 +00003892 }
Jakob Stoklund Olesene4b94b42010-03-29 23:24:21 +00003893 assert(table && "Cannot change domain");
3894 MI->setDesc(get(table[Domain-1]));
Jakob Stoklund Olesen352aa502010-03-25 17:25:00 +00003895}
Chris Lattneree9eb412010-04-26 23:37:21 +00003896
3897/// getNoopForMachoTarget - Return the noop instruction to use for a noop.
3898void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
3899 NopInst.setOpcode(X86::NOOP);
3900}
Dan Gohman84023e02010-07-10 09:00:22 +00003901
Andrew Tricke0ef5092011-03-05 08:00:22 +00003902bool X86InstrInfo::isHighLatencyDef(int opc) const {
3903 switch (opc) {
Evan Cheng23128422010-10-19 18:58:51 +00003904 default: return false;
3905 case X86::DIVSDrm:
3906 case X86::DIVSDrm_Int:
3907 case X86::DIVSDrr:
3908 case X86::DIVSDrr_Int:
3909 case X86::DIVSSrm:
3910 case X86::DIVSSrm_Int:
3911 case X86::DIVSSrr:
3912 case X86::DIVSSrr_Int:
3913 case X86::SQRTPDm:
3914 case X86::SQRTPDm_Int:
3915 case X86::SQRTPDr:
3916 case X86::SQRTPDr_Int:
3917 case X86::SQRTPSm:
3918 case X86::SQRTPSm_Int:
3919 case X86::SQRTPSr:
3920 case X86::SQRTPSr_Int:
3921 case X86::SQRTSDm:
3922 case X86::SQRTSDm_Int:
3923 case X86::SQRTSDr:
3924 case X86::SQRTSDr_Int:
3925 case X86::SQRTSSm:
3926 case X86::SQRTSSm_Int:
3927 case X86::SQRTSSr:
3928 case X86::SQRTSSr_Int:
Bruno Cardoso Lopesb4e905d2011-09-15 22:15:52 +00003929 // AVX instructions with high latency
3930 case X86::VDIVSDrm:
3931 case X86::VDIVSDrm_Int:
3932 case X86::VDIVSDrr:
3933 case X86::VDIVSDrr_Int:
3934 case X86::VDIVSSrm:
3935 case X86::VDIVSSrm_Int:
3936 case X86::VDIVSSrr:
3937 case X86::VDIVSSrr_Int:
3938 case X86::VSQRTPDm:
3939 case X86::VSQRTPDm_Int:
3940 case X86::VSQRTPDr:
3941 case X86::VSQRTPDr_Int:
3942 case X86::VSQRTPSm:
3943 case X86::VSQRTPSm_Int:
3944 case X86::VSQRTPSr:
3945 case X86::VSQRTPSr_Int:
3946 case X86::VSQRTSDm:
3947 case X86::VSQRTSDm_Int:
3948 case X86::VSQRTSDr:
3949 case X86::VSQRTSSm:
3950 case X86::VSQRTSSm_Int:
3951 case X86::VSQRTSSr:
Evan Cheng23128422010-10-19 18:58:51 +00003952 return true;
3953 }
3954}
3955
Andrew Tricke0ef5092011-03-05 08:00:22 +00003956bool X86InstrInfo::
3957hasHighOperandLatency(const InstrItineraryData *ItinData,
3958 const MachineRegisterInfo *MRI,
3959 const MachineInstr *DefMI, unsigned DefIdx,
3960 const MachineInstr *UseMI, unsigned UseIdx) const {
3961 return isHighLatencyDef(DefMI->getOpcode());
3962}
3963
Dan Gohman84023e02010-07-10 09:00:22 +00003964namespace {
3965 /// CGBR - Create Global Base Reg pass. This initializes the PIC
3966 /// global base register for x86-32.
3967 struct CGBR : public MachineFunctionPass {
3968 static char ID;
Owen Anderson90c579d2010-08-06 18:33:48 +00003969 CGBR() : MachineFunctionPass(ID) {}
Dan Gohman84023e02010-07-10 09:00:22 +00003970
3971 virtual bool runOnMachineFunction(MachineFunction &MF) {
3972 const X86TargetMachine *TM =
3973 static_cast<const X86TargetMachine *>(&MF.getTarget());
3974
3975 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
3976 "X86-64 PIC uses RIP relative addressing");
3977
3978 // Only emit a global base reg in PIC mode.
3979 if (TM->getRelocationModel() != Reloc::PIC_)
3980 return false;
3981
Dan Gohmand8c0a512010-09-17 20:24:24 +00003982 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
3983 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg();
3984
3985 // If we didn't need a GlobalBaseReg, don't insert code.
3986 if (GlobalBaseReg == 0)
3987 return false;
3988
Dan Gohman84023e02010-07-10 09:00:22 +00003989 // Insert the set of GlobalBaseReg into the first MBB of the function
3990 MachineBasicBlock &FirstMBB = MF.front();
3991 MachineBasicBlock::iterator MBBI = FirstMBB.begin();
3992 DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
3993 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3994 const X86InstrInfo *TII = TM->getInstrInfo();
3995
3996 unsigned PC;
3997 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
Craig Topperc9099502012-04-20 06:31:50 +00003998 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass);
Dan Gohman84023e02010-07-10 09:00:22 +00003999 else
Dan Gohmand8c0a512010-09-17 20:24:24 +00004000 PC = GlobalBaseReg;
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00004001
Dan Gohman84023e02010-07-10 09:00:22 +00004002 // Operand of MovePCtoStack is completely ignored by asm printer. It's
4003 // only used in JIT code emission as displacement to pc.
4004 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
NAKAMURA Takumie5fffe92011-01-26 02:03:37 +00004005
Dan Gohman84023e02010-07-10 09:00:22 +00004006 // If we're using vanilla 'GOT' PIC style, we should use relative addressing
4007 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
4008 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
Dan Gohman84023e02010-07-10 09:00:22 +00004009 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
4010 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
4011 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
4012 X86II::MO_GOT_ABSOLUTE_ADDRESS);
4013 }
4014
4015 return true;
4016 }
4017
4018 virtual const char *getPassName() const {
4019 return "X86 PIC Global Base Reg Initialization";
4020 }
4021
4022 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4023 AU.setPreservesCFG();
4024 MachineFunctionPass::getAnalysisUsage(AU);
4025 }
4026 };
4027}
4028
4029char CGBR::ID = 0;
4030FunctionPass*
4031llvm::createGlobalBaseRegPass() { return new CGBR(); }
Hans Wennborgf0234fc2012-06-01 16:27:21 +00004032
4033namespace {
4034 struct LDTLSCleanup : public MachineFunctionPass {
4035 static char ID;
4036 LDTLSCleanup() : MachineFunctionPass(ID) {}
4037
4038 virtual bool runOnMachineFunction(MachineFunction &MF) {
4039 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>();
4040 if (MFI->getNumLocalDynamicTLSAccesses() < 2) {
4041 // No point folding accesses if there isn't at least two.
4042 return false;
4043 }
4044
4045 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>();
4046 return VisitNode(DT->getRootNode(), 0);
4047 }
4048
4049 // Visit the dominator subtree rooted at Node in pre-order.
4050 // If TLSBaseAddrReg is non-null, then use that to replace any
4051 // TLS_base_addr instructions. Otherwise, create the register
4052 // when the first such instruction is seen, and then use it
4053 // as we encounter more instructions.
4054 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) {
4055 MachineBasicBlock *BB = Node->getBlock();
4056 bool Changed = false;
4057
4058 // Traverse the current block.
4059 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;
4060 ++I) {
4061 switch (I->getOpcode()) {
4062 case X86::TLS_base_addr32:
4063 case X86::TLS_base_addr64:
4064 if (TLSBaseAddrReg)
4065 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg);
4066 else
4067 I = SetRegister(I, &TLSBaseAddrReg);
4068 Changed = true;
4069 break;
4070 default:
4071 break;
4072 }
4073 }
4074
4075 // Visit the children of this block in the dominator tree.
4076 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end();
4077 I != E; ++I) {
4078 Changed |= VisitNode(*I, TLSBaseAddrReg);
4079 }
4080
4081 return Changed;
4082 }
4083
4084 // Replace the TLS_base_addr instruction I with a copy from
4085 // TLSBaseAddrReg, returning the new instruction.
4086 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I,
4087 unsigned TLSBaseAddrReg) {
4088 MachineFunction *MF = I->getParent()->getParent();
4089 const X86TargetMachine *TM =
4090 static_cast<const X86TargetMachine *>(&MF->getTarget());
4091 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4092 const X86InstrInfo *TII = TM->getInstrInfo();
4093
4094 // Insert a Copy from TLSBaseAddrReg to RAX/EAX.
4095 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(),
4096 TII->get(TargetOpcode::COPY),
4097 is64Bit ? X86::RAX : X86::EAX)
4098 .addReg(TLSBaseAddrReg);
4099
4100 // Erase the TLS_base_addr instruction.
4101 I->eraseFromParent();
4102
4103 return Copy;
4104 }
4105
4106 // Create a virtal register in *TLSBaseAddrReg, and populate it by
4107 // inserting a copy instruction after I. Returns the new instruction.
4108 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) {
4109 MachineFunction *MF = I->getParent()->getParent();
4110 const X86TargetMachine *TM =
4111 static_cast<const X86TargetMachine *>(&MF->getTarget());
4112 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit();
4113 const X86InstrInfo *TII = TM->getInstrInfo();
4114
4115 // Create a virtual register for the TLS base address.
4116 MachineRegisterInfo &RegInfo = MF->getRegInfo();
4117 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit
4118 ? &X86::GR64RegClass
4119 : &X86::GR32RegClass);
4120
4121 // Insert a copy from RAX/EAX to TLSBaseAddrReg.
4122 MachineInstr *Next = I->getNextNode();
4123 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(),
4124 TII->get(TargetOpcode::COPY),
4125 *TLSBaseAddrReg)
4126 .addReg(is64Bit ? X86::RAX : X86::EAX);
4127
4128 return Copy;
4129 }
4130
4131 virtual const char *getPassName() const {
4132 return "Local Dynamic TLS Access Clean-up";
4133 }
4134
4135 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
4136 AU.setPreservesCFG();
4137 AU.addRequired<MachineDominatorTree>();
4138 MachineFunctionPass::getAnalysisUsage(AU);
4139 }
4140 };
4141}
4142
4143char LDTLSCleanup::ID = 0;
4144FunctionPass*
4145llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); }